2 * Copyright 2011 INRIA Saclay
4 * Use of this software is governed by the GNU LGPLv2.1 license
6 * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
7 * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
15 #include <isl/options.h>
16 #include <isl/schedule.h>
19 #include "ppcg_options.h"
24 struct isl_options
*isl
;
25 struct pet_options
*pet
;
26 struct ppcg_options
*ppcg
;
30 const char *ppcg_version(void);
31 static void print_version(void)
33 printf("%s", ppcg_version());
36 ISL_ARGS_START(struct options
, options_args
)
37 ISL_ARG_CHILD(struct options
, isl
, "isl", &isl_options_args
, "isl options")
38 ISL_ARG_CHILD(struct options
, pet
, "pet", &pet_options_args
, "pet options")
39 ISL_ARG_CHILD(struct options
, ppcg
, NULL
, &ppcg_options_args
, "ppcg options")
40 ISL_ARG_ARG(struct options
, input
, "input", NULL
)
41 ISL_ARG_VERSION(print_version
)
44 ISL_ARG_DEF(options
, struct options
, options_args
)
46 /* Is "stmt" a kill statement?
48 static int is_kill(struct pet_stmt
*stmt
)
50 if (stmt
->body
->type
!= pet_expr_unary
)
52 return stmt
->body
->op
== pet_op_kill
;
55 /* Is "stmt" not a kill statement?
57 static int is_not_kill(struct pet_stmt
*stmt
)
59 return !is_kill(stmt
);
62 /* Collect the iteration domains of the statements in "scop" that
65 static __isl_give isl_union_set
*collect_domains(struct pet_scop
*scop
,
66 int (*pred
)(struct pet_stmt
*stmt
))
70 isl_union_set
*domain
;
75 domain
= isl_union_set_empty(isl_set_get_space(scop
->context
));
77 for (i
= 0; i
< scop
->n_stmt
; ++i
) {
78 struct pet_stmt
*stmt
= scop
->stmts
[i
];
82 domain_i
= isl_set_copy(scop
->stmts
[i
]->domain
);
83 domain
= isl_union_set_add_set(domain
, domain_i
);
89 /* Collect the iteration domains of the statements in "scop",
90 * skipping kill statements.
92 static __isl_give isl_union_set
*collect_non_kill_domains(struct pet_scop
*scop
)
94 return collect_domains(scop
, &is_not_kill
);
97 /* Does "expr" contain any call expressions?
99 static int expr_has_call(struct pet_expr
*expr
)
103 if (expr
->type
== pet_expr_call
)
106 for (i
= 0; i
< expr
->n_arg
; ++i
)
107 if (expr_has_call(expr
->args
[i
]))
113 /* Does "stmt" contain any call expressions?
115 static int has_call(struct pet_stmt
*stmt
)
117 return expr_has_call(stmt
->body
);
120 /* Collect the iteration domains of the statements in "scop"
121 * that contain a call expression.
123 static __isl_give isl_union_set
*collect_call_domains(struct pet_scop
*scop
)
125 return collect_domains(scop
, &has_call
);
128 /* Collect all kill accesses in "scop".
130 static __isl_give isl_union_map
*collect_kills(struct pet_scop
*scop
)
133 isl_union_map
*kills
;
138 kills
= isl_union_map_empty(isl_set_get_space(scop
->context
));
140 for (i
= 0; i
< scop
->n_stmt
; ++i
) {
141 struct pet_stmt
*stmt
= scop
->stmts
[i
];
146 kill_i
= isl_map_copy(stmt
->body
->args
[0]->acc
.access
);
147 kills
= isl_union_map_add_map(kills
, kill_i
);
153 /* Compute (flow) dependences and store the resulting flow dependences
154 * in scop->dep_flow and the reads with no corresponding writes in
157 static void compute_dependences(struct ppcg_scop
*scop
)
159 isl_union_map
*empty
;
164 empty
= isl_union_map_empty(isl_union_set_get_space(scop
->domain
));
165 isl_union_map_compute_flow(isl_union_map_copy(scop
->reads
),
166 isl_union_map_copy(scop
->writes
), empty
,
167 isl_union_map_copy(scop
->schedule
),
168 &scop
->dep_flow
, NULL
, &scop
->live_in
, NULL
);
171 /* Eliminate dead code from ps->domain.
173 * In particular, intersect ps->domain with the (parts of) iteration
174 * domains that are needed to produce the output or for statement
175 * iterations that call functions.
177 * We start with the iteration domains that call functions
178 * and the set of iterations that last write to an array
179 * (except those that are later killed).
181 * Then we add those statement iterations that produce
182 * something needed by the "live" statements iterations.
183 * We keep doing this until no more statement iterations can be added.
184 * To ensure that the procedure terminates, we compute the affine
185 * hull of the live iterations (bounded to the original iteration
186 * domains) each time we have added extra iterations.
188 static void eliminate_dead_code(struct ppcg_scop
*ps
)
190 isl_union_map
*exposed
;
194 exposed
= isl_union_map_union(isl_union_map_copy(ps
->writes
),
195 isl_union_map_copy(ps
->kills
));
196 exposed
= isl_union_map_reverse(exposed
);
197 exposed
= isl_union_map_apply_range(exposed
,
198 isl_union_map_copy(ps
->schedule
));
199 exposed
= isl_union_map_lexmax(exposed
);
200 exposed
= isl_union_map_apply_range(exposed
,
201 isl_union_map_reverse(isl_union_map_copy(ps
->schedule
)));
203 live
= isl_union_map_range(exposed
);
204 live
= isl_union_set_union(live
, isl_union_set_copy(ps
->call
));
206 dep
= isl_union_map_copy(ps
->dep_flow
);
207 dep
= isl_union_map_reverse(dep
);
210 isl_union_set
*extra
;
212 extra
= isl_union_set_apply(isl_union_set_copy(live
),
213 isl_union_map_copy(dep
));
214 if (isl_union_set_is_subset(extra
, live
)) {
215 isl_union_set_free(extra
);
219 live
= isl_union_set_union(live
, extra
);
220 live
= isl_union_set_affine_hull(live
);
221 live
= isl_union_set_intersect(live
,
222 isl_union_set_copy(ps
->domain
));
225 isl_union_map_free(dep
);
227 ps
->domain
= isl_union_set_intersect(ps
->domain
, live
);
230 /* Extract a ppcg_scop from a pet_scop.
232 * The constructed ppcg_scop refers to elements from the pet_scop
233 * so the pet_scop should not be freed before the ppcg_scop.
235 static struct ppcg_scop
*ppcg_scop_from_pet_scop(struct pet_scop
*scop
)
238 struct ppcg_scop
*ps
;
243 ctx
= isl_set_get_ctx(scop
->context
);
245 ps
= isl_calloc_type(ctx
, struct ppcg_scop
);
249 ps
->context
= isl_set_copy(scop
->context
);
250 ps
->domain
= collect_non_kill_domains(scop
);
251 ps
->call
= collect_call_domains(scop
);
252 ps
->reads
= pet_scop_collect_reads(scop
);
253 ps
->writes
= pet_scop_collect_writes(scop
);
254 ps
->kills
= collect_kills(scop
);
255 ps
->schedule
= pet_scop_collect_schedule(scop
);
256 ps
->n_array
= scop
->n_array
;
257 ps
->arrays
= scop
->arrays
;
258 ps
->n_stmt
= scop
->n_stmt
;
259 ps
->stmts
= scop
->stmts
;
261 compute_dependences(ps
);
262 eliminate_dead_code(ps
);
267 static void ppcg_scop_free(struct ppcg_scop
*ps
)
272 isl_set_free(ps
->context
);
273 isl_union_set_free(ps
->domain
);
274 isl_union_set_free(ps
->call
);
275 isl_union_map_free(ps
->reads
);
276 isl_union_map_free(ps
->live_in
);
277 isl_union_map_free(ps
->writes
);
278 isl_union_map_free(ps
->kills
);
279 isl_union_map_free(ps
->dep_flow
);
280 isl_union_map_free(ps
->schedule
);
285 int main(int argc
, char **argv
)
289 struct options
*options
;
290 struct pet_scop
*scop
;
291 struct ppcg_scop
*ps
;
293 options
= options_new_with_defaults();
296 ctx
= isl_ctx_alloc_with_options(&options_args
, options
);
297 isl_options_set_schedule_outer_zero_distance(ctx
, 1);
298 argc
= options_parse(options
, argc
, argv
, ISL_ARG_ALL
);
300 scop
= pet_scop_extract_from_C_source(ctx
, options
->input
, NULL
);
301 scop
= pet_scop_align_params(scop
);
302 ps
= ppcg_scop_from_pet_scop(scop
);
304 if (options
->ppcg
->target
== PPCG_TARGET_CUDA
)
305 r
= generate_cuda(ctx
, ps
, options
->ppcg
, options
->input
);
307 r
= generate_cpu(ctx
, ps
, options
->ppcg
, options
->input
);