TensorPrimitives.IndexOfXx methods sometimes return the wrong index in the face of equality

[stephentoub]

Description

When there are multiple equal elements that could have their index returned, the one with the smallest index is supposed to be returned. But it isn't always.

Reproduction Steps

using System.Numerics.Tensors;

ReadOnlySpan<float> floats = [10f, -20f, -30f, 1f, 1f, -40f];
Console.WriteLine(TensorPrimitives.IndexOfMinMagnitude(floats));

or

using System.Numerics.Tensors;

ReadOnlySpan<float> floats = [10f, 20f, 30f, 1f, 1f, 40f];
Console.WriteLine(TensorPrimitives.IndexOfMin(floats));

Expected behavior

Those should both print 3.

Actual behavior

They print 4.

Regression?

No

Known Workarounds

No response

Configuration

System.Numerics.Tensors v8.0.0 nuget package

Other information

No response

[ghost]

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

Issue Details

---

Description

When there are multiple equal elements that could have their index returned, the one with the smallest index is supposed to be returned. But it isn't always.

Reproduction Steps

using System.Numerics.Tensors;

ReadOnlySpan<float> floats = [10f, -20f, -30f, 1f, 1f, -40f];
Console.WriteLine(TensorPrimitives.IndexOfMinMagnitude(floats));

Expected behavior

That should print 3.

Actual behavior

It prints 4.

Regression?

No

Known Workarounds

No response

Configuration

System.Numerics.Tensors v8.0.0 nuget package

Other information

No response

Author:	stephentoub
Assignees:	-
Labels:	area-System.Numerics, untriaged
Milestone:	-

[tannergooding]

I think the bug is here in the equality handling check: https://source.dot.net/#System.Numerics.Tensors/System/Numerics/Tensors/netcore/TensorPrimitives.netcore.cs,10125

We'll get down to the MinAcross logic and end up comparing result = [10, 1, 1, 1]; resultIndex = [0, 3, 4, 3] to tmpResult = [1, 1, 10, 1]; tmpIndex = [4, 3, 0, 3].

The lessThanMask will be [false, false, true, false], the equalMask will be [false, true, false, true], the negativeMask will be [false, false, false, false], and the lessThanIndexMask will be [true, false, false, false].

We will then modify the lessThanMask using the other masks, giving us [false, true, true, true]. This then picks result = [1, 1, 1, 1]; resultIndex = [4, 3, 4, 3] and will compare that to tmpResult = [1, 1, 1, 1]; tmpIndex = [3, 4, 3, 4].

The lessThanMask will be [false, false, false, false], the equalMask will be [true, true, true, true], the negativeMask will be [false, false, false, false], and the lessThanIndexMask will be [false, true, false, true].

We will then modify the lessThanMask using the other masks, giving us [true, true, true, true] which causes us to preserve result = [1, 1, 1, 1]; resultIndex = [4, 3, 4, 3] and return 4 since its the first entry.

---

So the failure here looks to come about because lessThanIndexMask is only used when there is some negative in result. We'd need to modify this to use it in more than just this scenario to ensure that the lesser index is picked.

What we have now is effectively:

Vector128<float> currentNegativeMask = IsNegative(current);

// Take the entry from current if it is equal to result and negative
lessThanMask |= Vector128.AndNot(equalMask, currentNegativeMask);

// Take the entry from result if it is equal to current and negative and its index is lesser
lessThanMask |= equalMask & IsNegative(result) & lessThanIndexMask;

What we need to say is:

• If both are equal and have the same sign, pick the lesser index
• If both are equal and only one is negative, pick the negative one

This is required to ensure +0 vs -0 is handled correctly.