HLSL实现简单的图像处理功能

    由于对于dxva2解码得到的数据不宜copy回内存给CPU处理，所以最好的办法是在GPU上直接进行处理。D3D的像素着色器能够对像素直接进行操作，实现点运算极其简单方便，简单的卷积运算效果也非常好。但D3D9的限制也很多，对于过于复杂的图像处理则显得有些不能胜任。

1.点运算

    点运算用HLSL非常容易实现，几乎是公式怎么写，代码就怎么写。以RGB转灰度图显示为例：

texture Tex0 ;

int iFlag = 0 ;
float aValue= 0.0 ;
float bValue= 0.0 ;

sampler2D YTex = 
sampler_state
{
    Texture = <Tex0> ;
    MipFilter = LINEAR ;
    MinFilter = LINEAR ;
    MagFilter = LINEAR ;

    AddressU = CLAMP ;
    AddressV = CLAMP ;
};

struct PS_INPUT
{
    float2 uvCoords0 : TEXCOORD0 ;
};

float4 Main( PS_INPUT input ) : COLOR0
{
    float4 yuvColor ;
    //rgb to gray 不知道是不是这么显示的，姑且这么认为
    float gray = tex2D( YTex, input.uvCoords0 ).r * 0.299 + tex2D( YTex, input.uvCoords0 ).g * 0.587 + tex2D( YTex, input.uvCoords0 ).b * 0.114 ;
    float s = 0 ;
    if(iFlag == 0)
    {
        s = aValue * gray + bValue/255 ;
    }
    else if(iFlag == 1)
    {
        s = aValue * log(1+gray) ;
    }
    else if(iFlag == 2)
    {
        s = aValue * pow(abs(gray),bValue) ;
    }
    yuvColor.r = s ;
    yuvColor.g = s ;
    yuvColor.b = s ;
    yuvColor.a = 1.0 ;

    return yuvColor  ;
}

    点运算如此简单是因为GPU是并行运算的，我个人认为可以看成是每一个像素点（BGRA）对应一个线程，这大概就是OpenCL中所谓的数据并行。这是一个非常简单的程序，指令数少，程序结构也很简单，shader 的版本用2.0就可以轻松编过。

2.卷积运算举例

    指令数较多的情况2.0版本的shader就搞不定了，上3.0版本可以做一些简单的卷积运算。以中值滤波为例：

texture Tex0 ;

matrix WorldMatrix;
matrix ViewMatrix;
matrix ProjMatrix;

sampler2D YTex = 
sampler_state
{
    Texture = <Tex0> ;
    MipFilter = LINEAR ;
    MinFilter = LINEAR ;
    MagFilter = LINEAR ;

    AddressU = CLAMP ;
    AddressV = CLAMP ;
};

struct VS_INPUT
  {
    float4 pos    : POSITION;
    float4 color  : COLOR0;
    float2 tex    : TEXCOORD0;
  };
// 
struct VS_OUTPUT
  {
    float4 pos     : POSITION;
    float4 color   : COLOR0;
    float2 tex     : TEXCOORD0;
  };

float2 g_v4ScreenSize;
int ksize = 1 ;

float fLeft = -1.0f ;
float fTop = -1.0f ;
float fRight = -1.0f ;
float fBottom = -1.0f ;
//--------------------------------- BurTechnique --------------------------------------


VS_OUTPUT MainVS_Screen( VS_INPUT In )
{
    VS_OUTPUT Out = ( VS_OUTPUT )0;

    float4x4 matWorldView = mul(WorldMatrix,ViewMatrix);
    float4x4 matProject = mul(matWorldView,ProjMatrix);
    Out.pos  = mul(In.pos,matProject);
    Out.tex  = In.tex;
    Out.color = In.color;
    return Out;
}
  
float4 MainPS_Screen( VS_INPUT In ) : COLOR0
{
    float4 outColor = tex2D( YTex, In.tex ).rgba ;
    
    if(ksize <= 1 || ksize%2 == 0)
    {
        return outColor ;
    }
    
    if( ksize > 11 || ksize < 3)
    {
        return outColor ;
    }
    
    if(!(In.tex.x < fRight && In.tex.y < fBottom && In.tex.x > fLeft && In.tex.y > fTop))
    {
        return outColor ;
    }
    
    // 纹理大小
    float2 TexSize = float2( g_v4ScreenSize.x , g_v4ScreenSize.y );

    float x_off = 1.0f / TexSize.x;
    float y_off = 1.0f / TexSize.y;
    
    float2 fX0Y0 = In.tex - float2(x_off * ksize/2, y_off*ksize/2) ;

    float3 sum = {0.0f, 0.0f, 0.0f} ;

    if(ksize >= 3)
    {
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 0, y_off*0)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 0, y_off*1)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 0, y_off*2)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 1, y_off*0)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 1, y_off*1)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 1, y_off*2)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 2, y_off*0)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 2, y_off*1)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 2, y_off*2)).rgb;
    }
    
    if(ksize >= 5)
    {
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 3, y_off*0)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 3, y_off*1)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 3, y_off*2)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 3, y_off*3)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 3, y_off*4)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 4, y_off*0)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 4, y_off*1)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 4, y_off*2)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 4, y_off*3)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 4, y_off*4)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 0, y_off*3)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 1, y_off*3)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 2, y_off*3)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 0, y_off*4)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 1, y_off*4)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 2, y_off*4)).rgb;
    }
    
    if(ksize >= 7)
    {
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 5, y_off*0)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 5, y_off*1)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 5, y_off*2)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 5, y_off*3)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 5, y_off*4)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 5, y_off*5)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 5, y_off*6)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 6, y_off*0)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 6, y_off*1)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 6, y_off*2)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 6, y_off*3)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 6, y_off*4)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 6, y_off*5)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 6, y_off*6)).rgb;
        
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 0, y_off*5)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 1, y_off*5)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 2, y_off*5)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 3, y_off*5)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 4, y_off*5)).rgb;
        
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 0, y_off*6)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 1, y_off*6)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 2, y_off*6)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 3, y_off*6)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 4, y_off*6)).rgb;
    }
    
    if(ksize >= 9)
    {
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 7, y_off*0)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 7, y_off*1)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 7, y_off*2)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 7, y_off*3)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 7, y_off*4)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 7, y_off*5)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 7, y_off*6)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 7, y_off*7)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 7, y_off*8)).rgb;
        
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 8, y_off*0)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 8, y_off*1)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 8, y_off*2)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 8, y_off*3)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 8, y_off*4)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 8, y_off*5)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 8, y_off*6)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 8, y_off*7)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 8, y_off*8)).rgb;
        
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 0, y_off*7)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 1, y_off*7)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 2, y_off*7)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 3, y_off*7)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 4, y_off*7)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 5, y_off*7)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 6, y_off*7)).rgb;
        
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 0, y_off*8)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 1, y_off*8)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 2, y_off*8)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 3, y_off*8)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 4, y_off*8)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 5, y_off*8)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 6, y_off*8)).rgb;
    }
    
    if(ksize >= 11)
    {
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 9, y_off*0)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 9, y_off*1)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 9, y_off*2)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 9, y_off*3)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 9, y_off*4)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 9, y_off*5)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 9, y_off*6)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 9, y_off*7)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 9, y_off*8)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 9, y_off*9)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 9, y_off*10)).rgb;
        
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 10, y_off*0)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 10, y_off*1)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 10, y_off*2)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 10, y_off*3)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 10, y_off*4)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 10, y_off*5)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 10, y_off*6)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 10, y_off*7)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 10, y_off*8)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 10, y_off*9)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 10, y_off*10)).rgb;
        
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 0, y_off*9)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 1, y_off*9)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 2, y_off*9)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 3, y_off*9)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 4, y_off*9)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 5, y_off*9)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 6, y_off*9)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 7, y_off*9)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 8, y_off*9)).rgb;
        
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 0, y_off*10)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 1, y_off*10)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 2, y_off*10)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 3, y_off*10)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 4, y_off*10)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 5, y_off*10)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 6, y_off*10)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 7, y_off*10)).rgb;
        sum += tex2D( YTex , fX0Y0 + float2(x_off * 8, y_off*10)).rgb;
    }

    outColor = float4(sum/(ksize*ksize),1.0f);

    return outColor ;
}

//--------------------------- 技术---------------------------

technique BurTechnique
{
    pass P0
    {    
        LightEnable[0] = false;
        
        VertexShader = compile vs_3_0 MainVS_Screen();
        PixelShader  = compile ps_3_0 MainPS_Screen();
    }
}

     由于3.0版本的shader似乎不允许pixel shader单独出现，所以我从点运算用像素着色器实现改为用特效来实现。HLSL语法中有if语句，也有for语句，可是这个程序却不厌其烦的把所有的都给列出来来，而没有使用for循环。这是因为在实际使用中发现有一些限制，比如if语句的if（A>B）,A与B中必须有一个是常量，就像上面见到的那种形式；for循环中间的判断也是如此，只是在第二层j循环中可以是第一层循环的i,即不可以

for(int i=0;i<ksize;i++)
{
    for(int j=0;j<ksize1;j++)
        {
    ..........
        }
}

以上代码的ksize与ksize1都必须为常数，例外的情况是ksize1可以为第一层循环的 i 。这个问题不知道后续版本的shader有没有，反正我当前使用的版本有。

另外有一个需要注意的地方是指令数，2.0版本的shader支持的指令数相当少，3.0版本则要多好多，我最长写到了400多条快500条时才导致编译失败。 还有一个需要提醒的是3.0版本的shader只支持D3D 9.0C以后的。如果要求做更为复杂的图像处理，可以的话建议上D3D11，compute shader虽然我没用过，但从介绍来说，应该可以处理一些更为复杂的图像处理。

结合我的上一篇博客（DXVA2解码数据用texture纹理渲染http://www.cnblogs.com/betterwgo/p/6327422.html），就算是实现了从硬解到简单图像处理的完整过程。显卡加速效果非常好，在我的intel 5200上即使4K视频也可以实现比正常播放略快的效果。

工程源码：http://download.csdn.net/download/qq_33892166/9755307