[Date Prev][Date Next][Thread Prev][Thread Next]
[Date Index]
[Thread Index]
**Subject**: **Re: Poor man's benchmarking issue - help needed**
**From**: Lorenzo Donati <lorenzodonatibz@...>
**Date**: Wed, 31 Aug 2011 13:46:39 +0200

On 31/08/2011 11.32, TNHarris wrote:

Lorenzo Donati<lorenzodonatibz@interfree.it> wrote:

According to the following benchmarks, built-in math.floor and
math.ceil
are slower than their counterpart rewritten using the core % operator!

Partly because you're measuring table lookups. Alias math.floor and math.ceil to locals and the times I get go from 40% slower to just 25% slower.

Ouch! Silly me! I knew that!
My code is "littered" by local foo = lib.foo statements! :-)

`My original benchmark did have locals. I did that mistake when pulling
``it out of context in order to post it :-(
`

Also, that ce2 function is wrong when x%1 equals 0.

Uh! Right. I was too hasty.
Let's try again:
----------------------------------------------
local StatTime = require 'ld.debug'.StatTime
local floor = math.floor
local ceil = math.ceil
local function flo( x ) return floor( x ) end
local function flo2( x ) return x - x%1 end
local function ce( x ) return ceil( x ) end
local function ce2( x )
local fp = x%1
if fp == 0 then return x end
return x - fp + 1
end
print( StatTime( 1e8, { flo, flo2 }, 5.5 ) )
--> 2.275e-007 1 0.675
print( StatTime( 1e8, { ce, ce2 }, 5.5 ) )
--> 2.3765e-007 1 0.745
----------------------------------------------
Still quite slower.
BTW I was contacted by PM and I got this objection:
[...]

Your function "flo2" calculates the
floor of a number. But the function "flo" calls another function which
then calculates the floor. That is adding an extra function call to
"flo", which distorts the benchmark. You should compare the two
functions which calculate floor directly, i.e, math.floor on the one
hand and your function flo2 on the other.
Instead of this
print(StatTime(1e8,{ flo, flo2 }, 5.5))
print(StatTime(1e8,{ ce, ce2 }, 5.5))
Try this
print(StatTime(1e8,{ math.floor, flo2 }, 5.5))
print(StatTime(1e8,{ math.ceil, ce2 }, 5.5))
and you will see that the bult in functions are indeed faster.

[...]

`But I don't think this objection is correct, since I did it on purpose
``to wrap floor and ceil in flo and ce, in order to compare them on par
``with flo2 and ce2, because I wanted to compare them against the idioms
`using %, not to an hypothetical reimplementation of ceil and floor using %.

`And so it seems that my conclusion is valid: there is no advantage in
``using floor and ceil instead of using % directly, except when
``readability is important (admittedly most of the times when one is using
``Lua and not LuaJIT).
`

`Anyway this is a big gotcha, IMHO: one wouldn't expect that a library
``function would be slower than a do-it-yourself alternative.
`

`But the question about "useless fat" in Lua remains: why then log10 was
``dropped and ceil and floor remain? It's only a matter of statistical
``relevance? (I suppose that most people use floor and ceil much more
``often than log10). I remember Lua team, replying to people complaining
``about the missing log10 that they could roll their own. Why floor and
``ceil are so special?
`
thanks!
-- Lorenzo

`P.S.: has anyone given a look at my benchmarking function? I hope I
``didn't do something too stupid there (especially with the collectgarbage
``sequence etc.)
`