Add SSE optimizations to AudioRendererMixer.

media/base/audio_renderer_mixer.cc

Index: media/base/audio_renderer_mixer.cc
diff --git a/media/base/audio_renderer_mixer.cc b/media/base/audio_renderer_mixer.cc
index 6d23faad35d6257274d0738ffdbb903104e7006f..88b7ac0f624772d2c0902c4d489d50eb5df306b0 100644
--- a/media/base/audio_renderer_mixer.cc
+++ b/media/base/audio_renderer_mixer.cc
@@ -6,17 +6,57 @@
 
 #include "base/bind.h"
 #include "base/bind_helpers.h"
+#include "base/cpu.h"
 #include "base/logging.h"
 #include "media/audio/audio_util.h"
 #include "media/base/limits.h"
 
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+#include <xmmintrin.h>
+#endif
+
 namespace media {
 
+static void VectorFMAC_C(float* src, float* dest, float scale, int len) {
+  for (int i = 0; i < len; ++i)
+    dest[i] += src[i] * scale;
+}
+
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+static const int kFloatsPerPass = sizeof(__m128) / sizeof(float);
+static void VectorFMAC_SSE(float* src, float* dest, float scale, int len) {
+  // Ensure |src| and |dest| are aligned properly.
+  DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(src) & (sizeof(__m128) - 1));
+  DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(dest) & (sizeof(__m128) - 1));
+
+  __m128 m_scale = _mm_set_ps1(scale);
+  int rem = len % kFloatsPerPass;
+  for (int i = 0; i < len - rem; i += kFloatsPerPass) {
+    _mm_store_ps(dest + i, _mm_add_ps(_mm_load_ps(dest + i),
+                 _mm_mul_ps(_mm_load_ps(src + i), m_scale)));
+  }
+
+  // Handle any remaining values that wouldn't fit in an SSE pass.
+  if (rem)
+    VectorFMAC_C(src + len - rem, dest + len - rem, scale, rem);
+}
+#endif
+
+typedef void (*VectorFMACProc)(float* src, float* dest, float scale, int len);
+static VectorFMACProc vector_fmac_proc = NULL;
+
 AudioRendererMixer::AudioRendererMixer(
     const AudioParameters& input_params, const AudioParameters& output_params,
     const scoped_refptr<AudioRendererSink>& sink)
     : audio_sink_(sink),
       current_audio_delay_milliseconds_(0) {
+  if (!vector_fmac_proc) {
+    vector_fmac_proc = VectorFMAC_C;
+    base::CPU cpu;
+    if (cpu.has_sse())
+      vector_fmac_proc = VectorFMAC_SSE;
+  }
+
   // Sanity check sample rates.
   DCHECK_LE(input_params.sample_rate(), limits::kMaxSampleRate);
   DCHECK_GE(input_params.sample_rate(), limits::kMinSampleRate);
@@ -120,12 +160,9 @@ void AudioRendererMixer::ProvideInput(const std::vector<float*>& audio_data,
       continue;
 
     // Volume adjust and mix each mixer input into |audio_data| after rendering.
-    // TODO(dalecurtis): Optimize with NEON/SSE/AVX vector_fmac from FFmpeg.
     for (size_t j = 0; j < audio_data.size(); ++j) {
-      float* dest = audio_data[j];
-      float* source = mixer_input_audio_data_[j];
-      for (int k = 0; k < frames_filled; ++k)
-        dest[k] += source[k] * static_cast<float>(volume);
+       vector_fmac_proc(
+           mixer_input_audio_data_[j], audio_data[j], volume, frames_filled);
     }
 
     // No need to clamp values as InterleaveFloatToInt() will take care of this