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         /* Is the array local and should it be declared on the host? */
  44         int declare_local;
  45
  46         /* Is the corresponding global device memory accessed in any way? */
  47         int global;
  48
  49         /* Should the array be linearized? */
  50         int linearize;
  51
  52         /* Order dependences on this array.
  53          * Only used if live_range_reordering option is set.
  54          * It is set to NULL otherwise.
  55          */
  56         isl_union_map *dep_order;
  57 };
  58
  59 /* Represents an outer array accessed by a ppcg_kernel, localized
  60  * to the context of this kernel.
  61  *
  62  * "array" points to the corresponding array in the gpu_prog.
  63  * The "n_group" "groups" are the reference groups associated to the array.
  64  * If "force_private" is set, then the array (in practice a scalar)
  65  * must be mapped to a register.
  66  * "global" is set if the global device memory corresponding
  67  * to this array is accessed by the kernel.
  68  * For each index i with 0 <= i < n_index,
  69  * bound[i] is equal to array->bound[i] specialized to the current kernel.
  70  */
  71 struct gpu_local_array_info {
  72         struct gpu_array_info *array;
  73
  74         int n_group;
  75         struct gpu_array_ref_group **groups;
  76
  77         int force_private;
  78         int global;
  79
  80         unsigned n_index;
  81         isl_pw_aff_list *bound;
  82 };
  83
  84 __isl_give isl_ast_expr *gpu_local_array_info_linearize_index(
  85         struct gpu_local_array_info *array, __isl_take isl_ast_expr *expr);
  86
  87 /* A sequence of "n" names of types.
  88  */
  89 struct gpu_types {
  90         int n;
  91         char **name;
  92 };
  93
  94 /* "read" and "write" contain the original access relations, possibly
  95  * involving member accesses.
  96  *
  97  * The elements of "array", as well as the ranges of "copy_in" and "copy_out"
  98  * only refer to the outer arrays of any possible member accesses.
  99  */
 100 struct gpu_prog {
 101         isl_ctx *ctx;
 102
 103         struct ppcg_scop *scop;
 104
 105         /* Set of parameter values */
 106         isl_set *context;
 107
 108         /* All potential read accesses in the entire program */
 109         isl_union_map *read;
 110
 111         /* All potential write accesses in the entire program */
 112         isl_union_map *may_write;
 113         /* All definite write accesses in the entire program */
 114         isl_union_map *must_write;
 115         /* All tagged definite kills in the entire program */
 116         isl_union_map *tagged_must_kill;
 117
 118         /* The set of inner array elements that may be preserved. */
 119         isl_union_set *may_persist;
 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 };
 167
 168 enum ppcg_kernel_access_type {
 169         ppcg_access_global,
 170         ppcg_access_shared,
 171         ppcg_access_private
 172 };
 173
 174 enum ppcg_kernel_stmt_type {
 175         ppcg_kernel_copy,
 176         ppcg_kernel_domain,
 177         ppcg_kernel_sync
 178 };
 179
 180 /* Representation of special statements, in particular copy statements
 181  * and __syncthreads statements, inside a kernel.
 182  *
 183  * type represents the kind of statement
 184  *
 185  *
 186  * for ppcg_kernel_copy statements we have
 187  *
 188  * read is set if the statement should copy data from global memory
 189  * to shared memory or registers.
 190  *
 191  * index expresses an access to the array element that needs to be copied
 192  * local_index expresses the corresponding element in the tile
 193  *
 194  * array refers to the original array being copied
 195  * local_array is a pointer to the appropriate element in the "array"
 196  *      array of the ppcg_kernel to which this copy access belongs
 197  *
 198  *
 199  * for ppcg_kernel_domain statements we have
 200  *
 201  * stmt is the corresponding input statement
 202  *
 203  * n_access is the number of accesses in stmt
 204  * access is an array of local information about the accesses
 205  */
 206 struct ppcg_kernel_stmt {
 207         enum ppcg_kernel_stmt_type type;
 208
 209         union {
 210                 struct {
 211                         int read;
 212                         isl_ast_expr *index;
 213                         isl_ast_expr *local_index;
 214                         struct gpu_array_info *array;
 215                         struct gpu_local_array_info *local_array;
 216                 } c;
 217                 struct {
 218                         struct gpu_stmt *stmt;
 219                         isl_id_to_ast_expr *ref2expr;
 220                 } d;
 221         } u;
 222 };
 223
 224 /* Representation of a local variable in a kernel.
 225  */
 226 struct ppcg_kernel_var {
 227         struct gpu_array_info *array;
 228         enum ppcg_kernel_access_type type;
 229         char *name;
 230         isl_vec *size;
 231 };
 232
 233 /* Representation of a kernel.
 234  *
 235  * prog describes the original code from which the kernel is extracted.
 236  *
 237  * id is the sequence number of the kernel.
 238  *
 239  * block_ids contains the list of block identifiers for this kernel.
 240  * thread_ids contains the list of thread identifiers for this kernel.
 241  *
 242  * the first n_grid elements of grid_dim represent the specified size
 243  * of the grid.
 244  * the first n_block elements of block_dim represent the specified or
 245  * effective size of the block.
 246  * Note that in the input file, the sizes of the grid and the blocks
 247  * are specified in the order x, y, z, but internally, the sizes
 248  * are stored in reverse order, so that the last element always
 249  * refers to the x dimension.
 250  *
 251  * grid_size reflects the effective grid size.
 252  *
 253  * context contains the values of the parameters and outer schedule dimensions
 254  * for which any statement instance in this kernel needs to be executed.
 255  *
 256  * n_sync is the number of synchronization operations that have
 257  * been introduced in the schedule tree corresponding to this kernel (so far).
 258  *
 259  * core contains the spaces of the statement domains that form
 260  * the core computation of the kernel.  It is used to navigate
 261  * the tree during the construction of the device part of the schedule
 262  * tree in create_kernel.
 263  *
 264  * arrays is the set of possibly accessed outer array elements.
 265  *
 266  * space is the schedule space of the AST context.  That is, it represents
 267  * the loops of the generated host code containing the kernel launch.
 268  *
 269  * n_array is the total number of arrays in the input program and also
 270  * the number of element in the array array.
 271  * array contains information about each array that is local
 272  * to the current kernel.  If an array is not used in a kernel,
 273  * then the corresponding entry does not contain any information.
 274  *
 275  * any_force_private is set if any array in the kernel is marked force_private
 276  *
 277  * block_filter contains constraints on the domain elements in the kernel
 278  * that encode the mapping to block identifiers, where the block identifiers
 279  * are represented by "n_grid" parameters with as names the elements
 280  * of "block_ids".
 281  *
 282  * thread_filter contains constraints on the domain elements in the kernel
 283  * that encode the mapping to thread identifiers, where the thread identifiers
 284  * are represented by "n_block" parameters with as names the elements
 285  * of "thread_ids".
 286  *
 287  * shared_schedule corresponds to the schedule dimensions of
 288  * the (tiled) schedule for this kernel that have been taken into account
 289  * for computing private/shared memory tiles.
 290  * shared_schedule_dim is the dimension of this schedule.
 291  *
 292  * sync_writes contains write references that require synchronization.
 293  * Each reference is represented by a universe set in a space [S[i,j] -> R[]]
 294  * with S[i,j] the statement instance space and R[] the array reference.
 295  */
 296 struct ppcg_kernel {
 297         isl_ctx *ctx;
 298         struct ppcg_options *options;
 299
 300         struct gpu_prog *prog;
 301
 302         int id;
 303
 304         isl_id_list *block_ids;
 305         isl_id_list *thread_ids;
 306
 307         int n_grid;
 308         int n_block;
 309         int grid_dim[2];
 310         int block_dim[3];
 311
 312         isl_multi_pw_aff *grid_size;
 313         isl_set *context;
 314
 315         int n_sync;
 316         isl_union_set *core;
 317         isl_union_set *arrays;
 318
 319         isl_space *space;
 320
 321         int n_array;
 322         struct gpu_local_array_info *array;
 323
 324         int n_var;
 325         struct ppcg_kernel_var *var;
 326
 327         int any_force_private;
 328
 329         isl_union_set *block_filter;
 330         isl_union_set *thread_filter;
 331         isl_union_pw_multi_aff *shared_schedule;
 332         int shared_schedule_dim;
 333
 334         isl_union_set *sync_writes;
 335
 336         isl_ast_node *tree;
 337 };
 338
 339 int gpu_array_is_scalar(struct gpu_array_info *array);
 340 int gpu_array_is_read_only_scalar(struct gpu_array_info *array);
 341 int gpu_array_requires_device_allocation(struct gpu_array_info *array);
 342 __isl_give isl_set *gpu_array_positive_size_guard(struct gpu_array_info *array);
 343
 344 struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct ppcg_scop *scop);
 345 void *gpu_prog_free(struct gpu_prog *prog);
 346
 347 int ppcg_kernel_requires_array_argument(struct ppcg_kernel *kernel, int i);
 348
 349 int generate_gpu(isl_ctx *ctx, const char *input, FILE *out,
 350         struct ppcg_options *options,
 351         __isl_give isl_printer *(*print)(__isl_take isl_printer *p,
 352                 struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 353                 struct gpu_types *types, void *user), void *user);
 354
 355 #endif