Computational Speed of MLDIVIDE When Called From MEX Function

4 次查看（过去 30 天）

显示更早的评论

Bård Skaflestad 2015-6-9

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/223070-computational-speed-of-mldivide-when-called-from-mex-function

编辑： James Tursa 2024-1-28

在 MATLAB Online 中打开

Has anyone ever seen significant performance reduction on

x = A \ b

when A is created in a mexFunction using mxCreateSparse and filling in the Ir, Jc, and Pr arrays?

The reason I ask is that I'm using mexCallMATLAB to invoke mldivide on a moderately sized system (about 416,000 unknowns, 1.63 million non-zeros) and there is about a factor of 3.5 difference in computational speed (35 seconds vs. 10 seconds) in calling mldivide from MEX rather than directly from within the MATLAB environment.

Moreover, if I replace the direct call to mldivide with a call to a simple wrapper that does

[i, j, v] = find(A);
x = sparse(i, j, v) \ b;

I recover MATLAB's "inline" 10 second run-time and produce the same numerical solution (relative difference in the order of eps). I am quite probably doing something wrong and/or unusual in my mexFunction that's causing the slowdown, but I can't think of anything obvious. I have verified that the row indices appear in sorted (ascending) order within each column.

Any hints, explanations or suggestions would be greatly appreciated.

5 个评论
显示 3更早的评论隐藏 3更早的评论

Bård Skaflestad 2015-6-9

> I was asking for the code that built the sparse matrix.

Ah, right. Like I said, part of that is proprietary and in any case, it runs into a couple of thousand lines of code all included. If it helps, I can share the actual matrix and right-hand side.

> What version of MATLAB are you running?

In this case, R2014b. I haven't tried 2015a.

> You might try running SPOK on it to see if there is anything amiss

Thank you for the tip. I'll do that!

Bård Skaflestad 2015-6-9

> You might try running SPOK on it to see if there is anything amiss

That turned out to be fantastically useful advice. Thank you so much, James.

My matrix had about 1.6 million explicit zero entries, which spok helpfully pointed out. I was using a static sparsity pattern derived from the connection structure of the underlying grid. The entries are structurally non-zero, but in any given problem some of them will be zero.

Removing the zero entries before calling mldivide recovers the expected runtime. My initial suspicion that I was doing something wrong was therefore confirmed and this naturally also explains why the find call cured the problem.

请先登录，再进行评论。

请先登录，再回答此问题。