Optimizing Athlon/Pentium3 Code

[Joe O]

Prime95 said:

Actually, I'm pretty sure I can get another 10% for Athlons and Pentium 3 computers by employing one of the tricks I learned in the P4 optimizations. Unfortunately, it requires a major rewrite of the x86 FFT code. This is not something I relish doing!

Sadly, I also know of a way to improve the proth primes search (seventeen-or-bust) by a substantial amount, but again it requires a major coding effort.

I was wondering if you had started on either of these. Do you need any help?

[Prime95]

I started and stopped the effort. My P3 testbed (actually Celeron) died and I've also been kind of busy.

If you look around you'll see where I asked for some P3 benchmarks on two FFT sizes. The results were mixed - less than I had hoped for. The Athlon improvement was more than 10% if I recall correctly.

I've not started the Proth speedup, but the two FFT sizes I did change were coded in such a way that the Proth speedup can be implemented too.

What would be required to complete the effort? All x87 FFT sizes would need to be recoded too. The auxillary add and subtract routines need to be rewritten and tested. The Proth mod routine needs to be rewritten, etc. Even the P4 proth mod routine needs rewriting.

The downside is the new version seems to be slower for P2 and older CPUs. Maintaining both code paths is not reasonable, so these slower CPUs would be stuck with version 23 or suffer with a slower newer version.

[outlnder]

What about an AMD client and an Intel client??

[PageFault]

Aren't PII's by default put on trial factoring anyways, i.e., doublecheck cutoff at 500 MHz minimum? Besides, the enthusiaist who would still run archaic hardware would likely know which client to use anyways ... I doubt these machines contribute much to LL testing in the first place ... I don't want to think of my 350 crunching a 20000000M exponent ...

Athlons however do have a significant impact on production ... despite our mass adoption of Northwood technology, TPR still has hundreds of AMD machines ... they need the optimization ...

Originally posted by Prime95

The downside is the new version seems to be slower for P2 and older CPUs. Maintaining both code paths is not reasonable, so these slower CPUs would be stuck with version 23 or suffer with a slower newer version.

[Joe O]

Originally posted by Prime95
...
What would be required to complete the effort? All x87 FFT sizes would need to be recoded too.
...
The downside is the new version seems to be slower for P2 and older CPUs. Maintaining both code paths is not reasonable, so these slower CPUs would be stuck with version 23 or suffer with a slower newer version.

How about starting with the larger FFT sizes? Especially the new SSE2 one for 77M. I don't know how many P2s are in use on LL/DC, but presumably they are working on the smaller FFT sizes. If not, then they will have to stay with the current client. This would give the project the boost in the larger FFT ranges. 10%+ is nothing to sneeze at!

[cheesehead]

Originally posted by PageFault
Aren't PII's by default put on trial factoring anyways, i.e., doublecheck cutoff at 500 MHz minimum? Besides, the enthusiaist who would still run archaic hardware would likely know which client to use anyways ... I doubt these machines contribute much to LL testing in the first place

But trial is not the only method of factoring, and LL testing is not the only application of FFT multiplication.

P-1 factoring is included in PrimeNet factoring assignments, and it uses FFT multiplication.

So the effect of an FFT change on pre-P3 models does need to be considered.

[cheesehead]

Originally posted by cheesehead
P-1 factoring is included in PrimeNet factoring assignments

... or, rather, I thought that a PrimeNet factoring assignment was for both trial and P-1 factoring of the assigned exponent. But I find that the worktodo line is "Factor=...", which is only trial factoring.

Sorry, PageFault ... I mis-remembered, and I should've checked before posting.