summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorAlbert Cheng <acheng@hdfgroup.org>2004-08-05 05:15:08 (GMT)
committerAlbert Cheng <acheng@hdfgroup.org>2004-08-05 05:15:08 (GMT)
commit5390a63a77772ef0f91a0fd286128be9d5a84e8c (patch)
treeba9dbc00c5c69fc2ffc0f4957f661fff03440c37
parenta5bd3e42ea0913d6b652c26f82bd688ed6741cd1 (diff)
downloadhdf5-5390a63a77772ef0f91a0fd286128be9d5a84e8c.zip
hdf5-5390a63a77772ef0f91a0fd286128be9d5a84e8c.tar.gz
hdf5-5390a63a77772ef0f91a0fd286128be9d5a84e8c.tar.bz2
[svn-r9018] Purpose:
Bug fix (Failures when dataset size >= 1 GB, reported by Bill Loewe.) Description: In the IBM AIX system using 32bit mode, if a dataset size was 1GB or larger, when the "end" of the dataset was selected, MPI would complain it could not keep the Upper bound of a datatype within the range of MPI_Aint. This was because the old algorithm would derive the selection with extent of each row first. After all dimensions were processed, it then calculate the start position and just displace the whole MPI derived type. So, the final MPI type was actually the start position plus the whole dataset. Since the start can be as big as the whole dataset, this made the final derived twice as big as 1GB. That would hit the 2GB MPI_Aint range limit in the 32 bit mode. Solution: Use a different algorithm to include the start position in the defining of MPI type for each dimension. When all dimensions are processed, the MPI type represents the selection exactly. Platforms tested: h5committested Misc. update:
-rw-r--r--src/H5Smpio.c280
1 files changed, 276 insertions, 4 deletions
diff --git a/src/H5Smpio.c b/src/H5Smpio.c
index b04faeb..b46d347 100644
--- a/src/H5Smpio.c
+++ b/src/H5Smpio.c
@@ -181,12 +181,280 @@ H5S_mpio_none_type( const H5S_t UNUSED *space, size_t UNUSED elmt_size,
* akc, rky 2000-11-16 Replaced hard coded dimension size with
* H5S_MAX_RANK.
*
- * Quincey Koziol, June 18, 2002
- * Added 'extra_offset' parameter. Also accomodate
- * selection offset in MPI type built.
+ * Quincey Koziol, June 18, 2002
+ * Added 'extra_offset' parameter. Also accomodate selection
+ * offset in MPI type built.
+ *
+ * Albert Cheng, August 4, 2004
+ * Reimplemented the algorithm of forming the outer_type by
+ * defining it as (start, vector, extent) in one call.
*
*-------------------------------------------------------------------------
*/
+#ifndef AKC_OLD
+static herr_t
+H5S_mpio_hyper_type( const H5S_t *space, size_t elmt_size,
+ /* out: */
+ MPI_Datatype *new_type,
+ size_t *count,
+ hsize_t *extra_offset,
+ hbool_t *is_derived_type )
+{
+ H5S_sel_iter_t sel_iter; /* Selection iteration info */
+ hbool_t sel_iter_init=0; /* Selection iteration info has been initialized */
+
+ struct dim { /* less hassle than malloc/free & ilk */
+ hssize_t start;
+ hsize_t strid;
+ hsize_t block;
+ hsize_t xtent;
+ hsize_t count;
+ } d[H5S_MAX_RANK];
+
+ int i;
+ int offset[H5S_MAX_RANK];
+ int max_xtent[H5S_MAX_RANK];
+ H5S_hyper_dim_t *diminfo; /* [rank] */
+ int rank;
+ int block_length[3];
+ MPI_Datatype inner_type, outer_type, old_types[3];
+ MPI_Aint extent_len, displacement[3];
+ int mpi_code; /* MPI return code */
+ herr_t ret_value = SUCCEED;
+
+ FUNC_ENTER_NOAPI_NOINIT(H5S_mpio_hyper_type);
+
+ /* Check args */
+ assert (space);
+ assert(sizeof(MPI_Aint) >= sizeof(elmt_size));
+ if (0==elmt_size)
+ goto empty;
+
+ /* Initialize selection iterator */
+ if (H5S_select_iter_init(&sel_iter, space, elmt_size)<0)
+ HGOTO_ERROR (H5E_DATASPACE, H5E_CANTINIT, FAIL, "unable to initialize selection iterator");
+ sel_iter_init=1; /* Selection iteration info has been initialized */
+
+ /* Abbreviate args */
+ diminfo=sel_iter.u.hyp.diminfo;
+ assert (diminfo);
+
+ /* make a local copy of the dimension info so we can operate with them */
+
+ /* Check if this is a "flattened" regular hyperslab selection */
+ if(sel_iter.u.hyp.iter_rank!=0 && sel_iter.u.hyp.iter_rank<space->extent.rank) {
+ /* Flattened selection */
+ rank=sel_iter.u.hyp.iter_rank;
+ assert (rank >= 0 && rank<=H5S_MAX_RANK); /* within array bounds */
+ if (0==rank)
+ goto empty;
+
+#ifdef H5Smpi_DEBUG
+ HDfprintf(stderr, "%s: Flattened selection\n",FUNC);
+#endif
+ for ( i=0; i<rank; ++i) {
+ d[i].start = diminfo[i].start+sel_iter.u.hyp.sel_off[i];
+ d[i].strid = diminfo[i].stride;
+ d[i].block = diminfo[i].block;
+ d[i].count = diminfo[i].count;
+ d[i].xtent = sel_iter.u.hyp.size[i];
+#ifdef H5Smpi_DEBUG
+ HDfprintf(stderr, "%s: start=%Hd stride=%Hu count=%Hu block=%Hu xtent=%Hu",
+ FUNC, d[i].start, d[i].strid, d[i].count, d[i].block, d[i].xtent );
+ if (i==0)
+ HDfprintf(stderr, " rank=%d\n", rank );
+ else
+ HDfprintf(stderr, "\n" );
+#endif
+ if (0==d[i].block)
+ goto empty;
+ if (0==d[i].count)
+ goto empty;
+ if (0==d[i].xtent)
+ goto empty;
+ }
+ } /* end if */
+ else {
+ /* Non-flattened selection */
+ rank = space->extent.rank;
+ assert (rank >= 0 && rank<=H5S_MAX_RANK); /* within array bounds */
+ if (0==rank)
+ goto empty;
+
+#ifdef H5Smpi_DEBUG
+ HDfprintf(stderr, "%s: Non-flattened selection\n",FUNC);
+#endif
+ for ( i=0; i<rank; ++i) {
+ d[i].start = diminfo[i].start+space->select.offset[i];
+ d[i].strid = diminfo[i].stride;
+ d[i].block = diminfo[i].block;
+ d[i].count = diminfo[i].count;
+ d[i].xtent = space->extent.size[i];
+#ifdef H5Smpi_DEBUG
+ HDfprintf(stderr, "%s: start=%Hd stride=%Hu count=%Hu block=%Hu xtent=%Hu",
+ FUNC, d[i].start, d[i].strid, d[i].count, d[i].block, d[i].xtent );
+ if (i==0)
+ HDfprintf(stderr, " rank=%d\n", rank );
+ else
+ HDfprintf(stderr, "\n" );
+#endif
+ if (0==d[i].block)
+ goto empty;
+ if (0==d[i].count)
+ goto empty;
+ if (0==d[i].xtent)
+ goto empty;
+ }
+ } /* end else */
+
+/**********************************************************************
+ Compute array "offset[rank]" which gives the offsets for a multi-
+ dimensional array with dimensions "d[i].xtent" (i=0,1,...,rank-1).
+**********************************************************************/
+ offset[rank-1] = 1;
+ max_xtent[rank-1] = d[rank-1].xtent;
+#ifdef H5Smpi_DEBUG
+ i=rank-1;
+ HDfprintf(stderr, " offset[%2d]=%d; max_xtent[%2d]=%d\n",
+ i, offset[i], i, max_xtent[i]);
+#endif
+ for (i=rank-2; i>=0; --i) {
+ offset[i] = offset[i+1]*d[i+1].xtent;
+ max_xtent[i] = max_xtent[i+1]*d[i].xtent;
+#ifdef H5Smpi_DEBUG
+ HDfprintf(stderr, " offset[%2d]=%d; max_xtent[%2d]=%d\n",
+ i, offset[i], i, max_xtent[i]);
+#endif
+ }
+
+ /* Create a type covering the selected hyperslab.
+ * Multidimensional dataspaces are stored in row-major order.
+ * The type is built from the inside out, going from the
+ * fastest-changing (i.e., inner) dimension * to the slowest (outer). */
+
+/*******************************************************
+* Construct contig type for inner contig dims:
+*******************************************************/
+#ifdef H5Smpi_DEBUG
+ HDfprintf(stderr, "%s: Making contig type %d MPI_BYTEs\n", FUNC,elmt_size );
+ for (i=rank-1; i>=0; --i)
+ HDfprintf(stderr, "d[%d].xtent=%Hu \n", i, d[i].xtent);
+#endif
+ if (MPI_SUCCESS != (mpi_code= MPI_Type_contiguous( (int)elmt_size, MPI_BYTE, &inner_type )))
+ HMPI_GOTO_ERROR(FAIL, "MPI_Type_contiguous failed", mpi_code);
+
+/*******************************************************
+* Construct the type by walking the hyperslab dims
+* from the inside out:
+*******************************************************/
+ for ( i=rank-1; i>=0; --i) {
+#ifdef H5Smpi_DEBUG
+ HDfprintf(stderr, "%s: Dimension i=%d \n"
+ "start=%Hd count=%Hu block=%Hu stride=%Hu, xtent=%Hu max_xtent=%d\n",
+ FUNC, i, d[i].start, d[i].count, d[i].block, d[i].strid, d[i].xtent, max_xtent[i]);
+#endif
+
+#ifdef H5Smpi_DEBUG
+ HDfprintf(stderr, "%s: i=%d Making vector-type \n", FUNC,i);
+#endif
+ /****************************************
+ * Build vector type of the selection.
+ ****************************************/
+ mpi_code =MPI_Type_vector((int)(d[i].count), /* count */
+ (int)(d[i].block), /* blocklength */
+ (int)(d[i].strid), /* stride */
+ inner_type, /* old type */
+ &outer_type); /* new type */
+
+ MPI_Type_free( &inner_type );
+ if (mpi_code!=MPI_SUCCESS)
+ HMPI_GOTO_ERROR(FAIL, "couldn't create MPI vector type", mpi_code);
+
+ /****************************************
+ * Then build the dimension type as (start, vector type, xtent).
+ ****************************************/
+ /* calculate start and extent values of this dimension */
+ displacement[1] = d[i].start * offset[i] * elmt_size;
+ displacement[2] = (MPI_Aint)elmt_size * max_xtent[i];
+ if(MPI_SUCCESS != (mpi_code = MPI_Type_extent(outer_type, &extent_len)))
+ HMPI_GOTO_ERROR(FAIL, "MPI_Type_extent failed", mpi_code);
+
+ /*************************************************
+ * Restructure this datatype ("outer_type")
+ * so that it still starts at 0, but its extent
+ * is the full extent in this dimension.
+ *************************************************/
+ if (displacement[1] > 0 || (int)extent_len < displacement[2]) {
+
+ block_length[0] = 1;
+ block_length[1] = 1;
+ block_length[2] = 1;
+
+ displacement[0] = 0;
+
+ old_types[0] = MPI_LB;
+ old_types[1] = outer_type;
+ old_types[2] = MPI_UB;
+#ifdef H5Smpi_DEBUG
+ HDfprintf(stderr, "%s: i=%d Extending struct type\n"
+ "***displacements: %d, %d, %d\n",
+ FUNC, i, displacement[0], displacement[1], displacement[2]);
+#endif
+
+ mpi_code = MPI_Type_struct ( 3, /* count */
+ block_length, /* blocklengths */
+ displacement, /* displacements */
+ old_types, /* old types */
+ &inner_type); /* new type */
+
+ MPI_Type_free (&outer_type);
+ if (mpi_code!=MPI_SUCCESS)
+ HMPI_GOTO_ERROR(FAIL, "couldn't resize MPI vector type", mpi_code);
+ }
+ else {
+ inner_type = outer_type;
+ }
+ } /* end for */
+/***************************
+* End of loop, walking
+* thru dimensions.
+***************************/
+
+
+ /* At this point inner_type is actually the outermost type, even for 0-trip loop */
+
+ *new_type = inner_type;
+ if (MPI_SUCCESS != (mpi_code= MPI_Type_commit( new_type )))
+ HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code);
+
+ /* fill in the remaining return values */
+ *count = 1; /* only have to move one of these suckers! */
+ *extra_offset = 0;
+ *is_derived_type = 1;
+ HGOTO_DONE(SUCCEED);
+
+empty:
+ /* special case: empty hyperslab */
+ *new_type = MPI_BYTE;
+ *count = 0;
+ *extra_offset = 0;
+ *is_derived_type = 0;
+
+done:
+ /* Release selection iterator */
+ if(sel_iter_init) {
+ if (H5S_SELECT_ITER_RELEASE(&sel_iter)<0)
+ HDONE_ERROR (H5E_DATASPACE, H5E_CANTRELEASE, FAIL, "unable to release selection iterator");
+ } /* end if */
+
+#ifdef H5Smpi_DEBUG
+ HDfprintf(stderr, "Leave %s, count=%ld is_derived_type=%d\n",
+ FUNC, *count, *is_derived_type );
+#endif
+ FUNC_LEAVE_NOAPI(ret_value);
+}
+#else
+/* keep this old code for now. */
static herr_t
H5S_mpio_hyper_type( const H5S_t *space, size_t elmt_size,
/* out: */
@@ -427,6 +695,9 @@ H5S_mpio_hyper_type( const H5S_t *space, size_t elmt_size,
for (i=rank-1; i>=0; i--)
displacement[0] += d[i].start * offset[i];
+printf("dumping MPI_BYTE\n");
+printdatatype(MPI_INT);
+printdatatype(MPI_BYTE);
if (displacement[0] > 0) {
displacement[0] *= elmt_size;
block_length[0] = 1;
@@ -478,11 +749,12 @@ done:
} /* end if */
#ifdef H5Smpi_DEBUG
- HDfprintf(stderr, "Leave %s, count=%Hu is_derived_type=%d\n",
+ HDfprintf(stderr, "Leave %s, count=%ld is_derived_type=%d\n",
FUNC, *count, *is_derived_type );
#endif
FUNC_LEAVE_NOAPI(ret_value);
}
+#endif
/*-------------------------------------------------------------------------