/dports/science/qmcpack/qmcpack-3.11.0/src/QMCDrivers/ |
H A D | CloneManager.cpp | 80 wClones.resize(NumThreads); in makeClones() 81 psiClones.resize(NumThreads); in makeClones() 82 hClones.resize(NumThreads); in makeClones() 86 if (NumThreads == 1) in makeClones() 131 wClones.resize(NumThreads); in makeClones() 132 PsiPoolClones.resize(NumThreads); in makeClones() 133 HPoolClones.resize(NumThreads); in makeClones() 139 if (NumThreads == 1) in makeClones() 183 if (NumThreads == 1) in makeClones() 209 wgClones.resize(NumThreads, 0); in makeClones() [all …]
|
/dports/math/faiss/faiss-1.7.1/faiss/gpu/utils/ |
H A D | MergeNetworkBlock.cuh | 23 int NumThreads, 115 int NumThreads, 131 constexpr int kLoopPerThread = L / NumThreads; in blockMergeLarge() 137 int tid = loop * NumThreads + threadIdx.x; in blockMergeLarge() 208 int NumThreads, 221 int NumThreads, 243 NumThreads, in merge() 259 NumThreads, in merge() 275 int NumThreads, 297 int NumThreads, [all …]
|
/dports/devel/llvm-devel/llvm-project-f05c95f10fc1d8171071735af8ad3a9e87633120/openmp/libomptarget/DeviceRTL/src/ |
H A D | Parallelism.cpp | 52 uint32_t NumThreads = mapping::getBlockSize(); in determineNumberOfThreads() local 54 if (NThreadsICV != 0 && NThreadsICV < NumThreads) in determineNumberOfThreads() 55 NumThreads = NThreadsICV; in determineNumberOfThreads() 58 if (NumThreads < mapping::getWarpSize()) in determineNumberOfThreads() 59 NumThreads = 1; in determineNumberOfThreads() 61 NumThreads = (NumThreads & ~((uint32_t)mapping::getWarpSize() - 1)); in determineNumberOfThreads() 63 return NumThreads; in determineNumberOfThreads() 94 uint32_t NumThreads = determineNumberOfThreads(num_threads); in __kmpc_parallel_51() local 104 if (TId < NumThreads) in __kmpc_parallel_51() 117 bool IsActiveParallelRegion = NumThreads > 1; in __kmpc_parallel_51() [all …]
|
H A D | Reduction.cpp | 77 uint32_t NumThreads = omp_get_num_threads(); in nvptx_parallel_reduce_nowait() local 78 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 93 (NumThreads + mapping::getWarpSize() - 1) / mapping::getWarpSize(); in nvptx_parallel_reduce_nowait() 112 if (NumThreads > mapping::getWarpSize()) { in nvptx_parallel_reduce_nowait() 141 if (NumThreads > mapping::getWarpSize()) { in nvptx_parallel_reduce_nowait() 201 uint32_t NumThreads = omp_get_num_threads(); in __kmpc_nvptx_teams_reduce_nowait_v2() local 266 NumThreads = roundToWarpsize(kmpcMin(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 267 if (ThreadId >= NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 272 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 276 if (NumThreads > 1) { in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
/dports/devel/wasi-compiler-rt13/llvm-project-13.0.1.src/openmp/libomptarget/DeviceRTL/src/ |
H A D | Parallelism.cpp | 52 uint32_t NumThreads = mapping::getBlockSize(); in determineNumberOfThreads() local 54 if (NThreadsICV != 0 && NThreadsICV < NumThreads) in determineNumberOfThreads() 55 NumThreads = NThreadsICV; in determineNumberOfThreads() 58 if (NumThreads < mapping::getWarpSize()) in determineNumberOfThreads() 59 NumThreads = 1; in determineNumberOfThreads() 61 NumThreads = (NumThreads & ~((uint32_t)mapping::getWarpSize() - 1)); in determineNumberOfThreads() 63 return NumThreads; in determineNumberOfThreads() 94 uint32_t NumThreads = determineNumberOfThreads(num_threads); in __kmpc_parallel_51() local 103 if (TId < NumThreads) in __kmpc_parallel_51() 116 bool IsActiveParallelRegion = NumThreads > 1; in __kmpc_parallel_51() [all …]
|
H A D | Reduction.cpp | 77 uint32_t NumThreads = omp_get_num_threads(); in nvptx_parallel_reduce_nowait() local 78 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 93 (NumThreads + mapping::getWarpSize() - 1) / mapping::getWarpSize(); in nvptx_parallel_reduce_nowait() 112 if (NumThreads > mapping::getWarpSize()) { in nvptx_parallel_reduce_nowait() 141 if (NumThreads > mapping::getWarpSize()) { in nvptx_parallel_reduce_nowait() 201 uint32_t NumThreads = omp_get_num_threads(); in __kmpc_nvptx_teams_reduce_nowait_v2() local 266 NumThreads = roundToWarpsize(kmpcMin(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 267 if (ThreadId >= NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 272 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 276 if (NumThreads > 1) { in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
/dports/devel/wasi-libcxx/llvm-project-13.0.1.src/openmp/libomptarget/DeviceRTL/src/ |
H A D | Parallelism.cpp | 52 uint32_t NumThreads = mapping::getBlockSize(); in determineNumberOfThreads() local 54 if (NThreadsICV != 0 && NThreadsICV < NumThreads) in determineNumberOfThreads() 55 NumThreads = NThreadsICV; in determineNumberOfThreads() 58 if (NumThreads < mapping::getWarpSize()) in determineNumberOfThreads() 59 NumThreads = 1; in determineNumberOfThreads() 61 NumThreads = (NumThreads & ~((uint32_t)mapping::getWarpSize() - 1)); in determineNumberOfThreads() 63 return NumThreads; in determineNumberOfThreads() 94 uint32_t NumThreads = determineNumberOfThreads(num_threads); in __kmpc_parallel_51() local 103 if (TId < NumThreads) in __kmpc_parallel_51() 116 bool IsActiveParallelRegion = NumThreads > 1; in __kmpc_parallel_51() [all …]
|
H A D | Reduction.cpp | 77 uint32_t NumThreads = omp_get_num_threads(); in nvptx_parallel_reduce_nowait() local 78 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 93 (NumThreads + mapping::getWarpSize() - 1) / mapping::getWarpSize(); in nvptx_parallel_reduce_nowait() 112 if (NumThreads > mapping::getWarpSize()) { in nvptx_parallel_reduce_nowait() 141 if (NumThreads > mapping::getWarpSize()) { in nvptx_parallel_reduce_nowait() 201 uint32_t NumThreads = omp_get_num_threads(); in __kmpc_nvptx_teams_reduce_nowait_v2() local 266 NumThreads = roundToWarpsize(kmpcMin(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 267 if (ThreadId >= NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 272 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 276 if (NumThreads > 1) { in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
/dports/devel/llvm13/llvm-project-13.0.1.src/openmp/libomptarget/DeviceRTL/src/ |
H A D | Parallelism.cpp | 52 uint32_t NumThreads = mapping::getBlockSize(); 54 if (NThreadsICV != 0 && NThreadsICV < NumThreads) 55 NumThreads = NThreadsICV; 58 if (NumThreads < mapping::getWarpSize()) 59 NumThreads = 1; 61 NumThreads = (NumThreads & ~((uint32_t)mapping::getWarpSize() - 1)); 63 return NumThreads; 94 uint32_t NumThreads = determineNumberOfThreads(num_threads); 103 if (TId < NumThreads) 116 bool IsActiveParallelRegion = NumThreads > 1; [all …]
|
H A D | Reduction.cpp | 77 uint32_t NumThreads = omp_get_num_threads(); 78 if (NumThreads == 1) 93 (NumThreads + mapping::getWarpSize() - 1) / mapping::getWarpSize(); 112 if (NumThreads > mapping::getWarpSize()) { in DIRDEF() 141 if (NumThreads > mapping::getWarpSize()) { in GetAmType() 201 uint32_t NumThreads = omp_get_num_threads(); in GetProgMax() 266 NumThreads = roundToWarpsize(kmpcMin(NumThreads, NumRecs)); 267 if (ThreadId >= NumThreads) 272 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) 276 if (NumThreads > 1) { [all …]
|
/dports/sysutils/istio/istio-1.6.7/vendor/fortio.org/fortio/periodic/ |
H A D | periodic.go | 46 NumThreads: 4, 114 NumThreads int member 140 NumThreads int member 193 if r.NumThreads == 0 { 194 r.NumThreads = DefaultRunnerOptions.NumThreads 196 if r.NumThreads < 1 { 197 r.NumThreads = 1 317 r.NumThreads = 1 322 r.NumThreads = newN 380 if r.NumThreads > runnersLen { [all …]
|
/dports/devel/llvm-devel/llvm-project-f05c95f10fc1d8171071735af8ad3a9e87633120/openmp/libomptarget/deviceRTLs/common/src/ |
H A D | reduction.cu | 74 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 88 uint32_t WarpsNeeded = (NumThreads + WARPSIZE - 1) / WARPSIZE; in nvptx_parallel_reduce_nowait() 100 /*LaneCount=*/NumThreads % WARPSIZE, in nvptx_parallel_reduce_nowait() 107 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 135 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 198 uint32_t NumThreads = in __kmpc_nvptx_teams_reduce_nowait_v2() local 264 NumThreads = roundToWarpsize(kmpcMin(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 265 if (ThreadId >= NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 270 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 274 if (NumThreads > 1) { in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
/dports/devel/wasi-compiler-rt13/llvm-project-13.0.1.src/openmp/libomptarget/deviceRTLs/common/src/ |
H A D | reduction.cu | 74 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 88 uint32_t WarpsNeeded = (NumThreads + WARPSIZE - 1) / WARPSIZE; in nvptx_parallel_reduce_nowait() 100 /*LaneCount=*/NumThreads % WARPSIZE, in nvptx_parallel_reduce_nowait() 107 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 135 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 198 uint32_t NumThreads = in __kmpc_nvptx_teams_reduce_nowait_v2() local 264 NumThreads = roundToWarpsize(kmpcMin(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 265 if (ThreadId >= NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 270 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 274 if (NumThreads > 1) { in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
/dports/devel/wasi-libcxx/llvm-project-13.0.1.src/openmp/libomptarget/deviceRTLs/common/src/ |
H A D | reduction.cu | 74 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 88 uint32_t WarpsNeeded = (NumThreads + WARPSIZE - 1) / WARPSIZE; in nvptx_parallel_reduce_nowait() 100 /*LaneCount=*/NumThreads % WARPSIZE, in nvptx_parallel_reduce_nowait() 107 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 135 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 198 uint32_t NumThreads = in __kmpc_nvptx_teams_reduce_nowait_v2() local 264 NumThreads = roundToWarpsize(kmpcMin(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 265 if (ThreadId >= NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 270 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 274 if (NumThreads > 1) { in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
/dports/devel/llvm13/llvm-project-13.0.1.src/openmp/libomptarget/deviceRTLs/common/src/ |
H A D | reduction.cu | 74 if (NumThreads == 1) 88 uint32_t WarpsNeeded = (NumThreads + WARPSIZE - 1) / WARPSIZE; 100 /*LaneCount=*/NumThreads % WARPSIZE, 107 if (NumThreads > WARPSIZE) { 135 if (NumThreads > WARPSIZE) { 198 uint32_t NumThreads = 264 NumThreads = roundToWarpsize(kmpcMin(NumThreads, NumRecs)); 265 if (ThreadId >= NumThreads) 270 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) 274 if (NumThreads > 1) { [all …]
|
/dports/devel/llvm-cheri/llvm-project-37c49ff00e3eadce5d8703fdc4497f28458c64a8/openmp/libomptarget/deviceRTLs/common/src/ |
H A D | reduction.cu | 85 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 99 uint32_t WarpsNeeded = (NumThreads + WARPSIZE - 1) / WARPSIZE; in nvptx_parallel_reduce_nowait() 106 if ((NumThreads % WARPSIZE == 0) || (WarpId < WarpsNeeded - 1)) in nvptx_parallel_reduce_nowait() 110 /*LaneCount=*/NumThreads % WARPSIZE, in nvptx_parallel_reduce_nowait() 117 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 144 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 204 uint32_t NumThreads = in __kmpc_nvptx_teams_reduce_nowait_v2() local 270 NumThreads = roundToWarpsize(__kmpc_impl_min(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 271 if (ThreadId >= NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 276 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
/dports/devel/llvm11/llvm-11.0.1.src/projects/openmp/libomptarget/deviceRTLs/common/src/ |
H A D | reduction.cu | 85 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 99 uint32_t WarpsNeeded = (NumThreads + WARPSIZE - 1) / WARPSIZE; in nvptx_parallel_reduce_nowait() 106 if ((NumThreads % WARPSIZE == 0) || (WarpId < WarpsNeeded - 1)) in nvptx_parallel_reduce_nowait() 110 /*LaneCount=*/NumThreads % WARPSIZE, in nvptx_parallel_reduce_nowait() 117 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 144 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 204 uint32_t NumThreads = in __kmpc_nvptx_teams_reduce_nowait_v2() local 270 NumThreads = roundToWarpsize(__kmpc_impl_min(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 271 if (ThreadId >= NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 276 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
/dports/www/chromium-legacy/chromium-88.0.4324.182/third_party/llvm/openmp/libomptarget/deviceRTLs/common/src/ |
H A D | reduction.cu | 85 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 99 uint32_t WarpsNeeded = (NumThreads + WARPSIZE - 1) / WARPSIZE; in nvptx_parallel_reduce_nowait() 106 if ((NumThreads % WARPSIZE == 0) || (WarpId < WarpsNeeded - 1)) in nvptx_parallel_reduce_nowait() 110 /*LaneCount=*/NumThreads % WARPSIZE, in nvptx_parallel_reduce_nowait() 117 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 144 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 204 uint32_t NumThreads = in __kmpc_nvptx_teams_reduce_nowait_v2() local 270 NumThreads = roundToWarpsize(__kmpc_impl_min(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 271 if (ThreadId >= NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 276 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
/dports/devel/llvm10/llvm-10.0.1.src/projects/openmp/libomptarget/deviceRTLs/common/src/ |
H A D | reduction.cu | 99 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 131 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 158 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 228 uint32_t NumThreads = in nvptx_teams_reduce_nowait() local 270 if (NumTeams < NumThreads) { in nvptx_teams_reduce_nowait() 313 for (uint32_t i = NumThreads + ThreadId; i < NumTeams; i += NumThreads) in nvptx_teams_reduce_nowait() 327 uint32_t ActiveThreads = NumTeams < NumThreads ? NumTeams : NumThreads; in nvptx_teams_reduce_nowait() 421 uint32_t NumThreads = in __kmpc_nvptx_teams_reduce_nowait_v2() local 487 NumThreads = roundToWarpsize(__kmpc_impl_min(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 493 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
H A D | parallel.cu | 150 uint16_t NumThreads; in __kmpc_kernel_convergent_parallel() local 154 NumThreads = ConvergentSize; in __kmpc_kernel_convergent_parallel() 156 (int)NumThreads); in __kmpc_kernel_convergent_parallel() 161 if (OmpId < NumThreads) { in __kmpc_kernel_convergent_parallel() 219 if (NumThreads < WARPSIZE) { in determineNumberOfThreads() 220 NumThreads = 1; in determineNumberOfThreads() 222 NumThreads = (NumThreads & ~((uint16_t)WARPSIZE - 1)); in determineNumberOfThreads() 226 return NumThreads; in determineNumberOfThreads() 254 uint16_t NumThreads = in __kmpc_kernel_prepare_parallel() local 263 (int)NumThreads); in __kmpc_kernel_prepare_parallel() [all …]
|
/dports/devel/tinygo/tinygo-0.14.1/llvm-project/openmp/libomptarget/deviceRTLs/common/src/ |
H A D | reduction.cu | 99 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 131 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 158 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 228 uint32_t NumThreads = in nvptx_teams_reduce_nowait() local 270 if (NumTeams < NumThreads) { in nvptx_teams_reduce_nowait() 313 for (uint32_t i = NumThreads + ThreadId; i < NumTeams; i += NumThreads) in nvptx_teams_reduce_nowait() 327 uint32_t ActiveThreads = NumTeams < NumThreads ? NumTeams : NumThreads; in nvptx_teams_reduce_nowait() 421 uint32_t NumThreads = in __kmpc_nvptx_teams_reduce_nowait_v2() local 487 NumThreads = roundToWarpsize(__kmpc_impl_min(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 493 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
H A D | parallel.cu | 150 uint16_t NumThreads; in __kmpc_kernel_convergent_parallel() local 154 NumThreads = ConvergentSize; in __kmpc_kernel_convergent_parallel() 156 (int)NumThreads); in __kmpc_kernel_convergent_parallel() 161 if (OmpId < NumThreads) { in __kmpc_kernel_convergent_parallel() 219 if (NumThreads < WARPSIZE) { in determineNumberOfThreads() 220 NumThreads = 1; in determineNumberOfThreads() 222 NumThreads = (NumThreads & ~((uint16_t)WARPSIZE - 1)); in determineNumberOfThreads() 226 return NumThreads; in determineNumberOfThreads() 254 uint16_t NumThreads = in __kmpc_kernel_prepare_parallel() local 263 (int)NumThreads); in __kmpc_kernel_prepare_parallel() [all …]
|
/dports/devel/llvm90/llvm-9.0.1.src/projects/openmp/libomptarget/deviceRTLs/nvptx/src/ |
H A D | reduction.cu | 104 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 136 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 163 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 233 uint32_t NumThreads = in nvptx_teams_reduce_nowait() local 275 if (NumTeams < NumThreads) { in nvptx_teams_reduce_nowait() 318 for (uint32_t i = NumThreads + ThreadId; i < NumTeams; i += NumThreads) in nvptx_teams_reduce_nowait() 332 uint32_t ActiveThreads = NumTeams < NumThreads ? NumTeams : NumThreads; in nvptx_teams_reduce_nowait() 426 uint32_t NumThreads = in __kmpc_nvptx_teams_reduce_nowait_v2() local 492 NumThreads = roundToWarpsize(min(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 498 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
/dports/devel/llvm12/llvm-project-12.0.1.src/openmp/libomptarget/deviceRTLs/common/src/ |
H A D | reduction.cu | 85 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 99 uint32_t WarpsNeeded = (NumThreads + WARPSIZE - 1) / WARPSIZE; in nvptx_parallel_reduce_nowait() 106 if ((NumThreads % WARPSIZE == 0) || (WarpId < WarpsNeeded - 1)) in nvptx_parallel_reduce_nowait() 117 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 144 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 206 uint32_t NumThreads = in __kmpc_nvptx_teams_reduce_nowait_v2() local 272 NumThreads = roundToWarpsize(kmpcMin(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 273 if (ThreadId >= NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 278 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 282 if (NumThreads > 1) { in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|
/dports/devel/wasi-compiler-rt12/llvm-project-12.0.1.src/openmp/libomptarget/deviceRTLs/common/src/ |
H A D | reduction.cu | 85 if (NumThreads == 1) in nvptx_parallel_reduce_nowait() 99 uint32_t WarpsNeeded = (NumThreads + WARPSIZE - 1) / WARPSIZE; in nvptx_parallel_reduce_nowait() 106 if ((NumThreads % WARPSIZE == 0) || (WarpId < WarpsNeeded - 1)) in nvptx_parallel_reduce_nowait() 117 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 144 if (NumThreads > WARPSIZE) { in nvptx_parallel_reduce_nowait() 206 uint32_t NumThreads = in __kmpc_nvptx_teams_reduce_nowait_v2() local 272 NumThreads = roundToWarpsize(kmpcMin(NumThreads, NumRecs)); in __kmpc_nvptx_teams_reduce_nowait_v2() 273 if (ThreadId >= NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 278 for (uint32_t i = NumThreads + ThreadId; i < NumRecs; i += NumThreads) in __kmpc_nvptx_teams_reduce_nowait_v2() 282 if (NumThreads > 1) { in __kmpc_nvptx_teams_reduce_nowait_v2() [all …]
|