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"
37 struct gpu_array_info
;
39 /* Collect all references to the given array and store pointers to them
42 * If the array contains structures, then there is no need to collect
43 * the references since we will not be computing any reference groups.
45 static void collect_references(struct gpu_prog
*prog
,
46 struct gpu_array_info
*array
)
51 if (array
->has_compound_element
)
55 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
56 struct gpu_stmt
*stmt
= &prog
->stmts
[i
];
57 struct gpu_stmt_access
*access
;
59 for (access
= stmt
->accesses
; access
; access
= access
->next
) {
61 name
= isl_map_get_tuple_name(access
->access
,
63 if (name
&& !strcmp(array
->name
, name
))
69 array
->refs
= isl_alloc_array(prog
->ctx
, struct gpu_stmt_access
*, n
);
73 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
74 struct gpu_stmt
*stmt
= &prog
->stmts
[i
];
75 struct gpu_stmt_access
*access
;
77 for (access
= stmt
->accesses
; access
; access
= access
->next
) {
79 name
= isl_map_get_tuple_name(access
->access
,
81 if (!name
|| strcmp(array
->name
, name
))
84 array
->refs
[n
++] = access
;
89 /* Compute and return the extent of "array", taking into account the set of
92 * In particular, the extent in the outer dimension is taken
93 * from "accessed", while the extents in the remaining dimensions
94 * are taken from array->extent.
96 * The extent in the outer dimension cannot be taken from array->extent
97 * because that may be unbounded. Furthermore, even if it is bounded,
98 * it may be larger than the piece of the array that is being accessed.
100 static __isl_give isl_set
*compute_extent(struct pet_array
*array
,
101 __isl_keep isl_set
*accessed
)
108 extent
= isl_set_copy(array
->extent
);
110 n_index
= isl_set_dim(accessed
, isl_dim_set
);
114 extent
= isl_set_project_out(extent
, isl_dim_set
, 0, 1);
115 outer
= isl_set_copy(accessed
);
116 outer
= isl_set_project_out(outer
, isl_dim_set
, 1, n_index
- 1);
117 extent
= isl_set_flat_product(outer
, extent
);
118 id
= isl_set_get_tuple_id(accessed
);
119 extent
= isl_set_set_tuple_id(extent
, id
);
124 /* Is the array "array" being extracted a read-only scalar?
126 * That is, is "array" a scalar that is never possibly written to.
127 * An array containing structures is never considered to be a scalar.
129 static int is_read_only_scalar(struct gpu_array_info
*array
,
130 struct gpu_prog
*prog
)
133 isl_union_map
*write
;
136 if (array
->has_compound_element
)
138 if (array
->n_index
!= 0)
141 write
= isl_union_map_copy(prog
->may_write
);
142 space
= isl_set_universe(isl_space_copy(array
->space
));
143 write
= isl_union_map_intersect_range(write
,
144 isl_union_set_from_set(space
));
145 empty
= isl_union_map_is_empty(write
);
146 isl_union_map_free(write
);
151 /* Compute bounds on the host array "pa" based on the corresponding
152 * accessed elements in "arrays"
153 * and collect all references to the array.
154 * Store the results in "info".
156 * If the array is zero-dimensional and does not contain structures,
157 * i.e., if the array is a scalar, we check whether it is read-only.
158 * We also check whether the array is accessed at all.
160 static int extract_array_info(struct gpu_prog
*prog
,
161 struct gpu_array_info
*info
, struct pet_array
*pa
,
162 __isl_keep isl_union_set
*arrays
)
168 isl_set
*accessed
, *extent
;
170 n_index
= isl_set_dim(pa
->extent
, isl_dim_set
);
171 name
= isl_set_get_tuple_name(pa
->extent
);
172 bounds
= isl_alloc_array(prog
->ctx
, isl_pw_aff
*, n_index
);
176 info
->space
= isl_set_get_space(pa
->extent
);
177 info
->name
= strdup(name
);
178 info
->n_index
= n_index
;
179 info
->bound
= bounds
;
180 info
->linearize
= prog
->scop
->options
->linearize_device_arrays
;
182 info
->type
= strdup(pa
->element_type
);
183 info
->size
= pa
->element_size
;
184 info
->local
= pa
->declared
&& !pa
->exposed
;
185 info
->has_compound_element
= pa
->element_is_record
;
186 info
->read_only_scalar
= is_read_only_scalar(info
, prog
);
188 accessed
= isl_union_set_extract_set(arrays
,
189 isl_space_copy(info
->space
));
190 empty
= isl_set_is_empty(accessed
);
191 extent
= compute_extent(pa
, accessed
);
192 isl_set_free(accessed
);
193 info
->extent
= extent
;
196 info
->accessed
= !empty
;
197 for (i
= 0; i
< n_index
; ++i
) {
203 dom
= isl_set_copy(extent
);
204 dom
= isl_set_project_out(dom
, isl_dim_set
, i
+ 1,
206 dom
= isl_set_project_out(dom
, isl_dim_set
, 0, i
);
207 if (!isl_set_dim_has_upper_bound(dom
, isl_dim_set
, 0)) {
208 fprintf(stderr
, "unable to determine extent of '%s' "
209 "in dimension %d\n", info
->name
, i
);
210 dom
= isl_set_free(dom
);
212 bound
= isl_set_dim_max(dom
, 0);
213 dom
= isl_pw_aff_domain(isl_pw_aff_copy(bound
));
214 ls
= isl_local_space_from_space(isl_set_get_space(dom
));
215 one
= isl_aff_zero_on_domain(ls
);
216 one
= isl_aff_add_constant_si(one
, 1);
217 bound
= isl_pw_aff_add(bound
, isl_pw_aff_alloc(dom
, one
));
218 bound
= isl_pw_aff_gist(bound
, isl_set_copy(prog
->context
));
221 if (!isl_pw_aff_is_cst(bound
))
225 collect_references(prog
, info
);
230 /* Remove independence from the order constraints "order" on array "array".
231 * Since the pairs of iterations in the filter relation of an independence
232 * are guaranteed to be completely independent by the user, there is
233 * no need to ensure that live ranges are ordered along thong pairs.
234 * We make an exception for local variables, though, as the independence
235 * guarantee does not apply to those.
237 * The order constraints are used in two places.
238 * Those on scalars are used in check_scalar_live_ranges to check if
239 * we need to force the scalar to be private. Any non-local scalar
240 * should not be forced scalar if it only appears in independent loops.
241 * Those on non-scalars are added to the coincidence constraints
242 * in compute_schedule because we do not support any array expansion.
243 * Accesses to non-local arrays should not prevent a loop from being
244 * considered coincident so we should indeed remove those constraints
245 * from the order constraints.
247 static __isl_give isl_union_map
*remove_independences(struct gpu_prog
*prog
,
248 struct gpu_array_info
*array
, __isl_take isl_union_map
*order
)
252 for (i
= 0; i
< prog
->scop
->pet
->n_independence
; ++i
) {
253 struct pet_independence
*pi
= prog
->scop
->pet
->independences
[i
];
254 if (isl_union_set_contains(pi
->local
, array
->space
))
257 order
= isl_union_map_subtract(order
,
258 isl_union_map_copy(pi
->filter
));
264 /* For each array in "prog", store the (untagged) order dependences
265 * derived from the array in array->dep_order.
266 * In particular, consider all references that access the given array
267 * and take the order dependences that have one of these references
268 * as source. (Since an order dependence relates two references to
269 * the same array, the target of these order dependences will also
270 * be one of these references.)
271 * Additionally, store the union of these array->dep_order relations
272 * for all non-scalar arrays in prog->array_order.
274 void collect_order_dependences(struct gpu_prog
*prog
)
278 isl_union_map
*accesses
;
280 space
= isl_union_map_get_space(prog
->read
);
281 prog
->array_order
= isl_union_map_empty(space
);
283 accesses
= isl_union_map_copy(prog
->scop
->tagged_reads
);
284 accesses
= isl_union_map_union(accesses
,
285 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
286 accesses
= isl_union_map_universe(accesses
);
287 accesses
= isl_union_map_apply_range(accesses
,
288 isl_union_map_copy(prog
->to_outer
));
290 for (i
= 0; i
< prog
->n_array
; ++i
) {
291 struct gpu_array_info
*array
= &prog
->array
[i
];
294 isl_union_map
*order
;
296 set
= isl_set_universe(isl_space_copy(array
->space
));
297 uset
= isl_union_set_from_set(set
);
298 uset
= isl_union_map_domain(
299 isl_union_map_intersect_range(isl_union_map_copy(accesses
),
301 order
= isl_union_map_copy(prog
->scop
->tagged_dep_order
);
302 order
= isl_union_map_intersect_domain(order
, uset
);
303 order
= isl_union_map_zip(order
);
304 order
= isl_union_set_unwrap(isl_union_map_domain(order
));
305 order
= remove_independences(prog
, array
, order
);
306 array
->dep_order
= order
;
308 if (gpu_array_is_scalar(array
) && !array
->has_compound_element
)
311 prog
->array_order
= isl_union_map_union(prog
->array_order
,
312 isl_union_map_copy(array
->dep_order
));
315 isl_union_map_free(accesses
);
318 /* Construct a gpu_array_info for each array referenced by prog->scop and
319 * collect them in prog->array.
321 * The sizes are based on the extents and the set of possibly accessed
322 * elements by "prog".
323 * If there are any member accesses involved, then they are first mapped
324 * to the outer arrays of structs.
326 * If we are allowing live range reordering, then also set
327 * the dep_order field. Otherwise leave it NULL.
329 static int collect_array_info(struct gpu_prog
*prog
)
333 isl_union_set
*arrays
;
335 arrays
= isl_union_map_range(isl_union_map_copy(prog
->read
));
336 arrays
= isl_union_set_union(arrays
,
337 isl_union_map_range(isl_union_map_copy(prog
->may_write
)));
339 arrays
= isl_union_set_apply(arrays
,
340 isl_union_map_copy(prog
->to_outer
));
342 arrays
= isl_union_set_coalesce(arrays
);
344 prog
->n_array
= prog
->scop
->pet
->n_array
;
345 prog
->array
= isl_calloc_array(prog
->ctx
,
346 struct gpu_array_info
, prog
->n_array
);
348 for (i
= 0; i
< prog
->scop
->pet
->n_array
; ++i
)
349 if (extract_array_info(prog
, &prog
->array
[i
],
350 prog
->scop
->pet
->arrays
[i
], arrays
) < 0)
353 isl_union_set_free(arrays
);
355 if (prog
->scop
->options
->live_range_reordering
)
356 collect_order_dependences(prog
);
361 static void free_array_info(struct gpu_prog
*prog
)
365 for (i
= 0; i
< prog
->n_array
; ++i
) {
366 int n_index
= prog
->array
[i
].n_index
;
367 free(prog
->array
[i
].type
);
368 free(prog
->array
[i
].name
);
369 for (j
= 0; j
< n_index
; ++j
)
370 isl_pw_aff_free(prog
->array
[i
].bound
[j
]);
371 isl_space_free(prog
->array
[i
].space
);
372 isl_set_free(prog
->array
[i
].extent
);
373 free(prog
->array
[i
].bound
);
374 free(prog
->array
[i
].refs
);
375 isl_union_map_free(prog
->array
[i
].dep_order
);
380 /* Check if a gpu array is a scalar. A scalar is a value that is not stored
381 * as an array or through a pointer reference, but as a single data element.
382 * At the moment, scalars are represented as zero-dimensional arrays.
383 * Note that the single data element may be an entire structure.
385 int gpu_array_is_scalar(struct gpu_array_info
*array
)
387 return array
->n_index
== 0;
390 /* Is "array" a read-only scalar?
392 int gpu_array_is_read_only_scalar(struct gpu_array_info
*array
)
394 return array
->read_only_scalar
;
397 /* Return the set of parameter values for which the array has a positive
398 * size in all dimensions.
399 * If the sizes are only valid for some parameter values, then those
400 * constraints are also taken into account.
402 __isl_give isl_set
*gpu_array_positive_size_guard(struct gpu_array_info
*array
)
411 space
= isl_space_params(isl_space_copy(array
->space
));
412 guard
= isl_set_universe(space
);
414 for (i
= 0; i
< array
->n_index
; ++i
) {
416 isl_set
*guard_i
, *zero
;
418 bound
= isl_pw_aff_copy(array
->bound
[i
]);
419 guard_i
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(bound
));
420 zero
= isl_pw_aff_zero_set(bound
);
421 guard_i
= isl_set_subtract(guard_i
, zero
);
422 guard
= isl_set_intersect(guard
, guard_i
);
428 /* Internal data structure for extract_size_of_type.
429 * "type" specifies the name of the space that we want to extract.
430 * "res" is used to store the subset of that space.
432 struct ppcg_extract_size_data
{
437 /* This function is called for each set in a union_set.
438 * If the name of the set matches data->type, we store the
441 static int extract_size_of_type(__isl_take isl_set
*size
, void *user
)
443 struct ppcg_extract_size_data
*data
= user
;
446 name
= isl_set_get_tuple_name(size
);
447 if (name
&& !strcmp(name
, data
->type
)) {
456 /* Given a union map { kernel[i] -> *[...] },
457 * return the range in the space called "type" for the kernel with
458 * sequence number "id".
460 static __isl_give isl_set
*extract_sizes(__isl_keep isl_union_map
*sizes
,
461 const char *type
, int id
)
465 isl_union_set
*local_sizes
;
466 struct ppcg_extract_size_data data
= { type
, NULL
};
471 space
= isl_union_map_get_space(sizes
);
472 space
= isl_space_set_from_params(space
);
473 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
474 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
475 dom
= isl_set_universe(space
);
476 dom
= isl_set_fix_si(dom
, isl_dim_set
, 0, id
);
478 local_sizes
= isl_union_set_apply(isl_union_set_from_set(dom
),
479 isl_union_map_copy(sizes
));
480 isl_union_set_foreach_set(local_sizes
, &extract_size_of_type
, &data
);
481 isl_union_set_free(local_sizes
);
485 /* Given a singleton set, extract the first (at most *len) elements
486 * of the single integer tuple into *sizes and update *len if needed.
488 static void read_sizes_from_set(__isl_take isl_set
*set
, int *sizes
, int *len
)
496 dim
= isl_set_dim(set
, isl_dim_set
);
500 for (i
= 0; i
< *len
; ++i
) {
503 v
= isl_set_plain_get_val_if_fixed(set
, isl_dim_set
, i
);
506 sizes
[i
] = isl_val_get_num_si(v
);
513 /* Add the map { kernel[id] -> type[sizes] } to gen->used_sizes,
514 * if the option debug->dump_sizes is set.
516 static void set_used_sizes(struct gpu_gen
*gen
, const char *type
, int id
,
523 if (!gen
->options
->debug
->dump_sizes
)
526 space
= isl_union_map_get_space(gen
->used_sizes
);
527 space
= isl_space_set_from_params(space
);
528 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
529 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
530 space
= isl_space_from_domain(space
);
531 space
= isl_space_add_dims(space
, isl_dim_out
, len
);
532 space
= isl_space_set_tuple_name(space
, isl_dim_out
, type
);
534 map
= isl_map_universe(space
);
535 map
= isl_map_fix_si(map
, isl_dim_in
, 0, id
);
536 for (i
= 0; i
< len
; ++i
)
537 map
= isl_map_fix_si(map
, isl_dim_out
, i
, sizes
[i
]);
539 gen
->used_sizes
= isl_union_map_add_map(gen
->used_sizes
, map
);
542 /* Extract user specified "tile" sizes from the "sizes" command line option,
543 * defaulting to option->tile_size in each dimension.
544 * *tile_len contains the maximum number of tile sizes needed.
545 * Update *tile_len to the number of specified tile sizes, if any, and
546 * return a pointer to the tile sizes (or NULL on error).
547 * Add the effectively used sizes to gen->used_sizes.
549 static int *read_tile_sizes(struct gpu_gen
*gen
, int *tile_len
)
555 tile_size
= isl_alloc_array(gen
->ctx
, int, *tile_len
);
558 for (n
= 0; n
< *tile_len
; ++n
)
559 tile_size
[n
] = gen
->options
->tile_size
;
561 size
= extract_sizes(gen
->sizes
, "tile", gen
->kernel_id
);
562 read_sizes_from_set(size
, tile_size
, tile_len
);
563 set_used_sizes(gen
, "tile", gen
->kernel_id
, tile_size
, *tile_len
);
568 /* Extract user specified "block" sizes from the "sizes" command line option,
569 * after filling in some potentially useful defaults.
571 static void read_block_sizes(struct ppcg_kernel
*kernel
,
572 __isl_keep isl_union_map
*sizes
)
576 if (kernel
->n_block
> 3)
578 switch (kernel
->n_block
) {
580 kernel
->block_dim
[0] = 512;
583 kernel
->block_dim
[0] = 32;
584 kernel
->block_dim
[1] = 16;
587 kernel
->block_dim
[0] = 32;
588 kernel
->block_dim
[1] = 4;
589 kernel
->block_dim
[2] = 4;
593 size
= extract_sizes(sizes
, "block", kernel
->id
);
594 read_sizes_from_set(size
, kernel
->block_dim
, &kernel
->n_block
);
597 /* Extract user specified "grid" sizes from the "sizes" command line option,
598 * after filling in some potentially useful defaults.
600 static void read_grid_sizes(struct ppcg_kernel
*kernel
,
601 __isl_keep isl_union_map
*sizes
)
605 if (kernel
->n_grid
> 2)
607 switch (kernel
->n_grid
) {
609 kernel
->grid_dim
[0] = 32768;
612 kernel
->grid_dim
[0] = 256;
613 kernel
->grid_dim
[1] = 256;
617 size
= extract_sizes(sizes
, "grid", kernel
->id
);
618 read_sizes_from_set(size
, kernel
->grid_dim
, &kernel
->n_grid
);
621 /* Extract user specified grid and block sizes from the gen->sizes
622 * command line option after filling in some potentially useful defaults.
623 * Store the extracted sizes in "kernel".
624 * Add the effectively used sizes to gen->used_sizes.
626 static void read_grid_and_block_sizes(struct ppcg_kernel
*kernel
,
629 read_block_sizes(kernel
, gen
->sizes
);
630 read_grid_sizes(kernel
, gen
->sizes
);
631 set_used_sizes(gen
, "block", kernel
->id
,
632 kernel
->block_dim
, kernel
->n_block
);
633 set_used_sizes(gen
, "grid", kernel
->id
,
634 kernel
->grid_dim
, kernel
->n_grid
);
637 static void *free_stmts(struct gpu_stmt
*stmts
, int n
)
644 for (i
= 0; i
< n
; ++i
) {
645 struct gpu_stmt_access
*access
, *next
;
647 for (access
= stmts
[i
].accesses
; access
; access
= next
) {
649 isl_id_free(access
->ref_id
);
650 isl_map_free(access
->access
);
651 isl_map_free(access
->tagged_access
);
655 isl_id_free(stmts
[i
].id
);
662 /* Add parameters p[i] with identifiers "ids" to "set",
663 * with bounds to 0 <= p[i] < size[i].
665 __isl_give isl_set
*add_bounded_parameters(__isl_take isl_set
*set
,
666 int *size
, __isl_keep isl_id_list
*ids
)
671 len
= isl_id_list_n_id(ids
);
672 nparam
= isl_set_dim(set
, isl_dim_param
);
673 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
675 for (i
= 0; i
< len
; ++i
) {
678 id
= isl_id_list_get_id(ids
, i
);
679 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
680 set
= isl_set_lower_bound_si(set
, isl_dim_param
, nparam
+ i
, 0);
681 set
= isl_set_upper_bound_si(set
, isl_dim_param
,
682 nparam
+ i
, size
[i
] - 1);
688 /* Add "len" parameters p[i] with identifiers "ids" and intersect "set"
691 * { : 0 <= p[i] < size[i] }
693 * or an overapproximation.
695 static __isl_give isl_set
*add_bounded_parameters_dynamic(
696 __isl_take isl_set
*set
, __isl_keep isl_multi_pw_aff
*size
,
697 __isl_keep isl_id_list
*ids
)
704 len
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
705 nparam
= isl_set_dim(set
, isl_dim_param
);
706 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
708 for (i
= 0; i
< len
; ++i
) {
711 id
= isl_id_list_get_id(ids
, i
);
712 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
715 space
= isl_space_params(isl_set_get_space(set
));
716 ls
= isl_local_space_from_space(space
);
717 for (i
= 0; i
< len
; ++i
) {
718 isl_pw_aff
*param
, *size_i
, *zero
;
721 param
= isl_pw_aff_var_on_domain(isl_local_space_copy(ls
),
722 isl_dim_param
, nparam
+ i
);
724 size_i
= isl_multi_pw_aff_get_pw_aff(size
, i
);
725 bound
= isl_pw_aff_lt_set(isl_pw_aff_copy(param
), size_i
);
726 bound
= isl_set_from_basic_set(isl_set_simple_hull(bound
));
727 set
= isl_set_intersect_params(set
, bound
);
729 zero
= isl_pw_aff_zero_on_domain(isl_local_space_copy(ls
));
730 bound
= isl_pw_aff_ge_set(param
, zero
);
731 set
= isl_set_intersect_params(set
, bound
);
733 isl_local_space_free(ls
);
738 /* Return the union of all tagged access relations in the group.
740 static __isl_give isl_union_map
*group_tagged_access_relation(
741 struct gpu_array_ref_group
*group
)
744 isl_union_map
*access
;
746 access
= isl_union_map_empty(isl_map_get_space(group
->access
));
747 for (i
= 0; i
< group
->n_ref
; ++i
) {
750 map_i
= isl_map_copy(group
->refs
[i
]->tagged_access
);
751 access
= isl_union_map_union(access
,
752 isl_union_map_from_map(map_i
));
758 /* Return the extent of "array", recomputed from the bounds.
759 * The recomputed extent may be simpler than the original extent.
761 static __isl_give isl_set
*array_extent(struct gpu_array_info
*array
)
769 id
= isl_set_get_tuple_id(array
->extent
);
770 space
= isl_set_get_space(array
->extent
);
771 extent
= isl_set_universe(isl_space_copy(space
));
772 ls
= isl_local_space_from_space(space
);
773 for (i
= 0; i
< array
->n_index
; ++i
) {
779 extent
= isl_set_lower_bound_si(extent
, isl_dim_set
, i
, 0);
781 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
783 index
= isl_pw_aff_from_aff(aff
);
784 bound
= isl_pw_aff_copy(array
->bound
[i
]);
785 bound
= isl_pw_aff_from_range(bound
);
786 bound
= isl_pw_aff_add_dims(bound
, isl_dim_in
, array
->n_index
);
787 bound
= isl_pw_aff_set_tuple_id(bound
, isl_dim_in
,
789 lt
= isl_pw_aff_lt_set(index
, bound
);
790 extent
= isl_set_intersect(extent
, lt
);
792 isl_local_space_free(ls
);
798 /* Return a map from the first group->depth dimensions of the computed
799 * schedule to the array tile in
800 * global memory that corresponds to the shared memory copy.
802 * In particular, return a map
808 * tile_offset(i) <= a <= tile_offset(i) + tile_size - 1 (1)
812 * 0 <= a <= array_size - 1 (2)
814 * Note that if some stride has been detected (i.e., when
815 * group->shared_tile->bound[i].shift is set), then a in (1) refers
816 * to the shifted and scaled down version.
818 * Constraints (1) are obtained by mapping the size constraints on the
819 * shared/private memory tile back to the access relation.
820 * Constraints (2) are obtained from the (recomputed) extent.
822 static __isl_give isl_map
*group_tile(struct gpu_array_ref_group
*group
)
825 int n_index
= group
->array
->n_index
;
831 space
= isl_multi_aff_get_space(group
->shared_tile
->tiling
);
832 space
= isl_space_range(space
);
833 local
= isl_set_universe(space
);
834 for (i
= 0; i
< n_index
; ++i
) {
837 local
= isl_set_lower_bound_si(local
, isl_dim_set
, i
, 0);
838 bound
= isl_val_copy(group
->shared_tile
->bound
[i
].size
);
839 bound
= isl_val_sub_ui(bound
, 1);
840 local
= isl_set_upper_bound_val(local
, isl_dim_set
, i
, bound
);
842 local
= isl_set_preimage_multi_aff(local
,
843 isl_multi_aff_copy(group
->shared_tile
->tiling
));
844 tile
= isl_set_unwrap(local
);
845 extent
= array_extent(group
->array
);
846 tile
= isl_map_intersect_range(tile
, extent
);
851 /* Given a mapping "iterator_map" from the AST schedule to a domain,
852 * return the corresponding mapping from the AST schedule to
853 * to the outer kernel->shared_schedule_dim dimensions of
854 * the schedule computed by PPCG for this kernel.
856 * Note that kernel->shared_schedule_dim is at least as large as
857 * the largest depth of any array reference group associated to the kernel.
858 * This is needed as the returned schedule is used to extract a mapping
859 * to the outer group->depth dimensions in transform_index.
861 static __isl_give isl_pw_multi_aff
*compute_sched_to_shared(
862 struct ppcg_kernel
*kernel
, __isl_take isl_pw_multi_aff
*iterator_map
)
864 isl_union_pw_multi_aff
*upma
;
865 isl_pw_multi_aff
*pma
;
868 space
= isl_space_range(isl_pw_multi_aff_get_space(iterator_map
));
869 space
= isl_space_from_domain(space
);
870 space
= isl_space_add_dims(space
, isl_dim_out
,
871 kernel
->shared_schedule_dim
);
873 upma
= isl_union_pw_multi_aff_copy(kernel
->shared_schedule
);
874 pma
= isl_union_pw_multi_aff_extract_pw_multi_aff(upma
, space
);
875 isl_union_pw_multi_aff_free(upma
);
877 return isl_pw_multi_aff_pullback_pw_multi_aff(pma
, iterator_map
);
880 /* If max_shared_memory is not set to infinity (-1), then make
881 * sure that the total amount of shared memory required by the
882 * array reference groups mapped to shared memory by "kernel"
883 * is no larger than this maximum.
885 * We apply a greedy approach and discard (keep in global memory)
886 * those groups that would result in a total memory size that
887 * is larger than the maximum.
889 * This function should be called after any function that may
890 * affect the decision on whether to place a reference group
891 * in private, shared or global memory.
893 static void check_shared_memory_bound(struct ppcg_kernel
*kernel
)
896 isl_val
*left
, *size
;
898 if (kernel
->options
->max_shared_memory
< 0)
901 left
= isl_val_int_from_si(kernel
->ctx
,
902 kernel
->options
->max_shared_memory
);
904 for (i
= 0; i
< kernel
->n_array
; ++i
) {
905 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
907 for (j
= 0; j
< local
->n_group
; ++j
) {
908 struct gpu_array_ref_group
*group
;
910 group
= local
->groups
[j
];
911 if (group
->private_tile
)
913 if (!group
->shared_tile
)
916 size
= gpu_array_tile_size(group
->shared_tile
);
917 size
= isl_val_mul_ui(size
, local
->array
->size
);
919 if (isl_val_le(size
, left
)) {
920 left
= isl_val_sub(left
, size
);
926 gpu_array_tile_free(group
->shared_tile
);
933 /* Compute a tiling for all the array reference groups in "kernel".
935 static void compute_group_tilings(struct ppcg_kernel
*kernel
)
939 for (i
= 0; i
< kernel
->n_array
; ++i
) {
940 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
942 for (j
= 0; j
< array
->n_group
; ++j
)
943 gpu_array_ref_group_compute_tiling(array
->groups
[j
]);
947 /* Compute the size of a bounding box around the origin and "set",
948 * where "set" is assumed to contain only non-negative elements.
949 * In particular, compute the maximal value of "set" in each direction
952 static __isl_give isl_multi_pw_aff
*extract_size(__isl_take isl_set
*set
,
953 __isl_take isl_set
*context
)
956 isl_multi_pw_aff
*mpa
;
958 context
= isl_set_params(context
);
959 n
= isl_set_dim(set
, isl_dim_set
);
960 mpa
= isl_multi_pw_aff_zero(isl_set_get_space(set
));
961 for (i
= 0; i
< n
; ++i
) {
966 bound
= isl_set_dim_max(isl_set_copy(set
), i
);
967 bound
= isl_pw_aff_coalesce(bound
);
968 bound
= isl_pw_aff_gist(bound
, isl_set_copy(context
));
970 space
= isl_pw_aff_get_domain_space(bound
);
971 one
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
972 one
= isl_aff_add_constant_si(one
, 1);
973 bound
= isl_pw_aff_add(bound
, isl_pw_aff_from_aff(one
));
974 mpa
= isl_multi_pw_aff_set_pw_aff(mpa
, i
, bound
);
977 isl_set_free(context
);
982 /* Compute the effective grid size as a list of the sizes in each dimension.
984 * The grid size specified by the user or set by default
985 * in read_grid_sizes() and applied by the block filter,
986 * may be too large for the given code in the sense that
987 * it may contain blocks that don't need to execute anything.
988 * We therefore don't return this grid size, but instead the
989 * smallest grid size that ensures that all blocks that actually
990 * execute code are included in the grid.
992 * We first extract a description of the grid, i.e., the possible values
993 * of the block ids, from the domain elements in "domain" and
994 * kernel->block_filter.
995 * The block ids are parameters in kernel->block_filter.
996 * We simply need to change them into set dimensions.
998 * Then, for each block dimension, we compute the maximal value of the block id
1001 static __isl_give isl_multi_pw_aff
*extract_grid_size(
1002 struct ppcg_kernel
*kernel
, __isl_take isl_union_set
*domain
)
1007 domain
= isl_union_set_intersect(domain
,
1008 isl_union_set_copy(kernel
->block_filter
));
1009 grid
= isl_union_set_params(domain
);
1010 grid
= isl_set_from_params(grid
);
1011 grid
= isl_set_add_dims(grid
, isl_dim_set
, kernel
->n_grid
);
1012 for (i
= 0; i
< kernel
->n_grid
; ++i
) {
1016 id
= isl_id_list_get_id(kernel
->block_ids
, i
);
1017 pos
= isl_set_find_dim_by_id(grid
, isl_dim_param
, id
);
1020 grid
= isl_set_equate(grid
, isl_dim_param
, pos
, isl_dim_set
, i
);
1021 grid
= isl_set_project_out(grid
, isl_dim_param
, pos
, 1);
1024 return extract_size(grid
, isl_set_copy(kernel
->context
));
1027 /* Compute the size of a fixed bounding box around the origin and "set",
1028 * where "set" is assumed to contain only non-negative elements,
1029 * and store the results in "size".
1030 * In particular, compute the maximal value of "set" in each direction
1033 static void extract_fixed_size(__isl_take isl_set
*set
, int *size
)
1036 isl_local_space
*ls
;
1039 n
= isl_set_dim(set
, isl_dim_set
);
1040 ls
= isl_local_space_from_space(isl_set_get_space(set
));
1041 obj
= isl_aff_zero_on_domain(ls
);
1042 for (i
= 0; i
< n
; ++i
) {
1045 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 1);
1046 max
= isl_set_max_val(set
, obj
);
1047 size
[i
] = isl_val_get_num_si(max
) + 1;
1049 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 0);
1055 /* Compute the effective block size as a list of the sizes in each dimension
1056 * and store the sizes in kernel->block_dim.
1058 * The block size specified by the user or set by default
1059 * in read_block_sizes() and applied by the thread filter,
1060 * may be too large for the given code in the sense that
1061 * it may contain threads that don't need to execute anything.
1062 * We therefore update this block size in kernel->block_dim
1063 * to the smallest block size that ensures that all threads
1064 * that actually execute code are included in the block.
1066 * The possible values of the thread ids is obtained from
1067 * the domain elements "domain" and kernel->thread_filter.
1068 * The current implementation eliminates all parameters, ensuring
1069 * that the size is a fixed constant in each dimension.
1070 * In principle we could also compute parametric sizes.
1071 * We would have to make sure to project out all b%d and t%d parameters,
1074 static void extract_block_size(struct ppcg_kernel
*kernel
,
1075 __isl_take isl_union_set
*domain
)
1081 domain
= isl_union_set_intersect(domain
,
1082 isl_union_set_copy(kernel
->thread_filter
));
1083 block
= isl_union_set_params(domain
);
1084 block
= isl_set_from_params(block
);
1085 block
= isl_set_add_dims(block
, isl_dim_set
, kernel
->n_block
);
1086 for (i
= 0; i
< kernel
->n_block
; ++i
) {
1090 id
= isl_id_list_get_id(kernel
->thread_ids
, i
);
1091 pos
= isl_set_find_dim_by_id(block
, isl_dim_param
, id
);
1094 block
= isl_set_equate(block
, isl_dim_param
, pos
,
1097 nparam
= isl_set_dim(block
, isl_dim_param
);
1098 block
= isl_set_project_out(block
, isl_dim_param
, 0, nparam
);
1100 extract_fixed_size(block
, kernel
->block_dim
);
1103 struct ppcg_kernel
*ppcg_kernel_free(struct ppcg_kernel
*kernel
)
1110 isl_id_list_free(kernel
->block_ids
);
1111 isl_id_list_free(kernel
->thread_ids
);
1112 isl_multi_pw_aff_free(kernel
->grid_size
);
1113 isl_set_free(kernel
->context
);
1114 isl_union_set_free(kernel
->core
);
1115 isl_union_set_free(kernel
->arrays
);
1116 isl_space_free(kernel
->space
);
1117 isl_ast_node_free(kernel
->tree
);
1118 isl_union_set_free(kernel
->block_filter
);
1119 isl_union_set_free(kernel
->thread_filter
);
1120 isl_union_pw_multi_aff_free(kernel
->shared_schedule
);
1121 isl_union_set_free(kernel
->sync_writes
);
1123 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1124 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1126 for (j
= 0; j
< array
->n_group
; ++j
)
1127 gpu_array_ref_group_free(array
->groups
[j
]);
1128 free(array
->groups
);
1130 isl_pw_aff_list_free(array
->bound
);
1132 free(kernel
->array
);
1134 for (i
= 0; i
< kernel
->n_var
; ++i
) {
1135 free(kernel
->var
[i
].name
);
1136 isl_vec_free(kernel
->var
[i
].size
);
1145 /* Wrapper around ppcg_kernel_free for use as a isl_id_set_free_user callback.
1147 static void ppcg_kernel_free_wrap(void *user
)
1149 struct ppcg_kernel
*kernel
= user
;
1151 ppcg_kernel_free(kernel
);
1154 static void create_kernel_var(isl_ctx
*ctx
, struct gpu_array_ref_group
*group
,
1155 struct ppcg_kernel_var
*var
)
1158 struct gpu_array_tile
*tile
;
1162 var
->array
= group
->array
;
1164 tile
= group
->private_tile
;
1165 var
->type
= ppcg_access_private
;
1167 tile
= group
->shared_tile
;
1168 var
->type
= ppcg_access_shared
;
1171 p
= isl_printer_to_str(ctx
);
1172 p
= gpu_array_ref_group_print_name(group
, p
);
1173 var
->name
= isl_printer_get_str(p
);
1174 isl_printer_free(p
);
1176 var
->size
= isl_vec_alloc(ctx
, group
->array
->n_index
);
1178 for (j
= 0; j
< group
->array
->n_index
; ++j
)
1179 var
->size
= isl_vec_set_element_val(var
->size
, j
,
1180 isl_val_copy(tile
->bound
[j
].size
));
1183 static int create_kernel_vars(struct ppcg_kernel
*kernel
)
1188 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1189 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1191 for (j
= 0; j
< array
->n_group
; ++j
) {
1192 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1193 if (group
->private_tile
|| group
->shared_tile
)
1199 kernel
->var
= isl_calloc_array(kernel
->ctx
, struct ppcg_kernel_var
, n
);
1204 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1205 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1207 for (j
= 0; j
< array
->n_group
; ++j
) {
1208 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1209 if (!group
->private_tile
&& !group
->shared_tile
)
1211 create_kernel_var(kernel
->ctx
, group
, &kernel
->var
[n
]);
1219 /* Replace "pa" by the zero function defined over the universe domain
1220 * in the space of "pa".
1222 static __isl_give isl_pw_aff
*set_universally_zero(__isl_take isl_pw_aff
*pa
)
1227 space
= isl_space_domain(isl_pw_aff_get_space(pa
));
1228 isl_pw_aff_free(pa
);
1229 zero
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1231 return isl_pw_aff_from_aff(zero
);
1234 /* The sizes of the arrays on the host that have been computed by
1235 * extract_array_info may depend on the parameters. Use the extra
1236 * constraints on the parameters that are valid at "host_domain"
1237 * to simplify these expressions and store the results in kernel->array.
1239 * We only need these localized bounds for arrays that are accessed
1240 * by the current kernel. If we have found at least one reference group
1241 * then the array is accessed by the kernel. If the array has compound
1242 * elements then we skipped the construction of array reference groups.
1244 * The resulting sizes may be functions that are nowhere defined
1245 * in case the access function cannot possibly access anything inside
1246 * the kernel for some reason. If so, they are replaced by the zero
1247 * function. Since the access function cannot actually access anything,
1248 * there is no harm in printing the array sizes as zero.
1250 static void localize_bounds(struct ppcg_kernel
*kernel
,
1251 __isl_keep isl_set
*host_domain
)
1256 context
= isl_set_copy(host_domain
);
1257 context
= isl_set_params(context
);
1259 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1260 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
1261 isl_pw_aff_list
*bound
;
1264 if (local
->n_group
== 0 && !local
->array
->has_compound_element
)
1267 n_index
= local
->array
->n_index
;
1268 bound
= isl_pw_aff_list_alloc(kernel
->ctx
, n_index
);
1270 for (j
= 0; j
< n_index
; ++j
) {
1274 pwaff
= isl_pw_aff_copy(local
->array
->bound
[j
]);
1275 pwaff
= isl_pw_aff_gist(pwaff
, isl_set_copy(context
));
1276 empty
= isl_pw_aff_is_empty(pwaff
);
1278 pwaff
= isl_pw_aff_free(pwaff
);
1280 pwaff
= set_universally_zero(pwaff
);
1281 bound
= isl_pw_aff_list_add(bound
, pwaff
);
1284 local
->n_index
= n_index
;
1285 local
->bound
= bound
;
1287 isl_set_free(context
);
1290 /* Create the array of gpu_local_array_info structures "array"
1291 * inside "kernel". The number of elements in this array is
1292 * the same as the number of arrays in "prog".
1293 * Initialize the "array" field of each local array to point
1294 * to the corresponding array in "prog".
1296 static struct ppcg_kernel
*ppcg_kernel_create_local_arrays(
1297 struct ppcg_kernel
*kernel
, struct gpu_prog
*prog
)
1302 ctx
= isl_set_get_ctx(prog
->context
);
1303 kernel
->array
= isl_calloc_array(ctx
,
1304 struct gpu_local_array_info
, prog
->n_array
);
1306 return ppcg_kernel_free(kernel
);
1307 kernel
->n_array
= prog
->n_array
;
1309 for (i
= 0; i
< prog
->n_array
; ++i
)
1310 kernel
->array
[i
].array
= &prog
->array
[i
];
1315 /* Find the element in gen->stmt that has the given "id".
1316 * Return NULL if no such gpu_stmt can be found.
1318 static struct gpu_stmt
*find_stmt(struct gpu_prog
*prog
, __isl_keep isl_id
*id
)
1322 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
1323 if (id
== prog
->stmts
[i
].id
)
1327 return i
< prog
->n_stmts
? &prog
->stmts
[i
] : NULL
;
1330 void ppcg_kernel_stmt_free(void *user
)
1333 struct ppcg_kernel_stmt
*stmt
= user
;
1338 switch (stmt
->type
) {
1339 case ppcg_kernel_copy
:
1340 isl_ast_expr_free(stmt
->u
.c
.index
);
1341 isl_ast_expr_free(stmt
->u
.c
.local_index
);
1343 case ppcg_kernel_domain
:
1344 isl_id_to_ast_expr_free(stmt
->u
.d
.ref2expr
);
1346 case ppcg_kernel_sync
:
1353 /* Return the gpu_stmt_access in the list "accesses" that corresponds
1356 static struct gpu_stmt_access
*find_access(struct gpu_stmt_access
*accesses
,
1357 __isl_keep isl_id
*ref_id
)
1359 struct gpu_stmt_access
*access
;
1361 for (access
= accesses
; access
; access
= access
->next
)
1362 if (access
->ref_id
== ref_id
)
1368 /* Return the index of the array called "name" in the list of arrays.
1370 static int find_array_index(struct ppcg_kernel
*kernel
, const char *name
)
1374 for (i
= 0; i
< kernel
->n_array
; ++i
)
1375 if (!strcmp(name
, kernel
->array
[i
].array
->name
))
1381 /* Internal data structure for the index and AST expression transformation
1382 * callbacks for pet_stmt_build_ast_exprs.
1384 * "kernel" is the kernel for which are computing AST expressions.
1385 * "accesses" is the list of gpu_stmt_access in the statement.
1386 * "iterator_map" expresses the statement iterators in terms of
1387 * the AST loop iterators.
1388 * "sched2shared" expresses the outer shared_schedule_dim dimensions of
1389 * the kernel schedule in terms of the AST loop iterators.
1391 * The following fields are set in transform_index and used in transform_expr.
1392 * "array" is the array that is being accessed.
1393 * "global" is set if the global array is accessed (rather than
1394 * shared/private memory).
1395 * "local_array" refers to information on the array specialized
1396 * to the current kernel.
1398 struct ppcg_transform_data
{
1399 struct ppcg_kernel
*kernel
;
1400 struct gpu_stmt_access
*accesses
;
1401 isl_pw_multi_aff
*iterator_map
;
1402 isl_pw_multi_aff
*sched2shared
;
1404 struct gpu_array_info
*array
;
1406 struct gpu_local_array_info
*local_array
;
1409 /* Return the name of the outer array (of structs) accessed by "access".
1411 static const char *get_outer_array_name(__isl_keep isl_map
*access
)
1416 space
= isl_space_range(isl_map_get_space(access
));
1417 while (space
&& isl_space_is_wrapping(space
))
1418 space
= isl_space_domain(isl_space_unwrap(space
));
1419 name
= isl_space_get_tuple_name(space
, isl_dim_set
);
1420 isl_space_free(space
);
1425 /* Return a pointer to the gpu_array_ref_group in "local"
1426 * that contains the reference "access".
1427 * Return NULL if no such group can be found.
1429 static struct gpu_array_ref_group
*find_ref_group(
1430 struct gpu_local_array_info
*local
, struct gpu_stmt_access
*access
)
1434 for (i
= 0; i
< local
->n_group
; ++i
) {
1435 struct gpu_array_ref_group
*group
= local
->groups
[i
];
1437 for (j
= 0; j
< group
->n_ref
; ++j
)
1438 if (group
->refs
[j
] == access
)
1445 /* Index transformation callback for pet_stmt_build_ast_exprs.
1447 * "index" expresses the array indices in terms of statement iterators
1449 * We first reformulate "index" in terms of the AST loop iterators.
1450 * Then we check if we are accessing the global array or
1451 * a shared/private copy. In the former case, we simply return
1452 * the updated index. If "index" is an affine expression rather
1453 * than an array access, then we also return the updated index here.
1455 * If no reference groups have been computed for the array,
1456 * then we can only be accessing the global array.
1458 * Otherwise, we apply the tiling to the index.
1459 * This tiling is of the form
1463 * where D corresponds to the outer group->depth dimensions of
1464 * the kernel schedule.
1465 * The index is of the form
1469 * We update the tiling to refer to the AST loop iterators
1473 * and modify index to keep track of those iterators
1477 * Combining these two yields a tiled index expression in terms
1478 * of the AST loop iterators
1482 static __isl_give isl_multi_pw_aff
*transform_index(
1483 __isl_take isl_multi_pw_aff
*index
, __isl_keep isl_id
*ref_id
,
1486 struct ppcg_transform_data
*data
= user
;
1487 struct gpu_stmt_access
*access
;
1488 struct gpu_array_ref_group
*group
;
1489 struct gpu_array_tile
*tile
;
1490 isl_pw_multi_aff
*iterator_map
;
1495 isl_multi_pw_aff
*tiling
;
1496 isl_pw_multi_aff
*pma
;
1497 isl_multi_pw_aff
*mpa
;
1498 isl_pw_multi_aff
*sched2depth
;
1502 iterator_map
= isl_pw_multi_aff_copy(data
->iterator_map
);
1503 index
= isl_multi_pw_aff_pullback_pw_multi_aff(index
, iterator_map
);
1505 access
= find_access(data
->accesses
, ref_id
);
1508 if (!isl_map_has_tuple_name(access
->access
, isl_dim_out
))
1511 name
= get_outer_array_name(access
->access
);
1512 i
= find_array_index(data
->kernel
, name
);
1514 isl_die(isl_multi_pw_aff_get_ctx(index
), isl_error_internal
,
1515 "cannot find array",
1516 return isl_multi_pw_aff_free(index
));
1517 data
->local_array
= &data
->kernel
->array
[i
];
1518 data
->array
= data
->local_array
->array
;
1520 group
= find_ref_group(data
->local_array
, access
);
1526 tile
= group
->private_tile
;
1528 tile
= group
->shared_tile
;
1529 data
->global
= !tile
;
1533 space
= isl_space_range(isl_multi_pw_aff_get_space(index
));
1534 space
= isl_space_map_from_set(space
);
1535 pma
= isl_pw_multi_aff_identity(space
);
1536 sched2depth
= isl_pw_multi_aff_copy(data
->sched2shared
);
1537 dim
= isl_pw_multi_aff_dim(sched2depth
, isl_dim_out
);
1538 sched2depth
= isl_pw_multi_aff_drop_dims(sched2depth
, isl_dim_out
,
1539 group
->depth
, dim
- group
->depth
);
1540 pma
= isl_pw_multi_aff_product(sched2depth
, pma
);
1541 tiling
= isl_multi_pw_aff_from_multi_aff(
1542 isl_multi_aff_copy(tile
->tiling
));
1543 tiling
= isl_multi_pw_aff_pullback_pw_multi_aff(tiling
, pma
);
1545 space
= isl_space_domain(isl_multi_pw_aff_get_space(index
));
1546 space
= isl_space_map_from_set(space
);
1547 mpa
= isl_multi_pw_aff_identity(space
);
1548 index
= isl_multi_pw_aff_range_product(mpa
, index
);
1549 index
= isl_multi_pw_aff_pullback_multi_pw_aff(tiling
, index
);
1554 /* Dereference "expr" by adding an index [0].
1555 * The original "expr" is assumed not to have any indices.
1557 * If "expr" is a member access, then the dereferencing needs
1558 * to be applied to the structure argument of this member access.
1560 static __isl_give isl_ast_expr
*dereference(__isl_take isl_ast_expr
*expr
)
1563 isl_ast_expr
*arg0
, *res
;
1564 isl_ast_expr_list
*list
;
1566 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1568 return isl_ast_expr_free(expr
);
1569 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1570 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1573 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1574 arg
= dereference(arg
);
1575 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1576 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1580 isl_ast_expr_free(arg0
);
1582 ctx
= isl_ast_expr_get_ctx(expr
);
1583 res
= isl_ast_expr_from_val(isl_val_zero(ctx
));
1584 list
= isl_ast_expr_list_from_ast_expr(res
);
1585 res
= isl_ast_expr_get_op_arg(expr
, 0);
1586 res
= isl_ast_expr_access(res
, list
);
1587 isl_ast_expr_free(expr
);
1592 /* Linearize the index expression "expr" based on the array bounds
1595 * That is, transform expression
1597 * A[i_0][i_1]...[i_n]
1601 * A[(..((i_0 * b_1 + i_1) ... ) * b_n + i_n]
1603 * where b_0, b_1, ..., b_n are the bounds on the array.
1605 * If the base of "expr" is a member access, then the linearization needs
1606 * to be applied to the structure argument of this member access.
1608 * In the base case, if "expr" has no arguments (other than the name of
1609 * the array), then we are passing an entire array to a function.
1610 * In this case, there is nothing to linearize.
1611 * Note that at this point an expression with no arguments can
1612 * only be an entire array because the scalar case and
1613 * the case of single struct are handled by the caller.
1615 * If the number of specified index expressions in "expr"
1616 * is smaller than the dimension of the accessed array,
1617 * then the missing i_j also do not appear in the linearized expression.
1618 * Furthermore, since such an expression does not refer to a single
1619 * element while the default linearized expression would refer to
1620 * a single element, we return the expression
1622 * A + (..((i_0 * b_1 + i_1) ... ) * b_n]
1624 * instead. Note that because of the special case handling above,
1625 * we can assume here that here that there is at least one index expression.
1627 __isl_give isl_ast_expr
*gpu_local_array_info_linearize_index(
1628 struct gpu_local_array_info
*array
, __isl_take isl_ast_expr
*expr
)
1635 isl_ast_expr_list
*list
;
1636 isl_ast_build
*build
;
1638 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1639 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1640 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1643 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1644 arg
= gpu_local_array_info_linearize_index(array
, arg
);
1645 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1646 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1650 isl_ast_expr_free(arg0
);
1652 if (isl_ast_expr_get_op_n_arg(expr
) == 1)
1655 ctx
= isl_ast_expr_get_ctx(expr
);
1656 context
= isl_set_universe(isl_space_params_alloc(ctx
, 0));
1657 build
= isl_ast_build_from_context(context
);
1659 n
= isl_ast_expr_get_op_n_arg(expr
);
1660 res
= isl_ast_expr_get_op_arg(expr
, 1);
1661 for (i
= 1; i
< array
->n_index
; ++i
) {
1662 isl_pw_aff
*bound_i
;
1663 isl_ast_expr
*expr_i
;
1665 bound_i
= isl_pw_aff_list_get_pw_aff(array
->bound
, i
);
1666 expr_i
= isl_ast_build_expr_from_pw_aff(build
, bound_i
);
1667 res
= isl_ast_expr_mul(res
, expr_i
);
1671 expr_i
= isl_ast_expr_get_op_arg(expr
, i
+ 1);
1672 res
= isl_ast_expr_add(res
, expr_i
);
1675 isl_ast_build_free(build
);
1677 if (1 + array
->n_index
> n
) {
1678 res
= isl_ast_expr_add(isl_ast_expr_get_op_arg(expr
, 0), res
);
1680 list
= isl_ast_expr_list_from_ast_expr(res
);
1681 res
= isl_ast_expr_get_op_arg(expr
, 0);
1682 res
= isl_ast_expr_access(res
, list
);
1685 isl_ast_expr_free(expr
);
1690 /* AST expression transformation callback for pet_stmt_build_ast_exprs.
1692 * If the AST expression refers to an array that is not accessed
1693 * at all, then this means the value of the expression is not used,
1694 * so we might as well print zero (NULL pointer) instead.
1696 * If the AST expression refers to a global scalar that is not
1697 * a read-only scalar, then its address was passed to the kernel and
1698 * we need to dereference it.
1700 * If the AST expression refers to an access to a global array,
1701 * then we linearize the access exploiting the bounds in data->local_array.
1703 static __isl_give isl_ast_expr
*transform_expr(__isl_take isl_ast_expr
*expr
,
1704 __isl_keep isl_id
*id
, void *user
)
1706 struct ppcg_transform_data
*data
= user
;
1710 if (!data
->array
->accessed
) {
1713 ctx
= isl_ast_expr_get_ctx(expr
);
1714 isl_ast_expr_free(expr
);
1715 return isl_ast_expr_from_val(isl_val_zero(ctx
));
1717 if (gpu_array_is_read_only_scalar(data
->array
))
1721 if (data
->array
->n_index
== 0)
1722 return dereference(expr
);
1723 if (!data
->array
->linearize
)
1726 return gpu_local_array_info_linearize_index(data
->local_array
, expr
);
1729 /* This function is called for each instance of a user statement
1730 * in the kernel "kernel", identified by "gpu_stmt".
1732 * We attach a struct ppcg_kernel_stmt to the "node", containing
1733 * a computed AST expression for each access.
1734 * These AST expressions are computed from iterator_map,
1735 * which expresses the domain
1736 * elements in terms of the generated loops, and sched2shared,
1737 * which expresses the outer shared_schedule_dim dimensions of
1738 * the kernel schedule computed by PPCG in terms of the generated loops.
1740 static __isl_give isl_ast_node
*create_domain_leaf(
1741 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1742 __isl_keep isl_ast_build
*build
, struct gpu_stmt
*gpu_stmt
)
1744 struct ppcg_transform_data data
;
1745 struct ppcg_kernel_stmt
*stmt
;
1747 isl_pw_multi_aff
*sched2shared
;
1749 isl_pw_multi_aff
*iterator_map
;
1750 isl_union_map
*schedule
;
1752 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1754 return isl_ast_node_free(node
);
1756 schedule
= isl_ast_build_get_schedule(build
);
1757 map
= isl_map_reverse(isl_map_from_union_map(schedule
));
1758 iterator_map
= isl_pw_multi_aff_from_map(map
);
1759 sched2shared
= compute_sched_to_shared(kernel
,
1760 isl_pw_multi_aff_copy(iterator_map
));
1762 stmt
->type
= ppcg_kernel_domain
;
1763 stmt
->u
.d
.stmt
= gpu_stmt
;
1765 data
.kernel
= kernel
;
1766 data
.accesses
= stmt
->u
.d
.stmt
->accesses
;
1767 data
.iterator_map
= iterator_map
;
1768 data
.sched2shared
= sched2shared
;
1769 stmt
->u
.d
.ref2expr
= pet_stmt_build_ast_exprs(stmt
->u
.d
.stmt
->stmt
,
1770 build
, &transform_index
, &data
,
1771 &transform_expr
, &data
);
1773 isl_pw_multi_aff_free(iterator_map
);
1774 isl_pw_multi_aff_free(sched2shared
);
1776 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1777 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1778 return isl_ast_node_set_annotation(node
, id
);
1781 /* This function is called for each statement node in the AST
1782 * for copying to or from shared/private memory.
1783 * Attach a pointer to a ppcg_kernel_stmt representing the copy
1784 * statement to the node.
1785 * The statement name is "read" or "write", depending on whether we are
1786 * reading from global memory or writing to global memory.
1788 * The schedule is of the form
1792 * where D corresponds to the outer group->depth dimensions of
1793 * the kernel schedule, A to the global array and L to the outer
1794 * generated AST schedule.
1795 * We compute the inverse and strip off the type, resulting in
1799 * We combine this mapping with on the one hand the projection
1803 * and on the other hand the group tiling
1811 * and store the corresponding expressions in stmt->index and stmt->local_index,
1812 * where stmt points to the ppcg_kernel_stmt that is attached to the node.
1814 static __isl_give isl_ast_node
*create_access_leaf(struct ppcg_kernel
*kernel
,
1815 struct gpu_array_ref_group
*group
, __isl_take isl_ast_node
*node
,
1816 __isl_keep isl_ast_build
*build
)
1818 struct ppcg_kernel_stmt
*stmt
;
1819 struct gpu_array_tile
*tile
;
1824 isl_pw_multi_aff
*pma
, *pma2
;
1827 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1829 return isl_ast_node_free(node
);
1831 access
= isl_map_from_union_map(isl_ast_build_get_schedule(build
));
1832 type
= isl_map_get_tuple_name(access
, isl_dim_in
);
1833 stmt
->u
.c
.read
= !strcmp(type
, "read");
1834 access
= isl_map_reverse(access
);
1835 pma
= isl_pw_multi_aff_from_map(access
);
1836 pma
= isl_pw_multi_aff_reset_tuple_id(pma
, isl_dim_out
);
1838 space
= isl_space_range(isl_pw_multi_aff_get_space(pma
));
1839 space
= isl_space_unwrap(space
);
1840 pma2
= isl_pw_multi_aff_range_map(space
);
1841 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
,
1842 isl_pw_multi_aff_copy(pma
));
1843 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1844 stmt
->u
.c
.index
= expr
;
1846 tile
= gpu_array_ref_group_tile(group
);
1847 pma2
= isl_pw_multi_aff_from_multi_aff(
1848 isl_multi_aff_copy(tile
->tiling
));
1849 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
, pma
);
1850 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1851 stmt
->u
.c
.local_index
= expr
;
1853 stmt
->u
.c
.array
= group
->array
;
1854 stmt
->u
.c
.local_array
= group
->local_array
;
1855 stmt
->type
= ppcg_kernel_copy
;
1857 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1858 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1859 return isl_ast_node_set_annotation(node
, id
);
1862 /* Create a synchronization ppcg_kernel_stmt and
1863 * attach it to the node "node" representing the synchronization.
1865 static __isl_give isl_ast_node
*create_sync_leaf(
1866 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1867 __isl_keep isl_ast_build
*build
)
1869 struct ppcg_kernel_stmt
*stmt
;
1872 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1874 return isl_ast_node_free(node
);
1876 stmt
->type
= ppcg_kernel_sync
;
1877 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1878 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1879 return isl_ast_node_set_annotation(node
, id
);
1882 /* Internal data structure for at_domain.
1884 * "prog" represents the entire scop.
1885 * "kernel" points to the kernel to which the current schedule node
1886 * belongs. It is set by before_mark and reset by after_mark.
1888 struct ppcg_at_domain_data
{
1889 struct gpu_prog
*prog
;
1890 struct ppcg_kernel
*kernel
;
1893 /* This function is called for each instance of a user statement
1894 * in the kernel. This may be one of the original user statements
1895 * or a statement introduced by PPCG.
1897 * We assume that the original user statements only have a name
1898 * and no user pointer. The statements introduced by PPCG
1899 * on the other hand all have a user pointer.
1901 * If the user statement is one of the original user statements
1902 * (one with no user pointer), then we call create_domain_leaf. Otherwise,
1903 * we check if it is a copy or synchronization statement and
1904 * call the appropriate functions.
1905 * Statements that copy an array to/from the device do not need
1906 * any further treatment.
1908 static __isl_give isl_ast_node
*at_domain(__isl_take isl_ast_node
*node
,
1909 __isl_keep isl_ast_build
*build
, void *user
)
1911 struct ppcg_at_domain_data
*data
= user
;
1912 isl_ast_expr
*expr
, *arg
;
1918 expr
= isl_ast_node_user_get_expr(node
);
1919 arg
= isl_ast_expr_get_op_arg(expr
, 0);
1920 id
= isl_ast_expr_get_id(arg
);
1921 name
= isl_id_get_name(id
);
1922 p
= isl_id_get_user(id
);
1923 isl_ast_expr_free(expr
);
1924 isl_ast_expr_free(arg
);
1927 struct gpu_stmt
*gpu_stmt
;
1929 gpu_stmt
= find_stmt(data
->prog
, id
);
1932 isl_die(data
->prog
->ctx
, isl_error_internal
,
1933 "statement not found",
1934 return isl_ast_node_free(node
));
1936 return create_domain_leaf(data
->kernel
, node
, build
, gpu_stmt
);
1939 is_sync
= gpu_tree_id_is_sync(id
, data
->kernel
);
1941 if (!prefixcmp(name
, "to_device_") || !prefixcmp(name
, "from_device_"))
1944 return isl_ast_node_free(node
);
1945 if (!strcmp(name
, "read") || !strcmp(name
, "write")) {
1946 struct gpu_array_ref_group
*group
= p
;
1947 return create_access_leaf(data
->kernel
, group
, node
, build
);
1950 isl_die(data
->prog
->ctx
, isl_error_internal
,
1951 "unknown statement type",
1952 return isl_ast_node_free(node
));
1953 return create_sync_leaf(data
->kernel
, node
, build
);
1956 /* Given a set of wrapped references "ref", return the corresponding
1957 * access relations based on the tagged access relations "tagged".
1959 * The elements of "ref" are of the form
1963 * with D an iteration domains and R a reference.
1964 * The elements of "tagged" are of the form
1970 * Extend "tagged" to include the iteration domain in the range, i.e.,
1972 * [D -> R] -> [D -> A]
1974 * apply the result to "ref" and then unwrap the resulting set
1975 * to obtain relations of the form
1979 static __isl_give isl_union_map
*wrapped_reference_to_access(
1980 __isl_take isl_union_set
*ref
, __isl_take isl_union_map
*tagged
)
1982 isl_union_map
*tag2access
;
1984 tag2access
= isl_union_map_copy(tagged
);
1985 tag2access
= isl_union_map_universe(tag2access
);
1986 tag2access
= isl_union_set_unwrap(isl_union_map_domain(tag2access
));
1987 tag2access
= isl_union_map_domain_map(tag2access
);
1988 tag2access
= isl_union_map_range_product(tag2access
, tagged
);
1990 ref
= isl_union_set_coalesce(ref
);
1991 ref
= isl_union_set_apply(ref
, tag2access
);
1993 return isl_union_set_unwrap(ref
);
1996 /* Given an access relation "access" from one or more array reference groups,
1997 * remove those reads if ("read" is 1) or writes (if "read" is 0)
1998 * that are only needed to communicate data within
1999 * the same iteration of "sched".
2000 * "tagged" contains all tagged access relations to all
2001 * the array reference groups accessed by "access" from statement
2002 * instances scheduled by "sched".
2004 * If the access is a read then it is either an element of
2006 * live_in union (range flow)
2008 * where live_in and flow may be overapproximations, or
2009 * it reads an uninitialized value (that is not live-in because
2010 * there is an intermediate kill) or it reads a value that was
2011 * written within the same (compound) statement instance.
2012 * If the access is a write then it is either an element of
2014 * live_out union (domain flow)
2016 * or it writes a value that is never read (and is not live-out
2017 * because of an intermediate kill) or only
2018 * within the same (compound) statement instance.
2019 * In both cases, the access relation is also a subset of
2020 * the group access relation.
2022 * The cases where an uninitialized value is read or a value is written
2023 * that is never read or where the dataflow occurs within a statement
2024 * instance are also considered local and may also be removed.
2026 * Essentially, we compute the intersection of "access" with either
2028 * live_in union (range non-local-flow)
2032 * live_out union (domain non-local-flow)
2034 * We first construct a relation "local"
2036 * [[D -> R] -> [D' -> R']]
2038 * of pairs of domain iterations accessing the reference group
2039 * and references in the group that are coscheduled by "sched".
2041 * If this relation does not intersect the dataflow dependences,
2042 * then there is nothing we can possibly remove, unless the dataflow
2043 * dependences themselves only relate a subset of the accesses.
2044 * In particular, the accesses may not be involved in any dataflow
2045 * dependences, either because they are uninitialized reads/dead writes
2046 * or because the dataflow occurs inside a statement instance.
2048 * Since the computation below may break up the access relation
2049 * into smaller pieces, we only perform the intersection with
2050 * the non-local dependent accesses if the local pairs
2051 * intersect the dataflow dependences. Otherwise, we intersect
2052 * with the universe of the non-local dependent accesses.
2053 * This should at least remove accesses from statements that
2054 * do not participate in any dependences.
2056 * In particular, we remove the "local" dataflow dependences from
2057 * the set of all dataflow dependences.
2058 * Note that if the potential dataflow dependences are an overapproximation
2059 * of the actual dataflow dependences, then the result remains an
2060 * overapproximation of the non-local dataflow dependences.
2061 * Copying to/from global memory is only needed for the references
2062 * in the domain/range of the result or for accesses that are live out/in
2063 * for the entire scop.
2065 * We therefore map the domain/range of the "external" relation
2066 * to the corresponding access relation and take the union with
2067 * the live out/in relation.
2069 static __isl_give isl_union_map
*remove_local_accesses(
2070 struct gpu_prog
*prog
, __isl_take isl_union_map
*tagged
,
2071 __isl_take isl_union_map
*access
, __isl_take isl_union_map
*sched
,
2075 isl_union_pw_multi_aff
*tagger
;
2076 isl_union_set
*domain
;
2077 isl_union_map
*local
, *external
;
2078 isl_union_set
*tag_set
;
2080 if (isl_union_map_is_empty(access
)) {
2081 isl_union_map_free(sched
);
2082 isl_union_map_free(tagged
);
2086 tagger
= isl_union_pw_multi_aff_copy(prog
->scop
->tagger
);
2087 domain
= isl_union_map_domain(isl_union_map_copy(tagged
));
2088 tagger
= isl_union_pw_multi_aff_intersect_domain(tagger
, domain
);
2089 sched
= isl_union_map_preimage_domain_union_pw_multi_aff(sched
, tagger
);
2091 local
= isl_union_map_apply_range(sched
,
2092 isl_union_map_reverse(isl_union_map_copy(sched
)));
2093 local
= isl_union_map_intersect(local
,
2094 isl_union_map_copy(prog
->scop
->tagged_dep_flow
));
2096 empty
= isl_union_map_is_empty(local
);
2098 external
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
2099 external
= isl_union_map_intersect_params(external
,
2100 isl_set_copy(prog
->scop
->context
));
2101 external
= isl_union_map_subtract(external
, local
);
2104 tag_set
= isl_union_map_range(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_in
));
2109 tag_set
= isl_union_map_domain(external
);
2110 external
= wrapped_reference_to_access(tag_set
, tagged
);
2111 external
= isl_union_map_union(external
,
2112 isl_union_map_copy(prog
->scop
->live_out
));
2116 external
= isl_union_map_free(external
);
2118 external
= isl_union_map_universe(external
);
2120 access
= isl_union_map_intersect(access
, external
);
2125 /* Given an access relation "access" from "group", remove those reads
2126 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
2127 * communicate data within the same iteration of the schedule at the
2128 * position where the copying of the group is inserted.
2129 * "node" points to this position, i.e., the depth at "node"
2130 * is equal to group->depth.
2132 * We extract a schedule that picks out the iterations of the outer
2133 * group->depth dimensions and call remove_local_accesses.
2135 static __isl_give isl_union_map
*remove_local_accesses_group(
2136 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2137 __isl_take isl_union_map
*access
, __isl_keep isl_schedule_node
*node
,
2140 isl_union_map
*sched
, *tagged
;
2142 if (isl_union_map_is_empty(access
))
2145 tagged
= group_tagged_access_relation(group
);
2146 sched
= isl_schedule_node_get_prefix_schedule_relation(node
);
2148 return remove_local_accesses(kernel
->prog
, tagged
, access
, sched
, read
);
2151 /* This function is called before the AST generator starts traversing
2152 * the schedule subtree of a node with mark "mark".
2154 * If the mark is called "kernel", store the kernel pointer in data->kernel
2155 * for use in at_domain.
2157 static int before_mark(__isl_keep isl_id
*mark
,
2158 __isl_keep isl_ast_build
*build
, void *user
)
2160 struct ppcg_at_domain_data
*data
= user
;
2164 if (!strcmp(isl_id_get_name(mark
), "kernel"))
2165 data
->kernel
= isl_id_get_user(mark
);
2169 /* This function is called after the AST generator has finished traversing
2170 * the schedule subtree of a mark node. "node" points to the corresponding
2173 * If the mark is called "kernel", then replace "node" by a user node
2174 * that "calls" the kernel, representing the launch of the kernel.
2175 * The original "node" is stored inside the kernel object so that
2176 * it can be used to print the device code.
2177 * Note that this assumes that a kernel is only launched once.
2178 * Also clear data->kernel.
2180 static __isl_give isl_ast_node
*after_mark(__isl_take isl_ast_node
*node
,
2181 __isl_keep isl_ast_build
*build
, void *user
)
2186 isl_ast_expr_list
*list
;
2187 struct ppcg_kernel
*kernel
;
2188 struct ppcg_at_domain_data
*data
= user
;
2190 ctx
= isl_ast_node_get_ctx(node
);
2191 id
= isl_ast_node_mark_get_id(node
);
2193 return isl_ast_node_free(node
);
2194 if (strcmp(isl_id_get_name(id
), "kernel") || !data
->kernel
) {
2198 kernel
= data
->kernel
;
2199 data
->kernel
= NULL
;
2200 kernel
->space
= isl_ast_build_get_schedule_space(build
);
2201 kernel
->tree
= isl_ast_node_mark_get_node(node
);
2202 isl_ast_node_free(node
);
2204 expr
= isl_ast_expr_from_id(isl_id_copy(id
));
2205 list
= isl_ast_expr_list_alloc(ctx
, 0);
2206 expr
= isl_ast_expr_call(expr
, list
);
2207 node
= isl_ast_node_alloc_user(expr
);
2208 node
= isl_ast_node_set_annotation(node
, id
);
2213 static int update_depth(__isl_keep isl_schedule_node
*node
, void *user
)
2218 if (isl_schedule_node_get_type(node
) != isl_schedule_node_leaf
)
2220 node_depth
= isl_schedule_node_get_schedule_depth(node
);
2221 if (node_depth
> *depth
)
2222 *depth
= node_depth
;
2227 /* Use isl to generate code for both the host and the device
2229 * The device code is marked by "kernel" mark nodes in the schedule tree,
2230 * containing a pointer to a ppcg_kernel object.
2231 * The returned AST only contains the AST for the host code.
2232 * The ASTs for the device code are embedded in ppcg_kernel objects
2233 * attached to the leaf nodes that call "kernel".
2235 static __isl_give isl_ast_node
*generate_code(struct gpu_gen
*gen
,
2236 __isl_take isl_schedule
*schedule
)
2238 struct ppcg_at_domain_data data
;
2239 isl_ast_build
*build
;
2241 isl_id_list
*iterators
;
2244 data
.prog
= gen
->prog
;
2248 if (isl_schedule_foreach_schedule_node(schedule
, &update_depth
,
2251 build
= isl_ast_build_alloc(gen
->prog
->ctx
);
2252 iterators
= ppcg_scop_generate_names(gen
->prog
->scop
, depth
, "c");
2253 build
= isl_ast_build_set_iterators(build
, iterators
);
2254 build
= isl_ast_build_set_at_each_domain(build
, &at_domain
, &data
);
2255 build
= isl_ast_build_set_before_each_mark(build
, &before_mark
, &data
);
2256 build
= isl_ast_build_set_after_each_mark(build
, &after_mark
, &data
);
2257 if (gen
->prog
->scop
->options
->debug
->dump_final_schedule
)
2258 isl_schedule_dump(schedule
);
2259 tree
= isl_ast_build_node_from_schedule(build
, schedule
);
2260 isl_ast_build_free(build
);
2265 __isl_give isl_union_map
*extract_sizes_from_str(isl_ctx
*ctx
, const char *str
)
2269 return isl_union_map_read_from_str(ctx
, str
);
2272 /* Can "node" be tiled and then mapped to block and thread identifiers?
2273 * That is, is it permutable with at least one coincident dimension?
2275 static int is_permutable(__isl_keep isl_schedule_node
*node
)
2280 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
2282 if (!isl_schedule_node_band_get_permutable(node
))
2284 if (isl_schedule_node_band_n_member(node
) < 1)
2286 if (!isl_schedule_node_band_member_get_coincident(node
, 0))
2292 /* A isl_schedule_foreach_schedule_node callback
2293 * for setting *any_permutable and aborting the search
2294 * if "node" is a permutable band with coincident dimensions.
2295 * Otherwise, continue searching.
2297 static int set_permutable(__isl_keep isl_schedule_node
*node
, void *user
)
2299 int *any_permutable
= user
;
2302 permutable
= is_permutable(node
);
2308 *any_permutable
= 1;
2313 /* Does "schedule" contain any permutable band with at least one coincident
2316 static int has_any_permutable_node(__isl_keep isl_schedule
*schedule
)
2318 int any_permutable
= 0;
2320 if (isl_schedule_foreach_schedule_node(schedule
, &set_permutable
,
2321 &any_permutable
) < 0 &&
2325 return any_permutable
;
2328 /* Is "node" a leaf or can it be tiled and then mapped to
2329 * block and thread identifiers?
2331 static int is_leaf_or_tilable(__isl_keep isl_schedule_node
*node
)
2333 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
2335 return is_permutable(node
);
2338 /* Is "node" the outermost node in its branch that can be tiled
2339 * and then mapped to block and thread identifiers?
2340 * If there are no such nodes in the branch and if "node" is a leaf,
2341 * then it is accepted too.
2343 static int is_outer_tilable(__isl_keep isl_schedule_node
*node
)
2346 isl_schedule_node
*ancestor
;
2348 tilable
= is_leaf_or_tilable(node
);
2355 ancestor
= isl_schedule_node_copy(node
);
2356 while (isl_schedule_node_has_parent(ancestor
)) {
2357 ancestor
= isl_schedule_node_parent(ancestor
);
2359 tilable
= is_permutable(ancestor
);
2360 if (tilable
< 0 || tilable
)
2364 isl_schedule_node_free(ancestor
);
2365 return tilable
< 0 ? -1 : !tilable
;
2368 /* Collect the references to all writes in "group".
2369 * Each reference is represented by a universe set in a space
2373 * with S[i,j] the statement instance space and R[] the array reference.
2375 static __isl_give isl_union_set
*group_tagged_writes(
2376 struct gpu_array_ref_group
*group
)
2380 isl_union_set
*writes
;
2382 space
= isl_map_get_space(group
->access
);
2383 writes
= isl_union_set_empty(space
);
2384 for (i
= 0; i
< group
->n_ref
; ++i
) {
2388 if (!group
->refs
[i
]->write
)
2391 space
= isl_map_get_space(group
->refs
[i
]->tagged_access
);
2392 space
= isl_space_domain(space
);
2393 writes_i
= isl_set_universe(space
);
2394 writes
= isl_union_set_add_set(writes
, writes_i
);
2400 /* Is there any write access in "group" that requires synchronization
2401 * on a write to global memory?
2402 * We currently take into account all writes that would require
2403 * synchronization at the thread level depth, but if the copying
2404 * for this group is performed at an outer level, then we do not
2405 * actually need to take into account dependences at intermediate levels.
2407 static int any_sync_writes_in_group(struct ppcg_kernel
*kernel
,
2408 struct gpu_array_ref_group
*group
)
2410 isl_union_set
*writes
;
2411 int empty
, disjoint
;
2413 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2419 writes
= group_tagged_writes(group
);
2420 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2421 isl_union_set_free(writes
);
2423 return disjoint
< 0 ? -1 : !disjoint
;
2426 /* Collect the references to all writes in "kernel" that write directly
2427 * to global or shared memory, i.e., that are not mapped to private memory.
2428 * Each reference is represented by a universe set in a space
2432 * with S[i,j] the statement instance space and R[] the array reference.
2434 static __isl_give isl_union_set
*collect_non_private_tagged_writes(
2435 struct ppcg_kernel
*kernel
)
2437 isl_union_set
*writes
;
2440 writes
= isl_union_set_empty(isl_union_set_get_space(kernel
->arrays
));
2442 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2443 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2445 for (j
= 0; j
< array
->n_group
; ++j
) {
2446 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2447 isl_union_set
*writes_ij
;
2451 if (group
->private_tile
)
2453 writes_ij
= group_tagged_writes(group
);
2454 writes
= isl_union_set_union(writes
, writes_ij
);
2461 /* Are there any direct writes to global memory that require
2464 static int any_global_or_shared_sync_writes(struct ppcg_kernel
*kernel
)
2466 isl_union_set
*writes
;
2467 int empty
, disjoint
;
2469 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2475 writes
= collect_non_private_tagged_writes(kernel
);
2476 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2477 isl_union_set_free(writes
);
2479 return disjoint
< 0 ? -1 : !disjoint
;
2482 /* Construct an isl_multi_val for use as tile sizes for tiling "node"
2483 * from the elements in "tile_size".
2485 static __isl_give isl_multi_val
*construct_band_tiles_sizes(
2486 __isl_keep isl_schedule_node
*node
, int *tile_size
)
2496 ctx
= isl_schedule_node_get_ctx(node
);
2497 space
= isl_schedule_node_band_get_space(node
);
2498 n
= isl_schedule_node_band_n_member(node
);
2499 mv
= isl_multi_val_zero(space
);
2500 for (i
= 0; i
< n
; ++i
) {
2503 v
= isl_val_int_from_si(ctx
, tile_size
[i
]);
2504 mv
= isl_multi_val_set_val(mv
, i
, v
);
2510 /* Replace the partial schedule S of the band node "node" by
2518 * if scale_tile_loops is set, with f the integers in "factor".
2519 * The list that "factor" points to is assumed to contain at least
2520 * as many elements as the number of members in the band.
2522 static __isl_give isl_schedule_node
*snap_band_to_sizes(
2523 __isl_take isl_schedule_node
*node
, int *factor
,
2524 struct ppcg_options
*options
)
2528 mv
= construct_band_tiles_sizes(node
, factor
);
2529 node
= isl_schedule_node_band_scale_down(node
, isl_multi_val_copy(mv
));
2530 if (options
->scale_tile_loops
)
2531 node
= isl_schedule_node_band_scale(node
,
2532 isl_multi_val_copy(mv
));
2533 isl_multi_val_free(mv
);
2538 /* Tile "band" with tile size specified by "sizes".
2540 * Since the tile loops will be mapped to block ids, we forcibly
2541 * turn off tile loop scaling. We may want to enable tile loop scaling
2542 * at some later point, but then we would have to support the detection
2543 * of strides during the mapping to block ids.
2544 * Similarly, since the point loops will be mapped to thread ids,
2545 * we forcibly shift the point loops so that they start at zero.
2547 static __isl_give isl_schedule_node
*tile_band(
2548 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2550 isl_ctx
*ctx
= isl_schedule_node_get_ctx(node
);
2554 scale_tile
= isl_options_get_tile_scale_tile_loops(ctx
);
2555 isl_options_set_tile_scale_tile_loops(ctx
, 0);
2556 shift_point
= isl_options_get_tile_shift_point_loops(ctx
);
2557 isl_options_set_tile_shift_point_loops(ctx
, 1);
2559 node
= isl_schedule_node_band_tile(node
, sizes
);
2561 isl_options_set_tile_scale_tile_loops(ctx
, scale_tile
);
2562 isl_options_set_tile_shift_point_loops(ctx
, shift_point
);
2567 /* Extract the set of parameter values and outer schedule dimensions
2568 * for which any statement instance
2569 * in the kernel inserted at "node" needs to be executed.
2570 * Intersect the set of parameter values derived from the host schedule
2571 * relation with the context of "prog".
2573 static __isl_give isl_set
*extract_context(__isl_keep isl_schedule_node
*node
,
2574 struct gpu_prog
*prog
)
2576 isl_union_map
*schedule
;
2577 isl_union_set
*schedule_domain
;
2581 schedule
= isl_schedule_node_get_prefix_schedule_relation(node
);
2582 schedule_domain
= isl_union_map_range(schedule
);
2583 empty
= isl_union_set_is_empty(schedule_domain
);
2585 isl_union_set_free(schedule_domain
);
2592 space
= isl_union_set_get_space(schedule_domain
);
2593 isl_union_set_free(schedule_domain
);
2594 space
= isl_space_set_from_params(space
);
2595 depth
= isl_schedule_node_get_schedule_depth(node
);
2596 space
= isl_space_add_dims(space
, isl_dim_set
, depth
);
2597 context
= isl_set_empty(space
);
2599 context
= isl_set_from_union_set(schedule_domain
);
2601 context
= isl_set_intersect_params(context
,
2602 isl_set_copy(prog
->context
));
2607 /* Return the set of outer array elements accessed by
2608 * by the statement instance in "domain" in "prog".
2610 static __isl_give isl_union_set
*accessed_by_domain(
2611 __isl_take isl_union_set
*domain
, struct gpu_prog
*prog
)
2613 isl_union_map
*access
;
2614 isl_union_set
*arrays
;
2616 access
= isl_union_map_union(isl_union_map_copy(prog
->read
),
2617 isl_union_map_copy(prog
->may_write
));
2618 access
= isl_union_map_intersect_domain(access
, domain
);
2619 arrays
= isl_union_map_range(access
);
2620 arrays
= isl_union_set_apply(arrays
,
2621 isl_union_map_copy(prog
->to_outer
));
2626 /* Return the number of outer band members of the band node "node"
2627 * that are marked coincident.
2629 static int n_outer_coincidence(__isl_keep isl_schedule_node
*node
)
2633 n
= isl_schedule_node_band_n_member(node
);
2635 for (i
= 0; i
< n
; ++i
)
2636 if (!isl_schedule_node_band_member_get_coincident(node
, i
))
2642 /* If the band node "node" has more than "n" members, then split off
2643 * the first "n" of them.
2645 static __isl_give isl_schedule_node
*split_band(
2646 __isl_take isl_schedule_node
*node
, int n
)
2650 dim
= isl_schedule_node_band_n_member(node
);
2652 node
= isl_schedule_node_band_split(node
, n
);
2657 /* Scale a band node that may have been split by split_band.
2658 * "sizes" are the scaling factors for the original node.
2659 * "node" either points to the original band node, or the outer
2660 * of the two pieces after splitting.
2662 * If the number of elements in "node" is smaller than the number of
2663 * elements in "sizes", then some splitting has occurred and we split
2664 * "sizes" in the same way.
2666 static __isl_give isl_schedule_node
*scale_band(
2667 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2671 n
= isl_multi_val_dim(sizes
, isl_dim_set
);
2672 dim
= isl_schedule_node_band_n_member(node
);
2674 isl_multi_val
*sizes2
;
2676 sizes2
= isl_multi_val_copy(sizes
);
2677 sizes
= isl_multi_val_drop_dims(sizes
,
2678 isl_dim_set
, dim
, n
- dim
);
2679 sizes2
= isl_multi_val_drop_dims(sizes2
, isl_dim_set
, 0, dim
);
2680 node
= isl_schedule_node_child(node
, 0);
2681 node
= isl_schedule_node_band_scale(node
, sizes2
);
2682 node
= isl_schedule_node_parent(node
);
2685 return isl_schedule_node_band_scale(node
, sizes
);
2688 /* Return an isl_multi_aff, with as elements the parameters in "space"
2689 * that have the names specified by the elements in "names".
2690 * If (some of) these parameters do not already appear in "space",
2691 * then they are added first.
2693 static __isl_give isl_multi_aff
*parameter_vector(__isl_take isl_space
*space
,
2694 __isl_keep isl_id_list
*names
)
2697 isl_local_space
*ls
;
2701 space
= isl_space_free(space
);
2703 n
= isl_id_list_n_id(names
);
2704 for (i
= 0; i
< n
; ++i
) {
2708 id
= isl_id_list_get_id(names
, i
);
2709 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2714 pos
= isl_space_dim(space
, isl_dim_param
);
2715 space
= isl_space_add_dims(space
, isl_dim_param
, 1);
2716 space
= isl_space_set_dim_id(space
, isl_dim_param
, pos
, id
);
2718 ma
= isl_multi_aff_zero(isl_space_copy(space
));
2719 ls
= isl_local_space_from_space(isl_space_domain(space
));
2720 for (i
= 0; i
< n
; ++i
) {
2725 id
= isl_id_list_get_id(names
, i
);
2726 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2728 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
2729 isl_dim_param
, pos
);
2730 ma
= isl_multi_aff_set_aff(ma
, i
, aff
);
2732 isl_local_space_free(ls
);
2737 /* Return constraints on the domain elements that equate a sequence of
2738 * parameters called "names", to the partial schedule
2739 * of "node" modulo the integers in "size".
2740 * The number of elements in the array "size" should be equal
2741 * to the number of elements in "names".
2742 * The number of members of the band node "node" should be smaller
2743 * than or equal to this number. If it is smaller, then the first
2744 * elements of "names" are equated to zero.
2746 static __isl_give isl_union_set
*set_schedule_modulo(
2747 __isl_keep isl_schedule_node
*node
, __isl_keep isl_id_list
*names
,
2753 isl_multi_union_pw_aff
*mupa
, *mupa2
;
2755 isl_union_set
*domain
;
2759 n
= isl_id_list_n_id(names
);
2761 return isl_schedule_node_get_universe_domain(node
);
2762 n_zero
= n
- isl_schedule_node_band_n_member(node
);
2764 mupa
= isl_schedule_node_band_get_partial_schedule(node
);
2765 mv
= construct_band_tiles_sizes(node
, size
+ n_zero
);
2766 mupa
= isl_multi_union_pw_aff_mod_multi_val(mupa
, mv
);
2768 space
= isl_multi_union_pw_aff_get_space(mupa
);
2769 space
= isl_space_params(space
);
2770 space
= isl_space_set_from_params(space
);
2771 space
= isl_space_add_dims(space
, isl_dim_set
, n_zero
);
2772 ma
= isl_multi_aff_zero(space
);
2774 domain
= isl_schedule_node_get_universe_domain(node
);
2775 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(
2776 isl_union_set_copy(domain
), ma
);
2777 mupa
= isl_multi_union_pw_aff_range_product(mupa2
, mupa
);
2779 space
= isl_multi_union_pw_aff_get_space(mupa
);
2780 ma
= parameter_vector(space
, names
);
2782 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(domain
, ma
);
2783 mupa
= isl_multi_union_pw_aff_sub(mupa
, mupa2
);
2785 return isl_multi_union_pw_aff_zero_union_set(mupa
);
2788 /* Insert a context node at "node" introducing the block and thread
2789 * identifiers along with their bounds, which are stored in kernel->grid_size
2790 * and kernel->block_dim.
2791 * Note that the bounds on the block identifiers may implicitly impose
2792 * constraints on the parameters. A guard needs to be inserted
2793 * in the schedule tree to ensure that those bounds hold at "node".
2794 * This guard is inserted in insert_guard.
2796 static __isl_give isl_schedule_node
*insert_context(struct ppcg_kernel
*kernel
,
2797 __isl_take isl_schedule_node
*node
)
2801 context
= isl_set_universe(isl_set_get_space(kernel
->context
));
2803 context
= add_bounded_parameters_dynamic(context
,
2804 kernel
->grid_size
, kernel
->block_ids
);
2805 context
= add_bounded_parameters(context
,
2806 kernel
->block_dim
, kernel
->thread_ids
);
2808 node
= isl_schedule_node_insert_context(node
, context
);
2813 /* Insert a guard that eliminates kernel launches where the kernel
2814 * obviously does not have any work to do.
2816 * In particular, eliminate kernel launches where there are obviously
2818 * Use the same block size constraints that are used to create the context
2819 * to ensure that all constraints implicit in the constructed context
2820 * are imposed by the guard.
2822 * Additionally, add other constraints that are valid
2823 * for each executed instance ("context"), as long as this does not result
2826 static __isl_give isl_schedule_node
*insert_guard(
2827 __isl_take isl_schedule_node
*node
, __isl_keep isl_set
*context
,
2828 __isl_keep isl_multi_pw_aff
*size
, struct ppcg_scop
*scop
)
2834 guard
= isl_set_copy(context
);
2835 guard
= isl_set_compute_divs(guard
);
2836 guard
= isl_set_from_basic_set(isl_set_simple_hull(guard
));
2838 nparam
= isl_set_dim(guard
, isl_dim_param
);
2839 n
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
2840 ids
= ppcg_scop_generate_names(scop
, n
, "__ppcg_tmp");
2841 guard
= add_bounded_parameters_dynamic(guard
, size
, ids
);
2842 isl_id_list_free(ids
);
2843 guard
= isl_set_project_out(guard
, isl_dim_param
, nparam
, n
);
2845 node
= isl_schedule_node_insert_guard(node
, guard
);
2850 /* Does any array reference group mapping require the band that is mapped
2851 * to threads to be unrolled?
2853 static int kernel_requires_unroll(struct ppcg_kernel
*kernel
)
2857 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2858 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2860 for (j
= 0; j
< array
->n_group
; ++j
) {
2861 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2862 if (gpu_array_ref_group_requires_unroll(group
))
2870 /* Mark the given band node "node" for unrolling by the AST generator and
2871 * then sink it to the leaves of the schedule tree.
2872 * All dimensions of "node" are assumed to be coincident, such that this
2873 * sinking is a valid operation.
2875 static __isl_give isl_schedule_node
*unroll(__isl_take isl_schedule_node
*node
)
2879 n
= isl_schedule_node_band_n_member(node
);
2880 for (i
= 0; i
< n
; ++i
)
2881 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
2882 isl_ast_loop_unroll
);
2884 node
= isl_schedule_node_band_sink(node
);
2889 /* Insert a synchronization node in the schedule tree of "node"
2890 * after the core computation of "kernel" at the level of the band
2891 * that is mapped to threads, except if that level is equal to
2892 * that of the band that is mapped to blocks or if there are no writes
2893 * to global or shared memory in the core computation that require
2895 * If there are any writes to shared memory and the shared memory
2896 * copying is performed at the same level, then synchronization
2897 * is needed between the core and the copying anyway, so we might
2898 * as well add it here. If the copying is performed at a higher
2899 * level, then different iterations of intermediate schedule dimensions
2900 * may have a different mapping from between shared memory elements and
2901 * threads, such that synchronization is required after the core.
2902 * "node" is assumed to point to the kernel node.
2904 static __isl_give isl_schedule_node
*add_sync(struct ppcg_kernel
*kernel
,
2905 __isl_take isl_schedule_node
*node
)
2910 need_sync
= any_global_or_shared_sync_writes(kernel
);
2912 return isl_schedule_node_free(node
);
2916 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
2918 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
2919 if (kernel_depth
== isl_schedule_node_get_schedule_depth(node
))
2920 return gpu_tree_move_up_to_kernel(node
);
2922 node
= gpu_tree_ensure_following_sync(node
, kernel
);
2924 node
= gpu_tree_move_up_to_kernel(node
);
2929 /* Return a read ("read" is 1) or write access relation for "group"
2930 * with those accesses removed that are only needed to communicate data
2931 * within the subtree of the schedule rooted at "node".
2932 * Furthermore, include the prefix schedule at "node".
2933 * That is, return a relation of the form
2937 * with D the outer schedule dimensions at "node".
2939 static __isl_give isl_union_map
*anchored_non_local_accesses(
2940 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2941 __isl_take isl_schedule_node
*node
, int read
)
2943 isl_union_map
*access
;
2944 isl_union_map
*prefix
;
2946 access
= gpu_array_ref_group_access_relation(group
, read
, !read
);
2947 access
= remove_local_accesses_group(kernel
, group
, access
, node
, read
);
2948 prefix
= isl_schedule_node_get_prefix_schedule_relation(node
);
2949 access
= isl_union_map_range_product(prefix
, access
);
2954 /* Given an array reference group "group", create a mapping
2956 * read[D -> A] -> [D -> A]
2958 * if "read" is set or
2960 * write[D -> A] -> [D -> A]
2962 * if "read" is not set.
2963 * D corresponds to the outer group->depth dimensions of
2964 * the kernel schedule.
2966 static __isl_give isl_multi_aff
*create_from_access(isl_ctx
*ctx
,
2967 struct gpu_array_ref_group
*group
, int read
)
2972 space
= isl_space_copy(group
->array
->space
);
2973 space
= isl_space_from_range(space
);
2974 space
= isl_space_add_dims(space
, isl_dim_in
, group
->depth
);
2975 space
= isl_space_wrap(space
);
2976 space
= isl_space_map_from_set(space
);
2978 id
= isl_id_alloc(ctx
, read
? "read" : "write", group
);
2979 space
= isl_space_set_tuple_id(space
, isl_dim_in
, id
);
2981 return isl_multi_aff_identity(space
);
2984 /* If any writes in "group" require synchronization, then make sure
2985 * that there is a synchronization node for "kernel" after the node
2986 * following "node" in a sequence.
2988 * If "shared" is set and no synchronization is needed for
2989 * the writes to global memory, then add synchronization before
2990 * the kernel to protect shared memory from being overwritten
2991 * by the next iteration of the core computation.
2992 * No additional synchronization is needed to protect against
2993 * the next copy into shared memory because each element of
2994 * the shared memory tile is always copied by the same thread.
2996 static __isl_give isl_schedule_node
*add_group_write_sync(
2997 __isl_take isl_schedule_node
*node
, struct ppcg_kernel
*kernel
,
2998 struct gpu_array_ref_group
*group
, int shared
)
3002 need_sync
= any_sync_writes_in_group(kernel
, group
);
3004 return isl_schedule_node_free(node
);
3006 node
= isl_schedule_node_parent(node
);
3007 node
= isl_schedule_node_next_sibling(node
);
3008 node
= isl_schedule_node_child(node
, 0);
3009 node
= gpu_tree_ensure_following_sync(node
, kernel
);
3010 } else if (shared
) {
3011 node
= isl_schedule_node_parent(node
);
3012 node
= isl_schedule_node_parent(node
);
3013 node
= gpu_tree_move_down_to_depth(node
, group
->depth
,
3015 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3021 /* Add copy statements to the schedule tree of "node"
3022 * for reading from global memory to private memory (if "read" is set) or
3023 * for writing back from private memory to global memory
3024 * (if "read" is not set) for the array reference group "group" that
3025 * is mapped to private memory.
3026 * On input, "node" points to the kernel node, and it is moved
3027 * back there on output.
3029 * The copies are performed in the order of the array elements.
3030 * The copy statement instances include a reference to the outer
3031 * group->depth dimensions of the kernel schedule for ease of
3032 * combining them with the group tiling.
3034 * That is, the extra schedule is of the form
3038 * where D corresponds to the outer group->depth dimensions of
3039 * the kernel schedule and A to the global array.
3040 * This schedule is unrolled because registers are not addressable.
3042 * The copying is inserted in the schedule tree through an extension
3047 * where the extra domain elements type[D -> A] are those accessed
3049 * A filter is inserted on type[D -> A] to ensure that the element
3050 * is read/written by the same thread that needs the element.
3051 * This filter is obtained by applying
3055 * to the thread filter for the core statements.
3057 * The extension is inserted before the core computation in case of a read
3058 * and after the core computation in case of a write.
3059 * In the latter case, we also make sure that there is a synchronization
3060 * node after the write to global memory, unless this write is performed
3061 * at the outer level of the kernel.
3062 * In principle, this synchronization could be inserted higher
3063 * in the schedule tree depending on where the corresponding reads
3064 * from global memory are performed.
3066 static __isl_give isl_schedule_node
*add_copies_group_private(
3067 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3068 __isl_take isl_schedule_node
*node
, int read
)
3070 isl_union_map
*access
;
3071 isl_union_map
*prefix
;
3072 isl_union_set
*domain
;
3074 isl_multi_aff
*from_access
;
3075 isl_multi_pw_aff
*mpa
;
3076 isl_multi_union_pw_aff
*mupa
;
3077 isl_schedule_node
*graft
;
3078 isl_union_set
*filter
;
3082 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3083 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3085 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3086 empty
= isl_union_map_is_empty(access
);
3087 if (empty
< 0 || empty
) {
3088 isl_union_map_free(access
);
3090 return isl_schedule_node_free(node
);
3091 return gpu_tree_move_up_to_kernel(node
);
3094 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3095 space
= isl_space_domain(isl_multi_aff_get_space(from_access
));
3096 access
= isl_union_map_preimage_range_multi_aff(access
, from_access
);
3098 filter
= isl_union_set_copy(kernel
->thread_filter
);
3099 filter
= isl_union_set_apply(filter
, isl_union_map_copy(access
));
3100 filter
= isl_union_set_detect_equalities(filter
);
3101 filter
= isl_union_set_coalesce(filter
);
3103 domain
= isl_union_map_range(access
);
3104 access
= isl_union_set_wrapped_domain_map(domain
);
3105 access
= isl_union_map_reverse(access
);
3106 access
= isl_union_map_coalesce(access
);
3107 graft
= isl_schedule_node_from_extension(access
);
3109 space
= isl_space_map_from_set(space
);
3110 mpa
= isl_multi_pw_aff_identity(space
);
3111 mpa
= isl_multi_pw_aff_range_factor_range(mpa
);
3112 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3114 graft
= isl_schedule_node_child(graft
, 0);
3115 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3116 graft
= unroll(graft
);
3118 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3120 graft
= isl_schedule_node_parent(graft
);
3123 node
= isl_schedule_node_graft_before(node
, graft
);
3125 node
= isl_schedule_node_graft_after(node
, graft
);
3126 if (kernel_depth
< group
->depth
)
3127 node
= add_group_write_sync(node
, kernel
, group
, 0);
3130 node
= gpu_tree_move_up_to_kernel(node
);
3135 /* Add copy statements to the schedule tree of "node"
3136 * for reading from global memory to shared memory (if "read" is set) or
3137 * for writing back from shared memory to global memory
3138 * (if "read" is not set) for the array reference group "group" that
3139 * is mapped to shared memory.
3140 * On input, "node" points to the kernel node, and it is moved
3141 * back there on output.
3143 * The copies are performed in the order of the corresponding shared
3145 * The copy statement instances include a reference to the outer
3146 * group->depth dimensions of the kernel schedule for ease of
3147 * combining them with the group tiling.
3149 * If we are performing a read from global memory to shared memory and
3150 * if the array involved is not a scalar, then we copy
3151 * the entire tile to shared memory. This may result in some extra
3152 * elements getting copied, but it should lead to simpler code
3153 * (which means that fewer registers may be needed) and less divergence.
3155 * Otherwise, we only copy the elements that will be read or have been written
3158 * That is, the extra schedule is of the form
3162 * where D corresponds to the outer group->depth dimensions of
3163 * the kernel schedule, A to the global array and T is the corresponding
3164 * shared memory tile.
3166 * The copying is inserted in the schedule tree through an extension
3171 * where the extra domain elements type[D -> A] are those accessed
3172 * by the group. In the case of read from a non-scalar, this set
3173 * is replaced by the entire shared memory tile.
3175 * A filter is inserted on type[D -> A] to map the copy instances
3176 * to the threads. In particular, the thread identifiers are
3177 * equated to the position inside the shared memory tile (T)
3178 * modulo the block size.
3179 * We try to align the innermost tile dimension with the innermost
3180 * thread identifier (x) as a heuristic to improve coalescing.
3181 * In particular, if the dimension of the tile is greater than
3182 * the dimension of the block, then the schedule mapping to the tile
3183 * is broken up into two pieces and the filter is applied to the inner part.
3184 * If, on the other hand, the dimension of the tile is smaller than
3185 * the dimension of the block, then the initial thread identifiers
3186 * are equated to zero and the remaining thread identifiers are
3187 * matched to the memory tile.
3189 * The extension is inserted before the core computation in case of a read
3190 * and after the core computation in case of a write.
3191 * In the case of a read, we first need to make sure there is some
3192 * synchronization before the core computation such that we can put the read
3193 * from global memory to shared memory before that synchronization.
3194 * This ensures that all threads have finished copying into shared memory
3195 * before the shared memory is used.
3196 * We also need to make sure that there is a synchronization node after
3197 * the core computation to ensure that the next load into shared memory
3198 * only happens after all data has been used. There is no need for
3199 * this synchronization if we are at the outer level since then there
3200 * won't be a next load.
3201 * In the case of a write, we need to make sure there is some synchronization
3202 * after the core computation such taht we can put the write from shared
3203 * memory to global memory after that synchronization.
3204 * Unless we are at the outer level, we also need a synchronization node
3205 * after the write to ensure the data is saved to global memory
3206 * before the next iteration write to the same shared memory.
3207 * It also makes sure the data has arrived in global memory before
3208 * it is read in a subsequent iteration.
3210 static __isl_give isl_schedule_node
*add_copies_group_shared(
3211 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3212 __isl_take isl_schedule_node
*node
, int read
)
3214 struct gpu_array_tile
*tile
;
3215 isl_union_map
*access
;
3216 isl_union_set
*domain
;
3217 isl_union_set
*sync
;
3219 isl_multi_aff
*from_access
;
3220 isl_multi_pw_aff
*mpa
;
3221 isl_multi_union_pw_aff
*mupa
;
3222 isl_schedule_node
*graft
;
3223 isl_union_set
*filter
;
3228 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3229 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3231 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3232 empty
= isl_union_map_is_empty(access
);
3233 if (empty
< 0 || empty
) {
3234 isl_union_map_free(access
);
3236 return isl_schedule_node_free(node
);
3237 return gpu_tree_move_up_to_kernel(node
);
3240 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3242 tile
= gpu_array_ref_group_tile(group
);
3243 ma
= isl_multi_aff_copy(tile
->tiling
);
3244 ma
= isl_multi_aff_pullback_multi_aff(ma
,
3245 isl_multi_aff_copy(from_access
));
3246 mpa
= isl_multi_pw_aff_from_multi_aff(ma
);
3247 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3249 domain
= isl_union_map_range(access
);
3251 if (read
&& !gpu_array_is_scalar(group
->array
)) {
3253 isl_union_set_free(domain
);
3254 map
= group_tile(group
);
3255 domain
= isl_union_set_from_set(isl_map_wrap(map
));
3258 domain
= isl_union_set_preimage_multi_aff(domain
, from_access
);
3259 access
= isl_union_set_wrapped_domain_map(domain
);
3260 access
= isl_union_map_reverse(access
);
3261 access
= isl_union_map_coalesce(access
);
3262 graft
= isl_schedule_node_from_extension(access
);
3264 graft
= isl_schedule_node_child(graft
, 0);
3266 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3268 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0) {
3269 graft
= isl_schedule_node_band_split(graft
,
3270 tile
->n
- kernel
->n_block
);
3271 graft
= isl_schedule_node_child(graft
, 0);
3273 if (tile
->n
< kernel
->n_block
)
3274 skip
= kernel
->n_block
- tile
->n
;
3277 filter
= set_schedule_modulo(graft
, kernel
->thread_ids
,
3279 if (!kernel
->options
->wrap
)
3280 graft
= snap_band_to_sizes(graft
, kernel
->block_dim
+ skip
,
3282 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0)
3283 graft
= isl_schedule_node_parent(graft
);
3284 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3286 while (graft
&& isl_schedule_node_has_parent(graft
))
3287 graft
= isl_schedule_node_parent(graft
);
3290 if (kernel_depth
< group
->depth
)
3291 node
= gpu_tree_ensure_sync_after_core(node
, kernel
);
3292 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3293 node
= isl_schedule_node_graft_before(node
, graft
);
3295 node
= gpu_tree_move_right_to_sync(node
, kernel
);
3296 node
= isl_schedule_node_graft_after(node
, graft
);
3297 if (kernel_depth
< group
->depth
)
3298 node
= add_group_write_sync(node
, kernel
, group
, 1);
3301 node
= gpu_tree_move_up_to_kernel(node
);
3306 /* Check whether the array reference group "group" is mapped to
3307 * private or shared memory and, if so,
3308 * add copy statements to the schedule tree of "node"
3309 * for reading from global memory to private or shared memory
3310 * (if "read" is set) or for writing back from private or shared memory
3311 * to global memory (if "read" is not set) for this group.
3312 * On input, "node" points to the kernel node, and it is moved
3313 * back there on output.
3315 static __isl_give isl_schedule_node
*add_copies_group(
3316 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3317 __isl_take isl_schedule_node
*node
, int read
)
3319 if (group
->private_tile
)
3320 return add_copies_group_private(kernel
, group
, node
, read
);
3321 if (group
->shared_tile
)
3322 return add_copies_group_shared(kernel
, group
, node
, read
);
3326 /* For each array reference group that is mapped to private or shared memory,
3327 * add copy statements to the schedule tree of "node"
3328 * for reading from global memory to private or shared memory
3329 * and for writing back.
3330 * On input, "node" points to the kernel node, and it is moved
3331 * back there on output.
3333 static __isl_give isl_schedule_node
*add_copies(struct ppcg_kernel
*kernel
,
3334 __isl_take isl_schedule_node
*node
)
3338 for (i
= 0; i
< kernel
->n_array
; ++i
) {
3339 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
3341 for (j
= 0; j
< array
->n_group
; ++j
) {
3342 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3344 node
= add_copies_group(kernel
, group
, node
, 1);
3347 node
= add_copies_group(kernel
, group
, node
, 0);
3356 /* Mark all dimensions in the current band node atomic.
3358 static __isl_give isl_schedule_node
*atomic(__isl_take isl_schedule_node
*node
)
3362 n
= isl_schedule_node_band_n_member(node
);
3363 for (i
= 0; i
< n
; ++i
)
3364 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
3365 isl_ast_loop_atomic
);
3370 /* Mark "node" atomic, if it is a band node.
3371 * Do the same for all ancestors.
3372 * Return a pointer to "node" (in the updated schedule tree).
3374 static __isl_give isl_schedule_node
*atomic_ancestors(
3375 __isl_take isl_schedule_node
*node
)
3381 if (!isl_schedule_node_has_parent(node
))
3384 pos
= isl_schedule_node_get_child_position(node
);
3385 node
= isl_schedule_node_parent(node
);
3386 if (isl_schedule_node_get_type(node
) == isl_schedule_node_band
)
3387 node
= atomic(node
);
3388 node
= atomic_ancestors(node
);
3389 node
= isl_schedule_node_child(node
, pos
);
3394 /* Collect all write references that require synchronization.
3395 * "node" is assumed to point to the kernel node.
3396 * Each reference is represented by a universe set in a space
3400 * with S[i,j] the statement instance space and R[] the array reference.
3402 * This function should be called before block and thread filters are added.
3404 * Synchronization is needed after a write if there is a subsequent read
3405 * within the same block that may not be performed by the same thread.
3406 * There should not be any dependences between different blocks,
3407 * so we start with the flow dependences within the same kernel invocation
3408 * and we subtract from these those dependences that are mapped
3409 * to the same iteration of the bands where synchronization is inserted.
3410 * We do not remove pairs of instances that are known to map to
3411 * the same thread across different iterations of the intermediate
3412 * bands because the read may be performed by a different thread
3413 * than the one that needs the value if shared memory is involved.
3415 * We also consider all pairs of possible writes that access the same
3416 * memory location and that may be mapped to the same block but not
3417 * to the same iteration of the intermediate bands.
3418 * In theory, it would be possible for one thread to still be in
3419 * a previous iteration of a loop in these bands.
3420 * A write to global memory in this delayed thread could then overwrite
3421 * a write from another thread that has already moved on to
3422 * the next iteration.
3424 * After computing the above writes paired off with reads or writes
3425 * that depend on them, we project onto the domain writes.
3426 * Sychronization is needed after writes to global memory
3427 * through these references.
3429 static __isl_give isl_union_set
*compute_sync_writes(
3430 struct ppcg_kernel
*kernel
, __isl_keep isl_schedule_node
*node
)
3432 isl_union_map
*local
;
3433 isl_union_map
*may_writes
, *shared_access
;
3434 isl_union_map
*kernel_prefix
, *thread_prefix
;
3435 isl_union_map
*equal
;
3436 isl_union_set
*wrap
;
3437 isl_union_set
*domain
;
3439 domain
= isl_schedule_node_get_universe_domain(node
);
3440 kernel_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3441 node
= isl_schedule_node_copy(node
);
3442 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3443 thread_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3444 isl_schedule_node_free(node
);
3446 may_writes
= isl_union_map_copy(kernel
->prog
->scop
->tagged_may_writes
);
3447 may_writes
= isl_union_map_curry(may_writes
);
3448 may_writes
= isl_union_map_intersect_domain(may_writes
, domain
);
3449 may_writes
= isl_union_map_uncurry(may_writes
);
3450 shared_access
= isl_union_map_copy(may_writes
);
3451 shared_access
= isl_union_map_apply_range(shared_access
,
3452 isl_union_map_reverse(may_writes
));
3454 local
= isl_union_map_copy(kernel
->prog
->scop
->tagged_dep_flow
);
3455 local
= isl_union_map_union(local
, shared_access
);
3456 local
= isl_union_map_zip(local
);
3458 equal
= isl_union_map_apply_range(kernel_prefix
,
3459 isl_union_map_reverse(isl_union_map_copy(kernel_prefix
)));
3460 wrap
= isl_union_map_wrap(equal
);
3461 local
= isl_union_map_intersect_domain(local
, wrap
);
3462 equal
= isl_union_map_apply_range(thread_prefix
,
3463 isl_union_map_reverse(isl_union_map_copy(thread_prefix
)));
3464 wrap
= isl_union_map_wrap(equal
);
3465 local
= isl_union_map_subtract_domain(local
, wrap
);
3467 local
= isl_union_map_zip(local
);
3468 local
= isl_union_map_universe(local
);
3470 return isl_union_map_domain(local
);
3473 /* Group the domain elements into a single space, named kernelX,
3474 * with X the kernel sequence number "kernel_id".
3476 static __isl_give isl_schedule_node
*group_statements(
3477 __isl_take isl_schedule_node
*node
, int kernel_id
)
3485 snprintf(buffer
, sizeof(buffer
), "kernel%d", kernel_id
);
3486 id
= isl_id_alloc(isl_schedule_node_get_ctx(node
), buffer
, NULL
);
3487 return isl_schedule_node_group(node
, id
);
3490 /* Create a ppcg_kernel representing the domain instances that reach "node"
3491 * and insert a mark node pointing to the ppcg_kernel before "node".
3492 * The band that "node" points to is the band that needs to be mapped
3493 * to block identifiers. The band that needs to be mapped to thread
3494 * identifiers should be marked by a "thread" mark by the caller.
3495 * This mark is removed by this function.
3496 * If "scale" is set, then the band that "node" points to is scaled
3499 * Mark all outer band nodes as atomic to ensure each kernel is only
3501 * If the domain elements that reach "node" live in more than one space,
3502 * then group the domain elements into a single space, named kernelX,
3503 * with X the kernel sequence number.
3505 * Insert a guard node governing the kernel node to ensure that
3506 * no kernels with zero blocks are launched.
3508 * Insert a context node describing the block and thread
3509 * identifiers inside the kernel mark.
3510 * The context node needs to be inserted after the effective block size
3511 * has been determined such that the bounds on the thread identifiers
3512 * would reflect the effective block size.
3513 * Insert a filter node inside the context node mapping the statement
3514 * instances to block identifiers. In particular, the block identifiers
3515 * are equated to the partial schedule of band that was marked for mapping
3516 * to blocks modulo the grid size.
3517 * Insert a filter node inside the "thread" mark mapping the statement
3518 * instances to thread identifiers. In particular, the thread identifiers
3519 * are equated to the partial schedule of band that was marked for mapping
3520 * to threads modulo the block size.
3522 * Compute array reference groups for all arrays, set the local
3523 * array bounds based on the set of domain instances that reach
3524 * the kernel node, check the total amount of shared memory used
3525 * and compute all group tilings.
3526 * The array reference groups are computed after the block filter
3527 * has been inserted because it affects the mapping to shared or
3528 * private memory. This computation also requires the thread filter
3529 * (in the ppcg_kernel object), but this thread filter should not
3530 * have been added to the schedule tree yet since the computation
3531 * requires the schedule of the band that needs to be mapped to
3532 * threads before the privatization is applied.
3534 * If any array reference group requires the band mapped to threads
3535 * to be unrolled, then we perform the required unrolling.
3537 * We save a copy of the schedule that may influence the mappings
3538 * to shared or private memory in kernel->shared_schedule.
3540 * Finally, we add synchronization and copy statements to the schedule tree,
3541 * remove the "thread" mark and create representations for the local
3542 * variables in the kernel.
3544 * We keep a copy of the isl_id that points to the kernel to ensure
3545 * that the kernel does not get destroyed if the schedule node
3546 * is freed due to some error condition.
3548 static __isl_give isl_schedule_node
*create_kernel(struct gpu_gen
*gen
,
3549 __isl_take isl_schedule_node
*node
, int scale
,
3550 __isl_keep isl_multi_val
*sizes
)
3552 struct ppcg_kernel
*kernel
;
3554 isl_schedule_node
*node_thread
;
3555 isl_union_map
*host_schedule
;
3556 isl_set
*host_domain
;
3557 isl_union_set
*domain
;
3558 int single_statement
;
3560 kernel
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel
);
3561 kernel
= ppcg_kernel_create_local_arrays(kernel
, gen
->prog
);
3563 return isl_schedule_node_free(node
);
3565 domain
= isl_schedule_node_get_domain(node
);
3566 single_statement
= isl_union_set_n_set(domain
) == 1;
3568 kernel
->ctx
= gen
->ctx
;
3569 kernel
->prog
= gen
->prog
;
3570 kernel
->options
= gen
->options
;
3571 kernel
->context
= extract_context(node
, gen
->prog
);
3572 kernel
->core
= isl_union_set_universe(isl_union_set_copy(domain
));
3573 kernel
->arrays
= accessed_by_domain(isl_union_set_copy(domain
),
3575 kernel
->n_grid
= n_outer_coincidence(node
);
3576 node_thread
= isl_schedule_node_copy(node
);
3577 node_thread
= gpu_tree_move_down_to_thread(node_thread
, kernel
->core
);
3578 node_thread
= isl_schedule_node_child(node_thread
, 0);
3579 kernel
->n_block
= n_outer_coincidence(node_thread
);
3580 isl_schedule_node_free(node_thread
);
3581 kernel
->id
= gen
->kernel_id
++;
3582 read_grid_and_block_sizes(kernel
, gen
);
3584 kernel
->sync_writes
= compute_sync_writes(kernel
, node
);
3586 host_schedule
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3587 host_domain
= isl_set_from_union_set(isl_union_map_range(
3590 node
= atomic_ancestors(node
);
3592 id
= isl_id_alloc(gen
->ctx
, "kernel", kernel
);
3593 id
= isl_id_set_free_user(id
, &ppcg_kernel_free_wrap
);
3594 node
= isl_schedule_node_insert_mark(node
, isl_id_copy(id
));
3596 if (!single_statement
)
3597 node
= group_statements(node
, kernel
->id
);
3599 node
= isl_schedule_node_child(node
, 0);
3600 node
= split_band(node
, kernel
->n_grid
);
3601 kernel
->block_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3602 kernel
->n_grid
, "b");
3603 kernel
->block_filter
= set_schedule_modulo(node
, kernel
->block_ids
,
3605 kernel
->grid_size
= extract_grid_size(kernel
,
3606 isl_union_set_copy(domain
));
3607 if (!kernel
->options
->wrap
)
3608 node
= snap_band_to_sizes(node
, kernel
->grid_dim
,
3611 node
= scale_band(node
, isl_multi_val_copy(sizes
));
3612 node
= isl_schedule_node_parent(node
);
3613 if (!single_statement
)
3614 node
= isl_schedule_node_parent(node
);
3615 node
= insert_guard(node
, kernel
->context
, kernel
->grid_size
,
3617 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3618 node
= isl_schedule_node_child(node
, 0);
3619 node
= split_band(node
, kernel
->n_block
);
3620 kernel
->thread_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3621 kernel
->n_block
, "t");
3622 kernel
->thread_filter
= set_schedule_modulo(node
, kernel
->thread_ids
,
3624 extract_block_size(kernel
, domain
);
3626 node
= gpu_tree_move_up_to_kernel(node
);
3627 node
= isl_schedule_node_child(node
, 0);
3628 node
= insert_context(kernel
, node
);
3629 node
= isl_schedule_node_child(node
, 0);
3630 node
= isl_schedule_node_insert_filter(node
,
3631 isl_union_set_copy(kernel
->block_filter
));
3633 node
= gpu_tree_move_up_to_kernel(node
);
3635 if (gpu_group_references(kernel
, node
) < 0)
3636 node
= isl_schedule_node_free(node
);
3637 localize_bounds(kernel
, host_domain
);
3638 isl_set_free(host_domain
);
3640 check_shared_memory_bound(kernel
);
3641 compute_group_tilings(kernel
);
3643 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3644 node
= isl_schedule_node_child(node
, 0);
3645 if (!kernel
->options
->wrap
)
3646 node
= snap_band_to_sizes(node
, kernel
->block_dim
,
3648 node
= isl_schedule_node_insert_filter(node
,
3649 isl_union_set_copy(kernel
->thread_filter
));
3650 if (kernel_requires_unroll(kernel
)) {
3651 node
= isl_schedule_node_child(node
, 0);
3652 node
= unroll(node
);
3655 node
= gpu_tree_move_up_to_thread(node
);
3656 kernel
->shared_schedule_dim
=
3657 isl_schedule_node_get_schedule_depth(node
);
3658 kernel
->shared_schedule
=
3659 isl_schedule_node_get_prefix_schedule_union_pw_multi_aff(node
);
3661 node
= gpu_tree_move_up_to_kernel(node
);
3663 node
= add_sync(kernel
, node
);
3664 node
= add_copies(kernel
, node
);
3666 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3667 node
= isl_schedule_node_delete(node
);
3669 node
= gpu_tree_move_up_to_kernel(node
);
3671 if (create_kernel_vars(kernel
) < 0)
3672 node
= isl_schedule_node_free(node
);
3674 if (!single_statement
)
3675 node
= isl_schedule_node_parent(node
);
3676 node
= isl_schedule_node_parent(node
);
3682 /* Insert a zero-dimensional permutable band at "node".
3684 static __isl_give isl_schedule_node
*insert_empty_permutable_band(
3685 __isl_take isl_schedule_node
*node
)
3688 isl_schedule
*schedule
;
3689 isl_union_set
*domain
;
3690 isl_multi_union_pw_aff
*mupa
;
3692 schedule
= isl_schedule_node_get_schedule(node
);
3693 domain
= isl_schedule_get_domain(schedule
);
3694 space
= isl_union_set_get_space(domain
);
3695 isl_union_set_free(domain
);
3696 isl_schedule_free(schedule
);
3698 space
= isl_space_set_from_params(space
);
3699 mupa
= isl_multi_union_pw_aff_zero(space
);
3700 node
= isl_schedule_node_insert_partial_schedule(node
, mupa
);
3701 node
= isl_schedule_node_band_set_permutable(node
, 1);
3706 /* If "node" is the outermost permutable band that can be mapped to block and
3707 * thread identifiers in its branch (or a leaf with no such outer bands),
3708 * then mark the band as such, attaching a ppcg_kernel to the mark.
3710 * If "node" originally points to a leaf, then insert a zero-dimensional
3711 * permutable band such that we can assume that "node" always
3712 * points to a band node.
3714 * Tile "node" using user specified tile sizes, after splitting the band
3715 * if the number of specified tile sizes is smaller than the dimension
3716 * of the band. Mark the point band of this tiling as the band that
3717 * needs to be mapped to threads.
3718 * Create a kernel representing the domain instances that reach "node" and
3719 * insert a mark node pointing to the ppcg_kernel before the band node.
3721 static __isl_give isl_schedule_node
*mark_outer_permutable(
3722 __isl_take isl_schedule_node
*node
, void *user
)
3724 struct gpu_gen
*gen
= user
;
3730 isl_multi_val
*sizes
;
3732 outer
= is_outer_tilable(node
);
3734 return isl_schedule_node_free(node
);
3738 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
3739 node
= insert_empty_permutable_band(node
);
3741 tile_len
= isl_schedule_node_band_n_member(node
);
3742 tile_size
= read_tile_sizes(gen
, &tile_len
);
3744 return isl_schedule_node_free(node
);
3745 if (tile_len
< isl_schedule_node_band_n_member(node
))
3746 node
= isl_schedule_node_band_split(node
, tile_len
);
3747 sizes
= construct_band_tiles_sizes(node
, tile_size
);
3748 node
= tile_band(node
, isl_multi_val_copy(sizes
));
3749 node
= isl_schedule_node_child(node
, 0);
3750 id
= isl_id_alloc(gen
->ctx
, "thread", NULL
);
3751 node
= isl_schedule_node_insert_mark(node
, id
);
3752 node
= isl_schedule_node_parent(node
);
3754 scale
= gen
->options
->scale_tile_loops
;
3755 node
= create_kernel(gen
, node
, scale
, sizes
);
3756 isl_multi_val_free(sizes
);
3762 /* Replace any reference to an array element in the range of "copy"
3763 * by a reference to all array elements (defined by the extent of the array).
3765 static __isl_give isl_union_map
*approximate_copy_out(
3766 __isl_take isl_union_map
*copy
, struct gpu_prog
*prog
)
3771 res
= isl_union_map_empty(isl_union_map_get_space(copy
));
3773 for (i
= 0; i
< prog
->n_array
; ++i
) {
3776 isl_union_map
*copy_i
;
3777 isl_union_set
*extent
, *domain
;
3779 space
= isl_space_copy(prog
->array
[i
].space
);
3780 extent
= isl_union_set_from_set(isl_set_universe(space
));
3781 copy_i
= isl_union_map_copy(copy
);
3782 copy_i
= isl_union_map_intersect_range(copy_i
, extent
);
3783 set
= isl_set_copy(prog
->array
[i
].extent
);
3784 extent
= isl_union_set_from_set(set
);
3785 domain
= isl_union_map_domain(copy_i
);
3786 copy_i
= isl_union_map_from_domain_and_range(domain
, extent
);
3787 res
= isl_union_map_union(res
, copy_i
);
3790 isl_union_map_free(copy
);
3795 /* Insert "kernel" marks that point to a ppcg_kernel structure
3796 * in front of all outermost tilable band that (by construction)
3797 * have at least one parallel loop.
3799 static __isl_give isl_schedule_node
*mark_kernels(struct gpu_gen
*gen
,
3800 __isl_take isl_schedule_node
*node
)
3802 return isl_schedule_node_map_descendant(node
,
3803 &mark_outer_permutable
, gen
);
3806 /* Save the schedule "schedule" to a file called "filename".
3807 * The schedule is printed in block style.
3809 static void save_schedule(__isl_keep isl_schedule
*schedule
,
3810 const char *filename
)
3819 file
= fopen(filename
, "w");
3821 fprintf(stderr
, "Unable to open '%s' for writing\n", filename
);
3824 ctx
= isl_schedule_get_ctx(schedule
);
3825 p
= isl_printer_to_file(ctx
, file
);
3826 p
= isl_printer_set_yaml_style(p
, ISL_YAML_STYLE_BLOCK
);
3827 p
= isl_printer_print_schedule(p
, schedule
);
3828 isl_printer_free(p
);
3832 /* Load and return a schedule from a file called "filename".
3834 static __isl_give isl_schedule
*load_schedule(isl_ctx
*ctx
,
3835 const char *filename
)
3838 isl_schedule
*schedule
;
3840 file
= fopen(filename
, "r");
3842 fprintf(stderr
, "Unable to open '%s' for reading\n", filename
);
3845 schedule
= isl_schedule_read_from_file(ctx
, file
);
3851 /* Compute an appropriate schedule based on the accesses in
3852 * gen->read and gen->write.
3854 * We use the dependences in gen->prog->scop to compute
3855 * a schedule that has a parallel loop in each tilable band and
3856 * return this schedule.
3858 * If live range reordering is allowed, then we need to make sure
3859 * that live ranges on arrays are not run in parallel since doing
3860 * so would require array expansion. We therefore add the array
3861 * order dependences to the coincidence dependences. Non-zero array
3862 * order dependences will then prevent a schedule dimension from being
3863 * considered parallel.
3864 * Live ranges derived from scalars are allowed to be run in parallel
3865 * since we force the scalars to be mapped to private memory in
3866 * check_scalar_live_ranges.
3867 * If live range reordering is allowed, then the false dependences
3868 * are not added to the validity constraints as that would prevent
3869 * reordering. Instead, the external false dependences that enforce that reads
3870 * from potentially live-in data precede any later write and
3871 * that writes of potentially live-out data follow any other earlier write
3872 * are added to the validity and the coincidence constraints.
3873 * The false dependences are still added to the proximity constraints
3874 * for consistency with the case where live range reordering is not allowed.
3875 * The coincidence constraints then consist of flow dependences,
3876 * external false dependences and array order dependences.
3877 * The independences can be filtered out from the first two sets.
3878 * They have already been filtered out from the array order dependences
3879 * on a per array basis in collect_order_dependences.
3880 * There is no need for a per array handling of the other two sets
3881 * as there should be no flow or external false dependence on local
3882 * variables that can be filtered out.
3884 static __isl_give isl_schedule
*compute_schedule(struct gpu_gen
*gen
)
3886 isl_union_set
*domain
;
3887 isl_union_map
*dep_raw
, *dep
;
3888 isl_union_map
*validity
, *proximity
, *coincidence
;
3889 isl_schedule_constraints
*sc
;
3890 isl_schedule
*schedule
;
3892 domain
= isl_union_set_copy(gen
->prog
->scop
->domain
);
3893 sc
= isl_schedule_constraints_on_domain(domain
);
3894 sc
= isl_schedule_constraints_set_context(sc
,
3895 isl_set_copy(gen
->prog
->scop
->context
));
3896 if (gen
->options
->live_range_reordering
) {
3897 sc
= isl_schedule_constraints_set_conditional_validity(sc
,
3898 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_flow
),
3899 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_order
));
3900 proximity
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
3901 validity
= isl_union_map_copy(proximity
);
3902 validity
= isl_union_map_union(validity
,
3903 isl_union_map_copy(gen
->prog
->scop
->dep_forced
));
3904 proximity
= isl_union_map_union(proximity
,
3905 isl_union_map_copy(gen
->prog
->scop
->dep_false
));
3906 coincidence
= isl_union_map_copy(validity
);
3907 coincidence
= isl_union_map_subtract(coincidence
,
3908 isl_union_map_copy(gen
->prog
->scop
->independence
));
3909 coincidence
= isl_union_map_union(coincidence
,
3910 isl_union_map_copy(gen
->prog
->array_order
));
3912 dep_raw
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
3913 dep
= isl_union_map_copy(gen
->prog
->scop
->dep_false
);
3914 dep
= isl_union_map_union(dep
, dep_raw
);
3915 dep
= isl_union_map_coalesce(dep
);
3916 proximity
= isl_union_map_copy(dep
);
3917 coincidence
= isl_union_map_copy(dep
);
3920 sc
= isl_schedule_constraints_set_validity(sc
, validity
);
3921 sc
= isl_schedule_constraints_set_coincidence(sc
, coincidence
);
3922 sc
= isl_schedule_constraints_set_proximity(sc
, proximity
);
3924 if (gen
->options
->debug
->dump_schedule_constraints
)
3925 isl_schedule_constraints_dump(sc
);
3926 schedule
= isl_schedule_constraints_compute_schedule(sc
);
3931 /* Obtain a schedule for the scop, either by reading it from
3932 * a file or by computing one.
3934 static __isl_give isl_schedule
*get_schedule(struct gpu_gen
*gen
)
3936 isl_schedule
*schedule
;
3938 if (gen
->options
->load_schedule_file
) {
3939 schedule
= load_schedule(gen
->ctx
,
3940 gen
->options
->load_schedule_file
);
3942 schedule
= compute_schedule(gen
);
3943 if (gen
->options
->save_schedule_file
)
3944 save_schedule(schedule
,
3945 gen
->options
->save_schedule_file
);
3947 if (gen
->options
->debug
->dump_schedule
)
3948 isl_schedule_dump(schedule
);
3953 /* Construct the string "<a>_<b>".
3955 static char *concat(isl_ctx
*ctx
, const char *a
, const char *b
)
3960 p
= isl_printer_to_str(ctx
);
3961 p
= isl_printer_print_str(p
, a
);
3962 p
= isl_printer_print_str(p
, "_");
3963 p
= isl_printer_print_str(p
, b
);
3964 s
= isl_printer_get_str(p
);
3965 isl_printer_free(p
);
3970 /* For each array in "prog" of which an element appears in "accessed" and
3971 * that is not a read only scalar, create a zero-dimensional universe set
3972 * of which the tuple id has name "<prefix>_<name of array>" and a user
3973 * pointer pointing to the array (gpu_array_info).
3975 * If the array is local to "prog", then make sure it will be declared
3978 * Return the list of these universe sets.
3980 static __isl_give isl_union_set_list
*create_copy_filters(struct gpu_prog
*prog
,
3981 const char *prefix
, __isl_take isl_union_set
*accessed
)
3985 isl_union_set_list
*filters
;
3988 filters
= isl_union_set_list_alloc(ctx
, 0);
3989 for (i
= 0; i
< prog
->n_array
; ++i
) {
3990 struct gpu_array_info
*array
= &prog
->array
[i
];
3992 isl_set
*accessed_i
;
3996 isl_union_set
*uset
;
3998 if (gpu_array_is_read_only_scalar(array
))
4001 space
= isl_space_copy(array
->space
);
4002 accessed_i
= isl_union_set_extract_set(accessed
, space
);
4003 empty
= isl_set_plain_is_empty(accessed_i
);
4004 isl_set_free(accessed_i
);
4006 filters
= isl_union_set_list_free(filters
);
4013 array
->declare_local
= 1;
4015 name
= concat(ctx
, prefix
, array
->name
);
4016 id
= name
? isl_id_alloc(ctx
, name
, array
) : NULL
;
4018 space
= isl_space_set_alloc(ctx
, 0, 0);
4019 space
= isl_space_set_tuple_id(space
, isl_dim_set
, id
);
4020 uset
= isl_union_set_from_set(isl_set_universe(space
));
4022 filters
= isl_union_set_list_add(filters
, uset
);
4024 isl_union_set_free(accessed
);
4029 /* Make sure that code for the statements in "filters" that
4030 * copy arrays to or from the device is only generated when
4031 * the size of the corresponding array is positive.
4032 * That is, add a set node underneath "graft" with "filters" as children
4033 * and for each child add a guard that the selects the parameter
4034 * values for which the corresponding array has a positive size.
4035 * The array is available in the user pointer of the statement identifier.
4036 * "depth" is the schedule depth of the position where "graft"
4039 static __isl_give isl_schedule_node
*insert_positive_size_guards(
4040 __isl_take isl_schedule_node
*graft
,
4041 __isl_take isl_union_set_list
*filters
, int depth
)
4045 graft
= isl_schedule_node_child(graft
, 0);
4046 graft
= isl_schedule_node_insert_set(graft
, filters
);
4047 n
= isl_schedule_node_n_children(graft
);
4048 for (i
= 0; i
< n
; ++i
) {
4049 isl_union_set
*filter
;
4050 isl_set
*domain
, *guard
;
4052 struct gpu_array_info
*array
;
4054 graft
= isl_schedule_node_child(graft
, i
);
4055 filter
= isl_schedule_node_filter_get_filter(graft
);
4056 domain
= isl_set_from_union_set(filter
);
4057 id
= isl_set_get_tuple_id(domain
);
4058 array
= isl_id_get_user(id
);
4060 isl_set_free(domain
);
4061 guard
= gpu_array_positive_size_guard(array
);
4062 guard
= isl_set_from_params(guard
);
4063 guard
= isl_set_add_dims(guard
, isl_dim_set
, depth
);
4064 graft
= isl_schedule_node_child(graft
, 0);
4065 graft
= isl_schedule_node_insert_guard(graft
, guard
);
4066 graft
= isl_schedule_node_parent(graft
);
4067 graft
= isl_schedule_node_parent(graft
);
4069 graft
= isl_schedule_node_parent(graft
);
4074 /* Create a graft for copying arrays to or from the device,
4075 * whenever the size of the array is strictly positive.
4076 * Each statement is called "<prefix>_<name of array>" and
4077 * the identifier has a user pointer pointing to the array.
4078 * The graft will be added at the position specified by "node".
4079 * "copy" contains the array elements that need to be copied.
4080 * Only arrays of which some elements need to be copied
4081 * will have a corresponding statement in the graph.
4082 * Note though that each such statement will copy the entire array.
4084 static __isl_give isl_schedule_node
*create_copy_device(struct gpu_prog
*prog
,
4085 __isl_keep isl_schedule_node
*node
, const char *prefix
,
4086 __isl_take isl_union_set
*copy
)
4091 isl_union_set
*all
, *domain
;
4092 isl_union_set_list
*filters
;
4093 isl_union_map
*extension
;
4094 isl_schedule_node
*graft
;
4097 depth
= isl_schedule_node_get_schedule_depth(node
);
4098 filters
= create_copy_filters(prog
, prefix
, copy
);
4099 all
= isl_union_set_list_union(isl_union_set_list_copy(filters
));
4101 space
= depth
< 0 ? NULL
: isl_space_set_alloc(ctx
, 0, depth
);
4102 domain
= isl_union_set_from_set(isl_set_universe(space
));
4103 extension
= isl_union_map_from_domain_and_range(domain
, all
);
4104 graft
= isl_schedule_node_from_extension(extension
);
4107 return isl_schedule_node_free(graft
);
4108 if (isl_union_set_list_n_union_set(filters
) == 0) {
4109 isl_union_set_list_free(filters
);
4113 return insert_positive_size_guards(graft
, filters
, depth
);
4116 /* Return (the universe spaces of) the arrays that are declared
4117 * inside the scop corresponding to "prog" and for which all
4118 * potential writes inside the scop form a subset of "domain".
4120 static __isl_give isl_union_set
*extract_local_accesses(struct gpu_prog
*prog
,
4121 __isl_keep isl_union_set
*domain
)
4124 isl_union_set
*local
;
4126 local
= isl_union_set_empty(isl_union_set_get_space(domain
));
4128 for (i
= 0; i
< prog
->n_array
; ++i
) {
4130 isl_union_map
*to_outer
;
4131 isl_union_map
*may_write
;
4132 isl_union_set
*write_domain
;
4133 isl_union_set
*fields
;
4136 if (!prog
->array
[i
].local
)
4139 set
= isl_set_universe(isl_space_copy(prog
->array
[i
].space
));
4140 to_outer
= isl_union_map_copy(prog
->to_outer
);
4141 to_outer
= isl_union_map_intersect_range(to_outer
,
4142 isl_union_set_from_set(isl_set_copy(set
)));
4143 fields
= isl_union_map_domain(to_outer
);
4144 may_write
= isl_union_map_copy(prog
->may_write
);
4145 may_write
= isl_union_map_intersect_range(may_write
, fields
);
4146 write_domain
= isl_union_map_domain(may_write
);
4147 subset
= isl_union_set_is_subset(write_domain
, domain
);
4148 isl_union_set_free(write_domain
);
4152 return isl_union_set_free(local
);
4153 } else if (subset
) {
4154 local
= isl_union_set_add_set(local
, set
);
4163 /* Internal data structure for node_may_persist.
4165 * "tagger" maps tagged iteration domains to the corresponding untagged
4168 * "may_persist_flow" is the set of all tagged dataflow dependences
4169 * with those dependences removed that either precede or follow
4170 * the kernel launch in a sequence.
4171 * "inner_band_flow" is the set of all tagged dataflow dependences
4172 * that are local to a given iteration of the outer band nodes
4173 * with respect to the current node.
4174 * "local_flow" is equal to "inner_band_flow", except that the domain
4175 * and the range have been intersected with intermediate filters
4176 * on children of sets or sequences.
4178 struct ppcg_may_persist_data
{
4179 isl_union_pw_multi_aff
*tagger
;
4181 isl_union_map
*local_flow
;
4182 isl_union_map
*inner_band_flow
;
4183 isl_union_map
*may_persist_flow
;
4186 /* Update the information in "data" based on the band ancestor "node".
4188 * In particular, we restrict the dependences in data->local_flow
4189 * to those dependence where the source and the sink occur in
4190 * the same iteration of the given band node.
4191 * We also update data->inner_band_flow to the new value of
4194 static int update_may_persist_at_band(__isl_keep isl_schedule_node
*node
,
4195 struct ppcg_may_persist_data
*data
)
4197 isl_multi_union_pw_aff
*partial
;
4198 isl_union_pw_multi_aff
*contraction
;
4199 isl_union_map
*flow
;
4201 if (isl_schedule_node_band_n_member(node
) == 0)
4204 partial
= isl_schedule_node_band_get_partial_schedule(node
);
4205 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4206 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4208 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4209 isl_union_pw_multi_aff_copy(data
->tagger
));
4211 flow
= data
->local_flow
;
4212 flow
= isl_union_map_eq_at_multi_union_pw_aff(flow
, partial
);
4213 data
->local_flow
= flow
;
4215 isl_union_map_free(data
->inner_band_flow
);
4216 data
->inner_band_flow
= isl_union_map_copy(data
->local_flow
);
4221 /* Given a set of local reaching domain elements "domain",
4222 * expand them to the corresponding leaf domain elements using "contraction"
4223 * and insert the array references tags using data->tagger.
4225 static __isl_give isl_union_set
*expand_and_tag(
4226 __isl_take isl_union_set
*domain
,
4227 __isl_take isl_union_pw_multi_aff
*contraction
,
4228 struct ppcg_may_persist_data
*data
)
4230 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4232 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4233 isl_union_pw_multi_aff_copy(data
->tagger
));
4237 /* Given a filter node that is the child of a set or sequence node,
4238 * restrict data->local_flow to refer only to those elements
4239 * in the filter of the node.
4240 * "contraction" maps the leaf domain elements of the schedule tree
4241 * to the corresponding domain elements at (the parent of) "node".
4243 static int filter_flow(__isl_keep isl_schedule_node
*node
,
4244 struct ppcg_may_persist_data
*data
,
4245 __isl_take isl_union_pw_multi_aff
*contraction
)
4247 isl_union_set
*filter
;
4248 isl_union_map
*flow
;
4250 flow
= data
->local_flow
;
4251 filter
= isl_schedule_node_filter_get_filter(node
);
4252 filter
= expand_and_tag(filter
, contraction
, data
);
4253 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(filter
));
4254 flow
= isl_union_map_intersect_range(flow
, filter
);
4255 data
->local_flow
= flow
;
4260 /* Given a filter node "node", collect the filters on all preceding siblings
4261 * (which are also filter nodes), add them to "filters" and return the result.
4263 static __isl_give isl_union_set
*add_previous_filters(
4264 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4266 isl_schedule_node
*sibling
;
4268 sibling
= isl_schedule_node_copy(node
);
4269 while (sibling
&& isl_schedule_node_has_previous_sibling(sibling
)) {
4270 isl_union_set
*filter
;
4272 sibling
= isl_schedule_node_previous_sibling(sibling
);
4273 filter
= isl_schedule_node_filter_get_filter(sibling
);
4274 filters
= isl_union_set_union(filters
, filter
);
4276 isl_schedule_node_free(sibling
);
4278 return isl_union_set_free(filters
);
4283 /* Given a filter node "node", collect the filters on all following siblings
4284 * (which are also filter nodes), add them to "filters" and return the result.
4286 static __isl_give isl_union_set
*add_next_filters(
4287 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4289 isl_schedule_node
*sibling
;
4291 sibling
= isl_schedule_node_copy(node
);
4292 while (sibling
&& isl_schedule_node_has_next_sibling(sibling
)) {
4293 isl_union_set
*filter
;
4295 sibling
= isl_schedule_node_next_sibling(sibling
);
4296 filter
= isl_schedule_node_filter_get_filter(sibling
);
4297 filters
= isl_union_set_union(filters
, filter
);
4299 isl_schedule_node_free(sibling
);
4301 return isl_union_set_free(filters
);
4306 /* Remove those flow dependences from data->may_persist_flow
4307 * that flow between elements of "domain" within the same iteration
4308 * of all outer band nodes.
4309 * "contraction" maps the leaf domain elements of the schedule tree
4310 * to the corresponding elements "domain".
4312 static void remove_external_flow(struct ppcg_may_persist_data
*data
,
4313 __isl_take isl_union_set
*domain
,
4314 __isl_keep isl_union_pw_multi_aff
*contraction
)
4316 isl_union_map
*flow
;
4318 contraction
= isl_union_pw_multi_aff_copy(contraction
);
4319 domain
= expand_and_tag(domain
, contraction
, data
);
4320 flow
= isl_union_map_copy(data
->local_flow
);
4321 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(domain
));
4322 flow
= isl_union_map_intersect_range(flow
, domain
);
4324 data
->may_persist_flow
= isl_union_map_subtract(data
->may_persist_flow
,
4328 /* Update the information in "data" based on the filter ancestor "node".
4329 * We only need to modify anything if the filter is the child
4330 * of a set or sequence node.
4332 * In the case of a sequence, we remove the dependences between
4333 * statement instances that are both executed either before or
4334 * after the subtree that will be mapped to a kernel, within
4335 * the same iteration of outer bands.
4337 * In both cases, we restrict data->local_flow to the current child.
4339 static int update_may_persist_at_filter(__isl_keep isl_schedule_node
*node
,
4340 struct ppcg_may_persist_data
*data
)
4342 enum isl_schedule_node_type type
;
4343 isl_schedule_node
*parent
;
4345 isl_union_pw_multi_aff
*contraction
;
4346 isl_union_set
*before
, *after
, *filter
;
4347 isl_union_map
*flow
;
4349 type
= isl_schedule_node_get_parent_type(node
);
4350 if (type
!= isl_schedule_node_sequence
&& type
!= isl_schedule_node_set
)
4353 parent
= isl_schedule_node_copy(node
);
4354 parent
= isl_schedule_node_parent(parent
);
4355 contraction
= isl_schedule_node_get_subtree_contraction(parent
);
4356 isl_schedule_node_free(parent
);
4358 if (type
== isl_schedule_node_set
)
4359 return filter_flow(node
, data
, contraction
);
4361 filter
= isl_schedule_node_filter_get_filter(node
);
4362 space
= isl_union_set_get_space(filter
);
4363 isl_union_set_free(filter
);
4364 before
= isl_union_set_empty(space
);
4365 after
= isl_union_set_copy(before
);
4366 before
= add_previous_filters(before
, node
);
4367 after
= add_next_filters(after
, node
);
4369 remove_external_flow(data
, before
, contraction
);
4370 remove_external_flow(data
, after
, contraction
);
4372 return filter_flow(node
, data
, contraction
);
4375 /* Update the information in "data" based on the ancestor "node".
4377 static int update_may_persist_at(__isl_keep isl_schedule_node
*node
, void *user
)
4379 struct ppcg_may_persist_data
*data
= user
;
4381 switch (isl_schedule_node_get_type(node
)) {
4382 case isl_schedule_node_error
:
4384 case isl_schedule_node_context
:
4385 case isl_schedule_node_domain
:
4386 case isl_schedule_node_expansion
:
4387 case isl_schedule_node_extension
:
4388 case isl_schedule_node_guard
:
4389 case isl_schedule_node_leaf
:
4390 case isl_schedule_node_mark
:
4391 case isl_schedule_node_sequence
:
4392 case isl_schedule_node_set
:
4394 case isl_schedule_node_band
:
4395 if (update_may_persist_at_band(node
, data
) < 0)
4398 case isl_schedule_node_filter
:
4399 if (update_may_persist_at_filter(node
, data
) < 0)
4407 /* Determine the set of array elements that may need to be perserved
4408 * by a kernel constructed from the subtree at "node".
4409 * This includes the set of array elements that may need to be preserved
4410 * by the entire scop (prog->may_persist) and the elements for which
4411 * there is a potential flow dependence that may cross a kernel launch.
4413 * To determine the second set, we start from all flow dependences.
4414 * From this set of dependences, we remove those that cannot possibly
4415 * require data to be preserved by a kernel launch.
4416 * In particular, we consider the following sets of dependences.
4417 * - dependences of which the write occurs inside the kernel.
4418 * If the data is needed outside the kernel, then it will
4419 * be copied out immediately after the kernel launch, so there
4420 * is no need for any special care.
4421 * - dependences of which the read occurs inside the kernel and the
4422 * corresponding write occurs inside the same iteration of the
4423 * outer band nodes. This means that the data is needed in
4424 * the first kernel launch after the write, which is already
4425 * taken care of by the standard copy-in. That is, the data
4426 * do not need to be preserved by any intermediate call to
4428 * - dependences of which the write and the read either both occur
4429 * before the kernel launch or both occur after the kernel launch,
4430 * within the same iteration of the outer band nodes with respect
4431 * to the sequence that determines the ordering of the dependence
4432 * and the kernel launch. Such flow dependences cannot cross
4433 * any kernel launch.
4435 * For the remaining (tagged) dependences, we take the domain
4436 * (i.e., the tagged writes) and apply the tagged access relation
4437 * to obtain the accessed data elements.
4438 * These are then combined with the elements that may need to be
4439 * preserved by the entire scop.
4441 static __isl_give isl_union_set
*node_may_persist(
4442 __isl_keep isl_schedule_node
*node
, struct gpu_prog
*prog
)
4444 struct ppcg_may_persist_data data
;
4445 isl_schedule_node
*root
;
4446 isl_union_pw_multi_aff
*contraction
;
4447 isl_union_set
*domain
;
4448 isl_union_set
*persist
;
4449 isl_union_map
*flow
, *local_flow
;
4451 data
.tagger
= prog
->scop
->tagger
;
4453 flow
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
4454 data
.local_flow
= isl_union_map_copy(flow
);
4455 data
.inner_band_flow
= isl_union_map_copy(flow
);
4456 data
.may_persist_flow
= flow
;
4457 if (isl_schedule_node_foreach_ancestor_top_down(node
,
4458 &update_may_persist_at
, &data
) < 0)
4459 data
.may_persist_flow
=
4460 isl_union_map_free(data
.may_persist_flow
);
4461 flow
= data
.may_persist_flow
;
4462 isl_union_map_free(data
.local_flow
);
4464 domain
= isl_schedule_node_get_domain(node
);
4465 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4466 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4468 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4469 isl_union_pw_multi_aff_copy(data
.tagger
));
4470 flow
= isl_union_map_subtract_domain(flow
, isl_union_set_copy(domain
));
4471 local_flow
= data
.inner_band_flow
;
4472 local_flow
= isl_union_map_intersect_range(local_flow
, domain
);
4473 flow
= isl_union_map_subtract(flow
, local_flow
);
4475 persist
= isl_union_map_domain(flow
);
4476 persist
= isl_union_set_apply(persist
,
4477 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4478 persist
= isl_union_set_union(persist
,
4479 isl_union_set_copy(prog
->may_persist
));
4484 /* Add nodes for copying outer arrays in and out of the device
4485 * before and after the subtree "node", which contains one or more kernels.
4486 * "domain" contains the original reaching domain elements before
4487 * the kernels were created, i.e., before the contraction that
4488 * may have been performed in creating the kernels has been applied.
4489 * "prefix" contains the prefix schedule at that point, in terms
4490 * of the same original reaching domain elements.
4492 * We first compute the sets of outer array elements that need
4493 * to be copied in and out and then graft in the nodes for
4494 * performing this copying.
4496 * In particular, for each array that is possibly written anywhere in
4497 * the subtree "node" and that may be used after "node"
4498 * or that may be visible outside the corresponding scop,
4499 * we copy out its entire extent.
4501 * Any array elements that is read without first being written inside
4502 * the subtree "node" needs to be copied in.
4503 * Furthermore, if there are any array elements that
4504 * are copied out, but that may not be written inside "node, then
4505 * they also need to be copied in to ensure that the value after execution
4506 * is the same as the value before execution, at least for those array
4507 * elements that may have their values preserved by the scop or that
4508 * may be written before "node" and read after "node".
4509 * In case the array elements are structures, we need to take into
4510 * account that all members of the structures need to be written
4511 * by "node" before we can avoid copying the data structure in.
4513 * Note that the may_write relation is intersected with the domain,
4514 * which has been intersected with the context.
4515 * This helps in those cases where the arrays are declared with a fixed size,
4516 * while the accesses are parametric and the context assigns a fixed value
4517 * to the parameters.
4519 * If an element from a local array is read without first being written,
4520 * then there is no point in copying it in since it cannot have been
4521 * written prior to the scop. Warn about the uninitialized read instead.
4523 static __isl_give isl_schedule_node
*add_to_from_device(
4524 __isl_take isl_schedule_node
*node
, __isl_take isl_union_set
*domain
,
4525 __isl_take isl_union_map
*prefix
, struct gpu_prog
*prog
)
4527 isl_union_set
*local
;
4528 isl_union_set
*to_device
, *from_device
, *may_persist
;
4529 isl_union_map
*may_write
, *must_write
, *copy_out
, *not_written
;
4530 isl_union_map
*read
, *copy_in
;
4531 isl_union_map
*tagged
;
4532 isl_union_map
*local_uninitialized
;
4533 isl_schedule_node
*graft
;
4535 tagged
= isl_union_map_copy(prog
->scop
->tagged_reads
);
4536 tagged
= isl_union_map_union(tagged
,
4537 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4539 may_write
= isl_union_map_copy(prog
->may_write
);
4540 may_write
= isl_union_map_intersect_domain(may_write
,
4541 isl_union_set_copy(domain
));
4542 may_write
= remove_local_accesses(prog
,
4543 isl_union_map_copy(tagged
), may_write
,
4544 isl_union_map_copy(prefix
), 0);
4545 may_write
= isl_union_map_apply_range(may_write
,
4546 isl_union_map_copy(prog
->to_outer
));
4547 may_write
= isl_union_map_apply_domain(may_write
,
4548 isl_union_map_copy(prefix
));
4549 may_write
= approximate_copy_out(may_write
, prog
);
4550 copy_out
= isl_union_map_copy(may_write
);
4551 may_write
= isl_union_map_apply_range(may_write
,
4552 isl_union_map_copy(prog
->to_inner
));
4553 must_write
= isl_union_map_copy(prog
->must_write
);
4554 must_write
= isl_union_map_apply_domain(must_write
,
4555 isl_union_map_copy(prefix
));
4556 may_persist
= node_may_persist(node
, prog
);
4557 may_write
= isl_union_map_intersect_range(may_write
, may_persist
);
4558 not_written
= isl_union_map_subtract(may_write
, must_write
);
4560 local
= extract_local_accesses(prog
, domain
);
4561 read
= isl_union_map_copy(prog
->read
);
4562 read
= isl_union_map_intersect_domain(read
, domain
);
4563 read
= remove_local_accesses(prog
, tagged
, read
,
4564 isl_union_map_copy(prefix
), 1);
4565 local
= isl_union_set_apply(local
, isl_union_map_copy(prog
->to_inner
));
4566 local_uninitialized
= isl_union_map_copy(prog
->scop
->live_in
);
4567 local_uninitialized
= isl_union_map_intersect_range(local_uninitialized
,
4569 local_uninitialized
= isl_union_map_intersect(local_uninitialized
,
4570 isl_union_map_copy(read
));
4571 if (!isl_union_map_is_empty(local_uninitialized
)) {
4573 "possibly uninitialized reads (not copied in):\n");
4574 isl_union_map_dump(local_uninitialized
);
4576 read
= isl_union_map_subtract(read
, local_uninitialized
);
4577 read
= isl_union_map_apply_domain(read
, prefix
);
4578 copy_in
= isl_union_map_union(read
, not_written
);
4579 copy_in
= isl_union_map_apply_range(copy_in
,
4580 isl_union_map_copy(prog
->to_outer
));
4582 graft
= create_copy_device(prog
, node
, "to_device",
4583 isl_union_map_range(copy_in
));
4584 node
= isl_schedule_node_graft_before(node
, graft
);
4585 graft
= create_copy_device(prog
, node
, "from_device",
4586 isl_union_map_range(copy_out
));
4587 node
= isl_schedule_node_graft_after(node
, graft
);
4592 /* Update "schedule" for mapping to a GPU device.
4594 * In particular, insert a context node, create kernels for
4595 * each outermost tilable band and introduce node for copying array
4596 * in and out of the device.
4598 static __isl_give isl_schedule
*map_to_device(struct gpu_gen
*gen
,
4599 __isl_take isl_schedule
*schedule
)
4601 isl_schedule_node
*node
;
4603 isl_union_set
*domain
;
4604 isl_union_map
*prefix
;
4606 context
= isl_set_copy(gen
->prog
->context
);
4607 context
= isl_set_from_params(context
);
4608 schedule
= isl_schedule_insert_context(schedule
, context
);
4610 node
= isl_schedule_get_root(schedule
);
4611 isl_schedule_free(schedule
);
4612 node
= isl_schedule_node_child(node
, 0);
4613 if (isl_schedule_node_get_type(node
) == isl_schedule_node_context
)
4614 node
= isl_schedule_node_child(node
, 0);
4615 domain
= isl_schedule_node_get_domain(node
);
4616 prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
4617 node
= mark_kernels(gen
, node
);
4618 node
= add_to_from_device(node
, domain
, prefix
, gen
->prog
);
4619 schedule
= isl_schedule_node_get_schedule(node
);
4620 isl_schedule_node_free(node
);
4625 /* Internal data structure for extract_access.
4626 * "next_access" points to the end of a linked list that is extended
4627 * by extract_access.
4628 * "single_expression" is set if the access expressions belong to
4629 * an expression statement (i.e., a statement without internal control).
4630 * "any_to_outer" maps all intermediate arrays to their outer arrays.
4632 struct ppcg_extract_access_data
{
4633 struct gpu_stmt_access
**next_access
;
4634 int single_expression
;
4635 isl_union_map
*any_to_outer
;
4638 /* Given a tagged access relation to a single array "tagged", extract it
4639 * as a map, taking into account that the input may be empty.
4640 * If the access relation is empty, then it does not contain
4641 * any space information, so we try to recover it from the index
4643 * The space of the index expression is of the form I -> A,
4644 * with I the statement instances and A the array, or [I -> F] -> A,
4645 * with F the filters corresponding to arguments.
4646 * We first drop F, if present, obtaining I -> A.
4647 * Then we construct I -> R, with R the reference tag,
4648 * combine the two into I -> [R -> A] and uncurry to obtain
4649 * the final result [I -> R] -> A.
4650 * Note that the index expression may have a lower dimension
4651 * than that of the array, but this dimension is not used
4652 * if the access relation is empty.
4654 static __isl_give isl_map
*extract_single_tagged_access(
4655 __isl_take isl_union_map
*tagged
, __isl_keep pet_expr
*expr
)
4659 isl_space
*space
, *space2
;
4660 isl_multi_pw_aff
*index
;
4662 empty
= isl_union_map_is_empty(tagged
);
4666 return isl_map_from_union_map(tagged
);
4667 isl_union_map_free(tagged
);
4669 index
= pet_expr_access_get_index(expr
);
4670 space
= isl_multi_pw_aff_get_space(index
);
4671 isl_multi_pw_aff_free(index
);
4672 if (isl_space_domain_is_wrapping(space
))
4673 space
= isl_space_domain_factor_domain(space
);
4674 space2
= isl_space_copy(space
);
4675 space2
= isl_space_from_domain(isl_space_domain(space
));
4676 id
= pet_expr_access_get_ref_id(expr
);
4677 space2
= isl_space_set_tuple_id(space2
, isl_dim_out
, id
);
4678 space
= isl_space_range_product(space2
, space
);
4679 space
= isl_space_uncurry(space
);
4681 return isl_map_empty(space
);
4683 isl_union_map_free(tagged
);
4687 /* Extract a gpu_stmt_access from "expr", append it to the list
4688 * that ends in *data->next_access and update the end of the list.
4689 * If the access expression performs a write, then it is considered
4690 * exact only if it appears in a single expression statement and
4691 * if its may access relation is equal to its must access relation.
4693 * The combined set of may accesses may be union if member accesses
4694 * are involved, but the entire set is derived from a single reference and
4695 * therefore from a single index expression. These accesses therefore
4696 * all map to the same outer array.
4698 static int extract_access(__isl_keep pet_expr
*expr
, void *user
)
4700 struct ppcg_extract_access_data
*data
= user
;
4701 isl_union_map
*tagged
;
4702 struct gpu_stmt_access
*access
;
4703 isl_ctx
*ctx
= pet_expr_get_ctx(expr
);
4704 isl_multi_pw_aff
*index
;
4706 access
= isl_alloc_type(ctx
, struct gpu_stmt_access
);
4708 access
->next
= NULL
;
4709 access
->read
= pet_expr_access_is_read(expr
);
4710 access
->write
= pet_expr_access_is_write(expr
);
4711 tagged
= pet_expr_access_get_tagged_may_read(expr
);
4712 tagged
= isl_union_map_union(tagged
,
4713 pet_expr_access_get_tagged_may_write(expr
));
4714 tagged
= isl_union_map_apply_range(tagged
,
4715 isl_union_map_copy(data
->any_to_outer
));
4716 if (!access
->write
) {
4717 access
->exact_write
= 1;
4718 } else if (!data
->single_expression
) {
4719 access
->exact_write
= 0;
4721 isl_union_map
*must
, *may
;
4722 may
= isl_union_map_copy(tagged
);
4723 may
= isl_union_map_domain_factor_domain(may
);
4724 must
= pet_expr_access_get_must_write(expr
);
4725 access
->exact_write
= isl_union_map_is_equal(must
, may
);
4726 isl_union_map_free(must
);
4727 isl_union_map_free(may
);
4729 index
= pet_expr_access_get_index(expr
);
4730 access
->n_index
= isl_multi_pw_aff_dim(index
, isl_dim_out
);
4731 isl_multi_pw_aff_free(index
);
4732 access
->ref_id
= pet_expr_access_get_ref_id(expr
);
4733 access
->tagged_access
= extract_single_tagged_access(tagged
, expr
);
4734 access
->access
= isl_map_copy(access
->tagged_access
);
4735 access
->access
= isl_map_domain_factor_domain(access
->access
);
4737 *data
->next_access
= access
;
4738 data
->next_access
= &(*data
->next_access
)->next
;
4740 if (!access
->access
)
4746 /* Construct a linked list of gpu_stmt_access objects,
4747 * one for each access expression in the statement body.
4748 * "any_to_outer" maps all intermediate arrays to their outer arrays.
4750 static int pet_stmt_extract_accesses(struct gpu_stmt
*stmt
,
4751 __isl_keep isl_union_map
*any_to_outer
)
4753 struct ppcg_extract_access_data data
;
4755 stmt
->accesses
= NULL
;
4756 data
.next_access
= &stmt
->accesses
;
4757 data
.single_expression
=
4758 pet_tree_get_type(stmt
->stmt
->body
) == pet_tree_expr
;
4759 data
.any_to_outer
= any_to_outer
;
4760 return pet_tree_foreach_access_expr(stmt
->stmt
->body
,
4761 &extract_access
, &data
);
4764 /* Return an array of gpu_stmt representing the statements in "scop".
4766 static struct gpu_stmt
*extract_stmts(isl_ctx
*ctx
, struct ppcg_scop
*scop
,
4767 __isl_keep isl_set
*context
, __isl_keep isl_union_map
*any_to_outer
)
4770 struct gpu_stmt
*stmts
;
4772 stmts
= isl_calloc_array(ctx
, struct gpu_stmt
, scop
->pet
->n_stmt
);
4776 for (i
= 0; i
< scop
->pet
->n_stmt
; ++i
) {
4777 struct gpu_stmt
*s
= &stmts
[i
];
4779 s
->id
= isl_set_get_tuple_id(scop
->pet
->stmts
[i
]->domain
);
4780 s
->stmt
= scop
->pet
->stmts
[i
];
4781 if (pet_stmt_extract_accesses(s
, any_to_outer
) < 0)
4782 return free_stmts(stmts
, i
+ 1);
4788 /* Callback for ppcg_print_guarded that calls the callback for generate_gpu.
4790 static __isl_give isl_printer
*print_gpu(__isl_take isl_printer
*p
, void *user
)
4792 struct gpu_gen
*gen
= user
;
4794 return gen
->print(p
, gen
->prog
, gen
->tree
, &gen
->types
,
4798 /* Generate CUDA code for "scop" and print it to "p".
4799 * After generating an AST for the transformed scop as explained below,
4800 * we call "gen->print" to print the AST in the desired output format
4803 * If it turns out that it does not make sense to generate GPU code,
4804 * then we generate CPU code instead.
4806 * The GPU code is generated in a context where at least one
4807 * statement instance is executed. The corresponding guard (if any) is printed
4808 * around the entire generated GPU code, except for the declaration
4809 * of the arrays that are visible outside of the scop and that therefore
4810 * cannot be declared inside the body of any possible guard.
4812 * We first compute a schedule that respects the dependences
4813 * of the original program and select the outermost bands
4814 * of tilable dimensions that have at least one parallel loop.
4815 * If the --load-schedule is specified, then the loaded schedule
4816 * is used instead of a computed schedule.
4818 * Each of these bands B is then tiled according to "tile" sizes, resulting
4819 * in two nested bands, with a kernel marker on top
4827 * We then split off at most 2 parallel dimensions from the T band and
4828 * at most 3 parallel dimension from the P band
4841 * A filter is introduced in front of T1 that maps the domain instances
4842 * to block identifiers. Similarly, a filter is introduced in front of P1
4843 * that maps the domain instances to thread identifiers.
4845 * For each iteration of the T2 band and for each array, we compute
4846 * the array elements accessed by that iteration, construct a rectangular
4847 * box around it and shift it to the origin. The result is used
4848 * as shared memory for the array.
4850 * Copying and synchronization statements are added to this schedule tree.
4851 * In principle, these are added in front of the P1 band, but some of
4852 * them may get hoisted up to higher levels.
4854 * The entire AST is then generated from the single resulting schedule tree.
4855 * During the generation the subtrees at kernel nodes (K) are saved
4856 * aside and replaced by kernel calls. The result is printed as host code
4857 * while the saved subtrees are printed as device code.
4859 static __isl_give isl_printer
*generate(__isl_take isl_printer
*p
,
4860 struct gpu_gen
*gen
, struct ppcg_scop
*scop
,
4861 struct ppcg_options
*options
)
4863 struct gpu_prog
*prog
;
4865 isl_set
*context
, *guard
;
4866 isl_schedule
*schedule
;
4870 return isl_printer_free(p
);
4872 ctx
= isl_printer_get_ctx(p
);
4873 prog
= gpu_prog_alloc(ctx
, scop
);
4875 return isl_printer_free(p
);
4877 context
= isl_set_copy(prog
->context
);
4878 guard
= isl_union_set_params(isl_union_set_copy(prog
->scop
->domain
));
4879 prog
->context
= isl_set_intersect(prog
->context
, isl_set_copy(guard
));
4882 schedule
= get_schedule(gen
);
4884 any_permutable
= has_any_permutable_node(schedule
);
4885 if (any_permutable
< 0 || !any_permutable
) {
4886 isl_set_free(context
);
4887 isl_set_free(guard
);
4888 if (any_permutable
< 0)
4889 p
= isl_printer_free(p
);
4891 p
= print_cpu(p
, scop
, options
);
4892 isl_schedule_free(schedule
);
4894 schedule
= map_to_device(gen
, schedule
);
4895 gen
->tree
= generate_code(gen
, schedule
);
4896 p
= ppcg_print_exposed_declarations(p
, prog
->scop
);
4897 p
= ppcg_print_guarded(p
, guard
, context
, &print_gpu
, gen
);
4898 isl_ast_node_free(gen
->tree
);
4901 gpu_prog_free(prog
);
4906 /* Wrapper around generate for use as a ppcg_transform callback.
4908 static __isl_give isl_printer
*generate_wrap(__isl_take isl_printer
*p
,
4909 struct ppcg_scop
*scop
, void *user
)
4911 struct gpu_gen
*gen
= user
;
4913 return generate(p
, gen
, scop
, gen
->options
);
4916 /* Transform the code in the file called "input" by replacing
4917 * all scops by corresponding GPU code and write the results to "out".
4919 int generate_gpu(isl_ctx
*ctx
, const char *input
, FILE *out
,
4920 struct ppcg_options
*options
,
4921 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
4922 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
4923 struct gpu_types
*types
, void *user
), void *user
)
4930 gen
.sizes
= extract_sizes_from_str(ctx
, options
->sizes
);
4931 gen
.options
= options
;
4934 gen
.print_user
= user
;
4936 gen
.types
.name
= NULL
;
4938 if (options
->debug
->dump_sizes
) {
4939 isl_space
*space
= isl_space_params_alloc(ctx
, 0);
4940 gen
.used_sizes
= isl_union_map_empty(space
);
4943 r
= ppcg_transform(ctx
, input
, out
, options
, &generate_wrap
, &gen
);
4945 if (options
->debug
->dump_sizes
) {
4946 isl_union_map_dump(gen
.used_sizes
);
4947 isl_union_map_free(gen
.used_sizes
);
4950 isl_union_map_free(gen
.sizes
);
4951 for (i
= 0; i
< gen
.types
.n
; ++i
)
4952 free(gen
.types
.name
[i
]);
4953 free(gen
.types
.name
);
4958 /* Compute the set of inner array elements that may have their values
4959 * preserved by "prog". In particular, collect the array elements of
4960 * arrays that are not local to "prog" and remove those elements that
4961 * are definitely killed or definitely written by "prog".
4963 static __isl_give isl_union_set
*compute_may_persist(struct gpu_prog
*prog
)
4966 isl_union_set
*may_persist
, *killed
;
4967 isl_union_map
*must_kill
;
4969 may_persist
= isl_union_set_empty(isl_set_get_space(prog
->context
));
4970 for (i
= 0; i
< prog
->n_array
; ++i
) {
4973 if (prog
->array
[i
].local
)
4976 extent
= isl_set_copy(prog
->array
[i
].extent
);
4977 may_persist
= isl_union_set_add_set(may_persist
, extent
);
4980 may_persist
= isl_union_set_intersect_params(may_persist
,
4981 isl_set_copy(prog
->context
));
4982 may_persist
= isl_union_set_apply(may_persist
,
4983 isl_union_map_copy(prog
->to_inner
));
4984 must_kill
= isl_union_map_copy(prog
->tagged_must_kill
);
4985 killed
= isl_union_map_range(must_kill
);
4986 must_kill
= isl_union_map_copy(prog
->must_write
);
4987 killed
= isl_union_set_union(killed
, isl_union_map_range(must_kill
));
4989 may_persist
= isl_union_set_subtract(may_persist
, killed
);
4993 struct gpu_prog
*gpu_prog_alloc(isl_ctx
*ctx
, struct ppcg_scop
*scop
)
4995 struct gpu_prog
*prog
;
5002 prog
= isl_calloc_type(ctx
, struct gpu_prog
);
5007 prog
->context
= isl_set_copy(scop
->context
);
5008 prog
->n_stmts
= scop
->pet
->n_stmt
;
5009 prog
->any_to_outer
= pet_scop_compute_outer_to_any(scop
->pet
);
5010 prog
->any_to_outer
= isl_union_map_reverse(prog
->any_to_outer
);
5011 space
= isl_union_map_get_space(prog
->any_to_outer
);
5012 space
= isl_space_set_from_params(space
);
5013 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
5014 space
= isl_space_map_from_set(space
);
5015 id
= isl_map_identity(space
);
5016 prog
->any_to_outer
= isl_union_map_add_map(prog
->any_to_outer
, id
);
5017 prog
->stmts
= extract_stmts(ctx
, scop
,
5018 prog
->context
, prog
->any_to_outer
);
5019 prog
->read
= isl_union_map_copy(scop
->reads
);
5020 prog
->may_write
= isl_union_map_copy(scop
->may_writes
);
5021 prog
->must_write
= isl_union_map_copy(scop
->must_writes
);
5022 prog
->tagged_must_kill
= isl_union_map_copy(scop
->tagged_must_kills
);
5023 prog
->to_inner
= pet_scop_compute_outer_to_inner(scop
->pet
);
5024 prog
->to_outer
= isl_union_map_copy(prog
->to_inner
);
5025 prog
->to_outer
= isl_union_map_reverse(prog
->to_outer
);
5028 return gpu_prog_free(prog
);
5030 if (collect_array_info(prog
) < 0)
5031 return gpu_prog_free(prog
);
5032 prog
->may_persist
= compute_may_persist(prog
);
5037 void *gpu_prog_free(struct gpu_prog
*prog
)
5041 free_array_info(prog
);
5042 free_stmts(prog
->stmts
, prog
->n_stmts
);
5043 isl_union_map_free(prog
->any_to_outer
);
5044 isl_union_map_free(prog
->to_outer
);
5045 isl_union_map_free(prog
->to_inner
);
5046 isl_union_map_free(prog
->read
);
5047 isl_union_map_free(prog
->may_write
);
5048 isl_union_map_free(prog
->must_write
);
5049 isl_union_map_free(prog
->tagged_must_kill
);
5050 isl_union_map_free(prog
->array_order
);
5051 isl_union_set_free(prog
->may_persist
);
5052 isl_set_free(prog
->context
);