using StgSharp.Math;
using System;
using System.Collections.Generic;
using System.Runtime.InteropServices;
using System.Text;
using static StgSharp.Optimization.CpuInfo;

namespace StgSharp.Optimization
{
    public static unsafe partial class SSE
    {
        public static m128 sqrtssM128Cache = new m128();

        private static IntPtr sqrtssCodePointer;

        private static sqrtssDelegate sqrtssDele;

        private static byte[] sqrtssCodeBytes
            = new byte[]
        {
            0x48,0xb8,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,  // mov rax, [m128 location]
            0xf3,0x0f,0x6f,0x00,                                // movqdu xmm0,xmmword ptr [rax]
            0xf3,0x0f,0x51,0xc0,                                // sqrtss xmm0,xmm0
            0xf3,0x0f,0x7f,0x00                                 // movdqu xmmword ptr [rax],xmm0
        };

        [DllImport("kernel32.dll", SetLastError = true)]
        private static extern IntPtr VirtualAlloc(
            IntPtr lpAddress,
            UIntPtr dwSize,
            AllocationType flAllocationType,
            MemoryProtection flProtect
        );

        [DllImport("kernel32")]
        private static extern bool VirtualFree(
                IntPtr lpAddress,
                UInt32 dwSize,
                UInt32 dwFreeType
        );

        private unsafe static void InitSqrtss()
        {
            var marshalPtr = Marshal.AllocHGlobal(32);
            var m128ptr = new IntPtr(16*((long)marshalPtr/16+1));
            m128* sqrtssCachePointer = (m128*)m128ptr.ToPointer();

            sqrtssM128Cache = *sqrtssCachePointer;
            byte[] sqrtssM128PointerBytes = BitConverter.GetBytes((ulong)sqrtssCachePointer);
            for (int i = 0; i < sqrtssM128PointerBytes.Length; i++)
            {
                sqrtssCodeBytes[2+i] = sqrtssCodeBytes[7-i];
            }

            if (IntPtr.Size!=8)
            {
                throw new Exception("This SSE class support 64bits OS only presently!");
            }

            sqrtssCodePointer = VirtualAlloc(
                IntPtr.Zero,
                new UIntPtr((uint)sqrtssCodeBytes.Length),
                AllocationType.COMMIT | AllocationType.RESERVE,
                MemoryProtection.EXECUTE_READWRITE
            );

            Marshal.Copy(sqrtssCodeBytes, 0, sqrtssCodePointer, sqrtssCodeBytes.Length);
            sqrtssDele = (sqrtssDelegate)Marshal.GetDelegateForFunctionPointer(sqrtssCodePointer,typeof(sqrtssDelegate));
        }

        public static float Sqrtss(float a)
        {
            sqrtssM128Cache.float0 = a;
            sqrtssDele();
            return sqrtssM128Cache.float0;
        }

    }


    [UnmanagedFunctionPointerAttribute(CallingConvention.StdCall)]
    internal delegate void sqrtssDelegate();
}