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 and
1385 * may be NULL if we are not inside a kernel.
1386 * "accesses" is the list of gpu_stmt_access in the statement.
1387 * "iterator_map" expresses the statement iterators in terms of
1388 * the AST loop iterators.
1389 * "sched2shared" expresses the outer shared_schedule_dim dimensions of
1390 * the kernel schedule in terms of the AST loop iterators and
1391 * may be NULL if we are not inside a kernel.
1393 * The following fields are set in transform_index and used in transform_expr.
1394 * "array" is the array that is being accessed.
1395 * "global" is set if the global array is accessed (rather than
1396 * shared/private memory).
1397 * "local_array" refers to information on the array specialized
1398 * to the current kernel.
1400 struct ppcg_transform_data
{
1401 struct ppcg_kernel
*kernel
;
1402 struct gpu_stmt_access
*accesses
;
1403 isl_pw_multi_aff
*iterator_map
;
1404 isl_pw_multi_aff
*sched2shared
;
1406 struct gpu_array_info
*array
;
1408 struct gpu_local_array_info
*local_array
;
1411 /* Return the name of the outer array (of structs) accessed by "access".
1413 static const char *get_outer_array_name(__isl_keep isl_map
*access
)
1418 space
= isl_space_range(isl_map_get_space(access
));
1419 while (space
&& isl_space_is_wrapping(space
))
1420 space
= isl_space_domain(isl_space_unwrap(space
));
1421 name
= isl_space_get_tuple_name(space
, isl_dim_set
);
1422 isl_space_free(space
);
1427 /* Return a pointer to the gpu_array_ref_group in "local"
1428 * that contains the reference "access".
1429 * Return NULL if no such group can be found.
1431 static struct gpu_array_ref_group
*find_ref_group(
1432 struct gpu_local_array_info
*local
, struct gpu_stmt_access
*access
)
1436 for (i
= 0; i
< local
->n_group
; ++i
) {
1437 struct gpu_array_ref_group
*group
= local
->groups
[i
];
1439 for (j
= 0; j
< group
->n_ref
; ++j
)
1440 if (group
->refs
[j
] == access
)
1447 /* Index transformation callback for pet_stmt_build_ast_exprs.
1449 * "index" expresses the array indices in terms of statement iterators
1451 * We first reformulate "index" in terms of the AST loop iterators.
1452 * Then we check if we are accessing the global array or
1453 * a shared/private copy. In particular, if we are not inside a kernel
1454 * then we must be accessing a global array.
1455 * In the former case, we simply return
1456 * the updated index. If "index" is an affine expression rather
1457 * than an array access, then we also return the updated index here.
1459 * If no reference groups have been computed for the array,
1460 * then we can only be accessing the global array.
1462 * Otherwise, we apply the tiling to the index.
1463 * This tiling is of the form
1467 * where D corresponds to the outer group->depth dimensions of
1468 * the kernel schedule.
1469 * The index is of the form
1473 * We update the tiling to refer to the AST loop iterators
1477 * and modify index to keep track of those iterators
1481 * Combining these two yields a tiled index expression in terms
1482 * of the AST loop iterators
1486 static __isl_give isl_multi_pw_aff
*transform_index(
1487 __isl_take isl_multi_pw_aff
*index
, __isl_keep isl_id
*ref_id
,
1490 struct ppcg_transform_data
*data
= user
;
1491 struct gpu_stmt_access
*access
;
1492 struct gpu_array_ref_group
*group
;
1493 struct gpu_array_tile
*tile
;
1494 isl_pw_multi_aff
*iterator_map
;
1499 isl_multi_pw_aff
*tiling
;
1500 isl_pw_multi_aff
*pma
;
1501 isl_multi_pw_aff
*mpa
;
1502 isl_pw_multi_aff
*sched2depth
;
1506 iterator_map
= isl_pw_multi_aff_copy(data
->iterator_map
);
1507 index
= isl_multi_pw_aff_pullback_pw_multi_aff(index
, iterator_map
);
1512 access
= find_access(data
->accesses
, ref_id
);
1515 if (!isl_map_has_tuple_name(access
->access
, isl_dim_out
))
1518 name
= get_outer_array_name(access
->access
);
1519 i
= find_array_index(data
->kernel
, name
);
1521 isl_die(isl_multi_pw_aff_get_ctx(index
), isl_error_internal
,
1522 "cannot find array",
1523 return isl_multi_pw_aff_free(index
));
1524 data
->local_array
= &data
->kernel
->array
[i
];
1525 data
->array
= data
->local_array
->array
;
1527 group
= find_ref_group(data
->local_array
, access
);
1533 tile
= group
->private_tile
;
1535 tile
= group
->shared_tile
;
1536 data
->global
= !tile
;
1540 space
= isl_space_range(isl_multi_pw_aff_get_space(index
));
1541 space
= isl_space_map_from_set(space
);
1542 pma
= isl_pw_multi_aff_identity(space
);
1543 sched2depth
= isl_pw_multi_aff_copy(data
->sched2shared
);
1544 dim
= isl_pw_multi_aff_dim(sched2depth
, isl_dim_out
);
1545 sched2depth
= isl_pw_multi_aff_drop_dims(sched2depth
, isl_dim_out
,
1546 group
->depth
, dim
- group
->depth
);
1547 pma
= isl_pw_multi_aff_product(sched2depth
, pma
);
1548 tiling
= isl_multi_pw_aff_from_multi_aff(
1549 isl_multi_aff_copy(tile
->tiling
));
1550 tiling
= isl_multi_pw_aff_pullback_pw_multi_aff(tiling
, pma
);
1552 space
= isl_space_domain(isl_multi_pw_aff_get_space(index
));
1553 space
= isl_space_map_from_set(space
);
1554 mpa
= isl_multi_pw_aff_identity(space
);
1555 index
= isl_multi_pw_aff_range_product(mpa
, index
);
1556 index
= isl_multi_pw_aff_pullback_multi_pw_aff(tiling
, index
);
1561 /* Dereference "expr" by adding an index [0].
1562 * The original "expr" is assumed not to have any indices.
1564 * If "expr" is a member access, then the dereferencing needs
1565 * to be applied to the structure argument of this member access.
1567 static __isl_give isl_ast_expr
*dereference(__isl_take isl_ast_expr
*expr
)
1570 isl_ast_expr
*arg0
, *res
;
1571 isl_ast_expr_list
*list
;
1573 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1575 return isl_ast_expr_free(expr
);
1576 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1577 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1580 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1581 arg
= dereference(arg
);
1582 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1583 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1587 isl_ast_expr_free(arg0
);
1589 ctx
= isl_ast_expr_get_ctx(expr
);
1590 res
= isl_ast_expr_from_val(isl_val_zero(ctx
));
1591 list
= isl_ast_expr_list_from_ast_expr(res
);
1592 res
= isl_ast_expr_get_op_arg(expr
, 0);
1593 res
= isl_ast_expr_access(res
, list
);
1594 isl_ast_expr_free(expr
);
1599 /* Linearize the index expression "expr" based on the array bounds
1602 * That is, transform expression
1604 * A[i_0][i_1]...[i_n]
1608 * A[(..((i_0 * b_1 + i_1) ... ) * b_n + i_n]
1610 * where b_0, b_1, ..., b_n are the bounds on the array.
1612 * If the base of "expr" is a member access, then the linearization needs
1613 * to be applied to the structure argument of this member access.
1615 * In the base case, if "expr" has no arguments (other than the name of
1616 * the array), then we are passing an entire array to a function.
1617 * In this case, there is nothing to linearize.
1618 * Note that at this point an expression with no arguments can
1619 * only be an entire array because the scalar case and
1620 * the case of single struct are handled by the caller.
1622 * If the number of specified index expressions in "expr"
1623 * is smaller than the dimension of the accessed array,
1624 * then the missing i_j also do not appear in the linearized expression.
1625 * Furthermore, since such an expression does not refer to a single
1626 * element while the default linearized expression would refer to
1627 * a single element, we return the expression
1629 * A + (..((i_0 * b_1 + i_1) ... ) * b_n]
1631 * instead. Note that because of the special case handling above,
1632 * we can assume here that here that there is at least one index expression.
1634 __isl_give isl_ast_expr
*gpu_local_array_info_linearize_index(
1635 struct gpu_local_array_info
*array
, __isl_take isl_ast_expr
*expr
)
1642 isl_ast_expr_list
*list
;
1643 isl_ast_build
*build
;
1645 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1646 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1647 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1650 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1651 arg
= gpu_local_array_info_linearize_index(array
, arg
);
1652 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1653 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1657 isl_ast_expr_free(arg0
);
1659 if (isl_ast_expr_get_op_n_arg(expr
) == 1)
1662 ctx
= isl_ast_expr_get_ctx(expr
);
1663 context
= isl_set_universe(isl_space_params_alloc(ctx
, 0));
1664 build
= isl_ast_build_from_context(context
);
1666 n
= isl_ast_expr_get_op_n_arg(expr
);
1667 res
= isl_ast_expr_get_op_arg(expr
, 1);
1668 for (i
= 1; i
< array
->n_index
; ++i
) {
1669 isl_pw_aff
*bound_i
;
1670 isl_ast_expr
*expr_i
;
1672 bound_i
= isl_pw_aff_list_get_pw_aff(array
->bound
, i
);
1673 expr_i
= isl_ast_build_expr_from_pw_aff(build
, bound_i
);
1674 res
= isl_ast_expr_mul(res
, expr_i
);
1678 expr_i
= isl_ast_expr_get_op_arg(expr
, i
+ 1);
1679 res
= isl_ast_expr_add(res
, expr_i
);
1682 isl_ast_build_free(build
);
1684 if (1 + array
->n_index
> n
) {
1685 res
= isl_ast_expr_add(isl_ast_expr_get_op_arg(expr
, 0), res
);
1687 list
= isl_ast_expr_list_from_ast_expr(res
);
1688 res
= isl_ast_expr_get_op_arg(expr
, 0);
1689 res
= isl_ast_expr_access(res
, list
);
1692 isl_ast_expr_free(expr
);
1697 /* AST expression transformation callback for pet_stmt_build_ast_exprs.
1699 * If the AST expression refers to an array that is not accessed
1700 * at all, then this means the value of the expression is not used,
1701 * so we might as well print zero (NULL pointer) instead.
1703 * If the AST expression refers to a global scalar that is not
1704 * a read-only scalar, then its address was passed to the kernel and
1705 * we need to dereference it.
1707 * If the AST expression refers to an access to a global array,
1708 * then we linearize the access exploiting the bounds in data->local_array.
1710 static __isl_give isl_ast_expr
*transform_expr(__isl_take isl_ast_expr
*expr
,
1711 __isl_keep isl_id
*id
, void *user
)
1713 struct ppcg_transform_data
*data
= user
;
1717 if (!data
->array
->accessed
) {
1720 ctx
= isl_ast_expr_get_ctx(expr
);
1721 isl_ast_expr_free(expr
);
1722 return isl_ast_expr_from_val(isl_val_zero(ctx
));
1724 if (gpu_array_is_read_only_scalar(data
->array
))
1728 if (data
->array
->n_index
== 0)
1729 return dereference(expr
);
1730 if (!data
->array
->linearize
)
1733 return gpu_local_array_info_linearize_index(data
->local_array
, expr
);
1736 /* This function is called for each instance of a user statement
1737 * in the kernel "kernel", identified by "gpu_stmt".
1738 * "kernel" may be NULL if we are not inside a kernel.
1740 * We attach a struct ppcg_kernel_stmt to the "node", containing
1741 * a computed AST expression for each access, through an annotation
1743 * These AST expressions are computed from iterator_map,
1744 * which expresses the domain
1745 * elements in terms of the generated loops, and sched2shared,
1746 * which expresses the outer shared_schedule_dim dimensions of
1747 * the kernel schedule computed by PPCG in terms of the generated loops.
1749 static __isl_give isl_ast_node
*create_domain_leaf(
1750 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1751 __isl_keep isl_ast_build
*build
, struct gpu_stmt
*gpu_stmt
)
1753 struct ppcg_transform_data data
;
1754 struct ppcg_kernel_stmt
*stmt
;
1757 isl_pw_multi_aff
*sched2shared
;
1759 isl_pw_multi_aff
*iterator_map
;
1760 isl_union_map
*schedule
;
1764 ctx
= isl_ast_node_get_ctx(node
);
1766 stmt
= isl_calloc_type(ctx
, struct ppcg_kernel_stmt
);
1768 return isl_ast_node_free(node
);
1770 schedule
= isl_ast_build_get_schedule(build
);
1771 map
= isl_map_reverse(isl_map_from_union_map(schedule
));
1772 iterator_map
= isl_pw_multi_aff_from_map(map
);
1774 sched2shared
= compute_sched_to_shared(kernel
,
1775 isl_pw_multi_aff_copy(iterator_map
));
1777 sched2shared
= NULL
;
1779 stmt
->type
= ppcg_kernel_domain
;
1780 stmt
->u
.d
.stmt
= gpu_stmt
;
1782 data
.kernel
= kernel
;
1783 data
.accesses
= stmt
->u
.d
.stmt
->accesses
;
1784 data
.iterator_map
= iterator_map
;
1785 data
.sched2shared
= sched2shared
;
1786 stmt
->u
.d
.ref2expr
= pet_stmt_build_ast_exprs(stmt
->u
.d
.stmt
->stmt
,
1787 build
, &transform_index
, &data
,
1788 &transform_expr
, &data
);
1790 isl_pw_multi_aff_free(iterator_map
);
1791 isl_pw_multi_aff_free(sched2shared
);
1793 id
= isl_id_alloc(ctx
, "user", stmt
);
1794 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1795 return isl_ast_node_set_annotation(node
, id
);
1798 /* This function is called for each statement node in the AST
1799 * for copying to or from shared/private memory.
1800 * Attach a pointer to a ppcg_kernel_stmt representing the copy
1801 * statement to the node.
1802 * The statement name is "read" or "write", depending on whether we are
1803 * reading from global memory or writing to global memory.
1805 * The schedule is of the form
1809 * where D corresponds to the outer group->depth dimensions of
1810 * the kernel schedule, A to the global array and L to the outer
1811 * generated AST schedule.
1812 * We compute the inverse and strip off the type, resulting in
1816 * We combine this mapping with on the one hand the projection
1820 * and on the other hand the group tiling
1828 * and store the corresponding expressions in stmt->index and stmt->local_index,
1829 * where stmt points to the ppcg_kernel_stmt that is attached to the node.
1831 static __isl_give isl_ast_node
*create_access_leaf(struct ppcg_kernel
*kernel
,
1832 struct gpu_array_ref_group
*group
, __isl_take isl_ast_node
*node
,
1833 __isl_keep isl_ast_build
*build
)
1835 struct ppcg_kernel_stmt
*stmt
;
1836 struct gpu_array_tile
*tile
;
1841 isl_pw_multi_aff
*pma
, *pma2
;
1844 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1846 return isl_ast_node_free(node
);
1848 access
= isl_map_from_union_map(isl_ast_build_get_schedule(build
));
1849 type
= isl_map_get_tuple_name(access
, isl_dim_in
);
1850 stmt
->u
.c
.read
= !strcmp(type
, "read");
1851 access
= isl_map_reverse(access
);
1852 pma
= isl_pw_multi_aff_from_map(access
);
1853 pma
= isl_pw_multi_aff_reset_tuple_id(pma
, isl_dim_out
);
1855 space
= isl_space_range(isl_pw_multi_aff_get_space(pma
));
1856 space
= isl_space_unwrap(space
);
1857 pma2
= isl_pw_multi_aff_range_map(space
);
1858 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
,
1859 isl_pw_multi_aff_copy(pma
));
1860 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1861 stmt
->u
.c
.index
= expr
;
1863 tile
= gpu_array_ref_group_tile(group
);
1864 pma2
= isl_pw_multi_aff_from_multi_aff(
1865 isl_multi_aff_copy(tile
->tiling
));
1866 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
, pma
);
1867 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1868 stmt
->u
.c
.local_index
= expr
;
1870 stmt
->u
.c
.array
= group
->array
;
1871 stmt
->u
.c
.local_array
= group
->local_array
;
1872 stmt
->type
= ppcg_kernel_copy
;
1874 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1875 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1876 return isl_ast_node_set_annotation(node
, id
);
1879 /* Create a synchronization ppcg_kernel_stmt and
1880 * attach it to the node "node" representing the synchronization.
1882 static __isl_give isl_ast_node
*create_sync_leaf(
1883 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1884 __isl_keep isl_ast_build
*build
)
1886 struct ppcg_kernel_stmt
*stmt
;
1889 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1891 return isl_ast_node_free(node
);
1893 stmt
->type
= ppcg_kernel_sync
;
1894 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1895 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1896 return isl_ast_node_set_annotation(node
, id
);
1899 /* Internal data structure for at_domain.
1901 * "prog" represents the entire scop.
1902 * "kernel" points to the kernel to which the current schedule node
1903 * belongs. It is set by before_mark and reset by after_mark.
1904 * It may be NULL if we are outside any kernel.
1906 struct ppcg_at_domain_data
{
1907 struct gpu_prog
*prog
;
1908 struct ppcg_kernel
*kernel
;
1911 /* This function is called for each instance of a user statement
1912 * in the kernel. This may be one of the original user statements
1913 * or a statement introduced by PPCG.
1915 * We assume that the original user statements only have a name
1916 * and no user pointer. The statements introduced by PPCG
1917 * on the other hand all have a user pointer.
1919 * If the user statement is one of the original user statements
1920 * (one with no user pointer), then we call create_domain_leaf. Otherwise,
1921 * we check if it is a copy or synchronization statement and
1922 * call the appropriate functions.
1923 * Statements that copy an array to/from the device do not need
1924 * any further treatment.
1926 static __isl_give isl_ast_node
*at_domain(__isl_take isl_ast_node
*node
,
1927 __isl_keep isl_ast_build
*build
, void *user
)
1929 struct ppcg_at_domain_data
*data
= user
;
1930 isl_ast_expr
*expr
, *arg
;
1936 expr
= isl_ast_node_user_get_expr(node
);
1937 arg
= isl_ast_expr_get_op_arg(expr
, 0);
1938 id
= isl_ast_expr_get_id(arg
);
1939 name
= isl_id_get_name(id
);
1940 p
= isl_id_get_user(id
);
1941 isl_ast_expr_free(expr
);
1942 isl_ast_expr_free(arg
);
1945 struct gpu_stmt
*gpu_stmt
;
1947 gpu_stmt
= find_stmt(data
->prog
, id
);
1950 isl_die(data
->prog
->ctx
, isl_error_internal
,
1951 "statement not found",
1952 return isl_ast_node_free(node
));
1954 return create_domain_leaf(data
->kernel
, node
, build
, gpu_stmt
);
1957 is_sync
= gpu_tree_id_is_sync(id
, data
->kernel
);
1959 if (!prefixcmp(name
, "to_device_") || !prefixcmp(name
, "from_device_"))
1962 return isl_ast_node_free(node
);
1963 if (!strcmp(name
, "read") || !strcmp(name
, "write")) {
1964 struct gpu_array_ref_group
*group
= p
;
1965 return create_access_leaf(data
->kernel
, group
, node
, build
);
1968 isl_die(data
->prog
->ctx
, isl_error_internal
,
1969 "unknown statement type",
1970 return isl_ast_node_free(node
));
1971 return create_sync_leaf(data
->kernel
, node
, build
);
1974 /* Given a set of wrapped references "ref", return the corresponding
1975 * access relations based on the tagged access relations "tagged".
1977 * The elements of "ref" are of the form
1981 * with D an iteration domains and R a reference.
1982 * The elements of "tagged" are of the form
1988 * Extend "tagged" to include the iteration domain in the range, i.e.,
1990 * [D -> R] -> [D -> A]
1992 * apply the result to "ref" and then unwrap the resulting set
1993 * to obtain relations of the form
1997 static __isl_give isl_union_map
*wrapped_reference_to_access(
1998 __isl_take isl_union_set
*ref
, __isl_take isl_union_map
*tagged
)
2000 isl_union_map
*tag2access
;
2002 tag2access
= isl_union_map_copy(tagged
);
2003 tag2access
= isl_union_map_universe(tag2access
);
2004 tag2access
= isl_union_set_unwrap(isl_union_map_domain(tag2access
));
2005 tag2access
= isl_union_map_domain_map(tag2access
);
2006 tag2access
= isl_union_map_range_product(tag2access
, tagged
);
2008 ref
= isl_union_set_coalesce(ref
);
2009 ref
= isl_union_set_apply(ref
, tag2access
);
2011 return isl_union_set_unwrap(ref
);
2014 /* Given an access relation "access" from one or more array reference groups,
2015 * remove those reads if ("read" is 1) or writes (if "read" is 0)
2016 * that are only needed to communicate data within
2017 * the same iteration of "sched".
2018 * "tagged" contains all tagged access relations to all
2019 * the array reference groups accessed by "access" from statement
2020 * instances scheduled by "sched".
2022 * If the access is a read then it is either an element of
2024 * live_in union (range flow)
2026 * where live_in and flow may be overapproximations, or
2027 * it reads an uninitialized value (that is not live-in because
2028 * there is an intermediate kill) or it reads a value that was
2029 * written within the same (compound) statement instance.
2030 * If the access is a write then it is either an element of
2032 * live_out union (domain flow)
2034 * or it writes a value that is never read (and is not live-out
2035 * because of an intermediate kill) or only
2036 * within the same (compound) statement instance.
2037 * In both cases, the access relation is also a subset of
2038 * the group access relation.
2040 * The cases where an uninitialized value is read or a value is written
2041 * that is never read or where the dataflow occurs within a statement
2042 * instance are also considered local and may also be removed.
2044 * Essentially, we compute the intersection of "access" with either
2046 * live_in union (range non-local-flow)
2050 * live_out union (domain non-local-flow)
2052 * We first construct a relation "local"
2054 * [[D -> R] -> [D' -> R']]
2056 * of pairs of domain iterations accessing the reference group
2057 * and references in the group that are coscheduled by "sched".
2059 * If this relation does not intersect the dataflow dependences,
2060 * then there is nothing we can possibly remove, unless the dataflow
2061 * dependences themselves only relate a subset of the accesses.
2062 * In particular, the accesses may not be involved in any dataflow
2063 * dependences, either because they are uninitialized reads/dead writes
2064 * or because the dataflow occurs inside a statement instance.
2066 * Since the computation below may break up the access relation
2067 * into smaller pieces, we only perform the intersection with
2068 * the non-local dependent accesses if the local pairs
2069 * intersect the dataflow dependences. Otherwise, we intersect
2070 * with the universe of the non-local dependent accesses.
2071 * This should at least remove accesses from statements that
2072 * do not participate in any dependences.
2074 * In particular, we remove the "local" dataflow dependences from
2075 * the set of all dataflow dependences.
2076 * Note that if the potential dataflow dependences are an overapproximation
2077 * of the actual dataflow dependences, then the result remains an
2078 * overapproximation of the non-local dataflow dependences.
2079 * Copying to/from global memory is only needed for the references
2080 * in the domain/range of the result or for accesses that are live out/in
2081 * for the entire scop.
2083 * We therefore map the domain/range of the "external" relation
2084 * to the corresponding access relation and take the union with
2085 * the live out/in relation.
2087 static __isl_give isl_union_map
*remove_local_accesses(
2088 struct gpu_prog
*prog
, __isl_take isl_union_map
*tagged
,
2089 __isl_take isl_union_map
*access
, __isl_take isl_union_map
*sched
,
2093 isl_union_pw_multi_aff
*tagger
;
2094 isl_union_set
*domain
;
2095 isl_union_map
*local
, *external
;
2096 isl_union_set
*tag_set
;
2098 if (isl_union_map_is_empty(access
)) {
2099 isl_union_map_free(sched
);
2100 isl_union_map_free(tagged
);
2104 tagger
= isl_union_pw_multi_aff_copy(prog
->scop
->tagger
);
2105 domain
= isl_union_map_domain(isl_union_map_copy(tagged
));
2106 tagger
= isl_union_pw_multi_aff_intersect_domain(tagger
, domain
);
2107 sched
= isl_union_map_preimage_domain_union_pw_multi_aff(sched
, tagger
);
2109 local
= isl_union_map_apply_range(sched
,
2110 isl_union_map_reverse(isl_union_map_copy(sched
)));
2111 local
= isl_union_map_intersect(local
,
2112 isl_union_map_copy(prog
->scop
->tagged_dep_flow
));
2114 empty
= isl_union_map_is_empty(local
);
2116 external
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
2117 external
= isl_union_map_intersect_params(external
,
2118 isl_set_copy(prog
->scop
->context
));
2119 external
= isl_union_map_subtract(external
, local
);
2122 tag_set
= isl_union_map_range(external
);
2123 external
= wrapped_reference_to_access(tag_set
, tagged
);
2124 external
= isl_union_map_union(external
,
2125 isl_union_map_copy(prog
->scop
->live_in
));
2127 tag_set
= isl_union_map_domain(external
);
2128 external
= wrapped_reference_to_access(tag_set
, tagged
);
2129 external
= isl_union_map_union(external
,
2130 isl_union_map_copy(prog
->scop
->live_out
));
2134 external
= isl_union_map_free(external
);
2136 external
= isl_union_map_universe(external
);
2138 access
= isl_union_map_intersect(access
, external
);
2143 /* Given an access relation "access" from "group", remove those reads
2144 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
2145 * communicate data within the same iteration of the schedule at the
2146 * position where the copying of the group is inserted.
2147 * "node" points to this position, i.e., the depth at "node"
2148 * is equal to group->depth.
2150 * We extract a schedule that picks out the iterations of the outer
2151 * group->depth dimensions and call remove_local_accesses.
2153 static __isl_give isl_union_map
*remove_local_accesses_group(
2154 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2155 __isl_take isl_union_map
*access
, __isl_keep isl_schedule_node
*node
,
2158 isl_union_map
*sched
, *tagged
;
2160 if (isl_union_map_is_empty(access
))
2163 tagged
= group_tagged_access_relation(group
);
2164 sched
= isl_schedule_node_get_prefix_schedule_relation(node
);
2166 return remove_local_accesses(kernel
->prog
, tagged
, access
, sched
, read
);
2169 /* This function is called before the AST generator starts traversing
2170 * the schedule subtree of a node with mark "mark".
2172 * If the mark is called "kernel", store the kernel pointer in data->kernel
2173 * for use in at_domain.
2175 static int before_mark(__isl_keep isl_id
*mark
,
2176 __isl_keep isl_ast_build
*build
, void *user
)
2178 struct ppcg_at_domain_data
*data
= user
;
2182 if (!strcmp(isl_id_get_name(mark
), "kernel"))
2183 data
->kernel
= isl_id_get_user(mark
);
2187 /* This function is called after the AST generator has finished traversing
2188 * the schedule subtree of a mark node. "node" points to the corresponding
2191 * If the mark is called "kernel", then replace "node" by a user node
2192 * that "calls" the kernel, representing the launch of the kernel.
2193 * The original "node" is stored inside the kernel object so that
2194 * it can be used to print the device code.
2195 * Note that this assumes that a kernel is only launched once.
2196 * Also clear data->kernel.
2198 static __isl_give isl_ast_node
*after_mark(__isl_take isl_ast_node
*node
,
2199 __isl_keep isl_ast_build
*build
, void *user
)
2204 isl_ast_expr_list
*list
;
2205 struct ppcg_kernel
*kernel
;
2206 struct ppcg_at_domain_data
*data
= user
;
2208 ctx
= isl_ast_node_get_ctx(node
);
2209 id
= isl_ast_node_mark_get_id(node
);
2211 return isl_ast_node_free(node
);
2212 if (strcmp(isl_id_get_name(id
), "kernel") || !data
->kernel
) {
2216 kernel
= data
->kernel
;
2217 data
->kernel
= NULL
;
2218 kernel
->space
= isl_ast_build_get_schedule_space(build
);
2219 kernel
->tree
= isl_ast_node_mark_get_node(node
);
2220 isl_ast_node_free(node
);
2222 expr
= isl_ast_expr_from_id(isl_id_copy(id
));
2223 list
= isl_ast_expr_list_alloc(ctx
, 0);
2224 expr
= isl_ast_expr_call(expr
, list
);
2225 node
= isl_ast_node_alloc_user(expr
);
2226 node
= isl_ast_node_set_annotation(node
, id
);
2231 static int update_depth(__isl_keep isl_schedule_node
*node
, void *user
)
2236 if (isl_schedule_node_get_type(node
) != isl_schedule_node_leaf
)
2238 node_depth
= isl_schedule_node_get_schedule_depth(node
);
2239 if (node_depth
> *depth
)
2240 *depth
= node_depth
;
2245 /* Use isl to generate code for both the host and the device
2247 * The device code is marked by "kernel" mark nodes in the schedule tree,
2248 * containing a pointer to a ppcg_kernel object.
2249 * The returned AST only contains the AST for the host code.
2250 * The ASTs for the device code are embedded in ppcg_kernel objects
2251 * attached to the leaf nodes that call "kernel".
2253 static __isl_give isl_ast_node
*generate_code(struct gpu_gen
*gen
,
2254 __isl_take isl_schedule
*schedule
)
2256 struct ppcg_at_domain_data data
;
2257 isl_ast_build
*build
;
2259 isl_id_list
*iterators
;
2262 data
.prog
= gen
->prog
;
2266 if (isl_schedule_foreach_schedule_node(schedule
, &update_depth
,
2269 build
= isl_ast_build_alloc(gen
->prog
->ctx
);
2270 iterators
= ppcg_scop_generate_names(gen
->prog
->scop
, depth
, "c");
2271 build
= isl_ast_build_set_iterators(build
, iterators
);
2272 build
= isl_ast_build_set_at_each_domain(build
, &at_domain
, &data
);
2273 build
= isl_ast_build_set_before_each_mark(build
, &before_mark
, &data
);
2274 build
= isl_ast_build_set_after_each_mark(build
, &after_mark
, &data
);
2275 if (gen
->prog
->scop
->options
->debug
->dump_final_schedule
)
2276 isl_schedule_dump(schedule
);
2277 tree
= isl_ast_build_node_from_schedule(build
, schedule
);
2278 isl_ast_build_free(build
);
2283 __isl_give isl_union_map
*extract_sizes_from_str(isl_ctx
*ctx
, const char *str
)
2287 return isl_union_map_read_from_str(ctx
, str
);
2290 /* Can "node" be tiled and then mapped to block and thread identifiers?
2291 * That is, is it permutable with at least one coincident dimension?
2293 static int is_permutable(__isl_keep isl_schedule_node
*node
)
2298 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
2300 if (!isl_schedule_node_band_get_permutable(node
))
2302 if (isl_schedule_node_band_n_member(node
) < 1)
2304 if (!isl_schedule_node_band_member_get_coincident(node
, 0))
2310 /* A isl_schedule_foreach_schedule_node callback
2311 * for setting *any_permutable and aborting the search
2312 * if "node" is a permutable band with coincident dimensions.
2313 * Otherwise, continue searching.
2315 static int set_permutable(__isl_keep isl_schedule_node
*node
, void *user
)
2317 int *any_permutable
= user
;
2320 permutable
= is_permutable(node
);
2326 *any_permutable
= 1;
2331 /* Does "schedule" contain any permutable band with at least one coincident
2334 static int has_any_permutable_node(__isl_keep isl_schedule
*schedule
)
2336 int any_permutable
= 0;
2338 if (isl_schedule_foreach_schedule_node(schedule
, &set_permutable
,
2339 &any_permutable
) < 0 &&
2343 return any_permutable
;
2346 /* Is "node" a leaf or can it be tiled and then mapped to
2347 * block and thread identifiers?
2349 static int is_leaf_or_tilable(__isl_keep isl_schedule_node
*node
)
2351 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
2353 return is_permutable(node
);
2356 /* Is "node" the outermost node in its branch that can be tiled
2357 * and then mapped to block and thread identifiers?
2358 * If there are no such nodes in the branch and if "node" is a leaf,
2359 * then it is accepted too.
2361 static int is_outer_tilable(__isl_keep isl_schedule_node
*node
)
2364 isl_schedule_node
*ancestor
;
2366 tilable
= is_leaf_or_tilable(node
);
2373 ancestor
= isl_schedule_node_copy(node
);
2374 while (isl_schedule_node_has_parent(ancestor
)) {
2375 ancestor
= isl_schedule_node_parent(ancestor
);
2377 tilable
= is_permutable(ancestor
);
2378 if (tilable
< 0 || tilable
)
2382 isl_schedule_node_free(ancestor
);
2383 return tilable
< 0 ? -1 : !tilable
;
2386 /* Collect the references to all writes in "group".
2387 * Each reference is represented by a universe set in a space
2391 * with S[i,j] the statement instance space and R[] the array reference.
2393 static __isl_give isl_union_set
*group_tagged_writes(
2394 struct gpu_array_ref_group
*group
)
2398 isl_union_set
*writes
;
2400 space
= isl_map_get_space(group
->access
);
2401 writes
= isl_union_set_empty(space
);
2402 for (i
= 0; i
< group
->n_ref
; ++i
) {
2406 if (!group
->refs
[i
]->write
)
2409 space
= isl_map_get_space(group
->refs
[i
]->tagged_access
);
2410 space
= isl_space_domain(space
);
2411 writes_i
= isl_set_universe(space
);
2412 writes
= isl_union_set_add_set(writes
, writes_i
);
2418 /* Is there any write access in "group" that requires synchronization
2419 * on a write to global memory?
2420 * We currently take into account all writes that would require
2421 * synchronization at the thread level depth, but if the copying
2422 * for this group is performed at an outer level, then we do not
2423 * actually need to take into account dependences at intermediate levels.
2425 static int any_sync_writes_in_group(struct ppcg_kernel
*kernel
,
2426 struct gpu_array_ref_group
*group
)
2428 isl_union_set
*writes
;
2429 int empty
, disjoint
;
2431 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2437 writes
= group_tagged_writes(group
);
2438 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2439 isl_union_set_free(writes
);
2441 return disjoint
< 0 ? -1 : !disjoint
;
2444 /* Collect the references to all writes in "kernel" that write directly
2445 * to global or shared memory, i.e., that are not mapped to private memory.
2446 * Each reference is represented by a universe set in a space
2450 * with S[i,j] the statement instance space and R[] the array reference.
2452 static __isl_give isl_union_set
*collect_non_private_tagged_writes(
2453 struct ppcg_kernel
*kernel
)
2455 isl_union_set
*writes
;
2458 writes
= isl_union_set_empty(isl_union_set_get_space(kernel
->arrays
));
2460 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2461 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2463 for (j
= 0; j
< array
->n_group
; ++j
) {
2464 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2465 isl_union_set
*writes_ij
;
2469 if (group
->private_tile
)
2471 writes_ij
= group_tagged_writes(group
);
2472 writes
= isl_union_set_union(writes
, writes_ij
);
2479 /* Are there any direct writes to global memory that require
2482 static int any_global_or_shared_sync_writes(struct ppcg_kernel
*kernel
)
2484 isl_union_set
*writes
;
2485 int empty
, disjoint
;
2487 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2493 writes
= collect_non_private_tagged_writes(kernel
);
2494 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2495 isl_union_set_free(writes
);
2497 return disjoint
< 0 ? -1 : !disjoint
;
2500 /* Construct an isl_multi_val for use as tile sizes for tiling "node"
2501 * from the elements in "tile_size".
2503 static __isl_give isl_multi_val
*construct_band_tiles_sizes(
2504 __isl_keep isl_schedule_node
*node
, int *tile_size
)
2514 ctx
= isl_schedule_node_get_ctx(node
);
2515 space
= isl_schedule_node_band_get_space(node
);
2516 n
= isl_schedule_node_band_n_member(node
);
2517 mv
= isl_multi_val_zero(space
);
2518 for (i
= 0; i
< n
; ++i
) {
2521 v
= isl_val_int_from_si(ctx
, tile_size
[i
]);
2522 mv
= isl_multi_val_set_val(mv
, i
, v
);
2528 /* Replace the partial schedule S of the band node "node" by
2536 * if scale_tile_loops is set, with f the integers in "factor".
2537 * The list that "factor" points to is assumed to contain at least
2538 * as many elements as the number of members in the band.
2540 static __isl_give isl_schedule_node
*snap_band_to_sizes(
2541 __isl_take isl_schedule_node
*node
, int *factor
,
2542 struct ppcg_options
*options
)
2546 mv
= construct_band_tiles_sizes(node
, factor
);
2547 node
= isl_schedule_node_band_scale_down(node
, isl_multi_val_copy(mv
));
2548 if (options
->scale_tile_loops
)
2549 node
= isl_schedule_node_band_scale(node
,
2550 isl_multi_val_copy(mv
));
2551 isl_multi_val_free(mv
);
2556 /* Tile "band" with tile size specified by "sizes".
2558 * Since the tile loops will be mapped to block ids, we forcibly
2559 * turn off tile loop scaling. We may want to enable tile loop scaling
2560 * at some later point, but then we would have to support the detection
2561 * of strides during the mapping to block ids.
2562 * Similarly, since the point loops will be mapped to thread ids,
2563 * we forcibly shift the point loops so that they start at zero.
2565 static __isl_give isl_schedule_node
*tile_band(
2566 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2568 isl_ctx
*ctx
= isl_schedule_node_get_ctx(node
);
2572 scale_tile
= isl_options_get_tile_scale_tile_loops(ctx
);
2573 isl_options_set_tile_scale_tile_loops(ctx
, 0);
2574 shift_point
= isl_options_get_tile_shift_point_loops(ctx
);
2575 isl_options_set_tile_shift_point_loops(ctx
, 1);
2577 node
= isl_schedule_node_band_tile(node
, sizes
);
2579 isl_options_set_tile_scale_tile_loops(ctx
, scale_tile
);
2580 isl_options_set_tile_shift_point_loops(ctx
, shift_point
);
2585 /* Extract the set of parameter values and outer schedule dimensions
2586 * for which any statement instance
2587 * in the kernel inserted at "node" needs to be executed.
2588 * Intersect the set of parameter values derived from the host schedule
2589 * relation with the context of "prog".
2591 static __isl_give isl_set
*extract_context(__isl_keep isl_schedule_node
*node
,
2592 struct gpu_prog
*prog
)
2594 isl_union_map
*schedule
;
2595 isl_union_set
*schedule_domain
;
2599 schedule
= isl_schedule_node_get_prefix_schedule_relation(node
);
2600 schedule_domain
= isl_union_map_range(schedule
);
2601 empty
= isl_union_set_is_empty(schedule_domain
);
2603 isl_union_set_free(schedule_domain
);
2610 space
= isl_union_set_get_space(schedule_domain
);
2611 isl_union_set_free(schedule_domain
);
2612 space
= isl_space_set_from_params(space
);
2613 depth
= isl_schedule_node_get_schedule_depth(node
);
2614 space
= isl_space_add_dims(space
, isl_dim_set
, depth
);
2615 context
= isl_set_empty(space
);
2617 context
= isl_set_from_union_set(schedule_domain
);
2619 context
= isl_set_intersect_params(context
,
2620 isl_set_copy(prog
->context
));
2625 /* Return the set of outer array elements accessed by
2626 * by the statement instance in "domain" in "prog".
2628 static __isl_give isl_union_set
*accessed_by_domain(
2629 __isl_take isl_union_set
*domain
, struct gpu_prog
*prog
)
2631 isl_union_map
*access
;
2632 isl_union_set
*arrays
;
2634 access
= isl_union_map_union(isl_union_map_copy(prog
->read
),
2635 isl_union_map_copy(prog
->may_write
));
2636 access
= isl_union_map_intersect_domain(access
, domain
);
2637 arrays
= isl_union_map_range(access
);
2638 arrays
= isl_union_set_apply(arrays
,
2639 isl_union_map_copy(prog
->to_outer
));
2644 /* Return the number of outer band members of the band node "node"
2645 * that are marked coincident.
2647 static int n_outer_coincidence(__isl_keep isl_schedule_node
*node
)
2651 n
= isl_schedule_node_band_n_member(node
);
2653 for (i
= 0; i
< n
; ++i
)
2654 if (!isl_schedule_node_band_member_get_coincident(node
, i
))
2660 /* If the band node "node" has more than "n" members, then split off
2661 * the first "n" of them.
2663 static __isl_give isl_schedule_node
*split_band(
2664 __isl_take isl_schedule_node
*node
, int n
)
2668 dim
= isl_schedule_node_band_n_member(node
);
2670 node
= isl_schedule_node_band_split(node
, n
);
2675 /* Scale a band node that may have been split by split_band.
2676 * "sizes" are the scaling factors for the original node.
2677 * "node" either points to the original band node, or the outer
2678 * of the two pieces after splitting.
2680 * If the number of elements in "node" is smaller than the number of
2681 * elements in "sizes", then some splitting has occurred and we split
2682 * "sizes" in the same way.
2684 static __isl_give isl_schedule_node
*scale_band(
2685 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2689 n
= isl_multi_val_dim(sizes
, isl_dim_set
);
2690 dim
= isl_schedule_node_band_n_member(node
);
2692 isl_multi_val
*sizes2
;
2694 sizes2
= isl_multi_val_copy(sizes
);
2695 sizes
= isl_multi_val_drop_dims(sizes
,
2696 isl_dim_set
, dim
, n
- dim
);
2697 sizes2
= isl_multi_val_drop_dims(sizes2
, isl_dim_set
, 0, dim
);
2698 node
= isl_schedule_node_child(node
, 0);
2699 node
= isl_schedule_node_band_scale(node
, sizes2
);
2700 node
= isl_schedule_node_parent(node
);
2703 return isl_schedule_node_band_scale(node
, sizes
);
2706 /* Return an isl_multi_aff, with as elements the parameters in "space"
2707 * that have the names specified by the elements in "names".
2708 * If (some of) these parameters do not already appear in "space",
2709 * then they are added first.
2711 static __isl_give isl_multi_aff
*parameter_vector(__isl_take isl_space
*space
,
2712 __isl_keep isl_id_list
*names
)
2715 isl_local_space
*ls
;
2719 space
= isl_space_free(space
);
2721 n
= isl_id_list_n_id(names
);
2722 for (i
= 0; i
< n
; ++i
) {
2726 id
= isl_id_list_get_id(names
, i
);
2727 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2732 pos
= isl_space_dim(space
, isl_dim_param
);
2733 space
= isl_space_add_dims(space
, isl_dim_param
, 1);
2734 space
= isl_space_set_dim_id(space
, isl_dim_param
, pos
, id
);
2736 ma
= isl_multi_aff_zero(isl_space_copy(space
));
2737 ls
= isl_local_space_from_space(isl_space_domain(space
));
2738 for (i
= 0; i
< n
; ++i
) {
2743 id
= isl_id_list_get_id(names
, i
);
2744 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2746 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
2747 isl_dim_param
, pos
);
2748 ma
= isl_multi_aff_set_aff(ma
, i
, aff
);
2750 isl_local_space_free(ls
);
2755 /* Return constraints on the domain elements that equate a sequence of
2756 * parameters called "names", to the partial schedule
2757 * of "node" modulo the integers in "size".
2758 * The number of elements in the array "size" should be equal
2759 * to the number of elements in "names".
2760 * The number of members of the band node "node" should be smaller
2761 * than or equal to this number. If it is smaller, then the first
2762 * elements of "names" are equated to zero.
2764 static __isl_give isl_union_set
*set_schedule_modulo(
2765 __isl_keep isl_schedule_node
*node
, __isl_keep isl_id_list
*names
,
2771 isl_multi_union_pw_aff
*mupa
, *mupa2
;
2773 isl_union_set
*domain
;
2777 n
= isl_id_list_n_id(names
);
2779 return isl_schedule_node_get_universe_domain(node
);
2780 n_zero
= n
- isl_schedule_node_band_n_member(node
);
2782 mupa
= isl_schedule_node_band_get_partial_schedule(node
);
2783 mv
= construct_band_tiles_sizes(node
, size
+ n_zero
);
2784 mupa
= isl_multi_union_pw_aff_mod_multi_val(mupa
, mv
);
2786 space
= isl_multi_union_pw_aff_get_space(mupa
);
2787 space
= isl_space_params(space
);
2788 space
= isl_space_set_from_params(space
);
2789 space
= isl_space_add_dims(space
, isl_dim_set
, n_zero
);
2790 ma
= isl_multi_aff_zero(space
);
2792 domain
= isl_schedule_node_get_universe_domain(node
);
2793 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(
2794 isl_union_set_copy(domain
), ma
);
2795 mupa
= isl_multi_union_pw_aff_range_product(mupa2
, mupa
);
2797 space
= isl_multi_union_pw_aff_get_space(mupa
);
2798 ma
= parameter_vector(space
, names
);
2800 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(domain
, ma
);
2801 mupa
= isl_multi_union_pw_aff_sub(mupa
, mupa2
);
2803 return isl_multi_union_pw_aff_zero_union_set(mupa
);
2806 /* Insert a context node at "node" introducing the block and thread
2807 * identifiers along with their bounds, which are stored in kernel->grid_size
2808 * and kernel->block_dim.
2809 * Note that the bounds on the block identifiers may implicitly impose
2810 * constraints on the parameters. A guard needs to be inserted
2811 * in the schedule tree to ensure that those bounds hold at "node".
2812 * This guard is inserted in insert_guard.
2814 static __isl_give isl_schedule_node
*insert_context(struct ppcg_kernel
*kernel
,
2815 __isl_take isl_schedule_node
*node
)
2819 context
= isl_set_universe(isl_set_get_space(kernel
->context
));
2821 context
= add_bounded_parameters_dynamic(context
,
2822 kernel
->grid_size
, kernel
->block_ids
);
2823 context
= add_bounded_parameters(context
,
2824 kernel
->block_dim
, kernel
->thread_ids
);
2826 node
= isl_schedule_node_insert_context(node
, context
);
2831 /* Insert a guard that eliminates kernel launches where the kernel
2832 * obviously does not have any work to do.
2834 * In particular, eliminate kernel launches where there are obviously
2836 * Use the same block size constraints that are used to create the context
2837 * to ensure that all constraints implicit in the constructed context
2838 * are imposed by the guard.
2840 * Additionally, add other constraints that are valid
2841 * for each executed instance ("context"), as long as this does not result
2844 static __isl_give isl_schedule_node
*insert_guard(
2845 __isl_take isl_schedule_node
*node
, __isl_keep isl_set
*context
,
2846 __isl_keep isl_multi_pw_aff
*size
, struct ppcg_scop
*scop
)
2852 guard
= isl_set_copy(context
);
2853 guard
= isl_set_compute_divs(guard
);
2854 guard
= isl_set_from_basic_set(isl_set_simple_hull(guard
));
2856 nparam
= isl_set_dim(guard
, isl_dim_param
);
2857 n
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
2858 ids
= ppcg_scop_generate_names(scop
, n
, "__ppcg_tmp");
2859 guard
= add_bounded_parameters_dynamic(guard
, size
, ids
);
2860 isl_id_list_free(ids
);
2861 guard
= isl_set_project_out(guard
, isl_dim_param
, nparam
, n
);
2863 node
= isl_schedule_node_insert_guard(node
, guard
);
2868 /* Does any array reference group mapping require the band that is mapped
2869 * to threads to be unrolled?
2871 static int kernel_requires_unroll(struct ppcg_kernel
*kernel
)
2875 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2876 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2878 for (j
= 0; j
< array
->n_group
; ++j
) {
2879 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2880 if (gpu_array_ref_group_requires_unroll(group
))
2888 /* Mark the given band node "node" for unrolling by the AST generator and
2889 * then sink it to the leaves of the schedule tree.
2890 * All dimensions of "node" are assumed to be coincident, such that this
2891 * sinking is a valid operation.
2893 static __isl_give isl_schedule_node
*unroll(__isl_take isl_schedule_node
*node
)
2897 n
= isl_schedule_node_band_n_member(node
);
2898 for (i
= 0; i
< n
; ++i
)
2899 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
2900 isl_ast_loop_unroll
);
2902 node
= isl_schedule_node_band_sink(node
);
2907 /* Insert a synchronization node in the schedule tree of "node"
2908 * after the core computation of "kernel" at the level of the band
2909 * that is mapped to threads, except if that level is equal to
2910 * that of the band that is mapped to blocks or if there are no writes
2911 * to global or shared memory in the core computation that require
2913 * If there are any writes to shared memory and the shared memory
2914 * copying is performed at the same level, then synchronization
2915 * is needed between the core and the copying anyway, so we might
2916 * as well add it here. If the copying is performed at a higher
2917 * level, then different iterations of intermediate schedule dimensions
2918 * may have a different mapping from between shared memory elements and
2919 * threads, such that synchronization is required after the core.
2920 * "node" is assumed to point to the kernel node.
2922 static __isl_give isl_schedule_node
*add_sync(struct ppcg_kernel
*kernel
,
2923 __isl_take isl_schedule_node
*node
)
2928 need_sync
= any_global_or_shared_sync_writes(kernel
);
2930 return isl_schedule_node_free(node
);
2934 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
2936 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
2937 if (kernel_depth
== isl_schedule_node_get_schedule_depth(node
))
2938 return gpu_tree_move_up_to_kernel(node
);
2940 node
= gpu_tree_ensure_following_sync(node
, kernel
);
2942 node
= gpu_tree_move_up_to_kernel(node
);
2947 /* Return a read ("read" is 1) or write access relation for "group"
2948 * with those accesses removed that are only needed to communicate data
2949 * within the subtree of the schedule rooted at "node".
2950 * Furthermore, include the prefix schedule at "node".
2951 * That is, return a relation of the form
2955 * with D the outer schedule dimensions at "node".
2957 static __isl_give isl_union_map
*anchored_non_local_accesses(
2958 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2959 __isl_take isl_schedule_node
*node
, int read
)
2961 isl_union_map
*access
;
2962 isl_union_map
*prefix
;
2964 access
= gpu_array_ref_group_access_relation(group
, read
, !read
);
2965 access
= remove_local_accesses_group(kernel
, group
, access
, node
, read
);
2966 prefix
= isl_schedule_node_get_prefix_schedule_relation(node
);
2967 access
= isl_union_map_range_product(prefix
, access
);
2972 /* Given an array reference group "group", create a mapping
2974 * read[D -> A] -> [D -> A]
2976 * if "read" is set or
2978 * write[D -> A] -> [D -> A]
2980 * if "read" is not set.
2981 * D corresponds to the outer group->depth dimensions of
2982 * the kernel schedule.
2984 static __isl_give isl_multi_aff
*create_from_access(isl_ctx
*ctx
,
2985 struct gpu_array_ref_group
*group
, int read
)
2990 space
= isl_space_copy(group
->array
->space
);
2991 space
= isl_space_from_range(space
);
2992 space
= isl_space_add_dims(space
, isl_dim_in
, group
->depth
);
2993 space
= isl_space_wrap(space
);
2994 space
= isl_space_map_from_set(space
);
2996 id
= isl_id_alloc(ctx
, read
? "read" : "write", group
);
2997 space
= isl_space_set_tuple_id(space
, isl_dim_in
, id
);
2999 return isl_multi_aff_identity(space
);
3002 /* If any writes in "group" require synchronization, then make sure
3003 * that there is a synchronization node for "kernel" after the node
3004 * following "node" in a sequence.
3006 * If "shared" is set and no synchronization is needed for
3007 * the writes to global memory, then add synchronization before
3008 * the kernel to protect shared memory from being overwritten
3009 * by the next iteration of the core computation.
3010 * No additional synchronization is needed to protect against
3011 * the next copy into shared memory because each element of
3012 * the shared memory tile is always copied by the same thread.
3014 static __isl_give isl_schedule_node
*add_group_write_sync(
3015 __isl_take isl_schedule_node
*node
, struct ppcg_kernel
*kernel
,
3016 struct gpu_array_ref_group
*group
, int shared
)
3020 need_sync
= any_sync_writes_in_group(kernel
, group
);
3022 return isl_schedule_node_free(node
);
3024 node
= isl_schedule_node_parent(node
);
3025 node
= isl_schedule_node_next_sibling(node
);
3026 node
= isl_schedule_node_child(node
, 0);
3027 node
= gpu_tree_ensure_following_sync(node
, kernel
);
3028 } else if (shared
) {
3029 node
= isl_schedule_node_parent(node
);
3030 node
= isl_schedule_node_parent(node
);
3031 node
= gpu_tree_move_down_to_depth(node
, group
->depth
,
3033 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3039 /* Add copy statements to the schedule tree of "node"
3040 * for reading from global memory to private memory (if "read" is set) or
3041 * for writing back from private memory to global memory
3042 * (if "read" is not set) for the array reference group "group" that
3043 * is mapped to private memory.
3044 * On input, "node" points to the kernel node, and it is moved
3045 * back there on output.
3047 * The copies are performed in the order of the array elements.
3048 * The copy statement instances include a reference to the outer
3049 * group->depth dimensions of the kernel schedule for ease of
3050 * combining them with the group tiling.
3052 * That is, the extra schedule is of the form
3056 * where D corresponds to the outer group->depth dimensions of
3057 * the kernel schedule and A to the global array.
3058 * This schedule is unrolled because registers are not addressable.
3060 * The copying is inserted in the schedule tree through an extension
3065 * where the extra domain elements type[D -> A] are those accessed
3067 * A filter is inserted on type[D -> A] to ensure that the element
3068 * is read/written by the same thread that needs the element.
3069 * This filter is obtained by applying
3073 * to the thread filter for the core statements.
3075 * The extension is inserted before the core computation in case of a read
3076 * and after the core computation in case of a write.
3077 * In the latter case, we also make sure that there is a synchronization
3078 * node after the write to global memory, unless this write is performed
3079 * at the outer level of the kernel.
3080 * In principle, this synchronization could be inserted higher
3081 * in the schedule tree depending on where the corresponding reads
3082 * from global memory are performed.
3084 static __isl_give isl_schedule_node
*add_copies_group_private(
3085 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3086 __isl_take isl_schedule_node
*node
, int read
)
3088 isl_union_map
*access
;
3089 isl_union_map
*prefix
;
3090 isl_union_set
*domain
;
3092 isl_multi_aff
*from_access
;
3093 isl_multi_pw_aff
*mpa
;
3094 isl_multi_union_pw_aff
*mupa
;
3095 isl_schedule_node
*graft
;
3096 isl_union_set
*filter
;
3100 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3101 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3103 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3104 empty
= isl_union_map_is_empty(access
);
3105 if (empty
< 0 || empty
) {
3106 isl_union_map_free(access
);
3108 return isl_schedule_node_free(node
);
3109 return gpu_tree_move_up_to_kernel(node
);
3112 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3113 space
= isl_space_domain(isl_multi_aff_get_space(from_access
));
3114 access
= isl_union_map_preimage_range_multi_aff(access
, from_access
);
3116 filter
= isl_union_set_copy(kernel
->thread_filter
);
3117 filter
= isl_union_set_apply(filter
, isl_union_map_copy(access
));
3118 filter
= isl_union_set_detect_equalities(filter
);
3119 filter
= isl_union_set_coalesce(filter
);
3121 domain
= isl_union_map_range(access
);
3122 access
= isl_union_set_wrapped_domain_map(domain
);
3123 access
= isl_union_map_reverse(access
);
3124 access
= isl_union_map_coalesce(access
);
3125 graft
= isl_schedule_node_from_extension(access
);
3127 space
= isl_space_map_from_set(space
);
3128 mpa
= isl_multi_pw_aff_identity(space
);
3129 mpa
= isl_multi_pw_aff_range_factor_range(mpa
);
3130 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3132 graft
= isl_schedule_node_child(graft
, 0);
3133 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3134 graft
= unroll(graft
);
3136 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3138 graft
= isl_schedule_node_parent(graft
);
3141 node
= isl_schedule_node_graft_before(node
, graft
);
3143 node
= isl_schedule_node_graft_after(node
, graft
);
3144 if (kernel_depth
< group
->depth
)
3145 node
= add_group_write_sync(node
, kernel
, group
, 0);
3148 node
= gpu_tree_move_up_to_kernel(node
);
3153 /* Add copy statements to the schedule tree of "node"
3154 * for reading from global memory to shared memory (if "read" is set) or
3155 * for writing back from shared memory to global memory
3156 * (if "read" is not set) for the array reference group "group" that
3157 * is mapped to shared memory.
3158 * On input, "node" points to the kernel node, and it is moved
3159 * back there on output.
3161 * The copies are performed in the order of the corresponding shared
3163 * The copy statement instances include a reference to the outer
3164 * group->depth dimensions of the kernel schedule for ease of
3165 * combining them with the group tiling.
3167 * If we are performing a read from global memory to shared memory and
3168 * if the array involved is not a scalar, then we copy
3169 * the entire tile to shared memory. This may result in some extra
3170 * elements getting copied, but it should lead to simpler code
3171 * (which means that fewer registers may be needed) and less divergence.
3173 * Otherwise, we only copy the elements that will be read or have been written
3176 * That is, the extra schedule is of the form
3180 * where D corresponds to the outer group->depth dimensions of
3181 * the kernel schedule, A to the global array and T is the corresponding
3182 * shared memory tile.
3184 * The copying is inserted in the schedule tree through an extension
3189 * where the extra domain elements type[D -> A] are those accessed
3190 * by the group. In the case of read from a non-scalar, this set
3191 * is replaced by the entire shared memory tile.
3193 * A filter is inserted on type[D -> A] to map the copy instances
3194 * to the threads. In particular, the thread identifiers are
3195 * equated to the position inside the shared memory tile (T)
3196 * modulo the block size.
3197 * We try to align the innermost tile dimension with the innermost
3198 * thread identifier (x) as a heuristic to improve coalescing.
3199 * In particular, if the dimension of the tile is greater than
3200 * the dimension of the block, then the schedule mapping to the tile
3201 * is broken up into two pieces and the filter is applied to the inner part.
3202 * If, on the other hand, the dimension of the tile is smaller than
3203 * the dimension of the block, then the initial thread identifiers
3204 * are equated to zero and the remaining thread identifiers are
3205 * matched to the memory tile.
3207 * The extension is inserted before the core computation in case of a read
3208 * and after the core computation in case of a write.
3209 * In the case of a read, we first need to make sure there is some
3210 * synchronization before the core computation such that we can put the read
3211 * from global memory to shared memory before that synchronization.
3212 * This ensures that all threads have finished copying into shared memory
3213 * before the shared memory is used.
3214 * We also need to make sure that there is a synchronization node after
3215 * the core computation to ensure that the next load into shared memory
3216 * only happens after all data has been used. There is no need for
3217 * this synchronization if we are at the outer level since then there
3218 * won't be a next load.
3219 * In the case of a write, we need to make sure there is some synchronization
3220 * after the core computation such taht we can put the write from shared
3221 * memory to global memory after that synchronization.
3222 * Unless we are at the outer level, we also need a synchronization node
3223 * after the write to ensure the data is saved to global memory
3224 * before the next iteration write to the same shared memory.
3225 * It also makes sure the data has arrived in global memory before
3226 * it is read in a subsequent iteration.
3228 static __isl_give isl_schedule_node
*add_copies_group_shared(
3229 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3230 __isl_take isl_schedule_node
*node
, int read
)
3232 struct gpu_array_tile
*tile
;
3233 isl_union_map
*access
;
3234 isl_union_set
*domain
;
3235 isl_union_set
*sync
;
3237 isl_multi_aff
*from_access
;
3238 isl_multi_pw_aff
*mpa
;
3239 isl_multi_union_pw_aff
*mupa
;
3240 isl_schedule_node
*graft
;
3241 isl_union_set
*filter
;
3246 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3247 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3249 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3250 empty
= isl_union_map_is_empty(access
);
3251 if (empty
< 0 || empty
) {
3252 isl_union_map_free(access
);
3254 return isl_schedule_node_free(node
);
3255 return gpu_tree_move_up_to_kernel(node
);
3258 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3260 tile
= gpu_array_ref_group_tile(group
);
3261 ma
= isl_multi_aff_copy(tile
->tiling
);
3262 ma
= isl_multi_aff_pullback_multi_aff(ma
,
3263 isl_multi_aff_copy(from_access
));
3264 mpa
= isl_multi_pw_aff_from_multi_aff(ma
);
3265 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3267 domain
= isl_union_map_range(access
);
3269 if (read
&& !gpu_array_is_scalar(group
->array
)) {
3271 isl_union_set_free(domain
);
3272 map
= group_tile(group
);
3273 domain
= isl_union_set_from_set(isl_map_wrap(map
));
3276 domain
= isl_union_set_preimage_multi_aff(domain
, from_access
);
3277 access
= isl_union_set_wrapped_domain_map(domain
);
3278 access
= isl_union_map_reverse(access
);
3279 access
= isl_union_map_coalesce(access
);
3280 graft
= isl_schedule_node_from_extension(access
);
3282 graft
= isl_schedule_node_child(graft
, 0);
3284 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3286 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0) {
3287 graft
= isl_schedule_node_band_split(graft
,
3288 tile
->n
- kernel
->n_block
);
3289 graft
= isl_schedule_node_child(graft
, 0);
3291 if (tile
->n
< kernel
->n_block
)
3292 skip
= kernel
->n_block
- tile
->n
;
3295 filter
= set_schedule_modulo(graft
, kernel
->thread_ids
,
3297 if (!kernel
->options
->wrap
)
3298 graft
= snap_band_to_sizes(graft
, kernel
->block_dim
+ skip
,
3300 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0)
3301 graft
= isl_schedule_node_parent(graft
);
3302 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3304 while (graft
&& isl_schedule_node_has_parent(graft
))
3305 graft
= isl_schedule_node_parent(graft
);
3308 if (kernel_depth
< group
->depth
)
3309 node
= gpu_tree_ensure_sync_after_core(node
, kernel
);
3310 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3311 node
= isl_schedule_node_graft_before(node
, graft
);
3313 node
= gpu_tree_move_right_to_sync(node
, kernel
);
3314 node
= isl_schedule_node_graft_after(node
, graft
);
3315 if (kernel_depth
< group
->depth
)
3316 node
= add_group_write_sync(node
, kernel
, group
, 1);
3319 node
= gpu_tree_move_up_to_kernel(node
);
3324 /* Check whether the array reference group "group" is mapped to
3325 * private or shared memory and, if so,
3326 * add copy statements to the schedule tree of "node"
3327 * for reading from global memory to private or shared memory
3328 * (if "read" is set) or for writing back from private or shared memory
3329 * to global memory (if "read" is not set) for this group.
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_group(
3334 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3335 __isl_take isl_schedule_node
*node
, int read
)
3337 if (group
->private_tile
)
3338 return add_copies_group_private(kernel
, group
, node
, read
);
3339 if (group
->shared_tile
)
3340 return add_copies_group_shared(kernel
, group
, node
, read
);
3344 /* For each array reference group that is mapped to private or shared memory,
3345 * add copy statements to the schedule tree of "node"
3346 * for reading from global memory to private or shared memory
3347 * and for writing back.
3348 * On input, "node" points to the kernel node, and it is moved
3349 * back there on output.
3351 static __isl_give isl_schedule_node
*add_copies(struct ppcg_kernel
*kernel
,
3352 __isl_take isl_schedule_node
*node
)
3356 for (i
= 0; i
< kernel
->n_array
; ++i
) {
3357 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
3359 for (j
= 0; j
< array
->n_group
; ++j
) {
3360 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3362 node
= add_copies_group(kernel
, group
, node
, 1);
3365 node
= add_copies_group(kernel
, group
, node
, 0);
3374 /* Mark all dimensions in the current band node atomic.
3376 static __isl_give isl_schedule_node
*atomic(__isl_take isl_schedule_node
*node
)
3380 n
= isl_schedule_node_band_n_member(node
);
3381 for (i
= 0; i
< n
; ++i
)
3382 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
3383 isl_ast_loop_atomic
);
3388 /* Mark "node" atomic, if it is a band node.
3389 * Do the same for all ancestors.
3390 * Return a pointer to "node" (in the updated schedule tree).
3392 static __isl_give isl_schedule_node
*atomic_ancestors(
3393 __isl_take isl_schedule_node
*node
)
3399 if (!isl_schedule_node_has_parent(node
))
3402 pos
= isl_schedule_node_get_child_position(node
);
3403 node
= isl_schedule_node_parent(node
);
3404 if (isl_schedule_node_get_type(node
) == isl_schedule_node_band
)
3405 node
= atomic(node
);
3406 node
= atomic_ancestors(node
);
3407 node
= isl_schedule_node_child(node
, pos
);
3412 /* Collect all write references that require synchronization.
3413 * "node" is assumed to point to the kernel node.
3414 * Each reference is represented by a universe set in a space
3418 * with S[i,j] the statement instance space and R[] the array reference.
3420 * This function should be called before block and thread filters are added.
3422 * Synchronization is needed after a write if there is a subsequent read
3423 * within the same block that may not be performed by the same thread.
3424 * There should not be any dependences between different blocks,
3425 * so we start with the flow dependences within the same kernel invocation
3426 * and we subtract from these those dependences that are mapped
3427 * to the same iteration of the bands where synchronization is inserted.
3428 * We do not remove pairs of instances that are known to map to
3429 * the same thread across different iterations of the intermediate
3430 * bands because the read may be performed by a different thread
3431 * than the one that needs the value if shared memory is involved.
3433 * We also consider all pairs of possible writes that access the same
3434 * memory location and that may be mapped to the same block but not
3435 * to the same iteration of the intermediate bands.
3436 * In theory, it would be possible for one thread to still be in
3437 * a previous iteration of a loop in these bands.
3438 * A write to global memory in this delayed thread could then overwrite
3439 * a write from another thread that has already moved on to
3440 * the next iteration.
3442 * After computing the above writes paired off with reads or writes
3443 * that depend on them, we project onto the domain writes.
3444 * Sychronization is needed after writes to global memory
3445 * through these references.
3447 static __isl_give isl_union_set
*compute_sync_writes(
3448 struct ppcg_kernel
*kernel
, __isl_keep isl_schedule_node
*node
)
3450 isl_union_map
*local
;
3451 isl_union_map
*may_writes
, *shared_access
;
3452 isl_union_map
*kernel_prefix
, *thread_prefix
;
3453 isl_union_map
*equal
;
3454 isl_union_set
*wrap
;
3455 isl_union_set
*domain
;
3457 domain
= isl_schedule_node_get_universe_domain(node
);
3458 kernel_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3459 node
= isl_schedule_node_copy(node
);
3460 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3461 thread_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3462 isl_schedule_node_free(node
);
3464 may_writes
= isl_union_map_copy(kernel
->prog
->scop
->tagged_may_writes
);
3465 may_writes
= isl_union_map_curry(may_writes
);
3466 may_writes
= isl_union_map_intersect_domain(may_writes
, domain
);
3467 may_writes
= isl_union_map_uncurry(may_writes
);
3468 shared_access
= isl_union_map_copy(may_writes
);
3469 shared_access
= isl_union_map_apply_range(shared_access
,
3470 isl_union_map_reverse(may_writes
));
3472 local
= isl_union_map_copy(kernel
->prog
->scop
->tagged_dep_flow
);
3473 local
= isl_union_map_union(local
, shared_access
);
3474 local
= isl_union_map_zip(local
);
3476 equal
= isl_union_map_apply_range(kernel_prefix
,
3477 isl_union_map_reverse(isl_union_map_copy(kernel_prefix
)));
3478 wrap
= isl_union_map_wrap(equal
);
3479 local
= isl_union_map_intersect_domain(local
, wrap
);
3480 equal
= isl_union_map_apply_range(thread_prefix
,
3481 isl_union_map_reverse(isl_union_map_copy(thread_prefix
)));
3482 wrap
= isl_union_map_wrap(equal
);
3483 local
= isl_union_map_subtract_domain(local
, wrap
);
3485 local
= isl_union_map_zip(local
);
3486 local
= isl_union_map_universe(local
);
3488 return isl_union_map_domain(local
);
3491 /* Group the domain elements into a single space, named kernelX,
3492 * with X the kernel sequence number "kernel_id".
3494 static __isl_give isl_schedule_node
*group_statements(
3495 __isl_take isl_schedule_node
*node
, int kernel_id
)
3503 snprintf(buffer
, sizeof(buffer
), "kernel%d", kernel_id
);
3504 id
= isl_id_alloc(isl_schedule_node_get_ctx(node
), buffer
, NULL
);
3505 return isl_schedule_node_group(node
, id
);
3508 /* Create a ppcg_kernel representing the domain instances that reach "node"
3509 * and insert a mark node pointing to the ppcg_kernel before "node".
3510 * The band that "node" points to is the band that needs to be mapped
3511 * to block identifiers. The band that needs to be mapped to thread
3512 * identifiers should be marked by a "thread" mark by the caller.
3513 * This mark is removed by this function.
3514 * If "scale" is set, then the band that "node" points to is scaled
3517 * Mark all outer band nodes as atomic to ensure each kernel is only
3519 * If the domain elements that reach "node" live in more than one space,
3520 * then group the domain elements into a single space, named kernelX,
3521 * with X the kernel sequence number.
3523 * Insert a guard node governing the kernel node to ensure that
3524 * no kernels with zero blocks are launched.
3526 * Insert a context node describing the block and thread
3527 * identifiers inside the kernel mark.
3528 * The context node needs to be inserted after the effective block size
3529 * has been determined such that the bounds on the thread identifiers
3530 * would reflect the effective block size.
3531 * Insert a filter node inside the context node mapping the statement
3532 * instances to block identifiers. In particular, the block identifiers
3533 * are equated to the partial schedule of band that was marked for mapping
3534 * to blocks modulo the grid size.
3535 * Insert a filter node inside the "thread" mark mapping the statement
3536 * instances to thread identifiers. In particular, the thread identifiers
3537 * are equated to the partial schedule of band that was marked for mapping
3538 * to threads modulo the block size.
3540 * Compute array reference groups for all arrays, set the local
3541 * array bounds based on the set of domain instances that reach
3542 * the kernel node, check the total amount of shared memory used
3543 * and compute all group tilings.
3544 * The array reference groups are computed after the block filter
3545 * has been inserted because it affects the mapping to shared or
3546 * private memory. This computation also requires the thread filter
3547 * (in the ppcg_kernel object), but this thread filter should not
3548 * have been added to the schedule tree yet since the computation
3549 * requires the schedule of the band that needs to be mapped to
3550 * threads before the privatization is applied.
3552 * If any array reference group requires the band mapped to threads
3553 * to be unrolled, then we perform the required unrolling.
3555 * We save a copy of the schedule that may influence the mappings
3556 * to shared or private memory in kernel->shared_schedule.
3558 * Finally, we add synchronization and copy statements to the schedule tree,
3559 * remove the "thread" mark and create representations for the local
3560 * variables in the kernel.
3562 * We keep a copy of the isl_id that points to the kernel to ensure
3563 * that the kernel does not get destroyed if the schedule node
3564 * is freed due to some error condition.
3566 static __isl_give isl_schedule_node
*create_kernel(struct gpu_gen
*gen
,
3567 __isl_take isl_schedule_node
*node
, int scale
,
3568 __isl_keep isl_multi_val
*sizes
)
3570 struct ppcg_kernel
*kernel
;
3572 isl_schedule_node
*node_thread
;
3573 isl_union_map
*host_schedule
;
3574 isl_set
*host_domain
;
3575 isl_union_set
*domain
;
3576 int single_statement
;
3578 kernel
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel
);
3579 kernel
= ppcg_kernel_create_local_arrays(kernel
, gen
->prog
);
3581 return isl_schedule_node_free(node
);
3583 domain
= isl_schedule_node_get_domain(node
);
3584 single_statement
= isl_union_set_n_set(domain
) == 1;
3586 kernel
->ctx
= gen
->ctx
;
3587 kernel
->prog
= gen
->prog
;
3588 kernel
->options
= gen
->options
;
3589 kernel
->context
= extract_context(node
, gen
->prog
);
3590 kernel
->core
= isl_union_set_universe(isl_union_set_copy(domain
));
3591 kernel
->arrays
= accessed_by_domain(isl_union_set_copy(domain
),
3593 kernel
->n_grid
= n_outer_coincidence(node
);
3594 node_thread
= isl_schedule_node_copy(node
);
3595 node_thread
= gpu_tree_move_down_to_thread(node_thread
, kernel
->core
);
3596 node_thread
= isl_schedule_node_child(node_thread
, 0);
3597 kernel
->n_block
= n_outer_coincidence(node_thread
);
3598 isl_schedule_node_free(node_thread
);
3599 kernel
->id
= gen
->kernel_id
++;
3600 read_grid_and_block_sizes(kernel
, gen
);
3602 kernel
->sync_writes
= compute_sync_writes(kernel
, node
);
3604 host_schedule
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3605 host_domain
= isl_set_from_union_set(isl_union_map_range(
3608 node
= atomic_ancestors(node
);
3610 id
= isl_id_alloc(gen
->ctx
, "kernel", kernel
);
3611 id
= isl_id_set_free_user(id
, &ppcg_kernel_free_wrap
);
3612 node
= isl_schedule_node_insert_mark(node
, isl_id_copy(id
));
3614 if (!single_statement
)
3615 node
= group_statements(node
, kernel
->id
);
3617 node
= isl_schedule_node_child(node
, 0);
3618 node
= split_band(node
, kernel
->n_grid
);
3619 kernel
->block_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3620 kernel
->n_grid
, "b");
3621 kernel
->block_filter
= set_schedule_modulo(node
, kernel
->block_ids
,
3623 kernel
->grid_size
= extract_grid_size(kernel
,
3624 isl_union_set_copy(domain
));
3625 if (!kernel
->options
->wrap
)
3626 node
= snap_band_to_sizes(node
, kernel
->grid_dim
,
3629 node
= scale_band(node
, isl_multi_val_copy(sizes
));
3630 node
= isl_schedule_node_parent(node
);
3631 if (!single_statement
)
3632 node
= isl_schedule_node_parent(node
);
3633 node
= insert_guard(node
, kernel
->context
, kernel
->grid_size
,
3635 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3636 node
= isl_schedule_node_child(node
, 0);
3637 node
= split_band(node
, kernel
->n_block
);
3638 kernel
->thread_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3639 kernel
->n_block
, "t");
3640 kernel
->thread_filter
= set_schedule_modulo(node
, kernel
->thread_ids
,
3642 extract_block_size(kernel
, domain
);
3644 node
= gpu_tree_move_up_to_kernel(node
);
3645 node
= isl_schedule_node_child(node
, 0);
3646 node
= insert_context(kernel
, node
);
3647 node
= isl_schedule_node_child(node
, 0);
3648 node
= isl_schedule_node_insert_filter(node
,
3649 isl_union_set_copy(kernel
->block_filter
));
3651 node
= gpu_tree_move_up_to_kernel(node
);
3653 if (gpu_group_references(kernel
, node
) < 0)
3654 node
= isl_schedule_node_free(node
);
3655 localize_bounds(kernel
, host_domain
);
3656 isl_set_free(host_domain
);
3658 check_shared_memory_bound(kernel
);
3659 compute_group_tilings(kernel
);
3661 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3662 node
= isl_schedule_node_child(node
, 0);
3663 if (!kernel
->options
->wrap
)
3664 node
= snap_band_to_sizes(node
, kernel
->block_dim
,
3666 node
= isl_schedule_node_insert_filter(node
,
3667 isl_union_set_copy(kernel
->thread_filter
));
3668 if (kernel_requires_unroll(kernel
)) {
3669 node
= isl_schedule_node_child(node
, 0);
3670 node
= unroll(node
);
3673 node
= gpu_tree_move_up_to_thread(node
);
3674 kernel
->shared_schedule_dim
=
3675 isl_schedule_node_get_schedule_depth(node
);
3676 kernel
->shared_schedule
=
3677 isl_schedule_node_get_prefix_schedule_union_pw_multi_aff(node
);
3679 node
= gpu_tree_move_up_to_kernel(node
);
3681 node
= add_sync(kernel
, node
);
3682 node
= add_copies(kernel
, node
);
3684 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3685 node
= isl_schedule_node_delete(node
);
3687 node
= gpu_tree_move_up_to_kernel(node
);
3689 if (create_kernel_vars(kernel
) < 0)
3690 node
= isl_schedule_node_free(node
);
3692 if (!single_statement
)
3693 node
= isl_schedule_node_parent(node
);
3694 node
= isl_schedule_node_parent(node
);
3700 /* Insert a zero-dimensional permutable band at "node".
3702 static __isl_give isl_schedule_node
*insert_empty_permutable_band(
3703 __isl_take isl_schedule_node
*node
)
3706 isl_schedule
*schedule
;
3707 isl_union_set
*domain
;
3708 isl_multi_union_pw_aff
*mupa
;
3710 schedule
= isl_schedule_node_get_schedule(node
);
3711 domain
= isl_schedule_get_domain(schedule
);
3712 space
= isl_union_set_get_space(domain
);
3713 isl_union_set_free(domain
);
3714 isl_schedule_free(schedule
);
3716 space
= isl_space_set_from_params(space
);
3717 mupa
= isl_multi_union_pw_aff_zero(space
);
3718 node
= isl_schedule_node_insert_partial_schedule(node
, mupa
);
3719 node
= isl_schedule_node_band_set_permutable(node
, 1);
3724 /* If "node" is the outermost permutable band that can be mapped to block and
3725 * thread identifiers in its branch (or a leaf with no such outer bands),
3726 * then mark the band as such, attaching a ppcg_kernel to the mark.
3728 * If "node" originally points to a leaf, then insert a zero-dimensional
3729 * permutable band such that we can assume that "node" always
3730 * points to a band node.
3732 * Tile "node" using user specified tile sizes, after splitting the band
3733 * if the number of specified tile sizes is smaller than the dimension
3734 * of the band. Mark the point band of this tiling as the band that
3735 * needs to be mapped to threads.
3736 * Create a kernel representing the domain instances that reach "node" and
3737 * insert a mark node pointing to the ppcg_kernel before the band node.
3739 static __isl_give isl_schedule_node
*mark_outer_permutable(
3740 __isl_take isl_schedule_node
*node
, void *user
)
3742 struct gpu_gen
*gen
= user
;
3748 isl_multi_val
*sizes
;
3750 outer
= is_outer_tilable(node
);
3752 return isl_schedule_node_free(node
);
3756 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
3757 node
= insert_empty_permutable_band(node
);
3759 tile_len
= isl_schedule_node_band_n_member(node
);
3760 tile_size
= read_tile_sizes(gen
, &tile_len
);
3762 return isl_schedule_node_free(node
);
3763 if (tile_len
< isl_schedule_node_band_n_member(node
))
3764 node
= isl_schedule_node_band_split(node
, tile_len
);
3765 sizes
= construct_band_tiles_sizes(node
, tile_size
);
3766 node
= tile_band(node
, isl_multi_val_copy(sizes
));
3767 node
= isl_schedule_node_child(node
, 0);
3768 id
= isl_id_alloc(gen
->ctx
, "thread", NULL
);
3769 node
= isl_schedule_node_insert_mark(node
, id
);
3770 node
= isl_schedule_node_parent(node
);
3772 scale
= gen
->options
->scale_tile_loops
;
3773 node
= create_kernel(gen
, node
, scale
, sizes
);
3774 isl_multi_val_free(sizes
);
3780 /* Does the subtree rooted at "node" have any suitably permutable band nodes?
3781 * That is, does it have any nodes that are permutable and that
3782 * have a least one coincident dimension?
3784 static int subtree_has_permutable_bands(__isl_keep isl_schedule_node
*node
)
3786 int any_parallelism
= 0;
3788 if (isl_schedule_node_foreach_descendant(node
, &set_permutable
,
3789 &any_parallelism
) < 0 &&
3793 return any_parallelism
;
3796 /* Mark all variables that are accessed by the statement instances in "domain"
3797 * and that are local to "prog" as requiring a declaration in the host code.
3799 static int declare_accessed_local_variables(struct gpu_prog
*prog
,
3800 __isl_keep isl_union_set
*domain
)
3802 isl_union_set
*arrays
;
3805 if (!ppcg_scop_any_hidden_declarations(prog
->scop
))
3807 arrays
= accessed_by_domain(isl_union_set_copy(domain
), prog
);
3809 for (i
= 0; i
< prog
->n_array
; ++i
) {
3814 if (!prog
->array
[i
].local
)
3816 space
= isl_set_get_space(prog
->array
[i
].extent
);
3817 set
= isl_union_set_extract_set(arrays
, space
);
3818 empty
= isl_set_plain_is_empty(set
);
3823 prog
->array
[i
].declare_local
= 1;
3826 isl_union_set_free(arrays
);
3829 isl_union_set_free(arrays
);
3833 /* If "node" points to a set node, then separate its children
3834 * into subtrees that have suitably permutable bands and
3835 * those that do not.
3836 * Adjust the schedule tree in order to execute the second group
3837 * after the first group and return a pointer to the first group,
3838 * assuming there are any such subtrees.
3839 * Mark all local variables in "prog" that are accessed by
3840 * the second group as requiring a declaration on the host.
3842 static __isl_give isl_schedule_node
*isolate_permutable_subtrees(
3843 __isl_take isl_schedule_node
*node
, struct gpu_prog
*prog
)
3846 isl_union_set
*filter
;
3851 if (isl_schedule_node_get_type(node
) != isl_schedule_node_set
)
3854 n
= isl_schedule_node_n_children(node
);
3856 return isl_schedule_node_free(node
);
3858 node
= isl_schedule_node_child(node
, 0);
3859 filter
= isl_schedule_node_filter_get_filter(node
);
3860 node
= isl_schedule_node_parent(node
);
3861 space
= isl_union_set_get_space(filter
);
3862 isl_union_set_free(filter
);
3863 filter
= isl_union_set_empty(space
);
3865 for (i
= 0; i
< n
; ++i
) {
3868 node
= isl_schedule_node_child(node
, i
);
3869 parallelism
= subtree_has_permutable_bands(node
);
3870 if (parallelism
< 0) {
3871 node
= isl_schedule_node_free(node
);
3872 } else if (!parallelism
) {
3873 isl_union_set
*filter_i
;
3874 filter_i
= isl_schedule_node_filter_get_filter(node
);
3875 filter
= isl_union_set_union(filter
, filter_i
);
3877 node
= isl_schedule_node_parent(node
);
3880 if (declare_accessed_local_variables(prog
, filter
) < 0)
3881 node
= isl_schedule_node_free(node
);
3882 node
= isl_schedule_node_order_after(node
, filter
);
3887 /* Replace any reference to an array element in the range of "copy"
3888 * by a reference to all array elements (defined by the extent of the array).
3890 static __isl_give isl_union_map
*approximate_copy_out(
3891 __isl_take isl_union_map
*copy
, struct gpu_prog
*prog
)
3896 res
= isl_union_map_empty(isl_union_map_get_space(copy
));
3898 for (i
= 0; i
< prog
->n_array
; ++i
) {
3901 isl_union_map
*copy_i
;
3902 isl_union_set
*extent
, *domain
;
3904 space
= isl_space_copy(prog
->array
[i
].space
);
3905 extent
= isl_union_set_from_set(isl_set_universe(space
));
3906 copy_i
= isl_union_map_copy(copy
);
3907 copy_i
= isl_union_map_intersect_range(copy_i
, extent
);
3908 set
= isl_set_copy(prog
->array
[i
].extent
);
3909 extent
= isl_union_set_from_set(set
);
3910 domain
= isl_union_map_domain(copy_i
);
3911 copy_i
= isl_union_map_from_domain_and_range(domain
, extent
);
3912 res
= isl_union_map_union(res
, copy_i
);
3915 isl_union_map_free(copy
);
3920 /* Insert "kernel" marks that point to a ppcg_kernel structure
3921 * in front of all outermost tilable band that (by construction)
3922 * have at least one parallel loop.
3924 static __isl_give isl_schedule_node
*mark_kernels(struct gpu_gen
*gen
,
3925 __isl_take isl_schedule_node
*node
)
3927 return isl_schedule_node_map_descendant(node
,
3928 &mark_outer_permutable
, gen
);
3931 /* Save the schedule "schedule" to a file called "filename".
3932 * The schedule is printed in block style.
3934 static void save_schedule(__isl_keep isl_schedule
*schedule
,
3935 const char *filename
)
3944 file
= fopen(filename
, "w");
3946 fprintf(stderr
, "Unable to open '%s' for writing\n", filename
);
3949 ctx
= isl_schedule_get_ctx(schedule
);
3950 p
= isl_printer_to_file(ctx
, file
);
3951 p
= isl_printer_set_yaml_style(p
, ISL_YAML_STYLE_BLOCK
);
3952 p
= isl_printer_print_schedule(p
, schedule
);
3953 isl_printer_free(p
);
3957 /* Load and return a schedule from a file called "filename".
3959 static __isl_give isl_schedule
*load_schedule(isl_ctx
*ctx
,
3960 const char *filename
)
3963 isl_schedule
*schedule
;
3965 file
= fopen(filename
, "r");
3967 fprintf(stderr
, "Unable to open '%s' for reading\n", filename
);
3970 schedule
= isl_schedule_read_from_file(ctx
, file
);
3976 /* Compute an appropriate schedule based on the accesses in
3977 * gen->read and gen->write.
3979 * We use the dependences in gen->prog->scop to compute
3980 * a schedule that has a parallel loop in each tilable band and
3981 * return this schedule.
3983 * If live range reordering is allowed, then we need to make sure
3984 * that live ranges on arrays are not run in parallel since doing
3985 * so would require array expansion. We therefore add the array
3986 * order dependences to the coincidence dependences. Non-zero array
3987 * order dependences will then prevent a schedule dimension from being
3988 * considered parallel.
3989 * Live ranges derived from scalars are allowed to be run in parallel
3990 * since we force the scalars to be mapped to private memory in
3991 * check_scalar_live_ranges.
3992 * If live range reordering is allowed, then the false dependences
3993 * are not added to the validity constraints as that would prevent
3994 * reordering. Instead, the external false dependences that enforce that reads
3995 * from potentially live-in data precede any later write and
3996 * that writes of potentially live-out data follow any other earlier write
3997 * are added to the validity and the coincidence constraints.
3998 * The false dependences are still added to the proximity constraints
3999 * for consistency with the case where live range reordering is not allowed.
4000 * The coincidence constraints then consist of flow dependences,
4001 * external false dependences and array order dependences.
4002 * The independences can be filtered out from the first two sets.
4003 * They have already been filtered out from the array order dependences
4004 * on a per array basis in collect_order_dependences.
4005 * There is no need for a per array handling of the other two sets
4006 * as there should be no flow or external false dependence on local
4007 * variables that can be filtered out.
4009 static __isl_give isl_schedule
*compute_schedule(struct gpu_gen
*gen
)
4011 isl_union_set
*domain
;
4012 isl_union_map
*dep_raw
, *dep
;
4013 isl_union_map
*validity
, *proximity
, *coincidence
;
4014 isl_schedule_constraints
*sc
;
4015 isl_schedule
*schedule
;
4017 domain
= isl_union_set_copy(gen
->prog
->scop
->domain
);
4018 sc
= isl_schedule_constraints_on_domain(domain
);
4019 sc
= isl_schedule_constraints_set_context(sc
,
4020 isl_set_copy(gen
->prog
->scop
->context
));
4021 if (gen
->options
->live_range_reordering
) {
4022 sc
= isl_schedule_constraints_set_conditional_validity(sc
,
4023 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_flow
),
4024 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_order
));
4025 proximity
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
4026 validity
= isl_union_map_copy(proximity
);
4027 validity
= isl_union_map_union(validity
,
4028 isl_union_map_copy(gen
->prog
->scop
->dep_forced
));
4029 proximity
= isl_union_map_union(proximity
,
4030 isl_union_map_copy(gen
->prog
->scop
->dep_false
));
4031 coincidence
= isl_union_map_copy(validity
);
4032 coincidence
= isl_union_map_subtract(coincidence
,
4033 isl_union_map_copy(gen
->prog
->scop
->independence
));
4034 coincidence
= isl_union_map_union(coincidence
,
4035 isl_union_map_copy(gen
->prog
->array_order
));
4037 dep_raw
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
4038 dep
= isl_union_map_copy(gen
->prog
->scop
->dep_false
);
4039 dep
= isl_union_map_union(dep
, dep_raw
);
4040 dep
= isl_union_map_coalesce(dep
);
4041 proximity
= isl_union_map_copy(dep
);
4042 coincidence
= isl_union_map_copy(dep
);
4045 sc
= isl_schedule_constraints_set_validity(sc
, validity
);
4046 sc
= isl_schedule_constraints_set_coincidence(sc
, coincidence
);
4047 sc
= isl_schedule_constraints_set_proximity(sc
, proximity
);
4049 if (gen
->options
->debug
->dump_schedule_constraints
)
4050 isl_schedule_constraints_dump(sc
);
4051 schedule
= isl_schedule_constraints_compute_schedule(sc
);
4056 /* Obtain a schedule for the scop, either by reading it from
4057 * a file or by computing one.
4059 static __isl_give isl_schedule
*get_schedule(struct gpu_gen
*gen
)
4061 isl_schedule
*schedule
;
4063 if (gen
->options
->load_schedule_file
) {
4064 schedule
= load_schedule(gen
->ctx
,
4065 gen
->options
->load_schedule_file
);
4067 schedule
= compute_schedule(gen
);
4068 if (gen
->options
->save_schedule_file
)
4069 save_schedule(schedule
,
4070 gen
->options
->save_schedule_file
);
4072 if (gen
->options
->debug
->dump_schedule
)
4073 isl_schedule_dump(schedule
);
4078 /* Construct the string "<a>_<b>".
4080 static char *concat(isl_ctx
*ctx
, const char *a
, const char *b
)
4085 p
= isl_printer_to_str(ctx
);
4086 p
= isl_printer_print_str(p
, a
);
4087 p
= isl_printer_print_str(p
, "_");
4088 p
= isl_printer_print_str(p
, b
);
4089 s
= isl_printer_get_str(p
);
4090 isl_printer_free(p
);
4095 /* For each array in "prog" of which an element appears in "accessed" and
4096 * that is not a read only scalar, create a zero-dimensional universe set
4097 * of which the tuple id has name "<prefix>_<name of array>" and a user
4098 * pointer pointing to the array (gpu_array_info).
4100 * If the array is local to "prog", then make sure it will be declared
4103 * Return the list of these universe sets.
4105 static __isl_give isl_union_set_list
*create_copy_filters(struct gpu_prog
*prog
,
4106 const char *prefix
, __isl_take isl_union_set
*accessed
)
4110 isl_union_set_list
*filters
;
4113 filters
= isl_union_set_list_alloc(ctx
, 0);
4114 for (i
= 0; i
< prog
->n_array
; ++i
) {
4115 struct gpu_array_info
*array
= &prog
->array
[i
];
4117 isl_set
*accessed_i
;
4121 isl_union_set
*uset
;
4123 if (gpu_array_is_read_only_scalar(array
))
4126 space
= isl_space_copy(array
->space
);
4127 accessed_i
= isl_union_set_extract_set(accessed
, space
);
4128 empty
= isl_set_plain_is_empty(accessed_i
);
4129 isl_set_free(accessed_i
);
4131 filters
= isl_union_set_list_free(filters
);
4138 array
->declare_local
= 1;
4140 name
= concat(ctx
, prefix
, array
->name
);
4141 id
= name
? isl_id_alloc(ctx
, name
, array
) : NULL
;
4143 space
= isl_space_set_alloc(ctx
, 0, 0);
4144 space
= isl_space_set_tuple_id(space
, isl_dim_set
, id
);
4145 uset
= isl_union_set_from_set(isl_set_universe(space
));
4147 filters
= isl_union_set_list_add(filters
, uset
);
4149 isl_union_set_free(accessed
);
4154 /* Make sure that code for the statements in "filters" that
4155 * copy arrays to or from the device is only generated when
4156 * the size of the corresponding array is positive.
4157 * That is, add a set node underneath "graft" with "filters" as children
4158 * and for each child add a guard that the selects the parameter
4159 * values for which the corresponding array has a positive size.
4160 * The array is available in the user pointer of the statement identifier.
4161 * "depth" is the schedule depth of the position where "graft"
4164 static __isl_give isl_schedule_node
*insert_positive_size_guards(
4165 __isl_take isl_schedule_node
*graft
,
4166 __isl_take isl_union_set_list
*filters
, int depth
)
4170 graft
= isl_schedule_node_child(graft
, 0);
4171 graft
= isl_schedule_node_insert_set(graft
, filters
);
4172 n
= isl_schedule_node_n_children(graft
);
4173 for (i
= 0; i
< n
; ++i
) {
4174 isl_union_set
*filter
;
4175 isl_set
*domain
, *guard
;
4177 struct gpu_array_info
*array
;
4179 graft
= isl_schedule_node_child(graft
, i
);
4180 filter
= isl_schedule_node_filter_get_filter(graft
);
4181 domain
= isl_set_from_union_set(filter
);
4182 id
= isl_set_get_tuple_id(domain
);
4183 array
= isl_id_get_user(id
);
4185 isl_set_free(domain
);
4186 guard
= gpu_array_positive_size_guard(array
);
4187 guard
= isl_set_from_params(guard
);
4188 guard
= isl_set_add_dims(guard
, isl_dim_set
, depth
);
4189 graft
= isl_schedule_node_child(graft
, 0);
4190 graft
= isl_schedule_node_insert_guard(graft
, guard
);
4191 graft
= isl_schedule_node_parent(graft
);
4192 graft
= isl_schedule_node_parent(graft
);
4194 graft
= isl_schedule_node_parent(graft
);
4199 /* Create a graft for copying arrays to or from the device,
4200 * whenever the size of the array is strictly positive.
4201 * Each statement is called "<prefix>_<name of array>" and
4202 * the identifier has a user pointer pointing to the array.
4203 * The graft will be added at the position specified by "node".
4204 * "copy" contains the array elements that need to be copied.
4205 * Only arrays of which some elements need to be copied
4206 * will have a corresponding statement in the graph.
4207 * Note though that each such statement will copy the entire array.
4209 static __isl_give isl_schedule_node
*create_copy_device(struct gpu_prog
*prog
,
4210 __isl_keep isl_schedule_node
*node
, const char *prefix
,
4211 __isl_take isl_union_set
*copy
)
4216 isl_union_set
*all
, *domain
;
4217 isl_union_set_list
*filters
;
4218 isl_union_map
*extension
;
4219 isl_schedule_node
*graft
;
4222 depth
= isl_schedule_node_get_schedule_depth(node
);
4223 filters
= create_copy_filters(prog
, prefix
, copy
);
4224 all
= isl_union_set_list_union(isl_union_set_list_copy(filters
));
4226 space
= depth
< 0 ? NULL
: isl_space_set_alloc(ctx
, 0, depth
);
4227 domain
= isl_union_set_from_set(isl_set_universe(space
));
4228 extension
= isl_union_map_from_domain_and_range(domain
, all
);
4229 graft
= isl_schedule_node_from_extension(extension
);
4232 return isl_schedule_node_free(graft
);
4233 if (isl_union_set_list_n_union_set(filters
) == 0) {
4234 isl_union_set_list_free(filters
);
4238 return insert_positive_size_guards(graft
, filters
, depth
);
4241 /* Return (the universe spaces of) the arrays that are declared
4242 * inside the scop corresponding to "prog" and for which all
4243 * potential writes inside the scop form a subset of "domain".
4245 static __isl_give isl_union_set
*extract_local_accesses(struct gpu_prog
*prog
,
4246 __isl_keep isl_union_set
*domain
)
4249 isl_union_set
*local
;
4251 local
= isl_union_set_empty(isl_union_set_get_space(domain
));
4253 for (i
= 0; i
< prog
->n_array
; ++i
) {
4255 isl_union_map
*to_outer
;
4256 isl_union_map
*may_write
;
4257 isl_union_set
*write_domain
;
4258 isl_union_set
*fields
;
4261 if (!prog
->array
[i
].local
)
4264 set
= isl_set_universe(isl_space_copy(prog
->array
[i
].space
));
4265 to_outer
= isl_union_map_copy(prog
->to_outer
);
4266 to_outer
= isl_union_map_intersect_range(to_outer
,
4267 isl_union_set_from_set(isl_set_copy(set
)));
4268 fields
= isl_union_map_domain(to_outer
);
4269 may_write
= isl_union_map_copy(prog
->may_write
);
4270 may_write
= isl_union_map_intersect_range(may_write
, fields
);
4271 write_domain
= isl_union_map_domain(may_write
);
4272 subset
= isl_union_set_is_subset(write_domain
, domain
);
4273 isl_union_set_free(write_domain
);
4277 return isl_union_set_free(local
);
4278 } else if (subset
) {
4279 local
= isl_union_set_add_set(local
, set
);
4288 /* Internal data structure for node_may_persist.
4290 * "tagger" maps tagged iteration domains to the corresponding untagged
4293 * "may_persist_flow" is the set of all tagged dataflow dependences
4294 * with those dependences removed that either precede or follow
4295 * the kernel launch in a sequence.
4296 * "inner_band_flow" is the set of all tagged dataflow dependences
4297 * that are local to a given iteration of the outer band nodes
4298 * with respect to the current node.
4299 * "local_flow" is equal to "inner_band_flow", except that the domain
4300 * and the range have been intersected with intermediate filters
4301 * on children of sets or sequences.
4303 struct ppcg_may_persist_data
{
4304 isl_union_pw_multi_aff
*tagger
;
4306 isl_union_map
*local_flow
;
4307 isl_union_map
*inner_band_flow
;
4308 isl_union_map
*may_persist_flow
;
4311 /* Update the information in "data" based on the band ancestor "node".
4313 * In particular, we restrict the dependences in data->local_flow
4314 * to those dependence where the source and the sink occur in
4315 * the same iteration of the given band node.
4316 * We also update data->inner_band_flow to the new value of
4319 static int update_may_persist_at_band(__isl_keep isl_schedule_node
*node
,
4320 struct ppcg_may_persist_data
*data
)
4322 isl_multi_union_pw_aff
*partial
;
4323 isl_union_pw_multi_aff
*contraction
;
4324 isl_union_map
*flow
;
4326 if (isl_schedule_node_band_n_member(node
) == 0)
4329 partial
= isl_schedule_node_band_get_partial_schedule(node
);
4330 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4331 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4333 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4334 isl_union_pw_multi_aff_copy(data
->tagger
));
4336 flow
= data
->local_flow
;
4337 flow
= isl_union_map_eq_at_multi_union_pw_aff(flow
, partial
);
4338 data
->local_flow
= flow
;
4340 isl_union_map_free(data
->inner_band_flow
);
4341 data
->inner_band_flow
= isl_union_map_copy(data
->local_flow
);
4346 /* Given a set of local reaching domain elements "domain",
4347 * expand them to the corresponding leaf domain elements using "contraction"
4348 * and insert the array references tags using data->tagger.
4350 static __isl_give isl_union_set
*expand_and_tag(
4351 __isl_take isl_union_set
*domain
,
4352 __isl_take isl_union_pw_multi_aff
*contraction
,
4353 struct ppcg_may_persist_data
*data
)
4355 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4357 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4358 isl_union_pw_multi_aff_copy(data
->tagger
));
4362 /* Given a filter node that is the child of a set or sequence node,
4363 * restrict data->local_flow to refer only to those elements
4364 * in the filter of the node.
4365 * "contraction" maps the leaf domain elements of the schedule tree
4366 * to the corresponding domain elements at (the parent of) "node".
4368 static int filter_flow(__isl_keep isl_schedule_node
*node
,
4369 struct ppcg_may_persist_data
*data
,
4370 __isl_take isl_union_pw_multi_aff
*contraction
)
4372 isl_union_set
*filter
;
4373 isl_union_map
*flow
;
4375 flow
= data
->local_flow
;
4376 filter
= isl_schedule_node_filter_get_filter(node
);
4377 filter
= expand_and_tag(filter
, contraction
, data
);
4378 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(filter
));
4379 flow
= isl_union_map_intersect_range(flow
, filter
);
4380 data
->local_flow
= flow
;
4385 /* Given a filter node "node", collect the filters on all preceding siblings
4386 * (which are also filter nodes), add them to "filters" and return the result.
4388 static __isl_give isl_union_set
*add_previous_filters(
4389 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4391 isl_schedule_node
*sibling
;
4393 sibling
= isl_schedule_node_copy(node
);
4394 while (sibling
&& isl_schedule_node_has_previous_sibling(sibling
)) {
4395 isl_union_set
*filter
;
4397 sibling
= isl_schedule_node_previous_sibling(sibling
);
4398 filter
= isl_schedule_node_filter_get_filter(sibling
);
4399 filters
= isl_union_set_union(filters
, filter
);
4401 isl_schedule_node_free(sibling
);
4403 return isl_union_set_free(filters
);
4408 /* Given a filter node "node", collect the filters on all following siblings
4409 * (which are also filter nodes), add them to "filters" and return the result.
4411 static __isl_give isl_union_set
*add_next_filters(
4412 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4414 isl_schedule_node
*sibling
;
4416 sibling
= isl_schedule_node_copy(node
);
4417 while (sibling
&& isl_schedule_node_has_next_sibling(sibling
)) {
4418 isl_union_set
*filter
;
4420 sibling
= isl_schedule_node_next_sibling(sibling
);
4421 filter
= isl_schedule_node_filter_get_filter(sibling
);
4422 filters
= isl_union_set_union(filters
, filter
);
4424 isl_schedule_node_free(sibling
);
4426 return isl_union_set_free(filters
);
4431 /* Remove those flow dependences from data->may_persist_flow
4432 * that flow between elements of "domain" within the same iteration
4433 * of all outer band nodes.
4434 * "contraction" maps the leaf domain elements of the schedule tree
4435 * to the corresponding elements "domain".
4437 static void remove_external_flow(struct ppcg_may_persist_data
*data
,
4438 __isl_take isl_union_set
*domain
,
4439 __isl_keep isl_union_pw_multi_aff
*contraction
)
4441 isl_union_map
*flow
;
4443 contraction
= isl_union_pw_multi_aff_copy(contraction
);
4444 domain
= expand_and_tag(domain
, contraction
, data
);
4445 flow
= isl_union_map_copy(data
->local_flow
);
4446 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(domain
));
4447 flow
= isl_union_map_intersect_range(flow
, domain
);
4449 data
->may_persist_flow
= isl_union_map_subtract(data
->may_persist_flow
,
4453 /* Update the information in "data" based on the filter ancestor "node".
4454 * We only need to modify anything if the filter is the child
4455 * of a set or sequence node.
4457 * In the case of a sequence, we remove the dependences between
4458 * statement instances that are both executed either before or
4459 * after the subtree that will be mapped to a kernel, within
4460 * the same iteration of outer bands.
4462 * In both cases, we restrict data->local_flow to the current child.
4464 static int update_may_persist_at_filter(__isl_keep isl_schedule_node
*node
,
4465 struct ppcg_may_persist_data
*data
)
4467 enum isl_schedule_node_type type
;
4468 isl_schedule_node
*parent
;
4470 isl_union_pw_multi_aff
*contraction
;
4471 isl_union_set
*before
, *after
, *filter
;
4472 isl_union_map
*flow
;
4474 type
= isl_schedule_node_get_parent_type(node
);
4475 if (type
!= isl_schedule_node_sequence
&& type
!= isl_schedule_node_set
)
4478 parent
= isl_schedule_node_copy(node
);
4479 parent
= isl_schedule_node_parent(parent
);
4480 contraction
= isl_schedule_node_get_subtree_contraction(parent
);
4481 isl_schedule_node_free(parent
);
4483 if (type
== isl_schedule_node_set
)
4484 return filter_flow(node
, data
, contraction
);
4486 filter
= isl_schedule_node_filter_get_filter(node
);
4487 space
= isl_union_set_get_space(filter
);
4488 isl_union_set_free(filter
);
4489 before
= isl_union_set_empty(space
);
4490 after
= isl_union_set_copy(before
);
4491 before
= add_previous_filters(before
, node
);
4492 after
= add_next_filters(after
, node
);
4494 remove_external_flow(data
, before
, contraction
);
4495 remove_external_flow(data
, after
, contraction
);
4497 return filter_flow(node
, data
, contraction
);
4500 /* Update the information in "data" based on the ancestor "node".
4502 static int update_may_persist_at(__isl_keep isl_schedule_node
*node
, void *user
)
4504 struct ppcg_may_persist_data
*data
= user
;
4506 switch (isl_schedule_node_get_type(node
)) {
4507 case isl_schedule_node_error
:
4509 case isl_schedule_node_context
:
4510 case isl_schedule_node_domain
:
4511 case isl_schedule_node_expansion
:
4512 case isl_schedule_node_extension
:
4513 case isl_schedule_node_guard
:
4514 case isl_schedule_node_leaf
:
4515 case isl_schedule_node_mark
:
4516 case isl_schedule_node_sequence
:
4517 case isl_schedule_node_set
:
4519 case isl_schedule_node_band
:
4520 if (update_may_persist_at_band(node
, data
) < 0)
4523 case isl_schedule_node_filter
:
4524 if (update_may_persist_at_filter(node
, data
) < 0)
4532 /* Determine the set of array elements that may need to be perserved
4533 * by a kernel constructed from the subtree at "node".
4534 * This includes the set of array elements that may need to be preserved
4535 * by the entire scop (prog->may_persist) and the elements for which
4536 * there is a potential flow dependence that may cross a kernel launch.
4538 * To determine the second set, we start from all flow dependences.
4539 * From this set of dependences, we remove those that cannot possibly
4540 * require data to be preserved by a kernel launch.
4541 * In particular, we consider the following sets of dependences.
4542 * - dependences of which the write occurs inside the kernel.
4543 * If the data is needed outside the kernel, then it will
4544 * be copied out immediately after the kernel launch, so there
4545 * is no need for any special care.
4546 * - dependences of which the read occurs inside the kernel and the
4547 * corresponding write occurs inside the same iteration of the
4548 * outer band nodes. This means that the data is needed in
4549 * the first kernel launch after the write, which is already
4550 * taken care of by the standard copy-in. That is, the data
4551 * do not need to be preserved by any intermediate call to
4553 * - dependences of which the write and the read either both occur
4554 * before the kernel launch or both occur after the kernel launch,
4555 * within the same iteration of the outer band nodes with respect
4556 * to the sequence that determines the ordering of the dependence
4557 * and the kernel launch. Such flow dependences cannot cross
4558 * any kernel launch.
4560 * For the remaining (tagged) dependences, we take the domain
4561 * (i.e., the tagged writes) and apply the tagged access relation
4562 * to obtain the accessed data elements.
4563 * These are then combined with the elements that may need to be
4564 * preserved by the entire scop.
4566 static __isl_give isl_union_set
*node_may_persist(
4567 __isl_keep isl_schedule_node
*node
, struct gpu_prog
*prog
)
4569 struct ppcg_may_persist_data data
;
4570 isl_schedule_node
*root
;
4571 isl_union_pw_multi_aff
*contraction
;
4572 isl_union_set
*domain
;
4573 isl_union_set
*persist
;
4574 isl_union_map
*flow
, *local_flow
;
4576 data
.tagger
= prog
->scop
->tagger
;
4578 flow
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
4579 data
.local_flow
= isl_union_map_copy(flow
);
4580 data
.inner_band_flow
= isl_union_map_copy(flow
);
4581 data
.may_persist_flow
= flow
;
4582 if (isl_schedule_node_foreach_ancestor_top_down(node
,
4583 &update_may_persist_at
, &data
) < 0)
4584 data
.may_persist_flow
=
4585 isl_union_map_free(data
.may_persist_flow
);
4586 flow
= data
.may_persist_flow
;
4587 isl_union_map_free(data
.local_flow
);
4589 domain
= isl_schedule_node_get_domain(node
);
4590 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4591 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4593 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4594 isl_union_pw_multi_aff_copy(data
.tagger
));
4595 flow
= isl_union_map_subtract_domain(flow
, isl_union_set_copy(domain
));
4596 local_flow
= data
.inner_band_flow
;
4597 local_flow
= isl_union_map_intersect_range(local_flow
, domain
);
4598 flow
= isl_union_map_subtract(flow
, local_flow
);
4600 persist
= isl_union_map_domain(flow
);
4601 persist
= isl_union_set_apply(persist
,
4602 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4603 persist
= isl_union_set_union(persist
,
4604 isl_union_set_copy(prog
->may_persist
));
4609 /* Add nodes for copying outer arrays in and out of the device
4610 * before and after the subtree "node", which contains one or more kernels.
4611 * "domain" contains the original reaching domain elements before
4612 * the kernels were created, i.e., before the contraction that
4613 * may have been performed in creating the kernels has been applied.
4614 * "prefix" contains the prefix schedule at that point, in terms
4615 * of the same original reaching domain elements.
4617 * We first compute the sets of outer array elements that need
4618 * to be copied in and out and then graft in the nodes for
4619 * performing this copying.
4621 * In particular, for each array that is possibly written anywhere in
4622 * the subtree "node" and that may be used after "node"
4623 * or that may be visible outside the corresponding scop,
4624 * we copy out its entire extent.
4626 * Any array elements that is read without first being written inside
4627 * the subtree "node" needs to be copied in.
4628 * Furthermore, if there are any array elements that
4629 * are copied out, but that may not be written inside "node, then
4630 * they also need to be copied in to ensure that the value after execution
4631 * is the same as the value before execution, at least for those array
4632 * elements that may have their values preserved by the scop or that
4633 * may be written before "node" and read after "node".
4634 * In case the array elements are structures, we need to take into
4635 * account that all members of the structures need to be written
4636 * by "node" before we can avoid copying the data structure in.
4638 * Note that the may_write relation is intersected with the domain,
4639 * which has been intersected with the context.
4640 * This helps in those cases where the arrays are declared with a fixed size,
4641 * while the accesses are parametric and the context assigns a fixed value
4642 * to the parameters.
4644 * If an element from a local array is read without first being written,
4645 * then there is no point in copying it in since it cannot have been
4646 * written prior to the scop. Warn about the uninitialized read instead.
4648 static __isl_give isl_schedule_node
*add_to_from_device(
4649 __isl_take isl_schedule_node
*node
, __isl_take isl_union_set
*domain
,
4650 __isl_take isl_union_map
*prefix
, struct gpu_prog
*prog
)
4652 isl_union_set
*local
;
4653 isl_union_set
*to_device
, *from_device
, *may_persist
;
4654 isl_union_map
*may_write
, *must_write
, *copy_out
, *not_written
;
4655 isl_union_map
*read
, *copy_in
;
4656 isl_union_map
*tagged
;
4657 isl_union_map
*local_uninitialized
;
4658 isl_schedule_node
*graft
;
4660 tagged
= isl_union_map_copy(prog
->scop
->tagged_reads
);
4661 tagged
= isl_union_map_union(tagged
,
4662 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4664 may_write
= isl_union_map_copy(prog
->may_write
);
4665 may_write
= isl_union_map_intersect_domain(may_write
,
4666 isl_union_set_copy(domain
));
4667 may_write
= remove_local_accesses(prog
,
4668 isl_union_map_copy(tagged
), may_write
,
4669 isl_union_map_copy(prefix
), 0);
4670 may_write
= isl_union_map_apply_range(may_write
,
4671 isl_union_map_copy(prog
->to_outer
));
4672 may_write
= isl_union_map_apply_domain(may_write
,
4673 isl_union_map_copy(prefix
));
4674 may_write
= approximate_copy_out(may_write
, prog
);
4675 copy_out
= isl_union_map_copy(may_write
);
4676 may_write
= isl_union_map_apply_range(may_write
,
4677 isl_union_map_copy(prog
->to_inner
));
4678 must_write
= isl_union_map_copy(prog
->must_write
);
4679 must_write
= isl_union_map_apply_domain(must_write
,
4680 isl_union_map_copy(prefix
));
4681 may_persist
= node_may_persist(node
, prog
);
4682 may_write
= isl_union_map_intersect_range(may_write
, may_persist
);
4683 not_written
= isl_union_map_subtract(may_write
, must_write
);
4685 local
= extract_local_accesses(prog
, domain
);
4686 read
= isl_union_map_copy(prog
->read
);
4687 read
= isl_union_map_intersect_domain(read
, domain
);
4688 read
= remove_local_accesses(prog
, tagged
, read
,
4689 isl_union_map_copy(prefix
), 1);
4690 local
= isl_union_set_apply(local
, isl_union_map_copy(prog
->to_inner
));
4691 local_uninitialized
= isl_union_map_copy(prog
->scop
->live_in
);
4692 local_uninitialized
= isl_union_map_intersect_range(local_uninitialized
,
4694 local_uninitialized
= isl_union_map_intersect(local_uninitialized
,
4695 isl_union_map_copy(read
));
4696 if (!isl_union_map_is_empty(local_uninitialized
)) {
4698 "possibly uninitialized reads (not copied in):\n");
4699 isl_union_map_dump(local_uninitialized
);
4701 read
= isl_union_map_subtract(read
, local_uninitialized
);
4702 read
= isl_union_map_apply_domain(read
, prefix
);
4703 copy_in
= isl_union_map_union(read
, not_written
);
4704 copy_in
= isl_union_map_apply_range(copy_in
,
4705 isl_union_map_copy(prog
->to_outer
));
4707 graft
= create_copy_device(prog
, node
, "to_device",
4708 isl_union_map_range(copy_in
));
4709 node
= isl_schedule_node_graft_before(node
, graft
);
4710 graft
= create_copy_device(prog
, node
, "from_device",
4711 isl_union_map_range(copy_out
));
4712 node
= isl_schedule_node_graft_after(node
, graft
);
4717 /* Update "schedule" for mapping to a GPU device.
4719 * In particular, insert a context node, create kernels for
4720 * each outermost tilable band and introduce node for copying array
4721 * in and out of the device.
4722 * If the child of the initial root points to a set node,
4723 * then children of this node that do not contain any tilable bands
4724 * are separated from the other children and are not mapped to
4727 static __isl_give isl_schedule
*map_to_device(struct gpu_gen
*gen
,
4728 __isl_take isl_schedule
*schedule
)
4730 isl_schedule_node
*node
;
4732 isl_union_set
*domain
;
4733 isl_union_map
*prefix
;
4735 context
= isl_set_copy(gen
->prog
->context
);
4736 context
= isl_set_from_params(context
);
4737 schedule
= isl_schedule_insert_context(schedule
, context
);
4739 node
= isl_schedule_get_root(schedule
);
4740 isl_schedule_free(schedule
);
4741 node
= isl_schedule_node_child(node
, 0);
4742 if (isl_schedule_node_get_type(node
) == isl_schedule_node_context
)
4743 node
= isl_schedule_node_child(node
, 0);
4744 node
= isolate_permutable_subtrees(node
, gen
->prog
);
4745 domain
= isl_schedule_node_get_domain(node
);
4746 prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
4747 node
= mark_kernels(gen
, node
);
4748 node
= add_to_from_device(node
, domain
, prefix
, gen
->prog
);
4749 schedule
= isl_schedule_node_get_schedule(node
);
4750 isl_schedule_node_free(node
);
4755 /* Internal data structure for extract_access.
4756 * "next_access" points to the end of a linked list that is extended
4757 * by extract_access.
4758 * "single_expression" is set if the access expressions belong to
4759 * an expression statement (i.e., a statement without internal control).
4760 * "any_to_outer" maps all intermediate arrays to their outer arrays.
4762 struct ppcg_extract_access_data
{
4763 struct gpu_stmt_access
**next_access
;
4764 int single_expression
;
4765 isl_union_map
*any_to_outer
;
4768 /* Given a tagged access relation to a single array "tagged", extract it
4769 * as a map, taking into account that the input may be empty.
4770 * If the access relation is empty, then it does not contain
4771 * any space information, so we try to recover it from the index
4773 * The space of the index expression is of the form I -> A,
4774 * with I the statement instances and A the array, or [I -> F] -> A,
4775 * with F the filters corresponding to arguments.
4776 * We first drop F, if present, obtaining I -> A.
4777 * Then we construct I -> R, with R the reference tag,
4778 * combine the two into I -> [R -> A] and uncurry to obtain
4779 * the final result [I -> R] -> A.
4780 * Note that the index expression may have a lower dimension
4781 * than that of the array, but this dimension is not used
4782 * if the access relation is empty.
4784 static __isl_give isl_map
*extract_single_tagged_access(
4785 __isl_take isl_union_map
*tagged
, __isl_keep pet_expr
*expr
)
4789 isl_space
*space
, *space2
;
4790 isl_multi_pw_aff
*index
;
4792 empty
= isl_union_map_is_empty(tagged
);
4796 return isl_map_from_union_map(tagged
);
4797 isl_union_map_free(tagged
);
4799 index
= pet_expr_access_get_index(expr
);
4800 space
= isl_multi_pw_aff_get_space(index
);
4801 isl_multi_pw_aff_free(index
);
4802 if (isl_space_domain_is_wrapping(space
))
4803 space
= isl_space_domain_factor_domain(space
);
4804 space2
= isl_space_copy(space
);
4805 space2
= isl_space_from_domain(isl_space_domain(space
));
4806 id
= pet_expr_access_get_ref_id(expr
);
4807 space2
= isl_space_set_tuple_id(space2
, isl_dim_out
, id
);
4808 space
= isl_space_range_product(space2
, space
);
4809 space
= isl_space_uncurry(space
);
4811 return isl_map_empty(space
);
4813 isl_union_map_free(tagged
);
4817 /* Extract a gpu_stmt_access from "expr", append it to the list
4818 * that ends in *data->next_access and update the end of the list.
4819 * If the access expression performs a write, then it is considered
4820 * exact only if it appears in a single expression statement and
4821 * if its may access relation is equal to its must access relation.
4823 * The combined set of may accesses may be union if member accesses
4824 * are involved, but the entire set is derived from a single reference and
4825 * therefore from a single index expression. These accesses therefore
4826 * all map to the same outer array.
4828 static int extract_access(__isl_keep pet_expr
*expr
, void *user
)
4830 struct ppcg_extract_access_data
*data
= user
;
4831 isl_union_map
*tagged
;
4832 struct gpu_stmt_access
*access
;
4833 isl_ctx
*ctx
= pet_expr_get_ctx(expr
);
4834 isl_multi_pw_aff
*index
;
4836 access
= isl_alloc_type(ctx
, struct gpu_stmt_access
);
4838 access
->next
= NULL
;
4839 access
->read
= pet_expr_access_is_read(expr
);
4840 access
->write
= pet_expr_access_is_write(expr
);
4841 tagged
= pet_expr_access_get_tagged_may_read(expr
);
4842 tagged
= isl_union_map_union(tagged
,
4843 pet_expr_access_get_tagged_may_write(expr
));
4844 tagged
= isl_union_map_apply_range(tagged
,
4845 isl_union_map_copy(data
->any_to_outer
));
4846 if (!access
->write
) {
4847 access
->exact_write
= 1;
4848 } else if (!data
->single_expression
) {
4849 access
->exact_write
= 0;
4851 isl_union_map
*must
, *may
;
4852 may
= isl_union_map_copy(tagged
);
4853 may
= isl_union_map_domain_factor_domain(may
);
4854 must
= pet_expr_access_get_must_write(expr
);
4855 access
->exact_write
= isl_union_map_is_equal(must
, may
);
4856 isl_union_map_free(must
);
4857 isl_union_map_free(may
);
4859 index
= pet_expr_access_get_index(expr
);
4860 access
->n_index
= isl_multi_pw_aff_dim(index
, isl_dim_out
);
4861 isl_multi_pw_aff_free(index
);
4862 access
->ref_id
= pet_expr_access_get_ref_id(expr
);
4863 access
->tagged_access
= extract_single_tagged_access(tagged
, expr
);
4864 access
->access
= isl_map_copy(access
->tagged_access
);
4865 access
->access
= isl_map_domain_factor_domain(access
->access
);
4867 *data
->next_access
= access
;
4868 data
->next_access
= &(*data
->next_access
)->next
;
4870 if (!access
->access
)
4876 /* Construct a linked list of gpu_stmt_access objects,
4877 * one for each access expression in the statement body.
4878 * "any_to_outer" maps all intermediate arrays to their outer arrays.
4880 static int pet_stmt_extract_accesses(struct gpu_stmt
*stmt
,
4881 __isl_keep isl_union_map
*any_to_outer
)
4883 struct ppcg_extract_access_data data
;
4885 stmt
->accesses
= NULL
;
4886 data
.next_access
= &stmt
->accesses
;
4887 data
.single_expression
=
4888 pet_tree_get_type(stmt
->stmt
->body
) == pet_tree_expr
;
4889 data
.any_to_outer
= any_to_outer
;
4890 return pet_tree_foreach_access_expr(stmt
->stmt
->body
,
4891 &extract_access
, &data
);
4894 /* Return an array of gpu_stmt representing the statements in "scop".
4896 static struct gpu_stmt
*extract_stmts(isl_ctx
*ctx
, struct ppcg_scop
*scop
,
4897 __isl_keep isl_set
*context
, __isl_keep isl_union_map
*any_to_outer
)
4900 struct gpu_stmt
*stmts
;
4902 stmts
= isl_calloc_array(ctx
, struct gpu_stmt
, scop
->pet
->n_stmt
);
4906 for (i
= 0; i
< scop
->pet
->n_stmt
; ++i
) {
4907 struct gpu_stmt
*s
= &stmts
[i
];
4909 s
->id
= isl_set_get_tuple_id(scop
->pet
->stmts
[i
]->domain
);
4910 s
->stmt
= scop
->pet
->stmts
[i
];
4911 if (pet_stmt_extract_accesses(s
, any_to_outer
) < 0)
4912 return free_stmts(stmts
, i
+ 1);
4918 /* Callback for ppcg_print_guarded that calls the callback for generate_gpu.
4920 static __isl_give isl_printer
*print_gpu(__isl_take isl_printer
*p
, void *user
)
4922 struct gpu_gen
*gen
= user
;
4924 return gen
->print(p
, gen
->prog
, gen
->tree
, &gen
->types
,
4928 /* Generate CUDA code for "scop" and print it to "p".
4929 * After generating an AST for the transformed scop as explained below,
4930 * we call "gen->print" to print the AST in the desired output format
4933 * If it turns out that it does not make sense to generate GPU code,
4934 * then we generate CPU code instead.
4936 * The GPU code is generated in a context where at least one
4937 * statement instance is executed. The corresponding guard (if any) is printed
4938 * around the entire generated GPU code, except for the declaration
4939 * of the arrays that are visible outside of the scop and that therefore
4940 * cannot be declared inside the body of any possible guard.
4942 * We first compute a schedule that respects the dependences
4943 * of the original program and select the outermost bands
4944 * of tilable dimensions that have at least one parallel loop.
4945 * If the --load-schedule is specified, then the loaded schedule
4946 * is used instead of a computed schedule.
4948 * Each of these bands B is then tiled according to "tile" sizes, resulting
4949 * in two nested bands, with a kernel marker on top
4957 * We then split off at most 2 parallel dimensions from the T band and
4958 * at most 3 parallel dimension from the P band
4971 * A filter is introduced in front of T1 that maps the domain instances
4972 * to block identifiers. Similarly, a filter is introduced in front of P1
4973 * that maps the domain instances to thread identifiers.
4975 * For each iteration of the T2 band and for each array, we compute
4976 * the array elements accessed by that iteration, construct a rectangular
4977 * box around it and shift it to the origin. The result is used
4978 * as shared memory for the array.
4980 * Copying and synchronization statements are added to this schedule tree.
4981 * In principle, these are added in front of the P1 band, but some of
4982 * them may get hoisted up to higher levels.
4984 * The entire AST is then generated from the single resulting schedule tree.
4985 * During the generation the subtrees at kernel nodes (K) are saved
4986 * aside and replaced by kernel calls. The result is printed as host code
4987 * while the saved subtrees are printed as device code.
4989 static __isl_give isl_printer
*generate(__isl_take isl_printer
*p
,
4990 struct gpu_gen
*gen
, struct ppcg_scop
*scop
,
4991 struct ppcg_options
*options
)
4993 struct gpu_prog
*prog
;
4995 isl_set
*context
, *guard
;
4996 isl_schedule
*schedule
;
5000 return isl_printer_free(p
);
5002 ctx
= isl_printer_get_ctx(p
);
5003 prog
= gpu_prog_alloc(ctx
, scop
);
5005 return isl_printer_free(p
);
5007 context
= isl_set_copy(prog
->context
);
5008 guard
= isl_union_set_params(isl_union_set_copy(prog
->scop
->domain
));
5009 prog
->context
= isl_set_intersect(prog
->context
, isl_set_copy(guard
));
5012 schedule
= get_schedule(gen
);
5014 any_permutable
= has_any_permutable_node(schedule
);
5015 if (any_permutable
< 0 || !any_permutable
) {
5016 isl_set_free(context
);
5017 isl_set_free(guard
);
5018 if (any_permutable
< 0)
5019 p
= isl_printer_free(p
);
5021 p
= print_cpu(p
, scop
, options
);
5022 isl_schedule_free(schedule
);
5024 schedule
= map_to_device(gen
, schedule
);
5025 gen
->tree
= generate_code(gen
, schedule
);
5026 p
= ppcg_print_exposed_declarations(p
, prog
->scop
);
5027 p
= ppcg_print_guarded(p
, guard
, context
, &print_gpu
, gen
);
5028 isl_ast_node_free(gen
->tree
);
5031 gpu_prog_free(prog
);
5036 /* Wrapper around generate for use as a ppcg_transform callback.
5038 static __isl_give isl_printer
*generate_wrap(__isl_take isl_printer
*p
,
5039 struct ppcg_scop
*scop
, void *user
)
5041 struct gpu_gen
*gen
= user
;
5043 return generate(p
, gen
, scop
, gen
->options
);
5046 /* Transform the code in the file called "input" by replacing
5047 * all scops by corresponding GPU code and write the results to "out".
5049 int generate_gpu(isl_ctx
*ctx
, const char *input
, FILE *out
,
5050 struct ppcg_options
*options
,
5051 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
5052 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
5053 struct gpu_types
*types
, void *user
), void *user
)
5060 gen
.sizes
= extract_sizes_from_str(ctx
, options
->sizes
);
5061 gen
.options
= options
;
5064 gen
.print_user
= user
;
5066 gen
.types
.name
= NULL
;
5068 if (options
->debug
->dump_sizes
) {
5069 isl_space
*space
= isl_space_params_alloc(ctx
, 0);
5070 gen
.used_sizes
= isl_union_map_empty(space
);
5073 r
= ppcg_transform(ctx
, input
, out
, options
, &generate_wrap
, &gen
);
5075 if (options
->debug
->dump_sizes
) {
5076 isl_union_map_dump(gen
.used_sizes
);
5077 isl_union_map_free(gen
.used_sizes
);
5080 isl_union_map_free(gen
.sizes
);
5081 for (i
= 0; i
< gen
.types
.n
; ++i
)
5082 free(gen
.types
.name
[i
]);
5083 free(gen
.types
.name
);
5088 /* Compute the set of inner array elements that may have their values
5089 * preserved by "prog". In particular, collect the array elements of
5090 * arrays that are not local to "prog" and remove those elements that
5091 * are definitely killed or definitely written by "prog".
5093 static __isl_give isl_union_set
*compute_may_persist(struct gpu_prog
*prog
)
5096 isl_union_set
*may_persist
, *killed
;
5097 isl_union_map
*must_kill
;
5099 may_persist
= isl_union_set_empty(isl_set_get_space(prog
->context
));
5100 for (i
= 0; i
< prog
->n_array
; ++i
) {
5103 if (prog
->array
[i
].local
)
5106 extent
= isl_set_copy(prog
->array
[i
].extent
);
5107 may_persist
= isl_union_set_add_set(may_persist
, extent
);
5110 may_persist
= isl_union_set_intersect_params(may_persist
,
5111 isl_set_copy(prog
->context
));
5112 may_persist
= isl_union_set_apply(may_persist
,
5113 isl_union_map_copy(prog
->to_inner
));
5114 must_kill
= isl_union_map_copy(prog
->tagged_must_kill
);
5115 killed
= isl_union_map_range(must_kill
);
5116 must_kill
= isl_union_map_copy(prog
->must_write
);
5117 killed
= isl_union_set_union(killed
, isl_union_map_range(must_kill
));
5119 may_persist
= isl_union_set_subtract(may_persist
, killed
);
5123 struct gpu_prog
*gpu_prog_alloc(isl_ctx
*ctx
, struct ppcg_scop
*scop
)
5125 struct gpu_prog
*prog
;
5132 prog
= isl_calloc_type(ctx
, struct gpu_prog
);
5137 prog
->context
= isl_set_copy(scop
->context
);
5138 prog
->n_stmts
= scop
->pet
->n_stmt
;
5139 prog
->any_to_outer
= pet_scop_compute_outer_to_any(scop
->pet
);
5140 prog
->any_to_outer
= isl_union_map_reverse(prog
->any_to_outer
);
5141 space
= isl_union_map_get_space(prog
->any_to_outer
);
5142 space
= isl_space_set_from_params(space
);
5143 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
5144 space
= isl_space_map_from_set(space
);
5145 id
= isl_map_identity(space
);
5146 prog
->any_to_outer
= isl_union_map_add_map(prog
->any_to_outer
, id
);
5147 prog
->stmts
= extract_stmts(ctx
, scop
,
5148 prog
->context
, prog
->any_to_outer
);
5149 prog
->read
= isl_union_map_copy(scop
->reads
);
5150 prog
->may_write
= isl_union_map_copy(scop
->may_writes
);
5151 prog
->must_write
= isl_union_map_copy(scop
->must_writes
);
5152 prog
->tagged_must_kill
= isl_union_map_copy(scop
->tagged_must_kills
);
5153 prog
->to_inner
= pet_scop_compute_outer_to_inner(scop
->pet
);
5154 prog
->to_outer
= isl_union_map_copy(prog
->to_inner
);
5155 prog
->to_outer
= isl_union_map_reverse(prog
->to_outer
);
5158 return gpu_prog_free(prog
);
5160 if (collect_array_info(prog
) < 0)
5161 return gpu_prog_free(prog
);
5162 prog
->may_persist
= compute_may_persist(prog
);
5167 void *gpu_prog_free(struct gpu_prog
*prog
)
5171 free_array_info(prog
);
5172 free_stmts(prog
->stmts
, prog
->n_stmts
);
5173 isl_union_map_free(prog
->any_to_outer
);
5174 isl_union_map_free(prog
->to_outer
);
5175 isl_union_map_free(prog
->to_inner
);
5176 isl_union_map_free(prog
->read
);
5177 isl_union_map_free(prog
->may_write
);
5178 isl_union_map_free(prog
->must_write
);
5179 isl_union_map_free(prog
->tagged_must_kill
);
5180 isl_union_map_free(prog
->array_order
);
5181 isl_union_set_free(prog
->may_persist
);
5182 isl_set_free(prog
->context
);