1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libyuv/unit_test/scale_argb_test.cc Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,270 @@
1.4 +/*
1.5 + * Copyright 2011 The LibYuv Project Authors. All rights reserved.
1.6 + *
1.7 + * Use of this source code is governed by a BSD-style license
1.8 + * that can be found in the LICENSE file in the root of the source
1.9 + * tree. An additional intellectual property rights grant can be found
1.10 + * in the file PATENTS. All contributing project authors may
1.11 + * be found in the AUTHORS file in the root of the source tree.
1.12 + */
1.13 +
1.14 +#include <stdlib.h>
1.15 +#include <time.h>
1.16 +
1.17 +#include "libyuv/cpu_id.h"
1.18 +#include "libyuv/scale_argb.h"
1.19 +#include "libyuv/row.h"
1.20 +#include "../unit_test/unit_test.h"
1.21 +
1.22 +namespace libyuv {
1.23 +
1.24 +// Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
1.25 +static int ARGBTestFilter(int src_width, int src_height,
1.26 + int dst_width, int dst_height,
1.27 + FilterMode f, int benchmark_iterations) {
1.28 + const int b = 128;
1.29 + int i, j;
1.30 + int src_argb_plane_size = (Abs(src_width) + b * 2) *
1.31 + (Abs(src_height) + b * 2) * 4;
1.32 + int src_stride_argb = (b * 2 + Abs(src_width)) * 4;
1.33 +
1.34 + align_buffer_64(src_argb, src_argb_plane_size);
1.35 + srandom(time(NULL));
1.36 + MemRandomize(src_argb, src_argb_plane_size);
1.37 +
1.38 + int dst_argb_plane_size = (dst_width + b * 2) * (dst_height + b * 2) * 4;
1.39 + int dst_stride_argb = (b * 2 + dst_width) * 4;
1.40 +
1.41 + align_buffer_64(dst_argb_c, dst_argb_plane_size);
1.42 + align_buffer_64(dst_argb_opt, dst_argb_plane_size);
1.43 + memset(dst_argb_c, 2, dst_argb_plane_size);
1.44 + memset(dst_argb_opt, 3, dst_argb_plane_size);
1.45 +
1.46 + // Warm up both versions for consistent benchmarks.
1.47 + MaskCpuFlags(0); // Disable all CPU optimization.
1.48 + ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb,
1.49 + src_width, src_height,
1.50 + dst_argb_c + (dst_stride_argb * b) + b * 4, dst_stride_argb,
1.51 + dst_width, dst_height, f);
1.52 + MaskCpuFlags(-1); // Enable all CPU optimization.
1.53 + ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb,
1.54 + src_width, src_height,
1.55 + dst_argb_opt + (dst_stride_argb * b) + b * 4, dst_stride_argb,
1.56 + dst_width, dst_height, f);
1.57 +
1.58 + MaskCpuFlags(0); // Disable all CPU optimization.
1.59 + double c_time = get_time();
1.60 + ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb,
1.61 + src_width, src_height,
1.62 + dst_argb_c + (dst_stride_argb * b) + b * 4, dst_stride_argb,
1.63 + dst_width, dst_height, f);
1.64 +
1.65 + c_time = (get_time() - c_time);
1.66 +
1.67 + MaskCpuFlags(-1); // Enable all CPU optimization.
1.68 + double opt_time = get_time();
1.69 + for (i = 0; i < benchmark_iterations; ++i) {
1.70 + ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb,
1.71 + src_width, src_height,
1.72 + dst_argb_opt + (dst_stride_argb * b) + b * 4, dst_stride_argb,
1.73 + dst_width, dst_height, f);
1.74 + }
1.75 + opt_time = (get_time() - opt_time) / benchmark_iterations;
1.76 +
1.77 + // Report performance of C vs OPT
1.78 + printf("filter %d - %8d us C - %8d us OPT\n",
1.79 + f, static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
1.80 +
1.81 + // C version may be a little off from the optimized. Order of
1.82 + // operations may introduce rounding somewhere. So do a difference
1.83 + // of the buffers and look to see that the max difference isn't
1.84 + // over 2.
1.85 + int max_diff = 0;
1.86 + for (i = b; i < (dst_height + b); ++i) {
1.87 + for (j = b * 4; j < (dst_width + b) * 4; ++j) {
1.88 + int abs_diff = Abs(dst_argb_c[(i * dst_stride_argb) + j] -
1.89 + dst_argb_opt[(i * dst_stride_argb) + j]);
1.90 + if (abs_diff > max_diff) {
1.91 + max_diff = abs_diff;
1.92 + }
1.93 + }
1.94 + }
1.95 +
1.96 + free_aligned_buffer_64(dst_argb_c);
1.97 + free_aligned_buffer_64(dst_argb_opt);
1.98 + free_aligned_buffer_64(src_argb);
1.99 + return max_diff;
1.100 +}
1.101 +
1.102 +static const int kTileX = 8;
1.103 +static const int kTileY = 8;
1.104 +
1.105 +static int TileARGBScale(const uint8* src_argb, int src_stride_argb,
1.106 + int src_width, int src_height,
1.107 + uint8* dst_argb, int dst_stride_argb,
1.108 + int dst_width, int dst_height,
1.109 + FilterMode filtering) {
1.110 + for (int y = 0; y < dst_height; y += kTileY) {
1.111 + for (int x = 0; x < dst_width; x += kTileX) {
1.112 + int clip_width = kTileX;
1.113 + if (x + clip_width > dst_width) {
1.114 + clip_width = dst_width - x;
1.115 + }
1.116 + int clip_height = kTileY;
1.117 + if (y + clip_height > dst_height) {
1.118 + clip_height = dst_height - y;
1.119 + }
1.120 + int r = ARGBScaleClip(src_argb, src_stride_argb,
1.121 + src_width, src_height,
1.122 + dst_argb, dst_stride_argb,
1.123 + dst_width, dst_height,
1.124 + x, y, clip_width, clip_height, filtering);
1.125 + if (r) {
1.126 + return r;
1.127 + }
1.128 + }
1.129 + }
1.130 + return 0;
1.131 +}
1.132 +
1.133 +static int ARGBClipTestFilter(int src_width, int src_height,
1.134 + int dst_width, int dst_height,
1.135 + FilterMode f, int benchmark_iterations) {
1.136 + const int b = 128;
1.137 + int src_argb_plane_size = (Abs(src_width) + b * 2) *
1.138 + (Abs(src_height) + b * 2) * 4;
1.139 + int src_stride_argb = (b * 2 + Abs(src_width)) * 4;
1.140 +
1.141 + align_buffer_64(src_argb, src_argb_plane_size);
1.142 + memset(src_argb, 1, src_argb_plane_size);
1.143 +
1.144 + int dst_argb_plane_size = (dst_width + b * 2) * (dst_height + b * 2) * 4;
1.145 + int dst_stride_argb = (b * 2 + dst_width) * 4;
1.146 +
1.147 + srandom(time(NULL));
1.148 +
1.149 + int i, j;
1.150 + for (i = b; i < (Abs(src_height) + b); ++i) {
1.151 + for (j = b; j < (Abs(src_width) + b) * 4; ++j) {
1.152 + src_argb[(i * src_stride_argb) + j] = (random() & 0xff);
1.153 + }
1.154 + }
1.155 +
1.156 + align_buffer_64(dst_argb_c, dst_argb_plane_size);
1.157 + align_buffer_64(dst_argb_opt, dst_argb_plane_size);
1.158 + memset(dst_argb_c, 2, dst_argb_plane_size);
1.159 + memset(dst_argb_opt, 3, dst_argb_plane_size);
1.160 +
1.161 + // Do full image, no clipping.
1.162 + double c_time = get_time();
1.163 + ARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb,
1.164 + src_width, src_height,
1.165 + dst_argb_c + (dst_stride_argb * b) + b * 4, dst_stride_argb,
1.166 + dst_width, dst_height, f);
1.167 + c_time = (get_time() - c_time);
1.168 +
1.169 + // Do tiled image, clipping scale to a tile at a time.
1.170 + double opt_time = get_time();
1.171 + for (i = 0; i < benchmark_iterations; ++i) {
1.172 + TileARGBScale(src_argb + (src_stride_argb * b) + b * 4, src_stride_argb,
1.173 + src_width, src_height,
1.174 + dst_argb_opt + (dst_stride_argb * b) + b * 4, dst_stride_argb,
1.175 + dst_width, dst_height, f);
1.176 + }
1.177 + opt_time = (get_time() - opt_time) / benchmark_iterations;
1.178 +
1.179 + // Report performance of Full vs Tiled.
1.180 + printf("filter %d - %8d us Full - %8d us Tiled\n",
1.181 + f, static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
1.182 +
1.183 + // Compare full scaled image vs tiled image.
1.184 + int max_diff = 0;
1.185 + for (i = b; i < (dst_height + b); ++i) {
1.186 + for (j = b * 4; j < (dst_width + b) * 4; ++j) {
1.187 + int abs_diff = Abs(dst_argb_c[(i * dst_stride_argb) + j] -
1.188 + dst_argb_opt[(i * dst_stride_argb) + j]);
1.189 + if (abs_diff > max_diff) {
1.190 + max_diff = abs_diff;
1.191 + }
1.192 + }
1.193 + }
1.194 +
1.195 + free_aligned_buffer_64(dst_argb_c);
1.196 + free_aligned_buffer_64(dst_argb_opt);
1.197 + free_aligned_buffer_64(src_argb);
1.198 + return max_diff;
1.199 +}
1.200 +
1.201 +#define TEST_FACTOR1(name, filter, hfactor, vfactor, max_diff) \
1.202 + TEST_F(libyuvTest, ARGBScaleDownBy##name##_##filter) { \
1.203 + int diff = ARGBTestFilter(benchmark_width_, benchmark_height_, \
1.204 + Abs(benchmark_width_) * hfactor, \
1.205 + Abs(benchmark_height_) * vfactor, \
1.206 + kFilter##filter, benchmark_iterations_); \
1.207 + EXPECT_LE(diff, max_diff); \
1.208 + } \
1.209 + TEST_F(libyuvTest, ARGBScaleDownClipBy##name##_##filter) { \
1.210 + int diff = ARGBClipTestFilter(benchmark_width_, benchmark_height_, \
1.211 + Abs(benchmark_width_) * hfactor, \
1.212 + Abs(benchmark_height_) * vfactor, \
1.213 + kFilter##filter, benchmark_iterations_); \
1.214 + EXPECT_LE(diff, max_diff); \
1.215 + }
1.216 +
1.217 +// Test a scale factor with 2 filters. Expect unfiltered to be exact, but
1.218 +// filtering is different fixed point implementations for SSSE3, Neon and C.
1.219 +#define TEST_FACTOR(name, hfactor, vfactor) \
1.220 + TEST_FACTOR1(name, None, hfactor, vfactor, 2) \
1.221 + TEST_FACTOR1(name, Linear, hfactor, vfactor, 2) \
1.222 + TEST_FACTOR1(name, Bilinear, hfactor, vfactor, 2) \
1.223 + TEST_FACTOR1(name, Box, hfactor, vfactor, 2)
1.224 +
1.225 +TEST_FACTOR(2, 1 / 2, 1 / 2)
1.226 +TEST_FACTOR(4, 1 / 4, 1 / 4)
1.227 +TEST_FACTOR(8, 1 / 8, 1 / 8)
1.228 +TEST_FACTOR(3by4, 3 / 4, 3 / 4)
1.229 +#undef TEST_FACTOR1
1.230 +#undef TEST_FACTOR
1.231 +
1.232 +#define TEST_SCALETO1(name, width, height, filter, max_diff) \
1.233 + TEST_F(libyuvTest, name##To##width##x##height##_##filter) { \
1.234 + int diff = ARGBTestFilter(benchmark_width_, benchmark_height_, \
1.235 + width, height, \
1.236 + kFilter##filter, benchmark_iterations_); \
1.237 + EXPECT_LE(diff, max_diff); \
1.238 + } \
1.239 + TEST_F(libyuvTest, name##From##width##x##height##_##filter) { \
1.240 + int diff = ARGBTestFilter(width, height, \
1.241 + Abs(benchmark_width_), Abs(benchmark_height_), \
1.242 + kFilter##filter, benchmark_iterations_); \
1.243 + EXPECT_LE(diff, max_diff); \
1.244 + } \
1.245 + TEST_F(libyuvTest, name##ClipTo##width##x##height##_##filter) { \
1.246 + int diff = ARGBClipTestFilter(benchmark_width_, benchmark_height_, \
1.247 + width, height, \
1.248 + kFilter##filter, benchmark_iterations_); \
1.249 + EXPECT_LE(diff, max_diff); \
1.250 + } \
1.251 + TEST_F(libyuvTest, name##ClipFrom##width##x##height##_##filter) { \
1.252 + int diff = ARGBClipTestFilter(width, height, \
1.253 + Abs(benchmark_width_), Abs(benchmark_height_), \
1.254 + kFilter##filter, benchmark_iterations_); \
1.255 + EXPECT_LE(diff, max_diff); \
1.256 + }
1.257 +
1.258 +/// Test scale to a specified size with all 4 filters.
1.259 +#define TEST_SCALETO(name, width, height) \
1.260 + TEST_SCALETO1(name, width, height, None, 0) \
1.261 + TEST_SCALETO1(name, width, height, Linear, 3) \
1.262 + TEST_SCALETO1(name, width, height, Bilinear, 3) \
1.263 + TEST_SCALETO1(name, width, height, Box, 3)
1.264 +
1.265 +TEST_SCALETO(ARGBScale, 1, 1)
1.266 +TEST_SCALETO(ARGBScale, 320, 240)
1.267 +TEST_SCALETO(ARGBScale, 352, 288)
1.268 +TEST_SCALETO(ARGBScale, 640, 360)
1.269 +TEST_SCALETO(ARGBScale, 1280, 720)
1.270 +#undef TEST_SCALETO1
1.271 +#undef TEST_SCALETO
1.272 +
1.273 +} // namespace libyuv
