michael@0: /*
michael@0:  * Copyright (C) 2010 Google Inc. All rights reserved.
michael@0:  *
michael@0:  * Redistribution and use in source and binary forms, with or without
michael@0:  * modification, are permitted provided that the following conditions
michael@0:  * are met:
michael@0:  *
michael@0:  * 1.  Redistributions of source code must retain the above copyright
michael@0:  *     notice, this list of conditions and the following disclaimer.
michael@0:  * 2.  Redistributions in binary form must reproduce the above copyright
michael@0:  *     notice, this list of conditions and the following disclaimer in the
michael@0:  *     documentation and/or other materials provided with the distribution.
michael@0:  * 3.  Neither the name of Apple Computer, Inc. ("Apple") nor the names of
michael@0:  *     its contributors may be used to endorse or promote products derived
michael@0:  *     from this software without specific prior written permission.
michael@0:  *
michael@0:  * THIS SOFTWARE IS PROVIDED BY APPLE AND ITS CONTRIBUTORS "AS IS" AND ANY
michael@0:  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
michael@0:  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
michael@0:  * DISCLAIMED. IN NO EVENT SHALL APPLE OR ITS CONTRIBUTORS BE LIABLE FOR ANY
michael@0:  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
michael@0:  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
michael@0:  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
michael@0:  * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
michael@0:  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
michael@0:  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
michael@0:  */
michael@0: 
michael@0: #include "ReverbConvolver.h"
michael@0: #include "ReverbConvolverStage.h"
michael@0: 
michael@0: using namespace mozilla;
michael@0: 
michael@0: template<>
michael@0: struct RunnableMethodTraits<WebCore::ReverbConvolver>
michael@0: {
michael@0:   static void RetainCallee(WebCore::ReverbConvolver* obj) {}
michael@0:   static void ReleaseCallee(WebCore::ReverbConvolver* obj) {}
michael@0: };
michael@0: 
michael@0: namespace WebCore {
michael@0: 
michael@0: const int InputBufferSize = 8 * 16384;
michael@0: 
michael@0: // We only process the leading portion of the impulse response in the real-time thread.  We don't exceed this length.
michael@0: // It turns out then, that the background thread has about 278msec of scheduling slop.
michael@0: // Empirically, this has been found to be a good compromise between giving enough time for scheduling slop,
michael@0: // while still minimizing the amount of processing done in the primary (high-priority) thread.
michael@0: // This was found to be a good value on Mac OS X, and may work well on other platforms as well, assuming
michael@0: // the very rough scheduling latencies are similar on these time-scales.  Of course, this code may need to be
michael@0: // tuned for individual platforms if this assumption is found to be incorrect.
michael@0: const size_t RealtimeFrameLimit = 8192  + 4096; // ~278msec @ 44.1KHz
michael@0: 
michael@0: const size_t MinFFTSize = 128;
michael@0: const size_t MaxRealtimeFFTSize = 2048;
michael@0: 
michael@0: ReverbConvolver::ReverbConvolver(const float* impulseResponseData, size_t impulseResponseLength, size_t renderSliceSize, size_t maxFFTSize, size_t convolverRenderPhase, bool useBackgroundThreads)
michael@0:     : m_impulseResponseLength(impulseResponseLength)
michael@0:     , m_accumulationBuffer(impulseResponseLength + renderSliceSize)
michael@0:     , m_inputBuffer(InputBufferSize)
michael@0:     , m_minFFTSize(MinFFTSize) // First stage will have this size - successive stages will double in size each time
michael@0:     , m_maxFFTSize(maxFFTSize) // until we hit m_maxFFTSize
michael@0:     , m_backgroundThread("ConvolverWorker")
michael@0:     , m_backgroundThreadCondition(&m_backgroundThreadLock)
michael@0:     , m_useBackgroundThreads(useBackgroundThreads)
michael@0:     , m_wantsToExit(false)
michael@0:     , m_moreInputBuffered(false)
michael@0: {
michael@0:     // If we are using background threads then don't exceed this FFT size for the
michael@0:     // stages which run in the real-time thread.  This avoids having only one or two
michael@0:     // large stages (size 16384 or so) at the end which take a lot of time every several
michael@0:     // processing slices.  This way we amortize the cost over more processing slices.
michael@0:     m_maxRealtimeFFTSize = MaxRealtimeFFTSize;
michael@0: 
michael@0:     // For the moment, a good way to know if we have real-time constraint is to check if we're using background threads.
michael@0:     // Otherwise, assume we're being run from a command-line tool.
michael@0:     bool hasRealtimeConstraint = useBackgroundThreads;
michael@0: 
michael@0:     const float* response = impulseResponseData;
michael@0:     size_t totalResponseLength = impulseResponseLength;
michael@0: 
michael@0:     // The total latency is zero because the direct-convolution is used in the leading portion.
michael@0:     size_t reverbTotalLatency = 0;
michael@0: 
michael@0:     size_t stageOffset = 0;
michael@0:     int i = 0;
michael@0:     size_t fftSize = m_minFFTSize;
michael@0:     while (stageOffset < totalResponseLength) {
michael@0:         size_t stageSize = fftSize / 2;
michael@0: 
michael@0:         // For the last stage, it's possible that stageOffset is such that we're straddling the end
michael@0:         // of the impulse response buffer (if we use stageSize), so reduce the last stage's length...
michael@0:         if (stageSize + stageOffset > totalResponseLength)
michael@0:             stageSize = totalResponseLength - stageOffset;
michael@0: 
michael@0:         // This "staggers" the time when each FFT happens so they don't all happen at the same time
michael@0:         int renderPhase = convolverRenderPhase + i * renderSliceSize;
michael@0: 
michael@0:         bool useDirectConvolver = !stageOffset;
michael@0: 
michael@0:         nsAutoPtr<ReverbConvolverStage> stage(new ReverbConvolverStage(response, totalResponseLength, reverbTotalLatency, stageOffset, stageSize, fftSize, renderPhase, renderSliceSize, &m_accumulationBuffer, useDirectConvolver));
michael@0: 
michael@0:         bool isBackgroundStage = false;
michael@0: 
michael@0:         if (this->useBackgroundThreads() && stageOffset > RealtimeFrameLimit) {
michael@0:             m_backgroundStages.AppendElement(stage.forget());
michael@0:             isBackgroundStage = true;
michael@0:         } else
michael@0:             m_stages.AppendElement(stage.forget());
michael@0: 
michael@0:         stageOffset += stageSize;
michael@0:         ++i;
michael@0: 
michael@0:         if (!useDirectConvolver) {
michael@0:             // Figure out next FFT size
michael@0:             fftSize *= 2;
michael@0:         }
michael@0: 
michael@0:         if (hasRealtimeConstraint && !isBackgroundStage && fftSize > m_maxRealtimeFFTSize)
michael@0:             fftSize = m_maxRealtimeFFTSize;
michael@0:         if (fftSize > m_maxFFTSize)
michael@0:             fftSize = m_maxFFTSize;
michael@0:     }
michael@0: 
michael@0:     // Start up background thread
michael@0:     // FIXME: would be better to up the thread priority here.  It doesn't need to be real-time, but higher than the default...
michael@0:     if (this->useBackgroundThreads() && m_backgroundStages.Length() > 0) {
michael@0:         if (!m_backgroundThread.Start()) {
michael@0:           NS_WARNING("Cannot start convolver thread.");
michael@0:           return;
michael@0:         }
michael@0:         CancelableTask* task = NewRunnableMethod(this, &ReverbConvolver::backgroundThreadEntry);
michael@0:         m_backgroundThread.message_loop()->PostTask(FROM_HERE, task);
michael@0:     }
michael@0: }
michael@0: 
michael@0: ReverbConvolver::~ReverbConvolver()
michael@0: {
michael@0:     // Wait for background thread to stop
michael@0:     if (useBackgroundThreads() && m_backgroundThread.IsRunning()) {
michael@0:         m_wantsToExit = true;
michael@0: 
michael@0:         // Wake up thread so it can return
michael@0:         {
michael@0:             AutoLock locker(m_backgroundThreadLock);
michael@0:             m_moreInputBuffered = true;
michael@0:             m_backgroundThreadCondition.Signal();
michael@0:         }
michael@0: 
michael@0:         m_backgroundThread.Stop();
michael@0:     }
michael@0: }
michael@0: 
michael@0: size_t ReverbConvolver::sizeOfIncludingThis(mozilla::MallocSizeOf aMallocSizeOf) const
michael@0: {
michael@0:     size_t amount = aMallocSizeOf(this);
michael@0:     amount += m_stages.SizeOfExcludingThis(aMallocSizeOf);
michael@0:     for (size_t i = 0; i < m_stages.Length(); i++) {
michael@0:         if (m_stages[i]) {
michael@0:             amount += m_stages[i]->sizeOfIncludingThis(aMallocSizeOf);
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     amount += m_backgroundStages.SizeOfExcludingThis(aMallocSizeOf);
michael@0:     for (size_t i = 0; i < m_backgroundStages.Length(); i++) {
michael@0:         if (m_backgroundStages[i]) {
michael@0:             amount += m_backgroundStages[i]->sizeOfIncludingThis(aMallocSizeOf);
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     // NB: The buffer sizes are static, so even though they might be accessed
michael@0:     //     in another thread it's safe to measure them.
michael@0:     amount += m_accumulationBuffer.sizeOfExcludingThis(aMallocSizeOf);
michael@0:     amount += m_inputBuffer.sizeOfExcludingThis(aMallocSizeOf);
michael@0: 
michael@0:     // Possible future measurements:
michael@0:     // - m_backgroundThread
michael@0:     // - m_backgroundThreadLock
michael@0:     // - m_backgroundThreadCondition
michael@0:     return amount;
michael@0: }
michael@0: 
michael@0: void ReverbConvolver::backgroundThreadEntry()
michael@0: {
michael@0:     while (!m_wantsToExit) {
michael@0:         // Wait for realtime thread to give us more input
michael@0:         m_moreInputBuffered = false;
michael@0:         {
michael@0:             AutoLock locker(m_backgroundThreadLock);
michael@0:             while (!m_moreInputBuffered && !m_wantsToExit)
michael@0:                 m_backgroundThreadCondition.Wait();
michael@0:         }
michael@0: 
michael@0:         // Process all of the stages until their read indices reach the input buffer's write index
michael@0:         int writeIndex = m_inputBuffer.writeIndex();
michael@0: 
michael@0:         // Even though it doesn't seem like every stage needs to maintain its own version of readIndex 
michael@0:         // we do this in case we want to run in more than one background thread.
michael@0:         int readIndex;
michael@0: 
michael@0:         while ((readIndex = m_backgroundStages[0]->inputReadIndex()) != writeIndex) { // FIXME: do better to detect buffer overrun...
michael@0:             // The ReverbConvolverStages need to process in amounts which evenly divide half the FFT size
michael@0:             const int SliceSize = MinFFTSize / 2;
michael@0: 
michael@0:             // Accumulate contributions from each stage
michael@0:             for (size_t i = 0; i < m_backgroundStages.Length(); ++i)
michael@0:                 m_backgroundStages[i]->processInBackground(this, SliceSize);
michael@0:         }
michael@0:     }
michael@0: }
michael@0: 
michael@0: void ReverbConvolver::process(const float* sourceChannelData, size_t sourceChannelLength,
michael@0:                               float* destinationChannelData, size_t destinationChannelLength,
michael@0:                               size_t framesToProcess)
michael@0: {
michael@0:     bool isSafe = sourceChannelData && destinationChannelData && sourceChannelLength >= framesToProcess && destinationChannelLength >= framesToProcess;
michael@0:     MOZ_ASSERT(isSafe);
michael@0:     if (!isSafe)
michael@0:         return;
michael@0: 
michael@0:     const float* source = sourceChannelData;
michael@0:     float* destination = destinationChannelData;
michael@0:     bool isDataSafe = source && destination;
michael@0:     MOZ_ASSERT(isDataSafe);
michael@0:     if (!isDataSafe)
michael@0:         return;
michael@0: 
michael@0:     // Feed input buffer (read by all threads)
michael@0:     m_inputBuffer.write(source, framesToProcess);
michael@0: 
michael@0:     // Accumulate contributions from each stage
michael@0:     for (size_t i = 0; i < m_stages.Length(); ++i)
michael@0:         m_stages[i]->process(source, framesToProcess);
michael@0: 
michael@0:     // Finally read from accumulation buffer
michael@0:     m_accumulationBuffer.readAndClear(destination, framesToProcess);
michael@0: 
michael@0:     // Now that we've buffered more input, wake up our background thread.
michael@0: 
michael@0:     // Not using a MutexLocker looks strange, but we use a tryLock() instead because this is run on the real-time
michael@0:     // thread where it is a disaster for the lock to be contended (causes audio glitching).  It's OK if we fail to
michael@0:     // signal from time to time, since we'll get to it the next time we're called.  We're called repeatedly
michael@0:     // and frequently (around every 3ms).  The background thread is processing well into the future and has a considerable amount of 
michael@0:     // leeway here...
michael@0:     if (m_backgroundThreadLock.Try()) {
michael@0:         m_moreInputBuffered = true;
michael@0:         m_backgroundThreadCondition.Signal();
michael@0:         m_backgroundThreadLock.Release();
michael@0:     }
michael@0: }
michael@0: 
michael@0: void ReverbConvolver::reset()
michael@0: {
michael@0:     for (size_t i = 0; i < m_stages.Length(); ++i)
michael@0:         m_stages[i]->reset();
michael@0: 
michael@0:     for (size_t i = 0; i < m_backgroundStages.Length(); ++i)
michael@0:         m_backgroundStages[i]->reset();
michael@0: 
michael@0:     m_accumulationBuffer.reset();
michael@0:     m_inputBuffer.reset();
michael@0: }
michael@0: 
michael@0: size_t ReverbConvolver::latencyFrames() const
michael@0: {
michael@0:     return 0;
michael@0: }
michael@0: 
michael@0: } // namespace WebCore