EffectStage.cpp

Go to the documentation of this file.

/**********************************************************************
 
  Audacity: A Digital Audio Editor
 
  @file EffectStage.cpp
 
  Dominic Mazzoni
  Vaughan Johnson
  Martyn Shaw
 
  Paul Licameli split from PerTrackEffect.cpp
 
**********************************************************************/
#include "EffectStage.h"
#include "AudacityException.h"
#include "AudioGraphBuffers.h"
#include <cassert>
 
namespace {
std::vector<std::shared_ptr<EffectInstance>> MakeInstances(
   const EffectStage::Factory& factory, EffectSettings& settings,
   double sampleRate, std::optional<sampleCount> genLength, int channel,
   int nInputChannels)
{
   std::vector<std::shared_ptr<EffectInstance>> instances;
   // Make as many instances as needed for the channels of the source, which
   // depends on how the instances report how many channels they accept
   const auto nChannels = (channel < 0) ? nInputChannels : 1;
   for (size_t ii = 0; ii < nChannels;)
   {
      auto pInstance = factory();
      if (!pInstance)
         // A constructor that can't satisfy its post should throw instead
         throw std::exception{};
      auto count = pInstance->GetAudioInCount();
      ChannelName map[3]{ ChannelNameEOL, ChannelNameEOL, ChannelNameEOL };
      MakeChannelMap(nInputChannels, channel, map);
      // Give the plugin a chance to initialize
      if (!pInstance->ProcessInitialize(settings, sampleRate, map))
         throw std::exception{};
      instances.resize(ii);
 
      // Beware generators with zero in count
      if (genLength)
         count = nChannels;
 
      instances.push_back(move(pInstance));
 
      // Advance ii
      if (count == 0)
         // What? Can't make progress
         throw std::exception();
      ii += count;
   }
   return instances;
}
}
 
EffectStage::EffectStage(
   CreateToken, int channel, int nInputChannels, Source& upstream,
   Buffers& inBuffers, const Factory& factory, EffectSettings& settings,
   double sampleRate, std::optional<sampleCount> genLength)
    : mUpstream { upstream }
    , mInBuffers { inBuffers }
    , mInstances { MakeInstances(
         factory, settings, sampleRate, genLength, channel, nInputChannels) }
    , mSettings { settings }
    , mSampleRate { sampleRate }
    , mIsProcessor { !genLength.has_value() }
    , mDelayRemaining { genLength ? *genLength : sampleCount::max() }
{
   assert(upstream.AcceptsBlockSize(inBuffers.BlockSize()));
   assert(this->AcceptsBlockSize(inBuffers.BlockSize()));
 
   // Establish invariant
   mInBuffers.Rewind();
}
 
auto EffectStage::Create(
   int channel, int nInputChannels, Source& upstream, Buffers& inBuffers,
   const Factory& factory, EffectSettings& settings, double sampleRate,
   std::optional<sampleCount> genLength) -> std::unique_ptr<EffectStage>
{
   try {
      return std::make_unique<EffectStage>(
         CreateToken {}, channel, nInputChannels, upstream, inBuffers, factory,
         settings, sampleRate, genLength);
   }
   catch (const std::exception &) {
      return nullptr;
   }
}
 
EffectStage::~EffectStage()
{
   // Allow the instances to cleanup
   for (auto &pInstance : mInstances)
      if (pInstance)
         pInstance->ProcessFinalize();
}
 
bool EffectStage::AcceptsBuffers(const Buffers &buffers) const
{
   return true;
}
 
bool EffectStage::AcceptsBlockSize(size_t size) const
{
   // Test the equality of input and output block sizes
   return mInBuffers.BlockSize() == size;
}
 
std::optional<size_t> EffectStage::Acquire(Buffers &data, size_t bound)
{
   assert(AcceptsBuffers(data));
   assert(AcceptsBlockSize(data.BlockSize()));
   // pre, needed for Process() and Discard()
   assert(bound <= std::min(data.BlockSize(), data.Remaining()));
 
   // For each input block of samples, we pass it to the effect along with a
   // variable output location.  This output location is simply a pointer into a
   // much larger buffer.  This reduces the number of calls required to add the
   // samples to the output track.
   //
   // Upon return from the effect, the output samples are "moved to the left" by
   // the number of samples in the current latency setting, effectively removing any
   // delay introduced by the effect.
   //
   // At the same time the total number of delayed samples are gathered and when
   // there is no further input data to process, the loop continues to call the
   // effect with an empty input buffer until the effect has had a chance to
   // return all of the remaining delayed samples.
 
   // Invariant satisfies pre for mUpstream.Acquire() and for Process()
   assert(mInBuffers.BlockSize() <= mInBuffers.Remaining());
 
   size_t curBlockSize = 0;
 
   if (auto oCurBlockSize = FetchProcessAndAdvance(data, bound, false)
      ; !oCurBlockSize
   )
      return {};
   else {
      curBlockSize = *oCurBlockSize;
      if (mIsProcessor && !mLatencyDone) {
         // Come here only in the first call to Acquire()
         // Some effects (like ladspa/lv2 swh plug-ins) don't report latency
         // until at least one block of samples is processed.  Find latency
         // once only for the track and assume it doesn't vary
         auto delay = mDelayRemaining =
            mInstances[0]->GetLatency(mSettings, mSampleRate);
         for (size_t ii = 1, nn = mInstances.size(); ii < nn; ++ii)
            if (mInstances[ii] &&
               mInstances[ii]->GetLatency(mSettings, mSampleRate) != delay)
               // This mismatch is unexpected.  Fail
               return {};
         // Discard all the latency
         while (delay > 0 && curBlockSize > 0) {
            auto discard = limitSampleBufferSize(curBlockSize, delay);
            data.Discard(discard, curBlockSize - discard);
            delay -= discard;
            curBlockSize -= discard;
            if (curBlockSize == 0) {
               if (!(oCurBlockSize = FetchProcessAndAdvance(data, bound, false)
               ))
                  return {};
               else
                  curBlockSize = *oCurBlockSize;
            }
            mLastProduced -= discard;
         }
         if (curBlockSize > 0) {
            assert(delay == 0);
            if (curBlockSize < bound) {
               // Discarded all the latency, while upstream may still be
               // producing.  Try to fill the buffer up to the bound.
               if (!(oCurBlockSize = FetchProcessAndAdvance(
                  data, bound - curBlockSize, false, curBlockSize)
               ))
                  return {};
               else
                  curBlockSize += *oCurBlockSize;
            }
         }
         else while (delay > 0) {
            assert(curBlockSize == 0);
            // Finish one-time delay in case it exceeds entire upstream length
            // Upstream must have been exhausted
            assert(mUpstream.Remaining() == 0);
            // Feed zeroes to the effect
            auto zeroes = limitSampleBufferSize(data.BlockSize(), delay);
            if (!(FetchProcessAndAdvance(data, zeroes, true)))
               return {};
            delay -= zeroes;
            // Debit mDelayRemaining later in Release()
         }
         mLatencyDone = true;
      }
   }
 
   if (mIsProcessor && curBlockSize < bound) {
      // If there is still a short buffer by this point, upstream must have
      // been exhausted
      assert(mUpstream.Remaining() == 0);
 
      // Continue feeding zeroes; this code block will produce as many zeroes
      // at the end as were discarded at the beginning (over one or more visits)
      auto zeroes =
         limitSampleBufferSize(bound - curBlockSize, mDelayRemaining);
      if (!FetchProcessAndAdvance(data, zeroes, true, curBlockSize))
         return {};
      // Debit mDelayRemaining later in Release()
   }
 
   auto result = mLastProduced + mLastZeroes;
   // assert the post
   assert(data.Remaining() > 0);
   assert(result <= bound);
   assert(result <= data.Remaining());
   assert(result <= Remaining());
   assert(bound == 0 || Remaining() == 0 || result > 0);
   return { result };
}
 
std::optional<size_t> EffectStage::FetchProcessAndAdvance(
   Buffers &data, size_t bound, bool doZeroes, size_t outBufferOffset)
{
   std::optional<size_t> oCurBlockSize;
   // Generator always supplies zeroes in
   doZeroes = doZeroes || !mIsProcessor;
   if (!doZeroes)
      oCurBlockSize = mUpstream.Acquire(mInBuffers, bound);
   else {
      if (!mCleared) {
         // Need to do this the first time, only, that we begin to give zeroes
         // to the processor
         mInBuffers.Rewind();
         const auto blockSize = mInBuffers.BlockSize();
         for (size_t ii = 0; ii < mInBuffers.Channels(); ++ii) {
            auto p = &mInBuffers.GetWritePosition(ii);
            std::fill(p, p + blockSize, 0);
         }
         mCleared = true;
      }
      oCurBlockSize = {
         mIsProcessor ? bound : limitSampleBufferSize(bound, mDelayRemaining) };
      if (!mIsProcessor)
         // Do this (ignoring result) so we can correctly Release() upstream
         mUpstream.Acquire(mInBuffers, bound);
   }
   if (!oCurBlockSize)
      return {};
 
   const auto curBlockSize = *oCurBlockSize;
   if (curBlockSize == 0)
      assert(doZeroes || mUpstream.Remaining() == 0); // post of Acquire()
   else {
      // Called only in Acquire()
      // invariant or post of mUpstream.Acquire() satisfies pre of Process()
      // because curBlockSize <= bound <= mInBuffers.blockSize()
      //    == data.BlockSize()
      // and mInBuffers.BlockSize() <= mInBuffers.Remaining() by invariant
      // and data.BlockSize() <= data.Remaining() by pre of Acquire()
      for (size_t ii = 0, nn = mInstances.size(); ii < nn; ++ii) {
         auto &pInstance = mInstances[ii];
         if (!pInstance)
            continue;
         if (!Process(*pInstance, ii, data, curBlockSize, outBufferOffset))
            return {};
      }
 
      if (doZeroes) {
         // Either a generator or doing the tail; will count down delay
         mLastZeroes = limitSampleBufferSize(curBlockSize, DelayRemaining());
         if (!mIsProcessor) {
            // This allows polling the progress meter for a generator
            if (!mUpstream.Release())
               return {};
         }
      }
      else {
         // Will count down the upstream
         mLastProduced += curBlockSize;
         if (!mUpstream.Release())
            return {};
         mInBuffers.Advance(curBlockSize);
         if (mInBuffers.Remaining() < mInBuffers.BlockSize())
            // Maintain invariant minimum availability
            mInBuffers.Rotate();
      }
   }
   return oCurBlockSize;
}
 
bool EffectStage::Process(EffectInstance &instance,
   size_t channel, const Buffers &data, size_t curBlockSize,
   size_t outBufferOffset) const
{
   size_t processed{};
   try {
      const auto positions = mInBuffers.Positions();
      const auto nPositions = mInBuffers.Channels();
      // channel may be nonzero in the case of a plug-in that only reads
      // one channel at a time, so multiple instances are made to mix stereo
      assert(channel <= nPositions);
      std::vector<float *> inPositions(
         positions + channel, positions + nPositions - channel);
      // When the plug-in expects many input channels, replicate the last
      // buffer (assumed to be zero-filled) as dummy input
      inPositions.resize(
         instance.GetAudioInCount() - channel, inPositions.back());
 
      std::vector<float *> advancedOutPositions;
      const auto size = instance.GetAudioOutCount() - channel;
      advancedOutPositions.reserve(size);
 
      auto outPositions = data.Positions();
      // It is assumed that data has at least one dummy buffer last
      auto channels = data.Channels();
      // channel may be nonzero in the case of a plug-in that only writes
      // one channel at a time, so multiple instances are made to mix stereo
      for (size_t ii = channel; ii < channels; ++ii)
         advancedOutPositions.push_back(outPositions[ii] + outBufferOffset);
      // When the plug-in expects many output channels, replicate the last
      // as dummy output
      advancedOutPositions.resize(size, advancedOutPositions.back());
 
      processed = instance.ProcessBlock(mSettings,
         inPositions.data(), advancedOutPositions.data(), curBlockSize);
   }
   catch (const AudacityException &) {
      // PRL: Bug 437:
      // Pass this along to our application-level handler
      throw;
   }
   catch (...) {
      // PRL:
      // Exceptions for other reasons, maybe in third-party code...
      // Continue treating them as we used to, but I wonder if these
      // should now be treated the same way.
      return false;
   }
 
   return (processed == curBlockSize);
}
 
sampleCount EffectStage::Remaining() const
{
   // Not correct until at least one call to Acquire() so that mDelayRemaining
   // is assigned.
   // mLastProduced will have the up-front latency discarding deducted.
   // mDelayRemaining later decreases to 0 as zeroes are supplied to the
   // processor at the end, compensating for the discarding.
   return mLastProduced
      + (mIsProcessor ? mUpstream.Remaining() : 0)
      + DelayRemaining();
}
 
bool EffectStage::Release()
{
   // Progress toward termination (Remaining() == 0),
   // if mLastProduced + mLastZeroes > 0,
   // which is what Acquire() last returned
   mDelayRemaining -= mLastZeroes;
   assert(mDelayRemaining >= 0);
   mLastProduced = mLastZeroes = 0;
   return true;
}
 
unsigned MakeChannelMap(int nChannels, int channel, ChannelName map[3])
{
   assert(channel < static_cast<int>(nChannels)); // precondition
 
   size_t index = 0;
   if (nChannels == 1)
      map[index++] = ChannelNameMono;
   else {
      // TODO: more-than-two-channels
      if (channel < 1)
         map[index++] = ChannelNameFrontLeft;
      if (channel != 0)
         map[index++] = ChannelNameFrontRight;
   }
   map[index] = ChannelNameEOL;
   return index;
}