Searched refs:threadsPerGroup (Results 1 – 4 of 4) sorted by relevance
30 MTLSize threadsPerGroup = [device maxThreadsPerThreadgroup]; in CreateGpuInfoFromMetalDevice() local31 gpu_info.metal_info.max_work_group_size_x = threadsPerGroup.width; in CreateGpuInfoFromMetalDevice()32 gpu_info.metal_info.max_work_group_size_y = threadsPerGroup.height; in CreateGpuInfoFromMetalDevice()33 gpu_info.metal_info.max_work_group_size_z = threadsPerGroup.depth; in CreateGpuInfoFromMetalDevice()
1422 void getLaunchVals(int &threadsPerGroup, int &num_groups, int ConstWGSize, in getLaunchVals() argument1443 fprintf(stderr, "threadsPerGroup: %d\n", threadsPerGroup); in getLaunchVals()1448 threadsPerGroup = thread_limit; in getLaunchVals()1451 threadsPerGroup += RTLDeviceInfoTy::Warp_Size; in getLaunchVals()1454 if (threadsPerGroup > RTLDeviceInfoTy::Max_WG_Size) { // limit to max in getLaunchVals()1455 threadsPerGroup = RTLDeviceInfoTy::Max_WG_Size; in getLaunchVals()1456 DP("Setting threads per block to maximum %d\n", threadsPerGroup); in getLaunchVals()1460 if (threadsPerGroup > ConstWGSize) { in getLaunchVals()1461 threadsPerGroup = ConstWGSize; in getLaunchVals()1463 threadsPerGroup); in getLaunchVals()[all …]
619 const MTLSize &threadsPerGroup);622 const MTLSize &threadsPerGroup);
2255 const MTLSize &threadsPerGroup)2257 [get() dispatchThreadgroups:threadGroupsPerGrid threadsPerThreadgroup:threadsPerGroup];2262 const MTLSize &threadsPerGroup)2267 [get() dispatchThreads:threadsPerGrid threadsPerThreadgroup:threadsPerGroup];