[API Proposal]: Add AVX10.2 MoveScalar APIs: VMOVD / VMOVW

[khushal1996]

Background and motivation

This API proposal introduces remaining API surface for AVX10v2 in .NET.
Spec doc - link
Previous Discussion - comment1, comment2, comment3, comment4

AVX10.2 original proposal - link

We will be adding these APIs to existing Avx10v2 API surface

Note: We are not doing the StoreScalar API since it already exists. link

API Proposal

// VMOVD xmm1, xmm2/m32
public static Vector128<uint> MoveScalar(Vector128<uint>)

// VMOVD xmm1, xmm2/m32
public static Vector128<int> MoveScalar(Vector128<int>)

// VMOVW xmm1, xmm2/m16
public static Vector128<ushort> MoveScalar(Vector128<ushort>)

// VMOVW xmm1, xmm2/m16
public static Vector128<short> MoveScalar(Vector128<short>)

// VMOVW xmm1/m16, xmm2
public static unsafe void StoreScalar(Int16* address, Vector128<Int16>)

// VMOVW xmm1/m16, xmm2
public static unsafe void StoreScalar(UInt16* address, Vector128<UInt16>)

API Usage

Vector128<int> v1 = Vector128.Create((int)someParam1);
if (Avx10v2.IsSupported()) {
  Vector128<int> v2 = Avx10v2.MoveScalar(v1);
  // etc
}

Alternative Designs

No response

Risks

No response

[dotnet-policy-service]

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

[khushal1996]

@tannergooding This proposal covers the remaining APIs for Avx10v2. Can you please review this based on our previous discussion?

Previous discussion - #109083 (comment)

[tannergooding]

LGTM. Could you add the spaces in the proposal, that is change:

- public static Vector128<uint>MoveScalar(Vector128<uint>)
+ public static Vector128<uint> MoveScalar(Vector128<uint>)

and do we need to mark the previously approved ConvertToVector128UInt32 APIs as removed here, so that should be included as part of the proposal?

[khushal1996]

We have not added ConvertToVector128UInt32 to Avx10v2 API surface. So even though we approved it, because it was under discussion, I left it out of the API surface. So we do not need to include removal of ConvertToVetor128UInt32.

I will wait for API review.

[khushal1996]

@tannergooding
For the following instruction

Shall we introduce

StoreScalar(Int16** address, Vector128<Int16>);
StoreScalar(UInt16* address, Vector128<UInt16>);

??

I see we already have the following Stores

public static unsafe void StoreScalar(int* address, Vector128<int> source)
public static unsafe void StoreScalar(uint* address, Vector128<uint> source)
public static unsafe void StoreScalar(long* address, Vector128<long> source)
public static unsafe void StoreScalar(ulong* address, Vector128<ulong> source)
public static unsafe void StoreScalar(float* address, Vector128<float> source)
public static unsafe void StoreScalar(double* address, Vector128<double> source)

But we do not have one for short/ushort. I am not sure why we did not discuss this before. Let me know if we should introduce the remaining 2 APIs as well here.

[tannergooding]

Yes, I'd go ahead and add them for consistency. They're not "required" since we can just pattern match the relevant local = MoveScalar(value), but since we already have the other overloads the consistency makes sense.

[khushal1996]

Yes, I'd go ahead and add them for consistency. They're not "required" since we can just pattern match the relevant local = MoveScalar(value), but since we already have the other overloads the consistency makes sense.

Updated the API doc.

[bartonjs]

Video

• Looks good as proposed

// VMOVD xmm1, xmm2/m32
public static Vector128<uint> MoveScalar(Vector128<uint> value);

// VMOVD xmm1, xmm2/m32
public static Vector128<int> MoveScalar(Vector128<int> value);

// VMOVW xmm1, xmm2/m16
public static Vector128<ushort> MoveScalar(Vector128<ushort> value);

// VMOVW xmm1, xmm2/m16
public static Vector128<short> MoveScalar(Vector128<short> value);

// VMOVW xmm1/m16, xmm2
public static unsafe void StoreScalar(Int16* address, Vector128<Int16> source);

// VMOVW xmm1/m16, xmm2
public static unsafe void StoreScalar(UInt16* address, Vector128<UInt16> source);