2 * Copyright 2010-2011 INRIA Saclay
3 * Copyright 2012-2013 Ecole Normale Superieure
5 * Use of this software is governed by the MIT license
7 * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
8 * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
10 * and Ecole Normale Superieure, 45 rue d’Ulm, 75230 Paris, France
17 #include <isl/polynomial.h>
18 #include <isl/union_set.h>
22 #include <isl/schedule.h>
23 #include <isl/schedule_node.h>
24 #include <isl/options.h>
25 #include <isl/ast_build.h>
29 #include "gpu_array_tile.h"
30 #include "gpu_group.h"
33 #include "ppcg_options.h"
36 struct gpu_array_info
;
38 /* Collect all references to the given array and store pointers to them
41 * If the array contains structures, then there is no need to collect
42 * the references since we will not be computing any reference groups.
44 static void collect_references(struct gpu_prog
*prog
,
45 struct gpu_array_info
*array
)
50 if (array
->has_compound_element
)
54 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
55 struct gpu_stmt
*stmt
= &prog
->stmts
[i
];
56 struct gpu_stmt_access
*access
;
58 for (access
= stmt
->accesses
; access
; access
= access
->next
) {
60 name
= isl_map_get_tuple_name(access
->access
,
62 if (name
&& !strcmp(array
->name
, name
))
68 array
->refs
= isl_alloc_array(prog
->ctx
, struct gpu_stmt_access
*, n
);
72 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
73 struct gpu_stmt
*stmt
= &prog
->stmts
[i
];
74 struct gpu_stmt_access
*access
;
76 for (access
= stmt
->accesses
; access
; access
= access
->next
) {
78 name
= isl_map_get_tuple_name(access
->access
,
80 if (!name
|| strcmp(array
->name
, name
))
83 array
->refs
[n
++] = access
;
88 /* Compute and return the extent of "array", taking into account the set of
91 * In particular, the extent in the outer dimension is taken
92 * from "accessed", while the extents in the remaining dimensions
93 * are taken from array->extent.
95 * The extent in the outer dimension cannot be taken from array->extent
96 * because that may be unbounded. Furthermore, even if it is bounded,
97 * it may be larger than the piece of the array that is being accessed.
99 static __isl_give isl_set
*compute_extent(struct pet_array
*array
,
100 __isl_keep isl_set
*accessed
)
107 extent
= isl_set_copy(array
->extent
);
109 n_index
= isl_set_dim(accessed
, isl_dim_set
);
113 extent
= isl_set_project_out(extent
, isl_dim_set
, 0, 1);
114 outer
= isl_set_copy(accessed
);
115 outer
= isl_set_project_out(outer
, isl_dim_set
, 1, n_index
- 1);
116 extent
= isl_set_flat_product(outer
, extent
);
117 id
= isl_set_get_tuple_id(accessed
);
118 extent
= isl_set_set_tuple_id(extent
, id
);
123 /* Is the array "array" being extracted a read-only scalar?
125 * That is, is "array" a scalar that is never possibly written to.
126 * An array containing structures is never considered to be a scalar.
128 static int is_read_only_scalar(struct gpu_array_info
*array
,
129 struct gpu_prog
*prog
)
132 isl_union_map
*write
;
135 if (array
->has_compound_element
)
137 if (array
->n_index
!= 0)
140 write
= isl_union_map_copy(prog
->may_write
);
141 space
= isl_set_universe(isl_space_copy(array
->space
));
142 write
= isl_union_map_intersect_range(write
,
143 isl_union_set_from_set(space
));
144 empty
= isl_union_map_is_empty(write
);
145 isl_union_map_free(write
);
150 /* Compute bounds on the host array "pa" based on the corresponding
151 * accessed elements in "arrays"
152 * and collect all references to the array.
153 * Store the results in "info".
155 * If the array is zero-dimensional and does not contain structures,
156 * i.e., if the array is a scalar, we check whether it is read-only.
157 * We also check whether the array is accessed at all.
159 static int extract_array_info(struct gpu_prog
*prog
,
160 struct gpu_array_info
*info
, struct pet_array
*pa
,
161 __isl_keep isl_union_set
*arrays
)
167 isl_set
*accessed
, *extent
;
169 n_index
= isl_set_dim(pa
->extent
, isl_dim_set
);
170 name
= isl_set_get_tuple_name(pa
->extent
);
171 bounds
= isl_alloc_array(prog
->ctx
, isl_pw_aff
*, n_index
);
175 info
->space
= isl_set_get_space(pa
->extent
);
176 info
->name
= strdup(name
);
177 info
->n_index
= n_index
;
178 info
->bound
= bounds
;
179 info
->linearize
= prog
->scop
->options
->linearize_device_arrays
;
181 info
->type
= strdup(pa
->element_type
);
182 info
->size
= pa
->element_size
;
183 info
->local
= pa
->declared
&& !pa
->exposed
;
184 info
->has_compound_element
= pa
->element_is_record
;
185 info
->read_only_scalar
= is_read_only_scalar(info
, prog
);
187 accessed
= isl_union_set_extract_set(arrays
,
188 isl_space_copy(info
->space
));
189 empty
= isl_set_is_empty(accessed
);
190 extent
= compute_extent(pa
, accessed
);
191 isl_set_free(accessed
);
192 info
->extent
= extent
;
195 info
->accessed
= !empty
;
196 for (i
= 0; i
< n_index
; ++i
) {
202 dom
= isl_set_copy(extent
);
203 dom
= isl_set_project_out(dom
, isl_dim_set
, i
+ 1,
205 dom
= isl_set_project_out(dom
, isl_dim_set
, 0, i
);
206 if (!isl_set_dim_has_upper_bound(dom
, isl_dim_set
, 0)) {
207 fprintf(stderr
, "unable to determine extent of '%s' "
208 "in dimension %d\n", info
->name
, i
);
209 dom
= isl_set_free(dom
);
211 bound
= isl_set_dim_max(dom
, 0);
212 dom
= isl_pw_aff_domain(isl_pw_aff_copy(bound
));
213 ls
= isl_local_space_from_space(isl_set_get_space(dom
));
214 one
= isl_aff_zero_on_domain(ls
);
215 one
= isl_aff_add_constant_si(one
, 1);
216 bound
= isl_pw_aff_add(bound
, isl_pw_aff_alloc(dom
, one
));
217 bound
= isl_pw_aff_gist(bound
, isl_set_copy(prog
->context
));
220 if (!isl_pw_aff_is_cst(bound
))
224 collect_references(prog
, info
);
229 /* Remove independence from the order constraints "order" on array "array".
230 * Since the pairs of iterations in the filter relation of an independence
231 * are guaranteed to be completely independent by the user, there is
232 * no need to ensure that live ranges are ordered along thong pairs.
233 * We make an exception for local variables, though, as the independence
234 * guarantee does not apply to those.
236 * The order constraints are used in two places.
237 * Those on scalars are used in check_scalar_live_ranges to check if
238 * we need to force the scalar to be private. Any non-local scalar
239 * should not be forced scalar if it only appears in independent loops.
240 * Those on non-scalars are added to the coincidence constraints
241 * in compute_schedule because we do not support any array expansion.
242 * Accesses to non-local arrays should not prevent a loop from being
243 * considered coincident so we should indeed remove those constraints
244 * from the order constraints.
246 static __isl_give isl_union_map
*remove_independences(struct gpu_prog
*prog
,
247 struct gpu_array_info
*array
, __isl_take isl_union_map
*order
)
251 for (i
= 0; i
< prog
->scop
->pet
->n_independence
; ++i
) {
252 struct pet_independence
*pi
= prog
->scop
->pet
->independences
[i
];
253 if (isl_union_set_contains(pi
->local
, array
->space
))
256 order
= isl_union_map_subtract(order
,
257 isl_union_map_copy(pi
->filter
));
263 /* For each array in "prog", store the (untagged) order dependences
264 * derived from the array in array->dep_order.
265 * In particular, consider all references that access the given array
266 * and take the order dependences that have one of these references
267 * as source. (Since an order dependence relates two references to
268 * the same array, the target of these order dependences will also
269 * be one of these references.)
270 * Additionally, store the union of these array->dep_order relations
271 * for all non-scalar arrays in prog->array_order.
273 void collect_order_dependences(struct gpu_prog
*prog
)
277 isl_union_map
*accesses
;
279 space
= isl_union_map_get_space(prog
->read
);
280 prog
->array_order
= isl_union_map_empty(space
);
282 accesses
= isl_union_map_copy(prog
->scop
->tagged_reads
);
283 accesses
= isl_union_map_union(accesses
,
284 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
285 accesses
= isl_union_map_universe(accesses
);
286 accesses
= isl_union_map_apply_range(accesses
,
287 isl_union_map_copy(prog
->to_outer
));
289 for (i
= 0; i
< prog
->n_array
; ++i
) {
290 struct gpu_array_info
*array
= &prog
->array
[i
];
293 isl_union_map
*order
;
295 set
= isl_set_universe(isl_space_copy(array
->space
));
296 uset
= isl_union_set_from_set(set
);
297 uset
= isl_union_map_domain(
298 isl_union_map_intersect_range(isl_union_map_copy(accesses
),
300 order
= isl_union_map_copy(prog
->scop
->tagged_dep_order
);
301 order
= isl_union_map_intersect_domain(order
, uset
);
302 order
= isl_union_map_zip(order
);
303 order
= isl_union_set_unwrap(isl_union_map_domain(order
));
304 order
= remove_independences(prog
, array
, order
);
305 array
->dep_order
= order
;
307 if (gpu_array_is_scalar(array
) && !array
->has_compound_element
)
310 prog
->array_order
= isl_union_map_union(prog
->array_order
,
311 isl_union_map_copy(array
->dep_order
));
314 isl_union_map_free(accesses
);
317 /* Construct a gpu_array_info for each array referenced by prog->scop and
318 * collect them in prog->array.
320 * The sizes are based on the extents and the set of possibly accessed
321 * elements by "prog".
322 * If there are any member accesses involved, then they are first mapped
323 * to the outer arrays of structs.
325 * If we are allowing live range reordering, then also set
326 * the dep_order field. Otherwise leave it NULL.
328 static int collect_array_info(struct gpu_prog
*prog
)
332 isl_union_set
*arrays
;
334 arrays
= isl_union_map_range(isl_union_map_copy(prog
->read
));
335 arrays
= isl_union_set_union(arrays
,
336 isl_union_map_range(isl_union_map_copy(prog
->may_write
)));
338 arrays
= isl_union_set_apply(arrays
,
339 isl_union_map_copy(prog
->to_outer
));
341 arrays
= isl_union_set_coalesce(arrays
);
343 prog
->n_array
= prog
->scop
->pet
->n_array
;
344 prog
->array
= isl_calloc_array(prog
->ctx
,
345 struct gpu_array_info
, prog
->n_array
);
347 for (i
= 0; i
< prog
->scop
->pet
->n_array
; ++i
)
348 if (extract_array_info(prog
, &prog
->array
[i
],
349 prog
->scop
->pet
->arrays
[i
], arrays
) < 0)
352 isl_union_set_free(arrays
);
354 if (prog
->scop
->options
->live_range_reordering
)
355 collect_order_dependences(prog
);
360 static void free_array_info(struct gpu_prog
*prog
)
364 for (i
= 0; i
< prog
->n_array
; ++i
) {
365 int n_index
= prog
->array
[i
].n_index
;
366 free(prog
->array
[i
].type
);
367 free(prog
->array
[i
].name
);
368 for (j
= 0; j
< n_index
; ++j
)
369 isl_pw_aff_free(prog
->array
[i
].bound
[j
]);
370 isl_space_free(prog
->array
[i
].space
);
371 isl_set_free(prog
->array
[i
].extent
);
372 free(prog
->array
[i
].bound
);
373 free(prog
->array
[i
].refs
);
374 isl_union_map_free(prog
->array
[i
].dep_order
);
379 /* Check if a gpu array is a scalar. A scalar is a value that is not stored
380 * as an array or through a pointer reference, but as a single data element.
381 * At the moment, scalars are represented as zero-dimensional arrays.
382 * Note that the single data element may be an entire structure.
384 int gpu_array_is_scalar(struct gpu_array_info
*array
)
386 return array
->n_index
== 0;
389 /* Is "array" a read-only scalar?
391 int gpu_array_is_read_only_scalar(struct gpu_array_info
*array
)
393 return array
->read_only_scalar
;
396 /* Return the set of parameter values for which the array has a positive
397 * size in all dimensions.
398 * If the sizes are only valid for some parameter values, then those
399 * constraints are also taken into account.
401 __isl_give isl_set
*gpu_array_positive_size_guard(struct gpu_array_info
*array
)
407 space
= isl_space_params(isl_space_copy(array
->space
));
408 guard
= isl_set_universe(space
);
410 for (i
= 0; i
< array
->n_index
; ++i
) {
412 isl_set
*guard_i
, *zero
;
414 bound
= isl_pw_aff_copy(array
->bound
[i
]);
415 guard_i
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(bound
));
416 zero
= isl_pw_aff_zero_set(bound
);
417 guard_i
= isl_set_subtract(guard_i
, zero
);
418 guard
= isl_set_intersect(guard
, guard_i
);
424 /* Internal data structure for extract_size_of_type.
425 * "type" specifies the name of the space that we want to extract.
426 * "res" is used to store the subset of that space.
428 struct ppcg_extract_size_data
{
433 /* This function is called for each set in a union_set.
434 * If the name of the set matches data->type, we store the
437 static int extract_size_of_type(__isl_take isl_set
*size
, void *user
)
439 struct ppcg_extract_size_data
*data
= user
;
442 name
= isl_set_get_tuple_name(size
);
443 if (name
&& !strcmp(name
, data
->type
)) {
452 /* Given a union map { kernel[i] -> *[...] },
453 * return the range in the space called "type" for the kernel with
454 * sequence number "id".
456 static __isl_give isl_set
*extract_sizes(__isl_keep isl_union_map
*sizes
,
457 const char *type
, int id
)
461 isl_union_set
*local_sizes
;
462 struct ppcg_extract_size_data data
= { type
, NULL
};
467 space
= isl_union_map_get_space(sizes
);
468 space
= isl_space_set_from_params(space
);
469 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
470 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
471 dom
= isl_set_universe(space
);
472 dom
= isl_set_fix_si(dom
, isl_dim_set
, 0, id
);
474 local_sizes
= isl_union_set_apply(isl_union_set_from_set(dom
),
475 isl_union_map_copy(sizes
));
476 isl_union_set_foreach_set(local_sizes
, &extract_size_of_type
, &data
);
477 isl_union_set_free(local_sizes
);
481 /* Given a singleton set, extract the first (at most *len) elements
482 * of the single integer tuple into *sizes and update *len if needed.
484 static void read_sizes_from_set(__isl_take isl_set
*set
, int *sizes
, int *len
)
492 dim
= isl_set_dim(set
, isl_dim_set
);
496 for (i
= 0; i
< *len
; ++i
) {
499 v
= isl_set_plain_get_val_if_fixed(set
, isl_dim_set
, i
);
502 sizes
[i
] = isl_val_get_num_si(v
);
509 /* Add the map { kernel[id] -> type[sizes] } to gen->used_sizes,
510 * if the option debug->dump_sizes is set.
512 static void set_used_sizes(struct gpu_gen
*gen
, const char *type
, int id
,
519 if (!gen
->options
->debug
->dump_sizes
)
522 space
= isl_union_map_get_space(gen
->used_sizes
);
523 space
= isl_space_set_from_params(space
);
524 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
525 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
526 space
= isl_space_from_domain(space
);
527 space
= isl_space_add_dims(space
, isl_dim_out
, len
);
528 space
= isl_space_set_tuple_name(space
, isl_dim_out
, type
);
530 map
= isl_map_universe(space
);
531 map
= isl_map_fix_si(map
, isl_dim_in
, 0, id
);
532 for (i
= 0; i
< len
; ++i
)
533 map
= isl_map_fix_si(map
, isl_dim_out
, i
, sizes
[i
]);
535 gen
->used_sizes
= isl_union_map_add_map(gen
->used_sizes
, map
);
538 /* Extract user specified "tile" sizes from the "sizes" command line option,
539 * defaulting to option->tile_size in each dimension.
540 * *tile_len contains the maximum number of tile sizes needed.
541 * Update *tile_len to the number of specified tile sizes, if any, and
542 * return a pointer to the tile sizes (or NULL on error).
543 * Add the effectively used sizes to gen->used_sizes.
545 static int *read_tile_sizes(struct gpu_gen
*gen
, int *tile_len
)
551 tile_size
= isl_alloc_array(gen
->ctx
, int, *tile_len
);
554 for (n
= 0; n
< *tile_len
; ++n
)
555 tile_size
[n
] = gen
->options
->tile_size
;
557 size
= extract_sizes(gen
->sizes
, "tile", gen
->kernel_id
);
558 read_sizes_from_set(size
, tile_size
, tile_len
);
559 set_used_sizes(gen
, "tile", gen
->kernel_id
, tile_size
, *tile_len
);
564 /* Extract user specified "block" sizes from the "sizes" command line option,
565 * after filling in some potentially useful defaults.
567 static void read_block_sizes(struct ppcg_kernel
*kernel
,
568 __isl_keep isl_union_map
*sizes
)
572 if (kernel
->n_block
> 3)
574 switch (kernel
->n_block
) {
576 kernel
->block_dim
[0] = 512;
579 kernel
->block_dim
[0] = 32;
580 kernel
->block_dim
[1] = 16;
583 kernel
->block_dim
[0] = 32;
584 kernel
->block_dim
[1] = 4;
585 kernel
->block_dim
[2] = 4;
589 size
= extract_sizes(sizes
, "block", kernel
->id
);
590 read_sizes_from_set(size
, kernel
->block_dim
, &kernel
->n_block
);
593 /* Extract user specified "grid" sizes from the "sizes" command line option,
594 * after filling in some potentially useful defaults.
596 static void read_grid_sizes(struct ppcg_kernel
*kernel
,
597 __isl_keep isl_union_map
*sizes
)
601 if (kernel
->n_grid
> 2)
603 switch (kernel
->n_grid
) {
605 kernel
->grid_dim
[0] = 32768;
608 kernel
->grid_dim
[0] = 256;
609 kernel
->grid_dim
[1] = 256;
613 size
= extract_sizes(sizes
, "grid", kernel
->id
);
614 read_sizes_from_set(size
, kernel
->grid_dim
, &kernel
->n_grid
);
617 /* Extract user specified grid and block sizes from the gen->sizes
618 * command line option after filling in some potentially useful defaults.
619 * Store the extracted sizes in "kernel".
620 * Add the effectively used sizes to gen->used_sizes.
622 static void read_grid_and_block_sizes(struct ppcg_kernel
*kernel
,
625 read_block_sizes(kernel
, gen
->sizes
);
626 read_grid_sizes(kernel
, gen
->sizes
);
627 set_used_sizes(gen
, "block", kernel
->id
,
628 kernel
->block_dim
, kernel
->n_block
);
629 set_used_sizes(gen
, "grid", kernel
->id
,
630 kernel
->grid_dim
, kernel
->n_grid
);
633 static void *free_stmts(struct gpu_stmt
*stmts
, int n
)
640 for (i
= 0; i
< n
; ++i
) {
641 struct gpu_stmt_access
*access
, *next
;
643 for (access
= stmts
[i
].accesses
; access
; access
= next
) {
645 isl_id_free(access
->ref_id
);
646 isl_map_free(access
->access
);
647 isl_map_free(access
->tagged_access
);
651 isl_id_free(stmts
[i
].id
);
658 /* Add parameters p[i] with identifiers "ids" to "set",
659 * with bounds to 0 <= p[i] < size[i].
661 __isl_give isl_set
*add_bounded_parameters(__isl_take isl_set
*set
,
662 int *size
, __isl_keep isl_id_list
*ids
)
667 len
= isl_id_list_n_id(ids
);
668 nparam
= isl_set_dim(set
, isl_dim_param
);
669 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
671 for (i
= 0; i
< len
; ++i
) {
674 id
= isl_id_list_get_id(ids
, i
);
675 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
676 set
= isl_set_lower_bound_si(set
, isl_dim_param
, nparam
+ i
, 0);
677 set
= isl_set_upper_bound_si(set
, isl_dim_param
,
678 nparam
+ i
, size
[i
] - 1);
684 /* Add "len" parameters p[i] with identifiers "ids" and intersect "set"
687 * { : 0 <= p[i] < size[i] }
689 * or an overapproximation.
691 static __isl_give isl_set
*add_bounded_parameters_dynamic(
692 __isl_take isl_set
*set
, __isl_keep isl_multi_pw_aff
*size
,
693 __isl_keep isl_id_list
*ids
)
700 len
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
701 nparam
= isl_set_dim(set
, isl_dim_param
);
702 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
704 for (i
= 0; i
< len
; ++i
) {
707 id
= isl_id_list_get_id(ids
, i
);
708 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
711 space
= isl_space_params(isl_set_get_space(set
));
712 ls
= isl_local_space_from_space(space
);
713 for (i
= 0; i
< len
; ++i
) {
714 isl_pw_aff
*param
, *size_i
, *zero
;
717 param
= isl_pw_aff_var_on_domain(isl_local_space_copy(ls
),
718 isl_dim_param
, nparam
+ i
);
720 size_i
= isl_multi_pw_aff_get_pw_aff(size
, i
);
721 bound
= isl_pw_aff_lt_set(isl_pw_aff_copy(param
), size_i
);
722 bound
= isl_set_from_basic_set(isl_set_simple_hull(bound
));
723 set
= isl_set_intersect_params(set
, bound
);
725 zero
= isl_pw_aff_zero_on_domain(isl_local_space_copy(ls
));
726 bound
= isl_pw_aff_ge_set(param
, zero
);
727 set
= isl_set_intersect_params(set
, bound
);
729 isl_local_space_free(ls
);
734 /* Return the union of all tagged access relations in the group.
736 static __isl_give isl_union_map
*group_tagged_access_relation(
737 struct gpu_array_ref_group
*group
)
740 isl_union_map
*access
;
742 access
= isl_union_map_empty(isl_map_get_space(group
->access
));
743 for (i
= 0; i
< group
->n_ref
; ++i
) {
746 map_i
= isl_map_copy(group
->refs
[i
]->tagged_access
);
747 access
= isl_union_map_union(access
,
748 isl_union_map_from_map(map_i
));
754 /* Return the extent of "array", recomputed from the bounds.
755 * The recomputed extent may be simpler than the original extent.
757 static __isl_give isl_set
*array_extent(struct gpu_array_info
*array
)
765 id
= isl_set_get_tuple_id(array
->extent
);
766 space
= isl_set_get_space(array
->extent
);
767 extent
= isl_set_universe(isl_space_copy(space
));
768 ls
= isl_local_space_from_space(space
);
769 for (i
= 0; i
< array
->n_index
; ++i
) {
775 extent
= isl_set_lower_bound_si(extent
, isl_dim_set
, i
, 0);
777 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
779 index
= isl_pw_aff_from_aff(aff
);
780 bound
= isl_pw_aff_copy(array
->bound
[i
]);
781 bound
= isl_pw_aff_from_range(bound
);
782 bound
= isl_pw_aff_add_dims(bound
, isl_dim_in
, array
->n_index
);
783 bound
= isl_pw_aff_set_tuple_id(bound
, isl_dim_in
,
785 lt
= isl_pw_aff_lt_set(index
, bound
);
786 extent
= isl_set_intersect(extent
, lt
);
788 isl_local_space_free(ls
);
794 /* Return a map from the first group->depth dimensions of the computed
795 * schedule to the array tile in
796 * global memory that corresponds to the shared memory copy.
798 * In particular, return a map
804 * tile_offset(i) <= a <= tile_offset(i) + tile_size - 1 (1)
808 * 0 <= a <= array_size - 1 (2)
810 * Note that if some stride has been detected (i.e., when
811 * group->shared_tile->bound[i].shift is set), then a in (1) refers
812 * to the shifted and scaled down version.
814 * Constraints (1) are obtained by mapping the size constraints on the
815 * shared/private memory tile back to the access relation.
816 * Constraints (2) are obtained from the (recomputed) extent.
818 static __isl_give isl_map
*group_tile(struct gpu_array_ref_group
*group
)
821 int n_index
= group
->array
->n_index
;
827 space
= isl_multi_aff_get_space(group
->shared_tile
->tiling
);
828 space
= isl_space_range(space
);
829 local
= isl_set_universe(space
);
830 for (i
= 0; i
< n_index
; ++i
) {
833 local
= isl_set_lower_bound_si(local
, isl_dim_set
, i
, 0);
834 bound
= isl_val_copy(group
->shared_tile
->bound
[i
].size
);
835 bound
= isl_val_sub_ui(bound
, 1);
836 local
= isl_set_upper_bound_val(local
, isl_dim_set
, i
, bound
);
838 local
= isl_set_preimage_multi_aff(local
,
839 isl_multi_aff_copy(group
->shared_tile
->tiling
));
840 tile
= isl_set_unwrap(local
);
841 extent
= array_extent(group
->array
);
842 tile
= isl_map_intersect_range(tile
, extent
);
847 /* Given a mapping "iterator_map" from the AST schedule to a domain,
848 * return the corresponding mapping from the AST schedule to
849 * to the outer kernel->shared_schedule_dim dimensions of
850 * the schedule computed by PPCG for this kernel.
852 * Note that kernel->shared_schedule_dim is at least as large as
853 * the largest depth of any array reference group associated to the kernel.
854 * This is needed as the returned schedule is used to extract a mapping
855 * to the outer group->depth dimensions in transform_index.
857 static __isl_give isl_pw_multi_aff
*compute_sched_to_shared(
858 struct ppcg_kernel
*kernel
, __isl_take isl_pw_multi_aff
*iterator_map
)
860 isl_union_pw_multi_aff
*upma
;
861 isl_pw_multi_aff
*pma
;
864 space
= isl_space_range(isl_pw_multi_aff_get_space(iterator_map
));
865 space
= isl_space_from_domain(space
);
866 space
= isl_space_add_dims(space
, isl_dim_out
,
867 kernel
->shared_schedule_dim
);
869 upma
= isl_union_pw_multi_aff_copy(kernel
->shared_schedule
);
870 pma
= isl_union_pw_multi_aff_extract_pw_multi_aff(upma
, space
);
871 isl_union_pw_multi_aff_free(upma
);
873 return isl_pw_multi_aff_pullback_pw_multi_aff(pma
, iterator_map
);
876 /* If max_shared_memory is not set to infinity (-1), then make
877 * sure that the total amount of shared memory required by the
878 * array reference groups mapped to shared memory by "kernel"
879 * is no larger than this maximum.
881 * We apply a greedy approach and discard (keep in global memory)
882 * those groups that would result in a total memory size that
883 * is larger than the maximum.
885 * This function should be called after any function that may
886 * affect the decision on whether to place a reference group
887 * in private, shared or global memory.
889 static void check_shared_memory_bound(struct ppcg_kernel
*kernel
)
892 isl_val
*left
, *size
;
894 if (kernel
->options
->max_shared_memory
< 0)
897 left
= isl_val_int_from_si(kernel
->ctx
,
898 kernel
->options
->max_shared_memory
);
900 for (i
= 0; i
< kernel
->n_array
; ++i
) {
901 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
903 for (j
= 0; j
< local
->n_group
; ++j
) {
904 struct gpu_array_ref_group
*group
;
906 group
= local
->groups
[j
];
907 if (group
->private_tile
)
909 if (!group
->shared_tile
)
912 size
= gpu_array_tile_size(group
->shared_tile
);
913 size
= isl_val_mul_ui(size
, local
->array
->size
);
915 if (isl_val_le(size
, left
)) {
916 left
= isl_val_sub(left
, size
);
922 gpu_array_tile_free(group
->shared_tile
);
929 /* Compute a tiling for all the array reference groups in "kernel".
931 static void compute_group_tilings(struct ppcg_kernel
*kernel
)
935 for (i
= 0; i
< kernel
->n_array
; ++i
) {
936 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
938 for (j
= 0; j
< array
->n_group
; ++j
)
939 gpu_array_ref_group_compute_tiling(array
->groups
[j
]);
943 /* Compute the size of a bounding box around the origin and "set",
944 * where "set" is assumed to contain only non-negative elements.
945 * In particular, compute the maximal value of "set" in each direction
948 static __isl_give isl_multi_pw_aff
*extract_size(__isl_take isl_set
*set
,
949 __isl_take isl_set
*context
)
952 isl_multi_pw_aff
*mpa
;
954 context
= isl_set_params(context
);
955 n
= isl_set_dim(set
, isl_dim_set
);
956 mpa
= isl_multi_pw_aff_zero(isl_set_get_space(set
));
957 for (i
= 0; i
< n
; ++i
) {
962 bound
= isl_set_dim_max(isl_set_copy(set
), i
);
963 bound
= isl_pw_aff_coalesce(bound
);
964 bound
= isl_pw_aff_gist(bound
, isl_set_copy(context
));
966 space
= isl_pw_aff_get_domain_space(bound
);
967 one
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
968 one
= isl_aff_add_constant_si(one
, 1);
969 bound
= isl_pw_aff_add(bound
, isl_pw_aff_from_aff(one
));
970 mpa
= isl_multi_pw_aff_set_pw_aff(mpa
, i
, bound
);
973 isl_set_free(context
);
978 /* Compute the effective grid size as a list of the sizes in each dimension.
980 * The grid size specified by the user or set by default
981 * in read_grid_sizes() and applied by the block filter,
982 * may be too large for the given code in the sense that
983 * it may contain blocks that don't need to execute anything.
984 * We therefore don't return this grid size, but instead the
985 * smallest grid size that ensures that all blocks that actually
986 * execute code are included in the grid.
988 * We first extract a description of the grid, i.e., the possible values
989 * of the block ids, from the domain elements in "domain" and
990 * kernel->block_filter.
991 * The block ids are parameters in kernel->block_filter.
992 * We simply need to change them into set dimensions.
994 * Then, for each block dimension, we compute the maximal value of the block id
997 static __isl_give isl_multi_pw_aff
*extract_grid_size(
998 struct ppcg_kernel
*kernel
, __isl_take isl_union_set
*domain
)
1003 domain
= isl_union_set_intersect(domain
,
1004 isl_union_set_copy(kernel
->block_filter
));
1005 grid
= isl_union_set_params(domain
);
1006 grid
= isl_set_from_params(grid
);
1007 grid
= isl_set_add_dims(grid
, isl_dim_set
, kernel
->n_grid
);
1008 for (i
= 0; i
< kernel
->n_grid
; ++i
) {
1012 id
= isl_id_list_get_id(kernel
->block_ids
, i
);
1013 pos
= isl_set_find_dim_by_id(grid
, isl_dim_param
, id
);
1016 grid
= isl_set_equate(grid
, isl_dim_param
, pos
, isl_dim_set
, i
);
1017 grid
= isl_set_project_out(grid
, isl_dim_param
, pos
, 1);
1020 return extract_size(grid
, isl_set_copy(kernel
->context
));
1023 /* Compute the size of a fixed bounding box around the origin and "set",
1024 * where "set" is assumed to contain only non-negative elements,
1025 * and store the results in "size".
1026 * In particular, compute the maximal value of "set" in each direction
1029 static void extract_fixed_size(__isl_take isl_set
*set
, int *size
)
1032 isl_local_space
*ls
;
1035 n
= isl_set_dim(set
, isl_dim_set
);
1036 ls
= isl_local_space_from_space(isl_set_get_space(set
));
1037 obj
= isl_aff_zero_on_domain(ls
);
1038 for (i
= 0; i
< n
; ++i
) {
1041 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 1);
1042 max
= isl_set_max_val(set
, obj
);
1043 size
[i
] = isl_val_get_num_si(max
) + 1;
1045 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 0);
1051 /* Compute the effective block size as a list of the sizes in each dimension
1052 * and store the sizes in kernel->block_dim.
1054 * The block size specified by the user or set by default
1055 * in read_block_sizes() and applied by the thread filter,
1056 * may be too large for the given code in the sense that
1057 * it may contain threads that don't need to execute anything.
1058 * We therefore update this block size in kernel->block_dim
1059 * to the smallest block size that ensures that all threads
1060 * that actually execute code are included in the block.
1062 * The possible values of the thread ids is obtained from
1063 * the domain elements "domain" and kernel->thread_filter.
1064 * The current implementation eliminates all parameters, ensuring
1065 * that the size is a fixed constant in each dimension.
1066 * In principle we could also compute parametric sizes.
1067 * We would have to make sure to project out all b%d and t%d parameters,
1070 static void extract_block_size(struct ppcg_kernel
*kernel
,
1071 __isl_take isl_union_set
*domain
)
1077 domain
= isl_union_set_intersect(domain
,
1078 isl_union_set_copy(kernel
->thread_filter
));
1079 block
= isl_union_set_params(domain
);
1080 block
= isl_set_from_params(block
);
1081 block
= isl_set_add_dims(block
, isl_dim_set
, kernel
->n_block
);
1082 for (i
= 0; i
< kernel
->n_block
; ++i
) {
1086 id
= isl_id_list_get_id(kernel
->thread_ids
, i
);
1087 pos
= isl_set_find_dim_by_id(block
, isl_dim_param
, id
);
1090 block
= isl_set_equate(block
, isl_dim_param
, pos
,
1093 nparam
= isl_set_dim(block
, isl_dim_param
);
1094 block
= isl_set_project_out(block
, isl_dim_param
, 0, nparam
);
1096 extract_fixed_size(block
, kernel
->block_dim
);
1099 struct ppcg_kernel
*ppcg_kernel_free(struct ppcg_kernel
*kernel
)
1106 isl_id_list_free(kernel
->block_ids
);
1107 isl_id_list_free(kernel
->thread_ids
);
1108 isl_multi_pw_aff_free(kernel
->grid_size
);
1109 isl_set_free(kernel
->context
);
1110 isl_union_set_free(kernel
->core
);
1111 isl_union_set_free(kernel
->arrays
);
1112 isl_space_free(kernel
->space
);
1113 isl_ast_node_free(kernel
->tree
);
1114 isl_union_set_free(kernel
->block_filter
);
1115 isl_union_set_free(kernel
->thread_filter
);
1116 isl_union_pw_multi_aff_free(kernel
->shared_schedule
);
1117 isl_union_set_free(kernel
->sync_writes
);
1119 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1120 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1122 for (j
= 0; j
< array
->n_group
; ++j
)
1123 gpu_array_ref_group_free(array
->groups
[j
]);
1124 free(array
->groups
);
1126 isl_pw_aff_list_free(array
->bound
);
1128 free(kernel
->array
);
1130 for (i
= 0; i
< kernel
->n_var
; ++i
) {
1131 free(kernel
->var
[i
].name
);
1132 isl_vec_free(kernel
->var
[i
].size
);
1141 /* Wrapper around ppcg_kernel_free for use as a isl_id_set_free_user callback.
1143 static void ppcg_kernel_free_wrap(void *user
)
1145 struct ppcg_kernel
*kernel
= user
;
1147 ppcg_kernel_free(kernel
);
1150 static void create_kernel_var(isl_ctx
*ctx
, struct gpu_array_ref_group
*group
,
1151 struct ppcg_kernel_var
*var
)
1154 struct gpu_array_tile
*tile
;
1158 var
->array
= group
->array
;
1160 tile
= group
->private_tile
;
1161 var
->type
= ppcg_access_private
;
1163 tile
= group
->shared_tile
;
1164 var
->type
= ppcg_access_shared
;
1167 p
= isl_printer_to_str(ctx
);
1168 p
= gpu_array_ref_group_print_name(group
, p
);
1169 var
->name
= isl_printer_get_str(p
);
1170 isl_printer_free(p
);
1172 var
->size
= isl_vec_alloc(ctx
, group
->array
->n_index
);
1174 for (j
= 0; j
< group
->array
->n_index
; ++j
)
1175 var
->size
= isl_vec_set_element_val(var
->size
, j
,
1176 isl_val_copy(tile
->bound
[j
].size
));
1179 static int create_kernel_vars(struct ppcg_kernel
*kernel
)
1184 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1185 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1187 for (j
= 0; j
< array
->n_group
; ++j
) {
1188 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1189 if (group
->private_tile
|| group
->shared_tile
)
1195 kernel
->var
= isl_calloc_array(kernel
->ctx
, struct ppcg_kernel_var
, n
);
1200 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1201 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1203 for (j
= 0; j
< array
->n_group
; ++j
) {
1204 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1205 if (!group
->private_tile
&& !group
->shared_tile
)
1207 create_kernel_var(kernel
->ctx
, group
, &kernel
->var
[n
]);
1215 /* Replace "pa" by the zero function defined over the universe domain
1216 * in the space of "pa".
1218 static __isl_give isl_pw_aff
*set_universally_zero(__isl_take isl_pw_aff
*pa
)
1223 space
= isl_space_domain(isl_pw_aff_get_space(pa
));
1224 isl_pw_aff_free(pa
);
1225 zero
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1227 return isl_pw_aff_from_aff(zero
);
1230 /* The sizes of the arrays on the host that have been computed by
1231 * extract_array_info may depend on the parameters. Use the extra
1232 * constraints on the parameters that are valid at "host_domain"
1233 * to simplify these expressions and store the results in kernel->array.
1235 * We only need these localized bounds for arrays that are accessed
1236 * by the current kernel. If we have found at least one reference group
1237 * then the array is accessed by the kernel. If the array has compound
1238 * elements then we skipped the construction of array reference groups.
1240 * The resulting sizes may be functions that are nowhere defined
1241 * in case the access function cannot possibly access anything inside
1242 * the kernel for some reason. If so, they are replaced by the zero
1243 * function. Since the access function cannot actually access anything,
1244 * there is no harm in printing the array sizes as zero.
1246 static void localize_bounds(struct ppcg_kernel
*kernel
,
1247 __isl_keep isl_set
*host_domain
)
1252 context
= isl_set_copy(host_domain
);
1253 context
= isl_set_params(context
);
1255 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1256 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
1257 isl_pw_aff_list
*bound
;
1260 if (local
->n_group
== 0 && !local
->array
->has_compound_element
)
1263 n_index
= local
->array
->n_index
;
1264 bound
= isl_pw_aff_list_alloc(kernel
->ctx
, n_index
);
1266 for (j
= 0; j
< n_index
; ++j
) {
1270 pwaff
= isl_pw_aff_copy(local
->array
->bound
[j
]);
1271 pwaff
= isl_pw_aff_gist(pwaff
, isl_set_copy(context
));
1272 empty
= isl_pw_aff_is_empty(pwaff
);
1274 pwaff
= isl_pw_aff_free(pwaff
);
1276 pwaff
= set_universally_zero(pwaff
);
1277 bound
= isl_pw_aff_list_add(bound
, pwaff
);
1280 local
->n_index
= n_index
;
1281 local
->bound
= bound
;
1283 isl_set_free(context
);
1286 /* Create the array of gpu_local_array_info structures "array"
1287 * inside "kernel". The number of elements in this array is
1288 * the same as the number of arrays in "prog".
1289 * Initialize the "array" field of each local array to point
1290 * to the corresponding array in "prog".
1292 static struct ppcg_kernel
*ppcg_kernel_create_local_arrays(
1293 struct ppcg_kernel
*kernel
, struct gpu_prog
*prog
)
1298 ctx
= isl_set_get_ctx(prog
->context
);
1299 kernel
->array
= isl_calloc_array(ctx
,
1300 struct gpu_local_array_info
, prog
->n_array
);
1302 return ppcg_kernel_free(kernel
);
1303 kernel
->n_array
= prog
->n_array
;
1305 for (i
= 0; i
< prog
->n_array
; ++i
)
1306 kernel
->array
[i
].array
= &prog
->array
[i
];
1311 /* Find the element in gen->stmt that has the given "id".
1312 * Return NULL if no such gpu_stmt can be found.
1314 static struct gpu_stmt
*find_stmt(struct gpu_prog
*prog
, __isl_keep isl_id
*id
)
1318 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
1319 if (id
== prog
->stmts
[i
].id
)
1323 return i
< prog
->n_stmts
? &prog
->stmts
[i
] : NULL
;
1326 void ppcg_kernel_stmt_free(void *user
)
1329 struct ppcg_kernel_stmt
*stmt
= user
;
1334 switch (stmt
->type
) {
1335 case ppcg_kernel_copy
:
1336 isl_ast_expr_free(stmt
->u
.c
.index
);
1337 isl_ast_expr_free(stmt
->u
.c
.local_index
);
1339 case ppcg_kernel_domain
:
1340 isl_id_to_ast_expr_free(stmt
->u
.d
.ref2expr
);
1342 case ppcg_kernel_sync
:
1349 /* Return the gpu_stmt_access in the list "accesses" that corresponds
1352 static struct gpu_stmt_access
*find_access(struct gpu_stmt_access
*accesses
,
1353 __isl_keep isl_id
*ref_id
)
1355 struct gpu_stmt_access
*access
;
1357 for (access
= accesses
; access
; access
= access
->next
)
1358 if (access
->ref_id
== ref_id
)
1364 /* Return the index of the array called "name" in the list of arrays.
1366 static int find_array_index(struct ppcg_kernel
*kernel
, const char *name
)
1370 for (i
= 0; i
< kernel
->n_array
; ++i
)
1371 if (!strcmp(name
, kernel
->array
[i
].array
->name
))
1377 /* Internal data structure for the index and AST expression transformation
1378 * callbacks for pet_stmt_build_ast_exprs.
1380 * "kernel" is the kernel for which are computing AST expressions.
1381 * "accesses" is the list of gpu_stmt_access in the statement.
1382 * "iterator_map" expresses the statement iterators in terms of
1383 * the AST loop iterators.
1384 * "sched2shared" expresses the outer shared_schedule_dim dimensions of
1385 * the kernel schedule in terms of the AST loop iterators.
1387 * The following fields are set in transform_index and used in transform_expr.
1388 * "array" is the array that is being accessed.
1389 * "global" is set if the global array is accessed (rather than
1390 * shared/private memory).
1391 * "local_array" refers to information on the array specialized
1392 * to the current kernel.
1394 struct ppcg_transform_data
{
1395 struct ppcg_kernel
*kernel
;
1396 struct gpu_stmt_access
*accesses
;
1397 isl_pw_multi_aff
*iterator_map
;
1398 isl_pw_multi_aff
*sched2shared
;
1400 struct gpu_array_info
*array
;
1402 struct gpu_local_array_info
*local_array
;
1405 /* Return the name of the outer array (of structs) accessed by "access".
1407 static const char *get_outer_array_name(__isl_keep isl_map
*access
)
1412 space
= isl_space_range(isl_map_get_space(access
));
1413 while (space
&& isl_space_is_wrapping(space
))
1414 space
= isl_space_domain(isl_space_unwrap(space
));
1415 name
= isl_space_get_tuple_name(space
, isl_dim_set
);
1416 isl_space_free(space
);
1421 /* Return a pointer to the gpu_array_ref_group in "local"
1422 * that contains the reference "access".
1423 * Return NULL if no such group can be found.
1425 static struct gpu_array_ref_group
*find_ref_group(
1426 struct gpu_local_array_info
*local
, struct gpu_stmt_access
*access
)
1430 for (i
= 0; i
< local
->n_group
; ++i
) {
1431 struct gpu_array_ref_group
*group
= local
->groups
[i
];
1433 for (j
= 0; j
< group
->n_ref
; ++j
)
1434 if (group
->refs
[j
] == access
)
1441 /* Index transformation callback for pet_stmt_build_ast_exprs.
1443 * "index" expresses the array indices in terms of statement iterators
1445 * We first reformulate "index" in terms of the AST loop iterators.
1446 * Then we check if we are accessing the global array or
1447 * a shared/private copy. In the former case, we simply return
1448 * the updated index. If "index" is an affine expression rather
1449 * than an array access, then we also return the updated index here.
1451 * If no reference groups have been computed for the array,
1452 * then we can only be accessing the global array.
1454 * Otherwise, we apply the tiling to the index.
1455 * This tiling is of the form
1459 * where D corresponds to the outer group->depth dimensions of
1460 * the kernel schedule.
1461 * The index is of the form
1465 * We update the tiling to refer to the AST loop iterators
1469 * and modify index to keep track of those iterators
1473 * Combining these two yields a tiled index expression in terms
1474 * of the AST loop iterators
1478 static __isl_give isl_multi_pw_aff
*transform_index(
1479 __isl_take isl_multi_pw_aff
*index
, __isl_keep isl_id
*ref_id
,
1482 struct ppcg_transform_data
*data
= user
;
1483 struct gpu_stmt_access
*access
;
1484 struct gpu_array_ref_group
*group
;
1485 struct gpu_array_tile
*tile
;
1486 isl_pw_multi_aff
*iterator_map
;
1491 isl_multi_pw_aff
*tiling
;
1492 isl_pw_multi_aff
*pma
;
1493 isl_multi_pw_aff
*mpa
;
1494 isl_pw_multi_aff
*sched2depth
;
1498 iterator_map
= isl_pw_multi_aff_copy(data
->iterator_map
);
1499 index
= isl_multi_pw_aff_pullback_pw_multi_aff(index
, iterator_map
);
1501 access
= find_access(data
->accesses
, ref_id
);
1504 if (!isl_map_has_tuple_name(access
->access
, isl_dim_out
))
1507 name
= get_outer_array_name(access
->access
);
1508 i
= find_array_index(data
->kernel
, name
);
1510 isl_die(isl_multi_pw_aff_get_ctx(index
), isl_error_internal
,
1511 "cannot find array",
1512 return isl_multi_pw_aff_free(index
));
1513 data
->local_array
= &data
->kernel
->array
[i
];
1514 data
->array
= data
->local_array
->array
;
1516 group
= find_ref_group(data
->local_array
, access
);
1522 tile
= group
->private_tile
;
1524 tile
= group
->shared_tile
;
1525 data
->global
= !tile
;
1529 space
= isl_space_range(isl_multi_pw_aff_get_space(index
));
1530 space
= isl_space_map_from_set(space
);
1531 pma
= isl_pw_multi_aff_identity(space
);
1532 sched2depth
= isl_pw_multi_aff_copy(data
->sched2shared
);
1533 dim
= isl_pw_multi_aff_dim(sched2depth
, isl_dim_out
);
1534 sched2depth
= isl_pw_multi_aff_drop_dims(sched2depth
, isl_dim_out
,
1535 group
->depth
, dim
- group
->depth
);
1536 pma
= isl_pw_multi_aff_product(sched2depth
, pma
);
1537 tiling
= isl_multi_pw_aff_from_multi_aff(
1538 isl_multi_aff_copy(tile
->tiling
));
1539 tiling
= isl_multi_pw_aff_pullback_pw_multi_aff(tiling
, pma
);
1541 space
= isl_space_domain(isl_multi_pw_aff_get_space(index
));
1542 space
= isl_space_map_from_set(space
);
1543 mpa
= isl_multi_pw_aff_identity(space
);
1544 index
= isl_multi_pw_aff_range_product(mpa
, index
);
1545 index
= isl_multi_pw_aff_pullback_multi_pw_aff(tiling
, index
);
1550 /* Dereference "expr" by adding an index [0].
1551 * The original "expr" is assumed not to have any indices.
1553 * If "expr" is a member access, then the dereferencing needs
1554 * to be applied to the structure argument of this member access.
1556 static __isl_give isl_ast_expr
*dereference(__isl_take isl_ast_expr
*expr
)
1559 isl_ast_expr
*arg0
, *res
;
1560 isl_ast_expr_list
*list
;
1562 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1564 return isl_ast_expr_free(expr
);
1565 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1566 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1569 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1570 arg
= dereference(arg
);
1571 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1572 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1576 isl_ast_expr_free(arg0
);
1578 ctx
= isl_ast_expr_get_ctx(expr
);
1579 res
= isl_ast_expr_from_val(isl_val_zero(ctx
));
1580 list
= isl_ast_expr_list_from_ast_expr(res
);
1581 res
= isl_ast_expr_get_op_arg(expr
, 0);
1582 res
= isl_ast_expr_access(res
, list
);
1583 isl_ast_expr_free(expr
);
1588 /* Linearize the index expression "expr" based on the array bounds
1591 * That is, transform expression
1593 * A[i_0][i_1]...[i_n]
1597 * A[(..((i_0 * b_1 + i_1) ... ) * b_n + i_n]
1599 * where b_0, b_1, ..., b_n are the bounds on the array.
1601 * If the base of "expr" is a member access, then the linearization needs
1602 * to be applied to the structure argument of this member access.
1604 * In the base case, if "expr" has no arguments (other than the name of
1605 * the array), then we are passing an entire array to a function.
1606 * In this case, there is nothing to linearize.
1607 * Note that at this point an expression with no arguments can
1608 * only be an entire array because the scalar case and
1609 * the case of single struct are handled by the caller.
1611 * If the number of specified index expressions in "expr"
1612 * is smaller than the dimension of the accessed array,
1613 * then the missing i_j also do not appear in the linearized expression.
1614 * Furthermore, since such an expression does not refer to a single
1615 * element while the default linearized expression would refer to
1616 * a single element, we return the expression
1618 * A + (..((i_0 * b_1 + i_1) ... ) * b_n]
1620 * instead. Note that because of the special case handling above,
1621 * we can assume here that here that there is at least one index expression.
1623 __isl_give isl_ast_expr
*gpu_local_array_info_linearize_index(
1624 struct gpu_local_array_info
*array
, __isl_take isl_ast_expr
*expr
)
1631 isl_ast_expr_list
*list
;
1632 isl_ast_build
*build
;
1634 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1635 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1636 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1639 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1640 arg
= gpu_local_array_info_linearize_index(array
, arg
);
1641 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1642 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1646 isl_ast_expr_free(arg0
);
1648 if (isl_ast_expr_get_op_n_arg(expr
) == 1)
1651 ctx
= isl_ast_expr_get_ctx(expr
);
1652 context
= isl_set_universe(isl_space_params_alloc(ctx
, 0));
1653 build
= isl_ast_build_from_context(context
);
1655 n
= isl_ast_expr_get_op_n_arg(expr
);
1656 res
= isl_ast_expr_get_op_arg(expr
, 1);
1657 for (i
= 1; i
< array
->n_index
; ++i
) {
1658 isl_pw_aff
*bound_i
;
1659 isl_ast_expr
*expr_i
;
1661 bound_i
= isl_pw_aff_list_get_pw_aff(array
->bound
, i
);
1662 expr_i
= isl_ast_build_expr_from_pw_aff(build
, bound_i
);
1663 res
= isl_ast_expr_mul(res
, expr_i
);
1667 expr_i
= isl_ast_expr_get_op_arg(expr
, i
+ 1);
1668 res
= isl_ast_expr_add(res
, expr_i
);
1671 isl_ast_build_free(build
);
1673 if (1 + array
->n_index
> n
) {
1674 res
= isl_ast_expr_add(isl_ast_expr_get_op_arg(expr
, 0), res
);
1676 list
= isl_ast_expr_list_from_ast_expr(res
);
1677 res
= isl_ast_expr_get_op_arg(expr
, 0);
1678 res
= isl_ast_expr_access(res
, list
);
1681 isl_ast_expr_free(expr
);
1686 /* AST expression transformation callback for pet_stmt_build_ast_exprs.
1688 * If the AST expression refers to an array that is not accessed
1689 * at all, then this means the value of the expression is not used,
1690 * so we might as well print zero (NULL pointer) instead.
1692 * If the AST expression refers to a global scalar that is not
1693 * a read-only scalar, then its address was passed to the kernel and
1694 * we need to dereference it.
1696 * If the AST expression refers to an access to a global array,
1697 * then we linearize the access exploiting the bounds in data->local_array.
1699 static __isl_give isl_ast_expr
*transform_expr(__isl_take isl_ast_expr
*expr
,
1700 __isl_keep isl_id
*id
, void *user
)
1702 struct ppcg_transform_data
*data
= user
;
1706 if (!data
->array
->accessed
) {
1709 ctx
= isl_ast_expr_get_ctx(expr
);
1710 isl_ast_expr_free(expr
);
1711 return isl_ast_expr_from_val(isl_val_zero(ctx
));
1713 if (gpu_array_is_read_only_scalar(data
->array
))
1717 if (data
->array
->n_index
== 0)
1718 return dereference(expr
);
1719 if (!data
->array
->linearize
)
1722 return gpu_local_array_info_linearize_index(data
->local_array
, expr
);
1725 /* This function is called for each instance of a user statement
1726 * in the kernel "kernel", identified by "gpu_stmt".
1728 * We attach a struct ppcg_kernel_stmt to the "node", containing
1729 * a computed AST expression for each access.
1730 * These AST expressions are computed from iterator_map,
1731 * which expresses the domain
1732 * elements in terms of the generated loops, and sched2shared,
1733 * which expresses the outer shared_schedule_dim dimensions of
1734 * the kernel schedule computed by PPCG in terms of the generated loops.
1736 static __isl_give isl_ast_node
*create_domain_leaf(
1737 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1738 __isl_keep isl_ast_build
*build
, struct gpu_stmt
*gpu_stmt
)
1740 struct ppcg_transform_data data
;
1741 struct ppcg_kernel_stmt
*stmt
;
1743 isl_pw_multi_aff
*sched2shared
;
1745 isl_pw_multi_aff
*iterator_map
;
1746 isl_union_map
*schedule
;
1748 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1750 return isl_ast_node_free(node
);
1752 schedule
= isl_ast_build_get_schedule(build
);
1753 map
= isl_map_reverse(isl_map_from_union_map(schedule
));
1754 iterator_map
= isl_pw_multi_aff_from_map(map
);
1755 sched2shared
= compute_sched_to_shared(kernel
,
1756 isl_pw_multi_aff_copy(iterator_map
));
1758 stmt
->type
= ppcg_kernel_domain
;
1759 stmt
->u
.d
.stmt
= gpu_stmt
;
1761 data
.kernel
= kernel
;
1762 data
.accesses
= stmt
->u
.d
.stmt
->accesses
;
1763 data
.iterator_map
= iterator_map
;
1764 data
.sched2shared
= sched2shared
;
1765 stmt
->u
.d
.ref2expr
= pet_stmt_build_ast_exprs(stmt
->u
.d
.stmt
->stmt
,
1766 build
, &transform_index
, &data
,
1767 &transform_expr
, &data
);
1769 isl_pw_multi_aff_free(iterator_map
);
1770 isl_pw_multi_aff_free(sched2shared
);
1772 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1773 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1774 return isl_ast_node_set_annotation(node
, id
);
1777 /* This function is called for each statement node in the AST
1778 * for copying to or from shared/private memory.
1779 * Attach a pointer to a ppcg_kernel_stmt representing the copy
1780 * statement to the node.
1781 * The statement name is "read" or "write", depending on whether we are
1782 * reading from global memory or writing to global memory.
1784 * The schedule is of the form
1788 * where D corresponds to the outer group->depth dimensions of
1789 * the kernel schedule, A to the global array and L to the outer
1790 * generated AST schedule.
1791 * We compute the inverse and strip off the type, resulting in
1795 * We combine this mapping with on the one hand the projection
1799 * and on the other hand the group tiling
1807 * and store the corresponding expressions in stmt->index and stmt->local_index,
1808 * where stmt points to the ppcg_kernel_stmt that is attached to the node.
1810 static __isl_give isl_ast_node
*create_access_leaf(struct ppcg_kernel
*kernel
,
1811 struct gpu_array_ref_group
*group
, __isl_take isl_ast_node
*node
,
1812 __isl_keep isl_ast_build
*build
)
1814 struct ppcg_kernel_stmt
*stmt
;
1815 struct gpu_array_tile
*tile
;
1820 isl_pw_multi_aff
*pma
, *pma2
;
1823 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1825 return isl_ast_node_free(node
);
1827 access
= isl_map_from_union_map(isl_ast_build_get_schedule(build
));
1828 type
= isl_map_get_tuple_name(access
, isl_dim_in
);
1829 stmt
->u
.c
.read
= !strcmp(type
, "read");
1830 access
= isl_map_reverse(access
);
1831 pma
= isl_pw_multi_aff_from_map(access
);
1832 pma
= isl_pw_multi_aff_reset_tuple_id(pma
, isl_dim_out
);
1834 space
= isl_space_range(isl_pw_multi_aff_get_space(pma
));
1835 space
= isl_space_unwrap(space
);
1836 pma2
= isl_pw_multi_aff_range_map(space
);
1837 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
,
1838 isl_pw_multi_aff_copy(pma
));
1839 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1840 stmt
->u
.c
.index
= expr
;
1842 tile
= gpu_array_ref_group_tile(group
);
1843 pma2
= isl_pw_multi_aff_from_multi_aff(
1844 isl_multi_aff_copy(tile
->tiling
));
1845 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
, pma
);
1846 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1847 stmt
->u
.c
.local_index
= expr
;
1849 stmt
->u
.c
.array
= group
->array
;
1850 stmt
->u
.c
.local_array
= group
->local_array
;
1851 stmt
->type
= ppcg_kernel_copy
;
1853 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1854 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1855 return isl_ast_node_set_annotation(node
, id
);
1858 /* Create a synchronization ppcg_kernel_stmt and
1859 * attach it to the node "node" representing the synchronization.
1861 static __isl_give isl_ast_node
*create_sync_leaf(
1862 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1863 __isl_keep isl_ast_build
*build
)
1865 struct ppcg_kernel_stmt
*stmt
;
1868 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1870 return isl_ast_node_free(node
);
1872 stmt
->type
= ppcg_kernel_sync
;
1873 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1874 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1875 return isl_ast_node_set_annotation(node
, id
);
1878 /* Internal data structure for at_domain.
1880 * "prog" represents the entire scop.
1881 * "kernel" points to the kernel to which the current schedule node
1882 * belongs. It is set by before_mark and reset by after_mark.
1884 struct ppcg_at_domain_data
{
1885 struct gpu_prog
*prog
;
1886 struct ppcg_kernel
*kernel
;
1889 /* This function is called for each instance of a user statement
1890 * in the kernel. This may be one of the original user statements
1891 * or a statement introduced by PPCG.
1893 * We assume that the original user statements only have a name
1894 * and no user pointer. The statements introduced by PPCG
1895 * on the other hand all have a user pointer.
1897 * If the user statement is one of the original user statements
1898 * (one with no user pointer), then we call create_domain_leaf. Otherwise,
1899 * we check if it is a copy or synchronization statement and
1900 * call the appropriate functions.
1902 static __isl_give isl_ast_node
*at_domain(__isl_take isl_ast_node
*node
,
1903 __isl_keep isl_ast_build
*build
, void *user
)
1905 struct ppcg_at_domain_data
*data
= user
;
1906 isl_ast_expr
*expr
, *arg
;
1912 expr
= isl_ast_node_user_get_expr(node
);
1913 arg
= isl_ast_expr_get_op_arg(expr
, 0);
1914 id
= isl_ast_expr_get_id(arg
);
1915 name
= isl_id_get_name(id
);
1916 p
= isl_id_get_user(id
);
1917 isl_ast_expr_free(expr
);
1918 isl_ast_expr_free(arg
);
1921 struct gpu_stmt
*gpu_stmt
;
1923 gpu_stmt
= find_stmt(data
->prog
, id
);
1926 isl_die(data
->prog
->ctx
, isl_error_internal
,
1927 "statement not found",
1928 return isl_ast_node_free(node
));
1930 return create_domain_leaf(data
->kernel
, node
, build
, gpu_stmt
);
1933 is_sync
= gpu_tree_id_is_sync(id
, data
->kernel
);
1936 return isl_ast_node_free(node
);
1937 if (!strcmp(name
, "read") || !strcmp(name
, "write")) {
1938 struct gpu_array_ref_group
*group
= p
;
1939 return create_access_leaf(data
->kernel
, group
, node
, build
);
1942 isl_die(data
->prog
->ctx
, isl_error_internal
,
1943 "unknown statement type",
1944 return isl_ast_node_free(node
));
1945 return create_sync_leaf(data
->kernel
, node
, build
);
1948 /* Given a set of wrapped references "ref", return the corresponding
1949 * access relations based on the tagged access relations "tagged".
1951 * The elements of "ref" are of the form
1955 * with D an iteration domains and R a reference.
1956 * The elements of "tagged" are of the form
1962 * Extend "tagged" to include the iteration domain in the range, i.e.,
1964 * [D -> R] -> [D -> A]
1966 * apply the result to "ref" and then unwrap the resulting set
1967 * to obtain relations of the form
1971 static __isl_give isl_union_map
*wrapped_reference_to_access(
1972 __isl_take isl_union_set
*ref
, __isl_take isl_union_map
*tagged
)
1974 isl_union_map
*tag2access
;
1976 tag2access
= isl_union_map_copy(tagged
);
1977 tag2access
= isl_union_map_universe(tag2access
);
1978 tag2access
= isl_union_set_unwrap(isl_union_map_domain(tag2access
));
1979 tag2access
= isl_union_map_domain_map(tag2access
);
1980 tag2access
= isl_union_map_range_product(tag2access
, tagged
);
1982 ref
= isl_union_set_coalesce(ref
);
1983 ref
= isl_union_set_apply(ref
, tag2access
);
1985 return isl_union_set_unwrap(ref
);
1988 /* Given an access relation "access" from "group", remove those reads
1989 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
1990 * communicate data within the same iteration of "sched".
1992 * If the access is a read then it is either an element of
1994 * live_in union (range flow)
1996 * where live_in and flow may be overapproximations, or
1997 * it reads an uninitialized value (that is not live-in because
1998 * there is an intermediate kill) or it reads a value that was
1999 * written within the same (compound) statement instance.
2000 * If the access is a write then it is either an element of
2002 * live_out union (domain flow)
2004 * or it writes a value that is never read (and is not live-out
2005 * because of an intermediate kill) or only
2006 * within the same (compound) statement instance.
2007 * In both cases, the access relation is also a subset of
2008 * the group access relation.
2010 * The cases where an uninitialized value is read or a value is written
2011 * that is never read or where the dataflow occurs within a statement
2012 * instance are also considered local and may also be removed.
2014 * Essentially, we compute the intersection of "access" with either
2016 * live_in union (range non-local-flow)
2020 * live_out union (domain non-local-flow)
2022 * We first construct a relation "local"
2024 * [[D -> R] -> [D' -> R']]
2026 * of pairs of domain iterations accessing the reference group
2027 * and references in the group that are coscheduled by "sched".
2029 * If this relation does not intersect the dataflow dependences,
2030 * then there is nothing we can possibly remove, unless the dataflow
2031 * dependences themselves only relate a subset of the accesses.
2032 * In particular, the accesses may not be involved in any dataflow
2033 * dependences, either because they are uninitialized reads/dead writes
2034 * or because the dataflow occurs inside a statement instance.
2036 * Since the computation below may break up the access relation
2037 * into smaller pieces, we only perform the intersection with
2038 * the non-local dependent accesses if the local pairs
2039 * intersect the dataflow dependences. Otherwise, we intersect
2040 * with the universe of the non-local dependent accesses.
2041 * This should at least remove accesses from statements that
2042 * do not participate in any dependences.
2044 * In particular, we remove the "local" dataflow dependences from
2045 * the set of all dataflow dependences.
2046 * Note that if the potential dataflow dependences are an overapproximation
2047 * of the actual dataflow dependences, then the result remains an
2048 * overapproximation of the non-local dataflow dependences.
2049 * Copying to/from global memory is only needed for the references
2050 * in the domain/range of the result or for accesses that are live out/in
2051 * for the entire scop.
2053 * We therefore map the domain/range of the "external" relation
2054 * to the corresponding access relation and take the union with
2055 * the live out/in relation.
2057 static __isl_give isl_union_map
*remove_local_accesses(
2058 struct gpu_prog
*prog
, struct gpu_array_ref_group
*group
,
2059 __isl_take isl_union_map
*access
, __isl_take isl_union_map
*sched
,
2063 isl_union_pw_multi_aff
*tagger
;
2064 isl_union_set
*domain
;
2065 isl_union_map
*local
, *tagged
, *external
;
2066 isl_union_set
*tag_set
;
2068 if (isl_union_map_is_empty(access
)) {
2069 isl_union_map_free(sched
);
2073 tagged
= group_tagged_access_relation(group
);
2075 tagger
= isl_union_pw_multi_aff_copy(prog
->scop
->tagger
);
2076 domain
= isl_union_map_domain(isl_union_map_copy(tagged
));
2077 tagger
= isl_union_pw_multi_aff_intersect_domain(tagger
, domain
);
2078 sched
= isl_union_map_preimage_domain_union_pw_multi_aff(sched
, tagger
);
2080 local
= isl_union_map_apply_range(sched
,
2081 isl_union_map_reverse(isl_union_map_copy(sched
)));
2082 local
= isl_union_map_intersect(local
,
2083 isl_union_map_copy(prog
->scop
->tagged_dep_flow
));
2085 empty
= isl_union_map_is_empty(local
);
2087 external
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
2088 external
= isl_union_map_intersect_params(external
,
2089 isl_set_copy(prog
->scop
->context
));
2090 external
= isl_union_map_subtract(external
, local
);
2093 tag_set
= isl_union_map_range(external
);
2094 external
= wrapped_reference_to_access(tag_set
, tagged
);
2095 external
= isl_union_map_union(external
,
2096 isl_union_map_copy(prog
->scop
->live_in
));
2098 tag_set
= isl_union_map_domain(external
);
2099 external
= wrapped_reference_to_access(tag_set
, tagged
);
2100 external
= isl_union_map_union(external
,
2101 isl_union_map_copy(prog
->scop
->live_out
));
2105 external
= isl_union_map_free(external
);
2107 external
= isl_union_map_universe(external
);
2109 access
= isl_union_map_intersect(access
, external
);
2114 /* Given an access relation "access" from "group", remove those reads
2115 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
2116 * communicate data within the same iteration of the schedule at the
2117 * position where the copying of the group is inserted.
2118 * "node" points to this position, i.e., the depth at "node"
2119 * is equal to group->depth.
2121 * We extract a schedule that picks out the iterations of the outer
2122 * group->depth dimensions and call remove_local_accesses.
2124 static __isl_give isl_union_map
*remove_local_accesses_group(
2125 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2126 __isl_take isl_union_map
*access
, __isl_keep isl_schedule_node
*node
,
2129 isl_union_map
*sched
;
2131 if (isl_union_map_is_empty(access
))
2134 sched
= isl_schedule_node_get_prefix_schedule_relation(node
);
2136 return remove_local_accesses(kernel
->prog
, group
, access
, sched
, read
);
2139 /* This function is called before the AST generator starts traversing
2140 * the schedule subtree of a node with mark "mark".
2142 * If the mark is called "kernel", store the kernel pointer in data->kernel
2143 * for use in at_domain.
2145 static int before_mark(__isl_keep isl_id
*mark
,
2146 __isl_keep isl_ast_build
*build
, void *user
)
2148 struct ppcg_at_domain_data
*data
= user
;
2152 if (!strcmp(isl_id_get_name(mark
), "kernel"))
2153 data
->kernel
= isl_id_get_user(mark
);
2157 /* This function is called after the AST generator has finished traversing
2158 * the schedule subtree of a mark node. "node" points to the corresponding
2161 * If the mark is called "kernel", then replace "node" by a user node
2162 * that "calls" the kernel, representing the launch of the kernel.
2163 * The original "node" is stored inside the kernel object so that
2164 * it can be used to print the device code.
2165 * Note that this assumes that a kernel is only launched once.
2166 * Also clear data->kernel.
2168 static __isl_give isl_ast_node
*after_mark(__isl_take isl_ast_node
*node
,
2169 __isl_keep isl_ast_build
*build
, void *user
)
2174 isl_ast_expr_list
*list
;
2175 struct ppcg_kernel
*kernel
;
2176 struct ppcg_at_domain_data
*data
= user
;
2178 ctx
= isl_ast_node_get_ctx(node
);
2179 id
= isl_ast_node_mark_get_id(node
);
2181 return isl_ast_node_free(node
);
2182 if (strcmp(isl_id_get_name(id
), "kernel") || !data
->kernel
) {
2186 kernel
= data
->kernel
;
2187 data
->kernel
= NULL
;
2188 kernel
->space
= isl_ast_build_get_schedule_space(build
);
2189 kernel
->tree
= isl_ast_node_mark_get_node(node
);
2190 isl_ast_node_free(node
);
2192 expr
= isl_ast_expr_from_id(isl_id_copy(id
));
2193 list
= isl_ast_expr_list_alloc(ctx
, 0);
2194 expr
= isl_ast_expr_call(expr
, list
);
2195 node
= isl_ast_node_alloc_user(expr
);
2196 node
= isl_ast_node_set_annotation(node
, id
);
2201 static int update_depth(__isl_keep isl_schedule_node
*node
, void *user
)
2206 if (isl_schedule_node_get_type(node
) != isl_schedule_node_leaf
)
2208 node_depth
= isl_schedule_node_get_schedule_depth(node
);
2209 if (node_depth
> *depth
)
2210 *depth
= node_depth
;
2215 /* Use isl to generate code for both the host and the device
2217 * The device code is marked by "kernel" mark nodes in the schedule tree,
2218 * containing a pointer to a ppcg_kernel object.
2219 * The returned AST only contains the AST for the host code.
2220 * The ASTs for the device code are embedded in ppcg_kernel objects
2221 * attached to the leaf nodes that call "kernel".
2223 static __isl_give isl_ast_node
*generate_code(struct gpu_gen
*gen
,
2224 __isl_take isl_schedule
*schedule
)
2226 struct ppcg_at_domain_data data
;
2227 isl_ast_build
*build
;
2229 isl_id_list
*iterators
;
2232 data
.prog
= gen
->prog
;
2236 if (isl_schedule_foreach_schedule_node(schedule
, &update_depth
,
2239 build
= isl_ast_build_alloc(gen
->prog
->ctx
);
2240 iterators
= ppcg_scop_generate_names(gen
->prog
->scop
, depth
, "c");
2241 build
= isl_ast_build_set_iterators(build
, iterators
);
2242 build
= isl_ast_build_set_at_each_domain(build
, &at_domain
, &data
);
2243 build
= isl_ast_build_set_before_each_mark(build
, &before_mark
, &data
);
2244 build
= isl_ast_build_set_after_each_mark(build
, &after_mark
, &data
);
2245 if (gen
->prog
->scop
->options
->debug
->dump_final_schedule
)
2246 isl_schedule_dump(schedule
);
2247 tree
= isl_ast_build_node_from_schedule(build
, schedule
);
2248 isl_ast_build_free(build
);
2253 __isl_give isl_union_map
*extract_sizes_from_str(isl_ctx
*ctx
, const char *str
)
2257 return isl_union_map_read_from_str(ctx
, str
);
2260 /* Can "node" be tiled and then mapped to block and thread identifiers?
2261 * That is, is it permutable with at least one coincident dimension?
2263 static int is_permutable(__isl_keep isl_schedule_node
*node
)
2268 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
2270 if (!isl_schedule_node_band_get_permutable(node
))
2272 if (isl_schedule_node_band_n_member(node
) < 1)
2274 if (!isl_schedule_node_band_member_get_coincident(node
, 0))
2280 /* A isl_schedule_foreach_schedule_node callback
2281 * for setting *any_permutable and aborting the search
2282 * if "node" is a permutable band with coincident dimensions.
2283 * Otherwise, continue searching.
2285 static int set_permutable(__isl_keep isl_schedule_node
*node
, void *user
)
2287 int *any_permutable
= user
;
2290 permutable
= is_permutable(node
);
2296 *any_permutable
= 1;
2301 /* Does "schedule" contain any permutable band with at least one coincident
2304 static int has_any_permutable_node(__isl_keep isl_schedule
*schedule
)
2306 int any_permutable
= 0;
2308 if (isl_schedule_foreach_schedule_node(schedule
, &set_permutable
,
2309 &any_permutable
) < 0 &&
2313 return any_permutable
;
2316 /* Is "node" a leaf or can it be tiled and then mapped to
2317 * block and thread identifiers?
2319 static int is_leaf_or_tilable(__isl_keep isl_schedule_node
*node
)
2321 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
2323 return is_permutable(node
);
2326 /* Is "node" the outermost node in its branch that can be tiled
2327 * and then mapped to block and thread identifiers?
2328 * If there are no such nodes in the branch and if "node" is a leaf,
2329 * then it is accepted too.
2331 static int is_outer_tilable(__isl_keep isl_schedule_node
*node
)
2334 isl_schedule_node
*ancestor
;
2336 tilable
= is_leaf_or_tilable(node
);
2343 ancestor
= isl_schedule_node_copy(node
);
2344 while (isl_schedule_node_has_parent(ancestor
)) {
2345 ancestor
= isl_schedule_node_parent(ancestor
);
2347 tilable
= is_permutable(ancestor
);
2348 if (tilable
< 0 || tilable
)
2352 isl_schedule_node_free(ancestor
);
2353 return tilable
< 0 ? -1 : !tilable
;
2356 /* Collect the references to all writes in "group".
2357 * Each reference is represented by a universe set in a space
2361 * with S[i,j] the statement instance space and R[] the array reference.
2363 static __isl_give isl_union_set
*group_tagged_writes(
2364 struct gpu_array_ref_group
*group
)
2368 isl_union_set
*writes
;
2370 space
= isl_map_get_space(group
->access
);
2371 writes
= isl_union_set_empty(space
);
2372 for (i
= 0; i
< group
->n_ref
; ++i
) {
2376 if (!group
->refs
[i
]->write
)
2379 space
= isl_map_get_space(group
->refs
[i
]->tagged_access
);
2380 space
= isl_space_domain(space
);
2381 writes_i
= isl_set_universe(space
);
2382 writes
= isl_union_set_add_set(writes
, writes_i
);
2388 /* Is there any write access in "group" that requires synchronization
2389 * on a write to global memory?
2390 * We currently take into account all writes that would require
2391 * synchronization at the thread level depth, but if the copying
2392 * for this group is performed at an outer level, then we do not
2393 * actually need to take into account dependences at intermediate levels.
2395 static int any_sync_writes_in_group(struct ppcg_kernel
*kernel
,
2396 struct gpu_array_ref_group
*group
)
2398 isl_union_set
*writes
;
2399 int empty
, disjoint
;
2401 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2407 writes
= group_tagged_writes(group
);
2408 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2409 isl_union_set_free(writes
);
2411 return disjoint
< 0 ? -1 : !disjoint
;
2414 /* Collect the references to all writes in "kernel" that write directly
2415 * to global or shared memory, i.e., that are not mapped to private memory.
2416 * Each reference is represented by a universe set in a space
2420 * with S[i,j] the statement instance space and R[] the array reference.
2422 static __isl_give isl_union_set
*collect_non_private_tagged_writes(
2423 struct ppcg_kernel
*kernel
)
2425 isl_union_set
*writes
;
2428 writes
= isl_union_set_empty(isl_union_set_get_space(kernel
->arrays
));
2430 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2431 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2433 for (j
= 0; j
< array
->n_group
; ++j
) {
2434 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2435 isl_union_set
*writes_ij
;
2439 if (group
->private_tile
)
2441 writes_ij
= group_tagged_writes(group
);
2442 writes
= isl_union_set_union(writes
, writes_ij
);
2449 /* Are there any direct writes to global memory that require
2452 static int any_global_or_shared_sync_writes(struct ppcg_kernel
*kernel
)
2454 isl_union_set
*writes
;
2455 int empty
, disjoint
;
2457 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2463 writes
= collect_non_private_tagged_writes(kernel
);
2464 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2465 isl_union_set_free(writes
);
2467 return disjoint
< 0 ? -1 : !disjoint
;
2470 /* Construct an isl_multi_val for use as tile sizes for tiling "node"
2471 * from the elements in "tile_size".
2473 static __isl_give isl_multi_val
*construct_band_tiles_sizes(
2474 __isl_keep isl_schedule_node
*node
, int *tile_size
)
2484 ctx
= isl_schedule_node_get_ctx(node
);
2485 space
= isl_schedule_node_band_get_space(node
);
2486 n
= isl_schedule_node_band_n_member(node
);
2487 mv
= isl_multi_val_zero(space
);
2488 for (i
= 0; i
< n
; ++i
) {
2491 v
= isl_val_int_from_si(ctx
, tile_size
[i
]);
2492 mv
= isl_multi_val_set_val(mv
, i
, v
);
2498 /* Replace the partial schedule S of the band node "node" by
2506 * if scale_tile_loops is set, with f the integers in "factor".
2507 * The list that "factor" points to is assumed to contain at least
2508 * as many elements as the number of members in the band.
2510 static __isl_give isl_schedule_node
*snap_band_to_sizes(
2511 __isl_take isl_schedule_node
*node
, int *factor
,
2512 struct ppcg_options
*options
)
2516 mv
= construct_band_tiles_sizes(node
, factor
);
2517 node
= isl_schedule_node_band_scale_down(node
, isl_multi_val_copy(mv
));
2518 if (options
->scale_tile_loops
)
2519 node
= isl_schedule_node_band_scale(node
,
2520 isl_multi_val_copy(mv
));
2521 isl_multi_val_free(mv
);
2526 /* Tile "band" with tile size specified by "sizes".
2528 * Since the tile loops will be mapped to block ids, we forcibly
2529 * turn off tile loop scaling. We may want to enable tile loop scaling
2530 * at some later point, but then we would have to support the detection
2531 * of strides during the mapping to block ids.
2532 * Similarly, since the point loops will be mapped to thread ids,
2533 * we forcibly shift the point loops so that they start at zero.
2535 static __isl_give isl_schedule_node
*tile_band(
2536 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2538 isl_ctx
*ctx
= isl_schedule_node_get_ctx(node
);
2542 scale_tile
= isl_options_get_tile_scale_tile_loops(ctx
);
2543 isl_options_set_tile_scale_tile_loops(ctx
, 0);
2544 shift_point
= isl_options_get_tile_shift_point_loops(ctx
);
2545 isl_options_set_tile_shift_point_loops(ctx
, 1);
2547 node
= isl_schedule_node_band_tile(node
, sizes
);
2549 isl_options_set_tile_scale_tile_loops(ctx
, scale_tile
);
2550 isl_options_set_tile_shift_point_loops(ctx
, shift_point
);
2555 /* Extract the set of parameter values and outer schedule dimensions
2556 * for which any statement instance
2557 * in the kernel inserted at "node" needs to be executed.
2558 * Intersect the set of parameter values derived from the host schedule
2559 * relation with the context of "prog".
2561 static __isl_give isl_set
*extract_context(__isl_keep isl_schedule_node
*node
,
2562 struct gpu_prog
*prog
)
2564 isl_union_map
*schedule
;
2565 isl_union_set
*schedule_domain
;
2569 schedule
= isl_schedule_node_get_prefix_schedule_relation(node
);
2570 schedule_domain
= isl_union_map_range(schedule
);
2571 empty
= isl_union_set_is_empty(schedule_domain
);
2573 isl_union_set_free(schedule_domain
);
2580 space
= isl_union_set_get_space(schedule_domain
);
2581 isl_union_set_free(schedule_domain
);
2582 space
= isl_space_set_from_params(space
);
2583 depth
= isl_schedule_node_get_schedule_depth(node
);
2584 space
= isl_space_add_dims(space
, isl_dim_set
, depth
);
2585 context
= isl_set_empty(space
);
2587 context
= isl_set_from_union_set(schedule_domain
);
2589 context
= isl_set_intersect_params(context
,
2590 isl_set_copy(prog
->context
));
2595 /* Return the set of outer array elements accessed by
2596 * by the statement instance in "domain" in "prog".
2598 static __isl_give isl_union_set
*accessed_by_domain(
2599 __isl_take isl_union_set
*domain
, struct gpu_prog
*prog
)
2601 isl_union_map
*access
;
2602 isl_union_set
*arrays
;
2604 access
= isl_union_map_union(isl_union_map_copy(prog
->read
),
2605 isl_union_map_copy(prog
->may_write
));
2606 access
= isl_union_map_intersect_domain(access
, domain
);
2607 arrays
= isl_union_map_range(access
);
2608 arrays
= isl_union_set_apply(arrays
,
2609 isl_union_map_copy(prog
->to_outer
));
2614 /* Return the number of outer band members of the band node "node"
2615 * that are marked coincident.
2617 static int n_outer_coincidence(__isl_keep isl_schedule_node
*node
)
2621 n
= isl_schedule_node_band_n_member(node
);
2623 for (i
= 0; i
< n
; ++i
)
2624 if (!isl_schedule_node_band_member_get_coincident(node
, i
))
2630 /* If the band node "node" has more than "n" members, then split off
2631 * the first "n" of them.
2633 static __isl_give isl_schedule_node
*split_band(
2634 __isl_take isl_schedule_node
*node
, int n
)
2638 dim
= isl_schedule_node_band_n_member(node
);
2640 node
= isl_schedule_node_band_split(node
, n
);
2645 /* Scale a band node that may have been split by split_band.
2646 * "sizes" are the scaling factors for the original node.
2647 * "node" either points to the original band node, or the outer
2648 * of the two pieces after splitting.
2650 * If the number of elements in "node" is smaller than the number of
2651 * elements in "sizes", then some splitting has occurred and we split
2652 * "sizes" in the same way.
2654 static __isl_give isl_schedule_node
*scale_band(
2655 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2659 n
= isl_multi_val_dim(sizes
, isl_dim_set
);
2660 dim
= isl_schedule_node_band_n_member(node
);
2662 isl_multi_val
*sizes2
;
2664 sizes2
= isl_multi_val_copy(sizes
);
2665 sizes
= isl_multi_val_drop_dims(sizes
,
2666 isl_dim_set
, dim
, n
- dim
);
2667 sizes2
= isl_multi_val_drop_dims(sizes2
, isl_dim_set
, 0, dim
);
2668 node
= isl_schedule_node_child(node
, 0);
2669 node
= isl_schedule_node_band_scale(node
, sizes2
);
2670 node
= isl_schedule_node_parent(node
);
2673 return isl_schedule_node_band_scale(node
, sizes
);
2676 /* Return an isl_multi_aff, with as elements the parameters in "space"
2677 * that have the names specified by the elements in "names".
2678 * If (some of) these parameters do not already appear in "space",
2679 * then they are added first.
2681 static __isl_give isl_multi_aff
*parameter_vector(__isl_take isl_space
*space
,
2682 __isl_keep isl_id_list
*names
)
2685 isl_local_space
*ls
;
2689 space
= isl_space_free(space
);
2691 n
= isl_id_list_n_id(names
);
2692 for (i
= 0; i
< n
; ++i
) {
2696 id
= isl_id_list_get_id(names
, i
);
2697 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2702 pos
= isl_space_dim(space
, isl_dim_param
);
2703 space
= isl_space_add_dims(space
, isl_dim_param
, 1);
2704 space
= isl_space_set_dim_id(space
, isl_dim_param
, pos
, id
);
2706 ma
= isl_multi_aff_zero(isl_space_copy(space
));
2707 ls
= isl_local_space_from_space(isl_space_domain(space
));
2708 for (i
= 0; i
< n
; ++i
) {
2713 id
= isl_id_list_get_id(names
, i
);
2714 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2716 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
2717 isl_dim_param
, pos
);
2718 ma
= isl_multi_aff_set_aff(ma
, i
, aff
);
2720 isl_local_space_free(ls
);
2725 /* Return constraints on the domain elements that equate a sequence of
2726 * parameters called "names", to the partial schedule
2727 * of "node" modulo the integers in "size".
2728 * The number of elements in the array "size" should be equal
2729 * to the number of elements in "names".
2730 * The number of members of the band node "node" should be smaller
2731 * than or equal to this number. If it is smaller, then the first
2732 * elements of "names" are equated to zero.
2734 static __isl_give isl_union_set
*set_schedule_modulo(
2735 __isl_keep isl_schedule_node
*node
, __isl_keep isl_id_list
*names
,
2741 isl_multi_union_pw_aff
*mupa
, *mupa2
;
2743 isl_union_set
*domain
;
2747 n
= isl_id_list_n_id(names
);
2749 return isl_schedule_node_get_universe_domain(node
);
2750 n_zero
= n
- isl_schedule_node_band_n_member(node
);
2752 mupa
= isl_schedule_node_band_get_partial_schedule(node
);
2753 mv
= construct_band_tiles_sizes(node
, size
+ n_zero
);
2754 mupa
= isl_multi_union_pw_aff_mod_multi_val(mupa
, mv
);
2756 space
= isl_multi_union_pw_aff_get_space(mupa
);
2757 space
= isl_space_params(space
);
2758 space
= isl_space_set_from_params(space
);
2759 space
= isl_space_add_dims(space
, isl_dim_set
, n_zero
);
2760 ma
= isl_multi_aff_zero(space
);
2762 domain
= isl_schedule_node_get_universe_domain(node
);
2763 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(
2764 isl_union_set_copy(domain
), ma
);
2765 mupa
= isl_multi_union_pw_aff_range_product(mupa2
, mupa
);
2767 space
= isl_multi_union_pw_aff_get_space(mupa
);
2768 ma
= parameter_vector(space
, names
);
2770 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(domain
, ma
);
2771 mupa
= isl_multi_union_pw_aff_sub(mupa
, mupa2
);
2773 return isl_multi_union_pw_aff_zero_union_set(mupa
);
2776 /* Insert a context node at "node" introducing the block and thread
2777 * identifiers along with their bounds, which are stored in kernel->grid_size
2778 * and kernel->block_dim.
2779 * Note that the bounds on the block identifiers may implicitly impose
2780 * constraints on the parameters. A guard needs to be inserted
2781 * in the schedule tree to ensure that those bounds hold at "node".
2782 * This guard is inserted in insert_guard.
2784 static __isl_give isl_schedule_node
*insert_context(struct ppcg_kernel
*kernel
,
2785 __isl_take isl_schedule_node
*node
)
2789 context
= isl_set_universe(isl_set_get_space(kernel
->context
));
2791 context
= add_bounded_parameters_dynamic(context
,
2792 kernel
->grid_size
, kernel
->block_ids
);
2793 context
= add_bounded_parameters(context
,
2794 kernel
->block_dim
, kernel
->thread_ids
);
2796 node
= isl_schedule_node_insert_context(node
, context
);
2801 /* Insert a guard that eliminates kernel launches where the kernel
2802 * obviously does not have any work to do.
2804 * In particular, eliminate kernel launches where there are obviously
2806 * Use the same block size constraints that are used to create the context
2807 * to ensure that all constraints implicit in the constructed context
2808 * are imposed by the guard.
2810 * Additionally, add other constraints that are valid
2811 * for each executed instance ("context"), as long as this does not result
2814 static __isl_give isl_schedule_node
*insert_guard(
2815 __isl_take isl_schedule_node
*node
, __isl_keep isl_set
*context
,
2816 __isl_keep isl_multi_pw_aff
*size
, struct ppcg_scop
*scop
)
2822 guard
= isl_set_copy(context
);
2823 guard
= isl_set_compute_divs(guard
);
2824 guard
= isl_set_from_basic_set(isl_set_simple_hull(guard
));
2826 nparam
= isl_set_dim(guard
, isl_dim_param
);
2827 n
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
2828 ids
= ppcg_scop_generate_names(scop
, n
, "__ppcg_tmp");
2829 guard
= add_bounded_parameters_dynamic(guard
, size
, ids
);
2830 isl_id_list_free(ids
);
2831 guard
= isl_set_project_out(guard
, isl_dim_param
, nparam
, n
);
2833 node
= isl_schedule_node_insert_guard(node
, guard
);
2838 /* Does any array reference group mapping require the band that is mapped
2839 * to threads to be unrolled?
2841 static int kernel_requires_unroll(struct ppcg_kernel
*kernel
)
2845 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2846 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2848 for (j
= 0; j
< array
->n_group
; ++j
) {
2849 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2850 if (gpu_array_ref_group_requires_unroll(group
))
2858 /* Mark the given band node "node" for unrolling by the AST generator and
2859 * then sink it to the leaves of the schedule tree.
2860 * All dimensions of "node" are assumed to be coincident, such that this
2861 * sinking is a valid operation.
2863 static __isl_give isl_schedule_node
*unroll(__isl_take isl_schedule_node
*node
)
2867 n
= isl_schedule_node_band_n_member(node
);
2868 for (i
= 0; i
< n
; ++i
)
2869 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
2870 isl_ast_loop_unroll
);
2872 node
= isl_schedule_node_band_sink(node
);
2877 /* Insert a synchronization node in the schedule tree of "node"
2878 * after the core computation of "kernel" at the level of the band
2879 * that is mapped to threads, except if that level is equal to
2880 * that of the band that is mapped to blocks or if there are no writes
2881 * to global or shared memory in the core computation that require
2883 * If there are any writes to shared memory and the shared memory
2884 * copying is performed at the same level, then synchronization
2885 * is needed between the core and the copying anyway, so we might
2886 * as well add it here. If the copying is performed at a higher
2887 * level, then different iterations of intermediate schedule dimensions
2888 * may have a different mapping from between shared memory elements and
2889 * threads, such that synchronization is required after the core.
2890 * "node" is assumed to point to the kernel node.
2892 static __isl_give isl_schedule_node
*add_sync(struct ppcg_kernel
*kernel
,
2893 __isl_take isl_schedule_node
*node
)
2898 need_sync
= any_global_or_shared_sync_writes(kernel
);
2900 return isl_schedule_node_free(node
);
2904 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
2906 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
2907 if (kernel_depth
== isl_schedule_node_get_schedule_depth(node
))
2908 return gpu_tree_move_up_to_kernel(node
);
2910 node
= gpu_tree_ensure_following_sync(node
, kernel
);
2912 node
= gpu_tree_move_up_to_kernel(node
);
2917 /* Return a read ("read" is 1) or write access relation for "group"
2918 * with those accesses removed that are only needed to communicate data
2919 * within the subtree of the schedule rooted at "node".
2920 * Furthermore, include the prefix schedule at "node".
2921 * That is, return a relation of the form
2925 * with D the outer schedule dimensions at "node".
2927 static __isl_give isl_union_map
*anchored_non_local_accesses(
2928 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2929 __isl_take isl_schedule_node
*node
, int read
)
2931 isl_union_map
*access
;
2932 isl_union_map
*prefix
;
2934 access
= gpu_array_ref_group_access_relation(group
, read
, !read
);
2935 access
= remove_local_accesses_group(kernel
, group
, access
, node
, read
);
2936 prefix
= isl_schedule_node_get_prefix_schedule_relation(node
);
2937 access
= isl_union_map_range_product(prefix
, access
);
2942 /* Given an array reference group "group", create a mapping
2944 * read[D -> A] -> [D -> A]
2946 * if "read" is set or
2948 * write[D -> A] -> [D -> A]
2950 * if "read" is not set.
2951 * D corresponds to the outer group->depth dimensions of
2952 * the kernel schedule.
2954 static __isl_give isl_multi_aff
*create_from_access(isl_ctx
*ctx
,
2955 struct gpu_array_ref_group
*group
, int read
)
2960 space
= isl_space_copy(group
->array
->space
);
2961 space
= isl_space_from_range(space
);
2962 space
= isl_space_add_dims(space
, isl_dim_in
, group
->depth
);
2963 space
= isl_space_wrap(space
);
2964 space
= isl_space_map_from_set(space
);
2966 id
= isl_id_alloc(ctx
, read
? "read" : "write", group
);
2967 space
= isl_space_set_tuple_id(space
, isl_dim_in
, id
);
2969 return isl_multi_aff_identity(space
);
2972 /* If any writes in "group" require synchronization, then make sure
2973 * that there is a synchronization node for "kernel" after the node
2974 * following "node" in a sequence.
2976 * If "shared" is set and no synchronization is needed for
2977 * the writes to global memory, then add synchronization before
2978 * the kernel to protect shared memory from being overwritten
2979 * by the next iteration of the core computation.
2980 * No additional synchronization is needed to protect against
2981 * the next copy into shared memory because each element of
2982 * the shared memory tile is always copied by the same thread.
2984 static __isl_give isl_schedule_node
*add_group_write_sync(
2985 __isl_take isl_schedule_node
*node
, struct ppcg_kernel
*kernel
,
2986 struct gpu_array_ref_group
*group
, int shared
)
2990 need_sync
= any_sync_writes_in_group(kernel
, group
);
2992 return isl_schedule_node_free(node
);
2994 node
= isl_schedule_node_parent(node
);
2995 node
= isl_schedule_node_next_sibling(node
);
2996 node
= isl_schedule_node_child(node
, 0);
2997 node
= gpu_tree_ensure_following_sync(node
, kernel
);
2998 } else if (shared
) {
2999 node
= isl_schedule_node_parent(node
);
3000 node
= isl_schedule_node_parent(node
);
3001 node
= gpu_tree_move_down_to_depth(node
, group
->depth
,
3003 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3009 /* Add copy statements to the schedule tree of "node"
3010 * for reading from global memory to private memory (if "read" is set) or
3011 * for writing back from private memory to global memory
3012 * (if "read" is not set) for the array reference group "group" that
3013 * is mapped to private memory.
3014 * On input, "node" points to the kernel node, and it is moved
3015 * back there on output.
3017 * The copies are performed in the order of the array elements.
3018 * The copy statement instances include a reference to the outer
3019 * group->depth dimensions of the kernel schedule for ease of
3020 * combining them with the group tiling.
3022 * That is, the extra schedule is of the form
3026 * where D corresponds to the outer group->depth dimensions of
3027 * the kernel schedule and A to the global array.
3028 * This schedule is unrolled because registers are not addressable.
3030 * The copying is inserted in the schedule tree through an extension
3035 * where the extra domain elements type[D -> A] are those accessed
3037 * A filter is inserted on type[D -> A] to ensure that the element
3038 * is read/written by the same thread that needs the element.
3039 * This filter is obtained by applying
3043 * to the thread filter for the core statements.
3045 * The extension is inserted before the core computation in case of a read
3046 * and after the core computation in case of a write.
3047 * In the latter case, we also make sure that there is a synchronization
3048 * node after the write to global memory, unless this write is performed
3049 * at the outer level of the kernel.
3050 * In principle, this synchronization could be inserted higher
3051 * in the schedule tree depending on where the corresponding reads
3052 * from global memory are performed.
3054 static __isl_give isl_schedule_node
*add_copies_group_private(
3055 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3056 __isl_take isl_schedule_node
*node
, int read
)
3058 isl_union_map
*access
;
3059 isl_union_map
*prefix
;
3060 isl_union_set
*domain
;
3062 isl_multi_aff
*from_access
;
3063 isl_multi_pw_aff
*mpa
;
3064 isl_multi_union_pw_aff
*mupa
;
3065 isl_schedule_node
*graft
;
3066 isl_union_set
*filter
;
3070 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3071 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3073 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3074 empty
= isl_union_map_is_empty(access
);
3075 if (empty
< 0 || empty
) {
3076 isl_union_map_free(access
);
3078 return isl_schedule_node_free(node
);
3079 return gpu_tree_move_up_to_kernel(node
);
3082 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3083 space
= isl_space_domain(isl_multi_aff_get_space(from_access
));
3084 access
= isl_union_map_preimage_range_multi_aff(access
, from_access
);
3086 filter
= isl_union_set_copy(kernel
->thread_filter
);
3087 filter
= isl_union_set_apply(filter
, isl_union_map_copy(access
));
3088 filter
= isl_union_set_detect_equalities(filter
);
3089 filter
= isl_union_set_coalesce(filter
);
3091 domain
= isl_union_map_range(access
);
3092 access
= isl_union_set_wrapped_domain_map(domain
);
3093 access
= isl_union_map_reverse(access
);
3094 access
= isl_union_map_coalesce(access
);
3095 graft
= isl_schedule_node_from_extension(access
);
3097 space
= isl_space_map_from_set(space
);
3098 mpa
= isl_multi_pw_aff_identity(space
);
3099 mpa
= isl_multi_pw_aff_range_factor_range(mpa
);
3100 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3102 graft
= isl_schedule_node_child(graft
, 0);
3103 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3104 graft
= unroll(graft
);
3106 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3108 graft
= isl_schedule_node_parent(graft
);
3111 node
= isl_schedule_node_graft_before(node
, graft
);
3113 node
= isl_schedule_node_graft_after(node
, graft
);
3114 if (kernel_depth
< group
->depth
)
3115 node
= add_group_write_sync(node
, kernel
, group
, 0);
3118 node
= gpu_tree_move_up_to_kernel(node
);
3123 /* Add copy statements to the schedule tree of "node"
3124 * for reading from global memory to shared memory (if "read" is set) or
3125 * for writing back from shared memory to global memory
3126 * (if "read" is not set) for the array reference group "group" that
3127 * is mapped to shared memory.
3128 * On input, "node" points to the kernel node, and it is moved
3129 * back there on output.
3131 * The copies are performed in the order of the corresponding shared
3133 * The copy statement instances include a reference to the outer
3134 * group->depth dimensions of the kernel schedule for ease of
3135 * combining them with the group tiling.
3137 * If we are performing a read from global memory to shared memory and
3138 * if the array involved is not a scalar, then we copy
3139 * the entire tile to shared memory. This may result in some extra
3140 * elements getting copied, but it should lead to simpler code
3141 * (which means that fewer registers may be needed) and less divergence.
3143 * Otherwise, we only copy the elements that will be read or have been written
3146 * That is, the extra schedule is of the form
3150 * where D corresponds to the outer group->depth dimensions of
3151 * the kernel schedule, A to the global array and T is the corresponding
3152 * shared memory tile.
3154 * The copying is inserted in the schedule tree through an extension
3159 * where the extra domain elements type[D -> A] are those accessed
3160 * by the group. In the case of read from a non-scalar, this set
3161 * is replaced by the entire shared memory tile.
3163 * A filter is inserted on type[D -> A] to map the copy instances
3164 * to the threads. In particular, the thread identifiers are
3165 * equated to the position inside the shared memory tile (T)
3166 * modulo the block size.
3167 * We try to align the innermost tile dimension with the innermost
3168 * thread identifier (x) as a heuristic to improve coalescing.
3169 * In particular, if the dimension of the tile is greater than
3170 * the dimension of the block, then the schedule mapping to the tile
3171 * is broken up into two pieces and the filter is applied to the inner part.
3172 * If, on the other hand, the dimension of the tile is smaller than
3173 * the dimension of the block, then the initial thread identifiers
3174 * are equated to zero and the remaining thread identifiers are
3175 * matched to the memory tile.
3177 * The extension is inserted before the core computation in case of a read
3178 * and after the core computation in case of a write.
3179 * In the case of a read, we first need to make sure there is some
3180 * synchronization before the core computation such that we can put the read
3181 * from global memory to shared memory before that synchronization.
3182 * This ensures that all threads have finished copying into shared memory
3183 * before the shared memory is used.
3184 * We also need to make sure that there is a synchronization node after
3185 * the core computation to ensure that the next load into shared memory
3186 * only happens after all data has been used. There is no need for
3187 * this synchronization if we are at the outer level since then there
3188 * won't be a next load.
3189 * In the case of a write, we need to make sure there is some synchronization
3190 * after the core computation such taht we can put the write from shared
3191 * memory to global memory after that synchronization.
3192 * Unless we are at the outer level, we also need a synchronization node
3193 * after the write to ensure the data is saved to global memory
3194 * before the next iteration write to the same shared memory.
3195 * It also makes sure the data has arrived in global memory before
3196 * it is read in a subsequent iteration.
3198 static __isl_give isl_schedule_node
*add_copies_group_shared(
3199 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3200 __isl_take isl_schedule_node
*node
, int read
)
3202 struct gpu_array_tile
*tile
;
3203 isl_union_map
*access
;
3204 isl_union_set
*domain
;
3205 isl_union_set
*sync
;
3207 isl_multi_aff
*from_access
;
3208 isl_multi_pw_aff
*mpa
;
3209 isl_multi_union_pw_aff
*mupa
;
3210 isl_schedule_node
*graft
;
3211 isl_union_set
*filter
;
3216 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3217 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3219 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3220 empty
= isl_union_map_is_empty(access
);
3221 if (empty
< 0 || empty
) {
3222 isl_union_map_free(access
);
3224 return isl_schedule_node_free(node
);
3225 return gpu_tree_move_up_to_kernel(node
);
3228 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3230 tile
= gpu_array_ref_group_tile(group
);
3231 ma
= isl_multi_aff_copy(tile
->tiling
);
3232 ma
= isl_multi_aff_pullback_multi_aff(ma
,
3233 isl_multi_aff_copy(from_access
));
3234 mpa
= isl_multi_pw_aff_from_multi_aff(ma
);
3235 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3237 domain
= isl_union_map_range(access
);
3239 if (read
&& !gpu_array_is_scalar(group
->array
)) {
3241 isl_union_set_free(domain
);
3242 map
= group_tile(group
);
3243 domain
= isl_union_set_from_set(isl_map_wrap(map
));
3246 domain
= isl_union_set_preimage_multi_aff(domain
, from_access
);
3247 access
= isl_union_set_wrapped_domain_map(domain
);
3248 access
= isl_union_map_reverse(access
);
3249 access
= isl_union_map_coalesce(access
);
3250 graft
= isl_schedule_node_from_extension(access
);
3252 graft
= isl_schedule_node_child(graft
, 0);
3254 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3256 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0) {
3257 graft
= isl_schedule_node_band_split(graft
,
3258 tile
->n
- kernel
->n_block
);
3259 graft
= isl_schedule_node_child(graft
, 0);
3261 if (tile
->n
< kernel
->n_block
)
3262 skip
= kernel
->n_block
- tile
->n
;
3265 filter
= set_schedule_modulo(graft
, kernel
->thread_ids
,
3267 if (!kernel
->options
->wrap
)
3268 graft
= snap_band_to_sizes(graft
, kernel
->block_dim
+ skip
,
3270 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0)
3271 graft
= isl_schedule_node_parent(graft
);
3272 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3274 while (graft
&& isl_schedule_node_has_parent(graft
))
3275 graft
= isl_schedule_node_parent(graft
);
3278 if (kernel_depth
< group
->depth
)
3279 node
= gpu_tree_ensure_sync_after_core(node
, kernel
);
3280 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3281 node
= isl_schedule_node_graft_before(node
, graft
);
3283 node
= gpu_tree_move_right_to_sync(node
, kernel
);
3284 node
= isl_schedule_node_graft_after(node
, graft
);
3285 if (kernel_depth
< group
->depth
)
3286 node
= add_group_write_sync(node
, kernel
, group
, 1);
3289 node
= gpu_tree_move_up_to_kernel(node
);
3294 /* Check whether the array reference group "group" is mapped to
3295 * private or shared memory and, if so,
3296 * add copy statements to the schedule tree of "node"
3297 * for reading from global memory to private or shared memory
3298 * (if "read" is set) or for writing back from private or shared memory
3299 * to global memory (if "read" is not set) for this group.
3300 * On input, "node" points to the kernel node, and it is moved
3301 * back there on output.
3303 static __isl_give isl_schedule_node
*add_copies_group(
3304 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3305 __isl_take isl_schedule_node
*node
, int read
)
3307 if (group
->private_tile
)
3308 return add_copies_group_private(kernel
, group
, node
, read
);
3309 if (group
->shared_tile
)
3310 return add_copies_group_shared(kernel
, group
, node
, read
);
3314 /* For each array reference group that is mapped to private or shared memory,
3315 * add copy statements to the schedule tree of "node"
3316 * for reading from global memory to private or shared memory
3317 * and for writing back.
3318 * On input, "node" points to the kernel node, and it is moved
3319 * back there on output.
3321 static __isl_give isl_schedule_node
*add_copies(struct ppcg_kernel
*kernel
,
3322 __isl_take isl_schedule_node
*node
)
3326 for (i
= 0; i
< kernel
->n_array
; ++i
) {
3327 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
3329 for (j
= 0; j
< array
->n_group
; ++j
) {
3330 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3332 node
= add_copies_group(kernel
, group
, node
, 1);
3335 node
= add_copies_group(kernel
, group
, node
, 0);
3344 /* Mark all dimensions in the current band node atomic.
3346 static __isl_give isl_schedule_node
*atomic(__isl_take isl_schedule_node
*node
)
3350 n
= isl_schedule_node_band_n_member(node
);
3351 for (i
= 0; i
< n
; ++i
)
3352 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
3353 isl_ast_loop_atomic
);
3358 /* Mark "node" atomic, if it is a band node.
3359 * Do the same for all ancestors.
3360 * Return a pointer to "node" (in the updated schedule tree).
3362 static __isl_give isl_schedule_node
*atomic_ancestors(
3363 __isl_take isl_schedule_node
*node
)
3369 if (!isl_schedule_node_has_parent(node
))
3372 pos
= isl_schedule_node_get_child_position(node
);
3373 node
= isl_schedule_node_parent(node
);
3374 if (isl_schedule_node_get_type(node
) == isl_schedule_node_band
)
3375 node
= atomic(node
);
3376 node
= atomic_ancestors(node
);
3377 node
= isl_schedule_node_child(node
, pos
);
3382 /* Collect all write references that require synchronization.
3383 * "node" is assumed to point to the kernel node.
3384 * Each reference is represented by a universe set in a space
3388 * with S[i,j] the statement instance space and R[] the array reference.
3390 * This function should be called before block and thread filters are added.
3392 * Synchronization is needed after a write if there is a subsequent read
3393 * within the same block that may not be performed by the same thread.
3394 * There should not be any dependences between different blocks,
3395 * so we start with the flow dependences within the same kernel invocation
3396 * and we subtract from these those dependences that are mapped
3397 * to the same iteration of the bands where synchronization is inserted.
3398 * We do not remove pairs of instances that are known to map to
3399 * the same thread across different iterations of the intermediate
3400 * bands because the read may be performed by a different thread
3401 * than the one that needs the value if shared memory is involved.
3403 * We also consider all pairs of possible writes that access the same
3404 * memory location and that may be mapped to the same block but not
3405 * to the same iteration of the intermediate bands.
3406 * In theory, it would be possible for one thread to still be in
3407 * a previous iteration of a loop in these bands.
3408 * A write to global memory in this delayed thread could then overwrite
3409 * a write from another thread that has already moved on to
3410 * the next iteration.
3412 * After computing the above writes paired off with reads or writes
3413 * that depend on them, we project onto the domain writes.
3414 * Sychronization is needed after writes to global memory
3415 * through these references.
3417 static __isl_give isl_union_set
*compute_sync_writes(
3418 struct ppcg_kernel
*kernel
, __isl_keep isl_schedule_node
*node
)
3420 isl_union_map
*local
;
3421 isl_union_map
*may_writes
, *shared_access
;
3422 isl_union_map
*kernel_prefix
, *thread_prefix
;
3423 isl_union_map
*equal
;
3424 isl_union_set
*wrap
;
3425 isl_union_set
*domain
;
3427 domain
= isl_schedule_node_get_universe_domain(node
);
3428 kernel_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3429 node
= isl_schedule_node_copy(node
);
3430 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3431 thread_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3432 isl_schedule_node_free(node
);
3434 may_writes
= isl_union_map_copy(kernel
->prog
->scop
->tagged_may_writes
);
3435 may_writes
= isl_union_map_curry(may_writes
);
3436 may_writes
= isl_union_map_intersect_domain(may_writes
, domain
);
3437 may_writes
= isl_union_map_uncurry(may_writes
);
3438 shared_access
= isl_union_map_copy(may_writes
);
3439 shared_access
= isl_union_map_apply_range(shared_access
,
3440 isl_union_map_reverse(may_writes
));
3442 local
= isl_union_map_copy(kernel
->prog
->scop
->tagged_dep_flow
);
3443 local
= isl_union_map_union(local
, shared_access
);
3444 local
= isl_union_map_zip(local
);
3446 equal
= isl_union_map_apply_range(kernel_prefix
,
3447 isl_union_map_reverse(isl_union_map_copy(kernel_prefix
)));
3448 wrap
= isl_union_map_wrap(equal
);
3449 local
= isl_union_map_intersect_domain(local
, wrap
);
3450 equal
= isl_union_map_apply_range(thread_prefix
,
3451 isl_union_map_reverse(isl_union_map_copy(thread_prefix
)));
3452 wrap
= isl_union_map_wrap(equal
);
3453 local
= isl_union_map_subtract_domain(local
, wrap
);
3455 local
= isl_union_map_zip(local
);
3456 local
= isl_union_map_universe(local
);
3458 return isl_union_map_domain(local
);
3461 /* Group the domain elements into a single space, named kernelX,
3462 * with X the kernel sequence number "kernel_id".
3464 static __isl_give isl_schedule_node
*group_statements(
3465 __isl_take isl_schedule_node
*node
, int kernel_id
)
3473 snprintf(buffer
, sizeof(buffer
), "kernel%d", kernel_id
);
3474 id
= isl_id_alloc(isl_schedule_node_get_ctx(node
), buffer
, NULL
);
3475 return isl_schedule_node_group(node
, id
);
3478 /* Create a ppcg_kernel representing the domain instances that reach "node"
3479 * and insert a mark node pointing to the ppcg_kernel before "node".
3480 * The band that "node" points to is the band that needs to be mapped
3481 * to block identifiers. The band that needs to be mapped to thread
3482 * identifiers should be marked by a "thread" mark by the caller.
3483 * This mark is removed by this function.
3484 * If "scale" is set, then the band that "node" points to is scaled
3487 * Mark all outer band nodes as atomic to ensure each kernel is only
3489 * If the domain elements that reach "node" live in more than one space,
3490 * then group the domain elements into a single space, named kernelX,
3491 * with X the kernel sequence number.
3493 * Insert a guard node governing the kernel node to ensure that
3494 * no kernels with zero blocks are launched.
3496 * Insert a context node describing the block and thread
3497 * identifiers inside the kernel mark.
3498 * The context node needs to be inserted after the effective block size
3499 * has been determined such that the bounds on the thread identifiers
3500 * would reflect the effective block size.
3501 * Insert a filter node inside the context node mapping the statement
3502 * instances to block identifiers. In particular, the block identifiers
3503 * are equated to the partial schedule of band that was marked for mapping
3504 * to blocks modulo the grid size.
3505 * Insert a filter node inside the "thread" mark mapping the statement
3506 * instances to thread identifiers. In particular, the thread identifiers
3507 * are equated to the partial schedule of band that was marked for mapping
3508 * to threads modulo the block size.
3510 * Compute array reference groups for all arrays, set the local
3511 * array bounds based on the set of domain instances that reach
3512 * the kernel node, check the total amount of shared memory used
3513 * and compute all group tilings.
3514 * The array reference groups are computed after the block filter
3515 * has been inserted because it affects the mapping to shared or
3516 * private memory. This computation also requires the thread filter
3517 * (in the ppcg_kernel object), but this thread filter should not
3518 * have been added to the schedule tree yet since the computation
3519 * requires the schedule of the band that needs to be mapped to
3520 * threads before the privatization is applied.
3522 * If any array reference group requires the band mapped to threads
3523 * to be unrolled, then we perform the required unrolling.
3525 * We save a copy of the schedule that may influence the mappings
3526 * to shared or private memory in kernel->shared_schedule.
3528 * Finally, we add synchronization and copy statements to the schedule tree,
3529 * remove the "thread" mark and create representations for the local
3530 * variables in the kernel.
3532 * We keep a copy of the isl_id that points to the kernel to ensure
3533 * that the kernel does not get destroyed if the schedule node
3534 * is freed due to some error condition.
3536 static __isl_give isl_schedule_node
*create_kernel(struct gpu_gen
*gen
,
3537 __isl_take isl_schedule_node
*node
, int scale
,
3538 __isl_keep isl_multi_val
*sizes
)
3540 struct ppcg_kernel
*kernel
;
3542 isl_schedule_node
*node_thread
;
3543 isl_union_map
*host_schedule
;
3544 isl_set
*host_domain
;
3545 isl_union_set
*domain
;
3546 int single_statement
;
3548 kernel
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel
);
3549 kernel
= ppcg_kernel_create_local_arrays(kernel
, gen
->prog
);
3551 return isl_schedule_node_free(node
);
3553 domain
= isl_schedule_node_get_domain(node
);
3554 single_statement
= isl_union_set_n_set(domain
) == 1;
3556 kernel
->ctx
= gen
->ctx
;
3557 kernel
->prog
= gen
->prog
;
3558 kernel
->options
= gen
->options
;
3559 kernel
->context
= extract_context(node
, gen
->prog
);
3560 kernel
->core
= isl_union_set_universe(isl_union_set_copy(domain
));
3561 kernel
->arrays
= accessed_by_domain(isl_union_set_copy(domain
),
3563 kernel
->n_grid
= n_outer_coincidence(node
);
3564 node_thread
= isl_schedule_node_copy(node
);
3565 node_thread
= gpu_tree_move_down_to_thread(node_thread
, kernel
->core
);
3566 node_thread
= isl_schedule_node_child(node_thread
, 0);
3567 kernel
->n_block
= n_outer_coincidence(node_thread
);
3568 isl_schedule_node_free(node_thread
);
3569 kernel
->id
= gen
->kernel_id
++;
3570 read_grid_and_block_sizes(kernel
, gen
);
3572 kernel
->sync_writes
= compute_sync_writes(kernel
, node
);
3574 host_schedule
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3575 host_domain
= isl_set_from_union_set(isl_union_map_range(
3578 node
= atomic_ancestors(node
);
3580 id
= isl_id_alloc(gen
->ctx
, "kernel", kernel
);
3581 id
= isl_id_set_free_user(id
, &ppcg_kernel_free_wrap
);
3582 node
= isl_schedule_node_insert_mark(node
, isl_id_copy(id
));
3584 if (!single_statement
)
3585 node
= group_statements(node
, kernel
->id
);
3587 node
= isl_schedule_node_child(node
, 0);
3588 node
= split_band(node
, kernel
->n_grid
);
3589 kernel
->block_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3590 kernel
->n_grid
, "b");
3591 kernel
->block_filter
= set_schedule_modulo(node
, kernel
->block_ids
,
3593 kernel
->grid_size
= extract_grid_size(kernel
,
3594 isl_union_set_copy(domain
));
3595 if (!kernel
->options
->wrap
)
3596 node
= snap_band_to_sizes(node
, kernel
->grid_dim
,
3599 node
= scale_band(node
, isl_multi_val_copy(sizes
));
3600 node
= isl_schedule_node_parent(node
);
3601 if (!single_statement
)
3602 node
= isl_schedule_node_parent(node
);
3603 node
= insert_guard(node
, kernel
->context
, kernel
->grid_size
,
3605 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3606 node
= isl_schedule_node_child(node
, 0);
3607 node
= split_band(node
, kernel
->n_block
);
3608 kernel
->thread_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3609 kernel
->n_block
, "t");
3610 kernel
->thread_filter
= set_schedule_modulo(node
, kernel
->thread_ids
,
3612 extract_block_size(kernel
, domain
);
3614 node
= gpu_tree_move_up_to_kernel(node
);
3615 node
= isl_schedule_node_child(node
, 0);
3616 node
= insert_context(kernel
, node
);
3617 node
= isl_schedule_node_child(node
, 0);
3618 node
= isl_schedule_node_insert_filter(node
,
3619 isl_union_set_copy(kernel
->block_filter
));
3621 node
= gpu_tree_move_up_to_kernel(node
);
3623 if (gpu_group_references(kernel
, node
) < 0)
3624 node
= isl_schedule_node_free(node
);
3625 localize_bounds(kernel
, host_domain
);
3626 isl_set_free(host_domain
);
3628 check_shared_memory_bound(kernel
);
3629 compute_group_tilings(kernel
);
3631 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3632 node
= isl_schedule_node_child(node
, 0);
3633 if (!kernel
->options
->wrap
)
3634 node
= snap_band_to_sizes(node
, kernel
->block_dim
,
3636 node
= isl_schedule_node_insert_filter(node
,
3637 isl_union_set_copy(kernel
->thread_filter
));
3638 if (kernel_requires_unroll(kernel
)) {
3639 node
= isl_schedule_node_child(node
, 0);
3640 node
= unroll(node
);
3643 node
= gpu_tree_move_up_to_thread(node
);
3644 kernel
->shared_schedule_dim
=
3645 isl_schedule_node_get_schedule_depth(node
);
3646 kernel
->shared_schedule
=
3647 isl_schedule_node_get_prefix_schedule_union_pw_multi_aff(node
);
3649 node
= gpu_tree_move_up_to_kernel(node
);
3651 node
= add_sync(kernel
, node
);
3652 node
= add_copies(kernel
, node
);
3654 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3655 node
= isl_schedule_node_delete(node
);
3657 node
= gpu_tree_move_up_to_kernel(node
);
3659 if (create_kernel_vars(kernel
) < 0)
3660 node
= isl_schedule_node_free(node
);
3662 if (!single_statement
)
3663 node
= isl_schedule_node_parent(node
);
3664 node
= isl_schedule_node_parent(node
);
3670 /* Insert a zero-dimensional permutable band at "node".
3672 static __isl_give isl_schedule_node
*insert_empty_permutable_band(
3673 __isl_take isl_schedule_node
*node
)
3676 isl_schedule
*schedule
;
3677 isl_union_set
*domain
;
3678 isl_multi_union_pw_aff
*mupa
;
3680 schedule
= isl_schedule_node_get_schedule(node
);
3681 domain
= isl_schedule_get_domain(schedule
);
3682 space
= isl_union_set_get_space(domain
);
3683 isl_union_set_free(domain
);
3684 isl_schedule_free(schedule
);
3686 space
= isl_space_set_from_params(space
);
3687 mupa
= isl_multi_union_pw_aff_zero(space
);
3688 node
= isl_schedule_node_insert_partial_schedule(node
, mupa
);
3689 node
= isl_schedule_node_band_set_permutable(node
, 1);
3694 /* If "node" is the outermost permutable band that can be mapped to block and
3695 * thread identifiers in its branch (or a leaf with no such outer bands),
3696 * then mark the band as such, attaching a ppcg_kernel to the mark.
3698 * If "node" originally points to a leaf, then insert a zero-dimensional
3699 * permutable band such that we can assume that "node" always
3700 * points to a band node.
3702 * Tile "node" using user specified tile sizes, after splitting the band
3703 * if the number of specified tile sizes is smaller than the dimension
3704 * of the band. Mark the point band of this tiling as the band that
3705 * needs to be mapped to threads.
3706 * Create a kernel representing the domain instances that reach "node" and
3707 * insert a mark node pointing to the ppcg_kernel before the band node.
3709 static __isl_give isl_schedule_node
*mark_outer_permutable(
3710 __isl_take isl_schedule_node
*node
, void *user
)
3712 struct gpu_gen
*gen
= user
;
3718 isl_multi_val
*sizes
;
3720 outer
= is_outer_tilable(node
);
3722 return isl_schedule_node_free(node
);
3726 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
3727 node
= insert_empty_permutable_band(node
);
3729 tile_len
= isl_schedule_node_band_n_member(node
);
3730 tile_size
= read_tile_sizes(gen
, &tile_len
);
3732 return isl_schedule_node_free(node
);
3733 if (tile_len
< isl_schedule_node_band_n_member(node
))
3734 node
= isl_schedule_node_band_split(node
, tile_len
);
3735 sizes
= construct_band_tiles_sizes(node
, tile_size
);
3736 node
= tile_band(node
, isl_multi_val_copy(sizes
));
3737 node
= isl_schedule_node_child(node
, 0);
3738 id
= isl_id_alloc(gen
->ctx
, "thread", NULL
);
3739 node
= isl_schedule_node_insert_mark(node
, id
);
3740 node
= isl_schedule_node_parent(node
);
3742 scale
= gen
->options
->scale_tile_loops
;
3743 node
= create_kernel(gen
, node
, scale
, sizes
);
3744 isl_multi_val_free(sizes
);
3750 /* Insert "kernel" marks that point to a ppcg_kernel structure
3751 * in front of all outermost tilable band that (by construction)
3752 * have at least one parallel loop.
3754 static __isl_give isl_schedule_node
*mark_kernels(struct gpu_gen
*gen
,
3755 __isl_take isl_schedule_node
*node
)
3757 return isl_schedule_node_map_descendant(node
,
3758 &mark_outer_permutable
, gen
);
3761 /* Save the schedule "schedule" to a file called "filename".
3762 * The schedule is printed in block style.
3764 static void save_schedule(__isl_keep isl_schedule
*schedule
,
3765 const char *filename
)
3774 file
= fopen(filename
, "w");
3776 fprintf(stderr
, "Unable to open '%s' for writing\n", filename
);
3779 ctx
= isl_schedule_get_ctx(schedule
);
3780 p
= isl_printer_to_file(ctx
, file
);
3781 p
= isl_printer_set_yaml_style(p
, ISL_YAML_STYLE_BLOCK
);
3782 p
= isl_printer_print_schedule(p
, schedule
);
3783 isl_printer_free(p
);
3787 /* Load and return a schedule from a file called "filename".
3789 static __isl_give isl_schedule
*load_schedule(isl_ctx
*ctx
,
3790 const char *filename
)
3793 isl_schedule
*schedule
;
3795 file
= fopen(filename
, "r");
3797 fprintf(stderr
, "Unable to open '%s' for reading\n", filename
);
3800 schedule
= isl_schedule_read_from_file(ctx
, file
);
3806 /* Compute an appropriate schedule based on the accesses in
3807 * gen->read and gen->write.
3809 * We use the dependences in gen->prog->scop to compute
3810 * a schedule that has a parallel loop in each tilable band and
3811 * return this schedule.
3813 * If live range reordering is allowed, then we need to make sure
3814 * that live ranges on arrays are not run in parallel since doing
3815 * so would require array expansion. We therefore add the array
3816 * order dependences to the coincidence dependences. Non-zero array
3817 * order dependences will then prevent a schedule dimension from being
3818 * considered parallel.
3819 * Live ranges derived from scalars are allowed to be run in parallel
3820 * since we force the scalars to be mapped to private memory in
3821 * check_scalar_live_ranges.
3822 * If live range reordering is allowed, then the false dependences
3823 * are not added to the validity constraints as that would prevent
3824 * reordering. Instead, the external false dependences that enforce that reads
3825 * from potentially live-in data precede any later write and
3826 * that writes of potentially live-out data follow any other earlier write
3827 * are added to the validity and the coincidence constraints.
3828 * The false dependences are still added to the proximity constraints
3829 * for consistency with the case where live range reordering is not allowed.
3830 * The coincidence constraints then consist of flow dependences,
3831 * external false dependences and array order dependences.
3832 * The independences can be filtered out from the first two sets.
3833 * They have already been filtered out from the array order dependences
3834 * on a per array basis in collect_order_dependences.
3835 * There is no need for a per array handling of the other two sets
3836 * as there should be no flow or external false dependence on local
3837 * variables that can be filtered out.
3839 static __isl_give isl_schedule
*compute_schedule(struct gpu_gen
*gen
)
3841 isl_union_set
*domain
;
3842 isl_union_map
*dep_raw
, *dep
;
3843 isl_union_map
*validity
, *proximity
, *coincidence
;
3844 isl_schedule_constraints
*sc
;
3845 isl_schedule
*schedule
;
3847 domain
= isl_union_set_copy(gen
->prog
->scop
->domain
);
3848 sc
= isl_schedule_constraints_on_domain(domain
);
3849 sc
= isl_schedule_constraints_set_context(sc
,
3850 isl_set_copy(gen
->prog
->scop
->context
));
3851 if (gen
->options
->live_range_reordering
) {
3852 sc
= isl_schedule_constraints_set_conditional_validity(sc
,
3853 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_flow
),
3854 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_order
));
3855 proximity
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
3856 validity
= isl_union_map_copy(proximity
);
3857 validity
= isl_union_map_union(validity
,
3858 isl_union_map_copy(gen
->prog
->scop
->dep_forced
));
3859 proximity
= isl_union_map_union(proximity
,
3860 isl_union_map_copy(gen
->prog
->scop
->dep_false
));
3861 coincidence
= isl_union_map_copy(validity
);
3862 coincidence
= isl_union_map_subtract(coincidence
,
3863 isl_union_map_copy(gen
->prog
->scop
->independence
));
3864 coincidence
= isl_union_map_union(coincidence
,
3865 isl_union_map_copy(gen
->prog
->array_order
));
3867 dep_raw
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
3868 dep
= isl_union_map_copy(gen
->prog
->scop
->dep_false
);
3869 dep
= isl_union_map_union(dep
, dep_raw
);
3870 dep
= isl_union_map_coalesce(dep
);
3871 proximity
= isl_union_map_copy(dep
);
3872 coincidence
= isl_union_map_copy(dep
);
3875 sc
= isl_schedule_constraints_set_validity(sc
, validity
);
3876 sc
= isl_schedule_constraints_set_coincidence(sc
, coincidence
);
3877 sc
= isl_schedule_constraints_set_proximity(sc
, proximity
);
3879 if (gen
->options
->debug
->dump_schedule_constraints
)
3880 isl_schedule_constraints_dump(sc
);
3881 schedule
= isl_schedule_constraints_compute_schedule(sc
);
3886 /* Obtain a schedule for the scop, either by reading it from
3887 * a file or by computing one.
3889 static __isl_give isl_schedule
*get_schedule(struct gpu_gen
*gen
)
3891 isl_schedule
*schedule
;
3893 if (gen
->options
->load_schedule_file
) {
3894 schedule
= load_schedule(gen
->ctx
,
3895 gen
->options
->load_schedule_file
);
3897 schedule
= compute_schedule(gen
);
3898 if (gen
->options
->save_schedule_file
)
3899 save_schedule(schedule
,
3900 gen
->options
->save_schedule_file
);
3902 if (gen
->options
->debug
->dump_schedule
)
3903 isl_schedule_dump(schedule
);
3908 /* Compute the sets of outer array elements that need to be copied in and out.
3910 * In particular, for each array that is possibly written anywhere in
3911 * "prog" and that is visible outside the corresponding scop,
3912 * we copy out its entire extent.
3914 * Any array elements that is read without first being written needs
3915 * to be copied in. Furthermore, if there are any array elements that
3916 * are copied out, but that may not be written inside "prog", then
3917 * they also need to be copied in to ensure that the value after execution
3918 * is the same as the value before execution, at least for those array
3919 * elements that may have their values preserved by the scop.
3920 * In case the array elements are structures, we need to take into
3921 * account that all members of the structures need to be written
3922 * by "prog" before we can avoid copying the data structure in.
3924 * While computing the set of array elements that are copied out but
3925 * not necessarily written, we intersect both sets with the context.
3926 * This helps in those cases where the arrays are declared with a fixed size,
3927 * while the accesses are parametric and the context assigns a fixed value
3928 * to the parameters.
3930 * If an element from a local array is read without first being written,
3931 * then there is no point in copying it in since it cannot have been
3932 * written prior to the scop. Warn about the uninitialized read instead.
3934 static void compute_copy_in_and_out(struct gpu_prog
*prog
)
3937 isl_union_set
*local
;
3938 isl_union_set
*may_write
, *must_write
;
3939 isl_union_set
*copy_in
, *copy_out
;
3940 isl_union_set
*not_written
;
3941 isl_union_map
*uninitialized
;
3942 isl_union_map
*local_uninitialized
;
3944 must_write
= isl_union_map_range(
3945 isl_union_map_copy(prog
->must_write
));
3946 must_write
= isl_union_set_intersect_params(must_write
,
3947 isl_set_copy(prog
->context
));
3948 may_write
= isl_union_map_range(
3949 isl_union_map_copy(prog
->may_write
));
3950 may_write
= isl_union_set_intersect_params(may_write
,
3951 isl_set_copy(prog
->context
));
3952 may_write
= isl_union_set_universe(may_write
);
3953 may_write
= isl_union_set_apply(may_write
,
3954 isl_union_map_copy(prog
->to_outer
));
3955 copy_out
= isl_union_set_empty(isl_union_set_get_space(may_write
));
3956 local
= isl_union_set_copy(copy_out
);
3958 for (i
= 0; i
< prog
->n_array
; ++i
) {
3963 space
= isl_space_copy(prog
->array
[i
].space
);
3965 if (prog
->array
[i
].local
) {
3968 set
= isl_set_universe(space
);
3969 local
= isl_union_set_add_set(local
, set
);
3973 write_i
= isl_union_set_extract_set(may_write
, space
);
3974 empty
= isl_set_plain_is_empty(write_i
);
3975 isl_set_free(write_i
);
3979 write_i
= isl_set_copy(prog
->array
[i
].extent
);
3980 copy_out
= isl_union_set_add_set(copy_out
, write_i
);
3982 isl_union_set_free(may_write
);
3984 copy_out
= isl_union_set_intersect_params(copy_out
,
3985 isl_set_copy(prog
->context
));
3987 prog
->copy_out
= isl_union_set_copy(copy_out
);
3989 copy_out
= isl_union_set_apply(copy_out
,
3990 isl_union_map_copy(prog
->to_inner
));
3991 copy_out
= isl_union_set_intersect(copy_out
,
3992 isl_union_set_copy(prog
->may_persist
));
3993 not_written
= isl_union_set_subtract(copy_out
, must_write
);
3995 uninitialized
= isl_union_map_copy(prog
->scop
->live_in
);
3996 local_uninitialized
= isl_union_map_copy(uninitialized
);
3998 local
= isl_union_set_apply(local
, isl_union_map_copy(prog
->to_inner
));
3999 local_uninitialized
= isl_union_map_intersect_range(local_uninitialized
,
4001 if (!isl_union_map_is_empty(local_uninitialized
)) {
4003 "possibly uninitialized reads (not copied in):\n");
4004 isl_union_map_dump(local_uninitialized
);
4006 uninitialized
= isl_union_map_subtract(uninitialized
,
4007 local_uninitialized
);
4008 copy_in
= isl_union_map_range(uninitialized
);
4009 copy_in
= isl_union_set_union(copy_in
, not_written
);
4010 copy_in
= isl_union_set_apply(copy_in
,
4011 isl_union_map_copy(prog
->to_outer
));
4013 prog
->copy_in
= copy_in
;
4016 /* Update "schedule" for mapping to a GPU device.
4018 * In particular, insert a context node and create kernels for
4019 * each outermost tilable band.
4021 static __isl_give isl_schedule
*map_to_device(struct gpu_gen
*gen
,
4022 __isl_take isl_schedule
*schedule
)
4024 isl_schedule_node
*node
;
4027 context
= isl_set_copy(gen
->prog
->context
);
4028 context
= isl_set_from_params(context
);
4029 schedule
= isl_schedule_insert_context(schedule
, context
);
4031 node
= isl_schedule_get_root(schedule
);
4032 isl_schedule_free(schedule
);
4033 node
= isl_schedule_node_child(node
, 0);
4034 if (isl_schedule_node_get_type(node
) == isl_schedule_node_context
)
4035 node
= isl_schedule_node_child(node
, 0);
4036 node
= mark_kernels(gen
, node
);
4037 schedule
= isl_schedule_node_get_schedule(node
);
4038 isl_schedule_node_free(node
);
4043 /* Internal data structure for extract_access.
4044 * "next_access" points to the end of a linked list that is extended
4045 * by extract_access.
4046 * "single_expression" is set if the access expressions belong to
4047 * an expression statement (i.e., a statement without internal control).
4048 * "any_to_outer" maps all intermediate arrays to their outer arrays.
4050 struct ppcg_extract_access_data
{
4051 struct gpu_stmt_access
**next_access
;
4052 int single_expression
;
4053 isl_union_map
*any_to_outer
;
4056 /* Given a tagged access relation to a single array "tagged", extract it
4057 * as a map, taking into account that the input may be empty.
4058 * If the access relation is empty, then it does not contain
4059 * any space information, so we try to recover it from the index
4061 * The space of the index expression is of the form I -> A,
4062 * with I the statement instances and A the array, or [I -> F] -> A,
4063 * with F the filters corresponding to arguments.
4064 * We first drop F, if present, obtaining I -> A.
4065 * Then we construct I -> R, with R the reference tag,
4066 * combine the two into I -> [R -> A] and uncurry to obtain
4067 * the final result [I -> R] -> A.
4068 * Note that the index expression may have a lower dimension
4069 * than that of the array, but this dimension is not used
4070 * if the access relation is empty.
4072 static __isl_give isl_map
*extract_single_tagged_access(
4073 __isl_take isl_union_map
*tagged
, __isl_keep pet_expr
*expr
)
4077 isl_space
*space
, *space2
;
4078 isl_multi_pw_aff
*index
;
4080 empty
= isl_union_map_is_empty(tagged
);
4084 return isl_map_from_union_map(tagged
);
4085 isl_union_map_free(tagged
);
4087 index
= pet_expr_access_get_index(expr
);
4088 space
= isl_multi_pw_aff_get_space(index
);
4089 isl_multi_pw_aff_free(index
);
4090 if (isl_space_domain_is_wrapping(space
))
4091 space
= isl_space_domain_factor_domain(space
);
4092 space2
= isl_space_copy(space
);
4093 space2
= isl_space_from_domain(isl_space_domain(space
));
4094 id
= pet_expr_access_get_ref_id(expr
);
4095 space2
= isl_space_set_tuple_id(space2
, isl_dim_out
, id
);
4096 space
= isl_space_range_product(space2
, space
);
4097 space
= isl_space_uncurry(space
);
4099 return isl_map_empty(space
);
4101 isl_union_map_free(tagged
);
4105 /* Extract a gpu_stmt_access from "expr", append it to the list
4106 * that ends in *data->next_access and update the end of the list.
4107 * If the access expression performs a write, then it is considered
4108 * exact only if it appears in a single expression statement and
4109 * if its may access relation is equal to its must access relation.
4111 * The combined set of may accesses may be union if member accesses
4112 * are involved, but the entire set is derived from a single reference and
4113 * therefore from a single index expression. These accesses therefore
4114 * all map to the same outer array.
4116 static int extract_access(__isl_keep pet_expr
*expr
, void *user
)
4118 struct ppcg_extract_access_data
*data
= user
;
4119 isl_union_map
*tagged
;
4120 struct gpu_stmt_access
*access
;
4121 isl_ctx
*ctx
= pet_expr_get_ctx(expr
);
4122 isl_multi_pw_aff
*index
;
4124 access
= isl_alloc_type(ctx
, struct gpu_stmt_access
);
4126 access
->next
= NULL
;
4127 access
->read
= pet_expr_access_is_read(expr
);
4128 access
->write
= pet_expr_access_is_write(expr
);
4129 tagged
= pet_expr_access_get_tagged_may_read(expr
);
4130 tagged
= isl_union_map_union(tagged
,
4131 pet_expr_access_get_tagged_may_write(expr
));
4132 tagged
= isl_union_map_apply_range(tagged
,
4133 isl_union_map_copy(data
->any_to_outer
));
4134 if (!access
->write
) {
4135 access
->exact_write
= 1;
4136 } else if (!data
->single_expression
) {
4137 access
->exact_write
= 0;
4139 isl_union_map
*must
, *may
;
4140 may
= isl_union_map_copy(tagged
);
4141 may
= isl_union_map_domain_factor_domain(may
);
4142 must
= pet_expr_access_get_must_write(expr
);
4143 access
->exact_write
= isl_union_map_is_equal(must
, may
);
4144 isl_union_map_free(must
);
4145 isl_union_map_free(may
);
4147 index
= pet_expr_access_get_index(expr
);
4148 access
->n_index
= isl_multi_pw_aff_dim(index
, isl_dim_out
);
4149 isl_multi_pw_aff_free(index
);
4150 access
->ref_id
= pet_expr_access_get_ref_id(expr
);
4151 access
->tagged_access
= extract_single_tagged_access(tagged
, expr
);
4152 access
->access
= isl_map_copy(access
->tagged_access
);
4153 access
->access
= isl_map_domain_factor_domain(access
->access
);
4155 *data
->next_access
= access
;
4156 data
->next_access
= &(*data
->next_access
)->next
;
4158 if (!access
->access
)
4164 /* Construct a linked list of gpu_stmt_access objects,
4165 * one for each access expression in the statement body.
4166 * "any_to_outer" maps all intermediate arrays to their outer arrays.
4168 static int pet_stmt_extract_accesses(struct gpu_stmt
*stmt
,
4169 __isl_keep isl_union_map
*any_to_outer
)
4171 struct ppcg_extract_access_data data
;
4173 stmt
->accesses
= NULL
;
4174 data
.next_access
= &stmt
->accesses
;
4175 data
.single_expression
=
4176 pet_tree_get_type(stmt
->stmt
->body
) == pet_tree_expr
;
4177 data
.any_to_outer
= any_to_outer
;
4178 return pet_tree_foreach_access_expr(stmt
->stmt
->body
,
4179 &extract_access
, &data
);
4182 /* Return an array of gpu_stmt representing the statements in "scop".
4184 static struct gpu_stmt
*extract_stmts(isl_ctx
*ctx
, struct ppcg_scop
*scop
,
4185 __isl_keep isl_set
*context
, __isl_keep isl_union_map
*any_to_outer
)
4188 struct gpu_stmt
*stmts
;
4190 stmts
= isl_calloc_array(ctx
, struct gpu_stmt
, scop
->pet
->n_stmt
);
4194 for (i
= 0; i
< scop
->pet
->n_stmt
; ++i
) {
4195 struct gpu_stmt
*s
= &stmts
[i
];
4197 s
->id
= isl_set_get_tuple_id(scop
->pet
->stmts
[i
]->domain
);
4198 s
->stmt
= scop
->pet
->stmts
[i
];
4199 if (pet_stmt_extract_accesses(s
, any_to_outer
) < 0)
4200 return free_stmts(stmts
, i
+ 1);
4206 /* Callback for ppcg_print_guarded that calls the callback for generate_gpu.
4208 static __isl_give isl_printer
*print_gpu(__isl_take isl_printer
*p
, void *user
)
4210 struct gpu_gen
*gen
= user
;
4212 return gen
->print(p
, gen
->prog
, gen
->tree
, &gen
->types
,
4216 /* Generate CUDA code for "scop" and print it to "p".
4217 * After generating an AST for the transformed scop as explained below,
4218 * we call "gen->print" to print the AST in the desired output format
4221 * If it turns out that it does not make sense to generate GPU code,
4222 * then we generate CPU code instead.
4224 * The GPU code is generated in a context where at least one
4225 * statement instance is executed. The corresponding guard (if any) is printed
4226 * around the entire generated GPU code, except for the declaration
4227 * of the arrays that are visible outside of the scop and that therefore
4228 * cannot be declared inside the body of any possible guard.
4230 * We first compute a schedule that respects the dependences
4231 * of the original program and select the outermost bands
4232 * of tilable dimensions that have at least one parallel loop.
4233 * If the --load-schedule is specified, then the loaded schedule
4234 * is used instead of a computed schedule.
4236 * Each of these bands B is then tiled according to "tile" sizes, resulting
4237 * in two nested bands, with a kernel marker on top
4245 * We then split off at most 2 parallel dimensions from the T band and
4246 * at most 3 parallel dimension from the P band
4259 * A filter is introduced in front of T1 that maps the domain instances
4260 * to block identifiers. Similarly, a filter is introduced in front of P1
4261 * that maps the domain instances to thread identifiers.
4263 * For each iteration of the T2 band and for each array, we compute
4264 * the array elements accessed by that iteration, construct a rectangular
4265 * box around it and shift it to the origin. The result is used
4266 * as shared memory for the array.
4268 * Copying and synchronization statements are added to this schedule tree.
4269 * In principle, these are added in front of the P1 band, but some of
4270 * them may get hoisted up to higher levels.
4272 * The entire AST is then generated from the single resulting schedule tree.
4273 * During the generation the subtrees at kernel nodes (K) are saved
4274 * aside and replaced by kernel calls. The result is printed as host code
4275 * while the saved subtrees are printed as device code.
4277 static __isl_give isl_printer
*generate(__isl_take isl_printer
*p
,
4278 struct gpu_gen
*gen
, struct ppcg_scop
*scop
,
4279 struct ppcg_options
*options
)
4281 struct gpu_prog
*prog
;
4283 isl_set
*context
, *guard
;
4284 isl_schedule
*schedule
;
4288 return isl_printer_free(p
);
4290 ctx
= isl_printer_get_ctx(p
);
4291 prog
= gpu_prog_alloc(ctx
, scop
);
4293 return isl_printer_free(p
);
4295 context
= isl_set_copy(prog
->context
);
4296 guard
= isl_union_set_params(isl_union_set_copy(prog
->scop
->domain
));
4297 prog
->context
= isl_set_intersect(prog
->context
, isl_set_copy(guard
));
4300 schedule
= get_schedule(gen
);
4302 any_permutable
= has_any_permutable_node(schedule
);
4303 if (any_permutable
< 0 || !any_permutable
) {
4304 isl_set_free(context
);
4305 isl_set_free(guard
);
4306 if (any_permutable
< 0)
4307 p
= isl_printer_free(p
);
4309 p
= print_cpu(p
, scop
, options
);
4310 isl_schedule_free(schedule
);
4312 compute_copy_in_and_out(prog
);
4313 schedule
= map_to_device(gen
, schedule
);
4314 gen
->tree
= generate_code(gen
, schedule
);
4315 p
= ppcg_print_exposed_declarations(p
, prog
->scop
);
4316 p
= ppcg_print_guarded(p
, guard
, context
, &print_gpu
, gen
);
4317 isl_ast_node_free(gen
->tree
);
4320 gpu_prog_free(prog
);
4325 /* Wrapper around generate for use as a ppcg_transform callback.
4327 static __isl_give isl_printer
*generate_wrap(__isl_take isl_printer
*p
,
4328 struct ppcg_scop
*scop
, void *user
)
4330 struct gpu_gen
*gen
= user
;
4332 return generate(p
, gen
, scop
, gen
->options
);
4335 /* Transform the code in the file called "input" by replacing
4336 * all scops by corresponding GPU code and write the results to "out".
4338 int generate_gpu(isl_ctx
*ctx
, const char *input
, FILE *out
,
4339 struct ppcg_options
*options
,
4340 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
4341 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
4342 struct gpu_types
*types
, void *user
), void *user
)
4349 gen
.sizes
= extract_sizes_from_str(ctx
, options
->sizes
);
4350 gen
.options
= options
;
4353 gen
.print_user
= user
;
4355 gen
.types
.name
= NULL
;
4357 if (options
->debug
->dump_sizes
) {
4358 isl_space
*space
= isl_space_params_alloc(ctx
, 0);
4359 gen
.used_sizes
= isl_union_map_empty(space
);
4362 r
= ppcg_transform(ctx
, input
, out
, options
, &generate_wrap
, &gen
);
4364 if (options
->debug
->dump_sizes
) {
4365 isl_union_map_dump(gen
.used_sizes
);
4366 isl_union_map_free(gen
.used_sizes
);
4369 isl_union_map_free(gen
.sizes
);
4370 for (i
= 0; i
< gen
.types
.n
; ++i
)
4371 free(gen
.types
.name
[i
]);
4372 free(gen
.types
.name
);
4377 /* Compute the set of inner array elements that may have their values
4378 * preserved by "prog". In particular, collect the array elements of
4379 * arrays that are not local to "prog" and remove those elements that
4380 * are definitely killed or definitely written by "prog".
4382 static __isl_give isl_union_set
*compute_may_persist(struct gpu_prog
*prog
)
4385 isl_union_set
*may_persist
, *killed
;
4386 isl_union_map
*must_kill
;
4388 may_persist
= isl_union_set_empty(isl_set_get_space(prog
->context
));
4389 for (i
= 0; i
< prog
->n_array
; ++i
) {
4392 if (prog
->array
[i
].local
)
4395 extent
= isl_set_copy(prog
->array
[i
].extent
);
4396 may_persist
= isl_union_set_add_set(may_persist
, extent
);
4399 may_persist
= isl_union_set_intersect_params(may_persist
,
4400 isl_set_copy(prog
->context
));
4401 may_persist
= isl_union_set_apply(may_persist
,
4402 isl_union_map_copy(prog
->to_inner
));
4403 must_kill
= isl_union_map_copy(prog
->tagged_must_kill
);
4404 killed
= isl_union_map_range(must_kill
);
4405 must_kill
= isl_union_map_copy(prog
->must_write
);
4406 killed
= isl_union_set_union(killed
, isl_union_map_range(must_kill
));
4408 may_persist
= isl_union_set_subtract(may_persist
, killed
);
4412 struct gpu_prog
*gpu_prog_alloc(isl_ctx
*ctx
, struct ppcg_scop
*scop
)
4414 struct gpu_prog
*prog
;
4421 prog
= isl_calloc_type(ctx
, struct gpu_prog
);
4426 prog
->context
= isl_set_copy(scop
->context
);
4427 prog
->n_stmts
= scop
->pet
->n_stmt
;
4428 prog
->any_to_outer
= pet_scop_compute_outer_to_any(scop
->pet
);
4429 prog
->any_to_outer
= isl_union_map_reverse(prog
->any_to_outer
);
4430 space
= isl_union_map_get_space(prog
->any_to_outer
);
4431 space
= isl_space_set_from_params(space
);
4432 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
4433 space
= isl_space_map_from_set(space
);
4434 id
= isl_map_identity(space
);
4435 prog
->any_to_outer
= isl_union_map_add_map(prog
->any_to_outer
, id
);
4436 prog
->stmts
= extract_stmts(ctx
, scop
,
4437 prog
->context
, prog
->any_to_outer
);
4438 prog
->read
= isl_union_map_copy(scop
->reads
);
4439 prog
->may_write
= isl_union_map_copy(scop
->may_writes
);
4440 prog
->must_write
= isl_union_map_copy(scop
->must_writes
);
4441 prog
->tagged_must_kill
= isl_union_map_copy(scop
->tagged_must_kills
);
4442 prog
->to_inner
= pet_scop_compute_outer_to_inner(scop
->pet
);
4443 prog
->to_outer
= isl_union_map_copy(prog
->to_inner
);
4444 prog
->to_outer
= isl_union_map_reverse(prog
->to_outer
);
4447 return gpu_prog_free(prog
);
4449 if (collect_array_info(prog
) < 0)
4450 return gpu_prog_free(prog
);
4451 prog
->may_persist
= compute_may_persist(prog
);
4456 void *gpu_prog_free(struct gpu_prog
*prog
)
4460 free_array_info(prog
);
4461 free_stmts(prog
->stmts
, prog
->n_stmts
);
4462 isl_union_map_free(prog
->any_to_outer
);
4463 isl_union_map_free(prog
->to_outer
);
4464 isl_union_map_free(prog
->to_inner
);
4465 isl_union_set_free(prog
->copy_in
);
4466 isl_union_set_free(prog
->copy_out
);
4467 isl_union_map_free(prog
->read
);
4468 isl_union_map_free(prog
->may_write
);
4469 isl_union_map_free(prog
->must_write
);
4470 isl_union_map_free(prog
->tagged_must_kill
);
4471 isl_union_map_free(prog
->array_order
);
4472 isl_union_set_free(prog
->may_persist
);
4473 isl_set_free(prog
->context
);