Acovea-4.0.0 : Try out my ebuilds (and scripts)

[scoobydu]

I'm building everything with gcc3.4 atm, so I'll wait for that to finish, then kick it off.

Note that my script only does annotations for gcc3.3 options ("! -O1", "-O2", etc.), not gcc3.4.

Have fun!

Can you expand pls? So your report script will not display any 3.4 specific options that the test script may have tried ...?

[Hypnos]

Can you expand pls? So your report script will not display any 3.4 specific options that the test script may have tried ...?

No, it will display the switch scores just fine, but it won't display any helpful comments in the rightmost column.

[scoobydu]

Can you expand pls? So your report script will not display any 3.4 specific options that the test script may have tried ...?

No, it will display the switch scores just fine, but it won't display any helpful comments in the rightmost column.

Ok, thanks

[taskara]

where does one find the latest scripts? they all seem to be "obsolete" now?

forgive my ignorance..

also is acovea 4.0 gcc 3.4 compliant?

[Useful Idiot]

The latest scripts are in the second page of this thread. Like Hypnos said, that script doesn't yet indicate what options were added to -O2 and -O3 in 3.4 but I did make some changes and now it should.(I took them from the 3.4 config scripts)

Code: Select all

#!/usr/bin/env perl

#----------------------------------------------------------------------------
#
# Settings
#

# Names of benchmarks run by Acovea
@benches = ("alma", "evo", "fft", "huff", "lin", "mat1", "tree");

# Number of populations evolved per benchmark by Acovea
$pops = 5;
# Population size
$popsize = 40;

# Number of generations of evolution by Acovea
$gens = 20;

# Suppress output of statistical mumbo-jumbo?
$suppress = "1";

#----------------------------------------------------------------------------
#
# Digest data from *.run files
#

@switches = ();
$listflag = "";
foreach $bench (@benches) {
   open(BENCHRUN, "$bench.run");
   $genflag = "";
   $countflag = "";
   $blankflag = "";
   $linecount = 0;
   while (<BENCHRUN>) {
      # Look for final iteration
      if(/^iteration $gens/) {$genflag = "1"; next;}
      # Look for start of option counts table
      if($genflag && /^Option counts:/) {$countflag = "1"; next;}
      # Look for end of option counts table (blank line)
      if($genflag && $countflag && /^$/) {$blankflag = "1"; next;}
      # Parse counts lines
      if($genflag && $countflag && (! $blankflag)) {
         # first bench
         s/\n$//;
         if($listflag) {
            s/(^\s*?-\S+?\s+?)(\S)/$2/;
            s/\s+/ /g;
            s/\s$//;
            my @line2 = split(/\s/);
            $ref2 = $switches[$linecount];
            push @$ref2, @line2;
         }
         # subsequent benches
         else {
            s/^\s*?-/-/;
            s/\s+/ /g;
            s/\s$//;
            my @line = split(/\s/);
            $ref= \@line;
            @switches = (@switches, $ref);
         }
         $linecount += 1;
      }
   }
   $listflag = "1";
}

#----------------------------------------------------------------------------
#
# Calculate mean (total/bench), standard deviation, confidence interval,
# and score for each switch across all the benchmarks
#


#
# Implementation of erf()
# (continuing fraction by Laplace, Legendre, Ramanujan)
#
$pi = 3.14159;
sub erffrac {
   my($x, $count, $max) = @_;
   if ($count > $max) {1;}
   else{
      if ($count % 2 == 0) {$a = 2;}
      else {$a = 1;}
      $a*$x+$count/&erffrac($x, $count + 1, $max);
   };
}
sub erf {
   my($x) = @_ ;
   1-exp(-$x**2)/sqrt($pi)/&erffrac($x, 1, 10.*(1 + 1/$x));
}
#
# End implementation of erf()
#

@tswitches = ();
foreach $line (@switches) {
   # Mean
   $col = 1 + $pops;
   $total = 0;
   @line2 = @$line;
   while ($col < $#line2 + 1) {
      $total += $line2[$col];
      $col += 1 + $pops ;
   }
   $mean = $total/$#benches;
   # Standard deviation
   $col = 1 + $pops;
   $total = 0;
   while ($col < $#line2 + 1) {
      $total += ($line2[$col] - $mean)**2;
      $col += 1 + $pops ;
   }
   $stddev = sqrt($total)/$pops/$popsize;
   $mean = $mean/$pops/$popsize;  # renormalize
   # Confidence interval
   $x=$mean/$stddev/sqrt(2); # cum. prob. = erf(n/sqrt(2))
   # Gaussian approximation to Poisson distribution only works
   # for counts >= 5 (or so)
   if ($mean < 5/$popsize) {$confidence = 0.;}
   else {$confidence = &erf($x);}
   # score
   $score = $mean * $confidence * 100;
   # Store data
   my @foo = ($line2[0], $mean, $stddev, $confidence, $score);
   push @tswitches, \@foo;
}

#----------------------------------------------------------------------------
#
# Sort by score
#

sub byScore {
   $b2 = $b; $a2 = $a;
   @b3 = @$b2; @a3 = @$a2;
   $b4 = @b3[4]; $a4 = @a3[4];
   $b4 <=> $a4;
}

@sswitches = sort byScore @tswitches;

#----------------------------------------------------------------------------
#
# Set up annotation hash (GCC 3.3.x)
#

%annhash = ("-fno-merge-constants", "! -O1", "-fno-defer-pop", "! -O1", "-fno-thread-jumps", "! -O1", "-fno-omit-frame-pointer", "! -O1", "-fno-guess-branch-probability", "! -O1", "-fno-cprop-registers", "! -O1", "-fno-if-conversion", "! -O1", "-fno-if-conversion2", "! -O1", "-fno-delayed-branch", "! -O1", "-fno-loop-optimize", "! -O1", "-fno-crossjumping", "! -O1", "-foptimize-sibling-calls", "-O2", "-fcse-follow-jumps", "-O2", "-fcse-skip-blocks", "-O2", "-fgcse", "-O2", "-fexpensive-optimizations", "-O2", "-fstrength-reduce", "-O2", "-frerun-cse-after-loop", "-O2", "-frerun-loop-opt", "-O2", "-fcaller-saves", "-O2", "-fforce-mem", "-O2", "-fpeephole2", "-O2", "-fschedule-insns", "-O2", "-fschedule-insns2", "-O2", "-fregmove", "-O2", "-fstrict-aliasing", "-O2", "-fdelete-null-pointer-checks", "-O2", "-freorder-blocks", "-O2", "-fsched-interblock", "-O2 GCC 3.3", "-fsched-spec", "-O2 GCC 3.3", "-freorder-functions", "-O2 GCC 3.3", "-falign-loops", "-O2 GCC 3.3", "-falign-jumps", "-O2 GCC 3.3", "-falign-labels", "-O2 GCC 3.3", "-finline-functions", "-O3", "-frename-registers", "-O3", "-ffloat-store", "", "-fprefetch-loop-arrays", "", "-fmove-all-movables", "", "-freduce-all-givs", "", "-fno-inline", "", "-ftracer", "", "-fnew-ra", "",  "-funroll-loops", "", "-funroll-all-loops", "",  "-mieee-fp", "", "-malign-double", "", "-mno-push-args", "", "-maccumulate-outgoing-args", "", "-mno-align-stringops", "", "-minline-all-stringops", "", "-mfpmath=387", "", "-mfpmath=sse", "",  "-mfpmath=sse,387", "", "-fomit-frame-pointer", "", "-momit-leaf-frame-pointer", "", "-fno-math-errno", "fast math", "-funsafe-math-optimizations", "fast math", "-fno-trapping-math", "fast math",  "-ffinite-math-only", "fast math", "-fno-signaling-nans", "fast math", "-finline-limit", "" , "-fweb" , "-O3 GCC 3.4", "-funit-at-a-time", "-O2 GCC 3.4", "-falign-functions", "-O2 GCC 3.4");

#----------------------------------------------------------------------------
#
# Display results
#

sub printbar {
   print "------------------------------------------------------------------------------\n";
}

if (not $suppress) {print "  Mean  | Std. Dev. | Conf. |";}
print " Score |  So?  | Switch (annotation)\n";
&printbar;
foreach $line (@sswitches) {
   if (not $suppress) {
      print sprintf("  %5.3f",@$line[1]);
      print " | ";
      print sprintf("  %5.3f  ",@$line[2]);
      print " | ";
      print sprintf("%5.3f", @$line[3]);
      print " |";
   }
   print " ";
   print sprintf(" %4.1f", @$line[4]);
   print " | ";
   # Recommend based on confidence
   if (@$line[3] < 0.683) {print "  No ";}
      elsif (@$line[3] < 0.866) {print "Maybe";}
         else {print " Yes ";}
   print " | ";
   print @$line[0];
   if ($annhash{@$line[0]}) {print " (".$annhash{@$line[0]}.")";}
   print "\n";
}

Edit: s/-funit-at-time/-funit-at-a-time/g

[taskara]

ahh yes, sorry for the inconvenience, and thanks for the post

[Lepaca Kliffoth]

Now this is probably the most interesting thread I ever read. You're saying that with a 48 hours long computation you can get the best optimizations for a cpu wich will most probably have a 3 years long life. I'm definitely going to do this but I couldn't understand a couple of things, so I'm going to ask:

1) You all seem to have dated systems, except amybe an athlon 2100 I've seen in a post. My hardware: P4 2.6 without hyperthreading on my desktop and Athlon XP-M 2.5+ on my laptop. How long will the benchmark take? I can't run anything longer than 3 days, not until exams are over at least.

2) Someone said "I've run it 4 times and got 4 different results". Is this bench REALLY reliable?

3) Because of the CODE tag, the scripts you posted are messed up - to say the least. I have no knowledge of scripting languages (darn) so I'll need links to the 2 files containing the code. Right now there simply is no way to run those scripts and modifying them is out of the question - what if I run the test and after 48 hours I find out the "run the benchmark" script was messed up because of my lack of knowledge (by the way, I think the problem is in the line containing that "1>" and "2>". Maybe it's part of the previous line? Beats me.).

4) A couple of guys couldn't get any result in the end. Someone said something about populations not being enough. How do I make sure the benchmark is configured properly before I run it?

Thanks for the work you've done!

[nmcsween]

Yes I ran it a while ago and got 4 slightly different results but nothing unexpected after looking it over. Acovea should be ran at least twice in my opinion before anyone changes there cflags.

[Lepaca Kliffoth]

Good

[Lepaca Kliffoth]

OK I was impatient so I tried and found out how to make the scripts work. Here is my result:

Code: Select all

 Score |  So?  | Switch (annotation)
------------------------------------------------------------------------------
  34.3 | Maybe | -fforce-mem (-O2)
  31.9 |  Yes  | -fregmove (-O2)
  31.0 |  Yes  | -fcse-skip-blocks (-O2)
  30.6 | Maybe | -fno-merge-constants (! -O1)
  30.1 |  Yes  | -freduce-all-givs
  30.0 |  Yes  | -mieee-fp
  29.6 | Maybe | -falign-labels (-O2 GCC 3.3)
  29.4 | Maybe | -fgcse (-O2)
  29.4 |  Yes  | -frerun-cse-after-loop (-O2)
  29.1 |  Yes  | -malign-double
  29.1 |  Yes  | -fno-cprop-registers (! -O1)
  28.8 |  Yes  | -minline-all-stringops
  27.9 |  Yes  | -fno-thread-jumps (! -O1)
  27.8 | Maybe | -mno-align-stringops
  27.8 |  Yes  | -foptimize-sibling-calls (-O2)
  27.5 | Maybe | -finline-functions (-O3)
  27.4 | Maybe | -fno-delayed-branch (! -O1)
  27.4 |  Yes  | -fsched-spec (-O2 GCC 3.3)
  27.4 |  Yes  | -frename-registers (-O3)
  27.2 | Maybe | -finline-limit
  27.0 |  Yes  | -fdelete-null-pointer-checks (-O2)
  26.9 | Maybe | -fcse-follow-jumps (-O2)
  26.6 |  Yes  | -falign-jumps (-O2 GCC 3.3)
  26.6 | Maybe | -fno-crossjumping (! -O1)
  26.4 |  Yes  | -fno-omit-frame-pointer (! -O1)
  25.7 | Maybe | -freorder-blocks (-O2)
  24.2 | Maybe | -fmove-all-movables
  24.1 | Maybe | -fno-signaling-nans (fast math)
  24.1 | Maybe | -ffinite-math-only (fast math)
  23.9 |  Yes  | -fno-trapping-math (fast math)
  23.8 |  Yes  | -maccumulate-outgoing-args
  23.8 | Maybe | -fsched-interblock (-O2 GCC 3.3)
  23.7 | Maybe | -fpeephole2 (-O2)
  22.5 | Maybe | -frerun-loop-opt (-O2)
  22.2 |  Yes  | -fno-defer-pop (! -O1)
  22.1 | Maybe | -fno-math-errno (fast math)
  22.0 | Maybe | -mno-push-args
  21.8 | Maybe | -funsafe-math-optimizations (fast math)
  20.0 | Maybe | -fno-if-conversion2 (! -O1)
  19.3 | Maybe | -ftracer
  19.1 | Maybe | -fcaller-saves (-O2)
  18.4 |  Yes  | -falign-loops (-O2 GCC 3.3)
  18.4 |   No  | -fexpensive-optimizations (-O2)
  18.0 |   No  | -fstrength-reduce (-O2)
  17.1 |   No  | -fno-if-conversion (! -O1)
  16.9 |   No  | -fschedule-insns (-O2)
  16.8 |   No  | -freorder-functions (-O2 GCC 3.3)
  14.9 |   No  | -fschedule-insns2 (-O2)
  14.7 |   No  | -fprefetch-loop-arrays
  11.9 |   No  | -fno-inline
  11.7 |   No  | -fstrict-aliasing (-O2)
  10.3 |   No  | -funroll-loops
   9.9 |   No  | -fno-guess-branch-probability (! -O1)
   9.4 |   No  | -ffloat-store
   7.8 |   No  | -fomit-frame-pointer
   7.0 |   No  | -fnew-ra
   6.7 |   No  | -funroll-all-loops
   4.7 |   No  | -fno-loop-optimize (! -O1)
   0.0 |   No  | -mfpmath=387
   0.0 |   No  | -mfpmath=sse
   0.0 |   No  | -mfpmath=sse,387
   0.0 |   No  | -momit-leaf-frame-pointer

Any advice? (is this thread dead alredy??)

[Daagar]

The output you already posted _is_ the advice :) Simply set your CFLAGS to be -O1 or -O2, the correct -march for your platform, along with everything that is 'Yes' (except dangerous flags like -mieee-fp and -malign-double). And tack on -pipe to the end. That's basically all you need to have 'acovea cflags'.

[Lepaca Kliffoth]

Thanks!
What's with the gcc online manual? Can't connect +_*

[Lepaca Kliffoth]

-O2 seems the one that's turning on most Yes, but it turns on a lot of No too. How should I behave? Sorry, I need some _more_ advice. And the gcc manual is down *darn*

[darkless]

Use -O1 and the options of your choise (preferably the "yes" options). If using -O2, you'll enable a lot of "no" options, as you pointed out yourself.

[djm]

question 1: since updatedb runs everyday, and seems to use a fair amount of system resources, isn't that going to make the results unreliable unless you stop it from running? (since scanning this thread and searching for acovea and updatedb together on these forums didn't show anyone mentioning this I doubt everyone did this)

question 2: how do I (temporarily) stop updatedb from running everyday? Is it controlled by cron?

[Hypnos]

question 1: since updatedb runs everyday, and seems to use a fair amount of system resources, isn't that going to make the results unreliable unless you stop it from running? (since scanning this thread and searching for acovea and updatedb together on these forums didn't show anyone mentioning this I doubt everyone did this)

question 2: how do I (temporarily) stop updatedb from running everyday? Is it controlled by cron?

I haven't had updatedb on my machine for ages; didn't think of it. Yeah, I think you deactivate it by tweaking the cron config.

[MikeP]

1: probably, the question is how much more unreliable it gets
2: look in /etc/cron.daily/ and just move everything out of there to disable it, move it back in when youre done.

[Daagar]

Another suggestion (mentioned a few pages back, but in a different context) is that you might want to edit the config file a bit and remove options you know you won't use, so Acovea doesn't waste evolutionary time adding in options that you don't care about.

For example, mieee-fp and malign-double (and even ffast-math and fnew-ra) have all been targeted as possible 'dangerous' flags. By removing them from the acovea configs, you improve the chances of the other flags getting a shot of being mutated in rather than having generations spend time trying to deal with 'useless' flags. This would potentially lead to a 'stronger' set of flags after a single run of acovea (and hopefully cause a quicker convergence to a single set of flags across multiple acovea runs).

[mayday147]

It may be a stupid question, but i read all this thread, I didn't yet figured out how to run acovea for athlon-xp. My specs: [AMD Athlon(tm) XP 1600+] [392MB RAM].

[Hypnos]

It may be a stupid question, but i read all this thread, I didn't yet figured out how to run acovea for athlon-xp. My specs: [AMD Athlon(tm) XP 1600+] [392MB RAM].

/usr/share/doc/acovea-4.0.0/README.gz

[ikaro]

The next version of acovea will include the configs I sent the author for AMD cpus.


and as a side note, anyone got this running with gcc 3.4 ?
here it just segfaults....

[Lepaca Kliffoth]

OK I thought I'd just come back here and post my CFLAGS. Thanks for all the help

Code: Select all

CFLAGS="-march=pentium4 -O1 -fregmove -fcse-skip-blocks -frerun-cse-after-loop -fno-cprop-registers -minline-all-stringops -foptimize-sibling-calls -fsched-spec -frename-registers -fdelete-null-pointer-checks -falign-jumps -fno-omit-frame-pointer -fno-trapping-math -maccumulate-outgoing-args -fno-defer-pop -falign-loops -pipe"

[mayday147]

It may be a stupid question, but i read all this thread, I didn't yet figured out how to run acovea for athlon-xp. My specs: [AMD Athlon(tm) XP 1600+] [392MB RAM].

/usr/share/doc/acovea-4.0.0/README.gz

I read that file but i still don't know how to do it. It says nothing about athlon-xp.

[ikaro]

these are the CFLAGS I use for amd xp.
These were the ones i choose after i ran the acovea with gcc 3.3.3, im trying to run it again with gcc 3.4 .. but havent had much luck yet.

Code: Select all

CFLAGS="-march=athlon-xp -O3 -pipe -fno-cprop-registers -fno-thread-jumps -fno-defer-pop -maccumulate-outgoing-args -fno-if-conversion2 -fno-delayed-branch -fno-crossjumping -fno-merge-constants -fno-omit-frame-pointer -ftracer -finline-limit=600 -minline-all-stringops -mno-push-args -fmove-all-movables -mno-align-stringops"

been running these for some time now, and it runs stable.

[ikaro]

doubble posted.
sorry