how can I make this code faster on the GPU?

Hi.. This is my first experiment with GPU. When I tried to compare the executed time between the cpu and gpu work I found that GPU code take long time . I have an INTEL i5 processor and a NVIDIA 'GeForce GT 740M' with ComputeCapability: '3.5' in my computer, and MATLAB 2013a. I used different size of images but I found same result My current code runs significantly faster on the CPU, even without parfor or spmd, than it does on the GPU. how can I make this faster on the GPU?

The following is CPU code

tic

for i=1:4:size(f,1)

for j=1:4:size(f,2)

gg(i,j)=f(i,j);

end

end;toc

Elapsed time is 0.078421 seconds.

The following is GPU code

ff=gpuArray(f);

tic

for i=1:4:size(ff,1)

for j=1:4:size(ff,2)

gg(i,j)=ff(i,j);

end end;toc

Elapsed time is 1.461808 seconds.

[Thomas]

MY guess is your image sizes are still small to compensate for the overhead of moving to the GPU. Also, Why would you wnat to run sometihng on the GPU that only takes about 0.07 sec on the CPU anyway?

[Matt J]

In addition to what the others have told you, tic...toc is not a good way to measure computation time on the GPU. You should use gputimeit().

[mona]

@Thomas we try the code on different sizes but we found the same result

Answer 1

[Anand]

There is no operation in the above code that can take advantage of the GPU. All you are doing is looping over every 4th pixel serially. The GPU is well-suited to compute-heavy, parallelizable tasks. In order to take advantage of the GPU in MATLAB, you need to either use functions like arrayfun,<http://www.mathworks.com/help/distcomp/bsxfun.html bsxfun> which will work on multiple elements of the array in parallel or functions that are accelerated for the GPU.

Answer 2

[Royi Avital]

Hopefully, next versions of MATLAB will support OpenCL 2.0. When that happens, CPU and GPU will use the same memory space which will remove the Overhead of GPU Computing.

Assuming using CPU's from AMD which includes GPU which share the same memory with the CPU.

[Michal]

OpenCL support at next MATLAB version?? No way!! Mathworks did not express eny intention to support OpenCL in the near future.