TL/CUDA: fix build

openucx · Mar 22, 2024 · 7c22cb5 · 7c22cb5
1 parent 189120d
commit 7c22cb5
Show file tree

Hide file tree

Showing 3 changed files with 18 additions and 15 deletions.
diff --git a/src/components/tl/cuda/Makefile.am b/src/components/tl/cuda/Makefile.am
@@ -59,6 +59,7 @@ sources = \
  $(allgatherv) \
  $(alltoall) \
  $(alltoallv) \
+ $(bcast) \
  $(reduce_scatter) \
  $(reduce_scatterv)
 

diff --git a/src/components/tl/cuda/bcast/bcast_linear.c b/src/components/tl/cuda/bcast/bcast_linear.c
@@ -29,7 +29,8 @@ void ucc_tl_cuda_bcast_linear_progress(ucc_coll_task_t *coll_task)
  ucc_tl_cuda_task_t *task = ucc_derived_of(coll_task, ucc_tl_cuda_task_t);
  ucc_tl_cuda_team_t *team = TASK_TEAM(task);
  ucc_status_t st;
-
+ (void) team;
+ (void) st;
  task->super.status = UCC_INPROGRESS;
 }
 
@@ -40,11 +41,12 @@ ucc_status_t ucc_tl_cuda_bcast_linear_start(ucc_coll_task_t *coll_task)
  ucc_coll_args_t * args = &TASK_ARGS(task);
  ucc_rank_t tsize = UCC_TL_TEAM_SIZE(team);
  ucc_datatype_t dt = task->allgatherv_linear.dt;
- ucc_rank_t i;
- size_t send_size, frag_size, ssize;
 
+ (void) tsize;
+ (void) args;
+ (void) dt;
  task->bcast_linear.stage = STAGE_SYNC;
- task->allgatherv_linear.sbuf = args->src.info.buffer;
+ // task->bcast_linear.sbuf = args->src.info.buffer;
 
 
  return ucc_progress_queue_enqueue(UCC_TL_CORE_CTX(team)->pq, &task->super);
@@ -75,9 +77,9 @@ ucc_status_t ucc_tl_cuda_bcast_linear_init(ucc_base_coll_args_t *coll_args,
  // task->allgatherv_linear.rbuf = coll_args->args.dst.info.buffer;
 
  task->super.flags |= UCC_COLL_TASK_FLAG_EXECUTOR;
- task->super.post = ucc_tl_cuda_allgatherv_linear_start;
- task->super.progress = ucc_tl_cuda_allgatherv_linear_progress;
- task->super.finalize = ucc_tl_cuda_allgatherv_linear_finalize;
+ task->super.post = ucc_tl_cuda_bcast_linear_start;
+ task->super.progress = ucc_tl_cuda_bcast_linear_progress;
+ task->super.finalize = ucc_tl_cuda_bcast_linear_finalize;
  task->bar = TASK_BAR(task);
 
  *task_p = &task->super;

diff --git a/src/components/tl/cuda/tl_cuda_context.c b/src/components/tl/cuda/tl_cuda_context.c
@@ -20,8 +20,8 @@ UCC_CLASS_INIT_FUNC(ucc_tl_cuda_context_t,
  ucc_status_t status;
  int num_devices;
  cudaError_t cuda_st;
- CUcontext cu_ctx;
- CUresult cu_st;
+ // CUcontext cu_ctx;
+ // CUresult cu_st;
 
  UCC_CLASS_CALL_SUPER_INIT(ucc_tl_context_t, &tl_cuda_config->super,
  params->context);
@@ -37,12 +37,12 @@ UCC_CLASS_INIT_FUNC(ucc_tl_cuda_context_t,
  return UCC_ERR_NO_RESOURCE;
  }
 
- cu_st = cuCtxGetCurrent(&cu_ctx);
- if (cu_ctx == NULL || cu_st != CUDA_SUCCESS) {
- tl_debug(self->super.super.lib,
- "cannot create CUDA TL context without active CUDA context");
- return UCC_ERR_NO_RESOURCE;
- }
+ // cu_st = cuCtxGetCurrent(&cu_ctx);
+ // if (cu_ctx == NULL || cu_st != CUDA_SUCCESS) {
+ //  tl_debug(self->super.super.lib,
+ //  "cannot create CUDA TL context without active CUDA context");
+ //  return UCC_ERR_NO_RESOURCE;
+ // }
 
  status = ucc_mpool_init(&self->req_mp, 0, sizeof(ucc_tl_cuda_task_t), 0,
  UCC_CACHE_LINE_SIZE, 8, UINT_MAX,