From 188d58eaa99ed5b47af77b354e303a82614ce62a Mon Sep 17 00:00:00 2001 From: Johann Date: Thu, 4 May 2017 08:39:12 -0700 Subject: [PATCH] sub pel variance neon: 4x block sizes Add optimizations for blocks of width 4 BUG=webm:1423 Change-Id: Idfb458d36db3014d48fbfbe7f5462aa6eb249938 --- test/variance_test.cc | 4 ++- vpx_dsp/arm/mem_neon.h | 26 ++++++++++++++ vpx_dsp/arm/subpel_variance_neon.c | 73 +++++++++++++++++++++++++++----------- vpx_dsp/vpx_dsp_rtcd_defs.pl | 4 +-- 4 files changed, 84 insertions(+), 23 deletions(-) diff --git a/test/variance_test.cc b/test/variance_test.cc index 206130b..ff9369d 100644 --- a/test/variance_test.cc +++ b/test/variance_test.cc @@ -1275,7 +1275,9 @@ INSTANTIATE_TEST_CASE_P( make_tuple(4, 3, &vpx_sub_pixel_variance16x8_neon, 0), make_tuple(3, 4, &vpx_sub_pixel_variance8x16_neon, 0), make_tuple(3, 3, &vpx_sub_pixel_variance8x8_neon, 0), - make_tuple(3, 2, &vpx_sub_pixel_variance8x4_neon, 0))); + make_tuple(3, 2, &vpx_sub_pixel_variance8x4_neon, 0), + make_tuple(2, 3, &vpx_sub_pixel_variance4x8_neon, 0), + make_tuple(2, 2, &vpx_sub_pixel_variance4x4_neon, 0))); INSTANTIATE_TEST_CASE_P( NEON, VpxSubpelAvgVarianceTest, diff --git a/vpx_dsp/arm/mem_neon.h b/vpx_dsp/arm/mem_neon.h index 37b89b2..4efad53 100644 --- a/vpx_dsp/arm/mem_neon.h +++ b/vpx_dsp/arm/mem_neon.h @@ -79,6 +79,32 @@ static INLINE void uint32_to_mem(uint8_t *buf, uint32_t a) { memcpy(buf, &a, 4); } +// Load 2 sets of 4 bytes when alignment is not guaranteed. +static INLINE uint8x8_t load_unaligned_u8(const uint8_t *buf, int stride) { + uint32_t a; + uint32x2_t a_u32 = vdup_n_u32(0); + if (stride == 4) return vld1_u8(buf); + memcpy(&a, buf, 4); + buf += stride; + a_u32 = vld1_lane_u32(&a, a_u32, 0); + memcpy(&a, buf, 4); + a_u32 = vld1_lane_u32(&a, a_u32, 1); + return vreinterpret_u8_u32(a_u32); +} + +// Store 2 sets of 4 bytes when alignment is not guaranteed. +static INLINE void store_unaligned_u8(uint8_t *buf, int stride, + const uint8x8_t a) { + const uint32x2_t a_u32 = vreinterpret_u32_u8(a); + if (stride == 4) { + vst1_u8(buf, a); + return; + } + uint32_to_mem(buf, vget_lane_u32(a_u32, 0)); + buf += stride; + uint32_to_mem(buf, vget_lane_u32(a_u32, 1)); +} + // Load 4 sets of 4 bytes when alignment is not guaranteed. static INLINE uint8x16_t load_unaligned_u8q(const uint8_t *buf, int stride) { uint32_t a; diff --git a/vpx_dsp/arm/subpel_variance_neon.c b/vpx_dsp/arm/subpel_variance_neon.c index 3089c36..00a83b1 100644 --- a/vpx_dsp/arm/subpel_variance_neon.c +++ b/vpx_dsp/arm/subpel_variance_neon.c @@ -16,12 +16,37 @@ #include "vpx/vpx_integer.h" #include "vpx_dsp/variance.h" +#include "vpx_dsp/arm/mem_neon.h" static const uint8_t bilinear_filters[8][2] = { { 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 }, { 64, 64 }, { 48, 80 }, { 32, 96 }, { 16, 112 }, }; +// Process a block exactly 4 wide and a multiple of 2 high. +static void var_filter_block2d_bil_w4(const uint8_t *src_ptr, + uint8_t *output_ptr, + unsigned int src_pixels_per_line, + int pixel_step, + unsigned int output_height, + const uint8_t *filter) { + const uint8x8_t f0 = vmov_n_u8(filter[0]); + const uint8x8_t f1 = vmov_n_u8(filter[1]); + unsigned int i; + for (i = 0; i < output_height; i += 2) { + const uint8x8_t src_0 = load_unaligned_u8(src_ptr, src_pixels_per_line); + const uint8x8_t src_1 = + load_unaligned_u8(src_ptr + pixel_step, src_pixels_per_line); + const uint16x8_t a = vmull_u8(src_0, f0); + const uint16x8_t b = vmlal_u8(a, src_1, f1); + const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS); + store_unaligned_u8(output_ptr, 4, out); + // Next row... + src_ptr += 2 * src_pixels_per_line; + output_ptr += 8; + } +} + // Process a block exactly 8 wide and any height. static void var_filter_block2d_bil_w8(const uint8_t *src_ptr, uint8_t *output_ptr, @@ -74,28 +99,36 @@ static void var_filter_block2d_bil_w16(const uint8_t *src_ptr, } } -// TODO(johannkoenig): support 4xM block sizes. -#define sub_pixel_varianceNxM(n, m) \ - uint32_t vpx_sub_pixel_variance##n##x##m##_neon( \ - const uint8_t *a, int a_stride, int xoffset, int yoffset, \ - const uint8_t *b, int b_stride, uint32_t *sse) { \ - DECLARE_ALIGNED(16, uint8_t, fdata3[n * (m + 1)]); \ - DECLARE_ALIGNED(16, uint8_t, temp2[n * m]); \ - \ - if (n == 8) { \ - var_filter_block2d_bil_w8(a, fdata3, a_stride, 1, (m + 1), \ - bilinear_filters[xoffset]); \ - var_filter_block2d_bil_w8(fdata3, temp2, n, n, m, \ - bilinear_filters[yoffset]); \ - } else { \ - var_filter_block2d_bil_w16(a, fdata3, a_stride, 1, (m + 1), n, \ - bilinear_filters[xoffset]); \ - var_filter_block2d_bil_w16(fdata3, temp2, n, n, m, n, \ - bilinear_filters[yoffset]); \ - } \ - return vpx_variance##n##x##m(temp2, n, b, b_stride, sse); \ +// 4xM filter writes an extra row to fdata because it processes two rows at a +// time. +#define sub_pixel_varianceNxM(n, m) \ + uint32_t vpx_sub_pixel_variance##n##x##m##_neon( \ + const uint8_t *a, int a_stride, int xoffset, int yoffset, \ + const uint8_t *b, int b_stride, uint32_t *sse) { \ + DECLARE_ALIGNED(16, uint8_t, fdata3[n * (m + (n == 4 ? 2 : 1))]); \ + DECLARE_ALIGNED(16, uint8_t, temp2[n * m]); \ + \ + if (n == 4) { \ + var_filter_block2d_bil_w4(a, fdata3, a_stride, 1, (m + 2), \ + bilinear_filters[xoffset]); \ + var_filter_block2d_bil_w4(fdata3, temp2, n, n, m, \ + bilinear_filters[yoffset]); \ + } else if (n == 8) { \ + var_filter_block2d_bil_w8(a, fdata3, a_stride, 1, (m + 1), \ + bilinear_filters[xoffset]); \ + var_filter_block2d_bil_w8(fdata3, temp2, n, n, m, \ + bilinear_filters[yoffset]); \ + } else { \ + var_filter_block2d_bil_w16(a, fdata3, a_stride, 1, (m + 1), n, \ + bilinear_filters[xoffset]); \ + var_filter_block2d_bil_w16(fdata3, temp2, n, n, m, n, \ + bilinear_filters[yoffset]); \ + } \ + return vpx_variance##n##x##m(temp2, n, b, b_stride, sse); \ } +sub_pixel_varianceNxM(4, 4); +sub_pixel_varianceNxM(4, 8); sub_pixel_varianceNxM(8, 4); sub_pixel_varianceNxM(8, 8); sub_pixel_varianceNxM(8, 16); diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl index 09c697f..38424da 100644 --- a/vpx_dsp/vpx_dsp_rtcd_defs.pl +++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl @@ -1214,10 +1214,10 @@ add_proto qw/uint32_t vpx_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int s specialize qw/vpx_sub_pixel_variance8x4 neon msa sse2 ssse3/; add_proto qw/uint32_t vpx_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; - specialize qw/vpx_sub_pixel_variance4x8 msa sse2 ssse3/; + specialize qw/vpx_sub_pixel_variance4x8 neon msa sse2 ssse3/; add_proto qw/uint32_t vpx_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse"; - specialize qw/vpx_sub_pixel_variance4x4 msa sse2 ssse3/; + specialize qw/vpx_sub_pixel_variance4x4 neon msa sse2 ssse3/; add_proto qw/uint32_t vpx_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; specialize qw/vpx_sub_pixel_avg_variance64x64 neon avx2 msa sse2 ssse3/; -- 2.7.4