@marker68 , OpenBLAS didn't support GPU so far.
@wernsaar, Werner has optimized some kernels for GPU by OpenCL. We plan to release GPU optimization codes in the near future.
When you will release GPU like cuda or through OpenCL?
@guangwei , there are some experiment codes at https://github.com/xianyi/clOpenBLAS
@xianyi thank you for your information
Any updates on this? It would be awesome to have an OpenCL implementation of OpenBLAS.
By the looks of it, the last work on clOpenBLAS happened in April 2015 (and wernsaar its primary developer was last seen on github in January 2017, even xianyi himself appears to be absorbed in his own business lately). Have you looked at clBLAS and/or clBLAST ?
@martin-frbg Sad to hear that the main developers have abandoned their work.
I'm actually aware of several OpenCL implementations of BLAS API: CLBlast, clBLAS, clMAGMA, ArrayFire and ViennaCL, to mention a few. Although they are great, OpenBLAS seems to be the most prominent BLAS implementation of the API, used by a great majority of other projects. If we should have true cross platform and vendor natural GPGPU accelerated BLAS, OpenBLAS is the best one to invest in.
P.S. Unfortunately CUDA is eating away the whole GPGPU world and there will be a monopoly if the community keeps ignoring the issue.
You can certainly fork and continue the project (I do not have collaborator status there, so would not be able to merge PRs in the original clOpenBLAS repository). I think it is possible that the block algorithms originally introduced by K.Goto to make the most use of (x86) CPU caches are less well suited for GPU hardware. (This seems to be a major performance obstacle in the RISC V port of OpenBLAS at least, I know very little about OpenCL coding)
Most helpful comment
You can certainly fork and continue the project (I do not have collaborator status there, so would not be able to merge PRs in the original clOpenBLAS repository). I think it is possible that the block algorithms originally introduced by K.Goto to make the most use of (x86) CPU caches are less well suited for GPU hardware. (This seems to be a major performance obstacle in the RISC V port of OpenBLAS at least, I know very little about OpenCL coding)