| OLD | NEW |
| 1 // Copyright 2014 The Chromium Authors. All rights reserved. | 1 // Copyright 2014 The Chromium Authors. All rights reserved. |
| 2 // Use of this source code is governed by a BSD-style license that can be | 2 // Use of this source code is governed by a BSD-style license that can be |
| 3 // found in the LICENSE file. | 3 // found in the LICENSE file. |
| 4 | 4 |
| 5 #include "base/base_paths.h" | 5 #include "base/base_paths.h" |
| 6 #include "base/cpu.h" | 6 #include "base/cpu.h" |
| 7 #include "base/file_util.h" | 7 #include "base/file_util.h" |
| 8 #include "base/logging.h" | 8 #include "base/logging.h" |
| 9 #include "base/path_service.h" | 9 #include "base/path_service.h" |
| 10 #include "base/time/time.h" | 10 #include "base/time/time.h" |
| 11 #include "media/base/simd/convert_yuv_to_rgb.h" | 11 #include "media/base/simd/convert_yuv_to_rgb.h" |
| 12 #include "media/base/simd/yuv_to_rgb_table.h" | |
| 13 #include "media/base/yuv_convert.h" | 12 #include "media/base/yuv_convert.h" |
| 14 #include "testing/gtest/include/gtest/gtest.h" | 13 #include "testing/gtest/include/gtest/gtest.h" |
| 15 #include "testing/perf/perf_test.h" | 14 #include "testing/perf/perf_test.h" |
| 16 | 15 |
| 17 namespace media { | 16 namespace media { |
| 18 #if !defined(ARCH_CPU_ARM_FAMILY) && !defined(ARCH_CPU_MIPS_FAMILY) | 17 #if !defined(ARCH_CPU_ARM_FAMILY) && !defined(ARCH_CPU_MIPS_FAMILY) |
| 19 // Size of raw image. | 18 // Size of raw image. |
| 20 static const int kSourceWidth = 640; | 19 static const int kSourceWidth = 640; |
| 21 static const int kSourceHeight = 360; | 20 static const int kSourceHeight = 360; |
| 22 static const int kSourceYSize = kSourceWidth * kSourceHeight; | 21 static const int kSourceYSize = kSourceWidth * kSourceHeight; |
| (...skipping 47 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 70 | 69 |
| 71 base::TimeTicks start = base::TimeTicks::HighResNow(); | 70 base::TimeTicks start = base::TimeTicks::HighResNow(); |
| 72 for (int i = 0; i < kPerfTestIterations; ++i) { | 71 for (int i = 0; i < kPerfTestIterations; ++i) { |
| 73 for (int row = 0; row < kSourceHeight; ++row) { | 72 for (int row = 0; row < kSourceHeight; ++row) { |
| 74 int chroma_row = row / 2; | 73 int chroma_row = row / 2; |
| 75 ConvertYUVToRGB32Row_MMX( | 74 ConvertYUVToRGB32Row_MMX( |
| 76 yuv_bytes_.get() + row * kSourceWidth, | 75 yuv_bytes_.get() + row * kSourceWidth, |
| 77 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), | 76 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), |
| 78 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), | 77 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), |
| 79 rgb_bytes_converted_.get(), | 78 rgb_bytes_converted_.get(), |
| 80 kWidth, | 79 kWidth); |
| 81 kCoefficientsRgbY); | |
| 82 } | 80 } |
| 83 } | 81 } |
| 84 double total_time_seconds = | 82 double total_time_seconds = |
| 85 (base::TimeTicks::HighResNow() - start).InSecondsF(); | 83 (base::TimeTicks::HighResNow() - start).InSecondsF(); |
| 86 perf_test::PrintResult( | 84 perf_test::PrintResult( |
| 87 "yuv_convert_perftest", "", "ConvertYUVToRGB32Row_MMX", | 85 "yuv_convert_perftest", "", "ConvertYUVToRGB32Row_MMX", |
| 88 kPerfTestIterations / total_time_seconds, "runs/s", true); | 86 kPerfTestIterations / total_time_seconds, "runs/s", true); |
| 89 | 87 |
| 90 media::EmptyRegisterState(); | 88 media::EmptyRegisterState(); |
| 91 } | 89 } |
| 92 | 90 |
| 93 TEST_F(YUVConvertPerfTest, ConvertYUVToRGB32Row_SSE) { | 91 TEST_F(YUVConvertPerfTest, ConvertYUVToRGB32Row_SSE) { |
| 94 ASSERT_TRUE(base::CPU().has_sse()); | 92 ASSERT_TRUE(base::CPU().has_sse()); |
| 95 | 93 |
| 96 base::TimeTicks start = base::TimeTicks::HighResNow(); | 94 base::TimeTicks start = base::TimeTicks::HighResNow(); |
| 97 for (int i = 0; i < kPerfTestIterations; ++i) { | 95 for (int i = 0; i < kPerfTestIterations; ++i) { |
| 98 for (int row = 0; row < kSourceHeight; ++row) { | 96 for (int row = 0; row < kSourceHeight; ++row) { |
| 99 int chroma_row = row / 2; | 97 int chroma_row = row / 2; |
| 100 ConvertYUVToRGB32Row_SSE( | 98 ConvertYUVToRGB32Row_SSE( |
| 101 yuv_bytes_.get() + row * kSourceWidth, | 99 yuv_bytes_.get() + row * kSourceWidth, |
| 102 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), | 100 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), |
| 103 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), | 101 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), |
| 104 rgb_bytes_converted_.get(), | 102 rgb_bytes_converted_.get(), |
| 105 kWidth, | 103 kWidth); |
| 106 kCoefficientsRgbY); | |
| 107 } | 104 } |
| 108 } | 105 } |
| 109 double total_time_seconds = | 106 double total_time_seconds = |
| 110 (base::TimeTicks::HighResNow() - start).InSecondsF(); | 107 (base::TimeTicks::HighResNow() - start).InSecondsF(); |
| 111 perf_test::PrintResult( | 108 perf_test::PrintResult( |
| 112 "yuv_convert_perftest", "", "ConvertYUVToRGB32Row_SSE", | 109 "yuv_convert_perftest", "", "ConvertYUVToRGB32Row_SSE", |
| 113 kPerfTestIterations / total_time_seconds, "runs/s", true); | 110 kPerfTestIterations / total_time_seconds, "runs/s", true); |
| 114 media::EmptyRegisterState(); | 111 media::EmptyRegisterState(); |
| 115 } | 112 } |
| 116 | 113 |
| 117 TEST_F(YUVConvertPerfTest, ScaleYUVToRGB32Row_MMX) { | 114 TEST_F(YUVConvertPerfTest, ScaleYUVToRGB32Row_MMX) { |
| 118 ASSERT_TRUE(base::CPU().has_mmx()); | 115 ASSERT_TRUE(base::CPU().has_mmx()); |
| 119 | 116 |
| 120 const int kSourceDx = 80000; // This value means a scale down. | 117 const int kSourceDx = 80000; // This value means a scale down. |
| 121 | 118 |
| 122 base::TimeTicks start = base::TimeTicks::HighResNow(); | 119 base::TimeTicks start = base::TimeTicks::HighResNow(); |
| 123 for (int i = 0; i < kPerfTestIterations; ++i) { | 120 for (int i = 0; i < kPerfTestIterations; ++i) { |
| 124 for (int row = 0; row < kSourceHeight; ++row) { | 121 for (int row = 0; row < kSourceHeight; ++row) { |
| 125 int chroma_row = row / 2; | 122 int chroma_row = row / 2; |
| 126 ScaleYUVToRGB32Row_MMX( | 123 ScaleYUVToRGB32Row_MMX( |
| 127 yuv_bytes_.get() + row * kSourceWidth, | 124 yuv_bytes_.get() + row * kSourceWidth, |
| 128 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), | 125 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), |
| 129 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), | 126 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), |
| 130 rgb_bytes_converted_.get(), | 127 rgb_bytes_converted_.get(), |
| 131 kWidth, | 128 kWidth, |
| 132 kSourceDx, | 129 kSourceDx); |
| 133 kCoefficientsRgbY); | |
| 134 } | 130 } |
| 135 } | 131 } |
| 136 double total_time_seconds = | 132 double total_time_seconds = |
| 137 (base::TimeTicks::HighResNow() - start).InSecondsF(); | 133 (base::TimeTicks::HighResNow() - start).InSecondsF(); |
| 138 perf_test::PrintResult( | 134 perf_test::PrintResult( |
| 139 "yuv_convert_perftest", "", "ScaleYUVToRGB32Row_MMX", | 135 "yuv_convert_perftest", "", "ScaleYUVToRGB32Row_MMX", |
| 140 kPerfTestIterations / total_time_seconds, "runs/s", true); | 136 kPerfTestIterations / total_time_seconds, "runs/s", true); |
| 141 media::EmptyRegisterState(); | 137 media::EmptyRegisterState(); |
| 142 } | 138 } |
| 143 | 139 |
| 144 TEST_F(YUVConvertPerfTest, ScaleYUVToRGB32Row_SSE) { | 140 TEST_F(YUVConvertPerfTest, ScaleYUVToRGB32Row_SSE) { |
| 145 ASSERT_TRUE(base::CPU().has_sse()); | 141 ASSERT_TRUE(base::CPU().has_sse()); |
| 146 | 142 |
| 147 const int kSourceDx = 80000; // This value means a scale down. | 143 const int kSourceDx = 80000; // This value means a scale down. |
| 148 | 144 |
| 149 base::TimeTicks start = base::TimeTicks::HighResNow(); | 145 base::TimeTicks start = base::TimeTicks::HighResNow(); |
| 150 for (int i = 0; i < kPerfTestIterations; ++i) { | 146 for (int i = 0; i < kPerfTestIterations; ++i) { |
| 151 for (int row = 0; row < kSourceHeight; ++row) { | 147 for (int row = 0; row < kSourceHeight; ++row) { |
| 152 int chroma_row = row / 2; | 148 int chroma_row = row / 2; |
| 153 ScaleYUVToRGB32Row_SSE( | 149 ScaleYUVToRGB32Row_SSE( |
| 154 yuv_bytes_.get() + row * kSourceWidth, | 150 yuv_bytes_.get() + row * kSourceWidth, |
| 155 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), | 151 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), |
| 156 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), | 152 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), |
| 157 rgb_bytes_converted_.get(), | 153 rgb_bytes_converted_.get(), |
| 158 kWidth, | 154 kWidth, |
| 159 kSourceDx, | 155 kSourceDx); |
| 160 kCoefficientsRgbY); | |
| 161 } | 156 } |
| 162 } | 157 } |
| 163 double total_time_seconds = | 158 double total_time_seconds = |
| 164 (base::TimeTicks::HighResNow() - start).InSecondsF(); | 159 (base::TimeTicks::HighResNow() - start).InSecondsF(); |
| 165 perf_test::PrintResult( | 160 perf_test::PrintResult( |
| 166 "yuv_convert_perftest", "", "ScaleYUVToRGB32Row_SSE", | 161 "yuv_convert_perftest", "", "ScaleYUVToRGB32Row_SSE", |
| 167 kPerfTestIterations / total_time_seconds, "runs/s", true); | 162 kPerfTestIterations / total_time_seconds, "runs/s", true); |
| 168 media::EmptyRegisterState(); | 163 media::EmptyRegisterState(); |
| 169 } | 164 } |
| 170 | 165 |
| 171 TEST_F(YUVConvertPerfTest, LinearScaleYUVToRGB32Row_MMX) { | 166 TEST_F(YUVConvertPerfTest, LinearScaleYUVToRGB32Row_MMX) { |
| 172 ASSERT_TRUE(base::CPU().has_mmx()); | 167 ASSERT_TRUE(base::CPU().has_mmx()); |
| 173 | 168 |
| 174 const int kSourceDx = 80000; // This value means a scale down. | 169 const int kSourceDx = 80000; // This value means a scale down. |
| 175 | 170 |
| 176 base::TimeTicks start = base::TimeTicks::HighResNow(); | 171 base::TimeTicks start = base::TimeTicks::HighResNow(); |
| 177 for (int i = 0; i < kPerfTestIterations; ++i) { | 172 for (int i = 0; i < kPerfTestIterations; ++i) { |
| 178 for (int row = 0; row < kSourceHeight; ++row) { | 173 for (int row = 0; row < kSourceHeight; ++row) { |
| 179 int chroma_row = row / 2; | 174 int chroma_row = row / 2; |
| 180 LinearScaleYUVToRGB32Row_MMX( | 175 LinearScaleYUVToRGB32Row_MMX( |
| 181 yuv_bytes_.get() + row * kSourceWidth, | 176 yuv_bytes_.get() + row * kSourceWidth, |
| 182 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), | 177 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), |
| 183 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), | 178 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), |
| 184 rgb_bytes_converted_.get(), | 179 rgb_bytes_converted_.get(), |
| 185 kWidth, | 180 kWidth, |
| 186 kSourceDx, | 181 kSourceDx); |
| 187 kCoefficientsRgbY); | |
| 188 } | 182 } |
| 189 } | 183 } |
| 190 double total_time_seconds = | 184 double total_time_seconds = |
| 191 (base::TimeTicks::HighResNow() - start).InSecondsF(); | 185 (base::TimeTicks::HighResNow() - start).InSecondsF(); |
| 192 perf_test::PrintResult( | 186 perf_test::PrintResult( |
| 193 "yuv_convert_perftest", "", "LinearScaleYUVToRGB32Row_MMX", | 187 "yuv_convert_perftest", "", "LinearScaleYUVToRGB32Row_MMX", |
| 194 kPerfTestIterations / total_time_seconds, "runs/s", true); | 188 kPerfTestIterations / total_time_seconds, "runs/s", true); |
| 195 media::EmptyRegisterState(); | 189 media::EmptyRegisterState(); |
| 196 } | 190 } |
| 197 | 191 |
| 198 TEST_F(YUVConvertPerfTest, LinearScaleYUVToRGB32Row_SSE) { | 192 TEST_F(YUVConvertPerfTest, LinearScaleYUVToRGB32Row_SSE) { |
| 199 ASSERT_TRUE(base::CPU().has_sse()); | 193 ASSERT_TRUE(base::CPU().has_sse()); |
| 200 | 194 |
| 201 const int kSourceDx = 80000; // This value means a scale down. | 195 const int kSourceDx = 80000; // This value means a scale down. |
| 202 | 196 |
| 203 base::TimeTicks start = base::TimeTicks::HighResNow(); | 197 base::TimeTicks start = base::TimeTicks::HighResNow(); |
| 204 for (int i = 0; i < kPerfTestIterations; ++i) { | 198 for (int i = 0; i < kPerfTestIterations; ++i) { |
| 205 for (int row = 0; row < kSourceHeight; ++row) { | 199 for (int row = 0; row < kSourceHeight; ++row) { |
| 206 int chroma_row = row / 2; | 200 int chroma_row = row / 2; |
| 207 LinearScaleYUVToRGB32Row_SSE( | 201 LinearScaleYUVToRGB32Row_SSE( |
| 208 yuv_bytes_.get() + row * kSourceWidth, | 202 yuv_bytes_.get() + row * kSourceWidth, |
| 209 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), | 203 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), |
| 210 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), | 204 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), |
| 211 rgb_bytes_converted_.get(), | 205 rgb_bytes_converted_.get(), |
| 212 kWidth, | 206 kWidth, |
| 213 kSourceDx, | 207 kSourceDx); |
| 214 kCoefficientsRgbY); | |
| 215 } | 208 } |
| 216 } | 209 } |
| 217 double total_time_seconds = | 210 double total_time_seconds = |
| 218 (base::TimeTicks::HighResNow() - start).InSecondsF(); | 211 (base::TimeTicks::HighResNow() - start).InSecondsF(); |
| 219 perf_test::PrintResult( | 212 perf_test::PrintResult( |
| 220 "yuv_convert_perftest", "", "LinearScaleYUVToRGB32Row_SSE", | 213 "yuv_convert_perftest", "", "LinearScaleYUVToRGB32Row_SSE", |
| 221 kPerfTestIterations / total_time_seconds, "runs/s", true); | 214 kPerfTestIterations / total_time_seconds, "runs/s", true); |
| 222 media::EmptyRegisterState(); | 215 media::EmptyRegisterState(); |
| 223 } | 216 } |
| 224 | 217 |
| 225 #endif // !defined(ARCH_CPU_ARM_FAMILY) && !defined(ARCH_CPU_MIPS_FAMILY) | 218 #endif // !defined(ARCH_CPU_ARM_FAMILY) && !defined(ARCH_CPU_MIPS_FAMILY) |
| 226 | 219 |
| 227 } // namespace media | 220 } // namespace media |
| OLD | NEW |