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 rows in the untiled schedule. */
 175         int untiled_len;
 176
 177         /* A schedule tree corresponding to both the host and device code. */
 178         isl_schedule *schedule;
 179         /* Global untiled schedule. */
 180         isl_union_map *sched;
 181 };
 182
 183 enum ppcg_kernel_access_type {
 184         ppcg_access_global,
 185         ppcg_access_shared,
 186         ppcg_access_private
 187 };
 188
 189 enum ppcg_kernel_stmt_type {
 190         ppcg_kernel_copy,
 191         ppcg_kernel_domain,
 192         ppcg_kernel_sync
 193 };
 194
 195 /* Representation of special statements, in particular copy statements
 196  * and __syncthreads statements, inside a kernel.
 197  *
 198  * type represents the kind of statement
 199  *
 200  *
 201  * for ppcg_kernel_copy statements we have
 202  *
 203  * read is set if the statement should copy data from global memory
 204  * to shared memory or registers.
 205  *
 206  * index expresses an access to the array element that needs to be copied
 207  * local_index expresses the corresponding element in the tile
 208  *
 209  * array refers to the original array being copied
 210  * local_array is a pointer to the appropriate element in the "array"
 211  *      array of the ppcg_kernel to which this copy access belongs
 212  *
 213  *
 214  * for ppcg_kernel_domain statements we have
 215  *
 216  * stmt is the corresponding input statement
 217  *
 218  * n_access is the number of accesses in stmt
 219  * access is an array of local information about the accesses
 220  */
 221 struct ppcg_kernel_stmt {
 222         enum ppcg_kernel_stmt_type type;
 223
 224         union {
 225                 struct {
 226                         int read;
 227                         isl_ast_expr *index;
 228                         isl_ast_expr *local_index;
 229                         struct gpu_array_info *array;
 230                         struct gpu_local_array_info *local_array;
 231                 } c;
 232                 struct {
 233                         struct gpu_stmt *stmt;
 234                         isl_id_to_ast_expr *ref2expr;
 235                 } d;
 236         } u;
 237 };
 238
 239 /* Representation of a local variable in a kernel.
 240  */
 241 struct ppcg_kernel_var {
 242         struct gpu_array_info *array;
 243         enum ppcg_kernel_access_type type;
 244         char *name;
 245         isl_vec *size;
 246 };
 247
 248 /* Representation of a kernel.
 249  *
 250  * prog describes the original code from which the kernel is extracted.
 251  *
 252  * id is the sequence number of the kernel.
 253  *
 254  * block_ids contains the list of block identifiers for this kernel.
 255  * thread_ids contains the list of thread identifiers for this kernel.
 256  *
 257  * tile_len is the number of tile dimensions and
 258  * n_parallel is the number of initial parallel loops among those
 259  * tile dimensions.
 260  * tile_size is an array of length tile_len containing the tile sizes.
 261  *
 262  * the first n_grid elements of grid_dim represent the specified size
 263  * of the grid.
 264  * the first n_block elements of block_dim represent the specified or
 265  * effective size of the block.
 266  * Note that in the input file, the sizes of the grid and the blocks
 267  * are specified in the order x, y, z, but internally, the sizes
 268  * are stored in reverse order, so that the last element always
 269  * refers to the x dimension.
 270  *
 271  * grid_size reflects the effective grid size.
 272  *
 273  * context contains the values of the parameters and outer schedule dimensions
 274  * for which any statement instance in this kernel needs to be executed.
 275  *
 276  * n_sync is the number of synchronization operations that have
 277  * been introduced in the schedule tree corresponding to this kernel (so far).
 278  *
 279  * core contains the spaces of the statement domains that form
 280  * the core computation of the kernel.  It is used to navigate
 281  * the tree during the construction of the device part of the schedule
 282  * tree in create_kernel.
 283  *
 284  * arrays is the set of possibly accessed outer array elements.
 285  *
 286  * space is the schedule space of the AST context.  That is, it represents
 287  * the loops of the generated host code containing the kernel launch.
 288  *
 289  * n_array is the total number of arrays in the input program and also
 290  * the number of element in the array array.
 291  * array contains information about each array that is local
 292  * to the current kernel.  If an array is not used in a kernel,
 293  * then the corresponding entry does not contain any information.
 294  *
 295  * any_force_private is set if any array in the kernel is marked force_private
 296  *
 297  * block_filter contains constraints on the domain elements in the kernel
 298  * that encode the mapping to block identifiers, where the block identifiers
 299  * are represented by "n_grid" parameters with as names the elements
 300  * of "block_ids".
 301  *
 302  * thread_filter contains constraints on the domain elements in the kernel
 303  * that encode the mapping to thread identifiers, where the thread identifiers
 304  * are represented by "n_block" parameters with as names the elements
 305  * of "thread_ids".
 306  *
 307  * shared_schedule corresponds to the schedule dimensions of
 308  * the (tiled) schedule for this kernel that have been taken into account
 309  * for computing private/shared memory tiles.
 310  * shared_schedule_dim is the dimension of this schedule.
 311  */
 312 struct ppcg_kernel {
 313         isl_ctx *ctx;
 314         struct ppcg_options *options;
 315
 316         struct gpu_prog *prog;
 317
 318         int id;
 319
 320         isl_id_list *block_ids;
 321         isl_id_list *thread_ids;
 322
 323         int tile_len;
 324         int *tile_size;
 325         int n_parallel;
 326
 327         int n_grid;
 328         int n_block;
 329         int grid_dim[2];
 330         int block_dim[3];
 331
 332         isl_multi_pw_aff *grid_size;
 333         isl_set *context;
 334
 335         int n_sync;
 336         isl_union_set *core;
 337         isl_union_set *arrays;
 338
 339         isl_space *space;
 340
 341         int n_array;
 342         struct gpu_local_array_info *array;
 343
 344         int n_var;
 345         struct ppcg_kernel_var *var;
 346
 347         int any_force_private;
 348
 349         isl_union_set *block_filter;
 350         isl_union_set *thread_filter;
 351         isl_union_pw_multi_aff *shared_schedule;
 352         int shared_schedule_dim;
 353
 354         isl_ast_node *tree;
 355 };
 356
 357 int gpu_array_is_scalar(struct gpu_array_info *array);
 358 int gpu_array_is_read_only_scalar(struct gpu_array_info *array);
 359 __isl_give isl_set *gpu_array_positive_size_guard(struct gpu_array_info *array);
 360
 361 struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct ppcg_scop *scop);
 362 void *gpu_prog_free(struct gpu_prog *prog);
 363
 364 int generate_gpu(isl_ctx *ctx, const char *input, FILE *out,
 365         struct ppcg_options *options,
 366         __isl_give isl_printer *(*print)(__isl_take isl_printer *p,
 367                 struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 368                 struct gpu_types *types, void *user), void *user);
 369
 370 #endif