michael@0: /* michael@0: * Copyright 2011 The LibYuv Project Authors. All rights reserved. michael@0: * michael@0: * Use of this source code is governed by a BSD-style license michael@0: * that can be found in the LICENSE file in the root of the source michael@0: * tree. An additional intellectual property rights grant can be found michael@0: * in the file PATENTS. All contributing project authors may michael@0: * be found in the AUTHORS file in the root of the source tree. michael@0: */ michael@0: michael@0: #include michael@0: #include michael@0: michael@0: #include "libyuv/cpu_id.h" michael@0: #include "libyuv/scale_argb.h" michael@0: #include "libyuv/row.h" michael@0: #include "../unit_test/unit_test.h" michael@0: michael@0: namespace libyuv { michael@0: michael@0: // Test scaling with C vs Opt and return maximum pixel difference. 0 = exact. michael@0: static int ARGBTestFilter(int src_width, int src_height, michael@0: int dst_width, int dst_height, michael@0: FilterMode f, int benchmark_iterations) { michael@0: const int b = 128; michael@0: int i, j; michael@0: int src_argb_plane_size = (Abs(src_width) + b * 2) * michael@0: (Abs(src_height) + b * 2) * 4; michael@0: int src_stride_argb = (b * 2 + Abs(src_width)) * 4; michael@0: michael@0: align_buffer_64(src_argb, src_argb_plane_size); michael@0: srandom(time(NULL)); michael@0: MemRandomize(src_argb, src_argb_plane_size); michael@0: michael@0: int dst_argb_plane_size = (dst_width + b * 2) * (dst_height + b * 2) * 4; michael@0: int dst_stride_argb = (b * 2 + dst_width) * 4; michael@0: michael@0: align_buffer_64(dst_argb_c, dst_argb_plane_size); michael@0: align_buffer_64(dst_argb_opt, dst_argb_plane_size); michael@0: memset(dst_argb_c, 2, dst_argb_plane_size); michael@0: memset(dst_argb_opt, 3, dst_argb_plane_size); michael@0: michael@0: // Warm up both versions for consistent benchmarks. michael@0: MaskCpuFlags(0); // Disable all CPU optimization. michael@0: ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb, michael@0: src_width, src_height, michael@0: dst_argb_c + (dst_stride_argb * b) + b * 4, dst_stride_argb, michael@0: dst_width, dst_height, f); michael@0: MaskCpuFlags(-1); // Enable all CPU optimization. michael@0: ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb, michael@0: src_width, src_height, michael@0: dst_argb_opt + (dst_stride_argb * b) + b * 4, dst_stride_argb, michael@0: dst_width, dst_height, f); michael@0: michael@0: MaskCpuFlags(0); // Disable all CPU optimization. michael@0: double c_time = get_time(); michael@0: ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb, michael@0: src_width, src_height, michael@0: dst_argb_c + (dst_stride_argb * b) + b * 4, dst_stride_argb, michael@0: dst_width, dst_height, f); michael@0: michael@0: c_time = (get_time() - c_time); michael@0: michael@0: MaskCpuFlags(-1); // Enable all CPU optimization. michael@0: double opt_time = get_time(); michael@0: for (i = 0; i < benchmark_iterations; ++i) { michael@0: ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb, michael@0: src_width, src_height, michael@0: dst_argb_opt + (dst_stride_argb * b) + b * 4, dst_stride_argb, michael@0: dst_width, dst_height, f); michael@0: } michael@0: opt_time = (get_time() - opt_time) / benchmark_iterations; michael@0: michael@0: // Report performance of C vs OPT michael@0: printf("filter %d - %8d us C - %8d us OPT\n", michael@0: f, static_cast(c_time * 1e6), static_cast(opt_time * 1e6)); michael@0: michael@0: // C version may be a little off from the optimized. Order of michael@0: // operations may introduce rounding somewhere. So do a difference michael@0: // of the buffers and look to see that the max difference isn't michael@0: // over 2. michael@0: int max_diff = 0; michael@0: for (i = b; i < (dst_height + b); ++i) { michael@0: for (j = b * 4; j < (dst_width + b) * 4; ++j) { michael@0: int abs_diff = Abs(dst_argb_c[(i * dst_stride_argb) + j] - michael@0: dst_argb_opt[(i * dst_stride_argb) + j]); michael@0: if (abs_diff > max_diff) { michael@0: max_diff = abs_diff; michael@0: } michael@0: } michael@0: } michael@0: michael@0: free_aligned_buffer_64(dst_argb_c); michael@0: free_aligned_buffer_64(dst_argb_opt); michael@0: free_aligned_buffer_64(src_argb); michael@0: return max_diff; michael@0: } michael@0: michael@0: static const int kTileX = 8; michael@0: static const int kTileY = 8; michael@0: michael@0: static int TileARGBScale(const uint8* src_argb, int src_stride_argb, michael@0: int src_width, int src_height, michael@0: uint8* dst_argb, int dst_stride_argb, michael@0: int dst_width, int dst_height, michael@0: FilterMode filtering) { michael@0: for (int y = 0; y < dst_height; y += kTileY) { michael@0: for (int x = 0; x < dst_width; x += kTileX) { michael@0: int clip_width = kTileX; michael@0: if (x + clip_width > dst_width) { michael@0: clip_width = dst_width - x; michael@0: } michael@0: int clip_height = kTileY; michael@0: if (y + clip_height > dst_height) { michael@0: clip_height = dst_height - y; michael@0: } michael@0: int r = ARGBScaleClip(src_argb, src_stride_argb, michael@0: src_width, src_height, michael@0: dst_argb, dst_stride_argb, michael@0: dst_width, dst_height, michael@0: x, y, clip_width, clip_height, filtering); michael@0: if (r) { michael@0: return r; michael@0: } michael@0: } michael@0: } michael@0: return 0; michael@0: } michael@0: michael@0: static int ARGBClipTestFilter(int src_width, int src_height, michael@0: int dst_width, int dst_height, michael@0: FilterMode f, int benchmark_iterations) { michael@0: const int b = 128; michael@0: int src_argb_plane_size = (Abs(src_width) + b * 2) * michael@0: (Abs(src_height) + b * 2) * 4; michael@0: int src_stride_argb = (b * 2 + Abs(src_width)) * 4; michael@0: michael@0: align_buffer_64(src_argb, src_argb_plane_size); michael@0: memset(src_argb, 1, src_argb_plane_size); michael@0: michael@0: int dst_argb_plane_size = (dst_width + b * 2) * (dst_height + b * 2) * 4; michael@0: int dst_stride_argb = (b * 2 + dst_width) * 4; michael@0: michael@0: srandom(time(NULL)); michael@0: michael@0: int i, j; michael@0: for (i = b; i < (Abs(src_height) + b); ++i) { michael@0: for (j = b; j < (Abs(src_width) + b) * 4; ++j) { michael@0: src_argb[(i * src_stride_argb) + j] = (random() & 0xff); michael@0: } michael@0: } michael@0: michael@0: align_buffer_64(dst_argb_c, dst_argb_plane_size); michael@0: align_buffer_64(dst_argb_opt, dst_argb_plane_size); michael@0: memset(dst_argb_c, 2, dst_argb_plane_size); michael@0: memset(dst_argb_opt, 3, dst_argb_plane_size); michael@0: michael@0: // Do full image, no clipping. michael@0: double c_time = get_time(); michael@0: ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb, michael@0: src_width, src_height, michael@0: dst_argb_c + (dst_stride_argb * b) + b * 4, dst_stride_argb, michael@0: dst_width, dst_height, f); michael@0: c_time = (get_time() - c_time); michael@0: michael@0: // Do tiled image, clipping scale to a tile at a time. michael@0: double opt_time = get_time(); michael@0: for (i = 0; i < benchmark_iterations; ++i) { michael@0: TileARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb, michael@0: src_width, src_height, michael@0: dst_argb_opt + (dst_stride_argb * b) + b * 4, dst_stride_argb, michael@0: dst_width, dst_height, f); michael@0: } michael@0: opt_time = (get_time() - opt_time) / benchmark_iterations; michael@0: michael@0: // Report performance of Full vs Tiled. michael@0: printf("filter %d - %8d us Full - %8d us Tiled\n", michael@0: f, static_cast(c_time * 1e6), static_cast(opt_time * 1e6)); michael@0: michael@0: // Compare full scaled image vs tiled image. michael@0: int max_diff = 0; michael@0: for (i = b; i < (dst_height + b); ++i) { michael@0: for (j = b * 4; j < (dst_width + b) * 4; ++j) { michael@0: int abs_diff = Abs(dst_argb_c[(i * dst_stride_argb) + j] - michael@0: dst_argb_opt[(i * dst_stride_argb) + j]); michael@0: if (abs_diff > max_diff) { michael@0: max_diff = abs_diff; michael@0: } michael@0: } michael@0: } michael@0: michael@0: free_aligned_buffer_64(dst_argb_c); michael@0: free_aligned_buffer_64(dst_argb_opt); michael@0: free_aligned_buffer_64(src_argb); michael@0: return max_diff; michael@0: } michael@0: michael@0: #define TEST_FACTOR1(name, filter, hfactor, vfactor, max_diff) \ michael@0: TEST_F(libyuvTest, ARGBScaleDownBy##name##_##filter) { \ michael@0: int diff = ARGBTestFilter(benchmark_width_, benchmark_height_, \ michael@0: Abs(benchmark_width_) * hfactor, \ michael@0: Abs(benchmark_height_) * vfactor, \ michael@0: kFilter##filter, benchmark_iterations_); \ michael@0: EXPECT_LE(diff, max_diff); \ michael@0: } \ michael@0: TEST_F(libyuvTest, ARGBScaleDownClipBy##name##_##filter) { \ michael@0: int diff = ARGBClipTestFilter(benchmark_width_, benchmark_height_, \ michael@0: Abs(benchmark_width_) * hfactor, \ michael@0: Abs(benchmark_height_) * vfactor, \ michael@0: kFilter##filter, benchmark_iterations_); \ michael@0: EXPECT_LE(diff, max_diff); \ michael@0: } michael@0: michael@0: // Test a scale factor with 2 filters. Expect unfiltered to be exact, but michael@0: // filtering is different fixed point implementations for SSSE3, Neon and C. michael@0: #define TEST_FACTOR(name, hfactor, vfactor) \ michael@0: TEST_FACTOR1(name, None, hfactor, vfactor, 2) \ michael@0: TEST_FACTOR1(name, Linear, hfactor, vfactor, 2) \ michael@0: TEST_FACTOR1(name, Bilinear, hfactor, vfactor, 2) \ michael@0: TEST_FACTOR1(name, Box, hfactor, vfactor, 2) michael@0: michael@0: TEST_FACTOR(2, 1 / 2, 1 / 2) michael@0: TEST_FACTOR(4, 1 / 4, 1 / 4) michael@0: TEST_FACTOR(8, 1 / 8, 1 / 8) michael@0: TEST_FACTOR(3by4, 3 / 4, 3 / 4) michael@0: #undef TEST_FACTOR1 michael@0: #undef TEST_FACTOR michael@0: michael@0: #define TEST_SCALETO1(name, width, height, filter, max_diff) \ michael@0: TEST_F(libyuvTest, name##To##width##x##height##_##filter) { \ michael@0: int diff = ARGBTestFilter(benchmark_width_, benchmark_height_, \ michael@0: width, height, \ michael@0: kFilter##filter, benchmark_iterations_); \ michael@0: EXPECT_LE(diff, max_diff); \ michael@0: } \ michael@0: TEST_F(libyuvTest, name##From##width##x##height##_##filter) { \ michael@0: int diff = ARGBTestFilter(width, height, \ michael@0: Abs(benchmark_width_), Abs(benchmark_height_), \ michael@0: kFilter##filter, benchmark_iterations_); \ michael@0: EXPECT_LE(diff, max_diff); \ michael@0: } \ michael@0: TEST_F(libyuvTest, name##ClipTo##width##x##height##_##filter) { \ michael@0: int diff = ARGBClipTestFilter(benchmark_width_, benchmark_height_, \ michael@0: width, height, \ michael@0: kFilter##filter, benchmark_iterations_); \ michael@0: EXPECT_LE(diff, max_diff); \ michael@0: } \ michael@0: TEST_F(libyuvTest, name##ClipFrom##width##x##height##_##filter) { \ michael@0: int diff = ARGBClipTestFilter(width, height, \ michael@0: Abs(benchmark_width_), Abs(benchmark_height_), \ michael@0: kFilter##filter, benchmark_iterations_); \ michael@0: EXPECT_LE(diff, max_diff); \ michael@0: } michael@0: michael@0: /// Test scale to a specified size with all 4 filters. michael@0: #define TEST_SCALETO(name, width, height) \ michael@0: TEST_SCALETO1(name, width, height, None, 0) \ michael@0: TEST_SCALETO1(name, width, height, Linear, 3) \ michael@0: TEST_SCALETO1(name, width, height, Bilinear, 3) \ michael@0: TEST_SCALETO1(name, width, height, Box, 3) michael@0: michael@0: TEST_SCALETO(ARGBScale, 1, 1) michael@0: TEST_SCALETO(ARGBScale, 320, 240) michael@0: TEST_SCALETO(ARGBScale, 352, 288) michael@0: TEST_SCALETO(ARGBScale, 640, 360) michael@0: TEST_SCALETO(ARGBScale, 1280, 720) michael@0: #undef TEST_SCALETO1 michael@0: #undef TEST_SCALETO michael@0: michael@0: } // namespace libyuv