Python is darn fast (was: How fast is Python)

[Michele Simionato]

I posted this few weeks ago (remember the C Sharp thread?) but it went
unnoticed on the large mass of posts, so let me retry. Here I get Python+
Psyco twice as fast as optimized C, so I would like to now if something
is wrong on my old laptop and if anybody can reproduce my results.
Here are I my numbers for calling the error function a million times
(Python 2.3, Psyco 1.0, Red Hat Linux 7.3, Pentium II 366 MHz):

$ time p23 erf.py
real 0m0.614s
user 0m0.551s
sys 0m0.029s

This is twice as fast as optimized C:

$ gcc erf.c -lm -o3
$ time ./a.out
real 0m1.125s
user 0m1.086s
sys 0m0.006s

Here is the situation for pure Python

$time p23 erf.jy
real 0m25.761s
user 0m25.012s
sys 0m0.049s

and, just for fun, here is Jython performance:

$ time jython erf.jy
real 0m42.979s
user 0m41.430s
sys 0m0.361s

The source code follows (copied from Alex Martelli's post):

----------------------------------------------------------------------

$ cat erf.py
import math
import psyco
psyco.full()

def erfc(x):
exp = math.exp

p = 0.3275911
a1 = 0.254829592
a2 = -0.284496736
a3 = 1.421413741
a4 = -1.453152027
a5 = 1.061405429

t = 1.0 / (1.0 + p*x)
erfcx = ( (a1 + (a2 + (a3 +
(a4 + a5*t)*t)*t)*t)* t ) * exp(-x*x)
return erfcx

def main():
erg = 0.0

for i in xrange(1000000) :
erg += erfc(0.456)

if __name__ == '__main__':
main()

--------------------------------------------------------------------------

# python/jython version = same without "import psyco; psyco.full()"

--------------------------------------------------------------------------

$cat erf.c
#include <stdio.h>
#include <math.h>

double erfc( double x )
{
double p, a1, a2, a3, a4, a5;
double t, erfcx;

p = 0.3275911;
a1 = 0.254829592;
a2 = -0.284496736;
a3 = 1.421413741;
a4 = -1.453152027;
a5 = 1.061405429;

t = 1.0 / (1.0 + p*x);
erfcx = ( (a1 + (a2 + (a3 +
(a4 + a5*t)*t)*t)*t)* t ) * exp(-x*x);

return erfcx;
}

int main()
{
double erg=0.0;
int i;

for(i=0; i<1000000; i++)
{
erg = erg + erfc(0.456);
}

return 0;
}

Michele Simionato, Ph. D.
Mi************* *@libero.it
http://www.phyast.pitt.edu/~micheles
--- Currently looking for a job ---

[Simon Burton]

On Sun, 24 Aug 2003 00:31:15 +0100, John J. Lee wrote:

Irmen de Jong <[e-mail]> writes:

[e-mail] m wrote:
.... but Python+Psyco still wins :-)

So, the interesting part is: why?
John

My suspicion is that when psyco looks at erfc, it
finds that nothing changes and so replaces the
function call with the resulting number (am i right? it's the
same each time?). This is what a "specializi ng compiler"
would do, me thinks. So, try using a different number
with each call.

Simon.

[Van Gale]

Lawrence Oluyede wrote:

[e-mail] m wrote:

If you want to try different machines
then http://www.pixelbeat.org/scripts/gcccpuopt will give
you the appropriate machine specific gcc options to use.

Very cool script, thanks :) Anyway it didn't change so much with erf.c
erfCPU is compiled with the flags suggested by gcccpuopt script:

$ gcccpuopt
-march=athlon-xp -mfpmath=sse -msse -mmmx -m3dnow

You still need some -O optimization flags. The -m options just let gcc
generate some nice instructions specific to your Athlon CPU.

Also, I don't think that script is all that useful because at least some
(if not all) of those -m options are already implied by -march=athlon-xp
(I don't recall which ones off the top of my head but I'll find a
reference for anyone interested... you can also find out by looking at
the gcc command line option parsing code).

Anyone who wants some other good ideas for the best flags on their
machine check out ccbench:

http://www.rocklinux.net/packages/ccbench.html

The problem here of course is that not all applications behave like the
benchmarks :(

Van Gale

[Lawrence Oluyede]

Van Gale wrote:

You still need some -O optimization flags. The -m options just let gcc
generate some nice instructions specific to your Athlon CPU.

I didn't mention but I also used -O3 flag. I don't know why but on my
machine C code is faster than psyco code in this test

--
Lawrence "Rhymes" Oluyede
http://loluyede.blogspot.com
rh****@NOSPAMmy self.com

[Michele Simionato]

Van Gale <ne**@exultants .org> wrote in message news:<XK******* **********@news svr27.news.prod igy.com>...

Michele Simionato wrote:

I posted this few weeks ago (remember the C Sharp thread?) but it went
unnoticed on the large mass of posts, so let me retry. Here I get Python+
Psyco twice as fast as optimized C, so I would like to now if something
is wrong on my old laptop and if anybody can reproduce my results.
Here are I my numbers for calling the error function a million times
(Python 2.3, Psyco 1.0, Red Hat Linux 7.3, Pentium II 366 MHz):

$ gcc erf.c -lm -o3

Did you really use "-o3" instead of "-O3"? The lowercase -o3 will
produce object code file named "3" instead of doing optimization.

Yes, I used -O3, this was a misprint in the e-email. The compiler was
gcc 2.96.

Michele Simionato, Ph. D.
Mi************* *@libero.it
http://www.phyast.pitt.edu/~micheles
--- Currently looking for a job ---

[Michele Simionato]

I finally came to the conclusion that the exceeding good performance
of Psyco was due to the fact that the function was called a million
times with the *same* argument. Evidently Psyco is smart enough to
notice that. Changing the argument at each call
(erfc(0.456) -> i/1000000.0) slows down Python+Psyco at 1/4 of C speed.
Psyco improves Python performance by an order of magnitude, but still it
is not enough :-(

I was too optimistic!

Here I my numbers for Python 2.3, Psyco 1.0, Red Hat Linux 7.3,
Pentium II 366 MHz:

$ time p23 erf.py
real 0m3.245s
user 0m3.164s
sys 0m0.037s

This is more than four times slower than optimized C:

$ gcc erf.c -lm -O3
$ time ./a.out
real 0m0.742s
user 0m0.725s
sys 0m0.002s

Here is the situation for pure Python

$time p23 erf.jy
real 0m27.470s
user 0m27.162s
sys 0m0.023s

and, just for fun, here is Jython performance:

$ time jython erf.jy
real 0m44.395s
user 0m42.602s
sys 0m0.389s

----------------------------------------------------------------------

$ cat erf.py
import math
import psyco
psyco.full()

def erfc(x):
exp = math.exp

p = 0.3275911
a1 = 0.254829592
a2 = -0.284496736
a3 = 1.421413741
a4 = -1.453152027
a5 = 1.061405429

t = 1.0 / (1.0 + p*x)
erfcx = ( (a1 + (a2 + (a3 +
(a4 + a5*t)*t)*t)*t)* t ) * exp(-x*x)
return erfcx

def main():
erg = 0.0

for i in xrange(1000000) :
erg += erfc(i/1000000.0)

if __name__ == '__main__':
main()

--------------------------------------------------------------------------

# python/jython version = same without "import psyco; psyco.full()"

--------------------------------------------------------------------------

$cat erf.c
#include <stdio.h>
#include <math.h>

double erfc( double x )
{
double p, a1, a2, a3, a4, a5;
double t, erfcx;

p = 0.3275911;
a1 = 0.254829592;
a2 = -0.284496736;
a3 = 1.421413741;
a4 = -1.453152027;
a5 = 1.061405429;

t = 1.0 / (1.0 + p*x);
erfcx = ( (a1 + (a2 + (a3 +
(a4 + a5*t)*t)*t)*t)* t ) * exp(-x*x);

return erfcx;
}

int main()
{
double erg=0.0;
int i;

for(i=0; i<1000000; i++)
{
erg = erg + erfc(i/1000000.0);
}

return 0;
}

Michele Simionato, Ph. D.
Mi************* *@libero.it
http://www.phyast.pitt.edu/~micheles/
---- Currently looking for a job ----

[Tim Hochberg]

Michele Simionato wrote:

I finally came to the conclusion that the exceeding good performance
of Psyco was due to the fact that the function was called a million
times with the *same* argument. Evidently Psyco is smart enough to
notice that. Changing the argument at each call
(erfc(0.456) -> i/1000000.0) slows down Python+Psyco at 1/4 of C speed.
Psyco improves Python performance by an order of magnitude, but still it
is not enough :-(
This is not suprising. Last I checked, Psyco does not fully compile
floating point expressions. If, I rememeber correctly (though every time
try to delve too deeply into Psyco my brains start oozing out my ears),
there are three ways a in which a given chunk of code evaluated. At one
level, which I'll call #1, Psyco generates the machine code(*) for the
expression. At a second level, Psyco calls out to C helper functions,
but still works with unboxed values. At the third level, Psyco punts and
creates a Python object and hands things off to the interpreter.

Most integer functions operate at level #1, so they tend to be quite
fast. Most floating point operations operate at level #2, so they have a
certain amount of overhead, but are still much faster than unpsyco
(sane?) Python. I believe the reason for this is that x86 floating point
operations are very messy, so Armin punted...

(*) Armin is working on virtual machine implementation of Psyco, so it
should be available on non x86 machines soon.

FWIW,

-tim

I was too optimistic!

Here I my numbers for Python 2.3, Psyco 1.0, Red Hat Linux 7.3,
Pentium II 366 MHz:

$ time p23 erf.py
real 0m3.245s
user 0m3.164s
sys 0m0.037s

This is more than four times slower than optimized C:

$ gcc erf.c -lm -O3
$ time ./a.out
real 0m0.742s
user 0m0.725s
sys 0m0.002s

Here is the situation for pure Python

$time p23 erf.jy
real 0m27.470s
user 0m27.162s
sys 0m0.023s

and, just for fun, here is Jython performance:

$ time jython erf.jy
real 0m44.395s
user 0m42.602s
sys 0m0.389s

----------------------------------------------------------------------

$ cat erf.py
import math
import psyco
psyco.full()

def erfc(x):
exp = math.exp

p = 0.3275911
a1 = 0.254829592
a2 = -0.284496736
a3 = 1.421413741
a4 = -1.453152027
a5 = 1.061405429

t = 1.0 / (1.0 + p*x)
erfcx = ( (a1 + (a2 + (a3 +
(a4 + a5*t)*t)*t)*t)* t ) * exp(-x*x)
return erfcx

def main():
erg = 0.0

for i in xrange(1000000) :
erg += erfc(i/1000000.0)

if __name__ == '__main__':
main()

--------------------------------------------------------------------------

# python/jython version = same without "import psyco; psyco.full()"

--------------------------------------------------------------------------

$cat erf.c
#include <stdio.h>
#include <math.h>

double erfc( double x )
{
double p, a1, a2, a3, a4, a5;
double t, erfcx;

p = 0.3275911;
a1 = 0.254829592;
a2 = -0.284496736;
a3 = 1.421413741;
a4 = -1.453152027;
a5 = 1.061405429;

t = 1.0 / (1.0 + p*x);
erfcx = ( (a1 + (a2 + (a3 +
(a4 + a5*t)*t)*t)*t)* t ) * exp(-x*x);

return erfcx;
}

int main()
{
double erg=0.0;
int i;

for(i=0; i<1000000; i++)
{
erg = erg + erfc(i/1000000.0);
}

return 0;
}

Michele Simionato, Ph. D.
Mi************* *@libero.it
http://www.phyast.pitt.edu/~micheles/
---- Currently looking for a job ----

[dan]

mi**@pitt.edu (Michele Simionato) wrote in message
news:<22******* *************** ****@posting.go ogle.com>...

I finally came to the conclusion that the exceeding good performance
of Psyco was due to the fact that the function was called a million
times with the *same* argument. Evidently Psyco is smart enough to
notice that. Changing the argument at each call
(erfc(0.456) -> i/1000000.0) slows down Python+Psyco at 1/4 of C speed.
Psyco improves Python performance by an order of magnitude, but still it
is not enough :-(

It's plenty! A factor of 4 from optimized C, considering the newness
and limited resources behind psyco, is very encouraging, and good
enough for most tasks. Java JIT compilers are still around a factor
of 2 slower than C, and they've had at least 2 orders of magnitude
more whumpage.

This is a far cry from the factor of 10-30 I've been seeing with pure
python. For performance-critical code, this could be the difference
between hand-coding 5% versus 20% of your code.

Excellent news!!

[John J. Lee]

da*******@yahoo .com (dan) writes:

mi**@pitt.edu (Michele Simionato) wrote in message
news:<22******* *************** ****@posting.go ogle.com>... [...] This is a far cry from the factor of 10-30 I've been seeing with pure
python. For performance-critical code, this could be the difference
between hand-coding 5% versus 20% of your code.

Excellent news!!

If you care about this a lot, don't forget Pyrex.
John

[dan]

right, pyrex -- looked at that a while ago. Compiled Python with
C-style type declarations, right? Kinda like common lisp??? (I'm
stretching my memory cells now)

will review

jj*@pobox.com (John J. Lee) wrote in message news:<87******* *****@pobox.com >...

da*******@yahoo .com (dan) writes:

mi**@pitt.edu (Michele Simionato) wrote in message
news:<22******* *************** ****@posting.go ogle.com>...

[...]

This is a far cry from the factor of 10-30 I've been seeing with pure
python. For performance-critical code, this could be the difference
between hand-coding 5% versus 20% of your code.

Excellent news!!

If you care about this a lot, don't forget Pyrex.
John