2 * Copyright 2011 The LibYuv Project Authors. All rights reserved.
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.
14 #include "libyuv/cpu_id.h"
15 #include "libyuv/scale.h"
16 #include "../unit_test/unit_test.h"
20 // Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
21 static int TestFilter(int src_width, int src_height,
22 int dst_width, int dst_height,
23 FilterMode f, int benchmark_iterations) {
25 const int b = 0; // 128 to test for padding/stride.
26 int src_width_uv = (Abs(src_width) + 1) >> 1;
27 int src_height_uv = (Abs(src_height) + 1) >> 1;
29 int src_y_plane_size = (Abs(src_width) + b * 2) * (Abs(src_height) + b * 2);
30 int src_uv_plane_size = (src_width_uv + b * 2) * (src_height_uv + b * 2);
32 int src_stride_y = b * 2 + Abs(src_width);
33 int src_stride_uv = b * 2 + src_width_uv;
35 align_buffer_page_end(src_y, src_y_plane_size)
36 align_buffer_page_end(src_u, src_uv_plane_size)
37 align_buffer_page_end(src_v, src_uv_plane_size)
39 MemRandomize(src_y, src_y_plane_size);
40 MemRandomize(src_u, src_uv_plane_size);
41 MemRandomize(src_v, src_uv_plane_size);
43 int dst_width_uv = (dst_width + 1) >> 1;
44 int dst_height_uv = (dst_height + 1) >> 1;
46 int dst_y_plane_size = (dst_width + b * 2) * (dst_height + b * 2);
47 int dst_uv_plane_size = (dst_width_uv + b * 2) * (dst_height_uv + b * 2);
49 int dst_stride_y = b * 2 + dst_width;
50 int dst_stride_uv = b * 2 + dst_width_uv;
52 align_buffer_page_end(dst_y_c, dst_y_plane_size)
53 align_buffer_page_end(dst_u_c, dst_uv_plane_size)
54 align_buffer_page_end(dst_v_c, dst_uv_plane_size)
55 align_buffer_page_end(dst_y_opt, dst_y_plane_size)
56 align_buffer_page_end(dst_u_opt, dst_uv_plane_size)
57 align_buffer_page_end(dst_v_opt, dst_uv_plane_size)
60 MaskCpuFlags(0); // Disable all CPU optimization.
61 double c_time = get_time();
62 I420Scale(src_y + (src_stride_y * b) + b, src_stride_y,
63 src_u + (src_stride_uv * b) + b, src_stride_uv,
64 src_v + (src_stride_uv * b) + b, src_stride_uv,
65 src_width, src_height,
66 dst_y_c + (dst_stride_y * b) + b, dst_stride_y,
67 dst_u_c + (dst_stride_uv * b) + b, dst_stride_uv,
68 dst_v_c + (dst_stride_uv * b) + b, dst_stride_uv,
69 dst_width, dst_height, f);
70 c_time = (get_time() - c_time);
72 MaskCpuFlags(-1); // Enable all CPU optimization.
73 double opt_time = get_time();
74 for (i = 0; i < benchmark_iterations; ++i) {
75 I420Scale(src_y + (src_stride_y * b) + b, src_stride_y,
76 src_u + (src_stride_uv * b) + b, src_stride_uv,
77 src_v + (src_stride_uv * b) + b, src_stride_uv,
78 src_width, src_height,
79 dst_y_opt + (dst_stride_y * b) + b, dst_stride_y,
80 dst_u_opt + (dst_stride_uv * b) + b, dst_stride_uv,
81 dst_v_opt + (dst_stride_uv * b) + b, dst_stride_uv,
82 dst_width, dst_height, f);
84 opt_time = (get_time() - opt_time) / benchmark_iterations;
85 // Report performance of C vs OPT
86 printf("filter %d - %8d us C - %8d us OPT\n",
88 static_cast<int>(c_time * 1e6),
89 static_cast<int>(opt_time * 1e6));
91 // C version may be a little off from the optimized. Order of
92 // operations may introduce rounding somewhere. So do a difference
93 // of the buffers and look to see that the max difference isn't
96 for (i = b; i < (dst_height + b); ++i) {
97 for (j = b; j < (dst_width + b); ++j) {
98 int abs_diff = Abs(dst_y_c[(i * dst_stride_y) + j] -
99 dst_y_opt[(i * dst_stride_y) + j]);
100 if (abs_diff > max_diff) {
106 for (i = b; i < (dst_height_uv + b); ++i) {
107 for (j = b; j < (dst_width_uv + b); ++j) {
108 int abs_diff = Abs(dst_u_c[(i * dst_stride_uv) + j] -
109 dst_u_opt[(i * dst_stride_uv) + j]);
110 if (abs_diff > max_diff) {
113 abs_diff = Abs(dst_v_c[(i * dst_stride_uv) + j] -
114 dst_v_opt[(i * dst_stride_uv) + j]);
115 if (abs_diff > max_diff) {
121 free_aligned_buffer_page_end(dst_y_c)
122 free_aligned_buffer_page_end(dst_u_c)
123 free_aligned_buffer_page_end(dst_v_c)
124 free_aligned_buffer_page_end(dst_y_opt)
125 free_aligned_buffer_page_end(dst_u_opt)
126 free_aligned_buffer_page_end(dst_v_opt)
128 free_aligned_buffer_page_end(src_y)
129 free_aligned_buffer_page_end(src_u)
130 free_aligned_buffer_page_end(src_v)
135 // Test scaling with 8 bit C vs 16 bit C and return maximum pixel difference.
137 static int TestFilter_16(int src_width, int src_height,
138 int dst_width, int dst_height,
139 FilterMode f, int benchmark_iterations) {
141 const int b = 0; // 128 to test for padding/stride.
142 int src_width_uv = (Abs(src_width) + 1) >> 1;
143 int src_height_uv = (Abs(src_height) + 1) >> 1;
145 int src_y_plane_size = (Abs(src_width) + b * 2) * (Abs(src_height) + b * 2);
146 int src_uv_plane_size = (src_width_uv + b * 2) * (src_height_uv + b * 2);
148 int src_stride_y = b * 2 + Abs(src_width);
149 int src_stride_uv = b * 2 + src_width_uv;
151 align_buffer_page_end(src_y, src_y_plane_size)
152 align_buffer_page_end(src_u, src_uv_plane_size)
153 align_buffer_page_end(src_v, src_uv_plane_size)
154 align_buffer_page_end(src_y_16, src_y_plane_size * 2)
155 align_buffer_page_end(src_u_16, src_uv_plane_size * 2)
156 align_buffer_page_end(src_v_16, src_uv_plane_size * 2)
157 uint16* p_src_y_16 = reinterpret_cast<uint16*>(src_y_16);
158 uint16* p_src_u_16 = reinterpret_cast<uint16*>(src_u_16);
159 uint16* p_src_v_16 = reinterpret_cast<uint16*>(src_v_16);
162 MemRandomize(src_y, src_y_plane_size);
163 MemRandomize(src_u, src_uv_plane_size);
164 MemRandomize(src_v, src_uv_plane_size);
166 for (i = b; i < src_height + b; ++i) {
167 for (j = b; j < src_width + b; ++j) {
168 p_src_y_16[(i * src_stride_y) + j] = src_y[(i * src_stride_y) + j];
172 for (i = b; i < (src_height_uv + b); ++i) {
173 for (j = b; j < (src_width_uv + b); ++j) {
174 p_src_u_16[(i * src_stride_uv) + j] = src_u[(i * src_stride_uv) + j];
175 p_src_v_16[(i * src_stride_uv) + j] = src_v[(i * src_stride_uv) + j];
179 int dst_width_uv = (dst_width + 1) >> 1;
180 int dst_height_uv = (dst_height + 1) >> 1;
182 int dst_y_plane_size = (dst_width + b * 2) * (dst_height + b * 2);
183 int dst_uv_plane_size = (dst_width_uv + b * 2) * (dst_height_uv + b * 2);
185 int dst_stride_y = b * 2 + dst_width;
186 int dst_stride_uv = b * 2 + dst_width_uv;
188 align_buffer_page_end(dst_y_8, dst_y_plane_size)
189 align_buffer_page_end(dst_u_8, dst_uv_plane_size)
190 align_buffer_page_end(dst_v_8, dst_uv_plane_size)
191 align_buffer_page_end(dst_y_16, dst_y_plane_size * 2)
192 align_buffer_page_end(dst_u_16, dst_uv_plane_size * 2)
193 align_buffer_page_end(dst_v_16, dst_uv_plane_size * 2)
195 uint16* p_dst_y_16 = reinterpret_cast<uint16*>(dst_y_16);
196 uint16* p_dst_u_16 = reinterpret_cast<uint16*>(dst_u_16);
197 uint16* p_dst_v_16 = reinterpret_cast<uint16*>(dst_v_16);
199 I420Scale(src_y + (src_stride_y * b) + b, src_stride_y,
200 src_u + (src_stride_uv * b) + b, src_stride_uv,
201 src_v + (src_stride_uv * b) + b, src_stride_uv,
202 src_width, src_height,
203 dst_y_8 + (dst_stride_y * b) + b, dst_stride_y,
204 dst_u_8 + (dst_stride_uv * b) + b, dst_stride_uv,
205 dst_v_8 + (dst_stride_uv * b) + b, dst_stride_uv,
206 dst_width, dst_height, f);
208 for (i = 0; i < benchmark_iterations; ++i) {
209 I420Scale_16(p_src_y_16 + (src_stride_y * b) + b, src_stride_y,
210 p_src_u_16 + (src_stride_uv * b) + b, src_stride_uv,
211 p_src_v_16 + (src_stride_uv * b) + b, src_stride_uv,
212 src_width, src_height,
213 p_dst_y_16 + (dst_stride_y * b) + b, dst_stride_y,
214 p_dst_u_16 + (dst_stride_uv * b) + b, dst_stride_uv,
215 p_dst_v_16 + (dst_stride_uv * b) + b, dst_stride_uv,
216 dst_width, dst_height, f);
219 // Expect an exact match
221 for (i = b; i < (dst_height + b); ++i) {
222 for (j = b; j < (dst_width + b); ++j) {
223 int abs_diff = Abs(dst_y_8[(i * dst_stride_y) + j] -
224 p_dst_y_16[(i * dst_stride_y) + j]);
225 if (abs_diff > max_diff) {
231 for (i = b; i < (dst_height_uv + b); ++i) {
232 for (j = b; j < (dst_width_uv + b); ++j) {
233 int abs_diff = Abs(dst_u_8[(i * dst_stride_uv) + j] -
234 p_dst_u_16[(i * dst_stride_uv) + j]);
235 if (abs_diff > max_diff) {
238 abs_diff = Abs(dst_v_8[(i * dst_stride_uv) + j] -
239 p_dst_v_16[(i * dst_stride_uv) + j]);
240 if (abs_diff > max_diff) {
246 free_aligned_buffer_page_end(dst_y_8)
247 free_aligned_buffer_page_end(dst_u_8)
248 free_aligned_buffer_page_end(dst_v_8)
249 free_aligned_buffer_page_end(dst_y_16)
250 free_aligned_buffer_page_end(dst_u_16)
251 free_aligned_buffer_page_end(dst_v_16)
253 free_aligned_buffer_page_end(src_y)
254 free_aligned_buffer_page_end(src_u)
255 free_aligned_buffer_page_end(src_v)
256 free_aligned_buffer_page_end(src_y_16)
257 free_aligned_buffer_page_end(src_u_16)
258 free_aligned_buffer_page_end(src_v_16)
263 #define TEST_FACTOR1(name, filter, hfactor, vfactor, max_diff) \
264 TEST_F(libyuvTest, ScaleDownBy##name##_##filter) { \
265 int diff = TestFilter(benchmark_width_, benchmark_height_, \
266 Abs(benchmark_width_) * hfactor, \
267 Abs(benchmark_height_) * vfactor, \
268 kFilter##filter, benchmark_iterations_); \
269 EXPECT_LE(diff, max_diff); \
271 TEST_F(libyuvTest, ScaleDownBy##name##_##filter##_16) { \
272 int diff = TestFilter_16(benchmark_width_, benchmark_height_, \
273 Abs(benchmark_width_) * hfactor, \
274 Abs(benchmark_height_) * vfactor, \
275 kFilter##filter, benchmark_iterations_); \
276 EXPECT_LE(diff, max_diff); \
279 // Test a scale factor with all 4 filters. Expect unfiltered to be exact, but
280 // filtering is different fixed point implementations for SSSE3, Neon and C.
281 #define TEST_FACTOR(name, hfactor, vfactor) \
282 TEST_FACTOR1(name, None, hfactor, vfactor, 0) \
283 TEST_FACTOR1(name, Linear, hfactor, vfactor, 3) \
284 TEST_FACTOR1(name, Bilinear, hfactor, vfactor, 3) \
285 TEST_FACTOR1(name, Box, hfactor, vfactor, 3) \
287 TEST_FACTOR(2, 1 / 2, 1 / 2)
288 TEST_FACTOR(4, 1 / 4, 1 / 4)
289 TEST_FACTOR(8, 1 / 8, 1 / 8)
290 TEST_FACTOR(3by4, 3 / 4, 3 / 4)
291 TEST_FACTOR(3by8, 3 / 8, 3 / 8)
295 #define TEST_SCALETO1(name, width, height, filter, max_diff) \
296 TEST_F(libyuvTest, name##To##width##x##height##_##filter) { \
297 int diff = TestFilter(benchmark_width_, benchmark_height_, \
299 kFilter##filter, benchmark_iterations_); \
300 EXPECT_LE(diff, max_diff); \
302 TEST_F(libyuvTest, name##From##width##x##height##_##filter) { \
303 int diff = TestFilter(width, height, \
304 Abs(benchmark_width_), Abs(benchmark_height_), \
305 kFilter##filter, benchmark_iterations_); \
306 EXPECT_LE(diff, max_diff); \
308 TEST_F(libyuvTest, name##To##width##x##height##_##filter##_16) { \
309 int diff = TestFilter_16(benchmark_width_, benchmark_height_, \
311 kFilter##filter, benchmark_iterations_); \
312 EXPECT_LE(diff, max_diff); \
314 TEST_F(libyuvTest, name##From##width##x##height##_##filter##_16) { \
315 int diff = TestFilter_16(width, height, \
316 Abs(benchmark_width_), Abs(benchmark_height_), \
317 kFilter##filter, benchmark_iterations_); \
318 EXPECT_LE(diff, max_diff); \
321 // Test scale to a specified size with all 4 filters.
322 #define TEST_SCALETO(name, width, height) \
323 TEST_SCALETO1(name, width, height, None, 0) \
324 TEST_SCALETO1(name, width, height, Linear, 3) \
325 TEST_SCALETO1(name, width, height, Bilinear, 3) \
326 TEST_SCALETO1(name, width, height, Box, 3)
328 TEST_SCALETO(Scale, 1, 1)
329 TEST_SCALETO(Scale, 320, 240)
330 TEST_SCALETO(Scale, 352, 288)
331 TEST_SCALETO(Scale, 569, 480)
332 TEST_SCALETO(Scale, 640, 360)
333 TEST_SCALETO(Scale, 1280, 720)
337 } // namespace libyuv