btl/ofi: use libfabric 1.18 API if available

This patch try to use 1.18 API when it is available. This is because
1.18 API clearly define provider's CUDA API behavior to be that
provider can call CUDA API by default if application uses 1.18 API.

When using older version of API, some libfabric will not claim
support of FI_HMEM even if it is capable of supporting because
the provider does not know whether CUDA calls are permitted.

Signed-off-by: Wei Zhang <[email protected]>

Author: wzamazon

opal/mca/btl/ofi/btl_ofi_component.c

@@ -350,8 +350,12 @@ static mca_btl_base_module_t **mca_btl_ofi_component_init(int *num_btl_modules,
 no_hmem:
 #endif
 
-    /* Do the query. The earliest version that supports FI_HMEM hints is 1.9  */
-    rc = fi_getinfo(FI_VERSION(1, 9), NULL, NULL, 0, &hints, &info_list);
+    /* Do the query. The earliest version that supports FI_HMEM hints is 1.9.
+     * The earliest version the explictly allow provider to call CUDA API is 1.18  */
+    rc = fi_getinfo(FI_VERSION(1, 18), NULL, NULL, 0, &hints, &info_list);
+    if (FI_ENOSYS == -rc) {
+	rc = fi_getinfo(FI_VERSION(1, 9), NULL, NULL, 0, &hints, &info_list);
+    }
     if (0 != rc) {
 #if defined(FI_HMEM)
         if (hints.caps & FI_HMEM) {