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