sinc_resampler.cc revision 073dd7b423157020846872ca2398cc86d3ca9508
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/cpu_features_wrapper.h" 90#include "webrtc/typedefs.h" 91 92#include <assert.h> 93#include <math.h> 94#include <string.h> 95 96#include <limits> 97 98namespace webrtc { 99 100static double SincScaleFactor(double io_ratio) { 101 // |sinc_scale_factor| is basically the normalized cutoff frequency of the 102 // low-pass filter. 103 double sinc_scale_factor = io_ratio > 1.0 ? 1.0 / io_ratio : 1.0; 104 105 // 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; 112 113 return sinc_scale_factor; 114} 115 116// 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 120void SincResampler::InitializeCPUSpecificFeatures() {} 121#else 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_ 126 127void SincResampler::InitializeCPUSpecificFeatures() { 128 convolve_proc_ = WebRtc_GetCPUInfo(kSSE2) ? Convolve_SSE : Convolve_C; 129} 130#endif 131#elif defined(WEBRTC_DETECT_ARM_NEON) || defined(WEBRTC_ARCH_ARM_NEON) 132#if defined(WEBRTC_ARCH_ARM_NEON) 133#define CONVOLVE_FUNC Convolve_NEON 134void SincResampler::InitializeCPUSpecificFeatures() {} 135#else 136// ARM CPU detection required. Function will be set by 137// InitializeCPUSpecificFeatures(). 138#define CONVOLVE_FUNC convolve_proc_ 139 140void SincResampler::InitializeCPUSpecificFeatures() { 141 convolve_proc_ = WebRtc_GetCPUFeaturesARM() & kCPUFeatureNEON ? 142 Convolve_NEON : Convolve_C; 143} 144#endif 145#else 146// Unknown architecture. 147#define CONVOLVE_FUNC Convolve_C 148void SincResampler::InitializeCPUSpecificFeatures() {} 149#endif 150 151SincResampler::SincResampler(double io_sample_rate_ratio, 152 int request_frames, 153 SincResamplerCallback* read_cb) 154 : io_sample_rate_ratio_(io_sample_rate_ratio), 155 read_cb_(read_cb), 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), 169#endif 170 r1_(input_buffer_.get()), 171 r2_(input_buffer_.get() + kKernelSize / 2) { 172#if defined(WEBRTC_CPU_DETECTION) 173 InitializeCPUSpecificFeatures(); 174 assert(convolve_proc_); 175#endif 176 assert(request_frames_ > 0); 177 Flush(); 178 assert(block_size_ > kKernelSize); 179 180 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); 186 187 InitializeKernel(); 188} 189 190SincResampler::~SincResampler() {} 191 192void 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_; 199 200 // 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_); 204 // r2_ left of r3. 205 assert(r2_ < r3_); 206} 207 208void 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; 214 215 // 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; 221 222 for (int i = 0; i < kKernelSize; ++i) { 223 const int idx = i + offset_idx * kKernelSize; 224 const float pre_sinc = 225 static_cast<float>(M_PI * (i - kKernelSize / 2 - subsample_offset)); 226 kernel_pre_sinc_storage_[idx] = pre_sinc; 227 228 // Compute Blackman window, matching the offset of the sinc(). 229 const float x = (i - subsample_offset) / kKernelSize; 230 const float window = static_cast<float>(kA0 - kA1 * cos(2.0 * M_PI * x) + 231 kA2 * cos(4.0 * M_PI * x)); 232 kernel_window_storage_[idx] = window; 233 234 // Compute the sinc with offset, then window the sinc() function and store 235 // at the correct offset. 236 kernel_storage_[idx] = static_cast<float>(window * 237 ((pre_sinc == 0) ? 238 sinc_scale_factor : 239 (sin(sinc_scale_factor * pre_sinc) / pre_sinc))); 240 } 241 } 242} 243 244void SincResampler::SetRatio(double io_sample_rate_ratio) { 245 if (fabs(io_sample_rate_ratio_ - io_sample_rate_ratio) < 246 std::numeric_limits<double>::epsilon()) { 247 return; 248 } 249 250 io_sample_rate_ratio_ = io_sample_rate_ratio; 251 252 // Optimize reinitialization by reusing values which are independent of 253 // |sinc_scale_factor|. Provides a 3x speedup. 254 const double sinc_scale_factor = SincScaleFactor(io_sample_rate_ratio_); 255 for (int offset_idx = 0; offset_idx <= kKernelOffsetCount; ++offset_idx) { 256 for (int i = 0; i < kKernelSize; ++i) { 257 const int idx = i + offset_idx * kKernelSize; 258 const float window = kernel_window_storage_[idx]; 259 const float pre_sinc = kernel_pre_sinc_storage_[idx]; 260 261 kernel_storage_[idx] = static_cast<float>(window * 262 ((pre_sinc == 0) ? 263 sinc_scale_factor : 264 (sin(sinc_scale_factor * pre_sinc) / pre_sinc))); 265 } 266 } 267} 268 269void SincResampler::Resample(int frames, float* destination) { 270 int remaining_frames = frames; 271 272 // Step (1) -- Prime the input buffer at the start of the input stream. 273 if (!buffer_primed_ && remaining_frames) { 274 read_cb_->Run(request_frames_, r0_); 275 buffer_primed_ = true; 276 } 277 278 // Step (2) -- Resample! const what we can outside of the loop for speed. It 279 // actually has an impact on ARM performance. See inner loop comment below. 280 const double current_io_ratio = io_sample_rate_ratio_; 281 const float* const kernel_ptr = kernel_storage_.get(); 282 while (remaining_frames) { 283 // |i| may be negative if the last Resample() call ended on an iteration 284 // that put |virtual_source_idx_| over the limit. 285 // 286 // Note: The loop construct here can severely impact performance on ARM 287 // or when built with clang. See https://codereview.chromium.org/18566009/ 288 for (int i = static_cast<int>( 289 ceil((block_size_ - virtual_source_idx_) / current_io_ratio)); 290 i > 0; --i) { 291 assert(virtual_source_idx_ < block_size_); 292 293 // |virtual_source_idx_| lies in between two kernel offsets so figure out 294 // what they are. 295 const int source_idx = static_cast<int>(virtual_source_idx_); 296 const double subsample_remainder = virtual_source_idx_ - source_idx; 297 298 const double virtual_offset_idx = 299 subsample_remainder * kKernelOffsetCount; 300 const int offset_idx = static_cast<int>(virtual_offset_idx); 301 302 // We'll compute "convolutions" for the two kernels which straddle 303 // |virtual_source_idx_|. 304 const float* const k1 = kernel_ptr + offset_idx * kKernelSize; 305 const float* const k2 = k1 + kKernelSize; 306 307 // Ensure |k1|, |k2| are 16-byte aligned for SIMD usage. Should always be 308 // true so long as kKernelSize is a multiple of 16. 309 assert(0u == (reinterpret_cast<uintptr_t>(k1) & 0x0F)); 310 assert(0u == (reinterpret_cast<uintptr_t>(k2) & 0x0F)); 311 312 // Initialize input pointer based on quantized |virtual_source_idx_|. 313 const float* const input_ptr = r1_ + source_idx; 314 315 // Figure out how much to weight each kernel's "convolution". 316 const double kernel_interpolation_factor = 317 virtual_offset_idx - offset_idx; 318 *destination++ = CONVOLVE_FUNC( 319 input_ptr, k1, k2, kernel_interpolation_factor); 320 321 // Advance the virtual index. 322 virtual_source_idx_ += current_io_ratio; 323 324 if (!--remaining_frames) 325 return; 326 } 327 328 // Wrap back around to the start. 329 virtual_source_idx_ -= block_size_; 330 331 // Step (3) -- Copy r3_, r4_ to r1_, r2_. 332 // This wraps the last input frames back to the start of the buffer. 333 memcpy(r1_, r3_, sizeof(*input_buffer_.get()) * kKernelSize); 334 335 // Step (4) -- Reinitialize regions if necessary. 336 if (r0_ == r2_) 337 UpdateRegions(true); 338 339 // Step (5) -- Refresh the buffer with more input. 340 read_cb_->Run(request_frames_, r0_); 341 } 342} 343 344#undef CONVOLVE_FUNC 345 346int SincResampler::ChunkSize() const { 347 return static_cast<int>(block_size_ / io_sample_rate_ratio_); 348} 349 350void SincResampler::Flush() { 351 virtual_source_idx_ = 0; 352 buffer_primed_ = false; 353 memset(input_buffer_.get(), 0, 354 sizeof(*input_buffer_.get()) * input_buffer_size_); 355 UpdateRegions(false); 356} 357 358float SincResampler::Convolve_C(const float* input_ptr, const float* k1, 359 const float* k2, 360 double kernel_interpolation_factor) { 361 float sum1 = 0; 362 float sum2 = 0; 363 364 // Generate a single output sample. Unrolling this loop hurt performance in 365 // local testing. 366 int n = kKernelSize; 367 while (n--) { 368 sum1 += *input_ptr * *k1++; 369 sum2 += *input_ptr++ * *k2++; 370 } 371 372 // Linearly interpolate the two "convolutions". 373 return static_cast<float>((1.0 - kernel_interpolation_factor) * sum1 + 374 kernel_interpolation_factor * sum2); 375} 376 377} // namespace webrtc 378