forked from OSchip/llvm-project
[flang]Add support for do concurrent
[flang]Add support for do concurrent Upstreaming from fir-dev on https://github.com/flang-compiler/f18-llvm-project Support for concurrent execution in do-loops. A selection of tests are also added. Co-authored-by: V Donaldson <vdonaldson@nvidia.com> Reviewed By: kiranchandramohan Differential Revision: https://reviews.llvm.org/D127240
This commit is contained in:
Mats Petersson
parent
aebe24a856
commit
84b9ae6624
|
|
@ -87,7 +87,10 @@ struct IncrementLoopInfo {
|
|||
const Fortran::lower::SomeExpr *lowerExpr;
|
||||
const Fortran::lower::SomeExpr *upperExpr;
|
||||
const Fortran::lower::SomeExpr *stepExpr;
|
||||
const Fortran::lower::SomeExpr *maskExpr = nullptr;
|
||||
bool isUnordered; // do concurrent, forall
|
||||
llvm::SmallVector<const Fortran::semantics::Symbol *> localInitSymList;
|
||||
llvm::SmallVector<const Fortran::semantics::Symbol *> sharedSymList;
|
||||
mlir::Value loopVariable = nullptr;
|
||||
mlir::Value stepValue = nullptr; // possible uses in multiple blocks
|
||||
|
||||
|
|
@ -98,6 +101,7 @@ struct IncrementLoopInfo {
|
|||
bool hasRealControl = false;
|
||||
mlir::Value tripVariable = nullptr;
|
||||
mlir::Block *headerBlock = nullptr; // loop entry and test block
|
||||
mlir::Block *maskBlock = nullptr; // concurrent loop mask block
|
||||
mlir::Block *bodyBlock = nullptr; // first loop body block
|
||||
mlir::Block *exitBlock = nullptr; // loop exit target block
|
||||
};
|
||||
|
|
@ -636,9 +640,24 @@ private:
|
|||
}
|
||||
|
||||
/// Generate the address of loop variable \p sym.
|
||||
/// If \p sym is not mapped yet, allocate local storage for it.
|
||||
mlir::Value genLoopVariableAddress(mlir::Location loc,
|
||||
const Fortran::semantics::Symbol &sym) {
|
||||
assert(lookupSymbol(sym) && "loop control variable must already be in map");
|
||||
const Fortran::semantics::Symbol &sym,
|
||||
bool isUnordered) {
|
||||
if (isUnordered || sym.has<Fortran::semantics::HostAssocDetails>() ||
|
||||
sym.has<Fortran::semantics::UseDetails>()) {
|
||||
if (!shallowLookupSymbol(sym)) {
|
||||
// Do concurrent loop variables are not mapped yet since they are local
|
||||
// to the Do concurrent scope (same for OpenMP loops).
|
||||
auto newVal = builder->createTemporary(loc, genType(sym),
|
||||
toStringRef(sym.name()));
|
||||
bindIfNewSymbol(sym, newVal);
|
||||
return newVal;
|
||||
}
|
||||
}
|
||||
auto entry = lookupSymbol(sym);
|
||||
(void)entry;
|
||||
assert(entry && "loop control variable must already be in map");
|
||||
Fortran::lower::StatementContext stmtCtx;
|
||||
return fir::getBase(
|
||||
genExprAddr(Fortran::evaluate::AsGenericExpr(sym).value(), stmtCtx));
|
||||
|
|
@ -973,6 +992,34 @@ private:
|
|||
builder->create<fir::SelectOp>(loc, selectExpr, indexList, blockList);
|
||||
}
|
||||
|
||||
/// Collect DO CONCURRENT or FORALL loop control information.
|
||||
IncrementLoopNestInfo getConcurrentControl(
|
||||
const Fortran::parser::ConcurrentHeader &header,
|
||||
const std::list<Fortran::parser::LocalitySpec> &localityList = {}) {
|
||||
IncrementLoopNestInfo incrementLoopNestInfo;
|
||||
for (const Fortran::parser::ConcurrentControl &control :
|
||||
std::get<std::list<Fortran::parser::ConcurrentControl>>(header.t))
|
||||
incrementLoopNestInfo.emplace_back(
|
||||
*std::get<0>(control.t).symbol, std::get<1>(control.t),
|
||||
std::get<2>(control.t), std::get<3>(control.t), /*isUnordered=*/true);
|
||||
IncrementLoopInfo &info = incrementLoopNestInfo.back();
|
||||
info.maskExpr = Fortran::semantics::GetExpr(
|
||||
std::get<std::optional<Fortran::parser::ScalarLogicalExpr>>(header.t));
|
||||
for (const Fortran::parser::LocalitySpec &x : localityList) {
|
||||
if (const auto *localInitList =
|
||||
std::get_if<Fortran::parser::LocalitySpec::LocalInit>(&x.u))
|
||||
for (const Fortran::parser::Name &x : localInitList->v)
|
||||
info.localInitSymList.push_back(x.symbol);
|
||||
if (const auto *sharedList =
|
||||
std::get_if<Fortran::parser::LocalitySpec::Shared>(&x.u))
|
||||
for (const Fortran::parser::Name &x : sharedList->v)
|
||||
info.sharedSymList.push_back(x.symbol);
|
||||
if (std::get_if<Fortran::parser::LocalitySpec::Local>(&x.u))
|
||||
TODO(toLocation(), "do concurrent locality specs not implemented");
|
||||
}
|
||||
return incrementLoopNestInfo;
|
||||
}
|
||||
|
||||
/// Generate FIR for a DO construct. There are six variants:
|
||||
/// - unstructured infinite and while loops
|
||||
/// - structured and unstructured increment loops
|
||||
|
|
@ -1029,7 +1076,34 @@ private:
|
|||
info.exitBlock = exitBlock;
|
||||
}
|
||||
} else {
|
||||
TODO(toLocation(), "infinite/unstructured loop/concurrent loop");
|
||||
const auto *concurrent =
|
||||
std::get_if<Fortran::parser::LoopControl::Concurrent>(
|
||||
&loopControl->u);
|
||||
assert(concurrent && "invalid DO loop variant");
|
||||
incrementLoopNestInfo = getConcurrentControl(
|
||||
std::get<Fortran::parser::ConcurrentHeader>(concurrent->t),
|
||||
std::get<std::list<Fortran::parser::LocalitySpec>>(concurrent->t));
|
||||
if (unstructuredContext) {
|
||||
maybeStartBlock(preheaderBlock);
|
||||
for (IncrementLoopInfo &info : incrementLoopNestInfo) {
|
||||
// The original loop body provides the body and latch blocks of the
|
||||
// innermost dimension. The (first) body block of a non-innermost
|
||||
// dimension is the preheader block of the immediately enclosed
|
||||
// dimension. The latch block of a non-innermost dimension is the
|
||||
// exit block of the immediately enclosed dimension.
|
||||
auto createNextExitBlock = [&]() {
|
||||
// Create unstructured loop exit blocks, outermost to innermost.
|
||||
return exitBlock = insertBlock(exitBlock);
|
||||
};
|
||||
bool isInnermost = &info == &incrementLoopNestInfo.back();
|
||||
bool isOutermost = &info == &incrementLoopNestInfo.front();
|
||||
info.headerBlock = isOutermost ? headerBlock : createNextBeginBlock();
|
||||
info.bodyBlock = isInnermost ? bodyBlock : createNextBeginBlock();
|
||||
info.exitBlock = isOutermost ? exitBlock : createNextExitBlock();
|
||||
if (info.maskExpr)
|
||||
info.maskBlock = createNextBeginBlock();
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Increment loop begin code. (Infinite/while code was already generated.)
|
||||
|
|
@ -1065,8 +1139,28 @@ private:
|
|||
return builder->createRealConstant(loc, controlType, 1u);
|
||||
return builder->createIntegerConstant(loc, controlType, 1); // step
|
||||
};
|
||||
auto handleLocalitySpec = [&](IncrementLoopInfo &info) {
|
||||
// Generate Local Init Assignments
|
||||
for (const Fortran::semantics::Symbol *sym : info.localInitSymList) {
|
||||
const auto *hostDetails =
|
||||
sym->detailsIf<Fortran::semantics::HostAssocDetails>();
|
||||
assert(hostDetails && "missing local_init variable host variable");
|
||||
const Fortran::semantics::Symbol &hostSym = hostDetails->symbol();
|
||||
(void)hostSym;
|
||||
TODO(loc, "do concurrent locality specs not implemented");
|
||||
}
|
||||
// Handle shared locality spec
|
||||
for (const Fortran::semantics::Symbol *sym : info.sharedSymList) {
|
||||
const auto *hostDetails =
|
||||
sym->detailsIf<Fortran::semantics::HostAssocDetails>();
|
||||
assert(hostDetails && "missing shared variable host variable");
|
||||
const Fortran::semantics::Symbol &hostSym = hostDetails->symbol();
|
||||
copySymbolBinding(hostSym, *sym);
|
||||
}
|
||||
};
|
||||
for (IncrementLoopInfo &info : incrementLoopNestInfo) {
|
||||
info.loopVariable = genLoopVariableAddress(loc, info.loopVariableSym);
|
||||
info.loopVariable =
|
||||
genLoopVariableAddress(loc, info.loopVariableSym, info.isUnordered);
|
||||
mlir::Value lowerValue = genControlValue(info.lowerExpr, info);
|
||||
mlir::Value upperValue = genControlValue(info.upperExpr, info);
|
||||
info.stepValue = genControlValue(info.stepExpr, info);
|
||||
|
|
@ -1081,8 +1175,17 @@ private:
|
|||
mlir::Value value = builder->createConvert(
|
||||
loc, info.getLoopVariableType(), info.doLoop.getInductionVar());
|
||||
builder->create<fir::StoreOp>(loc, value, info.loopVariable);
|
||||
// TODO: Mask expr
|
||||
// TODO: handle Locality Spec
|
||||
if (info.maskExpr) {
|
||||
Fortran::lower::StatementContext stmtCtx;
|
||||
mlir::Value maskCond = createFIRExpr(loc, info.maskExpr, stmtCtx);
|
||||
stmtCtx.finalize();
|
||||
mlir::Value maskCondCast =
|
||||
builder->createConvert(loc, builder->getI1Type(), maskCond);
|
||||
auto ifOp = builder->create<fir::IfOp>(loc, maskCondCast,
|
||||
/*withElseRegion=*/false);
|
||||
builder->setInsertionPointToStart(&ifOp.getThenRegion().front());
|
||||
}
|
||||
handleLocalitySpec(info);
|
||||
continue;
|
||||
}
|
||||
|
||||
|
|
@ -1119,16 +1222,33 @@ private:
|
|||
builder->create<fir::StoreOp>(loc, lowerValue, info.loopVariable);
|
||||
|
||||
// Unstructured loop header - generate loop condition and mask.
|
||||
// Note - Currently there is no way to tag a loop as a concurrent loop.
|
||||
startBlock(info.headerBlock);
|
||||
tripCount = builder->create<fir::LoadOp>(loc, info.tripVariable);
|
||||
mlir::Value zero =
|
||||
builder->createIntegerConstant(loc, tripCount.getType(), 0);
|
||||
auto cond = builder->create<mlir::arith::CmpIOp>(
|
||||
loc, mlir::arith::CmpIPredicate::sgt, tripCount, zero);
|
||||
// TODO: mask expression
|
||||
genFIRConditionalBranch(cond, info.bodyBlock, info.exitBlock);
|
||||
if (&info != &incrementLoopNestInfo.back()) // not innermost
|
||||
startBlock(info.bodyBlock); // preheader block of enclosed dimension
|
||||
if (info.maskExpr) {
|
||||
genFIRConditionalBranch(cond, info.maskBlock, info.exitBlock);
|
||||
startBlock(info.maskBlock);
|
||||
mlir::Block *latchBlock = getEval().getLastNestedEvaluation().block;
|
||||
assert(latchBlock && "missing masked concurrent loop latch block");
|
||||
Fortran::lower::StatementContext stmtCtx;
|
||||
mlir::Value maskCond = createFIRExpr(loc, info.maskExpr, stmtCtx);
|
||||
stmtCtx.finalize();
|
||||
genFIRConditionalBranch(maskCond, info.bodyBlock, latchBlock);
|
||||
} else {
|
||||
genFIRConditionalBranch(cond, info.bodyBlock, info.exitBlock);
|
||||
if (&info != &incrementLoopNestInfo.back()) // not innermost
|
||||
startBlock(info.bodyBlock); // preheader block of enclosed dimension
|
||||
}
|
||||
if (!info.localInitSymList.empty()) {
|
||||
mlir::OpBuilder::InsertPoint insertPt = builder->saveInsertionPoint();
|
||||
builder->setInsertionPointToStart(info.bodyBlock);
|
||||
handleLocalitySpec(info);
|
||||
builder->restoreInsertionPoint(insertPt);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -59,8 +59,8 @@ subroutine ss2(n) ! unstructured OpenMP construct; loop exit inside construct
|
|||
end
|
||||
|
||||
! CHECK-LABEL: func @_QPss3{{.*}} {
|
||||
! CHECK: %[[ALLOCA_K:.*]] = fir.alloca i32 {bindc_name = "k", {{.*}}}
|
||||
! CHECK: omp.parallel {
|
||||
! CHECK: %[[ALLOCA_K:.*]] = fir.alloca i32 {bindc_name = "k", pinned}
|
||||
! CHECK: %[[ALLOCA_1:.*]] = fir.alloca i32 {{{.*}}, pinned}
|
||||
! CHECK: %[[ALLOCA_2:.*]] = fir.alloca i32 {{{.*}}, pinned}
|
||||
! CHECK: br ^bb1
|
||||
|
|
|
|||
|
|
@ -0,0 +1,109 @@
|
|||
! RUN: bbc -emit-fir -o - %s | FileCheck %s
|
||||
|
||||
! CHECK-LABEL: loop_test
|
||||
subroutine loop_test
|
||||
! CHECK: %[[VAL_2:.*]] = fir.alloca i16 {bindc_name = "i"}
|
||||
! CHECK: %[[VAL_3:.*]] = fir.alloca i16 {bindc_name = "i"}
|
||||
! CHECK: %[[VAL_4:.*]] = fir.alloca i16 {bindc_name = "i"}
|
||||
! CHECK: %[[VAL_5:.*]] = fir.alloca i8 {bindc_name = "k"}
|
||||
! CHECK: %[[VAL_6:.*]] = fir.alloca i8 {bindc_name = "j"}
|
||||
! CHECK: %[[VAL_7:.*]] = fir.alloca i8 {bindc_name = "i"}
|
||||
! CHECK: %[[VAL_8:.*]] = fir.alloca i32 {bindc_name = "k"}
|
||||
! CHECK: %[[VAL_9:.*]] = fir.alloca i32 {bindc_name = "j"}
|
||||
! CHECK: %[[VAL_10:.*]] = fir.alloca i32 {bindc_name = "i"}
|
||||
! CHECK: %[[VAL_11:.*]] = fir.alloca !fir.array<5x5x5xi32> {bindc_name = "a", uniq_name = "_QFloop_testEa"}
|
||||
! CHECK: %[[VAL_12:.*]] = fir.alloca i32 {bindc_name = "asum", uniq_name = "_QFloop_testEasum"}
|
||||
! CHECK: %[[VAL_13:.*]] = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFloop_testEi"}
|
||||
! CHECK: %[[VAL_14:.*]] = fir.alloca i32 {bindc_name = "j", uniq_name = "_QFloop_testEj"}
|
||||
! CHECK: %[[VAL_15:.*]] = fir.alloca i32 {bindc_name = "k", uniq_name = "_QFloop_testEk"}
|
||||
! CHECK: %[[VAL_16:.*]] = fir.alloca f32 {bindc_name = "x", uniq_name = "_QFloop_testEx"}
|
||||
! CHECK: %[[VAL_17:.*]] = fir.alloca i32 {bindc_name = "xsum", uniq_name = "_QFloop_testExsum"}
|
||||
|
||||
integer(4) :: a(5,5,5), i, j, k, asum, xsum
|
||||
|
||||
i = 100
|
||||
j = 200
|
||||
k = 300
|
||||
|
||||
! CHECK-COUNT-3: fir.do_loop {{.*}} unordered
|
||||
do concurrent (i=1:5, j=1:5, k=1:5) ! shared(a)
|
||||
! CHECK: fir.coordinate_of
|
||||
a(i,j,k) = 0
|
||||
enddo
|
||||
! CHECK: fir.call @_FortranAioBeginExternalListOutput
|
||||
print*, 'A:', i, j, k
|
||||
|
||||
! CHECK-COUNT-3: fir.do_loop {{.*}} unordered
|
||||
! CHECK: fir.if
|
||||
do concurrent (integer(1)::i=1:5, j=1:5, k=1:5, i.ne.j .and. k.ne.3) shared(a)
|
||||
! CHECK-COUNT-2: fir.coordinate_of
|
||||
a(i,j,k) = a(i,j,k) + 1
|
||||
enddo
|
||||
|
||||
! CHECK-COUNT-3: fir.do_loop {{[^un]*}} -> index
|
||||
asum = 0
|
||||
do i=1,5
|
||||
do j=1,5
|
||||
do k=1,5
|
||||
! CHECK: fir.coordinate_of
|
||||
asum = asum + a(i,j,k)
|
||||
enddo
|
||||
enddo
|
||||
enddo
|
||||
! CHECK: fir.call @_FortranAioBeginExternalListOutput
|
||||
print*, 'B:', i, j, k, '-', asum
|
||||
|
||||
! CHECK: fir.do_loop {{.*}} unordered
|
||||
! CHECK-COUNT-2: fir.if
|
||||
do concurrent (integer(2)::i=1:5, i.ne.3)
|
||||
if (i.eq.2 .or. i.eq.4) goto 5 ! fir.if
|
||||
! CHECK: fir.call @_FortranAioBeginExternalListOutput
|
||||
print*, 'C:', i
|
||||
5 continue
|
||||
enddo
|
||||
|
||||
! CHECK: fir.do_loop {{.*}} unordered
|
||||
! CHECK-COUNT-2: fir.if
|
||||
do concurrent (integer(2)::i=1:5, i.ne.3)
|
||||
if (i.eq.2 .or. i.eq.4) then ! fir.if
|
||||
goto 6
|
||||
endif
|
||||
! CHECK: fir.call @_FortranAioBeginExternalListOutput
|
||||
print*, 'D:', i
|
||||
6 continue
|
||||
enddo
|
||||
|
||||
! CHECK-NOT: fir.do_loop
|
||||
! CHECK-NOT: fir.if
|
||||
do concurrent (integer(2)::i=1:5, i.ne.3)
|
||||
goto (7, 7) i+1
|
||||
! CHECK: fir.call @_FortranAioBeginExternalListOutput
|
||||
print*, 'E:', i
|
||||
7 continue
|
||||
enddo
|
||||
|
||||
xsum = 0.0
|
||||
! CHECK-NOT: fir.do_loop
|
||||
do x = 1.5, 3.5, 0.3
|
||||
xsum = xsum + 1
|
||||
enddo
|
||||
! CHECK: fir.call @_FortranAioBeginExternalFormattedOutput
|
||||
print '(" F:",X,F3.1,A,I2)', x, ' -', xsum
|
||||
end subroutine loop_test
|
||||
|
||||
! CHECK-LABEL: print_nothing
|
||||
subroutine print_nothing(k1, k2)
|
||||
if (k1 > 0) then
|
||||
! CHECK: br [[header:\^bb[0-9]+]]
|
||||
! CHECK: [[header]]
|
||||
do while (k1 > k2)
|
||||
print*, k1, k2 ! no output
|
||||
k2 = k2 + 1
|
||||
! CHECK: br [[header]]
|
||||
end do
|
||||
end if
|
||||
end
|
||||
|
||||
call loop_test
|
||||
call print_nothing(2, 2)
|
||||
end
|
||||
Loading…
Reference in New Issue