[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: ATLAS code freeze


>Hi Clint!  Almost done with the L2.  Found a few possible issues with
>the ger tester.
>1)  Don't you need the following patch to r1test.c?

Indeed, yes.  Thanks for the bug report.  The fix I applied to each of those
lines was to replace aincY with (aincY SHIFT).

>2)  Without this, the tester runs over the very large numbers you
>    filled in the gaps, presumably to spot errors more readily.  It
>    revealed, however, (I think), an issue with the 'epsilon' used to
>    determine whether the test routine and the reference agree.  Here
>    is a small program which gives a small discrepancy between the
>    correct SSE and normal fpu, but which is (again I think) within
>    normal floating point tolerance:

Well, the error test is really more like a sanity test; it should probably
do a residual based on norms.  However, it looks like I screwed up the sanity
test even.  Looks like to me that emul should be 16 for complex.  Does this
fix your prob, or are you having real fail as well?