mathemathics lib's github functions not included by default

tassarho · August 1, 2021, 2:36pm

a was checking mathematics library on Unity.Mathematics/src/Unity.Mathematics/math.cs at master · Unity-Technologies/Unity.Mathematics · GitHub

And it seems some function are not included by default like this one :

The question is: Are thoses functions not added because Burst won’t recognize them or just a miss?

PS: I checked my version of unity mathematics and reset my library to be sure the problem was not on my side.

apkdev · August 1, 2021, 2:42pm

According to the change log on GitHub these additions are currently still unreleased:

github.com

Unity-Technologies/Unity.Mathematics/blob/master/src/CHANGELOG.md

# Changelog

## [Unreleased]

### Fixed
* Fixed `math.hash` crash when using IL2CPP builds on Arm 32 bit devices.
* Fixed obsolete method usage warnings for `MatrixDrawer.CanCacheInspectorGUI` and `PrimitiveVectorDrawer.CanCacheInspectorGUI` in UNITY_2023_2_OR_NEWER.

## [1.3.1] - 2023-07-12

### Added
* Added `math.square` to compute the square (x * x).
* Added `math.orthonormal_basis` to compute an orthonormal basis from a single unit length vector.
* Added `math.sign` for int, int2, int3 and int4.
* Added `math.chgsign` for float, float2, float3, and float4.
* Added `math.Euler` to convert a quaternion to Euler angles.
* Added `math.angle` to compute the angle between two unit quaternions.
* Added `math.rotation` to extract a quaternion rotation from a float3x3 (that may have scale).
* Added `math.mulScale` to scale columns of a float3x3 with scaling coefficients in a float3.
* Added `math.scaleMul` to scale rows of a float3x3 with scaling coefficients in a float3.

This file has been truncated. show original

tassarho · August 1, 2021, 2:45pm

so we can assume Burst will have no effect on these function for now?

apkdev · August 1, 2021, 3:03pm

With simple functions like math.square, I think it’s pretty unlikely Burst would need special intrinsics to optimize the code, so you can almost certainly just implement it yourself (or copy the implementation above) and get equal performance.

As far as I understand, most of the code in Unity.Mathematics doesn’t get any special treatment by Burst. You can write your own math library and as long as you use it in a Burst-compiled job, it will get optimized nicely.

Mortuus17 · August 2, 2021, 6:07am

math.square(myVar) seems more verbose than myVar * myVar but ok ^^
In fact, @apkdev summed it up nicely with

[quote=“apkdev, post:4, topic: 850547, username:apkdev”]
most of the code in Unity.Mathematics doesn’t get any special treatment by Burst
[/quote]Even stuff like myfloat4.xxyy is vectorized optimally by LLVM (the backbone of Burst).
Some major exceptions to that rule are, for instance the math.shuffle functions, half conversions etc. etc. and most importantly all the trig functions and stuff like math.log, math.exp and all that, which actually calls into the Sleef library when using Burst.

[quote=“apkdev, post:4, topic: 850547, username:apkdev”]
You can write your own math library and as long as you use it in a Burst-compiled job, it will get optimized nicely.
[/quote].
While that is generally true, I’ve had to fight the compiler so very often when writing my own libraries, that I mostly stick to Burst intrinsics when necessary - especially when handling vectors which don’t fill up an entire SIMD register (128/256 bits), since LLVM has some major problems with such types. That makes me think that Unity.Mathematics’ types like int3 also get special treatment by Burst.
And also: even trivial control flow, which could be easily vectorized by hand, is most often not vectorized by Burst/LLVM, which is why one should stick to math.select, which is also an actual intrinsic function.

tassarho · August 2, 2021, 3:55pm

thanks for thoses informations it really helps