1 /*
2 * Copyright 2022 The LibYuv Project Authors. All rights reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include <stdlib.h>
12 #include <time.h>
13
14 #include "../unit_test/unit_test.h"
15 #include "libyuv/cpu_id.h"
16 #include "libyuv/scale_rgb.h"
17
18 namespace libyuv {
19
20 #define STRINGIZE(line) #line
21 #define FILELINESTR(file, line) file ":" STRINGIZE(line)
22
23 #if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
24 // SLOW TESTS are those that are unoptimized C code.
25 // FULL TESTS are optimized but test many variations of the same code.
26 #define ENABLE_FULL_TESTS
27 #endif
28
29 // Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
RGBTestFilter(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)30 static int RGBTestFilter(int src_width,
31 int src_height,
32 int dst_width,
33 int dst_height,
34 FilterMode f,
35 int benchmark_iterations,
36 int disable_cpu_flags,
37 int benchmark_cpu_info) {
38 if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
39 return 0;
40 }
41
42 int i, j;
43 const int b = 0; // 128 to test for padding/stride.
44 int64_t src_rgb_plane_size =
45 (Abs(src_width) + b * 3) * (Abs(src_height) + b * 3) * 3LL;
46 int src_stride_rgb = (b * 3 + Abs(src_width)) * 3;
47
48 align_buffer_page_end(src_rgb, src_rgb_plane_size);
49 if (!src_rgb) {
50 printf("Skipped. Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
51 return 0;
52 }
53 MemRandomize(src_rgb, src_rgb_plane_size);
54
55 int64_t dst_rgb_plane_size = (dst_width + b * 3) * (dst_height + b * 3) * 3LL;
56 int dst_stride_rgb = (b * 3 + dst_width) * 3;
57
58 align_buffer_page_end(dst_rgb_c, dst_rgb_plane_size);
59 align_buffer_page_end(dst_rgb_opt, dst_rgb_plane_size);
60 if (!dst_rgb_c || !dst_rgb_opt) {
61 printf("Skipped. Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
62 return 0;
63 }
64 memset(dst_rgb_c, 2, dst_rgb_plane_size);
65 memset(dst_rgb_opt, 3, dst_rgb_plane_size);
66
67 // Warm up both versions for consistent benchmarks.
68 MaskCpuFlags(disable_cpu_flags); // Disable all CPU optimization.
69 RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
70 src_height, dst_rgb_c + (dst_stride_rgb * b) + b * 3, dst_stride_rgb,
71 dst_width, dst_height, f);
72 MaskCpuFlags(benchmark_cpu_info); // Enable all CPU optimization.
73 RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
74 src_height, dst_rgb_opt + (dst_stride_rgb * b) + b * 3,
75 dst_stride_rgb, dst_width, dst_height, f);
76
77 MaskCpuFlags(disable_cpu_flags); // Disable all CPU optimization.
78 double c_time = get_time();
79 RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
80 src_height, dst_rgb_c + (dst_stride_rgb * b) + b * 3, dst_stride_rgb,
81 dst_width, dst_height, f);
82
83 c_time = (get_time() - c_time);
84
85 MaskCpuFlags(benchmark_cpu_info); // Enable all CPU optimization.
86 double opt_time = get_time();
87 for (i = 0; i < benchmark_iterations; ++i) {
88 RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
89 src_height, dst_rgb_opt + (dst_stride_rgb * b) + b * 3,
90 dst_stride_rgb, dst_width, dst_height, f);
91 }
92 opt_time = (get_time() - opt_time) / benchmark_iterations;
93
94 // Report performance of C vs OPT
95 printf("filter %d - %8d us C - %8d us OPT\n", f,
96 static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
97
98 // C version may be a little off from the optimized. Order of
99 // operations may introduce rounding somewhere. So do a difference
100 // of the buffers and look to see that the max difference isn't
101 // over 2.
102 int max_diff = 0;
103 for (i = b; i < (dst_height + b); ++i) {
104 for (j = b * 3; j < (dst_width + b) * 3; ++j) {
105 int abs_diff = Abs(dst_rgb_c[(i * dst_stride_rgb) + j] -
106 dst_rgb_opt[(i * dst_stride_rgb) + j]);
107 if (abs_diff > max_diff) {
108 max_diff = abs_diff;
109 }
110 }
111 }
112
113 free_aligned_buffer_page_end(dst_rgb_c);
114 free_aligned_buffer_page_end(dst_rgb_opt);
115 free_aligned_buffer_page_end(src_rgb);
116 return max_diff;
117 }
118
119 // The following adjustments in dimensions ensure the scale factor will be
120 // exactly achieved.
121 #define DX(x, nom, denom) static_cast<int>((Abs(x) / nom) * nom)
122 #define SX(x, nom, denom) static_cast<int>((x / nom) * denom)
123
124 #define TEST_FACTOR1(name, filter, nom, denom, max_diff) \
125 TEST_F(LibYUVScaleTest, RGBScaleDownBy##name##_##filter) { \
126 int diff = RGBTestFilter( \
127 SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
128 DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
129 kFilter##filter, benchmark_iterations_, disable_cpu_flags_, \
130 benchmark_cpu_info_); \
131 EXPECT_LE(diff, max_diff); \
132 }
133
134 #if defined(ENABLE_FULL_TESTS)
135 // Test a scale factor with all 4 filters. Expect unfiltered to be exact, but
136 // filtering is different fixed point implementations for SSSE3, Neon and C.
137 #define TEST_FACTOR(name, nom, denom) \
138 TEST_FACTOR1(name, None, nom, denom, 0) \
139 TEST_FACTOR1(name, Linear, nom, denom, 3) \
140 TEST_FACTOR1(name, Bilinear, nom, denom, 3) \
141 TEST_FACTOR1(name, Box, nom, denom, 3)
142 #else
143 // Test a scale factor with Bilinear.
144 #define TEST_FACTOR(name, nom, denom) \
145 TEST_FACTOR1(name, Bilinear, nom, denom, 3)
146 #endif
147
148 TEST_FACTOR(2, 1, 2)
149 #ifndef DISABLE_SLOW_TESTS
150 TEST_FACTOR(4, 1, 4)
151 // TEST_FACTOR(8, 1, 8) Disable for benchmark performance.
152 TEST_FACTOR(3by4, 3, 4)
153 TEST_FACTOR(3by8, 3, 8)
154 TEST_FACTOR(3, 1, 3)
155 #endif
156 #undef TEST_FACTOR1
157 #undef TEST_FACTOR
158 #undef SX
159 #undef DX
160
161 #define TEST_SCALETO1(name, width, height, filter, max_diff) \
162 TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter) { \
163 int diff = RGBTestFilter(benchmark_width_, benchmark_height_, width, \
164 height, kFilter##filter, benchmark_iterations_, \
165 disable_cpu_flags_, benchmark_cpu_info_); \
166 EXPECT_LE(diff, max_diff); \
167 } \
168 TEST_F(LibYUVScaleTest, name##From##width##x##height##_##filter) { \
169 int diff = RGBTestFilter(width, height, Abs(benchmark_width_), \
170 Abs(benchmark_height_), kFilter##filter, \
171 benchmark_iterations_, disable_cpu_flags_, \
172 benchmark_cpu_info_); \
173 EXPECT_LE(diff, max_diff); \
174 }
175
176 #if defined(ENABLE_FULL_TESTS)
177 /// Test scale to a specified size with all 4 filters.
178 #define TEST_SCALETO(name, width, height) \
179 TEST_SCALETO1(name, width, height, None, 0) \
180 TEST_SCALETO1(name, width, height, Linear, 3) \
181 TEST_SCALETO1(name, width, height, Bilinear, 3)
182 #else
183 #define TEST_SCALETO(name, width, height) \
184 TEST_SCALETO1(name, width, height, Bilinear, 3)
185 #endif
186
187 TEST_SCALETO(RGBScale, 640, 360)
188 #ifndef DISABLE_SLOW_TESTS
189 TEST_SCALETO(RGBScale, 1, 1)
190 TEST_SCALETO(RGBScale, 256, 144) /* 128x72 * 3 */
191 TEST_SCALETO(RGBScale, 320, 240)
192 TEST_SCALETO(RGBScale, 569, 480)
193 TEST_SCALETO(RGBScale, 1280, 720)
194 TEST_SCALETO(RGBScale, 1920, 1080)
195 #endif // DISABLE_SLOW_TESTS
196 #undef TEST_SCALETO1
197 #undef TEST_SCALETO
198
199 #define TEST_SCALESWAPXY1(name, filter, max_diff) \
200 TEST_F(LibYUVScaleTest, name##SwapXY_##filter) { \
201 int diff = RGBTestFilter(benchmark_width_, benchmark_height_, \
202 benchmark_height_, benchmark_width_, \
203 kFilter##filter, benchmark_iterations_, \
204 disable_cpu_flags_, benchmark_cpu_info_); \
205 EXPECT_LE(diff, max_diff); \
206 }
207
208 #if defined(ENABLE_FULL_TESTS)
209 // Test scale with swapped width and height with all 3 filters.
210 TEST_SCALESWAPXY1(RGBScale, None, 0)
211 TEST_SCALESWAPXY1(RGBScale, Linear, 0)
212 TEST_SCALESWAPXY1(RGBScale, Bilinear, 0)
213 #else
214 TEST_SCALESWAPXY1(RGBScale, Bilinear, 0)
215 #endif
216 #undef TEST_SCALESWAPXY1
217
TEST_F(LibYUVScaleTest,RGBTest3x)218 TEST_F(LibYUVScaleTest, RGBTest3x) {
219 const int kSrcStride = 480 * 3;
220 const int kDstStride = 160 * 3;
221 const int kSize = kSrcStride * 3;
222 align_buffer_page_end(orig_pixels, kSize);
223 for (int i = 0; i < 480 * 3; ++i) {
224 orig_pixels[i * 3 + 0] = i;
225 orig_pixels[i * 3 + 1] = 255 - i;
226 }
227 align_buffer_page_end(dest_pixels, kDstStride);
228
229 int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
230 benchmark_iterations_;
231 for (int i = 0; i < iterations160; ++i) {
232 RGBScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
233 kFilterBilinear);
234 }
235
236 EXPECT_EQ(225, dest_pixels[0]);
237 EXPECT_EQ(255 - 225, dest_pixels[1]);
238
239 RGBScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
240 kFilterNone);
241
242 EXPECT_EQ(225, dest_pixels[0]);
243 EXPECT_EQ(255 - 225, dest_pixels[1]);
244
245 free_aligned_buffer_page_end(dest_pixels);
246 free_aligned_buffer_page_end(orig_pixels);
247 }
248
TEST_F(LibYUVScaleTest,RGBTest4x)249 TEST_F(LibYUVScaleTest, RGBTest4x) {
250 const int kSrcStride = 640 * 3;
251 const int kDstStride = 160 * 3;
252 const int kSize = kSrcStride * 4;
253 align_buffer_page_end(orig_pixels, kSize);
254 for (int i = 0; i < 640 * 4; ++i) {
255 orig_pixels[i * 3 + 0] = i;
256 orig_pixels[i * 3 + 1] = 255 - i;
257 }
258 align_buffer_page_end(dest_pixels, kDstStride);
259
260 int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
261 benchmark_iterations_;
262 for (int i = 0; i < iterations160; ++i) {
263 RGBScale(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
264 kFilterBilinear);
265 }
266
267 EXPECT_EQ(66, dest_pixels[0]);
268 EXPECT_EQ(190, dest_pixels[1]);
269
270 RGBScale(orig_pixels, kSrcStride, 64, 4, dest_pixels, kDstStride, 16, 1,
271 kFilterNone);
272
273 EXPECT_EQ(2, dest_pixels[0]); // expect the 3rd pixel of the 3rd row
274 EXPECT_EQ(255 - 2, dest_pixels[1]);
275
276 free_aligned_buffer_page_end(dest_pixels);
277 free_aligned_buffer_page_end(orig_pixels);
278 }
279
280 } // namespace libyuv
281