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         /* Should the array be linearized? */
  47         int linearize;
  48
  49         /* Order dependences on this array.
  50          * Only used if live_range_reordering option is set.
  51          * It is set to NULL otherwise.
  52          */
  53         isl_union_map *dep_order;
  54 };
  55
  56 /* Represents an outer array accessed by a ppcg_kernel, localized
  57  * to the context of this kernel.
  58  *
  59  * "array" points to the corresponding array in the gpu_prog.
  60  * The "n_group" "groups" are the reference groups associated to the array.
  61  * If "force_private" is set, then the array (in practice a scalar)
  62  * must be mapped to a register.
  63  * For each index i with 0 <= i < n_index,
  64  * bound[i] is equal to array->bound[i] specialized to the current kernel.
  65  */
  66 struct gpu_local_array_info {
  67         struct gpu_array_info *array;
  68
  69         int n_group;
  70         struct gpu_array_ref_group **groups;
  71
  72         int force_private;
  73
  74         unsigned n_index;
  75         isl_pw_aff_list *bound;
  76 };
  77
  78 __isl_give isl_ast_expr *gpu_local_array_info_linearize_index(
  79         struct gpu_local_array_info *array, __isl_take isl_ast_expr *expr);
  80
  81 /* A sequence of "n" names of types.
  82  */
  83 struct gpu_types {
  84         int n;
  85         char **name;
  86 };
  87
  88 /* "read" and "write" contain the original access relations, possibly
  89  * involving member accesses.
  90  *
  91  * The elements of "array", as well as the ranges of "copy_in" and "copy_out"
  92  * only refer to the outer arrays of any possible member accesses.
  93  */
  94 struct gpu_prog {
  95         isl_ctx *ctx;
  96
  97         struct ppcg_scop *scop;
  98
  99         /* Set of parameter values */
 100         isl_set *context;
 101
 102         /* All potential read accesses in the entire program */
 103         isl_union_map *read;
 104
 105         /* All potential write accesses in the entire program */
 106         isl_union_map *may_write;
 107         /* All definite write accesses in the entire program */
 108         isl_union_map *must_write;
 109         /* All tagged definite kills in the entire program */
 110         isl_union_map *tagged_must_kill;
 111
 112         /* The set of inner array elements that may be preserved. */
 113         isl_union_set *may_persist;
 114
 115         /* A mapping from all innermost arrays to their outer arrays. */
 116         isl_union_map *to_outer;
 117         /* A mapping from the outer arrays to all corresponding inner arrays. */
 118         isl_union_map *to_inner;
 119         /* A mapping from all intermediate arrays to their outer arrays,
 120          * including an identity mapping from the anoymous 1D space to itself.
 121          */
 122         isl_union_map *any_to_outer;
 123
 124         /* Order dependences on non-scalars. */
 125         isl_union_map *array_order;
 126
 127         /* Array of statements */
 128         int n_stmts;
 129         struct gpu_stmt *stmts;
 130
 131         int n_array;
 132         struct gpu_array_info *array;
 133 };
 134
 135 struct gpu_gen {
 136         isl_ctx *ctx;
 137         struct ppcg_options *options;
 138
 139         /* Callback for printing of AST in appropriate format. */
 140         __isl_give isl_printer *(*print)(__isl_take isl_printer *p,
 141                 struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 142                 struct gpu_types *types, void *user);
 143         void *print_user;
 144
 145         struct gpu_prog *prog;
 146         /* The generated AST. */
 147         isl_ast_node *tree;
 148
 149         /* The sequence of types for which a definition has been printed. */
 150         struct gpu_types types;
 151
 152         /* User specified tile, grid and block sizes for each kernel */
 153         isl_union_map *sizes;
 154
 155         /* Effectively used tile, grid and block sizes for each kernel */
 156         isl_union_map *used_sizes;
 157
 158         /* Identifier of the next kernel. */
 159         int kernel_id;
 160 };
 161
 162 enum ppcg_kernel_access_type {
 163         ppcg_access_global,
 164         ppcg_access_shared,
 165         ppcg_access_private
 166 };
 167
 168 enum ppcg_kernel_stmt_type {
 169         ppcg_kernel_copy,
 170         ppcg_kernel_domain,
 171         ppcg_kernel_sync
 172 };
 173
 174 /* Representation of special statements, in particular copy statements
 175  * and __syncthreads statements, inside a kernel.
 176  *
 177  * type represents the kind of statement
 178  *
 179  *
 180  * for ppcg_kernel_copy statements we have
 181  *
 182  * read is set if the statement should copy data from global memory
 183  * to shared memory or registers.
 184  *
 185  * index expresses an access to the array element that needs to be copied
 186  * local_index expresses the corresponding element in the tile
 187  *
 188  * array refers to the original array being copied
 189  * local_array is a pointer to the appropriate element in the "array"
 190  *      array of the ppcg_kernel to which this copy access belongs
 191  *
 192  *
 193  * for ppcg_kernel_domain statements we have
 194  *
 195  * stmt is the corresponding input statement
 196  *
 197  * n_access is the number of accesses in stmt
 198  * access is an array of local information about the accesses
 199  */
 200 struct ppcg_kernel_stmt {
 201         enum ppcg_kernel_stmt_type type;
 202
 203         union {
 204                 struct {
 205                         int read;
 206                         isl_ast_expr *index;
 207                         isl_ast_expr *local_index;
 208                         struct gpu_array_info *array;
 209                         struct gpu_local_array_info *local_array;
 210                 } c;
 211                 struct {
 212                         struct gpu_stmt *stmt;
 213                         isl_id_to_ast_expr *ref2expr;
 214                 } d;
 215         } u;
 216 };
 217
 218 /* Representation of a local variable in a kernel.
 219  */
 220 struct ppcg_kernel_var {
 221         struct gpu_array_info *array;
 222         enum ppcg_kernel_access_type type;
 223         char *name;
 224         isl_vec *size;
 225 };
 226
 227 /* Representation of a kernel.
 228  *
 229  * prog describes the original code from which the kernel is extracted.
 230  *
 231  * id is the sequence number of the kernel.
 232  *
 233  * block_ids contains the list of block identifiers for this kernel.
 234  * thread_ids contains the list of thread identifiers for this kernel.
 235  *
 236  * the first n_grid elements of grid_dim represent the specified size
 237  * of the grid.
 238  * the first n_block elements of block_dim represent the specified or
 239  * effective size of the block.
 240  * Note that in the input file, the sizes of the grid and the blocks
 241  * are specified in the order x, y, z, but internally, the sizes
 242  * are stored in reverse order, so that the last element always
 243  * refers to the x dimension.
 244  *
 245  * grid_size reflects the effective grid size.
 246  *
 247  * context contains the values of the parameters and outer schedule dimensions
 248  * for which any statement instance in this kernel needs to be executed.
 249  *
 250  * n_sync is the number of synchronization operations that have
 251  * been introduced in the schedule tree corresponding to this kernel (so far).
 252  *
 253  * core contains the spaces of the statement domains that form
 254  * the core computation of the kernel.  It is used to navigate
 255  * the tree during the construction of the device part of the schedule
 256  * tree in create_kernel.
 257  *
 258  * arrays is the set of possibly accessed outer array elements.
 259  *
 260  * space is the schedule space of the AST context.  That is, it represents
 261  * the loops of the generated host code containing the kernel launch.
 262  *
 263  * n_array is the total number of arrays in the input program and also
 264  * the number of element in the array array.
 265  * array contains information about each array that is local
 266  * to the current kernel.  If an array is not used in a kernel,
 267  * then the corresponding entry does not contain any information.
 268  *
 269  * any_force_private is set if any array in the kernel is marked force_private
 270  *
 271  * block_filter contains constraints on the domain elements in the kernel
 272  * that encode the mapping to block identifiers, where the block identifiers
 273  * are represented by "n_grid" parameters with as names the elements
 274  * of "block_ids".
 275  *
 276  * thread_filter contains constraints on the domain elements in the kernel
 277  * that encode the mapping to thread identifiers, where the thread identifiers
 278  * are represented by "n_block" parameters with as names the elements
 279  * of "thread_ids".
 280  *
 281  * shared_schedule corresponds to the schedule dimensions of
 282  * the (tiled) schedule for this kernel that have been taken into account
 283  * for computing private/shared memory tiles.
 284  * shared_schedule_dim is the dimension of this schedule.
 285  *
 286  * sync_writes contains write references that require synchronization.
 287  * Each reference is represented by a universe set in a space [S[i,j] -> R[]]
 288  * with S[i,j] the statement instance space and R[] the array reference.
 289  */
 290 struct ppcg_kernel {
 291         isl_ctx *ctx;
 292         struct ppcg_options *options;
 293
 294         struct gpu_prog *prog;
 295
 296         int id;
 297
 298         isl_id_list *block_ids;
 299         isl_id_list *thread_ids;
 300
 301         int n_grid;
 302         int n_block;
 303         int grid_dim[2];
 304         int block_dim[3];
 305
 306         isl_multi_pw_aff *grid_size;
 307         isl_set *context;
 308
 309         int n_sync;
 310         isl_union_set *core;
 311         isl_union_set *arrays;
 312
 313         isl_space *space;
 314
 315         int n_array;
 316         struct gpu_local_array_info *array;
 317
 318         int n_var;
 319         struct ppcg_kernel_var *var;
 320
 321         int any_force_private;
 322
 323         isl_union_set *block_filter;
 324         isl_union_set *thread_filter;
 325         isl_union_pw_multi_aff *shared_schedule;
 326         int shared_schedule_dim;
 327
 328         isl_union_set *sync_writes;
 329
 330         isl_ast_node *tree;
 331 };
 332
 333 int gpu_array_is_scalar(struct gpu_array_info *array);
 334 int gpu_array_is_read_only_scalar(struct gpu_array_info *array);
 335 __isl_give isl_set *gpu_array_positive_size_guard(struct gpu_array_info *array);
 336
 337 struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct ppcg_scop *scop);
 338 void *gpu_prog_free(struct gpu_prog *prog);
 339
 340 int generate_gpu(isl_ctx *ctx, const char *input, FILE *out,
 341         struct ppcg_options *options,
 342         __isl_give isl_printer *(*print)(__isl_take isl_printer *p,
 343                 struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 344                 struct gpu_types *types, void *user), void *user);
 345
 346 #endif