nrm2.hpp

Go to the documentation of this file.

#pragma once
 
namespace monolish {
 
namespace {
template <typename F1> double Dnrm2_core(const F1 &x) {
  Logger &logger = Logger::get_instance();
  logger.func_in(monolish_func);
 
  double ans = 0;
  const double *xd = x.data();
  size_t size = x.size();
  const size_t xoffset = x.get_offset();
 
  if (x.get_device_mem_stat() == true) {
#if MONOLISH_USE_NVIDIA_GPU
    cublasHandle_t h;
    internal::check_CUDA(cublasCreate(&h));
#pragma omp target data use_device_ptr(xd)
    {
      internal::check_CUDA(
          cublasDdot(h, size, xd + xoffset, 1, xd + xoffset, 1, &ans));
    }
    cublasDestroy(h);
#else
    throw std::runtime_error(
        "error USE_GPU is false, but get_device_mem_stat() == true");
#endif
  } else {
    ans = cblas_ddot(size, xd + xoffset, 1, xd + xoffset, 1);
  }
 
#if MONOLISH_USE_MPI
  mpi::comm &comm = mpi::comm::get_instance();
  ans = comm.Allreduce(ans);
#endif
 
  logger.func_out();
  return sqrt(ans);
}
 
template <typename F1> float Snrm2_core(const F1 &x) {
  Logger &logger = Logger::get_instance();
  logger.func_in(monolish_func);
 
  float ans = 0;
  const float *xd = x.data();
  size_t size = x.size();
  const size_t xoffset = x.get_offset();
 
  if (x.get_device_mem_stat() == true) {
#if MONOLISH_USE_NVIDIA_GPU
    cublasHandle_t h;
    internal::check_CUDA(cublasCreate(&h));
#pragma omp target data use_device_ptr(xd)
    {
      internal::check_CUDA(
          cublasSdot(h, size, xd + xoffset, 1, xd + xoffset, 1, &ans));
    }
    cublasDestroy(h);
#else
    throw std::runtime_error(
        "error USE_GPU is false, but get_device_mem_stat() == true");
#endif
  } else {
    ans = cblas_sdot(size, xd + xoffset, 1, xd + xoffset, 1);
  }
 
#if MONOLISH_USE_MPI
  mpi::comm &comm = mpi::comm::get_instance();
  ans = comm.Allreduce(ans);
#endif
 
  logger.func_out();
  return sqrt(ans);
}
} // namespace
 
} // namespace monolish