BASELINE: Update Chromium to 104.0.5112.120

Change-Id: I5d2726c2ab018d75d055739b6ba64317904f05bb
Reviewed-on: https://codereview.qt-project.org/c/qt/qtwebengine-chromium/+/438935
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>

37 files changed, 10234 insertions, 1035 deletions

diff --git a/chromium/third_party/libaom/source/config/config/aom_version.h b/chromium/third_party/libaom/source/config/config/aom_version.h
index 790736ffafa..3f163b8db5f 100644
--- a/chromium/third_party/libaom/source/config/config/aom_version.h
+++ b/chromium/third_party/libaom/source/config/config/aom_version.h
@@ -12,8 +12,8 @@
 #define VERSION_MAJOR 3
 #define VERSION_MINOR 3
 #define VERSION_PATCH 0
-#define VERSION_EXTRA "475-ge24a83a72"
+#define VERSION_EXTRA "648-g32d6783c4"
 #define VERSION_PACKED \
   ((VERSION_MAJOR << 16) | (VERSION_MINOR << 8) | (VERSION_PATCH))
-#define VERSION_STRING_NOSP "3.3.0-475-ge24a83a72"
-#define VERSION_STRING " 3.3.0-475-ge24a83a72"
+#define VERSION_STRING_NOSP "3.3.0-648-g32d6783c4"
+#define VERSION_STRING " 3.3.0-648-g32d6783c4"
diff --git a/chromium/third_party/libaom/source/config/ios/arm-neon/config/aom_config.c b/chromium/third_party/libaom/source/config/ios/arm-neon/config/aom_config.c
index b1084c5aef7..e81aaa8bc76 100644
--- a/chromium/third_party/libaom/source/config/ios/arm-neon/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/ios/arm-neon/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/armv7-ios.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/armv7-ios.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/ios/arm-neon/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/ios/arm-neon/config/aom_dsp_rtcd.h
index 155c11857c0..49fb79d7f18 100644
--- a/chromium/third_party/libaom/source/config/ios/arm-neon/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/ios/arm-neon/config/aom_dsp_rtcd.h
@@ -34,7 +34,12 @@ void aom_avg_8x8_quad_c(const uint8_t* s,
                         int x16_idx,
                         int y16_idx,
                         int* avg);
-#define aom_avg_8x8_quad aom_avg_8x8_quad_c
+void aom_avg_8x8_quad_neon(const uint8_t* s,
+                           int p,
+                           int x16_idx,
+                           int y16_idx,
+                           int* avg);
+#define aom_avg_8x8_quad aom_avg_8x8_quad_neon
 
 void aom_blend_a64_hmask_c(uint8_t* dst,
                            uint32_t dst_stride,
@@ -176,18 +181,24 @@ void aom_dc_128_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_c
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_c
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -210,6 +221,18 @@ void aom_dc_128_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_32x64 aom_dc_128_predictor_32x64_c
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_c
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_c
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -226,6 +249,12 @@ void aom_dc_128_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_64x16 aom_dc_128_predictor_64x16_c
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -244,6 +273,12 @@ void aom_dc_128_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_c
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -276,18 +311,24 @@ void aom_dc_left_predictor_16x32_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_c
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_c
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -310,6 +351,18 @@ void aom_dc_left_predictor_32x64_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_32x64 aom_dc_left_predictor_32x64_c
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_c
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_c
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -326,6 +379,12 @@ void aom_dc_left_predictor_4x8_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_64x16 aom_dc_left_predictor_64x16_c
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -344,6 +403,12 @@ void aom_dc_left_predictor_8x16_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_c
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -376,18 +441,24 @@ void aom_dc_predictor_16x32_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_c
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_c
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
                              const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -410,6 +481,18 @@ void aom_dc_predictor_32x64_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_32x64 aom_dc_predictor_32x64_c
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_c
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_c
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -426,6 +509,12 @@ void aom_dc_predictor_4x8_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_64x16 aom_dc_predictor_64x16_c
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -444,6 +533,12 @@ void aom_dc_predictor_8x16_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_c
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -476,18 +571,24 @@ void aom_dc_top_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_c
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_c
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -510,6 +611,18 @@ void aom_dc_top_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_32x64 aom_dc_top_predictor_32x64_c
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_c
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_c
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -526,6 +639,12 @@ void aom_dc_top_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_64x16 aom_dc_top_predictor_64x16_c
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -544,6 +663,12 @@ void aom_dc_top_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_c
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -1014,18 +1139,24 @@ void aom_h_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_c
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_c
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_c
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_c
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1048,6 +1179,18 @@ void aom_h_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_c
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_c
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_c
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1064,6 +1207,12 @@ void aom_h_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_c
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_c
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1082,6 +1231,12 @@ void aom_h_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_c
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_c
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1127,7 +1282,10 @@ void aom_hadamard_8x8_neon(const int16_t* src_diff,
 void aom_hadamard_8x8_dual_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
                              int16_t* coeff);
-#define aom_hadamard_8x8_dual aom_hadamard_8x8_dual_c
+void aom_hadamard_8x8_dual_neon(const int16_t* src_diff,
+                                ptrdiff_t src_stride,
+                                int16_t* coeff);
+#define aom_hadamard_8x8_dual aom_hadamard_8x8_dual_neon
 
 void aom_hadamard_lp_16x16_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
@@ -1237,7 +1395,12 @@ void aom_lpf_horizontal_14_quad_c(uint8_t* s,
                                   const uint8_t* blimit0,
                                   const uint8_t* limit0,
                                   const uint8_t* thresh0);
-#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_c
+void aom_lpf_horizontal_14_quad_neon(uint8_t* s,
+                                     int pitch,
+                                     const uint8_t* blimit0,
+                                     const uint8_t* limit0,
+                                     const uint8_t* thresh0);
+#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_neon
 
 void aom_lpf_horizontal_4_c(uint8_t* s,
                             int pitch,
@@ -1274,7 +1437,12 @@ void aom_lpf_horizontal_4_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_c
+void aom_lpf_horizontal_4_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_neon
 
 void aom_lpf_horizontal_6_c(uint8_t* s,
                             int pitch,
@@ -1311,7 +1479,12 @@ void aom_lpf_horizontal_6_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_c
+void aom_lpf_horizontal_6_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_neon
 
 void aom_lpf_horizontal_8_c(uint8_t* s,
                             int pitch,
@@ -1348,7 +1521,12 @@ void aom_lpf_horizontal_8_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_c
+void aom_lpf_horizontal_8_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_neon
 
 void aom_lpf_vertical_14_c(uint8_t* s,
                            int pitch,
@@ -1385,7 +1563,12 @@ void aom_lpf_vertical_14_quad_c(uint8_t* s,
                                 const uint8_t* blimit0,
                                 const uint8_t* limit0,
                                 const uint8_t* thresh0);
-#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_c
+void aom_lpf_vertical_14_quad_neon(uint8_t* s,
+                                   int pitch,
+                                   const uint8_t* blimit0,
+                                   const uint8_t* limit0,
+                                   const uint8_t* thresh0);
+#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_neon
 
 void aom_lpf_vertical_4_c(uint8_t* s,
                           int pitch,
@@ -1422,7 +1605,12 @@ void aom_lpf_vertical_4_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_c
+void aom_lpf_vertical_4_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_neon
 
 void aom_lpf_vertical_6_c(uint8_t* s,
                           int pitch,
@@ -1459,7 +1647,12 @@ void aom_lpf_vertical_6_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_c
+void aom_lpf_vertical_6_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_neon
 
 void aom_lpf_vertical_8_c(uint8_t* s,
                           int pitch,
@@ -1496,7 +1689,12 @@ void aom_lpf_vertical_8_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_c
+void aom_lpf_vertical_8_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_neon
 
 unsigned int aom_masked_sad128x128_c(const uint8_t* src,
                                      int src_stride,
@@ -2098,85 +2296,191 @@ void aom_paeth_predictor_16x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_c
+void aom_paeth_predictor_16x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_neon
 
 void aom_paeth_predictor_16x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
+void aom_paeth_predictor_16x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_neon
 
-void aom_paeth_predictor_16x8_c(uint8_t* dst,
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
+void aom_paeth_predictor_16x4_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_16x4 aom_paeth_predictor_16x4_neon
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_16x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x64 aom_paeth_predictor_16x64_neon
+
+void aom_paeth_predictor_16x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_16x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_neon
 
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c
+void aom_paeth_predictor_32x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_neon
 
 void aom_paeth_predictor_32x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_c
+void aom_paeth_predictor_32x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_neon
 
 void aom_paeth_predictor_32x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_c
+void aom_paeth_predictor_32x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_neon
+
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_32x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_32x8 aom_paeth_predictor_32x8_neon
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_4x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_4x16 aom_paeth_predictor_4x16_neon
 
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_c
+void aom_paeth_predictor_4x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_neon
 
 void aom_paeth_predictor_4x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
+void aom_paeth_predictor_4x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_neon
+
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_64x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x16 aom_paeth_predictor_64x16_neon
 
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_c
+void aom_paeth_predictor_64x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_neon
 
 void aom_paeth_predictor_64x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_c
+void aom_paeth_predictor_64x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_neon
 
 void aom_paeth_predictor_8x16_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
+void aom_paeth_predictor_8x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_neon
+
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_8x32_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_8x32 aom_paeth_predictor_8x32_neon
 
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_c
+void aom_paeth_predictor_8x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_neon
 
 void aom_paeth_predictor_8x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_c
+void aom_paeth_predictor_8x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_neon
 
 void aom_pixel_scale_c(const int16_t* src_diff,
                        ptrdiff_t src_stride,
@@ -3219,85 +3523,191 @@ void aom_smooth_h_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_c
+void aom_smooth_h_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_neon
 
 void aom_smooth_h_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c
+void aom_smooth_h_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_neon
 
-void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c
+void aom_smooth_h_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_16x4 aom_smooth_h_predictor_16x4_neon
 
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x64 aom_smooth_h_predictor_16x64_neon
+
+void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_neon
 
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_c
+void aom_smooth_h_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_neon
 
 void aom_smooth_h_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_c
+void aom_smooth_h_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_neon
 
 void aom_smooth_h_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_c
+void aom_smooth_h_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_neon
+
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_32x8 aom_smooth_h_predictor_32x8_neon
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_4x16 aom_smooth_h_predictor_4x16_neon
 
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_c
+void aom_smooth_h_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_neon
 
 void aom_smooth_h_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c
+void aom_smooth_h_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_neon
+
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x16 aom_smooth_h_predictor_64x16_neon
 
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_c
+void aom_smooth_h_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_neon
 
 void aom_smooth_h_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_c
+void aom_smooth_h_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_neon
 
 void aom_smooth_h_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c
+void aom_smooth_h_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_neon
+
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_8x32 aom_smooth_h_predictor_8x32_neon
 
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_c
+void aom_smooth_h_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_neon
 
 void aom_smooth_h_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_c
+void aom_smooth_h_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_neon
 
 void aom_smooth_predictor_16x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
@@ -3319,6 +3729,26 @@ void aom_smooth_predictor_16x32_neon(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_neon
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_16x4_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_16x4 aom_smooth_predictor_16x4_neon
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_16x64_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_predictor_16x64 aom_smooth_predictor_16x64_neon
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -3329,12 +3759,6 @@ void aom_smooth_predictor_16x8_neon(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_neon
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3365,6 +3789,26 @@ void aom_smooth_predictor_32x64_neon(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_smooth_predictor_32x64 aom_smooth_predictor_32x64_neon
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_32x8_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_32x8 aom_smooth_predictor_32x8_neon
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_4x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_4x16 aom_smooth_predictor_4x16_neon
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3385,6 +3829,16 @@ void aom_smooth_predictor_4x8_neon(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_neon
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_64x16_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_predictor_64x16 aom_smooth_predictor_64x16_neon
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3415,6 +3869,16 @@ void aom_smooth_predictor_8x16_neon(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_neon
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_8x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_8x32 aom_smooth_predictor_8x32_neon
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3439,85 +3903,191 @@ void aom_smooth_v_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_c
+void aom_smooth_v_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_neon
 
 void aom_smooth_v_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c
+void aom_smooth_v_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_neon
 
-void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c
+void aom_smooth_v_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_16x4 aom_smooth_v_predictor_16x4_neon
 
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x64 aom_smooth_v_predictor_16x64_neon
+
+void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_neon
 
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_c
+void aom_smooth_v_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_neon
 
 void aom_smooth_v_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_c
+void aom_smooth_v_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_neon
 
 void aom_smooth_v_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_c
+void aom_smooth_v_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_neon
+
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_32x8 aom_smooth_v_predictor_32x8_neon
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_4x16 aom_smooth_v_predictor_4x16_neon
 
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_c
+void aom_smooth_v_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_neon
 
 void aom_smooth_v_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c
+void aom_smooth_v_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_neon
+
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x16 aom_smooth_v_predictor_64x16_neon
 
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_c
+void aom_smooth_v_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_neon
 
 void aom_smooth_v_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_c
+void aom_smooth_v_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_neon
 
 void aom_smooth_v_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c
+void aom_smooth_v_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_neon
+
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_8x32 aom_smooth_v_predictor_8x32_neon
 
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_c
+void aom_smooth_v_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_neon
 
 void aom_smooth_v_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_c
+void aom_smooth_v_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_neon
 
 int64_t aom_sse_c(const uint8_t* a,
                   int a_stride,
@@ -4014,18 +4584,24 @@ void aom_v_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_c
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_c
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_c
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_c
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4048,6 +4624,18 @@ void aom_v_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_32x64 aom_v_predictor_32x64_c
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_c
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_c
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -4064,6 +4652,12 @@ void aom_v_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_c
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_64x16 aom_v_predictor_64x16_c
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4082,6 +4676,12 @@ void aom_v_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_c
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_c
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
diff --git a/chromium/third_party/libaom/source/config/ios/arm-neon/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/ios/arm-neon/config/av1_rtcd.h
index 23e831cc591..7c2e8affa18 100644
--- a/chromium/third_party/libaom/source/config/ios/arm-neon/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/ios/arm-neon/config/av1_rtcd.h
@@ -183,6 +183,30 @@ void aom_upsampled_pred_c(MACROBLOCKD* xd,
                           int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_c
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+void av1_apply_selfguided_restoration_neon(const uint8_t* dat,
+                                           int width,
+                                           int height,
+                                           int stride,
+                                           int eps,
+                                           const int* xqd,
+                                           uint8_t* dst,
+                                           int dst_stride,
+                                           int32_t* tmpbuf,
+                                           int bit_depth,
+                                           int highbd);
+#define av1_apply_selfguided_restoration av1_apply_selfguided_restoration_neon
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -234,6 +258,14 @@ void av1_build_compound_diffwtd_mask_d16_neon(uint8_t* mask,
 #define av1_build_compound_diffwtd_mask_d16 \
   av1_build_compound_diffwtd_mask_d16_neon
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+#define av1_calc_frame_error av1_calc_frame_error_c
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -1314,6 +1346,28 @@ void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 void av1_round_shift_array_neon(int32_t* arr, int size, int bit);
 #define av1_round_shift_array av1_round_shift_array_neon
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+int av1_selfguided_restoration_neon(const uint8_t* dgd8,
+                                    int width,
+                                    int height,
+                                    int dgd_stride,
+                                    int32_t* flt0,
+                                    int32_t* flt1,
+                                    int flt_stride,
+                                    int sgr_params_idx,
+                                    int bit_depth,
+                                    int highbd);
+#define av1_selfguided_restoration av1_selfguided_restoration_neon
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -1330,6 +1384,44 @@ void av1_upsample_intra_edge_c(uint8_t* p, int sz);
 void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 #define av1_upsample_intra_edge_high av1_upsample_intra_edge_high_c
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+void av1_warp_affine_neon(const int32_t* mat,
+                          const uint8_t* ref,
+                          int width,
+                          int height,
+                          int stride,
+                          uint8_t* pred,
+                          int p_col,
+                          int p_row,
+                          int p_width,
+                          int p_height,
+                          int p_stride,
+                          int subsampling_x,
+                          int subsampling_y,
+                          ConvolveParams* conv_params,
+                          int16_t alpha,
+                          int16_t beta,
+                          int16_t gamma,
+                          int16_t delta);
+#define av1_warp_affine av1_warp_affine_neon
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
@@ -1610,7 +1702,15 @@ void cdef_find_dir_dual_c(const uint16_t* img1,
                           int coeff_shift,
                           int* out1,
                           int* out2);
-#define cdef_find_dir_dual cdef_find_dir_dual_c
+void cdef_find_dir_dual_neon(const uint16_t* img1,
+                             const uint16_t* img2,
+                             int stride,
+                             int32_t* var1,
+                             int32_t* var2,
+                             int coeff_shift,
+                             int* out1,
+                             int* out2);
+#define cdef_find_dir_dual cdef_find_dir_dual_neon
 
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_neon(TX_SIZE tx_size);
diff --git a/chromium/third_party/libaom/source/config/ios/arm64/config/aom_config.c b/chromium/third_party/libaom/source/config/ios/arm64/config/aom_config.c
index cd0a632b72a..7f23b547ab1 100644
--- a/chromium/third_party/libaom/source/config/ios/arm64/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/ios/arm64/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/arm64-ios.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/arm64-ios.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/ios/arm64/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/ios/arm64/config/aom_dsp_rtcd.h
index 155c11857c0..49fb79d7f18 100644
--- a/chromium/third_party/libaom/source/config/ios/arm64/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/ios/arm64/config/aom_dsp_rtcd.h
@@ -34,7 +34,12 @@ void aom_avg_8x8_quad_c(const uint8_t* s,
                         int x16_idx,
                         int y16_idx,
                         int* avg);
-#define aom_avg_8x8_quad aom_avg_8x8_quad_c
+void aom_avg_8x8_quad_neon(const uint8_t* s,
+                           int p,
+                           int x16_idx,
+                           int y16_idx,
+                           int* avg);
+#define aom_avg_8x8_quad aom_avg_8x8_quad_neon
 
 void aom_blend_a64_hmask_c(uint8_t* dst,
                            uint32_t dst_stride,
@@ -176,18 +181,24 @@ void aom_dc_128_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_c
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_c
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -210,6 +221,18 @@ void aom_dc_128_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_32x64 aom_dc_128_predictor_32x64_c
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_c
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_c
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -226,6 +249,12 @@ void aom_dc_128_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_64x16 aom_dc_128_predictor_64x16_c
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -244,6 +273,12 @@ void aom_dc_128_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_c
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -276,18 +311,24 @@ void aom_dc_left_predictor_16x32_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_c
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_c
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -310,6 +351,18 @@ void aom_dc_left_predictor_32x64_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_32x64 aom_dc_left_predictor_32x64_c
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_c
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_c
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -326,6 +379,12 @@ void aom_dc_left_predictor_4x8_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_64x16 aom_dc_left_predictor_64x16_c
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -344,6 +403,12 @@ void aom_dc_left_predictor_8x16_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_c
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -376,18 +441,24 @@ void aom_dc_predictor_16x32_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_c
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_c
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
                              const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -410,6 +481,18 @@ void aom_dc_predictor_32x64_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_32x64 aom_dc_predictor_32x64_c
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_c
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_c
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -426,6 +509,12 @@ void aom_dc_predictor_4x8_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_64x16 aom_dc_predictor_64x16_c
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -444,6 +533,12 @@ void aom_dc_predictor_8x16_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_c
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -476,18 +571,24 @@ void aom_dc_top_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_c
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_c
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -510,6 +611,18 @@ void aom_dc_top_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_32x64 aom_dc_top_predictor_32x64_c
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_c
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_c
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -526,6 +639,12 @@ void aom_dc_top_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_64x16 aom_dc_top_predictor_64x16_c
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -544,6 +663,12 @@ void aom_dc_top_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_c
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -1014,18 +1139,24 @@ void aom_h_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_c
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_c
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_c
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_c
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1048,6 +1179,18 @@ void aom_h_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_c
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_c
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_c
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1064,6 +1207,12 @@ void aom_h_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_c
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_c
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1082,6 +1231,12 @@ void aom_h_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_c
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_c
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1127,7 +1282,10 @@ void aom_hadamard_8x8_neon(const int16_t* src_diff,
 void aom_hadamard_8x8_dual_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
                              int16_t* coeff);
-#define aom_hadamard_8x8_dual aom_hadamard_8x8_dual_c
+void aom_hadamard_8x8_dual_neon(const int16_t* src_diff,
+                                ptrdiff_t src_stride,
+                                int16_t* coeff);
+#define aom_hadamard_8x8_dual aom_hadamard_8x8_dual_neon
 
 void aom_hadamard_lp_16x16_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
@@ -1237,7 +1395,12 @@ void aom_lpf_horizontal_14_quad_c(uint8_t* s,
                                   const uint8_t* blimit0,
                                   const uint8_t* limit0,
                                   const uint8_t* thresh0);
-#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_c
+void aom_lpf_horizontal_14_quad_neon(uint8_t* s,
+                                     int pitch,
+                                     const uint8_t* blimit0,
+                                     const uint8_t* limit0,
+                                     const uint8_t* thresh0);
+#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_neon
 
 void aom_lpf_horizontal_4_c(uint8_t* s,
                             int pitch,
@@ -1274,7 +1437,12 @@ void aom_lpf_horizontal_4_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_c
+void aom_lpf_horizontal_4_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_neon
 
 void aom_lpf_horizontal_6_c(uint8_t* s,
                             int pitch,
@@ -1311,7 +1479,12 @@ void aom_lpf_horizontal_6_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_c
+void aom_lpf_horizontal_6_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_neon
 
 void aom_lpf_horizontal_8_c(uint8_t* s,
                             int pitch,
@@ -1348,7 +1521,12 @@ void aom_lpf_horizontal_8_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_c
+void aom_lpf_horizontal_8_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_neon
 
 void aom_lpf_vertical_14_c(uint8_t* s,
                            int pitch,
@@ -1385,7 +1563,12 @@ void aom_lpf_vertical_14_quad_c(uint8_t* s,
                                 const uint8_t* blimit0,
                                 const uint8_t* limit0,
                                 const uint8_t* thresh0);
-#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_c
+void aom_lpf_vertical_14_quad_neon(uint8_t* s,
+                                   int pitch,
+                                   const uint8_t* blimit0,
+                                   const uint8_t* limit0,
+                                   const uint8_t* thresh0);
+#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_neon
 
 void aom_lpf_vertical_4_c(uint8_t* s,
                           int pitch,
@@ -1422,7 +1605,12 @@ void aom_lpf_vertical_4_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_c
+void aom_lpf_vertical_4_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_neon
 
 void aom_lpf_vertical_6_c(uint8_t* s,
                           int pitch,
@@ -1459,7 +1647,12 @@ void aom_lpf_vertical_6_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_c
+void aom_lpf_vertical_6_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_neon
 
 void aom_lpf_vertical_8_c(uint8_t* s,
                           int pitch,
@@ -1496,7 +1689,12 @@ void aom_lpf_vertical_8_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_c
+void aom_lpf_vertical_8_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_neon
 
 unsigned int aom_masked_sad128x128_c(const uint8_t* src,
                                      int src_stride,
@@ -2098,85 +2296,191 @@ void aom_paeth_predictor_16x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_c
+void aom_paeth_predictor_16x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_neon
 
 void aom_paeth_predictor_16x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
+void aom_paeth_predictor_16x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_neon
 
-void aom_paeth_predictor_16x8_c(uint8_t* dst,
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
+void aom_paeth_predictor_16x4_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_16x4 aom_paeth_predictor_16x4_neon
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_16x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x64 aom_paeth_predictor_16x64_neon
+
+void aom_paeth_predictor_16x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_16x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_neon
 
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c
+void aom_paeth_predictor_32x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_neon
 
 void aom_paeth_predictor_32x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_c
+void aom_paeth_predictor_32x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_neon
 
 void aom_paeth_predictor_32x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_c
+void aom_paeth_predictor_32x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_neon
+
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_32x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_32x8 aom_paeth_predictor_32x8_neon
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_4x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_4x16 aom_paeth_predictor_4x16_neon
 
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_c
+void aom_paeth_predictor_4x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_neon
 
 void aom_paeth_predictor_4x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
+void aom_paeth_predictor_4x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_neon
+
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_64x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x16 aom_paeth_predictor_64x16_neon
 
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_c
+void aom_paeth_predictor_64x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_neon
 
 void aom_paeth_predictor_64x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_c
+void aom_paeth_predictor_64x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_neon
 
 void aom_paeth_predictor_8x16_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
+void aom_paeth_predictor_8x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_neon
+
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_8x32_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_8x32 aom_paeth_predictor_8x32_neon
 
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_c
+void aom_paeth_predictor_8x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_neon
 
 void aom_paeth_predictor_8x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_c
+void aom_paeth_predictor_8x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_neon
 
 void aom_pixel_scale_c(const int16_t* src_diff,
                        ptrdiff_t src_stride,
@@ -3219,85 +3523,191 @@ void aom_smooth_h_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_c
+void aom_smooth_h_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_neon
 
 void aom_smooth_h_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c
+void aom_smooth_h_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_neon
 
-void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c
+void aom_smooth_h_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_16x4 aom_smooth_h_predictor_16x4_neon
 
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x64 aom_smooth_h_predictor_16x64_neon
+
+void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_neon
 
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_c
+void aom_smooth_h_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_neon
 
 void aom_smooth_h_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_c
+void aom_smooth_h_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_neon
 
 void aom_smooth_h_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_c
+void aom_smooth_h_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_neon
+
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_32x8 aom_smooth_h_predictor_32x8_neon
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_4x16 aom_smooth_h_predictor_4x16_neon
 
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_c
+void aom_smooth_h_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_neon
 
 void aom_smooth_h_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c
+void aom_smooth_h_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_neon
+
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x16 aom_smooth_h_predictor_64x16_neon
 
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_c
+void aom_smooth_h_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_neon
 
 void aom_smooth_h_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_c
+void aom_smooth_h_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_neon
 
 void aom_smooth_h_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c
+void aom_smooth_h_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_neon
+
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_8x32 aom_smooth_h_predictor_8x32_neon
 
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_c
+void aom_smooth_h_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_neon
 
 void aom_smooth_h_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_c
+void aom_smooth_h_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_neon
 
 void aom_smooth_predictor_16x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
@@ -3319,6 +3729,26 @@ void aom_smooth_predictor_16x32_neon(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_neon
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_16x4_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_16x4 aom_smooth_predictor_16x4_neon
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_16x64_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_predictor_16x64 aom_smooth_predictor_16x64_neon
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -3329,12 +3759,6 @@ void aom_smooth_predictor_16x8_neon(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_neon
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3365,6 +3789,26 @@ void aom_smooth_predictor_32x64_neon(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_smooth_predictor_32x64 aom_smooth_predictor_32x64_neon
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_32x8_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_32x8 aom_smooth_predictor_32x8_neon
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_4x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_4x16 aom_smooth_predictor_4x16_neon
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3385,6 +3829,16 @@ void aom_smooth_predictor_4x8_neon(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_neon
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_64x16_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_predictor_64x16 aom_smooth_predictor_64x16_neon
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3415,6 +3869,16 @@ void aom_smooth_predictor_8x16_neon(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_neon
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_8x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_8x32 aom_smooth_predictor_8x32_neon
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3439,85 +3903,191 @@ void aom_smooth_v_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_c
+void aom_smooth_v_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_neon
 
 void aom_smooth_v_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c
+void aom_smooth_v_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_neon
 
-void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c
+void aom_smooth_v_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_16x4 aom_smooth_v_predictor_16x4_neon
 
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x64 aom_smooth_v_predictor_16x64_neon
+
+void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_neon
 
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_c
+void aom_smooth_v_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_neon
 
 void aom_smooth_v_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_c
+void aom_smooth_v_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_neon
 
 void aom_smooth_v_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_c
+void aom_smooth_v_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_neon
+
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_32x8 aom_smooth_v_predictor_32x8_neon
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_4x16 aom_smooth_v_predictor_4x16_neon
 
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_c
+void aom_smooth_v_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_neon
 
 void aom_smooth_v_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c
+void aom_smooth_v_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_neon
+
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x16 aom_smooth_v_predictor_64x16_neon
 
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_c
+void aom_smooth_v_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_neon
 
 void aom_smooth_v_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_c
+void aom_smooth_v_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_neon
 
 void aom_smooth_v_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c
+void aom_smooth_v_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_neon
+
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_8x32 aom_smooth_v_predictor_8x32_neon
 
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_c
+void aom_smooth_v_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_neon
 
 void aom_smooth_v_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_c
+void aom_smooth_v_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_neon
 
 int64_t aom_sse_c(const uint8_t* a,
                   int a_stride,
@@ -4014,18 +4584,24 @@ void aom_v_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_c
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_c
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_c
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_c
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4048,6 +4624,18 @@ void aom_v_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_32x64 aom_v_predictor_32x64_c
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_c
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_c
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -4064,6 +4652,12 @@ void aom_v_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_c
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_64x16 aom_v_predictor_64x16_c
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4082,6 +4676,12 @@ void aom_v_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_c
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_c
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
diff --git a/chromium/third_party/libaom/source/config/ios/arm64/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/ios/arm64/config/av1_rtcd.h
index 23e831cc591..7c2e8affa18 100644
--- a/chromium/third_party/libaom/source/config/ios/arm64/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/ios/arm64/config/av1_rtcd.h
@@ -183,6 +183,30 @@ void aom_upsampled_pred_c(MACROBLOCKD* xd,
                           int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_c
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+void av1_apply_selfguided_restoration_neon(const uint8_t* dat,
+                                           int width,
+                                           int height,
+                                           int stride,
+                                           int eps,
+                                           const int* xqd,
+                                           uint8_t* dst,
+                                           int dst_stride,
+                                           int32_t* tmpbuf,
+                                           int bit_depth,
+                                           int highbd);
+#define av1_apply_selfguided_restoration av1_apply_selfguided_restoration_neon
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -234,6 +258,14 @@ void av1_build_compound_diffwtd_mask_d16_neon(uint8_t* mask,
 #define av1_build_compound_diffwtd_mask_d16 \
   av1_build_compound_diffwtd_mask_d16_neon
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+#define av1_calc_frame_error av1_calc_frame_error_c
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -1314,6 +1346,28 @@ void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 void av1_round_shift_array_neon(int32_t* arr, int size, int bit);
 #define av1_round_shift_array av1_round_shift_array_neon
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+int av1_selfguided_restoration_neon(const uint8_t* dgd8,
+                                    int width,
+                                    int height,
+                                    int dgd_stride,
+                                    int32_t* flt0,
+                                    int32_t* flt1,
+                                    int flt_stride,
+                                    int sgr_params_idx,
+                                    int bit_depth,
+                                    int highbd);
+#define av1_selfguided_restoration av1_selfguided_restoration_neon
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -1330,6 +1384,44 @@ void av1_upsample_intra_edge_c(uint8_t* p, int sz);
 void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 #define av1_upsample_intra_edge_high av1_upsample_intra_edge_high_c
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+void av1_warp_affine_neon(const int32_t* mat,
+                          const uint8_t* ref,
+                          int width,
+                          int height,
+                          int stride,
+                          uint8_t* pred,
+                          int p_col,
+                          int p_row,
+                          int p_width,
+                          int p_height,
+                          int p_stride,
+                          int subsampling_x,
+                          int subsampling_y,
+                          ConvolveParams* conv_params,
+                          int16_t alpha,
+                          int16_t beta,
+                          int16_t gamma,
+                          int16_t delta);
+#define av1_warp_affine av1_warp_affine_neon
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
@@ -1610,7 +1702,15 @@ void cdef_find_dir_dual_c(const uint16_t* img1,
                           int coeff_shift,
                           int* out1,
                           int* out2);
-#define cdef_find_dir_dual cdef_find_dir_dual_c
+void cdef_find_dir_dual_neon(const uint16_t* img1,
+                             const uint16_t* img2,
+                             int stride,
+                             int32_t* var1,
+                             int32_t* var2,
+                             int coeff_shift,
+                             int* out1,
+                             int* out2);
+#define cdef_find_dir_dual cdef_find_dir_dual_neon
 
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_neon(TX_SIZE tx_size);
diff --git a/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/aom_config.c b/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/aom_config.c
index c20c00bb8b1..237bcef9085 100644
--- a/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/armv7-linux-gcc.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/armv7-linux-gcc.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/aom_dsp_rtcd.h
index 2a73acd6108..59e91c56bce 100644
--- a/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/aom_dsp_rtcd.h
@@ -34,7 +34,16 @@ void aom_avg_8x8_quad_c(const uint8_t* s,
                         int x16_idx,
                         int y16_idx,
                         int* avg);
-#define aom_avg_8x8_quad aom_avg_8x8_quad_c
+void aom_avg_8x8_quad_neon(const uint8_t* s,
+                           int p,
+                           int x16_idx,
+                           int y16_idx,
+                           int* avg);
+RTCD_EXTERN void (*aom_avg_8x8_quad)(const uint8_t* s,
+                                     int p,
+                                     int x16_idx,
+                                     int y16_idx,
+                                     int* avg);
 
 void aom_blend_a64_hmask_c(uint8_t* dst,
                            uint32_t dst_stride,
@@ -200,18 +209,24 @@ void aom_dc_128_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_c
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_c
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -237,6 +252,18 @@ void aom_dc_128_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_32x64 aom_dc_128_predictor_32x64_c
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_c
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_c
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -256,6 +283,12 @@ void aom_dc_128_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_64x16 aom_dc_128_predictor_64x16_c
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -274,6 +307,12 @@ void aom_dc_128_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_c
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -312,18 +351,24 @@ void aom_dc_left_predictor_16x32_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_c
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_c
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -349,6 +394,18 @@ void aom_dc_left_predictor_32x64_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_32x64 aom_dc_left_predictor_32x64_c
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_c
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_c
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -368,6 +425,12 @@ void aom_dc_left_predictor_4x8_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_64x16 aom_dc_left_predictor_64x16_c
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -386,6 +449,12 @@ void aom_dc_left_predictor_8x16_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_c
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -424,18 +493,24 @@ void aom_dc_predictor_16x32_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_c
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_c
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
                              const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -461,6 +536,18 @@ void aom_dc_predictor_32x64_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_32x64 aom_dc_predictor_32x64_c
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_c
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_c
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -480,6 +567,12 @@ void aom_dc_predictor_4x8_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_64x16 aom_dc_predictor_64x16_c
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -498,6 +591,12 @@ void aom_dc_predictor_8x16_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_c
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -536,18 +635,24 @@ void aom_dc_top_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_c
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_c
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -573,6 +678,18 @@ void aom_dc_top_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_32x64 aom_dc_top_predictor_32x64_c
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_c
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_c
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -592,6 +709,12 @@ void aom_dc_top_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_64x16 aom_dc_top_predictor_64x16_c
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -610,6 +733,12 @@ void aom_dc_top_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_c
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -1110,18 +1239,24 @@ void aom_h_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_c
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_c
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_c
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_c
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1147,6 +1282,18 @@ void aom_h_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_c
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_c
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_c
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1166,6 +1313,12 @@ void aom_h_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_c
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_c
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1184,6 +1337,12 @@ void aom_h_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_c
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_c
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1236,7 +1395,12 @@ RTCD_EXTERN void (*aom_hadamard_8x8)(const int16_t* src_diff,
 void aom_hadamard_8x8_dual_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
                              int16_t* coeff);
-#define aom_hadamard_8x8_dual aom_hadamard_8x8_dual_c
+void aom_hadamard_8x8_dual_neon(const int16_t* src_diff,
+                                ptrdiff_t src_stride,
+                                int16_t* coeff);
+RTCD_EXTERN void (*aom_hadamard_8x8_dual)(const int16_t* src_diff,
+                                          ptrdiff_t src_stride,
+                                          int16_t* coeff);
 
 void aom_hadamard_lp_16x16_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
@@ -1376,7 +1540,16 @@ void aom_lpf_horizontal_14_quad_c(uint8_t* s,
                                   const uint8_t* blimit0,
                                   const uint8_t* limit0,
                                   const uint8_t* thresh0);
-#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_c
+void aom_lpf_horizontal_14_quad_neon(uint8_t* s,
+                                     int pitch,
+                                     const uint8_t* blimit0,
+                                     const uint8_t* limit0,
+                                     const uint8_t* thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t* s,
+                                               int pitch,
+                                               const uint8_t* blimit0,
+                                               const uint8_t* limit0,
+                                               const uint8_t* thresh0);
 
 void aom_lpf_horizontal_4_c(uint8_t* s,
                             int pitch,
@@ -1424,7 +1597,16 @@ void aom_lpf_horizontal_4_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_c
+void aom_lpf_horizontal_4_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_4_quad)(uint8_t* s,
+                                              int pitch,
+                                              const uint8_t* blimit0,
+                                              const uint8_t* limit0,
+                                              const uint8_t* thresh0);
 
 void aom_lpf_horizontal_6_c(uint8_t* s,
                             int pitch,
@@ -1472,7 +1654,16 @@ void aom_lpf_horizontal_6_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_c
+void aom_lpf_horizontal_6_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t* s,
+                                              int pitch,
+                                              const uint8_t* blimit0,
+                                              const uint8_t* limit0,
+                                              const uint8_t* thresh0);
 
 void aom_lpf_horizontal_8_c(uint8_t* s,
                             int pitch,
@@ -1520,7 +1711,16 @@ void aom_lpf_horizontal_8_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_c
+void aom_lpf_horizontal_8_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t* s,
+                                              int pitch,
+                                              const uint8_t* blimit0,
+                                              const uint8_t* limit0,
+                                              const uint8_t* thresh0);
 
 void aom_lpf_vertical_14_c(uint8_t* s,
                            int pitch,
@@ -1568,7 +1768,16 @@ void aom_lpf_vertical_14_quad_c(uint8_t* s,
                                 const uint8_t* blimit0,
                                 const uint8_t* limit0,
                                 const uint8_t* thresh0);
-#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_c
+void aom_lpf_vertical_14_quad_neon(uint8_t* s,
+                                   int pitch,
+                                   const uint8_t* blimit0,
+                                   const uint8_t* limit0,
+                                   const uint8_t* thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_14_quad)(uint8_t* s,
+                                             int pitch,
+                                             const uint8_t* blimit0,
+                                             const uint8_t* limit0,
+                                             const uint8_t* thresh0);
 
 void aom_lpf_vertical_4_c(uint8_t* s,
                           int pitch,
@@ -1616,7 +1825,16 @@ void aom_lpf_vertical_4_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_c
+void aom_lpf_vertical_4_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_4_quad)(uint8_t* s,
+                                            int pitch,
+                                            const uint8_t* blimit0,
+                                            const uint8_t* limit0,
+                                            const uint8_t* thresh0);
 
 void aom_lpf_vertical_6_c(uint8_t* s,
                           int pitch,
@@ -1664,7 +1882,16 @@ void aom_lpf_vertical_6_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_c
+void aom_lpf_vertical_6_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_6_quad)(uint8_t* s,
+                                            int pitch,
+                                            const uint8_t* blimit0,
+                                            const uint8_t* limit0,
+                                            const uint8_t* thresh0);
 
 void aom_lpf_vertical_8_c(uint8_t* s,
                           int pitch,
@@ -1712,7 +1939,16 @@ void aom_lpf_vertical_8_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_c
+void aom_lpf_vertical_8_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_8_quad)(uint8_t* s,
+                                            int pitch,
+                                            const uint8_t* blimit0,
+                                            const uint8_t* limit0,
+                                            const uint8_t* thresh0);
 
 unsigned int aom_masked_sad128x128_c(const uint8_t* src,
                                      int src_stride,
@@ -2318,85 +2554,248 @@ void aom_paeth_predictor_16x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_c
+void aom_paeth_predictor_16x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
 
 void aom_paeth_predictor_16x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
+void aom_paeth_predictor_16x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
 
-void aom_paeth_predictor_16x8_c(uint8_t* dst,
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
+void aom_paeth_predictor_16x4_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_16x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_paeth_predictor_16x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_16x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
 
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c
+void aom_paeth_predictor_32x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
 
 void aom_paeth_predictor_32x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_c
+void aom_paeth_predictor_32x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x32)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
 
 void aom_paeth_predictor_32x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_c
+void aom_paeth_predictor_32x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_32x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_4x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
 
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_c
+void aom_paeth_predictor_4x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x4)(uint8_t* dst,
+                                            ptrdiff_t y_stride,
+                                            const uint8_t* above,
+                                            const uint8_t* left);
 
 void aom_paeth_predictor_4x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
+void aom_paeth_predictor_4x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t* dst,
+                                            ptrdiff_t y_stride,
+                                            const uint8_t* above,
+                                            const uint8_t* left);
+
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_64x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
 
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_c
+void aom_paeth_predictor_64x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x32)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
 
 void aom_paeth_predictor_64x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_c
+void aom_paeth_predictor_64x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x64)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
 
 void aom_paeth_predictor_8x16_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
+void aom_paeth_predictor_8x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_8x32_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
 
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_c
+void aom_paeth_predictor_8x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x4)(uint8_t* dst,
+                                            ptrdiff_t y_stride,
+                                            const uint8_t* above,
+                                            const uint8_t* left);
 
 void aom_paeth_predictor_8x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_c
+void aom_paeth_predictor_8x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t* dst,
+                                            ptrdiff_t y_stride,
+                                            const uint8_t* above,
+                                            const uint8_t* left);
 
 void aom_pixel_scale_c(const int16_t* src_diff,
                        ptrdiff_t src_stride,
@@ -3616,85 +4015,248 @@ void aom_smooth_h_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_c
+void aom_smooth_h_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_h_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c
+void aom_smooth_h_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c
-
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
+void aom_smooth_h_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
 
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_c
+void aom_smooth_h_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_h_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_c
+void aom_smooth_h_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x32)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_h_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_c
+void aom_smooth_h_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
 
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_c
+void aom_smooth_h_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x4)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
 
 void aom_smooth_h_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c
+void aom_smooth_h_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_c
+void aom_smooth_h_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x32)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_h_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_c
+void aom_smooth_h_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_h_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c
+void aom_smooth_h_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
 
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_c
+void aom_smooth_h_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x4)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
 
 void aom_smooth_h_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_c
+void aom_smooth_h_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x8)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
 
 void aom_smooth_predictor_16x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
@@ -3722,6 +4284,32 @@ RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_16x4_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_16x64_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -3735,12 +4323,6 @@ RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3780,6 +4362,32 @@ RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_32x8_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_4x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3806,6 +4414,19 @@ RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_64x16_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3845,6 +4466,19 @@ RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_8x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3875,85 +4509,248 @@ void aom_smooth_v_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_c
+void aom_smooth_v_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_v_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c
+void aom_smooth_v_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c
-
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
+void aom_smooth_v_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
 
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_c
+void aom_smooth_v_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_v_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_c
+void aom_smooth_v_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x32)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_v_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_c
+void aom_smooth_v_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
 
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_c
+void aom_smooth_v_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x4)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
 
 void aom_smooth_v_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c
+void aom_smooth_v_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_c
+void aom_smooth_v_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x32)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_v_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_c
+void aom_smooth_v_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
 
 void aom_smooth_v_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c
+void aom_smooth_v_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
 
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_c
+void aom_smooth_v_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x4)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
 
 void aom_smooth_v_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_c
+void aom_smooth_v_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x8)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
 
 int64_t aom_sse_c(const uint8_t* a,
                   int a_stride,
@@ -4564,18 +5361,24 @@ void aom_v_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_c
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_c
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_c
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_c
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4601,6 +5404,18 @@ void aom_v_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_32x64 aom_v_predictor_32x64_c
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_c
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_c
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -4620,6 +5435,12 @@ void aom_v_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_c
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_64x16 aom_v_predictor_64x16_c
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4638,6 +5459,12 @@ void aom_v_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_c
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_c
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -4963,6 +5790,9 @@ static void setup_rtcd_internal(void) {
   aom_avg_8x8 = aom_avg_8x8_c;
   if (flags & HAS_NEON)
     aom_avg_8x8 = aom_avg_8x8_neon;
+  aom_avg_8x8_quad = aom_avg_8x8_quad_c;
+  if (flags & HAS_NEON)
+    aom_avg_8x8_quad = aom_avg_8x8_quad_neon;
   aom_blend_a64_hmask = aom_blend_a64_hmask_c;
   if (flags & HAS_NEON)
     aom_blend_a64_hmask = aom_blend_a64_hmask_neon;
@@ -5059,6 +5889,9 @@ static void setup_rtcd_internal(void) {
   aom_hadamard_8x8 = aom_hadamard_8x8_c;
   if (flags & HAS_NEON)
     aom_hadamard_8x8 = aom_hadamard_8x8_neon;
+  aom_hadamard_8x8_dual = aom_hadamard_8x8_dual_c;
+  if (flags & HAS_NEON)
+    aom_hadamard_8x8_dual = aom_hadamard_8x8_dual_neon;
   aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_c;
   if (flags & HAS_NEON)
     aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_neon;
@@ -5080,51 +5913,132 @@ static void setup_rtcd_internal(void) {
   aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_c;
   if (flags & HAS_NEON)
     aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_neon;
+  aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_c;
+  if (flags & HAS_NEON)
+    aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_neon;
   aom_lpf_horizontal_4 = aom_lpf_horizontal_4_c;
   if (flags & HAS_NEON)
     aom_lpf_horizontal_4 = aom_lpf_horizontal_4_neon;
   aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_c;
   if (flags & HAS_NEON)
     aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_neon;
+  aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_c;
+  if (flags & HAS_NEON)
+    aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_neon;
   aom_lpf_horizontal_6 = aom_lpf_horizontal_6_c;
   if (flags & HAS_NEON)
     aom_lpf_horizontal_6 = aom_lpf_horizontal_6_neon;
   aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_c;
   if (flags & HAS_NEON)
     aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_neon;
+  aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_c;
+  if (flags & HAS_NEON)
+    aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_neon;
   aom_lpf_horizontal_8 = aom_lpf_horizontal_8_c;
   if (flags & HAS_NEON)
     aom_lpf_horizontal_8 = aom_lpf_horizontal_8_neon;
   aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_c;
   if (flags & HAS_NEON)
     aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_neon;
+  aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_c;
+  if (flags & HAS_NEON)
+    aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_neon;
   aom_lpf_vertical_14 = aom_lpf_vertical_14_c;
   if (flags & HAS_NEON)
     aom_lpf_vertical_14 = aom_lpf_vertical_14_neon;
   aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_c;
   if (flags & HAS_NEON)
     aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_neon;
+  aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_c;
+  if (flags & HAS_NEON)
+    aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_neon;
   aom_lpf_vertical_4 = aom_lpf_vertical_4_c;
   if (flags & HAS_NEON)
     aom_lpf_vertical_4 = aom_lpf_vertical_4_neon;
   aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_c;
   if (flags & HAS_NEON)
     aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_neon;
+  aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_c;
+  if (flags & HAS_NEON)
+    aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_neon;
   aom_lpf_vertical_6 = aom_lpf_vertical_6_c;
   if (flags & HAS_NEON)
     aom_lpf_vertical_6 = aom_lpf_vertical_6_neon;
   aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_c;
   if (flags & HAS_NEON)
     aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_neon;
+  aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_c;
+  if (flags & HAS_NEON)
+    aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_neon;
   aom_lpf_vertical_8 = aom_lpf_vertical_8_c;
   if (flags & HAS_NEON)
     aom_lpf_vertical_8 = aom_lpf_vertical_8_neon;
   aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_c;
   if (flags & HAS_NEON)
     aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_neon;
+  aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_c;
+  if (flags & HAS_NEON)
+    aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_neon;
   aom_mse16x16 = aom_mse16x16_c;
   if (flags & HAS_NEON)
     aom_mse16x16 = aom_mse16x16_neon;
+  aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_neon;
+  aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_neon;
+  aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_neon;
+  aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_neon;
+  aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_neon;
+  aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_neon;
+  aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_neon;
+  aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_neon;
+  aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_neon;
+  aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_neon;
+  aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_neon;
+  aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_neon;
+  aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_neon;
+  aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_neon;
+  aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_neon;
+  aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_neon;
+  aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_neon;
+  aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_neon;
+  aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c;
+  if (flags & HAS_NEON)
+    aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_neon;
   aom_quantize_b = aom_quantize_b_c;
   if (flags & HAS_NEON)
     aom_quantize_b = aom_quantize_b_neon;
@@ -5260,12 +6174,75 @@ static void setup_rtcd_internal(void) {
   aom_scaled_2d = aom_scaled_2d_c;
   if (flags & HAS_NEON)
     aom_scaled_2d = aom_scaled_2d_neon;
+  aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_neon;
+  aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_neon;
+  aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_neon;
+  aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_neon;
+  aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_neon;
+  aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_neon;
+  aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_neon;
+  aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_neon;
+  aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_neon;
+  aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_neon;
+  aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_neon;
+  aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_neon;
+  aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_neon;
+  aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_neon;
+  aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_neon;
+  aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_neon;
+  aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_neon;
+  aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_neon;
+  aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_c;
+  if (flags & HAS_NEON)
+    aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_neon;
   aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_c;
   if (flags & HAS_NEON)
     aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_neon;
   aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
   if (flags & HAS_NEON)
     aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_neon;
+  aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c;
+  if (flags & HAS_NEON)
+    aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_neon;
+  aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c;
+  if (flags & HAS_NEON)
+    aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_neon;
   aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
   if (flags & HAS_NEON)
     aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_neon;
@@ -5278,12 +6255,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c;
   if (flags & HAS_NEON)
     aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_neon;
+  aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c;
+  if (flags & HAS_NEON)
+    aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_neon;
+  aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_neon;
   aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
   if (flags & HAS_NEON)
     aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_neon;
   aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
   if (flags & HAS_NEON)
     aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_neon;
+  aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_neon;
   aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c;
   if (flags & HAS_NEON)
     aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_neon;
@@ -5293,12 +6279,72 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
   if (flags & HAS_NEON)
     aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_neon;
+  aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c;
+  if (flags & HAS_NEON)
+    aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_neon;
   aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
   if (flags & HAS_NEON)
     aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_neon;
   aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_c;
   if (flags & HAS_NEON)
     aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_neon;
+  aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_neon;
+  aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_neon;
+  aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_neon;
+  aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_neon;
+  aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_neon;
+  aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_neon;
+  aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_neon;
+  aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_neon;
+  aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_neon;
+  aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_neon;
+  aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_neon;
+  aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_neon;
+  aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_neon;
+  aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_neon;
+  aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_neon;
+  aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_neon;
+  aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_neon;
+  aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_neon;
+  aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_c;
+  if (flags & HAS_NEON)
+    aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_neon;
   aom_sse = aom_sse_c;
   if (flags & HAS_NEON)
     aom_sse = aom_sse_neon;
diff --git a/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/av1_rtcd.h
index 90c222a2a20..185c01d4dd0 100644
--- a/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/arm-neon-cpu-detect/config/av1_rtcd.h
@@ -197,6 +197,40 @@ void aom_upsampled_pred_c(MACROBLOCKD* xd,
                           int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_c
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+void av1_apply_selfguided_restoration_neon(const uint8_t* dat,
+                                           int width,
+                                           int height,
+                                           int stride,
+                                           int eps,
+                                           const int* xqd,
+                                           uint8_t* dst,
+                                           int dst_stride,
+                                           int32_t* tmpbuf,
+                                           int bit_depth,
+                                           int highbd);
+RTCD_EXTERN void (*av1_apply_selfguided_restoration)(const uint8_t* dat,
+                                                     int width,
+                                                     int height,
+                                                     int stride,
+                                                     int eps,
+                                                     const int* xqd,
+                                                     uint8_t* dst,
+                                                     int dst_stride,
+                                                     int32_t* tmpbuf,
+                                                     int bit_depth,
+                                                     int highbd);
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -262,6 +296,14 @@ RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(
     ConvolveParams* conv_params,
     int bd);
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+#define av1_calc_frame_error av1_calc_frame_error_c
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -1683,6 +1725,37 @@ void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 void av1_round_shift_array_neon(int32_t* arr, int size, int bit);
 RTCD_EXTERN void (*av1_round_shift_array)(int32_t* arr, int size, int bit);
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+int av1_selfguided_restoration_neon(const uint8_t* dgd8,
+                                    int width,
+                                    int height,
+                                    int dgd_stride,
+                                    int32_t* flt0,
+                                    int32_t* flt1,
+                                    int flt_stride,
+                                    int sgr_params_idx,
+                                    int bit_depth,
+                                    int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t* dgd8,
+                                              int width,
+                                              int height,
+                                              int dgd_stride,
+                                              int32_t* flt0,
+                                              int32_t* flt1,
+                                              int flt_stride,
+                                              int sgr_params_idx,
+                                              int bit_depth,
+                                              int highbd);
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -1702,6 +1775,61 @@ void av1_upsample_intra_edge_c(uint8_t* p, int sz);
 void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 #define av1_upsample_intra_edge_high av1_upsample_intra_edge_high_c
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+void av1_warp_affine_neon(const int32_t* mat,
+                          const uint8_t* ref,
+                          int width,
+                          int height,
+                          int stride,
+                          uint8_t* pred,
+                          int p_col,
+                          int p_row,
+                          int p_width,
+                          int p_height,
+                          int p_stride,
+                          int subsampling_x,
+                          int subsampling_y,
+                          ConvolveParams* conv_params,
+                          int16_t alpha,
+                          int16_t beta,
+                          int16_t gamma,
+                          int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t* mat,
+                                    const uint8_t* ref,
+                                    int width,
+                                    int height,
+                                    int stride,
+                                    uint8_t* pred,
+                                    int p_col,
+                                    int p_row,
+                                    int p_width,
+                                    int p_height,
+                                    int p_stride,
+                                    int subsampling_x,
+                                    int subsampling_y,
+                                    ConvolveParams* conv_params,
+                                    int16_t alpha,
+                                    int16_t beta,
+                                    int16_t gamma,
+                                    int16_t delta);
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
@@ -2086,7 +2214,22 @@ void cdef_find_dir_dual_c(const uint16_t* img1,
                           int coeff_shift,
                           int* out1,
                           int* out2);
-#define cdef_find_dir_dual cdef_find_dir_dual_c
+void cdef_find_dir_dual_neon(const uint16_t* img1,
+                             const uint16_t* img2,
+                             int stride,
+                             int32_t* var1,
+                             int32_t* var2,
+                             int coeff_shift,
+                             int* out1,
+                             int* out2);
+RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t* img1,
+                                       const uint16_t* img2,
+                                       int stride,
+                                       int32_t* var1,
+                                       int32_t* var2,
+                                       int coeff_shift,
+                                       int* out1,
+                                       int* out2);
 
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_neon(TX_SIZE tx_size);
@@ -2126,6 +2269,9 @@ static void setup_rtcd_internal(void) {
   aom_quantize_b_helper = aom_quantize_b_helper_c;
   if (flags & HAS_NEON)
     aom_quantize_b_helper = aom_quantize_b_helper_neon;
+  av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_c;
+  if (flags & HAS_NEON)
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_neon;
   av1_block_error = av1_block_error_c;
   if (flags & HAS_NEON)
     av1_block_error = av1_block_error_neon;
@@ -2352,9 +2498,15 @@ static void setup_rtcd_internal(void) {
   av1_round_shift_array = av1_round_shift_array_c;
   if (flags & HAS_NEON)
     av1_round_shift_array = av1_round_shift_array_neon;
+  av1_selfguided_restoration = av1_selfguided_restoration_c;
+  if (flags & HAS_NEON)
+    av1_selfguided_restoration = av1_selfguided_restoration_neon;
   av1_txb_init_levels = av1_txb_init_levels_c;
   if (flags & HAS_NEON)
     av1_txb_init_levels = av1_txb_init_levels_neon;
+  av1_warp_affine = av1_warp_affine_c;
+  if (flags & HAS_NEON)
+    av1_warp_affine = av1_warp_affine_neon;
   av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_c;
   if (flags & HAS_NEON)
     av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_neon;
@@ -2391,6 +2543,9 @@ static void setup_rtcd_internal(void) {
   cdef_find_dir = cdef_find_dir_c;
   if (flags & HAS_NEON)
     cdef_find_dir = cdef_find_dir_neon;
+  cdef_find_dir_dual = cdef_find_dir_dual_c;
+  if (flags & HAS_NEON)
+    cdef_find_dir_dual = cdef_find_dir_dual_neon;
   cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_c;
   if (flags & HAS_NEON)
     cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_neon;
diff --git a/chromium/third_party/libaom/source/config/linux/arm-neon/config/aom_config.c b/chromium/third_party/libaom/source/config/linux/arm-neon/config/aom_config.c
index c20c00bb8b1..237bcef9085 100644
--- a/chromium/third_party/libaom/source/config/linux/arm-neon/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/linux/arm-neon/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/armv7-linux-gcc.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/armv7-linux-gcc.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/linux/arm-neon/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/linux/arm-neon/config/aom_dsp_rtcd.h
index 155c11857c0..49fb79d7f18 100644
--- a/chromium/third_party/libaom/source/config/linux/arm-neon/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/arm-neon/config/aom_dsp_rtcd.h
@@ -34,7 +34,12 @@ void aom_avg_8x8_quad_c(const uint8_t* s,
                         int x16_idx,
                         int y16_idx,
                         int* avg);
-#define aom_avg_8x8_quad aom_avg_8x8_quad_c
+void aom_avg_8x8_quad_neon(const uint8_t* s,
+                           int p,
+                           int x16_idx,
+                           int y16_idx,
+                           int* avg);
+#define aom_avg_8x8_quad aom_avg_8x8_quad_neon
 
 void aom_blend_a64_hmask_c(uint8_t* dst,
                            uint32_t dst_stride,
@@ -176,18 +181,24 @@ void aom_dc_128_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_c
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_c
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -210,6 +221,18 @@ void aom_dc_128_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_32x64 aom_dc_128_predictor_32x64_c
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_c
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_c
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -226,6 +249,12 @@ void aom_dc_128_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_64x16 aom_dc_128_predictor_64x16_c
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -244,6 +273,12 @@ void aom_dc_128_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_c
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -276,18 +311,24 @@ void aom_dc_left_predictor_16x32_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_c
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_c
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -310,6 +351,18 @@ void aom_dc_left_predictor_32x64_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_32x64 aom_dc_left_predictor_32x64_c
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_c
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_c
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -326,6 +379,12 @@ void aom_dc_left_predictor_4x8_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_64x16 aom_dc_left_predictor_64x16_c
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -344,6 +403,12 @@ void aom_dc_left_predictor_8x16_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_c
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -376,18 +441,24 @@ void aom_dc_predictor_16x32_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_c
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_c
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
                              const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -410,6 +481,18 @@ void aom_dc_predictor_32x64_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_32x64 aom_dc_predictor_32x64_c
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_c
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_c
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -426,6 +509,12 @@ void aom_dc_predictor_4x8_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_64x16 aom_dc_predictor_64x16_c
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -444,6 +533,12 @@ void aom_dc_predictor_8x16_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_c
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -476,18 +571,24 @@ void aom_dc_top_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_c
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_c
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -510,6 +611,18 @@ void aom_dc_top_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_32x64 aom_dc_top_predictor_32x64_c
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_c
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_c
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -526,6 +639,12 @@ void aom_dc_top_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_64x16 aom_dc_top_predictor_64x16_c
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -544,6 +663,12 @@ void aom_dc_top_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_c
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -1014,18 +1139,24 @@ void aom_h_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_c
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_c
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_c
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_c
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1048,6 +1179,18 @@ void aom_h_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_c
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_c
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_c
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1064,6 +1207,12 @@ void aom_h_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_c
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_c
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1082,6 +1231,12 @@ void aom_h_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_c
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_c
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1127,7 +1282,10 @@ void aom_hadamard_8x8_neon(const int16_t* src_diff,
 void aom_hadamard_8x8_dual_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
                              int16_t* coeff);
-#define aom_hadamard_8x8_dual aom_hadamard_8x8_dual_c
+void aom_hadamard_8x8_dual_neon(const int16_t* src_diff,
+                                ptrdiff_t src_stride,
+                                int16_t* coeff);
+#define aom_hadamard_8x8_dual aom_hadamard_8x8_dual_neon
 
 void aom_hadamard_lp_16x16_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
@@ -1237,7 +1395,12 @@ void aom_lpf_horizontal_14_quad_c(uint8_t* s,
                                   const uint8_t* blimit0,
                                   const uint8_t* limit0,
                                   const uint8_t* thresh0);
-#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_c
+void aom_lpf_horizontal_14_quad_neon(uint8_t* s,
+                                     int pitch,
+                                     const uint8_t* blimit0,
+                                     const uint8_t* limit0,
+                                     const uint8_t* thresh0);
+#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_neon
 
 void aom_lpf_horizontal_4_c(uint8_t* s,
                             int pitch,
@@ -1274,7 +1437,12 @@ void aom_lpf_horizontal_4_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_c
+void aom_lpf_horizontal_4_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_neon
 
 void aom_lpf_horizontal_6_c(uint8_t* s,
                             int pitch,
@@ -1311,7 +1479,12 @@ void aom_lpf_horizontal_6_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_c
+void aom_lpf_horizontal_6_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_neon
 
 void aom_lpf_horizontal_8_c(uint8_t* s,
                             int pitch,
@@ -1348,7 +1521,12 @@ void aom_lpf_horizontal_8_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_c
+void aom_lpf_horizontal_8_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_neon
 
 void aom_lpf_vertical_14_c(uint8_t* s,
                            int pitch,
@@ -1385,7 +1563,12 @@ void aom_lpf_vertical_14_quad_c(uint8_t* s,
                                 const uint8_t* blimit0,
                                 const uint8_t* limit0,
                                 const uint8_t* thresh0);
-#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_c
+void aom_lpf_vertical_14_quad_neon(uint8_t* s,
+                                   int pitch,
+                                   const uint8_t* blimit0,
+                                   const uint8_t* limit0,
+                                   const uint8_t* thresh0);
+#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_neon
 
 void aom_lpf_vertical_4_c(uint8_t* s,
                           int pitch,
@@ -1422,7 +1605,12 @@ void aom_lpf_vertical_4_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_c
+void aom_lpf_vertical_4_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_neon
 
 void aom_lpf_vertical_6_c(uint8_t* s,
                           int pitch,
@@ -1459,7 +1647,12 @@ void aom_lpf_vertical_6_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_c
+void aom_lpf_vertical_6_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_neon
 
 void aom_lpf_vertical_8_c(uint8_t* s,
                           int pitch,
@@ -1496,7 +1689,12 @@ void aom_lpf_vertical_8_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_c
+void aom_lpf_vertical_8_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_neon
 
 unsigned int aom_masked_sad128x128_c(const uint8_t* src,
                                      int src_stride,
@@ -2098,85 +2296,191 @@ void aom_paeth_predictor_16x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_c
+void aom_paeth_predictor_16x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_neon
 
 void aom_paeth_predictor_16x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
+void aom_paeth_predictor_16x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_neon
 
-void aom_paeth_predictor_16x8_c(uint8_t* dst,
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
+void aom_paeth_predictor_16x4_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_16x4 aom_paeth_predictor_16x4_neon
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_16x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x64 aom_paeth_predictor_16x64_neon
+
+void aom_paeth_predictor_16x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_16x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_neon
 
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c
+void aom_paeth_predictor_32x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_neon
 
 void aom_paeth_predictor_32x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_c
+void aom_paeth_predictor_32x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_neon
 
 void aom_paeth_predictor_32x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_c
+void aom_paeth_predictor_32x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_neon
+
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_32x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_32x8 aom_paeth_predictor_32x8_neon
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_4x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_4x16 aom_paeth_predictor_4x16_neon
 
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_c
+void aom_paeth_predictor_4x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_neon
 
 void aom_paeth_predictor_4x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
+void aom_paeth_predictor_4x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_neon
+
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_64x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x16 aom_paeth_predictor_64x16_neon
 
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_c
+void aom_paeth_predictor_64x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_neon
 
 void aom_paeth_predictor_64x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_c
+void aom_paeth_predictor_64x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_neon
 
 void aom_paeth_predictor_8x16_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
+void aom_paeth_predictor_8x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_neon
+
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_8x32_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_8x32 aom_paeth_predictor_8x32_neon
 
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_c
+void aom_paeth_predictor_8x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_neon
 
 void aom_paeth_predictor_8x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_c
+void aom_paeth_predictor_8x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_neon
 
 void aom_pixel_scale_c(const int16_t* src_diff,
                        ptrdiff_t src_stride,
@@ -3219,85 +3523,191 @@ void aom_smooth_h_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_c
+void aom_smooth_h_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_neon
 
 void aom_smooth_h_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c
+void aom_smooth_h_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_neon
 
-void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c
+void aom_smooth_h_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_16x4 aom_smooth_h_predictor_16x4_neon
 
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x64 aom_smooth_h_predictor_16x64_neon
+
+void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_neon
 
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_c
+void aom_smooth_h_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_neon
 
 void aom_smooth_h_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_c
+void aom_smooth_h_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_neon
 
 void aom_smooth_h_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_c
+void aom_smooth_h_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_neon
+
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_32x8 aom_smooth_h_predictor_32x8_neon
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_4x16 aom_smooth_h_predictor_4x16_neon
 
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_c
+void aom_smooth_h_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_neon
 
 void aom_smooth_h_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c
+void aom_smooth_h_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_neon
+
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x16 aom_smooth_h_predictor_64x16_neon
 
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_c
+void aom_smooth_h_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_neon
 
 void aom_smooth_h_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_c
+void aom_smooth_h_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_neon
 
 void aom_smooth_h_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c
+void aom_smooth_h_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_neon
+
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_8x32 aom_smooth_h_predictor_8x32_neon
 
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_c
+void aom_smooth_h_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_neon
 
 void aom_smooth_h_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_c
+void aom_smooth_h_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_neon
 
 void aom_smooth_predictor_16x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
@@ -3319,6 +3729,26 @@ void aom_smooth_predictor_16x32_neon(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_neon
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_16x4_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_16x4 aom_smooth_predictor_16x4_neon
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_16x64_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_predictor_16x64 aom_smooth_predictor_16x64_neon
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -3329,12 +3759,6 @@ void aom_smooth_predictor_16x8_neon(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_neon
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3365,6 +3789,26 @@ void aom_smooth_predictor_32x64_neon(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_smooth_predictor_32x64 aom_smooth_predictor_32x64_neon
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_32x8_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_32x8 aom_smooth_predictor_32x8_neon
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_4x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_4x16 aom_smooth_predictor_4x16_neon
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3385,6 +3829,16 @@ void aom_smooth_predictor_4x8_neon(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_neon
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_64x16_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_predictor_64x16 aom_smooth_predictor_64x16_neon
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3415,6 +3869,16 @@ void aom_smooth_predictor_8x16_neon(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_neon
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_8x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_8x32 aom_smooth_predictor_8x32_neon
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3439,85 +3903,191 @@ void aom_smooth_v_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_c
+void aom_smooth_v_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_neon
 
 void aom_smooth_v_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c
+void aom_smooth_v_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_neon
 
-void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c
+void aom_smooth_v_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_16x4 aom_smooth_v_predictor_16x4_neon
 
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x64 aom_smooth_v_predictor_16x64_neon
+
+void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_neon
 
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_c
+void aom_smooth_v_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_neon
 
 void aom_smooth_v_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_c
+void aom_smooth_v_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_neon
 
 void aom_smooth_v_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_c
+void aom_smooth_v_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_neon
+
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_32x8 aom_smooth_v_predictor_32x8_neon
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_4x16 aom_smooth_v_predictor_4x16_neon
 
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_c
+void aom_smooth_v_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_neon
 
 void aom_smooth_v_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c
+void aom_smooth_v_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_neon
+
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x16 aom_smooth_v_predictor_64x16_neon
 
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_c
+void aom_smooth_v_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_neon
 
 void aom_smooth_v_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_c
+void aom_smooth_v_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_neon
 
 void aom_smooth_v_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c
+void aom_smooth_v_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_neon
+
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_8x32 aom_smooth_v_predictor_8x32_neon
 
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_c
+void aom_smooth_v_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_neon
 
 void aom_smooth_v_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_c
+void aom_smooth_v_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_neon
 
 int64_t aom_sse_c(const uint8_t* a,
                   int a_stride,
@@ -4014,18 +4584,24 @@ void aom_v_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_c
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_c
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_c
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_c
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4048,6 +4624,18 @@ void aom_v_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_32x64 aom_v_predictor_32x64_c
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_c
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_c
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -4064,6 +4652,12 @@ void aom_v_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_c
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_64x16 aom_v_predictor_64x16_c
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4082,6 +4676,12 @@ void aom_v_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_c
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_c
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
diff --git a/chromium/third_party/libaom/source/config/linux/arm-neon/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/linux/arm-neon/config/av1_rtcd.h
index 23e831cc591..7c2e8affa18 100644
--- a/chromium/third_party/libaom/source/config/linux/arm-neon/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/arm-neon/config/av1_rtcd.h
@@ -183,6 +183,30 @@ void aom_upsampled_pred_c(MACROBLOCKD* xd,
                           int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_c
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+void av1_apply_selfguided_restoration_neon(const uint8_t* dat,
+                                           int width,
+                                           int height,
+                                           int stride,
+                                           int eps,
+                                           const int* xqd,
+                                           uint8_t* dst,
+                                           int dst_stride,
+                                           int32_t* tmpbuf,
+                                           int bit_depth,
+                                           int highbd);
+#define av1_apply_selfguided_restoration av1_apply_selfguided_restoration_neon
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -234,6 +258,14 @@ void av1_build_compound_diffwtd_mask_d16_neon(uint8_t* mask,
 #define av1_build_compound_diffwtd_mask_d16 \
   av1_build_compound_diffwtd_mask_d16_neon
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+#define av1_calc_frame_error av1_calc_frame_error_c
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -1314,6 +1346,28 @@ void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 void av1_round_shift_array_neon(int32_t* arr, int size, int bit);
 #define av1_round_shift_array av1_round_shift_array_neon
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+int av1_selfguided_restoration_neon(const uint8_t* dgd8,
+                                    int width,
+                                    int height,
+                                    int dgd_stride,
+                                    int32_t* flt0,
+                                    int32_t* flt1,
+                                    int flt_stride,
+                                    int sgr_params_idx,
+                                    int bit_depth,
+                                    int highbd);
+#define av1_selfguided_restoration av1_selfguided_restoration_neon
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -1330,6 +1384,44 @@ void av1_upsample_intra_edge_c(uint8_t* p, int sz);
 void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 #define av1_upsample_intra_edge_high av1_upsample_intra_edge_high_c
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+void av1_warp_affine_neon(const int32_t* mat,
+                          const uint8_t* ref,
+                          int width,
+                          int height,
+                          int stride,
+                          uint8_t* pred,
+                          int p_col,
+                          int p_row,
+                          int p_width,
+                          int p_height,
+                          int p_stride,
+                          int subsampling_x,
+                          int subsampling_y,
+                          ConvolveParams* conv_params,
+                          int16_t alpha,
+                          int16_t beta,
+                          int16_t gamma,
+                          int16_t delta);
+#define av1_warp_affine av1_warp_affine_neon
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
@@ -1610,7 +1702,15 @@ void cdef_find_dir_dual_c(const uint16_t* img1,
                           int coeff_shift,
                           int* out1,
                           int* out2);
-#define cdef_find_dir_dual cdef_find_dir_dual_c
+void cdef_find_dir_dual_neon(const uint16_t* img1,
+                             const uint16_t* img2,
+                             int stride,
+                             int32_t* var1,
+                             int32_t* var2,
+                             int coeff_shift,
+                             int* out1,
+                             int* out2);
+#define cdef_find_dir_dual cdef_find_dir_dual_neon
 
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_neon(TX_SIZE tx_size);
diff --git a/chromium/third_party/libaom/source/config/linux/arm/config/aom_config.c b/chromium/third_party/libaom/source/config/linux/arm/config/aom_config.c
index c6697d9573f..becd745fe8b 100644
--- a/chromium/third_party/libaom/source/config/linux/arm/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/linux/arm/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/armv7-linux-gcc.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384 -DENABLE_NEON=0";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/armv7-linux-gcc.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384 -DENABLE_NEON=0";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/linux/arm/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/linux/arm/config/aom_dsp_rtcd.h
index a87855d6013..df533799ed2 100644
--- a/chromium/third_party/libaom/source/config/linux/arm/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/arm/config/aom_dsp_rtcd.h
@@ -146,18 +146,24 @@ void aom_dc_128_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_c
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_c
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -176,6 +182,18 @@ void aom_dc_128_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_32x64 aom_dc_128_predictor_32x64_c
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_c
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_c
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -188,6 +206,12 @@ void aom_dc_128_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_64x16 aom_dc_128_predictor_64x16_c
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -206,6 +230,12 @@ void aom_dc_128_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_c
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -230,18 +260,24 @@ void aom_dc_left_predictor_16x32_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_c
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_c
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -260,6 +296,18 @@ void aom_dc_left_predictor_32x64_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_32x64 aom_dc_left_predictor_32x64_c
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_c
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_c
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -272,6 +320,12 @@ void aom_dc_left_predictor_4x8_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_64x16 aom_dc_left_predictor_64x16_c
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -290,6 +344,12 @@ void aom_dc_left_predictor_8x16_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_c
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -314,18 +374,24 @@ void aom_dc_predictor_16x32_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_c
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_c
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
                              const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -344,6 +410,18 @@ void aom_dc_predictor_32x64_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_32x64 aom_dc_predictor_32x64_c
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_c
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_c
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -356,6 +434,12 @@ void aom_dc_predictor_4x8_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_64x16 aom_dc_predictor_64x16_c
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -374,6 +458,12 @@ void aom_dc_predictor_8x16_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_c
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -398,18 +488,24 @@ void aom_dc_top_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_c
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_c
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -428,6 +524,18 @@ void aom_dc_top_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_32x64 aom_dc_top_predictor_32x64_c
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_c
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_c
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -440,6 +548,12 @@ void aom_dc_top_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_64x16 aom_dc_top_predictor_64x16_c
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -458,6 +572,12 @@ void aom_dc_top_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_c
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -895,18 +1015,24 @@ void aom_h_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_c
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_c
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_c
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_c
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -925,6 +1051,18 @@ void aom_h_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_c
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_c
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_c
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -937,6 +1075,12 @@ void aom_h_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_c
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_c
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -955,6 +1099,12 @@ void aom_h_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_c
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_c
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1836,18 +1986,24 @@ void aom_paeth_predictor_16x32_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
 
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_paeth_predictor_16x4 aom_paeth_predictor_16x4_c
+
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_paeth_predictor_16x64 aom_paeth_predictor_16x64_c
+
 void aom_paeth_predictor_16x8_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
 #define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
-
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -1866,6 +2022,18 @@ void aom_paeth_predictor_32x64_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_c
 
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_paeth_predictor_32x8 aom_paeth_predictor_32x8_c
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_paeth_predictor_4x16 aom_paeth_predictor_4x16_c
+
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -1878,6 +2046,12 @@ void aom_paeth_predictor_4x8_c(uint8_t* dst,
                                const uint8_t* left);
 #define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
 
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_paeth_predictor_64x16 aom_paeth_predictor_64x16_c
+
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -1896,6 +2070,12 @@ void aom_paeth_predictor_8x16_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
 
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_paeth_predictor_8x32 aom_paeth_predictor_8x32_c
+
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -2735,18 +2915,24 @@ void aom_smooth_h_predictor_16x32_c(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c
 
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_h_predictor_16x4 aom_smooth_h_predictor_16x4_c
+
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_h_predictor_16x64 aom_smooth_h_predictor_16x64_c
+
 void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
 #define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c
 
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
-
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -2765,6 +2951,18 @@ void aom_smooth_h_predictor_32x64_c(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_c
 
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_h_predictor_32x8 aom_smooth_h_predictor_32x8_c
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_h_predictor_4x16 aom_smooth_h_predictor_4x16_c
+
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -2777,6 +2975,12 @@ void aom_smooth_h_predictor_4x8_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c
 
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_h_predictor_64x16 aom_smooth_h_predictor_64x16_c
+
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -2795,6 +2999,12 @@ void aom_smooth_h_predictor_8x16_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c
 
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_h_predictor_8x32 aom_smooth_h_predictor_8x32_c
+
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -2819,18 +3029,24 @@ void aom_smooth_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_c
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_smooth_predictor_16x4 aom_smooth_predictor_16x4_c
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_smooth_predictor_16x64 aom_smooth_predictor_16x64_c
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_c
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -2849,6 +3065,18 @@ void aom_smooth_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_smooth_predictor_32x64 aom_smooth_predictor_32x64_c
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_smooth_predictor_32x8 aom_smooth_predictor_32x8_c
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_smooth_predictor_4x16 aom_smooth_predictor_4x16_c
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -2861,6 +3089,12 @@ void aom_smooth_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_c
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_smooth_predictor_64x16 aom_smooth_predictor_64x16_c
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -2879,6 +3113,12 @@ void aom_smooth_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_c
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_smooth_predictor_8x32 aom_smooth_predictor_8x32_c
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -2903,18 +3143,24 @@ void aom_smooth_v_predictor_16x32_c(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c
 
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_v_predictor_16x4 aom_smooth_v_predictor_16x4_c
+
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_v_predictor_16x64 aom_smooth_v_predictor_16x64_c
+
 void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
 #define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c
 
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
-
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -2933,6 +3179,18 @@ void aom_smooth_v_predictor_32x64_c(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_c
 
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_v_predictor_32x8 aom_smooth_v_predictor_32x8_c
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_v_predictor_4x16 aom_smooth_v_predictor_4x16_c
+
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -2945,6 +3203,12 @@ void aom_smooth_v_predictor_4x8_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c
 
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_v_predictor_64x16 aom_smooth_v_predictor_64x16_c
+
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -2963,6 +3227,12 @@ void aom_smooth_v_predictor_8x16_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c
 
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_v_predictor_8x32 aom_smooth_v_predictor_8x32_c
+
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3336,18 +3606,24 @@ void aom_v_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_c
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_c
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_c
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_c
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -3366,6 +3642,18 @@ void aom_v_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_32x64 aom_v_predictor_32x64_c
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_c
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_c
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -3378,6 +3666,12 @@ void aom_v_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_c
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_64x16 aom_v_predictor_64x16_c
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -3396,6 +3690,12 @@ void aom_v_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_c
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_c
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
diff --git a/chromium/third_party/libaom/source/config/linux/arm/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/linux/arm/config/av1_rtcd.h
index eb237126baa..dc3e04f11a4 100644
--- a/chromium/third_party/libaom/source/config/linux/arm/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/arm/config/av1_rtcd.h
@@ -168,6 +168,19 @@ void aom_upsampled_pred_c(MACROBLOCKD* xd,
                           int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_c
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+#define av1_apply_selfguided_restoration av1_apply_selfguided_restoration_c
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -202,6 +215,14 @@ void av1_build_compound_diffwtd_mask_d16_c(uint8_t* mask,
 #define av1_build_compound_diffwtd_mask_d16 \
   av1_build_compound_diffwtd_mask_d16_c
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+#define av1_calc_frame_error av1_calc_frame_error_c
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -875,6 +896,18 @@ void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG* src,
 void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 #define av1_round_shift_array av1_round_shift_array_c
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+#define av1_selfguided_restoration av1_selfguided_restoration_c
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -887,6 +920,26 @@ void av1_upsample_intra_edge_c(uint8_t* p, int sz);
 void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 #define av1_upsample_intra_edge_high av1_upsample_intra_edge_high_c
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+#define av1_warp_affine av1_warp_affine_c
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
diff --git a/chromium/third_party/libaom/source/config/linux/arm64/config/aom_config.c b/chromium/third_party/libaom/source/config/linux/arm64/config/aom_config.c
index 7c8f15b9d65..ff9196735fc 100644
--- a/chromium/third_party/libaom/source/config/linux/arm64/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/linux/arm64/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/arm64-linux-gcc.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/arm64-linux-gcc.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/linux/arm64/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/linux/arm64/config/aom_dsp_rtcd.h
index 155c11857c0..49fb79d7f18 100644
--- a/chromium/third_party/libaom/source/config/linux/arm64/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/arm64/config/aom_dsp_rtcd.h
@@ -34,7 +34,12 @@ void aom_avg_8x8_quad_c(const uint8_t* s,
                         int x16_idx,
                         int y16_idx,
                         int* avg);
-#define aom_avg_8x8_quad aom_avg_8x8_quad_c
+void aom_avg_8x8_quad_neon(const uint8_t* s,
+                           int p,
+                           int x16_idx,
+                           int y16_idx,
+                           int* avg);
+#define aom_avg_8x8_quad aom_avg_8x8_quad_neon
 
 void aom_blend_a64_hmask_c(uint8_t* dst,
                            uint32_t dst_stride,
@@ -176,18 +181,24 @@ void aom_dc_128_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_c
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_c
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -210,6 +221,18 @@ void aom_dc_128_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_32x64 aom_dc_128_predictor_32x64_c
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_c
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_c
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -226,6 +249,12 @@ void aom_dc_128_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_64x16 aom_dc_128_predictor_64x16_c
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -244,6 +273,12 @@ void aom_dc_128_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_c
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -276,18 +311,24 @@ void aom_dc_left_predictor_16x32_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_c
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_c
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -310,6 +351,18 @@ void aom_dc_left_predictor_32x64_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_32x64 aom_dc_left_predictor_32x64_c
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_c
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_c
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -326,6 +379,12 @@ void aom_dc_left_predictor_4x8_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_64x16 aom_dc_left_predictor_64x16_c
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -344,6 +403,12 @@ void aom_dc_left_predictor_8x16_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_c
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -376,18 +441,24 @@ void aom_dc_predictor_16x32_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_c
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_c
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
                              const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -410,6 +481,18 @@ void aom_dc_predictor_32x64_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_32x64 aom_dc_predictor_32x64_c
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_c
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_c
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -426,6 +509,12 @@ void aom_dc_predictor_4x8_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_64x16 aom_dc_predictor_64x16_c
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -444,6 +533,12 @@ void aom_dc_predictor_8x16_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_c
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -476,18 +571,24 @@ void aom_dc_top_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_c
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_c
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -510,6 +611,18 @@ void aom_dc_top_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_32x64 aom_dc_top_predictor_32x64_c
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_c
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_c
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -526,6 +639,12 @@ void aom_dc_top_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_64x16 aom_dc_top_predictor_64x16_c
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -544,6 +663,12 @@ void aom_dc_top_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_c
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -1014,18 +1139,24 @@ void aom_h_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_c
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_c
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_c
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_c
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1048,6 +1179,18 @@ void aom_h_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_c
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_c
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_c
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1064,6 +1207,12 @@ void aom_h_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_c
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_c
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1082,6 +1231,12 @@ void aom_h_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_c
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_c
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1127,7 +1282,10 @@ void aom_hadamard_8x8_neon(const int16_t* src_diff,
 void aom_hadamard_8x8_dual_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
                              int16_t* coeff);
-#define aom_hadamard_8x8_dual aom_hadamard_8x8_dual_c
+void aom_hadamard_8x8_dual_neon(const int16_t* src_diff,
+                                ptrdiff_t src_stride,
+                                int16_t* coeff);
+#define aom_hadamard_8x8_dual aom_hadamard_8x8_dual_neon
 
 void aom_hadamard_lp_16x16_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
@@ -1237,7 +1395,12 @@ void aom_lpf_horizontal_14_quad_c(uint8_t* s,
                                   const uint8_t* blimit0,
                                   const uint8_t* limit0,
                                   const uint8_t* thresh0);
-#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_c
+void aom_lpf_horizontal_14_quad_neon(uint8_t* s,
+                                     int pitch,
+                                     const uint8_t* blimit0,
+                                     const uint8_t* limit0,
+                                     const uint8_t* thresh0);
+#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_neon
 
 void aom_lpf_horizontal_4_c(uint8_t* s,
                             int pitch,
@@ -1274,7 +1437,12 @@ void aom_lpf_horizontal_4_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_c
+void aom_lpf_horizontal_4_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_neon
 
 void aom_lpf_horizontal_6_c(uint8_t* s,
                             int pitch,
@@ -1311,7 +1479,12 @@ void aom_lpf_horizontal_6_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_c
+void aom_lpf_horizontal_6_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_neon
 
 void aom_lpf_horizontal_8_c(uint8_t* s,
                             int pitch,
@@ -1348,7 +1521,12 @@ void aom_lpf_horizontal_8_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_c
+void aom_lpf_horizontal_8_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_neon
 
 void aom_lpf_vertical_14_c(uint8_t* s,
                            int pitch,
@@ -1385,7 +1563,12 @@ void aom_lpf_vertical_14_quad_c(uint8_t* s,
                                 const uint8_t* blimit0,
                                 const uint8_t* limit0,
                                 const uint8_t* thresh0);
-#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_c
+void aom_lpf_vertical_14_quad_neon(uint8_t* s,
+                                   int pitch,
+                                   const uint8_t* blimit0,
+                                   const uint8_t* limit0,
+                                   const uint8_t* thresh0);
+#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_neon
 
 void aom_lpf_vertical_4_c(uint8_t* s,
                           int pitch,
@@ -1422,7 +1605,12 @@ void aom_lpf_vertical_4_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_c
+void aom_lpf_vertical_4_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_neon
 
 void aom_lpf_vertical_6_c(uint8_t* s,
                           int pitch,
@@ -1459,7 +1647,12 @@ void aom_lpf_vertical_6_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_c
+void aom_lpf_vertical_6_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_neon
 
 void aom_lpf_vertical_8_c(uint8_t* s,
                           int pitch,
@@ -1496,7 +1689,12 @@ void aom_lpf_vertical_8_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_c
+void aom_lpf_vertical_8_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_neon
 
 unsigned int aom_masked_sad128x128_c(const uint8_t* src,
                                      int src_stride,
@@ -2098,85 +2296,191 @@ void aom_paeth_predictor_16x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_c
+void aom_paeth_predictor_16x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_neon
 
 void aom_paeth_predictor_16x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
+void aom_paeth_predictor_16x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_neon
 
-void aom_paeth_predictor_16x8_c(uint8_t* dst,
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
+void aom_paeth_predictor_16x4_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_16x4 aom_paeth_predictor_16x4_neon
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_16x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x64 aom_paeth_predictor_16x64_neon
+
+void aom_paeth_predictor_16x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_16x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_neon
 
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c
+void aom_paeth_predictor_32x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_neon
 
 void aom_paeth_predictor_32x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_c
+void aom_paeth_predictor_32x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_neon
 
 void aom_paeth_predictor_32x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_c
+void aom_paeth_predictor_32x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_neon
+
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_32x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_32x8 aom_paeth_predictor_32x8_neon
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_4x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_4x16 aom_paeth_predictor_4x16_neon
 
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_c
+void aom_paeth_predictor_4x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_neon
 
 void aom_paeth_predictor_4x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
+void aom_paeth_predictor_4x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_neon
+
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_64x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x16 aom_paeth_predictor_64x16_neon
 
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_c
+void aom_paeth_predictor_64x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_neon
 
 void aom_paeth_predictor_64x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_c
+void aom_paeth_predictor_64x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_neon
 
 void aom_paeth_predictor_8x16_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
+void aom_paeth_predictor_8x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_neon
+
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_8x32_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_8x32 aom_paeth_predictor_8x32_neon
 
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_c
+void aom_paeth_predictor_8x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_neon
 
 void aom_paeth_predictor_8x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_c
+void aom_paeth_predictor_8x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_neon
 
 void aom_pixel_scale_c(const int16_t* src_diff,
                        ptrdiff_t src_stride,
@@ -3219,85 +3523,191 @@ void aom_smooth_h_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_c
+void aom_smooth_h_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_neon
 
 void aom_smooth_h_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c
+void aom_smooth_h_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_neon
 
-void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c
+void aom_smooth_h_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_16x4 aom_smooth_h_predictor_16x4_neon
 
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x64 aom_smooth_h_predictor_16x64_neon
+
+void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_neon
 
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_c
+void aom_smooth_h_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_neon
 
 void aom_smooth_h_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_c
+void aom_smooth_h_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_neon
 
 void aom_smooth_h_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_c
+void aom_smooth_h_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_neon
+
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_32x8 aom_smooth_h_predictor_32x8_neon
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_4x16 aom_smooth_h_predictor_4x16_neon
 
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_c
+void aom_smooth_h_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_neon
 
 void aom_smooth_h_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c
+void aom_smooth_h_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_neon
+
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x16 aom_smooth_h_predictor_64x16_neon
 
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_c
+void aom_smooth_h_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_neon
 
 void aom_smooth_h_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_c
+void aom_smooth_h_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_neon
 
 void aom_smooth_h_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c
+void aom_smooth_h_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_neon
+
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_8x32 aom_smooth_h_predictor_8x32_neon
 
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_c
+void aom_smooth_h_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_neon
 
 void aom_smooth_h_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_c
+void aom_smooth_h_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_neon
 
 void aom_smooth_predictor_16x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
@@ -3319,6 +3729,26 @@ void aom_smooth_predictor_16x32_neon(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_neon
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_16x4_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_16x4 aom_smooth_predictor_16x4_neon
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_16x64_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_predictor_16x64 aom_smooth_predictor_16x64_neon
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -3329,12 +3759,6 @@ void aom_smooth_predictor_16x8_neon(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_neon
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3365,6 +3789,26 @@ void aom_smooth_predictor_32x64_neon(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_smooth_predictor_32x64 aom_smooth_predictor_32x64_neon
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_32x8_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_32x8 aom_smooth_predictor_32x8_neon
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_4x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_4x16 aom_smooth_predictor_4x16_neon
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3385,6 +3829,16 @@ void aom_smooth_predictor_4x8_neon(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_neon
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_64x16_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_predictor_64x16 aom_smooth_predictor_64x16_neon
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3415,6 +3869,16 @@ void aom_smooth_predictor_8x16_neon(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_neon
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_8x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_8x32 aom_smooth_predictor_8x32_neon
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3439,85 +3903,191 @@ void aom_smooth_v_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_c
+void aom_smooth_v_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_neon
 
 void aom_smooth_v_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c
+void aom_smooth_v_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_neon
 
-void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c
+void aom_smooth_v_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_16x4 aom_smooth_v_predictor_16x4_neon
 
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x64 aom_smooth_v_predictor_16x64_neon
+
+void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_neon
 
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_c
+void aom_smooth_v_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_neon
 
 void aom_smooth_v_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_c
+void aom_smooth_v_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_neon
 
 void aom_smooth_v_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_c
+void aom_smooth_v_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_neon
+
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_32x8 aom_smooth_v_predictor_32x8_neon
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_4x16 aom_smooth_v_predictor_4x16_neon
 
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_c
+void aom_smooth_v_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_neon
 
 void aom_smooth_v_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c
+void aom_smooth_v_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_neon
+
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x16 aom_smooth_v_predictor_64x16_neon
 
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_c
+void aom_smooth_v_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_neon
 
 void aom_smooth_v_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_c
+void aom_smooth_v_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_neon
 
 void aom_smooth_v_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c
+void aom_smooth_v_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_neon
+
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_8x32 aom_smooth_v_predictor_8x32_neon
 
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_c
+void aom_smooth_v_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_neon
 
 void aom_smooth_v_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_c
+void aom_smooth_v_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_neon
 
 int64_t aom_sse_c(const uint8_t* a,
                   int a_stride,
@@ -4014,18 +4584,24 @@ void aom_v_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_c
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_c
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_c
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_c
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4048,6 +4624,18 @@ void aom_v_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_32x64 aom_v_predictor_32x64_c
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_c
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_c
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -4064,6 +4652,12 @@ void aom_v_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_c
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_64x16 aom_v_predictor_64x16_c
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4082,6 +4676,12 @@ void aom_v_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_c
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_c
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
diff --git a/chromium/third_party/libaom/source/config/linux/arm64/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/linux/arm64/config/av1_rtcd.h
index 23e831cc591..7c2e8affa18 100644
--- a/chromium/third_party/libaom/source/config/linux/arm64/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/arm64/config/av1_rtcd.h
@@ -183,6 +183,30 @@ void aom_upsampled_pred_c(MACROBLOCKD* xd,
                           int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_c
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+void av1_apply_selfguided_restoration_neon(const uint8_t* dat,
+                                           int width,
+                                           int height,
+                                           int stride,
+                                           int eps,
+                                           const int* xqd,
+                                           uint8_t* dst,
+                                           int dst_stride,
+                                           int32_t* tmpbuf,
+                                           int bit_depth,
+                                           int highbd);
+#define av1_apply_selfguided_restoration av1_apply_selfguided_restoration_neon
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -234,6 +258,14 @@ void av1_build_compound_diffwtd_mask_d16_neon(uint8_t* mask,
 #define av1_build_compound_diffwtd_mask_d16 \
   av1_build_compound_diffwtd_mask_d16_neon
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+#define av1_calc_frame_error av1_calc_frame_error_c
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -1314,6 +1346,28 @@ void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 void av1_round_shift_array_neon(int32_t* arr, int size, int bit);
 #define av1_round_shift_array av1_round_shift_array_neon
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+int av1_selfguided_restoration_neon(const uint8_t* dgd8,
+                                    int width,
+                                    int height,
+                                    int dgd_stride,
+                                    int32_t* flt0,
+                                    int32_t* flt1,
+                                    int flt_stride,
+                                    int sgr_params_idx,
+                                    int bit_depth,
+                                    int highbd);
+#define av1_selfguided_restoration av1_selfguided_restoration_neon
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -1330,6 +1384,44 @@ void av1_upsample_intra_edge_c(uint8_t* p, int sz);
 void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 #define av1_upsample_intra_edge_high av1_upsample_intra_edge_high_c
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+void av1_warp_affine_neon(const int32_t* mat,
+                          const uint8_t* ref,
+                          int width,
+                          int height,
+                          int stride,
+                          uint8_t* pred,
+                          int p_col,
+                          int p_row,
+                          int p_width,
+                          int p_height,
+                          int p_stride,
+                          int subsampling_x,
+                          int subsampling_y,
+                          ConvolveParams* conv_params,
+                          int16_t alpha,
+                          int16_t beta,
+                          int16_t gamma,
+                          int16_t delta);
+#define av1_warp_affine av1_warp_affine_neon
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
@@ -1610,7 +1702,15 @@ void cdef_find_dir_dual_c(const uint16_t* img1,
                           int coeff_shift,
                           int* out1,
                           int* out2);
-#define cdef_find_dir_dual cdef_find_dir_dual_c
+void cdef_find_dir_dual_neon(const uint16_t* img1,
+                             const uint16_t* img2,
+                             int stride,
+                             int32_t* var1,
+                             int32_t* var2,
+                             int coeff_shift,
+                             int* out1,
+                             int* out2);
+#define cdef_find_dir_dual cdef_find_dir_dual_neon
 
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_neon(TX_SIZE tx_size);
diff --git a/chromium/third_party/libaom/source/config/linux/generic/config/aom_config.c b/chromium/third_party/libaom/source/config/linux/generic/config/aom_config.c
index d4d7e87a64e..d7a0494ec2c 100644
--- a/chromium/third_party/libaom/source/config/linux/generic/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/linux/generic/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DAOM_TARGET_CPU=generic -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DAOM_TARGET_CPU=generic -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/linux/generic/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/linux/generic/config/aom_dsp_rtcd.h
index 5fbcf9c3384..7ec152b1f42 100644
--- a/chromium/third_party/libaom/source/config/linux/generic/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/generic/config/aom_dsp_rtcd.h
@@ -146,18 +146,24 @@ void aom_dc_128_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_c
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_c
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -176,6 +182,18 @@ void aom_dc_128_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_32x64 aom_dc_128_predictor_32x64_c
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_c
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_c
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -188,6 +206,12 @@ void aom_dc_128_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_64x16 aom_dc_128_predictor_64x16_c
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -206,6 +230,12 @@ void aom_dc_128_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_c
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -230,18 +260,24 @@ void aom_dc_left_predictor_16x32_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_c
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_c
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -260,6 +296,18 @@ void aom_dc_left_predictor_32x64_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_32x64 aom_dc_left_predictor_32x64_c
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_c
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_c
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -272,6 +320,12 @@ void aom_dc_left_predictor_4x8_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_64x16 aom_dc_left_predictor_64x16_c
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -290,6 +344,12 @@ void aom_dc_left_predictor_8x16_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_c
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -314,18 +374,24 @@ void aom_dc_predictor_16x32_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_c
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_c
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
                              const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -344,6 +410,18 @@ void aom_dc_predictor_32x64_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_32x64 aom_dc_predictor_32x64_c
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_c
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_c
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -356,6 +434,12 @@ void aom_dc_predictor_4x8_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_64x16 aom_dc_predictor_64x16_c
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -374,6 +458,12 @@ void aom_dc_predictor_8x16_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_c
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -398,18 +488,24 @@ void aom_dc_top_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_c
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_c
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -428,6 +524,18 @@ void aom_dc_top_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_32x64 aom_dc_top_predictor_32x64_c
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_c
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_c
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -440,6 +548,12 @@ void aom_dc_top_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_64x16 aom_dc_top_predictor_64x16_c
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -458,6 +572,12 @@ void aom_dc_top_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_c
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -895,18 +1015,24 @@ void aom_h_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_c
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_c
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_c
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_c
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -925,6 +1051,18 @@ void aom_h_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_c
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_c
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_c
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -937,6 +1075,12 @@ void aom_h_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_c
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_c
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -955,6 +1099,12 @@ void aom_h_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_c
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_c
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1836,18 +1986,24 @@ void aom_paeth_predictor_16x32_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
 
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_paeth_predictor_16x4 aom_paeth_predictor_16x4_c
+
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_paeth_predictor_16x64 aom_paeth_predictor_16x64_c
+
 void aom_paeth_predictor_16x8_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
 #define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
-
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -1866,6 +2022,18 @@ void aom_paeth_predictor_32x64_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_c
 
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_paeth_predictor_32x8 aom_paeth_predictor_32x8_c
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_paeth_predictor_4x16 aom_paeth_predictor_4x16_c
+
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -1878,6 +2046,12 @@ void aom_paeth_predictor_4x8_c(uint8_t* dst,
                                const uint8_t* left);
 #define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
 
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_paeth_predictor_64x16 aom_paeth_predictor_64x16_c
+
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -1896,6 +2070,12 @@ void aom_paeth_predictor_8x16_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
 
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_paeth_predictor_8x32 aom_paeth_predictor_8x32_c
+
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -2735,18 +2915,24 @@ void aom_smooth_h_predictor_16x32_c(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c
 
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_h_predictor_16x4 aom_smooth_h_predictor_16x4_c
+
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_h_predictor_16x64 aom_smooth_h_predictor_16x64_c
+
 void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
 #define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c
 
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
-
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -2765,6 +2951,18 @@ void aom_smooth_h_predictor_32x64_c(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_c
 
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_h_predictor_32x8 aom_smooth_h_predictor_32x8_c
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_h_predictor_4x16 aom_smooth_h_predictor_4x16_c
+
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -2777,6 +2975,12 @@ void aom_smooth_h_predictor_4x8_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c
 
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_h_predictor_64x16 aom_smooth_h_predictor_64x16_c
+
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -2795,6 +2999,12 @@ void aom_smooth_h_predictor_8x16_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c
 
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_h_predictor_8x32 aom_smooth_h_predictor_8x32_c
+
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -2819,18 +3029,24 @@ void aom_smooth_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_c
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_smooth_predictor_16x4 aom_smooth_predictor_16x4_c
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_smooth_predictor_16x64 aom_smooth_predictor_16x64_c
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_c
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -2849,6 +3065,18 @@ void aom_smooth_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_smooth_predictor_32x64 aom_smooth_predictor_32x64_c
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_smooth_predictor_32x8 aom_smooth_predictor_32x8_c
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_smooth_predictor_4x16 aom_smooth_predictor_4x16_c
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -2861,6 +3089,12 @@ void aom_smooth_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_c
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_smooth_predictor_64x16 aom_smooth_predictor_64x16_c
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -2879,6 +3113,12 @@ void aom_smooth_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_c
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_smooth_predictor_8x32 aom_smooth_predictor_8x32_c
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -2903,18 +3143,24 @@ void aom_smooth_v_predictor_16x32_c(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c
 
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_v_predictor_16x4 aom_smooth_v_predictor_16x4_c
+
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_v_predictor_16x64 aom_smooth_v_predictor_16x64_c
+
 void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
 #define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c
 
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
-
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -2933,6 +3179,18 @@ void aom_smooth_v_predictor_32x64_c(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_c
 
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_v_predictor_32x8 aom_smooth_v_predictor_32x8_c
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_v_predictor_4x16 aom_smooth_v_predictor_4x16_c
+
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -2945,6 +3203,12 @@ void aom_smooth_v_predictor_4x8_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c
 
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_v_predictor_64x16 aom_smooth_v_predictor_64x16_c
+
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -2963,6 +3227,12 @@ void aom_smooth_v_predictor_8x16_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c
 
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_smooth_v_predictor_8x32 aom_smooth_v_predictor_8x32_c
+
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3336,18 +3606,24 @@ void aom_v_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_c
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_c
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_c
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_c
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -3366,6 +3642,18 @@ void aom_v_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_32x64 aom_v_predictor_32x64_c
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_c
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_c
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -3378,6 +3666,12 @@ void aom_v_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_c
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_64x16 aom_v_predictor_64x16_c
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -3396,6 +3690,12 @@ void aom_v_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_c
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_c
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
diff --git a/chromium/third_party/libaom/source/config/linux/generic/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/linux/generic/config/av1_rtcd.h
index eb649a65e9c..c0ed827d1fd 100644
--- a/chromium/third_party/libaom/source/config/linux/generic/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/generic/config/av1_rtcd.h
@@ -168,6 +168,19 @@ void aom_upsampled_pred_c(MACROBLOCKD* xd,
                           int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_c
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+#define av1_apply_selfguided_restoration av1_apply_selfguided_restoration_c
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -202,6 +215,14 @@ void av1_build_compound_diffwtd_mask_d16_c(uint8_t* mask,
 #define av1_build_compound_diffwtd_mask_d16 \
   av1_build_compound_diffwtd_mask_d16_c
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+#define av1_calc_frame_error av1_calc_frame_error_c
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -875,6 +896,18 @@ void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG* src,
 void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 #define av1_round_shift_array av1_round_shift_array_c
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+#define av1_selfguided_restoration av1_selfguided_restoration_c
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -887,6 +920,26 @@ void av1_upsample_intra_edge_c(uint8_t* p, int sz);
 void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 #define av1_upsample_intra_edge_high av1_upsample_intra_edge_high_c
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+#define av1_warp_affine av1_warp_affine_c
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
diff --git a/chromium/third_party/libaom/source/config/linux/ia32/config/aom_config.c b/chromium/third_party/libaom/source/config/linux/ia32/config/aom_config.c
index aaa8142b8b8..dadb41a0d60 100644
--- a/chromium/third_party/libaom/source/config/linux/ia32/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/linux/ia32/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/x86-linux.cmake\" -DAOM_RTCD_FLAGS=--require-mmx;--require-sse;--require-sse2 -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_PIC=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/x86-linux.cmake\" -DAOM_RTCD_FLAGS=--require-mmx;--require-sse;--require-sse2 -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_PIC=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/linux/ia32/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/linux/ia32/config/aom_dsp_rtcd.h
index aa189c45f4a..a76e597b134 100644
--- a/chromium/third_party/libaom/source/config/linux/ia32/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/ia32/config/aom_dsp_rtcd.h
@@ -360,6 +360,26 @@ void aom_dc_128_predictor_16x32_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_sse2
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_16x4_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_sse2
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_128_predictor_16x64_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_sse2
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -370,12 +390,6 @@ void aom_dc_128_predictor_16x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -427,6 +441,26 @@ RTCD_EXTERN void (*aom_dc_128_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_32x8_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_sse2
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_4x16_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_sse2
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -447,6 +481,23 @@ void aom_dc_128_predictor_4x8_sse2(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_sse2
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_128_predictor_64x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_dc_128_predictor_64x16_avx2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -491,6 +542,16 @@ void aom_dc_128_predictor_8x16_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_sse2
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_8x32_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_sse2
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -531,6 +592,26 @@ void aom_dc_left_predictor_16x32_sse2(uint8_t* dst,
                                       const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_sse2
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_16x4_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_sse2
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_dc_left_predictor_16x64_sse2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_sse2
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -541,12 +622,6 @@ void aom_dc_left_predictor_16x8_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -598,6 +673,26 @@ RTCD_EXTERN void (*aom_dc_left_predictor_32x64)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_32x8_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_sse2
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_4x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_sse2
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -618,6 +713,23 @@ void aom_dc_left_predictor_4x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_sse2
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_dc_left_predictor_64x16_sse2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+void aom_dc_left_predictor_64x16_avx2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -662,6 +774,16 @@ void aom_dc_left_predictor_8x16_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_sse2
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_8x32_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_sse2
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -702,6 +824,26 @@ void aom_dc_predictor_16x32_sse2(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_sse2
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_16x4_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_sse2
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+void aom_dc_predictor_16x64_sse2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_sse2
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -712,12 +854,6 @@ void aom_dc_predictor_16x8_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -769,6 +905,26 @@ RTCD_EXTERN void (*aom_dc_predictor_32x64)(uint8_t* dst,
                                            const uint8_t* above,
                                            const uint8_t* left);
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_32x8_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_sse2
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_4x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_sse2
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -789,6 +945,23 @@ void aom_dc_predictor_4x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_sse2
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+void aom_dc_predictor_64x16_sse2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_predictor_64x16_avx2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_predictor_64x16)(uint8_t* dst,
+                                           ptrdiff_t y_stride,
+                                           const uint8_t* above,
+                                           const uint8_t* left);
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -833,6 +1006,16 @@ void aom_dc_predictor_8x16_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_sse2
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_8x32_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_sse2
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -873,6 +1056,26 @@ void aom_dc_top_predictor_16x32_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_sse2
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_16x4_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_sse2
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_top_predictor_16x64_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_sse2
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -883,12 +1086,6 @@ void aom_dc_top_predictor_16x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -940,6 +1137,26 @@ RTCD_EXTERN void (*aom_dc_top_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_32x8_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_sse2
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_4x16_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_sse2
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -960,6 +1177,23 @@ void aom_dc_top_predictor_4x8_sse2(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_sse2
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_top_predictor_64x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_dc_top_predictor_64x16_avx2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -1004,6 +1238,16 @@ void aom_dc_top_predictor_8x16_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_sse2
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_8x32_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_sse2
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -2024,6 +2268,26 @@ void aom_h_predictor_16x32_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_sse2
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_16x4_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_sse2
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_h_predictor_16x64_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_sse2
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -2034,12 +2298,6 @@ void aom_h_predictor_16x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -2077,6 +2335,26 @@ void aom_h_predictor_32x64_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_sse2
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_32x8_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_sse2
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_4x16_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_sse2
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -2097,6 +2375,16 @@ void aom_h_predictor_4x8_sse2(uint8_t* dst,
                               const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_sse2
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_h_predictor_64x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_sse2
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -2127,6 +2415,16 @@ void aom_h_predictor_8x16_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_sse2
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_8x32_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_sse2
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -4357,6 +4655,36 @@ RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_16x4_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_16x64_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_paeth_predictor_16x64_avx2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_paeth_predictor_16x8_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -4374,12 +4702,6 @@ RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
-
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -4431,6 +4753,32 @@ RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_32x8_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_4x16_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -4457,6 +4805,23 @@ RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t* dst,
                                             const uint8_t* above,
                                             const uint8_t* left);
 
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_64x16_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_paeth_predictor_64x16_avx2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -4504,6 +4869,19 @@ RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_8x32_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -6256,6 +6634,32 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_16x4_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_16x64_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -6269,12 +6673,6 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
-
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6314,6 +6712,32 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_32x8_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_4x16_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6340,6 +6764,19 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_64x16_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6379,6 +6816,19 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_8x32_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6431,6 +6881,32 @@ RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_16x4_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_16x64_ssse3(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -6444,12 +6920,6 @@ RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6489,6 +6959,32 @@ RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_32x8_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_4x16_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -6515,6 +7011,19 @@ RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_64x16_ssse3(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6554,6 +7063,19 @@ RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_8x32_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -6606,6 +7128,32 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_16x4_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_16x64_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -6619,12 +7167,6 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
-
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6664,6 +7206,32 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_32x8_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_4x16_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6690,6 +7258,19 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_64x16_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6729,6 +7310,19 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_8x32_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -8035,6 +8629,26 @@ void aom_v_predictor_16x32_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_sse2
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_16x4_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_sse2
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_v_predictor_16x64_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_sse2
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -8045,12 +8659,6 @@ void aom_v_predictor_16x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -8102,6 +8710,26 @@ RTCD_EXTERN void (*aom_v_predictor_32x64)(uint8_t* dst,
                                           const uint8_t* above,
                                           const uint8_t* left);
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_32x8_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_sse2
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_4x16_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_sse2
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -8122,6 +8750,23 @@ void aom_v_predictor_4x8_sse2(uint8_t* dst,
                               const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_sse2
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_v_predictor_64x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_v_predictor_64x16_avx2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+RTCD_EXTERN void (*aom_v_predictor_64x16)(uint8_t* dst,
+                                          ptrdiff_t y_stride,
+                                          const uint8_t* above,
+                                          const uint8_t* left);
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -8166,6 +8811,16 @@ void aom_v_predictor_8x16_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_sse2
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_8x32_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_sse2
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -8579,6 +9234,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_avx2;
+  aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_avx2;
   aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_avx2;
@@ -8594,6 +9252,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_avx2;
+  aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_avx2;
   aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_avx2;
@@ -8609,6 +9270,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_predictor_32x64 = aom_dc_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_predictor_32x64 = aom_dc_predictor_32x64_avx2;
+  aom_dc_predictor_64x16 = aom_dc_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_predictor_64x16 = aom_dc_predictor_64x16_avx2;
   aom_dc_predictor_64x32 = aom_dc_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_predictor_64x32 = aom_dc_predictor_64x32_avx2;
@@ -8624,6 +9288,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_avx2;
+  aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_avx2;
   aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_avx2;
@@ -9016,6 +9683,14 @@ static void setup_rtcd_internal(void) {
     aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_ssse3;
   if (flags & HAS_AVX2)
     aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_avx2;
+  aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_ssse3;
+  aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_ssse3;
+  if (flags & HAS_AVX2)
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_avx2;
   aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_ssse3;
@@ -9036,12 +9711,23 @@ static void setup_rtcd_internal(void) {
     aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_ssse3;
   if (flags & HAS_AVX2)
     aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_avx2;
+  aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_ssse3;
+  aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_ssse3;
   aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_ssse3;
   aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_ssse3;
+  aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_ssse3;
+  if (flags & HAS_AVX2)
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_avx2;
   aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_ssse3;
@@ -9055,6 +9741,9 @@ static void setup_rtcd_internal(void) {
   aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_ssse3;
+  aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_ssse3;
   aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3;
@@ -9208,6 +9897,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_ssse3;
+  aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_ssse3;
+  aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_ssse3;
   aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_ssse3;
@@ -9220,12 +9915,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_ssse3;
+  aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_ssse3;
+  aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_ssse3;
   aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_ssse3;
   aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_ssse3;
+  aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_ssse3;
   aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_ssse3;
@@ -9235,6 +9939,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_ssse3;
+  aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_ssse3;
   aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_ssse3;
@@ -9247,6 +9954,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_ssse3;
+  aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_ssse3;
+  aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_ssse3;
   aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_ssse3;
@@ -9259,12 +9972,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_ssse3;
+  aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_ssse3;
+  aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_ssse3;
   aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_ssse3;
   aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_ssse3;
+  aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_ssse3;
   aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_ssse3;
@@ -9274,6 +9996,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_ssse3;
+  aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_ssse3;
   aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_ssse3;
@@ -9286,6 +10011,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_ssse3;
+  aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_ssse3;
+  aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_ssse3;
   aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_ssse3;
@@ -9298,12 +10029,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_ssse3;
+  aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_ssse3;
+  aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_ssse3;
   aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_ssse3;
   aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_ssse3;
+  aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_ssse3;
   aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_ssse3;
@@ -9313,6 +10053,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_ssse3;
+  aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_ssse3;
   aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_ssse3;
@@ -9476,6 +10219,9 @@ static void setup_rtcd_internal(void) {
   aom_v_predictor_32x64 = aom_v_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_v_predictor_32x64 = aom_v_predictor_32x64_avx2;
+  aom_v_predictor_64x16 = aom_v_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_v_predictor_64x16 = aom_v_predictor_64x16_avx2;
   aom_v_predictor_64x32 = aom_v_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_v_predictor_64x32 = aom_v_predictor_64x32_avx2;
diff --git a/chromium/third_party/libaom/source/config/linux/ia32/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/linux/ia32/config/av1_rtcd.h
index 8d516d90bde..6d8a64d1270 100644
--- a/chromium/third_party/libaom/source/config/linux/ia32/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/ia32/config/av1_rtcd.h
@@ -242,6 +242,51 @@ void aom_upsampled_pred_sse2(MACROBLOCKD* xd,
                              int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_sse2
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+void av1_apply_selfguided_restoration_sse4_1(const uint8_t* dat,
+                                             int width,
+                                             int height,
+                                             int stride,
+                                             int eps,
+                                             const int* xqd,
+                                             uint8_t* dst,
+                                             int dst_stride,
+                                             int32_t* tmpbuf,
+                                             int bit_depth,
+                                             int highbd);
+void av1_apply_selfguided_restoration_avx2(const uint8_t* dat,
+                                           int width,
+                                           int height,
+                                           int stride,
+                                           int eps,
+                                           const int* xqd,
+                                           uint8_t* dst,
+                                           int dst_stride,
+                                           int32_t* tmpbuf,
+                                           int bit_depth,
+                                           int highbd);
+RTCD_EXTERN void (*av1_apply_selfguided_restoration)(const uint8_t* dat,
+                                                     int width,
+                                                     int height,
+                                                     int stride,
+                                                     int eps,
+                                                     const int* xqd,
+                                                     uint8_t* dst,
+                                                     int dst_stride,
+                                                     int32_t* tmpbuf,
+                                                     int bit_depth,
+                                                     int highbd);
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -347,6 +392,31 @@ RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(
     ConvolveParams* conv_params,
     int bd);
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+int64_t av1_calc_frame_error_sse2(const uint8_t* const ref,
+                                  int stride,
+                                  const uint8_t* const dst,
+                                  int p_width,
+                                  int p_height,
+                                  int p_stride);
+int64_t av1_calc_frame_error_avx2(const uint8_t* const ref,
+                                  int stride,
+                                  const uint8_t* const dst,
+                                  int p_width,
+                                  int p_height,
+                                  int p_stride);
+RTCD_EXTERN int64_t (*av1_calc_frame_error)(const uint8_t* const ref,
+                                            int stride,
+                                            const uint8_t* const dst,
+                                            int p_width,
+                                            int p_height,
+                                            int p_stride);
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -1839,6 +1909,47 @@ void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 void av1_round_shift_array_sse4_1(int32_t* arr, int size, int bit);
 RTCD_EXTERN void (*av1_round_shift_array)(int32_t* arr, int size, int bit);
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+int av1_selfguided_restoration_sse4_1(const uint8_t* dgd8,
+                                      int width,
+                                      int height,
+                                      int dgd_stride,
+                                      int32_t* flt0,
+                                      int32_t* flt1,
+                                      int flt_stride,
+                                      int sgr_params_idx,
+                                      int bit_depth,
+                                      int highbd);
+int av1_selfguided_restoration_avx2(const uint8_t* dgd8,
+                                    int width,
+                                    int height,
+                                    int dgd_stride,
+                                    int32_t* flt0,
+                                    int32_t* flt1,
+                                    int flt_stride,
+                                    int sgr_params_idx,
+                                    int bit_depth,
+                                    int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t* dgd8,
+                                              int width,
+                                              int height,
+                                              int dgd_stride,
+                                              int32_t* flt0,
+                                              int32_t* flt1,
+                                              int flt_stride,
+                                              int sgr_params_idx,
+                                              int bit_depth,
+                                              int highbd);
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -1864,6 +1975,79 @@ void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 void av1_upsample_intra_edge_high_sse4_1(uint16_t* p, int sz, int bd);
 RTCD_EXTERN void (*av1_upsample_intra_edge_high)(uint16_t* p, int sz, int bd);
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+void av1_warp_affine_sse4_1(const int32_t* mat,
+                            const uint8_t* ref,
+                            int width,
+                            int height,
+                            int stride,
+                            uint8_t* pred,
+                            int p_col,
+                            int p_row,
+                            int p_width,
+                            int p_height,
+                            int p_stride,
+                            int subsampling_x,
+                            int subsampling_y,
+                            ConvolveParams* conv_params,
+                            int16_t alpha,
+                            int16_t beta,
+                            int16_t gamma,
+                            int16_t delta);
+void av1_warp_affine_avx2(const int32_t* mat,
+                          const uint8_t* ref,
+                          int width,
+                          int height,
+                          int stride,
+                          uint8_t* pred,
+                          int p_col,
+                          int p_row,
+                          int p_width,
+                          int p_height,
+                          int p_stride,
+                          int subsampling_x,
+                          int subsampling_y,
+                          ConvolveParams* conv_params,
+                          int16_t alpha,
+                          int16_t beta,
+                          int16_t gamma,
+                          int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t* mat,
+                                    const uint8_t* ref,
+                                    int width,
+                                    int height,
+                                    int stride,
+                                    uint8_t* pred,
+                                    int p_col,
+                                    int p_row,
+                                    int p_width,
+                                    int p_height,
+                                    int p_stride,
+                                    int subsampling_x,
+                                    int subsampling_y,
+                                    ConvolveParams* conv_params,
+                                    int16_t alpha,
+                                    int16_t beta,
+                                    int16_t gamma,
+                                    int16_t delta);
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
@@ -2687,6 +2871,11 @@ static void setup_rtcd_internal(void) {
   if (flags & HAS_SSSE3)
     aom_dist_wtd_comp_avg_upsampled_pred =
         aom_dist_wtd_comp_avg_upsampled_pred_ssse3;
+  av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_c;
+  if (flags & HAS_SSE4_1)
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_avx2;
   av1_block_error = av1_block_error_sse2;
   if (flags & HAS_AVX2)
     av1_block_error = av1_block_error_avx2;
@@ -2705,6 +2894,9 @@ static void setup_rtcd_internal(void) {
   if (flags & HAS_AVX2)
     av1_build_compound_diffwtd_mask_d16 =
         av1_build_compound_diffwtd_mask_d16_avx2;
+  av1_calc_frame_error = av1_calc_frame_error_sse2;
+  if (flags & HAS_AVX2)
+    av1_calc_frame_error = av1_calc_frame_error_avx2;
   av1_calc_indices_dim1 = av1_calc_indices_dim1_sse2;
   if (flags & HAS_AVX2)
     av1_calc_indices_dim1 = av1_calc_indices_dim1_avx2;
@@ -2903,6 +3095,11 @@ static void setup_rtcd_internal(void) {
   av1_round_shift_array = av1_round_shift_array_c;
   if (flags & HAS_SSE4_1)
     av1_round_shift_array = av1_round_shift_array_sse4_1;
+  av1_selfguided_restoration = av1_selfguided_restoration_c;
+  if (flags & HAS_SSE4_1)
+    av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_selfguided_restoration = av1_selfguided_restoration_avx2;
   av1_txb_init_levels = av1_txb_init_levels_c;
   if (flags & HAS_SSE4_1)
     av1_txb_init_levels = av1_txb_init_levels_sse4_1;
@@ -2914,6 +3111,11 @@ static void setup_rtcd_internal(void) {
   av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_c;
   if (flags & HAS_SSE4_1)
     av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_sse4_1;
+  av1_warp_affine = av1_warp_affine_c;
+  if (flags & HAS_SSE4_1)
+    av1_warp_affine = av1_warp_affine_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_warp_affine = av1_warp_affine_avx2;
   av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_sse2;
   if (flags & HAS_AVX2)
     av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_avx2;
diff --git a/chromium/third_party/libaom/source/config/linux/x64/config/aom_config.c b/chromium/third_party/libaom/source/config/linux/x64/config/aom_config.c
index 8786cc6957d..4c037ce451b 100644
--- a/chromium/third_party/libaom/source/config/linux/x64/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/linux/x64/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DAOM_TARGET_CPU=x86_64 -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DAOM_TARGET_CPU=x86_64 -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/linux/x64/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/linux/x64/config/aom_dsp_rtcd.h
index b04c0cfea23..9f1f936e240 100644
--- a/chromium/third_party/libaom/source/config/linux/x64/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/x64/config/aom_dsp_rtcd.h
@@ -360,6 +360,26 @@ void aom_dc_128_predictor_16x32_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_sse2
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_16x4_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_sse2
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_128_predictor_16x64_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_sse2
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -370,12 +390,6 @@ void aom_dc_128_predictor_16x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -427,6 +441,26 @@ RTCD_EXTERN void (*aom_dc_128_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_32x8_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_sse2
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_4x16_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_sse2
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -447,6 +481,23 @@ void aom_dc_128_predictor_4x8_sse2(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_sse2
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_128_predictor_64x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_dc_128_predictor_64x16_avx2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -491,6 +542,16 @@ void aom_dc_128_predictor_8x16_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_sse2
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_8x32_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_sse2
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -531,6 +592,26 @@ void aom_dc_left_predictor_16x32_sse2(uint8_t* dst,
                                       const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_sse2
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_16x4_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_sse2
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_dc_left_predictor_16x64_sse2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_sse2
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -541,12 +622,6 @@ void aom_dc_left_predictor_16x8_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -598,6 +673,26 @@ RTCD_EXTERN void (*aom_dc_left_predictor_32x64)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_32x8_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_sse2
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_4x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_sse2
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -618,6 +713,23 @@ void aom_dc_left_predictor_4x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_sse2
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_dc_left_predictor_64x16_sse2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+void aom_dc_left_predictor_64x16_avx2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -662,6 +774,16 @@ void aom_dc_left_predictor_8x16_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_sse2
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_8x32_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_sse2
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -702,6 +824,26 @@ void aom_dc_predictor_16x32_sse2(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_sse2
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_16x4_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_sse2
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+void aom_dc_predictor_16x64_sse2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_sse2
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -712,12 +854,6 @@ void aom_dc_predictor_16x8_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -769,6 +905,26 @@ RTCD_EXTERN void (*aom_dc_predictor_32x64)(uint8_t* dst,
                                            const uint8_t* above,
                                            const uint8_t* left);
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_32x8_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_sse2
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_4x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_sse2
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -789,6 +945,23 @@ void aom_dc_predictor_4x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_sse2
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+void aom_dc_predictor_64x16_sse2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_predictor_64x16_avx2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_predictor_64x16)(uint8_t* dst,
+                                           ptrdiff_t y_stride,
+                                           const uint8_t* above,
+                                           const uint8_t* left);
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -833,6 +1006,16 @@ void aom_dc_predictor_8x16_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_sse2
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_8x32_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_sse2
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -873,6 +1056,26 @@ void aom_dc_top_predictor_16x32_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_sse2
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_16x4_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_sse2
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_top_predictor_16x64_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_sse2
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -883,12 +1086,6 @@ void aom_dc_top_predictor_16x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -940,6 +1137,26 @@ RTCD_EXTERN void (*aom_dc_top_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_32x8_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_sse2
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_4x16_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_sse2
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -960,6 +1177,23 @@ void aom_dc_top_predictor_4x8_sse2(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_sse2
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_top_predictor_64x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_dc_top_predictor_64x16_avx2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -1004,6 +1238,16 @@ void aom_dc_top_predictor_8x16_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_sse2
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_8x32_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_sse2
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -2027,6 +2271,26 @@ void aom_h_predictor_16x32_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_sse2
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_16x4_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_sse2
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_h_predictor_16x64_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_sse2
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -2037,12 +2301,6 @@ void aom_h_predictor_16x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -2080,6 +2338,26 @@ void aom_h_predictor_32x64_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_sse2
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_32x8_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_sse2
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_4x16_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_sse2
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -2100,6 +2378,16 @@ void aom_h_predictor_4x8_sse2(uint8_t* dst,
                               const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_sse2
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_h_predictor_64x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_sse2
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -2130,6 +2418,16 @@ void aom_h_predictor_8x16_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_sse2
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_8x32_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_sse2
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -4360,6 +4658,36 @@ RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_16x4_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_16x64_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_paeth_predictor_16x64_avx2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_paeth_predictor_16x8_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -4377,12 +4705,6 @@ RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
-
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -4434,6 +4756,32 @@ RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_32x8_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_4x16_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -4460,6 +4808,23 @@ RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t* dst,
                                             const uint8_t* above,
                                             const uint8_t* left);
 
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_64x16_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_paeth_predictor_64x16_avx2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -4507,6 +4872,19 @@ RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_8x32_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -6283,6 +6661,32 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_16x4_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_16x64_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -6296,12 +6700,6 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
-
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6341,6 +6739,32 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_32x8_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_4x16_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6367,6 +6791,19 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_64x16_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6406,6 +6843,19 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_8x32_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6458,6 +6908,32 @@ RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_16x4_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_16x64_ssse3(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -6471,12 +6947,6 @@ RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6516,6 +6986,32 @@ RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_32x8_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_4x16_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -6542,6 +7038,19 @@ RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_64x16_ssse3(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6581,6 +7090,19 @@ RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_8x32_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -6633,6 +7155,32 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_16x4_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_16x64_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -6646,12 +7194,6 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
-
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6691,6 +7233,32 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_32x8_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_4x16_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6717,6 +7285,19 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_64x16_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6756,6 +7337,19 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_8x32_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -8071,6 +8665,26 @@ void aom_v_predictor_16x32_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_sse2
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_16x4_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_sse2
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_v_predictor_16x64_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_sse2
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -8081,12 +8695,6 @@ void aom_v_predictor_16x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -8138,6 +8746,26 @@ RTCD_EXTERN void (*aom_v_predictor_32x64)(uint8_t* dst,
                                           const uint8_t* above,
                                           const uint8_t* left);
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_32x8_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_sse2
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_4x16_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_sse2
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -8158,6 +8786,23 @@ void aom_v_predictor_4x8_sse2(uint8_t* dst,
                               const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_sse2
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_v_predictor_64x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_v_predictor_64x16_avx2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+RTCD_EXTERN void (*aom_v_predictor_64x16)(uint8_t* dst,
+                                          ptrdiff_t y_stride,
+                                          const uint8_t* above,
+                                          const uint8_t* left);
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -8202,6 +8847,16 @@ void aom_v_predictor_8x16_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_sse2
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_8x32_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_sse2
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -8615,6 +9270,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_avx2;
+  aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_avx2;
   aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_avx2;
@@ -8630,6 +9288,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_avx2;
+  aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_avx2;
   aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_avx2;
@@ -8645,6 +9306,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_predictor_32x64 = aom_dc_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_predictor_32x64 = aom_dc_predictor_32x64_avx2;
+  aom_dc_predictor_64x16 = aom_dc_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_predictor_64x16 = aom_dc_predictor_64x16_avx2;
   aom_dc_predictor_64x32 = aom_dc_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_predictor_64x32 = aom_dc_predictor_64x32_avx2;
@@ -8660,6 +9324,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_avx2;
+  aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_avx2;
   aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_avx2;
@@ -9055,6 +9722,14 @@ static void setup_rtcd_internal(void) {
     aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_ssse3;
   if (flags & HAS_AVX2)
     aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_avx2;
+  aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_ssse3;
+  aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_ssse3;
+  if (flags & HAS_AVX2)
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_avx2;
   aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_ssse3;
@@ -9075,12 +9750,23 @@ static void setup_rtcd_internal(void) {
     aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_ssse3;
   if (flags & HAS_AVX2)
     aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_avx2;
+  aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_ssse3;
+  aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_ssse3;
   aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_ssse3;
   aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_ssse3;
+  aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_ssse3;
+  if (flags & HAS_AVX2)
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_avx2;
   aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_ssse3;
@@ -9094,6 +9780,9 @@ static void setup_rtcd_internal(void) {
   aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_ssse3;
+  aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_ssse3;
   aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3;
@@ -9251,6 +9940,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_ssse3;
+  aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_ssse3;
+  aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_ssse3;
   aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_ssse3;
@@ -9263,12 +9958,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_ssse3;
+  aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_ssse3;
+  aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_ssse3;
   aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_ssse3;
   aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_ssse3;
+  aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_ssse3;
   aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_ssse3;
@@ -9278,6 +9982,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_ssse3;
+  aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_ssse3;
   aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_ssse3;
@@ -9290,6 +9997,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_ssse3;
+  aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_ssse3;
+  aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_ssse3;
   aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_ssse3;
@@ -9302,12 +10015,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_ssse3;
+  aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_ssse3;
+  aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_ssse3;
   aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_ssse3;
   aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_ssse3;
+  aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_ssse3;
   aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_ssse3;
@@ -9317,6 +10039,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_ssse3;
+  aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_ssse3;
   aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_ssse3;
@@ -9329,6 +10054,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_ssse3;
+  aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_ssse3;
+  aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_ssse3;
   aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_ssse3;
@@ -9341,12 +10072,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_ssse3;
+  aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_ssse3;
+  aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_ssse3;
   aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_ssse3;
   aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_ssse3;
+  aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_ssse3;
   aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_ssse3;
@@ -9356,6 +10096,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_ssse3;
+  aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_ssse3;
   aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_ssse3;
@@ -9519,6 +10262,9 @@ static void setup_rtcd_internal(void) {
   aom_v_predictor_32x64 = aom_v_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_v_predictor_32x64 = aom_v_predictor_32x64_avx2;
+  aom_v_predictor_64x16 = aom_v_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_v_predictor_64x16 = aom_v_predictor_64x16_avx2;
   aom_v_predictor_64x32 = aom_v_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_v_predictor_64x32 = aom_v_predictor_64x32_avx2;
diff --git a/chromium/third_party/libaom/source/config/linux/x64/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/linux/x64/config/av1_rtcd.h
index 8d516d90bde..6d8a64d1270 100644
--- a/chromium/third_party/libaom/source/config/linux/x64/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/linux/x64/config/av1_rtcd.h
@@ -242,6 +242,51 @@ void aom_upsampled_pred_sse2(MACROBLOCKD* xd,
                              int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_sse2
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+void av1_apply_selfguided_restoration_sse4_1(const uint8_t* dat,
+                                             int width,
+                                             int height,
+                                             int stride,
+                                             int eps,
+                                             const int* xqd,
+                                             uint8_t* dst,
+                                             int dst_stride,
+                                             int32_t* tmpbuf,
+                                             int bit_depth,
+                                             int highbd);
+void av1_apply_selfguided_restoration_avx2(const uint8_t* dat,
+                                           int width,
+                                           int height,
+                                           int stride,
+                                           int eps,
+                                           const int* xqd,
+                                           uint8_t* dst,
+                                           int dst_stride,
+                                           int32_t* tmpbuf,
+                                           int bit_depth,
+                                           int highbd);
+RTCD_EXTERN void (*av1_apply_selfguided_restoration)(const uint8_t* dat,
+                                                     int width,
+                                                     int height,
+                                                     int stride,
+                                                     int eps,
+                                                     const int* xqd,
+                                                     uint8_t* dst,
+                                                     int dst_stride,
+                                                     int32_t* tmpbuf,
+                                                     int bit_depth,
+                                                     int highbd);
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -347,6 +392,31 @@ RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(
     ConvolveParams* conv_params,
     int bd);
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+int64_t av1_calc_frame_error_sse2(const uint8_t* const ref,
+                                  int stride,
+                                  const uint8_t* const dst,
+                                  int p_width,
+                                  int p_height,
+                                  int p_stride);
+int64_t av1_calc_frame_error_avx2(const uint8_t* const ref,
+                                  int stride,
+                                  const uint8_t* const dst,
+                                  int p_width,
+                                  int p_height,
+                                  int p_stride);
+RTCD_EXTERN int64_t (*av1_calc_frame_error)(const uint8_t* const ref,
+                                            int stride,
+                                            const uint8_t* const dst,
+                                            int p_width,
+                                            int p_height,
+                                            int p_stride);
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -1839,6 +1909,47 @@ void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 void av1_round_shift_array_sse4_1(int32_t* arr, int size, int bit);
 RTCD_EXTERN void (*av1_round_shift_array)(int32_t* arr, int size, int bit);
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+int av1_selfguided_restoration_sse4_1(const uint8_t* dgd8,
+                                      int width,
+                                      int height,
+                                      int dgd_stride,
+                                      int32_t* flt0,
+                                      int32_t* flt1,
+                                      int flt_stride,
+                                      int sgr_params_idx,
+                                      int bit_depth,
+                                      int highbd);
+int av1_selfguided_restoration_avx2(const uint8_t* dgd8,
+                                    int width,
+                                    int height,
+                                    int dgd_stride,
+                                    int32_t* flt0,
+                                    int32_t* flt1,
+                                    int flt_stride,
+                                    int sgr_params_idx,
+                                    int bit_depth,
+                                    int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t* dgd8,
+                                              int width,
+                                              int height,
+                                              int dgd_stride,
+                                              int32_t* flt0,
+                                              int32_t* flt1,
+                                              int flt_stride,
+                                              int sgr_params_idx,
+                                              int bit_depth,
+                                              int highbd);
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -1864,6 +1975,79 @@ void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 void av1_upsample_intra_edge_high_sse4_1(uint16_t* p, int sz, int bd);
 RTCD_EXTERN void (*av1_upsample_intra_edge_high)(uint16_t* p, int sz, int bd);
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+void av1_warp_affine_sse4_1(const int32_t* mat,
+                            const uint8_t* ref,
+                            int width,
+                            int height,
+                            int stride,
+                            uint8_t* pred,
+                            int p_col,
+                            int p_row,
+                            int p_width,
+                            int p_height,
+                            int p_stride,
+                            int subsampling_x,
+                            int subsampling_y,
+                            ConvolveParams* conv_params,
+                            int16_t alpha,
+                            int16_t beta,
+                            int16_t gamma,
+                            int16_t delta);
+void av1_warp_affine_avx2(const int32_t* mat,
+                          const uint8_t* ref,
+                          int width,
+                          int height,
+                          int stride,
+                          uint8_t* pred,
+                          int p_col,
+                          int p_row,
+                          int p_width,
+                          int p_height,
+                          int p_stride,
+                          int subsampling_x,
+                          int subsampling_y,
+                          ConvolveParams* conv_params,
+                          int16_t alpha,
+                          int16_t beta,
+                          int16_t gamma,
+                          int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t* mat,
+                                    const uint8_t* ref,
+                                    int width,
+                                    int height,
+                                    int stride,
+                                    uint8_t* pred,
+                                    int p_col,
+                                    int p_row,
+                                    int p_width,
+                                    int p_height,
+                                    int p_stride,
+                                    int subsampling_x,
+                                    int subsampling_y,
+                                    ConvolveParams* conv_params,
+                                    int16_t alpha,
+                                    int16_t beta,
+                                    int16_t gamma,
+                                    int16_t delta);
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
@@ -2687,6 +2871,11 @@ static void setup_rtcd_internal(void) {
   if (flags & HAS_SSSE3)
     aom_dist_wtd_comp_avg_upsampled_pred =
         aom_dist_wtd_comp_avg_upsampled_pred_ssse3;
+  av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_c;
+  if (flags & HAS_SSE4_1)
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_avx2;
   av1_block_error = av1_block_error_sse2;
   if (flags & HAS_AVX2)
     av1_block_error = av1_block_error_avx2;
@@ -2705,6 +2894,9 @@ static void setup_rtcd_internal(void) {
   if (flags & HAS_AVX2)
     av1_build_compound_diffwtd_mask_d16 =
         av1_build_compound_diffwtd_mask_d16_avx2;
+  av1_calc_frame_error = av1_calc_frame_error_sse2;
+  if (flags & HAS_AVX2)
+    av1_calc_frame_error = av1_calc_frame_error_avx2;
   av1_calc_indices_dim1 = av1_calc_indices_dim1_sse2;
   if (flags & HAS_AVX2)
     av1_calc_indices_dim1 = av1_calc_indices_dim1_avx2;
@@ -2903,6 +3095,11 @@ static void setup_rtcd_internal(void) {
   av1_round_shift_array = av1_round_shift_array_c;
   if (flags & HAS_SSE4_1)
     av1_round_shift_array = av1_round_shift_array_sse4_1;
+  av1_selfguided_restoration = av1_selfguided_restoration_c;
+  if (flags & HAS_SSE4_1)
+    av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_selfguided_restoration = av1_selfguided_restoration_avx2;
   av1_txb_init_levels = av1_txb_init_levels_c;
   if (flags & HAS_SSE4_1)
     av1_txb_init_levels = av1_txb_init_levels_sse4_1;
@@ -2914,6 +3111,11 @@ static void setup_rtcd_internal(void) {
   av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_c;
   if (flags & HAS_SSE4_1)
     av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_sse4_1;
+  av1_warp_affine = av1_warp_affine_c;
+  if (flags & HAS_SSE4_1)
+    av1_warp_affine = av1_warp_affine_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_warp_affine = av1_warp_affine_avx2;
   av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_sse2;
   if (flags & HAS_AVX2)
     av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_avx2;
diff --git a/chromium/third_party/libaom/source/config/win/arm64/config/aom_config.c b/chromium/third_party/libaom/source/config/win/arm64/config/aom_config.c
index 7c8f15b9d65..ff9196735fc 100644
--- a/chromium/third_party/libaom/source/config/win/arm64/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/win/arm64/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/arm64-linux-gcc.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/arm64-linux-gcc.cmake\" -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/win/arm64/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/win/arm64/config/aom_dsp_rtcd.h
index 155c11857c0..49fb79d7f18 100644
--- a/chromium/third_party/libaom/source/config/win/arm64/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/win/arm64/config/aom_dsp_rtcd.h
@@ -34,7 +34,12 @@ void aom_avg_8x8_quad_c(const uint8_t* s,
                         int x16_idx,
                         int y16_idx,
                         int* avg);
-#define aom_avg_8x8_quad aom_avg_8x8_quad_c
+void aom_avg_8x8_quad_neon(const uint8_t* s,
+                           int p,
+                           int x16_idx,
+                           int y16_idx,
+                           int* avg);
+#define aom_avg_8x8_quad aom_avg_8x8_quad_neon
 
 void aom_blend_a64_hmask_c(uint8_t* dst,
                            uint32_t dst_stride,
@@ -176,18 +181,24 @@ void aom_dc_128_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_c
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_c
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -210,6 +221,18 @@ void aom_dc_128_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_128_predictor_32x64 aom_dc_128_predictor_32x64_c
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_c
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_c
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -226,6 +249,12 @@ void aom_dc_128_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_128_predictor_64x16 aom_dc_128_predictor_64x16_c
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -244,6 +273,12 @@ void aom_dc_128_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_c
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -276,18 +311,24 @@ void aom_dc_left_predictor_16x32_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_c
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_c
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -310,6 +351,18 @@ void aom_dc_left_predictor_32x64_c(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_left_predictor_32x64 aom_dc_left_predictor_32x64_c
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_c
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_c
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -326,6 +379,12 @@ void aom_dc_left_predictor_4x8_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_dc_left_predictor_64x16 aom_dc_left_predictor_64x16_c
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -344,6 +403,12 @@ void aom_dc_left_predictor_8x16_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_c
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -376,18 +441,24 @@ void aom_dc_predictor_16x32_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_c
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_c
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
                              const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -410,6 +481,18 @@ void aom_dc_predictor_32x64_c(uint8_t* dst,
                               const uint8_t* left);
 #define aom_dc_predictor_32x64 aom_dc_predictor_32x64_c
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_c
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_c
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -426,6 +509,12 @@ void aom_dc_predictor_4x8_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+#define aom_dc_predictor_64x16 aom_dc_predictor_64x16_c
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -444,6 +533,12 @@ void aom_dc_predictor_8x16_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_c
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -476,18 +571,24 @@ void aom_dc_top_predictor_16x32_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_c
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_c
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -510,6 +611,18 @@ void aom_dc_top_predictor_32x64_c(uint8_t* dst,
                                   const uint8_t* left);
 #define aom_dc_top_predictor_32x64 aom_dc_top_predictor_32x64_c
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_c
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_c
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -526,6 +639,12 @@ void aom_dc_top_predictor_4x8_c(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_dc_top_predictor_64x16 aom_dc_top_predictor_64x16_c
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -544,6 +663,12 @@ void aom_dc_top_predictor_8x16_c(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_c
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -1014,18 +1139,24 @@ void aom_h_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_c
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_c
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_c
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_c
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1048,6 +1179,18 @@ void aom_h_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_c
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_c
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_c
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1064,6 +1207,12 @@ void aom_h_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_c
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_c
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -1082,6 +1231,12 @@ void aom_h_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_c
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_c
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -1127,7 +1282,10 @@ void aom_hadamard_8x8_neon(const int16_t* src_diff,
 void aom_hadamard_8x8_dual_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
                              int16_t* coeff);
-#define aom_hadamard_8x8_dual aom_hadamard_8x8_dual_c
+void aom_hadamard_8x8_dual_neon(const int16_t* src_diff,
+                                ptrdiff_t src_stride,
+                                int16_t* coeff);
+#define aom_hadamard_8x8_dual aom_hadamard_8x8_dual_neon
 
 void aom_hadamard_lp_16x16_c(const int16_t* src_diff,
                              ptrdiff_t src_stride,
@@ -1237,7 +1395,12 @@ void aom_lpf_horizontal_14_quad_c(uint8_t* s,
                                   const uint8_t* blimit0,
                                   const uint8_t* limit0,
                                   const uint8_t* thresh0);
-#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_c
+void aom_lpf_horizontal_14_quad_neon(uint8_t* s,
+                                     int pitch,
+                                     const uint8_t* blimit0,
+                                     const uint8_t* limit0,
+                                     const uint8_t* thresh0);
+#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_neon
 
 void aom_lpf_horizontal_4_c(uint8_t* s,
                             int pitch,
@@ -1274,7 +1437,12 @@ void aom_lpf_horizontal_4_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_c
+void aom_lpf_horizontal_4_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_neon
 
 void aom_lpf_horizontal_6_c(uint8_t* s,
                             int pitch,
@@ -1311,7 +1479,12 @@ void aom_lpf_horizontal_6_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_c
+void aom_lpf_horizontal_6_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_neon
 
 void aom_lpf_horizontal_8_c(uint8_t* s,
                             int pitch,
@@ -1348,7 +1521,12 @@ void aom_lpf_horizontal_8_quad_c(uint8_t* s,
                                  const uint8_t* blimit0,
                                  const uint8_t* limit0,
                                  const uint8_t* thresh0);
-#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_c
+void aom_lpf_horizontal_8_quad_neon(uint8_t* s,
+                                    int pitch,
+                                    const uint8_t* blimit0,
+                                    const uint8_t* limit0,
+                                    const uint8_t* thresh0);
+#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_neon
 
 void aom_lpf_vertical_14_c(uint8_t* s,
                            int pitch,
@@ -1385,7 +1563,12 @@ void aom_lpf_vertical_14_quad_c(uint8_t* s,
                                 const uint8_t* blimit0,
                                 const uint8_t* limit0,
                                 const uint8_t* thresh0);
-#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_c
+void aom_lpf_vertical_14_quad_neon(uint8_t* s,
+                                   int pitch,
+                                   const uint8_t* blimit0,
+                                   const uint8_t* limit0,
+                                   const uint8_t* thresh0);
+#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_neon
 
 void aom_lpf_vertical_4_c(uint8_t* s,
                           int pitch,
@@ -1422,7 +1605,12 @@ void aom_lpf_vertical_4_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_c
+void aom_lpf_vertical_4_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_neon
 
 void aom_lpf_vertical_6_c(uint8_t* s,
                           int pitch,
@@ -1459,7 +1647,12 @@ void aom_lpf_vertical_6_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_c
+void aom_lpf_vertical_6_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_neon
 
 void aom_lpf_vertical_8_c(uint8_t* s,
                           int pitch,
@@ -1496,7 +1689,12 @@ void aom_lpf_vertical_8_quad_c(uint8_t* s,
                                const uint8_t* blimit0,
                                const uint8_t* limit0,
                                const uint8_t* thresh0);
-#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_c
+void aom_lpf_vertical_8_quad_neon(uint8_t* s,
+                                  int pitch,
+                                  const uint8_t* blimit0,
+                                  const uint8_t* limit0,
+                                  const uint8_t* thresh0);
+#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_neon
 
 unsigned int aom_masked_sad128x128_c(const uint8_t* src,
                                      int src_stride,
@@ -2098,85 +2296,191 @@ void aom_paeth_predictor_16x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_c
+void aom_paeth_predictor_16x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_neon
 
 void aom_paeth_predictor_16x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
+void aom_paeth_predictor_16x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_neon
 
-void aom_paeth_predictor_16x8_c(uint8_t* dst,
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
+void aom_paeth_predictor_16x4_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_16x4 aom_paeth_predictor_16x4_neon
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_16x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_16x64 aom_paeth_predictor_16x64_neon
+
+void aom_paeth_predictor_16x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_16x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_neon
 
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c
+void aom_paeth_predictor_32x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_neon
 
 void aom_paeth_predictor_32x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_c
+void aom_paeth_predictor_32x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_neon
 
 void aom_paeth_predictor_32x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_c
+void aom_paeth_predictor_32x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_neon
+
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_32x8_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_32x8 aom_paeth_predictor_32x8_neon
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_4x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_4x16 aom_paeth_predictor_4x16_neon
 
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_c
+void aom_paeth_predictor_4x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_neon
 
 void aom_paeth_predictor_4x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
+void aom_paeth_predictor_4x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_neon
+
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_64x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x16 aom_paeth_predictor_64x16_neon
 
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_c
+void aom_paeth_predictor_64x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_neon
 
 void aom_paeth_predictor_64x64_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
                                  const uint8_t* left);
-#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_c
+void aom_paeth_predictor_64x64_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_neon
 
 void aom_paeth_predictor_8x16_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
                                 const uint8_t* left);
-#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
+void aom_paeth_predictor_8x16_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_neon
+
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_8x32_neon(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+#define aom_paeth_predictor_8x32 aom_paeth_predictor_8x32_neon
 
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_c
+void aom_paeth_predictor_8x4_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_neon
 
 void aom_paeth_predictor_8x8_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
                                const uint8_t* left);
-#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_c
+void aom_paeth_predictor_8x8_neon(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_neon
 
 void aom_pixel_scale_c(const int16_t* src_diff,
                        ptrdiff_t src_stride,
@@ -3219,85 +3523,191 @@ void aom_smooth_h_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_c
+void aom_smooth_h_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_neon
 
 void aom_smooth_h_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c
+void aom_smooth_h_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_neon
 
-void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c
+void aom_smooth_h_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_16x4 aom_smooth_h_predictor_16x4_neon
 
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_16x64 aom_smooth_h_predictor_16x64_neon
+
+void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_neon
 
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_c
+void aom_smooth_h_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_neon
 
 void aom_smooth_h_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_c
+void aom_smooth_h_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_neon
 
 void aom_smooth_h_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_c
+void aom_smooth_h_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_neon
+
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_32x8 aom_smooth_h_predictor_32x8_neon
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_4x16 aom_smooth_h_predictor_4x16_neon
 
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_c
+void aom_smooth_h_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_neon
 
 void aom_smooth_h_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c
+void aom_smooth_h_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_neon
+
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x16 aom_smooth_h_predictor_64x16_neon
 
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_c
+void aom_smooth_h_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_neon
 
 void aom_smooth_h_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_c
+void aom_smooth_h_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_neon
 
 void aom_smooth_h_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c
+void aom_smooth_h_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_neon
+
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_h_predictor_8x32 aom_smooth_h_predictor_8x32_neon
 
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_c
+void aom_smooth_h_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_neon
 
 void aom_smooth_h_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_c
+void aom_smooth_h_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_neon
 
 void aom_smooth_predictor_16x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
@@ -3319,6 +3729,26 @@ void aom_smooth_predictor_16x32_neon(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_neon
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_16x4_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_16x4 aom_smooth_predictor_16x4_neon
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_16x64_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_predictor_16x64 aom_smooth_predictor_16x64_neon
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -3329,12 +3759,6 @@ void aom_smooth_predictor_16x8_neon(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_neon
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3365,6 +3789,26 @@ void aom_smooth_predictor_32x64_neon(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_smooth_predictor_32x64 aom_smooth_predictor_32x64_neon
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_32x8_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_32x8 aom_smooth_predictor_32x8_neon
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_4x16_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_4x16 aom_smooth_predictor_4x16_neon
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3385,6 +3829,16 @@ void aom_smooth_predictor_4x8_neon(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_neon
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_64x16_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_predictor_64x16 aom_smooth_predictor_64x16_neon
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -3415,6 +3869,16 @@ void aom_smooth_predictor_8x16_neon(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_neon
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_8x32_neon(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_smooth_predictor_8x32 aom_smooth_predictor_8x32_neon
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -3439,85 +3903,191 @@ void aom_smooth_v_predictor_16x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_c
+void aom_smooth_v_predictor_16x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_neon
 
 void aom_smooth_v_predictor_16x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c
+void aom_smooth_v_predictor_16x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_neon
 
-void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c
+void aom_smooth_v_predictor_16x4_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_16x4 aom_smooth_v_predictor_16x4_neon
 
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_16x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_16x64 aom_smooth_v_predictor_16x64_neon
+
+void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_16x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_neon
 
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_c
+void aom_smooth_v_predictor_32x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_neon
 
 void aom_smooth_v_predictor_32x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_c
+void aom_smooth_v_predictor_32x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_neon
 
 void aom_smooth_v_predictor_32x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_c
+void aom_smooth_v_predictor_32x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_neon
+
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_32x8_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_32x8 aom_smooth_v_predictor_32x8_neon
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_4x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_4x16 aom_smooth_v_predictor_4x16_neon
 
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_c
+void aom_smooth_v_predictor_4x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_neon
 
 void aom_smooth_v_predictor_4x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c
+void aom_smooth_v_predictor_4x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_neon
+
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_64x16_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x16 aom_smooth_v_predictor_64x16_neon
 
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_c
+void aom_smooth_v_predictor_64x32_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_neon
 
 void aom_smooth_v_predictor_64x64_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
                                     const uint8_t* left);
-#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_c
+void aom_smooth_v_predictor_64x64_neon(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_neon
 
 void aom_smooth_v_predictor_8x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
                                    const uint8_t* left);
-#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c
+void aom_smooth_v_predictor_8x16_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_neon
+
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_8x32_neon(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_smooth_v_predictor_8x32 aom_smooth_v_predictor_8x32_neon
 
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_c
+void aom_smooth_v_predictor_8x4_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_neon
 
 void aom_smooth_v_predictor_8x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
                                   const uint8_t* left);
-#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_c
+void aom_smooth_v_predictor_8x8_neon(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_neon
 
 int64_t aom_sse_c(const uint8_t* a,
                   int a_stride,
@@ -4014,18 +4584,24 @@ void aom_v_predictor_16x32_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_c
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_c
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_c
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
                             const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_c
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4048,6 +4624,18 @@ void aom_v_predictor_32x64_c(uint8_t* dst,
                              const uint8_t* left);
 #define aom_v_predictor_32x64 aom_v_predictor_32x64_c
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_c
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_c
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -4064,6 +4652,12 @@ void aom_v_predictor_4x8_c(uint8_t* dst,
                            const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_c
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+#define aom_v_predictor_64x16 aom_v_predictor_64x16_c
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -4082,6 +4676,12 @@ void aom_v_predictor_8x16_c(uint8_t* dst,
                             const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_c
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_c
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
diff --git a/chromium/third_party/libaom/source/config/win/arm64/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/win/arm64/config/av1_rtcd.h
index 23e831cc591..7c2e8affa18 100644
--- a/chromium/third_party/libaom/source/config/win/arm64/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/win/arm64/config/av1_rtcd.h
@@ -183,6 +183,30 @@ void aom_upsampled_pred_c(MACROBLOCKD* xd,
                           int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_c
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+void av1_apply_selfguided_restoration_neon(const uint8_t* dat,
+                                           int width,
+                                           int height,
+                                           int stride,
+                                           int eps,
+                                           const int* xqd,
+                                           uint8_t* dst,
+                                           int dst_stride,
+                                           int32_t* tmpbuf,
+                                           int bit_depth,
+                                           int highbd);
+#define av1_apply_selfguided_restoration av1_apply_selfguided_restoration_neon
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -234,6 +258,14 @@ void av1_build_compound_diffwtd_mask_d16_neon(uint8_t* mask,
 #define av1_build_compound_diffwtd_mask_d16 \
   av1_build_compound_diffwtd_mask_d16_neon
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+#define av1_calc_frame_error av1_calc_frame_error_c
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -1314,6 +1346,28 @@ void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 void av1_round_shift_array_neon(int32_t* arr, int size, int bit);
 #define av1_round_shift_array av1_round_shift_array_neon
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+int av1_selfguided_restoration_neon(const uint8_t* dgd8,
+                                    int width,
+                                    int height,
+                                    int dgd_stride,
+                                    int32_t* flt0,
+                                    int32_t* flt1,
+                                    int flt_stride,
+                                    int sgr_params_idx,
+                                    int bit_depth,
+                                    int highbd);
+#define av1_selfguided_restoration av1_selfguided_restoration_neon
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -1330,6 +1384,44 @@ void av1_upsample_intra_edge_c(uint8_t* p, int sz);
 void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 #define av1_upsample_intra_edge_high av1_upsample_intra_edge_high_c
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+void av1_warp_affine_neon(const int32_t* mat,
+                          const uint8_t* ref,
+                          int width,
+                          int height,
+                          int stride,
+                          uint8_t* pred,
+                          int p_col,
+                          int p_row,
+                          int p_width,
+                          int p_height,
+                          int p_stride,
+                          int subsampling_x,
+                          int subsampling_y,
+                          ConvolveParams* conv_params,
+                          int16_t alpha,
+                          int16_t beta,
+                          int16_t gamma,
+                          int16_t delta);
+#define av1_warp_affine av1_warp_affine_neon
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
@@ -1610,7 +1702,15 @@ void cdef_find_dir_dual_c(const uint16_t* img1,
                           int coeff_shift,
                           int* out1,
                           int* out2);
-#define cdef_find_dir_dual cdef_find_dir_dual_c
+void cdef_find_dir_dual_neon(const uint16_t* img1,
+                             const uint16_t* img2,
+                             int stride,
+                             int32_t* var1,
+                             int32_t* var2,
+                             int coeff_shift,
+                             int* out1,
+                             int* out2);
+#define cdef_find_dir_dual cdef_find_dir_dual_neon
 
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_neon(TX_SIZE tx_size);
diff --git a/chromium/third_party/libaom/source/config/win/ia32/config/aom_config.c b/chromium/third_party/libaom/source/config/win/ia32/config/aom_config.c
index aaa8142b8b8..dadb41a0d60 100644
--- a/chromium/third_party/libaom/source/config/win/ia32/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/win/ia32/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/x86-linux.cmake\" -DAOM_RTCD_FLAGS=--require-mmx;--require-sse;--require-sse2 -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_PIC=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=\"../source/libaom/build/cmake/toolchains/x86-linux.cmake\" -DAOM_RTCD_FLAGS=--require-mmx;--require-sse;--require-sse2 -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_PIC=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/win/ia32/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/win/ia32/config/aom_dsp_rtcd.h
index aa189c45f4a..a76e597b134 100644
--- a/chromium/third_party/libaom/source/config/win/ia32/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/win/ia32/config/aom_dsp_rtcd.h
@@ -360,6 +360,26 @@ void aom_dc_128_predictor_16x32_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_sse2
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_16x4_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_sse2
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_128_predictor_16x64_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_sse2
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -370,12 +390,6 @@ void aom_dc_128_predictor_16x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -427,6 +441,26 @@ RTCD_EXTERN void (*aom_dc_128_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_32x8_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_sse2
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_4x16_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_sse2
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -447,6 +481,23 @@ void aom_dc_128_predictor_4x8_sse2(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_sse2
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_128_predictor_64x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_dc_128_predictor_64x16_avx2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -491,6 +542,16 @@ void aom_dc_128_predictor_8x16_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_sse2
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_8x32_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_sse2
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -531,6 +592,26 @@ void aom_dc_left_predictor_16x32_sse2(uint8_t* dst,
                                       const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_sse2
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_16x4_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_sse2
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_dc_left_predictor_16x64_sse2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_sse2
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -541,12 +622,6 @@ void aom_dc_left_predictor_16x8_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -598,6 +673,26 @@ RTCD_EXTERN void (*aom_dc_left_predictor_32x64)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_32x8_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_sse2
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_4x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_sse2
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -618,6 +713,23 @@ void aom_dc_left_predictor_4x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_sse2
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_dc_left_predictor_64x16_sse2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+void aom_dc_left_predictor_64x16_avx2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -662,6 +774,16 @@ void aom_dc_left_predictor_8x16_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_sse2
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_8x32_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_sse2
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -702,6 +824,26 @@ void aom_dc_predictor_16x32_sse2(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_sse2
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_16x4_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_sse2
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+void aom_dc_predictor_16x64_sse2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_sse2
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -712,12 +854,6 @@ void aom_dc_predictor_16x8_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -769,6 +905,26 @@ RTCD_EXTERN void (*aom_dc_predictor_32x64)(uint8_t* dst,
                                            const uint8_t* above,
                                            const uint8_t* left);
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_32x8_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_sse2
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_4x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_sse2
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -789,6 +945,23 @@ void aom_dc_predictor_4x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_sse2
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+void aom_dc_predictor_64x16_sse2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_predictor_64x16_avx2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_predictor_64x16)(uint8_t* dst,
+                                           ptrdiff_t y_stride,
+                                           const uint8_t* above,
+                                           const uint8_t* left);
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -833,6 +1006,16 @@ void aom_dc_predictor_8x16_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_sse2
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_8x32_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_sse2
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -873,6 +1056,26 @@ void aom_dc_top_predictor_16x32_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_sse2
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_16x4_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_sse2
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_top_predictor_16x64_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_sse2
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -883,12 +1086,6 @@ void aom_dc_top_predictor_16x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -940,6 +1137,26 @@ RTCD_EXTERN void (*aom_dc_top_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_32x8_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_sse2
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_4x16_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_sse2
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -960,6 +1177,23 @@ void aom_dc_top_predictor_4x8_sse2(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_sse2
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_top_predictor_64x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_dc_top_predictor_64x16_avx2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -1004,6 +1238,16 @@ void aom_dc_top_predictor_8x16_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_sse2
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_8x32_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_sse2
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -2024,6 +2268,26 @@ void aom_h_predictor_16x32_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_sse2
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_16x4_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_sse2
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_h_predictor_16x64_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_sse2
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -2034,12 +2298,6 @@ void aom_h_predictor_16x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -2077,6 +2335,26 @@ void aom_h_predictor_32x64_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_sse2
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_32x8_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_sse2
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_4x16_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_sse2
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -2097,6 +2375,16 @@ void aom_h_predictor_4x8_sse2(uint8_t* dst,
                               const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_sse2
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_h_predictor_64x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_sse2
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -2127,6 +2415,16 @@ void aom_h_predictor_8x16_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_sse2
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_8x32_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_sse2
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -4357,6 +4655,36 @@ RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_16x4_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_16x64_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_paeth_predictor_16x64_avx2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_paeth_predictor_16x8_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -4374,12 +4702,6 @@ RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
-
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -4431,6 +4753,32 @@ RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_32x8_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_4x16_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -4457,6 +4805,23 @@ RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t* dst,
                                             const uint8_t* above,
                                             const uint8_t* left);
 
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_64x16_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_paeth_predictor_64x16_avx2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -4504,6 +4869,19 @@ RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_8x32_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -6256,6 +6634,32 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_16x4_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_16x64_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -6269,12 +6673,6 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
-
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6314,6 +6712,32 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_32x8_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_4x16_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6340,6 +6764,19 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_64x16_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6379,6 +6816,19 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_8x32_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6431,6 +6881,32 @@ RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_16x4_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_16x64_ssse3(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -6444,12 +6920,6 @@ RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6489,6 +6959,32 @@ RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_32x8_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_4x16_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -6515,6 +7011,19 @@ RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_64x16_ssse3(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6554,6 +7063,19 @@ RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_8x32_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -6606,6 +7128,32 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_16x4_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_16x64_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -6619,12 +7167,6 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
-
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6664,6 +7206,32 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_32x8_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_4x16_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6690,6 +7258,19 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_64x16_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6729,6 +7310,19 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_8x32_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -8035,6 +8629,26 @@ void aom_v_predictor_16x32_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_sse2
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_16x4_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_sse2
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_v_predictor_16x64_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_sse2
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -8045,12 +8659,6 @@ void aom_v_predictor_16x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -8102,6 +8710,26 @@ RTCD_EXTERN void (*aom_v_predictor_32x64)(uint8_t* dst,
                                           const uint8_t* above,
                                           const uint8_t* left);
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_32x8_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_sse2
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_4x16_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_sse2
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -8122,6 +8750,23 @@ void aom_v_predictor_4x8_sse2(uint8_t* dst,
                               const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_sse2
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_v_predictor_64x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_v_predictor_64x16_avx2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+RTCD_EXTERN void (*aom_v_predictor_64x16)(uint8_t* dst,
+                                          ptrdiff_t y_stride,
+                                          const uint8_t* above,
+                                          const uint8_t* left);
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -8166,6 +8811,16 @@ void aom_v_predictor_8x16_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_sse2
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_8x32_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_sse2
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -8579,6 +9234,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_avx2;
+  aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_avx2;
   aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_avx2;
@@ -8594,6 +9252,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_avx2;
+  aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_avx2;
   aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_avx2;
@@ -8609,6 +9270,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_predictor_32x64 = aom_dc_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_predictor_32x64 = aom_dc_predictor_32x64_avx2;
+  aom_dc_predictor_64x16 = aom_dc_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_predictor_64x16 = aom_dc_predictor_64x16_avx2;
   aom_dc_predictor_64x32 = aom_dc_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_predictor_64x32 = aom_dc_predictor_64x32_avx2;
@@ -8624,6 +9288,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_avx2;
+  aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_avx2;
   aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_avx2;
@@ -9016,6 +9683,14 @@ static void setup_rtcd_internal(void) {
     aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_ssse3;
   if (flags & HAS_AVX2)
     aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_avx2;
+  aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_ssse3;
+  aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_ssse3;
+  if (flags & HAS_AVX2)
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_avx2;
   aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_ssse3;
@@ -9036,12 +9711,23 @@ static void setup_rtcd_internal(void) {
     aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_ssse3;
   if (flags & HAS_AVX2)
     aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_avx2;
+  aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_ssse3;
+  aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_ssse3;
   aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_ssse3;
   aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_ssse3;
+  aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_ssse3;
+  if (flags & HAS_AVX2)
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_avx2;
   aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_ssse3;
@@ -9055,6 +9741,9 @@ static void setup_rtcd_internal(void) {
   aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_ssse3;
+  aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_ssse3;
   aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3;
@@ -9208,6 +9897,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_ssse3;
+  aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_ssse3;
+  aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_ssse3;
   aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_ssse3;
@@ -9220,12 +9915,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_ssse3;
+  aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_ssse3;
+  aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_ssse3;
   aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_ssse3;
   aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_ssse3;
+  aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_ssse3;
   aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_ssse3;
@@ -9235,6 +9939,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_ssse3;
+  aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_ssse3;
   aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_ssse3;
@@ -9247,6 +9954,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_ssse3;
+  aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_ssse3;
+  aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_ssse3;
   aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_ssse3;
@@ -9259,12 +9972,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_ssse3;
+  aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_ssse3;
+  aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_ssse3;
   aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_ssse3;
   aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_ssse3;
+  aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_ssse3;
   aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_ssse3;
@@ -9274,6 +9996,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_ssse3;
+  aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_ssse3;
   aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_ssse3;
@@ -9286,6 +10011,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_ssse3;
+  aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_ssse3;
+  aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_ssse3;
   aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_ssse3;
@@ -9298,12 +10029,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_ssse3;
+  aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_ssse3;
+  aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_ssse3;
   aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_ssse3;
   aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_ssse3;
+  aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_ssse3;
   aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_ssse3;
@@ -9313,6 +10053,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_ssse3;
+  aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_ssse3;
   aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_ssse3;
@@ -9476,6 +10219,9 @@ static void setup_rtcd_internal(void) {
   aom_v_predictor_32x64 = aom_v_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_v_predictor_32x64 = aom_v_predictor_32x64_avx2;
+  aom_v_predictor_64x16 = aom_v_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_v_predictor_64x16 = aom_v_predictor_64x16_avx2;
   aom_v_predictor_64x32 = aom_v_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_v_predictor_64x32 = aom_v_predictor_64x32_avx2;
diff --git a/chromium/third_party/libaom/source/config/win/ia32/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/win/ia32/config/av1_rtcd.h
index 8d516d90bde..6d8a64d1270 100644
--- a/chromium/third_party/libaom/source/config/win/ia32/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/win/ia32/config/av1_rtcd.h
@@ -242,6 +242,51 @@ void aom_upsampled_pred_sse2(MACROBLOCKD* xd,
                              int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_sse2
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+void av1_apply_selfguided_restoration_sse4_1(const uint8_t* dat,
+                                             int width,
+                                             int height,
+                                             int stride,
+                                             int eps,
+                                             const int* xqd,
+                                             uint8_t* dst,
+                                             int dst_stride,
+                                             int32_t* tmpbuf,
+                                             int bit_depth,
+                                             int highbd);
+void av1_apply_selfguided_restoration_avx2(const uint8_t* dat,
+                                           int width,
+                                           int height,
+                                           int stride,
+                                           int eps,
+                                           const int* xqd,
+                                           uint8_t* dst,
+                                           int dst_stride,
+                                           int32_t* tmpbuf,
+                                           int bit_depth,
+                                           int highbd);
+RTCD_EXTERN void (*av1_apply_selfguided_restoration)(const uint8_t* dat,
+                                                     int width,
+                                                     int height,
+                                                     int stride,
+                                                     int eps,
+                                                     const int* xqd,
+                                                     uint8_t* dst,
+                                                     int dst_stride,
+                                                     int32_t* tmpbuf,
+                                                     int bit_depth,
+                                                     int highbd);
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -347,6 +392,31 @@ RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(
     ConvolveParams* conv_params,
     int bd);
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+int64_t av1_calc_frame_error_sse2(const uint8_t* const ref,
+                                  int stride,
+                                  const uint8_t* const dst,
+                                  int p_width,
+                                  int p_height,
+                                  int p_stride);
+int64_t av1_calc_frame_error_avx2(const uint8_t* const ref,
+                                  int stride,
+                                  const uint8_t* const dst,
+                                  int p_width,
+                                  int p_height,
+                                  int p_stride);
+RTCD_EXTERN int64_t (*av1_calc_frame_error)(const uint8_t* const ref,
+                                            int stride,
+                                            const uint8_t* const dst,
+                                            int p_width,
+                                            int p_height,
+                                            int p_stride);
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -1839,6 +1909,47 @@ void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 void av1_round_shift_array_sse4_1(int32_t* arr, int size, int bit);
 RTCD_EXTERN void (*av1_round_shift_array)(int32_t* arr, int size, int bit);
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+int av1_selfguided_restoration_sse4_1(const uint8_t* dgd8,
+                                      int width,
+                                      int height,
+                                      int dgd_stride,
+                                      int32_t* flt0,
+                                      int32_t* flt1,
+                                      int flt_stride,
+                                      int sgr_params_idx,
+                                      int bit_depth,
+                                      int highbd);
+int av1_selfguided_restoration_avx2(const uint8_t* dgd8,
+                                    int width,
+                                    int height,
+                                    int dgd_stride,
+                                    int32_t* flt0,
+                                    int32_t* flt1,
+                                    int flt_stride,
+                                    int sgr_params_idx,
+                                    int bit_depth,
+                                    int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t* dgd8,
+                                              int width,
+                                              int height,
+                                              int dgd_stride,
+                                              int32_t* flt0,
+                                              int32_t* flt1,
+                                              int flt_stride,
+                                              int sgr_params_idx,
+                                              int bit_depth,
+                                              int highbd);
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -1864,6 +1975,79 @@ void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 void av1_upsample_intra_edge_high_sse4_1(uint16_t* p, int sz, int bd);
 RTCD_EXTERN void (*av1_upsample_intra_edge_high)(uint16_t* p, int sz, int bd);
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+void av1_warp_affine_sse4_1(const int32_t* mat,
+                            const uint8_t* ref,
+                            int width,
+                            int height,
+                            int stride,
+                            uint8_t* pred,
+                            int p_col,
+                            int p_row,
+                            int p_width,
+                            int p_height,
+                            int p_stride,
+                            int subsampling_x,
+                            int subsampling_y,
+                            ConvolveParams* conv_params,
+                            int16_t alpha,
+                            int16_t beta,
+                            int16_t gamma,
+                            int16_t delta);
+void av1_warp_affine_avx2(const int32_t* mat,
+                          const uint8_t* ref,
+                          int width,
+                          int height,
+                          int stride,
+                          uint8_t* pred,
+                          int p_col,
+                          int p_row,
+                          int p_width,
+                          int p_height,
+                          int p_stride,
+                          int subsampling_x,
+                          int subsampling_y,
+                          ConvolveParams* conv_params,
+                          int16_t alpha,
+                          int16_t beta,
+                          int16_t gamma,
+                          int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t* mat,
+                                    const uint8_t* ref,
+                                    int width,
+                                    int height,
+                                    int stride,
+                                    uint8_t* pred,
+                                    int p_col,
+                                    int p_row,
+                                    int p_width,
+                                    int p_height,
+                                    int p_stride,
+                                    int subsampling_x,
+                                    int subsampling_y,
+                                    ConvolveParams* conv_params,
+                                    int16_t alpha,
+                                    int16_t beta,
+                                    int16_t gamma,
+                                    int16_t delta);
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
@@ -2687,6 +2871,11 @@ static void setup_rtcd_internal(void) {
   if (flags & HAS_SSSE3)
     aom_dist_wtd_comp_avg_upsampled_pred =
         aom_dist_wtd_comp_avg_upsampled_pred_ssse3;
+  av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_c;
+  if (flags & HAS_SSE4_1)
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_avx2;
   av1_block_error = av1_block_error_sse2;
   if (flags & HAS_AVX2)
     av1_block_error = av1_block_error_avx2;
@@ -2705,6 +2894,9 @@ static void setup_rtcd_internal(void) {
   if (flags & HAS_AVX2)
     av1_build_compound_diffwtd_mask_d16 =
         av1_build_compound_diffwtd_mask_d16_avx2;
+  av1_calc_frame_error = av1_calc_frame_error_sse2;
+  if (flags & HAS_AVX2)
+    av1_calc_frame_error = av1_calc_frame_error_avx2;
   av1_calc_indices_dim1 = av1_calc_indices_dim1_sse2;
   if (flags & HAS_AVX2)
     av1_calc_indices_dim1 = av1_calc_indices_dim1_avx2;
@@ -2903,6 +3095,11 @@ static void setup_rtcd_internal(void) {
   av1_round_shift_array = av1_round_shift_array_c;
   if (flags & HAS_SSE4_1)
     av1_round_shift_array = av1_round_shift_array_sse4_1;
+  av1_selfguided_restoration = av1_selfguided_restoration_c;
+  if (flags & HAS_SSE4_1)
+    av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_selfguided_restoration = av1_selfguided_restoration_avx2;
   av1_txb_init_levels = av1_txb_init_levels_c;
   if (flags & HAS_SSE4_1)
     av1_txb_init_levels = av1_txb_init_levels_sse4_1;
@@ -2914,6 +3111,11 @@ static void setup_rtcd_internal(void) {
   av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_c;
   if (flags & HAS_SSE4_1)
     av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_sse4_1;
+  av1_warp_affine = av1_warp_affine_c;
+  if (flags & HAS_SSE4_1)
+    av1_warp_affine = av1_warp_affine_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_warp_affine = av1_warp_affine_avx2;
   av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_sse2;
   if (flags & HAS_AVX2)
     av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_avx2;
diff --git a/chromium/third_party/libaom/source/config/win/x64/config/aom_config.c b/chromium/third_party/libaom/source/config/win/x64/config/aom_config.c
index 8786cc6957d..4c037ce451b 100644
--- a/chromium/third_party/libaom/source/config/win/x64/config/aom_config.c
+++ b/chromium/third_party/libaom/source/config/win/x64/config/aom_config.c
@@ -9,5 +9,5 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "aom/aom_codec.h"
-static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DAOM_TARGET_CPU=x86_64 -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
+static const char* const cfg = "cmake ../source/libaom -G \"Unix Makefiles\" -DAOM_TARGET_CPU=x86_64 -DCONFIG_AV1_ENCODER=1 -DCONFIG_LIBYUV=0 -DCONFIG_AV1_HIGHBITDEPTH=0 -DCONFIG_AV1_TEMPORAL_DENOISING=1 -DCONFIG_REALTIME_ONLY=1 -DCONFIG_MAX_DECODE_PROFILE=0 -DCONFIG_NORMAL_TILE_MODE=1 -DCONFIG_SIZE_LIMIT=1 -DDECODE_HEIGHT_LIMIT=16384 -DDECODE_WIDTH_LIMIT=16384";
 const char *aom_codec_build_config(void) {return cfg;}
diff --git a/chromium/third_party/libaom/source/config/win/x64/config/aom_dsp_rtcd.h b/chromium/third_party/libaom/source/config/win/x64/config/aom_dsp_rtcd.h
index b04c0cfea23..9f1f936e240 100644
--- a/chromium/third_party/libaom/source/config/win/x64/config/aom_dsp_rtcd.h
+++ b/chromium/third_party/libaom/source/config/win/x64/config/aom_dsp_rtcd.h
@@ -360,6 +360,26 @@ void aom_dc_128_predictor_16x32_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_sse2
 
+void aom_dc_128_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_16x4_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_sse2
+
+void aom_dc_128_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_128_predictor_16x64_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_sse2
+
 void aom_dc_128_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -370,12 +390,6 @@ void aom_dc_128_predictor_16x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2
 
-void aom_dc_128_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
-
 void aom_dc_128_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -427,6 +441,26 @@ RTCD_EXTERN void (*aom_dc_128_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_dc_128_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_32x8_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_sse2
+
+void aom_dc_128_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_4x16_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_sse2
+
 void aom_dc_128_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -447,6 +481,23 @@ void aom_dc_128_predictor_4x8_sse2(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_sse2
 
+void aom_dc_128_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_128_predictor_64x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_dc_128_predictor_64x16_avx2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_dc_128_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -491,6 +542,16 @@ void aom_dc_128_predictor_8x16_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_sse2
 
+void aom_dc_128_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_128_predictor_8x32_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_sse2
+
 void aom_dc_128_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -531,6 +592,26 @@ void aom_dc_left_predictor_16x32_sse2(uint8_t* dst,
                                       const uint8_t* left);
 #define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_sse2
 
+void aom_dc_left_predictor_16x4_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_16x4_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_sse2
+
+void aom_dc_left_predictor_16x64_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_dc_left_predictor_16x64_sse2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_sse2
+
 void aom_dc_left_predictor_16x8_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -541,12 +622,6 @@ void aom_dc_left_predictor_16x8_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2
 
-void aom_dc_left_predictor_2x2_c(uint8_t* dst,
-                                 ptrdiff_t y_stride,
-                                 const uint8_t* above,
-                                 const uint8_t* left);
-#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
-
 void aom_dc_left_predictor_32x16_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -598,6 +673,26 @@ RTCD_EXTERN void (*aom_dc_left_predictor_32x64)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_dc_left_predictor_32x8_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_32x8_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_sse2
+
+void aom_dc_left_predictor_4x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_4x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_sse2
+
 void aom_dc_left_predictor_4x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -618,6 +713,23 @@ void aom_dc_left_predictor_4x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_sse2
 
+void aom_dc_left_predictor_64x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_dc_left_predictor_64x16_sse2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+void aom_dc_left_predictor_64x16_avx2(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_dc_left_predictor_64x32_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -662,6 +774,16 @@ void aom_dc_left_predictor_8x16_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_sse2
 
+void aom_dc_left_predictor_8x32_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_left_predictor_8x32_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_sse2
+
 void aom_dc_left_predictor_8x4_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -702,6 +824,26 @@ void aom_dc_predictor_16x32_sse2(uint8_t* dst,
                                  const uint8_t* left);
 #define aom_dc_predictor_16x32 aom_dc_predictor_16x32_sse2
 
+void aom_dc_predictor_16x4_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_16x4_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_sse2
+
+void aom_dc_predictor_16x64_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+void aom_dc_predictor_16x64_sse2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_sse2
+
 void aom_dc_predictor_16x8_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -712,12 +854,6 @@ void aom_dc_predictor_16x8_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2
 
-void aom_dc_predictor_2x2_c(uint8_t* dst,
-                            ptrdiff_t y_stride,
-                            const uint8_t* above,
-                            const uint8_t* left);
-#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
-
 void aom_dc_predictor_32x16_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -769,6 +905,26 @@ RTCD_EXTERN void (*aom_dc_predictor_32x64)(uint8_t* dst,
                                            const uint8_t* above,
                                            const uint8_t* left);
 
+void aom_dc_predictor_32x8_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_32x8_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_sse2
+
+void aom_dc_predictor_4x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_4x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_sse2
+
 void aom_dc_predictor_4x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -789,6 +945,23 @@ void aom_dc_predictor_4x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_dc_predictor_4x8 aom_dc_predictor_4x8_sse2
 
+void aom_dc_predictor_64x16_c(uint8_t* dst,
+                              ptrdiff_t y_stride,
+                              const uint8_t* above,
+                              const uint8_t* left);
+void aom_dc_predictor_64x16_sse2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_predictor_64x16_avx2(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_predictor_64x16)(uint8_t* dst,
+                                           ptrdiff_t y_stride,
+                                           const uint8_t* above,
+                                           const uint8_t* left);
+
 void aom_dc_predictor_64x32_c(uint8_t* dst,
                               ptrdiff_t y_stride,
                               const uint8_t* above,
@@ -833,6 +1006,16 @@ void aom_dc_predictor_8x16_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_dc_predictor_8x16 aom_dc_predictor_8x16_sse2
 
+void aom_dc_predictor_8x32_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_dc_predictor_8x32_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_sse2
+
 void aom_dc_predictor_8x4_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -873,6 +1056,26 @@ void aom_dc_top_predictor_16x32_sse2(uint8_t* dst,
                                      const uint8_t* left);
 #define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_sse2
 
+void aom_dc_top_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_16x4_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_sse2
+
+void aom_dc_top_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_top_predictor_16x64_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_sse2
+
 void aom_dc_top_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -883,12 +1086,6 @@ void aom_dc_top_predictor_16x8_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2
 
-void aom_dc_top_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
-
 void aom_dc_top_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -940,6 +1137,26 @@ RTCD_EXTERN void (*aom_dc_top_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_dc_top_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_32x8_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_sse2
+
+void aom_dc_top_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_4x16_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_sse2
+
 void aom_dc_top_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -960,6 +1177,23 @@ void aom_dc_top_predictor_4x8_sse2(uint8_t* dst,
                                    const uint8_t* left);
 #define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_sse2
 
+void aom_dc_top_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_dc_top_predictor_64x16_sse2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_dc_top_predictor_64x16_avx2(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_dc_top_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -1004,6 +1238,16 @@ void aom_dc_top_predictor_8x16_sse2(uint8_t* dst,
                                     const uint8_t* left);
 #define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_sse2
 
+void aom_dc_top_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_dc_top_predictor_8x32_sse2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_sse2
+
 void aom_dc_top_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -2027,6 +2271,26 @@ void aom_h_predictor_16x32_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_h_predictor_16x32 aom_h_predictor_16x32_sse2
 
+void aom_h_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_16x4_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_sse2
+
+void aom_h_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_h_predictor_16x64_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_sse2
+
 void aom_h_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -2037,12 +2301,6 @@ void aom_h_predictor_16x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2
 
-void aom_h_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
-
 void aom_h_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -2080,6 +2338,26 @@ void aom_h_predictor_32x64_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_h_predictor_32x64 aom_h_predictor_32x64_sse2
 
+void aom_h_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_32x8_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_sse2
+
+void aom_h_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_4x16_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_sse2
+
 void aom_h_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -2100,6 +2378,16 @@ void aom_h_predictor_4x8_sse2(uint8_t* dst,
                               const uint8_t* left);
 #define aom_h_predictor_4x8 aom_h_predictor_4x8_sse2
 
+void aom_h_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_h_predictor_64x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_sse2
+
 void aom_h_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -2130,6 +2418,16 @@ void aom_h_predictor_8x16_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_h_predictor_8x16 aom_h_predictor_8x16_sse2
 
+void aom_h_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_h_predictor_8x32_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_sse2
+
 void aom_h_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -4360,6 +4658,36 @@ RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_paeth_predictor_16x4_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_16x4_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
+void aom_paeth_predictor_16x64_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_16x64_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_paeth_predictor_16x64_avx2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_paeth_predictor_16x8_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -4377,12 +4705,6 @@ RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
-void aom_paeth_predictor_2x2_c(uint8_t* dst,
-                               ptrdiff_t y_stride,
-                               const uint8_t* above,
-                               const uint8_t* left);
-#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
-
 void aom_paeth_predictor_32x16_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -4434,6 +4756,32 @@ RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_paeth_predictor_32x8_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_32x8_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
+void aom_paeth_predictor_4x16_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_4x16_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
 void aom_paeth_predictor_4x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -4460,6 +4808,23 @@ RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t* dst,
                                             const uint8_t* above,
                                             const uint8_t* left);
 
+void aom_paeth_predictor_64x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_paeth_predictor_64x16_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+void aom_paeth_predictor_64x16_avx2(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_paeth_predictor_64x32_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -4507,6 +4872,19 @@ RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
+void aom_paeth_predictor_8x32_c(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_paeth_predictor_8x32_ssse3(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t* dst,
+                                             ptrdiff_t y_stride,
+                                             const uint8_t* above,
+                                             const uint8_t* left);
+
 void aom_paeth_predictor_8x4_c(uint8_t* dst,
                                ptrdiff_t y_stride,
                                const uint8_t* above,
@@ -6283,6 +6661,32 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_h_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_16x4_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_h_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_16x64_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_h_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -6296,12 +6700,6 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
-void aom_smooth_h_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
-
 void aom_smooth_h_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6341,6 +6739,32 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_h_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_32x8_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_h_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_4x16_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_h_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6367,6 +6791,19 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_h_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_h_predictor_64x16_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_h_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6406,6 +6843,19 @@ RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_smooth_h_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_h_predictor_8x32_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_h_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6458,6 +6908,32 @@ RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_predictor_16x4_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_16x4_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_smooth_predictor_16x64_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_16x64_ssse3(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_smooth_predictor_16x8_c(uint8_t* dst,
                                  ptrdiff_t y_stride,
                                  const uint8_t* above,
@@ -6471,12 +6947,6 @@ RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
-void aom_smooth_predictor_2x2_c(uint8_t* dst,
-                                ptrdiff_t y_stride,
-                                const uint8_t* above,
-                                const uint8_t* left);
-#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
-
 void aom_smooth_predictor_32x16_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6516,6 +6986,32 @@ RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_predictor_32x8_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_32x8_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
+void aom_smooth_predictor_4x16_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_4x16_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_smooth_predictor_4x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -6542,6 +7038,19 @@ RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t* dst,
                                              const uint8_t* above,
                                              const uint8_t* left);
 
+void aom_smooth_predictor_64x16_c(uint8_t* dst,
+                                  ptrdiff_t y_stride,
+                                  const uint8_t* above,
+                                  const uint8_t* left);
+void aom_smooth_predictor_64x16_ssse3(uint8_t* dst,
+                                      ptrdiff_t y_stride,
+                                      const uint8_t* above,
+                                      const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t* dst,
+                                               ptrdiff_t y_stride,
+                                               const uint8_t* above,
+                                               const uint8_t* left);
+
 void aom_smooth_predictor_64x32_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6581,6 +7090,19 @@ RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t* dst,
                                               const uint8_t* above,
                                               const uint8_t* left);
 
+void aom_smooth_predictor_8x32_c(uint8_t* dst,
+                                 ptrdiff_t y_stride,
+                                 const uint8_t* above,
+                                 const uint8_t* left);
+void aom_smooth_predictor_8x32_ssse3(uint8_t* dst,
+                                     ptrdiff_t y_stride,
+                                     const uint8_t* above,
+                                     const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t* dst,
+                                              ptrdiff_t y_stride,
+                                              const uint8_t* above,
+                                              const uint8_t* left);
+
 void aom_smooth_predictor_8x4_c(uint8_t* dst,
                                 ptrdiff_t y_stride,
                                 const uint8_t* above,
@@ -6633,6 +7155,32 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_v_predictor_16x4_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_16x4_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_v_predictor_16x64_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_16x64_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_v_predictor_16x8_c(uint8_t* dst,
                                    ptrdiff_t y_stride,
                                    const uint8_t* above,
@@ -6646,12 +7194,6 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
-void aom_smooth_v_predictor_2x2_c(uint8_t* dst,
-                                  ptrdiff_t y_stride,
-                                  const uint8_t* above,
-                                  const uint8_t* left);
-#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
-
 void aom_smooth_v_predictor_32x16_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6691,6 +7233,32 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t* dst,
                                                  const uint8_t* above,
                                                  const uint8_t* left);
 
+void aom_smooth_v_predictor_32x8_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_32x8_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
+void aom_smooth_v_predictor_4x16_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_4x16_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_v_predictor_4x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -6717,6 +7285,19 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t* dst,
                                                const uint8_t* above,
                                                const uint8_t* left);
 
+void aom_smooth_v_predictor_64x16_c(uint8_t* dst,
+                                    ptrdiff_t y_stride,
+                                    const uint8_t* above,
+                                    const uint8_t* left);
+void aom_smooth_v_predictor_64x16_ssse3(uint8_t* dst,
+                                        ptrdiff_t y_stride,
+                                        const uint8_t* above,
+                                        const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t* dst,
+                                                 ptrdiff_t y_stride,
+                                                 const uint8_t* above,
+                                                 const uint8_t* left);
+
 void aom_smooth_v_predictor_64x32_c(uint8_t* dst,
                                     ptrdiff_t y_stride,
                                     const uint8_t* above,
@@ -6756,6 +7337,19 @@ RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t* dst,
                                                 const uint8_t* above,
                                                 const uint8_t* left);
 
+void aom_smooth_v_predictor_8x32_c(uint8_t* dst,
+                                   ptrdiff_t y_stride,
+                                   const uint8_t* above,
+                                   const uint8_t* left);
+void aom_smooth_v_predictor_8x32_ssse3(uint8_t* dst,
+                                       ptrdiff_t y_stride,
+                                       const uint8_t* above,
+                                       const uint8_t* left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t* dst,
+                                                ptrdiff_t y_stride,
+                                                const uint8_t* above,
+                                                const uint8_t* left);
+
 void aom_smooth_v_predictor_8x4_c(uint8_t* dst,
                                   ptrdiff_t y_stride,
                                   const uint8_t* above,
@@ -8071,6 +8665,26 @@ void aom_v_predictor_16x32_sse2(uint8_t* dst,
                                 const uint8_t* left);
 #define aom_v_predictor_16x32 aom_v_predictor_16x32_sse2
 
+void aom_v_predictor_16x4_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_16x4_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_sse2
+
+void aom_v_predictor_16x64_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_v_predictor_16x64_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_sse2
+
 void aom_v_predictor_16x8_c(uint8_t* dst,
                             ptrdiff_t y_stride,
                             const uint8_t* above,
@@ -8081,12 +8695,6 @@ void aom_v_predictor_16x8_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2
 
-void aom_v_predictor_2x2_c(uint8_t* dst,
-                           ptrdiff_t y_stride,
-                           const uint8_t* above,
-                           const uint8_t* left);
-#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
-
 void aom_v_predictor_32x16_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -8138,6 +8746,26 @@ RTCD_EXTERN void (*aom_v_predictor_32x64)(uint8_t* dst,
                                           const uint8_t* above,
                                           const uint8_t* left);
 
+void aom_v_predictor_32x8_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_32x8_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_sse2
+
+void aom_v_predictor_4x16_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_4x16_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_sse2
+
 void aom_v_predictor_4x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -8158,6 +8786,23 @@ void aom_v_predictor_4x8_sse2(uint8_t* dst,
                               const uint8_t* left);
 #define aom_v_predictor_4x8 aom_v_predictor_4x8_sse2
 
+void aom_v_predictor_64x16_c(uint8_t* dst,
+                             ptrdiff_t y_stride,
+                             const uint8_t* above,
+                             const uint8_t* left);
+void aom_v_predictor_64x16_sse2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+void aom_v_predictor_64x16_avx2(uint8_t* dst,
+                                ptrdiff_t y_stride,
+                                const uint8_t* above,
+                                const uint8_t* left);
+RTCD_EXTERN void (*aom_v_predictor_64x16)(uint8_t* dst,
+                                          ptrdiff_t y_stride,
+                                          const uint8_t* above,
+                                          const uint8_t* left);
+
 void aom_v_predictor_64x32_c(uint8_t* dst,
                              ptrdiff_t y_stride,
                              const uint8_t* above,
@@ -8202,6 +8847,16 @@ void aom_v_predictor_8x16_sse2(uint8_t* dst,
                                const uint8_t* left);
 #define aom_v_predictor_8x16 aom_v_predictor_8x16_sse2
 
+void aom_v_predictor_8x32_c(uint8_t* dst,
+                            ptrdiff_t y_stride,
+                            const uint8_t* above,
+                            const uint8_t* left);
+void aom_v_predictor_8x32_sse2(uint8_t* dst,
+                               ptrdiff_t y_stride,
+                               const uint8_t* above,
+                               const uint8_t* left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_sse2
+
 void aom_v_predictor_8x4_c(uint8_t* dst,
                            ptrdiff_t y_stride,
                            const uint8_t* above,
@@ -8615,6 +9270,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_avx2;
+  aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_avx2;
   aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_avx2;
@@ -8630,6 +9288,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_avx2;
+  aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_avx2;
   aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_avx2;
@@ -8645,6 +9306,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_predictor_32x64 = aom_dc_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_predictor_32x64 = aom_dc_predictor_32x64_avx2;
+  aom_dc_predictor_64x16 = aom_dc_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_predictor_64x16 = aom_dc_predictor_64x16_avx2;
   aom_dc_predictor_64x32 = aom_dc_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_predictor_64x32 = aom_dc_predictor_64x32_avx2;
@@ -8660,6 +9324,9 @@ static void setup_rtcd_internal(void) {
   aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_avx2;
+  aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_avx2;
   aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_avx2;
@@ -9055,6 +9722,14 @@ static void setup_rtcd_internal(void) {
     aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_ssse3;
   if (flags & HAS_AVX2)
     aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_avx2;
+  aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_ssse3;
+  aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_ssse3;
+  if (flags & HAS_AVX2)
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_avx2;
   aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_ssse3;
@@ -9075,12 +9750,23 @@ static void setup_rtcd_internal(void) {
     aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_ssse3;
   if (flags & HAS_AVX2)
     aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_avx2;
+  aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_ssse3;
+  aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_ssse3;
   aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_ssse3;
   aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_ssse3;
+  aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_ssse3;
+  if (flags & HAS_AVX2)
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_avx2;
   aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_ssse3;
@@ -9094,6 +9780,9 @@ static void setup_rtcd_internal(void) {
   aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_ssse3;
+  aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_ssse3;
   aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3;
@@ -9251,6 +9940,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_ssse3;
+  aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_ssse3;
+  aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_ssse3;
   aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_ssse3;
@@ -9263,12 +9958,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_ssse3;
+  aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_ssse3;
+  aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_ssse3;
   aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_ssse3;
   aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_ssse3;
+  aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_ssse3;
   aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_ssse3;
@@ -9278,6 +9982,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_ssse3;
+  aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_ssse3;
   aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_ssse3;
@@ -9290,6 +9997,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_ssse3;
+  aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_ssse3;
+  aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_ssse3;
   aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_ssse3;
@@ -9302,12 +10015,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_ssse3;
+  aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_ssse3;
+  aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_ssse3;
   aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_ssse3;
   aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_ssse3;
+  aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_ssse3;
   aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_ssse3;
@@ -9317,6 +10039,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_ssse3;
+  aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_ssse3;
   aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_ssse3;
@@ -9329,6 +10054,12 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_ssse3;
+  aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_ssse3;
+  aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_ssse3;
   aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_ssse3;
@@ -9341,12 +10072,21 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_ssse3;
+  aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_ssse3;
+  aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_ssse3;
   aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_ssse3;
   aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_ssse3;
+  aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_ssse3;
   aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_ssse3;
@@ -9356,6 +10096,9 @@ static void setup_rtcd_internal(void) {
   aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_ssse3;
+  aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c;
+  if (flags & HAS_SSSE3)
+    aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_ssse3;
   aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c;
   if (flags & HAS_SSSE3)
     aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_ssse3;
@@ -9519,6 +10262,9 @@ static void setup_rtcd_internal(void) {
   aom_v_predictor_32x64 = aom_v_predictor_32x64_sse2;
   if (flags & HAS_AVX2)
     aom_v_predictor_32x64 = aom_v_predictor_32x64_avx2;
+  aom_v_predictor_64x16 = aom_v_predictor_64x16_sse2;
+  if (flags & HAS_AVX2)
+    aom_v_predictor_64x16 = aom_v_predictor_64x16_avx2;
   aom_v_predictor_64x32 = aom_v_predictor_64x32_sse2;
   if (flags & HAS_AVX2)
     aom_v_predictor_64x32 = aom_v_predictor_64x32_avx2;
diff --git a/chromium/third_party/libaom/source/config/win/x64/config/av1_rtcd.h b/chromium/third_party/libaom/source/config/win/x64/config/av1_rtcd.h
index 8d516d90bde..6d8a64d1270 100644
--- a/chromium/third_party/libaom/source/config/win/x64/config/av1_rtcd.h
+++ b/chromium/third_party/libaom/source/config/win/x64/config/av1_rtcd.h
@@ -242,6 +242,51 @@ void aom_upsampled_pred_sse2(MACROBLOCKD* xd,
                              int subpel_search);
 #define aom_upsampled_pred aom_upsampled_pred_sse2
 
+void av1_apply_selfguided_restoration_c(const uint8_t* dat,
+                                        int width,
+                                        int height,
+                                        int stride,
+                                        int eps,
+                                        const int* xqd,
+                                        uint8_t* dst,
+                                        int dst_stride,
+                                        int32_t* tmpbuf,
+                                        int bit_depth,
+                                        int highbd);
+void av1_apply_selfguided_restoration_sse4_1(const uint8_t* dat,
+                                             int width,
+                                             int height,
+                                             int stride,
+                                             int eps,
+                                             const int* xqd,
+                                             uint8_t* dst,
+                                             int dst_stride,
+                                             int32_t* tmpbuf,
+                                             int bit_depth,
+                                             int highbd);
+void av1_apply_selfguided_restoration_avx2(const uint8_t* dat,
+                                           int width,
+                                           int height,
+                                           int stride,
+                                           int eps,
+                                           const int* xqd,
+                                           uint8_t* dst,
+                                           int dst_stride,
+                                           int32_t* tmpbuf,
+                                           int bit_depth,
+                                           int highbd);
+RTCD_EXTERN void (*av1_apply_selfguided_restoration)(const uint8_t* dat,
+                                                     int width,
+                                                     int height,
+                                                     int stride,
+                                                     int eps,
+                                                     const int* xqd,
+                                                     uint8_t* dst,
+                                                     int dst_stride,
+                                                     int32_t* tmpbuf,
+                                                     int bit_depth,
+                                                     int highbd);
+
 int64_t av1_block_error_c(const tran_low_t* coeff,
                           const tran_low_t* dqcoeff,
                           intptr_t block_size,
@@ -347,6 +392,31 @@ RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(
     ConvolveParams* conv_params,
     int bd);
 
+int64_t av1_calc_frame_error_c(const uint8_t* const ref,
+                               int stride,
+                               const uint8_t* const dst,
+                               int p_width,
+                               int p_height,
+                               int p_stride);
+int64_t av1_calc_frame_error_sse2(const uint8_t* const ref,
+                                  int stride,
+                                  const uint8_t* const dst,
+                                  int p_width,
+                                  int p_height,
+                                  int p_stride);
+int64_t av1_calc_frame_error_avx2(const uint8_t* const ref,
+                                  int stride,
+                                  const uint8_t* const dst,
+                                  int p_width,
+                                  int p_height,
+                                  int p_stride);
+RTCD_EXTERN int64_t (*av1_calc_frame_error)(const uint8_t* const ref,
+                                            int stride,
+                                            const uint8_t* const dst,
+                                            int p_width,
+                                            int p_height,
+                                            int p_stride);
+
 void av1_calc_indices_dim1_c(const int* data,
                              const int* centroids,
                              uint8_t* indices,
@@ -1839,6 +1909,47 @@ void av1_round_shift_array_c(int32_t* arr, int size, int bit);
 void av1_round_shift_array_sse4_1(int32_t* arr, int size, int bit);
 RTCD_EXTERN void (*av1_round_shift_array)(int32_t* arr, int size, int bit);
 
+int av1_selfguided_restoration_c(const uint8_t* dgd8,
+                                 int width,
+                                 int height,
+                                 int dgd_stride,
+                                 int32_t* flt0,
+                                 int32_t* flt1,
+                                 int flt_stride,
+                                 int sgr_params_idx,
+                                 int bit_depth,
+                                 int highbd);
+int av1_selfguided_restoration_sse4_1(const uint8_t* dgd8,
+                                      int width,
+                                      int height,
+                                      int dgd_stride,
+                                      int32_t* flt0,
+                                      int32_t* flt1,
+                                      int flt_stride,
+                                      int sgr_params_idx,
+                                      int bit_depth,
+                                      int highbd);
+int av1_selfguided_restoration_avx2(const uint8_t* dgd8,
+                                    int width,
+                                    int height,
+                                    int dgd_stride,
+                                    int32_t* flt0,
+                                    int32_t* flt1,
+                                    int flt_stride,
+                                    int sgr_params_idx,
+                                    int bit_depth,
+                                    int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t* dgd8,
+                                              int width,
+                                              int height,
+                                              int dgd_stride,
+                                              int32_t* flt0,
+                                              int32_t* flt1,
+                                              int flt_stride,
+                                              int sgr_params_idx,
+                                              int bit_depth,
+                                              int highbd);
+
 void av1_txb_init_levels_c(const tran_low_t* const coeff,
                            const int width,
                            const int height,
@@ -1864,6 +1975,79 @@ void av1_upsample_intra_edge_high_c(uint16_t* p, int sz, int bd);
 void av1_upsample_intra_edge_high_sse4_1(uint16_t* p, int sz, int bd);
 RTCD_EXTERN void (*av1_upsample_intra_edge_high)(uint16_t* p, int sz, int bd);
 
+void av1_warp_affine_c(const int32_t* mat,
+                       const uint8_t* ref,
+                       int width,
+                       int height,
+                       int stride,
+                       uint8_t* pred,
+                       int p_col,
+                       int p_row,
+                       int p_width,
+                       int p_height,
+                       int p_stride,
+                       int subsampling_x,
+                       int subsampling_y,
+                       ConvolveParams* conv_params,
+                       int16_t alpha,
+                       int16_t beta,
+                       int16_t gamma,
+                       int16_t delta);
+void av1_warp_affine_sse4_1(const int32_t* mat,
+                            const uint8_t* ref,
+                            int width,
+                            int height,
+                            int stride,
+                            uint8_t* pred,
+                            int p_col,
+                            int p_row,
+                            int p_width,
+                            int p_height,
+                            int p_stride,
+                            int subsampling_x,
+                            int subsampling_y,
+                            ConvolveParams* conv_params,
+                            int16_t alpha,
+                            int16_t beta,
+                            int16_t gamma,
+                            int16_t delta);
+void av1_warp_affine_avx2(const int32_t* mat,
+                          const uint8_t* ref,
+                          int width,
+                          int height,
+                          int stride,
+                          uint8_t* pred,
+                          int p_col,
+                          int p_row,
+                          int p_width,
+                          int p_height,
+                          int p_stride,
+                          int subsampling_x,
+                          int subsampling_y,
+                          ConvolveParams* conv_params,
+                          int16_t alpha,
+                          int16_t beta,
+                          int16_t gamma,
+                          int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t* mat,
+                                    const uint8_t* ref,
+                                    int width,
+                                    int height,
+                                    int stride,
+                                    uint8_t* pred,
+                                    int p_col,
+                                    int p_row,
+                                    int p_width,
+                                    int p_height,
+                                    int p_stride,
+                                    int subsampling_x,
+                                    int subsampling_y,
+                                    ConvolveParams* conv_params,
+                                    int16_t alpha,
+                                    int16_t beta,
+                                    int16_t gamma,
+                                    int16_t delta);
+
 void av1_wedge_compute_delta_squares_c(int16_t* d,
                                        const int16_t* a,
                                        const int16_t* b,
@@ -2687,6 +2871,11 @@ static void setup_rtcd_internal(void) {
   if (flags & HAS_SSSE3)
     aom_dist_wtd_comp_avg_upsampled_pred =
         aom_dist_wtd_comp_avg_upsampled_pred_ssse3;
+  av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_c;
+  if (flags & HAS_SSE4_1)
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_avx2;
   av1_block_error = av1_block_error_sse2;
   if (flags & HAS_AVX2)
     av1_block_error = av1_block_error_avx2;
@@ -2705,6 +2894,9 @@ static void setup_rtcd_internal(void) {
   if (flags & HAS_AVX2)
     av1_build_compound_diffwtd_mask_d16 =
         av1_build_compound_diffwtd_mask_d16_avx2;
+  av1_calc_frame_error = av1_calc_frame_error_sse2;
+  if (flags & HAS_AVX2)
+    av1_calc_frame_error = av1_calc_frame_error_avx2;
   av1_calc_indices_dim1 = av1_calc_indices_dim1_sse2;
   if (flags & HAS_AVX2)
     av1_calc_indices_dim1 = av1_calc_indices_dim1_avx2;
@@ -2903,6 +3095,11 @@ static void setup_rtcd_internal(void) {
   av1_round_shift_array = av1_round_shift_array_c;
   if (flags & HAS_SSE4_1)
     av1_round_shift_array = av1_round_shift_array_sse4_1;
+  av1_selfguided_restoration = av1_selfguided_restoration_c;
+  if (flags & HAS_SSE4_1)
+    av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_selfguided_restoration = av1_selfguided_restoration_avx2;
   av1_txb_init_levels = av1_txb_init_levels_c;
   if (flags & HAS_SSE4_1)
     av1_txb_init_levels = av1_txb_init_levels_sse4_1;
@@ -2914,6 +3111,11 @@ static void setup_rtcd_internal(void) {
   av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_c;
   if (flags & HAS_SSE4_1)
     av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_sse4_1;
+  av1_warp_affine = av1_warp_affine_c;
+  if (flags & HAS_SSE4_1)
+    av1_warp_affine = av1_warp_affine_sse4_1;
+  if (flags & HAS_AVX2)
+    av1_warp_affine = av1_warp_affine_avx2;
   av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_sse2;
   if (flags & HAS_AVX2)
     av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_avx2;