From 46594be7588d21cbe7a5620ff1ecee269e46c498 Mon Sep 17 00:00:00 2001 From: Frank Barchard Date: Wed, 15 Nov 2017 17:07:12 -0800 Subject: [PATCH] add ScalePlane_16 unit tests Tests ScalePlane vs ScalePlane_16 match. Bug: libyuv:749 Test: LibYUVScaleTest.ScalePlaneDownBy4_Box_16 Change-Id: I3f71748da404982d5d48bfb11bbd3ae95a1d021c Reviewed-on: https://chromium-review.googlesource.com/765045 Reviewed-by: Frank Barchard Reviewed-by: richard winterton Reviewed-by: Weiyong Yao --- README.chromium | 2 +- include/libyuv/row.h | 5 +- include/libyuv/version.h | 2 +- source/row_gcc.cc | 1 - unit_test/scale_test.cc | 284 +++++++++++++++++++++++++-------------- 5 files changed, 183 insertions(+), 111 deletions(-) diff --git a/README.chromium b/README.chromium index be05cba3a..7c42013b9 100644 --- a/README.chromium +++ b/README.chromium @@ -1,6 +1,6 @@ Name: libyuv URL: http://code.google.com/p/libyuv/ -Version: 1676 +Version: 1677 License: BSD License File: LICENSE diff --git a/include/libyuv/row.h b/include/libyuv/row.h index fc5caba4c..34d727641 100644 --- a/include/libyuv/row.h +++ b/include/libyuv/row.h @@ -1537,10 +1537,7 @@ void MultiplyRow_16_AVX2(const uint16* src_y, uint16* dst_y, int scale, int width); -void MultiplyRow_16_C(const uint16* src_y, - uint16* dst_y, - int scale, - int width); +void MultiplyRow_16_C(const uint16* src_y, uint16* dst_y, int scale, int width); void CopyRow_SSE2(const uint8* src, uint8* dst, int count); void CopyRow_AVX(const uint8* src, uint8* dst, int count); diff --git a/include/libyuv/version.h b/include/libyuv/version.h index a8d5f6d5f..1b6fd773b 100644 --- a/include/libyuv/version.h +++ b/include/libyuv/version.h @@ -11,6 +11,6 @@ #ifndef INCLUDE_LIBYUV_VERSION_H_ #define INCLUDE_LIBYUV_VERSION_H_ -#define LIBYUV_VERSION 1676 +#define LIBYUV_VERSION 1677 #endif // INCLUDE_LIBYUV_VERSION_H_ diff --git a/source/row_gcc.cc b/source/row_gcc.cc index ca220a22b..b5c2e65c9 100644 --- a/source/row_gcc.cc +++ b/source/row_gcc.cc @@ -2800,7 +2800,6 @@ void MergeUVRow_16_AVX2(const uint16* src_u, } #endif // HAS_MERGEUVROW_AVX2 - #ifdef HAS_MULTIPLYROW_16_AVX2 void MultiplyRow_16_AVX2(const uint16* src_y, uint16* dst_y, diff --git a/unit_test/scale_test.cc b/unit_test/scale_test.cc index a9d061069..c72000ce8 100644 --- a/unit_test/scale_test.cc +++ b/unit_test/scale_test.cc @@ -35,15 +35,14 @@ static int TestFilter(int src_width, } int i, j; - const int b = 0; // 128 to test for padding/stride. int src_width_uv = (Abs(src_width) + 1) >> 1; int src_height_uv = (Abs(src_height) + 1) >> 1; - int64 src_y_plane_size = (Abs(src_width) + b * 2) * (Abs(src_height) + b * 2); - int64 src_uv_plane_size = (src_width_uv + b * 2) * (src_height_uv + b * 2); + int64 src_y_plane_size = (Abs(src_width)) * (Abs(src_height)); + int64 src_uv_plane_size = (src_width_uv) * (src_height_uv); - int src_stride_y = b * 2 + Abs(src_width); - int src_stride_uv = b * 2 + src_width_uv; + int src_stride_y = Abs(src_width); + int src_stride_uv = src_width_uv; align_buffer_page_end(src_y, src_y_plane_size); align_buffer_page_end(src_u, src_uv_plane_size); @@ -59,11 +58,11 @@ static int TestFilter(int src_width, int dst_width_uv = (dst_width + 1) >> 1; int dst_height_uv = (dst_height + 1) >> 1; - int64 dst_y_plane_size = (dst_width + b * 2) * (dst_height + b * 2); - int64 dst_uv_plane_size = (dst_width_uv + b * 2) * (dst_height_uv + b * 2); + int64 dst_y_plane_size = (dst_width) * (dst_height); + int64 dst_uv_plane_size = (dst_width_uv) * (dst_height_uv); - int dst_stride_y = b * 2 + dst_width; - int dst_stride_uv = b * 2 + dst_width_uv; + int dst_stride_y = dst_width; + int dst_stride_uv = dst_width_uv; align_buffer_page_end(dst_y_c, dst_y_plane_size); align_buffer_page_end(dst_u_c, dst_uv_plane_size); @@ -79,38 +78,31 @@ static int TestFilter(int src_width, MaskCpuFlags(disable_cpu_flags); // Disable all CPU optimization. double c_time = get_time(); - I420Scale(src_y + (src_stride_y * b) + b, src_stride_y, - src_u + (src_stride_uv * b) + b, src_stride_uv, - src_v + (src_stride_uv * b) + b, src_stride_uv, src_width, - src_height, dst_y_c + (dst_stride_y * b) + b, dst_stride_y, - dst_u_c + (dst_stride_uv * b) + b, dst_stride_uv, - dst_v_c + (dst_stride_uv * b) + b, dst_stride_uv, dst_width, - dst_height, f); + I420Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv, + src_width, src_height, dst_y_c, dst_stride_y, dst_u_c, + dst_stride_uv, dst_v_c, dst_stride_uv, dst_width, dst_height, f); c_time = (get_time() - c_time); MaskCpuFlags(benchmark_cpu_info); // Enable all CPU optimization. double opt_time = get_time(); for (i = 0; i < benchmark_iterations; ++i) { - I420Scale(src_y + (src_stride_y * b) + b, src_stride_y, - src_u + (src_stride_uv * b) + b, src_stride_uv, - src_v + (src_stride_uv * b) + b, src_stride_uv, src_width, - src_height, dst_y_opt + (dst_stride_y * b) + b, dst_stride_y, - dst_u_opt + (dst_stride_uv * b) + b, dst_stride_uv, - dst_v_opt + (dst_stride_uv * b) + b, dst_stride_uv, dst_width, - dst_height, f); + I420Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv, + src_width, src_height, dst_y_opt, dst_stride_y, dst_u_opt, + dst_stride_uv, dst_v_opt, dst_stride_uv, dst_width, dst_height, + f); } opt_time = (get_time() - opt_time) / benchmark_iterations; - // Report performance of C vs OPT + // Report performance of C vs OPT. printf("filter %d - %8d us C - %8d us OPT\n", f, static_cast(c_time * 1e6), static_cast(opt_time * 1e6)); // C version may be a little off from the optimized. Order of // operations may introduce rounding somewhere. So do a difference - // of the buffers and look to see that the max difference isn't - // over 2. + // of the buffers and look to see that the max difference is not + // over 3. int max_diff = 0; - for (i = b; i < (dst_height + b); ++i) { - for (j = b; j < (dst_width + b); ++j) { + for (i = 0; i < (dst_height); ++i) { + for (j = 0; j < (dst_width); ++j) { int abs_diff = Abs(dst_y_c[(i * dst_stride_y) + j] - dst_y_opt[(i * dst_stride_y) + j]); if (abs_diff > max_diff) { @@ -119,8 +111,8 @@ static int TestFilter(int src_width, } } - for (i = b; i < (dst_height_uv + b); ++i) { - for (j = b; j < (dst_width_uv + b); ++j) { + for (i = 0; i < (dst_height_uv); ++i) { + for (j = 0; j < (dst_width_uv); ++j) { int abs_diff = Abs(dst_u_c[(i * dst_stride_uv) + j] - dst_u_opt[(i * dst_stride_uv) + j]); if (abs_diff > max_diff) { @@ -154,21 +146,22 @@ static int TestFilter_16(int src_width, int dst_width, int dst_height, FilterMode f, - int benchmark_iterations) { + int benchmark_iterations, + int disable_cpu_flags, + int benchmark_cpu_info) { if (!SizeValid(src_width, src_height, dst_width, dst_height)) { return 0; } int i, j; - const int b = 0; // 128 to test for padding/stride. int src_width_uv = (Abs(src_width) + 1) >> 1; int src_height_uv = (Abs(src_height) + 1) >> 1; - int64 src_y_plane_size = (Abs(src_width) + b * 2) * (Abs(src_height) + b * 2); - int64 src_uv_plane_size = (src_width_uv + b * 2) * (src_height_uv + b * 2); + int64 src_y_plane_size = (Abs(src_width)) * (Abs(src_height)); + int64 src_uv_plane_size = (src_width_uv) * (src_height_uv); - int src_stride_y = b * 2 + Abs(src_width); - int src_stride_uv = b * 2 + src_width_uv; + int src_stride_y = Abs(src_width); + int src_stride_uv = src_width_uv; align_buffer_page_end(src_y, src_y_plane_size); align_buffer_page_end(src_u, src_uv_plane_size); @@ -176,6 +169,10 @@ static int TestFilter_16(int src_width, align_buffer_page_end(src_y_16, src_y_plane_size * 2); align_buffer_page_end(src_u_16, src_uv_plane_size * 2); align_buffer_page_end(src_v_16, src_uv_plane_size * 2); + if (!src_y || !src_u || !src_v || !src_y_16 || !src_u_16 || !src_v_16) { + printf("Skipped. Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n"); + return 0; + } uint16* p_src_y_16 = reinterpret_cast(src_y_16); uint16* p_src_u_16 = reinterpret_cast(src_u_16); uint16* p_src_v_16 = reinterpret_cast(src_v_16); @@ -184,27 +181,22 @@ static int TestFilter_16(int src_width, MemRandomize(src_u, src_uv_plane_size); MemRandomize(src_v, src_uv_plane_size); - for (i = b; i < src_height + b; ++i) { - for (j = b; j < src_width + b; ++j) { - p_src_y_16[(i * src_stride_y) + j] = src_y[(i * src_stride_y) + j]; - } + for (i = 0; i < src_y_plane_size; ++i) { + p_src_y_16[i] = src_y[i]; } - - for (i = b; i < (src_height_uv + b); ++i) { - for (j = b; j < (src_width_uv + b); ++j) { - p_src_u_16[(i * src_stride_uv) + j] = src_u[(i * src_stride_uv) + j]; - p_src_v_16[(i * src_stride_uv) + j] = src_v[(i * src_stride_uv) + j]; - } + for (i = 0; i < src_uv_plane_size; ++i) { + p_src_u_16[i] = src_u[i]; + p_src_v_16[i] = src_v[i]; } int dst_width_uv = (dst_width + 1) >> 1; int dst_height_uv = (dst_height + 1) >> 1; - int dst_y_plane_size = (dst_width + b * 2) * (dst_height + b * 2); - int dst_uv_plane_size = (dst_width_uv + b * 2) * (dst_height_uv + b * 2); + int dst_y_plane_size = (dst_width) * (dst_height); + int dst_uv_plane_size = (dst_width_uv) * (dst_height_uv); - int dst_stride_y = b * 2 + dst_width; - int dst_stride_uv = b * 2 + dst_width_uv; + int dst_stride_y = dst_width; + int dst_stride_uv = dst_width_uv; align_buffer_page_end(dst_y_8, dst_y_plane_size); align_buffer_page_end(dst_u_8, dst_uv_plane_size); @@ -217,48 +209,34 @@ static int TestFilter_16(int src_width, uint16* p_dst_u_16 = reinterpret_cast(dst_u_16); uint16* p_dst_v_16 = reinterpret_cast(dst_v_16); - I420Scale(src_y + (src_stride_y * b) + b, src_stride_y, - src_u + (src_stride_uv * b) + b, src_stride_uv, - src_v + (src_stride_uv * b) + b, src_stride_uv, src_width, - src_height, dst_y_8 + (dst_stride_y * b) + b, dst_stride_y, - dst_u_8 + (dst_stride_uv * b) + b, dst_stride_uv, - dst_v_8 + (dst_stride_uv * b) + b, dst_stride_uv, dst_width, - dst_height, f); - + MaskCpuFlags(disable_cpu_flags); // Disable all CPU optimization. + I420Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv, + src_width, src_height, dst_y_8, dst_stride_y, dst_u_8, + dst_stride_uv, dst_v_8, dst_stride_uv, dst_width, dst_height, f); + MaskCpuFlags(benchmark_cpu_info); // Enable all CPU optimization. for (i = 0; i < benchmark_iterations; ++i) { - I420Scale_16(p_src_y_16 + (src_stride_y * b) + b, src_stride_y, - p_src_u_16 + (src_stride_uv * b) + b, src_stride_uv, - p_src_v_16 + (src_stride_uv * b) + b, src_stride_uv, src_width, - src_height, p_dst_y_16 + (dst_stride_y * b) + b, dst_stride_y, - p_dst_u_16 + (dst_stride_uv * b) + b, dst_stride_uv, - p_dst_v_16 + (dst_stride_uv * b) + b, dst_stride_uv, dst_width, - dst_height, f); + I420Scale_16(p_src_y_16, src_stride_y, p_src_u_16, src_stride_uv, + p_src_v_16, src_stride_uv, src_width, src_height, p_dst_y_16, + dst_stride_y, p_dst_u_16, dst_stride_uv, p_dst_v_16, + dst_stride_uv, dst_width, dst_height, f); } - // Expect an exact match + // Expect an exact match. int max_diff = 0; - for (i = b; i < (dst_height + b); ++i) { - for (j = b; j < (dst_width + b); ++j) { - int abs_diff = Abs(dst_y_8[(i * dst_stride_y) + j] - - p_dst_y_16[(i * dst_stride_y) + j]); - if (abs_diff > max_diff) { - max_diff = abs_diff; - } + for (i = 0; i < dst_y_plane_size; ++i) { + int abs_diff = Abs(dst_y_8[i] - p_dst_y_16[i]); + if (abs_diff > max_diff) { + max_diff = abs_diff; } } - - for (i = b; i < (dst_height_uv + b); ++i) { - for (j = b; j < (dst_width_uv + b); ++j) { - int abs_diff = Abs(dst_u_8[(i * dst_stride_uv) + j] - - p_dst_u_16[(i * dst_stride_uv) + j]); - if (abs_diff > max_diff) { - max_diff = abs_diff; - } - abs_diff = Abs(dst_v_8[(i * dst_stride_uv) + j] - - p_dst_v_16[(i * dst_stride_uv) + j]); - if (abs_diff > max_diff) { - max_diff = abs_diff; - } + for (i = 0; i < dst_uv_plane_size; ++i) { + int abs_diff = Abs(dst_u_8[i] - p_dst_u_16[i]); + if (abs_diff > max_diff) { + max_diff = abs_diff; + } + abs_diff = Abs(dst_v_8[i] - p_dst_v_16[i]); + if (abs_diff > max_diff) { + max_diff = abs_diff; } } @@ -280,7 +258,7 @@ static int TestFilter_16(int src_width, // The following adjustments in dimensions ensure the scale factor will be // exactly achieved. -// 2 is chroma subsample +// 2 is chroma subsample. #define DX(x, nom, denom) static_cast(((Abs(x) / nom + 1) / 2) * nom * 2) #define SX(x, nom, denom) static_cast(((x / nom + 1) / 2) * denom * 2) @@ -293,11 +271,12 @@ static int TestFilter_16(int src_width, benchmark_cpu_info_); \ EXPECT_LE(diff, max_diff); \ } \ - TEST_F(LibYUVScaleTest, DISABLED_ScaleDownBy##name##_##filter##_16) { \ + TEST_F(LibYUVScaleTest, ScaleDownBy##name##_##filter##_16) { \ int diff = TestFilter_16( \ SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \ DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \ - kFilter##filter, benchmark_iterations_); \ + kFilter##filter, benchmark_iterations_, disable_cpu_flags_, \ + benchmark_cpu_info_); \ EXPECT_LE(diff, max_diff); \ } @@ -334,26 +313,26 @@ TEST_FACTOR(3, 1, 3, 0) benchmark_cpu_info_); \ EXPECT_LE(diff, max_diff); \ } \ - TEST_F(LibYUVScaleTest, \ - DISABLED_##name##To##width##x##height##_##filter##_16) { \ + TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter##_16) { \ int diff = TestFilter_16(benchmark_width_, benchmark_height_, width, \ - height, kFilter##filter, benchmark_iterations_); \ + height, kFilter##filter, benchmark_iterations_, \ + disable_cpu_flags_, benchmark_cpu_info_); \ EXPECT_LE(diff, max_diff); \ } \ - TEST_F(LibYUVScaleTest, \ - DISABLED_##name##From##width##x##height##_##filter##_16) { \ + TEST_F(LibYUVScaleTest, name##From##width##x##height##_##filter##_16) { \ int diff = TestFilter_16(width, height, Abs(benchmark_width_), \ Abs(benchmark_height_), kFilter##filter, \ - benchmark_iterations_); \ + benchmark_iterations_, disable_cpu_flags_, \ + benchmark_cpu_info_); \ EXPECT_LE(diff, max_diff); \ } // Test scale to a specified size with all 4 filters. #define TEST_SCALETO(name, width, height) \ TEST_SCALETO1(name, width, height, None, 0) \ - TEST_SCALETO1(name, width, height, Linear, 0) \ - TEST_SCALETO1(name, width, height, Bilinear, 0) \ - TEST_SCALETO1(name, width, height, Box, 0) + TEST_SCALETO1(name, width, height, Linear, 3) \ + TEST_SCALETO1(name, width, height, Bilinear, 3) \ + TEST_SCALETO1(name, width, height, Box, 3) TEST_SCALETO(Scale, 1, 1) TEST_SCALETO(Scale, 320, 240) @@ -377,27 +356,27 @@ TEST_F(LibYUVScaleTest, TestScaleRowDown2Box_Odd_SSSE3) { if (!has_ssse3) { printf("Warning SSSE3 not detected; Skipping test.\n"); } else { - // TL + // TL. orig_pixels[0] = 255u; orig_pixels[1] = 0u; orig_pixels[128 + 0] = 0u; orig_pixels[128 + 1] = 0u; - // TR + // TR. orig_pixels[2] = 0u; orig_pixels[3] = 100u; orig_pixels[128 + 2] = 0u; orig_pixels[128 + 3] = 0u; - // BL + // BL. orig_pixels[4] = 0u; orig_pixels[5] = 0u; orig_pixels[128 + 4] = 50u; orig_pixels[128 + 5] = 0u; - // BR + // BR. orig_pixels[6] = 0u; orig_pixels[7] = 0u; orig_pixels[128 + 6] = 0u; orig_pixels[128 + 7] = 20u; - // Odd + // Odd. orig_pixels[126] = 4u; orig_pixels[127] = 255u; orig_pixels[128 + 126] = 16u; @@ -464,7 +443,7 @@ extern "C" void ScaleRowUp2_16_C(const uint16* src_ptr, int dst_width); TEST_F(LibYUVScaleTest, TestScaleRowUp2_16) { - SIMD_ALIGNED(uint16 orig_pixels[640 * 2 + 1]); // 2 rows + 1 pixel overrun + SIMD_ALIGNED(uint16 orig_pixels[640 * 2 + 1]); // 2 rows + 1 pixel overrun. SIMD_ALIGNED(uint16 dst_pixels_opt[1280]); SIMD_ALIGNED(uint16 dst_pixels_c[1280]); @@ -535,4 +514,101 @@ TEST_F(LibYUVScaleTest, TestScaleRowDown2Box_16) { EXPECT_EQ(dst_pixels_c[1279], 3839); } +// Test scaling plane with 8 bit C vs 16 bit C and return maximum pixel +// difference. +// 0 = exact. +static int TestPlaneFilter_16(int src_width, + int src_height, + int dst_width, + int dst_height, + FilterMode f, + int benchmark_iterations, + int disable_cpu_flags, + int benchmark_cpu_info) { + if (!SizeValid(src_width, src_height, dst_width, dst_height)) { + return 0; + } + + int i; + int64 src_y_plane_size = (Abs(src_width)) * (Abs(src_height)); + int src_stride_y = Abs(src_width); + int dst_y_plane_size = dst_width * dst_height; + int dst_stride_y = dst_width; + + align_buffer_page_end(src_y, src_y_plane_size); + align_buffer_page_end(src_y_16, src_y_plane_size * 2); + align_buffer_page_end(dst_y_8, dst_y_plane_size); + align_buffer_page_end(dst_y_16, dst_y_plane_size * 2); + uint16* p_src_y_16 = reinterpret_cast(src_y_16); + uint16* p_dst_y_16 = reinterpret_cast(dst_y_16); + + MemRandomize(src_y, src_y_plane_size); + memset(dst_y_8, 0, dst_y_plane_size); + memset(dst_y_16, 1, dst_y_plane_size * 2); + + for (i = 0; i < src_y_plane_size; ++i) { + p_src_y_16[i] = src_y[i] & 255; + } + + MaskCpuFlags(disable_cpu_flags); // Disable all CPU optimization. + ScalePlane(src_y, src_stride_y, src_width, src_height, dst_y_8, dst_stride_y, + dst_width, dst_height, f); + MaskCpuFlags(benchmark_cpu_info); // Enable all CPU optimization. + + for (i = 0; i < benchmark_iterations; ++i) { + ScalePlane_16(p_src_y_16, src_stride_y, src_width, src_height, p_dst_y_16, + dst_stride_y, dst_width, dst_height, f); + } + + // Expect an exact match. + int max_diff = 0; + for (i = 0; i < dst_y_plane_size; ++i) { + int abs_diff = Abs(dst_y_8[i] - p_dst_y_16[i]); + if (abs_diff > max_diff) { + max_diff = abs_diff; + } + } + + free_aligned_buffer_page_end(dst_y_8); + free_aligned_buffer_page_end(dst_y_16); + free_aligned_buffer_page_end(src_y); + free_aligned_buffer_page_end(src_y_16); + + return max_diff; +} + +// The following adjustments in dimensions ensure the scale factor will be +// exactly achieved. +// 2 is chroma subsample. +#define DX(x, nom, denom) static_cast(((Abs(x) / nom + 1) / 2) * nom * 2) +#define SX(x, nom, denom) static_cast(((x / nom + 1) / 2) * denom * 2) + +#define TEST_FACTOR1(name, filter, nom, denom, max_diff) \ + TEST_F(LibYUVScaleTest, ScalePlaneDownBy##name##_##filter##_16) { \ + int diff = TestPlaneFilter_16( \ + SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \ + DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \ + kFilter##filter, benchmark_iterations_, disable_cpu_flags_, \ + benchmark_cpu_info_); \ + EXPECT_LE(diff, max_diff); \ + } + +// Test a scale factor with all 4 filters. Expect unfiltered to be exact, but +// filtering is different fixed point implementations for SSSE3, Neon and C. +#define TEST_FACTOR(name, nom, denom, boxdiff) \ + TEST_FACTOR1(name, None, nom, denom, 0) \ + TEST_FACTOR1(name, Linear, nom, denom, boxdiff) \ + TEST_FACTOR1(name, Bilinear, nom, denom, boxdiff) \ + TEST_FACTOR1(name, Box, nom, denom, boxdiff) + +TEST_FACTOR(2, 1, 2, 0) +TEST_FACTOR(4, 1, 4, 0) +TEST_FACTOR(8, 1, 8, 0) +TEST_FACTOR(3by4, 3, 4, 1) +TEST_FACTOR(3by8, 3, 8, 1) +TEST_FACTOR(3, 1, 3, 0) +#undef TEST_FACTOR1 +#undef TEST_FACTOR +#undef SX +#undef DX } // namespace libyuv