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!
+  // Step (2) -- Resample!  const what we can outside of the loop for speed.  It
+  // actually has an impact on ARM performance.  See inner loop comment below.
+  const double current_io_ratio = io_sample_rate_ratio_;
+  const float* const kernel_ptr = kernel_storage_.get();
   while (remaining_frames) {
-    while (virtual_source_idx_ < block_size_) {
+    // |i| may be negative if the last Resample() call ended on an iteration
+    // that put |virtual_source_idx_| over the limit.
+    //
+    // Note: The loop construct here can severely impact performance on ARM
+    // or when built with clang.  See https://codereview.chromium.org/18566009/
+    for (int i = ceil((block_size_ - virtual_source_idx_) / current_io_ratio);
+         i > 0; --i) {
+      DCHECK_LT(virtual_source_idx_, block_size_);
+
       // |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;
+      const float* const k1 = kernel_ptr + offset_idx * kKernelSize;
+      const float* const 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_|.
-      const float* input_ptr = r1_ + source_idx;
+      const float* const input_ptr = r1_ + source_idx;
 
       // Figure out how much to weight each kernel's "convolution".
       const double kernel_interpolation_factor =
           virtual_offset_idx - offset_idx;
       *destination++ = CONVOLVE_FUNC(
           input_ptr, k1, k2, kernel_interpolation_factor);
 
       // Advance the virtual index.
-      virtual_source_idx_ += io_sample_rate_ratio_;
+      virtual_source_idx_ += current_io_ratio;
 
       if (!--remaining_frames)
         return;
     }
 
     // Wrap back around to the start.
     virtual_source_idx_ -= block_size_;
 
     // Step (3) -- Copy r3_, r4_ to r1_, r2_.
     // This wraps the last input frames back to the start of the buffer.
@@ skipping 64 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