[petsc-dev] Bug in MatMatMultTranspose for sequantial AIJ matrices from petsc-dev?

[Mark F. Adams]

On Nov 9, 2011, at 11:02 AM, Barry Smith wrote:

> 
> On Nov 9, 2011, at 7:58 AM, Mark F. Adams wrote:
> 
>> FYI: I appear to be getting not great flop rates out of these methods on my Mac:
> 
>   Known issue. Do you have alternative algorithms that would crank it up? This is something we are actively working on.
> 

I don't know what you're doing now, Hong mentioned that you had some good ideas for optimizing the code from what we look at together when I was at Argonne.

One generic idea that has come to mind since we last talked, for RAP, is folding the two parts: 1) T = A*P, 2) RAP = R*T, together.  I have not looked at this in detail but perhaps instead of computing the whole "T" here, compute parts (a row, an element ...) call each one T_i, and use them right away, RAP += R*T_i, in one big loop and throw them away.  This might improve cache performance because T will be high in cache.

It sounds like the serial code is stable now.  I will dive into it this week, finally figure out what you are doing exactly, and see if I can come up with any ideas.  This is a hard problem and I'm not even sure what fast is here.

Mark

>   Barry
> 
>> 
>> MatMatMult             2 1.0 1.1382e+00 1.0 8.48e+07 1.0 0.0e+00 0.0e+00 4.0e+00  4  1  0  0  2  12  8  0  0  2    75
>> MatPtAPNumeric         2 1.0 4.3557e+00 1.0 7.82e+08 1.0 0.0e+00 0.0e+00 0.0e+00 14 11  0  0  0  45 74  0  0  0   180
>> MatTrnMatMult          2 1.0 6.0777e-01 1.0 3.31e+07 1.0 0.0e+00 0.0e+00 8.0e+00  2  0  0  0  4   6  3  0  0  4    55
>> 
>> KSPSolve               1 1.0 2.9164e+00 1.0 1.57e+09 1.0 0.0e+00 0.0e+00 0.0e+00 10 21  0  0  0 100100  0  0  0   538
>> 
>> Mark
>> 
>> On Nov 9, 2011, at 10:47 AM, Hong Zhang wrote:
>> 
>>>> Hong, please update src/docs/website/documentation/changes/dev.html when you
>>>> make API changes.
>>> Done and pushed to petsc-dev.
>>> 
>>> Hong
>>> 
>> 
> 
>