in reply to Re^2: Efficient bit-twiddling in Perl.
in thread Efficient bit-twiddling in Perl.
Your benchmark is not very realistic. As $n doesn't vary, you are completely ignoring the effect of the CPU cache misses that the lookup table may introduce.
If you use a random $n, you will see that the table approach becomes actually quite slower than the simple one.
I have tried encoding the table in other ways, but have not been able to find any one good enough:
#! perl -slw use strict; use Time::HiRes qw[ time ]; my @lookup; $#lookup = 0x3ffff; $lookup[ $_ ] = [ ( $_ & 0x3f000 ) >> 12, ( $_ & 0xfc0 ) >> 6, $_ & 0x +3f ] for 0 .. 0x3ffff; my( @nxt, @mid, @bot ); $nxt[ $_ ] = ( $_ & 0x3f000 ) >> 12, $mid[ $_ ] = ( $_ & 0xfc0 ) >> 6, $bot[ $_ ] = $_ & 0x3f for 0 .. 0x3ffff; my (@lookup3); $#lookup3 = 0x3ffff; $lookup3[$_ << 6] = [$_ >> 6, $_ & 0x3f] for 0 .. 0xfff; my $lookup4 = 'x' x (3 * (1<<18)); $lookup4 = ''; $lookup4 .= pack CCC => $_ >> 12, ($_>>6) & 0x3f, $_ & 0x3f for 0..0x3 +ffff; my $lookup6 = 'x' x (2 * (1<<12)); $lookup6 = ''; $lookup6 .= pack CC => $_ >> 6, $_ & 0x3f for 0..0xfff; print "tables generated"; our $ITERS //= 10e6; my @n = map int(rand(1<<18)), 1..$ITERS; print "sample data generated"; sub stuff{ # print "@_"; } my $start = time; for my $n (@n) { stuff( ( $n ) >> 18, ( $n & 0x0003f000 ) >> 12, ( $n & 0x00000fc0 ) >> 6, ( $n & 0x0000003f ) ); } printf "Shift&and took: %.12f seconds\n", ( time() - $start ) / $ITERS +; $start = time; for my $n (@n) { stuff( $n >> 18, @{ $lookup[ $n & 0x3ffff ] } ); } printf " Lookup took: %.12f seconds\n", ( time() - $start ) / $ITERS +; $start = time; for my $n (@n) { stuff( $n >> 18, @{ $lookup3[$n & 0x3ffc0] }, $n & 0x3f ); } printf " Lookup3 took: %.12f seconds\n", ( time() - $start ) / $ITERS +; $start = time; for my $n (@n) { stuff( $n >> 18, unpack CCC => substr($lookup4, 3 * ($n & 0x3ffff) +, 3)); } printf " Lookup4 took: %.12f seconds\n", ( time() - $start ) / $ITERS +; $start = time; for my $n (@n) { stuff( $n >> 18, unpack 'x'.(3 * ($n & 0x3ffff)).'CCC' => $lookup4 +); } printf " Lookup5 took: %.12f seconds\n", ( time() - $start ) / $ITERS +; $start = time; for my $n (@n) { stuff( $n >> 18, unpack(CC => substr($lookup6, ($n & 0x3ffc0) >> 5 +, 3)), $n & 0x3f); } printf " Lookup6 took: %.12f seconds\n", ( time() - $start ) / $ITERS +; $start = time; for my $n (@n) { stuff( $n >> 18, unpack('x'.(($n & 0x3ffc0) >> 5).'CC', $lookup6), + $n & 0x3f); } printf " Lookup7 took: %.12f seconds\n", ( time() - $start ) / $ITERS +; __END__ Shift&and took: 0.000000783860 seconds Lookup took: 0.000001267049 seconds Lookup3 took: 0.000001018672 seconds Lookup4 took: 0.000001903985 seconds Lookup5 took: 0.000002110766 seconds Lookup6 took: 0.000001607903 seconds Lookup7 took: 0.000001791258 seconds
|
---|
Replies are listed 'Best First'. | |
---|---|
Re^4: Efficient bit-twiddling in Perl.
by BrowserUk (Patriarch) on Mar 01, 2013 at 12:40 UTC | |
by salva (Canon) on Mar 01, 2013 at 13:25 UTC | |
by BrowserUk (Patriarch) on Mar 01, 2013 at 13:28 UTC | |
by salva (Canon) on Mar 01, 2013 at 14:11 UTC | |
by BrowserUk (Patriarch) on Mar 01, 2013 at 14:30 UTC | |
by Tux (Canon) on Mar 01, 2013 at 14:05 UTC | |
by BrowserUk (Patriarch) on Mar 01, 2013 at 14:49 UTC | |
by Tux (Canon) on Mar 01, 2013 at 14:07 UTC |
In Section
Seekers of Perl Wisdom