`TensorSpan<T>.FlattenTo` is 80x slower with the .NET 10 package

[MitchRazga]

Description

There is around an ~80x performance regression in System.Numerics.Tensors TensorSpan<T>.FlattenTo (and ReadOnlyTensorSpan<T>.FlattenTo).

Prior to 10.0.0-preview.4.25258.110 the implementation copied to the destination in blocks but now copies per-element.

Reproduction + benchmark solution here
Run FlattenTo.Before and FlattenTo.After console apps to benchmark against each package version.

Regression?

This performance issue was not present in the stable 9.0.10 release however the package was marked Experimental before .NET 10.

10.0.0-preview.3.25171.5 is the last version before the rewrite #114927 that introduced TensorOperation.

Data

Method	Mean	Error	StdDev	Allocated
FlattenToBefore	25.16 μs	0.136 μs	0.120 μs	48 B
FlattenToAfter	1.936 ms	0.0013 ms	0.0011 ms	-

Before-Benchmarks.log

After-Benchmarks.log

Analysis

release/9.0

runtime/src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/ReadOnlyTensorSpan.cs

Line 768 in 9c2fb4b

public void FlattenTo(scoped Span<T> destination)

release/10.0-rc2

runtime/src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/ReadOnlyTensorSpan_1.cs

Line 352 in 9a6bee3

public void FlattenTo(scoped Span<T> destination)

runtime/src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/TensorOperation.cs

Line 156 in 9a6bee3

public static void Invoke<TOperation, TArg, TResult>(in ReadOnlyTensorSpan<TArg> x, in Span<TResult> destination)

cc: @tannergooding

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-system-numerics-tensors
See info in area-owners.md if you want to be subscribed.

[tannergooding]

The .NET 9 preview had a number of bugs when dealing with various edge cases and slices.

The .NET 10 release was updated first and foremost to be correct. There was then known places where perf was regressed to achieve that and where we plan on getting a future patch up that improves it, which will be feasible since S.N.Tensors remains an "out of band" (rather than "in box") package.

The work to accelerate these so that we can copy by "row" or "longest contiguous region" is mostly setup and just not enabled or tested yet.

[MitchRazga]

@tannergooding just wondering if addressing this regression is on the radar for .NET 11?
You mentioned the contiguous region acceleration was mostly set up already but it doesn't look like it has been enabled in a preview yet.

[jeffhandley]

@MitchRazga Sorry, this didn't get attention during .NET 11. I'm marking as .NET 12.

[jeffhandley]

Related: #125663 tracks the broader work to avoid temporary buffer allocation in non-dense TensorSpan operations, which should cover the contiguous-region acceleration needed here. Both are milestoned 12.0.0.