Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Async multistream synchronization for CUDA and HIP backends #164

Open
wants to merge 2 commits into
base: develop
Choose a base branch
from
Open

Async multistream synchronization for CUDA and HIP backends #164

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6926048
async multistream synchronization
DejvBayer Mar 9, 2024
c5c557b
disable CUDA and HIP events timing
DejvBayer Mar 9, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 2 additions & 4 deletions vkFFT/vkFFT/vkFFT_AppManagement/vkFFT_InitializeApp.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -543,7 +543,6 @@ static inline VkFFTResult setConfigurationVkFFT(VkFFTApplication* app, VkFFTConf
app->configuration.device = inputLaunchConfiguration.device;
if (inputLaunchConfiguration.num_streams != 0) app->configuration.num_streams = inputLaunchConfiguration.num_streams;
if (inputLaunchConfiguration.stream != 0) app->configuration.stream = inputLaunchConfiguration.stream;
app->configuration.streamID = 0;
int value = 0;
res = cuDeviceGetAttribute(&value, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, app->configuration.device[0]);
if (res != CUDA_SUCCESS) {
Expand Down Expand Up @@ -635,7 +634,7 @@ static inline VkFFTResult setConfigurationVkFFT(VkFFTApplication* app, VkFFTConf
return VKFFT_ERROR_MALLOC_FAILED;
}
for (pfUINT i = 0; i < app->configuration.num_streams; i++) {
res_t = cudaEventCreate(&app->configuration.stream_event[i]);
res_t = cudaEventCreateWithFlags(&app->configuration.stream_event[i], cudaEventDisableTiming);
if (res_t != cudaSuccess) {
deleteVkFFT(app);
return VKFFT_ERROR_FAILED_TO_CREATE_EVENT;
Expand All @@ -659,7 +658,6 @@ static inline VkFFTResult setConfigurationVkFFT(VkFFTApplication* app, VkFFTConf
app->configuration.device = inputLaunchConfiguration.device;
if (inputLaunchConfiguration.num_streams != 0) app->configuration.num_streams = inputLaunchConfiguration.num_streams;
if (inputLaunchConfiguration.stream != 0) app->configuration.stream = inputLaunchConfiguration.stream;
app->configuration.streamID = 0;
int value = 0;
res = hipDeviceGetAttribute(&value, hipDeviceAttributeComputeCapabilityMajor, app->configuration.device[0]);
if (res != hipSuccess) {
Expand Down Expand Up @@ -740,7 +738,7 @@ static inline VkFFTResult setConfigurationVkFFT(VkFFTApplication* app, VkFFTConf
return VKFFT_ERROR_MALLOC_FAILED;
}
for (pfUINT i = 0; i < app->configuration.num_streams; i++) {
res = hipEventCreate(&app->configuration.stream_event[i]);
res = hipEventCreateWithFlags(&app->configuration.stream_event[i], hipEventDisableTiming);
if (res != hipSuccess) {
deleteVkFFT(app);
return VKFFT_ERROR_FAILED_TO_CREATE_EVENT;
Expand Down
64 changes: 46 additions & 18 deletions vkFFT/vkFFT/vkFFT_AppManagement/vkFFT_RunApp.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,23 +29,7 @@ static inline VkFFTResult VkFFTSync(VkFFTApplication* app) {
#if(VKFFT_BACKEND==0)
vkCmdPipelineBarrier(app->configuration.commandBuffer[0], VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 1, app->configuration.memory_barrier, 0, 0, 0, 0);
#elif(VKFFT_BACKEND==1)
if (app->configuration.num_streams > 1) {
cudaError_t res = cudaSuccess;
for (pfUINT s = 0; s < app->configuration.num_streams; s++) {
res = cudaEventSynchronize(app->configuration.stream_event[s]);
if (res != cudaSuccess) return VKFFT_ERROR_FAILED_TO_SYNCHRONIZE;
}
app->configuration.streamCounter = 0;
}
#elif(VKFFT_BACKEND==2)
if (app->configuration.num_streams > 1) {
hipError_t res = hipSuccess;
for (pfUINT s = 0; s < app->configuration.num_streams; s++) {
res = hipEventSynchronize(app->configuration.stream_event[s]);
if (res != hipSuccess) return VKFFT_ERROR_FAILED_TO_SYNCHRONIZE;
}
app->configuration.streamCounter = 0;
}
#elif(VKFFT_BACKEND==3)
#elif(VKFFT_BACKEND==4)
ze_result_t res = ZE_RESULT_SUCCESS;
Expand Down Expand Up @@ -88,9 +72,7 @@ static inline VkFFTResult VkFFTAppend(VkFFTApplication* app, int inverse, VkFFTL
};
app->configuration.memory_barrier = &memory_barrier;
#elif(VKFFT_BACKEND==1)
app->configuration.streamCounter = 0;
#elif(VKFFT_BACKEND==2)
app->configuration.streamCounter = 0;
#elif(VKFFT_BACKEND==3)
app->configuration.commandQueue = launchParams->commandQueue;
#elif(VKFFT_BACKEND==4)
Expand All @@ -108,6 +90,29 @@ static inline VkFFTResult VkFFTAppend(VkFFTApplication* app, int inverse, VkFFTL
if (resFFT != VKFFT_SUCCESS) {
return resFFT;
}

#if(VKFFT_BACKEND==1)
if (app->configuration.num_streams > 1) {
cudaError_t res = cudaSuccess;
for (pfUINT i = 1; i < app->configuration.num_streams; i++) {
res = cudaEventRecord(app->configuration.stream_event[i], app->configuration.stream[i]);
if (res != cudaSuccess) return VKFFT_ERROR_FAILED_TO_EVENT_RECORD;
res = cudaStreamWaitEvent(app->configuration.stream[0], app->configuration.stream_event[i], cudaEventWaitDefault);
if (res != cudaSuccess) return VKFFT_ERROR_FAILED_TO_SYNCHRONIZE;
}
}
#elif(VKFFT_BACKEND==2)
if (app->configuration.num_streams > 1) {
hipError_t result = hipSuccess;
for (pfUINT i = 1; i < app->configuration.num_streams; i++) {
res = hipEventRecord(app->configuration.stream_event[i], app->configuration.stream[i]);
if (res != hipSuccess) return VKFFT_ERROR_FAILED_TO_EVENT_RECORD;
res = hipStreamWaitEvent(app->configuration.stream[0], app->configuration.stream_event[i], 0);
if (res != hipSuccess) return VKFFT_ERROR_FAILED_TO_SYNCHRONIZE;
}
}
#endif

if (inverse != 1) {
//FFT axis 0
if (!app->configuration.omitDimension[0]) {
Expand Down Expand Up @@ -649,6 +654,29 @@ static inline VkFFTResult VkFFTAppend(VkFFTApplication* app, int inverse, VkFFTL
//if (app->localFFTPlan_inverse->bigSequenceEvenR2C) app->configuration.size[0] *= 2;

}

#if(VKFFT_BACKEND==1)
if (app->configuration.num_streams > 1) {
cudaError_t res = cudaSuccess;
res = cudaEventRecord(app->configuration.stream_event[0], app->configuration.stream[0]);
if (res != cudaSuccess) return VKFFT_ERROR_FAILED_TO_EVENT_RECORD;
for (pfUINT i = 1; i < app->configuration.num_streams; i++) {
res = cudaStreamWaitEvent(app->configuration.stream[i], app->configuration.stream_event[0], cudaEventWaitDefault);
if (res != cudaSuccess) return VKFFT_ERROR_FAILED_TO_SYNCHRONIZE;
}
}
#elif(VKFFT_BACKEND==2)
if (app->configuration.num_streams > 1) {
hipError_t res = hipSuccess;
res = hipEventRecord(app->configuration.stream_event[0], app->configuration.stream[0]);
if (res != hipSuccess) return VKFFT_ERROR_FAILED_TO_EVENT_RECORD;
for (pfUINT i = 1; i < app->configuration.num_streams; i++) {
res = hipStreamWaitEvent(app->configuration.stream[i], app->configuration.stream_event[0], 0);
if (res != hipSuccess) return VKFFT_ERROR_FAILED_TO_SYNCHRONIZE;
}
}
#endif

return resFFT;
}

Expand Down
56 changes: 12 additions & 44 deletions vkFFT/vkFFT/vkFFT_PlanManagement/vkFFT_API_handles/vkFFT_DispatchPlan.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -197,32 +197,16 @@ static inline VkFFTResult VkFFT_DispatchPlan(VkFFTApplication* app, VkFFTAxis* a
}
}
}*/
if (app->configuration.num_streams >= 1) {
result = cuLaunchKernel(axis->VkFFTKernel,
(unsigned int)dispatchSize[0], (unsigned int)dispatchSize[1], (unsigned int)dispatchSize[2], // grid dim
(unsigned int)axis->specializationConstants.localSize[0].data.i, (unsigned int)axis->specializationConstants.localSize[1].data.i, (unsigned int)axis->specializationConstants.localSize[2].data.i, // block dim
(unsigned int)axis->specializationConstants.usedSharedMemory.data.i, app->configuration.stream[app->configuration.streamID], // shared mem and stream
args, 0);
}
else {
result = cuLaunchKernel(axis->VkFFTKernel,
(unsigned int)dispatchSize[0], (unsigned int)dispatchSize[1], (unsigned int)dispatchSize[2], // grid dim
(unsigned int)axis->specializationConstants.localSize[0].data.i, (unsigned int)axis->specializationConstants.localSize[1].data.i, (unsigned int)axis->specializationConstants.localSize[2].data.i, // block dim
(unsigned int)axis->specializationConstants.usedSharedMemory.data.i, 0, // shared mem and stream
args, 0);
}
result = cuLaunchKernel(axis->VkFFTKernel,
(unsigned int)dispatchSize[0], (unsigned int)dispatchSize[1], (unsigned int)dispatchSize[2], // grid dim
(unsigned int)axis->specializationConstants.localSize[0].data.i, (unsigned int)axis->specializationConstants.localSize[1].data.i, (unsigned int)axis->specializationConstants.localSize[2].data.i, // block dim
(unsigned int)axis->specializationConstants.usedSharedMemory.data.i, (app->configuration.num_streams > 0) ? app->configuration.stream[0] : 0, // shared mem and stream
args, 0);

if (result != CUDA_SUCCESS) {
printf("cuLaunchKernel error: %d, %" PRIu64 " %" PRIu64 " %" PRIu64 " - %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", result, dispatchSize[0], dispatchSize[1], dispatchSize[2], axis->specializationConstants.localSize[0].data.i, axis->specializationConstants.localSize[1].data.i, axis->specializationConstants.localSize[2].data.i);
return VKFFT_ERROR_FAILED_TO_LAUNCH_KERNEL;
}
if (app->configuration.num_streams > 1) {
app->configuration.streamID = app->configuration.streamCounter % app->configuration.num_streams;
if (app->configuration.streamCounter == 0) {
cudaError_t res2 = cudaEventRecord(app->configuration.stream_event[app->configuration.streamID], app->configuration.stream[app->configuration.streamID]);
if (res2 != cudaSuccess) return VKFFT_ERROR_FAILED_TO_EVENT_RECORD;
}
app->configuration.streamCounter++;
}
#elif(VKFFT_BACKEND==2)
hipError_t result = hipSuccess;
void* args[10];
Expand Down Expand Up @@ -267,32 +251,16 @@ static inline VkFFTResult VkFFT_DispatchPlan(VkFFTApplication* app, VkFFTAxis* a
}
}*/
//printf("%" PRIu64 " %" PRIu64 " %" PRIu64 " %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",maxBlockSize[0], maxBlockSize[1], maxBlockSize[2], axis->specializationConstants.localSize[0], axis->specializationConstants.localSize[1], axis->specializationConstants.localSize[2]);
if (app->configuration.num_streams >= 1) {
result = hipModuleLaunchKernel(axis->VkFFTKernel,
(unsigned int)dispatchSize[0], (unsigned int)dispatchSize[1], (unsigned int)dispatchSize[2], // grid dim
(unsigned int)axis->specializationConstants.localSize[0].data.i, (unsigned int)axis->specializationConstants.localSize[1].data.i, (unsigned int)axis->specializationConstants.localSize[2].data.i, // block dim
(unsigned int)axis->specializationConstants.usedSharedMemory.data.i, app->configuration.stream[app->configuration.streamID], // shared mem and stream
args, 0);
}
else {
result = hipModuleLaunchKernel(axis->VkFFTKernel,
(unsigned int)dispatchSize[0], (unsigned int)dispatchSize[1], (unsigned int)dispatchSize[2], // grid dim
(unsigned int)axis->specializationConstants.localSize[0].data.i, (unsigned int)axis->specializationConstants.localSize[1].data.i, (unsigned int)axis->specializationConstants.localSize[2].data.i, // block dim
(unsigned int)axis->specializationConstants.usedSharedMemory.data.i, 0, // shared mem and stream
args, 0);
}
result = hipModuleLaunchKernel(axis->VkFFTKernel,
(unsigned int)dispatchSize[0], (unsigned int)dispatchSize[1], (unsigned int)dispatchSize[2], // grid dim
(unsigned int)axis->specializationConstants.localSize[0].data.i, (unsigned int)axis->specializationConstants.localSize[1].data.i, (unsigned int)axis->specializationConstants.localSize[2].data.i, // block dim
(unsigned int)axis->specializationConstants.usedSharedMemory.data.i, (app->configuration.num_streams > 0) ? app->configuration.stream[0] : 0, // shared mem and stream
args, 0);

if (result != hipSuccess) {
printf("hipModuleLaunchKernel error: %d, %" PRIu64 " %" PRIu64 " %" PRIu64 " - %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", result, dispatchSize[0], dispatchSize[1], dispatchSize[2], axis->specializationConstants.localSize[0].data.i, axis->specializationConstants.localSize[1].data.i, axis->specializationConstants.localSize[2].data.i);
return VKFFT_ERROR_FAILED_TO_LAUNCH_KERNEL;
}
if (app->configuration.num_streams > 1) {
app->configuration.streamID = app->configuration.streamCounter % app->configuration.num_streams;
if (app->configuration.streamCounter == 0) {
result = hipEventRecord(app->configuration.stream_event[app->configuration.streamID], app->configuration.stream[app->configuration.streamID]);
if (result != hipSuccess) return VKFFT_ERROR_FAILED_TO_EVENT_RECORD;
}
app->configuration.streamCounter++;
}
#elif(VKFFT_BACKEND==3)
cl_int result = CL_SUCCESS;
void* args[10];
Expand Down
4 changes: 0 additions & 4 deletions vkFFT/vkFFT/vkFFT_Structs/vkFFT_Structs.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -310,12 +310,8 @@ typedef struct {
VkMemoryBarrier* memory_barrier;//Filled at app creation
#elif(VKFFT_BACKEND==1)
cudaEvent_t* stream_event;//Filled at app creation
pfUINT streamCounter;//Filled at app creation
pfUINT streamID;//Filled at app creation
#elif(VKFFT_BACKEND==2)
hipEvent_t* stream_event;//Filled at app creation
pfUINT streamCounter;//Filled at app creation
pfUINT streamID;//Filled at app creation
pfINT useStrict32BitAddress; // guarantee 32 bit addresses in bytes instead of number of elements. This results in fewer instructions generated. -1: Disable, 0: Infer based on size, 1: enable. Has no effect with useUint64.
#elif(VKFFT_BACKEND==3)
cl_command_queue* commandQueue;
Expand Down