Home
last modified time | relevance | path

Searched defs:kWarpSize (Results 1 – 7 of 7) sorted by relevance

/external/pytorch/aten/src/ATen/native/cuda/
Dint4mm.cu132 constexpr int32_t kWarpSize = 64; variable
151 constexpr int32_t kWarpSize = 32; variable
607 __launch_bounds__(Warps* kWarpSize) void tinygemm_m16n8k16_chunk_kernel( in __launch_bounds__()
Dlayer_norm_kernel.cu36 constexpr unsigned int kWarpSize = C10_WARP_SIZE; variable
/external/tensorflow/tensorflow/core/kernels/
Dbias_op_gpu.cu.cc225 static constexpr int32 kWarpSize = 32; in compute() local
/external/pytorch/torch/csrc/distributed/c10d/
Dintra_node_comm.cu13 static constexpr size_t kWarpSize = 32; variable
/external/pytorch/aten/src/ATen/native/nested/cuda/
DNestedTensorTransformerFunctions.cu538 static constexpr int32_t kWarpSize = 64; variable
540 static constexpr int32_t kWarpSize = 32; variable
/external/pytorch/aten/src/ATen/native/transformers/cuda/mem_eff_attention/
Dkernel_forward.h125 static constexpr int kWarpSize = 32; member
Dkernel_backward.h245 static constexpr int64_t kWarpSize = 32; member