C++ HLSL實現簡單的圖像處理功能

2020-01-26 14:14:52

字體：大中小

來源：轉載

供稿：網友

由于對于dxva2解碼得到的數據不宜copy回內存給CPU處理，所以最好的辦法是在GPU上直接進行處理。D3D的像素著色器能夠對像素直接進行操作，實現點運算極其簡單方便，簡單的卷積運算效果也非常好。但D3D9的限制也很多，對于過于復雜的圖像處理則顯得有些不能勝任。

1.點運算

點運算用HLSL非常容易實現，幾乎是公式怎么寫，代碼就怎么寫。以RGB轉灰度圖顯示為例：

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 不知道是不是這么顯示的，姑且這么認為  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 ;}

點運算如此簡單是因為GPU是并行運算的，我個人認為可以看成是每一個像素點（BGRA）對應一個線程，這大概就是OpenCL中所謂的數據并行。這是一個非常簡單的程序，指令數少，程序結構也很簡單，shader 的版本用2.0就可以輕松編過。

2.卷積運算舉例

指令數較多的情況2.0版本的shader就搞不定了，上3.0版本可以做一些簡單的卷積運算。以中值濾波為例：

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 ;  }  // 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();  }}

由于3.0版本的shader似乎不允許pixel shader單獨出現，所以我從點運算用像素著色器實現改為用特效來實現。HLSL語法中有if語句，也有for語句，可是這個程序卻不厭其煩的把所有的都給列出來來，而沒有使用for循環。這是因為在實際使用中發現有一些限制，比如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循環中可以是第一層循環的i,即不可以

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

以上代碼的ksize與ksize1都必須為常數，例外的情況是ksize1可以為第一層循環的 i 。這個問題不知道后續版本的shader有沒有，反正我當前使用的版本有。

另外有一個需要注意的地方是指令數，2.0版本的shader支持的指令數相當少，3.0版本則要多好多，我最長寫到了400多條快500條時才導致編譯失敗。還有一個需要提醒的是3.0版本的shader只支持D3D 9.0C以后的。如果要求做更為復雜的圖像處理，可以的話建議上D3D11，compute shader雖然我沒用過，但從介紹來說，應該可以處理一些更為復雜的圖像處理。

以上就是本文的全部內容，希望本文的內容對大家的學習或者工作能帶來一定的幫助，同時也希望多多支持武林網！

上一篇：C/C++中宏定義(#define)

下一篇：C++ 基礎編程之十進制轉換為任意進制及操作符重載