Unclear results of timing tests, for branching methods

[Antypodish]

Hi,

Resolved, click to jump to results #3

Original problem [With burst issue]

Can someone correct me, where, or if I do something wrong in my test?
It looks in this test, main thread is faster than parallel for job with burst.
I must be missing something?

Testing timing of different branching approaches.
Using Unity 2020.1.3 Entities 0.14. But that shouldn’t make a difference.

Jobs Branching Timing test has following results:

Iterating 1 million of elements.

Main thread no burst

• • Branching using ? 2 ms

• z = i > b ? i : b ;

• • No branching 5 ms

• z = b ^ ((i ^ b) & -(i << b)) ;

• • No branching with if 6 ms

• z = i ; if ( i <= b ) z = b ;

• • Branching if else 7 ms

• if ( a > b ) { z = a ; } else { z = b ; }

Incorrect burst timing, as used BurstCompatible, instead BurstCompile :p:roll_eyes:
Job Parallel For With burst (equivalent)

• • Branching using ? 9 ms

• int a = i > na_i [1] ? i : na_i [1] ;

• • No branching 22 ms

• int a = na_i [1] ^ ((i ^ na_i [1]) & -(i << na_i [1])) ;

• • No branching with if 10 ms

• int a = i ;if ( i <= na_i [1] ) a = na_i [1] ;

• • Branching if else 10 ms

• if ( i > na_i [1] ) { a = i ; } else { a = na_i [1] ; }

Github link to system base cs file.

Antypodish/Unity_Jobs_TimingTest/blob/master/JobsBranchingTimingTestSystem.cs

In the system OnUpdates, repeating test after every Stop.

Profiler

Any thoughts?

[DreamingImLatios]

You have [BurstCompatible] instead of [BurstCompile].

[Antypodish]

@DreamingImLatios

You have [BurstCompatible] instead of [BurstCompile].

Oh lol, I am so dumb indeed
I think I was taking for granted auto complete
Massive thx.

Anyway, after correction I got new better results.

I have changed 1 million to 10 million of elements, as burst results were close to 0 ms.

Each bursted job takes approx 4.5 ms (according to profiler).

(10 mln elements)
Main thread no burst

• • Branching using ternary (?: ) 49 ms

• z = i > b ? i : b ;

• • No branching 28 ms

• z = b ^ ((i ^ b) & -(i << b)) ;

• • No branching with if 64 ms

• z = i ; if ( i <= b ) z = b ;

• • Branching if else 78 ms

• if ( a > b ) { z = a ; } else { z = b ; }

Job Parallel For With burst (equivalent)

• • Branching using ? ~4.1 ms

• int a = i > na_i [1] ? i : na_i [1] ;

• • No branching ~4.5 ms

• int a = na_i [1] ^ ((i ^ na_i [1]) & -(i << na_i [1])) ;

• • No branching with if ~4.5 ms

• int a = i ;if ( i <= na_i [1] ) a = na_i [1] ;

• • Branching if else ~4.5 ms

• if ( i > na_i [1] ) { a = i ; } else { a = na_i [1] ; }

(1 mln elements - corrected timings)
Main thread no burst

• • Branching using ? 2 ms

• z = i > b ? i : b ;

• • No branching 5 ms

• z = b ^ ((i ^ b) & -(i << b)) ;

• • No branching with if 6 ms

• z = i ; if ( i <= b ) z = b ;

• • Branching if else 7 ms

• if ( a > b ) { z = a ; } else { z = b ; }

Job Parallel For With burst (equivalent)

• • Branching using ? ~0 ms

• int a = i > na_i [1] ? i : na_i [1] ;

• • No branching ~0 ms

• int a = na_i [1] ^ ((i ^ na_i [1]) & -(i << na_i [1])) ;

• • No branching with if ~0 ms

• int a = i ;if ( i <= na_i [1] ) a = na_i [1] ;

• • Branching if else ~0 ms

• if ( i > na_i [1] ) { a = i ; } else { a = na_i [1] ; }

Corrected source code too.
https://github.com/Antypodish/Unity_Jobs_TimingTest/blob/master/JobsBranchingTimingTestSystem.cs

[gebbiz]

You are supposed to use math.select to avoid branching in burst. Though based on your results in this case burst might be smart enough to compile the ? case similarly.

[Antypodish]

@gebbiz interesting.
I never used it.
But decided quickly to test it with 10mln entries.

So far select is on 2nd place with 5ms, vs 3x 4ms and 1x 6ms jobs.

Further looking into profiler

Select job: ~5 ms, 8 instances, total of 25.87 ms

Ternary (? : ) job: ~4.5 ms, 8 instances, total of 18.34 ms

Updated the code on git

Antypodish/Unity_Jobs_TimingTest/blob/master/JobsBranchingTimingTestSystem.cs