[Fluxus] vector operations performance

[evan.raskob [lists]]

Interesting results... For vaddx, garbage collection time just about  
doubles when you add 3 random vectors, but vadd actually  
*decreases*?  That seems odd to me.



On May 18, 2009, at 4:10 AM, gabor papp wrote:

>>> As long as it doesn't impact performance, please feel free to  
>>> improve
> i am testing the performance of vector operators and comparing them to
> simple macros that can handle every vector sizes. i attached a vadd
> test. it turns out that it is even a bit faster than the c++ version,
> probably due to the scheme->c conversion overhead.
>
> adding 2 random vectors:
> vadd: 0.00564 0.00783 0.0007
> vaddx: 0.00401 0.00724 0.00056
>
> adding 3 random vectors:
> vadd: 0.00608 0.01074 0.00064
> vaddx: 0.00489 0.00951 0.00108
>
> the results are calculated by adding the vectors 100000 times and
> calculating the average time needed to execute the instructions.  
> vadd is
> the current vector addition fluxus uses, vaddx is the macro attached.
> the three numbers are: the number of milliseconds of CPU time required
> to obtain this result, the number of "real" milliseconds required for
> the result, and the number of milliseconds of CPU time spent on  
> garbage
> collection.
>
> what do you think? can we change the vector operations to more  
> flexible
> scheme versions or am i overlooking something?
>
> best,
> gabor
>
>
> ----------------------------
>
> (define (vadd2 a b)
>   (build-vector (vector-length a)
>                 (lambda (i)
>                   (+ (vector-ref a i)
>                      (vector-ref b i)))))
>
> (define-syntax vaddx
>   (syntax-rules ()
>     [(vaddx a b) (vadd2 a b)]
>     [(vaddx a b c ...) (vadd2 a (vaddx b c ...))]))
>
> (define (profile f n)
>   (let ([timing-info
>           (map (lambda (x)
>                   (exact->inexact (/ x n)))
>                (foldl (lambda (vl al)
>                     (map + vl al))
>                     (list 0 0 0)
>                     (for/list ([i (in-range n)])
>                         (let-values ([(v cpu user gc) (time-apply f  
> '())])
>                             (list cpu user gc)))))])
>     (printf "~a ~a ~a~n" (list-ref timing-info 0)
>                          (list-ref timing-info 1)
>                          (list-ref timing-info 2))))
>
> (display "vadd: ")
> (profile (lambda () (vadd (rndvec) (rndvec))) 100000)
>
> (display "vaddx: ")
> (profile (lambda () (vaddx (rndvec) (rndvec))) 100000)
>

Evan Raskob
Top Floor
4-8 Arcola Street
London E8 2DJ
United Kingdom