> Also note that jruby has a new version released now.

Unfortunatelly, i didn’t find the p129 for linux. Anyway, it didn’t seem to much stable compared newer versions so i don’t use it any more.

I’ve just noticed has been released jruby 1.5, thanks! I will try it as soon as possible.

I only used –server, however also adding –fast, there’s no difference on my system.
I have ubuntu in that notebook, nextly i could do another bench though i think on linux there will be less difference. Usually, on linux you make the binary by yourself, on windows instead, people are lazier and they downloads the binary.

NameError: global name ‘MAX_ITERATIONS’ is not defined

I added MAX_ITERATIONS = 50
and
NameError: global name ‘BAILOUT’ is not defined

I added BAILOUT = 10
and
NameError: global name ‘logical_not’ is not defined

Than i stopped.

Can you send us the correct code, please?

Bye Michele.

@Romuald
Thank you for your improvements. I update the compare as soon as possible

Here is a better perl version : http://pastebin.com/d74858d17

Two majors changes:
- use “my” instead of “local” (slower and doesn’t make sense in this context: it’s used to localize the value of one variable to a block, resetting is to it’s previous value once out of the block)
- use Time::HiRes module since time() does not return microseconds

class Bench2(object):
….def __init__(self):
……..print (‘Rendering…’)
……..y,x = numpy.mgrid[-39/40.:39/40.:78j, -39/40.:39/40.:78j]
……..i = self.start(x,y)
……..out = numpy.ones(i.shape, dtype=numpy.byte) * 42
……..out[i==0] = 32
……..for row in out:
…………stdout.write(row.tostring())
…………stdout.write(‘n’)
….
….def start(self, x, y):
……..”"”x & y are now arrays of shape (nx, ny)
……..”"”
……..c = (y – 0.5) + 1.j*x
……..c = c.ravel()
……..z = numpy.zeros(c.shape, numpy.complex128)
……..idx = numpy.arange(z.size)
……..out = numpy.ones(x.shape, numpy.bool)
……..out_ = out.ravel()
……..for i in xrange(MAX_ITERATIONS):
…………z **= 2
…………z += c
…………mask = (z.real + z.imag) <= BAILOUT
…………idx, IDX = idx[mask], idx[logical_not(mask)]
…………z = z[mask]
…………c = c[mask]
…………out_[IDX] = False
……..return out

This now goes ~23x faster than the original CPython version (py2.6), although at the cost of less readable code. So there you have it: python is faster than java!

(and this is why python rules for scientific and engineering type applications)

import numpy
from numpy import logical_or

class Bench2(object):
def __init__(self):
print (‘Rendering…’)
y,x = numpy.mgrid[-1.:1.:80j, -1.:1.:80j]
i = self.start(x,y)
out = numpy.ones(i.shape, dtype=numpy.byte) * 32
out[i==0] = 42
for row in out:
stdout.write(row.tostring())
stdout.write(‘n’)

def start(self, x, y):
“”"x & y are now arrays of shape (nx, ny)
“”"
c = (y – 0.5) + 1.j*x
z = numpy.zeros(c.shape, numpy.complex128)
mask = numpy.zeros(c.shape, numpy.bool)
for i in xrange(MAX_ITERATIONS):
z2 = z**2 + c
mask = logical_or(mask, (z2.real + z2.imag) > BAILOUT, mask)
z = z2
return mask

On my machine this runs 3.75x faster than the vanilla CPython version. In fact, even easier is to use Cython, add a few type-defs and get raw C-speed.