Allow btl/sm to memcpy from unified memory

Signed-off-by: Joseph Schuchart <schuchart@icl.utk.edu>

Author: devreal

opal/datatype/opal_convertor.h

@@ -192,6 +192,18 @@ static inline int32_t opal_convertor_on_device(const opal_convertor_t *pConverto
     return !!(pConvertor->flags & CONVERTOR_ACCELERATOR);
 }
 
+static inline int32_t opal_convertor_on_discrete_device(const opal_convertor_t *pConvertor)
+{
+    return (CONVERTOR_ACCELERATOR == ((pConvertor->flags & CONVERTOR_ACCELERATOR) |
+                                      (pConvertor->flags & CONVERTOR_ACCELERATOR_UNIFIED)));
+}
+
+static inline int32_t opal_convertor_on_unified_device(const opal_convertor_t *pConvertor)
+{
+    return (!!(pConvertor->flags & CONVERTOR_ACCELERATOR) &&
+            !!(pConvertor->flags & CONVERTOR_ACCELERATOR_UNIFIED));
+}
+
 /**
  * Update the size of the remote datatype representation. The size will
  * depend on the configuration of the master convertor. In homogeneous

opal/mca/btl/sm/btl_sm_module.c

@@ -430,7 +430,9 @@ static struct mca_btl_base_descriptor_t *sm_prepare_src(struct mca_btl_base_modu
 
     /* in place send fragment */
     if (OPAL_UNLIKELY(opal_convertor_need_buffers(convertor) ||
-                      opal_convertor_on_device(convertor))) {
+                      opal_convertor_on_discrete_device(convertor) ||
+                      (opal_convertor_on_unified_device(convertor) &&
+                       total_size > (size_t) mca_btl_sm_component.max_inline_send))) {
         uint32_t iov_count = 1;
         struct iovec iov;