2 * Copyright (c) 2013 The WebRTC project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
11 // Modified from the Chromium original:
12 // src/media/base/sinc_resampler.cc
14 // Initial input buffer layout, dividing into regions r0_ to r4_ (note: r0_, r3_
15 // and r4_ will move after the first load):
17 // |----------------|-----------------------------------------|----------------|
20 // <--------------------------------------------------------->
21 // r0_ (during first load)
23 // kKernelSize / 2 kKernelSize / 2 kKernelSize / 2 kKernelSize / 2
24 // <---------------> <---------------> <---------------> <--------------->
27 // block_size_ == r4_ - r2_
28 // <--------------------------------------->
31 // <------------------ ... ----------------->
32 // r0_ (during second load)
34 // On the second request r0_ slides to the right by kKernelSize / 2 and r3_, r4_
35 // and block_size_ are reinitialized via step (3) in the algorithm below.
37 // These new regions remain constant until a Flush() occurs. While complicated,
38 // this allows us to reduce jitter by always requesting the same amount from the
43 // 1) Allocate input_buffer of size: request_frames_ + kKernelSize; this ensures
44 // there's enough room to read request_frames_ from the callback into region
45 // r0_ (which will move between the first and subsequent passes).
47 // 2) Let r1_, r2_ each represent half the kernel centered around r0_:
49 // r0_ = input_buffer_ + kKernelSize / 2
50 // r1_ = input_buffer_
53 // r0_ is always request_frames_ in size. r1_, r2_ are kKernelSize / 2 in
54 // size. r1_ must be zero initialized to avoid convolution with garbage (see
57 // 3) Let r3_, r4_ each represent half the kernel right aligned with the end of
58 // r0_ and choose block_size_ as the distance in frames between r4_ and r2_:
60 // r3_ = r0_ + request_frames_ - kKernelSize
61 // r4_ = r0_ + request_frames_ - kKernelSize / 2
62 // block_size_ = r4_ - r2_ = request_frames_ - kKernelSize / 2
64 // 4) Consume request_frames_ frames into r0_.
66 // 5) Position kernel centered at start of r2_ and generate output frames until
67 // the kernel is centered at the start of r4_ or we've finished generating
68 // all the output frames.
70 // 6) Wrap left over data from the r3_ to r1_ and r4_ to r2_.
72 // 7) If we're on the second load, in order to avoid overwriting the frames we
73 // just wrapped from r4_ we need to slide r0_ to the right by the size of
74 // r4_, which is kKernelSize / 2:
76 // r0_ = r0_ + kKernelSize / 2 = input_buffer_ + kKernelSize
78 // r3_, r4_, and block_size_ then need to be reinitialized, so goto (3).
80 // 8) Else, if we're not on the second load, goto (4).
82 // Note: we're glossing over how the sub-sample handling works with
83 // |virtual_source_idx_|, etc.
85 // MSVC++ requires this to be set before any other includes to get M_PI.
86 #define _USE_MATH_DEFINES
88 #include "webrtc/common_audio/resampler/sinc_resampler.h"
89 #include "webrtc/system_wrappers/interface/compile_assert.h"
90 #include "webrtc/system_wrappers/interface/cpu_features_wrapper.h"
91 #include "webrtc/typedefs.h"
100 static double SincScaleFactor(double io_ratio) {
101 // |sinc_scale_factor| is basically the normalized cutoff frequency of the
103 double sinc_scale_factor = io_ratio > 1.0 ? 1.0 / io_ratio : 1.0;
105 // The sinc function is an idealized brick-wall filter, but since we're
106 // windowing it the transition from pass to stop does not happen right away.
107 // So we should adjust the low pass filter cutoff slightly downward to avoid
108 // some aliasing at the very high-end.
109 // TODO(crogers): this value is empirical and to be more exact should vary
110 // depending on kKernelSize.
111 sinc_scale_factor *= 0.9;
113 return sinc_scale_factor;
116 // If we know the minimum architecture at compile time, avoid CPU detection.
117 #if defined(WEBRTC_ARCH_X86_FAMILY)
118 #if defined(__SSE2__)
119 #define CONVOLVE_FUNC Convolve_SSE
120 void SincResampler::InitializeCPUSpecificFeatures() {}
122 // x86 CPU detection required. Function will be set by
123 // InitializeCPUSpecificFeatures().
124 // TODO(dalecurtis): Once Chrome moves to an SSE baseline this can be removed.
125 #define CONVOLVE_FUNC convolve_proc_
127 void SincResampler::InitializeCPUSpecificFeatures() {
128 convolve_proc_ = WebRtc_GetCPUInfo(kSSE2) ? Convolve_SSE : Convolve_C;
131 #elif defined(WEBRTC_ARCH_ARM_V7)
132 #if defined(WEBRTC_ARCH_ARM_NEON)
133 #define CONVOLVE_FUNC Convolve_NEON
134 void SincResampler::InitializeCPUSpecificFeatures() {}
136 // ARM CPU detection required. Function will be set by
137 // InitializeCPUSpecificFeatures().
138 #define CONVOLVE_FUNC convolve_proc_
140 void SincResampler::InitializeCPUSpecificFeatures() {
141 convolve_proc_ = WebRtc_GetCPUFeaturesARM() & kCPUFeatureNEON ?
142 Convolve_NEON : Convolve_C;
146 // Unknown architecture.
147 #define CONVOLVE_FUNC Convolve_C
148 void SincResampler::InitializeCPUSpecificFeatures() {}
151 SincResampler::SincResampler(double io_sample_rate_ratio,
153 SincResamplerCallback* read_cb)
154 : io_sample_rate_ratio_(io_sample_rate_ratio),
156 request_frames_(request_frames),
157 input_buffer_size_(request_frames_ + kKernelSize),
158 // Create input buffers with a 16-byte alignment for SSE optimizations.
159 kernel_storage_(static_cast<float*>(
160 AlignedMalloc(sizeof(float) * kKernelStorageSize, 16))),
161 kernel_pre_sinc_storage_(static_cast<float*>(
162 AlignedMalloc(sizeof(float) * kKernelStorageSize, 16))),
163 kernel_window_storage_(static_cast<float*>(
164 AlignedMalloc(sizeof(float) * kKernelStorageSize, 16))),
165 input_buffer_(static_cast<float*>(
166 AlignedMalloc(sizeof(float) * input_buffer_size_, 16))),
167 #if defined(WEBRTC_CPU_DETECTION)
168 convolve_proc_(NULL),
170 r1_(input_buffer_.get()),
171 r2_(input_buffer_.get() + kKernelSize / 2) {
172 #if defined(WEBRTC_CPU_DETECTION)
173 InitializeCPUSpecificFeatures();
174 assert(convolve_proc_);
176 assert(request_frames_ > 0);
178 assert(block_size_ > kKernelSize);
180 memset(kernel_storage_.get(), 0,
181 sizeof(*kernel_storage_.get()) * kKernelStorageSize);
182 memset(kernel_pre_sinc_storage_.get(), 0,
183 sizeof(*kernel_pre_sinc_storage_.get()) * kKernelStorageSize);
184 memset(kernel_window_storage_.get(), 0,
185 sizeof(*kernel_window_storage_.get()) * kKernelStorageSize);
190 SincResampler::~SincResampler() {}
192 void SincResampler::UpdateRegions(bool second_load) {
193 // Setup various region pointers in the buffer (see diagram above). If we're
194 // on the second load we need to slide r0_ to the right by kKernelSize / 2.
195 r0_ = input_buffer_.get() + (second_load ? kKernelSize : kKernelSize / 2);
196 r3_ = r0_ + request_frames_ - kKernelSize;
197 r4_ = r0_ + request_frames_ - kKernelSize / 2;
198 block_size_ = r4_ - r2_;
200 // r1_ at the beginning of the buffer.
201 assert(r1_ == input_buffer_.get());
202 // r1_ left of r2_, r4_ left of r3_ and size correct.
203 assert(r2_ - r1_ == r4_ - r3_);
208 void SincResampler::InitializeKernel() {
209 // Blackman window parameters.
210 static const double kAlpha = 0.16;
211 static const double kA0 = 0.5 * (1.0 - kAlpha);
212 static const double kA1 = 0.5;
213 static const double kA2 = 0.5 * kAlpha;
215 // Generates a set of windowed sinc() kernels.
216 // We generate a range of sub-sample offsets from 0.0 to 1.0.
217 const double sinc_scale_factor = SincScaleFactor(io_sample_rate_ratio_);
218 for (int offset_idx = 0; offset_idx <= kKernelOffsetCount; ++offset_idx) {
219 const float subsample_offset =
220 static_cast<float>(offset_idx) / kKernelOffsetCount;
222 for (int i = 0; i < kKernelSize; ++i) {
223 const int idx = i + offset_idx * kKernelSize;
224 const float pre_sinc = M_PI * (i - kKernelSize / 2 - subsample_offset);
225 kernel_pre_sinc_storage_[idx] = pre_sinc;
227 // Compute Blackman window, matching the offset of the sinc().
228 const float x = (i - subsample_offset) / kKernelSize;
229 const float window = kA0 - kA1 * cos(2.0 * M_PI * x) + kA2
230 * cos(4.0 * M_PI * x);
231 kernel_window_storage_[idx] = window;
233 // Compute the sinc with offset, then window the sinc() function and store
234 // at the correct offset.
236 kernel_storage_[idx] = sinc_scale_factor * window;
238 kernel_storage_[idx] =
239 window * sin(sinc_scale_factor * pre_sinc) / pre_sinc;
245 void SincResampler::SetRatio(double io_sample_rate_ratio) {
246 if (fabs(io_sample_rate_ratio_ - io_sample_rate_ratio) <
247 std::numeric_limits<double>::epsilon()) {
251 io_sample_rate_ratio_ = io_sample_rate_ratio;
253 // Optimize reinitialization by reusing values which are independent of
254 // |sinc_scale_factor|. Provides a 3x speedup.
255 const double sinc_scale_factor = SincScaleFactor(io_sample_rate_ratio_);
256 for (int offset_idx = 0; offset_idx <= kKernelOffsetCount; ++offset_idx) {
257 for (int i = 0; i < kKernelSize; ++i) {
258 const int idx = i + offset_idx * kKernelSize;
259 const float window = kernel_window_storage_[idx];
260 const float pre_sinc = kernel_pre_sinc_storage_[idx];
263 kernel_storage_[idx] = sinc_scale_factor * window;
265 kernel_storage_[idx] =
266 window * sin(sinc_scale_factor * pre_sinc) / pre_sinc;
272 void SincResampler::Resample(int frames, float* destination) {
273 int remaining_frames = frames;
275 // Step (1) -- Prime the input buffer at the start of the input stream.
276 if (!buffer_primed_ && remaining_frames) {
277 read_cb_->Run(request_frames_, r0_);
278 buffer_primed_ = true;
281 // Step (2) -- Resample! const what we can outside of the loop for speed. It
282 // actually has an impact on ARM performance. See inner loop comment below.
283 const double current_io_ratio = io_sample_rate_ratio_;
284 const float* const kernel_ptr = kernel_storage_.get();
285 while (remaining_frames) {
286 // |i| may be negative if the last Resample() call ended on an iteration
287 // that put |virtual_source_idx_| over the limit.
289 // Note: The loop construct here can severely impact performance on ARM
290 // or when built with clang. See https://codereview.chromium.org/18566009/
291 for (int i = ceil((block_size_ - virtual_source_idx_) / current_io_ratio);
293 assert(virtual_source_idx_ < block_size_);
295 // |virtual_source_idx_| lies in between two kernel offsets so figure out
297 const int source_idx = virtual_source_idx_;
298 const double subsample_remainder = virtual_source_idx_ - source_idx;
300 const double virtual_offset_idx =
301 subsample_remainder * kKernelOffsetCount;
302 const int offset_idx = virtual_offset_idx;
304 // We'll compute "convolutions" for the two kernels which straddle
305 // |virtual_source_idx_|.
306 const float* const k1 = kernel_ptr + offset_idx * kKernelSize;
307 const float* const k2 = k1 + kKernelSize;
309 // Ensure |k1|, |k2| are 16-byte aligned for SIMD usage. Should always be
310 // true so long as kKernelSize is a multiple of 16.
311 assert(0u == (reinterpret_cast<uintptr_t>(k1) & 0x0F));
312 assert(0u == (reinterpret_cast<uintptr_t>(k2) & 0x0F));
314 // Initialize input pointer based on quantized |virtual_source_idx_|.
315 const float* const input_ptr = r1_ + source_idx;
317 // Figure out how much to weight each kernel's "convolution".
318 const double kernel_interpolation_factor =
319 virtual_offset_idx - offset_idx;
320 *destination++ = CONVOLVE_FUNC(
321 input_ptr, k1, k2, kernel_interpolation_factor);
323 // Advance the virtual index.
324 virtual_source_idx_ += current_io_ratio;
326 if (!--remaining_frames)
330 // Wrap back around to the start.
331 virtual_source_idx_ -= block_size_;
333 // Step (3) -- Copy r3_, r4_ to r1_, r2_.
334 // This wraps the last input frames back to the start of the buffer.
335 memcpy(r1_, r3_, sizeof(*input_buffer_.get()) * kKernelSize);
337 // Step (4) -- Reinitialize regions if necessary.
341 // Step (5) -- Refresh the buffer with more input.
342 read_cb_->Run(request_frames_, r0_);
348 int SincResampler::ChunkSize() const {
349 return block_size_ / io_sample_rate_ratio_;
352 void SincResampler::Flush() {
353 virtual_source_idx_ = 0;
354 buffer_primed_ = false;
355 memset(input_buffer_.get(), 0,
356 sizeof(*input_buffer_.get()) * input_buffer_size_);
357 UpdateRegions(false);
360 float SincResampler::Convolve_C(const float* input_ptr, const float* k1,
362 double kernel_interpolation_factor) {
366 // Generate a single output sample. Unrolling this loop hurt performance in
370 sum1 += *input_ptr * *k1++;
371 sum2 += *input_ptr++ * *k2++;
374 // Linearly interpolate the two "convolutions".
375 return (1.0 - kernel_interpolation_factor) * sum1
376 + kernel_interpolation_factor * sum2;
379 } // namespace webrtc