wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

sgemm_wg: Fix possible divide-by-0

Browse Source
This commit is contained in:
Hansung Kim
2024-03-28 17:35:47 -07:00
parent 9555b790e7
commit 9673db4e8c
1 changed files with 7 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
tests/regression/sgemm_wg/kernel.cpp
View File

@@ -132,7 +132,7 @@ void kernel_body(int task_id, kernel_arg_t* __UNIFORM__ arg) {
const uint32_t threads_per_threadblock = (BM * BN) / (TM * TN); const uint32_t threads_per_threadblock = (BM * BN) / (TM * TN);
#ifdef RADIANCE #ifdef RADIANCE
const uint32_t threadblocks_per_core = const uint32_t threadblocks_per_core =
vx_num_threads() * vx_num_warps() / (threads_per_threadblock / CORES_PER_CLUSTER); vx_num_threads() * vx_num_warps() / threads_per_threadblock * CORES_PER_CLUSTER;
#else #else
const uint32_t threadblocks_per_core = const uint32_t threadblocks_per_core =
vx_num_threads() * vx_num_warps() / threads_per_threadblock; vx_num_threads() * vx_num_warps() / threads_per_threadblock;
@@ -149,13 +149,13 @@ void kernel_body(int task_id, kernel_arg_t* __UNIFORM__ arg) {
const int threadblock_id_x = threadblock_id % dim_n_in_blocks; const int threadblock_id_x = threadblock_id % dim_n_in_blocks;
const int threadblock_id_y = threadblock_id / dim_n_in_blocks; const int threadblock_id_y = threadblock_id / dim_n_in_blocks;
// "static" shared memory allocation. This would determine threadblock
// occupancy of a single cluster
float *sharedmem_per_threadblock = float *sharedmem_per_threadblock =
(float *)DEV_SMEM_START_ADDR + (float *)DEV_SMEM_START_ADDR + (2 * BM * BK) * threadblock_id_in_core;
(2 * BM * BK) * threadblock_id_in_core; thread_block_gemm(arg, tid_in_threadblock, threadblock_dim_x,
thread_block_gemm(arg, tid_in_threadblock, threadblock_dim_y, threadblock_id_x, threadblock_id_y,
threadblock_dim_x, threadblock_dim_y, threadblock_id_x, threadblock_id_in_core, sharedmem_per_threadblock);
threadblock_id_y, threadblock_id_in_core,
sharedmem_per_threadblock);
} }
int main() { int main() {