2 * Copyright 2010-2011 INRIA Saclay
3 * Copyright 2012-2013 Ecole Normale Superieure
4 * Copyright 2015-2016 Sven Verdoolaege
6 * Use of this software is governed by the MIT license
8 * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
9 * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
11 * 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"
31 #include "gpu_hybrid.h"
35 #include "ppcg_options.h"
39 struct gpu_array_info
;
41 /* Return the name of the outer array (of structs) accessed by "access".
43 static const char *get_outer_array_name(__isl_keep isl_map
*access
)
48 space
= isl_space_range(isl_map_get_space(access
));
49 while (space
&& isl_space_is_wrapping(space
))
50 space
= isl_space_domain(isl_space_unwrap(space
));
51 name
= isl_space_get_tuple_name(space
, isl_dim_set
);
52 isl_space_free(space
);
57 /* Collect all references to the given array and store pointers to them
60 static isl_stat
collect_references(struct gpu_prog
*prog
,
61 struct gpu_array_info
*array
)
67 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
68 struct gpu_stmt
*stmt
= &prog
->stmts
[i
];
69 struct gpu_stmt_access
*access
;
71 for (access
= stmt
->accesses
; access
; access
= access
->next
) {
73 name
= get_outer_array_name(access
->access
);
74 if (name
&& !strcmp(array
->name
, name
))
79 array
->refs
= isl_alloc_array(prog
->ctx
, struct gpu_stmt_access
*, n
);
81 return isl_stat_error
;
85 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
86 struct gpu_stmt
*stmt
= &prog
->stmts
[i
];
87 struct gpu_stmt_access
*access
;
89 for (access
= stmt
->accesses
; access
; access
= access
->next
) {
91 name
= get_outer_array_name(access
->access
);
92 if (!name
|| strcmp(array
->name
, name
))
95 array
->refs
[n
++] = access
;
102 /* Compute and return the extent of "array", taking into account the set of
105 * In particular, the extent in the outer dimension is taken
106 * from "accessed", while the extents in the remaining dimensions
107 * are taken from array->extent.
109 * The extent in the outer dimension cannot be taken from array->extent
110 * because that may be unbounded. Furthermore, even if it is bounded,
111 * it may be larger than the piece of the array that is being accessed.
113 static __isl_give isl_set
*compute_extent(struct pet_array
*array
,
114 __isl_keep isl_set
*accessed
)
121 extent
= isl_set_copy(array
->extent
);
123 n_index
= isl_set_dim(accessed
, isl_dim_set
);
127 extent
= isl_set_project_out(extent
, isl_dim_set
, 0, 1);
128 outer
= isl_set_copy(accessed
);
129 outer
= isl_set_project_out(outer
, isl_dim_set
, 1, n_index
- 1);
130 extent
= isl_set_flat_product(outer
, extent
);
131 id
= isl_set_get_tuple_id(accessed
);
132 extent
= isl_set_set_tuple_id(extent
, id
);
137 /* Is the array "array" being extracted a read-only scalar?
139 * That is, is "array" a scalar that is never possibly written to.
140 * An array containing structures is never considered to be a scalar.
142 static int is_read_only_scalar(struct gpu_array_info
*array
,
143 struct gpu_prog
*prog
)
146 isl_union_map
*write
;
149 if (array
->has_compound_element
)
151 if (array
->n_index
!= 0)
154 write
= isl_union_map_copy(prog
->may_write
);
155 space
= isl_set_universe(isl_space_copy(array
->space
));
156 write
= isl_union_map_intersect_range(write
,
157 isl_union_set_from_set(space
));
158 empty
= isl_union_map_is_empty(write
);
159 isl_union_map_free(write
);
164 /* Is "array" only accessed as individual, fixed elements?
165 * That is, does each access to "array" access a single, fixed element?
167 static isl_bool
only_fixed_element_accessed(struct gpu_array_info
*array
)
171 for (i
= 0; i
< array
->n_ref
; ++i
)
172 if (!array
->refs
[i
]->fixed_element
)
173 return isl_bool_false
;
175 return isl_bool_true
;
178 /* Compute bounds on the host array "pa" based on the corresponding
179 * accessed elements in "arrays"
180 * and collect all references to the array.
181 * Store the results in "info".
183 * If the array is zero-dimensional and does not contain structures,
184 * i.e., if the array is a scalar, we check whether it is read-only.
185 * We also check whether the array is accessed at all.
187 static isl_stat
extract_array_info(struct gpu_prog
*prog
,
188 struct gpu_array_info
*info
, struct pet_array
*pa
,
189 __isl_keep isl_union_set
*arrays
)
194 isl_multi_pw_aff
*bounds
;
195 isl_set
*accessed
, *extent
;
197 n_index
= isl_set_dim(pa
->extent
, isl_dim_set
);
198 name
= isl_set_get_tuple_name(pa
->extent
);
200 info
->space
= isl_set_get_space(pa
->extent
);
201 info
->name
= strdup(name
);
202 info
->n_index
= n_index
;
203 info
->linearize
= prog
->scop
->options
->linearize_device_arrays
;
205 info
->type
= strdup(pa
->element_type
);
206 info
->size
= pa
->element_size
;
207 info
->local
= pa
->declared
&& !pa
->exposed
;
208 info
->has_compound_element
= pa
->element_is_record
;
209 info
->read_only_scalar
= is_read_only_scalar(info
, prog
);
211 info
->declared_extent
= isl_set_copy(pa
->extent
);
212 accessed
= isl_union_set_extract_set(arrays
,
213 isl_space_copy(info
->space
));
214 empty
= isl_set_is_empty(accessed
);
215 extent
= compute_extent(pa
, accessed
);
216 isl_set_free(accessed
);
217 info
->extent
= extent
;
219 return isl_stat_error
;
220 info
->accessed
= !empty
;
221 bounds
= ppcg_size_from_extent(isl_set_copy(extent
));
222 bounds
= isl_multi_pw_aff_gist(bounds
, isl_set_copy(prog
->context
));
224 return isl_stat_error
;
225 if (!isl_multi_pw_aff_is_cst(bounds
))
227 info
->bound
= bounds
;
229 if (collect_references(prog
, info
) < 0)
230 return isl_stat_error
;
231 info
->only_fixed_element
= only_fixed_element_accessed(info
);
236 /* Remove independence from the order constraints "order" on array "array".
237 * Since the pairs of iterations in the filter relation of an independence
238 * are guaranteed to be completely independent by the user, there is
239 * no need to ensure that live ranges are ordered along those pairs.
240 * We make an exception for local variables, though, as the independence
241 * guarantee does not apply to those.
243 * The order constraints are used in two places.
244 * Those on scalars are used in check_scalar_live_ranges to check if
245 * we need to force the scalar to be private. Any non-local scalar
246 * should not be forced scalar if it only appears in independent loops.
247 * Those on non-scalars are added to the coincidence constraints
248 * in compute_schedule because we do not support any array expansion.
249 * Accesses to non-local arrays should not prevent a loop from being
250 * considered coincident so we should indeed remove those constraints
251 * from the order constraints.
253 static __isl_give isl_union_map
*remove_independences(struct gpu_prog
*prog
,
254 struct gpu_array_info
*array
, __isl_take isl_union_map
*order
)
258 for (i
= 0; i
< prog
->scop
->pet
->n_independence
; ++i
) {
259 struct pet_independence
*pi
= prog
->scop
->pet
->independences
[i
];
260 if (isl_union_set_contains(pi
->local
, array
->space
))
263 order
= isl_union_map_subtract(order
,
264 isl_union_map_copy(pi
->filter
));
270 /* For each array in "prog", store the (untagged) order dependences
271 * derived from the array in array->dep_order.
272 * In particular, consider all references that access the given array
273 * and take the order dependences that have one of these references
274 * as source. (Since an order dependence relates two references to
275 * the same array, the target of these order dependences will also
276 * be one of these references.)
277 * Additionally, store the union of these array->dep_order relations
278 * for all arrays that cannot be mapped to private memory in prog->array_order.
280 void collect_order_dependences(struct gpu_prog
*prog
)
284 isl_union_map
*accesses
;
286 space
= isl_union_map_get_space(prog
->read
);
287 prog
->array_order
= isl_union_map_empty(space
);
289 accesses
= isl_union_map_copy(prog
->scop
->tagged_reads
);
290 accesses
= isl_union_map_union(accesses
,
291 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
292 accesses
= isl_union_map_universe(accesses
);
293 accesses
= isl_union_map_apply_range(accesses
,
294 isl_union_map_copy(prog
->to_outer
));
296 for (i
= 0; i
< prog
->n_array
; ++i
) {
297 struct gpu_array_info
*array
= &prog
->array
[i
];
300 isl_union_map
*order
;
302 set
= isl_set_universe(isl_space_copy(array
->space
));
303 uset
= isl_union_set_from_set(set
);
304 uset
= isl_union_map_domain(
305 isl_union_map_intersect_range(isl_union_map_copy(accesses
),
307 order
= isl_union_map_copy(prog
->scop
->tagged_dep_order
);
308 order
= isl_union_map_intersect_domain(order
, uset
);
309 order
= isl_union_map_zip(order
);
310 order
= isl_union_set_unwrap(isl_union_map_domain(order
));
311 order
= remove_independences(prog
, array
, order
);
312 array
->dep_order
= order
;
314 if (gpu_array_can_be_private(array
))
317 prog
->array_order
= isl_union_map_union(prog
->array_order
,
318 isl_union_map_copy(array
->dep_order
));
321 isl_union_map_free(accesses
);
324 /* Construct a gpu_array_info for each array referenced by prog->scop and
325 * collect them in prog->array.
327 * The sizes are based on the extents and the set of possibly accessed
328 * elements by "prog".
329 * If there are any member accesses involved, then they are first mapped
330 * to the outer arrays of structs.
331 * Only extract gpu_array_info entries for these outer arrays.
333 * If we are allowing live range reordering, then also set
334 * the dep_order field. Otherwise leave it NULL.
336 static isl_stat
collect_array_info(struct gpu_prog
*prog
)
339 isl_stat r
= isl_stat_ok
;
340 isl_union_set
*arrays
;
343 prog
->array
= isl_calloc_array(prog
->ctx
,
344 struct gpu_array_info
, prog
->scop
->pet
->n_array
);
346 return isl_stat_error
;
348 arrays
= isl_union_map_range(isl_union_map_copy(prog
->read
));
349 arrays
= isl_union_set_union(arrays
,
350 isl_union_map_range(isl_union_map_copy(prog
->may_write
)));
352 arrays
= isl_union_set_apply(arrays
,
353 isl_union_map_copy(prog
->to_outer
));
355 arrays
= isl_union_set_coalesce(arrays
);
357 for (i
= 0; i
< prog
->scop
->pet
->n_array
; ++i
) {
360 field
= isl_set_is_wrapping(prog
->scop
->pet
->arrays
[i
]->extent
);
365 if (extract_array_info(prog
, &prog
->array
[prog
->n_array
++],
366 prog
->scop
->pet
->arrays
[i
], arrays
) < 0)
369 if (i
< prog
->scop
->pet
->n_array
)
372 isl_union_set_free(arrays
);
374 if (prog
->scop
->options
->live_range_reordering
)
375 collect_order_dependences(prog
);
380 static void free_array_info(struct gpu_prog
*prog
)
384 for (i
= 0; i
< prog
->n_array
; ++i
) {
385 free(prog
->array
[i
].type
);
386 free(prog
->array
[i
].name
);
387 isl_multi_pw_aff_free(prog
->array
[i
].bound
);
388 isl_ast_expr_free(prog
->array
[i
].bound_expr
);
389 isl_space_free(prog
->array
[i
].space
);
390 isl_set_free(prog
->array
[i
].declared_extent
);
391 isl_set_free(prog
->array
[i
].extent
);
392 isl_ast_expr_free(prog
->array
[i
].declared_size
);
393 free(prog
->array
[i
].refs
);
394 isl_union_map_free(prog
->array
[i
].dep_order
);
399 /* Check if a gpu array is a scalar. A scalar is a value that is not stored
400 * as an array or through a pointer reference, but as a single data element.
401 * At the moment, scalars are represented as zero-dimensional arrays.
402 * Note that the single data element may be an entire structure.
404 int gpu_array_is_scalar(struct gpu_array_info
*array
)
406 return array
->n_index
== 0;
409 /* Can "array" be mapped to private memory?
410 * That is, is it only accessed as individual elements with
411 * constant index expressions?
413 isl_bool
gpu_array_can_be_private(struct gpu_array_info
*array
)
416 return isl_bool_error
;
417 return array
->only_fixed_element
;
420 /* Is "array" a read-only scalar?
422 int gpu_array_is_read_only_scalar(struct gpu_array_info
*array
)
424 return array
->read_only_scalar
;
427 /* Does "array" need to be allocated on the device?
428 * If it is a read-only scalar, then it will be passed as an argument
429 * to the kernel and therefore does not require any allocation.
430 * If this device memory is not accessed at all, then it does not
431 * need to be allocated either.
433 int gpu_array_requires_device_allocation(struct gpu_array_info
*array
)
435 if (gpu_array_is_read_only_scalar(array
))
442 /* Return the set of parameter values for which the array has a positive
443 * size in all dimensions.
444 * If the sizes are only valid for some parameter values, then those
445 * constraints are also taken into account.
447 __isl_give isl_set
*gpu_array_positive_size_guard(struct gpu_array_info
*array
)
456 space
= isl_space_params(isl_space_copy(array
->space
));
457 guard
= isl_set_universe(space
);
459 for (i
= 0; i
< array
->n_index
; ++i
) {
461 isl_set
*guard_i
, *zero
;
463 bound
= isl_multi_pw_aff_get_pw_aff(array
->bound
, i
);
464 guard_i
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(bound
));
465 zero
= isl_pw_aff_zero_set(bound
);
466 guard_i
= isl_set_subtract(guard_i
, zero
);
467 guard
= isl_set_intersect(guard
, guard_i
);
473 /* Internal data structure for extract_size_of_type.
474 * "type" specifies the name of the space that we want to extract.
475 * "res" is used to store the subset of that space.
477 struct ppcg_extract_size_data
{
482 /* This function is called for each set in a union_set.
483 * If the name of the set matches data->type, we store the
486 static isl_stat
extract_size_of_type(__isl_take isl_set
*size
, void *user
)
488 struct ppcg_extract_size_data
*data
= user
;
491 name
= isl_set_get_tuple_name(size
);
492 if (name
&& !strcmp(name
, data
->type
)) {
494 return isl_stat_error
;
501 /* Given a union map { kernel[i] -> *[...] },
502 * return the range in the space called "type" for the kernel with
503 * sequence number "id".
505 static __isl_give isl_set
*extract_sizes(__isl_keep isl_union_map
*sizes
,
506 const char *type
, int id
)
510 isl_union_set
*local_sizes
;
511 struct ppcg_extract_size_data data
= { type
, NULL
};
516 space
= isl_union_map_get_space(sizes
);
517 space
= isl_space_set_from_params(space
);
518 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
519 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
520 dom
= isl_set_universe(space
);
521 dom
= isl_set_fix_si(dom
, isl_dim_set
, 0, id
);
523 local_sizes
= isl_union_set_apply(isl_union_set_from_set(dom
),
524 isl_union_map_copy(sizes
));
525 isl_union_set_foreach_set(local_sizes
, &extract_size_of_type
, &data
);
526 isl_union_set_free(local_sizes
);
530 /* Given a singleton set, extract the first (at most *len) elements
531 * of the single integer tuple into *sizes and update *len if needed.
533 * If "set" is NULL, then the "sizes" array is not updated.
535 static isl_stat
read_sizes_from_set(__isl_take isl_set
*set
, int *sizes
,
544 dim
= isl_set_dim(set
, isl_dim_set
);
548 for (i
= 0; i
< *len
; ++i
) {
551 v
= isl_set_plain_get_val_if_fixed(set
, isl_dim_set
, i
);
554 sizes
[i
] = isl_val_get_num_si(v
);
562 return isl_stat_error
;
565 /* Add the map { kernel[id] -> type[sizes] } to gen->used_sizes,
566 * if the option debug->dump_sizes is set.
568 static void set_used_sizes(struct gpu_gen
*gen
, const char *type
, int id
,
575 if (!gen
->options
->debug
->dump_sizes
)
578 space
= isl_union_map_get_space(gen
->used_sizes
);
579 space
= isl_space_set_from_params(space
);
580 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
581 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
582 space
= isl_space_from_domain(space
);
583 space
= isl_space_add_dims(space
, isl_dim_out
, len
);
584 space
= isl_space_set_tuple_name(space
, isl_dim_out
, type
);
586 map
= isl_map_universe(space
);
587 map
= isl_map_fix_si(map
, isl_dim_in
, 0, id
);
588 for (i
= 0; i
< len
; ++i
)
589 map
= isl_map_fix_si(map
, isl_dim_out
, i
, sizes
[i
]);
591 gen
->used_sizes
= isl_union_map_add_map(gen
->used_sizes
, map
);
594 /* Extract user specified "tile" sizes from the "sizes" command line option,
595 * defaulting to option->tile_size in each dimension.
596 * *tile_len contains the maximum number of tile sizes needed.
597 * Update *tile_len to the number of specified tile sizes, if any, and
598 * return a pointer to the tile sizes (or NULL on error).
599 * Add the effectively used sizes to gen->used_sizes.
601 static int *read_tile_sizes(struct gpu_gen
*gen
, int *tile_len
)
607 tile_size
= isl_alloc_array(gen
->ctx
, int, *tile_len
);
610 for (n
= 0; n
< *tile_len
; ++n
)
611 tile_size
[n
] = gen
->options
->tile_size
;
613 size
= extract_sizes(gen
->sizes
, "tile", gen
->kernel_id
);
614 if (read_sizes_from_set(size
, tile_size
, tile_len
) < 0)
616 set_used_sizes(gen
, "tile", gen
->kernel_id
, tile_size
, *tile_len
);
624 /* Extract user specified "block" sizes from the "sizes" command line option,
625 * after filling in some potentially useful defaults.
627 static isl_stat
read_block_sizes(struct ppcg_kernel
*kernel
,
628 __isl_keep isl_union_map
*sizes
)
632 if (kernel
->n_block
> 3)
634 switch (kernel
->n_block
) {
636 kernel
->block_dim
[0] = 512;
639 kernel
->block_dim
[0] = 32;
640 kernel
->block_dim
[1] = 16;
643 kernel
->block_dim
[0] = 32;
644 kernel
->block_dim
[1] = 4;
645 kernel
->block_dim
[2] = 4;
649 size
= extract_sizes(sizes
, "block", kernel
->id
);
650 return read_sizes_from_set(size
, kernel
->block_dim
, &kernel
->n_block
);
653 /* Extract user specified "grid" sizes from the "sizes" command line option,
654 * after filling in some potentially useful defaults.
656 static isl_stat
read_grid_sizes(struct ppcg_kernel
*kernel
,
657 __isl_keep isl_union_map
*sizes
)
661 if (kernel
->n_grid
> 2)
663 switch (kernel
->n_grid
) {
665 kernel
->grid_dim
[0] = 32768;
668 kernel
->grid_dim
[0] = 256;
669 kernel
->grid_dim
[1] = 256;
673 size
= extract_sizes(sizes
, "grid", kernel
->id
);
674 return read_sizes_from_set(size
, kernel
->grid_dim
, &kernel
->n_grid
);
677 /* Extract user specified grid and block sizes from the gen->sizes
678 * command line option after filling in some potentially useful defaults.
679 * Store the extracted sizes in "kernel".
680 * Add the effectively used sizes to gen->used_sizes.
682 static isl_stat
read_grid_and_block_sizes(struct ppcg_kernel
*kernel
,
685 if (read_block_sizes(kernel
, gen
->sizes
) < 0)
686 return isl_stat_error
;
687 if (read_grid_sizes(kernel
, gen
->sizes
) < 0)
688 return isl_stat_error
;
689 set_used_sizes(gen
, "block", kernel
->id
,
690 kernel
->block_dim
, kernel
->n_block
);
691 set_used_sizes(gen
, "grid", kernel
->id
,
692 kernel
->grid_dim
, kernel
->n_grid
);
696 static void *free_stmts(struct gpu_stmt
*stmts
, int n
)
703 for (i
= 0; i
< n
; ++i
) {
704 struct gpu_stmt_access
*access
, *next
;
706 for (access
= stmts
[i
].accesses
; access
; access
= next
) {
708 isl_id_free(access
->ref_id
);
709 isl_map_free(access
->access
);
710 isl_map_free(access
->tagged_access
);
714 isl_id_free(stmts
[i
].id
);
721 /* Add parameters p[i] with identifiers "ids" to "set",
722 * with bounds to 0 <= p[i] < size[i].
724 __isl_give isl_set
*add_bounded_parameters(__isl_take isl_set
*set
,
725 int *size
, __isl_keep isl_id_list
*ids
)
730 len
= isl_id_list_n_id(ids
);
731 nparam
= isl_set_dim(set
, isl_dim_param
);
732 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
734 for (i
= 0; i
< len
; ++i
) {
737 id
= isl_id_list_get_id(ids
, i
);
738 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
739 set
= isl_set_lower_bound_si(set
, isl_dim_param
, nparam
+ i
, 0);
740 set
= isl_set_upper_bound_si(set
, isl_dim_param
,
741 nparam
+ i
, size
[i
] - 1);
747 /* Add "len" parameters p[i] with identifiers "ids" and intersect "set"
750 * { : 0 <= p[i] < size[i] }
752 * or an overapproximation.
754 static __isl_give isl_set
*add_bounded_parameters_dynamic(
755 __isl_take isl_set
*set
, __isl_keep isl_multi_pw_aff
*size
,
756 __isl_keep isl_id_list
*ids
)
763 len
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
764 nparam
= isl_set_dim(set
, isl_dim_param
);
765 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
767 for (i
= 0; i
< len
; ++i
) {
770 id
= isl_id_list_get_id(ids
, i
);
771 set
= isl_set_set_dim_id(set
, isl_dim_param
, nparam
+ i
, id
);
774 space
= isl_space_params(isl_set_get_space(set
));
775 ls
= isl_local_space_from_space(space
);
776 for (i
= 0; i
< len
; ++i
) {
777 isl_pw_aff
*param
, *size_i
, *zero
;
780 param
= isl_pw_aff_var_on_domain(isl_local_space_copy(ls
),
781 isl_dim_param
, nparam
+ i
);
783 size_i
= isl_multi_pw_aff_get_pw_aff(size
, i
);
784 bound
= isl_pw_aff_lt_set(isl_pw_aff_copy(param
), size_i
);
785 bound
= isl_set_from_basic_set(isl_set_simple_hull(bound
));
786 set
= isl_set_intersect_params(set
, bound
);
788 zero
= isl_pw_aff_zero_on_domain(isl_local_space_copy(ls
));
789 bound
= isl_pw_aff_ge_set(param
, zero
);
790 set
= isl_set_intersect_params(set
, bound
);
792 isl_local_space_free(ls
);
797 /* Return the union of all tagged access relations in the group.
799 static __isl_give isl_union_map
*group_tagged_access_relation(
800 struct gpu_array_ref_group
*group
)
803 isl_union_map
*access
;
805 access
= isl_union_map_empty(isl_map_get_space(group
->access
));
806 for (i
= 0; i
< group
->n_ref
; ++i
) {
809 map_i
= isl_map_copy(group
->refs
[i
]->tagged_access
);
810 access
= isl_union_map_union(access
,
811 isl_union_map_from_map(map_i
));
817 /* Return the extent of "array", recomputed from the bounds.
818 * The recomputed extent may be simpler than the original extent.
820 static __isl_give isl_set
*array_extent(struct gpu_array_info
*array
)
828 id
= isl_set_get_tuple_id(array
->extent
);
829 space
= isl_set_get_space(array
->extent
);
830 extent
= isl_set_universe(isl_space_copy(space
));
831 ls
= isl_local_space_from_space(space
);
832 for (i
= 0; i
< array
->n_index
; ++i
) {
838 extent
= isl_set_lower_bound_si(extent
, isl_dim_set
, i
, 0);
840 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
842 index
= isl_pw_aff_from_aff(aff
);
843 bound
= isl_multi_pw_aff_get_pw_aff(array
->bound
, i
);
844 bound
= isl_pw_aff_from_range(bound
);
845 bound
= isl_pw_aff_add_dims(bound
, isl_dim_in
, array
->n_index
);
846 bound
= isl_pw_aff_set_tuple_id(bound
, isl_dim_in
,
848 lt
= isl_pw_aff_lt_set(index
, bound
);
849 extent
= isl_set_intersect(extent
, lt
);
851 isl_local_space_free(ls
);
857 /* Return a map from the first group->shared_tile->depth dimensions
858 * of the computed schedule to the array tile in
859 * global memory that corresponds to the shared memory copy.
861 * In particular, return a map
867 * tile_offset(i) <= a <= tile_offset(i) + tile_size - 1 (1)
871 * 0 <= a <= array_size - 1 (2)
873 * Note that if some stride has been detected (i.e., when
874 * group->shared_tile->bound[i].shift is set), then a in (1) refers
875 * to the shifted and scaled down version.
877 * Constraints (1) are obtained by mapping the size constraints on the
878 * shared/private memory tile back to the access relation.
879 * Constraints (2) are obtained from the (recomputed) extent.
881 static __isl_give isl_map
*group_tile(struct gpu_array_ref_group
*group
)
884 int n_index
= group
->array
->n_index
;
890 space
= isl_multi_aff_get_space(group
->shared_tile
->tiling
);
891 space
= isl_space_range(space
);
892 local
= isl_set_universe(space
);
893 for (i
= 0; i
< n_index
; ++i
) {
896 local
= isl_set_lower_bound_si(local
, isl_dim_set
, i
, 0);
897 bound
= isl_val_copy(group
->shared_tile
->bound
[i
].size
);
898 bound
= isl_val_sub_ui(bound
, 1);
899 local
= isl_set_upper_bound_val(local
, isl_dim_set
, i
, bound
);
901 local
= isl_set_preimage_multi_aff(local
,
902 isl_multi_aff_copy(group
->shared_tile
->tiling
));
903 tile
= isl_set_unwrap(local
);
904 extent
= array_extent(group
->array
);
905 tile
= isl_map_intersect_range(tile
, extent
);
910 /* Given a mapping "iterator_map" from the AST schedule to a domain,
911 * return the corresponding mapping from the AST schedule
912 * to the outer kernel->copy_schedule_dim dimensions of
913 * the schedule computed by PPCG for this kernel.
915 * Note that kernel->copy_schedule_dim is at least as large as
916 * the largest depth of any array reference group associated to the kernel.
917 * This is needed as the returned schedule is used to extract a mapping
918 * to the outer tile->depth dimensions in transform_index.
920 static __isl_give isl_pw_multi_aff
*compute_sched_to_copy(
921 struct ppcg_kernel
*kernel
, __isl_take isl_pw_multi_aff
*iterator_map
)
923 isl_union_pw_multi_aff
*upma
;
924 isl_pw_multi_aff
*pma
;
927 space
= isl_space_range(isl_pw_multi_aff_get_space(iterator_map
));
928 space
= isl_space_from_domain(space
);
929 space
= isl_space_add_dims(space
, isl_dim_out
,
930 kernel
->copy_schedule_dim
);
932 upma
= isl_union_pw_multi_aff_copy(kernel
->copy_schedule
);
933 pma
= isl_union_pw_multi_aff_extract_pw_multi_aff(upma
, space
);
934 isl_union_pw_multi_aff_free(upma
);
936 return isl_pw_multi_aff_pullback_pw_multi_aff(pma
, iterator_map
);
939 /* If max_shared_memory is not set to infinity (-1), then make
940 * sure that the total amount of shared memory required by the
941 * array reference groups mapped to shared memory by "kernel"
942 * is no larger than this maximum.
944 * We apply a greedy approach and discard (keep in global memory)
945 * those groups that would result in a total memory size that
946 * is larger than the maximum.
948 * This function should be called after any function that may
949 * affect the decision on whether to place a reference group
950 * in private, shared or global memory.
952 static void check_shared_memory_bound(struct ppcg_kernel
*kernel
)
955 isl_val
*left
, *size
;
957 if (kernel
->options
->max_shared_memory
< 0)
960 left
= isl_val_int_from_si(kernel
->ctx
,
961 kernel
->options
->max_shared_memory
);
963 for (i
= 0; i
< kernel
->n_array
; ++i
) {
964 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
966 for (j
= 0; j
< local
->n_group
; ++j
) {
967 struct gpu_array_ref_group
*group
;
968 enum ppcg_group_access_type type
;
970 group
= local
->groups
[j
];
971 type
= gpu_array_ref_group_type(group
);
972 if (type
!= ppcg_access_shared
)
975 size
= gpu_array_tile_size(group
->shared_tile
);
976 size
= isl_val_mul_ui(size
, local
->array
->size
);
978 if (isl_val_le(size
, left
)) {
979 left
= isl_val_sub(left
, size
);
985 gpu_array_tile_free(group
->shared_tile
);
992 /* Mark all arrays of "kernel" that have an array reference group
993 * that is not mapped to private or shared memory as
994 * accessing the corresponding global device memory.
996 static void mark_global_arrays(struct ppcg_kernel
*kernel
)
1000 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1001 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
1005 for (j
= 0; j
< local
->n_group
; ++j
) {
1006 if (gpu_array_ref_group_tile(local
->groups
[j
]))
1010 local
->array
->global
= 1;
1016 /* Compute a tiling for all the array reference groups in "kernel".
1018 static void compute_group_tilings(struct ppcg_kernel
*kernel
)
1022 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1023 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1025 for (j
= 0; j
< array
->n_group
; ++j
)
1026 gpu_array_ref_group_compute_tiling(array
->groups
[j
]);
1030 /* Compute the effective grid size as a list of the sizes in each dimension.
1032 * The grid size specified by the user or set by default
1033 * in read_grid_sizes() and applied by the block filter,
1034 * may be too large for the given code in the sense that
1035 * it may contain blocks that don't need to execute anything.
1036 * We therefore don't return this grid size, but instead the
1037 * smallest grid size that ensures that all blocks that actually
1038 * execute code are included in the grid.
1040 * We first extract a description of the grid, i.e., the possible values
1041 * of the block ids, from the domain elements in "domain" and
1042 * kernel->block_filter.
1043 * The block ids are parameters in kernel->block_filter.
1044 * We simply need to change them into set dimensions.
1046 * Then, for each block dimension, we compute the maximal value of the block id
1049 static __isl_give isl_multi_pw_aff
*extract_grid_size(
1050 struct ppcg_kernel
*kernel
, __isl_take isl_union_set
*domain
)
1055 isl_multi_pw_aff
*size
;
1057 domain
= isl_union_set_intersect(domain
,
1058 isl_union_set_copy(kernel
->block_filter
));
1059 grid
= isl_union_set_params(domain
);
1060 grid
= isl_set_from_params(grid
);
1061 grid
= isl_set_add_dims(grid
, isl_dim_set
, kernel
->n_grid
);
1062 for (i
= 0; i
< kernel
->n_grid
; ++i
) {
1069 id
= isl_id_list_get_id(kernel
->block_ids
, i
);
1070 pos
= isl_set_find_dim_by_id(grid
, isl_dim_param
, id
);
1073 isl_die(isl_set_get_ctx(grid
), isl_error_internal
,
1074 "missing constraints on block identifier",
1075 grid
= isl_set_free(grid
));
1076 grid
= isl_set_equate(grid
, isl_dim_param
, pos
, isl_dim_set
, i
);
1077 grid
= isl_set_project_out(grid
, isl_dim_param
, pos
, 1);
1080 grid
= isl_set_coalesce(grid
);
1081 size
= ppcg_size_from_extent(grid
);
1082 context
= isl_set_params(isl_set_copy(kernel
->context
));
1083 return isl_multi_pw_aff_gist(size
, context
);
1086 /* Compute the size of a fixed bounding box around the origin and "set",
1087 * where "set" is assumed to contain only non-negative elements,
1088 * and store the results in "size".
1089 * In particular, compute the maximal value of "set" in each direction
1092 static void extract_fixed_size(__isl_take isl_set
*set
, int *size
)
1095 isl_local_space
*ls
;
1098 n
= isl_set_dim(set
, isl_dim_set
);
1099 ls
= isl_local_space_from_space(isl_set_get_space(set
));
1100 obj
= isl_aff_zero_on_domain(ls
);
1101 for (i
= 0; i
< n
; ++i
) {
1104 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 1);
1105 max
= isl_set_max_val(set
, obj
);
1106 size
[i
] = isl_val_get_num_si(max
) + 1;
1108 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 0);
1114 /* Compute the effective block size as a list of the sizes in each dimension
1115 * and store the sizes in kernel->block_dim.
1117 * The block size specified by the user or set by default
1118 * in read_block_sizes() and applied by the thread filter,
1119 * may be too large for the given code in the sense that
1120 * it may contain threads that don't need to execute anything.
1121 * We therefore update this block size in kernel->block_dim
1122 * to the smallest block size that ensures that all threads
1123 * that actually execute code are included in the block.
1125 * The set of possible values of the thread ids is obtained from
1126 * the domain elements "domain" and kernel->thread_filter.
1127 * The current implementation eliminates all parameters, ensuring
1128 * that the size is a fixed constant in each dimension.
1129 * In principle we could also compute parametric sizes.
1130 * We would have to make sure to project out all b%d and t%d parameters,
1133 static isl_stat
extract_block_size(struct ppcg_kernel
*kernel
,
1134 __isl_take isl_union_set
*domain
)
1140 domain
= isl_union_set_intersect(domain
,
1141 isl_union_set_copy(kernel
->thread_filter
));
1142 block
= isl_union_set_params(domain
);
1143 block
= isl_set_from_params(block
);
1144 block
= isl_set_add_dims(block
, isl_dim_set
, kernel
->n_block
);
1145 for (i
= 0; i
< kernel
->n_block
; ++i
) {
1150 return isl_stat_error
;
1152 id
= isl_id_list_get_id(kernel
->thread_ids
, i
);
1153 pos
= isl_set_find_dim_by_id(block
, isl_dim_param
, id
);
1156 isl_die(isl_set_get_ctx(block
), isl_error_internal
,
1157 "missing constraints on thread identifier",
1158 block
= isl_set_free(block
));
1159 block
= isl_set_equate(block
, isl_dim_param
, pos
,
1162 nparam
= isl_set_dim(block
, isl_dim_param
);
1163 block
= isl_set_project_out(block
, isl_dim_param
, 0, nparam
);
1166 return isl_stat_error
;
1168 extract_fixed_size(block
, kernel
->block_dim
);
1173 struct ppcg_kernel
*ppcg_kernel_free(struct ppcg_kernel
*kernel
)
1180 isl_id_list_free(kernel
->block_ids
);
1181 isl_id_list_free(kernel
->thread_ids
);
1182 isl_multi_pw_aff_free(kernel
->grid_size
);
1183 isl_ast_expr_free(kernel
->grid_size_expr
);
1184 isl_set_free(kernel
->context
);
1185 isl_union_set_free(kernel
->core
);
1186 isl_union_set_free(kernel
->arrays
);
1187 isl_union_pw_multi_aff_free(kernel
->contraction
);
1188 isl_union_set_free(kernel
->expanded_domain
);
1189 isl_space_free(kernel
->space
);
1190 isl_ast_node_free(kernel
->tree
);
1191 isl_union_set_free(kernel
->block_filter
);
1192 isl_union_set_free(kernel
->thread_filter
);
1193 isl_union_pw_multi_aff_free(kernel
->copy_schedule
);
1194 isl_union_set_free(kernel
->sync_writes
);
1196 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1197 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1199 for (j
= 0; j
< array
->n_group
; ++j
)
1200 gpu_array_ref_group_free(array
->groups
[j
]);
1201 free(array
->groups
);
1203 isl_multi_pw_aff_free(array
->bound
);
1204 isl_ast_expr_free(array
->bound_expr
);
1206 free(kernel
->array
);
1208 for (i
= 0; i
< kernel
->n_var
; ++i
) {
1209 free(kernel
->var
[i
].name
);
1210 isl_vec_free(kernel
->var
[i
].size
);
1219 /* Wrapper around ppcg_kernel_free for use as a isl_id_set_free_user callback.
1221 static void ppcg_kernel_free_wrap(void *user
)
1223 struct ppcg_kernel
*kernel
= user
;
1225 ppcg_kernel_free(kernel
);
1228 static void create_kernel_var(isl_ctx
*ctx
, struct gpu_array_ref_group
*group
,
1229 struct ppcg_kernel_var
*var
)
1232 struct gpu_array_tile
*tile
;
1235 var
->array
= group
->array
;
1237 var
->type
= gpu_array_ref_group_type(group
);
1238 tile
= gpu_array_ref_group_tile(group
);
1240 p
= isl_printer_to_str(ctx
);
1241 p
= gpu_array_ref_group_print_name(group
, p
);
1242 var
->name
= isl_printer_get_str(p
);
1243 isl_printer_free(p
);
1245 var
->size
= isl_vec_alloc(ctx
, group
->array
->n_index
);
1247 for (j
= 0; j
< group
->array
->n_index
; ++j
)
1248 var
->size
= isl_vec_set_element_val(var
->size
, j
,
1249 isl_val_copy(tile
->bound
[j
].size
));
1252 static isl_stat
create_kernel_vars(struct ppcg_kernel
*kernel
)
1257 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1258 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1260 for (j
= 0; j
< array
->n_group
; ++j
) {
1261 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1262 enum ppcg_group_access_type type
;
1264 type
= gpu_array_ref_group_type(group
);
1265 if (type
!= ppcg_access_global
)
1270 kernel
->var
= isl_calloc_array(kernel
->ctx
, struct ppcg_kernel_var
, n
);
1272 return isl_stat_error
;
1276 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1277 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
1279 for (j
= 0; j
< array
->n_group
; ++j
) {
1280 struct gpu_array_ref_group
*group
= array
->groups
[j
];
1281 enum ppcg_group_access_type type
;
1283 type
= gpu_array_ref_group_type(group
);
1284 if (type
== ppcg_access_global
)
1286 create_kernel_var(kernel
->ctx
, group
, &kernel
->var
[n
]);
1294 /* Replace "pa" by the zero function defined over the universe domain
1295 * in the space of "pa".
1297 static __isl_give isl_pw_aff
*set_universally_zero(__isl_take isl_pw_aff
*pa
)
1302 space
= isl_space_domain(isl_pw_aff_get_space(pa
));
1303 isl_pw_aff_free(pa
);
1304 zero
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1306 return isl_pw_aff_from_aff(zero
);
1309 /* The sizes of the arrays on the host that have been computed by
1310 * extract_array_info may depend on the parameters. Use the extra
1311 * constraints on the parameters that are valid at "host_domain"
1312 * to simplify these expressions and store the results in kernel->array.
1314 * We only need these localized bounds for arrays that are accessed
1315 * by the current kernel. If we have found at least one reference group
1316 * then the array is accessed by the kernel.
1318 * The resulting sizes may be functions that are nowhere defined
1319 * in case the access function cannot possibly access anything inside
1320 * the kernel for some reason. If so, they are replaced by the zero
1321 * function. Since the access function cannot actually access anything,
1322 * there is no harm in printing the array sizes as zero.
1324 static void localize_bounds(struct ppcg_kernel
*kernel
,
1325 __isl_keep isl_set
*host_domain
)
1330 context
= isl_set_copy(host_domain
);
1331 context
= isl_set_params(context
);
1333 for (i
= 0; i
< kernel
->n_array
; ++i
) {
1334 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
1335 isl_multi_pw_aff
*bound
;
1338 if (local
->n_group
== 0)
1341 n_index
= local
->array
->n_index
;
1342 bound
= isl_multi_pw_aff_copy(local
->array
->bound
);
1344 for (j
= 0; j
< n_index
; ++j
) {
1348 pwaff
= isl_multi_pw_aff_get_pw_aff(bound
, j
);
1349 pwaff
= isl_pw_aff_gist(pwaff
, isl_set_copy(context
));
1350 empty
= isl_pw_aff_is_empty(pwaff
);
1352 pwaff
= isl_pw_aff_free(pwaff
);
1354 pwaff
= set_universally_zero(pwaff
);
1355 bound
= isl_multi_pw_aff_set_pw_aff(bound
, j
, pwaff
);
1358 local
->n_index
= n_index
;
1359 local
->bound
= bound
;
1361 isl_set_free(context
);
1364 /* Create the array of gpu_local_array_info structures "array"
1365 * inside "kernel". The number of elements in this array is
1366 * the same as the number of arrays in "prog".
1367 * Initialize the "array" field of each local array to point
1368 * to the corresponding array in "prog".
1370 static struct ppcg_kernel
*ppcg_kernel_create_local_arrays(
1371 struct ppcg_kernel
*kernel
, struct gpu_prog
*prog
)
1379 ctx
= isl_set_get_ctx(prog
->context
);
1380 kernel
->array
= isl_calloc_array(ctx
,
1381 struct gpu_local_array_info
, prog
->n_array
);
1383 return ppcg_kernel_free(kernel
);
1384 kernel
->n_array
= prog
->n_array
;
1386 for (i
= 0; i
< prog
->n_array
; ++i
)
1387 kernel
->array
[i
].array
= &prog
->array
[i
];
1392 /* Does "kernel" need to be passed an argument corresponding to array "i"?
1394 * The argument is only needed if the kernel accesses this device memory.
1396 int ppcg_kernel_requires_array_argument(struct ppcg_kernel
*kernel
, int i
)
1398 return kernel
->array
[i
].global
;
1401 /* Find the element in gen->stmt that has the given "id".
1402 * Return NULL if no such gpu_stmt can be found.
1404 static struct gpu_stmt
*find_stmt(struct gpu_prog
*prog
, __isl_keep isl_id
*id
)
1408 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
1409 if (id
== prog
->stmts
[i
].id
)
1413 return i
< prog
->n_stmts
? &prog
->stmts
[i
] : NULL
;
1416 void ppcg_kernel_stmt_free(void *user
)
1418 struct ppcg_kernel_stmt
*stmt
= user
;
1423 switch (stmt
->type
) {
1424 case ppcg_kernel_copy
:
1425 isl_ast_expr_free(stmt
->u
.c
.index
);
1426 isl_ast_expr_free(stmt
->u
.c
.local_index
);
1428 case ppcg_kernel_domain
:
1429 isl_id_to_ast_expr_free(stmt
->u
.d
.ref2expr
);
1431 case ppcg_kernel_sync
:
1438 /* Return the gpu_stmt_access in the list "accesses" that corresponds
1441 static struct gpu_stmt_access
*find_access(struct gpu_stmt_access
*accesses
,
1442 __isl_keep isl_id
*ref_id
)
1444 struct gpu_stmt_access
*access
;
1446 for (access
= accesses
; access
; access
= access
->next
)
1447 if (access
->ref_id
== ref_id
)
1453 /* Return the index of the array called "name" in the list of arrays.
1455 static int find_array_index(struct ppcg_kernel
*kernel
, const char *name
)
1459 for (i
= 0; i
< kernel
->n_array
; ++i
)
1460 if (!strcmp(name
, kernel
->array
[i
].array
->name
))
1466 /* Internal data structure for the index and AST expression transformation
1467 * callbacks for pet_stmt_build_ast_exprs.
1469 * "kernel" is the kernel for which are computing AST expressions and
1470 * may be NULL if we are not inside a kernel.
1471 * "accesses" is the list of gpu_stmt_access in the statement.
1472 * "iterator_map" expresses the statement iterators in terms of
1473 * the AST loop iterators.
1474 * "sched2copy" expresses the outer copy_schedule_dim dimensions of
1475 * the kernel schedule in terms of the AST loop iterators and
1476 * may be NULL if we are not inside a kernel.
1478 * The following fields are set in transform_index and used in transform_expr.
1479 * "array" is the array that is being accessed.
1480 * "global" is set if the global array is accessed (rather than
1481 * shared/private memory).
1482 * "local_array" refers to information on the array specialized
1483 * to the current kernel.
1485 struct ppcg_transform_data
{
1486 struct ppcg_kernel
*kernel
;
1487 struct gpu_stmt_access
*accesses
;
1488 isl_pw_multi_aff
*iterator_map
;
1489 isl_pw_multi_aff
*sched2copy
;
1491 struct gpu_array_info
*array
;
1493 struct gpu_local_array_info
*local_array
;
1496 /* Return a pointer to the gpu_array_ref_group in "local"
1497 * that contains the reference "access".
1498 * Return NULL if no such group can be found.
1500 static struct gpu_array_ref_group
*find_ref_group(
1501 struct gpu_local_array_info
*local
, struct gpu_stmt_access
*access
)
1505 for (i
= 0; i
< local
->n_group
; ++i
) {
1506 struct gpu_array_ref_group
*group
= local
->groups
[i
];
1508 for (j
= 0; j
< group
->n_ref
; ++j
)
1509 if (group
->refs
[j
] == access
)
1516 /* Given an index expression "index" of the form
1520 * with F(A) either A or some subfield of A and L the AST loop iterators,
1521 * and a tiling "tiling" of the form
1525 * apply the tiling to the outer array in the index expression to obtain
1529 * If F(A) is some subfield of A, then separate the member access
1530 * into the base index expression and the field index expression,
1531 * apply the tiling to the base index expression and combine the result
1532 * with the field index expression.
1534 * If F(A) is A, then modify index to keep track of the iterators
1538 * and combine the result with the tiling to obtain a tiled index expression
1539 * in terms of the AST loop iterators
1543 static __isl_give isl_multi_pw_aff
*tile_outer(
1544 __isl_take isl_multi_pw_aff
*index
, __isl_take isl_multi_pw_aff
*tiling
)
1546 isl_bool is_wrapping
;
1548 isl_multi_pw_aff
*mpa
;
1550 is_wrapping
= isl_multi_pw_aff_range_is_wrapping(index
);
1551 if (is_wrapping
< 0)
1554 isl_multi_pw_aff
*field
;
1556 field
= isl_multi_pw_aff_copy(index
);
1557 field
= isl_multi_pw_aff_range_factor_range(field
);
1558 index
= isl_multi_pw_aff_range_factor_domain(index
);
1559 index
= tile_outer(index
, tiling
);
1560 return isl_multi_pw_aff_range_product(index
, field
);
1563 space
= isl_space_domain(isl_multi_pw_aff_get_space(index
));
1564 space
= isl_space_map_from_set(space
);
1565 mpa
= isl_multi_pw_aff_identity(space
);
1566 index
= isl_multi_pw_aff_range_product(mpa
, index
);
1567 index
= isl_multi_pw_aff_pullback_multi_pw_aff(tiling
, index
);
1571 isl_multi_pw_aff_free(index
);
1572 isl_multi_pw_aff_free(tiling
);
1576 /* Index transformation callback for pet_stmt_build_ast_exprs.
1578 * "index" expresses the array indices in terms of statement iterators
1580 * We first reformulate "index" in terms of the AST loop iterators.
1581 * Then we check if we are accessing the global array or
1582 * a shared/private copy. In particular, if we are not inside a kernel
1583 * then we must be accessing a global array.
1584 * In the former case, we simply return
1585 * the updated index. If "index" is an affine expression rather
1586 * than an array access, then we also return the updated index here.
1588 * If no reference groups have been computed for the array,
1589 * then we can only be accessing the global array.
1591 * Otherwise, we apply the tiling to the index.
1592 * This tiling is of the form
1596 * where D corresponds to the outer tile->depth dimensions of
1597 * the kernel schedule.
1598 * The index is of the form
1602 * We update the tiling to refer to the AST loop iterators
1606 * and combine it with the index to obtain a tiled index expression in terms
1607 * of the AST loop iterators
1611 * Note that while the tiling applies directly to an outer array.
1612 * the index may refer to some subfield of this outer array.
1613 * In such cases, the result will refer to the same subfield of the tile.
1614 * That is, an index expression of the form L -> F(A) will be transformed
1615 * into an index expression of the form L -> F(T).
1617 static __isl_give isl_multi_pw_aff
*transform_index(
1618 __isl_take isl_multi_pw_aff
*index
, __isl_keep isl_id
*ref_id
,
1621 struct ppcg_transform_data
*data
= user
;
1622 struct gpu_stmt_access
*access
;
1623 struct gpu_array_ref_group
*group
;
1624 struct gpu_array_tile
*tile
;
1625 isl_pw_multi_aff
*iterator_map
;
1630 isl_multi_pw_aff
*tiling
;
1631 isl_pw_multi_aff
*pma
;
1632 isl_pw_multi_aff
*sched2depth
;
1636 iterator_map
= isl_pw_multi_aff_copy(data
->iterator_map
);
1637 index
= isl_multi_pw_aff_pullback_pw_multi_aff(index
, iterator_map
);
1642 access
= find_access(data
->accesses
, ref_id
);
1645 if (!isl_map_has_tuple_name(access
->access
, isl_dim_out
))
1648 name
= get_outer_array_name(access
->access
);
1650 return isl_multi_pw_aff_free(index
);
1651 i
= find_array_index(data
->kernel
, name
);
1653 isl_die(isl_multi_pw_aff_get_ctx(index
), isl_error_internal
,
1654 "cannot find array",
1655 return isl_multi_pw_aff_free(index
));
1656 data
->local_array
= &data
->kernel
->array
[i
];
1657 data
->array
= data
->local_array
->array
;
1659 group
= find_ref_group(data
->local_array
, access
);
1665 tile
= gpu_array_ref_group_tile(group
);
1666 data
->global
= !tile
;
1670 space
= isl_space_domain(isl_multi_aff_get_space(tile
->tiling
));
1671 space
= isl_space_range(isl_space_unwrap(space
));
1672 space
= isl_space_map_from_set(space
);
1673 pma
= isl_pw_multi_aff_identity(space
);
1674 sched2depth
= isl_pw_multi_aff_copy(data
->sched2copy
);
1675 dim
= isl_pw_multi_aff_dim(sched2depth
, isl_dim_out
);
1676 sched2depth
= isl_pw_multi_aff_drop_dims(sched2depth
, isl_dim_out
,
1677 tile
->depth
, dim
- tile
->depth
);
1678 pma
= isl_pw_multi_aff_product(sched2depth
, pma
);
1679 tiling
= isl_multi_pw_aff_from_multi_aff(
1680 isl_multi_aff_copy(tile
->tiling
));
1681 tiling
= isl_multi_pw_aff_pullback_pw_multi_aff(tiling
, pma
);
1683 index
= tile_outer(index
, tiling
);
1688 /* Dereference "expr" by adding an index [0].
1689 * The original "expr" is assumed not to have any indices.
1691 * If "expr" is a member access, then the dereferencing needs
1692 * to be applied to the structure argument of this member access.
1694 static __isl_give isl_ast_expr
*dereference(__isl_take isl_ast_expr
*expr
)
1697 isl_ast_expr
*arg0
, *res
;
1698 isl_ast_expr_list
*list
;
1700 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1702 return isl_ast_expr_free(expr
);
1703 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1704 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1707 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1708 arg
= dereference(arg
);
1709 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1710 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1714 isl_ast_expr_free(arg0
);
1716 ctx
= isl_ast_expr_get_ctx(expr
);
1717 res
= isl_ast_expr_from_val(isl_val_zero(ctx
));
1718 list
= isl_ast_expr_list_from_ast_expr(res
);
1719 res
= isl_ast_expr_get_op_arg(expr
, 0);
1720 res
= isl_ast_expr_access(res
, list
);
1721 isl_ast_expr_free(expr
);
1726 /* Linearize the index expression "expr" based on the array bounds
1729 * That is, transform expression
1731 * A[i_0][i_1]...[i_n]
1735 * A[(..((i_0 * b_1 + i_1) ... ) * b_n + i_n]
1737 * where b_0, b_1, ..., b_n are the bounds on the array.
1739 * If the base of "expr" is a member access, then the linearization needs
1740 * to be applied to the structure argument of this member access.
1742 * In the base case, if "expr" has no arguments (other than the name of
1743 * the array), then we are passing an entire array to a function.
1744 * In this case, there is nothing to linearize.
1745 * Note that at this point an expression with no arguments can
1746 * only be an entire array because the scalar case and
1747 * the case of single struct are handled by the caller.
1749 * If the number of specified index expressions in "expr"
1750 * is smaller than the dimension of the accessed array,
1751 * then the missing i_j also do not appear in the linearized expression.
1752 * Furthermore, since such an expression does not refer to a single
1753 * element while the default linearized expression would refer to
1754 * a single element, we return the expression
1756 * A + (..((i_0 * b_1 + i_1) ... ) * b_l + i_l)
1758 * instead. Note that because of the special case handling above,
1759 * we can assume here that there is at least one index expression.
1761 __isl_give isl_ast_expr
*gpu_local_array_info_linearize_index(
1762 struct gpu_local_array_info
*array
, __isl_take isl_ast_expr
*expr
)
1767 isl_ast_expr_list
*list
;
1769 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
1770 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
1771 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
1774 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
1775 arg
= gpu_local_array_info_linearize_index(array
, arg
);
1776 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
1777 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
1781 isl_ast_expr_free(arg0
);
1783 if (isl_ast_expr_get_op_n_arg(expr
) == 1)
1786 n
= isl_ast_expr_get_op_n_arg(expr
);
1787 res
= isl_ast_expr_get_op_arg(expr
, 1);
1788 for (i
= 1; i
< array
->n_index
; ++i
) {
1789 isl_ast_expr
*expr_i
;
1791 expr_i
= isl_ast_expr_get_op_arg(array
->bound_expr
, 1 + i
);
1792 res
= isl_ast_expr_mul(res
, expr_i
);
1796 expr_i
= isl_ast_expr_get_op_arg(expr
, i
+ 1);
1797 res
= isl_ast_expr_add(res
, expr_i
);
1800 if (1 + array
->n_index
> n
) {
1801 res
= isl_ast_expr_add(isl_ast_expr_get_op_arg(expr
, 0), res
);
1803 list
= isl_ast_expr_list_from_ast_expr(res
);
1804 res
= isl_ast_expr_get_op_arg(expr
, 0);
1805 res
= isl_ast_expr_access(res
, list
);
1808 isl_ast_expr_free(expr
);
1813 /* AST expression transformation callback for pet_stmt_build_ast_exprs.
1815 * If the AST expression refers to an array that is not accessed
1816 * at all, then this means the value of the expression is not used,
1817 * so we might as well print zero (NULL pointer) instead.
1819 * If the AST expression refers to a global scalar that is not
1820 * a read-only scalar, then its address was passed to the kernel and
1821 * we need to dereference it.
1823 * If the AST expression refers to an access to a global array,
1824 * then we linearize the access exploiting the bounds in data->local_array.
1826 static __isl_give isl_ast_expr
*transform_expr(__isl_take isl_ast_expr
*expr
,
1827 __isl_keep isl_id
*id
, void *user
)
1829 struct ppcg_transform_data
*data
= user
;
1833 if (!data
->array
->accessed
) {
1836 ctx
= isl_ast_expr_get_ctx(expr
);
1837 isl_ast_expr_free(expr
);
1838 return isl_ast_expr_from_val(isl_val_zero(ctx
));
1840 if (gpu_array_is_read_only_scalar(data
->array
))
1844 if (data
->array
->n_index
== 0)
1845 return dereference(expr
);
1846 if (!data
->array
->linearize
)
1849 return gpu_local_array_info_linearize_index(data
->local_array
, expr
);
1852 /* This function is called for each instance of a user statement
1853 * in the kernel "kernel", identified by "gpu_stmt".
1854 * "kernel" may be NULL if we are not inside a kernel.
1856 * We attach a struct ppcg_kernel_stmt to the "node", containing
1857 * a computed AST expression for each access, through an annotation
1859 * These AST expressions are computed from iterator_map,
1860 * which expresses the domain
1861 * elements in terms of the generated loops, and sched2copy,
1862 * which expresses the outer copy_schedule_dim dimensions of
1863 * the kernel schedule computed by PPCG in terms of the generated loops.
1865 static __isl_give isl_ast_node
*create_domain_leaf(
1866 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
1867 __isl_keep isl_ast_build
*build
, struct gpu_stmt
*gpu_stmt
)
1869 struct ppcg_transform_data data
;
1870 struct ppcg_kernel_stmt
*stmt
;
1873 isl_pw_multi_aff
*sched2copy
;
1875 isl_pw_multi_aff
*iterator_map
;
1876 isl_union_map
*schedule
;
1880 ctx
= isl_ast_node_get_ctx(node
);
1882 stmt
= isl_calloc_type(ctx
, struct ppcg_kernel_stmt
);
1884 return isl_ast_node_free(node
);
1886 schedule
= isl_ast_build_get_schedule(build
);
1887 map
= isl_map_reverse(isl_map_from_union_map(schedule
));
1888 iterator_map
= isl_pw_multi_aff_from_map(map
);
1890 sched2copy
= compute_sched_to_copy(kernel
,
1891 isl_pw_multi_aff_copy(iterator_map
));
1895 stmt
->type
= ppcg_kernel_domain
;
1896 stmt
->u
.d
.stmt
= gpu_stmt
;
1898 data
.kernel
= kernel
;
1899 data
.accesses
= stmt
->u
.d
.stmt
->accesses
;
1900 data
.iterator_map
= iterator_map
;
1901 data
.sched2copy
= sched2copy
;
1902 stmt
->u
.d
.ref2expr
= pet_stmt_build_ast_exprs(stmt
->u
.d
.stmt
->stmt
,
1903 build
, &transform_index
, &data
,
1904 &transform_expr
, &data
);
1906 isl_pw_multi_aff_free(iterator_map
);
1907 isl_pw_multi_aff_free(sched2copy
);
1909 id
= isl_id_alloc(ctx
, "user", stmt
);
1910 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1912 ppcg_kernel_stmt_free(stmt
);
1913 return isl_ast_node_set_annotation(node
, id
);
1916 /* This function is called for each statement node in the AST
1917 * for copying to or from shared/private memory.
1918 * Attach a pointer to a ppcg_kernel_stmt representing the copy
1919 * statement to the node.
1920 * The statement name is "read" or "write", depending on whether we are
1921 * reading from global memory or writing to global memory.
1923 * The schedule is of the form
1927 * where D corresponds to the outer tile->depth dimensions of
1928 * the kernel schedule, A to the global array and L to the outer
1929 * generated AST schedule.
1930 * We compute the inverse and strip off the type, resulting in
1934 * We combine this mapping with on the one hand the projection
1938 * and on the other hand the group tiling
1946 * and store the corresponding expressions in stmt->index and stmt->local_index,
1947 * where stmt points to the ppcg_kernel_stmt that is attached to the node.
1948 * stmt->index is linearized if the global memory array is linearized.
1950 static __isl_give isl_ast_node
*create_access_leaf(struct ppcg_kernel
*kernel
,
1951 struct gpu_array_ref_group
*group
, __isl_take isl_ast_node
*node
,
1952 __isl_keep isl_ast_build
*build
)
1954 struct ppcg_kernel_stmt
*stmt
;
1955 struct gpu_array_tile
*tile
;
1960 isl_pw_multi_aff
*pma
, *pma2
;
1963 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
1965 return isl_ast_node_free(node
);
1967 access
= isl_map_from_union_map(isl_ast_build_get_schedule(build
));
1968 type
= isl_map_get_tuple_name(access
, isl_dim_in
);
1969 stmt
->u
.c
.read
= type
&& !strcmp(type
, "read");
1970 access
= isl_map_reverse(access
);
1971 pma
= isl_pw_multi_aff_from_map(access
);
1972 pma
= isl_pw_multi_aff_reset_tuple_id(pma
, isl_dim_out
);
1974 space
= isl_space_range(isl_pw_multi_aff_get_space(pma
));
1975 space
= isl_space_unwrap(space
);
1976 pma2
= isl_pw_multi_aff_range_map(space
);
1977 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
,
1978 isl_pw_multi_aff_copy(pma
));
1979 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1980 if (group
->array
->linearize
)
1981 expr
= gpu_local_array_info_linearize_index(group
->local_array
,
1983 stmt
->u
.c
.index
= expr
;
1985 tile
= gpu_array_ref_group_tile(group
);
1986 pma2
= isl_pw_multi_aff_from_multi_aff(
1987 isl_multi_aff_copy(tile
->tiling
));
1988 pma2
= isl_pw_multi_aff_pullback_pw_multi_aff(pma2
, pma
);
1989 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma2
);
1990 stmt
->u
.c
.local_index
= expr
;
1992 stmt
->u
.c
.array
= group
->array
;
1993 stmt
->u
.c
.local_array
= group
->local_array
;
1994 stmt
->type
= ppcg_kernel_copy
;
1996 id
= isl_id_alloc(kernel
->ctx
, "copy", stmt
);
1997 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
1999 ppcg_kernel_stmt_free(stmt
);
2000 return isl_ast_node_set_annotation(node
, id
);
2003 /* Create a synchronization ppcg_kernel_stmt and
2004 * attach it to the node "node" representing the synchronization.
2006 static __isl_give isl_ast_node
*create_sync_leaf(
2007 struct ppcg_kernel
*kernel
, __isl_take isl_ast_node
*node
,
2008 __isl_keep isl_ast_build
*build
)
2010 struct ppcg_kernel_stmt
*stmt
;
2013 stmt
= isl_calloc_type(kernel
->ctx
, struct ppcg_kernel_stmt
);
2015 return isl_ast_node_free(node
);
2017 stmt
->type
= ppcg_kernel_sync
;
2018 id
= isl_id_alloc(kernel
->ctx
, "sync", stmt
);
2019 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
2021 ppcg_kernel_stmt_free(stmt
);
2022 return isl_ast_node_set_annotation(node
, id
);
2025 /* Build AST expressions for the device array sizes of all arrays in "prog"
2026 * that require allocation on the device using "build", as well as
2027 * for the original array sizes of all arrays that need to be declared
2029 * "node" is freed in case of error.
2031 static __isl_give isl_ast_node
*build_array_bounds(
2032 __isl_take isl_ast_node
*node
, struct gpu_prog
*prog
,
2033 __isl_keep isl_ast_build
*build
)
2037 for (i
= 0; i
< prog
->n_array
; ++i
) {
2038 struct gpu_array_info
*array
= &prog
->array
[i
];
2039 isl_multi_pw_aff
*size
;
2042 if (!gpu_array_requires_device_allocation(array
))
2045 size
= isl_multi_pw_aff_copy(array
->bound
);
2046 expr
= ppcg_build_size_expr(size
, build
);
2047 array
->bound_expr
= expr
;
2049 return isl_ast_node_free(node
);
2052 for (i
= 0; i
< prog
->n_array
; ++i
) {
2053 struct gpu_array_info
*array
= &prog
->array
[i
];
2055 isl_multi_pw_aff
*size
;
2058 if (!array
->declare_local
)
2060 extent
= isl_set_copy(array
->declared_extent
);
2061 size
= ppcg_size_from_extent(extent
);
2062 expr
= ppcg_build_size_expr(size
, build
);
2063 array
->declared_size
= expr
;
2065 return isl_ast_node_free(node
);
2071 /* Internal data structure for at_domain.
2073 * "prog" represents the entire scop.
2074 * "kernel" points to the kernel to which the current schedule node
2075 * belongs. It is set by before_mark and reset by after_mark.
2076 * It may be NULL if we are outside any kernel.
2078 struct ppcg_at_domain_data
{
2079 struct gpu_prog
*prog
;
2080 struct ppcg_kernel
*kernel
;
2083 /* This function is called for each instance of a user statement
2084 * in the kernel. This may be one of the original user statements
2085 * or a statement introduced by PPCG.
2087 * We first check if the statement id corresponds to a gpu statement,
2088 * which indicates the statement is an original user statement. Any statement
2089 * that is not an original user statement has been introduced by PPCG and
2090 * requires special handling.
2092 * If the user statement is one of the original user statements, then we call
2093 * create_domain_leaf. If it is "init_device", then we call
2094 * build_array_bounds. Otherwise, we check if it is a copy or synchronization
2095 * statement and call the appropriate functions. Statements that copy an array
2096 * to/from the device do not need any further treatment.
2097 * Neither does "clear_device".
2099 static __isl_give isl_ast_node
*at_domain(__isl_take isl_ast_node
*node
,
2100 __isl_keep isl_ast_build
*build
, void *user
)
2102 struct ppcg_at_domain_data
*data
= user
;
2103 struct gpu_stmt
*gpu_stmt
;
2104 isl_ast_expr
*expr
, *arg
;
2110 expr
= isl_ast_node_user_get_expr(node
);
2111 arg
= isl_ast_expr_get_op_arg(expr
, 0);
2112 id
= isl_ast_expr_get_id(arg
);
2113 name
= isl_id_get_name(id
);
2114 p
= isl_id_get_user(id
);
2115 isl_ast_expr_free(expr
);
2116 isl_ast_expr_free(arg
);
2118 gpu_stmt
= find_stmt(data
->prog
, id
);
2119 is_sync
= gpu_tree_id_is_sync(id
, data
->kernel
);
2123 return create_domain_leaf(data
->kernel
, node
, build
, gpu_stmt
);
2125 if (!prefixcmp(name
, "to_device_") || !prefixcmp(name
, "from_device_"))
2127 if (!strcmp(name
, "init_device"))
2128 return build_array_bounds(node
, data
->prog
, build
);
2129 if (!strcmp(name
, "clear_device"))
2132 return isl_ast_node_free(node
);
2133 if (!strcmp(name
, "read") || !strcmp(name
, "write")) {
2134 struct gpu_array_ref_group
*group
= p
;
2135 return create_access_leaf(data
->kernel
, group
, node
, build
);
2138 isl_die(data
->prog
->ctx
, isl_error_internal
,
2139 "unknown statement type",
2140 return isl_ast_node_free(node
));
2141 return create_sync_leaf(data
->kernel
, node
, build
);
2144 /* Given a set of wrapped references "ref", return the corresponding
2145 * access relations based on the tagged access relations "tagged".
2147 * The elements of "ref" are of the form
2151 * with D an iteration domains and R a reference.
2152 * The elements of "tagged" are of the form
2158 * Extend "tagged" to include the iteration domain in the range, i.e.,
2160 * [D -> R] -> [D -> A]
2162 * apply the result to "ref" and then unwrap the resulting set
2163 * to obtain relations of the form
2167 static __isl_give isl_union_map
*wrapped_reference_to_access(
2168 __isl_take isl_union_set
*ref
, __isl_take isl_union_map
*tagged
)
2170 isl_union_map
*tag2access
;
2172 tag2access
= isl_union_map_copy(tagged
);
2173 tag2access
= isl_union_map_universe(tag2access
);
2174 tag2access
= isl_union_set_unwrap(isl_union_map_domain(tag2access
));
2175 tag2access
= isl_union_map_domain_map(tag2access
);
2176 tag2access
= isl_union_map_range_product(tag2access
, tagged
);
2178 ref
= isl_union_set_coalesce(ref
);
2179 ref
= isl_union_set_apply(ref
, tag2access
);
2181 return isl_union_set_unwrap(ref
);
2184 /* Given an access relation "access" from one or more array reference groups,
2185 * remove those reads if ("read" is 1) or writes (if "read" is 0)
2186 * that are only needed to communicate data within
2187 * the same iteration of "sched".
2188 * The domain of "sched" corresponds to the original statement instances,
2189 * i.e., those that appear in the domains of the access relations.
2190 * "tagged" contains all tagged access relations to all
2191 * the array reference groups accessed by "access" from statement
2192 * instances scheduled by "sched".
2194 * If the access is a read then it is either an element of
2196 * live_in union (range flow)
2198 * where live_in and flow may be overapproximations, or
2199 * it reads an uninitialized value (that is not live-in because
2200 * there is an intermediate kill) or it reads a value that was
2201 * written within the same (compound) statement instance.
2202 * If the access is a write then it is either an element of
2204 * live_out union (domain flow)
2206 * or it writes a value that is never read (and is not live-out
2207 * because of an intermediate kill) or only
2208 * within the same (compound) statement instance.
2209 * In both cases, the access relation is also a subset of
2210 * the group access relation.
2212 * The cases where an uninitialized value is read or a value is written
2213 * that is never read or where the dataflow occurs within a statement
2214 * instance are also considered local and may also be removed.
2216 * Essentially, we compute the intersection of "access" with either
2218 * live_in union (range non-local-flow)
2222 * live_out union (domain non-local-flow)
2224 * We first construct a relation "local"
2226 * [[D -> R] -> [D' -> R']]
2228 * of pairs of domain iterations accessing the reference group
2229 * and references in the group that are coscheduled by "sched".
2231 * If this relation does not intersect the dataflow dependences,
2232 * then there is nothing we can possibly remove, unless the dataflow
2233 * dependences themselves only relate a subset of the accesses.
2234 * In particular, the accesses may not be involved in any dataflow
2235 * dependences, either because they are uninitialized reads/dead writes
2236 * or because the dataflow occurs inside a statement instance.
2238 * Since the computation below may break up the access relation
2239 * into smaller pieces, we only perform the intersection with
2240 * the non-local dependent accesses if the local pairs
2241 * intersect the dataflow dependences. Otherwise, we intersect
2242 * with the universe of the non-local dependent accesses.
2243 * This should at least remove accesses from statements that
2244 * do not participate in any dependences.
2246 * In particular, we remove the "local" dataflow dependences from
2247 * the set of all dataflow dependences, or at least those
2248 * that may contribute to a domain/range that intersects
2249 * the domain of "access".
2250 * Note that if the potential dataflow dependences are an overapproximation
2251 * of the actual dataflow dependences, then the result remains an
2252 * overapproximation of the non-local dataflow dependences.
2253 * Copying to/from global memory is only needed for the references
2254 * in the domain/range of the result or for accesses that are live out/in
2255 * for the entire scop.
2257 * We therefore map the domain/range of the "external" relation
2258 * to the corresponding access relation and take the union with
2259 * the live out/in relation.
2261 static __isl_give isl_union_map
*remove_local_accesses(
2262 struct gpu_prog
*prog
, __isl_take isl_union_map
*tagged
,
2263 __isl_take isl_union_map
*access
, __isl_take isl_union_map
*sched
,
2267 isl_union_pw_multi_aff
*tagger
;
2268 isl_union_set
*domain
, *access_domain
;
2269 isl_union_map
*local
, *external
, *universe
;
2270 isl_union_set
*tag_set
;
2272 if (isl_union_map_is_empty(access
)) {
2273 isl_union_map_free(sched
);
2274 isl_union_map_free(tagged
);
2278 tagger
= isl_union_pw_multi_aff_copy(prog
->scop
->tagger
);
2279 domain
= isl_union_map_domain(isl_union_map_copy(tagged
));
2280 tagger
= isl_union_pw_multi_aff_intersect_domain(tagger
,
2281 isl_union_set_copy(domain
));
2282 sched
= isl_union_map_preimage_domain_union_pw_multi_aff(sched
, tagger
);
2284 local
= isl_union_map_apply_range(sched
,
2285 isl_union_map_reverse(isl_union_map_copy(sched
)));
2286 local
= isl_union_map_intersect(local
,
2287 isl_union_map_copy(prog
->scop
->tagged_dep_flow
));
2289 empty
= isl_union_map_is_empty(local
);
2291 external
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
2292 universe
= isl_union_map_universe(isl_union_map_copy(access
));
2293 access_domain
= isl_union_map_domain(universe
);
2294 domain
= isl_union_set_universe(domain
);
2295 universe
= isl_union_set_unwrap(domain
);
2296 universe
= isl_union_map_intersect_domain(universe
, access_domain
);
2297 domain
= isl_union_map_wrap(universe
);
2299 external
= isl_union_map_intersect_range(external
, domain
);
2301 external
= isl_union_map_intersect_domain(external
, domain
);
2302 external
= isl_union_map_intersect_params(external
,
2303 isl_set_copy(prog
->scop
->context
));
2304 external
= isl_union_map_subtract(external
, local
);
2307 tag_set
= isl_union_map_range(external
);
2308 external
= wrapped_reference_to_access(tag_set
, tagged
);
2309 external
= isl_union_map_union(external
,
2310 isl_union_map_copy(prog
->scop
->live_in
));
2312 tag_set
= isl_union_map_domain(external
);
2313 external
= wrapped_reference_to_access(tag_set
, tagged
);
2314 external
= isl_union_map_union(external
,
2315 isl_union_map_copy(prog
->scop
->live_out
));
2319 external
= isl_union_map_free(external
);
2321 external
= isl_union_map_universe(external
);
2323 access
= isl_union_map_intersect(access
, external
);
2328 /* Given an access relation "access" from "group", remove those reads
2329 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
2330 * communicate data within the same iteration of the schedule "prefix"
2331 * at the position where the copying of the group is inserted.
2332 * That is, the output dimension of "prefix"
2333 * is equal to tile->depth.
2334 * The domain of "prefix" corresponds to the original statement instances,
2335 * i.e., those that appear in the domains of the access relations.
2337 * Extract the tagged access relation of "group" and
2338 * then call remove_local_accesses.
2340 static __isl_give isl_union_map
*remove_local_accesses_group(
2341 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
2342 __isl_take isl_union_map
*access
, __isl_keep isl_union_map
*prefix
,
2345 isl_union_map
*sched
, *tagged
;
2347 if (isl_union_map_is_empty(access
))
2350 tagged
= group_tagged_access_relation(group
);
2351 sched
= isl_union_map_copy(prefix
);
2353 return remove_local_accesses(kernel
->prog
, tagged
, access
, sched
, read
);
2356 /* Build an access AST expression for the effective grid size using "build".
2357 * Store the result in kernel->grid_size_expr.
2359 static isl_stat
build_grid_size(struct ppcg_kernel
*kernel
,
2360 __isl_keep isl_ast_build
*build
)
2362 isl_multi_pw_aff
*size
;
2364 size
= isl_multi_pw_aff_copy(kernel
->grid_size
);
2365 size
= isl_multi_pw_aff_set_tuple_name(size
, isl_dim_out
, "grid");
2366 kernel
->grid_size_expr
= ppcg_build_size_expr(size
, build
);
2368 if (!kernel
->grid_size_expr
)
2369 return isl_stat_error
;
2373 /* Build access AST expressions for the localized array sizes using "build".
2374 * Store the result in local->bound_expr.
2375 * Only do this for arrays for which localized bounds have been computed.
2377 static isl_stat
build_local_array_sizes(struct ppcg_kernel
*kernel
,
2378 __isl_keep isl_ast_build
*build
)
2382 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2383 struct gpu_local_array_info
*local
= &kernel
->array
[i
];
2384 isl_multi_pw_aff
*size
;
2386 if (local
->n_group
== 0)
2388 size
= isl_multi_pw_aff_copy(local
->bound
);
2389 local
->bound_expr
= ppcg_build_size_expr(size
, build
);
2390 if (!local
->bound_expr
)
2391 return isl_stat_error
;
2397 /* Build access AST expressions for the effective grid size and
2398 * the localized array sizes using "build".
2400 static isl_stat
build_grid_and_local_array_sizes(struct ppcg_kernel
*kernel
,
2401 __isl_keep isl_ast_build
*build
)
2403 if (build_grid_size(kernel
, build
) < 0)
2404 return isl_stat_error
;
2405 if (build_local_array_sizes(kernel
, build
) < 0)
2406 return isl_stat_error
;
2410 /* This function is called before the AST generator starts traversing
2411 * the schedule subtree of a node with mark "mark".
2413 * If the mark is called "kernel", store the kernel pointer in data->kernel
2414 * for use in at_domain and build AST expressions for the grid size and
2415 * the localized array sizes.
2417 static isl_stat
before_mark(__isl_keep isl_id
*mark
,
2418 __isl_keep isl_ast_build
*build
, void *user
)
2420 struct ppcg_at_domain_data
*data
= user
;
2423 return isl_stat_error
;
2424 if (!strcmp(isl_id_get_name(mark
), "kernel")) {
2425 data
->kernel
= isl_id_get_user(mark
);
2426 if (build_grid_and_local_array_sizes(data
->kernel
, build
) < 0)
2427 return isl_stat_error
;
2432 /* This function is called after the AST generator has finished traversing
2433 * the schedule subtree of a mark node. "node" points to the corresponding
2436 * If the mark is called "kernel", then replace "node" by a user node
2437 * that "calls" the kernel, representing the launch of the kernel.
2438 * The original "node" is stored inside the kernel object so that
2439 * it can be used to print the device code.
2440 * Note that this assumes that a kernel is only launched once.
2441 * Also clear data->kernel.
2443 static __isl_give isl_ast_node
*after_mark(__isl_take isl_ast_node
*node
,
2444 __isl_keep isl_ast_build
*build
, void *user
)
2449 isl_ast_expr_list
*list
;
2450 struct ppcg_kernel
*kernel
;
2451 struct ppcg_at_domain_data
*data
= user
;
2453 ctx
= isl_ast_node_get_ctx(node
);
2454 id
= isl_ast_node_mark_get_id(node
);
2456 return isl_ast_node_free(node
);
2457 if (strcmp(isl_id_get_name(id
), "kernel") || !data
->kernel
) {
2461 kernel
= data
->kernel
;
2462 data
->kernel
= NULL
;
2463 kernel
->space
= isl_ast_build_get_schedule_space(build
);
2464 kernel
->tree
= isl_ast_node_mark_get_node(node
);
2465 isl_ast_node_free(node
);
2467 expr
= isl_ast_expr_from_id(isl_id_copy(id
));
2468 list
= isl_ast_expr_list_alloc(ctx
, 0);
2469 expr
= isl_ast_expr_call(expr
, list
);
2470 node
= isl_ast_node_alloc_user(expr
);
2471 node
= isl_ast_node_set_annotation(node
, id
);
2476 static isl_bool
update_depth(__isl_keep isl_schedule_node
*node
, void *user
)
2481 if (isl_schedule_node_get_type(node
) != isl_schedule_node_leaf
)
2482 return isl_bool_true
;
2483 node_depth
= isl_schedule_node_get_schedule_depth(node
);
2484 if (node_depth
> *depth
)
2485 *depth
= node_depth
;
2487 return isl_bool_false
;
2490 /* Use isl to generate code for both the host and the device
2492 * The device code is marked by "kernel" mark nodes in the schedule tree,
2493 * containing a pointer to a ppcg_kernel object.
2494 * The returned AST only contains the AST for the host code.
2495 * The ASTs for the device code are embedded in ppcg_kernel objects
2496 * attached to the leaf nodes that call "kernel".
2498 static __isl_give isl_ast_node
*generate_code(struct gpu_gen
*gen
,
2499 __isl_take isl_schedule
*schedule
)
2501 struct ppcg_at_domain_data data
;
2502 isl_ast_build
*build
;
2504 isl_id_list
*iterators
;
2507 data
.prog
= gen
->prog
;
2511 if (isl_schedule_foreach_schedule_node_top_down(schedule
, &update_depth
,
2513 schedule
= isl_schedule_free(schedule
);
2514 build
= isl_ast_build_alloc(gen
->prog
->ctx
);
2515 iterators
= ppcg_scop_generate_names(gen
->prog
->scop
, depth
, "c");
2516 build
= isl_ast_build_set_iterators(build
, iterators
);
2517 build
= isl_ast_build_set_at_each_domain(build
, &at_domain
, &data
);
2518 build
= isl_ast_build_set_before_each_mark(build
, &before_mark
, &data
);
2519 build
= isl_ast_build_set_after_each_mark(build
, &after_mark
, &data
);
2520 if (gen
->prog
->scop
->options
->debug
->dump_final_schedule
)
2521 isl_schedule_dump(schedule
);
2522 tree
= isl_ast_build_node_from_schedule(build
, schedule
);
2523 isl_ast_build_free(build
);
2528 __isl_give isl_union_map
*extract_sizes_from_str(isl_ctx
*ctx
, const char *str
)
2532 return isl_union_map_read_from_str(ctx
, str
);
2535 /* Can "node" be tiled and then mapped to block and thread identifiers?
2536 * That is, is it permutable with at least one coincident dimension?
2538 static isl_bool
is_permutable(__isl_keep isl_schedule_node
*node
)
2541 return isl_bool_error
;
2543 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
2544 return isl_bool_false
;
2545 if (!isl_schedule_node_band_get_permutable(node
))
2546 return isl_bool_false
;
2547 if (isl_schedule_node_band_n_member(node
) < 1)
2548 return isl_bool_false
;
2549 if (!isl_schedule_node_band_member_get_coincident(node
, 0))
2550 return isl_bool_false
;
2552 return isl_bool_true
;
2555 /* Is "node" not a suitably permutable band?
2557 static isl_bool
not_permutable(__isl_keep isl_schedule_node
*node
, void *user
)
2559 return isl_bool_not(is_permutable(node
));
2562 /* Does the subtree rooted at "node" have any suitably permutable band nodes?
2563 * That is, does it have any nodes that are permutable and that
2564 * have a least one coincident dimension?
2566 static isl_bool
subtree_has_permutable_bands(__isl_keep isl_schedule_node
*node
)
2568 isl_bool all_non_permutable
;
2570 all_non_permutable
= isl_schedule_node_every_descendant(node
,
2571 ¬_permutable
, NULL
);
2572 return isl_bool_not(all_non_permutable
);
2575 /* Does "schedule" contain any permutable band with at least one coincident
2578 static isl_bool
has_any_permutable_node(__isl_keep isl_schedule
*schedule
)
2580 isl_schedule_node
*root
;
2581 isl_bool any_permutable
;
2583 root
= isl_schedule_get_root(schedule
);
2584 any_permutable
= subtree_has_permutable_bands(root
);
2585 isl_schedule_node_free(root
);
2587 return any_permutable
;
2590 /* Is "node" a candidate for mapping to block and thread identifiers?
2591 * In particular, is it permutable with at least one coincident dimension?
2592 * Alternatively, does the subtree rooted at "node" not contain
2593 * any such permutable node? Filter nodes are skipped in this case,
2594 * because a band node will be inserted in front of the returned
2595 * node and this is not possible for filter nodes that are children
2596 * of set or sequence nodes.
2598 static int is_candidate(__isl_keep isl_schedule_node
*node
)
2600 isl_bool permutable
;
2602 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
)
2604 permutable
= is_permutable(node
);
2605 if (permutable
< 0 || permutable
)
2607 if (isl_schedule_node_get_type(node
) == isl_schedule_node_filter
)
2609 permutable
= subtree_has_permutable_bands(node
);
2615 /* Is "node" the outermost node in its branch that can be tiled
2616 * and then mapped to block and thread identifiers?
2617 * If there are no such nodes in the subtree at "node" and
2618 * if "node" is not a filter node, then it is accepted too.
2620 static int is_outer_tilable(__isl_keep isl_schedule_node
*node
)
2623 isl_schedule_node
*ancestor
;
2625 tilable
= is_candidate(node
);
2632 ancestor
= isl_schedule_node_copy(node
);
2633 while (isl_schedule_node_has_parent(ancestor
)) {
2634 ancestor
= isl_schedule_node_parent(ancestor
);
2636 tilable
= is_candidate(ancestor
);
2637 if (tilable
< 0 || tilable
)
2641 isl_schedule_node_free(ancestor
);
2642 return tilable
< 0 ? -1 : !tilable
;
2645 /* Collect the references to all writes in "group".
2646 * Each reference is represented by a universe set in a space
2650 * with S[i,j] the statement instance space and R[] the array reference.
2652 static __isl_give isl_union_set
*group_tagged_writes(
2653 struct gpu_array_ref_group
*group
)
2657 isl_union_set
*writes
;
2659 space
= isl_map_get_space(group
->access
);
2660 writes
= isl_union_set_empty(space
);
2661 for (i
= 0; i
< group
->n_ref
; ++i
) {
2665 if (!group
->refs
[i
]->write
)
2668 space
= isl_map_get_space(group
->refs
[i
]->tagged_access
);
2669 space
= isl_space_domain(space
);
2670 writes_i
= isl_set_universe(space
);
2671 writes
= isl_union_set_add_set(writes
, writes_i
);
2677 /* Is there any write access in "group" that requires synchronization
2678 * on a write to global memory?
2679 * We currently take into account all writes that would require
2680 * synchronization at the thread level depth, but if the copying
2681 * for this group is performed at an outer level, then we do not
2682 * actually need to take into account dependences at intermediate levels.
2684 static int any_sync_writes_in_group(struct ppcg_kernel
*kernel
,
2685 struct gpu_array_ref_group
*group
)
2687 isl_union_set
*writes
;
2688 int empty
, disjoint
;
2690 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2696 writes
= group_tagged_writes(group
);
2697 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2698 isl_union_set_free(writes
);
2700 return disjoint
< 0 ? -1 : !disjoint
;
2703 /* Collect the references to all writes in "kernel" that write directly
2704 * to global or shared memory, i.e., that are not mapped to private memory.
2705 * Each reference is represented by a universe set in a space
2709 * with S[i,j] the statement instance space and R[] the array reference.
2711 static __isl_give isl_union_set
*collect_non_private_tagged_writes(
2712 struct ppcg_kernel
*kernel
)
2714 isl_union_set
*writes
;
2717 writes
= isl_union_set_empty(isl_union_set_get_space(kernel
->arrays
));
2719 for (i
= 0; i
< kernel
->n_array
; ++i
) {
2720 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
2722 for (j
= 0; j
< array
->n_group
; ++j
) {
2723 struct gpu_array_ref_group
*group
= array
->groups
[j
];
2724 enum ppcg_group_access_type type
;
2725 isl_union_set
*writes_ij
;
2729 type
= gpu_array_ref_group_type(group
);
2730 if (type
== ppcg_access_private
)
2732 writes_ij
= group_tagged_writes(group
);
2733 writes
= isl_union_set_union(writes
, writes_ij
);
2740 /* Are there any direct writes to global memory that require
2743 static int any_global_or_shared_sync_writes(struct ppcg_kernel
*kernel
)
2745 isl_union_set
*writes
;
2746 int empty
, disjoint
;
2748 empty
= isl_union_set_is_empty(kernel
->sync_writes
);
2754 writes
= collect_non_private_tagged_writes(kernel
);
2755 disjoint
= isl_union_set_is_disjoint(kernel
->sync_writes
, writes
);
2756 isl_union_set_free(writes
);
2758 return disjoint
< 0 ? -1 : !disjoint
;
2761 /* Construct an isl_multi_val for use as tile sizes for tiling "node"
2762 * from the elements in "tile_size".
2764 static __isl_give isl_multi_val
*construct_band_tiles_sizes(
2765 __isl_keep isl_schedule_node
*node
, int *tile_size
)
2772 space
= isl_schedule_node_band_get_space(node
);
2773 return ppcg_multi_val_from_int_list(space
, tile_size
);
2776 /* Replace the partial schedule S of the band node "node" by
2784 * if scale_tile_loops is set, with f the integers in "factor".
2785 * The list that "factor" points to is assumed to contain at least
2786 * as many elements as the number of members in the band.
2788 static __isl_give isl_schedule_node
*snap_band_to_sizes(
2789 __isl_take isl_schedule_node
*node
, int *factor
,
2790 struct ppcg_options
*options
)
2794 mv
= construct_band_tiles_sizes(node
, factor
);
2795 node
= isl_schedule_node_band_scale_down(node
, isl_multi_val_copy(mv
));
2796 if (options
->scale_tile_loops
)
2797 node
= isl_schedule_node_band_scale(node
,
2798 isl_multi_val_copy(mv
));
2799 isl_multi_val_free(mv
);
2804 /* Tile "band" with tile size specified by "sizes".
2806 * Since the tile loops will be mapped to block ids, we forcibly
2807 * turn off tile loop scaling. We may want to enable tile loop scaling
2808 * at some later point, but then we would have to support the detection
2809 * of strides during the mapping to block ids.
2810 * Similarly, since the point loops will be mapped to thread ids,
2811 * we forcibly shift the point loops so that they start at zero.
2813 static __isl_give isl_schedule_node
*tile_band(
2814 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2816 isl_ctx
*ctx
= isl_schedule_node_get_ctx(node
);
2820 scale_tile
= isl_options_get_tile_scale_tile_loops(ctx
);
2821 isl_options_set_tile_scale_tile_loops(ctx
, 0);
2822 shift_point
= isl_options_get_tile_shift_point_loops(ctx
);
2823 isl_options_set_tile_shift_point_loops(ctx
, 1);
2825 node
= ppcg_tile(node
, sizes
);
2827 isl_options_set_tile_scale_tile_loops(ctx
, scale_tile
);
2828 isl_options_set_tile_shift_point_loops(ctx
, shift_point
);
2833 /* Extract the set of parameter values and outer schedule dimensions
2834 * for which any statement instance
2835 * in the kernel inserted at "node" needs to be executed.
2836 * Intersect the set of parameter values derived from the host schedule
2837 * relation with the context of "prog".
2839 static __isl_give isl_set
*extract_context(__isl_keep isl_schedule_node
*node
,
2840 struct gpu_prog
*prog
)
2842 isl_union_map
*schedule
;
2843 isl_union_set
*schedule_domain
;
2847 schedule
= isl_schedule_node_get_prefix_schedule_relation(node
);
2848 schedule_domain
= isl_union_map_range(schedule
);
2849 empty
= isl_union_set_is_empty(schedule_domain
);
2851 isl_union_set_free(schedule_domain
);
2858 space
= isl_union_set_get_space(schedule_domain
);
2859 isl_union_set_free(schedule_domain
);
2860 space
= isl_space_set_from_params(space
);
2861 depth
= isl_schedule_node_get_schedule_depth(node
);
2862 space
= isl_space_add_dims(space
, isl_dim_set
, depth
);
2863 context
= isl_set_empty(space
);
2865 context
= isl_set_from_union_set(schedule_domain
);
2867 context
= isl_set_intersect_params(context
,
2868 isl_set_copy(prog
->context
));
2873 /* Return the set of outer array elements accessed by
2874 * by the statement instances in "domain" in "prog".
2875 * The instances in "domain" are those that appear
2876 * in the domains of the access relations in "prog".
2878 static __isl_give isl_union_set
*accessed_by_domain(
2879 __isl_take isl_union_set
*domain
, struct gpu_prog
*prog
)
2881 isl_union_map
*access
;
2882 isl_union_set
*arrays
;
2884 access
= isl_union_map_union(isl_union_map_copy(prog
->read
),
2885 isl_union_map_copy(prog
->may_write
));
2886 access
= isl_union_map_intersect_domain(access
, domain
);
2887 arrays
= isl_union_map_range(access
);
2888 arrays
= isl_union_set_apply(arrays
,
2889 isl_union_map_copy(prog
->to_outer
));
2894 /* Return the number of outer band members of the band node "node"
2895 * that are marked coincident.
2897 static int n_outer_coincidence(__isl_keep isl_schedule_node
*node
)
2901 n
= isl_schedule_node_band_n_member(node
);
2903 for (i
= 0; i
< n
; ++i
)
2904 if (!isl_schedule_node_band_member_get_coincident(node
, i
))
2910 /* If the band node "node" has more than "n" members, then split off
2911 * the first "n" of them.
2913 static __isl_give isl_schedule_node
*split_band(
2914 __isl_take isl_schedule_node
*node
, int n
)
2918 dim
= isl_schedule_node_band_n_member(node
);
2920 node
= isl_schedule_node_band_split(node
, n
);
2925 /* Scale a band node that may have been split by split_band.
2926 * "sizes" are the scaling factors for the original node.
2927 * "node" either points to the original band node, or the outer
2928 * of the two pieces after splitting.
2930 * If the number of elements in "node" is smaller than the number of
2931 * elements in "sizes", then some splitting has occurred and we split
2932 * "sizes" in the same way.
2934 static __isl_give isl_schedule_node
*scale_band(
2935 __isl_take isl_schedule_node
*node
, __isl_take isl_multi_val
*sizes
)
2939 n
= isl_multi_val_dim(sizes
, isl_dim_set
);
2940 dim
= isl_schedule_node_band_n_member(node
);
2942 isl_multi_val
*sizes2
;
2944 sizes2
= isl_multi_val_copy(sizes
);
2945 sizes
= isl_multi_val_drop_dims(sizes
,
2946 isl_dim_set
, dim
, n
- dim
);
2947 sizes2
= isl_multi_val_drop_dims(sizes2
, isl_dim_set
, 0, dim
);
2948 node
= isl_schedule_node_child(node
, 0);
2949 node
= isl_schedule_node_band_scale(node
, sizes2
);
2950 node
= isl_schedule_node_parent(node
);
2953 return isl_schedule_node_band_scale(node
, sizes
);
2956 /* Return an isl_multi_aff, with as elements the parameters in "space"
2957 * that have the names specified by the elements in "names".
2958 * If (some of) these parameters do not already appear in "space",
2959 * then they are added first.
2961 static __isl_give isl_multi_aff
*parameter_vector(__isl_take isl_space
*space
,
2962 __isl_keep isl_id_list
*names
)
2965 isl_local_space
*ls
;
2969 space
= isl_space_free(space
);
2971 n
= isl_id_list_n_id(names
);
2972 for (i
= 0; i
< n
; ++i
) {
2976 id
= isl_id_list_get_id(names
, i
);
2977 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2982 pos
= isl_space_dim(space
, isl_dim_param
);
2983 space
= isl_space_add_dims(space
, isl_dim_param
, 1);
2984 space
= isl_space_set_dim_id(space
, isl_dim_param
, pos
, id
);
2986 ma
= isl_multi_aff_zero(isl_space_copy(space
));
2987 ls
= isl_local_space_from_space(isl_space_domain(space
));
2988 for (i
= 0; i
< n
; ++i
) {
2993 id
= isl_id_list_get_id(names
, i
);
2994 pos
= isl_space_find_dim_by_id(space
, isl_dim_param
, id
);
2996 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
2997 isl_dim_param
, pos
);
2998 ma
= isl_multi_aff_set_aff(ma
, i
, aff
);
3000 isl_local_space_free(ls
);
3005 /* Return constraints on the domain elements that equate a sequence of
3006 * parameters called "names", to the partial schedule
3007 * of "node" modulo the integers in "size".
3008 * The number of elements in the array "size" should be equal
3009 * to the number of elements in "names".
3010 * The number of members of the band node "node" should be smaller
3011 * than or equal to this number. If it is smaller, then the first
3012 * elements of "names" are equated to zero.
3014 static __isl_give isl_union_set
*set_schedule_modulo(
3015 __isl_keep isl_schedule_node
*node
, __isl_keep isl_id_list
*names
,
3021 isl_multi_union_pw_aff
*mupa
, *mupa2
;
3023 isl_union_set
*domain
;
3027 n
= isl_id_list_n_id(names
);
3029 return isl_schedule_node_get_universe_domain(node
);
3030 n_zero
= n
- isl_schedule_node_band_n_member(node
);
3032 mupa
= isl_schedule_node_band_get_partial_schedule(node
);
3033 mv
= construct_band_tiles_sizes(node
, size
+ n_zero
);
3034 mupa
= isl_multi_union_pw_aff_mod_multi_val(mupa
, mv
);
3036 space
= isl_multi_union_pw_aff_get_space(mupa
);
3037 space
= isl_space_params(space
);
3038 space
= isl_space_set_from_params(space
);
3039 space
= isl_space_add_dims(space
, isl_dim_set
, n_zero
);
3040 ma
= isl_multi_aff_zero(space
);
3042 domain
= isl_schedule_node_get_universe_domain(node
);
3043 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(
3044 isl_union_set_copy(domain
), ma
);
3045 mupa
= isl_multi_union_pw_aff_range_product(mupa2
, mupa
);
3047 space
= isl_multi_union_pw_aff_get_space(mupa
);
3048 ma
= parameter_vector(space
, names
);
3050 mupa2
= isl_multi_union_pw_aff_multi_aff_on_domain(domain
, ma
);
3051 mupa
= isl_multi_union_pw_aff_sub(mupa
, mupa2
);
3053 return isl_multi_union_pw_aff_zero_union_set(mupa
);
3056 /* Insert a context node at "node" introducing the block and thread
3057 * identifiers along with their bounds, which are stored in kernel->grid_size
3058 * and kernel->block_dim.
3059 * Note that the bounds on the block identifiers may implicitly impose
3060 * constraints on the parameters. A guard needs to be inserted
3061 * in the schedule tree to ensure that those bounds hold at "node".
3062 * This guard is inserted in insert_guard.
3064 static __isl_give isl_schedule_node
*insert_context(struct ppcg_kernel
*kernel
,
3065 __isl_take isl_schedule_node
*node
)
3069 context
= isl_set_universe(isl_set_get_space(kernel
->context
));
3071 context
= add_bounded_parameters_dynamic(context
,
3072 kernel
->grid_size
, kernel
->block_ids
);
3073 context
= add_bounded_parameters(context
,
3074 kernel
->block_dim
, kernel
->thread_ids
);
3076 node
= isl_schedule_node_insert_context(node
, context
);
3081 /* Insert a guard that eliminates kernel launches where the kernel
3082 * obviously does not have any work to do.
3084 * In particular, eliminate kernel launches where there are obviously
3086 * Use the same block size constraints that are used to create the context
3087 * to ensure that all constraints implicit in the constructed context
3088 * are imposed by the guard.
3090 * Additionally, add other constraints that are valid
3091 * for each executed instance ("context"), as long as this does not result
3094 static __isl_give isl_schedule_node
*insert_guard(
3095 __isl_take isl_schedule_node
*node
, __isl_keep isl_set
*context
,
3096 __isl_keep isl_multi_pw_aff
*size
, struct ppcg_scop
*scop
)
3102 guard
= isl_set_copy(context
);
3103 guard
= isl_set_compute_divs(guard
);
3104 guard
= isl_set_from_basic_set(isl_set_simple_hull(guard
));
3106 nparam
= isl_set_dim(guard
, isl_dim_param
);
3107 n
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
3108 ids
= ppcg_scop_generate_names(scop
, n
, "__ppcg_tmp");
3109 guard
= add_bounded_parameters_dynamic(guard
, size
, ids
);
3110 isl_id_list_free(ids
);
3111 guard
= isl_set_project_out(guard
, isl_dim_param
, nparam
, n
);
3113 node
= isl_schedule_node_insert_guard(node
, guard
);
3118 /* Does any array reference group mapping require the band that is mapped
3119 * to threads to be unrolled?
3121 static int kernel_requires_unroll(struct ppcg_kernel
*kernel
)
3125 for (i
= 0; i
< kernel
->n_array
; ++i
) {
3126 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
3128 for (j
= 0; j
< array
->n_group
; ++j
) {
3129 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3130 if (gpu_array_ref_group_requires_unroll(group
))
3138 /* Mark the given band node "node" for unrolling by the AST generator and
3139 * then sink it to the leaves of the schedule tree.
3140 * All dimensions of "node" are assumed to be coincident, such that this
3141 * sinking is a valid operation.
3143 static __isl_give isl_schedule_node
*unroll(__isl_take isl_schedule_node
*node
)
3145 node
= ppcg_set_schedule_node_type(node
, isl_ast_loop_unroll
);
3147 node
= isl_schedule_node_band_sink(node
);
3152 /* Insert a synchronization node in the schedule tree of "node"
3153 * after the core computation of "kernel" at the level of the band
3154 * that is mapped to threads, except if that level is equal to
3155 * that of the band that is mapped to blocks or if there are no writes
3156 * to global or shared memory in the core computation that require
3158 * If there are any writes to shared memory and the shared memory
3159 * copying is performed at the same level, then synchronization
3160 * is needed between the core and the copying anyway, so we might
3161 * as well add it here. If the copying is performed at a higher
3162 * level, then different iterations of intermediate schedule dimensions
3163 * may have a different mapping from between shared memory elements and
3164 * threads, such that synchronization is required after the core.
3165 * "node" is assumed to point to the kernel node.
3167 * If the shared and the thread mark point to the same node, then make
3168 * sure the synchronization is inserted outside of the shared mark.
3170 static __isl_give isl_schedule_node
*add_sync(struct ppcg_kernel
*kernel
,
3171 __isl_take isl_schedule_node
*node
)
3176 need_sync
= any_global_or_shared_sync_writes(kernel
);
3178 return isl_schedule_node_free(node
);
3182 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3183 depth
= isl_schedule_node_get_schedule_depth(node
);
3184 node
= gpu_tree_move_up_to_kernel(node
);
3185 if (depth
== isl_schedule_node_get_schedule_depth(node
))
3188 node
= gpu_tree_move_down_to_depth(node
, depth
, kernel
->core
);
3189 node
= gpu_tree_ensure_following_sync(node
, kernel
);
3191 node
= gpu_tree_move_up_to_kernel(node
);
3196 /* Return a read ("read" is 1) or write access relation for "group"
3197 * with those accesses removed that are only needed to communicate data
3198 * within the subtree of the schedule rooted at "node".
3199 * Furthermore, include the prefix schedule at "node".
3200 * That is, return a relation of the form
3204 * with D the outer schedule dimensions at "node".
3206 static __isl_give isl_union_map
*anchored_non_local_accesses(
3207 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3208 __isl_take isl_schedule_node
*node
, int read
)
3210 isl_union_map
*access
;
3211 isl_union_map
*prefix
;
3213 prefix
= isl_schedule_node_get_prefix_schedule_relation(node
);
3214 prefix
= isl_union_map_preimage_domain_union_pw_multi_aff(prefix
,
3215 isl_union_pw_multi_aff_copy(kernel
->contraction
));
3216 access
= gpu_array_ref_group_access_relation(group
, read
, !read
);
3217 access
= remove_local_accesses_group(kernel
, group
, access
, prefix
,
3219 access
= isl_union_map_range_product(prefix
, access
);
3224 /* Given an array reference group "group", create a mapping
3226 * read[D -> A] -> [D -> A]
3228 * if "read" is set or
3230 * write[D -> A] -> [D -> A]
3232 * if "read" is not set.
3233 * D corresponds to the outer tile->depth dimensions of
3234 * the kernel schedule.
3236 static __isl_give isl_multi_aff
*create_from_access(isl_ctx
*ctx
,
3237 struct gpu_array_ref_group
*group
, int read
)
3239 struct gpu_array_tile
*tile
;
3243 tile
= gpu_array_ref_group_tile(group
);
3244 space
= isl_space_copy(group
->array
->space
);
3245 space
= isl_space_from_range(space
);
3246 space
= isl_space_add_dims(space
, isl_dim_in
, tile
->depth
);
3247 space
= isl_space_wrap(space
);
3248 space
= isl_space_map_from_set(space
);
3250 id
= isl_id_alloc(ctx
, read
? "read" : "write", group
);
3251 space
= isl_space_set_tuple_id(space
, isl_dim_in
, id
);
3253 return isl_multi_aff_identity(space
);
3256 /* If any writes in "group" require synchronization, then make sure
3257 * that there is a synchronization node for "kernel" after the node
3258 * following "node" in a sequence.
3260 * If "shared" is set and no synchronization is needed for
3261 * the writes to global memory, then add synchronization before
3262 * the kernel to protect shared memory from being overwritten
3263 * by the next iteration of the core computation.
3264 * No additional synchronization is needed to protect against
3265 * the next copy into shared memory because each element of
3266 * the shared memory tile is always copied by the same thread.
3268 static __isl_give isl_schedule_node
*add_group_write_sync(
3269 __isl_take isl_schedule_node
*node
, struct ppcg_kernel
*kernel
,
3270 struct gpu_array_ref_group
*group
, int shared
)
3274 need_sync
= any_sync_writes_in_group(kernel
, group
);
3276 return isl_schedule_node_free(node
);
3278 node
= isl_schedule_node_parent(node
);
3279 node
= isl_schedule_node_next_sibling(node
);
3280 node
= isl_schedule_node_child(node
, 0);
3281 node
= gpu_tree_ensure_following_sync(node
, kernel
);
3282 } else if (shared
) {
3283 struct gpu_array_tile
*tile
;
3285 tile
= gpu_array_ref_group_tile(group
);
3286 node
= isl_schedule_node_parent(node
);
3287 node
= isl_schedule_node_parent(node
);
3288 node
= gpu_tree_move_down_to_depth(node
, tile
->depth
,
3290 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3296 /* Add copy statements to the schedule tree of "node"
3297 * for reading from global memory to private memory (if "read" is set) or
3298 * for writing back from private memory to global memory
3299 * (if "read" is not set) for the array reference group "group" that
3300 * is mapped to private memory.
3301 * On input, "node" points to the kernel node, and it is moved
3302 * back there on output.
3304 * The copies are performed in the order of the array elements.
3305 * The copy statement instances include a reference to the outer
3306 * tile->depth dimensions of the kernel schedule for ease of
3307 * combining them with the group tiling.
3309 * That is, the extra schedule is of the form
3313 * where D corresponds to the outer tile->depth dimensions of
3314 * the kernel schedule and A to the global array.
3315 * This schedule is unrolled because registers are not addressable.
3317 * The copying is inserted in the schedule tree through an extension
3322 * where the extra domain elements type[D -> A] are those accessed
3324 * A filter is inserted on type[D -> A] to ensure that the element
3325 * is read/written by the same thread that needs the element.
3326 * This filter is obtained by applying
3330 * to the thread filter for the core statements.
3332 * The extension is inserted before the core computation in case of a read
3333 * and after the core computation in case of a write.
3334 * In the latter case, we also make sure that there is a synchronization
3335 * node after the write to global memory, unless this write is performed
3336 * at the outer level of the kernel.
3337 * In principle, this synchronization could be inserted higher
3338 * in the schedule tree depending on where the corresponding reads
3339 * from global memory are performed.
3341 static __isl_give isl_schedule_node
*add_copies_group_private(
3342 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3343 __isl_take isl_schedule_node
*node
, int read
)
3345 struct gpu_array_tile
*tile
;
3346 isl_union_map
*access
;
3347 isl_union_set
*domain
;
3349 isl_multi_aff
*from_access
;
3350 isl_multi_pw_aff
*mpa
;
3351 isl_multi_union_pw_aff
*mupa
;
3352 isl_union_pw_multi_aff
*contraction
;
3353 isl_schedule_node
*graft
;
3354 isl_union_set
*filter
;
3358 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3359 tile
= gpu_array_ref_group_tile(group
);
3360 node
= gpu_tree_move_down_to_depth(node
, tile
->depth
, kernel
->core
);
3362 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3363 empty
= isl_union_map_is_empty(access
);
3364 if (empty
< 0 || empty
) {
3365 isl_union_map_free(access
);
3367 return isl_schedule_node_free(node
);
3368 return gpu_tree_move_up_to_kernel(node
);
3371 group
->array
->global
= 1;
3372 group
->local_array
->global
= 1;
3374 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3375 space
= isl_space_domain(isl_multi_aff_get_space(from_access
));
3376 access
= isl_union_map_preimage_range_multi_aff(access
, from_access
);
3378 filter
= isl_union_set_copy(kernel
->thread_filter
);
3379 contraction
= isl_union_pw_multi_aff_copy(kernel
->contraction
);
3380 filter
= isl_union_set_preimage_union_pw_multi_aff(filter
, contraction
);
3381 filter
= isl_union_set_apply(filter
, isl_union_map_copy(access
));
3382 filter
= isl_union_set_detect_equalities(filter
);
3383 filter
= isl_union_set_coalesce(filter
);
3385 domain
= isl_union_map_range(access
);
3386 access
= isl_union_set_wrapped_domain_map(domain
);
3387 access
= isl_union_map_reverse(access
);
3388 access
= isl_union_map_coalesce(access
);
3389 graft
= isl_schedule_node_from_extension(access
);
3391 space
= isl_space_map_from_set(space
);
3392 mpa
= isl_multi_pw_aff_identity(space
);
3393 mpa
= isl_multi_pw_aff_range_factor_range(mpa
);
3394 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3396 graft
= isl_schedule_node_child(graft
, 0);
3397 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3398 graft
= unroll(graft
);
3400 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3402 graft
= isl_schedule_node_parent(graft
);
3405 node
= isl_schedule_node_graft_before(node
, graft
);
3407 node
= isl_schedule_node_graft_after(node
, graft
);
3408 if (kernel_depth
< tile
->depth
)
3409 node
= add_group_write_sync(node
, kernel
, group
, 0);
3412 node
= gpu_tree_move_up_to_kernel(node
);
3417 /* Add copy statements to the schedule tree of "node"
3418 * for reading from global memory to shared memory (if "read" is set) or
3419 * for writing back from shared memory to global memory
3420 * (if "read" is not set) for the array reference group "group" that
3421 * is mapped to shared memory.
3422 * On input, "node" points to the kernel node, and it is moved
3423 * back there on output.
3425 * The copies are performed in the order of the corresponding shared
3427 * The copy statement instances include a reference to the outer
3428 * tile->depth dimensions of the kernel schedule for ease of
3429 * combining them with the group tiling.
3431 * If we are performing a read from global memory to shared memory and
3432 * if the array involved is not a scalar, then we copy
3433 * the entire tile to shared memory. This may result in some extra
3434 * elements getting copied, but it should lead to simpler code
3435 * (which means that fewer registers may be needed) and less divergence.
3437 * Otherwise, we only copy the elements that will be read or have been written
3440 * That is, the extra schedule is of the form
3444 * where D corresponds to the outer tile->depth dimensions of
3445 * the kernel schedule, A to the global array and T is the corresponding
3446 * shared memory tile.
3448 * The copying is inserted in the schedule tree through an extension
3453 * where the extra domain elements type[D -> A] are those accessed
3454 * by the group. In the case of read from a non-scalar, this set
3455 * is replaced by the entire shared memory tile.
3457 * If the "unroll_copy_shared" option is set, then the AST generator
3458 * is instructed to unroll the copying code.
3460 * A filter is inserted on type[D -> A] to map the copy instances
3461 * to the threads. In particular, the thread identifiers are
3462 * equated to the position inside the shared memory tile (T)
3463 * modulo the block size.
3464 * We try to align the innermost tile dimension with the innermost
3465 * thread identifier (x) as a heuristic to improve coalescing.
3466 * In particular, if the dimension of the tile is greater than
3467 * the dimension of the block, then the schedule mapping to the tile
3468 * is broken up into two pieces and the filter is applied to the inner part.
3469 * If, on the other hand, the dimension of the tile is smaller than
3470 * the dimension of the block, then the initial thread identifiers
3471 * are equated to zero and the remaining thread identifiers are
3472 * matched to the memory tile.
3474 * The extension is inserted before the core computation in case of a read
3475 * and after the core computation in case of a write.
3476 * In the case of a read, we first need to make sure there is some
3477 * synchronization before the core computation such that we can put the read
3478 * from global memory to shared memory before that synchronization.
3479 * This ensures that all threads have finished copying into shared memory
3480 * before the shared memory is used.
3481 * We also need to make sure that there is a synchronization node after
3482 * the core computation to ensure that the next load into shared memory
3483 * only happens after all data has been used. There is no need for
3484 * this synchronization if we are at the outer level since then there
3485 * won't be a next load.
3486 * In the case of a write, we need to make sure there is some synchronization
3487 * after the core computation such that we can put the write from shared
3488 * memory to global memory after that synchronization.
3489 * Unless we are at the outer level, we also need a synchronization node
3490 * after the write to ensure the data is saved to global memory
3491 * before the next iteration writes to the same shared memory.
3492 * It also makes sure the data has arrived in global memory before
3493 * it is read in a subsequent iteration.
3495 static __isl_give isl_schedule_node
*add_copies_group_shared(
3496 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3497 __isl_take isl_schedule_node
*node
, int read
)
3499 struct gpu_array_tile
*tile
;
3500 isl_union_map
*access
;
3501 isl_union_set
*domain
;
3503 isl_multi_aff
*from_access
;
3504 isl_multi_pw_aff
*mpa
;
3505 isl_multi_union_pw_aff
*mupa
;
3506 isl_schedule_node
*graft
;
3507 isl_union_set
*filter
;
3512 tile
= gpu_array_ref_group_tile(group
);
3513 kernel_depth
= isl_schedule_node_get_schedule_depth(node
);
3514 node
= gpu_tree_move_down_to_depth(node
, tile
->depth
, kernel
->core
);
3516 access
= anchored_non_local_accesses(kernel
, group
, node
, read
);
3517 empty
= isl_union_map_is_empty(access
);
3518 if (empty
< 0 || empty
) {
3519 isl_union_map_free(access
);
3521 return isl_schedule_node_free(node
);
3522 return gpu_tree_move_up_to_kernel(node
);
3525 group
->array
->global
= 1;
3526 group
->local_array
->global
= 1;
3528 from_access
= create_from_access(kernel
->ctx
, group
, read
);
3530 ma
= isl_multi_aff_copy(tile
->tiling
);
3531 ma
= isl_multi_aff_pullback_multi_aff(ma
,
3532 isl_multi_aff_copy(from_access
));
3533 mpa
= isl_multi_pw_aff_from_multi_aff(ma
);
3534 mupa
= isl_multi_union_pw_aff_from_multi_pw_aff(mpa
);
3536 domain
= isl_union_map_range(access
);
3538 if (read
&& !gpu_array_is_scalar(group
->array
)) {
3540 isl_union_set_free(domain
);
3541 map
= group_tile(group
);
3542 domain
= isl_union_set_from_set(isl_map_wrap(map
));
3545 domain
= isl_union_set_preimage_multi_aff(domain
, from_access
);
3546 access
= isl_union_set_wrapped_domain_map(domain
);
3547 access
= isl_union_map_reverse(access
);
3548 access
= isl_union_map_coalesce(access
);
3549 graft
= isl_schedule_node_from_extension(access
);
3551 graft
= isl_schedule_node_child(graft
, 0);
3553 graft
= isl_schedule_node_insert_partial_schedule(graft
, mupa
);
3554 if (kernel
->options
->unroll_copy_shared
)
3555 graft
= ppcg_set_schedule_node_type(graft
, isl_ast_loop_unroll
);
3557 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0) {
3558 graft
= isl_schedule_node_band_split(graft
,
3559 tile
->n
- kernel
->n_block
);
3560 graft
= isl_schedule_node_child(graft
, 0);
3562 if (tile
->n
< kernel
->n_block
)
3563 skip
= kernel
->n_block
- tile
->n
;
3566 filter
= set_schedule_modulo(graft
, kernel
->thread_ids
,
3568 if (!kernel
->options
->wrap
)
3569 graft
= snap_band_to_sizes(graft
, kernel
->block_dim
+ skip
,
3571 if (tile
->n
> kernel
->n_block
&& kernel
->n_block
> 0)
3572 graft
= isl_schedule_node_parent(graft
);
3573 graft
= isl_schedule_node_insert_filter(graft
, filter
);
3575 while (graft
&& isl_schedule_node_has_parent(graft
))
3576 graft
= isl_schedule_node_parent(graft
);
3579 if (kernel_depth
< tile
->depth
)
3580 node
= gpu_tree_ensure_sync_after_core(node
, kernel
);
3581 node
= gpu_tree_move_left_to_sync(node
, kernel
);
3582 node
= isl_schedule_node_graft_before(node
, graft
);
3584 node
= gpu_tree_move_right_to_sync(node
, kernel
);
3585 node
= isl_schedule_node_graft_after(node
, graft
);
3586 if (kernel_depth
< tile
->depth
)
3587 node
= add_group_write_sync(node
, kernel
, group
, 1);
3590 node
= gpu_tree_move_up_to_kernel(node
);
3595 /* Check whether the array reference group "group" is mapped to
3596 * private or shared memory and, if so,
3597 * add copy statements to the schedule tree of "node"
3598 * for reading from global memory to private or shared memory
3599 * (if "read" is set) or for writing back from private or shared memory
3600 * to global memory (if "read" is not set) for this group.
3601 * On input, "node" points to the kernel node, and it is moved
3602 * back there on output.
3604 static __isl_give isl_schedule_node
*add_copies_group(
3605 struct ppcg_kernel
*kernel
, struct gpu_array_ref_group
*group
,
3606 __isl_take isl_schedule_node
*node
, int read
)
3608 enum ppcg_group_access_type type
;
3610 type
= gpu_array_ref_group_type(group
);
3611 if (type
== ppcg_access_private
)
3612 return add_copies_group_private(kernel
, group
, node
, read
);
3613 if (type
== ppcg_access_shared
)
3614 return add_copies_group_shared(kernel
, group
, node
, read
);
3618 /* For each array reference group that is mapped to private or shared memory,
3619 * add copy statements to the schedule tree of "node"
3620 * for reading from global memory to private or shared memory
3621 * and for writing back.
3622 * On input, "node" points to the kernel node, and it is moved
3623 * back there on output.
3625 static __isl_give isl_schedule_node
*add_copies(struct ppcg_kernel
*kernel
,
3626 __isl_take isl_schedule_node
*node
)
3630 for (i
= 0; i
< kernel
->n_array
; ++i
) {
3631 struct gpu_local_array_info
*array
= &kernel
->array
[i
];
3633 for (j
= 0; j
< array
->n_group
; ++j
) {
3634 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3636 node
= add_copies_group(kernel
, group
, node
, 1);
3639 node
= add_copies_group(kernel
, group
, node
, 0);
3648 /* Mark all dimensions in the current band node atomic.
3650 static __isl_give isl_schedule_node
*atomic(__isl_take isl_schedule_node
*node
)
3652 return ppcg_set_schedule_node_type(node
, isl_ast_loop_atomic
);
3655 /* Mark "node" atomic, if it is a band node.
3656 * Do the same for all ancestors.
3657 * Return a pointer to "node" (in the updated schedule tree).
3659 static __isl_give isl_schedule_node
*atomic_ancestors(
3660 __isl_take isl_schedule_node
*node
)
3666 if (!isl_schedule_node_has_parent(node
))
3669 pos
= isl_schedule_node_get_child_position(node
);
3670 node
= isl_schedule_node_parent(node
);
3671 if (isl_schedule_node_get_type(node
) == isl_schedule_node_band
)
3672 node
= atomic(node
);
3673 node
= atomic_ancestors(node
);
3674 node
= isl_schedule_node_child(node
, pos
);
3679 /* Collect all write references that require synchronization.
3680 * "node" is assumed to point to the kernel node.
3681 * Each reference is represented by a universe set in a space
3685 * with S[i,j] the statement instance space and R[] the array reference.
3687 * This function should be called before block and thread filters are added.
3689 * Synchronization is needed after a write if there is a subsequent read
3690 * within the same block that may not be performed by the same thread.
3691 * There should not be any dependences between different blocks,
3692 * so we start with the flow dependences within the same kernel invocation
3693 * and we subtract from these those dependences that are mapped
3694 * to the same iteration of the bands where synchronization is inserted.
3695 * We do not remove pairs of instances that are known to map to
3696 * the same thread across different iterations of the intermediate
3697 * bands because the read may be performed by a different thread
3698 * than the one that needs the value if shared memory is involved.
3700 * We also consider all pairs of possible writes that access the same
3701 * memory location and that may be mapped to the same block but not
3702 * to the same iteration of the intermediate bands.
3703 * In theory, it would be possible for one thread to still be in
3704 * a previous iteration of a loop in these bands.
3705 * A write to global memory in this delayed thread could then overwrite
3706 * a write from another thread that has already moved on to
3707 * the next iteration.
3709 * After computing the above writes paired off with reads or writes
3710 * that depend on them, we project onto the domain writes.
3711 * Sychronization is needed after writes to global memory
3712 * through these references.
3714 static __isl_give isl_union_set
*compute_sync_writes(
3715 struct ppcg_kernel
*kernel
, __isl_keep isl_schedule_node
*node
)
3717 isl_union_map
*local
;
3718 isl_union_map
*may_writes
, *shared_access
;
3719 isl_union_map
*kernel_prefix
, *thread_prefix
;
3720 isl_union_map
*equal
;
3721 isl_union_set
*wrap
;
3722 isl_union_set
*domain
;
3723 isl_union_pw_multi_aff
*contraction
;
3725 kernel_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3726 node
= isl_schedule_node_copy(node
);
3727 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3728 thread_prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3729 isl_schedule_node_free(node
);
3731 contraction
= kernel
->contraction
;
3732 kernel_prefix
= isl_union_map_preimage_domain_union_pw_multi_aff(
3733 kernel_prefix
, isl_union_pw_multi_aff_copy(contraction
));
3734 thread_prefix
= isl_union_map_preimage_domain_union_pw_multi_aff(
3735 thread_prefix
, isl_union_pw_multi_aff_copy(contraction
));
3736 domain
= isl_union_set_copy(kernel
->expanded_domain
);
3737 domain
= isl_union_set_universe(domain
);
3739 may_writes
= isl_union_map_copy(kernel
->prog
->scop
->tagged_may_writes
);
3740 may_writes
= isl_union_map_curry(may_writes
);
3741 may_writes
= isl_union_map_intersect_domain(may_writes
, domain
);
3742 may_writes
= isl_union_map_uncurry(may_writes
);
3743 shared_access
= isl_union_map_copy(may_writes
);
3744 shared_access
= isl_union_map_apply_range(shared_access
,
3745 isl_union_map_reverse(may_writes
));
3747 local
= isl_union_map_copy(kernel
->prog
->scop
->tagged_dep_flow
);
3748 local
= isl_union_map_union(local
, shared_access
);
3749 local
= isl_union_map_zip(local
);
3751 equal
= isl_union_map_apply_range(kernel_prefix
,
3752 isl_union_map_reverse(isl_union_map_copy(kernel_prefix
)));
3753 wrap
= isl_union_map_wrap(equal
);
3754 local
= isl_union_map_intersect_domain(local
, wrap
);
3755 equal
= isl_union_map_apply_range(thread_prefix
,
3756 isl_union_map_reverse(isl_union_map_copy(thread_prefix
)));
3757 wrap
= isl_union_map_wrap(equal
);
3758 local
= isl_union_map_subtract_domain(local
, wrap
);
3760 local
= isl_union_map_zip(local
);
3761 local
= isl_union_map_universe(local
);
3763 return isl_union_map_domain(local
);
3766 /* Group the domain elements into a single space, named kernelX,
3767 * with X the kernel sequence number "kernel_id".
3769 static __isl_give isl_schedule_node
*group_statements(
3770 __isl_take isl_schedule_node
*node
, int kernel_id
)
3778 snprintf(buffer
, sizeof(buffer
), "kernel%d", kernel_id
);
3779 id
= isl_id_alloc(isl_schedule_node_get_ctx(node
), buffer
, NULL
);
3780 return isl_schedule_node_group(node
, id
);
3783 /* Create a ppcg_kernel representing the domain instances that reach "node"
3784 * and insert a mark node pointing to the ppcg_kernel before "node".
3785 * The band that "node" points to is the band that needs to be mapped
3786 * to block identifiers. The band that needs to be mapped to thread
3787 * identifiers should be marked by a "thread" mark by the caller.
3788 * The linear branch between the current node and the "thread" mark
3789 * may also have a "shared" mark. If present, the mapping to shared
3790 * memory is computed at that point.
3791 * Both marks are removed by this function.
3792 * If "scale" is set, then the band that "node" points to is scaled
3795 * Mark all outer band nodes as atomic to ensure each kernel is only
3797 * If the domain elements that reach "node" live in more than one space,
3798 * then group the domain elements into a single space, named kernelX,
3799 * with X the kernel sequence number.
3801 * Insert a guard node governing the kernel node to ensure that
3802 * no kernels with zero blocks are launched.
3804 * Insert a context node describing the block and thread
3805 * identifiers inside the kernel mark.
3806 * The context node needs to be inserted after the effective block size
3807 * has been determined such that the bounds on the thread identifiers
3808 * would reflect the effective block size.
3809 * Insert a filter node inside the context node mapping the statement
3810 * instances to block identifiers. In particular, the block identifiers
3811 * are equated to the partial schedule of band that was marked for mapping
3812 * to blocks modulo the grid size.
3813 * Insert a filter node inside the "thread" mark mapping the statement
3814 * instances to thread identifiers. In particular, the thread identifiers
3815 * are equated to the partial schedule of band that was marked for mapping
3816 * to threads modulo the block size.
3818 * Compute array reference groups for all arrays, set the local
3819 * array bounds based on the set of domain instances that reach
3820 * the kernel node, check the total amount of shared memory used
3821 * and compute all group tilings.
3822 * The array reference groups are computed after the block filter
3823 * has been inserted because it affects the mapping to shared or
3824 * private memory. This computation also requires the thread filter
3825 * (in the ppcg_kernel object), but this thread filter should not
3826 * have been added to the schedule tree yet since the computation
3827 * requires the schedule of the band that needs to be mapped to
3828 * threads before the privatization is applied.
3830 * If any array reference group requires the band mapped to threads
3831 * to be unrolled, then we perform the required unrolling.
3833 * We save a copy of the schedule that may influence the mappings
3834 * to shared or private memory in kernel->copy_schedule.
3836 * Finally, we add synchronization and copy statements to the schedule tree,
3837 * remove the "thread" mark and create representations for the local
3838 * variables in the kernel.
3840 * We keep a copy of the isl_id that points to the kernel to ensure
3841 * that the kernel does not get destroyed if the schedule node
3842 * is freed due to some error condition.
3844 __isl_give isl_schedule_node
*gpu_create_kernel(struct gpu_gen
*gen
,
3845 __isl_take isl_schedule_node
*node
, int scale
,
3846 __isl_keep isl_multi_val
*sizes
)
3848 struct ppcg_kernel
*kernel
;
3850 isl_schedule_node
*node_thread
;
3851 isl_union_map
*host_schedule
;
3852 isl_union_pw_multi_aff
*contraction
;
3853 isl_set
*host_domain
;
3854 isl_union_set
*domain
, *expanded
;
3855 int single_statement
;
3857 node
= gpu_tree_insert_shared_before_thread(node
);
3861 kernel
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel
);
3862 kernel
= ppcg_kernel_create_local_arrays(kernel
, gen
->prog
);
3864 return isl_schedule_node_free(node
);
3866 domain
= isl_schedule_node_get_domain(node
);
3867 single_statement
= isl_union_set_n_set(domain
) == 1;
3869 kernel
->ctx
= gen
->ctx
;
3870 kernel
->prog
= gen
->prog
;
3871 kernel
->options
= gen
->options
;
3872 kernel
->context
= extract_context(node
, gen
->prog
);
3873 kernel
->core
= isl_union_set_universe(isl_union_set_copy(domain
));
3874 contraction
= isl_schedule_node_get_subtree_contraction(node
);
3875 kernel
->contraction
= isl_union_pw_multi_aff_copy(contraction
);
3876 expanded
= isl_union_set_copy(domain
);
3877 expanded
= isl_union_set_preimage_union_pw_multi_aff(expanded
,
3879 kernel
->expanded_domain
= isl_union_set_copy(expanded
);
3880 kernel
->arrays
= accessed_by_domain(expanded
, gen
->prog
);
3881 kernel
->n_grid
= n_outer_coincidence(node
);
3882 node_thread
= isl_schedule_node_copy(node
);
3883 node_thread
= gpu_tree_move_down_to_thread(node_thread
, kernel
->core
);
3884 node_thread
= isl_schedule_node_child(node_thread
, 0);
3885 kernel
->n_block
= n_outer_coincidence(node_thread
);
3886 isl_schedule_node_free(node_thread
);
3887 kernel
->id
= gen
->kernel_id
++;
3888 if (read_grid_and_block_sizes(kernel
, gen
) < 0)
3889 node
= isl_schedule_node_free(node
);
3891 kernel
->sync_writes
= compute_sync_writes(kernel
, node
);
3893 host_schedule
= isl_schedule_node_get_prefix_schedule_union_map(node
);
3894 host_domain
= isl_set_from_union_set(isl_union_map_range(
3897 node
= atomic_ancestors(node
);
3899 id
= isl_id_alloc(gen
->ctx
, "kernel", kernel
);
3900 id
= isl_id_set_free_user(id
, &ppcg_kernel_free_wrap
);
3901 node
= isl_schedule_node_insert_mark(node
, isl_id_copy(id
));
3903 if (!single_statement
)
3904 node
= group_statements(node
, kernel
->id
);
3906 node
= isl_schedule_node_child(node
, 0);
3907 node
= split_band(node
, kernel
->n_grid
);
3908 kernel
->block_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3909 kernel
->n_grid
, "b");
3910 kernel
->block_filter
= set_schedule_modulo(node
, kernel
->block_ids
,
3912 kernel
->grid_size
= extract_grid_size(kernel
,
3913 isl_union_set_copy(domain
));
3914 if (!kernel
->options
->wrap
)
3915 node
= snap_band_to_sizes(node
, kernel
->grid_dim
,
3918 node
= scale_band(node
, isl_multi_val_copy(sizes
));
3919 node
= isl_schedule_node_parent(node
);
3920 if (!single_statement
)
3921 node
= isl_schedule_node_parent(node
);
3922 node
= insert_guard(node
, kernel
->context
, kernel
->grid_size
,
3924 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3925 node
= isl_schedule_node_child(node
, 0);
3926 node
= split_band(node
, kernel
->n_block
);
3927 kernel
->thread_ids
= ppcg_scop_generate_names(gen
->prog
->scop
,
3928 kernel
->n_block
, "t");
3929 kernel
->thread_filter
= set_schedule_modulo(node
, kernel
->thread_ids
,
3931 if (extract_block_size(kernel
, domain
) < 0)
3932 node
= isl_schedule_node_free(node
);
3934 node
= gpu_tree_move_up_to_kernel(node
);
3935 node
= isl_schedule_node_child(node
, 0);
3936 node
= insert_context(kernel
, node
);
3937 node
= isl_schedule_node_child(node
, 0);
3938 node
= isl_schedule_node_insert_filter(node
,
3939 isl_union_set_copy(kernel
->block_filter
));
3941 node
= gpu_tree_move_up_to_kernel(node
);
3943 if (gpu_group_references(kernel
, node
) < 0)
3944 node
= isl_schedule_node_free(node
);
3945 localize_bounds(kernel
, host_domain
);
3946 isl_set_free(host_domain
);
3948 check_shared_memory_bound(kernel
);
3949 mark_global_arrays(kernel
);
3950 compute_group_tilings(kernel
);
3952 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3953 node
= isl_schedule_node_child(node
, 0);
3954 if (!kernel
->options
->wrap
)
3955 node
= snap_band_to_sizes(node
, kernel
->block_dim
,
3957 node
= isl_schedule_node_insert_filter(node
,
3958 isl_union_set_copy(kernel
->thread_filter
));
3959 if (kernel_requires_unroll(kernel
)) {
3960 node
= isl_schedule_node_child(node
, 0);
3961 node
= unroll(node
);
3964 node
= gpu_tree_move_up_to_thread(node
);
3965 kernel
->copy_schedule_dim
= isl_schedule_node_get_schedule_depth(node
);
3966 kernel
->copy_schedule
=
3967 isl_schedule_node_get_prefix_schedule_union_pw_multi_aff(node
);
3968 contraction
= isl_union_pw_multi_aff_copy(kernel
->contraction
);
3969 kernel
->copy_schedule
=
3970 isl_union_pw_multi_aff_pullback_union_pw_multi_aff(
3971 kernel
->copy_schedule
, contraction
);
3973 node
= gpu_tree_move_up_to_kernel(node
);
3975 node
= add_sync(kernel
, node
);
3976 node
= add_copies(kernel
, node
);
3978 node
= gpu_tree_move_down_to_shared(node
, kernel
->core
);
3979 node
= isl_schedule_node_delete(node
);
3981 node
= gpu_tree_move_down_to_thread(node
, kernel
->core
);
3982 node
= isl_schedule_node_delete(node
);
3984 node
= gpu_tree_move_up_to_kernel(node
);
3986 if (create_kernel_vars(kernel
) < 0)
3987 node
= isl_schedule_node_free(node
);
3989 if (!single_statement
)
3990 node
= isl_schedule_node_parent(node
);
3991 node
= isl_schedule_node_parent(node
);
3995 ppcg_kernel_free(kernel
);
3999 /* Insert a zero-dimensional permutable band at "node",
4000 * defined over the spaces of the statements that reach this node.
4002 static __isl_give isl_schedule_node
*insert_empty_permutable_band(
4003 __isl_take isl_schedule_node
*node
)
4006 isl_union_set
*domain
;
4007 isl_multi_union_pw_aff
*mupa
;
4009 domain
= isl_schedule_node_get_universe_domain(node
);
4010 space
= isl_union_set_get_space(domain
);
4012 space
= isl_space_set_from_params(space
);
4013 mupa
= isl_multi_union_pw_aff_zero(space
);
4014 mupa
= isl_multi_union_pw_aff_intersect_domain(mupa
, domain
);
4015 node
= isl_schedule_node_insert_partial_schedule(node
, mupa
);
4016 node
= isl_schedule_node_band_set_permutable(node
, 1);
4021 /* See if hybrid tiling can be performed on "node" and its parent.
4022 * If so, apply hybrid tiling and return the updated schedule tree.
4023 * If not, return the original schedule tree.
4024 * Return NULL on error.
4026 * First check if "node", together with its parent, meets
4027 * the basic requirements for hybrid tiling.
4028 * If so, compute the relative dependence distances of "node"
4029 * with respect to its parent and check if they are sufficiently bounded.
4030 * If so, apply hybrid tiling using user specified tile sizes.
4032 * The tile sizes are read before the dependence distance bounds are
4033 * computed, because the user may have specified fewer dimensions
4034 * than are available. In this case, the remaining schedule dimensions
4035 * are split off and the dependence distances should be computed
4036 * after these dimensions have been split off.
4038 static __isl_give isl_schedule_node
*try_hybrid_tile(struct gpu_gen
*gen
,
4039 __isl_take isl_schedule_node
*node
)
4044 isl_schedule_node
*orig
= node
;
4045 ppcg_ht_bounds
*bounds
;
4047 ok
= ppcg_ht_parent_has_input_pattern(node
);
4049 return isl_schedule_node_free(node
);
4053 tile_len
= 1 + isl_schedule_node_band_n_member(node
);
4054 tile_size
= read_tile_sizes(gen
, &tile_len
);
4056 return isl_schedule_node_free(node
);
4058 node
= isl_schedule_node_copy(node
);
4059 node
= split_band(node
, tile_len
- 1);
4060 node
= isl_schedule_node_parent(node
);
4061 bounds
= ppcg_ht_compute_bounds(gen
->prog
->scop
, node
);
4062 node
= isl_schedule_node_child(node
, 0);
4064 ok
= ppcg_ht_bounds_is_valid(bounds
);
4066 node
= gpu_hybrid_tile(gen
, node
, bounds
, tile_size
);
4068 ppcg_ht_bounds_free(bounds
);
4071 if (ok
>= 0 && !ok
) {
4072 isl_schedule_node_free(node
);
4075 isl_schedule_node_free(orig
);
4077 return isl_schedule_node_free(node
);
4081 /* If "node" is the outermost permutable band that can be mapped to block and
4082 * thread identifiers in its branch (or the root of a subtree with
4083 * no such outer bands),
4084 * then mark the band as such, attaching a ppcg_kernel to the mark.
4086 * If hybrid tiling is allowed, then first try and apply it
4087 * to "node" and its parent.
4089 * If "node" is the root of a subtree without permutable bands,
4090 * then insert a zero-dimensional permutable band such that
4091 * we can assume that "node" always points to a band node.
4092 * This includes the case where "node" already points to a band node,
4093 * but one without any coincident dimension. In this case,
4094 * the extra node ensures that this original node does not get tiled.
4096 * Tile "node" using user specified tile sizes, after splitting the band
4097 * if the number of specified tile sizes is smaller than the dimension
4098 * of the band. Mark the point band of this tiling as the band that
4099 * needs to be mapped to threads and instruct the AST generator to unroll
4100 * the band if the "unroll_gpu_tile" option is set.
4101 * Create a kernel representing the domain instances that reach "node" and
4102 * insert a mark node pointing to the ppcg_kernel before the band node.
4104 static __isl_give isl_schedule_node
*mark_outer_permutable(
4105 __isl_take isl_schedule_node
*node
, void *user
)
4107 struct gpu_gen
*gen
= user
;
4113 isl_multi_val
*sizes
;
4115 outer
= is_outer_tilable(node
);
4117 return isl_schedule_node_free(node
);
4121 if (gen
->options
->hybrid
) {
4122 isl_schedule_node
*saved
= isl_schedule_node_copy(node
);
4123 node
= try_hybrid_tile(gen
, node
);
4124 isl_schedule_node_free(saved
);
4129 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
||
4130 !isl_schedule_node_band_member_get_coincident(node
, 0))
4131 node
= insert_empty_permutable_band(node
);
4133 tile_len
= isl_schedule_node_band_n_member(node
);
4134 tile_size
= read_tile_sizes(gen
, &tile_len
);
4136 return isl_schedule_node_free(node
);
4137 if (tile_len
< isl_schedule_node_band_n_member(node
))
4138 node
= isl_schedule_node_band_split(node
, tile_len
);
4139 sizes
= construct_band_tiles_sizes(node
, tile_size
);
4140 node
= tile_band(node
, isl_multi_val_copy(sizes
));
4141 node
= isl_schedule_node_child(node
, 0);
4142 if (gen
->options
->unroll_gpu_tile
)
4143 node
= ppcg_set_schedule_node_type(node
, isl_ast_loop_unroll
);
4144 id
= isl_id_alloc(gen
->ctx
, "thread", NULL
);
4145 node
= isl_schedule_node_insert_mark(node
, id
);
4146 node
= isl_schedule_node_parent(node
);
4148 scale
= gen
->options
->scale_tile_loops
;
4149 node
= gpu_create_kernel(gen
, node
, scale
, sizes
);
4150 isl_multi_val_free(sizes
);
4156 /* Given a set or sequence node, return the union the filters of either all
4157 * (if "only_initial" is not set) or the initial (if "only_initial" is set)
4158 * direct subtrees that do not contain any suitably permutable bands
4159 * (according to subtree_has_permutable_bands).
4161 static __isl_give isl_union_set
*get_non_parallel_subtree_filters(
4162 __isl_keep isl_schedule_node
*node
, int only_initial
)
4165 isl_union_set
*filter
;
4168 n
= isl_schedule_node_n_children(node
);
4172 node
= isl_schedule_node_copy(node
);
4173 node
= isl_schedule_node_child(node
, 0);
4174 filter
= isl_schedule_node_filter_get_filter(node
);
4175 node
= isl_schedule_node_parent(node
);
4176 space
= isl_union_set_get_space(filter
);
4177 isl_union_set_free(filter
);
4178 filter
= isl_union_set_empty(space
);
4180 for (i
= 0; i
< n
; ++i
) {
4183 node
= isl_schedule_node_child(node
, i
);
4184 parallelism
= subtree_has_permutable_bands(node
);
4185 if (parallelism
< 0) {
4186 filter
= isl_union_set_free(filter
);
4187 } else if (!parallelism
) {
4188 isl_union_set
*filter_i
;
4189 filter_i
= isl_schedule_node_filter_get_filter(node
);
4190 filter
= isl_union_set_union(filter
, filter_i
);
4191 } else if (only_initial
)
4193 node
= isl_schedule_node_parent(node
);
4196 isl_schedule_node_free(node
);
4201 /* Given a set or sequence node, return the union of the filters of
4202 * the direct subtrees that do not contain any suitably permutable bands
4203 * (according to subtree_has_permutable_bands).
4205 static __isl_give isl_union_set
*get_all_non_parallel_subtree_filters(
4206 __isl_keep isl_schedule_node
*node
)
4208 return get_non_parallel_subtree_filters(node
, 0);
4211 /* Given a set or sequence node, return the union of the filters of
4212 * the initial direct subtrees that do not contain any suitably permutable
4213 * bands (according to subtree_has_permutable_bands).
4215 static __isl_give isl_union_set
*get_initial_non_parallel_subtree_filters(
4216 __isl_keep isl_schedule_node
*node
)
4218 return get_non_parallel_subtree_filters(node
, 1);
4221 /* Mark all variables that are accessed by the statement instances in "domain"
4222 * and that are local to "prog" as requiring a declaration in the host code.
4223 * The statement instances in "domain" correspond to (a subset of)
4224 * the active instances at "node".
4225 * "node" is not modified by this function, except that NULL is returned
4228 static __isl_give isl_schedule_node
*declare_accessed_local_variables(
4229 __isl_take isl_schedule_node
*node
, struct gpu_prog
*prog
,
4230 __isl_keep isl_union_set
*domain
)
4232 isl_union_pw_multi_aff
*contraction
;
4233 isl_union_set
*arrays
;
4236 if (!ppcg_scop_any_hidden_declarations(prog
->scop
))
4238 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4239 domain
= isl_union_set_copy(domain
);
4240 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
, contraction
);
4241 arrays
= accessed_by_domain(domain
, prog
);
4243 for (i
= 0; i
< prog
->n_array
; ++i
) {
4248 if (!prog
->array
[i
].local
)
4250 space
= isl_set_get_space(prog
->array
[i
].extent
);
4251 set
= isl_union_set_extract_set(arrays
, space
);
4252 empty
= isl_set_plain_is_empty(set
);
4257 prog
->array
[i
].declare_local
= 1;
4260 isl_union_set_free(arrays
);
4263 isl_union_set_free(arrays
);
4264 return isl_schedule_node_free(node
);
4267 /* If "node" points to a set node, then separate its children
4268 * into subtrees that have suitably permutable bands and
4269 * those that do not.
4270 * Adjust the schedule tree in order to execute the second group
4271 * after the first group and return a pointer to the first group,
4272 * assuming there are any such subtrees.
4273 * If "node" points to a sequence node, then separate the initial
4274 * children that do not have suitably permutable bands and
4275 * return a pointer to the subsequence of children that do have such bands,
4276 * assuming there are any such subtrees.
4278 * In both cases, mark all local variables in "prog" that are accessed by
4279 * the group without permutable bands as requiring a declaration on the host.
4281 static __isl_give isl_schedule_node
*isolate_permutable_subtrees(
4282 __isl_take isl_schedule_node
*node
, struct gpu_prog
*prog
)
4284 isl_union_set
*filter
;
4285 enum isl_schedule_node_type type
;
4289 type
= isl_schedule_node_get_type(node
);
4290 if (type
== isl_schedule_node_set
) {
4291 filter
= get_all_non_parallel_subtree_filters(node
);
4292 node
= declare_accessed_local_variables(node
, prog
, filter
);
4293 node
= isl_schedule_node_order_after(node
, filter
);
4294 } else if (type
== isl_schedule_node_sequence
) {
4295 filter
= get_initial_non_parallel_subtree_filters(node
);
4296 node
= declare_accessed_local_variables(node
, prog
, filter
);
4297 node
= isl_schedule_node_order_before(node
, filter
);
4303 /* Replace any reference to an array element in the range of "copy"
4304 * by a reference to all array elements (defined by the extent of the array).
4306 static __isl_give isl_union_map
*approximate_copy_out(
4307 __isl_take isl_union_map
*copy
, struct gpu_prog
*prog
)
4312 res
= isl_union_map_empty(isl_union_map_get_space(copy
));
4314 for (i
= 0; i
< prog
->n_array
; ++i
) {
4317 isl_union_map
*copy_i
;
4318 isl_union_set
*extent
, *domain
;
4320 space
= isl_space_copy(prog
->array
[i
].space
);
4321 extent
= isl_union_set_from_set(isl_set_universe(space
));
4322 copy_i
= isl_union_map_copy(copy
);
4323 copy_i
= isl_union_map_intersect_range(copy_i
, extent
);
4324 set
= isl_set_copy(prog
->array
[i
].extent
);
4325 extent
= isl_union_set_from_set(set
);
4326 domain
= isl_union_map_domain(copy_i
);
4327 copy_i
= isl_union_map_from_domain_and_range(domain
, extent
);
4328 res
= isl_union_map_union(res
, copy_i
);
4331 isl_union_map_free(copy
);
4336 /* Insert "kernel" marks that point to a ppcg_kernel structure
4337 * in front of all outermost tilable band that (by construction)
4338 * have at least one parallel loop.
4340 static __isl_give isl_schedule_node
*mark_kernels(struct gpu_gen
*gen
,
4341 __isl_take isl_schedule_node
*node
)
4343 return isl_schedule_node_map_descendant_bottom_up(node
,
4344 &mark_outer_permutable
, gen
);
4347 /* Construct schedule constraints from the dependences in prog->scop and
4348 * the array order dependences in prog->array_order.
4350 * If live range reordering is allowed, then we need to make sure
4351 * that live ranges on arrays are not run in parallel since doing
4352 * so would require array expansion. We therefore add the array
4353 * order dependences to the coincidence dependences. Non-zero array
4354 * order dependences will then prevent a schedule dimension from being
4355 * considered parallel.
4356 * Live ranges derived from scalars are allowed to be run in parallel
4357 * since we force the scalars to be mapped to private memory in
4358 * check_scalar_live_ranges.
4359 * If live range reordering is allowed, then the false dependences
4360 * are not added to the validity constraints as that would prevent
4361 * reordering. Instead, the external false dependences that enforce that reads
4362 * from potentially live-in data precede any later write and
4363 * that writes of potentially live-out data follow any other earlier write
4364 * are added to the validity and the coincidence constraints.
4365 * The false dependences are still added to the proximity constraints
4366 * for consistency with the case where live range reordering is not allowed.
4367 * The coincidence constraints then consist of flow dependences,
4368 * external false dependences and array order dependences.
4369 * The independences can be filtered out from the first two sets.
4370 * They have already been filtered out from the array order dependences
4371 * on a per array basis in collect_order_dependences.
4372 * There is no need for a per array handling of the other two sets
4373 * as there should be no flow or external false dependence on local
4374 * variables that can be filtered out.
4376 static __isl_give isl_schedule_constraints
*construct_schedule_constraints(
4377 struct gpu_prog
*prog
)
4379 isl_union_set
*domain
;
4380 isl_union_map
*dep_raw
, *dep
;
4381 isl_union_map
*validity
, *proximity
, *coincidence
;
4382 isl_schedule_constraints
*sc
;
4384 domain
= isl_union_set_copy(prog
->scop
->domain
);
4385 sc
= isl_schedule_constraints_on_domain(domain
);
4386 sc
= isl_schedule_constraints_set_context(sc
,
4387 isl_set_copy(prog
->scop
->context
));
4388 if (prog
->scop
->options
->live_range_reordering
) {
4389 sc
= isl_schedule_constraints_set_conditional_validity(sc
,
4390 isl_union_map_copy(prog
->scop
->tagged_dep_flow
),
4391 isl_union_map_copy(prog
->scop
->tagged_dep_order
));
4392 proximity
= isl_union_map_copy(prog
->scop
->dep_flow
);
4393 validity
= isl_union_map_copy(proximity
);
4394 validity
= isl_union_map_union(validity
,
4395 isl_union_map_copy(prog
->scop
->dep_forced
));
4396 proximity
= isl_union_map_union(proximity
,
4397 isl_union_map_copy(prog
->scop
->dep_false
));
4398 coincidence
= isl_union_map_copy(validity
);
4399 coincidence
= isl_union_map_subtract(coincidence
,
4400 isl_union_map_copy(prog
->scop
->independence
));
4401 coincidence
= isl_union_map_union(coincidence
,
4402 isl_union_map_copy(prog
->array_order
));
4404 dep_raw
= isl_union_map_copy(prog
->scop
->dep_flow
);
4405 dep
= isl_union_map_copy(prog
->scop
->dep_false
);
4406 dep
= isl_union_map_union(dep
, dep_raw
);
4407 dep
= isl_union_map_coalesce(dep
);
4408 proximity
= isl_union_map_copy(dep
);
4409 coincidence
= isl_union_map_copy(dep
);
4412 sc
= isl_schedule_constraints_set_validity(sc
, validity
);
4413 sc
= isl_schedule_constraints_set_coincidence(sc
, coincidence
);
4414 sc
= isl_schedule_constraints_set_proximity(sc
, proximity
);
4419 /* Compute an appropriate schedule based on the accesses in
4420 * gen->read and gen->write.
4422 * We derive schedule constraints from the dependences in gen->prog->scop
4423 * and then use isl to compute a schedule that has a parallel loop
4424 * in each tilable band.
4425 * During the schedule construction, some statement instances
4426 * may be grouped first based on the input schedule.
4428 static __isl_give isl_schedule
*compute_schedule(struct gpu_gen
*gen
)
4430 isl_schedule_constraints
*sc
;
4431 isl_schedule
*schedule
;
4433 sc
= construct_schedule_constraints(gen
->prog
);
4434 schedule
= gen
->prog
->scop
->schedule
;
4435 schedule
= ppcg_compute_schedule(sc
, schedule
, gen
->options
);
4440 /* If the band node "node" has exactly one member then mark it permutable.
4442 static __isl_give isl_schedule_node
*band_set_permutable(
4443 __isl_take isl_schedule_node
*node
,
4444 __isl_keep isl_schedule_constraints
*sc
)
4446 if (isl_schedule_node_band_n_member(node
) == 1)
4447 node
= isl_schedule_node_band_set_permutable(node
, 1);
4452 /* Return the coincidence constraints between pairs of instances
4453 * that are scheduled together by the ancestors of "node".
4454 * That is, select those coincidence constraints that relate
4455 * pairs of instances that have the same value for the prefix schedule.
4456 * If the schedule depth is zero, then the prefix schedule does not
4457 * contain any information, so we intersect domain and range
4458 * of the schedule constraints with the reaching domain elements instead.
4460 static __isl_give isl_union_map
*get_local_coincidence(
4461 __isl_keep isl_schedule_node
*node
,
4462 __isl_keep isl_schedule_constraints
*sc
)
4464 isl_union_map
*coincidence
;
4465 isl_multi_union_pw_aff
*prefix
;
4466 isl_union_pw_multi_aff
*contraction
;
4468 coincidence
= isl_schedule_constraints_get_coincidence(sc
);
4469 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4470 if (isl_schedule_node_get_schedule_depth(node
) == 0) {
4471 isl_union_set
*domain
;
4473 domain
= isl_schedule_node_get_domain(node
);
4474 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4476 coincidence
= isl_union_map_intersect_domain(coincidence
,
4477 isl_union_set_copy(domain
));
4478 coincidence
= isl_union_map_intersect_range(coincidence
,
4483 prefix
= isl_schedule_node_get_prefix_schedule_multi_union_pw_aff(node
);
4484 prefix
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(prefix
,
4486 return isl_union_map_eq_at_multi_union_pw_aff(coincidence
, prefix
);
4489 /* For each member in the band node "node", determine whether
4490 * it is coincident with respect to the outer nodes and mark
4493 * That is, for each coincidence constraint between pairs
4494 * of instances that are scheduled together by the outer nodes,
4495 * check that domain and range are assigned the same value
4496 * by the band member. This test is performed by checking
4497 * that imposing the same value for the band member does not
4498 * remove any elements from the set of coincidence constraints.
4500 static __isl_give isl_schedule_node
*band_set_coincident(
4501 __isl_take isl_schedule_node
*node
,
4502 __isl_keep isl_schedule_constraints
*sc
)
4504 isl_union_map
*coincidence
;
4505 isl_union_pw_multi_aff
*contraction
;
4506 isl_multi_union_pw_aff
*partial
;
4509 coincidence
= get_local_coincidence(node
, sc
);
4511 partial
= isl_schedule_node_band_get_partial_schedule(node
);
4512 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4513 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4515 n
= isl_schedule_node_band_n_member(node
);
4516 for (i
= 0; i
< n
; ++i
) {
4517 isl_union_map
*coincidence_i
;
4518 isl_union_pw_aff
*upa
;
4519 isl_multi_union_pw_aff
*partial_i
;
4522 upa
= isl_multi_union_pw_aff_get_union_pw_aff(partial
, i
);
4523 partial_i
= isl_multi_union_pw_aff_from_union_pw_aff(upa
);
4524 coincidence_i
= isl_union_map_copy(coincidence
);
4525 coincidence_i
= isl_union_map_eq_at_multi_union_pw_aff(
4526 coincidence_i
, partial_i
);
4527 subset
= isl_union_map_is_subset(coincidence
, coincidence_i
);
4528 isl_union_map_free(coincidence_i
);
4532 node
= isl_schedule_node_band_member_set_coincident(node
, i
,
4536 node
= isl_schedule_node_free(node
);
4537 isl_multi_union_pw_aff_free(partial
);
4538 isl_union_map_free(coincidence
);
4543 /* If "node" is a band, then set its properties.
4545 * In particular, if the band has exactly one member, then mark it permutable.
4546 * Mark the band members coincident based on the coincidence constraints
4549 static __isl_give isl_schedule_node
*set_band_properties(
4550 __isl_take isl_schedule_node
*node
, void *user
)
4552 isl_schedule_constraints
*sc
= user
;
4554 if (isl_schedule_node_get_type(node
) != isl_schedule_node_band
)
4556 if (isl_schedule_node_band_n_member(node
) == 0)
4559 node
= band_set_permutable(node
, sc
);
4560 node
= band_set_coincident(node
, sc
);
4565 /* Return the original schedule with all bands marked permutable and
4566 * all band members marked coincident based on the coincidence constraints.
4567 * The bands are explicitly marked permutable so that they will be considered
4568 * by mark_outer_permutable.
4570 static __isl_give isl_schedule
*determine_properties_original_schedule(
4571 struct gpu_gen
*gen
)
4573 isl_schedule
*schedule
;
4574 isl_schedule_constraints
*sc
;
4576 schedule
= isl_schedule_copy(gen
->prog
->scop
->schedule
);
4577 sc
= construct_schedule_constraints(gen
->prog
);
4578 schedule
= isl_schedule_map_schedule_node_bottom_up(schedule
,
4579 &set_band_properties
, sc
);
4580 isl_schedule_constraints_free(sc
);
4585 /* Compute a schedule or determine the properties of the original schedule
4586 * depending on the value of the "reschedule" option.
4588 static __isl_give isl_schedule
*compute_or_set_properties(void *user
)
4590 struct gpu_gen
*gen
= user
;
4592 if (gen
->options
->reschedule
)
4593 return compute_schedule(gen
);
4595 return determine_properties_original_schedule(gen
);
4598 /* Obtain a schedule for the scop, by reading it from
4599 * a file, by computing one or by determining the properties
4600 * of the original schedule.
4602 static __isl_give isl_schedule
*get_schedule(struct gpu_gen
*gen
)
4604 return ppcg_get_schedule(gen
->ctx
, gen
->options
,
4605 &compute_or_set_properties
, gen
);
4608 /* Construct the string "<a>_<b>".
4610 static char *concat(isl_ctx
*ctx
, const char *a
, const char *b
)
4615 p
= isl_printer_to_str(ctx
);
4616 p
= isl_printer_print_str(p
, a
);
4617 p
= isl_printer_print_str(p
, "_");
4618 p
= isl_printer_print_str(p
, b
);
4619 s
= isl_printer_get_str(p
);
4620 isl_printer_free(p
);
4625 /* For each array in "prog" of which an element appears in "accessed" and
4626 * that is not a read only scalar, create a zero-dimensional universe set
4627 * of which the tuple id has name "<prefix>_<name of array>" and a user
4628 * pointer pointing to the array (gpu_array_info).
4630 * If the array is local to "prog", then make sure it will be declared
4633 * Return the list of these universe sets.
4635 static __isl_give isl_union_set_list
*create_copy_filters(struct gpu_prog
*prog
,
4636 const char *prefix
, __isl_take isl_union_set
*accessed
)
4640 isl_union_set_list
*filters
;
4643 filters
= isl_union_set_list_alloc(ctx
, 0);
4644 for (i
= 0; i
< prog
->n_array
; ++i
) {
4645 struct gpu_array_info
*array
= &prog
->array
[i
];
4647 isl_set
*accessed_i
;
4651 isl_union_set
*uset
;
4653 if (gpu_array_is_read_only_scalar(array
))
4656 space
= isl_space_copy(array
->space
);
4657 accessed_i
= isl_union_set_extract_set(accessed
, space
);
4658 empty
= isl_set_plain_is_empty(accessed_i
);
4659 isl_set_free(accessed_i
);
4661 filters
= isl_union_set_list_free(filters
);
4669 array
->declare_local
= 1;
4671 name
= concat(ctx
, prefix
, array
->name
);
4672 id
= name
? isl_id_alloc(ctx
, name
, array
) : NULL
;
4674 space
= isl_space_set_alloc(ctx
, 0, 0);
4675 space
= isl_space_set_tuple_id(space
, isl_dim_set
, id
);
4676 uset
= isl_union_set_from_set(isl_set_universe(space
));
4678 filters
= isl_union_set_list_add(filters
, uset
);
4680 isl_union_set_free(accessed
);
4685 /* Make sure that code for the statements in "filters" that
4686 * copy arrays to or from the device is only generated when
4687 * the size of the corresponding array is positive.
4688 * That is, add a set node underneath "graft" with "filters" as children
4689 * and for each child add a guard that the selects the parameter
4690 * values for which the corresponding array has a positive size.
4691 * The array is available in the user pointer of the statement identifier.
4692 * "depth" is the schedule depth of the position where "graft"
4695 static __isl_give isl_schedule_node
*insert_positive_size_guards(
4696 __isl_take isl_schedule_node
*graft
,
4697 __isl_take isl_union_set_list
*filters
, int depth
)
4701 graft
= isl_schedule_node_child(graft
, 0);
4702 graft
= isl_schedule_node_insert_set(graft
, filters
);
4703 n
= isl_schedule_node_n_children(graft
);
4704 for (i
= 0; i
< n
; ++i
) {
4705 isl_union_set
*filter
;
4706 isl_set
*domain
, *guard
;
4708 struct gpu_array_info
*array
;
4710 graft
= isl_schedule_node_child(graft
, i
);
4711 filter
= isl_schedule_node_filter_get_filter(graft
);
4712 domain
= isl_set_from_union_set(filter
);
4713 id
= isl_set_get_tuple_id(domain
);
4714 array
= isl_id_get_user(id
);
4716 isl_set_free(domain
);
4717 guard
= gpu_array_positive_size_guard(array
);
4718 guard
= isl_set_from_params(guard
);
4719 guard
= isl_set_add_dims(guard
, isl_dim_set
, depth
);
4720 graft
= isl_schedule_node_child(graft
, 0);
4721 graft
= isl_schedule_node_insert_guard(graft
, guard
);
4722 graft
= isl_schedule_node_parent(graft
);
4723 graft
= isl_schedule_node_parent(graft
);
4725 graft
= isl_schedule_node_parent(graft
);
4730 /* Create a graft for copying arrays to or from the device,
4731 * whenever the size of the array is strictly positive.
4732 * Each statement is called "<prefix>_<name of array>" and
4733 * the identifier has a user pointer pointing to the array.
4734 * The graft will be added at the position specified by "node".
4735 * "copy" contains the array elements that need to be copied.
4736 * Only arrays of which some elements need to be copied
4737 * will have a corresponding statement in the graph.
4738 * Note though that each such statement will copy the entire array.
4740 static __isl_give isl_schedule_node
*create_copy_device(struct gpu_prog
*prog
,
4741 __isl_keep isl_schedule_node
*node
, const char *prefix
,
4742 __isl_take isl_union_set
*copy
)
4747 isl_union_set
*all
, *domain
;
4748 isl_union_set_list
*filters
;
4749 isl_union_map
*extension
;
4750 isl_schedule_node
*graft
;
4753 depth
= isl_schedule_node_get_schedule_depth(node
);
4754 filters
= create_copy_filters(prog
, prefix
, copy
);
4755 all
= isl_union_set_list_union(isl_union_set_list_copy(filters
));
4757 space
= depth
< 0 ? NULL
: isl_space_set_alloc(ctx
, 0, depth
);
4758 domain
= isl_union_set_from_set(isl_set_universe(space
));
4759 extension
= isl_union_map_from_domain_and_range(domain
, all
);
4760 graft
= isl_schedule_node_from_extension(extension
);
4763 return isl_schedule_node_free(graft
);
4764 if (isl_union_set_list_n_union_set(filters
) == 0) {
4765 isl_union_set_list_free(filters
);
4769 return insert_positive_size_guards(graft
, filters
, depth
);
4772 /* Return (the universe spaces of) the arrays that are declared
4773 * inside the scop corresponding to "prog" and for which all
4774 * potential writes inside the scop form a subset of "domain".
4776 static __isl_give isl_union_set
*extract_local_accesses(struct gpu_prog
*prog
,
4777 __isl_keep isl_union_set
*domain
)
4780 isl_union_set
*local
;
4782 local
= isl_union_set_empty(isl_union_set_get_space(domain
));
4784 for (i
= 0; i
< prog
->n_array
; ++i
) {
4786 isl_union_map
*to_outer
;
4787 isl_union_map
*may_write
;
4788 isl_union_set
*write_domain
;
4789 isl_union_set
*fields
;
4792 if (!prog
->array
[i
].local
)
4795 set
= isl_set_universe(isl_space_copy(prog
->array
[i
].space
));
4796 to_outer
= isl_union_map_copy(prog
->to_outer
);
4797 to_outer
= isl_union_map_intersect_range(to_outer
,
4798 isl_union_set_from_set(isl_set_copy(set
)));
4799 fields
= isl_union_map_domain(to_outer
);
4800 may_write
= isl_union_map_copy(prog
->may_write
);
4801 may_write
= isl_union_map_intersect_range(may_write
, fields
);
4802 write_domain
= isl_union_map_domain(may_write
);
4803 subset
= isl_union_set_is_subset(write_domain
, domain
);
4804 isl_union_set_free(write_domain
);
4808 return isl_union_set_free(local
);
4809 } else if (subset
) {
4810 local
= isl_union_set_add_set(local
, set
);
4819 /* Internal data structure for node_may_persist.
4821 * "tagger" maps tagged iteration domains to the corresponding untagged
4824 * "may_persist_flow" is the set of all tagged dataflow dependences
4825 * with those dependences removed that either precede or follow
4826 * the kernel launch in a sequence.
4827 * "inner_band_flow" is the set of all tagged dataflow dependences
4828 * that are local to a given iteration of the outer band nodes
4829 * with respect to the current node.
4830 * "local_flow" is equal to "inner_band_flow", except that the domain
4831 * and the range have been intersected with intermediate filters
4832 * on children of sets or sequences.
4834 struct ppcg_may_persist_data
{
4835 isl_union_pw_multi_aff
*tagger
;
4837 isl_union_map
*local_flow
;
4838 isl_union_map
*inner_band_flow
;
4839 isl_union_map
*may_persist_flow
;
4842 /* Update the information in "data" based on the band ancestor "node".
4844 * In particular, we restrict the dependences in data->local_flow
4845 * to those dependence where the source and the sink occur in
4846 * the same iteration of the given band node.
4847 * We also update data->inner_band_flow to the new value of
4850 static int update_may_persist_at_band(__isl_keep isl_schedule_node
*node
,
4851 struct ppcg_may_persist_data
*data
)
4853 isl_multi_union_pw_aff
*partial
;
4854 isl_union_pw_multi_aff
*contraction
;
4855 isl_union_map
*flow
;
4857 if (isl_schedule_node_band_n_member(node
) == 0)
4860 partial
= isl_schedule_node_band_get_partial_schedule(node
);
4861 contraction
= isl_schedule_node_get_subtree_contraction(node
);
4862 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4864 partial
= isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial
,
4865 isl_union_pw_multi_aff_copy(data
->tagger
));
4867 flow
= data
->local_flow
;
4868 flow
= isl_union_map_eq_at_multi_union_pw_aff(flow
, partial
);
4869 data
->local_flow
= flow
;
4871 isl_union_map_free(data
->inner_band_flow
);
4872 data
->inner_band_flow
= isl_union_map_copy(data
->local_flow
);
4877 /* Given a set of local reaching domain elements "domain",
4878 * expand them to the corresponding leaf domain elements using "contraction"
4879 * and insert the array references tags using data->tagger.
4881 static __isl_give isl_union_set
*expand_and_tag(
4882 __isl_take isl_union_set
*domain
,
4883 __isl_take isl_union_pw_multi_aff
*contraction
,
4884 struct ppcg_may_persist_data
*data
)
4886 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4888 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
4889 isl_union_pw_multi_aff_copy(data
->tagger
));
4893 /* Given a filter node that is the child of a set or sequence node,
4894 * restrict data->local_flow to refer only to those elements
4895 * in the filter of the node.
4896 * "contraction" maps the leaf domain elements of the schedule tree
4897 * to the corresponding domain elements at (the parent of) "node".
4899 static int filter_flow(__isl_keep isl_schedule_node
*node
,
4900 struct ppcg_may_persist_data
*data
,
4901 __isl_take isl_union_pw_multi_aff
*contraction
)
4903 isl_union_set
*filter
;
4904 isl_union_map
*flow
;
4906 flow
= data
->local_flow
;
4907 filter
= isl_schedule_node_filter_get_filter(node
);
4908 filter
= expand_and_tag(filter
, contraction
, data
);
4909 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(filter
));
4910 flow
= isl_union_map_intersect_range(flow
, filter
);
4911 data
->local_flow
= flow
;
4916 /* Given a filter node "node", collect the filters on all preceding siblings
4917 * (which are also filter nodes), add them to "filters" and return the result.
4919 static __isl_give isl_union_set
*add_previous_filters(
4920 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4922 isl_schedule_node
*sibling
;
4924 sibling
= isl_schedule_node_copy(node
);
4925 while (sibling
&& isl_schedule_node_has_previous_sibling(sibling
)) {
4926 isl_union_set
*filter
;
4928 sibling
= isl_schedule_node_previous_sibling(sibling
);
4929 filter
= isl_schedule_node_filter_get_filter(sibling
);
4930 filters
= isl_union_set_union(filters
, filter
);
4932 isl_schedule_node_free(sibling
);
4934 return isl_union_set_free(filters
);
4939 /* Given a filter node "node", collect the filters on all following siblings
4940 * (which are also filter nodes), add them to "filters" and return the result.
4942 static __isl_give isl_union_set
*add_next_filters(
4943 __isl_take isl_union_set
*filters
, __isl_keep isl_schedule_node
*node
)
4945 isl_schedule_node
*sibling
;
4947 sibling
= isl_schedule_node_copy(node
);
4948 while (sibling
&& isl_schedule_node_has_next_sibling(sibling
)) {
4949 isl_union_set
*filter
;
4951 sibling
= isl_schedule_node_next_sibling(sibling
);
4952 filter
= isl_schedule_node_filter_get_filter(sibling
);
4953 filters
= isl_union_set_union(filters
, filter
);
4955 isl_schedule_node_free(sibling
);
4957 return isl_union_set_free(filters
);
4962 /* Remove those flow dependences from data->may_persist_flow
4963 * that flow between elements of "domain" within the same iteration
4964 * of all outer band nodes.
4965 * "contraction" maps the leaf domain elements of the schedule tree
4966 * to the corresponding elements "domain".
4968 static void remove_external_flow(struct ppcg_may_persist_data
*data
,
4969 __isl_take isl_union_set
*domain
,
4970 __isl_keep isl_union_pw_multi_aff
*contraction
)
4972 isl_union_map
*flow
;
4974 contraction
= isl_union_pw_multi_aff_copy(contraction
);
4975 domain
= expand_and_tag(domain
, contraction
, data
);
4976 flow
= isl_union_map_copy(data
->local_flow
);
4977 flow
= isl_union_map_intersect_domain(flow
, isl_union_set_copy(domain
));
4978 flow
= isl_union_map_intersect_range(flow
, domain
);
4980 data
->may_persist_flow
= isl_union_map_subtract(data
->may_persist_flow
,
4984 /* Update the information in "data" based on the filter ancestor "node".
4985 * We only need to modify anything if the filter is the child
4986 * of a set or sequence node.
4988 * In the case of a sequence, we remove the dependences between
4989 * statement instances that are both executed either before or
4990 * after the subtree that will be mapped to a kernel, within
4991 * the same iteration of outer bands.
4993 * In both cases, we restrict data->local_flow to the current child.
4995 static int update_may_persist_at_filter(__isl_keep isl_schedule_node
*node
,
4996 struct ppcg_may_persist_data
*data
)
4998 enum isl_schedule_node_type type
;
4999 isl_schedule_node
*parent
;
5001 isl_union_pw_multi_aff
*contraction
;
5002 isl_union_set
*before
, *after
, *filter
;
5004 type
= isl_schedule_node_get_parent_type(node
);
5005 if (type
!= isl_schedule_node_sequence
&& type
!= isl_schedule_node_set
)
5008 parent
= isl_schedule_node_copy(node
);
5009 parent
= isl_schedule_node_parent(parent
);
5010 contraction
= isl_schedule_node_get_subtree_contraction(parent
);
5011 isl_schedule_node_free(parent
);
5013 if (type
== isl_schedule_node_set
)
5014 return filter_flow(node
, data
, contraction
);
5016 filter
= isl_schedule_node_filter_get_filter(node
);
5017 space
= isl_union_set_get_space(filter
);
5018 isl_union_set_free(filter
);
5019 before
= isl_union_set_empty(space
);
5020 after
= isl_union_set_copy(before
);
5021 before
= add_previous_filters(before
, node
);
5022 after
= add_next_filters(after
, node
);
5024 remove_external_flow(data
, before
, contraction
);
5025 remove_external_flow(data
, after
, contraction
);
5027 return filter_flow(node
, data
, contraction
);
5030 /* Update the information in "data" based on the ancestor "node".
5032 static isl_stat
update_may_persist_at(__isl_keep isl_schedule_node
*node
,
5035 struct ppcg_may_persist_data
*data
= user
;
5037 switch (isl_schedule_node_get_type(node
)) {
5038 case isl_schedule_node_error
:
5039 return isl_stat_error
;
5040 case isl_schedule_node_context
:
5041 case isl_schedule_node_domain
:
5042 case isl_schedule_node_expansion
:
5043 case isl_schedule_node_extension
:
5044 case isl_schedule_node_guard
:
5045 case isl_schedule_node_leaf
:
5046 case isl_schedule_node_mark
:
5047 case isl_schedule_node_sequence
:
5048 case isl_schedule_node_set
:
5050 case isl_schedule_node_band
:
5051 if (update_may_persist_at_band(node
, data
) < 0)
5052 return isl_stat_error
;
5054 case isl_schedule_node_filter
:
5055 if (update_may_persist_at_filter(node
, data
) < 0)
5056 return isl_stat_error
;
5063 /* Determine the set of array elements that may need to be perserved
5064 * by a kernel constructed from the subtree at "node".
5065 * This includes the set of array elements that may need to be preserved
5066 * by the entire scop (prog->may_persist) and the elements for which
5067 * there is a potential flow dependence that may cross a kernel launch.
5069 * To determine the second set, we start from all flow dependences.
5070 * From this set of dependences, we remove those that cannot possibly
5071 * require data to be preserved by a kernel launch.
5072 * In particular, we consider the following sets of dependences.
5073 * - dependences of which the write occurs inside the kernel.
5074 * If the data is needed outside the kernel, then it will
5075 * be copied out immediately after the kernel launch, so there
5076 * is no need for any special care.
5077 * - dependences of which the read occurs inside the kernel and the
5078 * corresponding write occurs inside the same iteration of the
5079 * outer band nodes. This means that the data is needed in
5080 * the first kernel launch after the write, which is already
5081 * taken care of by the standard copy-in. That is, the data
5082 * do not need to be preserved by any intermediate call to
5084 * - dependences of which the write and the read either both occur
5085 * before the kernel launch or both occur after the kernel launch,
5086 * within the same iteration of the outer band nodes with respect
5087 * to the sequence that determines the ordering of the dependence
5088 * and the kernel launch. Such flow dependences cannot cross
5089 * any kernel launch.
5091 * For the remaining (tagged) dependences, we take the domain
5092 * (i.e., the tagged writes) and apply the tagged access relation
5093 * to obtain the accessed data elements.
5094 * These are then combined with the elements that may need to be
5095 * preserved by the entire scop.
5097 static __isl_give isl_union_set
*node_may_persist(
5098 __isl_keep isl_schedule_node
*node
, struct gpu_prog
*prog
)
5100 struct ppcg_may_persist_data data
;
5101 isl_union_pw_multi_aff
*contraction
;
5102 isl_union_set
*domain
;
5103 isl_union_set
*persist
;
5104 isl_union_map
*flow
, *local_flow
;
5106 data
.tagger
= prog
->scop
->tagger
;
5108 flow
= isl_union_map_copy(prog
->scop
->tagged_dep_flow
);
5109 data
.local_flow
= isl_union_map_copy(flow
);
5110 data
.inner_band_flow
= isl_union_map_copy(flow
);
5111 data
.may_persist_flow
= flow
;
5112 if (isl_schedule_node_foreach_ancestor_top_down(node
,
5113 &update_may_persist_at
, &data
) < 0)
5114 data
.may_persist_flow
=
5115 isl_union_map_free(data
.may_persist_flow
);
5116 flow
= data
.may_persist_flow
;
5117 isl_union_map_free(data
.local_flow
);
5119 domain
= isl_schedule_node_get_domain(node
);
5120 contraction
= isl_schedule_node_get_subtree_contraction(node
);
5121 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
5123 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
5124 isl_union_pw_multi_aff_copy(data
.tagger
));
5125 flow
= isl_union_map_subtract_domain(flow
, isl_union_set_copy(domain
));
5126 local_flow
= data
.inner_band_flow
;
5127 local_flow
= isl_union_map_intersect_range(local_flow
, domain
);
5128 flow
= isl_union_map_subtract(flow
, local_flow
);
5130 persist
= isl_union_map_domain(flow
);
5131 persist
= isl_union_set_apply(persist
,
5132 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
5133 persist
= isl_union_set_union(persist
,
5134 isl_union_set_copy(prog
->may_persist
));
5139 /* Add nodes for copying outer arrays in and out of the device
5140 * before and after the subtree "node", which contains one or more kernels.
5141 * "domain" contains the original statement instances, i.e.,
5142 * those that correspond to the domains of the access relations in "prog".
5143 * In particular, the domain has not been contracted in any way.
5144 * "prefix" contains the prefix schedule at that point, in terms
5145 * of the same original statement instances.
5147 * We first compute the sets of outer array elements that need
5148 * to be copied in and out and then graft in the nodes for
5149 * performing this copying.
5151 * In particular, for each array that is possibly written anywhere in
5152 * the subtree "node" and that may be used after "node"
5153 * or that may be visible outside the corresponding scop,
5154 * we copy out its entire extent.
5156 * Any array elements that is read without first being written inside
5157 * the subtree "node" needs to be copied in.
5158 * Furthermore, if there are any array elements that
5159 * are copied out, but that may not be written inside "node, then
5160 * they also need to be copied in to ensure that the value after execution
5161 * is the same as the value before execution, at least for those array
5162 * elements that may have their values preserved by the scop or that
5163 * may be written before "node" and read after "node".
5164 * In case the array elements are structures, we need to take into
5165 * account that all members of the structures need to be written
5166 * by "node" before we can avoid copying the data structure in.
5168 * Note that the may_write relation is intersected with the domain,
5169 * which has been intersected with the context.
5170 * This helps in those cases where the arrays are declared with a fixed size,
5171 * while the accesses are parametric and the context assigns a fixed value
5172 * to the parameters.
5174 * If an element from a local array is read without first being written,
5175 * then there is no point in copying it in since it cannot have been
5176 * written prior to the scop. Warn about the uninitialized read instead.
5178 static __isl_give isl_schedule_node
*add_to_from_device(
5179 __isl_take isl_schedule_node
*node
, __isl_take isl_union_set
*domain
,
5180 __isl_take isl_union_map
*prefix
, struct gpu_prog
*prog
)
5182 isl_union_set
*local
;
5183 isl_union_set
*may_persist
;
5184 isl_union_map
*may_write
, *must_write
, *copy_out
, *not_written
;
5185 isl_union_map
*read
, *copy_in
;
5186 isl_union_map
*tagged
;
5187 isl_union_map
*local_uninitialized
;
5188 isl_schedule_node
*graft
;
5190 tagged
= isl_union_map_copy(prog
->scop
->tagged_reads
);
5191 tagged
= isl_union_map_union(tagged
,
5192 isl_union_map_copy(prog
->scop
->tagged_may_writes
));
5194 may_write
= isl_union_map_copy(prog
->may_write
);
5195 may_write
= isl_union_map_intersect_domain(may_write
,
5196 isl_union_set_copy(domain
));
5197 may_write
= remove_local_accesses(prog
,
5198 isl_union_map_copy(tagged
), may_write
,
5199 isl_union_map_copy(prefix
), 0);
5200 may_write
= isl_union_map_apply_range(may_write
,
5201 isl_union_map_copy(prog
->to_outer
));
5202 may_write
= isl_union_map_apply_domain(may_write
,
5203 isl_union_map_copy(prefix
));
5204 may_write
= approximate_copy_out(may_write
, prog
);
5205 copy_out
= isl_union_map_copy(may_write
);
5206 may_write
= isl_union_map_apply_range(may_write
,
5207 isl_union_map_copy(prog
->to_inner
));
5208 must_write
= isl_union_map_copy(prog
->must_write
);
5209 must_write
= isl_union_map_apply_domain(must_write
,
5210 isl_union_map_copy(prefix
));
5211 may_persist
= node_may_persist(node
, prog
);
5212 may_write
= isl_union_map_intersect_range(may_write
, may_persist
);
5213 not_written
= isl_union_map_subtract(may_write
, must_write
);
5215 local
= extract_local_accesses(prog
, domain
);
5216 read
= isl_union_map_copy(prog
->read
);
5217 read
= isl_union_map_intersect_domain(read
, domain
);
5218 read
= remove_local_accesses(prog
, tagged
, read
,
5219 isl_union_map_copy(prefix
), 1);
5220 local
= isl_union_set_apply(local
, isl_union_map_copy(prog
->to_inner
));
5221 local_uninitialized
= isl_union_map_copy(prog
->scop
->live_in
);
5222 local_uninitialized
= isl_union_map_intersect_range(local_uninitialized
,
5224 local_uninitialized
= isl_union_map_intersect(local_uninitialized
,
5225 isl_union_map_copy(read
));
5226 if (!isl_union_map_is_empty(local_uninitialized
)) {
5228 "possibly uninitialized reads (not copied in):\n");
5229 isl_union_map_dump(local_uninitialized
);
5231 read
= isl_union_map_subtract(read
, local_uninitialized
);
5232 read
= isl_union_map_apply_domain(read
, prefix
);
5233 copy_in
= isl_union_map_union(read
, not_written
);
5234 copy_in
= isl_union_map_apply_range(copy_in
,
5235 isl_union_map_copy(prog
->to_outer
));
5237 graft
= create_copy_device(prog
, node
, "to_device",
5238 isl_union_map_range(copy_in
));
5239 node
= isl_schedule_node_graft_before(node
, graft
);
5240 graft
= create_copy_device(prog
, node
, "from_device",
5241 isl_union_map_range(copy_out
));
5242 node
= isl_schedule_node_graft_after(node
, graft
);
5247 /* Add nodes for initializing ("init_device") and clearing ("clear_device")
5248 * the device before and after "node".
5250 static __isl_give isl_schedule_node
*add_init_clear_device(
5251 __isl_take isl_schedule_node
*node
)
5255 isl_union_set
*domain
;
5256 isl_schedule_node
*graft
;
5258 ctx
= isl_schedule_node_get_ctx(node
);
5260 space
= isl_space_set_alloc(ctx
, 0, 0);
5261 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "init_device");
5262 domain
= isl_union_set_from_set(isl_set_universe(space
));
5263 graft
= isl_schedule_node_from_domain(domain
);
5265 node
= isl_schedule_node_graft_before(node
, graft
);
5267 space
= isl_space_set_alloc(ctx
, 0, 0);
5268 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "clear_device");
5269 domain
= isl_union_set_from_set(isl_set_universe(space
));
5270 graft
= isl_schedule_node_from_domain(domain
);
5272 node
= isl_schedule_node_graft_after(node
, graft
);
5277 /* Update "schedule" for mapping to a GPU device.
5279 * In particular, insert a context node, create kernels for
5280 * each outermost tilable band and introduce nodes for copying arrays
5281 * in and out of the device and for initializing and clearing the device.
5282 * If the child of the initial root points to a set node,
5283 * then children of this node that do not contain any tilable bands
5284 * are separated from the other children and are not mapped to
5287 * The GPU code is generated in a context where at least one
5288 * statement instance is executed. The corresponding guard is inserted
5289 * around the entire schedule.
5291 static __isl_give isl_schedule
*map_to_device(struct gpu_gen
*gen
,
5292 __isl_take isl_schedule
*schedule
)
5294 isl_schedule_node
*node
;
5297 isl_union_set
*domain
;
5298 isl_union_map
*prefix
;
5299 isl_union_pw_multi_aff
*contraction
;
5300 struct gpu_prog
*prog
;
5302 context
= isl_set_copy(gen
->prog
->context
);
5303 context
= isl_set_from_params(context
);
5304 schedule
= isl_schedule_insert_context(schedule
, context
);
5307 guard
= isl_union_set_params(isl_union_set_copy(prog
->scop
->domain
));
5308 prog
->context
= isl_set_intersect(prog
->context
, isl_set_copy(guard
));
5309 guard
= isl_set_from_params(guard
);
5311 node
= isl_schedule_get_root(schedule
);
5312 isl_schedule_free(schedule
);
5313 node
= isl_schedule_node_child(node
, 0);
5314 node
= isl_schedule_node_child(node
, 0);
5315 node
= isolate_permutable_subtrees(node
, gen
->prog
);
5316 domain
= isl_schedule_node_get_domain(node
);
5317 contraction
= isl_schedule_node_get_subtree_contraction(node
);
5318 domain
= isl_union_set_preimage_union_pw_multi_aff(domain
,
5319 isl_union_pw_multi_aff_copy(contraction
));
5320 prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
5321 prefix
= isl_union_map_preimage_domain_union_pw_multi_aff(prefix
,
5323 node
= mark_kernels(gen
, node
);
5324 node
= add_to_from_device(node
, domain
, prefix
, gen
->prog
);
5325 node
= isl_schedule_node_root(node
);
5326 node
= isl_schedule_node_child(node
, 0);
5327 node
= isl_schedule_node_child(node
, 0);
5328 node
= isl_schedule_node_insert_guard(node
, guard
);
5329 node
= isl_schedule_node_child(node
, 0);
5330 node
= add_init_clear_device(node
);
5331 schedule
= isl_schedule_node_get_schedule(node
);
5332 isl_schedule_node_free(node
);
5337 /* Internal data structure for extract_access.
5338 * "next_access" points to the end of a linked list that is extended
5339 * by extract_access.
5340 * "single_expression" is set if the access expressions belong to
5341 * an expression statement (i.e., a statement without internal control).
5342 * "any_to_outer" maps all intermediate arrays to their outer arrays.
5344 struct ppcg_extract_access_data
{
5345 struct gpu_stmt_access
**next_access
;
5346 int single_expression
;
5347 isl_union_map
*any_to_outer
;
5350 /* Given a tagged access relation to a single array "tagged", extract it
5351 * as a map, taking into account that the input may be empty.
5352 * If the access relation is empty, then it does not contain
5353 * any space information, so we try to recover it from the index
5355 * The space of the index expression is of the form I -> A,
5356 * with I the statement instances and A the array, or [I -> F] -> A,
5357 * with F the filters corresponding to arguments.
5358 * We first drop F, if present, obtaining I -> A.
5359 * Then we construct I -> R, with R the reference tag,
5360 * combine the two into I -> [R -> A] and uncurry to obtain
5361 * the final result [I -> R] -> A.
5362 * Note that the index expression may have a lower dimension
5363 * than that of the array, but this dimension is not used
5364 * if the access relation is empty.
5366 static __isl_give isl_map
*extract_single_tagged_access(
5367 __isl_take isl_union_map
*tagged
, __isl_keep pet_expr
*expr
)
5371 isl_space
*space
, *space2
;
5372 isl_multi_pw_aff
*index
;
5374 empty
= isl_union_map_is_empty(tagged
);
5378 return isl_map_from_union_map(tagged
);
5379 isl_union_map_free(tagged
);
5381 index
= pet_expr_access_get_index(expr
);
5382 space
= isl_multi_pw_aff_get_space(index
);
5383 isl_multi_pw_aff_free(index
);
5384 if (isl_space_domain_is_wrapping(space
))
5385 space
= isl_space_domain_factor_domain(space
);
5386 space2
= isl_space_from_domain(isl_space_domain(isl_space_copy(space
)));
5387 id
= pet_expr_access_get_ref_id(expr
);
5388 space2
= isl_space_set_tuple_id(space2
, isl_dim_out
, id
);
5389 space
= isl_space_range_product(space2
, space
);
5390 space
= isl_space_uncurry(space
);
5392 return isl_map_empty(space
);
5394 isl_union_map_free(tagged
);
5398 /* Does the index expression "index" of "expr" represent an access
5399 * to a single element?
5400 * That is, is "index" completely specified?
5402 * If "expr" accesses elements from different spaces (i.e., fields
5403 * of a structure), then it does not access a single element.
5404 * Otherwise, if the single space of the access matches the space
5405 * of "index", then the index expression is completely specified
5406 * (no pointer to a lower-dimensional slice of the accessed array)
5407 * and a single element is being accessed.
5409 static isl_bool
complete_index(__isl_keep pet_expr
*expr
,
5410 __isl_keep isl_multi_pw_aff
*index
)
5412 isl_union_map
*read
, *write
, *all
;
5414 isl_space
*space1
, *space2
;
5417 read
= pet_expr_access_get_may_read(expr
);
5418 write
= pet_expr_access_get_may_write(expr
);
5419 all
= isl_union_map_union(read
, write
);
5421 return isl_bool_error
;
5422 if (isl_union_map_n_map(all
) != 1) {
5423 isl_union_map_free(all
);
5424 return isl_bool_false
;
5426 map
= isl_map_from_union_map(all
);
5427 space1
= isl_map_get_space(map
);
5429 space2
= isl_multi_pw_aff_get_space(index
);
5430 complete
= isl_space_tuple_is_equal(space1
, isl_dim_out
,
5431 space2
, isl_dim_out
);
5432 isl_space_free(space1
);
5433 isl_space_free(space2
);
5438 /* Does "expr" access a single, fixed element (independently of the statement
5440 * That is, does it have a completely specified constant index expression?
5442 * Note that it is not sufficient for the index expression to be
5443 * piecewise constant. isl_multi_pw_aff_is_cst can therefore not be used.
5445 static isl_bool
accesses_fixed_element(__isl_keep pet_expr
*expr
)
5448 isl_multi_pw_aff
*index
;
5449 isl_bool fixed
= isl_bool_true
;
5451 index
= pet_expr_access_get_index(expr
);
5453 return isl_bool_error
;
5454 n
= isl_multi_pw_aff_dim(index
, isl_dim_out
);
5455 for (i
= 0; i
< n
; ++i
) {
5458 pa
= isl_multi_pw_aff_get_pw_aff(index
, 0);
5459 fixed
= isl_pw_aff_n_piece(pa
) == 1;
5461 fixed
= isl_pw_aff_is_cst(pa
);
5462 isl_pw_aff_free(pa
);
5463 if (fixed
< 0 || !fixed
)
5466 if (fixed
>= 0 && fixed
)
5467 fixed
= complete_index(expr
, index
);
5468 isl_multi_pw_aff_free(index
);
5473 /* Extract a gpu_stmt_access from "expr", append it to the list
5474 * that ends in *data->next_access and update the end of the list.
5475 * If the access expression performs a write, then it is considered
5476 * exact only if it appears in a single expression statement and
5477 * if its may access relation is equal to its must access relation.
5479 * The combined set of may accesses may be a union if member accesses
5480 * are involved, but the entire set is derived from a single reference and
5481 * therefore from a single index expression. These accesses therefore
5482 * all map to the same outer array.
5484 static int extract_access(__isl_keep pet_expr
*expr
, void *user
)
5486 struct ppcg_extract_access_data
*data
= user
;
5487 isl_union_map
*tagged
;
5488 struct gpu_stmt_access
*access
;
5489 isl_ctx
*ctx
= pet_expr_get_ctx(expr
);
5490 isl_multi_pw_aff
*index
;
5492 access
= isl_alloc_type(ctx
, struct gpu_stmt_access
);
5495 access
->next
= NULL
;
5496 access
->read
= pet_expr_access_is_read(expr
);
5497 access
->write
= pet_expr_access_is_write(expr
);
5498 tagged
= pet_expr_access_get_tagged_may_read(expr
);
5499 tagged
= isl_union_map_union(tagged
,
5500 pet_expr_access_get_tagged_may_write(expr
));
5501 tagged
= isl_union_map_apply_range(tagged
,
5502 isl_union_map_copy(data
->any_to_outer
));
5503 if (!access
->write
) {
5504 access
->exact_write
= 1;
5505 } else if (!data
->single_expression
) {
5506 access
->exact_write
= 0;
5508 isl_union_map
*must
, *may
;
5509 may
= isl_union_map_copy(tagged
);
5510 may
= isl_union_map_domain_factor_domain(may
);
5511 must
= pet_expr_access_get_must_write(expr
);
5512 access
->exact_write
= isl_union_map_is_equal(must
, may
);
5513 isl_union_map_free(must
);
5514 isl_union_map_free(may
);
5516 index
= pet_expr_access_get_index(expr
);
5517 access
->n_index
= isl_multi_pw_aff_dim(index
, isl_dim_out
);
5518 isl_multi_pw_aff_free(index
);
5519 access
->ref_id
= pet_expr_access_get_ref_id(expr
);
5520 access
->tagged_access
= extract_single_tagged_access(tagged
, expr
);
5521 access
->access
= isl_map_copy(access
->tagged_access
);
5522 access
->access
= isl_map_domain_factor_domain(access
->access
);
5523 access
->fixed_element
= accesses_fixed_element(expr
);
5525 *data
->next_access
= access
;
5526 data
->next_access
= &(*data
->next_access
)->next
;
5528 if (!access
->access
|| access
->fixed_element
< 0)
5534 /* Construct a linked list of gpu_stmt_access objects,
5535 * one for each access expression in the statement body.
5536 * "any_to_outer" maps all intermediate arrays to their outer arrays.
5538 static int pet_stmt_extract_accesses(struct gpu_stmt
*stmt
,
5539 __isl_keep isl_union_map
*any_to_outer
)
5541 struct ppcg_extract_access_data data
;
5543 stmt
->accesses
= NULL
;
5544 data
.next_access
= &stmt
->accesses
;
5545 data
.single_expression
=
5546 pet_tree_get_type(stmt
->stmt
->body
) == pet_tree_expr
;
5547 data
.any_to_outer
= any_to_outer
;
5548 return pet_tree_foreach_access_expr(stmt
->stmt
->body
,
5549 &extract_access
, &data
);
5552 /* Has statement "stmt" been killed from "scop"?
5553 * That is, is the instance set of "scop" free from any
5554 * instances of "stmt"?
5556 static isl_bool
is_stmt_killed(struct ppcg_scop
*scop
, struct pet_stmt
*stmt
)
5563 return isl_bool_error
;
5564 space
= isl_set_get_space(stmt
->domain
);
5565 left
= isl_union_set_extract_set(scop
->domain
, space
);
5566 empty
= isl_set_plain_is_empty(left
);
5572 /* Return an array of gpu_stmt representing the statements in "scop".
5573 * Do not collect array accesses for statements that have been killed.
5575 static struct gpu_stmt
*extract_stmts(isl_ctx
*ctx
, struct ppcg_scop
*scop
,
5576 __isl_keep isl_union_map
*any_to_outer
)
5579 struct gpu_stmt
*stmts
;
5581 stmts
= isl_calloc_array(ctx
, struct gpu_stmt
, scop
->pet
->n_stmt
);
5585 for (i
= 0; i
< scop
->pet
->n_stmt
; ++i
) {
5586 struct gpu_stmt
*s
= &stmts
[i
];
5589 s
->id
= isl_set_get_tuple_id(scop
->pet
->stmts
[i
]->domain
);
5590 s
->stmt
= scop
->pet
->stmts
[i
];
5591 killed
= is_stmt_killed(scop
, scop
->pet
->stmts
[i
]);
5593 return free_stmts(stmts
, i
+ 1);
5596 if (pet_stmt_extract_accesses(s
, any_to_outer
) < 0)
5597 return free_stmts(stmts
, i
+ 1);
5603 /* Generate CUDA code for "scop" and print it to "p".
5604 * After generating an AST for the transformed scop as explained below,
5605 * we call "gen->print" to print the AST in the desired output format
5608 * If it turns out that it does not make sense to generate GPU code,
5609 * then we generate CPU code instead.
5611 * The declarations of the arrays that are visible outside of the scop
5612 * are printed outside of the code generated from the schedule,
5613 * because the generated code may involve a guard around the entire code.
5615 * We first compute a schedule that respects the dependences
5616 * of the original program and select the outermost bands
5617 * of tilable dimensions that have at least one parallel loop.
5618 * If the --load-schedule is specified, then the loaded schedule
5619 * is used instead of a computed schedule.
5621 * Each of these bands B is then tiled according to "tile" sizes, resulting
5622 * in two nested bands, with a kernel marker on top
5630 * We then split off at most 2 parallel dimensions from the T band and
5631 * at most 3 parallel dimension from the P band
5644 * A filter is introduced in front of T1 that maps the domain instances
5645 * to block identifiers. Similarly, a filter is introduced in front of P1
5646 * that maps the domain instances to thread identifiers.
5648 * For each iteration of the T2 band and for each array, we compute
5649 * the array elements accessed by that iteration, construct a rectangular
5650 * box around it and shift it to the origin. The result is used
5651 * as shared memory for the array.
5653 * Copying and synchronization statements are added to this schedule tree.
5654 * In principle, these are added in front of the P1 band, but some of
5655 * them may get hoisted up to higher levels.
5657 * The entire AST is then generated from the single resulting schedule tree.
5658 * During the generation the subtrees at kernel nodes (K) are saved
5659 * aside and replaced by kernel calls. The result is printed as host code
5660 * while the saved subtrees are printed as device code.
5662 static __isl_give isl_printer
*generate(__isl_take isl_printer
*p
,
5663 struct gpu_gen
*gen
, struct ppcg_scop
*scop
,
5664 struct ppcg_options
*options
)
5666 struct gpu_prog
*prog
;
5668 isl_schedule
*schedule
;
5669 isl_bool any_permutable
;
5672 return isl_printer_free(p
);
5674 ctx
= isl_printer_get_ctx(p
);
5675 prog
= gpu_prog_alloc(ctx
, scop
);
5677 return isl_printer_free(p
);
5680 schedule
= get_schedule(gen
);
5682 any_permutable
= has_any_permutable_node(schedule
);
5683 if (any_permutable
< 0 || !any_permutable
) {
5684 if (any_permutable
< 0)
5685 p
= isl_printer_free(p
);
5687 p
= print_cpu(p
, scop
, options
);
5688 isl_schedule_free(schedule
);
5690 schedule
= map_to_device(gen
, schedule
);
5691 gen
->tree
= generate_code(gen
, schedule
);
5692 p
= ppcg_set_macro_names(p
);
5693 p
= ppcg_print_exposed_declarations(p
, prog
->scop
);
5694 p
= gen
->print(p
, gen
->prog
, gen
->tree
, &gen
->types
,
5696 isl_ast_node_free(gen
->tree
);
5699 gpu_prog_free(prog
);
5704 /* Wrapper around generate for use as a ppcg_transform callback.
5706 static __isl_give isl_printer
*generate_wrap(__isl_take isl_printer
*p
,
5707 struct ppcg_scop
*scop
, void *user
)
5709 struct gpu_gen
*gen
= user
;
5711 return generate(p
, gen
, scop
, gen
->options
);
5714 /* Transform the code in the file called "input" by replacing
5715 * all scops by corresponding GPU code and write the results to "out".
5717 int generate_gpu(isl_ctx
*ctx
, const char *input
, FILE *out
,
5718 struct ppcg_options
*options
,
5719 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
5720 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
5721 struct gpu_types
*types
, void *user
), void *user
)
5728 gen
.sizes
= extract_sizes_from_str(ctx
, options
->sizes
);
5729 gen
.options
= options
;
5732 gen
.print_user
= user
;
5734 gen
.types
.name
= NULL
;
5736 if (options
->debug
->dump_sizes
) {
5737 isl_space
*space
= isl_space_params_alloc(ctx
, 0);
5738 gen
.used_sizes
= isl_union_map_empty(space
);
5741 r
= ppcg_transform(ctx
, input
, out
, options
, &generate_wrap
, &gen
);
5743 if (options
->debug
->dump_sizes
) {
5744 isl_union_map_dump(gen
.used_sizes
);
5745 isl_union_map_free(gen
.used_sizes
);
5748 isl_union_map_free(gen
.sizes
);
5749 for (i
= 0; i
< gen
.types
.n
; ++i
)
5750 free(gen
.types
.name
[i
]);
5751 free(gen
.types
.name
);
5756 /* Compute the set of inner array elements that may have their values
5757 * preserved by "prog". In particular, collect the array elements of
5758 * arrays that are not local to "prog" and remove those elements that
5759 * are definitely killed or definitely written by "prog".
5761 static __isl_give isl_union_set
*compute_may_persist(struct gpu_prog
*prog
)
5764 isl_union_set
*may_persist
, *killed
;
5765 isl_union_map
*must_kill
;
5767 may_persist
= isl_union_set_empty(isl_set_get_space(prog
->context
));
5768 for (i
= 0; i
< prog
->n_array
; ++i
) {
5771 if (prog
->array
[i
].local
)
5774 extent
= isl_set_copy(prog
->array
[i
].extent
);
5775 may_persist
= isl_union_set_add_set(may_persist
, extent
);
5778 may_persist
= isl_union_set_intersect_params(may_persist
,
5779 isl_set_copy(prog
->context
));
5780 may_persist
= isl_union_set_apply(may_persist
,
5781 isl_union_map_copy(prog
->to_inner
));
5782 must_kill
= isl_union_map_copy(prog
->tagged_must_kill
);
5783 killed
= isl_union_map_range(must_kill
);
5784 must_kill
= isl_union_map_copy(prog
->must_write
);
5785 killed
= isl_union_set_union(killed
, isl_union_map_range(must_kill
));
5787 may_persist
= isl_union_set_subtract(may_persist
, killed
);
5791 struct gpu_prog
*gpu_prog_alloc(isl_ctx
*ctx
, struct ppcg_scop
*scop
)
5793 struct gpu_prog
*prog
;
5800 prog
= isl_calloc_type(ctx
, struct gpu_prog
);
5806 prog
->context
= isl_set_copy(scop
->context
);
5807 prog
->n_stmts
= scop
->pet
->n_stmt
;
5808 prog
->any_to_outer
= pet_scop_compute_outer_to_any(scop
->pet
);
5809 prog
->any_to_outer
= isl_union_map_reverse(prog
->any_to_outer
);
5810 space
= isl_union_map_get_space(prog
->any_to_outer
);
5811 space
= isl_space_set_from_params(space
);
5812 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
5813 space
= isl_space_map_from_set(space
);
5814 id
= isl_map_identity(space
);
5815 prog
->any_to_outer
= isl_union_map_add_map(prog
->any_to_outer
, id
);
5816 prog
->stmts
= extract_stmts(ctx
, scop
, prog
->any_to_outer
);
5817 prog
->read
= isl_union_map_copy(scop
->reads
);
5818 prog
->may_write
= isl_union_map_copy(scop
->may_writes
);
5819 prog
->must_write
= isl_union_map_copy(scop
->must_writes
);
5820 prog
->tagged_must_kill
= isl_union_map_copy(scop
->tagged_must_kills
);
5821 prog
->to_inner
= pet_scop_compute_outer_to_inner(scop
->pet
);
5822 prog
->to_outer
= isl_union_map_copy(prog
->to_inner
);
5823 prog
->to_outer
= isl_union_map_reverse(prog
->to_outer
);
5826 return gpu_prog_free(prog
);
5828 if (collect_array_info(prog
) < 0)
5829 return gpu_prog_free(prog
);
5830 prog
->may_persist
= compute_may_persist(prog
);
5835 void *gpu_prog_free(struct gpu_prog
*prog
)
5839 free_array_info(prog
);
5840 free_stmts(prog
->stmts
, prog
->n_stmts
);
5841 isl_union_map_free(prog
->any_to_outer
);
5842 isl_union_map_free(prog
->to_outer
);
5843 isl_union_map_free(prog
->to_inner
);
5844 isl_union_map_free(prog
->read
);
5845 isl_union_map_free(prog
->may_write
);
5846 isl_union_map_free(prog
->must_write
);
5847 isl_union_map_free(prog
->tagged_must_kill
);
5848 isl_union_map_free(prog
->array_order
);
5849 isl_union_set_free(prog
->may_persist
);
5850 isl_set_free(prog
->context
);