Skip to content

Commit

Permalink
bugfix
Browse files Browse the repository at this point in the history 
-fix to some systems generating shaders with workgroup sizes out of maxComputeWorkGroupSize limits
  • Loading branch information
@DTolm
DTolm committed Jan 21, 2021
1 parent 45d8611 commit 9e1eca1
Showing 1 changed file with 23 additions and 6 deletions.
29 changes: 23 additions & 6 deletions vkFFT/vkFFT.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@ extern "C" {
uint32_t inputBufferStride[3];
uint32_t outputBufferStride[3];
uint32_t maxComputeWorkGroupCount[3]; // maxComputeWorkGroupCount from VkPhysicalDeviceLimits
uint32_t maxComputeWorkGroupSize[3]; // maxComputeWorkGroupCount from VkPhysicalDeviceLimits
uint32_t coordinateFeatures; // C - coordinate, or dimension of features vector. In matrix convolution - size of vector
uint32_t matrixConvolution; //if equal to 2 perform 2x2, if equal to 3 perform 3x3 matrix-vector convolution. Overrides coordinateFeatures

Expand Down Expand Up @@ -92,7 +93,7 @@ extern "C" {
uint32_t halfThreads;
} VkFFTConfiguration;

static VkFFTConfiguration defaultVkFFTConfiguration = { {1,1,1}, {1,1,1}, {1,1,1}, {1,1,1}, {65535,65535,65535},1,1,1,1,1,8,0,{0,0,0},{0,0,0},{0,0,0}, {0,0},0,0,0,0,0,0,0,0,0, 0, 0, 0, 0, 32768, 32768, 32, 1, 1, 0, 1,"shaders/", 32, 0,0,0,0,0, 1,1,1,1,1, 0,0,0,0,0, 0,0,0,0,0,0, 0 };
static VkFFTConfiguration defaultVkFFTConfiguration = { {1,1,1}, {1,1,1}, {1,1,1}, {1,1,1}, {65535,65535,65535},{1024,1024,64}, 1,1,1,1,1,8,0,{0,0,0},{0,0,0},{0,0,0}, {0,0},0,0,0,0,0,0,0,0,0, 0, 0, 0, 0, 32768, 32768, 32, 1, 1, 0, 1,"shaders/", 32, 0,0,0,0,0, 1,1,1,1,1, 0,0,0,0,0, 0,0,0,0,0,0, 0 };

typedef struct {
uint32_t size[3];
Expand Down Expand Up @@ -3654,7 +3655,7 @@ layout(std430, binding = %d) readonly buffer DataLUT {\n\
else
sprintf(output + strlen(output), " outputBlocks[indexOutput(inoutID%s%s) / %d].outputs[indexOutput(inoutID%s%s) %% %d] = %stemp_%d%s;\n", requestCoordinate, requestBatch, sc.outputBufferBlockSize, requestCoordinate, requestBatch, sc.outputBufferBlockSize, convTypeLeft, i, convTypeRight);
*/
sprintf(output + strlen(output), " }n");
sprintf(output + strlen(output), " }");
}
}
}
Expand Down Expand Up @@ -5639,6 +5640,8 @@ layout(std430, binding = %d) readonly buffer DataLUT {\n\
if (axis_upload_id == 0) {
axis->axisBlock[0] = (axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread > 1) ? axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread : 1;
if (axis->axisBlock[0] > maxThreadNum) axis->axisBlock[0] = maxThreadNum;
if (axis->axisBlock[0] > app->configuration.maxComputeWorkGroupSize[0]) axis->axisBlock[0] = app->configuration.maxComputeWorkGroupSize[0];

axis->axisBlock[1] = 1;
axis->axisBlock[2] = 1;
axis->axisBlock[3] = axis->specializationConstants.fftDim;
Expand All @@ -5647,7 +5650,7 @@ layout(std430, binding = %d) readonly buffer DataLUT {\n\
axis->axisBlock[1] = (axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread > 1) ? axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread : 1;

axis->axisBlock[0] = (axis->specializationConstants.stageStartSize > axis->groupedBatch) ? axis->groupedBatch : axis->specializationConstants.stageStartSize;

if (axis->axisBlock[0] > app->configuration.maxComputeWorkGroupSize[0]) axis->axisBlock[0] = app->configuration.maxComputeWorkGroupSize[0];
axis->axisBlock[2] = 1;
axis->axisBlock[3] = axis->specializationConstants.fftDim;
}
Expand All @@ -5656,6 +5659,8 @@ layout(std430, binding = %d) readonly buffer DataLUT {\n\
axis->axisBlock[1] = (axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread > 1) ? axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread : 1;

axis->axisBlock[0] = (app->configuration.size[1] > axis->groupedBatch) ? axis->groupedBatch : app->configuration.size[1];
if (axis->axisBlock[0] > app->configuration.maxComputeWorkGroupSize[0]) axis->axisBlock[0] = app->configuration.maxComputeWorkGroupSize[0];

/*if (axis->axisBlock[0] * axis->axisBlock[1] < 64)
if (app->configuration.size[1] > 64 / axis->axisBlock[1])
axis->axisBlock[0] = 64 / axis->axisBlock[1];
Expand Down Expand Up @@ -6912,20 +6917,22 @@ layout(std430, binding = %d) readonly buffer DataLUT {\n\
if (axis_upload_id == 0) {
axis->axisBlock[0] = (axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread > 1) ? axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread : 1;
if (axis->axisBlock[0] > maxThreadNum) axis->axisBlock[0] = maxThreadNum;
if (axis->axisBlock[0] > app->configuration.maxComputeWorkGroupSize[0]) axis->axisBlock[0] = app->configuration.maxComputeWorkGroupSize[0];
if (app->configuration.reorderFourStep && (FFTPlan->numAxisUploads[axis_id] > 1))
axis->axisBlock[1] = axis->groupedBatch;
else {
axis->axisBlock[1] = (axis->axisBlock[0] < app->configuration.warpSize) ? app->configuration.warpSize / axis->axisBlock[0] : 1;
if (app->configuration.size[1] < axis->axisBlock[1]) axis->axisBlock[1] = app->configuration.size[1];
}
if (axis->axisBlock[1] > app->configuration.maxComputeWorkGroupSize[1]) axis->axisBlock[1] = app->configuration.maxComputeWorkGroupSize[1];
axis->axisBlock[2] = 1;
axis->axisBlock[3] = axis->specializationConstants.fftDim;
}
else {
axis->axisBlock[1] = (axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread > 1) ? axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread : 1;

axis->axisBlock[0] = (axis->specializationConstants.stageStartSize > axis->groupedBatch) ? axis->groupedBatch : axis->specializationConstants.stageStartSize;

if (axis->axisBlock[0] > app->configuration.maxComputeWorkGroupSize[0]) axis->axisBlock[0] = app->configuration.maxComputeWorkGroupSize[0];
axis->axisBlock[2] = 1;
axis->axisBlock[3] = axis->specializationConstants.fftDim;
}
Expand Down Expand Up @@ -6963,6 +6970,7 @@ layout(std430, binding = %d) readonly buffer DataLUT {\n\
else
axis->axisBlock[0] = app->configuration.size[0];*/
}
if (axis->axisBlock[0] > app->configuration.maxComputeWorkGroupSize[0]) axis->axisBlock[0] = app->configuration.maxComputeWorkGroupSize[0];

axis->axisBlock[2] = 1;
axis->axisBlock[3] = axis->specializationConstants.fftDim;
Expand Down Expand Up @@ -7000,6 +7008,8 @@ layout(std430, binding = %d) readonly buffer DataLUT {\n\
else
axis->axisBlock[0] = app->configuration.size[0];*/
}
if (axis->axisBlock[0] > app->configuration.maxComputeWorkGroupSize[0]) axis->axisBlock[0] = app->configuration.maxComputeWorkGroupSize[0];

axis->axisBlock[2] = 1;
axis->axisBlock[3] = axis->specializationConstants.fftDim;
}
Expand All @@ -7009,21 +7019,23 @@ layout(std430, binding = %d) readonly buffer DataLUT {\n\
if (axis_upload_id == 0) {
axis->axisBlock[0] = (axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread > 1) ? axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread : 1;
if (axis->axisBlock[0] > maxThreadNum) axis->axisBlock[0] = maxThreadNum;

if (axis->axisBlock[0] > app->configuration.maxComputeWorkGroupSize[0]) axis->axisBlock[0] = app->configuration.maxComputeWorkGroupSize[0];
if (app->configuration.reorderFourStep && (FFTPlan->numAxisUploads[axis_id] > 1))
axis->axisBlock[1] = axis->groupedBatch;
else {
axis->axisBlock[1] = (axis->axisBlock[0] < app->configuration.warpSize) ? app->configuration.warpSize / axis->axisBlock[0] : 1;
if (app->configuration.size[1] < axis->axisBlock[1]) axis->axisBlock[1] = app->configuration.size[1];
}
if (axis->axisBlock[1] > app->configuration.maxComputeWorkGroupSize[1]) axis->axisBlock[1] = app->configuration.maxComputeWorkGroupSize[1];

axis->axisBlock[2] = 1;
axis->axisBlock[3] = axis->specializationConstants.fftDim;
}
else {
axis->axisBlock[1] = (axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread > 1) ? axis->specializationConstants.fftDim / axis->specializationConstants.min_registers_per_thread : 1;

axis->axisBlock[0] = (axis->specializationConstants.stageStartSize > axis->groupedBatch) ? axis->groupedBatch : axis->specializationConstants.stageStartSize;

if (axis->axisBlock[0] > app->configuration.maxComputeWorkGroupSize[0]) axis->axisBlock[0] = app->configuration.maxComputeWorkGroupSize[0];
axis->axisBlock[2] = 1;
axis->axisBlock[3] = axis->specializationConstants.fftDim;
}
Expand Down Expand Up @@ -7061,6 +7073,7 @@ layout(std430, binding = %d) readonly buffer DataLUT {\n\
else
axis->axisBlock[0] = app->configuration.size[0];*/
}
if (axis->axisBlock[0] > app->configuration.maxComputeWorkGroupSize[0]) axis->axisBlock[0] = app->configuration.maxComputeWorkGroupSize[0];
axis->axisBlock[2] = 1;
axis->axisBlock[3] = axis->specializationConstants.fftDim;

Expand Down Expand Up @@ -7098,6 +7111,7 @@ layout(std430, binding = %d) readonly buffer DataLUT {\n\
else
axis->axisBlock[0] = app->configuration.size[0];*/
}
if (axis->axisBlock[0] > app->configuration.maxComputeWorkGroupSize[0]) axis->axisBlock[0] = app->configuration.maxComputeWorkGroupSize[0];
axis->axisBlock[2] = 1;
axis->axisBlock[3] = axis->specializationConstants.fftDim;

Expand Down Expand Up @@ -7530,6 +7544,9 @@ layout(std430, binding = %d) readonly buffer DataLUT {\n\
app->configuration.maxComputeWorkGroupCount[0] = physicalDeviceProperties.limits.maxComputeWorkGroupCount[0];
app->configuration.maxComputeWorkGroupCount[1] = physicalDeviceProperties.limits.maxComputeWorkGroupCount[1];
app->configuration.maxComputeWorkGroupCount[2] = physicalDeviceProperties.limits.maxComputeWorkGroupCount[2];
app->configuration.maxComputeWorkGroupSize[0] = physicalDeviceProperties.limits.maxComputeWorkGroupSize[0];
app->configuration.maxComputeWorkGroupSize[1] = physicalDeviceProperties.limits.maxComputeWorkGroupSize[1];
app->configuration.maxComputeWorkGroupSize[2] = physicalDeviceProperties.limits.maxComputeWorkGroupSize[2];
if ((physicalDeviceProperties.vendorID == 0x8086) && (!app->configuration.doublePrecision)) app->configuration.halfThreads = 1;
// if ((physicalDeviceProperties.vendorID == 0x8086) && (!app->configuration.doublePrecision)) app->configuration.sharedMemorySize /= 2;//Temporary measure, until L1 overutilization is enabled
app->configuration.sharedMemorySize = physicalDeviceProperties.limits.maxComputeSharedMemorySize;
Expand Down

0 comments on commit 9e1eca1

Please sign in to comment.