launch_kernel.cu

#include <stdio.h>
#include <chrono>

__global__ void noop() { }

int main(int argc, const char *argv[]) {
   cudaDeviceProp prop;
  auto rc = cudaGetDeviceProperties(&prop, 0);
  printf("Running on %s sm%d.%d  multiProcessorCount = %d maxBlocksPerMultiProcessor  = %d maxThreadsPerBlock  = %d\n",
         prop.name, prop.major, prop.minor, prop.multiProcessorCount, prop.maxBlocksPerMultiProcessor, prop.maxThreadsPerBlock);
  if (rc != cudaSuccess) {
    printf("cudaGetDeviceProperties()=%d (%s)\n", rc, cudaGetErrorString(rc));
    return rc;
  }
  if (argc > 1) {
    rc = cudaSetDeviceFlags(cudaDeviceScheduleYield);
    if (rc != cudaSuccess) {
      printf("cudaSetDeviceFlags)=%d (%s)\n", rc, cudaGetErrorString(rc));
    }
  }
  auto start = std::chrono::system_clock::now();
  constexpr auto launch_count = 16384;
  for(auto cnt = 0; cnt < launch_count; ++cnt) {
      noop<<<prop.multiProcessorCount * 10, 16384>>>();
  }
  rc = cudaDeviceSynchronize();
  auto end = std::chrono::system_clock::now();
  auto duration =  std::chrono::duration_cast<std::chrono::microseconds>(end-start).count();
  if (rc != cudaSuccess) {
    printf("cudaDeviceSynchronie()=%d (%s)\n", rc, cudaGetErrorString(rc));
  }
  printf("Launching CUDA kernel takes %.2f ms\n", 1. * duration / launch_count);
  return rc;
}