Vectorize TensorPrimitives.PopCount

[stephentoub]

Contributes to #97193

[ghost]

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

Issue Details

---

Use popcount hardware intrinsics to vectorize TensorPrimitives.PopCount for sizeof(T) == 1 on platforms that have intrinsics.

I tried doing this more generally on other platforms with:

Vector128<byte> lookup = Vector128.Create((byte)0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
Vector128<byte> cnt1 = Vector128.Shuffle(lookup, x.AsByte() & Vector128.Create((byte)0xF));
Vector128<byte> cnt2 = Vector128.Shuffle(lookup, (x.AsByte() >> 4) & Vector128.Create((byte)0xF));
return (cnt1 + cnt2).As<byte, T>();

and the equivalent for Vector256/512, but it ended up being ~5x slower than just using the scalar byte.PopCount for each element. I also tried using this with SumAbsoluteDifferences in support of sizeof(T) == 8, and it ended up being an order of magnitude worse.

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