System blas/lapack

[fommil]

for OS X this would be veclib, for Linux this would be /usr/lib/libblas.so etc (allowing ATLAS to be swapped in).

The biggest stumbling block here is working out if all the routines are implemented by the system library or not. We might have to pick the lowest common denominator for ease of packaging.

[fommil]

maven native plugin seems to ignore the excludes tag. sigh

[fommil]

http://jira.codehaus.org/browse/MOJO-1967

[fommil]

I worked around it AND got performance charts for OS X veclib as the backend :-D DGEMM results are impressive but DGETRI is a bit disappointing compared to the pure C version.

[fommil]

yay! it was just a regression in the test parameters

[fommil]

@dlwh snapshots are now available with veclib support in OS X (Linux to come soon, I'm not sure about Windows... will have to investigate further) as discussed in scalanlp/breeze#85

[dlwh]

great thanks!

On Wed, Sep 4, 2013 at 2:10 PM, Sam Halliday [email protected]:

@dlwh https://github.com/dlwh snapshots are now available with veclib
support in OS X (Linux to come soon, I'm not sure about Windows... will
have to investigate further) as discussed in scalanlp/breeze#85scalanlp/breeze#85

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/27#issuecomment-23824263
.

[fommil]

@dlwh this is insane, I'm seeing x10 performance for larger DBGEMM and DGETRI etc vs even the native ref

[fommil]

now windows builds available... but due to limitations in the availability of system BLAS/LAPACK, I had to link against libopenblas.

[fommil]

snapshots available on sonatype

[fommil]

reopening to get proper windows linking against a libblas3.dll and liblapack3.dll and cuBLAS support documented for OS X.

[fommil]

I have simple instructions now to build our own CBLAS/LAPACK to link against.

[fommil]

getting the following working are part of this ticket:

https://developer.nvidia.com/cublas

https://github.com/clMathLibraries/clBLAS

[fommil]

tested builds on osx, linux 32/64, win 32/64 and Raspberry Pi :-D

Unfortunately, the OpenCL BLAS implementations are not quite there yet... but in theory end users can set it up long after we release (all they need to do is to implement CBLAS).