Using PTX with Runtime Possible?

Accelerated Computing CUDA CUDA Programming and Performance
1

I am using the Runtime environment (not the driver level) in Visual Studio. I am wondering if it is possible to compile an executable with a PTX embedded, which will compile (using JIT) to a device specific binary at runtime. The documentation I have read seems to say this is possible, but I cannot find any examples in the SDK which use the Runtime API–only the Driver API. I tried changing the custom build rule for the kernel.cu file to only compile .ptx and turned on ‘keep’, so the file would not be deleted. When I try to compile the main .cpp file, I get unresolved externals if I do not specify this .ptx file in the Linker. If I do specify the .ptx file in the Linker, I get an ‘invalid or corrupt file’ error. Could someone explain the process of doing this in Visual Studio, if it is possible using the Runtime API?

Thanks

2

The runtime will JIT automatically.

3

tmurray,

Thank you very much for the reply. So, to enforce a PTX compilation of kernel.cu code only (no cubin) in Visual Studio, do I just enter something like: ‘gencode=arch=compute_10,code=compute_10’ as an argument/flag to nvcc or do I need to change the custom build rule for the kernel.cu file to generate .ptx code only? Since either is a change to the custom build rule, do I modify only the kernel.cu build rule arguments or the Solution build rule arguments or both? (I know I am showing my ignorance here about MSVS in general).

4

bump to tmurray

5

From my experience, nvcc always embeds ptx in compiled .cu files. Whether or not it is actually loaded rather than an embedded cubin is up to the runtime…