Optimize loop condition for 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 1df8d4d08f894201a823c15c2066ca3ad43904ec..0ceb7a85e271a155a244fbe0ad9ca982dbdf1967 100644
--- a/media/base/sinc_resampler.cc
+++ b/media/base/sinc_resampler.cc
@@ -138,7 +138,7 @@ 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),
@@ -155,8 +155,9 @@ SincResampler::SincResampler(double io_sample_rate_ratio,
           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,
@@ -255,17 +256,26 @@ 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 =