Hey folks - I have succeeded in transferring data from Perl to my video card, calling a kernel to operate on that data, and transferring the data back to Perl. Herein I describe how I got it working. Comments, suggestions, advice for moving forward, and general enthusiasm would be greatly appreciated.
The first question I asked myself was this: Could I simply use Inline::C and specify nvcc as my compiler? Answer: No. ExtUtils::MakeMaker sends a whole boatload of compiler flags to gcc that makes nvcc choke. However, there is a way to send special compiler flags to the compiler backend, using the -Xcompiler=flag1,flag2,... option for nvcc. Thence came the next question: Could I use a perl script for Inline::C's compiler and linker? Answer: Yes. So, I wrote a little perl script that takes all of the command-line options and checks if they are valid nvcc options or flags. If so, it passes them along to nvcc unmodified, and if not it collects them and packages them into a -Xcompiler=... or -Xlinker=... argument. It then calls nvcc with the revised set of flags and options. And after a bit of fiddling, it actually worked. I've posted my code on github at this address: https://github.com/run4flat/perl_nvcc My plans are to create CPAN distributions ExtUtils::nvcc (which would install perl_nvcc) and Inline::CUDA (which would be derived from Inline::C). If I still have energy, I may work with Module::Build to include some sort of CUDA support. Eventually, I'd like to start working on the CUDA-Perl bindings, which I'll take elsewhere, probably to the PDL list. Does these sound like a good ideas? Is anybody else interested? David
