MemoryCopy.cs

namespace CSharpSpeedTest
{
    public static class MemoryCopy
    {
        public static unsafe void Fast(void* dst, void* src, ulong n)
        {
            var offset = 0;
            while (n != 0)
            {
                if (n % sizeof(ulong) * 8 == 0)
                {
                    //8 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //16 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //24 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);

                    //32 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);

                    //40 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //48 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //56 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);

                    //64 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);

                    n -= 64;
                } else if (n % sizeof(ulong) * 7 == 0)
                {
                    //8 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //16 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //24 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);

                    //32 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);

                    //40 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //48 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //56 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                } else if (n % sizeof(ulong) * 6 == 0)
                {
                    //8 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //16 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //24 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);

                    //32 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);

                    //40 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //48 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                } else if (n % sizeof(ulong) * 5 == 0)
                {
                    //8 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //16 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //24 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);

                    //32 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //40 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                } else if (n % sizeof(ulong) * 4 == 0)
                {
                    //8 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //16 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //24 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);

                    //32 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                } else if (n % sizeof(ulong) * 3 == 0)
                {
                    //8 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //16 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //24 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                } else if (n % sizeof(ulong) * 2 == 0)
                {
                    //8 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                    
                    //16 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                } else if (n % sizeof(ulong) == 0)
                {
                    //8 bytes
                    *((ulong*) dst + offset) = *((ulong*) src + offset);
                    offset += sizeof(ulong);
                } else if (n % sizeof(uint) + sizeof(ushort) == 0)
                {
                    //4 bytes
                    *((uint*) dst + offset) = *((uint*) src + offset);
                    offset += sizeof(uint);
                    
                    //2 bytes
                    *((ushort*) dst + offset) = *((ushort*) src + offset);
                    offset += sizeof(ushort);
                } else if (n % sizeof(uint) == 0)
                {
                    //4 bytes
                    *((uint*) dst + offset) = *((uint*) src + offset);
                    offset += sizeof(uint);
                } else if (n % sizeof(ushort) == 0)
                {
                    //2 bytes
                    *((ushort*) dst + offset) = *((ushort*) src + offset);
                    offset += sizeof(ushort);
                }
                else
                {
                    //1 byte
                    *((byte*) dst + offset) = *((byte*) src + offset);
                    offset += sizeof(byte);
                }
            }
        }

        public static unsafe void Normal(void* dst, void* src, ulong n)
        {
            for (ulong i = 0; i < n; i++)
                *((byte*) dst + i) = *((byte*) src + i);
        }
    }
}