ompi_datatype_t* test_struct( void )
{
ompi_datatype_t* types[] = { &ompi_mpi_float.dt /* ompi_datatype_basicDatatypes[DT_FLOAT] */,
NULL,
&ompi_mpi_char.dt /* ompi_datatype_basicDatatypes[DT_CHAR] */ };
int lengths[] = { 2, 1, 3 };
MPI_Aint disp[] = { 0, 16, 26 };
ompi_datatype_t* pdt, *pdt1;
printf( "test struct\n" );
ompi_datatype_create_contiguous(0, &ompi_mpi_datatype_null.dt, &pdt1);
ompi_datatype_add( pdt1, &ompi_mpi_double.dt, 1, 0, -1 );
ompi_datatype_add( pdt1, &ompi_mpi_char.dt, 1, 8, -1 );
if( outputFlags & DUMP_DATA_AFTER_COMMIT ) {
ompi_datatype_dump( pdt1 );
}
types[1] = pdt1;
ompi_datatype_create_struct( 3, lengths, disp, types, &pdt );
OBJ_RELEASE( pdt1 ); /*assert( pdt1 == NULL );*/
if( outputFlags & DUMP_DATA_AFTER_COMMIT ) {
ompi_datatype_dump( pdt );
}
return pdt;
}
ompi_datatype_t* create_strange_dt( void )
{
sdata_intern v[2];
MPI_Aint displ[3];
ompi_datatype_t* types[3] = { &ompi_mpi_int.dt };
sstrange t[2];
int pBlock[3] = {1, 10, 1}, dispi[3];
ompi_datatype_t *pdt, *pdt1, *pdt2, *pdtTemp;
dispi[0] = (int)((char*)&(v[0].i1) - (char*)&(v[0])); /* 0 */
dispi[1] = (int)(((char*)(&(v[0].i2)) - (char*)&(v[0])) / sizeof(int)); /* 2 */
ompi_datatype_create_indexed_block( 2, 1, dispi, &ompi_mpi_int.dt, &pdtTemp );
#ifdef USE_RESIZED
/* optional */
displ[0] = 0;
displ[1] = (char*)&(v[1]) - (char*)&(v[0]);
ompi_datatype_create_resized( pdtTemp, displ[0], displ[1], &pdt1 );
OBJ_RELEASE( pdtTemp ); assert( pdtTemp == NULL );
#else
pdt1 = pdtTemp;
#endif /* USE_RESIZED */
types[1] = pdt1;
types[2] = &ompi_mpi_int.dt;
displ[0] = 0;
displ[1] = (long)((char*)&(t[0].v[0]) - (char*)&(t[0]));
displ[2] = (long)((char*)&(t[0].last) - (char*)&(t[0]));
ompi_datatype_create_struct( 3, pBlock, displ, types, &pdtTemp );
#ifdef USE_RESIZED
/* optional */
displ[1] = (char*)&(t[1]) - (char*)&(t[0]);
ompi_datatype_create_resized( pdtTemp, displ[0], displ[1], &pdt2 );
OBJ_RELEASE( pdtTemp ); assert( pdtTemp == NULL );
#else
pdt2 = pdtTemp;
#endif /* USE_RESIZED */
ompi_datatype_create_contiguous( SSTRANGE_CNT, pdt2, &pdt );
OBJ_RELEASE( pdt1 );
OBJ_RELEASE( pdt2 );
printf( "\nStrange datatype BEFORE COMMIT\n" );
if( outputFlags & DUMP_DATA_AFTER_COMMIT ) {
ompi_datatype_dump( pdt );
}
ompi_datatype_commit( &pdt );
printf( "\nStrange datatype AFTER COMMIT\n" );
if( outputFlags & DUMP_DATA_AFTER_COMMIT ) {
ompi_datatype_dump( pdt );
}
return pdt;
}
ompi_datatype_t* test_struct_char_double( void )
{
char_double_t data;
int lengths[] = {1, 1};
MPI_Aint displ[] = {0, 0};
ompi_datatype_t *pdt;
ompi_datatype_t* types[] = { &ompi_mpi_char.dt, &ompi_mpi_double.dt};
displ[0] = (char*)&(data.c) - (char*)&(data);
displ[1] = (char*)&(data.d) - (char*)&(data);
ompi_datatype_create_struct( 2, lengths, displ, types, &pdt );
ompi_datatype_commit( &pdt );
if( outputFlags & DUMP_DATA_AFTER_COMMIT ) {
ompi_datatype_dump( pdt );
}
return pdt;
}
ompi_datatype_t* create_struct_constant_gap_resized_ddt( ompi_datatype_t* type )
{
struct structure data[1];
ompi_datatype_t *struct_type, *temp_type;
ompi_datatype_t *types[2] = {type, type};
int blocklens[2] = {1, 1};
MPI_Aint disps[3];
MPI_Get_address(&data[0].transfered_1, &disps[0]);
MPI_Get_address(&data[0].transfered_2, &disps[1]);
MPI_Get_address(&data[0], &disps[2]);
disps[1] -= disps[2]; /* 8 */
disps[0] -= disps[2]; /* 16 */
ompi_datatype_create_struct(2, blocklens, disps, types, &temp_type);
ompi_datatype_create_resized(temp_type, 0, sizeof(data[0]), &struct_type);
ompi_datatype_commit(&struct_type);
OBJ_RELEASE(temp_type); assert( temp_type == NULL );
if( outputFlags & DUMP_DATA_AFTER_COMMIT ) {
ompi_datatype_dump( struct_type );
}
return struct_type;
}
static int
cyclic(const int *gsize_array, int dim, int ndims, int nprocs,
int rank, int darg, int order, ptrdiff_t orig_extent,
ompi_datatype_t* type_old, ompi_datatype_t **type_new,
ptrdiff_t *st_offset)
{
int blksize, i, blklens[2], st_index, end_index, local_size, rem, count, rc;
ptrdiff_t stride, disps[2];
ompi_datatype_t *type_tmp, *types[2];
if (darg == MPI_DISTRIBUTE_DFLT_DARG) {
blksize = 1;
} else {
blksize = darg;
}
st_index = rank * blksize;
end_index = gsize_array[dim] - 1;
if (end_index < st_index) {
local_size = 0;
} else {
local_size = ((end_index - st_index + 1)/(nprocs*blksize))*blksize;
rem = (end_index - st_index + 1) % (nprocs*blksize);
local_size += rem < blksize ? rem : blksize;
}
count = local_size / blksize;
rem = local_size % blksize;
stride = nprocs*blksize*orig_extent;
if (order == MPI_ORDER_FORTRAN) {
for (i=0; i<dim; i++) {
stride *= gsize_array[i];
}
} else {
for (i=ndims-1; i>dim; i--) {
stride *= gsize_array[i];
}
}
rc = ompi_datatype_create_hvector(count, blksize, stride, type_old, type_new);
if (OMPI_SUCCESS != rc) return rc;
if (rem) {
/* if the last block is of size less than blksize, include
it separately using MPI_Type_struct */
types [0] = *type_new; types [1] = type_old;
disps [0] = 0; disps [1] = count*stride;
blklens[0] = 1; blklens[1] = rem;
rc = ompi_datatype_create_struct(2, blklens, disps, types, &type_tmp);
ompi_datatype_destroy(type_new);
/* even in error condition, need to destroy type_new, so check
for error after destroy. */
if (OMPI_SUCCESS != rc) return rc;
*type_new = type_tmp;
}
/* need to set the UB for block-cyclic to work */
disps[0] = 0; disps[1] = orig_extent;
if (order == MPI_ORDER_FORTRAN) {
for(i=0; i<=dim; i++) {
disps[1] *= gsize_array[i];
}
} else {
for(i=ndims-1; i>=dim; i--) {
disps[1] *= gsize_array[i];
}
}
rc = opal_datatype_resize( &(*type_new)->super, disps[0], disps[1] );
if (OMPI_SUCCESS != rc) return rc;
*st_offset = rank * blksize;
/* in terms of no. of elements of type oldtype in this dimension */
if (local_size == 0) *st_offset = 0;
return OMPI_SUCCESS;
}
int32_t ompi_datatype_create_darray(int size,
int rank,
int ndims,
int const* gsize_array,
int const* distrib_array,
int const* darg_array,
int const* psize_array,
int order,
const ompi_datatype_t* oldtype,
ompi_datatype_t** newtype)
{
ompi_datatype_t *lastType;
ptrdiff_t orig_extent, *st_offsets = NULL;
int i, start_loop, end_loop, step;
int *coords = NULL, rc = OMPI_SUCCESS;
/* speedy corner case */
if (ndims < 1) {
/* Don't just return MPI_DATATYPE_NULL as that can't be
MPI_TYPE_FREE()ed, and that seems bad */
*newtype = ompi_datatype_create(0);
ompi_datatype_add(*newtype, &ompi_mpi_datatype_null.dt, 0, 0, 0);
return MPI_SUCCESS;
}
rc = ompi_datatype_type_extent(oldtype, &orig_extent);
if (MPI_SUCCESS != rc) goto cleanup;
/* calculate position in grid using row-major ordering */
{
int tmp_rank = rank, procs = size;
coords = (int *) malloc(ndims * sizeof(int));
for (i = 0 ; i < ndims ; i++) {
procs = procs / psize_array[i];
coords[i] = tmp_rank / procs;
tmp_rank = tmp_rank % procs;
}
}
st_offsets = (ptrdiff_t *) malloc(ndims * sizeof(ptrdiff_t));
/* duplicate type to here to 1) deal with constness without
casting and 2) eliminate need to for conditional destroy below.
Lame, yes. But cleaner code all around. */
rc = ompi_datatype_duplicate(oldtype, &lastType);
if (OMPI_SUCCESS != rc) goto cleanup;
/* figure out ordering issues */
if (MPI_ORDER_C == order) {
start_loop = ndims - 1 ; step = -1; end_loop = -1;
} else {
start_loop = 0 ; step = 1; end_loop = ndims;
}
/* Build up array */
for (i = start_loop; i != end_loop; i += step) {
int nprocs, tmp_rank;
switch(distrib_array[i]) {
case MPI_DISTRIBUTE_BLOCK:
rc = block(gsize_array, i, ndims, psize_array[i], coords[i],
darg_array[i], order, orig_extent,
lastType, newtype, st_offsets+i);
break;
case MPI_DISTRIBUTE_CYCLIC:
rc = cyclic(gsize_array, i, ndims, psize_array[i], coords[i],
darg_array[i], order, orig_extent,
lastType, newtype, st_offsets+i);
break;
case MPI_DISTRIBUTE_NONE:
/* treat it as a block distribution on 1 process */
if (order == MPI_ORDER_C) {
nprocs = psize_array[i]; tmp_rank = coords[i];
} else {
nprocs = 1; tmp_rank = 0;
}
rc = block(gsize_array, i, ndims, nprocs, tmp_rank,
MPI_DISTRIBUTE_DFLT_DARG, order, orig_extent,
lastType, newtype, st_offsets+i);
break;
default:
rc = MPI_ERR_ARG;
}
ompi_datatype_destroy(&lastType);
/* need to destroy the old type even in error condition, so
don't check return code from above until after cleanup. */
if (MPI_SUCCESS != rc) goto cleanup;
lastType = *newtype;
}
/* set displacement and UB correctly. Use struct instead of
resized for same reason as subarray */
{
ptrdiff_t displs[3], tmp_size;
ompi_datatype_t *types[3];
int blength[3] = { 1, 1, 1};
displs[1] = st_offsets[start_loop];
tmp_size = 1;
for (i = start_loop + step ; i != end_loop ; i += step) {
tmp_size *= gsize_array[i - step];
displs[1] += tmp_size * st_offsets[i];
}
displs[0] = 0;
displs[1] *= orig_extent;
displs[2] = orig_extent;
for (i = 0 ; i < ndims ; i++) {
displs[2] *= gsize_array[i];
}
types[0] = MPI_LB; types[1] = lastType; types[2] = MPI_UB;
rc = ompi_datatype_create_struct(3, blength, displs, types, newtype);
ompi_datatype_destroy(&lastType);
/* need to destroy the old type even in error condition, so
don't check return code from above until after cleanup. */
if (MPI_SUCCESS != rc) goto cleanup;
}
cleanup:
if (NULL != st_offsets) free(st_offsets);
if (NULL != coords) free(coords);
return OMPI_SUCCESS;
}
static ompi_datatype_t* __ompi_datatype_create_from_args( int32_t* i, MPI_Aint* a,
ompi_datatype_t** d, int32_t type )
{
ompi_datatype_t* datatype = NULL;
switch(type){
/******************************************************************/
case MPI_COMBINER_DUP:
/* should we duplicate d[0]? */
/* ompi_datatype_set_args( datatype, 0, NULL, 0, NULL, 1, d[0], MPI_COMBINER_DUP ); */
assert(0); /* shouldn't happen */
break;
/******************************************************************/
case MPI_COMBINER_CONTIGUOUS:
ompi_datatype_create_contiguous( i[0], d[0], &datatype );
ompi_datatype_set_args( datatype, 1, (const int **) &i, 0, NULL, 1, d, MPI_COMBINER_CONTIGUOUS );
break;
/******************************************************************/
case MPI_COMBINER_VECTOR:
ompi_datatype_create_vector( i[0], i[1], i[2], d[0], &datatype );
{
const int* a_i[3] = {&i[0], &i[1], &i[2]};
ompi_datatype_set_args( datatype, 3, a_i, 0, NULL, 1, d, MPI_COMBINER_VECTOR );
}
break;
/******************************************************************/
case MPI_COMBINER_HVECTOR_INTEGER:
case MPI_COMBINER_HVECTOR:
ompi_datatype_create_hvector( i[0], i[1], a[0], d[0], &datatype );
{
const int* a_i[2] = {&i[0], &i[1]};
ompi_datatype_set_args( datatype, 2, a_i, 1, a, 1, d, MPI_COMBINER_HVECTOR );
}
break;
/******************************************************************/
case MPI_COMBINER_INDEXED: /* TO CHECK */
ompi_datatype_create_indexed( i[0], &(i[1]), &(i[1+i[0]]), d[0], &datatype );
{
const int* a_i[3] = {&i[0], &i[1], &(i[1+i[0]])};
ompi_datatype_set_args( datatype, 2 * i[0] + 1, a_i, 0, NULL, 1, d, MPI_COMBINER_INDEXED );
}
break;
/******************************************************************/
case MPI_COMBINER_HINDEXED_INTEGER:
case MPI_COMBINER_HINDEXED:
ompi_datatype_create_hindexed( i[0], &(i[1]), a, d[0], &datatype );
{
const int* a_i[2] = {&i[0], &i[1]};
ompi_datatype_set_args( datatype, i[0] + 1, a_i, i[0], a, 1, d, MPI_COMBINER_HINDEXED );
}
break;
/******************************************************************/
case MPI_COMBINER_INDEXED_BLOCK:
ompi_datatype_create_indexed_block( i[0], i[1], &(i[2]), d[0], &datatype );
{
const int* a_i[3] = {&i[0], &i[1], &i[2]};
ompi_datatype_set_args( datatype, i[0] + 2, a_i, 0, NULL, 1, d, MPI_COMBINER_INDEXED_BLOCK );
}
break;
/******************************************************************/
case MPI_COMBINER_STRUCT_INTEGER:
case MPI_COMBINER_STRUCT:
ompi_datatype_create_struct( i[0], &(i[1]), a, d, &datatype );
{
const int* a_i[2] = {&i[0], &i[1]};
ompi_datatype_set_args( datatype, i[0] + 1, a_i, i[0], a, i[0], d, MPI_COMBINER_STRUCT );
}
break;
/******************************************************************/
case MPI_COMBINER_SUBARRAY:
ompi_datatype_create_subarray( i[0], &i[1 + 0 * i[0]], &i[1 + 1 * i[0]],
&i[1 + 2 * i[0]], i[1 + 3 * i[0]],
d[0], &datatype );
{
const int* a_i[5] = {&i[0], &i[1 + 0 * i[0]], &i[1 + 1 * i[0]], &i[1 + 2 * i[0]], &i[1 + 3 * i[0]]};
ompi_datatype_set_args( datatype, 3 * i[0] + 2, a_i, 0, NULL, 1, d, MPI_COMBINER_SUBARRAY);
}
break;
/******************************************************************/
case MPI_COMBINER_DARRAY:
ompi_datatype_create_darray( i[0] /* size */, i[1] /* rank */, i[2] /* ndims */,
&i[3 + 0 * i[0]], &i[3 + 1 * i[0]],
&i[3 + 2 * i[0]], &i[3 + 3 * i[0]],
i[3 + 4 * i[0]], d[0], &datatype );
{
const int* a_i[8] = {&i[0], &i[1], &i[2], &i[3 + 0 * i[0]], &i[3 + 1 * i[0]], &i[3 + 2 * i[0]],
&i[3 + 3 * i[0]], &i[3 + 4 * i[0]]};
ompi_datatype_set_args( datatype, 4 * i[0] + 4,a_i, 0, NULL, 1, d, MPI_COMBINER_DARRAY);
}
break;
/******************************************************************/
case MPI_COMBINER_F90_REAL:
case MPI_COMBINER_F90_COMPLEX:
/*pArgs->i[0] = i[0][0];
pArgs->i[1] = i[1][0];
*/
break;
/******************************************************************/
case MPI_COMBINER_F90_INTEGER:
/*pArgs->i[0] = i[0][0];*/
break;
/******************************************************************/
case MPI_COMBINER_RESIZED:
ompi_datatype_create_resized(d[0], a[0], a[1], &datatype);
ompi_datatype_set_args( datatype, 0, NULL, 2, a, 1, d, MPI_COMBINER_RESIZED );
break;
/******************************************************************/
case MPI_COMBINER_HINDEXED_BLOCK:
ompi_datatype_create_hindexed_block( i[0], i[1], a, d[0], &datatype );
{
const int* a_i[2] = {&i[0], &i[1]};
ompi_datatype_set_args( datatype, 2 + i[0], a_i, i[0], a, 1, d, MPI_COMBINER_HINDEXED_BLOCK );
}
break;
/******************************************************************/
default:
break;
}
return datatype;
}
int32_t ompi_datatype_create_subarray(int ndims,
int const* size_array,
int const* subsize_array,
int const* start_array,
int order,
const ompi_datatype_t* oldtype,
ompi_datatype_t** newtype)
{
MPI_Datatype last_type;
int32_t i, step, end_loop;
MPI_Aint size, displ, extent;
/**
* If the oldtype contains the original MPI_LB and MPI_UB markers then we
* are forced to follow the MPI standard suggestion and reset these 2
* markers (MPI 3.0 page 96 line 37). Otherwise we can simply resize the
* datatype.
*/
ompi_datatype_type_extent( oldtype, &extent );
/* If the ndims is zero then return the NULL datatype */
if( ndims < 2 ) {
if( 0 == ndims ) {
*newtype = &ompi_mpi_datatype_null.dt;
return MPI_SUCCESS;
}
ompi_datatype_create_contiguous( subsize_array[0], oldtype, &last_type );
size = size_array[0];
displ = start_array[0];
goto replace_subarray_type;
}
if( MPI_ORDER_C == order ) {
i = ndims - 1;
step = -1;
end_loop = -1;
} else {
i = 0;
step = 1;
end_loop = ndims;
}
/* As we know that the ndims is at least 1 we can start by creating the
* first dimension data outside the loop, such that we dont have to create
* a duplicate of the oldtype just to be able to free it.
*/
ompi_datatype_create_vector( subsize_array[i+step], subsize_array[i], size_array[i],
oldtype, newtype );
last_type = *newtype;
size = (MPI_Aint)size_array[i] * (MPI_Aint)size_array[i+step];
displ = (MPI_Aint)start_array[i] + (MPI_Aint)start_array[i+step] * (MPI_Aint)size_array[i];
for( i += 2 * step; i != end_loop; i += step ) {
ompi_datatype_create_hvector( subsize_array[i], 1, size * extent,
last_type, newtype );
ompi_datatype_destroy( &last_type );
displ += size * start_array[i];
size *= size_array[i];
last_type = *newtype;
}
replace_subarray_type:
/*
* Resized will only set the soft lb and ub markers without moving the real
* data inside. Thus, in case the original data contains the hard markers
* (MPI_LB or MPI_UB) we must force the displacement of the data upward to
* the right position AND set the hard markers LB and UB.
*
* NTH: ompi_datatype_create_resized() does not do enough for the general
* pack/unpack functions to work correctly. Until this is fixed always use
* ompi_datatype_create_struct(). Once this is fixed remove 1 || below. To
* verify that the regression is fixed run the subarray test in the Open MPI
* ibm testsuite.
*/
if(1 || oldtype->super.flags & (OPAL_DATATYPE_FLAG_USER_LB | OPAL_DATATYPE_FLAG_USER_UB) ) {
MPI_Aint displs[3];
MPI_Datatype types[3];
int blength[3] = { 1, 1, 1 };
displs[0] = 0; displs[1] = displ * extent; displs[2] = size * extent;
types[0] = MPI_LB; types[1] = last_type; types[2] = MPI_UB;
ompi_datatype_create_struct( 3, blength, displs, types, newtype );
} else {
ompi_datatype_create_resized(last_type, displ * extent, size * extent, newtype);
}
ompi_datatype_destroy( &last_type );
return OMPI_SUCCESS;
}
/*
* file_open_pvfs2: This is the same strategy as ROMIO's pvfs2 open
*
* Function: - opens a new file
* Accepts: - same arguments as MPI_File_open()
* Returns: - Success if new file handle
*/
int
mca_fs_pvfs2_file_open (struct ompi_communicator_t *comm,
const char* filename,
int access_mode,
struct ompi_info_t *info,
mca_io_ompio_file_t *fh)
{
int ret;
mca_fs_pvfs2 *pvfs2_fs;
PVFS_fs_id pvfs2_id;
char pvfs2_path[OMPIO_MAX_NAME] = {0};
char * ncache_timeout;
open_status o_status = {0, {0, 0}};
struct ompi_datatype_t *open_status_type;
struct ompi_datatype_t *types[2] = {&ompi_mpi_int.dt, &ompi_mpi_byte.dt};
int lens[2] = {1, sizeof(PVFS_object_ref)};
OPAL_PTRDIFF_TYPE offsets[2];
char char_stripe[MPI_MAX_INFO_KEY];
int flag;
int fs_pvfs2_stripe_size = -1;
int fs_pvfs2_stripe_width = -1;
/* We are going to do what ROMIO does with one process resolving
* the name and broadcasting to others */
pvfs2_fs = (mca_fs_pvfs2 *) malloc(sizeof(mca_fs_pvfs2));
if (NULL == pvfs2_fs) {
opal_output (1, "OUT OF MEMORY\n");
return OMPI_ERR_OUT_OF_RESOURCE;
}
if (!mca_fs_pvfs2_IS_INITIALIZED) {
/* disable the pvfs2 ncache */
ncache_timeout = getenv("PVFS2_NCACHE_TIMEOUT");
if (ncache_timeout == NULL ) {
setenv("PVFS2_NCACHE_TIMEOUT", "0", 1);
}
ret = PVFS_util_init_defaults();
if (ret < 0) {
PVFS_perror("PVFS_util_init_defaults", ret);
return OMPI_ERROR;
}
mca_fs_pvfs2_IS_INITIALIZED = 1;
}
memset(&(pvfs2_fs->credentials), 0, sizeof(PVFS_credentials));
PVFS_util_gen_credentials(&(pvfs2_fs->credentials));
/* check for stripe size and stripe depth in the info object and
update mca_fs_pvfs2_stripe_width and mca_fs_pvfs2_stripe_size
before calling fake_an_open() */
ompi_info_get (info, "stripe_size", MPI_MAX_INFO_VAL, char_stripe, &flag);
if ( flag ) {
sscanf ( char_stripe, "%d", &fs_pvfs2_stripe_size );
}
ompi_info_get (info, "stripe_width", MPI_MAX_INFO_VAL, char_stripe, &flag);
if ( flag ) {
sscanf ( char_stripe, "%d", &fs_pvfs2_stripe_width );
}
if (fs_pvfs2_stripe_size < 0) {
fs_pvfs2_stripe_size = mca_fs_pvfs2_stripe_size;
}
if (fs_pvfs2_stripe_width < 0) {
fs_pvfs2_stripe_width = mca_fs_pvfs2_stripe_width;
}
if (OMPIO_ROOT == fh->f_rank) {
ret = PVFS_util_resolve(filename, &pvfs2_id, pvfs2_path, OMPIO_MAX_NAME);
if (ret < 0 ) {
PVFS_perror("PVFS_util_resolve", ret);
o_status.error = -1;
}
else {
fake_an_open (pvfs2_id,
pvfs2_path,
access_mode,
fs_pvfs2_stripe_width,
(PVFS_size)fs_pvfs2_stripe_size,
pvfs2_fs,
&o_status);
}
pvfs2_fs->object_ref = o_status.object_ref;
fh->f_fs_ptr = pvfs2_fs;
}
/* broadcast status and (possibly valid) object reference */
offsets[0] = (MPI_Aint)(&o_status.error);
offsets[1] = (MPI_Aint)(&o_status.object_ref);
ompi_datatype_create_struct (2, lens, offsets, types, &open_status_type);
ompi_datatype_commit (&open_status_type);
fh->f_comm->c_coll.coll_bcast (MPI_BOTTOM,
1,
open_status_type,
OMPIO_ROOT,
fh->f_comm,
fh->f_comm->c_coll.coll_bcast_module);
ompi_datatype_destroy (&open_status_type);
if (o_status.error != 0) {
/* No need to free the pvfs2_fs structure, since it will
be deallocated in file_close in case of an error */
fh->f_fs_ptr = NULL;
return OMPI_ERROR;
}
pvfs2_fs->object_ref = o_status.object_ref;
fh->f_fs_ptr = pvfs2_fs;
/* update the internal ompio structure to store stripe
size and stripe depth correctly.
Hadi(to be done): For this read the stripe size and stripe depth from
the file itself
*/
if (fs_pvfs2_stripe_size > 0 && fs_pvfs2_stripe_width > 0) {
fh->f_stripe_size = fs_pvfs2_stripe_size;
fh->f_stripe_count = fs_pvfs2_stripe_width;
}
return OMPI_SUCCESS;
}
int32_t ompi_datatype_create_subarray(int ndims,
int const* size_array,
int const* subsize_array,
int const* start_array,
int order,
const ompi_datatype_t* oldtype,
ompi_datatype_t** newtype)
{
MPI_Datatype last_type;
int32_t i, step, end_loop;
MPI_Aint size, displ, extent;
ompi_datatype_type_extent( oldtype, &extent );
/* If the ndims is zero then return the NULL datatype */
if( ndims < 2 ) {
if( 0 == ndims ) {
*newtype = &ompi_mpi_datatype_null.dt;
return MPI_SUCCESS;
}
ompi_datatype_create_contiguous( subsize_array[0], oldtype, &last_type );
size = size_array[0];
displ = start_array[0];
goto replace_subarray_type;
}
if( MPI_ORDER_C == order ) {
i = ndims - 1;
step = -1;
end_loop = -1;
} else {
i = 0;
step = 1;
end_loop = ndims;
}
/* As we know that the ndims is at least 1 we can start by creating the
* first dimension data outside the loop, such that we dont have to create
* a duplicate of the oldtype just to be able to free it.
*/
ompi_datatype_create_vector( subsize_array[i+step], subsize_array[i], size_array[i],
oldtype, newtype );
last_type = *newtype;
size = size_array[i] * size_array[i+step];
displ = start_array[i] + start_array[i+step] * size_array[i];
for( i += 2 * step; i != end_loop; i += step ) {
ompi_datatype_create_hvector( subsize_array[i], 1, size * extent,
last_type, newtype );
ompi_datatype_destroy( &last_type );
displ += size * start_array[i];
size *= size_array[i];
last_type = *newtype;
}
replace_subarray_type:
/**
* We cannot use resized here. Resized will only set the soft lb and ub markers
* without moving the real data inside. What we need is to force the displacement
* of the data upward to the right position AND set the LB and UB. A type
* struct is the function we need.
*/
{
MPI_Aint displs[3];
MPI_Datatype types[3];
int blength[3] = { 1, 1, 1 };
displs[0] = 0;
displs[1] = displ * extent;
displs[2] = size * extent;
types[0] = MPI_LB;
types[1] = last_type;
types[2] = MPI_UB;
ompi_datatype_create_struct( 3, blength, displs, types, newtype );
}
ompi_datatype_destroy( &last_type );
return OMPI_SUCCESS;
}
int
mca_fcoll_static_file_write_all (mca_io_ompio_file_t *fh,
void *buf,
int count,
struct ompi_datatype_t *datatype,
ompi_status_public_t *status)
{
size_t max_data = 0, bytes_per_cycle=0;
struct iovec *iov=NULL, *decoded_iov=NULL;
uint32_t iov_count=0, iov_index=0;
int i=0,j=0,l=0, temp_index;
int ret=OMPI_SUCCESS, cycles, local_cycles, *bytes_per_process=NULL;
int index, *disp_index=NULL, **blocklen_per_process=NULL;
int *iovec_count_per_process=NULL, *displs=NULL;
size_t total_bytes_written=0;
MPI_Aint **displs_per_process=NULL, *memory_displacements=NULL;
MPI_Aint bytes_to_write_in_cycle=0, global_iov_count=0, global_count=0;
local_io_array *local_iov_array =NULL, *global_iov_array=NULL;
local_io_array *file_offsets_for_agg=NULL;
int *sorted=NULL, *sorted_file_offsets=NULL, temp_pindex, *temp_disp_index=NULL;
char *send_buf=NULL, *global_buf=NULL;
int iov_size=0, current_position=0, *current_index=NULL;
int *bytes_remaining=NULL, entries_per_aggregator=0;
ompi_datatype_t **recvtype = NULL;
MPI_Request *send_req=NULL, *recv_req=NULL;
/* For creating datatype of type io_array */
int blocklen[3] = {1, 1, 1};
int static_num_io_procs=1;
OPAL_PTRDIFF_TYPE d[3], base;
ompi_datatype_t *types[3];
ompi_datatype_t *io_array_type=MPI_DATATYPE_NULL;
/*----------------------------------------------*/
#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 DEBUG_ON
MPI_Aint gc_in;
#endif
// if (opal_datatype_is_contiguous_memory_layout(&datatype->super,1)) {
// fh->f_flags |= OMPIO_CONTIGUOUS_MEMORY;
// }
/* In case the data is not contigous in memory, decode it into an iovec */
if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
fh->f_decode_datatype ((struct mca_io_ompio_file_t *)fh,
datatype,
count,
buf,
&max_data,
&decoded_iov,
&iov_count);
}
else {
max_data = count * datatype->super.size;
}
if ( MPI_STATUS_IGNORE != status ) {
status->_ucount = max_data;
}
fh->f_get_num_aggregators ( & static_num_io_procs );
fh->f_set_aggregator_props ((struct mca_io_ompio_file_t *)fh,
static_num_io_procs,
max_data);
/* io_array datatype for using in communication*/
types[0] = &ompi_mpi_long.dt;
types[1] = &ompi_mpi_long.dt;
types[2] = &ompi_mpi_int.dt;
d[0] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0];
d[1] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0].length;
d[2] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0].process_id;
base = d[0];
for (i=0 ; i<3 ; i++) {
d[i] -= base;
}
ompi_datatype_create_struct (3,
blocklen,
d,
types,
&io_array_type);
ompi_datatype_commit (&io_array_type);
/* #########################################################*/
ret = fh->f_generate_current_file_view((struct mca_io_ompio_file_t *)fh,
max_data,
&iov,
&iov_size);
if (ret != OMPI_SUCCESS){
fprintf(stderr,"Current File View Generation Error\n");
goto exit;
}
if (0 == iov_size){
iov_size = 1;
}
local_iov_array = (local_io_array *)malloc (iov_size * sizeof(local_io_array));
if ( NULL == local_iov_array){
fprintf(stderr,"local_iov_array allocation error\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for (j=0; j < iov_size; j++){
local_iov_array[j].offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t)
iov[j].iov_base;
local_iov_array[j].length = (size_t)iov[j].iov_len;
local_iov_array[j].process_id = fh->f_rank;
}
fh->f_get_bytes_per_agg ( (int *) &bytes_per_cycle);
local_cycles = ceil((double)max_data/bytes_per_cycle);
ret = fh->f_comm->c_coll.coll_allreduce (&local_cycles,
&cycles,
1,
MPI_INT,
MPI_MAX,
fh->f_comm,
fh->f_comm->c_coll.coll_allreduce_module);
if (OMPI_SUCCESS != ret){
fprintf(stderr,"local cycles allreduce!\n");
goto exit;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
disp_index = (int *)malloc (fh->f_procs_per_group * sizeof (int));
if (NULL == disp_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
bytes_per_process = (int *) malloc (fh->f_procs_per_group * sizeof(int ));
if (NULL == bytes_per_process){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
bytes_remaining = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == bytes_remaining){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
current_index = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == current_index){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
blocklen_per_process = (int **)malloc (fh->f_procs_per_group * sizeof (int*));
if (NULL == blocklen_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs_per_process = (MPI_Aint **)
malloc (fh->f_procs_per_group * sizeof (MPI_Aint*));
if (NULL == displs_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for(i=0;i<fh->f_procs_per_group;i++){
current_index[i] = 0;
bytes_remaining[i] =0;
blocklen_per_process[i] = NULL;
displs_per_process[i] = NULL;
}
}
iovec_count_per_process = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == iovec_count_per_process){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == displs){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
ret = fh->f_allgather_array (&iov_size,
1,
MPI_INT,
iovec_count_per_process,
1,
MPI_INT,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if( OMPI_SUCCESS != ret){
fprintf(stderr,"iov size allgatherv array!\n");
goto exit;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
displs[0] = 0;
global_iov_count = iovec_count_per_process[0];
for (i=1 ; i<fh->f_procs_per_group ; i++) {
global_iov_count += iovec_count_per_process[i];
displs[i] = displs[i-1] + iovec_count_per_process[i-1];
}
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
global_iov_array = (local_io_array *) malloc (global_iov_count *
sizeof(local_io_array));
if (NULL == global_iov_array){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
ret = fh->f_gatherv_array (local_iov_array,
iov_size,
io_array_type,
global_iov_array,
iovec_count_per_process,
displs,
io_array_type,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if (OMPI_SUCCESS != ret){
fprintf(stderr,"global_iov_array gather error!\n");
goto exit;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
if ( 0 == global_iov_count){
global_iov_count = 1;
}
sorted = (int *)malloc (global_iov_count * sizeof(int));
if (NULL == sorted) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
local_heap_sort (global_iov_array, global_iov_count, sorted);
}
#if DEBUG_ON
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
for (gc_in=0; gc_in<global_iov_count; gc_in++){
printf("%d: Offset[%ld]: %lld, Length[%ld]: %ld\n",
global_iov_array[gc_in].process_id,
gc_in, global_iov_array[gc_in].offset,
gc_in, global_iov_array[gc_in].length);
}
}
#endif
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_exch = MPI_Wtime();
#endif
for (index = 0; index < cycles; index++){
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
if (NULL == recvtype){
recvtype = (ompi_datatype_t **)
malloc (fh->f_procs_per_group * sizeof(ompi_datatype_t *));
if (NULL == recvtype) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
for(l=0;l<fh->f_procs_per_group;l++){
disp_index[l] = 1;
if (NULL != blocklen_per_process[l]){
free(blocklen_per_process[l]);
blocklen_per_process[l] = NULL;
}
if (NULL != displs_per_process[l]){
free(displs_per_process[l]);
displs_per_process[l] = NULL;
}
blocklen_per_process[l] = (int *) calloc (1, sizeof(int));
if (NULL == blocklen_per_process[l]) {
opal_output (1, "OUT OF MEMORY for blocklen\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs_per_process[l] = (MPI_Aint *) calloc (1, sizeof(MPI_Aint));
if (NULL == displs_per_process[l]){
opal_output (1, "OUT OF MEMORY for displs\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
if (NULL != sorted_file_offsets){
free(sorted_file_offsets);
sorted_file_offsets = NULL;
}
if(NULL != file_offsets_for_agg){
free(file_offsets_for_agg);
file_offsets_for_agg = NULL;
}
if (NULL != memory_displacements){
free(memory_displacements);
memory_displacements = NULL;
}
}
if (local_cycles > index) {
if ((index == local_cycles-1) && (max_data % bytes_per_cycle)) {
bytes_to_write_in_cycle = max_data % bytes_per_cycle;
}
else if (max_data <= bytes_per_cycle) {
bytes_to_write_in_cycle = max_data;
}
else {
bytes_to_write_in_cycle = bytes_per_cycle;
}
}
else {
bytes_to_write_in_cycle = 0;
}
#if DEBUG_ON
/* if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {*/
printf ("***%d: CYCLE %d Bytes %ld**********\n",
fh->f_rank,
index,
bytes_to_write_in_cycle);
/* }*/
#endif
/**********************************************************
**Gather the Data from all the processes at the writers **
*********************************************************/
/* gather from each process how many bytes each will be sending */
fh->f_gather_array (&bytes_to_write_in_cycle,
1,
MPI_INT,
bytes_per_process,
1,
MPI_INT,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
/*
For each aggregator
it needs to get bytes_to_write_in_cycle from each process
in group which adds up to bytes_per_cycle
*/
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
for (i=0;i<fh->f_procs_per_group; i++){
/* printf("bytes_per_process[%d]: %d\n", i, bytes_per_process[i]);
*/
#if DEBUG_ON
printf ("%d : bytes_per_process : %d\n",
fh->f_procs_in_group[i],
bytes_per_process[i]);
#endif
while (bytes_per_process[i] > 0){
if (get_process_id(global_iov_array[sorted[current_index[i]]].process_id,
fh) == i){ /* current id owns this entry!*/
/*Add and subtract length and create
blocklength and displs array*/
if (bytes_remaining[i]){ /*Remaining bytes in the current entry of
the global offset array*/
if (bytes_remaining[i] <= bytes_per_process[i]){
blocklen_per_process[i][disp_index[i] - 1] = bytes_remaining[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset +
(global_iov_array[sorted[current_index[i]]].length
- bytes_remaining[i]);
blocklen_per_process[i] = (int *) realloc
((void *)blocklen_per_process[i], (disp_index[i]+1)*sizeof(int));
displs_per_process[i] = (MPI_Aint *)realloc
((void *)displs_per_process[i], (disp_index[i]+1)*sizeof(MPI_Aint));
bytes_per_process[i] -= bytes_remaining[i];
blocklen_per_process[i][disp_index[i]] = 0;
displs_per_process[i][disp_index[i]] = 0;
bytes_remaining[i] = 0;
disp_index[i] += 1;
/* This entry has been used up, we need to move to the
next entry of this process and make current_index point there*/
current_index[i] = find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1){
/* No more entries left, so Its all done! exit!*/
break;
}
continue;
}
else{
blocklen_per_process[i][disp_index[i] - 1] = bytes_per_process[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset +
(global_iov_array[sorted[current_index[i]]].length
- bytes_remaining[i]);
bytes_remaining[i] -= bytes_per_process[i];
bytes_per_process[i] = 0;
break;
}
}
else{
if (bytes_per_process[i] <
global_iov_array[sorted[current_index[i]]].length){
blocklen_per_process[i][disp_index[i] - 1] =
bytes_per_process[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset;
bytes_remaining[i] =
global_iov_array[sorted[current_index[i]]].length -
bytes_per_process[i];
bytes_per_process[i] = 0;
break;
}
else {
blocklen_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].length;
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset;
blocklen_per_process[i] =
(int *) realloc ((void *)blocklen_per_process[i], (disp_index[i]+1)*sizeof(int));
displs_per_process[i] = (MPI_Aint *)realloc
((void *)displs_per_process[i], (disp_index[i]+1)*sizeof(MPI_Aint));
blocklen_per_process[i][disp_index[i]] = 0;
displs_per_process[i][disp_index[i]] = 0;
disp_index[i] += 1;
bytes_per_process[i] -=
global_iov_array[sorted[current_index[i]]].length;
current_index[i] = find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1){
break;
}
}
}
}
else{
current_index[i] = find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1){
bytes_per_process[i] = 0; /* no more entries left
to service this request*/
continue;
}
}
}
}
entries_per_aggregator=0;
for (i=0;i<fh->f_procs_per_group;i++){
for (j=0;j<disp_index[i];j++){
if (blocklen_per_process[i][j] > 0){
entries_per_aggregator++;
#if DEBUG_ON
printf("%d sends blocklen[%d]: %d, disp[%d]: %ld to %d\n",
fh->f_procs_in_group[i],j,
blocklen_per_process[i][j],j,
displs_per_process[i][j],
fh->f_rank);
#endif
}
}
}
if (entries_per_aggregator > 0){
file_offsets_for_agg = (local_io_array *)
malloc(entries_per_aggregator*sizeof(local_io_array));
if (NULL == file_offsets_for_agg) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
sorted_file_offsets = (int *)
malloc (entries_per_aggregator*sizeof(int));
if (NULL == sorted_file_offsets){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
temp_index = 0;
for (i=0;i<fh->f_procs_per_group; i++){
for(j=0;j<disp_index[i];j++){
if (blocklen_per_process[i][j] > 0){
file_offsets_for_agg[temp_index].length =
blocklen_per_process[i][j];
file_offsets_for_agg[temp_index].process_id = i;
file_offsets_for_agg[temp_index].offset =
displs_per_process[i][j];
temp_index++;
}
}
}
}
else{
continue;
}
local_heap_sort (file_offsets_for_agg,
entries_per_aggregator,
sorted_file_offsets);
memory_displacements = (MPI_Aint *) malloc
(entries_per_aggregator * sizeof(MPI_Aint));
memory_displacements[sorted_file_offsets[0]] = 0;
for (i=1; i<entries_per_aggregator; i++){
memory_displacements[sorted_file_offsets[i]] =
memory_displacements[sorted_file_offsets[i-1]] +
file_offsets_for_agg[sorted_file_offsets[i-1]].length;
}
temp_disp_index = (int *)calloc (1, fh->f_procs_per_group * sizeof (int));
if (NULL == temp_disp_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
global_count = 0;
for (i=0;i<entries_per_aggregator;i++){
temp_pindex =
file_offsets_for_agg[sorted_file_offsets[i]].process_id;
displs_per_process[temp_pindex][temp_disp_index[temp_pindex]] =
memory_displacements[sorted_file_offsets[i]];
if (temp_disp_index[temp_pindex] < disp_index[temp_pindex])
temp_disp_index[temp_pindex] += 1;
else{
printf("temp_disp_index[%d]: %d is greater than disp_index[%d]: %d\n",
temp_pindex, temp_disp_index[temp_pindex],
temp_pindex, disp_index[temp_pindex]);
}
global_count +=
file_offsets_for_agg[sorted_file_offsets[i]].length;
}
if (NULL != temp_disp_index){
free(temp_disp_index);
temp_disp_index = NULL;
}
#if DEBUG_ON
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
for (i=0; i<entries_per_aggregator;i++){
printf("%d: OFFSET: %lld LENGTH: %ld, Mem-offset: %ld, disp : %d\n",
file_offsets_for_agg[sorted_file_offsets[i]].process_id,
file_offsets_for_agg[sorted_file_offsets[i]].offset,
file_offsets_for_agg[sorted_file_offsets[i]].length,
memory_displacements[sorted_file_offsets[i]],
disp_index[ file_offsets_for_agg[sorted_file_offsets[i]].process_id]);
}
#endif
#if DEBUG_ON
printf("%d: global_count : %ld, bytes_to_write_in_cycle : %ld, procs_per_group: %d\n",
fh->f_rank,
global_count,
bytes_to_write_in_cycle,
fh->f_procs_per_group);
#endif
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_comm_time = MPI_Wtime();
#endif
global_buf = (char *) malloc (global_count);
if (NULL == global_buf){
opal_output(1, "OUT OF MEMORY");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
recv_req = (MPI_Request *)
malloc (fh->f_procs_per_group * sizeof(MPI_Request));
if (NULL == recv_req){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for (i=0;i<fh->f_procs_per_group; i++){
ompi_datatype_create_hindexed(disp_index[i],
blocklen_per_process[i],
displs_per_process[i],
MPI_BYTE,
&recvtype[i]);
ompi_datatype_commit(&recvtype[i]);
ret = MCA_PML_CALL(irecv(global_buf,
1,
recvtype[i],
fh->f_procs_in_group[i],
123,
fh->f_comm,
&recv_req[i]));
if (OMPI_SUCCESS != ret){
fprintf(stderr,"irecv Error!\n");
goto exit;
}
}
}
if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) {
send_buf = &((char*)buf)[total_bytes_written];
}
else if (bytes_to_write_in_cycle) {
/* allocate a send buffer and copy the data that needs
to be sent into it in case the data is non-contigous
in memory */
OPAL_PTRDIFF_TYPE mem_address;
size_t remaining = 0;
size_t temp_position = 0;
send_buf = malloc (bytes_to_write_in_cycle);
if (NULL == send_buf) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
remaining = bytes_to_write_in_cycle;
while (remaining) {
mem_address = (OPAL_PTRDIFF_TYPE)
(decoded_iov[iov_index].iov_base) + current_position;
if (remaining >=
(decoded_iov[iov_index].iov_len - current_position)) {
memcpy (send_buf+temp_position,
(IOVBASE_TYPE *)mem_address,
decoded_iov[iov_index].iov_len - current_position);
remaining = remaining -
(decoded_iov[iov_index].iov_len - current_position);
temp_position = temp_position +
(decoded_iov[iov_index].iov_len - current_position);
iov_index = iov_index + 1;
current_position = 0;
}
else {
memcpy (send_buf+temp_position,
(IOVBASE_TYPE *)mem_address,
remaining);
current_position = current_position + remaining;
remaining = 0;
}
}
}
total_bytes_written += bytes_to_write_in_cycle;
send_req = (MPI_Request *) malloc (sizeof(MPI_Request));
if (NULL == send_req){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
ret = MCA_PML_CALL(isend(send_buf,
bytes_to_write_in_cycle,
MPI_BYTE,
fh->f_procs_in_group[fh->f_aggregator_index],
123,
MCA_PML_BASE_SEND_STANDARD,
fh->f_comm,
send_req));
if ( OMPI_SUCCESS != ret ){
fprintf(stderr,"isend error!\n");
goto exit;
}
ret = ompi_request_wait (send_req, MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret){
goto exit;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
ret = ompi_request_wait_all (fh->f_procs_per_group,
recv_req,
MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret){
goto exit;
}
#if DEBUG_ON
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank){
for (i=0 ; i<global_count/4 ; i++)
printf (" RECV %d \n",((int *)global_buf)[i]);
}
#endif
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_comm_time = MPI_Wtime();
comm_time += end_comm_time - start_comm_time;
#endif
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
fh->f_io_array = (mca_io_ompio_io_array_t *) malloc
(entries_per_aggregator * sizeof (mca_io_ompio_io_array_t));
if (NULL == fh->f_io_array) {
opal_output(1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
fh->f_num_of_io_entries = 0;
/*First entry for every aggregator*/
fh->f_io_array[fh->f_num_of_io_entries].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[0]].offset;
fh->f_io_array[fh->f_num_of_io_entries].length =
file_offsets_for_agg[sorted_file_offsets[0]].length;
fh->f_io_array[fh->f_num_of_io_entries].memory_address =
global_buf+memory_displacements[sorted_file_offsets[0]];
fh->f_num_of_io_entries++;
for (i=1;i<entries_per_aggregator;i++){
if (file_offsets_for_agg[sorted_file_offsets[i-1]].offset +
file_offsets_for_agg[sorted_file_offsets[i-1]].length ==
file_offsets_for_agg[sorted_file_offsets[i]].offset){
fh->f_io_array[fh->f_num_of_io_entries - 1].length +=
file_offsets_for_agg[sorted_file_offsets[i]].length;
}
else {
fh->f_io_array[fh->f_num_of_io_entries].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[i]].offset;
fh->f_io_array[fh->f_num_of_io_entries].length =
file_offsets_for_agg[sorted_file_offsets[i]].length;
fh->f_io_array[fh->f_num_of_io_entries].memory_address =
global_buf+memory_displacements[sorted_file_offsets[i]];
fh->f_num_of_io_entries++;
}
}
#if DEBUG_ON
printf("*************************** %d\n", fh->f_num_of_io_entries);
for (i=0 ; i<fh->f_num_of_io_entries ; i++) {
printf(" ADDRESS: %p OFFSET: %ld LENGTH: %ld\n",
fh->f_io_array[i].memory_address,
(OPAL_PTRDIFF_TYPE)fh->f_io_array[i].offset,
fh->f_io_array[i].length);
}
#endif
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_write_time = MPI_Wtime();
#endif
if (fh->f_num_of_io_entries) {
if ( 0 > fh->f_fbtl->fbtl_pwritev (fh)) {
opal_output (1, "WRITE FAILED\n");
ret = OMPI_ERROR;
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_write_time = MPI_Wtime();
write_time += end_write_time - start_write_time;
#endif
}
if (NULL != send_req){
free(send_req);
send_req = NULL;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
fh->f_num_of_io_entries = 0;
if (NULL != fh->f_io_array) {
free (fh->f_io_array);
fh->f_io_array = NULL;
}
for (i = 0; i < fh->f_procs_per_group; i++)
ompi_datatype_destroy(recvtype+i);
if (NULL != recvtype){
free(recvtype);
recvtype=NULL;
}
if (NULL != recv_req){
free(recv_req);
recv_req = NULL;
}
if (NULL != global_buf) {
free (global_buf);
global_buf = NULL;
}
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_exch = MPI_Wtime();
exch_write += end_exch - start_exch;
nentry.time[0] = write_time;
nentry.time[1] = comm_time;
nentry.time[2] = exch_write;
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank)
nentry.aggregator = 1;
else
nentry.aggregator = 0;
nentry.nprocs_for_coll = static_num_io_procs;
if (!fh->f_full_print_queue(WRITE_PRINT_QUEUE)){
fh->f_register_print_entry(WRITE_PRINT_QUEUE,
nentry);
}
#endif
exit:
if (NULL != decoded_iov){
free(decoded_iov);
decoded_iov = NULL;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
if (NULL != local_iov_array){
free(local_iov_array);
local_iov_array = NULL;
}
for(l=0;l<fh->f_procs_per_group;l++){
if (NULL != blocklen_per_process[l]){
free(blocklen_per_process[l]);
blocklen_per_process[l] = NULL;
}
if (NULL != displs_per_process[l]){
free(displs_per_process[l]);
displs_per_process[l] = NULL;
}
}
}
if (NULL != send_buf){
free(send_buf);
send_buf = NULL;
}
if (NULL != global_buf){
free(global_buf);
global_buf = NULL;
}
if (NULL != recvtype){
free(recvtype);
recvtype = NULL;
}
if (NULL != sorted_file_offsets){
free(sorted_file_offsets);
sorted_file_offsets = NULL;
}
if (NULL != file_offsets_for_agg){
free(file_offsets_for_agg);
file_offsets_for_agg = NULL;
}
if (NULL != memory_displacements){
free(memory_displacements);
memory_displacements = NULL;
}
if (NULL != displs_per_process){
free(displs_per_process);
displs_per_process = NULL;
}
if (NULL != blocklen_per_process){
free(blocklen_per_process);
blocklen_per_process = NULL;
}
if(NULL != current_index){
free(current_index);
current_index = NULL;
}
if(NULL != bytes_remaining){
free(bytes_remaining);
bytes_remaining = NULL;
}
if (NULL != disp_index){
free(disp_index);
disp_index = NULL;
}
if (NULL != sorted) {
free(sorted);
sorted = NULL;
}
return ret;
}
static int unpack_ooo(void)
{
ompi_datatype_t * t1;
ompi_datatype_t * t2;
ompi_datatype_t * type[2];
ompi_datatype_t * newtype;
MPI_Aint disp[2];
int len[2], rc;
rc = ompi_datatype_create_vector(2, 1, 2, MPI_INT, &t1);
if (OMPI_SUCCESS != rc) {
fprintf(stderr, "could not create vector t1\n");
return 1;
}
rc = ompi_datatype_commit (&t1);
if (OMPI_SUCCESS != rc) {
fprintf(stderr, "could not commit vector t1\n");
return 1;
}
rc = ompi_datatype_create_vector(2, 1, 2, MPI_DOUBLE, &t2);
if (OMPI_SUCCESS != rc) {
fprintf(stderr, "could not create vector t2\n");
return 1;
}
rc = ompi_datatype_commit (&t2);
if (OMPI_SUCCESS != rc) {
fprintf(stderr, "could not commit vector t2\n");
return 1;
}
/*
* btl=0x7f7823672580 bytes_received=992 data_offset=0
* btl=0x7f7823260420 bytes_received=1325 data_offset=992
* btl=0x7f7823672580 bytes_received=992 data_offset=2317
* btl=0x7f7823672580 bytes_received=992 data_offset=3309
* btl=0x7f7823672580 bytes_received=992 data_offset=4301
* btl=0x7f7823672580 bytes_received=992 data_offset=5293
* btl=0x7f7823672580 bytes_received=992 data_offset=6285
* btl=0x7f7823672580 bytes_received=667 data_offset=7277
*/
size_t test1[9][2] = {
{992, 0},
{1325, 992},
{992, 2317},
{992, 3309},
{992, 4301},
{992, 5293},
{992, 6285},
{667, 7277},
{0, -1},
};
/*
* btl=0x7f80bc545580 bytes_received=992 data_offset=0
* btl=0x7f80bc545580 bytes_received=992 data_offset=2317
* btl=0x7f80bc545580 bytes_received=992 data_offset=3309
* btl=0x7f80bc545580 bytes_received=992 data_offset=4301
* btl=0x7f80bc545580 bytes_received=992 data_offset=5293
* btl=0x7f80bc545580 bytes_received=992 data_offset=6285
* btl=0x7f80bc133420 bytes_received=1325 data_offset=992
* btl=0x7f80bc545580 bytes_received=667 data_offset=7277
*/
size_t test2[9][2] = {
{992, 0},
{992, 2317},
{992, 3309},
{992, 4301},
{992, 5293},
{992, 6285},
{1325, 992},
{667, 7277},
{0, -1},
};
/* trimmed version of test2 */
size_t test3[9][2] = {
{992, 0},
{4960, 2317},
{1325, 992},
{667, 7277},
{0, -1},
};
/* an other test case */
size_t test4[9][2] = {
{992, 0},
{992, 2976},
{992, 1984},
{992, 992},
{3976, 3968},
{0, -1},
};
disp[0] = (long)(&foo.i[0]) - (long)&foo;
disp[1] = (long)(&foo.d[0]) - (long)&foo;
type[0] = t1;
type[1] = t2;
len[0] = 1;
len[1] = 1;
rc = ompi_datatype_create_struct(2, len, disp, type, &newtype);
if (OMPI_SUCCESS != rc) {
fprintf(stderr, "could not create struct\n");
return 1;
}
rc = ompi_datatype_commit (&newtype);
if (OMPI_SUCCESS != rc) {
fprintf(stderr, "could not create struct\n");
return 1;
}
pbar = (struct pfoo_t *)malloc (N * sizeof(struct pfoo_t));
if (NULL == pbar) {
fprintf(stderr, "could not malloc pbar\n");
return 1;
}
bar = (struct foo_t *)malloc (N * sizeof(struct foo_t));
if (NULL == bar) {
fprintf(stderr, "could not malloc bar\n");
return 1;
}
if (0 != testcase(newtype, test1)) {
printf ("test1 failed\n");
return 2;
}
if (0 != testcase(newtype, test2)) {
printf ("test2 failed\n");
return 2;
}
if (0 != testcase(newtype, test3)) {
printf ("test3 failed\n");
return 2;
}
if (0 != testcase(newtype, test4)) {
printf ("test4 failed\n");
return 2;
}
/* test the automatic destruction pf the data */
ompi_datatype_destroy( &newtype ); assert( newtype == NULL );
return rc;
}
int mca_common_ompio_set_file_defaults (mca_io_ompio_file_t *fh)
{
if (NULL != fh) {
ompi_datatype_t *types[2];
int blocklen[2] = {1, 1};
OPAL_PTRDIFF_TYPE d[2], base;
int i;
fh->f_io_array = NULL;
fh->f_perm = OMPIO_PERM_NULL;
fh->f_flags = 0;
fh->f_bytes_per_agg = mca_io_ompio_bytes_per_agg;
fh->f_datarep = strdup ("native");
fh->f_offset = 0;
fh->f_disp = 0;
fh->f_position_in_file_view = 0;
fh->f_index_in_file_view = 0;
fh->f_total_bytes = 0;
fh->f_init_procs_per_group = -1;
fh->f_init_procs_in_group = NULL;
fh->f_procs_per_group = -1;
fh->f_procs_in_group = NULL;
fh->f_init_num_aggrs = -1;
fh->f_init_aggr_list = NULL;
/* Default file View */
fh->f_iov_type = MPI_DATATYPE_NULL;
fh->f_stripe_size = mca_io_ompio_bytes_per_agg;
/*Decoded iovec of the file-view*/
fh->f_decoded_iov = NULL;
fh->f_etype = NULL;
fh->f_filetype = NULL;
fh->f_orig_filetype = NULL;
mca_common_ompio_set_view(fh,
0,
&ompi_mpi_byte.dt,
&ompi_mpi_byte.dt,
"native",
fh->f_info);
/*Create a derived datatype for the created iovec */
types[0] = &ompi_mpi_long.dt;
types[1] = &ompi_mpi_long.dt;
d[0] = (OPAL_PTRDIFF_TYPE) fh->f_decoded_iov;
d[1] = (OPAL_PTRDIFF_TYPE) &fh->f_decoded_iov[0].iov_len;
base = d[0];
for (i=0 ; i<2 ; i++) {
d[i] -= base;
}
ompi_datatype_create_struct (2,
blocklen,
d,
types,
&fh->f_iov_type);
ompi_datatype_commit (&fh->f_iov_type);
return OMPI_SUCCESS;
}
else {
return OMPI_ERROR;
}
}
int
mca_fcoll_static_file_read_all (mca_io_ompio_file_t *fh,
void *buf,
int count,
struct ompi_datatype_t *datatype,
ompi_status_public_t *status)
{
int ret = OMPI_SUCCESS, iov_size=0, *bytes_remaining=NULL;
int i, j, l,cycles=0, local_cycles=0, *current_index=NULL;
int index, *disp_index=NULL, *bytes_per_process=NULL, current_position=0;
int **blocklen_per_process=NULL, *iovec_count_per_process=NULL;
int *displs=NULL, *sorted=NULL ,entries_per_aggregator=0;
int *sorted_file_offsets=NULL, temp_index=0, position=0, *temp_disp_index=NULL;
MPI_Aint **displs_per_process=NULL, global_iov_count=0, global_count=0;
MPI_Aint *memory_displacements=NULL;
int bytes_to_read_in_cycle=0;
size_t max_data=0, bytes_per_cycle=0;
uint32_t iov_count=0, iov_index=0;
struct iovec *decoded_iov=NULL, *iov=NULL;
mca_fcoll_static_local_io_array *local_iov_array=NULL, *global_iov_array=NULL;
mca_fcoll_static_local_io_array *file_offsets_for_agg=NULL;
char *global_buf=NULL, *receive_buf=NULL;
int blocklen[3] = {1, 1, 1};
int static_num_io_procs=1;
OPAL_PTRDIFF_TYPE d[3], base;
ompi_datatype_t *types[3];
ompi_datatype_t *io_array_type=MPI_DATATYPE_NULL;
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;
double rcomm_time = 0.0, start_rcomm_time = 0.0, end_rcomm_time = 0.0;
double read_exch = 0.0, start_rexch = 0.0, end_rexch = 0.0;
mca_common_ompio_print_entry nentry;
#endif
#if DEBUG_ON
MPI_Aint gc_in;
#endif
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 ( ( ftype_extent == (OPAL_PTRDIFF_TYPE) ftype_size) &&
opal_datatype_is_contiguous_memory_layout(&datatype->super,1) &&
0 == lb ) {
recvbuf_is_contiguous = true;
}
/* In case the data is not contigous in memory, decode it into an iovec */
if (!recvbuf_is_contiguous ) {
fh->f_decode_datatype ( (struct mca_io_ompio_file_t *)fh,
datatype,
count,
buf,
&max_data,
&decoded_iov,
&iov_count);
}
else {
max_data = count * datatype->super.size;
}
if ( MPI_STATUS_IGNORE != status ) {
status->_ucount = max_data;
}
fh->f_get_num_aggregators ( &static_num_io_procs );
fh->f_set_aggregator_props ((struct mca_io_ompio_file_t *) fh,
static_num_io_procs,
max_data);
my_aggregator = fh->f_procs_in_group[fh->f_aggregator_index];
/* printf("max_data %ld\n", max_data); */
ret = fh->f_generate_current_file_view((struct mca_io_ompio_file_t *)fh,
max_data,
&iov,
&iov_size);
if (ret != OMPI_SUCCESS){
goto exit;
}
if ( iov_size > 0 ) {
local_iov_array = (mca_fcoll_static_local_io_array *)malloc (iov_size * sizeof(mca_fcoll_static_local_io_array));
if ( NULL == local_iov_array){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for (j=0; j < iov_size; j++){
local_iov_array[j].offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t)
iov[j].iov_base;
local_iov_array[j].length = (size_t)iov[j].iov_len;
local_iov_array[j].process_id = fh->f_rank;
}
}
else {
/* Allocate at least one element to correctly create the derived
data type */
local_iov_array = (mca_fcoll_static_local_io_array *)malloc (sizeof(mca_fcoll_static_local_io_array));
if ( NULL == local_iov_array){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
local_iov_array[0].offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t) 0;
local_iov_array[0].length = (size_t) 0;
local_iov_array[0].process_id = fh->f_rank;
}
d[0] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0];
d[1] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0].length;
d[2] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0].process_id;
base = d[0];
for (i=0 ; i<3 ; i++) {
d[i] -= base;
}
/* io_array datatype for using in communication*/
types[0] = &ompi_mpi_long.dt;
types[1] = &ompi_mpi_long.dt;
types[2] = &ompi_mpi_int.dt;
ompi_datatype_create_struct (3,
blocklen,
d,
types,
&io_array_type);
ompi_datatype_commit (&io_array_type);
/* #########################################################*/
fh->f_get_bytes_per_agg ( (int*) &bytes_per_cycle);
local_cycles = ceil((double)max_data*fh->f_procs_per_group/bytes_per_cycle);
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
ret = fh->f_comm->c_coll.coll_allreduce (&local_cycles,
&cycles,
1,
MPI_INT,
MPI_MAX,
fh->f_comm,
fh->f_comm->c_coll.coll_allreduce_module);
if (OMPI_SUCCESS != ret){
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
if (my_aggregator == fh->f_rank) {
disp_index = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == disp_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
bytes_per_process = (int *) malloc (fh->f_procs_per_group * sizeof(int ));
if (NULL == bytes_per_process){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
bytes_remaining = (int *) calloc (fh->f_procs_per_group, sizeof(int));
if (NULL == bytes_remaining){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
current_index = (int *) calloc (fh->f_procs_per_group, sizeof(int));
if (NULL == current_index){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
blocklen_per_process = (int **)calloc (fh->f_procs_per_group, sizeof (int*));
if (NULL == blocklen_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs_per_process = (MPI_Aint **)calloc (fh->f_procs_per_group, sizeof (MPI_Aint*));
if (NULL == displs_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
iovec_count_per_process = (int *) calloc (fh->f_procs_per_group, sizeof(int));
if (NULL == iovec_count_per_process){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs = (int *) calloc (fh->f_procs_per_group, sizeof(int));
if (NULL == displs){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
ret = fcoll_base_coll_allgather_array (&iov_size,
1,
MPI_INT,
iovec_count_per_process,
1,
MPI_INT,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if( OMPI_SUCCESS != ret){
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
if (my_aggregator == fh->f_rank) {
displs[0] = 0;
global_iov_count = iovec_count_per_process[0];
for (i=1 ; i<fh->f_procs_per_group ; i++) {
global_iov_count += iovec_count_per_process[i];
displs[i] = displs[i-1] + iovec_count_per_process[i-1];
}
}
if ( (my_aggregator == fh->f_rank) &&
(global_iov_count > 0 )) {
global_iov_array = (mca_fcoll_static_local_io_array *) malloc (global_iov_count *
sizeof(mca_fcoll_static_local_io_array));
if (NULL == global_iov_array){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
ret = fcoll_base_coll_gatherv_array (local_iov_array,
iov_size,
io_array_type,
global_iov_array,
iovec_count_per_process,
displs,
io_array_type,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if (OMPI_SUCCESS != ret){
fprintf(stderr,"global_iov_array gather error!\n");
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
if (NULL != local_iov_array){
free(local_iov_array);
local_iov_array = NULL;
}
if ( ( my_aggregator == fh->f_rank) &&
( global_iov_count > 0 )) {
sorted = (int *)malloc (global_iov_count * sizeof(int));
if (NULL == sorted) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
read_local_heap_sort (global_iov_array, global_iov_count, sorted);
send_req = (MPI_Request *) malloc (fh->f_procs_per_group * sizeof(MPI_Request));
if (NULL == send_req){
opal_output ( 1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
sendtype = (ompi_datatype_t **) malloc (fh->f_procs_per_group * sizeof(ompi_datatype_t *));
if (NULL == sendtype) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for ( i=0; i<fh->f_procs_per_group; i++ ) {
sendtype[i] = MPI_DATATYPE_NULL;
}
if (NULL == bytes_per_process){
bytes_per_process = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == bytes_per_process){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
}
#if DEBUG_ON
if (my_aggregator == fh->f_rank) {
for (gc_in=0; gc_in<global_iov_count; gc_in++){
printf("%d: Offset[%ld]: %lld, Length[%ld]: %ld\n",
global_iov_array[sorted[gc_in]].process_id,
gc_in, global_iov_array[sorted[gc_in]].offset,
gc_in, global_iov_array[sorted[gc_in]].length);
}
}
#endif
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
for (index = 0; index < cycles; index++){
if (my_aggregator == fh->f_rank) {
fh->f_num_of_io_entries = 0;
if (NULL != fh->f_io_array) {
free (fh->f_io_array);
fh->f_io_array = NULL;
}
if (NULL != global_buf) {
free (global_buf);
global_buf = NULL;
}
if (NULL != sorted_file_offsets){
free(sorted_file_offsets);
sorted_file_offsets = NULL;
}
if (NULL != file_offsets_for_agg){
free(file_offsets_for_agg);
file_offsets_for_agg = NULL;
}
if (NULL != memory_displacements){
free(memory_displacements);
memory_displacements= NULL;
}
if ( NULL != sendtype ) {
for ( i=0; i<fh->f_procs_per_group; i++ ) {
if ( MPI_DATATYPE_NULL != sendtype[i] ) {
ompi_datatype_destroy (&sendtype[i] );
sendtype[i] = MPI_DATATYPE_NULL;
}
}
}
for(l=0;l<fh->f_procs_per_group;l++){
disp_index[l] = 1;
if (NULL != blocklen_per_process[l]){
free(blocklen_per_process[l]);
blocklen_per_process[l] = NULL;
}
if (NULL != displs_per_process[l]){
free(displs_per_process[l]);
displs_per_process[l] = NULL;
}
blocklen_per_process[l] = (int *) calloc (1, sizeof(int));
if (NULL == blocklen_per_process[l]) {
opal_output (1, "OUT OF MEMORY for blocklen\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs_per_process[l] = (MPI_Aint *) calloc (1, sizeof(MPI_Aint));
if (NULL == displs_per_process[l]){
opal_output (1, "OUT OF MEMORY for displs\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
}
if (index < local_cycles ) {
if ((index == local_cycles-1) && (max_data % (bytes_per_cycle/fh->f_procs_per_group))) {
bytes_to_read_in_cycle = max_data - position;
}
else if (max_data <= bytes_per_cycle/fh->f_procs_per_group) {
bytes_to_read_in_cycle = max_data;
}
else {
bytes_to_read_in_cycle = bytes_per_cycle/fh->f_procs_per_group;
}
}
else {
bytes_to_read_in_cycle = 0;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
fcoll_base_coll_gather_array (&bytes_to_read_in_cycle,
1,
MPI_INT,
bytes_per_process,
1,
MPI_INT,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
if (recvbuf_is_contiguous ) {
receive_buf = &((char*)buf)[position];
}
else if (bytes_to_read_in_cycle) {
receive_buf = (char *) malloc (bytes_to_read_in_cycle * sizeof(char));
if ( NULL == receive_buf){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rcomm_time = MPI_Wtime();
#endif
ret = MCA_PML_CALL(irecv(receive_buf,
bytes_to_read_in_cycle,
MPI_BYTE,
my_aggregator,
123,
fh->f_comm,
&recv_req));
if (OMPI_SUCCESS != ret){
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
if (my_aggregator == fh->f_rank) {
for (i=0;i<fh->f_procs_per_group; i++){
while (bytes_per_process[i] > 0){
/*printf("%d: bytes_per_process[%d]: %d, bytes_remaining[%d]: %d\n",
index, i, bytes_per_process[i], i, bytes_remaining[i]);*/
if (read_get_process_id(global_iov_array[sorted[current_index[i]]].process_id,
fh) == i){ /* current id owns this entry!*/
if (bytes_remaining[i]){ /*Remaining bytes in the current entry of
the global offset array*/
if (bytes_remaining[i] <= bytes_per_process[i]){
blocklen_per_process[i][disp_index[i] - 1] = bytes_remaining[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset +
(global_iov_array[sorted[current_index[i]]].length
- bytes_remaining[i]);
blocklen_per_process[i] = (int *) realloc
((void *)blocklen_per_process[i], (disp_index[i]+1)*sizeof(int));
displs_per_process[i] = (MPI_Aint *)realloc
((void *)displs_per_process[i], (disp_index[i]+1)*sizeof(MPI_Aint));
bytes_per_process[i] -= bytes_remaining[i];
blocklen_per_process[i][disp_index[i]] = 0;
displs_per_process[i][disp_index[i]] = 0;
disp_index[i] += 1;
bytes_remaining[i] = 0;
/* This entry has been used up, we need to move to the
next entry of this process and make current_index point there*/
current_index[i] = read_find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1){
break;
}
continue;
}
else{
blocklen_per_process[i][disp_index[i] - 1] = bytes_per_process[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset +
(global_iov_array[sorted[current_index[i]]].length
- bytes_remaining[i]);
bytes_remaining[i] -= bytes_per_process[i];
bytes_per_process[i] = 0;
break;
}
}
else{
if (bytes_per_process[i] <
global_iov_array[sorted[current_index[i]]].length){
blocklen_per_process[i][disp_index[i] - 1] =
bytes_per_process[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset;
bytes_remaining[i] =
global_iov_array[sorted[current_index[i]]].length -
bytes_per_process[i];
bytes_per_process[i] = 0;
break;
}
else {
blocklen_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].length;
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset;
blocklen_per_process[i] =
(int *) realloc ((void *)blocklen_per_process[i], (disp_index[i]+1)*sizeof(int));
displs_per_process[i] = (MPI_Aint *)realloc
((void *)displs_per_process[i], (disp_index[i]+1)*sizeof(MPI_Aint));
blocklen_per_process[i][disp_index[i]] = 0;
displs_per_process[i][disp_index[i]] = 0;
disp_index[i] += 1;
bytes_per_process[i] -=
global_iov_array[sorted[current_index[i]]].length;
current_index[i] = read_find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1){
break;
}
}
}
}
else{
current_index[i] = read_find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1){
bytes_per_process[i] = 0; /* no more entries left
to service this request*/
continue;
}
}
}
}
entries_per_aggregator=0;
for (i=0;i<fh->f_procs_per_group;i++){
for (j=0;j<disp_index[i];j++){
if (blocklen_per_process[i][j] > 0){
entries_per_aggregator++;
#if DEBUG_ON
printf("%d sends blocklen[%d]: %d, disp[%d]: %ld to %d\n",
fh->f_procs_in_group[i],j,
blocklen_per_process[i][j],j,
displs_per_process[i][j],
fh->f_rank);
#endif
}
}
}
if (entries_per_aggregator > 0){
file_offsets_for_agg = (mca_fcoll_static_local_io_array *)
malloc(entries_per_aggregator*sizeof(mca_fcoll_static_local_io_array));
if (NULL == file_offsets_for_agg) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
sorted_file_offsets = (int *) malloc (entries_per_aggregator * sizeof(int));
if (NULL == sorted_file_offsets){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
temp_index=0;
global_count = 0;
for (i=0;i<fh->f_procs_per_group; i++){
for(j=0;j<disp_index[i]; j++){
if (blocklen_per_process[i][j] > 0){
file_offsets_for_agg[temp_index].length =
blocklen_per_process[i][j];
global_count += blocklen_per_process[i][j];
file_offsets_for_agg[temp_index].process_id = i;
file_offsets_for_agg[temp_index].offset =
displs_per_process[i][j];
temp_index++;
}
}
}
}
else{
continue;
}
read_local_heap_sort (file_offsets_for_agg,
entries_per_aggregator,
sorted_file_offsets);
memory_displacements = (MPI_Aint *) malloc
(entries_per_aggregator * sizeof(MPI_Aint));
memory_displacements[sorted_file_offsets[0]] = 0;
for (i=1; i<entries_per_aggregator; i++){
memory_displacements[sorted_file_offsets[i]] =
memory_displacements[sorted_file_offsets[i-1]] +
file_offsets_for_agg[sorted_file_offsets[i-1]].length;
}
global_buf = (char *) malloc (global_count * sizeof(char));
if (NULL == global_buf){
opal_output(1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
#if DEBUG_ON
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
for (i=0; i<entries_per_aggregator;i++){
printf("%d: OFFSET: %lld LENGTH: %ld, Mem-offset: %ld, disp_index :%d\n",
file_offsets_for_agg[sorted_file_offsets[i]].process_id,
file_offsets_for_agg[sorted_file_offsets[i]].offset,
file_offsets_for_agg[sorted_file_offsets[i]].length,
memory_displacements[sorted_file_offsets[i]],
disp_index[i]);
}
#endif
fh->f_io_array = (mca_io_ompio_io_array_t *) malloc
(entries_per_aggregator * sizeof (mca_io_ompio_io_array_t));
if (NULL == fh->f_io_array) {
opal_output(1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
fh->f_num_of_io_entries = 0;
fh->f_io_array[0].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[0]].offset;
fh->f_io_array[0].length = file_offsets_for_agg[sorted_file_offsets[0]].length;
fh->f_io_array[0].memory_address = global_buf+memory_displacements[sorted_file_offsets[0]];
fh->f_num_of_io_entries++;
for (i=1;i<entries_per_aggregator;i++){
if (file_offsets_for_agg[sorted_file_offsets[i-1]].offset +
file_offsets_for_agg[sorted_file_offsets[i-1]].length ==
file_offsets_for_agg[sorted_file_offsets[i]].offset){
fh->f_io_array[fh->f_num_of_io_entries - 1].length +=
file_offsets_for_agg[sorted_file_offsets[i]].length;
}
else{
fh->f_io_array[fh->f_num_of_io_entries].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[i]].offset;
fh->f_io_array[fh->f_num_of_io_entries].length =
file_offsets_for_agg[sorted_file_offsets[i]].length;
fh->f_io_array[fh->f_num_of_io_entries].memory_address =
global_buf+memory_displacements[sorted_file_offsets[i]];
fh->f_num_of_io_entries++;
}
}
#if DEBUG_ON
printf("*************************** %d\n", fh->f_num_of_io_entries);
for (i=0 ; i<fh->f_num_of_io_entries ; i++) {
printf(" ADDRESS: %p OFFSET: %ld LENGTH: %ld\n",
fh->f_io_array[i].memory_address,
(OPAL_PTRDIFF_TYPE)fh->f_io_array[i].offset,
fh->f_io_array[i].length);
}
#endif
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_read_time = MPI_Wtime();
#endif
if (fh->f_num_of_io_entries) {
if ( 0 > fh->f_fbtl->fbtl_preadv (fh)) {
opal_output (1, "READ FAILED\n");
ret = OMPI_ERROR;
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_read_time = MPI_Wtime();
read_time += end_read_time - start_read_time;
#endif
#if DEBUG_ON
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
if (my_aggregator == fh->f_rank){
for (i=0 ; i<global_count/4 ; i++)
printf (" READ %d \n",((int *)global_buf)[i]);
}
#endif
temp_disp_index = (int *)calloc (1, fh->f_procs_per_group * sizeof (int));
if (NULL == temp_disp_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for (i=0; i<entries_per_aggregator; i++){
temp_index =
file_offsets_for_agg[sorted_file_offsets[i]].process_id;
displs_per_process[temp_index][temp_disp_index[temp_index]] =
memory_displacements[sorted_file_offsets[i]];
if (temp_disp_index[temp_index] < disp_index[temp_index]){
temp_disp_index[temp_index] += 1;
}
else{
printf("temp_disp_index[%d]: %d is greater than disp_index[%d]: %d\n",
temp_index, temp_disp_index[temp_index],
temp_index, disp_index[temp_index]);
}
}
if (NULL != temp_disp_index){
free(temp_disp_index);
temp_disp_index = NULL;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rcomm_time = MPI_Wtime();
#endif
for (i=0;i<fh->f_procs_per_group; i++){
send_req[i] = MPI_REQUEST_NULL;
ompi_datatype_create_hindexed(disp_index[i],
blocklen_per_process[i],
displs_per_process[i],
MPI_BYTE,
&sendtype[i]);
ompi_datatype_commit(&sendtype[i]);
ret = MCA_PML_CALL (isend(global_buf,
1,
sendtype[i],
fh->f_procs_in_group[i],
123,
MCA_PML_BASE_SEND_STANDARD,
fh->f_comm,
&send_req[i]));
if(OMPI_SUCCESS != ret){
goto exit;
}
}
ret = ompi_request_wait_all (fh->f_procs_per_group,
send_req,
MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret){
goto exit;
}
} /* if ( my_aggregator == fh->f_rank ) */
ret = ompi_request_wait (&recv_req, MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret){
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
position += bytes_to_read_in_cycle;
if (!recvbuf_is_contiguous) {
OPAL_PTRDIFF_TYPE mem_address;
size_t remaining = 0;
size_t temp_position = 0;
remaining = bytes_to_read_in_cycle;
while (remaining && (iov_count > iov_index)){
mem_address = (OPAL_PTRDIFF_TYPE)
(decoded_iov[iov_index].iov_base) + current_position;
if (remaining >=
(decoded_iov[iov_index].iov_len - current_position)) {
memcpy ((IOVBASE_TYPE *) mem_address,
receive_buf+temp_position,
decoded_iov[iov_index].iov_len - current_position);
remaining = remaining -
(decoded_iov[iov_index].iov_len - current_position);
temp_position = temp_position +
(decoded_iov[iov_index].iov_len - current_position);
iov_index = iov_index + 1;
current_position = 0;
}
else{
memcpy ((IOVBASE_TYPE *) mem_address,
receive_buf+temp_position,
remaining);
current_position = current_position + remaining;
remaining = 0;
}
}
if (NULL != receive_buf) {
free (receive_buf);
receive_buf = NULL;
}
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rexch = MPI_Wtime();
read_exch += end_rexch - start_rexch;
nentry.time[0] = read_time;
nentry.time[1] = rcomm_time;
nentry.time[2] = read_exch;
if (my_aggregator == fh->f_rank)
nentry.aggregator = 1;
else
nentry.aggregator = 0;
nentry.nprocs_for_coll = static_num_io_procs;
if (!mca_common_ompio_full_print_queue(fh->f_coll_read_time)){
mca_common_ompio_register_print_entry(fh->f_coll_read_time,
nentry);
}
#endif
exit:
if (NULL != decoded_iov){
free(decoded_iov);
decoded_iov = NULL;
}
if (NULL != displs){
free(displs);
displs = NULL;
}
if (NULL != iovec_count_per_process){
free(iovec_count_per_process);
iovec_count_per_process=NULL;
}
if (NULL != local_iov_array){
free(local_iov_array);
local_iov_array=NULL;
}
if (NULL != global_iov_array){
free(global_iov_array);
global_iov_array=NULL;
}
if (my_aggregator == fh->f_rank) {
for(l=0;l<fh->f_procs_per_group;l++){
if (blocklen_per_process) {
free(blocklen_per_process[l]);
}
if (NULL != displs_per_process[l]){
free(displs_per_process[l]);
displs_per_process[l] = NULL;
}
}
}
if (NULL != bytes_per_process){
free(bytes_per_process);
bytes_per_process =NULL;
}
if (NULL != disp_index){
free(disp_index);
disp_index =NULL;
}
if (NULL != displs_per_process){
free(displs_per_process);
displs_per_process = NULL;
}
if(NULL != bytes_remaining){
free(bytes_remaining);
bytes_remaining = NULL;
}
if(NULL != current_index){
free(current_index);
current_index = NULL;
}
if (NULL != blocklen_per_process){
free(blocklen_per_process);
blocklen_per_process =NULL;
}
if (NULL != bytes_remaining){
free(bytes_remaining);
bytes_remaining =NULL;
}
if (NULL != memory_displacements){
free(memory_displacements);
memory_displacements= NULL;
}
if (NULL != file_offsets_for_agg){
free(file_offsets_for_agg);
file_offsets_for_agg = NULL;
}
if (NULL != sorted_file_offsets){
free(sorted_file_offsets);
sorted_file_offsets = NULL;
}
if (NULL != sendtype){
free(sendtype);
sendtype=NULL;
}
if ( !recvbuf_is_contiguous ) {
if (NULL != receive_buf){
free(receive_buf);
receive_buf=NULL;
}
}
if (NULL != global_buf) {
free(global_buf);
global_buf = NULL;
}
if (NULL != sorted) {
free(sorted);
sorted = NULL;
}
if (NULL != send_req){
free(send_req);
send_req = NULL;
}
return ret;
}
