Hi Gexarcha, I'm the author of the OpenCL package for Julia. Although things have stalled lately do to other commitments, I think we are slowly building a pretty good story at least on the OpenCL side. I know that the library has already been used sucesfully in a handful of reasearch projects. Right now I'm working on finishing up Image and OpenGL support, which are last missing features in the OpenCL bindings. I've implemented wrappers for CLFFT and another contributor has preliminarily CLBLAS bindings. As I wrote in the GSOC project proposal, I think what we are missing is an GPU Array datatype similar to what you get with PyCuda / PyOpenCL. I hope to work on that in the future but impementing rich mulit-dim array support is a lot of work.
I've experimented with compiling Julia's lowered AST down to OpenCL. It works, but larger kernels quickly cause register pressure leading to a drop in performance over what you would get if you wrote the kernel by hand. I'm eagerly waiting to hear any results from the ptx backend, as compiling down to llvm ir / ptx would seem more efficient. However, the authors of Numba's (python) cuda jit compiler have stated that you ususally get a performance loss of 2x+ over hand written kernels with their library, so there seems to be a lot of room for improvment in this area. Jake I am trying to figure out what kind of options do I have for utilizing the GPU for high performance computing in Julia. I have seen some old threads<https://groups.google.com/forum/#%21msg/julia-dev/YeF2mu2Ii-k/zcr78Ml_6IIJ>discussing an nvptx backend and the CUDA.jl package. I would like to know if there is any other effort out there dealing with GPU computing. Also, does anybody know if there is a plan of integrating GPU support in the language? Thanks! On Friday, April 11, 2014 10:47:41 AM UTC-4, andrew cooke wrote: > > > it's included at http://julialang.org/gsoc/2014/ > > On Friday, 11 April 2014 05:09:11 UTC-3, gexarcha wrote: >> >> Hi, >> >> I am trying to figure out what kind of options do I have for utilizing >> the GPU for high performance computing in Julia. >> I have seen some old >> threads<https://groups.google.com/forum/#!msg/julia-dev/YeF2mu2Ii-k/zcr78Ml_6IIJ>discussing >> an nvptx backend and the CUDA.jl package. I would like to know >> if there is any other effort out there dealing with GPU computing. Also, >> does anybody know if there is a plan of integrating GPU support in the >> language? >> Thanks! >> >> - Georgios >> >
