[MUMmerGPU-devel] Building MummerGPU 2.0 with CUDA 3.2

[Meehan, Joe* <Joe...@fd...>]

This doesn't seem to be a very active list, but I did want to report that I was able to build MummerGPU 2.0 on CUDA 3.2 after making a few small changes to overcome the problems mentioned in other posts:

The error on line 468 of mummergpu.cu regarding an "unsigned int *" argument being incompatible with a "size_t *" parameter is due to a change in the parameters of cuMemGetInfo() in CUDA 3.2.  See the following PDF for more info:

http://developer.download.nvidia.com/compute/cuda/3_2/toolkit/docs/CUDA_3.2_Readiness_Tech_Brief.pdf

NVIDIA says that fixing this with a cast is unsafe.  The solution is to change the variables used for memory size from "unsigned int" to "size_t", which has some ripple effects throughout the source code.

The multiple definition of GetRef is due to it being defined once for the CPU and once for the GPU.  I'm not sure why this is a problem, but I just renamed the GPU version in mummergpu_kernel.cu to getRefDev and changed the function calls in that file to match.

Finally, the inability to find -lcudart in my case was due to the Makefile specifying /usr/local/lib instead of /usr/local/lib64.  I changed the definition of LIB in the makefile to use lib64 instead of lib, and it worked.

I haven't tried to upload any of these changes to SourceForge, but if anyone is interested I'll look into it.

Regards,
Joe Meehan