Searched refs:WarpSize (Results 1 – 5 of 5) sorted by relevance
| /llvm-project-15.0.7/openmp/libomptarget/plugins/cuda/src/ |
| H A D | rtl.cpp | 164 unsigned int WarpSize = 0; member
692 int WarpSize; in initDevice() local
694 cuDeviceGetAttribute(&WarpSize, CU_DEVICE_ATTRIBUTE_WARP_SIZE, Device); in initDevice()
697 DeviceData[DeviceId].WarpSize = 32; in initDevice()
699 DP("Using warp size %d\n", WarpSize); in initDevice()
700 DeviceData[DeviceId].WarpSize = WarpSize; in initDevice()
733 DeviceData[DeviceId].ThreadsPerBlock, DeviceData[DeviceId].WarpSize); in initDevice()
1141 DP("Adding master warp: +%d threads\n", DeviceData[DeviceId].WarpSize); in runTargetTeamRegion()
1142 CudaThreadsPerBlock += DeviceData[DeviceId].WarpSize; in runTargetTeamRegion()
|
| /llvm-project-15.0.7/openmp/libomptarget/plugins/amdgpu/src/ |
| H A D | rtl.cpp | 470 std::vector<int> WarpSize; member in RTLDeviceInfoTy
1023 WarpSize.resize(NumberOfDevices); in RTLDeviceInfoTy()
1206 LaunchVals getLaunchVals(int WarpSize, EnvironmentVariables Env, in getLaunchVals() argument
1222 DP("RTLDeviceInfoTy::Warp_Size: %d\n", WarpSize); in getLaunchVals()
1237 ThreadsPerGroup += WarpSize; in getLaunchVals()
1238 DP("Adding master wavefront: +%d threads\n", WarpSize); in getLaunchVals()
1405 getLaunchVals(DeviceInfo().WarpSize[DeviceId], DeviceInfo().Env, in runRegionLocked()
2157 DeviceInfo().WarpSize[DeviceId] = WavefrontSize; in __tgt_rtl_init_device()
2161 DeviceInfo().WarpSize[DeviceId] = 64; in __tgt_rtl_init_device()
2220 DeviceInfo().WarpSize[DeviceId], DeviceInfo().ThreadsPerGroup[DeviceId], in __tgt_rtl_init_device()
|
| /llvm-project-15.0.7/openmp/libomptarget/DeviceRTL/src/ |
| H A D | Synchronization.cpp | 117 uint32_t WarpSize = mapping::getWarpSize(); in namedBarrier() local
118 uint32_t NumWaves = NumThreads / WarpSize; in namedBarrier()
|
| /llvm-project-15.0.7/clang/lib/CodeGen/ |
| H A D | CGOpenMPRuntimeGPU.cpp | 340 unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; in buildRecordForGlobalizedVars() local
347 MappedDeclsFields, WarpSize); in buildRecordForGlobalizedVars()
1325 unsigned WarpSize = CGM.getTarget().getGridValue().GV_Warp_Size; in emitTeamsOutlinedFunction() local
1334 MappedDeclsFields, WarpSize); in emitTeamsOutlinedFunction()
1755 llvm::Value *WarpSize = in createRuntimeShuffleFunction() local
1760 {ElemCast, Offset, WarpSize}); in createRuntimeShuffleFunction()
2135 unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; in emitInterWarpCopyFunction() local
2137 auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize); in emitInterWarpCopyFunction()
|
| /llvm-project-15.0.7/llvm/lib/Transforms/IPO/ |
| H A D | OpenMPOpt.cpp | 2131 ExternalizationRAII WarpSize(OMPInfoCache, OMPRTL___kmpc_get_warp_size); in runAttributor() local
3830 CallInst *WarpSize = in buildCustomStateMachine() local
3832 OMPInfoCache.setCallingConvention(WarpSizeFn, WarpSize); in buildCustomStateMachine()
3833 WarpSize->setDebugLoc(DLoc); in buildCustomStateMachine()
3835 BlockHwSize, WarpSize, "block.size", IsWorkerCheckBB); in buildCustomStateMachine()
|