Beefy Boxes and Bandwidth Generously Provided by pair Networks
No such thing as a small change
 
PerlMonks  

Re: code optimization

by patcat88 (Deacon)
on Nov 03, 2011 at 20:52 UTC ( [id://935771]=note: print w/replies, xml ) Need Help??


in reply to code optimization

Spent an hour or 2 optimizing. Used Concise to look at perl assembly. "perl -MO=Concise,-src,-exec,-stash="main" pmonks.pl > t.txt" I always roll my own benchmark. I dont trust benchmarking perl modules. Got rid of nextstates (the op triggers debugger break points and line numbers in caller and die) and padsvs (i guess copies pointers to scalars from pad to stack positions) by compounding operations whenever possible. Got rid of all regular expressions and file handle operations. You can sysread the file yourself. Dont use buffered I/O or regular expressions, very slow. Do list to list assigns less nextstates and overhead. Dont declare/my variables in loop body, perl engine must create and delete them every loop around.
# 17: $minimumAmount = $1 / $2; 14 <;> nextstate(main 941 pmonks.pl:17) v:*,&,$ 15 <#> gvsv[*1] s 16 <#> gvsv[*2] s 17 <2> divide[$minimumAmount:939,947] sK/TARGMY,2 # 18: $amount = $1; 18 <;> nextstate(main 941 pmonks.pl:18) v:*,&,$ 19 <#> gvsv[*1] s 1a <0> padsv[$amount:940,947] sRM* 1b <2> sassign vKS/2 # 19: $weight = $2; 1c <;> nextstate(main 941 pmonks.pl:19) v:*,&,{,$ 1d <#> gvsv[*2] s 1e <0> padsv[$weight:940,947] sRM* 1f <2> sassign vKS/2
vs,
20 <0> pushmark s 21 <0> padsv[$div:961,963] l 22 <0> padsv[$amount:961,963] l 23 <0> padsv[$weight:961,963] l 24 <0> pushmark sRM* 25 <0> padsv[$minimumAmount:961,963] lRM* 26 <0> padsv[$gamount:961,963] lRM* 27 <0> padsv[$gweight:961,963] lRM* 28 <2> aassign[t27] vKS
also use direct @_ slices, not shift, unless you have variable args or something. Compare
# 69: my $input = ${$_[0]}; 1 <;> nextstate(main 970 pmonks.pl:69) v:*,&,$ 2 <#> aelemfast[*_] s 3 <1> rv2sv sK/3 4 <0> padsv[$input:970,973] sRM*/LVINTRO 5 <2> sassign vKS/2
to
# 59: my $input = ${(shift)}; 1 <;> nextstate(main 965 pmonks.pl:59) v:*,&,$ 2 <#> gv[*_] s 3 <1> rv2av[t3] sKRM/3 4 <1> shift sKP/1 5 <1> rv2sv sK/3 6 <0> padsv[$input:965,968] sRM*/LVINTRO 7 <2> sassign vKS/2
4 heavy ops, compared to 2 heavy ops.

For "my( $start, $minimumAmount, $end, $amount, $weight, $div, $gamount, $gweight) = (0, 1e12, length($input));" dont pad right hand list with undefs, those are extra ops. The scalars are undef to begin with.

Here are my times. badarivu2 is to get rid of a skewing delay in c std lib doing the first console print and maybe some CPU cache preloaded. arivu4 is the fastest. I should add, that this is an excellent candidate for some C code, since we are processing character by character and using pointer arithmetic in Perl. Only optimization I can think of but dont know how to do, is to alias, not copy, $_[0] to $input. Might involve typeglobs, but I dont know how to use them.
C:\Documents and Settings\Owner\Desktop>perl pmonks.pl badarivu2 0.921875 arivu2 8.9530200958252 arivu3 8.59366011619568 arivu4 8.54677510261536 choroba 22.8123989105225 arivu 13.078125 C:\Documents and Settings\Owner\Desktop>
Full code.
#!/usr/bin/perl use warnings; use strict; #use Benchmark 'cmpthese'; use Time::HiRes 'time'; sub arivu { my $input = shift; open my $FH, '<', $input or die; my $minimumAmount = 1e12; my ($amount, $weight); while (<$FH>) { if (/(\d+) (\d+)/) { if ($minimumAmount > ($1/$2)) { $minimumAmount = $1 / $2; $amount = $1; $weight = $2; } } } close $FH; return "$amount\t$weight\n"; } # arivu sub choroba { my $input = shift; open my $FH, '<', $input or die; my $minimumAmount = 1e12; my ($amount, $weight); while (<$FH>) { my ($a, $w) = split; if (defined $a and defined $w and (my $ratio = $a / $w) < $minimumAmount) { $minimumAmount = $ratio; $amount = $a; $weight = $w; } } close $FH; return "$amount\t$weight\n"; } # choroba sub arivu2 { my $input = ${(shift)}; my( $start, $minimumAmount, $end, $amount, $weight, $div, $gamount +, $gweight) = (0, 1e12, length($input)); do{ $amount = substr($input, $start, index($input, ' ', $start)-$s +tart); $start += length($amount) + 1; $weight = substr($input, $start, index($input, "\n", $start)-$ +start); $start += length($weight) + 1; ($minimumAmount, $gamount, $gweight) = ($div, $amount, $weight +) if ($minimumAmount > ($div = $amount/$weight)); }while($start != $end); return "$gamount\t$gweight\n"; } # arivu2 sub arivu3 { my $input = ${(shift)}; my( $start, $minimumAmount, $end, $amount, $weight, $div, $gamount +, $gweight) = (0, 1e12, length($input)); do{ $start += length($amount = substr($input, $start, index($input +, ' ', $start)-$start)) + 1; $start += length($weight = substr($input, $start, index($input +, "\n", $start)-$start)) + 1; ($minimumAmount, $gamount, $gweight) = ($div, $amount, $weight +) if ($minimumAmount > ($div = $amount/$weight)); }while($start != $end); return "$gamount\t$gweight\n"; } # arivu3 sub arivu4 { my $input = ${$_[0]}; my( $start, $minimumAmount, $end, $amount, $weight, $div, $gamount +, $gweight) = (0, 1e12, length($input)); do{ $start += length($amount = substr($input, $start, index($input +, ' ', $start)-$start)) + 1; $start += length($weight = substr($input, $start, index($input +, "\n", $start)-$start)) + 1; ($minimumAmount, $gamount, $gweight) = ($div, $amount, $weight +) if ($minimumAmount > ($div = $amount/$weight)); }while($start != $end); return "$gamount\t$gweight\n"; } # arivu4 my $input; $input .= (join ' ', map int(1 + rand 1000), 1 .. 2) . "\n" for 1 .. 1 +000; #cmpthese(-1, {arivu => sub {arivu \$input}, # choroba => sub {choroba \$input}, # choroba2 => sub {choroba2 \$input}, # }); my $time; $time = time; for(0..500) { arivu2 \$input; }#ignore first time, arivu1 and arivu2 same body, diff times print "badarivu2 ".(time-$time)."\n"; $time = time; for(0..5000) { arivu2 \$input; } print "arivu2 ".(time-$time)."\n"; $time = time; for(0..5000) { arivu3 \$input; } print "arivu3 ".(time-$time)."\n"; $time = time; for(0..5000) { arivu4 \$input; } print "arivu4 ".(time-$time)."\n"; $time = time; for(0..5000) { choroba \$input; } print "choroba ".(time-$time)."\n"; $time = time; for(0..5000) { arivu \$input; } print "arivu ".(time-$time)."\n"; #print arivu \$input; #print choroba \$input; #print choroba2 \$input;

Replies are listed 'Best First'.
Re^2: code optimization
by Anonymous Monk on Nov 04, 2011 at 01:39 UTC

    I always roll my own benchmark. I dont trust benchmarking perl modules.

    Why?

      Profilers can lead to false optimizations because your measuring the overhead of recording the timings (OS interrupts/OS calls/context switches/mutex hits/evals/method resolution) or 1 run if the func is below the OS timer resolution. I took nytprof to these subs and nytprof changed which is the fastest sub which is no good.

        Profilers can lead to false optimizations because your measuring the overhead of recording the timings (OS interrupts/OS calls/context switches/mutex hits/evals/method resolution) or 1 run if the func is below the OS timer resolution. I took nytprof to these subs and nytprof changed which is the fastest sub which is no good.

        Well, Benchmark isn't exactly a profiler, and reinventing Benchmark seems like a particularly pedestrian thing to reinvent, and the overhead is accounted for

        But, regarding NYTProf, I don't get the argument, its not as if it is Benchmark :)

Log In?
Username:
Password:

What's my password?
Create A New User
Domain Nodelet?
Node Status?
node history
Node Type: note [id://935771]
help
Chatterbox?
and the web crawler heard nothing...

How do I use this?Last hourOther CB clients
Other Users?
Others pondering the Monastery: (8)
As of 2024-03-28 10:10 GMT
Sections?
Information?
Find Nodes?
Leftovers?
    Voting Booth?

    No recent polls found