int boundingBoxWidth = maxx - minx;
    int boundingBoxHeight = maxy - miny;
    int numPixels = boundingBoxWidth * boundingBoxHeight;
    int parallelThreshold = 32;
    bool overThreshold = numPixels > parallelThreshold; // true if over threshold
    int votes = __ballot_sync(0xFFFFFFFF, overThreshold);
    int count = __popc(votes);


    for (int bigTriangle = 0; bigTriangle < count; bigTriangle++){ // 1.
         int position = __ffs(votes);
         //votes &= ~(1UL << position);  // Set bit at position in votes to 0
         votes ^= 1UL << position;
         unsigned int triangleIndexP = __shfl_sync(0xFFFFFFFF, triangleIndex, position);
         int minxP = __shfl_sync(0xFFFFFFFF, minx, position);
         int maxxP = __shfl_sync(0xFFFFFFFF, maxx, position);
         int minyP = __shfl_sync(0xFFFFFFFF, miny, position);
         int maxyP = __shfl_sync(0xFFFFFFFF, maxy, position);
         int numPixelsP = __shfl_sync(0xFFFFFFFF, numPixels, position);
         float4 v0P = make_float4(0,0,0,0);
         float4 v1P = make_float4(0,0,0,0);
         float4 v2P = make_float4(0,0,0,0);
         v0P.x = __shfl_sync(0xFFFFFFFF, v0.x, position);
         v0P.y = __shfl_sync(0xFFFFFFFF, v0.y, position);
         v0P.z = __shfl_sync(0xFFFFFFFF, v0.z, position);
         v0P.w = __shfl_sync(0xFFFFFFFF, v0.w, position);
         v1P.x = __shfl_sync(0xFFFFFFFF, v1.x, position);
         v1P.y = __shfl_sync(0xFFFFFFFF, v1.y, position);
         v1P.z = __shfl_sync(0xFFFFFFFF, v1.z, position);
         v1P.w = __shfl_sync(0xFFFFFFFF, v1.w, position);
         v2P.x = __shfl_sync(0xFFFFFFFF, v2.x, position);
         v2P.y = __shfl_sync(0xFFFFFFFF, v2.y, position);
         v2P.z = __shfl_sync(0xFFFFFFFF, v2.z, position);
         v2P.w = __shfl_sync(0xFFFFFFFF, v2.w, position);
         for (int id = 0; id < numPixelsP; id += 32 ){
             // int currentIndex = id + (int(fmod(float(threadIdx.x), float(32))));
             int threadId = threadIdx.x + threadIdx.y * blockDim.x;
             int currentIndex = id + (threadId % 32);
             if (currentIndex > numPixelsP) {break;}
             //int xP = minxP + int(fmod(float(currentIndex), float(maxxP - minxP)));//(currentIndex % (maxxP - minxP));
             int xP = minxP + (currentIndex % (maxxP - minxP));
             int yP = minyP + int(currentIndex / (maxxP - minxP));
             rasteriseSinglePixel(v0P, v1P, v2P, xP, yP, mesh, triangleIndexP, frameBuffer, depthBuffer, width, height);
         }
    }