Hi guys,
today I finally completed the update of the bindings to ViennaCL 1.5.1,
which brings, among some performance improvements, also the missing
implementation of distributed MPI/OpenCL-based matrix-vector products
using the compressed CSR format. As a consequence, PETSc can now be used
with any OpenCL-capable device (most notably AMD and NVIDIA GPUs, but
also Intel's MIC architecture) in a full MPI environment. Since there is
no need to hide our OpenCL capabilities, particularly as it is far less
a burden on the build system, can be used with any host compiler, and is
not restricted to a single vendor, I updated the webpage accordingly here:
https://bitbucket.org/petsc/petsc/commits/c738b73485b3ba70a7722f0a4f5092a7fe93ee87
Please let me know if there are any objections, otherwise I will merge
this to master soon (i.e. once all nightly-tests pass).
There is of course still plenty of work and optimizations ahead,
particularly the vector scatters (both CUDA/CUSP and OpenCL/ViennaCL).
@Jed: Can you please update the repository description here:
https://bitbucket.org/petsc/petsc/ by replacing "NVIDIA GPUs"
accordingly, e.g. "GPUs through CUDA and OpenCL" (or any other more
elegant formulation)?
Thanks and best regards,
Karli
