Resolved Unity SIMD Acceleration

In Unity, are any .NET types SIMD-accelerated? https://learn.microsoft.com/en-us/dotnet/standard/simd

If not, does this exist on the roadmap? Thanks!

 

Have a look at the (new) Unity Math-library.
Unity.Mathematics | Package Manager UI website (unity3d.com)
Designed for ECS/Jobs, it's (supposed to be?) SIMD-accelerated.

Edit: Correction.. It's sort of 'IL'. Burst-compiling using the lib will provide/create the proper SIMD-accelerated types.

 

Thank you!

I'd like to refine my question. Will (or can) this support be added to Transform, Matrix and Vector3? If it can't (or shouldn't be), what is the thinking behind that?

 

Awesome, thank you!

 

xoofx and melvyn may, man. These guys are on F***ing fire.

 

Hello @xoofx , I also have a question about SIMD in Unity.

Does this mean that `System.Numerics.Vector` is not(yet) optimized when built in IL2CPP?

`System.Numerics.Vector` has some public methods like Length() or Distance(). It's implementation contains a call to `Vector.IsHardwareAccelerated` to check if the system supports SIMD or not.

Code (CSharp):

1.  // System.Numerics.Vertor2.cs
2. /// <summary>
3. /// Returns the length of the vector.
4. /// </summary>
5. /// <returns>The vector's length.</returns>
6. [MethodImpl(MethodImplOptions.AggressiveInlining)]
7. public float Length()
8. {
9.     if (Vector.IsHardwareAccelerated)
10.     {
11.         float ls = Vector2.Dot(this, this);
12.         return (float)Math.Sqrt(ls);
13.     }
14.     else
15.     {
16.         float ls = X * X + Y * Y;
17.         return (float)Math.Sqrt((double)ls);
18.     }
19. }

When it is built with IL2CPP(iOS Xcode project, for example), the following code is generated:

Code (CSharp):

1. // System.Single System.Numerics.Vector2::Length()
2. IL2CPP_EXTERN_C IL2CPP_METHOD_ATTR float Vector2_Length_mC99CCF8D33432D4C3392E1390C7605A053251E18 (Vector2_tD99316096CAA530734E8AEB77A0353568D6DC575* __this, const RuntimeMethod* method)
3. {
4.     static bool s_Il2CppMethodInitialized;
5.     if (!s_Il2CppMethodInitialized)
6.     {
7.         il2cpp_codegen_initialize_runtime_metadata((uintptr_t*)&MathF_t54810B9E7503B5BD0850A3D6E6336922E1553C52_il2cpp_TypeInfo_var);
8.         s_Il2CppMethodInitialized = true;
9.     }
10.     {
11.         bool L_0;
12.         L_0 = Vector_get_IsHardwareAccelerated_m783509258751EBED64CBD9F387EC1BB4A15088AA(NULL);
13.         if (!L_0)
14.         {
15.             goto IL_001e;
16.         }
17.     }
18.     {
19.         Vector2_tD99316096CAA530734E8AEB77A0353568D6DC575 L_1 = (*(Vector2_tD99316096CAA530734E8AEB77A0353568D6DC575*)__this);
20.         Vector2_tD99316096CAA530734E8AEB77A0353568D6DC575 L_2 = (*(Vector2_tD99316096CAA530734E8AEB77A0353568D6DC575*)__this);
21.         float L_3;
22.         L_3 = Vector2_Dot_m3DBD983D4FEBCC0FE4E598FE56C6DDFC51C0250B_inline(L_1, L_2, NULL);
23.         il2cpp_codegen_runtime_class_init_inline(MathF_t54810B9E7503B5BD0850A3D6E6336922E1553C52_il2cpp_TypeInfo_var);
24.         float L_4;
25.         L_4 = sqrtf(L_3);
26.         return L_4;
27.     }
28. IL_001e:
29.     {
30.         float L_5 = __this->___X_0;
31.         float L_6 = __this->___X_0;
32.         float L_7 = __this->___Y_1;
33.         float L_8 = __this->___Y_1;
34.         il2cpp_codegen_runtime_class_init_inline(MathF_t54810B9E7503B5BD0850A3D6E6336922E1553C52_il2cpp_TypeInfo_var);
35.         float L_9;
36.         L_9 = sqrtf(((float)il2cpp_codegen_add(((float)il2cpp_codegen_multiply(L_5, L_6)), ((float)il2cpp_codegen_multiply(L_7, L_8)))));
37.         return L_9;
38.     }
39. }

I expected that those call to

Code (CSharp):

1.         L_0 = Vector_get_IsHardwareAccelerated_m783509258751EBED64CBD9F387EC1BB4A15088AA(NULL);
2.         if (!L_0)
3.         {
4.             goto IL_001e;
5.         }

would be optimized depending on the platform I'm targeting, but it wasn't, I can confirm that Vector.IsHardwareAccelerated was called with Xcode Instruments.