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 /* Return the name of the outer array (of structs) accessed by "access".
41 static const char *get_outer_array_name(__isl_keep isl_map
*access
)
46 space
= isl_space_range(isl_map_get_space(access
));
47 while (space
&& isl_space_is_wrapping(space
))
48 space
= isl_space_domain(isl_space_unwrap(space
));
49 name
= isl_space_get_tuple_name(space
, isl_dim_set
);
50 isl_space_free(space
);
55 /* Collect all references to the given array and store pointers to them
58 static void collect_references(struct gpu_prog
*prog
,
59 struct gpu_array_info
*array
)
65 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
66 struct gpu_stmt
*stmt
= &prog
->stmts
[i
];
67 struct gpu_stmt_access
*access
;
69 for (access
= stmt
->accesses
; access
; access
= access
->next
) {
71 name
= get_outer_array_name(access
->access
);
72 if (name
&& !strcmp(array
->name
, name
))
78 array
->refs
= isl_alloc_array(prog
->ctx
, struct gpu_stmt_access
*, n
);
82 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
83 struct gpu_stmt
*stmt
= &prog
->stmts
[i
];
84 struct gpu_stmt_access
*access
;
86 for (access
= stmt
->accesses
; access
; access
= access
->next
) {
88 name
= get_outer_array_name(access
->access
);
89 if (!name
|| strcmp(array
->name
, name
))
92 array
->refs
[n
++] = access
;
97 /* Compute and return the extent of "array", taking into account the set of
100 * In particular, the extent in the outer dimension is taken
101 * from "accessed", while the extents in the remaining dimensions
102 * are taken from array->extent.
104 * The extent in the outer dimension cannot be taken from array->extent
105 * because that may be unbounded. Furthermore, even if it is bounded,
106 * it may be larger than the piece of the array that is being accessed.
108 static __isl_give isl_set
*compute_extent(struct pet_array
*array
,
109 __isl_keep isl_set
*accessed
)
116 extent
= isl_set_copy(array
->extent
);
118 n_index
= isl_set_dim(accessed
, isl_dim_set
);
122 extent
= isl_set_project_out(extent
, isl_dim_set
, 0, 1);
123 outer
= isl_set_copy(accessed
);
124 outer
= isl_set_project_out(outer
, isl_dim_set
, 1, n_index
- 1);
125 extent
= isl_set_flat_product(outer
, extent
);
126 id
= isl_set_get_tuple_id(accessed
);
127 extent
= isl_set_set_tuple_id(extent
, id
);
132 /* Is the array "array" being extracted a read-only scalar?
134 * That is, is "array" a scalar that is never possibly written to.
135 * An array containing structures is never considered to be a scalar.
137 static int is_read_only_scalar(struct gpu_array_info
*array
,
138 struct gpu_prog
*prog
)
141 isl_union_map
*write
;
144 if (array
->has_compound_element
)
146 if (array
->n_index
!= 0)
149 write
= isl_union_map_copy(prog
->may_write
);
150 space
= isl_set_universe(isl_space_copy(array
->space
));
151 write
= isl_union_map_intersect_range(write
,
152 isl_union_set_from_set(space
));
153 empty
= isl_union_map_is_empty(write
);
154 isl_union_map_free(write
);
159 /* Compute bounds on the host array "pa" based on the corresponding
160 * accessed elements in "arrays"
161 * and collect all references to the array.
162 * Store the results in "info".
164 * If the array is zero-dimensional and does not contain structures,
165 * i.e., if the array is a scalar, we check whether it is read-only.
166 * We also check whether the array is accessed at all.
168 static int extract_array_info(struct gpu_prog
*prog
,
169 struct gpu_array_info
*info
, struct pet_array
*pa
,
170 __isl_keep isl_union_set
*arrays
)
175 isl_multi_pw_aff
*bounds
;
176 isl_set
*accessed
, *extent
;
178 n_index
= isl_set_dim(pa
->extent
, isl_dim_set
);
179 name
= isl_set_get_tuple_name(pa
->extent
);
181 info
->space
= isl_set_get_space(pa
->extent
);
182 info
->name
= strdup(name
);
183 info
->n_index
= n_index
;
184 info
->linearize
= prog
->scop
->options
->linearize_device_arrays
;
186 info
->type
= strdup(pa
->element_type
);
187 info
->size
= pa
->element_size
;
188 info
->local
= pa
->declared
&& !pa
->exposed
;
189 info
->has_compound_element
= pa
->element_is_record
;
190 info
->read_only_scalar
= is_read_only_scalar(info
, prog
);
192 accessed
= isl_union_set_extract_set(arrays
,
193 isl_space_copy(info
->space
));
194 empty
= isl_set_is_empty(accessed
);
195 extent
= compute_extent(pa
, accessed
);
196 isl_set_free(accessed
);
197 info
->extent
= extent
;
200 info
->accessed
= !empty
;
201 bounds
= ppcg_size_from_extent(isl_set_copy(extent
));
202 bounds
= isl_multi_pw_aff_gist(bounds
, isl_set_copy(prog
->context
));
205 if (!isl_multi_pw_aff_is_cst(bounds
))
207 info
->bound
= bounds
;
209 collect_references(prog
, info
);
214 /* Remove independence from the order constraints "order" on array "array".
215 * Since the pairs of iterations in the filter relation of an independence
216 * are guaranteed to be completely independent by the user, there is
217 * no need to ensure that live ranges are ordered along thong pairs.
218 * We make an exception for local variables, though, as the independence
219 * guarantee does not apply to those.
221 * The order constraints are used in two places.
222 * Those on scalars are used in check_scalar_live_ranges to check if
223 * we need to force the scalar to be private. Any non-local scalar
224 * should not be forced scalar if it only appears in independent loops.
225 * Those on non-scalars are added to the coincidence constraints
226 * in compute_schedule because we do not support any array expansion.
227 * Accesses to non-local arrays should not prevent a loop from being
228 * considered coincident so we should indeed remove those constraints
229 * from the order constraints.
231 static __isl_give isl_union_map
*remove_independences(struct gpu_prog
*prog
,
232 struct gpu_array_info
*array
, __isl_take isl_union_map
*order
)
236 for (i
= 0; i
< prog
->scop
->pet
->n_independence
; ++i
) {
237 struct pet_independence
*pi
= prog
->scop
->pet
->independences
[i
];
238 if (isl_union_set_contains(pi
->local
, array
->space
))
241 order
= isl_union_map_subtract(order
,
242 isl_union_map_copy(pi
->filter
));
248 /* For each array in "prog", store the (untagged) order dependences
249 * derived from the array in array->dep_order.
250 * In particular, consider all references that access the given array
251 * and take the order dependences that have one of these references
252 * as source. (Since an order dependence relates two references to
253 * the same array, the target of these order dependences will also
254 * be one of these references.)
255 * Additionally, store the union of these array->dep_order relations
256 * for all non-scalar arrays in prog->array_order.
258 void collect_order_dependences(struct gpu_prog
*prog
)
262 isl_union_map
*accesses
;
264 space
= isl_union_map_get_space(prog
->read
);
265 prog
->array_order
= isl_union_map_empty(space
);
267 accesses
= isl_union_map_copy(prog
->scop
->tagged_reads
);
268 accesses
= isl_union_map_union(accesses
,
269 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
270 accesses
= isl_union_map_universe(accesses
);
271 accesses
= isl_union_map_apply_range(accesses
,
272 isl_union_map_copy(prog
->to_outer
));
274 for (i
= 0; i
< prog
->n_array
; ++i
) {
275 struct gpu_array_info
*array
= &prog
->array
[i
];
278 isl_union_map
*order
;
280 set
= isl_set_universe(isl_space_copy(array
->space
));
281 uset
= isl_union_set_from_set(set
);
282 uset
= isl_union_map_domain(
283 isl_union_map_intersect_range(isl_union_map_copy(accesses
),
285 order
= isl_union_map_copy(prog
->scop
->tagged_dep_order
);
286 order
= isl_union_map_intersect_domain(order
, uset
);
287 order
= isl_union_map_zip(order
);
288 order
= isl_union_set_unwrap(isl_union_map_domain(order
));
289 order
= remove_independences(prog
, array
, order
);
290 array
->dep_order
= order
;
292 if (gpu_array_is_scalar(array
) && !array
->has_compound_element
)
295 prog
->array_order
= isl_union_map_union(prog
->array_order
,
296 isl_union_map_copy(array
->dep_order
));
299 isl_union_map_free(accesses
);
302 /* Construct a gpu_array_info for each array referenced by prog->scop and
303 * collect them in prog->array.
305 * The sizes are based on the extents and the set of possibly accessed
306 * elements by "prog".
307 * If there are any member accesses involved, then they are first mapped
308 * to the outer arrays of structs.
310 * If we are allowing live range reordering, then also set
311 * the dep_order field. Otherwise leave it NULL.
313 static int collect_array_info(struct gpu_prog
*prog
)
317 isl_union_set
*arrays
;
319 arrays
= isl_union_map_range(isl_union_map_copy(prog
->read
));
320 arrays
= isl_union_set_union(arrays
,
321 isl_union_map_range(isl_union_map_copy(prog
->may_write
)));
323 arrays
= isl_union_set_apply(arrays
,
324 isl_union_map_copy(prog
->to_outer
));
326 arrays
= isl_union_set_coalesce(arrays
);
328 prog
->n_array
= prog
->scop
->pet
->n_array
;
329 prog
->array
= isl_calloc_array(prog
->ctx
,
330 struct gpu_array_info
, prog
->n_array
);
332 for (i
= 0; i
< prog
->scop
->pet
->n_array
; ++i
)
333 if (extract_array_info(prog
, &prog
->array
[i
],
334 prog
->scop
->pet
->arrays
[i
], arrays
) < 0)
337 isl_union_set_free(arrays
);
339 if (prog
->scop
->options
->live_range_reordering
)
340 collect_order_dependences(prog
);
345 static void free_array_info(struct gpu_prog
*prog
)
349 for (i
= 0; i
< prog
->n_array
; ++i
) {
350 int n_index
= prog
->array
[i
].n_index
;
351 free(prog
->array
[i
].type
);
352 free(prog
->array
[i
].name
);
353 isl_multi_pw_aff_free(prog
->array
[i
].bound
);
354 isl_ast_expr_free(prog
->array
[i
].bound_expr
);
355 isl_space_free(prog
->array
[i
].space
);
356 isl_set_free(prog
->array
[i
].extent
);
357 isl_ast_expr_free(prog
->array
[i
].declared_size
);
358 free(prog
->array
[i
].refs
);
359 isl_union_map_free(prog
->array
[i
].dep_order
);
364 /* Check if a gpu array is a scalar. A scalar is a value that is not stored
365 * as an array or through a pointer reference, but as a single data element.
366 * At the moment, scalars are represented as zero-dimensional arrays.
367 * Note that the single data element may be an entire structure.
369 int gpu_array_is_scalar(struct gpu_array_info
*array
)
371 return array
->n_index
== 0;
374 /* Is "array" a read-only scalar?
376 int gpu_array_is_read_only_scalar(struct gpu_array_info
*array
)
378 return array
->read_only_scalar
;
381 /* Does "array" need to be allocated on the device?
382 * If it is a read-only scalar, then it will be passed as an argument
383 * to the kernel and therefore does not require any allocation.
384 * If this device memory is not accessed at all, then it does not
385 * need to be allocated either.
387 int gpu_array_requires_device_allocation(struct gpu_array_info
*array
)
389 if (gpu_array_is_read_only_scalar(array
))
396 /* Return the set of parameter values for which the array has a positive
397 * size in all dimensions.
398 * If the sizes are only valid for some parameter values, then those
399 * constraints are also taken into account.
401 __isl_give isl_set
*gpu_array_positive_size_guard(struct gpu_array_info
*array
)
410 space
= isl_space_params(isl_space_copy(array
->space
));
411 guard
= isl_set_universe(space
);
413 for (i
= 0; i
< array
->n_index
; ++i
) {
415 isl_set
*guard_i
, *zero
;
417 bound
= isl_multi_pw_aff_get_pw_aff(array
->bound
, i
);
418 guard_i
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(bound
));
419 zero
= isl_pw_aff_zero_set(bound
);
420 guard_i
= isl_set_subtract(guard_i
, zero
);
421 guard
= isl_set_intersect(guard
, guard_i
);
427 /* Internal data structure for extract_size_of_type.
428 * "type" specifies the name of the space that we want to extract.
429 * "res" is used to store the subset of that space.
431 struct ppcg_extract_size_data
{
436 /* This function is called for each set in a union_set.
437 * If the name of the set matches data->type, we store the
440 static isl_stat
extract_size_of_type(__isl_take isl_set
*size
, void *user
)
442 struct ppcg_extract_size_data
*data
= user
;
445 name
= isl_set_get_tuple_name(size
);
446 if (name
&& !strcmp(name
, data
->type
)) {
448 return isl_stat_error
;
455 /* Given a union map { kernel[i] -> *[...] },
456 * return the range in the space called "type" for the kernel with
457 * sequence number "id".
459 static __isl_give isl_set
*extract_sizes(__isl_keep isl_union_map
*sizes
,
460 const char *type
, int id
)
464 isl_union_set
*local_sizes
;
465 struct ppcg_extract_size_data data
= { type
, NULL
};
470 space
= isl_union_map_get_space(sizes
);
471 space
= isl_space_set_from_params(space
);
472 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
473 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
474 dom
= isl_set_universe(space
);
475 dom
= isl_set_fix_si(dom
, isl_dim_set
, 0, id
);
477 local_sizes
= isl_union_set_apply(isl_union_set_from_set(dom
),
478 isl_union_map_copy(sizes
));
479 isl_union_set_foreach_set(local_sizes
, &extract_size_of_type
, &data
);
480 isl_union_set_free(local_sizes
);
484 /* Given a singleton set, extract the first (at most *len) elements
485 * of the single integer tuple into *sizes and update *len if needed.
487 static void read_sizes_from_set(__isl_take isl_set
*set
, int *sizes
, int *len
)
495 dim
= isl_set_dim(set
, isl_dim_set
);
499 for (i
= 0; i
< *len
; ++i
) {
502 v
= isl_set_plain_get_val_if_fixed(set
, isl_dim_set
, i
);
505 sizes
[i
] = isl_val_get_num_si(v
);
512 /* Add the map { kernel[id] -> type[sizes] } to gen->used_sizes,
513 * if the option debug->dump_sizes is set.
515 static void set_used_sizes(struct gpu_gen
*gen
, const char *type
, int id
,
522 if (!gen
->options
->debug
->dump_sizes
)
525 space
= isl_union_map_get_space(gen
->used_sizes
);
526 space
= isl_space_set_from_params(space
);
527 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
528 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
529 space
= isl_space_from_domain(space
);
530 space
= isl_space_add_dims(space
, isl_dim_out
, len
);
531 space
= isl_space_set_tuple_name(space
, isl_dim_out
, type
);
533 map
= isl_map_universe(space
);
534 map
= isl_map_fix_si(map
, isl_dim_in
, 0, id
);
535 for (i
= 0; i
< len
; ++i
)
536 map
= isl_map_fix_si(map
, isl_dim_out
, i
, sizes
[i
]);
538 gen
->used_sizes
= isl_union_map_add_map(gen
->used_sizes
, map
);
541 /* Extract user specified "tile" sizes from the "sizes" command line option,
542 * defaulting to option->tile_size in each dimension.
543 * *tile_len contains the maximum number of tile sizes needed.
544 * Update *tile_len to the number of specified tile sizes, if any, and
545 * return a pointer to the tile sizes (or NULL on error).
546 * Add the effectively used sizes to gen->used_sizes.
548 static int *read_tile_sizes(struct gpu_gen
*gen
, int *tile_len
)
554 tile_size
= isl_alloc_array(gen
->ctx
, int, *tile_len
);
557 for (n
= 0; n
< *tile_len
; ++n
)
558 tile_size
[n
] = gen
->options
->tile_size
;
560 size
= extract_sizes(gen
->sizes
, "tile", gen
->kernel_id
);
561 read_sizes_from_set(size
, tile_size
, tile_len
);
562 set_used_sizes(gen
, "tile", gen
->kernel_id
, tile_size
, *tile_len
);
567 /* Extract user specified "block" sizes from the "sizes" command line option,
568 * after filling in some potentially useful defaults.
570 static void read_block_sizes(struct ppcg_kernel
*kernel
,
571 __isl_keep isl_union_map
*sizes
)
575 if (kernel
->n_block
> 3)
577 switch (kernel
->n_block
) {
579 kernel
->block_dim
[0] = 512;
582 kernel
->block_dim
[0] = 32;
583 kernel
->block_dim
[1] = 16;
586 kernel
->block_dim
[0] = 32;
587 kernel
->block_dim
[1] = 4;
588 kernel
->block_dim
[2] = 4;
592 size
= extract_sizes(sizes
, "block", kernel
->id
);
593 read_sizes_from_set(size
, kernel
->block_dim
, &kernel
->n_block
);
596 /* Extract user specified "grid" sizes from the "sizes" command line option,
597 * after filling in some potentially useful defaults.
599 static void read_grid_sizes(struct ppcg_kernel
*kernel
,
600 __isl_keep isl_union_map
*sizes
)
604 if (kernel
->n_grid
> 2)
606 switch (kernel
->n_grid
) {
608 kernel
->grid_dim
[0] = 32768;
611 kernel
->grid_dim
[0] = 256;
612 kernel
->grid_dim
[1] = 256;
616 size
= extract_sizes(sizes
, "grid", kernel
->id
);
617 read_sizes_from_set(size
, kernel
->grid_dim
, &kernel
->n_grid
);
620 /* Extract user specified grid and block sizes from the gen->sizes
621 * command line option after filling in some potentially useful defaults.
622 * Store the extracted sizes in "kernel".
623 * Add the effectively used sizes to gen->used_sizes.
625 static void read_grid_and_block_sizes(struct ppcg_kernel
*kernel
,
628 read_block_sizes(kernel
, gen
->sizes
);
629 read_grid_sizes(kernel
, gen
->sizes
);
630 set_used_sizes(gen
, "block", kernel
->id
,
631 kernel
->block_dim
, kernel
->n_block
);
632 set_used_sizes(gen
, "grid", kernel
->id
,
633 kernel
->grid_dim
, kernel
->n_grid
);
636 static void *free_stmts(struct gpu_stmt
*stmts
, int n
)
643 for (i
= 0; i
< n
; ++i
) {
644 struct gpu_stmt_access
*access
, *next
;
646 for (access
= stmts
[i
].accesses
; access
; access
= next
) {
648 isl_id_free(access
->ref_id
);
649 isl_map_free(access
->access
);
650 isl_map_free(access
->tagged_access
);
654 isl_id_free(stmts
[i
].id
);
661 /* Add parameters p[i] with identifiers "ids" to "set",
662 * with bounds to 0 <= p[i] < size[i].
664 __isl_give isl_set
*add_bounded_parameters(__isl_take isl_set
*set
,
665 int *size
, __isl_keep isl_id_list
*ids
)
670 len
= isl_id_list_n_id(ids
);
671 nparam
= isl_set_dim(set
, isl_dim_param
);
672 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
674 for (i
= 0; i
< len
; ++i
) {
677 id
= isl_id_list_get_id(ids
, i
);
678 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
679 set
= isl_set_lower_bound_si(set
, isl_dim_param
, nparam
+ i
, 0);
680 set
= isl_set_upper_bound_si(set
, isl_dim_param
,
681 nparam
+ i
, size
[i
] - 1);
687 /* Add "len" parameters p[i] with identifiers "ids" and intersect "set"
690 * { : 0 <= p[i] < size[i] }
692 * or an overapproximation.
694 static __isl_give isl_set
*add_bounded_parameters_dynamic(
695 __isl_take isl_set
*set
, __isl_keep isl_multi_pw_aff
*size
,
696 __isl_keep isl_id_list
*ids
)
703 len
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
704 nparam
= isl_set_dim(set
, isl_dim_param
);
705 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
707 for (i
= 0; i
< len
; ++i
) {
710 id
= isl_id_list_get_id(ids
, i
);
711 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
714 space
= isl_space_params(isl_set_get_space(set
));
715 ls
= isl_local_space_from_space(space
);
716 for (i
= 0; i
< len
; ++i
) {
717 isl_pw_aff
*param
, *size_i
, *zero
;
720 param
= isl_pw_aff_var_on_domain(isl_local_space_copy(ls
),
721 isl_dim_param
, nparam
+ i
);
723 size_i
= isl_multi_pw_aff_get_pw_aff(size
, i
);
724 bound
= isl_pw_aff_lt_set(isl_pw_aff_copy(param
), size_i
);
725 bound
= isl_set_from_basic_set(isl_set_simple_hull(bound
));
726 set
= isl_set_intersect_params(set
, bound
);
728 zero
= isl_pw_aff_zero_on_domain(isl_local_space_copy(ls
));
729 bound
= isl_pw_aff_ge_set(param
, zero
);
730 set
= isl_set_intersect_params(set
, bound
);
732 isl_local_space_free(ls
);
737 /* Return the union of all tagged access relations in the group.
739 static __isl_give isl_union_map
*group_tagged_access_relation(
740 struct gpu_array_ref_group
*group
)
743 isl_union_map
*access
;
745 access
= isl_union_map_empty(isl_map_get_space(group
->access
));
746 for (i
= 0; i
< group
->n_ref
; ++i
) {
749 map_i
= isl_map_copy(group
->refs
[i
]->tagged_access
);
750 access
= isl_union_map_union(access
,
751 isl_union_map_from_map(map_i
));
757 /* Return the extent of "array", recomputed from the bounds.
758 * The recomputed extent may be simpler than the original extent.
760 static __isl_give isl_set
*array_extent(struct gpu_array_info
*array
)
768 id
= isl_set_get_tuple_id(array
->extent
);
769 space
= isl_set_get_space(array
->extent
);
770 extent
= isl_set_universe(isl_space_copy(space
));
771 ls
= isl_local_space_from_space(space
);
772 for (i
= 0; i
< array
->n_index
; ++i
) {
778 extent
= isl_set_lower_bound_si(extent
, isl_dim_set
, i
, 0);
780 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
782 index
= isl_pw_aff_from_aff(aff
);
783 bound
= isl_multi_pw_aff_get_pw_aff(array
->bound
, i
);
784 bound
= isl_pw_aff_from_range(bound
);
785 bound
= isl_pw_aff_add_dims(bound
, isl_dim_in
, array
->n_index
);
786 bound
= isl_pw_aff_set_tuple_id(bound
, isl_dim_in
,
788 lt
= isl_pw_aff_lt_set(index
, bound
);
789 extent
= isl_set_intersect(extent
, lt
);
791 isl_local_space_free(ls
);
797 /* Return a map from the first group->shared_tile->depth dimensions
798 * of the computed schedule to the array tile in
799 * global memory that corresponds to the shared memory copy.
801 * In particular, return a map
807 * tile_offset(i) <= a <= tile_offset(i) + tile_size - 1 (1)
811 * 0 <= a <= array_size - 1 (2)
813 * Note that if some stride has been detected (i.e., when
814 * group->shared_tile->bound[i].shift is set), then a in (1) refers
815 * to the shifted and scaled down version.
817 * Constraints (1) are obtained by mapping the size constraints on the
818 * shared/private memory tile back to the access relation.
819 * Constraints (2) are obtained from the (recomputed) extent.
821 static __isl_give isl_map
*group_tile(struct gpu_array_ref_group
*group
)
824 int n_index
= group
->array
->n_index
;
830 space
= isl_multi_aff_get_space(group
->shared_tile
->tiling
);
831 space
= isl_space_range(space
);
832 local
= isl_set_universe(space
);
833 for (i
= 0; i
< n_index
; ++i
) {
836 local
= isl_set_lower_bound_si(local
, isl_dim_set
, i
, 0);
837 bound
= isl_val_copy(group
->shared_tile
->bound
[i
].size
);
838 bound
= isl_val_sub_ui(bound
, 1);
839 local
= isl_set_upper_bound_val(local
, isl_dim_set
, i
, bound
);
841 local
= isl_set_preimage_multi_aff(local
,
842 isl_multi_aff_copy(group
->shared_tile
->tiling
));
843 tile
= isl_set_unwrap(local
);
844 extent
= array_extent(group
->array
);
845 tile
= isl_map_intersect_range(tile
, extent
);
850 /* Given a mapping "iterator_map" from the AST schedule to a domain,
851 * return the corresponding mapping from the AST schedule to
852 * to the outer kernel->copy_schedule_dim dimensions of
853 * the schedule computed by PPCG for this kernel.
855 * Note that kernel->copy_schedule_dim is at least as large as
856 * the largest depth of any array reference group associated to the kernel.
857 * This is needed as the returned schedule is used to extract a mapping
858 * to the outer tile->depth dimensions in transform_index.
860 static __isl_give isl_pw_multi_aff
*compute_sched_to_copy(
861 struct ppcg_kernel
*kernel
, __isl_take isl_pw_multi_aff
*iterator_map
)
863 isl_union_pw_multi_aff
*upma
;
864 isl_pw_multi_aff
*pma
;
867 space
= isl_space_range(isl_pw_multi_aff_get_space(iterator_map
));
868 space
= isl_space_from_domain(space
);
869 space
= isl_space_add_dims(space
, isl_dim_out
,
870 kernel
->copy_schedule_dim
);
872 upma
= isl_union_pw_multi_aff_copy(kernel
->copy_schedule
);
873 pma
= isl_union_pw_multi_aff_extract_pw_multi_aff(upma
, space
);
874 isl_union_pw_multi_aff_free(upma
);
876 return isl_pw_multi_aff_pullback_pw_multi_aff(pma
, iterator_map
);
879 /* If max_shared_memory is not set to infinity (-1), then make
880 * sure that the total amount of shared memory required by the
881 * array reference groups mapped to shared memory by "kernel"
882 * is no larger than this maximum.
884 * We apply a greedy approach and discard (keep in global memory)
885 * those groups that would result in a total memory size that
886 * is larger than the maximum.
888 * This function should be called after any function that may
889 * affect the decision on whether to place a reference group
890 * in private, shared or global memory.
892 static void check_shared_memory_bound(struct ppcg_kernel
*kernel
)
895 isl_val
*left
, *size
;
897 if (kernel
->options
->max_shared_memory
< 0)
900 left
= isl_val_int_from_si(kernel
->ctx
,
901 kernel
->options
->max_shared_memory
);
903 for (i
= 0; i
< kernel
->n_array
; ++i
) {
904 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
906 for (j
= 0; j
< local
->n_group
; ++j
) {
907 struct gpu_array_ref_group
*group
;
908 enum ppcg_group_access_type type
;
910 group
= local
->groups
[j
];
911 type
= gpu_array_ref_group_type(group
);
912 if (type
!= ppcg_access_shared
)
915 size
= gpu_array_tile_size(group
->shared_tile
);
916 size
= isl_val_mul_ui(size
, local
->array
->size
);
918 if (isl_val_le(size
, left
)) {
919 left
= isl_val_sub(left
, size
);
925 gpu_array_tile_free(group
->shared_tile
);
932 /* Mark all arrays of "kernel" that have an array reference group
933 * that is not mapped to private or shared memory as
934 * accessing the corresponding global device memory.
936 static void mark_global_arrays(struct ppcg_kernel
*kernel
)
940 for (i
= 0; i
< kernel
->n_array
; ++i
) {
941 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
945 for (j
= 0; j
< local
->n_group
; ++j
) {
946 if (gpu_array_ref_group_tile(local
->groups
[j
]))
950 local
->array
->global
= 1;
956 /* Compute a tiling for all the array reference groups in "kernel".
958 static void compute_group_tilings(struct ppcg_kernel
*kernel
)
962 for (i
= 0; i
< kernel
->n_array
; ++i
) {
963 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
965 for (j
= 0; j
< array
->n_group
; ++j
)
966 gpu_array_ref_group_compute_tiling(array
->groups
[j
]);
970 /* Compute the effective grid size as a list of the sizes in each dimension.
972 * The grid size specified by the user or set by default
973 * in read_grid_sizes() and applied by the block filter,
974 * may be too large for the given code in the sense that
975 * it may contain blocks that don't need to execute anything.
976 * We therefore don't return this grid size, but instead the
977 * smallest grid size that ensures that all blocks that actually
978 * execute code are included in the grid.
980 * We first extract a description of the grid, i.e., the possible values
981 * of the block ids, from the domain elements in "domain" and
982 * kernel->block_filter.
983 * The block ids are parameters in kernel->block_filter.
984 * We simply need to change them into set dimensions.
986 * Then, for each block dimension, we compute the maximal value of the block id
989 static __isl_give isl_multi_pw_aff
*extract_grid_size(
990 struct ppcg_kernel
*kernel
, __isl_take isl_union_set
*domain
)
995 isl_multi_pw_aff
*size
;
997 domain
= isl_union_set_intersect(domain
,
998 isl_union_set_copy(kernel
->block_filter
));
999 grid
= isl_union_set_params(domain
);
1000 grid
= isl_set_from_params(grid
);
1001 grid
= isl_set_add_dims(grid
, isl_dim_set
, kernel
->n_grid
);
1002 for (i
= 0; i
< kernel
->n_grid
; ++i
) {
1006 id
= isl_id_list_get_id(kernel
->block_ids
, i
);
1007 pos
= isl_set_find_dim_by_id(grid
, isl_dim_param
, id
);
1010 grid
= isl_set_equate(grid
, isl_dim_param
, pos
, isl_dim_set
, i
);
1011 grid
= isl_set_project_out(grid
, isl_dim_param
, pos
, 1);
1014 grid
= isl_set_coalesce(grid
);
1015 size
= ppcg_size_from_extent(grid
);
1016 context
= isl_set_params(isl_set_copy(kernel
->context
));
1017 return isl_multi_pw_aff_gist(size
, context
);
1020 /* Compute the size of a fixed bounding box around the origin and "set",
1021 * where "set" is assumed to contain only non-negative elements,
1022 * and store the results in "size".
1023 * In particular, compute the maximal value of "set" in each direction
1026 static void extract_fixed_size(__isl_take isl_set
*set
, int *size
)
1029 isl_local_space
*ls
;
1032 n
= isl_set_dim(set
, isl_dim_set
);
1033 ls
= isl_local_space_from_space(isl_set_get_space(set
));
1034 obj
= isl_aff_zero_on_domain(ls
);
1035 for (i
= 0; i
< n
; ++i
) {
1038 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 1);
1039 max
= isl_set_max_val(set
, obj
);
1040 size
[i
] = isl_val_get_num_si(max
) + 1;
1042 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 0);
1048 /* Compute the effective block size as a list of the sizes in each dimension
1049 * and store the sizes in kernel->block_dim.
1051 * The block size specified by the user or set by default
1052 * in read_block_sizes() and applied by the thread filter,
1053 * may be too large for the given code in the sense that
1054 * it may contain threads that don't need to execute anything.
1055 * We therefore update this block size in kernel->block_dim
1056 * to the smallest block size that ensures that all threads
1057 * that actually execute code are included in the block.
1059 * The possible values of the thread ids is obtained from
1060 * the domain elements "domain" and kernel->thread_filter.
1061 * The current implementation eliminates all parameters, ensuring
1062 * that the size is a fixed constant in each dimension.
1063 * In principle we could also compute parametric sizes.
1064 * We would have to make sure to project out all b%d and t%d parameters,
1067 static isl_stat
extract_block_size(struct ppcg_kernel
*kernel
,
1068 __isl_take isl_union_set
*domain
)
1074 domain
= isl_union_set_intersect(domain
,
1075 isl_union_set_copy(kernel
->thread_filter
));
1076 block
= isl_union_set_params(domain
);
1077 block
= isl_set_from_params(block
);
1078 block
= isl_set_add_dims(block
, isl_dim_set
, kernel
->n_block
);
1079 for (i
= 0; i
< kernel
->n_block
; ++i
) {
1084 return isl_stat_error
;
1086 id
= isl_id_list_get_id(kernel
->thread_ids
, i
);
1087 pos
= isl_set_find_dim_by_id(block
, isl_dim_param
, id
);
1090 isl_die(isl_set_get_ctx(block
), isl_error_internal
,
1091 "missing constraints on thread identifier",
1092 block
= isl_set_free(block
));
1093 block
= isl_set_equate(block
, isl_dim_param
, pos
,
1096 nparam
= isl_set_dim(block
, isl_dim_param
);
1097 block
= isl_set_project_out(block
, isl_dim_param
, 0, nparam
);
1100 return isl_stat_error
;
1102 extract_fixed_size(block
, kernel
->block_dim
);
1107 struct ppcg_kernel
*ppcg_kernel_free(struct ppcg_kernel
*kernel
)
1114 isl_id_list_free(kernel
->block_ids
);
1115 isl_id_list_free(kernel
->thread_ids
);
1116 isl_multi_pw_aff_free(kernel
->grid_size
);
1117 isl_ast_expr_free(kernel
->grid_size_expr
);
1118 isl_set_free(kernel
->context
);
1119 isl_union_set_free(kernel
->core
);
1120 isl_union_set_free(kernel
->arrays
);
1121 isl_space_free(kernel
->space
);
1122 isl_ast_node_free(kernel
->tree
);
1123 isl_union_set_free(kernel
->block_filter
);
1124 isl_union_set_free(kernel
->thread_filter
);
1125 isl_union_pw_multi_aff_free(kernel
->copy_schedule
);
1126 isl_union_set_free(kernel
->sync_writes
);
1128 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1129 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1131 for (j
= 0; j
< array
->n_group
; ++j
)
1132 gpu_array_ref_group_free(array
->groups
[j
]);
1133 free(array
->groups
);
1135 isl_multi_pw_aff_free(array
->bound
);
1136 isl_ast_expr_free(array
->bound_expr
);
1138 free(kernel
->array
);
1140 for (i
= 0; i
< kernel
->n_var
; ++i
) {
1141 free(kernel
->var
[i
].name
);
1142 isl_vec_free(kernel
->var
[i
].size
);
1151 /* Wrapper around ppcg_kernel_free for use as a isl_id_set_free_user callback.
1153 static void ppcg_kernel_free_wrap(void *user
)
1155 struct ppcg_kernel
*kernel
= user
;
1157 ppcg_kernel_free(kernel
);
1160 static void create_kernel_var(isl_ctx
*ctx
, struct gpu_array_ref_group
*group
,
1161 struct ppcg_kernel_var
*var
)
1164 struct gpu_array_tile
*tile
;
1168 var
->array
= group
->array
;
1170 var
->type
= gpu_array_ref_group_type(group
);
1171 tile
= gpu_array_ref_group_tile(group
);
1173 p
= isl_printer_to_str(ctx
);
1174 p
= gpu_array_ref_group_print_name(group
, p
);
1175 var
->name
= isl_printer_get_str(p
);
1176 isl_printer_free(p
);
1178 var
->size
= isl_vec_alloc(ctx
, group
->array
->n_index
);
1180 for (j
= 0; j
< group
->array
->n_index
; ++j
)
1181 var
->size
= isl_vec_set_element_val(var
->size
, j
,
1182 isl_val_copy(tile
->bound
[j
].size
));
1185 static int create_kernel_vars(struct ppcg_kernel
*kernel
)
1190 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1191 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1193 for (j
= 0; j
< array
->n_group
; ++j
) {
1194 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1195 enum ppcg_group_access_type type
;
1197 type
= gpu_array_ref_group_type(group
);
1198 if (type
!= ppcg_access_global
)
1204 kernel
->var
= isl_calloc_array(kernel
->ctx
, struct ppcg_kernel_var
, n
);
1209 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1210 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1212 for (j
= 0; j
< array
->n_group
; ++j
) {
1213 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1214 enum ppcg_group_access_type type
;
1216 type
= gpu_array_ref_group_type(group
);
1217 if (type
== ppcg_access_global
)
1219 create_kernel_var(kernel
->ctx
, group
, &kernel
->var
[n
]);
1227 /* Replace "pa" by the zero function defined over the universe domain
1228 * in the space of "pa".
1230 static __isl_give isl_pw_aff
*set_universally_zero(__isl_take isl_pw_aff
*pa
)
1235 space
= isl_space_domain(isl_pw_aff_get_space(pa
));
1236 isl_pw_aff_free(pa
);
1237 zero
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1239 return isl_pw_aff_from_aff(zero
);
1242 /* The sizes of the arrays on the host that have been computed by
1243 * extract_array_info may depend on the parameters. Use the extra
1244 * constraints on the parameters that are valid at "host_domain"
1245 * to simplify these expressions and store the results in kernel->array.
1247 * We only need these localized bounds for arrays that are accessed
1248 * by the current kernel. If we have found at least one reference group
1249 * then the array is accessed by the kernel.
1251 * The resulting sizes may be functions that are nowhere defined
1252 * in case the access function cannot possibly access anything inside
1253 * the kernel for some reason. If so, they are replaced by the zero
1254 * function. Since the access function cannot actually access anything,
1255 * there is no harm in printing the array sizes as zero.
1257 static void localize_bounds(struct ppcg_kernel
*kernel
,
1258 __isl_keep isl_set
*host_domain
)
1263 context
= isl_set_copy(host_domain
);
1264 context
= isl_set_params(context
);
1266 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1267 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
1268 isl_multi_pw_aff
*bound
;
1271 if (local
->n_group
== 0)
1274 n_index
= local
->array
->n_index
;
1275 bound
= isl_multi_pw_aff_copy(local
->array
->bound
);
1277 for (j
= 0; j
< n_index
; ++j
) {
1281 pwaff
= isl_multi_pw_aff_get_pw_aff(bound
, j
);
1282 pwaff
= isl_pw_aff_gist(pwaff
, isl_set_copy(context
));
1283 empty
= isl_pw_aff_is_empty(pwaff
);
1285 pwaff
= isl_pw_aff_free(pwaff
);
1287 pwaff
= set_universally_zero(pwaff
);
1288 bound
= isl_multi_pw_aff_set_pw_aff(bound
, j
, pwaff
);
1291 local
->n_index
= n_index
;
1292 local
->bound
= bound
;
1294 isl_set_free(context
);
1297 /* Create the array of gpu_local_array_info structures "array"
1298 * inside "kernel". The number of elements in this array is
1299 * the same as the number of arrays in "prog".
1300 * Initialize the "array" field of each local array to point
1301 * to the corresponding array in "prog".
1303 static struct ppcg_kernel
*ppcg_kernel_create_local_arrays(
1304 struct ppcg_kernel
*kernel
, struct gpu_prog
*prog
)
1309 ctx
= isl_set_get_ctx(prog
->context
);
1310 kernel
->array
= isl_calloc_array(ctx
,
1311 struct gpu_local_array_info
, prog
->n_array
);
1313 return ppcg_kernel_free(kernel
);
1314 kernel
->n_array
= prog
->n_array
;
1316 for (i
= 0; i
< prog
->n_array
; ++i
)
1317 kernel
->array
[i
].array
= &prog
->array
[i
];
1322 /* Does "kernel" need to be passed an argument corresponding to array "i"?
1324 * The argument is only needed if the kernel accesses this device memory.
1326 int ppcg_kernel_requires_array_argument(struct ppcg_kernel
*kernel
, int i
)
1328 return kernel
->array
[i
].global
;
1331 /* Find the element in gen->stmt that has the given "id".
1332 * Return NULL if no such gpu_stmt can be found.
1334 static struct gpu_stmt
*find_stmt(struct gpu_prog
*prog
, __isl_keep isl_id
*id
)
1338 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
1339 if (id
== prog
->stmts
[i
].id
)
1343 return i
< prog
->n_stmts
? &prog
->stmts
[i
] : NULL
;
1346 void ppcg_kernel_stmt_free(void *user
)
1349 struct ppcg_kernel_stmt
*stmt
= user
;
1354 switch (stmt
->type
) {
1355 case ppcg_kernel_copy
:
1356 isl_ast_expr_free(stmt
->u
.c
.index
);
1357 isl_ast_expr_free(stmt
->u
.c
.local_index
);
1359 case ppcg_kernel_domain
:
1360 isl_id_to_ast_expr_free(stmt
->u
.d
.ref2expr
);
1362 case ppcg_kernel_sync
:
1369 /* Return the gpu_stmt_access in the list "accesses" that corresponds
1372 static struct gpu_stmt_access
*find_access(struct gpu_stmt_access
*accesses
,
1373 __isl_keep isl_id
*ref_id
)
1375 struct gpu_stmt_access
*access
;
1377 for (access
= accesses
; access
; access
= access
->next
)
1378 if (access
->ref_id
== ref_id
)
1384 /* Return the index of the array called "name" in the list of arrays.
1386 static int find_array_index(struct ppcg_kernel
*kernel
, const char *name
)
1390 for (i
= 0; i
< kernel
->n_array
; ++i
)
1391 if (!strcmp(name
, kernel
->array
[i
].array
->name
))
1397 /* Internal data structure for the index and AST expression transformation
1398 * callbacks for pet_stmt_build_ast_exprs.
1400 * "kernel" is the kernel for which are computing AST expressions and
1401 * may be NULL if we are not inside a kernel.
1402 * "accesses" is the list of gpu_stmt_access in the statement.
1403 * "iterator_map" expresses the statement iterators in terms of
1404 * the AST loop iterators.
1405 * "sched2copy" expresses the outer copy_schedule_dim dimensions of
1406 * the kernel schedule in terms of the AST loop iterators and
1407 * may be NULL if we are not inside a kernel.
1409 * The following fields are set in transform_index and used in transform_expr.
1410 * "array" is the array that is being accessed.
1411 * "global" is set if the global array is accessed (rather than
1412 * shared/private memory).
1413 * "local_array" refers to information on the array specialized
1414 * to the current kernel.
1416 struct ppcg_transform_data
{
1417 struct ppcg_kernel
*kernel
;
1418 struct gpu_stmt_access
*accesses
;
1419 isl_pw_multi_aff
*iterator_map
;
1420 isl_pw_multi_aff
*sched2copy
;
1422 struct gpu_array_info
*array
;
1424 struct gpu_local_array_info
*local_array
;
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 tile->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
= gpu_array_ref_group_tile(group
);
1534 data
->global
= !tile
;
1538 space
= isl_space_range(isl_multi_pw_aff_get_space(index
));
1539 space
= isl_space_map_from_set(space
);
1540 pma
= isl_pw_multi_aff_identity(space
);
1541 sched2depth
= isl_pw_multi_aff_copy(data
->sched2copy
);
1542 dim
= isl_pw_multi_aff_dim(sched2depth
, isl_dim_out
);
1543 sched2depth
= isl_pw_multi_aff_drop_dims(sched2depth
, isl_dim_out
,
1544 tile
->depth
, dim
- tile
->depth
);
1545 pma
= isl_pw_multi_aff_product(sched2depth
, pma
);
1546 tiling
= isl_multi_pw_aff_from_multi_aff(
1547 isl_multi_aff_copy(tile
->tiling
));
1548 tiling
= isl_multi_pw_aff_pullback_pw_multi_aff(tiling
, pma
);
1550 space
= isl_space_domain(isl_multi_pw_aff_get_space(index
));
1551 space
= isl_space_map_from_set(space
);
1552 mpa
= isl_multi_pw_aff_identity(space
);
1553 index
= isl_multi_pw_aff_range_product(mpa
, index
);
1554 index
= isl_multi_pw_aff_pullback_multi_pw_aff(tiling
, index
);
1559 /* Dereference "expr" by adding an index [0].
1560 * The original "expr" is assumed not to have any indices.
1562 * If "expr" is a member access, then the dereferencing needs
1563 * to be applied to the structure argument of this member access.
1565 static __isl_give isl_ast_expr
*dereference(__isl_take isl_ast_expr
*expr
)
1568 isl_ast_expr
*arg0
, *res
;
1569 isl_ast_expr_list
*list
;
1571 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1573 return isl_ast_expr_free(expr
);
1574 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1575 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1578 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1579 arg
= dereference(arg
);
1580 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1581 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1585 isl_ast_expr_free(arg0
);
1587 ctx
= isl_ast_expr_get_ctx(expr
);
1588 res
= isl_ast_expr_from_val(isl_val_zero(ctx
));
1589 list
= isl_ast_expr_list_from_ast_expr(res
);
1590 res
= isl_ast_expr_get_op_arg(expr
, 0);
1591 res
= isl_ast_expr_access(res
, list
);
1592 isl_ast_expr_free(expr
);
1597 /* Linearize the index expression "expr" based on the array bounds
1600 * That is, transform expression
1602 * A[i_0][i_1]...[i_n]
1606 * A[(..((i_0 * b_1 + i_1) ... ) * b_n + i_n]
1608 * where b_0, b_1, ..., b_n are the bounds on the array.
1610 * If the base of "expr" is a member access, then the linearization needs
1611 * to be applied to the structure argument of this member access.
1613 * In the base case, if "expr" has no arguments (other than the name of
1614 * the array), then we are passing an entire array to a function.
1615 * In this case, there is nothing to linearize.
1616 * Note that at this point an expression with no arguments can
1617 * only be an entire array because the scalar case and
1618 * the case of single struct are handled by the caller.
1620 * If the number of specified index expressions in "expr"
1621 * is smaller than the dimension of the accessed array,
1622 * then the missing i_j also do not appear in the linearized expression.
1623 * Furthermore, since such an expression does not refer to a single
1624 * element while the default linearized expression would refer to
1625 * a single element, we return the expression
1627 * A + (..((i_0 * b_1 + i_1) ... ) * b_l + i_l)
1629 * instead. Note that because of the special case handling above,
1630 * we can assume here that there is at least one index expression.
1632 __isl_give isl_ast_expr
*gpu_local_array_info_linearize_index(
1633 struct gpu_local_array_info
*array
, __isl_take isl_ast_expr
*expr
)
1638 isl_ast_expr_list
*list
;
1640 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1641 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1642 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1645 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1646 arg
= gpu_local_array_info_linearize_index(array
, arg
);
1647 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1648 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1652 isl_ast_expr_free(arg0
);
1654 if (isl_ast_expr_get_op_n_arg(expr
) == 1)
1657 n
= isl_ast_expr_get_op_n_arg(expr
);
1658 res
= isl_ast_expr_get_op_arg(expr
, 1);
1659 for (i
= 1; i
< array
->n_index
; ++i
) {
1660 isl_ast_expr
*expr_i
;
1662 expr_i
= isl_ast_expr_get_op_arg(array
->bound_expr
, 1 + i
);
1663 res
= isl_ast_expr_mul(res
, expr_i
);
1667 expr_i
= isl_ast_expr_get_op_arg(expr
, i
+ 1);
1668 res
= isl_ast_expr_add(res
, expr_i
);
1671 if (1 + array
->n_index
> n
) {
1672 res
= isl_ast_expr_add(isl_ast_expr_get_op_arg(expr
, 0), res
);
1674 list
= isl_ast_expr_list_from_ast_expr(res
);
1675 res
= isl_ast_expr_get_op_arg(expr
, 0);
1676 res
= isl_ast_expr_access(res
, list
);
1679 isl_ast_expr_free(expr
);
1684 /* AST expression transformation callback for pet_stmt_build_ast_exprs.
1686 * If the AST expression refers to an array that is not accessed
1687 * at all, then this means the value of the expression is not used,
1688 * so we might as well print zero (NULL pointer) instead.
1690 * If the AST expression refers to a global scalar that is not
1691 * a read-only scalar, then its address was passed to the kernel and
1692 * we need to dereference it.
1694 * If the AST expression refers to an access to a global array,
1695 * then we linearize the access exploiting the bounds in data->local_array.
1697 static __isl_give isl_ast_expr
*transform_expr(__isl_take isl_ast_expr
*expr
,
1698 __isl_keep isl_id
*id
, void *user
)
1700 struct ppcg_transform_data
*data
= user
;
1704 if (!data
->array
->accessed
) {
1707 ctx
= isl_ast_expr_get_ctx(expr
);
1708 isl_ast_expr_free(expr
);
1709 return isl_ast_expr_from_val(isl_val_zero(ctx
));
1711 if (gpu_array_is_read_only_scalar(data
->array
))
1715 if (data
->array
->n_index
== 0)
1716 return dereference(expr
);
1717 if (!data
->array
->linearize
)
1720 return gpu_local_array_info_linearize_index(data
->local_array
, expr
);
1723 /* This function is called for each instance of a user statement
1724 * in the kernel "kernel", identified by "gpu_stmt".
1725 * "kernel" may be NULL if we are not inside a kernel.
1727 * We attach a struct ppcg_kernel_stmt to the "node", containing
1728 * a computed AST expression for each access, through an annotation
1730 * These AST expressions are computed from iterator_map,
1731 * which expresses the domain
1732 * elements in terms of the generated loops, and sched2copy,
1733 * which expresses the outer copy_schedule_dim dimensions of
1734 * the kernel schedule computed by PPCG in terms of the generated loops.
1736 static __isl_give isl_ast_node
*create_domain_leaf(
1737 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1738 __isl_keep isl_ast_build
*build
, struct gpu_stmt
*gpu_stmt
)
1740 struct ppcg_transform_data data
;
1741 struct ppcg_kernel_stmt
*stmt
;
1744 isl_pw_multi_aff
*sched2copy
;
1746 isl_pw_multi_aff
*iterator_map
;
1747 isl_union_map
*schedule
;
1751 ctx
= isl_ast_node_get_ctx(node
);
1753 stmt
= isl_calloc_type(ctx
, struct ppcg_kernel_stmt
);
1755 return isl_ast_node_free(node
);
1757 schedule
= isl_ast_build_get_schedule(build
);
1758 map
= isl_map_reverse(isl_map_from_union_map(schedule
));
1759 iterator_map
= isl_pw_multi_aff_from_map(map
);
1761 sched2copy
= compute_sched_to_copy(kernel
,
1762 isl_pw_multi_aff_copy(iterator_map
));
1766 stmt
->type
= ppcg_kernel_domain
;
1767 stmt
->u
.d
.stmt
= gpu_stmt
;
1769 data
.kernel
= kernel
;
1770 data
.accesses
= stmt
->u
.d
.stmt
->accesses
;
1771 data
.iterator_map
= iterator_map
;
1772 data
.sched2copy
= sched2copy
;
1773 stmt
->u
.d
.ref2expr
= pet_stmt_build_ast_exprs(stmt
->u
.d
.stmt
->stmt
,
1774 build
, &transform_index
, &data
,
1775 &transform_expr
, &data
);
1777 isl_pw_multi_aff_free(iterator_map
);
1778 isl_pw_multi_aff_free(sched2copy
);
1780 id
= isl_id_alloc(ctx
, "user", stmt
);
1781 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1782 return isl_ast_node_set_annotation(node
, id
);
1785 /* This function is called for each statement node in the AST
1786 * for copying to or from shared/private memory.
1787 * Attach a pointer to a ppcg_kernel_stmt representing the copy
1788 * statement to the node.
1789 * The statement name is "read" or "write", depending on whether we are
1790 * reading from global memory or writing to global memory.
1792 * The schedule is of the form
1796 * where D corresponds to the outer tile->depth dimensions of
1797 * the kernel schedule, A to the global array and L to the outer
1798 * generated AST schedule.
1799 * We compute the inverse and strip off the type, resulting in
1803 * We combine this mapping with on the one hand the projection
1807 * and on the other hand the group tiling
1815 * and store the corresponding expressions in stmt->index and stmt->local_index,
1816 * where stmt points to the ppcg_kernel_stmt that is attached to the node.
1817 * stmt->index is linearized if the global memory array is linearized.
1819 static __isl_give isl_ast_node
*create_access_leaf(struct ppcg_kernel
*kernel
,
1820 struct gpu_array_ref_group
*group
, __isl_take isl_ast_node
*node
,
1821 __isl_keep isl_ast_build
*build
)
1823 struct ppcg_kernel_stmt
*stmt
;
1824 struct gpu_array_tile
*tile
;
1829 isl_pw_multi_aff
*pma
, *pma2
;
1832 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1834 return isl_ast_node_free(node
);
1836 access
= isl_map_from_union_map(isl_ast_build_get_schedule(build
));
1837 type
= isl_map_get_tuple_name(access
, isl_dim_in
);
1838 stmt
->u
.c
.read
= !strcmp(type
, "read");
1839 access
= isl_map_reverse(access
);
1840 pma
= isl_pw_multi_aff_from_map(access
);
1841 pma
= isl_pw_multi_aff_reset_tuple_id(pma
, isl_dim_out
);
1843 space
= isl_space_range(isl_pw_multi_aff_get_space(pma
));
1844 space
= isl_space_unwrap(space
);
1845 pma2
= isl_pw_multi_aff_range_map(space
);
1846 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
,
1847 isl_pw_multi_aff_copy(pma
));
1848 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1849 if (group
->array
->linearize
)
1850 expr
= gpu_local_array_info_linearize_index(group
->local_array
,
1852 stmt
->u
.c
.index
= expr
;
1854 tile
= gpu_array_ref_group_tile(group
);
1855 pma2
= isl_pw_multi_aff_from_multi_aff(
1856 isl_multi_aff_copy(tile
->tiling
));
1857 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
, pma
);
1858 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1859 stmt
->u
.c
.local_index
= expr
;
1861 stmt
->u
.c
.array
= group
->array
;
1862 stmt
->u
.c
.local_array
= group
->local_array
;
1863 stmt
->type
= ppcg_kernel_copy
;
1865 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1866 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1867 return isl_ast_node_set_annotation(node
, id
);
1870 /* Create a synchronization ppcg_kernel_stmt and
1871 * attach it to the node "node" representing the synchronization.
1873 static __isl_give isl_ast_node
*create_sync_leaf(
1874 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1875 __isl_keep isl_ast_build
*build
)
1877 struct ppcg_kernel_stmt
*stmt
;
1880 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1882 return isl_ast_node_free(node
);
1884 stmt
->type
= ppcg_kernel_sync
;
1885 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1886 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1887 return isl_ast_node_set_annotation(node
, id
);
1890 /* Build AST expressions for the device array sizes of all arrays in "prog"
1891 * that require allocation on the device using "build", as well as
1892 * for the original array sizes of all arrays that need to be declared
1894 * "node" is freed in case of error.
1896 static __isl_give isl_ast_node
*build_array_bounds(
1897 __isl_take isl_ast_node
*node
, struct gpu_prog
*prog
,
1898 __isl_keep isl_ast_build
*build
)
1902 for (i
= 0; i
< prog
->n_array
; ++i
) {
1903 struct gpu_array_info
*array
= &prog
->array
[i
];
1904 isl_multi_pw_aff
*size
;
1907 if (!gpu_array_requires_device_allocation(array
))
1910 size
= isl_multi_pw_aff_copy(array
->bound
);
1911 expr
= ppcg_build_size_expr(size
, build
);
1912 array
->bound_expr
= expr
;
1914 return isl_ast_node_free(node
);
1917 for (i
= 0; i
< prog
->n_array
; ++i
) {
1918 struct gpu_array_info
*array
= &prog
->array
[i
];
1919 struct pet_array
*pet_array
= prog
->scop
->pet
->arrays
[i
];
1920 isl_multi_pw_aff
*size
;
1923 if (!array
->declare_local
)
1925 size
= ppcg_size_from_extent(isl_set_copy(pet_array
->extent
));
1926 expr
= ppcg_build_size_expr(size
, build
);
1927 array
->declared_size
= expr
;
1929 return isl_ast_node_free(node
);
1935 /* Internal data structure for at_domain.
1937 * "prog" represents the entire scop.
1938 * "kernel" points to the kernel to which the current schedule node
1939 * belongs. It is set by before_mark and reset by after_mark.
1940 * It may be NULL if we are outside any kernel.
1942 struct ppcg_at_domain_data
{
1943 struct gpu_prog
*prog
;
1944 struct ppcg_kernel
*kernel
;
1947 /* This function is called for each instance of a user statement
1948 * in the kernel. This may be one of the original user statements
1949 * or a statement introduced by PPCG.
1951 * We first check if the statement id corresponds to a gpu statement,
1952 * which indicates the statement is an original user statement. Any statement
1953 * that is not an original user statement has been introduced by PPCG and
1954 * requires special handling.
1956 * If the user statement is one of the original user statements, then we call
1957 * create_domain_leaf. If it is "init_device", then we call
1958 * build_array_bounds. Otherwise, we check if it is a copy or synchronization
1959 * statement and call the appropriate functions. Statements that copy an array
1960 * to/from the device do not need any further treatment.
1961 * Neither does "clear_device".
1963 static __isl_give isl_ast_node
*at_domain(__isl_take isl_ast_node
*node
,
1964 __isl_keep isl_ast_build
*build
, void *user
)
1966 struct ppcg_at_domain_data
*data
= user
;
1967 struct gpu_stmt
*gpu_stmt
;
1968 isl_ast_expr
*expr
, *arg
;
1974 expr
= isl_ast_node_user_get_expr(node
);
1975 arg
= isl_ast_expr_get_op_arg(expr
, 0);
1976 id
= isl_ast_expr_get_id(arg
);
1977 name
= isl_id_get_name(id
);
1978 p
= isl_id_get_user(id
);
1979 isl_ast_expr_free(expr
);
1980 isl_ast_expr_free(arg
);
1982 gpu_stmt
= find_stmt(data
->prog
, id
);
1983 is_sync
= gpu_tree_id_is_sync(id
, data
->kernel
);
1987 return create_domain_leaf(data
->kernel
, node
, build
, gpu_stmt
);
1989 if (!prefixcmp(name
, "to_device_") || !prefixcmp(name
, "from_device_"))
1991 if (!strcmp(name
, "init_device"))
1992 return build_array_bounds(node
, data
->prog
, build
);
1993 if (!strcmp(name
, "clear_device"))
1996 return isl_ast_node_free(node
);
1997 if (!strcmp(name
, "read") || !strcmp(name
, "write")) {
1998 struct gpu_array_ref_group
*group
= p
;
1999 return create_access_leaf(data
->kernel
, group
, node
, build
);
2002 isl_die(data
->prog
->ctx
, isl_error_internal
,
2003 "unknown statement type",
2004 return isl_ast_node_free(node
));
2005 return create_sync_leaf(data
->kernel
, node
, build
);
2008 /* Given a set of wrapped references "ref", return the corresponding
2009 * access relations based on the tagged access relations "tagged".
2011 * The elements of "ref" are of the form
2015 * with D an iteration domains and R a reference.
2016 * The elements of "tagged" are of the form
2022 * Extend "tagged" to include the iteration domain in the range, i.e.,
2024 * [D -> R] -> [D -> A]
2026 * apply the result to "ref" and then unwrap the resulting set
2027 * to obtain relations of the form
2031 static __isl_give isl_union_map
*wrapped_reference_to_access(
2032 __isl_take isl_union_set
*ref
, __isl_take isl_union_map
*tagged
)
2034 isl_union_map
*tag2access
;
2036 tag2access
= isl_union_map_copy(tagged
);
2037 tag2access
= isl_union_map_universe(tag2access
);
2038 tag2access
= isl_union_set_unwrap(isl_union_map_domain(tag2access
));
2039 tag2access
= isl_union_map_domain_map(tag2access
);
2040 tag2access
= isl_union_map_range_product(tag2access
, tagged
);
2042 ref
= isl_union_set_coalesce(ref
);
2043 ref
= isl_union_set_apply(ref
, tag2access
);
2045 return isl_union_set_unwrap(ref
);
2048 /* Given an access relation "access" from one or more array reference groups,
2049 * remove those reads if ("read" is 1) or writes (if "read" is 0)
2050 * that are only needed to communicate data within
2051 * the same iteration of "sched".
2052 * "tagged" contains all tagged access relations to all
2053 * the array reference groups accessed by "access" from statement
2054 * instances scheduled by "sched".
2056 * If the access is a read then it is either an element of
2058 * live_in union (range flow)
2060 * where live_in and flow may be overapproximations, or
2061 * it reads an uninitialized value (that is not live-in because
2062 * there is an intermediate kill) or it reads a value that was
2063 * written within the same (compound) statement instance.
2064 * If the access is a write then it is either an element of
2066 * live_out union (domain flow)
2068 * or it writes a value that is never read (and is not live-out
2069 * because of an intermediate kill) or only
2070 * within the same (compound) statement instance.
2071 * In both cases, the access relation is also a subset of
2072 * the group access relation.
2074 * The cases where an uninitialized value is read or a value is written
2075 * that is never read or where the dataflow occurs within a statement
2076 * instance are also considered local and may also be removed.
2078 * Essentially, we compute the intersection of "access" with either
2080 * live_in union (range non-local-flow)
2084 * live_out union (domain non-local-flow)
2086 * We first construct a relation "local"
2088 * [[D -> R] -> [D' -> R']]
2090 * of pairs of domain iterations accessing the reference group
2091 * and references in the group that are coscheduled by "sched".
2093 * If this relation does not intersect the dataflow dependences,
2094 * then there is nothing we can possibly remove, unless the dataflow
2095 * dependences themselves only relate a subset of the accesses.
2096 * In particular, the accesses may not be involved in any dataflow
2097 * dependences, either because they are uninitialized reads/dead writes
2098 * or because the dataflow occurs inside a statement instance.
2100 * Since the computation below may break up the access relation
2101 * into smaller pieces, we only perform the intersection with
2102 * the non-local dependent accesses if the local pairs
2103 * intersect the dataflow dependences. Otherwise, we intersect
2104 * with the universe of the non-local dependent accesses.
2105 * This should at least remove accesses from statements that
2106 * do not participate in any dependences.
2108 * In particular, we remove the "local" dataflow dependences from
2109 * the set of all dataflow dependences, or at least those
2110 * that may contribute to a domain/range that intersects
2111 * the domain of "access".
2112 * Note that if the potential dataflow dependences are an overapproximation
2113 * of the actual dataflow dependences, then the result remains an
2114 * overapproximation of the non-local dataflow dependences.
2115 * Copying to/from global memory is only needed for the references
2116 * in the domain/range of the result or for accesses that are live out/in
2117 * for the entire scop.
2119 * We therefore map the domain/range of the "external" relation
2120 * to the corresponding access relation and take the union with
2121 * the live out/in relation.
2123 static __isl_give isl_union_map
*remove_local_accesses(
2124 struct gpu_prog
*prog
, __isl_take isl_union_map
*tagged
,
2125 __isl_take isl_union_map
*access
, __isl_take isl_union_map
*sched
,
2129 isl_union_pw_multi_aff
*tagger
;
2130 isl_union_set
*domain
, *access_domain
;
2131 isl_union_map
*local
, *external
, *universe
;
2132 isl_union_set
*tag_set
;
2134 if (isl_union_map_is_empty(access
)) {
2135 isl_union_map_free(sched
);
2136 isl_union_map_free(tagged
);
2140 tagger
= isl_union_pw_multi_aff_copy(prog
->scop
->tagger
);
2141 domain
= isl_union_map_domain(isl_union_map_copy(tagged
));
2142 tagger
= isl_union_pw_multi_aff_intersect_domain(tagger
,
2143 isl_union_set_copy(domain
));
2144 sched
= isl_union_map_preimage_domain_union_pw_multi_aff(sched
, tagger
);
2146 local
= isl_union_map_apply_range(sched
,
2147 isl_union_map_reverse(isl_union_map_copy(sched
)));
2148 local
= isl_union_map_intersect(local
,
2149 isl_union_map_copy(prog
->scop
->tagged_dep_flow
));
2151 empty
= isl_union_map_is_empty(local
);
2153 external
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
2154 universe
= isl_union_map_universe(isl_union_map_copy(access
));
2155 access_domain
= isl_union_map_domain(universe
);
2156 domain
= isl_union_set_universe(domain
);
2157 universe
= isl_union_set_unwrap(domain
);
2158 universe
= isl_union_map_intersect_domain(universe
, access_domain
);
2159 domain
= isl_union_map_wrap(universe
);
2161 external
= isl_union_map_intersect_range(external
, domain
);
2163 external
= isl_union_map_intersect_domain(external
, domain
);
2164 external
= isl_union_map_intersect_params(external
,
2165 isl_set_copy(prog
->scop
->context
));
2166 external
= isl_union_map_subtract(external
, local
);
2169 tag_set
= isl_union_map_range(external
);
2170 external
= wrapped_reference_to_access(tag_set
, tagged
);
2171 external
= isl_union_map_union(external
,
2172 isl_union_map_copy(prog
->scop
->live_in
));
2174 tag_set
= isl_union_map_domain(external
);
2175 external
= wrapped_reference_to_access(tag_set
, tagged
);
2176 external
= isl_union_map_union(external
,
2177 isl_union_map_copy(prog
->scop
->live_out
));
2181 external
= isl_union_map_free(external
);
2183 external
= isl_union_map_universe(external
);
2185 access
= isl_union_map_intersect(access
, external
);
2190 /* Given an access relation "access" from "group", remove those reads
2191 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
2192 * communicate data within the same iteration of the schedule at the
2193 * position where the copying of the group is inserted.
2194 * "node" points to this position, i.e., the depth at "node"
2195 * is equal to tile->depth.
2197 * We extract a schedule that picks out the iterations of the outer
2198 * tile->depth dimensions and call remove_local_accesses.
2200 static __isl_give isl_union_map
*remove_local_accesses_group(
2201 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2202 __isl_take isl_union_map
*access
, __isl_keep isl_schedule_node
*node
,
2205 isl_union_map
*sched
, *tagged
;
2207 if (isl_union_map_is_empty(access
))
2210 tagged
= group_tagged_access_relation(group
);
2211 sched
= isl_schedule_node_get_prefix_schedule_relation(node
);
2213 return remove_local_accesses(kernel
->prog
, tagged
, access
, sched
, read
);
2216 /* Build an access AST expression for the effective grid size using "build".
2217 * Store the result in kernel->grid_size_expr.
2219 static isl_stat
build_grid_size(struct ppcg_kernel
*kernel
,
2220 __isl_keep isl_ast_build
*build
)
2222 isl_multi_pw_aff
*size
;
2224 size
= isl_multi_pw_aff_copy(kernel
->grid_size
);
2225 size
= isl_multi_pw_aff_set_tuple_name(size
, isl_dim_out
, "grid");
2226 kernel
->grid_size_expr
= ppcg_build_size_expr(size
, build
);
2228 if (!kernel
->grid_size_expr
)
2229 return isl_stat_error
;
2233 /* Build access AST expressions for the localized array sizes using "build".
2234 * Store the result in local->bound_expr.
2235 * Only do this for arrays for which localized bounds have been computed.
2237 static isl_stat
build_local_array_sizes(struct ppcg_kernel
*kernel
,
2238 __isl_keep isl_ast_build
*build
)
2242 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2243 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
2244 isl_multi_pw_aff
*size
;
2246 if (local
->n_group
== 0)
2248 size
= isl_multi_pw_aff_copy(local
->bound
);
2249 local
->bound_expr
= ppcg_build_size_expr(size
, build
);
2250 if (!local
->bound_expr
)
2251 return isl_stat_error
;
2257 /* Build access AST expressions for the effective grid size and
2258 * the localized array sizes using "build".
2260 static isl_stat
build_grid_and_local_array_sizes(struct ppcg_kernel
*kernel
,
2261 __isl_keep isl_ast_build
*build
)
2263 if (build_grid_size(kernel
, build
) < 0)
2264 return isl_stat_error
;
2265 if (build_local_array_sizes(kernel
, build
) < 0)
2266 return isl_stat_error
;
2270 /* This function is called before the AST generator starts traversing
2271 * the schedule subtree of a node with mark "mark".
2273 * If the mark is called "kernel", store the kernel pointer in data->kernel
2274 * for use in at_domain and build AST expressions for the grid size and
2275 * the localized array sizes.
2277 static isl_stat
before_mark(__isl_keep isl_id
*mark
,
2278 __isl_keep isl_ast_build
*build
, void *user
)
2280 struct ppcg_at_domain_data
*data
= user
;
2283 return isl_stat_error
;
2284 if (!strcmp(isl_id_get_name(mark
), "kernel")) {
2285 data
->kernel
= isl_id_get_user(mark
);
2286 if (build_grid_and_local_array_sizes(data
->kernel
, build
) < 0)
2287 return isl_stat_error
;
2292 /* This function is called after the AST generator has finished traversing
2293 * the schedule subtree of a mark node. "node" points to the corresponding
2296 * If the mark is called "kernel", then replace "node" by a user node
2297 * that "calls" the kernel, representing the launch of the kernel.
2298 * The original "node" is stored inside the kernel object so that
2299 * it can be used to print the device code.
2300 * Note that this assumes that a kernel is only launched once.
2301 * Also clear data->kernel.
2303 static __isl_give isl_ast_node
*after_mark(__isl_take isl_ast_node
*node
,
2304 __isl_keep isl_ast_build
*build
, void *user
)
2309 isl_ast_expr_list
*list
;
2310 struct ppcg_kernel
*kernel
;
2311 struct ppcg_at_domain_data
*data
= user
;
2313 ctx
= isl_ast_node_get_ctx(node
);
2314 id
= isl_ast_node_mark_get_id(node
);
2316 return isl_ast_node_free(node
);
2317 if (strcmp(isl_id_get_name(id
), "kernel") || !data
->kernel
) {
2321 kernel
= data
->kernel
;
2322 data
->kernel
= NULL
;
2323 kernel
->space
= isl_ast_build_get_schedule_space(build
);
2324 kernel
->tree
= isl_ast_node_mark_get_node(node
);
2325 isl_ast_node_free(node
);
2327 expr
= isl_ast_expr_from_id(isl_id_copy(id
));
2328 list
= isl_ast_expr_list_alloc(ctx
, 0);
2329 expr
= isl_ast_expr_call(expr
, list
);
2330 node
= isl_ast_node_alloc_user(expr
);
2331 node
= isl_ast_node_set_annotation(node
, id
);
2336 static isl_bool
update_depth(__isl_keep isl_schedule_node
*node
, void *user
)
2341 if (isl_schedule_node_get_type(node
) != isl_schedule_node_leaf
)
2342 return isl_bool_true
;
2343 node_depth
= isl_schedule_node_get_schedule_depth(node
);
2344 if (node_depth
> *depth
)
2345 *depth
= node_depth
;
2347 return isl_bool_false
;
2350 /* Use isl to generate code for both the host and the device
2352 * The device code is marked by "kernel" mark nodes in the schedule tree,
2353 * containing a pointer to a ppcg_kernel object.
2354 * The returned AST only contains the AST for the host code.
2355 * The ASTs for the device code are embedded in ppcg_kernel objects
2356 * attached to the leaf nodes that call "kernel".
2358 static __isl_give isl_ast_node
*generate_code(struct gpu_gen
*gen
,
2359 __isl_take isl_schedule
*schedule
)
2361 struct ppcg_at_domain_data data
;
2362 isl_ast_build
*build
;
2364 isl_id_list
*iterators
;
2367 data
.prog
= gen
->prog
;
2371 if (isl_schedule_foreach_schedule_node_top_down(schedule
, &update_depth
,
2374 build
= isl_ast_build_alloc(gen
->prog
->ctx
);
2375 iterators
= ppcg_scop_generate_names(gen
->prog
->scop
, depth
, "c");
2376 build
= isl_ast_build_set_iterators(build
, iterators
);
2377 build
= isl_ast_build_set_at_each_domain(build
, &at_domain
, &data
);
2378 build
= isl_ast_build_set_before_each_mark(build
, &before_mark
, &data
);
2379 build
= isl_ast_build_set_after_each_mark(build
, &after_mark
, &data
);
2380 if (gen
->prog
->scop
->options
->debug
->dump_final_schedule
)
2381 isl_schedule_dump(schedule
);
2382 tree
= isl_ast_build_node_from_schedule(build
, schedule
);
2383 isl_ast_build_free(build
);
2388 __isl_give isl_union_map
*extract_sizes_from_str(isl_ctx
*ctx
, const char *str
)
2392 return isl_union_map_read_from_str(ctx
, str
);
2395 /* Can "node" be tiled and then mapped to block and thread identifiers?
2396 * That is, is it permutable with at least one coincident dimension?
2398 static int is_permutable(__isl_keep isl_schedule_node
*node
)
2403 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
2405 if (!isl_schedule_node_band_get_permutable(node
))
2407 if (isl_schedule_node_band_n_member(node
) < 1)
2409 if (!isl_schedule_node_band_member_get_coincident(node
, 0))
2415 /* A isl_schedule_foreach_schedule_node_top_down callback
2416 * for setting *any_permutable and aborting the search
2417 * if "node" is a permutable band with coincident dimensions.
2418 * Otherwise, continue searching.
2420 static isl_bool
set_permutable(__isl_keep isl_schedule_node
*node
, void *user
)
2422 int *any_permutable
= user
;
2425 permutable
= is_permutable(node
);
2427 return isl_bool_error
;
2429 return isl_bool_true
;
2431 *any_permutable
= 1;
2433 return isl_bool_error
;
2436 /* Does the subtree rooted at "node" have any suitably permutable band nodes?
2437 * That is, does it have any nodes that are permutable and that
2438 * have a least one coincident dimension?
2440 static int subtree_has_permutable_bands(__isl_keep isl_schedule_node
*node
)
2442 int any_parallelism
= 0;
2444 if (isl_schedule_node_foreach_descendant_top_down(node
, &set_permutable
,
2445 &any_parallelism
) < 0 &&
2449 return any_parallelism
;
2452 /* Does "schedule" contain any permutable band with at least one coincident
2455 static int has_any_permutable_node(__isl_keep isl_schedule
*schedule
)
2457 isl_schedule_node
*root
;
2460 root
= isl_schedule_get_root(schedule
);
2461 any_permutable
= subtree_has_permutable_bands(root
);
2462 isl_schedule_node_free(root
);
2464 return any_permutable
;
2467 /* Is "node" a candidate for mapping to block and thread identifiers?
2468 * In particular, is it permutable with at least one coincident dimension?
2469 * Alternatively, does the subtree rooted at "node" not contain
2470 * any such permutable node? Filter nodes are skipped in this case,
2471 * because a band node will be inserted in front of the returned
2472 * node and this is not possible for filter nodes that are children
2473 * of set or sequence nodes.
2475 static int is_candidate(__isl_keep isl_schedule_node
*node
)
2479 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
2481 permutable
= is_permutable(node
);
2482 if (permutable
< 0 || permutable
)
2484 if (isl_schedule_node_get_type(node
) == isl_schedule_node_filter
)
2486 permutable
= subtree_has_permutable_bands(node
);
2492 /* Is "node" the outermost node in its branch that can be tiled
2493 * and then mapped to block and thread identifiers?
2494 * If there are no such nodes in the subtree at "node" and
2495 * if "node" is not a filter node, then it is accepted too.
2497 static int is_outer_tilable(__isl_keep isl_schedule_node
*node
)
2500 isl_schedule_node
*ancestor
;
2502 tilable
= is_candidate(node
);
2509 ancestor
= isl_schedule_node_copy(node
);
2510 while (isl_schedule_node_has_parent(ancestor
)) {
2511 ancestor
= isl_schedule_node_parent(ancestor
);
2513 tilable
= is_candidate(ancestor
);
2514 if (tilable
< 0 || tilable
)
2518 isl_schedule_node_free(ancestor
);
2519 return tilable
< 0 ? -1 : !tilable
;
2522 /* Collect the references to all writes in "group".
2523 * Each reference is represented by a universe set in a space
2527 * with S[i,j] the statement instance space and R[] the array reference.
2529 static __isl_give isl_union_set
*group_tagged_writes(
2530 struct gpu_array_ref_group
*group
)
2534 isl_union_set
*writes
;
2536 space
= isl_map_get_space(group
->access
);
2537 writes
= isl_union_set_empty(space
);
2538 for (i
= 0; i
< group
->n_ref
; ++i
) {
2542 if (!group
->refs
[i
]->write
)
2545 space
= isl_map_get_space(group
->refs
[i
]->tagged_access
);
2546 space
= isl_space_domain(space
);
2547 writes_i
= isl_set_universe(space
);
2548 writes
= isl_union_set_add_set(writes
, writes_i
);
2554 /* Is there any write access in "group" that requires synchronization
2555 * on a write to global memory?
2556 * We currently take into account all writes that would require
2557 * synchronization at the thread level depth, but if the copying
2558 * for this group is performed at an outer level, then we do not
2559 * actually need to take into account dependences at intermediate levels.
2561 static int any_sync_writes_in_group(struct ppcg_kernel
*kernel
,
2562 struct gpu_array_ref_group
*group
)
2564 isl_union_set
*writes
;
2565 int empty
, disjoint
;
2567 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2573 writes
= group_tagged_writes(group
);
2574 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2575 isl_union_set_free(writes
);
2577 return disjoint
< 0 ? -1 : !disjoint
;
2580 /* Collect the references to all writes in "kernel" that write directly
2581 * to global or shared memory, i.e., that are not mapped to private memory.
2582 * Each reference is represented by a universe set in a space
2586 * with S[i,j] the statement instance space and R[] the array reference.
2588 static __isl_give isl_union_set
*collect_non_private_tagged_writes(
2589 struct ppcg_kernel
*kernel
)
2591 isl_union_set
*writes
;
2594 writes
= isl_union_set_empty(isl_union_set_get_space(kernel
->arrays
));
2596 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2597 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2599 for (j
= 0; j
< array
->n_group
; ++j
) {
2600 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2601 enum ppcg_group_access_type type
;
2602 isl_union_set
*writes_ij
;
2606 type
= gpu_array_ref_group_type(group
);
2607 if (type
== ppcg_access_private
)
2609 writes_ij
= group_tagged_writes(group
);
2610 writes
= isl_union_set_union(writes
, writes_ij
);
2617 /* Are there any direct writes to global memory that require
2620 static int any_global_or_shared_sync_writes(struct ppcg_kernel
*kernel
)
2622 isl_union_set
*writes
;
2623 int empty
, disjoint
;
2625 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2631 writes
= collect_non_private_tagged_writes(kernel
);
2632 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2633 isl_union_set_free(writes
);
2635 return disjoint
< 0 ? -1 : !disjoint
;
2638 /* Construct an isl_multi_val for use as tile sizes for tiling "node"
2639 * from the elements in "tile_size".
2641 static __isl_give isl_multi_val
*construct_band_tiles_sizes(
2642 __isl_keep isl_schedule_node
*node
, int *tile_size
)
2652 ctx
= isl_schedule_node_get_ctx(node
);
2653 space
= isl_schedule_node_band_get_space(node
);
2654 n
= isl_schedule_node_band_n_member(node
);
2655 mv
= isl_multi_val_zero(space
);
2656 for (i
= 0; i
< n
; ++i
) {
2659 v
= isl_val_int_from_si(ctx
, tile_size
[i
]);
2660 mv
= isl_multi_val_set_val(mv
, i
, v
);
2666 /* Replace the partial schedule S of the band node "node" by
2674 * if scale_tile_loops is set, with f the integers in "factor".
2675 * The list that "factor" points to is assumed to contain at least
2676 * as many elements as the number of members in the band.
2678 static __isl_give isl_schedule_node
*snap_band_to_sizes(
2679 __isl_take isl_schedule_node
*node
, int *factor
,
2680 struct ppcg_options
*options
)
2684 mv
= construct_band_tiles_sizes(node
, factor
);
2685 node
= isl_schedule_node_band_scale_down(node
, isl_multi_val_copy(mv
));
2686 if (options
->scale_tile_loops
)
2687 node
= isl_schedule_node_band_scale(node
,
2688 isl_multi_val_copy(mv
));
2689 isl_multi_val_free(mv
);
2694 /* Tile "band" with tile size specified by "sizes".
2696 * Since the tile loops will be mapped to block ids, we forcibly
2697 * turn off tile loop scaling. We may want to enable tile loop scaling
2698 * at some later point, but then we would have to support the detection
2699 * of strides during the mapping to block ids.
2700 * Similarly, since the point loops will be mapped to thread ids,
2701 * we forcibly shift the point loops so that they start at zero.
2703 static __isl_give isl_schedule_node
*tile_band(
2704 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2706 isl_ctx
*ctx
= isl_schedule_node_get_ctx(node
);
2710 scale_tile
= isl_options_get_tile_scale_tile_loops(ctx
);
2711 isl_options_set_tile_scale_tile_loops(ctx
, 0);
2712 shift_point
= isl_options_get_tile_shift_point_loops(ctx
);
2713 isl_options_set_tile_shift_point_loops(ctx
, 1);
2715 node
= isl_schedule_node_band_tile(node
, sizes
);
2717 isl_options_set_tile_scale_tile_loops(ctx
, scale_tile
);
2718 isl_options_set_tile_shift_point_loops(ctx
, shift_point
);
2723 /* Extract the set of parameter values and outer schedule dimensions
2724 * for which any statement instance
2725 * in the kernel inserted at "node" needs to be executed.
2726 * Intersect the set of parameter values derived from the host schedule
2727 * relation with the context of "prog".
2729 static __isl_give isl_set
*extract_context(__isl_keep isl_schedule_node
*node
,
2730 struct gpu_prog
*prog
)
2732 isl_union_map
*schedule
;
2733 isl_union_set
*schedule_domain
;
2737 schedule
= isl_schedule_node_get_prefix_schedule_relation(node
);
2738 schedule_domain
= isl_union_map_range(schedule
);
2739 empty
= isl_union_set_is_empty(schedule_domain
);
2741 isl_union_set_free(schedule_domain
);
2748 space
= isl_union_set_get_space(schedule_domain
);
2749 isl_union_set_free(schedule_domain
);
2750 space
= isl_space_set_from_params(space
);
2751 depth
= isl_schedule_node_get_schedule_depth(node
);
2752 space
= isl_space_add_dims(space
, isl_dim_set
, depth
);
2753 context
= isl_set_empty(space
);
2755 context
= isl_set_from_union_set(schedule_domain
);
2757 context
= isl_set_intersect_params(context
,
2758 isl_set_copy(prog
->context
));
2763 /* Return the set of outer array elements accessed by
2764 * by the statement instance in "domain" in "prog".
2766 static __isl_give isl_union_set
*accessed_by_domain(
2767 __isl_take isl_union_set
*domain
, struct gpu_prog
*prog
)
2769 isl_union_map
*access
;
2770 isl_union_set
*arrays
;
2772 access
= isl_union_map_union(isl_union_map_copy(prog
->read
),
2773 isl_union_map_copy(prog
->may_write
));
2774 access
= isl_union_map_intersect_domain(access
, domain
);
2775 arrays
= isl_union_map_range(access
);
2776 arrays
= isl_union_set_apply(arrays
,
2777 isl_union_map_copy(prog
->to_outer
));
2782 /* Return the number of outer band members of the band node "node"
2783 * that are marked coincident.
2785 static int n_outer_coincidence(__isl_keep isl_schedule_node
*node
)
2789 n
= isl_schedule_node_band_n_member(node
);
2791 for (i
= 0; i
< n
; ++i
)
2792 if (!isl_schedule_node_band_member_get_coincident(node
, i
))
2798 /* If the band node "node" has more than "n" members, then split off
2799 * the first "n" of them.
2801 static __isl_give isl_schedule_node
*split_band(
2802 __isl_take isl_schedule_node
*node
, int n
)
2806 dim
= isl_schedule_node_band_n_member(node
);
2808 node
= isl_schedule_node_band_split(node
, n
);
2813 /* Scale a band node that may have been split by split_band.
2814 * "sizes" are the scaling factors for the original node.
2815 * "node" either points to the original band node, or the outer
2816 * of the two pieces after splitting.
2818 * If the number of elements in "node" is smaller than the number of
2819 * elements in "sizes", then some splitting has occurred and we split
2820 * "sizes" in the same way.
2822 static __isl_give isl_schedule_node
*scale_band(
2823 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2827 n
= isl_multi_val_dim(sizes
, isl_dim_set
);
2828 dim
= isl_schedule_node_band_n_member(node
);
2830 isl_multi_val
*sizes2
;
2832 sizes2
= isl_multi_val_copy(sizes
);
2833 sizes
= isl_multi_val_drop_dims(sizes
,
2834 isl_dim_set
, dim
, n
- dim
);
2835 sizes2
= isl_multi_val_drop_dims(sizes2
, isl_dim_set
, 0, dim
);
2836 node
= isl_schedule_node_child(node
, 0);
2837 node
= isl_schedule_node_band_scale(node
, sizes2
);
2838 node
= isl_schedule_node_parent(node
);
2841 return isl_schedule_node_band_scale(node
, sizes
);
2844 /* Return an isl_multi_aff, with as elements the parameters in "space"
2845 * that have the names specified by the elements in "names".
2846 * If (some of) these parameters do not already appear in "space",
2847 * then they are added first.
2849 static __isl_give isl_multi_aff
*parameter_vector(__isl_take isl_space
*space
,
2850 __isl_keep isl_id_list
*names
)
2853 isl_local_space
*ls
;
2857 space
= isl_space_free(space
);
2859 n
= isl_id_list_n_id(names
);
2860 for (i
= 0; i
< n
; ++i
) {
2864 id
= isl_id_list_get_id(names
, i
);
2865 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2870 pos
= isl_space_dim(space
, isl_dim_param
);
2871 space
= isl_space_add_dims(space
, isl_dim_param
, 1);
2872 space
= isl_space_set_dim_id(space
, isl_dim_param
, pos
, id
);
2874 ma
= isl_multi_aff_zero(isl_space_copy(space
));
2875 ls
= isl_local_space_from_space(isl_space_domain(space
));
2876 for (i
= 0; i
< n
; ++i
) {
2881 id
= isl_id_list_get_id(names
, i
);
2882 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2884 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
2885 isl_dim_param
, pos
);
2886 ma
= isl_multi_aff_set_aff(ma
, i
, aff
);
2888 isl_local_space_free(ls
);
2893 /* Return constraints on the domain elements that equate a sequence of
2894 * parameters called "names", to the partial schedule
2895 * of "node" modulo the integers in "size".
2896 * The number of elements in the array "size" should be equal
2897 * to the number of elements in "names".
2898 * The number of members of the band node "node" should be smaller
2899 * than or equal to this number. If it is smaller, then the first
2900 * elements of "names" are equated to zero.
2902 static __isl_give isl_union_set
*set_schedule_modulo(
2903 __isl_keep isl_schedule_node
*node
, __isl_keep isl_id_list
*names
,
2909 isl_multi_union_pw_aff
*mupa
, *mupa2
;
2911 isl_union_set
*domain
;
2915 n
= isl_id_list_n_id(names
);
2917 return isl_schedule_node_get_universe_domain(node
);
2918 n_zero
= n
- isl_schedule_node_band_n_member(node
);
2920 mupa
= isl_schedule_node_band_get_partial_schedule(node
);
2921 mv
= construct_band_tiles_sizes(node
, size
+ n_zero
);
2922 mupa
= isl_multi_union_pw_aff_mod_multi_val(mupa
, mv
);
2924 space
= isl_multi_union_pw_aff_get_space(mupa
);
2925 space
= isl_space_params(space
);
2926 space
= isl_space_set_from_params(space
);
2927 space
= isl_space_add_dims(space
, isl_dim_set
, n_zero
);
2928 ma
= isl_multi_aff_zero(space
);
2930 domain
= isl_schedule_node_get_universe_domain(node
);
2931 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(
2932 isl_union_set_copy(domain
), ma
);
2933 mupa
= isl_multi_union_pw_aff_range_product(mupa2
, mupa
);
2935 space
= isl_multi_union_pw_aff_get_space(mupa
);
2936 ma
= parameter_vector(space
, names
);
2938 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(domain
, ma
);
2939 mupa
= isl_multi_union_pw_aff_sub(mupa
, mupa2
);
2941 return isl_multi_union_pw_aff_zero_union_set(mupa
);
2944 /* Insert a context node at "node" introducing the block and thread
2945 * identifiers along with their bounds, which are stored in kernel->grid_size
2946 * and kernel->block_dim.
2947 * Note that the bounds on the block identifiers may implicitly impose
2948 * constraints on the parameters. A guard needs to be inserted
2949 * in the schedule tree to ensure that those bounds hold at "node".
2950 * This guard is inserted in insert_guard.
2952 static __isl_give isl_schedule_node
*insert_context(struct ppcg_kernel
*kernel
,
2953 __isl_take isl_schedule_node
*node
)
2957 context
= isl_set_universe(isl_set_get_space(kernel
->context
));
2959 context
= add_bounded_parameters_dynamic(context
,
2960 kernel
->grid_size
, kernel
->block_ids
);
2961 context
= add_bounded_parameters(context
,
2962 kernel
->block_dim
, kernel
->thread_ids
);
2964 node
= isl_schedule_node_insert_context(node
, context
);
2969 /* Insert a guard that eliminates kernel launches where the kernel
2970 * obviously does not have any work to do.
2972 * In particular, eliminate kernel launches where there are obviously
2974 * Use the same block size constraints that are used to create the context
2975 * to ensure that all constraints implicit in the constructed context
2976 * are imposed by the guard.
2978 * Additionally, add other constraints that are valid
2979 * for each executed instance ("context"), as long as this does not result
2982 static __isl_give isl_schedule_node
*insert_guard(
2983 __isl_take isl_schedule_node
*node
, __isl_keep isl_set
*context
,
2984 __isl_keep isl_multi_pw_aff
*size
, struct ppcg_scop
*scop
)
2990 guard
= isl_set_copy(context
);
2991 guard
= isl_set_compute_divs(guard
);
2992 guard
= isl_set_from_basic_set(isl_set_simple_hull(guard
));
2994 nparam
= isl_set_dim(guard
, isl_dim_param
);
2995 n
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
2996 ids
= ppcg_scop_generate_names(scop
, n
, "__ppcg_tmp");
2997 guard
= add_bounded_parameters_dynamic(guard
, size
, ids
);
2998 isl_id_list_free(ids
);
2999 guard
= isl_set_project_out(guard
, isl_dim_param
, nparam
, n
);
3001 node
= isl_schedule_node_insert_guard(node
, guard
);
3006 /* Does any array reference group mapping require the band that is mapped
3007 * to threads to be unrolled?
3009 static int kernel_requires_unroll(struct ppcg_kernel
*kernel
)
3013 for (i
= 0; i
< kernel
->n_array
; ++i
) {
3014 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
3016 for (j
= 0; j
< array
->n_group
; ++j
) {
3017 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3018 if (gpu_array_ref_group_requires_unroll(group
))
3026 /* Mark the given band node "node" for unrolling by the AST generator and
3027 * then sink it to the leaves of the schedule tree.
3028 * All dimensions of "node" are assumed to be coincident, such that this
3029 * sinking is a valid operation.
3031 static __isl_give isl_schedule_node
*unroll(__isl_take isl_schedule_node
*node
)
3033 node
= ppcg_set_schedule_node_type(node
, isl_ast_loop_unroll
);
3035 node
= isl_schedule_node_band_sink(node
);
3040 /* Insert a synchronization node in the schedule tree of "node"
3041 * after the core computation of "kernel" at the level of the band
3042 * that is mapped to threads, except if that level is equal to
3043 * that of the band that is mapped to blocks or if there are no writes
3044 * to global or shared memory in the core computation that require
3046 * If there are any writes to shared memory and the shared memory
3047 * copying is performed at the same level, then synchronization
3048 * is needed between the core and the copying anyway, so we might
3049 * as well add it here. If the copying is performed at a higher
3050 * level, then different iterations of intermediate schedule dimensions
3051 * may have a different mapping from between shared memory elements and
3052 * threads, such that synchronization is required after the core.
3053 * "node" is assumed to point to the kernel node.
3055 static __isl_give isl_schedule_node
*add_sync(struct ppcg_kernel
*kernel
,
3056 __isl_take isl_schedule_node
*node
)
3061 need_sync
= any_global_or_shared_sync_writes(kernel
);
3063 return isl_schedule_node_free(node
);
3067 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3069 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3070 if (kernel_depth
== isl_schedule_node_get_schedule_depth(node
))
3071 return gpu_tree_move_up_to_kernel(node
);
3073 node
= gpu_tree_ensure_following_sync(node
, kernel
);
3075 node
= gpu_tree_move_up_to_kernel(node
);
3080 /* Return a read ("read" is 1) or write access relation for "group"
3081 * with those accesses removed that are only needed to communicate data
3082 * within the subtree of the schedule rooted at "node".
3083 * Furthermore, include the prefix schedule at "node".
3084 * That is, return a relation of the form
3088 * with D the outer schedule dimensions at "node".
3090 static __isl_give isl_union_map
*anchored_non_local_accesses(
3091 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3092 __isl_take isl_schedule_node
*node
, int read
)
3094 isl_union_map
*access
;
3095 isl_union_map
*prefix
;
3097 access
= gpu_array_ref_group_access_relation(group
, read
, !read
);
3098 access
= remove_local_accesses_group(kernel
, group
, access
, node
, read
);
3099 prefix
= isl_schedule_node_get_prefix_schedule_relation(node
);
3100 access
= isl_union_map_range_product(prefix
, access
);
3105 /* Given an array reference group "group", create a mapping
3107 * read[D -> A] -> [D -> A]
3109 * if "read" is set or
3111 * write[D -> A] -> [D -> A]
3113 * if "read" is not set.
3114 * D corresponds to the outer tile->depth dimensions of
3115 * the kernel schedule.
3117 static __isl_give isl_multi_aff
*create_from_access(isl_ctx
*ctx
,
3118 struct gpu_array_ref_group
*group
, int read
)
3120 struct gpu_array_tile
*tile
;
3124 tile
= gpu_array_ref_group_tile(group
);
3125 space
= isl_space_copy(group
->array
->space
);
3126 space
= isl_space_from_range(space
);
3127 space
= isl_space_add_dims(space
, isl_dim_in
, tile
->depth
);
3128 space
= isl_space_wrap(space
);
3129 space
= isl_space_map_from_set(space
);
3131 id
= isl_id_alloc(ctx
, read
? "read" : "write", group
);
3132 space
= isl_space_set_tuple_id(space
, isl_dim_in
, id
);
3134 return isl_multi_aff_identity(space
);
3137 /* If any writes in "group" require synchronization, then make sure
3138 * that there is a synchronization node for "kernel" after the node
3139 * following "node" in a sequence.
3141 * If "shared" is set and no synchronization is needed for
3142 * the writes to global memory, then add synchronization before
3143 * the kernel to protect shared memory from being overwritten
3144 * by the next iteration of the core computation.
3145 * No additional synchronization is needed to protect against
3146 * the next copy into shared memory because each element of
3147 * the shared memory tile is always copied by the same thread.
3149 static __isl_give isl_schedule_node
*add_group_write_sync(
3150 __isl_take isl_schedule_node
*node
, struct ppcg_kernel
*kernel
,
3151 struct gpu_array_ref_group
*group
, int shared
)
3155 need_sync
= any_sync_writes_in_group(kernel
, group
);
3157 return isl_schedule_node_free(node
);
3159 node
= isl_schedule_node_parent(node
);
3160 node
= isl_schedule_node_next_sibling(node
);
3161 node
= isl_schedule_node_child(node
, 0);
3162 node
= gpu_tree_ensure_following_sync(node
, kernel
);
3163 } else if (shared
) {
3164 struct gpu_array_tile
*tile
;
3166 tile
= gpu_array_ref_group_tile(group
);
3167 node
= isl_schedule_node_parent(node
);
3168 node
= isl_schedule_node_parent(node
);
3169 node
= gpu_tree_move_down_to_depth(node
, tile
->depth
,
3171 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3177 /* Add copy statements to the schedule tree of "node"
3178 * for reading from global memory to private memory (if "read" is set) or
3179 * for writing back from private memory to global memory
3180 * (if "read" is not set) for the array reference group "group" that
3181 * is mapped to private memory.
3182 * On input, "node" points to the kernel node, and it is moved
3183 * back there on output.
3185 * The copies are performed in the order of the array elements.
3186 * The copy statement instances include a reference to the outer
3187 * tile->depth dimensions of the kernel schedule for ease of
3188 * combining them with the group tiling.
3190 * That is, the extra schedule is of the form
3194 * where D corresponds to the outer tile->depth dimensions of
3195 * the kernel schedule and A to the global array.
3196 * This schedule is unrolled because registers are not addressable.
3198 * The copying is inserted in the schedule tree through an extension
3203 * where the extra domain elements type[D -> A] are those accessed
3205 * A filter is inserted on type[D -> A] to ensure that the element
3206 * is read/written by the same thread that needs the element.
3207 * This filter is obtained by applying
3211 * to the thread filter for the core statements.
3213 * The extension is inserted before the core computation in case of a read
3214 * and after the core computation in case of a write.
3215 * In the latter case, we also make sure that there is a synchronization
3216 * node after the write to global memory, unless this write is performed
3217 * at the outer level of the kernel.
3218 * In principle, this synchronization could be inserted higher
3219 * in the schedule tree depending on where the corresponding reads
3220 * from global memory are performed.
3222 static __isl_give isl_schedule_node
*add_copies_group_private(
3223 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3224 __isl_take isl_schedule_node
*node
, int read
)
3226 struct gpu_array_tile
*tile
;
3227 isl_union_map
*access
;
3228 isl_union_map
*prefix
;
3229 isl_union_set
*domain
;
3231 isl_multi_aff
*from_access
;
3232 isl_multi_pw_aff
*mpa
;
3233 isl_multi_union_pw_aff
*mupa
;
3234 isl_schedule_node
*graft
;
3235 isl_union_set
*filter
;
3239 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3240 tile
= gpu_array_ref_group_tile(group
);
3241 node
= gpu_tree_move_down_to_depth(node
, tile
->depth
, kernel
->core
);
3243 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3244 empty
= isl_union_map_is_empty(access
);
3245 if (empty
< 0 || empty
) {
3246 isl_union_map_free(access
);
3248 return isl_schedule_node_free(node
);
3249 return gpu_tree_move_up_to_kernel(node
);
3252 group
->array
->global
= 1;
3253 group
->local_array
->global
= 1;
3255 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3256 space
= isl_space_domain(isl_multi_aff_get_space(from_access
));
3257 access
= isl_union_map_preimage_range_multi_aff(access
, from_access
);
3259 filter
= isl_union_set_copy(kernel
->thread_filter
);
3260 filter
= isl_union_set_apply(filter
, isl_union_map_copy(access
));
3261 filter
= isl_union_set_detect_equalities(filter
);
3262 filter
= isl_union_set_coalesce(filter
);
3264 domain
= isl_union_map_range(access
);
3265 access
= isl_union_set_wrapped_domain_map(domain
);
3266 access
= isl_union_map_reverse(access
);
3267 access
= isl_union_map_coalesce(access
);
3268 graft
= isl_schedule_node_from_extension(access
);
3270 space
= isl_space_map_from_set(space
);
3271 mpa
= isl_multi_pw_aff_identity(space
);
3272 mpa
= isl_multi_pw_aff_range_factor_range(mpa
);
3273 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3275 graft
= isl_schedule_node_child(graft
, 0);
3276 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3277 graft
= unroll(graft
);
3279 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3281 graft
= isl_schedule_node_parent(graft
);
3284 node
= isl_schedule_node_graft_before(node
, graft
);
3286 node
= isl_schedule_node_graft_after(node
, graft
);
3287 if (kernel_depth
< tile
->depth
)
3288 node
= add_group_write_sync(node
, kernel
, group
, 0);
3291 node
= gpu_tree_move_up_to_kernel(node
);
3296 /* Add copy statements to the schedule tree of "node"
3297 * for reading from global memory to shared memory (if "read" is set) or
3298 * for writing back from shared memory to global memory
3299 * (if "read" is not set) for the array reference group "group" that
3300 * is mapped to shared memory.
3301 * On input, "node" points to the kernel node, and it is moved
3302 * back there on output.
3304 * The copies are performed in the order of the corresponding shared
3306 * The copy statement instances include a reference to the outer
3307 * tile->depth dimensions of the kernel schedule for ease of
3308 * combining them with the group tiling.
3310 * If we are performing a read from global memory to shared memory and
3311 * if the array involved is not a scalar, then we copy
3312 * the entire tile to shared memory. This may result in some extra
3313 * elements getting copied, but it should lead to simpler code
3314 * (which means that fewer registers may be needed) and less divergence.
3316 * Otherwise, we only copy the elements that will be read or have been written
3319 * That is, the extra schedule is of the form
3323 * where D corresponds to the outer tile->depth dimensions of
3324 * the kernel schedule, A to the global array and T is the corresponding
3325 * shared memory tile.
3327 * The copying is inserted in the schedule tree through an extension
3332 * where the extra domain elements type[D -> A] are those accessed
3333 * by the group. In the case of read from a non-scalar, this set
3334 * is replaced by the entire shared memory tile.
3336 * A filter is inserted on type[D -> A] to map the copy instances
3337 * to the threads. In particular, the thread identifiers are
3338 * equated to the position inside the shared memory tile (T)
3339 * modulo the block size.
3340 * We try to align the innermost tile dimension with the innermost
3341 * thread identifier (x) as a heuristic to improve coalescing.
3342 * In particular, if the dimension of the tile is greater than
3343 * the dimension of the block, then the schedule mapping to the tile
3344 * is broken up into two pieces and the filter is applied to the inner part.
3345 * If, on the other hand, the dimension of the tile is smaller than
3346 * the dimension of the block, then the initial thread identifiers
3347 * are equated to zero and the remaining thread identifiers are
3348 * matched to the memory tile.
3350 * The extension is inserted before the core computation in case of a read
3351 * and after the core computation in case of a write.
3352 * In the case of a read, we first need to make sure there is some
3353 * synchronization before the core computation such that we can put the read
3354 * from global memory to shared memory before that synchronization.
3355 * This ensures that all threads have finished copying into shared memory
3356 * before the shared memory is used.
3357 * We also need to make sure that there is a synchronization node after
3358 * the core computation to ensure that the next load into shared memory
3359 * only happens after all data has been used. There is no need for
3360 * this synchronization if we are at the outer level since then there
3361 * won't be a next load.
3362 * In the case of a write, we need to make sure there is some synchronization
3363 * after the core computation such taht we can put the write from shared
3364 * memory to global memory after that synchronization.
3365 * Unless we are at the outer level, we also need a synchronization node
3366 * after the write to ensure the data is saved to global memory
3367 * before the next iteration write to the same shared memory.
3368 * It also makes sure the data has arrived in global memory before
3369 * it is read in a subsequent iteration.
3371 static __isl_give isl_schedule_node
*add_copies_group_shared(
3372 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3373 __isl_take isl_schedule_node
*node
, int read
)
3375 struct gpu_array_tile
*tile
;
3376 isl_union_map
*access
;
3377 isl_union_set
*domain
;
3378 isl_union_set
*sync
;
3380 isl_multi_aff
*from_access
;
3381 isl_multi_pw_aff
*mpa
;
3382 isl_multi_union_pw_aff
*mupa
;
3383 isl_schedule_node
*graft
;
3384 isl_union_set
*filter
;
3389 tile
= gpu_array_ref_group_tile(group
);
3390 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3391 node
= gpu_tree_move_down_to_depth(node
, tile
->depth
, kernel
->core
);
3393 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3394 empty
= isl_union_map_is_empty(access
);
3395 if (empty
< 0 || empty
) {
3396 isl_union_map_free(access
);
3398 return isl_schedule_node_free(node
);
3399 return gpu_tree_move_up_to_kernel(node
);
3402 group
->array
->global
= 1;
3403 group
->local_array
->global
= 1;
3405 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3407 ma
= isl_multi_aff_copy(tile
->tiling
);
3408 ma
= isl_multi_aff_pullback_multi_aff(ma
,
3409 isl_multi_aff_copy(from_access
));
3410 mpa
= isl_multi_pw_aff_from_multi_aff(ma
);
3411 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3413 domain
= isl_union_map_range(access
);
3415 if (read
&& !gpu_array_is_scalar(group
->array
)) {
3417 isl_union_set_free(domain
);
3418 map
= group_tile(group
);
3419 domain
= isl_union_set_from_set(isl_map_wrap(map
));
3422 domain
= isl_union_set_preimage_multi_aff(domain
, from_access
);
3423 access
= isl_union_set_wrapped_domain_map(domain
);
3424 access
= isl_union_map_reverse(access
);
3425 access
= isl_union_map_coalesce(access
);
3426 graft
= isl_schedule_node_from_extension(access
);
3428 graft
= isl_schedule_node_child(graft
, 0);
3430 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3432 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0) {
3433 graft
= isl_schedule_node_band_split(graft
,
3434 tile
->n
- kernel
->n_block
);
3435 graft
= isl_schedule_node_child(graft
, 0);
3437 if (tile
->n
< kernel
->n_block
)
3438 skip
= kernel
->n_block
- tile
->n
;
3441 filter
= set_schedule_modulo(graft
, kernel
->thread_ids
,
3443 if (!kernel
->options
->wrap
)
3444 graft
= snap_band_to_sizes(graft
, kernel
->block_dim
+ skip
,
3446 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0)
3447 graft
= isl_schedule_node_parent(graft
);
3448 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3450 while (graft
&& isl_schedule_node_has_parent(graft
))
3451 graft
= isl_schedule_node_parent(graft
);
3454 if (kernel_depth
< tile
->depth
)
3455 node
= gpu_tree_ensure_sync_after_core(node
, kernel
);
3456 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3457 node
= isl_schedule_node_graft_before(node
, graft
);
3459 node
= gpu_tree_move_right_to_sync(node
, kernel
);
3460 node
= isl_schedule_node_graft_after(node
, graft
);
3461 if (kernel_depth
< tile
->depth
)
3462 node
= add_group_write_sync(node
, kernel
, group
, 1);
3465 node
= gpu_tree_move_up_to_kernel(node
);
3470 /* Check whether the array reference group "group" is mapped to
3471 * private or shared memory and, if so,
3472 * add copy statements to the schedule tree of "node"
3473 * for reading from global memory to private or shared memory
3474 * (if "read" is set) or for writing back from private or shared memory
3475 * to global memory (if "read" is not set) for this group.
3476 * On input, "node" points to the kernel node, and it is moved
3477 * back there on output.
3479 static __isl_give isl_schedule_node
*add_copies_group(
3480 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3481 __isl_take isl_schedule_node
*node
, int read
)
3483 enum ppcg_group_access_type type
;
3485 type
= gpu_array_ref_group_type(group
);
3486 if (type
== ppcg_access_private
)
3487 return add_copies_group_private(kernel
, group
, node
, read
);
3488 if (type
== ppcg_access_shared
)
3489 return add_copies_group_shared(kernel
, group
, node
, read
);
3493 /* For each array reference group that is mapped to private or shared memory,
3494 * add copy statements to the schedule tree of "node"
3495 * for reading from global memory to private or shared memory
3496 * and for writing back.
3497 * On input, "node" points to the kernel node, and it is moved
3498 * back there on output.
3500 static __isl_give isl_schedule_node
*add_copies(struct ppcg_kernel
*kernel
,
3501 __isl_take isl_schedule_node
*node
)
3505 for (i
= 0; i
< kernel
->n_array
; ++i
) {
3506 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
3508 for (j
= 0; j
< array
->n_group
; ++j
) {
3509 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3511 node
= add_copies_group(kernel
, group
, node
, 1);
3514 node
= add_copies_group(kernel
, group
, node
, 0);
3523 /* Mark all dimensions in the current band node atomic.
3525 static __isl_give isl_schedule_node
*atomic(__isl_take isl_schedule_node
*node
)
3527 return ppcg_set_schedule_node_type(node
, isl_ast_loop_atomic
);
3530 /* Mark "node" atomic, if it is a band node.
3531 * Do the same for all ancestors.
3532 * Return a pointer to "node" (in the updated schedule tree).
3534 static __isl_give isl_schedule_node
*atomic_ancestors(
3535 __isl_take isl_schedule_node
*node
)
3541 if (!isl_schedule_node_has_parent(node
))
3544 pos
= isl_schedule_node_get_child_position(node
);
3545 node
= isl_schedule_node_parent(node
);
3546 if (isl_schedule_node_get_type(node
) == isl_schedule_node_band
)
3547 node
= atomic(node
);
3548 node
= atomic_ancestors(node
);
3549 node
= isl_schedule_node_child(node
, pos
);
3554 /* Collect all write references that require synchronization.
3555 * "node" is assumed to point to the kernel node.
3556 * Each reference is represented by a universe set in a space
3560 * with S[i,j] the statement instance space and R[] the array reference.
3562 * This function should be called before block and thread filters are added.
3564 * Synchronization is needed after a write if there is a subsequent read
3565 * within the same block that may not be performed by the same thread.
3566 * There should not be any dependences between different blocks,
3567 * so we start with the flow dependences within the same kernel invocation
3568 * and we subtract from these those dependences that are mapped
3569 * to the same iteration of the bands where synchronization is inserted.
3570 * We do not remove pairs of instances that are known to map to
3571 * the same thread across different iterations of the intermediate
3572 * bands because the read may be performed by a different thread
3573 * than the one that needs the value if shared memory is involved.
3575 * We also consider all pairs of possible writes that access the same
3576 * memory location and that may be mapped to the same block but not
3577 * to the same iteration of the intermediate bands.
3578 * In theory, it would be possible for one thread to still be in
3579 * a previous iteration of a loop in these bands.
3580 * A write to global memory in this delayed thread could then overwrite
3581 * a write from another thread that has already moved on to
3582 * the next iteration.
3584 * After computing the above writes paired off with reads or writes
3585 * that depend on them, we project onto the domain writes.
3586 * Sychronization is needed after writes to global memory
3587 * through these references.
3589 static __isl_give isl_union_set
*compute_sync_writes(
3590 struct ppcg_kernel
*kernel
, __isl_keep isl_schedule_node
*node
)
3592 isl_union_map
*local
;
3593 isl_union_map
*may_writes
, *shared_access
;
3594 isl_union_map
*kernel_prefix
, *thread_prefix
;
3595 isl_union_map
*equal
;
3596 isl_union_set
*wrap
;
3597 isl_union_set
*domain
;
3599 domain
= isl_schedule_node_get_universe_domain(node
);
3600 kernel_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3601 node
= isl_schedule_node_copy(node
);
3602 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3603 thread_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3604 isl_schedule_node_free(node
);
3606 may_writes
= isl_union_map_copy(kernel
->prog
->scop
->tagged_may_writes
);
3607 may_writes
= isl_union_map_curry(may_writes
);
3608 may_writes
= isl_union_map_intersect_domain(may_writes
, domain
);
3609 may_writes
= isl_union_map_uncurry(may_writes
);
3610 shared_access
= isl_union_map_copy(may_writes
);
3611 shared_access
= isl_union_map_apply_range(shared_access
,
3612 isl_union_map_reverse(may_writes
));
3614 local
= isl_union_map_copy(kernel
->prog
->scop
->tagged_dep_flow
);
3615 local
= isl_union_map_union(local
, shared_access
);
3616 local
= isl_union_map_zip(local
);
3618 equal
= isl_union_map_apply_range(kernel_prefix
,
3619 isl_union_map_reverse(isl_union_map_copy(kernel_prefix
)));
3620 wrap
= isl_union_map_wrap(equal
);
3621 local
= isl_union_map_intersect_domain(local
, wrap
);
3622 equal
= isl_union_map_apply_range(thread_prefix
,
3623 isl_union_map_reverse(isl_union_map_copy(thread_prefix
)));
3624 wrap
= isl_union_map_wrap(equal
);
3625 local
= isl_union_map_subtract_domain(local
, wrap
);
3627 local
= isl_union_map_zip(local
);
3628 local
= isl_union_map_universe(local
);
3630 return isl_union_map_domain(local
);
3633 /* Group the domain elements into a single space, named kernelX,
3634 * with X the kernel sequence number "kernel_id".
3636 static __isl_give isl_schedule_node
*group_statements(
3637 __isl_take isl_schedule_node
*node
, int kernel_id
)
3645 snprintf(buffer
, sizeof(buffer
), "kernel%d", kernel_id
);
3646 id
= isl_id_alloc(isl_schedule_node_get_ctx(node
), buffer
, NULL
);
3647 return isl_schedule_node_group(node
, id
);
3650 /* Create a ppcg_kernel representing the domain instances that reach "node"
3651 * and insert a mark node pointing to the ppcg_kernel before "node".
3652 * The band that "node" points to is the band that needs to be mapped
3653 * to block identifiers. The band that needs to be mapped to thread
3654 * identifiers should be marked by a "thread" mark by the caller.
3655 * This mark is removed by this function.
3656 * If "scale" is set, then the band that "node" points to is scaled
3659 * Mark all outer band nodes as atomic to ensure each kernel is only
3661 * If the domain elements that reach "node" live in more than one space,
3662 * then group the domain elements into a single space, named kernelX,
3663 * with X the kernel sequence number.
3665 * Insert a guard node governing the kernel node to ensure that
3666 * no kernels with zero blocks are launched.
3668 * Insert a context node describing the block and thread
3669 * identifiers inside the kernel mark.
3670 * The context node needs to be inserted after the effective block size
3671 * has been determined such that the bounds on the thread identifiers
3672 * would reflect the effective block size.
3673 * Insert a filter node inside the context node mapping the statement
3674 * instances to block identifiers. In particular, the block identifiers
3675 * are equated to the partial schedule of band that was marked for mapping
3676 * to blocks modulo the grid size.
3677 * Insert a filter node inside the "thread" mark mapping the statement
3678 * instances to thread identifiers. In particular, the thread identifiers
3679 * are equated to the partial schedule of band that was marked for mapping
3680 * to threads modulo the block size.
3682 * Compute array reference groups for all arrays, set the local
3683 * array bounds based on the set of domain instances that reach
3684 * the kernel node, check the total amount of shared memory used
3685 * and compute all group tilings.
3686 * The array reference groups are computed after the block filter
3687 * has been inserted because it affects the mapping to shared or
3688 * private memory. This computation also requires the thread filter
3689 * (in the ppcg_kernel object), but this thread filter should not
3690 * have been added to the schedule tree yet since the computation
3691 * requires the schedule of the band that needs to be mapped to
3692 * threads before the privatization is applied.
3694 * If any array reference group requires the band mapped to threads
3695 * to be unrolled, then we perform the required unrolling.
3697 * We save a copy of the schedule that may influence the mappings
3698 * to shared or private memory in kernel->copy_schedule.
3700 * Finally, we add synchronization and copy statements to the schedule tree,
3701 * remove the "thread" mark and create representations for the local
3702 * variables in the kernel.
3704 * We keep a copy of the isl_id that points to the kernel to ensure
3705 * that the kernel does not get destroyed if the schedule node
3706 * is freed due to some error condition.
3708 static __isl_give isl_schedule_node
*create_kernel(struct gpu_gen
*gen
,
3709 __isl_take isl_schedule_node
*node
, int scale
,
3710 __isl_keep isl_multi_val
*sizes
)
3712 struct ppcg_kernel
*kernel
;
3714 isl_schedule_node
*node_thread
;
3715 isl_union_map
*host_schedule
;
3716 isl_set
*host_domain
;
3717 isl_union_set
*domain
;
3718 int single_statement
;
3720 kernel
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel
);
3721 kernel
= ppcg_kernel_create_local_arrays(kernel
, gen
->prog
);
3723 return isl_schedule_node_free(node
);
3725 domain
= isl_schedule_node_get_domain(node
);
3726 single_statement
= isl_union_set_n_set(domain
) == 1;
3728 kernel
->ctx
= gen
->ctx
;
3729 kernel
->prog
= gen
->prog
;
3730 kernel
->options
= gen
->options
;
3731 kernel
->context
= extract_context(node
, gen
->prog
);
3732 kernel
->core
= isl_union_set_universe(isl_union_set_copy(domain
));
3733 kernel
->arrays
= accessed_by_domain(isl_union_set_copy(domain
),
3735 kernel
->n_grid
= n_outer_coincidence(node
);
3736 node_thread
= isl_schedule_node_copy(node
);
3737 node_thread
= gpu_tree_move_down_to_thread(node_thread
, kernel
->core
);
3738 node_thread
= isl_schedule_node_child(node_thread
, 0);
3739 kernel
->n_block
= n_outer_coincidence(node_thread
);
3740 isl_schedule_node_free(node_thread
);
3741 kernel
->id
= gen
->kernel_id
++;
3742 read_grid_and_block_sizes(kernel
, gen
);
3744 kernel
->sync_writes
= compute_sync_writes(kernel
, node
);
3746 host_schedule
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3747 host_domain
= isl_set_from_union_set(isl_union_map_range(
3750 node
= atomic_ancestors(node
);
3752 id
= isl_id_alloc(gen
->ctx
, "kernel", kernel
);
3753 id
= isl_id_set_free_user(id
, &ppcg_kernel_free_wrap
);
3754 node
= isl_schedule_node_insert_mark(node
, isl_id_copy(id
));
3756 if (!single_statement
)
3757 node
= group_statements(node
, kernel
->id
);
3759 node
= isl_schedule_node_child(node
, 0);
3760 node
= split_band(node
, kernel
->n_grid
);
3761 kernel
->block_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3762 kernel
->n_grid
, "b");
3763 kernel
->block_filter
= set_schedule_modulo(node
, kernel
->block_ids
,
3765 kernel
->grid_size
= extract_grid_size(kernel
,
3766 isl_union_set_copy(domain
));
3767 if (!kernel
->options
->wrap
)
3768 node
= snap_band_to_sizes(node
, kernel
->grid_dim
,
3771 node
= scale_band(node
, isl_multi_val_copy(sizes
));
3772 node
= isl_schedule_node_parent(node
);
3773 if (!single_statement
)
3774 node
= isl_schedule_node_parent(node
);
3775 node
= insert_guard(node
, kernel
->context
, kernel
->grid_size
,
3777 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3778 node
= isl_schedule_node_child(node
, 0);
3779 node
= split_band(node
, kernel
->n_block
);
3780 kernel
->thread_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3781 kernel
->n_block
, "t");
3782 kernel
->thread_filter
= set_schedule_modulo(node
, kernel
->thread_ids
,
3784 if (extract_block_size(kernel
, domain
) < 0)
3785 node
= isl_schedule_node_free(node
);
3787 node
= gpu_tree_move_up_to_kernel(node
);
3788 node
= isl_schedule_node_child(node
, 0);
3789 node
= insert_context(kernel
, node
);
3790 node
= isl_schedule_node_child(node
, 0);
3791 node
= isl_schedule_node_insert_filter(node
,
3792 isl_union_set_copy(kernel
->block_filter
));
3794 node
= gpu_tree_move_up_to_kernel(node
);
3796 if (gpu_group_references(kernel
, node
) < 0)
3797 node
= isl_schedule_node_free(node
);
3798 localize_bounds(kernel
, host_domain
);
3799 isl_set_free(host_domain
);
3801 check_shared_memory_bound(kernel
);
3802 mark_global_arrays(kernel
);
3803 compute_group_tilings(kernel
);
3805 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3806 node
= isl_schedule_node_child(node
, 0);
3807 if (!kernel
->options
->wrap
)
3808 node
= snap_band_to_sizes(node
, kernel
->block_dim
,
3810 node
= isl_schedule_node_insert_filter(node
,
3811 isl_union_set_copy(kernel
->thread_filter
));
3812 if (kernel_requires_unroll(kernel
)) {
3813 node
= isl_schedule_node_child(node
, 0);
3814 node
= unroll(node
);
3817 node
= gpu_tree_move_up_to_thread(node
);
3818 kernel
->copy_schedule_dim
= isl_schedule_node_get_schedule_depth(node
);
3819 kernel
->copy_schedule
=
3820 isl_schedule_node_get_prefix_schedule_union_pw_multi_aff(node
);
3822 node
= gpu_tree_move_up_to_kernel(node
);
3824 node
= add_sync(kernel
, node
);
3825 node
= add_copies(kernel
, node
);
3827 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3828 node
= isl_schedule_node_delete(node
);
3830 node
= gpu_tree_move_up_to_kernel(node
);
3832 if (create_kernel_vars(kernel
) < 0)
3833 node
= isl_schedule_node_free(node
);
3835 if (!single_statement
)
3836 node
= isl_schedule_node_parent(node
);
3837 node
= isl_schedule_node_parent(node
);
3843 /* Insert a zero-dimensional permutable band at "node".
3845 static __isl_give isl_schedule_node
*insert_empty_permutable_band(
3846 __isl_take isl_schedule_node
*node
)
3849 isl_schedule
*schedule
;
3850 isl_union_set
*domain
;
3851 isl_multi_union_pw_aff
*mupa
;
3853 schedule
= isl_schedule_node_get_schedule(node
);
3854 domain
= isl_schedule_get_domain(schedule
);
3855 space
= isl_union_set_get_space(domain
);
3856 isl_union_set_free(domain
);
3857 isl_schedule_free(schedule
);
3859 space
= isl_space_set_from_params(space
);
3860 mupa
= isl_multi_union_pw_aff_zero(space
);
3861 node
= isl_schedule_node_insert_partial_schedule(node
, mupa
);
3862 node
= isl_schedule_node_band_set_permutable(node
, 1);
3867 /* If "node" is the outermost permutable band that can be mapped to block and
3868 * thread identifiers in its branch (or the root of a subtree with
3869 * no such outer bands),
3870 * then mark the band as such, attaching a ppcg_kernel to the mark.
3872 * If "node" is the root of a subtree without permutable bands,
3873 * then insert a zero-dimensional permutable band such that
3874 * we can assume that "node" always points to a band node.
3875 * This includes the case where "node" already points to a band node,
3876 * but one without any coincident dimension. In this case,
3877 * the extra node ensures that this original node does not get tiled.
3879 * Tile "node" using user specified tile sizes, after splitting the band
3880 * if the number of specified tile sizes is smaller than the dimension
3881 * of the band. Mark the point band of this tiling as the band that
3882 * needs to be mapped to threads.
3883 * Create a kernel representing the domain instances that reach "node" and
3884 * insert a mark node pointing to the ppcg_kernel before the band node.
3886 static __isl_give isl_schedule_node
*mark_outer_permutable(
3887 __isl_take isl_schedule_node
*node
, void *user
)
3889 struct gpu_gen
*gen
= user
;
3895 isl_multi_val
*sizes
;
3897 outer
= is_outer_tilable(node
);
3899 return isl_schedule_node_free(node
);
3903 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
||
3904 !isl_schedule_node_band_member_get_coincident(node
, 0))
3905 node
= insert_empty_permutable_band(node
);
3907 tile_len
= isl_schedule_node_band_n_member(node
);
3908 tile_size
= read_tile_sizes(gen
, &tile_len
);
3910 return isl_schedule_node_free(node
);
3911 if (tile_len
< isl_schedule_node_band_n_member(node
))
3912 node
= isl_schedule_node_band_split(node
, tile_len
);
3913 sizes
= construct_band_tiles_sizes(node
, tile_size
);
3914 node
= tile_band(node
, isl_multi_val_copy(sizes
));
3915 node
= isl_schedule_node_child(node
, 0);
3916 id
= isl_id_alloc(gen
->ctx
, "thread", NULL
);
3917 node
= isl_schedule_node_insert_mark(node
, id
);
3918 node
= isl_schedule_node_parent(node
);
3920 scale
= gen
->options
->scale_tile_loops
;
3921 node
= create_kernel(gen
, node
, scale
, sizes
);
3922 isl_multi_val_free(sizes
);
3928 /* Given a set or sequence node, return the union the filters of either all
3929 * (if "only_initial" is not set) or the initial (if "only_initial" is set)
3930 * direct subtrees that do not contain any suitably permutable bands
3931 * (according to subtree_has_permutable_bands).
3933 static __isl_give isl_union_set
*get_non_parallel_subtree_filters(
3934 __isl_keep isl_schedule_node
*node
, int only_initial
)
3937 isl_union_set
*filter
;
3940 n
= isl_schedule_node_n_children(node
);
3944 node
= isl_schedule_node_copy(node
);
3945 node
= isl_schedule_node_child(node
, 0);
3946 filter
= isl_schedule_node_filter_get_filter(node
);
3947 node
= isl_schedule_node_parent(node
);
3948 space
= isl_union_set_get_space(filter
);
3949 isl_union_set_free(filter
);
3950 filter
= isl_union_set_empty(space
);
3952 for (i
= 0; i
< n
; ++i
) {
3955 node
= isl_schedule_node_child(node
, i
);
3956 parallelism
= subtree_has_permutable_bands(node
);
3957 if (parallelism
< 0) {
3958 filter
= isl_union_set_free(filter
);
3959 } else if (!parallelism
) {
3960 isl_union_set
*filter_i
;
3961 filter_i
= isl_schedule_node_filter_get_filter(node
);
3962 filter
= isl_union_set_union(filter
, filter_i
);
3963 } else if (only_initial
)
3965 node
= isl_schedule_node_parent(node
);
3968 isl_schedule_node_free(node
);
3973 /* Given a set or sequence node, return the union of the filters of
3974 * the direct subtrees that do not contain any suitably permutable bands
3975 * (according to subtree_has_permutable_bands).
3977 static __isl_give isl_union_set
*get_all_non_parallel_subtree_filters(
3978 __isl_keep isl_schedule_node
*node
)
3980 return get_non_parallel_subtree_filters(node
, 0);
3983 /* Given a set or sequence node, return the union of the filters of
3984 * the initial direct subtrees that do not contain any suitably permutable
3985 * bands (according to subtree_has_permutable_bands).
3987 static __isl_give isl_union_set
*get_initial_non_parallel_subtree_filters(
3988 __isl_keep isl_schedule_node
*node
)
3990 return get_non_parallel_subtree_filters(node
, 1);
3993 /* Mark all variables that are accessed by the statement instances in "domain"
3994 * and that are local to "prog" as requiring a declaration in the host code.
3996 static int declare_accessed_local_variables(struct gpu_prog
*prog
,
3997 __isl_keep isl_union_set
*domain
)
3999 isl_union_set
*arrays
;
4002 if (!ppcg_scop_any_hidden_declarations(prog
->scop
))
4004 arrays
= accessed_by_domain(isl_union_set_copy(domain
), prog
);
4006 for (i
= 0; i
< prog
->n_array
; ++i
) {
4011 if (!prog
->array
[i
].local
)
4013 space
= isl_set_get_space(prog
->array
[i
].extent
);
4014 set
= isl_union_set_extract_set(arrays
, space
);
4015 empty
= isl_set_plain_is_empty(set
);
4020 prog
->array
[i
].declare_local
= 1;
4023 isl_union_set_free(arrays
);
4026 isl_union_set_free(arrays
);
4030 /* If "node" points to a set node, then separate its children
4031 * into subtrees that have suitably permutable bands and
4032 * those that do not.
4033 * Adjust the schedule tree in order to execute the second group
4034 * after the first group and return a pointer to the first group,
4035 * assuming there are any such subtrees.
4036 * If "node" points to a sequence node, then separate the initial
4037 * children that do not have suitably permutable bands and
4038 * return a pointer to the subsequence of children that do have such bands,
4039 * assuming there are any such subtrees.
4041 * In both cases, mark all local variables in "prog" that are accessed by
4042 * the group without permutable bands as requiring a declaration on the host.
4044 static __isl_give isl_schedule_node
*isolate_permutable_subtrees(
4045 __isl_take isl_schedule_node
*node
, struct gpu_prog
*prog
)
4047 isl_union_set
*filter
;
4048 enum isl_schedule_node_type type
;
4052 type
= isl_schedule_node_get_type(node
);
4053 if (type
== isl_schedule_node_set
) {
4054 filter
= get_all_non_parallel_subtree_filters(node
);
4056 node
= isl_schedule_node_free(node
);
4058 if (declare_accessed_local_variables(prog
, filter
) < 0)
4059 node
= isl_schedule_node_free(node
);
4060 node
= isl_schedule_node_order_after(node
, filter
);
4061 } else if (type
== isl_schedule_node_sequence
) {
4062 filter
= get_initial_non_parallel_subtree_filters(node
);
4064 node
= isl_schedule_node_free(node
);
4066 if (declare_accessed_local_variables(prog
, filter
) < 0)
4067 node
= isl_schedule_node_free(node
);
4068 node
= isl_schedule_node_order_before(node
, filter
);
4074 /* Replace any reference to an array element in the range of "copy"
4075 * by a reference to all array elements (defined by the extent of the array).
4077 static __isl_give isl_union_map
*approximate_copy_out(
4078 __isl_take isl_union_map
*copy
, struct gpu_prog
*prog
)
4083 res
= isl_union_map_empty(isl_union_map_get_space(copy
));
4085 for (i
= 0; i
< prog
->n_array
; ++i
) {
4088 isl_union_map
*copy_i
;
4089 isl_union_set
*extent
, *domain
;
4091 space
= isl_space_copy(prog
->array
[i
].space
);
4092 extent
= isl_union_set_from_set(isl_set_universe(space
));
4093 copy_i
= isl_union_map_copy(copy
);
4094 copy_i
= isl_union_map_intersect_range(copy_i
, extent
);
4095 set
= isl_set_copy(prog
->array
[i
].extent
);
4096 extent
= isl_union_set_from_set(set
);
4097 domain
= isl_union_map_domain(copy_i
);
4098 copy_i
= isl_union_map_from_domain_and_range(domain
, extent
);
4099 res
= isl_union_map_union(res
, copy_i
);
4102 isl_union_map_free(copy
);
4107 /* Insert "kernel" marks that point to a ppcg_kernel structure
4108 * in front of all outermost tilable band that (by construction)
4109 * have at least one parallel loop.
4111 static __isl_give isl_schedule_node
*mark_kernels(struct gpu_gen
*gen
,
4112 __isl_take isl_schedule_node
*node
)
4114 return isl_schedule_node_map_descendant_bottom_up(node
,
4115 &mark_outer_permutable
, gen
);
4118 /* Construct schedule constraints from the dependences in prog->scop and
4119 * the array order dependences in prog->array_order.
4121 * If live range reordering is allowed, then we need to make sure
4122 * that live ranges on arrays are not run in parallel since doing
4123 * so would require array expansion. We therefore add the array
4124 * order dependences to the coincidence dependences. Non-zero array
4125 * order dependences will then prevent a schedule dimension from being
4126 * considered parallel.
4127 * Live ranges derived from scalars are allowed to be run in parallel
4128 * since we force the scalars to be mapped to private memory in
4129 * check_scalar_live_ranges.
4130 * If live range reordering is allowed, then the false dependences
4131 * are not added to the validity constraints as that would prevent
4132 * reordering. Instead, the external false dependences that enforce that reads
4133 * from potentially live-in data precede any later write and
4134 * that writes of potentially live-out data follow any other earlier write
4135 * are added to the validity and the coincidence constraints.
4136 * The false dependences are still added to the proximity constraints
4137 * for consistency with the case where live range reordering is not allowed.
4138 * The coincidence constraints then consist of flow dependences,
4139 * external false dependences and array order dependences.
4140 * The independences can be filtered out from the first two sets.
4141 * They have already been filtered out from the array order dependences
4142 * on a per array basis in collect_order_dependences.
4143 * There is no need for a per array handling of the other two sets
4144 * as there should be no flow or external false dependence on local
4145 * variables that can be filtered out.
4147 static __isl_give isl_schedule_constraints
*construct_schedule_constraints(
4148 struct gpu_prog
*prog
)
4150 isl_union_set
*domain
;
4151 isl_union_map
*dep_raw
, *dep
;
4152 isl_union_map
*validity
, *proximity
, *coincidence
;
4153 isl_schedule_constraints
*sc
;
4155 domain
= isl_union_set_copy(prog
->scop
->domain
);
4156 sc
= isl_schedule_constraints_on_domain(domain
);
4157 sc
= isl_schedule_constraints_set_context(sc
,
4158 isl_set_copy(prog
->scop
->context
));
4159 if (prog
->scop
->options
->live_range_reordering
) {
4160 sc
= isl_schedule_constraints_set_conditional_validity(sc
,
4161 isl_union_map_copy(prog
->scop
->tagged_dep_flow
),
4162 isl_union_map_copy(prog
->scop
->tagged_dep_order
));
4163 proximity
= isl_union_map_copy(prog
->scop
->dep_flow
);
4164 validity
= isl_union_map_copy(proximity
);
4165 validity
= isl_union_map_union(validity
,
4166 isl_union_map_copy(prog
->scop
->dep_forced
));
4167 proximity
= isl_union_map_union(proximity
,
4168 isl_union_map_copy(prog
->scop
->dep_false
));
4169 coincidence
= isl_union_map_copy(validity
);
4170 coincidence
= isl_union_map_subtract(coincidence
,
4171 isl_union_map_copy(prog
->scop
->independence
));
4172 coincidence
= isl_union_map_union(coincidence
,
4173 isl_union_map_copy(prog
->array_order
));
4175 dep_raw
= isl_union_map_copy(prog
->scop
->dep_flow
);
4176 dep
= isl_union_map_copy(prog
->scop
->dep_false
);
4177 dep
= isl_union_map_union(dep
, dep_raw
);
4178 dep
= isl_union_map_coalesce(dep
);
4179 proximity
= isl_union_map_copy(dep
);
4180 coincidence
= isl_union_map_copy(dep
);
4183 sc
= isl_schedule_constraints_set_validity(sc
, validity
);
4184 sc
= isl_schedule_constraints_set_coincidence(sc
, coincidence
);
4185 sc
= isl_schedule_constraints_set_proximity(sc
, proximity
);
4187 if (prog
->scop
->options
->debug
->dump_schedule_constraints
)
4188 isl_schedule_constraints_dump(sc
);
4192 /* Compute an appropriate schedule based on the accesses in
4193 * gen->read and gen->write.
4195 * We derive schedule constraints from the dependences in gen->prog->scop
4196 * and then use isl to compute a schedule that has a parallel loop
4197 * in each tilable band.
4199 static __isl_give isl_schedule
*compute_schedule(struct gpu_gen
*gen
)
4201 isl_schedule_constraints
*sc
;
4202 isl_schedule
*schedule
;
4204 sc
= construct_schedule_constraints(gen
->prog
);
4205 schedule
= isl_schedule_constraints_compute_schedule(sc
);
4210 /* If the band node "node" has exactly one member then mark it permutable.
4212 static __isl_give isl_schedule_node
*band_set_permutable(
4213 __isl_take isl_schedule_node
*node
,
4214 __isl_keep isl_schedule_constraints
*sc
)
4216 if (isl_schedule_node_band_n_member(node
) == 1)
4217 node
= isl_schedule_node_band_set_permutable(node
, 1);
4222 /* Return the coincidence constraints between pairs of instances
4223 * that are scheduled together by the ancestors of "node".
4224 * That is, select those coincidence constraints that relate
4225 * pairs of instances that have the same value for the prefix schedule.
4226 * If the schedule depth is zero, then the prefix schedule does not
4227 * contain any information, so we intersect domain and range
4228 * of the schedule constraints with the reaching domain elements instead.
4230 static __isl_give isl_union_map
*get_local_coincidence(
4231 __isl_keep isl_schedule_node
*node
,
4232 __isl_keep isl_schedule_constraints
*sc
)
4234 isl_union_map
*coincidence
;
4235 isl_multi_union_pw_aff
*prefix
;
4236 isl_union_pw_multi_aff
*contraction
;
4238 coincidence
= isl_schedule_constraints_get_coincidence(sc
);
4239 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4240 if (isl_schedule_node_get_schedule_depth(node
) == 0) {
4241 isl_union_set
*domain
;
4243 domain
= isl_schedule_node_get_domain(node
);
4244 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4246 coincidence
= isl_union_map_intersect_domain(coincidence
,
4247 isl_union_set_copy(domain
));
4248 coincidence
= isl_union_map_intersect_range(coincidence
,
4253 prefix
= isl_schedule_node_get_prefix_schedule_multi_union_pw_aff(node
);
4254 prefix
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(prefix
,
4256 return isl_union_map_eq_at_multi_union_pw_aff(coincidence
, prefix
);
4259 /* For each member in the band node "node", determine whether
4260 * it is coincident with respect to the outer nodes and mark
4263 * That is, for each coincidence constraint between pairs
4264 * of instances that are scheduled together by the outer nodes,
4265 * check that domain and range are assigned the same value
4266 * by the band member. This test is performed by checking
4267 * that imposing the same value for the band member does not
4268 * remove any elements from the set of coincidence constraints.
4270 static __isl_give isl_schedule_node
*band_set_coincident(
4271 __isl_take isl_schedule_node
*node
,
4272 __isl_keep isl_schedule_constraints
*sc
)
4274 isl_union_map
*coincidence
;
4275 isl_union_pw_multi_aff
*contraction
;
4276 isl_multi_union_pw_aff
*partial
;
4279 coincidence
= get_local_coincidence(node
, sc
);
4281 partial
= isl_schedule_node_band_get_partial_schedule(node
);
4282 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4283 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4285 n
= isl_schedule_node_band_n_member(node
);
4286 for (i
= 0; i
< n
; ++i
) {
4287 isl_union_map
*coincidence_i
;
4288 isl_union_pw_aff
*upa
;
4289 isl_multi_union_pw_aff
*partial_i
;
4292 upa
= isl_multi_union_pw_aff_get_union_pw_aff(partial
, i
);
4293 partial_i
= isl_multi_union_pw_aff_from_union_pw_aff(upa
);
4294 coincidence_i
= isl_union_map_copy(coincidence
);
4295 coincidence_i
= isl_union_map_eq_at_multi_union_pw_aff(
4296 coincidence_i
, partial_i
);
4297 subset
= isl_union_map_is_subset(coincidence
, coincidence_i
);
4298 isl_union_map_free(coincidence_i
);
4302 node
= isl_schedule_node_band_member_set_coincident(node
, i
,
4306 node
= isl_schedule_node_free(node
);
4307 isl_multi_union_pw_aff_free(partial
);
4308 isl_union_map_free(coincidence
);
4313 /* If "node" is a band, then set its properties.
4315 * In particular, if the band has exactly one member, then mark it permutable.
4316 * Mark the band member coincident based on the coincidence constraints
4319 static __isl_give isl_schedule_node
*set_band_properties(
4320 __isl_take isl_schedule_node
*node
, void *user
)
4322 isl_schedule_constraints
*sc
= user
;
4324 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
4326 if (isl_schedule_node_band_n_member(node
) == 0)
4329 node
= band_set_permutable(node
, sc
);
4330 node
= band_set_coincident(node
, sc
);
4335 /* Return the original schedule with all bands marked permutable and
4336 * all band members marked coincident based on the coincidence constraints.
4337 * The bands are explicitly marked permutable so that they will be considered
4338 * by mark_outer_permutable.
4340 static __isl_give isl_schedule
*determine_properties_original_schedule(
4341 struct gpu_gen
*gen
)
4343 isl_schedule
*schedule
;
4344 isl_schedule_constraints
*sc
;
4346 schedule
= isl_schedule_copy(gen
->prog
->scop
->schedule
);
4347 sc
= construct_schedule_constraints(gen
->prog
);
4348 schedule
= isl_schedule_map_schedule_node_bottom_up(schedule
,
4349 &set_band_properties
, sc
);
4350 isl_schedule_constraints_free(sc
);
4355 /* Compute a schedule or determine the properties of the original schedule
4356 * depending on the value of the "reschedule" option.
4358 static __isl_give isl_schedule
*compute_or_set_properties(void *user
)
4360 struct gpu_gen
*gen
= user
;
4362 if (gen
->options
->reschedule
)
4363 return compute_schedule(gen
);
4365 return determine_properties_original_schedule(gen
);
4368 /* Obtain a schedule for the scop, by reading it from
4369 * a file, by computing one or by determining the properties
4370 * of the original schedule.
4372 static __isl_give isl_schedule
*get_schedule(struct gpu_gen
*gen
)
4374 return ppcg_get_schedule(gen
->ctx
, gen
->options
,
4375 &compute_or_set_properties
, gen
);
4378 /* Construct the string "<a>_<b>".
4380 static char *concat(isl_ctx
*ctx
, const char *a
, const char *b
)
4385 p
= isl_printer_to_str(ctx
);
4386 p
= isl_printer_print_str(p
, a
);
4387 p
= isl_printer_print_str(p
, "_");
4388 p
= isl_printer_print_str(p
, b
);
4389 s
= isl_printer_get_str(p
);
4390 isl_printer_free(p
);
4395 /* For each array in "prog" of which an element appears in "accessed" and
4396 * that is not a read only scalar, create a zero-dimensional universe set
4397 * of which the tuple id has name "<prefix>_<name of array>" and a user
4398 * pointer pointing to the array (gpu_array_info).
4400 * If the array is local to "prog", then make sure it will be declared
4403 * Return the list of these universe sets.
4405 static __isl_give isl_union_set_list
*create_copy_filters(struct gpu_prog
*prog
,
4406 const char *prefix
, __isl_take isl_union_set
*accessed
)
4410 isl_union_set_list
*filters
;
4413 filters
= isl_union_set_list_alloc(ctx
, 0);
4414 for (i
= 0; i
< prog
->n_array
; ++i
) {
4415 struct gpu_array_info
*array
= &prog
->array
[i
];
4417 isl_set
*accessed_i
;
4421 isl_union_set
*uset
;
4423 if (gpu_array_is_read_only_scalar(array
))
4426 space
= isl_space_copy(array
->space
);
4427 accessed_i
= isl_union_set_extract_set(accessed
, space
);
4428 empty
= isl_set_plain_is_empty(accessed_i
);
4429 isl_set_free(accessed_i
);
4431 filters
= isl_union_set_list_free(filters
);
4439 array
->declare_local
= 1;
4441 name
= concat(ctx
, prefix
, array
->name
);
4442 id
= name
? isl_id_alloc(ctx
, name
, array
) : NULL
;
4444 space
= isl_space_set_alloc(ctx
, 0, 0);
4445 space
= isl_space_set_tuple_id(space
, isl_dim_set
, id
);
4446 uset
= isl_union_set_from_set(isl_set_universe(space
));
4448 filters
= isl_union_set_list_add(filters
, uset
);
4450 isl_union_set_free(accessed
);
4455 /* Make sure that code for the statements in "filters" that
4456 * copy arrays to or from the device is only generated when
4457 * the size of the corresponding array is positive.
4458 * That is, add a set node underneath "graft" with "filters" as children
4459 * and for each child add a guard that the selects the parameter
4460 * values for which the corresponding array has a positive size.
4461 * The array is available in the user pointer of the statement identifier.
4462 * "depth" is the schedule depth of the position where "graft"
4465 static __isl_give isl_schedule_node
*insert_positive_size_guards(
4466 __isl_take isl_schedule_node
*graft
,
4467 __isl_take isl_union_set_list
*filters
, int depth
)
4471 graft
= isl_schedule_node_child(graft
, 0);
4472 graft
= isl_schedule_node_insert_set(graft
, filters
);
4473 n
= isl_schedule_node_n_children(graft
);
4474 for (i
= 0; i
< n
; ++i
) {
4475 isl_union_set
*filter
;
4476 isl_set
*domain
, *guard
;
4478 struct gpu_array_info
*array
;
4480 graft
= isl_schedule_node_child(graft
, i
);
4481 filter
= isl_schedule_node_filter_get_filter(graft
);
4482 domain
= isl_set_from_union_set(filter
);
4483 id
= isl_set_get_tuple_id(domain
);
4484 array
= isl_id_get_user(id
);
4486 isl_set_free(domain
);
4487 guard
= gpu_array_positive_size_guard(array
);
4488 guard
= isl_set_from_params(guard
);
4489 guard
= isl_set_add_dims(guard
, isl_dim_set
, depth
);
4490 graft
= isl_schedule_node_child(graft
, 0);
4491 graft
= isl_schedule_node_insert_guard(graft
, guard
);
4492 graft
= isl_schedule_node_parent(graft
);
4493 graft
= isl_schedule_node_parent(graft
);
4495 graft
= isl_schedule_node_parent(graft
);
4500 /* Create a graft for copying arrays to or from the device,
4501 * whenever the size of the array is strictly positive.
4502 * Each statement is called "<prefix>_<name of array>" and
4503 * the identifier has a user pointer pointing to the array.
4504 * The graft will be added at the position specified by "node".
4505 * "copy" contains the array elements that need to be copied.
4506 * Only arrays of which some elements need to be copied
4507 * will have a corresponding statement in the graph.
4508 * Note though that each such statement will copy the entire array.
4510 static __isl_give isl_schedule_node
*create_copy_device(struct gpu_prog
*prog
,
4511 __isl_keep isl_schedule_node
*node
, const char *prefix
,
4512 __isl_take isl_union_set
*copy
)
4517 isl_union_set
*all
, *domain
;
4518 isl_union_set_list
*filters
;
4519 isl_union_map
*extension
;
4520 isl_schedule_node
*graft
;
4523 depth
= isl_schedule_node_get_schedule_depth(node
);
4524 filters
= create_copy_filters(prog
, prefix
, copy
);
4525 all
= isl_union_set_list_union(isl_union_set_list_copy(filters
));
4527 space
= depth
< 0 ? NULL
: isl_space_set_alloc(ctx
, 0, depth
);
4528 domain
= isl_union_set_from_set(isl_set_universe(space
));
4529 extension
= isl_union_map_from_domain_and_range(domain
, all
);
4530 graft
= isl_schedule_node_from_extension(extension
);
4533 return isl_schedule_node_free(graft
);
4534 if (isl_union_set_list_n_union_set(filters
) == 0) {
4535 isl_union_set_list_free(filters
);
4539 return insert_positive_size_guards(graft
, filters
, depth
);
4542 /* Return (the universe spaces of) the arrays that are declared
4543 * inside the scop corresponding to "prog" and for which all
4544 * potential writes inside the scop form a subset of "domain".
4546 static __isl_give isl_union_set
*extract_local_accesses(struct gpu_prog
*prog
,
4547 __isl_keep isl_union_set
*domain
)
4550 isl_union_set
*local
;
4552 local
= isl_union_set_empty(isl_union_set_get_space(domain
));
4554 for (i
= 0; i
< prog
->n_array
; ++i
) {
4556 isl_union_map
*to_outer
;
4557 isl_union_map
*may_write
;
4558 isl_union_set
*write_domain
;
4559 isl_union_set
*fields
;
4562 if (!prog
->array
[i
].local
)
4565 set
= isl_set_universe(isl_space_copy(prog
->array
[i
].space
));
4566 to_outer
= isl_union_map_copy(prog
->to_outer
);
4567 to_outer
= isl_union_map_intersect_range(to_outer
,
4568 isl_union_set_from_set(isl_set_copy(set
)));
4569 fields
= isl_union_map_domain(to_outer
);
4570 may_write
= isl_union_map_copy(prog
->may_write
);
4571 may_write
= isl_union_map_intersect_range(may_write
, fields
);
4572 write_domain
= isl_union_map_domain(may_write
);
4573 subset
= isl_union_set_is_subset(write_domain
, domain
);
4574 isl_union_set_free(write_domain
);
4578 return isl_union_set_free(local
);
4579 } else if (subset
) {
4580 local
= isl_union_set_add_set(local
, set
);
4589 /* Internal data structure for node_may_persist.
4591 * "tagger" maps tagged iteration domains to the corresponding untagged
4594 * "may_persist_flow" is the set of all tagged dataflow dependences
4595 * with those dependences removed that either precede or follow
4596 * the kernel launch in a sequence.
4597 * "inner_band_flow" is the set of all tagged dataflow dependences
4598 * that are local to a given iteration of the outer band nodes
4599 * with respect to the current node.
4600 * "local_flow" is equal to "inner_band_flow", except that the domain
4601 * and the range have been intersected with intermediate filters
4602 * on children of sets or sequences.
4604 struct ppcg_may_persist_data
{
4605 isl_union_pw_multi_aff
*tagger
;
4607 isl_union_map
*local_flow
;
4608 isl_union_map
*inner_band_flow
;
4609 isl_union_map
*may_persist_flow
;
4612 /* Update the information in "data" based on the band ancestor "node".
4614 * In particular, we restrict the dependences in data->local_flow
4615 * to those dependence where the source and the sink occur in
4616 * the same iteration of the given band node.
4617 * We also update data->inner_band_flow to the new value of
4620 static int update_may_persist_at_band(__isl_keep isl_schedule_node
*node
,
4621 struct ppcg_may_persist_data
*data
)
4623 isl_multi_union_pw_aff
*partial
;
4624 isl_union_pw_multi_aff
*contraction
;
4625 isl_union_map
*flow
;
4627 if (isl_schedule_node_band_n_member(node
) == 0)
4630 partial
= isl_schedule_node_band_get_partial_schedule(node
);
4631 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4632 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4634 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4635 isl_union_pw_multi_aff_copy(data
->tagger
));
4637 flow
= data
->local_flow
;
4638 flow
= isl_union_map_eq_at_multi_union_pw_aff(flow
, partial
);
4639 data
->local_flow
= flow
;
4641 isl_union_map_free(data
->inner_band_flow
);
4642 data
->inner_band_flow
= isl_union_map_copy(data
->local_flow
);
4647 /* Given a set of local reaching domain elements "domain",
4648 * expand them to the corresponding leaf domain elements using "contraction"
4649 * and insert the array references tags using data->tagger.
4651 static __isl_give isl_union_set
*expand_and_tag(
4652 __isl_take isl_union_set
*domain
,
4653 __isl_take isl_union_pw_multi_aff
*contraction
,
4654 struct ppcg_may_persist_data
*data
)
4656 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4658 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4659 isl_union_pw_multi_aff_copy(data
->tagger
));
4663 /* Given a filter node that is the child of a set or sequence node,
4664 * restrict data->local_flow to refer only to those elements
4665 * in the filter of the node.
4666 * "contraction" maps the leaf domain elements of the schedule tree
4667 * to the corresponding domain elements at (the parent of) "node".
4669 static int filter_flow(__isl_keep isl_schedule_node
*node
,
4670 struct ppcg_may_persist_data
*data
,
4671 __isl_take isl_union_pw_multi_aff
*contraction
)
4673 isl_union_set
*filter
;
4674 isl_union_map
*flow
;
4676 flow
= data
->local_flow
;
4677 filter
= isl_schedule_node_filter_get_filter(node
);
4678 filter
= expand_and_tag(filter
, contraction
, data
);
4679 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(filter
));
4680 flow
= isl_union_map_intersect_range(flow
, filter
);
4681 data
->local_flow
= flow
;
4686 /* Given a filter node "node", collect the filters on all preceding siblings
4687 * (which are also filter nodes), add them to "filters" and return the result.
4689 static __isl_give isl_union_set
*add_previous_filters(
4690 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4692 isl_schedule_node
*sibling
;
4694 sibling
= isl_schedule_node_copy(node
);
4695 while (sibling
&& isl_schedule_node_has_previous_sibling(sibling
)) {
4696 isl_union_set
*filter
;
4698 sibling
= isl_schedule_node_previous_sibling(sibling
);
4699 filter
= isl_schedule_node_filter_get_filter(sibling
);
4700 filters
= isl_union_set_union(filters
, filter
);
4702 isl_schedule_node_free(sibling
);
4704 return isl_union_set_free(filters
);
4709 /* Given a filter node "node", collect the filters on all following siblings
4710 * (which are also filter nodes), add them to "filters" and return the result.
4712 static __isl_give isl_union_set
*add_next_filters(
4713 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4715 isl_schedule_node
*sibling
;
4717 sibling
= isl_schedule_node_copy(node
);
4718 while (sibling
&& isl_schedule_node_has_next_sibling(sibling
)) {
4719 isl_union_set
*filter
;
4721 sibling
= isl_schedule_node_next_sibling(sibling
);
4722 filter
= isl_schedule_node_filter_get_filter(sibling
);
4723 filters
= isl_union_set_union(filters
, filter
);
4725 isl_schedule_node_free(sibling
);
4727 return isl_union_set_free(filters
);
4732 /* Remove those flow dependences from data->may_persist_flow
4733 * that flow between elements of "domain" within the same iteration
4734 * of all outer band nodes.
4735 * "contraction" maps the leaf domain elements of the schedule tree
4736 * to the corresponding elements "domain".
4738 static void remove_external_flow(struct ppcg_may_persist_data
*data
,
4739 __isl_take isl_union_set
*domain
,
4740 __isl_keep isl_union_pw_multi_aff
*contraction
)
4742 isl_union_map
*flow
;
4744 contraction
= isl_union_pw_multi_aff_copy(contraction
);
4745 domain
= expand_and_tag(domain
, contraction
, data
);
4746 flow
= isl_union_map_copy(data
->local_flow
);
4747 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(domain
));
4748 flow
= isl_union_map_intersect_range(flow
, domain
);
4750 data
->may_persist_flow
= isl_union_map_subtract(data
->may_persist_flow
,
4754 /* Update the information in "data" based on the filter ancestor "node".
4755 * We only need to modify anything if the filter is the child
4756 * of a set or sequence node.
4758 * In the case of a sequence, we remove the dependences between
4759 * statement instances that are both executed either before or
4760 * after the subtree that will be mapped to a kernel, within
4761 * the same iteration of outer bands.
4763 * In both cases, we restrict data->local_flow to the current child.
4765 static int update_may_persist_at_filter(__isl_keep isl_schedule_node
*node
,
4766 struct ppcg_may_persist_data
*data
)
4768 enum isl_schedule_node_type type
;
4769 isl_schedule_node
*parent
;
4771 isl_union_pw_multi_aff
*contraction
;
4772 isl_union_set
*before
, *after
, *filter
;
4773 isl_union_map
*flow
;
4775 type
= isl_schedule_node_get_parent_type(node
);
4776 if (type
!= isl_schedule_node_sequence
&& type
!= isl_schedule_node_set
)
4779 parent
= isl_schedule_node_copy(node
);
4780 parent
= isl_schedule_node_parent(parent
);
4781 contraction
= isl_schedule_node_get_subtree_contraction(parent
);
4782 isl_schedule_node_free(parent
);
4784 if (type
== isl_schedule_node_set
)
4785 return filter_flow(node
, data
, contraction
);
4787 filter
= isl_schedule_node_filter_get_filter(node
);
4788 space
= isl_union_set_get_space(filter
);
4789 isl_union_set_free(filter
);
4790 before
= isl_union_set_empty(space
);
4791 after
= isl_union_set_copy(before
);
4792 before
= add_previous_filters(before
, node
);
4793 after
= add_next_filters(after
, node
);
4795 remove_external_flow(data
, before
, contraction
);
4796 remove_external_flow(data
, after
, contraction
);
4798 return filter_flow(node
, data
, contraction
);
4801 /* Update the information in "data" based on the ancestor "node".
4803 static isl_stat
update_may_persist_at(__isl_keep isl_schedule_node
*node
,
4806 struct ppcg_may_persist_data
*data
= user
;
4808 switch (isl_schedule_node_get_type(node
)) {
4809 case isl_schedule_node_error
:
4810 return isl_stat_error
;
4811 case isl_schedule_node_context
:
4812 case isl_schedule_node_domain
:
4813 case isl_schedule_node_expansion
:
4814 case isl_schedule_node_extension
:
4815 case isl_schedule_node_guard
:
4816 case isl_schedule_node_leaf
:
4817 case isl_schedule_node_mark
:
4818 case isl_schedule_node_sequence
:
4819 case isl_schedule_node_set
:
4821 case isl_schedule_node_band
:
4822 if (update_may_persist_at_band(node
, data
) < 0)
4823 return isl_stat_error
;
4825 case isl_schedule_node_filter
:
4826 if (update_may_persist_at_filter(node
, data
) < 0)
4827 return isl_stat_error
;
4834 /* Determine the set of array elements that may need to be perserved
4835 * by a kernel constructed from the subtree at "node".
4836 * This includes the set of array elements that may need to be preserved
4837 * by the entire scop (prog->may_persist) and the elements for which
4838 * there is a potential flow dependence that may cross a kernel launch.
4840 * To determine the second set, we start from all flow dependences.
4841 * From this set of dependences, we remove those that cannot possibly
4842 * require data to be preserved by a kernel launch.
4843 * In particular, we consider the following sets of dependences.
4844 * - dependences of which the write occurs inside the kernel.
4845 * If the data is needed outside the kernel, then it will
4846 * be copied out immediately after the kernel launch, so there
4847 * is no need for any special care.
4848 * - dependences of which the read occurs inside the kernel and the
4849 * corresponding write occurs inside the same iteration of the
4850 * outer band nodes. This means that the data is needed in
4851 * the first kernel launch after the write, which is already
4852 * taken care of by the standard copy-in. That is, the data
4853 * do not need to be preserved by any intermediate call to
4855 * - dependences of which the write and the read either both occur
4856 * before the kernel launch or both occur after the kernel launch,
4857 * within the same iteration of the outer band nodes with respect
4858 * to the sequence that determines the ordering of the dependence
4859 * and the kernel launch. Such flow dependences cannot cross
4860 * any kernel launch.
4862 * For the remaining (tagged) dependences, we take the domain
4863 * (i.e., the tagged writes) and apply the tagged access relation
4864 * to obtain the accessed data elements.
4865 * These are then combined with the elements that may need to be
4866 * preserved by the entire scop.
4868 static __isl_give isl_union_set
*node_may_persist(
4869 __isl_keep isl_schedule_node
*node
, struct gpu_prog
*prog
)
4871 struct ppcg_may_persist_data data
;
4872 isl_schedule_node
*root
;
4873 isl_union_pw_multi_aff
*contraction
;
4874 isl_union_set
*domain
;
4875 isl_union_set
*persist
;
4876 isl_union_map
*flow
, *local_flow
;
4878 data
.tagger
= prog
->scop
->tagger
;
4880 flow
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
4881 data
.local_flow
= isl_union_map_copy(flow
);
4882 data
.inner_band_flow
= isl_union_map_copy(flow
);
4883 data
.may_persist_flow
= flow
;
4884 if (isl_schedule_node_foreach_ancestor_top_down(node
,
4885 &update_may_persist_at
, &data
) < 0)
4886 data
.may_persist_flow
=
4887 isl_union_map_free(data
.may_persist_flow
);
4888 flow
= data
.may_persist_flow
;
4889 isl_union_map_free(data
.local_flow
);
4891 domain
= isl_schedule_node_get_domain(node
);
4892 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4893 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4895 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4896 isl_union_pw_multi_aff_copy(data
.tagger
));
4897 flow
= isl_union_map_subtract_domain(flow
, isl_union_set_copy(domain
));
4898 local_flow
= data
.inner_band_flow
;
4899 local_flow
= isl_union_map_intersect_range(local_flow
, domain
);
4900 flow
= isl_union_map_subtract(flow
, local_flow
);
4902 persist
= isl_union_map_domain(flow
);
4903 persist
= isl_union_set_apply(persist
,
4904 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4905 persist
= isl_union_set_union(persist
,
4906 isl_union_set_copy(prog
->may_persist
));
4911 /* Add nodes for copying outer arrays in and out of the device
4912 * before and after the subtree "node", which contains one or more kernels.
4913 * "domain" contains the original reaching domain elements before
4914 * the kernels were created, i.e., before the contraction that
4915 * may have been performed in creating the kernels has been applied.
4916 * "prefix" contains the prefix schedule at that point, in terms
4917 * of the same original reaching domain elements.
4919 * We first compute the sets of outer array elements that need
4920 * to be copied in and out and then graft in the nodes for
4921 * performing this copying.
4923 * In particular, for each array that is possibly written anywhere in
4924 * the subtree "node" and that may be used after "node"
4925 * or that may be visible outside the corresponding scop,
4926 * we copy out its entire extent.
4928 * Any array elements that is read without first being written inside
4929 * the subtree "node" needs to be copied in.
4930 * Furthermore, if there are any array elements that
4931 * are copied out, but that may not be written inside "node, then
4932 * they also need to be copied in to ensure that the value after execution
4933 * is the same as the value before execution, at least for those array
4934 * elements that may have their values preserved by the scop or that
4935 * may be written before "node" and read after "node".
4936 * In case the array elements are structures, we need to take into
4937 * account that all members of the structures need to be written
4938 * by "node" before we can avoid copying the data structure in.
4940 * Note that the may_write relation is intersected with the domain,
4941 * which has been intersected with the context.
4942 * This helps in those cases where the arrays are declared with a fixed size,
4943 * while the accesses are parametric and the context assigns a fixed value
4944 * to the parameters.
4946 * If an element from a local array is read without first being written,
4947 * then there is no point in copying it in since it cannot have been
4948 * written prior to the scop. Warn about the uninitialized read instead.
4950 static __isl_give isl_schedule_node
*add_to_from_device(
4951 __isl_take isl_schedule_node
*node
, __isl_take isl_union_set
*domain
,
4952 __isl_take isl_union_map
*prefix
, struct gpu_prog
*prog
)
4954 isl_union_set
*local
;
4955 isl_union_set
*to_device
, *from_device
, *may_persist
;
4956 isl_union_map
*may_write
, *must_write
, *copy_out
, *not_written
;
4957 isl_union_map
*read
, *copy_in
;
4958 isl_union_map
*tagged
;
4959 isl_union_map
*local_uninitialized
;
4960 isl_schedule_node
*graft
;
4962 tagged
= isl_union_map_copy(prog
->scop
->tagged_reads
);
4963 tagged
= isl_union_map_union(tagged
,
4964 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4966 may_write
= isl_union_map_copy(prog
->may_write
);
4967 may_write
= isl_union_map_intersect_domain(may_write
,
4968 isl_union_set_copy(domain
));
4969 may_write
= remove_local_accesses(prog
,
4970 isl_union_map_copy(tagged
), may_write
,
4971 isl_union_map_copy(prefix
), 0);
4972 may_write
= isl_union_map_apply_range(may_write
,
4973 isl_union_map_copy(prog
->to_outer
));
4974 may_write
= isl_union_map_apply_domain(may_write
,
4975 isl_union_map_copy(prefix
));
4976 may_write
= approximate_copy_out(may_write
, prog
);
4977 copy_out
= isl_union_map_copy(may_write
);
4978 may_write
= isl_union_map_apply_range(may_write
,
4979 isl_union_map_copy(prog
->to_inner
));
4980 must_write
= isl_union_map_copy(prog
->must_write
);
4981 must_write
= isl_union_map_apply_domain(must_write
,
4982 isl_union_map_copy(prefix
));
4983 may_persist
= node_may_persist(node
, prog
);
4984 may_write
= isl_union_map_intersect_range(may_write
, may_persist
);
4985 not_written
= isl_union_map_subtract(may_write
, must_write
);
4987 local
= extract_local_accesses(prog
, domain
);
4988 read
= isl_union_map_copy(prog
->read
);
4989 read
= isl_union_map_intersect_domain(read
, domain
);
4990 read
= remove_local_accesses(prog
, tagged
, read
,
4991 isl_union_map_copy(prefix
), 1);
4992 local
= isl_union_set_apply(local
, isl_union_map_copy(prog
->to_inner
));
4993 local_uninitialized
= isl_union_map_copy(prog
->scop
->live_in
);
4994 local_uninitialized
= isl_union_map_intersect_range(local_uninitialized
,
4996 local_uninitialized
= isl_union_map_intersect(local_uninitialized
,
4997 isl_union_map_copy(read
));
4998 if (!isl_union_map_is_empty(local_uninitialized
)) {
5000 "possibly uninitialized reads (not copied in):\n");
5001 isl_union_map_dump(local_uninitialized
);
5003 read
= isl_union_map_subtract(read
, local_uninitialized
);
5004 read
= isl_union_map_apply_domain(read
, prefix
);
5005 copy_in
= isl_union_map_union(read
, not_written
);
5006 copy_in
= isl_union_map_apply_range(copy_in
,
5007 isl_union_map_copy(prog
->to_outer
));
5009 graft
= create_copy_device(prog
, node
, "to_device",
5010 isl_union_map_range(copy_in
));
5011 node
= isl_schedule_node_graft_before(node
, graft
);
5012 graft
= create_copy_device(prog
, node
, "from_device",
5013 isl_union_map_range(copy_out
));
5014 node
= isl_schedule_node_graft_after(node
, graft
);
5019 /* Add nodes for initializing ("init_device") and clearing ("clear_device")
5020 * the device before and after "node".
5022 static __isl_give isl_schedule_node
*add_init_clear_device(
5023 __isl_take isl_schedule_node
*node
)
5027 isl_union_set
*domain
;
5028 isl_schedule_node
*graft
;
5030 ctx
= isl_schedule_node_get_ctx(node
);
5032 space
= isl_space_set_alloc(ctx
, 0, 0);
5033 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "init_device");
5034 domain
= isl_union_set_from_set(isl_set_universe(space
));
5035 graft
= isl_schedule_node_from_domain(domain
);
5037 node
= isl_schedule_node_graft_before(node
, graft
);
5039 space
= isl_space_set_alloc(ctx
, 0, 0);
5040 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "clear_device");
5041 domain
= isl_union_set_from_set(isl_set_universe(space
));
5042 graft
= isl_schedule_node_from_domain(domain
);
5044 node
= isl_schedule_node_graft_after(node
, graft
);
5049 /* Update "schedule" for mapping to a GPU device.
5051 * In particular, insert a context node, create kernels for
5052 * each outermost tilable band and introduce nodes for copying arrays
5053 * in and out of the device and for initializing and clearing the device.
5054 * If the child of the initial root points to a set node,
5055 * then children of this node that do not contain any tilable bands
5056 * are separated from the other children and are not mapped to
5059 * The GPU code is generated in a context where at least one
5060 * statement instance is executed. The corresponding guard is inserted
5061 * around the entire schedule.
5063 static __isl_give isl_schedule
*map_to_device(struct gpu_gen
*gen
,
5064 __isl_take isl_schedule
*schedule
)
5066 isl_schedule_node
*node
;
5069 isl_union_set
*domain
;
5070 isl_union_map
*prefix
;
5071 struct gpu_prog
*prog
;
5073 context
= isl_set_copy(gen
->prog
->context
);
5074 context
= isl_set_from_params(context
);
5075 schedule
= isl_schedule_insert_context(schedule
, context
);
5078 guard
= isl_union_set_params(isl_union_set_copy(prog
->scop
->domain
));
5079 prog
->context
= isl_set_intersect(prog
->context
, isl_set_copy(guard
));
5080 guard
= isl_set_from_params(guard
);
5082 node
= isl_schedule_get_root(schedule
);
5083 isl_schedule_free(schedule
);
5084 node
= isl_schedule_node_child(node
, 0);
5085 node
= isl_schedule_node_child(node
, 0);
5086 node
= isolate_permutable_subtrees(node
, gen
->prog
);
5087 domain
= isl_schedule_node_get_domain(node
);
5088 prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
5089 node
= mark_kernels(gen
, node
);
5090 node
= add_to_from_device(node
, domain
, prefix
, gen
->prog
);
5091 node
= isl_schedule_node_root(node
);
5092 node
= isl_schedule_node_child(node
, 0);
5093 node
= isl_schedule_node_child(node
, 0);
5094 node
= isl_schedule_node_insert_guard(node
, guard
);
5095 node
= isl_schedule_node_child(node
, 0);
5096 node
= add_init_clear_device(node
);
5097 schedule
= isl_schedule_node_get_schedule(node
);
5098 isl_schedule_node_free(node
);
5103 /* Internal data structure for extract_access.
5104 * "next_access" points to the end of a linked list that is extended
5105 * by extract_access.
5106 * "single_expression" is set if the access expressions belong to
5107 * an expression statement (i.e., a statement without internal control).
5108 * "any_to_outer" maps all intermediate arrays to their outer arrays.
5110 struct ppcg_extract_access_data
{
5111 struct gpu_stmt_access
**next_access
;
5112 int single_expression
;
5113 isl_union_map
*any_to_outer
;
5116 /* Given a tagged access relation to a single array "tagged", extract it
5117 * as a map, taking into account that the input may be empty.
5118 * If the access relation is empty, then it does not contain
5119 * any space information, so we try to recover it from the index
5121 * The space of the index expression is of the form I -> A,
5122 * with I the statement instances and A the array, or [I -> F] -> A,
5123 * with F the filters corresponding to arguments.
5124 * We first drop F, if present, obtaining I -> A.
5125 * Then we construct I -> R, with R the reference tag,
5126 * combine the two into I -> [R -> A] and uncurry to obtain
5127 * the final result [I -> R] -> A.
5128 * Note that the index expression may have a lower dimension
5129 * than that of the array, but this dimension is not used
5130 * if the access relation is empty.
5132 static __isl_give isl_map
*extract_single_tagged_access(
5133 __isl_take isl_union_map
*tagged
, __isl_keep pet_expr
*expr
)
5137 isl_space
*space
, *space2
;
5138 isl_multi_pw_aff
*index
;
5140 empty
= isl_union_map_is_empty(tagged
);
5144 return isl_map_from_union_map(tagged
);
5145 isl_union_map_free(tagged
);
5147 index
= pet_expr_access_get_index(expr
);
5148 space
= isl_multi_pw_aff_get_space(index
);
5149 isl_multi_pw_aff_free(index
);
5150 if (isl_space_domain_is_wrapping(space
))
5151 space
= isl_space_domain_factor_domain(space
);
5152 space2
= isl_space_copy(space
);
5153 space2
= isl_space_from_domain(isl_space_domain(space
));
5154 id
= pet_expr_access_get_ref_id(expr
);
5155 space2
= isl_space_set_tuple_id(space2
, isl_dim_out
, id
);
5156 space
= isl_space_range_product(space2
, space
);
5157 space
= isl_space_uncurry(space
);
5159 return isl_map_empty(space
);
5161 isl_union_map_free(tagged
);
5165 /* Extract a gpu_stmt_access from "expr", append it to the list
5166 * that ends in *data->next_access and update the end of the list.
5167 * If the access expression performs a write, then it is considered
5168 * exact only if it appears in a single expression statement and
5169 * if its may access relation is equal to its must access relation.
5171 * The combined set of may accesses may be union if member accesses
5172 * are involved, but the entire set is derived from a single reference and
5173 * therefore from a single index expression. These accesses therefore
5174 * all map to the same outer array.
5176 static int extract_access(__isl_keep pet_expr
*expr
, void *user
)
5178 struct ppcg_extract_access_data
*data
= user
;
5179 isl_union_map
*tagged
;
5180 struct gpu_stmt_access
*access
;
5181 isl_ctx
*ctx
= pet_expr_get_ctx(expr
);
5182 isl_multi_pw_aff
*index
;
5184 access
= isl_alloc_type(ctx
, struct gpu_stmt_access
);
5186 access
->next
= NULL
;
5187 access
->read
= pet_expr_access_is_read(expr
);
5188 access
->write
= pet_expr_access_is_write(expr
);
5189 tagged
= pet_expr_access_get_tagged_may_read(expr
);
5190 tagged
= isl_union_map_union(tagged
,
5191 pet_expr_access_get_tagged_may_write(expr
));
5192 tagged
= isl_union_map_apply_range(tagged
,
5193 isl_union_map_copy(data
->any_to_outer
));
5194 if (!access
->write
) {
5195 access
->exact_write
= 1;
5196 } else if (!data
->single_expression
) {
5197 access
->exact_write
= 0;
5199 isl_union_map
*must
, *may
;
5200 may
= isl_union_map_copy(tagged
);
5201 may
= isl_union_map_domain_factor_domain(may
);
5202 must
= pet_expr_access_get_must_write(expr
);
5203 access
->exact_write
= isl_union_map_is_equal(must
, may
);
5204 isl_union_map_free(must
);
5205 isl_union_map_free(may
);
5207 index
= pet_expr_access_get_index(expr
);
5208 access
->n_index
= isl_multi_pw_aff_dim(index
, isl_dim_out
);
5209 isl_multi_pw_aff_free(index
);
5210 access
->ref_id
= pet_expr_access_get_ref_id(expr
);
5211 access
->tagged_access
= extract_single_tagged_access(tagged
, expr
);
5212 access
->access
= isl_map_copy(access
->tagged_access
);
5213 access
->access
= isl_map_domain_factor_domain(access
->access
);
5215 *data
->next_access
= access
;
5216 data
->next_access
= &(*data
->next_access
)->next
;
5218 if (!access
->access
)
5224 /* Construct a linked list of gpu_stmt_access objects,
5225 * one for each access expression in the statement body.
5226 * "any_to_outer" maps all intermediate arrays to their outer arrays.
5228 static int pet_stmt_extract_accesses(struct gpu_stmt
*stmt
,
5229 __isl_keep isl_union_map
*any_to_outer
)
5231 struct ppcg_extract_access_data data
;
5233 stmt
->accesses
= NULL
;
5234 data
.next_access
= &stmt
->accesses
;
5235 data
.single_expression
=
5236 pet_tree_get_type(stmt
->stmt
->body
) == pet_tree_expr
;
5237 data
.any_to_outer
= any_to_outer
;
5238 return pet_tree_foreach_access_expr(stmt
->stmt
->body
,
5239 &extract_access
, &data
);
5242 /* Return an array of gpu_stmt representing the statements in "scop".
5244 static struct gpu_stmt
*extract_stmts(isl_ctx
*ctx
, struct ppcg_scop
*scop
,
5245 __isl_keep isl_set
*context
, __isl_keep isl_union_map
*any_to_outer
)
5248 struct gpu_stmt
*stmts
;
5250 stmts
= isl_calloc_array(ctx
, struct gpu_stmt
, scop
->pet
->n_stmt
);
5254 for (i
= 0; i
< scop
->pet
->n_stmt
; ++i
) {
5255 struct gpu_stmt
*s
= &stmts
[i
];
5257 s
->id
= isl_set_get_tuple_id(scop
->pet
->stmts
[i
]->domain
);
5258 s
->stmt
= scop
->pet
->stmts
[i
];
5259 if (pet_stmt_extract_accesses(s
, any_to_outer
) < 0)
5260 return free_stmts(stmts
, i
+ 1);
5266 /* Generate CUDA code for "scop" and print it to "p".
5267 * After generating an AST for the transformed scop as explained below,
5268 * we call "gen->print" to print the AST in the desired output format
5271 * If it turns out that it does not make sense to generate GPU code,
5272 * then we generate CPU code instead.
5274 * The declarations of the arrays that are visible outside of the scop
5275 * are printed outside of the code generated from the schedule,
5276 * because the generated code may involve a guard around the entire code.
5278 * We first compute a schedule that respects the dependences
5279 * of the original program and select the outermost bands
5280 * of tilable dimensions that have at least one parallel loop.
5281 * If the --load-schedule is specified, then the loaded schedule
5282 * is used instead of a computed schedule.
5284 * Each of these bands B is then tiled according to "tile" sizes, resulting
5285 * in two nested bands, with a kernel marker on top
5293 * We then split off at most 2 parallel dimensions from the T band and
5294 * at most 3 parallel dimension from the P band
5307 * A filter is introduced in front of T1 that maps the domain instances
5308 * to block identifiers. Similarly, a filter is introduced in front of P1
5309 * that maps the domain instances to thread identifiers.
5311 * For each iteration of the T2 band and for each array, we compute
5312 * the array elements accessed by that iteration, construct a rectangular
5313 * box around it and shift it to the origin. The result is used
5314 * as shared memory for the array.
5316 * Copying and synchronization statements are added to this schedule tree.
5317 * In principle, these are added in front of the P1 band, but some of
5318 * them may get hoisted up to higher levels.
5320 * The entire AST is then generated from the single resulting schedule tree.
5321 * During the generation the subtrees at kernel nodes (K) are saved
5322 * aside and replaced by kernel calls. The result is printed as host code
5323 * while the saved subtrees are printed as device code.
5325 static __isl_give isl_printer
*generate(__isl_take isl_printer
*p
,
5326 struct gpu_gen
*gen
, struct ppcg_scop
*scop
,
5327 struct ppcg_options
*options
)
5329 struct gpu_prog
*prog
;
5331 isl_schedule
*schedule
;
5335 return isl_printer_free(p
);
5337 ctx
= isl_printer_get_ctx(p
);
5338 prog
= gpu_prog_alloc(ctx
, scop
);
5340 return isl_printer_free(p
);
5343 schedule
= get_schedule(gen
);
5345 any_permutable
= has_any_permutable_node(schedule
);
5346 if (any_permutable
< 0 || !any_permutable
) {
5347 if (any_permutable
< 0)
5348 p
= isl_printer_free(p
);
5350 p
= print_cpu(p
, scop
, options
);
5351 isl_schedule_free(schedule
);
5353 schedule
= map_to_device(gen
, schedule
);
5354 gen
->tree
= generate_code(gen
, schedule
);
5355 p
= ppcg_set_macro_names(p
);
5356 p
= isl_ast_op_type_print_macro(isl_ast_op_fdiv_q
, p
);
5357 p
= ppcg_print_exposed_declarations(p
, prog
->scop
);
5358 p
= gen
->print(p
, gen
->prog
, gen
->tree
, &gen
->types
,
5360 isl_ast_node_free(gen
->tree
);
5363 gpu_prog_free(prog
);
5368 /* Wrapper around generate for use as a ppcg_transform callback.
5370 static __isl_give isl_printer
*generate_wrap(__isl_take isl_printer
*p
,
5371 struct ppcg_scop
*scop
, void *user
)
5373 struct gpu_gen
*gen
= user
;
5375 return generate(p
, gen
, scop
, gen
->options
);
5378 /* Transform the code in the file called "input" by replacing
5379 * all scops by corresponding GPU code and write the results to "out".
5381 int generate_gpu(isl_ctx
*ctx
, const char *input
, FILE *out
,
5382 struct ppcg_options
*options
,
5383 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
5384 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
5385 struct gpu_types
*types
, void *user
), void *user
)
5392 gen
.sizes
= extract_sizes_from_str(ctx
, options
->sizes
);
5393 gen
.options
= options
;
5396 gen
.print_user
= user
;
5398 gen
.types
.name
= NULL
;
5400 if (options
->debug
->dump_sizes
) {
5401 isl_space
*space
= isl_space_params_alloc(ctx
, 0);
5402 gen
.used_sizes
= isl_union_map_empty(space
);
5405 r
= ppcg_transform(ctx
, input
, out
, options
, &generate_wrap
, &gen
);
5407 if (options
->debug
->dump_sizes
) {
5408 isl_union_map_dump(gen
.used_sizes
);
5409 isl_union_map_free(gen
.used_sizes
);
5412 isl_union_map_free(gen
.sizes
);
5413 for (i
= 0; i
< gen
.types
.n
; ++i
)
5414 free(gen
.types
.name
[i
]);
5415 free(gen
.types
.name
);
5420 /* Compute the set of inner array elements that may have their values
5421 * preserved by "prog". In particular, collect the array elements of
5422 * arrays that are not local to "prog" and remove those elements that
5423 * are definitely killed or definitely written by "prog".
5425 static __isl_give isl_union_set
*compute_may_persist(struct gpu_prog
*prog
)
5428 isl_union_set
*may_persist
, *killed
;
5429 isl_union_map
*must_kill
;
5431 may_persist
= isl_union_set_empty(isl_set_get_space(prog
->context
));
5432 for (i
= 0; i
< prog
->n_array
; ++i
) {
5435 if (prog
->array
[i
].local
)
5438 extent
= isl_set_copy(prog
->array
[i
].extent
);
5439 may_persist
= isl_union_set_add_set(may_persist
, extent
);
5442 may_persist
= isl_union_set_intersect_params(may_persist
,
5443 isl_set_copy(prog
->context
));
5444 may_persist
= isl_union_set_apply(may_persist
,
5445 isl_union_map_copy(prog
->to_inner
));
5446 must_kill
= isl_union_map_copy(prog
->tagged_must_kill
);
5447 killed
= isl_union_map_range(must_kill
);
5448 must_kill
= isl_union_map_copy(prog
->must_write
);
5449 killed
= isl_union_set_union(killed
, isl_union_map_range(must_kill
));
5451 may_persist
= isl_union_set_subtract(may_persist
, killed
);
5455 struct gpu_prog
*gpu_prog_alloc(isl_ctx
*ctx
, struct ppcg_scop
*scop
)
5457 struct gpu_prog
*prog
;
5464 prog
= isl_calloc_type(ctx
, struct gpu_prog
);
5469 prog
->context
= isl_set_copy(scop
->context
);
5470 prog
->n_stmts
= scop
->pet
->n_stmt
;
5471 prog
->any_to_outer
= pet_scop_compute_outer_to_any(scop
->pet
);
5472 prog
->any_to_outer
= isl_union_map_reverse(prog
->any_to_outer
);
5473 space
= isl_union_map_get_space(prog
->any_to_outer
);
5474 space
= isl_space_set_from_params(space
);
5475 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
5476 space
= isl_space_map_from_set(space
);
5477 id
= isl_map_identity(space
);
5478 prog
->any_to_outer
= isl_union_map_add_map(prog
->any_to_outer
, id
);
5479 prog
->stmts
= extract_stmts(ctx
, scop
,
5480 prog
->context
, prog
->any_to_outer
);
5481 prog
->read
= isl_union_map_copy(scop
->reads
);
5482 prog
->may_write
= isl_union_map_copy(scop
->may_writes
);
5483 prog
->must_write
= isl_union_map_copy(scop
->must_writes
);
5484 prog
->tagged_must_kill
= isl_union_map_copy(scop
->tagged_must_kills
);
5485 prog
->to_inner
= pet_scop_compute_outer_to_inner(scop
->pet
);
5486 prog
->to_outer
= isl_union_map_copy(prog
->to_inner
);
5487 prog
->to_outer
= isl_union_map_reverse(prog
->to_outer
);
5490 return gpu_prog_free(prog
);
5492 if (collect_array_info(prog
) < 0)
5493 return gpu_prog_free(prog
);
5494 prog
->may_persist
= compute_may_persist(prog
);
5499 void *gpu_prog_free(struct gpu_prog
*prog
)
5503 free_array_info(prog
);
5504 free_stmts(prog
->stmts
, prog
->n_stmts
);
5505 isl_union_map_free(prog
->any_to_outer
);
5506 isl_union_map_free(prog
->to_outer
);
5507 isl_union_map_free(prog
->to_inner
);
5508 isl_union_map_free(prog
->read
);
5509 isl_union_map_free(prog
->may_write
);
5510 isl_union_map_free(prog
->must_write
);
5511 isl_union_map_free(prog
->tagged_must_kill
);
5512 isl_union_map_free(prog
->array_order
);
5513 isl_union_set_free(prog
->may_persist
);
5514 isl_set_free(prog
->context
);