accelerator: use correct copy type if known

fix the copy type passed to the accelerator framework if the information
is available. (Similarly use the correct device id if the info is
already available, that might be required at a later stage if we
want to support multiple GPUs per process).

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

Author: edgargabriel

opal/datatype/opal_datatype_copy.c

@@ -58,7 +58,8 @@
 static void *opal_datatype_accelerator_memcpy(void *dest, const void *src, size_t size)
 {
     int res;
-    int dev_id;
+    int dst_type, dst_dev, src_type, src_dev;
+    int copy_type = MCA_ACCELERATOR_TRANSFER_DTOD;
     uint64_t flags;
     /* If accelerator check addr returns an error, we can only
      * assume it is a host buffer. If device buffer checking fails,
@@ -67,12 +68,19 @@ static void *opal_datatype_accelerator_memcpy(void *dest, const void *src, size_
      * and retries are also unlikely to succeed. We identify these
      * buffers as host buffers as attempting a memcpy would provide
      * a chance to succeed. */
-    if (0 >= opal_accelerator.check_addr(dest, &dev_id, &flags) &&
-        0 >= opal_accelerator.check_addr(src, &dev_id, &flags)) {
+    dst_type = opal_accelerator.check_addr(dest, &dst_dev, &flags);
+    src_type = opal_accelerator.check_addr(src, &src_dev, &flags);
+    if (0 >= dst_type && 0 >= src_type) {
         return memcpy(dest, src, size);
     }
-    res = opal_accelerator.mem_copy(MCA_ACCELERATOR_NO_DEVICE_ID, MCA_ACCELERATOR_NO_DEVICE_ID,
-                                  dest, src, size, MCA_ACCELERATOR_TRANSFER_UNSPEC);
+    else if (0 >= dst_type && 0 < src_type) {
+	copy_type = MCA_ACCELERATOR_TRANSFER_DTOH;
+    }
+    else if (0 < dst_type && 0 >= dst_type) {
+	copy_type = MCA_ACCELERATOR_TRANSFER_HTOD;
+    }
+    res = opal_accelerator.mem_copy(dst_dev, src_dev,
+				    dest, src, size, copy_type);
     if (OPAL_SUCCESS != res) {
         opal_output(0, "Error in accelerator memcpy");
         abort();
@@ -83,7 +91,8 @@ static void *opal_datatype_accelerator_memcpy(void *dest, const void *src, size_
 static void *opal_datatype_accelerator_memmove(void *dest, const void *src, size_t size)
 {
     int res;
-    int dev_id;
+    int dst_type, dst_dev, src_type, src_dev;
+    int copy_type = MCA_ACCELERATOR_TRANSFER_DTOD;
     uint64_t flags;
     /* If accelerator check addr returns an error, we can only
      * assume it is a host buffer. If device buffer checking fails,
@@ -92,12 +101,19 @@ static void *opal_datatype_accelerator_memmove(void *dest, const void *src, size
      * and retries are also unlikely to succeed. We identify these
      * buffers as host buffers as attempting a memmove would provide
      * a chance to succeed. */
-    if (0 >= opal_accelerator.check_addr(dest, &dev_id, &flags) &&
-        0 >= opal_accelerator.check_addr(src, &dev_id, &flags)) {
+    dst_type = opal_accelerator.check_addr(dest, &dst_dev, &flags);
+    src_type = opal_accelerator.check_addr(src, &src_dev, &flags);
+    if (0 >= dst_type && 0 >= src_type) {
         return memmove(dest, src, size);
     }
-    res = opal_accelerator.mem_move(MCA_ACCELERATOR_NO_DEVICE_ID, MCA_ACCELERATOR_NO_DEVICE_ID,
-                                    dest, src, size, MCA_ACCELERATOR_TRANSFER_UNSPEC);
+    else if (0 >= dst_type && 0 < src_type) {
+	copy_type = MCA_ACCELERATOR_TRANSFER_DTOH;
+    }
+    else if (0 < dst_type && 0 >= dst_type) {
+	copy_type = MCA_ACCELERATOR_TRANSFER_HTOD;
+    }
+    res = opal_accelerator.mem_move(dst_dev, src_dev,
+                                    dest, src, size, copy_type);
     if (OPAL_SUCCESS != res) {
         opal_output(0, "Error in accelerator memmove");
         abort();

opal/mca/accelerator/rocm/accelerator_rocm_module.c

@@ -334,7 +334,9 @@ static int mca_accelerator_rocm_memcpy(int dest_dev_id, int src_dev_id, void *de
         return OPAL_SUCCESS;
     }
 
-    if (type == MCA_ACCELERATOR_TRANSFER_DTOH && size <= opal_accelerator_rocm_memcpyD2H_limit) {
+    if ((type == MCA_ACCELERATOR_TRANSFER_DTOH ||
+	 type == MCA_ACCELERATOR_TRANSFER_UNSPEC) &&
+	size <= opal_accelerator_rocm_memcpyD2H_limit) {
         memcpy(dest, src, size);
         return OPAL_SUCCESS;
     }

opal/mca/btl/smcuda/btl_smcuda_accelerator.c

@@ -65,7 +65,7 @@ int mca_btl_smcuda_accelerator_init(void)
     }
     /* Create the events since they can be reused. */
     for (i = 0; i < accelerator_event_max; i++) {
-        rc = opal_accelerator.create_event(MCA_ACCELERATOR_NO_DEVICE_ID, &accelerator_event_ipc_array[i], true);
+        rc = opal_accelerator.create_event(MCA_ACCELERATOR_NO_DEVICE_ID, &accelerator_event_ipc_array[i], opal_accelerator_use_sync_memops ? false : true);
         if (OPAL_SUCCESS != rc) {
             opal_output_verbose(1, mca_btl_smcuda_component.cuda_ipc_output, "Accelerator create event failed.");
             rc = OPAL_ERROR;
@@ -215,7 +215,7 @@ int mca_btl_smcuda_memcpy(void *dst, void *src, size_t amount, char *msg,
     }
 
     result = opal_accelerator.mem_copy_async(MCA_ACCELERATOR_NO_DEVICE_ID, MCA_ACCELERATOR_NO_DEVICE_ID,
-                                             dst, src, amount, ipc_stream, MCA_ACCELERATOR_TRANSFER_UNSPEC);
+                                             dst, src, amount, ipc_stream, MCA_ACCELERATOR_TRANSFER_DTOD);
     if (OPAL_UNLIKELY(OPAL_SUCCESS != result)) {
         opal_output_verbose(1, mca_btl_smcuda_component.cuda_ipc_output, "smcuda: memcpy async failed: %d",
                             result);