Re: [Sbcl-help] Performance: vector math, boxing

[Juho S. <js...@ik...>]

"Andreas Fredriksson" <dep...@gm...> writes:
> Thank you, that is very interesting. I'm toying with something similar
> here and seeing much less consing than what I did with arrays. There's
> still _some_ consing, though. Do multiple-values-bind and values cons
> in the general case or have I done something stupid, as in:
> 
> (defun f (a b c)
>   (declare
>    (type single-float a b c))
>   (values (+ a b) (- b a) (* c a)))
> 
> (defun test ()
>   (time
>    (dotimes (k 100000000)
>      (let ((i (coerce k 'single-float)))
>        (multiple-value-bind (x y z) (f i (+ i 10) (* i 20))
>          (progn))))))
> 
> CL-USER> (test)
> Evaluation took:
>   3.328 seconds of real time
>   3.328208 seconds of user run time
>   0.0 seconds of system run time
>   0 page faults and
>   16,320 bytes consed.
> 
> Why would this cons? Is it because my loop counter is becoming a large
> enough fixnum that it has to be promoted to some arbitrary precision
> integer?

It shouldn't [*]. My guess is that the small amount of consing there
is due to for example Slime doing something in another thread.

[*] I assume you're on x86-64, where single-floats can be passed
    around as immediate values. On x86 or if using double-floats, you
    should declare F as inlined to avoid that consing a few gigabytes
    of temporary floats.

-- 
Juho Snellman