| 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/files/file_util.h" | 7 #include "base/files/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/yuv_convert.h" | 12 #include "media/base/yuv_convert.h" |
| 13 #include "testing/gtest/include/gtest/gtest.h" | 13 #include "testing/gtest/include/gtest/gtest.h" |
| 14 #include "testing/perf/perf_test.h" | 14 #include "testing/perf/perf_test.h" |
| 15 #include "third_party/libyuv/include/libyuv/row.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; |
| 23 static const int kSourceUOffset = kSourceYSize; | 22 static const int kSourceUOffset = kSourceYSize; |
| 24 static const int kSourceVOffset = kSourceYSize * 5 / 4; | 23 static const int kSourceVOffset = kSourceYSize * 5 / 4; |
| 25 static const int kBpp = 4; | 24 static const int kBpp = 4; |
| (...skipping 32 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 58 CHECK_EQ(bytes_read, kYUV12Size); | 57 CHECK_EQ(bytes_read, kYUV12Size); |
| 59 } | 58 } |
| 60 | 59 |
| 61 scoped_ptr<uint8[]> yuv_bytes_; | 60 scoped_ptr<uint8[]> yuv_bytes_; |
| 62 scoped_ptr<uint8[]> rgb_bytes_converted_; | 61 scoped_ptr<uint8[]> rgb_bytes_converted_; |
| 63 | 62 |
| 64 private: | 63 private: |
| 65 DISALLOW_COPY_AND_ASSIGN(YUVConvertPerfTest); | 64 DISALLOW_COPY_AND_ASSIGN(YUVConvertPerfTest); |
| 66 }; | 65 }; |
| 67 | 66 |
| 68 TEST_F(YUVConvertPerfTest, I422ToARGBRow_SSSE3) { | |
| 69 ASSERT_TRUE(base::CPU().has_ssse3()); | |
| 70 | |
| 71 base::TimeTicks start = base::TimeTicks::Now(); | |
| 72 for (int i = 0; i < kPerfTestIterations; ++i) { | |
| 73 for (int row = 0; row < kSourceHeight; ++row) { | |
| 74 int chroma_row = row / 2; | |
| 75 libyuv::I422ToARGBRow_SSSE3( | |
| 76 yuv_bytes_.get() + row * kSourceWidth, | |
| 77 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), | |
| 78 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), | |
| 79 rgb_bytes_converted_.get(), | |
| 80 kWidth); | |
| 81 } | |
| 82 } | |
| 83 double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF(); | |
| 84 perf_test::PrintResult( | |
| 85 "yuv_convert_perftest", "", "I422ToARGBRow_SSSE3", | |
| 86 kPerfTestIterations / total_time_seconds, "runs/s", true); | |
| 87 } | |
| 88 | |
| 89 TEST_F(YUVConvertPerfTest, ConvertYUVToRGB32Row_SSE) { | 67 TEST_F(YUVConvertPerfTest, ConvertYUVToRGB32Row_SSE) { |
| 90 ASSERT_TRUE(base::CPU().has_sse()); | 68 ASSERT_TRUE(base::CPU().has_sse()); |
| 91 | 69 |
| 92 base::TimeTicks start = base::TimeTicks::Now(); | 70 base::TimeTicks start = base::TimeTicks::Now(); |
| 93 for (int i = 0; i < kPerfTestIterations; ++i) { | 71 for (int i = 0; i < kPerfTestIterations; ++i) { |
| 94 for (int row = 0; row < kSourceHeight; ++row) { | 72 for (int row = 0; row < kSourceHeight; ++row) { |
| 95 int chroma_row = row / 2; | 73 int chroma_row = row / 2; |
| 96 ConvertYUVToRGB32Row_SSE( | 74 ConvertYUVToRGB32Row_SSE( |
| 97 yuv_bytes_.get() + row * kSourceWidth, | 75 yuv_bytes_.get() + row * kSourceWidth, |
| 98 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), | 76 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), |
| 99 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), | 77 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), |
| 100 rgb_bytes_converted_.get(), | 78 rgb_bytes_converted_.get(), |
| 101 kWidth, | 79 kWidth, |
| 102 GetLookupTable(YV12)); | 80 GetLookupTable(YV12)); |
| 103 } | 81 } |
| 104 } | 82 } |
| 105 media::EmptyRegisterState(); | |
| 106 double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF(); | 83 double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF(); |
| 107 perf_test::PrintResult( | 84 perf_test::PrintResult( |
| 108 "yuv_convert_perftest", "", "ConvertYUVToRGB32Row_SSE", | 85 "yuv_convert_perftest", "", "ConvertYUVToRGB32Row_SSE", |
| 109 kPerfTestIterations / total_time_seconds, "runs/s", true); | 86 kPerfTestIterations / total_time_seconds, "runs/s", true); |
| 110 } | |
| 111 | |
| 112 TEST_F(YUVConvertPerfTest, ConvertYUVAToARGBRow_MMX) { | |
| 113 ASSERT_TRUE(base::CPU().has_sse()); | |
| 114 | |
| 115 base::TimeTicks start = base::TimeTicks::Now(); | |
| 116 for (int i = 0; i < kPerfTestIterations; ++i) { | |
| 117 for (int row = 0; row < kSourceHeight; ++row) { | |
| 118 int chroma_row = row / 2; | |
| 119 ConvertYUVAToARGBRow_MMX( | |
| 120 yuv_bytes_.get() + row * kSourceWidth, | |
| 121 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), | |
| 122 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), | |
| 123 yuv_bytes_.get() + row * kSourceWidth, // hack: use luma for alpha | |
| 124 rgb_bytes_converted_.get(), | |
| 125 kWidth, | |
| 126 GetLookupTable(YV12)); | |
| 127 } | |
| 128 } | |
| 129 media::EmptyRegisterState(); | 87 media::EmptyRegisterState(); |
| 130 double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF(); | |
| 131 perf_test::PrintResult( | |
| 132 "yuv_convert_perftest", "", "ConvertYUVAToARGBRow_MMX", | |
| 133 kPerfTestIterations / total_time_seconds, "runs/s", true); | |
| 134 } | 88 } |
| 135 | 89 |
| 136 // 64-bit release + component builds on Windows are too smart and optimizes | 90 // 64-bit release + component builds on Windows are too smart and optimizes |
| 137 // away the function being tested. | 91 // away the function being tested. |
| 138 #if defined(OS_WIN) && (defined(ARCH_CPU_X86) || !defined(COMPONENT_BUILD)) | 92 #if defined(OS_WIN) && (defined(ARCH_CPU_X86) || !defined(COMPONENT_BUILD)) |
| 139 TEST_F(YUVConvertPerfTest, ScaleYUVToRGB32Row_SSE) { | 93 TEST_F(YUVConvertPerfTest, ScaleYUVToRGB32Row_SSE) { |
| 140 ASSERT_TRUE(base::CPU().has_sse()); | 94 ASSERT_TRUE(base::CPU().has_sse()); |
| 141 | 95 |
| 142 const int kSourceDx = 80000; // This value means a scale down. | 96 const int kSourceDx = 80000; // This value means a scale down. |
| 143 | 97 |
| 144 base::TimeTicks start = base::TimeTicks::Now(); | 98 base::TimeTicks start = base::TimeTicks::Now(); |
| 145 for (int i = 0; i < kPerfTestIterations; ++i) { | 99 for (int i = 0; i < kPerfTestIterations; ++i) { |
| 146 for (int row = 0; row < kSourceHeight; ++row) { | 100 for (int row = 0; row < kSourceHeight; ++row) { |
| 147 int chroma_row = row / 2; | 101 int chroma_row = row / 2; |
| 148 ScaleYUVToRGB32Row_SSE( | 102 ScaleYUVToRGB32Row_SSE( |
| 149 yuv_bytes_.get() + row * kSourceWidth, | 103 yuv_bytes_.get() + row * kSourceWidth, |
| 150 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), | 104 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), |
| 151 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), | 105 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), |
| 152 rgb_bytes_converted_.get(), | 106 rgb_bytes_converted_.get(), |
| 153 kWidth, | 107 kWidth, |
| 154 kSourceDx, | 108 kSourceDx, |
| 155 GetLookupTable(YV12)); | 109 GetLookupTable(YV12)); |
| 156 } | 110 } |
| 157 } | 111 } |
| 158 media::EmptyRegisterState(); | |
| 159 double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF(); | 112 double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF(); |
| 160 perf_test::PrintResult( | 113 perf_test::PrintResult( |
| 161 "yuv_convert_perftest", "", "ScaleYUVToRGB32Row_SSE", | 114 "yuv_convert_perftest", "", "ScaleYUVToRGB32Row_SSE", |
| 162 kPerfTestIterations / total_time_seconds, "runs/s", true); | 115 kPerfTestIterations / total_time_seconds, "runs/s", true); |
| 116 media::EmptyRegisterState(); |
| 163 } | 117 } |
| 164 | 118 |
| 165 TEST_F(YUVConvertPerfTest, LinearScaleYUVToRGB32Row_SSE) { | 119 TEST_F(YUVConvertPerfTest, LinearScaleYUVToRGB32Row_SSE) { |
| 166 ASSERT_TRUE(base::CPU().has_sse()); | 120 ASSERT_TRUE(base::CPU().has_sse()); |
| 167 | 121 |
| 168 const int kSourceDx = 80000; // This value means a scale down. | 122 const int kSourceDx = 80000; // This value means a scale down. |
| 169 | 123 |
| 170 base::TimeTicks start = base::TimeTicks::Now(); | 124 base::TimeTicks start = base::TimeTicks::Now(); |
| 171 for (int i = 0; i < kPerfTestIterations; ++i) { | 125 for (int i = 0; i < kPerfTestIterations; ++i) { |
| 172 for (int row = 0; row < kSourceHeight; ++row) { | 126 for (int row = 0; row < kSourceHeight; ++row) { |
| 173 int chroma_row = row / 2; | 127 int chroma_row = row / 2; |
| 174 LinearScaleYUVToRGB32Row_SSE( | 128 LinearScaleYUVToRGB32Row_SSE( |
| 175 yuv_bytes_.get() + row * kSourceWidth, | 129 yuv_bytes_.get() + row * kSourceWidth, |
| 176 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), | 130 yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2), |
| 177 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), | 131 yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2), |
| 178 rgb_bytes_converted_.get(), | 132 rgb_bytes_converted_.get(), |
| 179 kWidth, | 133 kWidth, |
| 180 kSourceDx, | 134 kSourceDx, |
| 181 GetLookupTable(YV12)); | 135 GetLookupTable(YV12)); |
| 182 } | 136 } |
| 183 } | 137 } |
| 184 media::EmptyRegisterState(); | |
| 185 double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF(); | 138 double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF(); |
| 186 perf_test::PrintResult( | 139 perf_test::PrintResult( |
| 187 "yuv_convert_perftest", "", "LinearScaleYUVToRGB32Row_SSE", | 140 "yuv_convert_perftest", "", "LinearScaleYUVToRGB32Row_SSE", |
| 188 kPerfTestIterations / total_time_seconds, "runs/s", true); | 141 kPerfTestIterations / total_time_seconds, "runs/s", true); |
| 142 media::EmptyRegisterState(); |
| 189 } | 143 } |
| 190 #endif // defined(OS_WIN) && (ARCH_CPU_X86 || COMPONENT_BUILD) | 144 #endif // defined(OS_WIN) && (ARCH_CPU_X86 || COMPONENT_BUILD) |
| 191 | 145 |
| 192 #endif // !defined(ARCH_CPU_ARM_FAMILY) && !defined(ARCH_CPU_MIPS_FAMILY) | 146 #endif // !defined(ARCH_CPU_ARM_FAMILY) && !defined(ARCH_CPU_MIPS_FAMILY) |
| 193 | 147 |
| 194 } // namespace media | 148 } // namespace media |
| OLD | NEW |