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 /* Does "array" need to be allocated on the device?
406 * If it is a read-only scalar, then it will be passed as an argument
407 * to the kernel and therefore does not require any allocation.
408 * If this device memory is not accessed at all, then it does not
409 * need to be allocated either.
411 int gpu_array_requires_device_allocation(struct gpu_array_info
*array
)
413 if (gpu_array_is_read_only_scalar(array
))
420 /* Return the set of parameter values for which the array has a positive
421 * size in all dimensions.
422 * If the sizes are only valid for some parameter values, then those
423 * constraints are also taken into account.
425 __isl_give isl_set
*gpu_array_positive_size_guard(struct gpu_array_info
*array
)
434 space
= isl_space_params(isl_space_copy(array
->space
));
435 guard
= isl_set_universe(space
);
437 for (i
= 0; i
< array
->n_index
; ++i
) {
439 isl_set
*guard_i
, *zero
;
441 bound
= isl_pw_aff_copy(array
->bound
[i
]);
442 guard_i
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(bound
));
443 zero
= isl_pw_aff_zero_set(bound
);
444 guard_i
= isl_set_subtract(guard_i
, zero
);
445 guard
= isl_set_intersect(guard
, guard_i
);
451 /* Internal data structure for extract_size_of_type.
452 * "type" specifies the name of the space that we want to extract.
453 * "res" is used to store the subset of that space.
455 struct ppcg_extract_size_data
{
460 /* This function is called for each set in a union_set.
461 * If the name of the set matches data->type, we store the
464 static isl_stat
extract_size_of_type(__isl_take isl_set
*size
, void *user
)
466 struct ppcg_extract_size_data
*data
= user
;
469 name
= isl_set_get_tuple_name(size
);
470 if (name
&& !strcmp(name
, data
->type
)) {
472 return isl_stat_error
;
479 /* Given a union map { kernel[i] -> *[...] },
480 * return the range in the space called "type" for the kernel with
481 * sequence number "id".
483 static __isl_give isl_set
*extract_sizes(__isl_keep isl_union_map
*sizes
,
484 const char *type
, int id
)
488 isl_union_set
*local_sizes
;
489 struct ppcg_extract_size_data data
= { type
, NULL
};
494 space
= isl_union_map_get_space(sizes
);
495 space
= isl_space_set_from_params(space
);
496 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
497 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
498 dom
= isl_set_universe(space
);
499 dom
= isl_set_fix_si(dom
, isl_dim_set
, 0, id
);
501 local_sizes
= isl_union_set_apply(isl_union_set_from_set(dom
),
502 isl_union_map_copy(sizes
));
503 isl_union_set_foreach_set(local_sizes
, &extract_size_of_type
, &data
);
504 isl_union_set_free(local_sizes
);
508 /* Given a singleton set, extract the first (at most *len) elements
509 * of the single integer tuple into *sizes and update *len if needed.
511 static void read_sizes_from_set(__isl_take isl_set
*set
, int *sizes
, int *len
)
519 dim
= isl_set_dim(set
, isl_dim_set
);
523 for (i
= 0; i
< *len
; ++i
) {
526 v
= isl_set_plain_get_val_if_fixed(set
, isl_dim_set
, i
);
529 sizes
[i
] = isl_val_get_num_si(v
);
536 /* Add the map { kernel[id] -> type[sizes] } to gen->used_sizes,
537 * if the option debug->dump_sizes is set.
539 static void set_used_sizes(struct gpu_gen
*gen
, const char *type
, int id
,
546 if (!gen
->options
->debug
->dump_sizes
)
549 space
= isl_union_map_get_space(gen
->used_sizes
);
550 space
= isl_space_set_from_params(space
);
551 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
552 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
553 space
= isl_space_from_domain(space
);
554 space
= isl_space_add_dims(space
, isl_dim_out
, len
);
555 space
= isl_space_set_tuple_name(space
, isl_dim_out
, type
);
557 map
= isl_map_universe(space
);
558 map
= isl_map_fix_si(map
, isl_dim_in
, 0, id
);
559 for (i
= 0; i
< len
; ++i
)
560 map
= isl_map_fix_si(map
, isl_dim_out
, i
, sizes
[i
]);
562 gen
->used_sizes
= isl_union_map_add_map(gen
->used_sizes
, map
);
565 /* Extract user specified "tile" sizes from the "sizes" command line option,
566 * defaulting to option->tile_size in each dimension.
567 * *tile_len contains the maximum number of tile sizes needed.
568 * Update *tile_len to the number of specified tile sizes, if any, and
569 * return a pointer to the tile sizes (or NULL on error).
570 * Add the effectively used sizes to gen->used_sizes.
572 static int *read_tile_sizes(struct gpu_gen
*gen
, int *tile_len
)
578 tile_size
= isl_alloc_array(gen
->ctx
, int, *tile_len
);
581 for (n
= 0; n
< *tile_len
; ++n
)
582 tile_size
[n
] = gen
->options
->tile_size
;
584 size
= extract_sizes(gen
->sizes
, "tile", gen
->kernel_id
);
585 read_sizes_from_set(size
, tile_size
, tile_len
);
586 set_used_sizes(gen
, "tile", gen
->kernel_id
, tile_size
, *tile_len
);
591 /* Extract user specified "block" sizes from the "sizes" command line option,
592 * after filling in some potentially useful defaults.
594 static void read_block_sizes(struct ppcg_kernel
*kernel
,
595 __isl_keep isl_union_map
*sizes
)
599 if (kernel
->n_block
> 3)
601 switch (kernel
->n_block
) {
603 kernel
->block_dim
[0] = 512;
606 kernel
->block_dim
[0] = 32;
607 kernel
->block_dim
[1] = 16;
610 kernel
->block_dim
[0] = 32;
611 kernel
->block_dim
[1] = 4;
612 kernel
->block_dim
[2] = 4;
616 size
= extract_sizes(sizes
, "block", kernel
->id
);
617 read_sizes_from_set(size
, kernel
->block_dim
, &kernel
->n_block
);
620 /* Extract user specified "grid" sizes from the "sizes" command line option,
621 * after filling in some potentially useful defaults.
623 static void read_grid_sizes(struct ppcg_kernel
*kernel
,
624 __isl_keep isl_union_map
*sizes
)
628 if (kernel
->n_grid
> 2)
630 switch (kernel
->n_grid
) {
632 kernel
->grid_dim
[0] = 32768;
635 kernel
->grid_dim
[0] = 256;
636 kernel
->grid_dim
[1] = 256;
640 size
= extract_sizes(sizes
, "grid", kernel
->id
);
641 read_sizes_from_set(size
, kernel
->grid_dim
, &kernel
->n_grid
);
644 /* Extract user specified grid and block sizes from the gen->sizes
645 * command line option after filling in some potentially useful defaults.
646 * Store the extracted sizes in "kernel".
647 * Add the effectively used sizes to gen->used_sizes.
649 static void read_grid_and_block_sizes(struct ppcg_kernel
*kernel
,
652 read_block_sizes(kernel
, gen
->sizes
);
653 read_grid_sizes(kernel
, gen
->sizes
);
654 set_used_sizes(gen
, "block", kernel
->id
,
655 kernel
->block_dim
, kernel
->n_block
);
656 set_used_sizes(gen
, "grid", kernel
->id
,
657 kernel
->grid_dim
, kernel
->n_grid
);
660 static void *free_stmts(struct gpu_stmt
*stmts
, int n
)
667 for (i
= 0; i
< n
; ++i
) {
668 struct gpu_stmt_access
*access
, *next
;
670 for (access
= stmts
[i
].accesses
; access
; access
= next
) {
672 isl_id_free(access
->ref_id
);
673 isl_map_free(access
->access
);
674 isl_map_free(access
->tagged_access
);
678 isl_id_free(stmts
[i
].id
);
685 /* Add parameters p[i] with identifiers "ids" to "set",
686 * with bounds to 0 <= p[i] < size[i].
688 __isl_give isl_set
*add_bounded_parameters(__isl_take isl_set
*set
,
689 int *size
, __isl_keep isl_id_list
*ids
)
694 len
= isl_id_list_n_id(ids
);
695 nparam
= isl_set_dim(set
, isl_dim_param
);
696 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
698 for (i
= 0; i
< len
; ++i
) {
701 id
= isl_id_list_get_id(ids
, i
);
702 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
703 set
= isl_set_lower_bound_si(set
, isl_dim_param
, nparam
+ i
, 0);
704 set
= isl_set_upper_bound_si(set
, isl_dim_param
,
705 nparam
+ i
, size
[i
] - 1);
711 /* Add "len" parameters p[i] with identifiers "ids" and intersect "set"
714 * { : 0 <= p[i] < size[i] }
716 * or an overapproximation.
718 static __isl_give isl_set
*add_bounded_parameters_dynamic(
719 __isl_take isl_set
*set
, __isl_keep isl_multi_pw_aff
*size
,
720 __isl_keep isl_id_list
*ids
)
727 len
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
728 nparam
= isl_set_dim(set
, isl_dim_param
);
729 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
731 for (i
= 0; i
< len
; ++i
) {
734 id
= isl_id_list_get_id(ids
, i
);
735 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
738 space
= isl_space_params(isl_set_get_space(set
));
739 ls
= isl_local_space_from_space(space
);
740 for (i
= 0; i
< len
; ++i
) {
741 isl_pw_aff
*param
, *size_i
, *zero
;
744 param
= isl_pw_aff_var_on_domain(isl_local_space_copy(ls
),
745 isl_dim_param
, nparam
+ i
);
747 size_i
= isl_multi_pw_aff_get_pw_aff(size
, i
);
748 bound
= isl_pw_aff_lt_set(isl_pw_aff_copy(param
), size_i
);
749 bound
= isl_set_from_basic_set(isl_set_simple_hull(bound
));
750 set
= isl_set_intersect_params(set
, bound
);
752 zero
= isl_pw_aff_zero_on_domain(isl_local_space_copy(ls
));
753 bound
= isl_pw_aff_ge_set(param
, zero
);
754 set
= isl_set_intersect_params(set
, bound
);
756 isl_local_space_free(ls
);
761 /* Return the union of all tagged access relations in the group.
763 static __isl_give isl_union_map
*group_tagged_access_relation(
764 struct gpu_array_ref_group
*group
)
767 isl_union_map
*access
;
769 access
= isl_union_map_empty(isl_map_get_space(group
->access
));
770 for (i
= 0; i
< group
->n_ref
; ++i
) {
773 map_i
= isl_map_copy(group
->refs
[i
]->tagged_access
);
774 access
= isl_union_map_union(access
,
775 isl_union_map_from_map(map_i
));
781 /* Return the extent of "array", recomputed from the bounds.
782 * The recomputed extent may be simpler than the original extent.
784 static __isl_give isl_set
*array_extent(struct gpu_array_info
*array
)
792 id
= isl_set_get_tuple_id(array
->extent
);
793 space
= isl_set_get_space(array
->extent
);
794 extent
= isl_set_universe(isl_space_copy(space
));
795 ls
= isl_local_space_from_space(space
);
796 for (i
= 0; i
< array
->n_index
; ++i
) {
802 extent
= isl_set_lower_bound_si(extent
, isl_dim_set
, i
, 0);
804 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
806 index
= isl_pw_aff_from_aff(aff
);
807 bound
= isl_pw_aff_copy(array
->bound
[i
]);
808 bound
= isl_pw_aff_from_range(bound
);
809 bound
= isl_pw_aff_add_dims(bound
, isl_dim_in
, array
->n_index
);
810 bound
= isl_pw_aff_set_tuple_id(bound
, isl_dim_in
,
812 lt
= isl_pw_aff_lt_set(index
, bound
);
813 extent
= isl_set_intersect(extent
, lt
);
815 isl_local_space_free(ls
);
821 /* Return a map from the first group->depth dimensions of the computed
822 * schedule to the array tile in
823 * global memory that corresponds to the shared memory copy.
825 * In particular, return a map
831 * tile_offset(i) <= a <= tile_offset(i) + tile_size - 1 (1)
835 * 0 <= a <= array_size - 1 (2)
837 * Note that if some stride has been detected (i.e., when
838 * group->shared_tile->bound[i].shift is set), then a in (1) refers
839 * to the shifted and scaled down version.
841 * Constraints (1) are obtained by mapping the size constraints on the
842 * shared/private memory tile back to the access relation.
843 * Constraints (2) are obtained from the (recomputed) extent.
845 static __isl_give isl_map
*group_tile(struct gpu_array_ref_group
*group
)
848 int n_index
= group
->array
->n_index
;
854 space
= isl_multi_aff_get_space(group
->shared_tile
->tiling
);
855 space
= isl_space_range(space
);
856 local
= isl_set_universe(space
);
857 for (i
= 0; i
< n_index
; ++i
) {
860 local
= isl_set_lower_bound_si(local
, isl_dim_set
, i
, 0);
861 bound
= isl_val_copy(group
->shared_tile
->bound
[i
].size
);
862 bound
= isl_val_sub_ui(bound
, 1);
863 local
= isl_set_upper_bound_val(local
, isl_dim_set
, i
, bound
);
865 local
= isl_set_preimage_multi_aff(local
,
866 isl_multi_aff_copy(group
->shared_tile
->tiling
));
867 tile
= isl_set_unwrap(local
);
868 extent
= array_extent(group
->array
);
869 tile
= isl_map_intersect_range(tile
, extent
);
874 /* Given a mapping "iterator_map" from the AST schedule to a domain,
875 * return the corresponding mapping from the AST schedule to
876 * to the outer kernel->shared_schedule_dim dimensions of
877 * the schedule computed by PPCG for this kernel.
879 * Note that kernel->shared_schedule_dim is at least as large as
880 * the largest depth of any array reference group associated to the kernel.
881 * This is needed as the returned schedule is used to extract a mapping
882 * to the outer group->depth dimensions in transform_index.
884 static __isl_give isl_pw_multi_aff
*compute_sched_to_shared(
885 struct ppcg_kernel
*kernel
, __isl_take isl_pw_multi_aff
*iterator_map
)
887 isl_union_pw_multi_aff
*upma
;
888 isl_pw_multi_aff
*pma
;
891 space
= isl_space_range(isl_pw_multi_aff_get_space(iterator_map
));
892 space
= isl_space_from_domain(space
);
893 space
= isl_space_add_dims(space
, isl_dim_out
,
894 kernel
->shared_schedule_dim
);
896 upma
= isl_union_pw_multi_aff_copy(kernel
->shared_schedule
);
897 pma
= isl_union_pw_multi_aff_extract_pw_multi_aff(upma
, space
);
898 isl_union_pw_multi_aff_free(upma
);
900 return isl_pw_multi_aff_pullback_pw_multi_aff(pma
, iterator_map
);
903 /* If max_shared_memory is not set to infinity (-1), then make
904 * sure that the total amount of shared memory required by the
905 * array reference groups mapped to shared memory by "kernel"
906 * is no larger than this maximum.
908 * We apply a greedy approach and discard (keep in global memory)
909 * those groups that would result in a total memory size that
910 * is larger than the maximum.
912 * This function should be called after any function that may
913 * affect the decision on whether to place a reference group
914 * in private, shared or global memory.
916 static void check_shared_memory_bound(struct ppcg_kernel
*kernel
)
919 isl_val
*left
, *size
;
921 if (kernel
->options
->max_shared_memory
< 0)
924 left
= isl_val_int_from_si(kernel
->ctx
,
925 kernel
->options
->max_shared_memory
);
927 for (i
= 0; i
< kernel
->n_array
; ++i
) {
928 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
930 for (j
= 0; j
< local
->n_group
; ++j
) {
931 struct gpu_array_ref_group
*group
;
933 group
= local
->groups
[j
];
934 if (group
->private_tile
)
936 if (!group
->shared_tile
)
939 size
= gpu_array_tile_size(group
->shared_tile
);
940 size
= isl_val_mul_ui(size
, local
->array
->size
);
942 if (isl_val_le(size
, left
)) {
943 left
= isl_val_sub(left
, size
);
949 gpu_array_tile_free(group
->shared_tile
);
956 /* Mark all arrays of "kernel" that have an array reference group
957 * that is not mapped to private or shared memory as
958 * accessing the corresponding global device memory.
960 static void mark_global_arrays(struct ppcg_kernel
*kernel
)
964 for (i
= 0; i
< kernel
->n_array
; ++i
) {
965 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
969 for (j
= 0; j
< local
->n_group
; ++j
) {
970 if (gpu_array_ref_group_tile(local
->groups
[j
]))
974 local
->array
->global
= 1;
980 /* Compute a tiling for all the array reference groups in "kernel".
982 static void compute_group_tilings(struct ppcg_kernel
*kernel
)
986 for (i
= 0; i
< kernel
->n_array
; ++i
) {
987 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
989 for (j
= 0; j
< array
->n_group
; ++j
)
990 gpu_array_ref_group_compute_tiling(array
->groups
[j
]);
994 /* Compute the size of a bounding box around the origin and "set",
995 * where "set" is assumed to contain only non-negative elements.
996 * In particular, compute the maximal value of "set" in each direction
999 static __isl_give isl_multi_pw_aff
*extract_size(__isl_take isl_set
*set
,
1000 __isl_take isl_set
*context
)
1003 isl_multi_pw_aff
*mpa
;
1005 context
= isl_set_params(context
);
1006 n
= isl_set_dim(set
, isl_dim_set
);
1007 mpa
= isl_multi_pw_aff_zero(isl_set_get_space(set
));
1008 for (i
= 0; i
< n
; ++i
) {
1013 bound
= isl_set_dim_max(isl_set_copy(set
), i
);
1014 bound
= isl_pw_aff_coalesce(bound
);
1015 bound
= isl_pw_aff_gist(bound
, isl_set_copy(context
));
1017 space
= isl_pw_aff_get_domain_space(bound
);
1018 one
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1019 one
= isl_aff_add_constant_si(one
, 1);
1020 bound
= isl_pw_aff_add(bound
, isl_pw_aff_from_aff(one
));
1021 mpa
= isl_multi_pw_aff_set_pw_aff(mpa
, i
, bound
);
1024 isl_set_free(context
);
1029 /* Compute the effective grid size as a list of the sizes in each dimension.
1031 * The grid size specified by the user or set by default
1032 * in read_grid_sizes() and applied by the block filter,
1033 * may be too large for the given code in the sense that
1034 * it may contain blocks that don't need to execute anything.
1035 * We therefore don't return this grid size, but instead the
1036 * smallest grid size that ensures that all blocks that actually
1037 * execute code are included in the grid.
1039 * We first extract a description of the grid, i.e., the possible values
1040 * of the block ids, from the domain elements in "domain" and
1041 * kernel->block_filter.
1042 * The block ids are parameters in kernel->block_filter.
1043 * We simply need to change them into set dimensions.
1045 * Then, for each block dimension, we compute the maximal value of the block id
1048 static __isl_give isl_multi_pw_aff
*extract_grid_size(
1049 struct ppcg_kernel
*kernel
, __isl_take isl_union_set
*domain
)
1054 domain
= isl_union_set_intersect(domain
,
1055 isl_union_set_copy(kernel
->block_filter
));
1056 grid
= isl_union_set_params(domain
);
1057 grid
= isl_set_from_params(grid
);
1058 grid
= isl_set_add_dims(grid
, isl_dim_set
, kernel
->n_grid
);
1059 for (i
= 0; i
< kernel
->n_grid
; ++i
) {
1063 id
= isl_id_list_get_id(kernel
->block_ids
, i
);
1064 pos
= isl_set_find_dim_by_id(grid
, isl_dim_param
, id
);
1067 grid
= isl_set_equate(grid
, isl_dim_param
, pos
, isl_dim_set
, i
);
1068 grid
= isl_set_project_out(grid
, isl_dim_param
, pos
, 1);
1071 return extract_size(grid
, isl_set_copy(kernel
->context
));
1074 /* Compute the size of a fixed bounding box around the origin and "set",
1075 * where "set" is assumed to contain only non-negative elements,
1076 * and store the results in "size".
1077 * In particular, compute the maximal value of "set" in each direction
1080 static void extract_fixed_size(__isl_take isl_set
*set
, int *size
)
1083 isl_local_space
*ls
;
1086 n
= isl_set_dim(set
, isl_dim_set
);
1087 ls
= isl_local_space_from_space(isl_set_get_space(set
));
1088 obj
= isl_aff_zero_on_domain(ls
);
1089 for (i
= 0; i
< n
; ++i
) {
1092 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 1);
1093 max
= isl_set_max_val(set
, obj
);
1094 size
[i
] = isl_val_get_num_si(max
) + 1;
1096 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 0);
1102 /* Compute the effective block size as a list of the sizes in each dimension
1103 * and store the sizes in kernel->block_dim.
1105 * The block size specified by the user or set by default
1106 * in read_block_sizes() and applied by the thread filter,
1107 * may be too large for the given code in the sense that
1108 * it may contain threads that don't need to execute anything.
1109 * We therefore update this block size in kernel->block_dim
1110 * to the smallest block size that ensures that all threads
1111 * that actually execute code are included in the block.
1113 * The possible values of the thread ids is obtained from
1114 * the domain elements "domain" and kernel->thread_filter.
1115 * The current implementation eliminates all parameters, ensuring
1116 * that the size is a fixed constant in each dimension.
1117 * In principle we could also compute parametric sizes.
1118 * We would have to make sure to project out all b%d and t%d parameters,
1121 static isl_stat
extract_block_size(struct ppcg_kernel
*kernel
,
1122 __isl_take isl_union_set
*domain
)
1128 domain
= isl_union_set_intersect(domain
,
1129 isl_union_set_copy(kernel
->thread_filter
));
1130 block
= isl_union_set_params(domain
);
1131 block
= isl_set_from_params(block
);
1132 block
= isl_set_add_dims(block
, isl_dim_set
, kernel
->n_block
);
1133 for (i
= 0; i
< kernel
->n_block
; ++i
) {
1138 return isl_stat_error
;
1140 id
= isl_id_list_get_id(kernel
->thread_ids
, i
);
1141 pos
= isl_set_find_dim_by_id(block
, isl_dim_param
, id
);
1144 isl_die(isl_set_get_ctx(block
), isl_error_internal
,
1145 "missing constraints on thread identifier",
1146 block
= isl_set_free(block
));
1147 block
= isl_set_equate(block
, isl_dim_param
, pos
,
1150 nparam
= isl_set_dim(block
, isl_dim_param
);
1151 block
= isl_set_project_out(block
, isl_dim_param
, 0, nparam
);
1154 return isl_stat_error
;
1156 extract_fixed_size(block
, kernel
->block_dim
);
1161 struct ppcg_kernel
*ppcg_kernel_free(struct ppcg_kernel
*kernel
)
1168 isl_id_list_free(kernel
->block_ids
);
1169 isl_id_list_free(kernel
->thread_ids
);
1170 isl_multi_pw_aff_free(kernel
->grid_size
);
1171 isl_set_free(kernel
->context
);
1172 isl_union_set_free(kernel
->core
);
1173 isl_union_set_free(kernel
->arrays
);
1174 isl_space_free(kernel
->space
);
1175 isl_ast_node_free(kernel
->tree
);
1176 isl_union_set_free(kernel
->block_filter
);
1177 isl_union_set_free(kernel
->thread_filter
);
1178 isl_union_pw_multi_aff_free(kernel
->shared_schedule
);
1179 isl_union_set_free(kernel
->sync_writes
);
1181 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1182 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1184 for (j
= 0; j
< array
->n_group
; ++j
)
1185 gpu_array_ref_group_free(array
->groups
[j
]);
1186 free(array
->groups
);
1188 isl_pw_aff_list_free(array
->bound
);
1190 free(kernel
->array
);
1192 for (i
= 0; i
< kernel
->n_var
; ++i
) {
1193 free(kernel
->var
[i
].name
);
1194 isl_vec_free(kernel
->var
[i
].size
);
1203 /* Wrapper around ppcg_kernel_free for use as a isl_id_set_free_user callback.
1205 static void ppcg_kernel_free_wrap(void *user
)
1207 struct ppcg_kernel
*kernel
= user
;
1209 ppcg_kernel_free(kernel
);
1212 static void create_kernel_var(isl_ctx
*ctx
, struct gpu_array_ref_group
*group
,
1213 struct ppcg_kernel_var
*var
)
1216 struct gpu_array_tile
*tile
;
1220 var
->array
= group
->array
;
1222 tile
= group
->private_tile
;
1223 var
->type
= ppcg_access_private
;
1225 tile
= group
->shared_tile
;
1226 var
->type
= ppcg_access_shared
;
1229 p
= isl_printer_to_str(ctx
);
1230 p
= gpu_array_ref_group_print_name(group
, p
);
1231 var
->name
= isl_printer_get_str(p
);
1232 isl_printer_free(p
);
1234 var
->size
= isl_vec_alloc(ctx
, group
->array
->n_index
);
1236 for (j
= 0; j
< group
->array
->n_index
; ++j
)
1237 var
->size
= isl_vec_set_element_val(var
->size
, j
,
1238 isl_val_copy(tile
->bound
[j
].size
));
1241 static int create_kernel_vars(struct ppcg_kernel
*kernel
)
1246 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1247 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1249 for (j
= 0; j
< array
->n_group
; ++j
) {
1250 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1251 if (group
->private_tile
|| group
->shared_tile
)
1257 kernel
->var
= isl_calloc_array(kernel
->ctx
, struct ppcg_kernel_var
, n
);
1262 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1263 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1265 for (j
= 0; j
< array
->n_group
; ++j
) {
1266 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1267 if (!group
->private_tile
&& !group
->shared_tile
)
1269 create_kernel_var(kernel
->ctx
, group
, &kernel
->var
[n
]);
1277 /* Replace "pa" by the zero function defined over the universe domain
1278 * in the space of "pa".
1280 static __isl_give isl_pw_aff
*set_universally_zero(__isl_take isl_pw_aff
*pa
)
1285 space
= isl_space_domain(isl_pw_aff_get_space(pa
));
1286 isl_pw_aff_free(pa
);
1287 zero
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1289 return isl_pw_aff_from_aff(zero
);
1292 /* The sizes of the arrays on the host that have been computed by
1293 * extract_array_info may depend on the parameters. Use the extra
1294 * constraints on the parameters that are valid at "host_domain"
1295 * to simplify these expressions and store the results in kernel->array.
1297 * We only need these localized bounds for arrays that are accessed
1298 * by the current kernel. If we have found at least one reference group
1299 * then the array is accessed by the kernel.
1301 * The resulting sizes may be functions that are nowhere defined
1302 * in case the access function cannot possibly access anything inside
1303 * the kernel for some reason. If so, they are replaced by the zero
1304 * function. Since the access function cannot actually access anything,
1305 * there is no harm in printing the array sizes as zero.
1307 static void localize_bounds(struct ppcg_kernel
*kernel
,
1308 __isl_keep isl_set
*host_domain
)
1313 context
= isl_set_copy(host_domain
);
1314 context
= isl_set_params(context
);
1316 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1317 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
1318 isl_pw_aff_list
*bound
;
1321 if (local
->n_group
== 0)
1324 n_index
= local
->array
->n_index
;
1325 bound
= isl_pw_aff_list_alloc(kernel
->ctx
, n_index
);
1327 for (j
= 0; j
< n_index
; ++j
) {
1331 pwaff
= isl_pw_aff_copy(local
->array
->bound
[j
]);
1332 pwaff
= isl_pw_aff_gist(pwaff
, isl_set_copy(context
));
1333 empty
= isl_pw_aff_is_empty(pwaff
);
1335 pwaff
= isl_pw_aff_free(pwaff
);
1337 pwaff
= set_universally_zero(pwaff
);
1338 bound
= isl_pw_aff_list_add(bound
, pwaff
);
1341 local
->n_index
= n_index
;
1342 local
->bound
= bound
;
1344 isl_set_free(context
);
1347 /* Create the array of gpu_local_array_info structures "array"
1348 * inside "kernel". The number of elements in this array is
1349 * the same as the number of arrays in "prog".
1350 * Initialize the "array" field of each local array to point
1351 * to the corresponding array in "prog".
1353 static struct ppcg_kernel
*ppcg_kernel_create_local_arrays(
1354 struct ppcg_kernel
*kernel
, struct gpu_prog
*prog
)
1359 ctx
= isl_set_get_ctx(prog
->context
);
1360 kernel
->array
= isl_calloc_array(ctx
,
1361 struct gpu_local_array_info
, prog
->n_array
);
1363 return ppcg_kernel_free(kernel
);
1364 kernel
->n_array
= prog
->n_array
;
1366 for (i
= 0; i
< prog
->n_array
; ++i
)
1367 kernel
->array
[i
].array
= &prog
->array
[i
];
1372 /* Does "kernel" need to be passed an argument corresponding to array "i"?
1374 * The argument is only needed if the kernel accesses this device memory.
1376 int ppcg_kernel_requires_array_argument(struct ppcg_kernel
*kernel
, int i
)
1378 return kernel
->array
[i
].global
;
1381 /* Find the element in gen->stmt that has the given "id".
1382 * Return NULL if no such gpu_stmt can be found.
1384 static struct gpu_stmt
*find_stmt(struct gpu_prog
*prog
, __isl_keep isl_id
*id
)
1388 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
1389 if (id
== prog
->stmts
[i
].id
)
1393 return i
< prog
->n_stmts
? &prog
->stmts
[i
] : NULL
;
1396 void ppcg_kernel_stmt_free(void *user
)
1399 struct ppcg_kernel_stmt
*stmt
= user
;
1404 switch (stmt
->type
) {
1405 case ppcg_kernel_copy
:
1406 isl_ast_expr_free(stmt
->u
.c
.index
);
1407 isl_ast_expr_free(stmt
->u
.c
.local_index
);
1409 case ppcg_kernel_domain
:
1410 isl_id_to_ast_expr_free(stmt
->u
.d
.ref2expr
);
1412 case ppcg_kernel_sync
:
1419 /* Return the gpu_stmt_access in the list "accesses" that corresponds
1422 static struct gpu_stmt_access
*find_access(struct gpu_stmt_access
*accesses
,
1423 __isl_keep isl_id
*ref_id
)
1425 struct gpu_stmt_access
*access
;
1427 for (access
= accesses
; access
; access
= access
->next
)
1428 if (access
->ref_id
== ref_id
)
1434 /* Return the index of the array called "name" in the list of arrays.
1436 static int find_array_index(struct ppcg_kernel
*kernel
, const char *name
)
1440 for (i
= 0; i
< kernel
->n_array
; ++i
)
1441 if (!strcmp(name
, kernel
->array
[i
].array
->name
))
1447 /* Internal data structure for the index and AST expression transformation
1448 * callbacks for pet_stmt_build_ast_exprs.
1450 * "kernel" is the kernel for which are computing AST expressions and
1451 * may be NULL if we are not inside a kernel.
1452 * "accesses" is the list of gpu_stmt_access in the statement.
1453 * "iterator_map" expresses the statement iterators in terms of
1454 * the AST loop iterators.
1455 * "sched2shared" expresses the outer shared_schedule_dim dimensions of
1456 * the kernel schedule in terms of the AST loop iterators and
1457 * may be NULL if we are not inside a kernel.
1459 * The following fields are set in transform_index and used in transform_expr.
1460 * "array" is the array that is being accessed.
1461 * "global" is set if the global array is accessed (rather than
1462 * shared/private memory).
1463 * "local_array" refers to information on the array specialized
1464 * to the current kernel.
1466 struct ppcg_transform_data
{
1467 struct ppcg_kernel
*kernel
;
1468 struct gpu_stmt_access
*accesses
;
1469 isl_pw_multi_aff
*iterator_map
;
1470 isl_pw_multi_aff
*sched2shared
;
1472 struct gpu_array_info
*array
;
1474 struct gpu_local_array_info
*local_array
;
1477 /* Return a pointer to the gpu_array_ref_group in "local"
1478 * that contains the reference "access".
1479 * Return NULL if no such group can be found.
1481 static struct gpu_array_ref_group
*find_ref_group(
1482 struct gpu_local_array_info
*local
, struct gpu_stmt_access
*access
)
1486 for (i
= 0; i
< local
->n_group
; ++i
) {
1487 struct gpu_array_ref_group
*group
= local
->groups
[i
];
1489 for (j
= 0; j
< group
->n_ref
; ++j
)
1490 if (group
->refs
[j
] == access
)
1497 /* Index transformation callback for pet_stmt_build_ast_exprs.
1499 * "index" expresses the array indices in terms of statement iterators
1501 * We first reformulate "index" in terms of the AST loop iterators.
1502 * Then we check if we are accessing the global array or
1503 * a shared/private copy. In particular, if we are not inside a kernel
1504 * then we must be accessing a global array.
1505 * In the former case, we simply return
1506 * the updated index. If "index" is an affine expression rather
1507 * than an array access, then we also return the updated index here.
1509 * If no reference groups have been computed for the array,
1510 * then we can only be accessing the global array.
1512 * Otherwise, we apply the tiling to the index.
1513 * This tiling is of the form
1517 * where D corresponds to the outer group->depth dimensions of
1518 * the kernel schedule.
1519 * The index is of the form
1523 * We update the tiling to refer to the AST loop iterators
1527 * and modify index to keep track of those iterators
1531 * Combining these two yields a tiled index expression in terms
1532 * of the AST loop iterators
1536 static __isl_give isl_multi_pw_aff
*transform_index(
1537 __isl_take isl_multi_pw_aff
*index
, __isl_keep isl_id
*ref_id
,
1540 struct ppcg_transform_data
*data
= user
;
1541 struct gpu_stmt_access
*access
;
1542 struct gpu_array_ref_group
*group
;
1543 struct gpu_array_tile
*tile
;
1544 isl_pw_multi_aff
*iterator_map
;
1549 isl_multi_pw_aff
*tiling
;
1550 isl_pw_multi_aff
*pma
;
1551 isl_multi_pw_aff
*mpa
;
1552 isl_pw_multi_aff
*sched2depth
;
1556 iterator_map
= isl_pw_multi_aff_copy(data
->iterator_map
);
1557 index
= isl_multi_pw_aff_pullback_pw_multi_aff(index
, iterator_map
);
1562 access
= find_access(data
->accesses
, ref_id
);
1565 if (!isl_map_has_tuple_name(access
->access
, isl_dim_out
))
1568 name
= get_outer_array_name(access
->access
);
1569 i
= find_array_index(data
->kernel
, name
);
1571 isl_die(isl_multi_pw_aff_get_ctx(index
), isl_error_internal
,
1572 "cannot find array",
1573 return isl_multi_pw_aff_free(index
));
1574 data
->local_array
= &data
->kernel
->array
[i
];
1575 data
->array
= data
->local_array
->array
;
1577 group
= find_ref_group(data
->local_array
, access
);
1583 tile
= group
->private_tile
;
1585 tile
= group
->shared_tile
;
1586 data
->global
= !tile
;
1590 space
= isl_space_range(isl_multi_pw_aff_get_space(index
));
1591 space
= isl_space_map_from_set(space
);
1592 pma
= isl_pw_multi_aff_identity(space
);
1593 sched2depth
= isl_pw_multi_aff_copy(data
->sched2shared
);
1594 dim
= isl_pw_multi_aff_dim(sched2depth
, isl_dim_out
);
1595 sched2depth
= isl_pw_multi_aff_drop_dims(sched2depth
, isl_dim_out
,
1596 group
->depth
, dim
- group
->depth
);
1597 pma
= isl_pw_multi_aff_product(sched2depth
, pma
);
1598 tiling
= isl_multi_pw_aff_from_multi_aff(
1599 isl_multi_aff_copy(tile
->tiling
));
1600 tiling
= isl_multi_pw_aff_pullback_pw_multi_aff(tiling
, pma
);
1602 space
= isl_space_domain(isl_multi_pw_aff_get_space(index
));
1603 space
= isl_space_map_from_set(space
);
1604 mpa
= isl_multi_pw_aff_identity(space
);
1605 index
= isl_multi_pw_aff_range_product(mpa
, index
);
1606 index
= isl_multi_pw_aff_pullback_multi_pw_aff(tiling
, index
);
1611 /* Dereference "expr" by adding an index [0].
1612 * The original "expr" is assumed not to have any indices.
1614 * If "expr" is a member access, then the dereferencing needs
1615 * to be applied to the structure argument of this member access.
1617 static __isl_give isl_ast_expr
*dereference(__isl_take isl_ast_expr
*expr
)
1620 isl_ast_expr
*arg0
, *res
;
1621 isl_ast_expr_list
*list
;
1623 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1625 return isl_ast_expr_free(expr
);
1626 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1627 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1630 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1631 arg
= dereference(arg
);
1632 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1633 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1637 isl_ast_expr_free(arg0
);
1639 ctx
= isl_ast_expr_get_ctx(expr
);
1640 res
= isl_ast_expr_from_val(isl_val_zero(ctx
));
1641 list
= isl_ast_expr_list_from_ast_expr(res
);
1642 res
= isl_ast_expr_get_op_arg(expr
, 0);
1643 res
= isl_ast_expr_access(res
, list
);
1644 isl_ast_expr_free(expr
);
1649 /* Linearize the index expression "expr" based on the array bounds
1652 * That is, transform expression
1654 * A[i_0][i_1]...[i_n]
1658 * A[(..((i_0 * b_1 + i_1) ... ) * b_n + i_n]
1660 * where b_0, b_1, ..., b_n are the bounds on the array.
1662 * If the base of "expr" is a member access, then the linearization needs
1663 * to be applied to the structure argument of this member access.
1665 * In the base case, if "expr" has no arguments (other than the name of
1666 * the array), then we are passing an entire array to a function.
1667 * In this case, there is nothing to linearize.
1668 * Note that at this point an expression with no arguments can
1669 * only be an entire array because the scalar case and
1670 * the case of single struct are handled by the caller.
1672 * If the number of specified index expressions in "expr"
1673 * is smaller than the dimension of the accessed array,
1674 * then the missing i_j also do not appear in the linearized expression.
1675 * Furthermore, since such an expression does not refer to a single
1676 * element while the default linearized expression would refer to
1677 * a single element, we return the expression
1679 * A + (..((i_0 * b_1 + i_1) ... ) * b_n]
1681 * instead. Note that because of the special case handling above,
1682 * we can assume here that here that there is at least one index expression.
1684 __isl_give isl_ast_expr
*gpu_local_array_info_linearize_index(
1685 struct gpu_local_array_info
*array
, __isl_take isl_ast_expr
*expr
)
1692 isl_ast_expr_list
*list
;
1693 isl_ast_build
*build
;
1695 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1696 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1697 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1700 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1701 arg
= gpu_local_array_info_linearize_index(array
, arg
);
1702 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1703 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1707 isl_ast_expr_free(arg0
);
1709 if (isl_ast_expr_get_op_n_arg(expr
) == 1)
1712 ctx
= isl_ast_expr_get_ctx(expr
);
1713 context
= isl_set_universe(isl_space_params_alloc(ctx
, 0));
1714 build
= isl_ast_build_from_context(context
);
1716 n
= isl_ast_expr_get_op_n_arg(expr
);
1717 res
= isl_ast_expr_get_op_arg(expr
, 1);
1718 for (i
= 1; i
< array
->n_index
; ++i
) {
1719 isl_pw_aff
*bound_i
;
1720 isl_ast_expr
*expr_i
;
1722 bound_i
= isl_pw_aff_list_get_pw_aff(array
->bound
, i
);
1723 expr_i
= isl_ast_build_expr_from_pw_aff(build
, bound_i
);
1724 res
= isl_ast_expr_mul(res
, expr_i
);
1728 expr_i
= isl_ast_expr_get_op_arg(expr
, i
+ 1);
1729 res
= isl_ast_expr_add(res
, expr_i
);
1732 isl_ast_build_free(build
);
1734 if (1 + array
->n_index
> n
) {
1735 res
= isl_ast_expr_add(isl_ast_expr_get_op_arg(expr
, 0), res
);
1737 list
= isl_ast_expr_list_from_ast_expr(res
);
1738 res
= isl_ast_expr_get_op_arg(expr
, 0);
1739 res
= isl_ast_expr_access(res
, list
);
1742 isl_ast_expr_free(expr
);
1747 /* AST expression transformation callback for pet_stmt_build_ast_exprs.
1749 * If the AST expression refers to an array that is not accessed
1750 * at all, then this means the value of the expression is not used,
1751 * so we might as well print zero (NULL pointer) instead.
1753 * If the AST expression refers to a global scalar that is not
1754 * a read-only scalar, then its address was passed to the kernel and
1755 * we need to dereference it.
1757 * If the AST expression refers to an access to a global array,
1758 * then we linearize the access exploiting the bounds in data->local_array.
1760 static __isl_give isl_ast_expr
*transform_expr(__isl_take isl_ast_expr
*expr
,
1761 __isl_keep isl_id
*id
, void *user
)
1763 struct ppcg_transform_data
*data
= user
;
1767 if (!data
->array
->accessed
) {
1770 ctx
= isl_ast_expr_get_ctx(expr
);
1771 isl_ast_expr_free(expr
);
1772 return isl_ast_expr_from_val(isl_val_zero(ctx
));
1774 if (gpu_array_is_read_only_scalar(data
->array
))
1778 if (data
->array
->n_index
== 0)
1779 return dereference(expr
);
1780 if (!data
->array
->linearize
)
1783 return gpu_local_array_info_linearize_index(data
->local_array
, expr
);
1786 /* This function is called for each instance of a user statement
1787 * in the kernel "kernel", identified by "gpu_stmt".
1788 * "kernel" may be NULL if we are not inside a kernel.
1790 * We attach a struct ppcg_kernel_stmt to the "node", containing
1791 * a computed AST expression for each access, through an annotation
1793 * These AST expressions are computed from iterator_map,
1794 * which expresses the domain
1795 * elements in terms of the generated loops, and sched2shared,
1796 * which expresses the outer shared_schedule_dim dimensions of
1797 * the kernel schedule computed by PPCG in terms of the generated loops.
1799 static __isl_give isl_ast_node
*create_domain_leaf(
1800 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1801 __isl_keep isl_ast_build
*build
, struct gpu_stmt
*gpu_stmt
)
1803 struct ppcg_transform_data data
;
1804 struct ppcg_kernel_stmt
*stmt
;
1807 isl_pw_multi_aff
*sched2shared
;
1809 isl_pw_multi_aff
*iterator_map
;
1810 isl_union_map
*schedule
;
1814 ctx
= isl_ast_node_get_ctx(node
);
1816 stmt
= isl_calloc_type(ctx
, struct ppcg_kernel_stmt
);
1818 return isl_ast_node_free(node
);
1820 schedule
= isl_ast_build_get_schedule(build
);
1821 map
= isl_map_reverse(isl_map_from_union_map(schedule
));
1822 iterator_map
= isl_pw_multi_aff_from_map(map
);
1824 sched2shared
= compute_sched_to_shared(kernel
,
1825 isl_pw_multi_aff_copy(iterator_map
));
1827 sched2shared
= NULL
;
1829 stmt
->type
= ppcg_kernel_domain
;
1830 stmt
->u
.d
.stmt
= gpu_stmt
;
1832 data
.kernel
= kernel
;
1833 data
.accesses
= stmt
->u
.d
.stmt
->accesses
;
1834 data
.iterator_map
= iterator_map
;
1835 data
.sched2shared
= sched2shared
;
1836 stmt
->u
.d
.ref2expr
= pet_stmt_build_ast_exprs(stmt
->u
.d
.stmt
->stmt
,
1837 build
, &transform_index
, &data
,
1838 &transform_expr
, &data
);
1840 isl_pw_multi_aff_free(iterator_map
);
1841 isl_pw_multi_aff_free(sched2shared
);
1843 id
= isl_id_alloc(ctx
, "user", stmt
);
1844 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1845 return isl_ast_node_set_annotation(node
, id
);
1848 /* This function is called for each statement node in the AST
1849 * for copying to or from shared/private memory.
1850 * Attach a pointer to a ppcg_kernel_stmt representing the copy
1851 * statement to the node.
1852 * The statement name is "read" or "write", depending on whether we are
1853 * reading from global memory or writing to global memory.
1855 * The schedule is of the form
1859 * where D corresponds to the outer group->depth dimensions of
1860 * the kernel schedule, A to the global array and L to the outer
1861 * generated AST schedule.
1862 * We compute the inverse and strip off the type, resulting in
1866 * We combine this mapping with on the one hand the projection
1870 * and on the other hand the group tiling
1878 * and store the corresponding expressions in stmt->index and stmt->local_index,
1879 * where stmt points to the ppcg_kernel_stmt that is attached to the node.
1881 static __isl_give isl_ast_node
*create_access_leaf(struct ppcg_kernel
*kernel
,
1882 struct gpu_array_ref_group
*group
, __isl_take isl_ast_node
*node
,
1883 __isl_keep isl_ast_build
*build
)
1885 struct ppcg_kernel_stmt
*stmt
;
1886 struct gpu_array_tile
*tile
;
1891 isl_pw_multi_aff
*pma
, *pma2
;
1894 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1896 return isl_ast_node_free(node
);
1898 access
= isl_map_from_union_map(isl_ast_build_get_schedule(build
));
1899 type
= isl_map_get_tuple_name(access
, isl_dim_in
);
1900 stmt
->u
.c
.read
= !strcmp(type
, "read");
1901 access
= isl_map_reverse(access
);
1902 pma
= isl_pw_multi_aff_from_map(access
);
1903 pma
= isl_pw_multi_aff_reset_tuple_id(pma
, isl_dim_out
);
1905 space
= isl_space_range(isl_pw_multi_aff_get_space(pma
));
1906 space
= isl_space_unwrap(space
);
1907 pma2
= isl_pw_multi_aff_range_map(space
);
1908 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
,
1909 isl_pw_multi_aff_copy(pma
));
1910 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1911 stmt
->u
.c
.index
= expr
;
1913 tile
= gpu_array_ref_group_tile(group
);
1914 pma2
= isl_pw_multi_aff_from_multi_aff(
1915 isl_multi_aff_copy(tile
->tiling
));
1916 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
, pma
);
1917 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1918 stmt
->u
.c
.local_index
= expr
;
1920 stmt
->u
.c
.array
= group
->array
;
1921 stmt
->u
.c
.local_array
= group
->local_array
;
1922 stmt
->type
= ppcg_kernel_copy
;
1924 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1925 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1926 return isl_ast_node_set_annotation(node
, id
);
1929 /* Create a synchronization ppcg_kernel_stmt and
1930 * attach it to the node "node" representing the synchronization.
1932 static __isl_give isl_ast_node
*create_sync_leaf(
1933 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1934 __isl_keep isl_ast_build
*build
)
1936 struct ppcg_kernel_stmt
*stmt
;
1939 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1941 return isl_ast_node_free(node
);
1943 stmt
->type
= ppcg_kernel_sync
;
1944 id
= isl_id_alloc(kernel
->ctx
, NULL
, stmt
);
1945 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1946 return isl_ast_node_set_annotation(node
, id
);
1949 /* Internal data structure for at_domain.
1951 * "prog" represents the entire scop.
1952 * "kernel" points to the kernel to which the current schedule node
1953 * belongs. It is set by before_mark and reset by after_mark.
1954 * It may be NULL if we are outside any kernel.
1956 struct ppcg_at_domain_data
{
1957 struct gpu_prog
*prog
;
1958 struct ppcg_kernel
*kernel
;
1961 /* This function is called for each instance of a user statement
1962 * in the kernel. This may be one of the original user statements
1963 * or a statement introduced by PPCG.
1965 * We first check if the statement id corresponds to a gpu statement,
1966 * which indicates the statement is an original user statement. Any statement
1967 * that is not an original user statement has been introduced by PPCG and
1968 * requires special handling.
1970 * If the user statement is one of the original user statements, then we call
1971 * create_domain_leaf. Otherwise, we check if it is a copy or synchronization
1972 * statement and call the appropriate functions. Statements that copy an array
1973 * to/from the device do not need any further treatment.
1975 static __isl_give isl_ast_node
*at_domain(__isl_take isl_ast_node
*node
,
1976 __isl_keep isl_ast_build
*build
, void *user
)
1978 struct ppcg_at_domain_data
*data
= user
;
1979 struct gpu_stmt
*gpu_stmt
;
1980 isl_ast_expr
*expr
, *arg
;
1986 expr
= isl_ast_node_user_get_expr(node
);
1987 arg
= isl_ast_expr_get_op_arg(expr
, 0);
1988 id
= isl_ast_expr_get_id(arg
);
1989 name
= isl_id_get_name(id
);
1990 p
= isl_id_get_user(id
);
1991 isl_ast_expr_free(expr
);
1992 isl_ast_expr_free(arg
);
1994 gpu_stmt
= find_stmt(data
->prog
, id
);
1995 is_sync
= gpu_tree_id_is_sync(id
, data
->kernel
);
1999 return create_domain_leaf(data
->kernel
, node
, build
, gpu_stmt
);
2001 if (!prefixcmp(name
, "to_device_") || !prefixcmp(name
, "from_device_"))
2004 return isl_ast_node_free(node
);
2005 if (!strcmp(name
, "read") || !strcmp(name
, "write")) {
2006 struct gpu_array_ref_group
*group
= p
;
2007 return create_access_leaf(data
->kernel
, group
, node
, build
);
2010 isl_die(data
->prog
->ctx
, isl_error_internal
,
2011 "unknown statement type",
2012 return isl_ast_node_free(node
));
2013 return create_sync_leaf(data
->kernel
, node
, build
);
2016 /* Given a set of wrapped references "ref", return the corresponding
2017 * access relations based on the tagged access relations "tagged".
2019 * The elements of "ref" are of the form
2023 * with D an iteration domains and R a reference.
2024 * The elements of "tagged" are of the form
2030 * Extend "tagged" to include the iteration domain in the range, i.e.,
2032 * [D -> R] -> [D -> A]
2034 * apply the result to "ref" and then unwrap the resulting set
2035 * to obtain relations of the form
2039 static __isl_give isl_union_map
*wrapped_reference_to_access(
2040 __isl_take isl_union_set
*ref
, __isl_take isl_union_map
*tagged
)
2042 isl_union_map
*tag2access
;
2044 tag2access
= isl_union_map_copy(tagged
);
2045 tag2access
= isl_union_map_universe(tag2access
);
2046 tag2access
= isl_union_set_unwrap(isl_union_map_domain(tag2access
));
2047 tag2access
= isl_union_map_domain_map(tag2access
);
2048 tag2access
= isl_union_map_range_product(tag2access
, tagged
);
2050 ref
= isl_union_set_coalesce(ref
);
2051 ref
= isl_union_set_apply(ref
, tag2access
);
2053 return isl_union_set_unwrap(ref
);
2056 /* Given an access relation "access" from one or more array reference groups,
2057 * remove those reads if ("read" is 1) or writes (if "read" is 0)
2058 * that are only needed to communicate data within
2059 * the same iteration of "sched".
2060 * "tagged" contains all tagged access relations to all
2061 * the array reference groups accessed by "access" from statement
2062 * instances scheduled by "sched".
2064 * If the access is a read then it is either an element of
2066 * live_in union (range flow)
2068 * where live_in and flow may be overapproximations, or
2069 * it reads an uninitialized value (that is not live-in because
2070 * there is an intermediate kill) or it reads a value that was
2071 * written within the same (compound) statement instance.
2072 * If the access is a write then it is either an element of
2074 * live_out union (domain flow)
2076 * or it writes a value that is never read (and is not live-out
2077 * because of an intermediate kill) or only
2078 * within the same (compound) statement instance.
2079 * In both cases, the access relation is also a subset of
2080 * the group access relation.
2082 * The cases where an uninitialized value is read or a value is written
2083 * that is never read or where the dataflow occurs within a statement
2084 * instance are also considered local and may also be removed.
2086 * Essentially, we compute the intersection of "access" with either
2088 * live_in union (range non-local-flow)
2092 * live_out union (domain non-local-flow)
2094 * We first construct a relation "local"
2096 * [[D -> R] -> [D' -> R']]
2098 * of pairs of domain iterations accessing the reference group
2099 * and references in the group that are coscheduled by "sched".
2101 * If this relation does not intersect the dataflow dependences,
2102 * then there is nothing we can possibly remove, unless the dataflow
2103 * dependences themselves only relate a subset of the accesses.
2104 * In particular, the accesses may not be involved in any dataflow
2105 * dependences, either because they are uninitialized reads/dead writes
2106 * or because the dataflow occurs inside a statement instance.
2108 * Since the computation below may break up the access relation
2109 * into smaller pieces, we only perform the intersection with
2110 * the non-local dependent accesses if the local pairs
2111 * intersect the dataflow dependences. Otherwise, we intersect
2112 * with the universe of the non-local dependent accesses.
2113 * This should at least remove accesses from statements that
2114 * do not participate in any dependences.
2116 * In particular, we remove the "local" dataflow dependences from
2117 * the set of all dataflow dependences, or at least those
2118 * that may contribute to a domain/range that intersects
2119 * the domain of "access".
2120 * Note that if the potential dataflow dependences are an overapproximation
2121 * of the actual dataflow dependences, then the result remains an
2122 * overapproximation of the non-local dataflow dependences.
2123 * Copying to/from global memory is only needed for the references
2124 * in the domain/range of the result or for accesses that are live out/in
2125 * for the entire scop.
2127 * We therefore map the domain/range of the "external" relation
2128 * to the corresponding access relation and take the union with
2129 * the live out/in relation.
2131 static __isl_give isl_union_map
*remove_local_accesses(
2132 struct gpu_prog
*prog
, __isl_take isl_union_map
*tagged
,
2133 __isl_take isl_union_map
*access
, __isl_take isl_union_map
*sched
,
2137 isl_union_pw_multi_aff
*tagger
;
2138 isl_union_set
*domain
, *access_domain
;
2139 isl_union_map
*local
, *external
, *universe
;
2140 isl_union_set
*tag_set
;
2142 if (isl_union_map_is_empty(access
)) {
2143 isl_union_map_free(sched
);
2144 isl_union_map_free(tagged
);
2148 tagger
= isl_union_pw_multi_aff_copy(prog
->scop
->tagger
);
2149 domain
= isl_union_map_domain(isl_union_map_copy(tagged
));
2150 tagger
= isl_union_pw_multi_aff_intersect_domain(tagger
,
2151 isl_union_set_copy(domain
));
2152 sched
= isl_union_map_preimage_domain_union_pw_multi_aff(sched
, tagger
);
2154 local
= isl_union_map_apply_range(sched
,
2155 isl_union_map_reverse(isl_union_map_copy(sched
)));
2156 local
= isl_union_map_intersect(local
,
2157 isl_union_map_copy(prog
->scop
->tagged_dep_flow
));
2159 empty
= isl_union_map_is_empty(local
);
2161 external
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
2162 universe
= isl_union_map_universe(isl_union_map_copy(access
));
2163 access_domain
= isl_union_map_domain(universe
);
2164 domain
= isl_union_set_universe(domain
);
2165 universe
= isl_union_set_unwrap(domain
);
2166 universe
= isl_union_map_intersect_domain(universe
, access_domain
);
2167 domain
= isl_union_map_wrap(universe
);
2169 external
= isl_union_map_intersect_range(external
, domain
);
2171 external
= isl_union_map_intersect_domain(external
, domain
);
2172 external
= isl_union_map_intersect_params(external
,
2173 isl_set_copy(prog
->scop
->context
));
2174 external
= isl_union_map_subtract(external
, local
);
2177 tag_set
= isl_union_map_range(external
);
2178 external
= wrapped_reference_to_access(tag_set
, tagged
);
2179 external
= isl_union_map_union(external
,
2180 isl_union_map_copy(prog
->scop
->live_in
));
2182 tag_set
= isl_union_map_domain(external
);
2183 external
= wrapped_reference_to_access(tag_set
, tagged
);
2184 external
= isl_union_map_union(external
,
2185 isl_union_map_copy(prog
->scop
->live_out
));
2189 external
= isl_union_map_free(external
);
2191 external
= isl_union_map_universe(external
);
2193 access
= isl_union_map_intersect(access
, external
);
2198 /* Given an access relation "access" from "group", remove those reads
2199 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
2200 * communicate data within the same iteration of the schedule at the
2201 * position where the copying of the group is inserted.
2202 * "node" points to this position, i.e., the depth at "node"
2203 * is equal to group->depth.
2205 * We extract a schedule that picks out the iterations of the outer
2206 * group->depth dimensions and call remove_local_accesses.
2208 static __isl_give isl_union_map
*remove_local_accesses_group(
2209 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2210 __isl_take isl_union_map
*access
, __isl_keep isl_schedule_node
*node
,
2213 isl_union_map
*sched
, *tagged
;
2215 if (isl_union_map_is_empty(access
))
2218 tagged
= group_tagged_access_relation(group
);
2219 sched
= isl_schedule_node_get_prefix_schedule_relation(node
);
2221 return remove_local_accesses(kernel
->prog
, tagged
, access
, sched
, read
);
2224 /* This function is called before the AST generator starts traversing
2225 * the schedule subtree of a node with mark "mark".
2227 * If the mark is called "kernel", store the kernel pointer in data->kernel
2228 * for use in at_domain.
2230 static int before_mark(__isl_keep isl_id
*mark
,
2231 __isl_keep isl_ast_build
*build
, void *user
)
2233 struct ppcg_at_domain_data
*data
= user
;
2237 if (!strcmp(isl_id_get_name(mark
), "kernel"))
2238 data
->kernel
= isl_id_get_user(mark
);
2242 /* This function is called after the AST generator has finished traversing
2243 * the schedule subtree of a mark node. "node" points to the corresponding
2246 * If the mark is called "kernel", then replace "node" by a user node
2247 * that "calls" the kernel, representing the launch of the kernel.
2248 * The original "node" is stored inside the kernel object so that
2249 * it can be used to print the device code.
2250 * Note that this assumes that a kernel is only launched once.
2251 * Also clear data->kernel.
2253 static __isl_give isl_ast_node
*after_mark(__isl_take isl_ast_node
*node
,
2254 __isl_keep isl_ast_build
*build
, void *user
)
2259 isl_ast_expr_list
*list
;
2260 struct ppcg_kernel
*kernel
;
2261 struct ppcg_at_domain_data
*data
= user
;
2263 ctx
= isl_ast_node_get_ctx(node
);
2264 id
= isl_ast_node_mark_get_id(node
);
2266 return isl_ast_node_free(node
);
2267 if (strcmp(isl_id_get_name(id
), "kernel") || !data
->kernel
) {
2271 kernel
= data
->kernel
;
2272 data
->kernel
= NULL
;
2273 kernel
->space
= isl_ast_build_get_schedule_space(build
);
2274 kernel
->tree
= isl_ast_node_mark_get_node(node
);
2275 isl_ast_node_free(node
);
2277 expr
= isl_ast_expr_from_id(isl_id_copy(id
));
2278 list
= isl_ast_expr_list_alloc(ctx
, 0);
2279 expr
= isl_ast_expr_call(expr
, list
);
2280 node
= isl_ast_node_alloc_user(expr
);
2281 node
= isl_ast_node_set_annotation(node
, id
);
2286 static isl_bool
update_depth(__isl_keep isl_schedule_node
*node
, void *user
)
2291 if (isl_schedule_node_get_type(node
) != isl_schedule_node_leaf
)
2292 return isl_bool_true
;
2293 node_depth
= isl_schedule_node_get_schedule_depth(node
);
2294 if (node_depth
> *depth
)
2295 *depth
= node_depth
;
2297 return isl_bool_false
;
2300 /* Use isl to generate code for both the host and the device
2302 * The device code is marked by "kernel" mark nodes in the schedule tree,
2303 * containing a pointer to a ppcg_kernel object.
2304 * The returned AST only contains the AST for the host code.
2305 * The ASTs for the device code are embedded in ppcg_kernel objects
2306 * attached to the leaf nodes that call "kernel".
2308 static __isl_give isl_ast_node
*generate_code(struct gpu_gen
*gen
,
2309 __isl_take isl_schedule
*schedule
)
2311 struct ppcg_at_domain_data data
;
2312 isl_ast_build
*build
;
2314 isl_id_list
*iterators
;
2317 data
.prog
= gen
->prog
;
2321 if (isl_schedule_foreach_schedule_node_top_down(schedule
, &update_depth
,
2324 build
= isl_ast_build_alloc(gen
->prog
->ctx
);
2325 iterators
= ppcg_scop_generate_names(gen
->prog
->scop
, depth
, "c");
2326 build
= isl_ast_build_set_iterators(build
, iterators
);
2327 build
= isl_ast_build_set_at_each_domain(build
, &at_domain
, &data
);
2328 build
= isl_ast_build_set_before_each_mark(build
, &before_mark
, &data
);
2329 build
= isl_ast_build_set_after_each_mark(build
, &after_mark
, &data
);
2330 if (gen
->prog
->scop
->options
->debug
->dump_final_schedule
)
2331 isl_schedule_dump(schedule
);
2332 tree
= isl_ast_build_node_from_schedule(build
, schedule
);
2333 isl_ast_build_free(build
);
2338 __isl_give isl_union_map
*extract_sizes_from_str(isl_ctx
*ctx
, const char *str
)
2342 return isl_union_map_read_from_str(ctx
, str
);
2345 /* Can "node" be tiled and then mapped to block and thread identifiers?
2346 * That is, is it permutable with at least one coincident dimension?
2348 static int is_permutable(__isl_keep isl_schedule_node
*node
)
2353 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
2355 if (!isl_schedule_node_band_get_permutable(node
))
2357 if (isl_schedule_node_band_n_member(node
) < 1)
2359 if (!isl_schedule_node_band_member_get_coincident(node
, 0))
2365 /* A isl_schedule_foreach_schedule_node_top_down callback
2366 * for setting *any_permutable and aborting the search
2367 * if "node" is a permutable band with coincident dimensions.
2368 * Otherwise, continue searching.
2370 static isl_bool
set_permutable(__isl_keep isl_schedule_node
*node
, void *user
)
2372 int *any_permutable
= user
;
2375 permutable
= is_permutable(node
);
2377 return isl_bool_error
;
2379 return isl_bool_true
;
2381 *any_permutable
= 1;
2383 return isl_bool_error
;
2386 /* Does "schedule" contain any permutable band with at least one coincident
2389 static int has_any_permutable_node(__isl_keep isl_schedule
*schedule
)
2391 int any_permutable
= 0;
2393 if (isl_schedule_foreach_schedule_node_top_down(schedule
,
2394 &set_permutable
, &any_permutable
) < 0 &&
2398 return any_permutable
;
2401 /* Is "node" a leaf or can it be tiled and then mapped to
2402 * block and thread identifiers?
2404 static int is_leaf_or_tilable(__isl_keep isl_schedule_node
*node
)
2406 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
2408 return is_permutable(node
);
2411 /* Is "node" the outermost node in its branch that can be tiled
2412 * and then mapped to block and thread identifiers?
2413 * If there are no such nodes in the branch and if "node" is a leaf,
2414 * then it is accepted too.
2416 static int is_outer_tilable(__isl_keep isl_schedule_node
*node
)
2419 isl_schedule_node
*ancestor
;
2421 tilable
= is_leaf_or_tilable(node
);
2428 ancestor
= isl_schedule_node_copy(node
);
2429 while (isl_schedule_node_has_parent(ancestor
)) {
2430 ancestor
= isl_schedule_node_parent(ancestor
);
2432 tilable
= is_permutable(ancestor
);
2433 if (tilable
< 0 || tilable
)
2437 isl_schedule_node_free(ancestor
);
2438 return tilable
< 0 ? -1 : !tilable
;
2441 /* Collect the references to all writes in "group".
2442 * Each reference is represented by a universe set in a space
2446 * with S[i,j] the statement instance space and R[] the array reference.
2448 static __isl_give isl_union_set
*group_tagged_writes(
2449 struct gpu_array_ref_group
*group
)
2453 isl_union_set
*writes
;
2455 space
= isl_map_get_space(group
->access
);
2456 writes
= isl_union_set_empty(space
);
2457 for (i
= 0; i
< group
->n_ref
; ++i
) {
2461 if (!group
->refs
[i
]->write
)
2464 space
= isl_map_get_space(group
->refs
[i
]->tagged_access
);
2465 space
= isl_space_domain(space
);
2466 writes_i
= isl_set_universe(space
);
2467 writes
= isl_union_set_add_set(writes
, writes_i
);
2473 /* Is there any write access in "group" that requires synchronization
2474 * on a write to global memory?
2475 * We currently take into account all writes that would require
2476 * synchronization at the thread level depth, but if the copying
2477 * for this group is performed at an outer level, then we do not
2478 * actually need to take into account dependences at intermediate levels.
2480 static int any_sync_writes_in_group(struct ppcg_kernel
*kernel
,
2481 struct gpu_array_ref_group
*group
)
2483 isl_union_set
*writes
;
2484 int empty
, disjoint
;
2486 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2492 writes
= group_tagged_writes(group
);
2493 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2494 isl_union_set_free(writes
);
2496 return disjoint
< 0 ? -1 : !disjoint
;
2499 /* Collect the references to all writes in "kernel" that write directly
2500 * to global or shared memory, i.e., that are not mapped to private memory.
2501 * Each reference is represented by a universe set in a space
2505 * with S[i,j] the statement instance space and R[] the array reference.
2507 static __isl_give isl_union_set
*collect_non_private_tagged_writes(
2508 struct ppcg_kernel
*kernel
)
2510 isl_union_set
*writes
;
2513 writes
= isl_union_set_empty(isl_union_set_get_space(kernel
->arrays
));
2515 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2516 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2518 for (j
= 0; j
< array
->n_group
; ++j
) {
2519 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2520 isl_union_set
*writes_ij
;
2524 if (group
->private_tile
)
2526 writes_ij
= group_tagged_writes(group
);
2527 writes
= isl_union_set_union(writes
, writes_ij
);
2534 /* Are there any direct writes to global memory that require
2537 static int any_global_or_shared_sync_writes(struct ppcg_kernel
*kernel
)
2539 isl_union_set
*writes
;
2540 int empty
, disjoint
;
2542 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2548 writes
= collect_non_private_tagged_writes(kernel
);
2549 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2550 isl_union_set_free(writes
);
2552 return disjoint
< 0 ? -1 : !disjoint
;
2555 /* Construct an isl_multi_val for use as tile sizes for tiling "node"
2556 * from the elements in "tile_size".
2558 static __isl_give isl_multi_val
*construct_band_tiles_sizes(
2559 __isl_keep isl_schedule_node
*node
, int *tile_size
)
2569 ctx
= isl_schedule_node_get_ctx(node
);
2570 space
= isl_schedule_node_band_get_space(node
);
2571 n
= isl_schedule_node_band_n_member(node
);
2572 mv
= isl_multi_val_zero(space
);
2573 for (i
= 0; i
< n
; ++i
) {
2576 v
= isl_val_int_from_si(ctx
, tile_size
[i
]);
2577 mv
= isl_multi_val_set_val(mv
, i
, v
);
2583 /* Replace the partial schedule S of the band node "node" by
2591 * if scale_tile_loops is set, with f the integers in "factor".
2592 * The list that "factor" points to is assumed to contain at least
2593 * as many elements as the number of members in the band.
2595 static __isl_give isl_schedule_node
*snap_band_to_sizes(
2596 __isl_take isl_schedule_node
*node
, int *factor
,
2597 struct ppcg_options
*options
)
2601 mv
= construct_band_tiles_sizes(node
, factor
);
2602 node
= isl_schedule_node_band_scale_down(node
, isl_multi_val_copy(mv
));
2603 if (options
->scale_tile_loops
)
2604 node
= isl_schedule_node_band_scale(node
,
2605 isl_multi_val_copy(mv
));
2606 isl_multi_val_free(mv
);
2611 /* Tile "band" with tile size specified by "sizes".
2613 * Since the tile loops will be mapped to block ids, we forcibly
2614 * turn off tile loop scaling. We may want to enable tile loop scaling
2615 * at some later point, but then we would have to support the detection
2616 * of strides during the mapping to block ids.
2617 * Similarly, since the point loops will be mapped to thread ids,
2618 * we forcibly shift the point loops so that they start at zero.
2620 static __isl_give isl_schedule_node
*tile_band(
2621 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2623 isl_ctx
*ctx
= isl_schedule_node_get_ctx(node
);
2627 scale_tile
= isl_options_get_tile_scale_tile_loops(ctx
);
2628 isl_options_set_tile_scale_tile_loops(ctx
, 0);
2629 shift_point
= isl_options_get_tile_shift_point_loops(ctx
);
2630 isl_options_set_tile_shift_point_loops(ctx
, 1);
2632 node
= isl_schedule_node_band_tile(node
, sizes
);
2634 isl_options_set_tile_scale_tile_loops(ctx
, scale_tile
);
2635 isl_options_set_tile_shift_point_loops(ctx
, shift_point
);
2640 /* Extract the set of parameter values and outer schedule dimensions
2641 * for which any statement instance
2642 * in the kernel inserted at "node" needs to be executed.
2643 * Intersect the set of parameter values derived from the host schedule
2644 * relation with the context of "prog".
2646 static __isl_give isl_set
*extract_context(__isl_keep isl_schedule_node
*node
,
2647 struct gpu_prog
*prog
)
2649 isl_union_map
*schedule
;
2650 isl_union_set
*schedule_domain
;
2654 schedule
= isl_schedule_node_get_prefix_schedule_relation(node
);
2655 schedule_domain
= isl_union_map_range(schedule
);
2656 empty
= isl_union_set_is_empty(schedule_domain
);
2658 isl_union_set_free(schedule_domain
);
2665 space
= isl_union_set_get_space(schedule_domain
);
2666 isl_union_set_free(schedule_domain
);
2667 space
= isl_space_set_from_params(space
);
2668 depth
= isl_schedule_node_get_schedule_depth(node
);
2669 space
= isl_space_add_dims(space
, isl_dim_set
, depth
);
2670 context
= isl_set_empty(space
);
2672 context
= isl_set_from_union_set(schedule_domain
);
2674 context
= isl_set_intersect_params(context
,
2675 isl_set_copy(prog
->context
));
2680 /* Return the set of outer array elements accessed by
2681 * by the statement instance in "domain" in "prog".
2683 static __isl_give isl_union_set
*accessed_by_domain(
2684 __isl_take isl_union_set
*domain
, struct gpu_prog
*prog
)
2686 isl_union_map
*access
;
2687 isl_union_set
*arrays
;
2689 access
= isl_union_map_union(isl_union_map_copy(prog
->read
),
2690 isl_union_map_copy(prog
->may_write
));
2691 access
= isl_union_map_intersect_domain(access
, domain
);
2692 arrays
= isl_union_map_range(access
);
2693 arrays
= isl_union_set_apply(arrays
,
2694 isl_union_map_copy(prog
->to_outer
));
2699 /* Return the number of outer band members of the band node "node"
2700 * that are marked coincident.
2702 static int n_outer_coincidence(__isl_keep isl_schedule_node
*node
)
2706 n
= isl_schedule_node_band_n_member(node
);
2708 for (i
= 0; i
< n
; ++i
)
2709 if (!isl_schedule_node_band_member_get_coincident(node
, i
))
2715 /* If the band node "node" has more than "n" members, then split off
2716 * the first "n" of them.
2718 static __isl_give isl_schedule_node
*split_band(
2719 __isl_take isl_schedule_node
*node
, int n
)
2723 dim
= isl_schedule_node_band_n_member(node
);
2725 node
= isl_schedule_node_band_split(node
, n
);
2730 /* Scale a band node that may have been split by split_band.
2731 * "sizes" are the scaling factors for the original node.
2732 * "node" either points to the original band node, or the outer
2733 * of the two pieces after splitting.
2735 * If the number of elements in "node" is smaller than the number of
2736 * elements in "sizes", then some splitting has occurred and we split
2737 * "sizes" in the same way.
2739 static __isl_give isl_schedule_node
*scale_band(
2740 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2744 n
= isl_multi_val_dim(sizes
, isl_dim_set
);
2745 dim
= isl_schedule_node_band_n_member(node
);
2747 isl_multi_val
*sizes2
;
2749 sizes2
= isl_multi_val_copy(sizes
);
2750 sizes
= isl_multi_val_drop_dims(sizes
,
2751 isl_dim_set
, dim
, n
- dim
);
2752 sizes2
= isl_multi_val_drop_dims(sizes2
, isl_dim_set
, 0, dim
);
2753 node
= isl_schedule_node_child(node
, 0);
2754 node
= isl_schedule_node_band_scale(node
, sizes2
);
2755 node
= isl_schedule_node_parent(node
);
2758 return isl_schedule_node_band_scale(node
, sizes
);
2761 /* Return an isl_multi_aff, with as elements the parameters in "space"
2762 * that have the names specified by the elements in "names".
2763 * If (some of) these parameters do not already appear in "space",
2764 * then they are added first.
2766 static __isl_give isl_multi_aff
*parameter_vector(__isl_take isl_space
*space
,
2767 __isl_keep isl_id_list
*names
)
2770 isl_local_space
*ls
;
2774 space
= isl_space_free(space
);
2776 n
= isl_id_list_n_id(names
);
2777 for (i
= 0; i
< n
; ++i
) {
2781 id
= isl_id_list_get_id(names
, i
);
2782 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2787 pos
= isl_space_dim(space
, isl_dim_param
);
2788 space
= isl_space_add_dims(space
, isl_dim_param
, 1);
2789 space
= isl_space_set_dim_id(space
, isl_dim_param
, pos
, id
);
2791 ma
= isl_multi_aff_zero(isl_space_copy(space
));
2792 ls
= isl_local_space_from_space(isl_space_domain(space
));
2793 for (i
= 0; i
< n
; ++i
) {
2798 id
= isl_id_list_get_id(names
, i
);
2799 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2801 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
2802 isl_dim_param
, pos
);
2803 ma
= isl_multi_aff_set_aff(ma
, i
, aff
);
2805 isl_local_space_free(ls
);
2810 /* Return constraints on the domain elements that equate a sequence of
2811 * parameters called "names", to the partial schedule
2812 * of "node" modulo the integers in "size".
2813 * The number of elements in the array "size" should be equal
2814 * to the number of elements in "names".
2815 * The number of members of the band node "node" should be smaller
2816 * than or equal to this number. If it is smaller, then the first
2817 * elements of "names" are equated to zero.
2819 static __isl_give isl_union_set
*set_schedule_modulo(
2820 __isl_keep isl_schedule_node
*node
, __isl_keep isl_id_list
*names
,
2826 isl_multi_union_pw_aff
*mupa
, *mupa2
;
2828 isl_union_set
*domain
;
2832 n
= isl_id_list_n_id(names
);
2834 return isl_schedule_node_get_universe_domain(node
);
2835 n_zero
= n
- isl_schedule_node_band_n_member(node
);
2837 mupa
= isl_schedule_node_band_get_partial_schedule(node
);
2838 mv
= construct_band_tiles_sizes(node
, size
+ n_zero
);
2839 mupa
= isl_multi_union_pw_aff_mod_multi_val(mupa
, mv
);
2841 space
= isl_multi_union_pw_aff_get_space(mupa
);
2842 space
= isl_space_params(space
);
2843 space
= isl_space_set_from_params(space
);
2844 space
= isl_space_add_dims(space
, isl_dim_set
, n_zero
);
2845 ma
= isl_multi_aff_zero(space
);
2847 domain
= isl_schedule_node_get_universe_domain(node
);
2848 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(
2849 isl_union_set_copy(domain
), ma
);
2850 mupa
= isl_multi_union_pw_aff_range_product(mupa2
, mupa
);
2852 space
= isl_multi_union_pw_aff_get_space(mupa
);
2853 ma
= parameter_vector(space
, names
);
2855 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(domain
, ma
);
2856 mupa
= isl_multi_union_pw_aff_sub(mupa
, mupa2
);
2858 return isl_multi_union_pw_aff_zero_union_set(mupa
);
2861 /* Insert a context node at "node" introducing the block and thread
2862 * identifiers along with their bounds, which are stored in kernel->grid_size
2863 * and kernel->block_dim.
2864 * Note that the bounds on the block identifiers may implicitly impose
2865 * constraints on the parameters. A guard needs to be inserted
2866 * in the schedule tree to ensure that those bounds hold at "node".
2867 * This guard is inserted in insert_guard.
2869 static __isl_give isl_schedule_node
*insert_context(struct ppcg_kernel
*kernel
,
2870 __isl_take isl_schedule_node
*node
)
2874 context
= isl_set_universe(isl_set_get_space(kernel
->context
));
2876 context
= add_bounded_parameters_dynamic(context
,
2877 kernel
->grid_size
, kernel
->block_ids
);
2878 context
= add_bounded_parameters(context
,
2879 kernel
->block_dim
, kernel
->thread_ids
);
2881 node
= isl_schedule_node_insert_context(node
, context
);
2886 /* Insert a guard that eliminates kernel launches where the kernel
2887 * obviously does not have any work to do.
2889 * In particular, eliminate kernel launches where there are obviously
2891 * Use the same block size constraints that are used to create the context
2892 * to ensure that all constraints implicit in the constructed context
2893 * are imposed by the guard.
2895 * Additionally, add other constraints that are valid
2896 * for each executed instance ("context"), as long as this does not result
2899 static __isl_give isl_schedule_node
*insert_guard(
2900 __isl_take isl_schedule_node
*node
, __isl_keep isl_set
*context
,
2901 __isl_keep isl_multi_pw_aff
*size
, struct ppcg_scop
*scop
)
2907 guard
= isl_set_copy(context
);
2908 guard
= isl_set_compute_divs(guard
);
2909 guard
= isl_set_from_basic_set(isl_set_simple_hull(guard
));
2911 nparam
= isl_set_dim(guard
, isl_dim_param
);
2912 n
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
2913 ids
= ppcg_scop_generate_names(scop
, n
, "__ppcg_tmp");
2914 guard
= add_bounded_parameters_dynamic(guard
, size
, ids
);
2915 isl_id_list_free(ids
);
2916 guard
= isl_set_project_out(guard
, isl_dim_param
, nparam
, n
);
2918 node
= isl_schedule_node_insert_guard(node
, guard
);
2923 /* Does any array reference group mapping require the band that is mapped
2924 * to threads to be unrolled?
2926 static int kernel_requires_unroll(struct ppcg_kernel
*kernel
)
2930 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2931 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2933 for (j
= 0; j
< array
->n_group
; ++j
) {
2934 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2935 if (gpu_array_ref_group_requires_unroll(group
))
2943 /* Mark the given band node "node" for unrolling by the AST generator and
2944 * then sink it to the leaves of the schedule tree.
2945 * All dimensions of "node" are assumed to be coincident, such that this
2946 * sinking is a valid operation.
2948 static __isl_give isl_schedule_node
*unroll(__isl_take isl_schedule_node
*node
)
2952 n
= isl_schedule_node_band_n_member(node
);
2953 for (i
= 0; i
< n
; ++i
)
2954 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
2955 isl_ast_loop_unroll
);
2957 node
= isl_schedule_node_band_sink(node
);
2962 /* Insert a synchronization node in the schedule tree of "node"
2963 * after the core computation of "kernel" at the level of the band
2964 * that is mapped to threads, except if that level is equal to
2965 * that of the band that is mapped to blocks or if there are no writes
2966 * to global or shared memory in the core computation that require
2968 * If there are any writes to shared memory and the shared memory
2969 * copying is performed at the same level, then synchronization
2970 * is needed between the core and the copying anyway, so we might
2971 * as well add it here. If the copying is performed at a higher
2972 * level, then different iterations of intermediate schedule dimensions
2973 * may have a different mapping from between shared memory elements and
2974 * threads, such that synchronization is required after the core.
2975 * "node" is assumed to point to the kernel node.
2977 static __isl_give isl_schedule_node
*add_sync(struct ppcg_kernel
*kernel
,
2978 __isl_take isl_schedule_node
*node
)
2983 need_sync
= any_global_or_shared_sync_writes(kernel
);
2985 return isl_schedule_node_free(node
);
2989 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
2991 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
2992 if (kernel_depth
== isl_schedule_node_get_schedule_depth(node
))
2993 return gpu_tree_move_up_to_kernel(node
);
2995 node
= gpu_tree_ensure_following_sync(node
, kernel
);
2997 node
= gpu_tree_move_up_to_kernel(node
);
3002 /* Return a read ("read" is 1) or write access relation for "group"
3003 * with those accesses removed that are only needed to communicate data
3004 * within the subtree of the schedule rooted at "node".
3005 * Furthermore, include the prefix schedule at "node".
3006 * That is, return a relation of the form
3010 * with D the outer schedule dimensions at "node".
3012 static __isl_give isl_union_map
*anchored_non_local_accesses(
3013 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3014 __isl_take isl_schedule_node
*node
, int read
)
3016 isl_union_map
*access
;
3017 isl_union_map
*prefix
;
3019 access
= gpu_array_ref_group_access_relation(group
, read
, !read
);
3020 access
= remove_local_accesses_group(kernel
, group
, access
, node
, read
);
3021 prefix
= isl_schedule_node_get_prefix_schedule_relation(node
);
3022 access
= isl_union_map_range_product(prefix
, access
);
3027 /* Given an array reference group "group", create a mapping
3029 * read[D -> A] -> [D -> A]
3031 * if "read" is set or
3033 * write[D -> A] -> [D -> A]
3035 * if "read" is not set.
3036 * D corresponds to the outer group->depth dimensions of
3037 * the kernel schedule.
3039 static __isl_give isl_multi_aff
*create_from_access(isl_ctx
*ctx
,
3040 struct gpu_array_ref_group
*group
, int read
)
3045 space
= isl_space_copy(group
->array
->space
);
3046 space
= isl_space_from_range(space
);
3047 space
= isl_space_add_dims(space
, isl_dim_in
, group
->depth
);
3048 space
= isl_space_wrap(space
);
3049 space
= isl_space_map_from_set(space
);
3051 id
= isl_id_alloc(ctx
, read
? "read" : "write", group
);
3052 space
= isl_space_set_tuple_id(space
, isl_dim_in
, id
);
3054 return isl_multi_aff_identity(space
);
3057 /* If any writes in "group" require synchronization, then make sure
3058 * that there is a synchronization node for "kernel" after the node
3059 * following "node" in a sequence.
3061 * If "shared" is set and no synchronization is needed for
3062 * the writes to global memory, then add synchronization before
3063 * the kernel to protect shared memory from being overwritten
3064 * by the next iteration of the core computation.
3065 * No additional synchronization is needed to protect against
3066 * the next copy into shared memory because each element of
3067 * the shared memory tile is always copied by the same thread.
3069 static __isl_give isl_schedule_node
*add_group_write_sync(
3070 __isl_take isl_schedule_node
*node
, struct ppcg_kernel
*kernel
,
3071 struct gpu_array_ref_group
*group
, int shared
)
3075 need_sync
= any_sync_writes_in_group(kernel
, group
);
3077 return isl_schedule_node_free(node
);
3079 node
= isl_schedule_node_parent(node
);
3080 node
= isl_schedule_node_next_sibling(node
);
3081 node
= isl_schedule_node_child(node
, 0);
3082 node
= gpu_tree_ensure_following_sync(node
, kernel
);
3083 } else if (shared
) {
3084 node
= isl_schedule_node_parent(node
);
3085 node
= isl_schedule_node_parent(node
);
3086 node
= gpu_tree_move_down_to_depth(node
, group
->depth
,
3088 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3094 /* Add copy statements to the schedule tree of "node"
3095 * for reading from global memory to private memory (if "read" is set) or
3096 * for writing back from private memory to global memory
3097 * (if "read" is not set) for the array reference group "group" that
3098 * is mapped to private memory.
3099 * On input, "node" points to the kernel node, and it is moved
3100 * back there on output.
3102 * The copies are performed in the order of the array elements.
3103 * The copy statement instances include a reference to the outer
3104 * group->depth dimensions of the kernel schedule for ease of
3105 * combining them with the group tiling.
3107 * That is, the extra schedule is of the form
3111 * where D corresponds to the outer group->depth dimensions of
3112 * the kernel schedule and A to the global array.
3113 * This schedule is unrolled because registers are not addressable.
3115 * The copying is inserted in the schedule tree through an extension
3120 * where the extra domain elements type[D -> A] are those accessed
3122 * A filter is inserted on type[D -> A] to ensure that the element
3123 * is read/written by the same thread that needs the element.
3124 * This filter is obtained by applying
3128 * to the thread filter for the core statements.
3130 * The extension is inserted before the core computation in case of a read
3131 * and after the core computation in case of a write.
3132 * In the latter case, we also make sure that there is a synchronization
3133 * node after the write to global memory, unless this write is performed
3134 * at the outer level of the kernel.
3135 * In principle, this synchronization could be inserted higher
3136 * in the schedule tree depending on where the corresponding reads
3137 * from global memory are performed.
3139 static __isl_give isl_schedule_node
*add_copies_group_private(
3140 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3141 __isl_take isl_schedule_node
*node
, int read
)
3143 isl_union_map
*access
;
3144 isl_union_map
*prefix
;
3145 isl_union_set
*domain
;
3147 isl_multi_aff
*from_access
;
3148 isl_multi_pw_aff
*mpa
;
3149 isl_multi_union_pw_aff
*mupa
;
3150 isl_schedule_node
*graft
;
3151 isl_union_set
*filter
;
3155 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3156 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3158 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3159 empty
= isl_union_map_is_empty(access
);
3160 if (empty
< 0 || empty
) {
3161 isl_union_map_free(access
);
3163 return isl_schedule_node_free(node
);
3164 return gpu_tree_move_up_to_kernel(node
);
3167 group
->array
->global
= 1;
3168 group
->local_array
->global
= 1;
3170 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3171 space
= isl_space_domain(isl_multi_aff_get_space(from_access
));
3172 access
= isl_union_map_preimage_range_multi_aff(access
, from_access
);
3174 filter
= isl_union_set_copy(kernel
->thread_filter
);
3175 filter
= isl_union_set_apply(filter
, isl_union_map_copy(access
));
3176 filter
= isl_union_set_detect_equalities(filter
);
3177 filter
= isl_union_set_coalesce(filter
);
3179 domain
= isl_union_map_range(access
);
3180 access
= isl_union_set_wrapped_domain_map(domain
);
3181 access
= isl_union_map_reverse(access
);
3182 access
= isl_union_map_coalesce(access
);
3183 graft
= isl_schedule_node_from_extension(access
);
3185 space
= isl_space_map_from_set(space
);
3186 mpa
= isl_multi_pw_aff_identity(space
);
3187 mpa
= isl_multi_pw_aff_range_factor_range(mpa
);
3188 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3190 graft
= isl_schedule_node_child(graft
, 0);
3191 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3192 graft
= unroll(graft
);
3194 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3196 graft
= isl_schedule_node_parent(graft
);
3199 node
= isl_schedule_node_graft_before(node
, graft
);
3201 node
= isl_schedule_node_graft_after(node
, graft
);
3202 if (kernel_depth
< group
->depth
)
3203 node
= add_group_write_sync(node
, kernel
, group
, 0);
3206 node
= gpu_tree_move_up_to_kernel(node
);
3211 /* Add copy statements to the schedule tree of "node"
3212 * for reading from global memory to shared memory (if "read" is set) or
3213 * for writing back from shared memory to global memory
3214 * (if "read" is not set) for the array reference group "group" that
3215 * is mapped to shared memory.
3216 * On input, "node" points to the kernel node, and it is moved
3217 * back there on output.
3219 * The copies are performed in the order of the corresponding shared
3221 * The copy statement instances include a reference to the outer
3222 * group->depth dimensions of the kernel schedule for ease of
3223 * combining them with the group tiling.
3225 * If we are performing a read from global memory to shared memory and
3226 * if the array involved is not a scalar, then we copy
3227 * the entire tile to shared memory. This may result in some extra
3228 * elements getting copied, but it should lead to simpler code
3229 * (which means that fewer registers may be needed) and less divergence.
3231 * Otherwise, we only copy the elements that will be read or have been written
3234 * That is, the extra schedule is of the form
3238 * where D corresponds to the outer group->depth dimensions of
3239 * the kernel schedule, A to the global array and T is the corresponding
3240 * shared memory tile.
3242 * The copying is inserted in the schedule tree through an extension
3247 * where the extra domain elements type[D -> A] are those accessed
3248 * by the group. In the case of read from a non-scalar, this set
3249 * is replaced by the entire shared memory tile.
3251 * A filter is inserted on type[D -> A] to map the copy instances
3252 * to the threads. In particular, the thread identifiers are
3253 * equated to the position inside the shared memory tile (T)
3254 * modulo the block size.
3255 * We try to align the innermost tile dimension with the innermost
3256 * thread identifier (x) as a heuristic to improve coalescing.
3257 * In particular, if the dimension of the tile is greater than
3258 * the dimension of the block, then the schedule mapping to the tile
3259 * is broken up into two pieces and the filter is applied to the inner part.
3260 * If, on the other hand, the dimension of the tile is smaller than
3261 * the dimension of the block, then the initial thread identifiers
3262 * are equated to zero and the remaining thread identifiers are
3263 * matched to the memory tile.
3265 * The extension is inserted before the core computation in case of a read
3266 * and after the core computation in case of a write.
3267 * In the case of a read, we first need to make sure there is some
3268 * synchronization before the core computation such that we can put the read
3269 * from global memory to shared memory before that synchronization.
3270 * This ensures that all threads have finished copying into shared memory
3271 * before the shared memory is used.
3272 * We also need to make sure that there is a synchronization node after
3273 * the core computation to ensure that the next load into shared memory
3274 * only happens after all data has been used. There is no need for
3275 * this synchronization if we are at the outer level since then there
3276 * won't be a next load.
3277 * In the case of a write, we need to make sure there is some synchronization
3278 * after the core computation such taht we can put the write from shared
3279 * memory to global memory after that synchronization.
3280 * Unless we are at the outer level, we also need a synchronization node
3281 * after the write to ensure the data is saved to global memory
3282 * before the next iteration write to the same shared memory.
3283 * It also makes sure the data has arrived in global memory before
3284 * it is read in a subsequent iteration.
3286 static __isl_give isl_schedule_node
*add_copies_group_shared(
3287 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3288 __isl_take isl_schedule_node
*node
, int read
)
3290 struct gpu_array_tile
*tile
;
3291 isl_union_map
*access
;
3292 isl_union_set
*domain
;
3293 isl_union_set
*sync
;
3295 isl_multi_aff
*from_access
;
3296 isl_multi_pw_aff
*mpa
;
3297 isl_multi_union_pw_aff
*mupa
;
3298 isl_schedule_node
*graft
;
3299 isl_union_set
*filter
;
3304 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3305 node
= gpu_tree_move_down_to_depth(node
, group
->depth
, kernel
->core
);
3307 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3308 empty
= isl_union_map_is_empty(access
);
3309 if (empty
< 0 || empty
) {
3310 isl_union_map_free(access
);
3312 return isl_schedule_node_free(node
);
3313 return gpu_tree_move_up_to_kernel(node
);
3316 group
->array
->global
= 1;
3317 group
->local_array
->global
= 1;
3319 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3321 tile
= gpu_array_ref_group_tile(group
);
3322 ma
= isl_multi_aff_copy(tile
->tiling
);
3323 ma
= isl_multi_aff_pullback_multi_aff(ma
,
3324 isl_multi_aff_copy(from_access
));
3325 mpa
= isl_multi_pw_aff_from_multi_aff(ma
);
3326 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3328 domain
= isl_union_map_range(access
);
3330 if (read
&& !gpu_array_is_scalar(group
->array
)) {
3332 isl_union_set_free(domain
);
3333 map
= group_tile(group
);
3334 domain
= isl_union_set_from_set(isl_map_wrap(map
));
3337 domain
= isl_union_set_preimage_multi_aff(domain
, from_access
);
3338 access
= isl_union_set_wrapped_domain_map(domain
);
3339 access
= isl_union_map_reverse(access
);
3340 access
= isl_union_map_coalesce(access
);
3341 graft
= isl_schedule_node_from_extension(access
);
3343 graft
= isl_schedule_node_child(graft
, 0);
3345 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3347 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0) {
3348 graft
= isl_schedule_node_band_split(graft
,
3349 tile
->n
- kernel
->n_block
);
3350 graft
= isl_schedule_node_child(graft
, 0);
3352 if (tile
->n
< kernel
->n_block
)
3353 skip
= kernel
->n_block
- tile
->n
;
3356 filter
= set_schedule_modulo(graft
, kernel
->thread_ids
,
3358 if (!kernel
->options
->wrap
)
3359 graft
= snap_band_to_sizes(graft
, kernel
->block_dim
+ skip
,
3361 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0)
3362 graft
= isl_schedule_node_parent(graft
);
3363 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3365 while (graft
&& isl_schedule_node_has_parent(graft
))
3366 graft
= isl_schedule_node_parent(graft
);
3369 if (kernel_depth
< group
->depth
)
3370 node
= gpu_tree_ensure_sync_after_core(node
, kernel
);
3371 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3372 node
= isl_schedule_node_graft_before(node
, graft
);
3374 node
= gpu_tree_move_right_to_sync(node
, kernel
);
3375 node
= isl_schedule_node_graft_after(node
, graft
);
3376 if (kernel_depth
< group
->depth
)
3377 node
= add_group_write_sync(node
, kernel
, group
, 1);
3380 node
= gpu_tree_move_up_to_kernel(node
);
3385 /* Check whether the array reference group "group" is mapped to
3386 * private or shared memory and, if so,
3387 * add copy statements to the schedule tree of "node"
3388 * for reading from global memory to private or shared memory
3389 * (if "read" is set) or for writing back from private or shared memory
3390 * to global memory (if "read" is not set) for this group.
3391 * On input, "node" points to the kernel node, and it is moved
3392 * back there on output.
3394 static __isl_give isl_schedule_node
*add_copies_group(
3395 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3396 __isl_take isl_schedule_node
*node
, int read
)
3398 if (group
->private_tile
)
3399 return add_copies_group_private(kernel
, group
, node
, read
);
3400 if (group
->shared_tile
)
3401 return add_copies_group_shared(kernel
, group
, node
, read
);
3405 /* For each array reference group that is mapped to private or shared memory,
3406 * add copy statements to the schedule tree of "node"
3407 * for reading from global memory to private or shared memory
3408 * and for writing back.
3409 * On input, "node" points to the kernel node, and it is moved
3410 * back there on output.
3412 static __isl_give isl_schedule_node
*add_copies(struct ppcg_kernel
*kernel
,
3413 __isl_take isl_schedule_node
*node
)
3417 for (i
= 0; i
< kernel
->n_array
; ++i
) {
3418 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
3420 for (j
= 0; j
< array
->n_group
; ++j
) {
3421 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3423 node
= add_copies_group(kernel
, group
, node
, 1);
3426 node
= add_copies_group(kernel
, group
, node
, 0);
3435 /* Mark all dimensions in the current band node atomic.
3437 static __isl_give isl_schedule_node
*atomic(__isl_take isl_schedule_node
*node
)
3441 n
= isl_schedule_node_band_n_member(node
);
3442 for (i
= 0; i
< n
; ++i
)
3443 node
= isl_schedule_node_band_member_set_ast_loop_type(node
, i
,
3444 isl_ast_loop_atomic
);
3449 /* Mark "node" atomic, if it is a band node.
3450 * Do the same for all ancestors.
3451 * Return a pointer to "node" (in the updated schedule tree).
3453 static __isl_give isl_schedule_node
*atomic_ancestors(
3454 __isl_take isl_schedule_node
*node
)
3460 if (!isl_schedule_node_has_parent(node
))
3463 pos
= isl_schedule_node_get_child_position(node
);
3464 node
= isl_schedule_node_parent(node
);
3465 if (isl_schedule_node_get_type(node
) == isl_schedule_node_band
)
3466 node
= atomic(node
);
3467 node
= atomic_ancestors(node
);
3468 node
= isl_schedule_node_child(node
, pos
);
3473 /* Collect all write references that require synchronization.
3474 * "node" is assumed to point to the kernel node.
3475 * Each reference is represented by a universe set in a space
3479 * with S[i,j] the statement instance space and R[] the array reference.
3481 * This function should be called before block and thread filters are added.
3483 * Synchronization is needed after a write if there is a subsequent read
3484 * within the same block that may not be performed by the same thread.
3485 * There should not be any dependences between different blocks,
3486 * so we start with the flow dependences within the same kernel invocation
3487 * and we subtract from these those dependences that are mapped
3488 * to the same iteration of the bands where synchronization is inserted.
3489 * We do not remove pairs of instances that are known to map to
3490 * the same thread across different iterations of the intermediate
3491 * bands because the read may be performed by a different thread
3492 * than the one that needs the value if shared memory is involved.
3494 * We also consider all pairs of possible writes that access the same
3495 * memory location and that may be mapped to the same block but not
3496 * to the same iteration of the intermediate bands.
3497 * In theory, it would be possible for one thread to still be in
3498 * a previous iteration of a loop in these bands.
3499 * A write to global memory in this delayed thread could then overwrite
3500 * a write from another thread that has already moved on to
3501 * the next iteration.
3503 * After computing the above writes paired off with reads or writes
3504 * that depend on them, we project onto the domain writes.
3505 * Sychronization is needed after writes to global memory
3506 * through these references.
3508 static __isl_give isl_union_set
*compute_sync_writes(
3509 struct ppcg_kernel
*kernel
, __isl_keep isl_schedule_node
*node
)
3511 isl_union_map
*local
;
3512 isl_union_map
*may_writes
, *shared_access
;
3513 isl_union_map
*kernel_prefix
, *thread_prefix
;
3514 isl_union_map
*equal
;
3515 isl_union_set
*wrap
;
3516 isl_union_set
*domain
;
3518 domain
= isl_schedule_node_get_universe_domain(node
);
3519 kernel_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3520 node
= isl_schedule_node_copy(node
);
3521 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3522 thread_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3523 isl_schedule_node_free(node
);
3525 may_writes
= isl_union_map_copy(kernel
->prog
->scop
->tagged_may_writes
);
3526 may_writes
= isl_union_map_curry(may_writes
);
3527 may_writes
= isl_union_map_intersect_domain(may_writes
, domain
);
3528 may_writes
= isl_union_map_uncurry(may_writes
);
3529 shared_access
= isl_union_map_copy(may_writes
);
3530 shared_access
= isl_union_map_apply_range(shared_access
,
3531 isl_union_map_reverse(may_writes
));
3533 local
= isl_union_map_copy(kernel
->prog
->scop
->tagged_dep_flow
);
3534 local
= isl_union_map_union(local
, shared_access
);
3535 local
= isl_union_map_zip(local
);
3537 equal
= isl_union_map_apply_range(kernel_prefix
,
3538 isl_union_map_reverse(isl_union_map_copy(kernel_prefix
)));
3539 wrap
= isl_union_map_wrap(equal
);
3540 local
= isl_union_map_intersect_domain(local
, wrap
);
3541 equal
= isl_union_map_apply_range(thread_prefix
,
3542 isl_union_map_reverse(isl_union_map_copy(thread_prefix
)));
3543 wrap
= isl_union_map_wrap(equal
);
3544 local
= isl_union_map_subtract_domain(local
, wrap
);
3546 local
= isl_union_map_zip(local
);
3547 local
= isl_union_map_universe(local
);
3549 return isl_union_map_domain(local
);
3552 /* Group the domain elements into a single space, named kernelX,
3553 * with X the kernel sequence number "kernel_id".
3555 static __isl_give isl_schedule_node
*group_statements(
3556 __isl_take isl_schedule_node
*node
, int kernel_id
)
3564 snprintf(buffer
, sizeof(buffer
), "kernel%d", kernel_id
);
3565 id
= isl_id_alloc(isl_schedule_node_get_ctx(node
), buffer
, NULL
);
3566 return isl_schedule_node_group(node
, id
);
3569 /* Create a ppcg_kernel representing the domain instances that reach "node"
3570 * and insert a mark node pointing to the ppcg_kernel before "node".
3571 * The band that "node" points to is the band that needs to be mapped
3572 * to block identifiers. The band that needs to be mapped to thread
3573 * identifiers should be marked by a "thread" mark by the caller.
3574 * This mark is removed by this function.
3575 * If "scale" is set, then the band that "node" points to is scaled
3578 * Mark all outer band nodes as atomic to ensure each kernel is only
3580 * If the domain elements that reach "node" live in more than one space,
3581 * then group the domain elements into a single space, named kernelX,
3582 * with X the kernel sequence number.
3584 * Insert a guard node governing the kernel node to ensure that
3585 * no kernels with zero blocks are launched.
3587 * Insert a context node describing the block and thread
3588 * identifiers inside the kernel mark.
3589 * The context node needs to be inserted after the effective block size
3590 * has been determined such that the bounds on the thread identifiers
3591 * would reflect the effective block size.
3592 * Insert a filter node inside the context node mapping the statement
3593 * instances to block identifiers. In particular, the block identifiers
3594 * are equated to the partial schedule of band that was marked for mapping
3595 * to blocks modulo the grid size.
3596 * Insert a filter node inside the "thread" mark mapping the statement
3597 * instances to thread identifiers. In particular, the thread identifiers
3598 * are equated to the partial schedule of band that was marked for mapping
3599 * to threads modulo the block size.
3601 * Compute array reference groups for all arrays, set the local
3602 * array bounds based on the set of domain instances that reach
3603 * the kernel node, check the total amount of shared memory used
3604 * and compute all group tilings.
3605 * The array reference groups are computed after the block filter
3606 * has been inserted because it affects the mapping to shared or
3607 * private memory. This computation also requires the thread filter
3608 * (in the ppcg_kernel object), but this thread filter should not
3609 * have been added to the schedule tree yet since the computation
3610 * requires the schedule of the band that needs to be mapped to
3611 * threads before the privatization is applied.
3613 * If any array reference group requires the band mapped to threads
3614 * to be unrolled, then we perform the required unrolling.
3616 * We save a copy of the schedule that may influence the mappings
3617 * to shared or private memory in kernel->shared_schedule.
3619 * Finally, we add synchronization and copy statements to the schedule tree,
3620 * remove the "thread" mark and create representations for the local
3621 * variables in the kernel.
3623 * We keep a copy of the isl_id that points to the kernel to ensure
3624 * that the kernel does not get destroyed if the schedule node
3625 * is freed due to some error condition.
3627 static __isl_give isl_schedule_node
*create_kernel(struct gpu_gen
*gen
,
3628 __isl_take isl_schedule_node
*node
, int scale
,
3629 __isl_keep isl_multi_val
*sizes
)
3631 struct ppcg_kernel
*kernel
;
3633 isl_schedule_node
*node_thread
;
3634 isl_union_map
*host_schedule
;
3635 isl_set
*host_domain
;
3636 isl_union_set
*domain
;
3637 int single_statement
;
3639 kernel
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel
);
3640 kernel
= ppcg_kernel_create_local_arrays(kernel
, gen
->prog
);
3642 return isl_schedule_node_free(node
);
3644 domain
= isl_schedule_node_get_domain(node
);
3645 single_statement
= isl_union_set_n_set(domain
) == 1;
3647 kernel
->ctx
= gen
->ctx
;
3648 kernel
->prog
= gen
->prog
;
3649 kernel
->options
= gen
->options
;
3650 kernel
->context
= extract_context(node
, gen
->prog
);
3651 kernel
->core
= isl_union_set_universe(isl_union_set_copy(domain
));
3652 kernel
->arrays
= accessed_by_domain(isl_union_set_copy(domain
),
3654 kernel
->n_grid
= n_outer_coincidence(node
);
3655 node_thread
= isl_schedule_node_copy(node
);
3656 node_thread
= gpu_tree_move_down_to_thread(node_thread
, kernel
->core
);
3657 node_thread
= isl_schedule_node_child(node_thread
, 0);
3658 kernel
->n_block
= n_outer_coincidence(node_thread
);
3659 isl_schedule_node_free(node_thread
);
3660 kernel
->id
= gen
->kernel_id
++;
3661 read_grid_and_block_sizes(kernel
, gen
);
3663 kernel
->sync_writes
= compute_sync_writes(kernel
, node
);
3665 host_schedule
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3666 host_domain
= isl_set_from_union_set(isl_union_map_range(
3669 node
= atomic_ancestors(node
);
3671 id
= isl_id_alloc(gen
->ctx
, "kernel", kernel
);
3672 id
= isl_id_set_free_user(id
, &ppcg_kernel_free_wrap
);
3673 node
= isl_schedule_node_insert_mark(node
, isl_id_copy(id
));
3675 if (!single_statement
)
3676 node
= group_statements(node
, kernel
->id
);
3678 node
= isl_schedule_node_child(node
, 0);
3679 node
= split_band(node
, kernel
->n_grid
);
3680 kernel
->block_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3681 kernel
->n_grid
, "b");
3682 kernel
->block_filter
= set_schedule_modulo(node
, kernel
->block_ids
,
3684 kernel
->grid_size
= extract_grid_size(kernel
,
3685 isl_union_set_copy(domain
));
3686 if (!kernel
->options
->wrap
)
3687 node
= snap_band_to_sizes(node
, kernel
->grid_dim
,
3690 node
= scale_band(node
, isl_multi_val_copy(sizes
));
3691 node
= isl_schedule_node_parent(node
);
3692 if (!single_statement
)
3693 node
= isl_schedule_node_parent(node
);
3694 node
= insert_guard(node
, kernel
->context
, kernel
->grid_size
,
3696 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3697 node
= isl_schedule_node_child(node
, 0);
3698 node
= split_band(node
, kernel
->n_block
);
3699 kernel
->thread_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3700 kernel
->n_block
, "t");
3701 kernel
->thread_filter
= set_schedule_modulo(node
, kernel
->thread_ids
,
3703 if (extract_block_size(kernel
, domain
) < 0)
3704 node
= isl_schedule_node_free(node
);
3706 node
= gpu_tree_move_up_to_kernel(node
);
3707 node
= isl_schedule_node_child(node
, 0);
3708 node
= insert_context(kernel
, node
);
3709 node
= isl_schedule_node_child(node
, 0);
3710 node
= isl_schedule_node_insert_filter(node
,
3711 isl_union_set_copy(kernel
->block_filter
));
3713 node
= gpu_tree_move_up_to_kernel(node
);
3715 if (gpu_group_references(kernel
, node
) < 0)
3716 node
= isl_schedule_node_free(node
);
3717 localize_bounds(kernel
, host_domain
);
3718 isl_set_free(host_domain
);
3720 check_shared_memory_bound(kernel
);
3721 mark_global_arrays(kernel
);
3722 compute_group_tilings(kernel
);
3724 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3725 node
= isl_schedule_node_child(node
, 0);
3726 if (!kernel
->options
->wrap
)
3727 node
= snap_band_to_sizes(node
, kernel
->block_dim
,
3729 node
= isl_schedule_node_insert_filter(node
,
3730 isl_union_set_copy(kernel
->thread_filter
));
3731 if (kernel_requires_unroll(kernel
)) {
3732 node
= isl_schedule_node_child(node
, 0);
3733 node
= unroll(node
);
3736 node
= gpu_tree_move_up_to_thread(node
);
3737 kernel
->shared_schedule_dim
=
3738 isl_schedule_node_get_schedule_depth(node
);
3739 kernel
->shared_schedule
=
3740 isl_schedule_node_get_prefix_schedule_union_pw_multi_aff(node
);
3742 node
= gpu_tree_move_up_to_kernel(node
);
3744 node
= add_sync(kernel
, node
);
3745 node
= add_copies(kernel
, node
);
3747 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3748 node
= isl_schedule_node_delete(node
);
3750 node
= gpu_tree_move_up_to_kernel(node
);
3752 if (create_kernel_vars(kernel
) < 0)
3753 node
= isl_schedule_node_free(node
);
3755 if (!single_statement
)
3756 node
= isl_schedule_node_parent(node
);
3757 node
= isl_schedule_node_parent(node
);
3763 /* Insert a zero-dimensional permutable band at "node".
3765 static __isl_give isl_schedule_node
*insert_empty_permutable_band(
3766 __isl_take isl_schedule_node
*node
)
3769 isl_schedule
*schedule
;
3770 isl_union_set
*domain
;
3771 isl_multi_union_pw_aff
*mupa
;
3773 schedule
= isl_schedule_node_get_schedule(node
);
3774 domain
= isl_schedule_get_domain(schedule
);
3775 space
= isl_union_set_get_space(domain
);
3776 isl_union_set_free(domain
);
3777 isl_schedule_free(schedule
);
3779 space
= isl_space_set_from_params(space
);
3780 mupa
= isl_multi_union_pw_aff_zero(space
);
3781 node
= isl_schedule_node_insert_partial_schedule(node
, mupa
);
3782 node
= isl_schedule_node_band_set_permutable(node
, 1);
3787 /* If "node" is the outermost permutable band that can be mapped to block and
3788 * thread identifiers in its branch (or a leaf with no such outer bands),
3789 * then mark the band as such, attaching a ppcg_kernel to the mark.
3791 * If "node" originally points to a leaf, then insert a zero-dimensional
3792 * permutable band such that we can assume that "node" always
3793 * points to a band node.
3795 * Tile "node" using user specified tile sizes, after splitting the band
3796 * if the number of specified tile sizes is smaller than the dimension
3797 * of the band. Mark the point band of this tiling as the band that
3798 * needs to be mapped to threads.
3799 * Create a kernel representing the domain instances that reach "node" and
3800 * insert a mark node pointing to the ppcg_kernel before the band node.
3802 static __isl_give isl_schedule_node
*mark_outer_permutable(
3803 __isl_take isl_schedule_node
*node
, void *user
)
3805 struct gpu_gen
*gen
= user
;
3811 isl_multi_val
*sizes
;
3813 outer
= is_outer_tilable(node
);
3815 return isl_schedule_node_free(node
);
3819 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
3820 node
= insert_empty_permutable_band(node
);
3822 tile_len
= isl_schedule_node_band_n_member(node
);
3823 tile_size
= read_tile_sizes(gen
, &tile_len
);
3825 return isl_schedule_node_free(node
);
3826 if (tile_len
< isl_schedule_node_band_n_member(node
))
3827 node
= isl_schedule_node_band_split(node
, tile_len
);
3828 sizes
= construct_band_tiles_sizes(node
, tile_size
);
3829 node
= tile_band(node
, isl_multi_val_copy(sizes
));
3830 node
= isl_schedule_node_child(node
, 0);
3831 id
= isl_id_alloc(gen
->ctx
, "thread", NULL
);
3832 node
= isl_schedule_node_insert_mark(node
, id
);
3833 node
= isl_schedule_node_parent(node
);
3835 scale
= gen
->options
->scale_tile_loops
;
3836 node
= create_kernel(gen
, node
, scale
, sizes
);
3837 isl_multi_val_free(sizes
);
3843 /* Does the subtree rooted at "node" have any suitably permutable band nodes?
3844 * That is, does it have any nodes that are permutable and that
3845 * have a least one coincident dimension?
3847 static int subtree_has_permutable_bands(__isl_keep isl_schedule_node
*node
)
3849 int any_parallelism
= 0;
3851 if (isl_schedule_node_foreach_descendant_top_down(node
, &set_permutable
,
3852 &any_parallelism
) < 0 &&
3856 return any_parallelism
;
3859 /* Mark all variables that are accessed by the statement instances in "domain"
3860 * and that are local to "prog" as requiring a declaration in the host code.
3862 static int declare_accessed_local_variables(struct gpu_prog
*prog
,
3863 __isl_keep isl_union_set
*domain
)
3865 isl_union_set
*arrays
;
3868 if (!ppcg_scop_any_hidden_declarations(prog
->scop
))
3870 arrays
= accessed_by_domain(isl_union_set_copy(domain
), prog
);
3872 for (i
= 0; i
< prog
->n_array
; ++i
) {
3877 if (!prog
->array
[i
].local
)
3879 space
= isl_set_get_space(prog
->array
[i
].extent
);
3880 set
= isl_union_set_extract_set(arrays
, space
);
3881 empty
= isl_set_plain_is_empty(set
);
3886 prog
->array
[i
].declare_local
= 1;
3889 isl_union_set_free(arrays
);
3892 isl_union_set_free(arrays
);
3896 /* If "node" points to a set node, then separate its children
3897 * into subtrees that have suitably permutable bands and
3898 * those that do not.
3899 * Adjust the schedule tree in order to execute the second group
3900 * after the first group and return a pointer to the first group,
3901 * assuming there are any such subtrees.
3902 * Mark all local variables in "prog" that are accessed by
3903 * the second group as requiring a declaration on the host.
3905 static __isl_give isl_schedule_node
*isolate_permutable_subtrees(
3906 __isl_take isl_schedule_node
*node
, struct gpu_prog
*prog
)
3909 isl_union_set
*filter
;
3914 if (isl_schedule_node_get_type(node
) != isl_schedule_node_set
)
3917 n
= isl_schedule_node_n_children(node
);
3919 return isl_schedule_node_free(node
);
3921 node
= isl_schedule_node_child(node
, 0);
3922 filter
= isl_schedule_node_filter_get_filter(node
);
3923 node
= isl_schedule_node_parent(node
);
3924 space
= isl_union_set_get_space(filter
);
3925 isl_union_set_free(filter
);
3926 filter
= isl_union_set_empty(space
);
3928 for (i
= 0; i
< n
; ++i
) {
3931 node
= isl_schedule_node_child(node
, i
);
3932 parallelism
= subtree_has_permutable_bands(node
);
3933 if (parallelism
< 0) {
3934 node
= isl_schedule_node_free(node
);
3935 } else if (!parallelism
) {
3936 isl_union_set
*filter_i
;
3937 filter_i
= isl_schedule_node_filter_get_filter(node
);
3938 filter
= isl_union_set_union(filter
, filter_i
);
3940 node
= isl_schedule_node_parent(node
);
3943 if (declare_accessed_local_variables(prog
, filter
) < 0)
3944 node
= isl_schedule_node_free(node
);
3945 node
= isl_schedule_node_order_after(node
, filter
);
3950 /* Replace any reference to an array element in the range of "copy"
3951 * by a reference to all array elements (defined by the extent of the array).
3953 static __isl_give isl_union_map
*approximate_copy_out(
3954 __isl_take isl_union_map
*copy
, struct gpu_prog
*prog
)
3959 res
= isl_union_map_empty(isl_union_map_get_space(copy
));
3961 for (i
= 0; i
< prog
->n_array
; ++i
) {
3964 isl_union_map
*copy_i
;
3965 isl_union_set
*extent
, *domain
;
3967 space
= isl_space_copy(prog
->array
[i
].space
);
3968 extent
= isl_union_set_from_set(isl_set_universe(space
));
3969 copy_i
= isl_union_map_copy(copy
);
3970 copy_i
= isl_union_map_intersect_range(copy_i
, extent
);
3971 set
= isl_set_copy(prog
->array
[i
].extent
);
3972 extent
= isl_union_set_from_set(set
);
3973 domain
= isl_union_map_domain(copy_i
);
3974 copy_i
= isl_union_map_from_domain_and_range(domain
, extent
);
3975 res
= isl_union_map_union(res
, copy_i
);
3978 isl_union_map_free(copy
);
3983 /* Insert "kernel" marks that point to a ppcg_kernel structure
3984 * in front of all outermost tilable band that (by construction)
3985 * have at least one parallel loop.
3987 static __isl_give isl_schedule_node
*mark_kernels(struct gpu_gen
*gen
,
3988 __isl_take isl_schedule_node
*node
)
3990 return isl_schedule_node_map_descendant_bottom_up(node
,
3991 &mark_outer_permutable
, gen
);
3994 /* Save the schedule "schedule" to a file called "filename".
3995 * The schedule is printed in block style.
3997 static void save_schedule(__isl_keep isl_schedule
*schedule
,
3998 const char *filename
)
4007 file
= fopen(filename
, "w");
4009 fprintf(stderr
, "Unable to open '%s' for writing\n", filename
);
4012 ctx
= isl_schedule_get_ctx(schedule
);
4013 p
= isl_printer_to_file(ctx
, file
);
4014 p
= isl_printer_set_yaml_style(p
, ISL_YAML_STYLE_BLOCK
);
4015 p
= isl_printer_print_schedule(p
, schedule
);
4016 isl_printer_free(p
);
4020 /* Load and return a schedule from a file called "filename".
4022 static __isl_give isl_schedule
*load_schedule(isl_ctx
*ctx
,
4023 const char *filename
)
4026 isl_schedule
*schedule
;
4028 file
= fopen(filename
, "r");
4030 fprintf(stderr
, "Unable to open '%s' for reading\n", filename
);
4033 schedule
= isl_schedule_read_from_file(ctx
, file
);
4039 /* Construct schedule constraints from the dependences in prog->scop and
4040 * the array order dependences in prog->array_order.
4042 * If live range reordering is allowed, then we need to make sure
4043 * that live ranges on arrays are not run in parallel since doing
4044 * so would require array expansion. We therefore add the array
4045 * order dependences to the coincidence dependences. Non-zero array
4046 * order dependences will then prevent a schedule dimension from being
4047 * considered parallel.
4048 * Live ranges derived from scalars are allowed to be run in parallel
4049 * since we force the scalars to be mapped to private memory in
4050 * check_scalar_live_ranges.
4051 * If live range reordering is allowed, then the false dependences
4052 * are not added to the validity constraints as that would prevent
4053 * reordering. Instead, the external false dependences that enforce that reads
4054 * from potentially live-in data precede any later write and
4055 * that writes of potentially live-out data follow any other earlier write
4056 * are added to the validity and the coincidence constraints.
4057 * The false dependences are still added to the proximity constraints
4058 * for consistency with the case where live range reordering is not allowed.
4059 * The coincidence constraints then consist of flow dependences,
4060 * external false dependences and array order dependences.
4061 * The independences can be filtered out from the first two sets.
4062 * They have already been filtered out from the array order dependences
4063 * on a per array basis in collect_order_dependences.
4064 * There is no need for a per array handling of the other two sets
4065 * as there should be no flow or external false dependence on local
4066 * variables that can be filtered out.
4068 static __isl_give isl_schedule_constraints
*construct_schedule_constraints(
4069 struct gpu_prog
*prog
)
4071 isl_union_set
*domain
;
4072 isl_union_map
*dep_raw
, *dep
;
4073 isl_union_map
*validity
, *proximity
, *coincidence
;
4074 isl_schedule_constraints
*sc
;
4076 domain
= isl_union_set_copy(prog
->scop
->domain
);
4077 sc
= isl_schedule_constraints_on_domain(domain
);
4078 sc
= isl_schedule_constraints_set_context(sc
,
4079 isl_set_copy(prog
->scop
->context
));
4080 if (prog
->scop
->options
->live_range_reordering
) {
4081 sc
= isl_schedule_constraints_set_conditional_validity(sc
,
4082 isl_union_map_copy(prog
->scop
->tagged_dep_flow
),
4083 isl_union_map_copy(prog
->scop
->tagged_dep_order
));
4084 proximity
= isl_union_map_copy(prog
->scop
->dep_flow
);
4085 validity
= isl_union_map_copy(proximity
);
4086 validity
= isl_union_map_union(validity
,
4087 isl_union_map_copy(prog
->scop
->dep_forced
));
4088 proximity
= isl_union_map_union(proximity
,
4089 isl_union_map_copy(prog
->scop
->dep_false
));
4090 coincidence
= isl_union_map_copy(validity
);
4091 coincidence
= isl_union_map_subtract(coincidence
,
4092 isl_union_map_copy(prog
->scop
->independence
));
4093 coincidence
= isl_union_map_union(coincidence
,
4094 isl_union_map_copy(prog
->array_order
));
4096 dep_raw
= isl_union_map_copy(prog
->scop
->dep_flow
);
4097 dep
= isl_union_map_copy(prog
->scop
->dep_false
);
4098 dep
= isl_union_map_union(dep
, dep_raw
);
4099 dep
= isl_union_map_coalesce(dep
);
4100 proximity
= isl_union_map_copy(dep
);
4101 coincidence
= isl_union_map_copy(dep
);
4104 sc
= isl_schedule_constraints_set_validity(sc
, validity
);
4105 sc
= isl_schedule_constraints_set_coincidence(sc
, coincidence
);
4106 sc
= isl_schedule_constraints_set_proximity(sc
, proximity
);
4108 if (prog
->scop
->options
->debug
->dump_schedule_constraints
)
4109 isl_schedule_constraints_dump(sc
);
4113 /* Compute an appropriate schedule based on the accesses in
4114 * gen->read and gen->write.
4116 * We derive schedule constraints from the dependences in gen->prog->scop
4117 * and then use isl to compute a schedule that has a parallel loop
4118 * in each tilable band.
4120 static __isl_give isl_schedule
*compute_schedule(struct gpu_gen
*gen
)
4122 isl_schedule_constraints
*sc
;
4123 isl_schedule
*schedule
;
4125 sc
= construct_schedule_constraints(gen
->prog
);
4126 schedule
= isl_schedule_constraints_compute_schedule(sc
);
4131 /* If the band node "node" has exactly one member then mark it permutable.
4133 static __isl_give isl_schedule_node
*band_set_permutable(
4134 __isl_take isl_schedule_node
*node
,
4135 __isl_keep isl_schedule_constraints
*sc
)
4137 if (isl_schedule_node_band_n_member(node
) == 1)
4138 node
= isl_schedule_node_band_set_permutable(node
, 1);
4143 /* Return the coincidence constraints between pairs of instances
4144 * that are scheduled together by the ancestors of "node".
4145 * That is, select those coincidence constraints that relate
4146 * pairs of instances that have the same value for the prefix schedule.
4147 * If the schedule depth is zero, then the prefix schedule does not
4148 * contain any information, so we intersect domain and range
4149 * of the schedule constraints with the reaching domain elements instead.
4151 static __isl_give isl_union_map
*get_local_coincidence(
4152 __isl_keep isl_schedule_node
*node
,
4153 __isl_keep isl_schedule_constraints
*sc
)
4155 isl_union_map
*coincidence
;
4156 isl_multi_union_pw_aff
*prefix
;
4157 isl_union_pw_multi_aff
*contraction
;
4159 coincidence
= isl_schedule_constraints_get_coincidence(sc
);
4160 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4161 if (isl_schedule_node_get_schedule_depth(node
) == 0) {
4162 isl_union_set
*domain
;
4164 domain
= isl_schedule_node_get_domain(node
);
4165 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4167 coincidence
= isl_union_map_intersect_domain(coincidence
,
4168 isl_union_set_copy(domain
));
4169 coincidence
= isl_union_map_intersect_range(coincidence
,
4174 prefix
= isl_schedule_node_get_prefix_schedule_multi_union_pw_aff(node
);
4175 prefix
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(prefix
,
4177 return isl_union_map_eq_at_multi_union_pw_aff(coincidence
, prefix
);
4180 /* For each member in the band node "node", determine whether
4181 * it is coincident with respect to the outer nodes and mark
4184 * That is, for each coincidence constraint between pairs
4185 * of instances that are scheduled together by the outer nodes,
4186 * check that domain and range are assigned the same value
4187 * by the band member. This test is performed by checking
4188 * that imposing the same value for the band member does not
4189 * remove any elements from the set of coincidence constraints.
4191 static __isl_give isl_schedule_node
*band_set_coincident(
4192 __isl_take isl_schedule_node
*node
,
4193 __isl_keep isl_schedule_constraints
*sc
)
4195 isl_union_map
*coincidence
;
4196 isl_union_pw_multi_aff
*contraction
;
4197 isl_multi_union_pw_aff
*partial
;
4200 coincidence
= get_local_coincidence(node
, sc
);
4202 partial
= isl_schedule_node_band_get_partial_schedule(node
);
4203 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4204 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4206 n
= isl_schedule_node_band_n_member(node
);
4207 for (i
= 0; i
< n
; ++i
) {
4208 isl_union_map
*coincidence_i
;
4209 isl_union_pw_aff
*upa
;
4210 isl_multi_union_pw_aff
*partial_i
;
4213 upa
= isl_multi_union_pw_aff_get_union_pw_aff(partial
, i
);
4214 partial_i
= isl_multi_union_pw_aff_from_union_pw_aff(upa
);
4215 coincidence_i
= isl_union_map_copy(coincidence
);
4216 coincidence_i
= isl_union_map_eq_at_multi_union_pw_aff(
4217 coincidence_i
, partial_i
);
4218 subset
= isl_union_map_is_subset(coincidence
, coincidence_i
);
4219 isl_union_map_free(coincidence_i
);
4223 node
= isl_schedule_node_band_member_set_coincident(node
, i
,
4227 node
= isl_schedule_node_free(node
);
4228 isl_multi_union_pw_aff_free(partial
);
4229 isl_union_map_free(coincidence
);
4234 /* If "node" is a band, then set its properties.
4236 * In particular, if the band has exactly one member, then mark it permutable.
4237 * Mark the band member coincident based on the coincidence constraints
4240 static __isl_give isl_schedule_node
*set_band_properties(
4241 __isl_take isl_schedule_node
*node
, void *user
)
4243 isl_schedule_constraints
*sc
= user
;
4245 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
4247 if (isl_schedule_node_band_n_member(node
) == 0)
4250 node
= band_set_permutable(node
, sc
);
4251 node
= band_set_coincident(node
, sc
);
4256 /* Return the original schedule with all bands marked permutable and
4257 * all band members marked coincident based on the coincidence constraints.
4258 * The bands are explicitly marked permutable so that they will be considered
4259 * by mark_outer_permutable.
4261 static __isl_give isl_schedule
*determine_properties_original_schedule(
4262 struct gpu_gen
*gen
)
4264 isl_schedule
*schedule
;
4265 isl_schedule_constraints
*sc
;
4267 schedule
= isl_schedule_copy(gen
->prog
->scop
->schedule
);
4268 sc
= construct_schedule_constraints(gen
->prog
);
4269 schedule
= isl_schedule_map_schedule_node_bottom_up(schedule
,
4270 &set_band_properties
, sc
);
4271 isl_schedule_constraints_free(sc
);
4276 /* Obtain a schedule for the scop, by reading it from
4277 * a file, by computing one or by determining the properties
4278 * of the original schedule.
4280 static __isl_give isl_schedule
*get_schedule(struct gpu_gen
*gen
)
4282 isl_schedule
*schedule
;
4284 if (gen
->options
->load_schedule_file
) {
4285 schedule
= load_schedule(gen
->ctx
,
4286 gen
->options
->load_schedule_file
);
4288 if (gen
->options
->reschedule
)
4289 schedule
= compute_schedule(gen
);
4291 schedule
= determine_properties_original_schedule(gen
);
4292 if (gen
->options
->save_schedule_file
)
4293 save_schedule(schedule
,
4294 gen
->options
->save_schedule_file
);
4296 if (gen
->options
->debug
->dump_schedule
)
4297 isl_schedule_dump(schedule
);
4302 /* Construct the string "<a>_<b>".
4304 static char *concat(isl_ctx
*ctx
, const char *a
, const char *b
)
4309 p
= isl_printer_to_str(ctx
);
4310 p
= isl_printer_print_str(p
, a
);
4311 p
= isl_printer_print_str(p
, "_");
4312 p
= isl_printer_print_str(p
, b
);
4313 s
= isl_printer_get_str(p
);
4314 isl_printer_free(p
);
4319 /* For each array in "prog" of which an element appears in "accessed" and
4320 * that is not a read only scalar, create a zero-dimensional universe set
4321 * of which the tuple id has name "<prefix>_<name of array>" and a user
4322 * pointer pointing to the array (gpu_array_info).
4324 * If the array is local to "prog", then make sure it will be declared
4327 * Return the list of these universe sets.
4329 static __isl_give isl_union_set_list
*create_copy_filters(struct gpu_prog
*prog
,
4330 const char *prefix
, __isl_take isl_union_set
*accessed
)
4334 isl_union_set_list
*filters
;
4337 filters
= isl_union_set_list_alloc(ctx
, 0);
4338 for (i
= 0; i
< prog
->n_array
; ++i
) {
4339 struct gpu_array_info
*array
= &prog
->array
[i
];
4341 isl_set
*accessed_i
;
4345 isl_union_set
*uset
;
4347 if (gpu_array_is_read_only_scalar(array
))
4350 space
= isl_space_copy(array
->space
);
4351 accessed_i
= isl_union_set_extract_set(accessed
, space
);
4352 empty
= isl_set_plain_is_empty(accessed_i
);
4353 isl_set_free(accessed_i
);
4355 filters
= isl_union_set_list_free(filters
);
4363 array
->declare_local
= 1;
4365 name
= concat(ctx
, prefix
, array
->name
);
4366 id
= name
? isl_id_alloc(ctx
, name
, array
) : NULL
;
4368 space
= isl_space_set_alloc(ctx
, 0, 0);
4369 space
= isl_space_set_tuple_id(space
, isl_dim_set
, id
);
4370 uset
= isl_union_set_from_set(isl_set_universe(space
));
4372 filters
= isl_union_set_list_add(filters
, uset
);
4374 isl_union_set_free(accessed
);
4379 /* Make sure that code for the statements in "filters" that
4380 * copy arrays to or from the device is only generated when
4381 * the size of the corresponding array is positive.
4382 * That is, add a set node underneath "graft" with "filters" as children
4383 * and for each child add a guard that the selects the parameter
4384 * values for which the corresponding array has a positive size.
4385 * The array is available in the user pointer of the statement identifier.
4386 * "depth" is the schedule depth of the position where "graft"
4389 static __isl_give isl_schedule_node
*insert_positive_size_guards(
4390 __isl_take isl_schedule_node
*graft
,
4391 __isl_take isl_union_set_list
*filters
, int depth
)
4395 graft
= isl_schedule_node_child(graft
, 0);
4396 graft
= isl_schedule_node_insert_set(graft
, filters
);
4397 n
= isl_schedule_node_n_children(graft
);
4398 for (i
= 0; i
< n
; ++i
) {
4399 isl_union_set
*filter
;
4400 isl_set
*domain
, *guard
;
4402 struct gpu_array_info
*array
;
4404 graft
= isl_schedule_node_child(graft
, i
);
4405 filter
= isl_schedule_node_filter_get_filter(graft
);
4406 domain
= isl_set_from_union_set(filter
);
4407 id
= isl_set_get_tuple_id(domain
);
4408 array
= isl_id_get_user(id
);
4410 isl_set_free(domain
);
4411 guard
= gpu_array_positive_size_guard(array
);
4412 guard
= isl_set_from_params(guard
);
4413 guard
= isl_set_add_dims(guard
, isl_dim_set
, depth
);
4414 graft
= isl_schedule_node_child(graft
, 0);
4415 graft
= isl_schedule_node_insert_guard(graft
, guard
);
4416 graft
= isl_schedule_node_parent(graft
);
4417 graft
= isl_schedule_node_parent(graft
);
4419 graft
= isl_schedule_node_parent(graft
);
4424 /* Create a graft for copying arrays to or from the device,
4425 * whenever the size of the array is strictly positive.
4426 * Each statement is called "<prefix>_<name of array>" and
4427 * the identifier has a user pointer pointing to the array.
4428 * The graft will be added at the position specified by "node".
4429 * "copy" contains the array elements that need to be copied.
4430 * Only arrays of which some elements need to be copied
4431 * will have a corresponding statement in the graph.
4432 * Note though that each such statement will copy the entire array.
4434 static __isl_give isl_schedule_node
*create_copy_device(struct gpu_prog
*prog
,
4435 __isl_keep isl_schedule_node
*node
, const char *prefix
,
4436 __isl_take isl_union_set
*copy
)
4441 isl_union_set
*all
, *domain
;
4442 isl_union_set_list
*filters
;
4443 isl_union_map
*extension
;
4444 isl_schedule_node
*graft
;
4447 depth
= isl_schedule_node_get_schedule_depth(node
);
4448 filters
= create_copy_filters(prog
, prefix
, copy
);
4449 all
= isl_union_set_list_union(isl_union_set_list_copy(filters
));
4451 space
= depth
< 0 ? NULL
: isl_space_set_alloc(ctx
, 0, depth
);
4452 domain
= isl_union_set_from_set(isl_set_universe(space
));
4453 extension
= isl_union_map_from_domain_and_range(domain
, all
);
4454 graft
= isl_schedule_node_from_extension(extension
);
4457 return isl_schedule_node_free(graft
);
4458 if (isl_union_set_list_n_union_set(filters
) == 0) {
4459 isl_union_set_list_free(filters
);
4463 return insert_positive_size_guards(graft
, filters
, depth
);
4466 /* Return (the universe spaces of) the arrays that are declared
4467 * inside the scop corresponding to "prog" and for which all
4468 * potential writes inside the scop form a subset of "domain".
4470 static __isl_give isl_union_set
*extract_local_accesses(struct gpu_prog
*prog
,
4471 __isl_keep isl_union_set
*domain
)
4474 isl_union_set
*local
;
4476 local
= isl_union_set_empty(isl_union_set_get_space(domain
));
4478 for (i
= 0; i
< prog
->n_array
; ++i
) {
4480 isl_union_map
*to_outer
;
4481 isl_union_map
*may_write
;
4482 isl_union_set
*write_domain
;
4483 isl_union_set
*fields
;
4486 if (!prog
->array
[i
].local
)
4489 set
= isl_set_universe(isl_space_copy(prog
->array
[i
].space
));
4490 to_outer
= isl_union_map_copy(prog
->to_outer
);
4491 to_outer
= isl_union_map_intersect_range(to_outer
,
4492 isl_union_set_from_set(isl_set_copy(set
)));
4493 fields
= isl_union_map_domain(to_outer
);
4494 may_write
= isl_union_map_copy(prog
->may_write
);
4495 may_write
= isl_union_map_intersect_range(may_write
, fields
);
4496 write_domain
= isl_union_map_domain(may_write
);
4497 subset
= isl_union_set_is_subset(write_domain
, domain
);
4498 isl_union_set_free(write_domain
);
4502 return isl_union_set_free(local
);
4503 } else if (subset
) {
4504 local
= isl_union_set_add_set(local
, set
);
4513 /* Internal data structure for node_may_persist.
4515 * "tagger" maps tagged iteration domains to the corresponding untagged
4518 * "may_persist_flow" is the set of all tagged dataflow dependences
4519 * with those dependences removed that either precede or follow
4520 * the kernel launch in a sequence.
4521 * "inner_band_flow" is the set of all tagged dataflow dependences
4522 * that are local to a given iteration of the outer band nodes
4523 * with respect to the current node.
4524 * "local_flow" is equal to "inner_band_flow", except that the domain
4525 * and the range have been intersected with intermediate filters
4526 * on children of sets or sequences.
4528 struct ppcg_may_persist_data
{
4529 isl_union_pw_multi_aff
*tagger
;
4531 isl_union_map
*local_flow
;
4532 isl_union_map
*inner_band_flow
;
4533 isl_union_map
*may_persist_flow
;
4536 /* Update the information in "data" based on the band ancestor "node".
4538 * In particular, we restrict the dependences in data->local_flow
4539 * to those dependence where the source and the sink occur in
4540 * the same iteration of the given band node.
4541 * We also update data->inner_band_flow to the new value of
4544 static int update_may_persist_at_band(__isl_keep isl_schedule_node
*node
,
4545 struct ppcg_may_persist_data
*data
)
4547 isl_multi_union_pw_aff
*partial
;
4548 isl_union_pw_multi_aff
*contraction
;
4549 isl_union_map
*flow
;
4551 if (isl_schedule_node_band_n_member(node
) == 0)
4554 partial
= isl_schedule_node_band_get_partial_schedule(node
);
4555 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4556 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4558 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4559 isl_union_pw_multi_aff_copy(data
->tagger
));
4561 flow
= data
->local_flow
;
4562 flow
= isl_union_map_eq_at_multi_union_pw_aff(flow
, partial
);
4563 data
->local_flow
= flow
;
4565 isl_union_map_free(data
->inner_band_flow
);
4566 data
->inner_band_flow
= isl_union_map_copy(data
->local_flow
);
4571 /* Given a set of local reaching domain elements "domain",
4572 * expand them to the corresponding leaf domain elements using "contraction"
4573 * and insert the array references tags using data->tagger.
4575 static __isl_give isl_union_set
*expand_and_tag(
4576 __isl_take isl_union_set
*domain
,
4577 __isl_take isl_union_pw_multi_aff
*contraction
,
4578 struct ppcg_may_persist_data
*data
)
4580 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4582 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4583 isl_union_pw_multi_aff_copy(data
->tagger
));
4587 /* Given a filter node that is the child of a set or sequence node,
4588 * restrict data->local_flow to refer only to those elements
4589 * in the filter of the node.
4590 * "contraction" maps the leaf domain elements of the schedule tree
4591 * to the corresponding domain elements at (the parent of) "node".
4593 static int filter_flow(__isl_keep isl_schedule_node
*node
,
4594 struct ppcg_may_persist_data
*data
,
4595 __isl_take isl_union_pw_multi_aff
*contraction
)
4597 isl_union_set
*filter
;
4598 isl_union_map
*flow
;
4600 flow
= data
->local_flow
;
4601 filter
= isl_schedule_node_filter_get_filter(node
);
4602 filter
= expand_and_tag(filter
, contraction
, data
);
4603 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(filter
));
4604 flow
= isl_union_map_intersect_range(flow
, filter
);
4605 data
->local_flow
= flow
;
4610 /* Given a filter node "node", collect the filters on all preceding siblings
4611 * (which are also filter nodes), add them to "filters" and return the result.
4613 static __isl_give isl_union_set
*add_previous_filters(
4614 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4616 isl_schedule_node
*sibling
;
4618 sibling
= isl_schedule_node_copy(node
);
4619 while (sibling
&& isl_schedule_node_has_previous_sibling(sibling
)) {
4620 isl_union_set
*filter
;
4622 sibling
= isl_schedule_node_previous_sibling(sibling
);
4623 filter
= isl_schedule_node_filter_get_filter(sibling
);
4624 filters
= isl_union_set_union(filters
, filter
);
4626 isl_schedule_node_free(sibling
);
4628 return isl_union_set_free(filters
);
4633 /* Given a filter node "node", collect the filters on all following siblings
4634 * (which are also filter nodes), add them to "filters" and return the result.
4636 static __isl_give isl_union_set
*add_next_filters(
4637 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4639 isl_schedule_node
*sibling
;
4641 sibling
= isl_schedule_node_copy(node
);
4642 while (sibling
&& isl_schedule_node_has_next_sibling(sibling
)) {
4643 isl_union_set
*filter
;
4645 sibling
= isl_schedule_node_next_sibling(sibling
);
4646 filter
= isl_schedule_node_filter_get_filter(sibling
);
4647 filters
= isl_union_set_union(filters
, filter
);
4649 isl_schedule_node_free(sibling
);
4651 return isl_union_set_free(filters
);
4656 /* Remove those flow dependences from data->may_persist_flow
4657 * that flow between elements of "domain" within the same iteration
4658 * of all outer band nodes.
4659 * "contraction" maps the leaf domain elements of the schedule tree
4660 * to the corresponding elements "domain".
4662 static void remove_external_flow(struct ppcg_may_persist_data
*data
,
4663 __isl_take isl_union_set
*domain
,
4664 __isl_keep isl_union_pw_multi_aff
*contraction
)
4666 isl_union_map
*flow
;
4668 contraction
= isl_union_pw_multi_aff_copy(contraction
);
4669 domain
= expand_and_tag(domain
, contraction
, data
);
4670 flow
= isl_union_map_copy(data
->local_flow
);
4671 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(domain
));
4672 flow
= isl_union_map_intersect_range(flow
, domain
);
4674 data
->may_persist_flow
= isl_union_map_subtract(data
->may_persist_flow
,
4678 /* Update the information in "data" based on the filter ancestor "node".
4679 * We only need to modify anything if the filter is the child
4680 * of a set or sequence node.
4682 * In the case of a sequence, we remove the dependences between
4683 * statement instances that are both executed either before or
4684 * after the subtree that will be mapped to a kernel, within
4685 * the same iteration of outer bands.
4687 * In both cases, we restrict data->local_flow to the current child.
4689 static int update_may_persist_at_filter(__isl_keep isl_schedule_node
*node
,
4690 struct ppcg_may_persist_data
*data
)
4692 enum isl_schedule_node_type type
;
4693 isl_schedule_node
*parent
;
4695 isl_union_pw_multi_aff
*contraction
;
4696 isl_union_set
*before
, *after
, *filter
;
4697 isl_union_map
*flow
;
4699 type
= isl_schedule_node_get_parent_type(node
);
4700 if (type
!= isl_schedule_node_sequence
&& type
!= isl_schedule_node_set
)
4703 parent
= isl_schedule_node_copy(node
);
4704 parent
= isl_schedule_node_parent(parent
);
4705 contraction
= isl_schedule_node_get_subtree_contraction(parent
);
4706 isl_schedule_node_free(parent
);
4708 if (type
== isl_schedule_node_set
)
4709 return filter_flow(node
, data
, contraction
);
4711 filter
= isl_schedule_node_filter_get_filter(node
);
4712 space
= isl_union_set_get_space(filter
);
4713 isl_union_set_free(filter
);
4714 before
= isl_union_set_empty(space
);
4715 after
= isl_union_set_copy(before
);
4716 before
= add_previous_filters(before
, node
);
4717 after
= add_next_filters(after
, node
);
4719 remove_external_flow(data
, before
, contraction
);
4720 remove_external_flow(data
, after
, contraction
);
4722 return filter_flow(node
, data
, contraction
);
4725 /* Update the information in "data" based on the ancestor "node".
4727 static isl_stat
update_may_persist_at(__isl_keep isl_schedule_node
*node
,
4730 struct ppcg_may_persist_data
*data
= user
;
4732 switch (isl_schedule_node_get_type(node
)) {
4733 case isl_schedule_node_error
:
4734 return isl_stat_error
;
4735 case isl_schedule_node_context
:
4736 case isl_schedule_node_domain
:
4737 case isl_schedule_node_expansion
:
4738 case isl_schedule_node_extension
:
4739 case isl_schedule_node_guard
:
4740 case isl_schedule_node_leaf
:
4741 case isl_schedule_node_mark
:
4742 case isl_schedule_node_sequence
:
4743 case isl_schedule_node_set
:
4745 case isl_schedule_node_band
:
4746 if (update_may_persist_at_band(node
, data
) < 0)
4747 return isl_stat_error
;
4749 case isl_schedule_node_filter
:
4750 if (update_may_persist_at_filter(node
, data
) < 0)
4751 return isl_stat_error
;
4758 /* Determine the set of array elements that may need to be perserved
4759 * by a kernel constructed from the subtree at "node".
4760 * This includes the set of array elements that may need to be preserved
4761 * by the entire scop (prog->may_persist) and the elements for which
4762 * there is a potential flow dependence that may cross a kernel launch.
4764 * To determine the second set, we start from all flow dependences.
4765 * From this set of dependences, we remove those that cannot possibly
4766 * require data to be preserved by a kernel launch.
4767 * In particular, we consider the following sets of dependences.
4768 * - dependences of which the write occurs inside the kernel.
4769 * If the data is needed outside the kernel, then it will
4770 * be copied out immediately after the kernel launch, so there
4771 * is no need for any special care.
4772 * - dependences of which the read occurs inside the kernel and the
4773 * corresponding write occurs inside the same iteration of the
4774 * outer band nodes. This means that the data is needed in
4775 * the first kernel launch after the write, which is already
4776 * taken care of by the standard copy-in. That is, the data
4777 * do not need to be preserved by any intermediate call to
4779 * - dependences of which the write and the read either both occur
4780 * before the kernel launch or both occur after the kernel launch,
4781 * within the same iteration of the outer band nodes with respect
4782 * to the sequence that determines the ordering of the dependence
4783 * and the kernel launch. Such flow dependences cannot cross
4784 * any kernel launch.
4786 * For the remaining (tagged) dependences, we take the domain
4787 * (i.e., the tagged writes) and apply the tagged access relation
4788 * to obtain the accessed data elements.
4789 * These are then combined with the elements that may need to be
4790 * preserved by the entire scop.
4792 static __isl_give isl_union_set
*node_may_persist(
4793 __isl_keep isl_schedule_node
*node
, struct gpu_prog
*prog
)
4795 struct ppcg_may_persist_data data
;
4796 isl_schedule_node
*root
;
4797 isl_union_pw_multi_aff
*contraction
;
4798 isl_union_set
*domain
;
4799 isl_union_set
*persist
;
4800 isl_union_map
*flow
, *local_flow
;
4802 data
.tagger
= prog
->scop
->tagger
;
4804 flow
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
4805 data
.local_flow
= isl_union_map_copy(flow
);
4806 data
.inner_band_flow
= isl_union_map_copy(flow
);
4807 data
.may_persist_flow
= flow
;
4808 if (isl_schedule_node_foreach_ancestor_top_down(node
,
4809 &update_may_persist_at
, &data
) < 0)
4810 data
.may_persist_flow
=
4811 isl_union_map_free(data
.may_persist_flow
);
4812 flow
= data
.may_persist_flow
;
4813 isl_union_map_free(data
.local_flow
);
4815 domain
= isl_schedule_node_get_domain(node
);
4816 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4817 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4819 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4820 isl_union_pw_multi_aff_copy(data
.tagger
));
4821 flow
= isl_union_map_subtract_domain(flow
, isl_union_set_copy(domain
));
4822 local_flow
= data
.inner_band_flow
;
4823 local_flow
= isl_union_map_intersect_range(local_flow
, domain
);
4824 flow
= isl_union_map_subtract(flow
, local_flow
);
4826 persist
= isl_union_map_domain(flow
);
4827 persist
= isl_union_set_apply(persist
,
4828 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4829 persist
= isl_union_set_union(persist
,
4830 isl_union_set_copy(prog
->may_persist
));
4835 /* Add nodes for copying outer arrays in and out of the device
4836 * before and after the subtree "node", which contains one or more kernels.
4837 * "domain" contains the original reaching domain elements before
4838 * the kernels were created, i.e., before the contraction that
4839 * may have been performed in creating the kernels has been applied.
4840 * "prefix" contains the prefix schedule at that point, in terms
4841 * of the same original reaching domain elements.
4843 * We first compute the sets of outer array elements that need
4844 * to be copied in and out and then graft in the nodes for
4845 * performing this copying.
4847 * In particular, for each array that is possibly written anywhere in
4848 * the subtree "node" and that may be used after "node"
4849 * or that may be visible outside the corresponding scop,
4850 * we copy out its entire extent.
4852 * Any array elements that is read without first being written inside
4853 * the subtree "node" needs to be copied in.
4854 * Furthermore, if there are any array elements that
4855 * are copied out, but that may not be written inside "node, then
4856 * they also need to be copied in to ensure that the value after execution
4857 * is the same as the value before execution, at least for those array
4858 * elements that may have their values preserved by the scop or that
4859 * may be written before "node" and read after "node".
4860 * In case the array elements are structures, we need to take into
4861 * account that all members of the structures need to be written
4862 * by "node" before we can avoid copying the data structure in.
4864 * Note that the may_write relation is intersected with the domain,
4865 * which has been intersected with the context.
4866 * This helps in those cases where the arrays are declared with a fixed size,
4867 * while the accesses are parametric and the context assigns a fixed value
4868 * to the parameters.
4870 * If an element from a local array is read without first being written,
4871 * then there is no point in copying it in since it cannot have been
4872 * written prior to the scop. Warn about the uninitialized read instead.
4874 static __isl_give isl_schedule_node
*add_to_from_device(
4875 __isl_take isl_schedule_node
*node
, __isl_take isl_union_set
*domain
,
4876 __isl_take isl_union_map
*prefix
, struct gpu_prog
*prog
)
4878 isl_union_set
*local
;
4879 isl_union_set
*to_device
, *from_device
, *may_persist
;
4880 isl_union_map
*may_write
, *must_write
, *copy_out
, *not_written
;
4881 isl_union_map
*read
, *copy_in
;
4882 isl_union_map
*tagged
;
4883 isl_union_map
*local_uninitialized
;
4884 isl_schedule_node
*graft
;
4886 tagged
= isl_union_map_copy(prog
->scop
->tagged_reads
);
4887 tagged
= isl_union_map_union(tagged
,
4888 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
4890 may_write
= isl_union_map_copy(prog
->may_write
);
4891 may_write
= isl_union_map_intersect_domain(may_write
,
4892 isl_union_set_copy(domain
));
4893 may_write
= remove_local_accesses(prog
,
4894 isl_union_map_copy(tagged
), may_write
,
4895 isl_union_map_copy(prefix
), 0);
4896 may_write
= isl_union_map_apply_range(may_write
,
4897 isl_union_map_copy(prog
->to_outer
));
4898 may_write
= isl_union_map_apply_domain(may_write
,
4899 isl_union_map_copy(prefix
));
4900 may_write
= approximate_copy_out(may_write
, prog
);
4901 copy_out
= isl_union_map_copy(may_write
);
4902 may_write
= isl_union_map_apply_range(may_write
,
4903 isl_union_map_copy(prog
->to_inner
));
4904 must_write
= isl_union_map_copy(prog
->must_write
);
4905 must_write
= isl_union_map_apply_domain(must_write
,
4906 isl_union_map_copy(prefix
));
4907 may_persist
= node_may_persist(node
, prog
);
4908 may_write
= isl_union_map_intersect_range(may_write
, may_persist
);
4909 not_written
= isl_union_map_subtract(may_write
, must_write
);
4911 local
= extract_local_accesses(prog
, domain
);
4912 read
= isl_union_map_copy(prog
->read
);
4913 read
= isl_union_map_intersect_domain(read
, domain
);
4914 read
= remove_local_accesses(prog
, tagged
, read
,
4915 isl_union_map_copy(prefix
), 1);
4916 local
= isl_union_set_apply(local
, isl_union_map_copy(prog
->to_inner
));
4917 local_uninitialized
= isl_union_map_copy(prog
->scop
->live_in
);
4918 local_uninitialized
= isl_union_map_intersect_range(local_uninitialized
,
4920 local_uninitialized
= isl_union_map_intersect(local_uninitialized
,
4921 isl_union_map_copy(read
));
4922 if (!isl_union_map_is_empty(local_uninitialized
)) {
4924 "possibly uninitialized reads (not copied in):\n");
4925 isl_union_map_dump(local_uninitialized
);
4927 read
= isl_union_map_subtract(read
, local_uninitialized
);
4928 read
= isl_union_map_apply_domain(read
, prefix
);
4929 copy_in
= isl_union_map_union(read
, not_written
);
4930 copy_in
= isl_union_map_apply_range(copy_in
,
4931 isl_union_map_copy(prog
->to_outer
));
4933 graft
= create_copy_device(prog
, node
, "to_device",
4934 isl_union_map_range(copy_in
));
4935 node
= isl_schedule_node_graft_before(node
, graft
);
4936 graft
= create_copy_device(prog
, node
, "from_device",
4937 isl_union_map_range(copy_out
));
4938 node
= isl_schedule_node_graft_after(node
, graft
);
4943 /* Update "schedule" for mapping to a GPU device.
4945 * In particular, insert a context node, create kernels for
4946 * each outermost tilable band and introduce node for copying array
4947 * in and out of the device.
4948 * If the child of the initial root points to a set node,
4949 * then children of this node that do not contain any tilable bands
4950 * are separated from the other children and are not mapped to
4953 static __isl_give isl_schedule
*map_to_device(struct gpu_gen
*gen
,
4954 __isl_take isl_schedule
*schedule
)
4956 isl_schedule_node
*node
;
4958 isl_union_set
*domain
;
4959 isl_union_map
*prefix
;
4961 context
= isl_set_copy(gen
->prog
->context
);
4962 context
= isl_set_from_params(context
);
4963 schedule
= isl_schedule_insert_context(schedule
, context
);
4965 node
= isl_schedule_get_root(schedule
);
4966 isl_schedule_free(schedule
);
4967 node
= isl_schedule_node_child(node
, 0);
4968 if (isl_schedule_node_get_type(node
) == isl_schedule_node_context
)
4969 node
= isl_schedule_node_child(node
, 0);
4970 node
= isolate_permutable_subtrees(node
, gen
->prog
);
4971 domain
= isl_schedule_node_get_domain(node
);
4972 prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
4973 node
= mark_kernels(gen
, node
);
4974 node
= add_to_from_device(node
, domain
, prefix
, gen
->prog
);
4975 schedule
= isl_schedule_node_get_schedule(node
);
4976 isl_schedule_node_free(node
);
4981 /* Internal data structure for extract_access.
4982 * "next_access" points to the end of a linked list that is extended
4983 * by extract_access.
4984 * "single_expression" is set if the access expressions belong to
4985 * an expression statement (i.e., a statement without internal control).
4986 * "any_to_outer" maps all intermediate arrays to their outer arrays.
4988 struct ppcg_extract_access_data
{
4989 struct gpu_stmt_access
**next_access
;
4990 int single_expression
;
4991 isl_union_map
*any_to_outer
;
4994 /* Given a tagged access relation to a single array "tagged", extract it
4995 * as a map, taking into account that the input may be empty.
4996 * If the access relation is empty, then it does not contain
4997 * any space information, so we try to recover it from the index
4999 * The space of the index expression is of the form I -> A,
5000 * with I the statement instances and A the array, or [I -> F] -> A,
5001 * with F the filters corresponding to arguments.
5002 * We first drop F, if present, obtaining I -> A.
5003 * Then we construct I -> R, with R the reference tag,
5004 * combine the two into I -> [R -> A] and uncurry to obtain
5005 * the final result [I -> R] -> A.
5006 * Note that the index expression may have a lower dimension
5007 * than that of the array, but this dimension is not used
5008 * if the access relation is empty.
5010 static __isl_give isl_map
*extract_single_tagged_access(
5011 __isl_take isl_union_map
*tagged
, __isl_keep pet_expr
*expr
)
5015 isl_space
*space
, *space2
;
5016 isl_multi_pw_aff
*index
;
5018 empty
= isl_union_map_is_empty(tagged
);
5022 return isl_map_from_union_map(tagged
);
5023 isl_union_map_free(tagged
);
5025 index
= pet_expr_access_get_index(expr
);
5026 space
= isl_multi_pw_aff_get_space(index
);
5027 isl_multi_pw_aff_free(index
);
5028 if (isl_space_domain_is_wrapping(space
))
5029 space
= isl_space_domain_factor_domain(space
);
5030 space2
= isl_space_copy(space
);
5031 space2
= isl_space_from_domain(isl_space_domain(space
));
5032 id
= pet_expr_access_get_ref_id(expr
);
5033 space2
= isl_space_set_tuple_id(space2
, isl_dim_out
, id
);
5034 space
= isl_space_range_product(space2
, space
);
5035 space
= isl_space_uncurry(space
);
5037 return isl_map_empty(space
);
5039 isl_union_map_free(tagged
);
5043 /* Extract a gpu_stmt_access from "expr", append it to the list
5044 * that ends in *data->next_access and update the end of the list.
5045 * If the access expression performs a write, then it is considered
5046 * exact only if it appears in a single expression statement and
5047 * if its may access relation is equal to its must access relation.
5049 * The combined set of may accesses may be union if member accesses
5050 * are involved, but the entire set is derived from a single reference and
5051 * therefore from a single index expression. These accesses therefore
5052 * all map to the same outer array.
5054 static int extract_access(__isl_keep pet_expr
*expr
, void *user
)
5056 struct ppcg_extract_access_data
*data
= user
;
5057 isl_union_map
*tagged
;
5058 struct gpu_stmt_access
*access
;
5059 isl_ctx
*ctx
= pet_expr_get_ctx(expr
);
5060 isl_multi_pw_aff
*index
;
5062 access
= isl_alloc_type(ctx
, struct gpu_stmt_access
);
5064 access
->next
= NULL
;
5065 access
->read
= pet_expr_access_is_read(expr
);
5066 access
->write
= pet_expr_access_is_write(expr
);
5067 tagged
= pet_expr_access_get_tagged_may_read(expr
);
5068 tagged
= isl_union_map_union(tagged
,
5069 pet_expr_access_get_tagged_may_write(expr
));
5070 tagged
= isl_union_map_apply_range(tagged
,
5071 isl_union_map_copy(data
->any_to_outer
));
5072 if (!access
->write
) {
5073 access
->exact_write
= 1;
5074 } else if (!data
->single_expression
) {
5075 access
->exact_write
= 0;
5077 isl_union_map
*must
, *may
;
5078 may
= isl_union_map_copy(tagged
);
5079 may
= isl_union_map_domain_factor_domain(may
);
5080 must
= pet_expr_access_get_must_write(expr
);
5081 access
->exact_write
= isl_union_map_is_equal(must
, may
);
5082 isl_union_map_free(must
);
5083 isl_union_map_free(may
);
5085 index
= pet_expr_access_get_index(expr
);
5086 access
->n_index
= isl_multi_pw_aff_dim(index
, isl_dim_out
);
5087 isl_multi_pw_aff_free(index
);
5088 access
->ref_id
= pet_expr_access_get_ref_id(expr
);
5089 access
->tagged_access
= extract_single_tagged_access(tagged
, expr
);
5090 access
->access
= isl_map_copy(access
->tagged_access
);
5091 access
->access
= isl_map_domain_factor_domain(access
->access
);
5093 *data
->next_access
= access
;
5094 data
->next_access
= &(*data
->next_access
)->next
;
5096 if (!access
->access
)
5102 /* Construct a linked list of gpu_stmt_access objects,
5103 * one for each access expression in the statement body.
5104 * "any_to_outer" maps all intermediate arrays to their outer arrays.
5106 static int pet_stmt_extract_accesses(struct gpu_stmt
*stmt
,
5107 __isl_keep isl_union_map
*any_to_outer
)
5109 struct ppcg_extract_access_data data
;
5111 stmt
->accesses
= NULL
;
5112 data
.next_access
= &stmt
->accesses
;
5113 data
.single_expression
=
5114 pet_tree_get_type(stmt
->stmt
->body
) == pet_tree_expr
;
5115 data
.any_to_outer
= any_to_outer
;
5116 return pet_tree_foreach_access_expr(stmt
->stmt
->body
,
5117 &extract_access
, &data
);
5120 /* Return an array of gpu_stmt representing the statements in "scop".
5122 static struct gpu_stmt
*extract_stmts(isl_ctx
*ctx
, struct ppcg_scop
*scop
,
5123 __isl_keep isl_set
*context
, __isl_keep isl_union_map
*any_to_outer
)
5126 struct gpu_stmt
*stmts
;
5128 stmts
= isl_calloc_array(ctx
, struct gpu_stmt
, scop
->pet
->n_stmt
);
5132 for (i
= 0; i
< scop
->pet
->n_stmt
; ++i
) {
5133 struct gpu_stmt
*s
= &stmts
[i
];
5135 s
->id
= isl_set_get_tuple_id(scop
->pet
->stmts
[i
]->domain
);
5136 s
->stmt
= scop
->pet
->stmts
[i
];
5137 if (pet_stmt_extract_accesses(s
, any_to_outer
) < 0)
5138 return free_stmts(stmts
, i
+ 1);
5144 /* Callback for ppcg_print_guarded that calls the callback for generate_gpu.
5146 static __isl_give isl_printer
*print_gpu(__isl_take isl_printer
*p
, void *user
)
5148 struct gpu_gen
*gen
= user
;
5150 return gen
->print(p
, gen
->prog
, gen
->tree
, &gen
->types
,
5154 /* Generate CUDA code for "scop" and print it to "p".
5155 * After generating an AST for the transformed scop as explained below,
5156 * we call "gen->print" to print the AST in the desired output format
5159 * If it turns out that it does not make sense to generate GPU code,
5160 * then we generate CPU code instead.
5162 * The GPU code is generated in a context where at least one
5163 * statement instance is executed. The corresponding guard (if any) is printed
5164 * around the entire generated GPU code, except for the declaration
5165 * of the arrays that are visible outside of the scop and that therefore
5166 * cannot be declared inside the body of any possible guard.
5168 * We first compute a schedule that respects the dependences
5169 * of the original program and select the outermost bands
5170 * of tilable dimensions that have at least one parallel loop.
5171 * If the --load-schedule is specified, then the loaded schedule
5172 * is used instead of a computed schedule.
5174 * Each of these bands B is then tiled according to "tile" sizes, resulting
5175 * in two nested bands, with a kernel marker on top
5183 * We then split off at most 2 parallel dimensions from the T band and
5184 * at most 3 parallel dimension from the P band
5197 * A filter is introduced in front of T1 that maps the domain instances
5198 * to block identifiers. Similarly, a filter is introduced in front of P1
5199 * that maps the domain instances to thread identifiers.
5201 * For each iteration of the T2 band and for each array, we compute
5202 * the array elements accessed by that iteration, construct a rectangular
5203 * box around it and shift it to the origin. The result is used
5204 * as shared memory for the array.
5206 * Copying and synchronization statements are added to this schedule tree.
5207 * In principle, these are added in front of the P1 band, but some of
5208 * them may get hoisted up to higher levels.
5210 * The entire AST is then generated from the single resulting schedule tree.
5211 * During the generation the subtrees at kernel nodes (K) are saved
5212 * aside and replaced by kernel calls. The result is printed as host code
5213 * while the saved subtrees are printed as device code.
5215 static __isl_give isl_printer
*generate(__isl_take isl_printer
*p
,
5216 struct gpu_gen
*gen
, struct ppcg_scop
*scop
,
5217 struct ppcg_options
*options
)
5219 struct gpu_prog
*prog
;
5221 isl_set
*context
, *guard
;
5222 isl_schedule
*schedule
;
5226 return isl_printer_free(p
);
5228 ctx
= isl_printer_get_ctx(p
);
5229 prog
= gpu_prog_alloc(ctx
, scop
);
5231 return isl_printer_free(p
);
5233 context
= isl_set_copy(prog
->context
);
5234 guard
= isl_union_set_params(isl_union_set_copy(prog
->scop
->domain
));
5235 prog
->context
= isl_set_intersect(prog
->context
, isl_set_copy(guard
));
5238 schedule
= get_schedule(gen
);
5240 any_permutable
= has_any_permutable_node(schedule
);
5241 if (any_permutable
< 0 || !any_permutable
) {
5242 isl_set_free(context
);
5243 isl_set_free(guard
);
5244 if (any_permutable
< 0)
5245 p
= isl_printer_free(p
);
5247 p
= print_cpu(p
, scop
, options
);
5248 isl_schedule_free(schedule
);
5250 schedule
= map_to_device(gen
, schedule
);
5251 gen
->tree
= generate_code(gen
, schedule
);
5252 p
= isl_ast_op_type_print_macro(isl_ast_op_fdiv_q
, p
);
5253 p
= ppcg_print_exposed_declarations(p
, prog
->scop
);
5254 p
= ppcg_print_guarded(p
, guard
, context
, &print_gpu
, gen
);
5255 isl_ast_node_free(gen
->tree
);
5258 gpu_prog_free(prog
);
5263 /* Wrapper around generate for use as a ppcg_transform callback.
5265 static __isl_give isl_printer
*generate_wrap(__isl_take isl_printer
*p
,
5266 struct ppcg_scop
*scop
, void *user
)
5268 struct gpu_gen
*gen
= user
;
5270 return generate(p
, gen
, scop
, gen
->options
);
5273 /* Transform the code in the file called "input" by replacing
5274 * all scops by corresponding GPU code and write the results to "out".
5276 int generate_gpu(isl_ctx
*ctx
, const char *input
, FILE *out
,
5277 struct ppcg_options
*options
,
5278 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
5279 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
5280 struct gpu_types
*types
, void *user
), void *user
)
5287 gen
.sizes
= extract_sizes_from_str(ctx
, options
->sizes
);
5288 gen
.options
= options
;
5291 gen
.print_user
= user
;
5293 gen
.types
.name
= NULL
;
5295 if (options
->debug
->dump_sizes
) {
5296 isl_space
*space
= isl_space_params_alloc(ctx
, 0);
5297 gen
.used_sizes
= isl_union_map_empty(space
);
5300 r
= ppcg_transform(ctx
, input
, out
, options
, &generate_wrap
, &gen
);
5302 if (options
->debug
->dump_sizes
) {
5303 isl_union_map_dump(gen
.used_sizes
);
5304 isl_union_map_free(gen
.used_sizes
);
5307 isl_union_map_free(gen
.sizes
);
5308 for (i
= 0; i
< gen
.types
.n
; ++i
)
5309 free(gen
.types
.name
[i
]);
5310 free(gen
.types
.name
);
5315 /* Compute the set of inner array elements that may have their values
5316 * preserved by "prog". In particular, collect the array elements of
5317 * arrays that are not local to "prog" and remove those elements that
5318 * are definitely killed or definitely written by "prog".
5320 static __isl_give isl_union_set
*compute_may_persist(struct gpu_prog
*prog
)
5323 isl_union_set
*may_persist
, *killed
;
5324 isl_union_map
*must_kill
;
5326 may_persist
= isl_union_set_empty(isl_set_get_space(prog
->context
));
5327 for (i
= 0; i
< prog
->n_array
; ++i
) {
5330 if (prog
->array
[i
].local
)
5333 extent
= isl_set_copy(prog
->array
[i
].extent
);
5334 may_persist
= isl_union_set_add_set(may_persist
, extent
);
5337 may_persist
= isl_union_set_intersect_params(may_persist
,
5338 isl_set_copy(prog
->context
));
5339 may_persist
= isl_union_set_apply(may_persist
,
5340 isl_union_map_copy(prog
->to_inner
));
5341 must_kill
= isl_union_map_copy(prog
->tagged_must_kill
);
5342 killed
= isl_union_map_range(must_kill
);
5343 must_kill
= isl_union_map_copy(prog
->must_write
);
5344 killed
= isl_union_set_union(killed
, isl_union_map_range(must_kill
));
5346 may_persist
= isl_union_set_subtract(may_persist
, killed
);
5350 struct gpu_prog
*gpu_prog_alloc(isl_ctx
*ctx
, struct ppcg_scop
*scop
)
5352 struct gpu_prog
*prog
;
5359 prog
= isl_calloc_type(ctx
, struct gpu_prog
);
5364 prog
->context
= isl_set_copy(scop
->context
);
5365 prog
->n_stmts
= scop
->pet
->n_stmt
;
5366 prog
->any_to_outer
= pet_scop_compute_outer_to_any(scop
->pet
);
5367 prog
->any_to_outer
= isl_union_map_reverse(prog
->any_to_outer
);
5368 space
= isl_union_map_get_space(prog
->any_to_outer
);
5369 space
= isl_space_set_from_params(space
);
5370 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
5371 space
= isl_space_map_from_set(space
);
5372 id
= isl_map_identity(space
);
5373 prog
->any_to_outer
= isl_union_map_add_map(prog
->any_to_outer
, id
);
5374 prog
->stmts
= extract_stmts(ctx
, scop
,
5375 prog
->context
, prog
->any_to_outer
);
5376 prog
->read
= isl_union_map_copy(scop
->reads
);
5377 prog
->may_write
= isl_union_map_copy(scop
->may_writes
);
5378 prog
->must_write
= isl_union_map_copy(scop
->must_writes
);
5379 prog
->tagged_must_kill
= isl_union_map_copy(scop
->tagged_must_kills
);
5380 prog
->to_inner
= pet_scop_compute_outer_to_inner(scop
->pet
);
5381 prog
->to_outer
= isl_union_map_copy(prog
->to_inner
);
5382 prog
->to_outer
= isl_union_map_reverse(prog
->to_outer
);
5385 return gpu_prog_free(prog
);
5387 if (collect_array_info(prog
) < 0)
5388 return gpu_prog_free(prog
);
5389 prog
->may_persist
= compute_may_persist(prog
);
5394 void *gpu_prog_free(struct gpu_prog
*prog
)
5398 free_array_info(prog
);
5399 free_stmts(prog
->stmts
, prog
->n_stmts
);
5400 isl_union_map_free(prog
->any_to_outer
);
5401 isl_union_map_free(prog
->to_outer
);
5402 isl_union_map_free(prog
->to_inner
);
5403 isl_union_map_free(prog
->read
);
5404 isl_union_map_free(prog
->may_write
);
5405 isl_union_map_free(prog
->must_write
);
5406 isl_union_map_free(prog
->tagged_must_kill
);
5407 isl_union_map_free(prog
->array_order
);
5408 isl_union_set_free(prog
->may_persist
);
5409 isl_set_free(prog
->context
);