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
)
410 space
= isl_space_params(isl_space_copy(array
->space
));
411 guard
= isl_set_universe(space
);
413 for (i
= 0; i
< array
->n_index
; ++i
) {
415 isl_set
*guard_i
, *zero
;
417 bound
= isl_pw_aff_copy(array
->bound
[i
]);
418 guard_i
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(bound
));
419 zero
= isl_pw_aff_zero_set(bound
);
420 guard_i
= isl_set_subtract(guard_i
, zero
);
421 guard
= isl_set_intersect(guard
, guard_i
);
427 /* Internal data structure for extract_size_of_type.
428 * "type" specifies the name of the space that we want to extract.
429 * "res" is used to store the subset of that space.
431 struct ppcg_extract_size_data
{
436 /* This function is called for each set in a union_set.
437 * If the name of the set matches data->type, we store the
440 static int extract_size_of_type(__isl_take isl_set
*size
, void *user
)
442 struct ppcg_extract_size_data
*data
= user
;
445 name
= isl_set_get_tuple_name(size
);
446 if (name
&& !strcmp(name
, data
->type
)) {
455 /* Given a union map { kernel[i] -> *[...] },
456 * return the range in the space called "type" for the kernel with
457 * sequence number "id".
459 static __isl_give isl_set
*extract_sizes(__isl_keep isl_union_map
*sizes
,
460 const char *type
, int id
)
464 isl_union_set
*local_sizes
;
465 struct ppcg_extract_size_data data
= { type
, NULL
};
470 space
= isl_union_map_get_space(sizes
);
471 space
= isl_space_set_from_params(space
);
472 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
473 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
474 dom
= isl_set_universe(space
);
475 dom
= isl_set_fix_si(dom
, isl_dim_set
, 0, id
);
477 local_sizes
= isl_union_set_apply(isl_union_set_from_set(dom
),
478 isl_union_map_copy(sizes
));
479 isl_union_set_foreach_set(local_sizes
, &extract_size_of_type
, &data
);
480 isl_union_set_free(local_sizes
);
484 /* Given a singleton set, extract the first (at most *len) elements
485 * of the single integer tuple into *sizes and update *len if needed.
487 static void read_sizes_from_set(__isl_take isl_set
*set
, int *sizes
, int *len
)
495 dim
= isl_set_dim(set
, isl_dim_set
);
499 for (i
= 0; i
< *len
; ++i
) {
502 v
= isl_set_plain_get_val_if_fixed(set
, isl_dim_set
, i
);
505 sizes
[i
] = isl_val_get_num_si(v
);
512 /* Add the map { kernel[id] -> type[sizes] } to gen->used_sizes,
513 * if the option debug->dump_sizes is set.
515 static void set_used_sizes(struct gpu_gen
*gen
, const char *type
, int id
,
522 if (!gen
->options
->debug
->dump_sizes
)
525 space
= isl_union_map_get_space(gen
->used_sizes
);
526 space
= isl_space_set_from_params(space
);
527 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
528 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
529 space
= isl_space_from_domain(space
);
530 space
= isl_space_add_dims(space
, isl_dim_out
, len
);
531 space
= isl_space_set_tuple_name(space
, isl_dim_out
, type
);
533 map
= isl_map_universe(space
);
534 map
= isl_map_fix_si(map
, isl_dim_in
, 0, id
);
535 for (i
= 0; i
< len
; ++i
)
536 map
= isl_map_fix_si(map
, isl_dim_out
, i
, sizes
[i
]);
538 gen
->used_sizes
= isl_union_map_add_map(gen
->used_sizes
, map
);
541 /* Extract user specified "tile" sizes from the "sizes" command line option,
542 * defaulting to option->tile_size in each dimension.
543 * *tile_len contains the maximum number of tile sizes needed.
544 * Update *tile_len to the number of specified tile sizes, if any, and
545 * return a pointer to the tile sizes (or NULL on error).
546 * Add the effectively used sizes to gen->used_sizes.
548 static int *read_tile_sizes(struct gpu_gen
*gen
, int *tile_len
)
554 tile_size
= isl_alloc_array(gen
->ctx
, int, *tile_len
);
557 for (n
= 0; n
< *tile_len
; ++n
)
558 tile_size
[n
] = gen
->options
->tile_size
;
560 size
= extract_sizes(gen
->sizes
, "tile", gen
->kernel_id
);
561 read_sizes_from_set(size
, tile_size
, tile_len
);
562 set_used_sizes(gen
, "tile", gen
->kernel_id
, tile_size
, *tile_len
);
567 /* Extract user specified "block" sizes from the "sizes" command line option,
568 * after filling in some potentially useful defaults.
570 static void read_block_sizes(struct ppcg_kernel
*kernel
,
571 __isl_keep isl_union_map
*sizes
)
575 if (kernel
->n_block
> 3)
577 switch (kernel
->n_block
) {
579 kernel
->block_dim
[0] = 512;
582 kernel
->block_dim
[0] = 32;
583 kernel
->block_dim
[1] = 16;
586 kernel
->block_dim
[0] = 32;
587 kernel
->block_dim
[1] = 4;
588 kernel
->block_dim
[2] = 4;
592 size
= extract_sizes(sizes
, "block", kernel
->id
);
593 read_sizes_from_set(size
, kernel
->block_dim
, &kernel
->n_block
);
596 /* Extract user specified "grid" sizes from the "sizes" command line option,
597 * after filling in some potentially useful defaults.
599 static void read_grid_sizes(struct ppcg_kernel
*kernel
,
600 __isl_keep isl_union_map
*sizes
)
604 if (kernel
->n_grid
> 2)
606 switch (kernel
->n_grid
) {
608 kernel
->grid_dim
[0] = 32768;
611 kernel
->grid_dim
[0] = 256;
612 kernel
->grid_dim
[1] = 256;
616 size
= extract_sizes(sizes
, "grid", kernel
->id
);
617 read_sizes_from_set(size
, kernel
->grid_dim
, &kernel
->n_grid
);
620 /* Extract user specified grid and block sizes from the gen->sizes
621 * command line option after filling in some potentially useful defaults.
622 * Store the extracted sizes in "kernel".
623 * Add the effectively used sizes to gen->used_sizes.
625 static void read_grid_and_block_sizes(struct ppcg_kernel
*kernel
,
628 read_block_sizes(kernel
, gen
->sizes
);
629 read_grid_sizes(kernel
, gen
->sizes
);
630 set_used_sizes(gen
, "block", kernel
->id
,
631 kernel
->block_dim
, kernel
->n_block
);
632 set_used_sizes(gen
, "grid", kernel
->id
,
633 kernel
->grid_dim
, kernel
->n_grid
);
636 static void *free_stmts(struct gpu_stmt
*stmts
, int n
)
643 for (i
= 0; i
< n
; ++i
) {
644 struct gpu_stmt_access
*access
, *next
;
646 for (access
= stmts
[i
].accesses
; access
; access
= next
) {
648 isl_id_free(access
->ref_id
);
649 isl_map_free(access
->access
);
650 isl_map_free(access
->tagged_access
);
654 isl_id_free(stmts
[i
].id
);
661 /* Add parameters p[i] with identifiers "ids" to "set",
662 * with bounds to 0 <= p[i] < size[i].
664 __isl_give isl_set
*add_bounded_parameters(__isl_take isl_set
*set
,
665 int *size
, __isl_keep isl_id_list
*ids
)
670 len
= isl_id_list_n_id(ids
);
671 nparam
= isl_set_dim(set
, isl_dim_param
);
672 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
674 for (i
= 0; i
< len
; ++i
) {
677 id
= isl_id_list_get_id(ids
, i
);
678 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
679 set
= isl_set_lower_bound_si(set
, isl_dim_param
, nparam
+ i
, 0);
680 set
= isl_set_upper_bound_si(set
, isl_dim_param
,
681 nparam
+ i
, size
[i
] - 1);
687 /* Add "len" parameters p[i] with identifiers "ids" and intersect "set"
690 * { : 0 <= p[i] < size[i] }
692 * or an overapproximation.
694 static __isl_give isl_set
*add_bounded_parameters_dynamic(
695 __isl_take isl_set
*set
, __isl_keep isl_multi_pw_aff
*size
,
696 __isl_keep isl_id_list
*ids
)
703 len
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
704 nparam
= isl_set_dim(set
, isl_dim_param
);
705 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
707 for (i
= 0; i
< len
; ++i
) {
710 id
= isl_id_list_get_id(ids
, i
);
711 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
714 space
= isl_space_params(isl_set_get_space(set
));
715 ls
= isl_local_space_from_space(space
);
716 for (i
= 0; i
< len
; ++i
) {
717 isl_pw_aff
*param
, *size_i
, *zero
;
720 param
= isl_pw_aff_var_on_domain(isl_local_space_copy(ls
),
721 isl_dim_param
, nparam
+ i
);
723 size_i
= isl_multi_pw_aff_get_pw_aff(size
, i
);
724 bound
= isl_pw_aff_lt_set(isl_pw_aff_copy(param
), size_i
);
725 bound
= isl_set_from_basic_set(isl_set_simple_hull(bound
));
726 set
= isl_set_intersect_params(set
, bound
);
728 zero
= isl_pw_aff_zero_on_domain(isl_local_space_copy(ls
));
729 bound
= isl_pw_aff_ge_set(param
, zero
);
730 set
= isl_set_intersect_params(set
, bound
);
732 isl_local_space_free(ls
);
737 /* Return the union of all tagged access relations in the group.
739 static __isl_give isl_union_map
*group_tagged_access_relation(
740 struct gpu_array_ref_group
*group
)
743 isl_union_map
*access
;
745 access
= isl_union_map_empty(isl_map_get_space(group
->access
));
746 for (i
= 0; i
< group
->n_ref
; ++i
) {
749 map_i
= isl_map_copy(group
->refs
[i
]->tagged_access
);
750 access
= isl_union_map_union(access
,
751 isl_union_map_from_map(map_i
));
757 /* Return the extent of "array", recomputed from the bounds.
758 * The recomputed extent may be simpler than the original extent.
760 static __isl_give isl_set
*array_extent(struct gpu_array_info
*array
)
768 id
= isl_set_get_tuple_id(array
->extent
);
769 space
= isl_set_get_space(array
->extent
);
770 extent
= isl_set_universe(isl_space_copy(space
));
771 ls
= isl_local_space_from_space(space
);
772 for (i
= 0; i
< array
->n_index
; ++i
) {
778 extent
= isl_set_lower_bound_si(extent
, isl_dim_set
, i
, 0);
780 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
782 index
= isl_pw_aff_from_aff(aff
);
783 bound
= isl_pw_aff_copy(array
->bound
[i
]);
784 bound
= isl_pw_aff_from_range(bound
);
785 bound
= isl_pw_aff_add_dims(bound
, isl_dim_in
, array
->n_index
);
786 bound
= isl_pw_aff_set_tuple_id(bound
, isl_dim_in
,
788 lt
= isl_pw_aff_lt_set(index
, bound
);
789 extent
= isl_set_intersect(extent
, lt
);
791 isl_local_space_free(ls
);
797 /* Return a map from the first group->depth dimensions of the computed
798 * schedule to the array tile in
799 * global memory that corresponds to the shared memory copy.
801 * In particular, return a map
807 * tile_offset(i) <= a <= tile_offset(i) + tile_size - 1 (1)
811 * 0 <= a <= array_size - 1 (2)
813 * Note that if some stride has been detected (i.e., when
814 * group->shared_tile->bound[i].shift is set), then a in (1) refers
815 * to the shifted and scaled down version.
817 * Constraints (1) are obtained by mapping the size constraints on the
818 * shared/private memory tile back to the access relation.
819 * Constraints (2) are obtained from the (recomputed) extent.
821 static __isl_give isl_map
*group_tile(struct gpu_array_ref_group
*group
)
824 int n_index
= group
->array
->n_index
;
830 space
= isl_multi_aff_get_space(group
->shared_tile
->tiling
);
831 space
= isl_space_range(space
);
832 local
= isl_set_universe(space
);
833 for (i
= 0; i
< n_index
; ++i
) {
836 local
= isl_set_lower_bound_si(local
, isl_dim_set
, i
, 0);
837 bound
= isl_val_copy(group
->shared_tile
->bound
[i
].size
);
838 bound
= isl_val_sub_ui(bound
, 1);
839 local
= isl_set_upper_bound_val(local
, isl_dim_set
, i
, bound
);
841 local
= isl_set_preimage_multi_aff(local
,
842 isl_multi_aff_copy(group
->shared_tile
->tiling
));
843 tile
= isl_set_unwrap(local
);
844 extent
= array_extent(group
->array
);
845 tile
= isl_map_intersect_range(tile
, extent
);
850 /* Given a mapping "iterator_map" from the AST schedule to a domain,
851 * return the corresponding mapping from the AST schedule to
852 * to the outer kernel->shared_schedule_dim dimensions of
853 * the schedule computed by PPCG for this kernel.
855 * Note that kernel->shared_schedule_dim is at least as large as
856 * the largest depth of any array reference group associated to the kernel.
857 * This is needed as the returned schedule is used to extract a mapping
858 * to the outer group->depth dimensions in transform_index.
860 static __isl_give isl_pw_multi_aff
*compute_sched_to_shared(
861 struct ppcg_kernel
*kernel
, __isl_take isl_pw_multi_aff
*iterator_map
)
863 isl_union_pw_multi_aff
*upma
;
864 isl_pw_multi_aff
*pma
;
867 space
= isl_space_range(isl_pw_multi_aff_get_space(iterator_map
));
868 space
= isl_space_from_domain(space
);
869 space
= isl_space_add_dims(space
, isl_dim_out
,
870 kernel
->shared_schedule_dim
);
872 upma
= isl_union_pw_multi_aff_copy(kernel
->shared_schedule
);
873 pma
= isl_union_pw_multi_aff_extract_pw_multi_aff(upma
, space
);
874 isl_union_pw_multi_aff_free(upma
);
876 return isl_pw_multi_aff_pullback_pw_multi_aff(pma
, iterator_map
);
879 /* If max_shared_memory is not set to infinity (-1), then make
880 * sure that the total amount of shared memory required by the
881 * array reference groups mapped to shared memory by "kernel"
882 * is no larger than this maximum.
884 * We apply a greedy approach and discard (keep in global memory)
885 * those groups that would result in a total memory size that
886 * is larger than the maximum.
888 * This function should be called after any function that may
889 * affect the decision on whether to place a reference group
890 * in private, shared or global memory.
892 static void check_shared_memory_bound(struct ppcg_kernel
*kernel
)
895 isl_val
*left
, *size
;
897 if (kernel
->options
->max_shared_memory
< 0)
900 left
= isl_val_int_from_si(kernel
->ctx
,
901 kernel
->options
->max_shared_memory
);
903 for (i
= 0; i
< kernel
->n_array
; ++i
) {
904 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
906 for (j
= 0; j
< local
->n_group
; ++j
) {
907 struct gpu_array_ref_group
*group
;
909 group
= local
->groups
[j
];
910 if (group
->private_tile
)
912 if (!group
->shared_tile
)
915 size
= gpu_array_tile_size(group
->shared_tile
);
916 size
= isl_val_mul_ui(size
, local
->array
->size
);
918 if (isl_val_le(size
, left
)) {
919 left
= isl_val_sub(left
, size
);
925 gpu_array_tile_free(group
->shared_tile
);
932 /* Compute a tiling for all the array reference groups in "kernel".
934 static void compute_group_tilings(struct ppcg_kernel
*kernel
)
938 for (i
= 0; i
< kernel
->n_array
; ++i
) {
939 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
941 for (j
= 0; j
< array
->n_group
; ++j
)
942 gpu_array_ref_group_compute_tiling(array
->groups
[j
]);
946 /* Compute the size of a bounding box around the origin and "set",
947 * where "set" is assumed to contain only non-negative elements.
948 * In particular, compute the maximal value of "set" in each direction
951 static __isl_give isl_multi_pw_aff
*extract_size(__isl_take isl_set
*set
,
952 __isl_take isl_set
*context
)
955 isl_multi_pw_aff
*mpa
;
957 context
= isl_set_params(context
);
958 n
= isl_set_dim(set
, isl_dim_set
);
959 mpa
= isl_multi_pw_aff_zero(isl_set_get_space(set
));
960 for (i
= 0; i
< n
; ++i
) {
965 bound
= isl_set_dim_max(isl_set_copy(set
), i
);
966 bound
= isl_pw_aff_coalesce(bound
);
967 bound
= isl_pw_aff_gist(bound
, isl_set_copy(context
));
969 space
= isl_pw_aff_get_domain_space(bound
);
970 one
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
971 one
= isl_aff_add_constant_si(one
, 1);
972 bound
= isl_pw_aff_add(bound
, isl_pw_aff_from_aff(one
));
973 mpa
= isl_multi_pw_aff_set_pw_aff(mpa
, i
, bound
);
976 isl_set_free(context
);
981 /* Compute the effective grid size as a list of the sizes in each dimension.
983 * The grid size specified by the user or set by default
984 * in read_grid_sizes() and applied by the block filter,
985 * may be too large for the given code in the sense that
986 * it may contain blocks that don't need to execute anything.
987 * We therefore don't return this grid size, but instead the
988 * smallest grid size that ensures that all blocks that actually
989 * execute code are included in the grid.
991 * We first extract a description of the grid, i.e., the possible values
992 * of the block ids, from the domain elements in "domain" and
993 * kernel->block_filter.
994 * The block ids are parameters in kernel->block_filter.
995 * We simply need to change them into set dimensions.
997 * Then, for each block dimension, we compute the maximal value of the block id
1000 static __isl_give isl_multi_pw_aff
*extract_grid_size(
1001 struct ppcg_kernel
*kernel
, __isl_take isl_union_set
*domain
)
1006 domain
= isl_union_set_intersect(domain
,
1007 isl_union_set_copy(kernel
->block_filter
));
1008 grid
= isl_union_set_params(domain
);
1009 grid
= isl_set_from_params(grid
);
1010 grid
= isl_set_add_dims(grid
, isl_dim_set
, kernel
->n_grid
);
1011 for (i
= 0; i
< kernel
->n_grid
; ++i
) {
1015 id
= isl_id_list_get_id(kernel
->block_ids
, i
);
1016 pos
= isl_set_find_dim_by_id(grid
, isl_dim_param
, id
);
1019 grid
= isl_set_equate(grid
, isl_dim_param
, pos
, isl_dim_set
, i
);
1020 grid
= isl_set_project_out(grid
, isl_dim_param
, pos
, 1);
1023 return extract_size(grid
, isl_set_copy(kernel
->context
));
1026 /* Compute the size of a fixed bounding box around the origin and "set",
1027 * where "set" is assumed to contain only non-negative elements,
1028 * and store the results in "size".
1029 * In particular, compute the maximal value of "set" in each direction
1032 static void extract_fixed_size(__isl_take isl_set
*set
, int *size
)
1035 isl_local_space
*ls
;
1038 n
= isl_set_dim(set
, isl_dim_set
);
1039 ls
= isl_local_space_from_space(isl_set_get_space(set
));
1040 obj
= isl_aff_zero_on_domain(ls
);
1041 for (i
= 0; i
< n
; ++i
) {
1044 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 1);
1045 max
= isl_set_max_val(set
, obj
);
1046 size
[i
] = isl_val_get_num_si(max
) + 1;
1048 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 0);
1054 /* Compute the effective block size as a list of the sizes in each dimension
1055 * and store the sizes in kernel->block_dim.
1057 * The block size specified by the user or set by default
1058 * in read_block_sizes() and applied by the thread filter,
1059 * may be too large for the given code in the sense that
1060 * it may contain threads that don't need to execute anything.
1061 * We therefore update this block size in kernel->block_dim
1062 * to the smallest block size that ensures that all threads
1063 * that actually execute code are included in the block.
1065 * The possible values of the thread ids is obtained from
1066 * the domain elements "domain" and kernel->thread_filter.
1067 * The current implementation eliminates all parameters, ensuring
1068 * that the size is a fixed constant in each dimension.
1069 * In principle we could also compute parametric sizes.
1070 * We would have to make sure to project out all b%d and t%d parameters,
1073 static void extract_block_size(struct ppcg_kernel
*kernel
,
1074 __isl_take isl_union_set
*domain
)
1080 domain
= isl_union_set_intersect(domain
,
1081 isl_union_set_copy(kernel
->thread_filter
));
1082 block
= isl_union_set_params(domain
);
1083 block
= isl_set_from_params(block
);
1084 block
= isl_set_add_dims(block
, isl_dim_set
, kernel
->n_block
);
1085 for (i
= 0; i
< kernel
->n_block
; ++i
) {
1089 id
= isl_id_list_get_id(kernel
->thread_ids
, i
);
1090 pos
= isl_set_find_dim_by_id(block
, isl_dim_param
, id
);
1093 block
= isl_set_equate(block
, isl_dim_param
, pos
,
1096 nparam
= isl_set_dim(block
, isl_dim_param
);
1097 block
= isl_set_project_out(block
, isl_dim_param
, 0, nparam
);
1099 extract_fixed_size(block
, kernel
->block_dim
);
1102 struct ppcg_kernel
*ppcg_kernel_free(struct ppcg_kernel
*kernel
)
1109 isl_id_list_free(kernel
->block_ids
);
1110 isl_id_list_free(kernel
->thread_ids
);
1111 isl_multi_pw_aff_free(kernel
->grid_size
);
1112 isl_set_free(kernel
->context
);
1113 isl_union_set_free(kernel
->core
);
1114 isl_union_set_free(kernel
->arrays
);
1115 isl_space_free(kernel
->space
);
1116 isl_ast_node_free(kernel
->tree
);
1117 isl_union_set_free(kernel
->block_filter
);
1118 isl_union_set_free(kernel
->thread_filter
);
1119 isl_union_pw_multi_aff_free(kernel
->shared_schedule
);
1120 isl_union_set_free(kernel
->sync_writes
);
1122 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1123 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1125 for (j
= 0; j
< array
->n_group
; ++j
)
1126 gpu_array_ref_group_free(array
->groups
[j
]);
1127 free(array
->groups
);
1129 isl_pw_aff_list_free(array
->bound
);
1131 free(kernel
->array
);
1133 for (i
= 0; i
< kernel
->n_var
; ++i
) {
1134 free(kernel
->var
[i
].name
);
1135 isl_vec_free(kernel
->var
[i
].size
);
1144 /* Wrapper around ppcg_kernel_free for use as a isl_id_set_free_user callback.
1146 static void ppcg_kernel_free_wrap(void *user
)
1148 struct ppcg_kernel
*kernel
= user
;
1150 ppcg_kernel_free(kernel
);
1153 static void create_kernel_var(isl_ctx
*ctx
, struct gpu_array_ref_group
*group
,
1154 struct ppcg_kernel_var
*var
)
1157 struct gpu_array_tile
*tile
;
1161 var
->array
= group
->array
;
1163 tile
= group
->private_tile
;
1164 var
->type
= ppcg_access_private
;
1166 tile
= group
->shared_tile
;
1167 var
->type
= ppcg_access_shared
;
1170 p
= isl_printer_to_str(ctx
);
1171 p
= gpu_array_ref_group_print_name(group
, p
);
1172 var
->name
= isl_printer_get_str(p
);
1173 isl_printer_free(p
);
1175 var
->size
= isl_vec_alloc(ctx
, group
->array
->n_index
);
1177 for (j
= 0; j
< group
->array
->n_index
; ++j
)
1178 var
->size
= isl_vec_set_element_val(var
->size
, j
,
1179 isl_val_copy(tile
->bound
[j
].size
));
1182 static int create_kernel_vars(struct ppcg_kernel
*kernel
)
1187 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1188 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1190 for (j
= 0; j
< array
->n_group
; ++j
) {
1191 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1192 if (group
->private_tile
|| group
->shared_tile
)
1198 kernel
->var
= isl_calloc_array(kernel
->ctx
, struct ppcg_kernel_var
, n
);
1203 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1204 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1206 for (j
= 0; j
< array
->n_group
; ++j
) {
1207 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1208 if (!group
->private_tile
&& !group
->shared_tile
)
1210 create_kernel_var(kernel
->ctx
, group
, &kernel
->var
[n
]);
1218 /* Replace "pa" by the zero function defined over the universe domain
1219 * in the space of "pa".
1221 static __isl_give isl_pw_aff
*set_universally_zero(__isl_take isl_pw_aff
*pa
)
1226 space
= isl_space_domain(isl_pw_aff_get_space(pa
));
1227 isl_pw_aff_free(pa
);
1228 zero
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1230 return isl_pw_aff_from_aff(zero
);
1233 /* The sizes of the arrays on the host that have been computed by
1234 * extract_array_info may depend on the parameters. Use the extra
1235 * constraints on the parameters that are valid at "host_domain"
1236 * to simplify these expressions and store the results in kernel->array.
1238 * We only need these localized bounds for arrays that are accessed
1239 * by the current kernel. If we have found at least one reference group
1240 * then the array is accessed by the kernel. If the array has compound
1241 * elements then we skipped the construction of array reference groups.
1243 * The resulting sizes may be functions that are nowhere defined
1244 * in case the access function cannot possibly access anything inside
1245 * the kernel for some reason. If so, they are replaced by the zero
1246 * function. Since the access function cannot actually access anything,
1247 * there is no harm in printing the array sizes as zero.
1249 static void localize_bounds(struct ppcg_kernel
*kernel
,
1250 __isl_keep isl_set
*host_domain
)
1255 context
= isl_set_copy(host_domain
);
1256 context
= isl_set_params(context
);
1258 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1259 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
1260 isl_pw_aff_list
*bound
;
1263 if (local
->n_group
== 0 && !local
->array
->has_compound_element
)
1266 n_index
= local
->array
->n_index
;
1267 bound
= isl_pw_aff_list_alloc(kernel
->ctx
, n_index
);
1269 for (j
= 0; j
< n_index
; ++j
) {
1273 pwaff
= isl_pw_aff_copy(local
->array
->bound
[j
]);
1274 pwaff
= isl_pw_aff_gist(pwaff
, isl_set_copy(context
));
1275 empty
= isl_pw_aff_is_empty(pwaff
);
1277 pwaff
= isl_pw_aff_free(pwaff
);
1279 pwaff
= set_universally_zero(pwaff
);
1280 bound
= isl_pw_aff_list_add(bound
, pwaff
);
1283 local
->n_index
= n_index
;
1284 local
->bound
= bound
;
1286 isl_set_free(context
);
1289 /* Create the array of gpu_local_array_info structures "array"
1290 * inside "kernel". The number of elements in this array is
1291 * the same as the number of arrays in "prog".
1292 * Initialize the "array" field of each local array to point
1293 * to the corresponding array in "prog".
1295 static struct ppcg_kernel
*ppcg_kernel_create_local_arrays(
1296 struct ppcg_kernel
*kernel
, struct gpu_prog
*prog
)
1301 ctx
= isl_set_get_ctx(prog
->context
);
1302 kernel
->array
= isl_calloc_array(ctx
,
1303 struct gpu_local_array_info
, prog
->n_array
);
1305 return ppcg_kernel_free(kernel
);
1306 kernel
->n_array
= prog
->n_array
;
1308 for (i
= 0; i
< prog
->n_array
; ++i
)
1309 kernel
->array
[i
].array
= &prog
->array
[i
];
1314 /* Find the element in gen->stmt that has the given "id".
1315 * Return NULL if no such gpu_stmt can be found.
1317 static struct gpu_stmt
*find_stmt(struct gpu_prog
*prog
, __isl_keep isl_id
*id
)
1321 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
1322 if (id
== prog
->stmts
[i
].id
)
1326 return i
< prog
->n_stmts
? &prog
->stmts
[i
] : NULL
;
1329 void ppcg_kernel_stmt_free(void *user
)
1332 struct ppcg_kernel_stmt
*stmt
= user
;
1337 switch (stmt
->type
) {
1338 case ppcg_kernel_copy
:
1339 isl_ast_expr_free(stmt
->u
.c
.index
);
1340 isl_ast_expr_free(stmt
->u
.c
.local_index
);
1342 case ppcg_kernel_domain
:
1343 isl_id_to_ast_expr_free(stmt
->u
.d
.ref2expr
);
1345 case ppcg_kernel_sync
:
1352 /* Return the gpu_stmt_access in the list "accesses" that corresponds
1355 static struct gpu_stmt_access
*find_access(struct gpu_stmt_access
*accesses
,
1356 __isl_keep isl_id
*ref_id
)
1358 struct gpu_stmt_access
*access
;
1360 for (access
= accesses
; access
; access
= access
->next
)
1361 if (access
->ref_id
== ref_id
)
1367 /* Return the index of the array called "name" in the list of arrays.
1369 static int find_array_index(struct ppcg_kernel
*kernel
, const char *name
)
1373 for (i
= 0; i
< kernel
->n_array
; ++i
)
1374 if (!strcmp(name
, kernel
->array
[i
].array
->name
))
1380 /* Internal data structure for the index and AST expression transformation
1381 * callbacks for pet_stmt_build_ast_exprs.
1383 * "kernel" is the kernel for which are computing AST expressions.
1384 * "accesses" is the list of gpu_stmt_access in the statement.
1385 * "iterator_map" expresses the statement iterators in terms of
1386 * the AST loop iterators.
1387 * "sched2shared" expresses the outer shared_schedule_dim dimensions of
1388 * the kernel schedule in terms of the AST loop iterators.
1390 * The following fields are set in transform_index and used in transform_expr.
1391 * "array" is the array that is being accessed.
1392 * "global" is set if the global array is accessed (rather than
1393 * shared/private memory).
1394 * "local_array" refers to information on the array specialized
1395 * to the current kernel.
1397 struct ppcg_transform_data
{
1398 struct ppcg_kernel
*kernel
;
1399 struct gpu_stmt_access
*accesses
;
1400 isl_pw_multi_aff
*iterator_map
;
1401 isl_pw_multi_aff
*sched2shared
;
1403 struct gpu_array_info
*array
;
1405 struct gpu_local_array_info
*local_array
;
1408 /* Return the name of the outer array (of structs) accessed by "access".
1410 static const char *get_outer_array_name(__isl_keep isl_map
*access
)
1415 space
= isl_space_range(isl_map_get_space(access
));
1416 while (space
&& isl_space_is_wrapping(space
))
1417 space
= isl_space_domain(isl_space_unwrap(space
));
1418 name
= isl_space_get_tuple_name(space
, isl_dim_set
);
1419 isl_space_free(space
);
1424 /* Return a pointer to the gpu_array_ref_group in "local"
1425 * that contains the reference "access".
1426 * Return NULL if no such group can be found.
1428 static struct gpu_array_ref_group
*find_ref_group(
1429 struct gpu_local_array_info
*local
, struct gpu_stmt_access
*access
)
1433 for (i
= 0; i
< local
->n_group
; ++i
) {
1434 struct gpu_array_ref_group
*group
= local
->groups
[i
];
1436 for (j
= 0; j
< group
->n_ref
; ++j
)
1437 if (group
->refs
[j
] == access
)
1444 /* Index transformation callback for pet_stmt_build_ast_exprs.
1446 * "index" expresses the array indices in terms of statement iterators
1448 * We first reformulate "index" in terms of the AST loop iterators.
1449 * Then we check if we are accessing the global array or
1450 * a shared/private copy. In the former case, we simply return
1451 * the updated index. If "index" is an affine expression rather
1452 * than an array access, then we also return the updated index here.
1454 * If no reference groups have been computed for the array,
1455 * then we can only be accessing the global array.
1457 * Otherwise, we apply the tiling to the index.
1458 * This tiling is of the form
1462 * where D corresponds to the outer group->depth dimensions of
1463 * the kernel schedule.
1464 * The index is of the form
1468 * We update the tiling to refer to the AST loop iterators
1472 * and modify index to keep track of those iterators
1476 * Combining these two yields a tiled index expression in terms
1477 * of the AST loop iterators
1481 static __isl_give isl_multi_pw_aff
*transform_index(
1482 __isl_take isl_multi_pw_aff
*index
, __isl_keep isl_id
*ref_id
,
1485 struct ppcg_transform_data
*data
= user
;
1486 struct gpu_stmt_access
*access
;
1487 struct gpu_array_ref_group
*group
;
1488 struct gpu_array_tile
*tile
;
1489 isl_pw_multi_aff
*iterator_map
;
1494 isl_multi_pw_aff
*tiling
;
1495 isl_pw_multi_aff
*pma
;
1496 isl_multi_pw_aff
*mpa
;
1497 isl_pw_multi_aff
*sched2depth
;
1501 iterator_map
= isl_pw_multi_aff_copy(data
->iterator_map
);
1502 index
= isl_multi_pw_aff_pullback_pw_multi_aff(index
, iterator_map
);
1504 access
= find_access(data
->accesses
, ref_id
);
1507 if (!isl_map_has_tuple_name(access
->access
, isl_dim_out
))
1510 name
= get_outer_array_name(access
->access
);
1511 i
= find_array_index(data
->kernel
, name
);
1513 isl_die(isl_multi_pw_aff_get_ctx(index
), isl_error_internal
,
1514 "cannot find array",
1515 return isl_multi_pw_aff_free(index
));
1516 data
->local_array
= &data
->kernel
->array
[i
];
1517 data
->array
= data
->local_array
->array
;
1519 group
= find_ref_group(data
->local_array
, access
);
1525 tile
= group
->private_tile
;
1527 tile
= group
->shared_tile
;
1528 data
->global
= !tile
;
1532 space
= isl_space_range(isl_multi_pw_aff_get_space(index
));
1533 space
= isl_space_map_from_set(space
);
1534 pma
= isl_pw_multi_aff_identity(space
);
1535 sched2depth
= isl_pw_multi_aff_copy(data
->sched2shared
);
1536 dim
= isl_pw_multi_aff_dim(sched2depth
, isl_dim_out
);
1537 sched2depth
= isl_pw_multi_aff_drop_dims(sched2depth
, isl_dim_out
,
1538 group
->depth
, dim
- group
->depth
);
1539 pma
= isl_pw_multi_aff_product(sched2depth
, pma
);
1540 tiling
= isl_multi_pw_aff_from_multi_aff(
1541 isl_multi_aff_copy(tile
->tiling
));
1542 tiling
= isl_multi_pw_aff_pullback_pw_multi_aff(tiling
, pma
);
1544 space
= isl_space_domain(isl_multi_pw_aff_get_space(index
));
1545 space
= isl_space_map_from_set(space
);
1546 mpa
= isl_multi_pw_aff_identity(space
);
1547 index
= isl_multi_pw_aff_range_product(mpa
, index
);
1548 index
= isl_multi_pw_aff_pullback_multi_pw_aff(tiling
, index
);
1553 /* Dereference "expr" by adding an index [0].
1554 * The original "expr" is assumed not to have any indices.
1556 * If "expr" is a member access, then the dereferencing needs
1557 * to be applied to the structure argument of this member access.
1559 static __isl_give isl_ast_expr
*dereference(__isl_take isl_ast_expr
*expr
)
1562 isl_ast_expr
*arg0
, *res
;
1563 isl_ast_expr_list
*list
;
1565 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1567 return isl_ast_expr_free(expr
);
1568 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1569 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1572 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1573 arg
= dereference(arg
);
1574 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1575 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1579 isl_ast_expr_free(arg0
);
1581 ctx
= isl_ast_expr_get_ctx(expr
);
1582 res
= isl_ast_expr_from_val(isl_val_zero(ctx
));
1583 list
= isl_ast_expr_list_from_ast_expr(res
);
1584 res
= isl_ast_expr_get_op_arg(expr
, 0);
1585 res
= isl_ast_expr_access(res
, list
);
1586 isl_ast_expr_free(expr
);
1591 /* Linearize the index expression "expr" based on the array bounds
1594 * That is, transform expression
1596 * A[i_0][i_1]...[i_n]
1600 * A[(..((i_0 * b_1 + i_1) ... ) * b_n + i_n]
1602 * where b_0, b_1, ..., b_n are the bounds on the array.
1604 * If the base of "expr" is a member access, then the linearization needs
1605 * to be applied to the structure argument of this member access.
1607 * In the base case, if "expr" has no arguments (other than the name of
1608 * the array), then we are passing an entire array to a function.
1609 * In this case, there is nothing to linearize.
1610 * Note that at this point an expression with no arguments can
1611 * only be an entire array because the scalar case and
1612 * the case of single struct are handled by the caller.
1614 * If the number of specified index expressions in "expr"
1615 * is smaller than the dimension of the accessed array,
1616 * then the missing i_j also do not appear in the linearized expression.
1617 * Furthermore, since such an expression does not refer to a single
1618 * element while the default linearized expression would refer to
1619 * a single element, we return the expression
1621 * A + (..((i_0 * b_1 + i_1) ... ) * b_n]
1623 * instead. Note that because of the special case handling above,
1624 * we can assume here that here that there is at least one index expression.
1626 __isl_give isl_ast_expr
*gpu_local_array_info_linearize_index(
1627 struct gpu_local_array_info
*array
, __isl_take isl_ast_expr
*expr
)
1634 isl_ast_expr_list
*list
;
1635 isl_ast_build
*build
;
1637 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1638 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1639 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1642 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1643 arg
= gpu_local_array_info_linearize_index(array
, arg
);
1644 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1645 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1649 isl_ast_expr_free(arg0
);
1651 if (isl_ast_expr_get_op_n_arg(expr
) == 1)
1654 ctx
= isl_ast_expr_get_ctx(expr
);
1655 context
= isl_set_universe(isl_space_params_alloc(ctx
, 0));
1656 build
= isl_ast_build_from_context(context
);
1658 n
= isl_ast_expr_get_op_n_arg(expr
);
1659 res
= isl_ast_expr_get_op_arg(expr
, 1);
1660 for (i
= 1; i
< array
->n_index
; ++i
) {
1661 isl_pw_aff
*bound_i
;
1662 isl_ast_expr
*expr_i
;
1664 bound_i
= isl_pw_aff_list_get_pw_aff(array
->bound
, i
);
1665 expr_i
= isl_ast_build_expr_from_pw_aff(build
, bound_i
);
1666 res
= isl_ast_expr_mul(res
, expr_i
);
1670 expr_i
= isl_ast_expr_get_op_arg(expr
, i
+ 1);
1671 res
= isl_ast_expr_add(res
, expr_i
);
1674 isl_ast_build_free(build
);
1676 if (1 + array
->n_index
> n
) {
1677 res
= isl_ast_expr_add(isl_ast_expr_get_op_arg(expr
, 0), res
);
1679 list
= isl_ast_expr_list_from_ast_expr(res
);
1680 res
= isl_ast_expr_get_op_arg(expr
, 0);
1681 res
= isl_ast_expr_access(res
, list
);
1684 isl_ast_expr_free(expr
);
1689 /* AST expression transformation callback for pet_stmt_build_ast_exprs.
1691 * If the AST expression refers to an array that is not accessed
1692 * at all, then this means the value of the expression is not used,
1693 * so we might as well print zero (NULL pointer) instead.
1695 * If the AST expression refers to a global scalar that is not
1696 * a read-only scalar, then its address was passed to the kernel and
1697 * we need to dereference it.
1699 * If the AST expression refers to an access to a global array,
1700 * then we linearize the access exploiting the bounds in data->local_array.
1702 static __isl_give isl_ast_expr
*transform_expr(__isl_take isl_ast_expr
*expr
,
1703 __isl_keep isl_id
*id
, void *user
)
1705 struct ppcg_transform_data
*data
= user
;
1709 if (!data
->array
->accessed
) {
1712 ctx
= isl_ast_expr_get_ctx(expr
);
1713 isl_ast_expr_free(expr
);
1714 return isl_ast_expr_from_val(isl_val_zero(ctx
));
1716 if (gpu_array_is_read_only_scalar(data
->array
))
1720 if (data
->array
->n_index
== 0)
1721 return dereference(expr
);
1722 if (!data
->array
->linearize
)
1725 return gpu_local_array_info_linearize_index(data
->local_array
, expr
);
1728 /* This function is called for each instance of a user statement
1729 * in the kernel "kernel", identified by "gpu_stmt".
1731 * We attach a struct ppcg_kernel_stmt to the "node", containing
1732 * a computed AST expression for each access.
1733 * These AST expressions are computed from iterator_map,
1734 * which expresses the domain
1735 * elements in terms of the generated loops, and sched2shared,
1736 * which expresses the outer shared_schedule_dim dimensions of
1737 * the kernel schedule computed by PPCG in terms of the generated loops.
1739 static __isl_give isl_ast_node
*create_domain_leaf(
1740 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1741 __isl_keep isl_ast_build
*build
, struct gpu_stmt
*gpu_stmt
)
1743 struct ppcg_transform_data data
;
1744 struct ppcg_kernel_stmt
*stmt
;
1746 isl_pw_multi_aff
*sched2shared
;
1748 isl_pw_multi_aff
*iterator_map
;
1749 isl_union_map
*schedule
;
1751 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1753 return isl_ast_node_free(node
);
1755 schedule
= isl_ast_build_get_schedule(build
);
1756 map
= isl_map_reverse(isl_map_from_union_map(schedule
));
1757 iterator_map
= isl_pw_multi_aff_from_map(map
);
1758 sched2shared
= compute_sched_to_shared(kernel
,
1759 isl_pw_multi_aff_copy(iterator_map
));
1761 stmt
->type
= ppcg_kernel_domain
;
1762 stmt
->u
.d
.stmt
= gpu_stmt
;
1764 data
.kernel
= kernel
;
1765 data
.accesses
= stmt
->u
.d
.stmt
->accesses
;
1766 data
.iterator_map
= iterator_map
;
1767 data
.sched2shared
= sched2shared
;
1768 stmt
->u
.d
.ref2expr
= pet_stmt_build_ast_exprs(stmt
->u
.d
.stmt
->stmt
,
1769 build
, &transform_index
, &data
,
1770 &transform_expr
, &data
);
1772 isl_pw_multi_aff_free(iterator_map
);
1773 isl_pw_multi_aff_free(sched2shared
);
1775 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1776 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1777 return isl_ast_node_set_annotation(node
, id
);
1780 /* This function is called for each statement node in the AST
1781 * for copying to or from shared/private memory.
1782 * Attach a pointer to a ppcg_kernel_stmt representing the copy
1783 * statement to the node.
1784 * The statement name is "read" or "write", depending on whether we are
1785 * reading from global memory or writing to global memory.
1787 * The schedule is of the form
1791 * where D corresponds to the outer group->depth dimensions of
1792 * the kernel schedule, A to the global array and L to the outer
1793 * generated AST schedule.
1794 * We compute the inverse and strip off the type, resulting in
1798 * We combine this mapping with on the one hand the projection
1802 * and on the other hand the group tiling
1810 * and store the corresponding expressions in stmt->index and stmt->local_index,
1811 * where stmt points to the ppcg_kernel_stmt that is attached to the node.
1813 static __isl_give isl_ast_node
*create_access_leaf(struct ppcg_kernel
*kernel
,
1814 struct gpu_array_ref_group
*group
, __isl_take isl_ast_node
*node
,
1815 __isl_keep isl_ast_build
*build
)
1817 struct ppcg_kernel_stmt
*stmt
;
1818 struct gpu_array_tile
*tile
;
1823 isl_pw_multi_aff
*pma
, *pma2
;
1826 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1828 return isl_ast_node_free(node
);
1830 access
= isl_map_from_union_map(isl_ast_build_get_schedule(build
));
1831 type
= isl_map_get_tuple_name(access
, isl_dim_in
);
1832 stmt
->u
.c
.read
= !strcmp(type
, "read");
1833 access
= isl_map_reverse(access
);
1834 pma
= isl_pw_multi_aff_from_map(access
);
1835 pma
= isl_pw_multi_aff_reset_tuple_id(pma
, isl_dim_out
);
1837 space
= isl_space_range(isl_pw_multi_aff_get_space(pma
));
1838 space
= isl_space_unwrap(space
);
1839 pma2
= isl_pw_multi_aff_range_map(space
);
1840 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
,
1841 isl_pw_multi_aff_copy(pma
));
1842 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1843 stmt
->u
.c
.index
= expr
;
1845 tile
= gpu_array_ref_group_tile(group
);
1846 pma2
= isl_pw_multi_aff_from_multi_aff(
1847 isl_multi_aff_copy(tile
->tiling
));
1848 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
, pma
);
1849 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1850 stmt
->u
.c
.local_index
= expr
;
1852 stmt
->u
.c
.array
= group
->array
;
1853 stmt
->u
.c
.local_array
= group
->local_array
;
1854 stmt
->type
= ppcg_kernel_copy
;
1856 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1857 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1858 return isl_ast_node_set_annotation(node
, id
);
1861 /* Create a synchronization ppcg_kernel_stmt and
1862 * attach it to the node "node" representing the synchronization.
1864 static __isl_give isl_ast_node
*create_sync_leaf(
1865 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1866 __isl_keep isl_ast_build
*build
)
1868 struct ppcg_kernel_stmt
*stmt
;
1871 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1873 return isl_ast_node_free(node
);
1875 stmt
->type
= ppcg_kernel_sync
;
1876 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1877 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1878 return isl_ast_node_set_annotation(node
, id
);
1881 /* Internal data structure for at_domain.
1883 * "prog" represents the entire scop.
1884 * "kernel" points to the kernel to which the current schedule node
1885 * belongs. It is set by before_mark and reset by after_mark.
1887 struct ppcg_at_domain_data
{
1888 struct gpu_prog
*prog
;
1889 struct ppcg_kernel
*kernel
;
1892 /* This function is called for each instance of a user statement
1893 * in the kernel. This may be one of the original user statements
1894 * or a statement introduced by PPCG.
1896 * We assume that the original user statements only have a name
1897 * and no user pointer. The statements introduced by PPCG
1898 * on the other hand all have a user pointer.
1900 * If the user statement is one of the original user statements
1901 * (one with no user pointer), then we call create_domain_leaf. Otherwise,
1902 * we check if it is a copy or synchronization statement and
1903 * call the appropriate functions.
1905 static __isl_give isl_ast_node
*at_domain(__isl_take isl_ast_node
*node
,
1906 __isl_keep isl_ast_build
*build
, void *user
)
1908 struct ppcg_at_domain_data
*data
= user
;
1909 isl_ast_expr
*expr
, *arg
;
1915 expr
= isl_ast_node_user_get_expr(node
);
1916 arg
= isl_ast_expr_get_op_arg(expr
, 0);
1917 id
= isl_ast_expr_get_id(arg
);
1918 name
= isl_id_get_name(id
);
1919 p
= isl_id_get_user(id
);
1920 isl_ast_expr_free(expr
);
1921 isl_ast_expr_free(arg
);
1924 struct gpu_stmt
*gpu_stmt
;
1926 gpu_stmt
= find_stmt(data
->prog
, id
);
1929 isl_die(data
->prog
->ctx
, isl_error_internal
,
1930 "statement not found",
1931 return isl_ast_node_free(node
));
1933 return create_domain_leaf(data
->kernel
, node
, build
, gpu_stmt
);
1936 is_sync
= gpu_tree_id_is_sync(id
, data
->kernel
);
1939 return isl_ast_node_free(node
);
1940 if (!strcmp(name
, "read") || !strcmp(name
, "write")) {
1941 struct gpu_array_ref_group
*group
= p
;
1942 return create_access_leaf(data
->kernel
, group
, node
, build
);
1945 isl_die(data
->prog
->ctx
, isl_error_internal
,
1946 "unknown statement type",
1947 return isl_ast_node_free(node
));
1948 return create_sync_leaf(data
->kernel
, node
, build
);
1951 /* Given a set of wrapped references "ref", return the corresponding
1952 * access relations based on the tagged access relations "tagged".
1954 * The elements of "ref" are of the form
1958 * with D an iteration domains and R a reference.
1959 * The elements of "tagged" are of the form
1965 * Extend "tagged" to include the iteration domain in the range, i.e.,
1967 * [D -> R] -> [D -> A]
1969 * apply the result to "ref" and then unwrap the resulting set
1970 * to obtain relations of the form
1974 static __isl_give isl_union_map
*wrapped_reference_to_access(
1975 __isl_take isl_union_set
*ref
, __isl_take isl_union_map
*tagged
)
1977 isl_union_map
*tag2access
;
1979 tag2access
= isl_union_map_copy(tagged
);
1980 tag2access
= isl_union_map_universe(tag2access
);
1981 tag2access
= isl_union_set_unwrap(isl_union_map_domain(tag2access
));
1982 tag2access
= isl_union_map_domain_map(tag2access
);
1983 tag2access
= isl_union_map_range_product(tag2access
, tagged
);
1985 ref
= isl_union_set_coalesce(ref
);
1986 ref
= isl_union_set_apply(ref
, tag2access
);
1988 return isl_union_set_unwrap(ref
);
1991 /* Given an access relation "access" from one or more array reference groups,
1992 * remove those reads if ("read" is 1) or writes (if "read" is 0)
1993 * that are only needed to communicate data within
1994 * the same iteration of "sched".
1995 * "tagged" contains all tagged access relations to all
1996 * the array reference groups accessed by "access" from statement
1997 * instances scheduled by "sched".
1999 * If the access is a read then it is either an element of
2001 * live_in union (range flow)
2003 * where live_in and flow may be overapproximations, or
2004 * it reads an uninitialized value (that is not live-in because
2005 * there is an intermediate kill) or it reads a value that was
2006 * written within the same (compound) statement instance.
2007 * If the access is a write then it is either an element of
2009 * live_out union (domain flow)
2011 * or it writes a value that is never read (and is not live-out
2012 * because of an intermediate kill) or only
2013 * within the same (compound) statement instance.
2014 * In both cases, the access relation is also a subset of
2015 * the group access relation.
2017 * The cases where an uninitialized value is read or a value is written
2018 * that is never read or where the dataflow occurs within a statement
2019 * instance are also considered local and may also be removed.
2021 * Essentially, we compute the intersection of "access" with either
2023 * live_in union (range non-local-flow)
2027 * live_out union (domain non-local-flow)
2029 * We first construct a relation "local"
2031 * [[D -> R] -> [D' -> R']]
2033 * of pairs of domain iterations accessing the reference group
2034 * and references in the group that are coscheduled by "sched".
2036 * If this relation does not intersect the dataflow dependences,
2037 * then there is nothing we can possibly remove, unless the dataflow
2038 * dependences themselves only relate a subset of the accesses.
2039 * In particular, the accesses may not be involved in any dataflow
2040 * dependences, either because they are uninitialized reads/dead writes
2041 * or because the dataflow occurs inside a statement instance.
2043 * Since the computation below may break up the access relation
2044 * into smaller pieces, we only perform the intersection with
2045 * the non-local dependent accesses if the local pairs
2046 * intersect the dataflow dependences. Otherwise, we intersect
2047 * with the universe of the non-local dependent accesses.
2048 * This should at least remove accesses from statements that
2049 * do not participate in any dependences.
2051 * In particular, we remove the "local" dataflow dependences from
2052 * the set of all dataflow dependences.
2053 * Note that if the potential dataflow dependences are an overapproximation
2054 * of the actual dataflow dependences, then the result remains an
2055 * overapproximation of the non-local dataflow dependences.
2056 * Copying to/from global memory is only needed for the references
2057 * in the domain/range of the result or for accesses that are live out/in
2058 * for the entire scop.
2060 * We therefore map the domain/range of the "external" relation
2061 * to the corresponding access relation and take the union with
2062 * the live out/in relation.
2064 static __isl_give isl_union_map
*remove_local_accesses(
2065 struct gpu_prog
*prog
, __isl_take isl_union_map
*tagged
,
2066 __isl_take isl_union_map
*access
, __isl_take isl_union_map
*sched
,
2070 isl_union_pw_multi_aff
*tagger
;
2071 isl_union_set
*domain
;
2072 isl_union_map
*local
, *external
;
2073 isl_union_set
*tag_set
;
2075 if (isl_union_map_is_empty(access
)) {
2076 isl_union_map_free(sched
);
2077 isl_union_map_free(tagged
);
2081 tagger
= isl_union_pw_multi_aff_copy(prog
->scop
->tagger
);
2082 domain
= isl_union_map_domain(isl_union_map_copy(tagged
));
2083 tagger
= isl_union_pw_multi_aff_intersect_domain(tagger
, domain
);
2084 sched
= isl_union_map_preimage_domain_union_pw_multi_aff(sched
, tagger
);
2086 local
= isl_union_map_apply_range(sched
,
2087 isl_union_map_reverse(isl_union_map_copy(sched
)));
2088 local
= isl_union_map_intersect(local
,
2089 isl_union_map_copy(prog
->scop
->tagged_dep_flow
));
2091 empty
= isl_union_map_is_empty(local
);
2093 external
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
2094 external
= isl_union_map_intersect_params(external
,
2095 isl_set_copy(prog
->scop
->context
));
2096 external
= isl_union_map_subtract(external
, local
);
2099 tag_set
= isl_union_map_range(external
);
2100 external
= wrapped_reference_to_access(tag_set
, tagged
);
2101 external
= isl_union_map_union(external
,
2102 isl_union_map_copy(prog
->scop
->live_in
));
2104 tag_set
= isl_union_map_domain(external
);
2105 external
= wrapped_reference_to_access(tag_set
, tagged
);
2106 external
= isl_union_map_union(external
,
2107 isl_union_map_copy(prog
->scop
->live_out
));
2111 external
= isl_union_map_free(external
);
2113 external
= isl_union_map_universe(external
);
2115 access
= isl_union_map_intersect(access
, external
);
2120 /* Given an access relation "access" from "group", remove those reads
2121 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
2122 * communicate data within the same iteration of the schedule at the
2123 * position where the copying of the group is inserted.
2124 * "node" points to this position, i.e., the depth at "node"
2125 * is equal to group->depth.
2127 * We extract a schedule that picks out the iterations of the outer
2128 * group->depth dimensions and call remove_local_accesses.
2130 static __isl_give isl_union_map
*remove_local_accesses_group(
2131 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2132 __isl_take isl_union_map
*access
, __isl_keep isl_schedule_node
*node
,
2135 isl_union_map
*sched
, *tagged
;
2137 if (isl_union_map_is_empty(access
))
2140 tagged
= group_tagged_access_relation(group
);
2141 sched
= isl_schedule_node_get_prefix_schedule_relation(node
);
2143 return remove_local_accesses(kernel
->prog
, tagged
, access
, sched
, read
);
2146 /* This function is called before the AST generator starts traversing
2147 * the schedule subtree of a node with mark "mark".
2149 * If the mark is called "kernel", store the kernel pointer in data->kernel
2150 * for use in at_domain.
2152 static int before_mark(__isl_keep isl_id
*mark
,
2153 __isl_keep isl_ast_build
*build
, void *user
)
2155 struct ppcg_at_domain_data
*data
= user
;
2159 if (!strcmp(isl_id_get_name(mark
), "kernel"))
2160 data
->kernel
= isl_id_get_user(mark
);
2164 /* This function is called after the AST generator has finished traversing
2165 * the schedule subtree of a mark node. "node" points to the corresponding
2168 * If the mark is called "kernel", then replace "node" by a user node
2169 * that "calls" the kernel, representing the launch of the kernel.
2170 * The original "node" is stored inside the kernel object so that
2171 * it can be used to print the device code.
2172 * Note that this assumes that a kernel is only launched once.
2173 * Also clear data->kernel.
2175 static __isl_give isl_ast_node
*after_mark(__isl_take isl_ast_node
*node
,
2176 __isl_keep isl_ast_build
*build
, void *user
)
2181 isl_ast_expr_list
*list
;
2182 struct ppcg_kernel
*kernel
;
2183 struct ppcg_at_domain_data
*data
= user
;
2185 ctx
= isl_ast_node_get_ctx(node
);
2186 id
= isl_ast_node_mark_get_id(node
);
2188 return isl_ast_node_free(node
);
2189 if (strcmp(isl_id_get_name(id
), "kernel") || !data
->kernel
) {
2193 kernel
= data
->kernel
;
2194 data
->kernel
= NULL
;
2195 kernel
->space
= isl_ast_build_get_schedule_space(build
);
2196 kernel
->tree
= isl_ast_node_mark_get_node(node
);
2197 isl_ast_node_free(node
);
2199 expr
= isl_ast_expr_from_id(isl_id_copy(id
));
2200 list
= isl_ast_expr_list_alloc(ctx
, 0);
2201 expr
= isl_ast_expr_call(expr
, list
);
2202 node
= isl_ast_node_alloc_user(expr
);
2203 node
= isl_ast_node_set_annotation(node
, id
);
2208 static int update_depth(__isl_keep isl_schedule_node
*node
, void *user
)
2213 if (isl_schedule_node_get_type(node
) != isl_schedule_node_leaf
)
2215 node_depth
= isl_schedule_node_get_schedule_depth(node
);
2216 if (node_depth
> *depth
)
2217 *depth
= node_depth
;
2222 /* Use isl to generate code for both the host and the device
2224 * The device code is marked by "kernel" mark nodes in the schedule tree,
2225 * containing a pointer to a ppcg_kernel object.
2226 * The returned AST only contains the AST for the host code.
2227 * The ASTs for the device code are embedded in ppcg_kernel objects
2228 * attached to the leaf nodes that call "kernel".
2230 static __isl_give isl_ast_node
*generate_code(struct gpu_gen
*gen
,
2231 __isl_take isl_schedule
*schedule
)
2233 struct ppcg_at_domain_data data
;
2234 isl_ast_build
*build
;
2236 isl_id_list
*iterators
;
2239 data
.prog
= gen
->prog
;
2243 if (isl_schedule_foreach_schedule_node(schedule
, &update_depth
,
2246 build
= isl_ast_build_alloc(gen
->prog
->ctx
);
2247 iterators
= ppcg_scop_generate_names(gen
->prog
->scop
, depth
, "c");
2248 build
= isl_ast_build_set_iterators(build
, iterators
);
2249 build
= isl_ast_build_set_at_each_domain(build
, &at_domain
, &data
);
2250 build
= isl_ast_build_set_before_each_mark(build
, &before_mark
, &data
);
2251 build
= isl_ast_build_set_after_each_mark(build
, &after_mark
, &data
);
2252 if (gen
->prog
->scop
->options
->debug
->dump_final_schedule
)
2253 isl_schedule_dump(schedule
);
2254 tree
= isl_ast_build_node_from_schedule(build
, schedule
);
2255 isl_ast_build_free(build
);
2260 __isl_give isl_union_map
*extract_sizes_from_str(isl_ctx
*ctx
, const char *str
)
2264 return isl_union_map_read_from_str(ctx
, str
);
2267 /* Can "node" be tiled and then mapped to block and thread identifiers?
2268 * That is, is it permutable with at least one coincident dimension?
2270 static int is_permutable(__isl_keep isl_schedule_node
*node
)
2275 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
2277 if (!isl_schedule_node_band_get_permutable(node
))
2279 if (isl_schedule_node_band_n_member(node
) < 1)
2281 if (!isl_schedule_node_band_member_get_coincident(node
, 0))
2287 /* A isl_schedule_foreach_schedule_node callback
2288 * for setting *any_permutable and aborting the search
2289 * if "node" is a permutable band with coincident dimensions.
2290 * Otherwise, continue searching.
2292 static int set_permutable(__isl_keep isl_schedule_node
*node
, void *user
)
2294 int *any_permutable
= user
;
2297 permutable
= is_permutable(node
);
2303 *any_permutable
= 1;
2308 /* Does "schedule" contain any permutable band with at least one coincident
2311 static int has_any_permutable_node(__isl_keep isl_schedule
*schedule
)
2313 int any_permutable
= 0;
2315 if (isl_schedule_foreach_schedule_node(schedule
, &set_permutable
,
2316 &any_permutable
) < 0 &&
2320 return any_permutable
;
2323 /* Is "node" a leaf or can it be tiled and then mapped to
2324 * block and thread identifiers?
2326 static int is_leaf_or_tilable(__isl_keep isl_schedule_node
*node
)
2328 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
2330 return is_permutable(node
);
2333 /* Is "node" the outermost node in its branch that can be tiled
2334 * and then mapped to block and thread identifiers?
2335 * If there are no such nodes in the branch and if "node" is a leaf,
2336 * then it is accepted too.
2338 static int is_outer_tilable(__isl_keep isl_schedule_node
*node
)
2341 isl_schedule_node
*ancestor
;
2343 tilable
= is_leaf_or_tilable(node
);
2350 ancestor
= isl_schedule_node_copy(node
);
2351 while (isl_schedule_node_has_parent(ancestor
)) {
2352 ancestor
= isl_schedule_node_parent(ancestor
);
2354 tilable
= is_permutable(ancestor
);
2355 if (tilable
< 0 || tilable
)
2359 isl_schedule_node_free(ancestor
);
2360 return tilable
< 0 ? -1 : !tilable
;
2363 /* Collect the references to all writes in "group".
2364 * Each reference is represented by a universe set in a space
2368 * with S[i,j] the statement instance space and R[] the array reference.
2370 static __isl_give isl_union_set
*group_tagged_writes(
2371 struct gpu_array_ref_group
*group
)
2375 isl_union_set
*writes
;
2377 space
= isl_map_get_space(group
->access
);
2378 writes
= isl_union_set_empty(space
);
2379 for (i
= 0; i
< group
->n_ref
; ++i
) {
2383 if (!group
->refs
[i
]->write
)
2386 space
= isl_map_get_space(group
->refs
[i
]->tagged_access
);
2387 space
= isl_space_domain(space
);
2388 writes_i
= isl_set_universe(space
);
2389 writes
= isl_union_set_add_set(writes
, writes_i
);
2395 /* Is there any write access in "group" that requires synchronization
2396 * on a write to global memory?
2397 * We currently take into account all writes that would require
2398 * synchronization at the thread level depth, but if the copying
2399 * for this group is performed at an outer level, then we do not
2400 * actually need to take into account dependences at intermediate levels.
2402 static int any_sync_writes_in_group(struct ppcg_kernel
*kernel
,
2403 struct gpu_array_ref_group
*group
)
2405 isl_union_set
*writes
;
2406 int empty
, disjoint
;
2408 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2414 writes
= group_tagged_writes(group
);
2415 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2416 isl_union_set_free(writes
);
2418 return disjoint
< 0 ? -1 : !disjoint
;
2421 /* Collect the references to all writes in "kernel" that write directly
2422 * to global or shared memory, i.e., that are not mapped to private memory.
2423 * Each reference is represented by a universe set in a space
2427 * with S[i,j] the statement instance space and R[] the array reference.
2429 static __isl_give isl_union_set
*collect_non_private_tagged_writes(
2430 struct ppcg_kernel
*kernel
)
2432 isl_union_set
*writes
;
2435 writes
= isl_union_set_empty(isl_union_set_get_space(kernel
->arrays
));
2437 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2438 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2440 for (j
= 0; j
< array
->n_group
; ++j
) {
2441 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2442 isl_union_set
*writes_ij
;
2446 if (group
->private_tile
)
2448 writes_ij
= group_tagged_writes(group
);
2449 writes
= isl_union_set_union(writes
, writes_ij
);
2456 /* Are there any direct writes to global memory that require
2459 static int any_global_or_shared_sync_writes(struct ppcg_kernel
*kernel
)
2461 isl_union_set
*writes
;
2462 int empty
, disjoint
;
2464 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2470 writes
= collect_non_private_tagged_writes(kernel
);
2471 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2472 isl_union_set_free(writes
);
2474 return disjoint
< 0 ? -1 : !disjoint
;
2477 /* Construct an isl_multi_val for use as tile sizes for tiling "node"
2478 * from the elements in "tile_size".
2480 static __isl_give isl_multi_val
*construct_band_tiles_sizes(
2481 __isl_keep isl_schedule_node
*node
, int *tile_size
)
2491 ctx
= isl_schedule_node_get_ctx(node
);
2492 space
= isl_schedule_node_band_get_space(node
);
2493 n
= isl_schedule_node_band_n_member(node
);
2494 mv
= isl_multi_val_zero(space
);
2495 for (i
= 0; i
< n
; ++i
) {
2498 v
= isl_val_int_from_si(ctx
, tile_size
[i
]);
2499 mv
= isl_multi_val_set_val(mv
, i
, v
);
2505 /* Replace the partial schedule S of the band node "node" by
2513 * if scale_tile_loops is set, with f the integers in "factor".
2514 * The list that "factor" points to is assumed to contain at least
2515 * as many elements as the number of members in the band.
2517 static __isl_give isl_schedule_node
*snap_band_to_sizes(
2518 __isl_take isl_schedule_node
*node
, int *factor
,
2519 struct ppcg_options
*options
)
2523 mv
= construct_band_tiles_sizes(node
, factor
);
2524 node
= isl_schedule_node_band_scale_down(node
, isl_multi_val_copy(mv
));
2525 if (options
->scale_tile_loops
)
2526 node
= isl_schedule_node_band_scale(node
,
2527 isl_multi_val_copy(mv
));
2528 isl_multi_val_free(mv
);
2533 /* Tile "band" with tile size specified by "sizes".
2535 * Since the tile loops will be mapped to block ids, we forcibly
2536 * turn off tile loop scaling. We may want to enable tile loop scaling
2537 * at some later point, but then we would have to support the detection
2538 * of strides during the mapping to block ids.
2539 * Similarly, since the point loops will be mapped to thread ids,
2540 * we forcibly shift the point loops so that they start at zero.
2542 static __isl_give isl_schedule_node
*tile_band(
2543 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2545 isl_ctx
*ctx
= isl_schedule_node_get_ctx(node
);
2549 scale_tile
= isl_options_get_tile_scale_tile_loops(ctx
);
2550 isl_options_set_tile_scale_tile_loops(ctx
, 0);
2551 shift_point
= isl_options_get_tile_shift_point_loops(ctx
);
2552 isl_options_set_tile_shift_point_loops(ctx
, 1);
2554 node
= isl_schedule_node_band_tile(node
, sizes
);
2556 isl_options_set_tile_scale_tile_loops(ctx
, scale_tile
);
2557 isl_options_set_tile_shift_point_loops(ctx
, shift_point
);
2562 /* Extract the set of parameter values and outer schedule dimensions
2563 * for which any statement instance
2564 * in the kernel inserted at "node" needs to be executed.
2565 * Intersect the set of parameter values derived from the host schedule
2566 * relation with the context of "prog".
2568 static __isl_give isl_set
*extract_context(__isl_keep isl_schedule_node
*node
,
2569 struct gpu_prog
*prog
)
2571 isl_union_map
*schedule
;
2572 isl_union_set
*schedule_domain
;
2576 schedule
= isl_schedule_node_get_prefix_schedule_relation(node
);
2577 schedule_domain
= isl_union_map_range(schedule
);
2578 empty
= isl_union_set_is_empty(schedule_domain
);
2580 isl_union_set_free(schedule_domain
);
2587 space
= isl_union_set_get_space(schedule_domain
);
2588 isl_union_set_free(schedule_domain
);
2589 space
= isl_space_set_from_params(space
);
2590 depth
= isl_schedule_node_get_schedule_depth(node
);
2591 space
= isl_space_add_dims(space
, isl_dim_set
, depth
);
2592 context
= isl_set_empty(space
);
2594 context
= isl_set_from_union_set(schedule_domain
);
2596 context
= isl_set_intersect_params(context
,
2597 isl_set_copy(prog
->context
));
2602 /* Return the set of outer array elements accessed by
2603 * by the statement instance in "domain" in "prog".
2605 static __isl_give isl_union_set
*accessed_by_domain(
2606 __isl_take isl_union_set
*domain
, struct gpu_prog
*prog
)
2608 isl_union_map
*access
;
2609 isl_union_set
*arrays
;
2611 access
= isl_union_map_union(isl_union_map_copy(prog
->read
),
2612 isl_union_map_copy(prog
->may_write
));
2613 access
= isl_union_map_intersect_domain(access
, domain
);
2614 arrays
= isl_union_map_range(access
);
2615 arrays
= isl_union_set_apply(arrays
,
2616 isl_union_map_copy(prog
->to_outer
));
2621 /* Return the number of outer band members of the band node "node"
2622 * that are marked coincident.
2624 static int n_outer_coincidence(__isl_keep isl_schedule_node
*node
)
2628 n
= isl_schedule_node_band_n_member(node
);
2630 for (i
= 0; i
< n
; ++i
)
2631 if (!isl_schedule_node_band_member_get_coincident(node
, i
))
2637 /* If the band node "node" has more than "n" members, then split off
2638 * the first "n" of them.
2640 static __isl_give isl_schedule_node
*split_band(
2641 __isl_take isl_schedule_node
*node
, int n
)
2645 dim
= isl_schedule_node_band_n_member(node
);
2647 node
= isl_schedule_node_band_split(node
, n
);
2652 /* Scale a band node that may have been split by split_band.
2653 * "sizes" are the scaling factors for the original node.
2654 * "node" either points to the original band node, or the outer
2655 * of the two pieces after splitting.
2657 * If the number of elements in "node" is smaller than the number of
2658 * elements in "sizes", then some splitting has occurred and we split
2659 * "sizes" in the same way.
2661 static __isl_give isl_schedule_node
*scale_band(
2662 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2666 n
= isl_multi_val_dim(sizes
, isl_dim_set
);
2667 dim
= isl_schedule_node_band_n_member(node
);
2669 isl_multi_val
*sizes2
;
2671 sizes2
= isl_multi_val_copy(sizes
);
2672 sizes
= isl_multi_val_drop_dims(sizes
,
2673 isl_dim_set
, dim
, n
- dim
);
2674 sizes2
= isl_multi_val_drop_dims(sizes2
, isl_dim_set
, 0, dim
);
2675 node
= isl_schedule_node_child(node
, 0);
2676 node
= isl_schedule_node_band_scale(node
, sizes2
);
2677 node
= isl_schedule_node_parent(node
);
2680 return isl_schedule_node_band_scale(node
, sizes
);
2683 /* Return an isl_multi_aff, with as elements the parameters in "space"
2684 * that have the names specified by the elements in "names".
2685 * If (some of) these parameters do not already appear in "space",
2686 * then they are added first.
2688 static __isl_give isl_multi_aff
*parameter_vector(__isl_take isl_space
*space
,
2689 __isl_keep isl_id_list
*names
)
2692 isl_local_space
*ls
;
2696 space
= isl_space_free(space
);
2698 n
= isl_id_list_n_id(names
);
2699 for (i
= 0; i
< n
; ++i
) {
2703 id
= isl_id_list_get_id(names
, i
);
2704 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2709 pos
= isl_space_dim(space
, isl_dim_param
);
2710 space
= isl_space_add_dims(space
, isl_dim_param
, 1);
2711 space
= isl_space_set_dim_id(space
, isl_dim_param
, pos
, id
);
2713 ma
= isl_multi_aff_zero(isl_space_copy(space
));
2714 ls
= isl_local_space_from_space(isl_space_domain(space
));
2715 for (i
= 0; i
< n
; ++i
) {
2720 id
= isl_id_list_get_id(names
, i
);
2721 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2723 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
2724 isl_dim_param
, pos
);
2725 ma
= isl_multi_aff_set_aff(ma
, i
, aff
);
2727 isl_local_space_free(ls
);
2732 /* Return constraints on the domain elements that equate a sequence of
2733 * parameters called "names", to the partial schedule
2734 * of "node" modulo the integers in "size".
2735 * The number of elements in the array "size" should be equal
2736 * to the number of elements in "names".
2737 * The number of members of the band node "node" should be smaller
2738 * than or equal to this number. If it is smaller, then the first
2739 * elements of "names" are equated to zero.
2741 static __isl_give isl_union_set
*set_schedule_modulo(
2742 __isl_keep isl_schedule_node
*node
, __isl_keep isl_id_list
*names
,
2748 isl_multi_union_pw_aff
*mupa
, *mupa2
;
2750 isl_union_set
*domain
;
2754 n
= isl_id_list_n_id(names
);
2756 return isl_schedule_node_get_universe_domain(node
);
2757 n_zero
= n
- isl_schedule_node_band_n_member(node
);
2759 mupa
= isl_schedule_node_band_get_partial_schedule(node
);
2760 mv
= construct_band_tiles_sizes(node
, size
+ n_zero
);
2761 mupa
= isl_multi_union_pw_aff_mod_multi_val(mupa
, mv
);
2763 space
= isl_multi_union_pw_aff_get_space(mupa
);
2764 space
= isl_space_params(space
);
2765 space
= isl_space_set_from_params(space
);
2766 space
= isl_space_add_dims(space
, isl_dim_set
, n_zero
);
2767 ma
= isl_multi_aff_zero(space
);
2769 domain
= isl_schedule_node_get_universe_domain(node
);
2770 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(
2771 isl_union_set_copy(domain
), ma
);
2772 mupa
= isl_multi_union_pw_aff_range_product(mupa2
, mupa
);
2774 space
= isl_multi_union_pw_aff_get_space(mupa
);
2775 ma
= parameter_vector(space
, names
);
2777 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(domain
, ma
);
2778 mupa
= isl_multi_union_pw_aff_sub(mupa
, mupa2
);
2780 return isl_multi_union_pw_aff_zero_union_set(mupa
);
2783 /* Insert a context node at "node" introducing the block and thread
2784 * identifiers along with their bounds, which are stored in kernel->grid_size
2785 * and kernel->block_dim.
2786 * Note that the bounds on the block identifiers may implicitly impose
2787 * constraints on the parameters. A guard needs to be inserted
2788 * in the schedule tree to ensure that those bounds hold at "node".
2789 * This guard is inserted in insert_guard.
2791 static __isl_give isl_schedule_node
*insert_context(struct ppcg_kernel
*kernel
,
2792 __isl_take isl_schedule_node
*node
)
2796 context
= isl_set_universe(isl_set_get_space(kernel
->context
));
2798 context
= add_bounded_parameters_dynamic(context
,
2799 kernel
->grid_size
, kernel
->block_ids
);
2800 context
= add_bounded_parameters(context
,
2801 kernel
->block_dim
, kernel
->thread_ids
);
2803 node
= isl_schedule_node_insert_context(node
, context
);
2808 /* Insert a guard that eliminates kernel launches where the kernel
2809 * obviously does not have any work to do.
2811 * In particular, eliminate kernel launches where there are obviously
2813 * Use the same block size constraints that are used to create the context
2814 * to ensure that all constraints implicit in the constructed context
2815 * are imposed by the guard.
2817 * Additionally, add other constraints that are valid
2818 * for each executed instance ("context"), as long as this does not result
2821 static __isl_give isl_schedule_node
*insert_guard(
2822 __isl_take isl_schedule_node
*node
, __isl_keep isl_set
*context
,
2823 __isl_keep isl_multi_pw_aff
*size
, struct ppcg_scop
*scop
)
2829 guard
= isl_set_copy(context
);
2830 guard
= isl_set_compute_divs(guard
);
2831 guard
= isl_set_from_basic_set(isl_set_simple_hull(guard
));
2833 nparam
= isl_set_dim(guard
, isl_dim_param
);
2834 n
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
2835 ids
= ppcg_scop_generate_names(scop
, n
, "__ppcg_tmp");
2836 guard
= add_bounded_parameters_dynamic(guard
, size
, ids
);
2837 isl_id_list_free(ids
);
2838 guard
= isl_set_project_out(guard
, isl_dim_param
, nparam
, n
);
2840 node
= isl_schedule_node_insert_guard(node
, guard
);
2845 /* Does any array reference group mapping require the band that is mapped
2846 * to threads to be unrolled?
2848 static int kernel_requires_unroll(struct ppcg_kernel
*kernel
)
2852 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2853 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2855 for (j
= 0; j
< array
->n_group
; ++j
) {
2856 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2857 if (gpu_array_ref_group_requires_unroll(group
))
2865 /* Mark the given band node "node" for unrolling by the AST generator and
2866 * then sink it to the leaves of the schedule tree.
2867 * All dimensions of "node" are assumed to be coincident, such that this
2868 * sinking is a valid operation.
2870 static __isl_give isl_schedule_node
*unroll(__isl_take isl_schedule_node
*node
)
2874 n
= isl_schedule_node_band_n_member(node
);
2875 for (i
= 0; i
< n
; ++i
)
2876 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
2877 isl_ast_loop_unroll
);
2879 node
= isl_schedule_node_band_sink(node
);
2884 /* Insert a synchronization node in the schedule tree of "node"
2885 * after the core computation of "kernel" at the level of the band
2886 * that is mapped to threads, except if that level is equal to
2887 * that of the band that is mapped to blocks or if there are no writes
2888 * to global or shared memory in the core computation that require
2890 * If there are any writes to shared memory and the shared memory
2891 * copying is performed at the same level, then synchronization
2892 * is needed between the core and the copying anyway, so we might
2893 * as well add it here. If the copying is performed at a higher
2894 * level, then different iterations of intermediate schedule dimensions
2895 * may have a different mapping from between shared memory elements and
2896 * threads, such that synchronization is required after the core.
2897 * "node" is assumed to point to the kernel node.
2899 static __isl_give isl_schedule_node
*add_sync(struct ppcg_kernel
*kernel
,
2900 __isl_take isl_schedule_node
*node
)
2905 need_sync
= any_global_or_shared_sync_writes(kernel
);
2907 return isl_schedule_node_free(node
);
2911 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
2913 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
2914 if (kernel_depth
== isl_schedule_node_get_schedule_depth(node
))
2915 return gpu_tree_move_up_to_kernel(node
);
2917 node
= gpu_tree_ensure_following_sync(node
, kernel
);
2919 node
= gpu_tree_move_up_to_kernel(node
);
2924 /* Return a read ("read" is 1) or write access relation for "group"
2925 * with those accesses removed that are only needed to communicate data
2926 * within the subtree of the schedule rooted at "node".
2927 * Furthermore, include the prefix schedule at "node".
2928 * That is, return a relation of the form
2932 * with D the outer schedule dimensions at "node".
2934 static __isl_give isl_union_map
*anchored_non_local_accesses(
2935 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2936 __isl_take isl_schedule_node
*node
, int read
)
2938 isl_union_map
*access
;
2939 isl_union_map
*prefix
;
2941 access
= gpu_array_ref_group_access_relation(group
, read
, !read
);
2942 access
= remove_local_accesses_group(kernel
, group
, access
, node
, read
);
2943 prefix
= isl_schedule_node_get_prefix_schedule_relation(node
);
2944 access
= isl_union_map_range_product(prefix
, access
);
2949 /* Given an array reference group "group", create a mapping
2951 * read[D -> A] -> [D -> A]
2953 * if "read" is set or
2955 * write[D -> A] -> [D -> A]
2957 * if "read" is not set.
2958 * D corresponds to the outer group->depth dimensions of
2959 * the kernel schedule.
2961 static __isl_give isl_multi_aff
*create_from_access(isl_ctx
*ctx
,
2962 struct gpu_array_ref_group
*group
, int read
)
2967 space
= isl_space_copy(group
->array
->space
);
2968 space
= isl_space_from_range(space
);
2969 space
= isl_space_add_dims(space
, isl_dim_in
, group
->depth
);
2970 space
= isl_space_wrap(space
);
2971 space
= isl_space_map_from_set(space
);
2973 id
= isl_id_alloc(ctx
, read
? "read" : "write", group
);
2974 space
= isl_space_set_tuple_id(space
, isl_dim_in
, id
);
2976 return isl_multi_aff_identity(space
);
2979 /* If any writes in "group" require synchronization, then make sure
2980 * that there is a synchronization node for "kernel" after the node
2981 * following "node" in a sequence.
2983 * If "shared" is set and no synchronization is needed for
2984 * the writes to global memory, then add synchronization before
2985 * the kernel to protect shared memory from being overwritten
2986 * by the next iteration of the core computation.
2987 * No additional synchronization is needed to protect against
2988 * the next copy into shared memory because each element of
2989 * the shared memory tile is always copied by the same thread.
2991 static __isl_give isl_schedule_node
*add_group_write_sync(
2992 __isl_take isl_schedule_node
*node
, struct ppcg_kernel
*kernel
,
2993 struct gpu_array_ref_group
*group
, int shared
)
2997 need_sync
= any_sync_writes_in_group(kernel
, group
);
2999 return isl_schedule_node_free(node
);
3001 node
= isl_schedule_node_parent(node
);
3002 node
= isl_schedule_node_next_sibling(node
);
3003 node
= isl_schedule_node_child(node
, 0);
3004 node
= gpu_tree_ensure_following_sync(node
, kernel
);
3005 } else if (shared
) {
3006 node
= isl_schedule_node_parent(node
);
3007 node
= isl_schedule_node_parent(node
);
3008 node
= gpu_tree_move_down_to_depth(node
, group
->depth
,
3010 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3016 /* Add copy statements to the schedule tree of "node"
3017 * for reading from global memory to private memory (if "read" is set) or
3018 * for writing back from private memory to global memory
3019 * (if "read" is not set) for the array reference group "group" that
3020 * is mapped to private memory.
3021 * On input, "node" points to the kernel node, and it is moved
3022 * back there on output.
3024 * The copies are performed in the order of the array elements.
3025 * The copy statement instances include a reference to the outer
3026 * group->depth dimensions of the kernel schedule for ease of
3027 * combining them with the group tiling.
3029 * That is, the extra schedule is of the form
3033 * where D corresponds to the outer group->depth dimensions of
3034 * the kernel schedule and A to the global array.
3035 * This schedule is unrolled because registers are not addressable.
3037 * The copying is inserted in the schedule tree through an extension
3042 * where the extra domain elements type[D -> A] are those accessed
3044 * A filter is inserted on type[D -> A] to ensure that the element
3045 * is read/written by the same thread that needs the element.
3046 * This filter is obtained by applying
3050 * to the thread filter for the core statements.
3052 * The extension is inserted before the core computation in case of a read
3053 * and after the core computation in case of a write.
3054 * In the latter case, we also make sure that there is a synchronization
3055 * node after the write to global memory, unless this write is performed
3056 * at the outer level of the kernel.
3057 * In principle, this synchronization could be inserted higher
3058 * in the schedule tree depending on where the corresponding reads
3059 * from global memory are performed.
3061 static __isl_give isl_schedule_node
*add_copies_group_private(
3062 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3063 __isl_take isl_schedule_node
*node
, int read
)
3065 isl_union_map
*access
;
3066 isl_union_map
*prefix
;
3067 isl_union_set
*domain
;
3069 isl_multi_aff
*from_access
;
3070 isl_multi_pw_aff
*mpa
;
3071 isl_multi_union_pw_aff
*mupa
;
3072 isl_schedule_node
*graft
;
3073 isl_union_set
*filter
;
3077 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3078 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3080 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3081 empty
= isl_union_map_is_empty(access
);
3082 if (empty
< 0 || empty
) {
3083 isl_union_map_free(access
);
3085 return isl_schedule_node_free(node
);
3086 return gpu_tree_move_up_to_kernel(node
);
3089 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3090 space
= isl_space_domain(isl_multi_aff_get_space(from_access
));
3091 access
= isl_union_map_preimage_range_multi_aff(access
, from_access
);
3093 filter
= isl_union_set_copy(kernel
->thread_filter
);
3094 filter
= isl_union_set_apply(filter
, isl_union_map_copy(access
));
3095 filter
= isl_union_set_detect_equalities(filter
);
3096 filter
= isl_union_set_coalesce(filter
);
3098 domain
= isl_union_map_range(access
);
3099 access
= isl_union_set_wrapped_domain_map(domain
);
3100 access
= isl_union_map_reverse(access
);
3101 access
= isl_union_map_coalesce(access
);
3102 graft
= isl_schedule_node_from_extension(access
);
3104 space
= isl_space_map_from_set(space
);
3105 mpa
= isl_multi_pw_aff_identity(space
);
3106 mpa
= isl_multi_pw_aff_range_factor_range(mpa
);
3107 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3109 graft
= isl_schedule_node_child(graft
, 0);
3110 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3111 graft
= unroll(graft
);
3113 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3115 graft
= isl_schedule_node_parent(graft
);
3118 node
= isl_schedule_node_graft_before(node
, graft
);
3120 node
= isl_schedule_node_graft_after(node
, graft
);
3121 if (kernel_depth
< group
->depth
)
3122 node
= add_group_write_sync(node
, kernel
, group
, 0);
3125 node
= gpu_tree_move_up_to_kernel(node
);
3130 /* Add copy statements to the schedule tree of "node"
3131 * for reading from global memory to shared memory (if "read" is set) or
3132 * for writing back from shared memory to global memory
3133 * (if "read" is not set) for the array reference group "group" that
3134 * is mapped to shared memory.
3135 * On input, "node" points to the kernel node, and it is moved
3136 * back there on output.
3138 * The copies are performed in the order of the corresponding shared
3140 * The copy statement instances include a reference to the outer
3141 * group->depth dimensions of the kernel schedule for ease of
3142 * combining them with the group tiling.
3144 * If we are performing a read from global memory to shared memory and
3145 * if the array involved is not a scalar, then we copy
3146 * the entire tile to shared memory. This may result in some extra
3147 * elements getting copied, but it should lead to simpler code
3148 * (which means that fewer registers may be needed) and less divergence.
3150 * Otherwise, we only copy the elements that will be read or have been written
3153 * That is, the extra schedule is of the form
3157 * where D corresponds to the outer group->depth dimensions of
3158 * the kernel schedule, A to the global array and T is the corresponding
3159 * shared memory tile.
3161 * The copying is inserted in the schedule tree through an extension
3166 * where the extra domain elements type[D -> A] are those accessed
3167 * by the group. In the case of read from a non-scalar, this set
3168 * is replaced by the entire shared memory tile.
3170 * A filter is inserted on type[D -> A] to map the copy instances
3171 * to the threads. In particular, the thread identifiers are
3172 * equated to the position inside the shared memory tile (T)
3173 * modulo the block size.
3174 * We try to align the innermost tile dimension with the innermost
3175 * thread identifier (x) as a heuristic to improve coalescing.
3176 * In particular, if the dimension of the tile is greater than
3177 * the dimension of the block, then the schedule mapping to the tile
3178 * is broken up into two pieces and the filter is applied to the inner part.
3179 * If, on the other hand, the dimension of the tile is smaller than
3180 * the dimension of the block, then the initial thread identifiers
3181 * are equated to zero and the remaining thread identifiers are
3182 * matched to the memory tile.
3184 * The extension is inserted before the core computation in case of a read
3185 * and after the core computation in case of a write.
3186 * In the case of a read, we first need to make sure there is some
3187 * synchronization before the core computation such that we can put the read
3188 * from global memory to shared memory before that synchronization.
3189 * This ensures that all threads have finished copying into shared memory
3190 * before the shared memory is used.
3191 * We also need to make sure that there is a synchronization node after
3192 * the core computation to ensure that the next load into shared memory
3193 * only happens after all data has been used. There is no need for
3194 * this synchronization if we are at the outer level since then there
3195 * won't be a next load.
3196 * In the case of a write, we need to make sure there is some synchronization
3197 * after the core computation such taht we can put the write from shared
3198 * memory to global memory after that synchronization.
3199 * Unless we are at the outer level, we also need a synchronization node
3200 * after the write to ensure the data is saved to global memory
3201 * before the next iteration write to the same shared memory.
3202 * It also makes sure the data has arrived in global memory before
3203 * it is read in a subsequent iteration.
3205 static __isl_give isl_schedule_node
*add_copies_group_shared(
3206 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3207 __isl_take isl_schedule_node
*node
, int read
)
3209 struct gpu_array_tile
*tile
;
3210 isl_union_map
*access
;
3211 isl_union_set
*domain
;
3212 isl_union_set
*sync
;
3214 isl_multi_aff
*from_access
;
3215 isl_multi_pw_aff
*mpa
;
3216 isl_multi_union_pw_aff
*mupa
;
3217 isl_schedule_node
*graft
;
3218 isl_union_set
*filter
;
3223 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3224 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3226 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3227 empty
= isl_union_map_is_empty(access
);
3228 if (empty
< 0 || empty
) {
3229 isl_union_map_free(access
);
3231 return isl_schedule_node_free(node
);
3232 return gpu_tree_move_up_to_kernel(node
);
3235 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3237 tile
= gpu_array_ref_group_tile(group
);
3238 ma
= isl_multi_aff_copy(tile
->tiling
);
3239 ma
= isl_multi_aff_pullback_multi_aff(ma
,
3240 isl_multi_aff_copy(from_access
));
3241 mpa
= isl_multi_pw_aff_from_multi_aff(ma
);
3242 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3244 domain
= isl_union_map_range(access
);
3246 if (read
&& !gpu_array_is_scalar(group
->array
)) {
3248 isl_union_set_free(domain
);
3249 map
= group_tile(group
);
3250 domain
= isl_union_set_from_set(isl_map_wrap(map
));
3253 domain
= isl_union_set_preimage_multi_aff(domain
, from_access
);
3254 access
= isl_union_set_wrapped_domain_map(domain
);
3255 access
= isl_union_map_reverse(access
);
3256 access
= isl_union_map_coalesce(access
);
3257 graft
= isl_schedule_node_from_extension(access
);
3259 graft
= isl_schedule_node_child(graft
, 0);
3261 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3263 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0) {
3264 graft
= isl_schedule_node_band_split(graft
,
3265 tile
->n
- kernel
->n_block
);
3266 graft
= isl_schedule_node_child(graft
, 0);
3268 if (tile
->n
< kernel
->n_block
)
3269 skip
= kernel
->n_block
- tile
->n
;
3272 filter
= set_schedule_modulo(graft
, kernel
->thread_ids
,
3274 if (!kernel
->options
->wrap
)
3275 graft
= snap_band_to_sizes(graft
, kernel
->block_dim
+ skip
,
3277 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0)
3278 graft
= isl_schedule_node_parent(graft
);
3279 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3281 while (graft
&& isl_schedule_node_has_parent(graft
))
3282 graft
= isl_schedule_node_parent(graft
);
3285 if (kernel_depth
< group
->depth
)
3286 node
= gpu_tree_ensure_sync_after_core(node
, kernel
);
3287 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3288 node
= isl_schedule_node_graft_before(node
, graft
);
3290 node
= gpu_tree_move_right_to_sync(node
, kernel
);
3291 node
= isl_schedule_node_graft_after(node
, graft
);
3292 if (kernel_depth
< group
->depth
)
3293 node
= add_group_write_sync(node
, kernel
, group
, 1);
3296 node
= gpu_tree_move_up_to_kernel(node
);
3301 /* Check whether the array reference group "group" is mapped to
3302 * private or shared memory and, if so,
3303 * add copy statements to the schedule tree of "node"
3304 * for reading from global memory to private or shared memory
3305 * (if "read" is set) or for writing back from private or shared memory
3306 * to global memory (if "read" is not set) for this group.
3307 * On input, "node" points to the kernel node, and it is moved
3308 * back there on output.
3310 static __isl_give isl_schedule_node
*add_copies_group(
3311 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3312 __isl_take isl_schedule_node
*node
, int read
)
3314 if (group
->private_tile
)
3315 return add_copies_group_private(kernel
, group
, node
, read
);
3316 if (group
->shared_tile
)
3317 return add_copies_group_shared(kernel
, group
, node
, read
);
3321 /* For each array reference group that is mapped to private or shared memory,
3322 * add copy statements to the schedule tree of "node"
3323 * for reading from global memory to private or shared memory
3324 * and for writing back.
3325 * On input, "node" points to the kernel node, and it is moved
3326 * back there on output.
3328 static __isl_give isl_schedule_node
*add_copies(struct ppcg_kernel
*kernel
,
3329 __isl_take isl_schedule_node
*node
)
3333 for (i
= 0; i
< kernel
->n_array
; ++i
) {
3334 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
3336 for (j
= 0; j
< array
->n_group
; ++j
) {
3337 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3339 node
= add_copies_group(kernel
, group
, node
, 1);
3342 node
= add_copies_group(kernel
, group
, node
, 0);
3351 /* Mark all dimensions in the current band node atomic.
3353 static __isl_give isl_schedule_node
*atomic(__isl_take isl_schedule_node
*node
)
3357 n
= isl_schedule_node_band_n_member(node
);
3358 for (i
= 0; i
< n
; ++i
)
3359 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
3360 isl_ast_loop_atomic
);
3365 /* Mark "node" atomic, if it is a band node.
3366 * Do the same for all ancestors.
3367 * Return a pointer to "node" (in the updated schedule tree).
3369 static __isl_give isl_schedule_node
*atomic_ancestors(
3370 __isl_take isl_schedule_node
*node
)
3376 if (!isl_schedule_node_has_parent(node
))
3379 pos
= isl_schedule_node_get_child_position(node
);
3380 node
= isl_schedule_node_parent(node
);
3381 if (isl_schedule_node_get_type(node
) == isl_schedule_node_band
)
3382 node
= atomic(node
);
3383 node
= atomic_ancestors(node
);
3384 node
= isl_schedule_node_child(node
, pos
);
3389 /* Collect all write references that require synchronization.
3390 * "node" is assumed to point to the kernel node.
3391 * Each reference is represented by a universe set in a space
3395 * with S[i,j] the statement instance space and R[] the array reference.
3397 * This function should be called before block and thread filters are added.
3399 * Synchronization is needed after a write if there is a subsequent read
3400 * within the same block that may not be performed by the same thread.
3401 * There should not be any dependences between different blocks,
3402 * so we start with the flow dependences within the same kernel invocation
3403 * and we subtract from these those dependences that are mapped
3404 * to the same iteration of the bands where synchronization is inserted.
3405 * We do not remove pairs of instances that are known to map to
3406 * the same thread across different iterations of the intermediate
3407 * bands because the read may be performed by a different thread
3408 * than the one that needs the value if shared memory is involved.
3410 * We also consider all pairs of possible writes that access the same
3411 * memory location and that may be mapped to the same block but not
3412 * to the same iteration of the intermediate bands.
3413 * In theory, it would be possible for one thread to still be in
3414 * a previous iteration of a loop in these bands.
3415 * A write to global memory in this delayed thread could then overwrite
3416 * a write from another thread that has already moved on to
3417 * the next iteration.
3419 * After computing the above writes paired off with reads or writes
3420 * that depend on them, we project onto the domain writes.
3421 * Sychronization is needed after writes to global memory
3422 * through these references.
3424 static __isl_give isl_union_set
*compute_sync_writes(
3425 struct ppcg_kernel
*kernel
, __isl_keep isl_schedule_node
*node
)
3427 isl_union_map
*local
;
3428 isl_union_map
*may_writes
, *shared_access
;
3429 isl_union_map
*kernel_prefix
, *thread_prefix
;
3430 isl_union_map
*equal
;
3431 isl_union_set
*wrap
;
3432 isl_union_set
*domain
;
3434 domain
= isl_schedule_node_get_universe_domain(node
);
3435 kernel_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3436 node
= isl_schedule_node_copy(node
);
3437 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3438 thread_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3439 isl_schedule_node_free(node
);
3441 may_writes
= isl_union_map_copy(kernel
->prog
->scop
->tagged_may_writes
);
3442 may_writes
= isl_union_map_curry(may_writes
);
3443 may_writes
= isl_union_map_intersect_domain(may_writes
, domain
);
3444 may_writes
= isl_union_map_uncurry(may_writes
);
3445 shared_access
= isl_union_map_copy(may_writes
);
3446 shared_access
= isl_union_map_apply_range(shared_access
,
3447 isl_union_map_reverse(may_writes
));
3449 local
= isl_union_map_copy(kernel
->prog
->scop
->tagged_dep_flow
);
3450 local
= isl_union_map_union(local
, shared_access
);
3451 local
= isl_union_map_zip(local
);
3453 equal
= isl_union_map_apply_range(kernel_prefix
,
3454 isl_union_map_reverse(isl_union_map_copy(kernel_prefix
)));
3455 wrap
= isl_union_map_wrap(equal
);
3456 local
= isl_union_map_intersect_domain(local
, wrap
);
3457 equal
= isl_union_map_apply_range(thread_prefix
,
3458 isl_union_map_reverse(isl_union_map_copy(thread_prefix
)));
3459 wrap
= isl_union_map_wrap(equal
);
3460 local
= isl_union_map_subtract_domain(local
, wrap
);
3462 local
= isl_union_map_zip(local
);
3463 local
= isl_union_map_universe(local
);
3465 return isl_union_map_domain(local
);
3468 /* Group the domain elements into a single space, named kernelX,
3469 * with X the kernel sequence number "kernel_id".
3471 static __isl_give isl_schedule_node
*group_statements(
3472 __isl_take isl_schedule_node
*node
, int kernel_id
)
3480 snprintf(buffer
, sizeof(buffer
), "kernel%d", kernel_id
);
3481 id
= isl_id_alloc(isl_schedule_node_get_ctx(node
), buffer
, NULL
);
3482 return isl_schedule_node_group(node
, id
);
3485 /* Create a ppcg_kernel representing the domain instances that reach "node"
3486 * and insert a mark node pointing to the ppcg_kernel before "node".
3487 * The band that "node" points to is the band that needs to be mapped
3488 * to block identifiers. The band that needs to be mapped to thread
3489 * identifiers should be marked by a "thread" mark by the caller.
3490 * This mark is removed by this function.
3491 * If "scale" is set, then the band that "node" points to is scaled
3494 * Mark all outer band nodes as atomic to ensure each kernel is only
3496 * If the domain elements that reach "node" live in more than one space,
3497 * then group the domain elements into a single space, named kernelX,
3498 * with X the kernel sequence number.
3500 * Insert a guard node governing the kernel node to ensure that
3501 * no kernels with zero blocks are launched.
3503 * Insert a context node describing the block and thread
3504 * identifiers inside the kernel mark.
3505 * The context node needs to be inserted after the effective block size
3506 * has been determined such that the bounds on the thread identifiers
3507 * would reflect the effective block size.
3508 * Insert a filter node inside the context node mapping the statement
3509 * instances to block identifiers. In particular, the block identifiers
3510 * are equated to the partial schedule of band that was marked for mapping
3511 * to blocks modulo the grid size.
3512 * Insert a filter node inside the "thread" mark mapping the statement
3513 * instances to thread identifiers. In particular, the thread identifiers
3514 * are equated to the partial schedule of band that was marked for mapping
3515 * to threads modulo the block size.
3517 * Compute array reference groups for all arrays, set the local
3518 * array bounds based on the set of domain instances that reach
3519 * the kernel node, check the total amount of shared memory used
3520 * and compute all group tilings.
3521 * The array reference groups are computed after the block filter
3522 * has been inserted because it affects the mapping to shared or
3523 * private memory. This computation also requires the thread filter
3524 * (in the ppcg_kernel object), but this thread filter should not
3525 * have been added to the schedule tree yet since the computation
3526 * requires the schedule of the band that needs to be mapped to
3527 * threads before the privatization is applied.
3529 * If any array reference group requires the band mapped to threads
3530 * to be unrolled, then we perform the required unrolling.
3532 * We save a copy of the schedule that may influence the mappings
3533 * to shared or private memory in kernel->shared_schedule.
3535 * Finally, we add synchronization and copy statements to the schedule tree,
3536 * remove the "thread" mark and create representations for the local
3537 * variables in the kernel.
3539 * We keep a copy of the isl_id that points to the kernel to ensure
3540 * that the kernel does not get destroyed if the schedule node
3541 * is freed due to some error condition.
3543 static __isl_give isl_schedule_node
*create_kernel(struct gpu_gen
*gen
,
3544 __isl_take isl_schedule_node
*node
, int scale
,
3545 __isl_keep isl_multi_val
*sizes
)
3547 struct ppcg_kernel
*kernel
;
3549 isl_schedule_node
*node_thread
;
3550 isl_union_map
*host_schedule
;
3551 isl_set
*host_domain
;
3552 isl_union_set
*domain
;
3553 int single_statement
;
3555 kernel
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel
);
3556 kernel
= ppcg_kernel_create_local_arrays(kernel
, gen
->prog
);
3558 return isl_schedule_node_free(node
);
3560 domain
= isl_schedule_node_get_domain(node
);
3561 single_statement
= isl_union_set_n_set(domain
) == 1;
3563 kernel
->ctx
= gen
->ctx
;
3564 kernel
->prog
= gen
->prog
;
3565 kernel
->options
= gen
->options
;
3566 kernel
->context
= extract_context(node
, gen
->prog
);
3567 kernel
->core
= isl_union_set_universe(isl_union_set_copy(domain
));
3568 kernel
->arrays
= accessed_by_domain(isl_union_set_copy(domain
),
3570 kernel
->n_grid
= n_outer_coincidence(node
);
3571 node_thread
= isl_schedule_node_copy(node
);
3572 node_thread
= gpu_tree_move_down_to_thread(node_thread
, kernel
->core
);
3573 node_thread
= isl_schedule_node_child(node_thread
, 0);
3574 kernel
->n_block
= n_outer_coincidence(node_thread
);
3575 isl_schedule_node_free(node_thread
);
3576 kernel
->id
= gen
->kernel_id
++;
3577 read_grid_and_block_sizes(kernel
, gen
);
3579 kernel
->sync_writes
= compute_sync_writes(kernel
, node
);
3581 host_schedule
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3582 host_domain
= isl_set_from_union_set(isl_union_map_range(
3585 node
= atomic_ancestors(node
);
3587 id
= isl_id_alloc(gen
->ctx
, "kernel", kernel
);
3588 id
= isl_id_set_free_user(id
, &ppcg_kernel_free_wrap
);
3589 node
= isl_schedule_node_insert_mark(node
, isl_id_copy(id
));
3591 if (!single_statement
)
3592 node
= group_statements(node
, kernel
->id
);
3594 node
= isl_schedule_node_child(node
, 0);
3595 node
= split_band(node
, kernel
->n_grid
);
3596 kernel
->block_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3597 kernel
->n_grid
, "b");
3598 kernel
->block_filter
= set_schedule_modulo(node
, kernel
->block_ids
,
3600 kernel
->grid_size
= extract_grid_size(kernel
,
3601 isl_union_set_copy(domain
));
3602 if (!kernel
->options
->wrap
)
3603 node
= snap_band_to_sizes(node
, kernel
->grid_dim
,
3606 node
= scale_band(node
, isl_multi_val_copy(sizes
));
3607 node
= isl_schedule_node_parent(node
);
3608 if (!single_statement
)
3609 node
= isl_schedule_node_parent(node
);
3610 node
= insert_guard(node
, kernel
->context
, kernel
->grid_size
,
3612 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3613 node
= isl_schedule_node_child(node
, 0);
3614 node
= split_band(node
, kernel
->n_block
);
3615 kernel
->thread_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3616 kernel
->n_block
, "t");
3617 kernel
->thread_filter
= set_schedule_modulo(node
, kernel
->thread_ids
,
3619 extract_block_size(kernel
, domain
);
3621 node
= gpu_tree_move_up_to_kernel(node
);
3622 node
= isl_schedule_node_child(node
, 0);
3623 node
= insert_context(kernel
, node
);
3624 node
= isl_schedule_node_child(node
, 0);
3625 node
= isl_schedule_node_insert_filter(node
,
3626 isl_union_set_copy(kernel
->block_filter
));
3628 node
= gpu_tree_move_up_to_kernel(node
);
3630 if (gpu_group_references(kernel
, node
) < 0)
3631 node
= isl_schedule_node_free(node
);
3632 localize_bounds(kernel
, host_domain
);
3633 isl_set_free(host_domain
);
3635 check_shared_memory_bound(kernel
);
3636 compute_group_tilings(kernel
);
3638 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3639 node
= isl_schedule_node_child(node
, 0);
3640 if (!kernel
->options
->wrap
)
3641 node
= snap_band_to_sizes(node
, kernel
->block_dim
,
3643 node
= isl_schedule_node_insert_filter(node
,
3644 isl_union_set_copy(kernel
->thread_filter
));
3645 if (kernel_requires_unroll(kernel
)) {
3646 node
= isl_schedule_node_child(node
, 0);
3647 node
= unroll(node
);
3650 node
= gpu_tree_move_up_to_thread(node
);
3651 kernel
->shared_schedule_dim
=
3652 isl_schedule_node_get_schedule_depth(node
);
3653 kernel
->shared_schedule
=
3654 isl_schedule_node_get_prefix_schedule_union_pw_multi_aff(node
);
3656 node
= gpu_tree_move_up_to_kernel(node
);
3658 node
= add_sync(kernel
, node
);
3659 node
= add_copies(kernel
, node
);
3661 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3662 node
= isl_schedule_node_delete(node
);
3664 node
= gpu_tree_move_up_to_kernel(node
);
3666 if (create_kernel_vars(kernel
) < 0)
3667 node
= isl_schedule_node_free(node
);
3669 if (!single_statement
)
3670 node
= isl_schedule_node_parent(node
);
3671 node
= isl_schedule_node_parent(node
);
3677 /* Insert a zero-dimensional permutable band at "node".
3679 static __isl_give isl_schedule_node
*insert_empty_permutable_band(
3680 __isl_take isl_schedule_node
*node
)
3683 isl_schedule
*schedule
;
3684 isl_union_set
*domain
;
3685 isl_multi_union_pw_aff
*mupa
;
3687 schedule
= isl_schedule_node_get_schedule(node
);
3688 domain
= isl_schedule_get_domain(schedule
);
3689 space
= isl_union_set_get_space(domain
);
3690 isl_union_set_free(domain
);
3691 isl_schedule_free(schedule
);
3693 space
= isl_space_set_from_params(space
);
3694 mupa
= isl_multi_union_pw_aff_zero(space
);
3695 node
= isl_schedule_node_insert_partial_schedule(node
, mupa
);
3696 node
= isl_schedule_node_band_set_permutable(node
, 1);
3701 /* If "node" is the outermost permutable band that can be mapped to block and
3702 * thread identifiers in its branch (or a leaf with no such outer bands),
3703 * then mark the band as such, attaching a ppcg_kernel to the mark.
3705 * If "node" originally points to a leaf, then insert a zero-dimensional
3706 * permutable band such that we can assume that "node" always
3707 * points to a band node.
3709 * Tile "node" using user specified tile sizes, after splitting the band
3710 * if the number of specified tile sizes is smaller than the dimension
3711 * of the band. Mark the point band of this tiling as the band that
3712 * needs to be mapped to threads.
3713 * Create a kernel representing the domain instances that reach "node" and
3714 * insert a mark node pointing to the ppcg_kernel before the band node.
3716 static __isl_give isl_schedule_node
*mark_outer_permutable(
3717 __isl_take isl_schedule_node
*node
, void *user
)
3719 struct gpu_gen
*gen
= user
;
3725 isl_multi_val
*sizes
;
3727 outer
= is_outer_tilable(node
);
3729 return isl_schedule_node_free(node
);
3733 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
3734 node
= insert_empty_permutable_band(node
);
3736 tile_len
= isl_schedule_node_band_n_member(node
);
3737 tile_size
= read_tile_sizes(gen
, &tile_len
);
3739 return isl_schedule_node_free(node
);
3740 if (tile_len
< isl_schedule_node_band_n_member(node
))
3741 node
= isl_schedule_node_band_split(node
, tile_len
);
3742 sizes
= construct_band_tiles_sizes(node
, tile_size
);
3743 node
= tile_band(node
, isl_multi_val_copy(sizes
));
3744 node
= isl_schedule_node_child(node
, 0);
3745 id
= isl_id_alloc(gen
->ctx
, "thread", NULL
);
3746 node
= isl_schedule_node_insert_mark(node
, id
);
3747 node
= isl_schedule_node_parent(node
);
3749 scale
= gen
->options
->scale_tile_loops
;
3750 node
= create_kernel(gen
, node
, scale
, sizes
);
3751 isl_multi_val_free(sizes
);
3757 /* Replace any reference to an array element in the range of "copy"
3758 * by a reference to all array elements (defined by the extent of the array).
3760 static __isl_give isl_union_map
*approximate_copy_out(
3761 __isl_take isl_union_map
*copy
, struct gpu_prog
*prog
)
3766 res
= isl_union_map_empty(isl_union_map_get_space(copy
));
3768 for (i
= 0; i
< prog
->n_array
; ++i
) {
3771 isl_union_map
*copy_i
;
3772 isl_union_set
*extent
, *domain
;
3774 space
= isl_space_copy(prog
->array
[i
].space
);
3775 extent
= isl_union_set_from_set(isl_set_universe(space
));
3776 copy_i
= isl_union_map_copy(copy
);
3777 copy_i
= isl_union_map_intersect_range(copy_i
, extent
);
3778 set
= isl_set_copy(prog
->array
[i
].extent
);
3779 extent
= isl_union_set_from_set(set
);
3780 domain
= isl_union_map_domain(copy_i
);
3781 copy_i
= isl_union_map_from_domain_and_range(domain
, extent
);
3782 res
= isl_union_map_union(res
, copy_i
);
3785 isl_union_map_free(copy
);
3790 /* Insert "kernel" marks that point to a ppcg_kernel structure
3791 * in front of all outermost tilable band that (by construction)
3792 * have at least one parallel loop.
3794 static __isl_give isl_schedule_node
*mark_kernels(struct gpu_gen
*gen
,
3795 __isl_take isl_schedule_node
*node
)
3797 return isl_schedule_node_map_descendant(node
,
3798 &mark_outer_permutable
, gen
);
3801 /* Save the schedule "schedule" to a file called "filename".
3802 * The schedule is printed in block style.
3804 static void save_schedule(__isl_keep isl_schedule
*schedule
,
3805 const char *filename
)
3814 file
= fopen(filename
, "w");
3816 fprintf(stderr
, "Unable to open '%s' for writing\n", filename
);
3819 ctx
= isl_schedule_get_ctx(schedule
);
3820 p
= isl_printer_to_file(ctx
, file
);
3821 p
= isl_printer_set_yaml_style(p
, ISL_YAML_STYLE_BLOCK
);
3822 p
= isl_printer_print_schedule(p
, schedule
);
3823 isl_printer_free(p
);
3827 /* Load and return a schedule from a file called "filename".
3829 static __isl_give isl_schedule
*load_schedule(isl_ctx
*ctx
,
3830 const char *filename
)
3833 isl_schedule
*schedule
;
3835 file
= fopen(filename
, "r");
3837 fprintf(stderr
, "Unable to open '%s' for reading\n", filename
);
3840 schedule
= isl_schedule_read_from_file(ctx
, file
);
3846 /* Compute an appropriate schedule based on the accesses in
3847 * gen->read and gen->write.
3849 * We use the dependences in gen->prog->scop to compute
3850 * a schedule that has a parallel loop in each tilable band and
3851 * return this schedule.
3853 * If live range reordering is allowed, then we need to make sure
3854 * that live ranges on arrays are not run in parallel since doing
3855 * so would require array expansion. We therefore add the array
3856 * order dependences to the coincidence dependences. Non-zero array
3857 * order dependences will then prevent a schedule dimension from being
3858 * considered parallel.
3859 * Live ranges derived from scalars are allowed to be run in parallel
3860 * since we force the scalars to be mapped to private memory in
3861 * check_scalar_live_ranges.
3862 * If live range reordering is allowed, then the false dependences
3863 * are not added to the validity constraints as that would prevent
3864 * reordering. Instead, the external false dependences that enforce that reads
3865 * from potentially live-in data precede any later write and
3866 * that writes of potentially live-out data follow any other earlier write
3867 * are added to the validity and the coincidence constraints.
3868 * The false dependences are still added to the proximity constraints
3869 * for consistency with the case where live range reordering is not allowed.
3870 * The coincidence constraints then consist of flow dependences,
3871 * external false dependences and array order dependences.
3872 * The independences can be filtered out from the first two sets.
3873 * They have already been filtered out from the array order dependences
3874 * on a per array basis in collect_order_dependences.
3875 * There is no need for a per array handling of the other two sets
3876 * as there should be no flow or external false dependence on local
3877 * variables that can be filtered out.
3879 static __isl_give isl_schedule
*compute_schedule(struct gpu_gen
*gen
)
3881 isl_union_set
*domain
;
3882 isl_union_map
*dep_raw
, *dep
;
3883 isl_union_map
*validity
, *proximity
, *coincidence
;
3884 isl_schedule_constraints
*sc
;
3885 isl_schedule
*schedule
;
3887 domain
= isl_union_set_copy(gen
->prog
->scop
->domain
);
3888 sc
= isl_schedule_constraints_on_domain(domain
);
3889 sc
= isl_schedule_constraints_set_context(sc
,
3890 isl_set_copy(gen
->prog
->scop
->context
));
3891 if (gen
->options
->live_range_reordering
) {
3892 sc
= isl_schedule_constraints_set_conditional_validity(sc
,
3893 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_flow
),
3894 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_order
));
3895 proximity
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
3896 validity
= isl_union_map_copy(proximity
);
3897 validity
= isl_union_map_union(validity
,
3898 isl_union_map_copy(gen
->prog
->scop
->dep_forced
));
3899 proximity
= isl_union_map_union(proximity
,
3900 isl_union_map_copy(gen
->prog
->scop
->dep_false
));
3901 coincidence
= isl_union_map_copy(validity
);
3902 coincidence
= isl_union_map_subtract(coincidence
,
3903 isl_union_map_copy(gen
->prog
->scop
->independence
));
3904 coincidence
= isl_union_map_union(coincidence
,
3905 isl_union_map_copy(gen
->prog
->array_order
));
3907 dep_raw
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
3908 dep
= isl_union_map_copy(gen
->prog
->scop
->dep_false
);
3909 dep
= isl_union_map_union(dep
, dep_raw
);
3910 dep
= isl_union_map_coalesce(dep
);
3911 proximity
= isl_union_map_copy(dep
);
3912 coincidence
= isl_union_map_copy(dep
);
3915 sc
= isl_schedule_constraints_set_validity(sc
, validity
);
3916 sc
= isl_schedule_constraints_set_coincidence(sc
, coincidence
);
3917 sc
= isl_schedule_constraints_set_proximity(sc
, proximity
);
3919 if (gen
->options
->debug
->dump_schedule_constraints
)
3920 isl_schedule_constraints_dump(sc
);
3921 schedule
= isl_schedule_constraints_compute_schedule(sc
);
3926 /* Obtain a schedule for the scop, either by reading it from
3927 * a file or by computing one.
3929 static __isl_give isl_schedule
*get_schedule(struct gpu_gen
*gen
)
3931 isl_schedule
*schedule
;
3933 if (gen
->options
->load_schedule_file
) {
3934 schedule
= load_schedule(gen
->ctx
,
3935 gen
->options
->load_schedule_file
);
3937 schedule
= compute_schedule(gen
);
3938 if (gen
->options
->save_schedule_file
)
3939 save_schedule(schedule
,
3940 gen
->options
->save_schedule_file
);
3942 if (gen
->options
->debug
->dump_schedule
)
3943 isl_schedule_dump(schedule
);
3948 /* Return (the universe spaces of) the arrays that are declared
3949 * inside the scop corresponding to "prog" and for which all
3950 * potential writes inside the scop form a subset of "domain".
3952 static __isl_give isl_union_set
*extract_local_accesses(struct gpu_prog
*prog
,
3953 __isl_keep isl_union_set
*domain
)
3956 isl_union_set
*local
;
3958 local
= isl_union_set_empty(isl_union_set_get_space(domain
));
3960 for (i
= 0; i
< prog
->n_array
; ++i
) {
3962 isl_union_map
*to_outer
;
3963 isl_union_map
*may_write
;
3964 isl_union_set
*write_domain
;
3965 isl_union_set
*fields
;
3968 if (!prog
->array
[i
].local
)
3971 set
= isl_set_universe(isl_space_copy(prog
->array
[i
].space
));
3972 to_outer
= isl_union_map_copy(prog
->to_outer
);
3973 to_outer
= isl_union_map_intersect_range(to_outer
,
3974 isl_union_set_from_set(isl_set_copy(set
)));
3975 fields
= isl_union_map_domain(to_outer
);
3976 may_write
= isl_union_map_copy(prog
->may_write
);
3977 may_write
= isl_union_map_intersect_range(may_write
, fields
);
3978 write_domain
= isl_union_map_domain(may_write
);
3979 subset
= isl_union_set_is_subset(write_domain
, domain
);
3980 isl_union_set_free(write_domain
);
3984 return isl_union_set_free(local
);
3985 } else if (subset
) {
3986 local
= isl_union_set_add_set(local
, set
);
3995 /* Internal data structure for node_may_persist.
3997 * "tagger" maps tagged iteration domains to the corresponding untagged
4000 * "may_persist_flow" is the set of all tagged dataflow dependences
4001 * with those dependences removed that either precede or follow
4002 * the kernel launch in a sequence.
4003 * "inner_band_flow" is the set of all tagged dataflow dependences
4004 * that are local to a given iteration of the outer band nodes
4005 * with respect to the current node.
4006 * "local_flow" is equal to "inner_band_flow", except that the domain
4007 * and the range have been intersected with intermediate filters
4008 * on children of sets or sequences.
4010 struct ppcg_may_persist_data
{
4011 isl_union_pw_multi_aff
*tagger
;
4013 isl_union_map
*local_flow
;
4014 isl_union_map
*inner_band_flow
;
4015 isl_union_map
*may_persist_flow
;
4018 /* Update the information in "data" based on the band ancestor "node".
4020 * In particular, we restrict the dependences in data->local_flow
4021 * to those dependence where the source and the sink occur in
4022 * the same iteration of the given band node.
4023 * We also update data->inner_band_flow to the new value of
4026 static int update_may_persist_at_band(__isl_keep isl_schedule_node
*node
,
4027 struct ppcg_may_persist_data
*data
)
4029 isl_multi_union_pw_aff
*partial
;
4030 isl_union_pw_multi_aff
*contraction
;
4031 isl_union_map
*flow
;
4033 if (isl_schedule_node_band_n_member(node
) == 0)
4036 partial
= isl_schedule_node_band_get_partial_schedule(node
);
4037 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4038 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4040 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4041 isl_union_pw_multi_aff_copy(data
->tagger
));
4043 flow
= data
->local_flow
;
4044 flow
= isl_union_map_eq_at_multi_union_pw_aff(flow
, partial
);
4045 data
->local_flow
= flow
;
4047 isl_union_map_free(data
->inner_band_flow
);
4048 data
->inner_band_flow
= isl_union_map_copy(data
->local_flow
);
4053 /* Given a set of local reaching domain elements "domain",
4054 * expand them to the corresponding leaf domain elements using "contraction"
4055 * and insert the array references tags using data->tagger.
4057 static __isl_give isl_union_set
*expand_and_tag(
4058 __isl_take isl_union_set
*domain
,
4059 __isl_take isl_union_pw_multi_aff
*contraction
,
4060 struct ppcg_may_persist_data
*data
)
4062 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4064 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4065 isl_union_pw_multi_aff_copy(data
->tagger
));
4069 /* Given a filter node that is the child of a set or sequence node,
4070 * restrict data->local_flow to refer only to those elements
4071 * in the filter of the node.
4072 * "contraction" maps the leaf domain elements of the schedule tree
4073 * to the corresponding domain elements at (the parent of) "node".
4075 static int filter_flow(__isl_keep isl_schedule_node
*node
,
4076 struct ppcg_may_persist_data
*data
,
4077 __isl_take isl_union_pw_multi_aff
*contraction
)
4079 isl_union_set
*filter
;
4080 isl_union_map
*flow
;
4082 flow
= data
->local_flow
;
4083 filter
= isl_schedule_node_filter_get_filter(node
);
4084 filter
= expand_and_tag(filter
, contraction
, data
);
4085 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(filter
));
4086 flow
= isl_union_map_intersect_range(flow
, filter
);
4087 data
->local_flow
= flow
;
4092 /* Given a filter node "node", collect the filters on all preceding siblings
4093 * (which are also filter nodes), add them to "filters" and return the result.
4095 static __isl_give isl_union_set
*add_previous_filters(
4096 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4098 isl_schedule_node
*sibling
;
4100 sibling
= isl_schedule_node_copy(node
);
4101 while (sibling
&& isl_schedule_node_has_previous_sibling(sibling
)) {
4102 isl_union_set
*filter
;
4104 sibling
= isl_schedule_node_previous_sibling(sibling
);
4105 filter
= isl_schedule_node_filter_get_filter(sibling
);
4106 filters
= isl_union_set_union(filters
, filter
);
4108 isl_schedule_node_free(sibling
);
4110 return isl_union_set_free(filters
);
4115 /* Given a filter node "node", collect the filters on all following siblings
4116 * (which are also filter nodes), add them to "filters" and return the result.
4118 static __isl_give isl_union_set
*add_next_filters(
4119 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4121 isl_schedule_node
*sibling
;
4123 sibling
= isl_schedule_node_copy(node
);
4124 while (sibling
&& isl_schedule_node_has_next_sibling(sibling
)) {
4125 isl_union_set
*filter
;
4127 sibling
= isl_schedule_node_next_sibling(sibling
);
4128 filter
= isl_schedule_node_filter_get_filter(sibling
);
4129 filters
= isl_union_set_union(filters
, filter
);
4131 isl_schedule_node_free(sibling
);
4133 return isl_union_set_free(filters
);
4138 /* Remove those flow dependences from data->may_persist_flow
4139 * that flow between elements of "domain" within the same iteration
4140 * of all outer band nodes.
4141 * "contraction" maps the leaf domain elements of the schedule tree
4142 * to the corresponding elements "domain".
4144 static void remove_external_flow(struct ppcg_may_persist_data
*data
,
4145 __isl_take isl_union_set
*domain
,
4146 __isl_keep isl_union_pw_multi_aff
*contraction
)
4148 isl_union_map
*flow
;
4150 contraction
= isl_union_pw_multi_aff_copy(contraction
);
4151 domain
= expand_and_tag(domain
, contraction
, data
);
4152 flow
= isl_union_map_copy(data
->local_flow
);
4153 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(domain
));
4154 flow
= isl_union_map_intersect_range(flow
, domain
);
4156 data
->may_persist_flow
= isl_union_map_subtract(data
->may_persist_flow
,
4160 /* Update the information in "data" based on the filter ancestor "node".
4161 * We only need to modify anything if the filter is the child
4162 * of a set or sequence node.
4164 * In the case of a sequence, we remove the dependences between
4165 * statement instances that are both executed either before or
4166 * after the subtree that will be mapped to a kernel, within
4167 * the same iteration of outer bands.
4169 * In both cases, we restrict data->local_flow to the current child.
4171 static int update_may_persist_at_filter(__isl_keep isl_schedule_node
*node
,
4172 struct ppcg_may_persist_data
*data
)
4174 enum isl_schedule_node_type type
;
4175 isl_schedule_node
*parent
;
4177 isl_union_pw_multi_aff
*contraction
;
4178 isl_union_set
*before
, *after
, *filter
;
4179 isl_union_map
*flow
;
4181 type
= isl_schedule_node_get_parent_type(node
);
4182 if (type
!= isl_schedule_node_sequence
&& type
!= isl_schedule_node_set
)
4185 parent
= isl_schedule_node_copy(node
);
4186 parent
= isl_schedule_node_parent(parent
);
4187 contraction
= isl_schedule_node_get_subtree_contraction(parent
);
4188 isl_schedule_node_free(parent
);
4190 if (type
== isl_schedule_node_set
)
4191 return filter_flow(node
, data
, contraction
);
4193 filter
= isl_schedule_node_filter_get_filter(node
);
4194 space
= isl_union_set_get_space(filter
);
4195 isl_union_set_free(filter
);
4196 before
= isl_union_set_empty(space
);
4197 after
= isl_union_set_copy(before
);
4198 before
= add_previous_filters(before
, node
);
4199 after
= add_next_filters(after
, node
);
4201 remove_external_flow(data
, before
, contraction
);
4202 remove_external_flow(data
, after
, contraction
);
4204 return filter_flow(node
, data
, contraction
);
4207 /* Update the information in "data" based on the ancestor "node".
4209 static int update_may_persist_at(__isl_keep isl_schedule_node
*node
, void *user
)
4211 struct ppcg_may_persist_data
*data
= user
;
4213 switch (isl_schedule_node_get_type(node
)) {
4214 case isl_schedule_node_error
:
4216 case isl_schedule_node_context
:
4217 case isl_schedule_node_domain
:
4218 case isl_schedule_node_expansion
:
4219 case isl_schedule_node_extension
:
4220 case isl_schedule_node_guard
:
4221 case isl_schedule_node_leaf
:
4222 case isl_schedule_node_mark
:
4223 case isl_schedule_node_sequence
:
4224 case isl_schedule_node_set
:
4226 case isl_schedule_node_band
:
4227 if (update_may_persist_at_band(node
, data
) < 0)
4230 case isl_schedule_node_filter
:
4231 if (update_may_persist_at_filter(node
, data
) < 0)
4239 /* Determine the set of array elements that may need to be perserved
4240 * by a kernel constructed from the subtree at "node".
4241 * This includes the set of array elements that may need to be preserved
4242 * by the entire scop (prog->may_persist) and the elements for which
4243 * there is a potential flow dependence that may cross a kernel launch.
4245 * To determine the second set, we start from all flow dependences.
4246 * From this set of dependences, we remove those that cannot possibly
4247 * require data to be preserved by a kernel launch.
4248 * In particular, we consider the following sets of dependences.
4249 * - dependences of which the write occurs inside the kernel.
4250 * If the data is needed outside the kernel, then it will
4251 * be copied out immediately after the kernel launch, so there
4252 * is no need for any special care.
4253 * - dependences of which the read occurs inside the kernel and the
4254 * corresponding write occurs inside the same iteration of the
4255 * outer band nodes. This means that the data is needed in
4256 * the first kernel launch after the write, which is already
4257 * taken care of by the standard copy-in. That is, the data
4258 * do not need to be preserved by any intermediate call to
4260 * - dependences of which the write and the read either both occur
4261 * before the kernel launch or both occur after the kernel launch,
4262 * within the same iteration of the outer band nodes with respect
4263 * to the sequence that determines the ordering of the dependence
4264 * and the kernel launch. Such flow dependences cannot cross
4265 * any kernel launch.
4267 * For the remaining (tagged) dependences, we take the domain
4268 * (i.e., the tagged writes) and apply the tagged access relation
4269 * to obtain the accessed data elements.
4270 * These are then combined with the elements that may need to be
4271 * preserved by the entire scop.
4273 static __isl_give isl_union_set
*node_may_persist(
4274 __isl_keep isl_schedule_node
*node
, struct gpu_prog
*prog
)
4276 struct ppcg_may_persist_data data
;
4277 isl_schedule_node
*root
;
4278 isl_union_pw_multi_aff
*contraction
;
4279 isl_union_set
*domain
;
4280 isl_union_set
*persist
;
4281 isl_union_map
*flow
, *local_flow
;
4283 data
.tagger
= prog
->scop
->tagger
;
4285 flow
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
4286 data
.local_flow
= isl_union_map_copy(flow
);
4287 data
.inner_band_flow
= isl_union_map_copy(flow
);
4288 data
.may_persist_flow
= flow
;
4289 if (isl_schedule_node_foreach_ancestor_top_down(node
,
4290 &update_may_persist_at
, &data
) < 0)
4291 data
.may_persist_flow
=
4292 isl_union_map_free(data
.may_persist_flow
);
4293 flow
= data
.may_persist_flow
;
4294 isl_union_map_free(data
.local_flow
);
4296 domain
= isl_schedule_node_get_domain(node
);
4297 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4298 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4300 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4301 isl_union_pw_multi_aff_copy(data
.tagger
));
4302 flow
= isl_union_map_subtract_domain(flow
, isl_union_set_copy(domain
));
4303 local_flow
= data
.inner_band_flow
;
4304 local_flow
= isl_union_map_intersect_range(local_flow
, domain
);
4305 flow
= isl_union_map_subtract(flow
, local_flow
);
4307 persist
= isl_union_map_domain(flow
);
4308 persist
= isl_union_set_apply(persist
,
4309 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4310 persist
= isl_union_set_union(persist
,
4311 isl_union_set_copy(prog
->may_persist
));
4316 /* Compute the sets of outer array elements that need to be copied in and out
4317 * before and after the subtree "node", which will be mapped
4318 * to one or more kernels.
4319 * "domain" contains the original reaching domain elements before
4320 * the kernels were created, i.e., before the contraction that
4321 * may have been performed in creating the kernels has been applied.
4322 * "prefix" contains the prefix schedule at that point, in terms
4323 * of the same original reaching domain elements.
4325 * In particular, for each array that is possibly written anywhere in
4326 * the subtree "node" and that may be used after "node"
4327 * or that may be visible outside the corresponding scop,
4328 * we copy out its entire extent.
4330 * Any array elements that is read without first being written inside
4331 * the subtree "node" needs to be copied in.
4332 * Furthermore, if there are any array elements that
4333 * are copied out, but that may not be written inside "node, then
4334 * they also need to be copied in to ensure that the value after execution
4335 * is the same as the value before execution, at least for those array
4336 * elements that may have their values preserved by the scop or that
4337 * may be written before "node" and read after "node".
4338 * In case the array elements are structures, we need to take into
4339 * account that all members of the structures need to be written
4340 * by "node" before we can avoid copying the data structure in.
4342 * Note that the may_write relation is intersected with the domain,
4343 * which has been intersected with the context.
4344 * This helps in those cases where the arrays are declared with a fixed size,
4345 * while the accesses are parametric and the context assigns a fixed value
4346 * to the parameters.
4348 * If an element from a local array is read without first being written,
4349 * then there is no point in copying it in since it cannot have been
4350 * written prior to the scop. Warn about the uninitialized read instead.
4352 static void compute_copy_in_and_out(__isl_keep isl_schedule_node
*node
,
4353 __isl_take isl_union_set
*domain
, __isl_take isl_union_map
*prefix
,
4354 struct gpu_prog
*prog
)
4356 isl_union_set
*local
;
4357 isl_union_set
*to_device
, *from_device
, *may_persist
;
4358 isl_union_map
*may_write
, *must_write
, *copy_out
, *not_written
;
4359 isl_union_map
*read
, *copy_in
;
4360 isl_union_map
*tagged
;
4361 isl_union_map
*local_uninitialized
;
4363 tagged
= isl_union_map_copy(prog
->scop
->tagged_reads
);
4364 tagged
= isl_union_map_union(tagged
,
4365 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4367 may_write
= isl_union_map_copy(prog
->may_write
);
4368 may_write
= isl_union_map_intersect_domain(may_write
,
4369 isl_union_set_copy(domain
));
4370 may_write
= remove_local_accesses(prog
,
4371 isl_union_map_copy(tagged
), may_write
,
4372 isl_union_map_copy(prefix
), 0);
4373 may_write
= isl_union_map_apply_range(may_write
,
4374 isl_union_map_copy(prog
->to_outer
));
4375 may_write
= isl_union_map_apply_domain(may_write
,
4376 isl_union_map_copy(prefix
));
4377 may_write
= approximate_copy_out(may_write
, prog
);
4378 copy_out
= isl_union_map_copy(may_write
);
4379 may_write
= isl_union_map_apply_range(may_write
,
4380 isl_union_map_copy(prog
->to_inner
));
4381 must_write
= isl_union_map_copy(prog
->must_write
);
4382 must_write
= isl_union_map_apply_domain(must_write
,
4383 isl_union_map_copy(prefix
));
4384 may_persist
= node_may_persist(node
, prog
);
4385 may_write
= isl_union_map_intersect_range(may_write
, may_persist
);
4386 not_written
= isl_union_map_subtract(may_write
, must_write
);
4388 local
= extract_local_accesses(prog
, domain
);
4389 read
= isl_union_map_copy(prog
->read
);
4390 read
= isl_union_map_intersect_domain(read
, domain
);
4391 read
= remove_local_accesses(prog
, tagged
, read
,
4392 isl_union_map_copy(prefix
), 1);
4393 local
= isl_union_set_apply(local
, isl_union_map_copy(prog
->to_inner
));
4394 local_uninitialized
= isl_union_map_copy(prog
->scop
->live_in
);
4395 local_uninitialized
= isl_union_map_intersect_range(local_uninitialized
,
4397 local_uninitialized
= isl_union_map_intersect(local_uninitialized
,
4398 isl_union_map_copy(read
));
4399 if (!isl_union_map_is_empty(local_uninitialized
)) {
4401 "possibly uninitialized reads (not copied in):\n");
4402 isl_union_map_dump(local_uninitialized
);
4404 read
= isl_union_map_subtract(read
, local_uninitialized
);
4405 read
= isl_union_map_apply_domain(read
, prefix
);
4406 copy_in
= isl_union_map_union(read
, not_written
);
4407 copy_in
= isl_union_map_apply_range(copy_in
,
4408 isl_union_map_copy(prog
->to_outer
));
4410 prog
->copy_out
= isl_union_map_range(copy_out
);
4411 prog
->copy_in
= isl_union_map_range(copy_in
);
4414 /* Update "schedule" for mapping to a GPU device.
4416 * In particular, insert a context node, create kernels for
4417 * each outermost tilable band.
4418 * Also compute the sets of outer array elements that need
4419 * to be copied in and out.
4421 static __isl_give isl_schedule
*map_to_device(struct gpu_gen
*gen
,
4422 __isl_take isl_schedule
*schedule
)
4424 isl_schedule_node
*node
;
4426 isl_union_set
*domain
;
4427 isl_union_map
*prefix
;
4429 context
= isl_set_copy(gen
->prog
->context
);
4430 context
= isl_set_from_params(context
);
4431 schedule
= isl_schedule_insert_context(schedule
, context
);
4433 node
= isl_schedule_get_root(schedule
);
4434 isl_schedule_free(schedule
);
4435 node
= isl_schedule_node_child(node
, 0);
4436 if (isl_schedule_node_get_type(node
) == isl_schedule_node_context
)
4437 node
= isl_schedule_node_child(node
, 0);
4438 domain
= isl_schedule_node_get_domain(node
);
4439 prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
4440 compute_copy_in_and_out(node
, domain
, prefix
, gen
->prog
);
4441 node
= mark_kernels(gen
, node
);
4442 schedule
= isl_schedule_node_get_schedule(node
);
4443 isl_schedule_node_free(node
);
4448 /* Internal data structure for extract_access.
4449 * "next_access" points to the end of a linked list that is extended
4450 * by extract_access.
4451 * "single_expression" is set if the access expressions belong to
4452 * an expression statement (i.e., a statement without internal control).
4453 * "any_to_outer" maps all intermediate arrays to their outer arrays.
4455 struct ppcg_extract_access_data
{
4456 struct gpu_stmt_access
**next_access
;
4457 int single_expression
;
4458 isl_union_map
*any_to_outer
;
4461 /* Given a tagged access relation to a single array "tagged", extract it
4462 * as a map, taking into account that the input may be empty.
4463 * If the access relation is empty, then it does not contain
4464 * any space information, so we try to recover it from the index
4466 * The space of the index expression is of the form I -> A,
4467 * with I the statement instances and A the array, or [I -> F] -> A,
4468 * with F the filters corresponding to arguments.
4469 * We first drop F, if present, obtaining I -> A.
4470 * Then we construct I -> R, with R the reference tag,
4471 * combine the two into I -> [R -> A] and uncurry to obtain
4472 * the final result [I -> R] -> A.
4473 * Note that the index expression may have a lower dimension
4474 * than that of the array, but this dimension is not used
4475 * if the access relation is empty.
4477 static __isl_give isl_map
*extract_single_tagged_access(
4478 __isl_take isl_union_map
*tagged
, __isl_keep pet_expr
*expr
)
4482 isl_space
*space
, *space2
;
4483 isl_multi_pw_aff
*index
;
4485 empty
= isl_union_map_is_empty(tagged
);
4489 return isl_map_from_union_map(tagged
);
4490 isl_union_map_free(tagged
);
4492 index
= pet_expr_access_get_index(expr
);
4493 space
= isl_multi_pw_aff_get_space(index
);
4494 isl_multi_pw_aff_free(index
);
4495 if (isl_space_domain_is_wrapping(space
))
4496 space
= isl_space_domain_factor_domain(space
);
4497 space2
= isl_space_copy(space
);
4498 space2
= isl_space_from_domain(isl_space_domain(space
));
4499 id
= pet_expr_access_get_ref_id(expr
);
4500 space2
= isl_space_set_tuple_id(space2
, isl_dim_out
, id
);
4501 space
= isl_space_range_product(space2
, space
);
4502 space
= isl_space_uncurry(space
);
4504 return isl_map_empty(space
);
4506 isl_union_map_free(tagged
);
4510 /* Extract a gpu_stmt_access from "expr", append it to the list
4511 * that ends in *data->next_access and update the end of the list.
4512 * If the access expression performs a write, then it is considered
4513 * exact only if it appears in a single expression statement and
4514 * if its may access relation is equal to its must access relation.
4516 * The combined set of may accesses may be union if member accesses
4517 * are involved, but the entire set is derived from a single reference and
4518 * therefore from a single index expression. These accesses therefore
4519 * all map to the same outer array.
4521 static int extract_access(__isl_keep pet_expr
*expr
, void *user
)
4523 struct ppcg_extract_access_data
*data
= user
;
4524 isl_union_map
*tagged
;
4525 struct gpu_stmt_access
*access
;
4526 isl_ctx
*ctx
= pet_expr_get_ctx(expr
);
4527 isl_multi_pw_aff
*index
;
4529 access
= isl_alloc_type(ctx
, struct gpu_stmt_access
);
4531 access
->next
= NULL
;
4532 access
->read
= pet_expr_access_is_read(expr
);
4533 access
->write
= pet_expr_access_is_write(expr
);
4534 tagged
= pet_expr_access_get_tagged_may_read(expr
);
4535 tagged
= isl_union_map_union(tagged
,
4536 pet_expr_access_get_tagged_may_write(expr
));
4537 tagged
= isl_union_map_apply_range(tagged
,
4538 isl_union_map_copy(data
->any_to_outer
));
4539 if (!access
->write
) {
4540 access
->exact_write
= 1;
4541 } else if (!data
->single_expression
) {
4542 access
->exact_write
= 0;
4544 isl_union_map
*must
, *may
;
4545 may
= isl_union_map_copy(tagged
);
4546 may
= isl_union_map_domain_factor_domain(may
);
4547 must
= pet_expr_access_get_must_write(expr
);
4548 access
->exact_write
= isl_union_map_is_equal(must
, may
);
4549 isl_union_map_free(must
);
4550 isl_union_map_free(may
);
4552 index
= pet_expr_access_get_index(expr
);
4553 access
->n_index
= isl_multi_pw_aff_dim(index
, isl_dim_out
);
4554 isl_multi_pw_aff_free(index
);
4555 access
->ref_id
= pet_expr_access_get_ref_id(expr
);
4556 access
->tagged_access
= extract_single_tagged_access(tagged
, expr
);
4557 access
->access
= isl_map_copy(access
->tagged_access
);
4558 access
->access
= isl_map_domain_factor_domain(access
->access
);
4560 *data
->next_access
= access
;
4561 data
->next_access
= &(*data
->next_access
)->next
;
4563 if (!access
->access
)
4569 /* Construct a linked list of gpu_stmt_access objects,
4570 * one for each access expression in the statement body.
4571 * "any_to_outer" maps all intermediate arrays to their outer arrays.
4573 static int pet_stmt_extract_accesses(struct gpu_stmt
*stmt
,
4574 __isl_keep isl_union_map
*any_to_outer
)
4576 struct ppcg_extract_access_data data
;
4578 stmt
->accesses
= NULL
;
4579 data
.next_access
= &stmt
->accesses
;
4580 data
.single_expression
=
4581 pet_tree_get_type(stmt
->stmt
->body
) == pet_tree_expr
;
4582 data
.any_to_outer
= any_to_outer
;
4583 return pet_tree_foreach_access_expr(stmt
->stmt
->body
,
4584 &extract_access
, &data
);
4587 /* Return an array of gpu_stmt representing the statements in "scop".
4589 static struct gpu_stmt
*extract_stmts(isl_ctx
*ctx
, struct ppcg_scop
*scop
,
4590 __isl_keep isl_set
*context
, __isl_keep isl_union_map
*any_to_outer
)
4593 struct gpu_stmt
*stmts
;
4595 stmts
= isl_calloc_array(ctx
, struct gpu_stmt
, scop
->pet
->n_stmt
);
4599 for (i
= 0; i
< scop
->pet
->n_stmt
; ++i
) {
4600 struct gpu_stmt
*s
= &stmts
[i
];
4602 s
->id
= isl_set_get_tuple_id(scop
->pet
->stmts
[i
]->domain
);
4603 s
->stmt
= scop
->pet
->stmts
[i
];
4604 if (pet_stmt_extract_accesses(s
, any_to_outer
) < 0)
4605 return free_stmts(stmts
, i
+ 1);
4611 /* Callback for ppcg_print_guarded that calls the callback for generate_gpu.
4613 static __isl_give isl_printer
*print_gpu(__isl_take isl_printer
*p
, void *user
)
4615 struct gpu_gen
*gen
= user
;
4617 return gen
->print(p
, gen
->prog
, gen
->tree
, &gen
->types
,
4621 /* Generate CUDA code for "scop" and print it to "p".
4622 * After generating an AST for the transformed scop as explained below,
4623 * we call "gen->print" to print the AST in the desired output format
4626 * If it turns out that it does not make sense to generate GPU code,
4627 * then we generate CPU code instead.
4629 * The GPU code is generated in a context where at least one
4630 * statement instance is executed. The corresponding guard (if any) is printed
4631 * around the entire generated GPU code, except for the declaration
4632 * of the arrays that are visible outside of the scop and that therefore
4633 * cannot be declared inside the body of any possible guard.
4635 * We first compute a schedule that respects the dependences
4636 * of the original program and select the outermost bands
4637 * of tilable dimensions that have at least one parallel loop.
4638 * If the --load-schedule is specified, then the loaded schedule
4639 * is used instead of a computed schedule.
4641 * Each of these bands B is then tiled according to "tile" sizes, resulting
4642 * in two nested bands, with a kernel marker on top
4650 * We then split off at most 2 parallel dimensions from the T band and
4651 * at most 3 parallel dimension from the P band
4664 * A filter is introduced in front of T1 that maps the domain instances
4665 * to block identifiers. Similarly, a filter is introduced in front of P1
4666 * that maps the domain instances to thread identifiers.
4668 * For each iteration of the T2 band and for each array, we compute
4669 * the array elements accessed by that iteration, construct a rectangular
4670 * box around it and shift it to the origin. The result is used
4671 * as shared memory for the array.
4673 * Copying and synchronization statements are added to this schedule tree.
4674 * In principle, these are added in front of the P1 band, but some of
4675 * them may get hoisted up to higher levels.
4677 * The entire AST is then generated from the single resulting schedule tree.
4678 * During the generation the subtrees at kernel nodes (K) are saved
4679 * aside and replaced by kernel calls. The result is printed as host code
4680 * while the saved subtrees are printed as device code.
4682 static __isl_give isl_printer
*generate(__isl_take isl_printer
*p
,
4683 struct gpu_gen
*gen
, struct ppcg_scop
*scop
,
4684 struct ppcg_options
*options
)
4686 struct gpu_prog
*prog
;
4688 isl_set
*context
, *guard
;
4689 isl_schedule
*schedule
;
4693 return isl_printer_free(p
);
4695 ctx
= isl_printer_get_ctx(p
);
4696 prog
= gpu_prog_alloc(ctx
, scop
);
4698 return isl_printer_free(p
);
4700 context
= isl_set_copy(prog
->context
);
4701 guard
= isl_union_set_params(isl_union_set_copy(prog
->scop
->domain
));
4702 prog
->context
= isl_set_intersect(prog
->context
, isl_set_copy(guard
));
4705 schedule
= get_schedule(gen
);
4707 any_permutable
= has_any_permutable_node(schedule
);
4708 if (any_permutable
< 0 || !any_permutable
) {
4709 isl_set_free(context
);
4710 isl_set_free(guard
);
4711 if (any_permutable
< 0)
4712 p
= isl_printer_free(p
);
4714 p
= print_cpu(p
, scop
, options
);
4715 isl_schedule_free(schedule
);
4717 schedule
= map_to_device(gen
, schedule
);
4718 gen
->tree
= generate_code(gen
, schedule
);
4719 p
= ppcg_print_exposed_declarations(p
, prog
->scop
);
4720 p
= ppcg_print_guarded(p
, guard
, context
, &print_gpu
, gen
);
4721 isl_ast_node_free(gen
->tree
);
4724 gpu_prog_free(prog
);
4729 /* Wrapper around generate for use as a ppcg_transform callback.
4731 static __isl_give isl_printer
*generate_wrap(__isl_take isl_printer
*p
,
4732 struct ppcg_scop
*scop
, void *user
)
4734 struct gpu_gen
*gen
= user
;
4736 return generate(p
, gen
, scop
, gen
->options
);
4739 /* Transform the code in the file called "input" by replacing
4740 * all scops by corresponding GPU code and write the results to "out".
4742 int generate_gpu(isl_ctx
*ctx
, const char *input
, FILE *out
,
4743 struct ppcg_options
*options
,
4744 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
4745 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
4746 struct gpu_types
*types
, void *user
), void *user
)
4753 gen
.sizes
= extract_sizes_from_str(ctx
, options
->sizes
);
4754 gen
.options
= options
;
4757 gen
.print_user
= user
;
4759 gen
.types
.name
= NULL
;
4761 if (options
->debug
->dump_sizes
) {
4762 isl_space
*space
= isl_space_params_alloc(ctx
, 0);
4763 gen
.used_sizes
= isl_union_map_empty(space
);
4766 r
= ppcg_transform(ctx
, input
, out
, options
, &generate_wrap
, &gen
);
4768 if (options
->debug
->dump_sizes
) {
4769 isl_union_map_dump(gen
.used_sizes
);
4770 isl_union_map_free(gen
.used_sizes
);
4773 isl_union_map_free(gen
.sizes
);
4774 for (i
= 0; i
< gen
.types
.n
; ++i
)
4775 free(gen
.types
.name
[i
]);
4776 free(gen
.types
.name
);
4781 /* Compute the set of inner array elements that may have their values
4782 * preserved by "prog". In particular, collect the array elements of
4783 * arrays that are not local to "prog" and remove those elements that
4784 * are definitely killed or definitely written by "prog".
4786 static __isl_give isl_union_set
*compute_may_persist(struct gpu_prog
*prog
)
4789 isl_union_set
*may_persist
, *killed
;
4790 isl_union_map
*must_kill
;
4792 may_persist
= isl_union_set_empty(isl_set_get_space(prog
->context
));
4793 for (i
= 0; i
< prog
->n_array
; ++i
) {
4796 if (prog
->array
[i
].local
)
4799 extent
= isl_set_copy(prog
->array
[i
].extent
);
4800 may_persist
= isl_union_set_add_set(may_persist
, extent
);
4803 may_persist
= isl_union_set_intersect_params(may_persist
,
4804 isl_set_copy(prog
->context
));
4805 may_persist
= isl_union_set_apply(may_persist
,
4806 isl_union_map_copy(prog
->to_inner
));
4807 must_kill
= isl_union_map_copy(prog
->tagged_must_kill
);
4808 killed
= isl_union_map_range(must_kill
);
4809 must_kill
= isl_union_map_copy(prog
->must_write
);
4810 killed
= isl_union_set_union(killed
, isl_union_map_range(must_kill
));
4812 may_persist
= isl_union_set_subtract(may_persist
, killed
);
4816 struct gpu_prog
*gpu_prog_alloc(isl_ctx
*ctx
, struct ppcg_scop
*scop
)
4818 struct gpu_prog
*prog
;
4825 prog
= isl_calloc_type(ctx
, struct gpu_prog
);
4830 prog
->context
= isl_set_copy(scop
->context
);
4831 prog
->n_stmts
= scop
->pet
->n_stmt
;
4832 prog
->any_to_outer
= pet_scop_compute_outer_to_any(scop
->pet
);
4833 prog
->any_to_outer
= isl_union_map_reverse(prog
->any_to_outer
);
4834 space
= isl_union_map_get_space(prog
->any_to_outer
);
4835 space
= isl_space_set_from_params(space
);
4836 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
4837 space
= isl_space_map_from_set(space
);
4838 id
= isl_map_identity(space
);
4839 prog
->any_to_outer
= isl_union_map_add_map(prog
->any_to_outer
, id
);
4840 prog
->stmts
= extract_stmts(ctx
, scop
,
4841 prog
->context
, prog
->any_to_outer
);
4842 prog
->read
= isl_union_map_copy(scop
->reads
);
4843 prog
->may_write
= isl_union_map_copy(scop
->may_writes
);
4844 prog
->must_write
= isl_union_map_copy(scop
->must_writes
);
4845 prog
->tagged_must_kill
= isl_union_map_copy(scop
->tagged_must_kills
);
4846 prog
->to_inner
= pet_scop_compute_outer_to_inner(scop
->pet
);
4847 prog
->to_outer
= isl_union_map_copy(prog
->to_inner
);
4848 prog
->to_outer
= isl_union_map_reverse(prog
->to_outer
);
4851 return gpu_prog_free(prog
);
4853 if (collect_array_info(prog
) < 0)
4854 return gpu_prog_free(prog
);
4855 prog
->may_persist
= compute_may_persist(prog
);
4860 void *gpu_prog_free(struct gpu_prog
*prog
)
4864 free_array_info(prog
);
4865 free_stmts(prog
->stmts
, prog
->n_stmts
);
4866 isl_union_map_free(prog
->any_to_outer
);
4867 isl_union_map_free(prog
->to_outer
);
4868 isl_union_map_free(prog
->to_inner
);
4869 isl_union_set_free(prog
->copy_in
);
4870 isl_union_set_free(prog
->copy_out
);
4871 isl_union_map_free(prog
->read
);
4872 isl_union_map_free(prog
->may_write
);
4873 isl_union_map_free(prog
->must_write
);
4874 isl_union_map_free(prog
->tagged_must_kill
);
4875 isl_union_map_free(prog
->array_order
);
4876 isl_union_set_free(prog
->may_persist
);
4877 isl_set_free(prog
->context
);