Skip to content

[SYCL][CUDA] Fix context scope in kernel launch #4606

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

Merged
merged 2 commits into from
Sep 21, 2021
Merged

[SYCL][CUDA] Fix context scope in kernel launch #4606

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0cbc7af
[SYCL][CUDA] Fix context scope in kernel launch
npmiller Sep 20, 2021
55f2f4f
[SYCL][CUDA] Fix try/catch block
npmiller Sep 21, 2021
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
95 changes: 48 additions & 47 deletions sycl/plugins/cuda/pi_cuda.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2577,61 +2577,62 @@ pi_result cuda_piEnqueueKernelLaunch(
size_t maxThreadsPerBlock[3] = {};
bool providedLocalWorkGroupSize = (local_work_size != nullptr);
pi_uint32 local_size = kernel->get_local_size();
pi_result retError = PI_SUCCESS;

{
size_t *reqdThreadsPerBlock = kernel->reqdThreadsPerBlock_;
maxWorkGroupSize = command_queue->device_->get_max_work_group_size();
command_queue->device_->get_max_work_item_sizes(sizeof(maxThreadsPerBlock),
maxThreadsPerBlock);

if (providedLocalWorkGroupSize) {
auto isValid = [&](int dim) {
if (reqdThreadsPerBlock[dim] != 0 &&
local_work_size[dim] != reqdThreadsPerBlock[dim])
return PI_INVALID_WORK_GROUP_SIZE;

if (local_work_size[dim] > maxThreadsPerBlock[dim])
return PI_INVALID_WORK_ITEM_SIZE;
// Checks that local work sizes are a divisor of the global work sizes
// which includes that the local work sizes are neither larger than the
// global work sizes and not 0.
if (0u == local_work_size[dim])
return PI_INVALID_WORK_GROUP_SIZE;
if (0u != (global_work_size[dim] % local_work_size[dim]))
return PI_INVALID_WORK_GROUP_SIZE;
threadsPerBlock[dim] = static_cast<int>(local_work_size[dim]);
return PI_SUCCESS;
};

for (size_t dim = 0; dim < work_dim; dim++) {
auto err = isValid(dim);
if (err != PI_SUCCESS)
return err;
try {
// Set the active context here as guessLocalWorkSize needs an active context
ScopedContext active(command_queue->get_context());
{
size_t *reqdThreadsPerBlock = kernel->reqdThreadsPerBlock_;
maxWorkGroupSize = command_queue->device_->get_max_work_group_size();
command_queue->device_->get_max_work_item_sizes(
sizeof(maxThreadsPerBlock), maxThreadsPerBlock);

if (providedLocalWorkGroupSize) {
auto isValid = [&](int dim) {
if (reqdThreadsPerBlock[dim] != 0 &&
local_work_size[dim] != reqdThreadsPerBlock[dim])
return PI_INVALID_WORK_GROUP_SIZE;

if (local_work_size[dim] > maxThreadsPerBlock[dim])
return PI_INVALID_WORK_ITEM_SIZE;
// Checks that local work sizes are a divisor of the global work sizes
// which includes that the local work sizes are neither larger than
// the global work sizes and not 0.
if (0u == local_work_size[dim])
return PI_INVALID_WORK_GROUP_SIZE;
if (0u != (global_work_size[dim] % local_work_size[dim]))
return PI_INVALID_WORK_GROUP_SIZE;
threadsPerBlock[dim] = static_cast<int>(local_work_size[dim]);
return PI_SUCCESS;
};

for (size_t dim = 0; dim < work_dim; dim++) {
auto err = isValid(dim);
if (err != PI_SUCCESS)
return err;
}
} else {
guessLocalWorkSize(threadsPerBlock, global_work_size,
maxThreadsPerBlock, kernel, local_size);
}
} else {
guessLocalWorkSize(threadsPerBlock, global_work_size, maxThreadsPerBlock,
kernel, local_size);
}
}

if (maxWorkGroupSize <
size_t(threadsPerBlock[0] * threadsPerBlock[1] * threadsPerBlock[2])) {
return PI_INVALID_WORK_GROUP_SIZE;
}
if (maxWorkGroupSize <
size_t(threadsPerBlock[0] * threadsPerBlock[1] * threadsPerBlock[2])) {
return PI_INVALID_WORK_GROUP_SIZE;
}

int blocksPerGrid[3] = {1, 1, 1};
int blocksPerGrid[3] = {1, 1, 1};

for (size_t i = 0; i < work_dim; i++) {
blocksPerGrid[i] =
static_cast<int>(global_work_size[i] + threadsPerBlock[i] - 1) /
threadsPerBlock[i];
}
for (size_t i = 0; i < work_dim; i++) {
blocksPerGrid[i] =
static_cast<int>(global_work_size[i] + threadsPerBlock[i] - 1) /
threadsPerBlock[i];
}

pi_result retError = PI_SUCCESS;
std::unique_ptr<_pi_event> retImplEv{nullptr};
std::unique_ptr<_pi_event> retImplEv{nullptr};

try {
ScopedContext active(command_queue->get_context());
CUstream cuStream = command_queue->get();
CUfunction cuFunc = kernel->get();

Expand Down