Dears,
Is it possible to use the xeon-phi coprocessor using OPENCL via the GPU package?
regards
No, the GPU package is written with CUDA and only supports Nvidia GPUs. You can check out the KOKKOS package 
Anders
No, the GPU package is written with CUDA and only supports Nvidia GPUs. You
this is incorrect. you seem to be confusing the GPU package with the
(no longer available) USER-CUDA package.
the code in GPU library _can_ be compiled either for CUDA or OpenCL
and the OpenCL version can be used not only with Nvidia GPUs, but can
also run on AMD GPUs or it can run on CPUs. in fact, i have run it
with AMD GPUs quite successfully and we also tested it against
Intel's CPU OpenCL runtime.
however, this was quite a while ago, and the performance on the CPU
wasn't so good because the implementation is tuned for the hardware
layout of GPUs.
thus there is a chance to run the GPU package when compiled for OpenCL
on Xeon Phi, but i wouldn't expect as much acceleration as with
KOKKOS, since KOKKOS can adapt both computing/parallelization and data
access for different platforms.
axel.
Ahh sorry. Before answering I read in the docs for the GPU package:
Required hardware/software:
To use this package, you currently need to have an NVIDIA GPU and install the NVIDIA Cuda software on your system:
but I now see that there is more to it than this
Thanks,
Anders
Ahh sorry. Before answering I read in the docs for the GPU package:
Required hardware/software:
To use this package, you currently need to have an NVIDIA GPU and install
the NVIDIA Cuda software on your system:
well, in lib/gpu/README it says,
A Makefile for OpenCL compilation is provided, but support for OpenCL
use is not currently provided by the developers.
so you can say, that the GPU package has been designed to support both
CUDA and OpenCL (this is why it uses the lower level driver interface
and not the CUDA runtime interface). way back when i was testing it on
AMD GPUs, the AMD hardware had significant performance advantages on
individual GPUs, especially for double precision kernels, but the
interface of the driver into the Linux kernel was limited, so
multi-GPU machines and using multiple MPI tasks per GPU were not
scaling as well as with the nvidia toolchain.
axel.