Implement rocshmem_pe_quiet() (#282)

Co-authored-by: Aurelien Bouteiller <aurelien.bouteiller@amd.com>

src/rocshmem_gpu.cpp

@@ -107,7 +107,7 @@ __device__ void rocshmem_wg_finalize() {}
 
 
 /******************************************************************************
-* These host APIs use Device side symbol - ROCSHMEM_CTX_DEFAULT so it needs 
+* These host APIs use Device side symbol - ROCSHMEM_CTX_DEFAULT so it needs
 * to stay here to avoid getting pulled into other places in compilation
 ******************************************************************************/
 
@@ -188,6 +188,10 @@ __device__ void rocshmem_quiet() {
   rocshmem_ctx_quiet(ROCSHMEM_CTX_DEFAULT);
 }
 
+__device__ void rocshmem_pe_quiet(const int *target_pes, size_t npes) {
+  rocshmem_ctx_pe_quiet(ROCSHMEM_CTX_DEFAULT, target_pes, npes);
+}
+
 __device__ void rocshmem_threadfence_system() {
   rocshmem_ctx_threadfence_system(ROCSHMEM_CTX_DEFAULT);
 }
@@ -485,6 +489,16 @@ __device__ void rocshmem_ctx_quiet(rocshmem_ctx_t ctx) {
   get_internal_ctx(ctx)->quiet();
 }
 
+__device__ void rocshmem_ctx_pe_quiet(rocshmem_ctx_t ctx, const int *target_pes, size_t npes) {
+  GPU_DPRINTF("Function: %s (ctx=%zd)\n", __FUNC__, ctx.ctx_opaque);
+
+  ContextTy *internal_ctx = get_internal_ctx(ctx);
+
+  for (int i = 0; i < npes;  i++) {
+    internal_ctx->pe_quiet(translate_pe(ctx, target_pes[i]));
+  }
+}
+
 __device__ void *rocshmem_ptr(const void *dest, int pe) {
   GPU_DPRINTF("Function: rocshmem_ptr (dest=%p, pe=%d w%d\n",
     dest, pe, pe);