Re: [Boost-users] cyl_bessel_j performance

Subject: Re: [Boost-users] cyl_bessel_j performance
From: Matwey V. Kornilov (matwey.kornilov_at_[hidden])
Date: 2011-03-11 16:16:23

Thanks for your reply. Yes, fortunately my charges were completely 
unfounded. I was confused by gcc behaviour. When I compile with explicit -O2 
the timing is roughly same.

So, if I may, yet another question. I suppose math::tr1 bessels use math 
implementation, aren't they?

John Maddock wrote:

>> cyl_bessel_j is several times slower than jn from glibc for me. What do I
>> do
>> wrong? Bessel function from boost.math seems to be pretty good coded and
>> optimized but in a runtime it baffles me.
> 
> There are potentially several issue here:
> 
> * Are you building with all optimizations on?  If not it will be way way
> slower: remember the code is expanded inline in your program and not in a
> separate lib.
> * Are you calling cyl_bessel_j with the first parameter an integer type?
> This saves a few cycles over a floating-point argument call since the
> algorithm doesn't have to figure out whether the parameter is an integer
> or not.
> * I confess that's not a class of Math functions I've spent much time
> optimizing for speed.
> * There may be a few domains of the function that are deliberately slower
> in order to avoid the worst of the inaccuracies that the bessel functions
> can suffer from.
> 
> Even given all of the above, they shouldn't be "several times slower", so
> if you have a specific test case I'll investigate what's wrong.
> 
> HTH, John.