CUDA pointers: 64-bit or not?

Accelerated Computing CUDA CUDA Programming and Performance
1

I’ve run into some problems dealing with several pointers in a kernel parameter list. It might boil down to my misunderstanding of pointers in CUDA.

Are pointers sometimes/always 64-bit, on both 32-bit and 64-bit hosts?

The matrix multiply driver API example uses cuParamSeti (32-bit integer) and separates pointers by only 4 bytes. However, other posts on this forum say that the separation must be 8 bytes and this matches some of my past experience. In which case, what happens to the other 4 bytes that cuParamSeti doesn’t touch?

2

In my experience nvcc uses the default setting of the host, which was x64 in my case. I added -m32 in my makefiles to be sure everything is x32.

If you always want x64 I would suggest enforcing this by -m64

Please be sure to link with the right libs, I’ve had numerous problems with this and no error from the compiler/linker.

Sry, I’ve no experience with the driver api.

3

Thanks, I switched to the (hidden) flag -m32 and made the pointer separation 4 bytes. My parameter problem - basically when I added a few unused pointers as parameters to a kernel, separated by 8 bytes, CUDA was reading subsequent parameters from an incorrect offset - went away.

I wonder why there’s 64-bit support in there at all. CUdeviceptrs are 32-bit and CuParamSeti is 32-bit (I guess CuParamSetv could work).