http://qs1969.pair.com?node_id=513483


in reply to Speed/Efficiency tweaks for a fannkuch benchmark script?

I can obtain a modest (< 10%) but consistent speedup by modifying the permutation generator so that it doesn't even generate sequences that would fail the extended precheck. Like this:
sub fannkuch { my ( $a, $level, $split ) = ( @_, 0, 1 ); my ( $index, $ok, $copy ) = ( $level, $level + 1 == length( $a ), +$a ); if ($ok) { # print "Before munging: ($split) ", unpack('C*', $copy); $index = $split - 1; substr($copy, $index, 0) = chop($copy); } do { if ($ok) { # print "($split) ", unpack('C*', $copy); my $q = $copy; my ( $k, $flips ); for ( $flips = 0; ( $k = ord( $q ) ) != 1; $flips++ ) { substr( $q, 0, $k ) = reverse substr( $q, 0, $k ); } if ( $flips >= $maxflips ) { if ( $flips == $maxflips) { push @max_sequence, $copy; } else { $maxflips = $flips; @max_sequence = ($copy); } } } else { fannkuch( $copy, 1 + $level, $split ); $split = $level + 1 if $index == $split; } substr( $copy, $index - 1, 2 ) = reverse substr( $copy, $index + -1, 2 ); } while $index--; return $maxflips; }
The idea is that $split is the smallest positive number for which all the elements in substr($copy, 0, $split) are less than or equal to $split.

This table shows a few examples:
sequence $split
153421
213452
231453
234154
314254
315425
(The commented-out debugging statements may make it easier to follow what's going on. They certainly helped me to understand it as I was writing it.)

On each top-level run (i.e. where $ok is true) we can skip directly to the first permutation where the moving element (which starts out as the last one in the list) has passed the split-point.

I've also reordered the maxflips logic, which gives an additional small improvement.

On an unloaded Linux PC, I get the following results. This is the output from diff --side-by-side -W 80, with the output from BrowserUk's code on the left, and mine on the right:

[rpc142: /tmp]$ diff --side-by-side -W 80 fann-char-orig.out fann-char +.out Pfannkuchen(1) = 0 for: Pfannkuchen(1) = 0 for: 8.7e-05 elapsed seconds. | 1 > 0.000118 elapsed seconds. Pfannkuchen(2) = 1 for: Pfannkuchen(2) = 1 for: 21 21 7.7e-05 elapsed seconds. | 6.4e-05 elapsed seconds. Pfannkuchen(3) = 2 for: Pfannkuchen(3) = 2 for: 231 231 312 312 0.000134 elapsed seconds. | 0.000133 elapsed seconds. Pfannkuchen(4) = 4 for: Pfannkuchen(4) = 4 for: 2413 2413 3142 3142 0.000307 elapsed seconds. | 0.000281 elapsed seconds. Pfannkuchen(5) = 7 for: Pfannkuchen(5) = 7 for: 31452 31452 0.001292 elapsed seconds. | 0.001164 elapsed seconds. Pfannkuchen(6) = 10 for: Pfannkuchen(6) = 10 for: 365142 365142 415263 415263 416523 416523 456213 456213 564132 564132 0.008305 elapsed seconds. | 0.007341 elapsed seconds. Pfannkuchen(7) = 16 for: Pfannkuchen(7) = 16 for: 3146752 3146752 4762153 4762153 0.062369 elapsed seconds. | 0.056875 elapsed seconds. Pfannkuchen(8) = 22 for: Pfannkuchen(8) = 22 for: 61578324 61578324 0.541676 elapsed seconds. | 0.498606 elapsed seconds. Pfannkuchen(9) = 30 for: Pfannkuchen(9) = 30 for: 615972834 615972834 5.347029 elapsed seconds. | 4.943895 elapsed seconds. Pfannkuchen(10) = 38 for: Pfannkuchen(10) = 38 for: 59186210473 59186210473 58.118665 elapsed seconds. | 54.24672 elapsed seconds.