Memory mapping large files into the GPU or managed memory space?

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi all,

Is there a way of memory mapping large files into the address space of the GPU? Is this supported by Enterprise CUDA cards (say, a V100/A100) or even high end cards like the Titan V using stock Linux kernels?

Christian

#2

This may be of interest. I don’t really know how large the file can be, haven’t tried to find the upper bound. It seems to work at least up to 100MB.

Without knowing more about your goals, I think my general advice might be to just try using managed memory, although I acknowledge that is not a file mapping, per se. I don’t know of any method similar to above that has managed memory characteristics.