HLSL實現簡單的圖像處理功能

時間  2019-12-06
標籤 hlsl 實現 簡單 圖像處理 功能 简体版
原文   原文鏈接

    因爲對於dxva2解碼獲得的數據不宜copy回內存給CPU處理，因此最好的辦法是在GPU上直接進行處理。D3D的像素着色器可以對像素直接進行操做，實現點運算極其簡單方便，簡單的卷積運算效果也很是好。但D3D9的限制也不少，對於過於複雜的圖像處理則顯得有些不能勝任。html

1.點運算

    點運算用HLSL很是容易實現，幾乎是公式怎麼寫，代碼就怎麼寫。以RGB轉灰度圖顯示爲例：spa

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 不知道是否是這麼顯示的，姑且這麼認爲
    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  ;
}

    點運算如此簡單是由於GPU是並行運算的，我我的認爲能夠當作是每個像素點（BGRA）對應一個線程，這大概就是OpenCL中所謂的數據並行。這是一個很是簡單的程序，指令數少，程序結構也很簡單，shader 的版本用2.0就能夠輕鬆編過。.net

2.卷積運算舉例

    指令數較多的狀況2.0版本的shader就搞不定了，上3.0版本能夠作一些簡單的卷積運算。以中值濾波爲例：線程

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 ;
    }
    
    // 紋理大小
    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 ;
}

//--------------------------- 技術---------------------------

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

     因爲3.0版本的shader彷佛不容許pixel shader單獨出現，因此我從點運算用像素着色器實現改成用特效來實現。HLSL語法中有if語句，也有for語句，但是這個程序卻不厭其煩的把全部的都給列出來來，而沒有使用for循環。這是由於在實際使用中發現有一些限制，好比if語句的if（A>B）,A與B中必須有一個是常量，就像上面見到的那種形式；for循環中間的判斷也是如此，只是在第二層j循環中能夠是第一層循環的i,即不能夠code

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

以上代碼的ksize與ksize1都必須爲常數，例外的狀況是ksize1能夠爲第一層循環的 i 。這個問題不知道後續版本的shader有沒有，反正我當前使用的版本有。視頻

 

另外有一個須要注意的地方是指令數，2.0版本的shader支持的指令數至關少，3.0版本則要多好多，我最長寫到了400多條快500條時才致使編譯失敗。 還有一個須要提醒的是3.0版本的shader只支持D3D 9.0C之後的。若是要求作更爲複雜的圖像處理，能夠的話建議上D3D11，compute shader雖然我沒用過，但從介紹來講，應該能夠處理一些更爲複雜的圖像處理。htm

 

結合個人上一篇博客（DXVA2解碼數據用texture紋理渲染http://www.cnblogs.com/betterwgo/p/6327422.html），就算是實現了從硬解到簡單圖像處理的完整過程。顯卡加速效果很是好，在個人intel 5200上即便4K視頻也能夠實現比正常播放略快的效果。blog

 

工程源碼：http://download.csdn.net/download/qq_33892166/9755307ip

相關文章
相關標籤/搜索
每日一句
    每一个你不满意的现在，都有一个你没有努力的曾经。
本站公眾號
   歡迎關注本站公眾號,獲取更多信息
聯繫我們 最近搜索 最新文章 沪ICP备13005482号-10 MyBatis教程 SQL 教程 MySQL教程 Java 教程 Thymeleaf 教程 Hibernate教程 Spring教程 Redis教程