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

sgemm_wg: ifdef-guard cluster specific code

Browse Source
This commit is contained in:
Hansung Kim
2024-03-27 22:45:51 -07:00
parent 09822764e7
commit 9555b790e7
1 changed files with 12 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
tests/regression/sgemm_wg/kernel.cpp
View File

@@ -130,8 +130,13 @@ void kernel_body(int task_id, kernel_arg_t* __UNIFORM__ arg) {
// across the threadblock // across the threadblock
const uint32_t threads_per_threadblock = (BM * BN) / (TM * TN); const uint32_t threads_per_threadblock = (BM * BN) / (TM * TN);
#ifdef RADIANCE
const uint32_t threadblocks_per_core =
vx_num_threads() * vx_num_warps() / (threads_per_threadblock / CORES_PER_CLUSTER);
#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;
#endif
const uint32_t threadblock_dim_x = vx_num_threads(); const uint32_t threadblock_dim_x = vx_num_threads();
const uint32_t threadblock_dim_y = vx_num_warps() / threadblocks_per_core; const uint32_t threadblock_dim_y = vx_num_warps() / threadblocks_per_core;
const int threadblock_id = task_id / threads_per_threadblock; const int threadblock_id = task_id / threads_per_threadblock;
@@ -156,6 +161,12 @@ void kernel_body(int task_id, kernel_arg_t* __UNIFORM__ arg) {
int main() { int main() {
kernel_arg_t *arg = (kernel_arg_t *)KERNEL_ARG_DEV_MEM_ADDR; kernel_arg_t *arg = (kernel_arg_t *)KERNEL_ARG_DEV_MEM_ADDR;
const uint32_t grid_size = arg->dim_m * arg->dim_n / (TM * TN); const uint32_t grid_size = arg->dim_m * arg->dim_n / (TM * TN);
vx_spawn_tasks(grid_size, (vx_spawn_tasks_cb)kernel_body, arg); #ifdef RADIANCE
vx_spawn_tasks_cluster(grid_size, (vx_spawn_tasks_cb)kernel_body, arg);
#else
// NOTE: This kernel assumes contiguous thread scheduling for threadblock
// allocation, and therefore does not work with original vx_spawn_tasks
vx_spawn_tasks_contiguous(grid_size, (vx_spawn_tasks_cb)kernel_body, arg);
#endif
return 0; return 0;
} }