|
// Licensed to the .NET Foundation under one or more agreements.
// The .NET Foundation licenses this file to you under the MIT license.
using System.Diagnostics;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.Arm;
using System.Runtime.Intrinsics.X86;
namespace System.Numerics.Tensors
{
public static partial class TensorPrimitives
{
/// <summary>Computes the element-wise result of <c>(<paramref name="x" /> * <paramref name="y" />) + <paramref name="addend" /></c> for the specified tensors of numbers.</summary>
/// <param name="x">The first tensor, represented as a span.</param>
/// <param name="y">The second tensor, represented as a span.</param>
/// <param name="addend">The third tensor, represented as a span.</param>
/// <param name="destination">The destination tensor, represented as a span.</param>
/// <exception cref="ArgumentException">Length of <paramref name="x" /> must be same as length of <paramref name="y" /> and length of <paramref name="addend" />.</exception>
/// <exception cref="ArgumentException">Destination is too short.</exception>
/// <exception cref="ArgumentException"><paramref name="x"/> and <paramref name="destination"/> reference overlapping memory locations and do not begin at the same location.</exception>
/// <exception cref="ArgumentException"><paramref name="y"/> and <paramref name="destination"/> reference overlapping memory locations and do not begin at the same location.</exception>
/// <exception cref="ArgumentException"><paramref name="addend"/> and <paramref name="destination"/> reference overlapping memory locations and do not begin at the same location.</exception>
/// <remarks>
/// <para>
/// This method effectively computes <c><paramref name="destination" />[i] = (<paramref name="x" />[i] * <paramref name="y" />[i]) + <paramref name="addend" />[i]</c>.
/// </para>
/// <para>
/// If either of the element-wise input values is equal to <see cref="IFloatingPointIeee754{TSelf}.NaN"/>, the resulting element-wise value is also NaN.
/// </para>
/// <para>
/// This computes (<paramref name="x"/> * <paramref name="y"/>) as if to infinite precision, adds <paramref name="addend"/> to that result as if to
/// infinite precision, and finally rounds to the nearest representable value. This differs from the non-fused sequence which would compute
/// (<paramref name="x"/> * <paramref name="y"/>) as if to infinite precision, round the result to the nearest representable value, add <paramref name="addend"/> to the
/// rounded result as if to infinite precision, and finally round to the nearest representable value.
/// </para>
/// </remarks>
public static void FusedMultiplyAdd<T>(ReadOnlySpan<T> x, ReadOnlySpan<T> y, ReadOnlySpan<T> addend, Span<T> destination)
where T : IFloatingPointIeee754<T>
{
if (typeof(T) == typeof(Half) && TryTernaryInvokeHalfAsInt16<T, FusedMultiplyAddOperator<float>>(x, y, addend, destination))
{
return;
}
InvokeSpanSpanSpanIntoSpan<T, FusedMultiplyAddOperator<T>>(x, y, addend, destination);
}
/// <summary>Computes the element-wise result of <c>(<paramref name="x" /> * <paramref name="y" />) + <paramref name="addend" /></c> for the specified tensors of numbers.</summary>
/// <param name="x">The first tensor, represented as a span.</param>
/// <param name="y">The second tensor, represented as a span.</param>
/// <param name="addend">The third tensor, represented as a scalar.</param>
/// <param name="destination">The destination tensor, represented as a span.</param>
/// <exception cref="ArgumentException">Length of <paramref name="x" /> must be same as length of <paramref name="y" />.</exception>
/// <exception cref="ArgumentException">Destination is too short.</exception>
/// <exception cref="ArgumentException"><paramref name="x"/> and <paramref name="destination"/> reference overlapping memory locations and do not begin at the same location.</exception>
/// <exception cref="ArgumentException"><paramref name="y"/> and <paramref name="destination"/> reference overlapping memory locations and do not begin at the same location.</exception>
/// <remarks>
/// <para>
/// This method effectively computes <c><paramref name="destination" />[i] = (<paramref name="x" />[i] * <paramref name="y" />[i]) + <paramref name="addend" /></c>.
/// It corresponds to the <c>axpy</c> method defined by <c>BLAS1</c>.
/// </para>
/// <para>
/// If either of the element-wise input values is equal to <see cref="IFloatingPointIeee754{TSelf}.NaN"/>, the resulting element-wise value is also NaN.
/// </para>
/// <para>
/// This computes (<paramref name="x"/> * <paramref name="y"/>) as if to infinite precision, adds <paramref name="addend"/> to that result as if to
/// infinite precision, and finally rounds to the nearest representable value. This differs from the non-fused sequence which would compute
/// (<paramref name="x"/> * <paramref name="y"/>) as if to infinite precision, round the result to the nearest representable value, add <paramref name="addend"/> to the
/// rounded result as if to infinite precision, and finally round to the nearest representable value.
/// </para>
/// </remarks>
public static void FusedMultiplyAdd<T>(ReadOnlySpan<T> x, ReadOnlySpan<T> y, T addend, Span<T> destination)
where T : IFloatingPointIeee754<T>
{
if (typeof(T) == typeof(Half) && TryTernaryInvokeHalfAsInt16<T, FusedMultiplyAddOperator<float>>(x, y, addend, destination))
{
return;
}
InvokeSpanSpanScalarIntoSpan<T, FusedMultiplyAddOperator<T>>(x, y, addend, destination);
}
/// <summary>Computes the element-wise result of <c>(<paramref name="x" /> * <paramref name="y" />) + <paramref name="addend" /></c> for the specified tensors of numbers.</summary>
/// <param name="x">The first tensor, represented as a span.</param>
/// <param name="y">The second tensor, represented as a scalar.</param>
/// <param name="addend">The third tensor, represented as a span.</param>
/// <param name="destination">The destination tensor, represented as a span.</param>
/// <exception cref="ArgumentException">Length of <paramref name="x" /> must be same as length of <paramref name="addend" />.</exception>
/// <exception cref="ArgumentException">Destination is too short.</exception>
/// <exception cref="ArgumentException"><paramref name="x"/> and <paramref name="destination"/> reference overlapping memory locations and do not begin at the same location.</exception>
/// <exception cref="ArgumentException"><paramref name="addend"/> and <paramref name="destination"/> reference overlapping memory locations and do not begin at the same location.</exception>
/// <remarks>
/// <para>
/// This method effectively computes <c><paramref name="destination" />[i] = (<paramref name="x" />[i] * <paramref name="y" />) + <paramref name="addend" />[i]</c>.
/// </para>
/// <para>
/// If either of the element-wise input values is equal to <see cref="IFloatingPointIeee754{TSelf}.NaN"/>, the resulting element-wise value is also NaN.
/// </para>
/// <para>
/// This computes (<paramref name="x"/> * <paramref name="y"/>) as if to infinite precision, adds <paramref name="addend"/> to that result as if to
/// infinite precision, and finally rounds to the nearest representable value. This differs from the non-fused sequence which would compute
/// (<paramref name="x"/> * <paramref name="y"/>) as if to infinite precision, round the result to the nearest representable value, add <paramref name="addend"/> to the
/// rounded result as if to infinite precision, and finally round to the nearest representable value.
/// </para>
/// </remarks>
public static void FusedMultiplyAdd<T>(ReadOnlySpan<T> x, T y, ReadOnlySpan<T> addend, Span<T> destination)
where T : IFloatingPointIeee754<T>
{
if (typeof(T) == typeof(Half) && TryTernaryInvokeHalfAsInt16<T, FusedMultiplyAddOperator<float>>(x, y, addend, destination))
{
return;
}
InvokeSpanScalarSpanIntoSpan<T, FusedMultiplyAddOperator<T>>(x, y, addend, destination);
}
/// <summary>(x * y) + z</summary>
private readonly struct FusedMultiplyAddOperator<T> : ITernaryOperator<T> where T : IFloatingPointIeee754<T>
{
public static bool Vectorizable => true;
public static T Invoke(T x, T y, T z) => T.FusedMultiplyAdd(x, y, z);
public static Vector128<T> Invoke(Vector128<T> x, Vector128<T> y, Vector128<T> z)
{
if (typeof(T) == typeof(double))
{
return Vector128.FusedMultiplyAdd(x.AsDouble(), y.AsDouble(), z.AsDouble()).As<double, T>();
}
else
{
Debug.Assert(typeof(T) == typeof(float));
return Vector128.FusedMultiplyAdd(x.AsSingle(), y.AsSingle(), z.AsSingle()).As<float, T>();
}
}
public static Vector256<T> Invoke(Vector256<T> x, Vector256<T> y, Vector256<T> z)
{
if (typeof(T) == typeof(double))
{
return Vector256.FusedMultiplyAdd(x.AsDouble(), y.AsDouble(), z.AsDouble()).As<double, T>();
}
else
{
Debug.Assert(typeof(T) == typeof(float));
return Vector256.FusedMultiplyAdd(x.AsSingle(), y.AsSingle(), z.AsSingle()).As<float, T>();
}
}
public static Vector512<T> Invoke(Vector512<T> x, Vector512<T> y, Vector512<T> z)
{
if (typeof(T) == typeof(double))
{
return Vector512.FusedMultiplyAdd(x.AsDouble(), y.AsDouble(), z.AsDouble()).As<double, T>();
}
else
{
Debug.Assert(typeof(T) == typeof(float));
return Vector512.FusedMultiplyAdd(x.AsSingle(), y.AsSingle(), z.AsSingle()).As<float, T>();
}
}
}
}
}
|