Lines Matching defs:reduce_data

27 void gpu_regular_warp_reduce(void *reduce_data, ShuffleReductFnTy shflFct) {
29     shflFct(reduce_data, /*LaneId - not used= */ 0,
34 void gpu_irregular_warp_reduce(void *reduce_data, ShuffleReductFnTy shflFct,
41     shflFct(reduce_data, /*LaneId = */ tid, /*Offset=*/mask, /*AlgoVersion=*/1);
47 static uint32_t gpu_irregular_simd_reduce(void *reduce_data,
60     shflFct(reduce_data, /*LaneId =*/logical_lane_id,
66 static int32_t nvptx_parallel_reduce_nowait(void *reduce_data,
99       gpu_regular_warp_reduce(reduce_data, shflFct);
102           reduce_data, shflFct,
113       cpyFct(reduce_data, WarpsNeeded);
116         gpu_irregular_warp_reduce(reduce_data, shflFct, WarpsNeeded,
124     gpu_regular_warp_reduce(reduce_data, shflFct);
126     gpu_irregular_warp_reduce(reduce_data, shflFct,
133     return gpu_irregular_simd_reduce(reduce_data, shflFct);
145     cpyFct(reduce_data, WarpsNeeded);
149       gpu_irregular_warp_reduce(reduce_data, shflFct, WarpsNeeded,
173                                                void *reduce_data,
176   return nvptx_parallel_reduce_nowait(reduce_data, shflFct, cpyFct);
181     uint64_t reduce_data_size, void *reduce_data, ShuffleReductFnTy shflFct,
217       lgcpyFct(GlobalBuffer, ModBockId, reduce_data);
219       lgredFct(GlobalBuffer, ModBockId, reduce_data);
236   // reduce_data is global or shared so before being reduced within the
238   // local_reduce_data = reduce_data[i]
274     glcpyFct(GlobalBuffer, ThreadId, reduce_data);
276       glredFct(GlobalBuffer, i, reduce_data);
280       gpu_regular_warp_reduce(reduce_data, shflFct);
290         cpyFct(reduce_data, WarpsNeeded);
294           gpu_irregular_warp_reduce(reduce_data, shflFct, WarpsNeeded,