Skip to content

TL/UCP: use memcpy instead of sendrecv in agv #1255

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Sergei-Lebedev wants to merge 1 commit into
base: master
Choose a base branch
from
Open

TL/UCP: use memcpy instead of sendrecv in agv #1255

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
89 changes: 48 additions & 41 deletions src/components/tl/ucp/allgatherv/allgatherv_ring.c
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/**
* Copyright (c) 2021-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* Copyright (c) 2021-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
*
* See file LICENSE for terms.
*/
Expand Down Expand Up @@ -32,41 +32,53 @@ void ucc_tl_ucp_allgatherv_ring_progress(ucc_coll_task_t *coll_task)
sendto = ucc_ep_map_eval(task->subset.map, (trank + 1) % tsize);
recvfrom = ucc_ep_map_eval(task->subset.map, (trank - 1 + tsize) % tsize);

while (task->tagged.send_posted < tsize) {
send_idx =
ucc_ep_map_eval(task->subset.map, (trank -
task->tagged.send_posted + 1 +
tsize) % tsize);
while (task->tagged.send_posted < tsize - 1) {
send_idx = ucc_ep_map_eval(
task->subset.map,
(trank - task->tagged.send_posted + tsize) % tsize);
data_displ = ucc_coll_args_get_displacement(
args, args->dst.info_v.displacements, send_idx) *
rdt_size;
data_size =
ucc_coll_args_get_count(args, args->dst.info_v.counts, send_idx) *
rdt_size;
UCPCHECK_GOTO(ucc_tl_ucp_send_nb((void *)(rbuf + data_displ), data_size,
rmem, sendto, team, task),
task, out);
recv_idx =
ucc_ep_map_eval(task->subset.map, (trank -
task->tagged.recv_posted +
tsize) % tsize);
data_size = ucc_coll_args_get_count(
args, args->dst.info_v.counts, send_idx) *
rdt_size;
UCPCHECK_GOTO(
ucc_tl_ucp_send_nb(
(void *)(rbuf + data_displ),
data_size,
rmem,
sendto,
team,
task),
task,
out);
recv_idx = ucc_ep_map_eval(
task->subset.map,
(trank - task->tagged.recv_posted - 1 + tsize) % tsize);
data_displ = ucc_coll_args_get_displacement(
args, args->dst.info_v.displacements, recv_idx) *
rdt_size;
data_size =
ucc_coll_args_get_count(args, args->dst.info_v.counts, recv_idx) *
rdt_size;
UCPCHECK_GOTO(ucc_tl_ucp_recv_nb((void *)(rbuf + data_displ), data_size,
rmem, recvfrom, team, task),
task, out);
data_size = ucc_coll_args_get_count(
args, args->dst.info_v.counts, recv_idx) *
rdt_size;
UCPCHECK_GOTO(
ucc_tl_ucp_recv_nb(
(void *)(rbuf + data_displ),
data_size,
rmem,
recvfrom,
team,
task),
task,
out);
if (UCC_INPROGRESS == ucc_tl_ucp_test(task)) {
return;
}
}
ucc_assert(UCC_TL_UCP_TASK_P2P_COMPLETE(task));
task->super.status = UCC_OK;
out:
return;
UCC_TL_UCP_PROFILE_REQUEST_EVENT(coll_task, "ucp_allgatherv_ring_done", 0);
}

ucc_status_t ucc_tl_ucp_allgatherv_ring_start(ucc_coll_task_t *coll_task)
Expand All @@ -80,31 +92,26 @@ ucc_status_t ucc_tl_ucp_allgatherv_ring_start(ucc_coll_task_t *coll_task)
ucc_memory_type_t rmem = args->dst.info_v.mem_type;
ucc_rank_t grank = UCC_TL_TEAM_RANK(team);
size_t data_size, data_displ, rdt_size;
ucc_status_t status;

UCC_TL_UCP_PROFILE_REQUEST_EVENT(coll_task, "ucp_allgatherv_ring_start", 0);
ucc_tl_ucp_task_reset(task, UCC_INPROGRESS);

if (!UCC_IS_INPLACE(*args)) {
/* TODO replace local sendrecv with memcpy? */
rdt_size = ucc_dt_size(args->dst.info_v.datatype);
data_displ = ucc_coll_args_get_displacement(args,
args->dst.info_v.displacements, grank) * rdt_size;
data_size = ucc_coll_args_get_count(args,
args->dst.info_v.counts, grank) * rdt_size;
UCPCHECK_GOTO(ucc_tl_ucp_recv_nb(PTR_OFFSET(rbuf, data_displ), data_size,
rmem, grank, team, task),
task, error);
UCPCHECK_GOTO(ucc_tl_ucp_send_nb(sbuf, data_size, smem, grank, team, task),
task, error);
} else {
/* to simplify progress fucnction and make it identical for
in-place and non in-place */
task->tagged.send_posted = task->tagged.send_completed = 1;
task->tagged.recv_posted = task->tagged.recv_completed = 1;
data_displ = ucc_coll_args_get_displacement(
args, args->dst.info_v.displacements, grank) *
rdt_size;
data_size = ucc_coll_args_get_count(
args, args->dst.info_v.counts, grank) *
rdt_size;
status = ucc_mc_memcpy(
Copy link
Collaborator

@wfaderhold21 wfaderhold21 Jan 14, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM! question though: would ucc_mc_memcpy potentially cause a deadlock if using cuda buffers for a scenario of NCCL + UCC? should this be a local copy instead?

PTR_OFFSET(rbuf, data_displ), sbuf, data_size, rmem, smem);
if (ucc_unlikely(UCC_OK != status)) {
return status;
}
}

return ucc_progress_queue_enqueue(UCC_TL_CORE_CTX(team)->pq, &task->super);
error:
return task->super.status;
}

ucc_status_t ucc_tl_ucp_allgatherv_ring_init_common(ucc_tl_ucp_task_t *task)
Expand Down