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         /* Identifier of the next kernel. */
 165         int kernel_id;
 166         /* Pointer to the current kernel. */
 167         struct ppcg_kernel *kernel;
 168         /* Does the computed schedule exhibit any parallelism? */
 169         int any_parallelism;
 170
 171         /* First tile dimension. */
 172         int tile_first;
 173
 174         /* Number of dimensions determining shared memory. */
 175         int shared_len;
 176
 177         /* Number of rows in the untiled schedule. */
 178         int untiled_len;
 179         /* Number of rows in the tiled schedule. */
 180         int tiled_len;
 181         /* Number of rows in schedule after tiling/wrapping over threads. */
 182         int thread_tiled_len;
 183
 184         /* A schedule tree corresponding to both the host and device code. */
 185         isl_schedule *schedule;
 186         /* Global untiled schedule. */
 187         isl_union_map *sched;
 188         /* Local (per kernel launch) tiled schedule. */
 189         isl_union_map *tiled_sched;
 190
 191         /* Local tiled schedule projected onto the shared tile loops and
 192          * the loops that will be wrapped over the threads,
 193          * with all shared tile loops parametrized.
 194          */
 195         isl_union_map *shared_sched;
 196         /* Projects out the loops that will be wrapped over the threads
 197          * from shared_sched.
 198          */
 199         isl_union_map *shared_proj;
 200
 201         /* First loop to unroll (or -1 if none) in the current part of the
 202          * schedule.
 203          */
 204         int first_unroll;
 205 };
 206
 207 enum ppcg_kernel_access_type {
 208         ppcg_access_global,
 209         ppcg_access_shared,
 210         ppcg_access_private
 211 };
 212
 213 enum ppcg_kernel_stmt_type {
 214         ppcg_kernel_copy,
 215         ppcg_kernel_domain,
 216         ppcg_kernel_sync
 217 };
 218
 219 /* Representation of special statements, in particular copy statements
 220  * and __syncthreads statements, inside a kernel.
 221  *
 222  * type represents the kind of statement
 223  *
 224  *
 225  * for ppcg_kernel_copy statements we have
 226  *
 227  * read is set if the statement should copy data from global memory
 228  * to shared memory or registers.
 229  *
 230  * index expresses an access to the array element that needs to be copied
 231  * local_index expresses the corresponding element in the tile
 232  *
 233  * array refers to the original array being copied
 234  * local_array is a pointer to the appropriate element in the "array"
 235  *      array of the ppcg_kernel to which this copy access belongs
 236  *
 237  *
 238  * for ppcg_kernel_domain statements we have
 239  *
 240  * stmt is the corresponding input statement
 241  *
 242  * n_access is the number of accesses in stmt
 243  * access is an array of local information about the accesses
 244  */
 245 struct ppcg_kernel_stmt {
 246         enum ppcg_kernel_stmt_type type;
 247
 248         union {
 249                 struct {
 250                         int read;
 251                         isl_ast_expr *index;
 252                         isl_ast_expr *local_index;
 253                         struct gpu_array_info *array;
 254                         struct gpu_local_array_info *local_array;
 255                 } c;
 256                 struct {
 257                         struct gpu_stmt *stmt;
 258                         isl_id_to_ast_expr *ref2expr;
 259                 } d;
 260         } u;
 261 };
 262
 263 /* Representation of a local variable in a kernel.
 264  */
 265 struct ppcg_kernel_var {
 266         struct gpu_array_info *array;
 267         enum ppcg_kernel_access_type type;
 268         char *name;
 269         isl_vec *size;
 270 };
 271
 272 /* Representation of a kernel.
 273  *
 274  * prog describes the original code from which the kernel is extracted.
 275  *
 276  * id is the sequence number of the kernel.
 277  *
 278  * block_ids contains the list of block identifiers for this kernel.
 279  * thread_ids contains the list of thread identifiers for this kernel.
 280  *
 281  * tile_len is the number of tile dimensions and
 282  * n_parallel is the number of initial parallel loops among those
 283  * tile dimensions.
 284  * tile_size is an array of length tile_len containing the tile sizes.
 285  *
 286  * the first n_grid elements of grid_dim represent the specified size
 287  * of the grid.
 288  * the first n_block elements of block_dim represent the specified or
 289  * effective size of the block.
 290  * Note that in the input file, the sizes of the grid and the blocks
 291  * are specified in the order x, y, z, but internally, the sizes
 292  * are stored in reverse order, so that the last element always
 293  * refers to the x dimension.
 294  *
 295  * grid_size reflects the effective grid size.
 296  *
 297  * context contains the values of the parameters and outer schedule dimensions
 298  * for which any statement instance in this kernel needs to be executed.
 299  *
 300  * n_sync is the number of synchronization operations that have
 301  * been introduced in the schedule tree corresponding to this kernel (so far).
 302  *
 303  * core contains the spaces of the statement domains that form
 304  * the core computation of the kernel.  It is used to navigate
 305  * the tree during the construction of the device part of the schedule
 306  * tree in create_kernel.
 307  *
 308  * arrays is the set of possibly accessed outer array elements.
 309  *
 310  * space is the schedule space of the AST context.  That is, it represents
 311  * the loops of the generated host code containing the kernel launch.
 312  *
 313  * n_array is the total number of arrays in the input program and also
 314  * the number of element in the array array.
 315  * array contains information about each array that is local
 316  * to the current kernel.  If an array is not used in a kernel,
 317  * then the corresponding entry does not contain any information.
 318  *
 319  * any_force_private is set if any array in the kernel is marked force_private
 320  *
 321  * block_filter contains constraints on the domain elements in the kernel
 322  * that encode the mapping to block identifiers, where the block identifiers
 323  * are represented by "n_grid" parameters with as names the elements
 324  * of "block_ids".
 325  *
 326  * thread_filter contains constraints on the domain elements in the kernel
 327  * that encode the mapping to thread identifiers, where the thread identifiers
 328  * are represented by "n_block" parameters with as names the elements
 329  * of "thread_ids".
 330  *
 331  * shared_schedule corresponds to the schedule dimensions of
 332  * the (tiled) schedule for this kernel that have been taken into account
 333  * for computing private/shared memory tiles.
 334  * shared_schedule_dim is the dimension of this schedule.
 335  */
 336 struct ppcg_kernel {
 337         isl_ctx *ctx;
 338         struct ppcg_options *options;
 339
 340         struct gpu_prog *prog;
 341
 342         int id;
 343
 344         isl_id_list *block_ids;
 345         isl_id_list *thread_ids;
 346
 347         int tile_len;
 348         int *tile_size;
 349         int n_parallel;
 350
 351         int n_grid;
 352         int n_block;
 353         int grid_dim[2];
 354         int block_dim[3];
 355
 356         isl_multi_pw_aff *grid_size;
 357         isl_set *context;
 358
 359         int n_sync;
 360         isl_union_set *core;
 361         isl_union_set *arrays;
 362
 363         isl_space *space;
 364
 365         int n_array;
 366         struct gpu_local_array_info *array;
 367
 368         int n_var;
 369         struct ppcg_kernel_var *var;
 370
 371         int any_force_private;
 372
 373         isl_union_set *block_filter;
 374         isl_union_set *thread_filter;
 375         isl_union_pw_multi_aff *shared_schedule;
 376         int shared_schedule_dim;
 377
 378         isl_ast_node *tree;
 379 };
 380
 381 int gpu_array_is_scalar(struct gpu_array_info *array);
 382 int gpu_array_is_read_only_scalar(struct gpu_array_info *array);
 383 __isl_give isl_set *gpu_array_positive_size_guard(struct gpu_array_info *array);
 384
 385 struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct ppcg_scop *scop);
 386 void *gpu_prog_free(struct gpu_prog *prog);
 387
 388 int generate_gpu(isl_ctx *ctx, const char *input, FILE *out,
 389         struct ppcg_options *options,
 390         __isl_give isl_printer *(*print)(__isl_take isl_printer *p,
 391                 struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 392                 struct gpu_types *types, void *user), void *user);
 393
 394 #endif