scm2cpp 3 times faster than Gambit-C. More faster?

[niitsuma]

scm2cpp is compiler scheme subset to readable c++.

http://www.suri.cs.okayama-u.ac.jp/servlets2/scm2cpp.rkt
http://www.suri.cs.okayama-u.ac.jp/~niitsuma/scm2cpp/scm2cpp-0.7-commandline.tar.gz

Benchmark shows scm2cpp 3 times faster than Gambit-C.
(10 times faster with -O3 ?)


Gambit-C

$ time ./fft-gambit
real 0m16.797s
user 0m13.046s
sys 0m0.135s

scm2cpp
$ time ./fft.exe
real 0m5.202s
user 0m4.298s
sys 0m0.029s

scm2cpp with g++ -O3

$ time ./fft-O3
real 0m0.984s
user 0m0.807s
sys 0m0.005s

used code
http://www.suri.cs.okayama-u.ac.jp/~niitsuma/scm2cpp/benchmark/fft.scm

Environment MacBook core2duo 2.26Ghz

$ uname -a
Darwin 10.8.0 Darwin Kernel Version 10.8.0: Tue Jun 7 16:33:36 PDT
2011; RELEASE_I386 i386

[Brad Lucier]

On Feb 6, 12:20 am, niitsuma <hirotaka.niits...@gmail.com> wrote:
> scm2cpp is compiler scheme subset to readable c++.
>
> http://www.suri.cs.okayama-u.ac.jp/servlets2/scm2cpp.rkt

> http://www.suri.cs.okayama-u.ac.jp/~niitsuma/scm2cpp/scm2cpp-0.7-comm...

>
> Benchmark shows scm2cpp 3 times faster than Gambit-C.
> (10 times faster with -O3 ?)
>
> Gambit-C
>
> $ time ./fft-gambit
> real    0m16.797s
> user    0m13.046s
> sys     0m0.135s
>
> scm2cpp
> $ time ./fft.exe
> real    0m5.202s
> user    0m4.298s
> sys     0m0.029s
>
> scm2cpp with g++ -O3
>
> $ time ./fft-O3
> real    0m0.984s
> user    0m0.807s
> sys     0m0.005s
>

> used codehttp://www.suri.cs.okayama-u.ac.jp/~niitsuma/scm2cpp/benchmark/fft.scm

>
> Environment MacBook core2duo 2.26Ghz
>
> $ uname -a
> Darwin 10.8.0 Darwin Kernel Version 10.8.0: Tue Jun 7 16:33:36 PDT
> 2011; RELEASE_I386 i386

Yes, that's true, for that fft code from the Gabriel benchmarks, which
was written in a style for Common Lisp, and massaged to run fast on
Stalin. On my Mac laptop, I get running on Gambit:

[Bradley-Luciers-MacBook-Pro:~/Desktop] lucier% time ./fft
11.339u 0.111s 0:11.45 99.9% 0+0k 0+0io 0pf+0w

On the other hand, if you extract the FFT code from lib/_num.scm in
the Gambit source code, put

(declare (standard-bindings)
(extended-bindings)
(block)
(not safe)
)

at the top, and run the following benchmark:

(define (test n)
(define (flsquare x) (fl* x x))
(let ((two^n
(expt 2 n))
(table
(time (make-w (fx- n 1)))))
(let ((a
(make-f64vector (fx* two^n 2)))
(inexact-two^n
(fixnum->flonum two^n)))
(do ((i 0 (fx+ i 1)))
((fx= i (fx* two^n 2)))
(f64vector-set! a i (fixnum->flonum i)))
(time (do ((i 0 (fx+ i 1)))
((fx= i 1000))
(direct-fft-recursive-4 a table)
(inverse-fft-recursive-4 a table)
(do ((i 0 (fx+ i 1)))
((fx= i (fx* two^n 2)))
(f64vector-set! a i (fl/ (f64vector-ref a i)
inexact-two^n)))))
(do ((i 0 (fx+ i 1))
(error 0. (fl+ error (flsquare (fl- (f64vector-ref a i)
(fixnum->flonum i))))))
((fx= i (fx* two^n 2))
(display (flsqrt error))
(newline))))))

(test 10)

which does twice as much work as your benchmark (because it does 1000
forward and 1000 backward transforms), then it runs in


(time (do ((i 0 (fx+ i 1))) ((fx= i 1000)) (direct-fft-recursive-4 a
table) (inverse-fft-recursive-4 a table) (do ((i 0 (fx+ i 1))) ((fx= i
(fx* two^n 2))) (f64vector-set! a i (fl/ (f64vector-ref a i) inexact-
two^n)))))
60 ms real time
60 ms cpu time (60 user, 0 system)
no collections
160000 bytes allocated
22 minor faults
no major faults

on my machine.

So this program on Gambit runs over 300 times faster than the Gabriel
benchmark program on Gambit.

I don't see why someone interested in fast FFTs would care how fast
the Gabriel benchmark program runs on Gambit (or, perhaps, on any
other system).

Brad

[Brad Lucier]

On Feb 6, 7:29 pm, Brad Lucier <luc...@math.purdue.edu> wrote:

> On the other hand, if you extract the FFT code from lib/_num.scm in
> the Gambit source code, put

I put the code at

http://www.math.purdue.edu/~lucier/fft-gambit.scm

Brad