TL/UCP: add reduce srg knomial #1058
Conversation
| reduce/reduce.c \ | ||
| reduce/reduce_knomial.c \ | ||
| reduce/reduce_dbt.c \ | ||
| reduce/reduce_srg_knomial.c |
| @@ -499,6 +513,9 @@ ucc_tl_ucp_reduce_scatter_knomial_init_r(ucc_base_coll_args_t *coll_args, | |||
| ucc_kn_rsx_pattern_init(size, rank, radix, | |||
| count, &task->reduce_scatter_kn.p); | |||
|
|
|||
| } else if (ct == UCC_COLL_TYPE_REDUCE) { | |||
| ucc_kn_rsx_pattern_init(size, VRANK(rank, coll_args->args.root, size), | |||
There was a problem hiding this comment.
add reordering for reduce
| UCC_TL_TEAM_SIZE(tl_team), | ||
| count); | ||
|
|
||
| /* 1st step of reduce: knomial reduce_scatter */ |
There was a problem hiding this comment.
move comment up
| cfg_radix = ucc_tl_ucp_get_radix_from_range(tl_team, | ||
| count * ucc_dt_size(dt), | ||
| mt, p, 4); | ||
| radix = ucc_knomial_pattern_get_min_radix(cfg_radix, |
There was a problem hiding this comment.
add tl_trace
| args.args.dst.info.buffer = rs_rbuf; | ||
| args.args.mask |= UCC_COLL_ARGS_FIELD_FLAGS; | ||
| args.args.flags |= UCC_COLL_ARGS_FLAG_IN_PLACE; | ||
|
|
There was a problem hiding this comment.
rem empty line
| targs->dst.info.count = frag_count; | ||
|
|
||
| targs = &frag->tasks[1]->bargs.args; /* GATHER */ | ||
| targs->src.info.buffer = PTR_OFFSET(targs->src.info.buffer, offset * dt_size);; |
| } | ||
|
|
||
| if (!UCC_IS_ROOT(*args, trank) || UCC_IS_INPLACE(*args)) { | ||
| st = ucc_mc_alloc(&schedule->scratch_mc_header, count * dt_size, mt); |
There was a problem hiding this comment.
Can you avoid the scratch buffer allocation? For example, in the first call of this collective algorithm, allocate a large enough buffer for this team, and reuse this scratch buffer for the future calls into this collective.
| root = args->root; | ||
| rank = VRANK(rank, root, size); | ||
| } | ||
|
|
||
| UCC_KN_REDUCE_GOTO_PHASE(task->reduce_scatter_kn.phase); | ||
| block_count = ucc_sra_kn_compute_block_count(count, rank, p); | ||
| get_rs_work_buf(task, block_count, &wb); | ||
| if (KN_NODE_EXTRA == node_type) { | ||
| peer = ucc_ep_map_eval(task->subset.map, |
There was a problem hiding this comment.
I see that in many places we are trying to find out the actual peer. Can you add a new function like knomial_get_actual_peer() that unifies all these VRANK/INV_VRANK/ucc_ep_map_eval and use this function instead?
|
@Sergei-Lebedev very solid PR, Thanks for creating this new SRG reduce algorithm. Please see my comments. |
| @@ -1,5 +1,5 @@ | |||
| /** | |||
| * Copyright (c) 2021-2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. | |||
| * Copyright (c) 2021-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved. | |||
| @@ -1,5 +1,5 @@ | |||
| /** | |||
| * Copyright (c) 2021-2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. | |||
| * Copyright (c) 2021-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved. | |||
| size_t frag_count; | ||
| size_t offset; | ||
| ucc_coll_args_t *targs; | ||
| ucc_rank_t rank; |
| team = TASK_TEAM(&schedule_p->super); | ||
| rank = UCC_TL_TEAM_RANK(team); | ||
| if (UCC_IS_ROOT(*args, rank)) { | ||
| count = args->dst.info.count; |
| dt_size = ucc_dt_size(args->dst.info.datatype); | ||
| } else { | ||
| count = args->src.info.count; | ||
| dt_size = ucc_dt_size(args->src.info.datatype); |
| /* 1st step of reduce: knomial reduce_scatter */ | ||
| UCC_CHECK_GOTO(ucc_tl_ucp_reduce_scatter_knomial_init_r(&args, team, &task, | ||
| radix), | ||
| out, status); |
There was a problem hiding this comment.
not sure if it's just me, but looks like out, status); have a weird alignment, or should be no newline
What
Add reduce srg knomial algorithm to TL UCP
Why ?
Improves performance of large msg reduce collective
How ?
Implemented as schedule of reduce scatter knomial followed by gather knomial