swscale/arm: add ff_nv{12,21}_to_{argb,rgba,abgr,bgra}_neon_16

2 files changed, 167 insertions, 59 deletions

diff --git a/libswscale/arm/swscale_unscaled.c b/libswscale/arm/swscale_unscaled.c
index 5d0c8e9e4c..dbb0fb05e1 100644
--- a/libswscale/arm/swscale_unscaled.c
+++ b/libswscale/arm/swscale_unscaled.c
@@ -63,8 +63,8 @@ static int rgbx_to_nv12_neon_16_wrapper(SwsContext *context, const uint8_t *src[
 }
 #endif
 
-#define DECLARE_FF_NVX_TO_RGBX_FUNCS(ifmt, ofmt)                                            \
-int ff_##ifmt##_to_##ofmt##_neon(int w, int h,                                              \
+#define DECLARE_FF_NVX_TO_RGBX_FUNCS(ifmt, ofmt, precision)                                 \
+int ff_##ifmt##_to_##ofmt##_neon_##precision(int w, int h,                                  \
                                  uint8_t *dst, int linesize,                                \
                                  const uint8_t *srcY, int linesizeY,                        \
                                  const uint8_t *srcC, int linesizeC,                        \
@@ -72,59 +72,64 @@ int ff_##ifmt##_to_##ofmt##_neon(int w, int h,
                                  int y_offset,                                              \
                                  int y_coeff);                                              \
                                                                                             \
-static int ifmt##_to_##ofmt##_neon_wrapper(SwsContext *c, const uint8_t *src[],             \
+static int ifmt##_to_##ofmt##_neon_wrapper_##precision(SwsContext *c, const uint8_t *src[], \
                                            int srcStride[], int srcSliceY, int srcSliceH,   \
                                            uint8_t *dst[], int dstStride[]) {               \
     const int16_t yuv2rgb_table[] = {                                                       \
-        c->yuv2rgb_v2r_coeff,                                                               \
-        c->yuv2rgb_u2g_coeff,                                                               \
-        c->yuv2rgb_v2g_coeff,                                                               \
-        c->yuv2rgb_u2b_coeff,                                                               \
+        c->yuv2rgb_v2r_coeff / ((precision) == 16 ? 1 << 7 : 1),                            \
+        c->yuv2rgb_u2g_coeff / ((precision) == 16 ? 1 << 7 : 1),                            \
+        c->yuv2rgb_v2g_coeff / ((precision) == 16 ? 1 << 7 : 1),                            \
+        c->yuv2rgb_u2b_coeff / ((precision) == 16 ? 1 << 7 : 1),                            \
     };                                                                                      \
                                                                                             \
-    ff_##ifmt##_to_##ofmt##_neon(c->srcW, srcSliceH,                                        \
+    ff_##ifmt##_to_##ofmt##_neon_##precision(c->srcW, srcSliceH,                            \
                                  dst[0] +  srcSliceY      * dstStride[0], dstStride[0],     \
                                  src[0] +  srcSliceY      * srcStride[0], srcStride[0],     \
                                  src[1] + (srcSliceY / 2) * srcStride[1], srcStride[1],     \
                                  yuv2rgb_table,                                             \
                                  c->yuv2rgb_y_offset >> 9,                                  \
-                                 c->yuv2rgb_y_coeff);                                       \
+                                 c->yuv2rgb_y_coeff / ((precision) == 16 ? 1 << 7 : 1));    \
                                                                                             \
     return 0;                                                                               \
-}
+}                                                                                           \
+
+#define DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nvx, precision)                                    \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, argb, precision)                                          \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, rgba, precision)                                          \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, abgr, precision)                                          \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, bgra, precision)                                          \
 
-#define DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nvx)                                               \
-DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, argb)                                                     \
-DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, rgba)                                                     \
-DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, abgr)                                                     \
-DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, bgra)                                                     \
+#define DECLARE_FF_NVX_TO_ALL_RGBX_ALL_PRECISION_FUNCS(nvx)                                 \
+DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nvx, 16)                                                   \
+DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nvx, 32)                                                   \
 
-DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nv12)
-DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nv21)
+DECLARE_FF_NVX_TO_ALL_RGBX_ALL_PRECISION_FUNCS(nv12)
+DECLARE_FF_NVX_TO_ALL_RGBX_ALL_PRECISION_FUNCS(nv21)
 
 /* We need a 16 pixel width alignment. This constraint can easily be removed
  * for input reading but for the output which is 4-bytes per pixel (RGBA) the
  * assembly might be writing as much as 4*15=60 extra bytes at the end of the
  * line, which won't fit the 32-bytes buffer alignment. */
-#define SET_FF_NVX_TO_RGBX_FUNC(ifmt, IFMT, ofmt, OFMT) do {                                \
+#define SET_FF_NVX_TO_RGBX_FUNC(ifmt, IFMT, ofmt, OFMT, accurate_rnd) do {                  \
     if (c->srcFormat == AV_PIX_FMT_##IFMT                                                   \
         && c->dstFormat == AV_PIX_FMT_##OFMT                                                \
         && !(c->srcH & 1)                                                                   \
         && !(c->srcW & 15)) {                                                               \
-        c->swscale = ifmt##_to_##ofmt##_neon_wrapper;                                       \
+        c->swscale = (accurate_rnd) ? ifmt##_to_##ofmt##_neon_wrapper_32 :                  \
+                                      ifmt##_to_##ofmt##_neon_wrapper_16 ;                  \
     }                                                                                       \
 } while (0)
 
-#define SET_FF_NVX_TO_ALL_RGBX_FUNC(nvx, NVX) do {                                          \
-    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, argb, ARGB);                                          \
-    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, rgba, RGBA);                                          \
-    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, abgr, ABGR);                                          \
-    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, bgra, BGRA);                                          \
+#define SET_FF_NVX_TO_ALL_RGBX_FUNC(nvx, NVX, accurate_rnd) do {                            \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, argb, ARGB, accurate_rnd);                            \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, rgba, RGBA, accurate_rnd);                            \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, abgr, ABGR, accurate_rnd);                            \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, bgra, BGRA, accurate_rnd);                            \
 } while (0)
 
 static void get_unscaled_swscale_neon(SwsContext *c) {
-#if 0
     int accurate_rnd = c->flags & SWS_ACCURATE_RND;
+#if 0
     if (c->srcFormat == AV_PIX_FMT_RGBA
             && c->dstFormat == AV_PIX_FMT_NV12
             && (c->srcW >= 16)) {
@@ -133,8 +138,8 @@ static void get_unscaled_swscale_neon(SwsContext *c) {
     }
 #endif
 
-    SET_FF_NVX_TO_ALL_RGBX_FUNC(nv12, NV12);
-    SET_FF_NVX_TO_ALL_RGBX_FUNC(nv21, NV21);
+    SET_FF_NVX_TO_ALL_RGBX_FUNC(nv12, NV12, accurate_rnd);
+    SET_FF_NVX_TO_ALL_RGBX_FUNC(nv21, NV21, accurate_rnd);
 }
 
 void ff_get_unscaled_swscale_arm(SwsContext *c)
diff --git a/libswscale/arm/yuv2rgb_neon.S b/libswscale/arm/yuv2rgb_neon.S
index 39b18cd34c..01d8536cdb 100644
--- a/libswscale/arm/yuv2rgb_neon.S
+++ b/libswscale/arm/yuv2rgb_neon.S
@@ -21,7 +21,35 @@
 
 #include "libavutil/arm/asm.S"
 
-.macro compute_premult half_u half_v
+
+.macro compute_premult_16 half_u1, half_u2, half_v1, half_v2
+    vmov                d2, \half_u1                                   @ copy left q14 to left q1
+    vmov                d3, \half_u1                                   @ copy left q14 to right q1
+    vmov                d4, \half_u2                                   @ copy right q14 to left q2
+    vmov                d5, \half_u2                                   @ copy right q14 to right q2
+
+    vmov                d6, \half_v1                                   @ copy left q15 to left q3
+    vmov                d7, \half_v1                                   @ copy left q15 to right q3
+    vmov                d8, \half_v2                                   @ copy right q15 to left q4
+    vmov                d9, \half_v2                                   @ copy right q15 to right q4
+
+    vzip.16             d2, d3                                         @ U1U1U2U2U3U3U4U4
+    vzip.16             d4, d5                                         @ U5U5U6U6U7U7U8U8
+
+    vzip.16             d6, d7                                         @ V1V1V2V2V3V3V4V4
+    vzip.16             d8, d9                                         @ V5V5V6V6V7V7V8V8
+
+    vmul.s16            q8,  q3, d1[0]                                 @  V * v2r             (left,  red)
+    vmul.s16            q9,  q4, d1[0]                                 @  V * v2r             (right, red)
+    vmul.s16            q10, q1, d1[1]                                 @  U * u2g
+    vmul.s16            q11, q2, d1[1]                                 @  U * u2g
+    vmla.s16            q10, q3, d1[2]                                 @  U * u2g + V * v2g   (left,  green)
+    vmla.s16            q11, q4, d1[2]                                 @  U * u2g + V * v2g   (right, green)
+    vmul.s16            q12, q1, d1[3]                                 @  U * u2b             (left,  blue)
+    vmul.s16            q13, q2, d1[3]                                 @  U * u2b             (right, blue)
+.endm
+
+.macro compute_premult_32 half_u half_v
     vmov                d2, \half_u                                    @ copy left q14 to left q1
     vmov                d3, \half_u                                    @ copy left q14 to right q1
     vmov                d4, \half_v                                    @ copy left q15 to left q2
@@ -40,7 +68,14 @@
     vmull.s16           q13, d3, d1[3]                                 @  U * u2b             (right, blue)
 .endm
 
-.macro compute_color dst_comp pre1 pre2
+.macro compute_color_16 dst_comp1 dst_comp2 pre1 pre2
+    vadd.s16            q1, q14, \pre1
+    vadd.s16            q2, q15, \pre2
+    vqrshrun.s16        \dst_comp1, q1, #6
+    vqrshrun.s16        \dst_comp2, q2, #6
+.endm
+
+.macro compute_color_32 dst_comp pre1 pre2
     vadd.s32            q3, q1, \pre1
     vadd.s32            q4, q2, \pre2
     vqrshrun.s32        d10, q3, #13
@@ -48,14 +83,56 @@
     vqmovn.u16          \dst_comp, q5                                  @ saturate 16bit -> 8bit
 .endm
 
-.macro compute_rgba r g b a
-    compute_color       \r, q8,  q9
-    compute_color       \g, q10, q11
-    compute_color       \b, q12, q13
+.macro compute_rgba_16 r1 r2 g1 g2 b1 b2 a1 a2
+    compute_color_16    \r1, \r2, q8,  q9
+    compute_color_16    \g1, \g2, q10, q11
+    compute_color_16    \b1, \b2, q12, q13
+    vmov.u8             \a1, #255
+    vmov.u8             \a2, #255
+.endm
+
+.macro compute_rgba_32 r g b a
+    compute_color_32    \r, q8,  q9
+    compute_color_32    \g, q10, q11
+    compute_color_32    \b, q12, q13
     vmov.u8             \a, #255
 .endm
 
-.macro compute_half_line dst half_y ofmt
+.macro compute_16px_16 dst y0 y1 ofmt
+    vmovl.u8            q14, \y0                                       @ 8px of y
+    vmovl.u8            q15, \y1                                       @ 8px of y
+
+    vdup.16             q5, r9                                         @ q5  = y_offset
+    vdup.16             q7, r10                                        @ q7  = y_coeff
+
+    vsub.s16            q14, q5
+    vsub.s16            q15, q5
+
+    vmul.s16            q14, q7                                        @ q14 = (srcY - y_offset) * y_coeff (left)
+    vmul.s16            q15, q7                                        @ q15 = (srcY - y_offset) * y_coeff (right)
+
+
+.ifc \ofmt,argb
+    compute_rgba_16     d7, d11, d8, d12, d9, d13, d6, d10
+.endif
+
+.ifc \ofmt,rgba
+    compute_rgba_16     d6, d10, d7, d11, d8, d12, d9, d13
+.endif
+
+.ifc \ofmt,abgr
+    compute_rgba_16     d9, d13, d8, d12, d7, d11, d6, d10
+.endif
+
+.ifc \ofmt,bgra
+    compute_rgba_16     d8, d12, d7, d11, d6, d10, d9, d13
+.endif
+    vst4.8              {q3, q4}, [\dst,:128]!
+    vst4.8              {q5, q6}, [\dst,:128]!
+
+.endm
+
+.macro compute_8px_32 dst half_y ofmt
     vmovl.u8            q7, \half_y                                    @ 8px of Y
     vdup.16             q5, r9
     vsub.s16            q7, q5
@@ -63,26 +140,51 @@
     vmull.s16           q2, d15, d0                                    @ q2 = (srcY - y_offset) * y_coeff (right)
 
 .ifc \ofmt,argb
-    compute_rgba        d13, d14, d15, d12
+    compute_rgba_32     d13, d14, d15, d12
 .endif
 
 .ifc \ofmt,rgba
-    compute_rgba        d12, d13, d14, d15
+    compute_rgba_32     d12, d13, d14, d15
 .endif
 
 .ifc \ofmt,abgr
-    compute_rgba        d15, d14, d13, d12
+    compute_rgba_32     d15, d14, d13, d12
 .endif
 
 .ifc \ofmt,bgra
-    compute_rgba        d14, d13, d12, d15
+    compute_rgba_32     d14, d13, d12, d15
 .endif
 
     vst4.8              {q6, q7}, [\dst,:128]!
 .endm
 
-.macro declare_func ifmt ofmt
-function ff_\ifmt\()_to_\ofmt\()_neon, export=1
+.macro process_16px_16 ofmt
+    compute_premult_16  d28, d29, d30, d31
+
+    vld1.8              {q7}, [r4]!                                    @ first line of luma
+    compute_16px_16     r2, d14, d15, \ofmt
+
+    vld1.8              {q7}, [r12]!                                   @ second line of luma
+    compute_16px_16     r11, d14, d15, \ofmt
+.endm
+
+.macro process_16px_32 ofmt
+    compute_premult_32  d28, d30
+
+    vld1.8              {q7}, [r4]!                                    @ first line of luma
+    vmov                d28, d15                                       @ save right of the first line of luma for later use
+    compute_8px_32      r2, d14, \ofmt
+
+    vld1.8              {q7}, [r12]!                                   @ second line of luma
+    vmov                d30, d15                                       @ save right of the second line of luma for later use
+    compute_8px_32      r11, d14, \ofmt
+
+    compute_premult_32  d29, d31
+    compute_8px_32      r2,  d28, \ofmt
+    compute_8px_32      r11, d30, \ofmt
+.endm
+
+.macro load_args
     push                {r4-r12, lr}
     vpush               {q4-q7}
     ldr                 r4, [sp, #104]                                 @ r4  = srcY
@@ -102,6 +204,11 @@ function ff_\ifmt\()_to_\ofmt\()_neon, export=1
     sub                 r3, r3, r8                                     @ r3 = linesize  * 2 - width * 4 (padding)
     sub                 r5, r5, r0                                     @ r5 = linesizeY * 2 - width     (paddingY)
     sub                 r7, r7, r0                                     @ r7 = linesizeC     - width     (paddingC)
+.endm
+
+.macro declare_func ifmt ofmt precision
+function ff_\ifmt\()_to_\ofmt\()_neon_\precision\(), export=1
+    load_args
 1:
     mov                 r8, r0                                         @ r8 = width
 2:
@@ -119,19 +226,13 @@ function ff_\ifmt\()_to_\ofmt\()_neon, export=1
     vsubl.u8            q15, d2, d10                                   @ q15 = V - 128
 .endif
 
-    compute_premult     d28, d30
-
-    vld1.8              {q7}, [r4]!                                    @ first line of luma
-    vmov                d28, d15                                       @ save right of the first line of luma for later use
-    compute_half_line   r2, d14, \ofmt
-
-    vld1.8              {q7}, [r12]!                                   @ second line of luma
-    vmov                d30, d15                                       @ save right of the second line of luma for later use
-    compute_half_line   r11, d14, \ofmt
+.ifc \precision,16
+    process_16px_16     \ofmt
+.endif
 
-    compute_premult     d29, d31
-    compute_half_line   r2,  d28, \ofmt
-    compute_half_line   r11, d30, \ofmt
+.ifc \precision,32
+    process_16px_32     \ofmt
+.endif
 
     subs                r8, r8, #16                                    @ width -= 16
     bgt                 2b
@@ -151,12 +252,14 @@ function ff_\ifmt\()_to_\ofmt\()_neon, export=1
 endfunc
 .endm
 
-.macro declare_rgb_funcs ifmt
-    declare_func \ifmt, argb
-    declare_func \ifmt, rgba
-    declare_func \ifmt, abgr
-    declare_func \ifmt, bgra
+.macro declare_rgb_funcs ifmt precision
+    declare_func \ifmt, argb, \precision
+    declare_func \ifmt, rgba, \precision
+    declare_func \ifmt, abgr, \precision
+    declare_func \ifmt, bgra, \precision
 .endm
 
-declare_rgb_funcs nv12
-declare_rgb_funcs nv21
+declare_rgb_funcs nv12, 16
+declare_rgb_funcs nv21, 16
+declare_rgb_funcs nv12, 32
+declare_rgb_funcs nv21, 32