Cuda grid size limitations - cuda

Question

Are there limitations as to what I can set the grid size of a CUDA kernel to be? I ran into a problem where kernels were not launching with a grid size of 33 x 33 but were able to launch when the grid size was 32 x 32. Is there any reason for this to occur? Or is it likely that changing the number of blocks from 32 x 32 to 33 x 33 broke some other constraint?
dim3 blockSize(8, 8);
dim3 gridSize(32, 32);
cudaDeviceSynchronize();
set_start<<<gridSize, blockSize>>>(some_params);
The above works.
dim3 blockSize(8, 8);
dim3 gridSize(33, 33);
cudaDeviceSynchronize();
set_start<<<gridSize, blockSize>>>(some_params);
The above does not work.
kernel & main:
__global__
void set_start(double * const H , double * const HU , double * const HV ,
double * const E , const int Na)
{
int j = threadIdx.x + blockIdx.x*blockDim.x + 1;
int i = threadIdx.y + blockIdx.y*blockDim.y + 1;
if(i >= Na-1 || j >= Na-1)
return;
H[i*Na+j] = 1.0 + exp(-100.0*((E[j-1]-0.75)*(E[j-1]-0.75)+(E[i-1]-0.75)*(E[i-1]-0.75))) + 0.5*exp(-100.0*((E[j-1]-0.75)*(E[j-1]-0.75)+(E[i-1]-0.25)*(E[i-1]-0.25)));
HU[i*Na+j] = 0;
HV[i*Na+j] = 0;
}
int main(int argc, char** argv){
double* E_d;
cudaMalloc(&E_d, sizeof(double) * (Nh+1));
set_E<<<64, (Nh/64) + 1>>>(E_d, dx, Nh);
int Na = 259;
double *H_d, *HU_d, *HV_d, *Ht_d, *HUt_d, *HVt_d;
cudaMalloc(&H_d , sizeof(double) * Na * Na);
cudaMalloc(&HU_d, sizeof(double) * Na * Na);
cudaMalloc(&HV_d, sizeof(double) * Na * Na);
dim3 blockSize(8, 8);
//dim3 gridSize(((Na-1)/blockSize.x) + 1, ((Na-1)/blockSize.y) + 1);
//dim3 gridSize(33, 33);
dim3 gridSize(32, 32);
cudaDeviceSynchronize();
set_start<<<blockSize, gridSize>>>(H_d, HU_d, HV_d, E_d, Na);
}
This was on CUDA 7.0.

Answer 1

You have block size and grid size mixed up when calling the kernel.
set_start<<<blockSize, gridSize>>>(H_d, HU_d, HV_d, E_d, Na);
should read:
set_start<<<gridSize, blockSize>>>(H_d, HU_d, HV_d, E_d, Na);
Because of this bug you are actually trying to launch a grid of size blockSize, and blocks of size gridSize. It would appear that the maximum size of a block on your GPU is 1024 threads, so launching blocks of 33x33 fails.