Add a Fine Grained Completion Routine: shmem_pe_quiet

Author: avincigu

mpp/shmemx_c_func.h4

@@ -105,3 +105,7 @@ SHMEM_FUNCTION_ATTRIBUTES void SHPRE()shmemx_heap_create(void *base, size_t size
 SHMEM_FUNCTION_ATTRIBUTES void SHPRE()shmemx_heap_preinit(void);
 SHMEM_FUNCTION_ATTRIBUTES int SHPRE()shmemx_heap_preinit_thread(int requested, int *provided);
 SHMEM_FUNCTION_ATTRIBUTES void SHPRE()shmemx_heap_postinit(void);
+
+/* Memory Ordering Routines */
+SHMEM_FUNCTION_ATTRIBUTES void shmem_pe_quiet(const int *target_pes, int npes);
+SHMEM_FUNCTION_ATTRIBUTES void shmem_ctx_quiet(shmem_ctx_t ctx, const int *target_pes, int npes);

src/shmem_synchronization.h

@@ -40,6 +40,25 @@ shmem_internal_quiet(shmem_ctx_t ctx)
     shmem_transport_syncmem();
 }
 
+static inline void
+shmem_internal_pe_quiet(shmem_ctx_t ctx, const int *target_pes, int npes)
+{
+    int ret;
+
+    if (ctx == SHMEM_CTX_INVALID)
+        return;
+
+    ret = shmem_transport_pe_quiet((shmem_transport_ctx_t *)ctx, target_pes, npes);
+    if (0 != ret) { RAISE_ERROR(ret); }
+
+    shmem_internal_membar();
+
+    /* Transport level memory flush is required to make memory 
+     * changes (i.e. subsequent coherent load operations 
+     * performed via the shmem_ptr API, the result of atomics 
+     * that targeted the local process) visible */
+    shmem_transport_syncmem();
+}
 
 static inline void
 shmem_internal_fence(shmem_ctx_t ctx)

src/synchronization_f.c

@@ -37,6 +37,28 @@ FC_SHMEM_QUIET(void)
     shmem_internal_quiet(SHMEM_CTX_DEFAULT);
 }
 
+#define FC_SHMEM_PE_QUIET FC_FUNC_(shmem_pe_quiet, SHMEM_PE_QUIET)
+void SHMEM_FUNCTION_ATTRIBUTES
+FC_SHMEM_PE_QUIET(const int *target_pes, int npes);
+void
+FC_SHMEM_PE_QUIET(const int *target_pes, int npes)
+{
+    SHMEM_ERR_CHECK_INITIALIZED();
+
+    shmem_internal_pe_quiet(SHMEM_CTX_DEFAULT, target_pes, npes);
+}
+
+#define FC_SHMEM_CTX_QUIET FC_FUNC_(shmem_ctx_quiet, SHMEM_CTX_QUIET)
+void SHMEM_FUNCTION_ATTRIBUTES
+FC_SHMEM_CTX_QUIET(shmem_ctx_t ctx, const int *target_pes, int npes);
+void
+FC_SHMEM_CTX_QUIET(shmem_ctx_t ctx, const int *target_pes, int npes)
+{
+    SHMEM_ERR_CHECK_INITIALIZED();
+
+    shmem_internal_pe_quiet(ctx, target_pes, npes);
+}
+
 
 #define FC_SHMEM_FENCE FC_FUNC_(shmem_fence, SHMEM_FENCE)
 void SHMEM_FUNCTION_ATTRIBUTES

src/transport_none.h

@@ -103,6 +103,13 @@ shmem_transport_quiet(shmem_transport_ctx_t* ctx)
     return 0;
 }
 
+static inline
+int
+shmem_transport_pe_quiet(shmem_transport_ctx_t* ctx, const int *target_pes, int npes)
+{
+    return 0;
+}
+
 static inline
 int
 shmem_transport_fence(shmem_transport_ctx_t* ctx)

src/transport_ofi.c

@@ -2021,6 +2021,11 @@ int shmem_transport_ctx_create(struct shmem_internal_team_t *team, long options,
 #ifndef USE_CTX_LOCK
     shmem_internal_cntr_write(&ctxp->pending_put_cntr, 0);
     shmem_internal_cntr_write(&ctxp->pending_get_cntr, 0);
+    
+    for (int i = 0; i < shmem_internal_num_pes; i++) {
+        shmem_internal_cntr_write(&ctx->pending_get_per_pe[i], 0);
+        shmem_internal_cntr_write(&ctx->pending_put_per_pe[i], 0);
+    }
 #endif
 
     ctxp->stx_idx = -1;