CORE: fix bug for triggered colls in tl/ucp (openucx#757)

jeniaka · Jun 21, 2023 · f8516a9 · f8516a9
1 parent 4ec3bc9
commit f8516a9
Show file tree

Hide file tree

Showing 20 changed files with 18 additions and 38 deletions.
diff --git a/src/components/cl/hier/allreduce/allreduce_rab.c b/src/components/cl/hier/allreduce/allreduce_rab.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2021-2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2021-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -160,7 +160,6 @@ ucc_cl_hier_allreduce_rab_init_schedule(ucc_base_coll_args_t *coll_args,
     schedule->super.post           = ucc_cl_hier_allreduce_rab_start;
     schedule->super.progress       = NULL;
     schedule->super.finalize       = ucc_cl_hier_allreduce_rab_finalize;
-    schedule->super.triggered_post = ucc_triggered_post;
     *sched_p                       = schedule;
     return UCC_OK;
 
@@ -243,7 +242,6 @@ UCC_CL_HIER_PROFILE_FUNC(ucc_status_t, ucc_cl_hier_allreduce_rab_init,
     }
 
     schedule->super.super.super.post = ucc_cl_hier_rab_allreduce_start;
-    schedule->super.super.super.triggered_post = ucc_triggered_post;
     schedule->super.super.super.finalize = ucc_cl_hier_ar_rab_schedule_finalize;
     *task                                = &schedule->super.super.super;
     return UCC_OK;

diff --git a/src/components/cl/hier/allreduce/allreduce_split_rail.c b/src/components/cl/hier/allreduce/allreduce_split_rail.c
@@ -324,7 +324,6 @@ UCC_CL_HIER_PROFILE_FUNC(ucc_status_t, ucc_cl_hier_allreduce_split_rail_init,
     }
 
     schedule->super.super.super.post = ucc_cl_hier_split_rail_allreduce_start;
-    schedule->super.super.super.triggered_post = ucc_triggered_post;
     schedule->super.super.super.finalize =
         ucc_cl_hier_ar_split_rail_schedule_finalize;
     *task = &schedule->super.super.super;

diff --git a/src/components/cl/hier/alltoallv/alltoallv.c b/src/components/cl/hier/alltoallv/alltoallv.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * Copyright (c) Meta Platforms, Inc. and affiliates. 2022.
  *
  * See file LICENSE for terms.
@@ -260,7 +260,6 @@ UCC_CL_HIER_PROFILE_FUNC(ucc_status_t, ucc_cl_hier_alltoallv_init,
     schedule->super.post           = ucc_cl_hier_alltoallv_start;
     schedule->super.progress       = NULL;
     schedule->super.finalize       = ucc_cl_hier_alltoallv_finalize;
-    schedule->super.triggered_post = ucc_triggered_post;
     schedule->super.triggered_post_setup =
         ucc_cl_hier_alltoallv_triggered_post_setup;
     *task = &schedule->super;

diff --git a/src/components/cl/hier/bcast/bcast_2step.c b/src/components/cl/hier/bcast/bcast_2step.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -172,7 +172,6 @@ ucc_cl_hier_bcast_2step_init_schedule(ucc_base_coll_args_t *coll_args,
     schedule->super.post           = ucc_cl_hier_bcast_2step_start;
     schedule->super.progress       = NULL;
     schedule->super.finalize       = ucc_cl_hier_bcast_2step_finalize;
-    schedule->super.triggered_post = ucc_triggered_post;
     *sched_p                       = schedule;
     return UCC_OK;
 
@@ -251,7 +250,6 @@ UCC_CL_HIER_PROFILE_FUNC(ucc_status_t, ucc_cl_hier_bcast_2step_init,
     }
 
     schedule->super.super.super.post           = ucc_cl_hier_2step_bcast_start;
-    schedule->super.super.super.triggered_post = ucc_triggered_post;
     schedule->super.super.super.finalize       =
         ucc_cl_hier_bcast_2step_schedule_finalize;
     *task = &schedule->super.super.super;

diff --git a/src/components/tl/cuda/allgather/allgather_linear.c b/src/components/tl/cuda/allgather/allgather_linear.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -32,7 +32,6 @@ ucc_status_t ucc_tl_cuda_allgather_linear_init(ucc_base_coll_args_t *coll_args,
 
     task->super.flags |= UCC_COLL_TASK_FLAG_EXECUTOR;
     task->super.post           = ucc_tl_cuda_allgatherv_linear_start;
-    task->super.triggered_post = ucc_triggered_post;
     task->super.progress       = ucc_tl_cuda_allgatherv_linear_progress;
     task->super.finalize       = ucc_tl_cuda_allgatherv_linear_finalize;
     task->bar                  = TASK_BAR(task);

diff --git a/src/components/tl/cuda/allgather/allgather_ring.c b/src/components/tl/cuda/allgather/allgather_ring.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -29,7 +29,6 @@ ucc_status_t ucc_tl_cuda_allgather_ring_init(ucc_base_coll_args_t *coll_args,
 
     task->super.flags               |= UCC_COLL_TASK_FLAG_EXECUTOR;
     task->super.post                = ucc_tl_cuda_allgatherv_ring_start;
-    task->super.triggered_post      = ucc_triggered_post;
     task->super.progress            = ucc_tl_cuda_allgatherv_ring_progress;
     task->super.finalize            = ucc_tl_cuda_allgatherv_ring_finalize;
     task->bar                       = TASK_BAR(task);

diff --git a/src/components/tl/cuda/allgatherv/allgatherv_linear.c b/src/components/tl/cuda/allgatherv/allgatherv_linear.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -443,7 +443,6 @@ ucc_status_t ucc_tl_cuda_allgatherv_linear_init(ucc_base_coll_args_t *coll_args,
     task->allgatherv_linear.dt         = coll_args->args.dst.info_v.datatype;
     task->super.flags |= UCC_COLL_TASK_FLAG_EXECUTOR;
     task->super.post           = ucc_tl_cuda_allgatherv_linear_start;
-    task->super.triggered_post = ucc_triggered_post;
     task->super.progress       = ucc_tl_cuda_allgatherv_linear_progress;
     task->super.finalize       = ucc_tl_cuda_allgatherv_linear_finalize;
     task->bar                  = TASK_BAR(task);

diff --git a/src/components/tl/cuda/allgatherv/allgatherv_ring.c b/src/components/tl/cuda/allgatherv/allgatherv_ring.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -431,7 +431,6 @@ ucc_status_t ucc_tl_cuda_allgatherv_ring_init(ucc_base_coll_args_t *coll_args,
 
     task->super.flags               |= UCC_COLL_TASK_FLAG_EXECUTOR;
     task->super.post                 = ucc_tl_cuda_allgatherv_ring_start;
-    task->super.triggered_post       = ucc_triggered_post;
     task->super.progress             = ucc_tl_cuda_allgatherv_ring_progress;
     task->super.finalize             = ucc_tl_cuda_allgatherv_ring_finalize;
     task->bar                        = TASK_BAR(task);

diff --git a/src/components/tl/cuda/alltoall/alltoall_ce.c b/src/components/tl/cuda/alltoall/alltoall_ce.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2021-2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2021-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * Copyright (c) Meta Platforms, Inc. and affiliates. 2022.
  *
  * See file LICENSE for terms.
@@ -67,7 +67,6 @@ ucc_status_t ucc_tl_cuda_alltoall_ce_init(ucc_tl_cuda_task_t *task)
     }
 
     task->super.post           = ucc_tl_cuda_alltoallv_ce_start;
-    task->super.triggered_post = ucc_triggered_post;
     task->super.triggered_post_setup =
         ucc_tl_cuda_alltoallv_ce_triggered_post_setup;
     task->super.progress = ucc_tl_cuda_alltoallv_ce_progress;

diff --git a/src/components/tl/cuda/alltoallv/alltoallv_ce.c b/src/components/tl/cuda/alltoallv/alltoallv_ce.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2021-2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2021-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * Copyright (c) Meta Platforms, Inc. and affiliates. 2022.
  *
  * See file LICENSE for terms.
@@ -455,7 +455,6 @@ ucc_status_t ucc_tl_cuda_alltoallv_ce_init(ucc_tl_cuda_task_t *task)
 
     task->super.flags |= UCC_COLL_TASK_FLAG_EXECUTOR;
     task->super.post           = ucc_tl_cuda_alltoallv_ce_start;
-    task->super.triggered_post = ucc_triggered_post;
     task->super.triggered_post_setup =
         ucc_tl_cuda_alltoallv_ce_triggered_post_setup;
     task->super.progress = ucc_tl_cuda_alltoallv_ce_progress;

diff --git a/src/components/tl/cuda/reduce_scatter/reduce_scatter_linear.c b/src/components/tl/cuda/reduce_scatter/reduce_scatter_linear.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -36,7 +36,6 @@ ucc_status_t ucc_tl_cuda_reduce_scatter_linear_init(ucc_base_coll_args_t *coll_a
     task->reduce_scatterv_linear.rbuf       = coll_args->args.dst.info.buffer;
     task->super.flags          |= UCC_COLL_TASK_FLAG_EXECUTOR;
     task->super.post           = ucc_tl_cuda_reduce_scatterv_linear_start;
-    task->super.triggered_post = ucc_triggered_post;
     task->super.progress       = ucc_tl_cuda_reduce_scatterv_linear_progress;
     task->super.finalize       = ucc_tl_cuda_reduce_scatterv_linear_finalize;
     task->bar                  = TASK_BAR(task);

diff --git a/src/components/tl/cuda/reduce_scatter/reduce_scatter_ring.c b/src/components/tl/cuda/reduce_scatter/reduce_scatter_ring.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -39,7 +39,6 @@ ucc_status_t ucc_tl_cuda_reduce_scatter_ring_init(ucc_base_coll_args_t *coll_arg
     task->super.flags                    |= UCC_COLL_TASK_FLAG_EXECUTOR;
     task->reduce_scatterv_ring.num_frags = ucc_div_round_up(send_size, frag_size);
     task->super.post                     = ucc_tl_cuda_reduce_scatterv_ring_start;
-    task->super.triggered_post           = ucc_triggered_post;
     task->super.progress                 = ucc_tl_cuda_reduce_scatterv_ring_progress;
     task->super.finalize                 = ucc_tl_cuda_reduce_scatterv_ring_finalize;
     task->bar                            = TASK_BAR(task);

diff --git a/src/components/tl/cuda/reduce_scatterv/reduce_scatterv_linear.c b/src/components/tl/cuda/reduce_scatterv/reduce_scatterv_linear.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -452,7 +452,6 @@ ucc_tl_cuda_reduce_scatterv_linear_init(ucc_base_coll_args_t *coll_args,
     task->super.post           = ucc_tl_cuda_reduce_scatterv_linear_start;
     task->super.progress       = ucc_tl_cuda_reduce_scatterv_linear_progress;
     task->super.finalize       = ucc_tl_cuda_reduce_scatterv_linear_finalize;
-    task->super.triggered_post = ucc_triggered_post;
     task->bar                  = TASK_BAR(task);
 
     *task_p = &task->super;

diff --git a/src/components/tl/cuda/reduce_scatterv/reduce_scatterv_ring.c b/src/components/tl/cuda/reduce_scatterv/reduce_scatterv_ring.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -348,7 +348,6 @@ ucc_tl_cuda_reduce_scatterv_ring_init(ucc_base_coll_args_t *coll_args,
 
     task->super.flags                    |= UCC_COLL_TASK_FLAG_EXECUTOR;
     task->super.post                      = ucc_tl_cuda_reduce_scatterv_ring_start;
-    task->super.triggered_post            = ucc_triggered_post;
     task->super.progress                  = ucc_tl_cuda_reduce_scatterv_ring_progress;
     task->super.finalize                  = ucc_tl_cuda_reduce_scatterv_ring_finalize;
     task->reduce_scatterv_ring.get_count  = ucc_tl_cuda_reduce_scatterv_get_count;

diff --git a/src/components/tl/self/tl_self_coll.c b/src/components/tl/self/tl_self_coll.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2021-2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2021-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * Copyright (c) Meta Platforms, Inc. and affiliates. 2022.
  *
  * See file LICENSE for terms.
@@ -24,7 +24,6 @@ ucc_tl_self_coll_init_task(ucc_base_coll_args_t *coll_args,
     ucc_coll_task_init(&task->super, coll_args, team);
     UCC_TL_SELF_PROFILE_REQUEST_NEW(task, "tl_self_task", 0);
     task->super.finalize       = ucc_tl_self_coll_finalize;
-    task->super.triggered_post = ucc_triggered_post;
     task->src                  = NULL;
     task->dst                  = NULL;
     task->size                 = 0;

diff --git a/src/components/tl/sharp/tl_sharp_team.c b/src/components/tl/sharp/tl_sharp_team.c
@@ -217,7 +217,6 @@ ucc_status_t ucc_tl_sharp_coll_init(ucc_base_coll_args_t *coll_args,
 
     task->req_handle           = NULL;
     task->super.finalize       = ucc_tl_sharp_coll_finalize;
-    task->super.triggered_post = ucc_triggered_post;
 
     switch (coll_args->args.coll_type)
     {

diff --git a/src/components/tl/ucp/allreduce/allreduce_sra_knomial.c b/src/components/tl/ucp/allreduce/allreduce_sra_knomial.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2021-2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2021-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -214,7 +214,6 @@ ucc_tl_ucp_allreduce_sra_knomial_init(ucc_base_coll_args_t *coll_args,
     }
     schedule_p->super.super.finalize =
         ucc_tl_ucp_allreduce_sra_knomial_finalize;
-    schedule_p->super.super.triggered_post = ucc_triggered_post;
     schedule_p->super.super.post = ucc_tl_ucp_allreduce_sra_knomial_start;
     *task_h                                = &schedule_p->super.super;
     return UCC_OK;

diff --git a/src/components/tl/ucp/bcast/bcast_sag_knomial.c b/src/components/tl/ucp/bcast/bcast_sag_knomial.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2021-2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2021-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  *
  * See file LICENSE for terms.
  */
@@ -125,7 +125,6 @@ ucc_tl_ucp_bcast_sag_knomial_init(ucc_base_coll_args_t *coll_args,
     schedule->super.post           = ucc_tl_ucp_bcast_sag_knomial_start;
     schedule->super.progress       = NULL;
     schedule->super.finalize       = ucc_tl_ucp_bcast_sag_knomial_finalize;
-    schedule->super.triggered_post = ucc_triggered_post;
     *task_h                        = &schedule->super;
     return UCC_OK;
 out:

diff --git a/src/components/tl/ucp/tl_ucp_coll.h b/src/components/tl/ucp/tl_ucp_coll.h
@@ -323,7 +323,6 @@ ucc_tl_ucp_init_task(ucc_base_coll_args_t *coll_args, ucc_base_team_t *team)
     }
 
     task->super.finalize       = ucc_tl_ucp_coll_finalize;
-    task->super.triggered_post = ucc_triggered_post;
     return task;
 }
 

diff --git a/src/schedule/ucc_schedule.c b/src/schedule/ucc_schedule.c
@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2021-2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * Copyright (c) 2021-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * See file LICENSE for terms.
  */
 #include "ucc_schedule.h"
@@ -96,6 +96,7 @@ ucc_status_t ucc_coll_task_init(ucc_coll_task_t *task,
     task->executor             = NULL;
     task->super.status         = UCC_OPERATION_INITIALIZED;
     task->triggered_post_setup = NULL;
+    task->triggered_post       = ucc_triggered_post;
     if (bargs) {
         memcpy(&task->bargs, bargs, sizeof(*bargs));
     }
-Original file line number
+Diff line change
@@ Expand Up @@
         }
         task->super.finalize       = ucc_tl_ucp_coll_finalize;
-        task->super.triggered_post = ucc_triggered_post;
         return task;
     }
@@ Expand Down @@