frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 21:14:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

[AMD] use DC method for linalg.eigh (#140327)

Browse Source 
Summary: Jacobi method has larger numerical errors, see D64997718, use divide-and-conquer method instead.

Test Plan: CI

Differential Revision: D65786796

Pull Request resolved: https://github.com/pytorch/pytorch/pull/140327
Approved by: https://github.com/jianyuh
This commit is contained in:
Haoran Zhang
2024-11-12 19:17:25 +00:00
committed by PyTorch MergeBot
parent 726424f4de
commit cc8e832066
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
aten/src/ATen/native/cuda/linalg/BatchLinearAlgebraLib.cpp
View File

@ -1427,9 +1427,9 @@ static void linalg_eigh_cusolver_syevj_batched(const Tensor& eigenvalues, const
}
void linalg_eigh_cusolver(const Tensor& eigenvalues, const Tensor& eigenvectors, const Tensor& infos, bool upper, bool compute_eigenvectors) {
// for ROCm's hipSolver, syevj is fastest.
#ifdef USE_ROCM
linalg_eigh_cusolver_syevj(eigenvalues, eigenvectors, infos, upper, compute_eigenvectors);
// syevj has larger numerical errors than syevd
linalg_eigh_cusolver_syevd(eigenvalues, eigenvectors, infos, upper, compute_eigenvectors);
#else
if (use_cusolver_syevj_batched_ && batchCount(eigenvectors) > 1 && eigenvectors.size(-1) <= 32) {
// Use syevjBatched for batched matrix operation when matrix size <= 32