File: | tools/polly/lib/External/ppcg/gpu.c |
Warning: | line 1178, column 19 Access to field 'n_group' results in a dereference of a null pointer (loaded from variable 'array') |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | /* | |||
2 | * Copyright 2010-2011 INRIA Saclay | |||
3 | * Copyright 2012-2013 Ecole Normale Superieure | |||
4 | * Copyright 2015-2016 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 <assert.h> | |||
15 | #include <stdlib.h> | |||
16 | #include <string.h> | |||
17 | ||||
18 | #include <isl/polynomial.h> | |||
19 | #include <isl/union_set.h> | |||
20 | #include <isl/aff.h> | |||
21 | #include <isl/ilp.h> | |||
22 | #include <isl/flow.h> | |||
23 | #include <isl/schedule.h> | |||
24 | #include <isl/schedule_node.h> | |||
25 | #include <isl/options.h> | |||
26 | #include <isl/ast_build.h> | |||
27 | ||||
28 | #include "cpu.h" | |||
29 | #include "gpu.h" | |||
30 | #include "gpu_array_tile.h" | |||
31 | #include "gpu_group.h" | |||
32 | #include "gpu_hybrid.h" | |||
33 | #include "gpu_tree.h" | |||
34 | #include "hybrid.h" | |||
35 | #include "schedule.h" | |||
36 | #include "ppcg_options.h" | |||
37 | #include "print.h" | |||
38 | #include "util.h" | |||
39 | ||||
40 | struct gpu_array_info; | |||
41 | ||||
42 | /* Return the name of the outer array (of structs) accessed by "access". | |||
43 | */ | |||
44 | static const char *get_outer_array_name(__isl_keep isl_map *access) | |||
45 | { | |||
46 | isl_space *space; | |||
47 | const char *name; | |||
48 | ||||
49 | space = isl_space_range(isl_map_get_space(access)); | |||
50 | while (space && isl_space_is_wrapping(space)) | |||
51 | space = isl_space_domain(isl_space_unwrap(space)); | |||
52 | name = isl_space_get_tuple_name(space, isl_dim_set); | |||
53 | isl_space_free(space); | |||
54 | ||||
55 | return name; | |||
56 | } | |||
57 | ||||
58 | /* Collect all references to the given array and store pointers to them | |||
59 | * in array->refs. | |||
60 | */ | |||
61 | void collect_references(struct gpu_prog *prog, | |||
62 | struct gpu_array_info *array) | |||
63 | { | |||
64 | int i; | |||
65 | int n; | |||
66 | ||||
67 | n = 0; | |||
68 | for (i = 0; i < prog->n_stmts; ++i) { | |||
69 | struct gpu_stmt *stmt = &prog->stmts[i]; | |||
70 | struct gpu_stmt_access *access; | |||
71 | ||||
72 | for (access = stmt->accesses; access; access = access->next) { | |||
73 | const char *name; | |||
74 | name = get_outer_array_name(access->access); | |||
75 | if (name && !strcmp(array->name, name)) | |||
76 | n++; | |||
77 | } | |||
78 | } | |||
79 | ||||
80 | array->n_ref = n; | |||
81 | array->refs = isl_alloc_array(prog->ctx, struct gpu_stmt_access *, n)((struct gpu_stmt_access * *)isl_malloc_or_die(prog->ctx, ( n)*sizeof(struct gpu_stmt_access *))); | |||
82 | assert(array->refs)((void) sizeof ((array->refs) ? 1 : 0), __extension__ ({ if (array->refs) ; else __assert_fail ("array->refs", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 82, __extension__ __PRETTY_FUNCTION__); })); | |||
83 | ||||
84 | n = 0; | |||
85 | for (i = 0; i < prog->n_stmts; ++i) { | |||
86 | struct gpu_stmt *stmt = &prog->stmts[i]; | |||
87 | struct gpu_stmt_access *access; | |||
88 | ||||
89 | for (access = stmt->accesses; access; access = access->next) { | |||
90 | const char *name; | |||
91 | name = get_outer_array_name(access->access); | |||
92 | if (!name || strcmp(array->name, name)) | |||
93 | continue; | |||
94 | ||||
95 | array->refs[n++] = access; | |||
96 | } | |||
97 | } | |||
98 | } | |||
99 | ||||
100 | /* Compute and return the extent of "array", taking into account the set of | |||
101 | * accessed elements. | |||
102 | * | |||
103 | * In particular, the extent in the outer dimension is taken | |||
104 | * from "accessed", while the extents in the remaining dimensions | |||
105 | * are taken from array->extent. | |||
106 | * | |||
107 | * The extent in the outer dimension cannot be taken from array->extent | |||
108 | * because that may be unbounded. Furthermore, even if it is bounded, | |||
109 | * it may be larger than the piece of the array that is being accessed. | |||
110 | */ | |||
111 | static __isl_give isl_set *compute_extent(struct pet_array *array, | |||
112 | __isl_keep isl_set *accessed) | |||
113 | { | |||
114 | int n_index; | |||
115 | isl_id *id; | |||
116 | isl_set *outer; | |||
117 | isl_set *extent; | |||
118 | ||||
119 | extent = isl_set_copy(array->extent); | |||
120 | ||||
121 | n_index = isl_set_dim(accessed, isl_dim_set); | |||
122 | if (n_index == 0) | |||
123 | return extent; | |||
124 | ||||
125 | extent = isl_set_project_out(extent, isl_dim_set, 0, 1); | |||
126 | outer = isl_set_copy(accessed); | |||
127 | outer = isl_set_project_out(outer, isl_dim_set, 1, n_index - 1); | |||
128 | extent = isl_set_flat_product(outer, extent); | |||
129 | id = isl_set_get_tuple_id(accessed); | |||
130 | extent = isl_set_set_tuple_id(extent, id); | |||
131 | ||||
132 | return extent; | |||
133 | } | |||
134 | ||||
135 | /* Is the array "array" being extracted a read-only scalar? | |||
136 | * | |||
137 | * That is, is "array" a scalar that is never possibly written to. | |||
138 | * An array containing structures is never considered to be a scalar. | |||
139 | */ | |||
140 | static int is_read_only_scalar(struct gpu_array_info *array, | |||
141 | struct gpu_prog *prog) | |||
142 | { | |||
143 | isl_set *space; | |||
144 | isl_union_map *write; | |||
145 | int empty; | |||
146 | ||||
147 | if (array->has_compound_element) | |||
148 | return 0; | |||
149 | if (array->n_index != 0) | |||
150 | return 0; | |||
151 | ||||
152 | write = isl_union_map_copy(prog->may_write); | |||
153 | space = isl_set_universe(isl_space_copy(array->space)); | |||
154 | write = isl_union_map_intersect_range(write, | |||
155 | isl_union_set_from_set(space)); | |||
156 | empty = isl_union_map_is_empty(write); | |||
157 | isl_union_map_free(write); | |||
158 | ||||
159 | return empty; | |||
160 | } | |||
161 | ||||
162 | /* Is "array" only accessed as individual, fixed elements? | |||
163 | * That is, does each access to "array" access a single, fixed element? | |||
164 | */ | |||
165 | isl_bool only_fixed_element_accessed(struct gpu_array_info *array) | |||
166 | { | |||
167 | int i; | |||
168 | ||||
169 | for (i = 0; i < array->n_ref; ++i) | |||
170 | if (!array->refs[i]->fixed_element) | |||
171 | return isl_bool_false; | |||
172 | ||||
173 | return isl_bool_true; | |||
174 | } | |||
175 | ||||
176 | /* Compute bounds on the host array "pa" based on the corresponding | |||
177 | * accessed elements in "arrays" | |||
178 | * and collect all references to the array. | |||
179 | * Store the results in "info". | |||
180 | * | |||
181 | * If the array is zero-dimensional and does not contain structures, | |||
182 | * i.e., if the array is a scalar, we check whether it is read-only. | |||
183 | * We also check whether the array is accessed at all. | |||
184 | */ | |||
185 | static int extract_array_info(struct gpu_prog *prog, | |||
186 | struct gpu_array_info *info, struct pet_array *pa, | |||
187 | __isl_keep isl_union_set *arrays) | |||
188 | { | |||
189 | int empty; | |||
190 | const char *name; | |||
191 | int n_index; | |||
192 | isl_multi_pw_aff *bounds; | |||
193 | isl_set *accessed, *extent; | |||
194 | ||||
195 | n_index = isl_set_dim(pa->extent, isl_dim_set); | |||
196 | name = isl_set_get_tuple_name(pa->extent); | |||
197 | ||||
198 | info->space = isl_set_get_space(pa->extent); | |||
199 | info->name = strdup(name); | |||
200 | info->n_index = n_index; | |||
201 | info->linearize = prog->scop->options->linearize_device_arrays; | |||
202 | ||||
203 | info->type = strdup(pa->element_type); | |||
204 | info->size = pa->element_size; | |||
205 | info->local = pa->declared && !pa->exposed; | |||
206 | info->has_compound_element = pa->element_is_record; | |||
207 | info->read_only_scalar = is_read_only_scalar(info, prog); | |||
208 | ||||
209 | info->declared_extent = isl_set_copy(pa->extent); | |||
210 | accessed = isl_union_set_extract_set(arrays, | |||
211 | isl_space_copy(info->space)); | |||
212 | empty = isl_set_is_empty(accessed); | |||
213 | extent = compute_extent(pa, accessed); | |||
214 | isl_set_free(accessed); | |||
215 | info->extent = extent; | |||
216 | if (empty < 0) | |||
217 | return -1; | |||
218 | info->accessed = !empty; | |||
219 | bounds = ppcg_size_from_extent(isl_set_copy(extent)); | |||
220 | bounds = isl_multi_pw_aff_gist(bounds, isl_set_copy(prog->context)); | |||
221 | if (!bounds) | |||
222 | return -1; | |||
223 | if (!isl_multi_pw_aff_is_cst(bounds)) | |||
224 | info->linearize = 1; | |||
225 | info->bound = bounds; | |||
226 | ||||
227 | collect_references(prog, info); | |||
228 | info->only_fixed_element = only_fixed_element_accessed(info); | |||
229 | ||||
230 | return 0; | |||
231 | } | |||
232 | ||||
233 | /* Remove independence from the order constraints "order" on array "array". | |||
234 | * Since the pairs of iterations in the filter relation of an independence | |||
235 | * are guaranteed to be completely independent by the user, there is | |||
236 | * no need to ensure that live ranges are ordered along those pairs. | |||
237 | * We make an exception for local variables, though, as the independence | |||
238 | * guarantee does not apply to those. | |||
239 | * | |||
240 | * The order constraints are used in two places. | |||
241 | * Those on scalars are used in check_scalar_live_ranges to check if | |||
242 | * we need to force the scalar to be private. Any non-local scalar | |||
243 | * should not be forced scalar if it only appears in independent loops. | |||
244 | * Those on non-scalars are added to the coincidence constraints | |||
245 | * in compute_schedule because we do not support any array expansion. | |||
246 | * Accesses to non-local arrays should not prevent a loop from being | |||
247 | * considered coincident so we should indeed remove those constraints | |||
248 | * from the order constraints. | |||
249 | */ | |||
250 | static __isl_give isl_union_map *remove_independences(struct gpu_prog *prog, | |||
251 | struct gpu_array_info *array, __isl_take isl_union_map *order) | |||
252 | { | |||
253 | // We do not have independence information in Polly. Hence, make this | |||
254 | // function a no-op. | |||
255 | return order; | |||
256 | int i; | |||
257 | ||||
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)) | |||
261 | continue; | |||
262 | ||||
263 | order = isl_union_map_subtract(order, | |||
264 | isl_union_map_copy(pi->filter)); | |||
265 | } | |||
266 | ||||
267 | return order; | |||
268 | } | |||
269 | ||||
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. | |||
279 | */ | |||
280 | void collect_order_dependences(struct gpu_prog *prog) | |||
281 | { | |||
282 | int i; | |||
283 | isl_space *space; | |||
284 | isl_union_map *accesses; | |||
285 | ||||
286 | space = isl_union_map_get_space(prog->read); | |||
287 | prog->array_order = isl_union_map_empty(space); | |||
288 | ||||
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)); | |||
295 | ||||
296 | for (i = 0; i < prog->n_array; ++i) { | |||
297 | struct gpu_array_info *array = &prog->array[i]; | |||
298 | isl_set *set; | |||
299 | isl_union_set *uset; | |||
300 | isl_union_map *order; | |||
301 | ||||
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), | |||
306 | uset)); | |||
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; | |||
313 | ||||
314 | if (gpu_array_can_be_private(array)) | |||
315 | continue; | |||
316 | ||||
317 | prog->array_order = isl_union_map_union(prog->array_order, | |||
318 | isl_union_map_copy(array->dep_order)); | |||
319 | } | |||
320 | ||||
321 | isl_union_map_free(accesses); | |||
322 | } | |||
323 | ||||
324 | /* Construct a gpu_array_info for each array referenced by prog->scop and | |||
325 | * collect them in prog->array. | |||
326 | * | |||
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. | |||
332 | * | |||
333 | * If we are allowing live range reordering, then also set | |||
334 | * the dep_order field. Otherwise leave it NULL. | |||
335 | */ | |||
336 | static int collect_array_info(struct gpu_prog *prog) | |||
337 | { | |||
338 | int i; | |||
339 | int r = 0; | |||
340 | isl_union_set *arrays; | |||
341 | ||||
342 | arrays = isl_union_map_range(isl_union_map_copy(prog->read)); | |||
343 | arrays = isl_union_set_union(arrays, | |||
344 | isl_union_map_range(isl_union_map_copy(prog->may_write))); | |||
345 | ||||
346 | arrays = isl_union_set_apply(arrays, | |||
347 | isl_union_map_copy(prog->to_outer)); | |||
348 | ||||
349 | arrays = isl_union_set_coalesce(arrays); | |||
350 | ||||
351 | prog->n_array = prog->scop->pet->n_array; | |||
352 | prog->array = isl_calloc_array(prog->ctx,((struct gpu_array_info *)isl_calloc_or_die(prog->ctx, prog ->n_array, sizeof(struct gpu_array_info))) | |||
353 | struct gpu_array_info, prog->n_array)((struct gpu_array_info *)isl_calloc_or_die(prog->ctx, prog ->n_array, sizeof(struct gpu_array_info))); | |||
354 | assert(prog->array)((void) sizeof ((prog->array) ? 1 : 0), __extension__ ({ if (prog->array) ; else __assert_fail ("prog->array", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 354, __extension__ __PRETTY_FUNCTION__); })); | |||
355 | prog->n_array = 0; | |||
356 | for (i = 0; i < prog->scop->pet->n_array; ++i) { | |||
357 | isl_bool field; | |||
358 | ||||
359 | field = isl_set_is_wrapping(prog->scop->pet->arrays[i]->extent); | |||
360 | if (field < 0) | |||
361 | break; | |||
362 | if (field) | |||
363 | continue; | |||
364 | if (extract_array_info(prog, &prog->array[prog->n_array++], | |||
365 | prog->scop->pet->arrays[i], arrays) < 0) | |||
366 | r = -1; | |||
367 | } | |||
368 | if (i < prog->scop->pet->n_array) | |||
369 | r = -1; | |||
370 | ||||
371 | isl_union_set_free(arrays); | |||
372 | ||||
373 | if (prog->scop->options->live_range_reordering) | |||
374 | collect_order_dependences(prog); | |||
375 | ||||
376 | return r; | |||
377 | } | |||
378 | ||||
379 | static void free_array_info(struct gpu_prog *prog) | |||
380 | { | |||
381 | int i; | |||
382 | ||||
383 | for (i = 0; i < prog->n_array; ++i) { | |||
384 | free(prog->array[i].type); | |||
385 | free(prog->array[i].name); | |||
386 | isl_multi_pw_aff_free(prog->array[i].bound); | |||
387 | isl_ast_expr_free(prog->array[i].bound_expr); | |||
388 | isl_space_free(prog->array[i].space); | |||
389 | isl_set_free(prog->array[i].declared_extent); | |||
390 | isl_set_free(prog->array[i].extent); | |||
391 | isl_ast_expr_free(prog->array[i].declared_size); | |||
392 | free(prog->array[i].refs); | |||
393 | isl_union_map_free(prog->array[i].dep_order); | |||
394 | } | |||
395 | free(prog->array); | |||
396 | } | |||
397 | ||||
398 | /* Check if a gpu array is a scalar. A scalar is a value that is not stored | |||
399 | * as an array or through a pointer reference, but as a single data element. | |||
400 | * At the moment, scalars are represented as zero-dimensional arrays. | |||
401 | * Note that the single data element may be an entire structure. | |||
402 | */ | |||
403 | int gpu_array_is_scalar(struct gpu_array_info *array) | |||
404 | { | |||
405 | return array->n_index == 0; | |||
406 | } | |||
407 | ||||
408 | /* Can "array" be mapped to private memory? | |||
409 | * That is, is it only accessed as individual elements with | |||
410 | * constant index expressions? | |||
411 | */ | |||
412 | isl_bool gpu_array_can_be_private(struct gpu_array_info *array) | |||
413 | { | |||
414 | if (!array) | |||
415 | return isl_bool_error; | |||
416 | return array->only_fixed_element; | |||
417 | } | |||
418 | ||||
419 | /* Is "array" a read-only scalar? | |||
420 | */ | |||
421 | int gpu_array_is_read_only_scalar(struct gpu_array_info *array) | |||
422 | { | |||
423 | return array->read_only_scalar; | |||
424 | } | |||
425 | ||||
426 | /* Does "array" need to be allocated on the device? | |||
427 | * If it is a read-only scalar, then it will be passed as an argument | |||
428 | * to the kernel and therefore does not require any allocation. | |||
429 | * If this device memory is not accessed at all, then it does not | |||
430 | * need to be allocated either. | |||
431 | */ | |||
432 | int gpu_array_requires_device_allocation(struct gpu_array_info *array) | |||
433 | { | |||
434 | if (gpu_array_is_read_only_scalar(array)) | |||
435 | return 0; | |||
436 | if (!array->global) | |||
437 | return 0; | |||
438 | return 1; | |||
439 | } | |||
440 | ||||
441 | /* Return the set of parameter values for which the array has a positive | |||
442 | * size in all dimensions. | |||
443 | * If the sizes are only valid for some parameter values, then those | |||
444 | * constraints are also taken into account. | |||
445 | */ | |||
446 | __isl_give isl_set *gpu_array_positive_size_guard(struct gpu_array_info *array) | |||
447 | { | |||
448 | int i; | |||
449 | isl_space *space; | |||
450 | isl_set *guard; | |||
451 | ||||
452 | if (!array) | |||
453 | return NULL((void*)0); | |||
454 | ||||
455 | space = isl_space_params(isl_space_copy(array->space)); | |||
456 | guard = isl_set_universe(space); | |||
457 | ||||
458 | for (i = 0; i < array->n_index; ++i) { | |||
459 | isl_pw_aff *bound; | |||
460 | isl_set *guard_i, *zero; | |||
461 | ||||
462 | bound = isl_multi_pw_aff_get_pw_aff(array->bound, i); | |||
463 | guard_i = isl_pw_aff_nonneg_set(isl_pw_aff_copy(bound)); | |||
464 | zero = isl_pw_aff_zero_set(bound); | |||
465 | guard_i = isl_set_subtract(guard_i, zero); | |||
466 | guard = isl_set_intersect(guard, guard_i); | |||
467 | } | |||
468 | ||||
469 | return guard; | |||
470 | } | |||
471 | ||||
472 | /* Internal data structure for extract_size_of_type. | |||
473 | * "type" specifies the name of the space that we want to extract. | |||
474 | * "res" is used to store the subset of that space. | |||
475 | */ | |||
476 | struct ppcg_extract_size_data { | |||
477 | const char *type; | |||
478 | isl_set *res; | |||
479 | }; | |||
480 | ||||
481 | /* This function is called for each set in a union_set. | |||
482 | * If the name of the set matches data->type, we store the | |||
483 | * set in data->res. | |||
484 | */ | |||
485 | static isl_stat extract_size_of_type(__isl_take isl_set *size, void *user) | |||
486 | { | |||
487 | struct ppcg_extract_size_data *data = user; | |||
488 | const char *name; | |||
489 | ||||
490 | name = isl_set_get_tuple_name(size); | |||
491 | if (name && !strcmp(name, data->type)) { | |||
492 | data->res = size; | |||
493 | return isl_stat_error; | |||
494 | } | |||
495 | ||||
496 | isl_set_free(size); | |||
497 | return isl_stat_ok; | |||
498 | } | |||
499 | ||||
500 | /* Given a union map { kernel[i] -> *[...] }, | |||
501 | * return the range in the space called "type" for the kernel with | |||
502 | * sequence number "id". | |||
503 | */ | |||
504 | static __isl_give isl_set *extract_sizes(__isl_keep isl_union_map *sizes, | |||
505 | const char *type, int id) | |||
506 | { | |||
507 | isl_space *space; | |||
508 | isl_set *dom; | |||
509 | isl_union_set *local_sizes; | |||
510 | struct ppcg_extract_size_data data = { type, NULL((void*)0) }; | |||
511 | ||||
512 | if (!sizes) | |||
513 | return NULL((void*)0); | |||
514 | ||||
515 | space = isl_union_map_get_space(sizes); | |||
516 | space = isl_space_set_from_params(space); | |||
517 | space = isl_space_add_dims(space, isl_dim_set, 1); | |||
518 | space = isl_space_set_tuple_name(space, isl_dim_set, "kernel"); | |||
519 | dom = isl_set_universe(space); | |||
520 | dom = isl_set_fix_si(dom, isl_dim_set, 0, id); | |||
521 | ||||
522 | local_sizes = isl_union_set_apply(isl_union_set_from_set(dom), | |||
523 | isl_union_map_copy(sizes)); | |||
524 | isl_union_set_foreach_set(local_sizes, &extract_size_of_type, &data); | |||
525 | isl_union_set_free(local_sizes); | |||
526 | return data.res; | |||
527 | } | |||
528 | ||||
529 | /* Given a singleton set, extract the first (at most *len) elements | |||
530 | * of the single integer tuple into *sizes and update *len if needed. | |||
531 | */ | |||
532 | static void read_sizes_from_set(__isl_take isl_set *set, int *sizes, int *len) | |||
533 | { | |||
534 | int i; | |||
535 | int dim; | |||
536 | ||||
537 | if (!set) | |||
538 | return; | |||
539 | ||||
540 | dim = isl_set_dim(set, isl_dim_set); | |||
541 | if (dim < *len) | |||
542 | *len = dim; | |||
543 | ||||
544 | for (i = 0; i < *len; ++i) { | |||
545 | isl_val *v; | |||
546 | ||||
547 | v = isl_set_plain_get_val_if_fixed(set, isl_dim_set, i); | |||
548 | assert(v)((void) sizeof ((v) ? 1 : 0), __extension__ ({ if (v) ; else __assert_fail ("v", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 548, __extension__ __PRETTY_FUNCTION__); })); | |||
549 | ||||
550 | sizes[i] = isl_val_get_num_si(v); | |||
551 | isl_val_free(v); | |||
552 | } | |||
553 | ||||
554 | isl_set_free(set); | |||
555 | } | |||
556 | ||||
557 | /* Add the map { kernel[id] -> type[sizes] } to gen->used_sizes, | |||
558 | * if the option debug->dump_sizes is set. | |||
559 | */ | |||
560 | static void set_used_sizes(struct gpu_gen *gen, const char *type, int id, | |||
561 | int *sizes, int len) | |||
562 | { | |||
563 | int i; | |||
564 | isl_space *space; | |||
565 | isl_map *map; | |||
566 | ||||
567 | if (!gen->options->debug->dump_sizes) | |||
568 | return; | |||
569 | ||||
570 | space = isl_union_map_get_space(gen->used_sizes); | |||
571 | space = isl_space_set_from_params(space); | |||
572 | space = isl_space_add_dims(space, isl_dim_set, 1); | |||
573 | space = isl_space_set_tuple_name(space, isl_dim_set, "kernel"); | |||
574 | space = isl_space_from_domain(space); | |||
575 | space = isl_space_add_dims(space, isl_dim_out, len); | |||
576 | space = isl_space_set_tuple_name(space, isl_dim_out, type); | |||
577 | ||||
578 | map = isl_map_universe(space); | |||
579 | map = isl_map_fix_si(map, isl_dim_in, 0, id); | |||
580 | for (i = 0; i < len; ++i) | |||
581 | map = isl_map_fix_si(map, isl_dim_out, i, sizes[i]); | |||
582 | ||||
583 | gen->used_sizes = isl_union_map_add_map(gen->used_sizes, map); | |||
584 | } | |||
585 | ||||
586 | /* Extract user specified "tile" sizes from the "sizes" command line option, | |||
587 | * defaulting to option->tile_size in each dimension. | |||
588 | * *tile_len contains the maximum number of tile sizes needed. | |||
589 | * Update *tile_len to the number of specified tile sizes, if any, and | |||
590 | * return a pointer to the tile sizes (or NULL on error). | |||
591 | * Add the effectively used sizes to gen->used_sizes. | |||
592 | */ | |||
593 | static int *read_tile_sizes(struct gpu_gen *gen, int *tile_len) | |||
594 | { | |||
595 | int n; | |||
596 | int *tile_size; | |||
597 | isl_set *size; | |||
598 | ||||
599 | tile_size = isl_alloc_array(gen->ctx, int, *tile_len)((int *)isl_malloc_or_die(gen->ctx, (*tile_len)*sizeof(int ))); | |||
600 | if (!tile_size) | |||
601 | return NULL((void*)0); | |||
602 | for (n = 0; n < *tile_len; ++n) | |||
603 | tile_size[n] = gen->options->tile_size; | |||
604 | ||||
605 | size = extract_sizes(gen->sizes, "tile", gen->kernel_id); | |||
606 | read_sizes_from_set(size, tile_size, tile_len); | |||
607 | set_used_sizes(gen, "tile", gen->kernel_id, tile_size, *tile_len); | |||
608 | ||||
609 | return tile_size; | |||
610 | } | |||
611 | ||||
612 | /* Extract user specified "block" sizes from the "sizes" command line option, | |||
613 | * after filling in some potentially useful defaults. | |||
614 | */ | |||
615 | static void read_block_sizes(struct ppcg_kernel *kernel, | |||
616 | __isl_keep isl_union_map *sizes) | |||
617 | { | |||
618 | isl_set *size; | |||
619 | ||||
620 | if (kernel->n_block > 3) | |||
621 | kernel->n_block = 3; | |||
622 | switch (kernel->n_block) { | |||
623 | case 1: | |||
624 | kernel->block_dim[0] = 512; | |||
625 | break; | |||
626 | case 2: | |||
627 | kernel->block_dim[0] = 32; | |||
628 | kernel->block_dim[1] = 16; | |||
629 | break; | |||
630 | default: | |||
631 | kernel->block_dim[0] = 32; | |||
632 | kernel->block_dim[1] = 4; | |||
633 | kernel->block_dim[2] = 4; | |||
634 | break; | |||
635 | } | |||
636 | ||||
637 | size = extract_sizes(sizes, "block", kernel->id); | |||
638 | read_sizes_from_set(size, kernel->block_dim, &kernel->n_block); | |||
639 | } | |||
640 | ||||
641 | /* Extract user specified "grid" sizes from the "sizes" command line option, | |||
642 | * after filling in some potentially useful defaults. | |||
643 | */ | |||
644 | static void read_grid_sizes(struct ppcg_kernel *kernel, | |||
645 | __isl_keep isl_union_map *sizes) | |||
646 | { | |||
647 | isl_set *size; | |||
648 | ||||
649 | if (kernel->n_grid > 2) | |||
650 | kernel->n_grid = 2; | |||
651 | switch (kernel->n_grid) { | |||
652 | case 1: | |||
653 | kernel->grid_dim[0] = 32768; | |||
654 | break; | |||
655 | default: | |||
656 | kernel->grid_dim[0] = 256; | |||
657 | kernel->grid_dim[1] = 256; | |||
658 | break; | |||
659 | } | |||
660 | ||||
661 | size = extract_sizes(sizes, "grid", kernel->id); | |||
662 | read_sizes_from_set(size, kernel->grid_dim, &kernel->n_grid); | |||
663 | } | |||
664 | ||||
665 | /* Extract user specified grid and block sizes from the gen->sizes | |||
666 | * command line option after filling in some potentially useful defaults. | |||
667 | * Store the extracted sizes in "kernel". | |||
668 | * Add the effectively used sizes to gen->used_sizes. | |||
669 | */ | |||
670 | static void read_grid_and_block_sizes(struct ppcg_kernel *kernel, | |||
671 | struct gpu_gen *gen) | |||
672 | { | |||
673 | read_block_sizes(kernel, gen->sizes); | |||
674 | read_grid_sizes(kernel, gen->sizes); | |||
675 | set_used_sizes(gen, "block", kernel->id, | |||
676 | kernel->block_dim, kernel->n_block); | |||
677 | set_used_sizes(gen, "grid", kernel->id, | |||
678 | kernel->grid_dim, kernel->n_grid); | |||
679 | } | |||
680 | ||||
681 | static void *free_stmts(struct gpu_stmt *stmts, int n) | |||
682 | { | |||
683 | int i; | |||
684 | ||||
685 | if (!stmts) | |||
686 | return NULL((void*)0); | |||
687 | ||||
688 | for (i = 0; i < n; ++i) { | |||
689 | struct gpu_stmt_access *access, *next; | |||
690 | ||||
691 | for (access = stmts[i].accesses; access; access = next) { | |||
692 | next = access->next; | |||
693 | isl_id_free(access->ref_id); | |||
694 | isl_map_free(access->access); | |||
695 | isl_map_free(access->tagged_access); | |||
696 | free(access); | |||
697 | } | |||
698 | ||||
699 | isl_id_free(stmts[i].id); | |||
700 | } | |||
701 | free(stmts); | |||
702 | ||||
703 | return NULL((void*)0); | |||
704 | } | |||
705 | ||||
706 | /* Add parameters p[i] with identifiers "ids" to "set", | |||
707 | * with bounds to 0 <= p[i] < size[i]. | |||
708 | */ | |||
709 | __isl_give isl_set *add_bounded_parameters(__isl_take isl_set *set, | |||
710 | int *size, __isl_keep isl_id_list *ids) | |||
711 | { | |||
712 | int i, len; | |||
713 | unsigned nparam; | |||
714 | ||||
715 | len = isl_id_list_n_id(ids); | |||
716 | nparam = isl_set_dim(set, isl_dim_param); | |||
717 | set = isl_set_add_dims(set, isl_dim_param, len); | |||
718 | ||||
719 | for (i = 0; i < len; ++i) { | |||
720 | isl_id *id; | |||
721 | ||||
722 | id = isl_id_list_get_id(ids, i); | |||
723 | set = isl_set_set_dim_id(set, isl_dim_param, nparam + i, id); | |||
724 | set = isl_set_lower_bound_si(set, isl_dim_param, nparam + i, 0); | |||
725 | set = isl_set_upper_bound_si(set, isl_dim_param, | |||
726 | nparam + i, size[i] - 1); | |||
727 | } | |||
728 | ||||
729 | return set; | |||
730 | } | |||
731 | ||||
732 | /* Add "len" parameters p[i] with identifiers "ids" and intersect "set" | |||
733 | * with | |||
734 | * | |||
735 | * { : 0 <= p[i] < size[i] } | |||
736 | * | |||
737 | * or an overapproximation. | |||
738 | */ | |||
739 | static __isl_give isl_set *add_bounded_parameters_dynamic( | |||
740 | __isl_take isl_set *set, __isl_keep isl_multi_pw_aff *size, | |||
741 | __isl_keep isl_id_list *ids) | |||
742 | { | |||
743 | int i, len; | |||
744 | unsigned nparam; | |||
745 | isl_space *space; | |||
746 | isl_local_space *ls; | |||
747 | ||||
748 | len = isl_multi_pw_aff_dim(size, isl_dim_out); | |||
749 | nparam = isl_set_dim(set, isl_dim_param); | |||
750 | set = isl_set_add_dims(set, isl_dim_param, len); | |||
751 | ||||
752 | for (i = 0; i < len; ++i) { | |||
753 | isl_id *id; | |||
754 | ||||
755 | id = isl_id_list_get_id(ids, i); | |||
756 | set = isl_set_set_dim_id(set, isl_dim_param, nparam + i, id); | |||
757 | } | |||
758 | ||||
759 | space = isl_space_params(isl_set_get_space(set)); | |||
760 | ls = isl_local_space_from_space(space); | |||
761 | for (i = 0; i < len; ++i) { | |||
762 | isl_pw_aff *param, *size_i, *zero; | |||
763 | isl_set *bound; | |||
764 | ||||
765 | param = isl_pw_aff_var_on_domain(isl_local_space_copy(ls), | |||
766 | isl_dim_param, nparam + i); | |||
767 | ||||
768 | size_i = isl_multi_pw_aff_get_pw_aff(size, i); | |||
769 | bound = isl_pw_aff_lt_set(isl_pw_aff_copy(param), size_i); | |||
770 | bound = isl_set_from_basic_set(isl_set_simple_hull(bound)); | |||
771 | set = isl_set_intersect_params(set, bound); | |||
772 | ||||
773 | zero = isl_pw_aff_zero_on_domain(isl_local_space_copy(ls)); | |||
774 | bound = isl_pw_aff_ge_set(param, zero); | |||
775 | set = isl_set_intersect_params(set, bound); | |||
776 | } | |||
777 | isl_local_space_free(ls); | |||
778 | ||||
779 | return set; | |||
780 | } | |||
781 | ||||
782 | /* Return the union of all tagged access relations in the group. | |||
783 | */ | |||
784 | static __isl_give isl_union_map *group_tagged_access_relation( | |||
785 | struct gpu_array_ref_group *group) | |||
786 | { | |||
787 | int i; | |||
788 | isl_union_map *access; | |||
789 | ||||
790 | access = isl_union_map_empty(isl_map_get_space(group->access)); | |||
791 | for (i = 0; i < group->n_ref; ++i) { | |||
792 | isl_map *map_i; | |||
793 | ||||
794 | map_i = isl_map_copy(group->refs[i]->tagged_access); | |||
795 | access = isl_union_map_union(access, | |||
796 | isl_union_map_from_map(map_i)); | |||
797 | } | |||
798 | ||||
799 | return access; | |||
800 | } | |||
801 | ||||
802 | /* Return the extent of "array", recomputed from the bounds. | |||
803 | * The recomputed extent may be simpler than the original extent. | |||
804 | */ | |||
805 | static __isl_give isl_set *array_extent(struct gpu_array_info *array) | |||
806 | { | |||
807 | int i; | |||
808 | isl_id *id; | |||
809 | isl_space *space; | |||
810 | isl_local_space *ls; | |||
811 | isl_set *extent; | |||
812 | ||||
813 | id = isl_set_get_tuple_id(array->extent); | |||
814 | space = isl_set_get_space(array->extent); | |||
815 | extent = isl_set_universe(isl_space_copy(space)); | |||
816 | ls = isl_local_space_from_space(space); | |||
817 | for (i = 0; i < array->n_index; ++i) { | |||
818 | isl_pw_aff *bound; | |||
819 | isl_aff *aff; | |||
820 | isl_pw_aff *index; | |||
821 | isl_set *lt; | |||
822 | ||||
823 | extent = isl_set_lower_bound_si(extent, isl_dim_set, i, 0); | |||
824 | ||||
825 | aff = isl_aff_var_on_domain(isl_local_space_copy(ls), | |||
826 | isl_dim_set, i); | |||
827 | index = isl_pw_aff_from_aff(aff); | |||
828 | bound = isl_multi_pw_aff_get_pw_aff(array->bound, i); | |||
829 | bound = isl_pw_aff_from_range(bound); | |||
830 | bound = isl_pw_aff_add_dims(bound, isl_dim_in, array->n_index); | |||
831 | bound = isl_pw_aff_set_tuple_id(bound, isl_dim_in, | |||
832 | isl_id_copy(id)); | |||
833 | lt = isl_pw_aff_lt_set(index, bound); | |||
834 | extent = isl_set_intersect(extent, lt); | |||
835 | } | |||
836 | isl_local_space_free(ls); | |||
837 | isl_id_free(id); | |||
838 | ||||
839 | return extent; | |||
840 | } | |||
841 | ||||
842 | /* Return a map from the first group->shared_tile->depth dimensions | |||
843 | * of the computed schedule to the array tile in | |||
844 | * global memory that corresponds to the shared memory copy. | |||
845 | * | |||
846 | * In particular, return a map | |||
847 | * | |||
848 | * { D[i] -> A[a] } | |||
849 | * | |||
850 | * with constraints | |||
851 | * | |||
852 | * tile_offset(i) <= a <= tile_offset(i) + tile_size - 1 (1) | |||
853 | * | |||
854 | * and | |||
855 | * | |||
856 | * 0 <= a <= array_size - 1 (2) | |||
857 | * | |||
858 | * Note that if some stride has been detected (i.e., when | |||
859 | * group->shared_tile->bound[i].shift is set), then a in (1) refers | |||
860 | * to the shifted and scaled down version. | |||
861 | * | |||
862 | * Constraints (1) are obtained by mapping the size constraints on the | |||
863 | * shared/private memory tile back to the access relation. | |||
864 | * Constraints (2) are obtained from the (recomputed) extent. | |||
865 | */ | |||
866 | static __isl_give isl_map *group_tile(struct gpu_array_ref_group *group) | |||
867 | { | |||
868 | int i; | |||
869 | int n_index = group->array->n_index; | |||
870 | isl_map *tile; | |||
871 | isl_space *space; | |||
872 | isl_set *local; | |||
873 | isl_set *extent; | |||
874 | ||||
875 | space = isl_multi_aff_get_space(group->shared_tile->tiling); | |||
876 | space = isl_space_range(space); | |||
877 | local = isl_set_universe(space); | |||
878 | for (i = 0; i < n_index; ++i) { | |||
879 | isl_val *bound; | |||
880 | ||||
881 | local = isl_set_lower_bound_si(local, isl_dim_set, i, 0); | |||
882 | bound = isl_val_copy(group->shared_tile->bound[i].size); | |||
883 | bound = isl_val_sub_ui(bound, 1); | |||
884 | local = isl_set_upper_bound_val(local, isl_dim_set, i, bound); | |||
885 | } | |||
886 | local = isl_set_preimage_multi_aff(local, | |||
887 | isl_multi_aff_copy(group->shared_tile->tiling)); | |||
888 | tile = isl_set_unwrap(local); | |||
889 | extent = array_extent(group->array); | |||
890 | tile = isl_map_intersect_range(tile, extent); | |||
891 | ||||
892 | return tile; | |||
893 | } | |||
894 | ||||
895 | /* Given a mapping "iterator_map" from the AST schedule to a domain, | |||
896 | * return the corresponding mapping from the AST schedule to | |||
897 | * to the outer kernel->copy_schedule_dim dimensions of | |||
898 | * the schedule computed by PPCG for this kernel. | |||
899 | * | |||
900 | * Note that kernel->copy_schedule_dim is at least as large as | |||
901 | * the largest depth of any array reference group associated to the kernel. | |||
902 | * This is needed as the returned schedule is used to extract a mapping | |||
903 | * to the outer tile->depth dimensions in transform_index. | |||
904 | */ | |||
905 | static __isl_give isl_pw_multi_aff *compute_sched_to_copy( | |||
906 | struct ppcg_kernel *kernel, __isl_take isl_pw_multi_aff *iterator_map) | |||
907 | { | |||
908 | isl_union_pw_multi_aff *upma; | |||
909 | isl_pw_multi_aff *pma; | |||
910 | isl_space *space; | |||
911 | ||||
912 | space = isl_space_range(isl_pw_multi_aff_get_space(iterator_map)); | |||
913 | space = isl_space_from_domain(space); | |||
914 | space = isl_space_add_dims(space, isl_dim_out, | |||
915 | kernel->copy_schedule_dim); | |||
916 | ||||
917 | upma = isl_union_pw_multi_aff_copy(kernel->copy_schedule); | |||
918 | pma = isl_union_pw_multi_aff_extract_pw_multi_aff(upma, space); | |||
919 | isl_union_pw_multi_aff_free(upma); | |||
920 | ||||
921 | return isl_pw_multi_aff_pullback_pw_multi_aff(pma, iterator_map); | |||
922 | } | |||
923 | ||||
924 | /* If max_shared_memory is not set to infinity (-1), then make | |||
925 | * sure that the total amount of shared memory required by the | |||
926 | * array reference groups mapped to shared memory by "kernel" | |||
927 | * is no larger than this maximum. | |||
928 | * | |||
929 | * We apply a greedy approach and discard (keep in global memory) | |||
930 | * those groups that would result in a total memory size that | |||
931 | * is larger than the maximum. | |||
932 | * | |||
933 | * This function should be called after any function that may | |||
934 | * affect the decision on whether to place a reference group | |||
935 | * in private, shared or global memory. | |||
936 | */ | |||
937 | static void check_shared_memory_bound(struct ppcg_kernel *kernel) | |||
938 | { | |||
939 | int i, j; | |||
940 | isl_val *left, *size; | |||
941 | ||||
942 | if (kernel->options->max_shared_memory < 0) | |||
943 | return; | |||
944 | ||||
945 | left = isl_val_int_from_si(kernel->ctx, | |||
946 | kernel->options->max_shared_memory); | |||
947 | ||||
948 | for (i = 0; i < kernel->n_array; ++i) { | |||
949 | struct gpu_local_array_info *local = &kernel->array[i]; | |||
950 | ||||
951 | for (j = 0; j < local->n_group; ++j) { | |||
952 | struct gpu_array_ref_group *group; | |||
953 | enum ppcg_group_access_type type; | |||
954 | ||||
955 | group = local->groups[j]; | |||
956 | type = gpu_array_ref_group_type(group); | |||
957 | if (type != ppcg_access_shared) | |||
958 | continue; | |||
959 | ||||
960 | size = gpu_array_tile_size(group->shared_tile); | |||
961 | size = isl_val_mul_ui(size, local->array->size); | |||
962 | ||||
963 | if (isl_val_le(size, left)) { | |||
964 | left = isl_val_sub(left, size); | |||
965 | continue; | |||
966 | } | |||
967 | isl_val_free(size); | |||
968 | ||||
969 | group->shared_tile = | |||
970 | gpu_array_tile_free(group->shared_tile); | |||
971 | } | |||
972 | } | |||
973 | ||||
974 | isl_val_free(left); | |||
975 | } | |||
976 | ||||
977 | /* Mark all arrays of "kernel" that have an array reference group | |||
978 | * that is not mapped to private or shared memory as | |||
979 | * accessing the corresponding global device memory. | |||
980 | */ | |||
981 | static void mark_global_arrays(struct ppcg_kernel *kernel) | |||
982 | { | |||
983 | int i, j; | |||
984 | ||||
985 | for (i = 0; i < kernel->n_array; ++i) { | |||
986 | struct gpu_local_array_info *local = &kernel->array[i]; | |||
987 | ||||
988 | if (local->global) | |||
989 | continue; | |||
990 | for (j = 0; j < local->n_group; ++j) { | |||
991 | if (gpu_array_ref_group_tile(local->groups[j])) | |||
992 | continue; | |||
993 | ||||
994 | local->global = 1; | |||
995 | local->array->global = 1; | |||
996 | break; | |||
997 | } | |||
998 | } | |||
999 | } | |||
1000 | ||||
1001 | /* Compute a tiling for all the array reference groups in "kernel". | |||
1002 | */ | |||
1003 | static void compute_group_tilings(struct ppcg_kernel *kernel) | |||
1004 | { | |||
1005 | int i, j; | |||
1006 | ||||
1007 | for (i = 0; i < kernel->n_array; ++i) { | |||
1008 | struct gpu_local_array_info *array = &kernel->array[i]; | |||
1009 | ||||
1010 | for (j = 0; j < array->n_group; ++j) | |||
1011 | gpu_array_ref_group_compute_tiling(array->groups[j]); | |||
1012 | } | |||
1013 | } | |||
1014 | ||||
1015 | /* Compute the effective grid size as a list of the sizes in each dimension. | |||
1016 | * | |||
1017 | * The grid size specified by the user or set by default | |||
1018 | * in read_grid_sizes() and applied by the block filter, | |||
1019 | * may be too large for the given code in the sense that | |||
1020 | * it may contain blocks that don't need to execute anything. | |||
1021 | * We therefore don't return this grid size, but instead the | |||
1022 | * smallest grid size that ensures that all blocks that actually | |||
1023 | * execute code are included in the grid. | |||
1024 | * | |||
1025 | * We first extract a description of the grid, i.e., the possible values | |||
1026 | * of the block ids, from the domain elements in "domain" and | |||
1027 | * kernel->block_filter. | |||
1028 | * The block ids are parameters in kernel->block_filter. | |||
1029 | * We simply need to change them into set dimensions. | |||
1030 | * | |||
1031 | * Then, for each block dimension, we compute the maximal value of the block id | |||
1032 | * and add one. | |||
1033 | */ | |||
1034 | static __isl_give isl_multi_pw_aff *extract_grid_size( | |||
1035 | struct ppcg_kernel *kernel, __isl_take isl_union_set *domain) | |||
1036 | { | |||
1037 | int i; | |||
1038 | isl_set *grid; | |||
1039 | isl_set *context; | |||
1040 | isl_multi_pw_aff *size; | |||
1041 | ||||
1042 | domain = isl_union_set_intersect(domain, | |||
1043 | isl_union_set_copy(kernel->block_filter)); | |||
1044 | grid = isl_union_set_params(domain); | |||
1045 | grid = isl_set_from_params(grid); | |||
1046 | grid = isl_set_add_dims(grid, isl_dim_set, kernel->n_grid); | |||
1047 | for (i = 0; i < kernel->n_grid; ++i) { | |||
1048 | int pos; | |||
1049 | isl_id *id; | |||
1050 | ||||
1051 | id = isl_id_list_get_id(kernel->block_ids, i); | |||
1052 | pos = isl_set_find_dim_by_id(grid, isl_dim_param, id); | |||
1053 | isl_id_free(id); | |||
1054 | assert(pos >= 0)((void) sizeof ((pos >= 0) ? 1 : 0), __extension__ ({ if ( pos >= 0) ; else __assert_fail ("pos >= 0", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 1054, __extension__ __PRETTY_FUNCTION__); })); | |||
1055 | grid = isl_set_equate(grid, isl_dim_param, pos, isl_dim_set, i); | |||
1056 | grid = isl_set_project_out(grid, isl_dim_param, pos, 1); | |||
1057 | } | |||
1058 | ||||
1059 | grid = isl_set_coalesce(grid); | |||
1060 | size = ppcg_size_from_extent(grid); | |||
1061 | context = isl_set_params(isl_set_copy(kernel->context)); | |||
1062 | return isl_multi_pw_aff_gist(size, context); | |||
1063 | } | |||
1064 | ||||
1065 | /* Compute the size of a fixed bounding box around the origin and "set", | |||
1066 | * where "set" is assumed to contain only non-negative elements, | |||
1067 | * and store the results in "size". | |||
1068 | * In particular, compute the maximal value of "set" in each direction | |||
1069 | * and add one. | |||
1070 | */ | |||
1071 | static void extract_fixed_size(__isl_take isl_set *set, int *size) | |||
1072 | { | |||
1073 | int i, n; | |||
1074 | isl_local_space *ls; | |||
1075 | isl_aff *obj; | |||
1076 | ||||
1077 | n = isl_set_dim(set, isl_dim_set); | |||
1078 | ls = isl_local_space_from_space(isl_set_get_space(set)); | |||
1079 | obj = isl_aff_zero_on_domain(ls); | |||
1080 | for (i = 0; i < n; ++i) { | |||
1081 | isl_val *max; | |||
1082 | ||||
1083 | obj = isl_aff_set_coefficient_si(obj, isl_dim_in, i, 1); | |||
1084 | max = isl_set_max_val(set, obj); | |||
1085 | size[i] = isl_val_get_num_si(max) + 1; | |||
1086 | isl_val_free(max); | |||
1087 | obj = isl_aff_set_coefficient_si(obj, isl_dim_in, i, 0); | |||
1088 | } | |||
1089 | isl_aff_free(obj); | |||
1090 | isl_set_free(set); | |||
1091 | } | |||
1092 | ||||
1093 | /* Compute the effective block size as a list of the sizes in each dimension | |||
1094 | * and store the sizes in kernel->block_dim. | |||
1095 | * | |||
1096 | * The block size specified by the user or set by default | |||
1097 | * in read_block_sizes() and applied by the thread filter, | |||
1098 | * may be too large for the given code in the sense that | |||
1099 | * it may contain threads that don't need to execute anything. | |||
1100 | * We therefore update this block size in kernel->block_dim | |||
1101 | * to the smallest block size that ensures that all threads | |||
1102 | * that actually execute code are included in the block. | |||
1103 | * | |||
1104 | * The set of possible values of the thread ids is obtained from | |||
1105 | * the domain elements "domain" and kernel->thread_filter. | |||
1106 | * The current implementation eliminates all parameters, ensuring | |||
1107 | * that the size is a fixed constant in each dimension. | |||
1108 | * In principle we could also compute parametric sizes. | |||
1109 | * We would have to make sure to project out all b%d and t%d parameters, | |||
1110 | * however. | |||
1111 | */ | |||
1112 | static isl_stat extract_block_size(struct ppcg_kernel *kernel, | |||
1113 | __isl_take isl_union_set *domain) | |||
1114 | { | |||
1115 | int i; | |||
1116 | int nparam; | |||
1117 | isl_set *block; | |||
1118 | ||||
1119 | domain = isl_union_set_intersect(domain, | |||
1120 | isl_union_set_copy(kernel->thread_filter)); | |||
1121 | block = isl_union_set_params(domain); | |||
1122 | block = isl_set_from_params(block); | |||
1123 | block = isl_set_add_dims(block, isl_dim_set, kernel->n_block); | |||
1124 | for (i = 0; i < kernel->n_block; ++i) { | |||
1125 | int pos; | |||
1126 | isl_id *id; | |||
1127 | ||||
1128 | if (!block) | |||
1129 | return isl_stat_error; | |||
1130 | ||||
1131 | id = isl_id_list_get_id(kernel->thread_ids, i); | |||
1132 | pos = isl_set_find_dim_by_id(block, isl_dim_param, id); | |||
1133 | isl_id_free(id); | |||
1134 | if (pos < 0) | |||
1135 | isl_die(isl_set_get_ctx(block), isl_error_internal,do { isl_handle_error(isl_set_get_ctx(block), isl_error_internal , "missing constraints on thread identifier", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 1137); block = isl_set_free(block); } while (0) | |||
1136 | "missing constraints on thread identifier",do { isl_handle_error(isl_set_get_ctx(block), isl_error_internal , "missing constraints on thread identifier", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 1137); block = isl_set_free(block); } while (0) | |||
1137 | block = isl_set_free(block))do { isl_handle_error(isl_set_get_ctx(block), isl_error_internal , "missing constraints on thread identifier", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 1137); block = isl_set_free(block); } while (0); | |||
1138 | block = isl_set_equate(block, isl_dim_param, pos, | |||
1139 | isl_dim_set, i); | |||
1140 | } | |||
1141 | nparam = isl_set_dim(block, isl_dim_param); | |||
1142 | block = isl_set_project_out(block, isl_dim_param, 0, nparam); | |||
1143 | ||||
1144 | if (!block) | |||
1145 | return isl_stat_error; | |||
1146 | ||||
1147 | extract_fixed_size(block, kernel->block_dim); | |||
1148 | ||||
1149 | return isl_stat_ok; | |||
1150 | } | |||
1151 | ||||
1152 | struct ppcg_kernel *ppcg_kernel_free(struct ppcg_kernel *kernel) | |||
1153 | { | |||
1154 | int i, j; | |||
1155 | ||||
1156 | if (!kernel) | |||
1157 | return NULL((void*)0); | |||
1158 | ||||
1159 | isl_id_list_free(kernel->block_ids); | |||
1160 | isl_id_list_free(kernel->thread_ids); | |||
1161 | isl_multi_pw_aff_free(kernel->grid_size); | |||
1162 | isl_ast_expr_free(kernel->grid_size_expr); | |||
1163 | isl_set_free(kernel->context); | |||
1164 | isl_union_set_free(kernel->core); | |||
1165 | isl_union_set_free(kernel->arrays); | |||
1166 | isl_union_pw_multi_aff_free(kernel->contraction); | |||
1167 | isl_union_set_free(kernel->expanded_domain); | |||
1168 | isl_space_free(kernel->space); | |||
1169 | isl_ast_node_free(kernel->tree); | |||
1170 | isl_union_set_free(kernel->block_filter); | |||
1171 | isl_union_set_free(kernel->thread_filter); | |||
1172 | isl_union_pw_multi_aff_free(kernel->copy_schedule); | |||
1173 | isl_union_set_free(kernel->sync_writes); | |||
1174 | ||||
1175 | for (i = 0; i < kernel->n_array; ++i) { | |||
1176 | struct gpu_local_array_info *array = &kernel->array[i]; | |||
1177 | ||||
1178 | for (j = 0; j < array->n_group; ++j) | |||
| ||||
1179 | gpu_array_ref_group_free(array->groups[j]); | |||
1180 | free(array->groups); | |||
1181 | ||||
1182 | isl_multi_pw_aff_free(array->bound); | |||
1183 | isl_ast_expr_free(array->bound_expr); | |||
1184 | } | |||
1185 | free(kernel->array); | |||
1186 | ||||
1187 | for (i = 0; i < kernel->n_var; ++i) { | |||
1188 | free(kernel->var[i].name); | |||
1189 | isl_vec_free(kernel->var[i].size); | |||
1190 | } | |||
1191 | free(kernel->var); | |||
1192 | ||||
1193 | free(kernel); | |||
1194 | ||||
1195 | return NULL((void*)0); | |||
1196 | } | |||
1197 | ||||
1198 | /* Wrapper around ppcg_kernel_free for use as a isl_id_set_free_user callback. | |||
1199 | */ | |||
1200 | static void ppcg_kernel_free_wrap(void *user) | |||
1201 | { | |||
1202 | struct ppcg_kernel *kernel = user; | |||
1203 | ||||
1204 | ppcg_kernel_free(kernel); | |||
1205 | } | |||
1206 | ||||
1207 | static void create_kernel_var(isl_ctx *ctx, struct gpu_array_ref_group *group, | |||
1208 | struct ppcg_kernel_var *var) | |||
1209 | { | |||
1210 | int j; | |||
1211 | struct gpu_array_tile *tile; | |||
1212 | isl_printer *p; | |||
1213 | ||||
1214 | var->array = group->array; | |||
1215 | ||||
1216 | var->type = gpu_array_ref_group_type(group); | |||
1217 | tile = gpu_array_ref_group_tile(group); | |||
1218 | ||||
1219 | p = isl_printer_to_str(ctx); | |||
1220 | p = gpu_array_ref_group_print_name(group, p); | |||
1221 | var->name = isl_printer_get_str(p); | |||
1222 | isl_printer_free(p); | |||
1223 | ||||
1224 | var->size = isl_vec_alloc(ctx, group->array->n_index); | |||
1225 | ||||
1226 | for (j = 0; j < group->array->n_index; ++j) | |||
1227 | var->size = isl_vec_set_element_val(var->size, j, | |||
1228 | isl_val_copy(tile->bound[j].size)); | |||
1229 | } | |||
1230 | ||||
1231 | static int create_kernel_vars(struct ppcg_kernel *kernel) | |||
1232 | { | |||
1233 | int i, j, n; | |||
1234 | ||||
1235 | n = 0; | |||
1236 | for (i = 0; i < kernel->n_array; ++i) { | |||
1237 | struct gpu_local_array_info *array = &kernel->array[i]; | |||
1238 | ||||
1239 | for (j = 0; j < array->n_group; ++j) { | |||
1240 | struct gpu_array_ref_group *group = array->groups[j]; | |||
1241 | enum ppcg_group_access_type type; | |||
1242 | ||||
1243 | type = gpu_array_ref_group_type(group); | |||
1244 | if (type != ppcg_access_global) | |||
1245 | ++n; | |||
1246 | } | |||
1247 | } | |||
1248 | ||||
1249 | kernel->n_var = n; | |||
1250 | kernel->var = isl_calloc_array(kernel->ctx, struct ppcg_kernel_var, n)((struct ppcg_kernel_var *)isl_calloc_or_die(kernel->ctx, n , sizeof(struct ppcg_kernel_var))); | |||
1251 | if (!kernel->var) | |||
1252 | return -1; | |||
1253 | ||||
1254 | n = 0; | |||
1255 | for (i = 0; i < kernel->n_array; ++i) { | |||
1256 | struct gpu_local_array_info *array = &kernel->array[i]; | |||
1257 | ||||
1258 | for (j = 0; j < array->n_group; ++j) { | |||
1259 | struct gpu_array_ref_group *group = array->groups[j]; | |||
1260 | enum ppcg_group_access_type type; | |||
1261 | ||||
1262 | type = gpu_array_ref_group_type(group); | |||
1263 | if (type == ppcg_access_global) | |||
1264 | continue; | |||
1265 | create_kernel_var(kernel->ctx, group, &kernel->var[n]); | |||
1266 | ++n; | |||
1267 | } | |||
1268 | } | |||
1269 | ||||
1270 | return 0; | |||
1271 | } | |||
1272 | ||||
1273 | /* Replace "pa" by the zero function defined over the universe domain | |||
1274 | * in the space of "pa". | |||
1275 | */ | |||
1276 | static __isl_give isl_pw_aff *set_universally_zero(__isl_take isl_pw_aff *pa) | |||
1277 | { | |||
1278 | isl_space *space; | |||
1279 | isl_aff *zero; | |||
1280 | ||||
1281 | space = isl_space_domain(isl_pw_aff_get_space(pa)); | |||
1282 | isl_pw_aff_free(pa); | |||
1283 | zero = isl_aff_zero_on_domain(isl_local_space_from_space(space)); | |||
1284 | ||||
1285 | return isl_pw_aff_from_aff(zero); | |||
1286 | } | |||
1287 | ||||
1288 | /* The sizes of the arrays on the host that have been computed by | |||
1289 | * extract_array_info may depend on the parameters. Use the extra | |||
1290 | * constraints on the parameters that are valid at "host_domain" | |||
1291 | * to simplify these expressions and store the results in kernel->array. | |||
1292 | * | |||
1293 | * We only need these localized bounds for arrays that are accessed | |||
1294 | * by the current kernel. If we have found at least one reference group | |||
1295 | * then the array is accessed by the kernel. | |||
1296 | * | |||
1297 | * The resulting sizes may be functions that are nowhere defined | |||
1298 | * in case the access function cannot possibly access anything inside | |||
1299 | * the kernel for some reason. If so, they are replaced by the zero | |||
1300 | * function. Since the access function cannot actually access anything, | |||
1301 | * there is no harm in printing the array sizes as zero. | |||
1302 | */ | |||
1303 | static void localize_bounds(struct ppcg_kernel *kernel, | |||
1304 | __isl_keep isl_set *host_domain) | |||
1305 | { | |||
1306 | int i, j; | |||
1307 | isl_set *context; | |||
1308 | ||||
1309 | context = isl_set_copy(host_domain); | |||
1310 | context = isl_set_params(context); | |||
1311 | ||||
1312 | for (i = 0; i < kernel->n_array; ++i) { | |||
1313 | struct gpu_local_array_info *local = &kernel->array[i]; | |||
1314 | isl_multi_pw_aff *bound; | |||
1315 | int n_index; | |||
1316 | ||||
1317 | if (local->n_group == 0) | |||
1318 | continue; | |||
1319 | ||||
1320 | n_index = local->array->n_index; | |||
1321 | bound = isl_multi_pw_aff_copy(local->array->bound); | |||
1322 | ||||
1323 | for (j = 0; j < n_index; ++j) { | |||
1324 | isl_pw_aff *pwaff; | |||
1325 | int empty; | |||
1326 | ||||
1327 | pwaff = isl_multi_pw_aff_get_pw_aff(bound, j); | |||
1328 | pwaff = isl_pw_aff_gist(pwaff, isl_set_copy(context)); | |||
1329 | empty = isl_pw_aff_is_empty(pwaff); | |||
1330 | if (empty < 0) | |||
1331 | pwaff = isl_pw_aff_free(pwaff); | |||
1332 | else if (empty) | |||
1333 | pwaff = set_universally_zero(pwaff); | |||
1334 | bound = isl_multi_pw_aff_set_pw_aff(bound, j, pwaff); | |||
1335 | } | |||
1336 | ||||
1337 | local->n_index = n_index; | |||
1338 | local->bound = bound; | |||
1339 | } | |||
1340 | isl_set_free(context); | |||
1341 | } | |||
1342 | ||||
1343 | /* Create the array of gpu_local_array_info structures "array" | |||
1344 | * inside "kernel". The number of elements in this array is | |||
1345 | * the same as the number of arrays in "prog". | |||
1346 | * Initialize the "array" field of each local array to point | |||
1347 | * to the corresponding array in "prog". | |||
1348 | */ | |||
1349 | static struct ppcg_kernel *ppcg_kernel_create_local_arrays( | |||
1350 | struct ppcg_kernel *kernel, struct gpu_prog *prog) | |||
1351 | { | |||
1352 | int i; | |||
1353 | isl_ctx *ctx; | |||
1354 | ||||
1355 | ctx = isl_set_get_ctx(prog->context); | |||
1356 | kernel->array = isl_calloc_array(ctx,((struct gpu_local_array_info *)isl_calloc_or_die(ctx, prog-> n_array, sizeof(struct gpu_local_array_info))) | |||
1357 | struct gpu_local_array_info, prog->n_array)((struct gpu_local_array_info *)isl_calloc_or_die(ctx, prog-> n_array, sizeof(struct gpu_local_array_info))); | |||
1358 | if (!kernel->array) | |||
1359 | return ppcg_kernel_free(kernel); | |||
1360 | kernel->n_array = prog->n_array; | |||
1361 | ||||
1362 | for (i = 0; i < prog->n_array; ++i) | |||
1363 | kernel->array[i].array = &prog->array[i]; | |||
1364 | ||||
1365 | return kernel; | |||
1366 | } | |||
1367 | ||||
1368 | /* Does "kernel" need to be passed an argument corresponding to array "i"? | |||
1369 | * | |||
1370 | * The argument is only needed if the kernel accesses this device memory. | |||
1371 | */ | |||
1372 | int ppcg_kernel_requires_array_argument(struct ppcg_kernel *kernel, int i) | |||
1373 | { | |||
1374 | return kernel->array[i].global; | |||
1375 | } | |||
1376 | ||||
1377 | /* Find the element in gen->stmt that has the given "id". | |||
1378 | * Return NULL if no such gpu_stmt can be found. | |||
1379 | */ | |||
1380 | static struct gpu_stmt *find_stmt(struct gpu_prog *prog, __isl_keep isl_id *id) | |||
1381 | { | |||
1382 | int i; | |||
1383 | ||||
1384 | for (i = 0; i < prog->n_stmts; ++i) { | |||
1385 | if (id == prog->stmts[i].id) | |||
1386 | break; | |||
1387 | } | |||
1388 | ||||
1389 | return i < prog->n_stmts ? &prog->stmts[i] : NULL((void*)0); | |||
1390 | } | |||
1391 | ||||
1392 | void ppcg_kernel_stmt_free(void *user) | |||
1393 | { | |||
1394 | struct ppcg_kernel_stmt *stmt = user; | |||
1395 | ||||
1396 | if (!stmt) | |||
1397 | return; | |||
1398 | ||||
1399 | switch (stmt->type) { | |||
1400 | case ppcg_kernel_copy: | |||
1401 | isl_ast_expr_free(stmt->u.c.index); | |||
1402 | isl_ast_expr_free(stmt->u.c.local_index); | |||
1403 | break; | |||
1404 | case ppcg_kernel_domain: | |||
1405 | isl_id_to_ast_expr_free(stmt->u.d.ref2expr); | |||
1406 | break; | |||
1407 | case ppcg_kernel_sync: | |||
1408 | break; | |||
1409 | } | |||
1410 | ||||
1411 | free(stmt); | |||
1412 | } | |||
1413 | ||||
1414 | /* Return the gpu_stmt_access in the list "accesses" that corresponds | |||
1415 | * to "ref_id". | |||
1416 | */ | |||
1417 | static struct gpu_stmt_access *find_access(struct gpu_stmt_access *accesses, | |||
1418 | __isl_keep isl_id *ref_id) | |||
1419 | { | |||
1420 | struct gpu_stmt_access *access; | |||
1421 | ||||
1422 | for (access = accesses; access; access = access->next) | |||
1423 | if (access->ref_id == ref_id) | |||
1424 | return access; | |||
1425 | ||||
1426 | return NULL((void*)0); | |||
1427 | } | |||
1428 | ||||
1429 | /* Return the index of the array called "name" in the list of arrays. | |||
1430 | */ | |||
1431 | static int find_array_index(struct ppcg_kernel *kernel, const char *name) | |||
1432 | { | |||
1433 | int i; | |||
1434 | ||||
1435 | for (i = 0; i < kernel->n_array; ++i) | |||
1436 | if (!strcmp(name, kernel->array[i].array->name)) | |||
1437 | return i; | |||
1438 | ||||
1439 | return -1; | |||
1440 | } | |||
1441 | ||||
1442 | /* Internal data structure for the index and AST expression transformation | |||
1443 | * callbacks for pet_stmt_build_ast_exprs. | |||
1444 | * | |||
1445 | * "kernel" is the kernel for which are computing AST expressions and | |||
1446 | * may be NULL if we are not inside a kernel. | |||
1447 | * "accesses" is the list of gpu_stmt_access in the statement. | |||
1448 | * "iterator_map" expresses the statement iterators in terms of | |||
1449 | * the AST loop iterators. | |||
1450 | * "sched2copy" expresses the outer copy_schedule_dim dimensions of | |||
1451 | * the kernel schedule in terms of the AST loop iterators and | |||
1452 | * may be NULL if we are not inside a kernel. | |||
1453 | * | |||
1454 | * The following fields are set in transform_index and used in transform_expr. | |||
1455 | * "array" is the array that is being accessed. | |||
1456 | * "global" is set if the global array is accessed (rather than | |||
1457 | * shared/private memory). | |||
1458 | * "local_array" refers to information on the array specialized | |||
1459 | * to the current kernel. | |||
1460 | */ | |||
1461 | struct ppcg_transform_data { | |||
1462 | struct ppcg_options *options; | |||
1463 | struct ppcg_kernel *kernel; | |||
1464 | struct gpu_stmt_access *accesses; | |||
1465 | isl_pw_multi_aff *iterator_map; | |||
1466 | isl_pw_multi_aff *sched2copy; | |||
1467 | ||||
1468 | struct gpu_array_info *array; | |||
1469 | int global; | |||
1470 | struct gpu_local_array_info *local_array; | |||
1471 | }; | |||
1472 | ||||
1473 | /* Return a pointer to the gpu_array_ref_group in "local" | |||
1474 | * that contains the reference "access". | |||
1475 | * Return NULL if no such group can be found. | |||
1476 | */ | |||
1477 | static struct gpu_array_ref_group *find_ref_group( | |||
1478 | struct gpu_local_array_info *local, struct gpu_stmt_access *access) | |||
1479 | { | |||
1480 | int i, j; | |||
1481 | ||||
1482 | for (i = 0; i < local->n_group; ++i) { | |||
1483 | struct gpu_array_ref_group *group = local->groups[i]; | |||
1484 | ||||
1485 | for (j = 0; j < group->n_ref; ++j) | |||
1486 | if (group->refs[j] == access) | |||
1487 | return group; | |||
1488 | } | |||
1489 | ||||
1490 | return NULL((void*)0); | |||
1491 | } | |||
1492 | ||||
1493 | /* Given an index expression "index" of the form | |||
1494 | * | |||
1495 | * L -> F(A), | |||
1496 | * | |||
1497 | * with F(A) either A or some subfield of A and L the AST loop iterators, | |||
1498 | * and a tiling "tiling" of the form | |||
1499 | * | |||
1500 | * [L -> A] -> T | |||
1501 | * | |||
1502 | * apply the tiling to the outer array in the index expression to obtain | |||
1503 | * | |||
1504 | * L -> T(A) | |||
1505 | * | |||
1506 | * If F(A) is some subfield of A, then separate the member access | |||
1507 | * into the base index expression and the field index expression, | |||
1508 | * apply the tiling to the base index expression and combine the result | |||
1509 | * with the field index expression. | |||
1510 | * | |||
1511 | * If F(A) is A, then modify index to keep track of the iterators | |||
1512 | * | |||
1513 | * L -> [L -> A] | |||
1514 | * | |||
1515 | * and combine the result with the tiling to obtain a tiled index expression | |||
1516 | * in terms of the AST loop iterators | |||
1517 | * | |||
1518 | * L -> T | |||
1519 | */ | |||
1520 | static __isl_give isl_multi_pw_aff *tile_outer( | |||
1521 | __isl_take isl_multi_pw_aff *index, __isl_take isl_multi_pw_aff *tiling) | |||
1522 | { | |||
1523 | isl_bool is_wrapping; | |||
1524 | isl_space *space; | |||
1525 | isl_multi_pw_aff *mpa; | |||
1526 | ||||
1527 | is_wrapping = isl_multi_pw_aff_range_is_wrapping(index); | |||
1528 | if (is_wrapping < 0) | |||
1529 | goto error; | |||
1530 | if (is_wrapping) { | |||
1531 | isl_multi_pw_aff *field; | |||
1532 | ||||
1533 | field = isl_multi_pw_aff_copy(index); | |||
1534 | field = isl_multi_pw_aff_range_factor_range(field); | |||
1535 | index = isl_multi_pw_aff_range_factor_domain(index); | |||
1536 | index = tile_outer(index, tiling); | |||
1537 | return isl_multi_pw_aff_range_product(index, field); | |||
1538 | } | |||
1539 | ||||
1540 | space = isl_space_domain(isl_multi_pw_aff_get_space(index)); | |||
1541 | space = isl_space_map_from_set(space); | |||
1542 | mpa = isl_multi_pw_aff_identity(space); | |||
1543 | index = isl_multi_pw_aff_range_product(mpa, index); | |||
1544 | index = isl_multi_pw_aff_pullback_multi_pw_aff(tiling, index); | |||
1545 | ||||
1546 | return index; | |||
1547 | error: | |||
1548 | isl_multi_pw_aff_free(index); | |||
1549 | isl_multi_pw_aff_free(tiling); | |||
1550 | return NULL((void*)0); | |||
1551 | } | |||
1552 | ||||
1553 | /* Index transformation callback for pet_stmt_build_ast_exprs. | |||
1554 | * | |||
1555 | * "index" expresses the array indices in terms of statement iterators | |||
1556 | * | |||
1557 | * We first reformulate "index" in terms of the AST loop iterators. | |||
1558 | * Then we check if we are accessing the global array or | |||
1559 | * a shared/private copy. In particular, if we are not inside a kernel | |||
1560 | * then we must be accessing a global array. | |||
1561 | * In the former case, we simply return | |||
1562 | * the updated index. If "index" is an affine expression rather | |||
1563 | * than an array access, then we also return the updated index here. | |||
1564 | * | |||
1565 | * If no reference groups have been computed for the array, | |||
1566 | * then we can only be accessing the global array. | |||
1567 | * | |||
1568 | * Otherwise, we apply the tiling to the index. | |||
1569 | * This tiling is of the form | |||
1570 | * | |||
1571 | * [D -> A] -> T | |||
1572 | * | |||
1573 | * where D corresponds to the outer tile->depth dimensions of | |||
1574 | * the kernel schedule. | |||
1575 | * The index is of the form | |||
1576 | * | |||
1577 | * L -> A | |||
1578 | * | |||
1579 | * We update the tiling to refer to the AST loop iterators | |||
1580 | * | |||
1581 | * [L -> A] -> T | |||
1582 | * | |||
1583 | * and combine it with the index to obtain a tiled index expression in terms | |||
1584 | * of the AST loop iterators | |||
1585 | * | |||
1586 | * L -> T | |||
1587 | * | |||
1588 | * Note that while the tiling applies directly to an outer array. | |||
1589 | * the index may refer to some subfield of this outer array. | |||
1590 | * In such cases, the result will refer to the same subfield of the tile. | |||
1591 | * That is, an index expression of the form L -> F(A) will be transformed | |||
1592 | * into an index expression of the form L -> F(T). | |||
1593 | */ | |||
1594 | static __isl_give isl_multi_pw_aff *transform_index( | |||
1595 | __isl_take isl_multi_pw_aff *index, __isl_keep isl_id *ref_id, | |||
1596 | void *user) | |||
1597 | { | |||
1598 | struct ppcg_transform_data *data = user; | |||
1599 | struct gpu_stmt_access *access; | |||
1600 | struct gpu_array_ref_group *group; | |||
1601 | struct gpu_array_tile *tile; | |||
1602 | isl_pw_multi_aff *iterator_map; | |||
1603 | int i; | |||
1604 | int dim; | |||
1605 | const char *name; | |||
1606 | isl_space *space; | |||
1607 | isl_multi_pw_aff *tiling; | |||
1608 | isl_pw_multi_aff *pma; | |||
1609 | isl_pw_multi_aff *sched2depth; | |||
1610 | ||||
1611 | data->array = NULL((void*)0); | |||
1612 | ||||
1613 | iterator_map = isl_pw_multi_aff_copy(data->iterator_map); | |||
1614 | index = isl_multi_pw_aff_pullback_pw_multi_aff(index, iterator_map); | |||
1615 | ||||
1616 | if (!data->kernel) | |||
1617 | return index; | |||
1618 | ||||
1619 | access = find_access(data->accesses, ref_id); | |||
1620 | if (!access) | |||
1621 | return index; | |||
1622 | if (!isl_map_has_tuple_name(access->access, isl_dim_out)) | |||
1623 | return index; | |||
1624 | ||||
1625 | name = get_outer_array_name(access->access); | |||
1626 | i = find_array_index(data->kernel, name); | |||
1627 | if (i < 0) | |||
1628 | isl_die(isl_multi_pw_aff_get_ctx(index), isl_error_internal,do { isl_handle_error(isl_multi_pw_aff_get_ctx(index), isl_error_internal , "cannot find array", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 1630); return isl_multi_pw_aff_free(index); } while (0) | |||
1629 | "cannot find array",do { isl_handle_error(isl_multi_pw_aff_get_ctx(index), isl_error_internal , "cannot find array", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 1630); return isl_multi_pw_aff_free(index); } while (0) | |||
1630 | return isl_multi_pw_aff_free(index))do { isl_handle_error(isl_multi_pw_aff_get_ctx(index), isl_error_internal , "cannot find array", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 1630); return isl_multi_pw_aff_free(index); } while (0); | |||
1631 | data->local_array = &data->kernel->array[i]; | |||
1632 | data->array = data->local_array->array; | |||
1633 | ||||
1634 | group = find_ref_group(data->local_array, access); | |||
1635 | if (!group) { | |||
1636 | data->global = 1; | |||
1637 | return index; | |||
1638 | } | |||
1639 | ||||
1640 | tile = gpu_array_ref_group_tile(group); | |||
1641 | data->global = !tile; | |||
1642 | if (!tile) | |||
1643 | return index; | |||
1644 | ||||
1645 | space = isl_space_domain(isl_multi_aff_get_space(tile->tiling)); | |||
1646 | space = isl_space_range(isl_space_unwrap(space)); | |||
1647 | space = isl_space_map_from_set(space); | |||
1648 | pma = isl_pw_multi_aff_identity(space); | |||
1649 | sched2depth = isl_pw_multi_aff_copy(data->sched2copy); | |||
1650 | dim = isl_pw_multi_aff_dim(sched2depth, isl_dim_out); | |||
1651 | sched2depth = isl_pw_multi_aff_drop_dims(sched2depth, isl_dim_out, | |||
1652 | tile->depth, dim - tile->depth); | |||
1653 | pma = isl_pw_multi_aff_product(sched2depth, pma); | |||
1654 | tiling = isl_multi_pw_aff_from_multi_aff( | |||
1655 | isl_multi_aff_copy(tile->tiling)); | |||
1656 | tiling = isl_multi_pw_aff_pullback_pw_multi_aff(tiling, pma); | |||
1657 | ||||
1658 | index = tile_outer(index, tiling); | |||
1659 | ||||
1660 | return index; | |||
1661 | } | |||
1662 | ||||
1663 | /* Dereference "expr" by adding an index [0]. | |||
1664 | * The original "expr" is assumed not to have any indices. | |||
1665 | * | |||
1666 | * If "expr" is a member access, then the dereferencing needs | |||
1667 | * to be applied to the structure argument of this member access. | |||
1668 | */ | |||
1669 | static __isl_give isl_ast_expr *dereference(__isl_take isl_ast_expr *expr) | |||
1670 | { | |||
1671 | isl_ctx *ctx; | |||
1672 | isl_ast_expr *arg0, *res; | |||
1673 | isl_ast_expr_list *list; | |||
1674 | ||||
1675 | arg0 = isl_ast_expr_get_op_arg(expr, 0); | |||
1676 | if (!arg0) | |||
1677 | return isl_ast_expr_free(expr); | |||
1678 | if (isl_ast_expr_get_type(arg0) == isl_ast_expr_op && | |||
1679 | isl_ast_expr_get_op_type(arg0) == isl_ast_op_member) { | |||
1680 | isl_ast_expr *arg; | |||
1681 | ||||
1682 | arg = isl_ast_expr_get_op_arg(arg0, 0); | |||
1683 | arg = dereference(arg); | |||
1684 | arg0 = isl_ast_expr_set_op_arg(arg0, 0, arg); | |||
1685 | expr = isl_ast_expr_set_op_arg(expr, 0, arg0); | |||
1686 | ||||
1687 | return expr; | |||
1688 | } | |||
1689 | isl_ast_expr_free(arg0); | |||
1690 | ||||
1691 | ctx = isl_ast_expr_get_ctx(expr); | |||
1692 | res = isl_ast_expr_from_val(isl_val_zero(ctx)); | |||
1693 | list = isl_ast_expr_list_from_ast_expr(res); | |||
1694 | res = isl_ast_expr_get_op_arg(expr, 0); | |||
1695 | res = isl_ast_expr_access(res, list); | |||
1696 | isl_ast_expr_free(expr); | |||
1697 | ||||
1698 | return res; | |||
1699 | } | |||
1700 | ||||
1701 | /* Linearize the index expression "expr" based on the array bounds | |||
1702 | * of "array". | |||
1703 | * | |||
1704 | * That is, transform expression | |||
1705 | * | |||
1706 | * A[i_0][i_1]...[i_n] | |||
1707 | * | |||
1708 | * to | |||
1709 | * | |||
1710 | * A[(..((i_0 * b_1 + i_1) ... ) * b_n + i_n] | |||
1711 | * | |||
1712 | * where b_0, b_1, ..., b_n are the bounds on the array. | |||
1713 | * | |||
1714 | * If the base of "expr" is a member access, then the linearization needs | |||
1715 | * to be applied to the structure argument of this member access. | |||
1716 | * | |||
1717 | * In the base case, if "expr" has no arguments (other than the name of | |||
1718 | * the array), then we are passing an entire array to a function. | |||
1719 | * In this case, there is nothing to linearize. | |||
1720 | * Note that at this point an expression with no arguments can | |||
1721 | * only be an entire array because the scalar case and | |||
1722 | * the case of single struct are handled by the caller. | |||
1723 | * | |||
1724 | * If the number of specified index expressions in "expr" | |||
1725 | * is smaller than the dimension of the accessed array, | |||
1726 | * then the missing i_j also do not appear in the linearized expression. | |||
1727 | * Furthermore, since such an expression does not refer to a single | |||
1728 | * element while the default linearized expression would refer to | |||
1729 | * a single element, we return the expression | |||
1730 | * | |||
1731 | * A + (..((i_0 * b_1 + i_1) ... ) * b_l + i_l) | |||
1732 | * | |||
1733 | * instead. Note that because of the special case handling above, | |||
1734 | * we can assume here that there is at least one index expression. | |||
1735 | */ | |||
1736 | __isl_give isl_ast_expr *gpu_local_array_info_linearize_index( | |||
1737 | struct gpu_local_array_info *array, __isl_take isl_ast_expr *expr) | |||
1738 | { | |||
1739 | int i, n; | |||
1740 | isl_ast_expr *arg0; | |||
1741 | isl_ast_expr *res; | |||
1742 | isl_ast_expr_list *list; | |||
1743 | ||||
1744 | arg0 = isl_ast_expr_get_op_arg(expr, 0); | |||
1745 | if (isl_ast_expr_get_type(arg0) == isl_ast_expr_op && | |||
1746 | isl_ast_expr_get_op_type(arg0) == isl_ast_op_member) { | |||
1747 | isl_ast_expr *arg; | |||
1748 | ||||
1749 | arg = isl_ast_expr_get_op_arg(arg0, 0); | |||
1750 | arg = gpu_local_array_info_linearize_index(array, arg); | |||
1751 | arg0 = isl_ast_expr_set_op_arg(arg0, 0, arg); | |||
1752 | expr = isl_ast_expr_set_op_arg(expr, 0, arg0); | |||
1753 | ||||
1754 | return expr; | |||
1755 | } | |||
1756 | isl_ast_expr_free(arg0); | |||
1757 | ||||
1758 | if (isl_ast_expr_get_op_n_arg(expr) == 1) | |||
1759 | return expr; | |||
1760 | ||||
1761 | n = isl_ast_expr_get_op_n_arg(expr); | |||
1762 | res = isl_ast_expr_get_op_arg(expr, 1); | |||
1763 | for (i = 1; i < array->n_index; ++i) { | |||
1764 | isl_ast_expr *expr_i; | |||
1765 | ||||
1766 | expr_i = isl_ast_expr_get_op_arg(array->bound_expr, 1 + i); | |||
1767 | res = isl_ast_expr_mul(res, expr_i); | |||
1768 | ||||
1769 | if (i + 1 >= n) | |||
1770 | continue; | |||
1771 | expr_i = isl_ast_expr_get_op_arg(expr, i + 1); | |||
1772 | res = isl_ast_expr_add(res, expr_i); | |||
1773 | } | |||
1774 | ||||
1775 | if (1 + array->n_index > n) { | |||
1776 | res = isl_ast_expr_add(isl_ast_expr_get_op_arg(expr, 0), res); | |||
1777 | } else { | |||
1778 | list = isl_ast_expr_list_from_ast_expr(res); | |||
1779 | res = isl_ast_expr_get_op_arg(expr, 0); | |||
1780 | res = isl_ast_expr_access(res, list); | |||
1781 | } | |||
1782 | ||||
1783 | isl_ast_expr_free(expr); | |||
1784 | ||||
1785 | return res; | |||
1786 | } | |||
1787 | ||||
1788 | /* AST expression transformation callback for pet_stmt_build_ast_exprs. | |||
1789 | * | |||
1790 | * If the AST expression refers to an array that is not accessed | |||
1791 | * at all, then this means the value of the expression is not used, | |||
1792 | * so we might as well print zero (NULL pointer) instead. | |||
1793 | * | |||
1794 | * If the AST expression refers to a global scalar that is not | |||
1795 | * a read-only scalar, then its address was passed to the kernel and | |||
1796 | * we need to dereference it. | |||
1797 | * | |||
1798 | * If the AST expression refers to an access to a global array, | |||
1799 | * then we linearize the access exploiting the bounds in data->local_array. | |||
1800 | */ | |||
1801 | static __isl_give isl_ast_expr *transform_expr(__isl_take isl_ast_expr *expr, | |||
1802 | __isl_keep isl_id *id, void *user) | |||
1803 | { | |||
1804 | struct ppcg_transform_data *data = user; | |||
1805 | ||||
1806 | if (!data->array) | |||
1807 | return expr; | |||
1808 | if (!data->array->accessed) { | |||
1809 | isl_ctx *ctx; | |||
1810 | ||||
1811 | ctx = isl_ast_expr_get_ctx(expr); | |||
1812 | isl_ast_expr_free(expr); | |||
1813 | return isl_ast_expr_from_val(isl_val_zero(ctx)); | |||
1814 | } | |||
1815 | if (gpu_array_is_read_only_scalar(data->array)) | |||
1816 | return expr; | |||
1817 | if (!data->global) | |||
1818 | return expr; | |||
1819 | if (data->array->n_index == 0) | |||
1820 | return dereference(expr); | |||
1821 | if (!data->array->linearize) | |||
1822 | return expr; | |||
1823 | ||||
1824 | return gpu_local_array_info_linearize_index(data->local_array, expr); | |||
1825 | } | |||
1826 | ||||
1827 | /* This function is called for each instance of a user statement | |||
1828 | * in the kernel "kernel", identified by "gpu_stmt". | |||
1829 | * "kernel" may be NULL if we are not inside a kernel. | |||
1830 | * | |||
1831 | * We attach a struct ppcg_kernel_stmt to the "node", containing | |||
1832 | * a computed AST expression for each access, through an annotation | |||
1833 | * with name "user". | |||
1834 | * These AST expressions are computed from iterator_map, | |||
1835 | * which expresses the domain | |||
1836 | * elements in terms of the generated loops, and sched2copy, | |||
1837 | * which expresses the outer copy_schedule_dim dimensions of | |||
1838 | * the kernel schedule computed by PPCG in terms of the generated loops. | |||
1839 | */ | |||
1840 | static __isl_give isl_ast_node *create_domain_leaf( | |||
1841 | struct ppcg_kernel *kernel, __isl_take isl_ast_node *node, | |||
1842 | __isl_keep isl_ast_build *build, struct gpu_stmt *gpu_stmt, | |||
1843 | struct gpu_gen *gen) | |||
1844 | { | |||
1845 | struct ppcg_transform_data data; | |||
1846 | struct ppcg_kernel_stmt *stmt; | |||
1847 | isl_ctx *ctx; | |||
1848 | isl_id *id; | |||
1849 | isl_pw_multi_aff *sched2copy; | |||
1850 | isl_map *map; | |||
1851 | isl_pw_multi_aff *iterator_map; | |||
1852 | isl_union_map *schedule; | |||
1853 | ||||
1854 | if (!node) | |||
1855 | return NULL((void*)0); | |||
1856 | ctx = isl_ast_node_get_ctx(node); | |||
1857 | ||||
1858 | stmt = isl_calloc_type(ctx, struct ppcg_kernel_stmt)((struct ppcg_kernel_stmt *)isl_calloc_or_die(ctx, 1, sizeof( struct ppcg_kernel_stmt))); | |||
1859 | if (!stmt) | |||
1860 | return isl_ast_node_free(node); | |||
1861 | ||||
1862 | schedule = isl_ast_build_get_schedule(build); | |||
1863 | map = isl_map_reverse(isl_map_from_union_map(schedule)); | |||
1864 | iterator_map = isl_pw_multi_aff_from_map(map); | |||
1865 | if (kernel) | |||
1866 | sched2copy = compute_sched_to_copy(kernel, | |||
1867 | isl_pw_multi_aff_copy(iterator_map)); | |||
1868 | else | |||
1869 | sched2copy = NULL((void*)0); | |||
1870 | ||||
1871 | stmt->type = ppcg_kernel_domain; | |||
1872 | stmt->u.d.stmt = gpu_stmt; | |||
1873 | ||||
1874 | data.kernel = kernel; | |||
1875 | data.accesses = stmt->u.d.stmt->accesses; | |||
1876 | data.iterator_map = iterator_map; | |||
1877 | data.sched2copy = sched2copy; | |||
1878 | stmt->u.d.ref2expr = gen->build_ast_expr(stmt->u.d.stmt->stmt, | |||
1879 | build, &transform_index, &data, | |||
1880 | &transform_expr, &data); | |||
1881 | ||||
1882 | isl_pw_multi_aff_free(iterator_map); | |||
1883 | isl_pw_multi_aff_free(sched2copy); | |||
1884 | ||||
1885 | id = isl_id_alloc(ctx, "user", stmt); | |||
1886 | id = isl_id_set_free_user(id, &ppcg_kernel_stmt_free); | |||
1887 | return isl_ast_node_set_annotation(node, id); | |||
1888 | } | |||
1889 | ||||
1890 | /* This function is called for each statement node in the AST | |||
1891 | * for copying to or from shared/private memory. | |||
1892 | * Attach a pointer to a ppcg_kernel_stmt representing the copy | |||
1893 | * statement to the node. | |||
1894 | * The statement name is "read" or "write", depending on whether we are | |||
1895 | * reading from global memory or writing to global memory. | |||
1896 | * | |||
1897 | * The schedule is of the form | |||
1898 | * | |||
1899 | * type[D -> A] -> L | |||
1900 | * | |||
1901 | * where D corresponds to the outer tile->depth dimensions of | |||
1902 | * the kernel schedule, A to the global array and L to the outer | |||
1903 | * generated AST schedule. | |||
1904 | * We compute the inverse and strip off the type, resulting in | |||
1905 | * | |||
1906 | * L -> [D -> A] | |||
1907 | * | |||
1908 | * We combine this mapping with on the one hand the projection | |||
1909 | * | |||
1910 | * [D -> A] -> A | |||
1911 | * | |||
1912 | * and on the other hand the group tiling | |||
1913 | * | |||
1914 | * [D -> A] -> T | |||
1915 | * | |||
1916 | * resulting in | |||
1917 | * | |||
1918 | * L -> A and L -> T | |||
1919 | * | |||
1920 | * and store the corresponding expressions in stmt->index and stmt->local_index, | |||
1921 | * where stmt points to the ppcg_kernel_stmt that is attached to the node. | |||
1922 | * stmt->index is linearized if the global memory array is linearized. | |||
1923 | */ | |||
1924 | static __isl_give isl_ast_node *create_access_leaf(struct ppcg_kernel *kernel, | |||
1925 | struct gpu_array_ref_group *group, __isl_take isl_ast_node *node, | |||
1926 | __isl_keep isl_ast_build *build) | |||
1927 | { | |||
1928 | struct ppcg_kernel_stmt *stmt; | |||
1929 | struct gpu_array_tile *tile; | |||
1930 | isl_id *id; | |||
1931 | isl_ast_expr *expr; | |||
1932 | isl_space *space; | |||
1933 | isl_map *access; | |||
1934 | isl_pw_multi_aff *pma, *pma2; | |||
1935 | const char *type; | |||
1936 | ||||
1937 | stmt = isl_calloc_type(kernel->ctx, struct ppcg_kernel_stmt)((struct ppcg_kernel_stmt *)isl_calloc_or_die(kernel->ctx, 1, sizeof(struct ppcg_kernel_stmt))); | |||
1938 | if (!stmt) | |||
1939 | return isl_ast_node_free(node); | |||
1940 | ||||
1941 | access = isl_map_from_union_map(isl_ast_build_get_schedule(build)); | |||
1942 | type = isl_map_get_tuple_name(access, isl_dim_in); | |||
1943 | stmt->u.c.read = !strcmp(type, "read"); | |||
1944 | access = isl_map_reverse(access); | |||
1945 | pma = isl_pw_multi_aff_from_map(access); | |||
1946 | pma = isl_pw_multi_aff_reset_tuple_id(pma, isl_dim_out); | |||
1947 | ||||
1948 | space = isl_space_range(isl_pw_multi_aff_get_space(pma)); | |||
1949 | space = isl_space_unwrap(space); | |||
1950 | pma2 = isl_pw_multi_aff_range_map(space); | |||
1951 | pma2 = isl_pw_multi_aff_pullback_pw_multi_aff(pma2, | |||
1952 | isl_pw_multi_aff_copy(pma)); | |||
1953 | expr = isl_ast_build_access_from_pw_multi_aff(build, pma2); | |||
1954 | if (group->array->linearize) | |||
1955 | expr = gpu_local_array_info_linearize_index(group->local_array, | |||
1956 | expr); | |||
1957 | stmt->u.c.index = expr; | |||
1958 | ||||
1959 | tile = gpu_array_ref_group_tile(group); | |||
1960 | pma2 = isl_pw_multi_aff_from_multi_aff( | |||
1961 | isl_multi_aff_copy(tile->tiling)); | |||
1962 | pma2 = isl_pw_multi_aff_pullback_pw_multi_aff(pma2, pma); | |||
1963 | expr = isl_ast_build_access_from_pw_multi_aff(build, pma2); | |||
1964 | stmt->u.c.local_index = expr; | |||
1965 | ||||
1966 | stmt->u.c.array = group->array; | |||
1967 | stmt->u.c.local_array = group->local_array; | |||
1968 | stmt->type = ppcg_kernel_copy; | |||
1969 | ||||
1970 | id = isl_id_alloc(kernel->ctx, "copy", stmt); | |||
1971 | id = isl_id_set_free_user(id, &ppcg_kernel_stmt_free); | |||
1972 | return isl_ast_node_set_annotation(node, id); | |||
1973 | } | |||
1974 | ||||
1975 | /* Create a synchronization ppcg_kernel_stmt and | |||
1976 | * attach it to the node "node" representing the synchronization. | |||
1977 | */ | |||
1978 | static __isl_give isl_ast_node *create_sync_leaf( | |||
1979 | struct ppcg_kernel *kernel, __isl_take isl_ast_node *node, | |||
1980 | __isl_keep isl_ast_build *build) | |||
1981 | { | |||
1982 | struct ppcg_kernel_stmt *stmt; | |||
1983 | isl_id *id; | |||
1984 | ||||
1985 | stmt = isl_calloc_type(kernel->ctx, struct ppcg_kernel_stmt)((struct ppcg_kernel_stmt *)isl_calloc_or_die(kernel->ctx, 1, sizeof(struct ppcg_kernel_stmt))); | |||
1986 | if (!stmt) | |||
1987 | return isl_ast_node_free(node); | |||
1988 | ||||
1989 | stmt->type = ppcg_kernel_sync; | |||
1990 | id = isl_id_alloc(kernel->ctx, "sync", stmt); | |||
1991 | id = isl_id_set_free_user(id, &ppcg_kernel_stmt_free); | |||
1992 | return isl_ast_node_set_annotation(node, id); | |||
1993 | } | |||
1994 | ||||
1995 | /* Build AST expressions for the device array sizes of all arrays in "prog" | |||
1996 | * that require allocation on the device using "build", as well as | |||
1997 | * for the original array sizes of all arrays that need to be declared | |||
1998 | * on the host. | |||
1999 | * "node" is freed in case of error. | |||
2000 | */ | |||
2001 | static __isl_give isl_ast_node *build_array_bounds( | |||
2002 | __isl_take isl_ast_node *node, struct gpu_prog *prog, | |||
2003 | __isl_keep isl_ast_build *build) | |||
2004 | { | |||
2005 | int i; | |||
2006 | ||||
2007 | for (i = 0; i < prog->n_array; ++i) { | |||
2008 | struct gpu_array_info *array = &prog->array[i]; | |||
2009 | isl_multi_pw_aff *size; | |||
2010 | isl_ast_expr *expr; | |||
2011 | ||||
2012 | if (!gpu_array_requires_device_allocation(array)) | |||
2013 | continue; | |||
2014 | ||||
2015 | size = isl_multi_pw_aff_copy(array->bound); | |||
2016 | expr = ppcg_build_size_expr(size, build); | |||
2017 | array->bound_expr = expr; | |||
2018 | if (!expr) | |||
2019 | return isl_ast_node_free(node); | |||
2020 | } | |||
2021 | ||||
2022 | for (i = 0; i < prog->n_array; ++i) { | |||
2023 | struct gpu_array_info *array = &prog->array[i]; | |||
2024 | isl_set *extent; | |||
2025 | isl_multi_pw_aff *size; | |||
2026 | isl_ast_expr *expr; | |||
2027 | ||||
2028 | if (!array->declare_local) | |||
2029 | continue; | |||
2030 | extent = isl_set_copy(array->declared_extent); | |||
2031 | size = ppcg_size_from_extent(extent); | |||
2032 | expr = ppcg_build_size_expr(size, build); | |||
2033 | array->declared_size = expr; | |||
2034 | if (!expr) | |||
2035 | return isl_ast_node_free(node); | |||
2036 | } | |||
2037 | ||||
2038 | return node; | |||
2039 | } | |||
2040 | ||||
2041 | /* Internal data structure for at_domain. | |||
2042 | * | |||
2043 | * "prog" represents the entire scop. | |||
2044 | * "kernel" points to the kernel to which the current schedule node | |||
2045 | * belongs. It is set by before_mark and reset by after_mark. | |||
2046 | * It may be NULL if we are outside any kernel. | |||
2047 | */ | |||
2048 | struct ppcg_at_domain_data { | |||
2049 | struct gpu_prog *prog; | |||
2050 | struct gpu_gen *gen; | |||
2051 | struct ppcg_kernel *kernel; | |||
2052 | }; | |||
2053 | ||||
2054 | /* This function is called for each instance of a user statement | |||
2055 | * in the kernel. This may be one of the original user statements | |||
2056 | * or a statement introduced by PPCG. | |||
2057 | * | |||
2058 | * We first check if the statement id corresponds to a gpu statement, | |||
2059 | * which indicates the statement is an original user statement. Any statement | |||
2060 | * that is not an original user statement has been introduced by PPCG and | |||
2061 | * requires special handling. | |||
2062 | * | |||
2063 | * If the user statement is one of the original user statements, then we call | |||
2064 | * create_domain_leaf. If it is "init_device", then we call | |||
2065 | * build_array_bounds. Otherwise, we check if it is a copy or synchronization | |||
2066 | * statement and call the appropriate functions. Statements that copy an array | |||
2067 | * to/from the device do not need any further treatment. | |||
2068 | * Neither does "clear_device". | |||
2069 | */ | |||
2070 | static __isl_give isl_ast_node *at_domain(__isl_take isl_ast_node *node, | |||
2071 | __isl_keep isl_ast_build *build, void *user) | |||
2072 | { | |||
2073 | struct ppcg_at_domain_data *data = user; | |||
2074 | struct gpu_stmt *gpu_stmt; | |||
2075 | isl_ast_expr *expr, *arg; | |||
2076 | isl_id *id; | |||
2077 | int is_sync; | |||
2078 | const char *name; | |||
2079 | void *p; | |||
2080 | ||||
2081 | expr = isl_ast_node_user_get_expr(node); | |||
2082 | arg = isl_ast_expr_get_op_arg(expr, 0); | |||
2083 | id = isl_ast_expr_get_id(arg); | |||
2084 | name = isl_id_get_name(id); | |||
2085 | p = isl_id_get_user(id); | |||
2086 | isl_ast_expr_free(expr); | |||
2087 | isl_ast_expr_free(arg); | |||
2088 | ||||
2089 | gpu_stmt = find_stmt(data->prog, id); | |||
2090 | is_sync = gpu_tree_id_is_sync(id, data->kernel); | |||
2091 | isl_id_free(id); | |||
2092 | ||||
2093 | if (gpu_stmt) | |||
2094 | return create_domain_leaf(data->kernel, node, build, gpu_stmt, | |||
2095 | data->gen); | |||
2096 | ||||
2097 | if (!prefixcmp(name, "to_device_") || !prefixcmp(name, "from_device_")) | |||
2098 | return node; | |||
2099 | if (!strcmp(name, "init_device")) | |||
2100 | return build_array_bounds(node, data->prog, build); | |||
2101 | if (!strcmp(name, "clear_device")) | |||
2102 | return node; | |||
2103 | if (is_sync < 0) | |||
2104 | return isl_ast_node_free(node); | |||
2105 | if (!strcmp(name, "read") || !strcmp(name, "write")) { | |||
2106 | struct gpu_array_ref_group *group = p; | |||
2107 | return create_access_leaf(data->kernel, group, node, build); | |||
2108 | } | |||
2109 | if (!is_sync) | |||
2110 | isl_die(data->prog->ctx, isl_error_internal,do { isl_handle_error(data->prog->ctx, isl_error_internal , "unknown statement type", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 2112); return isl_ast_node_free(node); } while (0) | |||
2111 | "unknown statement type",do { isl_handle_error(data->prog->ctx, isl_error_internal , "unknown statement type", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 2112); return isl_ast_node_free(node); } while (0) | |||
2112 | return isl_ast_node_free(node))do { isl_handle_error(data->prog->ctx, isl_error_internal , "unknown statement type", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 2112); return isl_ast_node_free(node); } while (0); | |||
2113 | return create_sync_leaf(data->kernel, node, build); | |||
2114 | } | |||
2115 | ||||
2116 | /* Given a set of wrapped references "ref", return the corresponding | |||
2117 | * access relations based on the tagged access relations "tagged". | |||
2118 | * | |||
2119 | * The elements of "ref" are of the form | |||
2120 | * | |||
2121 | * [D -> R] | |||
2122 | * | |||
2123 | * with D an iteration domains and R a reference. | |||
2124 | * The elements of "tagged" are of the form | |||
2125 | * | |||
2126 | * [D -> R] -> A | |||
2127 | * | |||
2128 | * with A an array. | |||
2129 | * | |||
2130 | * Extend "tagged" to include the iteration domain in the range, i.e., | |||
2131 | * | |||
2132 | * [D -> R] -> [D -> A] | |||
2133 | * | |||
2134 | * apply the result to "ref" and then unwrap the resulting set | |||
2135 | * to obtain relations of the form | |||
2136 | * | |||
2137 | * D -> A | |||
2138 | */ | |||
2139 | static __isl_give isl_union_map *wrapped_reference_to_access( | |||
2140 | __isl_take isl_union_set *ref, __isl_take isl_union_map *tagged) | |||
2141 | { | |||
2142 | isl_union_map *tag2access; | |||
2143 | ||||
2144 | tag2access = isl_union_map_copy(tagged); | |||
2145 | tag2access = isl_union_map_universe(tag2access); | |||
2146 | tag2access = isl_union_set_unwrap(isl_union_map_domain(tag2access)); | |||
2147 | tag2access = isl_union_map_domain_map(tag2access); | |||
2148 | tag2access = isl_union_map_range_product(tag2access, tagged); | |||
2149 | ||||
2150 | ref = isl_union_set_coalesce(ref); | |||
2151 | ref = isl_union_set_apply(ref, tag2access); | |||
2152 | ||||
2153 | return isl_union_set_unwrap(ref); | |||
2154 | } | |||
2155 | ||||
2156 | /* Given an access relation "access" from one or more array reference groups, | |||
2157 | * remove those reads if ("read" is 1) or writes (if "read" is 0) | |||
2158 | * that are only needed to communicate data within | |||
2159 | * the same iteration of "sched". | |||
2160 | * The domain of "sched" corresponds to the original statement instances, | |||
2161 | * i.e., those that appear in the domains of the access relations. | |||
2162 | * "tagged" contains all tagged access relations to all | |||
2163 | * the array reference groups accessed by "access" from statement | |||
2164 | * instances scheduled by "sched". | |||
2165 | * | |||
2166 | * If the access is a read then it is either an element of | |||
2167 | * | |||
2168 | * live_in union (range flow) | |||
2169 | * | |||
2170 | * where live_in and flow may be overapproximations, or | |||
2171 | * it reads an uninitialized value (that is not live-in because | |||
2172 | * there is an intermediate kill) or it reads a value that was | |||
2173 | * written within the same (compound) statement instance. | |||
2174 | * If the access is a write then it is either an element of | |||
2175 | * | |||
2176 | * live_out union (domain flow) | |||
2177 | * | |||
2178 | * or it writes a value that is never read (and is not live-out | |||
2179 | * because of an intermediate kill) or only | |||
2180 | * within the same (compound) statement instance. | |||
2181 | * In both cases, the access relation is also a subset of | |||
2182 | * the group access relation. | |||
2183 | * | |||
2184 | * The cases where an uninitialized value is read or a value is written | |||
2185 | * that is never read or where the dataflow occurs within a statement | |||
2186 | * instance are also considered local and may also be removed. | |||
2187 | * | |||
2188 | * Essentially, we compute the intersection of "access" with either | |||
2189 | * | |||
2190 | * live_in union (range non-local-flow) | |||
2191 | * | |||
2192 | * or | |||
2193 | * | |||
2194 | * live_out union (domain non-local-flow) | |||
2195 | * | |||
2196 | * We first construct a relation "local" | |||
2197 | * | |||
2198 | * [[D -> R] -> [D' -> R']] | |||
2199 | * | |||
2200 | * of pairs of domain iterations accessing the reference group | |||
2201 | * and references in the group that are coscheduled by "sched". | |||
2202 | * | |||
2203 | * If this relation does not intersect the dataflow dependences, | |||
2204 | * then there is nothing we can possibly remove, unless the dataflow | |||
2205 | * dependences themselves only relate a subset of the accesses. | |||
2206 | * In particular, the accesses may not be involved in any dataflow | |||
2207 | * dependences, either because they are uninitialized reads/dead writes | |||
2208 | * or because the dataflow occurs inside a statement instance. | |||
2209 | * | |||
2210 | * Since the computation below may break up the access relation | |||
2211 | * into smaller pieces, we only perform the intersection with | |||
2212 | * the non-local dependent accesses if the local pairs | |||
2213 | * intersect the dataflow dependences. Otherwise, we intersect | |||
2214 | * with the universe of the non-local dependent accesses. | |||
2215 | * This should at least remove accesses from statements that | |||
2216 | * do not participate in any dependences. | |||
2217 | * | |||
2218 | * In particular, we remove the "local" dataflow dependences from | |||
2219 | * the set of all dataflow dependences, or at least those | |||
2220 | * that may contribute to a domain/range that intersects | |||
2221 | * the domain of "access". | |||
2222 | * Note that if the potential dataflow dependences are an overapproximation | |||
2223 | * of the actual dataflow dependences, then the result remains an | |||
2224 | * overapproximation of the non-local dataflow dependences. | |||
2225 | * Copying to/from global memory is only needed for the references | |||
2226 | * in the domain/range of the result or for accesses that are live out/in | |||
2227 | * for the entire scop. | |||
2228 | * | |||
2229 | * We therefore map the domain/range of the "external" relation | |||
2230 | * to the corresponding access relation and take the union with | |||
2231 | * the live out/in relation. | |||
2232 | */ | |||
2233 | static __isl_give isl_union_map *remove_local_accesses( | |||
2234 | struct gpu_prog *prog, __isl_take isl_union_map *tagged, | |||
2235 | __isl_take isl_union_map *access, __isl_take isl_union_map *sched, | |||
2236 | int read) | |||
2237 | { | |||
2238 | int empty; | |||
2239 | isl_union_pw_multi_aff *tagger; | |||
2240 | isl_union_set *domain, *access_domain; | |||
2241 | isl_union_map *local, *external, *universe; | |||
2242 | isl_union_set *tag_set; | |||
2243 | ||||
2244 | if (isl_union_map_is_empty(access)) { | |||
2245 | isl_union_map_free(sched); | |||
2246 | isl_union_map_free(tagged); | |||
2247 | return access; | |||
2248 | } | |||
2249 | ||||
2250 | tagger = isl_union_pw_multi_aff_copy(prog->scop->tagger); | |||
2251 | domain = isl_union_map_domain(isl_union_map_copy(tagged)); | |||
2252 | tagger = isl_union_pw_multi_aff_intersect_domain(tagger, | |||
2253 | isl_union_set_copy(domain)); | |||
2254 | sched = isl_union_map_preimage_domain_union_pw_multi_aff(sched, tagger); | |||
2255 | ||||
2256 | local = isl_union_map_apply_range(sched, | |||
2257 | isl_union_map_reverse(isl_union_map_copy(sched))); | |||
2258 | local = isl_union_map_intersect(local, | |||
2259 | isl_union_map_copy(prog->scop->tagged_dep_flow)); | |||
2260 | ||||
2261 | empty = isl_union_map_is_empty(local); | |||
2262 | ||||
2263 | external = isl_union_map_copy(prog->scop->tagged_dep_flow); | |||
2264 | universe = isl_union_map_universe(isl_union_map_copy(access)); | |||
2265 | access_domain = isl_union_map_domain(universe); | |||
2266 | domain = isl_union_set_universe(domain); | |||
2267 | universe = isl_union_set_unwrap(domain); | |||
2268 | universe = isl_union_map_intersect_domain(universe, access_domain); | |||
2269 | domain = isl_union_map_wrap(universe); | |||
2270 | if (read) | |||
2271 | external = isl_union_map_intersect_range(external, domain); | |||
2272 | else | |||
2273 | external = isl_union_map_intersect_domain(external, domain); | |||
2274 | external = isl_union_map_intersect_params(external, | |||
2275 | isl_set_copy(prog->scop->context)); | |||
2276 | external = isl_union_map_subtract(external, local); | |||
2277 | ||||
2278 | if (read) { | |||
2279 | tag_set = isl_union_map_range(external); | |||
2280 | external = wrapped_reference_to_access(tag_set, tagged); | |||
2281 | external = isl_union_map_union(external, | |||
2282 | isl_union_map_copy(prog->scop->live_in)); | |||
2283 | } else { | |||
2284 | tag_set = isl_union_map_domain(external); | |||
2285 | external = wrapped_reference_to_access(tag_set, tagged); | |||
2286 | external = isl_union_map_union(external, | |||
2287 | isl_union_map_copy(prog->scop->live_out)); | |||
2288 | } | |||
2289 | ||||
2290 | if (empty < 0) | |||
2291 | external = isl_union_map_free(external); | |||
2292 | else if (empty) | |||
2293 | external = isl_union_map_universe(external); | |||
2294 | ||||
2295 | access = isl_union_map_intersect(access, external); | |||
2296 | ||||
2297 | return access; | |||
2298 | } | |||
2299 | ||||
2300 | /* Given an access relation "access" from "group", remove those reads | |||
2301 | * if ("read" is 1) or writes (if "read" is 0) that are only needed to | |||
2302 | * communicate data within the same iteration of the schedule "prefix" | |||
2303 | * at the position where the copying of the group is inserted. | |||
2304 | * That is, the output dimension of "prefix" | |||
2305 | * is equal to tile->depth. | |||
2306 | * The domain of "prefix" corresponds to the original statement instances, | |||
2307 | * i.e., those that appear in the domains of the access relations. | |||
2308 | * | |||
2309 | * Extract the tagged access relation of "group" and | |||
2310 | * then call remove_local_accesses. | |||
2311 | */ | |||
2312 | static __isl_give isl_union_map *remove_local_accesses_group( | |||
2313 | struct ppcg_kernel *kernel, struct gpu_array_ref_group *group, | |||
2314 | __isl_take isl_union_map *access, __isl_keep isl_union_map *prefix, | |||
2315 | int read) | |||
2316 | { | |||
2317 | isl_union_map *sched, *tagged; | |||
2318 | ||||
2319 | if (isl_union_map_is_empty(access)) | |||
2320 | return access; | |||
2321 | ||||
2322 | tagged = group_tagged_access_relation(group); | |||
2323 | sched = isl_union_map_copy(prefix); | |||
2324 | ||||
2325 | return remove_local_accesses(kernel->prog, tagged, access, sched, read); | |||
2326 | } | |||
2327 | ||||
2328 | /* Build an access AST expression for the effective grid size using "build". | |||
2329 | * Store the result in kernel->grid_size_expr. | |||
2330 | */ | |||
2331 | static isl_stat build_grid_size(struct ppcg_kernel *kernel, | |||
2332 | __isl_keep isl_ast_build *build) | |||
2333 | { | |||
2334 | isl_multi_pw_aff *size; | |||
2335 | ||||
2336 | size = isl_multi_pw_aff_copy(kernel->grid_size); | |||
2337 | size = isl_multi_pw_aff_set_tuple_name(size, isl_dim_out, "grid"); | |||
2338 | kernel->grid_size_expr = ppcg_build_size_expr(size, build); | |||
2339 | ||||
2340 | if (!kernel->grid_size_expr) | |||
2341 | return isl_stat_error; | |||
2342 | return isl_stat_ok; | |||
2343 | } | |||
2344 | ||||
2345 | /* Build access AST expressions for the localized array sizes using "build". | |||
2346 | * Store the result in local->bound_expr. | |||
2347 | * Only do this for arrays for which localized bounds have been computed. | |||
2348 | */ | |||
2349 | static isl_stat build_local_array_sizes(struct ppcg_kernel *kernel, | |||
2350 | __isl_keep isl_ast_build *build) | |||
2351 | { | |||
2352 | int i; | |||
2353 | ||||
2354 | for (i = 0; i < kernel->n_array; ++i) { | |||
2355 | struct gpu_local_array_info *local = &kernel->array[i]; | |||
2356 | isl_multi_pw_aff *size; | |||
2357 | ||||
2358 | if (local->n_group == 0) | |||
2359 | continue; | |||
2360 | size = isl_multi_pw_aff_copy(local->bound); | |||
2361 | local->bound_expr = ppcg_build_size_expr(size, build); | |||
2362 | if (!local->bound_expr) | |||
2363 | return isl_stat_error; | |||
2364 | } | |||
2365 | ||||
2366 | return isl_stat_ok; | |||
2367 | } | |||
2368 | ||||
2369 | /* Build access AST expressions for the effective grid size and | |||
2370 | * the localized array sizes using "build". | |||
2371 | */ | |||
2372 | static isl_stat build_grid_and_local_array_sizes(struct ppcg_kernel *kernel, | |||
2373 | __isl_keep isl_ast_build *build) | |||
2374 | { | |||
2375 | if (build_grid_size(kernel, build) < 0) | |||
2376 | return isl_stat_error; | |||
2377 | if (build_local_array_sizes(kernel, build) < 0) | |||
2378 | return isl_stat_error; | |||
2379 | return isl_stat_ok; | |||
2380 | } | |||
2381 | ||||
2382 | /* This function is called before the AST generator starts traversing | |||
2383 | * the schedule subtree of a node with mark "mark". | |||
2384 | * | |||
2385 | * If the mark is called "kernel", store the kernel pointer in data->kernel | |||
2386 | * for use in at_domain and build AST expressions for the grid size and | |||
2387 | * the localized array sizes. | |||
2388 | */ | |||
2389 | static isl_stat before_mark(__isl_keep isl_id *mark, | |||
2390 | __isl_keep isl_ast_build *build, void *user) | |||
2391 | { | |||
2392 | struct ppcg_at_domain_data *data = user; | |||
2393 | ||||
2394 | if (!mark) | |||
2395 | return isl_stat_error; | |||
2396 | if (!strcmp(isl_id_get_name(mark), "kernel")) { | |||
2397 | data->kernel = isl_id_get_user(mark); | |||
2398 | if (build_grid_and_local_array_sizes(data->kernel, build) < 0) | |||
2399 | return isl_stat_error; | |||
2400 | } | |||
2401 | return isl_stat_ok; | |||
2402 | } | |||
2403 | ||||
2404 | /* This function is called after the AST generator has finished traversing | |||
2405 | * the schedule subtree of a mark node. "node" points to the corresponding | |||
2406 | * mark AST node. | |||
2407 | * | |||
2408 | * If the mark is called "kernel", then replace "node" by a user node | |||
2409 | * that "calls" the kernel, representing the launch of the kernel. | |||
2410 | * The original "node" is stored inside the kernel object so that | |||
2411 | * it can be used to print the device code. | |||
2412 | * Note that this assumes that a kernel is only launched once. | |||
2413 | * Also clear data->kernel. | |||
2414 | */ | |||
2415 | static __isl_give isl_ast_node *after_mark(__isl_take isl_ast_node *node, | |||
2416 | __isl_keep isl_ast_build *build, void *user) | |||
2417 | { | |||
2418 | isl_ctx *ctx; | |||
2419 | isl_id *id; | |||
2420 | isl_ast_expr *expr; | |||
2421 | isl_ast_expr_list *list; | |||
2422 | struct ppcg_kernel *kernel; | |||
2423 | struct ppcg_at_domain_data *data = user; | |||
2424 | ||||
2425 | ctx = isl_ast_node_get_ctx(node); | |||
2426 | id = isl_ast_node_mark_get_id(node); | |||
2427 | if (!id) | |||
2428 | return isl_ast_node_free(node); | |||
2429 | if (strcmp(isl_id_get_name(id), "kernel") || !data->kernel) { | |||
2430 | isl_id_free(id); | |||
2431 | return node; | |||
2432 | } | |||
2433 | kernel = data->kernel; | |||
2434 | data->kernel = NULL((void*)0); | |||
2435 | kernel->space = isl_ast_build_get_schedule_space(build); | |||
2436 | kernel->tree = isl_ast_node_mark_get_node(node); | |||
2437 | isl_ast_node_free(node); | |||
2438 | ||||
2439 | expr = isl_ast_expr_from_id(isl_id_copy(id)); | |||
2440 | list = isl_ast_expr_list_alloc(ctx, 0); | |||
2441 | expr = isl_ast_expr_call(expr, list); | |||
2442 | node = isl_ast_node_alloc_user(expr); | |||
2443 | node = isl_ast_node_set_annotation(node, id); | |||
2444 | ||||
2445 | return node; | |||
2446 | } | |||
2447 | ||||
2448 | static isl_bool update_depth(__isl_keep isl_schedule_node *node, void *user) | |||
2449 | { | |||
2450 | int *depth = user; | |||
2451 | int node_depth; | |||
2452 | ||||
2453 | if (isl_schedule_node_get_type(node) != isl_schedule_node_leaf) | |||
2454 | return isl_bool_true; | |||
2455 | node_depth = isl_schedule_node_get_schedule_depth(node); | |||
2456 | if (node_depth > *depth) | |||
2457 | *depth = node_depth; | |||
2458 | ||||
2459 | return isl_bool_false; | |||
2460 | } | |||
2461 | ||||
2462 | /* Use isl to generate code for both the host and the device | |||
2463 | * from "schedule". | |||
2464 | * The device code is marked by "kernel" mark nodes in the schedule tree, | |||
2465 | * containing a pointer to a ppcg_kernel object. | |||
2466 | * The returned AST only contains the AST for the host code. | |||
2467 | * The ASTs for the device code are embedded in ppcg_kernel objects | |||
2468 | * attached to the leaf nodes that call "kernel". | |||
2469 | */ | |||
2470 | __isl_give isl_ast_node *generate_code(struct gpu_gen *gen, | |||
2471 | __isl_take isl_schedule *schedule) | |||
2472 | { | |||
2473 | struct ppcg_at_domain_data data; | |||
2474 | isl_ast_build *build; | |||
2475 | isl_ast_node *tree; | |||
2476 | isl_id_list *iterators; | |||
2477 | int depth; | |||
2478 | ||||
2479 | data.prog = gen->prog; | |||
2480 | data.gen = gen; | |||
2481 | data.kernel = NULL((void*)0); | |||
2482 | ||||
2483 | depth = 0; | |||
2484 | if (isl_schedule_foreach_schedule_node_top_down(schedule, &update_depth, | |||
2485 | &depth) < 0) | |||
2486 | return NULL((void*)0); | |||
2487 | build = isl_ast_build_alloc(gen->prog->ctx); | |||
2488 | iterators = ppcg_scop_generate_names(gen->prog->scop, depth, "c"); | |||
2489 | build = isl_ast_build_set_iterators(build, iterators); | |||
2490 | build = isl_ast_build_set_at_each_domain(build, &at_domain, &data); | |||
2491 | build = isl_ast_build_set_before_each_mark(build, &before_mark, &data); | |||
2492 | build = isl_ast_build_set_after_each_mark(build, &after_mark, &data); | |||
2493 | if (gen->prog->scop->options->debug->dump_final_schedule) | |||
2494 | isl_schedule_dump(schedule); | |||
2495 | tree = isl_ast_build_node_from_schedule(build, schedule); | |||
2496 | isl_ast_build_free(build); | |||
2497 | ||||
2498 | return tree; | |||
2499 | } | |||
2500 | ||||
2501 | __isl_give isl_union_map *extract_sizes_from_str(isl_ctx *ctx, const char *str) | |||
2502 | { | |||
2503 | if (!str) | |||
2504 | return NULL((void*)0); | |||
2505 | return isl_union_map_read_from_str(ctx, str); | |||
2506 | } | |||
2507 | ||||
2508 | /* Can "node" be tiled and then mapped to block and thread identifiers? | |||
2509 | * That is, is it permutable with at least one coincident dimension? | |||
2510 | */ | |||
2511 | static int is_permutable(__isl_keep isl_schedule_node *node) | |||
2512 | { | |||
2513 | if (!node) | |||
2514 | return -1; | |||
2515 | ||||
2516 | if (isl_schedule_node_get_type(node) != isl_schedule_node_band) | |||
2517 | return 0; | |||
2518 | if (!isl_schedule_node_band_get_permutable(node)) | |||
2519 | return 0; | |||
2520 | if (isl_schedule_node_band_n_member(node) < 1) | |||
2521 | return 0; | |||
2522 | if (!isl_schedule_node_band_member_get_coincident(node, 0)) | |||
2523 | return 0; | |||
2524 | ||||
2525 | return 1; | |||
2526 | } | |||
2527 | ||||
2528 | /* A isl_schedule_foreach_schedule_node_top_down callback | |||
2529 | * for setting *any_permutable and aborting the search | |||
2530 | * if "node" is a permutable band with coincident dimensions. | |||
2531 | * Otherwise, continue searching. | |||
2532 | */ | |||
2533 | static isl_bool set_permutable(__isl_keep isl_schedule_node *node, void *user) | |||
2534 | { | |||
2535 | int *any_permutable = user; | |||
2536 | int permutable; | |||
2537 | ||||
2538 | permutable = is_permutable(node); | |||
2539 | if (permutable < 0) | |||
2540 | return isl_bool_error; | |||
2541 | if (!permutable) | |||
2542 | return isl_bool_true; | |||
2543 | ||||
2544 | *any_permutable = 1; | |||
2545 | ||||
2546 | return isl_bool_error; | |||
2547 | } | |||
2548 | ||||
2549 | /* Does the subtree rooted at "node" have any suitably permutable band nodes? | |||
2550 | * That is, does it have any nodes that are permutable and that | |||
2551 | * have a least one coincident dimension? | |||
2552 | */ | |||
2553 | static int subtree_has_permutable_bands(__isl_keep isl_schedule_node *node) | |||
2554 | { | |||
2555 | int any_parallelism = 0; | |||
2556 | ||||
2557 | if (isl_schedule_node_foreach_descendant_top_down(node, &set_permutable, | |||
2558 | &any_parallelism) < 0 && | |||
2559 | !any_parallelism) | |||
2560 | return -1; | |||
2561 | ||||
2562 | return any_parallelism; | |||
2563 | } | |||
2564 | ||||
2565 | /* Does "schedule" contain any permutable band with at least one coincident | |||
2566 | * member? | |||
2567 | */ | |||
2568 | int has_any_permutable_node(__isl_keep isl_schedule *schedule) | |||
2569 | { | |||
2570 | isl_schedule_node *root; | |||
2571 | int any_permutable; | |||
2572 | ||||
2573 | root = isl_schedule_get_root(schedule); | |||
2574 | any_permutable = subtree_has_permutable_bands(root); | |||
2575 | isl_schedule_node_free(root); | |||
2576 | ||||
2577 | return any_permutable; | |||
2578 | } | |||
2579 | ||||
2580 | /* Is "node" a candidate for mapping to block and thread identifiers? | |||
2581 | * In particular, is it permutable with at least one coincident dimension? | |||
2582 | * Alternatively, does the subtree rooted at "node" not contain | |||
2583 | * any such permutable node? Filter nodes are skipped in this case, | |||
2584 | * because a band node will be inserted in front of the returned | |||
2585 | * node and this is not possible for filter nodes that are children | |||
2586 | * of set or sequence nodes. | |||
2587 | */ | |||
2588 | static int is_candidate(__isl_keep isl_schedule_node *node) | |||
2589 | { | |||
2590 | int permutable; | |||
2591 | ||||
2592 | if (isl_schedule_node_get_type(node) == isl_schedule_node_leaf) | |||
2593 | return 1; | |||
2594 | permutable = is_permutable(node); | |||
2595 | if (permutable < 0 || permutable) | |||
2596 | return permutable; | |||
2597 | if (isl_schedule_node_get_type(node) == isl_schedule_node_filter) | |||
2598 | return 0; | |||
2599 | permutable = subtree_has_permutable_bands(node); | |||
2600 | if (permutable < 0) | |||
2601 | return -1; | |||
2602 | return !permutable; | |||
2603 | } | |||
2604 | ||||
2605 | /* Is "node" the outermost node in its branch that can be tiled | |||
2606 | * and then mapped to block and thread identifiers? | |||
2607 | * If there are no such nodes in the subtree at "node" and | |||
2608 | * if "node" is not a filter node, then it is accepted too. | |||
2609 | */ | |||
2610 | static int is_outer_tilable(__isl_keep isl_schedule_node *node) | |||
2611 | { | |||
2612 | int tilable; | |||
2613 | isl_schedule_node *ancestor; | |||
2614 | ||||
2615 | tilable = is_candidate(node); | |||
2616 | if (tilable < 0) | |||
2617 | return -1; | |||
2618 | if (!tilable) | |||
2619 | return 0; | |||
2620 | ||||
2621 | tilable = 0; | |||
2622 | ancestor = isl_schedule_node_copy(node); | |||
2623 | while (isl_schedule_node_has_parent(ancestor)) { | |||
2624 | ancestor = isl_schedule_node_parent(ancestor); | |||
2625 | ||||
2626 | tilable = is_candidate(ancestor); | |||
2627 | if (tilable < 0 || tilable) | |||
2628 | break; | |||
2629 | } | |||
2630 | ||||
2631 | isl_schedule_node_free(ancestor); | |||
2632 | return tilable < 0 ? -1 : !tilable; | |||
2633 | } | |||
2634 | ||||
2635 | /* Collect the references to all writes in "group". | |||
2636 | * Each reference is represented by a universe set in a space | |||
2637 | * | |||
2638 | * [S[i,j] -> R[]] | |||
2639 | * | |||
2640 | * with S[i,j] the statement instance space and R[] the array reference. | |||
2641 | */ | |||
2642 | static __isl_give isl_union_set *group_tagged_writes( | |||
2643 | struct gpu_array_ref_group *group) | |||
2644 | { | |||
2645 | int i; | |||
2646 | isl_space *space; | |||
2647 | isl_union_set *writes; | |||
2648 | ||||
2649 | space = isl_map_get_space(group->access); | |||
2650 | writes = isl_union_set_empty(space); | |||
2651 | for (i = 0; i < group->n_ref; ++i) { | |||
2652 | isl_space *space; | |||
2653 | isl_set *writes_i; | |||
2654 | ||||
2655 | if (!group->refs[i]->write) | |||
2656 | continue; | |||
2657 | ||||
2658 | space = isl_map_get_space(group->refs[i]->tagged_access); | |||
2659 | space = isl_space_domain(space); | |||
2660 | writes_i = isl_set_universe(space); | |||
2661 | writes = isl_union_set_add_set(writes, writes_i); | |||
2662 | } | |||
2663 | ||||
2664 | return writes; | |||
2665 | } | |||
2666 | ||||
2667 | /* Is there any write access in "group" that requires synchronization | |||
2668 | * on a write to global memory? | |||
2669 | * We currently take into account all writes that would require | |||
2670 | * synchronization at the thread level depth, but if the copying | |||
2671 | * for this group is performed at an outer level, then we do not | |||
2672 | * actually need to take into account dependences at intermediate levels. | |||
2673 | */ | |||
2674 | static int any_sync_writes_in_group(struct ppcg_kernel *kernel, | |||
2675 | struct gpu_array_ref_group *group) | |||
2676 | { | |||
2677 | isl_union_set *writes; | |||
2678 | int empty, disjoint; | |||
2679 | ||||
2680 | empty = isl_union_set_is_empty(kernel->sync_writes); | |||
2681 | if (empty < 0) | |||
2682 | return -1; | |||
2683 | if (empty) | |||
2684 | return 0; | |||
2685 | ||||
2686 | writes = group_tagged_writes(group); | |||
2687 | disjoint = isl_union_set_is_disjoint(kernel->sync_writes, writes); | |||
2688 | isl_union_set_free(writes); | |||
2689 | ||||
2690 | return disjoint < 0 ? -1 : !disjoint; | |||
2691 | } | |||
2692 | ||||
2693 | /* Collect the references to all writes in "kernel" that write directly | |||
2694 | * to global or shared memory, i.e., that are not mapped to private memory. | |||
2695 | * Each reference is represented by a universe set in a space | |||
2696 | * | |||
2697 | * [S[i,j] -> R[]] | |||
2698 | * | |||
2699 | * with S[i,j] the statement instance space and R[] the array reference. | |||
2700 | */ | |||
2701 | static __isl_give isl_union_set *collect_non_private_tagged_writes( | |||
2702 | struct ppcg_kernel *kernel) | |||
2703 | { | |||
2704 | isl_union_set *writes; | |||
2705 | int i, j; | |||
2706 | ||||
2707 | writes = isl_union_set_empty(isl_union_set_get_space(kernel->arrays)); | |||
2708 | ||||
2709 | for (i = 0; i < kernel->n_array; ++i) { | |||
2710 | struct gpu_local_array_info *array = &kernel->array[i]; | |||
2711 | ||||
2712 | for (j = 0; j < array->n_group; ++j) { | |||
2713 | struct gpu_array_ref_group *group = array->groups[j]; | |||
2714 | enum ppcg_group_access_type type; | |||
2715 | isl_union_set *writes_ij; | |||
2716 | ||||
2717 | if (!group->write) | |||
2718 | continue; | |||
2719 | type = gpu_array_ref_group_type(group); | |||
2720 | if (type == ppcg_access_private) | |||
2721 | continue; | |||
2722 | writes_ij = group_tagged_writes(group); | |||
2723 | writes = isl_union_set_union(writes, writes_ij); | |||
2724 | } | |||
2725 | } | |||
2726 | ||||
2727 | return writes; | |||
2728 | } | |||
2729 | ||||
2730 | /* Are there any direct writes to global memory that require | |||
2731 | * synchronization? | |||
2732 | */ | |||
2733 | static int any_global_or_shared_sync_writes(struct ppcg_kernel *kernel) | |||
2734 | { | |||
2735 | isl_union_set *writes; | |||
2736 | int empty, disjoint; | |||
2737 | ||||
2738 | empty = isl_union_set_is_empty(kernel->sync_writes); | |||
2739 | if (empty < 0) | |||
2740 | return -1; | |||
2741 | if (empty) | |||
2742 | return 0; | |||
2743 | ||||
2744 | writes = collect_non_private_tagged_writes(kernel); | |||
2745 | disjoint = isl_union_set_is_disjoint(kernel->sync_writes, writes); | |||
2746 | isl_union_set_free(writes); | |||
2747 | ||||
2748 | return disjoint < 0 ? -1 : !disjoint; | |||
2749 | } | |||
2750 | ||||
2751 | /* Construct an isl_multi_val for use as tile sizes for tiling "node" | |||
2752 | * from the elements in "tile_size". | |||
2753 | */ | |||
2754 | static __isl_give isl_multi_val *construct_band_tiles_sizes( | |||
2755 | __isl_keep isl_schedule_node *node, int *tile_size) | |||
2756 | { | |||
2757 | isl_space *space; | |||
2758 | ||||
2759 | if (!node) | |||
2760 | return NULL((void*)0); | |||
2761 | ||||
2762 | space = isl_schedule_node_band_get_space(node); | |||
2763 | return ppcg_multi_val_from_int_list(space, tile_size); | |||
2764 | } | |||
2765 | ||||
2766 | /* Replace the partial schedule S of the band node "node" by | |||
2767 | * | |||
2768 | * floor(S/f) | |||
2769 | * | |||
2770 | * or | |||
2771 | * | |||
2772 | * f * floor(S/f) | |||
2773 | * | |||
2774 | * if scale_tile_loops is set, with f the integers in "factor". | |||
2775 | * The list that "factor" points to is assumed to contain at least | |||
2776 | * as many elements as the number of members in the band. | |||
2777 | */ | |||
2778 | static __isl_give isl_schedule_node *snap_band_to_sizes( | |||
2779 | __isl_take isl_schedule_node *node, int *factor, | |||
2780 | struct ppcg_options *options) | |||
2781 | { | |||
2782 | isl_multi_val *mv; | |||
2783 | ||||
2784 | mv = construct_band_tiles_sizes(node, factor); | |||
2785 | node = isl_schedule_node_band_scale_down(node, isl_multi_val_copy(mv)); | |||
2786 | if (options->scale_tile_loops) | |||
2787 | node = isl_schedule_node_band_scale(node, | |||
2788 | isl_multi_val_copy(mv)); | |||
2789 | isl_multi_val_free(mv); | |||
2790 | ||||
2791 | return node; | |||
2792 | } | |||
2793 | ||||
2794 | /* Tile "band" with tile size specified by "sizes". | |||
2795 | * | |||
2796 | * Since the tile loops will be mapped to block ids, we forcibly | |||
2797 | * turn off tile loop scaling. We may want to enable tile loop scaling | |||
2798 | * at some later point, but then we would have to support the detection | |||
2799 | * of strides during the mapping to block ids. | |||
2800 | * Similarly, since the point loops will be mapped to thread ids, | |||
2801 | * we forcibly shift the point loops so that they start at zero. | |||
2802 | */ | |||
2803 | static __isl_give isl_schedule_node *tile_band( | |||
2804 | __isl_take isl_schedule_node *node, __isl_take isl_multi_val *sizes) | |||
2805 | { | |||
2806 | isl_ctx *ctx = isl_schedule_node_get_ctx(node); | |||
2807 | int scale_tile; | |||
2808 | int shift_point; | |||
2809 | ||||
2810 | scale_tile = isl_options_get_tile_scale_tile_loops(ctx); | |||
2811 | isl_options_set_tile_scale_tile_loops(ctx, 0); | |||
2812 | shift_point = isl_options_get_tile_shift_point_loops(ctx); | |||
2813 | isl_options_set_tile_shift_point_loops(ctx, 1); | |||
2814 | ||||
2815 | node = isl_schedule_node_band_tile(node, sizes); | |||
2816 | ||||
2817 | isl_options_set_tile_scale_tile_loops(ctx, scale_tile); | |||
2818 | isl_options_set_tile_shift_point_loops(ctx, shift_point); | |||
2819 | ||||
2820 | return node; | |||
2821 | } | |||
2822 | ||||
2823 | /* Extract the set of parameter values and outer schedule dimensions | |||
2824 | * for which any statement instance | |||
2825 | * in the kernel inserted at "node" needs to be executed. | |||
2826 | * Intersect the set of parameter values derived from the host schedule | |||
2827 | * relation with the context of "prog". | |||
2828 | */ | |||
2829 | static __isl_give isl_set *extract_context(__isl_keep isl_schedule_node *node, | |||
2830 | struct gpu_prog *prog) | |||
2831 | { | |||
2832 | isl_union_map *schedule; | |||
2833 | isl_union_set *schedule_domain; | |||
2834 | isl_set *context; | |||
2835 | int empty; | |||
2836 | ||||
2837 | schedule = isl_schedule_node_get_prefix_schedule_relation(node); | |||
2838 | schedule_domain = isl_union_map_range(schedule); | |||
2839 | empty = isl_union_set_is_empty(schedule_domain); | |||
2840 | if (empty < 0) { | |||
2841 | isl_union_set_free(schedule_domain); | |||
2842 | return NULL((void*)0); | |||
2843 | } | |||
2844 | if (empty) { | |||
2845 | int depth; | |||
2846 | isl_space *space; | |||
2847 | ||||
2848 | space = isl_union_set_get_space(schedule_domain); | |||
2849 | isl_union_set_free(schedule_domain); | |||
2850 | space = isl_space_set_from_params(space); | |||
2851 | depth = isl_schedule_node_get_schedule_depth(node); | |||
2852 | space = isl_space_add_dims(space, isl_dim_set, depth); | |||
2853 | context = isl_set_empty(space); | |||
2854 | } else { | |||
2855 | context = isl_set_from_union_set(schedule_domain); | |||
2856 | } | |||
2857 | context = isl_set_intersect_params(context, | |||
2858 | isl_set_copy(prog->context)); | |||
2859 | ||||
2860 | return context; | |||
2861 | } | |||
2862 | ||||
2863 | /* Return the set of outer array elements accessed by | |||
2864 | * by the statement instances in "domain" in "prog". | |||
2865 | * The instances in "domain" are those that appear | |||
2866 | * in the domains of the access relations in "prog". | |||
2867 | */ | |||
2868 | static __isl_give isl_union_set *accessed_by_domain( | |||
2869 | __isl_take isl_union_set *domain, struct gpu_prog *prog) | |||
2870 | { | |||
2871 | isl_union_map *access; | |||
2872 | isl_union_set *arrays; | |||
2873 | ||||
2874 | access = isl_union_map_union(isl_union_map_copy(prog->read), | |||
2875 | isl_union_map_copy(prog->may_write)); | |||
2876 | access = isl_union_map_intersect_domain(access, domain); | |||
2877 | arrays = isl_union_map_range(access); | |||
2878 | arrays = isl_union_set_apply(arrays, | |||
2879 | isl_union_map_copy(prog->to_outer)); | |||
2880 | ||||
2881 | return arrays; | |||
2882 | } | |||
2883 | ||||
2884 | /* Return the number of outer band members of the band node "node" | |||
2885 | * that are marked coincident. | |||
2886 | */ | |||
2887 | static int n_outer_coincidence(__isl_keep isl_schedule_node *node) | |||
2888 | { | |||
2889 | int i, n; | |||
2890 | ||||
2891 | n = isl_schedule_node_band_n_member(node); | |||
2892 | ||||
2893 | for (i = 0; i < n; ++i) | |||
2894 | if (!isl_schedule_node_band_member_get_coincident(node, i)) | |||
2895 | break; | |||
2896 | ||||
2897 | return i; | |||
2898 | } | |||
2899 | ||||
2900 | /* If the band node "node" has more than "n" members, then split off | |||
2901 | * the first "n" of them. | |||
2902 | */ | |||
2903 | static __isl_give isl_schedule_node *split_band( | |||
2904 | __isl_take isl_schedule_node *node, int n) | |||
2905 | { | |||
2906 | int dim; | |||
2907 | ||||
2908 | dim = isl_schedule_node_band_n_member(node); | |||
2909 | if (n < dim) | |||
2910 | node = isl_schedule_node_band_split(node, n); | |||
2911 | ||||
2912 | return node; | |||
2913 | } | |||
2914 | ||||
2915 | /* Scale a band node that may have been split by split_band. | |||
2916 | * "sizes" are the scaling factors for the original node. | |||
2917 | * "node" either points to the original band node, or the outer | |||
2918 | * of the two pieces after splitting. | |||
2919 | * | |||
2920 | * If the number of elements in "node" is smaller than the number of | |||
2921 | * elements in "sizes", then some splitting has occurred and we split | |||
2922 | * "sizes" in the same way. | |||
2923 | */ | |||
2924 | static __isl_give isl_schedule_node *scale_band( | |||
2925 | __isl_take isl_schedule_node *node, __isl_take isl_multi_val *sizes) | |||
2926 | { | |||
2927 | int n, dim; | |||
2928 | ||||
2929 | n = isl_multi_val_dim(sizes, isl_dim_set); | |||
2930 | dim = isl_schedule_node_band_n_member(node); | |||
2931 | if (n > dim) { | |||
2932 | isl_multi_val *sizes2; | |||
2933 | ||||
2934 | sizes2 = isl_multi_val_copy(sizes); | |||
2935 | sizes = isl_multi_val_drop_dims(sizes, | |||
2936 | isl_dim_set, dim, n - dim); | |||
2937 | sizes2 = isl_multi_val_drop_dims(sizes2, isl_dim_set, 0, dim); | |||
2938 | node = isl_schedule_node_child(node, 0); | |||
2939 | node = isl_schedule_node_band_scale(node, sizes2); | |||
2940 | node = isl_schedule_node_parent(node); | |||
2941 | } | |||
2942 | ||||
2943 | return isl_schedule_node_band_scale(node, sizes); | |||
2944 | } | |||
2945 | ||||
2946 | /* Return an isl_multi_aff, with as elements the parameters in "space" | |||
2947 | * that have the names specified by the elements in "names". | |||
2948 | * If (some of) these parameters do not already appear in "space", | |||
2949 | * then they are added first. | |||
2950 | */ | |||
2951 | static __isl_give isl_multi_aff *parameter_vector(__isl_take isl_space *space, | |||
2952 | __isl_keep isl_id_list *names) | |||
2953 | { | |||
2954 | int i, n; | |||
2955 | isl_local_space *ls; | |||
2956 | isl_multi_aff *ma; | |||
2957 | ||||
2958 | if (!names) | |||
2959 | space = isl_space_free(space); | |||
2960 | ||||
2961 | n = isl_id_list_n_id(names); | |||
2962 | for (i = 0; i < n; ++i) { | |||
2963 | int pos; | |||
2964 | isl_id *id; | |||
2965 | ||||
2966 | id = isl_id_list_get_id(names, i); | |||
2967 | pos = isl_space_find_dim_by_id(space, isl_dim_param, id); | |||
2968 | if (pos >= 0) { | |||
2969 | isl_id_free(id); | |||
2970 | continue; | |||
2971 | } | |||
2972 | pos = isl_space_dim(space, isl_dim_param); | |||
2973 | space = isl_space_add_dims(space, isl_dim_param, 1); | |||
2974 | space = isl_space_set_dim_id(space, isl_dim_param, pos, id); | |||
2975 | } | |||
2976 | ma = isl_multi_aff_zero(isl_space_copy(space)); | |||
2977 | ls = isl_local_space_from_space(isl_space_domain(space)); | |||
2978 | for (i = 0; i < n; ++i) { | |||
2979 | int pos; | |||
2980 | isl_id *id; | |||
2981 | isl_aff *aff; | |||
2982 | ||||
2983 | id = isl_id_list_get_id(names, i); | |||
2984 | pos = isl_space_find_dim_by_id(space, isl_dim_param, id); | |||
2985 | isl_id_free(id); | |||
2986 | aff = isl_aff_var_on_domain(isl_local_space_copy(ls), | |||
2987 | isl_dim_param, pos); | |||
2988 | ma = isl_multi_aff_set_aff(ma, i, aff); | |||
2989 | } | |||
2990 | isl_local_space_free(ls); | |||
2991 | ||||
2992 | return ma; | |||
2993 | } | |||
2994 | ||||
2995 | /* Return constraints on the domain elements that equate a sequence of | |||
2996 | * parameters called "names", to the partial schedule | |||
2997 | * of "node" modulo the integers in "size". | |||
2998 | * The number of elements in the array "size" should be equal | |||
2999 | * to the number of elements in "names". | |||
3000 | * The number of members of the band node "node" should be smaller | |||
3001 | * than or equal to this number. If it is smaller, then the first | |||
3002 | * elements of "names" are equated to zero. | |||
3003 | */ | |||
3004 | static __isl_give isl_union_set *set_schedule_modulo( | |||
3005 | __isl_keep isl_schedule_node *node, __isl_keep isl_id_list *names, | |||
3006 | int *size) | |||
3007 | { | |||
3008 | int n, n_zero; | |||
3009 | isl_space *space; | |||
3010 | isl_multi_aff *ma; | |||
3011 | isl_multi_union_pw_aff *mupa, *mupa2; | |||
3012 | isl_multi_val *mv; | |||
3013 | isl_union_set *domain; | |||
3014 | ||||
3015 | if (!node) | |||
3016 | return NULL((void*)0); | |||
3017 | n = isl_id_list_n_id(names); | |||
3018 | if (n == 0) | |||
3019 | return isl_schedule_node_get_universe_domain(node); | |||
3020 | n_zero = n - isl_schedule_node_band_n_member(node); | |||
3021 | ||||
3022 | mupa = isl_schedule_node_band_get_partial_schedule(node); | |||
3023 | mv = construct_band_tiles_sizes(node, size + n_zero); | |||
3024 | mupa = isl_multi_union_pw_aff_mod_multi_val(mupa, mv); | |||
3025 | ||||
3026 | space = isl_multi_union_pw_aff_get_space(mupa); | |||
3027 | space = isl_space_params(space); | |||
3028 | space = isl_space_set_from_params(space); | |||
3029 | space = isl_space_add_dims(space, isl_dim_set, n_zero); | |||
3030 | ma = isl_multi_aff_zero(space); | |||
3031 | ||||
3032 | domain = isl_schedule_node_get_universe_domain(node); | |||
3033 | mupa2 = isl_multi_union_pw_aff_multi_aff_on_domain( | |||
3034 | isl_union_set_copy(domain), ma); | |||
3035 | mupa = isl_multi_union_pw_aff_range_product(mupa2, mupa); | |||
3036 | ||||
3037 | space = isl_multi_union_pw_aff_get_space(mupa); | |||
3038 | ma = parameter_vector(space, names); | |||
3039 | ||||
3040 | mupa2 = isl_multi_union_pw_aff_multi_aff_on_domain(domain, ma); | |||
3041 | mupa = isl_multi_union_pw_aff_sub(mupa, mupa2); | |||
3042 | ||||
3043 | return isl_multi_union_pw_aff_zero_union_set(mupa); | |||
3044 | } | |||
3045 | ||||
3046 | /* Insert a context node at "node" introducing the block and thread | |||
3047 | * identifiers along with their bounds, which are stored in kernel->grid_size | |||
3048 | * and kernel->block_dim. | |||
3049 | * Note that the bounds on the block identifiers may implicitly impose | |||
3050 | * constraints on the parameters. A guard needs to be inserted | |||
3051 | * in the schedule tree to ensure that those bounds hold at "node". | |||
3052 | * This guard is inserted in insert_guard. | |||
3053 | */ | |||
3054 | static __isl_give isl_schedule_node *insert_context(struct ppcg_kernel *kernel, | |||
3055 | __isl_take isl_schedule_node *node) | |||
3056 | { | |||
3057 | isl_set *context; | |||
3058 | ||||
3059 | context = isl_set_universe(isl_set_get_space(kernel->context)); | |||
3060 | ||||
3061 | context = add_bounded_parameters_dynamic(context, | |||
3062 | kernel->grid_size, kernel->block_ids); | |||
3063 | context = add_bounded_parameters(context, | |||
3064 | kernel->block_dim, kernel->thread_ids); | |||
3065 | ||||
3066 | node = isl_schedule_node_insert_context(node, context); | |||
3067 | ||||
3068 | return node; | |||
3069 | } | |||
3070 | ||||
3071 | /* Insert a guard that eliminates kernel launches where the kernel | |||
3072 | * obviously does not have any work to do. | |||
3073 | * | |||
3074 | * In particular, eliminate kernel launches where there are obviously | |||
3075 | * zero blocks. | |||
3076 | * Use the same block size constraints that are used to create the context | |||
3077 | * to ensure that all constraints implicit in the constructed context | |||
3078 | * are imposed by the guard. | |||
3079 | * | |||
3080 | * Additionally, add other constraints that are valid | |||
3081 | * for each executed instance ("context"), as long as this does not result | |||
3082 | * in a disjunction. | |||
3083 | */ | |||
3084 | static __isl_give isl_schedule_node *insert_guard( | |||
3085 | __isl_take isl_schedule_node *node, __isl_keep isl_set *context, | |||
3086 | __isl_keep isl_multi_pw_aff *size, struct ppcg_scop *scop) | |||
3087 | { | |||
3088 | unsigned nparam, n; | |||
3089 | isl_set *guard; | |||
3090 | isl_id_list *ids; | |||
3091 | ||||
3092 | guard = isl_set_copy(context); | |||
3093 | guard = isl_set_compute_divs(guard); | |||
3094 | guard = isl_set_from_basic_set(isl_set_simple_hull(guard)); | |||
3095 | ||||
3096 | nparam = isl_set_dim(guard, isl_dim_param); | |||
3097 | n = isl_multi_pw_aff_dim(size, isl_dim_out); | |||
3098 | ids = ppcg_scop_generate_names(scop, n, "__ppcg_tmp"); | |||
3099 | guard = add_bounded_parameters_dynamic(guard, size, ids); | |||
3100 | isl_id_list_free(ids); | |||
3101 | guard = isl_set_project_out(guard, isl_dim_param, nparam, n); | |||
3102 | ||||
3103 | node = isl_schedule_node_insert_guard(node, guard); | |||
3104 | ||||
3105 | return node; | |||
3106 | } | |||
3107 | ||||
3108 | /* Does any array reference group mapping require the band that is mapped | |||
3109 | * to threads to be unrolled? | |||
3110 | */ | |||
3111 | static int kernel_requires_unroll(struct ppcg_kernel *kernel) | |||
3112 | { | |||
3113 | int i, j; | |||
3114 | ||||
3115 | for (i = 0; i < kernel->n_array; ++i) { | |||
3116 | struct gpu_local_array_info *array = &kernel->array[i]; | |||
3117 | ||||
3118 | for (j = 0; j < array->n_group; ++j) { | |||
3119 | struct gpu_array_ref_group *group = array->groups[j]; | |||
3120 | if (gpu_array_ref_group_requires_unroll(group)) | |||
3121 | return 1; | |||
3122 | } | |||
3123 | } | |||
3124 | ||||
3125 | return 0; | |||
3126 | } | |||
3127 | ||||
3128 | /* Mark the given band node "node" for unrolling by the AST generator and | |||
3129 | * then sink it to the leaves of the schedule tree. | |||
3130 | * All dimensions of "node" are assumed to be coincident, such that this | |||
3131 | * sinking is a valid operation. | |||
3132 | */ | |||
3133 | static __isl_give isl_schedule_node *unroll(__isl_take isl_schedule_node *node) | |||
3134 | { | |||
3135 | node = ppcg_set_schedule_node_type(node, isl_ast_loop_unroll); | |||
3136 | ||||
3137 | node = isl_schedule_node_band_sink(node); | |||
3138 | ||||
3139 | return node; | |||
3140 | } | |||
3141 | ||||
3142 | /* Insert a synchronization node in the schedule tree of "node" | |||
3143 | * after the core computation of "kernel" at the level of the band | |||
3144 | * that is mapped to threads, except if that level is equal to | |||
3145 | * that of the band that is mapped to blocks or if there are no writes | |||
3146 | * to global or shared memory in the core computation that require | |||
3147 | * synchronization. | |||
3148 | * If there are any writes to shared memory and the shared memory | |||
3149 | * copying is performed at the same level, then synchronization | |||
3150 | * is needed between the core and the copying anyway, so we might | |||
3151 | * as well add it here. If the copying is performed at a higher | |||
3152 | * level, then different iterations of intermediate schedule dimensions | |||
3153 | * may have a different mapping from between shared memory elements and | |||
3154 | * threads, such that synchronization is required after the core. | |||
3155 | * "node" is assumed to point to the kernel node. | |||
3156 | * | |||
3157 | * If the shared and the thread mark point to the same node, then make | |||
3158 | * sure the synchronization is inserted outside of the shared mark. | |||
3159 | */ | |||
3160 | static __isl_give isl_schedule_node *add_sync(struct ppcg_kernel *kernel, | |||
3161 | __isl_take isl_schedule_node *node) | |||
3162 | { | |||
3163 | int depth; | |||
3164 | int need_sync; | |||
3165 | ||||
3166 | need_sync = any_global_or_shared_sync_writes(kernel); | |||
3167 | if (need_sync < 0) | |||
3168 | return isl_schedule_node_free(node); | |||
3169 | if (!need_sync) | |||
3170 | return node; | |||
3171 | ||||
3172 | node = gpu_tree_move_down_to_thread(node, kernel->core); | |||
3173 | depth = isl_schedule_node_get_schedule_depth(node); | |||
3174 | node = gpu_tree_move_up_to_kernel(node); | |||
3175 | if (depth == isl_schedule_node_get_schedule_depth(node)) | |||
3176 | return node; | |||
3177 | ||||
3178 | node = gpu_tree_move_down_to_depth(node, depth, kernel->core); | |||
3179 | node = gpu_tree_ensure_following_sync(node, kernel); | |||
3180 | ||||
3181 | node = gpu_tree_move_up_to_kernel(node); | |||
3182 | ||||
3183 | return node; | |||
3184 | } | |||
3185 | ||||
3186 | /* Return a read ("read" is 1) or write access relation for "group" | |||
3187 | * with those accesses removed that are only needed to communicate data | |||
3188 | * within the subtree of the schedule rooted at "node". | |||
3189 | * Furthermore, include the prefix schedule at "node". | |||
3190 | * That is, return a relation of the form | |||
3191 | * | |||
3192 | * S -> [D -> A] | |||
3193 | * | |||
3194 | * with D the outer schedule dimensions at "node". | |||
3195 | */ | |||
3196 | static __isl_give isl_union_map *anchored_non_local_accesses( | |||
3197 | struct ppcg_kernel *kernel, struct gpu_array_ref_group *group, | |||
3198 | __isl_take isl_schedule_node *node, int read) | |||
3199 | { | |||
3200 | isl_union_map *access; | |||
3201 | isl_union_map *prefix; | |||
3202 | ||||
3203 | prefix = isl_schedule_node_get_prefix_schedule_relation(node); | |||
3204 | prefix = isl_union_map_preimage_domain_union_pw_multi_aff(prefix, | |||
3205 | isl_union_pw_multi_aff_copy(kernel->contraction)); | |||
3206 | access = gpu_array_ref_group_access_relation(group, read, !read); | |||
3207 | access = remove_local_accesses_group(kernel, group, access, prefix, | |||
3208 | read); | |||
3209 | access = isl_union_map_range_product(prefix, access); | |||
3210 | ||||
3211 | return access; | |||
3212 | } | |||
3213 | ||||
3214 | /* Given an array reference group "group", create a mapping | |||
3215 | * | |||
3216 | * read[D -> A] -> [D -> A] | |||
3217 | * | |||
3218 | * if "read" is set or | |||
3219 | * | |||
3220 | * write[D -> A] -> [D -> A] | |||
3221 | * | |||
3222 | * if "read" is not set. | |||
3223 | * D corresponds to the outer tile->depth dimensions of | |||
3224 | * the kernel schedule. | |||
3225 | */ | |||
3226 | static __isl_give isl_multi_aff *create_from_access(isl_ctx *ctx, | |||
3227 | struct gpu_array_ref_group *group, int read) | |||
3228 | { | |||
3229 | struct gpu_array_tile *tile; | |||
3230 | isl_space *space; | |||
3231 | isl_id *id; | |||
3232 | ||||
3233 | tile = gpu_array_ref_group_tile(group); | |||
3234 | space = isl_space_copy(group->array->space); | |||
3235 | space = isl_space_from_range(space); | |||
3236 | space = isl_space_add_dims(space, isl_dim_in, tile->depth); | |||
3237 | space = isl_space_wrap(space); | |||
3238 | space = isl_space_map_from_set(space); | |||
3239 | ||||
3240 | id = isl_id_alloc(ctx, read ? "read" : "write", group); | |||
3241 | space = isl_space_set_tuple_id(space, isl_dim_in, id); | |||
3242 | ||||
3243 | return isl_multi_aff_identity(space); | |||
3244 | } | |||
3245 | ||||
3246 | /* If any writes in "group" require synchronization, then make sure | |||
3247 | * that there is a synchronization node for "kernel" after the node | |||
3248 | * following "node" in a sequence. | |||
3249 | * | |||
3250 | * If "shared" is set and no synchronization is needed for | |||
3251 | * the writes to global memory, then add synchronization before | |||
3252 | * the kernel to protect shared memory from being overwritten | |||
3253 | * by the next iteration of the core computation. | |||
3254 | * No additional synchronization is needed to protect against | |||
3255 | * the next copy into shared memory because each element of | |||
3256 | * the shared memory tile is always copied by the same thread. | |||
3257 | */ | |||
3258 | static __isl_give isl_schedule_node *add_group_write_sync( | |||
3259 | __isl_take isl_schedule_node *node, struct ppcg_kernel *kernel, | |||
3260 | struct gpu_array_ref_group *group, int shared) | |||
3261 | { | |||
3262 | int need_sync; | |||
3263 | ||||
3264 | need_sync = any_sync_writes_in_group(kernel, group); | |||
3265 | if (need_sync < 0) | |||
3266 | return isl_schedule_node_free(node); | |||
3267 | if (need_sync) { | |||
3268 | node = isl_schedule_node_parent(node); | |||
3269 | node = isl_schedule_node_next_sibling(node); | |||
3270 | node = isl_schedule_node_child(node, 0); | |||
3271 | node = gpu_tree_ensure_following_sync(node, kernel); | |||
3272 | } else if (shared) { | |||
3273 | struct gpu_array_tile *tile; | |||
3274 | ||||
3275 | tile = gpu_array_ref_group_tile(group); | |||
3276 | node = isl_schedule_node_parent(node); | |||
3277 | node = isl_schedule_node_parent(node); | |||
3278 | node = gpu_tree_move_down_to_depth(node, tile->depth, | |||
3279 | kernel->core); | |||
3280 | node = gpu_tree_move_left_to_sync(node, kernel); | |||
3281 | } | |||
3282 | ||||
3283 | return node; | |||
3284 | } | |||
3285 | ||||
3286 | /* Add copy statements to the schedule tree of "node" | |||
3287 | * for reading from global memory to private memory (if "read" is set) or | |||
3288 | * for writing back from private memory to global memory | |||
3289 | * (if "read" is not set) for the array reference group "group" that | |||
3290 | * is mapped to private memory. | |||
3291 | * On input, "node" points to the kernel node, and it is moved | |||
3292 | * back there on output. | |||
3293 | * | |||
3294 | * The copies are performed in the order of the array elements. | |||
3295 | * The copy statement instances include a reference to the outer | |||
3296 | * tile->depth dimensions of the kernel schedule for ease of | |||
3297 | * combining them with the group tiling. | |||
3298 | * | |||
3299 | * That is, the extra schedule is of the form | |||
3300 | * | |||
3301 | * type[D -> A] -> A | |||
3302 | * | |||
3303 | * where D corresponds to the outer tile->depth dimensions of | |||
3304 | * the kernel schedule and A to the global array. | |||
3305 | * This schedule is unrolled because registers are not addressable. | |||
3306 | * | |||
3307 | * The copying is inserted in the schedule tree through an extension | |||
3308 | * of the form | |||
3309 | * | |||
3310 | * D -> type[D -> A] | |||
3311 | * | |||
3312 | * where the extra domain elements type[D -> A] are those accessed | |||
3313 | * by the group. | |||
3314 | * A filter is inserted on type[D -> A] to ensure that the element | |||
3315 | * is read/written by the same thread that needs the element. | |||
3316 | * This filter is obtained by applying | |||
3317 | * | |||
3318 | * S -> type[D -> A] | |||
3319 | * | |||
3320 | * to the thread filter for the core statements. | |||
3321 | * | |||
3322 | * The extension is inserted before the core computation in case of a read | |||
3323 | * and after the core computation in case of a write. | |||
3324 | * In the latter case, we also make sure that there is a synchronization | |||
3325 | * node after the write to global memory, unless this write is performed | |||
3326 | * at the outer level of the kernel. | |||
3327 | * In principle, this synchronization could be inserted higher | |||
3328 | * in the schedule tree depending on where the corresponding reads | |||
3329 | * from global memory are performed. | |||
3330 | */ | |||
3331 | static __isl_give isl_schedule_node *add_copies_group_private( | |||
3332 | struct ppcg_kernel *kernel, struct gpu_array_ref_group *group, | |||
3333 | __isl_take isl_schedule_node *node, int read) | |||
3334 | { | |||
3335 | struct gpu_array_tile *tile; | |||
3336 | isl_union_map *access; | |||
3337 | isl_union_set *domain; | |||
3338 | isl_space *space; | |||
3339 | isl_multi_aff *from_access; | |||
3340 | isl_multi_pw_aff *mpa; | |||
3341 | isl_multi_union_pw_aff *mupa; | |||
3342 | isl_union_pw_multi_aff *contraction; | |||
3343 | isl_schedule_node *graft; | |||
3344 | isl_union_set *filter; | |||
3345 | int kernel_depth; | |||
3346 | int empty; | |||
3347 | ||||
3348 | kernel_depth = isl_schedule_node_get_schedule_depth(node); | |||
3349 | tile = gpu_array_ref_group_tile(group); | |||
3350 | node = gpu_tree_move_down_to_depth(node, tile->depth, kernel->core); | |||
3351 | ||||
3352 | access = anchored_non_local_accesses(kernel, group, node, read); | |||
3353 | empty = isl_union_map_is_empty(access); | |||
3354 | if (empty < 0 || empty) { | |||
3355 | isl_union_map_free(access); | |||
3356 | if (empty < 0) | |||
3357 | return isl_schedule_node_free(node); | |||
3358 | return gpu_tree_move_up_to_kernel(node); | |||
3359 | } | |||
3360 | ||||
3361 | group->array->global = 1; | |||
3362 | group->local_array->global = 1; | |||
3363 | ||||
3364 | from_access = create_from_access(kernel->ctx, group, read); | |||
3365 | space = isl_space_domain(isl_multi_aff_get_space(from_access)); | |||
3366 | access = isl_union_map_preimage_range_multi_aff(access, from_access); | |||
3367 | ||||
3368 | filter = isl_union_set_copy(kernel->thread_filter); | |||
3369 | contraction = isl_union_pw_multi_aff_copy(kernel->contraction); | |||
3370 | filter = isl_union_set_preimage_union_pw_multi_aff(filter, contraction); | |||
3371 | filter = isl_union_set_apply(filter, isl_union_map_copy(access)); | |||
3372 | filter = isl_union_set_detect_equalities(filter); | |||
3373 | filter = isl_union_set_coalesce(filter); | |||
3374 | ||||
3375 | domain = isl_union_map_range(access); | |||
3376 | access = isl_union_set_wrapped_domain_map(domain); | |||
3377 | access = isl_union_map_reverse(access); | |||
3378 | access = isl_union_map_coalesce(access); | |||
3379 | graft = isl_schedule_node_from_extension(access); | |||
3380 | ||||
3381 | space = isl_space_map_from_set(space); | |||
3382 | mpa = isl_multi_pw_aff_identity(space); | |||
3383 | mpa = isl_multi_pw_aff_range_factor_range(mpa); | |||
3384 | mupa = isl_multi_union_pw_aff_from_multi_pw_aff(mpa); | |||
3385 | ||||
3386 | graft = isl_schedule_node_child(graft, 0); | |||
3387 | graft = isl_schedule_node_insert_partial_schedule(graft, mupa); | |||
3388 | graft = unroll(graft); | |||
3389 | ||||
3390 | graft = isl_schedule_node_insert_filter(graft, filter); | |||
3391 | ||||
3392 | graft = isl_schedule_node_parent(graft); | |||
3393 | ||||
3394 | if (read) | |||
3395 | node = isl_schedule_node_graft_before(node, graft); | |||
3396 | else { | |||
3397 | node = isl_schedule_node_graft_after(node, graft); | |||
3398 | if (kernel_depth < tile->depth) | |||
3399 | node = add_group_write_sync(node, kernel, group, 0); | |||
3400 | } | |||
3401 | ||||
3402 | node = gpu_tree_move_up_to_kernel(node); | |||
3403 | ||||
3404 | return node; | |||
3405 | } | |||
3406 | ||||
3407 | /* Add copy statements to the schedule tree of "node" | |||
3408 | * for reading from global memory to shared memory (if "read" is set) or | |||
3409 | * for writing back from shared memory to global memory | |||
3410 | * (if "read" is not set) for the array reference group "group" that | |||
3411 | * is mapped to shared memory. | |||
3412 | * On input, "node" points to the kernel node, and it is moved | |||
3413 | * back there on output. | |||
3414 | * | |||
3415 | * The copies are performed in the order of the corresponding shared | |||
3416 | * memory tile. | |||
3417 | * The copy statement instances include a reference to the outer | |||
3418 | * tile->depth dimensions of the kernel schedule for ease of | |||
3419 | * combining them with the group tiling. | |||
3420 | * | |||
3421 | * If we are performing a read from global memory to shared memory and | |||
3422 | * if the array involved is not a scalar, then we copy | |||
3423 | * the entire tile to shared memory. This may result in some extra | |||
3424 | * elements getting copied, but it should lead to simpler code | |||
3425 | * (which means that fewer registers may be needed) and less divergence. | |||
3426 | * | |||
3427 | * Otherwise, we only copy the elements that will be read or have been written | |||
3428 | * in the kernel. | |||
3429 | * | |||
3430 | * That is, the extra schedule is of the form | |||
3431 | * | |||
3432 | * type[D -> A] -> T | |||
3433 | * | |||
3434 | * where D corresponds to the outer tile->depth dimensions of | |||
3435 | * the kernel schedule, A to the global array and T is the corresponding | |||
3436 | * shared memory tile. | |||
3437 | * | |||
3438 | * The copying is inserted in the schedule tree through an extension | |||
3439 | * of the form | |||
3440 | * | |||
3441 | * D -> type[D -> A] | |||
3442 | * | |||
3443 | * where the extra domain elements type[D -> A] are those accessed | |||
3444 | * by the group. In the case of read from a non-scalar, this set | |||
3445 | * is replaced by the entire shared memory tile. | |||
3446 | * | |||
3447 | * If the "unroll_copy_shared" option is set, then the AST generator | |||
3448 | * is instructed to unroll the copying code. | |||
3449 | * | |||
3450 | * A filter is inserted on type[D -> A] to map the copy instances | |||
3451 | * to the threads. In particular, the thread identifiers are | |||
3452 | * equated to the position inside the shared memory tile (T) | |||
3453 | * modulo the block size. | |||
3454 | * We try to align the innermost tile dimension with the innermost | |||
3455 | * thread identifier (x) as a heuristic to improve coalescing. | |||
3456 | * In particular, if the dimension of the tile is greater than | |||
3457 | * the dimension of the block, then the schedule mapping to the tile | |||
3458 | * is broken up into two pieces and the filter is applied to the inner part. | |||
3459 | * If, on the other hand, the dimension of the tile is smaller than | |||
3460 | * the dimension of the block, then the initial thread identifiers | |||
3461 | * are equated to zero and the remaining thread identifiers are | |||
3462 | * matched to the memory tile. | |||
3463 | * | |||
3464 | * The extension is inserted before the core computation in case of a read | |||
3465 | * and after the core computation in case of a write. | |||
3466 | * In the case of a read, we first need to make sure there is some | |||
3467 | * synchronization before the core computation such that we can put the read | |||
3468 | * from global memory to shared memory before that synchronization. | |||
3469 | * This ensures that all threads have finished copying into shared memory | |||
3470 | * before the shared memory is used. | |||
3471 | * We also need to make sure that there is a synchronization node after | |||
3472 | * the core computation to ensure that the next load into shared memory | |||
3473 | * only happens after all data has been used. There is no need for | |||
3474 | * this synchronization if we are at the outer level since then there | |||
3475 | * won't be a next load. | |||
3476 | * In the case of a write, we need to make sure there is some synchronization | |||
3477 | * after the core computation such taht we can put the write from shared | |||
3478 | * memory to global memory after that synchronization. | |||
3479 | * Unless we are at the outer level, we also need a synchronization node | |||
3480 | * after the write to ensure the data is saved to global memory | |||
3481 | * before the next iteration write to the same shared memory. | |||
3482 | * It also makes sure the data has arrived in global memory before | |||
3483 | * it is read in a subsequent iteration. | |||
3484 | */ | |||
3485 | static __isl_give isl_schedule_node *add_copies_group_shared( | |||
3486 | struct ppcg_kernel *kernel, struct gpu_array_ref_group *group, | |||
3487 | __isl_take isl_schedule_node *node, int read) | |||
3488 | { | |||
3489 | struct gpu_array_tile *tile; | |||
3490 | isl_union_map *access; | |||
3491 | isl_union_set *domain; | |||
3492 | isl_multi_aff *ma; | |||
3493 | isl_multi_aff *from_access; | |||
3494 | isl_multi_pw_aff *mpa; | |||
3495 | isl_multi_union_pw_aff *mupa; | |||
3496 | isl_schedule_node *graft; | |||
3497 | isl_union_set *filter; | |||
3498 | int skip; | |||
3499 | int kernel_depth; | |||
3500 | int empty; | |||
3501 | ||||
3502 | tile = gpu_array_ref_group_tile(group); | |||
3503 | kernel_depth = isl_schedule_node_get_schedule_depth(node); | |||
3504 | node = gpu_tree_move_down_to_depth(node, tile->depth, kernel->core); | |||
3505 | ||||
3506 | access = anchored_non_local_accesses(kernel, group, node, read); | |||
3507 | empty = isl_union_map_is_empty(access); | |||
3508 | if (empty < 0 || empty) { | |||
3509 | isl_union_map_free(access); | |||
3510 | if (empty < 0) | |||
3511 | return isl_schedule_node_free(node); | |||
3512 | return gpu_tree_move_up_to_kernel(node); | |||
3513 | } | |||
3514 | ||||
3515 | group->array->global = 1; | |||
3516 | group->local_array->global = 1; | |||
3517 | ||||
3518 | from_access = create_from_access(kernel->ctx, group, read); | |||
3519 | ||||
3520 | ma = isl_multi_aff_copy(tile->tiling); | |||
3521 | ma = isl_multi_aff_pullback_multi_aff(ma, | |||
3522 | isl_multi_aff_copy(from_access)); | |||
3523 | mpa = isl_multi_pw_aff_from_multi_aff(ma); | |||
3524 | mupa = isl_multi_union_pw_aff_from_multi_pw_aff(mpa); | |||
3525 | ||||
3526 | domain = isl_union_map_range(access); | |||
3527 | ||||
3528 | if (read && !gpu_array_is_scalar(group->array)) { | |||
3529 | isl_map *map; | |||
3530 | isl_union_set_free(domain); | |||
3531 | map = group_tile(group); | |||
3532 | domain = isl_union_set_from_set(isl_map_wrap(map)); | |||
3533 | } | |||
3534 | ||||
3535 | domain = isl_union_set_preimage_multi_aff(domain, from_access); | |||
3536 | access = isl_union_set_wrapped_domain_map(domain); | |||
3537 | access = isl_union_map_reverse(access); | |||
3538 | access = isl_union_map_coalesce(access); | |||
3539 | graft = isl_schedule_node_from_extension(access); | |||
3540 | ||||
3541 | graft = isl_schedule_node_child(graft, 0); | |||
3542 | ||||
3543 | graft = isl_schedule_node_insert_partial_schedule(graft, mupa); | |||
3544 | if (kernel->options->unroll_copy_shared) | |||
3545 | graft = ppcg_set_schedule_node_type(graft, isl_ast_loop_unroll); | |||
3546 | ||||
3547 | if (tile->n > kernel->n_block && kernel->n_block > 0) { | |||
3548 | graft = isl_schedule_node_band_split(graft, | |||
3549 | tile->n - kernel->n_block); | |||
3550 | graft = isl_schedule_node_child(graft, 0); | |||
3551 | } | |||
3552 | if (tile->n < kernel->n_block) | |||
3553 | skip = kernel->n_block - tile->n; | |||
3554 | else | |||
3555 | skip = 0; | |||
3556 | filter = set_schedule_modulo(graft, kernel->thread_ids, | |||
3557 | kernel->block_dim); | |||
3558 | if (!kernel->options->wrap) | |||
3559 | graft = snap_band_to_sizes(graft, kernel->block_dim + skip, | |||
3560 | kernel->options); | |||
3561 | if (tile->n > kernel->n_block && kernel->n_block > 0) | |||
3562 | graft = isl_schedule_node_parent(graft); | |||
3563 | graft = isl_schedule_node_insert_filter(graft, filter); | |||
3564 | ||||
3565 | while (graft && isl_schedule_node_has_parent(graft)) | |||
3566 | graft = isl_schedule_node_parent(graft); | |||
3567 | ||||
3568 | if (read) { | |||
3569 | if (kernel_depth < tile->depth) | |||
3570 | node = gpu_tree_ensure_sync_after_core(node, kernel); | |||
3571 | node = gpu_tree_move_left_to_sync(node, kernel); | |||
3572 | node = isl_schedule_node_graft_before(node, graft); | |||
3573 | } else { | |||
3574 | node = gpu_tree_move_right_to_sync(node, kernel); | |||
3575 | node = isl_schedule_node_graft_after(node, graft); | |||
3576 | if (kernel_depth < tile->depth) | |||
3577 | node = add_group_write_sync(node, kernel, group, 1); | |||
3578 | } | |||
3579 | ||||
3580 | node = gpu_tree_move_up_to_kernel(node); | |||
3581 | ||||
3582 | return node; | |||
3583 | } | |||
3584 | ||||
3585 | /* Check whether the array reference group "group" is mapped to | |||
3586 | * private or shared memory and, if so, | |||
3587 | * add copy statements to the schedule tree of "node" | |||
3588 | * for reading from global memory to private or shared memory | |||
3589 | * (if "read" is set) or for writing back from private or shared memory | |||
3590 | * to global memory (if "read" is not set) for this group. | |||
3591 | * On input, "node" points to the kernel node, and it is moved | |||
3592 | * back there on output. | |||
3593 | */ | |||
3594 | static __isl_give isl_schedule_node *add_copies_group( | |||
3595 | struct ppcg_kernel *kernel, struct gpu_array_ref_group *group, | |||
3596 | __isl_take isl_schedule_node *node, int read) | |||
3597 | { | |||
3598 | enum ppcg_group_access_type type; | |||
3599 | ||||
3600 | type = gpu_array_ref_group_type(group); | |||
3601 | if (type == ppcg_access_private) | |||
3602 | return add_copies_group_private(kernel, group, node, read); | |||
3603 | if (type == ppcg_access_shared) | |||
3604 | return add_copies_group_shared(kernel, group, node, read); | |||
3605 | return node; | |||
3606 | } | |||
3607 | ||||
3608 | /* For each array reference group that is mapped to private or shared memory, | |||
3609 | * add copy statements to the schedule tree of "node" | |||
3610 | * for reading from global memory to private or shared memory | |||
3611 | * and for writing back. | |||
3612 | * On input, "node" points to the kernel node, and it is moved | |||
3613 | * back there on output. | |||
3614 | */ | |||
3615 | static __isl_give isl_schedule_node *add_copies(struct ppcg_kernel *kernel, | |||
3616 | __isl_take isl_schedule_node *node) | |||
3617 | { | |||
3618 | int i, j; | |||
3619 | ||||
3620 | for (i = 0; i < kernel->n_array; ++i) { | |||
3621 | struct gpu_local_array_info *array = &kernel->array[i]; | |||
3622 | ||||
3623 | for (j = 0; j < array->n_group; ++j) { | |||
3624 | struct gpu_array_ref_group *group = array->groups[j]; | |||
3625 | ||||
3626 | node = add_copies_group(kernel, group, node, 1); | |||
3627 | if (!node) | |||
3628 | return NULL((void*)0); | |||
3629 | node = add_copies_group(kernel, group, node, 0); | |||
3630 | if (!node) | |||
3631 | return NULL((void*)0); | |||
3632 | } | |||
3633 | } | |||
3634 | ||||
3635 | return node; | |||
3636 | } | |||
3637 | ||||
3638 | /* Mark all dimensions in the current band node atomic. | |||
3639 | */ | |||
3640 | static __isl_give isl_schedule_node *atomic(__isl_take isl_schedule_node *node) | |||
3641 | { | |||
3642 | return ppcg_set_schedule_node_type(node, isl_ast_loop_atomic); | |||
3643 | } | |||
3644 | ||||
3645 | /* Mark "node" atomic, if it is a band node. | |||
3646 | * Do the same for all ancestors. | |||
3647 | * Return a pointer to "node" (in the updated schedule tree). | |||
3648 | */ | |||
3649 | static __isl_give isl_schedule_node *atomic_ancestors( | |||
3650 | __isl_take isl_schedule_node *node) | |||
3651 | { | |||
3652 | int pos; | |||
3653 | ||||
3654 | if (!node) | |||
3655 | return NULL((void*)0); | |||
3656 | if (!isl_schedule_node_has_parent(node)) | |||
3657 | return node; | |||
3658 | ||||
3659 | pos = isl_schedule_node_get_child_position(node); | |||
3660 | node = isl_schedule_node_parent(node); | |||
3661 | if (isl_schedule_node_get_type(node) == isl_schedule_node_band) | |||
3662 | node = atomic(node); | |||
3663 | node = atomic_ancestors(node); | |||
3664 | node = isl_schedule_node_child(node, pos); | |||
3665 | ||||
3666 | return node; | |||
3667 | } | |||
3668 | ||||
3669 | /* Collect all write references that require synchronization. | |||
3670 | * "node" is assumed to point to the kernel node. | |||
3671 | * Each reference is represented by a universe set in a space | |||
3672 | * | |||
3673 | * [S[i,j] -> R[]] | |||
3674 | * | |||
3675 | * with S[i,j] the statement instance space and R[] the array reference. | |||
3676 | * | |||
3677 | * This function should be called before block and thread filters are added. | |||
3678 | * | |||
3679 | * Synchronization is needed after a write if there is a subsequent read | |||
3680 | * within the same block that may not be performed by the same thread. | |||
3681 | * There should not be any dependences between different blocks, | |||
3682 | * so we start with the flow dependences within the same kernel invocation | |||
3683 | * and we subtract from these those dependences that are mapped | |||
3684 | * to the same iteration of the bands where synchronization is inserted. | |||
3685 | * We do not remove pairs of instances that are known to map to | |||
3686 | * the same thread across different iterations of the intermediate | |||
3687 | * bands because the read may be performed by a different thread | |||
3688 | * than the one that needs the value if shared memory is involved. | |||
3689 | * | |||
3690 | * We also consider all pairs of possible writes that access the same | |||
3691 | * memory location and that may be mapped to the same block but not | |||
3692 | * to the same iteration of the intermediate bands. | |||
3693 | * In theory, it would be possible for one thread to still be in | |||
3694 | * a previous iteration of a loop in these bands. | |||
3695 | * A write to global memory in this delayed thread could then overwrite | |||
3696 | * a write from another thread that has already moved on to | |||
3697 | * the next iteration. | |||
3698 | * | |||
3699 | * After computing the above writes paired off with reads or writes | |||
3700 | * that depend on them, we project onto the domain writes. | |||
3701 | * Sychronization is needed after writes to global memory | |||
3702 | * through these references. | |||
3703 | */ | |||
3704 | static __isl_give isl_union_set *compute_sync_writes( | |||
3705 | struct ppcg_kernel *kernel, __isl_keep isl_schedule_node *node) | |||
3706 | { | |||
3707 | isl_union_map *local; | |||
3708 | isl_union_map *may_writes, *shared_access; | |||
3709 | isl_union_map *kernel_prefix, *thread_prefix; | |||
3710 | isl_union_map *equal; | |||
3711 | isl_union_set *wrap; | |||
3712 | isl_union_set *domain; | |||
3713 | isl_union_pw_multi_aff *contraction; | |||
3714 | ||||
3715 | kernel_prefix = isl_schedule_node_get_prefix_schedule_union_map(node); | |||
3716 | node = isl_schedule_node_copy(node); | |||
3717 | node = gpu_tree_move_down_to_thread(node, kernel->core); | |||
3718 | thread_prefix = isl_schedule_node_get_prefix_schedule_union_map(node); | |||
3719 | isl_schedule_node_free(node); | |||
3720 | ||||
3721 | contraction = kernel->contraction; | |||
3722 | kernel_prefix = isl_union_map_preimage_domain_union_pw_multi_aff( | |||
3723 | kernel_prefix, isl_union_pw_multi_aff_copy(contraction)); | |||
3724 | thread_prefix = isl_union_map_preimage_domain_union_pw_multi_aff( | |||
3725 | thread_prefix, isl_union_pw_multi_aff_copy(contraction)); | |||
3726 | domain = isl_union_set_copy(kernel->expanded_domain); | |||
3727 | domain = isl_union_set_universe(domain); | |||
3728 | ||||
3729 | may_writes = isl_union_map_copy(kernel->prog->scop->tagged_may_writes); | |||
3730 | may_writes = isl_union_map_curry(may_writes); | |||
3731 | may_writes = isl_union_map_intersect_domain(may_writes, domain); | |||
3732 | may_writes = isl_union_map_uncurry(may_writes); | |||
3733 | shared_access = isl_union_map_copy(may_writes); | |||
3734 | shared_access = isl_union_map_apply_range(shared_access, | |||
3735 | isl_union_map_reverse(may_writes)); | |||
3736 | ||||
3737 | local = isl_union_map_copy(kernel->prog->scop->tagged_dep_flow); | |||
3738 | local = isl_union_map_union(local, shared_access); | |||
3739 | local = isl_union_map_zip(local); | |||
3740 | ||||
3741 | equal = isl_union_map_apply_range(kernel_prefix, | |||
3742 | isl_union_map_reverse(isl_union_map_copy(kernel_prefix))); | |||
3743 | wrap = isl_union_map_wrap(equal); | |||
3744 | local = isl_union_map_intersect_domain(local, wrap); | |||
3745 | equal = isl_union_map_apply_range(thread_prefix, | |||
3746 | isl_union_map_reverse(isl_union_map_copy(thread_prefix))); | |||
3747 | wrap = isl_union_map_wrap(equal); | |||
3748 | local = isl_union_map_subtract_domain(local, wrap); | |||
3749 | ||||
3750 | local = isl_union_map_zip(local); | |||
3751 | local = isl_union_map_universe(local); | |||
3752 | ||||
3753 | return isl_union_map_domain(local); | |||
3754 | } | |||
3755 | ||||
3756 | /* Group the domain elements into a single space, named kernelX, | |||
3757 | * with X the kernel sequence number "kernel_id". | |||
3758 | */ | |||
3759 | static __isl_give isl_schedule_node *group_statements( | |||
3760 | __isl_take isl_schedule_node *node, int kernel_id) | |||
3761 | { | |||
3762 | char buffer[20]; | |||
3763 | isl_id *id; | |||
3764 | ||||
3765 | if (!node) | |||
3766 | return NULL((void*)0); | |||
3767 | ||||
3768 | snprintf(buffer, sizeof(buffer), "kernel%d", kernel_id); | |||
3769 | id = isl_id_alloc(isl_schedule_node_get_ctx(node), buffer, NULL((void*)0)); | |||
3770 | return isl_schedule_node_group(node, id); | |||
3771 | } | |||
3772 | ||||
3773 | /* Create a ppcg_kernel representing the domain instances that reach "node" | |||
3774 | * and insert a mark node pointing to the ppcg_kernel before "node". | |||
3775 | * The band that "node" points to is the band that needs to be mapped | |||
3776 | * to block identifiers. The band that needs to be mapped to thread | |||
3777 | * identifiers should be marked by a "thread" mark by the caller. | |||
3778 | * The linear branch between the current node and the "thread" mark | |||
3779 | * may also have a "shared" mark. If present, the mapping to shared | |||
3780 | * memory is computed at that point. | |||
3781 | * Both marks are removed by this function. | |||
3782 | * If "scale" is set, then the band that "node" points to is scaled | |||
3783 | * by "sizes". | |||
3784 | * | |||
3785 | * Mark all outer band nodes as atomic to ensure each kernel is only | |||
3786 | * scheduled once. | |||
3787 | * If the domain elements that reach "node" live in more than one space, | |||
3788 | * then group the domain elements into a single space, named kernelX, | |||
3789 | * with X the kernel sequence number. | |||
3790 | * | |||
3791 | * Insert a guard node governing the kernel node to ensure that | |||
3792 | * no kernels with zero blocks are launched. | |||
3793 | * | |||
3794 | * Insert a context node describing the block and thread | |||
3795 | * identifiers inside the kernel mark. | |||
3796 | * The context node needs to be inserted after the effective block size | |||
3797 | * has been determined such that the bounds on the thread identifiers | |||
3798 | * would reflect the effective block size. | |||
3799 | * Insert a filter node inside the context node mapping the statement | |||
3800 | * instances to block identifiers. In particular, the block identifiers | |||
3801 | * are equated to the partial schedule of band that was marked for mapping | |||
3802 | * to blocks modulo the grid size. | |||
3803 | * Insert a filter node inside the "thread" mark mapping the statement | |||
3804 | * instances to thread identifiers. In particular, the thread identifiers | |||
3805 | * are equated to the partial schedule of band that was marked for mapping | |||
3806 | * to threads modulo the block size. | |||
3807 | * | |||
3808 | * Compute array reference groups for all arrays, set the local | |||
3809 | * array bounds based on the set of domain instances that reach | |||
3810 | * the kernel node, check the total amount of shared memory used | |||
3811 | * and compute all group tilings. | |||
3812 | * The array reference groups are computed after the block filter | |||
3813 | * has been inserted because it affects the mapping to shared or | |||
3814 | * private memory. This computation also requires the thread filter | |||
3815 | * (in the ppcg_kernel object), but this thread filter should not | |||
3816 | * have been added to the schedule tree yet since the computation | |||
3817 | * requires the schedule of the band that needs to be mapped to | |||
3818 | * threads before the privatization is applied. | |||
3819 | * | |||
3820 | * If any array reference group requires the band mapped to threads | |||
3821 | * to be unrolled, then we perform the required unrolling. | |||
3822 | * | |||
3823 | * We save a copy of the schedule that may influence the mappings | |||
3824 | * to shared or private memory in kernel->copy_schedule. | |||
3825 | * | |||
3826 | * Finally, we add synchronization and copy statements to the schedule tree, | |||
3827 | * remove the "thread" mark and create representations for the local | |||
3828 | * variables in the kernel. | |||
3829 | * | |||
3830 | * We keep a copy of the isl_id that points to the kernel to ensure | |||
3831 | * that the kernel does not get destroyed if the schedule node | |||
3832 | * is freed due to some error condition. | |||
3833 | */ | |||
3834 | __isl_give isl_schedule_node *gpu_create_kernel(struct gpu_gen *gen, | |||
3835 | __isl_take isl_schedule_node *node, int scale, | |||
3836 | __isl_keep isl_multi_val *sizes) | |||
3837 | { | |||
3838 | struct ppcg_kernel *kernel; | |||
3839 | isl_id *id; | |||
3840 | isl_schedule_node *node_thread; | |||
3841 | isl_union_map *host_schedule; | |||
3842 | isl_union_pw_multi_aff *contraction; | |||
3843 | isl_set *host_domain; | |||
3844 | isl_union_set *domain, *expanded; | |||
3845 | int single_statement; | |||
3846 | ||||
3847 | node = gpu_tree_insert_shared_before_thread(node); | |||
3848 | if (!node) | |||
3849 | return NULL((void*)0); | |||
3850 | ||||
3851 | kernel = isl_calloc_type(gen->ctx, struct ppcg_kernel)((struct ppcg_kernel *)isl_calloc_or_die(gen->ctx, 1, sizeof (struct ppcg_kernel))); | |||
3852 | kernel = ppcg_kernel_create_local_arrays(kernel, gen->prog); | |||
3853 | if (!kernel) | |||
3854 | return isl_schedule_node_free(node); | |||
3855 | ||||
3856 | domain = isl_schedule_node_get_domain(node); | |||
3857 | single_statement = isl_union_set_n_set(domain) == 1; | |||
3858 | ||||
3859 | kernel->ctx = gen->ctx; | |||
3860 | kernel->prog = gen->prog; | |||
3861 | kernel->options = gen->options; | |||
3862 | kernel->context = extract_context(node, gen->prog); | |||
3863 | kernel->core = isl_union_set_universe(isl_union_set_copy(domain)); | |||
3864 | contraction = isl_schedule_node_get_subtree_contraction(node); | |||
3865 | kernel->contraction = isl_union_pw_multi_aff_copy(contraction); | |||
3866 | expanded = isl_union_set_copy(domain); | |||
3867 | expanded = isl_union_set_preimage_union_pw_multi_aff(expanded, | |||
3868 | contraction); | |||
3869 | kernel->expanded_domain = isl_union_set_copy(expanded); | |||
3870 | kernel->arrays = accessed_by_domain(expanded, gen->prog); | |||
3871 | kernel->n_grid = n_outer_coincidence(node); | |||
3872 | node_thread = isl_schedule_node_copy(node); | |||
3873 | node_thread = gpu_tree_move_down_to_thread(node_thread, kernel->core); | |||
3874 | node_thread = isl_schedule_node_child(node_thread, 0); | |||
3875 | kernel->n_block = n_outer_coincidence(node_thread); | |||
3876 | isl_schedule_node_free(node_thread); | |||
3877 | kernel->id = gen->kernel_id++; | |||
3878 | read_grid_and_block_sizes(kernel, gen); | |||
3879 | ||||
3880 | kernel->sync_writes = compute_sync_writes(kernel, node); | |||
3881 | ||||
3882 | host_schedule = isl_schedule_node_get_prefix_schedule_union_map(node); | |||
3883 | host_domain = isl_set_from_union_set(isl_union_map_range( | |||
3884 | host_schedule)); | |||
3885 | ||||
3886 | node = atomic_ancestors(node); | |||
3887 | ||||
3888 | id = isl_id_alloc(gen->ctx, "kernel", kernel); | |||
3889 | id = isl_id_set_free_user(id, &ppcg_kernel_free_wrap); | |||
3890 | node = isl_schedule_node_insert_mark(node, isl_id_copy(id)); | |||
3891 | ||||
3892 | if (!single_statement) | |||
3893 | node = group_statements(node, kernel->id); | |||
3894 | ||||
3895 | node = isl_schedule_node_child(node, 0); | |||
3896 | node = split_band(node, kernel->n_grid); | |||
3897 | kernel->block_ids = ppcg_scop_generate_names(gen->prog->scop, | |||
3898 | kernel->n_grid, "b"); | |||
3899 | kernel->block_filter = set_schedule_modulo(node, kernel->block_ids, | |||
3900 | kernel->grid_dim); | |||
3901 | kernel->grid_size = extract_grid_size(kernel, | |||
3902 | isl_union_set_copy(domain)); | |||
3903 | if (!kernel->options->wrap) | |||
3904 | node = snap_band_to_sizes(node, kernel->grid_dim, | |||
3905 | kernel->options); | |||
3906 | if (scale) | |||
3907 | node = scale_band(node, isl_multi_val_copy(sizes)); | |||
3908 | node = isl_schedule_node_parent(node); | |||
3909 | if (!single_statement) | |||
3910 | node = isl_schedule_node_parent(node); | |||
3911 | node = insert_guard(node, kernel->context, kernel->grid_size, | |||
3912 | gen->prog->scop); | |||
3913 | node = gpu_tree_move_down_to_thread(node, kernel->core); | |||
3914 | node = isl_schedule_node_child(node, 0); | |||
3915 | node = split_band(node, kernel->n_block); | |||
3916 | kernel->thread_ids = ppcg_scop_generate_names(gen->prog->scop, | |||
3917 | kernel->n_block, "t"); | |||
3918 | kernel->thread_filter = set_schedule_modulo(node, kernel->thread_ids, | |||
3919 | kernel->block_dim); | |||
3920 | if (extract_block_size(kernel, domain) < 0) | |||
3921 | node = isl_schedule_node_free(node); | |||
3922 | ||||
3923 | node = gpu_tree_move_up_to_kernel(node); | |||
3924 | node = isl_schedule_node_child(node, 0); | |||
3925 | node = insert_context(kernel, node); | |||
3926 | node = isl_schedule_node_child(node, 0); | |||
3927 | node = isl_schedule_node_insert_filter(node, | |||
3928 | isl_union_set_copy(kernel->block_filter)); | |||
3929 | ||||
3930 | node = gpu_tree_move_up_to_kernel(node); | |||
3931 | ||||
3932 | if (gpu_group_references(kernel, node) < 0) | |||
3933 | node = isl_schedule_node_free(node); | |||
3934 | localize_bounds(kernel, host_domain); | |||
3935 | isl_set_free(host_domain); | |||
3936 | ||||
3937 | check_shared_memory_bound(kernel); | |||
3938 | mark_global_arrays(kernel); | |||
3939 | compute_group_tilings(kernel); | |||
3940 | ||||
3941 | node = gpu_tree_move_down_to_thread(node, kernel->core); | |||
3942 | node = isl_schedule_node_child(node, 0); | |||
3943 | if (!kernel->options->wrap) | |||
3944 | node = snap_band_to_sizes(node, kernel->block_dim, | |||
3945 | kernel->options); | |||
3946 | node = isl_schedule_node_insert_filter(node, | |||
3947 | isl_union_set_copy(kernel->thread_filter)); | |||
3948 | if (kernel_requires_unroll(kernel)) { | |||
3949 | node = isl_schedule_node_child(node, 0); | |||
3950 | node = unroll(node); | |||
3951 | } | |||
3952 | ||||
3953 | node = gpu_tree_move_up_to_thread(node); | |||
3954 | kernel->copy_schedule_dim = isl_schedule_node_get_schedule_depth(node); | |||
3955 | kernel->copy_schedule = | |||
3956 | isl_schedule_node_get_prefix_schedule_union_pw_multi_aff(node); | |||
3957 | contraction = isl_union_pw_multi_aff_copy(kernel->contraction); | |||
3958 | kernel->copy_schedule = | |||
3959 | isl_union_pw_multi_aff_pullback_union_pw_multi_aff( | |||
3960 | kernel->copy_schedule, contraction); | |||
3961 | ||||
3962 | node = gpu_tree_move_up_to_kernel(node); | |||
3963 | ||||
3964 | node = add_sync(kernel, node); | |||
3965 | node = add_copies(kernel, node); | |||
3966 | ||||
3967 | node = gpu_tree_move_down_to_shared(node, kernel->core); | |||
3968 | node = isl_schedule_node_delete(node); | |||
3969 | ||||
3970 | node = gpu_tree_move_down_to_thread(node, kernel->core); | |||
3971 | node = isl_schedule_node_delete(node); | |||
3972 | ||||
3973 | node = gpu_tree_move_up_to_kernel(node); | |||
3974 | ||||
3975 | if (create_kernel_vars(kernel) < 0) | |||
3976 | node = isl_schedule_node_free(node); | |||
3977 | ||||
3978 | if (!single_statement) | |||
3979 | node = isl_schedule_node_parent(node); | |||
3980 | node = isl_schedule_node_parent(node); | |||
3981 | ||||
3982 | isl_id_free(id); | |||
3983 | return node; | |||
3984 | } | |||
3985 | ||||
3986 | /* Insert a zero-dimensional permutable band at "node". | |||
3987 | */ | |||
3988 | static __isl_give isl_schedule_node *insert_empty_permutable_band( | |||
3989 | __isl_take isl_schedule_node *node) | |||
3990 | { | |||
3991 | isl_space *space; | |||
3992 | isl_schedule *schedule; | |||
3993 | isl_union_set *domain; | |||
3994 | isl_multi_union_pw_aff *mupa; | |||
3995 | ||||
3996 | schedule = isl_schedule_node_get_schedule(node); | |||
3997 | domain = isl_schedule_get_domain(schedule); | |||
3998 | space = isl_union_set_get_space(domain); | |||
3999 | isl_union_set_free(domain); | |||
4000 | isl_schedule_free(schedule); | |||
4001 | ||||
4002 | space = isl_space_set_from_params(space); | |||
4003 | mupa = isl_multi_union_pw_aff_zero(space); | |||
4004 | node = isl_schedule_node_insert_partial_schedule(node, mupa); | |||
4005 | node = isl_schedule_node_band_set_permutable(node, 1); | |||
4006 | ||||
4007 | return node; | |||
4008 | } | |||
4009 | ||||
4010 | /* See if hybrid tiling can be performed on "node" and its parent. | |||
4011 | * If so, apply hybrid tiling and return the updated schedule tree. | |||
4012 | * If not, return the original schedule tree. | |||
4013 | * Return NULL on error. | |||
4014 | * | |||
4015 | * First check if "node", together with its parent, meets | |||
4016 | * the basic requirements for hybrid tiling. | |||
4017 | * If so, compute the relative dependence distances of "node" | |||
4018 | * with respect to its parent and check if they are sufficiently bounded. | |||
4019 | * If so, apply hybrid tiling using user specified tile sizes. | |||
4020 | * | |||
4021 | * The tile sizes are read before the dependence distance bounds are | |||
4022 | * computed, because the user may have specified fewer dimensions | |||
4023 | * than are available. In this case, the remaining schedule dimensions | |||
4024 | * are split off and the dependence distances should be computed | |||
4025 | * after these dimensions have been split off. | |||
4026 | */ | |||
4027 | static __isl_give isl_schedule_node *try_hybrid_tile(struct gpu_gen *gen, | |||
4028 | __isl_take isl_schedule_node *node) | |||
4029 | { | |||
4030 | int tile_len; | |||
4031 | int *tile_size; | |||
4032 | isl_bool ok; | |||
4033 | isl_schedule_node *orig = node; | |||
4034 | ppcg_ht_bounds *bounds; | |||
4035 | ||||
4036 | ok = ppcg_ht_parent_has_input_pattern(node); | |||
4037 | if (ok < 0) | |||
4038 | return isl_schedule_node_free(node); | |||
4039 | if (!ok) | |||
4040 | return orig; | |||
4041 | ||||
4042 | tile_len = 1 + isl_schedule_node_band_n_member(node); | |||
4043 | tile_size = read_tile_sizes(gen, &tile_len); | |||
4044 | if (!tile_size) | |||
4045 | return isl_schedule_node_free(node); | |||
4046 | ||||
4047 | node = isl_schedule_node_copy(node); | |||
4048 | node = split_band(node, tile_len - 1); | |||
4049 | node = isl_schedule_node_parent(node); | |||
4050 | bounds = ppcg_ht_compute_bounds(gen->prog->scop, node); | |||
4051 | node = isl_schedule_node_child(node, 0); | |||
4052 | ||||
4053 | ok = ppcg_ht_bounds_is_valid(bounds); | |||
4054 | if (ok >= 0 && ok) | |||
4055 | node = gpu_hybrid_tile(gen, node, bounds, tile_size); | |||
4056 | else | |||
4057 | ppcg_ht_bounds_free(bounds); | |||
4058 | free(tile_size); | |||
4059 | ||||
4060 | if (ok >= 0 && !ok) { | |||
4061 | isl_schedule_node_free(node); | |||
4062 | return orig; | |||
4063 | } | |||
4064 | isl_schedule_node_free(orig); | |||
4065 | if (ok < 0) | |||
4066 | return isl_schedule_node_free(node); | |||
4067 | return node; | |||
4068 | } | |||
4069 | ||||
4070 | /* If "node" is the outermost permutable band that can be mapped to block and | |||
4071 | * thread identifiers in its branch (or the root of a subtree with | |||
4072 | * no such outer bands), | |||
4073 | * then mark the band as such, attaching a ppcg_kernel to the mark. | |||
4074 | * | |||
4075 | * If hybrid tiling is allowed, then first try and apply it | |||
4076 | * to "node" and its parent. | |||
4077 | * | |||
4078 | * If "node" is the root of a subtree without permutable bands, | |||
4079 | * then insert a zero-dimensional permutable band such that | |||
4080 | * we can assume that "node" always points to a band node. | |||
4081 | * This includes the case where "node" already points to a band node, | |||
4082 | * but one without any coincident dimension. In this case, | |||
4083 | * the extra node ensures that this original node does not get tiled. | |||
4084 | * | |||
4085 | * Tile "node" using user specified tile sizes, after splitting the band | |||
4086 | * if the number of specified tile sizes is smaller than the dimension | |||
4087 | * of the band. Mark the point band of this tiling as the band that | |||
4088 | * needs to be mapped to threads and instruct the AST generator to unroll | |||
4089 | * the band if the "unroll_gpu_tile" option is set. | |||
4090 | * Create a kernel representing the domain instances that reach "node" and | |||
4091 | * insert a mark node pointing to the ppcg_kernel before the band node. | |||
4092 | */ | |||
4093 | static __isl_give isl_schedule_node *mark_outer_permutable( | |||
4094 | __isl_take isl_schedule_node *node, void *user) | |||
4095 | { | |||
4096 | struct gpu_gen *gen = user; | |||
4097 | int outer; | |||
4098 | int scale; | |||
4099 | int tile_len; | |||
4100 | int *tile_size; | |||
4101 | isl_id *id; | |||
4102 | isl_multi_val *sizes; | |||
4103 | ||||
4104 | outer = is_outer_tilable(node); | |||
4105 | if (outer < 0) | |||
| ||||
4106 | return isl_schedule_node_free(node); | |||
4107 | if (!outer) | |||
4108 | return node; | |||
4109 | ||||
4110 | if (gen->options->hybrid) { | |||
4111 | isl_schedule_node *saved = isl_schedule_node_copy(node); | |||
4112 | node = try_hybrid_tile(gen, node); | |||
4113 | isl_schedule_node_free(saved); | |||
4114 | if (node != saved) | |||
4115 | return node; | |||
4116 | } | |||
4117 | ||||
4118 | if (isl_schedule_node_get_type(node) != isl_schedule_node_band || | |||
4119 | !isl_schedule_node_band_member_get_coincident(node, 0)) | |||
4120 | node = insert_empty_permutable_band(node); | |||
4121 | ||||
4122 | tile_len = isl_schedule_node_band_n_member(node); | |||
4123 | tile_size = read_tile_sizes(gen, &tile_len); | |||
4124 | if (!tile_size) | |||
4125 | return isl_schedule_node_free(node); | |||
4126 | if (tile_len < isl_schedule_node_band_n_member(node)) | |||
4127 | node = isl_schedule_node_band_split(node, tile_len); | |||
4128 | sizes = construct_band_tiles_sizes(node, tile_size); | |||
4129 | node = tile_band(node, isl_multi_val_copy(sizes)); | |||
4130 | node = isl_schedule_node_child(node, 0); | |||
4131 | if (gen->options->unroll_gpu_tile) | |||
4132 | node = ppcg_set_schedule_node_type(node, isl_ast_loop_unroll); | |||
4133 | id = isl_id_alloc(gen->ctx, "thread", NULL((void*)0)); | |||
4134 | node = isl_schedule_node_insert_mark(node, id); | |||
4135 | node = isl_schedule_node_parent(node); | |||
4136 | ||||
4137 | scale = gen->options->scale_tile_loops; | |||
4138 | node = gpu_create_kernel(gen, node, scale, sizes); | |||
4139 | isl_multi_val_free(sizes); | |||
4140 | free(tile_size); | |||
4141 | ||||
4142 | return node; | |||
4143 | } | |||
4144 | ||||
4145 | /* Given a set or sequence node, return the union the filters of either all | |||
4146 | * (if "only_initial" is not set) or the initial (if "only_initial" is set) | |||
4147 | * direct subtrees that do not contain any suitably permutable bands | |||
4148 | * (according to subtree_has_permutable_bands). | |||
4149 | */ | |||
4150 | static __isl_give isl_union_set *get_non_parallel_subtree_filters( | |||
4151 | __isl_keep isl_schedule_node *node, int only_initial) | |||
4152 | { | |||
4153 | isl_space *space; | |||
4154 | isl_union_set *filter; | |||
4155 | int i, n; | |||
4156 | ||||
4157 | n = isl_schedule_node_n_children(node); | |||
4158 | if (n < 0) | |||
4159 | return NULL((void*)0); | |||
4160 | ||||
4161 | node = isl_schedule_node_copy(node); | |||
4162 | node = isl_schedule_node_child(node, 0); | |||
4163 | filter = isl_schedule_node_filter_get_filter(node); | |||
4164 | node = isl_schedule_node_parent(node); | |||
4165 | space = isl_union_set_get_space(filter); | |||
4166 | isl_union_set_free(filter); | |||
4167 | filter = isl_union_set_empty(space); | |||
4168 | ||||
4169 | for (i = 0; i < n; ++i) { | |||
4170 | int parallelism; | |||
4171 | ||||
4172 | node = isl_schedule_node_child(node, i); | |||
4173 | parallelism = subtree_has_permutable_bands(node); | |||
4174 | if (parallelism < 0) { | |||
4175 | filter = isl_union_set_free(filter); | |||
4176 | } else if (!parallelism) { | |||
4177 | isl_union_set *filter_i; | |||
4178 | filter_i = isl_schedule_node_filter_get_filter(node); | |||
4179 | filter = isl_union_set_union(filter, filter_i); | |||
4180 | } else if (only_initial) | |||
4181 | break; | |||
4182 | node = isl_schedule_node_parent(node); | |||
4183 | } | |||
4184 | ||||
4185 | isl_schedule_node_free(node); | |||
4186 | ||||
4187 | return filter; | |||
4188 | } | |||
4189 | ||||
4190 | /* Given a set or sequence node, return the union of the filters of | |||
4191 | * the direct subtrees that do not contain any suitably permutable bands | |||
4192 | * (according to subtree_has_permutable_bands). | |||
4193 | */ | |||
4194 | static __isl_give isl_union_set *get_all_non_parallel_subtree_filters( | |||
4195 | __isl_keep isl_schedule_node *node) | |||
4196 | { | |||
4197 | return get_non_parallel_subtree_filters(node, 0); | |||
4198 | } | |||
4199 | ||||
4200 | /* Given a set or sequence node, return the union of the filters of | |||
4201 | * the initial direct subtrees that do not contain any suitably permutable | |||
4202 | * bands (according to subtree_has_permutable_bands). | |||
4203 | */ | |||
4204 | static __isl_give isl_union_set *get_initial_non_parallel_subtree_filters( | |||
4205 | __isl_keep isl_schedule_node *node) | |||
4206 | { | |||
4207 | return get_non_parallel_subtree_filters(node, 1); | |||
4208 | } | |||
4209 | ||||
4210 | /* Mark all variables that are accessed by the statement instances in "domain" | |||
4211 | * and that are local to "prog" as requiring a declaration in the host code. | |||
4212 | * The statement instances in "domain" correspond to (a subset of) | |||
4213 | * the active instances at "node". | |||
4214 | * "node" is not modified by this function, except that NULL is returned | |||
4215 | * in case of error. | |||
4216 | */ | |||
4217 | static __isl_give isl_schedule_node *declare_accessed_local_variables( | |||
4218 | __isl_take isl_schedule_node *node, struct gpu_prog *prog, | |||
4219 | __isl_keep isl_union_set *domain) | |||
4220 | { | |||
4221 | isl_union_pw_multi_aff *contraction; | |||
4222 | isl_union_set *arrays; | |||
4223 | int i; | |||
4224 | ||||
4225 | if (!ppcg_scop_any_hidden_declarations(prog->scop)) | |||
4226 | return node; | |||
4227 | contraction = isl_schedule_node_get_subtree_contraction(node); | |||
4228 | domain = isl_union_set_copy(domain); | |||
4229 | domain = isl_union_set_preimage_union_pw_multi_aff(domain, contraction); | |||
4230 | arrays = accessed_by_domain(domain, prog); | |||
4231 | ||||
4232 | for (i = 0; i < prog->n_array; ++i) { | |||
4233 | isl_space *space; | |||
4234 | isl_set *set; | |||
4235 | int empty; | |||
4236 | ||||
4237 | if (!prog->array[i].local) | |||
4238 | continue; | |||
4239 | space = isl_set_get_space(prog->array[i].extent); | |||
4240 | set = isl_union_set_extract_set(arrays, space); | |||
4241 | empty = isl_set_plain_is_empty(set); | |||
4242 | isl_set_free(set); | |||
4243 | if (empty < 0) | |||
4244 | goto error; | |||
4245 | if (!empty) | |||
4246 | prog->array[i].declare_local = 1; | |||
4247 | } | |||
4248 | ||||
4249 | isl_union_set_free(arrays); | |||
4250 | return node; | |||
4251 | error: | |||
4252 | isl_union_set_free(arrays); | |||
4253 | return isl_schedule_node_free(node); | |||
4254 | } | |||
4255 | ||||
4256 | /* If "node" points to a set node, then separate its children | |||
4257 | * into subtrees that have suitably permutable bands and | |||
4258 | * those that do not. | |||
4259 | * Adjust the schedule tree in order to execute the second group | |||
4260 | * after the first group and return a pointer to the first group, | |||
4261 | * assuming there are any such subtrees. | |||
4262 | * If "node" points to a sequence node, then separate the initial | |||
4263 | * children that do not have suitably permutable bands and | |||
4264 | * return a pointer to the subsequence of children that do have such bands, | |||
4265 | * assuming there are any such subtrees. | |||
4266 | * | |||
4267 | * In both cases, mark all local variables in "prog" that are accessed by | |||
4268 | * the group without permutable bands as requiring a declaration on the host. | |||
4269 | */ | |||
4270 | static __isl_give isl_schedule_node *isolate_permutable_subtrees( | |||
4271 | __isl_take isl_schedule_node *node, struct gpu_prog *prog) | |||
4272 | { | |||
4273 | isl_union_set *filter; | |||
4274 | enum isl_schedule_node_type type; | |||
4275 | ||||
4276 | if (!node) | |||
4277 | return NULL((void*)0); | |||
4278 | type = isl_schedule_node_get_type(node); | |||
4279 | if (type == isl_schedule_node_set) { | |||
4280 | filter = get_all_non_parallel_subtree_filters(node); | |||
4281 | node = declare_accessed_local_variables(node, prog, filter); | |||
4282 | node = isl_schedule_node_order_after(node, filter); | |||
4283 | } else if (type == isl_schedule_node_sequence) { | |||
4284 | filter = get_initial_non_parallel_subtree_filters(node); | |||
4285 | node = declare_accessed_local_variables(node, prog, filter); | |||
4286 | node = isl_schedule_node_order_before(node, filter); | |||
4287 | } | |||
4288 | ||||
4289 | return node; | |||
4290 | } | |||
4291 | ||||
4292 | /* Replace any reference to an array element in the range of "copy" | |||
4293 | * by a reference to all array elements (defined by the extent of the array). | |||
4294 | */ | |||
4295 | static __isl_give isl_union_map *approximate_copy_out( | |||
4296 | __isl_take isl_union_map *copy, struct gpu_prog *prog) | |||
4297 | { | |||
4298 | int i; | |||
4299 | isl_union_map *res; | |||
4300 | ||||
4301 | res = isl_union_map_empty(isl_union_map_get_space(copy)); | |||
4302 | ||||
4303 | for (i = 0; i < prog->n_array; ++i) { | |||
4304 | isl_space *space; | |||
4305 | isl_set *set; | |||
4306 | isl_union_map *copy_i; | |||
4307 | isl_union_set *extent, *domain; | |||
4308 | ||||
4309 | space = isl_space_copy(prog->array[i].space); | |||
4310 | extent = isl_union_set_from_set(isl_set_universe(space)); | |||
4311 | copy_i = isl_union_map_copy(copy); | |||
4312 | copy_i = isl_union_map_intersect_range(copy_i, extent); | |||
4313 | set = isl_set_copy(prog->array[i].extent); | |||
4314 | extent = isl_union_set_from_set(set); | |||
4315 | domain = isl_union_map_domain(copy_i); | |||
4316 | copy_i = isl_union_map_from_domain_and_range(domain, extent); | |||
4317 | res = isl_union_map_union(res, copy_i); | |||
4318 | } | |||
4319 | ||||
4320 | isl_union_map_free(copy); | |||
4321 | ||||
4322 | return res; | |||
4323 | } | |||
4324 | ||||
4325 | /* Insert "kernel" marks that point to a ppcg_kernel structure | |||
4326 | * in front of all outermost tilable band that (by construction) | |||
4327 | * have at least one parallel loop. | |||
4328 | */ | |||
4329 | static __isl_give isl_schedule_node *mark_kernels(struct gpu_gen *gen, | |||
4330 | __isl_take isl_schedule_node *node) | |||
4331 | { | |||
4332 | return isl_schedule_node_map_descendant_bottom_up(node, | |||
4333 | &mark_outer_permutable, gen); | |||
4334 | } | |||
4335 | ||||
4336 | /* Construct schedule constraints from the dependences in prog->scop and | |||
4337 | * the array order dependences in prog->array_order. | |||
4338 | * | |||
4339 | * If live range reordering is allowed, then we need to make sure | |||
4340 | * that live ranges on arrays are not run in parallel since doing | |||
4341 | * so would require array expansion. We therefore add the array | |||
4342 | * order dependences to the coincidence dependences. Non-zero array | |||
4343 | * order dependences will then prevent a schedule dimension from being | |||
4344 | * considered parallel. | |||
4345 | * Live ranges derived from scalars are allowed to be run in parallel | |||
4346 | * since we force the scalars to be mapped to private memory in | |||
4347 | * check_scalar_live_ranges. | |||
4348 | * If live range reordering is allowed, then the false dependences | |||
4349 | * are not added to the validity constraints as that would prevent | |||
4350 | * reordering. Instead, the external false dependences that enforce that reads | |||
4351 | * from potentially live-in data precede any later write and | |||
4352 | * that writes of potentially live-out data follow any other earlier write | |||
4353 | * are added to the validity and the coincidence constraints. | |||
4354 | * The false dependences are still added to the proximity constraints | |||
4355 | * for consistency with the case where live range reordering is not allowed. | |||
4356 | * The coincidence constraints then consist of flow dependences, | |||
4357 | * external false dependences and array order dependences. | |||
4358 | * The independences can be filtered out from the first two sets. | |||
4359 | * They have already been filtered out from the array order dependences | |||
4360 | * on a per array basis in collect_order_dependences. | |||
4361 | * There is no need for a per array handling of the other two sets | |||
4362 | * as there should be no flow or external false dependence on local | |||
4363 | * variables that can be filtered out. | |||
4364 | */ | |||
4365 | static __isl_give isl_schedule_constraints *construct_schedule_constraints( | |||
4366 | struct gpu_prog *prog) | |||
4367 | { | |||
4368 | isl_union_set *domain; | |||
4369 | isl_union_map *dep_raw, *dep; | |||
4370 | isl_union_map *validity, *proximity, *coincidence; | |||
4371 | isl_schedule_constraints *sc; | |||
4372 | ||||
4373 | domain = isl_union_set_copy(prog->scop->domain); | |||
4374 | sc = isl_schedule_constraints_on_domain(domain); | |||
4375 | sc = isl_schedule_constraints_set_context(sc, | |||
4376 | isl_set_copy(prog->scop->context)); | |||
4377 | if (prog->scop->options->live_range_reordering) { | |||
4378 | sc = isl_schedule_constraints_set_conditional_validity(sc, | |||
4379 | isl_union_map_copy(prog->scop->tagged_dep_flow), | |||
4380 | isl_union_map_copy(prog->scop->tagged_dep_order)); | |||
4381 | proximity = isl_union_map_copy(prog->scop->dep_flow); | |||
4382 | validity = isl_union_map_copy(proximity); | |||
4383 | validity = isl_union_map_union(validity, | |||
4384 | isl_union_map_copy(prog->scop->dep_forced)); | |||
4385 | proximity = isl_union_map_union(proximity, | |||
4386 | isl_union_map_copy(prog->scop->dep_false)); | |||
4387 | coincidence = isl_union_map_copy(validity); | |||
4388 | coincidence = isl_union_map_subtract(coincidence, | |||
4389 | isl_union_map_copy(prog->scop->independence)); | |||
4390 | coincidence = isl_union_map_union(coincidence, | |||
4391 | isl_union_map_copy(prog->array_order)); | |||
4392 | } else { | |||
4393 | dep_raw = isl_union_map_copy(prog->scop->dep_flow); | |||
4394 | dep = isl_union_map_copy(prog->scop->dep_false); | |||
4395 | dep = isl_union_map_union(dep, dep_raw); | |||
4396 | dep = isl_union_map_coalesce(dep); | |||
4397 | proximity = isl_union_map_copy(dep); | |||
4398 | coincidence = isl_union_map_copy(dep); | |||
4399 | validity = dep; | |||
4400 | } | |||
4401 | sc = isl_schedule_constraints_set_validity(sc, validity); | |||
4402 | sc = isl_schedule_constraints_set_coincidence(sc, coincidence); | |||
4403 | sc = isl_schedule_constraints_set_proximity(sc, proximity); | |||
4404 | ||||
4405 | if (prog->scop->options->debug->dump_schedule_constraints) | |||
4406 | isl_schedule_constraints_dump(sc); | |||
4407 | return sc; | |||
4408 | } | |||
4409 | ||||
4410 | /* Compute an appropriate schedule based on the accesses in | |||
4411 | * gen->read and gen->write. | |||
4412 | * | |||
4413 | * We derive schedule constraints from the dependences in gen->prog->scop | |||
4414 | * and then use isl to compute a schedule that has a parallel loop | |||
4415 | * in each tilable band. | |||
4416 | * During the schedule construction, some statement instances | |||
4417 | * may be grouped first based on the input schedule. | |||
4418 | */ | |||
4419 | static __isl_give isl_schedule *compute_schedule(struct gpu_gen *gen) | |||
4420 | { | |||
4421 | isl_schedule_constraints *sc; | |||
4422 | isl_schedule *schedule; | |||
4423 | ||||
4424 | sc = construct_schedule_constraints(gen->prog); | |||
4425 | schedule = gen->prog->scop->schedule; | |||
4426 | schedule = ppcg_compute_schedule(sc, schedule, gen->options); | |||
4427 | ||||
4428 | return schedule; | |||
4429 | } | |||
4430 | ||||
4431 | /* If the band node "node" has exactly one member then mark it permutable. | |||
4432 | */ | |||
4433 | static __isl_give isl_schedule_node *band_set_permutable( | |||
4434 | __isl_take isl_schedule_node *node, | |||
4435 | __isl_keep isl_schedule_constraints *sc) | |||
4436 | { | |||
4437 | if (isl_schedule_node_band_n_member(node) == 1) | |||
4438 | node = isl_schedule_node_band_set_permutable(node, 1); | |||
4439 | ||||
4440 | return node; | |||
4441 | } | |||
4442 | ||||
4443 | /* Return the coincidence constraints between pairs of instances | |||
4444 | * that are scheduled together by the ancestors of "node". | |||
4445 | * That is, select those coincidence constraints that relate | |||
4446 | * pairs of instances that have the same value for the prefix schedule. | |||
4447 | * If the schedule depth is zero, then the prefix schedule does not | |||
4448 | * contain any information, so we intersect domain and range | |||
4449 | * of the schedule constraints with the reaching domain elements instead. | |||
4450 | */ | |||
4451 | static __isl_give isl_union_map *get_local_coincidence( | |||
4452 | __isl_keep isl_schedule_node *node, | |||
4453 | __isl_keep isl_schedule_constraints *sc) | |||
4454 | { | |||
4455 | isl_union_map *coincidence; | |||
4456 | isl_multi_union_pw_aff *prefix; | |||
4457 | isl_union_pw_multi_aff *contraction; | |||
4458 | ||||
4459 | coincidence = isl_schedule_constraints_get_coincidence(sc); | |||
4460 | contraction = isl_schedule_node_get_subtree_contraction(node); | |||
4461 | if (isl_schedule_node_get_schedule_depth(node) == 0) { | |||
4462 | isl_union_set *domain; | |||
4463 | ||||
4464 | domain = isl_schedule_node_get_domain(node); | |||
4465 | domain = isl_union_set_preimage_union_pw_multi_aff(domain, | |||
4466 | contraction); | |||
4467 | coincidence = isl_union_map_intersect_domain(coincidence, | |||
4468 | isl_union_set_copy(domain)); | |||
4469 | coincidence = isl_union_map_intersect_range(coincidence, | |||
4470 | domain); | |||
4471 | return coincidence; | |||
4472 | } | |||
4473 | ||||
4474 | prefix = isl_schedule_node_get_prefix_schedule_multi_union_pw_aff(node); | |||
4475 | prefix = isl_multi_union_pw_aff_pullback_union_pw_multi_aff(prefix, | |||
4476 | contraction); | |||
4477 | return isl_union_map_eq_at_multi_union_pw_aff(coincidence, prefix); | |||
4478 | } | |||
4479 | ||||
4480 | /* For each member in the band node "node", determine whether | |||
4481 | * it is coincident with respect to the outer nodes and mark | |||
4482 | * it accordingly. | |||
4483 | * | |||
4484 | * That is, for each coincidence constraint between pairs | |||
4485 | * of instances that are scheduled together by the outer nodes, | |||
4486 | * check that domain and range are assigned the same value | |||
4487 | * by the band member. This test is performed by checking | |||
4488 | * that imposing the same value for the band member does not | |||
4489 | * remove any elements from the set of coincidence constraints. | |||
4490 | */ | |||
4491 | static __isl_give isl_schedule_node *band_set_coincident( | |||
4492 | __isl_take isl_schedule_node *node, | |||
4493 | __isl_keep isl_schedule_constraints *sc) | |||
4494 | { | |||
4495 | isl_union_map *coincidence; | |||
4496 | isl_union_pw_multi_aff *contraction; | |||
4497 | isl_multi_union_pw_aff *partial; | |||
4498 | int i, n; | |||
4499 | ||||
4500 | coincidence = get_local_coincidence(node, sc); | |||
4501 | ||||
4502 | partial = isl_schedule_node_band_get_partial_schedule(node); | |||
4503 | contraction = isl_schedule_node_get_subtree_contraction(node); | |||
4504 | partial = isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial, | |||
4505 | contraction); | |||
4506 | n = isl_schedule_node_band_n_member(node); | |||
4507 | for (i = 0; i < n; ++i) { | |||
4508 | isl_union_map *coincidence_i; | |||
4509 | isl_union_pw_aff *upa; | |||
4510 | isl_multi_union_pw_aff *partial_i; | |||
4511 | int subset; | |||
4512 | ||||
4513 | upa = isl_multi_union_pw_aff_get_union_pw_aff(partial, i); | |||
4514 | partial_i = isl_multi_union_pw_aff_from_union_pw_aff(upa); | |||
4515 | coincidence_i = isl_union_map_copy(coincidence); | |||
4516 | coincidence_i = isl_union_map_eq_at_multi_union_pw_aff( | |||
4517 | coincidence_i, partial_i); | |||
4518 | subset = isl_union_map_is_subset(coincidence, coincidence_i); | |||
4519 | isl_union_map_free(coincidence_i); | |||
4520 | ||||
4521 | if (subset < 0) | |||
4522 | break; | |||
4523 | node = isl_schedule_node_band_member_set_coincident(node, i, | |||
4524 | subset); | |||
4525 | } | |||
4526 | if (i < n) | |||
4527 | node = isl_schedule_node_free(node); | |||
4528 | isl_multi_union_pw_aff_free(partial); | |||
4529 | isl_union_map_free(coincidence); | |||
4530 | ||||
4531 | return node; | |||
4532 | } | |||
4533 | ||||
4534 | /* If "node" is a band, then set its properties. | |||
4535 | * | |||
4536 | * In particular, if the band has exactly one member, then mark it permutable. | |||
4537 | * Mark the band member coincident based on the coincidence constraints | |||
4538 | * of "sc". | |||
4539 | */ | |||
4540 | static __isl_give isl_schedule_node *set_band_properties( | |||
4541 | __isl_take isl_schedule_node *node, void *user) | |||
4542 | { | |||
4543 | isl_schedule_constraints *sc = user; | |||
4544 | ||||
4545 | if (isl_schedule_node_get_type(node) != isl_schedule_node_band) | |||
4546 | return node; | |||
4547 | if (isl_schedule_node_band_n_member(node) == 0) | |||
4548 | return node; | |||
4549 | ||||
4550 | node = band_set_permutable(node, sc); | |||
4551 | node = band_set_coincident(node, sc); | |||
4552 | ||||
4553 | return node; | |||
4554 | } | |||
4555 | ||||
4556 | /* Return the original schedule with all bands marked permutable and | |||
4557 | * all band members marked coincident based on the coincidence constraints. | |||
4558 | * The bands are explicitly marked permutable so that they will be considered | |||
4559 | * by mark_outer_permutable. | |||
4560 | */ | |||
4561 | static __isl_give isl_schedule *determine_properties_original_schedule( | |||
4562 | struct gpu_gen *gen) | |||
4563 | { | |||
4564 | isl_schedule *schedule; | |||
4565 | isl_schedule_constraints *sc; | |||
4566 | ||||
4567 | schedule = isl_schedule_copy(gen->prog->scop->schedule); | |||
4568 | sc = construct_schedule_constraints(gen->prog); | |||
4569 | schedule = isl_schedule_map_schedule_node_bottom_up(schedule, | |||
4570 | &set_band_properties, sc); | |||
4571 | isl_schedule_constraints_free(sc); | |||
4572 | ||||
4573 | return schedule; | |||
4574 | } | |||
4575 | ||||
4576 | /* Compute a schedule or determine the properties of the original schedule | |||
4577 | * depending on the value of the "reschedule" option. | |||
4578 | */ | |||
4579 | static __isl_give isl_schedule *compute_or_set_properties(void *user) | |||
4580 | { | |||
4581 | struct gpu_gen *gen = user; | |||
4582 | ||||
4583 | if (gen->options->reschedule) | |||
4584 | return compute_schedule(gen); | |||
4585 | else | |||
4586 | return determine_properties_original_schedule(gen); | |||
4587 | } | |||
4588 | ||||
4589 | /* Obtain a schedule for the scop, by reading it from | |||
4590 | * a file, by computing one or by determining the properties | |||
4591 | * of the original schedule. | |||
4592 | */ | |||
4593 | __isl_give isl_schedule *get_schedule(struct gpu_gen *gen) | |||
4594 | { | |||
4595 | return ppcg_get_schedule(gen->ctx, gen->options, | |||
4596 | &compute_or_set_properties, gen); | |||
4597 | } | |||
4598 | ||||
4599 | /* Construct the string "<a>_<b>". | |||
4600 | */ | |||
4601 | static char *concat(isl_ctx *ctx, const char *a, const char *b) | |||
4602 | { | |||
4603 | isl_printer *p; | |||
4604 | char *s; | |||
4605 | ||||
4606 | p = isl_printer_to_str(ctx); | |||
4607 | p = isl_printer_print_str(p, a); | |||
4608 | p = isl_printer_print_str(p, "_"); | |||
4609 | p = isl_printer_print_str(p, b); | |||
4610 | s = isl_printer_get_str(p); | |||
4611 | isl_printer_free(p); | |||
4612 | ||||
4613 | return s; | |||
4614 | } | |||
4615 | ||||
4616 | /* For each array in "prog" of which an element appears in "accessed" and | |||
4617 | * that is not a read only scalar, create a zero-dimensional universe set | |||
4618 | * of which the tuple id has name "<prefix>_<name of array>" and a user | |||
4619 | * pointer pointing to the array (gpu_array_info). | |||
4620 | * | |||
4621 | * If the array is local to "prog", then make sure it will be declared | |||
4622 | * in the host code. | |||
4623 | * | |||
4624 | * Return the list of these universe sets. | |||
4625 | */ | |||
4626 | static __isl_give isl_union_set_list *create_copy_filters(struct gpu_prog *prog, | |||
4627 | const char *prefix, __isl_take isl_union_set *accessed) | |||
4628 | { | |||
4629 | int i; | |||
4630 | isl_ctx *ctx; | |||
4631 | isl_union_set_list *filters; | |||
4632 | ||||
4633 | ctx = prog->ctx; | |||
4634 | filters = isl_union_set_list_alloc(ctx, 0); | |||
4635 | for (i = 0; i < prog->n_array; ++i) { | |||
4636 | struct gpu_array_info *array = &prog->array[i]; | |||
4637 | isl_space *space; | |||
4638 | isl_set *accessed_i; | |||
4639 | int empty; | |||
4640 | char *name; | |||
4641 | isl_id *id; | |||
4642 | isl_union_set *uset; | |||
4643 | ||||
4644 | if (gpu_array_is_read_only_scalar(array)) | |||
4645 | continue; | |||
4646 | ||||
4647 | space = isl_space_copy(array->space); | |||
4648 | accessed_i = isl_union_set_extract_set(accessed, space); | |||
4649 | empty = isl_set_plain_is_empty(accessed_i); | |||
4650 | isl_set_free(accessed_i); | |||
4651 | if (empty < 0) { | |||
4652 | filters = isl_union_set_list_free(filters); | |||
4653 | break; | |||
4654 | } | |||
4655 | if (empty) | |||
4656 | continue; | |||
4657 | ||||
4658 | array->global = 1; | |||
4659 | if (array->local) | |||
4660 | array->declare_local = 1; | |||
4661 | ||||
4662 | name = concat(ctx, prefix, array->name); | |||
4663 | id = name ? isl_id_alloc(ctx, name, array) : NULL((void*)0); | |||
4664 | free(name); | |||
4665 | space = isl_space_set_alloc(ctx, 0, 0); | |||
4666 | space = isl_space_set_tuple_id(space, isl_dim_set, id); | |||
4667 | uset = isl_union_set_from_set(isl_set_universe(space)); | |||
4668 | ||||
4669 | filters = isl_union_set_list_add(filters, uset); | |||
4670 | } | |||
4671 | isl_union_set_free(accessed); | |||
4672 | ||||
4673 | return filters; | |||
4674 | } | |||
4675 | ||||
4676 | /* Make sure that code for the statements in "filters" that | |||
4677 | * copy arrays to or from the device is only generated when | |||
4678 | * the size of the corresponding array is positive. | |||
4679 | * That is, add a set node underneath "graft" with "filters" as children | |||
4680 | * and for each child add a guard that the selects the parameter | |||
4681 | * values for which the corresponding array has a positive size. | |||
4682 | * The array is available in the user pointer of the statement identifier. | |||
4683 | * "depth" is the schedule depth of the position where "graft" | |||
4684 | * will be added. | |||
4685 | */ | |||
4686 | static __isl_give isl_schedule_node *insert_positive_size_guards( | |||
4687 | __isl_take isl_schedule_node *graft, | |||
4688 | __isl_take isl_union_set_list *filters, int depth) | |||
4689 | { | |||
4690 | int i, n; | |||
4691 | ||||
4692 | graft = isl_schedule_node_child(graft, 0); | |||
4693 | graft = isl_schedule_node_insert_set(graft, filters); | |||
4694 | n = isl_schedule_node_n_children(graft); | |||
4695 | for (i = 0; i < n; ++i) { | |||
4696 | isl_union_set *filter; | |||
4697 | isl_set *domain, *guard; | |||
4698 | isl_id *id; | |||
4699 | struct gpu_array_info *array; | |||
4700 | ||||
4701 | graft = isl_schedule_node_child(graft, i); | |||
4702 | filter = isl_schedule_node_filter_get_filter(graft); | |||
4703 | domain = isl_set_from_union_set(filter); | |||
4704 | id = isl_set_get_tuple_id(domain); | |||
4705 | array = isl_id_get_user(id); | |||
4706 | isl_id_free(id); | |||
4707 | isl_set_free(domain); | |||
4708 | guard = gpu_array_positive_size_guard(array); | |||
4709 | guard = isl_set_from_params(guard); | |||
4710 | guard = isl_set_add_dims(guard, isl_dim_set, depth); | |||
4711 | graft = isl_schedule_node_child(graft, 0); | |||
4712 | graft = isl_schedule_node_insert_guard(graft, guard); | |||
4713 | graft = isl_schedule_node_parent(graft); | |||
4714 | graft = isl_schedule_node_parent(graft); | |||
4715 | } | |||
4716 | graft = isl_schedule_node_parent(graft); | |||
4717 | ||||
4718 | return graft; | |||
4719 | } | |||
4720 | ||||
4721 | /* Create a graft for copying arrays to or from the device, | |||
4722 | * whenever the size of the array is strictly positive. | |||
4723 | * Each statement is called "<prefix>_<name of array>" and | |||
4724 | * the identifier has a user pointer pointing to the array. | |||
4725 | * The graft will be added at the position specified by "node". | |||
4726 | * "copy" contains the array elements that need to be copied. | |||
4727 | * Only arrays of which some elements need to be copied | |||
4728 | * will have a corresponding statement in the graph. | |||
4729 | * Note though that each such statement will copy the entire array. | |||
4730 | */ | |||
4731 | static __isl_give isl_schedule_node *create_copy_device(struct gpu_prog *prog, | |||
4732 | __isl_keep isl_schedule_node *node, const char *prefix, | |||
4733 | __isl_take isl_union_set *copy) | |||
4734 | { | |||
4735 | int depth; | |||
4736 | isl_ctx *ctx; | |||
4737 | isl_space *space; | |||
4738 | isl_union_set *all, *domain; | |||
4739 | isl_union_set_list *filters; | |||
4740 | isl_union_map *extension; | |||
4741 | isl_schedule_node *graft; | |||
4742 | ||||
4743 | ctx = prog->ctx; | |||
4744 | depth = isl_schedule_node_get_schedule_depth(node); | |||
4745 | filters = create_copy_filters(prog, prefix, copy); | |||
4746 | all = isl_union_set_list_union(isl_union_set_list_copy(filters)); | |||
4747 | ||||
4748 | space = depth < 0 ? NULL((void*)0) : isl_space_set_alloc(ctx, 0, depth); | |||
4749 | domain = isl_union_set_from_set(isl_set_universe(space)); | |||
4750 | extension = isl_union_map_from_domain_and_range(domain, all); | |||
4751 | graft = isl_schedule_node_from_extension(extension); | |||
4752 | ||||
4753 | if (!filters) | |||
4754 | return isl_schedule_node_free(graft); | |||
4755 | if (isl_union_set_list_n_union_set(filters) == 0) { | |||
4756 | isl_union_set_list_free(filters); | |||
4757 | return graft; | |||
4758 | } | |||
4759 | ||||
4760 | return insert_positive_size_guards(graft, filters, depth); | |||
4761 | } | |||
4762 | ||||
4763 | /* Return (the universe spaces of) the arrays that are declared | |||
4764 | * inside the scop corresponding to "prog" and for which all | |||
4765 | * potential writes inside the scop form a subset of "domain". | |||
4766 | */ | |||
4767 | static __isl_give isl_union_set *extract_local_accesses(struct gpu_prog *prog, | |||
4768 | __isl_keep isl_union_set *domain) | |||
4769 | { | |||
4770 | int i; | |||
4771 | isl_union_set *local; | |||
4772 | ||||
4773 | local = isl_union_set_empty(isl_union_set_get_space(domain)); | |||
4774 | ||||
4775 | for (i = 0; i < prog->n_array; ++i) { | |||
4776 | isl_set *set; | |||
4777 | isl_union_map *to_outer; | |||
4778 | isl_union_map *may_write; | |||
4779 | isl_union_set *write_domain; | |||
4780 | isl_union_set *fields; | |||
4781 | int subset; | |||
4782 | ||||
4783 | if (!prog->array[i].local) | |||
4784 | continue; | |||
4785 | ||||
4786 | set = isl_set_universe(isl_space_copy(prog->array[i].space)); | |||
4787 | to_outer = isl_union_map_copy(prog->to_outer); | |||
4788 | to_outer = isl_union_map_intersect_range(to_outer, | |||
4789 | isl_union_set_from_set(isl_set_copy(set))); | |||
4790 | fields = isl_union_map_domain(to_outer); | |||
4791 | may_write = isl_union_map_copy(prog->may_write); | |||
4792 | may_write = isl_union_map_intersect_range(may_write, fields); | |||
4793 | write_domain = isl_union_map_domain(may_write); | |||
4794 | subset = isl_union_set_is_subset(write_domain, domain); | |||
4795 | isl_union_set_free(write_domain); | |||
4796 | ||||
4797 | if (subset < 0) { | |||
4798 | isl_set_free(set); | |||
4799 | return isl_union_set_free(local); | |||
4800 | } else if (subset) { | |||
4801 | local = isl_union_set_add_set(local, set); | |||
4802 | } else { | |||
4803 | isl_set_free(set); | |||
4804 | } | |||
4805 | } | |||
4806 | ||||
4807 | return local; | |||
4808 | } | |||
4809 | ||||
4810 | /* Internal data structure for node_may_persist. | |||
4811 | * | |||
4812 | * "tagger" maps tagged iteration domains to the corresponding untagged | |||
4813 | * iteration domain. | |||
4814 | * | |||
4815 | * "may_persist_flow" is the set of all tagged dataflow dependences | |||
4816 | * with those dependences removed that either precede or follow | |||
4817 | * the kernel launch in a sequence. | |||
4818 | * "inner_band_flow" is the set of all tagged dataflow dependences | |||
4819 | * that are local to a given iteration of the outer band nodes | |||
4820 | * with respect to the current node. | |||
4821 | * "local_flow" is equal to "inner_band_flow", except that the domain | |||
4822 | * and the range have been intersected with intermediate filters | |||
4823 | * on children of sets or sequences. | |||
4824 | */ | |||
4825 | struct ppcg_may_persist_data { | |||
4826 | isl_union_pw_multi_aff *tagger; | |||
4827 | ||||
4828 | isl_union_map *local_flow; | |||
4829 | isl_union_map *inner_band_flow; | |||
4830 | isl_union_map *may_persist_flow; | |||
4831 | }; | |||
4832 | ||||
4833 | /* Update the information in "data" based on the band ancestor "node". | |||
4834 | * | |||
4835 | * In particular, we restrict the dependences in data->local_flow | |||
4836 | * to those dependence where the source and the sink occur in | |||
4837 | * the same iteration of the given band node. | |||
4838 | * We also update data->inner_band_flow to the new value of | |||
4839 | * data->local_flow. | |||
4840 | */ | |||
4841 | static int update_may_persist_at_band(__isl_keep isl_schedule_node *node, | |||
4842 | struct ppcg_may_persist_data *data) | |||
4843 | { | |||
4844 | isl_multi_union_pw_aff *partial; | |||
4845 | isl_union_pw_multi_aff *contraction; | |||
4846 | isl_union_map *flow; | |||
4847 | ||||
4848 | if (isl_schedule_node_band_n_member(node) == 0) | |||
4849 | return 0; | |||
4850 | ||||
4851 | partial = isl_schedule_node_band_get_partial_schedule(node); | |||
4852 | contraction = isl_schedule_node_get_subtree_contraction(node); | |||
4853 | partial = isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial, | |||
4854 | contraction); | |||
4855 | partial = isl_multi_union_pw_aff_pullback_union_pw_multi_aff(partial, | |||
4856 | isl_union_pw_multi_aff_copy(data->tagger)); | |||
4857 | ||||
4858 | flow = data->local_flow; | |||
4859 | flow = isl_union_map_eq_at_multi_union_pw_aff(flow, partial); | |||
4860 | data->local_flow = flow; | |||
4861 | ||||
4862 | isl_union_map_free(data->inner_band_flow); | |||
4863 | data->inner_band_flow = isl_union_map_copy(data->local_flow); | |||
4864 | ||||
4865 | return 0; | |||
4866 | } | |||
4867 | ||||
4868 | /* Given a set of local reaching domain elements "domain", | |||
4869 | * expand them to the corresponding leaf domain elements using "contraction" | |||
4870 | * and insert the array references tags using data->tagger. | |||
4871 | */ | |||
4872 | static __isl_give isl_union_set *expand_and_tag( | |||
4873 | __isl_take isl_union_set *domain, | |||
4874 | __isl_take isl_union_pw_multi_aff *contraction, | |||
4875 | struct ppcg_may_persist_data *data) | |||
4876 | { | |||
4877 | domain = isl_union_set_preimage_union_pw_multi_aff(domain, | |||
4878 | contraction); | |||
4879 | domain = isl_union_set_preimage_union_pw_multi_aff(domain, | |||
4880 | isl_union_pw_multi_aff_copy(data->tagger)); | |||
4881 | return domain; | |||
4882 | } | |||
4883 | ||||
4884 | /* Given a filter node that is the child of a set or sequence node, | |||
4885 | * restrict data->local_flow to refer only to those elements | |||
4886 | * in the filter of the node. | |||
4887 | * "contraction" maps the leaf domain elements of the schedule tree | |||
4888 | * to the corresponding domain elements at (the parent of) "node". | |||
4889 | */ | |||
4890 | static int filter_flow(__isl_keep isl_schedule_node *node, | |||
4891 | struct ppcg_may_persist_data *data, | |||
4892 | __isl_take isl_union_pw_multi_aff *contraction) | |||
4893 | { | |||
4894 | isl_union_set *filter; | |||
4895 | isl_union_map *flow; | |||
4896 | ||||
4897 | flow = data->local_flow; | |||
4898 | filter = isl_schedule_node_filter_get_filter(node); | |||
4899 | filter = expand_and_tag(filter, contraction, data); | |||
4900 | flow = isl_union_map_intersect_domain(flow, isl_union_set_copy(filter)); | |||
4901 | flow = isl_union_map_intersect_range(flow, filter); | |||
4902 | data->local_flow = flow; | |||
4903 | ||||
4904 | return 0; | |||
4905 | } | |||
4906 | ||||
4907 | /* Given a filter node "node", collect the filters on all preceding siblings | |||
4908 | * (which are also filter nodes), add them to "filters" and return the result. | |||
4909 | */ | |||
4910 | static __isl_give isl_union_set *add_previous_filters( | |||
4911 | __isl_take isl_union_set *filters, __isl_keep isl_schedule_node *node) | |||
4912 | { | |||
4913 | isl_schedule_node *sibling; | |||
4914 | ||||
4915 | sibling = isl_schedule_node_copy(node); | |||
4916 | while (sibling && isl_schedule_node_has_previous_sibling(sibling)) { | |||
4917 | isl_union_set *filter; | |||
4918 | ||||
4919 | sibling = isl_schedule_node_previous_sibling(sibling); | |||
4920 | filter = isl_schedule_node_filter_get_filter(sibling); | |||
4921 | filters = isl_union_set_union(filters, filter); | |||
4922 | } | |||
4923 | isl_schedule_node_free(sibling); | |||
4924 | if (!sibling) | |||
4925 | return isl_union_set_free(filters); | |||
4926 | ||||
4927 | return filters; | |||
4928 | } | |||
4929 | ||||
4930 | /* Given a filter node "node", collect the filters on all following siblings | |||
4931 | * (which are also filter nodes), add them to "filters" and return the result. | |||
4932 | */ | |||
4933 | static __isl_give isl_union_set *add_next_filters( | |||
4934 | __isl_take isl_union_set *filters, __isl_keep isl_schedule_node *node) | |||
4935 | { | |||
4936 | isl_schedule_node *sibling; | |||
4937 | ||||
4938 | sibling = isl_schedule_node_copy(node); | |||
4939 | while (sibling && isl_schedule_node_has_next_sibling(sibling)) { | |||
4940 | isl_union_set *filter; | |||
4941 | ||||
4942 | sibling = isl_schedule_node_next_sibling(sibling); | |||
4943 | filter = isl_schedule_node_filter_get_filter(sibling); | |||
4944 | filters = isl_union_set_union(filters, filter); | |||
4945 | } | |||
4946 | isl_schedule_node_free(sibling); | |||
4947 | if (!sibling) | |||
4948 | return isl_union_set_free(filters); | |||
4949 | ||||
4950 | return filters; | |||
4951 | } | |||
4952 | ||||
4953 | /* Remove those flow dependences from data->may_persist_flow | |||
4954 | * that flow between elements of "domain" within the same iteration | |||
4955 | * of all outer band nodes. | |||
4956 | * "contraction" maps the leaf domain elements of the schedule tree | |||
4957 | * to the corresponding elements "domain". | |||
4958 | */ | |||
4959 | static void remove_external_flow(struct ppcg_may_persist_data *data, | |||
4960 | __isl_take isl_union_set *domain, | |||
4961 | __isl_keep isl_union_pw_multi_aff *contraction) | |||
4962 | { | |||
4963 | isl_union_map *flow; | |||
4964 | ||||
4965 | contraction = isl_union_pw_multi_aff_copy(contraction); | |||
4966 | domain = expand_and_tag(domain, contraction, data); | |||
4967 | flow = isl_union_map_copy(data->local_flow); | |||
4968 | flow = isl_union_map_intersect_domain(flow, isl_union_set_copy(domain)); | |||
4969 | flow = isl_union_map_intersect_range(flow, domain); | |||
4970 | ||||
4971 | data->may_persist_flow = isl_union_map_subtract(data->may_persist_flow, | |||
4972 | flow); | |||
4973 | } | |||
4974 | ||||
4975 | /* Update the information in "data" based on the filter ancestor "node". | |||
4976 | * We only need to modify anything if the filter is the child | |||
4977 | * of a set or sequence node. | |||
4978 | * | |||
4979 | * In the case of a sequence, we remove the dependences between | |||
4980 | * statement instances that are both executed either before or | |||
4981 | * after the subtree that will be mapped to a kernel, within | |||
4982 | * the same iteration of outer bands. | |||
4983 | * | |||
4984 | * In both cases, we restrict data->local_flow to the current child. | |||
4985 | */ | |||
4986 | static int update_may_persist_at_filter(__isl_keep isl_schedule_node *node, | |||
4987 | struct ppcg_may_persist_data *data) | |||
4988 | { | |||
4989 | enum isl_schedule_node_type type; | |||
4990 | isl_schedule_node *parent; | |||
4991 | isl_space *space; | |||
4992 | isl_union_pw_multi_aff *contraction; | |||
4993 | isl_union_set *before, *after, *filter; | |||
4994 | ||||
4995 | type = isl_schedule_node_get_parent_type(node); | |||
4996 | if (type != isl_schedule_node_sequence && type != isl_schedule_node_set) | |||
4997 | return 0; | |||
4998 | ||||
4999 | parent = isl_schedule_node_copy(node); | |||
5000 | parent = isl_schedule_node_parent(parent); | |||
5001 | contraction = isl_schedule_node_get_subtree_contraction(parent); | |||
5002 | isl_schedule_node_free(parent); | |||
5003 | ||||
5004 | if (type == isl_schedule_node_set) | |||
5005 | return filter_flow(node, data, contraction); | |||
5006 | ||||
5007 | filter = isl_schedule_node_filter_get_filter(node); | |||
5008 | space = isl_union_set_get_space(filter); | |||
5009 | isl_union_set_free(filter); | |||
5010 | before = isl_union_set_empty(space); | |||
5011 | after = isl_union_set_copy(before); | |||
5012 | before = add_previous_filters(before, node); | |||
5013 | after = add_next_filters(after, node); | |||
5014 | ||||
5015 | remove_external_flow(data, before, contraction); | |||
5016 | remove_external_flow(data, after, contraction); | |||
5017 | ||||
5018 | return filter_flow(node, data, contraction); | |||
5019 | } | |||
5020 | ||||
5021 | /* Update the information in "data" based on the ancestor "node". | |||
5022 | */ | |||
5023 | static isl_stat update_may_persist_at(__isl_keep isl_schedule_node *node, | |||
5024 | void *user) | |||
5025 | { | |||
5026 | struct ppcg_may_persist_data *data = user; | |||
5027 | ||||
5028 | switch (isl_schedule_node_get_type(node)) { | |||
5029 | case isl_schedule_node_error: | |||
5030 | return isl_stat_error; | |||
5031 | case isl_schedule_node_context: | |||
5032 | case isl_schedule_node_domain: | |||
5033 | case isl_schedule_node_expansion: | |||
5034 | case isl_schedule_node_extension: | |||
5035 | case isl_schedule_node_guard: | |||
5036 | case isl_schedule_node_leaf: | |||
5037 | case isl_schedule_node_mark: | |||
5038 | case isl_schedule_node_sequence: | |||
5039 | case isl_schedule_node_set: | |||
5040 | break; | |||
5041 | case isl_schedule_node_band: | |||
5042 | if (update_may_persist_at_band(node, data) < 0) | |||
5043 | return isl_stat_error; | |||
5044 | break; | |||
5045 | case isl_schedule_node_filter: | |||
5046 | if (update_may_persist_at_filter(node, data) < 0) | |||
5047 | return isl_stat_error; | |||
5048 | break; | |||
5049 | } | |||
5050 | ||||
5051 | return isl_stat_ok; | |||
5052 | } | |||
5053 | ||||
5054 | /* Determine the set of array elements that may need to be perserved | |||
5055 | * by a kernel constructed from the subtree at "node". | |||
5056 | * This includes the set of array elements that may need to be preserved | |||
5057 | * by the entire scop (prog->may_persist) and the elements for which | |||
5058 | * there is a potential flow dependence that may cross a kernel launch. | |||
5059 | * | |||
5060 | * To determine the second set, we start from all flow dependences. | |||
5061 | * From this set of dependences, we remove those that cannot possibly | |||
5062 | * require data to be preserved by a kernel launch. | |||
5063 | * In particular, we consider the following sets of dependences. | |||
5064 | * - dependences of which the write occurs inside the kernel. | |||
5065 | * If the data is needed outside the kernel, then it will | |||
5066 | * be copied out immediately after the kernel launch, so there | |||
5067 | * is no need for any special care. | |||
5068 | * - dependences of which the read occurs inside the kernel and the | |||
5069 | * corresponding write occurs inside the same iteration of the | |||
5070 | * outer band nodes. This means that the data is needed in | |||
5071 | * the first kernel launch after the write, which is already | |||
5072 | * taken care of by the standard copy-in. That is, the data | |||
5073 | * do not need to be preserved by any intermediate call to | |||
5074 | * the same kernel. | |||
5075 | * - dependences of which the write and the read either both occur | |||
5076 | * before the kernel launch or both occur after the kernel launch, | |||
5077 | * within the same iteration of the outer band nodes with respect | |||
5078 | * to the sequence that determines the ordering of the dependence | |||
5079 | * and the kernel launch. Such flow dependences cannot cross | |||
5080 | * any kernel launch. | |||
5081 | * | |||
5082 | * For the remaining (tagged) dependences, we take the domain | |||
5083 | * (i.e., the tagged writes) and apply the tagged access relation | |||
5084 | * to obtain the accessed data elements. | |||
5085 | * These are then combined with the elements that may need to be | |||
5086 | * preserved by the entire scop. | |||
5087 | */ | |||
5088 | static __isl_give isl_union_set *node_may_persist( | |||
5089 | __isl_keep isl_schedule_node *node, struct gpu_prog *prog) | |||
5090 | { | |||
5091 | struct ppcg_may_persist_data data; | |||
5092 | isl_union_pw_multi_aff *contraction; | |||
5093 | isl_union_set *domain; | |||
5094 | isl_union_set *persist; | |||
5095 | isl_union_map *flow, *local_flow; | |||
5096 | ||||
5097 | data.tagger = prog->scop->tagger; | |||
5098 | ||||
5099 | flow = isl_union_map_copy(prog->scop->tagged_dep_flow); | |||
5100 | data.local_flow = isl_union_map_copy(flow); | |||
5101 | data.inner_band_flow = isl_union_map_copy(flow); | |||
5102 | data.may_persist_flow = flow; | |||
5103 | if (isl_schedule_node_foreach_ancestor_top_down(node, | |||
5104 | &update_may_persist_at, &data) < 0) | |||
5105 | data.may_persist_flow = | |||
5106 | isl_union_map_free(data.may_persist_flow); | |||
5107 | flow = data.may_persist_flow; | |||
5108 | isl_union_map_free(data.local_flow); | |||
5109 | ||||
5110 | domain = isl_schedule_node_get_domain(node); | |||
5111 | contraction = isl_schedule_node_get_subtree_contraction(node); | |||
5112 | domain = isl_union_set_preimage_union_pw_multi_aff(domain, | |||
5113 | contraction); | |||
5114 | domain = isl_union_set_preimage_union_pw_multi_aff(domain, | |||
5115 | isl_union_pw_multi_aff_copy(data.tagger)); | |||
5116 | flow = isl_union_map_subtract_domain(flow, isl_union_set_copy(domain)); | |||
5117 | local_flow = data.inner_band_flow; | |||
5118 | local_flow = isl_union_map_intersect_range(local_flow, domain); | |||
5119 | flow = isl_union_map_subtract(flow, local_flow); | |||
5120 | ||||
5121 | persist = isl_union_map_domain(flow); | |||
5122 | persist = isl_union_set_apply(persist, | |||
5123 | isl_union_map_copy(prog->scop->tagged_may_writes)); | |||
5124 | persist = isl_union_set_union(persist, | |||
5125 | isl_union_set_copy(prog->may_persist)); | |||
5126 | ||||
5127 | return persist; | |||
5128 | } | |||
5129 | ||||
5130 | /* Add nodes for copying outer arrays in and out of the device | |||
5131 | * before and after the subtree "node", which contains one or more kernels. | |||
5132 | * "domain" contains the original statement instances, i.e., | |||
5133 | * those that correspond to the domains of the access relations in "prog". | |||
5134 | * In particular, the domain has not been contracted in any way. | |||
5135 | * "prefix" contains the prefix schedule at that point, in terms | |||
5136 | * of the same original statement instances. | |||
5137 | * | |||
5138 | * We first compute the sets of outer array elements that need | |||
5139 | * to be copied in and out and then graft in the nodes for | |||
5140 | * performing this copying. | |||
5141 | * | |||
5142 | * In particular, for each array that is possibly written anywhere in | |||
5143 | * the subtree "node" and that may be used after "node" | |||
5144 | * or that may be visible outside the corresponding scop, | |||
5145 | * we copy out its entire extent. | |||
5146 | * | |||
5147 | * Any array elements that is read without first being written inside | |||
5148 | * the subtree "node" needs to be copied in. | |||
5149 | * Furthermore, if there are any array elements that | |||
5150 | * are copied out, but that may not be written inside "node, then | |||
5151 | * they also need to be copied in to ensure that the value after execution | |||
5152 | * is the same as the value before execution, at least for those array | |||
5153 | * elements that may have their values preserved by the scop or that | |||
5154 | * may be written before "node" and read after "node". | |||
5155 | * In case the array elements are structures, we need to take into | |||
5156 | * account that all members of the structures need to be written | |||
5157 | * by "node" before we can avoid copying the data structure in. | |||
5158 | * | |||
5159 | * Note that the may_write relation is intersected with the domain, | |||
5160 | * which has been intersected with the context. | |||
5161 | * This helps in those cases where the arrays are declared with a fixed size, | |||
5162 | * while the accesses are parametric and the context assigns a fixed value | |||
5163 | * to the parameters. | |||
5164 | * | |||
5165 | * If an element from a local array is read without first being written, | |||
5166 | * then there is no point in copying it in since it cannot have been | |||
5167 | * written prior to the scop. Warn about the uninitialized read instead. | |||
5168 | */ | |||
5169 | static __isl_give isl_schedule_node *add_to_from_device( | |||
5170 | __isl_take isl_schedule_node *node, __isl_take isl_union_set *domain, | |||
5171 | __isl_take isl_union_map *prefix, struct gpu_prog *prog) | |||
5172 | { | |||
5173 | isl_union_set *local; | |||
5174 | isl_union_set *may_persist; | |||
5175 | isl_union_map *may_write, *must_write, *copy_out, *not_written; | |||
5176 | isl_union_map *read, *copy_in; | |||
5177 | isl_union_map *tagged; | |||
5178 | isl_union_map *local_uninitialized; | |||
5179 | isl_schedule_node *graft; | |||
5180 | ||||
5181 | tagged = isl_union_map_copy(prog->scop->tagged_reads); | |||
5182 | tagged = isl_union_map_union(tagged, | |||
5183 | isl_union_map_copy(prog->scop->tagged_may_writes)); | |||
5184 | ||||
5185 | may_write = isl_union_map_copy(prog->may_write); | |||
5186 | may_write = isl_union_map_intersect_domain(may_write, | |||
5187 | isl_union_set_copy(domain)); | |||
5188 | may_write = remove_local_accesses(prog, | |||
5189 | isl_union_map_copy(tagged), may_write, | |||
5190 | isl_union_map_copy(prefix), 0); | |||
5191 | may_write = isl_union_map_apply_range(may_write, | |||
5192 | isl_union_map_copy(prog->to_outer)); | |||
5193 | may_write = isl_union_map_apply_domain(may_write, | |||
5194 | isl_union_map_copy(prefix)); | |||
5195 | may_write = approximate_copy_out(may_write, prog); | |||
5196 | copy_out = isl_union_map_copy(may_write); | |||
5197 | may_write = isl_union_map_apply_range(may_write, | |||
5198 | isl_union_map_copy(prog->to_inner)); | |||
5199 | must_write = isl_union_map_copy(prog->must_write); | |||
5200 | must_write = isl_union_map_apply_domain(must_write, | |||
5201 | isl_union_map_copy(prefix)); | |||
5202 | may_persist = node_may_persist(node, prog); | |||
5203 | may_write = isl_union_map_intersect_range(may_write, may_persist); | |||
5204 | not_written = isl_union_map_subtract(may_write, must_write); | |||
5205 | ||||
5206 | local = extract_local_accesses(prog, domain); | |||
5207 | read = isl_union_map_copy(prog->read); | |||
5208 | read = isl_union_map_intersect_domain(read, domain); | |||
5209 | read = remove_local_accesses(prog, tagged, read, | |||
5210 | isl_union_map_copy(prefix), 1); | |||
5211 | local = isl_union_set_apply(local, isl_union_map_copy(prog->to_inner)); | |||
5212 | local_uninitialized = isl_union_map_copy(prog->scop->live_in); | |||
5213 | local_uninitialized = isl_union_map_intersect_range(local_uninitialized, | |||
5214 | local); | |||
5215 | local_uninitialized = isl_union_map_intersect(local_uninitialized, | |||
5216 | isl_union_map_copy(read)); | |||
5217 | if (!isl_union_map_is_empty(local_uninitialized)) { | |||
5218 | fprintf(stderrstderr, | |||
5219 | "possibly uninitialized reads (not copied in):\n"); | |||
5220 | isl_union_map_dump(local_uninitialized); | |||
5221 | } | |||
5222 | read = isl_union_map_subtract(read, local_uninitialized); | |||
5223 | read = isl_union_map_apply_domain(read, prefix); | |||
5224 | copy_in = isl_union_map_union(read, not_written); | |||
5225 | copy_in = isl_union_map_apply_range(copy_in, | |||
5226 | isl_union_map_copy(prog->to_outer)); | |||
5227 | ||||
5228 | graft = create_copy_device(prog, node, "to_device", | |||
5229 | isl_union_map_range(copy_in)); | |||
5230 | node = isl_schedule_node_graft_before(node, graft); | |||
5231 | graft = create_copy_device(prog, node, "from_device", | |||
5232 | isl_union_map_range(copy_out)); | |||
5233 | node = isl_schedule_node_graft_after(node, graft); | |||
5234 | ||||
5235 | return node; | |||
5236 | } | |||
5237 | ||||
5238 | /* Add nodes for initializing ("init_device") and clearing ("clear_device") | |||
5239 | * the device before and after "node". | |||
5240 | */ | |||
5241 | static __isl_give isl_schedule_node *add_init_clear_device( | |||
5242 | __isl_take isl_schedule_node *node) | |||
5243 | { | |||
5244 | isl_ctx *ctx; | |||
5245 | isl_space *space; | |||
5246 | isl_union_set *domain; | |||
5247 | isl_schedule_node *graft; | |||
5248 | ||||
5249 | ctx = isl_schedule_node_get_ctx(node); | |||
5250 | ||||
5251 | space = isl_space_set_alloc(ctx, 0, 0); | |||
5252 | space = isl_space_set_tuple_name(space, isl_dim_set, "init_device"); | |||
5253 | domain = isl_union_set_from_set(isl_set_universe(space)); | |||
5254 | graft = isl_schedule_node_from_domain(domain); | |||
5255 | ||||
5256 | node = isl_schedule_node_graft_before(node, graft); | |||
5257 | ||||
5258 | space = isl_space_set_alloc(ctx, 0, 0); | |||
5259 | space = isl_space_set_tuple_name(space, isl_dim_set, "clear_device"); | |||
5260 | domain = isl_union_set_from_set(isl_set_universe(space)); | |||
5261 | graft = isl_schedule_node_from_domain(domain); | |||
5262 | ||||
5263 | node = isl_schedule_node_graft_after(node, graft); | |||
5264 | ||||
5265 | return node; | |||
5266 | } | |||
5267 | ||||
5268 | /* Update "schedule" for mapping to a GPU device. | |||
5269 | * | |||
5270 | * In particular, insert a context node, create kernels for | |||
5271 | * each outermost tilable band and introduce nodes for copying arrays | |||
5272 | * in and out of the device and for initializing and clearing the device. | |||
5273 | * If the child of the initial root points to a set node, | |||
5274 | * then children of this node that do not contain any tilable bands | |||
5275 | * are separated from the other children and are not mapped to | |||
5276 | * the device. | |||
5277 | * | |||
5278 | * The GPU code is generated in a context where at least one | |||
5279 | * statement instance is executed. The corresponding guard is inserted | |||
5280 | * around the entire schedule. | |||
5281 | */ | |||
5282 | __isl_give isl_schedule *map_to_device(struct gpu_gen *gen, | |||
5283 | __isl_take isl_schedule *schedule, int to_from_device) | |||
5284 | { | |||
5285 | isl_schedule_node *node; | |||
5286 | isl_set *context; | |||
5287 | isl_set *guard; | |||
5288 | isl_union_set *domain; | |||
5289 | isl_union_map *prefix; | |||
5290 | isl_union_pw_multi_aff *contraction; | |||
5291 | struct gpu_prog *prog; | |||
5292 | ||||
5293 | context = isl_set_copy(gen->prog->context); | |||
5294 | context = isl_set_from_params(context); | |||
5295 | schedule = isl_schedule_insert_context(schedule, context); | |||
5296 | ||||
5297 | prog = gen->prog; | |||
5298 | guard = isl_union_set_params(isl_union_set_copy(prog->scop->domain)); | |||
5299 | prog->context = isl_set_intersect(prog->context, isl_set_copy(guard)); | |||
5300 | guard = isl_set_from_params(guard); | |||
5301 | ||||
5302 | node = isl_schedule_get_root(schedule); | |||
5303 | isl_schedule_free(schedule); | |||
5304 | node = isl_schedule_node_child(node, 0); | |||
5305 | node = isl_schedule_node_child(node, 0); | |||
5306 | node = isolate_permutable_subtrees(node, gen->prog); | |||
5307 | domain = isl_schedule_node_get_domain(node); | |||
5308 | contraction = isl_schedule_node_get_subtree_contraction(node); | |||
5309 | domain = isl_union_set_preimage_union_pw_multi_aff(domain, | |||
5310 | isl_union_pw_multi_aff_copy(contraction)); | |||
5311 | prefix = isl_schedule_node_get_prefix_schedule_union_map(node); | |||
5312 | prefix = isl_union_map_preimage_domain_union_pw_multi_aff(prefix, | |||
5313 | contraction); | |||
5314 | node = mark_kernels(gen, node); | |||
5315 | if (to_from_device) { | |||
5316 | node = add_to_from_device(node, domain, prefix, gen->prog); | |||
5317 | } else { | |||
5318 | isl_union_set_free(domain); | |||
5319 | isl_union_map_free(prefix); | |||
5320 | } | |||
5321 | node = isl_schedule_node_root(node); | |||
5322 | node = isl_schedule_node_child(node, 0); | |||
5323 | node = isl_schedule_node_child(node, 0); | |||
5324 | node = isl_schedule_node_insert_guard(node, guard); | |||
5325 | node = isl_schedule_node_child(node, 0); | |||
5326 | node = add_init_clear_device(node); | |||
5327 | schedule = isl_schedule_node_get_schedule(node); | |||
5328 | isl_schedule_node_free(node); | |||
5329 | ||||
5330 | return schedule; | |||
5331 | } | |||
5332 | ||||
5333 | /* Internal data structure for extract_access. | |||
5334 | * "next_access" points to the end of a linked list that is extended | |||
5335 | * by extract_access. | |||
5336 | * "single_expression" is set if the access expressions belong to | |||
5337 | * an expression statement (i.e., a statement without internal control). | |||
5338 | * "any_to_outer" maps all intermediate arrays to their outer arrays. | |||
5339 | */ | |||
5340 | struct ppcg_extract_access_data { | |||
5341 | struct gpu_stmt_access **next_access; | |||
5342 | int single_expression; | |||
5343 | isl_union_map *any_to_outer; | |||
5344 | }; | |||
5345 | ||||
5346 | /* Given a tagged access relation to a single array "tagged", extract it | |||
5347 | * as a map, taking into account that the input may be empty. | |||
5348 | * If the access relation is empty, then it does not contain | |||
5349 | * any space information, so we try to recover it from the index | |||
5350 | * expression. | |||
5351 | * The space of the index expression is of the form I -> A, | |||
5352 | * with I the statement instances and A the array, or [I -> F] -> A, | |||
5353 | * with F the filters corresponding to arguments. | |||
5354 | * We first drop F, if present, obtaining I -> A. | |||
5355 | * Then we construct I -> R, with R the reference tag, | |||
5356 | * combine the two into I -> [R -> A] and uncurry to obtain | |||
5357 | * the final result [I -> R] -> A. | |||
5358 | * Note that the index expression may have a lower dimension | |||
5359 | * than that of the array, but this dimension is not used | |||
5360 | * if the access relation is empty. | |||
5361 | */ | |||
5362 | static __isl_give isl_map *extract_single_tagged_access( | |||
5363 | __isl_take isl_union_map *tagged, __isl_keep pet_expr *expr) | |||
5364 | { | |||
5365 | int empty; | |||
5366 | isl_id *id; | |||
5367 | isl_space *space, *space2; | |||
5368 | isl_multi_pw_aff *index; | |||
5369 | ||||
5370 | empty = isl_union_map_is_empty(tagged); | |||
5371 | if (empty < 0) | |||
5372 | goto error; | |||
5373 | if (!empty) | |||
5374 | return isl_map_from_union_map(tagged); | |||
5375 | isl_union_map_free(tagged); | |||
5376 | ||||
5377 | index = pet_expr_access_get_index(expr); | |||
5378 | space = isl_multi_pw_aff_get_space(index); | |||
5379 | isl_multi_pw_aff_free(index); | |||
5380 | if (isl_space_domain_is_wrapping(space)) | |||
5381 | space = isl_space_domain_factor_domain(space); | |||
5382 | space2 = isl_space_copy(space); | |||
5383 | space2 = isl_space_from_domain(isl_space_domain(space)); | |||
5384 | id = pet_expr_access_get_ref_id(expr); | |||
5385 | space2 = isl_space_set_tuple_id(space2, isl_dim_out, id); | |||
5386 | space = isl_space_range_product(space2, space); | |||
5387 | space = isl_space_uncurry(space); | |||
5388 | ||||
5389 | return isl_map_empty(space); | |||
5390 | error: | |||
5391 | isl_union_map_free(tagged); | |||
5392 | return NULL((void*)0); | |||
5393 | } | |||
5394 | ||||
5395 | /* Does the index expression "index" of "expr" represent an access | |||
5396 | * to a single element? | |||
5397 | * That is, is "index" completely specified? | |||
5398 | * | |||
5399 | * If "expr" accesses elements from different spaces (i.e., fields | |||
5400 | * of a structure), then it does not access a single element. | |||
5401 | * Otherwise, if the single space of the access matches the space | |||
5402 | * of "index", then the index expression is completely specified | |||
5403 | * (no pointer to a lower-dimensional slice of the accessed array) | |||
5404 | * and a single element is being accessed. | |||
5405 | */ | |||
5406 | static isl_bool complete_index(__isl_keep pet_expr *expr, | |||
5407 | __isl_keep isl_multi_pw_aff *index) | |||
5408 | { | |||
5409 | isl_union_map *read, *write, *all; | |||
5410 | isl_map *map; | |||
5411 | isl_space *space1, *space2; | |||
5412 | isl_bool complete; | |||
5413 | ||||
5414 | read = pet_expr_access_get_may_read(expr); | |||
5415 | write = pet_expr_access_get_may_write(expr); | |||
5416 | all = isl_union_map_union(read, write); | |||
5417 | if (!all) | |||
5418 | return isl_bool_error; | |||
5419 | if (isl_union_map_n_map(all) != 1) { | |||
5420 | isl_union_map_free(all); | |||
5421 | return isl_bool_false; | |||
5422 | } | |||
5423 | map = isl_map_from_union_map(all); | |||
5424 | space1 = isl_map_get_space(map); | |||
5425 | isl_map_free(map); | |||
5426 | space2 = isl_multi_pw_aff_get_space(index); | |||
5427 | complete = isl_space_tuple_is_equal(space1, isl_dim_out, | |||
5428 | space2, isl_dim_out); | |||
5429 | isl_space_free(space1); | |||
5430 | isl_space_free(space2); | |||
5431 | ||||
5432 | return complete; | |||
5433 | } | |||
5434 | ||||
5435 | /* Does "expr" access a single, fixed element (independently of the statement | |||
5436 | * instance)? | |||
5437 | * That is, does it have a completely specified constant index expression? | |||
5438 | * | |||
5439 | * Note that it is not sufficient for the index expression to be | |||
5440 | * piecewise constant. isl_multi_pw_aff_is_cst can therefore not be used. | |||
5441 | */ | |||
5442 | static isl_bool accesses_fixed_element(__isl_keep pet_expr *expr) | |||
5443 | { | |||
5444 | int i, n; | |||
5445 | isl_multi_pw_aff *index; | |||
5446 | isl_bool fixed = isl_bool_true; | |||
5447 | ||||
5448 | index = pet_expr_access_get_index(expr); | |||
5449 | if (index < 0) | |||
5450 | return isl_bool_error; | |||
5451 | n = isl_multi_pw_aff_dim(index, isl_dim_out); | |||
5452 | for (i = 0; i < n; ++i) { | |||
5453 | isl_pw_aff *pa; | |||
5454 | ||||
5455 | pa = isl_multi_pw_aff_get_pw_aff(index, 0); | |||
5456 | fixed = isl_pw_aff_n_piece(pa) == 1; | |||
5457 | if (fixed) | |||
5458 | fixed = isl_pw_aff_is_cst(pa); | |||
5459 | isl_pw_aff_free(pa); | |||
5460 | if (fixed < 0 || !fixed) | |||
5461 | break; | |||
5462 | } | |||
5463 | if (fixed >= 0 && fixed) | |||
5464 | fixed = complete_index(expr, index); | |||
5465 | isl_multi_pw_aff_free(index); | |||
5466 | ||||
5467 | return fixed; | |||
5468 | } | |||
5469 | ||||
5470 | /* Extract a gpu_stmt_access from "expr", append it to the list | |||
5471 | * that ends in *data->next_access and update the end of the list. | |||
5472 | * If the access expression performs a write, then it is considered | |||
5473 | * exact only if it appears in a single expression statement and | |||
5474 | * if its may access relation is equal to its must access relation. | |||
5475 | * | |||
5476 | * The combined set of may accesses may be a union if member accesses | |||
5477 | * are involved, but the entire set is derived from a single reference and | |||
5478 | * therefore from a single index expression. These accesses therefore | |||
5479 | * all map to the same outer array. | |||
5480 | */ | |||
5481 | static int extract_access(__isl_keep pet_expr *expr, void *user) | |||
5482 | { | |||
5483 | struct ppcg_extract_access_data *data = user; | |||
5484 | isl_union_map *tagged; | |||
5485 | struct gpu_stmt_access *access; | |||
5486 | isl_ctx *ctx = pet_expr_get_ctx(expr); | |||
5487 | isl_multi_pw_aff *index; | |||
5488 | ||||
5489 | access = isl_alloc_type(ctx, struct gpu_stmt_access)((struct gpu_stmt_access *)isl_malloc_or_die(ctx, sizeof(struct gpu_stmt_access))); | |||
5490 | assert(access)((void) sizeof ((access) ? 1 : 0), __extension__ ({ if (access ) ; else __assert_fail ("access", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 5490, __extension__ __PRETTY_FUNCTION__); })); | |||
5491 | access->next = NULL((void*)0); | |||
5492 | access->read = pet_expr_access_is_read(expr); | |||
5493 | access->write = pet_expr_access_is_write(expr); | |||
5494 | tagged = pet_expr_access_get_tagged_may_read(expr); | |||
5495 | tagged = isl_union_map_union(tagged, | |||
5496 | pet_expr_access_get_tagged_may_write(expr)); | |||
5497 | tagged = isl_union_map_apply_range(tagged, | |||
5498 | isl_union_map_copy(data->any_to_outer)); | |||
5499 | if (!access->write) { | |||
5500 | access->exact_write = 1; | |||
5501 | } else if (!data->single_expression) { | |||
5502 | access->exact_write = 0; | |||
5503 | } else { | |||
5504 | isl_union_map *must, *may; | |||
5505 | may = isl_union_map_copy(tagged); | |||
5506 | may = isl_union_map_domain_factor_domain(may); | |||
5507 | must = pet_expr_access_get_must_write(expr); | |||
5508 | access->exact_write = isl_union_map_is_equal(must, may); | |||
5509 | isl_union_map_free(must); | |||
5510 | isl_union_map_free(may); | |||
5511 | } | |||
5512 | index = pet_expr_access_get_index(expr); | |||
5513 | access->n_index = isl_multi_pw_aff_dim(index, isl_dim_out); | |||
5514 | isl_multi_pw_aff_free(index); | |||
5515 | access->ref_id = pet_expr_access_get_ref_id(expr); | |||
5516 | access->tagged_access = extract_single_tagged_access(tagged, expr); | |||
5517 | access->access = isl_map_copy(access->tagged_access); | |||
5518 | access->access = isl_map_domain_factor_domain(access->access); | |||
5519 | access->fixed_element = accesses_fixed_element(expr); | |||
5520 | ||||
5521 | *data->next_access = access; | |||
5522 | data->next_access = &(*data->next_access)->next; | |||
5523 | ||||
5524 | if (!access->access || access->fixed_element < 0) | |||
5525 | return -1; | |||
5526 | ||||
5527 | return 0; | |||
5528 | } | |||
5529 | ||||
5530 | /* Construct a linked list of gpu_stmt_access objects, | |||
5531 | * one for each access expression in the statement body. | |||
5532 | * "any_to_outer" maps all intermediate arrays to their outer arrays. | |||
5533 | */ | |||
5534 | static int pet_stmt_extract_accesses(struct gpu_stmt *stmt, | |||
5535 | __isl_keep isl_union_map *any_to_outer) | |||
5536 | { | |||
5537 | struct ppcg_extract_access_data data; | |||
5538 | ||||
5539 | stmt->accesses = NULL((void*)0); | |||
5540 | data.next_access = &stmt->accesses; | |||
5541 | data.single_expression = | |||
5542 | pet_tree_get_type(stmt->stmt->body) == pet_tree_expr; | |||
5543 | data.any_to_outer = any_to_outer; | |||
5544 | return pet_tree_foreach_access_expr(stmt->stmt->body, | |||
5545 | &extract_access, &data); | |||
5546 | } | |||
5547 | ||||
5548 | /* Has statement "stmt" been killed from "scop"? | |||
5549 | * That is, is the instance set of "scop" free from any | |||
5550 | * instances of "stmt"? | |||
5551 | */ | |||
5552 | static isl_bool is_stmt_killed(struct ppcg_scop *scop, struct pet_stmt *stmt) | |||
5553 | { | |||
5554 | isl_space *space; | |||
5555 | isl_set *left; | |||
5556 | isl_bool empty; | |||
5557 | ||||
5558 | if (!scop || !stmt) | |||
5559 | return isl_bool_error; | |||
5560 | space = isl_set_get_space(stmt->domain); | |||
5561 | left = isl_union_set_extract_set(scop->domain, space); | |||
5562 | empty = isl_set_plain_is_empty(left); | |||
5563 | isl_set_free(left); | |||
5564 | ||||
5565 | return empty; | |||
5566 | } | |||
5567 | ||||
5568 | /* Return an array of gpu_stmt representing the statements in "scop". | |||
5569 | * Do not collect array accesses for statements that have been killed. | |||
5570 | */ | |||
5571 | static struct gpu_stmt *extract_stmts(isl_ctx *ctx, struct ppcg_scop *scop, | |||
5572 | __isl_keep isl_union_map *any_to_outer) | |||
5573 | { | |||
5574 | int i; | |||
5575 | struct gpu_stmt *stmts; | |||
5576 | ||||
5577 | stmts = isl_calloc_array(ctx, struct gpu_stmt, scop->pet->n_stmt)((struct gpu_stmt *)isl_calloc_or_die(ctx, scop->pet->n_stmt , sizeof(struct gpu_stmt))); | |||
5578 | if (!stmts) | |||
5579 | return NULL((void*)0); | |||
5580 | ||||
5581 | for (i = 0; i < scop->pet->n_stmt; ++i) { | |||
5582 | struct gpu_stmt *s = &stmts[i]; | |||
5583 | isl_bool killed; | |||
5584 | ||||
5585 | s->id = isl_set_get_tuple_id(scop->pet->stmts[i]->domain); | |||
5586 | s->stmt = scop->pet->stmts[i]; | |||
5587 | killed = is_stmt_killed(scop, scop->pet->stmts[i]); | |||
5588 | if (killed < 0) | |||
5589 | return free_stmts(stmts, i + 1); | |||
5590 | if (killed) | |||
5591 | continue; | |||
5592 | if (pet_stmt_extract_accesses(s, any_to_outer) < 0) | |||
5593 | return free_stmts(stmts, i + 1); | |||
5594 | } | |||
5595 | ||||
5596 | return stmts; | |||
5597 | } | |||
5598 | ||||
5599 | /* Generate CUDA code for "scop" and print it to "p". | |||
5600 | * After generating an AST for the transformed scop as explained below, | |||
5601 | * we call "gen->print" to print the AST in the desired output format | |||
5602 | * to "p". | |||
5603 | * | |||
5604 | * If it turns out that it does not make sense to generate GPU code, | |||
5605 | * then we generate CPU code instead. | |||
5606 | * | |||
5607 | * The declarations of the arrays that are visible outside of the scop | |||
5608 | * are printed outside of the code generated from the schedule, | |||
5609 | * because the generated code may involve a guard around the entire code. | |||
5610 | * | |||
5611 | * We first compute a schedule that respects the dependences | |||
5612 | * of the original program and select the outermost bands | |||
5613 | * of tilable dimensions that have at least one parallel loop. | |||
5614 | * If the --load-schedule is specified, then the loaded schedule | |||
5615 | * is used instead of a computed schedule. | |||
5616 | * | |||
5617 | * Each of these bands B is then tiled according to "tile" sizes, resulting | |||
5618 | * in two nested bands, with a kernel marker on top | |||
5619 | * | |||
5620 | * K | |||
5621 | * | | |||
5622 | * T | |||
5623 | * | | |||
5624 | * P | |||
5625 | * | |||
5626 | * We then split off at most 2 parallel dimensions from the T band and | |||
5627 | * at most 3 parallel dimension from the P band | |||
5628 | * | |||
5629 | * K | |||
5630 | * | | |||
5631 | * T | |||
5632 | * T1 | |||
5633 | * | | |||
5634 | * T2 | |||
5635 | * | | |||
5636 | * P1 | |||
5637 | * | | |||
5638 | * P2 | |||
5639 | * | |||
5640 | * A filter is introduced in front of T1 that maps the domain instances | |||
5641 | * to block identifiers. Similarly, a filter is introduced in front of P1 | |||
5642 | * that maps the domain instances to thread identifiers. | |||
5643 | * | |||
5644 | * For each iteration of the T2 band and for each array, we compute | |||
5645 | * the array elements accessed by that iteration, construct a rectangular | |||
5646 | * box around it and shift it to the origin. The result is used | |||
5647 | * as shared memory for the array. | |||
5648 | * | |||
5649 | * Copying and synchronization statements are added to this schedule tree. | |||
5650 | * In principle, these are added in front of the P1 band, but some of | |||
5651 | * them may get hoisted up to higher levels. | |||
5652 | * | |||
5653 | * The entire AST is then generated from the single resulting schedule tree. | |||
5654 | * During the generation the subtrees at kernel nodes (K) are saved | |||
5655 | * aside and replaced by kernel calls. The result is printed as host code | |||
5656 | * while the saved subtrees are printed as device code. | |||
5657 | */ | |||
5658 | static __isl_give isl_printer *generate(__isl_take isl_printer *p, | |||
5659 | struct gpu_gen *gen, struct ppcg_scop *scop, | |||
5660 | struct ppcg_options *options) | |||
5661 | { | |||
5662 | struct gpu_prog *prog; | |||
5663 | isl_ctx *ctx; | |||
5664 | isl_schedule *schedule; | |||
5665 | int any_permutable; | |||
5666 | ||||
5667 | if (!scop) | |||
5668 | return isl_printer_free(p); | |||
5669 | ||||
5670 | ctx = isl_printer_get_ctx(p); | |||
5671 | prog = gpu_prog_alloc(ctx, scop); | |||
5672 | if (!prog) | |||
5673 | return isl_printer_free(p); | |||
5674 | ||||
5675 | gen->prog = prog; | |||
5676 | schedule = get_schedule(gen); | |||
5677 | ||||
5678 | any_permutable = has_any_permutable_node(schedule); | |||
5679 | if (any_permutable < 0 || !any_permutable) { | |||
5680 | if (any_permutable < 0) | |||
5681 | p = isl_printer_free(p); | |||
5682 | else | |||
5683 | p = print_cpu(p, scop, options); | |||
5684 | isl_schedule_free(schedule); | |||
5685 | } else { | |||
5686 | const int create_to_from_device = 1; | |||
5687 | schedule = map_to_device(gen, schedule, create_to_from_device); | |||
5688 | gen->tree = generate_code(gen, schedule); | |||
5689 | p = ppcg_set_macro_names(p); | |||
5690 | p = ppcg_print_exposed_declarations(p, prog->scop); | |||
5691 | p = gen->print(p, gen->prog, gen->tree, &gen->types, | |||
5692 | gen->print_user); | |||
5693 | isl_ast_node_free(gen->tree); | |||
5694 | } | |||
5695 | ||||
5696 | gpu_prog_free(prog); | |||
5697 | ||||
5698 | return p; | |||
5699 | } | |||
5700 | ||||
5701 | /* Wrapper around generate for use as a ppcg_transform callback. | |||
5702 | */ | |||
5703 | static __isl_give isl_printer *generate_wrap(__isl_take isl_printer *p, | |||
5704 | struct ppcg_scop *scop, void *user) | |||
5705 | { | |||
5706 | struct gpu_gen *gen = user; | |||
5707 | ||||
5708 | return generate(p, gen, scop, gen->options); | |||
5709 | } | |||
5710 | ||||
5711 | /* Transform the code in the file called "input" by replacing | |||
5712 | * all scops by corresponding GPU code and write the results to "out". | |||
5713 | */ | |||
5714 | int generate_gpu(isl_ctx *ctx, const char *input, FILE *out, | |||
5715 | struct ppcg_options *options, | |||
5716 | __isl_give isl_printer *(*print)(__isl_take isl_printer *p, | |||
5717 | struct gpu_prog *prog, __isl_keep isl_ast_node *tree, | |||
5718 | struct gpu_types *types, void *user), void *user) | |||
5719 | { | |||
5720 | struct gpu_gen gen; | |||
5721 | int r; | |||
5722 | int i; | |||
5723 | ||||
5724 | gen.ctx = ctx; | |||
5725 | gen.sizes = extract_sizes_from_str(ctx, options->sizes); | |||
5726 | gen.options = options; | |||
5727 | gen.kernel_id = 0; | |||
5728 | gen.print = print; | |||
5729 | gen.print_user = user; | |||
5730 | gen.types.n = 0; | |||
5731 | gen.types.name = NULL((void*)0); | |||
5732 | ||||
5733 | if (options->debug->dump_sizes) { | |||
5734 | isl_space *space = isl_space_params_alloc(ctx, 0); | |||
5735 | gen.used_sizes = isl_union_map_empty(space); | |||
5736 | } | |||
5737 | ||||
5738 | r = ppcg_transform(ctx, input, out, options, &generate_wrap, &gen); | |||
5739 | ||||
5740 | if (options->debug->dump_sizes) { | |||
5741 | isl_union_map_dump(gen.used_sizes); | |||
5742 | isl_union_map_free(gen.used_sizes); | |||
5743 | } | |||
5744 | ||||
5745 | isl_union_map_free(gen.sizes); | |||
5746 | for (i = 0; i < gen.types.n; ++i) | |||
5747 | free(gen.types.name[i]); | |||
5748 | free(gen.types.name); | |||
5749 | ||||
5750 | return r; | |||
5751 | } | |||
5752 | ||||
5753 | /* Compute the set of inner array elements that may have their values | |||
5754 | * preserved by "prog". In particular, collect the array elements of | |||
5755 | * arrays that are not local to "prog" and remove those elements that | |||
5756 | * are definitely killed or definitely written by "prog". | |||
5757 | */ | |||
5758 | __isl_give isl_union_set *compute_may_persist(struct gpu_prog *prog) | |||
5759 | { | |||
5760 | int i; | |||
5761 | isl_union_set *may_persist, *killed; | |||
5762 | isl_union_map *must_kill; | |||
5763 | ||||
5764 | may_persist = isl_union_set_empty(isl_set_get_space(prog->context)); | |||
5765 | for (i = 0; i < prog->n_array; ++i) { | |||
5766 | isl_set *extent; | |||
5767 | ||||
5768 | if (prog->array[i].local) | |||
5769 | continue; | |||
5770 | ||||
5771 | extent = isl_set_copy(prog->array[i].extent); | |||
5772 | may_persist = isl_union_set_add_set(may_persist, extent); | |||
5773 | } | |||
5774 | ||||
5775 | may_persist = isl_union_set_intersect_params(may_persist, | |||
5776 | isl_set_copy(prog->context)); | |||
5777 | may_persist = isl_union_set_apply(may_persist, | |||
5778 | isl_union_map_copy(prog->to_inner)); | |||
5779 | must_kill = isl_union_map_copy(prog->tagged_must_kill); | |||
5780 | killed = isl_union_map_range(must_kill); | |||
5781 | must_kill = isl_union_map_copy(prog->must_write); | |||
5782 | killed = isl_union_set_union(killed, isl_union_map_range(must_kill)); | |||
5783 | ||||
5784 | may_persist = isl_union_set_subtract(may_persist, killed); | |||
5785 | return may_persist; | |||
5786 | } | |||
5787 | ||||
5788 | struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct ppcg_scop *scop) | |||
5789 | { | |||
5790 | struct gpu_prog *prog; | |||
5791 | isl_space *space; | |||
5792 | isl_map *id; | |||
5793 | ||||
5794 | if (!scop) | |||
5795 | return NULL((void*)0); | |||
5796 | ||||
5797 | prog = isl_calloc_type(ctx, struct gpu_prog)((struct gpu_prog *)isl_calloc_or_die(ctx, 1, sizeof(struct gpu_prog ))); | |||
5798 | assert(prog)((void) sizeof ((prog) ? 1 : 0), __extension__ ({ if (prog) ; else __assert_fail ("prog", "/build/llvm-toolchain-snapshot-7~svn329677/tools/polly/lib/External/ppcg/gpu.c" , 5798, __extension__ __PRETTY_FUNCTION__); })); | |||
5799 | ||||
5800 | prog->ctx = ctx; | |||
5801 | prog->scop = scop; | |||
5802 | prog->context = isl_set_copy(scop->context); | |||
5803 | prog->n_stmts = scop->pet->n_stmt; | |||
5804 | prog->any_to_outer = pet_scop_compute_outer_to_any(scop->pet); | |||
5805 | prog->any_to_outer = isl_union_map_reverse(prog->any_to_outer); | |||
5806 | space = isl_union_map_get_space(prog->any_to_outer); | |||
5807 | space = isl_space_set_from_params(space); | |||
5808 | space = isl_space_add_dims(space, isl_dim_set, 1); | |||
5809 | space = isl_space_map_from_set(space); | |||
5810 | id = isl_map_identity(space); | |||
5811 | prog->any_to_outer = isl_union_map_add_map(prog->any_to_outer, id); | |||
5812 | prog->stmts = extract_stmts(ctx, scop, prog->any_to_outer); | |||
5813 | prog->read = isl_union_map_copy(scop->reads); | |||
5814 | prog->may_write = isl_union_map_copy(scop->may_writes); | |||
5815 | prog->must_write = isl_union_map_copy(scop->must_writes); | |||
5816 | prog->tagged_must_kill = isl_union_map_copy(scop->tagged_must_kills); | |||
5817 | prog->to_inner = pet_scop_compute_outer_to_inner(scop->pet); | |||
5818 | prog->to_outer = isl_union_map_copy(prog->to_inner); | |||
5819 | prog->to_outer = isl_union_map_reverse(prog->to_outer); | |||
5820 | ||||
5821 | if (!prog->stmts) | |||
5822 | return gpu_prog_free(prog); | |||
5823 | ||||
5824 | if (collect_array_info(prog) < 0) | |||
5825 | return gpu_prog_free(prog); | |||
5826 | prog->may_persist = compute_may_persist(prog); | |||
5827 | ||||
5828 | return prog; | |||
5829 | } | |||
5830 | ||||
5831 | void *gpu_prog_free(struct gpu_prog *prog) | |||
5832 | { | |||
5833 | if (!prog) | |||
5834 | return NULL((void*)0); | |||
5835 | free_array_info(prog); | |||
5836 | free_stmts(prog->stmts, prog->n_stmts); | |||
5837 | isl_union_map_free(prog->any_to_outer); | |||
5838 | isl_union_map_free(prog->to_outer); | |||
5839 | isl_union_map_free(prog->to_inner); | |||
5840 | isl_union_map_free(prog->read); | |||
5841 | isl_union_map_free(prog->may_write); | |||
5842 | isl_union_map_free(prog->must_write); | |||
5843 | isl_union_map_free(prog->tagged_must_kill); | |||
5844 | isl_union_map_free(prog->array_order); | |||
5845 | isl_union_set_free(prog->may_persist); | |||
5846 | isl_set_free(prog->context); | |||
5847 | free(prog); | |||
5848 | return NULL((void*)0); | |||
5849 | } |