3D gpuArray, length of each dimension has influence on speed?

1 次查看（过去 30 天）
Hao Zhang
Hao Zhang 2014-5-21
评论： Joss Knight 2014-6-16
Hi,
I am doing 3D-FDTD simulations using gpuArray.
example 1: grid size: 32.000000 X 289.000000 X 289.000000, this means each field component is a 3D matrix of size 32.000000 X 289.000000 X 289.000000 In this case, the speed of the code is 17.1 million FDTD cells per second
example 2: simply change the grid size to 289.000000 X 289.000000 X 32.000000 The speed of the same code has increased to 24.6 million cells per second, nearly 50% speed gain!
Does anyone understand how the matlab distributes the matrix into each graphic processors? I always need the first dimension to be the smallest due to the geometry of the device I need to simulate, yes I can change my code to make the 3rd dimension to be the current first dimension of my code, but this needs a lot efforts to revise the code. So is there a way to distribute gpuarrays along a specified dimension like the codistributor1d does but with gpuArray? Any help would be appreciated. Thanks everybody!

请先登录，再回答此问题。

回答（1 个）

John D'Errico
John D'Errico 2014-5-21
编辑：John D'Errico 2014-5-21
No. I sincerely doubt that you could ever simply tell MATLAB to think of an array as being stored in memory in a different order WITHOUT actually changing that order.
You get the speed bump because of the order elements are stored in memory. If you need that speed bump, then do the work to get it. (In fact, all it takes is a permute when you make the appropriate call, so I don't totally see the problem. But maybe there is an issue in your code.)
  2 个评论
Hao Zhang
Hao Zhang 2014-5-21
Hi John, thanks for your answer, but I mean to distribute the gpuArray to each GPU processing unit like the codistributor1d does
Joss Knight
Joss Knight 2014-6-16
What do you mean by 'each GPU processing unit'? Do you have more than one GPU card? Or are you referring to the thousands individual GPU cores or SMs on that card? If the latter then this is perhaps the source of your confusion. These cores to all intents and purposes share memory; the only meaningful way to 'distribute' parts of the array to each thread or thread block is to permute the array.

请先登录，再进行评论。

类别

在 Help Center 和 File Exchange 中查找有关 GPU Computing in MATLAB 的更多信息

标签

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by