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
src/reverse_offload/context_ro_device.cpp
+29 -1
View File
@@ -170,6 +170,20 @@ __device__ void *ROContext::shmem_ptr(const void *dest, int pe) {
}
__device__ void ROContext::barrier_all() {
build_queue_element(RO_NET_BARRIER, nullptr, nullptr, 0, 0, 0, 0, 0,
nullptr, nullptr, (MPI_Comm)NULL, ro_net_win_id,
block_handle, true, get_status_flag(), is_default_ctx);
}
__device__ void ROContext::barrier_all_wave() {
if (is_thread_zero_in_wave()) {
build_queue_element(RO_NET_BARRIER, nullptr, nullptr, 0, 0, 0, 0, 0,
nullptr, nullptr, (MPI_Comm)NULL, ro_net_win_id,
block_handle, true, get_status_flag(), is_default_ctx);
}
}
__device__ void ROContext::barrier_all_wg() {
if (is_thread_zero_in_block()) {
build_queue_element(RO_NET_BARRIER, nullptr, nullptr, 0, 0, 0, 0, 0,
nullptr, nullptr, (MPI_Comm)NULL, ro_net_win_id,
@@ -189,6 +203,20 @@ __device__ void ROContext::barrier(rocshmem_team_t team) {
}
__device__ void ROContext::sync_all() {
build_queue_element(RO_NET_SYNC, nullptr, nullptr, 0, 0, 0, 0, 0,
nullptr, nullptr, (MPI_Comm)NULL, ro_net_win_id,
block_handle, true, get_status_flag(), is_default_ctx);
}
__device__ void ROContext::sync_all_wave() {
if (is_thread_zero_in_wave()) {
build_queue_element(RO_NET_SYNC, nullptr, nullptr, 0, 0, 0, 0, 0,
nullptr, nullptr, (MPI_Comm)NULL, ro_net_win_id,
block_handle, true, get_status_flag(), is_default_ctx);
}
}
__device__ void ROContext::sync_all_wg() {
if (is_thread_zero_in_block()) {
build_queue_element(RO_NET_SYNC, nullptr, nullptr, 0, 0, 0, 0, 0,
nullptr, nullptr, (MPI_Comm)NULL, ro_net_win_id,
@@ -197,7 +225,7 @@ __device__ void ROContext::sync_all() {
__syncthreads();
}
__device__ void ROContext::sync(rocshmem_team_t team) {
__device__ void ROContext::sync_wg(rocshmem_team_t team) {
ROTeam *team_obj = reinterpret_cast<ROTeam *>(team);
if (is_thread_zero_in_block()) {
build_queue_element(RO_NET_SYNC, nullptr, nullptr, 0, 0, 0, 0, 0, nullptr,
src/reverse_offload/context_ro_device.hpp
+9 -1
View File
@@ -65,11 +65,19 @@ class ROContext : public Context {
__device__ void barrier_all();
__device__ void barrier_all_wave();
__device__ void barrier_all_wg();
__device__ void barrier(rocshmem_team_t team);
__device__ void sync_all();
__device__ void sync(rocshmem_team_t team);
__device__ void sync_all_wave();
__device__ void sync_all_wg();
__device__ void sync_wg(rocshmem_team_t team);
template <typename T>
__device__ void p(T *dest, T value, int pe);