Add SSE optimizations to AudioRendererMixer.

media/base/audio_renderer_mixer_unittest.cc

Index: media/base/audio_renderer_mixer_unittest.cc
diff --git a/media/base/audio_renderer_mixer_unittest.cc b/media/base/audio_renderer_mixer_unittest.cc
index 51d06e0d63fc4da7626600972a3189f8028c14c2..6be81dba23e5d24de12576ce1839ff2737282fcc 100644
--- a/media/base/audio_renderer_mixer_unittest.cc
+++ b/media/base/audio_renderer_mixer_unittest.cc
@@ -8,8 +8,11 @@
 
 #include "base/bind.h"
 #include "base/bind_helpers.h"
+#include "base/command_line.h"
+#include "base/memory/aligned_memory.h"
 #include "base/memory/scoped_ptr.h"
 #include "base/memory/scoped_vector.h"
+#include "base/string_number_conversions.h"
 #include "media/base/audio_renderer_mixer.h"
 #include "media/base/audio_renderer_mixer_input.h"
 #include "media/base/fake_audio_render_callback.h"
@@ -17,6 +20,11 @@
 #include "testing/gmock/include/gmock/gmock.h"
 #include "testing/gtest/include/gtest/gtest.h"
 
+namespace switches {

[Ami GONE FROM CHROMIUM, 2012/07/27 17:33:11]

why?

[DaleCurtis, 2012/07/27 21:23:42]

Saw it done this way elsewhere. Changed.

+// Allows runtime adjustment of iterations for the VectorFMACBenchmark test.
+const char kVectorFMACIterations[] = "vector-fmac-iterations";
+}  // namespace switches
+
 namespace media {
 
 // Parameters which control the many input case tests.
@@ -28,10 +36,131 @@ static const int kBitsPerChannel = 16;
 static const ChannelLayout kChannelLayout = CHANNEL_LAYOUT_STEREO;
 static const int kHighLatencyBufferSize = 8192;
 static const int kLowLatencyBufferSize = 256;
+static const int kSampleRate = 48000;
 
 // Number of full sine wave cycles for each Render() call.
 static const int kSineCycles = 4;
 
+// Ensure various optimized VectorFMAC() methods return the right value.

[Ami GONE FROM CHROMIUM, 2012/07/27 17:33:11]

s/right/same/

[DaleCurtis, 2012/07/27 21:23:42]

Done.

+TEST(AudioRendererMixerTest, VectorFMAC) {
+  // Initialize a dummy mixer.
+  scoped_refptr<MockAudioRendererSink> sink = new MockAudioRendererSink();
+  EXPECT_CALL(*sink, Start());
+  EXPECT_CALL(*sink, Stop());
+  AudioParameters params(
+      AudioParameters::AUDIO_PCM_LINEAR, kChannelLayout, kSampleRate,
+      kBitsPerChannel, kHighLatencyBufferSize);
+  AudioRendererMixer mixer(params, params, sink);
+
+  // Initialize input and output vectors.
+  scoped_ptr_malloc<float, base::ScopedPtrAlignedFree> input_vector(
+      static_cast<float*>(
+          base::AlignedAlloc(sizeof(float) * kHighLatencyBufferSize, 16)));
+  scoped_ptr_malloc<float, base::ScopedPtrAlignedFree> output_vector(
+      static_cast<float*>(
+          base::AlignedAlloc(sizeof(float) * kHighLatencyBufferSize, 16)));
+
+  static const float kScale = 0.5;
+  static const float kFillValue = 1.0;
+
+  // Fill input vector with 1.0, zero output vector
+  std::fill(input_vector.get(), input_vector.get() + kHighLatencyBufferSize,
+            kFillValue);
+  memset(output_vector.get(), 0, sizeof(float) * kHighLatencyBufferSize);

[Ami GONE FROM CHROMIUM, 2012/07/27 17:33:11]

By init'd output with 0 you're avoiding testing the "accumulate" part of FMAC ;)
Replace memset of 0 with std::fill of 3?

[DaleCurtis, 2012/07/27 21:23:42]

Done.

+  mixer.VectorFMAC_C(
+      input_vector.get(), kScale, kHighLatencyBufferSize, output_vector.get());
+  for(int i = 0; i < kHighLatencyBufferSize; ++i)
+    ASSERT_FLOAT_EQ(output_vector.get()[i], kScale);
+
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+  // Reset vectors, and try with SSE.
+  std::fill(input_vector.get(), input_vector.get() + kHighLatencyBufferSize,

[Ami GONE FROM CHROMIUM, 2012/07/27 17:33:11]

Why is this necessary?

[DaleCurtis, 2012/07/27 21:23:42]

Whoops, don't need to reset input vector, but output vector needs to be reset,
otherwise it'll accumulate the previous VectorFMAC_C result.

+            kFillValue);
+  memset(output_vector.get(), 0, sizeof(float) * kHighLatencyBufferSize);

[Ami GONE FROM CHROMIUM, 2012/07/27 17:33:11]

ditto

[DaleCurtis, 2012/07/27 21:23:42]

Done.

+  mixer.VectorFMAC_SSE(
+      input_vector.get(), kScale, kHighLatencyBufferSize, output_vector.get());
+  for(int i = 0; i < kHighLatencyBufferSize; ++i)
+    ASSERT_FLOAT_EQ(output_vector.get()[i], kScale);
+#endif
+}
+
+// Benchmark for the various VectorFMAC() methods.  Make sure to build with
+// branding=Chrome so that DCHECKs are compiled out when benchmarking.  Original
+// benchmarks were run with --vector-fmac-iterations=200000.
+TEST(AudioRendererMixerTest, VectorFMACBenchmark) {
+  // Initialize a dummy mixer.
+  scoped_refptr<MockAudioRendererSink> sink = new MockAudioRendererSink();
+  EXPECT_CALL(*sink, Start());
+  EXPECT_CALL(*sink, Stop());
+  AudioParameters params(
+      AudioParameters::AUDIO_PCM_LINEAR, kChannelLayout, kSampleRate,
+      kBitsPerChannel, kHighLatencyBufferSize);
+  AudioRendererMixer mixer(params, params, sink);
+
+  // Initialize input and output vectors.
+  scoped_ptr_malloc<float, base::ScopedPtrAlignedFree> input_vector(
+      static_cast<float*>(
+          base::AlignedAlloc(sizeof(float) * kHighLatencyBufferSize, 16)));
+  scoped_ptr_malloc<float, base::ScopedPtrAlignedFree> output_vector(
+      static_cast<float*>(
+          base::AlignedAlloc(sizeof(float) * kHighLatencyBufferSize, 16)));
+
+  // Retrieve benchmark iterations from command line.
+  int vector_fmac_iterations = 10;
+  std::string iterations(CommandLine::ForCurrentProcess()->GetSwitchValueASCII(
+      switches::kVectorFMACIterations));
+  if (!iterations.empty())
+    base::StringToInt(iterations, &vector_fmac_iterations);
+
+  printf("Benchmarking %d iterations:\n", vector_fmac_iterations);
+
+  // Benchmark VectorFMAC_C().
+  memset(input_vector.get(), 1, sizeof(float) * kHighLatencyBufferSize);

[Ami GONE FROM CHROMIUM, 2012/07/27 17:33:11]

Are you intentionally testing inputs of inputs that aren't 1?
(this is setting bytes, not floats)

[DaleCurtis, 2012/07/27 21:23:42]

Yeah, just needed to initialize the array with something. Switched to
std::fill(3) for clarity.

+  memset(output_vector.get(), 0, sizeof(float) * kHighLatencyBufferSize);
+  base::TimeTicks start = base::TimeTicks::HighResNow();
+  for (int i = 0; i < vector_fmac_iterations; ++i) {
+    mixer.VectorFMAC_C(input_vector.get(), M_PI, kHighLatencyBufferSize,
+                       output_vector.get());
+  }
+  double total_time_c_ms =
+      (base::TimeTicks::HighResNow() - start).InMillisecondsF();
+  printf("VectorFMAC_C took %.2fms.\n", total_time_c_ms);
+
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+  // Benchmark VectorFMAC_SSE() with unaligned size; I.e., size % 4 != 0.

[Ami GONE FROM CHROMIUM, 2012/07/27 17:33:11]

Comment here that the API insists the head *is* aligned?

[DaleCurtis, 2012/07/27 21:23:42]

Seems redundant at this point, it's already in the header and we've got aligned
allocations above. It's clearer to me to focus on what's different here.

+  ASSERT_NE((kHighLatencyBufferSize - 1) % 4, 0);
+  memset(input_vector.get(), 1, sizeof(float) * kHighLatencyBufferSize);

[Ami GONE FROM CHROMIUM, 2012/07/27 17:33:11]

ditto

[DaleCurtis, 2012/07/27 21:23:42]

Done.

+  memset(output_vector.get(), 0, sizeof(float) * kHighLatencyBufferSize);
+  start = base::TimeTicks::HighResNow();
+  for (int j = 0; j < vector_fmac_iterations; ++j) {
+    mixer.VectorFMAC_SSE(input_vector.get(), M_PI, kHighLatencyBufferSize - 1,
+                         output_vector.get());
+  }
+  double total_time_sse_unaligned_ms =
+      (base::TimeTicks::HighResNow() - start).InMillisecondsF();
+  printf("VectorFMAC_SSE (unaligned size) took %.2fms; which is %.2fx faster"
+         " than VectorFMAC_C.\n", total_time_sse_unaligned_ms,
+         total_time_c_ms / total_time_sse_unaligned_ms);
+
+  // Benchmark VectorFMAC_SSE() with aligned size; I.e., size % 4 == 0.
+  ASSERT_EQ(kHighLatencyBufferSize % 4, 0);
+  memset(input_vector.get(), 1, sizeof(float) * kHighLatencyBufferSize);

[Ami GONE FROM CHROMIUM, 2012/07/27 17:33:11]

ditto

[DaleCurtis, 2012/07/27 21:23:42]

Done.

+  memset(output_vector.get(), 0, sizeof(float) * kHighLatencyBufferSize);
+  start = base::TimeTicks::HighResNow();
+  for (int j = 0; j < vector_fmac_iterations; ++j) {
+    mixer.VectorFMAC_SSE(input_vector.get(), M_PI, kHighLatencyBufferSize,
+                         output_vector.get());
+  }
+  double total_time_sse_aligned_ms =
+      (base::TimeTicks::HighResNow() - start).InMillisecondsF();
+  printf("VectorFMAC_SSE (aligned size) took %.2fms; which is %.2fx faster than"
+         " VectorFMAC_C and %.2fx faster than VectorFMAC_SSE (unaligned size)."
+         "\n",
+         total_time_sse_aligned_ms, total_time_c_ms / total_time_sse_aligned_ms,
+         total_time_sse_unaligned_ms / total_time_sse_aligned_ms);
+#endif
+}
+
 // Tuple of <input sampling rate, output sampling rate, epsilon>.
 typedef std::tr1::tuple<int, int, double> AudioRendererMixerTestData;
 class AudioRendererMixerTest