Invoking my kernel in Matlab affects fft2 speed on GPU?

Hi, I happened to test such kind of program in Matlab. The code snippet is just like this:

 tic; fft2(img_d); toc   % img_d is a gpuArray on GPU

 .......

 %Set the block and grid sizes of my kernel
 ........

% invoke my kernel
tic
img_d=feval(myKernelName, .......);
toc

% do fft2 on gpu again
tic; fft2(img_d); toc   % img_d is a gpuArray on GPU

The result is stange:

the first fft2 on gpu costs:

Elapsed time is 0.000557 seconds.

but after invoke my kernel ,the second fft2 on gpu costs:

Elapsed time is 0.074028 seconds.

If I add such line after invoking line:

img = gather(img_d)

and then measure the time for the second fft2 on gpu the time looks right:

Elapsed time is 0.000392 seconds.

What is wrong with my kernel? or what is wrong with feval function?

[Jill Reese]

What release of MATLAB are you using?

[Fred Sigworth]

I am using Matlab R2012a

The gpu card is Tesla M2090.

Answer 1

[Jill Reese]

In R2012a, all GPU calculations run asynchronously with MATLAB. Because of this a new command was introduced to facilitate accurate code timing by synchronizing MATLAB and the GPU (see the R2012a release notes for more details). The form of the command is

wait(gpudev)

where gpudev is the object representing the GPU device to wait for. You will need to change your timings like so:

gpudev = gpuDevice();
tic; 
f = fft2(img_d); 
wait(gpudev);
toc   

%Set the block and grid sizes of my kernel
% invoke my kernel
tic
img_d=feval(myKernelName, .......);
wait(gpudev);
toc

% do fft2 on gpu again
tic; 
f = fft2(img_d); 
wait(gpudev);
toc

The code above should result in more sensible timings.