Kron error when dealing with sparse gpuarrays

Question

Henry Abrahamson 2019-10-17

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/485793-kron-error-when-dealing-with-sparse-gpuarrays

编辑： Matt J 2019-10-17

When I attempt to use kron with a sparse gpu array as an input on R2019a, I get the following error:

"Error using matlab.internal.sparse.kronSparse

Input must be a sparse matrix."

This indicates to me that MATLAB is having issues either making or recognizing when gpuarrays are made sparse, since I've done the simple test case:

a = gpuarray([1 2]);
b = gpuarray([1 2]);
test1 = kron(a,b)
a = sparse(a)
test2 = kron(a,b)

and only the assignment of test2 gives the error. Does anyone know if there's a workaround for this, or am I stuck with using my cpu for these calculations?

3 个评论
显示 1更早的评论隐藏 1更早的评论

Matt J 2019-10-17

Also, what operations will be done subsequently with kron(a,b) once it is computed? Knowing that, one can often find workarounds that avoid computing the Kronecker product altogether.

Henry Abrahamson 2019-10-17

编辑：Henry Abrahamson 2019-10-17

I know that b isn't sparse, there's a line in the kron function (like 37 or something) that's supposed to make b sparse and then pass it to that internal function, but the fact that it doesn't work for me is my whole issue.

The Kronecker is just being used to simplify a bit of matrix multiplication I'd have to do. But now that you mention it, I think I see a way to do it the long way (in terms of some extra lines of code) that should get the same results without a terrible performance hit. Thanks for the tip!

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Matt J 2019-10-17

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/485793-kron-error-when-dealing-with-sparse-gpuarrays#answer_396724

编辑：Matt J 2019-10-17

You may want to look at,

https://www.mathworks.com/matlabcentral/fileexchange/25969-efficient-object-oriented-kronecker-product-manipulation

which allows you to do many kinds of manipulations with Kronecker products without computing them in explicit matrix form. It should work with gpuArray operands.