ch4/ofi: allocate pack buffers with alignment

allocate pack buffers with alignment to enable stream instructions
for GPU fast copy.

src/mpid/ch4/netmod/ofi/ofi_recv.h

@@ -282,7 +282,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_do_irecv(void *buf,
 
         /* Unpack */
         MPIDI_OFI_REQUEST(rreq, event_id) = MPIDI_OFI_EVENT_RECV_PACK;
-        MPIDI_OFI_REQUEST(rreq, noncontig.pack.pack_buffer) = MPL_malloc(data_sz, MPL_MEM_OTHER);
+        MPIDI_OFI_REQUEST(rreq, noncontig.pack.pack_buffer) =
+            MPL_aligned_alloc(64, data_sz, MPL_MEM_OTHER);
         MPIR_ERR_CHKANDJUMP1(MPIDI_OFI_REQUEST(rreq, noncontig.pack.pack_buffer) == NULL, mpi_errno,
                              MPI_ERR_OTHER, "**nomem", "**nomem %s", "Recv Pack Buffer alloc");
         recv_buf = MPIDI_OFI_REQUEST(rreq, noncontig.pack.pack_buffer);

src/mpid/ch4/netmod/ofi/ofi_send.h

@@ -625,7 +625,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_send(const void *buf, MPI_Aint count, MPI
         /* inject path */
         void *pack_buf = NULL;
         if (need_pack) {
-            pack_buf = MPL_malloc(data_sz, MPL_MEM_OTHER);
+            pack_buf = MPL_aligned_alloc(64, data_sz, MPL_MEM_OTHER);
             mpi_errno = MPIR_Localcopy_gpu(buf, count, datatype, 0, &attr,
                                            pack_buf, data_sz, MPI_BYTE, 0, MPIR_GPU_ATTR_HOST,
                                            MPL_GPU_COPY_DIRECTION_NONE,
@@ -649,7 +649,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_send(const void *buf, MPI_Aint count, MPI
 
         void *data = NULL;
         if (need_pack) {
-            void *pack_buf = MPL_malloc(data_sz, MPL_MEM_OTHER);
+            void *pack_buf = MPL_aligned_alloc(64, data_sz, MPL_MEM_OTHER);
             MPIR_ERR_CHKANDJUMP1(pack_buf == NULL, mpi_errno,
                                  MPI_ERR_OTHER, "**nomem", "**nomem %s", "Send Pack buffer alloc");