[API Proposal]: AVX-512 pointer overloads for VPMOVZXBD and pals

[saucecontrol]

Background and motivation

This is basically the AVX-512 version of https://github.com/dotnet/runtime/issues/28868. The pointer overloads were missed for AVX-512 as well since they deviate from the C intrinsic surface.

API Proposal

public abstract class Avx512F : Avx2
{
    /// <summary>
    /// __m512i _mm512_cvtepi8_epi32 (__m128i a)
    ///   VPMOVSXBD zmm1 {k1}{z}, xmm2/m128
    /// </summary>
    public static Vector512<int> ConvertToVector512Int32(sbyte* address) => ConvertToVector512Int32(address);
    /// <summary>
    /// __m512i _mm512_cvtepu8_epi32 (__m128i a)
    ///   VPMOVZXBD zmm1 {k1}{z}, xmm2/m128
    /// </summary>
    public static Vector512<int> ConvertToVector512Int32(byte* address) => ConvertToVector512Int32(address);
    /// <summary>
    /// __m512i _mm512_cvtepi16_epi32 (__m256i a)
    ///   VPMOVSXWD zmm1 {k1}{z}, ymm2/m256
    /// </summary>
    public static Vector512<int> ConvertToVector512Int32(short* address) => ConvertToVector512Int32(address);
    /// <summary>
    /// __m512i _mm512_cvtepu16_epi32 (__m256i a)
    ///   VPMOVZXWD zmm1 {k1}{z}, ymm2/m256
    /// </summary>
    public static Vector512<int> ConvertToVector512Int32(ushort* address) => ConvertToVector512Int32(address);
    /// <summary>
    /// __m512i _mm512_cvtepi8_epi64 (__m128i a)
    ///   VPMOVSXBQ zmm1 {k1}{z}, xmm2/m64
    /// </summary>
    public static Vector512<long> ConvertToVector512Int64(sbyte* address) => ConvertToVector512Int64(address);
    /// <summary>
    /// __m512i _mm512_cvtepu8_epi64 (__m128i a)
    ///   VPMOVZXBQ zmm1 {k1}{z}, xmm2/m64
    /// </summary>
    public static Vector512<long> ConvertToVector512Int64(byte* address) => ConvertToVector512Int64(address);
    /// <summary>
    /// __m512i _mm512_cvtepi16_epi64 (__m128i a)
    ///   VPMOVSXWQ zmm1 {k1}{z}, xmm2/m128
    /// </summary>
    public static Vector512<long> ConvertToVector512Int64(short* address) => ConvertToVector512Int64(address);
    /// <summary>
    /// __m512i _mm512_cvtepu16_epi64 (__m128i a)
    ///   VPMOVZXWQ zmm1 {k1}{z}, xmm2/m128
    /// </summary>
    public static Vector512<long> ConvertToVector512Int64(ushort* address) => ConvertToVector512Int64(address);
    /// <summary>
    /// __m512i _mm512_cvtepi32_epi64 (__m256i a)
    ///   VPMOVSXDQ zmm1 {k1}{z}, ymm2/m256
    /// </summary>
    public static Vector512<long> ConvertToVector512Int64(int* address) => ConvertToVector512Int64(address);
    /// <summary>
    /// __m512i _mm512_cvtepu32_epi64 (__m256i a)
    ///   VPMOVZXDQ zmm1 {k1}{z}, ymm2/m256
    /// </summary>
    public static Vector512<long> ConvertToVector512Int64(uint* address) => ConvertToVector512Int64(address);
}

public abstract class Avx512BW : Avx512F
{
    /// <summary>
    /// __m512i _mm512_cvtepi8_epi16 (__m256i a)
    ///   VPMOVSXBW zmm1 {k1}{z}, ymm2/m256
    /// </summary>
    public static Vector512<short> ConvertToVector512Int16(sbyte* address) => ConvertToVector512Int16(address);
    /// <summary>
    /// __m512i _mm512_cvtepu8_epi16 (__m256i a)
    ///   VPMOVZXBW zmm1 {k1}{z}, ymm2/m256
    /// </summary>
    public static Vector512<short> ConvertToVector512Int16(byte* address) => ConvertToVector512Int16(address);
}

API Usage

var zmm0 = Avx512F.ConvertToVector512Int32(ptr);

Alternative Designs

vpmovsxbq and vpmovzxbq consume 64 bits, and as such are the only instructions that can't be handled by containment of a Vector128 or Vector256 load without an implied overread. Those could be exposed without the others if containment were handled properly.

Risks

No response

[dotnet-policy-service[bot]]

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