These are chat archives for elemental/chat

30th
Oct 2016
Ryan H. Lewis
@rhl-
Oct 30 2016 05:52
@poulson: I have submitted the Elemental package for review and inclusion in the mainline fedora repositories: https://bugzilla.redhat.com/show_bug.cgi?id=1389971
Jack Poulson
@poulson
Oct 30 2016 20:21
@CsatiZoltan @rhl- Qt5 is only needed if one wants to visualize matrices from C++
but the Python interface's viz capabilities are much better
Ryan H. Lewis
@rhl-
Oct 30 2016 20:30
@poulson makes sense. I think fedora will be upset if we don't change the spec to point to a release
Now that you have merged the new bulge chase stuff can we create a release candidate ?
Jack Poulson
@poulson
Oct 30 2016 20:33
I'm debugging some performance issues with the sequential Hessenberg QR algorithm on my Mac
I am seeing a 20x difference with a particular build and am trying to figure out if it was something pathological with the build
Ryan H. Lewis
@rhl-
Oct 30 2016 20:37
Makes sense. I think that a release candidate can have these issues
It's necessarily not production ready
Jack Poulson
@poulson
Oct 30 2016 20:43
I'm not against there being a release candidate, but I'm traveling until tomorrow night and have limited time until then
Jack Poulson
@poulson
Oct 30 2016 21:11
it seems there are issues with the current usage of OpenBLAS's imatcopy and omatcopy routines
the non-CBLAS interfaces seem not to be defined on OSX
Ryan H. Lewis
@rhl-
Oct 30 2016 21:12
I'm not surprised. I've used them before they don't work correctly all the time
Also in place transpose
Jack Poulson
@poulson
Oct 30 2016 21:13
localhost:build-gcc-brew-openblas poulson$ nm /usr/local/gcc-6-brew-openblas/lib/libopenblas.dylib | grep cimatcopy
000000000002f250 T _cblas_cimatcopy
000000000002c7f0 t _cimatcopy_
0000000000548f20 t _cimatcopy_k_cn
0000000000549250 t _cimatcopy_k_cnc
0000000000549050 t _cimatcopy_k_ct
0000000000549390 t _cimatcopy_k_ctc
0000000000548fd0 t _cimatcopy_k_rn
0000000000549310 t _cimatcopy_k_rnc
0000000000549150 t _cimatcopy_k_rt
00000000005494a0 t _cimatcopy_k_rtc
also xianyi/OpenBLAS#899
agreed
I ran into a link error using them
so I am disabling them
Ryan H. Lewis
@rhl-
Oct 30 2016 21:15
OS X has accelerate
It's much better
Jack Poulson
@poulson
Oct 30 2016 21:18
perhaps, accelerate isn't open source
so I can't debug or instrument it
Jack Poulson
@poulson
Oct 30 2016 21:42
I don't mind so much for BLAS, but it is frustrating for LAPACK
Ryan H. Lewis
@rhl-
Oct 30 2016 21:43
I'm assuming accelerate is not all that specialized for LAPACK
But I could be wrong
Maybe for QR
Copy is a very generic routine. Surprised it's not any faster than memmove or something
Jack Poulson
@poulson
Oct 30 2016 22:09
it's more about being able to time specific components for comparison reasons to see where the deviation happened
every extra bit seems to help when dealing with homebrew stacks, which I've found to be very unstable