Vector3 and other structs optimization of operators

Hi,
While working on some real time mesh generating code, that did hundreds of thousands of Vector3 operations per frame, I was surprised to find that Vector3 (among other Unity structs) operators (*, +, ...) can be easily and massively optimized.

The current implementation of the * operator in Vector3 is:

Code (CSharp):

1. public static Vector3 operator *(Vector3 a, float d)
2. {
3.     return new Vector3(a.x * d, a.y * d, a.z * d);
4. }

The optimized implementation I suggest is:

Code (CSharp):

1. public static Vector3 operator *(Vector3 a, float d)
2. {
3.     Vector3 result;
4.     result.x = a.x * d;
5.     result.y = a.y * d;
6.     result.z = a.z * d;
7.     return result;
8. }

I run some simple comparison test that you can find here https://dropb.in/ponda.nimrod, and the result was:
When run 50000 times, the current Unity's operator took 18.9 ms to execute, while the optimized one took 2.5 ms.
The reason behind this difference is that the optimized version avoids calling unnecessarily the Vector3 constructor.

I opened a suggestion at Unity's feedback site, so please support it by voting for it so we can see this optimization integrated in Unity some day
https://feedback.unity3d.com/suggestions/vector3-and-other-structs-optimization-of-operators

Thanks and have a nice day.

 

Interesting, I never thought about checking the performance of unitys math library stuff.
I just tried out the operator you mentioned, and I do indeed get better performance when I roll my own struct and implement the operator as you suggest.
Unitys Vector3 seems to take 30% more time for me, both inside the editor and inside a build (on 5.6.1f1).

 

Thanks a lot ThomasTrenkwalder for your tests, and thanks for voting for my suggestion, I hope it will make the Unity team consider the suggestion.

 

Interesting. Considering that the current implementation is less verbose, and all that the constructor does is assign the corresponding fields, I wonder why the C# compiler doesn't optimize this?

Have you tested this in standalone release mode? Maybe it only affects debug mode?

When you tested it in standalone did you disable development build/script debugging?

 

I wonder the same thing. It seems to me to be something the compiler can handle, but I suppose things are more complicated than what I imagine.

I confirm the optimization works in those conditions as well. I used a heavier version of the script above, and used Fraps to get the FPS count (to exclude any Unity's profiler possible issue), and here are the results:
- Optimized version: 53 FPS
- Unoptimized version: 42 FPS

 

Wow I didn't realize that running the constructor in this case would make that big of difference. (I'm self taught BTW) but thanks for this insight. I will keep this in mind when developing my stuff. Thanks for the info!

 

yep, a constructor function is just that... a function.

So it allocates a stack frame to call it.

If you don't call the constructor though, it just allocates the memory needed for the struct with empty values.

This is why struct's don't allow field initializers, they MUST be empty values. Where as classes always have a constructor phase, so it doesn't have this restriction.

...

I find this a minor optimization, probably resulting from early Unity. I bet it came about because the unity devs were all C++ programmers first and foremost, and so didn't really consider the inner workings of the mono CLR. But it is a area of optimization that could potentially give a little oomph since vector construction is very common.

 

Yeah I knew that the constructor is a glorified method basically but that sheds some light on how C# allocates it's frame steps so thank you for that info!

 

Yup, no development build here. I measured the times using the .NET Stopwatch class.
One would think that the compilers (either the C# one or the JIT) should be able to inline this constructor call, but apparently they just don't.

Considering that working with Vector3s and other math structs is quite common in many games, optimizing these operators would provide a nice benefit, and it doesn't even look like a lot of work ^^

 

I completely agree. I think that implementing these optimizations could be done in less than a man-day.

You are welcome And please consider voting for the suggestion to hopefully make Unity's team implement it.
https://feedback.unity3d.com/suggestions/vector3-and-other-structs-optimization-of-operators

 

For those who are interested, here are the IL instructions for the optimized Vector3 multiplication

Code (CSharp):

1. .method public hidebysig static
2.     valuetype [UnityEngine]UnityEngine.Vector3 Optimized_Multiplication (
3.         valuetype [UnityEngine]UnityEngine.Vector3 a,
4.         float32 d
5.     ) cil managed
6. {
7.     // Method begins at RVA 0x20e0
8.     // Code size 58 (0x3a)
9.     .maxstack 3
10.     .locals init (
11.         [0] valuetype [UnityEngine]UnityEngine.Vector3,
12.         [1] valuetype [UnityEngine]UnityEngine.Vector3
13.     )
14.  
15.     IL_0000: nop
16.     IL_0001: ldloca.s 0
17.     IL_0003: ldarga.s a
18.     IL_0005: ldfld float32 [UnityEngine]UnityEngine.Vector3::x
19.     IL_000a: ldarg.1
20.     IL_000b: mul
21.     IL_000c: stfld float32 [UnityEngine]UnityEngine.Vector3::x
22.     IL_0011: ldloca.s 0
23.     IL_0013: ldarga.s a
24.     IL_0015: ldfld float32 [UnityEngine]UnityEngine.Vector3::y
25.     IL_001a: ldarg.1
26.     IL_001b: mul
27.     IL_001c: stfld float32 [UnityEngine]UnityEngine.Vector3::y
28.     IL_0021: ldloca.s 0
29.     IL_0023: ldarga.s a
30.     IL_0025: ldfld float32 [UnityEngine]UnityEngine.Vector3::z
31.     IL_002a: ldarg.1
32.     IL_002b: mul
33.     IL_002c: stfld float32 [UnityEngine]UnityEngine.Vector3::z
34.     IL_0031: ldloc.0
35.     IL_0032: stloc.1
36.     IL_0033: br IL_0038
37.  
38.     IL_0038: ldloc.1
39.     IL_0039: ret
40. } // end of method test::Optimized_Multiplication

and those for the unoptimized one

Code (CSharp):

1. .method public hidebysig specialname static
2.     valuetype UnityEngine.Vector3 op_Multiply (
3.         valuetype UnityEngine.Vector3 a,
4.         float32 d
5.     ) cil managed
6. {
7.     // Method begins at RVA 0xb5b8
8.     // Code size 41 (0x29)
9.     .maxstack 4
10.     .locals init (
11.         [0] valuetype UnityEngine.Vector3
12.     )
13.  
14.     IL_0000: nop
15.     IL_0001: ldarga.s a
16.     IL_0003: ldfld float32 UnityEngine.Vector3::x
17.     IL_0008: ldarg.1
18.     IL_0009: mul
19.     IL_000a: ldarga.s a
20.     IL_000c: ldfld float32 UnityEngine.Vector3::y
21.     IL_0011: ldarg.1
22.     IL_0012: mul
23.     IL_0013: ldarga.s a
24.     IL_0015: ldfld float32 UnityEngine.Vector3::z
25.     IL_001a: ldarg.1
26.     IL_001b: mul
27.     IL_001c: newobj instance void UnityEngine.Vector3::.ctor(float32, float32, float32)
28.     IL_0021: stloc.0
29.     IL_0022: br IL_0027
30.  
31.     IL_0027: ldloc.0
32.     IL_0028: ret
33. } // end of method Vector3::op_Multiply
34.  

 

Did a test because I was curious if the same issue would happen with the object initializer {} feature.

Code (CSharp):

1. public static Vector3 GetSomeVector3()
2. {
3.     Vector3 vec;
4.     vec.x = 3.4f; vec.y = 2.3f; vec.z = 55.5f;
5.     return vec;
6. }
7.  
8. IL_0000 nop
9. IL_0001 ldloca.s  vec
10. IL_0003 ldc.r4    3.4
11. IL_0008 stfld     System.Single UnityEngine.Vector3::x
12. IL_000D ldloca.s  vec
13. IL_000F ldc.r4    2.3
14. IL_0014 stfld     System.Single UnityEngine.Vector3::y
15. IL_0019 ldloca.s  vec
16. IL_001B ldc.r4    55.5
17. IL_0020 stfld     System.Single UnityEngine.Vector3::z
18. IL_0025 ldloc.0
19. IL_0026 stloc.1
20. IL_0027 br.s      IL_0029
21. IL_0029 ldloc.1
22. IL_002A ret
23.  
24. public static Vector3 SomeNewVector3()
25. {
26.     return new Vector3 {x = 3.4f, y = 2.3f, z = 55.5f };
27. }
28.  
29. IL_0000 nop
30. IL_0001 ldloca.s  V_0 //Extra Instruction
31. IL_0003 initobj   UnityEngine.Vector3 //Extra Instruction
32. IL_0009 ldloca.s  V_0
33. IL_000B ldc.r4    3.4
34. IL_0010 stfld     System.Single UnityEngine.Vector3::x
35. IL_0015 ldloca.s  V_0
36. IL_0017 ldc.r4    2.3
37. IL_001C stfld     System.Single UnityEngine.Vector3::y
38. IL_0021 ldloca.s  V_0
39. IL_0023 ldc.r4    55.5
40. IL_0028 stfld     System.Single UnityEngine.Vector3::z
41. IL_002D ldloc.0
42. IL_002E stloc.1
43. IL_002F br.s      IL_0031
44. IL_0031 ldloc.1
45. IL_0032 ret

The object initializer method does also avoid the call to the constructor, but it still has two extra instructions, the important one being an initobj call, which is going to cause a bit of extra work to be done in the form of it initializing all the values of the struct to zero or null. So while that should still be a lot better than the call to the constructor, the local declaration and assignment still wins out.

I'm really surprised the CLR doesn't optimize this initobj call out if it detects you're assigning to every value in the struct.

 

Thanks for that extra information. Didn't tough to test that as well.

 

https://forum.unity.com/threads/unity-mathematics-available-on-github.526100

They are also making a whole new library.

When working with huge amounts of Vectors, it can be benificial to avoid them at all. Just define three floats, its still a new object that is being created.

 

I linked to it over on this other thread earlier on. Vote count has gone from 152 to 165 in three hours.

I wonder why they cannot just fix the aforementioned request rather than create a whole new library that you have to know to get and integrate? I appreciate that a release of Unity (which is presumably what would be required) is no small matter. However, this does seem to be such a fundamental part of a 3D platform to reasonably have expectations of an efficient implementation.

But then maybe I am simply missing something here.

 

You are missing something
That mathematics library isn't the "solution" to this tiny little problem here, it's completely unrelated to it. That mathematics library is designed to help ensure highly efficient compilation of your complex vector/matrix/etc.. math in general, helping it be tightly packed and memory efficient in the burst compiler.
That mathematics library will be integrated in Unity... It's just that it's quite beta right now so people who want to mess with it right now can do so through the repository and also help find bugs or contribute improvements (at some point potentially).

 

Ah, ok. I've got my wires crossed. That means my vote for improved struct performance may not have been wasted then - assuming that ever gets looked at.

 

I rewrote the IL of some Unity's DLLs and measured performance of a few applications. My conclusion was that Unity Technologies can achieve quite some performance improvements, with very little work, with trivial changes only, without actually changing something in user-code.

Yes, they do provide a new math lib, but to make use of it, you need to change your project. This probably give better performance, but it might also not be a trivial change. Therefore, if Unity would just change some simple code in their Vector classes, every existing Unity project would actually benefit from those changes automagically.

Here are my findings:
https://forum.unity.com/threads/wip...faster-without-any-changes-in-seconds.531169/

 

Interesting video. Thumbs up from me.

 

The new Unity mathematics library has definitely its benefits, that are higher than the optimizations this forum thread is about. But using that library means you have to modify/rewrite parts of your code. The Vector3 (and similar) optimization works with 0 modification on your code.

What kills me the most is to know that this optimization should hardly take more than a man/day to Unity's developers to implement, which is peanuts knowing the increase of performance it creates (Peter77 spoke here about a 4% increase in his game). Knowing that people in Unity are aware of the existence of this optimization (Suggestion ticket + me writing to them), the most probable explanation I see is that the internal organization of the Unity company became so complicated that making such simple useful modifications became a daunting task.

 

Wow, that's some great tooling there. Thanks a lot Peter for making this, and pushing the idea beyond where I stopped.

 

Thanks a lot for spreading the word.

 

Hi again,
I implemented the optimizations in a free asset. Here is its link https://www.assetstore.unity3d.com/#!/content/120660?aid=1101l3N9P
It is easy to use, just import it and rebuild your game.
To be completely aware of its limitations, please read the asset description.
Thanks for your interest everyone and have a nice day.