This sounds very much like what I need right now. :) About running OpenCL on the CPU: I'm not really experienced with OpenCL yet. Do I need some additional extra flags so that it uses the CPU? What performance could I expect? Esp., can I expect better performance than the straightforward single-thread implementation of matrix-matrix multiplication?
I just tried VIENNACL_WITH_OPENCL and I get the exception invalid_buffer_size as written in my other mail. So I guess my matrix is too big for OpenCL? If OpenCL is using the CPU, how can that be? (Because without OpenCL, the code works fine.) Regards, Albert On Thu, Dec 12, 2013 at 3:47 PM, Karl Rupp <r...@iue.tuwien.ac.at> wrote: > Hey, > >> This seems like the good time to discuss something I was thinking about. >> Since OpenCL/CUDA is our main focus, wouldn't it be a good idea to >> provide the ability to link our routines against some CBLAS library? I >> think that most people are ready to compile/install for example OpenBlas >> for when they have no OpenCL/CUDA environment on their machine. I could >> add this to my todo-list. I myself use OpenBlas a lot, OpenMP is not the >> main focus of ViennaCL for now, and making it the main focus would be >> redundant with a lot of libraries which do an awesome job at this already. > > > Yes, this is definitely something we should provide in the future. The calls > into the backends are pretty clean already, so it is fairly simple to 'just > hook in' additional backends. > > Feel free to add it to your TODO list :-) We may also formulate this as a > GSoC project, because the entry barrier for this is pretty low. > > Best regards, > Karli > > > >> 2013/12/12 Karl Rupp <r...@iue.tuwien.ac.at >> <mailto:r...@iue.tuwien.ac.at>> >> >> >> Hi Albert, >> >> > The standard CPU backend doesn't seem to use multiple CPUs (e.g. >> for >> > matrix multiplication). >> >> This depends on the operation. Some are OpenMP-enabled, others aren't >> yet. So far we needed to focus on OpenCL and CUDA, with the CPU/OpenMP >> backends being used mostly for fallback purposes. If you're limited by >> matrix-matrix multiplications, you might want to give the OpenCL >> backend >> on the CPU a shot. >> >> >> > Is this planned? In some cases, I see some OpenMP pragmas in the >> code, >> > but not everywhere, esp. not on matrix-matrix products, which is my >> > most used operation. Has that just not been tested yet or is there >> a >> > reason? >> >> It has not been thoroughly tested and optimized yet. Our observation >> was >> that many people use vendor-libraries for big dense matrix-matrix >> multiplications anyway, so we better focused on the many 'other' >> linear >> algebra operations and the API. Adding OpenMP-pragmas to the current >> matrix-matrix multiplication would be easy, but we better improve >> cache >> utilization through blocking first. If there is user demand, we are >> happy to assign higher priority to this :-) >> >> Best regards, >> Karli >> >> >> >> ------------------------------------------------------------------------------ >> Rapidly troubleshoot problems before they affect your business. Most >> IT >> organizations don't have a clear picture of how application >> performance >> affects their revenue. With AppDynamics, you get 100% visibility >> into your >> Java,.NET, & PHP application. Start your 15-day FREE TRIAL of >> AppDynamics Pro! >> >> http://pubads.g.doubleclick.net/gampad/clk?id=84349831&iu=/4140/ostg.clktrk >> _______________________________________________ >> ViennaCL-devel mailing list >> ViennaCL-devel@lists.sourceforge.net >> <mailto:ViennaCL-devel@lists.sourceforge.net> >> https://lists.sourceforge.net/lists/listinfo/viennacl-devel >> >> > ------------------------------------------------------------------------------ Rapidly troubleshoot problems before they affect your business. Most IT organizations don't have a clear picture of how application performance affects their revenue. With AppDynamics, you get 100% visibility into your Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics Pro! http://pubads.g.doubleclick.net/gampad/clk?id=84349831&iu=/4140/ostg.clktrk _______________________________________________ ViennaCL-devel mailing list ViennaCL-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/viennacl-devel
