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 "../unit_test/unit_test.h"
15 #include "libyuv/cpu_id.h"
16 #include "libyuv/scale_uv.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.
UVTestFilter(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 UVTestFilter(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_uv_plane_size =
45 (Abs(src_width) + b * 2) * (Abs(src_height) + b * 2) * 2LL;
46 int src_stride_uv = (b * 2 + Abs(src_width)) * 2;
47
48 align_buffer_page_end(src_uv, src_uv_plane_size);
49 if (!src_uv) {
50 printf("Skipped. Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
51 return 0;
52 }
53 MemRandomize(src_uv, src_uv_plane_size);
54
55 int64_t dst_uv_plane_size = (dst_width + b * 2) * (dst_height + b * 2) * 2LL;
56 int dst_stride_uv = (b * 2 + dst_width) * 2;
57
58 align_buffer_page_end(dst_uv_c, dst_uv_plane_size);
59 align_buffer_page_end(dst_uv_opt, dst_uv_plane_size);
60 if (!dst_uv_c || !dst_uv_opt) {
61 printf("Skipped. Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
62 return 0;
63 }
64 memset(dst_uv_c, 2, dst_uv_plane_size);
65 memset(dst_uv_opt, 3, dst_uv_plane_size);
66
67 // Warm up both versions for consistent benchmarks.
68 MaskCpuFlags(disable_cpu_flags); // Disable all CPU optimization.
69 UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
70 src_height, dst_uv_c + (dst_stride_uv * b) + b * 2, dst_stride_uv,
71 dst_width, dst_height, f);
72 MaskCpuFlags(benchmark_cpu_info); // Enable all CPU optimization.
73 UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
74 src_height, dst_uv_opt + (dst_stride_uv * b) + b * 2, dst_stride_uv,
75 dst_width, dst_height, f);
76
77 MaskCpuFlags(disable_cpu_flags); // Disable all CPU optimization.
78 double c_time = get_time();
79 UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
80 src_height, dst_uv_c + (dst_stride_uv * b) + b * 2, dst_stride_uv,
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 UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
89 src_height, dst_uv_opt + (dst_stride_uv * b) + b * 2, dst_stride_uv,
90 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 * 2; j < (dst_width + b) * 2; ++j) {
105 int abs_diff = Abs(dst_uv_c[(i * dst_stride_uv) + j] -
106 dst_uv_opt[(i * dst_stride_uv) + j]);
107 if (abs_diff > max_diff) {
108 max_diff = abs_diff;
109 }
110 }
111 }
112
113 free_aligned_buffer_page_end(dst_uv_c);
114 free_aligned_buffer_page_end(dst_uv_opt);
115 free_aligned_buffer_page_end(src_uv);
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, UVScaleDownBy##name##_##filter) { \
126 int diff = UVTestFilter( \
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 TEST_FACTOR(4, 1, 4)
150 // TEST_FACTOR(8, 1, 8) Disable for benchmark performance.
151 TEST_FACTOR(3by4, 3, 4)
152 TEST_FACTOR(3by8, 3, 8)
153 TEST_FACTOR(3, 1, 3)
154 #undef TEST_FACTOR1
155 #undef TEST_FACTOR
156 #undef SX
157 #undef DX
158
159 #define TEST_SCALETO1(name, width, height, filter, max_diff) \
160 TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter) { \
161 int diff = UVTestFilter(benchmark_width_, benchmark_height_, width, \
162 height, kFilter##filter, benchmark_iterations_, \
163 disable_cpu_flags_, benchmark_cpu_info_); \
164 EXPECT_LE(diff, max_diff); \
165 } \
166 TEST_F(LibYUVScaleTest, name##From##width##x##height##_##filter) { \
167 int diff = UVTestFilter(width, height, Abs(benchmark_width_), \
168 Abs(benchmark_height_), kFilter##filter, \
169 benchmark_iterations_, disable_cpu_flags_, \
170 benchmark_cpu_info_); \
171 EXPECT_LE(diff, max_diff); \
172 }
173
174 #if defined(ENABLE_FULL_TESTS)
175 /// Test scale to a specified size with all 4 filters.
176 #define TEST_SCALETO(name, width, height) \
177 TEST_SCALETO1(name, width, height, None, 0) \
178 TEST_SCALETO1(name, width, height, Linear, 3) \
179 TEST_SCALETO1(name, width, height, Bilinear, 3)
180 #else
181 #define TEST_SCALETO(name, width, height) \
182 TEST_SCALETO1(name, width, height, Bilinear, 3)
183 #endif
184
185 TEST_SCALETO(UVScale, 1, 1)
186 TEST_SCALETO(UVScale, 569, 480)
187 TEST_SCALETO(UVScale, 640, 360)
188 #ifndef DISABLE_SLOW_TESTS
189 TEST_SCALETO(UVScale, 256, 144) /* 128x72 * 2 */
190 TEST_SCALETO(UVScale, 320, 240)
191 TEST_SCALETO(UVScale, 1280, 720)
192 TEST_SCALETO(UVScale, 1920, 1080)
193 #endif // DISABLE_SLOW_TESTS
194 #undef TEST_SCALETO1
195 #undef TEST_SCALETO
196
197 #define TEST_SCALESWAPXY1(name, filter, max_diff) \
198 TEST_F(LibYUVScaleTest, name##SwapXY_##filter) { \
199 int diff = \
200 UVTestFilter(benchmark_width_, benchmark_height_, benchmark_height_, \
201 benchmark_width_, kFilter##filter, benchmark_iterations_, \
202 disable_cpu_flags_, benchmark_cpu_info_); \
203 EXPECT_LE(diff, max_diff); \
204 }
205
206 #if defined(ENABLE_FULL_TESTS)
207 // Test scale with swapped width and height with all 3 filters.
208 TEST_SCALESWAPXY1(UVScale, None, 0)
209 TEST_SCALESWAPXY1(UVScale, Linear, 0)
210 TEST_SCALESWAPXY1(UVScale, Bilinear, 0)
211 #else
212 TEST_SCALESWAPXY1(UVScale, Bilinear, 0)
213 #endif
214 #undef TEST_SCALESWAPXY1
215
TEST_F(LibYUVScaleTest,UVTest3x)216 TEST_F(LibYUVScaleTest, UVTest3x) {
217 const int kSrcStride = 480 * 2;
218 const int kDstStride = 160 * 2;
219 const int kSize = kSrcStride * 3;
220 align_buffer_page_end(orig_pixels, kSize);
221 for (int i = 0; i < 480 * 3; ++i) {
222 orig_pixels[i * 2 + 0] = i;
223 orig_pixels[i * 2 + 1] = 255 - i;
224 }
225 align_buffer_page_end(dest_pixels, kDstStride);
226
227 int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
228 benchmark_iterations_;
229 for (int i = 0; i < iterations160; ++i) {
230 UVScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
231 kFilterBilinear);
232 }
233
234 EXPECT_EQ(225, dest_pixels[0]);
235 EXPECT_EQ(255 - 225, dest_pixels[1]);
236
237 UVScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
238 kFilterNone);
239
240 EXPECT_EQ(225, dest_pixels[0]);
241 EXPECT_EQ(255 - 225, dest_pixels[1]);
242
243 free_aligned_buffer_page_end(dest_pixels);
244 free_aligned_buffer_page_end(orig_pixels);
245 }
246
TEST_F(LibYUVScaleTest,UVTest4x)247 TEST_F(LibYUVScaleTest, UVTest4x) {
248 const int kSrcStride = 640 * 2;
249 const int kDstStride = 160 * 2;
250 const int kSize = kSrcStride * 4;
251 align_buffer_page_end(orig_pixels, kSize);
252 for (int i = 0; i < 640 * 4; ++i) {
253 orig_pixels[i * 2 + 0] = i;
254 orig_pixels[i * 2 + 1] = 255 - i;
255 }
256 align_buffer_page_end(dest_pixels, kDstStride);
257
258 int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
259 benchmark_iterations_;
260 for (int i = 0; i < iterations160; ++i) {
261 UVScale(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
262 kFilterBilinear);
263 }
264
265 EXPECT_EQ(66, dest_pixels[0]);
266 EXPECT_EQ(190, dest_pixels[1]);
267
268 UVScale(orig_pixels, kSrcStride, 64, 4, dest_pixels, kDstStride, 16, 1,
269 kFilterNone);
270
271 EXPECT_EQ(2, dest_pixels[0]); // expect the 3rd pixel of the 3rd row
272 EXPECT_EQ(255 - 2, dest_pixels[1]);
273
274 free_aligned_buffer_page_end(dest_pixels);
275 free_aligned_buffer_page_end(orig_pixels);
276 }
277
278 } // namespace libyuv
279