Merge from Chromium at DEPS revision r190564

This commit was generated by merge_to_master.py.

Change-Id: Icadecbce29854b8fa25fd335b2c1949b5ca5d170

1 files changed, 28 insertions, 98 deletions

diff --git a/media/base/sinc_resampler.cc b/media/base/sinc_resampler.cc
index d836fc7cbc..00f9314c61 100644
--- a/media/base/sinc_resampler.cc
+++ b/media/base/sinc_resampler.cc
@@ -40,11 +40,6 @@
 
 #include "base/cpu.h"
 #include "base/logging.h"
-#include "build/build_config.h"
-
-#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
-#include <xmmintrin.h>
-#endif
 
 #if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
 #include <arm_neon.h>
@@ -52,33 +47,6 @@
 
 namespace media {
 
-namespace {
-
-enum {
-  // The kernel size can be adjusted for quality (higher is better) at the
-  // expense of performance.  Must be a multiple of 32.
-  // TODO(dalecurtis): Test performance to see if we can jack this up to 64+.
-  kKernelSize = 32,
-
-  // The number of destination frames generated per processing pass.  Affects
-  // how often and for how much SincResampler calls back for input.  Must be
-  // greater than kKernelSize.
-  kBlockSize = 512,
-
-  // The kernel offset count is used for interpolation and is the number of
-  // sub-sample kernel shifts.  Can be adjusted for quality (higher is better)
-  // at the expense of allocating more memory.
-  kKernelOffsetCount = 32,
-  kKernelStorageSize = kKernelSize * (kKernelOffsetCount + 1),
-
-  // The size (in samples) of the internal buffer used by the resampler.
-  kBufferSize = kBlockSize + kKernelSize
-};
-
-}  // namespace
-
-const int SincResampler::kMaximumLookAheadSize = kBufferSize;
-
 SincResampler::SincResampler(double io_sample_rate_ratio, const ReadCB& read_cb)
     : io_sample_rate_ratio_(io_sample_rate_ratio),
       virtual_source_idx_(0),
@@ -89,6 +57,9 @@ SincResampler::SincResampler(double io_sample_rate_ratio, const ReadCB& read_cb)
           base::AlignedAlloc(sizeof(float) * kKernelStorageSize, 16))),
       input_buffer_(static_cast<float*>(
           base::AlignedAlloc(sizeof(float) * kBufferSize, 16))),
+#if defined(ARCH_CPU_X86_FAMILY) && !defined(__SSE__)
+      convolve_proc_(base::CPU().has_sse() ? Convolve_SSE : Convolve_C),
+#endif
       // Setup various region pointers in the buffer (see diagram above).
       r0_(input_buffer_.get() + kKernelSize / 2),
       r1_(input_buffer_.get()),
@@ -168,6 +139,22 @@ void SincResampler::InitializeKernel() {
   }
 }
 
+// If we know the minimum architecture avoid function hopping for CPU detection.
+#if defined(ARCH_CPU_X86_FAMILY)
+#if defined(__SSE__)
+#define CONVOLVE_FUNC Convolve_SSE
+#else
+// X86 CPU detection required.  |convolve_proc_| will be set upon construction.
+// TODO(dalecurtis): Once Chrome moves to a SSE baseline this can be removed.
+#define CONVOLVE_FUNC convolve_proc_
+#endif
+#elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+#define CONVOLVE_FUNC Convolve_NEON
+#else
+// Unknown architecture.
+#define CONVOLVE_FUNC Convolve_C
+#endif
+
 void SincResampler::Resample(float* destination, int frames) {
   int remaining_frames = frames;
 
@@ -193,12 +180,17 @@ void SincResampler::Resample(float* destination, int frames) {
       float* k1 = kernel_storage_.get() + offset_idx * kKernelSize;
       float* k2 = k1 + kKernelSize;
 
+      // Ensure |k1|, |k2| are 16-byte aligned for SIMD usage.  Should always be
+      // true so long as kKernelSize is a multiple of 16.
+      DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(k1) & 0x0F);
+      DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(k2) & 0x0F);
+
       // Initialize input pointer based on quantized |virtual_source_idx_|.
       float* input_ptr = r1_ + source_idx;
 
       // Figure out how much to weight each kernel's "convolution".
       double kernel_interpolation_factor = virtual_offset_idx - offset_idx;
-      *destination++ = Convolve(
+      *destination++ = CONVOLVE_FUNC(
           input_ptr, k1, k2, kernel_interpolation_factor);
 
       // Advance the virtual index.
@@ -222,7 +214,9 @@ void SincResampler::Resample(float* destination, int frames) {
   }
 }
 
-int SincResampler::ChunkSize() {
+#undef CONVOLVE_FUNC
+
+int SincResampler::ChunkSize() const {
   return kBlockSize / io_sample_rate_ratio_;
 }
 
@@ -232,26 +226,6 @@ void SincResampler::Flush() {
   memset(input_buffer_.get(), 0, sizeof(*input_buffer_.get()) * kBufferSize);
 }
 
-float SincResampler::Convolve(const float* input_ptr, const float* k1,
-                              const float* k2,
-                              double kernel_interpolation_factor) {
-  // Rely on function level static initialization to keep ConvolveProc selection
-  // thread safe.
-  typedef float (*ConvolveProc)(const float* src, const float* k1,
-                                const float* k2,
-                                double kernel_interpolation_factor);
-#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
-  static const ConvolveProc kConvolveProc =
-      base::CPU().has_sse() ? Convolve_SSE : Convolve_C;
-#elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
-  static const ConvolveProc kConvolveProc = Convolve_NEON;
-#else
-  static const ConvolveProc kConvolveProc = Convolve_C;
-#endif
-
-  return kConvolveProc(input_ptr, k1, k2, kernel_interpolation_factor);
-}
-
 float SincResampler::Convolve_C(const float* input_ptr, const float* k1,
                                 const float* k2,
                                 double kernel_interpolation_factor) {
@@ -271,50 +245,6 @@ float SincResampler::Convolve_C(const float* input_ptr, const float* k1,
       + kernel_interpolation_factor * sum2;
 }
 
-#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
-float SincResampler::Convolve_SSE(const float* input_ptr, const float* k1,
-                                  const float* k2,
-                                  double kernel_interpolation_factor) {
-  // Ensure |k1|, |k2| are 16-byte aligned for SSE usage.  Should always be true
-  // so long as kKernelSize is a multiple of 16.
-  DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(k1) & 0x0F);
-  DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(k2) & 0x0F);
-
-  __m128 m_input;
-  __m128 m_sums1 = _mm_setzero_ps();
-  __m128 m_sums2 = _mm_setzero_ps();
-
-  // Based on |input_ptr| alignment, we need to use loadu or load.  Unrolling
-  // these loops hurt performance in local testing.
-  if (reinterpret_cast<uintptr_t>(input_ptr) & 0x0F) {
-    for (int i = 0; i < kKernelSize; i += 4) {
-      m_input = _mm_loadu_ps(input_ptr + i);
-      m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i)));
-      m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i)));
-    }
-  } else {
-    for (int i = 0; i < kKernelSize; i += 4) {
-      m_input = _mm_load_ps(input_ptr + i);
-      m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i)));
-      m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i)));
-    }
-  }
-
-  // Linearly interpolate the two "convolutions".
-  m_sums1 = _mm_mul_ps(m_sums1, _mm_set_ps1(1.0 - kernel_interpolation_factor));
-  m_sums2 = _mm_mul_ps(m_sums2, _mm_set_ps1(kernel_interpolation_factor));
-  m_sums1 = _mm_add_ps(m_sums1, m_sums2);
-
-  // Sum components together.
-  float result;
-  m_sums2 = _mm_add_ps(_mm_movehl_ps(m_sums1, m_sums1), m_sums1);
-  _mm_store_ss(&result, _mm_add_ss(m_sums2, _mm_shuffle_ps(
-      m_sums2, m_sums2, 1)));
-
-  return result;
-}
-#endif
-
 #if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
 float SincResampler::Convolve_NEON(const float* input_ptr, const float* k1,
                                    const float* k2,