Skip to content

[SYCL][CUDA][PI] Fix infinite loop when parallel_for range exceeds INT_MAX #5095

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
Dec 9, 2021
Merged

[SYCL][CUDA][PI] Fix infinite loop when parallel_for range exceeds INT_MAX #5095

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
08edd13
[SYCL][CUDA][PI] Fix infinite loop when parallel_for range exceeds IN…
pgorlani Dec 6, 2021
7cba4a5
remove unnecessary parentheses
pgorlani Dec 7, 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
18 changes: 8 additions & 10 deletions sycl/plugins/cuda/pi_cuda.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -245,7 +245,7 @@ int getAttribute(pi_device device, CUdevice_attribute attribute) {
// Determine local work sizes that result in uniform work groups.
// The default threadsPerBlock only require handling the first work_dim
// dimension.
void guessLocalWorkSize(int *threadsPerBlock, const size_t *global_work_size,
void guessLocalWorkSize(size_t *threadsPerBlock, const size_t *global_work_size,
const size_t maxThreadsPerBlock[3], pi_kernel kernel,
pi_uint32 local_size) {
assert(threadsPerBlock != nullptr);
Expand All @@ -259,10 +259,9 @@ void guessLocalWorkSize(int *threadsPerBlock, const size_t *global_work_size,

(void)minGrid; // Not used, avoid warnings

threadsPerBlock[0] =
std::min(static_cast<int>(maxThreadsPerBlock[0]),
std::min(static_cast<int>(global_work_size[0]),
static_cast<int>(recommendedBlockSize)));
threadsPerBlock[0] = std::min(
maxThreadsPerBlock[0],
std::min(global_work_size[0], static_cast<size_t>(recommendedBlockSize)));

// Find a local work group size that is a divisor of the global
// work group size to produce uniform work groups.
Expand Down Expand Up @@ -2610,7 +2609,7 @@ pi_result cuda_piEnqueueKernelLaunch(

// Set the number of threads per block to the number of threads per warp
// by default unless user has provided a better number
int threadsPerBlock[3] = {32, 1, 1};
size_t threadsPerBlock[3] = {32u, 1u, 1u};
size_t maxWorkGroupSize = 0u;
size_t maxThreadsPerBlock[3] = {};
bool providedLocalWorkGroupSize = (local_work_size != nullptr);
Expand Down Expand Up @@ -2641,7 +2640,7 @@ pi_result cuda_piEnqueueKernelLaunch(
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]);
threadsPerBlock[dim] = local_work_size[dim];
return PI_SUCCESS;
};

Expand All @@ -2661,12 +2660,11 @@ pi_result cuda_piEnqueueKernelLaunch(
return PI_INVALID_WORK_GROUP_SIZE;
}

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

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

std::unique_ptr<_pi_event> retImplEv{nullptr};
Expand Down