/
AMDGPUAtomicOptimizer.cpp
910 lines (787 loc) · 33.5 KB
/
AMDGPUAtomicOptimizer.cpp
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
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
//===-- AMDGPUAtomicOptimizer.cpp -----------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
/// \file
/// This pass optimizes atomic operations by using a single lane of a wavefront
/// to perform the atomic operation, thus reducing contention on that memory
/// location.
/// Atomic optimizer uses following strategies to compute scan and reduced
/// values
/// 1. DPP -
/// This is the most efficient implementation for scan. DPP uses Whole Wave
/// Mode (WWM)
/// 2. Iterative -
// An alternative implementation iterates over all active lanes
/// of Wavefront using llvm.cttz and performs scan using readlane & writelane
/// intrinsics
//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
#include "GCNSubtarget.h"
#include "llvm/Analysis/DomTreeUpdater.h"
#include "llvm/Analysis/UniformityAnalysis.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstVisitor.h"
#include "llvm/IR/IntrinsicsAMDGPU.h"
#include "llvm/InitializePasses.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#define DEBUG_TYPE "amdgpu-atomic-optimizer"
using namespace llvm;
using namespace llvm::AMDGPU;
namespace {
struct ReplacementInfo {
Instruction *I;
AtomicRMWInst::BinOp Op;
unsigned ValIdx;
bool ValDivergent;
};
class AMDGPUAtomicOptimizer : public FunctionPass {
public:
static char ID;
ScanOptions ScanImpl;
AMDGPUAtomicOptimizer(ScanOptions ScanImpl)
: FunctionPass(ID), ScanImpl(ScanImpl) {}
bool runOnFunction(Function &F) override;
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.addPreserved<DominatorTreeWrapperPass>();
AU.addRequired<UniformityInfoWrapperPass>();
AU.addRequired<TargetPassConfig>();
}
};
class AMDGPUAtomicOptimizerImpl
: public InstVisitor<AMDGPUAtomicOptimizerImpl> {
private:
SmallVector<ReplacementInfo, 8> ToReplace;
const UniformityInfo *UA;
const DataLayout *DL;
DomTreeUpdater &DTU;
const GCNSubtarget *ST;
bool IsPixelShader;
ScanOptions ScanImpl;
Value *buildReduction(IRBuilder<> &B, AtomicRMWInst::BinOp Op, Value *V,
Value *const Identity) const;
Value *buildScan(IRBuilder<> &B, AtomicRMWInst::BinOp Op, Value *V,
Value *const Identity) const;
Value *buildShiftRight(IRBuilder<> &B, Value *V, Value *const Identity) const;
std::pair<Value *, Value *>
buildScanIteratively(IRBuilder<> &B, AtomicRMWInst::BinOp Op,
Value *const Identity, Value *V, Instruction &I,
BasicBlock *ComputeLoop, BasicBlock *ComputeEnd) const;
void optimizeAtomic(Instruction &I, AtomicRMWInst::BinOp Op, unsigned ValIdx,
bool ValDivergent) const;
public:
AMDGPUAtomicOptimizerImpl() = delete;
AMDGPUAtomicOptimizerImpl(const UniformityInfo *UA, const DataLayout *DL,
DomTreeUpdater &DTU, const GCNSubtarget *ST,
bool IsPixelShader, ScanOptions ScanImpl)
: UA(UA), DL(DL), DTU(DTU), ST(ST), IsPixelShader(IsPixelShader),
ScanImpl(ScanImpl) {}
bool run(Function &F);
void visitAtomicRMWInst(AtomicRMWInst &I);
void visitIntrinsicInst(IntrinsicInst &I);
};
} // namespace
char AMDGPUAtomicOptimizer::ID = 0;
char &llvm::AMDGPUAtomicOptimizerID = AMDGPUAtomicOptimizer::ID;
bool AMDGPUAtomicOptimizer::runOnFunction(Function &F) {
if (skipFunction(F)) {
return false;
}
const UniformityInfo *UA =
&getAnalysis<UniformityInfoWrapperPass>().getUniformityInfo();
const DataLayout *DL = &F.getParent()->getDataLayout();
DominatorTreeWrapperPass *const DTW =
getAnalysisIfAvailable<DominatorTreeWrapperPass>();
DomTreeUpdater DTU(DTW ? &DTW->getDomTree() : nullptr,
DomTreeUpdater::UpdateStrategy::Lazy);
const TargetPassConfig &TPC = getAnalysis<TargetPassConfig>();
const TargetMachine &TM = TPC.getTM<TargetMachine>();
const GCNSubtarget *ST = &TM.getSubtarget<GCNSubtarget>(F);
bool IsPixelShader = F.getCallingConv() == CallingConv::AMDGPU_PS;
return AMDGPUAtomicOptimizerImpl(UA, DL, DTU, ST, IsPixelShader, ScanImpl)
.run(F);
}
PreservedAnalyses AMDGPUAtomicOptimizerPass::run(Function &F,
FunctionAnalysisManager &AM) {
const auto *UA = &AM.getResult<UniformityInfoAnalysis>(F);
const DataLayout *DL = &F.getParent()->getDataLayout();
DomTreeUpdater DTU(&AM.getResult<DominatorTreeAnalysis>(F),
DomTreeUpdater::UpdateStrategy::Lazy);
const GCNSubtarget *ST = &TM.getSubtarget<GCNSubtarget>(F);
bool IsPixelShader = F.getCallingConv() == CallingConv::AMDGPU_PS;
return AMDGPUAtomicOptimizerImpl(UA, DL, DTU, ST, IsPixelShader, ScanImpl)
.run(F)
? PreservedAnalyses::none()
: PreservedAnalyses::all();
}
bool AMDGPUAtomicOptimizerImpl::run(Function &F) {
visit(F);
const bool Changed = !ToReplace.empty();
for (ReplacementInfo &Info : ToReplace) {
optimizeAtomic(*Info.I, Info.Op, Info.ValIdx, Info.ValDivergent);
}
ToReplace.clear();
return Changed;
}
void AMDGPUAtomicOptimizerImpl::visitAtomicRMWInst(AtomicRMWInst &I) {
// Early exit for unhandled address space atomic instructions.
switch (I.getPointerAddressSpace()) {
default:
return;
case AMDGPUAS::GLOBAL_ADDRESS:
case AMDGPUAS::LOCAL_ADDRESS:
break;
}
AtomicRMWInst::BinOp Op = I.getOperation();
switch (Op) {
default:
return;
case AtomicRMWInst::Add:
case AtomicRMWInst::Sub:
case AtomicRMWInst::And:
case AtomicRMWInst::Or:
case AtomicRMWInst::Xor:
case AtomicRMWInst::Max:
case AtomicRMWInst::Min:
case AtomicRMWInst::UMax:
case AtomicRMWInst::UMin:
break;
}
const unsigned PtrIdx = 0;
const unsigned ValIdx = 1;
// If the pointer operand is divergent, then each lane is doing an atomic
// operation on a different address, and we cannot optimize that.
if (UA->isDivergentUse(I.getOperandUse(PtrIdx))) {
return;
}
const bool ValDivergent = UA->isDivergentUse(I.getOperandUse(ValIdx));
// If the value operand is divergent, each lane is contributing a different
// value to the atomic calculation. We can only optimize divergent values if
// we have DPP available on our subtarget, and the atomic operation is 32
// bits.
if (ValDivergent &&
(!ST->hasDPP() || DL->getTypeSizeInBits(I.getType()) != 32)) {
return;
}
// If we get here, we can optimize the atomic using a single wavefront-wide
// atomic operation to do the calculation for the entire wavefront, so
// remember the instruction so we can come back to it.
const ReplacementInfo Info = {&I, Op, ValIdx, ValDivergent};
ToReplace.push_back(Info);
}
void AMDGPUAtomicOptimizerImpl::visitIntrinsicInst(IntrinsicInst &I) {
AtomicRMWInst::BinOp Op;
switch (I.getIntrinsicID()) {
default:
return;
case Intrinsic::amdgcn_buffer_atomic_add:
case Intrinsic::amdgcn_struct_buffer_atomic_add:
case Intrinsic::amdgcn_struct_ptr_buffer_atomic_add:
case Intrinsic::amdgcn_raw_buffer_atomic_add:
case Intrinsic::amdgcn_raw_ptr_buffer_atomic_add:
Op = AtomicRMWInst::Add;
break;
case Intrinsic::amdgcn_buffer_atomic_sub:
case Intrinsic::amdgcn_struct_buffer_atomic_sub:
case Intrinsic::amdgcn_struct_ptr_buffer_atomic_sub:
case Intrinsic::amdgcn_raw_buffer_atomic_sub:
case Intrinsic::amdgcn_raw_ptr_buffer_atomic_sub:
Op = AtomicRMWInst::Sub;
break;
case Intrinsic::amdgcn_buffer_atomic_and:
case Intrinsic::amdgcn_struct_buffer_atomic_and:
case Intrinsic::amdgcn_struct_ptr_buffer_atomic_and:
case Intrinsic::amdgcn_raw_buffer_atomic_and:
case Intrinsic::amdgcn_raw_ptr_buffer_atomic_and:
Op = AtomicRMWInst::And;
break;
case Intrinsic::amdgcn_buffer_atomic_or:
case Intrinsic::amdgcn_struct_buffer_atomic_or:
case Intrinsic::amdgcn_struct_ptr_buffer_atomic_or:
case Intrinsic::amdgcn_raw_buffer_atomic_or:
case Intrinsic::amdgcn_raw_ptr_buffer_atomic_or:
Op = AtomicRMWInst::Or;
break;
case Intrinsic::amdgcn_buffer_atomic_xor:
case Intrinsic::amdgcn_struct_buffer_atomic_xor:
case Intrinsic::amdgcn_struct_ptr_buffer_atomic_xor:
case Intrinsic::amdgcn_raw_buffer_atomic_xor:
case Intrinsic::amdgcn_raw_ptr_buffer_atomic_xor:
Op = AtomicRMWInst::Xor;
break;
case Intrinsic::amdgcn_buffer_atomic_smin:
case Intrinsic::amdgcn_struct_buffer_atomic_smin:
case Intrinsic::amdgcn_struct_ptr_buffer_atomic_smin:
case Intrinsic::amdgcn_raw_buffer_atomic_smin:
case Intrinsic::amdgcn_raw_ptr_buffer_atomic_smin:
Op = AtomicRMWInst::Min;
break;
case Intrinsic::amdgcn_buffer_atomic_umin:
case Intrinsic::amdgcn_struct_buffer_atomic_umin:
case Intrinsic::amdgcn_struct_ptr_buffer_atomic_umin:
case Intrinsic::amdgcn_raw_buffer_atomic_umin:
case Intrinsic::amdgcn_raw_ptr_buffer_atomic_umin:
Op = AtomicRMWInst::UMin;
break;
case Intrinsic::amdgcn_buffer_atomic_smax:
case Intrinsic::amdgcn_struct_buffer_atomic_smax:
case Intrinsic::amdgcn_struct_ptr_buffer_atomic_smax:
case Intrinsic::amdgcn_raw_buffer_atomic_smax:
case Intrinsic::amdgcn_raw_ptr_buffer_atomic_smax:
Op = AtomicRMWInst::Max;
break;
case Intrinsic::amdgcn_buffer_atomic_umax:
case Intrinsic::amdgcn_struct_buffer_atomic_umax:
case Intrinsic::amdgcn_struct_ptr_buffer_atomic_umax:
case Intrinsic::amdgcn_raw_buffer_atomic_umax:
case Intrinsic::amdgcn_raw_ptr_buffer_atomic_umax:
Op = AtomicRMWInst::UMax;
break;
}
const unsigned ValIdx = 0;
const bool ValDivergent = UA->isDivergentUse(I.getOperandUse(ValIdx));
// If the value operand is divergent, each lane is contributing a different
// value to the atomic calculation. We can only optimize divergent values if
// we have DPP available on our subtarget, and the atomic operation is 32
// bits.
if (ValDivergent &&
(!ST->hasDPP() || DL->getTypeSizeInBits(I.getType()) != 32)) {
return;
}
// If any of the other arguments to the intrinsic are divergent, we can't
// optimize the operation.
for (unsigned Idx = 1; Idx < I.getNumOperands(); Idx++) {
if (UA->isDivergentUse(I.getOperandUse(Idx))) {
return;
}
}
// If we get here, we can optimize the atomic using a single wavefront-wide
// atomic operation to do the calculation for the entire wavefront, so
// remember the instruction so we can come back to it.
const ReplacementInfo Info = {&I, Op, ValIdx, ValDivergent};
ToReplace.push_back(Info);
}
// Use the builder to create the non-atomic counterpart of the specified
// atomicrmw binary op.
static Value *buildNonAtomicBinOp(IRBuilder<> &B, AtomicRMWInst::BinOp Op,
Value *LHS, Value *RHS) {
CmpInst::Predicate Pred;
switch (Op) {
default:
llvm_unreachable("Unhandled atomic op");
case AtomicRMWInst::Add:
return B.CreateBinOp(Instruction::Add, LHS, RHS);
case AtomicRMWInst::Sub:
return B.CreateBinOp(Instruction::Sub, LHS, RHS);
case AtomicRMWInst::And:
return B.CreateBinOp(Instruction::And, LHS, RHS);
case AtomicRMWInst::Or:
return B.CreateBinOp(Instruction::Or, LHS, RHS);
case AtomicRMWInst::Xor:
return B.CreateBinOp(Instruction::Xor, LHS, RHS);
case AtomicRMWInst::Max:
Pred = CmpInst::ICMP_SGT;
break;
case AtomicRMWInst::Min:
Pred = CmpInst::ICMP_SLT;
break;
case AtomicRMWInst::UMax:
Pred = CmpInst::ICMP_UGT;
break;
case AtomicRMWInst::UMin:
Pred = CmpInst::ICMP_ULT;
break;
}
Value *Cond = B.CreateICmp(Pred, LHS, RHS);
return B.CreateSelect(Cond, LHS, RHS);
}
// Use the builder to create a reduction of V across the wavefront, with all
// lanes active, returning the same result in all lanes.
Value *AMDGPUAtomicOptimizerImpl::buildReduction(IRBuilder<> &B,
AtomicRMWInst::BinOp Op,
Value *V,
Value *const Identity) const {
Type *const Ty = V->getType();
Module *M = B.GetInsertBlock()->getModule();
Function *UpdateDPP =
Intrinsic::getDeclaration(M, Intrinsic::amdgcn_update_dpp, Ty);
// Reduce within each row of 16 lanes.
for (unsigned Idx = 0; Idx < 4; Idx++) {
V = buildNonAtomicBinOp(
B, Op, V,
B.CreateCall(UpdateDPP,
{Identity, V, B.getInt32(DPP::ROW_XMASK0 | 1 << Idx),
B.getInt32(0xf), B.getInt32(0xf), B.getFalse()}));
}
// Reduce within each pair of rows (i.e. 32 lanes).
assert(ST->hasPermLaneX16());
V = buildNonAtomicBinOp(
B, Op, V,
B.CreateIntrinsic(
Intrinsic::amdgcn_permlanex16, {},
{V, V, B.getInt32(-1), B.getInt32(-1), B.getFalse(), B.getFalse()}));
if (ST->isWave32())
return V;
if (ST->hasPermLane64()) {
// Reduce across the upper and lower 32 lanes.
return buildNonAtomicBinOp(
B, Op, V, B.CreateIntrinsic(Intrinsic::amdgcn_permlane64, {}, V));
}
// Pick an arbitrary lane from 0..31 and an arbitrary lane from 32..63 and
// combine them with a scalar operation.
Function *ReadLane =
Intrinsic::getDeclaration(M, Intrinsic::amdgcn_readlane, {});
Value *const Lane0 = B.CreateCall(ReadLane, {V, B.getInt32(0)});
Value *const Lane32 = B.CreateCall(ReadLane, {V, B.getInt32(32)});
return buildNonAtomicBinOp(B, Op, Lane0, Lane32);
}
// Use the builder to create an inclusive scan of V across the wavefront, with
// all lanes active.
Value *AMDGPUAtomicOptimizerImpl::buildScan(IRBuilder<> &B,
AtomicRMWInst::BinOp Op, Value *V,
Value *const Identity) const {
Type *const Ty = V->getType();
Module *M = B.GetInsertBlock()->getModule();
Function *UpdateDPP =
Intrinsic::getDeclaration(M, Intrinsic::amdgcn_update_dpp, Ty);
for (unsigned Idx = 0; Idx < 4; Idx++) {
V = buildNonAtomicBinOp(
B, Op, V,
B.CreateCall(UpdateDPP,
{Identity, V, B.getInt32(DPP::ROW_SHR0 | 1 << Idx),
B.getInt32(0xf), B.getInt32(0xf), B.getFalse()}));
}
if (ST->hasDPPBroadcasts()) {
// GFX9 has DPP row broadcast operations.
V = buildNonAtomicBinOp(
B, Op, V,
B.CreateCall(UpdateDPP,
{Identity, V, B.getInt32(DPP::BCAST15), B.getInt32(0xa),
B.getInt32(0xf), B.getFalse()}));
V = buildNonAtomicBinOp(
B, Op, V,
B.CreateCall(UpdateDPP,
{Identity, V, B.getInt32(DPP::BCAST31), B.getInt32(0xc),
B.getInt32(0xf), B.getFalse()}));
} else {
// On GFX10 all DPP operations are confined to a single row. To get cross-
// row operations we have to use permlane or readlane.
// Combine lane 15 into lanes 16..31 (and, for wave 64, lane 47 into lanes
// 48..63).
assert(ST->hasPermLaneX16());
Value *const PermX = B.CreateIntrinsic(
Intrinsic::amdgcn_permlanex16, {},
{V, V, B.getInt32(-1), B.getInt32(-1), B.getFalse(), B.getFalse()});
V = buildNonAtomicBinOp(
B, Op, V,
B.CreateCall(UpdateDPP,
{Identity, PermX, B.getInt32(DPP::QUAD_PERM_ID),
B.getInt32(0xa), B.getInt32(0xf), B.getFalse()}));
if (!ST->isWave32()) {
// Combine lane 31 into lanes 32..63.
Value *const Lane31 = B.CreateIntrinsic(Intrinsic::amdgcn_readlane, {},
{V, B.getInt32(31)});
V = buildNonAtomicBinOp(
B, Op, V,
B.CreateCall(UpdateDPP,
{Identity, Lane31, B.getInt32(DPP::QUAD_PERM_ID),
B.getInt32(0xc), B.getInt32(0xf), B.getFalse()}));
}
}
return V;
}
// Use the builder to create a shift right of V across the wavefront, with all
// lanes active, to turn an inclusive scan into an exclusive scan.
Value *AMDGPUAtomicOptimizerImpl::buildShiftRight(IRBuilder<> &B, Value *V,
Value *const Identity) const {
Type *const Ty = V->getType();
Module *M = B.GetInsertBlock()->getModule();
Function *UpdateDPP =
Intrinsic::getDeclaration(M, Intrinsic::amdgcn_update_dpp, Ty);
if (ST->hasDPPWavefrontShifts()) {
// GFX9 has DPP wavefront shift operations.
V = B.CreateCall(UpdateDPP,
{Identity, V, B.getInt32(DPP::WAVE_SHR1), B.getInt32(0xf),
B.getInt32(0xf), B.getFalse()});
} else {
Function *ReadLane =
Intrinsic::getDeclaration(M, Intrinsic::amdgcn_readlane, {});
Function *WriteLane =
Intrinsic::getDeclaration(M, Intrinsic::amdgcn_writelane, {});
// On GFX10 all DPP operations are confined to a single row. To get cross-
// row operations we have to use permlane or readlane.
Value *Old = V;
V = B.CreateCall(UpdateDPP,
{Identity, V, B.getInt32(DPP::ROW_SHR0 + 1),
B.getInt32(0xf), B.getInt32(0xf), B.getFalse()});
// Copy the old lane 15 to the new lane 16.
V = B.CreateCall(WriteLane, {B.CreateCall(ReadLane, {Old, B.getInt32(15)}),
B.getInt32(16), V});
if (!ST->isWave32()) {
// Copy the old lane 31 to the new lane 32.
V = B.CreateCall(
WriteLane,
{B.CreateCall(ReadLane, {Old, B.getInt32(31)}), B.getInt32(32), V});
// Copy the old lane 47 to the new lane 48.
V = B.CreateCall(
WriteLane,
{B.CreateCall(ReadLane, {Old, B.getInt32(47)}), B.getInt32(48), V});
}
}
return V;
}
// Use the builder to create an exclusive scan and compute the final reduced
// value using an iterative approach. This provides an alternative
// implementation to DPP which uses WMM for scan computations. This API iterate
// over active lanes to read, compute and update the value using
// readlane and writelane intrinsics.
std::pair<Value *, Value *> AMDGPUAtomicOptimizerImpl::buildScanIteratively(
IRBuilder<> &B, AtomicRMWInst::BinOp Op, Value *const Identity, Value *V,
Instruction &I, BasicBlock *ComputeLoop, BasicBlock *ComputeEnd) const {
auto *Ty = I.getType();
auto *WaveTy = B.getIntNTy(ST->getWavefrontSize());
auto *EntryBB = I.getParent();
auto NeedResult = !I.use_empty();
auto *Ballot =
B.CreateIntrinsic(Intrinsic::amdgcn_ballot, WaveTy, B.getTrue());
// Start inserting instructions for ComputeLoop block
B.SetInsertPoint(ComputeLoop);
// Phi nodes for Accumulator, Scan results destination, and Active Lanes
auto *Accumulator = B.CreatePHI(Ty, 2, "Accumulator");
Accumulator->addIncoming(Identity, EntryBB);
PHINode *OldValuePhi = nullptr;
if (NeedResult) {
OldValuePhi = B.CreatePHI(Ty, 2, "OldValuePhi");
OldValuePhi->addIncoming(PoisonValue::get(Ty), EntryBB);
}
auto *ActiveBits = B.CreatePHI(WaveTy, 2, "ActiveBits");
ActiveBits->addIncoming(Ballot, EntryBB);
// Use llvm.cttz instrinsic to find the lowest remaining active lane.
auto *FF1 =
B.CreateIntrinsic(Intrinsic::cttz, WaveTy, {ActiveBits, B.getTrue()});
auto *LaneIdxInt = B.CreateTrunc(FF1, Ty);
// Get the value required for atomic operation
auto *LaneValue =
B.CreateIntrinsic(Intrinsic::amdgcn_readlane, {}, {V, LaneIdxInt});
// Perform writelane if intermediate scan results are required later in the
// kernel computations
Value *OldValue = nullptr;
if (NeedResult) {
OldValue = B.CreateIntrinsic(Intrinsic::amdgcn_writelane, {},
{Accumulator, LaneIdxInt, OldValuePhi});
OldValuePhi->addIncoming(OldValue, ComputeLoop);
}
// Accumulate the results
auto *NewAccumulator = buildNonAtomicBinOp(B, Op, Accumulator, LaneValue);
Accumulator->addIncoming(NewAccumulator, ComputeLoop);
// Set bit to zero of current active lane so that for next iteration llvm.cttz
// return the next active lane
auto *Mask = B.CreateShl(ConstantInt::get(WaveTy, 1), FF1);
auto *InverseMask = B.CreateXor(Mask, ConstantInt::get(WaveTy, -1));
auto *NewActiveBits = B.CreateAnd(ActiveBits, InverseMask);
ActiveBits->addIncoming(NewActiveBits, ComputeLoop);
// Branch out of the loop when all lanes are processed.
auto *IsEnd = B.CreateICmpEQ(NewActiveBits, ConstantInt::get(WaveTy, 0));
B.CreateCondBr(IsEnd, ComputeEnd, ComputeLoop);
B.SetInsertPoint(ComputeEnd);
return {OldValue, NewAccumulator};
}
static APInt getIdentityValueForAtomicOp(AtomicRMWInst::BinOp Op,
unsigned BitWidth) {
switch (Op) {
default:
llvm_unreachable("Unhandled atomic op");
case AtomicRMWInst::Add:
case AtomicRMWInst::Sub:
case AtomicRMWInst::Or:
case AtomicRMWInst::Xor:
case AtomicRMWInst::UMax:
return APInt::getMinValue(BitWidth);
case AtomicRMWInst::And:
case AtomicRMWInst::UMin:
return APInt::getMaxValue(BitWidth);
case AtomicRMWInst::Max:
return APInt::getSignedMinValue(BitWidth);
case AtomicRMWInst::Min:
return APInt::getSignedMaxValue(BitWidth);
}
}
static Value *buildMul(IRBuilder<> &B, Value *LHS, Value *RHS) {
const ConstantInt *CI = dyn_cast<ConstantInt>(LHS);
return (CI && CI->isOne()) ? RHS : B.CreateMul(LHS, RHS);
}
void AMDGPUAtomicOptimizerImpl::optimizeAtomic(Instruction &I,
AtomicRMWInst::BinOp Op,
unsigned ValIdx,
bool ValDivergent) const {
// Start building just before the instruction.
IRBuilder<> B(&I);
// If we are in a pixel shader, because of how we have to mask out helper
// lane invocations, we need to record the entry and exit BB's.
BasicBlock *PixelEntryBB = nullptr;
BasicBlock *PixelExitBB = nullptr;
// If we're optimizing an atomic within a pixel shader, we need to wrap the
// entire atomic operation in a helper-lane check. We do not want any helper
// lanes that are around only for the purposes of derivatives to take part
// in any cross-lane communication, and we use a branch on whether the lane is
// live to do this.
if (IsPixelShader) {
// Record I's original position as the entry block.
PixelEntryBB = I.getParent();
Value *const Cond = B.CreateIntrinsic(Intrinsic::amdgcn_ps_live, {}, {});
Instruction *const NonHelperTerminator =
SplitBlockAndInsertIfThen(Cond, &I, false, nullptr, &DTU, nullptr);
// Record I's new position as the exit block.
PixelExitBB = I.getParent();
I.moveBefore(NonHelperTerminator);
B.SetInsertPoint(&I);
}
Type *const Ty = I.getType();
const unsigned TyBitWidth = DL->getTypeSizeInBits(Ty);
auto *const VecTy = FixedVectorType::get(B.getInt32Ty(), 2);
// This is the value in the atomic operation we need to combine in order to
// reduce the number of atomic operations.
Value *const V = I.getOperand(ValIdx);
// We need to know how many lanes are active within the wavefront, and we do
// this by doing a ballot of active lanes.
Type *const WaveTy = B.getIntNTy(ST->getWavefrontSize());
CallInst *const Ballot =
B.CreateIntrinsic(Intrinsic::amdgcn_ballot, WaveTy, B.getTrue());
// We need to know how many lanes are active within the wavefront that are
// below us. If we counted each lane linearly starting from 0, a lane is
// below us only if its associated index was less than ours. We do this by
// using the mbcnt intrinsic.
Value *Mbcnt;
if (ST->isWave32()) {
Mbcnt = B.CreateIntrinsic(Intrinsic::amdgcn_mbcnt_lo, {},
{Ballot, B.getInt32(0)});
} else {
Value *const BitCast = B.CreateBitCast(Ballot, VecTy);
Value *const ExtractLo = B.CreateExtractElement(BitCast, B.getInt32(0));
Value *const ExtractHi = B.CreateExtractElement(BitCast, B.getInt32(1));
Mbcnt = B.CreateIntrinsic(Intrinsic::amdgcn_mbcnt_lo, {},
{ExtractLo, B.getInt32(0)});
Mbcnt =
B.CreateIntrinsic(Intrinsic::amdgcn_mbcnt_hi, {}, {ExtractHi, Mbcnt});
}
Mbcnt = B.CreateIntCast(Mbcnt, Ty, false);
Value *const Identity = B.getInt(getIdentityValueForAtomicOp(Op, TyBitWidth));
Value *ExclScan = nullptr;
Value *NewV = nullptr;
const bool NeedResult = !I.use_empty();
Function *F = I.getFunction();
LLVMContext &C = F->getContext();
BasicBlock *ComputeLoop = nullptr;
BasicBlock *ComputeEnd = nullptr;
// If we have a divergent value in each lane, we need to combine the value
// using DPP.
if (ValDivergent) {
const AtomicRMWInst::BinOp ScanOp =
Op == AtomicRMWInst::Sub ? AtomicRMWInst::Add : Op;
if (ScanImpl == ScanOptions::DPP) {
// First we need to set all inactive invocations to the identity value, so
// that they can correctly contribute to the final result.
NewV =
B.CreateIntrinsic(Intrinsic::amdgcn_set_inactive, Ty, {V, Identity});
const AtomicRMWInst::BinOp ScanOp =
Op == AtomicRMWInst::Sub ? AtomicRMWInst::Add : Op;
if (!NeedResult && ST->hasPermLaneX16()) {
// On GFX10 the permlanex16 instruction helps us build a reduction
// without too many readlanes and writelanes, which are generally bad
// for performance.
NewV = buildReduction(B, ScanOp, NewV, Identity);
} else {
NewV = buildScan(B, ScanOp, NewV, Identity);
if (NeedResult)
ExclScan = buildShiftRight(B, NewV, Identity);
// Read the value from the last lane, which has accumulated the values
// of each active lane in the wavefront. This will be our new value
// which we will provide to the atomic operation.
Value *const LastLaneIdx = B.getInt32(ST->getWavefrontSize() - 1);
assert(TyBitWidth == 32);
NewV = B.CreateIntrinsic(Intrinsic::amdgcn_readlane, {},
{NewV, LastLaneIdx});
}
// Finally mark the readlanes in the WWM section.
NewV = B.CreateIntrinsic(Intrinsic::amdgcn_strict_wwm, Ty, NewV);
} else {
// Alternative implementation for scan
ComputeLoop = BasicBlock::Create(C, "ComputeLoop", F);
ComputeEnd = BasicBlock::Create(C, "ComputeEnd", F);
std::tie(ExclScan, NewV) = buildScanIteratively(B, ScanOp, Identity, V, I,
ComputeLoop, ComputeEnd);
}
} else {
switch (Op) {
default:
llvm_unreachable("Unhandled atomic op");
case AtomicRMWInst::Add:
case AtomicRMWInst::Sub: {
// The new value we will be contributing to the atomic operation is the
// old value times the number of active lanes.
Value *const Ctpop = B.CreateIntCast(
B.CreateUnaryIntrinsic(Intrinsic::ctpop, Ballot), Ty, false);
NewV = buildMul(B, V, Ctpop);
break;
}
case AtomicRMWInst::And:
case AtomicRMWInst::Or:
case AtomicRMWInst::Max:
case AtomicRMWInst::Min:
case AtomicRMWInst::UMax:
case AtomicRMWInst::UMin:
// These operations with a uniform value are idempotent: doing the atomic
// operation multiple times has the same effect as doing it once.
NewV = V;
break;
case AtomicRMWInst::Xor:
// The new value we will be contributing to the atomic operation is the
// old value times the parity of the number of active lanes.
Value *const Ctpop = B.CreateIntCast(
B.CreateUnaryIntrinsic(Intrinsic::ctpop, Ballot), Ty, false);
NewV = buildMul(B, V, B.CreateAnd(Ctpop, 1));
break;
}
}
// We only want a single lane to enter our new control flow, and we do this
// by checking if there are any active lanes below us. Only one lane will
// have 0 active lanes below us, so that will be the only one to progress.
Value *const Cond = B.CreateICmpEQ(Mbcnt, B.getIntN(TyBitWidth, 0));
// Store I's original basic block before we split the block.
BasicBlock *const EntryBB = I.getParent();
// We need to introduce some new control flow to force a single lane to be
// active. We do this by splitting I's basic block at I, and introducing the
// new block such that:
// entry --> single_lane -\
// \------------------> exit
Instruction *const SingleLaneTerminator =
SplitBlockAndInsertIfThen(Cond, &I, false, nullptr, &DTU, nullptr);
// At this point, we have split the I's block to allow one lane in wavefront
// to update the precomputed reduced value. Also, completed the codegen for
// new control flow i.e. iterative loop which perform reduction and scan using
// ComputeLoop and ComputeEnd.
// For the new control flow, we need to move branch instruction i.e.
// terminator created during SplitBlockAndInsertIfThen from I's block to
// ComputeEnd block. We also need to set up predecessor to next block when
// single lane done updating the final reduced value.
BasicBlock *Predecessor = nullptr;
if (ValDivergent && ScanImpl == ScanOptions::Iterative) {
// Move terminator from I's block to ComputeEnd block.
Instruction *Terminator = EntryBB->getTerminator();
B.SetInsertPoint(ComputeEnd);
Terminator->removeFromParent();
B.Insert(Terminator);
// Branch to ComputeLoop Block unconditionally from the I's block for
// iterative approach.
B.SetInsertPoint(EntryBB);
B.CreateBr(ComputeLoop);
// Update the dominator tree for new control flow.
DTU.applyUpdates(
{{DominatorTree::Insert, EntryBB, ComputeLoop},
{DominatorTree::Insert, ComputeLoop, ComputeEnd},
{DominatorTree::Delete, EntryBB, SingleLaneTerminator->getParent()}});
Predecessor = ComputeEnd;
} else {
Predecessor = EntryBB;
}
// Move the IR builder into single_lane next.
B.SetInsertPoint(SingleLaneTerminator);
// Clone the original atomic operation into single lane, replacing the
// original value with our newly created one.
Instruction *const NewI = I.clone();
B.Insert(NewI);
NewI->setOperand(ValIdx, NewV);
// Move the IR builder into exit next, and start inserting just before the
// original instruction.
B.SetInsertPoint(&I);
if (NeedResult) {
// Create a PHI node to get our new atomic result into the exit block.
PHINode *const PHI = B.CreatePHI(Ty, 2);
PHI->addIncoming(PoisonValue::get(Ty), Predecessor);
PHI->addIncoming(NewI, SingleLaneTerminator->getParent());
// We need to broadcast the value who was the lowest active lane (the first
// lane) to all other lanes in the wavefront. We use an intrinsic for this,
// but have to handle 64-bit broadcasts with two calls to this intrinsic.
Value *BroadcastI = nullptr;
if (TyBitWidth == 64) {
Value *const ExtractLo = B.CreateTrunc(PHI, B.getInt32Ty());
Value *const ExtractHi =
B.CreateTrunc(B.CreateLShr(PHI, 32), B.getInt32Ty());
CallInst *const ReadFirstLaneLo =
B.CreateIntrinsic(Intrinsic::amdgcn_readfirstlane, {}, ExtractLo);
CallInst *const ReadFirstLaneHi =
B.CreateIntrinsic(Intrinsic::amdgcn_readfirstlane, {}, ExtractHi);
Value *const PartialInsert = B.CreateInsertElement(
PoisonValue::get(VecTy), ReadFirstLaneLo, B.getInt32(0));
Value *const Insert =
B.CreateInsertElement(PartialInsert, ReadFirstLaneHi, B.getInt32(1));
BroadcastI = B.CreateBitCast(Insert, Ty);
} else if (TyBitWidth == 32) {
BroadcastI = B.CreateIntrinsic(Intrinsic::amdgcn_readfirstlane, {}, PHI);
} else {
llvm_unreachable("Unhandled atomic bit width");
}
// Now that we have the result of our single atomic operation, we need to
// get our individual lane's slice into the result. We use the lane offset
// we previously calculated combined with the atomic result value we got
// from the first lane, to get our lane's index into the atomic result.
Value *LaneOffset = nullptr;
if (ValDivergent) {
if (ScanImpl == ScanOptions::DPP) {
LaneOffset =
B.CreateIntrinsic(Intrinsic::amdgcn_strict_wwm, Ty, ExclScan);
} else {
LaneOffset = ExclScan;
}
} else {
switch (Op) {
default:
llvm_unreachable("Unhandled atomic op");
case AtomicRMWInst::Add:
case AtomicRMWInst::Sub:
LaneOffset = buildMul(B, V, Mbcnt);
break;
case AtomicRMWInst::And:
case AtomicRMWInst::Or:
case AtomicRMWInst::Max:
case AtomicRMWInst::Min:
case AtomicRMWInst::UMax:
case AtomicRMWInst::UMin:
LaneOffset = B.CreateSelect(Cond, Identity, V);
break;
case AtomicRMWInst::Xor:
LaneOffset = buildMul(B, V, B.CreateAnd(Mbcnt, 1));
break;
}
}
Value *const Result = buildNonAtomicBinOp(B, Op, BroadcastI, LaneOffset);
if (IsPixelShader) {
// Need a final PHI to reconverge to above the helper lane branch mask.
B.SetInsertPoint(PixelExitBB->getFirstNonPHI());
PHINode *const PHI = B.CreatePHI(Ty, 2);
PHI->addIncoming(PoisonValue::get(Ty), PixelEntryBB);
PHI->addIncoming(Result, I.getParent());
I.replaceAllUsesWith(PHI);
} else {
// Replace the original atomic instruction with the new one.
I.replaceAllUsesWith(Result);
}
}
// And delete the original.
I.eraseFromParent();
}
INITIALIZE_PASS_BEGIN(AMDGPUAtomicOptimizer, DEBUG_TYPE,
"AMDGPU atomic optimizations", false, false)
INITIALIZE_PASS_DEPENDENCY(UniformityInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
INITIALIZE_PASS_END(AMDGPUAtomicOptimizer, DEBUG_TYPE,
"AMDGPU atomic optimizations", false, false)
FunctionPass *llvm::createAMDGPUAtomicOptimizerPass(ScanOptions ScanStrategy) {
return new AMDGPUAtomicOptimizer(ScanStrategy);
}