last bug in dt fixed

kernels.cl

@@ -415,9 +415,9 @@ void DGVolume(int ie,                    // 1: macrocel index
 
   
   // ref coordinates
-  float hx = 1. / _RAF;
-  float hy = 1. / _RAF;
-  float hz = 1. / _RAF;
+  float hx = 1. / (float)_RAF;
+  float hy = 1. / (float)_RAF;
+  float hz = 1. / (float)_RAF;
 
   int offset[3] = {gauss_lob_offset[_DEG] + p[0],
 		   gauss_lob_offset[_DEG] + p[1],
@@ -484,10 +484,10 @@ void DGVolume(int ie,                    // 1: macrocel index
 
       NumFlux(wL, wL, dphi, flux); // 3m mults when using NumFlux
  
-      int ipgR = (ix + iq) % _NP + _NPG_ROW * (iy + _NPG_ROW * iz);
+      int ipgR = q[0] + icL[0] *  _NP + _NPG_ROW * (iy + _NPG_ROW * iz);
       for(int iv = 0; iv < _M; iv++) {
       int imemR =  Varindex(ipgR, iv) + woffset;
-      printf("imemR=%d wpg=%f\n",imemR,wpg);
+      //printf("imemR=%d wpg=%f\n",imemR,wpg);
 
      	dtwn[imemR] += flux[iv] * wpg;
       }
@@ -626,14 +626,14 @@ void DGFlux(int ie,                      // 1: macrocell index
   for(int iv = 0; iv < _M; ++iv) {
     if (ipgL >= 0){
       int imemL =  Varindex(ipgL, iv) + woffset;
-      printf("imemL=%d wpgs=%f\n",imemL,wpgs);
+      //printf("imemL=%d wpgs=%f\n",imemL,wpgs);
       dtwn[imemL] -= flux[iv] * wpgs;
     }
     
     
     if (ipgR >= 0){
       int imemR =  Varindex(ipgR, iv) + woffset;
-      printf("imemR=%d\n",imemR);
+      //printf("imemR=%d\n",imemR);
       dtwn[imemR] += flux[iv] * wpgs;
     }
   }

testtranspose.py

@@ -9,7 +9,7 @@ with open('kernels.cl', 'r') as f:
     source = f.read()
 
 m = 2
-raf = 1
+raf = 2
 deg = 1
 
 source = source.replace("THE_DEG", str(deg))