Re: of Mops, jit and perl6
Dan Sugalski wrote: At 10:44 AM +0200 7/28/02, Leopold Toetsch wrote: 2) Some Mops numbers, all on i386/linux Athlon 800, slightly shortend: (»make mops« in parrot root) Just out of curiosity, I presume the (rather abysmal) perl 6 numbers include time to generate the assembly and assemble it--have you tried running the generated code by itself as a test? (At the moment, the assembler's rather slow) No, these Mopsen are pure run time. As Sean alredy stated, code quality is to blame. Times for individual steps are: $ time perl6 --imc mops.p6 real0m1.794s user0m1.690s sys 0m0.100s $ time ../imcc/imcc mops.imc mops.pasm 327 lines compiled. Compiling assembly module mops.pasm real0m0.032s user0m0.020s sys 0m0.000s [lt@thu8:~/src/parrot-007/languages/perl6] $ time perl ../../assemble.pl mops.pasm mops.pbc real0m0.481s user0m0.470s sys 0m0.010s This adds up to 2.3s to get a perl6 program running. leo
Re: of Mops, jit and perl6
Dan Sugalski wrote: At 10:44 AM +0200 7/28/02, Leopold Toetsch wrote: 2) Some Mops numbers, all on i386/linux Athlon 800, slightly shortend: Just out of curiosity, I presume the (rather abysmal) perl 6 numbers After the bugfix in perlarray.pmc I can bring you new numbers, which are not that abysmal. Changing the while loop in mops.p6 to: $I4 = $I4 - $I3 # subI4, I4, I3 - while ($I40); # if I4, REDO + while ($I4); # if I4, REDO (which is more correct, comparing to e.g. mops.pl) $ time perl6 -Rj mops.p6 Iterations:100 Estimated ops: 200 Elapsed time: 1.032118 M op/s:1.937763 real0m3.357s user0m3.190s sys 0m0.150s We have already the same Mops as perl5, but additionaly 2.3 seconds overhead. Just running the byte code is as fast as perl5. Without jit, mops.p6 performs at 0.8 Mops. leo
Re: of Mops, jit and perl6
At 10:23 AM 7/30/2002 +0200, Leopold Toetsch wrote: Dan Sugalski wrote: Just out of curiosity, I presume the (rather abysmal) perl 6 numbers We have already the same Mops as perl5, but additionaly 2.3 seconds overhead. Just running the byte code is as fast as perl5. Without jit, mops.p6 performs at 0.8 Mops. While I don't mind worrying about the compile time numbers, it is probably way, way too early to be placing any value on MOPS numbers. We just barely got a compiler patched together with gum and tape before TPC and the code generated is t-t-t-terrible. Some of it is Sean's frontend and some is my backend. Toss in the fact that the VM still unoptimized in many areas and we basically have a LOT of work cut out for us. So lets archive these numbers, and see how much we can improve on them and not get too depressed. :) [Both the front and back end compilers are apt to be rewritten anyway] -Melvin
Re: of Mops, jit and perl6
At 10:44 AM +0200 7/28/02, Leopold Toetsch wrote: 2) Some Mops numbers, all on i386/linux Athlon 800, slightly shortend: (»make mops« in parrot root) Just out of curiosity, I presume the (rather abysmal) perl 6 numbers include time to generate the assembly and assemble it--have you tried running the generated code by itself as a test? (At the moment, the assembler's rather slow) -- Dan --it's like this--- Dan Sugalski even samurai [EMAIL PROTECTED] have teddy bears and even teddy bears get drunk
Re: of Mops, jit and perl6
On Mon, 29 Jul 2002, Dan Sugalski wrote: Just out of curiosity, I presume the (rather abysmal) perl 6 numbers include time to generate the assembly and assemble it--have you tried running the generated code by itself as a test? (At the moment, the assembler's rather slow) It's mostly the parser that's slow, not the assembler (about a second's worth of startup time to load the grammar and do a bunch of initialization). But even without that, I'd suspect that the perl6 numbers would suck, since the compiler does some pretty heavy pessimization. For example, I included the main loop at the end of the email, showing a number of painful things, including: - spilling inside the loop - re-initializing the constant zero on every iteration - putting things in PMC's only to take them out again Some of these would be easy to fix, but I think things need to advance more in the features and stability departments before going for speed. /s L_14: L_17: new P6, .PerlUndef set I0, 0 set P9, P31[1] #FETCH set N0, P9 new P0, .PerlUndef set P0, 0 set N1, P0 gt N0, N1, L_19 branch L_comparison18 L_19: set I0, 1 L_comparison18: set P6, I0 if P6, L_while_body20 branch L_while_end13 L_while_body20: L_15: new P6, .PerlUndef set P9, P31[1] #FETCH set P2, P31[3] #FETCH sub P6, P9, P2 clone P9, P6 set P31[1], P9 #STORE branch L_14 L_16: L_while_end13:
Re: of Mops, jit and perl6
At 07:57 PM 7/29/2002 -0700, Sean O'Rourke wrote: On Mon, 29 Jul 2002, Dan Sugalski wrote: Just out of curiosity, I presume the (rather abysmal) perl 6 numbers include time to generate the assembly and assemble it--have you tried running the generated code by itself as a test? (At the moment, the assembler's rather slow) It's mostly the parser that's slow, not the assembler (about a second's worth of startup time to load the grammar and do a bunch of The assembler _is_ still slow. It is faster than it was, but slow nonetheless, and unacceptable for on the fly assembly. Also, IMCC is to blame for some of the poor quality of the generated code. Eventually it will do proper loop invariant optimization among other things. Things can only get better from here. -Melvin
of Mops, jit and perl6
Hi all,
1) perl6 driver program arrived in CVS/languages/perl6
CAVEATS: it generates a lot of intermediate files:
($filename.{warn,imc,pbc,pasm[,c,o,tree,])
an may therefore clobber e.g. mops.c if you run
languages/perl6 perl6 -C ../../examples/mops/mops.p6
So please test it in a copy of your CVS tree.
NB: source is a mess, but I wanted a quick start to get it running.
Thanks to Sean, for prd-perl and many hints.
and finally: perl6 will currently run only in languages/perl6
2) Some Mops numbers, all on i386/linux Athlon 800, slightly shortend:
(»make mops« in parrot root)
[lt@thu8:~/src/parrot-007/examples/assembly]
$ ../../parrot mops.pbc
M op/s:18.270783
$ ../../parrot -j mops.pbc
M op/s:363.468516
$ ./mops
M op/s:196.355836
[lt@thu8:~/src/parrot-007/examples/mops]
$ ./mops
M op/s:366.106527
[lt@thu8:~/src/parrot-007/languages/perl6]
$ perl6 ../../examples/mops/mops.p6
Iterations:100
M op/s:0.303587
$ perl6 ../../examples/mops/mops.p6 -Rj # run jit
M op/s:1.148392
[lt@thu8:~/src/parrot-007/examples/mops]
$ perl mops.pl
Iterations:1000
M op/s:2.22
(Iteration count for mops.pl was reduced, patch sent)
Summary:
Program
Mops
perl5.005_03
2.2
perl6
0.3
perl6 jit 1.1
parrot interpreted 18
parrot compiled 200
parrot jit 363
plain c 366
Remarks:
- jit is a lot faster then compiled
- plain mops.c is only slightly faster the jit, i.e. 1%
- perl6-jit, albeit still totally unoptimized, doesn't compare too bad
to perl5
- mops is a silly test ;-)
3) jit detection was broken, patch is on the way to bugs-parrot.
Have fun,
leo
