Re-enable the contiguous buffer optimization to the read_all and the write_all routines.

After long debugging, I found last week the reason this optimization originally broke
some hdf5 tests. We now pass the hdf5 test suite with the optimization being actively used.

Author: edgargabriel

ompi/mca/fcoll/dynamic/fcoll_dynamic_file_read_all.c

@@ -94,6 +94,10 @@ mca_fcoll_dynamic_file_read_all (mca_io_ompio_file_t *fh,
     ompi_datatype_t **sendtype = NULL;
     MPI_Request *send_req=NULL, recv_req=NULL;
     int my_aggregator =-1;
+    bool recvbuf_is_contiguous=false;
+    size_t ftype_size;
+    OPAL_PTRDIFF_TYPE ftype_extent, lb; 
+
 
 #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
     double read_time = 0.0, start_read_time = 0.0, end_read_time = 0.0;
@@ -105,11 +109,18 @@ mca_fcoll_dynamic_file_read_all (mca_io_ompio_file_t *fh,
     /**************************************************************************
      ** 1. In case the data is not contigous in memory, decode it into an iovec 
      **************************************************************************/
+
+    opal_datatype_type_size ( &datatype->super, &ftype_size );
+    opal_datatype_get_extent ( &datatype->super, &lb, &ftype_extent );
+
+    if ( (ftype_extent == (OPAL_PTRDIFF_TYPE) ftype_size)             && 
+        opal_datatype_is_contiguous_memory_layout(&datatype->super,1) && 
+        0 == lb ) {
+        recvbuf_is_contiguous = true;
+    }
+    
 
-//     if (opal_datatype_is_contiguous_memory_layout(&datatype->super,1)) {
-//	 fh->f_flags |= OMPIO_CONTIGUOUS_MEMORY;
-//     }
-    if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
+    if (! recvbuf_is_contiguous ) {
         ret = fh->f_decode_datatype ((struct mca_io_ompio_file_t *)fh,
                                      datatype,
                                      count,
@@ -760,7 +771,7 @@ mca_fcoll_dynamic_file_read_all (mca_io_ompio_file_t *fh,
         /**********************************************************
          *** 7f.  Scatter the Data from the readers
          *********************************************************/
-        if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) {
+        if ( recvbuf_is_contiguous ) {
             receive_buf = &((char*)buf)[position];
         }
         else if (bytes_received) {
@@ -807,7 +818,7 @@ mca_fcoll_dynamic_file_read_all (mca_io_ompio_file_t *fh,
 
         /* If data is not contigous in memory, copy the data from the
            receive buffer into the buffer passed in */
-        if (!(fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
+        if (!recvbuf_is_contiguous ) {
             OPAL_PTRDIFF_TYPE mem_address;
             size_t remaining = 0;
             size_t temp_position = 0;
@@ -868,7 +879,7 @@ mca_fcoll_dynamic_file_read_all (mca_io_ompio_file_t *fh,
 #endif
 
 exit:
-    if (!(fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
+    if (!recvbuf_is_contiguous) {
         if (NULL != receive_buf) {
             free (receive_buf);
             receive_buf = NULL;

ompi/mca/fcoll/dynamic/fcoll_dynamic_file_write_all.c

@@ -96,23 +96,33 @@ mca_fcoll_dynamic_file_write_all (mca_io_ompio_file_t *fh,
     MPI_Aint *total_bytes_per_process = NULL;
     MPI_Request send_req=NULL, *recv_req=NULL;
     int my_aggregator=-1;
+    bool sendbuf_is_contiguous = false;
+    size_t ftype_size;
+    OPAL_PTRDIFF_TYPE ftype_extent, lb; 
+
 
 #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
     double write_time = 0.0, start_write_time = 0.0, end_write_time = 0.0;
     double comm_time = 0.0, start_comm_time = 0.0, end_comm_time = 0.0;
     double exch_write = 0.0, start_exch = 0.0, end_exch = 0.0;
     mca_io_ompio_print_entry nentry;
 #endif
-    
-    
-//    if (opal_datatype_is_contiguous_memory_layout(&datatype->super,1)) {
-//        fh->f_flags |= OMPIO_CONTIGUOUS_MEMORY;
-//    }
+     
+    opal_datatype_type_size ( &datatype->super, &ftype_size );
+    opal_datatype_get_extent ( &datatype->super, &lb, &ftype_extent );
 
     /**************************************************************************
      ** 1.  In case the data is not contigous in memory, decode it into an iovec
      **************************************************************************/
-    if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
+    if ( ( ftype_extent == (OPAL_PTRDIFF_TYPE) ftype_size)             && 
+         opal_datatype_is_contiguous_memory_layout(&datatype->super,1) && 
+         0 == lb ) {
+        sendbuf_is_contiguous = true;
+    }
+    
+
+
+    if (! sendbuf_is_contiguous ) {
         ret =   fh->f_decode_datatype ((struct mca_io_ompio_file_t *) fh,
                                        datatype,
                                        count,
@@ -791,7 +801,7 @@ mca_fcoll_dynamic_file_write_all (mca_io_ompio_file_t *fh,
         } /* end if (my_aggregator == fh->f_rank ) */
 
 
-        if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) {
+        if ( sendbuf_is_contiguous ) {
             send_buf = &((char*)buf)[total_bytes_written];
         }
         else if (bytes_sent) {
@@ -880,7 +890,7 @@ mca_fcoll_dynamic_file_write_all (mca_io_ompio_file_t *fh,
 	}
 #endif
 
-        if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
+        if (! sendbuf_is_contiguous) {
             if (NULL != send_buf) {
                 free (send_buf);
                 send_buf = NULL;
@@ -1050,7 +1060,7 @@ exit :
 
     }
 
-    if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
+    if (! sendbuf_is_contiguous) {
 	if (NULL != send_buf) {
 	    free (send_buf);
 	    send_buf = NULL;