经过前面的讨论，我对Image类进行了优化，代码如下：
//C#灰度图像处理类，作者：wmesci
//http://http://blog.csdn.net/wmesci unsafe class Image :CriticalHandle, IDisposable
{
[DllImport("kernel32.dll")]
static extern IntPtr LocalAlloc(int flags, int size);
[DllImport("kernel32.dll")]
static extern IntPtr LocalFree(IntPtr memBlock);
[DllImport("kernel32.dll", EntryPoint = "RtlMoveMemory")]
static extern unsafe void CopyMemory(void* dst, void* src, int count);
[DllImport("ntdll.dll")]
static extern unsafe void* memset(void* src, byte value, uint size);
const byte Max = 255;
const byte Min = 0;
public Image(int width, int height)
: base(IntPtr.Zero)
{
if (width <= 0 || height <= 0)
throw new ArgumentOutOfRangeException();
Width = width;
Height = height;
Stride = (width + 3) & ~3;
Length = Stride * Height;
base.SetHandle(LocalAlloc(0x40, Length));
Pointer = (byte*)handle.ToPointer();
}
public Image(int width, int height, byte* data)
: this(width, height)
{
SetData(data);
}
public void GetData(void* dst)
{
CopyMemory(dst, Pointer, Length);
}
public void SetData(void* src)
{
CopyMemory(Pointer, src, Length);
}
public readonly int Width;
public readonly int Height;
public readonly int Length;
public readonly int Stride;
public readonly byte* Pointer;
public byte this[int x, int y]
{
get
{
return *(Pointer + y * Stride + x);
}
set
{
*(Pointer + y * Stride + x) = value;
}
}
public Image Clone()
{
return new Image(Width, Height, Pointer);
}
public void Add(Image img)
{
Action<int> act = y =>
{
byte* p1 = Pointer + y * Stride, p2 = (byte*)img.Pointer + y * img.Stride;
for (int x = 0; x < Stride; x += 4, p1 += 4, p2 += 4)
{
int d = (int)p1[0] + (int)p2[0];
if (d < 0)
p1[0] = 0;
else if (d > 255)
p1[0] = 255;
else
p1[0] = (byte)d;
d = (int)p1[1] + (int)p2[1];
if (d < 0)
p1[1] = 0;
else if (d > 255)
p1[1] = 255;
else
p1[1] = (byte)d;
d = (int)p1[2] + (int)p2[2];
if (d < 0)
p1[2] = 0;
else if (d > 255)
p1[2] = 255;
else
p1[2] = (byte)d;
d = (int)p1[3] + (int)p2[3];
if (d < 0)
p1[3] = 0;
else if (d > 255)
p1[3] = 255;
else
p1[3] = (byte)d;
}
};
Parallel.For(0, Height, act);
}
public void Sub(Image img)
{
Action<int> act = y =>
{
byte* p1 = Pointer + y * Stride, p2 = (byte*)img.Pointer + y * img.Stride;
for (int x = 0; x < Stride; x += 4, p1 += 4, p2 += 4)
{
int d = (int)p1[0] - (int)p2[0];
if (d < 0)
p1[0] = 0;
else if (d > 255)
p1[0] = 255;
else
p1[0] = (byte)d;
d = (int)p1[1] - (int)p2[1];
if (d < 0)
p1[1] = 0;
else if (d > 255)
p1[1] = 255;
else
p1[1] = (byte)d;
d = (int)p1[2] - (int)p2[2];
if (d < 0)
p1[2] = 0;
else if (d > 255)
p1[2] = 255;
else
p1[2] = (byte)d;
d = (int)p1[3] - (int)p2[3];
if (d < 0)
p1[3] = 0;
else if (d > 255)
p1[3] = 255;
else
p1[3] = (byte)d;
}
};
Parallel.For(0, Height, act);
}
/// <summary>OK</summary>
public void Mul(Image img, double scale)
{
Action<int> act = y =>
{
byte* p1 = Pointer + y * Stride, p2 = (byte*)img.Pointer + y * img.Stride;
for (int x = 0; x < Stride; x+=4, p1+=4, p2+=4)
{
double d = (int)p1[0] * (int)p2[0] * scale;
if (d < 0)
p1[0] = 0;
else if (d > 255)
p1[0] = 255;
else
p1[0] = (byte)d;
d = (int)p1[1] * (int)p2[1] * scale;
if (d < 0)
p1[1] = 0;
else if (d > 255)
p1[1] = 255;
else
p1[1] = (byte)d;
d = (int)p1[2] * (int)p2[2] * scale;
if (d < 0)
p1[2] = 0;
else if (d > 255)
p1[2] = 255;
else
p1[2] = (byte)d;
d = (int)p1[3] * (int)p2[3] * scale;
if (d < 0)
p1[3] = 0;
else if (d > 255)
p1[3] = 255;
else
p1[3] = (byte)d;
}
};
Parallel.For(0, Height, act);
}
public void Threshold(byte threshold)
{
Action<int> act = y =>
{
byte* p = Pointer + y * Stride;
for (int x = 0; x < Stride; x+=4, p+=4)
{
p[0] = p[0] < threshold ? Min : Max;
p[1] = p[1] < threshold ? Min : Max;
p[2] = p[2] < threshold ? Min : Max;
p[3] = p[3] < threshold ? Min : Max;
}
};
Parallel.For(0, Height, act);
}
/// <summary>OK</summary>
public void AddWeighted(Image img, double a, double b)
{
int* taba = stackalloc int[256];
for (int i = 0; i < 256; i++)
taba[i] = (int)(i * a);
int* tabb = stackalloc int[256];
for (int i = 0; i < 256; i++)
tabb[i] = (int)(i * b);
Action<int> act = y =>
{
byte* p1 = this.Pointer + y * this.Stride, p2 = (byte*)img.Pointer + y * img.Stride;
for (int x = 0; x < this.Stride; x+=4, p1+=4, p2+=4)
{
int d = taba[p1[0]] + taba[p2[0]];
if (d < 0)
p1[0] = 0;
else if (d > 255)
p1[0] = 255;
else
p1[0] = (byte)d;
d = taba[p1[1]] + taba[p2[1]];
if (d < 0)
p1[1] = 0;
else if (d > 255)
p1[1] = 255;
else
p1[1] = (byte)d;
d = taba[p1[2]] + taba[p2[2]];
if (d < 0)
p1[2] = 0;
else if (d > 255)
p1[2] = 255;
else
p1[2] = (byte)d;
d = taba[p1[3]] + taba[p2[3]];
if (d < 0)
p1[3] = 0;
else if (d > 255)
p1[3] = 255;
else
p1[3] = (byte)d;
}
};
Parallel.For(0, this.Height, act);
}
public static void Smooth(Image src, Image dst, int n)
{
//分配一块临时存储区
int* tmp = (int*)Marshal.AllocHGlobal(src.Stride * src.Height * 4).ToPointer();
Action<int> act = y =>
{
byte* p = src.Pointer + y * src.Stride;
int d = 0;
for (int i = -n; i <= n; i++)
{
int xx = Clamp(i, src.Stride);
d += p[xx];
}
tmp[y * src.Stride] = d;
};
Parallel.For(0, src.Height, act);
act = y =>
{
int i = y * src.Stride;
byte* p = src.Pointer + y * src.Stride;
for (int x = 1; x < src.Stride; x++)
{
int d = tmp[i];
int x1 = Clamp(x - n - 1, src.Stride);
int x2 = Clamp(x + n, src.Stride);
d += (p[x2] - p[x1]);
tmp[++i] = d;
}
};
Parallel.For(0, src.Height, act);
double f = 1.0 / (2 * n + 1);
f *= f;
act = x =>
{
int d = 0;
byte* p = dst.Pointer + x;
for (int j = -n; j <= n; j++)
{
int yy = Clamp(j, src.Height);
d += tmp[x + yy * src.Stride];
}
*p = (byte)(d * f);
p += src.Stride;
for (int y = 1; y < src.Height; y++, p += src.Stride)
{
int y1 = Clamp(y - n - 1, src.Height);
int y2 = Clamp(y + n, src.Height);
d += (tmp[x + y2 * src.Stride] - tmp[x + y1 * src.Stride]);
*p = (byte)(d * f);
}
};
Parallel.For(0, src.Stride, act);
Marshal.FreeHGlobal(new IntPtr(tmp));
}
private static int Clamp(int i, int max)
{
if (i < 0) return 0;
if (i >= max) return max - 1;
return i;
}
public override bool IsInvalid
{
get { return handle == IntPtr.Zero; }
}
protected override bool ReleaseHandle()
{
LocalFree(handle);
return true;
}
}主要修改的地方如下：
1、将图像的每一行4字节对齐，增加Stribe属性，其值等于Width向上取最近的4的倍数，然后在所有的for循环里，每次操作4个字节。这样一来，减少了循环次数。
2、减少浮点运算
A：Add/Sub方法中的临时变量d改为int型
B：Mul方法中，调整运算顺序，由scale * *p1 * *p2改为p1[0] * p2[0] * scale，区别在于，前一种先算scale * *p0，是一个浮点乘法，其结果也是浮点数，然后再算和*p2的乘积，共两次浮点乘法；而后一种先算p1[0] * p2[0]，这是一次整数乘法，然后再算和scale的积，共一次整数乘法一次浮点乘法。由于浮点乘法比整数乘法慢，因此效率会有所提高。
3、AddWeighted改为使用查表法进行运算，首先算出0~255这256个数和a、b的积，放在数组taba、tabb中，其后的循环中只需查表再相加即可，效率大幅提高！
下面是优化后的测试结果（数值表示Image类方法和对应的OpenCV方法执行时间之比）：
CPU：AMD Athlon(tm) II X4 640 3.00GHz (四核)
样本：600 X 896
-------------------------------------
Add 1.446 1.315
Sub 1.171 1.109
Mul 0.651 0.580
Threshold 1.511 1.432
Smooth 0.938 0.908
AddWeighted 0.528 0.474
CPU：AMD Athlon(tm) II X4 640 3.00GHz (四核)
样本：1600 X 1200
-------------------------------------
Add 1.041 1.052
Sub 0.910 0.906
Mul 0.562 0.558
Threshold 1.277 1.236
Smooth 1.020 1.024
AddWeighted 0.462 0.461
CPU：AMD Athlon(tm) II X2 245 2.91GHz (双核)
样本：1600 X 1200
-------------------------------------
Add 1.514 1.533
Sub 1.225 1.163
Mul 1.085 1.095
Threshold 1.643 1.630
Smooth 1.847 1.867
AddWeighted 0.957 0.924
CPU：AMD Athlon(tm) II X2 245 2.91GHz (双核)
样本：600 X 896
-------------------------------------
Add 2.559 2.073 2.676
Sub 2.240 1.784 1.856
Mul 1.261 1.352 1.284
Threshold 2.453 2.511 3.101
Smooth 1.660 1.647 1.663
AddWeighted 0.978 1.017 0.961
CPU：Intel Core i3 M330 2.13GHz (双核四线程)
样本：1600 X 1200
-------------------------------------
Add 2.611
Sub 2.545
Mul 1.011
Threshold 2.882
Smooth 1.891
AddWeighted 0.525
CPU：Intel Core i3 M330 2.13GHz (双核四线程)
样本：600 X 896
-------------------------------------
Add 4.483
Sub 3.576
Mul 1.101
Threshold 5.953
Smooth 2.029
AddWeighted 0.581
CPU：Intel Core i7 2360QM 2.00GHz (四核八线程)
样本：600 X 896
-------------------------------------
Add 1.080 1.020
Sub 0.977 1.010
Mul 0.575 0.558
Threshold 0.842 0.898
Smooth 1.447 1.386
AddWeighted 0.325 0.366
CPU：Intel Core i7 2360QM 2.00GHz (四核八线程)
样本：1600 X 1200
-------------------------------------
Add 1.420
Sub 1.134
Mul 0.535
Threshold 0.878
Smooth 1.379
AddWeighted 0.325
分析以上数据，我们不难发现以下几点：
1、样本大小相同时，CPU核心数越多，Image/OpenCV就越小，这说明了多线程算法在多核CPU下的优势。
2、CPU相同时，样本大小越打，比值越小。
3、OpenCV针对Intel CPU使用IPP进行了优化，因此在Intel CPU上跑，比值会比在AMD CPU上打很多。
4、OpenCV里使用SSE优化过的方法，用C#实现时差距比较明显，怎么才能达到差不多的效率，这个暂时还没想到。但是OpenCV里没使用SSE优化的方法，如Mul、AddWeighted，使用C#完全可以达到相同的性能，甚至超过OpenCV，如AddWeighted，在双核CPU上也比OpenCV要快，在4核以上CPU上远超OpenCV！！
5、使用自写方法替代OpenCV是完全可行的！！
以上测试有几点需要说明：
1、CLR是在第一次运行某个方法时才进行编译，因此第一次执行某个方法时会慢很多，在计算时间时要排除第一次执行的时间。
2、C#调用OpenCV需要经过封送处理，但封送处理所消耗的时间在这里无法避免。
各位看官如有其它优化修改意见，还望不吝赐教！！！同时我会不断对这个类进行修改以及添加其它图像处理方法，对图像处理、OpenCV以及C#代码优化感兴趣的同学，请关注本贴！！