How to increase the speed of submatrix data transfer using gpuArray rather than CPU?

Question

Sung Hwan Heo 2016-4-6

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/277457-how-to-increase-the-speed-of-submatrix-data-transfer-using-gpuarray-rather-than-cpu

关闭： MATLAB Answer Bot 2021-8-20

Hello. My problem is... How can I reduce the difference btw. them? The size of matrix is smaller than 1000 by 1000. If possible (whether this is intrinsic limitation) please tell me a proper method.

A = rand(512);
gA = gpuArray(A);
tic; B = A(50:100,50:100); toc
tic; gB = gA(50:100,50:100); toc

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

此问题已关闭。

Answer 1

Walter Roberson 2016-4-6

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/277457-how-to-increase-the-speed-of-submatrix-data-transfer-using-gpuarray-rather-than-cpu#answer_216685

Instead of using tic and toc, use timeit() for the first case, and gputimeit() for the second case. Those will give you more accurate times.

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Sung Hwan Heo 2016-4-6

Already did that, but there was no big changes @.@;;;

Answer 2

Joss Knight 2016-4-15

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/277457-how-to-increase-the-speed-of-submatrix-data-transfer-using-gpuarray-rather-than-cpu#answer_218046

I think the problem is intrinsic. Even on my Kepler K20c the GPU is slower for this case. This is a memory intensive non-coherent operation and you don't have enough data.

Best solution: Try to do more! Perhaps if you're doing this same indexing to multiple matrices, you should do them all at once, for instance.