Re: 64-bit indexing for blas

[Kai Torben Ohlhus]

On 2/8/20 10:18 PM, JuanPi wrote:
> Hi all,
> 
> I followed the instructions in the manual to build 64-bit BLAS indexing
> https://octave.org/doc/interpreter/Compiling-Octave-with-64_002dbit-Indexing.html
> 
> I tried with the configure command there (directly modifying library
> and include search paths) but the configure logs shows
> 
> 64-bit array dims and indexing:       yes
> 64-bit BLAS array dims and indexing:  no
> 
> I opted for a more clear configure command
> 
> ./../../default/configure JAVA_HOME=/usr/lib/jvm/java-8-openjdk-amd64/
> --prefix=/opt/octave-dev64 CPPFLAGS=-I/usr/include/suitesparse
> --with-arpack-includedir=/opt/include/arpack
> --with-arpack-libdir=/opt/lib
> --with-lapack-libdir=/opt/lib/liblapack.a
> --with-blas=/opt/lib/libblas.a --with-qrupdate-libdir=/opt/lib
> 
> but still I can't get satisfactory config logs.
> 
> How can activate this?
> 


Juan,

I think you already got the state-of-art on this topic by Mike in

   https://savannah.gnu.org/bugs/?57771

Another interesting approach I follow is a combination of Spack and
Singularity which I recently discussed on a workshop [1].  Spack is a
nice Supercomputer PACKage manager, which allows to configure packages
rather than only installing some binary ones.  Mike recently got
maintainer of the GNU Octave package there.

Singularity can be considered as some more lightweight Docker with CUDA,
MPI, and InfiniBand support.  Older Singularity builds of mine not using
Spack are available here [2,3].

One drawback of this approach is, that building those images still takes
ages, like MXE.  But I work on a solution to reduce the repetitive build
time significantly.

Kai


[1] https://github.com/siko1056/slides_LSPANC_2020
[2] https://github.com/siko1056/GNU-Octave-64-Singularity
[3] https://cloud.sylabs.io/library/_container/5d663e4a7c33a87c6fa6a9c5