FM / cuda

/usr/local/lib64/python3.6/site-packages/torch/include/ATen/native/cuda/

/usr/local/lib64/python3.6/site-packages/torch/include/ATen/native/cuda

Name	Size	Mode	Actions
BatchLinearAlgebraLib.h	3114	0644	edit dl rm
block_reduce.cuh	2549	0644	edit dl rm
CompositeRandomAccessor.h	929	0644	edit dl rm
CUDALoops.cuh	7598	0644	edit dl rm
CuFFTPlanCache.h	19282	0644	edit dl rm
CuFFTUtils.h	1892	0644	edit dl rm
DeviceSqrt.cuh	585	0644	edit dl rm
DistributionTemplates.h	27435	0644	edit dl rm
EmbeddingBackwardKernel.cuh	715	0644	edit dl rm
ForeachFunctors.cuh	16851	0644	edit dl rm
GridSampler.cuh	11316	0644	edit dl rm
im2col.cuh	6577	0644	edit dl rm
KernelUtils.cuh	2553	0644	edit dl rm
LaunchUtils.h	306	0644	edit dl rm
Loops.cuh	9997	0644	edit dl rm
Math.cuh	13840	0644	edit dl rm
MemoryAccess.cuh	12463	0644	edit dl rm
MiscUtils.h	3341	0644	edit dl rm
MultiTensorApply.cuh	7552	0644	edit dl rm
Normalization.cuh	74441	0644	edit dl rm
PersistentSoftmax.cuh	14635	0644	edit dl rm
Randperm.cuh	2114	0644	edit dl rm
Reduce.cuh	38784	0644	edit dl rm
Resize.cuh	1919	0644	edit dl rm
ROCmLoops.cuh	13526	0644	edit dl rm
SortingCommon.cuh	5688	0644	edit dl rm
SortingRadixSelect.cuh	11918	0644	edit dl rm
SortUtils.cuh	5549	0644	edit dl rm
TensorModeKernel.cuh	14391	0644	edit dl rm
UniqueCub.cuh	345	0644	edit dl rm
UpSample.cuh	7552	0644	edit dl rm
vol2col.cuh	8297	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/include/ATen/native/cuda/BatchLinearAlgebraLib.h (3114B)

#pragma once

#include <ATen/Context.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDACachingAllocator.h>

#include <ATen/native/LinearAlgebraUtils.h>
#include <ATen/native/cuda/MiscUtils.h>

#if defined(CUDART_VERSION) && defined(CUSOLVER_VERSION)
#define USE_CUSOLVER
#endif

// cusolverDn<T>potrfBatched may have numerical issue before cuda 11.3 release,
// (which is cusolver version 11101 in the header), so we only use cusolver potrf batched
// if cuda version is >= 11.3
#if CUSOLVER_VERSION >= 11101
  constexpr bool use_cusolver_potrf_batched_ = true;
#else
  constexpr bool use_cusolver_potrf_batched_ = false;
#endif

// cusolverDn<T>syevjBatched may have numerical issue before cuda 11.3.1 release,
// (which is cusolver version 11102 in the header), so we only use cusolver syevj batched
// if cuda version is >= 11.3.1
// See https://github.com/pytorch/pytorch/pull/53040#issuecomment-793626268 and https://github.com/cupy/cupy/issues/4847
#if CUSOLVER_VERSION >= 11102
  constexpr bool use_cusolver_syevj_batched_ = true;
#else
  constexpr bool use_cusolver_syevj_batched_ = false;
#endif

namespace at {
namespace native {

void geqrf_batched_cublas(const Tensor& input, const Tensor& tau);
void triangular_solve_cublas(Tensor& A, Tensor& B, bool left, bool upper, TransposeType transpose, bool unitriangular);
void triangular_solve_batched_cublas(Tensor& A, Tensor& B, bool left, bool upper, TransposeType transpose, bool unitriangular);
void gels_batched_cublas(const Tensor& a, Tensor& b, Tensor& infos);
void lu_solve_batched_cublas(const Tensor& b, const Tensor& lu, const Tensor& pivots, TransposeType transpose);

#ifdef USE_CUSOLVER

// entrance of calculations of `inverse` using cusolver getrf + getrs, cublas getrfBatched + getriBatched
Tensor _inverse_helper_cuda_lib(const Tensor& self);
Tensor& _linalg_inv_out_helper_cuda_lib(Tensor& result, Tensor& infos_getrf, Tensor& infos_getrs);

// entrance of calculations of `svd` using cusolver gesvdj and gesvdjBatched
std::tuple<Tensor, Tensor, Tensor> _svd_helper_cuda_lib(const Tensor& self, bool some, bool compute_uv);

// entrance of calculations of `cholesky` using cusolver potrf and potrfBatched
void cholesky_helper_cusolver(const Tensor& input, bool upper, const Tensor& info);
Tensor _cholesky_solve_helper_cuda_cusolver(const Tensor& self, const Tensor& A, bool upper);
Tensor& cholesky_inverse_kernel_impl_cusolver(Tensor &result, Tensor& infos, bool upper);

void geqrf_cusolver(const Tensor& input, const Tensor& tau);
void ormqr_cusolver(const Tensor& input, const Tensor& tau, const Tensor& other, bool left, bool transpose);
Tensor& orgqr_helper_cusolver(Tensor& result, const Tensor& tau);

void linalg_eigh_cusolver(const Tensor& eigenvalues, const Tensor& eigenvectors, const Tensor& infos, bool upper, bool compute_eigenvectors);
void lu_solve_looped_cusolver(const Tensor& b, const Tensor& lu, const Tensor& pivots, TransposeType transpose);

void lu_looped_cusolver(const Tensor& self, const Tensor& pivots, const Tensor& infos, bool get_pivots);

#endif  // USE_CUSOLVER

}}  // namespace at::native

cmd: