Add SSE optimizations to SincResampler.

media/base/sinc_resampler.cc

 // Copyright (c) 2012 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.
 //
 // Input buffer layout, dividing the total buffer into regions (r0_ - r5_):
 //
 // |----------------|-----------------------------------------|----------------|
 //
 //                                   kBlockSize + kKernelSize / 2
 //                   <--------------------------------------------------------->
@@ skipping 20 matching lines @@
 // Note: we're glossing over how the sub-sample handling works with
 // |virtual_source_idx_|, etc.
 
 // MSVC++ requires this to be set before any other includes to get M_PI.
 #define _USE_MATH_DEFINES
 
 #include "media/base/sinc_resampler.h"
 
 #include <cmath>
 
+#include "base/cpu.h"
 #include "base/logging.h"
 
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+#include <xmmintrin.h>
+#endif
+
 namespace media {
 
 enum {
   // The kernel size can be adjusted for quality (higher is better) at the
   // expense of performance.  Must be an even number.
   // 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
 };
 
+
+#define CONVOLVE_ONE_SAMPLE \

[Ami GONE FROM CHROMIUM, 2012/07/18 04:40:35]

#undef this when done with it.
Except you only use this once, so why not just inline?

[DaleCurtis, 2012/07/21 02:35:14]

Good point. Leftovers from the WebKit version of the optimizations.

+    sum1 += *input_ptr * *k1++; \
+    sum2 += *input_ptr++ * *k2++;
+
+static double Convolve_C(float *input_ptr, float* k1, float* k2,
+                         double kernel_interpolation_factor) {
+  float sum1 = 0;
+  float sum2 = 0;
+
+  // Generate a single output sample.  Unrolling this loop hurt performance in
+  // local testing.
+  int n = kKernelSize;
+  while (n--) {
+    CONVOLVE_ONE_SAMPLE
+  }
+
+  // Linearly interpolate the two "convolutions".
+  return (1.0 - kernel_interpolation_factor) * sum1
+      + kernel_interpolation_factor * sum2;
+}
+
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+#define CONVOLVE_4_SAMPLES(load) \

[Ami GONE FROM CHROMIUM, 2012/07/18 04:40:35]

Any reason to use a #define instead of an inline function?

[DaleCurtis, 2012/07/21 02:35:14]

Not sure how that'd work since we're modifying the instruction used. In any case
I've moved this directly into each for loop for readability.

+    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)));
+
+static const int kFloatsPerPass = sizeof(__m128) / sizeof(float);

[Ami GONE FROM CHROMIUM, 2012/07/18 04:40:35]

Not sure what this generality buys you given you've hardcoded _4_ above
already...

[DaleCurtis, 2012/07/21 02:35:14]

Removed.

+static double Convolve_SSE(float *input_ptr, float* k1, float* k2,

[Ami GONE FROM CHROMIUM, 2012/07/18 04:40:35]

i can haz a test that shows the two Convolve_* functions return equal values?
(obvs., that test would only run on some machines)

[DaleCurtis, 2012/07/21 02:35:14]

Done.

+                            double kernel_interpolation_factor) {
+  // Ensure |k1|, |k2| are aligned for SSE usage.  Should always be true so long
+  // as kKernelSize is a power of 2.
+  DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(k1) & (sizeof(__m128) - 1));
+  DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(k2) & (sizeof(__m128) - 1));
+
+  __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) & (sizeof(__m128) - 1)) {
+    for (int i = 0; i < kKernelSize; i += kFloatsPerPass) {
+      CONVOLVE_4_SAMPLES(loadu)
+    }
+  } else {
+    for (int i = 0; i < kKernelSize; i += kFloatsPerPass) {
+      CONVOLVE_4_SAMPLES(load)
+    }
+  }
+
+  // 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
+
+typedef double (*ConvolveProc)(float* src, float* k1, float* k2,
+                               double kernel_interpolation_factor);
+static ConvolveProc convolve_proc = NULL;
+
 SincResampler::SincResampler(double io_sample_rate_ratio, const ReadCB& read_cb)
     : io_sample_rate_ratio_(io_sample_rate_ratio),
       virtual_source_idx_(0),
       buffer_primed_(false),
       read_cb_(read_cb),
-      // TODO(dalecurtis): When we switch to AVX/SSE optimization, we'll need to
-      // allocate with 32-byte alignment and ensure they're sized % 32 bytes.
+      // TODO(dalecurtis): When we switch to SSE optimization, we'll need to
+      // allocate with 16-byte alignment (default linux, mac, not win)

[Ami GONE FROM CHROMIUM, 2012/07/18 04:40:35]

IDK what this means

[DaleCurtis, 2012/07/21 02:35:14]

Changed. Was a reminder to me to land my AlignedMemory CL and change this before
this lands.

       kernel_storage_(new float[kKernelStorageSize]),
       input_buffer_(new float[kBufferSize]),
       // Setup various region pointers in the buffer (see diagram above).
       r0_(input_buffer_.get() + kKernelSize / 2),
       r1_(input_buffer_.get()),
       r2_(r0_),
       r3_(r0_ + kBlockSize - kKernelSize / 2),
       r4_(r0_ + kBlockSize),
       r5_(r0_ + kKernelSize / 2) {
   DCHECK_EQ(kKernelSize % 2, 0) << "kKernelSize must be even!";
   DCHECK_GT(kBlockSize, kKernelSize)
       << "kBlockSize must be greater than kKernelSize!";
   // Basic sanity checks to ensure buffer regions are laid out correctly:
   // r0_ and r2_ should always be the same position.
   DCHECK_EQ(r0_, r2_);
   // r1_ at the beginning of the buffer.
   DCHECK_EQ(r1_, input_buffer_.get());
   // r1_ left of r2_, r2_ left of r5_ and r1_, r2_ size correct.
   DCHECK_EQ(r2_ - r1_, r5_ - r2_);
   // r3_ left of r4_, r5_ left of r0_ and r3_ size correct.
   DCHECK_EQ(r4_ - r3_, r5_ - r0_);
   // r3_, r4_ size correct and r4_ at the end of the buffer.
   DCHECK_EQ(r4_ + (r4_ - r3_), r1_ + kBufferSize);
   // r5_ size correct and at the end of the buffer.
   DCHECK_EQ(r5_ + kBlockSize, r1_ + kBufferSize);
 
+  if (!convolve_proc) {

[Ami GONE FROM CHROMIUM, 2012/07/18 04:40:35]

This is racy.

[DaleCurtis, 2012/07/21 02:35:14]

Done.

+    convolve_proc = Convolve_C;
+    base::CPU cpu;
+    if (cpu.has_sse())
+      convolve_proc = Convolve_SSE;

[Ami GONE FROM CHROMIUM, 2012/07/18 04:40:35]

Does this compile on build platforms that don't qualify for the #ifdef above? 

[DaleCurtis, 2012/07/21 02:35:14]

Done.

+  }
+
   memset(kernel_storage_.get(), 0,
          sizeof(*kernel_storage_.get()) * kKernelStorageSize);
   memset(input_buffer_.get(), 0, sizeof(*input_buffer_.get()) * kBufferSize);
 
   InitializeKernel();
 }
 
 SincResampler::~SincResampler() {}
 
 void SincResampler::InitializeKernel() {
@@ skipping 52 matching lines @@
   while (remaining_frames) {
     while (virtual_source_idx_ < kBlockSize) {
       // |virtual_source_idx_| lies in between two kernel offsets so figure out
       // what they are.
       int source_idx = static_cast<int>(virtual_source_idx_);
       double subsample_remainder = virtual_source_idx_ - source_idx;
 
       double virtual_offset_idx = subsample_remainder * kKernelOffsetCount;
       int offset_idx = static_cast<int>(virtual_offset_idx);
 
+      // We'll compute "convolutions" for the two kernels which straddle
+      // |virtual_source_idx_|.
       float* k1 = kernel_storage_.get() + offset_idx * kKernelSize;
       float* k2 = k1 + kKernelSize;
 
       // Initialize input pointer based on quantized |virtual_source_idx_|.
       float* input_ptr = r1_ + source_idx;
 
-      // We'll compute "convolutions" for the two kernels which straddle
-      // |virtual_source_idx_|.
-      float sum1 = 0;
-      float sum2 = 0;
-
       // Figure out how much to weight each kernel's "convolution".
       double kernel_interpolation_factor = virtual_offset_idx - offset_idx;
-
-      // Generate a single output sample.
-      int n = kKernelSize;
-      float input;
-      // TODO(dalecurtis): For initial commit, I've ripped out all the SSE
-      // optimizations, these definitely need to go back in before release.
-      while (n--) {
-        input = *input_ptr++;
-        sum1 += input * *k1++;
-        sum2 += input * *k2++;
-      }
-
-      // Linearly interpolate the two "convolutions".
-      double result = (1.0 - kernel_interpolation_factor) * sum1
-          + kernel_interpolation_factor * sum2;
-
-      *destination++ = result;
+      *destination++ = convolve_proc(
+          input_ptr, k1, k2, kernel_interpolation_factor);
 
       // Advance the virtual index.
       virtual_source_idx_ += io_sample_rate_ratio_;
 
       if (!--remaining_frames)
         return;
     }
 
     // Wrap back around to the start.
     virtual_source_idx_ -= kBlockSize;
 
     // Step (3) Copy r3_ to r1_ and r4_ to r2_.
     // This wraps the last input frames back to the start of the buffer.
     memcpy(r1_, r3_, sizeof(*input_buffer_.get()) * (kKernelSize / 2));
     memcpy(r2_, r4_, sizeof(*input_buffer_.get()) * (kKernelSize / 2));
 
     // Step (4)
     // Refresh the buffer with more input.
     read_cb_.Run(r5_, kBlockSize);
   }
 }
 
 int SincResampler::ChunkSize() {
   return kBlockSize / io_sample_rate_ratio_;
 }
 
 }  // namespace media