About Host API of curand curandGenerateUniform() function

Accelerated Computing CUDA CUDA Programming and Performance
1

I’m faced with the problem with curandGenerateUniformDouble() function when using GPU to accelerate my program. Does this API run on GPU or CPU? IF it works on GPU using a kernel, how does it generate configurations? If I use this API on host and generate random numbers on host, does the random numbers firstly generate on GPU and then transfer to host memory?
Thank you

2

Does this API run on GPU or CPU?

both. see following sample.

#include <cuda_runtime.h>
#include <curand.h>

#include <iostream>

int main() {
  using namespace std;
  curandGenerator_t genGPU;
  curandGenerator_t genCPU;

  curandCreateGenerator(&genGPU, CURAND_RNG_PSEUDO_MTGP32);
  curandSetPseudoRandomGeneratorSeed(genGPU, 1234ULL);
  curandCreateGeneratorHost(&genCPU, CURAND_RNG_PSEUDO_MTGP32);
  curandSetPseudoRandomGeneratorSeed(genCPU, 1234ULL);

  const int n = 10;
  double CPU[n];
  double GPU[n];

  double* d_GPU;
  cudaMalloc(&d_GPU, n*sizeof(double));

  curandGenerateUniformDouble(genCPU, CPU, n);
  curandGenerateUniformDouble(genGPU, d_GPU, n);
  cudaMemcpy(GPU, d_GPU, n*sizeof(double), cudaMemcpyDeviceToHost);

  for ( int i = 0; i < n; ++i ) {
    cout << CPU[i] << ' ' << GPU[i] << endl;
  }

  curandDestroyGenerator(genGPU);
  curandDestroyGenerator(genCPU);
  cudaFree(d_GPU);

}
3

Thank you for answering my questions. I know that Host API can call on both sides. my question is that whether it total works on CPU when use curandCreateGeneratorHost(). That is ,when I disable my GPU, it does not work any more. The reason why I ask this question is that I find it is extremely slow when generating small size sequence(0.16s ,size=100). I read the manual which says that it need a kernel to generate random numbers. So I ask this question to confirm my conclusion. Thank you

4

run test codes using the profiler. It will immediately be obvious whether it is using the GPU. You can also observe host/device data transfers, size and direction.