• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 	/* Is a single, fixed element being accessed? */
27 	isl_bool fixed_element;
28 	/* The number of index expressions specified in the access. */
29 	int n_index;
30 
31 	/* May access relation */
32 	isl_map *access;
33 	/* May access relation with as domain a mapping from iteration domain
34 	 * to a reference identifier.
35 	 */
36 	isl_map *tagged_access;
37 	/* The reference id of the corresponding pet_expr. */
38 	isl_id *ref_id;
39 
40 	struct gpu_stmt_access *next;
41 };
42 
43 /* A representation of a user statement.
44  * "stmt" points to the corresponding pet statement.
45  * "id" is the identifier of the instance set of the statement.
46  * "accesses" is a linked list of accesses performed by the statement.
47  * If the statement has been killed, i.e., if it will not be scheduled,
48  * then this linked list may be empty even if the actual statement does
49  * perform accesses.
50  */
51 struct gpu_stmt {
52 	isl_id *id;
53 	struct pet_stmt *stmt;
54 
55 	struct gpu_stmt_access *accesses;
56 };
57 
58 /* Represents an outer array possibly accessed by a gpu_prog.
59  */
60 struct gpu_array_info {
61 	/* The array data space. */
62 	isl_space *space;
63 	/* Element type. */
64 	char *type;
65 	/* Element size. */
66 	int size;
67 	/* Name of the array. */
68 	char *name;
69 	/* Declared extent of original array. */
70 	isl_set *declared_extent;
71 	/* AST expression for declared size of original array. */
72 	isl_ast_expr *declared_size;
73 	/* Extent of the array that needs to be copied. */
74 	isl_set *extent;
75 	/* Number of indices. */
76 	unsigned n_index;
77 	/* For each index, a bound on "extent" in that direction. */
78 	isl_multi_pw_aff *bound;
79 	/* The corresponding access AST expression, if the array needs
80 	 * to be allocated on the device.
81 	 */
82 	isl_ast_expr *bound_expr;
83 
84 	/* All references to this array; point to elements of a linked list. */
85 	int n_ref;
86 	struct gpu_stmt_access **refs;
87 
88 	/* Is this array accessed at all by the program? */
89 	int accessed;
90 
91 	/* Is this a scalar that is read-only within the entire program? */
92 	int read_only_scalar;
93 
94 	/* Are the elements of the array structures? */
95 	int has_compound_element;
96 
97 	/* Are the elements only accessed through constant index expressions? */
98 	int only_fixed_element;
99 
100 	/* Is the array local to the scop? */
101 	int local;
102 	/* Is the array local and should it be declared on the host? */
103 	int declare_local;
104 
105 	/* Is the corresponding global device memory accessed in any way? */
106 	int global;
107 
108 	/* Should the array be linearized? */
109 	int linearize;
110 
111 	/* Order dependences on this array.
112 	 * Only used if live_range_reordering option is set.
113 	 * It is set to NULL otherwise.
114 	 */
115 	isl_union_map *dep_order;
116 
117     void *user;
118 };
119 
120 /* Represents an outer array accessed by a ppcg_kernel, localized
121  * to the context of this kernel.
122  *
123  * "array" points to the corresponding array in the gpu_prog.
124  * The "n_group" "groups" are the reference groups associated to the array.
125  * If "force_private" is set, then the array (in practice a scalar)
126  * must be mapped to a register.
127  * "global" is set if the global device memory corresponding
128  * to this array is accessed by the kernel.
129  * "bound" is equal to array->bound specialized to the current kernel.
130  * "bound_expr" is the corresponding access AST expression.
131  */
132 struct gpu_local_array_info {
133 	struct gpu_array_info *array;
134 
135 	int n_group;
136 	struct gpu_array_ref_group **groups;
137 
138 	int force_private;
139 	int global;
140 
141 	unsigned n_index;
142 	isl_multi_pw_aff *bound;
143 	isl_ast_expr *bound_expr;
144 };
145 
146 __isl_give isl_ast_expr *gpu_local_array_info_linearize_index(
147 	struct gpu_local_array_info *array, __isl_take isl_ast_expr *expr);
148 
149 /* A sequence of "n" names of types.
150  */
151 struct gpu_types {
152 	int n;
153 	char **name;
154 };
155 
156 /* "read" and "write" contain the original access relations, possibly
157  * involving member accesses.
158  *
159  * The elements of "array", as well as the ranges of "copy_in" and "copy_out"
160  * only refer to the outer arrays of any possible member accesses.
161  */
162 struct gpu_prog {
163 	isl_ctx *ctx;
164 
165 	struct ppcg_scop *scop;
166 
167 	/* Set of parameter values */
168 	isl_set *context;
169 
170 	/* All potential read accesses in the entire program */
171 	isl_union_map *read;
172 
173 	/* All potential write accesses in the entire program */
174 	isl_union_map *may_write;
175 	/* All definite write accesses in the entire program */
176 	isl_union_map *must_write;
177 	/* All tagged definite kills in the entire program */
178 	isl_union_map *tagged_must_kill;
179 
180 	/* The set of inner array elements that may be preserved. */
181 	isl_union_set *may_persist;
182 
183 	/* A mapping from all innermost arrays to their outer arrays. */
184 	isl_union_map *to_outer;
185 	/* A mapping from the outer arrays to all corresponding inner arrays. */
186 	isl_union_map *to_inner;
187 	/* A mapping from all intermediate arrays to their outer arrays,
188 	 * including an identity mapping from the anonymous 1D space to itself.
189 	 */
190 	isl_union_map *any_to_outer;
191 
192 	/* Order dependences on non-scalars. */
193 	isl_union_map *array_order;
194 
195 	/* Array of statements */
196 	int n_stmts;
197 	struct gpu_stmt *stmts;
198 
199 	int n_array;
200 	struct gpu_array_info *array;
201 };
202 
203 struct gpu_gen {
204 	isl_ctx *ctx;
205 	struct ppcg_options *options;
206 
207 	/* Callback for printing of AST in appropriate format. */
208 	__isl_give isl_printer *(*print)(__isl_take isl_printer *p,
209 		struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
210 		struct gpu_types *types, void *user);
211 	void *print_user;
212 
213     isl_id_to_ast_expr *(*build_ast_expr)(void *stmt,
214             isl_ast_build *build,
215             isl_multi_pw_aff *(*fn_index)(
216                 __isl_take isl_multi_pw_aff *mpa, isl_id *id,
217                 void *user),
218             void *user_index,
219             isl_ast_expr *(*fn_expr)(isl_ast_expr *expr,
220                 isl_id *id, void *user),
221         void *user_expr);
222 
223 	struct gpu_prog *prog;
224 	/* The generated AST. */
225 	isl_ast_node *tree;
226 
227 	/* The sequence of types for which a definition has been printed. */
228 	struct gpu_types types;
229 
230 	/* User specified tile, grid and block sizes for each kernel */
231 	isl_union_map *sizes;
232 
233 	/* Effectively used tile, grid and block sizes for each kernel */
234 	isl_union_map *used_sizes;
235 
236 	/* Identifier of the next kernel. */
237 	int kernel_id;
238 };
239 
240 enum ppcg_group_access_type {
241 	ppcg_access_global,
242 	ppcg_access_shared,
243 	ppcg_access_private
244 };
245 
246 enum ppcg_kernel_stmt_type {
247 	ppcg_kernel_copy,
248 	ppcg_kernel_domain,
249 	ppcg_kernel_sync
250 };
251 
252 /* Representation of special statements, in particular copy statements
253  * and __syncthreads statements, inside a kernel.
254  *
255  * type represents the kind of statement
256  *
257  *
258  * for ppcg_kernel_copy statements we have
259  *
260  * read is set if the statement should copy data from global memory
261  * to shared memory or registers.
262  *
263  * index expresses an access to the array element that needs to be copied
264  * local_index expresses the corresponding element in the tile
265  *
266  * array refers to the original array being copied
267  * local_array is a pointer to the appropriate element in the "array"
268  *	array of the ppcg_kernel to which this copy access belongs
269  *
270  *
271  * for ppcg_kernel_domain statements we have
272  *
273  * stmt is the corresponding input statement
274  *
275  * n_access is the number of accesses in stmt
276  * access is an array of local information about the accesses
277  */
278 struct ppcg_kernel_stmt {
279 	enum ppcg_kernel_stmt_type type;
280 
281 	union {
282 		struct {
283 			int read;
284 			isl_ast_expr *index;
285 			isl_ast_expr *local_index;
286 			struct gpu_array_info *array;
287 			struct gpu_local_array_info *local_array;
288 		} c;
289 		struct {
290 			struct gpu_stmt *stmt;
291 			isl_id_to_ast_expr *ref2expr;
292 		} d;
293 	} u;
294 };
295 
296 /* Representation of a local variable in a kernel.
297  */
298 struct ppcg_kernel_var {
299 	struct gpu_array_info *array;
300 	enum ppcg_group_access_type type;
301 	char *name;
302 	isl_vec *size;
303 };
304 
305 /* Representation of a kernel.
306  *
307  * prog describes the original code from which the kernel is extracted.
308  *
309  * id is the sequence number of the kernel.
310  *
311  * block_ids contains the list of block identifiers for this kernel.
312  * thread_ids contains the list of thread identifiers for this kernel.
313  *
314  * the first n_grid elements of grid_dim represent the specified size
315  * of the grid.
316  * the first n_block elements of block_dim represent the specified or
317  * effective size of the block.
318  * Note that in the input file, the sizes of the grid and the blocks
319  * are specified in the order x, y, z, but internally, the sizes
320  * are stored in reverse order, so that the last element always
321  * refers to the x dimension.
322  *
323  * grid_size reflects the effective grid size.
324  * grid_size_expr contains a corresponding access AST expression, built within
325  * the context where the launch appears.
326  *
327  * context contains the values of the parameters and outer schedule dimensions
328  * for which any statement instance in this kernel needs to be executed.
329  *
330  * n_sync is the number of synchronization operations that have
331  * been introduced in the schedule tree corresponding to this kernel (so far).
332  *
333  * core contains the spaces of the statement domains that form
334  * the core computation of the kernel.  It is used to navigate
335  * the tree during the construction of the device part of the schedule
336  * tree in gpu_create_kernel.
337  *
338  * expanded_domain contains the original statement instances,
339  * i.e., those that appear in the domains of access relations,
340  * that are involved in the kernel.
341  * contraction maps those original statement instances to
342  * the statement instances that are active at the point
343  * in the schedule tree where the kernel is created.
344  *
345  * arrays is the set of possibly accessed outer array elements.
346  *
347  * space is the schedule space of the AST context.  That is, it represents
348  * the loops of the generated host code containing the kernel launch.
349  *
350  * n_array is the total number of arrays in the input program and also
351  * the number of element in the array array.
352  * array contains information about each array that is local
353  * to the current kernel.  If an array is not used in a kernel,
354  * then the corresponding entry does not contain any information.
355  *
356  * any_force_private is set if any array in the kernel is marked force_private
357  *
358  * block_filter contains constraints on the domain elements in the kernel
359  * that encode the mapping to block identifiers, where the block identifiers
360  * are represented by "n_grid" parameters with as names the elements
361  * of "block_ids".
362  *
363  * thread_filter contains constraints on the domain elements in the kernel
364  * that encode the mapping to thread identifiers, where the thread identifiers
365  * are represented by "n_block" parameters with as names the elements
366  * of "thread_ids".
367  *
368  * copy_schedule corresponds to the schedule dimensions of
369  * the (tiled) schedule for this kernel that have been taken into account
370  * for computing private/shared memory tiles.
371  * The domain corresponds to the original statement instances, i.e.,
372  * those that appear in the leaves of the schedule tree.
373  * copy_schedule_dim is the dimension of this schedule.
374  *
375  * sync_writes contains write references that require synchronization.
376  * Each reference is represented by a universe set in a space [S[i,j] -> R[]]
377  * with S[i,j] the statement instance space and R[] the array reference.
378  */
379 struct ppcg_kernel {
380 	isl_ctx *ctx;
381 	struct ppcg_options *options;
382 
383 	struct gpu_prog *prog;
384 
385 	int id;
386 
387 	isl_id_list *block_ids;
388 	isl_id_list *thread_ids;
389 
390 	int n_grid;
391 	int n_block;
392 	int grid_dim[2];
393 	int block_dim[3];
394 
395 	isl_multi_pw_aff *grid_size;
396 	isl_ast_expr *grid_size_expr;
397 	isl_set *context;
398 
399 	int n_sync;
400 	isl_union_set *core;
401 	isl_union_set *arrays;
402 
403 	isl_union_pw_multi_aff *contraction;
404 	isl_union_set *expanded_domain;
405 
406 	isl_space *space;
407 
408 	int n_array;
409 	struct gpu_local_array_info *array;
410 
411 	int n_var;
412 	struct ppcg_kernel_var *var;
413 
414 	int any_force_private;
415 
416 	isl_union_set *block_filter;
417 	isl_union_set *thread_filter;
418 	isl_union_pw_multi_aff *copy_schedule;
419 	int copy_schedule_dim;
420 
421 	isl_union_set *sync_writes;
422 
423 	isl_ast_node *tree;
424 };
425 
426 int gpu_array_is_scalar(struct gpu_array_info *array);
427 int gpu_array_is_read_only_scalar(struct gpu_array_info *array);
428 int gpu_array_requires_device_allocation(struct gpu_array_info *array);
429 __isl_give isl_set *gpu_array_positive_size_guard(struct gpu_array_info *array);
430 isl_bool gpu_array_can_be_private(struct gpu_array_info *array);
431 
432 struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct ppcg_scop *scop);
433 void *gpu_prog_free(struct gpu_prog *prog);
434 
435 int ppcg_kernel_requires_array_argument(struct ppcg_kernel *kernel, int i);
436 
437 int generate_gpu(isl_ctx *ctx, const char *input, FILE *out,
438 	struct ppcg_options *options,
439 	__isl_give isl_printer *(*print)(__isl_take isl_printer *p,
440 		struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
441 		struct gpu_types *types, void *user), void *user);
442 
443 __isl_give isl_schedule_node *gpu_create_kernel(struct gpu_gen *gen,
444 	__isl_take isl_schedule_node *node, int scale,
445 	__isl_keep isl_multi_val *sizes);
446 
447 __isl_give isl_schedule *get_schedule(struct gpu_gen *gen);
448 int has_any_permutable_node(__isl_keep isl_schedule *schedule);
449 __isl_give isl_schedule *map_to_device(struct gpu_gen *gen,
450                                        __isl_take isl_schedule *schedule,
451                                       int to_from_device);
452 __isl_give isl_ast_node *generate_code(struct gpu_gen *gen,
453                                        __isl_take isl_schedule *schedule);
454 
455 __isl_give isl_union_set *compute_may_persist(struct gpu_prog *prog);
456 void collect_references(struct gpu_prog *prog, struct gpu_array_info *array);
457 void collect_order_dependences(struct gpu_prog *prog);
458 isl_bool only_fixed_element_accessed(struct gpu_array_info *array);
459 #endif
460