1 /*
2 * Copyright (c) 2013 The WebRTC project authors. All Rights Reserved.
3 *
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.
9 */
10
11 // Modified from the Chromium original:
12 // src/media/base/sinc_resampler.cc
13
14 // Initial input buffer layout, dividing into regions r0_ to r4_ (note: r0_, r3_
15 // and r4_ will move after the first load):
16 //
17 // |----------------|-----------------------------------------|----------------|
18 //
19 // request_frames_
20 // <--------------------------------------------------------->
21 // r0_ (during first load)
22 //
23 // kKernelSize / 2 kKernelSize / 2 kKernelSize / 2 kKernelSize / 2
24 // <---------------> <---------------> <---------------> <--------------->
25 // r1_ r2_ r3_ r4_
26 //
27 // block_size_ == r4_ - r2_
28 // <--------------------------------------->
29 //
30 // request_frames_
31 // <------------------ ... ----------------->
32 // r0_ (during second load)
33 //
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.
36 //
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
39 // provided callback.
40 //
41 // The algorithm:
42 //
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).
46 //
47 // 2) Let r1_, r2_ each represent half the kernel centered around r0_:
48 //
49 // r0_ = input_buffer_ + kKernelSize / 2
50 // r1_ = input_buffer_
51 // r2_ = r0_
52 //
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
55 // step (5) for why).
56 //
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_:
59 //
60 // r3_ = r0_ + request_frames_ - kKernelSize
61 // r4_ = r0_ + request_frames_ - kKernelSize / 2
62 // block_size_ = r4_ - r2_ = request_frames_ - kKernelSize / 2
63 //
64 // 4) Consume request_frames_ frames into r0_.
65 //
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.
69 //
70 // 6) Wrap left over data from the r3_ to r1_ and r4_ to r2_.
71 //
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:
75 //
76 // r0_ = r0_ + kKernelSize / 2 = input_buffer_ + kKernelSize
77 //
78 // r3_, r4_, and block_size_ then need to be reinitialized, so goto (3).
79 //
80 // 8) Else, if we're not on the second load, goto (4).
81 //
82 // Note: we're glossing over how the sub-sample handling works with
83 // |virtual_source_idx_|, etc.
84
85 // MSVC++ requires this to be set before any other includes to get M_PI.
86 #define _USE_MATH_DEFINES
87
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"
92
93 #include <assert.h>
94 #include <math.h>
95 #include <string.h>
96
97 #include <limits>
98
99 namespace webrtc {
100
SincScaleFactor(double io_ratio)101 static double SincScaleFactor(double io_ratio) {
102 // |sinc_scale_factor| is basically the normalized cutoff frequency of the
103 // low-pass filter.
104 double sinc_scale_factor = io_ratio > 1.0 ? 1.0 / io_ratio : 1.0;
105
106 // The sinc function is an idealized brick-wall filter, but since we're
107 // windowing it the transition from pass to stop does not happen right away.
108 // So we should adjust the low pass filter cutoff slightly downward to avoid
109 // some aliasing at the very high-end.
110 // TODO(crogers): this value is empirical and to be more exact should vary
111 // depending on kKernelSize.
112 sinc_scale_factor *= 0.9;
113
114 return sinc_scale_factor;
115 }
116
117 // If we know the minimum architecture at compile time, avoid CPU detection.
118 #if defined(WEBRTC_ARCH_X86_FAMILY)
119 #if defined(__SSE2__)
120 #define CONVOLVE_FUNC Convolve_SSE
InitializeCPUSpecificFeatures()121 void SincResampler::InitializeCPUSpecificFeatures() {}
122 #else
123 // x86 CPU detection required. Function will be set by
124 // InitializeCPUSpecificFeatures().
125 // TODO(dalecurtis): Once Chrome moves to an SSE baseline this can be removed.
126 #define CONVOLVE_FUNC convolve_proc_
127
InitializeCPUSpecificFeatures()128 void SincResampler::InitializeCPUSpecificFeatures() {
129 convolve_proc_ = WebRtc_GetCPUInfo(kSSE2) ? Convolve_SSE : Convolve_C;
130 }
131 #endif
132 #elif defined(WEBRTC_ARCH_ARM_V7)
133 #if defined(WEBRTC_ARCH_ARM_NEON)
134 #define CONVOLVE_FUNC Convolve_NEON
InitializeCPUSpecificFeatures()135 void SincResampler::InitializeCPUSpecificFeatures() {}
136 #else
137 // ARM CPU detection required. Function will be set by
138 // InitializeCPUSpecificFeatures().
139 #define CONVOLVE_FUNC convolve_proc_
140
InitializeCPUSpecificFeatures()141 void SincResampler::InitializeCPUSpecificFeatures() {
142 convolve_proc_ = WebRtc_GetCPUFeaturesARM() & kCPUFeatureNEON ?
143 Convolve_NEON : Convolve_C;
144 }
145 #endif
146 #else
147 // Unknown architecture.
148 #define CONVOLVE_FUNC Convolve_C
InitializeCPUSpecificFeatures()149 void SincResampler::InitializeCPUSpecificFeatures() {}
150 #endif
151
SincResampler(double io_sample_rate_ratio,int request_frames,SincResamplerCallback * read_cb)152 SincResampler::SincResampler(double io_sample_rate_ratio,
153 int request_frames,
154 SincResamplerCallback* read_cb)
155 : io_sample_rate_ratio_(io_sample_rate_ratio),
156 read_cb_(read_cb),
157 request_frames_(request_frames),
158 input_buffer_size_(request_frames_ + kKernelSize),
159 // Create input buffers with a 16-byte alignment for SSE optimizations.
160 kernel_storage_(static_cast<float*>(
161 AlignedMalloc(sizeof(float) * kKernelStorageSize, 16))),
162 kernel_pre_sinc_storage_(static_cast<float*>(
163 AlignedMalloc(sizeof(float) * kKernelStorageSize, 16))),
164 kernel_window_storage_(static_cast<float*>(
165 AlignedMalloc(sizeof(float) * kKernelStorageSize, 16))),
166 input_buffer_(static_cast<float*>(
167 AlignedMalloc(sizeof(float) * input_buffer_size_, 16))),
168 #if defined(WEBRTC_CPU_DETECTION)
169 convolve_proc_(NULL),
170 #endif
171 r1_(input_buffer_.get()),
172 r2_(input_buffer_.get() + kKernelSize / 2) {
173 #if defined(WEBRTC_CPU_DETECTION)
174 InitializeCPUSpecificFeatures();
175 assert(convolve_proc_);
176 #endif
177 assert(request_frames_ > 0);
178 Flush();
179 assert(block_size_ > kKernelSize);
180
181 memset(kernel_storage_.get(), 0,
182 sizeof(*kernel_storage_.get()) * kKernelStorageSize);
183 memset(kernel_pre_sinc_storage_.get(), 0,
184 sizeof(*kernel_pre_sinc_storage_.get()) * kKernelStorageSize);
185 memset(kernel_window_storage_.get(), 0,
186 sizeof(*kernel_window_storage_.get()) * kKernelStorageSize);
187
188 InitializeKernel();
189 }
190
~SincResampler()191 SincResampler::~SincResampler() {}
192
UpdateRegions(bool second_load)193 void SincResampler::UpdateRegions(bool second_load) {
194 // Setup various region pointers in the buffer (see diagram above). If we're
195 // on the second load we need to slide r0_ to the right by kKernelSize / 2.
196 r0_ = input_buffer_.get() + (second_load ? kKernelSize : kKernelSize / 2);
197 r3_ = r0_ + request_frames_ - kKernelSize;
198 r4_ = r0_ + request_frames_ - kKernelSize / 2;
199 block_size_ = r4_ - r2_;
200
201 // r1_ at the beginning of the buffer.
202 assert(r1_ == input_buffer_.get());
203 // r1_ left of r2_, r4_ left of r3_ and size correct.
204 assert(r2_ - r1_ == r4_ - r3_);
205 // r2_ left of r3.
206 assert(r2_ < r3_);
207 }
208
InitializeKernel()209 void SincResampler::InitializeKernel() {
210 // Blackman window parameters.
211 static const double kAlpha = 0.16;
212 static const double kA0 = 0.5 * (1.0 - kAlpha);
213 static const double kA1 = 0.5;
214 static const double kA2 = 0.5 * kAlpha;
215
216 // Generates a set of windowed sinc() kernels.
217 // We generate a range of sub-sample offsets from 0.0 to 1.0.
218 const double sinc_scale_factor = SincScaleFactor(io_sample_rate_ratio_);
219 for (int offset_idx = 0; offset_idx <= kKernelOffsetCount; ++offset_idx) {
220 const float subsample_offset =
221 static_cast<float>(offset_idx) / kKernelOffsetCount;
222
223 for (int i = 0; i < kKernelSize; ++i) {
224 const int idx = i + offset_idx * kKernelSize;
225 const float pre_sinc =
226 static_cast<float>(M_PI * (i - kKernelSize / 2 - subsample_offset));
227 kernel_pre_sinc_storage_[idx] = pre_sinc;
228
229 // Compute Blackman window, matching the offset of the sinc().
230 const float x = (i - subsample_offset) / kKernelSize;
231 const float window = static_cast<float>(kA0 - kA1 * cos(2.0 * M_PI * x) +
232 kA2 * cos(4.0 * M_PI * x));
233 kernel_window_storage_[idx] = window;
234
235 // Compute the sinc with offset, then window the sinc() function and store
236 // at the correct offset.
237 kernel_storage_[idx] = static_cast<float>(window *
238 ((pre_sinc == 0) ?
239 sinc_scale_factor :
240 (sin(sinc_scale_factor * pre_sinc) / pre_sinc)));
241 }
242 }
243 }
244
SetRatio(double io_sample_rate_ratio)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()) {
248 return;
249 }
250
251 io_sample_rate_ratio_ = io_sample_rate_ratio;
252
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];
261
262 kernel_storage_[idx] = static_cast<float>(window *
263 ((pre_sinc == 0) ?
264 sinc_scale_factor :
265 (sin(sinc_scale_factor * pre_sinc) / pre_sinc)));
266 }
267 }
268 }
269
Resample(int frames,float * destination)270 void SincResampler::Resample(int frames, float* destination) {
271 int remaining_frames = frames;
272
273 // Step (1) -- Prime the input buffer at the start of the input stream.
274 if (!buffer_primed_ && remaining_frames) {
275 read_cb_->Run(request_frames_, r0_);
276 buffer_primed_ = true;
277 }
278
279 // Step (2) -- Resample! const what we can outside of the loop for speed. It
280 // actually has an impact on ARM performance. See inner loop comment below.
281 const double current_io_ratio = io_sample_rate_ratio_;
282 const float* const kernel_ptr = kernel_storage_.get();
283 while (remaining_frames) {
284 // |i| may be negative if the last Resample() call ended on an iteration
285 // that put |virtual_source_idx_| over the limit.
286 //
287 // Note: The loop construct here can severely impact performance on ARM
288 // or when built with clang. See https://codereview.chromium.org/18566009/
289 for (int i = static_cast<int>(
290 ceil((block_size_ - virtual_source_idx_) / current_io_ratio));
291 i > 0; --i) {
292 assert(virtual_source_idx_ < block_size_);
293
294 // |virtual_source_idx_| lies in between two kernel offsets so figure out
295 // what they are.
296 const int source_idx = static_cast<int>(virtual_source_idx_);
297 const double subsample_remainder = virtual_source_idx_ - source_idx;
298
299 const double virtual_offset_idx =
300 subsample_remainder * kKernelOffsetCount;
301 const int offset_idx = static_cast<int>(virtual_offset_idx);
302
303 // We'll compute "convolutions" for the two kernels which straddle
304 // |virtual_source_idx_|.
305 const float* const k1 = kernel_ptr + offset_idx * kKernelSize;
306 const float* const k2 = k1 + kKernelSize;
307
308 // Ensure |k1|, |k2| are 16-byte aligned for SIMD usage. Should always be
309 // true so long as kKernelSize is a multiple of 16.
310 assert(0u == (reinterpret_cast<uintptr_t>(k1) & 0x0F));
311 assert(0u == (reinterpret_cast<uintptr_t>(k2) & 0x0F));
312
313 // Initialize input pointer based on quantized |virtual_source_idx_|.
314 const float* const input_ptr = r1_ + source_idx;
315
316 // Figure out how much to weight each kernel's "convolution".
317 const double kernel_interpolation_factor =
318 virtual_offset_idx - offset_idx;
319 *destination++ = CONVOLVE_FUNC(
320 input_ptr, k1, k2, kernel_interpolation_factor);
321
322 // Advance the virtual index.
323 virtual_source_idx_ += current_io_ratio;
324
325 if (!--remaining_frames)
326 return;
327 }
328
329 // Wrap back around to the start.
330 virtual_source_idx_ -= block_size_;
331
332 // Step (3) -- Copy r3_, r4_ to r1_, r2_.
333 // This wraps the last input frames back to the start of the buffer.
334 memcpy(r1_, r3_, sizeof(*input_buffer_.get()) * kKernelSize);
335
336 // Step (4) -- Reinitialize regions if necessary.
337 if (r0_ == r2_)
338 UpdateRegions(true);
339
340 // Step (5) -- Refresh the buffer with more input.
341 read_cb_->Run(request_frames_, r0_);
342 }
343 }
344
345 #undef CONVOLVE_FUNC
346
ChunkSize() const347 int SincResampler::ChunkSize() const {
348 return static_cast<int>(block_size_ / io_sample_rate_ratio_);
349 }
350
Flush()351 void SincResampler::Flush() {
352 virtual_source_idx_ = 0;
353 buffer_primed_ = false;
354 memset(input_buffer_.get(), 0,
355 sizeof(*input_buffer_.get()) * input_buffer_size_);
356 UpdateRegions(false);
357 }
358
Convolve_C(const float * input_ptr,const float * k1,const float * k2,double kernel_interpolation_factor)359 float SincResampler::Convolve_C(const float* input_ptr, const float* k1,
360 const float* k2,
361 double kernel_interpolation_factor) {
362 float sum1 = 0;
363 float sum2 = 0;
364
365 // Generate a single output sample. Unrolling this loop hurt performance in
366 // local testing.
367 int n = kKernelSize;
368 while (n--) {
369 sum1 += *input_ptr * *k1++;
370 sum2 += *input_ptr++ * *k2++;
371 }
372
373 // Linearly interpolate the two "convolutions".
374 return static_cast<float>((1.0 - kernel_interpolation_factor) * sum1 +
375 kernel_interpolation_factor * sum2);
376 }
377
378 } // namespace webrtc
379