[x86]: Add vpx_highbd_subtract_block_avx2().

Up to 4x faster than "sse2 vectorized C".

Change-Id: Ie9b3c12a437c5cddf92c4d5349c4f659ca6b82ea

3 files changed, 117 insertions, 1 deletions

diff --git a/test/vp9_subtract_test.cc b/test/vp9_subtract_test.cc
index 7c69a317c..a57082f1e 100644
--- a/test/vp9_subtract_test.cc
+++ b/test/vp9_subtract_test.cc
@@ -308,5 +308,14 @@ INSTANTIATE_TEST_SUITE_P(
                        ::testing::Values(&vpx_highbd_subtract_block_c),
                        ::testing::Values(&vpx_highbd_subtract_block_c)));
 
+#if HAVE_AVX2
+INSTANTIATE_TEST_SUITE_P(
+    AVX2, VPXHBDSubtractBlockTest,
+    ::testing::Combine(::testing::ValuesIn(kValidBlockSize),
+                       ::testing::Values(12),
+                       ::testing::Values(&vpx_highbd_subtract_block_avx2),
+                       ::testing::Values(&vpx_highbd_subtract_block_c)));
+#endif  // HAVE_AVX2
+
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 }  // namespace vp9
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 51f5ebedd..b6d656820 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -939,7 +939,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   # Block subtraction
   #
   add_proto qw/void vpx_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src8_ptr, ptrdiff_t src_stride, const uint8_t *pred8_ptr, ptrdiff_t pred_stride, int bd";
-  specialize qw/vpx_highbd_subtract_block neon/;
+  specialize qw/vpx_highbd_subtract_block neon avx2/;
 
   #
   # Single block SAD
diff --git a/vpx_dsp/x86/subtract_avx2.c b/vpx_dsp/x86/subtract_avx2.c
index 4d259ef5c..4849581ed 100644
--- a/vpx_dsp/x86/subtract_avx2.c
+++ b/vpx_dsp/x86/subtract_avx2.c
@@ -94,3 +94,110 @@ void vpx_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr,
       break;
   }
 }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr,
+                                    ptrdiff_t diff_stride,
+                                    const uint8_t *src8_ptr,
+                                    ptrdiff_t src_stride,
+                                    const uint8_t *pred8_ptr,
+                                    ptrdiff_t pred_stride, int bd) {
+  uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8_ptr);
+  uint16_t *pred_ptr = CONVERT_TO_SHORTPTR(pred8_ptr);
+  (void)bd;
+  if (cols == 64) {
+    int j = rows;
+    do {
+      const __m256i s0 = _mm256_lddqu_si256((const __m256i *)src_ptr);
+      const __m256i s1 = _mm256_lddqu_si256((const __m256i *)(src_ptr + 16));
+      const __m256i s2 = _mm256_lddqu_si256((const __m256i *)(src_ptr + 32));
+      const __m256i s3 = _mm256_lddqu_si256((const __m256i *)(src_ptr + 48));
+      const __m256i p0 = _mm256_lddqu_si256((const __m256i *)pred_ptr);
+      const __m256i p1 = _mm256_lddqu_si256((const __m256i *)(pred_ptr + 16));
+      const __m256i p2 = _mm256_lddqu_si256((const __m256i *)(pred_ptr + 32));
+      const __m256i p3 = _mm256_lddqu_si256((const __m256i *)(pred_ptr + 48));
+      const __m256i d0 = _mm256_sub_epi16(s0, p0);
+      const __m256i d1 = _mm256_sub_epi16(s1, p1);
+      const __m256i d2 = _mm256_sub_epi16(s2, p2);
+      const __m256i d3 = _mm256_sub_epi16(s3, p3);
+      _mm256_storeu_si256((__m256i *)diff_ptr, d0);
+      _mm256_storeu_si256((__m256i *)(diff_ptr + 16), d1);
+      _mm256_storeu_si256((__m256i *)(diff_ptr + 32), d2);
+      _mm256_storeu_si256((__m256i *)(diff_ptr + 48), d3);
+      src_ptr += src_stride;
+      pred_ptr += pred_stride;
+      diff_ptr += diff_stride;
+    } while (--j != 0);
+  } else if (cols == 32) {
+    int j = rows;
+    do {
+      const __m256i s0 = _mm256_lddqu_si256((const __m256i *)src_ptr);
+      const __m256i s1 = _mm256_lddqu_si256((const __m256i *)(src_ptr + 16));
+      const __m256i p0 = _mm256_lddqu_si256((const __m256i *)pred_ptr);
+      const __m256i p1 = _mm256_lddqu_si256((const __m256i *)(pred_ptr + 16));
+      const __m256i d0 = _mm256_sub_epi16(s0, p0);
+      const __m256i d1 = _mm256_sub_epi16(s1, p1);
+      _mm256_storeu_si256((__m256i *)diff_ptr, d0);
+      _mm256_storeu_si256((__m256i *)(diff_ptr + 16), d1);
+      src_ptr += src_stride;
+      pred_ptr += pred_stride;
+      diff_ptr += diff_stride;
+    } while (--j != 0);
+  } else if (cols == 16) {
+    int j = rows;
+    do {
+      const __m256i s0 = _mm256_lddqu_si256((const __m256i *)src_ptr);
+      const __m256i s1 =
+          _mm256_lddqu_si256((const __m256i *)(src_ptr + src_stride));
+      const __m256i p0 = _mm256_lddqu_si256((const __m256i *)pred_ptr);
+      const __m256i p1 =
+          _mm256_lddqu_si256((const __m256i *)(pred_ptr + pred_stride));
+      const __m256i d0 = _mm256_sub_epi16(s0, p0);
+      const __m256i d1 = _mm256_sub_epi16(s1, p1);
+      _mm256_storeu_si256((__m256i *)diff_ptr, d0);
+      _mm256_storeu_si256((__m256i *)(diff_ptr + diff_stride), d1);
+      src_ptr += src_stride << 1;
+      pred_ptr += pred_stride << 1;
+      diff_ptr += diff_stride << 1;
+      j -= 2;
+    } while (j != 0);
+  } else if (cols == 8) {
+    int j = rows;
+    do {
+      const __m128i s0 = _mm_lddqu_si128((const __m128i *)src_ptr);
+      const __m128i s1 =
+          _mm_lddqu_si128((const __m128i *)(src_ptr + src_stride));
+      const __m128i p0 = _mm_lddqu_si128((const __m128i *)pred_ptr);
+      const __m128i p1 =
+          _mm_lddqu_si128((const __m128i *)(pred_ptr + pred_stride));
+      const __m128i d0 = _mm_sub_epi16(s0, p0);
+      const __m128i d1 = _mm_sub_epi16(s1, p1);
+      _mm_storeu_si128((__m128i *)diff_ptr, d0);
+      _mm_storeu_si128((__m128i *)(diff_ptr + diff_stride), d1);
+      src_ptr += src_stride << 1;
+      pred_ptr += pred_stride << 1;
+      diff_ptr += diff_stride << 1;
+      j -= 2;
+    } while (j != 0);
+  } else {
+    int j = rows;
+    assert(cols == 4);
+    do {
+      const __m128i s0 = _mm_loadl_epi64((const __m128i *)src_ptr);
+      const __m128i s1 =
+          _mm_loadl_epi64((const __m128i *)(src_ptr + src_stride));
+      const __m128i p0 = _mm_loadl_epi64((const __m128i *)pred_ptr);
+      const __m128i p1 =
+          _mm_loadl_epi64((const __m128i *)(pred_ptr + pred_stride));
+      const __m128i d0 = _mm_sub_epi16(s0, p0);
+      const __m128i d1 = _mm_sub_epi16(s1, p1);
+      _mm_storel_epi64((__m128i *)diff_ptr, d0);
+      _mm_storel_epi64((__m128i *)(diff_ptr + diff_stride), d1);
+      src_ptr += src_stride << 1;
+      pred_ptr += pred_stride << 1;
+      diff_ptr += diff_stride << 1;
+      j -= 2;
+    } while (j != 0);
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH