260 lines
10 KiB
C#
260 lines
10 KiB
C#
using System.Runtime.Intrinsics.X86;
|
||
|
||
namespace QWERTYkez.Mensura.Extensions;
|
||
|
||
internal static partial class CollectionsRootOfSquareExtensions
|
||
{
|
||
// === SqrtCore === SIMD
|
||
internal static unsafe void SqrtCore<T, R>(this ReadOnlySpan<T> units, int len, Span<R> destination)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
ReadOnlySpan<double> srcDouble = MemoryMarshal.Cast<T, double>(units);
|
||
Span<double> dstDouble = MemoryMarshal.Cast<R, double>(destination);
|
||
|
||
int i = 0;
|
||
ref double srcRef = ref MemoryMarshal.GetReference(srcDouble);
|
||
ref double dstRef = ref MemoryMarshal.GetReference(dstDouble);
|
||
|
||
// 1. ПУТЬ AVX2 / AVX (x64 Процессоры) — обрабатываем по 4 элемента double за такт
|
||
if (Avx.IsSupported && len >= 4)
|
||
{
|
||
int simdEnd = len & ~3;
|
||
for (; i < simdEnd; i += 4)
|
||
{
|
||
ref double currentSrc = ref Unsafe.Add(ref srcRef, i);
|
||
ref double currentDst = ref Unsafe.Add(ref dstRef, i);
|
||
|
||
double* pSrc = (double*)Unsafe.AsPointer(ref currentSrc);
|
||
double* pDst = (double*)Unsafe.AsPointer(ref currentDst);
|
||
|
||
// Используем Load вместо LoadVector256 (Load безопасен к невыровненной по 32-байтам памяти)
|
||
var v = Avx.LoadVector256(pSrc);
|
||
var sqrtV = Avx.Sqrt(v);
|
||
Avx.Store(pDst, sqrtV);
|
||
}
|
||
}
|
||
// 2. КРОССПЛАТФОРМЕННЫЙ VECTOR (ARM64 / Apple Silicon / NEON / SIMD Fallback)
|
||
else if (Vector.IsHardwareAccelerated && len >= Vector<double>.Count)
|
||
{
|
||
int vCount = Vector<double>.Count;
|
||
int simdEnd = len & ~(vCount - 1);
|
||
|
||
for (; i < simdEnd; i += vCount)
|
||
{
|
||
var v = new Vector<double>(srcDouble.Slice(i, vCount));
|
||
var sqrtV = Vector.SquareRoot(v);
|
||
sqrtV.CopyTo(dstDouble.Slice(i, vCount));
|
||
}
|
||
}
|
||
|
||
// 3. ХВОСТ МАССИВА (или работа на старых CPU)
|
||
for (; i < len; i++)
|
||
{
|
||
Unsafe.Add(ref dstRef, i) = Math.Sqrt(Unsafe.Add(ref srcRef, i));
|
||
}
|
||
}
|
||
internal static void SqrtCore<T, R>(this ReadOnlySpan<T?> units, int len, Span<R?> destination)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
ref var srcRef = ref MemoryMarshal.GetReference(units);
|
||
ref var dstRef = ref MemoryMarshal.GetReference(destination);
|
||
|
||
int i = 0;
|
||
int unrollEnd = len & ~3;
|
||
|
||
// ОСНОВНОЙ ЦИКЛ: Конвейерное чтение по 4 элемента (Развертка цикла повышает ILP процессора)
|
||
for (; i < unrollEnd; i += 4)
|
||
{
|
||
ref readonly T? u0 = ref Unsafe.Add(ref srcRef, i);
|
||
ref readonly T? u1 = ref Unsafe.Add(ref srcRef, i + 1);
|
||
ref readonly T? u2 = ref Unsafe.Add(ref srcRef, i + 2);
|
||
ref readonly T? u3 = ref Unsafe.Add(ref srcRef, i + 3);
|
||
|
||
ref var d0 = ref Unsafe.Add(ref dstRef, i);
|
||
ref var d1 = ref Unsafe.Add(ref dstRef, i + 1);
|
||
ref var d2 = ref Unsafe.Add(ref dstRef, i + 2);
|
||
ref var d3 = ref Unsafe.Add(ref dstRef, i + 3);
|
||
|
||
d0 = Math.Sqrt(u0.Protected()).ToUnit<R>();
|
||
d1 = Math.Sqrt(u1.Protected()).ToUnit<R>();
|
||
d2 = Math.Sqrt(u2.Protected()).ToUnit<R>();
|
||
d3 = Math.Sqrt(u3.Protected()).ToUnit<R>();
|
||
}
|
||
|
||
// ХВОСТ ЦИКЛА: Остаток элементов
|
||
for (; i < len; i++)
|
||
{
|
||
ref readonly T? unit = ref Unsafe.Add(ref srcRef, i);
|
||
ref var dst = ref Unsafe.Add(ref dstRef, i);
|
||
dst = Math.Sqrt(unit.Protected()).ToUnit<R>();
|
||
}
|
||
}
|
||
|
||
// === ReadOnlySpan API ===
|
||
internal static void Sqrt<T, R>(this ReadOnlySpan<T> units, Span<R> destination)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
if (units.IsEmpty) return;
|
||
int len = units.Length;
|
||
if (len > destination.Length)
|
||
throw new ArgumentException("Целевой буфер destination меньше исходного source.");
|
||
|
||
units.SqrtCore(len, destination);
|
||
}
|
||
internal static void Sqrt<T, R>(this ReadOnlySpan<T?> units, Span<R?> destination)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
if (units.IsEmpty) return;
|
||
int len = units.Length;
|
||
if (len > destination.Length)
|
||
throw new ArgumentException("Целевой буфер destination меньше исходного source.");
|
||
|
||
units.SqrtCore(len, destination);
|
||
}
|
||
|
||
// === Array API ===
|
||
internal static R[] Sqrt<T, R>(this T[] units)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
if (units is null) return null!;
|
||
if (units.Length == 0) return [];
|
||
|
||
var result = new R[units.Length];
|
||
units.AsSpan().SqrtCore(units.Length, result.AsSpan());
|
||
return result;
|
||
}
|
||
internal static R?[] Sqrt<T, R>(this T?[] units)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
if (units is null) return null!;
|
||
if (units.Length == 0) return [];
|
||
|
||
var result = new R?[units.Length];
|
||
units.AsSpan().SqrtCore(units.Length, result.AsSpan());
|
||
return result;
|
||
}
|
||
|
||
// === List API ===
|
||
internal static List<R> Sqrt<T, R>(this List<T> units)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
if (units is null) return null!;
|
||
int count = units.Count;
|
||
if (count == 0) return [];
|
||
|
||
var result = new R[count];
|
||
CollectionsMarshal.AsSpan(units).SqrtCore(count, result.AsSpan());
|
||
return result.WrapAsList<R, R>();
|
||
}
|
||
internal static List<R?> Sqrt<T, R>(this List<T?> units)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
if (units is null) return null!;
|
||
int count = units.Count;
|
||
if (count == 0) return [];
|
||
|
||
var result = new R?[count];
|
||
CollectionsMarshal.AsSpan(units).SqrtCore(count, result.AsSpan());
|
||
return result.WrapAsList<R, R>();
|
||
}
|
||
|
||
// === IReadOnlyCollection ===
|
||
internal static void Sqrt<T, R>(this IReadOnlyCollection<T> units, Span<R> destination)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
if (units is null) return;
|
||
int count = units.Count;
|
||
if (count == 0) return;
|
||
if (destination.Length < count)
|
||
throw new ArgumentException("Destination too short");
|
||
|
||
if (units is T[] array) { array.AsSpan().SqrtCore(count, destination); return; }
|
||
if (units is List<T> list) { CollectionsMarshal.AsSpan(list).SqrtCore(count, destination); return; }
|
||
|
||
int i = 0;
|
||
ref R dstRef = ref MemoryMarshal.GetReference(destination);
|
||
foreach (T item in units)
|
||
{
|
||
if ((uint)i >= (uint)count) break;
|
||
Unsafe.Add(ref dstRef, i++) = Math.Sqrt(item.ToDouble()).ToUnit<R>();
|
||
}
|
||
}
|
||
internal static void Sqrt<T, R>(this IReadOnlyCollection<T?> units, Span<R?> destination)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
if (units is null) return;
|
||
int count = units.Count;
|
||
if (count == 0) return;
|
||
if (destination.Length < count)
|
||
throw new ArgumentException("Destination too short");
|
||
|
||
if (units is T?[] array) { array.AsSpan().SqrtCore(count, destination); return; }
|
||
if (units is List<T?> list) { CollectionsMarshal.AsSpan(list).SqrtCore(count, destination); return; }
|
||
|
||
int i = 0;
|
||
ref R? dstRef = ref MemoryMarshal.GetReference(destination);
|
||
foreach (T? item in units)
|
||
{
|
||
if ((uint)i >= (uint)count) break;
|
||
Unsafe.Add(ref dstRef, i++) = Math.Sqrt(item.Protected()).ToUnit<R>();
|
||
}
|
||
}
|
||
|
||
// === IEnumerable Iterators ===
|
||
internal static IEnumerable<R> SqrtIterator<T, R>(this IEnumerable<T> units)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
foreach (var item in units)
|
||
yield return Math.Sqrt(item.ToDouble()).ToUnit<R>();
|
||
}
|
||
internal static IEnumerable<R?> SqrtNullableIterator<T, R>(this IEnumerable<T?> units)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
foreach (var item in units)
|
||
yield return Math.Sqrt(item.Protected()).ToUnit<R>();
|
||
}
|
||
|
||
// === IEnumerable API ===
|
||
internal static IEnumerable<R> Sqrt<T, R>(this IEnumerable<T> units)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
if (units is null) return null!;
|
||
if (units is T[] array) return array.Sqrt<T, R>();
|
||
if (units is List<T> list) return list.Sqrt<T, R>();
|
||
if (units is IReadOnlyCollection<T> roc)
|
||
{
|
||
var arr = roc.ToArray();
|
||
arr.SqrtCore(arr.Length, arr);
|
||
return arr.ReCast<T, R>();
|
||
}
|
||
else return SqrtIterator<T, R>(units);
|
||
}
|
||
internal static IEnumerable<R?> Sqrt<T, R>(this IEnumerable<T?> units)
|
||
where T : struct, IMensuraUnit, IEquatable<T>
|
||
where R : struct, IMensuraUnit, IEquatable<R>
|
||
{
|
||
if (units is null) return null!;
|
||
if (units is T?[] array) return array.Sqrt<T, R>();
|
||
if (units is List<T?> list) return list.Sqrt<T, R>();
|
||
if (units is IReadOnlyCollection<T?> roc)
|
||
{
|
||
var arr = roc.ToArray();
|
||
arr.SqrtCore(arr.Length, arr);
|
||
return arr.ReCast<T, R>();
|
||
}
|
||
else return SqrtNullableIterator<T, R>(units);
|
||
}
|
||
} |