Merge branch 'ROCm:develop' into active_set_APIs

src/ipc/backend_ipc.hpp

@@ -32,10 +32,6 @@
 #include "ipc_context_proxy.hpp"
 #include "../ipc_policy.hpp"
 
-#ifndef USE_COOPERATIVE_GROUPS
-#include "../memory/notifier.hpp"
-#endif /* NOT DEFINED: USE_COOPERATIVE_GROUPS */
-
 namespace rocshmem {
 
 class IPCBackend : public Backend {
@@ -172,10 +168,6 @@ class IPCBackend : public Backend {
   */
   int *fence_pool{nullptr};
 
-#ifndef USE_COOPERATIVE_GROUPS
-  NotifierProxy<HIPAllocator, detail::atomic::memory_scope_agent> notifier_{};
-#endif /* NOT DEFINED: USE_COOPERATIVE_GROUPS */
-
  protected:
    /**
    * @copydoc Backend::dump_backend_stats()

src/ipc/context_ipc_device.cpp

@@ -51,10 +51,6 @@ __host__ IPCContext::IPCContext(Backend *b)
   Wrk_Sync_buffer_bases_ = backend->get_wrk_sync_bases();
 
   orders_.store = detail::atomic::rocshmem_memory_order::memory_order_seq_cst;
-
-#ifndef USE_COOPERATIVE_GROUPS
-  notifier_ = backend->notifier_.get();
-#endif /* NOT DEFINED: USE_COOPERATIVE_GROUPS */
 }
 
 __device__ void IPCContext::threadfence_system() {

src/ipc/context_ipc_device.hpp

@@ -27,13 +27,6 @@
 #include "../atomic.hpp"
 #include "../team.hpp"
 
-#ifdef USE_COOPERATIVE_GROUPS
-#include <hip/hip_cooperative_groups.h>
-namespace cg = cooperative_groups;
-#else
-#include "../memory/notifier.hpp"
-#endif /* USE_COOPERATIVE_GROUPS */
-
 namespace rocshmem {
 
 class IPCContext : public Context {
@@ -273,10 +266,6 @@ class IPCContext : public Context {
   //Temporary scratchpad memory used by internal barrier algorithms.
   int64_t *barrier_sync{nullptr};
 
-#ifndef USE_COOPERATIVE_GROUPS
-  Notifier<detail::atomic::memory_scope_agent> *notifier_{nullptr};
-#endif /* NOT DEFINED: USE_COOPERATIVE_GROUPS */
-
   //Struct defining memory ordering for atomic operations.
   detail::atomic::rocshmem_memory_orders orders_{};
 

src/ipc/context_ipc_device_coll.cpp

@@ -87,27 +87,15 @@ __device__ void IPCContext::internal_atomic_barrier(int pe, int PE_start,
 // Uses PE values that are relative to world
 __device__ void IPCContext::internal_sync(int pe, int PE_start, int stride,
                                           int PE_size, int64_t *pSync) {
-#ifdef USE_COOPERATIVE_GROUPS
-  cg::grid_group grid = cg::this_grid();
-  grid.sync();
-  if (0 == grid.thread_rank())
-#else
-  notifier_->sync();
-  if (0 == get_flat_id())
-#endif /* USE_COOPERATIVE_GROUPS */
-  {
+  __syncthreads();
+  if (is_thread_zero_in_block()) {
     if (PE_size < 64) {
       internal_direct_barrier(pe, PE_start, stride, PE_size, pSync);
     } else {
       internal_atomic_barrier(pe, PE_start, stride, PE_size, pSync);
     }
   }
-  __threadfence();
-#ifdef USE_COOPERATIVE_GROUPS
-  grid.sync();
-#else
-  notifier_->sync();
-#endif /* USE_COOPERATIVE_GROUPS */
+  __syncthreads();
 }
 
 __device__ void IPCContext::sync(roc_shmem_team_t team) {