Hi Philippe,

 > I've been obtaining recently significant performance improvement out of
> the kernel generator, which should bring ViennaCL 1.6 extremely close
> (>95%) to CuBLAS (on NVidia hardware) and clAmdBlas (on AMD hardware)
> for BLAS1/Dense BLAS2/Dense BLAS3.

Excellent, great news! :-)

> There are a lot of drawbacks to
> maintaining a blas-linking functionnality, which may not be worth it if
> we can improve the performance of ViennaCL to such an extent. I'd be in
> favor of dropping this idea away from the ViennaCL 1.6 roadmap. This
> would also free us some time to focus on distributed matrices/vectors.
>
> What do you think about it?

Since (currently) I'm the one who is doing all the release-testing and 
build-system-polishing, it would certainly save me a bunch of time not 
having to test the different BLAS libraries on the various operating 
systems. On the other hand, we cannot always assume that OpenCL is
  a) available on the target machine, and
  b) efficient (think of the poor CPU performance data we got recently).
There was a recent discussion at the uBLAS mailinglist about which 
direction to go, and one of the top items was the support for external 
BLAS libraries.

Overall, I still think that this is a fruitful direction to go, because 
it forces us to have clean interfaces at the lower level, and it also 
helps us in quickly producing performance comparisons. On the other 
hand, I don't see a necessity for providing the feature with 1.6.0 
already, if we can instead provide great built-in performance 
improvements such as faster kernels and a full scheduler<->generator 
integration.

Best regards,
Karli


------------------------------------------------------------------------------
Managing the Performance of Cloud-Based Applications
Take advantage of what the Cloud has to offer - Avoid Common Pitfalls.
Read the Whitepaper.
http://pubads.g.doubleclick.net/gampad/clk?id=121054471&iu=/4140/ostg.clktrk
_______________________________________________
ViennaCL-devel mailing list
ViennaCL-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/viennacl-devel

Reply via email to