diff options
Diffstat (limited to 'vpx_dsp')
| -rw-r--r-- | vpx_dsp/vpx_dsp.mk | 1 | ||||
| -rw-r--r-- | vpx_dsp/vpx_dsp_rtcd_defs.pl | 3 | ||||
| -rw-r--r-- | vpx_dsp/x86/sad4d_avx512.c | 83 |
3 files changed, 86 insertions, 1 deletions
diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk index 808ee36de..d18dd3107 100644 --- a/vpx_dsp/vpx_dsp.mk +++ b/vpx_dsp/vpx_dsp.mk @@ -327,6 +327,7 @@ DSP_SRCS-$(HAVE_SSSE3) += x86/sad_ssse3.asm DSP_SRCS-$(HAVE_SSE4_1) += x86/sad_sse4.asm DSP_SRCS-$(HAVE_AVX2) += x86/sad4d_avx2.c DSP_SRCS-$(HAVE_AVX2) += x86/sad_avx2.c +DSP_SRCS-$(HAVE_AVX512) += x86/sad4d_avx512.c DSP_SRCS-$(HAVE_SSE) += x86/sad4d_sse2.asm DSP_SRCS-$(HAVE_SSE) += x86/sad_sse2.asm diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl index bb54503fe..8ae847c3d 100644 --- a/vpx_dsp/vpx_dsp_rtcd_defs.pl +++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl @@ -20,6 +20,7 @@ if ($opts{arch} eq "x86_64") { $ssse3_x86_64 = 'ssse3'; $avx_x86_64 = 'avx'; $avx2_x86_64 = 'avx2'; + $avx512_x86_64 = 'avx512'; } # @@ -872,7 +873,7 @@ specialize qw/vpx_sad4x4x8 sse4_1 msa mmi/; # Multi-block SAD, comparing a reference to N independent blocks # add_proto qw/void vpx_sad64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array"; -specialize qw/vpx_sad64x64x4d avx2 neon msa sse2 vsx mmi/; +specialize qw/vpx_sad64x64x4d avx512 avx2 neon msa sse2 vsx mmi/; add_proto qw/void vpx_sad64x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array"; specialize qw/vpx_sad64x32x4d neon msa sse2 vsx mmi/; diff --git a/vpx_dsp/x86/sad4d_avx512.c b/vpx_dsp/x86/sad4d_avx512.c new file mode 100644 index 000000000..5f2ab6ea7 --- /dev/null +++ b/vpx_dsp/x86/sad4d_avx512.c @@ -0,0 +1,83 @@ +/* + * Copyright (c) 2017 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 <immintrin.h> // AVX512 +#include "./vpx_dsp_rtcd.h" +#include "vpx/vpx_integer.h" + +void vpx_sad64x64x4d_avx512(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + __m512i src_reg, ref0_reg, ref1_reg, ref2_reg, ref3_reg; + __m512i sum_ref0, sum_ref1, sum_ref2, sum_ref3; + __m512i sum_mlow, sum_mhigh; + int i; + const uint8_t *ref0, *ref1, *ref2, *ref3; + + ref0 = ref[0]; + ref1 = ref[1]; + ref2 = ref[2]; + ref3 = ref[3]; + sum_ref0 = _mm512_set1_epi16(0); + sum_ref1 = _mm512_set1_epi16(0); + sum_ref2 = _mm512_set1_epi16(0); + sum_ref3 = _mm512_set1_epi16(0); + for (i = 0; i < 64; i++) { + // load src and all refs + src_reg = _mm512_loadu_si512((const __m512i *)src); + ref0_reg = _mm512_loadu_si512((const __m512i *)ref0); + ref1_reg = _mm512_loadu_si512((const __m512i *)ref1); + ref2_reg = _mm512_loadu_si512((const __m512i *)ref2); + ref3_reg = _mm512_loadu_si512((const __m512i *)ref3); + // sum of the absolute differences between every ref-i to src + ref0_reg = _mm512_sad_epu8(ref0_reg, src_reg); + ref1_reg = _mm512_sad_epu8(ref1_reg, src_reg); + ref2_reg = _mm512_sad_epu8(ref2_reg, src_reg); + ref3_reg = _mm512_sad_epu8(ref3_reg, src_reg); + // sum every ref-i + sum_ref0 = _mm512_add_epi32(sum_ref0, ref0_reg); + sum_ref1 = _mm512_add_epi32(sum_ref1, ref1_reg); + sum_ref2 = _mm512_add_epi32(sum_ref2, ref2_reg); + sum_ref3 = _mm512_add_epi32(sum_ref3, ref3_reg); + + src += src_stride; + ref0 += ref_stride; + ref1 += ref_stride; + ref2 += ref_stride; + ref3 += ref_stride; + } + { + __m256i sum256; + __m128i sum128; + // in sum_ref-i the result is saved in the first 4 bytes + // the other 4 bytes are zeroed. + // sum_ref1 and sum_ref3 are shifted left by 4 bytes + sum_ref1 = _mm512_bslli_epi128(sum_ref1, 4); + sum_ref3 = _mm512_bslli_epi128(sum_ref3, 4); + + // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3 + sum_ref0 = _mm512_or_si512(sum_ref0, sum_ref1); + sum_ref2 = _mm512_or_si512(sum_ref2, sum_ref3); + + // merge every 64 bit from each sum_ref-i + sum_mlow = _mm512_unpacklo_epi64(sum_ref0, sum_ref2); + sum_mhigh = _mm512_unpackhi_epi64(sum_ref0, sum_ref2); + + // add the low 64 bit to the high 64 bit + sum_mlow = _mm512_add_epi32(sum_mlow, sum_mhigh); + + // add the low 128 bit to the high 128 bit + sum256 = _mm256_add_epi32(_mm512_castsi512_si256(sum_mlow), + _mm512_extracti32x8_epi32(sum_mlow, 1)); + sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum256), + _mm256_extractf128_si256(sum256, 1)); + + _mm_storeu_si128((__m128i *)(res), sum128); + } +} |