Re: Help needed to build libhmsbeagle using OpenCL

[Ghislain Vaillant <ghisvail@gmail.com>]

I am sorry Andreas, I can't help on that one.

Ghis

On 23/10/15 14:17, Andreas Tille wrote:
> Hi,
>
> I have a totally different question about OpenCL:  Assume I have
> successfully built it locally.  When checking the availability of
> libhmsbeagle features beast-mcmc (which uses libhmsbeagle) provides a
> check which in my case looks like:
>
>
> $ java -jar /usr/share/beast-mcmc/beast.jar -beagle -beagle_info ...
> Using BEAGLE library v2.1.2 for accelerated, parallel likelihood evaluation 2009-2013, BEAGLE Working Group - http://beagle-lib.googlecode.com/
> Citation: Ayres et al (2012) Systematic Biology 61: 170-173 | doi:10.1093/sysbio/syr100
>
> BEAGLE resources available:
> 0 : CPU
>      Flags: PRECISION_SINGLE PRECISION_DOUBLE COMPUTATION_SYNCH EIGEN_REAL EIGEN_COMPLEX SCALING_MANUAL SCALING_AUTO SCALING_ALWAYS SCALERS_RAW SCALERS_LOG VECTOR_SSE VECTOR_NONE THREADING_NONE PROCESSOR_CPU FRAMEWORK_CPU
>
> 1 : Intel(R) HD Graphics IvyBridge GT1 (OpenCL 1.2 beignet 1.0.3)
>      Global memory (MB): 2048
>      Clock speed (Ghz): 1,00
>      Number of multiprocessors: 6
>      Flags: PRECISION_SINGLE COMPUTATION_SYNCH EIGEN_REAL EIGEN_COMPLEX SCALING_MANUAL SCALING_AUTO SCALING_ALWAYS SCALERS_RAW SCALERS_LOG VECTOR_NONE THREADING_NONE PROCESSOR_GPU FRAMEWORK_OPENCL
>
>
> So the application sees the GPU via libhmsbeagle.
>
> However, a mac user told me that it looks like:
>
> > BEAGLE resources available:
> >
> > 0 : CPU
> >
> >      Flags: PRECISION_SINGLE PRECISION_DOUBLE COMPUTATION_SYNCH
> > EIGEN_REAL EIGEN_COMPLEX SCALING_MANUAL SCALING_AUTO SCALING_ALWAYS
> > SCALERS_RAW SCALERS_LOG VECTOR_SSE VECTOR_NONE THREADING_NONE
> > PROCESSOR_CPU FRAMEWORK_CPU
> >
> > 1 : Intel(R) Core(TM) i5-4258U CPU @ 2.40GHz (OpenCL 1.2 )
> >
> >      Global memory (MB): 8192
> >
> >      Clock speed (Ghz): 2,40
> >
> >      Number of multiprocessors: 4
> >
> >      Flags: PRECISION_SINGLE PRECISION_DOUBLE COMPUTATION_SYNCH
> > EIGEN_REAL EIGEN_COMPLEX SCALING_MANUAL SCALING_AUTO SCALING_ALWAYS
> > SCALERS_RAW SCALERS_LOG VECTOR_NONE THREADING_NONE PROCESSOR_CPU
> > FRAMEWORK_OPENCL
>
> !!!!
>
> > 2 : Iris (OpenCL 1.2 )
> >
> >      Global memory (MB): 1536
> >
> >      Clock speed (Ghz): 1,20
> >
> >      Number of multiprocessors: 280
> >
> >      Flags: PRECISION_SINGLE COMPUTATION_SYNCH EIGEN_REAL EIGEN_COMPLEX
> > SCALING_MANUAL SCALING_AUTO SCALING_ALWAYS SCALERS_RAW SCALERS_LOG
> > VECTOR_NONE THREADING_NONE PROCESSOR_GPU FRAMEWORK_OPENCL
>
> So the way more important point here is the multiprocessor feature which
> is actually wanted for beast-mcmc.  Any idea how to activate this?  I'm
> afraid this is rather a question for upstream but may be somebody here
> has some experience.
>
> Kind regards
>
>        Andreas.
>
> On Wed, Oct 21, 2015 at 03:38:12PM +0100, Ghislain Vaillant wrote:
> > > OpenCL error: CL_DEVICE_NOT_FOUND from file <GPUInterfaceOpenCL.cpp>, line 118.
> > > FAIL tinytest (exit status: 255)
> >
> > The error comes from their GPU interface abstraction. At that point, it may
> > be worth asking whether upstream have tested it with POCL. Chances are that
> > they did test with Beignet (since it worked on your machine?) but not on
> > POCL. It could also well be that POCL v0.10 is insufficient.
> >
> > Ghis