gpu.h

   1 #ifndef _GPU_H
   2 #define _GPU_H
   3
   4 #include <isl/ast.h>
   5 #include <isl/id.h>
   6 #include <isl/id_to_ast_expr.h>
   7
   8 #include <pet.h>
   9
  10 #include "ppcg.h"
  11 #include "ppcg_options.h"
  12
  13 /* An access to an outer array element or an iterator.
  14  * Accesses to iterators have an access relation that maps to an unnamed space.
  15  * An access may be both read and write.
  16  * If the access relation is empty, then the output dimension may
  17  * not be equal to the dimension of the corresponding array.
  18  */
  19 struct gpu_stmt_access {
  20         /* Access reads elements */
  21         int read;
  22         /* Access writes elements */
  23         int write;
  24         /* All writes are definite writes. */
  25         int exact_write;
  26         /* The number of index expressions specified in the access. */
  27         int n_index;
  28
  29         /* May access relation */
  30         isl_map *access;
  31         /* May access relation with as domain a mapping from iteration domain
  32          * to a reference identifier.
  33          */
  34         isl_map *tagged_access;
  35         /* The reference id of the corresponding pet_expr. */
  36         isl_id *ref_id;
  37
  38         struct gpu_stmt_access *next;
  39 };
  40
  41 struct gpu_stmt {
  42         isl_id *id;
  43         struct pet_stmt *stmt;
  44
  45         /* Linked list of accesses. */
  46         struct gpu_stmt_access *accesses;
  47 };
  48
  49 /* Represents an outer array possibly accessed by a gpu_prog.
  50  */
  51 struct gpu_array_info {
  52         /* The array data space. */
  53         isl_space *space;
  54         /* Element type. */
  55         char *type;
  56         /* Element size. */
  57         int size;
  58         /* Name of the array. */
  59         char *name;
  60         /* Extent of the array that needs to be copied. */
  61         isl_set *extent;
  62         /* Number of indices. */
  63         unsigned n_index;
  64         /* For each index, a bound on "extent" in that direction. */
  65         isl_pw_aff **bound;
  66
  67         /* All references to this array; point to elements of a linked list. */
  68         int n_ref;
  69         struct gpu_stmt_access **refs;
  70
  71         /* Is this array accessed at all by the program? */
  72         int accessed;
  73
  74         /* Is this a scalar that is read-only within the entire program? */
  75         int read_only_scalar;
  76
  77         /* Are the elements of the array structures? */
  78         int has_compound_element;
  79
  80         /* Is the array local to the scop? */
  81         int local;
  82         /* Is the array local and should it be declared on the host? */
  83         int declare_local;
  84
  85         /* Is the corresponding global device memory accessed in any way? */
  86         int global;
  87
  88         /* Should the array be linearized? */
  89         int linearize;
  90
  91         /* Order dependences on this array.
  92          * Only used if live_range_reordering option is set.
  93          * It is set to NULL otherwise.
  94          */
  95         isl_union_map *dep_order;
  96 };
  97
  98 /* Represents an outer array accessed by a ppcg_kernel, localized
  99  * to the context of this kernel.
 100  *
 101  * "array" points to the corresponding array in the gpu_prog.
 102  * The "n_group" "groups" are the reference groups associated to the array.
 103  * If "force_private" is set, then the array (in practice a scalar)
 104  * must be mapped to a register.
 105  * "global" is set if the global device memory corresponding
 106  * to this array is accessed by the kernel.
 107  * For each index i with 0 <= i < n_index,
 108  * bound[i] is equal to array->bound[i] specialized to the current kernel.
 109  */
 110 struct gpu_local_array_info {
 111         struct gpu_array_info *array;
 112
 113         int n_group;
 114         struct gpu_array_ref_group **groups;
 115
 116         int force_private;
 117         int global;
 118
 119         unsigned n_index;
 120         isl_pw_aff_list *bound;
 121 };
 122
 123 __isl_give isl_ast_expr *gpu_local_array_info_linearize_index(
 124         struct gpu_local_array_info *array, __isl_take isl_ast_expr *expr);
 125
 126 /* A sequence of "n" names of types.
 127  */
 128 struct gpu_types {
 129         int n;
 130         char **name;
 131 };
 132
 133 /* "read" and "write" contain the original access relations, possibly
 134  * involving member accesses.
 135  *
 136  * The elements of "array", as well as the ranges of "copy_in" and "copy_out"
 137  * only refer to the outer arrays of any possible member accesses.
 138  */
 139 struct gpu_prog {
 140         isl_ctx *ctx;
 141
 142         struct ppcg_scop *scop;
 143
 144         /* Set of parameter values */
 145         isl_set *context;
 146
 147         /* All potential read accesses in the entire program */
 148         isl_union_map *read;
 149
 150         /* All potential write accesses in the entire program */
 151         isl_union_map *may_write;
 152         /* All definite write accesses in the entire program */
 153         isl_union_map *must_write;
 154         /* All tagged definite kills in the entire program */
 155         isl_union_map *tagged_must_kill;
 156
 157         /* The set of inner array elements that may be preserved. */
 158         isl_union_set *may_persist;
 159
 160         /* A mapping from all innermost arrays to their outer arrays. */
 161         isl_union_map *to_outer;
 162         /* A mapping from the outer arrays to all corresponding inner arrays. */
 163         isl_union_map *to_inner;
 164         /* A mapping from all intermediate arrays to their outer arrays,
 165          * including an identity mapping from the anoymous 1D space to itself.
 166          */
 167         isl_union_map *any_to_outer;
 168
 169         /* Order dependences on non-scalars. */
 170         isl_union_map *array_order;
 171
 172         /* Array of statements */
 173         int n_stmts;
 174         struct gpu_stmt *stmts;
 175
 176         int n_array;
 177         struct gpu_array_info *array;
 178 };
 179
 180 struct gpu_gen {
 181         isl_ctx *ctx;
 182         struct ppcg_options *options;
 183
 184         /* Callback for printing of AST in appropriate format. */
 185         __isl_give isl_printer *(*print)(__isl_take isl_printer *p,
 186                 struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 187                 struct gpu_types *types, void *user);
 188         void *print_user;
 189
 190         struct gpu_prog *prog;
 191         /* The generated AST. */
 192         isl_ast_node *tree;
 193
 194         /* The sequence of types for which a definition has been printed. */
 195         struct gpu_types types;
 196
 197         /* User specified tile, grid and block sizes for each kernel */
 198         isl_union_map *sizes;
 199
 200         /* Effectively used tile, grid and block sizes for each kernel */
 201         isl_union_map *used_sizes;
 202
 203         /* Identifier of the next kernel. */
 204         int kernel_id;
 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  * the first n_grid elements of grid_dim represent the specified size
 282  * of the grid.
 283  * the first n_block elements of block_dim represent the specified or
 284  * effective size of the block.
 285  * Note that in the input file, the sizes of the grid and the blocks
 286  * are specified in the order x, y, z, but internally, the sizes
 287  * are stored in reverse order, so that the last element always
 288  * refers to the x dimension.
 289  *
 290  * grid_size reflects the effective grid size.
 291  *
 292  * context contains the values of the parameters and outer schedule dimensions
 293  * for which any statement instance in this kernel needs to be executed.
 294  *
 295  * n_sync is the number of synchronization operations that have
 296  * been introduced in the schedule tree corresponding to this kernel (so far).
 297  *
 298  * core contains the spaces of the statement domains that form
 299  * the core computation of the kernel.  It is used to navigate
 300  * the tree during the construction of the device part of the schedule
 301  * tree in create_kernel.
 302  *
 303  * arrays is the set of possibly accessed outer array elements.
 304  *
 305  * space is the schedule space of the AST context.  That is, it represents
 306  * the loops of the generated host code containing the kernel launch.
 307  *
 308  * n_array is the total number of arrays in the input program and also
 309  * the number of element in the array array.
 310  * array contains information about each array that is local
 311  * to the current kernel.  If an array is not used in a kernel,
 312  * then the corresponding entry does not contain any information.
 313  *
 314  * any_force_private is set if any array in the kernel is marked force_private
 315  *
 316  * block_filter contains constraints on the domain elements in the kernel
 317  * that encode the mapping to block identifiers, where the block identifiers
 318  * are represented by "n_grid" parameters with as names the elements
 319  * of "block_ids".
 320  *
 321  * thread_filter contains constraints on the domain elements in the kernel
 322  * that encode the mapping to thread identifiers, where the thread identifiers
 323  * are represented by "n_block" parameters with as names the elements
 324  * of "thread_ids".
 325  *
 326  * shared_schedule corresponds to the schedule dimensions of
 327  * the (tiled) schedule for this kernel that have been taken into account
 328  * for computing private/shared memory tiles.
 329  * shared_schedule_dim is the dimension of this schedule.
 330  *
 331  * sync_writes contains write references that require synchronization.
 332  * Each reference is represented by a universe set in a space [S[i,j] -> R[]]
 333  * with S[i,j] the statement instance space and R[] the array reference.
 334  */
 335 struct ppcg_kernel {
 336         isl_ctx *ctx;
 337         struct ppcg_options *options;
 338
 339         struct gpu_prog *prog;
 340
 341         int id;
 342
 343         isl_id_list *block_ids;
 344         isl_id_list *thread_ids;
 345
 346         int n_grid;
 347         int n_block;
 348         int grid_dim[2];
 349         int block_dim[3];
 350
 351         isl_multi_pw_aff *grid_size;
 352         isl_set *context;
 353
 354         int n_sync;
 355         isl_union_set *core;
 356         isl_union_set *arrays;
 357
 358         isl_space *space;
 359
 360         int n_array;
 361         struct gpu_local_array_info *array;
 362
 363         int n_var;
 364         struct ppcg_kernel_var *var;
 365
 366         int any_force_private;
 367
 368         isl_union_set *block_filter;
 369         isl_union_set *thread_filter;
 370         isl_union_pw_multi_aff *shared_schedule;
 371         int shared_schedule_dim;
 372
 373         isl_union_set *sync_writes;
 374
 375         isl_ast_node *tree;
 376 };
 377
 378 int gpu_array_is_scalar(struct gpu_array_info *array);
 379 int gpu_array_is_read_only_scalar(struct gpu_array_info *array);
 380 int gpu_array_requires_device_allocation(struct gpu_array_info *array);
 381 __isl_give isl_set *gpu_array_positive_size_guard(struct gpu_array_info *array);
 382
 383 struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct ppcg_scop *scop);
 384 void *gpu_prog_free(struct gpu_prog *prog);
 385
 386 int ppcg_kernel_requires_array_argument(struct ppcg_kernel *kernel, int i);
 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