vp9[loongarch]: Optimize vpx_quantize_b/b_32x32

1. vpx_quantize_b_lsx
2. vpx_quantize_b_32x32_lsx

Bug: webm:1755

Change-Id: I476c8677a2c2aed7248e088e62c3777c9bed2adb

4 files changed, 262 insertions, 2 deletions

diff --git a/test/vp9_quantize_test.cc b/test/vp9_quantize_test.cc
index d54f1bc9c..ca1062a76 100644
--- a/test/vp9_quantize_test.cc
+++ b/test/vp9_quantize_test.cc
@@ -568,6 +568,16 @@ INSTANTIATE_TEST_SUITE_P(
                                  VPX_BITS_8, 32, true)));
 #endif  // HAVE_VSX && !CONFIG_VP9_HIGHBITDEPTH
 
+#if HAVE_LSX && !CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_SUITE_P(LSX, VP9QuantizeTest,
+                         ::testing::Values(make_tuple(&vpx_quantize_b_lsx,
+                                                      &vpx_quantize_b_c,
+                                                      VPX_BITS_8, 16, false),
+                                           make_tuple(&vpx_quantize_b_32x32_lsx,
+                                                      &vpx_quantize_b_32x32_c,
+                                                      VPX_BITS_8, 32, false)));
+#endif  // HAVE_LSX && !CONFIG_VP9_HIGHBITDEPTH
+
 // Only useful to compare "Speed" test results.
 INSTANTIATE_TEST_SUITE_P(
     DISABLED_C, VP9QuantizeTest,
diff --git a/vpx_dsp/loongarch/quantize_lsx.c b/vpx_dsp/loongarch/quantize_lsx.c
new file mode 100644
index 000000000..e3fbb9e9e
--- /dev/null
+++ b/vpx_dsp/loongarch/quantize_lsx.c
@@ -0,0 +1,249 @@
+/*
+ *  Copyright (c) 2022 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_util/loongson_intrinsics.h"
+
+#define CALCULATE_QCOEFF(coeff, coeff_abs, round, quant, shift, cmp_mask) \
+  ({                                                                      \
+    __m128i rounded, qcoeff;                                              \
+                                                                          \
+    rounded = __lsx_vsadd_h(coeff_abs, round);                            \
+    qcoeff = __lsx_vmuh_h(rounded, quant);                                \
+    qcoeff = __lsx_vadd_h(rounded, qcoeff);                               \
+    qcoeff = __lsx_vmuh_h(qcoeff, shift);                                 \
+    qcoeff = __lsx_vsigncov_h(coeff, qcoeff);                             \
+    qcoeff = __lsx_vand_v(qcoeff, cmp_mask);                              \
+                                                                          \
+    qcoeff;                                                               \
+  })
+
+#define CALCULATE_DQCOEFF_AND_STORE(qcoeff, dequant, dqcoeff) \
+  {                                                           \
+    __m128i dqcoeff16 = __lsx_vmul_h(qcoeff, dequant);        \
+    __lsx_vst(dqcoeff16, dqcoeff, 0);                         \
+  }
+
+#define CALCULATE_DQCOEFF_AND_STORE_32x32(qcoeff, dequant, dqcoeff) \
+  {                                                                 \
+    __m128i low, high, dqcoeff32_0, dqcoeff32_1, res;               \
+    __m128i zero = __lsx_vldi(0);                                   \
+    __m128i coeff = __lsx_vabsd_h(qcoeff, zero);                    \
+                                                                    \
+    __m128i sign_0 = __lsx_vilvl_h(qcoeff, zero);                   \
+    __m128i sign_1 = __lsx_vilvh_h(qcoeff, zero);                   \
+                                                                    \
+    low = __lsx_vmul_h(coeff, dequant);                             \
+    high = __lsx_vmuh_h(coeff, dequant);                            \
+    dqcoeff32_0 = __lsx_vilvl_h(high, low);                         \
+    dqcoeff32_1 = __lsx_vilvh_h(high, low);                         \
+                                                                    \
+    dqcoeff32_0 = __lsx_vsrai_w(dqcoeff32_0, 1);                    \
+    dqcoeff32_1 = __lsx_vsrai_w(dqcoeff32_1, 1);                    \
+    dqcoeff32_0 = __lsx_vsigncov_w(sign_0, dqcoeff32_0);            \
+    dqcoeff32_1 = __lsx_vsigncov_w(sign_1, dqcoeff32_1);            \
+    res = __lsx_vpickev_h(dqcoeff32_1, dqcoeff32_0);                \
+    __lsx_vst(res, dqcoeff, 0);                                     \
+  }
+
+#define SCAN_FOR_EOB(coeff0, coeff1, zbin_mask0, zbin_mask1, scan, index, \
+                     zero)                                                \
+  ({                                                                      \
+    __m128i zero_coeff0 = __lsx_vseq_h(coeff0, zero);                     \
+    __m128i zero_coeff1 = __lsx_vseq_h(coeff1, zero);                     \
+    __m128i scan0 = __lsx_vld(scan + index, 0);                           \
+    __m128i scan1 = __lsx_vld(scan + index + 8, 0);                       \
+    __m128i eob0, eob1, eob_max;                                          \
+                                                                          \
+    scan0 = __lsx_vsub_h(scan0, zbin_mask0);                              \
+    scan1 = __lsx_vsub_h(scan1, zbin_mask1);                              \
+    eob0 = __lsx_vandn_v(zero_coeff0, scan0);                             \
+    eob1 = __lsx_vandn_v(zero_coeff1, scan1);                             \
+    eob_max = __lsx_vmax_h(eob0, eob1);                                   \
+    eob_max;                                                              \
+  })
+
+#define ACCUMULATE_EOB(eob)                   \
+  ({                                          \
+    __m128i eob_shuffled;                     \
+    int16_t res_m;                            \
+                                              \
+    eob_shuffled = __lsx_vshuf4i_w(eob, 0xe); \
+    eob = __lsx_vmax_h(eob, eob_shuffled);    \
+    eob_shuffled = __lsx_vshuf4i_h(eob, 0xe); \
+    eob = __lsx_vmax_h(eob, eob_shuffled);    \
+    eob_shuffled = __lsx_vshuf4i_h(eob, 0x1); \
+    eob = __lsx_vmax_h(eob, eob_shuffled);    \
+    res_m = __lsx_vpickve2gr_h(eob, 1);       \
+    res_m;                                    \
+  })
+
+#if !CONFIG_VP9_HIGHBITDEPTH
+void vpx_quantize_b_lsx(const int16_t *coeff_ptr, intptr_t n_coeffs,
+                        const int16_t *zbin_ptr, const int16_t *round_ptr,
+                        const int16_t *quant_ptr,
+                        const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
+                        int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                        uint16_t *eob_ptr, const int16_t *scan,
+                        const int16_t *iscan) {
+  __m128i zero = __lsx_vldi(0);
+  int index = 16;
+
+  __m128i zbin, round, quant, dequant, quant_shift;
+  __m128i coeff0, coeff1;
+  __m128i qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1;
+  __m128i eob, eob0;
+
+  (void)scan;
+
+  zbin = __lsx_vld(zbin_ptr, 0);
+  round = __lsx_vld(round_ptr, 0);
+  quant = __lsx_vld(quant_ptr, 0);
+  dequant = __lsx_vld(dequant_ptr, 0);
+  quant_shift = __lsx_vld(quant_shift_ptr, 0);
+  // Handle one DC and first 15 AC.
+  DUP2_ARG2(__lsx_vld, coeff_ptr, 0, coeff_ptr, 16, coeff0, coeff1);
+  qcoeff0 = __lsx_vabsd_h(coeff0, zero);
+  qcoeff1 = __lsx_vabsd_h(coeff1, zero);
+
+  cmp_mask0 = __lsx_vsle_h(zbin, qcoeff0);
+  zbin = __lsx_vilvh_d(zbin, zbin);
+  cmp_mask1 = __lsx_vsle_h(zbin, qcoeff1);
+
+  qcoeff0 =
+      CALCULATE_QCOEFF(coeff0, qcoeff0, round, quant, quant_shift, cmp_mask0);
+  round = __lsx_vilvh_d(round, round);
+  quant = __lsx_vilvh_d(quant, quant);
+  quant_shift = __lsx_vilvh_d(quant_shift, quant_shift);
+  qcoeff1 =
+      CALCULATE_QCOEFF(coeff1, qcoeff1, round, quant, quant_shift, cmp_mask1);
+
+  __lsx_vst(qcoeff0, qcoeff_ptr, 0);
+  __lsx_vst(qcoeff1, qcoeff_ptr, 16);
+
+  CALCULATE_DQCOEFF_AND_STORE(qcoeff0, dequant, dqcoeff_ptr);
+  dequant = __lsx_vilvh_d(dequant, dequant);
+  CALCULATE_DQCOEFF_AND_STORE(qcoeff1, dequant, dqcoeff_ptr + 8);
+
+  eob = SCAN_FOR_EOB(qcoeff0, qcoeff1, cmp_mask0, cmp_mask1, iscan, 0, zero);
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = __lsx_vld(coeff_ptr + index, 0);
+    coeff1 = __lsx_vld(coeff_ptr + index + 8, 0);
+
+    qcoeff0 = __lsx_vabsd_h(coeff0, zero);
+    qcoeff1 = __lsx_vabsd_h(coeff1, zero);
+
+    cmp_mask0 = __lsx_vsle_h(zbin, qcoeff0);
+    cmp_mask1 = __lsx_vsle_h(zbin, qcoeff1);
+
+    qcoeff0 =
+        CALCULATE_QCOEFF(coeff0, qcoeff0, round, quant, quant_shift, cmp_mask0);
+    qcoeff1 =
+        CALCULATE_QCOEFF(coeff1, qcoeff1, round, quant, quant_shift, cmp_mask1);
+
+    __lsx_vst(qcoeff0, qcoeff_ptr + index, 0);
+    __lsx_vst(qcoeff1, qcoeff_ptr + index + 8, 0);
+
+    CALCULATE_DQCOEFF_AND_STORE(qcoeff0, dequant, dqcoeff_ptr + index);
+    CALCULATE_DQCOEFF_AND_STORE(qcoeff1, dequant, dqcoeff_ptr + index + 8);
+
+    eob0 = SCAN_FOR_EOB(qcoeff0, qcoeff1, cmp_mask0, cmp_mask1, iscan, index,
+                        zero);
+    eob = __lsx_vmax_h(eob, eob0);
+
+    index += 16;
+  }
+
+  *eob_ptr = ACCUMULATE_EOB(eob);
+}
+
+void vpx_quantize_b_32x32_lsx(const int16_t *coeff_ptr, intptr_t n_coeffs,
+                              const int16_t *zbin_ptr, const int16_t *round_ptr,
+                              const int16_t *quant_ptr,
+                              const int16_t *quant_shift_ptr,
+                              int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr,
+                              const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                              const int16_t *scan, const int16_t *iscan) {
+  __m128i zero = __lsx_vldi(0);
+  int index;
+
+  __m128i zbin, round, quant, dequant, quant_shift;
+  __m128i coeff0, coeff1, qcoeff0, qcoeff1, cmp_mask0, cmp_mask1;
+  __m128i eob = zero, eob0;
+
+  (void)scan;
+  (void)n_coeffs;
+
+  zbin = __lsx_vld(zbin_ptr, 0);
+  zbin = __lsx_vsrari_h(zbin, 1);
+  round = __lsx_vld(round_ptr, 0);
+  round = __lsx_vsrari_h(round, 1);
+
+  quant = __lsx_vld(quant_ptr, 0);
+  dequant = __lsx_vld(dequant_ptr, 0);
+  quant_shift = __lsx_vld(quant_shift_ptr, 0);
+  quant_shift = __lsx_vslli_h(quant_shift, 1);
+  // Handle one DC and first 15 AC.
+  DUP2_ARG2(__lsx_vld, coeff_ptr, 0, coeff_ptr, 16, coeff0, coeff1);
+  qcoeff0 = __lsx_vabsd_h(coeff0, zero);
+  qcoeff1 = __lsx_vabsd_h(coeff1, zero);
+
+  cmp_mask0 = __lsx_vsle_h(zbin, qcoeff0);
+  // remove DC from zbin
+  zbin = __lsx_vilvh_d(zbin, zbin);
+  cmp_mask1 = __lsx_vsle_h(zbin, qcoeff1);
+
+  qcoeff0 =
+      CALCULATE_QCOEFF(coeff0, qcoeff0, round, quant, quant_shift, cmp_mask0);
+  // remove DC in quant_shift, quant, quant_shift
+  round = __lsx_vilvh_d(round, round);
+  quant = __lsx_vilvh_d(quant, quant);
+  quant_shift = __lsx_vilvh_d(quant_shift, quant_shift);
+  qcoeff1 =
+      CALCULATE_QCOEFF(coeff1, qcoeff1, round, quant, quant_shift, cmp_mask1);
+  __lsx_vst(qcoeff0, qcoeff_ptr, 0);
+  __lsx_vst(qcoeff1, qcoeff_ptr, 16);
+
+  CALCULATE_DQCOEFF_AND_STORE_32x32(qcoeff0, dequant, dqcoeff_ptr);
+  dequant = __lsx_vilvh_d(dequant, dequant);
+  CALCULATE_DQCOEFF_AND_STORE_32x32(qcoeff1, dequant, dqcoeff_ptr + 8);
+  eob = SCAN_FOR_EOB(qcoeff0, qcoeff1, cmp_mask0, cmp_mask1, iscan, 0, zero);
+  // AC only loop.
+  for (index = 16; index < 32 * 32; index += 16) {
+    coeff0 = __lsx_vld(coeff_ptr + index, 0);
+    coeff1 = __lsx_vld(coeff_ptr + index + 8, 0);
+
+    qcoeff0 = __lsx_vabsd_h(coeff0, zero);
+    qcoeff1 = __lsx_vabsd_h(coeff1, zero);
+
+    cmp_mask0 = __lsx_vsle_h(zbin, qcoeff0);
+    cmp_mask1 = __lsx_vsle_h(zbin, qcoeff1);
+
+    qcoeff0 =
+        CALCULATE_QCOEFF(coeff0, qcoeff0, round, quant, quant_shift, cmp_mask0);
+    qcoeff1 =
+        CALCULATE_QCOEFF(coeff1, qcoeff1, round, quant, quant_shift, cmp_mask1);
+    __lsx_vst(qcoeff0, qcoeff_ptr + index, 0);
+    __lsx_vst(qcoeff1, qcoeff_ptr + index + 8, 0);
+
+    CALCULATE_DQCOEFF_AND_STORE_32x32(qcoeff0, dequant, dqcoeff_ptr + index);
+    CALCULATE_DQCOEFF_AND_STORE_32x32(qcoeff1, dequant,
+                                      dqcoeff_ptr + 8 + index);
+    eob0 = SCAN_FOR_EOB(qcoeff0, qcoeff1, cmp_mask0, cmp_mask1, iscan, index,
+                        zero);
+    eob = __lsx_vmax_h(eob, eob0);
+  }
+
+  *eob_ptr = ACCUMULATE_EOB(eob);
+}
+#endif  // !CONFIG_VP9_HIGHBITDEPTH
diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index 4f5a7a190..13999af04 100644
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -328,6 +328,7 @@ DSP_SRCS-$(HAVE_SSSE3)  += x86/quantize_ssse3.h
 DSP_SRCS-$(HAVE_AVX)    += x86/quantize_avx.c
 DSP_SRCS-$(HAVE_NEON)   += arm/quantize_neon.c
 DSP_SRCS-$(HAVE_VSX)    += ppc/quantize_vsx.c
+DSP_SRCS-$(HAVE_LSX)    += loongarch/quantize_lsx.c
 ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
 DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_quantize_intrin_sse2.c
 endif
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index e82b487f1..d3c668f9a 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -711,10 +711,10 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 #
 if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
   add_proto qw/void vpx_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
-  specialize qw/vpx_quantize_b neon sse2 ssse3 avx vsx/;
+  specialize qw/vpx_quantize_b neon sse2 ssse3 avx vsx lsx/;
 
   add_proto qw/void vpx_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
-  specialize qw/vpx_quantize_b_32x32 neon ssse3 avx vsx/;
+  specialize qw/vpx_quantize_b_32x32 neon ssse3 avx vsx lsx/;
 
   if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
     add_proto qw/void vpx_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";