NVCC separate compilation with PTX output

Question

Just to see what kind of code CUDA is generating I like to compile to ptx in addition to an object file. Since some of my loop unrolling can take quite a while I'd like to be able to compile *.cu→*.ptx→*.o instead of wasting time with both *.cu→*.ptx and *.cu→*.o, which I'm currently doing.

Simply adding -ptx to the nvcc *.cu line gives the desired ptx output.

Using ptxas -c to compile *.ptx to *.o works, but causes an error in my executable linking: Relocations in generic ELF (EM: 190).

Attempting to compile the *.ptx with nvcc fails silently, outputting nothing.

this image is quite helpful:

Is there some option I need to pass to ptxas? How should I properly compile via ptx with separate compilation? Alternatively, can I just tell nvcc to keep the ptx?

Answer 1

Alternatively, can I just tell nvcc to keep the ptx?

Yes, you can tell nvcc to keep all intermediate files, one of which will be the .ptx file.

nvcc -keep ...

Keeping all the intermediate files is a bit messy, but I'm sure you can come up with a script to tidy things up, and only save the files you want.