From 01175463f5ba2216fd8cf1ee8bc891b3b3607782 Mon Sep 17 00:00:00 2001 From: Aurelien Bouteiller Date: Fri, 14 Mar 2025 14:11:31 -0400 Subject: [PATCH] Update for API 4.1 flow-mask icldisco/parsec#722 --- src/cuda/lapack_cuda_stage_in.c | 8 ++++---- src/dplasmajdf_lapack_dtt.h | 8 ++++---- 2 files changed, 8 insertions(+), 8 deletions(-) diff --git a/src/cuda/lapack_cuda_stage_in.c b/src/cuda/lapack_cuda_stage_in.c index f2f48ced..cb3b192f 100644 --- a/src/cuda/lapack_cuda_stage_in.c +++ b/src/cuda/lapack_cuda_stage_in.c @@ -17,7 +17,7 @@ int dplasma_cuda_lapack_stage_in(parsec_gpu_task_t *gtask, - uint32_t flow_mask, + parsec_flow_mask_t flow_mask, parsec_gpu_exec_stream_t *gpu_stream) { cudaError_t ret; @@ -30,7 +30,7 @@ dplasma_cuda_lapack_stage_in(parsec_gpu_task_t *gtask, int elem_sz; int i; for(i = 0; i < task->task_class->nb_flows; i++){ - if(flow_mask & (1U << i)){ + if(PARSEC_CHECK_FLOW_MASK(flow_mask, i)) { copy_in = task->data[i].data_in; copy_out = task->data[i].data_out; ddc = (dplasma_data_collection_t*)gtask->flow_dc[i]; @@ -93,7 +93,7 @@ dplasma_cuda_lapack_stage_in(parsec_gpu_task_t *gtask, int dplasma_cuda_lapack_stage_out(parsec_gpu_task_t *gtask, - uint32_t flow_mask, + parsec_flow_mask_t flow_mask, parsec_gpu_exec_stream_t *gpu_stream) { cudaError_t ret; @@ -106,7 +106,7 @@ dplasma_cuda_lapack_stage_out(parsec_gpu_task_t *gtask, int elem_sz; int i; for(i = 0; i < task->task_class->nb_flows; i++){ - if(flow_mask & (1U << i)){ + if(PARSEC_CHECK_FLOW_MASK(flow_mask, i)) { copy_in = task->data[i].data_out; copy_out = copy_in->original->device_copies[0]; ddc = (dplasma_data_collection_t*)gtask->flow_dc[i]; diff --git a/src/dplasmajdf_lapack_dtt.h b/src/dplasmajdf_lapack_dtt.h index 5fa568c8..9edf42ca 100644 --- a/src/dplasmajdf_lapack_dtt.h +++ b/src/dplasmajdf_lapack_dtt.h @@ -137,23 +137,23 @@ void ADTT_INFO_internal(parsec_data_copy_t *cp, const dplasma_data_collection_t #if defined(DPLASMA_HAVE_CUDA) int dplasma_cuda_lapack_stage_in(parsec_gpu_task_t *gtask, - uint32_t flow_mask, + parsec_flow_mask_t flow_mask, parsec_gpu_exec_stream_t *gpu_stream); int dplasma_cuda_lapack_stage_out(parsec_gpu_task_t *gtask, - uint32_t flow_mask, + parsec_flow_mask_t flow_mask, parsec_gpu_exec_stream_t *gpu_stream); #endif /* defined(DPLASMA_HAVE_CUDA) */ #if defined(DPLASMA_HAVE_HIP) int dplasma_hip_lapack_stage_in(parsec_gpu_task_t *gtask, - uint32_t flow_mask, + parsec_flow_mask_t flow_mask, parsec_gpu_exec_stream_t *gpu_stream); int dplasma_hip_lapack_stage_out(parsec_gpu_task_t *gtask, - uint32_t flow_mask, + parsec_flow_mask_t flow_mask, parsec_gpu_exec_stream_t *gpu_stream); #endif /* defined(DPLASMA_HAVE_HIP) */