Guides authoring and reviewing SIMD and hardware-intrinsics code in dotnet/runtime.
npx skills add https://github.com/dotnet/runtime --skill vectorization
Please help me install this Agent Skill. Skill: vectorization Source: https://github.com/dotnet/runtime/tree/main/.github/skills/vectorization Command: npx skills add https://github.com/dotnet/runtime --skill vectorization
Writing or reviewing vectorized .NET code is error-prone: remainder handling, unsigned offset underflow, GC holes from stray references, and untested fallback paths cause subtle bugs. This Skill distills the dotnet/runtime vectorization guidelines into concrete authoring, testing, and review checklists.
Ask the AI to review or write a vectorized implementation of your algorithm using Vector128<T> with proper remainder handling and hardware-acceleration fallbacks.
High-intent search queries and answers about installing and using this skill.
Start with Vector128<T>, the widest API accelerated on the broadest hardware, and structure code from that width down to a scalar fallback. Guard each path with IsHardwareAccelerated and Vector128<T>.IsSupported, then handle the remainder by reprocessing the last full vector.
Reprocess the last full vector's worth of elements, overlapping what the loop already did. For idempotent operations like searches, fold the overlap in directly; for non-idempotent operations like sums, mask the overlap to the identity with ConditionalSelect.
Prefer the cross-platform Vector128/Vector256 APIs, which lower to optimal instructions per target. Drop to System.Runtime.Intrinsics.X86/Arm/Wasm only when a specific instruction measurably beats the portable form, guarded by the class's IsSupported check.
Run the test suite with DOTNET_EnableAVX2=0 to disable Vector256 paths and DOTNET_EnableHWIntrinsic=0 to force the software fallback. Use BoundedMemory.Allocate<T> to place a no-access page after buffers so out-of-bounds reads fault instead of silently succeeding.
Common causes are nuint underflow when subtracting the vector count from a short buffer length, and transient references pointing outside their buffer during backwards iteration, which creates GC holes. Always check buffer length before offset arithmetic and never let a ref stray outside its buffer.
Avoid hand-rolling when higher-level APIs already vectorize the operation, such as Span<T> and string methods, TensorPrimitives, or LINQ operators like Sum, Max, Min, and Average. Also skip vectorization without benchmark evidence that it pays off, since small buffers can be slower.
HomericIntelligenceMeasure function execution time and memory usage to compare implementations.
sayedihashimiReplace scalar loops with SIMD intrinsics or TensorPrimitives in .NET 8+.
dotnetReview dotnet/runtime code changes for correctness, performance, and conventions.
ojrojasOptimize .NET 8+ math loops with SIMD intrinsics and TensorPrimitives.
