gpu array calculation limited memory

Hi, i have a quadro 2200 with 4 gig ram. I would like to do a calculation with a variable that is 6 gigabytes big. Since gpuarray is limited, I wanted to know if there is any other solutiob

[Matt J]

What do you mean "other" solution? What solution are you using now? And in what way isn't it working?

The obvious solution that comes to mind is just to split the data into smaller pieces and send them in separate jobs.

[AA]

That is what I had in my mind, but how do I have to reset my gpuarray after that each time? That is time consuming. Is there an overwrite function or something like that? For instance consider the variable x which is 6 gigabyte. That variable is split into x1 and x2, of which each of them is 3 gigabyte. Now I do gpuArray(x1) perform the calculation and then I have to do gpuArray(x2) but that is not accepted as the gpumemory is full with x1. How do I incorporate gpuArray(x2)?

Answer 1

[Matt J]

You should clear memory on the GPU before doing further calculations. For example, if y1 is the gpuArray resulting from the calculation with x1, then you could clear x1 and bring y1 back to the CPU with,

   clear x1
   y1=double(y1);

This will free memory on the GPU for x2.

[AA]

for x = 1:100
  for y =  1:61
             b = abs(bsxfun(@minus,gpuArray(q{x,y}),kevin))
               averagesAcrossColumns = mean(b, 2);
               b = abs(bsxfun(@minus,b,averagesAcrossColumns));
               s{x,y} = sum(b, 2);
               clear q{x,y}
   end
end

[Matt J]

clear q(x,y)

would give you an error, I'd expect. Also, If you want s{x,y} data held on the host device instead of the GPU, you could do instead,

   s{x,y} = double( sum(b, 2) );

[AA]

clear q command doesnt give me an error because it gets cleared after each loop and the result stays at the gpu. I have enough space to store all of s on the gpu but not to store q. So q gets thrown out of the GPU.