Mirrors/overte
3
0
Fork 0
mirror of https://github.com/overte-org/overte.git synced 2025-04-20 12:04:18 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge pull request #2331 from birarda/audio-scaling

Browse source 
AudioMixer optimizations with MMX intrinsics
This commit is contained in:
Brad Hefta-Gaub 2014-03-17 11:02:44 -07:00
commit 927006df91
4 changed files with 134 additions and 27 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

149
assignment-client/src/audio/AudioMixer.cpp
View file

@ -6,6 +6,7 @@
// Copyright (c) 2013 HighFidelity, Inc. All rights reserved.
//
#include <mmintrin.h>
#include <errno.h>
#include <fcntl.h>
#include <fstream>
@ -73,9 +74,7 @@ void AudioMixer::addBufferToMixForListeningNodeWithBuffer(PositionalAudioRingBuf
float attenuationCoefficient = 1.0f;
int numSamplesDelay = 0;
float weakChannelAmplitudeRatio = 1.0f;
const int PHASE_DELAY_AT_90 = 20;
if (bufferToAdd != listeningNodeBuffer) {
// if the two buffer pointers do not match then these are different buffers
@ -148,7 +147,7 @@ void AudioMixer::addBufferToMixForListeningNodeWithBuffer(PositionalAudioRingBuf
// figure out the number of samples of delay and the ratio of the amplitude
// in the weak channel for audio spatialization
float sinRatio = fabsf(sinf(bearingRelativeAngleToSource));
numSamplesDelay = PHASE_DELAY_AT_90 * sinRatio;
numSamplesDelay = SAMPLE_PHASE_DELAY_AT_90 * sinRatio;
weakChannelAmplitudeRatio = 1 - (PHASE_AMPLITUDE_RATIO_AT_90 * sinRatio);
}
}
@ -156,26 +155,130 @@ void AudioMixer::addBufferToMixForListeningNodeWithBuffer(PositionalAudioRingBuf
// if the bearing relative angle to source is > 0 then the delayed channel is the right one
int delayedChannelOffset = (bearingRelativeAngleToSource > 0.0f) ? 1 : 0;
int goodChannelOffset = delayedChannelOffset == 0 ? 1 : 0;
const int16_t* nextOutputStart = bufferToAdd->getNextOutput();
const int16_t* bufferStart = bufferToAdd->getBuffer();
int ringBufferSampleCapacity = bufferToAdd->getSampleCapacity();
for (int s = 0; s < NETWORK_BUFFER_LENGTH_SAMPLES_STEREO; s += 2) {
if ((s / 2) < numSamplesDelay) {
// pull the earlier sample for the delayed channel
int earlierSample = (*bufferToAdd)[(s / 2) - numSamplesDelay] * attenuationCoefficient * weakChannelAmplitudeRatio;
_clientSamples[s + delayedChannelOffset] = glm::clamp(_clientSamples[s + delayedChannelOffset] + earlierSample,
MIN_SAMPLE_VALUE, MAX_SAMPLE_VALUE);
int16_t correctBufferSample[2], delayBufferSample[2];
int delayedChannelIndex = 0;
const int SINGLE_STEREO_OFFSET = 2;
for (int s = 0; s < NETWORK_BUFFER_LENGTH_SAMPLES_STEREO; s += 4) {
// setup the int16_t variables for the two sample sets
correctBufferSample[0] = nextOutputStart[s / 2] * attenuationCoefficient;
correctBufferSample[1] = nextOutputStart[(s / 2) + 1] * attenuationCoefficient;
delayedChannelIndex = s + (numSamplesDelay * 2) + delayedChannelOffset;
delayBufferSample[0] = correctBufferSample[0] * weakChannelAmplitudeRatio;
delayBufferSample[1] = correctBufferSample[1] * weakChannelAmplitudeRatio;
__m64 bufferSamples = _mm_set_pi16(_clientSamples[s + goodChannelOffset],
_clientSamples[s + goodChannelOffset + SINGLE_STEREO_OFFSET],
_clientSamples[delayedChannelIndex],
_clientSamples[delayedChannelIndex + SINGLE_STEREO_OFFSET]);
__m64 addedSamples = _mm_set_pi16(correctBufferSample[0], correctBufferSample[1],
delayBufferSample[0], delayBufferSample[1]);
// perform the MMX add (with saturation) of two correct and delayed samples
__m64 mmxResult = _mm_adds_pi16(bufferSamples, addedSamples);
int16_t* shortResults = reinterpret_cast<int16_t*>(&mmxResult);
// assign the results from the result of the mmx arithmetic
_clientSamples[s + goodChannelOffset] = shortResults[3];
_clientSamples[s + goodChannelOffset + SINGLE_STEREO_OFFSET] = shortResults[2];
_clientSamples[delayedChannelIndex] = shortResults[1];
_clientSamples[delayedChannelIndex + SINGLE_STEREO_OFFSET] = shortResults[0];
}
// The following code is pretty gross and redundant, but AFAIK it's the best way to avoid
// too many conditionals in handling the delay samples at the beginning of _clientSamples.
// Basically we try to take the samples in batches of four, and then handle the remainder
// conditionally to get rid of the rest.
const int DOUBLE_STEREO_OFFSET = 4;
const int TRIPLE_STEREO_OFFSET = 6;
if (numSamplesDelay > 0) {
// if there was a sample delay for this buffer, we need to pull samples prior to the nextOutput
// to stick at the beginning
float attenuationAndWeakChannelRatio = attenuationCoefficient * weakChannelAmplitudeRatio;
const int16_t* delayNextOutputStart = nextOutputStart - numSamplesDelay;
if (delayNextOutputStart < bufferStart) {
delayNextOutputStart = bufferStart + ringBufferSampleCapacity - numSamplesDelay;
}
// pull the current sample for the good channel
int16_t currentSample = (*bufferToAdd)[s / 2] * attenuationCoefficient;
_clientSamples[s + goodChannelOffset] = glm::clamp(_clientSamples[s + goodChannelOffset] + currentSample,
MIN_SAMPLE_VALUE, MAX_SAMPLE_VALUE);
if ((s / 2) + numSamplesDelay < NETWORK_BUFFER_LENGTH_SAMPLES_PER_CHANNEL) {
// place the current sample at the right spot in the delayed channel
int16_t clampedSample = glm::clamp((int) (_clientSamples[s + (numSamplesDelay * 2) + delayedChannelOffset]
+ (currentSample * weakChannelAmplitudeRatio)),
MIN_SAMPLE_VALUE, MAX_SAMPLE_VALUE);
_clientSamples[s + (numSamplesDelay * 2) + delayedChannelOffset] = clampedSample;
int i = 0;
while (i + 3 < numSamplesDelay) {
// handle the first cases where we can MMX add four samples at once
int parentIndex = i * 2;
__m64 bufferSamples = _mm_setr_pi16(_clientSamples[parentIndex + delayedChannelOffset],
_clientSamples[parentIndex + SINGLE_STEREO_OFFSET + delayedChannelOffset],
_clientSamples[parentIndex + DOUBLE_STEREO_OFFSET + delayedChannelOffset],
_clientSamples[parentIndex + TRIPLE_STEREO_OFFSET + delayedChannelOffset]);
__m64 addSamples = _mm_set_pi16(delayNextOutputStart[i] * attenuationAndWeakChannelRatio,
delayNextOutputStart[i + 1] * attenuationAndWeakChannelRatio,
delayNextOutputStart[i + 2] * attenuationAndWeakChannelRatio,
delayNextOutputStart[i + 3] * attenuationAndWeakChannelRatio);
__m64 mmxResult = _mm_adds_pi16(bufferSamples, addSamples);
int16_t* shortResults = reinterpret_cast<int16_t*>(&mmxResult);
_clientSamples[parentIndex + delayedChannelOffset] = shortResults[3];
_clientSamples[parentIndex + SINGLE_STEREO_OFFSET + delayedChannelOffset] = shortResults[2];
_clientSamples[parentIndex + DOUBLE_STEREO_OFFSET + delayedChannelOffset] = shortResults[1];
_clientSamples[parentIndex + TRIPLE_STEREO_OFFSET + delayedChannelOffset] = shortResults[0];
// push the index
i += 4;
}
int parentIndex = i * 2;
if (i + 2 < numSamplesDelay) {
// MMX add only three delayed samples
__m64 bufferSamples = _mm_set_pi16(_clientSamples[parentIndex + delayedChannelOffset],
_clientSamples[parentIndex + SINGLE_STEREO_OFFSET + delayedChannelOffset],
_clientSamples[parentIndex + DOUBLE_STEREO_OFFSET + delayedChannelOffset],
0);
__m64 addSamples = _mm_set_pi16(delayNextOutputStart[i] * attenuationAndWeakChannelRatio,
delayNextOutputStart[i + 1] * attenuationAndWeakChannelRatio,
delayNextOutputStart[i + 2] * attenuationAndWeakChannelRatio,
0);
__m64 mmxResult = _mm_adds_pi16(bufferSamples, addSamples);
int16_t* shortResults = reinterpret_cast<int16_t*>(&mmxResult);
_clientSamples[parentIndex + delayedChannelOffset] = shortResults[3];
_clientSamples[parentIndex + SINGLE_STEREO_OFFSET + delayedChannelOffset] = shortResults[2];
_clientSamples[parentIndex + DOUBLE_STEREO_OFFSET + delayedChannelOffset] = shortResults[1];
} else if (i + 1 < numSamplesDelay) {
// MMX add two delayed samples
__m64 bufferSamples = _mm_set_pi16(_clientSamples[parentIndex + delayedChannelOffset],
_clientSamples[parentIndex + SINGLE_STEREO_OFFSET + delayedChannelOffset], 0, 0);
__m64 addSamples = _mm_set_pi16(delayNextOutputStart[i] * attenuationAndWeakChannelRatio,
delayNextOutputStart[i + 1] * attenuationAndWeakChannelRatio, 0, 0);
__m64 mmxResult = _mm_adds_pi16(bufferSamples, addSamples);
int16_t* shortResults = reinterpret_cast<int16_t*>(&mmxResult);
_clientSamples[parentIndex + delayedChannelOffset] = shortResults[3];
_clientSamples[parentIndex + SINGLE_STEREO_OFFSET + delayedChannelOffset] = shortResults[2];
} else if (i < numSamplesDelay) {
// MMX add a single delayed sample
__m64 bufferSamples = _mm_set_pi16(_clientSamples[parentIndex + delayedChannelOffset], 0, 0, 0);
__m64 addSamples = _mm_set_pi16(delayNextOutputStart[i] * attenuationAndWeakChannelRatio, 0, 0, 0);
__m64 mmxResult = _mm_adds_pi16(bufferSamples, addSamples);
int16_t* shortResults = reinterpret_cast<int16_t*>(&mmxResult);
_clientSamples[parentIndex + delayedChannelOffset] = shortResults[3];
}
}
}
@ -269,7 +372,7 @@ void AudioMixer::run() {
&& ((AudioMixerClientData*) node->getLinkedData())->getAvatarAudioRingBuffer()) {
prepareMixForListeningNode(node.data());
memcpy(_clientMixBuffer.data() + numBytesPacketHeader, _clientSamples, sizeof(_clientSamples));
memcpy(_clientMixBuffer.data() + numBytesPacketHeader, _clientSamples, NETWORK_BUFFER_LENGTH_BYTES_STEREO);
nodeList->writeDatagram(_clientMixBuffer, node);
}
}

5
assignment-client/src/audio/AudioMixer.h
View file

@ -16,6 +16,8 @@
class PositionalAudioRingBuffer;
class AvatarAudioRingBuffer;
const int SAMPLE_PHASE_DELAY_AT_90 = 20;
/// Handles assignments of type AudioMixer - mixing streams of audio and re-distributing to various clients.
class AudioMixer : public ThreadedAssignment {
Q_OBJECT
@ -38,7 +40,8 @@ private:
QByteArray _clientMixBuffer;
int16_t _clientSamples[NETWORK_BUFFER_LENGTH_SAMPLES_STEREO];
// client samples capacity is larger than what will be sent to optimize mixing
int16_t _clientSamples[NETWORK_BUFFER_LENGTH_SAMPLES_STEREO + SAMPLE_PHASE_DELAY_AT_90];
};
#endif /* defined(__hifi__AudioMixer__) */

3
libraries/audio/src/AudioRingBuffer.cpp
View file

@ -121,9 +121,6 @@ qint64 AudioRingBuffer::writeData(const char* data, qint64 maxSize) {
}
int16_t& AudioRingBuffer::operator[](const int index) {
// make sure this is a valid index
assert(index > -_sampleCapacity && index < _sampleCapacity);
return *shiftedPositionAccomodatingWrap(_nextOutput, index);
}

4
libraries/audio/src/AudioRingBuffer.h
View file

@ -45,6 +45,10 @@ public:
int getSampleCapacity() const { return _sampleCapacity; }
int parseData(const QByteArray& packet);
// assume callers using this will never wrap around the end
const int16_t* getNextOutput() { return _nextOutput; }
const int16_t* getBuffer() { return _buffer; }
qint64 readSamples(int16_t* destination, qint64 maxSamples);
qint64 writeSamples(const int16_t* source, qint64 maxSamples);