On 08/10/17 12:53, Rebecca N. Palmer wrote:
On 30/09/17 15:38, Rebecca N. Palmer wrote:
In Alioth (e49f225, not tagged as my other package's sponsor prefers
not doing that until upload).
Note that my hardware can't run the GPU tests due to #877316, so it
might be a good idea for someone else to do so before upload.
Has anyone tried this, and if not, should I try asking elsewhere (e.g.
the d-science list)?
On Nvidia systems, building theano includes the GPU tests if the GPU
is accessible (it may not be in chroots) and the required dependencies
are installed: python-pygpu python3-pygpu nvidia-cuda-dev libcuda1
nvidia-driver
(The -dev version is required because libgpuarray dlopen()s libraries
by their un-numbered names)
IMO, we should just focus on the CPU tests and hope for the best for the
GPU ones. GPU testing could be done within debci, but I am not even sure
how we could ensure a test machine is provisioned with the necessary
hardware for it (nvidia).
If you want something that doesn't take 5 hours, it's also possible to
run just the libgpuarray tests with nosetests3 -v <package
root>/theano/gpuarray/tests/ , but this will miss some GPU tests
elsewhere in the source tree.
Theano can also use OpenCL, but this is still in development and must
be explicitly enabled with THEANO_FLAGS='init_gpu_device=opencl0:0'
(dependencies: python-pygpu python3-pygpu libclblas-dev
ocl-icd-opencl-dev mesa-opencl-icd).
Considering upstream confirmed active development will stop next year, I
would not hold on the prospect of full OpenCL support.
Cheers,
Ghis
