Tensor primitives divide int32

[alexcovington]

Add a vector path for int in TensorPrimitives.Divide.

Improves performance in microbenchmarks:

| Method        | Job        | Toolchain        | BufferLength | Mean       | Error    | StdDev   | Median     | Min        | Max        | Ratio | Allocated | Alloc Ratio |
|-------------- |----------- |----------------- |------------- |-----------:|---------:|---------:|-----------:|-----------:|-----------:|------:|----------:|------------:|
| Divide_Vector | Job-SUKFCS | Base             | 128          |   174.4 ns |  0.83 ns |  0.69 ns |   174.2 ns |   173.5 ns |   176.0 ns |  1.00 |         - |          NA |
| Divide_Vector | Job-OOXWSE | Diff             | 128          |   111.5 ns |  0.56 ns |  0.50 ns |   111.5 ns |   110.7 ns |   112.3 ns |  0.64 |         - |          NA |
|               |            |                  |              |            |          |          |            |            |            |       |           |             |
| Divide_Scalar | Job-SUKFCS | Base             | 128          |   138.9 ns |  0.88 ns |  0.78 ns |   138.5 ns |   137.8 ns |   140.5 ns |  1.00 |         - |          NA |
| Divide_Scalar | Job-OOXWSE | Diff             | 128          |   104.8 ns |  0.50 ns |  0.42 ns |   104.7 ns |   104.3 ns |   105.8 ns |  0.75 |         - |          NA |
|               |            |                  |              |            |          |          |            |            |            |       |           |             |
| Divide_Vector | Job-SUKFCS | Base             | 3079         | 4,038.8 ns | 13.47 ns | 11.94 ns | 4,034.4 ns | 4,025.5 ns | 4,062.2 ns |  1.00 |         - |          NA |
| Divide_Vector | Job-OOXWSE | Diff             | 3079         | 1,358.5 ns |  5.78 ns |  4.83 ns | 1,356.7 ns | 1,351.8 ns | 1,367.6 ns |  0.34 |         - |          NA |
|               |            |                  |              |            |          |          |            |            |            |       |           |             |
| Divide_Scalar | Job-SUKFCS | Base             | 3079         | 3,315.1 ns | 12.55 ns | 11.12 ns | 3,313.0 ns | 3,301.0 ns | 3,337.9 ns |  1.00 |         - |          NA |
| Divide_Scalar | Job-OOXWSE | Diff             | 3079         | 1,294.8 ns |  7.84 ns |  6.95 ns | 1,293.1 ns | 1,286.2 ns | 1,309.0 ns |  0.39 |         - |          NA |

[huoyaoyuan]

All of these shouldn't be written in x86 intrinsics. Comparison, conversion and divide are all having cross platform intrinsic available.

[alexcovington]

Comparison and divide have appropriate cross platform instrinsics, but the conversion uses x86 intrinsics that convert Int32 -> Double and Double -> Int32. I did not see a cross platform convert that supports that case, but maybe I missed it? If there is a cross platform convert that supports the conversion, I'm happy to switch to that.

[tannergooding]

It would be for ConvertToDouble(WidenLower(vectorOfInt32)) which we don't currently optimize but would like to pattern match and have implicitly emit the more optimal Sse2.ConvertToVector128Double API

Correspondingly we'd want something like Create(ConvertToDouble(WidenLower(vectorOfInt32)), ConvertToDouble(WidenUpper(vectorOfInt32))) or one of the other patterns to work for Avx.ConvertToVector256Double

I think its fine to write the xplat code here first and then log the issue for optimizing it, which should also help prioritize getting that optimization implemented and ensure that it lights up on more hardware by default.

[stephentoub]

Why is this limited to .NET 10? Are the APIs being used newly introduced only in .NET 10?

[alexcovington]

I was running into issues with FloatRoundingMode not being available in .NET 8.0.

[stephentoub]

That was added in 9, not 10.

[tannergooding]

This can be the following on .NET 10:

if (Vector128.Any(y.AsInt32(), 0))
{
    ThrowHelper.DivideByZeroException();
}

or for .NET 8/9:

if (Vector128.EqualsAny(y.AsInt32(), Vector128<int>.Zero))
{
    ThrowHelper.DivideByZeroException();
}

[tannergooding]

I think it'd be better to use the xplat algorithm around ConvertToDouble(WidenLower(x.AsInt32())) / ConvertToDouble(WidenLower(y.AsInt32())) so it can lightup for .NET 8/9.

We should then ideally accelerate Vector128.operator / on .NET 10 and just use that there, so it can be directly handled and lightup anyone else using operator / as well.