GPU arrayfun is so slow, what is going on?

Asked by Hao Zhang on 11 Dec 2018

Latest activity Commented on by Derrick Ling on 20 Apr 2019

33 views (last 30 days)

Hi,

I am trying to understand what the GPU arrayfun is doing? The following is a test code.

clear;clc;close all
gd=gpuDevice();
reset(gd);
N=2e3;
a=rand(60,N,'single','gpuArray');
tic;
b=sum(a,1);
wait(gd);
toc;
tic;
c=arrayfun(@(i) sum(a(:,i),1),(1:N));
wait(gd);
toc;

The results are:

Elapsed time is 0.000468 seconds.
Elapsed time is 0.584521 seconds.

What is going on here? 1000 times difference?? I would expect similary runtime since GPU arrayfun is supposed to be executed parallel on GPU cores. Did I make stupid errors on using the arrayfun?

Thanks!

1 Comment

Hao Zhang on 13 Dec 2018

Hi, I come back for more updates. I have successfully vectorized and implemented my particle simulation on GPU. The speed up is astonishing, ~10 times faster than CPU code. Thanks Matt J and Joss Knight for their wonderful suggestions.

Now the other part of the code (except neighbor search but solving the fluid equations) is so fast that the limiting part now is the matlab function knnsearch, which uses kdtree algorithm runing on CPU. It takes 85% percent of the runtime (see the following code profiler results)

The function 'knnCPU_kdtree_func' uses the matlab built-in function knnsearch with kdtree algorithm runing on CPU. The other functions are doing the real math runing on GPU only consumes 10% of the total time.

I wonder is there any GPU implementation of k-nearest neighbor search that I can free download and using as a function call in my matlab code? Many thanks.

Products

Parallel Computing Toolbox

Release

R2017b

2 Answers

Link

Answer by Matt J on 11 Dec 2018

Edited by Matt J on 11 Dec 2018

Accepted Answer

What is the most efficient way to vectorize the above code

I would say, as follows,

idx_Neighbor=randi([1 N],60,N,'uint8');
temp=p(idx_Neighbor);
temp=temp+p.';
ax=sum(temp .* delW_x,1);

4 Comments

Show 1 older comment

Matt J on 11 Dec 2018

However it is a bit pity that the vectorized code consume more memory

It shouldn't. I conserved memory by creating idx_Neighbour as uint8 instead of double.

This is the first time I know one can do this eventhough the indexing is not the same dimension as the inferred array, is there any matlab documentation on this?

Oddly, I can't find any!

How about the efficiency of the vectorized code on GPU comparing with highly optimized cuda C code?

You would have to implement it to find out...

Hao Zhang on 11 Dec 2018

yes, but one can do uint8 to any code above. And if the indexing is larger then one has to use uint32. But the improvement of vectorizing is so much that one has to bear with a bit more memory usage.

Maybe a matlab developer can say something about the efficiency comparing to cuda C. I hope it will be really close.

Derrick Ling on 20 Apr 2019

Hi, which code did you use to run GPU? arrayfun? gpuArray?

Where or how did you insert the code?

And is there any advice you would give to make this code nicer? Undefined function error appears if I remove bbb = 0.

bbb = 0;
bbb = bbb + h*W_4';

Link

Answer by Joss Knight on 11 Dec 2018

You haven't called GPU arrayfun here, you've called CPU arrayfun and in the arrayfun function you are doing stuff on the GPU. This is because none of the arguments to your arrayfun call is a gpuArray.

You could force it to use GPU arrayfun by converting your input:

c = arrayfun(@(i) sum(a(:,i),1), gpuArray(1:N));

However, you'll immediately find it errors, because sum is not supported for GPU arrayfun. Obviously this is just a toy example, but the solution here is sum(a,1), not arrayfun.

4 Comments

Show 1 older comment

Matt J on 11 Dec 2018

I need to do c=sum(a.*repmat(b,1,2e3),1).

Your operation does not require repmat nor even sum, but simply

c=b.'*a;

But even if you were to use sum, recent Matlab no longer requires repmat, e.g.,

a=gpuArray.rand(60,2e3); b=a(:,1);
c=sum(a.*b,1);

will work fine.

Joss Knight on 11 Dec 2018

Thanks Matt.

GPU arrayfun is very special, you should read the documentation and list of supported functions. It only supports element-wise functionality, so you can't do any vector operations. That means you can't index an array unless you're indexing a single element or an up-level variable, you can't call sum or any other reduction or accumulation, and you can't output anything other than a scalar. This is because your arrayfun function gets compiled into a single CUDA kernel with no inter-thread communication. So it's incredibly useful and efficient when used within its limitations.

Nearly always (in my experience) when you want to do something more complex with vector operations, you can translate your code into a series of vectorized calls to normal MATLAB matrix functions, arrayfun, and pagefun.

Hao Zhang on 11 Dec 2018

This is a brillant solution! Thanks for making this happen.

However, if I want to do something even more complex (And this is what I actually need to do, instead of the toy examples before :)). So I need to do the following code:

clear;clc;close all;
N=2e3;
idx_Neighbor=randi([1 N],60,N);
p=rand(N,1);
delW_x=rand(60,N);
ax=zeros(N,1);
tic;
for j=1:N
    temp=idx_Neighbor(:,j);
    inner=p(temp)+p(j);
    ax(j)=sum(inner.*delW_x(:,j));
end
toc;

What is the most efficient way to vectorize the above code (so without using the for loop) and avoiding using repmat as much as possible? Thanks!

Products

Parallel Computing Toolbox

Release

R2017b

You are now following this question

Hao Zhang (view profile)