Sharing Graphic card for CUDA computing and graphic interface

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hello all.

I’m writting a software consisting on a Java-made graphic interface and a CUDA-based computing core.

For my objective, I need as much memory as possible, so, when the software starts, I check all the free GPU memory with cuMemGetInfo and allocate a big memory block, leaving a few free MB in the graphics card.

My software worked nicely on windows XP, but now, I’m testing it on windows vista/7 and have some errors. These errors are characterized by calculation errors, screen corruption and cudaUnknownError message when copy from GPU memory.

These errors usually appear when I resize the graphic interface or when I create and resize a 3D surface made with java3D (an optional part of my software). The frequency of these errors decreases drastically (but not dissapears) when I set more free memory in CUDA (maybe 128MB).

Well, I think that the problem is that CUDA core and my graphic interface fight for using the graphic card memory, but I don’t know if there is any way of manage this memory sharing.

So, if you know how to solve this in some way, please tell it to me.

Thank you!

Néstor

#2

Vista and Windows 7 use some GPU memory for the Aero effects. If you disable all those features then you free up around 60MB of GPU memory (as far as I remember).

#3

Thanks for your repply chippies.

Yes, I know that Aero uses a good amount of memory.

Finally I found a Kernel that, since its execution, the program started to fail. I checked it several times an I didn’t found any bug, so finally I decided to change the program algoritm in that part for another implementation approach and now It works nicely.

Sorry for this post, I was totally convinced it was not a bug of my software because it worked right on XP.

Cya!