On Tuesday, 31 May 2016 at 19:59:50 UTC, qznc wrote:
On Tuesday, 31 May 2016 at 19:29:25 UTC, Chris wrote:
Would it speed things up even more, if we put the function
`computeSkip` into the loop or is this done automatically by
the compiler?
LDC inlines it. DMD does not.
More numbers:
./benchmark.ldc
Search in Alice in Wonderland
std: 147 ±1
manual: 100 ±0
qznc: 121 ±1
Chris: 103 ±1
Andrei: 144 ±1
Andrei2: 105 ±1
Search in random short strings
std: 125 ±15
manual: 117 ±10
qznc: 104 ±6
Chris: 123 ±14
Andrei: 104 ±5
Andrei2: 103 ±4
Mismatch in random long strings
std: 140 ±22
manual: 164 ±64
qznc: 115 ±13
Chris: 167 ±63
Andrei: 161 ±68
Andrei2: 106 ±9
Search random haystack with random needle
std: 138 ±27
manual: 135 ±33
qznc: 116 ±16
Chris: 141 ±36
Andrei: 131 ±33
Andrei2: 109 ±12
(avg slowdown vs fastest; absolute deviation)
CPU ID: GenuineIntel Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz
Random short strings has haystacks of 10 to 300 characters and
needles of 2 to 10. Basically, no time for initialisation.
Random long strings has haystacks of size 1000, 10_000,
100_000, or 1_000_000 and needles 50 to 500. It inserts a
character into a random index of the needle to force a mismatch.
The last one is the configuration as before.
Overall, Andrei2 (the lazy compute skip) is really impressive.
:)
Yep. It's really impressive. I actually thought that dmd didn't
place `computeSkip` inside of the loop. This begs the question if
it should be moved to the loop, in case we use it in Phobos, to
make sure that it is as fast as possible even with dmd. However,
I like it the way it is now.
`Adrei2` is that it performs consistently well.
