Even though these libraries are open source, it cannot be built without their proprietary, binary-only "AMD APP SDK" [1], only available for Linux and Windows. Bummer.<p>1. <a href="http://developer.amd.com/tools-and-sdks/heterogeneous-computing/amd-accelerated-parallel-processing-app-sdk/downloads/" rel="nofollow">http://developer.amd.com/tools-and-sdks/heterogeneous-comput...</a>
I noticed this today as I was searching around for some opencl image routines, and thought it was of general interest to the community. I really think this is an awesome thing, and the availability of a high performance open source BLAS that can be compiled to a wide array of OpenCL capable hardware is just great news.
I have recently started using ViennaCL (<a href="http://viennacl.sourceforge.net/" rel="nofollow">http://viennacl.sourceforge.net/</a>). It has an Boost uBLAS-like interface and has backends for OpenMP, OpenCL, CUDA and uBLAS.<p>It is in very active development and the community is very nice and helpful.<p>I think there also doesn't exists something similar, i.e. a lib which can easily do the calculations on both the CPU and the GPU (via OpenCL).
as cool as this is, it was dated august.<p>it's still (practically) impossible to do a proper LINPACK benchmark with open source tools on AMD GPUs, although this is a step in the right direction, and more importantly, a big blow to CUDA.