Hi,
currently only the OpenCL backend is used in PETSc, as with CUDA there
is already the cusp bindings and for threading there is threadcomm.
Thus, you don't need to add any additional defines.
Best regards,
Karli
On 05/24/2013 10:04 PM, Mohammad Mirzadeh wrote:
sweet. I was in the middle of writing my own interfaces; i'll try those
as well. Do I still get to use the VIENNACL_WITH_XYZ macros?
On Fri, May 24, 2013 at 7:45 PM, Karl Rupp <[email protected]
<mailto:[email protected]>> wrote:
Hi Mohammad,
there is a first interface to ViennaCL in the next-branch already.
Configure with
--download-viennacl
--with-opencl-include=/path/__to/OpenCL/includes
--with-opencl-lib=/path/to/__libOpenCL.so
and then use
-vec_type viennacl
-mat_type aijviennacl
as runtime options.
As this resides in next, it is still work in progress. If you
encounter any problems during installation, please let us know.
Also, OpenCL typically shows larger latency than CUDA, so you should
have at least 100k unknowns to see any performance gain.
Best regards,
Karli
On 05/24/2013 09:08 PM, Mohammad Mirzadeh wrote:
Hi guys,
Speaking of interfaces, is there any plan to provide interfaces to
ViennaCL solvers?
Tnx
Mohammad