1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
|
//===- Schedule.cpp - Calculate an optimized schedule ---------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass the isl to calculate a schedule that is optimized for parallelism
// and tileablility. The algorithm used in isl is an optimized version of the
// algorithm described in following paper:
//
// U. Bondhugula, A. Hartono, J. Ramanujam, and P. Sadayappan.
// A Practical Automatic Polyhedral Parallelizer and Locality Optimizer.
// In Proceedings of the 2008 ACM SIGPLAN Conference On Programming Language
// Design and Implementation, PLDI ’08, pages 101–113. ACM, 2008.
//===----------------------------------------------------------------------===//
#include "polly/Cloog.h"
#include "polly/LinkAllPasses.h"
#include "polly/Support/GICHelper.h"
#include "polly/Dependences.h"
#include "polly/ScopInfo.h"
#include "isl/space.h"
#include "isl/map.h"
#include "isl/constraint.h"
#include "isl/schedule.h"
#include "isl/band.h"
#define DEBUG_TYPE "polly-optimize-isl"
#include "llvm/Support/Debug.h"
#include "llvm/Support/CommandLine.h"
using namespace llvm;
using namespace polly;
static cl::opt<bool>
DisableTiling("polly-no-tiling",
cl::desc("Disable tiling in the isl scheduler"), cl::Hidden,
cl::init(false));
static cl::opt<bool>
Prevector("polly-prevector",
cl::desc("Enable prevectorization in the isl scheduler"), cl::Hidden,
cl::value_desc("Prevectorization enabled"),
cl::init(false));
namespace {
class IslScheduleOptimizer : public ScopPass {
public:
static char ID;
explicit IslScheduleOptimizer() : ScopPass(ID) {}
virtual bool runOnScop(Scop &S);
void printScop(llvm::raw_ostream &OS) const;
void getAnalysisUsage(AnalysisUsage &AU) const;
};
}
char IslScheduleOptimizer::ID = 0;
static int getSingleMap(__isl_take isl_map *map, void *user) {
isl_map **singleMap = (isl_map **) user;
*singleMap = map;
return 0;
}
static void extendScattering(Scop &S, unsigned NewDimensions) {
for (Scop::iterator SI = S.begin(), SE = S.end(); SI != SE; ++SI) {
ScopStmt *Stmt = *SI;
if (Stmt->isFinalRead())
continue;
unsigned OldDimensions = Stmt->getNumScattering();
isl_space *Space;
isl_basic_map *ChangeScattering;
Space = isl_space_alloc(Stmt->getIslCtx(), 0, OldDimensions, NewDimensions);
ChangeScattering = isl_basic_map_universe(isl_space_copy(Space));
isl_local_space *LocalSpace = isl_local_space_from_space(Space);
for (unsigned i = 0; i < OldDimensions; i++) {
isl_constraint *c = isl_equality_alloc(isl_local_space_copy(LocalSpace));
isl_constraint_set_coefficient_si(c, isl_dim_in, i, 1);
isl_constraint_set_coefficient_si(c, isl_dim_out, i, -1);
ChangeScattering = isl_basic_map_add_constraint(ChangeScattering, c);
}
for (unsigned i = OldDimensions; i < NewDimensions; i++) {
isl_constraint *c = isl_equality_alloc(isl_local_space_copy(LocalSpace));
isl_constraint_set_coefficient_si(c, isl_dim_out, i, 1);
ChangeScattering = isl_basic_map_add_constraint(ChangeScattering, c);
}
isl_map *ChangeScatteringMap = isl_map_from_basic_map(ChangeScattering);
ChangeScatteringMap = isl_map_align_params(ChangeScatteringMap,
S.getParamSpace());
isl_map *NewScattering = isl_map_apply_range(Stmt->getScattering(),
ChangeScatteringMap);
Stmt->setScattering(NewScattering);
isl_local_space_free(LocalSpace);
}
}
// getTileMap - Create a map that describes a n-dimensonal tiling.
//
// getTileMap creates a map from a n-dimensional scattering space into an
// 2*n-dimensional scattering space. The map describes a rectangular tiling.
//
// Example:
// scheduleDimensions = 2, parameterDimensions = 1, tileSize = 32
//
// tileMap := [p0] -> {[s0, s1] -> [t0, t1, s0, s1]:
// t0 % 32 = 0 and t0 <= s0 < t0 + 32 and
// t1 % 32 = 0 and t1 <= s1 < t1 + 32}
//
// Before tiling:
//
// for (i = 0; i < N; i++)
// for (j = 0; j < M; j++)
// S(i,j)
//
// After tiling:
//
// for (t_i = 0; t_i < N; i+=32)
// for (t_j = 0; t_j < M; j+=32)
// for (i = t_i; i < min(t_i + 32, N); i++) | Unknown that N % 32 = 0
// for (j = t_j; j < t_j + 32; j++) | Known that M % 32 = 0
// S(i,j)
//
static isl_basic_map *getTileMap(isl_ctx *ctx, int scheduleDimensions,
isl_space *SpaceModel, int tileSize = 32) {
// We construct
//
// tileMap := [p0] -> {[s0, s1] -> [t0, t1, p0, p1, a0, a1]:
// s0 = a0 * 32 and s0 = p0 and t0 <= p0 < t0 + 32 and
// s1 = a1 * 32 and s1 = p1 and t1 <= p1 < t1 + 32}
//
// and project out the auxilary dimensions a0 and a1.
isl_space *Space = isl_space_alloc(ctx, 0, scheduleDimensions,
scheduleDimensions * 3);
isl_basic_map *tileMap = isl_basic_map_universe(isl_space_copy(Space));
isl_local_space *LocalSpace = isl_local_space_from_space(Space);
for (int x = 0; x < scheduleDimensions; x++) {
int sX = x;
int tX = x;
int pX = scheduleDimensions + x;
int aX = 2 * scheduleDimensions + x;
isl_constraint *c;
// sX = aX * tileSize;
c = isl_equality_alloc(isl_local_space_copy(LocalSpace));
isl_constraint_set_coefficient_si(c, isl_dim_out, sX, 1);
isl_constraint_set_coefficient_si(c, isl_dim_out, aX, -tileSize);
tileMap = isl_basic_map_add_constraint(tileMap, c);
// pX = sX;
c = isl_equality_alloc(isl_local_space_copy(LocalSpace));
isl_constraint_set_coefficient_si(c, isl_dim_out, pX, 1);
isl_constraint_set_coefficient_si(c, isl_dim_in, sX, -1);
tileMap = isl_basic_map_add_constraint(tileMap, c);
// tX <= pX
c = isl_inequality_alloc(isl_local_space_copy(LocalSpace));
isl_constraint_set_coefficient_si(c, isl_dim_out, pX, 1);
isl_constraint_set_coefficient_si(c, isl_dim_out, tX, -1);
tileMap = isl_basic_map_add_constraint(tileMap, c);
// pX <= tX + (tileSize - 1)
c = isl_inequality_alloc(isl_local_space_copy(LocalSpace));
isl_constraint_set_coefficient_si(c, isl_dim_out, tX, 1);
isl_constraint_set_coefficient_si(c, isl_dim_out, pX, -1);
isl_constraint_set_constant_si(c, tileSize - 1);
tileMap = isl_basic_map_add_constraint(tileMap, c);
}
// Project out auxilary dimensions.
//
// The auxilary dimensions are transformed into existentially quantified ones.
// This reduces the number of visible scattering dimensions and allows Cloog
// to produces better code.
tileMap = isl_basic_map_project_out(tileMap, isl_dim_out,
2 * scheduleDimensions,
scheduleDimensions);
isl_local_space_free(LocalSpace);
return tileMap;
}
isl_union_map *getTiledPartialSchedule(isl_band *band) {
isl_union_map *partialSchedule;
int scheduleDimensions;
isl_ctx *ctx;
isl_space *Space;
isl_basic_map *tileMap;
isl_union_map *tileUnionMap;
partialSchedule = isl_band_get_partial_schedule(band);
if (!DisableTiling) {
ctx = isl_union_map_get_ctx(partialSchedule);
Space= isl_union_map_get_space(partialSchedule);
scheduleDimensions = isl_band_n_member(band);
tileMap = getTileMap(ctx, scheduleDimensions, Space);
tileUnionMap = isl_union_map_from_map(isl_map_from_basic_map(tileMap));
tileUnionMap = isl_union_map_align_params(tileUnionMap, Space);
partialSchedule = isl_union_map_apply_range(partialSchedule, tileUnionMap);
}
return partialSchedule;
}
static isl_map *getPrevectorMap(isl_ctx *ctx, int vectorDimension,
int scheduleDimensions,
int parameterDimensions,
int vectorWidth = 4) {
assert (0 <= vectorDimension && vectorDimension < scheduleDimensions);
isl_space *Space = isl_space_alloc(ctx, parameterDimensions,
scheduleDimensions, scheduleDimensions + 2);
isl_basic_map *tilingMap = isl_basic_map_universe(isl_space_copy(Space));
isl_constraint *c;
isl_local_space *LocalSpace = isl_local_space_from_space(Space);
for (int i = 0; i < vectorDimension; i++) {
c = isl_equality_alloc(isl_local_space_copy(LocalSpace));
isl_constraint_set_coefficient_si(c, isl_dim_in, i, -1);
isl_constraint_set_coefficient_si(c, isl_dim_out, i, 1);
tilingMap = isl_basic_map_add_constraint(tilingMap, c);
}
for (int i = vectorDimension + 1; i < scheduleDimensions; i++) {
c = isl_equality_alloc(isl_local_space_copy(LocalSpace));
isl_constraint_set_coefficient_si(c, isl_dim_in, i, -1);
isl_constraint_set_coefficient_si(c, isl_dim_out, i, 1);
tilingMap = isl_basic_map_add_constraint(tilingMap, c);
}
int stepDimension = scheduleDimensions;
int auxilaryDimension = scheduleDimensions + 1;
c = isl_equality_alloc(isl_local_space_copy(LocalSpace));
isl_constraint_set_coefficient_si(c, isl_dim_out, vectorDimension, 1);
isl_constraint_set_coefficient_si(c, isl_dim_out, auxilaryDimension,
-vectorWidth);
tilingMap = isl_basic_map_add_constraint(tilingMap, c);
c = isl_equality_alloc(isl_local_space_copy(LocalSpace));
isl_constraint_set_coefficient_si(c, isl_dim_in, vectorDimension, -1);
isl_constraint_set_coefficient_si(c, isl_dim_out, stepDimension, 1);
tilingMap = isl_basic_map_add_constraint(tilingMap, c);
c = isl_inequality_alloc(isl_local_space_copy(LocalSpace));
isl_constraint_set_coefficient_si(c, isl_dim_out, vectorDimension, -1);
isl_constraint_set_coefficient_si(c, isl_dim_out, stepDimension, 1);
tilingMap = isl_basic_map_add_constraint(tilingMap, c);
c = isl_inequality_alloc(LocalSpace);
isl_constraint_set_coefficient_si(c, isl_dim_out, vectorDimension, 1);
isl_constraint_set_coefficient_si(c, isl_dim_out, stepDimension, -1);
isl_constraint_set_constant_si(c, vectorWidth- 1);
tilingMap = isl_basic_map_add_constraint(tilingMap, c);
// Project out auxilary dimensions (introduced to ensure 'ii % tileSize = 0')
//
// The real dimensions are transformed into existentially quantified ones.
// This reduces the number of visible scattering dimensions. Also, Cloog
// produces better code, if auxilary dimensions are existentially quantified.
tilingMap = isl_basic_map_project_out(tilingMap, isl_dim_out,
scheduleDimensions + 1, 1);
return isl_map_from_basic_map(tilingMap);
}
// tileBandList - Tile all bands contained in a band forest.
//
// Recursively walk the band forest and tile all bands in the forest. Return
// a schedule that describes the tiled scattering.
static isl_union_map *tileBandList(isl_band_list *blist) {
int numBands = isl_band_list_n_band(blist);
isl_union_map *finalSchedule = 0;
for (int i = 0; i < numBands; i++) {
isl_band *band;
isl_union_map *partialSchedule;
band = isl_band_list_get_band(blist, i);
partialSchedule = getTiledPartialSchedule(band);
int scheduleDimensions = isl_band_n_member(band);
isl_space *Space = isl_union_map_get_space(partialSchedule);
if (isl_band_has_children(band)) {
isl_band_list *children = isl_band_get_children(band);
isl_union_map *suffixSchedule = tileBandList(children);
partialSchedule = isl_union_map_flat_range_product(partialSchedule,
suffixSchedule);
isl_band_list_free(children);
} else if (Prevector) {
isl_map *tileMap;
isl_union_map *tileUnionMap;
isl_ctx *ctx;
ctx = isl_union_map_get_ctx(partialSchedule);
for (int i = scheduleDimensions - 1 ; i >= 0 ; i--) {
if (isl_band_member_is_zero_distance(band, i)) {
tileMap = getPrevectorMap(ctx, scheduleDimensions + i,
scheduleDimensions * 2, 0);
tileUnionMap = isl_union_map_from_map(tileMap);
tileUnionMap = isl_union_map_align_params(tileUnionMap,
isl_space_copy(Space));
partialSchedule = isl_union_map_apply_range(partialSchedule,
tileUnionMap);
break;
}
}
}
if (finalSchedule)
finalSchedule = isl_union_map_union(finalSchedule, partialSchedule);
else
finalSchedule = partialSchedule;
isl_band_free(band);
isl_space_free(Space);
}
return finalSchedule;
}
static isl_union_map *tileSchedule(isl_schedule *schedule) {
isl_band_list *blist = isl_schedule_get_band_forest(schedule);
isl_union_map *tiledSchedule = tileBandList(blist);
isl_band_list_free(blist);
return tiledSchedule;
}
bool IslScheduleOptimizer::runOnScop(Scop &S) {
Dependences *D = &getAnalysis<Dependences>();
// Build input data.
int dependencyKinds = Dependences::TYPE_RAW
| Dependences::TYPE_WAR
| Dependences::TYPE_WAW;
isl_union_map *validity = D->getDependences(dependencyKinds);
isl_union_map *proximity = D->getDependences(dependencyKinds);
isl_union_set *domain = NULL;
for (Scop::iterator SI = S.begin(), SE = S.end(); SI != SE; ++SI)
if ((*SI)->isFinalRead())
continue;
else if (!domain)
domain = isl_union_set_from_set((*SI)->getDomain());
else
domain = isl_union_set_union(domain,
isl_union_set_from_set((*SI)->getDomain()));
if (!domain)
return false;
DEBUG(dbgs() << "\n\nCompute schedule from: ");
DEBUG(dbgs() << "Domain := "; isl_union_set_dump(domain); dbgs() << ";\n");
DEBUG(dbgs() << "Proximity := "; isl_union_map_dump(proximity);
dbgs() << ";\n");
DEBUG(dbgs() << "Validity := "; isl_union_map_dump(validity);
dbgs() << ";\n");
isl_schedule *schedule;
schedule = isl_union_set_compute_schedule(domain, validity, proximity);
DEBUG(dbgs() << "Computed schedule: ");
DEBUG(dbgs() << stringFromIslObj(schedule));
DEBUG(dbgs() << "Individual bands: ");
isl_union_map *tiledSchedule = tileSchedule(schedule);
for (Scop::iterator SI = S.begin(), SE = S.end(); SI != SE; ++SI) {
ScopStmt *stmt = *SI;
if (stmt->isFinalRead())
continue;
isl_set *domain = stmt->getDomain();
isl_union_map *stmtBand;
stmtBand = isl_union_map_intersect_domain(isl_union_map_copy(tiledSchedule),
isl_union_set_from_set(domain));
isl_map *stmtSchedule;
isl_union_map_foreach_map(stmtBand, getSingleMap, &stmtSchedule);
stmt->setScattering(stmtSchedule);
isl_union_map_free(stmtBand);
}
isl_union_map_free(tiledSchedule);
isl_schedule_free(schedule);
unsigned maxScatDims = 0;
for (Scop::iterator SI = S.begin(), SE = S.end(); SI != SE; ++SI)
maxScatDims = std::max((*SI)->getNumScattering(), maxScatDims);
extendScattering(S, maxScatDims);
return false;
}
void IslScheduleOptimizer::printScop(raw_ostream &OS) const {
}
void IslScheduleOptimizer::getAnalysisUsage(AnalysisUsage &AU) const {
ScopPass::getAnalysisUsage(AU);
AU.addRequired<Dependences>();
}
INITIALIZE_PASS_BEGIN(IslScheduleOptimizer, "polly-optimize-isl",
"Polly - Optimize schedule of SCoP", false, false)
INITIALIZE_PASS_DEPENDENCY(Dependences)
INITIALIZE_PASS_DEPENDENCY(ScopInfo)
INITIALIZE_PASS_END(IslScheduleOptimizer, "polly-optimize-isl",
"Polly - Optimize schedule of SCoP", false, false)
Pass* polly::createIslScheduleOptimizerPass() {
return new IslScheduleOptimizer();
}
|