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