Add support for copying from Vector<T> to Span<T>

[surban]

Currently Vector<T> can be constructed from a Span<T>, but writing the result of a vectorized operation into a Span<T> is impossible, since the Vector<T>.CopyTo method lacks the necessary overload.

Rationale and usage

The following code computes the sum of two arrays using SIMD operations and stores it in the trgt parameter.

        static void VectorizedArrayAdd(double[] trgt, double[] x, double[] y)
        {
            Debug.Assert(x.Length == y.Length && x.Length == trgt.Length);
            Debug.Assert(x.Length % Vector<double>.Count == 0);

            for (int i=0; i < x.Length; i += Vector<double>.Count)
            {
                var xv = new Vector<double>(x, i);
                var yv = new Vector<double>(y, i);
                var tv = xv + yv;
                tv.CopyTo(trgt, i);
            }
        }

Converting this code to accept Span<double> instead of double[] makes sense, for example to support operation on slices of arrays or on external memory. A first conversion attempt leads to the following code.

        static void VectorizedSpanAdd(Span<double> trgt, Span<double> x, Span<double> y)
        {
            Debug.Assert(x.Length == y.Length && x.Length == trgt.Length);
            Debug.Assert(x.Length % Vector<double>.Count == 0);

            for (int i = 0; i < x.Length; i += Vector<double>.Count)
            {
                var xs = x.Slice(i, Vector<double>.Count);
                var xv = new Vector<double>(xs);

                var ys = y.Slice(i, Vector<double>.Count);
                var yv = new Vector<double>(ys);

                var tv = xv + yv;
                var ts = trgt.Slice(i, Vector<double>.Count);
                tv.CopyTo(ts); // ERROR: Argument 1: cannot convert from 'System.Span<double>' to 'double[]' 
            }
        }

However, this fails to compile since Vector<double>.CopyTo(Span<double> x) does not exist.

Propsed API

Extend Vector<T> as follows.

public struct Vector<T> : IEquatable<System.Numerics.Vector<T>>, IFormattable where T : struct
{
    public void CopyTo(Span<T> destination);
}

Details

• Support for constructing a Vector<T> from a Span<T> has been added in Issue #24343 Vector Ctor using Span corefx#26499, but without the ability to get the result back out into a Span<T> the usability of this feature is limited.

[scalablecory]

There's a standard feature for this already:

Span<byte> span;
Span<Vector<byte>> vector = MemoryMarshal.Cast<Vector<byte>, byte>(span);
vector[0] = ...;

We may need better documentation of this.

[tannergooding]

We already added this in: dotnet/coreclr#23333

[tannergooding]

It will be available in .NET Core 3.0 and should already be available in the latest public preview.