gpuArray for vpa (variable precision arithmetic)

Question

NICOLA BARBINI 2019-12-6

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/495258-gpuarray-for-vpa-variable-precision-arithmetic

编辑： Walter Roberson 2019-12-20

采纳的回答： JESUS DAVID ARIZA ROYETH

Problem:

I have big matrices with elements defined through variable precision arithmetic of the symbolic toolbox (vpa(x)).

Extracting the eigenvalues out of these matrices takes about 7 minutes.

operations that imply inversion-ish operations (C=A\B) also take some minutes.

Question:

Could I have any benefit from gpu computation?

To be clearer on what I don't understand:

I know that symbolic maths are not supported by GPU computing but in my case the variable is not really symbolic, it just has higher precision.

Thank you for any clarification

3 个评论
显示 1更早的评论隐藏 1更早的评论

NICOLA BARBINI 2019-12-20

编辑：NICOLA BARBINI 2019-12-20

I think this answers the last part of the question:

extended precision numbers can benefit from accelerated GPU computing.

Unfortunately, if I understand correctly, the linked library supports only basic operations

on multiple precision numbers, not on "compound" data as matrices and complex numbers.

Anyway thank you, I couldn't find this package on my own and it addreses part of the problem exactly.

Walter Roberson 2019-12-20

编辑：Walter Roberson 2019-12-20

https://devtalk.nvidia.com/default/topic/1052081/cuda-programming-and-performance/arbitrary-precision-arithmetic/ mentions two others.

I found the download page for CAMPARY but unfortunately there is no documentation included.

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

JESUS DAVID ARIZA ROYETH 2019-12-6

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/495258-gpuarray-for-vpa-variable-precision-arithmetic#answer_405233

In this case you are going to have to sacrifice something, if you use VPA the matrices become symbolic, therefore all the operations you do with them are going to slow down, GPU does not support these matrices so it would not give you anything.

On the other hand if you stay with "normal" matrices you will have a maximum of 15 digits after the decimal point, and using the GPU could give you something depending on the operations you are doing

If you really need too much precision you can search about Pari / GP that you could integrate with Matlab, or also Accelerating MATLAB Code Using Codegen and Mex Files