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



>Can't reproduce this one.  Can you give me the command that fails?
>I've included my runs below.

I used the stuff to build a full gemm.  Then, pretty much any M, N, or K 
that was not a multiple of NB failed.  In a few quick tests, I could not
get the cleanup to fail using mmutstcase, so it'll take some tracking
down.  I didn't figure it was worth my spending a bunch of time on, 
'cause you had said this was just an initial version; When you have 
some code closer to the finished product, that seems the best time to
really track it down . . .

Speaking of such things, I guess you noticed that camm_arith.h is now
almost 16MB in size?  The one we released was already 2MB, which too
large already, so we need to find a fix for this.  What does this file
do?  If I had to guess, it seems like it's something you use for indexing?
If so, is there a reason you don't use the NB1-8 that ATLAS provides?
I've found most places that to get all indices, NB2+3 can be used with
little loss of performance, rather than having to have an explicit macro
NB2_3, or some such . . .