__shared__ float tile[32][33];   // padded: row stride of 33 floats, not 32
float val = tile[threadIdx.x][0];  // now each thread lands in a different bank
