Add SincResampler ported from WebKit.

media/base/sinc_resampler.h

+// 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.
+
+#ifndef MEDIA_BASE_SINC_RESAMPLER_H_
+#define MEDIA_BASE_SINC_RESAMPLER_H_
+
+#include "base/callback.h"
+#include "base/memory/scoped_ptr.h"
+#include "media/base/media_export.h"
+
+namespace media {
+
+// SincResampler is a high-quality single-channel sample-rate converter.
+class MEDIA_EXPORT SincResampler {
+ public:
+  // Callback type for providing more data into the resampler.  Expects |frames|
+  // of data to be rendered into |destination|; zero padded if not enough frames
+  // are available to satisfy the request.
+  typedef base::Callback<void(float* destination, int frames)> ReadCB;
+
+  // Constructs a SincResampler with the specified |read_cb|, which is used to
+  // acquire audio data for resampling.  |io_sample_rate_ratio| is the ratio of
+  // input / output sample rates.
+  SincResampler(double io_sample_rate_ratio, const ReadCB& read_cb);
+  virtual ~SincResampler();
+
+  // Resample |frames| of data from |read_cb_| into |destination|.
+  void Resample(float* destination, int frames);
+
+  // The maximum size in frames that guarantees Resample() will only make a
+  // single call to |read_cb_| for more data.
+  int ChunkSize();
+
+ private:
+  enum {

[Ami GONE FROM CHROMIUM, 2012/07/10 03:23:00]

This entire enum can move to the top of the .cc file?

[DaleCurtis, 2012/07/10 21:38:55]

Done.

+    // The kernel size can be adjusted for quality (higher is better) at the
+    // expense of performance.  Must be an even number.
+    // TODO(dalecurtis): Test performance to see if we can jack this up to 64+.
+    kKernelSize = 32,
+
+    // The number of destination frames generated per processing pass.  Affects
+    // how often and for how much SincResampler calls back for input.  Must be
+    // greater than kKernelSize.
+    kBlockSize = 512,
+
+    // The kernel offset count is used for interpolation and is the number of
+    // sub-sample kernel shifts.  Can be adjusted for quality (higher is better)
+    // at the expense of allocating more memory.
+    kKernelOffsetCount = 32,
+    kKernelStorageSize = kKernelSize * (kKernelOffsetCount + 1),
+
+    // The size (in samples) of the internal buffer used by the resampler.
+    kBufferSize = kBlockSize + kKernelSize
+  };
+
+  // The ratio of input / output sample rates.
+  double io_sample_rate_ratio_;
+
+  // An index on the source input buffer with sub-sample precision.  It must be
+  // double precision to avoid drift.

[Ami GONE FROM CHROMIUM, 2012/07/10 03:23:00]

This gives me the heebie-jeebies.  Skimming the .cc file I'm led to believe this
can instead be an integral count of multiples of io_sample_rate_ratio_.  Would
that work for your math without requiring double-precision math to work like
infinite-precision math?

[DaleCurtis, 2012/07/10 05:24:48]

Not quite, it moves forward by io_sample_rate_ratio_ but backwards by
kBlockSize.  I'm not sure I follow your suggestion given that; tracking a count
seems prone to overflow and confusing.  Can you elaborate?

Keeping a count inside the outer loop, and saying virtual_source_idx = count *
io_sample_rate_ratio might yield some improvement in drift for little cost. I'll
test that tomorrow.

[Ami GONE FROM CHROMIUM, 2012/07/10 16:48:50]

int64 virtual_source_thingy_;
while (...) {
  int index = (virtual_source_thingy_++ * io_sample_rate_ratio_) % kBlockSize;
  ...
}

Possibly the fact that I don't have a good name for "thingy" is an indication
that there's a better way to think of all of this.

My worry is that in the current setup you're accumulating error with each
iteration and have no way to lose it.

My suggestion above also has a form of error accumulation, in that as thingy
grows, ratio's lower- and lower-order bits are used for the result.

It would be best to have a reset point so that error only accumulates during
processing of a single buffer/block/kernel.  If you track the *destination*
index count, and deduce the input index by dividing by the ratio_, then you can
reset the index when it hits the destination's sample rate (b/c you know at that
point the division by the ratio is precisely 1).  This would require this class
to take in the input & output sample rates (instead of the ratio), but would
allow you to contain the error accumulation to a single seconds' worth of data. 


WDYT?  Am I on crack?

[DaleCurtis, 2012/07/10 21:38:55]

So I experimented with this a couple different ways, without any measurable
effect in error over an hour of audio from 192kHz to 44.1kHz:
1. Adding an inner count.
2. Resetting virtual_source_idx_ to 0 when below machine epsilon after
subtracting kBlockSize. Which, surprisingly to me, happens very frequently.

I was sure #2 would have an impact on error, but the data doesn't show it. So my
conclusion is we don't need to worry about this.

+  double virtual_source_idx_;
+
+  // The buffer is primed once at the very beginning of processing.
+  bool buffer_primed_;
+
+  // Source of data for resampling.
+  ReadCB read_cb_;
+
+  // Contains kKernelOffsetCount kernels back-to-back, each of size kKernelSize.
+  // The kernel offsets are sub-sample shifts of a windowed sinc shifted from
+  // 0.0 to 1.0 sample.
+  scoped_array<float> kernel_storage_;
+
+  // Data from the source is copied into this buffer for each processing pass.
+  scoped_array<float> input_buffer_;
+
+  void InitializeKernel();

[Ami GONE FROM CHROMIUM, 2012/07/10 03:23:00]

Methods go above members.

[DaleCurtis, 2012/07/10 21:38:55]

Done.

+};
+
+}  // namespace media
+
+#endif  // MEDIA_BASE_SINC_RESAMPLER_H_