Add SSE optimizations to SincResampler.

media/base/sinc_resampler.cc

Index: media/base/sinc_resampler.cc
diff --git a/media/base/sinc_resampler.cc b/media/base/sinc_resampler.cc
index 88e62044d27c6a77e416f93c8355bbeb094ffd5d..7a67cb0b6593fbb6940a056eb6bc41cce8bea327 100644
--- a/media/base/sinc_resampler.cc
+++ b/media/base/sinc_resampler.cc
@@ -38,8 +38,13 @@
 
 #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 {
@@ -63,13 +68,84 @@ enum {
   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).
@@ -96,6 +172,13 @@ SincResampler::SincResampler(double io_sample_rate_ratio, const ReadCB& read_cb)
   // 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);
@@ -168,36 +251,18 @@ void SincResampler::Resample(float* destination, int frames) {
       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_;