new ch5 reread

[book_gpu.git] / BookGPU / Chapters / chapter4 / code / convoGene8x8pL3.cu

diff --git a/BookGPU/Chapters/chapter4/code/convoGene8x8pL3.cu b/BookGPU/Chapters/chapter4/code/convoGene8x8pL3.cu
index ce619d716f5997f53805c33fd23342feaff6c281..523dde5c418aeea9775ed69d9e067e9d5e15c69e 100644 (file)
--- a/BookGPU/Chapters/chapter4/code/convoGene8x8pL3.cu
+++ b/BookGPU/Chapters/chapter4/code/convoGene8x8pL3.cu
@@ -14,41 +14,41 @@ __global__ void kernel_convoGene8x8pL3( unsigned char  *output, int j_dim )
   for (ic=0 ; ic<k ; ic++)
        {
          pix = tex2D(tex_img_inc, j+1, i-1+ic) ;
   for (ic=0 ; ic<k ; ic++)
        {
          pix = tex2D(tex_img_inc, j+1, i-1+ic) ;

-         outval0 += masque[ __umul24(ic,k) +2 ]*pix ;
-         outval1 += masque[ __umul24(ic,k) +1 ]*pix ;
-         outval2 += masque[ __umul24(ic,k)    ]*pix ;
+         outval0 += mask[ __umul24(ic,k) +2 ]*pix ;
+         outval1 += mask[ __umul24(ic,k) +1 ]*pix ;
+         outval2 += mask[ __umul24(ic,k)    ]*pix ;

          pix = tex2D(tex_img_inc, j+2, i-1+ic) ;
          pix = tex2D(tex_img_inc, j+2, i-1+ic) ;

-         outval1 += masque[ __umul24(ic,k) +2 ]*pix ;
-         outval2 += masque[ __umul24(ic,k) +1 ]*pix ;
-         outval3 += masque[ __umul24(ic,k)    ]*pix ;
+         outval1 += mask[ __umul24(ic,k) +2 ]*pix ;
+         outval2 += mask[ __umul24(ic,k) +1 ]*pix ;
+         outval3 += mask[ __umul24(ic,k)    ]*pix ;

          pix = tex2D(tex_img_inc, j+3, i-1+ic) ;
          pix = tex2D(tex_img_inc, j+3, i-1+ic) ;

-         outval2 += masque[ __umul24(ic,k) +2 ]*pix ;
-         outval3 += masque[ __umul24(ic,k) +1 ]*pix ;
-         outval4 += masque[ __umul24(ic,k)    ]*pix ;
+         outval2 += mask[ __umul24(ic,k) +2 ]*pix ;
+         outval3 += mask[ __umul24(ic,k) +1 ]*pix ;
+         outval4 += mask[ __umul24(ic,k)    ]*pix ;

          pix = tex2D(tex_img_inc, j+4, i-1+ic) ;
          pix = tex2D(tex_img_inc, j+4, i-1+ic) ;

-         outval3 += masque[ __umul24(ic,k) +2 ]*pix ;
-         outval4 += masque[ __umul24(ic,k) +1 ]*pix ;
-         outval5 += masque[ __umul24(ic,k)    ]*pix ;
+         outval3 += mask[ __umul24(ic,k) +2 ]*pix ;
+         outval4 += mask[ __umul24(ic,k) +1 ]*pix ;
+         outval5 += mask[ __umul24(ic,k)    ]*pix ;

          pix = tex2D(tex_img_inc, j+5, i-1+ic) ;
          pix = tex2D(tex_img_inc, j+5, i-1+ic) ;

-         outval4 += masque[ __umul24(ic,k) +2 ]*pix ;
-         outval5 += masque[ __umul24(ic,k) +1 ]*pix ;
-         outval6 += masque[ __umul24(ic,k)    ]*pix ;
+         outval4 += mask[ __umul24(ic,k) +2 ]*pix ;
+         outval5 += mask[ __umul24(ic,k) +1 ]*pix ;
+         outval6 += mask[ __umul24(ic,k)    ]*pix ;

          pix = tex2D(tex_img_inc, j+6, i-1+ic) ;
          pix = tex2D(tex_img_inc, j+6, i-1+ic) ;

-         outval5 += masque[ __umul24(ic,k) +2 ]*pix ;
-         outval6 += masque[ __umul24(ic,k) +1 ]*pix ;
-         outval7 += masque[ __umul24(ic,k)    ]*pix ;
+         outval5 += mask[ __umul24(ic,k) +2 ]*pix ;
+         outval6 += mask[ __umul24(ic,k) +1 ]*pix ;
+         outval7 += mask[ __umul24(ic,k)    ]*pix ;

          // end zones
          pix = tex2D(tex_img_inc, j, i-1+ic) ;
          // end zones
          pix = tex2D(tex_img_inc, j, i-1+ic) ;

-         outval0 += masque[ __umul24(ic,k) +1 ]*pix ;
-         outval1 += masque[ __umul24(ic,k)    ]*pix ;
+         outval0 += mask[ __umul24(ic,k) +1 ]*pix ;
+         outval1 += mask[ __umul24(ic,k)    ]*pix ;

          pix = tex2D(tex_img_inc, j-1, i-1+ic) ;
          pix = tex2D(tex_img_inc, j-1, i-1+ic) ;

-         outval0 += masque[ __umul24(ic,k)  ]*pix ;
+         outval0 += mask[ __umul24(ic,k)  ]*pix ;

 
          pix = tex2D(tex_img_inc, j+7, i-1+ic) ;
 
          pix = tex2D(tex_img_inc, j+7, i-1+ic) ;

-         outval6 += masque[ __umul24(ic,k) +2 ]*pix ;
-         outval7 += masque[ __umul24(ic,k) +1 ]*pix ;
+         outval6 += mask[ __umul24(ic,k) +2 ]*pix ;
+         outval7 += mask[ __umul24(ic,k) +1 ]*pix ;

          pix = tex2D(tex_img_inc, j+8, i-1+ic) ;
          pix = tex2D(tex_img_inc, j+8, i-1+ic) ;

-         outval7 += masque[ __umul24(ic,k) +2 ]*pix ;
+         outval7 += mask[ __umul24(ic,k) +2 ]*pix ;

        }
   // multiple output 
   output[ __umul24(i, j_dim) + j   ] = outval0 ;
        }
   // multiple output 
   output[ __umul24(i, j_dim) + j   ] = outval0 ;