zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-08-02 00:36:49 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

no occupancy query for launch params (#2426)

Browse Source
This commit is contained in:
Awni Hannun 2025-07-28 09:09:41 -07:00 committed by GitHub
parent b9e88fb976
commit 1588659062
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 1 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
mlx/backend/cuda/kernel_utils.cuh
View File

@ -120,20 +120,6 @@ dim3 get_2d_grid_dims(
size_t divisor);
std::pair<dim3, dim3> get_grid_and_block(int dim0, int dim1, int dim2);
// Return a block size that achieves maximum potential occupancy for kernel.
template <typename T>
inline uint max_occupancy_block_dim(T kernel) {
int _, block_dim;
if constexpr (std::is_same_v<T, CUfunction>) {
CHECK_CUDA_ERROR(
cuOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel, 0, 0, 0));
} else {
CHECK_CUDA_ERROR(
cudaOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel));
}
return block_dim;
}
// Get the num_blocks and block_dims that maximize occupancy for |kernel|,
// assuming each thread handles |work_per_thread| elements of |arr|.
template <typename T>
@ -145,7 +131,7 @@ inline std::tuple<dim3, uint> get_launch_args(
bool large,
int work_per_thread = 1) {
size_t nthreads = cuda::ceil_div(size, work_per_thread);
uint block_dim = max_occupancy_block_dim(kernel);
uint block_dim = 1024;
if (block_dim > nthreads) {
block_dim = nthreads;
}