/usr/local/lib64/python3.6/site-packages/torch/include/ATen/native/cuda
NameSizeModeActions
BatchLinearAlgebraLib.h31140644editdlrm
block_reduce.cuh25490644editdlrm
CompositeRandomAccessor.h9290644editdlrm
CUDALoops.cuh75980644editdlrm
CuFFTPlanCache.h192820644editdlrm
CuFFTUtils.h18920644editdlrm
DeviceSqrt.cuh5850644editdlrm
DistributionTemplates.h274350644editdlrm
EmbeddingBackwardKernel.cuh7150644editdlrm
ForeachFunctors.cuh168510644editdlrm
GridSampler.cuh113160644editdlrm
im2col.cuh65770644editdlrm
KernelUtils.cuh25530644editdlrm
LaunchUtils.h3060644editdlrm
Loops.cuh99970644editdlrm
Math.cuh138400644editdlrm
MemoryAccess.cuh124630644editdlrm
MiscUtils.h33410644editdlrm
MultiTensorApply.cuh75520644editdlrm
Normalization.cuh744410644editdlrm
PersistentSoftmax.cuh146350644editdlrm
Randperm.cuh21140644editdlrm
Reduce.cuh387840644editdlrm
Resize.cuh19190644editdlrm
ROCmLoops.cuh135260644editdlrm
SortingCommon.cuh56880644editdlrm
SortingRadixSelect.cuh119180644editdlrm
SortUtils.cuh55490644editdlrm
TensorModeKernel.cuh143910644editdlrm
UniqueCub.cuh3450644editdlrm
UpSample.cuh75520644editdlrm
vol2col.cuh82970644editdlrm
Edit: /usr/local/lib64/python3.6/site-packages/torch/include/ATen/native/cuda/CUDALoops.cuh (7598B)
#pragma once // This file provides two functions to help write GPU elementwise kernels: // // gpu_kernel(TensorIterator iter, ) // gpu_kernel_with_scalars(TensorIterator iter, ) // // The gpu_kernel_with_scalars generates specializations that support a // single scalar CPU argument, such as from `cuda_tensor + 5`. The CPU scalar // is lifted to a kernel parameter instead of copying to device memory. // This should be used in conjunction with TensorIterator::allow_cpu_scalars_, // which is the default for TensorIterator::binary_op. Otherwise, all inputs // and the output must be on the GPU. // // For example, to write a reciprocal kernel for GPU float Tensors: // // gpu_kernel(iter, []GPU_LAMBDA(float a) { // return 1.0f / a; // }); // // To write a multiplication kernel for GPU float Tensors where one argument // may be a CPU scalar: // // gpu_kernel_with_scalars(iter, []GPU_LAMBDA(float a, float b) { // return a * b; // }); // // See BinaryOpsKernel.cu for the complete implementation // #include #include #include #include #include #include #include #include #include #include #include // Marks a lambda as executable on both the host and device. The __host__ // attribute is important so that we can access static type information from // the host, even if the function is typically only executed on the device. #ifndef GPU_LAMBDA #define GPU_LAMBDA __host__ __device__ #endif #ifdef __NVCC__ #define ASSERT_HOST_DEVICE_LAMBDA(type) \ static_assert(__nv_is_extended_host_device_lambda_closure_type(type), \ #type " must be a __host__ __device__ lambda") #else #define ASSERT_HOST_DEVICE_LAMBDA(type) #endif namespace at { namespace native { template C10_LAUNCH_BOUNDS_1(num_threads) __global__ void vectorized_elementwise_kernel(int N, func_t f, array_t data) { using traits = function_traits; int remaining = N - block_work_size * blockIdx.x; if (remaining < block_work_size) { // if this block handles the reminder, just do a naive unrolled loop auto input_calc = TrivialOffsetCalculator(); auto output_calc = TrivialOffsetCalculator<1>(); auto loader = memory::LoadWithoutCast(); auto storer = memory::StoreWithoutCast(); auto policy = memory::policies::unroll( data, remaining, input_calc, output_calc, loader, storer); elementwise_kernel_helper(f, policy); } else { // if this block has a full `block_work_size` data to handle, use vectorized memory access elementwise_kernel_helper(f, memory::policies::vectorized(data)); } } template C10_LAUNCH_BOUNDS_1(num_threads) __global__ void unrolled_elementwise_kernel(int N, func_t f, array_t data, inp_calc_t ic, out_calc_t oc, loader_t l, storer_t s) { int remaining = N - block_work_size * blockIdx.x; auto policy = memory::policies::unroll(data, remaining, ic, oc, l, s); elementwise_kernel_helper(f, policy); } // this function assume trivial 1d and no dynamic casting template static inline void launch_vectorized_kernel(int64_t N, const func_t& f, array_t data) { TORCH_INTERNAL_ASSERT(N > 0 && N <= std::numeric_limits::max()); using traits = function_traits; int64_t grid = (N + block_work_size - 1) / block_work_size; auto stream = at::cuda::getCurrentCUDAStream(); int vec_size = memory::can_vectorize_up_to(data); switch (vec_size) { case 4: vectorized_elementwise_kernel<4, func_t, array_t><<>>(N, f, data); C10_CUDA_KERNEL_LAUNCH_CHECK(); break; case 2: vectorized_elementwise_kernel<2, func_t, array_t><<>>(N, f, data); C10_CUDA_KERNEL_LAUNCH_CHECK(); break; case 1: { auto input_calc = TrivialOffsetCalculator(); auto output_calc = TrivialOffsetCalculator<1>(); auto loader = memory::LoadWithoutCast(); auto storer = memory::StoreWithoutCast(); unrolled_elementwise_kernel<<>>(N, f, data, input_calc, output_calc, loader, storer); C10_CUDA_KERNEL_LAUNCH_CHECK(); break; } default: TORCH_INTERNAL_ASSERT(false, "Unexpected vectorization size"); } } template static inline void launch_unrolled_kernel(int64_t N, const func_t& f, array_t data, inp_calc_t ic, out_calc_t oc, loader_t l, storer_t s) { TORCH_INTERNAL_ASSERT(N > 0 && N <= std::numeric_limits::max()); int64_t grid = (N + block_work_size - 1) / block_work_size; auto stream = at::cuda::getCurrentCUDAStream(); unrolled_elementwise_kernel<<>>(N, f, data, ic, oc, l, s); C10_CUDA_KERNEL_LAUNCH_CHECK(); } template void gpu_kernel_impl(TensorIteratorBase& iter, const func_t& f) { using traits = function_traits; using arg0_t = typename traits::result_type; constexpr int ntensors = traits::arity + 1; TORCH_INTERNAL_ASSERT(iter.can_use_32bit_indexing()); TORCH_INTERNAL_ASSERT(iter.ninputs() == traits::arity); TORCH_INTERNAL_ASSERT(iter.noutputs() == 1); at::detail::Array data; for (int i = 0; i < ntensors; i++) { data[i] = (char*)iter.data_ptr(i); } int64_t numel = iter.numel(); bool contiguous = iter.is_contiguous(); bool dynamic_casting = needs_dynamic_casting::check(iter); if (!dynamic_casting) { if (contiguous) { launch_vectorized_kernel(numel, f, data); } else { auto input_offset_calculator = make_input_offset_calculator(iter); auto output_offset_calculator = make_output_offset_calculator(iter); auto loader = memory::LoadWithoutCast(); auto storer = memory::StoreWithoutCast(); launch_unrolled_kernel(numel, f, data, input_offset_calculator, output_offset_calculator, loader, storer); } } else { at::detail::Array dtypes; for (int i = 0; i < traits::arity; i++) { dtypes[i] = iter.tensor(i + 1).scalar_type(); } auto loader = memory::LoadWithCast(dtypes); auto storer = memory::StoreWithCast(iter.tensor(0).scalar_type()); if (contiguous) { auto input_offset_calculator = TrivialOffsetCalculator(); auto output_offset_calculator = TrivialOffsetCalculator<1>(); launch_unrolled_kernel(numel, f, data, input_offset_calculator, output_offset_calculator, loader, storer); } else { auto input_offset_calculator = make_input_offset_calculator(iter); auto output_offset_calculator = make_output_offset_calculator(iter); launch_unrolled_kernel(numel, f, data, input_offset_calculator, output_offset_calculator, loader, storer); } } } }} // namespace at::native