TensorPrimitives.Exp<double> returns negative infinity instead of 0 for doubles between approximately -709.5 and -709.7

[BarionLP]

Description

TensorPrimitives.Exp returns negative infinity instead of 0 for doubles between approximately -709.5 and -709.7. When using Math.Exp() it correctly returns 0. when using float also everything is fine. I'm using System.Numerics.Tensors version 9.0.0-rc.1.24431.7

Reproduction Steps

double value = -709.5; //with -709.4 and -709.8 everything works fine
double[] test = [value, value, value, value, value];
double[] destination = new double[test.Length];
TensorPrimitives.Exp<double>(test, destination);

Console.WriteLine(string.Join(' ', destination));

Expected behavior

the above code should output 0 or values close to 0

Actual behavior

Output on my machine: -Infinity -Infinity -Infinity -Infinity -Infinity

Regression?

i haven't tried in the preview versions

Known Workarounds

No response

Configuration

.NET 9 rc1 Windows 10 x64 (AMD) Simd uses 256bits (4 doubles)

Other information

No response

[dotnet-policy-service[bot]]

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

[BarionLP]

Same on my Surface Pro X (ARM64, 128bit simd) and an old HP ProBook (Intel Core i7 x64, 256bit simd)

[stephentoub]

cc: @tannergooding

[stephentoub]

Something about the vectorized implementation of Exp, e.g. this:

using System.Numerics.Tensors;

for (int length = 1; length <= 2; length++)
{
    Console.WriteLine($"Length: {length}");
    double[] values = new double[length];
    double[] results = new double[values.Length];

    values.AsSpan().Fill(-709.7);
    TensorPrimitives.Exp<double>(values, results);

    Console.WriteLine(Math.Exp(values[0]));
    Console.WriteLine(results[0]);

    Console.WriteLine();
}

outputs this (vectorization will kick in for at least two doubles):

Length: 1
6.04235438695844E-309
6.04235438695844E-309

Length: 2
6.04235438695844E-309
-∞

[BarionLP]

I would appreciate if you can explain what happened here. I love low level performance and would like to learn more about it.

[dotnet-policy-service[bot]]

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

[stephentoub]

I would appreciate if you can explain what happened here.

The code has a fallback code path for edge case values, which these are. The vectorized implementation of that fallback is producing different answers than the scalar implementation.

[BarionLP]

can confirm fix in .NET 9 RC 2