- Gambit-list - mailman.iro.umontreal.ca

[gambit-list] Question about performance of shared libraries on MacOS X
by Bradley Lucier 06 Jan '05

06 Jan '05

Marc: As you guessed, once I downloaded the correct version of gambc40b12.tar.gz, my previously reported problems go away. I have a question now about the performance of applications built with shared libraries on MacOS X. When I configure gambit with --enable-shared --enable-single-host I get on my 2GHz G5 running MacOS 10.3.6 [descartes:~/programs/gambc40b12/bench] lucier% ./bench gambit fft Testing fft under Gambit-C Compiling... 0.95 real 0.77 user 0.16 sys -rwxr-xr-x 1 lucier lucier 30512 6 Jan 17:26 fft Running... (time (run-bench name count run ok?)) 3736 ms real time 3000 ms cpu time (2970 user, 30 system) 65 collections accounting for 29 ms real time (30 user, 0 system) 671999504 bytes allocated no minor faults no major faults 3.99 real 3.17 user 0.06 sys (This is with the version of bench at http://www.math.purdue.edu/~lucier/bench.tgz which has some changes to the number of iterations each benchmark is run so that I get at least about 20 clock ticks on my G5. It also changes ##fixnum.logand -> ##fixnum.bitwise-and, etc. You may be interested in integrating the changes into your bench directory.) I'm using gcc-3.4.3 since Apple's gcc-3.3 is just too slow compiling code generated by Gambit. When I configure with just --enable-single-host (so with static libraries), I get [descartes:~/programs/gambc40b12/bench] lucier% ./bench gambit fft Testing fft under Gambit-C Compiling... 1.13 real 0.62 user 0.26 sys -rwxr-xr-x 1 lucier lucier 2478500 6 Jan 17:36 fft Running... (time (run-bench name count run ok?)) 1473 ms real time 1140 ms cpu time (1130 user, 10 system) 65 collections accounting for 29 ms real time (30 user, 0 system) 671999504 bytes allocated no minor faults no major faults 1.75 real 1.13 user 0.04 sys Now that's a big difference. And it's a big difference in code that does very few calls to the Gambit runtime. Do you know why this is happening? Do you know whether I should, say, statically link the Meroon object system with applications that use it to avoid a performance problem? Or is there some type of prelink step for MacOS X that I can use to reduce this problem? Brad

1 0

Gambit-C 4.0 beta 12
by Marc Feeley 05 Jan '05

05 Jan '05

A new beta of Gambit-C 4.0 is now available in source form at this address: http://www.iro.umontreal.ca/~feeley/gambc40b12.tar.gz Here are the main changes: - The compiler now does constant folding and optimistic inlining of predefined procedures. Constant folding is performed for the call (predef arg1 arg2...) when arg1, arg2... are constants (of the right type) and predef is a global variable that is known to be equal to a predefined procedure which does not have side effects (i.e. when a "standard-bindings" declaration is used and predef is not vector-set!, cons, etc). So (declare (standard-bindings)) (define pi (* 4 (atan 1))) expands to (define pi 3.141592653589793) Note that "standard-bindings" is off by default. A new declaration, "run-time-bindings" (which is on by default), tells the compiler that it should check at run-time if the variable predef truly contains the corresponding predefined procedure, and call the predefined procedure if it is, otherwise perform a call using the global variable. This is useful for improving the performance of arithmetic in the absence of unsafe declarations. For example the following definition of fib (with no declarations) (define (fib n) (if (< n 2) 1 (+ (fib (- n 1)) (fib (- n 2))))) is expanded to this code (define fib (lambda (n) (if (if (and (##eq? < '#<procedure #2 <>) (and (##fixnum? n) #t)) (##fixnum.< n 2) (< n 2)) 1 (let ((t1 (fib (or (and (##eq? - '#<procedure #3 ->) (and (##fixnum? n) (##fixnum.-? n 2))) (- n 2)))) (t2 (fib (or (and (##eq? - '#<procedure #3 ->) (and (##fixnum? n) (##fixnum.-? n 1))) (- n 1))))) (or (and (##eq? + '#<procedure #4 +>) (and (##fixnum? t2) (and (##fixnum? t1) (##fixnum.+? t2 t1)))) (+ t2 t1)))))) which runs about 8 times faster than the previous beta version of Gambit and half as fast as when using the declarations (declare (standard-bindings) (fixnum) (not safe)) . - A problem with the configure script which prevented dynamic loading of object files on Mac OS X has been fixed. - A "web-repl" example has been added. It is a Java applet that implements an interface to Gambit's REPL (i.e. you can interact with the Gambit interpreter using a web browser). - The implementation of the C-interface has been improved to allow "struct", "union", and "class" (in C++) C types to be passed between Scheme and C. The C-interface documentation has been updated to explain this feature. An example is attached below. Marc ; C-interface example. (c-declare #<<c-declare-end #include <stdio.h> struct coord2d { int x; int y; }; struct coord2d make_coord2d (int x, int y) { struct coord2d c; c.x = x; c.y = y; return c; } int coord2d_x (struct coord2d p) { return p.x; } int coord2d_y (struct coord2d p) { return p.y; } ___SCMOBJ release_struct_coord2d (void *ptr) { printf ("deleting struct coord2d at %p\n", ptr); ___DELETE_STRUCT(coord2d,ptr); /* in C++ expands to */ /* delete ___CAST(struct coord2d*,ptr) */ /* in C expands to "___free_mem(ptr)" */ return ___FIX(___NO_ERR); } struct coord2d (*callback) (char *, struct coord2d) = NULL; void set_callback (struct coord2d (*c) (char *, struct coord2d)) { ___EXT(___addref_function) (___CAST(void*,c)); /* needed because a reference to the parameter c survives the */ /* activation of the function set_callback */ ___EXT(___release_function) (___CAST(void*,callback)); /* the reference to the old callback is no longer active */ callback = c; printf ("callback = %p\n", callback); } char *context = NULL; void set_context (char *s) { struct coord2d foo, ret; ___EXT(___addref_string) (___CAST(void*,s)); /* needed because a reference to the parameter s survives the */ /* activation of the function set_callback */ ___EXT(___release_string) (___CAST(void*,context)); /* the reference to the old context is no longer active */ context = s; printf ("context = %s\n", context); foo.x = 33; foo.y = 44; ret = callback (context, foo); printf ("ret.x = %d ret.y = %d\n", ret.x, ret.y); } c-declare-end ) (c-define-type coord2d (struct "coord2d" struct-coord2d "release_struct_coord2d")) ; note: had the simpler definition (c-define-type coord2d (struct "coord2d")) ; been used, the tag would have been |struct coord2d| and a default ; release function which calls ___DELETE_STRUCT(coord2d,p) would have ; been used. (define make-coord2d (c-lambda (int int) coord2d "make_coord2d")) (define coord2d-x (c-lambda (coord2d) int "coord2d_x")) (define coord2d-y (c-lambda (coord2d) int "coord2d_y")) (define p (make-coord2d 11 22)) (pp (list p (coord2d-x p) (coord2d-y p))) (pp (foreign-released? p)) (foreign-release! p) ; note: if the foreign object is not explicitly ; released with a call to foreign-release!, the ; runtime system will call the release function ; when the program terminates (pp (foreign-released? p)) (pp (list p)) (c-define (f s c) (char-string coord2d) coord2d "f" "" (pp (list 'in f 's= s 'c= c (coord2d-x c) (coord2d-y c))) (make-coord2d (* (coord2d-x c) 10) (* (coord2d-y c) 100))) (define set-callback (c-lambda ((function (char-string coord2d) coord2d)) void "set_callback")) (set-callback f) ; passing a c-define'd procedure is supported on all platforms ' (let ((z (cons 1 2))) (set-callback (lambda (x y) (pp (list x y z))))) ; passing other procedures ; (predefined procedures, ; closures, etc) only works ; on some platforms (define set-context (c-lambda (char-string) void "set_context")) (set-context "hello") (set-context "world") (display "triggering a garbage collection...\n") (##gc) (display "program is terminating...\n") ; This program's output is: ; ; (#<struct-coord2d #2 0x501660> 11 22) ; #f ; deleting struct coord2d at 0x501660 ; #t ; (#<struct-coord2d #2 0x0>) ; callback = 0x502ba0 ; context = hello ; (in #<procedure #3 f> s= "hello" c= #<struct-coord2d #4 0x501660> 33 44) ; ret.x = 330 ret.y = 4400 ; context = world ; (in #<procedure #3 f> s= "world" c= #<struct-coord2d #5 0x502D30> 33 44) ; ret.x = 330 ret.y = 4400 ; triggering a garbage collection... ; deleting struct coord2d at 0x502d70 ; deleting struct coord2d at 0x501670 ; program is terminating... ; deleting struct coord2d at 0x502d30 ; deleting struct coord2d at 0x501660

2 3

Problems with benchmark software
by Bradley Lucier 02 Jan '05

02 Jan '05

Marc: My bench/prefix-gambit.scm (admittedly perhaps modified) is (##declare (standard-bindings) (extended-bindings) (block) (not safe) ; (not interrupts-enabled) ; (fixnum) (inlining-limit 500) ) ;... (define (run-bench name count run ok?) (let loop ((i 0) (result (list 'undefined))) (if (< i count) (loop (+ i 1) (run)) result))) then a whole bunch of define-macros that include, e.g., (##define-macro (def-macro form . body) `(##define-macro ,form (let () ,@body))) (def-macro (< x y) `(##fixnum.< ,x ,y)) The problem is that the macros don't seem to be used to expand the operations in (the earlier-occuring) run-bench, so we're measuring the inter-module call time and the generic + and < operations in addition to the times for the calls to (run). This can be a problem when the number of iterations for a test is large enough. Brad

1 0

Re: C pointers
by Eduardo Cavazos 31 Déc '04

31 Déc '04

Hello Thomas, Below you provide the example function 'inc_int' and ask what you should pass to it. You should pass it a pointer to an int of course. :-) And where do you get such an object? You can make one with another c-lambda based function. For example: (define make-int-box (c-lambda () int* "___result = malloc(sizeof( int ));")) So now you can do: (define i (make-int-box)) (inc_int i) There is a problem... the int box doesn't contain anything to increment. So we should have a way to set the value inside the box. We also need a way to access the contents of the box. These functions do these: (define set-int-box (c-lambda (int* int) void "*___arg1 = ___arg2;")) (define val-int-box (c-lambda (int*) int "___result = *___arg1;")) That should get you started. The atomic code I announced a few days ago contains a macro called 'c-box' that works like this: (c-box int) Now you can do this: ;; Make a new int box (define i (new-int-box)) ;; Set the value of the box: [i val: 0] ;; Call your inc_int to increment the value in the box (inc_int [i box:]) ;; Print the new value: [i val:] Happy Hacking Ed Cavazos On 31 Dec 2004 17:39:20 +0100, Thomas Hafner <thomas(a)hafner.nl.eu.org> wrote: > Some given C functions receive pointers and manipulate the memory > where the pointers are pointing to. How should my c-lamda expression > look like to mach that requirement? > > Here's a simple example (not real, but to show my problem). If the > given C Funktion looks like that: > void inc_int(int *pI) > { > (*pI)++; > } > > Is it then Ok to use it from Gambit 4.0 beta 11 like follows? > (define inc-int > (c-lambda ((pointer int)) > void > "inc_int")) > > If yes: what can I pass to inc-int? How to express the Scheme > equivalent to the C pointer? > > Otherwise: how should my Scheme procedure definition look like? Of > course the example above is a very simple one. A good solution should > also work with more complicated C functions, e.g. of type > int () (int *, char **). > > Regards > Thomas Hafner > > _______________________________________________ > Gambit-list mailing list > Gambit-list(a)iro.umontreal.ca > http://mailman.iro.umontreal.ca/mailman/listinfo/gambit-list >

1 0

C pointers
by Thomas Hafner 31 Déc '04

31 Déc '04

Some given C functions receive pointers and manipulate the memory where the pointers are pointing to. How should my c-lamda expression look like to mach that requirement? Here's a simple example (not real, but to show my problem). If the given C Funktion looks like that: void inc_int(int *pI) { (*pI)++; } Is it then Ok to use it from Gambit 4.0 beta 11 like follows? (define inc-int (c-lambda ((pointer int)) void "inc_int")) If yes: what can I pass to inc-int? How to express the Scheme equivalent to the C pointer? Otherwise: how should my Scheme procedure definition look like? Of course the example above is a very simple one. A good solution should also work with more complicated C functions, e.g. of type int () (int *, char **). Regards Thomas Hafner

1 0

Announce: atomic
by Eduardo Cavazos 28 Déc '04

28 Déc '04

ATOMIC 0.1.3 http://proteus.freeshell.org/atomic/atomic-0.1.3.tar.gz A collection of code for the Gambit Scheme System. atomic.scm A base library for the rest of the system object.scm * A simple object system * It's relatively concise when compared to traditional Scheme record systems * Example: (class <point> x y) (def p0 (new <point> x: 4 y: 5)) [p0 x:] ;; Getting a slot [p0 x: 10] ;; Setting a slot Xlib * X Lib bindings * Based on the code that comes with the Gambit 4.0 beta distributions * The function and macro names are the same as the C library counterparts. she * A very simple UNIX shell * Supports some basic /bin/sh features like pipes * Readline support * Shell functions are written in Scheme winman * A simple window manager used to test the Xlib bindings and to explore window manager hacking * A REPL like window can be opened to interact with the system This release of Atomic is for Gambit 3.0. It will not work under 4.0. Atomic is in the early stages but is useful. I use the window manager and am close to using the shell full time. Right now it is a testbed for ideas. I welcome any mail regarding the system. Ed Cavazos

1 0

Re: [gambit-list] homogeneous vector performance
by Bradley Lucier 13 Déc '04

13 Déc '04

On Dec 11, 2004, at 4:14 PM, dev null wrote: > I've been playing around with implementations of simple algorithms in > C and gambit scheme, trying to get some idea of the latter's > usefulness for some of my scientific computing. My experience is that once you have problems that don't fit in first or second-level cache, the speed is the same for scientific computing. See http://www.math.purdue.edu/~lucier/615/software/ Also, if you're using gcc, the following options may improve the performance of code generated by Gambit: -O1 -fno-math-errno -fschedule-insns2 -fno-trapping-math -fno-strict-aliasing -fomit-frame-pointer (These options are chosen by configure when building gambit.) You may need the -fno-strict-aliasing for correctness of the compiled code. In my experience code compiled with -O2 goes slower. Brad

1 0

CPS and state machines
by Joel Reymont 13 Déc '04

13 Déc '04

Would anyone have an example of implementing a state machine in continuation passing style (CPS)? I'm looking for a simple example although it does not have to be. I'd like to implement a state machine for a game lobby where on receipt of event-1 I would send game summaries, on event-2 player summaries, etc. I will then build on this example to implement a poker logic state machine. In the game logic fsm there are events like player-joined and player-left and where the game would end when there are not enough players or begin when enough players joined and where it would otherwise state in the same state. Thanks, Joel -- OpenPoker: The Linux of poker software http://wagerlabs.com/forums

3 3

Gambit 4.0b11 and ICC 8.1
by Sven.Hartrumpf＠fernuni-hagen.de 12 Déc '04

12 Déc '04

Hi all. Has anybody compiled Gambit with ICC instead of gcc? My build process ends like this: _kernel.c(16881): remark #1418: external definition with no prior declaration ___BEGIN_MOD2 ^ _kernel.c(16881): remark #869: parameter "___gs" was never referenced^M ___BEGIN_MOD2 ^ icc: error: Fatal error in /opt/intel_cc_80/bin/mcpcom, terminated by segmentation violation compilation aborted for _kernel.c (code 1) make[1]: *** [_kernel.o] Error 1 make[1]: Leaving directory `tmp/gambc40b11/lib' make: *** [all] Error 1 Full typescript at ki220.fernuni-hagen.de/hartrumpf/gambit/ Greetings Sven

1 0

homogeneous vector performance
by dev null 11 Déc '04

11 Déc '04

Hi, I've been playing around with implementations of simple algorithms in C and gambit scheme, trying to get some idea of the latter's usefulness for some of my scientific computing. One of the things I did was code `insertion sort' in both languages: ------------------------------------------------------------------ #include <stdio.h> void isort(int a[], int len) { int i; for (i=1; i < len; i++) { int j; int key = a[i]; for (j=i-1; (j >= 0) && (a[j] > key); j--) { a[j+1] = a[j]; } a[j+1] = key; } } int main(void) { int i; int len = 100000; int * a = (int *) malloc(sizeof(int) * len); for (i=0; i < len; i++) { a[i] = len - i; } isort(a, len); free(a); return 0; } ------------------------------------------------------------- (declare (standard-bindings) (block) (not safe) (fixnum)) (define isort! (lambda (a) (let ((len (vector-length a))) (do ((i 1 (+ i 1))) ((>= i len) a) (let ((key (vector-ref a i))) (vector-set! a (do ((j (- i 1) (- j 1))) ((or (< j 0) (< (vector-ref a j) key)) (+ j 1)) (vector-set! a (+ j 1) (vector-ref a j))) key)))))) (define v (make-vector 100000)) (do ((i 0 (+ i 1)) (len (vector-length v))) ((>= i len)) (vector-set! v i (- len i))) (isort! v) --------------------------------------------------------------------------- For 100_000 elements, my laptop executes the C version in 11 secs and the scheme code in 20 secs. Since the two codes are written at similar abstractions levels, ideally I would like the scheme code to do a little better. Thinking that part of the problem might be the use of an overly-generic scheme vector vs a simple C integer array, I rewrote the scheme version using u32vector functions (just replacing vector-* above with u32vector-*). The result was not what I expected. I was able to sort a 10_000 element u32vector in 26 seconds, but the 100_000 element problem ran for 38 minutes before I got bored with it and killed it. Two questions: 1) Is it possible to make the scheme code run any faster? 2) Does anyone have an idea why the u32vector-based version runs so slow? Thanks very much, Ben

3 3