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 // 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)24 static int UVTestFilter(int src_width,
25 int src_height,
26 int dst_width,
27 int dst_height,
28 FilterMode f,
29 int benchmark_iterations,
30 int disable_cpu_flags,
31 int benchmark_cpu_info) {
32 if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
33 return 0;
34 }
35
36 int i, j;
37 const int b = 0; // 128 to test for padding/stride.
38 int64_t src_uv_plane_size =
39 (Abs(src_width) + b * 2) * (Abs(src_height) + b * 2) * 2LL;
40 int src_stride_uv = (b * 2 + Abs(src_width)) * 2;
41
42 align_buffer_page_end(src_uv, src_uv_plane_size);
43 if (!src_uv) {
44 printf("Skipped. Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
45 return 0;
46 }
47 MemRandomize(src_uv, src_uv_plane_size);
48
49 int64_t dst_uv_plane_size = (dst_width + b * 2) * (dst_height + b * 2) * 2LL;
50 int dst_stride_uv = (b * 2 + dst_width) * 2;
51
52 align_buffer_page_end(dst_uv_c, dst_uv_plane_size);
53 align_buffer_page_end(dst_uv_opt, dst_uv_plane_size);
54 if (!dst_uv_c || !dst_uv_opt) {
55 printf("Skipped. Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
56 return 0;
57 }
58 memset(dst_uv_c, 2, dst_uv_plane_size);
59 memset(dst_uv_opt, 3, dst_uv_plane_size);
60
61 // Warm up both versions for consistent benchmarks.
62 MaskCpuFlags(disable_cpu_flags); // Disable all CPU optimization.
63 UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
64 src_height, dst_uv_c + (dst_stride_uv * b) + b * 2, dst_stride_uv,
65 dst_width, dst_height, f);
66 MaskCpuFlags(benchmark_cpu_info); // Enable all CPU optimization.
67 UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
68 src_height, dst_uv_opt + (dst_stride_uv * b) + b * 2, dst_stride_uv,
69 dst_width, dst_height, f);
70
71 MaskCpuFlags(disable_cpu_flags); // Disable all CPU optimization.
72 double c_time = get_time();
73 UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
74 src_height, dst_uv_c + (dst_stride_uv * b) + b * 2, dst_stride_uv,
75 dst_width, dst_height, f);
76
77 c_time = (get_time() - c_time);
78
79 MaskCpuFlags(benchmark_cpu_info); // Enable all CPU optimization.
80 double opt_time = get_time();
81 for (i = 0; i < benchmark_iterations; ++i) {
82 UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
83 src_height, dst_uv_opt + (dst_stride_uv * b) + b * 2, dst_stride_uv,
84 dst_width, dst_height, f);
85 }
86 opt_time = (get_time() - opt_time) / benchmark_iterations;
87
88 // Report performance of C vs OPT
89 printf("filter %d - %8d us C - %8d us OPT\n", f,
90 static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
91
92 // C version may be a little off from the optimized. Order of
93 // operations may introduce rounding somewhere. So do a difference
94 // of the buffers and look to see that the max difference isn't
95 // over 2.
96 int max_diff = 0;
97 for (i = b; i < (dst_height + b); ++i) {
98 for (j = b * 2; j < (dst_width + b) * 2; ++j) {
99 int abs_diff = Abs(dst_uv_c[(i * dst_stride_uv) + j] -
100 dst_uv_opt[(i * dst_stride_uv) + j]);
101 if (abs_diff > max_diff) {
102 max_diff = abs_diff;
103 }
104 }
105 }
106
107 free_aligned_buffer_page_end(dst_uv_c);
108 free_aligned_buffer_page_end(dst_uv_opt);
109 free_aligned_buffer_page_end(src_uv);
110 return max_diff;
111 }
112
113 // The following adjustments in dimensions ensure the scale factor will be
114 // exactly achieved.
115 #define DX(x, nom, denom) static_cast<int>((Abs(x) / nom) * nom)
116 #define SX(x, nom, denom) static_cast<int>((x / nom) * denom)
117
118 #define TEST_FACTOR1(name, filter, nom, denom, max_diff) \
119 TEST_F(LibYUVScaleTest, UVScaleDownBy##name##_##filter) { \
120 int diff = UVTestFilter( \
121 SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
122 DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
123 kFilter##filter, benchmark_iterations_, disable_cpu_flags_, \
124 benchmark_cpu_info_); \
125 EXPECT_LE(diff, max_diff); \
126 }
127
128 // Test a scale factor with all 4 filters. Expect unfiltered to be exact, but
129 // filtering is different fixed point implementations for SSSE3, Neon and C.
130 #define TEST_FACTOR(name, nom, denom) \
131 TEST_FACTOR1(name, None, nom, denom, 0) \
132 TEST_FACTOR1(name, Linear, nom, denom, 3) \
133 TEST_FACTOR1(name, Bilinear, nom, denom, 3) \
134 TEST_FACTOR1(name, Box, nom, denom, 3)
135
136 TEST_FACTOR(2, 1, 2)
137 TEST_FACTOR(4, 1, 4)
138 // TEST_FACTOR(8, 1, 8) Disable for benchmark performance.
139 TEST_FACTOR(3by4, 3, 4)
140 TEST_FACTOR(3by8, 3, 8)
141 TEST_FACTOR(3, 1, 3)
142 #undef TEST_FACTOR1
143 #undef TEST_FACTOR
144 #undef SX
145 #undef DX
146
147 #define TEST_SCALETO1(name, width, height, filter, max_diff) \
148 TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter) { \
149 int diff = UVTestFilter(benchmark_width_, benchmark_height_, width, \
150 height, kFilter##filter, benchmark_iterations_, \
151 disable_cpu_flags_, benchmark_cpu_info_); \
152 EXPECT_LE(diff, max_diff); \
153 } \
154 TEST_F(LibYUVScaleTest, name##From##width##x##height##_##filter) { \
155 int diff = UVTestFilter(width, height, Abs(benchmark_width_), \
156 Abs(benchmark_height_), kFilter##filter, \
157 benchmark_iterations_, disable_cpu_flags_, \
158 benchmark_cpu_info_); \
159 EXPECT_LE(diff, max_diff); \
160 }
161
162 /// Test scale to a specified size with all 4 filters.
163 #define TEST_SCALETO(name, width, height) \
164 TEST_SCALETO1(name, width, height, None, 0) \
165 TEST_SCALETO1(name, width, height, Linear, 3) \
166 TEST_SCALETO1(name, width, height, Bilinear, 3)
167
168 TEST_SCALETO(UVScale, 1, 1)
169 TEST_SCALETO(UVScale, 256, 144) /* 128x72 * 2 */
170 TEST_SCALETO(UVScale, 320, 240)
171 TEST_SCALETO(UVScale, 569, 480)
172 TEST_SCALETO(UVScale, 640, 360)
173 #ifdef ENABLE_SLOW_TESTS
174 TEST_SCALETO(UVScale, 1280, 720)
175 TEST_SCALETO(UVScale, 1920, 1080)
176 #endif // ENABLE_SLOW_TESTS
177 #undef TEST_SCALETO1
178 #undef TEST_SCALETO
179
180 #define TEST_SCALESWAPXY1(name, filter, max_diff) \
181 TEST_F(LibYUVScaleTest, name##SwapXY_##filter) { \
182 int diff = \
183 UVTestFilter(benchmark_width_, benchmark_height_, benchmark_height_, \
184 benchmark_width_, kFilter##filter, benchmark_iterations_, \
185 disable_cpu_flags_, benchmark_cpu_info_); \
186 EXPECT_LE(diff, max_diff); \
187 }
188
189 // Test scale with swapped width and height with all 3 filters.
190 TEST_SCALESWAPXY1(UVScale, None, 0)
191 TEST_SCALESWAPXY1(UVScale, Linear, 0)
192 TEST_SCALESWAPXY1(UVScale, Bilinear, 0)
193 #undef TEST_SCALESWAPXY1
194
TEST_F(LibYUVScaleTest,UVTest3x)195 TEST_F(LibYUVScaleTest, UVTest3x) {
196 const int kSrcStride = 48 * 2;
197 const int kDstStride = 16 * 2;
198 const int kSize = kSrcStride * 3;
199 align_buffer_page_end(orig_pixels, kSize);
200 for (int i = 0; i < 48 * 3; ++i) {
201 orig_pixels[i * 2 + 0] = i;
202 orig_pixels[i * 2 + 1] = 255 - i;
203 }
204 align_buffer_page_end(dest_pixels, kDstStride);
205
206 int iterations16 =
207 benchmark_width_ * benchmark_height_ / (16 * 1) * benchmark_iterations_;
208 for (int i = 0; i < iterations16; ++i) {
209 UVScale(orig_pixels, kSrcStride, 48, 3, dest_pixels, kDstStride, 16, 1,
210 kFilterBilinear);
211 }
212
213 EXPECT_EQ(49, dest_pixels[0]);
214 EXPECT_EQ(255 - 49, dest_pixels[1]);
215
216 UVScale(orig_pixels, kSrcStride, 48, 3, dest_pixels, kDstStride, 16, 1,
217 kFilterNone);
218
219 EXPECT_EQ(49, dest_pixels[0]);
220 EXPECT_EQ(255 - 49, dest_pixels[1]);
221
222 free_aligned_buffer_page_end(dest_pixels);
223 free_aligned_buffer_page_end(orig_pixels);
224 }
225
TEST_F(LibYUVScaleTest,UVTest4x)226 TEST_F(LibYUVScaleTest, UVTest4x) {
227 const int kSrcStride = 64 * 2;
228 const int kDstStride = 16 * 2;
229 const int kSize = kSrcStride * 4;
230 align_buffer_page_end(orig_pixels, kSize);
231 for (int i = 0; i < 64 * 4; ++i) {
232 orig_pixels[i * 2 + 0] = i;
233 orig_pixels[i * 2 + 1] = 255 - i;
234 }
235 align_buffer_page_end(dest_pixels, kDstStride);
236
237 int iterations16 =
238 benchmark_width_ * benchmark_height_ / (16 * 1) * benchmark_iterations_;
239 for (int i = 0; i < iterations16; ++i) {
240 UVScale(orig_pixels, kSrcStride, 64, 4, dest_pixels, kDstStride, 16, 1,
241 kFilterBilinear);
242 }
243
244 EXPECT_EQ((65 + 66 + 129 + 130 + 2) / 4, dest_pixels[0]);
245 EXPECT_EQ((255 - 65 + 255 - 66 + 255 - 129 + 255 - 130 + 2) / 4,
246 dest_pixels[1]);
247
248 UVScale(orig_pixels, kSrcStride, 64, 4, dest_pixels, kDstStride, 16, 1,
249 kFilterNone);
250
251 EXPECT_EQ(130, dest_pixels[0]); // expect the 3rd pixel of the 3rd row
252 EXPECT_EQ(255 - 130, dest_pixels[1]);
253
254 free_aligned_buffer_page_end(dest_pixels);
255 free_aligned_buffer_page_end(orig_pixels);
256 }
257
258 } // namespace libyuv
259