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
);
1118 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1119 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1121 for (j
= 0; j
< array
->n_group
; ++j
)
1122 gpu_array_ref_group_free(array
->groups
[j
]);
1123 free(array
->groups
);
1125 isl_pw_aff_list_free(array
->bound
);
1127 free(kernel
->array
);
1129 for (i
= 0; i
< kernel
->n_var
; ++i
) {
1130 free(kernel
->var
[i
].name
);
1131 isl_vec_free(kernel
->var
[i
].size
);
1140 /* Wrapper around ppcg_kernel_free for use as a isl_id_set_free_user callback.
1142 static void ppcg_kernel_free_wrap(void *user
)
1144 struct ppcg_kernel
*kernel
= user
;
1146 ppcg_kernel_free(kernel
);
1149 static void create_kernel_var(isl_ctx
*ctx
, struct gpu_array_ref_group
*group
,
1150 struct ppcg_kernel_var
*var
)
1153 struct gpu_array_tile
*tile
;
1157 var
->array
= group
->array
;
1159 tile
= group
->private_tile
;
1160 var
->type
= ppcg_access_private
;
1162 tile
= group
->shared_tile
;
1163 var
->type
= ppcg_access_shared
;
1166 p
= isl_printer_to_str(ctx
);
1167 p
= gpu_array_ref_group_print_name(group
, p
);
1168 var
->name
= isl_printer_get_str(p
);
1169 isl_printer_free(p
);
1171 var
->size
= isl_vec_alloc(ctx
, group
->array
->n_index
);
1173 for (j
= 0; j
< group
->array
->n_index
; ++j
)
1174 var
->size
= isl_vec_set_element_val(var
->size
, j
,
1175 isl_val_copy(tile
->bound
[j
].size
));
1178 static int create_kernel_vars(struct ppcg_kernel
*kernel
)
1183 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1184 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1186 for (j
= 0; j
< array
->n_group
; ++j
) {
1187 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1188 if (group
->private_tile
|| group
->shared_tile
)
1194 kernel
->var
= isl_calloc_array(kernel
->ctx
, struct ppcg_kernel_var
, n
);
1199 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1200 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1202 for (j
= 0; j
< array
->n_group
; ++j
) {
1203 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1204 if (!group
->private_tile
&& !group
->shared_tile
)
1206 create_kernel_var(kernel
->ctx
, group
, &kernel
->var
[n
]);
1214 /* Replace "pa" by the zero function defined over the universe domain
1215 * in the space of "pa".
1217 static __isl_give isl_pw_aff
*set_universally_zero(__isl_take isl_pw_aff
*pa
)
1222 space
= isl_space_domain(isl_pw_aff_get_space(pa
));
1223 isl_pw_aff_free(pa
);
1224 zero
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1226 return isl_pw_aff_from_aff(zero
);
1229 /* The sizes of the arrays on the host that have been computed by
1230 * extract_array_info may depend on the parameters. Use the extra
1231 * constraints on the parameters that are valid at "host_domain"
1232 * to simplify these expressions and store the results in kernel->array.
1234 * We only need these localized bounds for arrays that are accessed
1235 * by the current kernel. If we have found at least one reference group
1236 * then the array is accessed by the kernel. If the array has compound
1237 * elements then we skipped the construction of array reference groups.
1239 * The resulting sizes may be functions that are nowhere defined
1240 * in case the access function cannot possibly access anything inside
1241 * the kernel for some reason. If so, they are replaced by the zero
1242 * function. Since the access function cannot actually access anything,
1243 * there is no harm in printing the array sizes as zero.
1245 static void localize_bounds(struct ppcg_kernel
*kernel
,
1246 __isl_keep isl_set
*host_domain
)
1251 context
= isl_set_copy(host_domain
);
1252 context
= isl_set_params(context
);
1254 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1255 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
1256 isl_pw_aff_list
*bound
;
1259 if (local
->n_group
== 0 && !local
->array
->has_compound_element
)
1262 n_index
= local
->array
->n_index
;
1263 bound
= isl_pw_aff_list_alloc(kernel
->ctx
, n_index
);
1265 for (j
= 0; j
< n_index
; ++j
) {
1269 pwaff
= isl_pw_aff_copy(local
->array
->bound
[j
]);
1270 pwaff
= isl_pw_aff_gist(pwaff
, isl_set_copy(context
));
1271 empty
= isl_pw_aff_is_empty(pwaff
);
1273 pwaff
= isl_pw_aff_free(pwaff
);
1275 pwaff
= set_universally_zero(pwaff
);
1276 bound
= isl_pw_aff_list_add(bound
, pwaff
);
1279 local
->n_index
= n_index
;
1280 local
->bound
= bound
;
1282 isl_set_free(context
);
1285 /* Create the array of gpu_local_array_info structures "array"
1286 * inside "kernel". The number of elements in this array is
1287 * the same as the number of arrays in "prog".
1288 * Initialize the "array" field of each local array to point
1289 * to the corresponding array in "prog".
1291 static struct ppcg_kernel
*ppcg_kernel_create_local_arrays(
1292 struct ppcg_kernel
*kernel
, struct gpu_prog
*prog
)
1297 ctx
= isl_set_get_ctx(prog
->context
);
1298 kernel
->array
= isl_calloc_array(ctx
,
1299 struct gpu_local_array_info
, prog
->n_array
);
1301 return ppcg_kernel_free(kernel
);
1302 kernel
->n_array
= prog
->n_array
;
1304 for (i
= 0; i
< prog
->n_array
; ++i
)
1305 kernel
->array
[i
].array
= &prog
->array
[i
];
1310 /* Find the element in gen->stmt that has the given "id".
1311 * Return NULL if no such gpu_stmt can be found.
1313 static struct gpu_stmt
*find_stmt(struct gpu_prog
*prog
, __isl_keep isl_id
*id
)
1317 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
1318 if (id
== prog
->stmts
[i
].id
)
1322 return i
< prog
->n_stmts
? &prog
->stmts
[i
] : NULL
;
1325 void ppcg_kernel_stmt_free(void *user
)
1328 struct ppcg_kernel_stmt
*stmt
= user
;
1333 switch (stmt
->type
) {
1334 case ppcg_kernel_copy
:
1335 isl_ast_expr_free(stmt
->u
.c
.index
);
1336 isl_ast_expr_free(stmt
->u
.c
.local_index
);
1338 case ppcg_kernel_domain
:
1339 isl_id_to_ast_expr_free(stmt
->u
.d
.ref2expr
);
1341 case ppcg_kernel_sync
:
1348 /* Return the gpu_stmt_access in the list "accesses" that corresponds
1351 static struct gpu_stmt_access
*find_access(struct gpu_stmt_access
*accesses
,
1352 __isl_keep isl_id
*ref_id
)
1354 struct gpu_stmt_access
*access
;
1356 for (access
= accesses
; access
; access
= access
->next
)
1357 if (access
->ref_id
== ref_id
)
1363 /* Return the index of the array called "name" in the list of arrays.
1365 static int find_array_index(struct ppcg_kernel
*kernel
, const char *name
)
1369 for (i
= 0; i
< kernel
->n_array
; ++i
)
1370 if (!strcmp(name
, kernel
->array
[i
].array
->name
))
1376 /* Internal data structure for the index and AST expression transformation
1377 * callbacks for pet_stmt_build_ast_exprs.
1379 * "accesses" is the list of gpu_stmt_access in the statement.
1380 * "iterator_map" expresses the statement iterators in terms of
1381 * the AST loop iterators.
1382 * "sched2shared" expresses the outer shared_schedule_dim dimensions of
1383 * the kernel schedule in terms of the AST loop iterators.
1385 * The following fields are set in transform_index and used in transform_expr.
1386 * "array" is the array that is being accessed.
1387 * "global" is set if the global array is accessed (rather than
1388 * shared/private memory).
1389 * "local_array" refers to information on the array specialized
1390 * to the current kernel.
1392 struct ppcg_transform_data
{
1393 struct gpu_gen
*gen
;
1394 struct gpu_stmt_access
*accesses
;
1395 isl_pw_multi_aff
*iterator_map
;
1396 isl_pw_multi_aff
*sched2shared
;
1398 struct gpu_array_info
*array
;
1400 struct gpu_local_array_info
*local_array
;
1403 /* Return the name of the outer array (of structs) accessed by "access".
1405 static const char *get_outer_array_name(__isl_keep isl_map
*access
)
1410 space
= isl_space_range(isl_map_get_space(access
));
1411 while (space
&& isl_space_is_wrapping(space
))
1412 space
= isl_space_domain(isl_space_unwrap(space
));
1413 name
= isl_space_get_tuple_name(space
, isl_dim_set
);
1414 isl_space_free(space
);
1419 /* Return a pointer to the gpu_array_ref_group in "local"
1420 * that contains the reference "access".
1421 * Return NULL if no such group can be found.
1423 static struct gpu_array_ref_group
*find_ref_group(
1424 struct gpu_local_array_info
*local
, struct gpu_stmt_access
*access
)
1428 for (i
= 0; i
< local
->n_group
; ++i
) {
1429 struct gpu_array_ref_group
*group
= local
->groups
[i
];
1431 for (j
= 0; j
< group
->n_ref
; ++j
)
1432 if (group
->refs
[j
] == access
)
1439 /* Index transformation callback for pet_stmt_build_ast_exprs.
1441 * "index" expresses the array indices in terms of statement iterators
1443 * We first reformulate "index" in terms of the AST loop iterators.
1444 * Then we check if we are accessing the global array or
1445 * a shared/private copy. In the former case, we simply return
1446 * the updated index. If "index" is an affine expression rather
1447 * than an array access, then we also return the updated index here.
1449 * If no reference groups have been computed for the array,
1450 * then we can only be accessing the global array.
1452 * Otherwise, we apply the tiling to the index.
1453 * This tiling is of the form
1457 * where D corresponds to the outer group->depth dimensions of
1458 * the kernel schedule.
1459 * The index is of the form
1463 * We update the tiling to refer to the AST loop iterators
1467 * and modify index to keep track of those iterators
1471 * Combining these two yields a tiled index expression in terms
1472 * of the AST loop iterators
1476 static __isl_give isl_multi_pw_aff
*transform_index(
1477 __isl_take isl_multi_pw_aff
*index
, __isl_keep isl_id
*ref_id
,
1480 struct ppcg_transform_data
*data
= user
;
1481 struct gpu_stmt_access
*access
;
1482 struct gpu_array_ref_group
*group
;
1483 struct gpu_array_tile
*tile
;
1484 isl_pw_multi_aff
*iterator_map
;
1489 isl_multi_pw_aff
*tiling
;
1490 isl_pw_multi_aff
*pma
;
1491 isl_multi_pw_aff
*mpa
;
1492 isl_pw_multi_aff
*sched2depth
;
1496 iterator_map
= isl_pw_multi_aff_copy(data
->iterator_map
);
1497 index
= isl_multi_pw_aff_pullback_pw_multi_aff(index
, iterator_map
);
1499 access
= find_access(data
->accesses
, ref_id
);
1502 if (!isl_map_has_tuple_name(access
->access
, isl_dim_out
))
1505 name
= get_outer_array_name(access
->access
);
1506 i
= find_array_index(data
->gen
->kernel
, name
);
1508 isl_die(isl_multi_pw_aff_get_ctx(index
), isl_error_internal
,
1509 "cannot find array",
1510 return isl_multi_pw_aff_free(index
));
1511 data
->array
= &data
->gen
->prog
->array
[i
];
1512 data
->local_array
= &data
->gen
->kernel
->array
[i
];
1514 group
= find_ref_group(data
->local_array
, access
);
1520 tile
= group
->private_tile
;
1522 tile
= group
->shared_tile
;
1523 data
->global
= !tile
;
1527 space
= isl_space_range(isl_multi_pw_aff_get_space(index
));
1528 space
= isl_space_map_from_set(space
);
1529 pma
= isl_pw_multi_aff_identity(space
);
1530 sched2depth
= isl_pw_multi_aff_copy(data
->sched2shared
);
1531 dim
= isl_pw_multi_aff_dim(sched2depth
, isl_dim_out
);
1532 sched2depth
= isl_pw_multi_aff_drop_dims(sched2depth
, isl_dim_out
,
1533 group
->depth
, dim
- group
->depth
);
1534 pma
= isl_pw_multi_aff_product(sched2depth
, pma
);
1535 tiling
= isl_multi_pw_aff_from_multi_aff(
1536 isl_multi_aff_copy(tile
->tiling
));
1537 tiling
= isl_multi_pw_aff_pullback_pw_multi_aff(tiling
, pma
);
1539 space
= isl_space_domain(isl_multi_pw_aff_get_space(index
));
1540 space
= isl_space_map_from_set(space
);
1541 mpa
= isl_multi_pw_aff_identity(space
);
1542 index
= isl_multi_pw_aff_range_product(mpa
, index
);
1543 index
= isl_multi_pw_aff_pullback_multi_pw_aff(tiling
, index
);
1548 /* Dereference "expr" by adding an index [0].
1549 * The original "expr" is assumed not to have any indices.
1551 * If "expr" is a member access, then the dereferencing needs
1552 * to be applied to the structure argument of this member access.
1554 static __isl_give isl_ast_expr
*dereference(__isl_take isl_ast_expr
*expr
)
1557 isl_ast_expr
*arg0
, *res
;
1558 isl_ast_expr_list
*list
;
1560 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1562 return isl_ast_expr_free(expr
);
1563 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1564 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1567 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1568 arg
= dereference(arg
);
1569 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1570 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1574 isl_ast_expr_free(arg0
);
1576 ctx
= isl_ast_expr_get_ctx(expr
);
1577 res
= isl_ast_expr_from_val(isl_val_zero(ctx
));
1578 list
= isl_ast_expr_list_from_ast_expr(res
);
1579 res
= isl_ast_expr_get_op_arg(expr
, 0);
1580 res
= isl_ast_expr_access(res
, list
);
1581 isl_ast_expr_free(expr
);
1586 /* Linearize the index expression "expr" based on the array bounds
1589 * That is, transform expression
1591 * A[i_0][i_1]...[i_n]
1595 * A[(..((i_0 * b_1 + i_1) ... ) * b_n + i_n]
1597 * where b_0, b_1, ..., b_n are the bounds on the array.
1599 * If the base of "expr" is a member access, then the linearization needs
1600 * to be applied to the structure argument of this member access.
1602 * In the base case, if "expr" has no arguments (other than the name of
1603 * the array), then we are passing an entire array to a function.
1604 * In this case, there is nothing to linearize.
1605 * Note that at this point an expression with no arguments can
1606 * only be an entire array because the scalar case and
1607 * the case of single struct are handled by the caller.
1609 * If the number of specified index expressions in "expr"
1610 * is smaller than the dimension of the accessed array,
1611 * then the missing i_j also do not appear in the linearized expression.
1612 * Furthermore, since such an expression does not refer to a single
1613 * element while the default linearized expression would refer to
1614 * a single element, we return the expression
1616 * A + (..((i_0 * b_1 + i_1) ... ) * b_n]
1618 * instead. Note that because of the special case handling above,
1619 * we can assume here that here that there is at least one index expression.
1621 __isl_give isl_ast_expr
*gpu_local_array_info_linearize_index(
1622 struct gpu_local_array_info
*array
, __isl_take isl_ast_expr
*expr
)
1629 isl_ast_expr_list
*list
;
1630 isl_ast_build
*build
;
1632 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1633 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1634 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1637 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1638 arg
= gpu_local_array_info_linearize_index(array
, arg
);
1639 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1640 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1644 isl_ast_expr_free(arg0
);
1646 if (isl_ast_expr_get_op_n_arg(expr
) == 1)
1649 ctx
= isl_ast_expr_get_ctx(expr
);
1650 context
= isl_set_universe(isl_space_params_alloc(ctx
, 0));
1651 build
= isl_ast_build_from_context(context
);
1653 n
= isl_ast_expr_get_op_n_arg(expr
);
1654 res
= isl_ast_expr_get_op_arg(expr
, 1);
1655 for (i
= 1; i
< array
->n_index
; ++i
) {
1656 isl_pw_aff
*bound_i
;
1657 isl_ast_expr
*expr_i
;
1659 bound_i
= isl_pw_aff_list_get_pw_aff(array
->bound
, i
);
1660 expr_i
= isl_ast_build_expr_from_pw_aff(build
, bound_i
);
1661 res
= isl_ast_expr_mul(res
, expr_i
);
1665 expr_i
= isl_ast_expr_get_op_arg(expr
, i
+ 1);
1666 res
= isl_ast_expr_add(res
, expr_i
);
1669 isl_ast_build_free(build
);
1671 if (1 + array
->n_index
> n
) {
1672 res
= isl_ast_expr_add(isl_ast_expr_get_op_arg(expr
, 0), res
);
1674 list
= isl_ast_expr_list_from_ast_expr(res
);
1675 res
= isl_ast_expr_get_op_arg(expr
, 0);
1676 res
= isl_ast_expr_access(res
, list
);
1679 isl_ast_expr_free(expr
);
1684 /* AST expression transformation callback for pet_stmt_build_ast_exprs.
1686 * If the AST expression refers to an array that is not accessed
1687 * at all, then this means the value of the expression is not used,
1688 * so we might as well print zero (NULL pointer) instead.
1690 * If the AST expression refers to a global scalar that is not
1691 * a read-only scalar, then its address was passed to the kernel and
1692 * we need to dereference it.
1694 * If the AST expression refers to an access to a global array,
1695 * then we linearize the access exploiting the bounds in data->local_array.
1697 static __isl_give isl_ast_expr
*transform_expr(__isl_take isl_ast_expr
*expr
,
1698 __isl_keep isl_id
*id
, void *user
)
1700 struct ppcg_transform_data
*data
= user
;
1704 if (!data
->array
->accessed
) {
1707 ctx
= isl_ast_expr_get_ctx(expr
);
1708 isl_ast_expr_free(expr
);
1709 return isl_ast_expr_from_val(isl_val_zero(ctx
));
1711 if (gpu_array_is_read_only_scalar(data
->array
))
1715 if (data
->array
->n_index
== 0)
1716 return dereference(expr
);
1717 if (!data
->array
->linearize
)
1720 return gpu_local_array_info_linearize_index(data
->local_array
, expr
);
1723 /* This function is called for each instance of a user statement
1724 * in the kernel, identified by "gpu_stmt".
1726 * We attach a struct ppcg_kernel_stmt to the "node", containing
1727 * a computed AST expression for each access.
1728 * These AST expressions are computed from iterator_map,
1729 * which expresses the domain
1730 * elements in terms of the generated loops, and sched2shared,
1731 * which expresses the outer shared_schedule_dim dimensions of
1732 * the kernel schedule computed by PPCG in terms of the generated loops.
1734 static __isl_give isl_ast_node
*create_domain_leaf(struct gpu_gen
*gen
,
1735 __isl_take isl_ast_node
*node
, __isl_keep isl_ast_build
*build
,
1736 struct gpu_stmt
*gpu_stmt
)
1738 struct ppcg_transform_data data
;
1739 struct ppcg_kernel_stmt
*stmt
;
1741 isl_pw_multi_aff
*sched2shared
;
1743 isl_pw_multi_aff
*iterator_map
;
1744 isl_union_map
*schedule
;
1746 stmt
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel_stmt
);
1748 return isl_ast_node_free(node
);
1750 schedule
= isl_ast_build_get_schedule(build
);
1751 map
= isl_map_reverse(isl_map_from_union_map(schedule
));
1752 iterator_map
= isl_pw_multi_aff_from_map(map
);
1753 sched2shared
= compute_sched_to_shared(gen
->kernel
,
1754 isl_pw_multi_aff_copy(iterator_map
));
1756 stmt
->type
= ppcg_kernel_domain
;
1757 stmt
->u
.d
.stmt
= gpu_stmt
;
1760 data
.accesses
= stmt
->u
.d
.stmt
->accesses
;
1761 data
.iterator_map
= iterator_map
;
1762 data
.sched2shared
= sched2shared
;
1763 stmt
->u
.d
.ref2expr
= pet_stmt_build_ast_exprs(stmt
->u
.d
.stmt
->stmt
,
1764 build
, &transform_index
, &data
,
1765 &transform_expr
, &data
);
1767 isl_pw_multi_aff_free(iterator_map
);
1768 isl_pw_multi_aff_free(sched2shared
);
1770 id
= isl_id_alloc(gen
->ctx
, NULL
, stmt
);
1771 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1772 return isl_ast_node_set_annotation(node
, id
);
1775 /* This function is called for each statement node in the AST
1776 * for copying to or from shared/private memory.
1777 * Attach a pointer to a ppcg_kernel_stmt representing the copy
1778 * statement to the node.
1779 * The statement name is "read" or "write", depending on whether we are
1780 * reading from global memory or writing to global memory.
1782 * The schedule is of the form
1786 * where D corresponds to the outer group->depth dimensions of
1787 * the kernel schedule, A to the global array and L to the outer
1788 * generated AST schedule.
1789 * We compute the inverse and strip off the type, resulting in
1793 * We combine this mapping with on the one hand the projection
1797 * and on the other hand the group tiling
1805 * and store the corresponding expressions in stmt->index and stmt->local_index,
1806 * where stmt points to the ppcg_kernel_stmt that is attached to the node.
1808 static __isl_give isl_ast_node
*create_access_leaf(struct gpu_gen
*gen
,
1809 struct gpu_array_ref_group
*group
, __isl_take isl_ast_node
*node
,
1810 __isl_keep isl_ast_build
*build
)
1812 struct ppcg_kernel_stmt
*stmt
;
1813 struct gpu_array_tile
*tile
;
1818 isl_pw_multi_aff
*pma
, *pma2
;
1821 stmt
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel_stmt
);
1823 return isl_ast_node_free(node
);
1825 access
= isl_map_from_union_map(isl_ast_build_get_schedule(build
));
1826 type
= isl_map_get_tuple_name(access
, isl_dim_in
);
1827 stmt
->u
.c
.read
= !strcmp(type
, "read");
1828 access
= isl_map_reverse(access
);
1829 pma
= isl_pw_multi_aff_from_map(access
);
1830 pma
= isl_pw_multi_aff_reset_tuple_id(pma
, isl_dim_out
);
1832 space
= isl_space_range(isl_pw_multi_aff_get_space(pma
));
1833 space
= isl_space_unwrap(space
);
1834 pma2
= isl_pw_multi_aff_range_map(space
);
1835 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
,
1836 isl_pw_multi_aff_copy(pma
));
1837 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1838 stmt
->u
.c
.index
= expr
;
1840 tile
= gpu_array_ref_group_tile(group
);
1841 pma2
= isl_pw_multi_aff_from_multi_aff(
1842 isl_multi_aff_copy(tile
->tiling
));
1843 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
, pma
);
1844 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1845 stmt
->u
.c
.local_index
= expr
;
1847 stmt
->u
.c
.array
= group
->array
;
1848 stmt
->u
.c
.local_array
= group
->local_array
;
1849 stmt
->type
= ppcg_kernel_copy
;
1851 id
= isl_id_alloc(gen
->ctx
, NULL
, stmt
);
1852 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1853 return isl_ast_node_set_annotation(node
, id
);
1856 /* Create a synchronization ppcg_kernel_stmt and
1857 * attach it to the node "node" representing the synchronization.
1859 static __isl_give isl_ast_node
*create_sync_leaf(
1860 struct gpu_gen
*gen
, __isl_take isl_ast_node
*node
,
1861 __isl_keep isl_ast_build
*build
)
1863 struct ppcg_kernel_stmt
*stmt
;
1866 stmt
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel_stmt
);
1868 return isl_ast_node_free(node
);
1870 stmt
->type
= ppcg_kernel_sync
;
1871 id
= isl_id_alloc(gen
->ctx
, NULL
, stmt
);
1872 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1873 return isl_ast_node_set_annotation(node
, id
);
1876 /* This function is called for each instance of a user statement
1877 * in the kernel. This may be one of the original user statements
1878 * or a statement introduced by PPCG.
1880 * We assume that the original user statements only have a name
1881 * and no user pointer. The statements introduced by PPCG
1882 * on the other hand all have a user pointer.
1884 * If the user statement is one of the original user statements
1885 * (one with no user pointer), then we call create_domain_leaf. Otherwise,
1886 * we check if it is a copy or synchronization statement and
1887 * call the appropriate functions.
1889 static __isl_give isl_ast_node
*at_domain(__isl_take isl_ast_node
*node
,
1890 __isl_keep isl_ast_build
*build
, void *user
)
1892 struct gpu_gen
*gen
= (struct gpu_gen
*) user
;
1893 isl_ast_expr
*expr
, *arg
;
1899 expr
= isl_ast_node_user_get_expr(node
);
1900 arg
= isl_ast_expr_get_op_arg(expr
, 0);
1901 id
= isl_ast_expr_get_id(arg
);
1902 name
= isl_id_get_name(id
);
1903 p
= isl_id_get_user(id
);
1904 isl_ast_expr_free(expr
);
1905 isl_ast_expr_free(arg
);
1908 struct gpu_stmt
*gpu_stmt
;
1910 gpu_stmt
= find_stmt(gen
->prog
, id
);
1913 isl_die(gen
->ctx
, isl_error_internal
,
1914 "statement not found",
1915 return isl_ast_node_free(node
));
1917 return create_domain_leaf(gen
, node
, build
, gpu_stmt
);
1920 is_sync
= gpu_tree_id_is_sync(id
, gen
->kernel
);
1923 return isl_ast_node_free(node
);
1924 if (!strcmp(name
, "read") || !strcmp(name
, "write")) {
1925 struct gpu_array_ref_group
*group
= p
;
1926 return create_access_leaf(gen
, group
, node
, build
);
1929 isl_die(gen
->ctx
, isl_error_internal
,
1930 "unknown statement type",
1931 return isl_ast_node_free(node
));
1932 return create_sync_leaf(gen
, node
, build
);
1935 /* Given a set of wrapped references "ref", return the corresponding
1936 * access relations based on the tagged access relations "tagged".
1938 * The elements of "ref" are of the form
1942 * with D an iteration domains and R a reference.
1943 * The elements of "tagged" are of the form
1949 * Extend "tagged" to include the iteration domain in the range, i.e.,
1951 * [D -> R] -> [D -> A]
1953 * apply the result to "ref" and then unwrap the resulting set
1954 * to obtain relations of the form
1958 static __isl_give isl_union_map
*wrapped_reference_to_access(
1959 __isl_take isl_union_set
*ref
, __isl_take isl_union_map
*tagged
)
1961 isl_union_map
*tag2access
;
1963 tag2access
= isl_union_map_copy(tagged
);
1964 tag2access
= isl_union_map_universe(tag2access
);
1965 tag2access
= isl_union_set_unwrap(isl_union_map_domain(tag2access
));
1966 tag2access
= isl_union_map_domain_map(tag2access
);
1967 tag2access
= isl_union_map_range_product(tag2access
, tagged
);
1969 ref
= isl_union_set_coalesce(ref
);
1970 ref
= isl_union_set_apply(ref
, tag2access
);
1972 return isl_union_set_unwrap(ref
);
1975 /* Given an access relation "access" from "group", remove those reads
1976 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
1977 * communicate data within the same iteration of "sched".
1979 * If the access is a read then it is either an element of
1981 * live_in union (range flow)
1983 * where live_in and flow may be overapproximations, or
1984 * it reads an uninitialized value (that is not live-in because
1985 * there is an intermediate kill) or it reads a value that was
1986 * written within the same (compound) statement instance.
1987 * If the access is a write then it is either an element of
1989 * live_out union (domain flow)
1991 * or it writes a value that is never read (and is not live-out
1992 * because of an intermediate kill) or only
1993 * within the same (compound) statement instance.
1994 * In both cases, the access relation is also a subset of
1995 * the group access relation.
1997 * The cases where an uninitialized value is read or a value is written
1998 * that is never read or where the dataflow occurs within a statement
1999 * instance are also considered local and may also be removed.
2001 * Essentially, we compute the intersection of "access" with either
2003 * live_in union (range non-local-flow)
2007 * live_out union (domain non-local-flow)
2009 * We first construct a relation "local"
2011 * [[D -> R] -> [D' -> R']]
2013 * of pairs of domain iterations accessing the reference group
2014 * and references in the group that are coscheduled by "sched".
2016 * If this relation does not intersect the dataflow dependences,
2017 * then there is nothing we can possibly remove, unless the dataflow
2018 * dependences themselves only relate a subset of the accesses.
2019 * In particular, the accesses may not be involved in any dataflow
2020 * dependences, either because they are uninitialized reads/dead writes
2021 * or because the dataflow occurs inside a statement instance.
2023 * Since the computation below may break up the access relation
2024 * into smaller pieces, we only perform the intersection with
2025 * the non-local dependent accesses if the local pairs
2026 * intersect the dataflow dependences. Otherwise, we intersect
2027 * with the universe of the non-local dependent accesses.
2028 * This should at least remove accesses from statements that
2029 * do not participate in any dependences.
2031 * In particular, we remove the "local" dataflow dependences from
2032 * the set of all dataflow dependences.
2033 * Note that if the potential dataflow dependences are an overapproximation
2034 * of the actual dataflow dependences, then the result remains an
2035 * overapproximation of the non-local dataflow dependences.
2036 * Copying to/from global memory is only needed for the references
2037 * in the domain/range of the result or for accesses that are live out/in
2038 * for the entire scop.
2040 * We therefore map the domain/range of the "external" relation
2041 * to the corresponding access relation and take the union with
2042 * the live out/in relation.
2044 static __isl_give isl_union_map
*remove_local_accesses(
2045 struct gpu_prog
*prog
, struct gpu_array_ref_group
*group
,
2046 __isl_take isl_union_map
*access
, __isl_take isl_union_map
*sched
,
2050 isl_union_pw_multi_aff
*tagger
;
2051 isl_union_set
*domain
;
2052 isl_union_map
*local
, *tagged
, *external
;
2053 isl_union_set
*tag_set
;
2055 if (isl_union_map_is_empty(access
)) {
2056 isl_union_map_free(sched
);
2060 tagged
= group_tagged_access_relation(group
);
2062 tagger
= isl_union_pw_multi_aff_copy(prog
->scop
->tagger
);
2063 domain
= isl_union_map_domain(isl_union_map_copy(tagged
));
2064 tagger
= isl_union_pw_multi_aff_intersect_domain(tagger
, domain
);
2065 sched
= isl_union_map_preimage_domain_union_pw_multi_aff(sched
, tagger
);
2067 local
= isl_union_map_apply_range(sched
,
2068 isl_union_map_reverse(isl_union_map_copy(sched
)));
2069 local
= isl_union_map_intersect(local
,
2070 isl_union_map_copy(prog
->scop
->tagged_dep_flow
));
2072 empty
= isl_union_map_is_empty(local
);
2074 external
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
2075 external
= isl_union_map_intersect_params(external
,
2076 isl_set_copy(prog
->scop
->context
));
2077 external
= isl_union_map_subtract(external
, local
);
2080 tag_set
= isl_union_map_range(external
);
2081 external
= wrapped_reference_to_access(tag_set
, tagged
);
2082 external
= isl_union_map_union(external
,
2083 isl_union_map_copy(prog
->scop
->live_in
));
2085 tag_set
= isl_union_map_domain(external
);
2086 external
= wrapped_reference_to_access(tag_set
, tagged
);
2087 external
= isl_union_map_union(external
,
2088 isl_union_map_copy(prog
->scop
->live_out
));
2092 external
= isl_union_map_free(external
);
2094 external
= isl_union_map_universe(external
);
2096 access
= isl_union_map_intersect(access
, external
);
2101 /* Given an access relation "access" from "group", remove those reads
2102 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
2103 * communicate data within the same iteration of the schedule at the
2104 * position where the copying of the group is inserted.
2105 * "node" points to this position, i.e., the depth at "node"
2106 * is equal to group->depth.
2108 * We extract a schedule that picks out the iterations of the outer
2109 * group->depth dimensions and call remove_local_accesses.
2111 static __isl_give isl_union_map
*remove_local_accesses_group(
2112 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2113 __isl_take isl_union_map
*access
, __isl_keep isl_schedule_node
*node
,
2116 isl_union_map
*sched
;
2118 if (isl_union_map_is_empty(access
))
2121 sched
= isl_schedule_node_get_prefix_schedule_relation(node
);
2123 return remove_local_accesses(kernel
->prog
, group
, access
, sched
, read
);
2126 /* This function is called before the AST generator starts traversing
2127 * the schedule subtree of a node with mark "mark".
2129 * If the mark is called "kernel", store the kernel pointer in gen->kernel
2130 * for use in at_domain.
2132 static int before_mark(__isl_keep isl_id
*mark
,
2133 __isl_keep isl_ast_build
*build
, void *user
)
2135 struct gpu_gen
*gen
= user
;
2139 if (!strcmp(isl_id_get_name(mark
), "kernel"))
2140 gen
->kernel
= isl_id_get_user(mark
);
2144 /* This function is called after the AST generator has finished traversing
2145 * the schedule subtree of a mark node. "node" points to the corresponding
2148 * If the mark is called "kernel", then replace "node" by a user node
2149 * that "calls" the kernel, representing the launch of the kernel.
2150 * The original "node" is stored inside the kernel object so that
2151 * it can be used to print the device code.
2152 * Note that this assumes that a kernel is only launched once.
2153 * Also clear the kernel field of gen.
2155 static __isl_give isl_ast_node
*after_mark(__isl_take isl_ast_node
*node
,
2156 __isl_keep isl_ast_build
*build
, void *user
)
2161 isl_ast_expr_list
*list
;
2162 struct ppcg_kernel
*kernel
;
2163 struct gpu_gen
*gen
= user
;
2165 ctx
= isl_ast_node_get_ctx(node
);
2166 id
= isl_ast_node_mark_get_id(node
);
2168 return isl_ast_node_free(node
);
2169 if (strcmp(isl_id_get_name(id
), "kernel") || !gen
->kernel
) {
2173 kernel
= gen
->kernel
;
2175 kernel
->space
= isl_ast_build_get_schedule_space(build
);
2176 kernel
->tree
= isl_ast_node_mark_get_node(node
);
2177 isl_ast_node_free(node
);
2179 expr
= isl_ast_expr_from_id(isl_id_copy(id
));
2180 list
= isl_ast_expr_list_alloc(ctx
, 0);
2181 expr
= isl_ast_expr_call(expr
, list
);
2182 node
= isl_ast_node_alloc_user(expr
);
2183 node
= isl_ast_node_set_annotation(node
, id
);
2188 static int update_depth(__isl_keep isl_schedule_node
*node
, void *user
)
2193 if (isl_schedule_node_get_type(node
) != isl_schedule_node_leaf
)
2195 node_depth
= isl_schedule_node_get_schedule_depth(node
);
2196 if (node_depth
> *depth
)
2197 *depth
= node_depth
;
2202 /* Use isl to generate code for both the host and the device
2204 * The device code is marked by "kernel" mark nodes in the schedule tree,
2205 * containing a pointer to a ppcg_kernel object.
2206 * The returned AST only contains the AST for the host code.
2207 * The ASTs for the device code are embedded in ppcg_kernel objects
2208 * attached to the leaf nodes that call "kernel".
2210 static __isl_give isl_ast_node
*generate_code(struct gpu_gen
*gen
,
2211 __isl_take isl_schedule
*schedule
)
2213 isl_ast_build
*build
;
2215 isl_id_list
*iterators
;
2219 if (isl_schedule_foreach_schedule_node(schedule
, &update_depth
,
2222 build
= isl_ast_build_from_context(isl_set_copy(gen
->prog
->context
));
2223 iterators
= ppcg_scop_generate_names(gen
->prog
->scop
, depth
, "c");
2224 build
= isl_ast_build_set_iterators(build
, iterators
);
2225 build
= isl_ast_build_set_at_each_domain(build
, &at_domain
, gen
);
2226 build
= isl_ast_build_set_before_each_mark(build
, &before_mark
, gen
);
2227 build
= isl_ast_build_set_after_each_mark(build
, &after_mark
, gen
);
2228 tree
= isl_ast_build_node_from_schedule(build
, schedule
);
2229 isl_ast_build_free(build
);
2234 __isl_give isl_union_map
*extract_sizes_from_str(isl_ctx
*ctx
, const char *str
)
2238 return isl_union_map_read_from_str(ctx
, str
);
2241 /* Can "node" be tiled and then mapped to block and thread identifiers?
2242 * That is, is it permutable with at least one coincident dimension?
2244 static int is_permutable(__isl_keep isl_schedule_node
*node
)
2249 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
2251 if (!isl_schedule_node_band_get_permutable(node
))
2253 if (isl_schedule_node_band_n_member(node
) < 1)
2255 if (!isl_schedule_node_band_member_get_coincident(node
, 0))
2261 /* A isl_schedule_foreach_schedule_node callback
2262 * for setting *any_permutable and aborting the search
2263 * if "node" is a permutable band with coincident dimensions.
2264 * Otherwise, continue searching.
2266 static int set_permutable(__isl_keep isl_schedule_node
*node
, void *user
)
2268 int *any_permutable
= user
;
2271 permutable
= is_permutable(node
);
2277 *any_permutable
= 1;
2282 /* Does "schedule" contain any permutable band with at least one coincident
2285 static int has_any_permutable_node(__isl_keep isl_schedule
*schedule
)
2287 int any_permutable
= 0;
2289 if (isl_schedule_foreach_schedule_node(schedule
, &set_permutable
,
2290 &any_permutable
) < 0 &&
2294 return any_permutable
;
2297 /* Is "node" a leaf or can it be tiled and then mapped to
2298 * block and thread identifiers?
2300 static int is_leaf_or_tilable(__isl_keep isl_schedule_node
*node
)
2302 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
2304 return is_permutable(node
);
2307 /* Is "node" the outermost node in its branch that can be tiled
2308 * and then mapped to block and thread identifiers?
2309 * If there are no such nodes in the branch and if "node" is a leaf,
2310 * then it is accepted too.
2312 static int is_outer_tilable(__isl_keep isl_schedule_node
*node
)
2315 isl_schedule_node
*ancestor
;
2317 tilable
= is_leaf_or_tilable(node
);
2324 ancestor
= isl_schedule_node_copy(node
);
2325 while (isl_schedule_node_has_parent(ancestor
)) {
2326 ancestor
= isl_schedule_node_parent(ancestor
);
2328 tilable
= is_permutable(ancestor
);
2329 if (tilable
< 0 || tilable
)
2333 isl_schedule_node_free(ancestor
);
2334 return tilable
< 0 ? -1 : !tilable
;
2337 /* Construct an isl_multi_val for use as tile sizes for tiling "node"
2338 * from the elements in "tile_size".
2340 static __isl_give isl_multi_val
*construct_band_tiles_sizes(
2341 __isl_keep isl_schedule_node
*node
, int *tile_size
)
2351 ctx
= isl_schedule_node_get_ctx(node
);
2352 space
= isl_schedule_node_band_get_space(node
);
2353 n
= isl_schedule_node_band_n_member(node
);
2354 mv
= isl_multi_val_zero(space
);
2355 for (i
= 0; i
< n
; ++i
) {
2358 v
= isl_val_int_from_si(ctx
, tile_size
[i
]);
2359 mv
= isl_multi_val_set_val(mv
, i
, v
);
2365 /* Replace the partial schedule S of the band node "node" by
2373 * if scale_tile_loops is set, with f the integers in "factor".
2374 * The list that "factor" points to is assumed to contain at least
2375 * as many elements as the number of members in the band.
2377 static __isl_give isl_schedule_node
*snap_band_to_sizes(
2378 __isl_take isl_schedule_node
*node
, int *factor
,
2379 struct ppcg_options
*options
)
2383 mv
= construct_band_tiles_sizes(node
, factor
);
2384 node
= isl_schedule_node_band_scale_down(node
, isl_multi_val_copy(mv
));
2385 if (options
->scale_tile_loops
)
2386 node
= isl_schedule_node_band_scale(node
,
2387 isl_multi_val_copy(mv
));
2388 isl_multi_val_free(mv
);
2393 /* Tile "band" with tile size specified by "sizes".
2395 * Since the tile loops will be mapped to block ids, we forcibly
2396 * turn off tile loop scaling. We may want to enable tile loop scaling
2397 * at some later point, but then we would have to support the detection
2398 * of strides during the mapping to block ids.
2399 * Similarly, since the point loops will be mapped to thread ids,
2400 * we forcibly shift the point loops so that they start at zero.
2402 static __isl_give isl_schedule_node
*tile_band(
2403 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2405 isl_ctx
*ctx
= isl_schedule_node_get_ctx(node
);
2409 scale_tile
= isl_options_get_tile_scale_tile_loops(ctx
);
2410 isl_options_set_tile_scale_tile_loops(ctx
, 0);
2411 shift_point
= isl_options_get_tile_shift_point_loops(ctx
);
2412 isl_options_set_tile_shift_point_loops(ctx
, 1);
2414 node
= isl_schedule_node_band_tile(node
, sizes
);
2416 isl_options_set_tile_scale_tile_loops(ctx
, scale_tile
);
2417 isl_options_set_tile_shift_point_loops(ctx
, shift_point
);
2422 /* Extract the set of parameter values and outer schedule dimensions
2423 * for which any statement instance
2424 * in the kernel inserted at "node" needs to be executed.
2425 * Intersect the set of parameter values derived from the host schedule
2426 * relation with the context of "prog".
2428 static __isl_give isl_set
*extract_context(__isl_keep isl_schedule_node
*node
,
2429 struct gpu_prog
*prog
)
2431 isl_union_map
*schedule
;
2432 isl_union_set
*schedule_domain
;
2436 schedule
= isl_schedule_node_get_prefix_schedule_relation(node
);
2437 schedule_domain
= isl_union_map_range(schedule
);
2438 empty
= isl_union_set_is_empty(schedule_domain
);
2440 isl_union_set_free(schedule_domain
);
2447 space
= isl_union_set_get_space(schedule_domain
);
2448 isl_union_set_free(schedule_domain
);
2449 space
= isl_space_set_from_params(space
);
2450 depth
= isl_schedule_node_get_schedule_depth(node
);
2451 space
= isl_space_add_dims(space
, isl_dim_set
, depth
);
2452 context
= isl_set_empty(space
);
2454 context
= isl_set_from_union_set(schedule_domain
);
2456 context
= isl_set_intersect_params(context
,
2457 isl_set_copy(prog
->context
));
2462 /* Return the set of outer array elements accessed by
2463 * by the statement instance in "domain" in "prog".
2465 static __isl_give isl_union_set
*accessed_by_domain(
2466 __isl_take isl_union_set
*domain
, struct gpu_prog
*prog
)
2468 isl_union_map
*access
;
2469 isl_union_set
*arrays
;
2471 access
= isl_union_map_union(isl_union_map_copy(prog
->read
),
2472 isl_union_map_copy(prog
->may_write
));
2473 access
= isl_union_map_intersect_domain(access
, domain
);
2474 arrays
= isl_union_map_range(access
);
2475 arrays
= isl_union_set_apply(arrays
,
2476 isl_union_map_copy(prog
->to_outer
));
2481 /* Return the number of outer band members of the band node "node"
2482 * that are marked coincident.
2484 static int n_outer_coincidence(__isl_keep isl_schedule_node
*node
)
2488 n
= isl_schedule_node_band_n_member(node
);
2490 for (i
= 0; i
< n
; ++i
)
2491 if (!isl_schedule_node_band_member_get_coincident(node
, i
))
2497 /* If the band node "node" has more than "n" members, then split off
2498 * the first "n" of them.
2500 static __isl_give isl_schedule_node
*split_band(
2501 __isl_take isl_schedule_node
*node
, int n
)
2505 dim
= isl_schedule_node_band_n_member(node
);
2507 node
= isl_schedule_node_band_split(node
, n
);
2512 /* Scale a band node that may have been split by split_band.
2513 * "sizes" are the scaling factors for the original node.
2514 * "node" either points to the original band node, or the outer
2515 * of the two pieces after splitting.
2517 * If the number of elements in "node" is smaller than the number of
2518 * elements in "sizes", then some splitting has occurred and we split
2519 * "sizes" in the same way.
2521 static __isl_give isl_schedule_node
*scale_band(
2522 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2526 n
= isl_multi_val_dim(sizes
, isl_dim_set
);
2527 dim
= isl_schedule_node_band_n_member(node
);
2529 isl_multi_val
*sizes2
;
2531 sizes2
= isl_multi_val_copy(sizes
);
2532 sizes
= isl_multi_val_drop_dims(sizes
,
2533 isl_dim_set
, dim
, n
- dim
);
2534 sizes2
= isl_multi_val_drop_dims(sizes2
, isl_dim_set
, 0, dim
);
2535 node
= isl_schedule_node_child(node
, 0);
2536 node
= isl_schedule_node_band_scale(node
, sizes2
);
2537 node
= isl_schedule_node_parent(node
);
2540 return isl_schedule_node_band_scale(node
, sizes
);
2543 /* Return an isl_multi_aff, with as elements the parameters in "space"
2544 * that have the names specified by the elements in "names".
2545 * If (some of) these parameters do not already appear in "space",
2546 * then they are added first.
2548 static __isl_give isl_multi_aff
*parameter_vector(__isl_take isl_space
*space
,
2549 __isl_keep isl_id_list
*names
)
2552 isl_local_space
*ls
;
2556 space
= isl_space_free(space
);
2558 n
= isl_id_list_n_id(names
);
2559 for (i
= 0; i
< n
; ++i
) {
2563 id
= isl_id_list_get_id(names
, i
);
2564 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2569 pos
= isl_space_dim(space
, isl_dim_param
);
2570 space
= isl_space_add_dims(space
, isl_dim_param
, 1);
2571 space
= isl_space_set_dim_id(space
, isl_dim_param
, pos
, id
);
2573 ma
= isl_multi_aff_zero(isl_space_copy(space
));
2574 ls
= isl_local_space_from_space(isl_space_domain(space
));
2575 for (i
= 0; i
< n
; ++i
) {
2580 id
= isl_id_list_get_id(names
, i
);
2581 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2583 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
2584 isl_dim_param
, pos
);
2585 ma
= isl_multi_aff_set_aff(ma
, i
, aff
);
2587 isl_local_space_free(ls
);
2592 /* Return constraints on the domain elements that equate a sequence of
2593 * parameters called "names", to the partial schedule
2594 * of "node" modulo the integers in "size".
2595 * The number of elements in the array "size" should be equal
2596 * to the number of elements in "names".
2597 * The number of members of the band node "node" should be smaller
2598 * than or equal to this number. If it is smaller, then the first
2599 * elements of "names" are equated to zero.
2601 static __isl_give isl_union_set
*set_schedule_modulo(
2602 __isl_keep isl_schedule_node
*node
, __isl_keep isl_id_list
*names
,
2608 isl_multi_union_pw_aff
*mupa
, *mupa2
;
2610 isl_union_set
*domain
;
2614 n
= isl_id_list_n_id(names
);
2616 return isl_schedule_node_get_universe_domain(node
);
2617 n_zero
= n
- isl_schedule_node_band_n_member(node
);
2619 mupa
= isl_schedule_node_band_get_partial_schedule(node
);
2620 mv
= construct_band_tiles_sizes(node
, size
+ n_zero
);
2621 mupa
= isl_multi_union_pw_aff_mod_multi_val(mupa
, mv
);
2623 space
= isl_multi_union_pw_aff_get_space(mupa
);
2624 space
= isl_space_params(space
);
2625 space
= isl_space_set_from_params(space
);
2626 space
= isl_space_add_dims(space
, isl_dim_set
, n_zero
);
2627 ma
= isl_multi_aff_zero(space
);
2629 domain
= isl_schedule_node_get_universe_domain(node
);
2630 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(
2631 isl_union_set_copy(domain
), ma
);
2632 mupa
= isl_multi_union_pw_aff_range_product(mupa2
, mupa
);
2634 space
= isl_multi_union_pw_aff_get_space(mupa
);
2635 ma
= parameter_vector(space
, names
);
2637 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(domain
, ma
);
2638 mupa
= isl_multi_union_pw_aff_sub(mupa
, mupa2
);
2640 return isl_multi_union_pw_aff_zero_union_set(mupa
);
2643 /* Insert a context node at "node" introducing the block and thread
2644 * identifiers along with their bounds, which are stored in kernel->grid_size
2645 * and kernel->block_dim.
2646 * Note that the bounds on the block identifiers may implicitly impose
2647 * constraints on the parameters. A guard needs to be inserted
2648 * in the schedule tree to ensure that those bounds hold at "node".
2649 * This guard is inserted in insert_guard.
2651 static __isl_give isl_schedule_node
*insert_context(struct ppcg_kernel
*kernel
,
2652 __isl_take isl_schedule_node
*node
)
2656 context
= isl_set_universe(isl_set_get_space(kernel
->context
));
2658 context
= add_bounded_parameters_dynamic(context
,
2659 kernel
->grid_size
, kernel
->block_ids
);
2660 context
= add_bounded_parameters(context
,
2661 kernel
->block_dim
, kernel
->thread_ids
);
2663 node
= isl_schedule_node_insert_context(node
, context
);
2668 /* Insert a guard that eliminates kernel launches where the kernel
2669 * obviously does not have any work to do.
2671 * In particular, eliminate kernel launches where there are obviously
2673 * Use the same block size constraints that are used to create the context
2674 * to ensure that all constraints implicit in the constructed context
2675 * are imposed by the guard.
2677 * Additionally, add other constraints that are valid
2678 * for each executed instance ("context"), as long as this does not result
2681 static __isl_give isl_schedule_node
*insert_guard(
2682 __isl_take isl_schedule_node
*node
, __isl_keep isl_set
*context
,
2683 __isl_keep isl_multi_pw_aff
*size
, struct ppcg_scop
*scop
)
2689 guard
= isl_set_copy(context
);
2690 guard
= isl_set_compute_divs(guard
);
2691 guard
= isl_set_from_basic_set(isl_set_simple_hull(guard
));
2693 nparam
= isl_set_dim(guard
, isl_dim_param
);
2694 n
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
2695 ids
= ppcg_scop_generate_names(scop
, n
, "__ppcg_tmp");
2696 guard
= add_bounded_parameters_dynamic(guard
, size
, ids
);
2697 isl_id_list_free(ids
);
2698 guard
= isl_set_project_out(guard
, isl_dim_param
, nparam
, n
);
2700 node
= isl_schedule_node_insert_guard(node
, guard
);
2705 /* Does any array reference group mapping require the band that is mapped
2706 * to threads to be unrolled?
2708 static int kernel_requires_unroll(struct ppcg_kernel
*kernel
)
2712 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2713 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2715 for (j
= 0; j
< array
->n_group
; ++j
) {
2716 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2717 if (gpu_array_ref_group_requires_unroll(group
))
2725 /* Mark the given band node "node" for unrolling by the AST generator and
2726 * then sink it to the leaves of the schedule tree.
2727 * All dimensions of "node" are assumed to be coincident, such that this
2728 * sinking is a valid operation.
2730 static __isl_give isl_schedule_node
*unroll(__isl_take isl_schedule_node
*node
)
2734 n
= isl_schedule_node_band_n_member(node
);
2735 for (i
= 0; i
< n
; ++i
)
2736 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
2737 isl_ast_loop_unroll
);
2739 node
= isl_schedule_node_band_sink(node
);
2744 /* Is there any write in "kernel" that writes directly to global memory?
2745 * That is, is there any array reference group that involves a write and
2746 * that is not mapped to private or shared memory?
2748 static int any_global_write(struct ppcg_kernel
*kernel
)
2752 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2753 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2755 for (j
= 0; j
< array
->n_group
; ++j
) {
2756 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2760 if (group
->private_tile
|| group
->shared_tile
)
2769 /* Insert a synchronization node in the schedule tree of "node"
2770 * after the core computation of "kernel" at the level of the band
2771 * that is mapped to threads, except if that level is equal to
2772 * that of the band that is mapped to blocks.
2773 * "node" is assumed to point to the kernel node.
2775 static __isl_give isl_schedule_node
*add_sync(struct ppcg_kernel
*kernel
,
2776 __isl_take isl_schedule_node
*node
)
2780 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
2782 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
2783 if (kernel_depth
== isl_schedule_node_get_schedule_depth(node
))
2784 return gpu_tree_move_up_to_kernel(node
);
2786 node
= gpu_tree_ensure_following_sync(node
, kernel
);
2788 node
= gpu_tree_move_up_to_kernel(node
);
2793 /* Return a read ("read" is 1) or write access relation for "group"
2794 * with those accesses removed that are only needed to communicate data
2795 * within the subtree of the schedule rooted at "node".
2796 * Furthermore, include the prefix schedule at "node".
2797 * That is, return a relation of the form
2801 * with D the outer schedule dimensions at "node".
2803 static __isl_give isl_union_map
*anchored_non_local_accesses(
2804 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2805 __isl_take isl_schedule_node
*node
, int read
)
2807 isl_union_map
*access
;
2808 isl_union_map
*prefix
;
2810 access
= gpu_array_ref_group_access_relation(group
, read
, !read
);
2811 access
= remove_local_accesses_group(kernel
, group
, access
, node
, read
);
2812 prefix
= isl_schedule_node_get_prefix_schedule_relation(node
);
2813 access
= isl_union_map_range_product(prefix
, access
);
2818 /* Given an array reference group "group", create a mapping
2820 * read[D -> A] -> [D -> A]
2822 * if "read" is set or
2824 * write[D -> A] -> [D -> A]
2826 * if "read" is not set.
2827 * D corresponds to the outer group->depth dimensions of
2828 * the kernel schedule.
2830 static __isl_give isl_multi_aff
*create_from_access(isl_ctx
*ctx
,
2831 struct gpu_array_ref_group
*group
, int read
)
2836 space
= isl_space_copy(group
->array
->space
);
2837 space
= isl_space_from_range(space
);
2838 space
= isl_space_add_dims(space
, isl_dim_in
, group
->depth
);
2839 space
= isl_space_wrap(space
);
2840 space
= isl_space_map_from_set(space
);
2842 id
= isl_id_alloc(ctx
, read
? "read" : "write", group
);
2843 space
= isl_space_set_tuple_id(space
, isl_dim_in
, id
);
2845 return isl_multi_aff_identity(space
);
2848 /* Add copy statements to the schedule tree of "node"
2849 * for reading from global memory to private memory (if "read" is set) or
2850 * for writing back from private memory to global memory
2851 * (if "read" is not set) for the array reference group "group" that
2852 * is mapped to private memory.
2853 * On input, "node" points to the kernel node, and it is moved
2854 * back there on output.
2856 * The copies are performed in the order of the array elements.
2857 * The copy statement instances include a reference to the outer
2858 * group->depth dimensions of the kernel schedule for ease of
2859 * combining them with the group tiling.
2861 * That is, the extra schedule is of the form
2865 * where D corresponds to the outer group->depth dimensions of
2866 * the kernel schedule and A to the global array.
2867 * This schedule is unrolled because registers are not addressable.
2869 * The copying is inserted in the schedule tree through an extension
2874 * where the extra domain elements type[D -> A] are those accessed
2876 * A filter is inserted on type[D -> A] to ensure that the element
2877 * is read/written by the same thread that needs the element.
2878 * This filter is obtained by applying
2882 * to the thread filter for the core statements.
2884 * The extension is inserted before the core computation in case of a read
2885 * and after the core computation in case of a write.
2886 * In the latter case, we also make sure that there is a synchronization
2887 * node after the write to global memory, unless this write is performed
2888 * at the outer level of the kernel.
2889 * In principle, this synchronization could be inserted higher
2890 * in the schedule tree depending on where the corresponding reads
2891 * from global memory are performed.
2893 static __isl_give isl_schedule_node
*add_copies_group_private(
2894 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2895 __isl_take isl_schedule_node
*node
, int read
)
2897 isl_union_map
*access
;
2898 isl_union_map
*prefix
;
2899 isl_union_set
*domain
;
2901 isl_multi_aff
*from_access
;
2902 isl_multi_pw_aff
*mpa
;
2903 isl_multi_union_pw_aff
*mupa
;
2904 isl_schedule_node
*graft
;
2905 isl_union_set
*filter
;
2909 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
2910 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
2912 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
2913 empty
= isl_union_map_is_empty(access
);
2914 if (empty
< 0 || empty
) {
2915 isl_union_map_free(access
);
2917 return isl_schedule_node_free(node
);
2918 return gpu_tree_move_up_to_kernel(node
);
2921 from_access
= create_from_access(kernel
->ctx
, group
, read
);
2922 space
= isl_space_domain(isl_multi_aff_get_space(from_access
));
2923 access
= isl_union_map_preimage_range_multi_aff(access
, from_access
);
2925 filter
= isl_union_set_copy(kernel
->thread_filter
);
2926 filter
= isl_union_set_apply(filter
, isl_union_map_copy(access
));
2927 filter
= isl_union_set_detect_equalities(filter
);
2928 filter
= isl_union_set_coalesce(filter
);
2930 domain
= isl_union_map_range(access
);
2931 access
= isl_union_set_wrapped_domain_map(domain
);
2932 access
= isl_union_map_reverse(access
);
2933 access
= isl_union_map_coalesce(access
);
2934 graft
= isl_schedule_node_from_extension(access
);
2936 space
= isl_space_map_from_set(space
);
2937 mpa
= isl_multi_pw_aff_identity(space
);
2938 mpa
= isl_multi_pw_aff_range_factor_range(mpa
);
2939 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
2941 graft
= isl_schedule_node_child(graft
, 0);
2942 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
2943 graft
= unroll(graft
);
2945 graft
= isl_schedule_node_insert_filter(graft
, filter
);
2947 graft
= isl_schedule_node_parent(graft
);
2950 node
= isl_schedule_node_graft_before(node
, graft
);
2952 node
= isl_schedule_node_graft_after(node
, graft
);
2953 if (kernel_depth
< group
->depth
) {
2954 node
= isl_schedule_node_parent(node
);
2955 node
= isl_schedule_node_next_sibling(node
);
2956 node
= isl_schedule_node_child(node
, 0);
2957 node
= gpu_tree_ensure_following_sync(node
, kernel
);
2961 node
= gpu_tree_move_up_to_kernel(node
);
2966 /* Add copy statements to the schedule tree of "node"
2967 * for reading from global memory to shared memory (if "read" is set) or
2968 * for writing back from shared memory to global memory
2969 * (if "read" is not set) for the array reference group "group" that
2970 * is mapped to shared memory.
2971 * On input, "node" points to the kernel node, and it is moved
2972 * back there on output.
2974 * The copies are performed in the order of the corresponding shared
2976 * The copy statement instances include a reference to the outer
2977 * group->depth dimensions of the kernel schedule for ease of
2978 * combining them with the group tiling.
2980 * If we are performing a read from global memory to shared memory and
2981 * if the array involved is not a scalar, then we copy
2982 * the entire tile to shared memory. This may result in some extra
2983 * elements getting copied, but it should lead to simpler code
2984 * (which means that fewer registers may be needed) and less divergence.
2986 * Otherwise, we only copy the elements that will be read or have been written
2989 * That is, the extra schedule is of the form
2993 * where D corresponds to the outer group->depth dimensions of
2994 * the kernel schedule, A to the global array and T is the corresponding
2995 * shared memory tile.
2997 * The copying is inserted in the schedule tree through an extension
3002 * where the extra domain elements type[D -> A] are those accessed
3003 * by the group. In the case of read from a non-scalar, this set
3004 * is replaced by the entire shared memory tile.
3006 * A filter is inserted on type[D -> A] to map the copy instances
3007 * to the threads. In particular, the thread identifiers are
3008 * equated to the position inside the shared memory tile (T)
3009 * modulo the block size.
3010 * We try to align the innermost tile dimension with the innermost
3011 * thread identifier (x) as a heuristic to improve coalescing.
3012 * In particular, if the dimension of the tile is greater than
3013 * the dimension of the block, then the schedule mapping to the tile
3014 * is broken up into two pieces and the filter is applied to the inner part.
3015 * If, on the other hand, the dimension of the tile is smaller than
3016 * the dimension of the block, then the initial thread identifiers
3017 * are equated to zero and the remaining thread identifiers are
3018 * matched to the memory tile.
3020 * The extension is inserted before the core computation in case of a read
3021 * and after the core computation in case of a write.
3022 * In the case of a read, we first need to make sure there is some
3023 * synchronization before the core computation such that we can put the read
3024 * from global memory to shared memory before that synchronization.
3025 * This ensures that all threads have finished copying into shared memory
3026 * before the shared memory is used.
3027 * We also need to make sure that there is a synchronization node after
3028 * the core computation to ensure that the next load into shared memory
3029 * only happens after all data has been used. There is no need for
3030 * this synchronization if we are at the outer level since then there
3031 * won't be a next load.
3032 * In the case of a write, we need to make sure there is some synchronization
3033 * after the core computation such taht we can put the write from shared
3034 * memory to global memory after that synchronization.
3035 * Unless we are at the outer level, we also need a synchronization node
3036 * after the write to ensure the data is saved to global memory
3037 * before the next iteration write to the same shared memory.
3038 * It also makes sure the data has arrived in global memory before
3039 * it is read in a subsequent iteration.
3041 static __isl_give isl_schedule_node
*add_copies_group_shared(
3042 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3043 __isl_take isl_schedule_node
*node
, int read
)
3045 struct gpu_array_tile
*tile
;
3046 isl_union_map
*access
;
3047 isl_union_set
*domain
;
3048 isl_union_set
*sync
;
3050 isl_multi_aff
*from_access
;
3051 isl_multi_pw_aff
*mpa
;
3052 isl_multi_union_pw_aff
*mupa
;
3053 isl_schedule_node
*graft
;
3054 isl_union_set
*filter
;
3059 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3060 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3062 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3063 empty
= isl_union_map_is_empty(access
);
3064 if (empty
< 0 || empty
) {
3065 isl_union_map_free(access
);
3067 return isl_schedule_node_free(node
);
3068 return gpu_tree_move_up_to_kernel(node
);
3071 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3073 tile
= gpu_array_ref_group_tile(group
);
3074 ma
= isl_multi_aff_copy(tile
->tiling
);
3075 ma
= isl_multi_aff_pullback_multi_aff(ma
,
3076 isl_multi_aff_copy(from_access
));
3077 mpa
= isl_multi_pw_aff_from_multi_aff(ma
);
3078 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3080 domain
= isl_union_map_range(access
);
3082 if (read
&& !gpu_array_is_scalar(group
->array
)) {
3084 isl_union_set_free(domain
);
3085 map
= group_tile(group
);
3086 domain
= isl_union_set_from_set(isl_map_wrap(map
));
3089 domain
= isl_union_set_preimage_multi_aff(domain
, from_access
);
3090 access
= isl_union_set_wrapped_domain_map(domain
);
3091 access
= isl_union_map_reverse(access
);
3092 access
= isl_union_map_coalesce(access
);
3093 graft
= isl_schedule_node_from_extension(access
);
3095 graft
= isl_schedule_node_child(graft
, 0);
3097 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3099 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0) {
3100 graft
= isl_schedule_node_band_split(graft
,
3101 tile
->n
- kernel
->n_block
);
3102 graft
= isl_schedule_node_child(graft
, 0);
3104 if (tile
->n
< kernel
->n_block
)
3105 skip
= kernel
->n_block
- tile
->n
;
3108 filter
= set_schedule_modulo(graft
, kernel
->thread_ids
,
3110 if (!kernel
->options
->wrap
)
3111 graft
= snap_band_to_sizes(graft
, kernel
->block_dim
+ skip
,
3113 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0)
3114 graft
= isl_schedule_node_parent(graft
);
3115 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3117 while (graft
&& isl_schedule_node_has_parent(graft
))
3118 graft
= isl_schedule_node_parent(graft
);
3121 if (kernel_depth
< group
->depth
)
3122 node
= gpu_tree_ensure_sync_after_core(node
, kernel
);
3123 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3124 node
= isl_schedule_node_graft_before(node
, graft
);
3126 node
= gpu_tree_move_right_to_sync(node
, kernel
);
3127 node
= isl_schedule_node_graft_after(node
, graft
);
3128 node
= isl_schedule_node_parent(node
);
3129 node
= isl_schedule_node_next_sibling(node
);
3130 node
= isl_schedule_node_child(node
, 0);
3131 if (kernel_depth
< group
->depth
)
3132 node
= gpu_tree_ensure_following_sync(node
, kernel
);
3135 node
= gpu_tree_move_up_to_kernel(node
);
3140 /* Check whether the array reference group "group" is mapped to
3141 * private or shared memory and, if so,
3142 * add copy statements to the schedule tree of "node"
3143 * for reading from global memory to private or shared memory
3144 * (if "read" is set) or for writing back from private or shared memory
3145 * to global memory (if "read" is not set) for this group.
3146 * On input, "node" points to the kernel node, and it is moved
3147 * back there on output.
3149 static __isl_give isl_schedule_node
*add_copies_group(
3150 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3151 __isl_take isl_schedule_node
*node
, int read
)
3153 if (group
->private_tile
)
3154 return add_copies_group_private(kernel
, group
, node
, read
);
3155 if (group
->shared_tile
)
3156 return add_copies_group_shared(kernel
, group
, node
, read
);
3160 /* For each array reference group that is mapped to private or shared memory,
3161 * add copy statements to the schedule tree of "node"
3162 * for reading from global memory to private or shared memory
3163 * and for writing back.
3164 * On input, "node" points to the kernel node, and it is moved
3165 * back there on output.
3167 static __isl_give isl_schedule_node
*add_copies(struct ppcg_kernel
*kernel
,
3168 __isl_take isl_schedule_node
*node
)
3172 for (i
= 0; i
< kernel
->n_array
; ++i
) {
3173 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
3175 for (j
= 0; j
< array
->n_group
; ++j
) {
3176 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3178 node
= add_copies_group(kernel
, group
, node
, 1);
3181 node
= add_copies_group(kernel
, group
, node
, 0);
3190 /* Mark all dimensions in the current band node atomic.
3192 static __isl_give isl_schedule_node
*atomic(__isl_take isl_schedule_node
*node
)
3196 n
= isl_schedule_node_band_n_member(node
);
3197 for (i
= 0; i
< n
; ++i
)
3198 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
3199 isl_ast_loop_atomic
);
3204 /* Mark "node" atomic, if it is a band node.
3205 * Do the same for all ancestors.
3206 * Return a pointer to "node" (in the updated schedule tree).
3208 static __isl_give isl_schedule_node
*atomic_ancestors(
3209 __isl_take isl_schedule_node
*node
)
3215 if (!isl_schedule_node_has_parent(node
))
3218 pos
= isl_schedule_node_get_child_position(node
);
3219 node
= isl_schedule_node_parent(node
);
3220 if (isl_schedule_node_get_type(node
) == isl_schedule_node_band
)
3221 node
= atomic(node
);
3222 node
= atomic_ancestors(node
);
3223 node
= isl_schedule_node_child(node
, pos
);
3228 /* Group the domain elements into a single space, named kernelX,
3229 * with X the kernel sequence number "kernel_id".
3231 static __isl_give isl_schedule_node
*group_statements(
3232 __isl_take isl_schedule_node
*node
, int kernel_id
)
3240 snprintf(buffer
, sizeof(buffer
), "kernel%d", kernel_id
);
3241 id
= isl_id_alloc(isl_schedule_node_get_ctx(node
), buffer
, NULL
);
3242 return isl_schedule_node_group(node
, id
);
3245 /* Create a ppcg_kernel representing the domain instances that reach "node"
3246 * and insert a mark node pointing to the ppcg_kernel before "node".
3247 * The band that "node" points to is the band that needs to be mapped
3248 * to block identifiers. The band that needs to be mapped to thread
3249 * identifiers should be marked by a "thread" mark by the caller.
3250 * This mark is removed by this function.
3251 * If "scale" is set, then the band that "node" points to is scaled
3254 * Mark all outer band nodes as atomic to ensure each kernel is only
3256 * If the domain elements that reach "node" live in more than one space,
3257 * then group the domain elements into a single space, named kernelX,
3258 * with X the kernel sequence number.
3260 * Insert a guard node governing the kernel node to ensure that
3261 * no kernels with zero blocks are launched.
3263 * Insert a context node describing the block and thread
3264 * identifiers inside the kernel mark.
3265 * The context node needs to be inserted after the effective block size
3266 * has been determined such that the bounds on the thread identifiers
3267 * would reflect the effective block size.
3268 * Insert a filter node inside the context node mapping the statement
3269 * instances to block identifiers. In particular, the block identifiers
3270 * are equated to the partial schedule of band that was marked for mapping
3271 * to blocks modulo the grid size.
3272 * Insert a filter node inside the "thread" mark mapping the statement
3273 * instances to thread identifiers. In particular, the thread identifiers
3274 * are equated to the partial schedule of band that was marked for mapping
3275 * to threads modulo the block size.
3277 * Compute array reference groups for all arrays, set the local
3278 * array bounds based on the set of domain instances that reach
3279 * the kernel node, check the total amount of shared memory used
3280 * and compute all group tilings.
3281 * The array reference groups are computed after the block filter
3282 * has been inserted because it affects the mapping to shared or
3283 * private memory. This computation also requires the thread filter
3284 * (in the ppcg_kernel object), but this thread filter should not
3285 * have been added to the schedule tree yet since the computation
3286 * requires the schedule of the band that needs to be mapped to
3287 * threads before the privatization is applied.
3289 * If any array reference group requires the band mapped to threads
3290 * to be unrolled, then we perform the required unrolling.
3292 * We save a copy of the schedule that may influence the mappings
3293 * to shared or private memory in kernel->shared_schedule.
3295 * Finally, we add synchronization and copy statements to the schedule tree,
3296 * remove the "thread" mark and create representations for the local
3297 * variables in the kernel.
3299 * We keep a copy of the isl_id that points to the kernel to ensure
3300 * that the kernel does not get destroyed if the schedule node
3301 * is freed due to some error condition.
3303 static __isl_give isl_schedule_node
*create_kernel(struct gpu_gen
*gen
,
3304 __isl_take isl_schedule_node
*node
, int scale
,
3305 __isl_keep isl_multi_val
*sizes
)
3307 struct ppcg_kernel
*kernel
;
3309 isl_schedule_node
*node_thread
;
3310 isl_union_map
*host_schedule
;
3311 isl_set
*host_domain
;
3312 isl_union_set
*domain
;
3313 int single_statement
;
3315 kernel
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel
);
3316 kernel
= ppcg_kernel_create_local_arrays(kernel
, gen
->prog
);
3318 return isl_schedule_node_free(node
);
3320 domain
= isl_schedule_node_get_domain(node
);
3321 single_statement
= isl_union_set_n_set(domain
) == 1;
3323 kernel
->ctx
= gen
->ctx
;
3324 kernel
->prog
= gen
->prog
;
3325 kernel
->options
= gen
->options
;
3326 kernel
->context
= extract_context(node
, gen
->prog
);
3327 kernel
->core
= isl_union_set_universe(isl_union_set_copy(domain
));
3328 kernel
->arrays
= accessed_by_domain(isl_union_set_copy(domain
),
3330 kernel
->n_grid
= n_outer_coincidence(node
);
3331 node_thread
= isl_schedule_node_copy(node
);
3332 node_thread
= gpu_tree_move_down_to_thread(node_thread
, kernel
->core
);
3333 node_thread
= isl_schedule_node_child(node_thread
, 0);
3334 kernel
->n_block
= n_outer_coincidence(node_thread
);
3335 isl_schedule_node_free(node_thread
);
3336 kernel
->id
= gen
->kernel_id
++;
3337 read_grid_and_block_sizes(kernel
, gen
);
3339 host_schedule
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3340 host_domain
= isl_set_from_union_set(isl_union_map_range(
3343 node
= atomic_ancestors(node
);
3345 id
= isl_id_alloc(gen
->ctx
, "kernel", kernel
);
3346 id
= isl_id_set_free_user(id
, &ppcg_kernel_free_wrap
);
3347 node
= isl_schedule_node_insert_mark(node
, isl_id_copy(id
));
3349 if (!single_statement
)
3350 node
= group_statements(node
, kernel
->id
);
3352 node
= isl_schedule_node_child(node
, 0);
3353 node
= split_band(node
, kernel
->n_grid
);
3354 kernel
->block_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3355 kernel
->n_grid
, "b");
3356 kernel
->block_filter
= set_schedule_modulo(node
, kernel
->block_ids
,
3358 kernel
->grid_size
= extract_grid_size(kernel
,
3359 isl_union_set_copy(domain
));
3360 if (!kernel
->options
->wrap
)
3361 node
= snap_band_to_sizes(node
, kernel
->grid_dim
,
3364 node
= scale_band(node
, isl_multi_val_copy(sizes
));
3365 node
= isl_schedule_node_parent(node
);
3366 if (!single_statement
)
3367 node
= isl_schedule_node_parent(node
);
3368 node
= insert_guard(node
, kernel
->context
, kernel
->grid_size
,
3370 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3371 node
= isl_schedule_node_child(node
, 0);
3372 node
= split_band(node
, kernel
->n_block
);
3373 kernel
->thread_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3374 kernel
->n_block
, "t");
3375 kernel
->thread_filter
= set_schedule_modulo(node
, kernel
->thread_ids
,
3377 extract_block_size(kernel
, domain
);
3379 node
= gpu_tree_move_up_to_kernel(node
);
3380 node
= isl_schedule_node_child(node
, 0);
3381 node
= insert_context(kernel
, node
);
3382 node
= isl_schedule_node_child(node
, 0);
3383 node
= isl_schedule_node_insert_filter(node
,
3384 isl_union_set_copy(kernel
->block_filter
));
3386 node
= gpu_tree_move_up_to_kernel(node
);
3388 if (gpu_group_references(kernel
, node
) < 0)
3389 node
= isl_schedule_node_free(node
);
3390 localize_bounds(kernel
, host_domain
);
3391 isl_set_free(host_domain
);
3393 check_shared_memory_bound(kernel
);
3394 compute_group_tilings(kernel
);
3396 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3397 node
= isl_schedule_node_child(node
, 0);
3398 if (!kernel
->options
->wrap
)
3399 node
= snap_band_to_sizes(node
, kernel
->block_dim
,
3401 node
= isl_schedule_node_insert_filter(node
,
3402 isl_union_set_copy(kernel
->thread_filter
));
3403 if (kernel_requires_unroll(kernel
)) {
3404 node
= isl_schedule_node_child(node
, 0);
3405 node
= unroll(node
);
3408 node
= gpu_tree_move_up_to_thread(node
);
3409 kernel
->shared_schedule_dim
=
3410 isl_schedule_node_get_schedule_depth(node
);
3411 kernel
->shared_schedule
=
3412 isl_schedule_node_get_prefix_schedule_union_pw_multi_aff(node
);
3414 node
= gpu_tree_move_up_to_kernel(node
);
3416 if (any_global_write(kernel
))
3417 node
= add_sync(kernel
, node
);
3418 node
= add_copies(kernel
, node
);
3420 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3421 node
= isl_schedule_node_delete(node
);
3423 node
= gpu_tree_move_up_to_kernel(node
);
3425 if (create_kernel_vars(kernel
) < 0)
3426 node
= isl_schedule_node_free(node
);
3428 if (!single_statement
)
3429 node
= isl_schedule_node_parent(node
);
3430 node
= isl_schedule_node_parent(node
);
3436 /* Insert a zero-dimensional permutable band at "node".
3438 static __isl_give isl_schedule_node
*insert_empty_permutable_band(
3439 __isl_take isl_schedule_node
*node
)
3442 isl_schedule
*schedule
;
3443 isl_union_set
*domain
;
3444 isl_multi_union_pw_aff
*mupa
;
3446 schedule
= isl_schedule_node_get_schedule(node
);
3447 domain
= isl_schedule_get_domain(schedule
);
3448 space
= isl_union_set_get_space(domain
);
3449 isl_union_set_free(domain
);
3450 isl_schedule_free(schedule
);
3452 space
= isl_space_set_from_params(space
);
3453 mupa
= isl_multi_union_pw_aff_zero(space
);
3454 node
= isl_schedule_node_insert_partial_schedule(node
, mupa
);
3455 node
= isl_schedule_node_band_set_permutable(node
, 1);
3460 /* If "node" is the outermost permutable band that can be mapped to block and
3461 * thread identifiers in its branch (or a leaf with no such outer bands),
3462 * then mark the band as such, attaching a ppcg_kernel to the mark.
3464 * If "node" originally points to a leaf, then insert a zero-dimensional
3465 * permutable band such that we can assume that "node" always
3466 * points to a band node.
3468 * Tile "node" using user specified tile sizes, after splitting the band
3469 * if the number of specified tile sizes is smaller than the dimension
3470 * of the band. Mark the point band of this tiling as the band that
3471 * needs to be mapped to threads.
3472 * Create a kernel representing the domain instances that reach "node" and
3473 * insert a mark node pointing to the ppcg_kernel before the band node.
3475 static __isl_give isl_schedule_node
*mark_outer_permutable(
3476 __isl_take isl_schedule_node
*node
, void *user
)
3478 struct gpu_gen
*gen
= user
;
3484 isl_multi_val
*sizes
;
3486 outer
= is_outer_tilable(node
);
3488 return isl_schedule_node_free(node
);
3492 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
3493 node
= insert_empty_permutable_band(node
);
3495 tile_len
= isl_schedule_node_band_n_member(node
);
3496 tile_size
= read_tile_sizes(gen
, &tile_len
);
3498 return isl_schedule_node_free(node
);
3499 if (tile_len
< isl_schedule_node_band_n_member(node
))
3500 node
= isl_schedule_node_band_split(node
, tile_len
);
3501 sizes
= construct_band_tiles_sizes(node
, tile_size
);
3502 node
= tile_band(node
, isl_multi_val_copy(sizes
));
3503 node
= isl_schedule_node_child(node
, 0);
3504 id
= isl_id_alloc(gen
->ctx
, "thread", NULL
);
3505 node
= isl_schedule_node_insert_mark(node
, id
);
3506 node
= isl_schedule_node_parent(node
);
3508 scale
= gen
->options
->scale_tile_loops
;
3509 node
= create_kernel(gen
, node
, scale
, sizes
);
3510 isl_multi_val_free(sizes
);
3516 /* Insert "kernel" marks that point to a ppcg_kernel structure
3517 * in front of all outermost tilable band that (by construction)
3518 * have at least one parallel loop.
3520 static __isl_give isl_schedule
*mark_kernels(struct gpu_gen
*gen
,
3521 __isl_take isl_schedule
*schedule
)
3523 isl_schedule_node
*node
;
3525 node
= isl_schedule_get_root(schedule
);
3526 isl_schedule_free(schedule
);
3527 node
= isl_schedule_node_child(node
, 0);
3528 node
= isl_schedule_node_map_descendant(node
,
3529 &mark_outer_permutable
, gen
);
3530 schedule
= isl_schedule_node_get_schedule(node
);
3531 isl_schedule_node_free(node
);
3535 /* Compute an appropriate schedule based on the accesses in
3536 * gen->read and gen->write.
3538 * We use the dependences in gen->prog->scop to compute
3539 * a schedule that has a parallel loop in each tilable band and
3540 * return this schedule.
3542 * If live range reordering is allowed, then we need to make sure
3543 * that live ranges on arrays are not run in parallel since doing
3544 * so would require array expansion. We therefore add the array
3545 * order dependences to the coincidence dependences. Non-zero array
3546 * order dependences will then prevent a schedule dimension from being
3547 * considered parallel.
3548 * Live ranges derived from scalars are allowed to be run in parallel
3549 * since we force the scalars to be mapped to private memory in
3550 * check_scalar_live_ranges.
3551 * If live range reordering is allowed, then the false dependences
3552 * are not added to the validity constraints as that would prevent
3553 * reordering. Instead, the external false dependences that enforce that reads
3554 * from potentially live-in data precede any later write and
3555 * that writes of potentially live-out data follow any other earlier write
3556 * are added to the validity and the coincidence constraints.
3557 * The false dependences are still added to the proximity constraints
3558 * for consistency with the case where live range reordering is not allowed.
3559 * The coincidence constraints then consist of flow dependences,
3560 * external false dependences and array order dependences.
3561 * The independences can be filtered out from the first two sets.
3562 * They have already been filtered out from the array order dependences
3563 * on a per array basis in collect_order_dependences.
3564 * There is no need for a per array handling of the other two sets
3565 * as there should be no flow or external false dependence on local
3566 * variables that can be filtered out.
3568 static __isl_give isl_schedule
*compute_schedule(struct gpu_gen
*gen
)
3570 isl_union_set
*domain
;
3571 isl_union_map
*dep_raw
, *dep
;
3572 isl_union_map
*validity
, *proximity
, *coincidence
;
3573 isl_schedule_constraints
*sc
;
3574 isl_schedule
*schedule
;
3576 domain
= isl_union_set_copy(gen
->prog
->scop
->domain
);
3577 sc
= isl_schedule_constraints_on_domain(domain
);
3578 sc
= isl_schedule_constraints_set_context(sc
,
3579 isl_set_copy(gen
->prog
->scop
->context
));
3580 if (gen
->options
->live_range_reordering
) {
3581 sc
= isl_schedule_constraints_set_conditional_validity(sc
,
3582 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_flow
),
3583 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_order
));
3584 proximity
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
3585 validity
= isl_union_map_copy(proximity
);
3586 validity
= isl_union_map_union(validity
,
3587 isl_union_map_copy(gen
->prog
->scop
->dep_forced
));
3588 proximity
= isl_union_map_union(proximity
,
3589 isl_union_map_copy(gen
->prog
->scop
->dep_false
));
3590 coincidence
= isl_union_map_copy(validity
);
3591 coincidence
= isl_union_map_subtract(coincidence
,
3592 isl_union_map_copy(gen
->prog
->scop
->independence
));
3593 coincidence
= isl_union_map_union(coincidence
,
3594 isl_union_map_copy(gen
->prog
->array_order
));
3596 dep_raw
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
3597 dep
= isl_union_map_copy(gen
->prog
->scop
->dep_false
);
3598 dep
= isl_union_map_union(dep
, dep_raw
);
3599 dep
= isl_union_map_coalesce(dep
);
3600 proximity
= isl_union_map_copy(dep
);
3601 coincidence
= isl_union_map_copy(dep
);
3604 sc
= isl_schedule_constraints_set_validity(sc
, validity
);
3605 sc
= isl_schedule_constraints_set_coincidence(sc
, coincidence
);
3606 sc
= isl_schedule_constraints_set_proximity(sc
, proximity
);
3608 if (gen
->options
->debug
->dump_schedule_constraints
)
3609 isl_schedule_constraints_dump(sc
);
3610 schedule
= isl_schedule_constraints_compute_schedule(sc
);
3611 if (gen
->options
->debug
->dump_schedule
)
3612 isl_schedule_dump(schedule
);
3617 /* Compute the sets of outer array elements that need to be copied in and out.
3619 * In particular, for each array that is possibly written anywhere in
3620 * gen->prog and that is visible outside the corresponding scop,
3621 * we copy out its entire extent.
3623 * Any array elements that is read without first being written needs
3624 * to be copied in. Furthermore, if there are any array elements that
3625 * are copied out, but that may not be written inside gen->prog, then
3626 * they also need to be copied in to ensure that the value after execution
3627 * is the same as the value before execution, at least for those array
3628 * elements that may have their values preserved by the scop.
3629 * In case the array elements are structures, we need to take into
3630 * account that all members of the structures need to be written
3631 * by gen->prog before we can avoid copying the data structure in.
3633 * While computing the set of array elements that are copied out but
3634 * not necessarily written, we intersect both sets with the context.
3635 * This helps in those cases where the arrays are declared with a fixed size,
3636 * while the accesses are parametric and the context assigns a fixed value
3637 * to the parameters.
3639 * If an element from a local array is read without first being written,
3640 * then there is no point in copying it in since it cannot have been
3641 * written prior to the scop. Warn about the uninitialized read instead.
3643 static void compute_copy_in_and_out(struct gpu_gen
*gen
)
3646 isl_union_set
*local
;
3647 isl_union_set
*may_write
, *must_write
;
3648 isl_union_set
*copy_in
, *copy_out
;
3649 isl_union_set
*not_written
;
3650 isl_union_map
*uninitialized
;
3651 isl_union_map
*local_uninitialized
;
3653 must_write
= isl_union_map_range(
3654 isl_union_map_copy(gen
->prog
->must_write
));
3655 must_write
= isl_union_set_intersect_params(must_write
,
3656 isl_set_copy(gen
->prog
->context
));
3657 may_write
= isl_union_map_range(
3658 isl_union_map_copy(gen
->prog
->may_write
));
3659 may_write
= isl_union_set_intersect_params(may_write
,
3660 isl_set_copy(gen
->prog
->context
));
3661 may_write
= isl_union_set_universe(may_write
);
3662 may_write
= isl_union_set_apply(may_write
,
3663 isl_union_map_copy(gen
->prog
->to_outer
));
3664 copy_out
= isl_union_set_empty(isl_union_set_get_space(may_write
));
3665 local
= isl_union_set_copy(copy_out
);
3667 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
3672 space
= isl_space_copy(gen
->prog
->array
[i
].space
);
3674 if (gen
->prog
->array
[i
].local
) {
3677 set
= isl_set_universe(space
);
3678 local
= isl_union_set_add_set(local
, set
);
3682 write_i
= isl_union_set_extract_set(may_write
, space
);
3683 empty
= isl_set_plain_is_empty(write_i
);
3684 isl_set_free(write_i
);
3688 write_i
= isl_set_copy(gen
->prog
->array
[i
].extent
);
3689 copy_out
= isl_union_set_add_set(copy_out
, write_i
);
3691 isl_union_set_free(may_write
);
3693 copy_out
= isl_union_set_intersect_params(copy_out
,
3694 isl_set_copy(gen
->prog
->context
));
3696 gen
->prog
->copy_out
= isl_union_set_copy(copy_out
);
3698 copy_out
= isl_union_set_apply(copy_out
,
3699 isl_union_map_copy(gen
->prog
->to_inner
));
3700 copy_out
= isl_union_set_intersect(copy_out
,
3701 isl_union_set_copy(gen
->prog
->may_persist
));
3702 not_written
= isl_union_set_subtract(copy_out
, must_write
);
3704 uninitialized
= isl_union_map_copy(gen
->prog
->scop
->live_in
);
3705 local_uninitialized
= isl_union_map_copy(uninitialized
);
3707 local
= isl_union_set_apply(local
,
3708 isl_union_map_copy(gen
->prog
->to_inner
));
3709 local_uninitialized
= isl_union_map_intersect_range(local_uninitialized
,
3711 if (!isl_union_map_is_empty(local_uninitialized
)) {
3713 "possibly uninitialized reads (not copied in):\n");
3714 isl_union_map_dump(local_uninitialized
);
3716 uninitialized
= isl_union_map_subtract(uninitialized
,
3717 local_uninitialized
);
3718 copy_in
= isl_union_map_range(uninitialized
);
3719 copy_in
= isl_union_set_union(copy_in
, not_written
);
3720 copy_in
= isl_union_set_apply(copy_in
,
3721 isl_union_map_copy(gen
->prog
->to_outer
));
3723 gen
->prog
->copy_in
= copy_in
;
3726 /* Internal data structure for extract_access.
3727 * "next_access" points to the end of a linked list that is extended
3728 * by extract_access.
3729 * "single_expression" is set if the access expressions belong to
3730 * an expression statement (i.e., a statement without internal control).
3731 * "any_to_outer" maps all intermediate arrays to their outer arrays.
3733 struct ppcg_extract_access_data
{
3734 struct gpu_stmt_access
**next_access
;
3735 int single_expression
;
3736 isl_union_map
*any_to_outer
;
3739 /* Given a tagged access relation to a single array "tagged", extract it
3740 * as a map, taking into account that the input may be empty.
3741 * If the access relation is empty, then it does not contain
3742 * any space information, so we try to recover it from the index
3744 * The space of the index expression is of the form I -> A,
3745 * with I the statement instances and A the array, or [I -> F] -> A,
3746 * with F the filters corresponding to arguments.
3747 * We first drop F, if present, obtaining I -> A.
3748 * Then we construct I -> R, with R the reference tag,
3749 * combine the two into I -> [R -> A] and uncurry to obtain
3750 * the final result [I -> R] -> A.
3751 * Note that the index expression may have a lower dimension
3752 * than that of the array, but this dimension is not used
3753 * if the access relation is empty.
3755 static __isl_give isl_map
*extract_single_tagged_access(
3756 __isl_take isl_union_map
*tagged
, __isl_keep pet_expr
*expr
)
3760 isl_space
*space
, *space2
;
3761 isl_multi_pw_aff
*index
;
3763 empty
= isl_union_map_is_empty(tagged
);
3767 return isl_map_from_union_map(tagged
);
3768 isl_union_map_free(tagged
);
3770 index
= pet_expr_access_get_index(expr
);
3771 space
= isl_multi_pw_aff_get_space(index
);
3772 isl_multi_pw_aff_free(index
);
3773 if (isl_space_domain_is_wrapping(space
))
3774 space
= isl_space_domain_factor_domain(space
);
3775 space2
= isl_space_copy(space
);
3776 space2
= isl_space_from_domain(isl_space_domain(space
));
3777 id
= pet_expr_access_get_ref_id(expr
);
3778 space2
= isl_space_set_tuple_id(space2
, isl_dim_out
, id
);
3779 space
= isl_space_range_product(space2
, space
);
3780 space
= isl_space_uncurry(space
);
3782 return isl_map_empty(space
);
3784 isl_union_map_free(tagged
);
3788 /* Extract a gpu_stmt_access from "expr", append it to the list
3789 * that ends in *data->next_access and update the end of the list.
3790 * If the access expression performs a write, then it is considered
3791 * exact only if it appears in a single expression statement and
3792 * if its may access relation is equal to its must access relation.
3794 * The combined set of may accesses may be union if member accesses
3795 * are involved, but the entire set is derived from a single reference and
3796 * therefore from a single index expression. These accesses therefore
3797 * all map to the same outer array.
3799 static int extract_access(__isl_keep pet_expr
*expr
, void *user
)
3801 struct ppcg_extract_access_data
*data
= user
;
3802 isl_union_map
*tagged
;
3803 struct gpu_stmt_access
*access
;
3804 isl_ctx
*ctx
= pet_expr_get_ctx(expr
);
3805 isl_multi_pw_aff
*index
;
3807 access
= isl_alloc_type(ctx
, struct gpu_stmt_access
);
3809 access
->next
= NULL
;
3810 access
->read
= pet_expr_access_is_read(expr
);
3811 access
->write
= pet_expr_access_is_write(expr
);
3812 tagged
= pet_expr_access_get_tagged_may_read(expr
);
3813 tagged
= isl_union_map_union(tagged
,
3814 pet_expr_access_get_tagged_may_write(expr
));
3815 tagged
= isl_union_map_apply_range(tagged
,
3816 isl_union_map_copy(data
->any_to_outer
));
3817 if (!access
->write
) {
3818 access
->exact_write
= 1;
3819 } else if (!data
->single_expression
) {
3820 access
->exact_write
= 0;
3822 isl_union_map
*must
, *may
;
3823 may
= isl_union_map_copy(tagged
);
3824 may
= isl_union_map_domain_factor_domain(may
);
3825 must
= pet_expr_access_get_must_write(expr
);
3826 access
->exact_write
= isl_union_map_is_equal(must
, may
);
3827 isl_union_map_free(must
);
3828 isl_union_map_free(may
);
3830 index
= pet_expr_access_get_index(expr
);
3831 access
->n_index
= isl_multi_pw_aff_dim(index
, isl_dim_out
);
3832 isl_multi_pw_aff_free(index
);
3833 access
->ref_id
= pet_expr_access_get_ref_id(expr
);
3834 access
->tagged_access
= extract_single_tagged_access(tagged
, expr
);
3835 access
->access
= isl_map_copy(access
->tagged_access
);
3836 access
->access
= isl_map_domain_factor_domain(access
->access
);
3838 *data
->next_access
= access
;
3839 data
->next_access
= &(*data
->next_access
)->next
;
3841 if (!access
->access
)
3847 /* Construct a linked list of gpu_stmt_access objects,
3848 * one for each access expression in the statement body.
3849 * "any_to_outer" maps all intermediate arrays to their outer arrays.
3851 static int pet_stmt_extract_accesses(struct gpu_stmt
*stmt
,
3852 __isl_keep isl_union_map
*any_to_outer
)
3854 struct ppcg_extract_access_data data
;
3856 stmt
->accesses
= NULL
;
3857 data
.next_access
= &stmt
->accesses
;
3858 data
.single_expression
=
3859 pet_tree_get_type(stmt
->stmt
->body
) == pet_tree_expr
;
3860 data
.any_to_outer
= any_to_outer
;
3861 return pet_tree_foreach_access_expr(stmt
->stmt
->body
,
3862 &extract_access
, &data
);
3865 /* Return an array of gpu_stmt representing the statements in "scop".
3867 static struct gpu_stmt
*extract_stmts(isl_ctx
*ctx
, struct ppcg_scop
*scop
,
3868 __isl_keep isl_set
*context
, __isl_keep isl_union_map
*any_to_outer
)
3871 struct gpu_stmt
*stmts
;
3873 stmts
= isl_calloc_array(ctx
, struct gpu_stmt
, scop
->pet
->n_stmt
);
3877 for (i
= 0; i
< scop
->pet
->n_stmt
; ++i
) {
3878 struct gpu_stmt
*s
= &stmts
[i
];
3880 s
->id
= isl_set_get_tuple_id(scop
->pet
->stmts
[i
]->domain
);
3881 s
->stmt
= scop
->pet
->stmts
[i
];
3882 if (pet_stmt_extract_accesses(s
, any_to_outer
) < 0)
3883 return free_stmts(stmts
, i
+ 1);
3889 /* Callback for ppcg_print_guarded that calls the callback for generate_gpu.
3891 static __isl_give isl_printer
*print_gpu(__isl_take isl_printer
*p
, void *user
)
3893 struct gpu_gen
*gen
= user
;
3895 return gen
->print(p
, gen
->prog
, gen
->tree
, &gen
->types
,
3899 /* Generate CUDA code for "scop" and print it to "p".
3900 * After generating an AST for the transformed scop as explained below,
3901 * we call "gen->print" to print the AST in the desired output format
3904 * If it turns out that it does not make sense to generate GPU code,
3905 * then we generate CPU code instead.
3907 * The GPU code is generated in a context where at least one
3908 * statement instance is executed. The corresponding guard (if any) is printed
3909 * around the entire generated GPU code, except for the declaration
3910 * of the arrays that are visible outside of the scop and that therefore
3911 * cannot be declared inside the body of any possible guard.
3913 * We first compute a schedule that respects the dependences
3914 * of the original program and select the outermost bands
3915 * of tilable dimensions that have at least one parallel loop.
3917 * Each of these bands B is then tiled according to "tile" sizes, resulting
3918 * in two nested bands, with a kernel marker on top
3926 * We then split off at most 2 parallel dimensions from the T band and
3927 * at most 3 parallel dimension from the P band
3940 * A filter is introduced in front of T1 that maps the domain instances
3941 * to block identifiers. Similarly, a filter is introduced in front of P1
3942 * that maps the domain instances to thread identifiers.
3944 * For each iteration of the T2 band and for each array, we compute
3945 * the array elements accessed by that iteration, construct a rectangular
3946 * box around it and shift it to the origin. The result is used
3947 * as shared memory for the array.
3949 * Copying and synchronization statements are added to this schedule tree.
3950 * In principle, these are added in front of the P1 band, but some of
3951 * them may get hoisted up to higher levels.
3953 * The entire AST is then generated from the single resulting schedule tree.
3954 * During the generation the subtrees at kernel nodes (K) are saved
3955 * aside and replaced by kernel calls. The result is printed as host code
3956 * while the saved subtrees are printed as device code.
3958 static __isl_give isl_printer
*generate(__isl_take isl_printer
*p
,
3959 struct gpu_gen
*gen
, struct ppcg_scop
*scop
,
3960 struct ppcg_options
*options
)
3962 struct gpu_prog
*prog
;
3964 isl_set
*context
, *guard
;
3965 isl_schedule
*schedule
;
3969 return isl_printer_free(p
);
3971 ctx
= isl_printer_get_ctx(p
);
3972 prog
= gpu_prog_alloc(ctx
, scop
);
3974 return isl_printer_free(p
);
3976 context
= isl_set_copy(prog
->context
);
3977 guard
= isl_union_set_params(isl_union_set_copy(prog
->scop
->domain
));
3978 prog
->context
= isl_set_intersect(prog
->context
, isl_set_copy(guard
));
3981 schedule
= compute_schedule(gen
);
3983 any_permutable
= has_any_permutable_node(schedule
);
3984 if (any_permutable
< 0 || !any_permutable
) {
3985 isl_set_free(context
);
3986 isl_set_free(guard
);
3987 if (any_permutable
< 0)
3988 p
= isl_printer_free(p
);
3990 p
= print_cpu(p
, scop
, options
);
3991 isl_schedule_free(schedule
);
3993 compute_copy_in_and_out(gen
);
3994 schedule
= mark_kernels(gen
, schedule
);
3995 gen
->tree
= generate_code(gen
, schedule
);
3996 p
= ppcg_print_exposed_declarations(p
, prog
->scop
);
3997 p
= ppcg_print_guarded(p
, guard
, context
, &print_gpu
, gen
);
3998 isl_ast_node_free(gen
->tree
);
4001 gpu_prog_free(prog
);
4006 /* Wrapper around generate for use as a ppcg_transform callback.
4008 static __isl_give isl_printer
*generate_wrap(__isl_take isl_printer
*p
,
4009 struct ppcg_scop
*scop
, void *user
)
4011 struct gpu_gen
*gen
= user
;
4013 return generate(p
, gen
, scop
, gen
->options
);
4016 /* Transform the code in the file called "input" by replacing
4017 * all scops by corresponding GPU code and write the results to "out".
4019 int generate_gpu(isl_ctx
*ctx
, const char *input
, FILE *out
,
4020 struct ppcg_options
*options
,
4021 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
4022 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
4023 struct gpu_types
*types
, void *user
), void *user
)
4030 gen
.sizes
= extract_sizes_from_str(ctx
, options
->sizes
);
4031 gen
.options
= options
;
4034 gen
.print_user
= user
;
4036 gen
.types
.name
= NULL
;
4038 if (options
->debug
->dump_sizes
) {
4039 isl_space
*space
= isl_space_params_alloc(ctx
, 0);
4040 gen
.used_sizes
= isl_union_map_empty(space
);
4043 r
= ppcg_transform(ctx
, input
, out
, options
, &generate_wrap
, &gen
);
4045 if (options
->debug
->dump_sizes
) {
4046 isl_union_map_dump(gen
.used_sizes
);
4047 isl_union_map_free(gen
.used_sizes
);
4050 isl_union_map_free(gen
.sizes
);
4051 for (i
= 0; i
< gen
.types
.n
; ++i
)
4052 free(gen
.types
.name
[i
]);
4053 free(gen
.types
.name
);
4058 /* Compute the set of inner array elements that may have their values
4059 * preserved by "prog". In particular, collect the array elements of
4060 * arrays that are not local to "prog" and remove those elements that
4061 * are definitely killed or definitely written by "prog".
4063 static __isl_give isl_union_set
*compute_may_persist(struct gpu_prog
*prog
)
4066 isl_union_set
*may_persist
, *killed
;
4067 isl_union_map
*must_kill
;
4069 may_persist
= isl_union_set_empty(isl_set_get_space(prog
->context
));
4070 for (i
= 0; i
< prog
->n_array
; ++i
) {
4073 if (prog
->array
[i
].local
)
4076 extent
= isl_set_copy(prog
->array
[i
].extent
);
4077 may_persist
= isl_union_set_add_set(may_persist
, extent
);
4080 may_persist
= isl_union_set_intersect_params(may_persist
,
4081 isl_set_copy(prog
->context
));
4082 may_persist
= isl_union_set_apply(may_persist
,
4083 isl_union_map_copy(prog
->to_inner
));
4084 must_kill
= isl_union_map_copy(prog
->tagged_must_kill
);
4085 killed
= isl_union_map_range(must_kill
);
4086 must_kill
= isl_union_map_copy(prog
->must_write
);
4087 killed
= isl_union_set_union(killed
, isl_union_map_range(must_kill
));
4089 may_persist
= isl_union_set_subtract(may_persist
, killed
);
4093 struct gpu_prog
*gpu_prog_alloc(isl_ctx
*ctx
, struct ppcg_scop
*scop
)
4095 struct gpu_prog
*prog
;
4102 prog
= isl_calloc_type(ctx
, struct gpu_prog
);
4107 prog
->context
= isl_set_copy(scop
->context
);
4108 prog
->n_stmts
= scop
->pet
->n_stmt
;
4109 prog
->any_to_outer
= pet_scop_compute_outer_to_any(scop
->pet
);
4110 prog
->any_to_outer
= isl_union_map_reverse(prog
->any_to_outer
);
4111 space
= isl_union_map_get_space(prog
->any_to_outer
);
4112 space
= isl_space_set_from_params(space
);
4113 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
4114 space
= isl_space_map_from_set(space
);
4115 id
= isl_map_identity(space
);
4116 prog
->any_to_outer
= isl_union_map_add_map(prog
->any_to_outer
, id
);
4117 prog
->stmts
= extract_stmts(ctx
, scop
,
4118 prog
->context
, prog
->any_to_outer
);
4119 prog
->read
= isl_union_map_copy(scop
->reads
);
4120 prog
->may_write
= isl_union_map_copy(scop
->may_writes
);
4121 prog
->must_write
= isl_union_map_copy(scop
->must_writes
);
4122 prog
->tagged_must_kill
= isl_union_map_copy(scop
->tagged_must_kills
);
4123 prog
->to_inner
= pet_scop_compute_outer_to_inner(scop
->pet
);
4124 prog
->to_outer
= isl_union_map_copy(prog
->to_inner
);
4125 prog
->to_outer
= isl_union_map_reverse(prog
->to_outer
);
4128 return gpu_prog_free(prog
);
4130 if (collect_array_info(prog
) < 0)
4131 return gpu_prog_free(prog
);
4132 prog
->may_persist
= compute_may_persist(prog
);
4137 void *gpu_prog_free(struct gpu_prog
*prog
)
4141 free_array_info(prog
);
4142 free_stmts(prog
->stmts
, prog
->n_stmts
);
4143 isl_union_map_free(prog
->any_to_outer
);
4144 isl_union_map_free(prog
->to_outer
);
4145 isl_union_map_free(prog
->to_inner
);
4146 isl_union_set_free(prog
->copy_in
);
4147 isl_union_set_free(prog
->copy_out
);
4148 isl_union_map_free(prog
->read
);
4149 isl_union_map_free(prog
->may_write
);
4150 isl_union_map_free(prog
->must_write
);
4151 isl_union_map_free(prog
->tagged_must_kill
);
4152 isl_union_map_free(prog
->array_order
);
4153 isl_union_set_free(prog
->may_persist
);
4154 isl_set_free(prog
->context
);