int ompi_osc_rdma_raccumulate (const void *origin_addr, int origin_count,
struct ompi_datatype_t *origin_datatype, int target_rank,
OPAL_PTRDIFF_TYPE target_disp, int target_count,
struct ompi_datatype_t *target_datatype, struct ompi_op_t *op,
struct ompi_win_t *win, struct ompi_request_t **request)
{
ompi_osc_rdma_module_t *module = GET_MODULE(win);
ompi_osc_rdma_peer_t *peer;
ompi_osc_rdma_request_t *rdma_request;
ompi_osc_rdma_sync_t *sync;
int ret;
sync = ompi_osc_rdma_module_sync_lookup (module, target_rank, &peer);
if (OPAL_UNLIKELY(NULL == sync)) {
return OMPI_ERR_RMA_SYNC;
}
OPAL_OUTPUT_VERBOSE((50, ompi_osc_base_framework.framework_output, "racc: 0x%lx, %d, %s, %d, 0x%lx, %d, %s, %s, %s",
(unsigned long) origin_addr, origin_count, origin_datatype->name, target_rank,
(unsigned long) target_disp, target_count, target_datatype->name, op->o_name, win->w_name));
OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, rdma_request);
ret = ompi_osc_rdma_rget_accumulate_internal (sync, origin_addr, origin_count, origin_datatype, NULL, 0,
NULL, peer, target_rank, target_disp, target_count, target_datatype,
op, rdma_request);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OMPI_OSC_RDMA_REQUEST_RETURN(rdma_request);
return ret;
}
*request = &rdma_request->super;
return OMPI_SUCCESS;
}
int ompi_osc_rdma_raccumulate(void *origin_addr, int origin_count,
struct ompi_datatype_t *origin_dt, int target,
OPAL_PTRDIFF_TYPE target_disp, int target_count,
struct ompi_datatype_t *target_dt, struct ompi_op_t *op,
struct ompi_win_t *win, struct ompi_request_t **request)
{
ompi_osc_rdma_request_t *rdma_request;
int ret;
OPAL_OUTPUT_VERBOSE((50, ompi_osc_base_framework.framework_output,
"raccumulate: 0x%lx, %d, %s, %d, %d, %d, %s, %s, %s",
(unsigned long) origin_addr, origin_count,
origin_dt->name, target, (int) target_disp,
target_count, target_dt->name, op->o_name,
win->w_name));
OMPI_OSC_RDMA_REQUEST_ALLOC(win, rdma_request);
if (NULL == rdma_request) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* short-circuit case */
if (0 == origin_count || 0 == target_count) {
ompi_osc_rdma_request_complete (rdma_request, MPI_SUCCESS);
*request = (ompi_request_t *) rdma_request;
return OMPI_SUCCESS;
}
rdma_request->type = OMPI_OSC_RDMA_HDR_TYPE_ACC;
ret = ompi_osc_rdma_accumulate_w_req (origin_addr, origin_count, origin_dt, target,
target_disp, target_count, target_dt, op, win,
rdma_request);
if (OMPI_SUCCESS != ret) {
OMPI_OSC_RDMA_REQUEST_RETURN(rdma_request);
return ret;
}
*request = (ompi_request_t *) rdma_request;
return OMPI_SUCCESS;
}
static inline int cas_rdma (ompi_osc_rdma_sync_t *sync, const void *source_buffer, const void *compare_buffer, void *result_buffer,
ompi_datatype_t *datatype, ompi_osc_rdma_peer_t *peer, uint64_t target_address,
mca_btl_base_registration_handle_t *target_handle)
{
ompi_osc_rdma_module_t *module = sync->module;
const size_t btl_alignment_mask = ALIGNMENT_MASK(module->selected_btl->btl_get_alignment);
unsigned long offset, aligned_len, len = datatype->super.size;
ompi_osc_rdma_frag_t *frag = NULL;
ompi_osc_rdma_request_t *request;
char *ptr = NULL;
int ret;
OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, request);
request->internal = true;
request->type = OMPI_OSC_RDMA_TYPE_CSWAP;
request->sync = sync;
OPAL_THREAD_LOCK(&module->lock);
/* to ensure order wait until the previous accumulate completes */
while (ompi_osc_rdma_peer_is_accumulating (peer)) {
OPAL_THREAD_UNLOCK(&module->lock);
ompi_osc_rdma_progress (module);
OPAL_THREAD_LOCK(&module->lock);
}
peer->flags |= OMPI_OSC_RDMA_PEER_ACCUMULATING;
OPAL_THREAD_UNLOCK(&module->lock);
offset = target_address & btl_alignment_mask;;
aligned_len = (len + offset + btl_alignment_mask) & ~btl_alignment_mask;
ret = ompi_osc_rdma_frag_alloc (module, aligned_len, &frag, &ptr);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
peer->flags &= ~OMPI_OSC_RDMA_PEER_ACCUMULATING;
OPAL_OUTPUT_VERBOSE((10, ompi_osc_base_framework.framework_output,
"Could not allocate an rdma fragment for get accumulate. Falling back on point-to-point"));
return OMPI_ERR_OUT_OF_RESOURCE;
}
if (!ompi_osc_rdma_peer_is_exclusive (peer)) {
(void) ompi_osc_rdma_lock_acquire_exclusive (module, peer, offsetof (ompi_osc_rdma_state_t, accumulate_lock));
}
/* set up the request */
request->frag = frag;
request->origin_addr = (void *) source_buffer;
request->ctx = (void *) target_handle;
request->result_addr = result_buffer;
request->compare_addr = compare_buffer;
request->result_dt = datatype;
request->offset = (ptrdiff_t) offset;
request->target_address = target_address;
request->len = len;
OPAL_OUTPUT_VERBOSE((60, ompi_osc_base_framework.framework_output, "initiating btl get..."));
ret = module->selected_btl->btl_get (module->selected_btl, peer->data_endpoint, ptr,
target_address, frag->handle, target_handle,
aligned_len, 0, MCA_BTL_NO_ORDER,
ompi_osc_rdma_cas_get_complete, request, NULL);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
ompi_osc_rdma_frag_complete (frag);
return ret;
}
ompi_osc_rdma_sync_rdma_inc (sync);
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;
}
static inline int ompi_osc_rdma_rget_internal (void *origin_addr, int origin_count,
struct ompi_datatype_t *origin_dt,
int target,
OPAL_PTRDIFF_TYPE target_disp,
int target_count,
struct ompi_datatype_t *target_dt,
struct ompi_win_t *win, bool release_req,
struct ompi_request_t **request)
{
int ret, tag;
ompi_osc_rdma_module_t *module = GET_MODULE(win);
bool is_long_datatype = false;
ompi_osc_rdma_frag_t *frag;
ompi_osc_rdma_header_get_t *header;
size_t ddt_len, frag_len;
char *ptr;
const void *packed_ddt;
ompi_osc_rdma_request_t *rdma_request;
OPAL_OUTPUT_VERBOSE((50, ompi_osc_base_framework.framework_output,
"get: 0x%lx, %d, %s, %d, %d, %d, %s, %s",
(unsigned long) origin_addr, origin_count,
origin_dt->name, target, (int) target_disp,
target_count, target_dt->name, win->w_name));
if (!ompi_osc_rdma_check_access_epoch (module, target)) {
return OMPI_ERR_RMA_SYNC;
}
/* gets are always request based, so that we know where to land the data */
OMPI_OSC_RDMA_REQUEST_ALLOC(win, rdma_request);
if (NULL == rdma_request) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
rdma_request->internal = release_req;
/* short-circuit case */
if (0 == origin_count || 0 == target_count) {
ompi_osc_rdma_request_complete (rdma_request, MPI_SUCCESS);
*request = &rdma_request->super;
return OMPI_SUCCESS;
}
/* optimize self communication. TODO: optimize local communication */
if (ompi_comm_rank (module->comm) == target) {
*request = &rdma_request->super;
return ompi_osc_rdma_get_self (origin_addr, origin_count, origin_dt,
target_disp, target_count, target_dt,
module, rdma_request);
}
rdma_request->type = OMPI_OSC_RDMA_HDR_TYPE_GET;
rdma_request->origin_addr = origin_addr;
rdma_request->origin_count = origin_count;
OBJ_RETAIN(origin_dt);
rdma_request->origin_dt = origin_dt;
/* Compute datatype length. Note that the datatype description
* must fit in a single frag */
ddt_len = ompi_datatype_pack_description_length(target_dt);
OPAL_THREAD_LOCK(&module->lock);
frag_len = sizeof(ompi_osc_rdma_header_get_t) + ddt_len;
ret = ompi_osc_rdma_frag_alloc(module, target, frag_len, &frag, &ptr);
if (OMPI_SUCCESS != ret) {
/* allocate space for the header plus space to store ddt_len */
frag_len = sizeof(ompi_osc_rdma_header_put_t) + 8;
ret = ompi_osc_rdma_frag_alloc(module, target, frag_len, &frag, &ptr);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OPAL_THREAD_UNLOCK(&module->lock);
return OMPI_ERR_OUT_OF_RESOURCE;
}
is_long_datatype = true;
}
tag = get_tag (module);
/* for bookkeeping the get is "outgoing" */
ompi_osc_signal_outgoing (module, target, 1);
/* flush will be called at the end of this function. make sure the post message has
* arrived. */
if (!release_req && module->sc_group) {
while (0 != module->num_post_msgs) {
OPAL_OUTPUT_VERBOSE((50, ompi_osc_base_framework.framework_output,
"waiting for post messages. num_post_msgs = %d", module->num_post_msgs));
opal_condition_wait(&module->cond, &module->lock);
}
}
OPAL_THREAD_UNLOCK(&module->lock);
header = (ompi_osc_rdma_header_get_t*) ptr;
header->base.type = OMPI_OSC_RDMA_HDR_TYPE_GET;
header->base.flags = 0;
header->len = frag_len;
header->count = target_count;
header->displacement = target_disp;
header->tag = tag;
ptr += sizeof(ompi_osc_rdma_header_get_t);
do {
ret = ompi_datatype_get_pack_description(target_dt, &packed_ddt);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
break;
}
if (is_long_datatype) {
/* the datatype does not fit in an eager message. send it seperately */
header->base.flags |= OMPI_OSC_RDMA_HDR_FLAG_LARGE_DATATYPE;
OBJ_RETAIN(target_dt);
ret = ompi_osc_rdma_isend_w_cb ((void *) packed_ddt, ddt_len, MPI_BYTE, target,
tag, module->comm, ompi_osc_rdma_dt_send_complete,
target_dt);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
break;
}
*((uint64_t *) ptr) = ddt_len;
ptr += 8;
} else {
memcpy((unsigned char*) ptr, packed_ddt, ddt_len);
ptr += ddt_len;
}
/* TODO -- store the request somewhere so we can cancel it on error */
rdma_request->outstanding_requests = 1;
ret = ompi_osc_rdma_irecv_w_cb (origin_addr, origin_count, origin_dt, target, tag,
module->comm, NULL, ompi_osc_rdma_req_comm_complete, rdma_request);
} while (0);
if (OMPI_SUCCESS == ret) {
header->base.flags |= OMPI_OSC_RDMA_HDR_FLAG_VALID;
*request = &rdma_request->super;
}
OPAL_THREAD_LOCK(&module->lock);
ret = ompi_osc_rdma_frag_finish(module, frag);
if (!release_req) {
/* need to flush now in case the caller decides to wait on the request */
ompi_osc_rdma_frag_flush_target (module, target);
}
OPAL_THREAD_UNLOCK(&module->lock);
return ret;
}
int ompi_osc_rdma_compare_and_swap (void *origin_addr, void *compare_addr,
void *result_addr, struct ompi_datatype_t *dt,
int target, OPAL_PTRDIFF_TYPE target_disp,
struct ompi_win_t *win)
{
ompi_osc_rdma_module_t *module = GET_MODULE(win);
ompi_proc_t *proc = ompi_comm_peer_lookup(module->comm, target);
ompi_osc_rdma_frag_t *frag;
ompi_osc_rdma_header_cswap_t *header;
size_t ddt_len, payload_len, frag_len;
ompi_osc_rdma_request_t *request;
const void *packed_ddt;
int ret, tag;
char *ptr;
OPAL_OUTPUT_VERBOSE((50, ompi_osc_base_framework.framework_output,
"cswap: 0x%lx, 0x%lx, 0x%lx, %s, %d, %d, %s",
(unsigned long) origin_addr, (unsigned long) compare_addr,
(unsigned long) result_addr, dt->name, target, (int) target_disp,
win->w_name));
if (!ompi_osc_rdma_check_access_epoch (module, target)) {
return OMPI_ERR_RMA_SYNC;
}
/* optimize self case. TODO: optimize local case */
if (ompi_comm_rank (module->comm) == target) {
return ompi_osc_rdma_cas_self (origin_addr, compare_addr, result_addr, dt, target_disp,
module);
}
/* compare-and-swaps are always request based, so that we know where to land the data */
OMPI_OSC_RDMA_REQUEST_ALLOC(win, request);
if (NULL == request) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
request->type = OMPI_OSC_RDMA_HDR_TYPE_CSWAP;
request->origin_addr = origin_addr;
request->internal = true;
OBJ_RETAIN(dt);
request->origin_dt = dt;
/* Compute datatype and payload lengths. Note that the datatype description
* must fit in a single frag. It should be small in this case. */
ddt_len = ompi_datatype_pack_description_length(dt);
/* we need to send both the origin and compare buffers */
payload_len = dt->super.size * 2;
OPAL_THREAD_LOCK(&module->lock);
frag_len = sizeof(ompi_osc_rdma_header_cswap_t) + ddt_len + payload_len;
ret = ompi_osc_rdma_frag_alloc(module, target, frag_len, &frag, &ptr);
if (OMPI_SUCCESS != ret) {
OPAL_THREAD_UNLOCK(&module->lock);
return OMPI_ERR_OUT_OF_RESOURCE;
}
tag = get_tag (module);
ompi_osc_signal_outgoing (module, target, 1);
header = (ompi_osc_rdma_header_cswap_t *) ptr;
header->base.type = OMPI_OSC_RDMA_HDR_TYPE_CSWAP;
header->base.flags = OMPI_OSC_RDMA_HDR_FLAG_VALID;
header->len = frag_len;
header->displacement = target_disp;
header->tag = tag;
ptr += sizeof(ompi_osc_rdma_header_cswap_t);
ret = ompi_datatype_get_pack_description(dt, &packed_ddt);
memcpy((unsigned char*) ptr, packed_ddt, ddt_len);
ptr += ddt_len;
/* pack the origin and compare data */
osc_rdma_copy_for_send (ptr, dt->super.size, origin_addr, proc, 1, dt);
ptr += dt->super.size;
osc_rdma_copy_for_send (ptr, dt->super.size, compare_addr, proc, 1, dt);
request->outstanding_requests = 1;
ret = ompi_osc_rdma_irecv_w_cb (result_addr, 1, dt, target, tag, module->comm,
NULL, ompi_osc_rdma_req_comm_complete, request);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OPAL_THREAD_UNLOCK(&module->lock);
return ret;
}
ret = ompi_osc_rdma_frag_finish(module, frag);
OPAL_THREAD_UNLOCK(&module->lock);
return ret;
}
static inline
int ompi_osc_rdma_rget_accumulate_internal (void *origin_addr, int origin_count,
struct ompi_datatype_t *origin_datatype,
void *result_addr, int result_count,
struct ompi_datatype_t *result_datatype,
int target_rank, MPI_Aint target_disp,
int target_count, struct ompi_datatype_t *target_datatype,
struct ompi_op_t *op, struct ompi_win_t *win,
bool release_req, struct ompi_request_t **request)
{
int ret;
ompi_osc_rdma_module_t *module = GET_MODULE(win);
ompi_proc_t *proc = ompi_comm_peer_lookup(module->comm, target_rank);
bool is_long_datatype = false;
bool is_long_msg = false;
ompi_osc_rdma_frag_t *frag;
ompi_osc_rdma_header_acc_t *header;
size_t ddt_len, payload_len, frag_len;
char *ptr;
const void *packed_ddt;
int tag;
ompi_osc_rdma_request_t *rdma_request;
OPAL_OUTPUT_VERBOSE((50, ompi_osc_base_framework.framework_output,
"rget_acc: 0x%lx, %d, %s, 0x%lx, %d, %s, 0x%x, %d, %d, %s, %s, %s",
(unsigned long) origin_addr, origin_count, origin_datatype->name,
(unsigned long) result_addr, result_count, result_datatype->name,
target_rank, (int) target_disp, target_count, target_datatype->name,
op->o_name, win->w_name));
if (!ompi_osc_rdma_check_access_epoch (module, target_rank)) {
return OMPI_ERR_RMA_SYNC;
}
/* get_accumulates are always request based, so that we know where to land the data */
OMPI_OSC_RDMA_REQUEST_ALLOC(win, rdma_request);
if (OPAL_UNLIKELY(NULL == rdma_request)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
rdma_request->internal = release_req;
/* short-circuit case. note that origin_count may be 0 if op is MPI_NO_OP */
if (0 == result_count || 0 == target_count) {
ompi_osc_rdma_request_complete (rdma_request, MPI_SUCCESS);
*request = &rdma_request->super;
return OMPI_SUCCESS;
}
/* optimize the self case. TODO: optimize the local case */
if (ompi_comm_rank (module->comm) == target_rank) {
*request = &rdma_request->super;
return ompi_osc_rdma_gacc_self (origin_addr, origin_count, origin_datatype,
result_addr, result_count, result_datatype,
target_disp, target_count, target_datatype,
op, module, rdma_request);
}
rdma_request->type = OMPI_OSC_RDMA_HDR_TYPE_GET_ACC;
rdma_request->origin_addr = origin_addr;
rdma_request->origin_count = origin_count;
OBJ_RETAIN(origin_datatype);
rdma_request->origin_dt = origin_datatype;
/* Compute datatype and payload lengths. Note that the datatype description
* must fit in a single frag */
ddt_len = ompi_datatype_pack_description_length(target_datatype);
if (&ompi_mpi_op_no_op.op != op) {
payload_len = origin_datatype->super.size * origin_count;
} else {
payload_len = 0;
}
OPAL_THREAD_LOCK(&module->lock);
frag_len = sizeof(*header) + ddt_len + payload_len;
ret = ompi_osc_rdma_frag_alloc(module, target_rank, frag_len, &frag, &ptr);
if (OMPI_SUCCESS != ret) {
frag_len = sizeof(*header) + ddt_len;
ret = ompi_osc_rdma_frag_alloc(module, target_rank, frag_len, &frag, &ptr);
if (OMPI_SUCCESS != ret) {
/* allocate space for the header plus space to store ddt_len */
frag_len = sizeof(*header) + 8;
ret = ompi_osc_rdma_frag_alloc(module, target_rank, frag_len, &frag, &ptr);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OPAL_THREAD_UNLOCK(&module->lock);
return OMPI_ERR_OUT_OF_RESOURCE;
}
is_long_datatype = true;
}
is_long_msg = true;
}
tag = get_tag (module);
/* If this is a long message then we need two completions before the
* request is complete (1 for the send, 1 for the receive) */
rdma_request->outstanding_requests = 1 + is_long_msg;
/* increment the number of outgoing fragments */
ompi_osc_signal_outgoing (module, target_rank, rdma_request->outstanding_requests);
/* flush will be called at the end of this function. make sure the post message has
* arrived. */
if (!release_req && module->sc_group) {
while (0 != module->num_post_msgs) {
OPAL_OUTPUT_VERBOSE((50, ompi_osc_base_framework.framework_output,
"waiting for post messages. num_post_msgs = %d", module->num_post_msgs));
opal_condition_wait(&module->cond, &module->lock);
}
}
OPAL_THREAD_UNLOCK(&module->lock);
header = (ompi_osc_rdma_header_acc_t *) ptr;
header->base.flags = 0;
header->len = frag_len;
header->count = target_count;
header->displacement = target_disp;
header->op = op->o_f_to_c_index;
header->tag = tag;
ptr = (char *)(header + 1);
do {
ret = ompi_datatype_get_pack_description(target_datatype, &packed_ddt);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
break;
}
if (is_long_datatype) {
/* the datatype does not fit in an eager message. send it seperately */
header->base.flags |= OMPI_OSC_RDMA_HDR_FLAG_LARGE_DATATYPE;
OBJ_RETAIN(target_datatype);
ret = ompi_osc_rdma_isend_w_cb ((void *) packed_ddt, ddt_len, MPI_BYTE, target_rank,
tag, module->comm, ompi_osc_rdma_dt_send_complete,
target_datatype);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
break;
}
*((uint64_t *) ptr) = ddt_len;
ptr += 8;
} else {
memcpy((unsigned char*) ptr, packed_ddt, ddt_len);
ptr += ddt_len;
}
ret = ompi_osc_rdma_irecv_w_cb (result_addr, result_count, result_datatype, target_rank, tag,
module->comm, NULL, ompi_osc_rdma_req_comm_complete, rdma_request);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
break;
}
if (!is_long_msg) {
header->base.type = OMPI_OSC_RDMA_HDR_TYPE_GET_ACC;
if (&ompi_mpi_op_no_op.op != op) {
osc_rdma_copy_for_send (ptr, payload_len, origin_addr, proc, origin_count,
origin_datatype);
}
} else {
header->base.type = OMPI_OSC_RDMA_HDR_TYPE_GET_ACC_LONG;
ret = ompi_osc_rdma_isend_w_cb (origin_addr, origin_count, origin_datatype, target_rank,
tag, module->comm, ompi_osc_rdma_req_comm_complete, rdma_request);
}
} while (0);
if (OMPI_SUCCESS == ret) {
header->base.flags |= OMPI_OSC_RDMA_HDR_FLAG_VALID;
*request = (ompi_request_t *) rdma_request;
}
OPAL_THREAD_LOCK(&module->lock);
ret = ompi_osc_rdma_frag_finish(module, frag);
if (!release_req) {
/* need to flush now in case the caller decides to wait on the request */
ompi_osc_rdma_frag_flush_target (module, target_rank);
}
OPAL_THREAD_UNLOCK(&module->lock);
return ret;
}
