Path: utzoo!utgpu!news-server.csri.toronto.edu!bonnie.concordia.ca!ccu.umanitoba.ca!herald.usask.ca!alberta!ubc-cs!uw-beaver!milton!dali.cs.montana.edu!uakari.primate.wisc.edu!crdgw1!uunet!mcsun!hp4nl!charon!dik
From: dik@cwi.nl (Dik T. Winter)
Newsgroups: comp.lang.c
Subject: Re: Execution time bottleneck: How to speed up execution?
Message-ID: <2942@charon.cwi.nl>
Date: 14 Feb 91 22:00:14 GMT
References: <24587:Feb1411:32:5391@kramden.acf.nyu.edu> <2940@charon.cwi.nl> <26862:Feb1416:46:4391@kramden.acf.nyu.edu>
Sender: news@cwi.nl
Organization: CWI, Amsterdam
Lines: 48

In article <26862:Feb1416:46:4391@kramden.acf.nyu.edu> brnstnd@kramden.acf.nyu.edu (Dan Bernstein) writes:
 > Dik, the optimizations I posted in a previous article are responsible
 > for between a 4% and a 50% speedup, depending on your machine, your
 > compiler, etc.
The optimizations you gave in your first article gave between 4 and 10% speedup,
not to 50%.  To break down the different speedups on a Sun SLC, comparing
O'Keefe's variant with your 5 versions (done with n=1000, 5 times; time in
seconds):
version	time	variant
OK	55.93	O'Keefe's variant
1	55.15	1.4% speedup (using tmp for xi-xj, including c)
2	53.37	4.6% speedup (looping down, not up)
3	53.30	4.7% speedup (register variable for a[j])
4	53.87	3.7% speedup (pointer for a+i and y+j)
5	53.82	3.8% speedup (vectorizable code)
And yes, I consider this micro optimization.

 > As I said the first time, the 10% is the speedup you get on a Convex
 > with the standard math library exp() when you apply the ``ludicrous''
 > optimizations I pointed out. It is not due to vectorization.
Might be.  Does the Convex vectorize all five variants?
 > 
 > Yes, and I could equally well have said ``buy a Cray.'' If the original
 > poster didn't have a Cray this would result in ``large improvements.''
Sure.  Times on a Cray Y/MP:
version	time	remarks
OK	0.3458	original
1	0.3282	5.1% speedup
2	0.3292	4.8% speedup
3	0.3312	4.2% speedup
4	0.3717	7.5% slowdown
5	0.3469	0.3% slowdown
Although the compiler needed a bit of persuasion to vectorize the original
and versions 1 and 2.  So also here: micro optimization.  (Calling version
4 and 5 optimized is even stressing the meaning of the term a bit!)

 > Similarly, the code becomes quite a lot faster if the poster uses a fast
 > exp()---but do you really think that ``use a fast exp()'' is any more
 > helpful than ``buy a Cray''? No. Neither one answers the question.
But 'use fast exp()' is much more cheaper than 'buy a Cray'!  And it can
be handled by an adqeuate programmer.

 > Furthermore, if the poster *does* have a fast exp() running on his Cray,
 > the optimizations I posted will give an even better speedup.
I doubt it very much.
--
dik t. winter, cwi, amsterdam, nederland
dik@cwi.nl