Lines Matching refs:reduce_data
27 void gpu_regular_warp_reduce(void *reduce_data, ShuffleReductFnTy shflFct) { in gpu_regular_warp_reduce() argument
29 shflFct(reduce_data, /*LaneId - not used= */ 0, in gpu_regular_warp_reduce()
34 void gpu_irregular_warp_reduce(void *reduce_data, ShuffleReductFnTy shflFct, in gpu_irregular_warp_reduce() argument
41 shflFct(reduce_data, /*LaneId = */ tid, /*Offset=*/mask, /*AlgoVersion=*/1); in gpu_irregular_warp_reduce()
48 static uint32_t gpu_irregular_simd_reduce(void *reduce_data, in gpu_irregular_simd_reduce() argument
61 shflFct(reduce_data, /*LaneId =*/logical_lane_id, in gpu_irregular_simd_reduce()
70 void *reduce_data, in nvptx_parallel_reduce_nowait() argument
101 gpu_regular_warp_reduce(reduce_data, shflFct); in nvptx_parallel_reduce_nowait()
103 gpu_irregular_warp_reduce(reduce_data, shflFct, in nvptx_parallel_reduce_nowait()
115 cpyFct(reduce_data, WarpsNeeded); in nvptx_parallel_reduce_nowait()
118 gpu_irregular_warp_reduce(reduce_data, shflFct, WarpsNeeded, in nvptx_parallel_reduce_nowait()
125 gpu_regular_warp_reduce(reduce_data, shflFct); in nvptx_parallel_reduce_nowait()
127 gpu_irregular_warp_reduce(reduce_data, shflFct, in nvptx_parallel_reduce_nowait()
134 return gpu_irregular_simd_reduce(reduce_data, shflFct); in nvptx_parallel_reduce_nowait()
146 cpyFct(reduce_data, WarpsNeeded); in nvptx_parallel_reduce_nowait()
150 gpu_irregular_warp_reduce(reduce_data, shflFct, WarpsNeeded, in nvptx_parallel_reduce_nowait()
178 void *reduce_data, ShuffleReductFnTy shflFct, InterWarpCopyFnTy cpyFct) { in __kmpc_nvptx_parallel_reduce_nowait_v2() argument
180 return nvptx_parallel_reduce_nowait(TId, num_vars, reduce_size, reduce_data, in __kmpc_nvptx_parallel_reduce_nowait_v2()
187 void *reduce_data, ShuffleReductFnTy shflFct, InterWarpCopyFnTy cpyFct, in __kmpc_nvptx_teams_reduce_nowait_v2() argument
222 lgcpyFct(GlobalBuffer, ModBockId, reduce_data); in __kmpc_nvptx_teams_reduce_nowait_v2()
224 lgredFct(GlobalBuffer, ModBockId, reduce_data); in __kmpc_nvptx_teams_reduce_nowait_v2()
272 glcpyFct(GlobalBuffer, ThreadId, reduce_data); in __kmpc_nvptx_teams_reduce_nowait_v2()
274 glredFct(GlobalBuffer, i, reduce_data); in __kmpc_nvptx_teams_reduce_nowait_v2()
278 gpu_regular_warp_reduce(reduce_data, shflFct); in __kmpc_nvptx_teams_reduce_nowait_v2()
288 cpyFct(reduce_data, WarpsNeeded); in __kmpc_nvptx_teams_reduce_nowait_v2()
292 gpu_irregular_warp_reduce(reduce_data, shflFct, WarpsNeeded, in __kmpc_nvptx_teams_reduce_nowait_v2()