1 /*
2 * Copyright 2011 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 "libyuv/cpu_id.h"
15 #include "libyuv/scale.h"
16 #include "../unit_test/unit_test.h"
17
18 namespace libyuv {
19
TestFilter(int src_width,int src_height,int dst_width,int dst_height,FilterMode f,int rounding,int benchmark_iterations)20 static int TestFilter(int src_width, int src_height,
21 int dst_width, int dst_height,
22 FilterMode f, int rounding, int benchmark_iterations) {
23 const int b = 128 * rounding;
24 int src_width_uv = (src_width + rounding) >> 1;
25 int src_height_uv = (src_height + rounding) >> 1;
26
27 int src_y_plane_size = (src_width + b * 2) * (src_height + b * 2);
28 int src_uv_plane_size = (src_width_uv + b * 2) * (src_height_uv + b * 2);
29
30 int src_stride_y = b * 2 + src_width;
31 int src_stride_uv = b * 2 + src_width_uv;
32
33 align_buffer_page_end(src_y, src_y_plane_size)
34 align_buffer_page_end(src_u, src_uv_plane_size)
35 align_buffer_page_end(src_v, src_uv_plane_size)
36
37 int dst_width_uv = (dst_width + rounding) >> 1;
38 int dst_height_uv = (dst_height + rounding) >> 1;
39
40 int dst_y_plane_size = (dst_width + b * 2) * (dst_height + b * 2);
41 int dst_uv_plane_size = (dst_width_uv + b * 2) * (dst_height_uv + b * 2);
42
43 int dst_stride_y = b * 2 + dst_width;
44 int dst_stride_uv = b * 2 + dst_width_uv;
45
46 srandom(time(NULL));
47
48 int i, j;
49 for (i = b; i < (src_height + b); ++i) {
50 for (j = b; j < (src_width + b); ++j) {
51 src_y[(i * src_stride_y) + j] = (random() & 0xff);
52 }
53 }
54
55 for (i = b; i < (src_height_uv + b); ++i) {
56 for (j = b; j < (src_width_uv + b); ++j) {
57 src_u[(i * src_stride_uv) + j] = (random() & 0xff);
58 src_v[(i * src_stride_uv) + j] = (random() & 0xff);
59 }
60 }
61
62 align_buffer_page_end(dst_y_c, dst_y_plane_size)
63 align_buffer_page_end(dst_u_c, dst_uv_plane_size)
64 align_buffer_page_end(dst_v_c, dst_uv_plane_size)
65 align_buffer_page_end(dst_y_opt, dst_y_plane_size)
66 align_buffer_page_end(dst_u_opt, dst_uv_plane_size)
67 align_buffer_page_end(dst_v_opt, dst_uv_plane_size)
68
69 // Warm up both versions for consistent benchmarks.
70 MaskCpuFlags(0); // Disable all CPU optimization.
71 I420Scale(src_y + (src_stride_y * b) + b, src_stride_y,
72 src_u + (src_stride_uv * b) + b, src_stride_uv,
73 src_v + (src_stride_uv * b) + b, src_stride_uv,
74 src_width, src_height,
75 dst_y_c + (dst_stride_y * b) + b, dst_stride_y,
76 dst_u_c + (dst_stride_uv * b) + b, dst_stride_uv,
77 dst_v_c + (dst_stride_uv * b) + b, dst_stride_uv,
78 dst_width, dst_height, f);
79 MaskCpuFlags(-1); // Enable all CPU optimization.
80 I420Scale(src_y + (src_stride_y * b) + b, src_stride_y,
81 src_u + (src_stride_uv * b) + b, src_stride_uv,
82 src_v + (src_stride_uv * b) + b, src_stride_uv,
83 src_width, src_height,
84 dst_y_opt + (dst_stride_y * b) + b, dst_stride_y,
85 dst_u_opt + (dst_stride_uv * b) + b, dst_stride_uv,
86 dst_v_opt + (dst_stride_uv * b) + b, dst_stride_uv,
87 dst_width, dst_height, f);
88
89 MaskCpuFlags(0); // Disable all CPU optimization.
90 double c_time = get_time();
91 for (i = 0; i < benchmark_iterations; ++i) {
92 I420Scale(src_y + (src_stride_y * b) + b, src_stride_y,
93 src_u + (src_stride_uv * b) + b, src_stride_uv,
94 src_v + (src_stride_uv * b) + b, src_stride_uv,
95 src_width, src_height,
96 dst_y_c + (dst_stride_y * b) + b, dst_stride_y,
97 dst_u_c + (dst_stride_uv * b) + b, dst_stride_uv,
98 dst_v_c + (dst_stride_uv * b) + b, dst_stride_uv,
99 dst_width, dst_height, f);
100 }
101 c_time = (get_time() - c_time) / benchmark_iterations;
102
103 MaskCpuFlags(-1); // Enable all CPU optimization.
104 double opt_time = get_time();
105 for (i = 0; i < benchmark_iterations; ++i) {
106 I420Scale(src_y + (src_stride_y * b) + b, src_stride_y,
107 src_u + (src_stride_uv * b) + b, src_stride_uv,
108 src_v + (src_stride_uv * b) + b, src_stride_uv,
109 src_width, src_height,
110 dst_y_opt + (dst_stride_y * b) + b, dst_stride_y,
111 dst_u_opt + (dst_stride_uv * b) + b, dst_stride_uv,
112 dst_v_opt + (dst_stride_uv * b) + b, dst_stride_uv,
113 dst_width, dst_height, f);
114 }
115 opt_time = (get_time() - opt_time) / benchmark_iterations;
116
117 // Report performance of C vs OPT
118 printf("filter %d - %8d us C - %8d us OPT\n",
119 f, static_cast<int>(c_time*1e6), static_cast<int>(opt_time*1e6));
120
121 // C version may be a little off from the optimized. Order of
122 // operations may introduce rounding somewhere. So do a difference
123 // of the buffers and look to see that the max difference isn't
124 // over 2.
125 int max_diff = 0;
126 for (i = b; i < (dst_height + b); ++i) {
127 for (j = b; j < (dst_width + b); ++j) {
128 int abs_diff = abs(dst_y_c[(i * dst_stride_y) + j] -
129 dst_y_opt[(i * dst_stride_y) + j]);
130 if (abs_diff > max_diff) {
131 max_diff = abs_diff;
132 }
133 }
134 }
135
136 for (i = b; i < (dst_height_uv + b); ++i) {
137 for (j = b; j < (dst_width_uv + b); ++j) {
138 int abs_diff = abs(dst_u_c[(i * dst_stride_uv) + j] -
139 dst_u_opt[(i * dst_stride_uv) + j]);
140 if (abs_diff > max_diff) {
141 max_diff = abs_diff;
142 }
143 abs_diff = abs(dst_v_c[(i * dst_stride_uv) + j] -
144 dst_v_opt[(i * dst_stride_uv) + j]);
145 if (abs_diff > max_diff) {
146 max_diff = abs_diff;
147 }
148 }
149 }
150
151 free_aligned_buffer_page_end(dst_y_c)
152 free_aligned_buffer_page_end(dst_u_c)
153 free_aligned_buffer_page_end(dst_v_c)
154 free_aligned_buffer_page_end(dst_y_opt)
155 free_aligned_buffer_page_end(dst_u_opt)
156 free_aligned_buffer_page_end(dst_v_opt)
157
158 free_aligned_buffer_page_end(src_y)
159 free_aligned_buffer_page_end(src_u)
160 free_aligned_buffer_page_end(src_v)
161
162 return max_diff;
163 }
164
TEST_F(libyuvTest,ScaleDownBy2)165 TEST_F(libyuvTest, ScaleDownBy2) {
166 const int src_width = 1280;
167 const int src_height = 720;
168 const int dst_width = src_width / 2;
169 const int dst_height = src_height / 2;
170
171 for (int f = 0; f < 3; ++f) {
172 int max_diff = TestFilter(src_width, src_height,
173 dst_width, dst_height,
174 static_cast<FilterMode>(f), 1,
175 benchmark_iterations_);
176 EXPECT_LE(max_diff, 1);
177 }
178 }
179
TEST_F(libyuvTest,ScaleDownBy4)180 TEST_F(libyuvTest, ScaleDownBy4) {
181 const int src_width = 1280;
182 const int src_height = 720;
183 const int dst_width = src_width / 4;
184 const int dst_height = src_height / 4;
185
186 for (int f = 0; f < 3; ++f) {
187 int max_diff = TestFilter(src_width, src_height,
188 dst_width, dst_height,
189 static_cast<FilterMode>(f), 1,
190 benchmark_iterations_);
191 EXPECT_LE(max_diff, 2); // This is the only scale factor with error of 2.
192 }
193 }
194
TEST_F(libyuvTest,ScaleDownBy5)195 TEST_F(libyuvTest, ScaleDownBy5) {
196 const int src_width = 1280;
197 const int src_height = 720;
198 const int dst_width = src_width / 5;
199 const int dst_height = src_height / 5;
200
201 for (int f = 0; f < 3; ++f) {
202 int max_diff = TestFilter(src_width, src_height,
203 dst_width, dst_height,
204 static_cast<FilterMode>(f), 1,
205 benchmark_iterations_);
206 EXPECT_LE(max_diff, 1);
207 }
208 }
209
TEST_F(libyuvTest,ScaleDownBy8)210 TEST_F(libyuvTest, ScaleDownBy8) {
211 const int src_width = 1280;
212 const int src_height = 720;
213 const int dst_width = src_width / 8;
214 const int dst_height = src_height / 8;
215
216 for (int f = 0; f < 3; ++f) {
217 int max_diff = TestFilter(src_width, src_height,
218 dst_width, dst_height,
219 static_cast<FilterMode>(f), 1,
220 benchmark_iterations_);
221 EXPECT_LE(max_diff, 1);
222 }
223 }
224
TEST_F(libyuvTest,ScaleDownBy16)225 TEST_F(libyuvTest, ScaleDownBy16) {
226 const int src_width = 1280;
227 const int src_height = 720;
228 const int dst_width = src_width / 16;
229 const int dst_height = src_height / 16;
230
231 for (int f = 0; f < 3; ++f) {
232 int max_diff = TestFilter(src_width, src_height,
233 dst_width, dst_height,
234 static_cast<FilterMode>(f), 1,
235 benchmark_iterations_);
236 EXPECT_LE(max_diff, 1);
237 }
238 }
239
TEST_F(libyuvTest,ScaleDownBy34)240 TEST_F(libyuvTest, ScaleDownBy34) {
241 const int src_width = 1280;
242 const int src_height = 720;
243 const int dst_width = src_width * 3 / 4;
244 const int dst_height = src_height * 3 / 4;
245
246 for (int f = 0; f < 3; ++f) {
247 int max_diff = TestFilter(src_width, src_height,
248 dst_width, dst_height,
249 static_cast<FilterMode>(f), 1,
250 benchmark_iterations_);
251 EXPECT_LE(max_diff, 1);
252 }
253 }
254
TEST_F(libyuvTest,ScaleDownBy38)255 TEST_F(libyuvTest, ScaleDownBy38) {
256 int src_width = 1280;
257 int src_height = 720;
258 int dst_width = src_width * 3 / 8;
259 int dst_height = src_height * 3 / 8;
260
261 for (int f = 0; f < 3; ++f) {
262 int max_diff = TestFilter(src_width, src_height,
263 dst_width, dst_height,
264 static_cast<FilterMode>(f), 1,
265 benchmark_iterations_);
266 EXPECT_LE(max_diff, 1);
267 }
268 }
269
TEST_F(libyuvTest,ScaleTo1366)270 TEST_F(libyuvTest, ScaleTo1366) {
271 int src_width = 1280;
272 int src_height = 720;
273 int dst_width = 1366;
274 int dst_height = 768;
275
276 for (int f = 0; f < 3; ++f) {
277 int max_diff = TestFilter(src_width, src_height,
278 dst_width, dst_height,
279 static_cast<FilterMode>(f), 1,
280 benchmark_iterations_);
281 EXPECT_LE(max_diff, 1);
282 }
283 }
284
TEST_F(libyuvTest,ScaleTo4074)285 TEST_F(libyuvTest, ScaleTo4074) {
286 int src_width = 2880 * 2;
287 int src_height = 1800;
288 int dst_width = 4074;
289 int dst_height = 1272;
290
291 for (int f = 0; f < 3; ++f) {
292 int max_diff = TestFilter(src_width, src_height,
293 dst_width, dst_height,
294 static_cast<FilterMode>(f), 1,
295 benchmark_iterations_);
296 EXPECT_LE(max_diff, 1);
297 }
298 }
299
TEST_F(libyuvTest,ScaleTo853)300 TEST_F(libyuvTest, ScaleTo853) {
301 int src_width = 1280;
302 int src_height = 720;
303 int dst_width = 853;
304 int dst_height = 480;
305
306 for (int f = 0; f < 3; ++f) {
307 int max_diff = TestFilter(src_width, src_height,
308 dst_width, dst_height,
309 static_cast<FilterMode>(f), 1,
310 benchmark_iterations_);
311 EXPECT_LE(max_diff, 1);
312 }
313 }
314
TEST_F(libyuvTest,ScaleTo853Wrong)315 TEST_F(libyuvTest, ScaleTo853Wrong) {
316 int src_width = 1280;
317 int src_height = 720;
318 int dst_width = 853;
319 int dst_height = 480;
320
321 for (int f = 0; f < 3; ++f) {
322 int max_diff = TestFilter(src_width, src_height,
323 dst_width, dst_height,
324 static_cast<FilterMode>(f), 0,
325 benchmark_iterations_);
326 EXPECT_LE(max_diff, 1);
327 }
328 }
329
330 // A one off test for a screen cast resolution scale.
TEST_F(libyuvTest,ScaleTo684)331 TEST_F(libyuvTest, ScaleTo684) {
332 int src_width = 686;
333 int src_height = 557;
334 int dst_width = 684;
335 int dst_height = 552;
336
337 for (int f = 0; f < 3; ++f) {
338 int max_diff = TestFilter(src_width, src_height,
339 dst_width, dst_height,
340 static_cast<FilterMode>(f), 1,
341 benchmark_iterations_);
342 EXPECT_LE(max_diff, 1);
343 }
344 }
345
TEST_F(libyuvTest,ScaleTo342)346 TEST_F(libyuvTest, ScaleTo342) {
347 int src_width = 686;
348 int src_height = 557;
349 int dst_width = 342;
350 int dst_height = 276;
351
352 for (int f = 0; f < 3; ++f) {
353 int max_diff = TestFilter(src_width, src_height,
354 dst_width, dst_height,
355 static_cast<FilterMode>(f), 1,
356 benchmark_iterations_);
357 EXPECT_LE(max_diff, 1);
358 }
359 }
360
TEST_F(libyuvTest,ScaleToHalf342)361 TEST_F(libyuvTest, ScaleToHalf342) {
362 int src_width = 684;
363 int src_height = 552;
364 int dst_width = 342;
365 int dst_height = 276;
366
367 for (int f = 0; f < 3; ++f) {
368 int max_diff = TestFilter(src_width, src_height,
369 dst_width, dst_height,
370 static_cast<FilterMode>(f), 1,
371 benchmark_iterations_);
372 EXPECT_LE(max_diff, 1);
373 }
374 }
375
376 } // namespace libyuv
377