/Users/buildslave/jenkins/workspace/coverage/llvm-project/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp
Line | Count | Source (jump to first uncovered line) |
1 | | //===---- CGOpenMPRuntimeGPU.cpp - Interface to OpenMP GPU Runtimes ----===// |
2 | | // |
3 | | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | | // See https://llvm.org/LICENSE.txt for license information. |
5 | | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | | // |
7 | | //===----------------------------------------------------------------------===// |
8 | | // |
9 | | // This provides a generalized class for OpenMP runtime code generation |
10 | | // specialized by GPU targets NVPTX and AMDGCN. |
11 | | // |
12 | | //===----------------------------------------------------------------------===// |
13 | | |
14 | | #include "CGOpenMPRuntimeGPU.h" |
15 | | #include "CodeGenFunction.h" |
16 | | #include "clang/AST/Attr.h" |
17 | | #include "clang/AST/DeclOpenMP.h" |
18 | | #include "clang/AST/StmtOpenMP.h" |
19 | | #include "clang/AST/StmtVisitor.h" |
20 | | #include "clang/Basic/Cuda.h" |
21 | | #include "llvm/ADT/SmallPtrSet.h" |
22 | | #include "llvm/Frontend/OpenMP/OMPGridValues.h" |
23 | | #include "llvm/Support/MathExtras.h" |
24 | | |
25 | | using namespace clang; |
26 | | using namespace CodeGen; |
27 | | using namespace llvm::omp; |
28 | | |
29 | | namespace { |
30 | | /// Pre(post)-action for different OpenMP constructs specialized for NVPTX. |
31 | | class NVPTXActionTy final : public PrePostActionTy { |
32 | | llvm::FunctionCallee EnterCallee = nullptr; |
33 | | ArrayRef<llvm::Value *> EnterArgs; |
34 | | llvm::FunctionCallee ExitCallee = nullptr; |
35 | | ArrayRef<llvm::Value *> ExitArgs; |
36 | | bool Conditional = false; |
37 | | llvm::BasicBlock *ContBlock = nullptr; |
38 | | |
39 | | public: |
40 | | NVPTXActionTy(llvm::FunctionCallee EnterCallee, |
41 | | ArrayRef<llvm::Value *> EnterArgs, |
42 | | llvm::FunctionCallee ExitCallee, |
43 | | ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false) |
44 | | : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee), |
45 | 28 | ExitArgs(ExitArgs), Conditional(Conditional) {} |
46 | 0 | void Enter(CodeGenFunction &CGF) override { |
47 | 0 | llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs); |
48 | 0 | if (Conditional) { |
49 | 0 | llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes); |
50 | 0 | auto *ThenBlock = CGF.createBasicBlock("omp_if.then"); |
51 | 0 | ContBlock = CGF.createBasicBlock("omp_if.end"); |
52 | | // Generate the branch (If-stmt) |
53 | 0 | CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock); |
54 | 0 | CGF.EmitBlock(ThenBlock); |
55 | 0 | } |
56 | 0 | } |
57 | 0 | void Done(CodeGenFunction &CGF) { |
58 | 0 | // Emit the rest of blocks/branches |
59 | 0 | CGF.EmitBranch(ContBlock); |
60 | 0 | CGF.EmitBlock(ContBlock, true); |
61 | 0 | } |
62 | 28 | void Exit(CodeGenFunction &CGF) override { |
63 | 28 | CGF.EmitRuntimeCall(ExitCallee, ExitArgs); |
64 | 28 | } |
65 | | }; |
66 | | |
67 | | /// A class to track the execution mode when codegening directives within |
68 | | /// a target region. The appropriate mode (SPMD|NON-SPMD) is set on entry |
69 | | /// to the target region and used by containing directives such as 'parallel' |
70 | | /// to emit optimized code. |
71 | | class ExecutionRuntimeModesRAII { |
72 | | private: |
73 | | CGOpenMPRuntimeGPU::ExecutionMode SavedExecMode = |
74 | | CGOpenMPRuntimeGPU::EM_Unknown; |
75 | | CGOpenMPRuntimeGPU::ExecutionMode &ExecMode; |
76 | | bool SavedRuntimeMode = false; |
77 | | bool *RuntimeMode = nullptr; |
78 | | |
79 | | public: |
80 | | /// Constructor for Non-SPMD mode. |
81 | | ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode) |
82 | 166 | : ExecMode(ExecMode) { |
83 | 166 | SavedExecMode = ExecMode; |
84 | 166 | ExecMode = CGOpenMPRuntimeGPU::EM_NonSPMD; |
85 | 166 | } |
86 | | /// Constructor for SPMD mode. |
87 | | ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode, |
88 | | bool &RuntimeMode, bool FullRuntimeMode) |
89 | 547 | : ExecMode(ExecMode), RuntimeMode(&RuntimeMode) { |
90 | 547 | SavedExecMode = ExecMode; |
91 | 547 | SavedRuntimeMode = RuntimeMode; |
92 | 547 | ExecMode = CGOpenMPRuntimeGPU::EM_SPMD; |
93 | 547 | RuntimeMode = FullRuntimeMode; |
94 | 547 | } |
95 | 713 | ~ExecutionRuntimeModesRAII() { |
96 | 713 | ExecMode = SavedExecMode; |
97 | 713 | if (RuntimeMode) |
98 | 547 | *RuntimeMode = SavedRuntimeMode; |
99 | 713 | } |
100 | | }; |
101 | | |
102 | | /// GPU Configuration: This information can be derived from cuda registers, |
103 | | /// however, providing compile time constants helps generate more efficient |
104 | | /// code. For all practical purposes this is fine because the configuration |
105 | | /// is the same for all known NVPTX architectures. |
106 | | enum MachineConfiguration : unsigned { |
107 | | /// See "llvm/Frontend/OpenMP/OMPGridValues.h" for various related target |
108 | | /// specific Grid Values like GV_Warp_Size, GV_Slot_Size |
109 | | |
110 | | /// Global memory alignment for performance. |
111 | | GlobalMemoryAlignment = 128, |
112 | | |
113 | | /// Maximal size of the shared memory buffer. |
114 | | SharedMemorySize = 128, |
115 | | }; |
116 | | |
117 | 31 | static const ValueDecl *getPrivateItem(const Expr *RefExpr) { |
118 | 31 | RefExpr = RefExpr->IgnoreParens(); |
119 | 31 | if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) { |
120 | 0 | const Expr *Base = ASE->getBase()->IgnoreParenImpCasts(); |
121 | 0 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) |
122 | 0 | Base = TempASE->getBase()->IgnoreParenImpCasts(); |
123 | 0 | RefExpr = Base; |
124 | 31 | } else if (auto *OASE = dyn_cast<OMPArraySectionExpr>(RefExpr)) { |
125 | 0 | const Expr *Base = OASE->getBase()->IgnoreParenImpCasts(); |
126 | 0 | while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base)) |
127 | 0 | Base = TempOASE->getBase()->IgnoreParenImpCasts(); |
128 | 0 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) |
129 | 0 | Base = TempASE->getBase()->IgnoreParenImpCasts(); |
130 | 0 | RefExpr = Base; |
131 | 0 | } |
132 | 31 | RefExpr = RefExpr->IgnoreParenImpCasts(); |
133 | 31 | if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr)) |
134 | 31 | return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl()); |
135 | 0 | const auto *ME = cast<MemberExpr>(RefExpr); |
136 | 0 | return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl()); |
137 | 31 | } |
138 | | |
139 | | |
140 | | static RecordDecl *buildRecordForGlobalizedVars( |
141 | | ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls, |
142 | | ArrayRef<const ValueDecl *> EscapedDeclsForTeams, |
143 | | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
144 | 776 | &MappedDeclsFields, int BufSize) { |
145 | 776 | using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>; |
146 | 776 | if (EscapedDecls.empty() && EscapedDeclsForTeams.empty()752 ) |
147 | 696 | return nullptr; |
148 | 80 | SmallVector<VarsDataTy, 4> GlobalizedVars; |
149 | 80 | for (const ValueDecl *D : EscapedDecls) |
150 | 30 | GlobalizedVars.emplace_back( |
151 | 30 | CharUnits::fromQuantity(std::max( |
152 | 30 | C.getDeclAlign(D).getQuantity(), |
153 | 30 | static_cast<CharUnits::QuantityType>(GlobalMemoryAlignment))), |
154 | 30 | D); |
155 | 80 | for (const ValueDecl *D : EscapedDeclsForTeams) |
156 | 78 | GlobalizedVars.emplace_back(C.getDeclAlign(D), D); |
157 | 80 | llvm::stable_sort(GlobalizedVars, [](VarsDataTy L, VarsDataTy R) { |
158 | 31 | return L.first > R.first; |
159 | 31 | }); |
160 | | |
161 | | // Build struct _globalized_locals_ty { |
162 | | // /* globalized vars */[WarSize] align (max(decl_align, |
163 | | // GlobalMemoryAlignment)) |
164 | | // /* globalized vars */ for EscapedDeclsForTeams |
165 | | // }; |
166 | 80 | RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty"); |
167 | 80 | GlobalizedRD->startDefinition(); |
168 | 80 | llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped( |
169 | 80 | EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end()); |
170 | 108 | for (const auto &Pair : GlobalizedVars) { |
171 | 108 | const ValueDecl *VD = Pair.second; |
172 | 108 | QualType Type = VD->getType(); |
173 | 108 | if (Type->isLValueReferenceType()) |
174 | 3 | Type = C.getPointerType(Type.getNonReferenceType()); |
175 | 105 | else |
176 | 105 | Type = Type.getNonReferenceType(); |
177 | 108 | SourceLocation Loc = VD->getLocation(); |
178 | 108 | FieldDecl *Field; |
179 | 108 | if (SingleEscaped.count(VD)) { |
180 | 78 | Field = FieldDecl::Create( |
181 | 78 | C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, |
182 | 78 | C.getTrivialTypeSourceInfo(Type, SourceLocation()), |
183 | 78 | /*BW=*/nullptr, /*Mutable=*/false, |
184 | 78 | /*InitStyle=*/ICIS_NoInit); |
185 | 78 | Field->setAccess(AS_public); |
186 | 78 | if (VD->hasAttrs()) { |
187 | 9 | for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()), |
188 | 9 | E(VD->getAttrs().end()); |
189 | 9 | I != E; ++I0 ) |
190 | 0 | Field->addAttr(*I); |
191 | 9 | } |
192 | 78 | } else { |
193 | 30 | llvm::APInt ArraySize(32, BufSize); |
194 | 30 | Type = C.getConstantArrayType(Type, ArraySize, nullptr, ArrayType::Normal, |
195 | 30 | 0); |
196 | 30 | Field = FieldDecl::Create( |
197 | 30 | C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, |
198 | 30 | C.getTrivialTypeSourceInfo(Type, SourceLocation()), |
199 | 30 | /*BW=*/nullptr, /*Mutable=*/false, |
200 | 30 | /*InitStyle=*/ICIS_NoInit); |
201 | 30 | Field->setAccess(AS_public); |
202 | 30 | llvm::APInt Align(32, std::max(C.getDeclAlign(VD).getQuantity(), |
203 | 30 | static_cast<CharUnits::QuantityType>( |
204 | 30 | GlobalMemoryAlignment))); |
205 | 30 | Field->addAttr(AlignedAttr::CreateImplicit( |
206 | 30 | C, /*IsAlignmentExpr=*/true, |
207 | 30 | IntegerLiteral::Create(C, Align, |
208 | 30 | C.getIntTypeForBitwidth(32, /*Signed=*/0), |
209 | 30 | SourceLocation()), |
210 | 30 | {}, AttributeCommonInfo::AS_GNU, AlignedAttr::GNU_aligned)); |
211 | 30 | } |
212 | 108 | GlobalizedRD->addDecl(Field); |
213 | 108 | MappedDeclsFields.try_emplace(VD, Field); |
214 | 108 | } |
215 | 80 | GlobalizedRD->completeDefinition(); |
216 | 80 | return GlobalizedRD; |
217 | 776 | } |
218 | | |
219 | | /// Get the list of variables that can escape their declaration context. |
220 | | class CheckVarsEscapingDeclContext final |
221 | | : public ConstStmtVisitor<CheckVarsEscapingDeclContext> { |
222 | | CodeGenFunction &CGF; |
223 | | llvm::SetVector<const ValueDecl *> EscapedDecls; |
224 | | llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls; |
225 | | llvm::SmallPtrSet<const Decl *, 4> EscapedParameters; |
226 | | RecordDecl *GlobalizedRD = nullptr; |
227 | | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; |
228 | | bool AllEscaped = false; |
229 | | bool IsForCombinedParallelRegion = false; |
230 | | |
231 | 292 | void markAsEscaped(const ValueDecl *VD) { |
232 | | // Do not globalize declare target variables. |
233 | 292 | if (!isa<VarDecl>(VD) || |
234 | 292 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD)274 ) |
235 | 18 | return; |
236 | 274 | VD = cast<ValueDecl>(VD->getCanonicalDecl()); |
237 | | // Use user-specified allocation. |
238 | 274 | if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>()0 ) |
239 | 0 | return; |
240 | | // Variables captured by value must be globalized. |
241 | 274 | if (auto *CSI = CGF.CapturedStmtInfo) { |
242 | 164 | if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) { |
243 | | // Check if need to capture the variable that was already captured by |
244 | | // value in the outer region. |
245 | 131 | if (!IsForCombinedParallelRegion) { |
246 | 131 | if (!FD->hasAttrs()) |
247 | 1 | return; |
248 | 130 | const auto *Attr = FD->getAttr<OMPCaptureKindAttr>(); |
249 | 130 | if (!Attr) |
250 | 0 | return; |
251 | 130 | if (((Attr->getCaptureKind() != OMPC_map) && |
252 | 130 | !isOpenMPPrivate(Attr->getCaptureKind())41 ) || |
253 | 130 | ((Attr->getCaptureKind() == OMPC_map) && |
254 | 130 | !FD->getType()->isAnyPointerType()89 )) |
255 | 89 | return; |
256 | 130 | } |
257 | 41 | if (!FD->getType()->isReferenceType()) { |
258 | 17 | assert(!VD->getType()->isVariablyModifiedType() && |
259 | 17 | "Parameter captured by value with variably modified type"); |
260 | 0 | EscapedParameters.insert(VD); |
261 | 24 | } else if (!IsForCombinedParallelRegion) { |
262 | 24 | return; |
263 | 24 | } |
264 | 41 | } |
265 | 164 | } |
266 | 160 | if ((!CGF.CapturedStmtInfo || |
267 | 160 | (50 IsForCombinedParallelRegion50 && CGF.CapturedStmtInfo0 )) && |
268 | 160 | VD->getType()->isReferenceType()110 ) |
269 | | // Do not globalize variables with reference type. |
270 | 84 | return; |
271 | 76 | if (VD->getType()->isVariablyModifiedType()) |
272 | 0 | EscapedVariableLengthDecls.insert(VD); |
273 | 76 | else |
274 | 76 | EscapedDecls.insert(VD); |
275 | 76 | } |
276 | | |
277 | 221 | void VisitValueDecl(const ValueDecl *VD) { |
278 | 221 | if (VD->getType()->isLValueReferenceType()) |
279 | 3 | markAsEscaped(VD); |
280 | 221 | if (const auto *VarD = dyn_cast<VarDecl>(VD)) { |
281 | 221 | if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) { |
282 | 202 | const bool SavedAllEscaped = AllEscaped; |
283 | 202 | AllEscaped = VD->getType()->isLValueReferenceType(); |
284 | 202 | Visit(VarD->getInit()); |
285 | 202 | AllEscaped = SavedAllEscaped; |
286 | 202 | } |
287 | 221 | } |
288 | 221 | } |
289 | | void VisitOpenMPCapturedStmt(const CapturedStmt *S, |
290 | | ArrayRef<OMPClause *> Clauses, |
291 | 65 | bool IsCombinedParallelRegion) { |
292 | 65 | if (!S) |
293 | 0 | return; |
294 | 65 | for (const CapturedStmt::Capture &C : S->captures()) { |
295 | 57 | if (C.capturesVariable() && !C.capturesVariableByCopy()) { |
296 | 57 | const ValueDecl *VD = C.getCapturedVar(); |
297 | 57 | bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion; |
298 | 57 | if (IsCombinedParallelRegion) { |
299 | | // Check if the variable is privatized in the combined construct and |
300 | | // those private copies must be shared in the inner parallel |
301 | | // directive. |
302 | 0 | IsForCombinedParallelRegion = false; |
303 | 0 | for (const OMPClause *C : Clauses) { |
304 | 0 | if (!isOpenMPPrivate(C->getClauseKind()) || |
305 | 0 | C->getClauseKind() == OMPC_reduction || |
306 | 0 | C->getClauseKind() == OMPC_linear || |
307 | 0 | C->getClauseKind() == OMPC_private) |
308 | 0 | continue; |
309 | 0 | ArrayRef<const Expr *> Vars; |
310 | 0 | if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C)) |
311 | 0 | Vars = PC->getVarRefs(); |
312 | 0 | else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C)) |
313 | 0 | Vars = PC->getVarRefs(); |
314 | 0 | else |
315 | 0 | llvm_unreachable("Unexpected clause."); |
316 | 0 | for (const auto *E : Vars) { |
317 | 0 | const Decl *D = |
318 | 0 | cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl(); |
319 | 0 | if (D == VD->getCanonicalDecl()) { |
320 | 0 | IsForCombinedParallelRegion = true; |
321 | 0 | break; |
322 | 0 | } |
323 | 0 | } |
324 | 0 | if (IsForCombinedParallelRegion) |
325 | 0 | break; |
326 | 0 | } |
327 | 0 | } |
328 | 57 | markAsEscaped(VD); |
329 | 57 | if (isa<OMPCapturedExprDecl>(VD)) |
330 | 0 | VisitValueDecl(VD); |
331 | 57 | IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion; |
332 | 57 | } |
333 | 57 | } |
334 | 65 | } |
335 | | |
336 | 745 | void buildRecordForGlobalizedVars(bool IsInTTDRegion) { |
337 | 745 | assert(!GlobalizedRD && |
338 | 745 | "Record for globalized variables is built already."); |
339 | 0 | ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams; |
340 | 745 | unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; |
341 | 745 | if (IsInTTDRegion) |
342 | 225 | EscapedDeclsForTeams = EscapedDecls.getArrayRef(); |
343 | 520 | else |
344 | 520 | EscapedDeclsForParallel = EscapedDecls.getArrayRef(); |
345 | 745 | GlobalizedRD = ::buildRecordForGlobalizedVars( |
346 | 745 | CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams, |
347 | 745 | MappedDeclsFields, WarpSize); |
348 | 745 | } |
349 | | |
350 | | public: |
351 | | CheckVarsEscapingDeclContext(CodeGenFunction &CGF, |
352 | | ArrayRef<const ValueDecl *> TeamsReductions) |
353 | 758 | : CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) { |
354 | 758 | } |
355 | 758 | virtual ~CheckVarsEscapingDeclContext() = default; |
356 | 226 | void VisitDeclStmt(const DeclStmt *S) { |
357 | 226 | if (!S) |
358 | 0 | return; |
359 | 226 | for (const Decl *D : S->decls()) |
360 | 226 | if (const auto *VD = dyn_cast_or_null<ValueDecl>(D)) |
361 | 221 | VisitValueDecl(VD); |
362 | 226 | } |
363 | 71 | void VisitOMPExecutableDirective(const OMPExecutableDirective *D) { |
364 | 71 | if (!D) |
365 | 0 | return; |
366 | 71 | if (!D->hasAssociatedStmt()) |
367 | 2 | return; |
368 | 69 | if (const auto *S = |
369 | 69 | dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) { |
370 | | // Do not analyze directives that do not actually require capturing, |
371 | | // like `omp for` or `omp simd` directives. |
372 | 67 | llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; |
373 | 67 | getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind()); |
374 | 67 | if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) { |
375 | 2 | VisitStmt(S->getCapturedStmt()); |
376 | 2 | return; |
377 | 2 | } |
378 | 65 | VisitOpenMPCapturedStmt( |
379 | 65 | S, D->clauses(), |
380 | 65 | CaptureRegions.back() == OMPD_parallel && |
381 | 65 | isOpenMPDistributeDirective(D->getDirectiveKind())51 ); |
382 | 65 | } |
383 | 69 | } |
384 | 26 | void VisitCapturedStmt(const CapturedStmt *S) { |
385 | 26 | if (!S) |
386 | 0 | return; |
387 | 26 | for (const CapturedStmt::Capture &C : S->captures()) { |
388 | 7 | if (C.capturesVariable() && !C.capturesVariableByCopy()1 ) { |
389 | 1 | const ValueDecl *VD = C.getCapturedVar(); |
390 | 1 | markAsEscaped(VD); |
391 | 1 | if (isa<OMPCapturedExprDecl>(VD)) |
392 | 0 | VisitValueDecl(VD); |
393 | 1 | } |
394 | 7 | } |
395 | 26 | } |
396 | 1 | void VisitLambdaExpr(const LambdaExpr *E) { |
397 | 1 | if (!E) |
398 | 0 | return; |
399 | 1 | for (const LambdaCapture &C : E->captures()) { |
400 | 0 | if (C.capturesVariable()) { |
401 | 0 | if (C.getCaptureKind() == LCK_ByRef) { |
402 | 0 | const ValueDecl *VD = C.getCapturedVar(); |
403 | 0 | markAsEscaped(VD); |
404 | 0 | if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD)) |
405 | 0 | VisitValueDecl(VD); |
406 | 0 | } |
407 | 0 | } |
408 | 0 | } |
409 | 1 | } |
410 | 0 | void VisitBlockExpr(const BlockExpr *E) { |
411 | 0 | if (!E) |
412 | 0 | return; |
413 | 0 | for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) { |
414 | 0 | if (C.isByRef()) { |
415 | 0 | const VarDecl *VD = C.getVariable(); |
416 | 0 | markAsEscaped(VD); |
417 | 0 | if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture()) |
418 | 0 | VisitValueDecl(VD); |
419 | 0 | } |
420 | 0 | } |
421 | 0 | } |
422 | 1.85k | void VisitCallExpr(const CallExpr *E) { |
423 | 1.85k | if (!E) |
424 | 0 | return; |
425 | 1.85k | for (const Expr *Arg : E->arguments()) { |
426 | 1.54k | if (!Arg) |
427 | 0 | continue; |
428 | 1.54k | if (Arg->isLValue()) { |
429 | 238 | const bool SavedAllEscaped = AllEscaped; |
430 | 238 | AllEscaped = true; |
431 | 238 | Visit(Arg); |
432 | 238 | AllEscaped = SavedAllEscaped; |
433 | 1.31k | } else { |
434 | 1.31k | Visit(Arg); |
435 | 1.31k | } |
436 | 1.54k | } |
437 | 1.85k | Visit(E->getCallee()); |
438 | 1.85k | } |
439 | 3.84k | void VisitDeclRefExpr(const DeclRefExpr *E) { |
440 | 3.84k | if (!E) |
441 | 0 | return; |
442 | 3.84k | const ValueDecl *VD = E->getDecl(); |
443 | 3.84k | if (AllEscaped) |
444 | 231 | markAsEscaped(VD); |
445 | 3.84k | if (isa<OMPCapturedExprDecl>(VD)) |
446 | 0 | VisitValueDecl(VD); |
447 | 3.84k | else if (const auto *VarD = dyn_cast<VarDecl>(VD)) |
448 | 2.43k | if (VarD->isInitCapture()) |
449 | 0 | VisitValueDecl(VD); |
450 | 3.84k | } |
451 | 244 | void VisitUnaryOperator(const UnaryOperator *E) { |
452 | 244 | if (!E) |
453 | 0 | return; |
454 | 244 | if (E->getOpcode() == UO_AddrOf) { |
455 | 1 | const bool SavedAllEscaped = AllEscaped; |
456 | 1 | AllEscaped = true; |
457 | 1 | Visit(E->getSubExpr()); |
458 | 1 | AllEscaped = SavedAllEscaped; |
459 | 243 | } else { |
460 | 243 | Visit(E->getSubExpr()); |
461 | 243 | } |
462 | 244 | } |
463 | 3.52k | void VisitImplicitCastExpr(const ImplicitCastExpr *E) { |
464 | 3.52k | if (!E) |
465 | 0 | return; |
466 | 3.52k | if (E->getCastKind() == CK_ArrayToPointerDecay) { |
467 | 98 | const bool SavedAllEscaped = AllEscaped; |
468 | 98 | AllEscaped = true; |
469 | 98 | Visit(E->getSubExpr()); |
470 | 98 | AllEscaped = SavedAllEscaped; |
471 | 3.42k | } else { |
472 | 3.42k | Visit(E->getSubExpr()); |
473 | 3.42k | } |
474 | 3.52k | } |
475 | 4.32k | void VisitExpr(const Expr *E) { |
476 | 4.32k | if (!E) |
477 | 0 | return; |
478 | 4.32k | bool SavedAllEscaped = AllEscaped; |
479 | 4.32k | if (!E->isLValue()) |
480 | 3.41k | AllEscaped = false; |
481 | 4.32k | for (const Stmt *Child : E->children()) |
482 | 5.13k | if (Child) |
483 | 5.13k | Visit(Child); |
484 | 4.32k | AllEscaped = SavedAllEscaped; |
485 | 4.32k | } |
486 | 1.47k | void VisitStmt(const Stmt *S) { |
487 | 1.47k | if (!S) |
488 | 0 | return; |
489 | 1.47k | for (const Stmt *Child : S->children()) |
490 | 2.34k | if (Child) |
491 | 2.32k | Visit(Child); |
492 | 1.47k | } |
493 | | |
494 | | /// Returns the record that handles all the escaped local variables and used |
495 | | /// instead of their original storage. |
496 | 745 | const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) { |
497 | 745 | if (!GlobalizedRD) |
498 | 745 | buildRecordForGlobalizedVars(IsInTTDRegion); |
499 | 745 | return GlobalizedRD; |
500 | 745 | } |
501 | | |
502 | | /// Returns the field in the globalized record for the escaped variable. |
503 | 0 | const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const { |
504 | 0 | assert(GlobalizedRD && |
505 | 0 | "Record for globalized variables must be generated already."); |
506 | 0 | auto I = MappedDeclsFields.find(VD); |
507 | 0 | if (I == MappedDeclsFields.end()) |
508 | 0 | return nullptr; |
509 | 0 | return I->getSecond(); |
510 | 0 | } |
511 | | |
512 | | /// Returns the list of the escaped local variables/parameters. |
513 | 62 | ArrayRef<const ValueDecl *> getEscapedDecls() const { |
514 | 62 | return EscapedDecls.getArrayRef(); |
515 | 62 | } |
516 | | |
517 | | /// Checks if the escaped local variable is actually a parameter passed by |
518 | | /// value. |
519 | 98 | const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const { |
520 | 98 | return EscapedParameters; |
521 | 98 | } |
522 | | |
523 | | /// Returns the list of the escaped variables with the variably modified |
524 | | /// types. |
525 | 745 | ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const { |
526 | 745 | return EscapedVariableLengthDecls.getArrayRef(); |
527 | 745 | } |
528 | | }; |
529 | | } // anonymous namespace |
530 | | |
531 | | /// Get the id of the warp in the block. |
532 | | /// We assume that the warp size is 32, which is always the case |
533 | | /// on the NVPTX device, to generate more efficient code. |
534 | 28 | static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) { |
535 | 28 | CGBuilderTy &Bld = CGF.Builder; |
536 | 28 | unsigned LaneIDBits = |
537 | 28 | llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size); |
538 | 28 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
539 | 28 | return Bld.CreateAShr(RT.getGPUThreadID(CGF), LaneIDBits, "nvptx_warp_id"); |
540 | 28 | } |
541 | | |
542 | | /// Get the id of the current lane in the Warp. |
543 | | /// We assume that the warp size is 32, which is always the case |
544 | | /// on the NVPTX device, to generate more efficient code. |
545 | 28 | static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) { |
546 | 28 | CGBuilderTy &Bld = CGF.Builder; |
547 | 28 | unsigned LaneIDBits = |
548 | 28 | llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size); |
549 | 28 | unsigned LaneIDMask = ~0u >> (32u - LaneIDBits); |
550 | 28 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
551 | 28 | return Bld.CreateAnd(RT.getGPUThreadID(CGF), Bld.getInt32(LaneIDMask), |
552 | 28 | "nvptx_lane_id"); |
553 | 28 | } |
554 | | |
555 | | CGOpenMPRuntimeGPU::ExecutionMode |
556 | 5.81k | CGOpenMPRuntimeGPU::getExecutionMode() const { |
557 | 5.81k | return CurrentExecutionMode; |
558 | 5.81k | } |
559 | | |
560 | | static CGOpenMPRuntimeGPU::DataSharingMode |
561 | 12.5k | getDataSharingMode(CodeGenModule &CGM) { |
562 | 12.5k | return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeGPU::CUDA2.61k |
563 | 12.5k | : CGOpenMPRuntimeGPU::Generic9.94k ; |
564 | 12.5k | } |
565 | | |
566 | | /// Check for inner (nested) SPMD construct, if any |
567 | | static bool hasNestedSPMDDirective(ASTContext &Ctx, |
568 | 616 | const OMPExecutableDirective &D) { |
569 | 616 | const auto *CS = D.getInnermostCapturedStmt(); |
570 | 616 | const auto *Body = |
571 | 616 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); |
572 | 616 | const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
573 | | |
574 | 616 | if (const auto *NestedDir = |
575 | 616 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
576 | 475 | OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); |
577 | 475 | switch (D.getDirectiveKind()) { |
578 | 289 | case OMPD_target: |
579 | 289 | if (isOpenMPParallelDirective(DKind)) |
580 | 180 | return true; |
581 | 109 | if (DKind == OMPD_teams) { |
582 | 107 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( |
583 | 107 | /*IgnoreCaptured=*/true); |
584 | 107 | if (!Body) |
585 | 0 | return false; |
586 | 107 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
587 | 107 | if (const auto *NND = |
588 | 107 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
589 | 93 | DKind = NND->getDirectiveKind(); |
590 | 93 | if (isOpenMPParallelDirective(DKind)) |
591 | 93 | return true; |
592 | 93 | } |
593 | 107 | } |
594 | 16 | return false; |
595 | 186 | case OMPD_target_teams: |
596 | 186 | return isOpenMPParallelDirective(DKind); |
597 | 0 | case OMPD_target_simd: |
598 | 0 | case OMPD_target_parallel: |
599 | 0 | case OMPD_target_parallel_for: |
600 | 0 | case OMPD_target_parallel_for_simd: |
601 | 0 | case OMPD_target_teams_distribute: |
602 | 0 | case OMPD_target_teams_distribute_simd: |
603 | 0 | case OMPD_target_teams_distribute_parallel_for: |
604 | 0 | case OMPD_target_teams_distribute_parallel_for_simd: |
605 | 0 | case OMPD_parallel: |
606 | 0 | case OMPD_for: |
607 | 0 | case OMPD_parallel_for: |
608 | 0 | case OMPD_parallel_master: |
609 | 0 | case OMPD_parallel_sections: |
610 | 0 | case OMPD_for_simd: |
611 | 0 | case OMPD_parallel_for_simd: |
612 | 0 | case OMPD_cancel: |
613 | 0 | case OMPD_cancellation_point: |
614 | 0 | case OMPD_ordered: |
615 | 0 | case OMPD_threadprivate: |
616 | 0 | case OMPD_allocate: |
617 | 0 | case OMPD_task: |
618 | 0 | case OMPD_simd: |
619 | 0 | case OMPD_sections: |
620 | 0 | case OMPD_section: |
621 | 0 | case OMPD_single: |
622 | 0 | case OMPD_master: |
623 | 0 | case OMPD_critical: |
624 | 0 | case OMPD_taskyield: |
625 | 0 | case OMPD_barrier: |
626 | 0 | case OMPD_taskwait: |
627 | 0 | case OMPD_taskgroup: |
628 | 0 | case OMPD_atomic: |
629 | 0 | case OMPD_flush: |
630 | 0 | case OMPD_depobj: |
631 | 0 | case OMPD_scan: |
632 | 0 | case OMPD_teams: |
633 | 0 | case OMPD_target_data: |
634 | 0 | case OMPD_target_exit_data: |
635 | 0 | case OMPD_target_enter_data: |
636 | 0 | case OMPD_distribute: |
637 | 0 | case OMPD_distribute_simd: |
638 | 0 | case OMPD_distribute_parallel_for: |
639 | 0 | case OMPD_distribute_parallel_for_simd: |
640 | 0 | case OMPD_teams_distribute: |
641 | 0 | case OMPD_teams_distribute_simd: |
642 | 0 | case OMPD_teams_distribute_parallel_for: |
643 | 0 | case OMPD_teams_distribute_parallel_for_simd: |
644 | 0 | case OMPD_target_update: |
645 | 0 | case OMPD_declare_simd: |
646 | 0 | case OMPD_declare_variant: |
647 | 0 | case OMPD_begin_declare_variant: |
648 | 0 | case OMPD_end_declare_variant: |
649 | 0 | case OMPD_declare_target: |
650 | 0 | case OMPD_end_declare_target: |
651 | 0 | case OMPD_declare_reduction: |
652 | 0 | case OMPD_declare_mapper: |
653 | 0 | case OMPD_taskloop: |
654 | 0 | case OMPD_taskloop_simd: |
655 | 0 | case OMPD_master_taskloop: |
656 | 0 | case OMPD_master_taskloop_simd: |
657 | 0 | case OMPD_parallel_master_taskloop: |
658 | 0 | case OMPD_parallel_master_taskloop_simd: |
659 | 0 | case OMPD_requires: |
660 | 0 | case OMPD_unknown: |
661 | 0 | default: |
662 | 0 | llvm_unreachable("Unexpected directive."); |
663 | 475 | } |
664 | 475 | } |
665 | | |
666 | 141 | return false; |
667 | 616 | } |
668 | | |
669 | | static bool supportsSPMDExecutionMode(ASTContext &Ctx, |
670 | 1.26k | const OMPExecutableDirective &D) { |
671 | 1.26k | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); |
672 | 1.26k | switch (DirectiveKind) { |
673 | 414 | case OMPD_target: |
674 | 616 | case OMPD_target_teams: |
675 | 616 | return hasNestedSPMDDirective(Ctx, D); |
676 | 261 | case OMPD_target_parallel: |
677 | 330 | case OMPD_target_parallel_for: |
678 | 330 | case OMPD_target_parallel_for_simd: |
679 | 459 | case OMPD_target_teams_distribute_parallel_for: |
680 | 546 | case OMPD_target_teams_distribute_parallel_for_simd: |
681 | 596 | case OMPD_target_simd: |
682 | 644 | case OMPD_target_teams_distribute_simd: |
683 | 644 | return true; |
684 | 9 | case OMPD_target_teams_distribute: |
685 | 9 | return false; |
686 | 0 | case OMPD_parallel: |
687 | 0 | case OMPD_for: |
688 | 0 | case OMPD_parallel_for: |
689 | 0 | case OMPD_parallel_master: |
690 | 0 | case OMPD_parallel_sections: |
691 | 0 | case OMPD_for_simd: |
692 | 0 | case OMPD_parallel_for_simd: |
693 | 0 | case OMPD_cancel: |
694 | 0 | case OMPD_cancellation_point: |
695 | 0 | case OMPD_ordered: |
696 | 0 | case OMPD_threadprivate: |
697 | 0 | case OMPD_allocate: |
698 | 0 | case OMPD_task: |
699 | 0 | case OMPD_simd: |
700 | 0 | case OMPD_sections: |
701 | 0 | case OMPD_section: |
702 | 0 | case OMPD_single: |
703 | 0 | case OMPD_master: |
704 | 0 | case OMPD_critical: |
705 | 0 | case OMPD_taskyield: |
706 | 0 | case OMPD_barrier: |
707 | 0 | case OMPD_taskwait: |
708 | 0 | case OMPD_taskgroup: |
709 | 0 | case OMPD_atomic: |
710 | 0 | case OMPD_flush: |
711 | 0 | case OMPD_depobj: |
712 | 0 | case OMPD_scan: |
713 | 0 | case OMPD_teams: |
714 | 0 | case OMPD_target_data: |
715 | 0 | case OMPD_target_exit_data: |
716 | 0 | case OMPD_target_enter_data: |
717 | 0 | case OMPD_distribute: |
718 | 0 | case OMPD_distribute_simd: |
719 | 0 | case OMPD_distribute_parallel_for: |
720 | 0 | case OMPD_distribute_parallel_for_simd: |
721 | 0 | case OMPD_teams_distribute: |
722 | 0 | case OMPD_teams_distribute_simd: |
723 | 0 | case OMPD_teams_distribute_parallel_for: |
724 | 0 | case OMPD_teams_distribute_parallel_for_simd: |
725 | 0 | case OMPD_target_update: |
726 | 0 | case OMPD_declare_simd: |
727 | 0 | case OMPD_declare_variant: |
728 | 0 | case OMPD_begin_declare_variant: |
729 | 0 | case OMPD_end_declare_variant: |
730 | 0 | case OMPD_declare_target: |
731 | 0 | case OMPD_end_declare_target: |
732 | 0 | case OMPD_declare_reduction: |
733 | 0 | case OMPD_declare_mapper: |
734 | 0 | case OMPD_taskloop: |
735 | 0 | case OMPD_taskloop_simd: |
736 | 0 | case OMPD_master_taskloop: |
737 | 0 | case OMPD_master_taskloop_simd: |
738 | 0 | case OMPD_parallel_master_taskloop: |
739 | 0 | case OMPD_parallel_master_taskloop_simd: |
740 | 0 | case OMPD_requires: |
741 | 0 | case OMPD_unknown: |
742 | 0 | default: |
743 | 0 | break; |
744 | 1.26k | } |
745 | 0 | llvm_unreachable( |
746 | 0 | "Unknown programming model for OpenMP directive on NVPTX target."); |
747 | 0 | } |
748 | | |
749 | | /// Check if the directive is loops based and has schedule clause at all or has |
750 | | /// static scheduling. |
751 | 237 | static bool hasStaticScheduling(const OMPExecutableDirective &D) { |
752 | 237 | assert(isOpenMPWorksharingDirective(D.getDirectiveKind()) && |
753 | 237 | isOpenMPLoopDirective(D.getDirectiveKind()) && |
754 | 237 | "Expected loop-based directive."); |
755 | 237 | return !D.hasClausesOfKind<OMPOrderedClause>() && |
756 | 237 | (234 !D.hasClausesOfKind<OMPScheduleClause>()234 || |
757 | 234 | llvm::any_of(D.getClausesOfKind<OMPScheduleClause>(), |
758 | 172 | [](const OMPScheduleClause *C) { |
759 | 172 | return C->getScheduleKind() == OMPC_SCHEDULE_static; |
760 | 172 | })); |
761 | 237 | } |
762 | | |
763 | | /// Check for inner (nested) lightweight runtime construct, if any |
764 | | static bool hasNestedLightweightDirective(ASTContext &Ctx, |
765 | 198 | const OMPExecutableDirective &D) { |
766 | 198 | assert(supportsSPMDExecutionMode(Ctx, D) && "Expected SPMD mode directive."); |
767 | 0 | const auto *CS = D.getInnermostCapturedStmt(); |
768 | 198 | const auto *Body = |
769 | 198 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); |
770 | 198 | const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
771 | | |
772 | 198 | if (const auto *NestedDir = |
773 | 198 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
774 | 139 | OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); |
775 | 139 | switch (D.getDirectiveKind()) { |
776 | 70 | case OMPD_target: |
777 | 70 | if (isOpenMPParallelDirective(DKind) && |
778 | 70 | isOpenMPWorksharingDirective(DKind)46 && isOpenMPLoopDirective(DKind)21 && |
779 | 70 | hasStaticScheduling(*NestedDir)21 ) |
780 | 9 | return true; |
781 | 61 | if (DKind == OMPD_teams_distribute_simd || DKind == OMPD_simd) |
782 | 0 | return true; |
783 | 61 | if (DKind == OMPD_parallel) { |
784 | 25 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( |
785 | 25 | /*IgnoreCaptured=*/true); |
786 | 25 | if (!Body) |
787 | 0 | return false; |
788 | 25 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
789 | 25 | if (const auto *NND = |
790 | 25 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
791 | 21 | DKind = NND->getDirectiveKind(); |
792 | 21 | if (isOpenMPWorksharingDirective(DKind) && |
793 | 21 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) |
794 | 6 | return true; |
795 | 21 | } |
796 | 36 | } else if (DKind == OMPD_teams) { |
797 | 24 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( |
798 | 24 | /*IgnoreCaptured=*/true); |
799 | 24 | if (!Body) |
800 | 0 | return false; |
801 | 24 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
802 | 24 | if (const auto *NND = |
803 | 24 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
804 | 24 | DKind = NND->getDirectiveKind(); |
805 | 24 | if (isOpenMPParallelDirective(DKind) && |
806 | 24 | isOpenMPWorksharingDirective(DKind) && |
807 | 24 | isOpenMPLoopDirective(DKind)21 && hasStaticScheduling(*NND)21 ) |
808 | 9 | return true; |
809 | 15 | if (DKind == OMPD_parallel) { |
810 | 3 | Body = NND->getInnermostCapturedStmt()->IgnoreContainers( |
811 | 3 | /*IgnoreCaptured=*/true); |
812 | 3 | if (!Body) |
813 | 0 | return false; |
814 | 3 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
815 | 3 | if (const auto *NND = |
816 | 3 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
817 | 0 | DKind = NND->getDirectiveKind(); |
818 | 0 | if (isOpenMPWorksharingDirective(DKind) && |
819 | 0 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) |
820 | 0 | return true; |
821 | 0 | } |
822 | 3 | } |
823 | 15 | } |
824 | 24 | } |
825 | 46 | return false; |
826 | 48 | case OMPD_target_teams: |
827 | 48 | if (isOpenMPParallelDirective(DKind) && |
828 | 48 | isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind)42 && |
829 | 48 | hasStaticScheduling(*NestedDir)42 ) |
830 | 18 | return true; |
831 | 30 | if (DKind == OMPD_distribute_simd || DKind == OMPD_simd) |
832 | 0 | return true; |
833 | 30 | if (DKind == OMPD_parallel) { |
834 | 6 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( |
835 | 6 | /*IgnoreCaptured=*/true); |
836 | 6 | if (!Body) |
837 | 0 | return false; |
838 | 6 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
839 | 6 | if (const auto *NND = |
840 | 6 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
841 | 3 | DKind = NND->getDirectiveKind(); |
842 | 3 | if (isOpenMPWorksharingDirective(DKind) && |
843 | 3 | isOpenMPLoopDirective(DKind)0 && hasStaticScheduling(*NND)0 ) |
844 | 0 | return true; |
845 | 3 | } |
846 | 6 | } |
847 | 30 | return false; |
848 | 21 | case OMPD_target_parallel: |
849 | 21 | if (DKind == OMPD_simd) |
850 | 0 | return true; |
851 | 21 | return isOpenMPWorksharingDirective(DKind) && |
852 | 21 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NestedDir); |
853 | 0 | case OMPD_target_teams_distribute: |
854 | 0 | case OMPD_target_simd: |
855 | 0 | case OMPD_target_parallel_for: |
856 | 0 | case OMPD_target_parallel_for_simd: |
857 | 0 | case OMPD_target_teams_distribute_simd: |
858 | 0 | case OMPD_target_teams_distribute_parallel_for: |
859 | 0 | case OMPD_target_teams_distribute_parallel_for_simd: |
860 | 0 | case OMPD_parallel: |
861 | 0 | case OMPD_for: |
862 | 0 | case OMPD_parallel_for: |
863 | 0 | case OMPD_parallel_master: |
864 | 0 | case OMPD_parallel_sections: |
865 | 0 | case OMPD_for_simd: |
866 | 0 | case OMPD_parallel_for_simd: |
867 | 0 | case OMPD_cancel: |
868 | 0 | case OMPD_cancellation_point: |
869 | 0 | case OMPD_ordered: |
870 | 0 | case OMPD_threadprivate: |
871 | 0 | case OMPD_allocate: |
872 | 0 | case OMPD_task: |
873 | 0 | case OMPD_simd: |
874 | 0 | case OMPD_sections: |
875 | 0 | case OMPD_section: |
876 | 0 | case OMPD_single: |
877 | 0 | case OMPD_master: |
878 | 0 | case OMPD_critical: |
879 | 0 | case OMPD_taskyield: |
880 | 0 | case OMPD_barrier: |
881 | 0 | case OMPD_taskwait: |
882 | 0 | case OMPD_taskgroup: |
883 | 0 | case OMPD_atomic: |
884 | 0 | case OMPD_flush: |
885 | 0 | case OMPD_depobj: |
886 | 0 | case OMPD_scan: |
887 | 0 | case OMPD_teams: |
888 | 0 | case OMPD_target_data: |
889 | 0 | case OMPD_target_exit_data: |
890 | 0 | case OMPD_target_enter_data: |
891 | 0 | case OMPD_distribute: |
892 | 0 | case OMPD_distribute_simd: |
893 | 0 | case OMPD_distribute_parallel_for: |
894 | 0 | case OMPD_distribute_parallel_for_simd: |
895 | 0 | case OMPD_teams_distribute: |
896 | 0 | case OMPD_teams_distribute_simd: |
897 | 0 | case OMPD_teams_distribute_parallel_for: |
898 | 0 | case OMPD_teams_distribute_parallel_for_simd: |
899 | 0 | case OMPD_target_update: |
900 | 0 | case OMPD_declare_simd: |
901 | 0 | case OMPD_declare_variant: |
902 | 0 | case OMPD_begin_declare_variant: |
903 | 0 | case OMPD_end_declare_variant: |
904 | 0 | case OMPD_declare_target: |
905 | 0 | case OMPD_end_declare_target: |
906 | 0 | case OMPD_declare_reduction: |
907 | 0 | case OMPD_declare_mapper: |
908 | 0 | case OMPD_taskloop: |
909 | 0 | case OMPD_taskloop_simd: |
910 | 0 | case OMPD_master_taskloop: |
911 | 0 | case OMPD_master_taskloop_simd: |
912 | 0 | case OMPD_parallel_master_taskloop: |
913 | 0 | case OMPD_parallel_master_taskloop_simd: |
914 | 0 | case OMPD_requires: |
915 | 0 | case OMPD_unknown: |
916 | 0 | default: |
917 | 0 | llvm_unreachable("Unexpected directive."); |
918 | 139 | } |
919 | 139 | } |
920 | | |
921 | 59 | return false; |
922 | 198 | } |
923 | | |
924 | | /// Checks if the construct supports lightweight runtime. It must be SPMD |
925 | | /// construct + inner loop-based construct with static scheduling. |
926 | | static bool supportsLightweightRuntime(ASTContext &Ctx, |
927 | 358 | const OMPExecutableDirective &D) { |
928 | 358 | if (!supportsSPMDExecutionMode(Ctx, D)) |
929 | 0 | return false; |
930 | 358 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); |
931 | 358 | switch (DirectiveKind) { |
932 | 70 | case OMPD_target: |
933 | 118 | case OMPD_target_teams: |
934 | 198 | case OMPD_target_parallel: |
935 | 198 | return hasNestedLightweightDirective(Ctx, D); |
936 | 24 | case OMPD_target_parallel_for: |
937 | 24 | case OMPD_target_parallel_for_simd: |
938 | 78 | case OMPD_target_teams_distribute_parallel_for: |
939 | 111 | case OMPD_target_teams_distribute_parallel_for_simd: |
940 | | // (Last|First)-privates must be shared in parallel region. |
941 | 111 | return hasStaticScheduling(D); |
942 | 25 | case OMPD_target_simd: |
943 | 49 | case OMPD_target_teams_distribute_simd: |
944 | 49 | return true; |
945 | 0 | case OMPD_target_teams_distribute: |
946 | 0 | return false; |
947 | 0 | case OMPD_parallel: |
948 | 0 | case OMPD_for: |
949 | 0 | case OMPD_parallel_for: |
950 | 0 | case OMPD_parallel_master: |
951 | 0 | case OMPD_parallel_sections: |
952 | 0 | case OMPD_for_simd: |
953 | 0 | case OMPD_parallel_for_simd: |
954 | 0 | case OMPD_cancel: |
955 | 0 | case OMPD_cancellation_point: |
956 | 0 | case OMPD_ordered: |
957 | 0 | case OMPD_threadprivate: |
958 | 0 | case OMPD_allocate: |
959 | 0 | case OMPD_task: |
960 | 0 | case OMPD_simd: |
961 | 0 | case OMPD_sections: |
962 | 0 | case OMPD_section: |
963 | 0 | case OMPD_single: |
964 | 0 | case OMPD_master: |
965 | 0 | case OMPD_critical: |
966 | 0 | case OMPD_taskyield: |
967 | 0 | case OMPD_barrier: |
968 | 0 | case OMPD_taskwait: |
969 | 0 | case OMPD_taskgroup: |
970 | 0 | case OMPD_atomic: |
971 | 0 | case OMPD_flush: |
972 | 0 | case OMPD_depobj: |
973 | 0 | case OMPD_scan: |
974 | 0 | case OMPD_teams: |
975 | 0 | case OMPD_target_data: |
976 | 0 | case OMPD_target_exit_data: |
977 | 0 | case OMPD_target_enter_data: |
978 | 0 | case OMPD_distribute: |
979 | 0 | case OMPD_distribute_simd: |
980 | 0 | case OMPD_distribute_parallel_for: |
981 | 0 | case OMPD_distribute_parallel_for_simd: |
982 | 0 | case OMPD_teams_distribute: |
983 | 0 | case OMPD_teams_distribute_simd: |
984 | 0 | case OMPD_teams_distribute_parallel_for: |
985 | 0 | case OMPD_teams_distribute_parallel_for_simd: |
986 | 0 | case OMPD_target_update: |
987 | 0 | case OMPD_declare_simd: |
988 | 0 | case OMPD_declare_variant: |
989 | 0 | case OMPD_begin_declare_variant: |
990 | 0 | case OMPD_end_declare_variant: |
991 | 0 | case OMPD_declare_target: |
992 | 0 | case OMPD_end_declare_target: |
993 | 0 | case OMPD_declare_reduction: |
994 | 0 | case OMPD_declare_mapper: |
995 | 0 | case OMPD_taskloop: |
996 | 0 | case OMPD_taskloop_simd: |
997 | 0 | case OMPD_master_taskloop: |
998 | 0 | case OMPD_master_taskloop_simd: |
999 | 0 | case OMPD_parallel_master_taskloop: |
1000 | 0 | case OMPD_parallel_master_taskloop_simd: |
1001 | 0 | case OMPD_requires: |
1002 | 0 | case OMPD_unknown: |
1003 | 0 | default: |
1004 | 0 | break; |
1005 | 358 | } |
1006 | 0 | llvm_unreachable( |
1007 | 0 | "Unknown programming model for OpenMP directive on NVPTX target."); |
1008 | 0 | } |
1009 | | |
1010 | | void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D, |
1011 | | StringRef ParentName, |
1012 | | llvm::Function *&OutlinedFn, |
1013 | | llvm::Constant *&OutlinedFnID, |
1014 | | bool IsOffloadEntry, |
1015 | 166 | const RegionCodeGenTy &CodeGen) { |
1016 | 166 | ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode); |
1017 | 166 | EntryFunctionState EST; |
1018 | 166 | WrapperFunctionsMap.clear(); |
1019 | | |
1020 | | // Emit target region as a standalone region. |
1021 | 166 | class NVPTXPrePostActionTy : public PrePostActionTy { |
1022 | 166 | CGOpenMPRuntimeGPU::EntryFunctionState &EST; |
1023 | | |
1024 | 166 | public: |
1025 | 166 | NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST) |
1026 | 166 | : EST(EST) {} |
1027 | 166 | void Enter(CodeGenFunction &CGF) override { |
1028 | 166 | auto &RT = |
1029 | 166 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
1030 | 166 | RT.emitKernelInit(CGF, EST, /* IsSPMD */ false); |
1031 | | // Skip target region initialization. |
1032 | 166 | RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); |
1033 | 166 | } |
1034 | 166 | void Exit(CodeGenFunction &CGF) override { |
1035 | 166 | auto &RT = |
1036 | 166 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
1037 | 166 | RT.clearLocThreadIdInsertPt(CGF); |
1038 | 166 | RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ false); |
1039 | 166 | } |
1040 | 166 | } Action(EST); |
1041 | 166 | CodeGen.setAction(Action); |
1042 | 166 | IsInTTDRegion = true; |
1043 | 166 | emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, |
1044 | 166 | IsOffloadEntry, CodeGen); |
1045 | 166 | IsInTTDRegion = false; |
1046 | 166 | } |
1047 | | |
1048 | | void CGOpenMPRuntimeGPU::emitKernelInit(CodeGenFunction &CGF, |
1049 | 713 | EntryFunctionState &EST, bool IsSPMD) { |
1050 | 713 | CGBuilderTy &Bld = CGF.Builder; |
1051 | 713 | Bld.restoreIP(OMPBuilder.createTargetInit(Bld, IsSPMD, requiresFullRuntime())); |
1052 | 713 | IsInTargetMasterThreadRegion = IsSPMD; |
1053 | 713 | if (!IsSPMD) |
1054 | 166 | emitGenericVarsProlog(CGF, EST.Loc); |
1055 | 713 | } |
1056 | | |
1057 | | void CGOpenMPRuntimeGPU::emitKernelDeinit(CodeGenFunction &CGF, |
1058 | | EntryFunctionState &EST, |
1059 | 713 | bool IsSPMD) { |
1060 | 713 | if (!IsSPMD) |
1061 | 166 | emitGenericVarsEpilog(CGF); |
1062 | | |
1063 | 713 | CGBuilderTy &Bld = CGF.Builder; |
1064 | 713 | OMPBuilder.createTargetDeinit(Bld, IsSPMD, requiresFullRuntime()); |
1065 | 713 | } |
1066 | | |
1067 | | void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D, |
1068 | | StringRef ParentName, |
1069 | | llvm::Function *&OutlinedFn, |
1070 | | llvm::Constant *&OutlinedFnID, |
1071 | | bool IsOffloadEntry, |
1072 | 547 | const RegionCodeGenTy &CodeGen) { |
1073 | 547 | ExecutionRuntimeModesRAII ModeRAII( |
1074 | 547 | CurrentExecutionMode, RequiresFullRuntime, |
1075 | 547 | CGM.getLangOpts().OpenMPCUDAForceFullRuntime || |
1076 | 547 | !supportsLightweightRuntime(CGM.getContext(), D)358 ); |
1077 | 547 | EntryFunctionState EST; |
1078 | | |
1079 | | // Emit target region as a standalone region. |
1080 | 547 | class NVPTXPrePostActionTy : public PrePostActionTy { |
1081 | 547 | CGOpenMPRuntimeGPU &RT; |
1082 | 547 | CGOpenMPRuntimeGPU::EntryFunctionState &EST; |
1083 | | |
1084 | 547 | public: |
1085 | 547 | NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT, |
1086 | 547 | CGOpenMPRuntimeGPU::EntryFunctionState &EST) |
1087 | 547 | : RT(RT), EST(EST) {} |
1088 | 547 | void Enter(CodeGenFunction &CGF) override { |
1089 | 547 | RT.emitKernelInit(CGF, EST, /* IsSPMD */ true); |
1090 | | // Skip target region initialization. |
1091 | 547 | RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); |
1092 | 547 | } |
1093 | 547 | void Exit(CodeGenFunction &CGF) override { |
1094 | 547 | RT.clearLocThreadIdInsertPt(CGF); |
1095 | 547 | RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ true); |
1096 | 547 | } |
1097 | 547 | } Action(*this, EST); |
1098 | 547 | CodeGen.setAction(Action); |
1099 | 547 | IsInTTDRegion = true; |
1100 | 547 | emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, |
1101 | 547 | IsOffloadEntry, CodeGen); |
1102 | 547 | IsInTTDRegion = false; |
1103 | 547 | } |
1104 | | |
1105 | | // Create a unique global variable to indicate the execution mode of this target |
1106 | | // region. The execution mode is either 'generic', or 'spmd' depending on the |
1107 | | // target directive. This variable is picked up by the offload library to setup |
1108 | | // the device appropriately before kernel launch. If the execution mode is |
1109 | | // 'generic', the runtime reserves one warp for the master, otherwise, all |
1110 | | // warps participate in parallel work. |
1111 | | static void setPropertyExecutionMode(CodeGenModule &CGM, StringRef Name, |
1112 | 713 | bool Mode) { |
1113 | 713 | auto *GVMode = new llvm::GlobalVariable( |
1114 | 713 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, |
1115 | 713 | llvm::GlobalValue::WeakAnyLinkage, |
1116 | 713 | llvm::ConstantInt::get(CGM.Int8Ty, Mode ? OMP_TGT_EXEC_MODE_SPMD547 |
1117 | 713 | : OMP_TGT_EXEC_MODE_GENERIC166 ), |
1118 | 713 | Twine(Name, "_exec_mode")); |
1119 | 713 | CGM.addCompilerUsedGlobal(GVMode); |
1120 | 713 | } |
1121 | | |
1122 | | void CGOpenMPRuntimeGPU::createOffloadEntry(llvm::Constant *ID, |
1123 | | llvm::Constant *Addr, |
1124 | | uint64_t Size, int32_t, |
1125 | 817 | llvm::GlobalValue::LinkageTypes) { |
1126 | | // TODO: Add support for global variables on the device after declare target |
1127 | | // support. |
1128 | 817 | llvm::Function *Fn = dyn_cast<llvm::Function>(Addr); |
1129 | 817 | if (!Fn) |
1130 | 40 | return; |
1131 | | |
1132 | 777 | llvm::Module &M = CGM.getModule(); |
1133 | 777 | llvm::LLVMContext &Ctx = CGM.getLLVMContext(); |
1134 | | |
1135 | | // Get "nvvm.annotations" metadata node. |
1136 | 777 | llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations"); |
1137 | | |
1138 | 777 | llvm::Metadata *MDVals[] = { |
1139 | 777 | llvm::ConstantAsMetadata::get(Fn), llvm::MDString::get(Ctx, "kernel"), |
1140 | 777 | llvm::ConstantAsMetadata::get( |
1141 | 777 | llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))}; |
1142 | | // Append metadata to nvvm.annotations. |
1143 | 777 | MD->addOperand(llvm::MDNode::get(Ctx, MDVals)); |
1144 | | |
1145 | | // Add a function attribute for the kernel. |
1146 | 777 | Fn->addFnAttr(llvm::Attribute::get(Ctx, "kernel")); |
1147 | 777 | } |
1148 | | |
1149 | | void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction( |
1150 | | const OMPExecutableDirective &D, StringRef ParentName, |
1151 | | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, |
1152 | 713 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { |
1153 | 713 | if (!IsOffloadEntry) // Nothing to do. |
1154 | 0 | return; |
1155 | | |
1156 | 713 | assert(!ParentName.empty() && "Invalid target region parent name!"); |
1157 | | |
1158 | 0 | bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D); |
1159 | 713 | if (Mode) |
1160 | 547 | emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, |
1161 | 547 | CodeGen); |
1162 | 166 | else |
1163 | 166 | emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, |
1164 | 166 | CodeGen); |
1165 | | |
1166 | 713 | setPropertyExecutionMode(CGM, OutlinedFn->getName(), Mode); |
1167 | 713 | } |
1168 | | |
1169 | | namespace { |
1170 | | LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE(); |
1171 | | /// Enum for accesseing the reserved_2 field of the ident_t struct. |
1172 | | enum ModeFlagsTy : unsigned { |
1173 | | /// Bit set to 1 when in SPMD mode. |
1174 | | KMP_IDENT_SPMD_MODE = 0x01, |
1175 | | /// Bit set to 1 when a simplified runtime is used. |
1176 | | KMP_IDENT_SIMPLE_RT_MODE = 0x02, |
1177 | | LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/KMP_IDENT_SIMPLE_RT_MODE) |
1178 | | }; |
1179 | | |
1180 | | /// Special mode Undefined. Is the combination of Non-SPMD mode + SimpleRuntime. |
1181 | | static const ModeFlagsTy UndefinedMode = |
1182 | | (~KMP_IDENT_SPMD_MODE) & KMP_IDENT_SIMPLE_RT_MODE; |
1183 | | } // anonymous namespace |
1184 | | |
1185 | 2.85k | unsigned CGOpenMPRuntimeGPU::getDefaultLocationReserved2Flags() const { |
1186 | 2.85k | switch (getExecutionMode()) { |
1187 | 2.62k | case EM_SPMD: |
1188 | 2.62k | if (requiresFullRuntime()) |
1189 | 1.84k | return KMP_IDENT_SPMD_MODE & (~KMP_IDENT_SIMPLE_RT_MODE); |
1190 | 786 | return KMP_IDENT_SPMD_MODE | KMP_IDENT_SIMPLE_RT_MODE; |
1191 | 201 | case EM_NonSPMD: |
1192 | 201 | assert(requiresFullRuntime() && "Expected full runtime."); |
1193 | 0 | return (~KMP_IDENT_SPMD_MODE) & (~KMP_IDENT_SIMPLE_RT_MODE); |
1194 | 30 | case EM_Unknown: |
1195 | 30 | return UndefinedMode; |
1196 | 2.85k | } |
1197 | 0 | llvm_unreachable("Unknown flags are requested."); |
1198 | 0 | } |
1199 | | |
1200 | | CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM) |
1201 | 178 | : CGOpenMPRuntime(CGM, "_", "$") { |
1202 | 178 | if (!CGM.getLangOpts().OpenMPIsDevice) |
1203 | 0 | llvm_unreachable("OpenMP can only handle device code."); |
1204 | | |
1205 | 178 | llvm::OpenMPIRBuilder &OMPBuilder = getOMPBuilder(); |
1206 | 178 | if (CGM.getLangOpts().NoGPULib || CGM.getLangOpts().OMPHostIRFile.empty()177 ) |
1207 | 4 | return; |
1208 | | |
1209 | 174 | OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTargetDebug, |
1210 | 174 | "__omp_rtl_debug_kind"); |
1211 | 174 | OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTeamSubscription, |
1212 | 174 | "__omp_rtl_assume_teams_oversubscription"); |
1213 | 174 | OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPThreadSubscription, |
1214 | 174 | "__omp_rtl_assume_threads_oversubscription"); |
1215 | 174 | OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoThreadState, |
1216 | 174 | "__omp_rtl_assume_no_thread_state"); |
1217 | 174 | } |
1218 | | |
1219 | | void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF, |
1220 | | ProcBindKind ProcBind, |
1221 | 25 | SourceLocation Loc) { |
1222 | | // Do nothing in case of SPMD mode and L0 parallel. |
1223 | 25 | if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) |
1224 | 25 | return; |
1225 | | |
1226 | 0 | CGOpenMPRuntime::emitProcBindClause(CGF, ProcBind, Loc); |
1227 | 0 | } |
1228 | | |
1229 | | void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF, |
1230 | | llvm::Value *NumThreads, |
1231 | 21 | SourceLocation Loc) { |
1232 | | // Nothing to do. |
1233 | 21 | } |
1234 | | |
1235 | | void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF, |
1236 | | const Expr *NumTeams, |
1237 | | const Expr *ThreadLimit, |
1238 | 17 | SourceLocation Loc) {} |
1239 | | |
1240 | | llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction( |
1241 | | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, |
1242 | 561 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { |
1243 | | // Emit target region as a standalone region. |
1244 | 561 | class NVPTXPrePostActionTy : public PrePostActionTy { |
1245 | 561 | bool &IsInParallelRegion; |
1246 | 561 | bool PrevIsInParallelRegion; |
1247 | | |
1248 | 561 | public: |
1249 | 561 | NVPTXPrePostActionTy(bool &IsInParallelRegion) |
1250 | 561 | : IsInParallelRegion(IsInParallelRegion) {} |
1251 | 561 | void Enter(CodeGenFunction &CGF) override { |
1252 | 561 | PrevIsInParallelRegion = IsInParallelRegion; |
1253 | 561 | IsInParallelRegion = true; |
1254 | 561 | } |
1255 | 561 | void Exit(CodeGenFunction &CGF) override { |
1256 | 561 | IsInParallelRegion = PrevIsInParallelRegion; |
1257 | 561 | } |
1258 | 561 | } Action(IsInParallelRegion); |
1259 | 561 | CodeGen.setAction(Action); |
1260 | 561 | bool PrevIsInTTDRegion = IsInTTDRegion; |
1261 | 561 | IsInTTDRegion = false; |
1262 | 561 | bool PrevIsInTargetMasterThreadRegion = IsInTargetMasterThreadRegion; |
1263 | 561 | IsInTargetMasterThreadRegion = false; |
1264 | 561 | auto *OutlinedFun = |
1265 | 561 | cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction( |
1266 | 561 | D, ThreadIDVar, InnermostKind, CodeGen)); |
1267 | 561 | IsInTargetMasterThreadRegion = PrevIsInTargetMasterThreadRegion; |
1268 | 561 | IsInTTDRegion = PrevIsInTTDRegion; |
1269 | 561 | if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD && |
1270 | 561 | !IsInParallelRegion57 ) { |
1271 | 57 | llvm::Function *WrapperFun = |
1272 | 57 | createParallelDataSharingWrapper(OutlinedFun, D); |
1273 | 57 | WrapperFunctionsMap[OutlinedFun] = WrapperFun; |
1274 | 57 | } |
1275 | | |
1276 | 561 | return OutlinedFun; |
1277 | 561 | } |
1278 | | |
1279 | | /// Get list of lastprivate variables from the teams distribute ... or |
1280 | | /// teams {distribute ...} directives. |
1281 | | static void |
1282 | | getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D, |
1283 | 288 | llvm::SmallVectorImpl<const ValueDecl *> &Vars) { |
1284 | 288 | assert(isOpenMPTeamsDirective(D.getDirectiveKind()) && |
1285 | 288 | "expected teams directive."); |
1286 | 0 | const OMPExecutableDirective *Dir = &D; |
1287 | 288 | if (!isOpenMPDistributeDirective(D.getDirectiveKind())) { |
1288 | 135 | if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild( |
1289 | 135 | Ctx, |
1290 | 135 | D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers( |
1291 | 135 | /*IgnoreCaptured=*/true))) { |
1292 | 135 | Dir = dyn_cast_or_null<OMPExecutableDirective>(S); |
1293 | 135 | if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind())) |
1294 | 9 | Dir = nullptr; |
1295 | 135 | } |
1296 | 135 | } |
1297 | 288 | if (!Dir) |
1298 | 9 | return; |
1299 | 279 | for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) { |
1300 | 22 | for (const Expr *E : C->getVarRefs()) |
1301 | 22 | Vars.push_back(getPrivateItem(E)); |
1302 | 22 | } |
1303 | 279 | } |
1304 | | |
1305 | | /// Get list of reduction variables from the teams ... directives. |
1306 | | static void |
1307 | | getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D, |
1308 | 39 | llvm::SmallVectorImpl<const ValueDecl *> &Vars) { |
1309 | 39 | assert(isOpenMPTeamsDirective(D.getDirectiveKind()) && |
1310 | 39 | "expected teams directive."); |
1311 | 9 | for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { |
1312 | 9 | for (const Expr *E : C->privates()) |
1313 | 9 | Vars.push_back(getPrivateItem(E)); |
1314 | 9 | } |
1315 | 39 | } |
1316 | | |
1317 | | llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction( |
1318 | | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, |
1319 | 327 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { |
1320 | 327 | SourceLocation Loc = D.getBeginLoc(); |
1321 | | |
1322 | 327 | const RecordDecl *GlobalizedRD = nullptr; |
1323 | 327 | llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions; |
1324 | 327 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; |
1325 | 327 | unsigned WarpSize = CGM.getTarget().getGridValue().GV_Warp_Size; |
1326 | | // Globalize team reductions variable unconditionally in all modes. |
1327 | 327 | if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) |
1328 | 39 | getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions); |
1329 | 327 | if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { |
1330 | 288 | getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions); |
1331 | 288 | if (!LastPrivatesReductions.empty()) { |
1332 | 22 | GlobalizedRD = ::buildRecordForGlobalizedVars( |
1333 | 22 | CGM.getContext(), llvm::None, LastPrivatesReductions, |
1334 | 22 | MappedDeclsFields, WarpSize); |
1335 | 22 | } |
1336 | 288 | } else if (39 !LastPrivatesReductions.empty()39 ) { |
1337 | 6 | assert(!TeamAndReductions.first && |
1338 | 6 | "Previous team declaration is not expected."); |
1339 | 0 | TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl(); |
1340 | 6 | std::swap(TeamAndReductions.second, LastPrivatesReductions); |
1341 | 6 | } |
1342 | | |
1343 | | // Emit target region as a standalone region. |
1344 | 0 | class NVPTXPrePostActionTy : public PrePostActionTy { |
1345 | 327 | SourceLocation &Loc; |
1346 | 327 | const RecordDecl *GlobalizedRD; |
1347 | 327 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
1348 | 327 | &MappedDeclsFields; |
1349 | | |
1350 | 327 | public: |
1351 | 327 | NVPTXPrePostActionTy( |
1352 | 327 | SourceLocation &Loc, const RecordDecl *GlobalizedRD, |
1353 | 327 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
1354 | 327 | &MappedDeclsFields) |
1355 | 327 | : Loc(Loc), GlobalizedRD(GlobalizedRD), |
1356 | 327 | MappedDeclsFields(MappedDeclsFields) {} |
1357 | 327 | void Enter(CodeGenFunction &CGF) override { |
1358 | 327 | auto &Rt = |
1359 | 327 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
1360 | 327 | if (GlobalizedRD) { |
1361 | 22 | auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; |
1362 | 22 | I->getSecond().MappedParams = |
1363 | 22 | std::make_unique<CodeGenFunction::OMPMapVars>(); |
1364 | 22 | DeclToAddrMapTy &Data = I->getSecond().LocalVarData; |
1365 | 22 | for (const auto &Pair : MappedDeclsFields) { |
1366 | 22 | assert(Pair.getFirst()->isCanonicalDecl() && |
1367 | 22 | "Expected canonical declaration"); |
1368 | 0 | Data.insert(std::make_pair(Pair.getFirst(), MappedVarData())); |
1369 | 22 | } |
1370 | 22 | } |
1371 | 327 | Rt.emitGenericVarsProlog(CGF, Loc); |
1372 | 327 | } |
1373 | 327 | void Exit(CodeGenFunction &CGF) override { |
1374 | 327 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) |
1375 | 327 | .emitGenericVarsEpilog(CGF); |
1376 | 327 | } |
1377 | 327 | } Action(Loc, GlobalizedRD, MappedDeclsFields); |
1378 | 327 | CodeGen.setAction(Action); |
1379 | 327 | llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction( |
1380 | 327 | D, ThreadIDVar, InnermostKind, CodeGen); |
1381 | | |
1382 | 327 | return OutlinedFun; |
1383 | 327 | } |
1384 | | |
1385 | | void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF, |
1386 | | SourceLocation Loc, |
1387 | 506 | bool WithSPMDCheck) { |
1388 | 506 | if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic && |
1389 | 506 | getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD68 ) |
1390 | 8 | return; |
1391 | | |
1392 | 498 | CGBuilderTy &Bld = CGF.Builder; |
1393 | | |
1394 | 498 | const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); |
1395 | 498 | if (I == FunctionGlobalizedDecls.end()) |
1396 | 429 | return; |
1397 | | |
1398 | 91 | for (auto &Rec : I->getSecond().LocalVarData)69 { |
1399 | 91 | const auto *VD = cast<VarDecl>(Rec.first); |
1400 | 91 | bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first); |
1401 | 91 | QualType VarTy = VD->getType(); |
1402 | | |
1403 | | // Get the local allocation of a firstprivate variable before sharing |
1404 | 91 | llvm::Value *ParValue; |
1405 | 91 | if (EscapedParam) { |
1406 | 17 | LValue ParLVal = |
1407 | 17 | CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType()); |
1408 | 17 | ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc); |
1409 | 17 | } |
1410 | | |
1411 | | // Allocate space for the variable to be globalized |
1412 | 91 | llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())}; |
1413 | 91 | llvm::CallBase *VoidPtr = |
1414 | 91 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
1415 | 91 | CGM.getModule(), OMPRTL___kmpc_alloc_shared), |
1416 | 91 | AllocArgs, VD->getName()); |
1417 | | // FIXME: We should use the variables actual alignment as an argument. |
1418 | 91 | VoidPtr->addRetAttr(llvm::Attribute::get( |
1419 | 91 | CGM.getLLVMContext(), llvm::Attribute::Alignment, |
1420 | 91 | CGM.getContext().getTargetInfo().getNewAlign() / 8)); |
1421 | | |
1422 | | // Cast the void pointer and get the address of the globalized variable. |
1423 | 91 | llvm::PointerType *VarPtrTy = CGF.ConvertTypeForMem(VarTy)->getPointerTo(); |
1424 | 91 | llvm::Value *CastedVoidPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
1425 | 91 | VoidPtr, VarPtrTy, VD->getName() + "_on_stack"); |
1426 | 91 | LValue VarAddr = CGF.MakeNaturalAlignAddrLValue(CastedVoidPtr, VarTy); |
1427 | 91 | Rec.second.PrivateAddr = VarAddr.getAddress(CGF); |
1428 | 91 | Rec.second.GlobalizedVal = VoidPtr; |
1429 | | |
1430 | | // Assign the local allocation to the newly globalized location. |
1431 | 91 | if (EscapedParam) { |
1432 | 17 | CGF.EmitStoreOfScalar(ParValue, VarAddr); |
1433 | 17 | I->getSecond().MappedParams->setVarAddr(CGF, VD, VarAddr.getAddress(CGF)); |
1434 | 17 | } |
1435 | 91 | if (auto *DI = CGF.getDebugInfo()) |
1436 | 0 | VoidPtr->setDebugLoc(DI->SourceLocToDebugLoc(VD->getLocation())); |
1437 | 91 | } |
1438 | 69 | for (const auto *VD : I->getSecond().EscapedVariableLengthDecls) { |
1439 | | // Use actual memory size of the VLA object including the padding |
1440 | | // for alignment purposes. |
1441 | 0 | llvm::Value *Size = CGF.getTypeSize(VD->getType()); |
1442 | 0 | CharUnits Align = CGM.getContext().getDeclAlign(VD); |
1443 | 0 | Size = Bld.CreateNUWAdd( |
1444 | 0 | Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1)); |
1445 | 0 | llvm::Value *AlignVal = |
1446 | 0 | llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity()); |
1447 | |
|
1448 | 0 | Size = Bld.CreateUDiv(Size, AlignVal); |
1449 | 0 | Size = Bld.CreateNUWMul(Size, AlignVal); |
1450 | | |
1451 | | // Allocate space for this VLA object to be globalized. |
1452 | 0 | llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())}; |
1453 | 0 | llvm::CallBase *VoidPtr = |
1454 | 0 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
1455 | 0 | CGM.getModule(), OMPRTL___kmpc_alloc_shared), |
1456 | 0 | AllocArgs, VD->getName()); |
1457 | 0 | VoidPtr->addRetAttr( |
1458 | 0 | llvm::Attribute::get(CGM.getLLVMContext(), llvm::Attribute::Alignment, |
1459 | 0 | CGM.getContext().getTargetInfo().getNewAlign())); |
1460 | |
|
1461 | 0 | I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back( |
1462 | 0 | std::pair<llvm::Value *, llvm::Value *>( |
1463 | 0 | {VoidPtr, CGF.getTypeSize(VD->getType())})); |
1464 | 0 | LValue Base = CGF.MakeAddrLValue(VoidPtr, VD->getType(), |
1465 | 0 | CGM.getContext().getDeclAlign(VD), |
1466 | 0 | AlignmentSource::Decl); |
1467 | 0 | I->getSecond().MappedParams->setVarAddr(CGF, cast<VarDecl>(VD), |
1468 | 0 | Base.getAddress(CGF)); |
1469 | 0 | } |
1470 | 69 | I->getSecond().MappedParams->apply(CGF); |
1471 | 69 | } |
1472 | | |
1473 | | void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF, |
1474 | 506 | bool WithSPMDCheck) { |
1475 | 506 | if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic && |
1476 | 506 | getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD68 ) |
1477 | 8 | return; |
1478 | | |
1479 | 498 | const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); |
1480 | 498 | if (I != FunctionGlobalizedDecls.end()) { |
1481 | | // Deallocate the memory for each globalized VLA object |
1482 | 69 | for (auto AddrSizePair : |
1483 | 69 | llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) { |
1484 | 0 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
1485 | 0 | CGM.getModule(), OMPRTL___kmpc_free_shared), |
1486 | 0 | {AddrSizePair.first, AddrSizePair.second}); |
1487 | 0 | } |
1488 | | // Deallocate the memory for each globalized value |
1489 | 91 | for (auto &Rec : llvm::reverse(I->getSecond().LocalVarData)) { |
1490 | 91 | const auto *VD = cast<VarDecl>(Rec.first); |
1491 | 91 | I->getSecond().MappedParams->restore(CGF); |
1492 | | |
1493 | 91 | llvm::Value *FreeArgs[] = {Rec.second.GlobalizedVal, |
1494 | 91 | CGF.getTypeSize(VD->getType())}; |
1495 | 91 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
1496 | 91 | CGM.getModule(), OMPRTL___kmpc_free_shared), |
1497 | 91 | FreeArgs); |
1498 | 91 | } |
1499 | 69 | } |
1500 | 498 | } |
1501 | | |
1502 | | void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF, |
1503 | | const OMPExecutableDirective &D, |
1504 | | SourceLocation Loc, |
1505 | | llvm::Function *OutlinedFn, |
1506 | 327 | ArrayRef<llvm::Value *> CapturedVars) { |
1507 | 327 | if (!CGF.HaveInsertPoint()) |
1508 | 0 | return; |
1509 | | |
1510 | 327 | Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, |
1511 | 327 | /*Name=*/".zero.addr"); |
1512 | 327 | CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr); |
1513 | 327 | llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; |
1514 | 327 | OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer()); |
1515 | 327 | OutlinedFnArgs.push_back(ZeroAddr.getPointer()); |
1516 | 327 | OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end()); |
1517 | 327 | emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs); |
1518 | 327 | } |
1519 | | |
1520 | | void CGOpenMPRuntimeGPU::emitParallelCall(CodeGenFunction &CGF, |
1521 | | SourceLocation Loc, |
1522 | | llvm::Function *OutlinedFn, |
1523 | | ArrayRef<llvm::Value *> CapturedVars, |
1524 | | const Expr *IfCond, |
1525 | 561 | llvm::Value *NumThreads) { |
1526 | 561 | if (!CGF.HaveInsertPoint()) |
1527 | 0 | return; |
1528 | | |
1529 | 561 | auto &&ParallelGen = [this, Loc, OutlinedFn, CapturedVars, IfCond, |
1530 | 561 | NumThreads](CodeGenFunction &CGF, |
1531 | 561 | PrePostActionTy &Action) { |
1532 | 561 | CGBuilderTy &Bld = CGF.Builder; |
1533 | 561 | llvm::Value *NumThreadsVal = NumThreads; |
1534 | 561 | llvm::Function *WFn = WrapperFunctionsMap[OutlinedFn]; |
1535 | 561 | llvm::Value *ID = llvm::ConstantPointerNull::get(CGM.Int8PtrTy); |
1536 | 561 | if (WFn) |
1537 | 57 | ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy); |
1538 | 561 | llvm::Value *FnPtr = Bld.CreateBitOrPointerCast(OutlinedFn, CGM.Int8PtrTy); |
1539 | | |
1540 | | // Create a private scope that will globalize the arguments |
1541 | | // passed from the outside of the target region. |
1542 | | // TODO: Is that needed? |
1543 | 561 | CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF); |
1544 | | |
1545 | 561 | Address CapturedVarsAddrs = CGF.CreateDefaultAlignTempAlloca( |
1546 | 561 | llvm::ArrayType::get(CGM.VoidPtrTy, CapturedVars.size()), |
1547 | 561 | "captured_vars_addrs"); |
1548 | | // There's something to share. |
1549 | 561 | if (!CapturedVars.empty()) { |
1550 | | // Prepare for parallel region. Indicate the outlined function. |
1551 | 350 | ASTContext &Ctx = CGF.getContext(); |
1552 | 350 | unsigned Idx = 0; |
1553 | 815 | for (llvm::Value *V : CapturedVars) { |
1554 | 815 | Address Dst = Bld.CreateConstArrayGEP(CapturedVarsAddrs, Idx); |
1555 | 815 | llvm::Value *PtrV; |
1556 | 815 | if (V->getType()->isIntegerTy()) |
1557 | 604 | PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy); |
1558 | 211 | else |
1559 | 211 | PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy); |
1560 | 815 | CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false, |
1561 | 815 | Ctx.getPointerType(Ctx.VoidPtrTy)); |
1562 | 815 | ++Idx; |
1563 | 815 | } |
1564 | 350 | } |
1565 | | |
1566 | 561 | llvm::Value *IfCondVal = nullptr; |
1567 | 561 | if (IfCond) |
1568 | 22 | IfCondVal = Bld.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.Int32Ty, |
1569 | 22 | /* isSigned */ false); |
1570 | 539 | else |
1571 | 539 | IfCondVal = llvm::ConstantInt::get(CGF.Int32Ty, 1); |
1572 | | |
1573 | 561 | if (!NumThreadsVal) |
1574 | 540 | NumThreadsVal = llvm::ConstantInt::get(CGF.Int32Ty, -1); |
1575 | 21 | else |
1576 | 21 | NumThreadsVal = Bld.CreateZExtOrTrunc(NumThreadsVal, CGF.Int32Ty), |
1577 | | |
1578 | 21 | assert(IfCondVal && "Expected a value"); |
1579 | 0 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); |
1580 | 561 | llvm::Value *Args[] = { |
1581 | 561 | RTLoc, |
1582 | 561 | getThreadID(CGF, Loc), |
1583 | 561 | IfCondVal, |
1584 | 561 | NumThreadsVal, |
1585 | 561 | llvm::ConstantInt::get(CGF.Int32Ty, -1), |
1586 | 561 | FnPtr, |
1587 | 561 | ID, |
1588 | 561 | Bld.CreateBitOrPointerCast(CapturedVarsAddrs.getPointer(), |
1589 | 561 | CGF.VoidPtrPtrTy), |
1590 | 561 | llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())}; |
1591 | 561 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
1592 | 561 | CGM.getModule(), OMPRTL___kmpc_parallel_51), |
1593 | 561 | Args); |
1594 | 561 | }; |
1595 | | |
1596 | 561 | RegionCodeGenTy RCG(ParallelGen); |
1597 | 561 | RCG(CGF); |
1598 | 561 | } |
1599 | | |
1600 | 0 | void CGOpenMPRuntimeGPU::syncCTAThreads(CodeGenFunction &CGF) { |
1601 | | // Always emit simple barriers! |
1602 | 0 | if (!CGF.HaveInsertPoint()) |
1603 | 0 | return; |
1604 | | // Build call __kmpc_barrier_simple_spmd(nullptr, 0); |
1605 | | // This function does not use parameters, so we can emit just default values. |
1606 | 0 | llvm::Value *Args[] = { |
1607 | 0 | llvm::ConstantPointerNull::get( |
1608 | 0 | cast<llvm::PointerType>(getIdentTyPointerTy())), |
1609 | 0 | llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)}; |
1610 | 0 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
1611 | 0 | CGM.getModule(), OMPRTL___kmpc_barrier_simple_spmd), |
1612 | 0 | Args); |
1613 | 0 | } |
1614 | | |
1615 | | void CGOpenMPRuntimeGPU::emitBarrierCall(CodeGenFunction &CGF, |
1616 | | SourceLocation Loc, |
1617 | | OpenMPDirectiveKind Kind, bool, |
1618 | 173 | bool) { |
1619 | | // Always emit simple barriers! |
1620 | 173 | if (!CGF.HaveInsertPoint()) |
1621 | 0 | return; |
1622 | | // Build call __kmpc_cancel_barrier(loc, thread_id); |
1623 | 173 | unsigned Flags = getDefaultFlagsForBarriers(Kind); |
1624 | 173 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags), |
1625 | 173 | getThreadID(CGF, Loc)}; |
1626 | | |
1627 | 173 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
1628 | 173 | CGM.getModule(), OMPRTL___kmpc_barrier), |
1629 | 173 | Args); |
1630 | 173 | } |
1631 | | |
1632 | | void CGOpenMPRuntimeGPU::emitCriticalRegion( |
1633 | | CodeGenFunction &CGF, StringRef CriticalName, |
1634 | | const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc, |
1635 | 2 | const Expr *Hint) { |
1636 | 2 | llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop"); |
1637 | 2 | llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test"); |
1638 | 2 | llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync"); |
1639 | 2 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body"); |
1640 | 2 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit"); |
1641 | | |
1642 | 2 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
1643 | | |
1644 | | // Get the mask of active threads in the warp. |
1645 | 2 | llvm::Value *Mask = CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
1646 | 2 | CGM.getModule(), OMPRTL___kmpc_warp_active_thread_mask)); |
1647 | | // Fetch team-local id of the thread. |
1648 | 2 | llvm::Value *ThreadID = RT.getGPUThreadID(CGF); |
1649 | | |
1650 | | // Get the width of the team. |
1651 | 2 | llvm::Value *TeamWidth = RT.getGPUNumThreads(CGF); |
1652 | | |
1653 | | // Initialize the counter variable for the loop. |
1654 | 2 | QualType Int32Ty = |
1655 | 2 | CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0); |
1656 | 2 | Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter"); |
1657 | 2 | LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty); |
1658 | 2 | CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal, |
1659 | 2 | /*isInit=*/true); |
1660 | | |
1661 | | // Block checks if loop counter exceeds upper bound. |
1662 | 2 | CGF.EmitBlock(LoopBB); |
1663 | 2 | llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); |
1664 | 2 | llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth); |
1665 | 2 | CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB); |
1666 | | |
1667 | | // Block tests which single thread should execute region, and which threads |
1668 | | // should go straight to synchronisation point. |
1669 | 2 | CGF.EmitBlock(TestBB); |
1670 | 2 | CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); |
1671 | 2 | llvm::Value *CmpThreadToCounter = |
1672 | 2 | CGF.Builder.CreateICmpEQ(ThreadID, CounterVal); |
1673 | 2 | CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB); |
1674 | | |
1675 | | // Block emits the body of the critical region. |
1676 | 2 | CGF.EmitBlock(BodyBB); |
1677 | | |
1678 | | // Output the critical statement. |
1679 | 2 | CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc, |
1680 | 2 | Hint); |
1681 | | |
1682 | | // After the body surrounded by the critical region, the single executing |
1683 | | // thread will jump to the synchronisation point. |
1684 | | // Block waits for all threads in current team to finish then increments the |
1685 | | // counter variable and returns to the loop. |
1686 | 2 | CGF.EmitBlock(SyncBB); |
1687 | | // Reconverge active threads in the warp. |
1688 | 2 | (void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
1689 | 2 | CGM.getModule(), OMPRTL___kmpc_syncwarp), |
1690 | 2 | Mask); |
1691 | | |
1692 | 2 | llvm::Value *IncCounterVal = |
1693 | 2 | CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1)); |
1694 | 2 | CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal); |
1695 | 2 | CGF.EmitBranch(LoopBB); |
1696 | | |
1697 | | // Block that is reached when all threads in the team complete the region. |
1698 | 2 | CGF.EmitBlock(ExitBB, /*IsFinished=*/true); |
1699 | 2 | } |
1700 | | |
1701 | | /// Cast value to the specified type. |
1702 | | static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val, |
1703 | | QualType ValTy, QualType CastTy, |
1704 | 86 | SourceLocation Loc) { |
1705 | 86 | assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && |
1706 | 86 | "Cast type must sized."); |
1707 | 0 | assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && |
1708 | 86 | "Val type must sized."); |
1709 | 0 | llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy); |
1710 | 86 | if (ValTy == CastTy) |
1711 | 56 | return Val; |
1712 | 30 | if (CGF.getContext().getTypeSizeInChars(ValTy) == |
1713 | 30 | CGF.getContext().getTypeSizeInChars(CastTy)) |
1714 | 0 | return CGF.Builder.CreateBitCast(Val, LLVMCastTy); |
1715 | 30 | if (CastTy->isIntegerType() && ValTy->isIntegerType()) |
1716 | 30 | return CGF.Builder.CreateIntCast(Val, LLVMCastTy, |
1717 | 30 | CastTy->hasSignedIntegerRepresentation()); |
1718 | 0 | Address CastItem = CGF.CreateMemTemp(CastTy); |
1719 | 0 | Address ValCastItem = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
1720 | 0 | CastItem, Val->getType()->getPointerTo(CastItem.getAddressSpace()), |
1721 | 0 | Val->getType()); |
1722 | 0 | CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy, |
1723 | 0 | LValueBaseInfo(AlignmentSource::Type), |
1724 | 0 | TBAAAccessInfo()); |
1725 | 0 | return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc, |
1726 | 0 | LValueBaseInfo(AlignmentSource::Type), |
1727 | 0 | TBAAAccessInfo()); |
1728 | 30 | } |
1729 | | |
1730 | | /// This function creates calls to one of two shuffle functions to copy |
1731 | | /// variables between lanes in a warp. |
1732 | | static llvm::Value *createRuntimeShuffleFunction(CodeGenFunction &CGF, |
1733 | | llvm::Value *Elem, |
1734 | | QualType ElemType, |
1735 | | llvm::Value *Offset, |
1736 | 43 | SourceLocation Loc) { |
1737 | 43 | CodeGenModule &CGM = CGF.CGM; |
1738 | 43 | CGBuilderTy &Bld = CGF.Builder; |
1739 | 43 | CGOpenMPRuntimeGPU &RT = |
1740 | 43 | *(static_cast<CGOpenMPRuntimeGPU *>(&CGM.getOpenMPRuntime())); |
1741 | 43 | llvm::OpenMPIRBuilder &OMPBuilder = RT.getOMPBuilder(); |
1742 | | |
1743 | 43 | CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); |
1744 | 43 | assert(Size.getQuantity() <= 8 && |
1745 | 43 | "Unsupported bitwidth in shuffle instruction."); |
1746 | | |
1747 | 43 | RuntimeFunction ShuffleFn = Size.getQuantity() <= 4 |
1748 | 43 | ? OMPRTL___kmpc_shuffle_int3230 |
1749 | 43 | : OMPRTL___kmpc_shuffle_int6413 ; |
1750 | | |
1751 | | // Cast all types to 32- or 64-bit values before calling shuffle routines. |
1752 | 43 | QualType CastTy = CGF.getContext().getIntTypeForBitwidth( |
1753 | 43 | Size.getQuantity() <= 4 ? 3230 : 6413 , /*Signed=*/1); |
1754 | 43 | llvm::Value *ElemCast = castValueToType(CGF, Elem, ElemType, CastTy, Loc); |
1755 | 43 | llvm::Value *WarpSize = |
1756 | 43 | Bld.CreateIntCast(RT.getGPUWarpSize(CGF), CGM.Int16Ty, /*isSigned=*/true); |
1757 | | |
1758 | 43 | llvm::Value *ShuffledVal = CGF.EmitRuntimeCall( |
1759 | 43 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), ShuffleFn), |
1760 | 43 | {ElemCast, Offset, WarpSize}); |
1761 | | |
1762 | 43 | return castValueToType(CGF, ShuffledVal, CastTy, ElemType, Loc); |
1763 | 43 | } |
1764 | | |
1765 | | static void shuffleAndStore(CodeGenFunction &CGF, Address SrcAddr, |
1766 | | Address DestAddr, QualType ElemType, |
1767 | 43 | llvm::Value *Offset, SourceLocation Loc) { |
1768 | 43 | CGBuilderTy &Bld = CGF.Builder; |
1769 | | |
1770 | 43 | CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); |
1771 | | // Create the loop over the big sized data. |
1772 | | // ptr = (void*)Elem; |
1773 | | // ptrEnd = (void*) Elem + 1; |
1774 | | // Step = 8; |
1775 | | // while (ptr + Step < ptrEnd) |
1776 | | // shuffle((int64_t)*ptr); |
1777 | | // Step = 4; |
1778 | | // while (ptr + Step < ptrEnd) |
1779 | | // shuffle((int32_t)*ptr); |
1780 | | // ... |
1781 | 43 | Address ElemPtr = DestAddr; |
1782 | 43 | Address Ptr = SrcAddr; |
1783 | 43 | Address PtrEnd = Bld.CreatePointerBitCastOrAddrSpaceCast( |
1784 | 43 | Bld.CreateConstGEP(SrcAddr, 1), CGF.VoidPtrTy, CGF.Int8Ty); |
1785 | 215 | for (int IntSize = 8; IntSize >= 1; IntSize /= 2172 ) { |
1786 | 172 | if (Size < CharUnits::fromQuantity(IntSize)) |
1787 | 129 | continue; |
1788 | 43 | QualType IntType = CGF.getContext().getIntTypeForBitwidth( |
1789 | 43 | CGF.getContext().toBits(CharUnits::fromQuantity(IntSize)), |
1790 | 43 | /*Signed=*/1); |
1791 | 43 | llvm::Type *IntTy = CGF.ConvertTypeForMem(IntType); |
1792 | 43 | Ptr = Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr, IntTy->getPointerTo(), |
1793 | 43 | IntTy); |
1794 | 43 | ElemPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
1795 | 43 | ElemPtr, IntTy->getPointerTo(), IntTy); |
1796 | 43 | if (Size.getQuantity() / IntSize > 1) { |
1797 | 3 | llvm::BasicBlock *PreCondBB = CGF.createBasicBlock(".shuffle.pre_cond"); |
1798 | 3 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".shuffle.then"); |
1799 | 3 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".shuffle.exit"); |
1800 | 3 | llvm::BasicBlock *CurrentBB = Bld.GetInsertBlock(); |
1801 | 3 | CGF.EmitBlock(PreCondBB); |
1802 | 3 | llvm::PHINode *PhiSrc = |
1803 | 3 | Bld.CreatePHI(Ptr.getType(), /*NumReservedValues=*/2); |
1804 | 3 | PhiSrc->addIncoming(Ptr.getPointer(), CurrentBB); |
1805 | 3 | llvm::PHINode *PhiDest = |
1806 | 3 | Bld.CreatePHI(ElemPtr.getType(), /*NumReservedValues=*/2); |
1807 | 3 | PhiDest->addIncoming(ElemPtr.getPointer(), CurrentBB); |
1808 | 3 | Ptr = Address(PhiSrc, Ptr.getElementType(), Ptr.getAlignment()); |
1809 | 3 | ElemPtr = |
1810 | 3 | Address(PhiDest, ElemPtr.getElementType(), ElemPtr.getAlignment()); |
1811 | 3 | llvm::Value *PtrDiff = Bld.CreatePtrDiff( |
1812 | 3 | CGF.Int8Ty, PtrEnd.getPointer(), |
1813 | 3 | Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr.getPointer(), |
1814 | 3 | CGF.VoidPtrTy)); |
1815 | 3 | Bld.CreateCondBr(Bld.CreateICmpSGT(PtrDiff, Bld.getInt64(IntSize - 1)), |
1816 | 3 | ThenBB, ExitBB); |
1817 | 3 | CGF.EmitBlock(ThenBB); |
1818 | 3 | llvm::Value *Res = createRuntimeShuffleFunction( |
1819 | 3 | CGF, |
1820 | 3 | CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc, |
1821 | 3 | LValueBaseInfo(AlignmentSource::Type), |
1822 | 3 | TBAAAccessInfo()), |
1823 | 3 | IntType, Offset, Loc); |
1824 | 3 | CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType, |
1825 | 3 | LValueBaseInfo(AlignmentSource::Type), |
1826 | 3 | TBAAAccessInfo()); |
1827 | 3 | Address LocalPtr = Bld.CreateConstGEP(Ptr, 1); |
1828 | 3 | Address LocalElemPtr = Bld.CreateConstGEP(ElemPtr, 1); |
1829 | 3 | PhiSrc->addIncoming(LocalPtr.getPointer(), ThenBB); |
1830 | 3 | PhiDest->addIncoming(LocalElemPtr.getPointer(), ThenBB); |
1831 | 3 | CGF.EmitBranch(PreCondBB); |
1832 | 3 | CGF.EmitBlock(ExitBB); |
1833 | 40 | } else { |
1834 | 40 | llvm::Value *Res = createRuntimeShuffleFunction( |
1835 | 40 | CGF, |
1836 | 40 | CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc, |
1837 | 40 | LValueBaseInfo(AlignmentSource::Type), |
1838 | 40 | TBAAAccessInfo()), |
1839 | 40 | IntType, Offset, Loc); |
1840 | 40 | CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType, |
1841 | 40 | LValueBaseInfo(AlignmentSource::Type), |
1842 | 40 | TBAAAccessInfo()); |
1843 | 40 | Ptr = Bld.CreateConstGEP(Ptr, 1); |
1844 | 40 | ElemPtr = Bld.CreateConstGEP(ElemPtr, 1); |
1845 | 40 | } |
1846 | 43 | Size = Size % IntSize; |
1847 | 43 | } |
1848 | 43 | } |
1849 | | |
1850 | | namespace { |
1851 | | enum CopyAction : unsigned { |
1852 | | // RemoteLaneToThread: Copy over a Reduce list from a remote lane in |
1853 | | // the warp using shuffle instructions. |
1854 | | RemoteLaneToThread, |
1855 | | // ThreadCopy: Make a copy of a Reduce list on the thread's stack. |
1856 | | ThreadCopy, |
1857 | | // ThreadToScratchpad: Copy a team-reduced array to the scratchpad. |
1858 | | ThreadToScratchpad, |
1859 | | // ScratchpadToThread: Copy from a scratchpad array in global memory |
1860 | | // containing team-reduced data to a thread's stack. |
1861 | | ScratchpadToThread, |
1862 | | }; |
1863 | | } // namespace |
1864 | | |
1865 | | struct CopyOptionsTy { |
1866 | | llvm::Value *RemoteLaneOffset; |
1867 | | llvm::Value *ScratchpadIndex; |
1868 | | llvm::Value *ScratchpadWidth; |
1869 | | }; |
1870 | | |
1871 | | /// Emit instructions to copy a Reduce list, which contains partially |
1872 | | /// aggregated values, in the specified direction. |
1873 | | static void emitReductionListCopy( |
1874 | | CopyAction Action, CodeGenFunction &CGF, QualType ReductionArrayTy, |
1875 | | ArrayRef<const Expr *> Privates, Address SrcBase, Address DestBase, |
1876 | 56 | CopyOptionsTy CopyOptions = {nullptr, nullptr, nullptr}) { |
1877 | | |
1878 | 56 | CodeGenModule &CGM = CGF.CGM; |
1879 | 56 | ASTContext &C = CGM.getContext(); |
1880 | 56 | CGBuilderTy &Bld = CGF.Builder; |
1881 | | |
1882 | 56 | llvm::Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset; |
1883 | 56 | llvm::Value *ScratchpadIndex = CopyOptions.ScratchpadIndex; |
1884 | 56 | llvm::Value *ScratchpadWidth = CopyOptions.ScratchpadWidth; |
1885 | | |
1886 | | // Iterates, element-by-element, through the source Reduce list and |
1887 | | // make a copy. |
1888 | 56 | unsigned Idx = 0; |
1889 | 56 | unsigned Size = Privates.size(); |
1890 | 86 | for (const Expr *Private : Privates) { |
1891 | 86 | Address SrcElementAddr = Address::invalid(); |
1892 | 86 | Address DestElementAddr = Address::invalid(); |
1893 | 86 | Address DestElementPtrAddr = Address::invalid(); |
1894 | | // Should we shuffle in an element from a remote lane? |
1895 | 86 | bool ShuffleInElement = false; |
1896 | | // Set to true to update the pointer in the dest Reduce list to a |
1897 | | // newly created element. |
1898 | 86 | bool UpdateDestListPtr = false; |
1899 | | // Increment the src or dest pointer to the scratchpad, for each |
1900 | | // new element. |
1901 | 86 | bool IncrScratchpadSrc = false; |
1902 | 86 | bool IncrScratchpadDest = false; |
1903 | 86 | QualType PrivatePtrType = C.getPointerType(Private->getType()); |
1904 | 86 | llvm::Type *PrivateLlvmPtrType = CGF.ConvertType(PrivatePtrType); |
1905 | | |
1906 | 86 | switch (Action) { |
1907 | 43 | case RemoteLaneToThread: { |
1908 | | // Step 1.1: Get the address for the src element in the Reduce list. |
1909 | 43 | Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); |
1910 | 43 | SrcElementAddr = |
1911 | 43 | CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
1912 | 43 | SrcElementPtrAddr, PrivateLlvmPtrType), |
1913 | 43 | PrivatePtrType->castAs<PointerType>()); |
1914 | | |
1915 | | // Step 1.2: Create a temporary to store the element in the destination |
1916 | | // Reduce list. |
1917 | 43 | DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); |
1918 | 43 | DestElementAddr = |
1919 | 43 | CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element"); |
1920 | 43 | ShuffleInElement = true; |
1921 | 43 | UpdateDestListPtr = true; |
1922 | 43 | break; |
1923 | 0 | } |
1924 | 43 | case ThreadCopy: { |
1925 | | // Step 1.1: Get the address for the src element in the Reduce list. |
1926 | 43 | Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); |
1927 | 43 | SrcElementAddr = |
1928 | 43 | CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
1929 | 43 | SrcElementPtrAddr, PrivateLlvmPtrType), |
1930 | 43 | PrivatePtrType->castAs<PointerType>()); |
1931 | | |
1932 | | // Step 1.2: Get the address for dest element. The destination |
1933 | | // element has already been created on the thread's stack. |
1934 | 43 | DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); |
1935 | 43 | DestElementAddr = |
1936 | 43 | CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
1937 | 43 | DestElementPtrAddr, PrivateLlvmPtrType), |
1938 | 43 | PrivatePtrType->castAs<PointerType>()); |
1939 | 43 | break; |
1940 | 0 | } |
1941 | 0 | case ThreadToScratchpad: { |
1942 | | // Step 1.1: Get the address for the src element in the Reduce list. |
1943 | 0 | Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); |
1944 | 0 | SrcElementAddr = |
1945 | 0 | CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
1946 | 0 | SrcElementPtrAddr, PrivateLlvmPtrType), |
1947 | 0 | PrivatePtrType->castAs<PointerType>()); |
1948 | | |
1949 | | // Step 1.2: Get the address for dest element: |
1950 | | // address = base + index * ElementSizeInChars. |
1951 | 0 | llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); |
1952 | 0 | llvm::Value *CurrentOffset = |
1953 | 0 | Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex); |
1954 | 0 | llvm::Value *ScratchPadElemAbsolutePtrVal = |
1955 | 0 | Bld.CreateNUWAdd(DestBase.getPointer(), CurrentOffset); |
1956 | 0 | ScratchPadElemAbsolutePtrVal = |
1957 | 0 | Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy); |
1958 | 0 | DestElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty, |
1959 | 0 | C.getTypeAlignInChars(Private->getType())); |
1960 | 0 | IncrScratchpadDest = true; |
1961 | 0 | break; |
1962 | 0 | } |
1963 | 0 | case ScratchpadToThread: { |
1964 | | // Step 1.1: Get the address for the src element in the scratchpad. |
1965 | | // address = base + index * ElementSizeInChars. |
1966 | 0 | llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); |
1967 | 0 | llvm::Value *CurrentOffset = |
1968 | 0 | Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex); |
1969 | 0 | llvm::Value *ScratchPadElemAbsolutePtrVal = |
1970 | 0 | Bld.CreateNUWAdd(SrcBase.getPointer(), CurrentOffset); |
1971 | 0 | ScratchPadElemAbsolutePtrVal = |
1972 | 0 | Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy); |
1973 | 0 | SrcElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty, |
1974 | 0 | C.getTypeAlignInChars(Private->getType())); |
1975 | 0 | IncrScratchpadSrc = true; |
1976 | | |
1977 | | // Step 1.2: Create a temporary to store the element in the destination |
1978 | | // Reduce list. |
1979 | 0 | DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); |
1980 | 0 | DestElementAddr = |
1981 | 0 | CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element"); |
1982 | 0 | UpdateDestListPtr = true; |
1983 | 0 | break; |
1984 | 0 | } |
1985 | 86 | } |
1986 | | |
1987 | | // Regardless of src and dest of copy, we emit the load of src |
1988 | | // element as this is required in all directions |
1989 | 86 | SrcElementAddr = Bld.CreateElementBitCast( |
1990 | 86 | SrcElementAddr, CGF.ConvertTypeForMem(Private->getType())); |
1991 | 86 | DestElementAddr = Bld.CreateElementBitCast(DestElementAddr, |
1992 | 86 | SrcElementAddr.getElementType()); |
1993 | | |
1994 | | // Now that all active lanes have read the element in the |
1995 | | // Reduce list, shuffle over the value from the remote lane. |
1996 | 86 | if (ShuffleInElement) { |
1997 | 43 | shuffleAndStore(CGF, SrcElementAddr, DestElementAddr, Private->getType(), |
1998 | 43 | RemoteLaneOffset, Private->getExprLoc()); |
1999 | 43 | } else { |
2000 | 43 | switch (CGF.getEvaluationKind(Private->getType())) { |
2001 | 37 | case TEK_Scalar: { |
2002 | 37 | llvm::Value *Elem = CGF.EmitLoadOfScalar( |
2003 | 37 | SrcElementAddr, /*Volatile=*/false, Private->getType(), |
2004 | 37 | Private->getExprLoc(), LValueBaseInfo(AlignmentSource::Type), |
2005 | 37 | TBAAAccessInfo()); |
2006 | | // Store the source element value to the dest element address. |
2007 | 37 | CGF.EmitStoreOfScalar( |
2008 | 37 | Elem, DestElementAddr, /*Volatile=*/false, Private->getType(), |
2009 | 37 | LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); |
2010 | 37 | break; |
2011 | 0 | } |
2012 | 0 | case TEK_Complex: { |
2013 | 0 | CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex( |
2014 | 0 | CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), |
2015 | 0 | Private->getExprLoc()); |
2016 | 0 | CGF.EmitStoreOfComplex( |
2017 | 0 | Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()), |
2018 | 0 | /*isInit=*/false); |
2019 | 0 | break; |
2020 | 0 | } |
2021 | 6 | case TEK_Aggregate: |
2022 | 6 | CGF.EmitAggregateCopy( |
2023 | 6 | CGF.MakeAddrLValue(DestElementAddr, Private->getType()), |
2024 | 6 | CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), |
2025 | 6 | Private->getType(), AggValueSlot::DoesNotOverlap); |
2026 | 6 | break; |
2027 | 43 | } |
2028 | 43 | } |
2029 | | |
2030 | | // Step 3.1: Modify reference in dest Reduce list as needed. |
2031 | | // Modifying the reference in Reduce list to point to the newly |
2032 | | // created element. The element is live in the current function |
2033 | | // scope and that of functions it invokes (i.e., reduce_function). |
2034 | | // RemoteReduceData[i] = (void*)&RemoteElem |
2035 | 86 | if (UpdateDestListPtr) { |
2036 | 43 | CGF.EmitStoreOfScalar(Bld.CreatePointerBitCastOrAddrSpaceCast( |
2037 | 43 | DestElementAddr.getPointer(), CGF.VoidPtrTy), |
2038 | 43 | DestElementPtrAddr, /*Volatile=*/false, |
2039 | 43 | C.VoidPtrTy); |
2040 | 43 | } |
2041 | | |
2042 | | // Step 4.1: Increment SrcBase/DestBase so that it points to the starting |
2043 | | // address of the next element in scratchpad memory, unless we're currently |
2044 | | // processing the last one. Memory alignment is also taken care of here. |
2045 | 86 | if ((IncrScratchpadDest || IncrScratchpadSrc) && (Idx + 1 < Size)0 ) { |
2046 | | // FIXME: This code doesn't make any sense, it's trying to perform |
2047 | | // integer arithmetic on pointers. |
2048 | 0 | llvm::Value *ScratchpadBasePtr = |
2049 | 0 | IncrScratchpadDest ? DestBase.getPointer() : SrcBase.getPointer(); |
2050 | 0 | llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); |
2051 | 0 | ScratchpadBasePtr = Bld.CreateNUWAdd( |
2052 | 0 | ScratchpadBasePtr, |
2053 | 0 | Bld.CreateNUWMul(ScratchpadWidth, ElementSizeInChars)); |
2054 | | |
2055 | | // Take care of global memory alignment for performance |
2056 | 0 | ScratchpadBasePtr = Bld.CreateNUWSub( |
2057 | 0 | ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1)); |
2058 | 0 | ScratchpadBasePtr = Bld.CreateUDiv( |
2059 | 0 | ScratchpadBasePtr, |
2060 | 0 | llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment)); |
2061 | 0 | ScratchpadBasePtr = Bld.CreateNUWAdd( |
2062 | 0 | ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1)); |
2063 | 0 | ScratchpadBasePtr = Bld.CreateNUWMul( |
2064 | 0 | ScratchpadBasePtr, |
2065 | 0 | llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment)); |
2066 | |
|
2067 | 0 | if (IncrScratchpadDest) |
2068 | 0 | DestBase = |
2069 | 0 | Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign()); |
2070 | 0 | else /* IncrScratchpadSrc = true */ |
2071 | 0 | SrcBase = |
2072 | 0 | Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign()); |
2073 | 0 | } |
2074 | | |
2075 | 86 | ++Idx; |
2076 | 86 | } |
2077 | 56 | } |
2078 | | |
2079 | | /// This function emits a helper that gathers Reduce lists from the first |
2080 | | /// lane of every active warp to lanes in the first warp. |
2081 | | /// |
2082 | | /// void inter_warp_copy_func(void* reduce_data, num_warps) |
2083 | | /// shared smem[warp_size]; |
2084 | | /// For all data entries D in reduce_data: |
2085 | | /// sync |
2086 | | /// If (I am the first lane in each warp) |
2087 | | /// Copy my local D to smem[warp_id] |
2088 | | /// sync |
2089 | | /// if (I am the first warp) |
2090 | | /// Copy smem[thread_id] to my local D |
2091 | | static llvm::Value *emitInterWarpCopyFunction(CodeGenModule &CGM, |
2092 | | ArrayRef<const Expr *> Privates, |
2093 | | QualType ReductionArrayTy, |
2094 | 28 | SourceLocation Loc) { |
2095 | 28 | ASTContext &C = CGM.getContext(); |
2096 | 28 | llvm::Module &M = CGM.getModule(); |
2097 | | |
2098 | | // ReduceList: thread local Reduce list. |
2099 | | // At the stage of the computation when this function is called, partially |
2100 | | // aggregated values reside in the first lane of every active warp. |
2101 | 28 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2102 | 28 | C.VoidPtrTy, ImplicitParamDecl::Other); |
2103 | | // NumWarps: number of warps active in the parallel region. This could |
2104 | | // be smaller than 32 (max warps in a CTA) for partial block reduction. |
2105 | 28 | ImplicitParamDecl NumWarpsArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2106 | 28 | C.getIntTypeForBitwidth(32, /* Signed */ true), |
2107 | 28 | ImplicitParamDecl::Other); |
2108 | 28 | FunctionArgList Args; |
2109 | 28 | Args.push_back(&ReduceListArg); |
2110 | 28 | Args.push_back(&NumWarpsArg); |
2111 | | |
2112 | 28 | const CGFunctionInfo &CGFI = |
2113 | 28 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
2114 | 28 | auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), |
2115 | 28 | llvm::GlobalValue::InternalLinkage, |
2116 | 28 | "_omp_reduction_inter_warp_copy_func", &M); |
2117 | 28 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
2118 | 28 | Fn->setDoesNotRecurse(); |
2119 | 28 | CodeGenFunction CGF(CGM); |
2120 | 28 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
2121 | | |
2122 | 28 | CGBuilderTy &Bld = CGF.Builder; |
2123 | | |
2124 | | // This array is used as a medium to transfer, one reduce element at a time, |
2125 | | // the data from the first lane of every warp to lanes in the first warp |
2126 | | // in order to perform the final step of a reduction in a parallel region |
2127 | | // (reduction across warps). The array is placed in NVPTX __shared__ memory |
2128 | | // for reduced latency, as well as to have a distinct copy for concurrently |
2129 | | // executing target regions. The array is declared with common linkage so |
2130 | | // as to be shared across compilation units. |
2131 | 28 | StringRef TransferMediumName = |
2132 | 28 | "__openmp_nvptx_data_transfer_temporary_storage"; |
2133 | 28 | llvm::GlobalVariable *TransferMedium = |
2134 | 28 | M.getGlobalVariable(TransferMediumName); |
2135 | 28 | unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; |
2136 | 28 | if (!TransferMedium) { |
2137 | 10 | auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize); |
2138 | 10 | unsigned SharedAddressSpace = C.getTargetAddressSpace(LangAS::cuda_shared); |
2139 | 10 | TransferMedium = new llvm::GlobalVariable( |
2140 | 10 | M, Ty, /*isConstant=*/false, llvm::GlobalVariable::WeakAnyLinkage, |
2141 | 10 | llvm::UndefValue::get(Ty), TransferMediumName, |
2142 | 10 | /*InsertBefore=*/nullptr, llvm::GlobalVariable::NotThreadLocal, |
2143 | 10 | SharedAddressSpace); |
2144 | 10 | CGM.addCompilerUsedGlobal(TransferMedium); |
2145 | 10 | } |
2146 | | |
2147 | 28 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
2148 | | // Get the CUDA thread id of the current OpenMP thread on the GPU. |
2149 | 28 | llvm::Value *ThreadID = RT.getGPUThreadID(CGF); |
2150 | | // nvptx_lane_id = nvptx_id % warpsize |
2151 | 28 | llvm::Value *LaneID = getNVPTXLaneID(CGF); |
2152 | | // nvptx_warp_id = nvptx_id / warpsize |
2153 | 28 | llvm::Value *WarpID = getNVPTXWarpID(CGF); |
2154 | | |
2155 | 28 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
2156 | 28 | llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
2157 | 28 | Address LocalReduceList( |
2158 | 28 | Bld.CreatePointerBitCastOrAddrSpaceCast( |
2159 | 28 | CGF.EmitLoadOfScalar( |
2160 | 28 | AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc, |
2161 | 28 | LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()), |
2162 | 28 | ElemTy->getPointerTo()), |
2163 | 28 | ElemTy, CGF.getPointerAlign()); |
2164 | | |
2165 | 28 | unsigned Idx = 0; |
2166 | 43 | for (const Expr *Private : Privates) { |
2167 | | // |
2168 | | // Warp master copies reduce element to transfer medium in __shared__ |
2169 | | // memory. |
2170 | | // |
2171 | 43 | unsigned RealTySize = |
2172 | 43 | C.getTypeSizeInChars(Private->getType()) |
2173 | 43 | .alignTo(C.getTypeAlignInChars(Private->getType())) |
2174 | 43 | .getQuantity(); |
2175 | 107 | for (unsigned TySize = 4; TySize > 0 && RealTySize > 0101 ; TySize /=264 ) { |
2176 | 64 | unsigned NumIters = RealTySize / TySize; |
2177 | 64 | if (NumIters == 0) |
2178 | 21 | continue; |
2179 | 43 | QualType CType = C.getIntTypeForBitwidth( |
2180 | 43 | C.toBits(CharUnits::fromQuantity(TySize)), /*Signed=*/1); |
2181 | 43 | llvm::Type *CopyType = CGF.ConvertTypeForMem(CType); |
2182 | 43 | CharUnits Align = CharUnits::fromQuantity(TySize); |
2183 | 43 | llvm::Value *Cnt = nullptr; |
2184 | 43 | Address CntAddr = Address::invalid(); |
2185 | 43 | llvm::BasicBlock *PrecondBB = nullptr; |
2186 | 43 | llvm::BasicBlock *ExitBB = nullptr; |
2187 | 43 | if (NumIters > 1) { |
2188 | 13 | CntAddr = CGF.CreateMemTemp(C.IntTy, ".cnt.addr"); |
2189 | 13 | CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.IntTy), CntAddr, |
2190 | 13 | /*Volatile=*/false, C.IntTy); |
2191 | 13 | PrecondBB = CGF.createBasicBlock("precond"); |
2192 | 13 | ExitBB = CGF.createBasicBlock("exit"); |
2193 | 13 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("body"); |
2194 | | // There is no need to emit line number for unconditional branch. |
2195 | 13 | (void)ApplyDebugLocation::CreateEmpty(CGF); |
2196 | 13 | CGF.EmitBlock(PrecondBB); |
2197 | 13 | Cnt = CGF.EmitLoadOfScalar(CntAddr, /*Volatile=*/false, C.IntTy, Loc); |
2198 | 13 | llvm::Value *Cmp = |
2199 | 13 | Bld.CreateICmpULT(Cnt, llvm::ConstantInt::get(CGM.IntTy, NumIters)); |
2200 | 13 | Bld.CreateCondBr(Cmp, BodyBB, ExitBB); |
2201 | 13 | CGF.EmitBlock(BodyBB); |
2202 | 13 | } |
2203 | | // kmpc_barrier. |
2204 | 43 | CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, |
2205 | 43 | /*EmitChecks=*/false, |
2206 | 43 | /*ForceSimpleCall=*/true); |
2207 | 43 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); |
2208 | 43 | llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); |
2209 | 43 | llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); |
2210 | | |
2211 | | // if (lane_id == 0) |
2212 | 43 | llvm::Value *IsWarpMaster = Bld.CreateIsNull(LaneID, "warp_master"); |
2213 | 43 | Bld.CreateCondBr(IsWarpMaster, ThenBB, ElseBB); |
2214 | 43 | CGF.EmitBlock(ThenBB); |
2215 | | |
2216 | | // Reduce element = LocalReduceList[i] |
2217 | 43 | Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
2218 | 43 | llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( |
2219 | 43 | ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); |
2220 | | // elemptr = ((CopyType*)(elemptrptr)) + I |
2221 | 43 | Address ElemPtr(ElemPtrPtr, CGF.Int8Ty, Align); |
2222 | 43 | ElemPtr = Bld.CreateElementBitCast(ElemPtr, CopyType); |
2223 | 43 | if (NumIters > 1) |
2224 | 13 | ElemPtr = Bld.CreateGEP(ElemPtr, Cnt); |
2225 | | |
2226 | | // Get pointer to location in transfer medium. |
2227 | | // MediumPtr = &medium[warp_id] |
2228 | 43 | llvm::Value *MediumPtrVal = Bld.CreateInBoundsGEP( |
2229 | 43 | TransferMedium->getValueType(), TransferMedium, |
2230 | 43 | {llvm::Constant::getNullValue(CGM.Int64Ty), WarpID}); |
2231 | | // Casting to actual data type. |
2232 | | // MediumPtr = (CopyType*)MediumPtrAddr; |
2233 | 43 | Address MediumPtr( |
2234 | 43 | Bld.CreateBitCast( |
2235 | 43 | MediumPtrVal, |
2236 | 43 | CopyType->getPointerTo( |
2237 | 43 | MediumPtrVal->getType()->getPointerAddressSpace())), |
2238 | 43 | CopyType, Align); |
2239 | | |
2240 | | // elem = *elemptr |
2241 | | //*MediumPtr = elem |
2242 | 43 | llvm::Value *Elem = CGF.EmitLoadOfScalar( |
2243 | 43 | ElemPtr, /*Volatile=*/false, CType, Loc, |
2244 | 43 | LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); |
2245 | | // Store the source element value to the dest element address. |
2246 | 43 | CGF.EmitStoreOfScalar(Elem, MediumPtr, /*Volatile=*/true, CType, |
2247 | 43 | LValueBaseInfo(AlignmentSource::Type), |
2248 | 43 | TBAAAccessInfo()); |
2249 | | |
2250 | 43 | Bld.CreateBr(MergeBB); |
2251 | | |
2252 | 43 | CGF.EmitBlock(ElseBB); |
2253 | 43 | Bld.CreateBr(MergeBB); |
2254 | | |
2255 | 43 | CGF.EmitBlock(MergeBB); |
2256 | | |
2257 | | // kmpc_barrier. |
2258 | 43 | CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, |
2259 | 43 | /*EmitChecks=*/false, |
2260 | 43 | /*ForceSimpleCall=*/true); |
2261 | | |
2262 | | // |
2263 | | // Warp 0 copies reduce element from transfer medium. |
2264 | | // |
2265 | 43 | llvm::BasicBlock *W0ThenBB = CGF.createBasicBlock("then"); |
2266 | 43 | llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else"); |
2267 | 43 | llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont"); |
2268 | | |
2269 | 43 | Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg); |
2270 | 43 | llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar( |
2271 | 43 | AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc); |
2272 | | |
2273 | | // Up to 32 threads in warp 0 are active. |
2274 | 43 | llvm::Value *IsActiveThread = |
2275 | 43 | Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread"); |
2276 | 43 | Bld.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB); |
2277 | | |
2278 | 43 | CGF.EmitBlock(W0ThenBB); |
2279 | | |
2280 | | // SrcMediumPtr = &medium[tid] |
2281 | 43 | llvm::Value *SrcMediumPtrVal = Bld.CreateInBoundsGEP( |
2282 | 43 | TransferMedium->getValueType(), TransferMedium, |
2283 | 43 | {llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID}); |
2284 | | // SrcMediumVal = *SrcMediumPtr; |
2285 | 43 | Address SrcMediumPtr( |
2286 | 43 | Bld.CreateBitCast( |
2287 | 43 | SrcMediumPtrVal, |
2288 | 43 | CopyType->getPointerTo( |
2289 | 43 | SrcMediumPtrVal->getType()->getPointerAddressSpace())), |
2290 | 43 | CopyType, Align); |
2291 | | |
2292 | | // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I |
2293 | 43 | Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
2294 | 43 | llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar( |
2295 | 43 | TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc); |
2296 | 43 | Address TargetElemPtr(TargetElemPtrVal, CGF.Int8Ty, Align); |
2297 | 43 | TargetElemPtr = Bld.CreateElementBitCast(TargetElemPtr, CopyType); |
2298 | 43 | if (NumIters > 1) |
2299 | 13 | TargetElemPtr = Bld.CreateGEP(TargetElemPtr, Cnt); |
2300 | | |
2301 | | // *TargetElemPtr = SrcMediumVal; |
2302 | 43 | llvm::Value *SrcMediumValue = |
2303 | 43 | CGF.EmitLoadOfScalar(SrcMediumPtr, /*Volatile=*/true, CType, Loc); |
2304 | 43 | CGF.EmitStoreOfScalar(SrcMediumValue, TargetElemPtr, /*Volatile=*/false, |
2305 | 43 | CType); |
2306 | 43 | Bld.CreateBr(W0MergeBB); |
2307 | | |
2308 | 43 | CGF.EmitBlock(W0ElseBB); |
2309 | 43 | Bld.CreateBr(W0MergeBB); |
2310 | | |
2311 | 43 | CGF.EmitBlock(W0MergeBB); |
2312 | | |
2313 | 43 | if (NumIters > 1) { |
2314 | 13 | Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1)); |
2315 | 13 | CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy); |
2316 | 13 | CGF.EmitBranch(PrecondBB); |
2317 | 13 | (void)ApplyDebugLocation::CreateEmpty(CGF); |
2318 | 13 | CGF.EmitBlock(ExitBB); |
2319 | 13 | } |
2320 | 43 | RealTySize %= TySize; |
2321 | 43 | } |
2322 | 43 | ++Idx; |
2323 | 43 | } |
2324 | | |
2325 | 28 | CGF.FinishFunction(); |
2326 | 28 | return Fn; |
2327 | 28 | } |
2328 | | |
2329 | | /// Emit a helper that reduces data across two OpenMP threads (lanes) |
2330 | | /// in the same warp. It uses shuffle instructions to copy over data from |
2331 | | /// a remote lane's stack. The reduction algorithm performed is specified |
2332 | | /// by the fourth parameter. |
2333 | | /// |
2334 | | /// Algorithm Versions. |
2335 | | /// Full Warp Reduce (argument value 0): |
2336 | | /// This algorithm assumes that all 32 lanes are active and gathers |
2337 | | /// data from these 32 lanes, producing a single resultant value. |
2338 | | /// Contiguous Partial Warp Reduce (argument value 1): |
2339 | | /// This algorithm assumes that only a *contiguous* subset of lanes |
2340 | | /// are active. This happens for the last warp in a parallel region |
2341 | | /// when the user specified num_threads is not an integer multiple of |
2342 | | /// 32. This contiguous subset always starts with the zeroth lane. |
2343 | | /// Partial Warp Reduce (argument value 2): |
2344 | | /// This algorithm gathers data from any number of lanes at any position. |
2345 | | /// All reduced values are stored in the lowest possible lane. The set |
2346 | | /// of problems every algorithm addresses is a super set of those |
2347 | | /// addressable by algorithms with a lower version number. Overhead |
2348 | | /// increases as algorithm version increases. |
2349 | | /// |
2350 | | /// Terminology |
2351 | | /// Reduce element: |
2352 | | /// Reduce element refers to the individual data field with primitive |
2353 | | /// data types to be combined and reduced across threads. |
2354 | | /// Reduce list: |
2355 | | /// Reduce list refers to a collection of local, thread-private |
2356 | | /// reduce elements. |
2357 | | /// Remote Reduce list: |
2358 | | /// Remote Reduce list refers to a collection of remote (relative to |
2359 | | /// the current thread) reduce elements. |
2360 | | /// |
2361 | | /// We distinguish between three states of threads that are important to |
2362 | | /// the implementation of this function. |
2363 | | /// Alive threads: |
2364 | | /// Threads in a warp executing the SIMT instruction, as distinguished from |
2365 | | /// threads that are inactive due to divergent control flow. |
2366 | | /// Active threads: |
2367 | | /// The minimal set of threads that has to be alive upon entry to this |
2368 | | /// function. The computation is correct iff active threads are alive. |
2369 | | /// Some threads are alive but they are not active because they do not |
2370 | | /// contribute to the computation in any useful manner. Turning them off |
2371 | | /// may introduce control flow overheads without any tangible benefits. |
2372 | | /// Effective threads: |
2373 | | /// In order to comply with the argument requirements of the shuffle |
2374 | | /// function, we must keep all lanes holding data alive. But at most |
2375 | | /// half of them perform value aggregation; we refer to this half of |
2376 | | /// threads as effective. The other half is simply handing off their |
2377 | | /// data. |
2378 | | /// |
2379 | | /// Procedure |
2380 | | /// Value shuffle: |
2381 | | /// In this step active threads transfer data from higher lane positions |
2382 | | /// in the warp to lower lane positions, creating Remote Reduce list. |
2383 | | /// Value aggregation: |
2384 | | /// In this step, effective threads combine their thread local Reduce list |
2385 | | /// with Remote Reduce list and store the result in the thread local |
2386 | | /// Reduce list. |
2387 | | /// Value copy: |
2388 | | /// In this step, we deal with the assumption made by algorithm 2 |
2389 | | /// (i.e. contiguity assumption). When we have an odd number of lanes |
2390 | | /// active, say 2k+1, only k threads will be effective and therefore k |
2391 | | /// new values will be produced. However, the Reduce list owned by the |
2392 | | /// (2k+1)th thread is ignored in the value aggregation. Therefore |
2393 | | /// we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so |
2394 | | /// that the contiguity assumption still holds. |
2395 | | static llvm::Function *emitShuffleAndReduceFunction( |
2396 | | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
2397 | 28 | QualType ReductionArrayTy, llvm::Function *ReduceFn, SourceLocation Loc) { |
2398 | 28 | ASTContext &C = CGM.getContext(); |
2399 | | |
2400 | | // Thread local Reduce list used to host the values of data to be reduced. |
2401 | 28 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2402 | 28 | C.VoidPtrTy, ImplicitParamDecl::Other); |
2403 | | // Current lane id; could be logical. |
2404 | 28 | ImplicitParamDecl LaneIDArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.ShortTy, |
2405 | 28 | ImplicitParamDecl::Other); |
2406 | | // Offset of the remote source lane relative to the current lane. |
2407 | 28 | ImplicitParamDecl RemoteLaneOffsetArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2408 | 28 | C.ShortTy, ImplicitParamDecl::Other); |
2409 | | // Algorithm version. This is expected to be known at compile time. |
2410 | 28 | ImplicitParamDecl AlgoVerArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2411 | 28 | C.ShortTy, ImplicitParamDecl::Other); |
2412 | 28 | FunctionArgList Args; |
2413 | 28 | Args.push_back(&ReduceListArg); |
2414 | 28 | Args.push_back(&LaneIDArg); |
2415 | 28 | Args.push_back(&RemoteLaneOffsetArg); |
2416 | 28 | Args.push_back(&AlgoVerArg); |
2417 | | |
2418 | 28 | const CGFunctionInfo &CGFI = |
2419 | 28 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
2420 | 28 | auto *Fn = llvm::Function::Create( |
2421 | 28 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
2422 | 28 | "_omp_reduction_shuffle_and_reduce_func", &CGM.getModule()); |
2423 | 28 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
2424 | 28 | Fn->setDoesNotRecurse(); |
2425 | | |
2426 | 28 | CodeGenFunction CGF(CGM); |
2427 | 28 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
2428 | | |
2429 | 28 | CGBuilderTy &Bld = CGF.Builder; |
2430 | | |
2431 | 28 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
2432 | 28 | llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
2433 | 28 | Address LocalReduceList( |
2434 | 28 | Bld.CreatePointerBitCastOrAddrSpaceCast( |
2435 | 28 | CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, |
2436 | 28 | C.VoidPtrTy, SourceLocation()), |
2437 | 28 | ElemTy->getPointerTo()), |
2438 | 28 | ElemTy, CGF.getPointerAlign()); |
2439 | | |
2440 | 28 | Address AddrLaneIDArg = CGF.GetAddrOfLocalVar(&LaneIDArg); |
2441 | 28 | llvm::Value *LaneIDArgVal = CGF.EmitLoadOfScalar( |
2442 | 28 | AddrLaneIDArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); |
2443 | | |
2444 | 28 | Address AddrRemoteLaneOffsetArg = CGF.GetAddrOfLocalVar(&RemoteLaneOffsetArg); |
2445 | 28 | llvm::Value *RemoteLaneOffsetArgVal = CGF.EmitLoadOfScalar( |
2446 | 28 | AddrRemoteLaneOffsetArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); |
2447 | | |
2448 | 28 | Address AddrAlgoVerArg = CGF.GetAddrOfLocalVar(&AlgoVerArg); |
2449 | 28 | llvm::Value *AlgoVerArgVal = CGF.EmitLoadOfScalar( |
2450 | 28 | AddrAlgoVerArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); |
2451 | | |
2452 | | // Create a local thread-private variable to host the Reduce list |
2453 | | // from a remote lane. |
2454 | 28 | Address RemoteReduceList = |
2455 | 28 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.remote_reduce_list"); |
2456 | | |
2457 | | // This loop iterates through the list of reduce elements and copies, |
2458 | | // element by element, from a remote lane in the warp to RemoteReduceList, |
2459 | | // hosted on the thread's stack. |
2460 | 28 | emitReductionListCopy(RemoteLaneToThread, CGF, ReductionArrayTy, Privates, |
2461 | 28 | LocalReduceList, RemoteReduceList, |
2462 | 28 | {/*RemoteLaneOffset=*/RemoteLaneOffsetArgVal, |
2463 | 28 | /*ScratchpadIndex=*/nullptr, |
2464 | 28 | /*ScratchpadWidth=*/nullptr}); |
2465 | | |
2466 | | // The actions to be performed on the Remote Reduce list is dependent |
2467 | | // on the algorithm version. |
2468 | | // |
2469 | | // if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 && |
2470 | | // LaneId % 2 == 0 && Offset > 0): |
2471 | | // do the reduction value aggregation |
2472 | | // |
2473 | | // The thread local variable Reduce list is mutated in place to host the |
2474 | | // reduced data, which is the aggregated value produced from local and |
2475 | | // remote lanes. |
2476 | | // |
2477 | | // Note that AlgoVer is expected to be a constant integer known at compile |
2478 | | // time. |
2479 | | // When AlgoVer==0, the first conjunction evaluates to true, making |
2480 | | // the entire predicate true during compile time. |
2481 | | // When AlgoVer==1, the second conjunction has only the second part to be |
2482 | | // evaluated during runtime. Other conjunctions evaluates to false |
2483 | | // during compile time. |
2484 | | // When AlgoVer==2, the third conjunction has only the second part to be |
2485 | | // evaluated during runtime. Other conjunctions evaluates to false |
2486 | | // during compile time. |
2487 | 28 | llvm::Value *CondAlgo0 = Bld.CreateIsNull(AlgoVerArgVal); |
2488 | | |
2489 | 28 | llvm::Value *Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); |
2490 | 28 | llvm::Value *CondAlgo1 = Bld.CreateAnd( |
2491 | 28 | Algo1, Bld.CreateICmpULT(LaneIDArgVal, RemoteLaneOffsetArgVal)); |
2492 | | |
2493 | 28 | llvm::Value *Algo2 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(2)); |
2494 | 28 | llvm::Value *CondAlgo2 = Bld.CreateAnd( |
2495 | 28 | Algo2, Bld.CreateIsNull(Bld.CreateAnd(LaneIDArgVal, Bld.getInt16(1)))); |
2496 | 28 | CondAlgo2 = Bld.CreateAnd( |
2497 | 28 | CondAlgo2, Bld.CreateICmpSGT(RemoteLaneOffsetArgVal, Bld.getInt16(0))); |
2498 | | |
2499 | 28 | llvm::Value *CondReduce = Bld.CreateOr(CondAlgo0, CondAlgo1); |
2500 | 28 | CondReduce = Bld.CreateOr(CondReduce, CondAlgo2); |
2501 | | |
2502 | 28 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); |
2503 | 28 | llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); |
2504 | 28 | llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); |
2505 | 28 | Bld.CreateCondBr(CondReduce, ThenBB, ElseBB); |
2506 | | |
2507 | 28 | CGF.EmitBlock(ThenBB); |
2508 | | // reduce_function(LocalReduceList, RemoteReduceList) |
2509 | 28 | llvm::Value *LocalReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
2510 | 28 | LocalReduceList.getPointer(), CGF.VoidPtrTy); |
2511 | 28 | llvm::Value *RemoteReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
2512 | 28 | RemoteReduceList.getPointer(), CGF.VoidPtrTy); |
2513 | 28 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
2514 | 28 | CGF, Loc, ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr}); |
2515 | 28 | Bld.CreateBr(MergeBB); |
2516 | | |
2517 | 28 | CGF.EmitBlock(ElseBB); |
2518 | 28 | Bld.CreateBr(MergeBB); |
2519 | | |
2520 | 28 | CGF.EmitBlock(MergeBB); |
2521 | | |
2522 | | // if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local |
2523 | | // Reduce list. |
2524 | 28 | Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); |
2525 | 28 | llvm::Value *CondCopy = Bld.CreateAnd( |
2526 | 28 | Algo1, Bld.CreateICmpUGE(LaneIDArgVal, RemoteLaneOffsetArgVal)); |
2527 | | |
2528 | 28 | llvm::BasicBlock *CpyThenBB = CGF.createBasicBlock("then"); |
2529 | 28 | llvm::BasicBlock *CpyElseBB = CGF.createBasicBlock("else"); |
2530 | 28 | llvm::BasicBlock *CpyMergeBB = CGF.createBasicBlock("ifcont"); |
2531 | 28 | Bld.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB); |
2532 | | |
2533 | 28 | CGF.EmitBlock(CpyThenBB); |
2534 | 28 | emitReductionListCopy(ThreadCopy, CGF, ReductionArrayTy, Privates, |
2535 | 28 | RemoteReduceList, LocalReduceList); |
2536 | 28 | Bld.CreateBr(CpyMergeBB); |
2537 | | |
2538 | 28 | CGF.EmitBlock(CpyElseBB); |
2539 | 28 | Bld.CreateBr(CpyMergeBB); |
2540 | | |
2541 | 28 | CGF.EmitBlock(CpyMergeBB); |
2542 | | |
2543 | 28 | CGF.FinishFunction(); |
2544 | 28 | return Fn; |
2545 | 28 | } |
2546 | | |
2547 | | /// This function emits a helper that copies all the reduction variables from |
2548 | | /// the team into the provided global buffer for the reduction variables. |
2549 | | /// |
2550 | | /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) |
2551 | | /// For all data entries D in reduce_data: |
2552 | | /// Copy local D to buffer.D[Idx] |
2553 | | static llvm::Value *emitListToGlobalCopyFunction( |
2554 | | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
2555 | | QualType ReductionArrayTy, SourceLocation Loc, |
2556 | | const RecordDecl *TeamReductionRec, |
2557 | | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
2558 | 9 | &VarFieldMap) { |
2559 | 9 | ASTContext &C = CGM.getContext(); |
2560 | | |
2561 | | // Buffer: global reduction buffer. |
2562 | 9 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2563 | 9 | C.VoidPtrTy, ImplicitParamDecl::Other); |
2564 | | // Idx: index of the buffer. |
2565 | 9 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
2566 | 9 | ImplicitParamDecl::Other); |
2567 | | // ReduceList: thread local Reduce list. |
2568 | 9 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2569 | 9 | C.VoidPtrTy, ImplicitParamDecl::Other); |
2570 | 9 | FunctionArgList Args; |
2571 | 9 | Args.push_back(&BufferArg); |
2572 | 9 | Args.push_back(&IdxArg); |
2573 | 9 | Args.push_back(&ReduceListArg); |
2574 | | |
2575 | 9 | const CGFunctionInfo &CGFI = |
2576 | 9 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
2577 | 9 | auto *Fn = llvm::Function::Create( |
2578 | 9 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
2579 | 9 | "_omp_reduction_list_to_global_copy_func", &CGM.getModule()); |
2580 | 9 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
2581 | 9 | Fn->setDoesNotRecurse(); |
2582 | 9 | CodeGenFunction CGF(CGM); |
2583 | 9 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
2584 | | |
2585 | 9 | CGBuilderTy &Bld = CGF.Builder; |
2586 | | |
2587 | 9 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
2588 | 9 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
2589 | 9 | llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
2590 | 9 | Address LocalReduceList( |
2591 | 9 | Bld.CreatePointerBitCastOrAddrSpaceCast( |
2592 | 9 | CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, |
2593 | 9 | C.VoidPtrTy, Loc), |
2594 | 9 | ElemTy->getPointerTo()), |
2595 | 9 | ElemTy, CGF.getPointerAlign()); |
2596 | 9 | QualType StaticTy = C.getRecordType(TeamReductionRec); |
2597 | 9 | llvm::Type *LLVMReductionsBufferTy = |
2598 | 9 | CGM.getTypes().ConvertTypeForMem(StaticTy); |
2599 | 9 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
2600 | 9 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
2601 | 9 | LLVMReductionsBufferTy->getPointerTo()); |
2602 | 9 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
2603 | 9 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
2604 | 9 | /*Volatile=*/false, C.IntTy, |
2605 | 9 | Loc)}; |
2606 | 9 | unsigned Idx = 0; |
2607 | 15 | for (const Expr *Private : Privates) { |
2608 | | // Reduce element = LocalReduceList[i] |
2609 | 15 | Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
2610 | 15 | llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( |
2611 | 15 | ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); |
2612 | | // elemptr = ((CopyType*)(elemptrptr)) + I |
2613 | 15 | ElemTy = CGF.ConvertTypeForMem(Private->getType()); |
2614 | 15 | ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
2615 | 15 | ElemPtrPtr, ElemTy->getPointerTo()); |
2616 | 15 | Address ElemPtr = |
2617 | 15 | Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType())); |
2618 | 15 | const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); |
2619 | | // Global = Buffer.VD[Idx]; |
2620 | 15 | const FieldDecl *FD = VarFieldMap.lookup(VD); |
2621 | 15 | LValue GlobLVal = CGF.EmitLValueForField( |
2622 | 15 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
2623 | 15 | Address GlobAddr = GlobLVal.getAddress(CGF); |
2624 | 15 | llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(), |
2625 | 15 | GlobAddr.getPointer(), Idxs); |
2626 | 15 | GlobLVal.setAddress(Address(BufferPtr, |
2627 | 15 | CGF.ConvertTypeForMem(Private->getType()), |
2628 | 15 | GlobAddr.getAlignment())); |
2629 | 15 | switch (CGF.getEvaluationKind(Private->getType())) { |
2630 | 15 | case TEK_Scalar: { |
2631 | 15 | llvm::Value *V = CGF.EmitLoadOfScalar( |
2632 | 15 | ElemPtr, /*Volatile=*/false, Private->getType(), Loc, |
2633 | 15 | LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); |
2634 | 15 | CGF.EmitStoreOfScalar(V, GlobLVal); |
2635 | 15 | break; |
2636 | 0 | } |
2637 | 0 | case TEK_Complex: { |
2638 | 0 | CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex( |
2639 | 0 | CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc); |
2640 | 0 | CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false); |
2641 | 0 | break; |
2642 | 0 | } |
2643 | 0 | case TEK_Aggregate: |
2644 | 0 | CGF.EmitAggregateCopy(GlobLVal, |
2645 | 0 | CGF.MakeAddrLValue(ElemPtr, Private->getType()), |
2646 | 0 | Private->getType(), AggValueSlot::DoesNotOverlap); |
2647 | 0 | break; |
2648 | 15 | } |
2649 | 15 | ++Idx; |
2650 | 15 | } |
2651 | | |
2652 | 9 | CGF.FinishFunction(); |
2653 | 9 | return Fn; |
2654 | 9 | } |
2655 | | |
2656 | | /// This function emits a helper that reduces all the reduction variables from |
2657 | | /// the team into the provided global buffer for the reduction variables. |
2658 | | /// |
2659 | | /// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data) |
2660 | | /// void *GlobPtrs[]; |
2661 | | /// GlobPtrs[0] = (void*)&buffer.D0[Idx]; |
2662 | | /// ... |
2663 | | /// GlobPtrs[N] = (void*)&buffer.DN[Idx]; |
2664 | | /// reduce_function(GlobPtrs, reduce_data); |
2665 | | static llvm::Value *emitListToGlobalReduceFunction( |
2666 | | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
2667 | | QualType ReductionArrayTy, SourceLocation Loc, |
2668 | | const RecordDecl *TeamReductionRec, |
2669 | | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
2670 | | &VarFieldMap, |
2671 | 9 | llvm::Function *ReduceFn) { |
2672 | 9 | ASTContext &C = CGM.getContext(); |
2673 | | |
2674 | | // Buffer: global reduction buffer. |
2675 | 9 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2676 | 9 | C.VoidPtrTy, ImplicitParamDecl::Other); |
2677 | | // Idx: index of the buffer. |
2678 | 9 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
2679 | 9 | ImplicitParamDecl::Other); |
2680 | | // ReduceList: thread local Reduce list. |
2681 | 9 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2682 | 9 | C.VoidPtrTy, ImplicitParamDecl::Other); |
2683 | 9 | FunctionArgList Args; |
2684 | 9 | Args.push_back(&BufferArg); |
2685 | 9 | Args.push_back(&IdxArg); |
2686 | 9 | Args.push_back(&ReduceListArg); |
2687 | | |
2688 | 9 | const CGFunctionInfo &CGFI = |
2689 | 9 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
2690 | 9 | auto *Fn = llvm::Function::Create( |
2691 | 9 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
2692 | 9 | "_omp_reduction_list_to_global_reduce_func", &CGM.getModule()); |
2693 | 9 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
2694 | 9 | Fn->setDoesNotRecurse(); |
2695 | 9 | CodeGenFunction CGF(CGM); |
2696 | 9 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
2697 | | |
2698 | 9 | CGBuilderTy &Bld = CGF.Builder; |
2699 | | |
2700 | 9 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
2701 | 9 | QualType StaticTy = C.getRecordType(TeamReductionRec); |
2702 | 9 | llvm::Type *LLVMReductionsBufferTy = |
2703 | 9 | CGM.getTypes().ConvertTypeForMem(StaticTy); |
2704 | 9 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
2705 | 9 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
2706 | 9 | LLVMReductionsBufferTy->getPointerTo()); |
2707 | | |
2708 | | // 1. Build a list of reduction variables. |
2709 | | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
2710 | 9 | Address ReductionList = |
2711 | 9 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); |
2712 | 9 | auto IPriv = Privates.begin(); |
2713 | 9 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
2714 | 9 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
2715 | 9 | /*Volatile=*/false, C.IntTy, |
2716 | 9 | Loc)}; |
2717 | 9 | unsigned Idx = 0; |
2718 | 24 | for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx15 ) { |
2719 | 15 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
2720 | | // Global = Buffer.VD[Idx]; |
2721 | 15 | const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); |
2722 | 15 | const FieldDecl *FD = VarFieldMap.lookup(VD); |
2723 | 15 | LValue GlobLVal = CGF.EmitLValueForField( |
2724 | 15 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
2725 | 15 | Address GlobAddr = GlobLVal.getAddress(CGF); |
2726 | 15 | llvm::Value *BufferPtr = Bld.CreateInBoundsGEP( |
2727 | 15 | GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs); |
2728 | 15 | llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr); |
2729 | 15 | CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy); |
2730 | 15 | if ((*IPriv)->getType()->isVariablyModifiedType()) { |
2731 | | // Store array size. |
2732 | 0 | ++Idx; |
2733 | 0 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
2734 | 0 | llvm::Value *Size = CGF.Builder.CreateIntCast( |
2735 | 0 | CGF.getVLASize( |
2736 | 0 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) |
2737 | 0 | .NumElts, |
2738 | 0 | CGF.SizeTy, /*isSigned=*/false); |
2739 | 0 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), |
2740 | 0 | Elem); |
2741 | 0 | } |
2742 | 15 | } |
2743 | | |
2744 | | // Call reduce_function(GlobalReduceList, ReduceList) |
2745 | 9 | llvm::Value *GlobalReduceList = |
2746 | 9 | CGF.EmitCastToVoidPtr(ReductionList.getPointer()); |
2747 | 9 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
2748 | 9 | llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( |
2749 | 9 | AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); |
2750 | 9 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
2751 | 9 | CGF, Loc, ReduceFn, {GlobalReduceList, ReducedPtr}); |
2752 | 9 | CGF.FinishFunction(); |
2753 | 9 | return Fn; |
2754 | 9 | } |
2755 | | |
2756 | | /// This function emits a helper that copies all the reduction variables from |
2757 | | /// the team into the provided global buffer for the reduction variables. |
2758 | | /// |
2759 | | /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) |
2760 | | /// For all data entries D in reduce_data: |
2761 | | /// Copy buffer.D[Idx] to local D; |
2762 | | static llvm::Value *emitGlobalToListCopyFunction( |
2763 | | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
2764 | | QualType ReductionArrayTy, SourceLocation Loc, |
2765 | | const RecordDecl *TeamReductionRec, |
2766 | | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
2767 | 9 | &VarFieldMap) { |
2768 | 9 | ASTContext &C = CGM.getContext(); |
2769 | | |
2770 | | // Buffer: global reduction buffer. |
2771 | 9 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2772 | 9 | C.VoidPtrTy, ImplicitParamDecl::Other); |
2773 | | // Idx: index of the buffer. |
2774 | 9 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
2775 | 9 | ImplicitParamDecl::Other); |
2776 | | // ReduceList: thread local Reduce list. |
2777 | 9 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2778 | 9 | C.VoidPtrTy, ImplicitParamDecl::Other); |
2779 | 9 | FunctionArgList Args; |
2780 | 9 | Args.push_back(&BufferArg); |
2781 | 9 | Args.push_back(&IdxArg); |
2782 | 9 | Args.push_back(&ReduceListArg); |
2783 | | |
2784 | 9 | const CGFunctionInfo &CGFI = |
2785 | 9 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
2786 | 9 | auto *Fn = llvm::Function::Create( |
2787 | 9 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
2788 | 9 | "_omp_reduction_global_to_list_copy_func", &CGM.getModule()); |
2789 | 9 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
2790 | 9 | Fn->setDoesNotRecurse(); |
2791 | 9 | CodeGenFunction CGF(CGM); |
2792 | 9 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
2793 | | |
2794 | 9 | CGBuilderTy &Bld = CGF.Builder; |
2795 | | |
2796 | 9 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
2797 | 9 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
2798 | 9 | llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
2799 | 9 | Address LocalReduceList( |
2800 | 9 | Bld.CreatePointerBitCastOrAddrSpaceCast( |
2801 | 9 | CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, |
2802 | 9 | C.VoidPtrTy, Loc), |
2803 | 9 | ElemTy->getPointerTo()), |
2804 | 9 | ElemTy, CGF.getPointerAlign()); |
2805 | 9 | QualType StaticTy = C.getRecordType(TeamReductionRec); |
2806 | 9 | llvm::Type *LLVMReductionsBufferTy = |
2807 | 9 | CGM.getTypes().ConvertTypeForMem(StaticTy); |
2808 | 9 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
2809 | 9 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
2810 | 9 | LLVMReductionsBufferTy->getPointerTo()); |
2811 | | |
2812 | 9 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
2813 | 9 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
2814 | 9 | /*Volatile=*/false, C.IntTy, |
2815 | 9 | Loc)}; |
2816 | 9 | unsigned Idx = 0; |
2817 | 15 | for (const Expr *Private : Privates) { |
2818 | | // Reduce element = LocalReduceList[i] |
2819 | 15 | Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
2820 | 15 | llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( |
2821 | 15 | ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); |
2822 | | // elemptr = ((CopyType*)(elemptrptr)) + I |
2823 | 15 | ElemTy = CGF.ConvertTypeForMem(Private->getType()); |
2824 | 15 | ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
2825 | 15 | ElemPtrPtr, ElemTy->getPointerTo()); |
2826 | 15 | Address ElemPtr = |
2827 | 15 | Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType())); |
2828 | 15 | const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); |
2829 | | // Global = Buffer.VD[Idx]; |
2830 | 15 | const FieldDecl *FD = VarFieldMap.lookup(VD); |
2831 | 15 | LValue GlobLVal = CGF.EmitLValueForField( |
2832 | 15 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
2833 | 15 | Address GlobAddr = GlobLVal.getAddress(CGF); |
2834 | 15 | llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(), |
2835 | 15 | GlobAddr.getPointer(), Idxs); |
2836 | 15 | GlobLVal.setAddress(Address(BufferPtr, |
2837 | 15 | CGF.ConvertTypeForMem(Private->getType()), |
2838 | 15 | GlobAddr.getAlignment())); |
2839 | 15 | switch (CGF.getEvaluationKind(Private->getType())) { |
2840 | 15 | case TEK_Scalar: { |
2841 | 15 | llvm::Value *V = CGF.EmitLoadOfScalar(GlobLVal, Loc); |
2842 | 15 | CGF.EmitStoreOfScalar(V, ElemPtr, /*Volatile=*/false, Private->getType(), |
2843 | 15 | LValueBaseInfo(AlignmentSource::Type), |
2844 | 15 | TBAAAccessInfo()); |
2845 | 15 | break; |
2846 | 0 | } |
2847 | 0 | case TEK_Complex: { |
2848 | 0 | CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc); |
2849 | 0 | CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()), |
2850 | 0 | /*isInit=*/false); |
2851 | 0 | break; |
2852 | 0 | } |
2853 | 0 | case TEK_Aggregate: |
2854 | 0 | CGF.EmitAggregateCopy(CGF.MakeAddrLValue(ElemPtr, Private->getType()), |
2855 | 0 | GlobLVal, Private->getType(), |
2856 | 0 | AggValueSlot::DoesNotOverlap); |
2857 | 0 | break; |
2858 | 15 | } |
2859 | 15 | ++Idx; |
2860 | 15 | } |
2861 | | |
2862 | 9 | CGF.FinishFunction(); |
2863 | 9 | return Fn; |
2864 | 9 | } |
2865 | | |
2866 | | /// This function emits a helper that reduces all the reduction variables from |
2867 | | /// the team into the provided global buffer for the reduction variables. |
2868 | | /// |
2869 | | /// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data) |
2870 | | /// void *GlobPtrs[]; |
2871 | | /// GlobPtrs[0] = (void*)&buffer.D0[Idx]; |
2872 | | /// ... |
2873 | | /// GlobPtrs[N] = (void*)&buffer.DN[Idx]; |
2874 | | /// reduce_function(reduce_data, GlobPtrs); |
2875 | | static llvm::Value *emitGlobalToListReduceFunction( |
2876 | | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
2877 | | QualType ReductionArrayTy, SourceLocation Loc, |
2878 | | const RecordDecl *TeamReductionRec, |
2879 | | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
2880 | | &VarFieldMap, |
2881 | 9 | llvm::Function *ReduceFn) { |
2882 | 9 | ASTContext &C = CGM.getContext(); |
2883 | | |
2884 | | // Buffer: global reduction buffer. |
2885 | 9 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2886 | 9 | C.VoidPtrTy, ImplicitParamDecl::Other); |
2887 | | // Idx: index of the buffer. |
2888 | 9 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
2889 | 9 | ImplicitParamDecl::Other); |
2890 | | // ReduceList: thread local Reduce list. |
2891 | 9 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
2892 | 9 | C.VoidPtrTy, ImplicitParamDecl::Other); |
2893 | 9 | FunctionArgList Args; |
2894 | 9 | Args.push_back(&BufferArg); |
2895 | 9 | Args.push_back(&IdxArg); |
2896 | 9 | Args.push_back(&ReduceListArg); |
2897 | | |
2898 | 9 | const CGFunctionInfo &CGFI = |
2899 | 9 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
2900 | 9 | auto *Fn = llvm::Function::Create( |
2901 | 9 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
2902 | 9 | "_omp_reduction_global_to_list_reduce_func", &CGM.getModule()); |
2903 | 9 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
2904 | 9 | Fn->setDoesNotRecurse(); |
2905 | 9 | CodeGenFunction CGF(CGM); |
2906 | 9 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
2907 | | |
2908 | 9 | CGBuilderTy &Bld = CGF.Builder; |
2909 | | |
2910 | 9 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
2911 | 9 | QualType StaticTy = C.getRecordType(TeamReductionRec); |
2912 | 9 | llvm::Type *LLVMReductionsBufferTy = |
2913 | 9 | CGM.getTypes().ConvertTypeForMem(StaticTy); |
2914 | 9 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
2915 | 9 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
2916 | 9 | LLVMReductionsBufferTy->getPointerTo()); |
2917 | | |
2918 | | // 1. Build a list of reduction variables. |
2919 | | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
2920 | 9 | Address ReductionList = |
2921 | 9 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); |
2922 | 9 | auto IPriv = Privates.begin(); |
2923 | 9 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
2924 | 9 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
2925 | 9 | /*Volatile=*/false, C.IntTy, |
2926 | 9 | Loc)}; |
2927 | 9 | unsigned Idx = 0; |
2928 | 24 | for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx15 ) { |
2929 | 15 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
2930 | | // Global = Buffer.VD[Idx]; |
2931 | 15 | const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); |
2932 | 15 | const FieldDecl *FD = VarFieldMap.lookup(VD); |
2933 | 15 | LValue GlobLVal = CGF.EmitLValueForField( |
2934 | 15 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
2935 | 15 | Address GlobAddr = GlobLVal.getAddress(CGF); |
2936 | 15 | llvm::Value *BufferPtr = Bld.CreateInBoundsGEP( |
2937 | 15 | GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs); |
2938 | 15 | llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr); |
2939 | 15 | CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy); |
2940 | 15 | if ((*IPriv)->getType()->isVariablyModifiedType()) { |
2941 | | // Store array size. |
2942 | 0 | ++Idx; |
2943 | 0 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
2944 | 0 | llvm::Value *Size = CGF.Builder.CreateIntCast( |
2945 | 0 | CGF.getVLASize( |
2946 | 0 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) |
2947 | 0 | .NumElts, |
2948 | 0 | CGF.SizeTy, /*isSigned=*/false); |
2949 | 0 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), |
2950 | 0 | Elem); |
2951 | 0 | } |
2952 | 15 | } |
2953 | | |
2954 | | // Call reduce_function(ReduceList, GlobalReduceList) |
2955 | 9 | llvm::Value *GlobalReduceList = |
2956 | 9 | CGF.EmitCastToVoidPtr(ReductionList.getPointer()); |
2957 | 9 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
2958 | 9 | llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( |
2959 | 9 | AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); |
2960 | 9 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
2961 | 9 | CGF, Loc, ReduceFn, {ReducedPtr, GlobalReduceList}); |
2962 | 9 | CGF.FinishFunction(); |
2963 | 9 | return Fn; |
2964 | 9 | } |
2965 | | |
2966 | | /// |
2967 | | /// Design of OpenMP reductions on the GPU |
2968 | | /// |
2969 | | /// Consider a typical OpenMP program with one or more reduction |
2970 | | /// clauses: |
2971 | | /// |
2972 | | /// float foo; |
2973 | | /// double bar; |
2974 | | /// #pragma omp target teams distribute parallel for \ |
2975 | | /// reduction(+:foo) reduction(*:bar) |
2976 | | /// for (int i = 0; i < N; i++) { |
2977 | | /// foo += A[i]; bar *= B[i]; |
2978 | | /// } |
2979 | | /// |
2980 | | /// where 'foo' and 'bar' are reduced across all OpenMP threads in |
2981 | | /// all teams. In our OpenMP implementation on the NVPTX device an |
2982 | | /// OpenMP team is mapped to a CUDA threadblock and OpenMP threads |
2983 | | /// within a team are mapped to CUDA threads within a threadblock. |
2984 | | /// Our goal is to efficiently aggregate values across all OpenMP |
2985 | | /// threads such that: |
2986 | | /// |
2987 | | /// - the compiler and runtime are logically concise, and |
2988 | | /// - the reduction is performed efficiently in a hierarchical |
2989 | | /// manner as follows: within OpenMP threads in the same warp, |
2990 | | /// across warps in a threadblock, and finally across teams on |
2991 | | /// the NVPTX device. |
2992 | | /// |
2993 | | /// Introduction to Decoupling |
2994 | | /// |
2995 | | /// We would like to decouple the compiler and the runtime so that the |
2996 | | /// latter is ignorant of the reduction variables (number, data types) |
2997 | | /// and the reduction operators. This allows a simpler interface |
2998 | | /// and implementation while still attaining good performance. |
2999 | | /// |
3000 | | /// Pseudocode for the aforementioned OpenMP program generated by the |
3001 | | /// compiler is as follows: |
3002 | | /// |
3003 | | /// 1. Create private copies of reduction variables on each OpenMP |
3004 | | /// thread: 'foo_private', 'bar_private' |
3005 | | /// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned |
3006 | | /// to it and writes the result in 'foo_private' and 'bar_private' |
3007 | | /// respectively. |
3008 | | /// 3. Call the OpenMP runtime on the GPU to reduce within a team |
3009 | | /// and store the result on the team master: |
3010 | | /// |
3011 | | /// __kmpc_nvptx_parallel_reduce_nowait_v2(..., |
3012 | | /// reduceData, shuffleReduceFn, interWarpCpyFn) |
3013 | | /// |
3014 | | /// where: |
3015 | | /// struct ReduceData { |
3016 | | /// double *foo; |
3017 | | /// double *bar; |
3018 | | /// } reduceData |
3019 | | /// reduceData.foo = &foo_private |
3020 | | /// reduceData.bar = &bar_private |
3021 | | /// |
3022 | | /// 'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two |
3023 | | /// auxiliary functions generated by the compiler that operate on |
3024 | | /// variables of type 'ReduceData'. They aid the runtime perform |
3025 | | /// algorithmic steps in a data agnostic manner. |
3026 | | /// |
3027 | | /// 'shuffleReduceFn' is a pointer to a function that reduces data |
3028 | | /// of type 'ReduceData' across two OpenMP threads (lanes) in the |
3029 | | /// same warp. It takes the following arguments as input: |
3030 | | /// |
3031 | | /// a. variable of type 'ReduceData' on the calling lane, |
3032 | | /// b. its lane_id, |
3033 | | /// c. an offset relative to the current lane_id to generate a |
3034 | | /// remote_lane_id. The remote lane contains the second |
3035 | | /// variable of type 'ReduceData' that is to be reduced. |
3036 | | /// d. an algorithm version parameter determining which reduction |
3037 | | /// algorithm to use. |
3038 | | /// |
3039 | | /// 'shuffleReduceFn' retrieves data from the remote lane using |
3040 | | /// efficient GPU shuffle intrinsics and reduces, using the |
3041 | | /// algorithm specified by the 4th parameter, the two operands |
3042 | | /// element-wise. The result is written to the first operand. |
3043 | | /// |
3044 | | /// Different reduction algorithms are implemented in different |
3045 | | /// runtime functions, all calling 'shuffleReduceFn' to perform |
3046 | | /// the essential reduction step. Therefore, based on the 4th |
3047 | | /// parameter, this function behaves slightly differently to |
3048 | | /// cooperate with the runtime to ensure correctness under |
3049 | | /// different circumstances. |
3050 | | /// |
3051 | | /// 'InterWarpCpyFn' is a pointer to a function that transfers |
3052 | | /// reduced variables across warps. It tunnels, through CUDA |
3053 | | /// shared memory, the thread-private data of type 'ReduceData' |
3054 | | /// from lane 0 of each warp to a lane in the first warp. |
3055 | | /// 4. Call the OpenMP runtime on the GPU to reduce across teams. |
3056 | | /// The last team writes the global reduced value to memory. |
3057 | | /// |
3058 | | /// ret = __kmpc_nvptx_teams_reduce_nowait(..., |
3059 | | /// reduceData, shuffleReduceFn, interWarpCpyFn, |
3060 | | /// scratchpadCopyFn, loadAndReduceFn) |
3061 | | /// |
3062 | | /// 'scratchpadCopyFn' is a helper that stores reduced |
3063 | | /// data from the team master to a scratchpad array in |
3064 | | /// global memory. |
3065 | | /// |
3066 | | /// 'loadAndReduceFn' is a helper that loads data from |
3067 | | /// the scratchpad array and reduces it with the input |
3068 | | /// operand. |
3069 | | /// |
3070 | | /// These compiler generated functions hide address |
3071 | | /// calculation and alignment information from the runtime. |
3072 | | /// 5. if ret == 1: |
3073 | | /// The team master of the last team stores the reduced |
3074 | | /// result to the globals in memory. |
3075 | | /// foo += reduceData.foo; bar *= reduceData.bar |
3076 | | /// |
3077 | | /// |
3078 | | /// Warp Reduction Algorithms |
3079 | | /// |
3080 | | /// On the warp level, we have three algorithms implemented in the |
3081 | | /// OpenMP runtime depending on the number of active lanes: |
3082 | | /// |
3083 | | /// Full Warp Reduction |
3084 | | /// |
3085 | | /// The reduce algorithm within a warp where all lanes are active |
3086 | | /// is implemented in the runtime as follows: |
3087 | | /// |
3088 | | /// full_warp_reduce(void *reduce_data, |
3089 | | /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { |
3090 | | /// for (int offset = WARPSIZE/2; offset > 0; offset /= 2) |
3091 | | /// ShuffleReduceFn(reduce_data, 0, offset, 0); |
3092 | | /// } |
3093 | | /// |
3094 | | /// The algorithm completes in log(2, WARPSIZE) steps. |
3095 | | /// |
3096 | | /// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is |
3097 | | /// not used therefore we save instructions by not retrieving lane_id |
3098 | | /// from the corresponding special registers. The 4th parameter, which |
3099 | | /// represents the version of the algorithm being used, is set to 0 to |
3100 | | /// signify full warp reduction. |
3101 | | /// |
3102 | | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
3103 | | /// |
3104 | | /// #reduce_elem refers to an element in the local lane's data structure |
3105 | | /// #remote_elem is retrieved from a remote lane |
3106 | | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
3107 | | /// reduce_elem = reduce_elem REDUCE_OP remote_elem; |
3108 | | /// |
3109 | | /// Contiguous Partial Warp Reduction |
3110 | | /// |
3111 | | /// This reduce algorithm is used within a warp where only the first |
3112 | | /// 'n' (n <= WARPSIZE) lanes are active. It is typically used when the |
3113 | | /// number of OpenMP threads in a parallel region is not a multiple of |
3114 | | /// WARPSIZE. The algorithm is implemented in the runtime as follows: |
3115 | | /// |
3116 | | /// void |
3117 | | /// contiguous_partial_reduce(void *reduce_data, |
3118 | | /// kmp_ShuffleReductFctPtr ShuffleReduceFn, |
3119 | | /// int size, int lane_id) { |
3120 | | /// int curr_size; |
3121 | | /// int offset; |
3122 | | /// curr_size = size; |
3123 | | /// mask = curr_size/2; |
3124 | | /// while (offset>0) { |
3125 | | /// ShuffleReduceFn(reduce_data, lane_id, offset, 1); |
3126 | | /// curr_size = (curr_size+1)/2; |
3127 | | /// offset = curr_size/2; |
3128 | | /// } |
3129 | | /// } |
3130 | | /// |
3131 | | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
3132 | | /// |
3133 | | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
3134 | | /// if (lane_id < offset) |
3135 | | /// reduce_elem = reduce_elem REDUCE_OP remote_elem |
3136 | | /// else |
3137 | | /// reduce_elem = remote_elem |
3138 | | /// |
3139 | | /// This algorithm assumes that the data to be reduced are located in a |
3140 | | /// contiguous subset of lanes starting from the first. When there is |
3141 | | /// an odd number of active lanes, the data in the last lane is not |
3142 | | /// aggregated with any other lane's dat but is instead copied over. |
3143 | | /// |
3144 | | /// Dispersed Partial Warp Reduction |
3145 | | /// |
3146 | | /// This algorithm is used within a warp when any discontiguous subset of |
3147 | | /// lanes are active. It is used to implement the reduction operation |
3148 | | /// across lanes in an OpenMP simd region or in a nested parallel region. |
3149 | | /// |
3150 | | /// void |
3151 | | /// dispersed_partial_reduce(void *reduce_data, |
3152 | | /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { |
3153 | | /// int size, remote_id; |
3154 | | /// int logical_lane_id = number_of_active_lanes_before_me() * 2; |
3155 | | /// do { |
3156 | | /// remote_id = next_active_lane_id_right_after_me(); |
3157 | | /// # the above function returns 0 of no active lane |
3158 | | /// # is present right after the current lane. |
3159 | | /// size = number_of_active_lanes_in_this_warp(); |
3160 | | /// logical_lane_id /= 2; |
3161 | | /// ShuffleReduceFn(reduce_data, logical_lane_id, |
3162 | | /// remote_id-1-threadIdx.x, 2); |
3163 | | /// } while (logical_lane_id % 2 == 0 && size > 1); |
3164 | | /// } |
3165 | | /// |
3166 | | /// There is no assumption made about the initial state of the reduction. |
3167 | | /// Any number of lanes (>=1) could be active at any position. The reduction |
3168 | | /// result is returned in the first active lane. |
3169 | | /// |
3170 | | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
3171 | | /// |
3172 | | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
3173 | | /// if (lane_id % 2 == 0 && offset > 0) |
3174 | | /// reduce_elem = reduce_elem REDUCE_OP remote_elem |
3175 | | /// else |
3176 | | /// reduce_elem = remote_elem |
3177 | | /// |
3178 | | /// |
3179 | | /// Intra-Team Reduction |
3180 | | /// |
3181 | | /// This function, as implemented in the runtime call |
3182 | | /// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP |
3183 | | /// threads in a team. It first reduces within a warp using the |
3184 | | /// aforementioned algorithms. We then proceed to gather all such |
3185 | | /// reduced values at the first warp. |
3186 | | /// |
3187 | | /// The runtime makes use of the function 'InterWarpCpyFn', which copies |
3188 | | /// data from each of the "warp master" (zeroth lane of each warp, where |
3189 | | /// warp-reduced data is held) to the zeroth warp. This step reduces (in |
3190 | | /// a mathematical sense) the problem of reduction across warp masters in |
3191 | | /// a block to the problem of warp reduction. |
3192 | | /// |
3193 | | /// |
3194 | | /// Inter-Team Reduction |
3195 | | /// |
3196 | | /// Once a team has reduced its data to a single value, it is stored in |
3197 | | /// a global scratchpad array. Since each team has a distinct slot, this |
3198 | | /// can be done without locking. |
3199 | | /// |
3200 | | /// The last team to write to the scratchpad array proceeds to reduce the |
3201 | | /// scratchpad array. One or more workers in the last team use the helper |
3202 | | /// 'loadAndReduceDataFn' to load and reduce values from the array, i.e., |
3203 | | /// the k'th worker reduces every k'th element. |
3204 | | /// |
3205 | | /// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to |
3206 | | /// reduce across workers and compute a globally reduced value. |
3207 | | /// |
3208 | | void CGOpenMPRuntimeGPU::emitReduction( |
3209 | | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates, |
3210 | | ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, |
3211 | 34 | ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) { |
3212 | 34 | if (!CGF.HaveInsertPoint()) |
3213 | 0 | return; |
3214 | | |
3215 | 34 | bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind); |
3216 | 34 | #ifndef NDEBUG |
3217 | 34 | bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind); |
3218 | 34 | #endif |
3219 | | |
3220 | 34 | if (Options.SimpleReduction) { |
3221 | 6 | assert(!TeamsReduction && !ParallelReduction && |
3222 | 6 | "Invalid reduction selection in emitReduction."); |
3223 | 0 | CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs, |
3224 | 6 | ReductionOps, Options); |
3225 | 6 | return; |
3226 | 6 | } |
3227 | | |
3228 | 28 | assert((TeamsReduction || ParallelReduction) && |
3229 | 28 | "Invalid reduction selection in emitReduction."); |
3230 | | |
3231 | | // Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList), |
3232 | | // RedList, shuffle_reduce_func, interwarp_copy_func); |
3233 | | // or |
3234 | | // Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>); |
3235 | 0 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); |
3236 | 28 | llvm::Value *ThreadId = getThreadID(CGF, Loc); |
3237 | | |
3238 | 28 | llvm::Value *Res; |
3239 | 28 | ASTContext &C = CGM.getContext(); |
3240 | | // 1. Build a list of reduction variables. |
3241 | | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
3242 | 28 | auto Size = RHSExprs.size(); |
3243 | 43 | for (const Expr *E : Privates) { |
3244 | 43 | if (E->getType()->isVariablyModifiedType()) |
3245 | | // Reserve place for array size. |
3246 | 0 | ++Size; |
3247 | 43 | } |
3248 | 28 | llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size); |
3249 | 28 | QualType ReductionArrayTy = |
3250 | 28 | C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal, |
3251 | 28 | /*IndexTypeQuals=*/0); |
3252 | 28 | Address ReductionList = |
3253 | 28 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); |
3254 | 28 | auto IPriv = Privates.begin(); |
3255 | 28 | unsigned Idx = 0; |
3256 | 71 | for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx43 ) { |
3257 | 43 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
3258 | 43 | CGF.Builder.CreateStore( |
3259 | 43 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
3260 | 43 | CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy), |
3261 | 43 | Elem); |
3262 | 43 | if ((*IPriv)->getType()->isVariablyModifiedType()) { |
3263 | | // Store array size. |
3264 | 0 | ++Idx; |
3265 | 0 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
3266 | 0 | llvm::Value *Size = CGF.Builder.CreateIntCast( |
3267 | 0 | CGF.getVLASize( |
3268 | 0 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) |
3269 | 0 | .NumElts, |
3270 | 0 | CGF.SizeTy, /*isSigned=*/false); |
3271 | 0 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), |
3272 | 0 | Elem); |
3273 | 0 | } |
3274 | 43 | } |
3275 | | |
3276 | 28 | llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
3277 | 28 | ReductionList.getPointer(), CGF.VoidPtrTy); |
3278 | 28 | llvm::Function *ReductionFn = |
3279 | 28 | emitReductionFunction(Loc, CGF.ConvertTypeForMem(ReductionArrayTy), |
3280 | 28 | Privates, LHSExprs, RHSExprs, ReductionOps); |
3281 | 28 | llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy); |
3282 | 28 | llvm::Function *ShuffleAndReduceFn = emitShuffleAndReduceFunction( |
3283 | 28 | CGM, Privates, ReductionArrayTy, ReductionFn, Loc); |
3284 | 28 | llvm::Value *InterWarpCopyFn = |
3285 | 28 | emitInterWarpCopyFunction(CGM, Privates, ReductionArrayTy, Loc); |
3286 | | |
3287 | 28 | if (ParallelReduction) { |
3288 | 19 | llvm::Value *Args[] = {RTLoc, |
3289 | 19 | ThreadId, |
3290 | 19 | CGF.Builder.getInt32(RHSExprs.size()), |
3291 | 19 | ReductionArrayTySize, |
3292 | 19 | RL, |
3293 | 19 | ShuffleAndReduceFn, |
3294 | 19 | InterWarpCopyFn}; |
3295 | | |
3296 | 19 | Res = CGF.EmitRuntimeCall( |
3297 | 19 | OMPBuilder.getOrCreateRuntimeFunction( |
3298 | 19 | CGM.getModule(), OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2), |
3299 | 19 | Args); |
3300 | 19 | } else { |
3301 | 9 | assert(TeamsReduction && "expected teams reduction."); |
3302 | 0 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap; |
3303 | 9 | llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size()); |
3304 | 9 | int Cnt = 0; |
3305 | 15 | for (const Expr *DRE : Privates) { |
3306 | 15 | PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl(); |
3307 | 15 | ++Cnt; |
3308 | 15 | } |
3309 | 9 | const RecordDecl *TeamReductionRec = ::buildRecordForGlobalizedVars( |
3310 | 9 | CGM.getContext(), PrivatesReductions, llvm::None, VarFieldMap, |
3311 | 9 | C.getLangOpts().OpenMPCUDAReductionBufNum); |
3312 | 9 | TeamsReductions.push_back(TeamReductionRec); |
3313 | 9 | if (!KernelTeamsReductionPtr) { |
3314 | 3 | KernelTeamsReductionPtr = new llvm::GlobalVariable( |
3315 | 3 | CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/true, |
3316 | 3 | llvm::GlobalValue::InternalLinkage, nullptr, |
3317 | 3 | "_openmp_teams_reductions_buffer_$_$ptr"); |
3318 | 3 | } |
3319 | 9 | llvm::Value *GlobalBufferPtr = CGF.EmitLoadOfScalar( |
3320 | 9 | Address(KernelTeamsReductionPtr, CGF.VoidPtrTy, CGM.getPointerAlign()), |
3321 | 9 | /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); |
3322 | 9 | llvm::Value *GlobalToBufferCpyFn = ::emitListToGlobalCopyFunction( |
3323 | 9 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap); |
3324 | 9 | llvm::Value *GlobalToBufferRedFn = ::emitListToGlobalReduceFunction( |
3325 | 9 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap, |
3326 | 9 | ReductionFn); |
3327 | 9 | llvm::Value *BufferToGlobalCpyFn = ::emitGlobalToListCopyFunction( |
3328 | 9 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap); |
3329 | 9 | llvm::Value *BufferToGlobalRedFn = ::emitGlobalToListReduceFunction( |
3330 | 9 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap, |
3331 | 9 | ReductionFn); |
3332 | | |
3333 | 9 | llvm::Value *Args[] = { |
3334 | 9 | RTLoc, |
3335 | 9 | ThreadId, |
3336 | 9 | GlobalBufferPtr, |
3337 | 9 | CGF.Builder.getInt32(C.getLangOpts().OpenMPCUDAReductionBufNum), |
3338 | 9 | RL, |
3339 | 9 | ShuffleAndReduceFn, |
3340 | 9 | InterWarpCopyFn, |
3341 | 9 | GlobalToBufferCpyFn, |
3342 | 9 | GlobalToBufferRedFn, |
3343 | 9 | BufferToGlobalCpyFn, |
3344 | 9 | BufferToGlobalRedFn}; |
3345 | | |
3346 | 9 | Res = CGF.EmitRuntimeCall( |
3347 | 9 | OMPBuilder.getOrCreateRuntimeFunction( |
3348 | 9 | CGM.getModule(), OMPRTL___kmpc_nvptx_teams_reduce_nowait_v2), |
3349 | 9 | Args); |
3350 | 9 | } |
3351 | | |
3352 | | // 5. Build if (res == 1) |
3353 | 0 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.reduction.done"); |
3354 | 28 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.then"); |
3355 | 28 | llvm::Value *Cond = CGF.Builder.CreateICmpEQ( |
3356 | 28 | Res, llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1)); |
3357 | 28 | CGF.Builder.CreateCondBr(Cond, ThenBB, ExitBB); |
3358 | | |
3359 | | // 6. Build then branch: where we have reduced values in the master |
3360 | | // thread in each team. |
3361 | | // __kmpc_end_reduce{_nowait}(<gtid>); |
3362 | | // break; |
3363 | 28 | CGF.EmitBlock(ThenBB); |
3364 | | |
3365 | | // Add emission of __kmpc_end_reduce{_nowait}(<gtid>); |
3366 | 28 | auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps, |
3367 | 28 | this](CodeGenFunction &CGF, PrePostActionTy &Action) { |
3368 | 28 | auto IPriv = Privates.begin(); |
3369 | 28 | auto ILHS = LHSExprs.begin(); |
3370 | 28 | auto IRHS = RHSExprs.begin(); |
3371 | 43 | for (const Expr *E : ReductionOps) { |
3372 | 43 | emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), |
3373 | 43 | cast<DeclRefExpr>(*IRHS)); |
3374 | 43 | ++IPriv; |
3375 | 43 | ++ILHS; |
3376 | 43 | ++IRHS; |
3377 | 43 | } |
3378 | 28 | }; |
3379 | 28 | llvm::Value *EndArgs[] = {ThreadId}; |
3380 | 28 | RegionCodeGenTy RCG(CodeGen); |
3381 | 28 | NVPTXActionTy Action( |
3382 | 28 | nullptr, llvm::None, |
3383 | 28 | OMPBuilder.getOrCreateRuntimeFunction( |
3384 | 28 | CGM.getModule(), OMPRTL___kmpc_nvptx_end_reduce_nowait), |
3385 | 28 | EndArgs); |
3386 | 28 | RCG.setAction(Action); |
3387 | 28 | RCG(CGF); |
3388 | | // There is no need to emit line number for unconditional branch. |
3389 | 28 | (void)ApplyDebugLocation::CreateEmpty(CGF); |
3390 | 28 | CGF.EmitBlock(ExitBB, /*IsFinished=*/true); |
3391 | 28 | } |
3392 | | |
3393 | | const VarDecl * |
3394 | | CGOpenMPRuntimeGPU::translateParameter(const FieldDecl *FD, |
3395 | 83 | const VarDecl *NativeParam) const { |
3396 | 83 | if (!NativeParam->getType()->isReferenceType()) |
3397 | 25 | return NativeParam; |
3398 | 58 | QualType ArgType = NativeParam->getType(); |
3399 | 58 | QualifierCollector QC; |
3400 | 58 | const Type *NonQualTy = QC.strip(ArgType); |
3401 | 58 | QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); |
3402 | 58 | if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) { |
3403 | 58 | if (Attr->getCaptureKind() == OMPC_map) { |
3404 | 38 | PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy, |
3405 | 38 | LangAS::opencl_global); |
3406 | 38 | } |
3407 | 58 | } |
3408 | 58 | ArgType = CGM.getContext().getPointerType(PointeeTy); |
3409 | 58 | QC.addRestrict(); |
3410 | 58 | enum { NVPTX_local_addr = 5 }; |
3411 | 58 | QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr)); |
3412 | 58 | ArgType = QC.apply(CGM.getContext(), ArgType); |
3413 | 58 | if (isa<ImplicitParamDecl>(NativeParam)) |
3414 | 0 | return ImplicitParamDecl::Create( |
3415 | 0 | CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(), |
3416 | 0 | NativeParam->getIdentifier(), ArgType, ImplicitParamDecl::Other); |
3417 | 58 | return ParmVarDecl::Create( |
3418 | 58 | CGM.getContext(), |
3419 | 58 | const_cast<DeclContext *>(NativeParam->getDeclContext()), |
3420 | 58 | NativeParam->getBeginLoc(), NativeParam->getLocation(), |
3421 | 58 | NativeParam->getIdentifier(), ArgType, |
3422 | 58 | /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr); |
3423 | 58 | } |
3424 | | |
3425 | | Address |
3426 | | CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF, |
3427 | | const VarDecl *NativeParam, |
3428 | 58 | const VarDecl *TargetParam) const { |
3429 | 58 | assert(NativeParam != TargetParam && |
3430 | 58 | NativeParam->getType()->isReferenceType() && |
3431 | 58 | "Native arg must not be the same as target arg."); |
3432 | 0 | Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam); |
3433 | 58 | QualType NativeParamType = NativeParam->getType(); |
3434 | 58 | QualifierCollector QC; |
3435 | 58 | const Type *NonQualTy = QC.strip(NativeParamType); |
3436 | 58 | QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); |
3437 | 58 | unsigned NativePointeeAddrSpace = |
3438 | 58 | CGF.getContext().getTargetAddressSpace(NativePointeeTy); |
3439 | 58 | QualType TargetTy = TargetParam->getType(); |
3440 | 58 | llvm::Value *TargetAddr = CGF.EmitLoadOfScalar( |
3441 | 58 | LocalAddr, /*Volatile=*/false, TargetTy, SourceLocation()); |
3442 | | // First cast to generic. |
3443 | 58 | TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
3444 | 58 | TargetAddr, llvm::PointerType::getWithSamePointeeType( |
3445 | 58 | cast<llvm::PointerType>(TargetAddr->getType()), /*AddrSpace=*/0)); |
3446 | | // Cast from generic to native address space. |
3447 | 58 | TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
3448 | 58 | TargetAddr, llvm::PointerType::getWithSamePointeeType( |
3449 | 58 | cast<llvm::PointerType>(TargetAddr->getType()), |
3450 | 58 | NativePointeeAddrSpace)); |
3451 | 58 | Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType); |
3452 | 58 | CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false, |
3453 | 58 | NativeParamType); |
3454 | 58 | return NativeParamAddr; |
3455 | 58 | } |
3456 | | |
3457 | | void CGOpenMPRuntimeGPU::emitOutlinedFunctionCall( |
3458 | | CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn, |
3459 | 460 | ArrayRef<llvm::Value *> Args) const { |
3460 | 460 | SmallVector<llvm::Value *, 4> TargetArgs; |
3461 | 460 | TargetArgs.reserve(Args.size()); |
3462 | 460 | auto *FnType = OutlinedFn.getFunctionType(); |
3463 | 1.66k | for (unsigned I = 0, E = Args.size(); I < E; ++I1.20k ) { |
3464 | 1.20k | if (FnType->isVarArg() && FnType->getNumParams() <= I0 ) { |
3465 | 0 | TargetArgs.append(std::next(Args.begin(), I), Args.end()); |
3466 | 0 | break; |
3467 | 0 | } |
3468 | 1.20k | llvm::Type *TargetType = FnType->getParamType(I); |
3469 | 1.20k | llvm::Value *NativeArg = Args[I]; |
3470 | 1.20k | if (!TargetType->isPointerTy()) { |
3471 | 121 | TargetArgs.emplace_back(NativeArg); |
3472 | 121 | continue; |
3473 | 121 | } |
3474 | 1.08k | llvm::Value *TargetArg = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
3475 | 1.08k | NativeArg, llvm::PointerType::getWithSamePointeeType( |
3476 | 1.08k | cast<llvm::PointerType>(NativeArg->getType()), /*AddrSpace*/ 0)); |
3477 | 1.08k | TargetArgs.emplace_back( |
3478 | 1.08k | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TargetArg, TargetType)); |
3479 | 1.08k | } |
3480 | 460 | CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs); |
3481 | 460 | } |
3482 | | |
3483 | | /// Emit function which wraps the outline parallel region |
3484 | | /// and controls the arguments which are passed to this function. |
3485 | | /// The wrapper ensures that the outlined function is called |
3486 | | /// with the correct arguments when data is shared. |
3487 | | llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper( |
3488 | 57 | llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) { |
3489 | 57 | ASTContext &Ctx = CGM.getContext(); |
3490 | 57 | const auto &CS = *D.getCapturedStmt(OMPD_parallel); |
3491 | | |
3492 | | // Create a function that takes as argument the source thread. |
3493 | 57 | FunctionArgList WrapperArgs; |
3494 | 57 | QualType Int16QTy = |
3495 | 57 | Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false); |
3496 | 57 | QualType Int32QTy = |
3497 | 57 | Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false); |
3498 | 57 | ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), |
3499 | 57 | /*Id=*/nullptr, Int16QTy, |
3500 | 57 | ImplicitParamDecl::Other); |
3501 | 57 | ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), |
3502 | 57 | /*Id=*/nullptr, Int32QTy, |
3503 | 57 | ImplicitParamDecl::Other); |
3504 | 57 | WrapperArgs.emplace_back(&ParallelLevelArg); |
3505 | 57 | WrapperArgs.emplace_back(&WrapperArg); |
3506 | | |
3507 | 57 | const CGFunctionInfo &CGFI = |
3508 | 57 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs); |
3509 | | |
3510 | 57 | auto *Fn = llvm::Function::Create( |
3511 | 57 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
3512 | 57 | Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule()); |
3513 | | |
3514 | | // Ensure we do not inline the function. This is trivially true for the ones |
3515 | | // passed to __kmpc_fork_call but the ones calles in serialized regions |
3516 | | // could be inlined. This is not a perfect but it is closer to the invariant |
3517 | | // we want, namely, every data environment starts with a new function. |
3518 | | // TODO: We should pass the if condition to the runtime function and do the |
3519 | | // handling there. Much cleaner code. |
3520 | 57 | Fn->addFnAttr(llvm::Attribute::NoInline); |
3521 | | |
3522 | 57 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
3523 | 57 | Fn->setLinkage(llvm::GlobalValue::InternalLinkage); |
3524 | 57 | Fn->setDoesNotRecurse(); |
3525 | | |
3526 | 57 | CodeGenFunction CGF(CGM, /*suppressNewContext=*/true); |
3527 | 57 | CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs, |
3528 | 57 | D.getBeginLoc(), D.getBeginLoc()); |
3529 | | |
3530 | 57 | const auto *RD = CS.getCapturedRecordDecl(); |
3531 | 57 | auto CurField = RD->field_begin(); |
3532 | | |
3533 | 57 | Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, |
3534 | 57 | /*Name=*/".zero.addr"); |
3535 | 57 | CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr); |
3536 | | // Get the array of arguments. |
3537 | 57 | SmallVector<llvm::Value *, 8> Args; |
3538 | | |
3539 | 57 | Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).getPointer()); |
3540 | 57 | Args.emplace_back(ZeroAddr.getPointer()); |
3541 | | |
3542 | 57 | CGBuilderTy &Bld = CGF.Builder; |
3543 | 57 | auto CI = CS.capture_begin(); |
3544 | | |
3545 | | // Use global memory for data sharing. |
3546 | | // Handle passing of global args to workers. |
3547 | 57 | Address GlobalArgs = |
3548 | 57 | CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args"); |
3549 | 57 | llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer(); |
3550 | 57 | llvm::Value *DataSharingArgs[] = {GlobalArgsPtr}; |
3551 | 57 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
3552 | 57 | CGM.getModule(), OMPRTL___kmpc_get_shared_variables), |
3553 | 57 | DataSharingArgs); |
3554 | | |
3555 | | // Retrieve the shared variables from the list of references returned |
3556 | | // by the runtime. Pass the variables to the outlined function. |
3557 | 57 | Address SharedArgListAddress = Address::invalid(); |
3558 | 57 | if (CS.capture_size() > 0 || |
3559 | 57 | isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())34 ) { |
3560 | 26 | SharedArgListAddress = CGF.EmitLoadOfPointer( |
3561 | 26 | GlobalArgs, CGF.getContext() |
3562 | 26 | .getPointerType(CGF.getContext().VoidPtrTy) |
3563 | 26 | .castAs<PointerType>()); |
3564 | 26 | } |
3565 | 57 | unsigned Idx = 0; |
3566 | 57 | if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) { |
3567 | 3 | Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); |
3568 | 3 | Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( |
3569 | 3 | Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy); |
3570 | 3 | llvm::Value *LB = CGF.EmitLoadOfScalar( |
3571 | 3 | TypedAddress, |
3572 | 3 | /*Volatile=*/false, |
3573 | 3 | CGF.getContext().getPointerType(CGF.getContext().getSizeType()), |
3574 | 3 | cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc()); |
3575 | 3 | Args.emplace_back(LB); |
3576 | 3 | ++Idx; |
3577 | 3 | Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); |
3578 | 3 | TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( |
3579 | 3 | Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy); |
3580 | 3 | llvm::Value *UB = CGF.EmitLoadOfScalar( |
3581 | 3 | TypedAddress, |
3582 | 3 | /*Volatile=*/false, |
3583 | 3 | CGF.getContext().getPointerType(CGF.getContext().getSizeType()), |
3584 | 3 | cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc()); |
3585 | 3 | Args.emplace_back(UB); |
3586 | 3 | ++Idx; |
3587 | 3 | } |
3588 | 57 | if (CS.capture_size() > 0) { |
3589 | 23 | ASTContext &CGFContext = CGF.getContext(); |
3590 | 63 | for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField40 ) { |
3591 | 40 | QualType ElemTy = CurField->getType(); |
3592 | 40 | Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx); |
3593 | 40 | Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( |
3594 | 40 | Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy)), |
3595 | 40 | CGF.ConvertTypeForMem(ElemTy)); |
3596 | 40 | llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress, |
3597 | 40 | /*Volatile=*/false, |
3598 | 40 | CGFContext.getPointerType(ElemTy), |
3599 | 40 | CI->getLocation()); |
3600 | 40 | if (CI->capturesVariableByCopy() && |
3601 | 40 | !CI->getCapturedVar()->getType()->isAnyPointerType()0 ) { |
3602 | 0 | Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(), |
3603 | 0 | CI->getLocation()); |
3604 | 0 | } |
3605 | 40 | Args.emplace_back(Arg); |
3606 | 40 | } |
3607 | 23 | } |
3608 | | |
3609 | 57 | emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args); |
3610 | 57 | CGF.FinishFunction(); |
3611 | 57 | return Fn; |
3612 | 57 | } |
3613 | | |
3614 | | void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF, |
3615 | 2.12k | const Decl *D) { |
3616 | 2.12k | if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic) |
3617 | 340 | return; |
3618 | | |
3619 | 1.78k | assert(D && "Expected function or captured|block decl."); |
3620 | 0 | assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && |
3621 | 1.78k | "Function is registered already."); |
3622 | 0 | assert((!TeamAndReductions.first || TeamAndReductions.first == D) && |
3623 | 1.78k | "Team is set but not processed."); |
3624 | 0 | const Stmt *Body = nullptr; |
3625 | 1.78k | bool NeedToDelayGlobalization = false; |
3626 | 1.78k | if (const auto *FD = dyn_cast<FunctionDecl>(D)) { |
3627 | 483 | Body = FD->getBody(); |
3628 | 1.29k | } else if (const auto *BD = dyn_cast<BlockDecl>(D)) { |
3629 | 0 | Body = BD->getBody(); |
3630 | 1.29k | } else if (const auto *CD = dyn_cast<CapturedDecl>(D)) { |
3631 | 1.29k | Body = CD->getBody(); |
3632 | 1.29k | NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP; |
3633 | 1.29k | if (NeedToDelayGlobalization && |
3634 | 1.29k | getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) |
3635 | 1.03k | return; |
3636 | 1.29k | } |
3637 | 745 | if (!Body) |
3638 | 0 | return; |
3639 | 745 | CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second); |
3640 | 745 | VarChecker.Visit(Body); |
3641 | 745 | const RecordDecl *GlobalizedVarsRecord = |
3642 | 745 | VarChecker.getGlobalizedRecord(IsInTTDRegion); |
3643 | 745 | TeamAndReductions.first = nullptr; |
3644 | 745 | TeamAndReductions.second.clear(); |
3645 | 745 | ArrayRef<const ValueDecl *> EscapedVariableLengthDecls = |
3646 | 745 | VarChecker.getEscapedVariableLengthDecls(); |
3647 | 745 | if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty()696 ) |
3648 | 696 | return; |
3649 | 49 | auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; |
3650 | 49 | I->getSecond().MappedParams = |
3651 | 49 | std::make_unique<CodeGenFunction::OMPMapVars>(); |
3652 | 49 | I->getSecond().EscapedParameters.insert( |
3653 | 49 | VarChecker.getEscapedParameters().begin(), |
3654 | 49 | VarChecker.getEscapedParameters().end()); |
3655 | 49 | I->getSecond().EscapedVariableLengthDecls.append( |
3656 | 49 | EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end()); |
3657 | 49 | DeclToAddrMapTy &Data = I->getSecond().LocalVarData; |
3658 | 71 | for (const ValueDecl *VD : VarChecker.getEscapedDecls()) { |
3659 | 71 | assert(VD->isCanonicalDecl() && "Expected canonical declaration"); |
3660 | 0 | Data.insert(std::make_pair(VD, MappedVarData())); |
3661 | 71 | } |
3662 | 49 | if (!IsInTTDRegion && !NeedToDelayGlobalization15 && !IsInParallelRegion13 ) { |
3663 | 13 | CheckVarsEscapingDeclContext VarChecker(CGF, llvm::None); |
3664 | 13 | VarChecker.Visit(Body); |
3665 | 13 | I->getSecond().SecondaryLocalVarData.emplace(); |
3666 | 13 | DeclToAddrMapTy &Data = *I->getSecond().SecondaryLocalVarData; |
3667 | 13 | for (const ValueDecl *VD : VarChecker.getEscapedDecls()) { |
3668 | 13 | assert(VD->isCanonicalDecl() && "Expected canonical declaration"); |
3669 | 0 | Data.insert(std::make_pair(VD, MappedVarData())); |
3670 | 13 | } |
3671 | 13 | } |
3672 | 49 | if (!NeedToDelayGlobalization) { |
3673 | 13 | emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true); |
3674 | 13 | struct GlobalizationScope final : EHScopeStack::Cleanup { |
3675 | 13 | GlobalizationScope() = default; |
3676 | | |
3677 | 13 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
3678 | 13 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) |
3679 | 13 | .emitGenericVarsEpilog(CGF, /*WithSPMDCheck=*/true); |
3680 | 13 | } |
3681 | 13 | }; |
3682 | 13 | CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup); |
3683 | 13 | } |
3684 | 49 | } |
3685 | | |
3686 | | Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF, |
3687 | 9.42k | const VarDecl *VD) { |
3688 | 9.42k | if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) { |
3689 | 6 | const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); |
3690 | 6 | auto AS = LangAS::Default; |
3691 | 6 | switch (A->getAllocatorType()) { |
3692 | | // Use the default allocator here as by default local vars are |
3693 | | // threadlocal. |
3694 | 0 | case OMPAllocateDeclAttr::OMPNullMemAlloc: |
3695 | 1 | case OMPAllocateDeclAttr::OMPDefaultMemAlloc: |
3696 | 2 | case OMPAllocateDeclAttr::OMPThreadMemAlloc: |
3697 | 2 | case OMPAllocateDeclAttr::OMPHighBWMemAlloc: |
3698 | 2 | case OMPAllocateDeclAttr::OMPLowLatMemAlloc: |
3699 | | // Follow the user decision - use default allocation. |
3700 | 2 | return Address::invalid(); |
3701 | 0 | case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: |
3702 | | // TODO: implement aupport for user-defined allocators. |
3703 | 0 | return Address::invalid(); |
3704 | 0 | case OMPAllocateDeclAttr::OMPConstMemAlloc: |
3705 | 0 | AS = LangAS::cuda_constant; |
3706 | 0 | break; |
3707 | 1 | case OMPAllocateDeclAttr::OMPPTeamMemAlloc: |
3708 | 1 | AS = LangAS::cuda_shared; |
3709 | 1 | break; |
3710 | 0 | case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: |
3711 | 3 | case OMPAllocateDeclAttr::OMPCGroupMemAlloc: |
3712 | 3 | break; |
3713 | 6 | } |
3714 | 4 | llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType()); |
3715 | 4 | auto *GV = new llvm::GlobalVariable( |
3716 | 4 | CGM.getModule(), VarTy, /*isConstant=*/false, |
3717 | 4 | llvm::GlobalValue::InternalLinkage, llvm::Constant::getNullValue(VarTy), |
3718 | 4 | VD->getName(), |
3719 | 4 | /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, |
3720 | 4 | CGM.getContext().getTargetAddressSpace(AS)); |
3721 | 4 | CharUnits Align = CGM.getContext().getDeclAlign(VD); |
3722 | 4 | GV->setAlignment(Align.getAsAlign()); |
3723 | 4 | return Address( |
3724 | 4 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
3725 | 4 | GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace( |
3726 | 4 | VD->getType().getAddressSpace()))), |
3727 | 4 | VarTy, Align); |
3728 | 6 | } |
3729 | | |
3730 | 9.42k | if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic) |
3731 | 2.13k | return Address::invalid(); |
3732 | | |
3733 | 7.28k | VD = VD->getCanonicalDecl(); |
3734 | 7.28k | auto I = FunctionGlobalizedDecls.find(CGF.CurFn); |
3735 | 7.28k | if (I == FunctionGlobalizedDecls.end()) |
3736 | 6.97k | return Address::invalid(); |
3737 | 305 | auto VDI = I->getSecond().LocalVarData.find(VD); |
3738 | 305 | if (VDI != I->getSecond().LocalVarData.end()) |
3739 | 58 | return VDI->second.PrivateAddr; |
3740 | 247 | if (VD->hasAttrs()) { |
3741 | 37 | for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()), |
3742 | 37 | E(VD->attr_end()); |
3743 | 61 | IT != E; ++IT24 ) { |
3744 | 37 | auto VDI = I->getSecond().LocalVarData.find( |
3745 | 37 | cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl()) |
3746 | 37 | ->getCanonicalDecl()); |
3747 | 37 | if (VDI != I->getSecond().LocalVarData.end()) |
3748 | 13 | return VDI->second.PrivateAddr; |
3749 | 37 | } |
3750 | 37 | } |
3751 | | |
3752 | 234 | return Address::invalid(); |
3753 | 247 | } |
3754 | | |
3755 | 2.36k | void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) { |
3756 | 2.36k | FunctionGlobalizedDecls.erase(CGF.CurFn); |
3757 | 2.36k | CGOpenMPRuntime::functionFinished(CGF); |
3758 | 2.36k | } |
3759 | | |
3760 | | void CGOpenMPRuntimeGPU::getDefaultDistScheduleAndChunk( |
3761 | | CodeGenFunction &CGF, const OMPLoopDirective &S, |
3762 | | OpenMPDistScheduleClauseKind &ScheduleKind, |
3763 | 278 | llvm::Value *&Chunk) const { |
3764 | 278 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
3765 | 278 | if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { |
3766 | 266 | ScheduleKind = OMPC_DIST_SCHEDULE_static; |
3767 | 266 | Chunk = CGF.EmitScalarConversion( |
3768 | 266 | RT.getGPUNumThreads(CGF), |
3769 | 266 | CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), |
3770 | 266 | S.getIterationVariable()->getType(), S.getBeginLoc()); |
3771 | 266 | return; |
3772 | 266 | } |
3773 | 12 | CGOpenMPRuntime::getDefaultDistScheduleAndChunk( |
3774 | 12 | CGF, S, ScheduleKind, Chunk); |
3775 | 12 | } |
3776 | | |
3777 | | void CGOpenMPRuntimeGPU::getDefaultScheduleAndChunk( |
3778 | | CodeGenFunction &CGF, const OMPLoopDirective &S, |
3779 | | OpenMPScheduleClauseKind &ScheduleKind, |
3780 | 102 | const Expr *&ChunkExpr) const { |
3781 | 102 | ScheduleKind = OMPC_SCHEDULE_static; |
3782 | | // Chunk size is 1 in this case. |
3783 | 102 | llvm::APInt ChunkSize(32, 1); |
3784 | 102 | ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize, |
3785 | 102 | CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), |
3786 | 102 | SourceLocation()); |
3787 | 102 | } |
3788 | | |
3789 | | void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas( |
3790 | 845 | CodeGenFunction &CGF, const OMPExecutableDirective &D) const { |
3791 | 845 | assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && |
3792 | 845 | " Expected target-based directive."); |
3793 | 0 | const CapturedStmt *CS = D.getCapturedStmt(OMPD_target); |
3794 | 845 | for (const CapturedStmt::Capture &C : CS->captures()) { |
3795 | | // Capture variables captured by reference in lambdas for target-based |
3796 | | // directives. |
3797 | 677 | if (!C.capturesVariable()) |
3798 | 340 | continue; |
3799 | 337 | const VarDecl *VD = C.getCapturedVar(); |
3800 | 337 | const auto *RD = VD->getType() |
3801 | 337 | .getCanonicalType() |
3802 | 337 | .getNonReferenceType() |
3803 | 337 | ->getAsCXXRecordDecl(); |
3804 | 337 | if (!RD || !RD->isLambda()19 ) |
3805 | 337 | continue; |
3806 | 0 | Address VDAddr = CGF.GetAddrOfLocalVar(VD); |
3807 | 0 | LValue VDLVal; |
3808 | 0 | if (VD->getType().getCanonicalType()->isReferenceType()) |
3809 | 0 | VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType()); |
3810 | 0 | else |
3811 | 0 | VDLVal = CGF.MakeAddrLValue( |
3812 | 0 | VDAddr, VD->getType().getCanonicalType().getNonReferenceType()); |
3813 | 0 | llvm::DenseMap<const VarDecl *, FieldDecl *> Captures; |
3814 | 0 | FieldDecl *ThisCapture = nullptr; |
3815 | 0 | RD->getCaptureFields(Captures, ThisCapture); |
3816 | 0 | if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) { |
3817 | 0 | LValue ThisLVal = |
3818 | 0 | CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture); |
3819 | 0 | llvm::Value *CXXThis = CGF.LoadCXXThis(); |
3820 | 0 | CGF.EmitStoreOfScalar(CXXThis, ThisLVal); |
3821 | 0 | } |
3822 | 0 | for (const LambdaCapture &LC : RD->captures()) { |
3823 | 0 | if (LC.getCaptureKind() != LCK_ByRef) |
3824 | 0 | continue; |
3825 | 0 | const VarDecl *VD = LC.getCapturedVar(); |
3826 | 0 | if (!CS->capturesVariable(VD)) |
3827 | 0 | continue; |
3828 | 0 | auto It = Captures.find(VD); |
3829 | 0 | assert(It != Captures.end() && "Found lambda capture without field."); |
3830 | 0 | LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second); |
3831 | 0 | Address VDAddr = CGF.GetAddrOfLocalVar(VD); |
3832 | 0 | if (VD->getType().getCanonicalType()->isReferenceType()) |
3833 | 0 | VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr, |
3834 | 0 | VD->getType().getCanonicalType()) |
3835 | 0 | .getAddress(CGF); |
3836 | 0 | CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal); |
3837 | 0 | } |
3838 | 0 | } |
3839 | 845 | } |
3840 | | |
3841 | | bool CGOpenMPRuntimeGPU::hasAllocateAttributeForGlobalVar(const VarDecl *VD, |
3842 | 101 | LangAS &AS) { |
3843 | 101 | if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>()) |
3844 | 88 | return false; |
3845 | 13 | const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); |
3846 | 13 | switch(A->getAllocatorType()) { |
3847 | 0 | case OMPAllocateDeclAttr::OMPNullMemAlloc: |
3848 | 1 | case OMPAllocateDeclAttr::OMPDefaultMemAlloc: |
3849 | | // Not supported, fallback to the default mem space. |
3850 | 2 | case OMPAllocateDeclAttr::OMPThreadMemAlloc: |
3851 | 4 | case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: |
3852 | 4 | case OMPAllocateDeclAttr::OMPCGroupMemAlloc: |
3853 | 8 | case OMPAllocateDeclAttr::OMPHighBWMemAlloc: |
3854 | 9 | case OMPAllocateDeclAttr::OMPLowLatMemAlloc: |
3855 | 9 | AS = LangAS::Default; |
3856 | 9 | return true; |
3857 | 2 | case OMPAllocateDeclAttr::OMPConstMemAlloc: |
3858 | 2 | AS = LangAS::cuda_constant; |
3859 | 2 | return true; |
3860 | 2 | case OMPAllocateDeclAttr::OMPPTeamMemAlloc: |
3861 | 2 | AS = LangAS::cuda_shared; |
3862 | 2 | return true; |
3863 | 0 | case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: |
3864 | 0 | llvm_unreachable("Expected predefined allocator for the variables with the " |
3865 | 13 | "static storage."); |
3866 | 13 | } |
3867 | 0 | return false; |
3868 | 13 | } |
3869 | | |
3870 | | // Get current CudaArch and ignore any unknown values |
3871 | 17 | static CudaArch getCudaArch(CodeGenModule &CGM) { |
3872 | 17 | if (!CGM.getTarget().hasFeature("ptx")) |
3873 | 0 | return CudaArch::UNKNOWN; |
3874 | 33 | for (const auto &Feature : CGM.getTarget().getTargetOpts().FeatureMap)17 { |
3875 | 33 | if (Feature.getValue()) { |
3876 | 33 | CudaArch Arch = StringToCudaArch(Feature.getKey()); |
3877 | 33 | if (Arch != CudaArch::UNKNOWN) |
3878 | 17 | return Arch; |
3879 | 33 | } |
3880 | 33 | } |
3881 | 0 | return CudaArch::UNKNOWN; |
3882 | 17 | } |
3883 | | |
3884 | | /// Check to see if target architecture supports unified addressing which is |
3885 | | /// a restriction for OpenMP requires clause "unified_shared_memory". |
3886 | | void CGOpenMPRuntimeGPU::processRequiresDirective( |
3887 | 17 | const OMPRequiresDecl *D) { |
3888 | 17 | for (const OMPClause *Clause : D->clauselists()) { |
3889 | 17 | if (Clause->getClauseKind() == OMPC_unified_shared_memory) { |
3890 | 17 | CudaArch Arch = getCudaArch(CGM); |
3891 | 17 | switch (Arch) { |
3892 | 1 | case CudaArch::SM_20: |
3893 | 2 | case CudaArch::SM_21: |
3894 | 3 | case CudaArch::SM_30: |
3895 | 4 | case CudaArch::SM_32: |
3896 | 5 | case CudaArch::SM_35: |
3897 | 6 | case CudaArch::SM_37: |
3898 | 7 | case CudaArch::SM_50: |
3899 | 8 | case CudaArch::SM_52: |
3900 | 9 | case CudaArch::SM_53: { |
3901 | 9 | SmallString<256> Buffer; |
3902 | 9 | llvm::raw_svector_ostream Out(Buffer); |
3903 | 9 | Out << "Target architecture " << CudaArchToString(Arch) |
3904 | 9 | << " does not support unified addressing"; |
3905 | 9 | CGM.Error(Clause->getBeginLoc(), Out.str()); |
3906 | 9 | return; |
3907 | 8 | } |
3908 | 1 | case CudaArch::SM_60: |
3909 | 2 | case CudaArch::SM_61: |
3910 | 3 | case CudaArch::SM_62: |
3911 | 6 | case CudaArch::SM_70: |
3912 | 7 | case CudaArch::SM_72: |
3913 | 8 | case CudaArch::SM_75: |
3914 | 8 | case CudaArch::SM_80: |
3915 | 8 | case CudaArch::SM_86: |
3916 | 8 | case CudaArch::GFX600: |
3917 | 8 | case CudaArch::GFX601: |
3918 | 8 | case CudaArch::GFX602: |
3919 | 8 | case CudaArch::GFX700: |
3920 | 8 | case CudaArch::GFX701: |
3921 | 8 | case CudaArch::GFX702: |
3922 | 8 | case CudaArch::GFX703: |
3923 | 8 | case CudaArch::GFX704: |
3924 | 8 | case CudaArch::GFX705: |
3925 | 8 | case CudaArch::GFX801: |
3926 | 8 | case CudaArch::GFX802: |
3927 | 8 | case CudaArch::GFX803: |
3928 | 8 | case CudaArch::GFX805: |
3929 | 8 | case CudaArch::GFX810: |
3930 | 8 | case CudaArch::GFX900: |
3931 | 8 | case CudaArch::GFX902: |
3932 | 8 | case CudaArch::GFX904: |
3933 | 8 | case CudaArch::GFX906: |
3934 | 8 | case CudaArch::GFX908: |
3935 | 8 | case CudaArch::GFX909: |
3936 | 8 | case CudaArch::GFX90a: |
3937 | 8 | case CudaArch::GFX90c: |
3938 | 8 | case CudaArch::GFX940: |
3939 | 8 | case CudaArch::GFX1010: |
3940 | 8 | case CudaArch::GFX1011: |
3941 | 8 | case CudaArch::GFX1012: |
3942 | 8 | case CudaArch::GFX1013: |
3943 | 8 | case CudaArch::GFX1030: |
3944 | 8 | case CudaArch::GFX1031: |
3945 | 8 | case CudaArch::GFX1032: |
3946 | 8 | case CudaArch::GFX1033: |
3947 | 8 | case CudaArch::GFX1034: |
3948 | 8 | case CudaArch::GFX1035: |
3949 | 8 | case CudaArch::GFX1036: |
3950 | 8 | case CudaArch::GFX1100: |
3951 | 8 | case CudaArch::GFX1101: |
3952 | 8 | case CudaArch::GFX1102: |
3953 | 8 | case CudaArch::GFX1103: |
3954 | 8 | case CudaArch::Generic: |
3955 | 8 | case CudaArch::UNUSED: |
3956 | 8 | case CudaArch::UNKNOWN: |
3957 | 8 | break; |
3958 | 0 | case CudaArch::LAST: |
3959 | 0 | llvm_unreachable("Unexpected Cuda arch."); |
3960 | 17 | } |
3961 | 17 | } |
3962 | 17 | } |
3963 | 8 | CGOpenMPRuntime::processRequiresDirective(D); |
3964 | 8 | } |
3965 | | |
3966 | 178 | void CGOpenMPRuntimeGPU::clear() { |
3967 | | |
3968 | 178 | if (!TeamsReductions.empty()) { |
3969 | 3 | ASTContext &C = CGM.getContext(); |
3970 | 3 | RecordDecl *StaticRD = C.buildImplicitRecord( |
3971 | 3 | "_openmp_teams_reduction_type_$_", RecordDecl::TagKind::TTK_Union); |
3972 | 3 | StaticRD->startDefinition(); |
3973 | 9 | for (const RecordDecl *TeamReductionRec : TeamsReductions) { |
3974 | 9 | QualType RecTy = C.getRecordType(TeamReductionRec); |
3975 | 9 | auto *Field = FieldDecl::Create( |
3976 | 9 | C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy, |
3977 | 9 | C.getTrivialTypeSourceInfo(RecTy, SourceLocation()), |
3978 | 9 | /*BW=*/nullptr, /*Mutable=*/false, |
3979 | 9 | /*InitStyle=*/ICIS_NoInit); |
3980 | 9 | Field->setAccess(AS_public); |
3981 | 9 | StaticRD->addDecl(Field); |
3982 | 9 | } |
3983 | 3 | StaticRD->completeDefinition(); |
3984 | 3 | QualType StaticTy = C.getRecordType(StaticRD); |
3985 | 3 | llvm::Type *LLVMReductionsBufferTy = |
3986 | 3 | CGM.getTypes().ConvertTypeForMem(StaticTy); |
3987 | | // FIXME: nvlink does not handle weak linkage correctly (object with the |
3988 | | // different size are reported as erroneous). |
3989 | | // Restore CommonLinkage as soon as nvlink is fixed. |
3990 | 3 | auto *GV = new llvm::GlobalVariable( |
3991 | 3 | CGM.getModule(), LLVMReductionsBufferTy, |
3992 | 3 | /*isConstant=*/false, llvm::GlobalValue::InternalLinkage, |
3993 | 3 | llvm::Constant::getNullValue(LLVMReductionsBufferTy), |
3994 | 3 | "_openmp_teams_reductions_buffer_$_"); |
3995 | 3 | KernelTeamsReductionPtr->setInitializer( |
3996 | 3 | llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, |
3997 | 3 | CGM.VoidPtrTy)); |
3998 | 3 | } |
3999 | 178 | CGOpenMPRuntime::clear(); |
4000 | 178 | } |
4001 | | |
4002 | 268 | llvm::Value *CGOpenMPRuntimeGPU::getGPUNumThreads(CodeGenFunction &CGF) { |
4003 | 268 | CGBuilderTy &Bld = CGF.Builder; |
4004 | 268 | llvm::Module *M = &CGF.CGM.getModule(); |
4005 | 268 | const char *LocSize = "__kmpc_get_hardware_num_threads_in_block"; |
4006 | 268 | llvm::Function *F = M->getFunction(LocSize); |
4007 | 268 | if (!F) { |
4008 | 33 | F = llvm::Function::Create( |
4009 | 33 | llvm::FunctionType::get(CGF.Int32Ty, llvm::None, false), |
4010 | 33 | llvm::GlobalVariable::ExternalLinkage, LocSize, &CGF.CGM.getModule()); |
4011 | 33 | } |
4012 | 268 | return Bld.CreateCall(F, llvm::None, "nvptx_num_threads"); |
4013 | 268 | } |
4014 | | |
4015 | 86 | llvm::Value *CGOpenMPRuntimeGPU::getGPUThreadID(CodeGenFunction &CGF) { |
4016 | 86 | ArrayRef<llvm::Value *> Args{}; |
4017 | 86 | return CGF.EmitRuntimeCall( |
4018 | 86 | OMPBuilder.getOrCreateRuntimeFunction( |
4019 | 86 | CGM.getModule(), OMPRTL___kmpc_get_hardware_thread_id_in_block), |
4020 | 86 | Args); |
4021 | 86 | } |
4022 | | |
4023 | 43 | llvm::Value *CGOpenMPRuntimeGPU::getGPUWarpSize(CodeGenFunction &CGF) { |
4024 | 43 | ArrayRef<llvm::Value *> Args{}; |
4025 | 43 | return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
4026 | 43 | CGM.getModule(), OMPRTL___kmpc_get_warp_size), |
4027 | 43 | Args); |
4028 | 43 | } |