Sparse array created from a cupy ndarray has incorrect values

I’m trying to build a sparse array from a reasonably large cupy ndarray using cp.sparse.csr_matrix(x), but when I test the correctness of the result, some values mismatch between the input array and the output sparse array.

Code to reproduce

import cupy as cp

dense_ary = cp.ones(shape=(50000, 1003))
sones = cp.sparse.csr_matrix(dense_ary)
cp.testing.assert_array_almost_equal(dense_ary, sones.toarray(), decimal=1)

Error messages, stack traces, or logs

AssertionError: 
Arrays are not almost equal to 1 decimals

Mismatched elements: 17973 / 50150000 (0.0358%)
Max absolute difference: 1.
Max relative difference: 0.
 x: array([[1., 1., 1., ..., 1., 1., 1.],
       [1., 1., 1., ..., 1., 1., 1.],
       [1., 1., 1., ..., 1., 1., 1.],...
 y: array([[1., 1., 1., ..., 1., 1., 1.],
       [1., 1., 1., ..., 1., 1., 1.],
       [1., 1., 1., ..., 1., 1., 1.],...`

Conditions

• CuPy Version : 7.2.0
• CUDA Root : /usr/local/cuda
• CUDA Build Version : 10000
• CUDA Driver Version : 10010
• CUDA Runtime Version : 10000
• cuBLAS Version : 10000
• cuFFT Version : 10000
• cuRAND Version : 10000
• cuSOLVER Version : (10, 0, 0)
• cuSPARSE Version : 10000
• NVRTC Version : (10, 0)
• cuDNN Build Version : 7605
• cuDNN Version : 7600
• NCCL Build Version : 2406
• NCCL Runtime Version : 2507

[emcastillo]

I have reproduced the issue, and I think this might be a cusparse bug. cc @anaruse @jakirkham @pentschev @leofang

When creating a csr matrix copy only calls the corresponding cusparseXdense2csr routine, the data array that is returned from there has some entries set to zero. This can be seen in cupy/cusparse.py line 690.

This is a reduced example

import cupy as cp

dense_ary = cp.ones(shape=(50000, 9))
sones = cp.sparse.csr_matrix(dense_ary)
print(dense_ary.sum(), sones.sum(), sones.toarray().sum())

If we look at the elements that are set to 0 in the cusparse dense2csr returned data array we see that starting from the row 40960 to row 49999 all elements in column 8 are set to 0. Even if we increase the number of columns to 1003, the 0s start always in the row 40960 and will be in the column 1002. This only happens when the column number is even.

Note that this error only happens when creating a csr matrix and not a csc one.

[anaruse]

This issue should be left open, in my opinion, but it is a issue of cuSparse, more specifically, cusparse<t>dense2csr().