-
Clint,
your fix solves the nan problem in our code.
Many, many thanks for your quick help!
Christoph
PS:
about whether to call cblas_zgemm with &alpha or only alpha: to my own big surprise, it is valid C++ to put the address operator or leave it out for fixed-size arrays (the Stroustrup book on C++ states that an array name can(!) be used as a pointer to its first element). The same...
2009-07-08 21:11:52 UTC in Automatically Tuned Linear Algebra Soft.
-
Clint,
I'll try the patch within the next 12 hours.
Christoph.
2009-07-08 14:42:38 UTC in Automatically Tuned Linear Algebra Soft.
-
Clint,
thanks for the answer to my vague bug report. I think I can back my case now with some hard facts.
To clarify the unclear points:
- yes, the patch applied is the "Complex gemm sometimes reads C for beta=0"
- the problem is that zgemm sometimes yields a NaN-containing result, although the exact answer is well behaved. So far, I could give no condition when "sometimes" is, but I...
2009-07-07 21:43:21 UTC in Automatically Tuned Linear Algebra Soft.
-
Hi,
we are using ATLAS for our electronic structure code SPHInX, and are really happy about its performance, in particular for non-quadratic matrices. We presently use the most recent stable release, ATLAS 3.8.3
Yet, we have encountered very strange problems in some cases for zgemm (beta=0 (but with your patch), alpha=1, no trans, N=12, M=12, K=79204, lda,ldb,ldc with their minimal value)...
2009-07-06 14:44:21 UTC in Automatically Tuned Linear Algebra Soft.
-
Hi,
we are using ATLAS for our density-functional theory code SPHInX (electronic structure theory), and its just great! Its big advantage over other BLAS implementations is for "thin" matrices (M >> N), which occur quite commonly in our code.
Unfortunately, I have now encountered a problem for certain cases in zgemm when beta=0 (alpha=1, transA='N', transB='N', lda=M, ldb=K). We do not...
2009-04-09 15:32:17 UTC in Automatically Tuned Linear Algebra Soft.
-
Clint,
as everyone in the game, also I am reluctant to set the stacksize because I have no idea how large it should be either ;-)
The assert will definitely be an improvement, although it might be not obvious to our code's users what is going on if they ever run into that problem. Changing the ATLAS locally is one option, but we are hesitant because we do not know how maintainable this is...
2007-07-23 20:40:39 UTC in Automatically Tuned Linear Algebra Soft.
-
Clint,
thanks for your quick answer. After more problems over the weekend, I finally digged deep enough to see what happened, and found a pragmatic solution. I also have a suggestion for the ATLAS thread handling.
The problems arise when the stacksize and virtual memory limits are set to similar values (as is done by the Sun Grid engine). When the stacksize limit is set, it appears that...
2007-07-23 13:22:36 UTC in Automatically Tuned Linear Algebra Soft.
-
Sorry for bothering. When I limit the vmemoryuse to 2048 megabytes during the 3.7.34 compile, it works for me. That may not be logic, but at least does the job.
2007-07-20 18:36:53 UTC in Automatically Tuned Linear Algebra Soft.
-
Hi,
I have encountered a strange segmentation fault in the threaded ATLAS version 3.7.26 when I limit the virtual memory to a huge (2 GB), but finite value.
We have been using the ATLAS successfully for years in our density-functional theory code (a branch in theoretical physics/chemistry). After we realized how fast matrix algebra can be, we have been trying to shuffle as much...
2007-07-20 17:41:17 UTC in Automatically Tuned Linear Algebra Soft.
