AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update Barrier_All and Sync_All APIs (#72)

Browse Source 
* Fix deadlock in `rocshmem_ctx_wg_barrier_all` API in IPC conduit by adding per-context pSync buffers and context IDs
  - Added separate pSync buffers for each device context
  - Resolved deadlock when invoking barrier API (`rocshmem_ctx_wg_barrier_all`) concurrently from multiple contexts

* Update barrier_all functional tests for multi-context support

* Add thread, wavefront, and workgroup-level barrier_all APIs in IPC and RO conduits
  - Implemented barrier_all APIs at thread, wavefront, and workgroup granularity
  - Added support in both IPC and RO conduits
  - Updated functional tests to cover all `barrier_all` APIs

* Add thread, wavefront, and workgroup-level sync_all APIs in IPC and RO conduits
  - Implemented sync_all APIs for thread, wavefront, and workgroup scopes
  - Added support into both IPC and RO conduits
  - Added functional tests to cover all `sync_all` APIs
This commit is contained in:
Avinash Kethineedi
2025-04-02 11:58:55 -05:00
committed by GitHub
parent e16ca7a1e3
commit c652f58cef
22 changed files with 508 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/functional_tests/barrier_all_tester.cpp
+24 -13
View File
@@ -30,9 +30,12 @@ using namespace rocshmem;
* DEVICE TEST KERNEL
*****************************************************************************/
__global__ void BarrierAllTest(int loop, int skip, long long int *start_time,
long long int *end_time) {
long long int *end_time, TestType type,
int wf_size) {
__shared__ rocshmem_ctx_t ctx;
int t_id = get_flat_block_id();
int wg_id = get_flat_grid_id();
int wf_id = t_id / wf_size;
rocshmem_wg_init();
rocshmem_wg_ctx_create(ROCSHMEM_CTX_WG_PRIVATE, &ctx);
@@ -42,17 +45,25 @@ __global__ void BarrierAllTest(int loop, int skip, long long int *start_time,
start_time[wg_id] = wall_clock64();
}
__syncthreads();
/**
* The function `rocshmem_ctx_wg_barrier_all` should be called from only
* one group within the grid to avoid unintended behavior.
*/
if (is_block_zero_in_grid()) {
rocshmem_ctx_wg_barrier_all(ctx);
switch (type) {
case BarrierAllTestType:
if(t_id == 0) {
rocshmem_ctx_barrier_all(ctx);
}
break;
case WAVEBarrierAllTestType:
if(wf_id == 0) {
rocshmem_ctx_wave_barrier_all(ctx);
}
break;
case WGBarrierAllTestType:
rocshmem_ctx_wg_barrier_all(ctx);
break;
default:
break;
}
__syncthreads();
}
__syncthreads();
if (hipThreadIdx_x == 0) {
end_time[wg_id] = wall_clock64();
@@ -74,10 +85,10 @@ void BarrierAllTester::launchKernel(dim3 gridSize, dim3 blockSize, int loop,
size_t shared_bytes = 0;
hipLaunchKernelGGL(BarrierAllTest, gridSize, blockSize, shared_bytes, stream,
loop, args.skip, start_time, end_time);
loop, args.skip, start_time, end_time, _type, wf_size);
num_msgs = loop + args.skip;
num_timed_msgs = loop;
num_msgs = (loop + args.skip) * gridSize.x;
num_timed_msgs = loop * gridSize.x;
}
void BarrierAllTester::resetBuffers(uint64_t size) {}