add ConvertYUVAToARGBRow_MMX to media_perftests

Add YUVConvertPerfTest.ConvertYUVAToARGBRow_MMX and YUVConvertPerfTest.I422ToARGBRow_SSSE3 to media_perftests.
Fix math error when printing results by clearing the MMX state before doing float math.
BUG=libyuv:473

Review URL: https://codereview.chromium.org/1425263002

Cr-Commit-Position: refs/heads/master@{#357525}

3 files changed, 68 insertions, 2 deletions

diff --git a/media/BUILD.gn b/media/BUILD.gn
index f15e590..52af493 100644
--- a/media/BUILD.gn
+++ b/media/BUILD.gn
@@ -791,6 +791,7 @@ test("media_perftests") {
     "//testing/gtest",
     "//testing/perf",
     "//third_party/widevine/cdm:version_h",
+    "//third_party/libyuv",
     "//ui/gfx:test_support",
   ]
   if (is_android) {
diff --git a/media/base/yuv_convert_perftest.cc b/media/base/yuv_convert_perftest.cc
index d676e4e..20a0da8 100644
--- a/media/base/yuv_convert_perftest.cc
+++ b/media/base/yuv_convert_perftest.cc
@@ -12,6 +12,7 @@
 #include "media/base/yuv_convert.h"
 #include "testing/gtest/include/gtest/gtest.h"
 #include "testing/perf/perf_test.h"
+#include "third_party/libyuv/include/libyuv/row.h"
 
 namespace media {
 #if !defined(ARCH_CPU_ARM_FAMILY) && !defined(ARCH_CPU_MIPS_FAMILY)
@@ -87,6 +88,69 @@ TEST_F(YUVConvertPerfTest, ConvertYUVToRGB32Row_SSE) {
       kPerfTestIterations / total_time_seconds, "runs/s", true);
 }
 
+TEST_F(YUVConvertPerfTest, I422ToARGBRow_SSSE3) {
+  ASSERT_TRUE(base::CPU().has_ssse3());
+
+  base::TimeTicks start = base::TimeTicks::Now();
+  for (int i = 0; i < kPerfTestIterations; ++i) {
+    for (int row = 0; row < kSourceHeight; ++row) {
+      int chroma_row = row / 2;
+      libyuv::I422ToARGBRow_SSSE3(
+          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(), &libyuv::kYuvIConstants, kWidth);
+    }
+  }
+  double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF();
+  perf_test::PrintResult("yuv_convert_perftest", "", "I422ToARGBRow_SSSE3",
+                         kPerfTestIterations / total_time_seconds, "runs/s",
+                         true);
+}
+
+TEST_F(YUVConvertPerfTest, ConvertYUVAToARGBRow_MMX) {
+  ASSERT_TRUE(base::CPU().has_sse());
+
+  base::TimeTicks start = base::TimeTicks::Now();
+  for (int i = 0; i < kPerfTestIterations; ++i) {
+    for (int row = 0; row < kSourceHeight; ++row) {
+      int chroma_row = row / 2;
+      ConvertYUVAToARGBRow_MMX(
+          yuv_bytes_.get() + row * kSourceWidth,
+          yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2),
+          yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2),
+          yuv_bytes_.get() + row * kSourceWidth,  // hack: use luma for alpha
+          rgb_bytes_converted_.get(), kWidth, GetLookupTable(YV12));
+    }
+  }
+  media::EmptyRegisterState();
+  double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF();
+  perf_test::PrintResult("yuv_convert_perftest", "", "ConvertYUVAToARGBRow_MMX",
+                         kPerfTestIterations / total_time_seconds, "runs/s",
+                         true);
+}
+
+TEST_F(YUVConvertPerfTest, I422AlphaToARGBRow_SSSE3) {
+  ASSERT_TRUE(base::CPU().has_ssse3());
+
+  base::TimeTicks start = base::TimeTicks::Now();
+  for (int i = 0; i < kPerfTestIterations; ++i) {
+    for (int row = 0; row < kSourceHeight; ++row) {
+      int chroma_row = row / 2;
+      libyuv::I422AlphaToARGBRow_SSSE3(
+          yuv_bytes_.get() + row * kSourceWidth,
+          yuv_bytes_.get() + kSourceUOffset + (chroma_row * kSourceWidth / 2),
+          yuv_bytes_.get() + kSourceVOffset + (chroma_row * kSourceWidth / 2),
+          yuv_bytes_.get() + row * kSourceWidth,  // hack: use luma for alpha
+          rgb_bytes_converted_.get(), &libyuv::kYuvIConstants, kWidth);
+    }
+  }
+  double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF();
+  perf_test::PrintResult("yuv_convert_perftest", "", "I422AlphaToARGBRow_SSSE3",
+                         kPerfTestIterations / total_time_seconds, "runs/s",
+                         true);
+}
+
 // 64-bit release + component builds on Windows are too smart and optimizes
 // away the function being tested.
 #if defined(OS_WIN) && (defined(ARCH_CPU_X86) || !defined(COMPONENT_BUILD))
@@ -135,11 +199,11 @@ TEST_F(YUVConvertPerfTest, LinearScaleYUVToRGB32Row_SSE) {
           GetLookupTable(YV12));
     }
   }
+  media::EmptyRegisterState();
   double total_time_seconds = (base::TimeTicks::Now() - start).InSecondsF();
   perf_test::PrintResult(
       "yuv_convert_perftest", "", "LinearScaleYUVToRGB32Row_SSE",
       kPerfTestIterations / total_time_seconds, "runs/s", true);
-  media::EmptyRegisterState();
 }
 #endif  // defined(OS_WIN) && (ARCH_CPU_X86 || COMPONENT_BUILD)
 
diff --git a/media/media.gyp b/media/media.gyp
index 9acf41e..661e8e7 100644
--- a/media/media.gyp
+++ b/media/media.gyp
@@ -1447,9 +1447,10 @@
         '../testing/gmock.gyp:gmock',
         '../testing/gtest.gyp:gtest',
         '../testing/perf/perf_test.gyp:perf_test',
+        '../third_party/libyuv/libyuv.gyp:libyuv',
         '../ui/gfx/gfx.gyp:gfx',
-        '../ui/gfx/gfx.gyp:gfx_test_support',
         '../ui/gfx/gfx.gyp:gfx_geometry',
+        '../ui/gfx/gfx.gyp:gfx_test_support',
         'media',
         'media_test_support',
         'shared_memory_support',