static inline int __ompi_datatype_pack_description( ompi_datatype_t* datatype,
void** packed_buffer, int* next_index )
{
int i, *position = (int*)*packed_buffer;
ompi_datatype_args_t* args = (ompi_datatype_args_t*)datatype->args;
char* next_packed = (char*)*packed_buffer;
if( ompi_datatype_is_predefined(datatype) ) {
position[0] = MPI_COMBINER_NAMED;
position[1] = datatype->id; /* On the OMPI - layer, copy the ompi_datatype.id */
next_packed += (2 * sizeof(int));
*packed_buffer = next_packed;
return OMPI_SUCCESS;
}
/* For duplicated datatype we don't have to store all the information */
if( MPI_COMBINER_DUP == args->create_type ) {
ompi_datatype_t* temp_data = args->d[0];
return __ompi_datatype_pack_description(temp_data,
packed_buffer,
next_index );
}
position[0] = args->create_type;
position[1] = args->ci;
position[2] = args->ca;
position[3] = args->cd;
next_packed += (4 * sizeof(int));
/* Spoiler: We will access the data in this storage structure, and thus we
* need to align it to the expected boundaries (special thanks to Sparc64).
* The simplest way is to ensure that prior to each type that must be 64
* bits aligned, we have a pointer that is 64 bits aligned. That will minimize
* the memory requirements in all cases where no displacements are stored.
*/
if( 0 < args->ca ) {
/* description of the displacements must be 64 bits aligned */
OMPI_DATATYPE_ALIGN_PTR(next_packed, char*);
memcpy( next_packed, args->a, sizeof(OPAL_PTRDIFF_TYPE) * args->ca );
next_packed += sizeof(OPAL_PTRDIFF_TYPE) * args->ca;
}
position = (int*)next_packed;
next_packed += sizeof(int) * args->cd;
/* copy the aray of counts (32 bits aligned) */
memcpy( next_packed, args->i, sizeof(int) * args->ci );
next_packed += args->ci * sizeof(int);
/* copy the rest of the data */
for( i = 0; i < args->cd; i++ ) {
ompi_datatype_t* temp_data = args->d[i];
if( ompi_datatype_is_predefined(temp_data) ) {
position[i] = temp_data->id; /* On the OMPI - layer, copy the ompi_datatype.id */
} else {
position[i] = *next_index;
(*next_index)++;
__ompi_datatype_pack_description( temp_data,
(void**)&next_packed,
next_index );
}
}
*packed_buffer = next_packed;
return OMPI_SUCCESS;
}
int ompi_datatype_get_pack_description( ompi_datatype_t* datatype,
const void** packed_buffer )
{
ompi_datatype_args_t* args = (ompi_datatype_args_t*)datatype->args;
int next_index = OMPI_DATATYPE_MAX_PREDEFINED;
void *packed_description = datatype->packed_description;
void* recursive_buffer;
if (NULL == packed_description) {
if (opal_atomic_cmpset (&datatype->packed_description, NULL, (void *) 1)) {
if( ompi_datatype_is_predefined(datatype) ) {
packed_description = malloc(2 * sizeof(int));
} else if( NULL == args ) {
return OMPI_ERROR;
} else {
packed_description = malloc(args->total_pack_size);
}
recursive_buffer = packed_description;
__ompi_datatype_pack_description( datatype, &recursive_buffer, &next_index );
if (!ompi_datatype_is_predefined(datatype)) {
args->total_pack_size = (uintptr_t)((char*)recursive_buffer - (char *) packed_description);
}
opal_atomic_wmb ();
datatype->packed_description = packed_description;
} else {
/* another thread beat us to it */
packed_description = datatype->packed_description;
}
}
if ((void *) 1 == packed_description) {
struct timespec interval = {.tv_sec = 0, .tv_nsec = 1000};
/* wait until the packed description is updated */
while ((void *) 1 == datatype->packed_description) {
nanosleep (&interval, NULL);
}
packed_description = datatype->packed_description;
}
*packed_buffer = (const void *) packed_description;
return OMPI_SUCCESS;
}
size_t ompi_datatype_pack_description_length( ompi_datatype_t* datatype )
{
void *packed_description = datatype->packed_description;
if( ompi_datatype_is_predefined(datatype) ) {
return 2 * sizeof(int);
}
if( NULL == packed_description || (void *) 1 == packed_description) {
const void* buf;
int rc;
rc = ompi_datatype_get_pack_description(datatype, &buf);
if( OMPI_SUCCESS != rc ) {
return 0;
}
}
assert( NULL != (ompi_datatype_args_t*)datatype->args );
assert( NULL != (ompi_datatype_args_t*)datatype->packed_description );
return ((ompi_datatype_args_t*)datatype->args)->total_pack_size;
}
int ompi_datatype_get_pack_description( ompi_datatype_t* datatype,
const void** packed_buffer )
{
ompi_datatype_args_t* args = (ompi_datatype_args_t*)datatype->args;
int next_index = OMPI_DATATYPE_MAX_PREDEFINED;
void *packed_description = (void *) datatype->packed_description;
void* recursive_buffer;
if (NULL == packed_description) {
void *_tmp_ptr = NULL;
if (opal_atomic_compare_exchange_strong_ptr (&datatype->packed_description, (intptr_t *) &_tmp_ptr, 1)) {
if( ompi_datatype_is_predefined(datatype) ) {
packed_description = malloc(2 * sizeof(int));
} else if( NULL == args ) {
return OMPI_ERROR;
} else {
packed_description = malloc(args->total_pack_size);
}
recursive_buffer = packed_description;
__ompi_datatype_pack_description( datatype, &recursive_buffer, &next_index );
if (!ompi_datatype_is_predefined(datatype)) {
/* If the precomputed size is not large enough we're already in troubles, we
* have overwritten outside of the allocated buffer. Raise the alarm !
* If not reassess the size of the packed buffer necessary for holding the
* datatype description.
*/
assert(args->total_pack_size >= (uintptr_t)((char*)recursive_buffer - (char *) packed_description));
args->total_pack_size = (uintptr_t)((char*)recursive_buffer - (char *) packed_description);
}
opal_atomic_wmb ();
datatype->packed_description = (intptr_t) packed_description;
} else {
/* another thread beat us to it */
packed_description = (void *) datatype->packed_description;
}
}
if ((void *) 1 == packed_description) {
struct timespec interval = {.tv_sec = 0, .tv_nsec = 1000};
/* wait until the packed description is updated */
while (1 == datatype->packed_description) {
nanosleep (&interval, NULL);
}
packed_description = (void *) datatype->packed_description;
}
*packed_buffer = (const void *) packed_description;
return OMPI_SUCCESS;
}
size_t ompi_datatype_pack_description_length( ompi_datatype_t* datatype )
{
void *packed_description = (void *) datatype->packed_description;
if( ompi_datatype_is_predefined(datatype) ) {
return 2 * sizeof(int);
}
if( NULL == packed_description || (void *) 1 == packed_description) {
const void* buf;
int rc;
rc = ompi_datatype_get_pack_description(datatype, &buf);
if( OMPI_SUCCESS != rc ) {
return 0;
}
}
assert( NULL != (ompi_datatype_args_t*)datatype->args );
assert( NULL != (ompi_datatype_args_t*)datatype->packed_description );
return ((ompi_datatype_args_t*)datatype->args)->total_pack_size;
}
