#include <stdio.h>

__global__ void kernel1(){ }
__global__ void kernel2(){ }
__global__ void kernel3(){ }

int main()
{
  cudaStream_t streamA, streamB;
  cudaStreamCreate(&streamA);
  cudaStreamCreate(&streamB);

  // Launch kernel1 and kernel2 in streamA.
  kernel1<<<32, 256, 0, streamA>>>();
  kernel2<<<32, 256, 0, streamA>>>();

  // Launch kernel3 in streamB.
  kernel3<<<32, 256, 0, streamB>>>();

  // Wait for both streams to finish
  cudaStreamSynchronize(streamA);
  cudaStreamSynchronize(streamB);

  cudaStreamDestroy(streamA);
  cudaStreamDestroy(streamB);

  return 0;
}