mirror of
https://github.com/HifiExperiments/overte.git
synced 2025-04-07 10:02:24 +02:00
Merge pull request #15307 from kencooke/audio-gain-interpolation
Case 22019: Audio clicks/pops when gain is rapidly changing
This commit is contained in:
Sam Gateau
GitHub
commit
6f4f7335dc
7 changed files with 205 additions and 93 deletions
|
|
@ -549,38 +549,28 @@ void AudioMixerSlave::addStream(AudioMixerClientData::MixableStream& mixableStre
|
|||
// grab the stream from the ring buffer
|
||||
AudioRingBuffer::ConstIterator streamPopOutput = streamToAdd->getLastPopOutput();
|
||||
|
||||
// stereo sources are not passed through HRTF
|
||||
if (streamToAdd->isStereo()) {
|
||||
|
||||
// apply the avatar gain adjustment
|
||||
gain *= mixableStream.hrtf->getGainAdjustment();
|
||||
streamPopOutput.readSamples(_bufferSamples, AudioConstants::NETWORK_FRAME_SAMPLES_STEREO);
|
||||
|
||||
const float scale = 1 / 32768.0f; // int16_t to float
|
||||
|
||||
for (int i = 0; i < AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL; i++) {
|
||||
_mixSamples[2*i+0] += (float)streamPopOutput[2*i+0] * gain * scale;
|
||||
_mixSamples[2*i+1] += (float)streamPopOutput[2*i+1] * gain * scale;
|
||||
}
|
||||
// stereo sources are not passed through HRTF
|
||||
mixableStream.hrtf->mixStereo(_bufferSamples, _mixSamples, gain, AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL);
|
||||
|
||||
++stats.manualStereoMixes;
|
||||
} else if (isEcho) {
|
||||
|
||||
streamPopOutput.readSamples(_bufferSamples, AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL);
|
||||
|
||||
// echo sources are not passed through HRTF
|
||||
|
||||
const float scale = 1/32768.0f; // int16_t to float
|
||||
|
||||
for (int i = 0; i < AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL; i++) {
|
||||
float sample = (float)streamPopOutput[i] * gain * scale;
|
||||
_mixSamples[2*i+0] += sample;
|
||||
_mixSamples[2*i+1] += sample;
|
||||
}
|
||||
mixableStream.hrtf->mixMono(_bufferSamples, _mixSamples, gain, AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL);
|
||||
|
||||
++stats.manualEchoMixes;
|
||||
} else {
|
||||
|
||||
streamPopOutput.readSamples(_bufferSamples, AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL);
|
||||
|
||||
mixableStream.hrtf->render(_bufferSamples, _mixSamples, HRTF_DATASET_INDEX, azimuth, distance, gain,
|
||||
AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL);
|
||||
|
||||
++stats.hrtfRenders;
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -1397,7 +1397,6 @@ bool AudioClient::mixLocalAudioInjectors(float* mixBuffer) {
|
|||
// spatialize into mixBuffer
|
||||
injector->getLocalFOA().render(_localScratchBuffer, mixBuffer, HRTF_DATASET_INDEX,
|
||||
qw, qx, qy, qz, gain, AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL);
|
||||
|
||||
} else if (options.stereo) {
|
||||
|
||||
if (options.positionSet) {
|
||||
|
|
@ -1409,11 +1408,8 @@ bool AudioClient::mixLocalAudioInjectors(float* mixBuffer) {
|
|||
}
|
||||
|
||||
// direct mix into mixBuffer
|
||||
for (int i = 0; i < AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL; i++) {
|
||||
mixBuffer[2*i+0] += convertToFloat(_localScratchBuffer[2*i+0]) * gain;
|
||||
mixBuffer[2*i+1] += convertToFloat(_localScratchBuffer[2*i+1]) * gain;
|
||||
}
|
||||
|
||||
injector->getLocalHRTF().mixStereo(_localScratchBuffer, mixBuffer, gain,
|
||||
AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL);
|
||||
} else { // injector is mono
|
||||
|
||||
if (options.positionSet) {
|
||||
|
|
@ -1431,11 +1427,8 @@ bool AudioClient::mixLocalAudioInjectors(float* mixBuffer) {
|
|||
} else {
|
||||
|
||||
// direct mix into mixBuffer
|
||||
for (int i = 0; i < AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL; i++) {
|
||||
float sample = convertToFloat(_localScratchBuffer[i]) * gain;
|
||||
mixBuffer[2*i+0] += sample;
|
||||
mixBuffer[2*i+1] += sample;
|
||||
}
|
||||
injector->getLocalHRTF().mixMono(_localScratchBuffer, mixBuffer, gain,
|
||||
AudioConstants::NETWORK_FRAME_SAMPLES_PER_CHANNEL);
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -882,14 +882,16 @@ static void convertInput_ref(int16_t* src, float *dst[4], float gain, int numFra
|
|||
|
||||
#endif
|
||||
|
||||
// in-place rotation of the soundfield
|
||||
// crossfade between old and new rotation, to prevent artifacts
|
||||
static void rotate_3x3_ref(float* buf[4], const float m0[3][3], const float m1[3][3], const float* win, int numFrames) {
|
||||
// in-place rotation and scaling of the soundfield
|
||||
// crossfade between old and new matrix, to prevent artifacts
|
||||
static void rotate_4x4_ref(float* buf[4], const float m0[4][4], const float m1[4][4], const float* win, int numFrames) {
|
||||
|
||||
const float md[3][3] = {
|
||||
{ m0[0][0] - m1[0][0], m0[0][1] - m1[0][1], m0[0][2] - m1[0][2] },
|
||||
{ m0[1][0] - m1[1][0], m0[1][1] - m1[1][1], m0[1][2] - m1[1][2] },
|
||||
{ m0[2][0] - m1[2][0], m0[2][1] - m1[2][1], m0[2][2] - m1[2][2] },
|
||||
// matrix difference
|
||||
const float md[4][4] = {
|
||||
{ m0[0][0] - m1[0][0], m0[0][1] - m1[0][1], m0[0][2] - m1[0][2], m0[0][3] - m1[0][3] },
|
||||
{ m0[1][0] - m1[1][0], m0[1][1] - m1[1][1], m0[1][2] - m1[1][2], m0[1][3] - m1[1][3] },
|
||||
{ m0[2][0] - m1[2][0], m0[2][1] - m1[2][1], m0[2][2] - m1[2][2], m0[2][3] - m1[2][3] },
|
||||
{ m0[3][0] - m1[3][0], m0[3][1] - m1[3][1], m0[3][2] - m1[3][2], m0[3][3] - m1[3][3] },
|
||||
};
|
||||
|
||||
for (int i = 0; i < numFrames; i++) {
|
||||
|
|
@ -898,22 +900,27 @@ static void rotate_3x3_ref(float* buf[4], const float m0[3][3], const float m1[3
|
|||
|
||||
// interpolate the matrix
|
||||
float m00 = m1[0][0] + frac * md[0][0];
|
||||
float m10 = m1[1][0] + frac * md[1][0];
|
||||
float m20 = m1[2][0] + frac * md[2][0];
|
||||
|
||||
float m01 = m1[0][1] + frac * md[0][1];
|
||||
float m11 = m1[1][1] + frac * md[1][1];
|
||||
float m21 = m1[2][1] + frac * md[2][1];
|
||||
float m31 = m1[3][1] + frac * md[3][1];
|
||||
|
||||
float m02 = m1[0][2] + frac * md[0][2];
|
||||
float m12 = m1[1][2] + frac * md[1][2];
|
||||
float m22 = m1[2][2] + frac * md[2][2];
|
||||
float m32 = m1[3][2] + frac * md[3][2];
|
||||
|
||||
float m13 = m1[1][3] + frac * md[1][3];
|
||||
float m23 = m1[2][3] + frac * md[2][3];
|
||||
float m33 = m1[3][3] + frac * md[3][3];
|
||||
|
||||
// matrix multiply
|
||||
float x = m00 * buf[1][i] + m01 * buf[2][i] + m02 * buf[3][i];
|
||||
float y = m10 * buf[1][i] + m11 * buf[2][i] + m12 * buf[3][i];
|
||||
float z = m20 * buf[1][i] + m21 * buf[2][i] + m22 * buf[3][i];
|
||||
float w = m00 * buf[0][i];
|
||||
|
||||
float x = m11 * buf[1][i] + m12 * buf[2][i] + m13 * buf[3][i];
|
||||
float y = m21 * buf[1][i] + m22 * buf[2][i] + m23 * buf[3][i];
|
||||
float z = m31 * buf[1][i] + m32 * buf[2][i] + m33 * buf[3][i];
|
||||
|
||||
buf[0][i] = w;
|
||||
buf[1][i] = x;
|
||||
buf[2][i] = y;
|
||||
buf[3][i] = z;
|
||||
|
|
@ -932,7 +939,7 @@ void rfft512_AVX2(float buf[512]);
|
|||
void rifft512_AVX2(float buf[512]);
|
||||
void rfft512_cmadd_1X2_AVX2(const float src[512], const float coef0[512], const float coef1[512], float dst0[512], float dst1[512]);
|
||||
void convertInput_AVX2(int16_t* src, float *dst[4], float gain, int numFrames);
|
||||
void rotate_3x3_AVX2(float* buf[4], const float m0[3][3], const float m1[3][3], const float* win, int numFrames);
|
||||
void rotate_4x4_AVX2(float* buf[4], const float m0[4][4], const float m1[4][4], const float* win, int numFrames);
|
||||
|
||||
static void rfft512(float buf[512]) {
|
||||
static auto f = cpuSupportsAVX2() ? rfft512_AVX2 : rfft512_ref;
|
||||
|
|
@ -954,8 +961,8 @@ static void convertInput(int16_t* src, float *dst[4], float gain, int numFrames)
|
|||
(*f)(src, dst, gain, numFrames); // dispatch
|
||||
}
|
||||
|
||||
static void rotate_3x3(float* buf[4], const float m0[3][3], const float m1[3][3], const float* win, int numFrames) {
|
||||
static auto f = cpuSupportsAVX2() ? rotate_3x3_AVX2 : rotate_3x3_ref;
|
||||
static void rotate_4x4(float* buf[4], const float m0[4][4], const float m1[4][4], const float* win, int numFrames) {
|
||||
static auto f = cpuSupportsAVX2() ? rotate_4x4_AVX2 : rotate_4x4_ref;
|
||||
(*f)(buf, m0, m1, win, numFrames); // dispatch
|
||||
}
|
||||
|
||||
|
|
@ -965,7 +972,7 @@ static auto& rfft512 = rfft512_ref;
|
|||
static auto& rifft512 = rifft512_ref;
|
||||
static auto& rfft512_cmadd_1X2 = rfft512_cmadd_1X2_ref;
|
||||
static auto& convertInput = convertInput_ref;
|
||||
static auto& rotate_3x3 = rotate_3x3_ref;
|
||||
static auto& rotate_4x4 = rotate_4x4_ref;
|
||||
|
||||
#endif
|
||||
|
||||
|
|
@ -1007,8 +1014,8 @@ ALIGN32 static const float crossfadeTable[FOA_BLOCK] = {
|
|||
0.0020975362f, 0.0015413331f, 0.0010705384f, 0.0006852326f, 0.0003854819f, 0.0001713375f, 0.0000428362f, 0.0000000000f,
|
||||
};
|
||||
|
||||
// convert quaternion to a column-major 3x3 rotation matrix
|
||||
static void quatToMatrix_3x3(float w, float x, float y, float z, float m[3][3]) {
|
||||
// convert quaternion to a column-major 4x4 rotation matrix
|
||||
static void quatToMatrix_4x4(float w, float x, float y, float z, float m[4][4]) {
|
||||
|
||||
float xx = x * (x + x);
|
||||
float xy = x * (y + y);
|
||||
|
|
@ -1022,17 +1029,33 @@ static void quatToMatrix_3x3(float w, float x, float y, float z, float m[3][3])
|
|||
float wy = w * (y + y);
|
||||
float wz = w * (z + z);
|
||||
|
||||
m[0][0] = 1.0f - (yy + zz);
|
||||
m[0][1] = xy - wz;
|
||||
m[0][2] = xz + wy;
|
||||
m[0][0] = 1.0f;
|
||||
m[0][1] = 0.0f;
|
||||
m[0][2] = 0.0f;
|
||||
m[0][3] = 0.0f;
|
||||
|
||||
m[1][0] = xy + wz;
|
||||
m[1][1] = 1.0f - (xx + zz);
|
||||
m[1][2] = yz - wx;
|
||||
m[1][0] = 0.0f;
|
||||
m[1][1] = 1.0f - (yy + zz);
|
||||
m[1][2] = xy - wz;
|
||||
m[1][3] = xz + wy;
|
||||
|
||||
m[2][0] = xz - wy;
|
||||
m[2][1] = yz + wx;
|
||||
m[2][2] = 1.0f - (xx + yy);
|
||||
m[2][0] = 0.0f;
|
||||
m[2][1] = xy + wz;
|
||||
m[2][2] = 1.0f - (xx + zz);
|
||||
m[2][3] = yz - wx;
|
||||
|
||||
m[3][0] = 0.0f;
|
||||
m[3][1] = xz - wy;
|
||||
m[3][2] = yz + wx;
|
||||
m[3][3] = 1.0f - (xx + yy);
|
||||
}
|
||||
|
||||
static void scaleMatrix_4x4(float scale, float m[4][4]) {
|
||||
for (int i = 0; i < 4; i++) {
|
||||
for (int j = 0; j < 4; j++) {
|
||||
m[i][j] *= scale;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Ambisonic to binaural render
|
||||
|
|
@ -1047,18 +1070,26 @@ void AudioFOA::render(int16_t* input, float* output, int index, float qw, float
|
|||
ALIGN32 float inBuffer[4][FOA_BLOCK]; // deinterleaved input buffers
|
||||
|
||||
float* in[4] = { inBuffer[0], inBuffer[1], inBuffer[2], inBuffer[3] };
|
||||
float rotation[3][3];
|
||||
float rotation[4][4];
|
||||
|
||||
// convert input to deinterleaved float
|
||||
convertInput(input, in, FOA_GAIN * gain, FOA_BLOCK);
|
||||
convertInput(input, in, FOA_GAIN, FOA_BLOCK);
|
||||
|
||||
// convert quaternion to 3x3 rotation
|
||||
quatToMatrix_3x3(qw, qx, qy, qz, rotation);
|
||||
// convert quaternion to 4x4 rotation
|
||||
quatToMatrix_4x4(qw, qx, qy, qz, rotation);
|
||||
|
||||
// rotate the soundfield
|
||||
rotate_3x3(in, _rotationState, rotation, crossfadeTable, FOA_BLOCK);
|
||||
// apply gain as uniform scale
|
||||
scaleMatrix_4x4(gain, rotation);
|
||||
|
||||
// rotation history update
|
||||
// disable interpolation from reset state
|
||||
if (_resetState) {
|
||||
memcpy(_rotationState, rotation, sizeof(_rotationState));
|
||||
}
|
||||
|
||||
// rotate and scale the soundfield
|
||||
rotate_4x4(in, _rotationState, rotation, crossfadeTable, FOA_BLOCK);
|
||||
|
||||
// new parameters become old
|
||||
memcpy(_rotationState, rotation, sizeof(_rotationState));
|
||||
|
||||
//
|
||||
|
|
@ -1093,4 +1124,6 @@ void AudioFOA::render(int16_t* input, float* output, int index, float qw, float
|
|||
output[2*i+0] += accBuffer[0][i + FOA_OVERLAP];
|
||||
output[2*i+1] += accBuffer[1][i + FOA_OVERLAP];
|
||||
}
|
||||
|
||||
_resetState = false;
|
||||
}
|
||||
|
|
|
|||
|
|
@ -28,12 +28,7 @@ static_assert((FOA_BLOCK + FOA_OVERLAP) == FOA_NFFT, "FFT convolution requires L
|
|||
class AudioFOA {
|
||||
|
||||
public:
|
||||
AudioFOA() {
|
||||
// identity matrix
|
||||
_rotationState[0][0] = 1.0f;
|
||||
_rotationState[1][1] = 1.0f;
|
||||
_rotationState[2][2] = 1.0f;
|
||||
};
|
||||
AudioFOA() {};
|
||||
|
||||
//
|
||||
// input: interleaved First-Order Ambisonic source
|
||||
|
|
@ -55,8 +50,10 @@ private:
|
|||
// input history, for overlap-save
|
||||
float _fftState[4][FOA_OVERLAP] = {};
|
||||
|
||||
// orientation history
|
||||
float _rotationState[3][3] = {};
|
||||
// orientation and gain history
|
||||
float _rotationState[4][4] = {};
|
||||
|
||||
bool _resetState = true;
|
||||
};
|
||||
|
||||
#endif // AudioFOA_h
|
||||
|
|
|
|||
|
|
@ -750,6 +750,43 @@ static void interpolate(const float* src0, const float* src1, float* dst, float
|
|||
|
||||
#endif
|
||||
|
||||
// apply gain crossfade with accumulation (interleaved)
|
||||
static void gainfade_1x2(int16_t* src, float* dst, const float* win, float gain0, float gain1, int numFrames) {
|
||||
|
||||
gain0 *= (1/32768.0f); // int16_t to float
|
||||
gain1 *= (1/32768.0f);
|
||||
|
||||
for (int i = 0; i < numFrames; i++) {
|
||||
|
||||
float frac = win[i];
|
||||
float gain = gain1 + frac * (gain0 - gain1);
|
||||
|
||||
float x0 = (float)src[i] * gain;
|
||||
|
||||
dst[2*i+0] += x0;
|
||||
dst[2*i+1] += x0;
|
||||
}
|
||||
}
|
||||
|
||||
// apply gain crossfade with accumulation (interleaved)
|
||||
static void gainfade_2x2(int16_t* src, float* dst, const float* win, float gain0, float gain1, int numFrames) {
|
||||
|
||||
gain0 *= (1/32768.0f); // int16_t to float
|
||||
gain1 *= (1/32768.0f);
|
||||
|
||||
for (int i = 0; i < numFrames; i++) {
|
||||
|
||||
float frac = win[i];
|
||||
float gain = gain1 + frac * (gain0 - gain1);
|
||||
|
||||
float x0 = (float)src[2*i+0] * gain;
|
||||
float x1 = (float)src[2*i+1] * gain;
|
||||
|
||||
dst[2*i+0] += x0;
|
||||
dst[2*i+1] += x1;
|
||||
}
|
||||
}
|
||||
|
||||
// design a 2nd order Thiran allpass
|
||||
static void ThiranBiquad(float f, float& b0, float& b1, float& b2, float& a1, float& a2) {
|
||||
|
||||
|
|
@ -1104,6 +1141,13 @@ void AudioHRTF::render(int16_t* input, float* output, int index, float azimuth,
|
|||
// apply global and local gain adjustment
|
||||
gain *= _gainAdjust;
|
||||
|
||||
// disable interpolation from reset state
|
||||
if (_resetState) {
|
||||
_azimuthState = azimuth;
|
||||
_distanceState = distance;
|
||||
_gainState = gain;
|
||||
}
|
||||
|
||||
// to avoid polluting the cache, old filters are recomputed instead of stored
|
||||
setFilters(firCoef, bqCoef, delay, index, _azimuthState, _distanceState, _gainState, L0);
|
||||
|
||||
|
|
@ -1175,3 +1219,45 @@ void AudioHRTF::render(int16_t* input, float* output, int index, float azimuth,
|
|||
|
||||
_resetState = false;
|
||||
}
|
||||
|
||||
void AudioHRTF::mixMono(int16_t* input, float* output, float gain, int numFrames) {
|
||||
|
||||
assert(numFrames == HRTF_BLOCK);
|
||||
|
||||
// apply global and local gain adjustment
|
||||
gain *= _gainAdjust;
|
||||
|
||||
// disable interpolation from reset state
|
||||
if (_resetState) {
|
||||
_gainState = gain;
|
||||
}
|
||||
|
||||
// crossfade gain and accumulate
|
||||
gainfade_1x2(input, output, crossfadeTable, _gainState, gain, HRTF_BLOCK);
|
||||
|
||||
// new parameters become old
|
||||
_gainState = gain;
|
||||
|
||||
_resetState = false;
|
||||
}
|
||||
|
||||
void AudioHRTF::mixStereo(int16_t* input, float* output, float gain, int numFrames) {
|
||||
|
||||
assert(numFrames == HRTF_BLOCK);
|
||||
|
||||
// apply global and local gain adjustment
|
||||
gain *= _gainAdjust;
|
||||
|
||||
// disable interpolation from reset state
|
||||
if (_resetState) {
|
||||
_gainState = gain;
|
||||
}
|
||||
|
||||
// crossfade gain and accumulate
|
||||
gainfade_2x2(input, output, crossfadeTable, _gainState, gain, HRTF_BLOCK);
|
||||
|
||||
// new parameters become old
|
||||
_gainState = gain;
|
||||
|
||||
_resetState = false;
|
||||
}
|
||||
|
|
|
|||
|
|
@ -50,6 +50,12 @@ public:
|
|||
//
|
||||
void render(int16_t* input, float* output, int index, float azimuth, float distance, float gain, int numFrames);
|
||||
|
||||
//
|
||||
// Non-spatialized direct mix (accumulates into existing output)
|
||||
//
|
||||
void mixMono(int16_t* input, float* output, float gain, int numFrames);
|
||||
void mixStereo(int16_t* input, float* output, float gain, int numFrames);
|
||||
|
||||
//
|
||||
// Fast path when input is known to be silent and state as been flushed
|
||||
//
|
||||
|
|
|
|||
|
|
@ -1289,14 +1289,16 @@ void convertInput_AVX2(int16_t* src, float *dst[4], float gain, int numFrames) {
|
|||
|
||||
#endif
|
||||
|
||||
// in-place rotation of the soundfield
|
||||
// crossfade between old and new rotation, to prevent artifacts
|
||||
void rotate_3x3_AVX2(float* buf[4], const float m0[3][3], const float m1[3][3], const float* win, int numFrames) {
|
||||
// in-place rotation and scaling of the soundfield
|
||||
// crossfade between old and new matrix, to prevent artifacts
|
||||
void rotate_4x4_AVX2(float* buf[4], const float m0[4][4], const float m1[4][4], const float* win, int numFrames) {
|
||||
|
||||
const float md[3][3] = {
|
||||
{ m0[0][0] - m1[0][0], m0[0][1] - m1[0][1], m0[0][2] - m1[0][2] },
|
||||
{ m0[1][0] - m1[1][0], m0[1][1] - m1[1][1], m0[1][2] - m1[1][2] },
|
||||
{ m0[2][0] - m1[2][0], m0[2][1] - m1[2][1], m0[2][2] - m1[2][2] },
|
||||
// matrix difference
|
||||
const float md[4][4] = {
|
||||
{ m0[0][0] - m1[0][0], m0[0][1] - m1[0][1], m0[0][2] - m1[0][2], m0[0][3] - m1[0][3] },
|
||||
{ m0[1][0] - m1[1][0], m0[1][1] - m1[1][1], m0[1][2] - m1[1][2], m0[1][3] - m1[1][3] },
|
||||
{ m0[2][0] - m1[2][0], m0[2][1] - m1[2][1], m0[2][2] - m1[2][2], m0[2][3] - m1[2][3] },
|
||||
{ m0[3][0] - m1[3][0], m0[3][1] - m1[3][1], m0[3][2] - m1[3][2], m0[3][3] - m1[3][3] },
|
||||
};
|
||||
|
||||
assert(numFrames % 8 == 0);
|
||||
|
|
@ -1307,30 +1309,35 @@ void rotate_3x3_AVX2(float* buf[4], const float m0[3][3], const float m1[3][3],
|
|||
|
||||
// interpolate the matrix
|
||||
__m256 m00 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[0][0]), _mm256_broadcast_ss(&m1[0][0]));
|
||||
__m256 m10 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[1][0]), _mm256_broadcast_ss(&m1[1][0]));
|
||||
__m256 m20 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[2][0]), _mm256_broadcast_ss(&m1[2][0]));
|
||||
|
||||
__m256 m01 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[0][1]), _mm256_broadcast_ss(&m1[0][1]));
|
||||
__m256 m11 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[1][1]), _mm256_broadcast_ss(&m1[1][1]));
|
||||
__m256 m21 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[2][1]), _mm256_broadcast_ss(&m1[2][1]));
|
||||
__m256 m31 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[3][1]), _mm256_broadcast_ss(&m1[3][1]));
|
||||
|
||||
__m256 m02 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[0][2]), _mm256_broadcast_ss(&m1[0][2]));
|
||||
__m256 m12 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[1][2]), _mm256_broadcast_ss(&m1[1][2]));
|
||||
__m256 m22 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[2][2]), _mm256_broadcast_ss(&m1[2][2]));
|
||||
__m256 m32 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[3][2]), _mm256_broadcast_ss(&m1[3][2]));
|
||||
|
||||
__m256 m13 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[1][3]), _mm256_broadcast_ss(&m1[1][3]));
|
||||
__m256 m23 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[2][3]), _mm256_broadcast_ss(&m1[2][3]));
|
||||
__m256 m33 = _mm256_fmadd_ps(frac, _mm256_broadcast_ss(&md[3][3]), _mm256_broadcast_ss(&m1[3][3]));
|
||||
|
||||
// matrix multiply
|
||||
__m256 x = _mm256_mul_ps(m00, _mm256_loadu_ps(&buf[1][i]));
|
||||
__m256 y = _mm256_mul_ps(m10, _mm256_loadu_ps(&buf[1][i]));
|
||||
__m256 z = _mm256_mul_ps(m20, _mm256_loadu_ps(&buf[1][i]));
|
||||
__m256 w = _mm256_mul_ps(m00, _mm256_loadu_ps(&buf[0][i]));
|
||||
|
||||
x = _mm256_fmadd_ps(m01, _mm256_loadu_ps(&buf[2][i]), x);
|
||||
y = _mm256_fmadd_ps(m11, _mm256_loadu_ps(&buf[2][i]), y);
|
||||
z = _mm256_fmadd_ps(m21, _mm256_loadu_ps(&buf[2][i]), z);
|
||||
__m256 x = _mm256_mul_ps(m11, _mm256_loadu_ps(&buf[1][i]));
|
||||
__m256 y = _mm256_mul_ps(m21, _mm256_loadu_ps(&buf[1][i]));
|
||||
__m256 z = _mm256_mul_ps(m31, _mm256_loadu_ps(&buf[1][i]));
|
||||
|
||||
x = _mm256_fmadd_ps(m02, _mm256_loadu_ps(&buf[3][i]), x);
|
||||
y = _mm256_fmadd_ps(m12, _mm256_loadu_ps(&buf[3][i]), y);
|
||||
z = _mm256_fmadd_ps(m22, _mm256_loadu_ps(&buf[3][i]), z);
|
||||
x = _mm256_fmadd_ps(m12, _mm256_loadu_ps(&buf[2][i]), x);
|
||||
y = _mm256_fmadd_ps(m22, _mm256_loadu_ps(&buf[2][i]), y);
|
||||
z = _mm256_fmadd_ps(m32, _mm256_loadu_ps(&buf[2][i]), z);
|
||||
|
||||
x = _mm256_fmadd_ps(m13, _mm256_loadu_ps(&buf[3][i]), x);
|
||||
y = _mm256_fmadd_ps(m23, _mm256_loadu_ps(&buf[3][i]), y);
|
||||
z = _mm256_fmadd_ps(m33, _mm256_loadu_ps(&buf[3][i]), z);
|
||||
|
||||
_mm256_storeu_ps(&buf[0][i], w);
|
||||
_mm256_storeu_ps(&buf[1][i], x);
|
||||
_mm256_storeu_ps(&buf[2][i], y);
|
||||
_mm256_storeu_ps(&buf[3][i], z);
|
||||
|
|
|
|||
Loading…
Reference in a new issue