src: Add DISABLE_NONFETCH_AMO support to op_to_all impls.

Sandia-OpenSHMEM · Jun 10, 2024 · 99609bf · 99609bf
1 parent fab23f4
commit 99609bf
Show file tree

Hide file tree

Showing 2 changed files with 26 additions and 0 deletions.
diff --git a/src/collectives.c b/src/collectives.c
@@ -613,8 +613,19 @@ shmem_internal_op_to_all_linear(void *target, const void *source, size_t count,
         SHMEM_WAIT_UNTIL(pSync, SHMEM_CMP_EQ, 0);
 
         /* send data, ack, and wait for completion */
+#ifdef DISABLE_NONFETCH_AMO
+        /* FIXME: This is a temporary workaround to resolve a known issue with non-fetching AMOs when using
+           the CXI provider */
+        unsigned long long tmp_fetch = 0;
+        for (size_t i =0; i < count; i++) {
+            shmem_internal_fetch_atomic(SHMEM_CTX_DEFAULT, ((uint8_t *) target) + (i * type_size),
+                                        ((uint8_t *) source) + (i * type_size), &tmp_fetch, type_size,
+                                        PE_start, op, datatype);
+        }
+#else
         shmem_internal_atomicv(SHMEM_CTX_DEFAULT, target, source, count * type_size,
                                PE_start, op, datatype, &completion);
+#endif
         shmem_internal_put_wait(SHMEM_CTX_DEFAULT, &completion);
         shmem_internal_fence(SHMEM_CTX_DEFAULT);
 
@@ -817,10 +828,21 @@ shmem_internal_op_to_all_tree(void *target, const void *source, size_t count, si
         SHMEM_WAIT_UNTIL(pSync + 1, SHMEM_CMP_EQ, 0);
 
         /* send data, ack, and wait for completion */
+#ifdef DISABLE_NONFETCH_AMO
+        /* FIXME: This is a temporary workaround to resolve a known issue with non-fetching AMOs when using
+           the CXI provider */
+        unsigned long long tmp_fetch = 0;
+        for (size_t i = 0; i < count; i++) {
+            shmem_internal_fetch_atomic(SHMEM_CTX_DEFAULT, ((uint8_t *) target) + (i * type_size),
+                                        (num_children == 0) ? ((uint8_t *) source) + (i * type_size) : ((uint8_t *) target) + (i * type_size),
+                                        &tmp_fetch, type_size, parent, op, datatype);
+        }
+#else
         shmem_internal_atomicv(SHMEM_CTX_DEFAULT, target,
                                (num_children == 0) ? source : target,
                                count * type_size, parent,
                                op, datatype, &completion);
+#endif
         shmem_internal_put_wait(SHMEM_CTX_DEFAULT, &completion);
         shmem_internal_fence(SHMEM_CTX_DEFAULT);
 

diff --git a/src/shmem_comm.h b/src/shmem_comm.h
@@ -245,6 +245,8 @@ shmem_internal_atomic(shmem_ctx_t ctx, void *target, const void *source, size_t
         shmem_shr_transport_atomic(ctx, target, source, len, pe, op, datatype);
     } else {
 #ifdef DISABLE_NONFETCH_AMO
+        /* FIXME: This is a temporary workaround to resolve a known issue with non-fetching AMOs when using
+           the CXI provider */
         unsigned long long tmp_fetch = 0;
         shmem_transport_fetch_atomic((shmem_transport_ctx_t *)ctx, target,
                                      source, &tmp_fetch, len, pe, op, datatype);
@@ -284,6 +286,8 @@ shmem_internal_atomic_set(shmem_ctx_t ctx, void *target, const void *source, siz
         shmem_shr_transport_atomic_set(ctx, target, source, len, pe, datatype);
     } else {
 #ifdef DISABLE_NONFETCH_AMO
+        /* FIXME: This is a temporary workaround to resolve a known issue with non-fetching AMOs when using
+           the CXI provider */
         unsigned long long tmp_fetch = 0;
         shmem_transport_fetch_atomic((shmem_transport_ctx_t *)ctx, target,
                                      source, &tmp_fetch, len, pe, FI_ATOMIC_WRITE, datatype);