Received from Evgeny Lazutkin on Tue, Feb 25, 2014 at 11:58:10AM EST: > Many thanks for your support, Lev! > > It works and looks good! > > I have already mentioned about parallelization: your answer was that > it happens automatically. > To be more detailed: > > 0. Could you please explain this mechanism?
Since the library is proprietary, I don't know what CULA is doing internally. You may wish to inquire further on the CULA forums [1], although I suspect that the developers will not wish to discuss the specifics of how they implement the algorithms. > 1. How many blocks/threads has been used in my program? > 2. How to obtain this numbers in program? How to manipulate? I don't believe those parameters can be modified. > 3. Could you give any literature, where I can read about it? If you want to get more insight into how LAPACK-type algorithms may be implemented using GPUs, you may wish to look into the MAGMA library [2]; it contains a number of similar functions, but since it is open-source you can see what it is doing internally. The authors have also published a number of articles discussing their algorithms. > I have such a question because I am new one in such theme. When > using the SourceModul - I give the number of blocks and threads. So > it is not clear for me - how it works automatically? Is that depends > from the Matrix size an so on? Although pycuda does do automated block and thread number selection for some of the kernels that it generates to support features such as elementwise computations, reductions, etc., you will have to figure out how to select the appropriate number of blocks and threads when you launch your own kernel via the SourceModule class depending on your specific problem. > I think such information will be useful for all new users! I hope, > you can help me (or someone else:-) ) to understand! There are a range of online GPU programming courses you may wish to consider for further information: https://developer.nvidia.com/cuda-training > Best regards, [1] http://www.culatools.com/forums/ [2] http://icl.cs.utk.edu/magma/ -- Lev Givon Bionet Group http://www.columbia.edu/~lev/ http://lebedov.github.io/ _______________________________________________ PyCUDA mailing list [email protected] http://lists.tiker.net/listinfo/pycuda
