| 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/OpenMPClause.h" |
| 19 |
#include "clang/AST/StmtOpenMP.h" |
| 20 |
#include "clang/AST/StmtVisitor.h" |
| 21 |
#include "clang/Basic/Cuda.h" |
| 22 |
#include "llvm/ADT/SmallPtrSet.h" |
| 23 |
#include "llvm/Frontend/OpenMP/OMPGridValues.h" |
| 24 |
#include "llvm/Support/MathExtras.h" |
| 25 |
|
| 26 |
using namespace clang; |
| 27 |
using namespace CodeGen; |
| 28 |
using namespace llvm::omp; |
| 29 |
|
| 30 |
namespace { |
| 31 |
/// Pre(post)-action for different OpenMP constructs specialized for NVPTX. |
| 32 |
class NVPTXActionTy final : public PrePostActionTy { |
| 33 |
llvm::FunctionCallee EnterCallee = nullptr; |
| 34 |
ArrayRef<llvm::Value *> EnterArgs; |
| 35 |
llvm::FunctionCallee ExitCallee = nullptr; |
| 36 |
ArrayRef<llvm::Value *> ExitArgs; |
| 37 |
bool Conditional = false; |
| 38 |
llvm::BasicBlock *ContBlock = nullptr; |
| 39 |
|
| 40 |
public: |
| 41 |
NVPTXActionTy(llvm::FunctionCallee EnterCallee, |
| 42 |
ArrayRef<llvm::Value *> EnterArgs, |
| 43 |
llvm::FunctionCallee ExitCallee, |
| 44 |
ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false) |
| 45 |
: EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee), |
| 46 |
ExitArgs(ExitArgs), Conditional(Conditional) {} |
| 47 |
void Enter(CodeGenFunction &CGF) override { |
| 48 |
llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs); |
| 49 |
if (Conditional) { |
| 50 |
llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes); |
| 51 |
auto *ThenBlock = CGF.createBasicBlock("omp_if.then"); |
| 52 |
ContBlock = CGF.createBasicBlock("omp_if.end"); |
| 53 |
// Generate the branch (If-stmt) |
| 54 |
CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock); |
| 55 |
CGF.EmitBlock(ThenBlock); |
| 56 |
} |
| 57 |
} |
| 58 |
void Done(CodeGenFunction &CGF) { |
| 59 |
// Emit the rest of blocks/branches |
| 60 |
CGF.EmitBranch(ContBlock); |
| 61 |
CGF.EmitBlock(ContBlock, true); |
| 62 |
} |
| 63 |
void Exit(CodeGenFunction &CGF) override { |
| 64 |
CGF.EmitRuntimeCall(ExitCallee, ExitArgs); |
| 65 |
} |
| 66 |
}; |
| 67 |
|
| 68 |
/// A class to track the execution mode when codegening directives within |
| 69 |
/// a target region. The appropriate mode (SPMD|NON-SPMD) is set on entry |
| 70 |
/// to the target region and used by containing directives such as 'parallel' |
| 71 |
/// to emit optimized code. |
| 72 |
class ExecutionRuntimeModesRAII { |
| 73 |
private: |
| 74 |
CGOpenMPRuntimeGPU::ExecutionMode SavedExecMode = |
| 75 |
CGOpenMPRuntimeGPU::EM_Unknown; |
| 76 |
CGOpenMPRuntimeGPU::ExecutionMode &ExecMode; |
| 77 |
|
| 78 |
public: |
| 79 |
ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode, |
| 80 |
CGOpenMPRuntimeGPU::ExecutionMode EntryMode) |
| 81 |
: ExecMode(ExecMode) { |
| 82 |
SavedExecMode = ExecMode; |
| 83 |
ExecMode = EntryMode; |
| 84 |
} |
| 85 |
~ExecutionRuntimeModesRAII() { ExecMode = SavedExecMode; } |
| 86 |
}; |
| 87 |
|
| 88 |
/// GPU Configuration: This information can be derived from cuda registers, |
| 89 |
/// however, providing compile time constants helps generate more efficient |
| 90 |
/// code. For all practical purposes this is fine because the configuration |
| 91 |
/// is the same for all known NVPTX architectures. |
| 92 |
enum MachineConfiguration : unsigned { |
| 93 |
/// See "llvm/Frontend/OpenMP/OMPGridValues.h" for various related target |
| 94 |
/// specific Grid Values like GV_Warp_Size, GV_Slot_Size |
| 95 |
|
| 96 |
/// Global memory alignment for performance. |
| 97 |
GlobalMemoryAlignment = 128, |
| 98 |
}; |
| 99 |
|
| 100 |
static const ValueDecl *getPrivateItem(const Expr *RefExpr) { |
| 101 |
RefExpr = RefExpr->IgnoreParens(); |
| 102 |
if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) { |
| 103 |
const Expr *Base = ASE->getBase()->IgnoreParenImpCasts(); |
| 104 |
while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) |
| 105 |
Base = TempASE->getBase()->IgnoreParenImpCasts(); |
| 106 |
RefExpr = Base; |
| 107 |
} else if (auto *OASE = dyn_cast<OMPArraySectionExpr>(RefExpr)) { |
| 108 |
const Expr *Base = OASE->getBase()->IgnoreParenImpCasts(); |
| 109 |
while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base)) |
| 110 |
Base = TempOASE->getBase()->IgnoreParenImpCasts(); |
| 111 |
while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) |
| 112 |
Base = TempASE->getBase()->IgnoreParenImpCasts(); |
| 113 |
RefExpr = Base; |
| 114 |
} |
| 115 |
RefExpr = RefExpr->IgnoreParenImpCasts(); |
| 116 |
if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr)) |
| 117 |
return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl()); |
| 118 |
const auto *ME = cast<MemberExpr>(RefExpr); |
| 119 |
return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl()); |
| 120 |
} |
| 121 |
|
| 122 |
|
| 123 |
static RecordDecl *buildRecordForGlobalizedVars( |
| 124 |
ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls, |
| 125 |
ArrayRef<const ValueDecl *> EscapedDeclsForTeams, |
| 126 |
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 127 |
&MappedDeclsFields, int BufSize) { |
| 128 |
using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>; |
| 129 |
if (EscapedDecls.empty() && EscapedDeclsForTeams.empty()) |
| 130 |
return nullptr; |
| 131 |
SmallVector<VarsDataTy, 4> GlobalizedVars; |
| 132 |
for (const ValueDecl *D : EscapedDecls) |
| 133 |
GlobalizedVars.emplace_back( |
| 134 |
CharUnits::fromQuantity(std::max( |
| 135 |
C.getDeclAlign(D).getQuantity(), |
| 136 |
static_cast<CharUnits::QuantityType>(GlobalMemoryAlignment))), |
| 137 |
D); |
| 138 |
for (const ValueDecl *D : EscapedDeclsForTeams) |
| 139 |
GlobalizedVars.emplace_back(C.getDeclAlign(D), D); |
| 140 |
llvm::stable_sort(GlobalizedVars, [](VarsDataTy L, VarsDataTy R) { |
| 141 |
return L.first > R.first; |
| 142 |
}); |
| 143 |
|
| 144 |
// Build struct _globalized_locals_ty { |
| 145 |
// /* globalized vars */[WarSize] align (max(decl_align, |
| 146 |
// GlobalMemoryAlignment)) |
| 147 |
// /* globalized vars */ for EscapedDeclsForTeams |
| 148 |
// }; |
| 149 |
RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty"); |
| 150 |
GlobalizedRD->startDefinition(); |
| 151 |
llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped( |
| 152 |
EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end()); |
| 153 |
for (const auto &Pair : GlobalizedVars) { |
| 154 |
const ValueDecl *VD = Pair.second; |
| 155 |
QualType Type = VD->getType(); |
| 156 |
if (Type->isLValueReferenceType()) |
| 157 |
Type = C.getPointerType(Type.getNonReferenceType()); |
| 158 |
else |
| 159 |
Type = Type.getNonReferenceType(); |
| 160 |
SourceLocation Loc = VD->getLocation(); |
| 161 |
FieldDecl *Field; |
| 162 |
if (SingleEscaped.count(VD)) { |
| 163 |
Field = FieldDecl::Create( |
| 164 |
C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, |
| 165 |
C.getTrivialTypeSourceInfo(Type, SourceLocation()), |
| 166 |
/*BW=*/nullptr, /*Mutable=*/false, |
| 167 |
/*InitStyle=*/ICIS_NoInit); |
| 168 |
Field->setAccess(AS_public); |
| 169 |
if (VD->hasAttrs()) { |
| 170 |
for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()), |
| 171 |
E(VD->getAttrs().end()); |
| 172 |
I != E; ++I) |
| 173 |
Field->addAttr(*I); |
| 174 |
} |
| 175 |
} else { |
| 176 |
llvm::APInt ArraySize(32, BufSize); |
| 177 |
Type = C.getConstantArrayType(Type, ArraySize, nullptr, ArrayType::Normal, |
| 178 |
0); |
| 179 |
Field = FieldDecl::Create( |
| 180 |
C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, |
| 181 |
C.getTrivialTypeSourceInfo(Type, SourceLocation()), |
| 182 |
/*BW=*/nullptr, /*Mutable=*/false, |
| 183 |
/*InitStyle=*/ICIS_NoInit); |
| 184 |
Field->setAccess(AS_public); |
| 185 |
llvm::APInt Align(32, std::max(C.getDeclAlign(VD).getQuantity(), |
| 186 |
static_cast<CharUnits::QuantityType>( |
| 187 |
GlobalMemoryAlignment))); |
| 188 |
Field->addAttr(AlignedAttr::CreateImplicit( |
| 189 |
C, /*IsAlignmentExpr=*/true, |
| 190 |
IntegerLiteral::Create(C, Align, |
| 191 |
C.getIntTypeForBitwidth(32, /*Signed=*/0), |
| 192 |
SourceLocation()), |
| 193 |
{}, AttributeCommonInfo::AS_GNU, AlignedAttr::GNU_aligned)); |
| 194 |
} |
| 195 |
GlobalizedRD->addDecl(Field); |
| 196 |
MappedDeclsFields.try_emplace(VD, Field); |
| 197 |
} |
| 198 |
GlobalizedRD->completeDefinition(); |
| 199 |
return GlobalizedRD; |
| 200 |
} |
| 201 |
|
| 202 |
/// Get the list of variables that can escape their declaration context. |
| 203 |
class CheckVarsEscapingDeclContext final |
| 204 |
: public ConstStmtVisitor<CheckVarsEscapingDeclContext> { |
| 205 |
CodeGenFunction &CGF; |
| 206 |
llvm::SetVector<const ValueDecl *> EscapedDecls; |
| 207 |
llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls; |
| 208 |
llvm::SmallPtrSet<const Decl *, 4> EscapedParameters; |
| 209 |
RecordDecl *GlobalizedRD = nullptr; |
| 210 |
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; |
| 211 |
bool AllEscaped = false; |
| 212 |
bool IsForCombinedParallelRegion = false; |
| 213 |
|
| 214 |
void markAsEscaped(const ValueDecl *VD) { |
| 215 |
// Do not globalize declare target variables. |
| 216 |
if (!isa<VarDecl>(VD) || |
| 217 |
OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD)) |
| 218 |
return; |
| 219 |
VD = cast<ValueDecl>(VD->getCanonicalDecl()); |
| 220 |
// Use user-specified allocation. |
| 221 |
if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>()) |
| 222 |
return; |
| 223 |
// Variables captured by value must be globalized. |
| 224 |
if (auto *CSI = CGF.CapturedStmtInfo) { |
| 225 |
if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) { |
| 226 |
// Check if need to capture the variable that was already captured by |
| 227 |
// value in the outer region. |
| 228 |
if (!IsForCombinedParallelRegion) { |
| 229 |
if (!FD->hasAttrs()) |
| 230 |
return; |
| 231 |
const auto *Attr = FD->getAttr<OMPCaptureKindAttr>(); |
| 232 |
if (!Attr) |
| 233 |
return; |
| 234 |
if (((Attr->getCaptureKind() != OMPC_map) && |
| 235 |
!isOpenMPPrivate(Attr->getCaptureKind())) || |
| 236 |
((Attr->getCaptureKind() == OMPC_map) && |
| 237 |
!FD->getType()->isAnyPointerType())) |
| 238 |
return; |
| 239 |
} |
| 240 |
if (!FD->getType()->isReferenceType()) { |
| 241 |
assert(!VD->getType()->isVariablyModifiedType() && |
| 242 |
"Parameter captured by value with variably modified type"); |
| 243 |
EscapedParameters.insert(VD); |
| 244 |
} else if (!IsForCombinedParallelRegion) { |
| 245 |
return; |
| 246 |
} |
| 247 |
} |
| 248 |
} |
| 249 |
if ((!CGF.CapturedStmtInfo || |
| 250 |
(IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) && |
| 251 |
VD->getType()->isReferenceType()) |
| 252 |
// Do not globalize variables with reference type. |
| 253 |
return; |
| 254 |
if (VD->getType()->isVariablyModifiedType()) |
| 255 |
EscapedVariableLengthDecls.insert(VD); |
| 256 |
else |
| 257 |
EscapedDecls.insert(VD); |
| 258 |
} |
| 259 |
|
| 260 |
void VisitValueDecl(const ValueDecl *VD) { |
| 261 |
if (VD->getType()->isLValueReferenceType()) |
| 262 |
markAsEscaped(VD); |
| 263 |
if (const auto *VarD = dyn_cast<VarDecl>(VD)) { |
| 264 |
if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) { |
| 265 |
const bool SavedAllEscaped = AllEscaped; |
| 266 |
AllEscaped = VD->getType()->isLValueReferenceType(); |
| 267 |
Visit(VarD->getInit()); |
| 268 |
AllEscaped = SavedAllEscaped; |
| 269 |
} |
| 270 |
} |
| 271 |
} |
| 272 |
void VisitOpenMPCapturedStmt(const CapturedStmt *S, |
| 273 |
ArrayRef<OMPClause *> Clauses, |
| 274 |
bool IsCombinedParallelRegion) { |
| 275 |
if (!S) |
| 276 |
return; |
| 277 |
for (const CapturedStmt::Capture &C : S->captures()) { |
| 278 |
if (C.capturesVariable() && !C.capturesVariableByCopy()) { |
| 279 |
const ValueDecl *VD = C.getCapturedVar(); |
| 280 |
bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion; |
| 281 |
if (IsCombinedParallelRegion) { |
| 282 |
// Check if the variable is privatized in the combined construct and |
| 283 |
// those private copies must be shared in the inner parallel |
| 284 |
// directive. |
| 285 |
IsForCombinedParallelRegion = false; |
| 286 |
for (const OMPClause *C : Clauses) { |
| 287 |
if (!isOpenMPPrivate(C->getClauseKind()) || |
| 288 |
C->getClauseKind() == OMPC_reduction || |
| 289 |
C->getClauseKind() == OMPC_linear || |
| 290 |
C->getClauseKind() == OMPC_private) |
| 291 |
continue; |
| 292 |
ArrayRef<const Expr *> Vars; |
| 293 |
if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C)) |
| 294 |
Vars = PC->getVarRefs(); |
| 295 |
else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C)) |
| 296 |
Vars = PC->getVarRefs(); |
| 297 |
else |
| 298 |
llvm_unreachable("Unexpected clause."); |
| 299 |
for (const auto *E : Vars) { |
| 300 |
const Decl *D = |
| 301 |
cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl(); |
| 302 |
if (D == VD->getCanonicalDecl()) { |
| 303 |
IsForCombinedParallelRegion = true; |
| 304 |
break; |
| 305 |
} |
| 306 |
} |
| 307 |
if (IsForCombinedParallelRegion) |
| 308 |
break; |
| 309 |
} |
| 310 |
} |
| 311 |
markAsEscaped(VD); |
| 312 |
if (isa<OMPCapturedExprDecl>(VD)) |
| 313 |
VisitValueDecl(VD); |
| 314 |
IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion; |
| 315 |
} |
| 316 |
} |
| 317 |
} |
| 318 |
|
| 319 |
void buildRecordForGlobalizedVars(bool IsInTTDRegion) { |
| 320 |
assert(!GlobalizedRD && |
| 321 |
"Record for globalized variables is built already."); |
| 322 |
ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams; |
| 323 |
unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; |
| 324 |
if (IsInTTDRegion) |
| 325 |
EscapedDeclsForTeams = EscapedDecls.getArrayRef(); |
| 326 |
else |
| 327 |
EscapedDeclsForParallel = EscapedDecls.getArrayRef(); |
| 328 |
GlobalizedRD = ::buildRecordForGlobalizedVars( |
| 329 |
CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams, |
| 330 |
MappedDeclsFields, WarpSize); |
| 331 |
} |
| 332 |
|
| 333 |
public: |
| 334 |
CheckVarsEscapingDeclContext(CodeGenFunction &CGF, |
| 335 |
ArrayRef<const ValueDecl *> TeamsReductions) |
| 336 |
: CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) { |
| 337 |
} |
| 338 |
virtual ~CheckVarsEscapingDeclContext() = default; |
| 339 |
void VisitDeclStmt(const DeclStmt *S) { |
| 340 |
if (!S) |
| 341 |
return; |
| 342 |
for (const Decl *D : S->decls()) |
| 343 |
if (const auto *VD = dyn_cast_or_null<ValueDecl>(D)) |
| 344 |
VisitValueDecl(VD); |
| 345 |
} |
| 346 |
void VisitOMPExecutableDirective(const OMPExecutableDirective *D) { |
| 347 |
if (!D) |
| 348 |
return; |
| 349 |
if (!D->hasAssociatedStmt()) |
| 350 |
return; |
| 351 |
if (const auto *S = |
| 352 |
dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) { |
| 353 |
// Do not analyze directives that do not actually require capturing, |
| 354 |
// like `omp for` or `omp simd` directives. |
| 355 |
llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; |
| 356 |
getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind()); |
| 357 |
if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) { |
| 358 |
VisitStmt(S->getCapturedStmt()); |
| 359 |
return; |
| 360 |
} |
| 361 |
VisitOpenMPCapturedStmt( |
| 362 |
S, D->clauses(), |
| 363 |
CaptureRegions.back() == OMPD_parallel && |
| 364 |
isOpenMPDistributeDirective(D->getDirectiveKind())); |
| 365 |
} |
| 366 |
} |
| 367 |
void VisitCapturedStmt(const CapturedStmt *S) { |
| 368 |
if (!S) |
| 369 |
return; |
| 370 |
for (const CapturedStmt::Capture &C : S->captures()) { |
| 371 |
if (C.capturesVariable() && !C.capturesVariableByCopy()) { |
| 372 |
const ValueDecl *VD = C.getCapturedVar(); |
| 373 |
markAsEscaped(VD); |
| 374 |
if (isa<OMPCapturedExprDecl>(VD)) |
| 375 |
VisitValueDecl(VD); |
| 376 |
} |
| 377 |
} |
| 378 |
} |
| 379 |
void VisitLambdaExpr(const LambdaExpr *E) { |
| 380 |
if (!E) |
| 381 |
return; |
| 382 |
for (const LambdaCapture &C : E->captures()) { |
| 383 |
if (C.capturesVariable()) { |
| 384 |
if (C.getCaptureKind() == LCK_ByRef) { |
| 385 |
const ValueDecl *VD = C.getCapturedVar(); |
| 386 |
markAsEscaped(VD); |
| 387 |
if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD)) |
| 388 |
VisitValueDecl(VD); |
| 389 |
} |
| 390 |
} |
| 391 |
} |
| 392 |
} |
| 393 |
void VisitBlockExpr(const BlockExpr *E) { |
| 394 |
if (!E) |
| 395 |
return; |
| 396 |
for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) { |
| 397 |
if (C.isByRef()) { |
| 398 |
const VarDecl *VD = C.getVariable(); |
| 399 |
markAsEscaped(VD); |
| 400 |
if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture()) |
| 401 |
VisitValueDecl(VD); |
| 402 |
} |
| 403 |
} |
| 404 |
} |
| 405 |
void VisitCallExpr(const CallExpr *E) { |
| 406 |
if (!E) |
| 407 |
return; |
| 408 |
for (const Expr *Arg : E->arguments()) { |
| 409 |
if (!Arg) |
| 410 |
continue; |
| 411 |
if (Arg->isLValue()) { |
| 412 |
const bool SavedAllEscaped = AllEscaped; |
| 413 |
AllEscaped = true; |
| 414 |
Visit(Arg); |
| 415 |
AllEscaped = SavedAllEscaped; |
| 416 |
} else { |
| 417 |
Visit(Arg); |
| 418 |
} |
| 419 |
} |
| 420 |
Visit(E->getCallee()); |
| 421 |
} |
| 422 |
void VisitDeclRefExpr(const DeclRefExpr *E) { |
| 423 |
if (!E) |
| 424 |
return; |
| 425 |
const ValueDecl *VD = E->getDecl(); |
| 426 |
if (AllEscaped) |
| 427 |
markAsEscaped(VD); |
| 428 |
if (isa<OMPCapturedExprDecl>(VD)) |
| 429 |
VisitValueDecl(VD); |
| 430 |
else if (VD->isInitCapture()) |
| 431 |
VisitValueDecl(VD); |
| 432 |
} |
| 433 |
void VisitUnaryOperator(const UnaryOperator *E) { |
| 434 |
if (!E) |
| 435 |
return; |
| 436 |
if (E->getOpcode() == UO_AddrOf) { |
| 437 |
const bool SavedAllEscaped = AllEscaped; |
| 438 |
AllEscaped = true; |
| 439 |
Visit(E->getSubExpr()); |
| 440 |
AllEscaped = SavedAllEscaped; |
| 441 |
} else { |
| 442 |
Visit(E->getSubExpr()); |
| 443 |
} |
| 444 |
} |
| 445 |
void VisitImplicitCastExpr(const ImplicitCastExpr *E) { |
| 446 |
if (!E) |
| 447 |
return; |
| 448 |
if (E->getCastKind() == CK_ArrayToPointerDecay) { |
| 449 |
const bool SavedAllEscaped = AllEscaped; |
| 450 |
AllEscaped = true; |
| 451 |
Visit(E->getSubExpr()); |
| 452 |
AllEscaped = SavedAllEscaped; |
| 453 |
} else { |
| 454 |
Visit(E->getSubExpr()); |
| 455 |
} |
| 456 |
} |
| 457 |
void VisitExpr(const Expr *E) { |
| 458 |
if (!E) |
| 459 |
return; |
| 460 |
bool SavedAllEscaped = AllEscaped; |
| 461 |
if (!E->isLValue()) |
| 462 |
AllEscaped = false; |
| 463 |
for (const Stmt *Child : E->children()) |
| 464 |
if (Child) |
| 465 |
Visit(Child); |
| 466 |
AllEscaped = SavedAllEscaped; |
| 467 |
} |
| 468 |
void VisitStmt(const Stmt *S) { |
| 469 |
if (!S) |
| 470 |
return; |
| 471 |
for (const Stmt *Child : S->children()) |
| 472 |
if (Child) |
| 473 |
Visit(Child); |
| 474 |
} |
| 475 |
|
| 476 |
/// Returns the record that handles all the escaped local variables and used |
| 477 |
/// instead of their original storage. |
| 478 |
const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) { |
| 479 |
if (!GlobalizedRD) |
| 480 |
buildRecordForGlobalizedVars(IsInTTDRegion); |
| 481 |
return GlobalizedRD; |
| 482 |
} |
| 483 |
|
| 484 |
/// Returns the field in the globalized record for the escaped variable. |
| 485 |
const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const { |
| 486 |
assert(GlobalizedRD && |
| 487 |
"Record for globalized variables must be generated already."); |
| 488 |
auto I = MappedDeclsFields.find(VD); |
| 489 |
if (I == MappedDeclsFields.end()) |
| 490 |
return nullptr; |
| 491 |
return I->getSecond(); |
| 492 |
} |
| 493 |
|
| 494 |
/// Returns the list of the escaped local variables/parameters. |
| 495 |
ArrayRef<const ValueDecl *> getEscapedDecls() const { |
| 496 |
return EscapedDecls.getArrayRef(); |
| 497 |
} |
| 498 |
|
| 499 |
/// Checks if the escaped local variable is actually a parameter passed by |
| 500 |
/// value. |
| 501 |
const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const { |
| 502 |
return EscapedParameters; |
| 503 |
} |
| 504 |
|
| 505 |
/// Returns the list of the escaped variables with the variably modified |
| 506 |
/// types. |
| 507 |
ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const { |
| 508 |
return EscapedVariableLengthDecls.getArrayRef(); |
| 509 |
} |
| 510 |
}; |
| 511 |
} // anonymous namespace |
| 512 |
|
| 513 |
/// Get the id of the warp in the block. |
| 514 |
/// We assume that the warp size is 32, which is always the case |
| 515 |
/// on the NVPTX device, to generate more efficient code. |
| 516 |
static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) { |
| 517 |
CGBuilderTy &Bld = CGF.Builder; |
| 518 |
unsigned LaneIDBits = |
| 519 |
llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size); |
| 520 |
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 521 |
return Bld.CreateAShr(RT.getGPUThreadID(CGF), LaneIDBits, "nvptx_warp_id"); |
| 522 |
} |
| 523 |
|
| 524 |
/// Get the id of the current lane in the Warp. |
| 525 |
/// We assume that the warp size is 32, which is always the case |
| 526 |
/// on the NVPTX device, to generate more efficient code. |
| 527 |
static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) { |
| 528 |
CGBuilderTy &Bld = CGF.Builder; |
| 529 |
unsigned LaneIDBits = |
| 530 |
llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size); |
| 531 |
unsigned LaneIDMask = ~0u >> (32u - LaneIDBits); |
| 532 |
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 533 |
return Bld.CreateAnd(RT.getGPUThreadID(CGF), Bld.getInt32(LaneIDMask), |
| 534 |
"nvptx_lane_id"); |
| 535 |
} |
| 536 |
|
| 537 |
CGOpenMPRuntimeGPU::ExecutionMode |
| 538 |
CGOpenMPRuntimeGPU::getExecutionMode() const { |
| 539 |
return CurrentExecutionMode; |
| 540 |
} |
| 541 |
|
| 542 |
static CGOpenMPRuntimeGPU::DataSharingMode |
| 543 |
getDataSharingMode(CodeGenModule &CGM) { |
| 544 |
return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeGPU::CUDA |
| 545 |
: CGOpenMPRuntimeGPU::Generic; |
| 546 |
} |
| 547 |
|
| 548 |
/// Check for inner (nested) SPMD construct, if any |
| 549 |
static bool hasNestedSPMDDirective(ASTContext &Ctx, |
| 550 |
const OMPExecutableDirective &D) { |
| 551 |
const auto *CS = D.getInnermostCapturedStmt(); |
| 552 |
const auto *Body = |
| 553 |
CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); |
| 554 |
const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
| 555 |
|
| 556 |
if (const auto *NestedDir = |
| 557 |
dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
| 558 |
OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); |
| 559 |
switch (D.getDirectiveKind()) { |
| 560 |
case OMPD_target: |
| 561 |
if (isOpenMPParallelDirective(DKind)) |
| 562 |
return true; |
| 563 |
if (DKind == OMPD_teams) { |
| 564 |
Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( |
| 565 |
/*IgnoreCaptured=*/true); |
| 566 |
if (!Body) |
| 567 |
return false; |
| 568 |
ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
| 569 |
if (const auto *NND = |
| 570 |
dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
| 571 |
DKind = NND->getDirectiveKind(); |
| 572 |
if (isOpenMPParallelDirective(DKind)) |
| 573 |
return true; |
| 574 |
} |
| 575 |
} |
| 576 |
return false; |
| 577 |
case OMPD_target_teams: |
| 578 |
return isOpenMPParallelDirective(DKind); |
| 579 |
case OMPD_target_simd: |
| 580 |
case OMPD_target_parallel: |
| 581 |
case OMPD_target_parallel_for: |
| 582 |
case OMPD_target_parallel_for_simd: |
| 583 |
case OMPD_target_teams_distribute: |
| 584 |
case OMPD_target_teams_distribute_simd: |
| 585 |
case OMPD_target_teams_distribute_parallel_for: |
| 586 |
case OMPD_target_teams_distribute_parallel_for_simd: |
| 587 |
case OMPD_parallel: |
| 588 |
case OMPD_for: |
| 589 |
case OMPD_parallel_for: |
| 590 |
case OMPD_parallel_master: |
| 591 |
case OMPD_parallel_sections: |
| 592 |
case OMPD_for_simd: |
| 593 |
case OMPD_parallel_for_simd: |
| 594 |
case OMPD_cancel: |
| 595 |
case OMPD_cancellation_point: |
| 596 |
case OMPD_ordered: |
| 597 |
case OMPD_threadprivate: |
| 598 |
case OMPD_allocate: |
| 599 |
case OMPD_task: |
| 600 |
case OMPD_simd: |
| 601 |
case OMPD_sections: |
| 602 |
case OMPD_section: |
| 603 |
case OMPD_single: |
| 604 |
case OMPD_master: |
| 605 |
case OMPD_critical: |
| 606 |
case OMPD_taskyield: |
| 607 |
case OMPD_barrier: |
| 608 |
case OMPD_taskwait: |
| 609 |
case OMPD_taskgroup: |
| 610 |
case OMPD_atomic: |
| 611 |
case OMPD_flush: |
| 612 |
case OMPD_depobj: |
| 613 |
case OMPD_scan: |
| 614 |
case OMPD_teams: |
| 615 |
case OMPD_target_data: |
| 616 |
case OMPD_target_exit_data: |
| 617 |
case OMPD_target_enter_data: |
| 618 |
case OMPD_distribute: |
| 619 |
case OMPD_distribute_simd: |
| 620 |
case OMPD_distribute_parallel_for: |
| 621 |
case OMPD_distribute_parallel_for_simd: |
| 622 |
case OMPD_teams_distribute: |
| 623 |
case OMPD_teams_distribute_simd: |
| 624 |
case OMPD_teams_distribute_parallel_for: |
| 625 |
case OMPD_teams_distribute_parallel_for_simd: |
| 626 |
case OMPD_target_update: |
| 627 |
case OMPD_declare_simd: |
| 628 |
case OMPD_declare_variant: |
| 629 |
case OMPD_begin_declare_variant: |
| 630 |
case OMPD_end_declare_variant: |
| 631 |
case OMPD_declare_target: |
| 632 |
case OMPD_end_declare_target: |
| 633 |
case OMPD_declare_reduction: |
| 634 |
case OMPD_declare_mapper: |
| 635 |
case OMPD_taskloop: |
| 636 |
case OMPD_taskloop_simd: |
| 637 |
case OMPD_master_taskloop: |
| 638 |
case OMPD_master_taskloop_simd: |
| 639 |
case OMPD_parallel_master_taskloop: |
| 640 |
case OMPD_parallel_master_taskloop_simd: |
| 641 |
case OMPD_requires: |
| 642 |
case OMPD_unknown: |
| 643 |
default: |
| 644 |
llvm_unreachable("Unexpected directive."); |
| 645 |
} |
| 646 |
} |
| 647 |
|
| 648 |
return false; |
| 649 |
} |
| 650 |
|
| 651 |
static bool supportsSPMDExecutionMode(ASTContext &Ctx, |
| 652 |
const OMPExecutableDirective &D) { |
| 653 |
OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); |
| 654 |
switch (DirectiveKind) { |
| 655 |
case OMPD_target: |
| 656 |
case OMPD_target_teams: |
| 657 |
return hasNestedSPMDDirective(Ctx, D); |
| 658 |
case OMPD_target_parallel: |
| 659 |
case OMPD_target_parallel_for: |
| 660 |
case OMPD_target_parallel_for_simd: |
| 661 |
case OMPD_target_teams_distribute_parallel_for: |
| 662 |
case OMPD_target_teams_distribute_parallel_for_simd: |
| 663 |
case OMPD_target_simd: |
| 664 |
case OMPD_target_teams_distribute_simd: |
| 665 |
return true; |
| 666 |
case OMPD_target_teams_distribute: |
| 667 |
return false; |
| 668 |
case OMPD_parallel: |
| 669 |
case OMPD_for: |
| 670 |
case OMPD_parallel_for: |
| 671 |
case OMPD_parallel_master: |
| 672 |
case OMPD_parallel_sections: |
| 673 |
case OMPD_for_simd: |
| 674 |
case OMPD_parallel_for_simd: |
| 675 |
case OMPD_cancel: |
| 676 |
case OMPD_cancellation_point: |
| 677 |
case OMPD_ordered: |
| 678 |
case OMPD_threadprivate: |
| 679 |
case OMPD_allocate: |
| 680 |
case OMPD_task: |
| 681 |
case OMPD_simd: |
| 682 |
case OMPD_sections: |
| 683 |
case OMPD_section: |
| 684 |
case OMPD_single: |
| 685 |
case OMPD_master: |
| 686 |
case OMPD_critical: |
| 687 |
case OMPD_taskyield: |
| 688 |
case OMPD_barrier: |
| 689 |
case OMPD_taskwait: |
| 690 |
case OMPD_taskgroup: |
| 691 |
case OMPD_atomic: |
| 692 |
case OMPD_flush: |
| 693 |
case OMPD_depobj: |
| 694 |
case OMPD_scan: |
| 695 |
case OMPD_teams: |
| 696 |
case OMPD_target_data: |
| 697 |
case OMPD_target_exit_data: |
| 698 |
case OMPD_target_enter_data: |
| 699 |
case OMPD_distribute: |
| 700 |
case OMPD_distribute_simd: |
| 701 |
case OMPD_distribute_parallel_for: |
| 702 |
case OMPD_distribute_parallel_for_simd: |
| 703 |
case OMPD_teams_distribute: |
| 704 |
case OMPD_teams_distribute_simd: |
| 705 |
case OMPD_teams_distribute_parallel_for: |
| 706 |
case OMPD_teams_distribute_parallel_for_simd: |
| 707 |
case OMPD_target_update: |
| 708 |
case OMPD_declare_simd: |
| 709 |
case OMPD_declare_variant: |
| 710 |
case OMPD_begin_declare_variant: |
| 711 |
case OMPD_end_declare_variant: |
| 712 |
case OMPD_declare_target: |
| 713 |
case OMPD_end_declare_target: |
| 714 |
case OMPD_declare_reduction: |
| 715 |
case OMPD_declare_mapper: |
| 716 |
case OMPD_taskloop: |
| 717 |
case OMPD_taskloop_simd: |
| 718 |
case OMPD_master_taskloop: |
| 719 |
case OMPD_master_taskloop_simd: |
| 720 |
case OMPD_parallel_master_taskloop: |
| 721 |
case OMPD_parallel_master_taskloop_simd: |
| 722 |
case OMPD_requires: |
| 723 |
case OMPD_unknown: |
| 724 |
default: |
| 725 |
break; |
| 726 |
} |
| 727 |
llvm_unreachable( |
| 728 |
"Unknown programming model for OpenMP directive on NVPTX target."); |
| 729 |
} |
| 730 |
|
| 731 |
void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D, |
| 732 |
StringRef ParentName, |
| 733 |
llvm::Function *&OutlinedFn, |
| 734 |
llvm::Constant *&OutlinedFnID, |
| 735 |
bool IsOffloadEntry, |
| 736 |
const RegionCodeGenTy &CodeGen) { |
| 737 |
ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_NonSPMD); |
| 738 |
EntryFunctionState EST; |
| 739 |
WrapperFunctionsMap.clear(); |
| 740 |
|
| 741 |
// Emit target region as a standalone region. |
| 742 |
class NVPTXPrePostActionTy : public PrePostActionTy { |
| 743 |
CGOpenMPRuntimeGPU::EntryFunctionState &EST; |
| 744 |
|
| 745 |
public: |
| 746 |
NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST) |
| 747 |
: EST(EST) {} |
| 748 |
void Enter(CodeGenFunction &CGF) override { |
| 749 |
auto &RT = |
| 750 |
static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 751 |
RT.emitKernelInit(CGF, EST, /* IsSPMD */ false); |
| 752 |
// Skip target region initialization. |
| 753 |
RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); |
| 754 |
} |
| 755 |
void Exit(CodeGenFunction &CGF) override { |
| 756 |
auto &RT = |
| 757 |
static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 758 |
RT.clearLocThreadIdInsertPt(CGF); |
| 759 |
RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ false); |
| 760 |
} |
| 761 |
} Action(EST); |
| 762 |
CodeGen.setAction(Action); |
| 763 |
IsInTTDRegion = true; |
| 764 |
emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, |
| 765 |
IsOffloadEntry, CodeGen); |
| 766 |
IsInTTDRegion = false; |
| 767 |
} |
| 768 |
|
| 769 |
void CGOpenMPRuntimeGPU::emitKernelInit(CodeGenFunction &CGF, |
| 770 |
EntryFunctionState &EST, bool IsSPMD) { |
| 771 |
CGBuilderTy &Bld = CGF.Builder; |
| 772 |
Bld.restoreIP(OMPBuilder.createTargetInit(Bld, IsSPMD)); |
| 773 |
if (!IsSPMD) |
| 774 |
emitGenericVarsProlog(CGF, EST.Loc); |
| 775 |
} |
| 776 |
|
| 777 |
void CGOpenMPRuntimeGPU::emitKernelDeinit(CodeGenFunction &CGF, |
| 778 |
EntryFunctionState &EST, |
| 779 |
bool IsSPMD) { |
| 780 |
if (!IsSPMD) |
| 781 |
emitGenericVarsEpilog(CGF); |
| 782 |
|
| 783 |
CGBuilderTy &Bld = CGF.Builder; |
| 784 |
OMPBuilder.createTargetDeinit(Bld, IsSPMD); |
| 785 |
} |
| 786 |
|
| 787 |
void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D, |
| 788 |
StringRef ParentName, |
| 789 |
llvm::Function *&OutlinedFn, |
| 790 |
llvm::Constant *&OutlinedFnID, |
| 791 |
bool IsOffloadEntry, |
| 792 |
const RegionCodeGenTy &CodeGen) { |
| 793 |
ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_SPMD); |
| 794 |
EntryFunctionState EST; |
| 795 |
|
| 796 |
// Emit target region as a standalone region. |
| 797 |
class NVPTXPrePostActionTy : public PrePostActionTy { |
| 798 |
CGOpenMPRuntimeGPU &RT; |
| 799 |
CGOpenMPRuntimeGPU::EntryFunctionState &EST; |
| 800 |
|
| 801 |
public: |
| 802 |
NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT, |
| 803 |
CGOpenMPRuntimeGPU::EntryFunctionState &EST) |
| 804 |
: RT(RT), EST(EST) {} |
| 805 |
void Enter(CodeGenFunction &CGF) override { |
| 806 |
RT.emitKernelInit(CGF, EST, /* IsSPMD */ true); |
| 807 |
// Skip target region initialization. |
| 808 |
RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); |
| 809 |
} |
| 810 |
void Exit(CodeGenFunction &CGF) override { |
| 811 |
RT.clearLocThreadIdInsertPt(CGF); |
| 812 |
RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ true); |
| 813 |
} |
| 814 |
} Action(*this, EST); |
| 815 |
CodeGen.setAction(Action); |
| 816 |
IsInTTDRegion = true; |
| 817 |
emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, |
| 818 |
IsOffloadEntry, CodeGen); |
| 819 |
IsInTTDRegion = false; |
| 820 |
} |
| 821 |
|
| 822 |
// Create a unique global variable to indicate the execution mode of this target |
| 823 |
// region. The execution mode is either 'generic', or 'spmd' depending on the |
| 824 |
// target directive. This variable is picked up by the offload library to setup |
| 825 |
// the device appropriately before kernel launch. If the execution mode is |
| 826 |
// 'generic', the runtime reserves one warp for the master, otherwise, all |
| 827 |
// warps participate in parallel work. |
| 828 |
static void setPropertyExecutionMode(CodeGenModule &CGM, StringRef Name, |
| 829 |
bool Mode) { |
| 830 |
auto *GVMode = new llvm::GlobalVariable( |
| 831 |
CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, |
| 832 |
llvm::GlobalValue::WeakAnyLinkage, |
| 833 |
llvm::ConstantInt::get(CGM.Int8Ty, Mode ? OMP_TGT_EXEC_MODE_SPMD |
| 834 |
: OMP_TGT_EXEC_MODE_GENERIC), |
| 835 |
Twine(Name, "_exec_mode")); |
| 836 |
GVMode->setVisibility(llvm::GlobalVariable::ProtectedVisibility); |
| 837 |
CGM.addCompilerUsedGlobal(GVMode); |
| 838 |
} |
| 839 |
|
| 840 |
void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction( |
| 841 |
const OMPExecutableDirective &D, StringRef ParentName, |
| 842 |
llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, |
| 843 |
bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { |
| 844 |
if (!IsOffloadEntry) // Nothing to do. |
| 845 |
return; |
| 846 |
|
| 847 |
assert(!ParentName.empty() && "Invalid target region parent name!"); |
| 848 |
|
| 849 |
bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D); |
| 850 |
if (Mode) |
| 851 |
emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, |
| 852 |
CodeGen); |
| 853 |
else |
| 854 |
emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, |
| 855 |
CodeGen); |
| 856 |
|
| 857 |
setPropertyExecutionMode(CGM, OutlinedFn->getName(), Mode); |
| 858 |
} |
| 859 |
|
| 860 |
CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM) |
| 861 |
: CGOpenMPRuntime(CGM) { |
| 862 |
llvm::OpenMPIRBuilderConfig Config(CGM.getLangOpts().OpenMPIsDevice, true, |
| 863 |
hasRequiresUnifiedSharedMemory(), |
| 864 |
CGM.getLangOpts().OpenMPOffloadMandatory); |
| 865 |
OMPBuilder.setConfig(Config); |
| 866 |
OffloadEntriesInfoManager.setConfig(Config); |
| 867 |
|
| 868 |
if (!CGM.getLangOpts().OpenMPIsDevice) |
| 869 |
llvm_unreachable("OpenMP can only handle device code."); |
| 870 |
|
| 871 |
llvm::OpenMPIRBuilder &OMPBuilder = getOMPBuilder(); |
| 872 |
if (CGM.getLangOpts().NoGPULib || CGM.getLangOpts().OMPHostIRFile.empty()) |
| 873 |
return; |
| 874 |
|
| 875 |
OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTargetDebug, |
| 876 |
"__omp_rtl_debug_kind"); |
| 877 |
OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTeamSubscription, |
| 878 |
"__omp_rtl_assume_teams_oversubscription"); |
| 879 |
OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPThreadSubscription, |
| 880 |
"__omp_rtl_assume_threads_oversubscription"); |
| 881 |
OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoThreadState, |
| 882 |
"__omp_rtl_assume_no_thread_state"); |
| 883 |
OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoNestedParallelism, |
| 884 |
"__omp_rtl_assume_no_nested_parallelism"); |
| 885 |
} |
| 886 |
|
| 887 |
void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF, |
| 888 |
ProcBindKind ProcBind, |
| 889 |
SourceLocation Loc) { |
| 890 |
// Do nothing in case of SPMD mode and L0 parallel. |
| 891 |
if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) |
| 892 |
return; |
| 893 |
|
| 894 |
CGOpenMPRuntime::emitProcBindClause(CGF, ProcBind, Loc); |
| 895 |
} |
| 896 |
|
| 897 |
void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF, |
| 898 |
llvm::Value *NumThreads, |
| 899 |
SourceLocation Loc) { |
| 900 |
// Nothing to do. |
| 901 |
} |
| 902 |
|
| 903 |
void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF, |
| 904 |
const Expr *NumTeams, |
| 905 |
const Expr *ThreadLimit, |
| 906 |
SourceLocation Loc) {} |
| 907 |
|
| 908 |
llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction( |
| 909 |
const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, |
| 910 |
OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { |
| 911 |
// Emit target region as a standalone region. |
| 912 |
bool PrevIsInTTDRegion = IsInTTDRegion; |
| 913 |
IsInTTDRegion = false; |
| 914 |
auto *OutlinedFun = |
| 915 |
cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction( |
| 916 |
D, ThreadIDVar, InnermostKind, CodeGen)); |
| 917 |
IsInTTDRegion = PrevIsInTTDRegion; |
| 918 |
if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) { |
| 919 |
llvm::Function *WrapperFun = |
| 920 |
createParallelDataSharingWrapper(OutlinedFun, D); |
| 921 |
WrapperFunctionsMap[OutlinedFun] = WrapperFun; |
| 922 |
} |
| 923 |
|
| 924 |
return OutlinedFun; |
| 925 |
} |
| 926 |
|
| 927 |
/// Get list of lastprivate variables from the teams distribute ... or |
| 928 |
/// teams {distribute ...} directives. |
| 929 |
static void |
| 930 |
getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D, |
| 931 |
llvm::SmallVectorImpl<const ValueDecl *> &Vars) { |
| 932 |
assert(isOpenMPTeamsDirective(D.getDirectiveKind()) && |
| 933 |
"expected teams directive."); |
| 934 |
const OMPExecutableDirective *Dir = &D; |
| 935 |
if (!isOpenMPDistributeDirective(D.getDirectiveKind())) { |
| 936 |
if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild( |
| 937 |
Ctx, |
| 938 |
D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers( |
| 939 |
/*IgnoreCaptured=*/true))) { |
| 940 |
Dir = dyn_cast_or_null<OMPExecutableDirective>(S); |
| 941 |
if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind())) |
| 942 |
Dir = nullptr; |
| 943 |
} |
| 944 |
} |
| 945 |
if (!Dir) |
| 946 |
return; |
| 947 |
for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) { |
| 948 |
for (const Expr *E : C->getVarRefs()) |
| 949 |
Vars.push_back(getPrivateItem(E)); |
| 950 |
} |
| 951 |
} |
| 952 |
|
| 953 |
/// Get list of reduction variables from the teams ... directives. |
| 954 |
static void |
| 955 |
getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D, |
| 956 |
llvm::SmallVectorImpl<const ValueDecl *> &Vars) { |
| 957 |
assert(isOpenMPTeamsDirective(D.getDirectiveKind()) && |
| 958 |
"expected teams directive."); |
| 959 |
for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { |
| 960 |
for (const Expr *E : C->privates()) |
| 961 |
Vars.push_back(getPrivateItem(E)); |
| 962 |
} |
| 963 |
} |
| 964 |
|
| 965 |
llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction( |
| 966 |
const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, |
| 967 |
OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { |
| 968 |
SourceLocation Loc = D.getBeginLoc(); |
| 969 |
|
| 970 |
const RecordDecl *GlobalizedRD = nullptr; |
| 971 |
llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions; |
| 972 |
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; |
| 973 |
unsigned WarpSize = CGM.getTarget().getGridValue().GV_Warp_Size; |
| 974 |
// Globalize team reductions variable unconditionally in all modes. |
| 975 |
if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) |
| 976 |
getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions); |
| 977 |
if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { |
| 978 |
getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions); |
| 979 |
if (!LastPrivatesReductions.empty()) { |
| 980 |
GlobalizedRD = ::buildRecordForGlobalizedVars( |
| 981 |
CGM.getContext(), std::nullopt, LastPrivatesReductions, |
| 982 |
MappedDeclsFields, WarpSize); |
| 983 |
} |
| 984 |
} else if (!LastPrivatesReductions.empty()) { |
| 985 |
assert(!TeamAndReductions.first && |
| 986 |
"Previous team declaration is not expected."); |
| 987 |
TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl(); |
| 988 |
std::swap(TeamAndReductions.second, LastPrivatesReductions); |
| 989 |
} |
| 990 |
|
| 991 |
// Emit target region as a standalone region. |
| 992 |
class NVPTXPrePostActionTy : public PrePostActionTy { |
| 993 |
SourceLocation &Loc; |
| 994 |
const RecordDecl *GlobalizedRD; |
| 995 |
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 996 |
&MappedDeclsFields; |
| 997 |
|
| 998 |
public: |
| 999 |
NVPTXPrePostActionTy( |
| 1000 |
SourceLocation &Loc, const RecordDecl *GlobalizedRD, |
| 1001 |
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 1002 |
&MappedDeclsFields) |
| 1003 |
: Loc(Loc), GlobalizedRD(GlobalizedRD), |
| 1004 |
MappedDeclsFields(MappedDeclsFields) {} |
| 1005 |
void Enter(CodeGenFunction &CGF) override { |
| 1006 |
auto &Rt = |
| 1007 |
static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 1008 |
if (GlobalizedRD) { |
| 1009 |
auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; |
| 1010 |
I->getSecond().MappedParams = |
| 1011 |
std::make_unique<CodeGenFunction::OMPMapVars>(); |
| 1012 |
DeclToAddrMapTy &Data = I->getSecond().LocalVarData; |
| 1013 |
for (const auto &Pair : MappedDeclsFields) { |
| 1014 |
assert(Pair.getFirst()->isCanonicalDecl() && |
| 1015 |
"Expected canonical declaration"); |
| 1016 |
Data.insert(std::make_pair(Pair.getFirst(), MappedVarData())); |
| 1017 |
} |
| 1018 |
} |
| 1019 |
Rt.emitGenericVarsProlog(CGF, Loc); |
| 1020 |
} |
| 1021 |
void Exit(CodeGenFunction &CGF) override { |
| 1022 |
static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) |
| 1023 |
.emitGenericVarsEpilog(CGF); |
| 1024 |
} |
| 1025 |
} Action(Loc, GlobalizedRD, MappedDeclsFields); |
| 1026 |
CodeGen.setAction(Action); |
| 1027 |
llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction( |
| 1028 |
D, ThreadIDVar, InnermostKind, CodeGen); |
| 1029 |
|
| 1030 |
return OutlinedFun; |
| 1031 |
} |
| 1032 |
|
| 1033 |
void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF, |
| 1034 |
SourceLocation Loc, |
| 1035 |
bool WithSPMDCheck) { |
| 1036 |
if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic && |
| 1037 |
getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) |
| 1038 |
return; |
| 1039 |
|
| 1040 |
CGBuilderTy &Bld = CGF.Builder; |
| 1041 |
|
| 1042 |
const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); |
| 1043 |
if (I == FunctionGlobalizedDecls.end()) |
| 1044 |
return; |
| 1045 |
|
| 1046 |
for (auto &Rec : I->getSecond().LocalVarData) { |
| 1047 |
const auto *VD = cast<VarDecl>(Rec.first); |
| 1048 |
bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first); |
| 1049 |
QualType VarTy = VD->getType(); |
| 1050 |
|
| 1051 |
// Get the local allocation of a firstprivate variable before sharing |
| 1052 |
llvm::Value *ParValue; |
| 1053 |
if (EscapedParam) { |
| 1054 |
LValue ParLVal = |
| 1055 |
CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType()); |
| 1056 |
ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc); |
| 1057 |
} |
| 1058 |
|
| 1059 |
// Allocate space for the variable to be globalized |
| 1060 |
llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())}; |
| 1061 |
llvm::CallBase *VoidPtr = |
| 1062 |
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1063 |
CGM.getModule(), OMPRTL___kmpc_alloc_shared), |
| 1064 |
AllocArgs, VD->getName()); |
| 1065 |
// FIXME: We should use the variables actual alignment as an argument. |
| 1066 |
VoidPtr->addRetAttr(llvm::Attribute::get( |
| 1067 |
CGM.getLLVMContext(), llvm::Attribute::Alignment, |
| 1068 |
CGM.getContext().getTargetInfo().getNewAlign() / 8)); |
| 1069 |
|
| 1070 |
// Cast the void pointer and get the address of the globalized variable. |
| 1071 |
llvm::PointerType *VarPtrTy = CGF.ConvertTypeForMem(VarTy)->getPointerTo(); |
| 1072 |
llvm::Value *CastedVoidPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 1073 |
VoidPtr, VarPtrTy, VD->getName() + "_on_stack"); |
| 1074 |
LValue VarAddr = CGF.MakeNaturalAlignAddrLValue(CastedVoidPtr, VarTy); |
| 1075 |
Rec.second.PrivateAddr = VarAddr.getAddress(CGF); |
| 1076 |
Rec.second.GlobalizedVal = VoidPtr; |
| 1077 |
|
| 1078 |
// Assign the local allocation to the newly globalized location. |
| 1079 |
if (EscapedParam) { |
| 1080 |
CGF.EmitStoreOfScalar(ParValue, VarAddr); |
| 1081 |
I->getSecond().MappedParams->setVarAddr(CGF, VD, VarAddr.getAddress(CGF)); |
| 1082 |
} |
| 1083 |
if (auto *DI = CGF.getDebugInfo()) |
| 1084 |
VoidPtr->setDebugLoc(DI->SourceLocToDebugLoc(VD->getLocation())); |
| 1085 |
} |
| 1086 |
for (const auto *VD : I->getSecond().EscapedVariableLengthDecls) { |
| 1087 |
// Use actual memory size of the VLA object including the padding |
| 1088 |
// for alignment purposes. |
| 1089 |
llvm::Value *Size = CGF.getTypeSize(VD->getType()); |
| 1090 |
CharUnits Align = CGM.getContext().getDeclAlign(VD); |
| 1091 |
Size = Bld.CreateNUWAdd( |
| 1092 |
Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1)); |
| 1093 |
llvm::Value *AlignVal = |
| 1094 |
llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity()); |
| 1095 |
|
| 1096 |
Size = Bld.CreateUDiv(Size, AlignVal); |
| 1097 |
Size = Bld.CreateNUWMul(Size, AlignVal); |
| 1098 |
|
| 1099 |
// Allocate space for this VLA object to be globalized. |
| 1100 |
llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())}; |
| 1101 |
llvm::CallBase *VoidPtr = |
| 1102 |
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1103 |
CGM.getModule(), OMPRTL___kmpc_alloc_shared), |
| 1104 |
AllocArgs, VD->getName()); |
| 1105 |
VoidPtr->addRetAttr( |
| 1106 |
llvm::Attribute::get(CGM.getLLVMContext(), llvm::Attribute::Alignment, |
| 1107 |
CGM.getContext().getTargetInfo().getNewAlign())); |
| 1108 |
|
| 1109 |
I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back( |
| 1110 |
std::pair<llvm::Value *, llvm::Value *>( |
| 1111 |
{VoidPtr, CGF.getTypeSize(VD->getType())})); |
| 1112 |
LValue Base = CGF.MakeAddrLValue(VoidPtr, VD->getType(), |
| 1113 |
CGM.getContext().getDeclAlign(VD), |
| 1114 |
AlignmentSource::Decl); |
| 1115 |
I->getSecond().MappedParams->setVarAddr(CGF, cast<VarDecl>(VD), |
| 1116 |
Base.getAddress(CGF)); |
| 1117 |
} |
| 1118 |
I->getSecond().MappedParams->apply(CGF); |
| 1119 |
} |
| 1120 |
|
| 1121 |
void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF, |
| 1122 |
bool WithSPMDCheck) { |
| 1123 |
if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic && |
| 1124 |
getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) |
| 1125 |
return; |
| 1126 |
|
| 1127 |
const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); |
| 1128 |
if (I != FunctionGlobalizedDecls.end()) { |
| 1129 |
// Deallocate the memory for each globalized VLA object |
| 1130 |
for (auto AddrSizePair : |
| 1131 |
llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) { |
| 1132 |
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1133 |
CGM.getModule(), OMPRTL___kmpc_free_shared), |
| 1134 |
{AddrSizePair.first, AddrSizePair.second}); |
| 1135 |
} |
| 1136 |
// Deallocate the memory for each globalized value |
| 1137 |
for (auto &Rec : llvm::reverse(I->getSecond().LocalVarData)) { |
| 1138 |
const auto *VD = cast<VarDecl>(Rec.first); |
| 1139 |
I->getSecond().MappedParams->restore(CGF); |
| 1140 |
|
| 1141 |
llvm::Value *FreeArgs[] = {Rec.second.GlobalizedVal, |
| 1142 |
CGF.getTypeSize(VD->getType())}; |
| 1143 |
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1144 |
CGM.getModule(), OMPRTL___kmpc_free_shared), |
| 1145 |
FreeArgs); |
| 1146 |
} |
| 1147 |
} |
| 1148 |
} |
| 1149 |
|
| 1150 |
void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF, |
| 1151 |
const OMPExecutableDirective &D, |
| 1152 |
SourceLocation Loc, |
| 1153 |
llvm::Function *OutlinedFn, |
| 1154 |
ArrayRef<llvm::Value *> CapturedVars) { |
| 1155 |
if (!CGF.HaveInsertPoint()) |
| 1156 |
return; |
| 1157 |
|
| 1158 |
Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, |
| 1159 |
/*Name=*/".zero.addr"); |
| 1160 |
CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr); |
| 1161 |
llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; |
| 1162 |
OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer()); |
| 1163 |
OutlinedFnArgs.push_back(ZeroAddr.getPointer()); |
| 1164 |
OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end()); |
| 1165 |
emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs); |
| 1166 |
} |
| 1167 |
|
| 1168 |
void CGOpenMPRuntimeGPU::emitParallelCall(CodeGenFunction &CGF, |
| 1169 |
SourceLocation Loc, |
| 1170 |
llvm::Function *OutlinedFn, |
| 1171 |
ArrayRef<llvm::Value *> CapturedVars, |
| 1172 |
const Expr *IfCond, |
| 1173 |
llvm::Value *NumThreads) { |
| 1174 |
if (!CGF.HaveInsertPoint()) |
| 1175 |
return; |
| 1176 |
|
| 1177 |
auto &&ParallelGen = [this, Loc, OutlinedFn, CapturedVars, IfCond, |
| 1178 |
NumThreads](CodeGenFunction &CGF, |
| 1179 |
PrePostActionTy &Action) { |
| 1180 |
CGBuilderTy &Bld = CGF.Builder; |
| 1181 |
llvm::Value *NumThreadsVal = NumThreads; |
| 1182 |
llvm::Function *WFn = WrapperFunctionsMap[OutlinedFn]; |
| 1183 |
llvm::Value *ID = llvm::ConstantPointerNull::get(CGM.Int8PtrTy); |
| 1184 |
if (WFn) |
| 1185 |
ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy); |
| 1186 |
llvm::Value *FnPtr = Bld.CreateBitOrPointerCast(OutlinedFn, CGM.Int8PtrTy); |
| 1187 |
|
| 1188 |
// Create a private scope that will globalize the arguments |
| 1189 |
// passed from the outside of the target region. |
| 1190 |
// TODO: Is that needed? |
| 1191 |
CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF); |
| 1192 |
|
| 1193 |
Address CapturedVarsAddrs = CGF.CreateDefaultAlignTempAlloca( |
| 1194 |
llvm::ArrayType::get(CGM.VoidPtrTy, CapturedVars.size()), |
| 1195 |
"captured_vars_addrs"); |
| 1196 |
// There's something to share. |
| 1197 |
if (!CapturedVars.empty()) { |
| 1198 |
// Prepare for parallel region. Indicate the outlined function. |
| 1199 |
ASTContext &Ctx = CGF.getContext(); |
| 1200 |
unsigned Idx = 0; |
| 1201 |
for (llvm::Value *V : CapturedVars) { |
| 1202 |
Address Dst = Bld.CreateConstArrayGEP(CapturedVarsAddrs, Idx); |
| 1203 |
llvm::Value *PtrV; |
| 1204 |
if (V->getType()->isIntegerTy()) |
| 1205 |
PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy); |
| 1206 |
else |
| 1207 |
PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy); |
| 1208 |
CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false, |
| 1209 |
Ctx.getPointerType(Ctx.VoidPtrTy)); |
| 1210 |
++Idx; |
| 1211 |
} |
| 1212 |
} |
| 1213 |
|
| 1214 |
llvm::Value *IfCondVal = nullptr; |
| 1215 |
if (IfCond) |
| 1216 |
IfCondVal = Bld.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.Int32Ty, |
| 1217 |
/* isSigned */ false); |
| 1218 |
else |
| 1219 |
IfCondVal = llvm::ConstantInt::get(CGF.Int32Ty, 1); |
| 1220 |
|
| 1221 |
if (!NumThreadsVal) |
| 1222 |
NumThreadsVal = llvm::ConstantInt::get(CGF.Int32Ty, -1); |
| 1223 |
else |
| 1224 |
NumThreadsVal = Bld.CreateZExtOrTrunc(NumThreadsVal, CGF.Int32Ty), |
| 1225 |
|
| 1226 |
assert(IfCondVal && "Expected a value"); |
| 1227 |
llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); |
| 1228 |
llvm::Value *Args[] = { |
| 1229 |
RTLoc, |
| 1230 |
getThreadID(CGF, Loc), |
| 1231 |
IfCondVal, |
| 1232 |
NumThreadsVal, |
| 1233 |
llvm::ConstantInt::get(CGF.Int32Ty, -1), |
| 1234 |
FnPtr, |
| 1235 |
ID, |
| 1236 |
Bld.CreateBitOrPointerCast(CapturedVarsAddrs.getPointer(), |
| 1237 |
CGF.VoidPtrPtrTy), |
| 1238 |
llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())}; |
| 1239 |
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1240 |
CGM.getModule(), OMPRTL___kmpc_parallel_51), |
| 1241 |
Args); |
| 1242 |
}; |
| 1243 |
|
| 1244 |
RegionCodeGenTy RCG(ParallelGen); |
| 1245 |
RCG(CGF); |
| 1246 |
} |
| 1247 |
|
| 1248 |
void CGOpenMPRuntimeGPU::syncCTAThreads(CodeGenFunction &CGF) { |
| 1249 |
// Always emit simple barriers! |
| 1250 |
if (!CGF.HaveInsertPoint()) |
| 1251 |
return; |
| 1252 |
// Build call __kmpc_barrier_simple_spmd(nullptr, 0); |
| 1253 |
// This function does not use parameters, so we can emit just default values. |
| 1254 |
llvm::Value *Args[] = { |
| 1255 |
llvm::ConstantPointerNull::get( |
| 1256 |
cast<llvm::PointerType>(getIdentTyPointerTy())), |
| 1257 |
llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)}; |
| 1258 |
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1259 |
CGM.getModule(), OMPRTL___kmpc_barrier_simple_spmd), |
| 1260 |
Args); |
| 1261 |
} |
| 1262 |
|
| 1263 |
void CGOpenMPRuntimeGPU::emitBarrierCall(CodeGenFunction &CGF, |
| 1264 |
SourceLocation Loc, |
| 1265 |
OpenMPDirectiveKind Kind, bool, |
| 1266 |
bool) { |
| 1267 |
// Always emit simple barriers! |
| 1268 |
if (!CGF.HaveInsertPoint()) |
| 1269 |
return; |
| 1270 |
// Build call __kmpc_cancel_barrier(loc, thread_id); |
| 1271 |
unsigned Flags = getDefaultFlagsForBarriers(Kind); |
| 1272 |
llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags), |
| 1273 |
getThreadID(CGF, Loc)}; |
| 1274 |
|
| 1275 |
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1276 |
CGM.getModule(), OMPRTL___kmpc_barrier), |
| 1277 |
Args); |
| 1278 |
} |
| 1279 |
|
| 1280 |
void CGOpenMPRuntimeGPU::emitCriticalRegion( |
| 1281 |
CodeGenFunction &CGF, StringRef CriticalName, |
| 1282 |
const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc, |
| 1283 |
const Expr *Hint) { |
| 1284 |
llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop"); |
| 1285 |
llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test"); |
| 1286 |
llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync"); |
| 1287 |
llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body"); |
| 1288 |
llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit"); |
| 1289 |
|
| 1290 |
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 1291 |
|
| 1292 |
// Get the mask of active threads in the warp. |
| 1293 |
llvm::Value *Mask = CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1294 |
CGM.getModule(), OMPRTL___kmpc_warp_active_thread_mask)); |
| 1295 |
// Fetch team-local id of the thread. |
| 1296 |
llvm::Value *ThreadID = RT.getGPUThreadID(CGF); |
| 1297 |
|
| 1298 |
// Get the width of the team. |
| 1299 |
llvm::Value *TeamWidth = RT.getGPUNumThreads(CGF); |
| 1300 |
|
| 1301 |
// Initialize the counter variable for the loop. |
| 1302 |
QualType Int32Ty = |
| 1303 |
CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0); |
| 1304 |
Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter"); |
| 1305 |
LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty); |
| 1306 |
CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal, |
| 1307 |
/*isInit=*/true); |
| 1308 |
|
| 1309 |
// Block checks if loop counter exceeds upper bound. |
| 1310 |
CGF.EmitBlock(LoopBB); |
| 1311 |
llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); |
| 1312 |
llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth); |
| 1313 |
CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB); |
| 1314 |
|
| 1315 |
// Block tests which single thread should execute region, and which threads |
| 1316 |
// should go straight to synchronisation point. |
| 1317 |
CGF.EmitBlock(TestBB); |
| 1318 |
CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); |
| 1319 |
llvm::Value *CmpThreadToCounter = |
| 1320 |
CGF.Builder.CreateICmpEQ(ThreadID, CounterVal); |
| 1321 |
CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB); |
| 1322 |
|
| 1323 |
// Block emits the body of the critical region. |
| 1324 |
CGF.EmitBlock(BodyBB); |
| 1325 |
|
| 1326 |
// Output the critical statement. |
| 1327 |
CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc, |
| 1328 |
Hint); |
| 1329 |
|
| 1330 |
// After the body surrounded by the critical region, the single executing |
| 1331 |
// thread will jump to the synchronisation point. |
| 1332 |
// Block waits for all threads in current team to finish then increments the |
| 1333 |
// counter variable and returns to the loop. |
| 1334 |
CGF.EmitBlock(SyncBB); |
| 1335 |
// Reconverge active threads in the warp. |
| 1336 |
(void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1337 |
CGM.getModule(), OMPRTL___kmpc_syncwarp), |
| 1338 |
Mask); |
| 1339 |
|
| 1340 |
llvm::Value *IncCounterVal = |
| 1341 |
CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1)); |
| 1342 |
CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal); |
| 1343 |
CGF.EmitBranch(LoopBB); |
| 1344 |
|
| 1345 |
// Block that is reached when all threads in the team complete the region. |
| 1346 |
CGF.EmitBlock(ExitBB, /*IsFinished=*/true); |
| 1347 |
} |
| 1348 |
|
| 1349 |
/// Cast value to the specified type. |
| 1350 |
static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val, |
| 1351 |
QualType ValTy, QualType CastTy, |
| 1352 |
SourceLocation Loc) { |
| 1353 |
assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && |
| 1354 |
"Cast type must sized."); |
| 1355 |
assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && |
| 1356 |
"Val type must sized."); |
| 1357 |
llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy); |
| 1358 |
if (ValTy == CastTy) |
| 1359 |
return Val; |
| 1360 |
if (CGF.getContext().getTypeSizeInChars(ValTy) == |
| 1361 |
CGF.getContext().getTypeSizeInChars(CastTy)) |
| 1362 |
return CGF.Builder.CreateBitCast(Val, LLVMCastTy); |
| 1363 |
if (CastTy->isIntegerType() && ValTy->isIntegerType()) |
| 1364 |
return CGF.Builder.CreateIntCast(Val, LLVMCastTy, |
| 1365 |
CastTy->hasSignedIntegerRepresentation()); |
| 1366 |
Address CastItem = CGF.CreateMemTemp(CastTy); |
| 1367 |
Address ValCastItem = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 1368 |
CastItem, Val->getType()->getPointerTo(CastItem.getAddressSpace()), |
| 1369 |
Val->getType()); |
| 1370 |
CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy, |
| 1371 |
LValueBaseInfo(AlignmentSource::Type), |
| 1372 |
TBAAAccessInfo()); |
| 1373 |
return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc, |
| 1374 |
LValueBaseInfo(AlignmentSource::Type), |
| 1375 |
TBAAAccessInfo()); |
| 1376 |
} |
| 1377 |
|
| 1378 |
/// This function creates calls to one of two shuffle functions to copy |
| 1379 |
/// variables between lanes in a warp. |
| 1380 |
static llvm::Value *createRuntimeShuffleFunction(CodeGenFunction &CGF, |
| 1381 |
llvm::Value *Elem, |
| 1382 |
QualType ElemType, |
| 1383 |
llvm::Value *Offset, |
| 1384 |
SourceLocation Loc) { |
| 1385 |
CodeGenModule &CGM = CGF.CGM; |
| 1386 |
CGBuilderTy &Bld = CGF.Builder; |
| 1387 |
CGOpenMPRuntimeGPU &RT = |
| 1388 |
*(static_cast<CGOpenMPRuntimeGPU *>(&CGM.getOpenMPRuntime())); |
| 1389 |
llvm::OpenMPIRBuilder &OMPBuilder = RT.getOMPBuilder(); |
| 1390 |
|
| 1391 |
CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); |
| 1392 |
assert(Size.getQuantity() <= 8 && |
| 1393 |
"Unsupported bitwidth in shuffle instruction."); |
| 1394 |
|
| 1395 |
RuntimeFunction ShuffleFn = Size.getQuantity() <= 4 |
| 1396 |
? OMPRTL___kmpc_shuffle_int32 |
| 1397 |
: OMPRTL___kmpc_shuffle_int64; |
| 1398 |
|
| 1399 |
// Cast all types to 32- or 64-bit values before calling shuffle routines. |
| 1400 |
QualType CastTy = CGF.getContext().getIntTypeForBitwidth( |
| 1401 |
Size.getQuantity() <= 4 ? 32 : 64, /*Signed=*/1); |
| 1402 |
llvm::Value *ElemCast = castValueToType(CGF, Elem, ElemType, CastTy, Loc); |
| 1403 |
llvm::Value *WarpSize = |
| 1404 |
Bld.CreateIntCast(RT.getGPUWarpSize(CGF), CGM.Int16Ty, /*isSigned=*/true); |
| 1405 |
|
| 1406 |
llvm::Value *ShuffledVal = CGF.EmitRuntimeCall( |
| 1407 |
OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), ShuffleFn), |
| 1408 |
{ElemCast, Offset, WarpSize}); |
| 1409 |
|
| 1410 |
return castValueToType(CGF, ShuffledVal, CastTy, ElemType, Loc); |
| 1411 |
} |
| 1412 |
|
| 1413 |
static void shuffleAndStore(CodeGenFunction &CGF, Address SrcAddr, |
| 1414 |
Address DestAddr, QualType ElemType, |
| 1415 |
llvm::Value *Offset, SourceLocation Loc) { |
| 1416 |
CGBuilderTy &Bld = CGF.Builder; |
| 1417 |
|
| 1418 |
CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); |
| 1419 |
// Create the loop over the big sized data. |
| 1420 |
// ptr = (void*)Elem; |
| 1421 |
// ptrEnd = (void*) Elem + 1; |
| 1422 |
// Step = 8; |
| 1423 |
// while (ptr + Step < ptrEnd) |
| 1424 |
// shuffle((int64_t)*ptr); |
| 1425 |
// Step = 4; |
| 1426 |
// while (ptr + Step < ptrEnd) |
| 1427 |
// shuffle((int32_t)*ptr); |
| 1428 |
// ... |
| 1429 |
Address ElemPtr = DestAddr; |
| 1430 |
Address Ptr = SrcAddr; |
| 1431 |
Address PtrEnd = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 1432 |
Bld.CreateConstGEP(SrcAddr, 1), CGF.VoidPtrTy, CGF.Int8Ty); |
| 1433 |
for (int IntSize = 8; IntSize >= 1; IntSize /= 2) { |
| 1434 |
if (Size < CharUnits::fromQuantity(IntSize)) |
| 1435 |
continue; |
| 1436 |
QualType IntType = CGF.getContext().getIntTypeForBitwidth( |
| 1437 |
CGF.getContext().toBits(CharUnits::fromQuantity(IntSize)), |
| 1438 |
/*Signed=*/1); |
| 1439 |
llvm::Type *IntTy = CGF.ConvertTypeForMem(IntType); |
| 1440 |
Ptr = Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr, IntTy->getPointerTo(), |
| 1441 |
IntTy); |
| 1442 |
ElemPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 1443 |
ElemPtr, IntTy->getPointerTo(), IntTy); |
| 1444 |
if (Size.getQuantity() / IntSize > 1) { |
| 1445 |
llvm::BasicBlock *PreCondBB = CGF.createBasicBlock(".shuffle.pre_cond"); |
| 1446 |
llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".shuffle.then"); |
| 1447 |
llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".shuffle.exit"); |
| 1448 |
llvm::BasicBlock *CurrentBB = Bld.GetInsertBlock(); |
| 1449 |
CGF.EmitBlock(PreCondBB); |
| 1450 |
llvm::PHINode *PhiSrc = |
| 1451 |
Bld.CreatePHI(Ptr.getType(), /*NumReservedValues=*/2); |
| 1452 |
PhiSrc->addIncoming(Ptr.getPointer(), CurrentBB); |
| 1453 |
llvm::PHINode *PhiDest = |
| 1454 |
Bld.CreatePHI(ElemPtr.getType(), /*NumReservedValues=*/2); |
| 1455 |
PhiDest->addIncoming(ElemPtr.getPointer(), CurrentBB); |
| 1456 |
Ptr = Address(PhiSrc, Ptr.getElementType(), Ptr.getAlignment()); |
| 1457 |
ElemPtr = |
| 1458 |
Address(PhiDest, ElemPtr.getElementType(), ElemPtr.getAlignment()); |
| 1459 |
llvm::Value *PtrDiff = Bld.CreatePtrDiff( |
| 1460 |
CGF.Int8Ty, PtrEnd.getPointer(), |
| 1461 |
Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr.getPointer(), |
| 1462 |
CGF.VoidPtrTy)); |
| 1463 |
Bld.CreateCondBr(Bld.CreateICmpSGT(PtrDiff, Bld.getInt64(IntSize - 1)), |
| 1464 |
ThenBB, ExitBB); |
| 1465 |
CGF.EmitBlock(ThenBB); |
| 1466 |
llvm::Value *Res = createRuntimeShuffleFunction( |
| 1467 |
CGF, |
| 1468 |
CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc, |
| 1469 |
LValueBaseInfo(AlignmentSource::Type), |
| 1470 |
TBAAAccessInfo()), |
| 1471 |
IntType, Offset, Loc); |
| 1472 |
CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType, |
| 1473 |
LValueBaseInfo(AlignmentSource::Type), |
| 1474 |
TBAAAccessInfo()); |
| 1475 |
Address LocalPtr = Bld.CreateConstGEP(Ptr, 1); |
| 1476 |
Address LocalElemPtr = Bld.CreateConstGEP(ElemPtr, 1); |
| 1477 |
PhiSrc->addIncoming(LocalPtr.getPointer(), ThenBB); |
| 1478 |
PhiDest->addIncoming(LocalElemPtr.getPointer(), ThenBB); |
| 1479 |
CGF.EmitBranch(PreCondBB); |
| 1480 |
CGF.EmitBlock(ExitBB); |
| 1481 |
} else { |
| 1482 |
llvm::Value *Res = createRuntimeShuffleFunction( |
| 1483 |
CGF, |
| 1484 |
CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc, |
| 1485 |
LValueBaseInfo(AlignmentSource::Type), |
| 1486 |
TBAAAccessInfo()), |
| 1487 |
IntType, Offset, Loc); |
| 1488 |
CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType, |
| 1489 |
LValueBaseInfo(AlignmentSource::Type), |
| 1490 |
TBAAAccessInfo()); |
| 1491 |
Ptr = Bld.CreateConstGEP(Ptr, 1); |
| 1492 |
ElemPtr = Bld.CreateConstGEP(ElemPtr, 1); |
| 1493 |
} |
| 1494 |
Size = Size % IntSize; |
| 1495 |
} |
| 1496 |
} |
| 1497 |
|
| 1498 |
namespace { |
| 1499 |
enum CopyAction : unsigned { |
| 1500 |
// RemoteLaneToThread: Copy over a Reduce list from a remote lane in |
| 1501 |
// the warp using shuffle instructions. |
| 1502 |
RemoteLaneToThread, |
| 1503 |
// ThreadCopy: Make a copy of a Reduce list on the thread's stack. |
| 1504 |
ThreadCopy, |
| 1505 |
// ThreadToScratchpad: Copy a team-reduced array to the scratchpad. |
| 1506 |
ThreadToScratchpad, |
| 1507 |
// ScratchpadToThread: Copy from a scratchpad array in global memory |
| 1508 |
// containing team-reduced data to a thread's stack. |
| 1509 |
ScratchpadToThread, |
| 1510 |
}; |
| 1511 |
} // namespace |
| 1512 |
|
| 1513 |
struct CopyOptionsTy { |
| 1514 |
llvm::Value *RemoteLaneOffset; |
| 1515 |
llvm::Value *ScratchpadIndex; |
| 1516 |
llvm::Value *ScratchpadWidth; |
| 1517 |
}; |
| 1518 |
|
| 1519 |
/// Emit instructions to copy a Reduce list, which contains partially |
| 1520 |
/// aggregated values, in the specified direction. |
| 1521 |
static void emitReductionListCopy( |
| 1522 |
CopyAction Action, CodeGenFunction &CGF, QualType ReductionArrayTy, |
| 1523 |
ArrayRef<const Expr *> Privates, Address SrcBase, Address DestBase, |
| 1524 |
CopyOptionsTy CopyOptions = {nullptr, nullptr, nullptr}) { |
| 1525 |
|
| 1526 |
CodeGenModule &CGM = CGF.CGM; |
| 1527 |
ASTContext &C = CGM.getContext(); |
| 1528 |
CGBuilderTy &Bld = CGF.Builder; |
| 1529 |
|
| 1530 |
llvm::Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset; |
| 1531 |
llvm::Value *ScratchpadIndex = CopyOptions.ScratchpadIndex; |
| 1532 |
llvm::Value *ScratchpadWidth = CopyOptions.ScratchpadWidth; |
| 1533 |
|
| 1534 |
// Iterates, element-by-element, through the source Reduce list and |
| 1535 |
// make a copy. |
| 1536 |
unsigned Idx = 0; |
| 1537 |
unsigned Size = Privates.size(); |
| 1538 |
for (const Expr *Private : Privates) { |
| 1539 |
Address SrcElementAddr = Address::invalid(); |
| 1540 |
Address DestElementAddr = Address::invalid(); |
| 1541 |
Address DestElementPtrAddr = Address::invalid(); |
| 1542 |
// Should we shuffle in an element from a remote lane? |
| 1543 |
bool ShuffleInElement = false; |
| 1544 |
// Set to true to update the pointer in the dest Reduce list to a |
| 1545 |
// newly created element. |
| 1546 |
bool UpdateDestListPtr = false; |
| 1547 |
// Increment the src or dest pointer to the scratchpad, for each |
| 1548 |
// new element. |
| 1549 |
bool IncrScratchpadSrc = false; |
| 1550 |
bool IncrScratchpadDest = false; |
| 1551 |
QualType PrivatePtrType = C.getPointerType(Private->getType()); |
| 1552 |
llvm::Type *PrivateLlvmPtrType = CGF.ConvertType(PrivatePtrType); |
| 1553 |
|
| 1554 |
switch (Action) { |
| 1555 |
case RemoteLaneToThread: { |
| 1556 |
// Step 1.1: Get the address for the src element in the Reduce list. |
| 1557 |
Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); |
| 1558 |
SrcElementAddr = |
| 1559 |
CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
| 1560 |
SrcElementPtrAddr, PrivateLlvmPtrType), |
| 1561 |
PrivatePtrType->castAs<PointerType>()); |
| 1562 |
|
| 1563 |
// Step 1.2: Create a temporary to store the element in the destination |
| 1564 |
// Reduce list. |
| 1565 |
DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); |
| 1566 |
DestElementAddr = |
| 1567 |
CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element"); |
| 1568 |
ShuffleInElement = true; |
| 1569 |
UpdateDestListPtr = true; |
| 1570 |
break; |
| 1571 |
} |
| 1572 |
case ThreadCopy: { |
| 1573 |
// Step 1.1: Get the address for the src element in the Reduce list. |
| 1574 |
Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); |
| 1575 |
SrcElementAddr = |
| 1576 |
CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
| 1577 |
SrcElementPtrAddr, PrivateLlvmPtrType), |
| 1578 |
PrivatePtrType->castAs<PointerType>()); |
| 1579 |
|
| 1580 |
// Step 1.2: Get the address for dest element. The destination |
| 1581 |
// element has already been created on the thread's stack. |
| 1582 |
DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); |
| 1583 |
DestElementAddr = |
| 1584 |
CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
| 1585 |
DestElementPtrAddr, PrivateLlvmPtrType), |
| 1586 |
PrivatePtrType->castAs<PointerType>()); |
| 1587 |
break; |
| 1588 |
} |
| 1589 |
case ThreadToScratchpad: { |
| 1590 |
// Step 1.1: Get the address for the src element in the Reduce list. |
| 1591 |
Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); |
| 1592 |
SrcElementAddr = |
| 1593 |
CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
| 1594 |
SrcElementPtrAddr, PrivateLlvmPtrType), |
| 1595 |
PrivatePtrType->castAs<PointerType>()); |
| 1596 |
|
| 1597 |
// Step 1.2: Get the address for dest element: |
| 1598 |
// address = base + index * ElementSizeInChars. |
| 1599 |
llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); |
| 1600 |
llvm::Value *CurrentOffset = |
| 1601 |
Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex); |
| 1602 |
llvm::Value *ScratchPadElemAbsolutePtrVal = |
| 1603 |
Bld.CreateNUWAdd(DestBase.getPointer(), CurrentOffset); |
| 1604 |
ScratchPadElemAbsolutePtrVal = |
| 1605 |
Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy); |
| 1606 |
DestElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty, |
| 1607 |
C.getTypeAlignInChars(Private->getType())); |
| 1608 |
IncrScratchpadDest = true; |
| 1609 |
break; |
| 1610 |
} |
| 1611 |
case ScratchpadToThread: { |
| 1612 |
// Step 1.1: Get the address for the src element in the scratchpad. |
| 1613 |
// address = base + index * ElementSizeInChars. |
| 1614 |
llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); |
| 1615 |
llvm::Value *CurrentOffset = |
| 1616 |
Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex); |
| 1617 |
llvm::Value *ScratchPadElemAbsolutePtrVal = |
| 1618 |
Bld.CreateNUWAdd(SrcBase.getPointer(), CurrentOffset); |
| 1619 |
ScratchPadElemAbsolutePtrVal = |
| 1620 |
Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy); |
| 1621 |
SrcElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty, |
| 1622 |
C.getTypeAlignInChars(Private->getType())); |
| 1623 |
IncrScratchpadSrc = true; |
| 1624 |
|
| 1625 |
// Step 1.2: Create a temporary to store the element in the destination |
| 1626 |
// Reduce list. |
| 1627 |
DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); |
| 1628 |
DestElementAddr = |
| 1629 |
CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element"); |
| 1630 |
UpdateDestListPtr = true; |
| 1631 |
break; |
| 1632 |
} |
| 1633 |
} |
| 1634 |
|
| 1635 |
// Regardless of src and dest of copy, we emit the load of src |
| 1636 |
// element as this is required in all directions |
| 1637 |
SrcElementAddr = Bld.CreateElementBitCast( |
| 1638 |
SrcElementAddr, CGF.ConvertTypeForMem(Private->getType())); |
| 1639 |
DestElementAddr = Bld.CreateElementBitCast(DestElementAddr, |
| 1640 |
SrcElementAddr.getElementType()); |
| 1641 |
|
| 1642 |
// Now that all active lanes have read the element in the |
| 1643 |
// Reduce list, shuffle over the value from the remote lane. |
| 1644 |
if (ShuffleInElement) { |
| 1645 |
shuffleAndStore(CGF, SrcElementAddr, DestElementAddr, Private->getType(), |
| 1646 |
RemoteLaneOffset, Private->getExprLoc()); |
| 1647 |
} else { |
| 1648 |
switch (CGF.getEvaluationKind(Private->getType())) { |
| 1649 |
case TEK_Scalar: { |
| 1650 |
llvm::Value *Elem = CGF.EmitLoadOfScalar( |
| 1651 |
SrcElementAddr, /*Volatile=*/false, Private->getType(), |
| 1652 |
Private->getExprLoc(), LValueBaseInfo(AlignmentSource::Type), |
| 1653 |
TBAAAccessInfo()); |
| 1654 |
// Store the source element value to the dest element address. |
| 1655 |
CGF.EmitStoreOfScalar( |
| 1656 |
Elem, DestElementAddr, /*Volatile=*/false, Private->getType(), |
| 1657 |
LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); |
| 1658 |
break; |
| 1659 |
} |
| 1660 |
case TEK_Complex: { |
| 1661 |
CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex( |
| 1662 |
CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), |
| 1663 |
Private->getExprLoc()); |
| 1664 |
CGF.EmitStoreOfComplex( |
| 1665 |
Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()), |
| 1666 |
/*isInit=*/false); |
| 1667 |
break; |
| 1668 |
} |
| 1669 |
case TEK_Aggregate: |
| 1670 |
CGF.EmitAggregateCopy( |
| 1671 |
CGF.MakeAddrLValue(DestElementAddr, Private->getType()), |
| 1672 |
CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), |
| 1673 |
Private->getType(), AggValueSlot::DoesNotOverlap); |
| 1674 |
break; |
| 1675 |
} |
| 1676 |
} |
| 1677 |
|
| 1678 |
// Step 3.1: Modify reference in dest Reduce list as needed. |
| 1679 |
// Modifying the reference in Reduce list to point to the newly |
| 1680 |
// created element. The element is live in the current function |
| 1681 |
// scope and that of functions it invokes (i.e., reduce_function). |
| 1682 |
// RemoteReduceData[i] = (void*)&RemoteElem |
| 1683 |
if (UpdateDestListPtr) { |
| 1684 |
CGF.EmitStoreOfScalar(Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 1685 |
DestElementAddr.getPointer(), CGF.VoidPtrTy), |
| 1686 |
DestElementPtrAddr, /*Volatile=*/false, |
| 1687 |
C.VoidPtrTy); |
| 1688 |
} |
| 1689 |
|
| 1690 |
// Step 4.1: Increment SrcBase/DestBase so that it points to the starting |
| 1691 |
// address of the next element in scratchpad memory, unless we're currently |
| 1692 |
// processing the last one. Memory alignment is also taken care of here. |
| 1693 |
if ((IncrScratchpadDest || IncrScratchpadSrc) && (Idx + 1 < Size)) { |
| 1694 |
// FIXME: This code doesn't make any sense, it's trying to perform |
| 1695 |
// integer arithmetic on pointers. |
| 1696 |
llvm::Value *ScratchpadBasePtr = |
| 1697 |
IncrScratchpadDest ? DestBase.getPointer() : SrcBase.getPointer(); |
| 1698 |
llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); |
| 1699 |
ScratchpadBasePtr = Bld.CreateNUWAdd( |
| 1700 |
ScratchpadBasePtr, |
| 1701 |
Bld.CreateNUWMul(ScratchpadWidth, ElementSizeInChars)); |
| 1702 |
|
| 1703 |
// Take care of global memory alignment for performance |
| 1704 |
ScratchpadBasePtr = Bld.CreateNUWSub( |
| 1705 |
ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1)); |
| 1706 |
ScratchpadBasePtr = Bld.CreateUDiv( |
| 1707 |
ScratchpadBasePtr, |
| 1708 |
llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment)); |
| 1709 |
ScratchpadBasePtr = Bld.CreateNUWAdd( |
| 1710 |
ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1)); |
| 1711 |
ScratchpadBasePtr = Bld.CreateNUWMul( |
| 1712 |
ScratchpadBasePtr, |
| 1713 |
llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment)); |
| 1714 |
|
| 1715 |
if (IncrScratchpadDest) |
| 1716 |
DestBase = |
| 1717 |
Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign()); |
| 1718 |
else /* IncrScratchpadSrc = true */ |
| 1719 |
SrcBase = |
| 1720 |
Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign()); |
| 1721 |
} |
| 1722 |
|
| 1723 |
++Idx; |
| 1724 |
} |
| 1725 |
} |
| 1726 |
|
| 1727 |
/// This function emits a helper that gathers Reduce lists from the first |
| 1728 |
/// lane of every active warp to lanes in the first warp. |
| 1729 |
/// |
| 1730 |
/// void inter_warp_copy_func(void* reduce_data, num_warps) |
| 1731 |
/// shared smem[warp_size]; |
| 1732 |
/// For all data entries D in reduce_data: |
| 1733 |
/// sync |
| 1734 |
/// If (I am the first lane in each warp) |
| 1735 |
/// Copy my local D to smem[warp_id] |
| 1736 |
/// sync |
| 1737 |
/// if (I am the first warp) |
| 1738 |
/// Copy smem[thread_id] to my local D |
| 1739 |
static llvm::Value *emitInterWarpCopyFunction(CodeGenModule &CGM, |
| 1740 |
ArrayRef<const Expr *> Privates, |
| 1741 |
QualType ReductionArrayTy, |
| 1742 |
SourceLocation Loc) { |
| 1743 |
ASTContext &C = CGM.getContext(); |
| 1744 |
llvm::Module &M = CGM.getModule(); |
| 1745 |
|
| 1746 |
// ReduceList: thread local Reduce list. |
| 1747 |
// At the stage of the computation when this function is called, partially |
| 1748 |
// aggregated values reside in the first lane of every active warp. |
| 1749 |
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 1750 |
C.VoidPtrTy, ImplicitParamDecl::Other); |
| 1751 |
// NumWarps: number of warps active in the parallel region. This could |
| 1752 |
// be smaller than 32 (max warps in a CTA) for partial block reduction. |
| 1753 |
ImplicitParamDecl NumWarpsArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 1754 |
C.getIntTypeForBitwidth(32, /* Signed */ true), |
| 1755 |
ImplicitParamDecl::Other); |
| 1756 |
FunctionArgList Args; |
| 1757 |
Args.push_back(&ReduceListArg); |
| 1758 |
Args.push_back(&NumWarpsArg); |
| 1759 |
|
| 1760 |
const CGFunctionInfo &CGFI = |
| 1761 |
CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 1762 |
auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), |
| 1763 |
llvm::GlobalValue::InternalLinkage, |
| 1764 |
"_omp_reduction_inter_warp_copy_func", &M); |
| 1765 |
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 1766 |
Fn->setDoesNotRecurse(); |
| 1767 |
CodeGenFunction CGF(CGM); |
| 1768 |
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 1769 |
|
| 1770 |
CGBuilderTy &Bld = CGF.Builder; |
| 1771 |
|
| 1772 |
// This array is used as a medium to transfer, one reduce element at a time, |
| 1773 |
// the data from the first lane of every warp to lanes in the first warp |
| 1774 |
// in order to perform the final step of a reduction in a parallel region |
| 1775 |
// (reduction across warps). The array is placed in NVPTX __shared__ memory |
| 1776 |
// for reduced latency, as well as to have a distinct copy for concurrently |
| 1777 |
// executing target regions. The array is declared with common linkage so |
| 1778 |
// as to be shared across compilation units. |
| 1779 |
StringRef TransferMediumName = |
| 1780 |
"__openmp_nvptx_data_transfer_temporary_storage"; |
| 1781 |
llvm::GlobalVariable *TransferMedium = |
| 1782 |
M.getGlobalVariable(TransferMediumName); |
| 1783 |
unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; |
| 1784 |
if (!TransferMedium) { |
| 1785 |
auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize); |
| 1786 |
unsigned SharedAddressSpace = C.getTargetAddressSpace(LangAS::cuda_shared); |
| 1787 |
TransferMedium = new llvm::GlobalVariable( |
| 1788 |
M, Ty, /*isConstant=*/false, llvm::GlobalVariable::WeakAnyLinkage, |
| 1789 |
llvm::UndefValue::get(Ty), TransferMediumName, |
| 1790 |
/*InsertBefore=*/nullptr, llvm::GlobalVariable::NotThreadLocal, |
| 1791 |
SharedAddressSpace); |
| 1792 |
CGM.addCompilerUsedGlobal(TransferMedium); |
| 1793 |
} |
| 1794 |
|
| 1795 |
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 1796 |
// Get the CUDA thread id of the current OpenMP thread on the GPU. |
| 1797 |
llvm::Value *ThreadID = RT.getGPUThreadID(CGF); |
| 1798 |
// nvptx_lane_id = nvptx_id % warpsize |
| 1799 |
llvm::Value *LaneID = getNVPTXLaneID(CGF); |
| 1800 |
// nvptx_warp_id = nvptx_id / warpsize |
| 1801 |
llvm::Value *WarpID = getNVPTXWarpID(CGF); |
| 1802 |
|
| 1803 |
Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 1804 |
llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
| 1805 |
Address LocalReduceList( |
| 1806 |
Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 1807 |
CGF.EmitLoadOfScalar( |
| 1808 |
AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc, |
| 1809 |
LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()), |
| 1810 |
ElemTy->getPointerTo()), |
| 1811 |
ElemTy, CGF.getPointerAlign()); |
| 1812 |
|
| 1813 |
unsigned Idx = 0; |
| 1814 |
for (const Expr *Private : Privates) { |
| 1815 |
// |
| 1816 |
// Warp master copies reduce element to transfer medium in __shared__ |
| 1817 |
// memory. |
| 1818 |
// |
| 1819 |
unsigned RealTySize = |
| 1820 |
C.getTypeSizeInChars(Private->getType()) |
| 1821 |
.alignTo(C.getTypeAlignInChars(Private->getType())) |
| 1822 |
.getQuantity(); |
| 1823 |
for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /=2) { |
| 1824 |
unsigned NumIters = RealTySize / TySize; |
| 1825 |
if (NumIters == 0) |
| 1826 |
continue; |
| 1827 |
QualType CType = C.getIntTypeForBitwidth( |
| 1828 |
C.toBits(CharUnits::fromQuantity(TySize)), /*Signed=*/1); |
| 1829 |
llvm::Type *CopyType = CGF.ConvertTypeForMem(CType); |
| 1830 |
CharUnits Align = CharUnits::fromQuantity(TySize); |
| 1831 |
llvm::Value *Cnt = nullptr; |
| 1832 |
Address CntAddr = Address::invalid(); |
| 1833 |
llvm::BasicBlock *PrecondBB = nullptr; |
| 1834 |
llvm::BasicBlock *ExitBB = nullptr; |
| 1835 |
if (NumIters > 1) { |
| 1836 |
CntAddr = CGF.CreateMemTemp(C.IntTy, ".cnt.addr"); |
| 1837 |
CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.IntTy), CntAddr, |
| 1838 |
/*Volatile=*/false, C.IntTy); |
| 1839 |
PrecondBB = CGF.createBasicBlock("precond"); |
| 1840 |
ExitBB = CGF.createBasicBlock("exit"); |
| 1841 |
llvm::BasicBlock *BodyBB = CGF.createBasicBlock("body"); |
| 1842 |
// There is no need to emit line number for unconditional branch. |
| 1843 |
(void)ApplyDebugLocation::CreateEmpty(CGF); |
| 1844 |
CGF.EmitBlock(PrecondBB); |
| 1845 |
Cnt = CGF.EmitLoadOfScalar(CntAddr, /*Volatile=*/false, C.IntTy, Loc); |
| 1846 |
llvm::Value *Cmp = |
| 1847 |
Bld.CreateICmpULT(Cnt, llvm::ConstantInt::get(CGM.IntTy, NumIters)); |
| 1848 |
Bld.CreateCondBr(Cmp, BodyBB, ExitBB); |
| 1849 |
CGF.EmitBlock(BodyBB); |
| 1850 |
} |
| 1851 |
// kmpc_barrier. |
| 1852 |
CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, |
| 1853 |
/*EmitChecks=*/false, |
| 1854 |
/*ForceSimpleCall=*/true); |
| 1855 |
llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); |
| 1856 |
llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); |
| 1857 |
llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); |
| 1858 |
|
| 1859 |
// if (lane_id == 0) |
| 1860 |
llvm::Value *IsWarpMaster = Bld.CreateIsNull(LaneID, "warp_master"); |
| 1861 |
Bld.CreateCondBr(IsWarpMaster, ThenBB, ElseBB); |
| 1862 |
CGF.EmitBlock(ThenBB); |
| 1863 |
|
| 1864 |
// Reduce element = LocalReduceList[i] |
| 1865 |
Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
| 1866 |
llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( |
| 1867 |
ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); |
| 1868 |
// elemptr = ((CopyType*)(elemptrptr)) + I |
| 1869 |
Address ElemPtr(ElemPtrPtr, CGF.Int8Ty, Align); |
| 1870 |
ElemPtr = Bld.CreateElementBitCast(ElemPtr, CopyType); |
| 1871 |
if (NumIters > 1) |
| 1872 |
ElemPtr = Bld.CreateGEP(ElemPtr, Cnt); |
| 1873 |
|
| 1874 |
// Get pointer to location in transfer medium. |
| 1875 |
// MediumPtr = &medium[warp_id] |
| 1876 |
llvm::Value *MediumPtrVal = Bld.CreateInBoundsGEP( |
| 1877 |
TransferMedium->getValueType(), TransferMedium, |
| 1878 |
{llvm::Constant::getNullValue(CGM.Int64Ty), WarpID}); |
| 1879 |
// Casting to actual data type. |
| 1880 |
// MediumPtr = (CopyType*)MediumPtrAddr; |
| 1881 |
Address MediumPtr( |
| 1882 |
Bld.CreateBitCast( |
| 1883 |
MediumPtrVal, |
| 1884 |
CopyType->getPointerTo( |
| 1885 |
MediumPtrVal->getType()->getPointerAddressSpace())), |
| 1886 |
CopyType, Align); |
| 1887 |
|
| 1888 |
// elem = *elemptr |
| 1889 |
//*MediumPtr = elem |
| 1890 |
llvm::Value *Elem = CGF.EmitLoadOfScalar( |
| 1891 |
ElemPtr, /*Volatile=*/false, CType, Loc, |
| 1892 |
LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); |
| 1893 |
// Store the source element value to the dest element address. |
| 1894 |
CGF.EmitStoreOfScalar(Elem, MediumPtr, /*Volatile=*/true, CType, |
| 1895 |
LValueBaseInfo(AlignmentSource::Type), |
| 1896 |
TBAAAccessInfo()); |
| 1897 |
|
| 1898 |
Bld.CreateBr(MergeBB); |
| 1899 |
|
| 1900 |
CGF.EmitBlock(ElseBB); |
| 1901 |
Bld.CreateBr(MergeBB); |
| 1902 |
|
| 1903 |
CGF.EmitBlock(MergeBB); |
| 1904 |
|
| 1905 |
// kmpc_barrier. |
| 1906 |
CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, |
| 1907 |
/*EmitChecks=*/false, |
| 1908 |
/*ForceSimpleCall=*/true); |
| 1909 |
|
| 1910 |
// |
| 1911 |
// Warp 0 copies reduce element from transfer medium. |
| 1912 |
// |
| 1913 |
llvm::BasicBlock *W0ThenBB = CGF.createBasicBlock("then"); |
| 1914 |
llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else"); |
| 1915 |
llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont"); |
| 1916 |
|
| 1917 |
Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg); |
| 1918 |
llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar( |
| 1919 |
AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc); |
| 1920 |
|
| 1921 |
// Up to 32 threads in warp 0 are active. |
| 1922 |
llvm::Value *IsActiveThread = |
| 1923 |
Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread"); |
| 1924 |
Bld.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB); |
| 1925 |
|
| 1926 |
CGF.EmitBlock(W0ThenBB); |
| 1927 |
|
| 1928 |
// SrcMediumPtr = &medium[tid] |
| 1929 |
llvm::Value *SrcMediumPtrVal = Bld.CreateInBoundsGEP( |
| 1930 |
TransferMedium->getValueType(), TransferMedium, |
| 1931 |
{llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID}); |
| 1932 |
// SrcMediumVal = *SrcMediumPtr; |
| 1933 |
Address SrcMediumPtr( |
| 1934 |
Bld.CreateBitCast( |
| 1935 |
SrcMediumPtrVal, |
| 1936 |
CopyType->getPointerTo( |
| 1937 |
SrcMediumPtrVal->getType()->getPointerAddressSpace())), |
| 1938 |
CopyType, Align); |
| 1939 |
|
| 1940 |
// TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I |
| 1941 |
Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
| 1942 |
llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar( |
| 1943 |
TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc); |
| 1944 |
Address TargetElemPtr(TargetElemPtrVal, CGF.Int8Ty, Align); |
| 1945 |
TargetElemPtr = Bld.CreateElementBitCast(TargetElemPtr, CopyType); |
| 1946 |
if (NumIters > 1) |
| 1947 |
TargetElemPtr = Bld.CreateGEP(TargetElemPtr, Cnt); |
| 1948 |
|
| 1949 |
// *TargetElemPtr = SrcMediumVal; |
| 1950 |
llvm::Value *SrcMediumValue = |
| 1951 |
CGF.EmitLoadOfScalar(SrcMediumPtr, /*Volatile=*/true, CType, Loc); |
| 1952 |
CGF.EmitStoreOfScalar(SrcMediumValue, TargetElemPtr, /*Volatile=*/false, |
| 1953 |
CType); |
| 1954 |
Bld.CreateBr(W0MergeBB); |
| 1955 |
|
| 1956 |
CGF.EmitBlock(W0ElseBB); |
| 1957 |
Bld.CreateBr(W0MergeBB); |
| 1958 |
|
| 1959 |
CGF.EmitBlock(W0MergeBB); |
| 1960 |
|
| 1961 |
if (NumIters > 1) { |
| 1962 |
Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1)); |
| 1963 |
CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy); |
| 1964 |
CGF.EmitBranch(PrecondBB); |
| 1965 |
(void)ApplyDebugLocation::CreateEmpty(CGF); |
| 1966 |
CGF.EmitBlock(ExitBB); |
| 1967 |
} |
| 1968 |
RealTySize %= TySize; |
| 1969 |
} |
| 1970 |
++Idx; |
| 1971 |
} |
| 1972 |
|
| 1973 |
CGF.FinishFunction(); |
| 1974 |
return Fn; |
| 1975 |
} |
| 1976 |
|
| 1977 |
/// Emit a helper that reduces data across two OpenMP threads (lanes) |
| 1978 |
/// in the same warp. It uses shuffle instructions to copy over data from |
| 1979 |
/// a remote lane's stack. The reduction algorithm performed is specified |
| 1980 |
/// by the fourth parameter. |
| 1981 |
/// |
| 1982 |
/// Algorithm Versions. |
| 1983 |
/// Full Warp Reduce (argument value 0): |
| 1984 |
/// This algorithm assumes that all 32 lanes are active and gathers |
| 1985 |
/// data from these 32 lanes, producing a single resultant value. |
| 1986 |
/// Contiguous Partial Warp Reduce (argument value 1): |
| 1987 |
/// This algorithm assumes that only a *contiguous* subset of lanes |
| 1988 |
/// are active. This happens for the last warp in a parallel region |
| 1989 |
/// when the user specified num_threads is not an integer multiple of |
| 1990 |
/// 32. This contiguous subset always starts with the zeroth lane. |
| 1991 |
/// Partial Warp Reduce (argument value 2): |
| 1992 |
/// This algorithm gathers data from any number of lanes at any position. |
| 1993 |
/// All reduced values are stored in the lowest possible lane. The set |
| 1994 |
/// of problems every algorithm addresses is a super set of those |
| 1995 |
/// addressable by algorithms with a lower version number. Overhead |
| 1996 |
/// increases as algorithm version increases. |
| 1997 |
/// |
| 1998 |
/// Terminology |
| 1999 |
/// Reduce element: |
| 2000 |
/// Reduce element refers to the individual data field with primitive |
| 2001 |
/// data types to be combined and reduced across threads. |
| 2002 |
/// Reduce list: |
| 2003 |
/// Reduce list refers to a collection of local, thread-private |
| 2004 |
/// reduce elements. |
| 2005 |
/// Remote Reduce list: |
| 2006 |
/// Remote Reduce list refers to a collection of remote (relative to |
| 2007 |
/// the current thread) reduce elements. |
| 2008 |
/// |
| 2009 |
/// We distinguish between three states of threads that are important to |
| 2010 |
/// the implementation of this function. |
| 2011 |
/// Alive threads: |
| 2012 |
/// Threads in a warp executing the SIMT instruction, as distinguished from |
| 2013 |
/// threads that are inactive due to divergent control flow. |
| 2014 |
/// Active threads: |
| 2015 |
/// The minimal set of threads that has to be alive upon entry to this |
| 2016 |
/// function. The computation is correct iff active threads are alive. |
| 2017 |
/// Some threads are alive but they are not active because they do not |
| 2018 |
/// contribute to the computation in any useful manner. Turning them off |
| 2019 |
/// may introduce control flow overheads without any tangible benefits. |
| 2020 |
/// Effective threads: |
| 2021 |
/// In order to comply with the argument requirements of the shuffle |
| 2022 |
/// function, we must keep all lanes holding data alive. But at most |
| 2023 |
/// half of them perform value aggregation; we refer to this half of |
| 2024 |
/// threads as effective. The other half is simply handing off their |
| 2025 |
/// data. |
| 2026 |
/// |
| 2027 |
/// Procedure |
| 2028 |
/// Value shuffle: |
| 2029 |
/// In this step active threads transfer data from higher lane positions |
| 2030 |
/// in the warp to lower lane positions, creating Remote Reduce list. |
| 2031 |
/// Value aggregation: |
| 2032 |
/// In this step, effective threads combine their thread local Reduce list |
| 2033 |
/// with Remote Reduce list and store the result in the thread local |
| 2034 |
/// Reduce list. |
| 2035 |
/// Value copy: |
| 2036 |
/// In this step, we deal with the assumption made by algorithm 2 |
| 2037 |
/// (i.e. contiguity assumption). When we have an odd number of lanes |
| 2038 |
/// active, say 2k+1, only k threads will be effective and therefore k |
| 2039 |
/// new values will be produced. However, the Reduce list owned by the |
| 2040 |
/// (2k+1)th thread is ignored in the value aggregation. Therefore |
| 2041 |
/// we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so |
| 2042 |
/// that the contiguity assumption still holds. |
| 2043 |
static llvm::Function *emitShuffleAndReduceFunction( |
| 2044 |
CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
| 2045 |
QualType ReductionArrayTy, llvm::Function *ReduceFn, SourceLocation Loc) { |
| 2046 |
ASTContext &C = CGM.getContext(); |
| 2047 |
|
| 2048 |
// Thread local Reduce list used to host the values of data to be reduced. |
| 2049 |
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2050 |
C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2051 |
// Current lane id; could be logical. |
| 2052 |
ImplicitParamDecl LaneIDArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.ShortTy, |
| 2053 |
ImplicitParamDecl::Other); |
| 2054 |
// Offset of the remote source lane relative to the current lane. |
| 2055 |
ImplicitParamDecl RemoteLaneOffsetArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2056 |
C.ShortTy, ImplicitParamDecl::Other); |
| 2057 |
// Algorithm version. This is expected to be known at compile time. |
| 2058 |
ImplicitParamDecl AlgoVerArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2059 |
C.ShortTy, ImplicitParamDecl::Other); |
| 2060 |
FunctionArgList Args; |
| 2061 |
Args.push_back(&ReduceListArg); |
| 2062 |
Args.push_back(&LaneIDArg); |
| 2063 |
Args.push_back(&RemoteLaneOffsetArg); |
| 2064 |
Args.push_back(&AlgoVerArg); |
| 2065 |
|
| 2066 |
const CGFunctionInfo &CGFI = |
| 2067 |
CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 2068 |
auto *Fn = llvm::Function::Create( |
| 2069 |
CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 2070 |
"_omp_reduction_shuffle_and_reduce_func", &CGM.getModule()); |
| 2071 |
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 2072 |
Fn->setDoesNotRecurse(); |
| 2073 |
|
| 2074 |
CodeGenFunction CGF(CGM); |
| 2075 |
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 2076 |
|
| 2077 |
CGBuilderTy &Bld = CGF.Builder; |
| 2078 |
|
| 2079 |
Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 2080 |
llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
| 2081 |
Address LocalReduceList( |
| 2082 |
Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2083 |
CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, |
| 2084 |
C.VoidPtrTy, SourceLocation()), |
| 2085 |
ElemTy->getPointerTo()), |
| 2086 |
ElemTy, CGF.getPointerAlign()); |
| 2087 |
|
| 2088 |
Address AddrLaneIDArg = CGF.GetAddrOfLocalVar(&LaneIDArg); |
| 2089 |
llvm::Value *LaneIDArgVal = CGF.EmitLoadOfScalar( |
| 2090 |
AddrLaneIDArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); |
| 2091 |
|
| 2092 |
Address AddrRemoteLaneOffsetArg = CGF.GetAddrOfLocalVar(&RemoteLaneOffsetArg); |
| 2093 |
llvm::Value *RemoteLaneOffsetArgVal = CGF.EmitLoadOfScalar( |
| 2094 |
AddrRemoteLaneOffsetArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); |
| 2095 |
|
| 2096 |
Address AddrAlgoVerArg = CGF.GetAddrOfLocalVar(&AlgoVerArg); |
| 2097 |
llvm::Value *AlgoVerArgVal = CGF.EmitLoadOfScalar( |
| 2098 |
AddrAlgoVerArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); |
| 2099 |
|
| 2100 |
// Create a local thread-private variable to host the Reduce list |
| 2101 |
// from a remote lane. |
| 2102 |
Address RemoteReduceList = |
| 2103 |
CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.remote_reduce_list"); |
| 2104 |
|
| 2105 |
// This loop iterates through the list of reduce elements and copies, |
| 2106 |
// element by element, from a remote lane in the warp to RemoteReduceList, |
| 2107 |
// hosted on the thread's stack. |
| 2108 |
emitReductionListCopy(RemoteLaneToThread, CGF, ReductionArrayTy, Privates, |
| 2109 |
LocalReduceList, RemoteReduceList, |
| 2110 |
{/*RemoteLaneOffset=*/RemoteLaneOffsetArgVal, |
| 2111 |
/*ScratchpadIndex=*/nullptr, |
| 2112 |
/*ScratchpadWidth=*/nullptr}); |
| 2113 |
|
| 2114 |
// The actions to be performed on the Remote Reduce list is dependent |
| 2115 |
// on the algorithm version. |
| 2116 |
// |
| 2117 |
// if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 && |
| 2118 |
// LaneId % 2 == 0 && Offset > 0): |
| 2119 |
// do the reduction value aggregation |
| 2120 |
// |
| 2121 |
// The thread local variable Reduce list is mutated in place to host the |
| 2122 |
// reduced data, which is the aggregated value produced from local and |
| 2123 |
// remote lanes. |
| 2124 |
// |
| 2125 |
// Note that AlgoVer is expected to be a constant integer known at compile |
| 2126 |
// time. |
| 2127 |
// When AlgoVer==0, the first conjunction evaluates to true, making |
| 2128 |
// the entire predicate true during compile time. |
| 2129 |
// When AlgoVer==1, the second conjunction has only the second part to be |
| 2130 |
// evaluated during runtime. Other conjunctions evaluates to false |
| 2131 |
// during compile time. |
| 2132 |
// When AlgoVer==2, the third conjunction has only the second part to be |
| 2133 |
// evaluated during runtime. Other conjunctions evaluates to false |
| 2134 |
// during compile time. |
| 2135 |
llvm::Value *CondAlgo0 = Bld.CreateIsNull(AlgoVerArgVal); |
| 2136 |
|
| 2137 |
llvm::Value *Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); |
| 2138 |
llvm::Value *CondAlgo1 = Bld.CreateAnd( |
| 2139 |
Algo1, Bld.CreateICmpULT(LaneIDArgVal, RemoteLaneOffsetArgVal)); |
| 2140 |
|
| 2141 |
llvm::Value *Algo2 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(2)); |
| 2142 |
llvm::Value *CondAlgo2 = Bld.CreateAnd( |
| 2143 |
Algo2, Bld.CreateIsNull(Bld.CreateAnd(LaneIDArgVal, Bld.getInt16(1)))); |
| 2144 |
CondAlgo2 = Bld.CreateAnd( |
| 2145 |
CondAlgo2, Bld.CreateICmpSGT(RemoteLaneOffsetArgVal, Bld.getInt16(0))); |
| 2146 |
|
| 2147 |
llvm::Value *CondReduce = Bld.CreateOr(CondAlgo0, CondAlgo1); |
| 2148 |
CondReduce = Bld.CreateOr(CondReduce, CondAlgo2); |
| 2149 |
|
| 2150 |
llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); |
| 2151 |
llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); |
| 2152 |
llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); |
| 2153 |
Bld.CreateCondBr(CondReduce, ThenBB, ElseBB); |
| 2154 |
|
| 2155 |
CGF.EmitBlock(ThenBB); |
| 2156 |
// reduce_function(LocalReduceList, RemoteReduceList) |
| 2157 |
llvm::Value *LocalReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2158 |
LocalReduceList.getPointer(), CGF.VoidPtrTy); |
| 2159 |
llvm::Value *RemoteReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2160 |
RemoteReduceList.getPointer(), CGF.VoidPtrTy); |
| 2161 |
CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
| 2162 |
CGF, Loc, ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr}); |
| 2163 |
Bld.CreateBr(MergeBB); |
| 2164 |
|
| 2165 |
CGF.EmitBlock(ElseBB); |
| 2166 |
Bld.CreateBr(MergeBB); |
| 2167 |
|
| 2168 |
CGF.EmitBlock(MergeBB); |
| 2169 |
|
| 2170 |
// if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local |
| 2171 |
// Reduce list. |
| 2172 |
Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); |
| 2173 |
llvm::Value *CondCopy = Bld.CreateAnd( |
| 2174 |
Algo1, Bld.CreateICmpUGE(LaneIDArgVal, RemoteLaneOffsetArgVal)); |
| 2175 |
|
| 2176 |
llvm::BasicBlock *CpyThenBB = CGF.createBasicBlock("then"); |
| 2177 |
llvm::BasicBlock *CpyElseBB = CGF.createBasicBlock("else"); |
| 2178 |
llvm::BasicBlock *CpyMergeBB = CGF.createBasicBlock("ifcont"); |
| 2179 |
Bld.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB); |
| 2180 |
|
| 2181 |
CGF.EmitBlock(CpyThenBB); |
| 2182 |
emitReductionListCopy(ThreadCopy, CGF, ReductionArrayTy, Privates, |
| 2183 |
RemoteReduceList, LocalReduceList); |
| 2184 |
Bld.CreateBr(CpyMergeBB); |
| 2185 |
|
| 2186 |
CGF.EmitBlock(CpyElseBB); |
| 2187 |
Bld.CreateBr(CpyMergeBB); |
| 2188 |
|
| 2189 |
CGF.EmitBlock(CpyMergeBB); |
| 2190 |
|
| 2191 |
CGF.FinishFunction(); |
| 2192 |
return Fn; |
| 2193 |
} |
| 2194 |
|
| 2195 |
/// This function emits a helper that copies all the reduction variables from |
| 2196 |
/// the team into the provided global buffer for the reduction variables. |
| 2197 |
/// |
| 2198 |
/// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) |
| 2199 |
/// For all data entries D in reduce_data: |
| 2200 |
/// Copy local D to buffer.D[Idx] |
| 2201 |
static llvm::Value *emitListToGlobalCopyFunction( |
| 2202 |
CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
| 2203 |
QualType ReductionArrayTy, SourceLocation Loc, |
| 2204 |
const RecordDecl *TeamReductionRec, |
| 2205 |
const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 2206 |
&VarFieldMap) { |
| 2207 |
ASTContext &C = CGM.getContext(); |
| 2208 |
|
| 2209 |
// Buffer: global reduction buffer. |
| 2210 |
ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2211 |
C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2212 |
// Idx: index of the buffer. |
| 2213 |
ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
| 2214 |
ImplicitParamDecl::Other); |
| 2215 |
// ReduceList: thread local Reduce list. |
| 2216 |
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2217 |
C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2218 |
FunctionArgList Args; |
| 2219 |
Args.push_back(&BufferArg); |
| 2220 |
Args.push_back(&IdxArg); |
| 2221 |
Args.push_back(&ReduceListArg); |
| 2222 |
|
| 2223 |
const CGFunctionInfo &CGFI = |
| 2224 |
CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 2225 |
auto *Fn = llvm::Function::Create( |
| 2226 |
CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 2227 |
"_omp_reduction_list_to_global_copy_func", &CGM.getModule()); |
| 2228 |
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 2229 |
Fn->setDoesNotRecurse(); |
| 2230 |
CodeGenFunction CGF(CGM); |
| 2231 |
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 2232 |
|
| 2233 |
CGBuilderTy &Bld = CGF.Builder; |
| 2234 |
|
| 2235 |
Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 2236 |
Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
| 2237 |
llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
| 2238 |
Address LocalReduceList( |
| 2239 |
Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2240 |
CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, |
| 2241 |
C.VoidPtrTy, Loc), |
| 2242 |
ElemTy->getPointerTo()), |
| 2243 |
ElemTy, CGF.getPointerAlign()); |
| 2244 |
QualType StaticTy = C.getRecordType(TeamReductionRec); |
| 2245 |
llvm::Type *LLVMReductionsBufferTy = |
| 2246 |
CGM.getTypes().ConvertTypeForMem(StaticTy); |
| 2247 |
llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2248 |
CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
| 2249 |
LLVMReductionsBufferTy->getPointerTo()); |
| 2250 |
llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
| 2251 |
CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
| 2252 |
/*Volatile=*/false, C.IntTy, |
| 2253 |
Loc)}; |
| 2254 |
unsigned Idx = 0; |
| 2255 |
for (const Expr *Private : Privates) { |
| 2256 |
// Reduce element = LocalReduceList[i] |
| 2257 |
Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
| 2258 |
llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( |
| 2259 |
ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); |
| 2260 |
// elemptr = ((CopyType*)(elemptrptr)) + I |
| 2261 |
ElemTy = CGF.ConvertTypeForMem(Private->getType()); |
| 2262 |
ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2263 |
ElemPtrPtr, ElemTy->getPointerTo()); |
| 2264 |
Address ElemPtr = |
| 2265 |
Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType())); |
| 2266 |
const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); |
| 2267 |
// Global = Buffer.VD[Idx]; |
| 2268 |
const FieldDecl *FD = VarFieldMap.lookup(VD); |
| 2269 |
LValue GlobLVal = CGF.EmitLValueForField( |
| 2270 |
CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
| 2271 |
Address GlobAddr = GlobLVal.getAddress(CGF); |
| 2272 |
llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(), |
| 2273 |
GlobAddr.getPointer(), Idxs); |
| 2274 |
GlobLVal.setAddress(Address(BufferPtr, |
| 2275 |
CGF.ConvertTypeForMem(Private->getType()), |
| 2276 |
GlobAddr.getAlignment())); |
| 2277 |
switch (CGF.getEvaluationKind(Private->getType())) { |
| 2278 |
case TEK_Scalar: { |
| 2279 |
llvm::Value *V = CGF.EmitLoadOfScalar( |
| 2280 |
ElemPtr, /*Volatile=*/false, Private->getType(), Loc, |
| 2281 |
LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); |
| 2282 |
CGF.EmitStoreOfScalar(V, GlobLVal); |
| 2283 |
break; |
| 2284 |
} |
| 2285 |
case TEK_Complex: { |
| 2286 |
CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex( |
| 2287 |
CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc); |
| 2288 |
CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false); |
| 2289 |
break; |
| 2290 |
} |
| 2291 |
case TEK_Aggregate: |
| 2292 |
CGF.EmitAggregateCopy(GlobLVal, |
| 2293 |
CGF.MakeAddrLValue(ElemPtr, Private->getType()), |
| 2294 |
Private->getType(), AggValueSlot::DoesNotOverlap); |
| 2295 |
break; |
| 2296 |
} |
| 2297 |
++Idx; |
| 2298 |
} |
| 2299 |
|
| 2300 |
CGF.FinishFunction(); |
| 2301 |
return Fn; |
| 2302 |
} |
| 2303 |
|
| 2304 |
/// This function emits a helper that reduces all the reduction variables from |
| 2305 |
/// the team into the provided global buffer for the reduction variables. |
| 2306 |
/// |
| 2307 |
/// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data) |
| 2308 |
/// void *GlobPtrs[]; |
| 2309 |
/// GlobPtrs[0] = (void*)&buffer.D0[Idx]; |
| 2310 |
/// ... |
| 2311 |
/// GlobPtrs[N] = (void*)&buffer.DN[Idx]; |
| 2312 |
/// reduce_function(GlobPtrs, reduce_data); |
| 2313 |
static llvm::Value *emitListToGlobalReduceFunction( |
| 2314 |
CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
| 2315 |
QualType ReductionArrayTy, SourceLocation Loc, |
| 2316 |
const RecordDecl *TeamReductionRec, |
| 2317 |
const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 2318 |
&VarFieldMap, |
| 2319 |
llvm::Function *ReduceFn) { |
| 2320 |
ASTContext &C = CGM.getContext(); |
| 2321 |
|
| 2322 |
// Buffer: global reduction buffer. |
| 2323 |
ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2324 |
C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2325 |
// Idx: index of the buffer. |
| 2326 |
ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
| 2327 |
ImplicitParamDecl::Other); |
| 2328 |
// ReduceList: thread local Reduce list. |
| 2329 |
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2330 |
C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2331 |
FunctionArgList Args; |
| 2332 |
Args.push_back(&BufferArg); |
| 2333 |
Args.push_back(&IdxArg); |
| 2334 |
Args.push_back(&ReduceListArg); |
| 2335 |
|
| 2336 |
const CGFunctionInfo &CGFI = |
| 2337 |
CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 2338 |
auto *Fn = llvm::Function::Create( |
| 2339 |
CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 2340 |
"_omp_reduction_list_to_global_reduce_func", &CGM.getModule()); |
| 2341 |
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 2342 |
Fn->setDoesNotRecurse(); |
| 2343 |
CodeGenFunction CGF(CGM); |
| 2344 |
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 2345 |
|
| 2346 |
CGBuilderTy &Bld = CGF.Builder; |
| 2347 |
|
| 2348 |
Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
| 2349 |
QualType StaticTy = C.getRecordType(TeamReductionRec); |
| 2350 |
llvm::Type *LLVMReductionsBufferTy = |
| 2351 |
CGM.getTypes().ConvertTypeForMem(StaticTy); |
| 2352 |
llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2353 |
CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
| 2354 |
LLVMReductionsBufferTy->getPointerTo()); |
| 2355 |
|
| 2356 |
// 1. Build a list of reduction variables. |
| 2357 |
// void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
| 2358 |
Address ReductionList = |
| 2359 |
CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); |
| 2360 |
auto IPriv = Privates.begin(); |
| 2361 |
llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
| 2362 |
CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
| 2363 |
/*Volatile=*/false, C.IntTy, |
| 2364 |
Loc)}; |
| 2365 |
unsigned Idx = 0; |
| 2366 |
for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) { |
| 2367 |
Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 2368 |
// Global = Buffer.VD[Idx]; |
| 2369 |
const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); |
| 2370 |
const FieldDecl *FD = VarFieldMap.lookup(VD); |
| 2371 |
LValue GlobLVal = CGF.EmitLValueForField( |
| 2372 |
CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
| 2373 |
Address GlobAddr = GlobLVal.getAddress(CGF); |
| 2374 |
llvm::Value *BufferPtr = Bld.CreateInBoundsGEP( |
| 2375 |
GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs); |
| 2376 |
llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr); |
| 2377 |
CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy); |
| 2378 |
if ((*IPriv)->getType()->isVariablyModifiedType()) { |
| 2379 |
// Store array size. |
| 2380 |
++Idx; |
| 2381 |
Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 2382 |
llvm::Value *Size = CGF.Builder.CreateIntCast( |
| 2383 |
CGF.getVLASize( |
| 2384 |
CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) |
| 2385 |
.NumElts, |
| 2386 |
CGF.SizeTy, /*isSigned=*/false); |
| 2387 |
CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), |
| 2388 |
Elem); |
| 2389 |
} |
| 2390 |
} |
| 2391 |
|
| 2392 |
// Call reduce_function(GlobalReduceList, ReduceList) |
| 2393 |
llvm::Value *GlobalReduceList = |
| 2394 |
CGF.EmitCastToVoidPtr(ReductionList.getPointer()); |
| 2395 |
Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 2396 |
llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( |
| 2397 |
AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); |
| 2398 |
CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
| 2399 |
CGF, Loc, ReduceFn, {GlobalReduceList, ReducedPtr}); |
| 2400 |
CGF.FinishFunction(); |
| 2401 |
return Fn; |
| 2402 |
} |
| 2403 |
|
| 2404 |
/// This function emits a helper that copies all the reduction variables from |
| 2405 |
/// the team into the provided global buffer for the reduction variables. |
| 2406 |
/// |
| 2407 |
/// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) |
| 2408 |
/// For all data entries D in reduce_data: |
| 2409 |
/// Copy buffer.D[Idx] to local D; |
| 2410 |
static llvm::Value *emitGlobalToListCopyFunction( |
| 2411 |
CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
| 2412 |
QualType ReductionArrayTy, SourceLocation Loc, |
| 2413 |
const RecordDecl *TeamReductionRec, |
| 2414 |
const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 2415 |
&VarFieldMap) { |
| 2416 |
ASTContext &C = CGM.getContext(); |
| 2417 |
|
| 2418 |
// Buffer: global reduction buffer. |
| 2419 |
ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2420 |
C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2421 |
// Idx: index of the buffer. |
| 2422 |
ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
| 2423 |
ImplicitParamDecl::Other); |
| 2424 |
// ReduceList: thread local Reduce list. |
| 2425 |
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2426 |
C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2427 |
FunctionArgList Args; |
| 2428 |
Args.push_back(&BufferArg); |
| 2429 |
Args.push_back(&IdxArg); |
| 2430 |
Args.push_back(&ReduceListArg); |
| 2431 |
|
| 2432 |
const CGFunctionInfo &CGFI = |
| 2433 |
CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 2434 |
auto *Fn = llvm::Function::Create( |
| 2435 |
CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 2436 |
"_omp_reduction_global_to_list_copy_func", &CGM.getModule()); |
| 2437 |
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 2438 |
Fn->setDoesNotRecurse(); |
| 2439 |
CodeGenFunction CGF(CGM); |
| 2440 |
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 2441 |
|
| 2442 |
CGBuilderTy &Bld = CGF.Builder; |
| 2443 |
|
| 2444 |
Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 2445 |
Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
| 2446 |
llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
| 2447 |
Address LocalReduceList( |
| 2448 |
Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2449 |
CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, |
| 2450 |
C.VoidPtrTy, Loc), |
| 2451 |
ElemTy->getPointerTo()), |
| 2452 |
ElemTy, CGF.getPointerAlign()); |
| 2453 |
QualType StaticTy = C.getRecordType(TeamReductionRec); |
| 2454 |
llvm::Type *LLVMReductionsBufferTy = |
| 2455 |
CGM.getTypes().ConvertTypeForMem(StaticTy); |
| 2456 |
llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2457 |
CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
| 2458 |
LLVMReductionsBufferTy->getPointerTo()); |
| 2459 |
|
| 2460 |
llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
| 2461 |
CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
| 2462 |
/*Volatile=*/false, C.IntTy, |
| 2463 |
Loc)}; |
| 2464 |
unsigned Idx = 0; |
| 2465 |
for (const Expr *Private : Privates) { |
| 2466 |
// Reduce element = LocalReduceList[i] |
| 2467 |
Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
| 2468 |
llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( |
| 2469 |
ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); |
| 2470 |
// elemptr = ((CopyType*)(elemptrptr)) + I |
| 2471 |
ElemTy = CGF.ConvertTypeForMem(Private->getType()); |
| 2472 |
ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2473 |
ElemPtrPtr, ElemTy->getPointerTo()); |
| 2474 |
Address ElemPtr = |
| 2475 |
Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType())); |
| 2476 |
const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); |
| 2477 |
// Global = Buffer.VD[Idx]; |
| 2478 |
const FieldDecl *FD = VarFieldMap.lookup(VD); |
| 2479 |
LValue GlobLVal = CGF.EmitLValueForField( |
| 2480 |
CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
| 2481 |
Address GlobAddr = GlobLVal.getAddress(CGF); |
| 2482 |
llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(), |
| 2483 |
GlobAddr.getPointer(), Idxs); |
| 2484 |
GlobLVal.setAddress(Address(BufferPtr, |
| 2485 |
CGF.ConvertTypeForMem(Private->getType()), |
| 2486 |
GlobAddr.getAlignment())); |
| 2487 |
switch (CGF.getEvaluationKind(Private->getType())) { |
| 2488 |
case TEK_Scalar: { |
| 2489 |
llvm::Value *V = CGF.EmitLoadOfScalar(GlobLVal, Loc); |
| 2490 |
CGF.EmitStoreOfScalar(V, ElemPtr, /*Volatile=*/false, Private->getType(), |
| 2491 |
LValueBaseInfo(AlignmentSource::Type), |
| 2492 |
TBAAAccessInfo()); |
| 2493 |
break; |
| 2494 |
} |
| 2495 |
case TEK_Complex: { |
| 2496 |
CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc); |
| 2497 |
CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()), |
| 2498 |
/*isInit=*/false); |
| 2499 |
break; |
| 2500 |
} |
| 2501 |
case TEK_Aggregate: |
| 2502 |
CGF.EmitAggregateCopy(CGF.MakeAddrLValue(ElemPtr, Private->getType()), |
| 2503 |
GlobLVal, Private->getType(), |
| 2504 |
AggValueSlot::DoesNotOverlap); |
| 2505 |
break; |
| 2506 |
} |
| 2507 |
++Idx; |
| 2508 |
} |
| 2509 |
|
| 2510 |
CGF.FinishFunction(); |
| 2511 |
return Fn; |
| 2512 |
} |
| 2513 |
|
| 2514 |
/// This function emits a helper that reduces all the reduction variables from |
| 2515 |
/// the team into the provided global buffer for the reduction variables. |
| 2516 |
/// |
| 2517 |
/// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data) |
| 2518 |
/// void *GlobPtrs[]; |
| 2519 |
/// GlobPtrs[0] = (void*)&buffer.D0[Idx]; |
| 2520 |
/// ... |
| 2521 |
/// GlobPtrs[N] = (void*)&buffer.DN[Idx]; |
| 2522 |
/// reduce_function(reduce_data, GlobPtrs); |
| 2523 |
static llvm::Value *emitGlobalToListReduceFunction( |
| 2524 |
CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
| 2525 |
QualType ReductionArrayTy, SourceLocation Loc, |
| 2526 |
const RecordDecl *TeamReductionRec, |
| 2527 |
const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 2528 |
&VarFieldMap, |
| 2529 |
llvm::Function *ReduceFn) { |
| 2530 |
ASTContext &C = CGM.getContext(); |
| 2531 |
|
| 2532 |
// Buffer: global reduction buffer. |
| 2533 |
ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2534 |
C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2535 |
// Idx: index of the buffer. |
| 2536 |
ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
| 2537 |
ImplicitParamDecl::Other); |
| 2538 |
// ReduceList: thread local Reduce list. |
| 2539 |
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2540 |
C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2541 |
FunctionArgList Args; |
| 2542 |
Args.push_back(&BufferArg); |
| 2543 |
Args.push_back(&IdxArg); |
| 2544 |
Args.push_back(&ReduceListArg); |
| 2545 |
|
| 2546 |
const CGFunctionInfo &CGFI = |
| 2547 |
CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 2548 |
auto *Fn = llvm::Function::Create( |
| 2549 |
CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 2550 |
"_omp_reduction_global_to_list_reduce_func", &CGM.getModule()); |
| 2551 |
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 2552 |
Fn->setDoesNotRecurse(); |
| 2553 |
CodeGenFunction CGF(CGM); |
| 2554 |
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 2555 |
|
| 2556 |
CGBuilderTy &Bld = CGF.Builder; |
| 2557 |
|
| 2558 |
Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
| 2559 |
QualType StaticTy = C.getRecordType(TeamReductionRec); |
| 2560 |
llvm::Type *LLVMReductionsBufferTy = |
| 2561 |
CGM.getTypes().ConvertTypeForMem(StaticTy); |
| 2562 |
llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2563 |
CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
| 2564 |
LLVMReductionsBufferTy->getPointerTo()); |
| 2565 |
|
| 2566 |
// 1. Build a list of reduction variables. |
| 2567 |
// void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
| 2568 |
Address ReductionList = |
| 2569 |
CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); |
| 2570 |
auto IPriv = Privates.begin(); |
| 2571 |
llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
| 2572 |
CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
| 2573 |
/*Volatile=*/false, C.IntTy, |
| 2574 |
Loc)}; |
| 2575 |
unsigned Idx = 0; |
| 2576 |
for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) { |
| 2577 |
Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 2578 |
// Global = Buffer.VD[Idx]; |
| 2579 |
const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); |
| 2580 |
const FieldDecl *FD = VarFieldMap.lookup(VD); |
| 2581 |
LValue GlobLVal = CGF.EmitLValueForField( |
| 2582 |
CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
| 2583 |
Address GlobAddr = GlobLVal.getAddress(CGF); |
| 2584 |
llvm::Value *BufferPtr = Bld.CreateInBoundsGEP( |
| 2585 |
GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs); |
| 2586 |
llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr); |
| 2587 |
CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy); |
| 2588 |
if ((*IPriv)->getType()->isVariablyModifiedType()) { |
| 2589 |
// Store array size. |
| 2590 |
++Idx; |
| 2591 |
Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 2592 |
llvm::Value *Size = CGF.Builder.CreateIntCast( |
| 2593 |
CGF.getVLASize( |
| 2594 |
CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) |
| 2595 |
.NumElts, |
| 2596 |
CGF.SizeTy, /*isSigned=*/false); |
| 2597 |
CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), |
| 2598 |
Elem); |
| 2599 |
} |
| 2600 |
} |
| 2601 |
|
| 2602 |
// Call reduce_function(ReduceList, GlobalReduceList) |
| 2603 |
llvm::Value *GlobalReduceList = |
| 2604 |
CGF.EmitCastToVoidPtr(ReductionList.getPointer()); |
| 2605 |
Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 2606 |
llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( |
| 2607 |
AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); |
| 2608 |
CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
| 2609 |
CGF, Loc, ReduceFn, {ReducedPtr, GlobalReduceList}); |
| 2610 |
CGF.FinishFunction(); |
| 2611 |
return Fn; |
| 2612 |
} |
| 2613 |
|
| 2614 |
/// |
| 2615 |
/// Design of OpenMP reductions on the GPU |
| 2616 |
/// |
| 2617 |
/// Consider a typical OpenMP program with one or more reduction |
| 2618 |
/// clauses: |
| 2619 |
/// |
| 2620 |
/// float foo; |
| 2621 |
/// double bar; |
| 2622 |
/// #pragma omp target teams distribute parallel for \ |
| 2623 |
/// reduction(+:foo) reduction(*:bar) |
| 2624 |
/// for (int i = 0; i < N; i++) { |
| 2625 |
/// foo += A[i]; bar *= B[i]; |
| 2626 |
/// } |
| 2627 |
/// |
| 2628 |
/// where 'foo' and 'bar' are reduced across all OpenMP threads in |
| 2629 |
/// all teams. In our OpenMP implementation on the NVPTX device an |
| 2630 |
/// OpenMP team is mapped to a CUDA threadblock and OpenMP threads |
| 2631 |
/// within a team are mapped to CUDA threads within a threadblock. |
| 2632 |
/// Our goal is to efficiently aggregate values across all OpenMP |
| 2633 |
/// threads such that: |
| 2634 |
/// |
| 2635 |
/// - the compiler and runtime are logically concise, and |
| 2636 |
/// - the reduction is performed efficiently in a hierarchical |
| 2637 |
/// manner as follows: within OpenMP threads in the same warp, |
| 2638 |
/// across warps in a threadblock, and finally across teams on |
| 2639 |
/// the NVPTX device. |
| 2640 |
/// |
| 2641 |
/// Introduction to Decoupling |
| 2642 |
/// |
| 2643 |
/// We would like to decouple the compiler and the runtime so that the |
| 2644 |
/// latter is ignorant of the reduction variables (number, data types) |
| 2645 |
/// and the reduction operators. This allows a simpler interface |
| 2646 |
/// and implementation while still attaining good performance. |
| 2647 |
/// |
| 2648 |
/// Pseudocode for the aforementioned OpenMP program generated by the |
| 2649 |
/// compiler is as follows: |
| 2650 |
/// |
| 2651 |
/// 1. Create private copies of reduction variables on each OpenMP |
| 2652 |
/// thread: 'foo_private', 'bar_private' |
| 2653 |
/// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned |
| 2654 |
/// to it and writes the result in 'foo_private' and 'bar_private' |
| 2655 |
/// respectively. |
| 2656 |
/// 3. Call the OpenMP runtime on the GPU to reduce within a team |
| 2657 |
/// and store the result on the team master: |
| 2658 |
/// |
| 2659 |
/// __kmpc_nvptx_parallel_reduce_nowait_v2(..., |
| 2660 |
/// reduceData, shuffleReduceFn, interWarpCpyFn) |
| 2661 |
/// |
| 2662 |
/// where: |
| 2663 |
/// struct ReduceData { |
| 2664 |
/// double *foo; |
| 2665 |
/// double *bar; |
| 2666 |
/// } reduceData |
| 2667 |
/// reduceData.foo = &foo_private |
| 2668 |
/// reduceData.bar = &bar_private |
| 2669 |
/// |
| 2670 |
/// 'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two |
| 2671 |
/// auxiliary functions generated by the compiler that operate on |
| 2672 |
/// variables of type 'ReduceData'. They aid the runtime perform |
| 2673 |
/// algorithmic steps in a data agnostic manner. |
| 2674 |
/// |
| 2675 |
/// 'shuffleReduceFn' is a pointer to a function that reduces data |
| 2676 |
/// of type 'ReduceData' across two OpenMP threads (lanes) in the |
| 2677 |
/// same warp. It takes the following arguments as input: |
| 2678 |
/// |
| 2679 |
/// a. variable of type 'ReduceData' on the calling lane, |
| 2680 |
/// b. its lane_id, |
| 2681 |
/// c. an offset relative to the current lane_id to generate a |
| 2682 |
/// remote_lane_id. The remote lane contains the second |
| 2683 |
/// variable of type 'ReduceData' that is to be reduced. |
| 2684 |
/// d. an algorithm version parameter determining which reduction |
| 2685 |
/// algorithm to use. |
| 2686 |
/// |
| 2687 |
/// 'shuffleReduceFn' retrieves data from the remote lane using |
| 2688 |
/// efficient GPU shuffle intrinsics and reduces, using the |
| 2689 |
/// algorithm specified by the 4th parameter, the two operands |
| 2690 |
/// element-wise. The result is written to the first operand. |
| 2691 |
/// |
| 2692 |
/// Different reduction algorithms are implemented in different |
| 2693 |
/// runtime functions, all calling 'shuffleReduceFn' to perform |
| 2694 |
/// the essential reduction step. Therefore, based on the 4th |
| 2695 |
/// parameter, this function behaves slightly differently to |
| 2696 |
/// cooperate with the runtime to ensure correctness under |
| 2697 |
/// different circumstances. |
| 2698 |
/// |
| 2699 |
/// 'InterWarpCpyFn' is a pointer to a function that transfers |
| 2700 |
/// reduced variables across warps. It tunnels, through CUDA |
| 2701 |
/// shared memory, the thread-private data of type 'ReduceData' |
| 2702 |
/// from lane 0 of each warp to a lane in the first warp. |
| 2703 |
/// 4. Call the OpenMP runtime on the GPU to reduce across teams. |
| 2704 |
/// The last team writes the global reduced value to memory. |
| 2705 |
/// |
| 2706 |
/// ret = __kmpc_nvptx_teams_reduce_nowait(..., |
| 2707 |
/// reduceData, shuffleReduceFn, interWarpCpyFn, |
| 2708 |
/// scratchpadCopyFn, loadAndReduceFn) |
| 2709 |
/// |
| 2710 |
/// 'scratchpadCopyFn' is a helper that stores reduced |
| 2711 |
/// data from the team master to a scratchpad array in |
| 2712 |
/// global memory. |
| 2713 |
/// |
| 2714 |
/// 'loadAndReduceFn' is a helper that loads data from |
| 2715 |
/// the scratchpad array and reduces it with the input |
| 2716 |
/// operand. |
| 2717 |
/// |
| 2718 |
/// These compiler generated functions hide address |
| 2719 |
/// calculation and alignment information from the runtime. |
| 2720 |
/// 5. if ret == 1: |
| 2721 |
/// The team master of the last team stores the reduced |
| 2722 |
/// result to the globals in memory. |
| 2723 |
/// foo += reduceData.foo; bar *= reduceData.bar |
| 2724 |
/// |
| 2725 |
/// |
| 2726 |
/// Warp Reduction Algorithms |
| 2727 |
/// |
| 2728 |
/// On the warp level, we have three algorithms implemented in the |
| 2729 |
/// OpenMP runtime depending on the number of active lanes: |
| 2730 |
/// |
| 2731 |
/// Full Warp Reduction |
| 2732 |
/// |
| 2733 |
/// The reduce algorithm within a warp where all lanes are active |
| 2734 |
/// is implemented in the runtime as follows: |
| 2735 |
/// |
| 2736 |
/// full_warp_reduce(void *reduce_data, |
| 2737 |
/// kmp_ShuffleReductFctPtr ShuffleReduceFn) { |
| 2738 |
/// for (int offset = WARPSIZE/2; offset > 0; offset /= 2) |
| 2739 |
/// ShuffleReduceFn(reduce_data, 0, offset, 0); |
| 2740 |
/// } |
| 2741 |
/// |
| 2742 |
/// The algorithm completes in log(2, WARPSIZE) steps. |
| 2743 |
/// |
| 2744 |
/// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is |
| 2745 |
/// not used therefore we save instructions by not retrieving lane_id |
| 2746 |
/// from the corresponding special registers. The 4th parameter, which |
| 2747 |
/// represents the version of the algorithm being used, is set to 0 to |
| 2748 |
/// signify full warp reduction. |
| 2749 |
/// |
| 2750 |
/// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
| 2751 |
/// |
| 2752 |
/// #reduce_elem refers to an element in the local lane's data structure |
| 2753 |
/// #remote_elem is retrieved from a remote lane |
| 2754 |
/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
| 2755 |
/// reduce_elem = reduce_elem REDUCE_OP remote_elem; |
| 2756 |
/// |
| 2757 |
/// Contiguous Partial Warp Reduction |
| 2758 |
/// |
| 2759 |
/// This reduce algorithm is used within a warp where only the first |
| 2760 |
/// 'n' (n <= WARPSIZE) lanes are active. It is typically used when the |
| 2761 |
/// number of OpenMP threads in a parallel region is not a multiple of |
| 2762 |
/// WARPSIZE. The algorithm is implemented in the runtime as follows: |
| 2763 |
/// |
| 2764 |
/// void |
| 2765 |
/// contiguous_partial_reduce(void *reduce_data, |
| 2766 |
/// kmp_ShuffleReductFctPtr ShuffleReduceFn, |
| 2767 |
/// int size, int lane_id) { |
| 2768 |
/// int curr_size; |
| 2769 |
/// int offset; |
| 2770 |
/// curr_size = size; |
| 2771 |
/// mask = curr_size/2; |
| 2772 |
/// while (offset>0) { |
| 2773 |
/// ShuffleReduceFn(reduce_data, lane_id, offset, 1); |
| 2774 |
/// curr_size = (curr_size+1)/2; |
| 2775 |
/// offset = curr_size/2; |
| 2776 |
/// } |
| 2777 |
/// } |
| 2778 |
/// |
| 2779 |
/// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
| 2780 |
/// |
| 2781 |
/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
| 2782 |
/// if (lane_id < offset) |
| 2783 |
/// reduce_elem = reduce_elem REDUCE_OP remote_elem |
| 2784 |
/// else |
| 2785 |
/// reduce_elem = remote_elem |
| 2786 |
/// |
| 2787 |
/// This algorithm assumes that the data to be reduced are located in a |
| 2788 |
/// contiguous subset of lanes starting from the first. When there is |
| 2789 |
/// an odd number of active lanes, the data in the last lane is not |
| 2790 |
/// aggregated with any other lane's dat but is instead copied over. |
| 2791 |
/// |
| 2792 |
/// Dispersed Partial Warp Reduction |
| 2793 |
/// |
| 2794 |
/// This algorithm is used within a warp when any discontiguous subset of |
| 2795 |
/// lanes are active. It is used to implement the reduction operation |
| 2796 |
/// across lanes in an OpenMP simd region or in a nested parallel region. |
| 2797 |
/// |
| 2798 |
/// void |
| 2799 |
/// dispersed_partial_reduce(void *reduce_data, |
| 2800 |
/// kmp_ShuffleReductFctPtr ShuffleReduceFn) { |
| 2801 |
/// int size, remote_id; |
| 2802 |
/// int logical_lane_id = number_of_active_lanes_before_me() * 2; |
| 2803 |
/// do { |
| 2804 |
/// remote_id = next_active_lane_id_right_after_me(); |
| 2805 |
/// # the above function returns 0 of no active lane |
| 2806 |
/// # is present right after the current lane. |
| 2807 |
/// size = number_of_active_lanes_in_this_warp(); |
| 2808 |
/// logical_lane_id /= 2; |
| 2809 |
/// ShuffleReduceFn(reduce_data, logical_lane_id, |
| 2810 |
/// remote_id-1-threadIdx.x, 2); |
| 2811 |
/// } while (logical_lane_id % 2 == 0 && size > 1); |
| 2812 |
/// } |
| 2813 |
/// |
| 2814 |
/// There is no assumption made about the initial state of the reduction. |
| 2815 |
/// Any number of lanes (>=1) could be active at any position. The reduction |
| 2816 |
/// result is returned in the first active lane. |
| 2817 |
/// |
| 2818 |
/// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
| 2819 |
/// |
| 2820 |
/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
| 2821 |
/// if (lane_id % 2 == 0 && offset > 0) |
| 2822 |
/// reduce_elem = reduce_elem REDUCE_OP remote_elem |
| 2823 |
/// else |
| 2824 |
/// reduce_elem = remote_elem |
| 2825 |
/// |
| 2826 |
/// |
| 2827 |
/// Intra-Team Reduction |
| 2828 |
/// |
| 2829 |
/// This function, as implemented in the runtime call |
| 2830 |
/// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP |
| 2831 |
/// threads in a team. It first reduces within a warp using the |
| 2832 |
/// aforementioned algorithms. We then proceed to gather all such |
| 2833 |
/// reduced values at the first warp. |
| 2834 |
/// |
| 2835 |
/// The runtime makes use of the function 'InterWarpCpyFn', which copies |
| 2836 |
/// data from each of the "warp master" (zeroth lane of each warp, where |
| 2837 |
/// warp-reduced data is held) to the zeroth warp. This step reduces (in |
| 2838 |
/// a mathematical sense) the problem of reduction across warp masters in |
| 2839 |
/// a block to the problem of warp reduction. |
| 2840 |
/// |
| 2841 |
/// |
| 2842 |
/// Inter-Team Reduction |
| 2843 |
/// |
| 2844 |
/// Once a team has reduced its data to a single value, it is stored in |
| 2845 |
/// a global scratchpad array. Since each team has a distinct slot, this |
| 2846 |
/// can be done without locking. |
| 2847 |
/// |
| 2848 |
/// The last team to write to the scratchpad array proceeds to reduce the |
| 2849 |
/// scratchpad array. One or more workers in the last team use the helper |
| 2850 |
/// 'loadAndReduceDataFn' to load and reduce values from the array, i.e., |
| 2851 |
/// the k'th worker reduces every k'th element. |
| 2852 |
/// |
| 2853 |
/// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to |
| 2854 |
/// reduce across workers and compute a globally reduced value. |
| 2855 |
/// |
| 2856 |
void CGOpenMPRuntimeGPU::emitReduction( |
| 2857 |
CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates, |
| 2858 |
ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, |
| 2859 |
ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) { |
| 2860 |
if (!CGF.HaveInsertPoint()) |
| 2861 |
return; |
| 2862 |
|
| 2863 |
bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind); |
| 2864 |
#ifndef NDEBUG |
| 2865 |
bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind); |
| 2866 |
#endif |
| 2867 |
|
| 2868 |
if (Options.SimpleReduction) { |
| 2869 |
assert(!TeamsReduction && !ParallelReduction && |
| 2870 |
"Invalid reduction selection in emitReduction."); |
| 2871 |
CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs, |
| 2872 |
ReductionOps, Options); |
| 2873 |
return; |
| 2874 |
} |
| 2875 |
|
| 2876 |
assert((TeamsReduction || ParallelReduction) && |
| 2877 |
"Invalid reduction selection in emitReduction."); |
| 2878 |
|
| 2879 |
// Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList), |
| 2880 |
// RedList, shuffle_reduce_func, interwarp_copy_func); |
| 2881 |
// or |
| 2882 |
// Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>); |
| 2883 |
llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); |
| 2884 |
llvm::Value *ThreadId = getThreadID(CGF, Loc); |
| 2885 |
|
| 2886 |
llvm::Value *Res; |
| 2887 |
ASTContext &C = CGM.getContext(); |
| 2888 |
// 1. Build a list of reduction variables. |
| 2889 |
// void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
| 2890 |
auto Size = RHSExprs.size(); |
| 2891 |
for (const Expr *E : Privates) { |
| 2892 |
if (E->getType()->isVariablyModifiedType()) |
| 2893 |
// Reserve place for array size. |
| 2894 |
++Size; |
| 2895 |
} |
| 2896 |
llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size); |
| 2897 |
QualType ReductionArrayTy = |
| 2898 |
C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal, |
| 2899 |
/*IndexTypeQuals=*/0); |
| 2900 |
Address ReductionList = |
| 2901 |
CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); |
| 2902 |
auto IPriv = Privates.begin(); |
| 2903 |
unsigned Idx = 0; |
| 2904 |
for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) { |
| 2905 |
Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 2906 |
CGF.Builder.CreateStore( |
| 2907 |
CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 2908 |
CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy), |
| 2909 |
Elem); |
| 2910 |
if ((*IPriv)->getType()->isVariablyModifiedType()) { |
| 2911 |
// Store array size. |
| 2912 |
++Idx; |
| 2913 |
Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 2914 |
llvm::Value *Size = CGF.Builder.CreateIntCast( |
| 2915 |
CGF.getVLASize( |
| 2916 |
CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) |
| 2917 |
.NumElts, |
| 2918 |
CGF.SizeTy, /*isSigned=*/false); |
| 2919 |
CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), |
| 2920 |
Elem); |
| 2921 |
} |
| 2922 |
} |
| 2923 |
|
| 2924 |
llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 2925 |
ReductionList.getPointer(), CGF.VoidPtrTy); |
| 2926 |
llvm::Function *ReductionFn = |
| 2927 |
emitReductionFunction(Loc, CGF.ConvertTypeForMem(ReductionArrayTy), |
| 2928 |
Privates, LHSExprs, RHSExprs, ReductionOps); |
| 2929 |
llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy); |
| 2930 |
llvm::Function *ShuffleAndReduceFn = emitShuffleAndReduceFunction( |
| 2931 |
CGM, Privates, ReductionArrayTy, ReductionFn, Loc); |
| 2932 |
llvm::Value *InterWarpCopyFn = |
| 2933 |
emitInterWarpCopyFunction(CGM, Privates, ReductionArrayTy, Loc); |
| 2934 |
|
| 2935 |
if (ParallelReduction) { |
| 2936 |
llvm::Value *Args[] = {RTLoc, |
| 2937 |
ThreadId, |
| 2938 |
CGF.Builder.getInt32(RHSExprs.size()), |
| 2939 |
ReductionArrayTySize, |
| 2940 |
RL, |
| 2941 |
ShuffleAndReduceFn, |
| 2942 |
InterWarpCopyFn}; |
| 2943 |
|
| 2944 |
Res = CGF.EmitRuntimeCall( |
| 2945 |
OMPBuilder.getOrCreateRuntimeFunction( |
| 2946 |
CGM.getModule(), OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2), |
| 2947 |
Args); |
| 2948 |
} else { |
| 2949 |
assert(TeamsReduction && "expected teams reduction."); |
| 2950 |
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap; |
| 2951 |
llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size()); |
| 2952 |
int Cnt = 0; |
| 2953 |
for (const Expr *DRE : Privates) { |
| 2954 |
PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl(); |
| 2955 |
++Cnt; |
| 2956 |
} |
| 2957 |
const RecordDecl *TeamReductionRec = ::buildRecordForGlobalizedVars( |
| 2958 |
CGM.getContext(), PrivatesReductions, std::nullopt, VarFieldMap, |
| 2959 |
C.getLangOpts().OpenMPCUDAReductionBufNum); |
| 2960 |
TeamsReductions.push_back(TeamReductionRec); |
| 2961 |
if (!KernelTeamsReductionPtr) { |
| 2962 |
KernelTeamsReductionPtr = new llvm::GlobalVariable( |
| 2963 |
CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/true, |
| 2964 |
llvm::GlobalValue::InternalLinkage, nullptr, |
| 2965 |
"_openmp_teams_reductions_buffer_$_$ptr"); |
| 2966 |
} |
| 2967 |
llvm::Value *GlobalBufferPtr = CGF.EmitLoadOfScalar( |
| 2968 |
Address(KernelTeamsReductionPtr, CGF.VoidPtrTy, CGM.getPointerAlign()), |
| 2969 |
/*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); |
| 2970 |
llvm::Value *GlobalToBufferCpyFn = ::emitListToGlobalCopyFunction( |
| 2971 |
CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap); |
| 2972 |
llvm::Value *GlobalToBufferRedFn = ::emitListToGlobalReduceFunction( |
| 2973 |
CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap, |
| 2974 |
ReductionFn); |
| 2975 |
llvm::Value *BufferToGlobalCpyFn = ::emitGlobalToListCopyFunction( |
| 2976 |
CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap); |
| 2977 |
llvm::Value *BufferToGlobalRedFn = ::emitGlobalToListReduceFunction( |
| 2978 |
CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap, |
| 2979 |
ReductionFn); |
| 2980 |
|
| 2981 |
llvm::Value *Args[] = { |
| 2982 |
RTLoc, |
| 2983 |
ThreadId, |
| 2984 |
GlobalBufferPtr, |
| 2985 |
CGF.Builder.getInt32(C.getLangOpts().OpenMPCUDAReductionBufNum), |
| 2986 |
RL, |
| 2987 |
ShuffleAndReduceFn, |
| 2988 |
InterWarpCopyFn, |
| 2989 |
GlobalToBufferCpyFn, |
| 2990 |
GlobalToBufferRedFn, |
| 2991 |
BufferToGlobalCpyFn, |
| 2992 |
BufferToGlobalRedFn}; |
| 2993 |
|
| 2994 |
Res = CGF.EmitRuntimeCall( |
| 2995 |
OMPBuilder.getOrCreateRuntimeFunction( |
| 2996 |
CGM.getModule(), OMPRTL___kmpc_nvptx_teams_reduce_nowait_v2), |
| 2997 |
Args); |
| 2998 |
} |
| 2999 |
|
| 3000 |
// 5. Build if (res == 1) |
| 3001 |
llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.reduction.done"); |
| 3002 |
llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.then"); |
| 3003 |
llvm::Value *Cond = CGF.Builder.CreateICmpEQ( |
| 3004 |
Res, llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1)); |
| 3005 |
CGF.Builder.CreateCondBr(Cond, ThenBB, ExitBB); |
| 3006 |
|
| 3007 |
// 6. Build then branch: where we have reduced values in the master |
| 3008 |
// thread in each team. |
| 3009 |
// __kmpc_end_reduce{_nowait}(<gtid>); |
| 3010 |
// break; |
| 3011 |
CGF.EmitBlock(ThenBB); |
| 3012 |
|
| 3013 |
// Add emission of __kmpc_end_reduce{_nowait}(<gtid>); |
| 3014 |
auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps, |
| 3015 |
this](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3016 |
auto IPriv = Privates.begin(); |
| 3017 |
auto ILHS = LHSExprs.begin(); |
| 3018 |
auto IRHS = RHSExprs.begin(); |
| 3019 |
for (const Expr *E : ReductionOps) { |
| 3020 |
emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), |
| 3021 |
cast<DeclRefExpr>(*IRHS)); |
| 3022 |
++IPriv; |
| 3023 |
++ILHS; |
| 3024 |
++IRHS; |
| 3025 |
} |
| 3026 |
}; |
| 3027 |
llvm::Value *EndArgs[] = {ThreadId}; |
| 3028 |
RegionCodeGenTy RCG(CodeGen); |
| 3029 |
NVPTXActionTy Action( |
| 3030 |
nullptr, std::nullopt, |
| 3031 |
OMPBuilder.getOrCreateRuntimeFunction( |
| 3032 |
CGM.getModule(), OMPRTL___kmpc_nvptx_end_reduce_nowait), |
| 3033 |
EndArgs); |
| 3034 |
RCG.setAction(Action); |
| 3035 |
RCG(CGF); |
| 3036 |
// There is no need to emit line number for unconditional branch. |
| 3037 |
(void)ApplyDebugLocation::CreateEmpty(CGF); |
| 3038 |
CGF.EmitBlock(ExitBB, /*IsFinished=*/true); |
| 3039 |
} |
| 3040 |
|
| 3041 |
const VarDecl * |
| 3042 |
CGOpenMPRuntimeGPU::translateParameter(const FieldDecl *FD, |
| 3043 |
const VarDecl *NativeParam) const { |
| 3044 |
if (!NativeParam->getType()->isReferenceType()) |
| 3045 |
return NativeParam; |
| 3046 |
QualType ArgType = NativeParam->getType(); |
| 3047 |
QualifierCollector QC; |
| 3048 |
const Type *NonQualTy = QC.strip(ArgType); |
| 3049 |
QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); |
| 3050 |
if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) { |
| 3051 |
if (Attr->getCaptureKind() == OMPC_map) { |
| 3052 |
PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy, |
| 3053 |
LangAS::opencl_global); |
| 3054 |
} |
| 3055 |
} |
| 3056 |
ArgType = CGM.getContext().getPointerType(PointeeTy); |
| 3057 |
QC.addRestrict(); |
| 3058 |
enum { NVPTX_local_addr = 5 }; |
| 3059 |
QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr)); |
| 3060 |
ArgType = QC.apply(CGM.getContext(), ArgType); |
| 3061 |
if (isa<ImplicitParamDecl>(NativeParam)) |
| 3062 |
return ImplicitParamDecl::Create( |
| 3063 |
CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(), |
| 3064 |
NativeParam->getIdentifier(), ArgType, ImplicitParamDecl::Other); |
| 3065 |
return ParmVarDecl::Create( |
| 3066 |
CGM.getContext(), |
| 3067 |
const_cast<DeclContext *>(NativeParam->getDeclContext()), |
| 3068 |
NativeParam->getBeginLoc(), NativeParam->getLocation(), |
| 3069 |
NativeParam->getIdentifier(), ArgType, |
| 3070 |
/*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr); |
| 3071 |
} |
| 3072 |
|
| 3073 |
Address |
| 3074 |
CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF, |
| 3075 |
const VarDecl *NativeParam, |
| 3076 |
const VarDecl *TargetParam) const { |
| 3077 |
assert(NativeParam != TargetParam && |
| 3078 |
NativeParam->getType()->isReferenceType() && |
| 3079 |
"Native arg must not be the same as target arg."); |
| 3080 |
Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam); |
| 3081 |
QualType NativeParamType = NativeParam->getType(); |
| 3082 |
QualifierCollector QC; |
| 3083 |
const Type *NonQualTy = QC.strip(NativeParamType); |
| 3084 |
QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); |
| 3085 |
unsigned NativePointeeAddrSpace = |
| 3086 |
CGF.getTypes().getTargetAddressSpace(NativePointeeTy); |
| 3087 |
QualType TargetTy = TargetParam->getType(); |
| 3088 |
llvm::Value *TargetAddr = CGF.EmitLoadOfScalar( |
| 3089 |
LocalAddr, /*Volatile=*/false, TargetTy, SourceLocation()); |
| 3090 |
// First cast to generic. |
| 3091 |
TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 3092 |
TargetAddr, llvm::PointerType::getWithSamePointeeType( |
| 3093 |
cast<llvm::PointerType>(TargetAddr->getType()), /*AddrSpace=*/0)); |
| 3094 |
// Cast from generic to native address space. |
| 3095 |
TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 3096 |
TargetAddr, llvm::PointerType::getWithSamePointeeType( |
| 3097 |
cast<llvm::PointerType>(TargetAddr->getType()), |
| 3098 |
NativePointeeAddrSpace)); |
| 3099 |
Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType); |
| 3100 |
CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false, |
| 3101 |
NativeParamType); |
| 3102 |
return NativeParamAddr; |
| 3103 |
} |
| 3104 |
|
| 3105 |
void CGOpenMPRuntimeGPU::emitOutlinedFunctionCall( |
| 3106 |
CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn, |
| 3107 |
ArrayRef<llvm::Value *> Args) const { |
| 3108 |
SmallVector<llvm::Value *, 4> TargetArgs; |
| 3109 |
TargetArgs.reserve(Args.size()); |
| 3110 |
auto *FnType = OutlinedFn.getFunctionType(); |
| 3111 |
for (unsigned I = 0, E = Args.size(); I < E; ++I) { |
| 3112 |
if (FnType->isVarArg() && FnType->getNumParams() <= I) { |
| 3113 |
TargetArgs.append(std::next(Args.begin(), I), Args.end()); |
| 3114 |
break; |
| 3115 |
} |
| 3116 |
llvm::Type *TargetType = FnType->getParamType(I); |
| 3117 |
llvm::Value *NativeArg = Args[I]; |
| 3118 |
if (!TargetType->isPointerTy()) { |
| 3119 |
TargetArgs.emplace_back(NativeArg); |
| 3120 |
continue; |
| 3121 |
} |
| 3122 |
llvm::Value *TargetArg = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 3123 |
NativeArg, llvm::PointerType::getWithSamePointeeType( |
| 3124 |
cast<llvm::PointerType>(NativeArg->getType()), /*AddrSpace*/ 0)); |
| 3125 |
TargetArgs.emplace_back( |
| 3126 |
CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TargetArg, TargetType)); |
| 3127 |
} |
| 3128 |
CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs); |
| 3129 |
} |
| 3130 |
|
| 3131 |
/// Emit function which wraps the outline parallel region |
| 3132 |
/// and controls the arguments which are passed to this function. |
| 3133 |
/// The wrapper ensures that the outlined function is called |
| 3134 |
/// with the correct arguments when data is shared. |
| 3135 |
llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper( |
| 3136 |
llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) { |
| 3137 |
ASTContext &Ctx = CGM.getContext(); |
| 3138 |
const auto &CS = *D.getCapturedStmt(OMPD_parallel); |
| 3139 |
|
| 3140 |
// Create a function that takes as argument the source thread. |
| 3141 |
FunctionArgList WrapperArgs; |
| 3142 |
QualType Int16QTy = |
| 3143 |
Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false); |
| 3144 |
QualType Int32QTy = |
| 3145 |
Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false); |
| 3146 |
ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), |
| 3147 |
/*Id=*/nullptr, Int16QTy, |
| 3148 |
ImplicitParamDecl::Other); |
| 3149 |
ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), |
| 3150 |
/*Id=*/nullptr, Int32QTy, |
| 3151 |
ImplicitParamDecl::Other); |
| 3152 |
WrapperArgs.emplace_back(&ParallelLevelArg); |
| 3153 |
WrapperArgs.emplace_back(&WrapperArg); |
| 3154 |
|
| 3155 |
const CGFunctionInfo &CGFI = |
| 3156 |
CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs); |
| 3157 |
|
| 3158 |
auto *Fn = llvm::Function::Create( |
| 3159 |
CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 3160 |
Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule()); |
| 3161 |
|
| 3162 |
// Ensure we do not inline the function. This is trivially true for the ones |
| 3163 |
// passed to __kmpc_fork_call but the ones calles in serialized regions |
| 3164 |
// could be inlined. This is not a perfect but it is closer to the invariant |
| 3165 |
// we want, namely, every data environment starts with a new function. |
| 3166 |
// TODO: We should pass the if condition to the runtime function and do the |
| 3167 |
// handling there. Much cleaner code. |
| 3168 |
Fn->addFnAttr(llvm::Attribute::NoInline); |
| 3169 |
|
| 3170 |
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 3171 |
Fn->setLinkage(llvm::GlobalValue::InternalLinkage); |
| 3172 |
Fn->setDoesNotRecurse(); |
| 3173 |
|
| 3174 |
CodeGenFunction CGF(CGM, /*suppressNewContext=*/true); |
| 3175 |
CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs, |
| 3176 |
D.getBeginLoc(), D.getBeginLoc()); |
| 3177 |
|
| 3178 |
const auto *RD = CS.getCapturedRecordDecl(); |
| 3179 |
auto CurField = RD->field_begin(); |
| 3180 |
|
| 3181 |
Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, |
| 3182 |
/*Name=*/".zero.addr"); |
| 3183 |
CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr); |
| 3184 |
// Get the array of arguments. |
| 3185 |
SmallVector<llvm::Value *, 8> Args; |
| 3186 |
|
| 3187 |
Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).getPointer()); |
| 3188 |
Args.emplace_back(ZeroAddr.getPointer()); |
| 3189 |
|
| 3190 |
CGBuilderTy &Bld = CGF.Builder; |
| 3191 |
auto CI = CS.capture_begin(); |
| 3192 |
|
| 3193 |
// Use global memory for data sharing. |
| 3194 |
// Handle passing of global args to workers. |
| 3195 |
Address GlobalArgs = |
| 3196 |
CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args"); |
| 3197 |
llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer(); |
| 3198 |
llvm::Value *DataSharingArgs[] = {GlobalArgsPtr}; |
| 3199 |
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 3200 |
CGM.getModule(), OMPRTL___kmpc_get_shared_variables), |
| 3201 |
DataSharingArgs); |
| 3202 |
|
| 3203 |
// Retrieve the shared variables from the list of references returned |
| 3204 |
// by the runtime. Pass the variables to the outlined function. |
| 3205 |
Address SharedArgListAddress = Address::invalid(); |
| 3206 |
if (CS.capture_size() > 0 || |
| 3207 |
isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) { |
| 3208 |
SharedArgListAddress = CGF.EmitLoadOfPointer( |
| 3209 |
GlobalArgs, CGF.getContext() |
| 3210 |
.getPointerType(CGF.getContext().VoidPtrTy) |
| 3211 |
.castAs<PointerType>()); |
| 3212 |
} |
| 3213 |
unsigned Idx = 0; |
| 3214 |
if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) { |
| 3215 |
Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); |
| 3216 |
Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 3217 |
Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy); |
| 3218 |
llvm::Value *LB = CGF.EmitLoadOfScalar( |
| 3219 |
TypedAddress, |
| 3220 |
/*Volatile=*/false, |
| 3221 |
CGF.getContext().getPointerType(CGF.getContext().getSizeType()), |
| 3222 |
cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc()); |
| 3223 |
Args.emplace_back(LB); |
| 3224 |
++Idx; |
| 3225 |
Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); |
| 3226 |
TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 3227 |
Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy); |
| 3228 |
llvm::Value *UB = CGF.EmitLoadOfScalar( |
| 3229 |
TypedAddress, |
| 3230 |
/*Volatile=*/false, |
| 3231 |
CGF.getContext().getPointerType(CGF.getContext().getSizeType()), |
| 3232 |
cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc()); |
| 3233 |
Args.emplace_back(UB); |
| 3234 |
++Idx; |
| 3235 |
} |
| 3236 |
if (CS.capture_size() > 0) { |
| 3237 |
ASTContext &CGFContext = CGF.getContext(); |
| 3238 |
for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) { |
| 3239 |
QualType ElemTy = CurField->getType(); |
| 3240 |
Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx); |
| 3241 |
Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 3242 |
Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy)), |
| 3243 |
CGF.ConvertTypeForMem(ElemTy)); |
| 3244 |
llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress, |
| 3245 |
/*Volatile=*/false, |
| 3246 |
CGFContext.getPointerType(ElemTy), |
| 3247 |
CI->getLocation()); |
| 3248 |
if (CI->capturesVariableByCopy() && |
| 3249 |
!CI->getCapturedVar()->getType()->isAnyPointerType()) { |
| 3250 |
Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(), |
| 3251 |
CI->getLocation()); |
| 3252 |
} |
| 3253 |
Args.emplace_back(Arg); |
| 3254 |
} |
| 3255 |
} |
| 3256 |
|
| 3257 |
emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args); |
| 3258 |
CGF.FinishFunction(); |
| 3259 |
return Fn; |
| 3260 |
} |
| 3261 |
|
| 3262 |
void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF, |
| 3263 |
const Decl *D) { |
| 3264 |
if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic) |
| 3265 |
return; |
| 3266 |
|
| 3267 |
assert(D && "Expected function or captured|block decl."); |
| 3268 |
assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && |
| 3269 |
"Function is registered already."); |
| 3270 |
assert((!TeamAndReductions.first || TeamAndReductions.first == D) && |
| 3271 |
"Team is set but not processed."); |
| 3272 |
const Stmt *Body = nullptr; |
| 3273 |
bool NeedToDelayGlobalization = false; |
| 3274 |
if (const auto *FD = dyn_cast<FunctionDecl>(D)) { |
| 3275 |
Body = FD->getBody(); |
| 3276 |
} else if (const auto *BD = dyn_cast<BlockDecl>(D)) { |
| 3277 |
Body = BD->getBody(); |
| 3278 |
} else if (const auto *CD = dyn_cast<CapturedDecl>(D)) { |
| 3279 |
Body = CD->getBody(); |
| 3280 |
NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP; |
| 3281 |
if (NeedToDelayGlobalization && |
| 3282 |
getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) |
| 3283 |
return; |
| 3284 |
} |
| 3285 |
if (!Body) |
| 3286 |
return; |
| 3287 |
CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second); |
| 3288 |
VarChecker.Visit(Body); |
| 3289 |
const RecordDecl *GlobalizedVarsRecord = |
| 3290 |
VarChecker.getGlobalizedRecord(IsInTTDRegion); |
| 3291 |
TeamAndReductions.first = nullptr; |
| 3292 |
TeamAndReductions.second.clear(); |
| 3293 |
ArrayRef<const ValueDecl *> EscapedVariableLengthDecls = |
| 3294 |
VarChecker.getEscapedVariableLengthDecls(); |
| 3295 |
if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty()) |
| 3296 |
return; |
| 3297 |
auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; |
| 3298 |
I->getSecond().MappedParams = |
| 3299 |
std::make_unique<CodeGenFunction::OMPMapVars>(); |
| 3300 |
I->getSecond().EscapedParameters.insert( |
| 3301 |
VarChecker.getEscapedParameters().begin(), |
| 3302 |
VarChecker.getEscapedParameters().end()); |
| 3303 |
I->getSecond().EscapedVariableLengthDecls.append( |
| 3304 |
EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end()); |
| 3305 |
DeclToAddrMapTy &Data = I->getSecond().LocalVarData; |
| 3306 |
for (const ValueDecl *VD : VarChecker.getEscapedDecls()) { |
| 3307 |
assert(VD->isCanonicalDecl() && "Expected canonical declaration"); |
| 3308 |
Data.insert(std::make_pair(VD, MappedVarData())); |
| 3309 |
} |
| 3310 |
if (!NeedToDelayGlobalization) { |
| 3311 |
emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true); |
| 3312 |
struct GlobalizationScope final : EHScopeStack::Cleanup { |
| 3313 |
GlobalizationScope() = default; |
| 3314 |
|
| 3315 |
void Emit(CodeGenFunction &CGF, Flags flags) override { |
| 3316 |
static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) |
| 3317 |
.emitGenericVarsEpilog(CGF, /*WithSPMDCheck=*/true); |
| 3318 |
} |
| 3319 |
}; |
| 3320 |
CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup); |
| 3321 |
} |
| 3322 |
} |
| 3323 |
|
| 3324 |
Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF, |
| 3325 |
const VarDecl *VD) { |
| 3326 |
if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) { |
| 3327 |
const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); |
| 3328 |
auto AS = LangAS::Default; |
| 3329 |
switch (A->getAllocatorType()) { |
| 3330 |
// Use the default allocator here as by default local vars are |
| 3331 |
// threadlocal. |
| 3332 |
case OMPAllocateDeclAttr::OMPNullMemAlloc: |
| 3333 |
case OMPAllocateDeclAttr::OMPDefaultMemAlloc: |
| 3334 |
case OMPAllocateDeclAttr::OMPThreadMemAlloc: |
| 3335 |
case OMPAllocateDeclAttr::OMPHighBWMemAlloc: |
| 3336 |
case OMPAllocateDeclAttr::OMPLowLatMemAlloc: |
| 3337 |
// Follow the user decision - use default allocation. |
| 3338 |
return Address::invalid(); |
| 3339 |
case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: |
| 3340 |
// TODO: implement aupport for user-defined allocators. |
| 3341 |
return Address::invalid(); |
| 3342 |
case OMPAllocateDeclAttr::OMPConstMemAlloc: |
| 3343 |
AS = LangAS::cuda_constant; |
| 3344 |
break; |
| 3345 |
case OMPAllocateDeclAttr::OMPPTeamMemAlloc: |
| 3346 |
AS = LangAS::cuda_shared; |
| 3347 |
break; |
| 3348 |
case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: |
| 3349 |
case OMPAllocateDeclAttr::OMPCGroupMemAlloc: |
| 3350 |
break; |
| 3351 |
} |
| 3352 |
llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType()); |
| 3353 |
auto *GV = new llvm::GlobalVariable( |
| 3354 |
CGM.getModule(), VarTy, /*isConstant=*/false, |
| 3355 |
llvm::GlobalValue::InternalLinkage, llvm::Constant::getNullValue(VarTy), |
| 3356 |
VD->getName(), |
| 3357 |
/*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, |
| 3358 |
CGM.getContext().getTargetAddressSpace(AS)); |
| 3359 |
CharUnits Align = CGM.getContext().getDeclAlign(VD); |
| 3360 |
GV->setAlignment(Align.getAsAlign()); |
| 3361 |
return Address( |
| 3362 |
CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 3363 |
GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace( |
| 3364 |
VD->getType().getAddressSpace()))), |
| 3365 |
VarTy, Align); |
| 3366 |
} |
| 3367 |
|
| 3368 |
if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic) |
| 3369 |
return Address::invalid(); |
| 3370 |
|
| 3371 |
VD = VD->getCanonicalDecl(); |
| 3372 |
auto I = FunctionGlobalizedDecls.find(CGF.CurFn); |
| 3373 |
if (I == FunctionGlobalizedDecls.end()) |
| 3374 |
return Address::invalid(); |
| 3375 |
auto VDI = I->getSecond().LocalVarData.find(VD); |
| 3376 |
if (VDI != I->getSecond().LocalVarData.end()) |
| 3377 |
return VDI->second.PrivateAddr; |
| 3378 |
if (VD->hasAttrs()) { |
| 3379 |
for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()), |
| 3380 |
E(VD->attr_end()); |
| 3381 |
IT != E; ++IT) { |
| 3382 |
auto VDI = I->getSecond().LocalVarData.find( |
| 3383 |
cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl()) |
| 3384 |
->getCanonicalDecl()); |
| 3385 |
if (VDI != I->getSecond().LocalVarData.end()) |
| 3386 |
return VDI->second.PrivateAddr; |
| 3387 |
} |
| 3388 |
} |
| 3389 |
|
| 3390 |
return Address::invalid(); |
| 3391 |
} |
| 3392 |
|
| 3393 |
void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) { |
| 3394 |
FunctionGlobalizedDecls.erase(CGF.CurFn); |
| 3395 |
CGOpenMPRuntime::functionFinished(CGF); |
| 3396 |
} |
| 3397 |
|
| 3398 |
void CGOpenMPRuntimeGPU::getDefaultDistScheduleAndChunk( |
| 3399 |
CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 3400 |
OpenMPDistScheduleClauseKind &ScheduleKind, |
| 3401 |
llvm::Value *&Chunk) const { |
| 3402 |
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 3403 |
if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { |
| 3404 |
ScheduleKind = OMPC_DIST_SCHEDULE_static; |
| 3405 |
Chunk = CGF.EmitScalarConversion( |
| 3406 |
RT.getGPUNumThreads(CGF), |
| 3407 |
CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), |
| 3408 |
S.getIterationVariable()->getType(), S.getBeginLoc()); |
| 3409 |
return; |
| 3410 |
} |
| 3411 |
CGOpenMPRuntime::getDefaultDistScheduleAndChunk( |
| 3412 |
CGF, S, ScheduleKind, Chunk); |
| 3413 |
} |
| 3414 |
|
| 3415 |
void CGOpenMPRuntimeGPU::getDefaultScheduleAndChunk( |
| 3416 |
CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 3417 |
OpenMPScheduleClauseKind &ScheduleKind, |
| 3418 |
const Expr *&ChunkExpr) const { |
| 3419 |
ScheduleKind = OMPC_SCHEDULE_static; |
| 3420 |
// Chunk size is 1 in this case. |
| 3421 |
llvm::APInt ChunkSize(32, 1); |
| 3422 |
ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize, |
| 3423 |
CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), |
| 3424 |
SourceLocation()); |
| 3425 |
} |
| 3426 |
|
| 3427 |
void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas( |
| 3428 |
CodeGenFunction &CGF, const OMPExecutableDirective &D) const { |
| 3429 |
assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && |
| 3430 |
" Expected target-based directive."); |
| 3431 |
const CapturedStmt *CS = D.getCapturedStmt(OMPD_target); |
| 3432 |
for (const CapturedStmt::Capture &C : CS->captures()) { |
| 3433 |
// Capture variables captured by reference in lambdas for target-based |
| 3434 |
// directives. |
| 3435 |
if (!C.capturesVariable()) |
| 3436 |
continue; |
| 3437 |
const VarDecl *VD = C.getCapturedVar(); |
| 3438 |
const auto *RD = VD->getType() |
| 3439 |
.getCanonicalType() |
| 3440 |
.getNonReferenceType() |
| 3441 |
->getAsCXXRecordDecl(); |
| 3442 |
if (!RD || !RD->isLambda()) |
| 3443 |
continue; |
| 3444 |
Address VDAddr = CGF.GetAddrOfLocalVar(VD); |
| 3445 |
LValue VDLVal; |
| 3446 |
if (VD->getType().getCanonicalType()->isReferenceType()) |
| 3447 |
VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType()); |
| 3448 |
else |
| 3449 |
VDLVal = CGF.MakeAddrLValue( |
| 3450 |
VDAddr, VD->getType().getCanonicalType().getNonReferenceType()); |
| 3451 |
llvm::DenseMap<const ValueDecl *, FieldDecl *> Captures; |
| 3452 |
FieldDecl *ThisCapture = nullptr; |
| 3453 |
RD->getCaptureFields(Captures, ThisCapture); |
| 3454 |
if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) { |
| 3455 |
LValue ThisLVal = |
| 3456 |
CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture); |
| 3457 |
llvm::Value *CXXThis = CGF.LoadCXXThis(); |
| 3458 |
CGF.EmitStoreOfScalar(CXXThis, ThisLVal); |
| 3459 |
} |
| 3460 |
for (const LambdaCapture &LC : RD->captures()) { |
| 3461 |
if (LC.getCaptureKind() != LCK_ByRef) |
| 3462 |
continue; |
| 3463 |
const ValueDecl *VD = LC.getCapturedVar(); |
| 3464 |
// FIXME: For now VD is always a VarDecl because OpenMP does not support |
| 3465 |
// capturing structured bindings in lambdas yet. |
| 3466 |
if (!CS->capturesVariable(cast<VarDecl>(VD))) |
| 3467 |
continue; |
| 3468 |
auto It = Captures.find(VD); |
| 3469 |
assert(It != Captures.end() && "Found lambda capture without field."); |
| 3470 |
LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second); |
| 3471 |
Address VDAddr = CGF.GetAddrOfLocalVar(cast<VarDecl>(VD)); |
| 3472 |
if (VD->getType().getCanonicalType()->isReferenceType()) |
| 3473 |
VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr, |
| 3474 |
VD->getType().getCanonicalType()) |
| 3475 |
.getAddress(CGF); |
| 3476 |
CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal); |
| 3477 |
} |
| 3478 |
} |
| 3479 |
} |
| 3480 |
|
| 3481 |
bool CGOpenMPRuntimeGPU::hasAllocateAttributeForGlobalVar(const VarDecl *VD, |
| 3482 |
LangAS &AS) { |
| 3483 |
if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>()) |
| 3484 |
return false; |
| 3485 |
const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); |
| 3486 |
switch(A->getAllocatorType()) { |
| 3487 |
case OMPAllocateDeclAttr::OMPNullMemAlloc: |
| 3488 |
case OMPAllocateDeclAttr::OMPDefaultMemAlloc: |
| 3489 |
// Not supported, fallback to the default mem space. |
| 3490 |
case OMPAllocateDeclAttr::OMPThreadMemAlloc: |
| 3491 |
case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: |
| 3492 |
case OMPAllocateDeclAttr::OMPCGroupMemAlloc: |
| 3493 |
case OMPAllocateDeclAttr::OMPHighBWMemAlloc: |
| 3494 |
case OMPAllocateDeclAttr::OMPLowLatMemAlloc: |
| 3495 |
AS = LangAS::Default; |
| 3496 |
return true; |
| 3497 |
case OMPAllocateDeclAttr::OMPConstMemAlloc: |
| 3498 |
AS = LangAS::cuda_constant; |
| 3499 |
return true; |
| 3500 |
case OMPAllocateDeclAttr::OMPPTeamMemAlloc: |
| 3501 |
AS = LangAS::cuda_shared; |
| 3502 |
return true; |
| 3503 |
case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: |
| 3504 |
llvm_unreachable("Expected predefined allocator for the variables with the " |
| 3505 |
"static storage."); |
| 3506 |
} |
| 3507 |
return false; |
| 3508 |
} |
| 3509 |
|
| 3510 |
// Get current CudaArch and ignore any unknown values |
| 3511 |
static CudaArch getCudaArch(CodeGenModule &CGM) { |
| 3512 |
if (!CGM.getTarget().hasFeature("ptx")) |
| 3513 |
return CudaArch::UNKNOWN; |
| 3514 |
for (const auto &Feature : CGM.getTarget().getTargetOpts().FeatureMap) { |
| 3515 |
if (Feature.getValue()) { |
| 3516 |
CudaArch Arch = StringToCudaArch(Feature.getKey()); |
| 3517 |
if (Arch != CudaArch::UNKNOWN) |
| 3518 |
return Arch; |
| 3519 |
} |
| 3520 |
} |
| 3521 |
return CudaArch::UNKNOWN; |
| 3522 |
} |
| 3523 |
|
| 3524 |
/// Check to see if target architecture supports unified addressing which is |
| 3525 |
/// a restriction for OpenMP requires clause "unified_shared_memory". |
| 3526 |
void CGOpenMPRuntimeGPU::processRequiresDirective( |
| 3527 |
const OMPRequiresDecl *D) { |
| 3528 |
for (const OMPClause *Clause : D->clauselists()) { |
| 3529 |
if (Clause->getClauseKind() == OMPC_unified_shared_memory) { |
| 3530 |
CudaArch Arch = getCudaArch(CGM); |
| 3531 |
switch (Arch) { |
| 3532 |
case CudaArch::SM_20: |
| 3533 |
case CudaArch::SM_21: |
| 3534 |
case CudaArch::SM_30: |
| 3535 |
case CudaArch::SM_32: |
| 3536 |
case CudaArch::SM_35: |
| 3537 |
case CudaArch::SM_37: |
| 3538 |
case CudaArch::SM_50: |
| 3539 |
case CudaArch::SM_52: |
| 3540 |
case CudaArch::SM_53: { |
| 3541 |
SmallString<256> Buffer; |
| 3542 |
llvm::raw_svector_ostream Out(Buffer); |
| 3543 |
Out << "Target architecture " << CudaArchToString(Arch) |
| 3544 |
<< " does not support unified addressing"; |
| 3545 |
CGM.Error(Clause->getBeginLoc(), Out.str()); |
| 3546 |
return; |
| 3547 |
} |
| 3548 |
case CudaArch::SM_60: |
| 3549 |
case CudaArch::SM_61: |
| 3550 |
case CudaArch::SM_62: |
| 3551 |
case CudaArch::SM_70: |
| 3552 |
case CudaArch::SM_72: |
| 3553 |
case CudaArch::SM_75: |
| 3554 |
case CudaArch::SM_80: |
| 3555 |
case CudaArch::SM_86: |
| 3556 |
case CudaArch::SM_87: |
| 3557 |
case CudaArch::SM_89: |
| 3558 |
case CudaArch::SM_90: |
| 3559 |
case CudaArch::GFX600: |
| 3560 |
case CudaArch::GFX601: |
| 3561 |
case CudaArch::GFX602: |
| 3562 |
case CudaArch::GFX700: |
| 3563 |
case CudaArch::GFX701: |
| 3564 |
case CudaArch::GFX702: |
| 3565 |
case CudaArch::GFX703: |
| 3566 |
case CudaArch::GFX704: |
| 3567 |
case CudaArch::GFX705: |
| 3568 |
case CudaArch::GFX801: |
| 3569 |
case CudaArch::GFX802: |
| 3570 |
case CudaArch::GFX803: |
| 3571 |
case CudaArch::GFX805: |
| 3572 |
case CudaArch::GFX810: |
| 3573 |
case CudaArch::GFX900: |
| 3574 |
case CudaArch::GFX902: |
| 3575 |
case CudaArch::GFX904: |
| 3576 |
case CudaArch::GFX906: |
| 3577 |
case CudaArch::GFX908: |
| 3578 |
case CudaArch::GFX909: |
| 3579 |
case CudaArch::GFX90a: |
| 3580 |
case CudaArch::GFX90c: |
| 3581 |
case CudaArch::GFX940: |
| 3582 |
case CudaArch::GFX1010: |
| 3583 |
case CudaArch::GFX1011: |
| 3584 |
case CudaArch::GFX1012: |
| 3585 |
case CudaArch::GFX1013: |
| 3586 |
case CudaArch::GFX1030: |
| 3587 |
case CudaArch::GFX1031: |
| 3588 |
case CudaArch::GFX1032: |
| 3589 |
case CudaArch::GFX1033: |
| 3590 |
case CudaArch::GFX1034: |
| 3591 |
case CudaArch::GFX1035: |
| 3592 |
case CudaArch::GFX1036: |
| 3593 |
case CudaArch::GFX1100: |
| 3594 |
case CudaArch::GFX1101: |
| 3595 |
case CudaArch::GFX1102: |
| 3596 |
case CudaArch::GFX1103: |
| 3597 |
case CudaArch::Generic: |
| 3598 |
case CudaArch::UNUSED: |
| 3599 |
case CudaArch::UNKNOWN: |
| 3600 |
break; |
| 3601 |
case CudaArch::LAST: |
| 3602 |
llvm_unreachable("Unexpected Cuda arch."); |
| 3603 |
} |
| 3604 |
} |
| 3605 |
} |
| 3606 |
CGOpenMPRuntime::processRequiresDirective(D); |
| 3607 |
} |
| 3608 |
|
| 3609 |
void CGOpenMPRuntimeGPU::clear() { |
| 3610 |
|
| 3611 |
if (!TeamsReductions.empty()) { |
| 3612 |
ASTContext &C = CGM.getContext(); |
| 3613 |
RecordDecl *StaticRD = C.buildImplicitRecord( |
| 3614 |
"_openmp_teams_reduction_type_$_", RecordDecl::TagKind::TTK_Union); |
| 3615 |
StaticRD->startDefinition(); |
| 3616 |
for (const RecordDecl *TeamReductionRec : TeamsReductions) { |
| 3617 |
QualType RecTy = C.getRecordType(TeamReductionRec); |
| 3618 |
auto *Field = FieldDecl::Create( |
| 3619 |
C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy, |
| 3620 |
C.getTrivialTypeSourceInfo(RecTy, SourceLocation()), |
| 3621 |
/*BW=*/nullptr, /*Mutable=*/false, |
| 3622 |
/*InitStyle=*/ICIS_NoInit); |
| 3623 |
Field->setAccess(AS_public); |
| 3624 |
StaticRD->addDecl(Field); |
| 3625 |
} |
| 3626 |
StaticRD->completeDefinition(); |
| 3627 |
QualType StaticTy = C.getRecordType(StaticRD); |
| 3628 |
llvm::Type *LLVMReductionsBufferTy = |
| 3629 |
CGM.getTypes().ConvertTypeForMem(StaticTy); |
| 3630 |
// FIXME: nvlink does not handle weak linkage correctly (object with the |
| 3631 |
// different size are reported as erroneous). |
| 3632 |
// Restore CommonLinkage as soon as nvlink is fixed. |
| 3633 |
auto *GV = new llvm::GlobalVariable( |
| 3634 |
CGM.getModule(), LLVMReductionsBufferTy, |
| 3635 |
/*isConstant=*/false, llvm::GlobalValue::InternalLinkage, |
| 3636 |
llvm::Constant::getNullValue(LLVMReductionsBufferTy), |
| 3637 |
"_openmp_teams_reductions_buffer_$_"); |
| 3638 |
KernelTeamsReductionPtr->setInitializer( |
| 3639 |
llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, |
| 3640 |
CGM.VoidPtrTy)); |
| 3641 |
} |
| 3642 |
CGOpenMPRuntime::clear(); |
| 3643 |
} |
| 3644 |
|
| 3645 |
llvm::Value *CGOpenMPRuntimeGPU::getGPUNumThreads(CodeGenFunction &CGF) { |
| 3646 |
CGBuilderTy &Bld = CGF.Builder; |
| 3647 |
llvm::Module *M = &CGF.CGM.getModule(); |
| 3648 |
const char *LocSize = "__kmpc_get_hardware_num_threads_in_block"; |
| 3649 |
llvm::Function *F = M->getFunction(LocSize); |
| 3650 |
if (!F) { |
| 3651 |
F = llvm::Function::Create( |
| 3652 |
llvm::FunctionType::get(CGF.Int32Ty, std::nullopt, false), |
| 3653 |
llvm::GlobalVariable::ExternalLinkage, LocSize, &CGF.CGM.getModule()); |
| 3654 |
} |
| 3655 |
return Bld.CreateCall(F, std::nullopt, "nvptx_num_threads"); |
| 3656 |
} |
| 3657 |
|
| 3658 |
llvm::Value *CGOpenMPRuntimeGPU::getGPUThreadID(CodeGenFunction &CGF) { |
| 3659 |
ArrayRef<llvm::Value *> Args{}; |
| 3660 |
return CGF.EmitRuntimeCall( |
| 3661 |
OMPBuilder.getOrCreateRuntimeFunction( |
| 3662 |
CGM.getModule(), OMPRTL___kmpc_get_hardware_thread_id_in_block), |
| 3663 |
Args); |
| 3664 |
} |
| 3665 |
|
| 3666 |
llvm::Value *CGOpenMPRuntimeGPU::getGPUWarpSize(CodeGenFunction &CGF) { |
| 3667 |
ArrayRef<llvm::Value *> Args{}; |
| 3668 |
return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 3669 |
CGM.getModule(), OMPRTL___kmpc_get_warp_size), |
| 3670 |
Args); |
| 3671 |
} |