Optimize loop condition for 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.
 //
 // Initial input buffer layout, dividing into regions r0_ to r4_ (note: r0_, r3_
 // and r4_ will move after the first load):
 //
 // |----------------|-----------------------------------------|----------------|
 //
 //                                        request_frames_
@@ skipping 120 matching lines @@
 #elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
 #define CONVOLVE_FUNC Convolve_NEON
 void SincResampler::InitializeCPUSpecificFeatures() {}
 #else
 // Unknown architecture.
 #define CONVOLVE_FUNC Convolve_C
 void SincResampler::InitializeCPUSpecificFeatures() {}
 #endif
 
 SincResampler::SincResampler(double io_sample_rate_ratio,
-                             size_t request_frames,
+                             int request_frames,
                              const ReadCB& read_cb)
     : io_sample_rate_ratio_(io_sample_rate_ratio),
       read_cb_(read_cb),
       request_frames_(request_frames),
       input_buffer_size_(request_frames_ + kKernelSize),
       // Create input buffers with a 16-byte alignment for SSE optimizations.
       kernel_storage_(static_cast<float*>(
           base::AlignedAlloc(sizeof(float) * kKernelStorageSize, 16))),
       kernel_pre_sinc_storage_(static_cast<float*>(
           base::AlignedAlloc(sizeof(float) * kKernelStorageSize, 16))),
       kernel_window_storage_(static_cast<float*>(
           base::AlignedAlloc(sizeof(float) * kKernelStorageSize, 16))),
       input_buffer_(static_cast<float*>(
           base::AlignedAlloc(sizeof(float) * input_buffer_size_, 16))),
       r1_(input_buffer_.get()),
       r2_(input_buffer_.get() + kKernelSize / 2) {
+  CHECK_GT(request_frames_, 0);
   Flush();
-  CHECK_GT(block_size_, static_cast<size_t>(kKernelSize))
+  CHECK_GT(block_size_, kKernelSize)
       << "block_size must be greater than kKernelSize!";
 
   memset(kernel_storage_.get(), 0,
          sizeof(*kernel_storage_.get()) * kKernelStorageSize);
   memset(kernel_pre_sinc_storage_.get(), 0,
          sizeof(*kernel_pre_sinc_storage_.get()) * kKernelStorageSize);
   memset(kernel_window_storage_.get(), 0,
          sizeof(*kernel_window_storage_.get()) * kKernelStorageSize);
 
   InitializeKernel();
@@ skipping 78 matching lines @@
             window * sin(sinc_scale_factor * pre_sinc) / pre_sinc;
       }
     }
   }
 }
 
 void SincResampler::Resample(int frames, float* destination) {
   int remaining_frames = frames;
 
   // Step (1) -- Prime the input buffer at the start of the input stream.
-  if (!buffer_primed_) {
+  if (!buffer_primed_ && remaining_frames) {
     read_cb_.Run(request_frames_, r0_);
     buffer_primed_ = true;
   }
 
   // Step (2) -- Resample!
   while (remaining_frames) {
+    // Avoid comparing a double to an int on most platforms.  Provides a 3% to
+    // 20% increase in some cases.  http://llvm.org/bugs/show_bug.cgi?id=16578
+#if defined(USE_NEON) || defined(__clang__)
     while (virtual_source_idx_ < block_size_) {

[Ami GONE FROM CHROMIUM, 2013/07/09 23:20:08]

what happens if you try:
while ((const int source_idx = virtual_source_idx) < block_size_)
on the various platforms?

[DaleCurtis, 2013/07/10 21:06:31]

As discussed offline, this doesn't compile, and moving the decl outside of the
while shows the same performance hit seen elsewhere.

[Ami GONE FROM CHROMIUM, 2013/07/10 21:18:00]

FTR that's only b/c of the "const"; declaring the var in loop scope should still
work.
:(

+      const int source_idx = virtual_source_idx_;
+#else
+    while (true) {
+      const int source_idx = virtual_source_idx_;
+      if (source_idx >= block_size_)
+        break;
+#endif

[Ami GONE FROM CHROMIUM, 2013/07/10 21:18:00]

Perverted sense of perversion leads me to wonder how this performs against

for (int loop_iters = (block_size_ - virtual_source_idx_) /
io_sample_rate_ratio_; loop_iters; --loop_iters)

[DaleCurtis, 2013/07/10 23:01:34]

Works well on my N4 w/ gcc and OSX w/ clang. I've pinged the guy with the A9 to
see how it does for him.  Strangely if I try to optimize further by making it
std::min(loop_iters, remaining_frames) I hit the 20% performance regression
previously seen with clang...

[DaleCurtis, 2013/07/11 21:04:34]

Switched to for(), but needs a ceil() to be accurate.  Still faster than the
existing approach.  Added some DCHECKs() to verify accuracy.

       // |virtual_source_idx_| lies in between two kernel offsets so figure out
       // what they are.
-      const int source_idx = virtual_source_idx_;
       const double subsample_remainder = virtual_source_idx_ - source_idx;
 
       const double virtual_offset_idx =
           subsample_remainder * kKernelOffsetCount;
       const int offset_idx = virtual_offset_idx;
 
       // We'll compute "convolutions" for the two kernels which straddle
       // |virtual_source_idx_|.
       const float* k1 = kernel_storage_.get() + offset_idx * kKernelSize;
       const float* k2 = k1 + kKernelSize;
@@ skipping 91 matching lines @@
       vmulq_f32(m_sums1, vmovq_n_f32(1.0 - kernel_interpolation_factor)),
       m_sums2, vmovq_n_f32(kernel_interpolation_factor));
 
   // Sum components together.
   float32x2_t m_half = vadd_f32(vget_high_f32(m_sums1), vget_low_f32(m_sums1));
   return vget_lane_f32(vpadd_f32(m_half, m_half), 0);
 }
 #endif
 
 }  // namespace media