Re: LNUM patch for Lua 5.1.3?

[Asko Kauppi <askok@...>]

Available at:  http://luaforge.net/frs/?group_id=214

performance.txt is particularily interesting, and I would warmly  
welcome a few field tests from you, both on FPU-having and FPU-less  
platforms (I might move the sheet to be an actual spreadsheet in the  
future).

Current results (double+int32 mode only):

OS X PowerPC:	
	4.13 user   0.01 sys    (-1.0%)
	5.49 user   0.03 sys    (-15.9%)

OS X Intel:
	2.82 user   0.01 sys    (+5.5%)
	4.52 user   0.02 sys    (+2.9%)

Windows XP (Pentium M):
	3.374	(+3.3%)
	4.656	(-10.9%)

Ubuntu Linux 7.10 (Pentium M):
	2.48 user   0.00 sys    (-0.8%)
	5.61 user   0.00 sys    (-2.3%)

ARM Linux no FPU (NSLU2):
	3.56 user   0.10 sys    (-38.8%)
	12.79 user   0.19 sys    (-85.2%)

The first lines are fibonacci calculations, which are not very integer  
oriented (really, it's the recursive variant).
The second lines are essentially running 'etc/life.lua' without output  
(_very_ integer oriented).

NSLU2 loops less than others, so don't compare its absolute speed.

Comparisons are to an unpatched Lua 5.1.3. Minus is better; reducing  
speeds by 40-85% should be rather straggering, really.
More real world results (on any machines) are requested.

-asko


Bogdan Marinescu kirjoitti 12.3.2008 kello 9:38:

> Excellent news! Thaks, your work is highly appreciated (and I don't  
> mean just by me :) ).
>
> On Tue, Mar 11, 2008 at 8:59 PM, Asko Kauppi <askok@dnainternet.net>  
> wrote:
>
> Good timing. :)
>
> I am just finishing a revise of the patch for Lua 5.1.3, and I'm glad
> to announce that the patch is smaller and performs faster, as ever
> before.
>
> I will send it out today, or tomorrow, once the last remaining tests
> are performed.
>
> What has changed on the inside is that floating point results are no
> longer tested for integer values (which was taking time). Instead,
> integer realm and floating point realm live happily together, and both
> are seen as type "number" on the outside. This is the right way to go,
> since calculations are often either integer-only or floating point
> based (with less integer accuracy).
>
> Anyways, thanks for asking.
>
> -asko
>
>
> Bogdan Marinescu kirjoitti 11.3.2008 kello 13:41:
>
> > Hi,
> >
> > I was wondering if anyone has a functional LNUM patch for LUA 5.1.3.
> > I tried to apply the 5.1.2 patch, but it fails in some places, and
> > I"m not familiar with the code.