Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

syevjBatched from cuSOLVER sometimes doesn't work correctly

See original GitHub issue

I didn’t find this issue being reported. I’d like to let CuPy team know that cuSOLVER’s syevjBatched (cupy.cusolver._syevj_batched) contains a bug leading to wrong results for some inputs. Eigenvalues are just the diagonal of the matrix, that is not correct. Single matrix input syevj doesn’t have this problem.

In [1]: import cupy

In [2]: cupy.random.seed(1)

In [3]: for i in range(3):
 ...:     a = cupy.random.normal(size=(2, 3, 3))
 ...: 

In [4]: cupy.cusolver._syevj_batched(a, "L", True)
Out[4]: 
(array([[-3.64909523, -1.09802557, -0.98772015],
      [-1.23675442,  1.13945238,  1.79567052]]),
array([[[ 0.        ,  1.        ,  0.        ],
       [ 1.        ,  0.        ,  0.        ],
       [ 0.        ,  0.        ,  1.        ]],

      [[-0.34828366, -0.32597124,  0.87888637],
       [ 0.90845376, -0.34853093,  0.23073352],
       [ 0.23110659,  0.87878834,  0.41751742]]]))

# eigenvalues above are just sorted diagonal elements
In [5]: a[0].diagonal()
Out[5]: array([-1.09802557, -3.64909523, -0.98772015])

# correct result for the 0th matrix is
In [6]: cupy.cusolver.syevj(a[0], "L", True)
Out[6]: 
(array([-3.74866036, -1.15578506, -0.83039553]),
 array([[-0.1051637 ,  0.8504732 , -0.51539881],
        [ 0.98278763,  0.00972565, -0.18448274],
        [ 0.15188504,  0.52592846,  0.83685733]]))

In [2]: cupy.show_config()
OS                           : Linux-5.4.0-65-generic-x86_64-with-glibc2.31
CuPy Version                 : 8.5.0
NumPy Version                : 1.19.4
SciPy Version                : 1.6.0
Cython Build Version         : 0.29.22
CUDA Build Version           : 11000
CUDA Driver Version          : 11020
CUDA Runtime Version         : 11000
cuBLAS Version               : 11200
cuFFT Version                : 10201
cuRAND Version               : 10201
cuSOLVER Version             : (10, 6, 0)
cuSPARSE Version             : 11101
NVRTC Version                : (11, 0)
Thrust Version               : 100909
CUB Build Version            : 100909
cuDNN Build Version          : 8100
cuDNN Version                : 8004
NCCL Build Version           : 2804
NCCL Runtime Version         : 2804
cuTENSOR Version             : None

Issue Analytics

State:
Created 3 years ago
Comments:8 (8 by maintainers)

Top GitHub Comments

4reactions

emcastillocommented, Mar 10, 2021

@anaruse might this be a cuda bug?

1reaction

IvanYashchukcommented, Mar 30, 2021

Oh, that’s right, my bad… I should sleep more.

Top Results From Across the Web

cuSOLVER API Reference - NVIDIA Documentation Center

To correct: call cusolverCreate() prior to the function call; and check that the hardware, an appropriate version of the driver, and the cuSolver...

cuSOLVER Library - Rice University

This function reports residual of syevj or sygvj. It does not support. syevjBatched. If the user calls this function after syevjBatched, the error....

Segmentation Fault in Pycuda using NVIDIA's cuSolver Library

Try modifying the cusolverLUFactorization() function as follows: ... m, n, int(mtx_gpu.gpudata), n, ctypes.pointer(work)) print status print work.value.

Accelerating Spectroscopic Data Processing Using Python ...

labels, it sometimes difficult to decipher the names of low-level ... of a syevjBatched function in CUDA cuSOLVER library which.

Wave-packet continuum discretisation for nucleon ... - arXiv

Speeding up computations very often comes at the expense of accuracy ... In this work we employ the standard MI method [17] which...