Reduce jitter from uneven SincResampler buffer size requests.

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_):
+// Initial input buffer layout, dividing into regions r0_ to r4_:
 //
 // |----------------|-----------------------------------------|----------------|
 //
-//                                   kBlockSize + kKernelSize / 2
+//                               1st request: request_frames_
 //                   <--------------------------------------------------------->
 //                                              r0_
 //
+//                 block_size_ = request_frames_ - kKernelSize / 2
+//                   <--------------------------------------->
+//
 //  kKernelSize / 2   kKernelSize / 2         kKernelSize / 2   kKernelSize / 2
 // <---------------> <--------------->       <---------------> <--------------->
 //        r1_               r2_                     r3_               r4_
 //
-//                                                     kBlockSize
-//                                     <--------------------------------------->
-//                                                        r5_
+// On the second request, block_size_ increases to request_frames_ while r0_,
+// r3_, and r4_ slide to the right by kKernelSize / 2:
+//
+// |----------------|-----------------------------------------|----------------|
+//
+//                                              2nd request: request_frames_
+//                                    <------------------ ... ----------------->
+//
+//                          block_size_ = request_frames_
+//                   <---------------- ... ------------------>
+//
+// These new regions remain constant until a Flush() occurs.  While complicated,
+// this allows us to reduce jitter by always requesting the same amount from the
+// provided callback.
 //
 // The algorithm:
 //
-// 1) Consume input frames into r0_ (r1_ is zero-initialized).
+// 1) Consume |request_frames_| frames into r0_ (r1_ is zero-initialized).
 // 2) Position kernel centered at start of r0_ (r2_) and generate output frames
 //    until kernel is centered at start of r4_ or we've finished generating all
 //    the output frames.
-// 3) Copy r3_ to r1_ and r4_ to r2_.
-// 4) Consume input frames into r5_ (zero-pad if we run out of input).
-// 5) Goto (2) until all of input is consumed.
+// 3) Copy r3_ to r1_, r4_ to r2_.
+// 4) If we're on the second load, set block_size_ equal to request_frames_
+//    and reinitialize r0_, r3_, and r4_ appropriately.

[Chris Rogers, 2013/05/06 21:17:31]

It's very hard for me to follow this new diagram, and verify exactly what's
happening with the code below.  I literally don't understand it.  It's not clear
what "appropriately" means, etc.

Is there some way to work on the diagram and explanation to help make it more
approachable?  Maybe there's a way to re-frame the description of the algorithm
in different terms (as far as diagram and explanation) without actually changing
the algorithm.

[DaleCurtis, 2013/05/06 22:52:09]

PTAL. I've rewritten the entire documentation section as well as simplified the
number of moving pieces to a single leader: r0_.

+// 5) Goto (1).
 //
 // 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>
@@ skipping 17 matching lines @@
   // windowing it the transition from pass to stop does not happen right away.
   // So we should adjust the low pass filter cutoff slightly downward to avoid
   // some aliasing at the very high-end.
   // TODO(crogers): this value is empirical and to be more exact should vary
   // depending on kKernelSize.
   sinc_scale_factor *= 0.9;
 
   return sinc_scale_factor;
 }
 
-SincResampler::SincResampler(double io_sample_rate_ratio, const ReadCB& read_cb)
+SincResampler::SincResampler(double io_sample_rate_ratio,
+                             size_t request_frames,
+                             const ReadCB& read_cb)
     : io_sample_rate_ratio_(io_sample_rate_ratio),
-      virtual_source_idx_(0),
-      buffer_primed_(false),
       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) * kBufferSize, 16))),
+          base::AlignedAlloc(sizeof(float) * input_buffer_size_, 16))),
 #if defined(ARCH_CPU_X86_FAMILY) && !defined(__SSE__)
       convolve_proc_(base::CPU().has_sse() ? Convolve_SSE : Convolve_C),
 #endif
-      // 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) {
-  // Ensure kKernelSize is a multiple of 32 for easy SSE optimizations; causes
-  // r0_ and r5_ (used for input) to always be 16-byte aligned by virtue of
-  // input_buffer_ being 16-byte aligned.
-  DCHECK_EQ(kKernelSize % 32, 0) << "kKernelSize must be a multiple of 32!";
-  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);
+      reinitialize_regions_(false) {
+  Flush();
+  CHECK_GT(block_size_, static_cast<size_t>(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);
-  memset(input_buffer_.get(), 0, sizeof(*input_buffer_.get()) * kBufferSize);
 
   InitializeKernel();
 }
 
 SincResampler::~SincResampler() {}
 
+void SincResampler::UpdateRegions() {
+  // Setup various region pointers in the buffer (see diagram above).
+  r1_ = input_buffer_.get();
+  r2_ = input_buffer_.get() + kKernelSize / 2;
+  r0_ = reinitialize_regions_ ? input_buffer_.get() + kKernelSize : r2_;
+  r3_ = r2_ + block_size_ - kKernelSize / 2;
+  r4_ = r2_ + block_size_;
+
+  // r1_ at the beginning of the buffer.
+  CHECK_EQ(r1_, input_buffer_.get());
+  // r1_ left of r2_, r4_ left of r3_ and size correct.
+  CHECK_EQ(r2_ - r1_, r4_ - r3_);
+  // r2_ left of r3.
+  CHECK_LT(r2_, r3_);
+}
+
 void SincResampler::InitializeKernel() {
   // Blackman window parameters.
   static const double kAlpha = 0.16;
   static const double kA0 = 0.5 * (1.0 - kAlpha);
   static const double kA1 = 0.5;
   static const double kA2 = 0.5 * kAlpha;
 
   // Generates a set of windowed sinc() kernels.
   // We generate a range of sub-sample offsets from 0.0 to 1.0.
   const double sinc_scale_factor = SincScaleFactor(io_sample_rate_ratio_);
@@ skipping 60 matching lines @@
 // TODO(dalecurtis): Once Chrome moves to a SSE baseline this can be removed.
 #define CONVOLVE_FUNC convolve_proc_
 #endif
 #elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
 #define CONVOLVE_FUNC Convolve_NEON
 #else
 // Unknown architecture.
 #define CONVOLVE_FUNC Convolve_C
 #endif
 
-void SincResampler::Resample(float* destination, int frames) {
+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_) {
-    read_cb_.Run(r0_, kBlockSize + kKernelSize / 2);
+    read_cb_.Run(request_frames_, r0_);
     buffer_primed_ = true;
+    reinitialize_regions_ = true;
   }
 
   // Step (2) -- Resample!
   while (remaining_frames) {
-    while (virtual_source_idx_ < kBlockSize) {
+    while (virtual_source_idx_ < block_size_) {
       // |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;
+      const int source_idx = virtual_source_idx_;
+      const double subsample_remainder = virtual_source_idx_ - source_idx;
 
-      double virtual_offset_idx = subsample_remainder * kKernelOffsetCount;
-      int offset_idx = static_cast<int>(virtual_offset_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_|.
-      float* k1 = kernel_storage_.get() + offset_idx * kKernelSize;
-      float* k2 = k1 + kKernelSize;
+      const float* k1 = kernel_storage_.get() + offset_idx * kKernelSize;
+      const float* 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_|.
-      float* input_ptr = r1_ + source_idx;
+      const float* input_ptr = r1_ + source_idx;
 
       // Figure out how much to weight each kernel's "convolution".
-      double kernel_interpolation_factor = virtual_offset_idx - offset_idx;
+      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_;
 
       if (!--remaining_frames)
         return;
     }
 
     // Wrap back around to the start.
-    virtual_source_idx_ -= kBlockSize;
+    virtual_source_idx_ -= block_size_;
 
-    // Step (3) Copy r3_ to r1_ and r4_ to r2_.
+    // Step (3) -- Copy r3_, r4_ to r1_, 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));
+    memcpy(r1_, r3_, sizeof(*input_buffer_.get()) * kKernelSize);
 
-    // Step (4)
-    // Refresh the buffer with more input.
-    read_cb_.Run(r5_, kBlockSize);
+    // Step (4) -- Reinitialize regions if necessary.
+    if (reinitialize_regions_) {
+      block_size_ = request_frames_;
+      UpdateRegions();
+      reinitialize_regions_ = false;
+    }
+
+    // Step (5) -- Refresh the buffer with more input.
+    read_cb_.Run(request_frames_, r0_);
   }
 }
 
 #undef CONVOLVE_FUNC
 
 int SincResampler::ChunkSize() const {
-  return kBlockSize / io_sample_rate_ratio_;
+  return block_size_ / io_sample_rate_ratio_;
 }
 
 void SincResampler::Flush() {
   virtual_source_idx_ = 0;
   buffer_primed_ = false;
-  memset(input_buffer_.get(), 0, sizeof(*input_buffer_.get()) * kBufferSize);
+  memset(input_buffer_.get(), 0,
+         sizeof(*input_buffer_.get()) * input_buffer_size_);
+
+  block_size_ = request_frames_ - kKernelSize / 2;
+  reinitialize_regions_ = false;
+  UpdateRegions();
 }
 
 float SincResampler::Convolve_C(const float* input_ptr, const float* k1,
                                 const 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.
@@ skipping 31 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