Repository navigation
Support BLAS-like extensions #51
Description
BLAS-like extensions,
e.g. mkl_?imatcopy and mkl_?omatcopy in Intel MKL
These functions support copy and transpose matrix.
{{ message }}
BLAS-like extensions,
e.g. mkl_?imatcopy and mkl_?omatcopy in Intel MKL
These functions support copy and transpose matrix.
Activity
ViralBShah commentedon Dec 15, 2011
These would be great to have. Also, would be nice to have an in-place transpose. Programming languages such as Julia (http://www.julialang.org) can immensely benefit from such things, if they are faster than the obvious implementations.
xianyi commentedon Dec 15, 2011
Hi,
Thank you for the suggestion.
Xianyi
certik commentedon Sep 26, 2013
Indeed, this would be fantastic to have.
certik commentedon Jun 4, 2014
@wernsaar could you point to a commit that fixed this issue?
wernsaar commentedon Jun 4, 2014
On 04.06.2014 18:55, Ondřej Čertík wrote:
I reopened this issue as a feature request, but at this time
I don't know when we have the time, to program this.
Best regards
Werner
wernsaar commentedon Jun 9, 2014
Hi,
I wrote the blas extension domatcopy ( like mkl_domatcopy).
Does anyone has some programs for testing.
Regards
Werner
Fulguritus commentedon Jun 9, 2014
@wernsaar
I have done some analysis about that in the past. I came to the conclusion that the
transposeintrinsic outperforms all tested compilers for matrix sizes up to 100k-150k elements and when using one thread only.Blocking and loop unrolling was faster than a simple nested loop, but letting the compiler do that was usually much faster.
I hope this helps...
Regards,
Alex
wernsaar commentedon Jun 9, 2014
On 09.06.2014 17:47, Alexander Vogt wrote:
thank you.
But first I need some programs to check the results.
If anyone still has such programs, I don't have to write
such porgrams.
Best regards
Werner
wernsaar commentedon Jun 10, 2014
The blas extensions (s,d,c,z)omatcopy and (s,d,c,z)imatcopy has benn added to my fork of the repository and will be pushed to the openblas develop repository. I close now this feature request.
andreasnoack commentedon Sep 6, 2016
MKL's
xgemmtis also worth considering. ComputingA*diag(x)*A'is really common.Merge pull request #51 from xianyi/develop