Re: problem with Real_Matrix*Real_Matrix

[Ken Thomas <kst@ecs.soton.ac.uk>]

On Wednesday, 19 September 2012 09:17:17 UTC+1, Simon Wright  wrote:
> reinkor <reinkor@gmail.com> writes:
> 
> 
> 
> > On Tuesday, September 18, 2012 10:33:42 PM UTC+2, AdaMagica wrote:
> 
> >> It's a rather scary bug, isn't it. In an operation like this!
> 
> >> 
> 
> >> Were AdaCore sleeping?
> 
> >
> 
> > But they have some formalized QA? N (>>1) tests to pass through at
> 
> > every release?
> 
> 
> 
> FSF GCC has the ACATS tests (not the latest version, though) + some
> 
> GNAT-specific tests (there are 947 source files, some tests require as
> 
> many as 3 sources).
> 
> 
> 
> AdaCore tell us that they have, in addition, test cases that they don't
> 
> make public (a lot of them are from bugs reported by customers and may
> 
> contain sensitive material); and that all these tests are run on a
> 
> regular basis.
> 
> 
> 
> I would have expected there to be a set of tests about numerics, and I'd
> 
> have expected these tests to be re-run after the refactoring that
> 
> happened at the GCC 4.6-4.7 change.

This is indeed a pity. I have run some performance tests on a Windows machine (XP Intel 2 CPU 2GB ram). On multiplying two 500 by 500 matrices (Long_Float), a rate of 966 MFlops/Sec was achieved. This is good.