1 /*
2  * Copyright 2010-2011 INRIA Saclay
3  * Copyright 2012-2014 Ecole Normale Superieure
4  * Copyright 2015      Sven Verdoolaege
5  *
6  * Use of this software is governed by the MIT license
7  *
8  * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
9  * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
10  * 91893 Orsay, France
11  * and Ecole Normale Superieure, 45 rue d'Ulm, 75230 Paris, France
12  */
13 
14 #include <isl/constraint.h>
15 #include <isl/ilp.h>
16 
17 #include "gpu_array_tile.h"
18 #include "gpu_group.h"
19 #include "gpu_tree.h"
20 #include "schedule.h"
21 
22 /* Print the name of the local copy of a given group of array references.
23  */
gpu_array_ref_group_print_name(struct gpu_array_ref_group * group,__isl_take isl_printer * p)24 __isl_give isl_printer *gpu_array_ref_group_print_name(
25 	struct gpu_array_ref_group *group, __isl_take isl_printer *p)
26 {
27 	int global = 0;
28 	enum ppcg_group_access_type type;
29 
30 	type = gpu_array_ref_group_type(group);
31 	if (type == ppcg_access_private)
32 		p = isl_printer_print_str(p, "private_");
33 	else if (type == ppcg_access_shared)
34 		p = isl_printer_print_str(p, "shared_");
35 	else
36 		global = 1;
37 	p = isl_printer_print_str(p, group->array->name);
38 	if (!global && group->local_array->n_group > 1) {
39 		p = isl_printer_print_str(p, "_");
40 		p = isl_printer_print_int(p, group->nr);
41 	}
42 
43 	return p;
44 }
45 
46 /* Return the union of all read (read = 1) and/or write (write = 1)
47  * access relations in the group.
48  */
gpu_array_ref_group_access_relation(struct gpu_array_ref_group * group,int read,int write)49 __isl_give isl_union_map *gpu_array_ref_group_access_relation(
50 	struct gpu_array_ref_group *group, int read, int write)
51 {
52 	int i;
53 	isl_union_map *access;
54 
55 	access = isl_union_map_empty(isl_map_get_space(group->access));
56 	for (i = 0; i < group->n_ref; ++i) {
57 		isl_map *map_i;
58 
59 		if (!((read && group->refs[i]->read) ||
60 		     (write && group->refs[i]->write)))
61 			continue;
62 		map_i = isl_map_copy(group->refs[i]->access);
63 		access = isl_union_map_union(access,
64 					    isl_union_map_from_map(map_i));
65 	}
66 
67 	return access;
68 }
69 
70 /* Should this array reference group be mapped to private, shared or global
71  * memory?
72  * If we have computed both a private and a shared tile, then
73  * the tile with the smallest depth is used.  If both have the same depth,
74  * then the private tile is used.
75  */
gpu_array_ref_group_type(struct gpu_array_ref_group * group)76 enum ppcg_group_access_type gpu_array_ref_group_type(
77 	struct gpu_array_ref_group *group)
78 {
79 	if (group->private_tile && group->shared_tile &&
80 	    group->shared_tile->depth < group->private_tile->depth)
81 		return ppcg_access_shared;
82 	if (group->private_tile)
83 		return ppcg_access_private;
84 	if (group->shared_tile)
85 		return ppcg_access_shared;
86 	return ppcg_access_global;
87 }
88 
89 
90 /* Return the effective gpu_array_tile associated to "group" or
91  * NULL if there is no such gpu_array_tile.
92  */
gpu_array_ref_group_tile(struct gpu_array_ref_group * group)93 struct gpu_array_tile *gpu_array_ref_group_tile(
94 	struct gpu_array_ref_group *group)
95 {
96 	switch (gpu_array_ref_group_type(group)) {
97 	case ppcg_access_global:
98 		return NULL;
99 	case ppcg_access_shared:
100 		return group->shared_tile;
101 	case ppcg_access_private:
102 		return group->private_tile;
103 	}
104 }
105 
106 /* Does the tile associated to "group" require unrolling of the schedule
107  * dimensions mapped to threads?
108  * Note that this can only happen for private tiles.
109  */
gpu_array_ref_group_requires_unroll(struct gpu_array_ref_group * group)110 int gpu_array_ref_group_requires_unroll(struct gpu_array_ref_group *group)
111 {
112 	struct gpu_array_tile *tile;
113 
114 	tile = gpu_array_ref_group_tile(group);
115 	if (!tile)
116 		return 0;
117 	return tile->requires_unroll;
118 }
119 
120 /* Given a constraint
121  *
122  *		a(p,i) + j = g f(e)
123  *
124  * or -a(p,i) - j = g f(e) if sign < 0,
125  * store a(p,i) in bound->shift and g (stride) in bound->stride.
126  * a(p,i) is assumed to be an expression in only the parameters
127  * and the input dimensions.
128  */
extract_stride(__isl_keep isl_constraint * c,struct gpu_array_bound * bound,__isl_keep isl_val * stride,int sign)129 static void extract_stride(__isl_keep isl_constraint *c,
130 	struct gpu_array_bound *bound, __isl_keep isl_val *stride, int sign)
131 {
132 	int i;
133 	isl_val *v;
134 	isl_space *space;
135 	unsigned nparam;
136 	unsigned nvar;
137 	isl_aff *aff;
138 
139 	isl_val_free(bound->stride);
140 	bound->stride = isl_val_copy(stride);
141 
142 	space = isl_constraint_get_space(c);
143 	space = isl_space_domain(space);
144 
145 	nparam = isl_space_dim(space, isl_dim_param);
146 	nvar = isl_space_dim(space, isl_dim_set);
147 
148 	v = isl_constraint_get_constant_val(c);
149 	if (sign < 0)
150 		v = isl_val_neg(v);
151 	aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
152 	aff = isl_aff_set_constant_val(aff, v);
153 
154 	for (i = 0; i < nparam; ++i) {
155 		if (!isl_constraint_involves_dims(c, isl_dim_param, i, 1))
156 			continue;
157 		v = isl_constraint_get_coefficient_val(c, isl_dim_param, i);
158 		if (sign < 0)
159 			v = isl_val_neg(v);
160 		aff = isl_aff_add_coefficient_val(aff, isl_dim_param, i, v);
161 	}
162 
163 	for (i = 0; i < nvar; ++i) {
164 		if (!isl_constraint_involves_dims(c, isl_dim_in, i, 1))
165 			continue;
166 		v = isl_constraint_get_coefficient_val(c, isl_dim_in, i);
167 		if (sign < 0)
168 			v = isl_val_neg(v);
169 		aff = isl_aff_add_coefficient_val(aff, isl_dim_in, i, v);
170 	}
171 
172 	bound->shift = aff;
173 }
174 
175 /* Given an equality constraint of a map with a single output dimension j,
176  * check if the constraint is of the form
177  *
178  *		a(p,i) + j = g f(e)
179  *
180  * with a(p,i) an expression in the parameters and input dimensions
181  * and f(e) an expression in the existentially quantified variables.
182  * If so, and if g is larger than any such g from a previously considered
183  * constraint, then call extract_stride to record the stride information
184  * in bound.
185  */
check_stride_constraint(__isl_take isl_constraint * c,void * user)186 static isl_stat check_stride_constraint(__isl_take isl_constraint *c,
187 	void *user)
188 {
189 	int i;
190 	isl_ctx *ctx;
191 	isl_val *v;
192 	unsigned n_div;
193 	struct gpu_array_bound *bound = user;
194 
195 	ctx = isl_constraint_get_ctx(c);
196 	n_div = isl_constraint_dim(c, isl_dim_div);
197 	v = isl_constraint_get_coefficient_val(c, isl_dim_out, 0);
198 
199 	if (n_div && (isl_val_is_one(v) || isl_val_is_negone(v))) {
200 		int s = isl_val_sgn(v);
201 		isl_val *stride = isl_val_zero(ctx);
202 
203 		isl_val_free(v);
204 		for (i = 0; i < n_div; ++i) {
205 			v = isl_constraint_get_coefficient_val(c,
206 								isl_dim_div, i);
207 			stride = isl_val_gcd(stride, v);
208 		}
209 		if (!isl_val_is_zero(stride) &&
210 		    isl_val_gt(stride, bound->stride))
211 			extract_stride(c, bound, stride, s);
212 
213 		isl_val_free(stride);
214 	} else
215 		isl_val_free(v);
216 
217 	isl_constraint_free(c);
218 	return isl_stat_ok;
219 }
220 
221 /* Given contraints on an array index i, check if we can find
222  * a shift a(p) and a stride g such that
223  *
224  *	a(p) + i = 0 mod g
225  *
226  * If so, record the information in bound and apply the mapping
227  * i -> (i + a(p))/g to the array index in bounds and return
228  * the new constraints.
229  * If not, simply return the original constraints.
230  *
231  * If bounds is a subset of the space
232  *
233  *	D -> i
234  *
235  * then the bound recorded in bound->shift is of the form
236  *
237  *	D -> s(D)
238  *
239  * with s(D) equal to a(p) above.
240  * Next, we construct a mapping of the form
241  *
242  *	[D -> i] -> [D -> (i + S(D))/g]
243  *
244  * This mapping is computed as follows.
245  * We first introduce "i" in the domain through precomposition
246  * with [D -> i] -> D obtaining
247  *
248  *	[D -> i] -> s(D)
249  *
250  * Adding [D -> i] -> i produces
251  *
252  *	[D -> i] -> i + s(D)
253  *
254  * and the domain product with [D -> i] -> D yields
255  *
256  *	[D -> i] -> [D -> i + s(D)]
257  *
258  * Composition with [D -> i] -> [D -> i/g] gives the desired result.
259  */
check_stride(struct gpu_array_bound * bound,__isl_take isl_basic_map * bounds)260 static __isl_give isl_basic_map *check_stride(struct gpu_array_bound *bound,
261 	__isl_take isl_basic_map *bounds)
262 {
263 	isl_space *space;
264 	isl_basic_map *hull;
265 	isl_basic_map *shift, *id, *bmap, *scale;
266 	isl_basic_set *bset;
267 	isl_aff *aff;
268 
269 	bound->stride = NULL;
270 
271 	hull = isl_basic_map_affine_hull(isl_basic_map_copy(bounds));
272 
273 	isl_basic_map_foreach_constraint(hull, &check_stride_constraint, bound);
274 
275 	isl_basic_map_free(hull);
276 
277 	if (!bound->stride)
278 		return bounds;
279 
280 	shift = isl_basic_map_from_aff(isl_aff_copy(bound->shift));
281 	space = isl_basic_map_get_space(bounds);
282 	bmap = isl_basic_map_domain_map(isl_basic_map_universe(space));
283 	shift = isl_basic_map_apply_range(bmap, shift);
284 	space = isl_basic_map_get_space(bounds);
285 	id = isl_basic_map_range_map(isl_basic_map_universe(space));
286 	shift = isl_basic_map_sum(id, shift);
287 	space = isl_basic_map_get_space(bounds);
288 	id = isl_basic_map_domain_map(isl_basic_map_universe(space));
289 	shift = isl_basic_map_range_product(id, shift);
290 
291 	space = isl_space_domain(isl_basic_map_get_space(bounds));
292 	id = isl_basic_map_identity(isl_space_map_from_set(space));
293 	space = isl_space_range(isl_basic_map_get_space(bounds));
294 	aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
295 	aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
296 	aff = isl_aff_scale_down_val(aff, isl_val_copy(bound->stride));
297 	scale = isl_basic_map_from_aff(aff);
298 	scale = isl_basic_map_product(id, scale);
299 
300 	bmap = isl_basic_map_apply_range(shift, scale);
301 	bset = isl_basic_set_apply(isl_basic_map_wrap(bounds), bmap);
302 	bounds = isl_basic_set_unwrap(bset);
303 
304 	return bounds;
305 }
306 
307 /* Data used in compute_array_dim_size and compute_size_in_direction.
308  *
309  * pos is the position of the variable representing the array index,
310  * i.e., the variable for which want to compute the size.  This variable
311  * is also the last variable in the set.
312  */
313 struct gpu_size_info {
314 	isl_basic_set *bset;
315 	struct gpu_array_bound *bound;
316 	int pos;
317 };
318 
319 /* Given a constraint from the basic set describing the bounds on
320  * an array index, check if it is a lower bound, say m i >= b(x), and,
321  * if so, check whether the expression "i - ceil(b(x)/m) + 1" has a constant
322  * upper bound.  If so, and if this bound is smaller than any bound
323  * derived from earlier constraints, set the size to this bound on
324  * the expression and the lower bound to ceil(b(x)/m).
325  */
compute_size_in_direction(__isl_take isl_constraint * c,void * user)326 static isl_stat compute_size_in_direction(__isl_take isl_constraint *c,
327 	void *user)
328 {
329 	struct gpu_size_info *size = user;
330 	unsigned nparam;
331 	unsigned n_div;
332 	isl_val *v;
333 	isl_aff *aff;
334 	isl_aff *lb;
335 
336 	nparam = isl_basic_set_dim(size->bset, isl_dim_param);
337 	n_div = isl_constraint_dim(c, isl_dim_div);
338 
339 	if (isl_constraint_involves_dims(c, isl_dim_div, 0, n_div) ||
340 	    !isl_constraint_is_lower_bound(c, isl_dim_set, size->pos)) {
341 		isl_constraint_free(c);
342 		return isl_stat_ok;
343 	}
344 
345 	aff = isl_constraint_get_bound(c, isl_dim_set, size->pos);
346 	aff = isl_aff_ceil(aff);
347 
348 	lb = isl_aff_copy(aff);
349 
350 	aff = isl_aff_neg(aff);
351 	aff = isl_aff_add_coefficient_si(aff, isl_dim_in, size->pos, 1);
352 
353 	v = isl_basic_set_max_val(size->bset, aff);
354 	isl_aff_free(aff);
355 
356 	if (isl_val_is_int(v)) {
357 		v = isl_val_add_ui(v, 1);
358 		if (!size->bound->size || isl_val_lt(v, size->bound->size)) {
359 			isl_val_free(size->bound->size);
360 			size->bound->size = isl_val_copy(v);
361 			lb = isl_aff_drop_dims(lb, isl_dim_in, size->pos, 1);
362 			isl_aff_free(size->bound->lb);
363 			size->bound->lb = isl_aff_copy(lb);
364 		}
365 	}
366 	isl_val_free(v);
367 	isl_aff_free(lb);
368 
369 	isl_constraint_free(c);
370 
371 	return isl_stat_ok;
372 }
373 
374 /* Given a basic map "bounds" that maps parameters and input dimensions
375  * to a single output dimension, look for an expression in the parameters
376  * and input dimensions such that the range of the output dimension shifted
377  * by this expression is a constant.
378  *
379  * In particular, we currently only consider lower bounds on the output
380  * dimension as candidate expressions.
381  */
compute_array_dim_size(struct gpu_array_bound * bound,__isl_take isl_basic_map * bounds)382 static int compute_array_dim_size(struct gpu_array_bound *bound,
383 	__isl_take isl_basic_map *bounds)
384 {
385 	struct gpu_size_info size;
386 
387 	bounds = isl_basic_map_detect_equalities(bounds);
388 	bounds = check_stride(bound, bounds);
389 
390 	bound->size = NULL;
391 	bound->lb = NULL;
392 
393 	size.bound = bound;
394 	size.pos = isl_basic_map_dim(bounds, isl_dim_in);
395 	size.bset = isl_basic_map_wrap(bounds);
396 	size.bset = isl_basic_set_flatten(size.bset);
397 	size.bset = isl_set_simple_hull(isl_basic_set_compute_divs(size.bset));
398 	isl_basic_set_foreach_constraint(size.bset, &compute_size_in_direction,
399 					&size);
400 	isl_basic_set_free(size.bset);
401 
402 	return bound->size ? 0 : -1;
403 }
404 
405 /* Check if we can find a memory tile for the given array
406  * based on the given accesses, and if so, put the results in "tile".
407  *
408  * We project the accesses on each index in turn and look for a parametric
409  * offset such that the size is constant.
410  *
411  * tile->depth is initialized to the input dimension of the computed bounds.
412  */
can_tile(__isl_keep isl_map * access,struct gpu_array_tile * tile)413 static int can_tile(__isl_keep isl_map *access, struct gpu_array_tile *tile)
414 {
415 	int i;
416 
417 	tile->depth = isl_map_dim(access, isl_dim_in);
418 
419 	for (i = 0; i < tile->n; ++i) {
420 		isl_map *access_i;
421 		isl_basic_map *hull;
422 
423 		access_i = isl_map_copy(access);
424 		access_i = isl_map_project_out(access_i, isl_dim_out, 0, i);
425 		access_i = isl_map_project_out(access_i, isl_dim_out,
426 					    1, tile->n - (i + 1));
427 		access_i = isl_map_compute_divs(access_i);
428 		hull = isl_map_simple_hull(access_i);
429 		if (compute_array_dim_size(&tile->bound[i], hull) < 0)
430 			return 0;
431 	}
432 
433 	return 1;
434 }
435 
436 /* Internal data structure for gpu_group_references.
437  *
438  * scop represents the input scop.
439  * kernel_depth is the schedule depth where the kernel launch will
440  * be introduced, i.e., it is the depth of the band that is mapped
441  * to blocks.
442  * shared_depth is the schedule depth at which the copying to/from
443  * shared memory is computed.  The copy operation may then
444  * later be hoisted to a higher level.
445  * thread_depth is the schedule depth where the thread mark is located,
446  * i.e., it is the depth of the band that is mapped to threads and also
447  * the schedule depth at which the copying to/from private memory
448  * is computed.  The copy operation may then later be hoisted to
449  * a higher level.
450  * n_thread is the number of schedule dimensions in the band that
451  * is mapped to threads.
452  * privatization lives in the range of thread_sched (i.e., it is
453  * of dimension thread_depth + n_thread) and encodes the mapping
454  * to thread identifiers (as parameters).
455  * host_sched contains the kernel_depth dimensions of the host schedule.
456  * shared_sched contains the first shared_depth dimensions of the
457  * kernel schedule.
458  * copy_sched contains the first thread_depth dimensions of the
459  * kernel schedule.
460  * thread_sched contains the first (thread_depth + n_thread) dimensions
461  * of the kernel schedule.
462  * full_sched is a union_map representation of the entire kernel schedule.
463  * The schedules are all formulated in terms of the original statement
464  * instances, i.e., those that appear in the domains of the access
465  * relations.
466  */
467 struct gpu_group_data {
468 	struct ppcg_scop *scop;
469 	int kernel_depth;
470 	int shared_depth;
471 	int thread_depth;
472 	int n_thread;
473 	isl_set *privatization;
474 	isl_union_map *host_sched;
475 	isl_union_map *shared_sched;
476 	isl_union_map *copy_sched;
477 	isl_union_map *thread_sched;
478 	isl_union_map *full_sched;
479 };
480 
481 /* Construct a map from domain_space to domain_space that increments
482  * the dimension at position "pos" and leaves all other dimensions
483  * constant.
484  */
next(__isl_take isl_space * domain_space,int pos)485 static __isl_give isl_map *next(__isl_take isl_space *domain_space, int pos)
486 {
487 	isl_space *space;
488 	isl_aff *aff;
489 	isl_multi_aff *next;
490 
491 	space = isl_space_map_from_set(domain_space);
492 	next = isl_multi_aff_identity(space);
493 	aff = isl_multi_aff_get_aff(next, pos);
494 	aff = isl_aff_add_constant_si(aff, 1);
495 	next = isl_multi_aff_set_aff(next, pos, aff);
496 
497 	return isl_map_from_multi_aff(next);
498 }
499 
500 /* Check if the given access is coalesced (or if there is no point
501  * in trying to coalesce the access by mapping the array to shared memory).
502  * That is, check whether incrementing the dimension that will get
503  * wrapped over the last thread index results in incrementing
504  * the last array index.
505  *
506  * If no two consecutive array elements are ever accessed by "access",
507  * then mapping the corresponding array to shared memory will not
508  * improve coalescing.  In fact, the copying will likely be performed
509  * by a single thread.  Consider the access as coalesced such that
510  * the caller will not try and map the array to shared memory just
511  * to improve coalescing.
512  *
513  * This function is only called for access relations without reuse and
514  * kernels with at least one thread identifier.
515  */
access_is_coalesced(struct gpu_group_data * data,__isl_keep isl_union_map * access)516 static int access_is_coalesced(struct gpu_group_data *data,
517 	__isl_keep isl_union_map *access)
518 {
519 	int dim;
520 	isl_space *space;
521 	isl_set *accessed;
522 	isl_map *access_map;
523 	isl_map *next_thread_x;
524 	isl_map *next_element;
525 	isl_map *map;
526 	int coalesced, empty;
527 
528 	access = isl_union_map_copy(access);
529 	access = isl_union_map_apply_domain(access,
530 				isl_union_map_copy(data->full_sched));
531 	access_map = isl_map_from_union_map(access);
532 
533 	space = isl_map_get_space(access_map);
534 	space = isl_space_range(space);
535 	dim = isl_space_dim(space, isl_dim_set);
536 	if (dim == 0)
537 		next_element = isl_map_empty(isl_space_map_from_set(space));
538 	else
539 		next_element = next(space, dim - 1);
540 
541 	accessed = isl_map_range(isl_map_copy(access_map));
542 	map = isl_map_copy(next_element);
543 	map = isl_map_intersect_domain(map, isl_set_copy(accessed));
544 	map = isl_map_intersect_range(map, accessed);
545 	empty = isl_map_is_empty(map);
546 	isl_map_free(map);
547 
548 	if (empty < 0 || empty) {
549 		isl_map_free(next_element);
550 		isl_map_free(access_map);
551 		return empty;
552 	}
553 
554 	space = isl_map_get_space(access_map);
555 	space = isl_space_domain(space);
556 	next_thread_x = next(space, data->thread_depth + data->n_thread - 1);
557 
558 	map = isl_map_apply_domain(next_thread_x, isl_map_copy(access_map));
559 	map = isl_map_apply_range(map, access_map);
560 
561 	coalesced = isl_map_is_subset(map, next_element);
562 
563 	isl_map_free(next_element);
564 	isl_map_free(map);
565 
566 	return coalesced;
567 }
568 
569 /* Replace the host schedule dimensions in the access relation "access"
570  * by parameters, so that they are treated as fixed when checking for reuse
571  * (within a kernel) or whether two consecutive elements are accessed
572  * (within a kernel).
573  */
localize_access(struct gpu_group_data * data,__isl_take isl_union_map * access)574 static __isl_give isl_union_map *localize_access(struct gpu_group_data *data,
575 	__isl_take isl_union_map *access)
576 {
577 	int n;
578 	isl_space *space;
579 	isl_set *param;
580 	isl_union_map *umap;
581 	isl_id_list *ids;
582 
583 	umap = isl_union_map_copy(data->host_sched);
584 	space = isl_union_map_get_space(umap);
585 	n = data->kernel_depth;
586 	ids = ppcg_scop_generate_names(data->scop, n, "__ppcg_host_");
587 	param = parametrization(space, n, 0, ids);
588 	isl_id_list_free(ids);
589 	umap = isl_union_map_intersect_range(umap,
590 						isl_union_set_from_set(param));
591 	access = isl_union_map_intersect_domain(access,
592 						isl_union_map_domain(umap));
593 
594 	return access;
595 }
596 
597 /* Given an access relation in terms of at least data->thread_depth initial
598  * dimensions of the computed schedule, check if it is bijective for
599  * fixed values of the first data->thread_depth dimensions.
600  * We perform this check by equating these dimensions to parameters.
601  */
access_is_bijective(struct gpu_group_data * data,__isl_keep isl_map * access)602 static int access_is_bijective(struct gpu_group_data *data,
603 	__isl_keep isl_map *access)
604 {
605 	int res;
606 	int dim;
607 	isl_set *par;
608 	isl_space *space;
609 	isl_id_list *ids;
610 
611 	access = isl_map_copy(access);
612 	space = isl_space_params(isl_map_get_space(access));
613 	ids = ppcg_scop_generate_names(data->scop, data->thread_depth, "s");
614 	dim = isl_map_dim(access, isl_dim_in);
615 	par = parametrization(space, dim, 0, ids);
616 	isl_id_list_free(ids);
617 	access = isl_map_intersect_domain(access, par);
618 	res = isl_map_is_bijective(access);
619 	isl_map_free(access);
620 
621 	return res;
622 }
623 
624 /* Compute the number of outer schedule tile dimensions that affect
625  * the offset of "tile".
626  * If there is no such dimension, then return the index
627  * of the first kernel dimension, i.e., data->kernel_depth.
628  */
compute_tile_depth(struct gpu_group_data * data,struct gpu_array_tile * tile)629 static int compute_tile_depth(struct gpu_group_data *data,
630 	struct gpu_array_tile *tile)
631 {
632 	int i, j;
633 
634 	for (j = tile->depth - 1; j >= data->kernel_depth; --j) {
635 		for (i = 0; i < tile->n; ++i) {
636 			isl_aff *lb;
637 			isl_aff *shift;
638 
639 			lb = tile->bound[i].lb;
640 			if (isl_aff_involves_dims(lb, isl_dim_in, j, 1))
641 				break;
642 
643 			shift = tile->bound[i].shift;
644 			if (!shift)
645 				continue;
646 			if (isl_aff_involves_dims(shift, isl_dim_in, j, 1))
647 				break;
648 		}
649 		if (i < tile->n)
650 			break;
651 	}
652 
653 	return ++j;
654 }
655 
656 /* Return the lowest depth between data->kernel_depth and data->thread_depth
657  * at which every array element accessed through "acc" is accessed
658  * by a single thread.  The input dimension of "acc" is
659  * data->thread_depth + data->n_thread, where the final data->n_thread
660  * dimensions are those that will be mapped to threads.
661  * If the values for these dimensions are uniquely determined
662  * by the array index and a given number of outer dimensions, then
663  * there is only one thread accessing that array element within those
664  * outer dimensions.
665  *
666  * The input space of "acc" is first split up, such that it has the form
667  *
668  *	[O -> T] -> A
669  *
670  * with O the outer dimensions, T the dimensions that will be mapped to threads
671  * and A the array index.
672  *
673  * Then the positions of T and A are interchanged to simplify the test
674  * whether T uniquely depends on O and A.
675  * In particular, the above access relation is first combined with
676  *
677  *	[O -> T] -> T
678  *
679  * to form
680  *
681  *	[O -> T] -> [A -> T]
682  *
683  * from which
684  *
685  *	O -> [A -> T]
686  *
687  * is extracted, which is then uncurried to
688  *
689  *	[O -> A] -> T
690  *
691  * Finally, the final dimensions of O are projected out one by one
692  * until T is no longer uniquely determined by A and the remaining
693  * dimensions in O.  The value returned is that of the last dimension
694  * that was successfully projected out.
695  * Note that there is no need to test whether [O -> A] -> T itself
696  * is single-valued as that was already tested in access_is_bijective.
697  */
compute_accessed_by_single_thread_depth(struct gpu_group_data * data,__isl_keep isl_map * acc)698 static int compute_accessed_by_single_thread_depth(struct gpu_group_data *data,
699 	__isl_keep isl_map *acc)
700 {
701 	int i;
702 	isl_space *space;
703 	isl_map *map;
704 	isl_bool sv;
705 
706 	if (data->thread_depth == data->kernel_depth)
707 		return data->thread_depth;
708 
709 	acc = isl_map_copy(acc);
710 
711 	space = isl_map_get_space(acc);
712 	space = isl_space_params(space);
713 	space = isl_space_set_from_params(space);
714 	space = isl_space_add_dims(space, isl_dim_set, data->thread_depth);
715 	space = isl_space_from_domain(space);
716 	space = isl_space_add_dims(space, isl_dim_out, data->n_thread);
717 	space = isl_space_wrap(space);
718 	map = isl_set_flatten_map(isl_set_universe(space));
719 	acc = isl_map_apply_range(map, acc);
720 
721 	space = isl_space_domain(isl_map_get_space(acc));
722 	map = isl_map_range_map(isl_map_universe(isl_space_unwrap(space)));
723 	acc = isl_map_range_product(acc, map);
724 	acc = isl_map_domain_factor_domain(acc);
725 	acc = isl_map_uncurry(acc);
726 
727 	for (i = data->thread_depth - 1; i >= data->kernel_depth; --i) {
728 		acc = isl_map_project_out(acc, isl_dim_in, i, 1);
729 		sv = isl_map_is_single_valued(acc);
730 		if (sv < 0)
731 			return -1;
732 		if (!sv)
733 			break;
734 	}
735 
736 	isl_map_free(acc);
737 
738 	return ++i;
739 }
740 
741 /* Adjust the fields of "tile" to reflect the new input dimension "depth".
742  * The dimension beyond "depth" are assumed not to affect the tile,
743  * so they can simply be dropped.
744  */
tile_adjust_depth(struct gpu_array_tile * tile,int depth)745 static int tile_adjust_depth(struct gpu_array_tile *tile, int depth)
746 {
747 	int i;
748 
749 	if (tile->depth == depth)
750 		return 0;
751 
752 	for (i = 0; i < tile->n; ++i) {
753 		tile->bound[i].lb = isl_aff_drop_dims(tile->bound[i].lb,
754 					isl_dim_in, depth, tile->depth - depth);
755 		if (!tile->bound[i].lb)
756 			return -1;
757 		if (!tile->bound[i].shift)
758 			continue;
759 		tile->bound[i].shift = isl_aff_drop_dims(tile->bound[i].shift,
760 					isl_dim_in, depth, tile->depth - depth);
761 		if (!tile->bound[i].shift)
762 			return -1;
763 	}
764 
765 	tile->depth = depth;
766 
767 	return 0;
768 }
769 
770 /* Determine the number of schedule dimensions that affect the offset of the
771  * shared or private tile "tile" and store the result in tile->depth, with
772  * a lower bound of data->kernel_depth.
773  * Also adjust the fields of the tile to only refer to the tile->depth
774  * outer schedule dimensions.
775  */
tile_set_depth(struct gpu_group_data * data,struct gpu_array_tile * tile)776 static isl_stat tile_set_depth(struct gpu_group_data *data,
777 	struct gpu_array_tile *tile)
778 {
779 	if (tile_adjust_depth(tile, compute_tile_depth(data, tile)) < 0)
780 		return isl_stat_error;
781 
782 	return isl_stat_ok;
783 }
784 
785 /* Determine the number of schedule dimensions that affect the offset of the
786  * shared tile and store the minimum of the private and shared tile depth
787  * in group->min_depth, with a lower bound of data->kernel_depth.
788  * If there is no tile defined on the array reference group,
789  * then set group->min_depth to data->thread_depth.
790  */
set_depth(struct gpu_group_data * data,struct gpu_array_ref_group * group)791 static int set_depth(struct gpu_group_data *data,
792 	struct gpu_array_ref_group *group)
793 {
794 	group->min_depth = data->thread_depth;
795 
796 	if (group->private_tile) {
797 		if (group->private_tile->depth < group->min_depth)
798 			group->min_depth = group->private_tile->depth;
799 	}
800 	if (group->shared_tile) {
801 		if (tile_set_depth(data, group->shared_tile) < 0)
802 			return -1;
803 		if (group->shared_tile->depth < group->min_depth)
804 			group->min_depth = group->shared_tile->depth;
805 	}
806 
807 	return 0;
808 }
809 
810 /* Fill up the groups array with singleton groups, i.e., one group
811  * per reference, initializing the array, access, write, n_ref and refs fields.
812  * In particular the access field is initialized to the scheduled
813  * access relation of the array reference.
814  *
815  * Return the number of elements initialized, i.e., the number of
816  * active references in the current kernel.
817  */
populate_array_references(struct gpu_local_array_info * local,struct gpu_array_ref_group ** groups,struct gpu_group_data * data)818 static int populate_array_references(struct gpu_local_array_info *local,
819 	struct gpu_array_ref_group **groups, struct gpu_group_data *data)
820 {
821 	int i;
822 	int n;
823 	isl_ctx *ctx = isl_union_map_get_ctx(data->copy_sched);
824 
825 	n = 0;
826 	for (i = 0; i < local->array->n_ref; ++i) {
827 		isl_union_map *umap;
828 		isl_map *map;
829 		struct gpu_array_ref_group *group;
830 		struct gpu_stmt_access *access = local->array->refs[i];
831 
832 		map = isl_map_copy(access->access);
833 		umap = isl_union_map_from_map(map);
834 		umap = isl_union_map_apply_domain(umap,
835 				isl_union_map_copy(data->copy_sched));
836 
837 		if (isl_union_map_is_empty(umap)) {
838 			isl_union_map_free(umap);
839 			continue;
840 		}
841 
842 		map = isl_map_from_union_map(umap);
843 		map = isl_map_detect_equalities(map);
844 
845 		group = isl_calloc_type(ctx, struct gpu_array_ref_group);
846 		if (!group)
847 			return -1;
848 		group->local_array = local;
849 		group->array = local->array;
850 		group->access = map;
851 		group->write = access->write;
852 		group->exact_write = access->exact_write;
853 		group->slice = access->n_index < local->array->n_index;
854 		group->refs = &local->array->refs[i];
855 		group->n_ref = 1;
856 
857 		groups[n++] = group;
858 	}
859 
860 	return n;
861 }
862 
863 /* If group->n_ref == 1, then group->refs was set by
864  * populate_array_references to point directly into
865  * group->array->refs and should not be freed.
866  * If group->n_ref > 1, then group->refs was set by join_groups
867  * to point to a newly allocated array.
868  */
gpu_array_ref_group_free(struct gpu_array_ref_group * group)869 struct gpu_array_ref_group *gpu_array_ref_group_free(
870 	struct gpu_array_ref_group *group)
871 {
872 	if (!group)
873 		return NULL;
874 	gpu_array_tile_free(group->shared_tile);
875 	gpu_array_tile_free(group->private_tile);
876 	isl_map_free(group->access);
877 	if (group->n_ref > 1)
878 		free(group->refs);
879 	free(group);
880 	return NULL;
881 }
882 
883 /* Check if the access relations of group1 and group2 overlap within
884  * copy_sched.
885  */
accesses_overlap(struct gpu_array_ref_group * group1,struct gpu_array_ref_group * group2)886 static int accesses_overlap(struct gpu_array_ref_group *group1,
887 	struct gpu_array_ref_group *group2)
888 {
889 	int disjoint;
890 
891 	disjoint = isl_map_is_disjoint(group1->access, group2->access);
892 	if (disjoint < 0)
893 		return -1;
894 
895 	return !disjoint;
896 }
897 
898 /* Combine the given two groups into a single group, containing
899  * the references of both groups.
900  */
join_groups(struct gpu_array_ref_group * group1,struct gpu_array_ref_group * group2)901 static struct gpu_array_ref_group *join_groups(
902 	struct gpu_array_ref_group *group1,
903 	struct gpu_array_ref_group *group2)
904 {
905 	int i;
906 	isl_ctx *ctx;
907 	struct gpu_array_ref_group *group;
908 
909 	if (!group1 || !group2)
910 		return NULL;
911 
912 	ctx = isl_map_get_ctx(group1->access);
913 	group = isl_calloc_type(ctx, struct gpu_array_ref_group);
914 	if (!group)
915 		return NULL;
916 	group->local_array = group1->local_array;
917 	group->array = group1->array;
918 	group->access = isl_map_union(isl_map_copy(group1->access),
919 					isl_map_copy(group2->access));
920 	group->write = group1->write || group2->write;
921 	group->exact_write = group1->exact_write && group2->exact_write;
922 	group->slice = group1->slice || group2->slice;
923 	group->n_ref = group1->n_ref + group2->n_ref;
924 	group->refs = isl_alloc_array(ctx, struct gpu_stmt_access *,
925 					group->n_ref);
926 	if (!group->refs)
927 		return gpu_array_ref_group_free(group);
928 	for (i = 0; i < group1->n_ref; ++i)
929 		group->refs[i] = group1->refs[i];
930 	for (i = 0; i < group2->n_ref; ++i)
931 		group->refs[group1->n_ref + i] = group2->refs[i];
932 
933 	return group;
934 }
935 
936 /* Combine the given two groups into a single group and free
937  * the original two groups.
938  */
join_groups_and_free(struct gpu_array_ref_group * group1,struct gpu_array_ref_group * group2)939 static struct gpu_array_ref_group *join_groups_and_free(
940 	struct gpu_array_ref_group *group1,
941 	struct gpu_array_ref_group *group2)
942 {
943 	struct gpu_array_ref_group *group;
944 
945 	group = join_groups(group1, group2);
946 	gpu_array_ref_group_free(group1);
947 	gpu_array_ref_group_free(group2);
948 	return group;
949 }
950 
951 /* Report that the array reference group with the given access relation
952  * is not mapped to shared memory in the given kernel because
953  * it does not exhibit any reuse and is considered to be coalesced.
954  */
report_no_reuse_and_coalesced(struct ppcg_kernel * kernel,__isl_keep isl_union_map * access)955 static void report_no_reuse_and_coalesced(struct ppcg_kernel *kernel,
956 	__isl_keep isl_union_map *access)
957 {
958 	isl_ctx *ctx;
959 	isl_printer *p;
960 
961 	ctx = isl_union_map_get_ctx(access);
962 	p = isl_printer_to_file(ctx, stdout);
963 	p = isl_printer_print_str(p, "Array reference group ");
964 	p = isl_printer_print_union_map(p, access);
965 	p = isl_printer_print_str(p,
966 	    " not considered for mapping to shared memory in kernel");
967 	p = isl_printer_print_int(p, kernel->id);
968 	p = isl_printer_print_str(p,
969 	    " because it exhibits no reuse and is considered to be coalesced");
970 	p = isl_printer_end_line(p);
971 	isl_printer_free(p);
972 }
973 
974 /* Given an access relation in terms of the data->thread_depth initial
975  * dimensions of the computed schedule and the thread identifiers
976  * (as parameters), check if the use of the corresponding private tile
977  * requires unrolling.
978  *
979  * If we are creating a private tile because we are forced to,
980  * then no unrolling is required.
981  * Otherwise we check if "access" is bijective and unrolling
982  * is required if it is not.  Note that the access relation
983  * has already been determined to be bijective before the introduction
984  * of the thread identifiers and the removal of the schedule dimensions
985  * that are mapped to these threads.  If the access relation is no longer
986  * bijective, then this means that more than one value of one of those
987  * schedule dimensions is mapped to the same thread and therefore
988  * unrolling is required.
989  */
check_requires_unroll(struct gpu_group_data * data,__isl_keep isl_map * access,int force_private)990 static int check_requires_unroll(struct gpu_group_data *data,
991 	__isl_keep isl_map *access, int force_private)
992 {
993 	int bijective;
994 
995 	if (force_private)
996 		return 0;
997 	bijective = access_is_bijective(data, access);
998 	if (bijective < 0)
999 		return -1;
1000 	return !bijective;
1001 }
1002 
1003 /* Map the domain of "access" to the outer data->shared_depth
1004  * schedule dimensions.  When data->shared_depth is equal to
1005  * data->thread_depth, this result is already available in group->access.
1006  */
shared_access(struct gpu_array_ref_group * group,__isl_keep isl_union_map * access,struct gpu_group_data * data)1007 static __isl_give isl_map *shared_access(struct gpu_array_ref_group *group,
1008 	__isl_keep isl_union_map *access, struct gpu_group_data *data)
1009 {
1010 	isl_union_map *shared;
1011 
1012 	if (data->shared_depth == data->thread_depth)
1013 		return isl_map_copy(group->access);
1014 
1015 	shared = isl_union_map_copy(access);
1016 	shared = isl_union_map_apply_domain(shared,
1017 			isl_union_map_copy(data->shared_sched));
1018 	return isl_map_from_union_map(shared);
1019 }
1020 
1021 /* Compute the private and/or shared memory tiles for the array
1022  * reference group "group" of array "array".
1023  * Return 0 on success and -1 on error.
1024  *
1025  * If the array is a read-only scalar or if the user requested
1026  * not to use shared or private memory, then we do not need to do anything.
1027  *
1028  * If any reference in the reference group accesses more than one element,
1029  * then we would have to make sure that the layout in shared memory
1030  * is the same as that in global memory.  Since we do not handle this yet
1031  * (and it may not even be possible), we refuse to map to private or
1032  * shared memory in such cases.
1033  *
1034  * If the array group involves any may writes (that are not must writes),
1035  * then we would have to make sure that we load the data into shared/private
1036  * memory first in case the data is not written by the kernel
1037  * (but still written back out to global memory).
1038  * Since we don't have any such mechanism at the moment, we don't
1039  * compute shared/private tiles for groups involving may writes.
1040  *
1041  * We only try to compute a shared memory tile if there is any reuse
1042  * or if the access is not coalesced.
1043  * Reuse and coalescing are checked within the given kernel.
1044  *
1045  * For computing a private memory tile, we also require that there is
1046  * some reuse.  Moreover, we require that the access is private
1047  * to the thread.  That is, we check that any given array element
1048  * is only accessed by a single thread.
1049  * We compute an access relation that maps the outer
1050  * data->thread_depth + data->n_thread schedule dimensions.
1051  * The latter data->n_thread will be mapped to thread identifiers.
1052  * We actually check that those iterators that will be wrapped
1053  * partition the array space.  This check is stricter than necessary
1054  * since several iterations may be mapped onto the same thread
1055  * and then they could be allowed to access the same memory elements,
1056  * but our check does not allow this situation.
1057  *
1058  * For private memory tiles, the number of schedule dimensions that
1059  * affect the offset is computed and stored in tile->depth, with
1060  * a lower bound of data->kernel_depth.  If this depth is smaller
1061  * than the minimal depth that still ensures that every element
1062  * is accessed by a single thread, then the depth is raised
1063  * to this minimal depth.
1064  * The fields of the tile are then adjusted to only refer to the tile->depth
1065  * outer schedule dimensions.
1066  *
1067  * We also check that the index expression only depends on parallel
1068  * loops.  That way, we can move those loops innermost and unroll them.
1069  * Again, we use a test that is stricter than necessary.
1070  * We actually check whether the index expression only depends
1071  * on the iterators that are wrapped over the threads.
1072  * These are necessarily parallel, but there may be more parallel loops.
1073  *
1074  * Combining the injectivity of the first test with the single-valuedness
1075  * of the second test, we simply test for bijectivity.
1076  *
1077  * If the use of the private tile requires unrolling, but some
1078  * of the other arrays are forcibly mapped to private memory,
1079  * then we do not allow the use of this private tile since
1080  * we cannot move the schedule dimensions that need to be unrolled down
1081  * without performing some kind of expansion on those arrays
1082  * that are forcibly mapped to private memory.
1083  *
1084  * If the array is marked force_private, then we bypass all checks
1085  * and assume we can (and should) use registers only.
1086  *
1087  * If it turns out we can (or have to) use registers, we compute
1088  * the private memory tile size using can_tile, after introducing a dependence
1089  * on the thread indices.
1090  */
compute_group_bounds_core(struct ppcg_kernel * kernel,struct gpu_array_ref_group * group,struct gpu_group_data * data)1091 static int compute_group_bounds_core(struct ppcg_kernel *kernel,
1092 	struct gpu_array_ref_group *group, struct gpu_group_data *data)
1093 {
1094 	isl_ctx *ctx = isl_space_get_ctx(group->array->space);
1095 	isl_union_map *access, *local;
1096 	int n_index = group->array->n_index;
1097 	int no_reuse, coalesced;
1098 	isl_map *acc;
1099 	int force_private = group->local_array->force_private;
1100 	int use_shared = !force_private && kernel->options->use_shared_memory &&
1101 				data->n_thread > 0;
1102 	int use_private = force_private || kernel->options->use_private_memory;
1103 	int r = 0;
1104 	int requires_unroll;
1105 	int unique_depth;
1106 
1107 	if (!use_shared && !use_private)
1108 		return 0;
1109 	if (gpu_array_is_read_only_scalar(group->array))
1110 		return 0;
1111 	if (!force_private && !group->exact_write)
1112 		return 0;
1113 	if (group->slice)
1114 		return 0;
1115 
1116 	access = gpu_array_ref_group_access_relation(group, 1, 1);
1117 	local = localize_access(data, isl_union_map_copy(access));
1118 	no_reuse = isl_union_map_is_injective(local);
1119 	if (no_reuse < 0)
1120 		r = -1;
1121 	if (use_shared && no_reuse)
1122 		coalesced = access_is_coalesced(data, local);
1123 	isl_union_map_free(local);
1124 
1125 	if (r >= 0 && kernel->options->debug->verbose &&
1126 	    use_shared && no_reuse && coalesced)
1127 		report_no_reuse_and_coalesced(kernel, access);
1128 
1129 	if (use_shared && (!no_reuse || !coalesced)) {
1130 		group->shared_tile = gpu_array_tile_create(ctx,
1131 							group->array->n_index);
1132 		acc = shared_access(group, access, data);
1133 		if (!group->shared_tile)
1134 			r = -1;
1135 		else if (!can_tile(acc, group->shared_tile))
1136 			group->shared_tile =
1137 					gpu_array_tile_free(group->shared_tile);
1138 		isl_map_free(acc);
1139 	}
1140 
1141 	if (r < 0 || (!force_private && (!use_private || no_reuse))) {
1142 		isl_union_map_free(access);
1143 		return r;
1144 	}
1145 
1146 	access = isl_union_map_apply_domain(access,
1147 					isl_union_map_copy(data->thread_sched));
1148 
1149 	acc = isl_map_from_union_map(access);
1150 
1151 	if (!force_private && !access_is_bijective(data, acc)) {
1152 		isl_map_free(acc);
1153 		return 0;
1154 	}
1155 
1156 	unique_depth = compute_accessed_by_single_thread_depth(data, acc);
1157 
1158 	acc = isl_map_intersect_domain(acc, isl_set_copy(data->privatization));
1159 	acc = isl_map_project_out(acc, isl_dim_in, data->thread_depth,
1160 								data->n_thread);
1161 	requires_unroll = check_requires_unroll(data, acc, force_private);
1162 	if (unique_depth < 0 || requires_unroll < 0 ||
1163 	    (requires_unroll && kernel->any_force_private)) {
1164 		isl_map_free(acc);
1165 		return requires_unroll < 0 ? -1 : 0;
1166 	}
1167 
1168 	group->private_tile = gpu_array_tile_create(ctx, n_index);
1169 	if (!group->private_tile) {
1170 		isl_map_free(acc);
1171 		return -1;
1172 	}
1173 	group->private_tile->requires_unroll = requires_unroll;
1174 	if (!can_tile(acc, group->private_tile))
1175 		group->private_tile = gpu_array_tile_free(group->private_tile);
1176 
1177 	isl_map_free(acc);
1178 
1179 	if (group->private_tile) {
1180 		struct gpu_array_tile *tile = group->private_tile;
1181 		int tile_depth = compute_tile_depth(data, tile);
1182 		if (tile_depth < unique_depth)
1183 			tile_depth = unique_depth;
1184 		if (tile_adjust_depth(tile, tile_depth) < 0)
1185 			return -1;
1186 	}
1187 
1188 	if (force_private && !group->private_tile)
1189 		isl_die(ctx, isl_error_internal,
1190 			"unable to map array reference group to registers",
1191 			return -1);
1192 
1193 	return 0;
1194 }
1195 
1196 /* Compute the private and/or shared memory tiles for the array
1197  * reference group "group" of array "array" and set the tile depth.
1198  * Return 0 on success and -1 on error.
1199  */
compute_group_bounds(struct ppcg_kernel * kernel,struct gpu_array_ref_group * group,struct gpu_group_data * data)1200 static int compute_group_bounds(struct ppcg_kernel *kernel,
1201 	struct gpu_array_ref_group *group, struct gpu_group_data *data)
1202 {
1203 	if (!group)
1204 		return -1;
1205 	if (compute_group_bounds_core(kernel, group, data) < 0)
1206 		return -1;
1207 	if (set_depth(data, group) < 0)
1208 		return -1;
1209 
1210 	return 0;
1211 }
1212 
1213 /* If two groups have overlapping access relations (as determined by
1214  * the "overlap" function) and if one of them involves a write,
1215  * then merge the two groups into one.
1216  * If "compute_bounds" is set, then call compute_group_bounds
1217  * on the merged groups.
1218  *
1219  * Return the updated number of groups.
1220  * Return -1 on error.
1221  */
group_writes(struct ppcg_kernel * kernel,int n,struct gpu_array_ref_group ** groups,int (* overlap)(struct gpu_array_ref_group * group1,struct gpu_array_ref_group * group2),int compute_bounds,struct gpu_group_data * data)1222 static int group_writes(struct ppcg_kernel *kernel,
1223 	int n, struct gpu_array_ref_group **groups,
1224 	int (*overlap)(struct gpu_array_ref_group *group1,
1225 		struct gpu_array_ref_group *group2), int compute_bounds,
1226 	struct gpu_group_data *data)
1227 {
1228 	int i, j;
1229 
1230 	for (i = 0; i < n; ++i) {
1231 		for (j = n - 1; j > i; --j) {
1232 			if (!groups[i]->write && !groups[j]->write)
1233 				continue;
1234 
1235 			if (!overlap(groups[i], groups[j]))
1236 				continue;
1237 
1238 			groups[i] = join_groups_and_free(groups[i], groups[j]);
1239 			if (j != n - 1)
1240 				groups[j] = groups[n - 1];
1241 			groups[n - 1] = NULL;
1242 			n--;
1243 
1244 			if (!groups[i])
1245 				return -1;
1246 			if (compute_bounds &&
1247 			    compute_group_bounds(kernel, groups[i], data) < 0)
1248 				return -1;
1249 		}
1250 	}
1251 
1252 	return n;
1253 }
1254 
1255 /* If two groups have overlapping access relations (within the innermost
1256  * loop) and if one of them involves a write, then merge the two groups
1257  * into one.
1258  *
1259  * Return the updated number of groups.
1260  */
group_overlapping_writes(struct ppcg_kernel * kernel,int n,struct gpu_array_ref_group ** groups,struct gpu_group_data * data)1261 static int group_overlapping_writes(struct ppcg_kernel *kernel,
1262 	int n, struct gpu_array_ref_group **groups,
1263 	struct gpu_group_data *data)
1264 {
1265 	return group_writes(kernel, n, groups, &accesses_overlap, 0, data);
1266 }
1267 
1268 /* Check if the access relations of group1 and group2 overlap within
1269  * the outermost min(group1->min_depth, group2->min_depth) loops.
1270  */
depth_accesses_overlap(struct gpu_array_ref_group * group1,struct gpu_array_ref_group * group2)1271 static int depth_accesses_overlap(struct gpu_array_ref_group *group1,
1272 	struct gpu_array_ref_group *group2)
1273 {
1274 	int depth;
1275 	int dim;
1276 	int empty;
1277 	isl_map *map_i, *map_j, *map;
1278 
1279 	depth = group1->min_depth;
1280 	if (group2->min_depth < depth)
1281 		depth = group2->min_depth;
1282 	map_i = isl_map_copy(group1->access);
1283 	dim = isl_map_dim(map_i, isl_dim_in);
1284 	map_i = isl_map_eliminate(map_i, isl_dim_in, depth, dim - depth);
1285 	map_j = isl_map_copy(group2->access);
1286 	map_j = isl_map_eliminate(map_j, isl_dim_in, depth, dim - depth);
1287 	map = isl_map_intersect(map_i, map_j);
1288 	empty = isl_map_is_empty(map);
1289 	isl_map_free(map);
1290 
1291 	return !empty;
1292 }
1293 
1294 /* If two groups have overlapping access relations (within the outer
1295  * depth loops) and if one of them involves a write,
1296  * then merge the two groups into one.
1297  *
1298  * Return the updated number of groups.
1299  */
group_depth_overlapping_writes(struct ppcg_kernel * kernel,int n,struct gpu_array_ref_group ** groups,struct gpu_group_data * data)1300 static int group_depth_overlapping_writes(struct ppcg_kernel *kernel,
1301 	int n, struct gpu_array_ref_group **groups, struct gpu_group_data *data)
1302 {
1303 	return group_writes(kernel, n, groups, &depth_accesses_overlap, 1,
1304 				data);
1305 }
1306 
1307 /* Is the size of the tile specified by "tile" smaller than the sum of
1308  * the sizes of the tiles specified by "tile1" and "tile2"?
1309  */
smaller_tile(struct gpu_array_tile * tile,struct gpu_array_tile * tile1,struct gpu_array_tile * tile2)1310 static int smaller_tile(struct gpu_array_tile *tile,
1311 	struct gpu_array_tile *tile1, struct gpu_array_tile *tile2)
1312 {
1313 	int smaller;
1314 	isl_val *size, *size1, *size2;
1315 
1316 	size = gpu_array_tile_size(tile);
1317 	size1 = gpu_array_tile_size(tile1);
1318 	size2 = gpu_array_tile_size(tile2);
1319 
1320 	size = isl_val_sub(size, size1);
1321 	size = isl_val_sub(size, size2);
1322 	smaller = isl_val_is_neg(size);
1323 
1324 	isl_val_free(size);
1325 
1326 	return smaller;
1327 }
1328 
1329 /* Given an initial grouping of array references and shared memory tiles
1330  * for each group that allows for a shared memory tile, merge two groups
1331  * if both have a shared memory tile, the merged group also has
1332  * a shared memory tile and the size of the tile for the merge group
1333  * is smaller than the sum of the tile sizes of the individual groups.
1334  *
1335  * If merging two groups decreases the depth of the tile of
1336  * one or both of the two groups, then we need to check for overlapping
1337  * writes again.
1338  *
1339  * Return the number of groups after merging.
1340  * Return -1 on error.
1341  */
group_common_shared_memory_tile(struct ppcg_kernel * kernel,struct gpu_array_info * array,int n,struct gpu_array_ref_group ** groups,struct gpu_group_data * data)1342 static int group_common_shared_memory_tile(struct ppcg_kernel *kernel,
1343 	struct gpu_array_info *array, int n,
1344 	struct gpu_array_ref_group **groups, struct gpu_group_data *data)
1345 {
1346 	int i, j;
1347 	int recompute_overlap = 0;
1348 
1349 	for (i = 0; i < n; ++i) {
1350 		if (!groups[i]->shared_tile)
1351 			continue;
1352 		for (j = n - 1; j > i; --j) {
1353 			struct gpu_array_ref_group *group;
1354 
1355 			if (!groups[j]->shared_tile)
1356 				continue;
1357 
1358 			if (!depth_accesses_overlap(groups[i], groups[j]))
1359 				continue;
1360 
1361 			group = join_groups(groups[i], groups[j]);
1362 			if (compute_group_bounds(kernel, group, data) < 0) {
1363 				gpu_array_ref_group_free(group);
1364 				return -1;
1365 			}
1366 			if (!group->shared_tile ||
1367 			    !smaller_tile(group->shared_tile,
1368 					groups[i]->shared_tile,
1369 					groups[j]->shared_tile)) {
1370 				gpu_array_ref_group_free(group);
1371 				continue;
1372 			}
1373 
1374 			if (group->min_depth < groups[i]->min_depth ||
1375 			    group->min_depth < groups[j]->min_depth)
1376 				recompute_overlap = 1;
1377 			gpu_array_ref_group_free(groups[i]);
1378 			gpu_array_ref_group_free(groups[j]);
1379 			groups[i] = group;
1380 			if (j != n - 1)
1381 				groups[j] = groups[n - 1];
1382 			n--;
1383 		}
1384 	}
1385 
1386 	if (recompute_overlap)
1387 		n = group_depth_overlapping_writes(kernel, n, groups, data);
1388 	return n;
1389 }
1390 
1391 /* Set array->n_group and array->groups to n and groups.
1392  *
1393  * Additionally, set the "nr" field of each group.
1394  */
set_array_groups(struct gpu_local_array_info * array,int n,struct gpu_array_ref_group ** groups)1395 static void set_array_groups(struct gpu_local_array_info *array,
1396 	int n, struct gpu_array_ref_group **groups)
1397 {
1398 	int i;
1399 
1400 	array->n_group = n;
1401 	array->groups = groups;
1402 
1403 	for (i = 0; i < n; ++i)
1404 		groups[i]->nr = i;
1405 }
1406 
1407 /* Combine all groups in "groups" into a single group and return
1408  * the new number of groups (1 or 0 if there were no groups to start with).
1409  */
join_all_groups(int n,struct gpu_array_ref_group ** groups)1410 static int join_all_groups(int n, struct gpu_array_ref_group **groups)
1411 {
1412 	int i;
1413 
1414 	for (i = n - 1; i > 0; --i) {
1415 		groups[0] = join_groups_and_free(groups[0], groups[i]);
1416 		groups[i] = NULL;
1417 		n--;
1418 	}
1419 
1420 	return n;
1421 }
1422 
1423 /* Group array references that should be considered together when
1424  * deciding whether to access them from private, shared or global memory.
1425  * Return -1 on error.
1426  *
1427  * In particular, if two array references overlap and if one of them
1428  * is a write, then the two references are grouped together.
1429  * We first perform an initial grouping based only on the access relation.
1430  * After computing shared and private memory tiles, we check for
1431  * overlapping writes again, but this time taking into account
1432  * the depth of the effective tile.
1433  *
1434  * Furthermore, if two groups admit a shared memory tile and if the
1435  * combination of the two also admits a shared memory tile, we merge
1436  * the two groups.
1437  *
1438  * If the array contains structures, then we compute a single
1439  * reference group without trying to find any tiles
1440  * since we do not map such arrays to private or shared
1441  * memory.  The only exception is when those arrays of structures
1442  * are required to be mapped to private memory.
1443  */
group_array_references(struct ppcg_kernel * kernel,struct gpu_local_array_info * local,struct gpu_group_data * data)1444 static int group_array_references(struct ppcg_kernel *kernel,
1445 	struct gpu_local_array_info *local, struct gpu_group_data *data)
1446 {
1447 	int i;
1448 	int n;
1449 	isl_ctx *ctx = isl_union_map_get_ctx(data->shared_sched);
1450 	struct gpu_array_ref_group **groups;
1451 
1452 	groups = isl_calloc_array(ctx, struct gpu_array_ref_group *,
1453 					local->array->n_ref);
1454 	if (!groups)
1455 		return -1;
1456 
1457 	n = populate_array_references(local, groups, data);
1458 
1459 	if (local->array->has_compound_element && !local->force_private) {
1460 		n = join_all_groups(n, groups);
1461 		set_array_groups(local, n, groups);
1462 		return 0;
1463 	}
1464 
1465 	n = group_overlapping_writes(kernel, n, groups, data);
1466 
1467 	for (i = 0; i < n; ++i)
1468 		if (compute_group_bounds(kernel, groups[i], data) < 0)
1469 			n = -1;
1470 
1471 	n = group_depth_overlapping_writes(kernel, n, groups, data);
1472 
1473 	n = group_common_shared_memory_tile(kernel, local->array,
1474 					    n, groups, data);
1475 
1476 	set_array_groups(local, n, groups);
1477 
1478 	if (n >= 0)
1479 		return 0;
1480 
1481 	for (i = 0; i < local->array->n_ref; ++i)
1482 		gpu_array_ref_group_free(groups[i]);
1483 	return -1;
1484 }
1485 
1486 /* For each array in the input program that can be mapped to private memory,
1487  * check if there are any order dependences active inside the current kernel,
1488  * within the same iteration of the host schedule, i.e., the prefix
1489  * schedule at "node".
1490  * If so, mark the array as force_private so that its reference groups will be
1491  * mapped to a registers.
1492  *
1493  * Note that the arrays that cannot be mapped to private memory have
1494  * had their order dependences added to prog->array_order and
1495  * subsequently to the coincidence constraints.
1496  */
check_can_be_private_live_ranges(struct ppcg_kernel * kernel,__isl_keep isl_schedule_node * node)1497 static void check_can_be_private_live_ranges(struct ppcg_kernel *kernel,
1498 	__isl_keep isl_schedule_node *node)
1499 {
1500 	int i;
1501 	isl_union_set *domain;
1502 	isl_multi_union_pw_aff *prefix;
1503 	isl_union_pw_multi_aff *contraction;
1504 
1505 	if (!kernel->options->live_range_reordering)
1506 		return;
1507 
1508 	kernel->any_force_private = 0;
1509 
1510 	prefix = isl_schedule_node_get_prefix_schedule_multi_union_pw_aff(node);
1511 	contraction = isl_union_pw_multi_aff_copy(kernel->contraction);
1512 	prefix = isl_multi_union_pw_aff_pullback_union_pw_multi_aff(prefix,
1513 								contraction);
1514 	domain = isl_union_set_copy(kernel->expanded_domain);
1515 	domain = isl_union_set_universe(domain);
1516 
1517 	for (i = 0; i < kernel->n_array; ++i) {
1518 		struct gpu_local_array_info *local = &kernel->array[i];
1519 		isl_union_map *order;
1520 
1521 		local->force_private = 0;
1522 		if (!gpu_array_can_be_private(local->array))
1523 			continue;
1524 		order = isl_union_map_copy(local->array->dep_order);
1525 		order = isl_union_map_intersect_domain(order,
1526 						    isl_union_set_copy(domain));
1527 		order = isl_union_map_intersect_range(order,
1528 						    isl_union_set_copy(domain));
1529 		order = isl_union_map_eq_at_multi_union_pw_aff(order,
1530 					isl_multi_union_pw_aff_copy(prefix));
1531 		if (!isl_union_map_is_empty(order)) {
1532 			local->force_private = 1;
1533 			kernel->any_force_private = 1;
1534 		}
1535 		isl_union_map_free(order);
1536 	}
1537 
1538 	isl_multi_union_pw_aff_free(prefix);
1539 	isl_union_set_free(domain);
1540 }
1541 
1542 /* Expand the domain of the schedule "s" by plugging in
1543  * the contraction "contraction" and return the result.
1544  */
expand(__isl_take isl_union_map * s,__isl_keep isl_union_pw_multi_aff * contraction)1545 static __isl_give isl_union_map *expand(__isl_take isl_union_map *s,
1546 	__isl_keep isl_union_pw_multi_aff *contraction)
1547 {
1548 	contraction = isl_union_pw_multi_aff_copy(contraction);
1549 	s = isl_union_map_preimage_domain_union_pw_multi_aff(s, contraction);
1550 	return s;
1551 }
1552 
1553 /* Create a set of dimension data->thread_depth + data->n_thread
1554  * that equates the residue of the final data->n_thread dimensions
1555  * modulo the kernel->block_dim sizes to the thread identifiers.
1556  * Store the computed set in data->privatization.
1557  *
1558  * The construction starts with the space of kernel->thread_filter,
1559  * which is known to reference all thread identifiers.
1560  */
compute_privatization(struct gpu_group_data * data,struct ppcg_kernel * kernel)1561 static void compute_privatization(struct gpu_group_data *data,
1562 	struct ppcg_kernel *kernel)
1563 {
1564 	int i;
1565 	isl_ctx *ctx;
1566 	isl_space *space;
1567 	isl_local_space *ls;
1568 	isl_set *set;
1569 
1570 	ctx = isl_union_map_get_ctx(data->shared_sched);
1571 	space = isl_union_set_get_space(kernel->thread_filter);
1572 	space = isl_space_set_from_params(space);
1573 	space = isl_space_add_dims(space, isl_dim_set,
1574 				    data->thread_depth + data->n_thread);
1575 	set = isl_set_universe(space);
1576 	space = isl_set_get_space(set);
1577 	ls = isl_local_space_from_space(space);
1578 
1579 	for (i = 0; i < data->n_thread; ++i) {
1580 		isl_aff *aff, *aff2;
1581 		isl_constraint *c;
1582 		isl_val *v;
1583 		isl_id *id;
1584 		int pos;
1585 
1586 		aff = isl_aff_var_on_domain(isl_local_space_copy(ls),
1587 					isl_dim_set, data->thread_depth + i);
1588 		v = isl_val_int_from_si(ctx, kernel->block_dim[i]);
1589 		aff = isl_aff_mod_val(aff, v);
1590 		id = isl_id_list_get_id(kernel->thread_ids, i);
1591 		pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
1592 		isl_id_free(id);
1593 		aff2 = isl_aff_var_on_domain(isl_local_space_copy(ls),
1594 					isl_dim_param, pos);
1595 		aff = isl_aff_sub(aff, aff2);
1596 		c = isl_equality_from_aff(aff);
1597 		set = isl_set_add_constraint(set, c);
1598 	}
1599 
1600 	isl_local_space_free(ls);
1601 	data->privatization = set;
1602 }
1603 
1604 /* Return the prefix schedule at "node" as a relation
1605  * between domain elements and schedule dimensions after detecting
1606  * equalities in this relation.
1607  */
prefix_with_equalities(__isl_keep isl_schedule_node * node)1608 static __isl_give isl_union_map *prefix_with_equalities(
1609 	__isl_keep isl_schedule_node *node)
1610 {
1611 	isl_union_map *schedule;
1612 
1613 	schedule = isl_schedule_node_get_prefix_schedule_relation(node);
1614 	schedule = isl_union_map_detect_equalities(schedule);
1615 
1616 	return schedule;
1617 }
1618 
1619 /* Group references of all arrays in "kernel".
1620  * "node" points to the kernel mark.
1621  * The mapping to shared memory in computed at the "shared" mark.
1622  *
1623  * We first extract all required schedule information into
1624  * a gpu_group_data structure and then consider each array
1625  * in turn.
1626  */
gpu_group_references(struct ppcg_kernel * kernel,__isl_keep isl_schedule_node * node)1627 int gpu_group_references(struct ppcg_kernel *kernel,
1628 	__isl_keep isl_schedule_node *node)
1629 {
1630 	int i;
1631 	int r = 0;
1632 	isl_union_pw_multi_aff *contraction;
1633 	struct gpu_group_data data;
1634 
1635 	check_can_be_private_live_ranges(kernel, node);
1636 
1637 	data.scop = kernel->prog->scop;
1638 
1639 	data.kernel_depth = isl_schedule_node_get_schedule_depth(node);
1640 	data.host_sched = isl_schedule_node_get_prefix_schedule_relation(node);
1641 
1642 	node = isl_schedule_node_copy(node);
1643 	node = gpu_tree_move_down_to_shared(node, kernel->core);
1644 	data.shared_depth = isl_schedule_node_get_schedule_depth(node);
1645 	data.shared_sched = prefix_with_equalities(node);
1646 
1647 	node = gpu_tree_move_down_to_thread(node, kernel->core);
1648 	node = isl_schedule_node_child(node, 0);
1649 	data.thread_depth = isl_schedule_node_get_schedule_depth(node);
1650 	data.n_thread = isl_schedule_node_band_n_member(node);
1651 	if (data.thread_depth == data.shared_depth)
1652 		data.copy_sched = isl_union_map_copy(data.shared_sched);
1653 	else
1654 		data.copy_sched = prefix_with_equalities(node);
1655 	data.thread_sched = isl_union_map_copy(data.copy_sched);
1656 	data.thread_sched = isl_union_map_flat_range_product(data.thread_sched,
1657 		isl_schedule_node_band_get_partial_schedule_union_map(node));
1658 	data.thread_sched = isl_union_map_detect_equalities(data.thread_sched);
1659 
1660 	contraction = isl_union_pw_multi_aff_copy(kernel->contraction);
1661 	data.host_sched = expand(data.host_sched, contraction);
1662 	data.shared_sched = expand(data.shared_sched, contraction);
1663 	if (data.thread_depth == data.shared_depth) {
1664 		isl_union_map_free(data.copy_sched);
1665 		data.copy_sched = isl_union_map_copy(data.shared_sched);
1666 	} else {
1667 		data.copy_sched = expand(data.copy_sched, contraction);
1668 	}
1669 	data.thread_sched = expand(data.thread_sched, contraction);
1670 	isl_union_pw_multi_aff_free(contraction);
1671 
1672 	node = isl_schedule_node_child(node, 0);
1673 	data.full_sched = isl_union_map_copy(data.thread_sched);
1674 	data.full_sched = isl_union_map_flat_range_product(data.full_sched,
1675 		isl_schedule_node_get_subtree_schedule_union_map(node));
1676 	isl_schedule_node_free(node);
1677 
1678 	compute_privatization(&data, kernel);
1679 
1680 	for (i = 0; i < kernel->n_array; ++i) {
1681 		r = group_array_references(kernel, &kernel->array[i], &data);
1682 		if (r < 0)
1683 			break;
1684 	}
1685 
1686 	isl_union_map_free(data.host_sched);
1687 	isl_union_map_free(data.shared_sched);
1688 	isl_union_map_free(data.copy_sched);
1689 	isl_union_map_free(data.thread_sched);
1690 	isl_union_map_free(data.full_sched);
1691 	isl_set_free(data.privatization);
1692 
1693 	return r;
1694 }
1695 
1696 /* Given a description of an array tile "tile" and the "space"
1697  *
1698  *	{ D -> A }
1699  *
1700  * where D represents the first tile->depth schedule dimensions
1701  * and A represents the array, construct an isl_multi_aff
1702  *
1703  *	{ [D[i] -> A[a]] -> A'[a'] }
1704  *
1705  * with A' a scaled down copy of A according to the shifts and strides
1706  * in "tile".  In particular,
1707  *
1708  *	a' = (a + shift(i))/stride
1709  *
1710  * "insert_array" represents
1711  *
1712  *	{ [D -> A] -> D }
1713  *
1714  * and is used to insert A into the domain of functions that only
1715  * reference D.
1716  */
strided_tile(struct gpu_array_tile * tile,__isl_keep isl_space * space,__isl_keep isl_multi_aff * insert_array)1717 static __isl_give isl_multi_aff *strided_tile(
1718 	struct gpu_array_tile *tile, __isl_keep isl_space *space,
1719 	__isl_keep isl_multi_aff *insert_array)
1720 {
1721 	int i;
1722 	isl_ctx *ctx;
1723 	isl_multi_aff *shift;
1724 	isl_multi_val *stride;
1725 	isl_space *space2;
1726 	isl_local_space *ls;
1727 	isl_multi_aff *tiling;
1728 
1729 	ctx = isl_space_get_ctx(space);
1730 	space2 = isl_space_domain(isl_space_copy(space));
1731 	ls = isl_local_space_from_space(space2);
1732 	space2 = isl_space_range(isl_space_copy(space));
1733 	stride = isl_multi_val_zero(space2);
1734 	shift = isl_multi_aff_zero(isl_space_copy(space));
1735 
1736 	for (i = 0; i < tile->n; ++i) {
1737 		struct gpu_array_bound *bound = &tile->bound[i];
1738 		isl_val *stride_i;
1739 		isl_aff *shift_i;
1740 
1741 		if (tile->bound[i].shift) {
1742 			stride_i = isl_val_copy(bound->stride);
1743 			shift_i = isl_aff_copy(bound->shift);
1744 		} else {
1745 			stride_i = isl_val_one(ctx);
1746 			shift_i = isl_aff_zero_on_domain(
1747 					isl_local_space_copy(ls));
1748 		}
1749 
1750 		stride = isl_multi_val_set_val(stride, i, stride_i);
1751 		shift = isl_multi_aff_set_aff(shift, i, shift_i);
1752 	}
1753 	isl_local_space_free(ls);
1754 
1755 	shift = isl_multi_aff_pullback_multi_aff(shift,
1756 				    isl_multi_aff_copy(insert_array));
1757 
1758 	tiling = isl_multi_aff_range_map(isl_space_copy(space));
1759 	tiling = isl_multi_aff_add(tiling, shift);
1760 	tiling = isl_multi_aff_scale_down_multi_val(tiling, stride);
1761 
1762 	return tiling;
1763 }
1764 
1765 /* Compute a tiling for the array reference group "group".
1766  *
1767  * The tiling is of the form
1768  *
1769  *	{ [D[i] -> A[a]] -> T[t] }
1770  *
1771  * where D represents the first tile->depth schedule dimensions,
1772  * A represents the global array and T represents the shared or
1773  * private memory tile.  The name of T is the name of the local
1774  * array.
1775  *
1776  * If there is any stride in the accesses, then the mapping is
1777  *
1778  *	t = (a + shift(i))/stride - lb(i)
1779  *
1780  * otherwise, it is simply
1781  *
1782  *	t = a - lb(i)
1783  */
gpu_array_ref_group_compute_tiling(struct gpu_array_ref_group * group)1784 void gpu_array_ref_group_compute_tiling(struct gpu_array_ref_group *group)
1785 {
1786 	int i;
1787 	struct gpu_array_tile *tile;
1788 	isl_space *space;
1789 	isl_multi_aff *tiling, *lb, *insert_array;
1790 	isl_printer *p;
1791 	char *local_name;
1792 
1793 	tile = gpu_array_ref_group_tile(group);
1794 	if (!tile)
1795 		return;
1796 
1797 	space = isl_map_get_space(group->access);
1798 	space = isl_space_from_range(isl_space_range(space));
1799 	space = isl_space_add_dims(space, isl_dim_in, tile->depth);
1800 	insert_array = isl_multi_aff_domain_map(isl_space_copy(space));
1801 
1802 	for (i = 0; i < tile->n; ++i)
1803 		if (tile->bound[i].shift)
1804 			break;
1805 
1806 	if (i < tile->n)
1807 		tiling = strided_tile(tile, space, insert_array);
1808 	else
1809 		tiling = isl_multi_aff_range_map(isl_space_copy(space));
1810 
1811 	lb = isl_multi_aff_zero(space);
1812 	for (i = 0; i < tile->n; ++i) {
1813 		isl_aff *lb_i = isl_aff_copy(tile->bound[i].lb);
1814 		lb = isl_multi_aff_set_aff(lb, i, lb_i);
1815 	}
1816 	lb = isl_multi_aff_pullback_multi_aff(lb, insert_array);
1817 
1818 	tiling = isl_multi_aff_sub(tiling, lb);
1819 
1820 	p = isl_printer_to_str(isl_multi_aff_get_ctx(tiling));
1821 	p = gpu_array_ref_group_print_name(group, p);
1822 	local_name = isl_printer_get_str(p);
1823 	isl_printer_free(p);
1824 	tiling = isl_multi_aff_set_tuple_name(tiling, isl_dim_out, local_name);
1825 	free(local_name);
1826 
1827 	tile->tiling = tiling;
1828 }
1829