[WIP] [2.0] A maths and hardware intrinsics library for Silk.NET and .NET 5 #173

Perksey · 2020-04-24T17:51:55Z

Summary of the PR

Adds a library for using SIMD instructions with .NET
Adds a library containing generic matrices, vectors, and quaternions; as well as their related maths ops.

What version does this PR target?

2.0

Related issues, Discord discussions, or proposals

#48

Further Comments

DO NOT SQUASH AND MERGE It must be merged in using a merge commit due to Gamma working on it too.

tannergooding · 2020-04-24T20:38:08Z

src/Maths/Silk.NET.Intrinsics/Avx/AvxRegister.Double.cs

+
+namespace Silk.NET.Intrinsics.Avx
+{
+    public partial struct AvxRegister : IRegister<double>


What does "Register" mean?

The name is poorly chosen, but in the context of the intrinsics library a register is a class capable of performing mathematical operations for a given type using a SIMD register such as AVX or SSE.

Maybe IVector or something would be a better name?

Is Avx referring to VEX encoded (128-bit or 256-bit) vectors or strictly 256-bit vectors? If the latter, it might also benefit from a clearer name.

Yeah VEX-encoded, but I don't want that leaking out to the user as I want this to be relatively easy to use and not try to introduce too many new concepts to the user - all the user needs to know is "fast maths ooh shiney"

Is it being VEX encoded an important detail? The operation is ultimately the same, just with better codegen under VEX.
It normally only gets interesting when the size changes, since that changes how much you are processing, etc.

tannergooding · 2020-04-24T20:39:11Z

src/Maths/Silk.NET.Intrinsics/Avx/AvxRegister.Float.cs

+            throw new System.NotImplementedException();
+        }
+
+        public WorkUnit<float> Normalize2(WorkUnit<float> vector)


I'm not sure Normalize2 is a "clear" name. I understand what it means with context, but it isn't immediately obvious.

XML docs will cover that when we add them post-development.

You might also consider renaming to NormalizeVector2 which would disambiguate and not force users to rely on docs.

tannergooding · 2020-04-24T20:39:59Z

src/Maths/Silk.NET.Intrinsics/Avx/AvxRegister.Float.cs

+            throw new System.NotImplementedException();
+        }
+
+        public WorkUnit<float> X(WorkUnit<float> vector)


Is this GetX?

Is there meant to be a SetX or WithX counterpart? (depending on if mutable or immutable)

tannergooding · 2020-04-24T20:41:03Z

src/Maths/Silk.NET.Intrinsics/Avx/AvxRegister.Float.cs

+            throw new System.NotImplementedException();
+        }
+
+        public WorkUnit<float> NegateMultiplyAddFused(WorkUnit<float> x, WorkUnit<float> y, WorkUnit<float> z)


Why is an explicit NegateMultiplyAdd needed? Seems like an optimization the JIT should (and does) do...

This will use the FMA register if applicable, otherwise it will use your everyday AVX register.

Sure, but I'm not sure that clarifies why it is needed?

I can't think of an optimization you can do knowing that it is -(a * b) + c vs (a * b) + c, in which case you should be able to just have MultiplyFusedAdd and let the JIT optimize MultiplyFusedAdd(Negate(a), b, c) when FMA is available and just do your normal math, including the negation, when it isn't

Very good point, cc @sunkin351

tannergooding · 2020-04-24T20:41:33Z

src/Maths/Silk.NET.Intrinsics/Avx/AvxRegister.Float.cs

+            throw new System.NotImplementedException();
+        }
+
+        public unsafe WorkUnit<float> ToVector4(float* ptr)


Is this supposed to be a Load counterpart to Store?

Yeah, probably poorly named.

tannergooding · 2020-04-24T20:42:22Z

src/Maths/Silk.NET.Intrinsics/Avx/AvxRegister.Helpers.cs

+    public partial struct AvxRegister
+    {
+        public static WorkUnitFlags Flags { get; } = GetFlags();
+        public static WorkUnitFlags Flags128F { get; } = Flags | WorkUnitFlags.Vector128 | WorkUnitFlags.TypeFloat;


Framework design guidelines recommends using the full non language specific name so there is no ambiguity.

That is, these should be Flags128Single, Flags128UInt64, etc.

Gotcha, will rename though these are meant to be private properties that for some reason I left public.

tannergooding · 2020-04-24T20:43:49Z

src/Maths/Silk.NET.Intrinsics/Common/WorkUnit128.cs

+    {   
+        public WorkUnitFlags Flags { get; set; }
+        public Vector128<T> Vector { get; set; }
+        public unsafe fixed byte Padding[16];


Why is fixed byte Padding and why does it need both the Vector and the padding?

This is to ensure that you can convert WorkUnit128 to WorkUnit and vice versa using Unsafe.As.

Would it be better to have a Vector128<T> Reserved { get; set; } or Vector128<T> Upper ?

I don't know, would it? As long as the space gets filled I suppose it doesn't really matter, the WorkUnitXXX types aren't really public-facing APIs.

The fixed sized buffers generally generate security cookies and other bits. Also worse codegen due to being many more fields. Having it be a Vector128<T> would remove that and clarify how the bits are reserved and may be interpreted.
It would also allow it to be interpreted as an HVA struct for ABI purposes (assuming you do get rid of Flags like you mentioned might be a consideration below).

Ah fair enough, the more you know :D

Will implement.

tannergooding · 2020-04-24T20:44:51Z

src/Maths/Silk.NET.Intrinsics/Common/WorkUnit256.cs

+    internal struct WorkUnit256<T> where T:unmanaged
+    {
+        public WorkUnitFlags Flags { get; set; }
+        public Vector256<T> Vector { get; set; }


This WorkUnit is going to be 64-bytes, seems like a lot of wasted space...

Possibly, however typically short-lived.

What's the purpose of WorkUnitFlags, they look to encode the type and size of the vector, but it isn't clear how its meant to be used in this context?

I'm considering removing it as I don't think it's worth having. Originally it was gonna hold which set of instructions to use (i.e. AVX or SSE) so that we can have a static class that redirects maths operations to the correct implementation, however I think we can do that without the flags and probably get better treatment from the JIT too.

That sounds reasonable. It would also help with throughput due to reduced memory usage, etc.

Perksey · 2020-05-27T14:07:01Z

Perksey · 2020-06-17T18:12:36Z

Replaced with #190

Perksey added 9 commits March 24, 2020 22:43

First commit of the maths library

4aaf97f

Second commit of the maths library

8779c6a

Wip of a rewrite, need to finalize apis first

f6db406

continued

ac6f88e

Merge branch 'master' into maths

4731832

Finalize API, add registers

02366a4

Commit latest

ab401ae

#ReadyToWork

990fd2f

Refactor incorrect signature

0869e1e

tannergooding reviewed Apr 24, 2020

View reviewed changes

Perksey added this to the 2.0 milestone Apr 30, 2020

Perksey closed this Jun 17, 2020

Perksey deleted the maths branch June 17, 2020 18:12

Perksey added the area-Maths label Apr 10, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[WIP] [2.0] A maths and hardware intrinsics library for Silk.NET and .NET 5 #173

[WIP] [2.0] A maths and hardware intrinsics library for Silk.NET and .NET 5 #173

Perksey commented Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

Perksey Apr 24, 2020

tannergooding Apr 24, 2020

Perksey commented May 27, 2020 •

edited by HurricanKai

Loading

Perksey commented Jun 17, 2020

[WIP] [2.0] A maths and hardware intrinsics library for Silk.NET and .NET 5 #173

[WIP] [2.0] A maths and hardware intrinsics library for Silk.NET and .NET 5 #173

Conversation

Perksey commented Apr 24, 2020

Summary of the PR

What version does this PR target?

Related issues, Discord discussions, or proposals

Further Comments

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Perksey commented May 27, 2020 • edited by HurricanKai Loading

Generator ops

Perksey commented Jun 17, 2020

Perksey commented May 27, 2020 •

edited by HurricanKai

Loading