Bill Allombert on Fri, 07 Nov 2003 13:01:29 +0100

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

Re: Experimenting with the Alpha build.

On Thu, Nov 06, 2003 at 12:25:32PM -0800, Phil Carmody wrote:
> --- Karim BELABAS <> wrote:
> I also recompiled GMP with -ffast-math too for the test. In theory GMP might
> perform some FP maths (FFT multiplication? or is it NTT?). I wasn't expecting
> massive changes, I just fancied experimenting. 

I don't thing GMP use FP on the alpha. FFT is implemented using modular
arithmetic. Also the bench never try to multiply large numbers so
FFT is not used.

> The only ones that had an improvement more than the 
> likely noise floor were:
> * Testing analyz        for gp-sta..TIME=445    for gp-dyn..TIME=399
> * Testing number        for gp-sta..TIME=226    for gp-dyn..TIME=231
> vs. the "slow" old
> * Testing analyz        for gp-sta..TIME=463    for gp-dyn..TIME=411
> * Testing number        for gp-sta..TIME=242    for gp-dyn..TIME=243
> analyz is a bit weird - I was expecting -dyn to by slower than -sta 
> across the board.

If you have a lot of cache, that is not unexpected: dynamic linking
resolve functions only when they are referenced and analyz use very few

The current bench basically run every GP functions on nearly trivial
input.  For tuning purpose, we need to write a special benchmark that
give a more accurate view of PARI performance on the platform.

Anyone interested in doing this ?