1 /* 2 * Copyright (C) 2010 Google Inc. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. Neither the name of Apple Computer, Inc. ("Apple") nor the names of 14 * its contributors may be used to endorse or promote products derived 15 * from this software without specific prior written permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY APPLE AND ITS CONTRIBUTORS "AS IS" AND ANY 18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 19 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 20 * DISCLAIMED. IN NO EVENT SHALL APPLE OR ITS CONTRIBUTORS BE LIABLE FOR ANY 21 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 22 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 23 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND 24 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 */ 28 29 #include "config.h" 30 31 #if ENABLE(WEB_AUDIO) 32 33 #include "ReverbConvolver.h" 34 35 #include "VectorMath.h" 36 #include "AudioBus.h" 37 38 namespace WebCore { 39 40 using namespace VectorMath; 41 42 const int InputBufferSize = 8 * 16384; 43 44 // We only process the leading portion of the impulse response in the real-time thread. We don't exceed this length. 45 // It turns out then, that the background thread has about 278msec of scheduling slop. 46 // Empirically, this has been found to be a good compromise between giving enough time for scheduling slop, 47 // while still minimizing the amount of processing done in the primary (high-priority) thread. 48 // This was found to be a good value on Mac OS X, and may work well on other platforms as well, assuming 49 // the very rough scheduling latencies are similar on these time-scales. Of course, this code may need to be 50 // tuned for individual platforms if this assumption is found to be incorrect. 51 const size_t RealtimeFrameLimit = 8192 + 4096; // ~278msec @ 44.1KHz 52 53 const size_t MinFFTSize = 256; 54 const size_t MaxRealtimeFFTSize = 2048; 55 56 static void* backgroundThreadEntry(void* threadData) 57 { 58 ReverbConvolver* reverbConvolver = static_cast<ReverbConvolver*>(threadData); 59 reverbConvolver->backgroundThreadEntry(); 60 return 0; 61 } 62 63 ReverbConvolver::ReverbConvolver(AudioChannel* impulseResponse, size_t renderSliceSize, size_t maxFFTSize, size_t convolverRenderPhase, bool useBackgroundThreads) 64 : m_impulseResponseLength(impulseResponse->length()) 65 , m_accumulationBuffer(impulseResponse->length() + renderSliceSize) 66 , m_inputBuffer(InputBufferSize) 67 , m_renderSliceSize(renderSliceSize) 68 , m_minFFTSize(MinFFTSize) // First stage will have this size - successive stages will double in size each time 69 , m_maxFFTSize(maxFFTSize) // until we hit m_maxFFTSize 70 , m_useBackgroundThreads(useBackgroundThreads) 71 , m_backgroundThread(0) 72 , m_wantsToExit(false) 73 , m_moreInputBuffered(false) 74 { 75 // If we are using background threads then don't exceed this FFT size for the 76 // stages which run in the real-time thread. This avoids having only one or two 77 // large stages (size 16384 or so) at the end which take a lot of time every several 78 // processing slices. This way we amortize the cost over more processing slices. 79 m_maxRealtimeFFTSize = MaxRealtimeFFTSize; 80 81 // For the moment, a good way to know if we have real-time constraint is to check if we're using background threads. 82 // Otherwise, assume we're being run from a command-line tool. 83 bool hasRealtimeConstraint = useBackgroundThreads; 84 85 float* response = impulseResponse->data(); 86 size_t totalResponseLength = impulseResponse->length(); 87 88 // Because we're not using direct-convolution in the leading portion, the reverb has an overall latency of half the first-stage FFT size 89 size_t reverbTotalLatency = m_minFFTSize / 2; 90 91 size_t stageOffset = 0; 92 int i = 0; 93 size_t fftSize = m_minFFTSize; 94 while (stageOffset < totalResponseLength) { 95 size_t stageSize = fftSize / 2; 96 97 // For the last stage, it's possible that stageOffset is such that we're straddling the end 98 // of the impulse response buffer (if we use stageSize), so reduce the last stage's length... 99 if (stageSize + stageOffset > totalResponseLength) 100 stageSize = totalResponseLength - stageOffset; 101 102 // This "staggers" the time when each FFT happens so they don't all happen at the same time 103 int renderPhase = convolverRenderPhase + i * renderSliceSize; 104 105 OwnPtr<ReverbConvolverStage> stage(new ReverbConvolverStage(response, totalResponseLength, reverbTotalLatency, stageOffset, stageSize, fftSize, renderPhase, renderSliceSize, &m_accumulationBuffer)); 106 107 bool isBackgroundStage = false; 108 109 if (this->useBackgroundThreads() && stageOffset > RealtimeFrameLimit) { 110 m_backgroundStages.append(stage.release()); 111 isBackgroundStage = true; 112 } else 113 m_stages.append(stage.release()); 114 115 stageOffset += stageSize; 116 ++i; 117 118 // Figure out next FFT size 119 fftSize *= 2; 120 if (hasRealtimeConstraint && !isBackgroundStage && fftSize > m_maxRealtimeFFTSize) 121 fftSize = m_maxRealtimeFFTSize; 122 if (fftSize > m_maxFFTSize) 123 fftSize = m_maxFFTSize; 124 } 125 126 // Start up background thread 127 // FIXME: would be better to up the thread priority here. It doesn't need to be real-time, but higher than the default... 128 if (this->useBackgroundThreads() && m_backgroundStages.size() > 0) 129 m_backgroundThread = createThread(WebCore::backgroundThreadEntry, this, "convolution background thread"); 130 } 131 132 ReverbConvolver::~ReverbConvolver() 133 { 134 // Wait for background thread to stop 135 if (useBackgroundThreads() && m_backgroundThread) { 136 m_wantsToExit = true; 137 138 // Wake up thread so it can return 139 { 140 MutexLocker locker(m_backgroundThreadLock); 141 m_moreInputBuffered = true; 142 m_backgroundThreadCondition.signal(); 143 } 144 145 waitForThreadCompletion(m_backgroundThread, 0); 146 } 147 } 148 149 void ReverbConvolver::backgroundThreadEntry() 150 { 151 while (!m_wantsToExit) { 152 // Wait for realtime thread to give us more input 153 m_moreInputBuffered = false; 154 { 155 MutexLocker locker(m_backgroundThreadLock); 156 while (!m_moreInputBuffered && !m_wantsToExit) 157 m_backgroundThreadCondition.wait(m_backgroundThreadLock); 158 } 159 160 // Process all of the stages until their read indices reach the input buffer's write index 161 int writeIndex = m_inputBuffer.writeIndex(); 162 163 // Even though it doesn't seem like every stage needs to maintain its own version of readIndex 164 // we do this in case we want to run in more than one background thread. 165 int readIndex; 166 167 while ((readIndex = m_backgroundStages[0]->inputReadIndex()) != writeIndex) { // FIXME: do better to detect buffer overrun... 168 // The ReverbConvolverStages need to process in amounts which evenly divide half the FFT size 169 const int SliceSize = MinFFTSize / 2; 170 171 // Accumulate contributions from each stage 172 for (size_t i = 0; i < m_backgroundStages.size(); ++i) 173 m_backgroundStages[i]->processInBackground(this, SliceSize); 174 } 175 } 176 } 177 178 void ReverbConvolver::process(AudioChannel* sourceChannel, AudioChannel* destinationChannel, size_t framesToProcess) 179 { 180 bool isSafe = sourceChannel && destinationChannel && sourceChannel->length() >= framesToProcess && destinationChannel->length() >= framesToProcess; 181 ASSERT(isSafe); 182 if (!isSafe) 183 return; 184 185 float* source = sourceChannel->data(); 186 float* destination = destinationChannel->data(); 187 bool isDataSafe = source && destination; 188 ASSERT(isDataSafe); 189 if (!isDataSafe) 190 return; 191 192 // Feed input buffer (read by all threads) 193 m_inputBuffer.write(source, framesToProcess); 194 195 // Accumulate contributions from each stage 196 for (size_t i = 0; i < m_stages.size(); ++i) 197 m_stages[i]->process(source, framesToProcess); 198 199 // Finally read from accumulation buffer 200 m_accumulationBuffer.readAndClear(destination, framesToProcess); 201 202 // Now that we've buffered more input, wake up our background thread. 203 204 // Not using a MutexLocker looks strange, but we use a tryLock() instead because this is run on the real-time 205 // thread where it is a disaster for the lock to be contended (causes audio glitching). It's OK if we fail to 206 // signal from time to time, since we'll get to it the next time we're called. We're called repeatedly 207 // and frequently (around every 3ms). The background thread is processing well into the future and has a considerable amount of 208 // leeway here... 209 if (m_backgroundThreadLock.tryLock()) { 210 m_moreInputBuffered = true; 211 m_backgroundThreadCondition.signal(); 212 m_backgroundThreadLock.unlock(); 213 } 214 } 215 216 void ReverbConvolver::reset() 217 { 218 for (size_t i = 0; i < m_stages.size(); ++i) 219 m_stages[i]->reset(); 220 221 for (size_t i = 0; i < m_backgroundStages.size(); ++i) 222 m_backgroundStages[i]->reset(); 223 224 m_accumulationBuffer.reset(); 225 m_inputBuffer.reset(); 226 } 227 228 } // namespace WebCore 229 230 #endif // ENABLE(WEB_AUDIO) 231
