Hello,
I have modified your Chicken benchmark code to use the BLAS
library. With that modification, and using SRFI-4 numeric vectors, the
benchmark results are:
0.024 seconds elapsed
0 seconds in (major) GC
22 mutations
0 minor GCs
10 major GCs
The modified code is attached.
-Ivan
On Sun, Oct 11, 2009 at 12:14 AM, Jeronimo Pellegrini <[email protected]> wrote:
> Hello!
>
> I've been looking for some Scheme implementation that is suitable
> for number crunching and also supports SRFIs like 42 (eager
> comprehension), 45 (laziness), 25 (multidimensional arrays),
> and some others.
>
> So, I have tested Chicken, Bigloo and Gambit. There seems to be a
> necessary trade-off between number of SRFIs and number-chunching speed,
> unfortunately.
>
> The time used to do some floating-point number crunching (this
> example repeatedly multiply matrices) is:
>
> bigloo 0.52s
> gambit 2.70s
> chicken 8.70s
>
> This is for compiled Scheme, with all possible optimizations turned
> on.
>
> The problem is that if it takes 1h for C code, I should then expect
> at least the same for bigloo, and at least 17 times more for Chicken
> (that's too much, unfortunately).
>
> So, going through the mailing list archives I learned about crunch,
> which seems to be awesome, but was being discontinued as of February:
> http://lists.gnu.org/archive/html/chicken-users/2009-02/msg00082.html
>
> (This would be my #1 option!)
>
> And I also found messages mentioning Pre-Scheme:
> http://lists.gnu.org/archive/html/chicken-users/2009-02/msg00109.html
>
> But it's not listed as an Egg in the eggs index here:
> http://chicken.wiki.br/chicken-projects/egg-index-4.html
>
> So, my question is -- was crunch really discontinued? What about the
> Pre-Scheme extension?
>
> Thank you a lot!
> J.
>
>
>
> _______________________________________________
> Chicken-users mailing list
> [email protected]
> http://lists.nongnu.org/mailman/listinfo/chicken-users
>
(require-extension srfi-4 blas )
(define matrix-size f64vector-length)
(define read-matrix
(lambda (port)
(let ((m (read port))
(n (read port)))
(let ((size (* m n))
(mat (make-f64vector (* m n))))
(do ((i 0 (+ 1 i)))
((= i size) mat)
(f64vector-set! mat i (read port)))))))
;; match-error is called to complain when mul receives a pair of
;; incompatible arguments.
(define match-error
(lambda (what1 what2)
(error 'mul
"~s and ~s are incompatible operands"
what1
what2)))
(define (zeros m n)
(let ((size (* m n)))
(make-f64vector size 0.0)))
(define bench-multiply
(lambda (a b times)
(let ((c #f))
(do ((i 0 (+ i 1)))
((> i times))
(set! c (blas:dgemm blas:RowMajor blas:NoTrans blas:NoTrans 100 100 4 1.0 a b 0.0 (zeros 100 100))))
c)))
(define main
(lambda (times)
(let ((port-a (open-input-file "x.mat"))
(port-b (open-input-file "y.mat")))
(let ((a (read-matrix port-a))
(b (read-matrix port-b))
(c #f))
(display (format "------------"))
(newline)
(time (set! c (bench-multiply a b times)))
(display (format "------------"))
(newline)
(if (< (matrix-size c) 100)
(display c))
(newline)
0))))
(newline)
(main (read))
_______________________________________________
Chicken-users mailing list
[email protected]
http://lists.nongnu.org/mailman/listinfo/chicken-users
