gpu.h

   1 #ifndef _GPU_H
   2 #define _GPU_H
   3
   4 #include <isl/ast.h>
   5 #include <isl/id_to_ast_expr.h>
   6
   7 #include "ppcg.h"
   8 #include "ppcg_options.h"
   9
  10 /* Represents an outer array possibly accessed by a gpu_prog.
  11  */
  12 struct gpu_array_info {
  13         /* The array data space. */
  14         isl_space *space;
  15         /* Element type. */
  16         char *type;
  17         /* Element size. */
  18         int size;
  19         /* Name of the array. */
  20         char *name;
  21         /* Extent of the array that needs to be copied. */
  22         isl_set *extent;
  23         /* Number of indices. */
  24         unsigned n_index;
  25         /* For each index, a bound on "extent" in that direction. */
  26         isl_pw_aff **bound;
  27
  28         /* All references to this array; point to elements of a linked list. */
  29         int n_ref;
  30         struct gpu_stmt_access **refs;
  31
  32         /* Is this array accessed at all by the program? */
  33         int accessed;
  34
  35         /* Is this a scalar that is read-only within the entire program? */
  36         int read_only_scalar;
  37
  38         /* Are the elements of the array structures? */
  39         int has_compound_element;
  40
  41         /* Is the array local to the scop? */
  42         int local;
  43
  44         /* Should the array be linearized? */
  45         int linearize;
  46
  47         /* Order dependences on this array.
  48          * Only used if live_range_reordering option is set.
  49          * It is set to NULL otherwise.
  50          */
  51         isl_union_map *dep_order;
  52 };
  53
  54 /* Represents an outer array accessed by a ppcg_kernel, localized
  55  * to the context of this kernel.
  56  *
  57  * "array" points to the corresponding array in the gpu_prog.
  58  * The "n_group" "groups" are the reference groups associated to the array.
  59  * If the outer array represented by the gpu_local_array_info
  60  * contains structures, then the references are not
  61  * collected and the reference groups are not computed.
  62  * If "force_private" is set, then the array (in practice a scalar)
  63  * must be mapped to a register.
  64  * For each index i with 0 <= i < n_index,
  65  * bound[i] is equal to array->bound[i] specialized to the current kernel.
  66  */
  67 struct gpu_local_array_info {
  68         struct gpu_array_info *array;
  69
  70         int n_group;
  71         struct gpu_array_ref_group **groups;
  72
  73         int force_private;
  74
  75         unsigned n_index;
  76         isl_pw_aff_list *bound;
  77 };
  78
  79 __isl_give isl_ast_expr *gpu_local_array_info_linearize_index(
  80         struct gpu_local_array_info *array, __isl_take isl_ast_expr *expr);
  81
  82 /* A sequence of "n" names of types.
  83  */
  84 struct gpu_types {
  85         int n;
  86         char **name;
  87 };
  88
  89 /* "read" and "write" contain the original access relations, possibly
  90  * involving member accesses.
  91  *
  92  * The elements of "array", as well as the ranges of "copy_in" and "copy_out"
  93  * only refer to the outer arrays of any possible member accesses.
  94  */
  95 struct gpu_prog {
  96         isl_ctx *ctx;
  97
  98         struct ppcg_scop *scop;
  99
 100         /* Set of parameter values */
 101         isl_set *context;
 102
 103         /* All potential read accesses in the entire program */
 104         isl_union_map *read;
 105
 106         /* All potential write accesses in the entire program */
 107         isl_union_map *may_write;
 108         /* All definite write accesses in the entire program */
 109         isl_union_map *must_write;
 110         /* All tagged definite kills in the entire program */
 111         isl_union_map *tagged_must_kill;
 112
 113         /* The set of inner array elements that may be preserved. */
 114         isl_union_set *may_persist;
 115
 116         /* Set of outer array elements that need to be copied in. */
 117         isl_union_set *copy_in;
 118         /* Set of outer array elements that need to be copied out. */
 119         isl_union_set *copy_out;
 120
 121         /* A mapping from all innermost arrays to their outer arrays. */
 122         isl_union_map *to_outer;
 123         /* A mapping from the outer arrays to all corresponding inner arrays. */
 124         isl_union_map *to_inner;
 125         /* A mapping from all intermediate arrays to their outer arrays,
 126          * including an identity mapping from the anoymous 1D space to itself.
 127          */
 128         isl_union_map *any_to_outer;
 129
 130         /* Order dependences on non-scalars. */
 131         isl_union_map *array_order;
 132
 133         /* Array of statements */
 134         int n_stmts;
 135         struct gpu_stmt *stmts;
 136
 137         int n_array;
 138         struct gpu_array_info *array;
 139 };
 140
 141 struct gpu_gen {
 142         isl_ctx *ctx;
 143         struct ppcg_options *options;
 144
 145         /* Callback for printing of AST in appropriate format. */
 146         __isl_give isl_printer *(*print)(__isl_take isl_printer *p,
 147                 struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 148                 struct gpu_types *types, void *user);
 149         void *print_user;
 150
 151         struct gpu_prog *prog;
 152         /* The generated AST. */
 153         isl_ast_node *tree;
 154
 155         /* The sequence of types for which a definition has been printed. */
 156         struct gpu_types types;
 157
 158         /* User specified tile, grid and block sizes for each kernel */
 159         isl_union_map *sizes;
 160
 161         /* Effectively used tile, grid and block sizes for each kernel */
 162         isl_union_map *used_sizes;
 163
 164         /* Pointer to current ppcg_kernel. */
 165         isl_id *kernel_mark;
 166         /* Identifier of the next kernel. */
 167         int kernel_id;
 168         /* Pointer to the current kernel. */
 169         struct ppcg_kernel *kernel;
 170         /* Does the computed schedule exhibit any parallelism? */
 171         int any_parallelism;
 172
 173         /* First tile dimension. */
 174         int tile_first;
 175         /* Number of tile dimensions. */
 176         int tile_len;
 177         /* Number of initial parallel loops among tile dimensions. */
 178         int n_parallel;
 179
 180         /* Number of dimensions determining shared memory. */
 181         int shared_len;
 182
 183         /* Number of rows in the untiled schedule. */
 184         int untiled_len;
 185         /* Number of rows in the tiled schedule. */
 186         int tiled_len;
 187         /* Number of rows in schedule after tiling/wrapping over threads. */
 188         int thread_tiled_len;
 189
 190         /* A schedule tree corresponding to the host code. */
 191         isl_schedule *host_schedule;
 192         /* Global untiled schedule. */
 193         isl_union_map *sched;
 194         /* Local (per kernel launch) tiled schedule. */
 195         isl_union_map *tiled_sched;
 196         /* Local schedule per shared memory tile loop iteration. */
 197         isl_union_map *local_sched;
 198
 199         /* Local tiled schedule projected onto the shared tile loops and
 200          * the loops that will be wrapped over the threads,
 201          * with all shared tile loops parametrized.
 202          */
 203         isl_union_map *shared_sched;
 204         /* Projects out the loops that will be wrapped over the threads
 205          * from shared_sched.
 206          */
 207         isl_union_map *shared_proj;
 208
 209         /* A map that takes the range of shared_sched as input,
 210          * wraps the appropriate loops over the threads and then projects
 211          * out these loops.
 212          */
 213         isl_map *privatization;
 214
 215         /* The array reference group corresponding to copy_sched. */
 216         struct gpu_array_ref_group *copy_group;
 217
 218         /* First loop to unroll (or -1 if none) in the current part of the
 219          * schedule.
 220          */
 221         int first_unroll;
 222
 223         int n_grid;
 224         int n_block;
 225         /* Note: in the input file, the sizes of the grid and the blocks
 226          * are specified in the order x, y, z, but internally, the sizes
 227          * are stored in reverse order, so that the last element always
 228          * refers to the x dimension.
 229          */
 230         int grid_dim[2];
 231         int block_dim[3];
 232         int *tile_size;
 233 };
 234
 235 enum ppcg_kernel_access_type {
 236         ppcg_access_global,
 237         ppcg_access_shared,
 238         ppcg_access_private
 239 };
 240
 241 enum ppcg_kernel_stmt_type {
 242         ppcg_kernel_copy,
 243         ppcg_kernel_domain,
 244         ppcg_kernel_sync
 245 };
 246
 247 /* Representation of special statements, in particular copy statements
 248  * and __syncthreads statements, inside a kernel.
 249  *
 250  * type represents the kind of statement
 251  *
 252  *
 253  * for ppcg_kernel_copy statements we have
 254  *
 255  * read is set if the statement should copy data from global memory
 256  * to shared memory or registers.
 257  *
 258  * index expresses an access to the array element that needs to be copied
 259  * local_index expresses the corresponding element in the tile
 260  *
 261  * array refers to the original array being copied
 262  * local_array is a pointer to the appropriate element in the "array"
 263  *      array of the ppcg_kernel to which this copy access belongs
 264  *
 265  *
 266  * for ppcg_kernel_domain statements we have
 267  *
 268  * stmt is the corresponding input statement
 269  *
 270  * n_access is the number of accesses in stmt
 271  * access is an array of local information about the accesses
 272  */
 273 struct ppcg_kernel_stmt {
 274         enum ppcg_kernel_stmt_type type;
 275
 276         union {
 277                 struct {
 278                         int read;
 279                         isl_ast_expr *index;
 280                         isl_ast_expr *local_index;
 281                         struct gpu_array_info *array;
 282                         struct gpu_local_array_info *local_array;
 283                 } c;
 284                 struct {
 285                         struct gpu_stmt *stmt;
 286                         isl_id_to_ast_expr *ref2expr;
 287                 } d;
 288         } u;
 289 };
 290
 291 /* Representation of a local variable in a kernel.
 292  */
 293 struct ppcg_kernel_var {
 294         struct gpu_array_info *array;
 295         enum ppcg_kernel_access_type type;
 296         char *name;
 297         isl_vec *size;
 298 };
 299
 300 /* Representation of a kernel.
 301  *
 302  * id is the sequence number of the kernel.
 303  *
 304  * block_ids contains the list of block identifiers for this kernel.
 305  * thread_ids contains the list of thread identifiers for this kernel.
 306  *
 307  * the first n_block elements of block_dim represent the effective size
 308  * of the block.
 309  *
 310  * grid_size reflects the effective grid size.
 311  *
 312  * context contains the values of the parameters and outer schedule dimensions
 313  * for which any statement instance in this kernel needs to be executed.
 314  *
 315  * arrays is the set of possibly accessed outer array elements.
 316  *
 317  * space is the schedule space of the AST context.  That is, it represents
 318  * the loops of the generated host code containing the kernel launch.
 319  *
 320  * n_array is the total number of arrays in the input program and also
 321  * the number of element in the array array.
 322  * array contains information about each array that is local
 323  * to the current kernel.  If an array is not used in a kernel,
 324  * then the corresponding entry does not contain any information.
 325  *
 326  * any_force_private is set if any array in the kernel is marked force_private
 327  */
 328 struct ppcg_kernel {
 329         isl_ctx *ctx;
 330         struct ppcg_options *options;
 331
 332         int id;
 333
 334         isl_id_list *block_ids;
 335         isl_id_list *thread_ids;
 336
 337         int n_block;
 338         int block_dim[3];
 339
 340         isl_multi_pw_aff *grid_size;
 341         isl_set *context;
 342
 343         isl_union_set *arrays;
 344
 345         isl_space *space;
 346
 347         int n_array;
 348         struct gpu_local_array_info *array;
 349
 350         int n_var;
 351         struct ppcg_kernel_var *var;
 352
 353         int any_force_private;
 354
 355         isl_ast_node *tree;
 356 };
 357
 358 int gpu_array_is_scalar(struct gpu_array_info *array);
 359 int gpu_array_is_read_only_scalar(struct gpu_array_info *array);
 360 __isl_give isl_set *gpu_array_positive_size_guard(struct gpu_array_info *array);
 361
 362 struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct ppcg_scop *scop);
 363 void *gpu_prog_free(struct gpu_prog *prog);
 364
 365 int generate_gpu(isl_ctx *ctx, const char *input, FILE *out,
 366         struct ppcg_options *options,
 367         __isl_give isl_printer *(*print)(__isl_take isl_printer *p,
 368                 struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 369                 struct gpu_types *types, void *user), void *user);
 370
 371 #endif