grid_dgemm_tensor_local.h

Go to the documentation of this file.

/*----------------------------------------------------------------------------*/
/*  CP2K: A general program to perform molecular dynamics simulations         */
/*  Copyright 2000-2024 CP2K developers group <https://cp2k.org>              */
/*                                                                            */
/*  SPDX-License-Identifier: BSD-3-Clause                                     */
/*----------------------------------------------------------------------------*/
 
#ifndef GRID_DGEMM_TENSOR_LOCAL_H
#define GRID_DGEMM_TENSOR_LOCAL_H
 
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
 
typedef struct tensor_ {
  int dim_;
  int size[4];
  size_t alloc_size_;
  size_t old_alloc_size_;
  int offsets[4];
  double *data;
  int ld_;
  int window_shift[4]; /* lower corner of the window. Should be between lower
                        * corner and upper corner of the local grid */
  int window_size[4];  /* size of the window where computations should be
                        * done */
  int full_size[4];    /* size of the global grid */
  int lower_corner[4]; /* coordinates of the lower corner of the local part of
                        * the grid. It can be different from the window where
                        * computations should be done. The upper corner can be
                        * deduced with the sum of the grid size and the lower
                        * corner */
  /* only relevant when the tensor represents a grid */
  double dh[3][3];
  double dh_inv[3][3];
  bool orthogonal[3];
} tensor;
 
extern void tensor_copy(tensor *const b, const tensor *const a);
 
/* initialize a tensor structure for a tensor of dimension dim <= 4 */
 
static inline void initialize_tensor(struct tensor_ *a, const int dim,
                                     const int *const sizes) {
  if (a == NULL)
    return;
 
  a->dim_ = dim;
  for (int d = 0; d < dim; d++)
    a->size[d] = sizes[d];
 
  // we need proper alignment here. But can be done later
  /* a->ld_ = (sizes[a->dim_ - 1] / 32 + 1) * 32; */
  a->ld_ = sizes[a->dim_ - 1];
  switch (a->dim_) {
  case 4: {
    a->offsets[0] = a->ld_ * a->size[1] * a->size[2];
    a->offsets[1] = a->ld_ * a->size[2];
    a->offsets[2] = a->ld_;
    break;
  }
  case 3: {
    a->offsets[0] = a->ld_ * a->size[1];
    a->offsets[1] = a->ld_;
  } break;
  case 2: { // matrix case
    a->offsets[0] = a->ld_;
  } break;
  case 1:
    break;
  }
 
  a->alloc_size_ = a->offsets[0] * a->size[0];
  return;
}
 
/* initialize a tensor structure for a tensor of dimension dim = 2 */
 
static inline void initialize_tensor_2(struct tensor_ *a, int n1, int n2) {
  if (a == NULL)
    return;
 
  int size_[2] = {n1, n2};
  initialize_tensor(a, 2, size_);
}
 
/* initialize a tensor structure for a tensor of dimension dim = 2 */
 
static inline void initialize_tensor_3(struct tensor_ *a, int n1, int n2,
                                       int n3) {
  if (a == NULL)
    return;
  int size_[3] = {n1, n2, n3};
  initialize_tensor(a, 3, size_);
}
 
/* initialize a tensor structure for a tensor of dimension dim = 2 */
 
static inline void initialize_tensor_4(struct tensor_ *a, int n1, int n2,
                                       int n3, int n4) {
  if (a == NULL)
    return;
  int size_[4] = {n1, n2, n3, n4};
  initialize_tensor(a, 4, size_);
}
 
/* initialize a tensor structure for a tensor of dimension dim = 2 */
 
static inline tensor *create_tensor(const int dim, const int *sizes) {
  tensor *a = (tensor *)malloc(sizeof(struct tensor_));
 
  if (a == NULL)
    abort();
 
  initialize_tensor(a, dim, sizes);
  a->data = (double *)malloc(sizeof(double) * a->alloc_size_);
  if (a->data == NULL)
    abort();
  a->old_alloc_size_ = a->alloc_size_;
  return a;
}
 
/* destroy a tensor created with the function above */
static inline void destroy_tensor(tensor *a) {
  if (a->data)
    free(a->data);
  free(a);
}
 
static inline size_t tensor_return_memory_size(const struct tensor_ *const a) {
  if (a == NULL)
    abort();
 
  return a->alloc_size_;
}
 
static inline void tensor_assign_memory(struct tensor_ *a, void *data) {
  if (a == NULL)
    abort();
  a->data = (double *)data;
}
 
static inline int tensor_get_leading_dimension(struct tensor_ *a) {
  if (a == NULL)
    abort();
  return a->ld_;
}
 
static inline void tensor_set_leading_dimension(struct tensor_ *a,
                                                const int ld) {
  if (a == NULL)
    abort();
  a->ld_ = ld;
}
 
static inline void recompute_tensor_offsets(struct tensor_ *a) {
  if (a == NULL)
    abort();
 
  switch (a->dim_) {
  case 5: {
    a->offsets[0] = a->ld_ * a->size[1] * a->size[2] * a->size[3];
    a->offsets[1] = a->ld_ * a->size[1] * a->size[2];
    a->offsets[2] = a->ld_ * a->size[2];
    a->offsets[3] = a->ld_;
    break;
  }
  case 4: {
    a->offsets[0] = a->ld_ * a->size[1] * a->size[2];
    a->offsets[1] = a->ld_ * a->size[2];
    a->offsets[2] = a->ld_;
    break;
  }
  case 3: {
    a->offsets[0] = a->ld_ * a->size[1];
    a->offsets[1] = a->ld_;
  } break;
  case 2: { // matrix case
    a->offsets[0] = a->ld_;
  } break;
  case 1:
    break;
  }
}
 
static inline size_t compute_memory_space_tensor_3(const int n1, const int n2,
                                                   const int n3) {
  return (n1 * n2 * n3);
}
 
static inline size_t compute_memory_space_tensor_4(const int n1, const int n2,
                                                   const int n3, const int n4) {
  return (n1 * n2 * n3 * n4);
}
 
static inline void setup_global_grid_size(tensor *const grid,
                                          const int *const full_size) {
  switch (grid->dim_) {
  case 1:
    grid->full_size[0] = full_size[0];
    break;
  case 2: {
    grid->full_size[1] = full_size[0];
    grid->full_size[0] = full_size[1];
  } break;
  case 3: {
    grid->full_size[0] = full_size[2];
    grid->full_size[1] = full_size[1];
    grid->full_size[2] = full_size[0];
  } break;
  default:
    for (int d = 0; d < grid->dim_; d++)
      grid->full_size[d] = full_size[grid->dim_ - d - 1];
    break;
  }
}
 
static inline void setup_grid_window(tensor *const grid,
                                     const int *const shift_local,
                                     const int *const border_width,
                                     const int border_mask) {
  for (int d = 0; d < grid->dim_; d++) {
    grid->lower_corner[d] = shift_local[grid->dim_ - d - 1];
    grid->window_shift[d] = 0;
    grid->window_size[d] = grid->size[d];
    if (grid->size[d] != grid->full_size[d]) {
      grid->window_size[d]--;
    }
  }
 
  if (border_width) {
    if (border_mask & (1 << 0))
      grid->window_shift[2] += border_width[0];
    if (border_mask & (1 << 1))
      grid->window_size[2] -= border_width[0];
    if (border_mask & (1 << 2))
      grid->window_shift[1] += border_width[1];
    if (border_mask & (1 << 3))
      grid->window_size[1] -= border_width[1];
    if (border_mask & (1 << 4))
      grid->window_shift[0] += border_width[2];
    if (border_mask & (1 << 5))
      grid->window_size[0] -= border_width[2];
  }
}
 
extern size_t realloc_tensor(tensor *t);
extern void alloc_tensor(tensor *t);
 
#define idx5(a, i, j, k, l, m)                                                 \
  a.data[(i)*a.offsets[0] + (j)*a.offsets[1] + (k)*a.offsets[2] + (l)*a.ld_ + m]
#define idx4(a, i, j, k, l)                                                    \
  a.data[(i)*a.offsets[0] + (j)*a.offsets[1] + (k)*a.ld_ + (l)]
#define idx3(a, i, j, k) a.data[(i)*a.offsets[0] + (j)*a.ld_ + (k)]
#define idx2(a, i, j) a.data[(i)*a.ld_ + (j)]
#endif