monolish_tensor_crs.hpp

Go to the documentation of this file.

#pragma once
#include "monolish_matrix.hpp"
#include "monolish_tensor.hpp"
#include "monolish_vector.hpp"
 
namespace monolish {
template <typename Float> class vector;
template <typename TYPE, typename Float> class view1D;
namespace tensor {
template <typename Float> class tensor_Dense;
template <typename Float> class tensor_COO;
template <typename Float> class tensor_CRS {
private:
  std::vector<size_t> shape;
 
  mutable std::shared_ptr<bool> gpu_status = std::make_shared<bool>(false);
 
  size_t structure_hash;
 
  size_t first = 0;
 
public:
  std::shared_ptr<Float> val;
 
  size_t val_nnz = 0;
 
  size_t alloc_nnz = 0;
 
  bool val_create_flag = false;
 
  std::vector<std::vector<int>> row_ptrs;
 
  std::vector<std::vector<int>> col_inds;
 
  tensor_CRS() : shape(), row_ptrs(), col_inds(), val_nnz(0) {
    val_create_flag = true;
  }
 
  tensor_CRS(const std::vector<size_t> &shape_)
      : shape(shape_), row_ptrs(), col_inds(), val_nnz(0) {
    val_create_flag = true;
  }
 
  tensor_CRS(const std::initializer_list<size_t> &shape_)
      : shape(shape_), row_ptrs(), col_inds(), val_nnz(0) {
    val_create_flag = true;
  }
 
  void convert(const tensor::tensor_COO<Float> &coo);
 
  tensor_CRS(const tensor::tensor_COO<Float> &coo) {
    val_create_flag = true;
    convert(coo);
  }
 
  void convert(const tensor::tensor_CRS<Float> &crs);
 
  tensor_CRS(const tensor::tensor_CRS<Float> &crs) {
    val_create_flag = true;
    convert(crs);
  }
 
  void convert(const matrix::CRS<Float> &crs);
 
  tensor_CRS(const matrix::CRS<Float> &crs) {
    val_create_flag = true;
    convert(crs);
  }
 
  tensor_CRS(const std::vector<size_t> &shape_,
             const std::vector<std::vector<int>> &row_ptrs_,
             const std::vector<std::vector<int>> &col_inds_,
             const Float *value);
 
  tensor_CRS(const tensor_CRS<Float> &crs, Float value);
 
  void print_all(bool force_cpu = false) const;
 
  // communication
  // ///////////////////////////////////////////////////////////////////////////
  void send() const;
 
  void recv();
 
  void nonfree_recv();
 
  void device_free() const;
 
  // TODO
  [[nodiscard]] double get_data_size() const {
    return get_nnz() * sizeof(Float) / 1.0e+9;
  }
 
  void set_ptr(const std::vector<size_t> &shape,
               const std::vector<std::vector<int>> &rowptrs,
               const std::vector<std::vector<int>> &colinds,
               const std::vector<Float> &value);
 
  void set_ptr(const std::vector<size_t> &shape,
               const std::vector<std::vector<int>> &rowptrs,
               const std::vector<std::vector<int>> &colinds, const size_t vsize,
               const Float *value);
 
  void set_ptr(const std::vector<size_t> &shape,
               const std::vector<std::vector<int>> &rowptrs,
               const std::vector<std::vector<int>> &colinds, const size_t vsize,
               const Float value);
 
  [[nodiscard]] std::vector<size_t> get_shape() const { return shape; }
 
  [[nodiscard]] std::shared_ptr<Float> get_val() { return val; }
 
  [[nodiscard]] const std::shared_ptr<Float> get_val() const { return val; }
 
  [[nodiscard]] size_t get_nnz() const { return val_nnz; }
 
  [[nodiscard]] size_t get_alloc_nnz() const { return alloc_nnz; }
 
  [[nodiscard]] size_t get_first() const { return first; }
 
  [[nodiscard]] size_t get_offset() const { return get_first(); }
 
  void set_first(size_t i) { first = i; }
 
  void set_shape(const std::vector<size_t> &shape) { this->shape = shape; }
 
  [[nodiscard]] bool get_device_mem_stat() const { return *gpu_status; }
 
  [[nodiscard]] std::shared_ptr<bool> get_gpu_status() const {
    return gpu_status;
  }
 
  ~tensor_CRS() {
    if (val_create_flag) {
      if (get_device_mem_stat()) {
        device_free();
      }
    }
  }
 
  [[nodiscard]] const Float *data() const { return val.get(); }
 
  [[nodiscard]] Float *data() { return val.get(); }
 
  void resize(const size_t N, Float Val = 0) {
    if (first + N < alloc_nnz) {
      for (size_t i = val_nnz; i < N; ++i) {
        begin()[i] = Val;
      }
      val_nnz = N;
      return;
    }
    if (get_device_mem_stat()) {
      throw std::runtime_error("Error, GPU matrix cant use resize");
    }
    if (val_create_flag) {
      std::shared_ptr<Float> tmp(new Float[N], std::default_delete<Float[]>());
      size_t copy_size = std::min(val_nnz, N);
      for (size_t i = 0; i < copy_size; ++i) {
        tmp.get()[i] = data()[i];
      }
      for (size_t i = copy_size; i < N; ++i) {
        tmp.get()[i] = Val;
      }
      val = tmp;
      alloc_nnz = N;
      val_nnz = N;
      first = 0;
    } else {
      throw std::runtime_error("Error, not create vector cant use resize");
    }
  }
 
  [[nodiscard]] std::string type() const { return "tensor_CRS"; }
 
  void compute_hash();
 
  [[nodiscard]] size_t get_hash() const { return structure_hash; }
 
  [[nodiscard]] const Float *begin() const { return data() + get_offset(); }
 
  [[nodiscard]] Float *begin() { return data() + get_offset(); }
 
  [[nodiscard]] const Float *end() const {
    return data() + get_offset() + get_nnz();
  }
 
  [[nodiscard]] Float *end() { return data() + get_offset() + get_nnz(); }
 
  void fill(Float value);
 
  void operator=(const tensor_CRS<Float> &mat);
 
  [[nodiscard]] Float &operator[](size_t i) {
    if (get_device_mem_stat()) {
      throw std::runtime_error("Error, GPU vector cant use operator[]");
    }
    return data()[first + i];
  }
 
  [[nodiscard]] bool equal(const tensor_CRS<Float> &mat,
                           bool compare_cpu_and_device = false) const;
 
  [[nodiscard]] bool operator==(const tensor_CRS<Float> &mat) const;
 
  [[nodiscard]] bool operator!=(const tensor_CRS<Float> &mat) const;
};
 
} // namespace tensor
 
} // namespace monolish