sinc_resampler.cc (13992B)
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 #include "common_audio/resampler/sinc_resampler.h" 86 87 #include <cmath> 88 #include <cstdint> 89 #include <cstring> 90 #include <limits> 91 #include <numbers> 92 93 #include "rtc_base/checks.h" 94 #include "rtc_base/cpu_info.h" 95 #include "rtc_base/memory/aligned_malloc.h" 96 #include "rtc_base/system/arch.h" 97 98 namespace webrtc { 99 100 namespace { 101 102 double SincScaleFactor(double io_ratio) { 103 // `sinc_scale_factor` is basically the normalized cutoff frequency of the 104 // low-pass filter. 105 double sinc_scale_factor = io_ratio > 1.0 ? 1.0 / io_ratio : 1.0; 106 107 // The sinc function is an idealized brick-wall filter, but since we're 108 // windowing it the transition from pass to stop does not happen right away. 109 // So we should adjust the low pass filter cutoff slightly downward to avoid 110 // some aliasing at the very high-end. 111 // TODO(crogers): this value is empirical and to be more exact should vary 112 // depending on kKernelSize. 113 sinc_scale_factor *= 0.9; 114 115 return sinc_scale_factor; 116 } 117 118 } // namespace 119 120 const size_t SincResampler::kKernelSize; 121 122 // If we know the minimum architecture at compile time, avoid CPU detection. 123 void SincResampler::InitializeCPUSpecificFeatures() { 124 #if defined(WEBRTC_HAS_NEON) 125 convolve_proc_ = Convolve_NEON; 126 #elif defined(WEBRTC_ARCH_X86_FAMILY) 127 // Using AVX2 instead of SSE2 when AVX2/FMA3 supported. 128 if (cpu_info::Supports(cpu_info::ISA::kAVX2) && 129 cpu_info::Supports(cpu_info::ISA::kFMA3)) 130 convolve_proc_ = Convolve_AVX2; 131 else if (cpu_info::Supports(cpu_info::ISA::kSSE2)) 132 convolve_proc_ = Convolve_SSE; 133 else 134 convolve_proc_ = Convolve_C; 135 #else 136 // Unknown architecture. 137 convolve_proc_ = Convolve_C; 138 #endif 139 } 140 141 SincResampler::SincResampler(double io_sample_rate_ratio, 142 size_t request_frames, 143 SincResamplerCallback* read_cb) 144 : io_sample_rate_ratio_(io_sample_rate_ratio), 145 read_cb_(read_cb), 146 request_frames_(request_frames), 147 input_buffer_size_(request_frames_ + kKernelSize), 148 // Create input buffers with a 32-byte alignment for SIMD optimizations. 149 kernel_storage_(static_cast<float*>( 150 AlignedMalloc(sizeof(float) * kKernelStorageSize, 32))), 151 kernel_pre_sinc_storage_(static_cast<float*>( 152 AlignedMalloc(sizeof(float) * kKernelStorageSize, 32))), 153 kernel_window_storage_(static_cast<float*>( 154 AlignedMalloc(sizeof(float) * kKernelStorageSize, 32))), 155 input_buffer_(static_cast<float*>( 156 AlignedMalloc(sizeof(float) * input_buffer_size_, 32))), 157 convolve_proc_(nullptr), 158 r1_(input_buffer_.get()), 159 r2_(input_buffer_.get() + kKernelSize / 2) { 160 InitializeCPUSpecificFeatures(); 161 RTC_DCHECK(convolve_proc_); 162 RTC_DCHECK_GT(request_frames_, 0); 163 Flush(); 164 RTC_DCHECK_GT(block_size_, kKernelSize); 165 166 memset(kernel_storage_.get(), 0, 167 sizeof(*kernel_storage_.get()) * kKernelStorageSize); 168 memset(kernel_pre_sinc_storage_.get(), 0, 169 sizeof(*kernel_pre_sinc_storage_.get()) * kKernelStorageSize); 170 memset(kernel_window_storage_.get(), 0, 171 sizeof(*kernel_window_storage_.get()) * kKernelStorageSize); 172 173 InitializeKernel(); 174 } 175 176 SincResampler::~SincResampler() {} 177 178 void SincResampler::UpdateRegions(bool second_load) { 179 // Setup various region pointers in the buffer (see diagram above). If we're 180 // on the second load we need to slide r0_ to the right by kKernelSize / 2. 181 r0_ = input_buffer_.get() + (second_load ? kKernelSize : kKernelSize / 2); 182 r3_ = r0_ + request_frames_ - kKernelSize; 183 r4_ = r0_ + request_frames_ - kKernelSize / 2; 184 block_size_ = r4_ - r2_; 185 186 // r1_ at the beginning of the buffer. 187 RTC_DCHECK_EQ(r1_, input_buffer_.get()); 188 // r1_ left of r2_, r4_ left of r3_ and size correct. 189 RTC_DCHECK_EQ(r2_ - r1_, r4_ - r3_); 190 // r2_ left of r3. 191 RTC_DCHECK_LT(r2_, r3_); 192 } 193 194 void SincResampler::InitializeKernel() { 195 // Blackman window parameters. 196 static const double kAlpha = 0.16; 197 static const double kA0 = 0.5 * (1.0 - kAlpha); 198 static const double kA1 = 0.5; 199 static const double kA2 = 0.5 * kAlpha; 200 201 // Generates a set of windowed sinc() kernels. 202 // We generate a range of sub-sample offsets from 0.0 to 1.0. 203 const double sinc_scale_factor = SincScaleFactor(io_sample_rate_ratio_); 204 for (size_t offset_idx = 0; offset_idx <= kKernelOffsetCount; ++offset_idx) { 205 const float subsample_offset = 206 static_cast<float>(offset_idx) / kKernelOffsetCount; 207 208 for (size_t i = 0; i < kKernelSize; ++i) { 209 const size_t idx = i + offset_idx * kKernelSize; 210 const float pre_sinc = static_cast<float>( 211 std::numbers::pi * 212 (static_cast<int>(i) - static_cast<int>(kKernelSize / 2) - 213 subsample_offset)); 214 kernel_pre_sinc_storage_[idx] = pre_sinc; 215 216 // Compute Blackman window, matching the offset of the sinc(). 217 const float x = (i - subsample_offset) / kKernelSize; 218 const float window = 219 static_cast<float>(kA0 - kA1 * cos(2.0 * std::numbers::pi * x) + 220 kA2 * cos(4.0 * std::numbers::pi * x)); 221 kernel_window_storage_[idx] = window; 222 223 // Compute the sinc with offset, then window the sinc() function and store 224 // at the correct offset. 225 kernel_storage_[idx] = static_cast<float>( 226 window * ((pre_sinc == 0) 227 ? sinc_scale_factor 228 : (sin(sinc_scale_factor * pre_sinc) / pre_sinc))); 229 } 230 } 231 } 232 233 void SincResampler::SetRatio(double io_sample_rate_ratio) { 234 if (fabs(io_sample_rate_ratio_ - io_sample_rate_ratio) < 235 std::numeric_limits<double>::epsilon()) { 236 return; 237 } 238 239 io_sample_rate_ratio_ = io_sample_rate_ratio; 240 241 // Optimize reinitialization by reusing values which are independent of 242 // `sinc_scale_factor`. Provides a 3x speedup. 243 const double sinc_scale_factor = SincScaleFactor(io_sample_rate_ratio_); 244 for (size_t offset_idx = 0; offset_idx <= kKernelOffsetCount; ++offset_idx) { 245 for (size_t i = 0; i < kKernelSize; ++i) { 246 const size_t idx = i + offset_idx * kKernelSize; 247 const float window = kernel_window_storage_[idx]; 248 const float pre_sinc = kernel_pre_sinc_storage_[idx]; 249 250 kernel_storage_[idx] = static_cast<float>( 251 window * ((pre_sinc == 0) 252 ? sinc_scale_factor 253 : (sin(sinc_scale_factor * pre_sinc) / pre_sinc))); 254 } 255 } 256 } 257 258 void SincResampler::Resample(size_t frames, float* destination) { 259 size_t remaining_frames = frames; 260 261 // Step (1) -- Prime the input buffer at the start of the input stream. 262 if (!buffer_primed_ && remaining_frames) { 263 read_cb_->Run(request_frames_, r0_); 264 buffer_primed_ = true; 265 } 266 267 // Step (2) -- Resample! const what we can outside of the loop for speed. It 268 // actually has an impact on ARM performance. See inner loop comment below. 269 const double current_io_ratio = io_sample_rate_ratio_; 270 const float* const kernel_ptr = kernel_storage_.get(); 271 while (remaining_frames) { 272 // `i` may be negative if the last Resample() call ended on an iteration 273 // that put `virtual_source_idx_` over the limit. 274 // 275 // Note: The loop construct here can severely impact performance on ARM 276 // or when built with clang. See https://codereview.chromium.org/18566009/ 277 for (int i = static_cast<int>( 278 ceil((block_size_ - virtual_source_idx_) / current_io_ratio)); 279 i > 0; --i) { 280 RTC_DCHECK_LT(virtual_source_idx_, block_size_); 281 282 // `virtual_source_idx_` lies in between two kernel offsets so figure out 283 // what they are. 284 const int source_idx = static_cast<int>(virtual_source_idx_); 285 const double subsample_remainder = virtual_source_idx_ - source_idx; 286 287 const double virtual_offset_idx = 288 subsample_remainder * kKernelOffsetCount; 289 const int offset_idx = static_cast<int>(virtual_offset_idx); 290 291 // We'll compute "convolutions" for the two kernels which straddle 292 // `virtual_source_idx_`. 293 const float* const k1 = kernel_ptr + offset_idx * kKernelSize; 294 const float* const k2 = k1 + kKernelSize; 295 296 // Ensure `k1`, `k2` are 32-byte aligned for SIMD usage. Should always be 297 // true so long as kKernelSize is a multiple of 32. 298 RTC_DCHECK_EQ(0, reinterpret_cast<uintptr_t>(k1) % 32); 299 RTC_DCHECK_EQ(0, reinterpret_cast<uintptr_t>(k2) % 32); 300 301 // Initialize input pointer based on quantized `virtual_source_idx_`. 302 const float* const input_ptr = r1_ + source_idx; 303 304 // Figure out how much to weight each kernel's "convolution". 305 const double kernel_interpolation_factor = 306 virtual_offset_idx - offset_idx; 307 *destination++ = 308 convolve_proc_(input_ptr, k1, k2, kernel_interpolation_factor); 309 310 // Advance the virtual index. 311 virtual_source_idx_ += current_io_ratio; 312 313 if (!--remaining_frames) 314 return; 315 } 316 317 // Wrap back around to the start. 318 virtual_source_idx_ -= block_size_; 319 320 // Step (3) -- Copy r3_, r4_ to r1_, r2_. 321 // This wraps the last input frames back to the start of the buffer. 322 memcpy(r1_, r3_, sizeof(*input_buffer_.get()) * kKernelSize); 323 324 // Step (4) -- Reinitialize regions if necessary. 325 if (r0_ == r2_) 326 UpdateRegions(true); 327 328 // Step (5) -- Refresh the buffer with more input. 329 read_cb_->Run(request_frames_, r0_); 330 } 331 } 332 333 #undef CONVOLVE_FUNC 334 335 size_t SincResampler::ChunkSize() const { 336 return static_cast<size_t>(block_size_ / io_sample_rate_ratio_); 337 } 338 339 void SincResampler::Flush() { 340 virtual_source_idx_ = 0; 341 buffer_primed_ = false; 342 memset(input_buffer_.get(), 0, 343 sizeof(*input_buffer_.get()) * input_buffer_size_); 344 UpdateRegions(false); 345 } 346 347 float SincResampler::Convolve_C(const float* input_ptr, 348 const float* k1, 349 const float* k2, 350 double kernel_interpolation_factor) { 351 float sum1 = 0; 352 float sum2 = 0; 353 354 // Generate a single output sample. Unrolling this loop hurt performance in 355 // local testing. 356 size_t n = kKernelSize; 357 while (n--) { 358 sum1 += *input_ptr * *k1++; 359 sum2 += *input_ptr++ * *k2++; 360 } 361 362 // Linearly interpolate the two "convolutions". 363 return static_cast<float>((1.0 - kernel_interpolation_factor) * sum1 + 364 kernel_interpolation_factor * sum2); 365 } 366 367 } // namespace webrtc