gpu_array_tile.h

   1 #ifndef GPU_ARRAY_TILE_H
   2 #define GPU_ARRAY_TILE_H
   3
   4 #include <isl/aff_type.h>
   5 #include <isl/map_type.h>
   6 #include <isl/val.h>
   7
   8 /* The fields stride and shift only contain valid information
   9  * if shift != NULL.
  10  * If so, they express that current index is such that if you add shift,
  11  * then the result is always a multiple of stride.
  12  * Let D represent the initial tile->depth dimensions of the computed schedule.
  13  * The spaces of "lb" and "shift" are of the form
  14  *
  15  *      D -> [b]
  16  */
  17 struct gpu_array_bound {
  18         isl_val *size;
  19         isl_aff *lb;
  20
  21         isl_val *stride;
  22         isl_aff *shift;
  23 };
  24
  25 /* A tile of an outer array.
  26  *
  27  * requires_unroll is set if the schedule dimensions that are mapped
  28  * to threads need to be unrolled for this (private) tile to be used.
  29  *
  30  * "depth" reflects the number of schedule dimensions that affect the tile.
  31  * The copying into and/or out of the tile is performed at that depth.
  32  *
  33  * n is the dimension of the array.
  34  * bound is an array of size "n" representing the lower bound
  35  *      and size for each index.
  36  *
  37  * tiling maps a tile in the global array to the corresponding
  38  * shared/private memory tile and is of the form
  39  *
  40  *      { [D[i] -> A[a]] -> T[(a + shift(i))/stride - lb(i)] }
  41  *
  42  * where D represents the initial "depth" dimensions
  43  * of the computed schedule.
  44  */
  45 struct gpu_array_tile {
  46         isl_ctx *ctx;
  47         int requires_unroll;
  48         int depth;
  49         int n;
  50         struct gpu_array_bound *bound;
  51         isl_multi_aff *tiling;
  52 };
  53
  54 struct gpu_array_tile *gpu_array_tile_create(isl_ctx *ctx, int n_index);
  55 struct gpu_array_tile *gpu_array_tile_free(struct gpu_array_tile *tile);
  56
  57 __isl_give isl_val *gpu_array_tile_size(struct gpu_array_tile *tile);
  58
  59 #endif