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
)
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
);
180 bounds
= isl_alloc_array(prog
->ctx
, isl_pw_aff
*, n_index
);
184 info
->space
= isl_set_get_space(pa
->extent
);
185 info
->name
= strdup(name
);
186 info
->n_index
= n_index
;
187 info
->bound
= bounds
;
188 info
->linearize
= prog
->scop
->options
->linearize_device_arrays
;
190 info
->type
= strdup(pa
->element_type
);
191 info
->size
= pa
->element_size
;
192 info
->local
= pa
->declared
&& !pa
->exposed
;
193 info
->has_compound_element
= pa
->element_is_record
;
194 info
->read_only_scalar
= is_read_only_scalar(info
, prog
);
196 accessed
= isl_union_set_extract_set(arrays
,
197 isl_space_copy(info
->space
));
198 empty
= isl_set_is_empty(accessed
);
199 extent
= compute_extent(pa
, accessed
);
200 isl_set_free(accessed
);
201 info
->extent
= extent
;
204 info
->accessed
= !empty
;
205 for (i
= 0; i
< n_index
; ++i
) {
211 dom
= isl_set_copy(extent
);
212 dom
= isl_set_project_out(dom
, isl_dim_set
, i
+ 1,
214 dom
= isl_set_project_out(dom
, isl_dim_set
, 0, i
);
215 if (!isl_set_dim_has_upper_bound(dom
, isl_dim_set
, 0)) {
216 fprintf(stderr
, "unable to determine extent of '%s' "
217 "in dimension %d\n", info
->name
, i
);
218 dom
= isl_set_free(dom
);
220 bound
= isl_set_dim_max(dom
, 0);
221 dom
= isl_pw_aff_domain(isl_pw_aff_copy(bound
));
222 ls
= isl_local_space_from_space(isl_set_get_space(dom
));
223 one
= isl_aff_zero_on_domain(ls
);
224 one
= isl_aff_add_constant_si(one
, 1);
225 bound
= isl_pw_aff_add(bound
, isl_pw_aff_alloc(dom
, one
));
226 bound
= isl_pw_aff_gist(bound
, isl_set_copy(prog
->context
));
229 if (!isl_pw_aff_is_cst(bound
))
233 collect_references(prog
, info
);
238 /* Remove independence from the order constraints "order" on array "array".
239 * Since the pairs of iterations in the filter relation of an independence
240 * are guaranteed to be completely independent by the user, there is
241 * no need to ensure that live ranges are ordered along thong pairs.
242 * We make an exception for local variables, though, as the independence
243 * guarantee does not apply to those.
245 * The order constraints are used in two places.
246 * Those on scalars are used in check_scalar_live_ranges to check if
247 * we need to force the scalar to be private. Any non-local scalar
248 * should not be forced scalar if it only appears in independent loops.
249 * Those on non-scalars are added to the coincidence constraints
250 * in compute_schedule because we do not support any array expansion.
251 * Accesses to non-local arrays should not prevent a loop from being
252 * considered coincident so we should indeed remove those constraints
253 * from the order constraints.
255 static __isl_give isl_union_map
*remove_independences(struct gpu_prog
*prog
,
256 struct gpu_array_info
*array
, __isl_take isl_union_map
*order
)
260 for (i
= 0; i
< prog
->scop
->pet
->n_independence
; ++i
) {
261 struct pet_independence
*pi
= prog
->scop
->pet
->independences
[i
];
262 if (isl_union_set_contains(pi
->local
, array
->space
))
265 order
= isl_union_map_subtract(order
,
266 isl_union_map_copy(pi
->filter
));
272 /* For each array in "prog", store the (untagged) order dependences
273 * derived from the array in array->dep_order.
274 * In particular, consider all references that access the given array
275 * and take the order dependences that have one of these references
276 * as source. (Since an order dependence relates two references to
277 * the same array, the target of these order dependences will also
278 * be one of these references.)
279 * Additionally, store the union of these array->dep_order relations
280 * for all non-scalar arrays in prog->array_order.
282 void collect_order_dependences(struct gpu_prog
*prog
)
286 isl_union_map
*accesses
;
288 space
= isl_union_map_get_space(prog
->read
);
289 prog
->array_order
= isl_union_map_empty(space
);
291 accesses
= isl_union_map_copy(prog
->scop
->tagged_reads
);
292 accesses
= isl_union_map_union(accesses
,
293 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
294 accesses
= isl_union_map_universe(accesses
);
295 accesses
= isl_union_map_apply_range(accesses
,
296 isl_union_map_copy(prog
->to_outer
));
298 for (i
= 0; i
< prog
->n_array
; ++i
) {
299 struct gpu_array_info
*array
= &prog
->array
[i
];
302 isl_union_map
*order
;
304 set
= isl_set_universe(isl_space_copy(array
->space
));
305 uset
= isl_union_set_from_set(set
);
306 uset
= isl_union_map_domain(
307 isl_union_map_intersect_range(isl_union_map_copy(accesses
),
309 order
= isl_union_map_copy(prog
->scop
->tagged_dep_order
);
310 order
= isl_union_map_intersect_domain(order
, uset
);
311 order
= isl_union_map_zip(order
);
312 order
= isl_union_set_unwrap(isl_union_map_domain(order
));
313 order
= remove_independences(prog
, array
, order
);
314 array
->dep_order
= order
;
316 if (gpu_array_is_scalar(array
) && !array
->has_compound_element
)
319 prog
->array_order
= isl_union_map_union(prog
->array_order
,
320 isl_union_map_copy(array
->dep_order
));
323 isl_union_map_free(accesses
);
326 /* Construct a gpu_array_info for each array referenced by prog->scop and
327 * collect them in prog->array.
329 * The sizes are based on the extents and the set of possibly accessed
330 * elements by "prog".
331 * If there are any member accesses involved, then they are first mapped
332 * to the outer arrays of structs.
334 * If we are allowing live range reordering, then also set
335 * the dep_order field. Otherwise leave it NULL.
337 static int collect_array_info(struct gpu_prog
*prog
)
341 isl_union_set
*arrays
;
343 arrays
= isl_union_map_range(isl_union_map_copy(prog
->read
));
344 arrays
= isl_union_set_union(arrays
,
345 isl_union_map_range(isl_union_map_copy(prog
->may_write
)));
347 arrays
= isl_union_set_apply(arrays
,
348 isl_union_map_copy(prog
->to_outer
));
350 arrays
= isl_union_set_coalesce(arrays
);
352 prog
->n_array
= prog
->scop
->pet
->n_array
;
353 prog
->array
= isl_calloc_array(prog
->ctx
,
354 struct gpu_array_info
, prog
->n_array
);
356 for (i
= 0; i
< prog
->scop
->pet
->n_array
; ++i
)
357 if (extract_array_info(prog
, &prog
->array
[i
],
358 prog
->scop
->pet
->arrays
[i
], arrays
) < 0)
361 isl_union_set_free(arrays
);
363 if (prog
->scop
->options
->live_range_reordering
)
364 collect_order_dependences(prog
);
369 static void free_array_info(struct gpu_prog
*prog
)
373 for (i
= 0; i
< prog
->n_array
; ++i
) {
374 int n_index
= prog
->array
[i
].n_index
;
375 free(prog
->array
[i
].type
);
376 free(prog
->array
[i
].name
);
377 for (j
= 0; j
< n_index
; ++j
)
378 isl_pw_aff_free(prog
->array
[i
].bound
[j
]);
379 isl_space_free(prog
->array
[i
].space
);
380 isl_set_free(prog
->array
[i
].extent
);
381 free(prog
->array
[i
].bound
);
382 free(prog
->array
[i
].refs
);
383 isl_union_map_free(prog
->array
[i
].dep_order
);
388 /* Check if a gpu array is a scalar. A scalar is a value that is not stored
389 * as an array or through a pointer reference, but as a single data element.
390 * At the moment, scalars are represented as zero-dimensional arrays.
391 * Note that the single data element may be an entire structure.
393 int gpu_array_is_scalar(struct gpu_array_info
*array
)
395 return array
->n_index
== 0;
398 /* Is "array" a read-only scalar?
400 int gpu_array_is_read_only_scalar(struct gpu_array_info
*array
)
402 return array
->read_only_scalar
;
405 /* Return the set of parameter values for which the array has a positive
406 * size in all dimensions.
407 * If the sizes are only valid for some parameter values, then those
408 * constraints are also taken into account.
410 __isl_give isl_set
*gpu_array_positive_size_guard(struct gpu_array_info
*array
)
419 space
= isl_space_params(isl_space_copy(array
->space
));
420 guard
= isl_set_universe(space
);
422 for (i
= 0; i
< array
->n_index
; ++i
) {
424 isl_set
*guard_i
, *zero
;
426 bound
= isl_pw_aff_copy(array
->bound
[i
]);
427 guard_i
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(bound
));
428 zero
= isl_pw_aff_zero_set(bound
);
429 guard_i
= isl_set_subtract(guard_i
, zero
);
430 guard
= isl_set_intersect(guard
, guard_i
);
436 /* Internal data structure for extract_size_of_type.
437 * "type" specifies the name of the space that we want to extract.
438 * "res" is used to store the subset of that space.
440 struct ppcg_extract_size_data
{
445 /* This function is called for each set in a union_set.
446 * If the name of the set matches data->type, we store the
449 static int extract_size_of_type(__isl_take isl_set
*size
, void *user
)
451 struct ppcg_extract_size_data
*data
= user
;
454 name
= isl_set_get_tuple_name(size
);
455 if (name
&& !strcmp(name
, data
->type
)) {
464 /* Given a union map { kernel[i] -> *[...] },
465 * return the range in the space called "type" for the kernel with
466 * sequence number "id".
468 static __isl_give isl_set
*extract_sizes(__isl_keep isl_union_map
*sizes
,
469 const char *type
, int id
)
473 isl_union_set
*local_sizes
;
474 struct ppcg_extract_size_data data
= { type
, NULL
};
479 space
= isl_union_map_get_space(sizes
);
480 space
= isl_space_set_from_params(space
);
481 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
482 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
483 dom
= isl_set_universe(space
);
484 dom
= isl_set_fix_si(dom
, isl_dim_set
, 0, id
);
486 local_sizes
= isl_union_set_apply(isl_union_set_from_set(dom
),
487 isl_union_map_copy(sizes
));
488 isl_union_set_foreach_set(local_sizes
, &extract_size_of_type
, &data
);
489 isl_union_set_free(local_sizes
);
493 /* Given a singleton set, extract the first (at most *len) elements
494 * of the single integer tuple into *sizes and update *len if needed.
496 static void read_sizes_from_set(__isl_take isl_set
*set
, int *sizes
, int *len
)
504 dim
= isl_set_dim(set
, isl_dim_set
);
508 for (i
= 0; i
< *len
; ++i
) {
511 v
= isl_set_plain_get_val_if_fixed(set
, isl_dim_set
, i
);
514 sizes
[i
] = isl_val_get_num_si(v
);
521 /* Add the map { kernel[id] -> type[sizes] } to gen->used_sizes,
522 * if the option debug->dump_sizes is set.
524 static void set_used_sizes(struct gpu_gen
*gen
, const char *type
, int id
,
531 if (!gen
->options
->debug
->dump_sizes
)
534 space
= isl_union_map_get_space(gen
->used_sizes
);
535 space
= isl_space_set_from_params(space
);
536 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
537 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
538 space
= isl_space_from_domain(space
);
539 space
= isl_space_add_dims(space
, isl_dim_out
, len
);
540 space
= isl_space_set_tuple_name(space
, isl_dim_out
, type
);
542 map
= isl_map_universe(space
);
543 map
= isl_map_fix_si(map
, isl_dim_in
, 0, id
);
544 for (i
= 0; i
< len
; ++i
)
545 map
= isl_map_fix_si(map
, isl_dim_out
, i
, sizes
[i
]);
547 gen
->used_sizes
= isl_union_map_add_map(gen
->used_sizes
, map
);
550 /* Extract user specified "tile" sizes from the "sizes" command line option,
551 * defaulting to option->tile_size in each dimension.
552 * *tile_len contains the maximum number of tile sizes needed.
553 * Update *tile_len to the number of specified tile sizes, if any, and
554 * return a pointer to the tile sizes (or NULL on error).
555 * Add the effectively used sizes to gen->used_sizes.
557 static int *read_tile_sizes(struct gpu_gen
*gen
, int *tile_len
)
563 tile_size
= isl_alloc_array(gen
->ctx
, int, *tile_len
);
566 for (n
= 0; n
< *tile_len
; ++n
)
567 tile_size
[n
] = gen
->options
->tile_size
;
569 size
= extract_sizes(gen
->sizes
, "tile", gen
->kernel_id
);
570 read_sizes_from_set(size
, tile_size
, tile_len
);
571 set_used_sizes(gen
, "tile", gen
->kernel_id
, tile_size
, *tile_len
);
576 /* Extract user specified "block" sizes from the "sizes" command line option,
577 * after filling in some potentially useful defaults.
579 static void read_block_sizes(struct ppcg_kernel
*kernel
,
580 __isl_keep isl_union_map
*sizes
)
584 if (kernel
->n_block
> 3)
586 switch (kernel
->n_block
) {
588 kernel
->block_dim
[0] = 512;
591 kernel
->block_dim
[0] = 32;
592 kernel
->block_dim
[1] = 16;
595 kernel
->block_dim
[0] = 32;
596 kernel
->block_dim
[1] = 4;
597 kernel
->block_dim
[2] = 4;
601 size
= extract_sizes(sizes
, "block", kernel
->id
);
602 read_sizes_from_set(size
, kernel
->block_dim
, &kernel
->n_block
);
605 /* Extract user specified "grid" sizes from the "sizes" command line option,
606 * after filling in some potentially useful defaults.
608 static void read_grid_sizes(struct ppcg_kernel
*kernel
,
609 __isl_keep isl_union_map
*sizes
)
613 if (kernel
->n_grid
> 2)
615 switch (kernel
->n_grid
) {
617 kernel
->grid_dim
[0] = 32768;
620 kernel
->grid_dim
[0] = 256;
621 kernel
->grid_dim
[1] = 256;
625 size
= extract_sizes(sizes
, "grid", kernel
->id
);
626 read_sizes_from_set(size
, kernel
->grid_dim
, &kernel
->n_grid
);
629 /* Extract user specified grid and block sizes from the gen->sizes
630 * command line option after filling in some potentially useful defaults.
631 * Store the extracted sizes in "kernel".
632 * Add the effectively used sizes to gen->used_sizes.
634 static void read_grid_and_block_sizes(struct ppcg_kernel
*kernel
,
637 read_block_sizes(kernel
, gen
->sizes
);
638 read_grid_sizes(kernel
, gen
->sizes
);
639 set_used_sizes(gen
, "block", kernel
->id
,
640 kernel
->block_dim
, kernel
->n_block
);
641 set_used_sizes(gen
, "grid", kernel
->id
,
642 kernel
->grid_dim
, kernel
->n_grid
);
645 static void *free_stmts(struct gpu_stmt
*stmts
, int n
)
652 for (i
= 0; i
< n
; ++i
) {
653 struct gpu_stmt_access
*access
, *next
;
655 for (access
= stmts
[i
].accesses
; access
; access
= next
) {
657 isl_id_free(access
->ref_id
);
658 isl_map_free(access
->access
);
659 isl_map_free(access
->tagged_access
);
663 isl_id_free(stmts
[i
].id
);
670 /* Add parameters p[i] with identifiers "ids" to "set",
671 * with bounds to 0 <= p[i] < size[i].
673 __isl_give isl_set
*add_bounded_parameters(__isl_take isl_set
*set
,
674 int *size
, __isl_keep isl_id_list
*ids
)
679 len
= isl_id_list_n_id(ids
);
680 nparam
= isl_set_dim(set
, isl_dim_param
);
681 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
683 for (i
= 0; i
< len
; ++i
) {
686 id
= isl_id_list_get_id(ids
, i
);
687 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
688 set
= isl_set_lower_bound_si(set
, isl_dim_param
, nparam
+ i
, 0);
689 set
= isl_set_upper_bound_si(set
, isl_dim_param
,
690 nparam
+ i
, size
[i
] - 1);
696 /* Add "len" parameters p[i] with identifiers "ids" and intersect "set"
699 * { : 0 <= p[i] < size[i] }
701 * or an overapproximation.
703 static __isl_give isl_set
*add_bounded_parameters_dynamic(
704 __isl_take isl_set
*set
, __isl_keep isl_multi_pw_aff
*size
,
705 __isl_keep isl_id_list
*ids
)
712 len
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
713 nparam
= isl_set_dim(set
, isl_dim_param
);
714 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
716 for (i
= 0; i
< len
; ++i
) {
719 id
= isl_id_list_get_id(ids
, i
);
720 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
723 space
= isl_space_params(isl_set_get_space(set
));
724 ls
= isl_local_space_from_space(space
);
725 for (i
= 0; i
< len
; ++i
) {
726 isl_pw_aff
*param
, *size_i
, *zero
;
729 param
= isl_pw_aff_var_on_domain(isl_local_space_copy(ls
),
730 isl_dim_param
, nparam
+ i
);
732 size_i
= isl_multi_pw_aff_get_pw_aff(size
, i
);
733 bound
= isl_pw_aff_lt_set(isl_pw_aff_copy(param
), size_i
);
734 bound
= isl_set_from_basic_set(isl_set_simple_hull(bound
));
735 set
= isl_set_intersect_params(set
, bound
);
737 zero
= isl_pw_aff_zero_on_domain(isl_local_space_copy(ls
));
738 bound
= isl_pw_aff_ge_set(param
, zero
);
739 set
= isl_set_intersect_params(set
, bound
);
741 isl_local_space_free(ls
);
746 /* Return the union of all tagged access relations in the group.
748 static __isl_give isl_union_map
*group_tagged_access_relation(
749 struct gpu_array_ref_group
*group
)
752 isl_union_map
*access
;
754 access
= isl_union_map_empty(isl_map_get_space(group
->access
));
755 for (i
= 0; i
< group
->n_ref
; ++i
) {
758 map_i
= isl_map_copy(group
->refs
[i
]->tagged_access
);
759 access
= isl_union_map_union(access
,
760 isl_union_map_from_map(map_i
));
766 /* Return the extent of "array", recomputed from the bounds.
767 * The recomputed extent may be simpler than the original extent.
769 static __isl_give isl_set
*array_extent(struct gpu_array_info
*array
)
777 id
= isl_set_get_tuple_id(array
->extent
);
778 space
= isl_set_get_space(array
->extent
);
779 extent
= isl_set_universe(isl_space_copy(space
));
780 ls
= isl_local_space_from_space(space
);
781 for (i
= 0; i
< array
->n_index
; ++i
) {
787 extent
= isl_set_lower_bound_si(extent
, isl_dim_set
, i
, 0);
789 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
791 index
= isl_pw_aff_from_aff(aff
);
792 bound
= isl_pw_aff_copy(array
->bound
[i
]);
793 bound
= isl_pw_aff_from_range(bound
);
794 bound
= isl_pw_aff_add_dims(bound
, isl_dim_in
, array
->n_index
);
795 bound
= isl_pw_aff_set_tuple_id(bound
, isl_dim_in
,
797 lt
= isl_pw_aff_lt_set(index
, bound
);
798 extent
= isl_set_intersect(extent
, lt
);
800 isl_local_space_free(ls
);
806 /* Return a map from the first group->depth dimensions of the computed
807 * schedule to the array tile in
808 * global memory that corresponds to the shared memory copy.
810 * In particular, return a map
816 * tile_offset(i) <= a <= tile_offset(i) + tile_size - 1 (1)
820 * 0 <= a <= array_size - 1 (2)
822 * Note that if some stride has been detected (i.e., when
823 * group->shared_tile->bound[i].shift is set), then a in (1) refers
824 * to the shifted and scaled down version.
826 * Constraints (1) are obtained by mapping the size constraints on the
827 * shared/private memory tile back to the access relation.
828 * Constraints (2) are obtained from the (recomputed) extent.
830 static __isl_give isl_map
*group_tile(struct gpu_array_ref_group
*group
)
833 int n_index
= group
->array
->n_index
;
839 space
= isl_multi_aff_get_space(group
->shared_tile
->tiling
);
840 space
= isl_space_range(space
);
841 local
= isl_set_universe(space
);
842 for (i
= 0; i
< n_index
; ++i
) {
845 local
= isl_set_lower_bound_si(local
, isl_dim_set
, i
, 0);
846 bound
= isl_val_copy(group
->shared_tile
->bound
[i
].size
);
847 bound
= isl_val_sub_ui(bound
, 1);
848 local
= isl_set_upper_bound_val(local
, isl_dim_set
, i
, bound
);
850 local
= isl_set_preimage_multi_aff(local
,
851 isl_multi_aff_copy(group
->shared_tile
->tiling
));
852 tile
= isl_set_unwrap(local
);
853 extent
= array_extent(group
->array
);
854 tile
= isl_map_intersect_range(tile
, extent
);
859 /* Given a mapping "iterator_map" from the AST schedule to a domain,
860 * return the corresponding mapping from the AST schedule to
861 * to the outer kernel->shared_schedule_dim dimensions of
862 * the schedule computed by PPCG for this kernel.
864 * Note that kernel->shared_schedule_dim is at least as large as
865 * the largest depth of any array reference group associated to the kernel.
866 * This is needed as the returned schedule is used to extract a mapping
867 * to the outer group->depth dimensions in transform_index.
869 static __isl_give isl_pw_multi_aff
*compute_sched_to_shared(
870 struct ppcg_kernel
*kernel
, __isl_take isl_pw_multi_aff
*iterator_map
)
872 isl_union_pw_multi_aff
*upma
;
873 isl_pw_multi_aff
*pma
;
876 space
= isl_space_range(isl_pw_multi_aff_get_space(iterator_map
));
877 space
= isl_space_from_domain(space
);
878 space
= isl_space_add_dims(space
, isl_dim_out
,
879 kernel
->shared_schedule_dim
);
881 upma
= isl_union_pw_multi_aff_copy(kernel
->shared_schedule
);
882 pma
= isl_union_pw_multi_aff_extract_pw_multi_aff(upma
, space
);
883 isl_union_pw_multi_aff_free(upma
);
885 return isl_pw_multi_aff_pullback_pw_multi_aff(pma
, iterator_map
);
888 /* If max_shared_memory is not set to infinity (-1), then make
889 * sure that the total amount of shared memory required by the
890 * array reference groups mapped to shared memory by "kernel"
891 * is no larger than this maximum.
893 * We apply a greedy approach and discard (keep in global memory)
894 * those groups that would result in a total memory size that
895 * is larger than the maximum.
897 * This function should be called after any function that may
898 * affect the decision on whether to place a reference group
899 * in private, shared or global memory.
901 static void check_shared_memory_bound(struct ppcg_kernel
*kernel
)
904 isl_val
*left
, *size
;
906 if (kernel
->options
->max_shared_memory
< 0)
909 left
= isl_val_int_from_si(kernel
->ctx
,
910 kernel
->options
->max_shared_memory
);
912 for (i
= 0; i
< kernel
->n_array
; ++i
) {
913 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
915 for (j
= 0; j
< local
->n_group
; ++j
) {
916 struct gpu_array_ref_group
*group
;
918 group
= local
->groups
[j
];
919 if (group
->private_tile
)
921 if (!group
->shared_tile
)
924 size
= gpu_array_tile_size(group
->shared_tile
);
925 size
= isl_val_mul_ui(size
, local
->array
->size
);
927 if (isl_val_le(size
, left
)) {
928 left
= isl_val_sub(left
, size
);
934 gpu_array_tile_free(group
->shared_tile
);
941 /* Compute a tiling for all the array reference groups in "kernel".
943 static void compute_group_tilings(struct ppcg_kernel
*kernel
)
947 for (i
= 0; i
< kernel
->n_array
; ++i
) {
948 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
950 for (j
= 0; j
< array
->n_group
; ++j
)
951 gpu_array_ref_group_compute_tiling(array
->groups
[j
]);
955 /* Compute the size of a bounding box around the origin and "set",
956 * where "set" is assumed to contain only non-negative elements.
957 * In particular, compute the maximal value of "set" in each direction
960 static __isl_give isl_multi_pw_aff
*extract_size(__isl_take isl_set
*set
,
961 __isl_take isl_set
*context
)
964 isl_multi_pw_aff
*mpa
;
966 context
= isl_set_params(context
);
967 n
= isl_set_dim(set
, isl_dim_set
);
968 mpa
= isl_multi_pw_aff_zero(isl_set_get_space(set
));
969 for (i
= 0; i
< n
; ++i
) {
974 bound
= isl_set_dim_max(isl_set_copy(set
), i
);
975 bound
= isl_pw_aff_coalesce(bound
);
976 bound
= isl_pw_aff_gist(bound
, isl_set_copy(context
));
978 space
= isl_pw_aff_get_domain_space(bound
);
979 one
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
980 one
= isl_aff_add_constant_si(one
, 1);
981 bound
= isl_pw_aff_add(bound
, isl_pw_aff_from_aff(one
));
982 mpa
= isl_multi_pw_aff_set_pw_aff(mpa
, i
, bound
);
985 isl_set_free(context
);
990 /* Compute the effective grid size as a list of the sizes in each dimension.
992 * The grid size specified by the user or set by default
993 * in read_grid_sizes() and applied by the block filter,
994 * may be too large for the given code in the sense that
995 * it may contain blocks that don't need to execute anything.
996 * We therefore don't return this grid size, but instead the
997 * smallest grid size that ensures that all blocks that actually
998 * execute code are included in the grid.
1000 * We first extract a description of the grid, i.e., the possible values
1001 * of the block ids, from the domain elements in "domain" and
1002 * kernel->block_filter.
1003 * The block ids are parameters in kernel->block_filter.
1004 * We simply need to change them into set dimensions.
1006 * Then, for each block dimension, we compute the maximal value of the block id
1009 static __isl_give isl_multi_pw_aff
*extract_grid_size(
1010 struct ppcg_kernel
*kernel
, __isl_take isl_union_set
*domain
)
1015 domain
= isl_union_set_intersect(domain
,
1016 isl_union_set_copy(kernel
->block_filter
));
1017 grid
= isl_union_set_params(domain
);
1018 grid
= isl_set_from_params(grid
);
1019 grid
= isl_set_add_dims(grid
, isl_dim_set
, kernel
->n_grid
);
1020 for (i
= 0; i
< kernel
->n_grid
; ++i
) {
1024 id
= isl_id_list_get_id(kernel
->block_ids
, i
);
1025 pos
= isl_set_find_dim_by_id(grid
, isl_dim_param
, id
);
1028 grid
= isl_set_equate(grid
, isl_dim_param
, pos
, isl_dim_set
, i
);
1029 grid
= isl_set_project_out(grid
, isl_dim_param
, pos
, 1);
1032 return extract_size(grid
, isl_set_copy(kernel
->context
));
1035 /* Compute the size of a fixed bounding box around the origin and "set",
1036 * where "set" is assumed to contain only non-negative elements,
1037 * and store the results in "size".
1038 * In particular, compute the maximal value of "set" in each direction
1041 static void extract_fixed_size(__isl_take isl_set
*set
, int *size
)
1044 isl_local_space
*ls
;
1047 n
= isl_set_dim(set
, isl_dim_set
);
1048 ls
= isl_local_space_from_space(isl_set_get_space(set
));
1049 obj
= isl_aff_zero_on_domain(ls
);
1050 for (i
= 0; i
< n
; ++i
) {
1053 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 1);
1054 max
= isl_set_max_val(set
, obj
);
1055 size
[i
] = isl_val_get_num_si(max
) + 1;
1057 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 0);
1063 /* Compute the effective block size as a list of the sizes in each dimension
1064 * and store the sizes in kernel->block_dim.
1066 * The block size specified by the user or set by default
1067 * in read_block_sizes() and applied by the thread filter,
1068 * may be too large for the given code in the sense that
1069 * it may contain threads that don't need to execute anything.
1070 * We therefore update this block size in kernel->block_dim
1071 * to the smallest block size that ensures that all threads
1072 * that actually execute code are included in the block.
1074 * The possible values of the thread ids is obtained from
1075 * the domain elements "domain" and kernel->thread_filter.
1076 * The current implementation eliminates all parameters, ensuring
1077 * that the size is a fixed constant in each dimension.
1078 * In principle we could also compute parametric sizes.
1079 * We would have to make sure to project out all b%d and t%d parameters,
1082 static void extract_block_size(struct ppcg_kernel
*kernel
,
1083 __isl_take isl_union_set
*domain
)
1089 domain
= isl_union_set_intersect(domain
,
1090 isl_union_set_copy(kernel
->thread_filter
));
1091 block
= isl_union_set_params(domain
);
1092 block
= isl_set_from_params(block
);
1093 block
= isl_set_add_dims(block
, isl_dim_set
, kernel
->n_block
);
1094 for (i
= 0; i
< kernel
->n_block
; ++i
) {
1098 id
= isl_id_list_get_id(kernel
->thread_ids
, i
);
1099 pos
= isl_set_find_dim_by_id(block
, isl_dim_param
, id
);
1102 block
= isl_set_equate(block
, isl_dim_param
, pos
,
1105 nparam
= isl_set_dim(block
, isl_dim_param
);
1106 block
= isl_set_project_out(block
, isl_dim_param
, 0, nparam
);
1108 extract_fixed_size(block
, kernel
->block_dim
);
1111 struct ppcg_kernel
*ppcg_kernel_free(struct ppcg_kernel
*kernel
)
1118 isl_id_list_free(kernel
->block_ids
);
1119 isl_id_list_free(kernel
->thread_ids
);
1120 isl_multi_pw_aff_free(kernel
->grid_size
);
1121 isl_set_free(kernel
->context
);
1122 isl_union_set_free(kernel
->core
);
1123 isl_union_set_free(kernel
->arrays
);
1124 isl_space_free(kernel
->space
);
1125 isl_ast_node_free(kernel
->tree
);
1126 isl_union_set_free(kernel
->block_filter
);
1127 isl_union_set_free(kernel
->thread_filter
);
1128 isl_union_pw_multi_aff_free(kernel
->shared_schedule
);
1129 isl_union_set_free(kernel
->sync_writes
);
1131 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1132 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1134 for (j
= 0; j
< array
->n_group
; ++j
)
1135 gpu_array_ref_group_free(array
->groups
[j
]);
1136 free(array
->groups
);
1138 isl_pw_aff_list_free(array
->bound
);
1140 free(kernel
->array
);
1142 for (i
= 0; i
< kernel
->n_var
; ++i
) {
1143 free(kernel
->var
[i
].name
);
1144 isl_vec_free(kernel
->var
[i
].size
);
1153 /* Wrapper around ppcg_kernel_free for use as a isl_id_set_free_user callback.
1155 static void ppcg_kernel_free_wrap(void *user
)
1157 struct ppcg_kernel
*kernel
= user
;
1159 ppcg_kernel_free(kernel
);
1162 static void create_kernel_var(isl_ctx
*ctx
, struct gpu_array_ref_group
*group
,
1163 struct ppcg_kernel_var
*var
)
1166 struct gpu_array_tile
*tile
;
1170 var
->array
= group
->array
;
1172 tile
= group
->private_tile
;
1173 var
->type
= ppcg_access_private
;
1175 tile
= group
->shared_tile
;
1176 var
->type
= ppcg_access_shared
;
1179 p
= isl_printer_to_str(ctx
);
1180 p
= gpu_array_ref_group_print_name(group
, p
);
1181 var
->name
= isl_printer_get_str(p
);
1182 isl_printer_free(p
);
1184 var
->size
= isl_vec_alloc(ctx
, group
->array
->n_index
);
1186 for (j
= 0; j
< group
->array
->n_index
; ++j
)
1187 var
->size
= isl_vec_set_element_val(var
->size
, j
,
1188 isl_val_copy(tile
->bound
[j
].size
));
1191 static int create_kernel_vars(struct ppcg_kernel
*kernel
)
1196 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1197 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1199 for (j
= 0; j
< array
->n_group
; ++j
) {
1200 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1201 if (group
->private_tile
|| group
->shared_tile
)
1207 kernel
->var
= isl_calloc_array(kernel
->ctx
, struct ppcg_kernel_var
, n
);
1212 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1213 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1215 for (j
= 0; j
< array
->n_group
; ++j
) {
1216 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1217 if (!group
->private_tile
&& !group
->shared_tile
)
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. If the array has compound
1250 * elements then we skipped the construction of array reference groups.
1252 * The resulting sizes may be functions that are nowhere defined
1253 * in case the access function cannot possibly access anything inside
1254 * the kernel for some reason. If so, they are replaced by the zero
1255 * function. Since the access function cannot actually access anything,
1256 * there is no harm in printing the array sizes as zero.
1258 static void localize_bounds(struct ppcg_kernel
*kernel
,
1259 __isl_keep isl_set
*host_domain
)
1264 context
= isl_set_copy(host_domain
);
1265 context
= isl_set_params(context
);
1267 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1268 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
1269 isl_pw_aff_list
*bound
;
1272 if (local
->n_group
== 0 && !local
->array
->has_compound_element
)
1275 n_index
= local
->array
->n_index
;
1276 bound
= isl_pw_aff_list_alloc(kernel
->ctx
, n_index
);
1278 for (j
= 0; j
< n_index
; ++j
) {
1282 pwaff
= isl_pw_aff_copy(local
->array
->bound
[j
]);
1283 pwaff
= isl_pw_aff_gist(pwaff
, isl_set_copy(context
));
1284 empty
= isl_pw_aff_is_empty(pwaff
);
1286 pwaff
= isl_pw_aff_free(pwaff
);
1288 pwaff
= set_universally_zero(pwaff
);
1289 bound
= isl_pw_aff_list_add(bound
, pwaff
);
1292 local
->n_index
= n_index
;
1293 local
->bound
= bound
;
1295 isl_set_free(context
);
1298 /* Create the array of gpu_local_array_info structures "array"
1299 * inside "kernel". The number of elements in this array is
1300 * the same as the number of arrays in "prog".
1301 * Initialize the "array" field of each local array to point
1302 * to the corresponding array in "prog".
1304 static struct ppcg_kernel
*ppcg_kernel_create_local_arrays(
1305 struct ppcg_kernel
*kernel
, struct gpu_prog
*prog
)
1310 ctx
= isl_set_get_ctx(prog
->context
);
1311 kernel
->array
= isl_calloc_array(ctx
,
1312 struct gpu_local_array_info
, prog
->n_array
);
1314 return ppcg_kernel_free(kernel
);
1315 kernel
->n_array
= prog
->n_array
;
1317 for (i
= 0; i
< prog
->n_array
; ++i
)
1318 kernel
->array
[i
].array
= &prog
->array
[i
];
1323 /* Find the element in gen->stmt that has the given "id".
1324 * Return NULL if no such gpu_stmt can be found.
1326 static struct gpu_stmt
*find_stmt(struct gpu_prog
*prog
, __isl_keep isl_id
*id
)
1330 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
1331 if (id
== prog
->stmts
[i
].id
)
1335 return i
< prog
->n_stmts
? &prog
->stmts
[i
] : NULL
;
1338 void ppcg_kernel_stmt_free(void *user
)
1341 struct ppcg_kernel_stmt
*stmt
= user
;
1346 switch (stmt
->type
) {
1347 case ppcg_kernel_copy
:
1348 isl_ast_expr_free(stmt
->u
.c
.index
);
1349 isl_ast_expr_free(stmt
->u
.c
.local_index
);
1351 case ppcg_kernel_domain
:
1352 isl_id_to_ast_expr_free(stmt
->u
.d
.ref2expr
);
1354 case ppcg_kernel_sync
:
1361 /* Return the gpu_stmt_access in the list "accesses" that corresponds
1364 static struct gpu_stmt_access
*find_access(struct gpu_stmt_access
*accesses
,
1365 __isl_keep isl_id
*ref_id
)
1367 struct gpu_stmt_access
*access
;
1369 for (access
= accesses
; access
; access
= access
->next
)
1370 if (access
->ref_id
== ref_id
)
1376 /* Return the index of the array called "name" in the list of arrays.
1378 static int find_array_index(struct ppcg_kernel
*kernel
, const char *name
)
1382 for (i
= 0; i
< kernel
->n_array
; ++i
)
1383 if (!strcmp(name
, kernel
->array
[i
].array
->name
))
1389 /* Internal data structure for the index and AST expression transformation
1390 * callbacks for pet_stmt_build_ast_exprs.
1392 * "kernel" is the kernel for which are computing AST expressions and
1393 * may be NULL if we are not inside a kernel.
1394 * "accesses" is the list of gpu_stmt_access in the statement.
1395 * "iterator_map" expresses the statement iterators in terms of
1396 * the AST loop iterators.
1397 * "sched2shared" expresses the outer shared_schedule_dim dimensions of
1398 * the kernel schedule in terms of the AST loop iterators and
1399 * may be NULL if we are not inside a kernel.
1401 * The following fields are set in transform_index and used in transform_expr.
1402 * "array" is the array that is being accessed.
1403 * "global" is set if the global array is accessed (rather than
1404 * shared/private memory).
1405 * "local_array" refers to information on the array specialized
1406 * to the current kernel.
1408 struct ppcg_transform_data
{
1409 struct ppcg_kernel
*kernel
;
1410 struct gpu_stmt_access
*accesses
;
1411 isl_pw_multi_aff
*iterator_map
;
1412 isl_pw_multi_aff
*sched2shared
;
1414 struct gpu_array_info
*array
;
1416 struct gpu_local_array_info
*local_array
;
1419 /* Return a pointer to the gpu_array_ref_group in "local"
1420 * that contains the reference "access".
1421 * Return NULL if no such group can be found.
1423 static struct gpu_array_ref_group
*find_ref_group(
1424 struct gpu_local_array_info
*local
, struct gpu_stmt_access
*access
)
1428 for (i
= 0; i
< local
->n_group
; ++i
) {
1429 struct gpu_array_ref_group
*group
= local
->groups
[i
];
1431 for (j
= 0; j
< group
->n_ref
; ++j
)
1432 if (group
->refs
[j
] == access
)
1439 /* Index transformation callback for pet_stmt_build_ast_exprs.
1441 * "index" expresses the array indices in terms of statement iterators
1443 * We first reformulate "index" in terms of the AST loop iterators.
1444 * Then we check if we are accessing the global array or
1445 * a shared/private copy. In particular, if we are not inside a kernel
1446 * then we must be accessing a global array.
1447 * In the former case, we simply return
1448 * the updated index. If "index" is an affine expression rather
1449 * than an array access, then we also return the updated index here.
1451 * If no reference groups have been computed for the array,
1452 * then we can only be accessing the global array.
1454 * Otherwise, we apply the tiling to the index.
1455 * This tiling is of the form
1459 * where D corresponds to the outer group->depth dimensions of
1460 * the kernel schedule.
1461 * The index is of the form
1465 * We update the tiling to refer to the AST loop iterators
1469 * and modify index to keep track of those iterators
1473 * Combining these two yields a tiled index expression in terms
1474 * of the AST loop iterators
1478 static __isl_give isl_multi_pw_aff
*transform_index(
1479 __isl_take isl_multi_pw_aff
*index
, __isl_keep isl_id
*ref_id
,
1482 struct ppcg_transform_data
*data
= user
;
1483 struct gpu_stmt_access
*access
;
1484 struct gpu_array_ref_group
*group
;
1485 struct gpu_array_tile
*tile
;
1486 isl_pw_multi_aff
*iterator_map
;
1491 isl_multi_pw_aff
*tiling
;
1492 isl_pw_multi_aff
*pma
;
1493 isl_multi_pw_aff
*mpa
;
1494 isl_pw_multi_aff
*sched2depth
;
1498 iterator_map
= isl_pw_multi_aff_copy(data
->iterator_map
);
1499 index
= isl_multi_pw_aff_pullback_pw_multi_aff(index
, iterator_map
);
1504 access
= find_access(data
->accesses
, ref_id
);
1507 if (!isl_map_has_tuple_name(access
->access
, isl_dim_out
))
1510 name
= get_outer_array_name(access
->access
);
1511 i
= find_array_index(data
->kernel
, name
);
1513 isl_die(isl_multi_pw_aff_get_ctx(index
), isl_error_internal
,
1514 "cannot find array",
1515 return isl_multi_pw_aff_free(index
));
1516 data
->local_array
= &data
->kernel
->array
[i
];
1517 data
->array
= data
->local_array
->array
;
1519 group
= find_ref_group(data
->local_array
, access
);
1525 tile
= group
->private_tile
;
1527 tile
= group
->shared_tile
;
1528 data
->global
= !tile
;
1532 space
= isl_space_range(isl_multi_pw_aff_get_space(index
));
1533 space
= isl_space_map_from_set(space
);
1534 pma
= isl_pw_multi_aff_identity(space
);
1535 sched2depth
= isl_pw_multi_aff_copy(data
->sched2shared
);
1536 dim
= isl_pw_multi_aff_dim(sched2depth
, isl_dim_out
);
1537 sched2depth
= isl_pw_multi_aff_drop_dims(sched2depth
, isl_dim_out
,
1538 group
->depth
, dim
- group
->depth
);
1539 pma
= isl_pw_multi_aff_product(sched2depth
, pma
);
1540 tiling
= isl_multi_pw_aff_from_multi_aff(
1541 isl_multi_aff_copy(tile
->tiling
));
1542 tiling
= isl_multi_pw_aff_pullback_pw_multi_aff(tiling
, pma
);
1544 space
= isl_space_domain(isl_multi_pw_aff_get_space(index
));
1545 space
= isl_space_map_from_set(space
);
1546 mpa
= isl_multi_pw_aff_identity(space
);
1547 index
= isl_multi_pw_aff_range_product(mpa
, index
);
1548 index
= isl_multi_pw_aff_pullback_multi_pw_aff(tiling
, index
);
1553 /* Dereference "expr" by adding an index [0].
1554 * The original "expr" is assumed not to have any indices.
1556 * If "expr" is a member access, then the dereferencing needs
1557 * to be applied to the structure argument of this member access.
1559 static __isl_give isl_ast_expr
*dereference(__isl_take isl_ast_expr
*expr
)
1562 isl_ast_expr
*arg0
, *res
;
1563 isl_ast_expr_list
*list
;
1565 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1567 return isl_ast_expr_free(expr
);
1568 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1569 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1572 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1573 arg
= dereference(arg
);
1574 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1575 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1579 isl_ast_expr_free(arg0
);
1581 ctx
= isl_ast_expr_get_ctx(expr
);
1582 res
= isl_ast_expr_from_val(isl_val_zero(ctx
));
1583 list
= isl_ast_expr_list_from_ast_expr(res
);
1584 res
= isl_ast_expr_get_op_arg(expr
, 0);
1585 res
= isl_ast_expr_access(res
, list
);
1586 isl_ast_expr_free(expr
);
1591 /* Linearize the index expression "expr" based on the array bounds
1594 * That is, transform expression
1596 * A[i_0][i_1]...[i_n]
1600 * A[(..((i_0 * b_1 + i_1) ... ) * b_n + i_n]
1602 * where b_0, b_1, ..., b_n are the bounds on the array.
1604 * If the base of "expr" is a member access, then the linearization needs
1605 * to be applied to the structure argument of this member access.
1607 * In the base case, if "expr" has no arguments (other than the name of
1608 * the array), then we are passing an entire array to a function.
1609 * In this case, there is nothing to linearize.
1610 * Note that at this point an expression with no arguments can
1611 * only be an entire array because the scalar case and
1612 * the case of single struct are handled by the caller.
1614 * If the number of specified index expressions in "expr"
1615 * is smaller than the dimension of the accessed array,
1616 * then the missing i_j also do not appear in the linearized expression.
1617 * Furthermore, since such an expression does not refer to a single
1618 * element while the default linearized expression would refer to
1619 * a single element, we return the expression
1621 * A + (..((i_0 * b_1 + i_1) ... ) * b_n]
1623 * instead. Note that because of the special case handling above,
1624 * we can assume here that here that there is at least one index expression.
1626 __isl_give isl_ast_expr
*gpu_local_array_info_linearize_index(
1627 struct gpu_local_array_info
*array
, __isl_take isl_ast_expr
*expr
)
1634 isl_ast_expr_list
*list
;
1635 isl_ast_build
*build
;
1637 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1638 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1639 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1642 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1643 arg
= gpu_local_array_info_linearize_index(array
, arg
);
1644 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1645 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1649 isl_ast_expr_free(arg0
);
1651 if (isl_ast_expr_get_op_n_arg(expr
) == 1)
1654 ctx
= isl_ast_expr_get_ctx(expr
);
1655 context
= isl_set_universe(isl_space_params_alloc(ctx
, 0));
1656 build
= isl_ast_build_from_context(context
);
1658 n
= isl_ast_expr_get_op_n_arg(expr
);
1659 res
= isl_ast_expr_get_op_arg(expr
, 1);
1660 for (i
= 1; i
< array
->n_index
; ++i
) {
1661 isl_pw_aff
*bound_i
;
1662 isl_ast_expr
*expr_i
;
1664 bound_i
= isl_pw_aff_list_get_pw_aff(array
->bound
, i
);
1665 expr_i
= isl_ast_build_expr_from_pw_aff(build
, bound_i
);
1666 res
= isl_ast_expr_mul(res
, expr_i
);
1670 expr_i
= isl_ast_expr_get_op_arg(expr
, i
+ 1);
1671 res
= isl_ast_expr_add(res
, expr_i
);
1674 isl_ast_build_free(build
);
1676 if (1 + array
->n_index
> n
) {
1677 res
= isl_ast_expr_add(isl_ast_expr_get_op_arg(expr
, 0), res
);
1679 list
= isl_ast_expr_list_from_ast_expr(res
);
1680 res
= isl_ast_expr_get_op_arg(expr
, 0);
1681 res
= isl_ast_expr_access(res
, list
);
1684 isl_ast_expr_free(expr
);
1689 /* AST expression transformation callback for pet_stmt_build_ast_exprs.
1691 * If the AST expression refers to an array that is not accessed
1692 * at all, then this means the value of the expression is not used,
1693 * so we might as well print zero (NULL pointer) instead.
1695 * If the AST expression refers to a global scalar that is not
1696 * a read-only scalar, then its address was passed to the kernel and
1697 * we need to dereference it.
1699 * If the AST expression refers to an access to a global array,
1700 * then we linearize the access exploiting the bounds in data->local_array.
1702 static __isl_give isl_ast_expr
*transform_expr(__isl_take isl_ast_expr
*expr
,
1703 __isl_keep isl_id
*id
, void *user
)
1705 struct ppcg_transform_data
*data
= user
;
1709 if (!data
->array
->accessed
) {
1712 ctx
= isl_ast_expr_get_ctx(expr
);
1713 isl_ast_expr_free(expr
);
1714 return isl_ast_expr_from_val(isl_val_zero(ctx
));
1716 if (gpu_array_is_read_only_scalar(data
->array
))
1720 if (data
->array
->n_index
== 0)
1721 return dereference(expr
);
1722 if (!data
->array
->linearize
)
1725 return gpu_local_array_info_linearize_index(data
->local_array
, expr
);
1728 /* This function is called for each instance of a user statement
1729 * in the kernel "kernel", identified by "gpu_stmt".
1730 * "kernel" may be NULL if we are not inside a kernel.
1732 * We attach a struct ppcg_kernel_stmt to the "node", containing
1733 * a computed AST expression for each access, through an annotation
1735 * These AST expressions are computed from iterator_map,
1736 * which expresses the domain
1737 * elements in terms of the generated loops, and sched2shared,
1738 * which expresses the outer shared_schedule_dim dimensions of
1739 * the kernel schedule computed by PPCG in terms of the generated loops.
1741 static __isl_give isl_ast_node
*create_domain_leaf(
1742 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1743 __isl_keep isl_ast_build
*build
, struct gpu_stmt
*gpu_stmt
)
1745 struct ppcg_transform_data data
;
1746 struct ppcg_kernel_stmt
*stmt
;
1749 isl_pw_multi_aff
*sched2shared
;
1751 isl_pw_multi_aff
*iterator_map
;
1752 isl_union_map
*schedule
;
1756 ctx
= isl_ast_node_get_ctx(node
);
1758 stmt
= isl_calloc_type(ctx
, struct ppcg_kernel_stmt
);
1760 return isl_ast_node_free(node
);
1762 schedule
= isl_ast_build_get_schedule(build
);
1763 map
= isl_map_reverse(isl_map_from_union_map(schedule
));
1764 iterator_map
= isl_pw_multi_aff_from_map(map
);
1766 sched2shared
= compute_sched_to_shared(kernel
,
1767 isl_pw_multi_aff_copy(iterator_map
));
1769 sched2shared
= NULL
;
1771 stmt
->type
= ppcg_kernel_domain
;
1772 stmt
->u
.d
.stmt
= gpu_stmt
;
1774 data
.kernel
= kernel
;
1775 data
.accesses
= stmt
->u
.d
.stmt
->accesses
;
1776 data
.iterator_map
= iterator_map
;
1777 data
.sched2shared
= sched2shared
;
1778 stmt
->u
.d
.ref2expr
= pet_stmt_build_ast_exprs(stmt
->u
.d
.stmt
->stmt
,
1779 build
, &transform_index
, &data
,
1780 &transform_expr
, &data
);
1782 isl_pw_multi_aff_free(iterator_map
);
1783 isl_pw_multi_aff_free(sched2shared
);
1785 id
= isl_id_alloc(ctx
, "user", stmt
);
1786 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1787 return isl_ast_node_set_annotation(node
, id
);
1790 /* This function is called for each statement node in the AST
1791 * for copying to or from shared/private memory.
1792 * Attach a pointer to a ppcg_kernel_stmt representing the copy
1793 * statement to the node.
1794 * The statement name is "read" or "write", depending on whether we are
1795 * reading from global memory or writing to global memory.
1797 * The schedule is of the form
1801 * where D corresponds to the outer group->depth dimensions of
1802 * the kernel schedule, A to the global array and L to the outer
1803 * generated AST schedule.
1804 * We compute the inverse and strip off the type, resulting in
1808 * We combine this mapping with on the one hand the projection
1812 * and on the other hand the group tiling
1820 * and store the corresponding expressions in stmt->index and stmt->local_index,
1821 * where stmt points to the ppcg_kernel_stmt that is attached to the node.
1823 static __isl_give isl_ast_node
*create_access_leaf(struct ppcg_kernel
*kernel
,
1824 struct gpu_array_ref_group
*group
, __isl_take isl_ast_node
*node
,
1825 __isl_keep isl_ast_build
*build
)
1827 struct ppcg_kernel_stmt
*stmt
;
1828 struct gpu_array_tile
*tile
;
1833 isl_pw_multi_aff
*pma
, *pma2
;
1836 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1838 return isl_ast_node_free(node
);
1840 access
= isl_map_from_union_map(isl_ast_build_get_schedule(build
));
1841 type
= isl_map_get_tuple_name(access
, isl_dim_in
);
1842 stmt
->u
.c
.read
= !strcmp(type
, "read");
1843 access
= isl_map_reverse(access
);
1844 pma
= isl_pw_multi_aff_from_map(access
);
1845 pma
= isl_pw_multi_aff_reset_tuple_id(pma
, isl_dim_out
);
1847 space
= isl_space_range(isl_pw_multi_aff_get_space(pma
));
1848 space
= isl_space_unwrap(space
);
1849 pma2
= isl_pw_multi_aff_range_map(space
);
1850 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
,
1851 isl_pw_multi_aff_copy(pma
));
1852 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1853 stmt
->u
.c
.index
= expr
;
1855 tile
= gpu_array_ref_group_tile(group
);
1856 pma2
= isl_pw_multi_aff_from_multi_aff(
1857 isl_multi_aff_copy(tile
->tiling
));
1858 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
, pma
);
1859 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1860 stmt
->u
.c
.local_index
= expr
;
1862 stmt
->u
.c
.array
= group
->array
;
1863 stmt
->u
.c
.local_array
= group
->local_array
;
1864 stmt
->type
= ppcg_kernel_copy
;
1866 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1867 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1868 return isl_ast_node_set_annotation(node
, id
);
1871 /* Create a synchronization ppcg_kernel_stmt and
1872 * attach it to the node "node" representing the synchronization.
1874 static __isl_give isl_ast_node
*create_sync_leaf(
1875 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1876 __isl_keep isl_ast_build
*build
)
1878 struct ppcg_kernel_stmt
*stmt
;
1881 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1883 return isl_ast_node_free(node
);
1885 stmt
->type
= ppcg_kernel_sync
;
1886 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1887 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1888 return isl_ast_node_set_annotation(node
, id
);
1891 /* Internal data structure for at_domain.
1893 * "prog" represents the entire scop.
1894 * "kernel" points to the kernel to which the current schedule node
1895 * belongs. It is set by before_mark and reset by after_mark.
1896 * It may be NULL if we are outside any kernel.
1898 struct ppcg_at_domain_data
{
1899 struct gpu_prog
*prog
;
1900 struct ppcg_kernel
*kernel
;
1903 /* This function is called for each instance of a user statement
1904 * in the kernel. This may be one of the original user statements
1905 * or a statement introduced by PPCG.
1907 * We assume that the original user statements only have a name
1908 * and no user pointer. The statements introduced by PPCG
1909 * on the other hand all have a user pointer.
1911 * If the user statement is one of the original user statements
1912 * (one with no user pointer), then we call create_domain_leaf. Otherwise,
1913 * we check if it is a copy or synchronization statement and
1914 * call the appropriate functions.
1915 * Statements that copy an array to/from the device do not need
1916 * any further treatment.
1918 static __isl_give isl_ast_node
*at_domain(__isl_take isl_ast_node
*node
,
1919 __isl_keep isl_ast_build
*build
, void *user
)
1921 struct ppcg_at_domain_data
*data
= user
;
1922 isl_ast_expr
*expr
, *arg
;
1928 expr
= isl_ast_node_user_get_expr(node
);
1929 arg
= isl_ast_expr_get_op_arg(expr
, 0);
1930 id
= isl_ast_expr_get_id(arg
);
1931 name
= isl_id_get_name(id
);
1932 p
= isl_id_get_user(id
);
1933 isl_ast_expr_free(expr
);
1934 isl_ast_expr_free(arg
);
1937 struct gpu_stmt
*gpu_stmt
;
1939 gpu_stmt
= find_stmt(data
->prog
, id
);
1942 isl_die(data
->prog
->ctx
, isl_error_internal
,
1943 "statement not found",
1944 return isl_ast_node_free(node
));
1946 return create_domain_leaf(data
->kernel
, node
, build
, gpu_stmt
);
1949 is_sync
= gpu_tree_id_is_sync(id
, data
->kernel
);
1951 if (!prefixcmp(name
, "to_device_") || !prefixcmp(name
, "from_device_"))
1954 return isl_ast_node_free(node
);
1955 if (!strcmp(name
, "read") || !strcmp(name
, "write")) {
1956 struct gpu_array_ref_group
*group
= p
;
1957 return create_access_leaf(data
->kernel
, group
, node
, build
);
1960 isl_die(data
->prog
->ctx
, isl_error_internal
,
1961 "unknown statement type",
1962 return isl_ast_node_free(node
));
1963 return create_sync_leaf(data
->kernel
, node
, build
);
1966 /* Given a set of wrapped references "ref", return the corresponding
1967 * access relations based on the tagged access relations "tagged".
1969 * The elements of "ref" are of the form
1973 * with D an iteration domains and R a reference.
1974 * The elements of "tagged" are of the form
1980 * Extend "tagged" to include the iteration domain in the range, i.e.,
1982 * [D -> R] -> [D -> A]
1984 * apply the result to "ref" and then unwrap the resulting set
1985 * to obtain relations of the form
1989 static __isl_give isl_union_map
*wrapped_reference_to_access(
1990 __isl_take isl_union_set
*ref
, __isl_take isl_union_map
*tagged
)
1992 isl_union_map
*tag2access
;
1994 tag2access
= isl_union_map_copy(tagged
);
1995 tag2access
= isl_union_map_universe(tag2access
);
1996 tag2access
= isl_union_set_unwrap(isl_union_map_domain(tag2access
));
1997 tag2access
= isl_union_map_domain_map(tag2access
);
1998 tag2access
= isl_union_map_range_product(tag2access
, tagged
);
2000 ref
= isl_union_set_coalesce(ref
);
2001 ref
= isl_union_set_apply(ref
, tag2access
);
2003 return isl_union_set_unwrap(ref
);
2006 /* Given an access relation "access" from one or more array reference groups,
2007 * remove those reads if ("read" is 1) or writes (if "read" is 0)
2008 * that are only needed to communicate data within
2009 * the same iteration of "sched".
2010 * "tagged" contains all tagged access relations to all
2011 * the array reference groups accessed by "access" from statement
2012 * instances scheduled by "sched".
2014 * If the access is a read then it is either an element of
2016 * live_in union (range flow)
2018 * where live_in and flow may be overapproximations, or
2019 * it reads an uninitialized value (that is not live-in because
2020 * there is an intermediate kill) or it reads a value that was
2021 * written within the same (compound) statement instance.
2022 * If the access is a write then it is either an element of
2024 * live_out union (domain flow)
2026 * or it writes a value that is never read (and is not live-out
2027 * because of an intermediate kill) or only
2028 * within the same (compound) statement instance.
2029 * In both cases, the access relation is also a subset of
2030 * the group access relation.
2032 * The cases where an uninitialized value is read or a value is written
2033 * that is never read or where the dataflow occurs within a statement
2034 * instance are also considered local and may also be removed.
2036 * Essentially, we compute the intersection of "access" with either
2038 * live_in union (range non-local-flow)
2042 * live_out union (domain non-local-flow)
2044 * We first construct a relation "local"
2046 * [[D -> R] -> [D' -> R']]
2048 * of pairs of domain iterations accessing the reference group
2049 * and references in the group that are coscheduled by "sched".
2051 * If this relation does not intersect the dataflow dependences,
2052 * then there is nothing we can possibly remove, unless the dataflow
2053 * dependences themselves only relate a subset of the accesses.
2054 * In particular, the accesses may not be involved in any dataflow
2055 * dependences, either because they are uninitialized reads/dead writes
2056 * or because the dataflow occurs inside a statement instance.
2058 * Since the computation below may break up the access relation
2059 * into smaller pieces, we only perform the intersection with
2060 * the non-local dependent accesses if the local pairs
2061 * intersect the dataflow dependences. Otherwise, we intersect
2062 * with the universe of the non-local dependent accesses.
2063 * This should at least remove accesses from statements that
2064 * do not participate in any dependences.
2066 * In particular, we remove the "local" dataflow dependences from
2067 * the set of all dataflow dependences.
2068 * Note that if the potential dataflow dependences are an overapproximation
2069 * of the actual dataflow dependences, then the result remains an
2070 * overapproximation of the non-local dataflow dependences.
2071 * Copying to/from global memory is only needed for the references
2072 * in the domain/range of the result or for accesses that are live out/in
2073 * for the entire scop.
2075 * We therefore map the domain/range of the "external" relation
2076 * to the corresponding access relation and take the union with
2077 * the live out/in relation.
2079 static __isl_give isl_union_map
*remove_local_accesses(
2080 struct gpu_prog
*prog
, __isl_take isl_union_map
*tagged
,
2081 __isl_take isl_union_map
*access
, __isl_take isl_union_map
*sched
,
2085 isl_union_pw_multi_aff
*tagger
;
2086 isl_union_set
*domain
;
2087 isl_union_map
*local
, *external
;
2088 isl_union_set
*tag_set
;
2090 if (isl_union_map_is_empty(access
)) {
2091 isl_union_map_free(sched
);
2092 isl_union_map_free(tagged
);
2096 tagger
= isl_union_pw_multi_aff_copy(prog
->scop
->tagger
);
2097 domain
= isl_union_map_domain(isl_union_map_copy(tagged
));
2098 tagger
= isl_union_pw_multi_aff_intersect_domain(tagger
, domain
);
2099 sched
= isl_union_map_preimage_domain_union_pw_multi_aff(sched
, tagger
);
2101 local
= isl_union_map_apply_range(sched
,
2102 isl_union_map_reverse(isl_union_map_copy(sched
)));
2103 local
= isl_union_map_intersect(local
,
2104 isl_union_map_copy(prog
->scop
->tagged_dep_flow
));
2106 empty
= isl_union_map_is_empty(local
);
2108 external
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
2109 external
= isl_union_map_intersect_params(external
,
2110 isl_set_copy(prog
->scop
->context
));
2111 external
= isl_union_map_subtract(external
, local
);
2114 tag_set
= isl_union_map_range(external
);
2115 external
= wrapped_reference_to_access(tag_set
, tagged
);
2116 external
= isl_union_map_union(external
,
2117 isl_union_map_copy(prog
->scop
->live_in
));
2119 tag_set
= isl_union_map_domain(external
);
2120 external
= wrapped_reference_to_access(tag_set
, tagged
);
2121 external
= isl_union_map_union(external
,
2122 isl_union_map_copy(prog
->scop
->live_out
));
2126 external
= isl_union_map_free(external
);
2128 external
= isl_union_map_universe(external
);
2130 access
= isl_union_map_intersect(access
, external
);
2135 /* Given an access relation "access" from "group", remove those reads
2136 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
2137 * communicate data within the same iteration of the schedule at the
2138 * position where the copying of the group is inserted.
2139 * "node" points to this position, i.e., the depth at "node"
2140 * is equal to group->depth.
2142 * We extract a schedule that picks out the iterations of the outer
2143 * group->depth dimensions and call remove_local_accesses.
2145 static __isl_give isl_union_map
*remove_local_accesses_group(
2146 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2147 __isl_take isl_union_map
*access
, __isl_keep isl_schedule_node
*node
,
2150 isl_union_map
*sched
, *tagged
;
2152 if (isl_union_map_is_empty(access
))
2155 tagged
= group_tagged_access_relation(group
);
2156 sched
= isl_schedule_node_get_prefix_schedule_relation(node
);
2158 return remove_local_accesses(kernel
->prog
, tagged
, access
, sched
, read
);
2161 /* This function is called before the AST generator starts traversing
2162 * the schedule subtree of a node with mark "mark".
2164 * If the mark is called "kernel", store the kernel pointer in data->kernel
2165 * for use in at_domain.
2167 static int before_mark(__isl_keep isl_id
*mark
,
2168 __isl_keep isl_ast_build
*build
, void *user
)
2170 struct ppcg_at_domain_data
*data
= user
;
2174 if (!strcmp(isl_id_get_name(mark
), "kernel"))
2175 data
->kernel
= isl_id_get_user(mark
);
2179 /* This function is called after the AST generator has finished traversing
2180 * the schedule subtree of a mark node. "node" points to the corresponding
2183 * If the mark is called "kernel", then replace "node" by a user node
2184 * that "calls" the kernel, representing the launch of the kernel.
2185 * The original "node" is stored inside the kernel object so that
2186 * it can be used to print the device code.
2187 * Note that this assumes that a kernel is only launched once.
2188 * Also clear data->kernel.
2190 static __isl_give isl_ast_node
*after_mark(__isl_take isl_ast_node
*node
,
2191 __isl_keep isl_ast_build
*build
, void *user
)
2196 isl_ast_expr_list
*list
;
2197 struct ppcg_kernel
*kernel
;
2198 struct ppcg_at_domain_data
*data
= user
;
2200 ctx
= isl_ast_node_get_ctx(node
);
2201 id
= isl_ast_node_mark_get_id(node
);
2203 return isl_ast_node_free(node
);
2204 if (strcmp(isl_id_get_name(id
), "kernel") || !data
->kernel
) {
2208 kernel
= data
->kernel
;
2209 data
->kernel
= NULL
;
2210 kernel
->space
= isl_ast_build_get_schedule_space(build
);
2211 kernel
->tree
= isl_ast_node_mark_get_node(node
);
2212 isl_ast_node_free(node
);
2214 expr
= isl_ast_expr_from_id(isl_id_copy(id
));
2215 list
= isl_ast_expr_list_alloc(ctx
, 0);
2216 expr
= isl_ast_expr_call(expr
, list
);
2217 node
= isl_ast_node_alloc_user(expr
);
2218 node
= isl_ast_node_set_annotation(node
, id
);
2223 static int update_depth(__isl_keep isl_schedule_node
*node
, void *user
)
2228 if (isl_schedule_node_get_type(node
) != isl_schedule_node_leaf
)
2230 node_depth
= isl_schedule_node_get_schedule_depth(node
);
2231 if (node_depth
> *depth
)
2232 *depth
= node_depth
;
2237 /* Use isl to generate code for both the host and the device
2239 * The device code is marked by "kernel" mark nodes in the schedule tree,
2240 * containing a pointer to a ppcg_kernel object.
2241 * The returned AST only contains the AST for the host code.
2242 * The ASTs for the device code are embedded in ppcg_kernel objects
2243 * attached to the leaf nodes that call "kernel".
2245 static __isl_give isl_ast_node
*generate_code(struct gpu_gen
*gen
,
2246 __isl_take isl_schedule
*schedule
)
2248 struct ppcg_at_domain_data data
;
2249 isl_ast_build
*build
;
2251 isl_id_list
*iterators
;
2254 data
.prog
= gen
->prog
;
2258 if (isl_schedule_foreach_schedule_node(schedule
, &update_depth
,
2261 build
= isl_ast_build_alloc(gen
->prog
->ctx
);
2262 iterators
= ppcg_scop_generate_names(gen
->prog
->scop
, depth
, "c");
2263 build
= isl_ast_build_set_iterators(build
, iterators
);
2264 build
= isl_ast_build_set_at_each_domain(build
, &at_domain
, &data
);
2265 build
= isl_ast_build_set_before_each_mark(build
, &before_mark
, &data
);
2266 build
= isl_ast_build_set_after_each_mark(build
, &after_mark
, &data
);
2267 if (gen
->prog
->scop
->options
->debug
->dump_final_schedule
)
2268 isl_schedule_dump(schedule
);
2269 tree
= isl_ast_build_node_from_schedule(build
, schedule
);
2270 isl_ast_build_free(build
);
2275 __isl_give isl_union_map
*extract_sizes_from_str(isl_ctx
*ctx
, const char *str
)
2279 return isl_union_map_read_from_str(ctx
, str
);
2282 /* Can "node" be tiled and then mapped to block and thread identifiers?
2283 * That is, is it permutable with at least one coincident dimension?
2285 static int is_permutable(__isl_keep isl_schedule_node
*node
)
2290 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
2292 if (!isl_schedule_node_band_get_permutable(node
))
2294 if (isl_schedule_node_band_n_member(node
) < 1)
2296 if (!isl_schedule_node_band_member_get_coincident(node
, 0))
2302 /* A isl_schedule_foreach_schedule_node callback
2303 * for setting *any_permutable and aborting the search
2304 * if "node" is a permutable band with coincident dimensions.
2305 * Otherwise, continue searching.
2307 static int set_permutable(__isl_keep isl_schedule_node
*node
, void *user
)
2309 int *any_permutable
= user
;
2312 permutable
= is_permutable(node
);
2318 *any_permutable
= 1;
2323 /* Does "schedule" contain any permutable band with at least one coincident
2326 static int has_any_permutable_node(__isl_keep isl_schedule
*schedule
)
2328 int any_permutable
= 0;
2330 if (isl_schedule_foreach_schedule_node(schedule
, &set_permutable
,
2331 &any_permutable
) < 0 &&
2335 return any_permutable
;
2338 /* Is "node" a leaf or can it be tiled and then mapped to
2339 * block and thread identifiers?
2341 static int is_leaf_or_tilable(__isl_keep isl_schedule_node
*node
)
2343 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
2345 return is_permutable(node
);
2348 /* Is "node" the outermost node in its branch that can be tiled
2349 * and then mapped to block and thread identifiers?
2350 * If there are no such nodes in the branch and if "node" is a leaf,
2351 * then it is accepted too.
2353 static int is_outer_tilable(__isl_keep isl_schedule_node
*node
)
2356 isl_schedule_node
*ancestor
;
2358 tilable
= is_leaf_or_tilable(node
);
2365 ancestor
= isl_schedule_node_copy(node
);
2366 while (isl_schedule_node_has_parent(ancestor
)) {
2367 ancestor
= isl_schedule_node_parent(ancestor
);
2369 tilable
= is_permutable(ancestor
);
2370 if (tilable
< 0 || tilable
)
2374 isl_schedule_node_free(ancestor
);
2375 return tilable
< 0 ? -1 : !tilable
;
2378 /* Collect the references to all writes in "group".
2379 * Each reference is represented by a universe set in a space
2383 * with S[i,j] the statement instance space and R[] the array reference.
2385 static __isl_give isl_union_set
*group_tagged_writes(
2386 struct gpu_array_ref_group
*group
)
2390 isl_union_set
*writes
;
2392 space
= isl_map_get_space(group
->access
);
2393 writes
= isl_union_set_empty(space
);
2394 for (i
= 0; i
< group
->n_ref
; ++i
) {
2398 if (!group
->refs
[i
]->write
)
2401 space
= isl_map_get_space(group
->refs
[i
]->tagged_access
);
2402 space
= isl_space_domain(space
);
2403 writes_i
= isl_set_universe(space
);
2404 writes
= isl_union_set_add_set(writes
, writes_i
);
2410 /* Is there any write access in "group" that requires synchronization
2411 * on a write to global memory?
2412 * We currently take into account all writes that would require
2413 * synchronization at the thread level depth, but if the copying
2414 * for this group is performed at an outer level, then we do not
2415 * actually need to take into account dependences at intermediate levels.
2417 static int any_sync_writes_in_group(struct ppcg_kernel
*kernel
,
2418 struct gpu_array_ref_group
*group
)
2420 isl_union_set
*writes
;
2421 int empty
, disjoint
;
2423 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2429 writes
= group_tagged_writes(group
);
2430 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2431 isl_union_set_free(writes
);
2433 return disjoint
< 0 ? -1 : !disjoint
;
2436 /* Collect the references to all writes in "kernel" that write directly
2437 * to global or shared memory, i.e., that are not mapped to private memory.
2438 * Each reference is represented by a universe set in a space
2442 * with S[i,j] the statement instance space and R[] the array reference.
2444 static __isl_give isl_union_set
*collect_non_private_tagged_writes(
2445 struct ppcg_kernel
*kernel
)
2447 isl_union_set
*writes
;
2450 writes
= isl_union_set_empty(isl_union_set_get_space(kernel
->arrays
));
2452 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2453 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2455 for (j
= 0; j
< array
->n_group
; ++j
) {
2456 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2457 isl_union_set
*writes_ij
;
2461 if (group
->private_tile
)
2463 writes_ij
= group_tagged_writes(group
);
2464 writes
= isl_union_set_union(writes
, writes_ij
);
2471 /* Are there any direct writes to global memory that require
2474 static int any_global_or_shared_sync_writes(struct ppcg_kernel
*kernel
)
2476 isl_union_set
*writes
;
2477 int empty
, disjoint
;
2479 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2485 writes
= collect_non_private_tagged_writes(kernel
);
2486 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2487 isl_union_set_free(writes
);
2489 return disjoint
< 0 ? -1 : !disjoint
;
2492 /* Construct an isl_multi_val for use as tile sizes for tiling "node"
2493 * from the elements in "tile_size".
2495 static __isl_give isl_multi_val
*construct_band_tiles_sizes(
2496 __isl_keep isl_schedule_node
*node
, int *tile_size
)
2506 ctx
= isl_schedule_node_get_ctx(node
);
2507 space
= isl_schedule_node_band_get_space(node
);
2508 n
= isl_schedule_node_band_n_member(node
);
2509 mv
= isl_multi_val_zero(space
);
2510 for (i
= 0; i
< n
; ++i
) {
2513 v
= isl_val_int_from_si(ctx
, tile_size
[i
]);
2514 mv
= isl_multi_val_set_val(mv
, i
, v
);
2520 /* Replace the partial schedule S of the band node "node" by
2528 * if scale_tile_loops is set, with f the integers in "factor".
2529 * The list that "factor" points to is assumed to contain at least
2530 * as many elements as the number of members in the band.
2532 static __isl_give isl_schedule_node
*snap_band_to_sizes(
2533 __isl_take isl_schedule_node
*node
, int *factor
,
2534 struct ppcg_options
*options
)
2538 mv
= construct_band_tiles_sizes(node
, factor
);
2539 node
= isl_schedule_node_band_scale_down(node
, isl_multi_val_copy(mv
));
2540 if (options
->scale_tile_loops
)
2541 node
= isl_schedule_node_band_scale(node
,
2542 isl_multi_val_copy(mv
));
2543 isl_multi_val_free(mv
);
2548 /* Tile "band" with tile size specified by "sizes".
2550 * Since the tile loops will be mapped to block ids, we forcibly
2551 * turn off tile loop scaling. We may want to enable tile loop scaling
2552 * at some later point, but then we would have to support the detection
2553 * of strides during the mapping to block ids.
2554 * Similarly, since the point loops will be mapped to thread ids,
2555 * we forcibly shift the point loops so that they start at zero.
2557 static __isl_give isl_schedule_node
*tile_band(
2558 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2560 isl_ctx
*ctx
= isl_schedule_node_get_ctx(node
);
2564 scale_tile
= isl_options_get_tile_scale_tile_loops(ctx
);
2565 isl_options_set_tile_scale_tile_loops(ctx
, 0);
2566 shift_point
= isl_options_get_tile_shift_point_loops(ctx
);
2567 isl_options_set_tile_shift_point_loops(ctx
, 1);
2569 node
= isl_schedule_node_band_tile(node
, sizes
);
2571 isl_options_set_tile_scale_tile_loops(ctx
, scale_tile
);
2572 isl_options_set_tile_shift_point_loops(ctx
, shift_point
);
2577 /* Extract the set of parameter values and outer schedule dimensions
2578 * for which any statement instance
2579 * in the kernel inserted at "node" needs to be executed.
2580 * Intersect the set of parameter values derived from the host schedule
2581 * relation with the context of "prog".
2583 static __isl_give isl_set
*extract_context(__isl_keep isl_schedule_node
*node
,
2584 struct gpu_prog
*prog
)
2586 isl_union_map
*schedule
;
2587 isl_union_set
*schedule_domain
;
2591 schedule
= isl_schedule_node_get_prefix_schedule_relation(node
);
2592 schedule_domain
= isl_union_map_range(schedule
);
2593 empty
= isl_union_set_is_empty(schedule_domain
);
2595 isl_union_set_free(schedule_domain
);
2602 space
= isl_union_set_get_space(schedule_domain
);
2603 isl_union_set_free(schedule_domain
);
2604 space
= isl_space_set_from_params(space
);
2605 depth
= isl_schedule_node_get_schedule_depth(node
);
2606 space
= isl_space_add_dims(space
, isl_dim_set
, depth
);
2607 context
= isl_set_empty(space
);
2609 context
= isl_set_from_union_set(schedule_domain
);
2611 context
= isl_set_intersect_params(context
,
2612 isl_set_copy(prog
->context
));
2617 /* Return the set of outer array elements accessed by
2618 * by the statement instance in "domain" in "prog".
2620 static __isl_give isl_union_set
*accessed_by_domain(
2621 __isl_take isl_union_set
*domain
, struct gpu_prog
*prog
)
2623 isl_union_map
*access
;
2624 isl_union_set
*arrays
;
2626 access
= isl_union_map_union(isl_union_map_copy(prog
->read
),
2627 isl_union_map_copy(prog
->may_write
));
2628 access
= isl_union_map_intersect_domain(access
, domain
);
2629 arrays
= isl_union_map_range(access
);
2630 arrays
= isl_union_set_apply(arrays
,
2631 isl_union_map_copy(prog
->to_outer
));
2636 /* Return the number of outer band members of the band node "node"
2637 * that are marked coincident.
2639 static int n_outer_coincidence(__isl_keep isl_schedule_node
*node
)
2643 n
= isl_schedule_node_band_n_member(node
);
2645 for (i
= 0; i
< n
; ++i
)
2646 if (!isl_schedule_node_band_member_get_coincident(node
, i
))
2652 /* If the band node "node" has more than "n" members, then split off
2653 * the first "n" of them.
2655 static __isl_give isl_schedule_node
*split_band(
2656 __isl_take isl_schedule_node
*node
, int n
)
2660 dim
= isl_schedule_node_band_n_member(node
);
2662 node
= isl_schedule_node_band_split(node
, n
);
2667 /* Scale a band node that may have been split by split_band.
2668 * "sizes" are the scaling factors for the original node.
2669 * "node" either points to the original band node, or the outer
2670 * of the two pieces after splitting.
2672 * If the number of elements in "node" is smaller than the number of
2673 * elements in "sizes", then some splitting has occurred and we split
2674 * "sizes" in the same way.
2676 static __isl_give isl_schedule_node
*scale_band(
2677 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2681 n
= isl_multi_val_dim(sizes
, isl_dim_set
);
2682 dim
= isl_schedule_node_band_n_member(node
);
2684 isl_multi_val
*sizes2
;
2686 sizes2
= isl_multi_val_copy(sizes
);
2687 sizes
= isl_multi_val_drop_dims(sizes
,
2688 isl_dim_set
, dim
, n
- dim
);
2689 sizes2
= isl_multi_val_drop_dims(sizes2
, isl_dim_set
, 0, dim
);
2690 node
= isl_schedule_node_child(node
, 0);
2691 node
= isl_schedule_node_band_scale(node
, sizes2
);
2692 node
= isl_schedule_node_parent(node
);
2695 return isl_schedule_node_band_scale(node
, sizes
);
2698 /* Return an isl_multi_aff, with as elements the parameters in "space"
2699 * that have the names specified by the elements in "names".
2700 * If (some of) these parameters do not already appear in "space",
2701 * then they are added first.
2703 static __isl_give isl_multi_aff
*parameter_vector(__isl_take isl_space
*space
,
2704 __isl_keep isl_id_list
*names
)
2707 isl_local_space
*ls
;
2711 space
= isl_space_free(space
);
2713 n
= isl_id_list_n_id(names
);
2714 for (i
= 0; i
< n
; ++i
) {
2718 id
= isl_id_list_get_id(names
, i
);
2719 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2724 pos
= isl_space_dim(space
, isl_dim_param
);
2725 space
= isl_space_add_dims(space
, isl_dim_param
, 1);
2726 space
= isl_space_set_dim_id(space
, isl_dim_param
, pos
, id
);
2728 ma
= isl_multi_aff_zero(isl_space_copy(space
));
2729 ls
= isl_local_space_from_space(isl_space_domain(space
));
2730 for (i
= 0; i
< n
; ++i
) {
2735 id
= isl_id_list_get_id(names
, i
);
2736 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2738 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
2739 isl_dim_param
, pos
);
2740 ma
= isl_multi_aff_set_aff(ma
, i
, aff
);
2742 isl_local_space_free(ls
);
2747 /* Return constraints on the domain elements that equate a sequence of
2748 * parameters called "names", to the partial schedule
2749 * of "node" modulo the integers in "size".
2750 * The number of elements in the array "size" should be equal
2751 * to the number of elements in "names".
2752 * The number of members of the band node "node" should be smaller
2753 * than or equal to this number. If it is smaller, then the first
2754 * elements of "names" are equated to zero.
2756 static __isl_give isl_union_set
*set_schedule_modulo(
2757 __isl_keep isl_schedule_node
*node
, __isl_keep isl_id_list
*names
,
2763 isl_multi_union_pw_aff
*mupa
, *mupa2
;
2765 isl_union_set
*domain
;
2769 n
= isl_id_list_n_id(names
);
2771 return isl_schedule_node_get_universe_domain(node
);
2772 n_zero
= n
- isl_schedule_node_band_n_member(node
);
2774 mupa
= isl_schedule_node_band_get_partial_schedule(node
);
2775 mv
= construct_band_tiles_sizes(node
, size
+ n_zero
);
2776 mupa
= isl_multi_union_pw_aff_mod_multi_val(mupa
, mv
);
2778 space
= isl_multi_union_pw_aff_get_space(mupa
);
2779 space
= isl_space_params(space
);
2780 space
= isl_space_set_from_params(space
);
2781 space
= isl_space_add_dims(space
, isl_dim_set
, n_zero
);
2782 ma
= isl_multi_aff_zero(space
);
2784 domain
= isl_schedule_node_get_universe_domain(node
);
2785 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(
2786 isl_union_set_copy(domain
), ma
);
2787 mupa
= isl_multi_union_pw_aff_range_product(mupa2
, mupa
);
2789 space
= isl_multi_union_pw_aff_get_space(mupa
);
2790 ma
= parameter_vector(space
, names
);
2792 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(domain
, ma
);
2793 mupa
= isl_multi_union_pw_aff_sub(mupa
, mupa2
);
2795 return isl_multi_union_pw_aff_zero_union_set(mupa
);
2798 /* Insert a context node at "node" introducing the block and thread
2799 * identifiers along with their bounds, which are stored in kernel->grid_size
2800 * and kernel->block_dim.
2801 * Note that the bounds on the block identifiers may implicitly impose
2802 * constraints on the parameters. A guard needs to be inserted
2803 * in the schedule tree to ensure that those bounds hold at "node".
2804 * This guard is inserted in insert_guard.
2806 static __isl_give isl_schedule_node
*insert_context(struct ppcg_kernel
*kernel
,
2807 __isl_take isl_schedule_node
*node
)
2811 context
= isl_set_universe(isl_set_get_space(kernel
->context
));
2813 context
= add_bounded_parameters_dynamic(context
,
2814 kernel
->grid_size
, kernel
->block_ids
);
2815 context
= add_bounded_parameters(context
,
2816 kernel
->block_dim
, kernel
->thread_ids
);
2818 node
= isl_schedule_node_insert_context(node
, context
);
2823 /* Insert a guard that eliminates kernel launches where the kernel
2824 * obviously does not have any work to do.
2826 * In particular, eliminate kernel launches where there are obviously
2828 * Use the same block size constraints that are used to create the context
2829 * to ensure that all constraints implicit in the constructed context
2830 * are imposed by the guard.
2832 * Additionally, add other constraints that are valid
2833 * for each executed instance ("context"), as long as this does not result
2836 static __isl_give isl_schedule_node
*insert_guard(
2837 __isl_take isl_schedule_node
*node
, __isl_keep isl_set
*context
,
2838 __isl_keep isl_multi_pw_aff
*size
, struct ppcg_scop
*scop
)
2844 guard
= isl_set_copy(context
);
2845 guard
= isl_set_compute_divs(guard
);
2846 guard
= isl_set_from_basic_set(isl_set_simple_hull(guard
));
2848 nparam
= isl_set_dim(guard
, isl_dim_param
);
2849 n
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
2850 ids
= ppcg_scop_generate_names(scop
, n
, "__ppcg_tmp");
2851 guard
= add_bounded_parameters_dynamic(guard
, size
, ids
);
2852 isl_id_list_free(ids
);
2853 guard
= isl_set_project_out(guard
, isl_dim_param
, nparam
, n
);
2855 node
= isl_schedule_node_insert_guard(node
, guard
);
2860 /* Does any array reference group mapping require the band that is mapped
2861 * to threads to be unrolled?
2863 static int kernel_requires_unroll(struct ppcg_kernel
*kernel
)
2867 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2868 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2870 for (j
= 0; j
< array
->n_group
; ++j
) {
2871 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2872 if (gpu_array_ref_group_requires_unroll(group
))
2880 /* Mark the given band node "node" for unrolling by the AST generator and
2881 * then sink it to the leaves of the schedule tree.
2882 * All dimensions of "node" are assumed to be coincident, such that this
2883 * sinking is a valid operation.
2885 static __isl_give isl_schedule_node
*unroll(__isl_take isl_schedule_node
*node
)
2889 n
= isl_schedule_node_band_n_member(node
);
2890 for (i
= 0; i
< n
; ++i
)
2891 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
2892 isl_ast_loop_unroll
);
2894 node
= isl_schedule_node_band_sink(node
);
2899 /* Insert a synchronization node in the schedule tree of "node"
2900 * after the core computation of "kernel" at the level of the band
2901 * that is mapped to threads, except if that level is equal to
2902 * that of the band that is mapped to blocks or if there are no writes
2903 * to global or shared memory in the core computation that require
2905 * If there are any writes to shared memory and the shared memory
2906 * copying is performed at the same level, then synchronization
2907 * is needed between the core and the copying anyway, so we might
2908 * as well add it here. If the copying is performed at a higher
2909 * level, then different iterations of intermediate schedule dimensions
2910 * may have a different mapping from between shared memory elements and
2911 * threads, such that synchronization is required after the core.
2912 * "node" is assumed to point to the kernel node.
2914 static __isl_give isl_schedule_node
*add_sync(struct ppcg_kernel
*kernel
,
2915 __isl_take isl_schedule_node
*node
)
2920 need_sync
= any_global_or_shared_sync_writes(kernel
);
2922 return isl_schedule_node_free(node
);
2926 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
2928 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
2929 if (kernel_depth
== isl_schedule_node_get_schedule_depth(node
))
2930 return gpu_tree_move_up_to_kernel(node
);
2932 node
= gpu_tree_ensure_following_sync(node
, kernel
);
2934 node
= gpu_tree_move_up_to_kernel(node
);
2939 /* Return a read ("read" is 1) or write access relation for "group"
2940 * with those accesses removed that are only needed to communicate data
2941 * within the subtree of the schedule rooted at "node".
2942 * Furthermore, include the prefix schedule at "node".
2943 * That is, return a relation of the form
2947 * with D the outer schedule dimensions at "node".
2949 static __isl_give isl_union_map
*anchored_non_local_accesses(
2950 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2951 __isl_take isl_schedule_node
*node
, int read
)
2953 isl_union_map
*access
;
2954 isl_union_map
*prefix
;
2956 access
= gpu_array_ref_group_access_relation(group
, read
, !read
);
2957 access
= remove_local_accesses_group(kernel
, group
, access
, node
, read
);
2958 prefix
= isl_schedule_node_get_prefix_schedule_relation(node
);
2959 access
= isl_union_map_range_product(prefix
, access
);
2964 /* Given an array reference group "group", create a mapping
2966 * read[D -> A] -> [D -> A]
2968 * if "read" is set or
2970 * write[D -> A] -> [D -> A]
2972 * if "read" is not set.
2973 * D corresponds to the outer group->depth dimensions of
2974 * the kernel schedule.
2976 static __isl_give isl_multi_aff
*create_from_access(isl_ctx
*ctx
,
2977 struct gpu_array_ref_group
*group
, int read
)
2982 space
= isl_space_copy(group
->array
->space
);
2983 space
= isl_space_from_range(space
);
2984 space
= isl_space_add_dims(space
, isl_dim_in
, group
->depth
);
2985 space
= isl_space_wrap(space
);
2986 space
= isl_space_map_from_set(space
);
2988 id
= isl_id_alloc(ctx
, read
? "read" : "write", group
);
2989 space
= isl_space_set_tuple_id(space
, isl_dim_in
, id
);
2991 return isl_multi_aff_identity(space
);
2994 /* If any writes in "group" require synchronization, then make sure
2995 * that there is a synchronization node for "kernel" after the node
2996 * following "node" in a sequence.
2998 * If "shared" is set and no synchronization is needed for
2999 * the writes to global memory, then add synchronization before
3000 * the kernel to protect shared memory from being overwritten
3001 * by the next iteration of the core computation.
3002 * No additional synchronization is needed to protect against
3003 * the next copy into shared memory because each element of
3004 * the shared memory tile is always copied by the same thread.
3006 static __isl_give isl_schedule_node
*add_group_write_sync(
3007 __isl_take isl_schedule_node
*node
, struct ppcg_kernel
*kernel
,
3008 struct gpu_array_ref_group
*group
, int shared
)
3012 need_sync
= any_sync_writes_in_group(kernel
, group
);
3014 return isl_schedule_node_free(node
);
3016 node
= isl_schedule_node_parent(node
);
3017 node
= isl_schedule_node_next_sibling(node
);
3018 node
= isl_schedule_node_child(node
, 0);
3019 node
= gpu_tree_ensure_following_sync(node
, kernel
);
3020 } else if (shared
) {
3021 node
= isl_schedule_node_parent(node
);
3022 node
= isl_schedule_node_parent(node
);
3023 node
= gpu_tree_move_down_to_depth(node
, group
->depth
,
3025 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3031 /* Add copy statements to the schedule tree of "node"
3032 * for reading from global memory to private memory (if "read" is set) or
3033 * for writing back from private memory to global memory
3034 * (if "read" is not set) for the array reference group "group" that
3035 * is mapped to private memory.
3036 * On input, "node" points to the kernel node, and it is moved
3037 * back there on output.
3039 * The copies are performed in the order of the array elements.
3040 * The copy statement instances include a reference to the outer
3041 * group->depth dimensions of the kernel schedule for ease of
3042 * combining them with the group tiling.
3044 * That is, the extra schedule is of the form
3048 * where D corresponds to the outer group->depth dimensions of
3049 * the kernel schedule and A to the global array.
3050 * This schedule is unrolled because registers are not addressable.
3052 * The copying is inserted in the schedule tree through an extension
3057 * where the extra domain elements type[D -> A] are those accessed
3059 * A filter is inserted on type[D -> A] to ensure that the element
3060 * is read/written by the same thread that needs the element.
3061 * This filter is obtained by applying
3065 * to the thread filter for the core statements.
3067 * The extension is inserted before the core computation in case of a read
3068 * and after the core computation in case of a write.
3069 * In the latter case, we also make sure that there is a synchronization
3070 * node after the write to global memory, unless this write is performed
3071 * at the outer level of the kernel.
3072 * In principle, this synchronization could be inserted higher
3073 * in the schedule tree depending on where the corresponding reads
3074 * from global memory are performed.
3076 static __isl_give isl_schedule_node
*add_copies_group_private(
3077 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3078 __isl_take isl_schedule_node
*node
, int read
)
3080 isl_union_map
*access
;
3081 isl_union_map
*prefix
;
3082 isl_union_set
*domain
;
3084 isl_multi_aff
*from_access
;
3085 isl_multi_pw_aff
*mpa
;
3086 isl_multi_union_pw_aff
*mupa
;
3087 isl_schedule_node
*graft
;
3088 isl_union_set
*filter
;
3092 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3093 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3095 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3096 empty
= isl_union_map_is_empty(access
);
3097 if (empty
< 0 || empty
) {
3098 isl_union_map_free(access
);
3100 return isl_schedule_node_free(node
);
3101 return gpu_tree_move_up_to_kernel(node
);
3104 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3105 space
= isl_space_domain(isl_multi_aff_get_space(from_access
));
3106 access
= isl_union_map_preimage_range_multi_aff(access
, from_access
);
3108 filter
= isl_union_set_copy(kernel
->thread_filter
);
3109 filter
= isl_union_set_apply(filter
, isl_union_map_copy(access
));
3110 filter
= isl_union_set_detect_equalities(filter
);
3111 filter
= isl_union_set_coalesce(filter
);
3113 domain
= isl_union_map_range(access
);
3114 access
= isl_union_set_wrapped_domain_map(domain
);
3115 access
= isl_union_map_reverse(access
);
3116 access
= isl_union_map_coalesce(access
);
3117 graft
= isl_schedule_node_from_extension(access
);
3119 space
= isl_space_map_from_set(space
);
3120 mpa
= isl_multi_pw_aff_identity(space
);
3121 mpa
= isl_multi_pw_aff_range_factor_range(mpa
);
3122 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3124 graft
= isl_schedule_node_child(graft
, 0);
3125 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3126 graft
= unroll(graft
);
3128 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3130 graft
= isl_schedule_node_parent(graft
);
3133 node
= isl_schedule_node_graft_before(node
, graft
);
3135 node
= isl_schedule_node_graft_after(node
, graft
);
3136 if (kernel_depth
< group
->depth
)
3137 node
= add_group_write_sync(node
, kernel
, group
, 0);
3140 node
= gpu_tree_move_up_to_kernel(node
);
3145 /* Add copy statements to the schedule tree of "node"
3146 * for reading from global memory to shared memory (if "read" is set) or
3147 * for writing back from shared memory to global memory
3148 * (if "read" is not set) for the array reference group "group" that
3149 * is mapped to shared memory.
3150 * On input, "node" points to the kernel node, and it is moved
3151 * back there on output.
3153 * The copies are performed in the order of the corresponding shared
3155 * The copy statement instances include a reference to the outer
3156 * group->depth dimensions of the kernel schedule for ease of
3157 * combining them with the group tiling.
3159 * If we are performing a read from global memory to shared memory and
3160 * if the array involved is not a scalar, then we copy
3161 * the entire tile to shared memory. This may result in some extra
3162 * elements getting copied, but it should lead to simpler code
3163 * (which means that fewer registers may be needed) and less divergence.
3165 * Otherwise, we only copy the elements that will be read or have been written
3168 * That is, the extra schedule is of the form
3172 * where D corresponds to the outer group->depth dimensions of
3173 * the kernel schedule, A to the global array and T is the corresponding
3174 * shared memory tile.
3176 * The copying is inserted in the schedule tree through an extension
3181 * where the extra domain elements type[D -> A] are those accessed
3182 * by the group. In the case of read from a non-scalar, this set
3183 * is replaced by the entire shared memory tile.
3185 * A filter is inserted on type[D -> A] to map the copy instances
3186 * to the threads. In particular, the thread identifiers are
3187 * equated to the position inside the shared memory tile (T)
3188 * modulo the block size.
3189 * We try to align the innermost tile dimension with the innermost
3190 * thread identifier (x) as a heuristic to improve coalescing.
3191 * In particular, if the dimension of the tile is greater than
3192 * the dimension of the block, then the schedule mapping to the tile
3193 * is broken up into two pieces and the filter is applied to the inner part.
3194 * If, on the other hand, the dimension of the tile is smaller than
3195 * the dimension of the block, then the initial thread identifiers
3196 * are equated to zero and the remaining thread identifiers are
3197 * matched to the memory tile.
3199 * The extension is inserted before the core computation in case of a read
3200 * and after the core computation in case of a write.
3201 * In the case of a read, we first need to make sure there is some
3202 * synchronization before the core computation such that we can put the read
3203 * from global memory to shared memory before that synchronization.
3204 * This ensures that all threads have finished copying into shared memory
3205 * before the shared memory is used.
3206 * We also need to make sure that there is a synchronization node after
3207 * the core computation to ensure that the next load into shared memory
3208 * only happens after all data has been used. There is no need for
3209 * this synchronization if we are at the outer level since then there
3210 * won't be a next load.
3211 * In the case of a write, we need to make sure there is some synchronization
3212 * after the core computation such taht we can put the write from shared
3213 * memory to global memory after that synchronization.
3214 * Unless we are at the outer level, we also need a synchronization node
3215 * after the write to ensure the data is saved to global memory
3216 * before the next iteration write to the same shared memory.
3217 * It also makes sure the data has arrived in global memory before
3218 * it is read in a subsequent iteration.
3220 static __isl_give isl_schedule_node
*add_copies_group_shared(
3221 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3222 __isl_take isl_schedule_node
*node
, int read
)
3224 struct gpu_array_tile
*tile
;
3225 isl_union_map
*access
;
3226 isl_union_set
*domain
;
3227 isl_union_set
*sync
;
3229 isl_multi_aff
*from_access
;
3230 isl_multi_pw_aff
*mpa
;
3231 isl_multi_union_pw_aff
*mupa
;
3232 isl_schedule_node
*graft
;
3233 isl_union_set
*filter
;
3238 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3239 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3241 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3242 empty
= isl_union_map_is_empty(access
);
3243 if (empty
< 0 || empty
) {
3244 isl_union_map_free(access
);
3246 return isl_schedule_node_free(node
);
3247 return gpu_tree_move_up_to_kernel(node
);
3250 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3252 tile
= gpu_array_ref_group_tile(group
);
3253 ma
= isl_multi_aff_copy(tile
->tiling
);
3254 ma
= isl_multi_aff_pullback_multi_aff(ma
,
3255 isl_multi_aff_copy(from_access
));
3256 mpa
= isl_multi_pw_aff_from_multi_aff(ma
);
3257 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3259 domain
= isl_union_map_range(access
);
3261 if (read
&& !gpu_array_is_scalar(group
->array
)) {
3263 isl_union_set_free(domain
);
3264 map
= group_tile(group
);
3265 domain
= isl_union_set_from_set(isl_map_wrap(map
));
3268 domain
= isl_union_set_preimage_multi_aff(domain
, from_access
);
3269 access
= isl_union_set_wrapped_domain_map(domain
);
3270 access
= isl_union_map_reverse(access
);
3271 access
= isl_union_map_coalesce(access
);
3272 graft
= isl_schedule_node_from_extension(access
);
3274 graft
= isl_schedule_node_child(graft
, 0);
3276 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3278 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0) {
3279 graft
= isl_schedule_node_band_split(graft
,
3280 tile
->n
- kernel
->n_block
);
3281 graft
= isl_schedule_node_child(graft
, 0);
3283 if (tile
->n
< kernel
->n_block
)
3284 skip
= kernel
->n_block
- tile
->n
;
3287 filter
= set_schedule_modulo(graft
, kernel
->thread_ids
,
3289 if (!kernel
->options
->wrap
)
3290 graft
= snap_band_to_sizes(graft
, kernel
->block_dim
+ skip
,
3292 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0)
3293 graft
= isl_schedule_node_parent(graft
);
3294 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3296 while (graft
&& isl_schedule_node_has_parent(graft
))
3297 graft
= isl_schedule_node_parent(graft
);
3300 if (kernel_depth
< group
->depth
)
3301 node
= gpu_tree_ensure_sync_after_core(node
, kernel
);
3302 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3303 node
= isl_schedule_node_graft_before(node
, graft
);
3305 node
= gpu_tree_move_right_to_sync(node
, kernel
);
3306 node
= isl_schedule_node_graft_after(node
, graft
);
3307 if (kernel_depth
< group
->depth
)
3308 node
= add_group_write_sync(node
, kernel
, group
, 1);
3311 node
= gpu_tree_move_up_to_kernel(node
);
3316 /* Check whether the array reference group "group" is mapped to
3317 * private or shared memory and, if so,
3318 * add copy statements to the schedule tree of "node"
3319 * for reading from global memory to private or shared memory
3320 * (if "read" is set) or for writing back from private or shared memory
3321 * to global memory (if "read" is not set) for this group.
3322 * On input, "node" points to the kernel node, and it is moved
3323 * back there on output.
3325 static __isl_give isl_schedule_node
*add_copies_group(
3326 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3327 __isl_take isl_schedule_node
*node
, int read
)
3329 if (group
->private_tile
)
3330 return add_copies_group_private(kernel
, group
, node
, read
);
3331 if (group
->shared_tile
)
3332 return add_copies_group_shared(kernel
, group
, node
, read
);
3336 /* For each array reference group that is mapped to private or shared memory,
3337 * add copy statements to the schedule tree of "node"
3338 * for reading from global memory to private or shared memory
3339 * and for writing back.
3340 * On input, "node" points to the kernel node, and it is moved
3341 * back there on output.
3343 static __isl_give isl_schedule_node
*add_copies(struct ppcg_kernel
*kernel
,
3344 __isl_take isl_schedule_node
*node
)
3348 for (i
= 0; i
< kernel
->n_array
; ++i
) {
3349 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
3351 for (j
= 0; j
< array
->n_group
; ++j
) {
3352 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3354 node
= add_copies_group(kernel
, group
, node
, 1);
3357 node
= add_copies_group(kernel
, group
, node
, 0);
3366 /* Mark all dimensions in the current band node atomic.
3368 static __isl_give isl_schedule_node
*atomic(__isl_take isl_schedule_node
*node
)
3372 n
= isl_schedule_node_band_n_member(node
);
3373 for (i
= 0; i
< n
; ++i
)
3374 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
3375 isl_ast_loop_atomic
);
3380 /* Mark "node" atomic, if it is a band node.
3381 * Do the same for all ancestors.
3382 * Return a pointer to "node" (in the updated schedule tree).
3384 static __isl_give isl_schedule_node
*atomic_ancestors(
3385 __isl_take isl_schedule_node
*node
)
3391 if (!isl_schedule_node_has_parent(node
))
3394 pos
= isl_schedule_node_get_child_position(node
);
3395 node
= isl_schedule_node_parent(node
);
3396 if (isl_schedule_node_get_type(node
) == isl_schedule_node_band
)
3397 node
= atomic(node
);
3398 node
= atomic_ancestors(node
);
3399 node
= isl_schedule_node_child(node
, pos
);
3404 /* Collect all write references that require synchronization.
3405 * "node" is assumed to point to the kernel node.
3406 * Each reference is represented by a universe set in a space
3410 * with S[i,j] the statement instance space and R[] the array reference.
3412 * This function should be called before block and thread filters are added.
3414 * Synchronization is needed after a write if there is a subsequent read
3415 * within the same block that may not be performed by the same thread.
3416 * There should not be any dependences between different blocks,
3417 * so we start with the flow dependences within the same kernel invocation
3418 * and we subtract from these those dependences that are mapped
3419 * to the same iteration of the bands where synchronization is inserted.
3420 * We do not remove pairs of instances that are known to map to
3421 * the same thread across different iterations of the intermediate
3422 * bands because the read may be performed by a different thread
3423 * than the one that needs the value if shared memory is involved.
3425 * We also consider all pairs of possible writes that access the same
3426 * memory location and that may be mapped to the same block but not
3427 * to the same iteration of the intermediate bands.
3428 * In theory, it would be possible for one thread to still be in
3429 * a previous iteration of a loop in these bands.
3430 * A write to global memory in this delayed thread could then overwrite
3431 * a write from another thread that has already moved on to
3432 * the next iteration.
3434 * After computing the above writes paired off with reads or writes
3435 * that depend on them, we project onto the domain writes.
3436 * Sychronization is needed after writes to global memory
3437 * through these references.
3439 static __isl_give isl_union_set
*compute_sync_writes(
3440 struct ppcg_kernel
*kernel
, __isl_keep isl_schedule_node
*node
)
3442 isl_union_map
*local
;
3443 isl_union_map
*may_writes
, *shared_access
;
3444 isl_union_map
*kernel_prefix
, *thread_prefix
;
3445 isl_union_map
*equal
;
3446 isl_union_set
*wrap
;
3447 isl_union_set
*domain
;
3449 domain
= isl_schedule_node_get_universe_domain(node
);
3450 kernel_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3451 node
= isl_schedule_node_copy(node
);
3452 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3453 thread_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3454 isl_schedule_node_free(node
);
3456 may_writes
= isl_union_map_copy(kernel
->prog
->scop
->tagged_may_writes
);
3457 may_writes
= isl_union_map_curry(may_writes
);
3458 may_writes
= isl_union_map_intersect_domain(may_writes
, domain
);
3459 may_writes
= isl_union_map_uncurry(may_writes
);
3460 shared_access
= isl_union_map_copy(may_writes
);
3461 shared_access
= isl_union_map_apply_range(shared_access
,
3462 isl_union_map_reverse(may_writes
));
3464 local
= isl_union_map_copy(kernel
->prog
->scop
->tagged_dep_flow
);
3465 local
= isl_union_map_union(local
, shared_access
);
3466 local
= isl_union_map_zip(local
);
3468 equal
= isl_union_map_apply_range(kernel_prefix
,
3469 isl_union_map_reverse(isl_union_map_copy(kernel_prefix
)));
3470 wrap
= isl_union_map_wrap(equal
);
3471 local
= isl_union_map_intersect_domain(local
, wrap
);
3472 equal
= isl_union_map_apply_range(thread_prefix
,
3473 isl_union_map_reverse(isl_union_map_copy(thread_prefix
)));
3474 wrap
= isl_union_map_wrap(equal
);
3475 local
= isl_union_map_subtract_domain(local
, wrap
);
3477 local
= isl_union_map_zip(local
);
3478 local
= isl_union_map_universe(local
);
3480 return isl_union_map_domain(local
);
3483 /* Group the domain elements into a single space, named kernelX,
3484 * with X the kernel sequence number "kernel_id".
3486 static __isl_give isl_schedule_node
*group_statements(
3487 __isl_take isl_schedule_node
*node
, int kernel_id
)
3495 snprintf(buffer
, sizeof(buffer
), "kernel%d", kernel_id
);
3496 id
= isl_id_alloc(isl_schedule_node_get_ctx(node
), buffer
, NULL
);
3497 return isl_schedule_node_group(node
, id
);
3500 /* Create a ppcg_kernel representing the domain instances that reach "node"
3501 * and insert a mark node pointing to the ppcg_kernel before "node".
3502 * The band that "node" points to is the band that needs to be mapped
3503 * to block identifiers. The band that needs to be mapped to thread
3504 * identifiers should be marked by a "thread" mark by the caller.
3505 * This mark is removed by this function.
3506 * If "scale" is set, then the band that "node" points to is scaled
3509 * Mark all outer band nodes as atomic to ensure each kernel is only
3511 * If the domain elements that reach "node" live in more than one space,
3512 * then group the domain elements into a single space, named kernelX,
3513 * with X the kernel sequence number.
3515 * Insert a guard node governing the kernel node to ensure that
3516 * no kernels with zero blocks are launched.
3518 * Insert a context node describing the block and thread
3519 * identifiers inside the kernel mark.
3520 * The context node needs to be inserted after the effective block size
3521 * has been determined such that the bounds on the thread identifiers
3522 * would reflect the effective block size.
3523 * Insert a filter node inside the context node mapping the statement
3524 * instances to block identifiers. In particular, the block identifiers
3525 * are equated to the partial schedule of band that was marked for mapping
3526 * to blocks modulo the grid size.
3527 * Insert a filter node inside the "thread" mark mapping the statement
3528 * instances to thread identifiers. In particular, the thread identifiers
3529 * are equated to the partial schedule of band that was marked for mapping
3530 * to threads modulo the block size.
3532 * Compute array reference groups for all arrays, set the local
3533 * array bounds based on the set of domain instances that reach
3534 * the kernel node, check the total amount of shared memory used
3535 * and compute all group tilings.
3536 * The array reference groups are computed after the block filter
3537 * has been inserted because it affects the mapping to shared or
3538 * private memory. This computation also requires the thread filter
3539 * (in the ppcg_kernel object), but this thread filter should not
3540 * have been added to the schedule tree yet since the computation
3541 * requires the schedule of the band that needs to be mapped to
3542 * threads before the privatization is applied.
3544 * If any array reference group requires the band mapped to threads
3545 * to be unrolled, then we perform the required unrolling.
3547 * We save a copy of the schedule that may influence the mappings
3548 * to shared or private memory in kernel->shared_schedule.
3550 * Finally, we add synchronization and copy statements to the schedule tree,
3551 * remove the "thread" mark and create representations for the local
3552 * variables in the kernel.
3554 * We keep a copy of the isl_id that points to the kernel to ensure
3555 * that the kernel does not get destroyed if the schedule node
3556 * is freed due to some error condition.
3558 static __isl_give isl_schedule_node
*create_kernel(struct gpu_gen
*gen
,
3559 __isl_take isl_schedule_node
*node
, int scale
,
3560 __isl_keep isl_multi_val
*sizes
)
3562 struct ppcg_kernel
*kernel
;
3564 isl_schedule_node
*node_thread
;
3565 isl_union_map
*host_schedule
;
3566 isl_set
*host_domain
;
3567 isl_union_set
*domain
;
3568 int single_statement
;
3570 kernel
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel
);
3571 kernel
= ppcg_kernel_create_local_arrays(kernel
, gen
->prog
);
3573 return isl_schedule_node_free(node
);
3575 domain
= isl_schedule_node_get_domain(node
);
3576 single_statement
= isl_union_set_n_set(domain
) == 1;
3578 kernel
->ctx
= gen
->ctx
;
3579 kernel
->prog
= gen
->prog
;
3580 kernel
->options
= gen
->options
;
3581 kernel
->context
= extract_context(node
, gen
->prog
);
3582 kernel
->core
= isl_union_set_universe(isl_union_set_copy(domain
));
3583 kernel
->arrays
= accessed_by_domain(isl_union_set_copy(domain
),
3585 kernel
->n_grid
= n_outer_coincidence(node
);
3586 node_thread
= isl_schedule_node_copy(node
);
3587 node_thread
= gpu_tree_move_down_to_thread(node_thread
, kernel
->core
);
3588 node_thread
= isl_schedule_node_child(node_thread
, 0);
3589 kernel
->n_block
= n_outer_coincidence(node_thread
);
3590 isl_schedule_node_free(node_thread
);
3591 kernel
->id
= gen
->kernel_id
++;
3592 read_grid_and_block_sizes(kernel
, gen
);
3594 kernel
->sync_writes
= compute_sync_writes(kernel
, node
);
3596 host_schedule
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3597 host_domain
= isl_set_from_union_set(isl_union_map_range(
3600 node
= atomic_ancestors(node
);
3602 id
= isl_id_alloc(gen
->ctx
, "kernel", kernel
);
3603 id
= isl_id_set_free_user(id
, &ppcg_kernel_free_wrap
);
3604 node
= isl_schedule_node_insert_mark(node
, isl_id_copy(id
));
3606 if (!single_statement
)
3607 node
= group_statements(node
, kernel
->id
);
3609 node
= isl_schedule_node_child(node
, 0);
3610 node
= split_band(node
, kernel
->n_grid
);
3611 kernel
->block_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3612 kernel
->n_grid
, "b");
3613 kernel
->block_filter
= set_schedule_modulo(node
, kernel
->block_ids
,
3615 kernel
->grid_size
= extract_grid_size(kernel
,
3616 isl_union_set_copy(domain
));
3617 if (!kernel
->options
->wrap
)
3618 node
= snap_band_to_sizes(node
, kernel
->grid_dim
,
3621 node
= scale_band(node
, isl_multi_val_copy(sizes
));
3622 node
= isl_schedule_node_parent(node
);
3623 if (!single_statement
)
3624 node
= isl_schedule_node_parent(node
);
3625 node
= insert_guard(node
, kernel
->context
, kernel
->grid_size
,
3627 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3628 node
= isl_schedule_node_child(node
, 0);
3629 node
= split_band(node
, kernel
->n_block
);
3630 kernel
->thread_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3631 kernel
->n_block
, "t");
3632 kernel
->thread_filter
= set_schedule_modulo(node
, kernel
->thread_ids
,
3634 extract_block_size(kernel
, domain
);
3636 node
= gpu_tree_move_up_to_kernel(node
);
3637 node
= isl_schedule_node_child(node
, 0);
3638 node
= insert_context(kernel
, node
);
3639 node
= isl_schedule_node_child(node
, 0);
3640 node
= isl_schedule_node_insert_filter(node
,
3641 isl_union_set_copy(kernel
->block_filter
));
3643 node
= gpu_tree_move_up_to_kernel(node
);
3645 if (gpu_group_references(kernel
, node
) < 0)
3646 node
= isl_schedule_node_free(node
);
3647 localize_bounds(kernel
, host_domain
);
3648 isl_set_free(host_domain
);
3650 check_shared_memory_bound(kernel
);
3651 compute_group_tilings(kernel
);
3653 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3654 node
= isl_schedule_node_child(node
, 0);
3655 if (!kernel
->options
->wrap
)
3656 node
= snap_band_to_sizes(node
, kernel
->block_dim
,
3658 node
= isl_schedule_node_insert_filter(node
,
3659 isl_union_set_copy(kernel
->thread_filter
));
3660 if (kernel_requires_unroll(kernel
)) {
3661 node
= isl_schedule_node_child(node
, 0);
3662 node
= unroll(node
);
3665 node
= gpu_tree_move_up_to_thread(node
);
3666 kernel
->shared_schedule_dim
=
3667 isl_schedule_node_get_schedule_depth(node
);
3668 kernel
->shared_schedule
=
3669 isl_schedule_node_get_prefix_schedule_union_pw_multi_aff(node
);
3671 node
= gpu_tree_move_up_to_kernel(node
);
3673 node
= add_sync(kernel
, node
);
3674 node
= add_copies(kernel
, node
);
3676 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3677 node
= isl_schedule_node_delete(node
);
3679 node
= gpu_tree_move_up_to_kernel(node
);
3681 if (create_kernel_vars(kernel
) < 0)
3682 node
= isl_schedule_node_free(node
);
3684 if (!single_statement
)
3685 node
= isl_schedule_node_parent(node
);
3686 node
= isl_schedule_node_parent(node
);
3692 /* Insert a zero-dimensional permutable band at "node".
3694 static __isl_give isl_schedule_node
*insert_empty_permutable_band(
3695 __isl_take isl_schedule_node
*node
)
3698 isl_schedule
*schedule
;
3699 isl_union_set
*domain
;
3700 isl_multi_union_pw_aff
*mupa
;
3702 schedule
= isl_schedule_node_get_schedule(node
);
3703 domain
= isl_schedule_get_domain(schedule
);
3704 space
= isl_union_set_get_space(domain
);
3705 isl_union_set_free(domain
);
3706 isl_schedule_free(schedule
);
3708 space
= isl_space_set_from_params(space
);
3709 mupa
= isl_multi_union_pw_aff_zero(space
);
3710 node
= isl_schedule_node_insert_partial_schedule(node
, mupa
);
3711 node
= isl_schedule_node_band_set_permutable(node
, 1);
3716 /* If "node" is the outermost permutable band that can be mapped to block and
3717 * thread identifiers in its branch (or a leaf with no such outer bands),
3718 * then mark the band as such, attaching a ppcg_kernel to the mark.
3720 * If "node" originally points to a leaf, then insert a zero-dimensional
3721 * permutable band such that we can assume that "node" always
3722 * points to a band node.
3724 * Tile "node" using user specified tile sizes, after splitting the band
3725 * if the number of specified tile sizes is smaller than the dimension
3726 * of the band. Mark the point band of this tiling as the band that
3727 * needs to be mapped to threads.
3728 * Create a kernel representing the domain instances that reach "node" and
3729 * insert a mark node pointing to the ppcg_kernel before the band node.
3731 static __isl_give isl_schedule_node
*mark_outer_permutable(
3732 __isl_take isl_schedule_node
*node
, void *user
)
3734 struct gpu_gen
*gen
= user
;
3740 isl_multi_val
*sizes
;
3742 outer
= is_outer_tilable(node
);
3744 return isl_schedule_node_free(node
);
3748 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
3749 node
= insert_empty_permutable_band(node
);
3751 tile_len
= isl_schedule_node_band_n_member(node
);
3752 tile_size
= read_tile_sizes(gen
, &tile_len
);
3754 return isl_schedule_node_free(node
);
3755 if (tile_len
< isl_schedule_node_band_n_member(node
))
3756 node
= isl_schedule_node_band_split(node
, tile_len
);
3757 sizes
= construct_band_tiles_sizes(node
, tile_size
);
3758 node
= tile_band(node
, isl_multi_val_copy(sizes
));
3759 node
= isl_schedule_node_child(node
, 0);
3760 id
= isl_id_alloc(gen
->ctx
, "thread", NULL
);
3761 node
= isl_schedule_node_insert_mark(node
, id
);
3762 node
= isl_schedule_node_parent(node
);
3764 scale
= gen
->options
->scale_tile_loops
;
3765 node
= create_kernel(gen
, node
, scale
, sizes
);
3766 isl_multi_val_free(sizes
);
3772 /* Does the subtree rooted at "node" have any suitably permutable band nodes?
3773 * That is, does it have any nodes that are permutable and that
3774 * have a least one coincident dimension?
3776 static int subtree_has_permutable_bands(__isl_keep isl_schedule_node
*node
)
3778 int any_parallelism
= 0;
3780 if (isl_schedule_node_foreach_descendant(node
, &set_permutable
,
3781 &any_parallelism
) < 0 &&
3785 return any_parallelism
;
3788 /* Mark all variables that are accessed by the statement instances in "domain"
3789 * and that are local to "prog" as requiring a declaration in the host code.
3791 static int declare_accessed_local_variables(struct gpu_prog
*prog
,
3792 __isl_keep isl_union_set
*domain
)
3794 isl_union_set
*arrays
;
3797 if (!ppcg_scop_any_hidden_declarations(prog
->scop
))
3799 arrays
= accessed_by_domain(isl_union_set_copy(domain
), prog
);
3801 for (i
= 0; i
< prog
->n_array
; ++i
) {
3806 if (!prog
->array
[i
].local
)
3808 space
= isl_set_get_space(prog
->array
[i
].extent
);
3809 set
= isl_union_set_extract_set(arrays
, space
);
3810 empty
= isl_set_plain_is_empty(set
);
3815 prog
->array
[i
].declare_local
= 1;
3818 isl_union_set_free(arrays
);
3821 isl_union_set_free(arrays
);
3825 /* If "node" points to a set node, then separate its children
3826 * into subtrees that have suitably permutable bands and
3827 * those that do not.
3828 * Adjust the schedule tree in order to execute the second group
3829 * after the first group and return a pointer to the first group,
3830 * assuming there are any such subtrees.
3831 * Mark all local variables in "prog" that are accessed by
3832 * the second group as requiring a declaration on the host.
3834 static __isl_give isl_schedule_node
*isolate_permutable_subtrees(
3835 __isl_take isl_schedule_node
*node
, struct gpu_prog
*prog
)
3838 isl_union_set
*filter
;
3843 if (isl_schedule_node_get_type(node
) != isl_schedule_node_set
)
3846 n
= isl_schedule_node_n_children(node
);
3848 return isl_schedule_node_free(node
);
3850 node
= isl_schedule_node_child(node
, 0);
3851 filter
= isl_schedule_node_filter_get_filter(node
);
3852 node
= isl_schedule_node_parent(node
);
3853 space
= isl_union_set_get_space(filter
);
3854 isl_union_set_free(filter
);
3855 filter
= isl_union_set_empty(space
);
3857 for (i
= 0; i
< n
; ++i
) {
3860 node
= isl_schedule_node_child(node
, i
);
3861 parallelism
= subtree_has_permutable_bands(node
);
3862 if (parallelism
< 0) {
3863 node
= isl_schedule_node_free(node
);
3864 } else if (!parallelism
) {
3865 isl_union_set
*filter_i
;
3866 filter_i
= isl_schedule_node_filter_get_filter(node
);
3867 filter
= isl_union_set_union(filter
, filter_i
);
3869 node
= isl_schedule_node_parent(node
);
3872 if (declare_accessed_local_variables(prog
, filter
) < 0)
3873 node
= isl_schedule_node_free(node
);
3874 node
= isl_schedule_node_order_after(node
, filter
);
3879 /* Replace any reference to an array element in the range of "copy"
3880 * by a reference to all array elements (defined by the extent of the array).
3882 static __isl_give isl_union_map
*approximate_copy_out(
3883 __isl_take isl_union_map
*copy
, struct gpu_prog
*prog
)
3888 res
= isl_union_map_empty(isl_union_map_get_space(copy
));
3890 for (i
= 0; i
< prog
->n_array
; ++i
) {
3893 isl_union_map
*copy_i
;
3894 isl_union_set
*extent
, *domain
;
3896 space
= isl_space_copy(prog
->array
[i
].space
);
3897 extent
= isl_union_set_from_set(isl_set_universe(space
));
3898 copy_i
= isl_union_map_copy(copy
);
3899 copy_i
= isl_union_map_intersect_range(copy_i
, extent
);
3900 set
= isl_set_copy(prog
->array
[i
].extent
);
3901 extent
= isl_union_set_from_set(set
);
3902 domain
= isl_union_map_domain(copy_i
);
3903 copy_i
= isl_union_map_from_domain_and_range(domain
, extent
);
3904 res
= isl_union_map_union(res
, copy_i
);
3907 isl_union_map_free(copy
);
3912 /* Insert "kernel" marks that point to a ppcg_kernel structure
3913 * in front of all outermost tilable band that (by construction)
3914 * have at least one parallel loop.
3916 static __isl_give isl_schedule_node
*mark_kernels(struct gpu_gen
*gen
,
3917 __isl_take isl_schedule_node
*node
)
3919 return isl_schedule_node_map_descendant(node
,
3920 &mark_outer_permutable
, gen
);
3923 /* Save the schedule "schedule" to a file called "filename".
3924 * The schedule is printed in block style.
3926 static void save_schedule(__isl_keep isl_schedule
*schedule
,
3927 const char *filename
)
3936 file
= fopen(filename
, "w");
3938 fprintf(stderr
, "Unable to open '%s' for writing\n", filename
);
3941 ctx
= isl_schedule_get_ctx(schedule
);
3942 p
= isl_printer_to_file(ctx
, file
);
3943 p
= isl_printer_set_yaml_style(p
, ISL_YAML_STYLE_BLOCK
);
3944 p
= isl_printer_print_schedule(p
, schedule
);
3945 isl_printer_free(p
);
3949 /* Load and return a schedule from a file called "filename".
3951 static __isl_give isl_schedule
*load_schedule(isl_ctx
*ctx
,
3952 const char *filename
)
3955 isl_schedule
*schedule
;
3957 file
= fopen(filename
, "r");
3959 fprintf(stderr
, "Unable to open '%s' for reading\n", filename
);
3962 schedule
= isl_schedule_read_from_file(ctx
, file
);
3968 /* Compute an appropriate schedule based on the accesses in
3969 * gen->read and gen->write.
3971 * We use the dependences in gen->prog->scop to compute
3972 * a schedule that has a parallel loop in each tilable band and
3973 * return this schedule.
3975 * If live range reordering is allowed, then we need to make sure
3976 * that live ranges on arrays are not run in parallel since doing
3977 * so would require array expansion. We therefore add the array
3978 * order dependences to the coincidence dependences. Non-zero array
3979 * order dependences will then prevent a schedule dimension from being
3980 * considered parallel.
3981 * Live ranges derived from scalars are allowed to be run in parallel
3982 * since we force the scalars to be mapped to private memory in
3983 * check_scalar_live_ranges.
3984 * If live range reordering is allowed, then the false dependences
3985 * are not added to the validity constraints as that would prevent
3986 * reordering. Instead, the external false dependences that enforce that reads
3987 * from potentially live-in data precede any later write and
3988 * that writes of potentially live-out data follow any other earlier write
3989 * are added to the validity and the coincidence constraints.
3990 * The false dependences are still added to the proximity constraints
3991 * for consistency with the case where live range reordering is not allowed.
3992 * The coincidence constraints then consist of flow dependences,
3993 * external false dependences and array order dependences.
3994 * The independences can be filtered out from the first two sets.
3995 * They have already been filtered out from the array order dependences
3996 * on a per array basis in collect_order_dependences.
3997 * There is no need for a per array handling of the other two sets
3998 * as there should be no flow or external false dependence on local
3999 * variables that can be filtered out.
4001 static __isl_give isl_schedule
*compute_schedule(struct gpu_gen
*gen
)
4003 isl_union_set
*domain
;
4004 isl_union_map
*dep_raw
, *dep
;
4005 isl_union_map
*validity
, *proximity
, *coincidence
;
4006 isl_schedule_constraints
*sc
;
4007 isl_schedule
*schedule
;
4009 domain
= isl_union_set_copy(gen
->prog
->scop
->domain
);
4010 sc
= isl_schedule_constraints_on_domain(domain
);
4011 sc
= isl_schedule_constraints_set_context(sc
,
4012 isl_set_copy(gen
->prog
->scop
->context
));
4013 if (gen
->options
->live_range_reordering
) {
4014 sc
= isl_schedule_constraints_set_conditional_validity(sc
,
4015 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_flow
),
4016 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_order
));
4017 proximity
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
4018 validity
= isl_union_map_copy(proximity
);
4019 validity
= isl_union_map_union(validity
,
4020 isl_union_map_copy(gen
->prog
->scop
->dep_forced
));
4021 proximity
= isl_union_map_union(proximity
,
4022 isl_union_map_copy(gen
->prog
->scop
->dep_false
));
4023 coincidence
= isl_union_map_copy(validity
);
4024 coincidence
= isl_union_map_subtract(coincidence
,
4025 isl_union_map_copy(gen
->prog
->scop
->independence
));
4026 coincidence
= isl_union_map_union(coincidence
,
4027 isl_union_map_copy(gen
->prog
->array_order
));
4029 dep_raw
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
4030 dep
= isl_union_map_copy(gen
->prog
->scop
->dep_false
);
4031 dep
= isl_union_map_union(dep
, dep_raw
);
4032 dep
= isl_union_map_coalesce(dep
);
4033 proximity
= isl_union_map_copy(dep
);
4034 coincidence
= isl_union_map_copy(dep
);
4037 sc
= isl_schedule_constraints_set_validity(sc
, validity
);
4038 sc
= isl_schedule_constraints_set_coincidence(sc
, coincidence
);
4039 sc
= isl_schedule_constraints_set_proximity(sc
, proximity
);
4041 if (gen
->options
->debug
->dump_schedule_constraints
)
4042 isl_schedule_constraints_dump(sc
);
4043 schedule
= isl_schedule_constraints_compute_schedule(sc
);
4048 /* Obtain a schedule for the scop, either by reading it from
4049 * a file or by computing one.
4051 static __isl_give isl_schedule
*get_schedule(struct gpu_gen
*gen
)
4053 isl_schedule
*schedule
;
4055 if (gen
->options
->load_schedule_file
) {
4056 schedule
= load_schedule(gen
->ctx
,
4057 gen
->options
->load_schedule_file
);
4059 schedule
= compute_schedule(gen
);
4060 if (gen
->options
->save_schedule_file
)
4061 save_schedule(schedule
,
4062 gen
->options
->save_schedule_file
);
4064 if (gen
->options
->debug
->dump_schedule
)
4065 isl_schedule_dump(schedule
);
4070 /* Construct the string "<a>_<b>".
4072 static char *concat(isl_ctx
*ctx
, const char *a
, const char *b
)
4077 p
= isl_printer_to_str(ctx
);
4078 p
= isl_printer_print_str(p
, a
);
4079 p
= isl_printer_print_str(p
, "_");
4080 p
= isl_printer_print_str(p
, b
);
4081 s
= isl_printer_get_str(p
);
4082 isl_printer_free(p
);
4087 /* For each array in "prog" of which an element appears in "accessed" and
4088 * that is not a read only scalar, create a zero-dimensional universe set
4089 * of which the tuple id has name "<prefix>_<name of array>" and a user
4090 * pointer pointing to the array (gpu_array_info).
4092 * If the array is local to "prog", then make sure it will be declared
4095 * Return the list of these universe sets.
4097 static __isl_give isl_union_set_list
*create_copy_filters(struct gpu_prog
*prog
,
4098 const char *prefix
, __isl_take isl_union_set
*accessed
)
4102 isl_union_set_list
*filters
;
4105 filters
= isl_union_set_list_alloc(ctx
, 0);
4106 for (i
= 0; i
< prog
->n_array
; ++i
) {
4107 struct gpu_array_info
*array
= &prog
->array
[i
];
4109 isl_set
*accessed_i
;
4113 isl_union_set
*uset
;
4115 if (gpu_array_is_read_only_scalar(array
))
4118 space
= isl_space_copy(array
->space
);
4119 accessed_i
= isl_union_set_extract_set(accessed
, space
);
4120 empty
= isl_set_plain_is_empty(accessed_i
);
4121 isl_set_free(accessed_i
);
4123 filters
= isl_union_set_list_free(filters
);
4130 array
->declare_local
= 1;
4132 name
= concat(ctx
, prefix
, array
->name
);
4133 id
= name
? isl_id_alloc(ctx
, name
, array
) : NULL
;
4135 space
= isl_space_set_alloc(ctx
, 0, 0);
4136 space
= isl_space_set_tuple_id(space
, isl_dim_set
, id
);
4137 uset
= isl_union_set_from_set(isl_set_universe(space
));
4139 filters
= isl_union_set_list_add(filters
, uset
);
4141 isl_union_set_free(accessed
);
4146 /* Make sure that code for the statements in "filters" that
4147 * copy arrays to or from the device is only generated when
4148 * the size of the corresponding array is positive.
4149 * That is, add a set node underneath "graft" with "filters" as children
4150 * and for each child add a guard that the selects the parameter
4151 * values for which the corresponding array has a positive size.
4152 * The array is available in the user pointer of the statement identifier.
4153 * "depth" is the schedule depth of the position where "graft"
4156 static __isl_give isl_schedule_node
*insert_positive_size_guards(
4157 __isl_take isl_schedule_node
*graft
,
4158 __isl_take isl_union_set_list
*filters
, int depth
)
4162 graft
= isl_schedule_node_child(graft
, 0);
4163 graft
= isl_schedule_node_insert_set(graft
, filters
);
4164 n
= isl_schedule_node_n_children(graft
);
4165 for (i
= 0; i
< n
; ++i
) {
4166 isl_union_set
*filter
;
4167 isl_set
*domain
, *guard
;
4169 struct gpu_array_info
*array
;
4171 graft
= isl_schedule_node_child(graft
, i
);
4172 filter
= isl_schedule_node_filter_get_filter(graft
);
4173 domain
= isl_set_from_union_set(filter
);
4174 id
= isl_set_get_tuple_id(domain
);
4175 array
= isl_id_get_user(id
);
4177 isl_set_free(domain
);
4178 guard
= gpu_array_positive_size_guard(array
);
4179 guard
= isl_set_from_params(guard
);
4180 guard
= isl_set_add_dims(guard
, isl_dim_set
, depth
);
4181 graft
= isl_schedule_node_child(graft
, 0);
4182 graft
= isl_schedule_node_insert_guard(graft
, guard
);
4183 graft
= isl_schedule_node_parent(graft
);
4184 graft
= isl_schedule_node_parent(graft
);
4186 graft
= isl_schedule_node_parent(graft
);
4191 /* Create a graft for copying arrays to or from the device,
4192 * whenever the size of the array is strictly positive.
4193 * Each statement is called "<prefix>_<name of array>" and
4194 * the identifier has a user pointer pointing to the array.
4195 * The graft will be added at the position specified by "node".
4196 * "copy" contains the array elements that need to be copied.
4197 * Only arrays of which some elements need to be copied
4198 * will have a corresponding statement in the graph.
4199 * Note though that each such statement will copy the entire array.
4201 static __isl_give isl_schedule_node
*create_copy_device(struct gpu_prog
*prog
,
4202 __isl_keep isl_schedule_node
*node
, const char *prefix
,
4203 __isl_take isl_union_set
*copy
)
4208 isl_union_set
*all
, *domain
;
4209 isl_union_set_list
*filters
;
4210 isl_union_map
*extension
;
4211 isl_schedule_node
*graft
;
4214 depth
= isl_schedule_node_get_schedule_depth(node
);
4215 filters
= create_copy_filters(prog
, prefix
, copy
);
4216 all
= isl_union_set_list_union(isl_union_set_list_copy(filters
));
4218 space
= depth
< 0 ? NULL
: isl_space_set_alloc(ctx
, 0, depth
);
4219 domain
= isl_union_set_from_set(isl_set_universe(space
));
4220 extension
= isl_union_map_from_domain_and_range(domain
, all
);
4221 graft
= isl_schedule_node_from_extension(extension
);
4224 return isl_schedule_node_free(graft
);
4225 if (isl_union_set_list_n_union_set(filters
) == 0) {
4226 isl_union_set_list_free(filters
);
4230 return insert_positive_size_guards(graft
, filters
, depth
);
4233 /* Return (the universe spaces of) the arrays that are declared
4234 * inside the scop corresponding to "prog" and for which all
4235 * potential writes inside the scop form a subset of "domain".
4237 static __isl_give isl_union_set
*extract_local_accesses(struct gpu_prog
*prog
,
4238 __isl_keep isl_union_set
*domain
)
4241 isl_union_set
*local
;
4243 local
= isl_union_set_empty(isl_union_set_get_space(domain
));
4245 for (i
= 0; i
< prog
->n_array
; ++i
) {
4247 isl_union_map
*to_outer
;
4248 isl_union_map
*may_write
;
4249 isl_union_set
*write_domain
;
4250 isl_union_set
*fields
;
4253 if (!prog
->array
[i
].local
)
4256 set
= isl_set_universe(isl_space_copy(prog
->array
[i
].space
));
4257 to_outer
= isl_union_map_copy(prog
->to_outer
);
4258 to_outer
= isl_union_map_intersect_range(to_outer
,
4259 isl_union_set_from_set(isl_set_copy(set
)));
4260 fields
= isl_union_map_domain(to_outer
);
4261 may_write
= isl_union_map_copy(prog
->may_write
);
4262 may_write
= isl_union_map_intersect_range(may_write
, fields
);
4263 write_domain
= isl_union_map_domain(may_write
);
4264 subset
= isl_union_set_is_subset(write_domain
, domain
);
4265 isl_union_set_free(write_domain
);
4269 return isl_union_set_free(local
);
4270 } else if (subset
) {
4271 local
= isl_union_set_add_set(local
, set
);
4280 /* Internal data structure for node_may_persist.
4282 * "tagger" maps tagged iteration domains to the corresponding untagged
4285 * "may_persist_flow" is the set of all tagged dataflow dependences
4286 * with those dependences removed that either precede or follow
4287 * the kernel launch in a sequence.
4288 * "inner_band_flow" is the set of all tagged dataflow dependences
4289 * that are local to a given iteration of the outer band nodes
4290 * with respect to the current node.
4291 * "local_flow" is equal to "inner_band_flow", except that the domain
4292 * and the range have been intersected with intermediate filters
4293 * on children of sets or sequences.
4295 struct ppcg_may_persist_data
{
4296 isl_union_pw_multi_aff
*tagger
;
4298 isl_union_map
*local_flow
;
4299 isl_union_map
*inner_band_flow
;
4300 isl_union_map
*may_persist_flow
;
4303 /* Update the information in "data" based on the band ancestor "node".
4305 * In particular, we restrict the dependences in data->local_flow
4306 * to those dependence where the source and the sink occur in
4307 * the same iteration of the given band node.
4308 * We also update data->inner_band_flow to the new value of
4311 static int update_may_persist_at_band(__isl_keep isl_schedule_node
*node
,
4312 struct ppcg_may_persist_data
*data
)
4314 isl_multi_union_pw_aff
*partial
;
4315 isl_union_pw_multi_aff
*contraction
;
4316 isl_union_map
*flow
;
4318 if (isl_schedule_node_band_n_member(node
) == 0)
4321 partial
= isl_schedule_node_band_get_partial_schedule(node
);
4322 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4323 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4325 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4326 isl_union_pw_multi_aff_copy(data
->tagger
));
4328 flow
= data
->local_flow
;
4329 flow
= isl_union_map_eq_at_multi_union_pw_aff(flow
, partial
);
4330 data
->local_flow
= flow
;
4332 isl_union_map_free(data
->inner_band_flow
);
4333 data
->inner_band_flow
= isl_union_map_copy(data
->local_flow
);
4338 /* Given a set of local reaching domain elements "domain",
4339 * expand them to the corresponding leaf domain elements using "contraction"
4340 * and insert the array references tags using data->tagger.
4342 static __isl_give isl_union_set
*expand_and_tag(
4343 __isl_take isl_union_set
*domain
,
4344 __isl_take isl_union_pw_multi_aff
*contraction
,
4345 struct ppcg_may_persist_data
*data
)
4347 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4349 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4350 isl_union_pw_multi_aff_copy(data
->tagger
));
4354 /* Given a filter node that is the child of a set or sequence node,
4355 * restrict data->local_flow to refer only to those elements
4356 * in the filter of the node.
4357 * "contraction" maps the leaf domain elements of the schedule tree
4358 * to the corresponding domain elements at (the parent of) "node".
4360 static int filter_flow(__isl_keep isl_schedule_node
*node
,
4361 struct ppcg_may_persist_data
*data
,
4362 __isl_take isl_union_pw_multi_aff
*contraction
)
4364 isl_union_set
*filter
;
4365 isl_union_map
*flow
;
4367 flow
= data
->local_flow
;
4368 filter
= isl_schedule_node_filter_get_filter(node
);
4369 filter
= expand_and_tag(filter
, contraction
, data
);
4370 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(filter
));
4371 flow
= isl_union_map_intersect_range(flow
, filter
);
4372 data
->local_flow
= flow
;
4377 /* Given a filter node "node", collect the filters on all preceding siblings
4378 * (which are also filter nodes), add them to "filters" and return the result.
4380 static __isl_give isl_union_set
*add_previous_filters(
4381 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4383 isl_schedule_node
*sibling
;
4385 sibling
= isl_schedule_node_copy(node
);
4386 while (sibling
&& isl_schedule_node_has_previous_sibling(sibling
)) {
4387 isl_union_set
*filter
;
4389 sibling
= isl_schedule_node_previous_sibling(sibling
);
4390 filter
= isl_schedule_node_filter_get_filter(sibling
);
4391 filters
= isl_union_set_union(filters
, filter
);
4393 isl_schedule_node_free(sibling
);
4395 return isl_union_set_free(filters
);
4400 /* Given a filter node "node", collect the filters on all following siblings
4401 * (which are also filter nodes), add them to "filters" and return the result.
4403 static __isl_give isl_union_set
*add_next_filters(
4404 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4406 isl_schedule_node
*sibling
;
4408 sibling
= isl_schedule_node_copy(node
);
4409 while (sibling
&& isl_schedule_node_has_next_sibling(sibling
)) {
4410 isl_union_set
*filter
;
4412 sibling
= isl_schedule_node_next_sibling(sibling
);
4413 filter
= isl_schedule_node_filter_get_filter(sibling
);
4414 filters
= isl_union_set_union(filters
, filter
);
4416 isl_schedule_node_free(sibling
);
4418 return isl_union_set_free(filters
);
4423 /* Remove those flow dependences from data->may_persist_flow
4424 * that flow between elements of "domain" within the same iteration
4425 * of all outer band nodes.
4426 * "contraction" maps the leaf domain elements of the schedule tree
4427 * to the corresponding elements "domain".
4429 static void remove_external_flow(struct ppcg_may_persist_data
*data
,
4430 __isl_take isl_union_set
*domain
,
4431 __isl_keep isl_union_pw_multi_aff
*contraction
)
4433 isl_union_map
*flow
;
4435 contraction
= isl_union_pw_multi_aff_copy(contraction
);
4436 domain
= expand_and_tag(domain
, contraction
, data
);
4437 flow
= isl_union_map_copy(data
->local_flow
);
4438 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(domain
));
4439 flow
= isl_union_map_intersect_range(flow
, domain
);
4441 data
->may_persist_flow
= isl_union_map_subtract(data
->may_persist_flow
,
4445 /* Update the information in "data" based on the filter ancestor "node".
4446 * We only need to modify anything if the filter is the child
4447 * of a set or sequence node.
4449 * In the case of a sequence, we remove the dependences between
4450 * statement instances that are both executed either before or
4451 * after the subtree that will be mapped to a kernel, within
4452 * the same iteration of outer bands.
4454 * In both cases, we restrict data->local_flow to the current child.
4456 static int update_may_persist_at_filter(__isl_keep isl_schedule_node
*node
,
4457 struct ppcg_may_persist_data
*data
)
4459 enum isl_schedule_node_type type
;
4460 isl_schedule_node
*parent
;
4462 isl_union_pw_multi_aff
*contraction
;
4463 isl_union_set
*before
, *after
, *filter
;
4464 isl_union_map
*flow
;
4466 type
= isl_schedule_node_get_parent_type(node
);
4467 if (type
!= isl_schedule_node_sequence
&& type
!= isl_schedule_node_set
)
4470 parent
= isl_schedule_node_copy(node
);
4471 parent
= isl_schedule_node_parent(parent
);
4472 contraction
= isl_schedule_node_get_subtree_contraction(parent
);
4473 isl_schedule_node_free(parent
);
4475 if (type
== isl_schedule_node_set
)
4476 return filter_flow(node
, data
, contraction
);
4478 filter
= isl_schedule_node_filter_get_filter(node
);
4479 space
= isl_union_set_get_space(filter
);
4480 isl_union_set_free(filter
);
4481 before
= isl_union_set_empty(space
);
4482 after
= isl_union_set_copy(before
);
4483 before
= add_previous_filters(before
, node
);
4484 after
= add_next_filters(after
, node
);
4486 remove_external_flow(data
, before
, contraction
);
4487 remove_external_flow(data
, after
, contraction
);
4489 return filter_flow(node
, data
, contraction
);
4492 /* Update the information in "data" based on the ancestor "node".
4494 static int update_may_persist_at(__isl_keep isl_schedule_node
*node
, void *user
)
4496 struct ppcg_may_persist_data
*data
= user
;
4498 switch (isl_schedule_node_get_type(node
)) {
4499 case isl_schedule_node_error
:
4501 case isl_schedule_node_context
:
4502 case isl_schedule_node_domain
:
4503 case isl_schedule_node_expansion
:
4504 case isl_schedule_node_extension
:
4505 case isl_schedule_node_guard
:
4506 case isl_schedule_node_leaf
:
4507 case isl_schedule_node_mark
:
4508 case isl_schedule_node_sequence
:
4509 case isl_schedule_node_set
:
4511 case isl_schedule_node_band
:
4512 if (update_may_persist_at_band(node
, data
) < 0)
4515 case isl_schedule_node_filter
:
4516 if (update_may_persist_at_filter(node
, data
) < 0)
4524 /* Determine the set of array elements that may need to be perserved
4525 * by a kernel constructed from the subtree at "node".
4526 * This includes the set of array elements that may need to be preserved
4527 * by the entire scop (prog->may_persist) and the elements for which
4528 * there is a potential flow dependence that may cross a kernel launch.
4530 * To determine the second set, we start from all flow dependences.
4531 * From this set of dependences, we remove those that cannot possibly
4532 * require data to be preserved by a kernel launch.
4533 * In particular, we consider the following sets of dependences.
4534 * - dependences of which the write occurs inside the kernel.
4535 * If the data is needed outside the kernel, then it will
4536 * be copied out immediately after the kernel launch, so there
4537 * is no need for any special care.
4538 * - dependences of which the read occurs inside the kernel and the
4539 * corresponding write occurs inside the same iteration of the
4540 * outer band nodes. This means that the data is needed in
4541 * the first kernel launch after the write, which is already
4542 * taken care of by the standard copy-in. That is, the data
4543 * do not need to be preserved by any intermediate call to
4545 * - dependences of which the write and the read either both occur
4546 * before the kernel launch or both occur after the kernel launch,
4547 * within the same iteration of the outer band nodes with respect
4548 * to the sequence that determines the ordering of the dependence
4549 * and the kernel launch. Such flow dependences cannot cross
4550 * any kernel launch.
4552 * For the remaining (tagged) dependences, we take the domain
4553 * (i.e., the tagged writes) and apply the tagged access relation
4554 * to obtain the accessed data elements.
4555 * These are then combined with the elements that may need to be
4556 * preserved by the entire scop.
4558 static __isl_give isl_union_set
*node_may_persist(
4559 __isl_keep isl_schedule_node
*node
, struct gpu_prog
*prog
)
4561 struct ppcg_may_persist_data data
;
4562 isl_schedule_node
*root
;
4563 isl_union_pw_multi_aff
*contraction
;
4564 isl_union_set
*domain
;
4565 isl_union_set
*persist
;
4566 isl_union_map
*flow
, *local_flow
;
4568 data
.tagger
= prog
->scop
->tagger
;
4570 flow
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
4571 data
.local_flow
= isl_union_map_copy(flow
);
4572 data
.inner_band_flow
= isl_union_map_copy(flow
);
4573 data
.may_persist_flow
= flow
;
4574 if (isl_schedule_node_foreach_ancestor_top_down(node
,
4575 &update_may_persist_at
, &data
) < 0)
4576 data
.may_persist_flow
=
4577 isl_union_map_free(data
.may_persist_flow
);
4578 flow
= data
.may_persist_flow
;
4579 isl_union_map_free(data
.local_flow
);
4581 domain
= isl_schedule_node_get_domain(node
);
4582 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4583 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4585 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4586 isl_union_pw_multi_aff_copy(data
.tagger
));
4587 flow
= isl_union_map_subtract_domain(flow
, isl_union_set_copy(domain
));
4588 local_flow
= data
.inner_band_flow
;
4589 local_flow
= isl_union_map_intersect_range(local_flow
, domain
);
4590 flow
= isl_union_map_subtract(flow
, local_flow
);
4592 persist
= isl_union_map_domain(flow
);
4593 persist
= isl_union_set_apply(persist
,
4594 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4595 persist
= isl_union_set_union(persist
,
4596 isl_union_set_copy(prog
->may_persist
));
4601 /* Add nodes for copying outer arrays in and out of the device
4602 * before and after the subtree "node", which contains one or more kernels.
4603 * "domain" contains the original reaching domain elements before
4604 * the kernels were created, i.e., before the contraction that
4605 * may have been performed in creating the kernels has been applied.
4606 * "prefix" contains the prefix schedule at that point, in terms
4607 * of the same original reaching domain elements.
4609 * We first compute the sets of outer array elements that need
4610 * to be copied in and out and then graft in the nodes for
4611 * performing this copying.
4613 * In particular, for each array that is possibly written anywhere in
4614 * the subtree "node" and that may be used after "node"
4615 * or that may be visible outside the corresponding scop,
4616 * we copy out its entire extent.
4618 * Any array elements that is read without first being written inside
4619 * the subtree "node" needs to be copied in.
4620 * Furthermore, if there are any array elements that
4621 * are copied out, but that may not be written inside "node, then
4622 * they also need to be copied in to ensure that the value after execution
4623 * is the same as the value before execution, at least for those array
4624 * elements that may have their values preserved by the scop or that
4625 * may be written before "node" and read after "node".
4626 * In case the array elements are structures, we need to take into
4627 * account that all members of the structures need to be written
4628 * by "node" before we can avoid copying the data structure in.
4630 * Note that the may_write relation is intersected with the domain,
4631 * which has been intersected with the context.
4632 * This helps in those cases where the arrays are declared with a fixed size,
4633 * while the accesses are parametric and the context assigns a fixed value
4634 * to the parameters.
4636 * If an element from a local array is read without first being written,
4637 * then there is no point in copying it in since it cannot have been
4638 * written prior to the scop. Warn about the uninitialized read instead.
4640 static __isl_give isl_schedule_node
*add_to_from_device(
4641 __isl_take isl_schedule_node
*node
, __isl_take isl_union_set
*domain
,
4642 __isl_take isl_union_map
*prefix
, struct gpu_prog
*prog
)
4644 isl_union_set
*local
;
4645 isl_union_set
*to_device
, *from_device
, *may_persist
;
4646 isl_union_map
*may_write
, *must_write
, *copy_out
, *not_written
;
4647 isl_union_map
*read
, *copy_in
;
4648 isl_union_map
*tagged
;
4649 isl_union_map
*local_uninitialized
;
4650 isl_schedule_node
*graft
;
4652 tagged
= isl_union_map_copy(prog
->scop
->tagged_reads
);
4653 tagged
= isl_union_map_union(tagged
,
4654 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4656 may_write
= isl_union_map_copy(prog
->may_write
);
4657 may_write
= isl_union_map_intersect_domain(may_write
,
4658 isl_union_set_copy(domain
));
4659 may_write
= remove_local_accesses(prog
,
4660 isl_union_map_copy(tagged
), may_write
,
4661 isl_union_map_copy(prefix
), 0);
4662 may_write
= isl_union_map_apply_range(may_write
,
4663 isl_union_map_copy(prog
->to_outer
));
4664 may_write
= isl_union_map_apply_domain(may_write
,
4665 isl_union_map_copy(prefix
));
4666 may_write
= approximate_copy_out(may_write
, prog
);
4667 copy_out
= isl_union_map_copy(may_write
);
4668 may_write
= isl_union_map_apply_range(may_write
,
4669 isl_union_map_copy(prog
->to_inner
));
4670 must_write
= isl_union_map_copy(prog
->must_write
);
4671 must_write
= isl_union_map_apply_domain(must_write
,
4672 isl_union_map_copy(prefix
));
4673 may_persist
= node_may_persist(node
, prog
);
4674 may_write
= isl_union_map_intersect_range(may_write
, may_persist
);
4675 not_written
= isl_union_map_subtract(may_write
, must_write
);
4677 local
= extract_local_accesses(prog
, domain
);
4678 read
= isl_union_map_copy(prog
->read
);
4679 read
= isl_union_map_intersect_domain(read
, domain
);
4680 read
= remove_local_accesses(prog
, tagged
, read
,
4681 isl_union_map_copy(prefix
), 1);
4682 local
= isl_union_set_apply(local
, isl_union_map_copy(prog
->to_inner
));
4683 local_uninitialized
= isl_union_map_copy(prog
->scop
->live_in
);
4684 local_uninitialized
= isl_union_map_intersect_range(local_uninitialized
,
4686 local_uninitialized
= isl_union_map_intersect(local_uninitialized
,
4687 isl_union_map_copy(read
));
4688 if (!isl_union_map_is_empty(local_uninitialized
)) {
4690 "possibly uninitialized reads (not copied in):\n");
4691 isl_union_map_dump(local_uninitialized
);
4693 read
= isl_union_map_subtract(read
, local_uninitialized
);
4694 read
= isl_union_map_apply_domain(read
, prefix
);
4695 copy_in
= isl_union_map_union(read
, not_written
);
4696 copy_in
= isl_union_map_apply_range(copy_in
,
4697 isl_union_map_copy(prog
->to_outer
));
4699 graft
= create_copy_device(prog
, node
, "to_device",
4700 isl_union_map_range(copy_in
));
4701 node
= isl_schedule_node_graft_before(node
, graft
);
4702 graft
= create_copy_device(prog
, node
, "from_device",
4703 isl_union_map_range(copy_out
));
4704 node
= isl_schedule_node_graft_after(node
, graft
);
4709 /* Update "schedule" for mapping to a GPU device.
4711 * In particular, insert a context node, create kernels for
4712 * each outermost tilable band and introduce node for copying array
4713 * in and out of the device.
4714 * If the child of the initial root points to a set node,
4715 * then children of this node that do not contain any tilable bands
4716 * are separated from the other children and are not mapped to
4719 static __isl_give isl_schedule
*map_to_device(struct gpu_gen
*gen
,
4720 __isl_take isl_schedule
*schedule
)
4722 isl_schedule_node
*node
;
4724 isl_union_set
*domain
;
4725 isl_union_map
*prefix
;
4727 context
= isl_set_copy(gen
->prog
->context
);
4728 context
= isl_set_from_params(context
);
4729 schedule
= isl_schedule_insert_context(schedule
, context
);
4731 node
= isl_schedule_get_root(schedule
);
4732 isl_schedule_free(schedule
);
4733 node
= isl_schedule_node_child(node
, 0);
4734 if (isl_schedule_node_get_type(node
) == isl_schedule_node_context
)
4735 node
= isl_schedule_node_child(node
, 0);
4736 node
= isolate_permutable_subtrees(node
, gen
->prog
);
4737 domain
= isl_schedule_node_get_domain(node
);
4738 prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
4739 node
= mark_kernels(gen
, node
);
4740 node
= add_to_from_device(node
, domain
, prefix
, gen
->prog
);
4741 schedule
= isl_schedule_node_get_schedule(node
);
4742 isl_schedule_node_free(node
);
4747 /* Internal data structure for extract_access.
4748 * "next_access" points to the end of a linked list that is extended
4749 * by extract_access.
4750 * "single_expression" is set if the access expressions belong to
4751 * an expression statement (i.e., a statement without internal control).
4752 * "any_to_outer" maps all intermediate arrays to their outer arrays.
4754 struct ppcg_extract_access_data
{
4755 struct gpu_stmt_access
**next_access
;
4756 int single_expression
;
4757 isl_union_map
*any_to_outer
;
4760 /* Given a tagged access relation to a single array "tagged", extract it
4761 * as a map, taking into account that the input may be empty.
4762 * If the access relation is empty, then it does not contain
4763 * any space information, so we try to recover it from the index
4765 * The space of the index expression is of the form I -> A,
4766 * with I the statement instances and A the array, or [I -> F] -> A,
4767 * with F the filters corresponding to arguments.
4768 * We first drop F, if present, obtaining I -> A.
4769 * Then we construct I -> R, with R the reference tag,
4770 * combine the two into I -> [R -> A] and uncurry to obtain
4771 * the final result [I -> R] -> A.
4772 * Note that the index expression may have a lower dimension
4773 * than that of the array, but this dimension is not used
4774 * if the access relation is empty.
4776 static __isl_give isl_map
*extract_single_tagged_access(
4777 __isl_take isl_union_map
*tagged
, __isl_keep pet_expr
*expr
)
4781 isl_space
*space
, *space2
;
4782 isl_multi_pw_aff
*index
;
4784 empty
= isl_union_map_is_empty(tagged
);
4788 return isl_map_from_union_map(tagged
);
4789 isl_union_map_free(tagged
);
4791 index
= pet_expr_access_get_index(expr
);
4792 space
= isl_multi_pw_aff_get_space(index
);
4793 isl_multi_pw_aff_free(index
);
4794 if (isl_space_domain_is_wrapping(space
))
4795 space
= isl_space_domain_factor_domain(space
);
4796 space2
= isl_space_copy(space
);
4797 space2
= isl_space_from_domain(isl_space_domain(space
));
4798 id
= pet_expr_access_get_ref_id(expr
);
4799 space2
= isl_space_set_tuple_id(space2
, isl_dim_out
, id
);
4800 space
= isl_space_range_product(space2
, space
);
4801 space
= isl_space_uncurry(space
);
4803 return isl_map_empty(space
);
4805 isl_union_map_free(tagged
);
4809 /* Extract a gpu_stmt_access from "expr", append it to the list
4810 * that ends in *data->next_access and update the end of the list.
4811 * If the access expression performs a write, then it is considered
4812 * exact only if it appears in a single expression statement and
4813 * if its may access relation is equal to its must access relation.
4815 * The combined set of may accesses may be union if member accesses
4816 * are involved, but the entire set is derived from a single reference and
4817 * therefore from a single index expression. These accesses therefore
4818 * all map to the same outer array.
4820 static int extract_access(__isl_keep pet_expr
*expr
, void *user
)
4822 struct ppcg_extract_access_data
*data
= user
;
4823 isl_union_map
*tagged
;
4824 struct gpu_stmt_access
*access
;
4825 isl_ctx
*ctx
= pet_expr_get_ctx(expr
);
4826 isl_multi_pw_aff
*index
;
4828 access
= isl_alloc_type(ctx
, struct gpu_stmt_access
);
4830 access
->next
= NULL
;
4831 access
->read
= pet_expr_access_is_read(expr
);
4832 access
->write
= pet_expr_access_is_write(expr
);
4833 tagged
= pet_expr_access_get_tagged_may_read(expr
);
4834 tagged
= isl_union_map_union(tagged
,
4835 pet_expr_access_get_tagged_may_write(expr
));
4836 tagged
= isl_union_map_apply_range(tagged
,
4837 isl_union_map_copy(data
->any_to_outer
));
4838 if (!access
->write
) {
4839 access
->exact_write
= 1;
4840 } else if (!data
->single_expression
) {
4841 access
->exact_write
= 0;
4843 isl_union_map
*must
, *may
;
4844 may
= isl_union_map_copy(tagged
);
4845 may
= isl_union_map_domain_factor_domain(may
);
4846 must
= pet_expr_access_get_must_write(expr
);
4847 access
->exact_write
= isl_union_map_is_equal(must
, may
);
4848 isl_union_map_free(must
);
4849 isl_union_map_free(may
);
4851 index
= pet_expr_access_get_index(expr
);
4852 access
->n_index
= isl_multi_pw_aff_dim(index
, isl_dim_out
);
4853 isl_multi_pw_aff_free(index
);
4854 access
->ref_id
= pet_expr_access_get_ref_id(expr
);
4855 access
->tagged_access
= extract_single_tagged_access(tagged
, expr
);
4856 access
->access
= isl_map_copy(access
->tagged_access
);
4857 access
->access
= isl_map_domain_factor_domain(access
->access
);
4859 *data
->next_access
= access
;
4860 data
->next_access
= &(*data
->next_access
)->next
;
4862 if (!access
->access
)
4868 /* Construct a linked list of gpu_stmt_access objects,
4869 * one for each access expression in the statement body.
4870 * "any_to_outer" maps all intermediate arrays to their outer arrays.
4872 static int pet_stmt_extract_accesses(struct gpu_stmt
*stmt
,
4873 __isl_keep isl_union_map
*any_to_outer
)
4875 struct ppcg_extract_access_data data
;
4877 stmt
->accesses
= NULL
;
4878 data
.next_access
= &stmt
->accesses
;
4879 data
.single_expression
=
4880 pet_tree_get_type(stmt
->stmt
->body
) == pet_tree_expr
;
4881 data
.any_to_outer
= any_to_outer
;
4882 return pet_tree_foreach_access_expr(stmt
->stmt
->body
,
4883 &extract_access
, &data
);
4886 /* Return an array of gpu_stmt representing the statements in "scop".
4888 static struct gpu_stmt
*extract_stmts(isl_ctx
*ctx
, struct ppcg_scop
*scop
,
4889 __isl_keep isl_set
*context
, __isl_keep isl_union_map
*any_to_outer
)
4892 struct gpu_stmt
*stmts
;
4894 stmts
= isl_calloc_array(ctx
, struct gpu_stmt
, scop
->pet
->n_stmt
);
4898 for (i
= 0; i
< scop
->pet
->n_stmt
; ++i
) {
4899 struct gpu_stmt
*s
= &stmts
[i
];
4901 s
->id
= isl_set_get_tuple_id(scop
->pet
->stmts
[i
]->domain
);
4902 s
->stmt
= scop
->pet
->stmts
[i
];
4903 if (pet_stmt_extract_accesses(s
, any_to_outer
) < 0)
4904 return free_stmts(stmts
, i
+ 1);
4910 /* Callback for ppcg_print_guarded that calls the callback for generate_gpu.
4912 static __isl_give isl_printer
*print_gpu(__isl_take isl_printer
*p
, void *user
)
4914 struct gpu_gen
*gen
= user
;
4916 return gen
->print(p
, gen
->prog
, gen
->tree
, &gen
->types
,
4920 /* Generate CUDA code for "scop" and print it to "p".
4921 * After generating an AST for the transformed scop as explained below,
4922 * we call "gen->print" to print the AST in the desired output format
4925 * If it turns out that it does not make sense to generate GPU code,
4926 * then we generate CPU code instead.
4928 * The GPU code is generated in a context where at least one
4929 * statement instance is executed. The corresponding guard (if any) is printed
4930 * around the entire generated GPU code, except for the declaration
4931 * of the arrays that are visible outside of the scop and that therefore
4932 * cannot be declared inside the body of any possible guard.
4934 * We first compute a schedule that respects the dependences
4935 * of the original program and select the outermost bands
4936 * of tilable dimensions that have at least one parallel loop.
4937 * If the --load-schedule is specified, then the loaded schedule
4938 * is used instead of a computed schedule.
4940 * Each of these bands B is then tiled according to "tile" sizes, resulting
4941 * in two nested bands, with a kernel marker on top
4949 * We then split off at most 2 parallel dimensions from the T band and
4950 * at most 3 parallel dimension from the P band
4963 * A filter is introduced in front of T1 that maps the domain instances
4964 * to block identifiers. Similarly, a filter is introduced in front of P1
4965 * that maps the domain instances to thread identifiers.
4967 * For each iteration of the T2 band and for each array, we compute
4968 * the array elements accessed by that iteration, construct a rectangular
4969 * box around it and shift it to the origin. The result is used
4970 * as shared memory for the array.
4972 * Copying and synchronization statements are added to this schedule tree.
4973 * In principle, these are added in front of the P1 band, but some of
4974 * them may get hoisted up to higher levels.
4976 * The entire AST is then generated from the single resulting schedule tree.
4977 * During the generation the subtrees at kernel nodes (K) are saved
4978 * aside and replaced by kernel calls. The result is printed as host code
4979 * while the saved subtrees are printed as device code.
4981 static __isl_give isl_printer
*generate(__isl_take isl_printer
*p
,
4982 struct gpu_gen
*gen
, struct ppcg_scop
*scop
,
4983 struct ppcg_options
*options
)
4985 struct gpu_prog
*prog
;
4987 isl_set
*context
, *guard
;
4988 isl_schedule
*schedule
;
4992 return isl_printer_free(p
);
4994 ctx
= isl_printer_get_ctx(p
);
4995 prog
= gpu_prog_alloc(ctx
, scop
);
4997 return isl_printer_free(p
);
4999 context
= isl_set_copy(prog
->context
);
5000 guard
= isl_union_set_params(isl_union_set_copy(prog
->scop
->domain
));
5001 prog
->context
= isl_set_intersect(prog
->context
, isl_set_copy(guard
));
5004 schedule
= get_schedule(gen
);
5006 any_permutable
= has_any_permutable_node(schedule
);
5007 if (any_permutable
< 0 || !any_permutable
) {
5008 isl_set_free(context
);
5009 isl_set_free(guard
);
5010 if (any_permutable
< 0)
5011 p
= isl_printer_free(p
);
5013 p
= print_cpu(p
, scop
, options
);
5014 isl_schedule_free(schedule
);
5016 schedule
= map_to_device(gen
, schedule
);
5017 gen
->tree
= generate_code(gen
, schedule
);
5018 p
= ppcg_print_exposed_declarations(p
, prog
->scop
);
5019 p
= ppcg_print_guarded(p
, guard
, context
, &print_gpu
, gen
);
5020 isl_ast_node_free(gen
->tree
);
5023 gpu_prog_free(prog
);
5028 /* Wrapper around generate for use as a ppcg_transform callback.
5030 static __isl_give isl_printer
*generate_wrap(__isl_take isl_printer
*p
,
5031 struct ppcg_scop
*scop
, void *user
)
5033 struct gpu_gen
*gen
= user
;
5035 return generate(p
, gen
, scop
, gen
->options
);
5038 /* Transform the code in the file called "input" by replacing
5039 * all scops by corresponding GPU code and write the results to "out".
5041 int generate_gpu(isl_ctx
*ctx
, const char *input
, FILE *out
,
5042 struct ppcg_options
*options
,
5043 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
5044 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
5045 struct gpu_types
*types
, void *user
), void *user
)
5052 gen
.sizes
= extract_sizes_from_str(ctx
, options
->sizes
);
5053 gen
.options
= options
;
5056 gen
.print_user
= user
;
5058 gen
.types
.name
= NULL
;
5060 if (options
->debug
->dump_sizes
) {
5061 isl_space
*space
= isl_space_params_alloc(ctx
, 0);
5062 gen
.used_sizes
= isl_union_map_empty(space
);
5065 r
= ppcg_transform(ctx
, input
, out
, options
, &generate_wrap
, &gen
);
5067 if (options
->debug
->dump_sizes
) {
5068 isl_union_map_dump(gen
.used_sizes
);
5069 isl_union_map_free(gen
.used_sizes
);
5072 isl_union_map_free(gen
.sizes
);
5073 for (i
= 0; i
< gen
.types
.n
; ++i
)
5074 free(gen
.types
.name
[i
]);
5075 free(gen
.types
.name
);
5080 /* Compute the set of inner array elements that may have their values
5081 * preserved by "prog". In particular, collect the array elements of
5082 * arrays that are not local to "prog" and remove those elements that
5083 * are definitely killed or definitely written by "prog".
5085 static __isl_give isl_union_set
*compute_may_persist(struct gpu_prog
*prog
)
5088 isl_union_set
*may_persist
, *killed
;
5089 isl_union_map
*must_kill
;
5091 may_persist
= isl_union_set_empty(isl_set_get_space(prog
->context
));
5092 for (i
= 0; i
< prog
->n_array
; ++i
) {
5095 if (prog
->array
[i
].local
)
5098 extent
= isl_set_copy(prog
->array
[i
].extent
);
5099 may_persist
= isl_union_set_add_set(may_persist
, extent
);
5102 may_persist
= isl_union_set_intersect_params(may_persist
,
5103 isl_set_copy(prog
->context
));
5104 may_persist
= isl_union_set_apply(may_persist
,
5105 isl_union_map_copy(prog
->to_inner
));
5106 must_kill
= isl_union_map_copy(prog
->tagged_must_kill
);
5107 killed
= isl_union_map_range(must_kill
);
5108 must_kill
= isl_union_map_copy(prog
->must_write
);
5109 killed
= isl_union_set_union(killed
, isl_union_map_range(must_kill
));
5111 may_persist
= isl_union_set_subtract(may_persist
, killed
);
5115 struct gpu_prog
*gpu_prog_alloc(isl_ctx
*ctx
, struct ppcg_scop
*scop
)
5117 struct gpu_prog
*prog
;
5124 prog
= isl_calloc_type(ctx
, struct gpu_prog
);
5129 prog
->context
= isl_set_copy(scop
->context
);
5130 prog
->n_stmts
= scop
->pet
->n_stmt
;
5131 prog
->any_to_outer
= pet_scop_compute_outer_to_any(scop
->pet
);
5132 prog
->any_to_outer
= isl_union_map_reverse(prog
->any_to_outer
);
5133 space
= isl_union_map_get_space(prog
->any_to_outer
);
5134 space
= isl_space_set_from_params(space
);
5135 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
5136 space
= isl_space_map_from_set(space
);
5137 id
= isl_map_identity(space
);
5138 prog
->any_to_outer
= isl_union_map_add_map(prog
->any_to_outer
, id
);
5139 prog
->stmts
= extract_stmts(ctx
, scop
,
5140 prog
->context
, prog
->any_to_outer
);
5141 prog
->read
= isl_union_map_copy(scop
->reads
);
5142 prog
->may_write
= isl_union_map_copy(scop
->may_writes
);
5143 prog
->must_write
= isl_union_map_copy(scop
->must_writes
);
5144 prog
->tagged_must_kill
= isl_union_map_copy(scop
->tagged_must_kills
);
5145 prog
->to_inner
= pet_scop_compute_outer_to_inner(scop
->pet
);
5146 prog
->to_outer
= isl_union_map_copy(prog
->to_inner
);
5147 prog
->to_outer
= isl_union_map_reverse(prog
->to_outer
);
5150 return gpu_prog_free(prog
);
5152 if (collect_array_info(prog
) < 0)
5153 return gpu_prog_free(prog
);
5154 prog
->may_persist
= compute_may_persist(prog
);
5159 void *gpu_prog_free(struct gpu_prog
*prog
)
5163 free_array_info(prog
);
5164 free_stmts(prog
->stmts
, prog
->n_stmts
);
5165 isl_union_map_free(prog
->any_to_outer
);
5166 isl_union_map_free(prog
->to_outer
);
5167 isl_union_map_free(prog
->to_inner
);
5168 isl_union_map_free(prog
->read
);
5169 isl_union_map_free(prog
->may_write
);
5170 isl_union_map_free(prog
->must_write
);
5171 isl_union_map_free(prog
->tagged_must_kill
);
5172 isl_union_map_free(prog
->array_order
);
5173 isl_union_set_free(prog
->may_persist
);
5174 isl_set_free(prog
->context
);