[Perf -50%] System.Numerics.Tests.Perf_Vector2.LengthSquaredBenchmark

[DrewScoggins]

Run Information

Architecture	x64
OS	Windows 10.0.18362
Changes	diff

Regressions in System.Numerics.Tests.Perf_Vector2

Benchmark	Baseline	Test	Test/Base	Modality	Baseline Outlier
LengthSquaredBenchmark	0.34 ns	0.51 ns	1.50		False

Historical Data in Reporting System

Repro

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f netcoreapp5.0 --filter 'System.Numerics.Tests.Perf_Vector2*';

Details

Histogram

LengthSquaredBenchmark

[0.305 ; 0.369) | @@@@@@@@@@@@@@@
[0.369 ; 0.434) | @
[0.434 ; 0.510) | @@@@@@@@@@@@@@@@@@@@@@@
[0.510 ; 0.539) | 
[0.539 ; 0.604) | @@@@@@@@@@@@@@@

Docs

Profiling workflow for dotnet/runtime repository
Benchmarking workflow for dotnet/runtime repository

category:cq
theme:vector-codegen
skill-level:intermediate
cost:medium

Tagging subscribers to this area: @tannergooding
Notify danmosemsft if you want to be subscribed.

[tannergooding]

The regression is due to Vector2.DotProduct not being a SimdAsHWIntrinsic yet and so there is some register shuffling when transitioning from the legacy GT_SIMD to the newer GT_HWINTRINSIC as the RMW and VEX support between them isn't the same.

[AndyAyersMS]

@tannergooding is this covered by the PR above?

Marking this as codegen.

[tannergooding]

Yes. This was due to some, but not all, of the System.Numerics intrinsics being transitioned to use HWIntrinsic internally.

The GT_HWINTRINSIC nodes are VEX aware while the GT_SIMD nodes are not, which caused a few additional moves to be inserted. The above PR transitions Dot and the other big functions to also use HWIntrinsics which allows the additional moves to be elided.