file: add no-accelerator assertion flag

- add a new flag indicating that the user guarantueed not to use
  accelerator memory with a particular file.
- check that flag in common_ompio_check_gpu_buf() and return immediatly
  `false` if the flag is set, i.e. bypassing the accelerator
  check_addr() routine (which is supposed to be the expensive part).

Signed-off-by: Edgar Gabriel <Edgar.Gabriel@amd.com>

Author: edgargabriel

ompi/file/file.c

@@ -18,7 +18,7 @@
  * Copyright (c) 2016-2017 IBM Corporation. All rights reserved.
  * Copyright (c) 2024      Triad National Security, LLC. All rights
  *                         reserved.
- * Copyright (c) 2024      Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2024-2025 Advanced Micro Devices, Inc. All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -126,7 +126,9 @@ int ompi_file_open(struct ompi_communicator_t *comm, const char *filename,
     }
     ompi_info_memkind_assert_type type;
     ompi_info_memkind_copy_or_set (&comm->instance->super, &file->super, info, &type);
-
+    if (OMPI_INFO_MEMKIND_ASSERT_NO_ACCEL == type) {
+        file->f_flags |= OMPI_FILE_ASSERT_NO_ACCEL_BUF;
+    }
     file->f_amode = amode;
     file->f_filename = strdup(filename);
     if (NULL == file->f_filename) {

ompi/file/file.h

@@ -39,9 +39,9 @@
 /*
  * Flags
  */
-#define OMPI_FILE_ISCLOSED     0x00000001
-#define OMPI_FILE_HIDDEN       0x00000002
-
+#define OMPI_FILE_ISCLOSED            0x00000001
+#define OMPI_FILE_HIDDEN              0x00000002
+#define OMPI_FILE_ASSERT_NO_ACCEL_BUF 0x00000004
 BEGIN_C_DECLS
 
 /**

ompi/mca/common/ompio/common_ompio_buffer.c

@@ -11,6 +11,7 @@
  *                          All rights reserved.
  *  Copyright (c) 2008-2019 University of Houston. All rights reserved.
  *  Copyright (c) 2022      Amazon.com, Inc. or its affiliates.  All Rights reserved.
+ *  Copyright (c) 2025      Advanced Micro Devices, Inc. All rights reserved.
  *  $COPYRIGHT$
  *
  *  Additional copyrights may follow
@@ -47,7 +48,11 @@ void mca_common_ompio_check_gpu_buf ( ompio_file_t *fh, const void *buf, int *is
 
     *is_gpu=0;
     *is_managed=0;
-    
+
+    if (fh->f_fh->f_flags & OMPI_FILE_ASSERT_NO_ACCEL_BUF) {
+        return;
+    }
+
     if (0 < opal_accelerator.check_addr(buf, &dev_id, &flags)) {
         *is_gpu = 1;
         if (flags & MCA_ACCELERATOR_FLAGS_UNIFIED_MEMORY) {
@@ -62,16 +67,14 @@ static void* mca_common_ompio_buffer_alloc_seg ( void*ctx, size_t *size )
 {
     char *buf=NULL;
     size_t realsize, numpages;
-    uint64_t flags = 0;
-    int dev_id;
 
     numpages = (*size + mca_common_ompio_pagesize -1 )/mca_common_ompio_pagesize;
     realsize = numpages * mca_common_ompio_pagesize;
 
-    buf = malloc ( realsize);
+    buf = malloc (realsize);
 
-    if (NULL != buf && 0 == opal_accelerator.check_addr(buf, &dev_id, &flags)) {
-        opal_accelerator.host_register(dev_id, (void *)buf, realsize);
+    if (NULL != buf) {
+        opal_accelerator.host_register(MCA_ACCELERATOR_NO_DEVICE_ID, (void *)buf, realsize);
     }
 
     *size = realsize;