> Given the benefits of massively parallel processors on the GPU and the > push by NVIDIA amongst others, is there any plan to incorporate the use of > OpenCL and/or CUDA in deal.II. I just thought it'd be a nice issue to > discuss.
Like so many other things, this is mainly a question of manpower of which we have little. But there's also the question of what operations would be worthwhile to put on a GPU. GPUs are really really good at a small number of operations, namely operating on large (or many smaller) dense matrices and vectors. This is not typically what one does in unstructured/adaptive finite element computations with implicit solvers, however, unless you use high-order methods (say, of order 8 or 10). So I'm not entirely sure how much one would gain. There is one matrix class in deal.II, the ChunkSparseMatrix, that stores its element but storing whole blocks of the matrix at once. This is the class that would probably benefit from experiments with GPUs. It would certainly be interesting to see if anyone tried this class with, say, OpenCL or CUDA. W. ------------------------------------------------------------------------- Wolfgang Bangerth email: [email protected] www: http://www.math.tamu.edu/~bangerth/ _______________________________________________ dealii mailing list http://poisson.dealii.org/mailman/listinfo/dealii
