Source High-Performance SIMD Vectorization in C# with System.Numerics.Vector<T> - UnreliableCode Forums

SimdSamurai

AVX-512 & SSE

MEMBER

Rep: 285

Join Date: Aug 2021

Posts: 7

Thanks: 36

1y ago · Aug 30, 2024 11:30 AM

#1

Did you know .NET has cross-platform SIMD vectorization built right into the standard library?

CODE

Vector<T>

automatically uses whatever vector width the host CPU supports (128-bit SSE, 256-bit AVX2, or 512-bit AVX-512)!

Vectorized Array Multiplication (Processing 8 or 16 floats per cycle):

CSHARP

using System.Numerics;

public static class SimdMath
{
    public static void MultiplyArrays(float[] a, float[] b, float[] result)
    {
        int simdLength = Vector<float>.Count; // e.g. 8 on AVX2
        int i = 0;

        // Process SIMD chunks in parallel
        for (; i <= a.Length - simdLength; i += simdLength)
        {
            var va = new Vector<float>(a, i);
            var vb = new Vector<float>(b, i);
            (va * vb).CopyTo(result, i); // Executes in 1 CPU instruction!
        }

        // Clean up remaining tail elements
        for (; i < a.Length; i++)
        {
            result[i] = a[i] * b[i];
        }
    }
}

Benchmark: **7.5x faster** than a standard for-loop on an Intel 13th Gen CPU!

Did you know .NET has cross-platform SIMD vectorization built right into the standard library? [code]Vector<T>[/code] automatically uses whatever vector width the host CPU supports (128-bit SSE, 256-bit AVX2, or 512-bit AVX-512)!

[b]Vectorized Array Multiplication (Processing 8 or 16 floats per cycle):[/b]
[code=csharp]using System.Numerics;

public static class SimdMath
{
    public static void MultiplyArrays(float[] a, float[] b, float[] result)
    {
        int simdLength = Vector<float>.Count; // e.g. 8 on AVX2
        int i = 0;

// Process SIMD chunks in parallel
        for (; i <= a.Length - simdLength; i += simdLength)
        {
            var va = new Vector<float>(a, i);
            var vb = new Vector<float>(b, i);
            (va * vb).CopyTo(result, i); // Executes in 1 CPU instruction!
        }

// Clean up remaining tail elements
        for (; i < a.Length; i++)
        {
            result[i] = a[i] * b[i];
        }
    }
}[/code]

[b]Benchmark:[/b] **7.5x faster** than a standard for-loop on an Intel 13th Gen CPU!

SimdSamurai · AVX-512 & SSE
Vectorizing matrix multiplications and particle engines with...

The following users thanked SimdSamurai for this post:

MathMagician

Vector Math & Physics

MEMBER

Rep: 219

Join Date: Nov 2024

Posts: 8

Thanks: 19

1y ago · Aug 30, 2024 1:40 PM

#2

CODE

Vector<T>

is great because it is completely portable across x86 and ARM64 Apple Silicon. If you need specific x86 instructions (like

CODE

Avx2.GatherVector256

or FMA), you can also use

CODE

System.Runtime.Intrinsics.X86

directly!

MathMagician · Vector Math & Physics
Quaternions, transformation matrices, and physics simulation...

NativeCoder

Low-Level C++ / ASM

VIP

Rep: 374

Join Date: Jun 2025

Posts: 12

Thanks: 20

1y ago · Aug 31, 2024 8:40 PM

#3

The automatic tail loop handling is clean. This is the exact pattern used internally in .NET's

CODE

Span.IndexOf

and

CODE

MemoryExtensions

!

NativeCoder · Low-Level C++ / ASM
Passionate about cache locality, SIMD instructions, and raw ...

UnreliableCode Community

High-Performance SIMD Vectorization in C# with System.Numerics.Vector<T>

Create Forum Thread

Add Reputation