| // Comment about https://mastodon.social/@[email protected]/113064780292323247 | |
| // https://brandewinder.com/2024/09/01/should-i-use-simd-vectors/ | |
| using System.Numerics; | |
| using System.Runtime.InteropServices; | |
| using System.Runtime.Intrinsics; | |
| public class VectorProcessor | |
| { | |
| public static float Take2(float[] left, float[] right) | |
| { |
| // From a discussion on mastodon https://mastodon.social/@[email protected]/112620103070756332 | |
| // This version: | |
| // * simplifies the setup (an inline array is already a Span) | |
| // * uses BinaryPrimitives.ReverseEndianness to batch writing the long to the buffer | |
| // * uses SkipLocalsInit on the ctor to avoid some zero init | |
| // Should be ~ 15% faster than the revised version with Span | |
| // Anyway most of the time is spent in DateTimeOffset.UtcNow and RandomNumberGenerator.Fill | |
| using System.Buffers.Binary; | |
| using System.Diagnostics.CodeAnalysis; | |
| using System.Runtime.CompilerServices; |
| <#@ template debug="false" hostspecific="false" language="C#" #> | |
| <#@ assembly name="System.Core" #> | |
| <#@ import namespace="System.Linq" #> | |
| <#@ import namespace="System.Text" #> | |
| <#@ import namespace="System.Collections.Generic" #> | |
| <#@ output extension=".txt" #> | |
| <# | |
| var vectors = new [] { "Vector64", "Vector128", "Vector256", "Vector512", "Vector", "Vector2", "Vector3", "Vector4" }; | |
| var templates = new string[0]; | |
| string genT; |
Commit https://github.com/nietras/Sep/tree/46d20139bf2e1f9ca01933ff9ca728873fe790cc
BenchmarkDotNet=v0.13.5, OS=Windows 11 (10.0.22621.819/22H2/2022Update/SunValley2)
AMD Ryzen 7 7840HS w/ Radeon 780M Graphics, 1 CPU, 16 logical and 8 physical cores
.NET SDK=8.0.100-preview.7.23376.3
[Host] : .NET 8.0.0 (8.0.23.37506), X64 RyuJIT AVX2
Job-WASKMK : .NET 7.0.10 (7.0.1023.36312), X64 RyuJIT AVX2
Job-PLCXMZ : .NET 8.0.0 (8.0.23.37506), X64 RyuJIT AVX2
| // Discussion about https://mastodon.social/@[email protected]/110644302160625267 | |
| // Optimized version using AVX. From 4x to 10x faster than a simple version. | |
| // - nint for indexing | |
| // - Unsafe.Add | |
| // - Unrolling of 4 Vector256 + Or of the results to have only 1 branch per loop | |
| // - Finding the local index within the Vector256 without a loop using AVX movemask | |
| // Similar code can be done for Vector128 | |
| using System.Numerics; | |
| using System.Runtime.CompilerServices; | |
| using System.Runtime.InteropServices; |
Commit c6a06151caca9a318891e583f6341b1a707ee408
BenchmarkDotNet=v0.13.5, OS=Windows 11 (10.0.22621.1702/22H2/2022Update/SunValley2)
Snapdragon Compute Platform, 1 CPU, 8 logical and 8 physical cores
.NET SDK=7.0.302
[Host] : .NET 7.0.5 (7.0.523.17405), Arm64 RyuJIT AdvSIMD
Job-BBZNUX : .NET 7.0.5 (7.0.523.17405), Arm64 RyuJIT AdvSIMD
Job-IJWXNS : .NET 7.0.5 (7.0.523.17405), Arm64 RyuJIT AdvSIMD
| using BenchmarkDotNet.Attributes; | |
| using System.Runtime.CompilerServices; | |
| /// <summary> | |
| /// Small benchmark to decide until which size I can use FNV-1A vs System.HashCode. | |
| /// | |
| /// Verdict: If the size is lower than 28, FNV-1A is better on .NET 7 | |
| /// On .NET 6, it's more above 64 bytes. | |
| /// | |
| /// BenchmarkDotNet=v0.13.2, OS=Windows 11 (10.0.22000.1098/21H2) | |
| /// AMD Ryzen 9 5950X, 1 CPU, 32 logical and 16 physical cores |
| // Bench Windows Terminal with a wall of 24-bits RGB ANSI scrolling chars | |
| using System.Diagnostics; | |
| using System.Runtime.InteropServices; | |
| using System.Text; | |
| var width = Console.WindowWidth; | |
| var height = Console.WindowHeight; | |
| var total = width * height; | |
| if (RuntimeInformation.IsOSPlatform(OSPlatform.Windows)) |
| // Twitter discussion https://twitter.com/nietras1/status/1469740763760181250 | |
| // Benchmark of DuffsDevice mem copy vs plain Unsafe.CopyBlock | |
| // | Method | BufferSize | Mean | Error | StdDev | | |
| // |--------------- |----------- |-----------:|----------:|----------:| | |
| // | BenchDuffs | 8 | 2.627 ns | 0.0227 ns | 0.0213 ns | | |
| // | BenchBlockCopy | 8 | 2.608 ns | 0.0160 ns | 0.0150 ns | | |
| // | BenchDuffs | 32 | 2.889 ns | 0.0184 ns | 0.0172 ns | | |
| // | BenchBlockCopy | 32 | 2.394 ns | 0.0132 ns | 0.0123 ns | | |
| // | BenchDuffs | 128 | 6.083 ns | 0.0261 ns | 0.0231 ns | | |
| // | BenchBlockCopy | 128 | 3.904 ns | 0.0223 ns | 0.0208 ns | |