Skip to content

Commit

Permalink
DCT-I and odd-length DCT-IV support
Browse files Browse the repository at this point in the history 
-Added support for the missing base Discrete Cosine Transforms.
-Fixed DCT-IV issues in OpenCL backend
-Improved coalescing logic for DCT and Bluestein's algorithm
-bugfixes
  • Loading branch information
@DTolm
DTolm committed Sep 21, 2021
1 parent 1af6cf7 commit 320ce4e
Showing 1 changed file with 8 additions and 8 deletions.
16 changes: 8 additions & 8 deletions vkFFT/vkFFT.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -10068,7 +10068,7 @@ static inline VkFFTResult appendRadixStageNonStrided(VkFFTSpecializationConstant
uint64_t logicalStoragePerThread = sc->registers_per_thread_per_radix[stageRadix] * sc->registerBoost;// (sc->registers_per_thread % stageRadix == 0) ? sc->registers_per_thread * sc->registerBoost : sc->min_registers_per_thread * sc->registerBoost;
uint64_t logicalRegistersPerThread = sc->registers_per_thread_per_radix[stageRadix];// (sc->registers_per_thread % stageRadix == 0) ? sc->registers_per_thread : sc->min_registers_per_thread;
uint64_t logicalGroupSize = sc->fftDim / logicalStoragePerThread;
if ((!((sc->readToRegisters == 1) && (stageSize == 1) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle > 0))))) && ((sc->localSize[0] * logicalStoragePerThread > sc->fftDim) || (stageSize > 1) || ((sc->localSize[1] > 1) && (!(sc->performR2C && (sc->actualInverse)))) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle > 0)) || (sc->performDCT)))
if ((!((sc->readToRegisters == 1) && (stageSize == 1) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle > 0) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)))))) && ((sc->localSize[0] * logicalStoragePerThread > sc->fftDim) || (stageSize > 1) || ((sc->localSize[1] > 1) && (!(sc->performR2C && (sc->actualInverse)))) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle > 0)) || (sc->performDCT)))
{
res = appendBarrierVkFFT(sc, 1);
if (res != VKFFT_SUCCESS) return res;
Expand Down Expand Up @@ -10096,7 +10096,7 @@ static inline VkFFTResult appendRadixStageNonStrided(VkFFTSpecializationConstant
sc->tempLen = sprintf(sc->tempStr, " angle = stageInvocationID * %.17f%s;\n", stageAngle, LFending);
res = VkAppendLine(sc);
if (res != VKFFT_SUCCESS) return res;
if ((!((sc->readToRegisters == 1) && (stageSize==1) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle > 0))))) && ((sc->registerBoost == 1) && ((sc->localSize[0] * logicalStoragePerThread > sc->fftDim) || (stageSize > 1) || ((sc->localSize[1] > 1) && (!(sc->performR2C && (sc->actualInverse)))) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle > 0)) || (sc->performDCT)))) {
if ((!((sc->readToRegisters == 1) && (stageSize==1) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle > 0) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)))))) && ((sc->registerBoost == 1) && ((sc->localSize[0] * logicalStoragePerThread > sc->fftDim) || (stageSize > 1) || ((sc->localSize[1] > 1) && (!(sc->performR2C && (sc->actualInverse)))) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle > 0)) || (sc->performDCT)))) {
//if(sc->readToRegisters==0){
for (uint64_t i = 0; i < stageRadix; i++) {
uint64_t id = j + i * logicalRegistersPerThread / stageRadix;
Expand Down Expand Up @@ -10228,7 +10228,7 @@ static inline VkFFTResult appendRadixStageStrided(VkFFTSpecializationConstantsLa
uint64_t logicalStoragePerThread = sc->registers_per_thread_per_radix[stageRadix] * sc->registerBoost;// (sc->registers_per_thread % stageRadix == 0) ? sc->registers_per_thread * sc->registerBoost : sc->min_registers_per_thread * sc->registerBoost;
uint64_t logicalRegistersPerThread = sc->registers_per_thread_per_radix[stageRadix];// (sc->registers_per_thread % stageRadix == 0) ? sc->registers_per_thread : sc->min_registers_per_thread;
uint64_t logicalGroupSize = sc->fftDim / logicalStoragePerThread;
if ((!((sc->readToRegisters == 1) && (stageSize == 1) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle > 0))))) && (((sc->axis_id == 0) && (sc->axis_upload_id == 0) && (!(sc->performR2C && (sc->actualInverse)))) || (sc->localSize[1] * logicalStoragePerThread > sc->fftDim) || (stageSize > 1) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle > 0)) || (sc->performDCT)))
if ((!((sc->readToRegisters == 1) && (stageSize == 1) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle > 0) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)))))) && (((sc->axis_id == 0) && (sc->axis_upload_id == 0) && (!(sc->performR2C && (sc->actualInverse)))) || (sc->localSize[1] * logicalStoragePerThread > sc->fftDim) || (stageSize > 1) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle > 0)) || (sc->performDCT)))
{
res = appendBarrierVkFFT(sc, 1);
if (res != VKFFT_SUCCESS) return res;
Expand All @@ -10255,7 +10255,7 @@ static inline VkFFTResult appendRadixStageStrided(VkFFTSpecializationConstantsLa
sc->tempLen = sprintf(sc->tempStr, " angle = stageInvocationID * %.17f%s;\n", stageAngle, LFending);
res = VkAppendLine(sc);
if (res != VKFFT_SUCCESS) return res;
if ((!((sc->readToRegisters == 1) && (stageSize == 1) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle > 0))))) && ((sc->registerBoost == 1) && (((sc->axis_id == 0) && (sc->axis_upload_id == 0) && (!(sc->performR2C && (sc->actualInverse)))) || (sc->localSize[1] * logicalStoragePerThread > sc->fftDim) || (stageSize > 1) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle > 0)) || (sc->performDCT)))) {
if ((!((sc->readToRegisters == 1) && (stageSize == 1) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle > 0) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)))))) && ((sc->registerBoost == 1) && (((sc->axis_id == 0) && (sc->axis_upload_id == 0) && (!(sc->performR2C && (sc->actualInverse)))) || (sc->localSize[1] * logicalStoragePerThread > sc->fftDim) || (stageSize > 1) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle > 0)) || (sc->performDCT)))) {
for (uint64_t i = 0; i < stageRadix; i++) {
uint64_t id = j + i * logicalRegistersPerThread / stageRadix;
id = (id / logicalRegistersPerThread) * sc->registers_per_thread + id % logicalRegistersPerThread;
Expand Down Expand Up @@ -10414,13 +10414,13 @@ static inline VkFFTResult appendRadixShuffleNonStrided(VkFFTSpecializationConsta

uint64_t logicalGroupSize = sc->fftDim / logicalStoragePerThread;
uint64_t logicalGroupSizeNext = sc->fftDim / logicalStoragePerThreadNext;
if ((!((sc->writeFromRegisters == 1) && (stageSize == sc->fftDim / stageRadix) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep))&&(stageAngle<0))))) && (((sc->registerBoost == 1) && ((sc->localSize[0] * logicalStoragePerThread > sc->fftDim) || (stageSize < sc->fftDim / stageRadix) || ((sc->reorderFourStep) && (sc->fftDim < sc->fft_dim_full) && (sc->localSize[1] > 1)) || (sc->localSize[1] > 1) || ((sc->performR2C) && (!sc->actualInverse) && (sc->axis_id == 0)) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle < 0)))) || (sc->performDCT)))
if ((!((sc->writeFromRegisters == 1) && (stageSize == sc->fftDim / stageRadix) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep))&&(stageAngle<0) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)))))) && (((sc->registerBoost == 1) && ((sc->localSize[0] * logicalStoragePerThread > sc->fftDim) || (stageSize < sc->fftDim / stageRadix) || ((sc->reorderFourStep) && (sc->fftDim < sc->fft_dim_full) && (sc->localSize[1] > 1)) || (sc->localSize[1] > 1) || ((sc->performR2C) && (!sc->actualInverse) && (sc->axis_id == 0)) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle < 0)))) || (sc->performDCT)))
{
res = appendBarrierVkFFT(sc, 1);
if (res != VKFFT_SUCCESS) return res;
}
//if ((sc->localSize[0] * logicalStoragePerThread > sc->fftDim) || (stageSize < sc->fftDim / stageRadix) || ((sc->reorderFourStep) && (sc->fftDim < sc->fft_dim_full) && (sc->localSize[1] > 1)) || (sc->localSize[1] > 1) || ((sc->performR2C) && (!sc->actualInverse) && (sc->axis_id == 0)) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle < 0)) || (sc->registerBoost > 1) || (sc->performDCT)) {
if ((!((sc->writeFromRegisters == 1) && (stageSize == sc->fftDim / stageRadix) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle < 0))))) && ((sc->localSize[0] * logicalStoragePerThread > sc->fftDim) || (stageSize < sc->fftDim / stageRadix) || ((sc->reorderFourStep) && (sc->fftDim < sc->fft_dim_full) && (sc->localSize[1] > 1)) || (sc->localSize[1] > 1) || ((sc->performR2C) && (!sc->actualInverse) && (sc->axis_id == 0)) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle < 0)) || (sc->registerBoost > 1) || (sc->performDCT))) {
if ((!((sc->writeFromRegisters == 1) && (stageSize == sc->fftDim / stageRadix) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle < 0) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)))))) && ((sc->localSize[0] * logicalStoragePerThread > sc->fftDim) || (stageSize < sc->fftDim / stageRadix) || ((sc->reorderFourStep) && (sc->fftDim < sc->fft_dim_full) && (sc->localSize[1] > 1)) || (sc->localSize[1] > 1) || ((sc->performR2C) && (!sc->actualInverse) && (sc->axis_id == 0)) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle < 0)) || (sc->registerBoost > 1) || (sc->performDCT))) {
if (!((sc->registerBoost > 1) && (stageSize * stageRadix == sc->fftDim / sc->stageRadix[sc->numStages - 1]) && (sc->stageRadix[sc->numStages - 1] == sc->registerBoost))) {
char** tempID;
tempID = (char**)malloc(sizeof(char*) * sc->registers_per_thread * sc->registerBoost);
Expand Down Expand Up @@ -10785,7 +10785,7 @@ static inline VkFFTResult appendRadixShuffleStrided(VkFFTSpecializationConstants
if (normalizationValue != 1) {
sprintf(stageNormalization, "%.17f%s", 1.0 / (double)(normalizationValue), LFending);
}
if ((!((sc->writeFromRegisters == 1) && (stageSize == sc->fftDim / stageRadix) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle < 0))))) && (((sc->axis_id == 0) && (sc->axis_upload_id == 0)) || (sc->localSize[1] * logicalStoragePerThread > sc->fftDim) || (stageSize < sc->fftDim / stageRadix) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle < 0)) || (sc->performDCT)))
if ((!((sc->writeFromRegisters == 1) && (stageSize == sc->fftDim / stageRadix) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle < 0) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)))))) && (((sc->axis_id == 0) && (sc->axis_upload_id == 0)) || (sc->localSize[1] * logicalStoragePerThread > sc->fftDim) || (stageSize < sc->fftDim / stageRadix) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle < 0)) || (sc->performDCT)))
{
res = appendBarrierVkFFT(sc, 2);
if (res != VKFFT_SUCCESS) return res;
Expand All @@ -10795,7 +10795,7 @@ static inline VkFFTResult appendRadixShuffleStrided(VkFFTSpecializationConstants
res = VkAppendLine(sc);
if (res != VKFFT_SUCCESS) return res;
}
if ((!((sc->writeFromRegisters == 1)&&(stageSize == sc->fftDim / stageRadix) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle < 0)))))&&(((sc->axis_id == 0) && (sc->axis_upload_id == 0)) || (sc->localSize[1] * logicalStoragePerThread > sc->fftDim) || (stageSize < sc->fftDim / stageRadix) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle < 0)) || (sc->performDCT))) {
if ((!((sc->writeFromRegisters == 1)&&(stageSize == sc->fftDim / stageRadix) && (!(((sc->convolutionStep) || (sc->useBluesteinFFT && sc->BluesteinConvolutionStep)) && (stageAngle < 0) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1))))))&&(((sc->axis_id == 0) && (sc->axis_upload_id == 0)) || (sc->localSize[1] * logicalStoragePerThread > sc->fftDim) || (stageSize < sc->fftDim / stageRadix) || ((sc->convolutionStep) && ((sc->matrixConvolution > 1) || (sc->numKernels > 1)) && (stageAngle < 0)) || (sc->performDCT))) {
//if (sc->writeFromRegisters == 0) {
//appendBarrierVkFFT(sc, 2);
if (!((sc->registerBoost > 1) && (stageSize * stageRadix == sc->fftDim / sc->stageRadix[sc->numStages - 1]) && (sc->stageRadix[sc->numStages - 1] == sc->registerBoost))) {
Expand Down

0 comments on commit 320ce4e

Please sign in to comment.