Vector{<T>|64|128|256|512}.Narrow with saturation

[gfoidl]

Summary

For SSE2 there's Sse2.PackSignedSaturate which packs the two vectors with saturation.
Such functionality is missing for the xplat-vectors -- i.e. Vector128.Narrow packs the two vectors, but cuts / masks off the values outside of the target range, e.g. for short -> sbyte bits above 0xFF is cut off resulting in (for me) unexpected values.

Maybe it's a bit astonishing that Narrow doesn't pack with saturation. This just bit me, and as it seems #73064 (comment) was similar.

There should be something like NarrowWithSaturation in order to don't need to fallback to Sse2 / AdvSimd.

API Proposal

namespace System.Runtime.Intrinsics;

public static class Vector64
{
    // Existing methods left out for brevity

    [CLSCompliant(false)]
    public static Vector64<sbyte> NarrowWithSaturation(Vector64<short> lower, Vector64<short> upper);

    public static Vector64<short> NarrowWithSaturation(Vector64<int> lower, Vector64<int> upper);

    public static Vector64<int> NarrowWithSaturation(Vector64<long> lower, Vector64<long> upper);

    [CLSCompliant(false)]
    public static Vector64<byte> NarrowWithSaturation(Vector64<ushort> lower, Vector64<ushort> upper);

    [CLSCompliant(false)]
    public static Vector64<ushort> NarrowWithSaturation(Vector64<uint> lower, Vector64<uint> upper);

    public static Vector64<uint> NarrowWithSaturation(Vector64<ulong> lower, Vector64<ulong> upper);

    public static Vector64<float> NarrowWithSaturation(Vector64<double> lower, Vector64<double> upper);
}

public static class Vector128
{
    // Existing methods left out for brevity

    [CLSCompliant(false)]
    public static Vector128<sbyte> NarrowWithSaturation(Vector128<short> lower, Vector128<short> upper);

    public static Vector128<short> NarrowWithSaturation(Vector128<int> lower, Vector128<int> upper);

    public static Vector128<int> NarrowWithSaturation(Vector128<long> lower, Vector128<long> upper);

    [CLSCompliant(false)]
    public static Vector128<byte> NarrowWithSaturation(Vector128<ushort> lower, Vector128<ushort> upper);

    [CLSCompliant(false)]
    public static Vector128<ushort> NarrowWithSaturation(Vector128<uint> lower, Vector128<uint> upper);

    public static Vector128<uint> NarrowWithSaturation(Vector128<ulong> lower, Vector128<ulong> upper);

    public static Vector128<float> NarrowWithSaturation(Vector128<double> lower, Vector128<double> upper);
}

public static class Vector256
{
    // Existing methods left out for brevity

    [CLSCompliant(false)]
    public static Vector256<sbyte> NarrowWithSaturation(Vector256<short> lower, Vector256<short> upper);

    public static Vector256<short> NarrowWithSaturation(Vector256<int> lower, Vector256<int> upper);

    public static Vector256<int> NarrowWithSaturation(Vector256<long> lower, Vector256<long> upper);

    [CLSCompliant(false)]
    public static Vector256<byte> NarrowWithSaturation(Vector256<ushort> lower, Vector256<ushort> upper);

    [CLSCompliant(false)]
    public static Vector256<ushort> NarrowWithSaturation(Vector256<uint> lower, Vector256<uint> upper);

    public static Vector256<uint> NarrowWithSaturation(Vector256<ulong> lower, Vector256<ulong> upper);

    public static Vector256<float> NarrowWithSaturation(Vector256<double> lower, Vector256<double> upper);
}

public static class Vector512
{
    // Existing methods left out for brevity

    [CLSCompliant(false)]
    public static Vector512<sbyte> NarrowWithSaturation(Vector512<short> lower, Vector512<short> upper);

    public static Vector512<short> NarrowWithSaturation(Vector512<int> lower, Vector512<int> upper);

    public static Vector512<int> NarrowWithSaturation(Vector512<long> lower, Vector512<long> upper);

    [CLSCompliant(false)]
    public static Vector512<byte> NarrowWithSaturation(Vector512<ushort> lower, Vector512<ushort> upper);

    [CLSCompliant(false)]
    public static Vector512<ushort> NarrowWithSaturation(Vector512<uint> lower, Vector512<uint> upper);

    public static Vector512<uint> NarrowWithSaturation(Vector512<ulong> lower, Vector512<ulong> upper);

    public static Vector512<float> NarrowWithSaturation(Vector512<double> lower, Vector512<double> upper);
}

[dotnet-issue-labeler]

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

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

Issue Details

---

For SSE2 there's Sse2.PackSignedSaturate which packs the two vectors with saturation.
Such functionality is missing (?) for the xplat-vectors -- i.e. Vector128.Narrow packs the two vectors, but cuts / masks off the values outside of the target range, e.g. for short -> sbyte bits above 0xFF is cut off resulting in (for me) unexpected values.

Maybe it's a bit astonishing that Narrow doesn't pack with saturation. This just bit me, and as it seems #73064 (comment) was similar.

Should there be something like NarrowWithSaturation in order to don't need to fallback to Sse2 / AdvSimd?

Author:	gfoidl
Assignees:	-
Labels:	area-System.Runtime.Intrinsics, untriaged
Milestone:	-

[tannergooding]

I'd be generally supportive of such new APIs, but this needs to be updated to be a "proper" API proposal for it to goto API review.

Basically, the proposal should minimally look like this: #73604

Provide a brief summary (which you already have) and then a section that lists the actual APIs being proposed (in this case just take all the Narrow APIs and expose NarrowSaturating or NarrowWithSaturation equivalents, not super picky on the name we can decide in API review).

This issue has been marked needs-author-action and may be missing some important information.

[gfoidl]

Updated to be a proposal -- just wanted to double-check with you before that (maybe there were good reasons why these APIs aren't available so far, etc.). Thanks.

[tannergooding]

Could you also update to cover Vector64 and Vector<T>? We want to try and maintain parity between all 4 types where possible.

Otherwise, LGTM and I can mark as api-ready-for-review.

[tannergooding]

Ah, and Vector512<T> as well since that is approved but not yet implemented.

[gfoidl]

Updated.