Add perf tests for simd yuv->rgb conversion functions.

media/base/yuv_convert_perftest.cc

Index: media/base/yuv_convert_perftest.cc
diff --git a/media/base/yuv_convert_perftest.cc b/media/base/yuv_convert_perftest.cc
new file mode 100644
index 0000000000000000000000000000000000000000..8a7c076aa76f527356eba3dae0a90ea5c5307c37
--- /dev/null
+++ b/media/base/yuv_convert_perftest.cc
@@ -0,0 +1,269 @@
+// Copyright 2014 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "base/base_paths.h"
+#include "base/cpu.h"
+#include "base/file_util.h"
+#include "base/logging.h"
+#include "base/path_service.h"
+#include "base/time/time.h"
+#include "media/base/djb2.h"

[scherkus (not reviewing), 2014/04/22 18:10:56]

this isn't used anywhere -- remove

(might be worth trimming a bunch of these headers if you yanked them from
another file...)

[rileya (GONE FROM CHROMIUM), 2014/04/23 18:09:21]

Whoops, yeah I grabbed this from yuv_convert_unittest.cc. Removed unnecessary
ones.

+#include "media/base/simd/convert_rgb_to_yuv.h"
+#include "media/base/simd/convert_yuv_to_rgb.h"
+#include "media/base/simd/filter_yuv.h"
+#include "media/base/yuv_convert.h"
+#include "testing/gtest/include/gtest/gtest.h"
+#include "testing/perf/perf_test.h"
+#include "ui/gfx/rect.h"
+

[scherkus (not reviewing), 2014/04/22 18:10:56]

namespace media should go here

[rileya (GONE FROM CHROMIUM), 2014/04/23 18:09:21]

Done.

+// Size of raw image.
+static const int kSourceWidth = 640;
+static const int kSourceHeight = 360;
+static const int kSourceYSize = kSourceWidth * kSourceHeight;
+static const int kSourceUOffset = kSourceYSize;
+static const int kSourceVOffset = kSourceYSize * 5 / 4;
+static const int kBpp = 4;
+
+// Width of the row to convert. Odd so that we exercise the ending
+// one-pixel-leftover case.
+static const int kWidth = 639;
+
+// Surface sizes for various test files.
+static const int kYUV12Size = kSourceYSize * 12 / 8;
+static const int kRGBSize = kSourceYSize * kBpp;
+
+static const int kPerfTestIterations = 2000;
+
+// Helper for reading test data into a scoped_ptr<uint8[]>.
+static void ReadData(const base::FilePath::CharType* filename,
+                     int expected_size,
+                     scoped_ptr<uint8[]>* data) {
+  data->reset(new uint8[expected_size]);
+
+  base::FilePath path;
+  CHECK(PathService::Get(base::DIR_SOURCE_ROOT, &path));
+  path = path.Append(FILE_PATH_LITERAL("media"))
+             .Append(FILE_PATH_LITERAL("test"))
+             .Append(FILE_PATH_LITERAL("data"))
+             .Append(filename);
+
+  // Verify file size is correct.
+  int64 actual_size = 0;
+  base::GetFileSize(path, &actual_size);
+  CHECK_EQ(actual_size, expected_size);
+
+  // Verify bytes read are correct.
+  int bytes_read = base::ReadFile(
+      path, reinterpret_cast<char*>(data->get()), expected_size);
+  CHECK_EQ(bytes_read, expected_size);
+}
+
+class YUVConvertPerfTest : public testing::Test {
+ public:
+  YUVConvertPerfTest()
+      : yuv_bytes_(new uint8[kYUV12Size]),
+        rgb_bytes_converted_(new uint8[kRGBSize]) {}
+
+  void InitYV12Data() {
+    ReadData(

[scherkus (not reviewing), 2014/04/22 18:10:56]

nit: is it worth having ReadData() if there's only one caller?

[rileya (GONE FROM CHROMIUM), 2014/04/23 18:09:21]

Removed.

+        FILE_PATH_LITERAL("bali_640x360_P420.yuv"), kYUV12Size, &yuv_bytes_);
+  }
+
+  scoped_ptr<uint8[]> yuv_bytes_;
+  scoped_ptr<uint8[]> rgb_bytes_converted_;
+};

[scherkus (not reviewing), 2014/04/22 18:10:56]

DISALLOW etc..

[rileya (GONE FROM CHROMIUM), 2014/04/23 18:09:21]

Added DISALLOW_COPY_AND_ASSIGN.

+
+namespace media {
+
+#if !defined(ARCH_CPU_ARM_FAMILY) && !defined(ARCH_CPU_MIPS_FAMILY)
+TEST_F(YUVConvertPerfTest, ConvertYUVToRGB32Row_MMX) {
+  base::CPU cpu;
+  if (!cpu.has_mmx()) {

[scherkus (not reviewing), 2014/04/22 18:10:56]

considering our minimum supported arch is SSE2 I'd rather turn these into
ASSERTs

are there machines out there that really don't have mmx/sse[2]?

[rileya (GONE FROM CHROMIUM), 2014/04/23 18:09:21]

Probably not many, but I was following the convention set by the yuv convert
unit tests. Turned these into ASSERT_TRUEs.

+    LOG(WARNING) << "System not supported. Test skipped.";
+    return;
+  }
+
+  InitYV12Data();
+
+  base::TimeTicks start = base::TimeTicks::HighResNow();
+  for (int i = 0; i < kPerfTestIterations; ++i) {
+    for (int row = 0; row < kSourceHeight; ++row) {
+      int chroma_row = row / 2;
+      ConvertYUVToRGB32Row_MMX(
+          yuv_bytes_.get() + row * kSourceWidth,
+          yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2),
+          yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2),
+          rgb_bytes_converted_.get(),
+          kWidth);
+    }
+  }
+  double total_time_seconds =
+      (base::TimeTicks::HighResNow() - start).InSecondsF();
+  perf_test::PrintResult(
+      "yuv_convert_perftest", "", "ConvertYUVToRGB32Row_MMX",
+      kPerfTestIterations / total_time_seconds, "runs/s", true);
+
+  media::EmptyRegisterState();
+}
+
+TEST_F(YUVConvertPerfTest, ConvertYUVToRGB32Row_SSE) {
+  base::CPU cpu;
+  if (!cpu.has_sse()) {
+    LOG(WARNING) << "System not supported. Test skipped.";
+    return;
+  }
+
+  InitYV12Data();

[scherkus (not reviewing), 2014/04/22 18:10:56]

this is called every time -- move into ctor?

[rileya (GONE FROM CHROMIUM), 2014/04/23 18:09:21]

Done.

+
+  base::TimeTicks start = base::TimeTicks::HighResNow();
+  for (int i = 0; i < kPerfTestIterations; ++i) {
+    for (int row = 0; row < kSourceHeight; ++row) {
+      int chroma_row = row / 2;
+      ConvertYUVToRGB32Row_SSE(
+          yuv_bytes_.get() + row * kSourceWidth,
+          yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2),
+          yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2),
+          rgb_bytes_converted_.get(),
+          kWidth);
+    }
+  }
+  double total_time_seconds =
+      (base::TimeTicks::HighResNow() - start).InSecondsF();
+  perf_test::PrintResult(
+      "yuv_convert_perftest", "", "ConvertYUVToRGB32Row_SSE",
+      kPerfTestIterations / total_time_seconds, "runs/s", true);
+  media::EmptyRegisterState();
+}
+
+TEST_F(YUVConvertPerfTest, ScaleYUVToRGB32Row_MMX) {
+  base::CPU cpu;
+  if (!cpu.has_mmx()) {
+    LOG(WARNING) << "System not supported. Test skipped.";
+    return;
+  }
+
+  InitYV12Data();
+
+  const int kSourceDx = 80000;  // This value means a scale down.
+
+  base::TimeTicks start = base::TimeTicks::HighResNow();
+  for (int i = 0; i < kPerfTestIterations; ++i) {
+    for (int row = 0; row < kSourceHeight; ++row) {
+      int chroma_row = row / 2;
+      ScaleYUVToRGB32Row_MMX(
+          yuv_bytes_.get() + row * kSourceWidth,
+          yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2),
+          yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2),
+          rgb_bytes_converted_.get(),
+          kWidth,
+          kSourceDx);
+    }
+  }
+  double total_time_seconds =
+      (base::TimeTicks::HighResNow() - start).InSecondsF();
+  perf_test::PrintResult(
+      "yuv_convert_perftest", "", "ScaleYUVToRGB32Row_MMX",
+      kPerfTestIterations / total_time_seconds, "runs/s", true);
+  media::EmptyRegisterState();
+}
+
+TEST_F(YUVConvertPerfTest, ScaleYUVToRGB32Row_SSE) {
+  base::CPU cpu;
+  if (!cpu.has_sse()) {
+    LOG(WARNING) << "System not supported. Test skipped.";
+    return;
+  }
+
+  InitYV12Data();
+
+  const int kSourceDx = 80000;  // This value means a scale down.
+
+  base::TimeTicks start = base::TimeTicks::HighResNow();
+  for (int i = 0; i < kPerfTestIterations; ++i) {
+    for (int row = 0; row < kSourceHeight; ++row) {
+      int chroma_row = row / 2;
+      ScaleYUVToRGB32Row_SSE(
+          yuv_bytes_.get() + row * kSourceWidth,
+          yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2),
+          yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2),
+          rgb_bytes_converted_.get(),
+          kWidth,
+          kSourceDx);
+    }
+  }
+  double total_time_seconds =
+      (base::TimeTicks::HighResNow() - start).InSecondsF();
+  perf_test::PrintResult(
+      "yuv_convert_perftest", "", "ScaleYUVToRGB32Row_SSE",
+      kPerfTestIterations / total_time_seconds, "runs/s", true);
+  media::EmptyRegisterState();
+}
+
+TEST_F(YUVConvertPerfTest, LinearScaleYUVToRGB32Row_MMX) {
+  base::CPU cpu;
+  if (!cpu.has_mmx()) {
+    LOG(WARNING) << "System not supported. Test skipped.";
+    return;
+  }
+
+  InitYV12Data();
+
+  const int kSourceDx = 80000;  // This value means a scale down.
+
+  base::TimeTicks start = base::TimeTicks::HighResNow();
+  for (int i = 0; i < kPerfTestIterations; ++i) {
+    for (int row = 0; row < kSourceHeight; ++row) {
+      int chroma_row = row / 2;
+      LinearScaleYUVToRGB32Row_MMX(
+          yuv_bytes_.get() + row * kSourceWidth,
+          yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2),
+          yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2),
+          rgb_bytes_converted_.get(),
+          kWidth,
+          kSourceDx);
+    }
+  }
+  double total_time_seconds =
+      (base::TimeTicks::HighResNow() - start).InSecondsF();
+  perf_test::PrintResult(
+      "yuv_convert_perftest", "", "LinearScaleYUVToRGB32Row_MMX",
+      kPerfTestIterations / total_time_seconds, "runs/s", true);
+  media::EmptyRegisterState();
+}
+
+TEST_F(YUVConvertPerfTest, LinearScaleYUVToRGB32Row_SSE) {
+  base::CPU cpu;
+  if (!cpu.has_sse()) {
+    LOG(WARNING) << "System not supported. Test skipped.";
+    return;
+  }
+
+  InitYV12Data();
+
+  const int kSourceDx = 80000;  // This value means a scale down.
+
+  base::TimeTicks start = base::TimeTicks::HighResNow();
+  for (int i = 0; i < kPerfTestIterations; ++i) {
+    for (int row = 0; row < kSourceHeight; ++row) {
+      int chroma_row = row / 2;
+      LinearScaleYUVToRGB32Row_SSE(
+          yuv_bytes_.get() + row * kSourceWidth,
+          yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2),
+          yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2),
+          rgb_bytes_converted_.get(),
+          kWidth,
+          kSourceDx);
+    }
+  }
+  double total_time_seconds =
+      (base::TimeTicks::HighResNow() - start).InSecondsF();
+  perf_test::PrintResult(
+      "yuv_convert_perftest", "", "LinearScaleYUVToRGB32Row_SSE",
+      kPerfTestIterations / total_time_seconds, "runs/s", true);
+  media::EmptyRegisterState();
+}
+
+#endif  // !defined(ARCH_CPU_ARM_FAMILY) && !defined(ARCH_CPU_MIPS_FAMILY)
+
+}  // namespace media