gpu_comm.cpp

Go to the documentation of this file.

#include "../../../include/monolish_blas.hpp"
#include "../../internal/monolish_internal.hpp"
 
namespace monolish {
// vec ///////////////////////////////////////
 
// send
template <typename T> void vector<T>::send() const {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  const T *d = val.data();
  const size_t N = val.size();
 
  if (gpu_status == true) {
#pragma omp target update to(d [0:N])
  } else {
#pragma omp target enter data map(to : d [0:N])
    gpu_status = true;
  }
#endif
  logger.util_out();
}
 
// recv
template <typename T> void vector<T>::recv() {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  if (gpu_status == true) {
    T *d = val.data();
    const size_t N = val.size();
 
#pragma omp target exit data map(from : d [0:N])
 
    gpu_status = false;
  }
#endif
  logger.util_out();
}
 
// nonfree_recv
template <typename T> void vector<T>::nonfree_recv() {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  if (gpu_status == true) {
    T *d = val.data();
    const size_t N = val.size();
#pragma omp target update from(d [0:N])
  }
#endif
  logger.util_out();
}
 
// device_free
template <typename T> void vector<T>::device_free() const {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  if (gpu_status == true) {
    const T *d = val.data();
    const size_t N = val.size();
#pragma omp target exit data map(release : d [0:N])
    gpu_status = false;
  }
#endif
  logger.util_out();
}
 
template void vector<float>::send() const;
template void vector<double>::send() const;
 
template void vector<float>::recv();
template void vector<double>::recv();
 
template void vector<float>::nonfree_recv();
template void vector<double>::nonfree_recv();
 
template void vector<float>::device_free() const;
template void vector<double>::device_free() const;
 
// CRS ///////////////////////////////////
// send
template <typename T> void matrix::CRS<T>::send() const {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  const T *vald = val.data();
  const int *cold = col_ind.data();
  const int *rowd = row_ptr.data();
  const size_t N = get_row();
  const size_t nnz = get_nnz();
 
  if (gpu_status == true) {
#pragma omp target update to(vald [0:nnz], cold [0:nnz], rowd [0:N + 1])
  } else {
#pragma omp target enter data map(to                                           \
                                  :                                            \
                                  vald [0:nnz], cold [0:nnz], rowd [0:N + 1])
    gpu_status = true;
  }
#endif
  logger.util_out();
}
 
// recv
template <typename T> void matrix::CRS<T>::recv() {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  if (gpu_status == true) {
    T *vald = val.data();
    int *cold = col_ind.data();
    int *rowd = row_ptr.data();
    size_t N = get_row();
    size_t nnz = get_nnz();
 
#pragma omp target exit data map(from                                          \
                                 : vald [0:nnz], cold [0:nnz], rowd [0:N + 1])
    gpu_status = false;
  }
#endif
  logger.util_out();
}
 
// nonfree_recv
template <typename T> void matrix::CRS<T>::nonfree_recv() {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  if (gpu_status == true) {
    T *vald = val.data();
    int *cold = col_ind.data();
    int *rowd = row_ptr.data();
    size_t N = get_row();
    size_t nnz = get_nnz();
 
#pragma omp target update from(vald [0:nnz], cold [0:nnz], rowd [0:N + 1])
  }
#endif
  logger.util_out();
}
 
// device_free
template <typename T> void matrix::CRS<T>::device_free() const {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  if (gpu_status == true) {
    const T *vald = val.data();
    const int *cold = col_ind.data();
    const int *rowd = row_ptr.data();
    const size_t N = get_row();
    const size_t nnz = get_nnz();
 
#pragma omp target exit data map(release                                       \
                                 : vald [0:nnz], cold [0:nnz], rowd [0:N + 1])
 
    gpu_status = false;
  }
#endif
  logger.util_out();
}
template void matrix::CRS<float>::send() const;
template void matrix::CRS<double>::send() const;
 
template void matrix::CRS<float>::recv();
template void matrix::CRS<double>::recv();
 
template void matrix::CRS<float>::nonfree_recv();
template void matrix::CRS<double>::nonfree_recv();
 
template void matrix::CRS<float>::device_free() const;
template void matrix::CRS<double>::device_free() const;
// Dense ///////////////////////////////////
// send
template <typename T> void matrix::Dense<T>::send() const {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  const T *vald = val.data();
  const size_t nnz = get_nnz();
 
  if (gpu_status == true) {
#pragma omp target update to(vald [0:nnz])
  } else {
#pragma omp target enter data map(to : vald [0:nnz])
    gpu_status = true;
  }
#endif
  logger.util_out();
}
 
// recv
template <typename T> void matrix::Dense<T>::recv() {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  if (gpu_status == true) {
    T *vald = val.data();
    size_t nnz = get_nnz();
 
#pragma omp target exit data map(from : vald [0:nnz])
    gpu_status = false;
  }
#endif
  logger.util_out();
}
 
// nonfree_recv
template <typename T> void matrix::Dense<T>::nonfree_recv() {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  if (gpu_status == true) {
    T *vald = val.data();
    size_t nnz = get_nnz();
 
#pragma omp target update from(vald [0:nnz])
  }
#endif
  logger.util_out();
}
 
// device_free
template <typename T> void matrix::Dense<T>::device_free() const {
  Logger &logger = Logger::get_instance();
  logger.util_in(monolish_func);
 
#if MONOLISH_USE_NVIDIA_GPU
  if (gpu_status == true) {
    const T *vald = val.data();
    const size_t nnz = get_nnz();
 
#pragma omp target exit data map(release : vald [0:nnz])
 
    gpu_status = false;
  }
#endif
  logger.util_out();
}
template void matrix::Dense<float>::send() const;
template void matrix::Dense<double>::send() const;
 
template void matrix::Dense<float>::recv();
template void matrix::Dense<double>::recv();
 
template void matrix::Dense<float>::nonfree_recv();
template void matrix::Dense<double>::nonfree_recv();
 
template void matrix::Dense<float>::device_free() const;
template void matrix::Dense<double>::device_free() const;
} // namespace monolish