static inline int create_iov_list(const void *addr, int count, ompi_datatype_t *datatype,
ucx_iovec_t **ucx_iov, uint32_t *ucx_iov_count) {
int ret = OMPI_SUCCESS;
size_t size;
bool done = false;
opal_convertor_t convertor;
uint32_t iov_count, iov_idx;
struct iovec iov[OSC_UCX_IOVEC_MAX];
uint32_t ucx_iov_idx;
OBJ_CONSTRUCT(&convertor, opal_convertor_t);
ret = opal_convertor_copy_and_prepare_for_send(ompi_mpi_local_convertor,
&datatype->super, count,
addr, 0, &convertor);
if (ret != OMPI_SUCCESS) {
return ret;
}
(*ucx_iov_count) = 0;
ucx_iov_idx = 0;
do {
iov_count = OSC_UCX_IOVEC_MAX;
iov_idx = 0;
done = opal_convertor_raw(&convertor, iov, &iov_count, &size);
(*ucx_iov_count) += iov_count;
(*ucx_iov) = (ucx_iovec_t *)realloc((*ucx_iov), (*ucx_iov_count) * sizeof(ucx_iovec_t));
if (*ucx_iov == NULL) {
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
while (iov_idx != iov_count) {
(*ucx_iov)[ucx_iov_idx].addr = iov[iov_idx].iov_base;
(*ucx_iov)[ucx_iov_idx].len = iov[iov_idx].iov_len;
ucx_iov_idx++;
iov_idx++;
}
assert((*ucx_iov_count) == ucx_iov_idx);
} while (!done);
opal_convertor_cleanup(&convertor);
OBJ_DESTRUCT(&convertor);
return ret;
}
static int test_upper( unsigned int length )
{
ompi_datatype_t *pdt;
opal_convertor_t * pConv;
int rc = OMPI_SUCCESS;
unsigned int i, iov_count, split_chunk, total_length;
size_t max_data;
struct iovec iov[5];
TIMER_DATA_TYPE start, end;
long total_time;
printf( "test upper matrix\n" );
pdt = upper_matrix( length );
/*dt_dump( pdt );*/
total_length = length * (length + 1) * ( sizeof(double) / 2);
pConv = opal_convertor_create( remote_arch, 0 );
if( OMPI_SUCCESS != opal_convertor_prepare_for_send( pConv, &(pdt->super), 1, NULL ) ) {
printf( "Cannot attach the datatype to a convertor\n" );
return OMPI_ERROR;
}
GET_TIME( start );
split_chunk = (length + 1) * sizeof(double);
/* split_chunk = (total_length + 1) * sizeof(double); */
for( i = total_length; i > 0; ) {
iov_count = 5;
max_data = 0;
opal_convertor_raw( pConv, iov, &iov_count, &max_data );
i -= max_data;
}
GET_TIME( end );
total_time = ELAPSED_TIME( start, end );
printf( "complete raw in %ld microsec\n", total_time );
/* test the automatic destruction pf the data */
ompi_datatype_destroy( &pdt );
assert( pdt == NULL );
OBJ_RELEASE( pConv );
return rc;
}
/**
* Conversion function. They deal with data-types in 3 ways, always making local copies.
* In order to allow performance testings, there are 3 functions:
* - one copying directly from one memory location to another one using the
* data-type copy function.
* - one which use a 2 convertors created with the same data-type
* - and one using 2 convertors created from different data-types.
*
*/
static int local_copy_ddt_raw( ompi_datatype_t* pdt, int count, int iov_num )
{
struct iovec* iov;
opal_convertor_t* convertor;
TIMER_DATA_TYPE start, end;
long total_time;
uint32_t iov_count = iov_num;
size_t max_data = 0, remaining_length;
iov = (struct iovec*)malloc(iov_num * sizeof(struct iovec));
convertor = opal_convertor_create( remote_arch, 0 );
if( OMPI_SUCCESS != opal_convertor_prepare_for_send( convertor, &(pdt->super), count, NULL ) ) {
printf( "Cannot attach the datatype to a convertor\n" );
return OMPI_ERROR;
}
remaining_length = count * pdt->super.size;
GET_TIME( start );
while( 0 == opal_convertor_raw(convertor, iov, &iov_count, &max_data) ) {
#if 0
printf( "New raw extraction (iov_count = %d, max_data = %zu)\n",
iov_count, max_data );
for( i = 0; i < iov_count; i++ ) {
printf( "\t{%p, %d}\n", iov[i].iov_base, iov[i].iov_len );
}
#endif
remaining_length -= max_data;
iov_count = iov_num;
}
remaining_length -= max_data;
GET_TIME( end );
total_time = ELAPSED_TIME( start, end );
printf( "raw extraction in %ld microsec\n", total_time );
OBJ_RELEASE( convertor );
if( remaining_length != 0 ) {
printf( "Not all raw description was been extracted (%lu bytes missing)\n",
(unsigned long) remaining_length );
}
free(iov);
return OMPI_SUCCESS;
}
static inline int ompi_osc_rdma_gacc_master (ompi_osc_rdma_sync_t *sync, const void *source_buffer, int source_count,
ompi_datatype_t *source_datatype, void *result_buffer, int result_count,
ompi_datatype_t *result_datatype, ompi_osc_rdma_peer_t *peer, uint64_t target_address,
mca_btl_base_registration_handle_t *target_handle, int target_count,
ompi_datatype_t *target_datatype, ompi_op_t *op, ompi_osc_rdma_request_t *request)
{
ompi_osc_rdma_module_t *module = sync->module;
struct iovec source_iovec[OMPI_OSC_RDMA_DECODE_MAX], target_iovec[OMPI_OSC_RDMA_DECODE_MAX];
const size_t acc_limit = (mca_osc_rdma_component.buffer_size >> 3);
uint32_t source_primitive_count, target_primitive_count;
opal_convertor_t source_convertor, target_convertor;
uint32_t source_iov_count, target_iov_count;
uint32_t source_iov_index, target_iov_index;
ompi_datatype_t *source_primitive, *target_primitive;
/* needed for opal_convertor_raw but not used */
size_t source_size, target_size;
ompi_osc_rdma_request_t *subreq;
size_t result_position;
ptrdiff_t lb, extent;
int ret, acc_len;
bool done;
(void) ompi_datatype_get_extent (target_datatype, &lb, &extent);
target_address += lb;
/* fast path for accumulate on built-in types */
if (OPAL_LIKELY((!source_count || ompi_datatype_is_predefined (source_datatype)) &&
ompi_datatype_is_predefined (target_datatype) &&
(!result_count || ompi_datatype_is_predefined (result_datatype)) &&
(target_datatype->super.size * target_count <= acc_limit))) {
if (NULL == request) {
OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, request);
request->internal = true;
request->type = result_datatype ? OMPI_OSC_RDMA_TYPE_GET_ACC : OMPI_OSC_RDMA_TYPE_ACC;
}
if (source_datatype) {
(void) ompi_datatype_get_extent (source_datatype, &lb, &extent);
source_buffer = (void *)((intptr_t) source_buffer + lb);
}
if (result_datatype) {
(void) ompi_datatype_get_extent (result_datatype, &lb, &extent);
result_buffer = (void *)((intptr_t) result_buffer + lb);
}
ret = ompi_osc_rdma_gacc_contig (sync, source_buffer, source_count, source_datatype, result_buffer,
result_count, result_datatype, peer, target_address,
target_handle, target_count, target_datatype, op,
request);
if (OPAL_LIKELY(OMPI_SUCCESS == ret)) {
return OMPI_SUCCESS;
}
if (source_datatype) {
/* the convertors will handle the lb */
(void) ompi_datatype_get_extent (source_datatype, &lb, &extent);
source_buffer = (void *)((intptr_t) source_buffer - lb);
}
if (result_datatype) {
(void) ompi_datatype_get_extent (result_datatype, &lb, &extent);
result_buffer = (void *)((intptr_t) result_buffer - lb);
}
}
/* the convertor will handle lb from here */
(void) ompi_datatype_get_extent (target_datatype, &lb, &extent);
target_address -= lb;
/* get the primitive datatype info */
ret = ompi_osc_base_get_primitive_type_info (target_datatype, &target_primitive, &target_primitive_count);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
/* target datatype is not made up of a single basic datatype */
return ret;
}
if (source_datatype) {
ret = ompi_osc_base_get_primitive_type_info (source_datatype, &source_primitive, &source_primitive_count);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
/* target datatype is not made up of a single basic datatype */
return ret;
}
if (OPAL_UNLIKELY(source_primitive != target_primitive)) {
return MPI_ERR_TYPE;
}
}
/* prepare convertors for the source and target. these convertors will be used to determine the
* contiguous segments within the source and target. */
/* the source may be NULL if using MPI_OP_NO_OP with MPI_Get_accumulate */
if (source_datatype) {
OBJ_CONSTRUCT(&source_convertor, opal_convertor_t);
ret = opal_convertor_copy_and_prepare_for_send (ompi_mpi_local_convertor, &source_datatype->super, source_count, source_buffer,
0, &source_convertor);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
}
/* target_datatype can never be NULL */
OBJ_CONSTRUCT(&target_convertor, opal_convertor_t);
ret = opal_convertor_copy_and_prepare_for_send (ompi_mpi_local_convertor, &target_datatype->super, target_count,
(void *) (intptr_t) target_address, 0, &target_convertor);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
if (request) {
/* keep the request from completing until all the transfers have started */
request->outstanding_requests = 1;
}
target_iov_index = 0;
target_iov_count = 0;
result_position = 0;
do {
/* decode segments of the source data */
source_iov_count = OMPI_OSC_RDMA_DECODE_MAX;
source_iov_index = 0;
/* opal_convertor_raw returns done when it has reached the end of the data */
if (!source_datatype) {
done = true;
source_iovec[0].iov_len = (size_t) -1;
source_iovec[0].iov_base = NULL;
source_iov_count = 1;
} else {
done = opal_convertor_raw (&source_convertor, source_iovec, &source_iov_count, &source_size);
}
/* loop on the target segments until we have exhaused the decoded source data */
while (source_iov_index != source_iov_count) {
if (target_iov_index == target_iov_count) {
/* decode segments of the target buffer */
target_iov_count = OMPI_OSC_RDMA_DECODE_MAX;
target_iov_index = 0;
(void) opal_convertor_raw (&target_convertor, target_iovec, &target_iov_count, &target_size);
}
/* we already checked that the target was large enough. this should be impossible */
assert (0 != target_iov_count);
/* determine how much to put in this operation */
acc_len = min(target_iovec[target_iov_index].iov_len, source_iovec[source_iov_index].iov_len);
acc_len = min((size_t) acc_len, acc_limit);
/* execute the get */
OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, subreq);
subreq->internal = true;
subreq->parent_request = request;
if (request) {
(void) OPAL_THREAD_ADD32 (&request->outstanding_requests, 1);
}
if (result_datatype) {
/* prepare a convertor for this part of the result */
opal_convertor_copy_and_prepare_for_recv (ompi_mpi_local_convertor, &result_datatype->super, result_count,
result_buffer, 0, &subreq->convertor);
opal_convertor_set_position (&subreq->convertor, &result_position);
subreq->type = OMPI_OSC_RDMA_TYPE_GET_ACC;
} else {
subreq->type = OMPI_OSC_RDMA_TYPE_ACC;
}
OPAL_OUTPUT_VERBOSE((60, ompi_osc_base_framework.framework_output,
"target index = %d, target = {%p, %lu}, source_index = %d, source = {%p, %lu}, result = %p, result position = %lu, "
"acc_len = %d, count = %lu",
target_iov_index, target_iovec[target_iov_index].iov_base, (unsigned long) target_iovec[target_iov_index].iov_len,
source_iov_index, source_iovec[source_iov_index].iov_base, (unsigned long) source_iovec[source_iov_index].iov_len,
result_buffer, (unsigned long) result_position, acc_len, (unsigned long)(acc_len / target_primitive->super.size)));
ret = ompi_osc_rdma_gacc_contig (sync, source_iovec[source_iov_index].iov_base, acc_len / target_primitive->super.size,
target_primitive, NULL, 0, NULL, peer, (uint64_t) (intptr_t) target_iovec[target_iov_index].iov_base,
target_handle, acc_len / target_primitive->super.size, target_primitive, op, subreq);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
if (OPAL_UNLIKELY(OMPI_ERR_OUT_OF_RESOURCE != ret)) {
/* something bad happened. need to figure out how to handle these errors */
return ret;
}
/* progress and try again */
ompi_osc_rdma_progress (module);
continue;
}
/* adjust io vectors */
target_iovec[target_iov_index].iov_len -= acc_len;
source_iovec[source_iov_index].iov_len -= acc_len;
target_iovec[target_iov_index].iov_base = (void *)((intptr_t) target_iovec[target_iov_index].iov_base + acc_len);
source_iovec[source_iov_index].iov_base = (void *)((intptr_t) source_iovec[source_iov_index].iov_base + acc_len);
result_position += acc_len;
source_iov_index += !source_datatype || (0 == source_iovec[source_iov_index].iov_len);
target_iov_index += (0 == target_iovec[target_iov_index].iov_len);
}
} while (!done);
if (request) {
/* release our reference so the request can complete */
(void) OPAL_THREAD_ADD32 (&request->outstanding_requests, -1);
}
if (source_datatype) {
opal_convertor_cleanup (&source_convertor);
OBJ_DESTRUCT(&source_convertor);
}
opal_convertor_cleanup (&target_convertor);
OBJ_DESTRUCT(&target_convertor);
return OMPI_SUCCESS;
}