VPX: Add vpx_subtract_block_avx2().

~1.3x faster than vpx_subtract_block_sse2().

Based on aom_subtract_block_avx2().

Bug: b/241580104

Change-Id: I17da036363f213d53c6546c3e858e4c3cba44a5b

diff --git a/test/vp9_subtract_test.cc b/test/vp9_subtract_test.cc
index 211cc6c7adfdd36dfaf785f042aa1b5675292104..f634a032d8dc034a581e8c191184a50945dc0997 100644 (file)
--- a/test/vp9_subtract_test.cc
+++ b/test/vp9_subtract_test.cc
@@ -133,6 +133,10 @@ INSTANTIATE_TEST_SUITE_P(C, VP9SubtractBlockTest,
 INSTANTIATE_TEST_SUITE_P(SSE2, VP9SubtractBlockTest,
                          ::testing::Values(vpx_subtract_block_sse2));
 #endif
+#if HAVE_AVX2
+INSTANTIATE_TEST_SUITE_P(AVX2, VP9SubtractBlockTest,
+                         ::testing::Values(vpx_subtract_block_avx2));
+#endif
 #if HAVE_NEON
 INSTANTIATE_TEST_SUITE_P(NEON, VP9SubtractBlockTest,
                          ::testing::Values(vpx_subtract_block_neon));

diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index dd667195f587ea2d90daf81ac4e4f71df2c192ab..ffe954832daeb90d19e81b4d90a5b8a4db50af2d 100644 (file)
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -376,6 +376,7 @@ DSP_SRCS-$(HAVE_MMI)    += mips/subtract_mmi.c
 
 DSP_SRCS-$(HAVE_AVX2)   += x86/sad4d_avx2.c
 DSP_SRCS-$(HAVE_AVX2)   += x86/sad_avx2.c
+DSP_SRCS-$(HAVE_AVX2)   += x86/subtract_avx2.c
 DSP_SRCS-$(HAVE_AVX512) += x86/sad4d_avx512.c
 
 DSP_SRCS-$(HAVE_SSE2)   += x86/sad4d_sse2.asm

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index e7ad640af8bb7afcc8fc93c72e95c2a6a285684c..db211ed8ce066649aee64a2c5b50ca1630bf25f2 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -730,7 +730,7 @@ if (vpx_config("CONFIG_ENCODERS") eq "yes") {
 # Block subtraction
 #
 add_proto qw/void vpx_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
-specialize qw/vpx_subtract_block neon msa mmi sse2 vsx lsx/;
+specialize qw/vpx_subtract_block neon msa mmi sse2 avx2 vsx lsx/;
 
 #
 # Single block SAD

diff --git a/vpx_dsp/x86/subtract_avx2.c b/vpx_dsp/x86/subtract_avx2.c
new file mode 100644 (file)
index 0000000..4d259ef
--- /dev/null
+++ b/vpx_dsp/x86/subtract_avx2.c
@@ -0,0 +1,96 @@
+/*
+ *  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 <assert.h>
+#include <immintrin.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+static VPX_FORCE_INLINE void subtract32_avx2(int16_t *diff_ptr,
+                                             const uint8_t *src_ptr,
+                                             const uint8_t *pred_ptr) {
+  const __m256i s = _mm256_lddqu_si256((const __m256i *)src_ptr);
+  const __m256i p = _mm256_lddqu_si256((const __m256i *)pred_ptr);
+  const __m256i s_0 = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(s));
+  const __m256i s_1 = _mm256_cvtepu8_epi16(_mm256_extracti128_si256(s, 1));
+  const __m256i p_0 = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(p));
+  const __m256i p_1 = _mm256_cvtepu8_epi16(_mm256_extracti128_si256(p, 1));
+  const __m256i d_0 = _mm256_sub_epi16(s_0, p_0);
+  const __m256i d_1 = _mm256_sub_epi16(s_1, p_1);
+  _mm256_storeu_si256((__m256i *)diff_ptr, d_0);
+  _mm256_storeu_si256((__m256i *)(diff_ptr + 16), d_1);
+}
+
+static VPX_FORCE_INLINE void subtract_block_16xn_avx2(
+    int rows, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr,
+    ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
+  int j;
+  for (j = 0; j < rows; ++j) {
+    const __m128i s = _mm_lddqu_si128((const __m128i *)src_ptr);
+    const __m128i p = _mm_lddqu_si128((const __m128i *)pred_ptr);
+    const __m256i s_0 = _mm256_cvtepu8_epi16(s);
+    const __m256i p_0 = _mm256_cvtepu8_epi16(p);
+    const __m256i d_0 = _mm256_sub_epi16(s_0, p_0);
+    _mm256_storeu_si256((__m256i *)diff_ptr, d_0);
+    src_ptr += src_stride;
+    pred_ptr += pred_stride;
+    diff_ptr += diff_stride;
+  }
+}
+
+static VPX_FORCE_INLINE void subtract_block_32xn_avx2(
+    int rows, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr,
+    ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
+  int j;
+  for (j = 0; j < rows; ++j) {
+    subtract32_avx2(diff_ptr, src_ptr, pred_ptr);
+    src_ptr += src_stride;
+    pred_ptr += pred_stride;
+    diff_ptr += diff_stride;
+  }
+}
+
+static VPX_FORCE_INLINE void subtract_block_64xn_avx2(
+    int rows, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr,
+    ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
+  int j;
+  for (j = 0; j < rows; ++j) {
+    subtract32_avx2(diff_ptr, src_ptr, pred_ptr);
+    subtract32_avx2(diff_ptr + 32, src_ptr + 32, pred_ptr + 32);
+    src_ptr += src_stride;
+    pred_ptr += pred_stride;
+    diff_ptr += diff_stride;
+  }
+}
+
+void vpx_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr,
+                             ptrdiff_t diff_stride, const uint8_t *src_ptr,
+                             ptrdiff_t src_stride, const uint8_t *pred_ptr,
+                             ptrdiff_t pred_stride) {
+  switch (cols) {
+    case 16:
+      subtract_block_16xn_avx2(rows, diff_ptr, diff_stride, src_ptr, src_stride,
+                               pred_ptr, pred_stride);
+      break;
+    case 32:
+      subtract_block_32xn_avx2(rows, diff_ptr, diff_stride, src_ptr, src_stride,
+                               pred_ptr, pred_stride);
+      break;
+    case 64:
+      subtract_block_64xn_avx2(rows, diff_ptr, diff_stride, src_ptr, src_stride,
+                               pred_ptr, pred_stride);
+      break;
+    default:
+      vpx_subtract_block_sse2(rows, cols, diff_ptr, diff_stride, src_ptr,
+                              src_stride, pred_ptr, pred_stride);
+      break;
+  }
+}