Skip to content

Fix calling convention and other small bugs #16

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 12 commits into from
Oct 15, 2019
Merged

Fix calling convention and other small bugs #16

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
d551bb1
Fixing code for call gradient
wsmoses Oct 10, 2019
1a613ad
fixing vector bug
wsmoses Oct 10, 2019
3741228
Fix various calling convention issues
wsmoses Oct 12, 2019
aa4f76d
further calling convention fixes
wsmoses Oct 12, 2019
a025080
lookat ifcrash
wsmoses Oct 12, 2019
098aa5b
resolve fn tests
wsmoses Oct 12, 2019
60820fd
look at weird fail
wsmoses Oct 12, 2019
d04d3e4
fix subtle if bug
wsmoses Oct 12, 2019
3412e9c
fix julia
wsmoses Oct 12, 2019
1498c0e
update tests
wsmoses Oct 12, 2019
a8bdf3b
revamp loop calc
wsmoses Oct 12, 2019
ae6bee0
add more caching layers to speed compilation
wsmoses Oct 12, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
fixing vector bug
  • Loading branch information
@wsmoses
wsmoses committed Oct 12, 2019
commit 1a613ad20d3b5d56a70d8cb0d315b27fb6ffc183
66 changes: 39 additions & 27 deletions enzyme/Enzyme/EnzymeLogic.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,7 @@ std::pair<Function*,StructType*> CreateAugmentedPrimal(Function* todiff, AAResul
if (cachedfunctions.find(tup) != cachedfunctions.end()) {
return cachedfunctions[tup];
}
llvm::errs() << "creating augmented primal for " << todiff->getName() << "\n";

if (constant_args.size() == 0 && hasMetadata(todiff, "enzyme_augment")) {
auto md = todiff->getMetadata("enzyme_augment");
Expand Down Expand Up @@ -109,6 +110,7 @@ std::pair<Function*,StructType*> CreateAugmentedPrimal(Function* todiff, AAResul
gutils->forceAugmentedReturns();

for(BasicBlock* BB: gutils->originalBlocks) {
llvm::errs() << " augment block" << BB->getName() << "\n";
auto term = BB->getTerminator();
assert(term);
if(auto ri = dyn_cast<ReturnInst>(term)) {
Expand All @@ -133,11 +135,13 @@ std::pair<Function*,StructType*> CreateAugmentedPrimal(Function* todiff, AAResul
}

if (!isa<UnreachableInst>(term))
for (auto I = BB->rbegin(), E = BB->rend(); I != E;) {
for (BasicBlock::reverse_iterator I = BB->rbegin(), E = BB->rend(); I != E;) {
Instruction* inst = &*I;
assert(inst);
I++;
if (gutils->originalInstructions.find(inst) == gutils->originalInstructions.end()) continue;

llvm::errs() << " augment inst" << *inst << "\n";

if(auto op = dyn_cast_or_null<IntrinsicInst>(inst)) {
switch(op->getIntrinsicID()) {
Expand Down Expand Up @@ -195,7 +199,11 @@ std::pair<Function*,StructType*> CreateAugmentedPrimal(Function* todiff, AAResul
case Intrinsic::lifetime_end:
case Intrinsic::assume:
case Intrinsic::fabs:
case Intrinsic::x86_sse_max_ss:
case Intrinsic::x86_sse_max_ps:
case Intrinsic::maxnum:
case Intrinsic::x86_sse_min_ss:
case Intrinsic::x86_sse_min_ps:
case Intrinsic::minnum:
case Intrinsic::log:
case Intrinsic::log2:
Expand Down Expand Up @@ -321,7 +329,7 @@ std::pair<Function*,StructType*> CreateAugmentedPrimal(Function* todiff, AAResul
gutils->nonconstant_values.insert(rv);
}
assert(op->getType() == rv->getType());
llvm::errs() << "augmented considering differential ip of " << called->getName() << " " << *op->getType() << " " << gutils->isConstantValue(op) << "\n";
//llvm::errs() << "augmented considering differential ip of " << called->getName() << " " << *op->getType() << " " << gutils->isConstantValue(op) << "\n";

if ((op->getType()->isPointerTy() || op->getType()->isIntegerTy()) && !gutils->isConstantValue(op)) {
auto antiptr = cast<Instruction>(BuilderZ.CreateExtractValue(augmentcall, {2}));
Expand Down Expand Up @@ -422,7 +430,13 @@ std::pair<Function*,StructType*> CreateAugmentedPrimal(Function* todiff, AAResul
MallocTypes.push_back(a->getType());
}

StructType* tapeType = StructType::get(nf->getContext(), MallocTypes);
StructType* tapeType;

//if (MallocTypes.size() > 1) {
// tapeType = StructType::create(MallocTypes, (todiff->getName()+"_tapeType").str(), false);
//} else {
tapeType = StructType::get(nf->getContext(), MallocTypes);
//}

bool recursive = cachedfunctions[tup].first->getNumUses() > 0;

Expand Down Expand Up @@ -674,8 +688,16 @@ void createInvertedTerminator(DiffeGradientUtils* gutils, BasicBlock *BB, Alloca
setphi = true;
}
}
Instruction* dif = cast<Instruction>(Builder.CreateSelect(replacePHIs[pred], prediff, Constant::getNullValue(prediff->getType())));
SelectInst* dif = cast<SelectInst>(Builder.CreateSelect(replacePHIs[pred], prediff, Constant::getNullValue(prediff->getType())));
auto addedSelects = gutils->addToDiffe(PN->getIncomingValueForBlock(pred), dif, Builder);
if (dif->getNumUses() != 0) {
llvm::errs() << "oldFunc: " << *gutils->oldFunc << "\n";
llvm::errs() << "newFunc: " << *gutils->newFunc << "\n";
for (auto use : dif->users()) {
llvm::errs() << "user: " << *use << "\n";
}
llvm::errs() << "dif: " << *dif << "\n";
}
assert(dif->getNumUses() == 0);
dif->eraseFromParent();
for (auto select : addedSelects)
Expand Down Expand Up @@ -1403,6 +1425,7 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
}
return cachedfunctions[tup] = foundcalled;
}
llvm::errs() << "creating gradient for " << todiff->getName() << " topLevel:" << topLevel << "\n";

assert(!todiff->empty());
auto M = todiff->getParent();
Expand Down Expand Up @@ -1445,6 +1468,7 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
std::map<ReturnInst*,StoreInst*> replacedReturns;

for(BasicBlock* BB: gutils->originalBlocks) {
llvm::errs() << " gradient block" << BB->getName() << "\n";

LoopContext loopContext;
bool inLoop = gutils->getContext(BB, loopContext);
Expand Down Expand Up @@ -1479,22 +1503,10 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
gutils->setDiffe(val, toset, Builder2);
};

auto addToDiffeIndexed = [&](Value* val, Value* dif, ArrayRef<Value*> idxs) -> void{
gutils->addToDiffeIndexed(val, dif, idxs, Builder2);
};

auto invertPointer = [&](Value* val) -> Value* {
return gutils->invertPointerM(val, Builder2);
};

auto addToPtrDiffe = [&](Value* val, Value* dif) {
gutils->addToPtrDiffe(val, dif, Builder2);
};

auto setPtrDiffe = [&](Value* val, Value* dif) {
gutils->setPtrDiffe(val, dif, Builder2);
};

auto term = BB->getTerminator();
assert(term);
bool unreachableTerminator = false;
Expand Down Expand Up @@ -1539,6 +1551,8 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
assert(inst);
I++;
if (gutils->originalInstructions.find(inst) == gutils->originalInstructions.end()) continue;

llvm::errs() << " gradient inst" << *inst << "\n";

if (auto op = dyn_cast<BinaryOperator>(inst)) {
if (gutils->isConstantInstruction(inst)) continue;
Expand Down Expand Up @@ -1687,6 +1701,8 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
}
break;
}
case Intrinsic::x86_sse_max_ss:
case Intrinsic::x86_sse_max_ps:
case Intrinsic::maxnum: {
if (!gutils->isConstantInstruction(op) && !gutils->isConstantValue(op->getOperand(0))) {
auto cmp = Builder2.CreateFCmpOLT(lookup(op->getOperand(0)), lookup(op->getOperand(1)));
Expand All @@ -1698,6 +1714,8 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
}
break;
}
case Intrinsic::x86_sse_min_ss:
case Intrinsic::x86_sse_min_ps:
case Intrinsic::minnum: {
if (!gutils->isConstantInstruction(op) && !gutils->isConstantValue(op->getOperand(0))) {
auto cmp = Builder2.CreateFCmpOLT(lookup(op->getOperand(0)), lookup(op->getOperand(1)));
Expand Down Expand Up @@ -1828,7 +1846,7 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
if (!op->getType()->isPointerTy()) {
auto prediff = diffe(inst);
setDiffe(inst, Constant::getNullValue(inst->getType()));
addToPtrDiffe(op->getOperand(0), prediff);
gutils->addToPtrDiffe(op->getOperand(0), prediff, Builder2);
} else {
//Builder2.CreateStore(diffe(inst), invertPointer(op->getOperand(0)));//, op->getName()+"'psweird");
//addToNPtrDiffe(op->getOperand(0), diffe(inst));
Expand All @@ -1844,7 +1862,7 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
if (! ( op->getValueOperand()->getType()->isPointerTy() || (op->getValueOperand()->getType()->isIntegerTy() && !isIntASecretFloat(op->getValueOperand()) ) ) ) {
if (!gutils->isConstantValue(op->getValueOperand())) {
auto dif1 = Builder2.CreateLoad(invertPointer(op->getPointerOperand()));
setPtrDiffe(op->getPointerOperand(), Constant::getNullValue(op->getValueOperand()->getType()));
gutils->setPtrDiffe(op->getPointerOperand(), Constant::getNullValue(op->getValueOperand()->getType()), Builder2);
addToDiffe(op->getValueOperand(), dif1);
}
} else if (topLevel) {
Expand All @@ -1856,12 +1874,6 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
storeBuilder.CreateStore(valueop, pointerop);
//llvm::errs() << "ignoring store bc pointer of " << *op << "\n";
}
//?necessary if pointer is readwrite
/*
IRBuilder<> BuilderZ(inst);
Builder2.CreateStore(
lookup(BuilderZ.CreateLoad(op->getPointerOperand())), lookup(op->getPointerOperand()));
*/
} else if(auto op = dyn_cast<ExtractValueInst>(inst)) {
if (gutils->isConstantValue(inst)) continue;
if (op->getType()->isPointerTy()) continue;
Expand All @@ -1872,7 +1884,7 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
SmallVector<Value*,4> sv;
for(auto i : op->getIndices())
sv.push_back(ConstantInt::get(Type::getInt32Ty(Context), i));
addToDiffeIndexed(op->getOperand(0), prediff, sv);
gutils->addToDiffeIndexed(op->getOperand(0), prediff, sv, Builder2);
}
setDiffe(inst, Constant::getNullValue(inst->getType()));
} else if(auto op = dyn_cast<InsertValueInst>(inst)) {
Expand Down Expand Up @@ -1911,7 +1923,7 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
SmallVector<Value*,4> sv;
sv.push_back(ConstantInt::get(Type::getInt32Ty(Context), opidx));
if (!gutils->isConstantValue(op->getOperand(opnum)))
addToDiffeIndexed(op->getOperand(opnum), Builder2.CreateExtractElement(loaded, instidx), sv);
gutils->addToDiffeIndexed(op->getOperand(opnum), Builder2.CreateExtractElement(loaded, instidx), sv, Builder2);
instidx++;
}
setDiffe(inst, Constant::getNullValue(inst->getType()));
Expand All @@ -1921,7 +1933,7 @@ Function* CreatePrimalAndGradient(Function* todiff, const std::set<unsigned>& co
if (!gutils->isConstantValue(op->getVectorOperand())) {
SmallVector<Value*,4> sv;
sv.push_back(op->getIndexOperand());
addToDiffeIndexed(op->getVectorOperand(), diffe(inst), sv);
gutils->addToDiffeIndexed(op->getVectorOperand(), diffe(inst), sv, Builder2);
}
setDiffe(inst, Constant::getNullValue(inst->getType()));
} else if(auto op = dyn_cast<InsertElementInst>(inst)) {
Expand Down
13 changes: 8 additions & 5 deletions enzyme/Enzyme/GradientUtils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -503,11 +503,14 @@ class GradientUtils {
erase(cast<Instruction>(malloc));
ret->setName(n);
}
llvm::errs() << " retrieved from malloc " << *ret << "\n";
return ret;
} else {
assert(malloc);
assert(!isa<PHINode>(malloc));

llvm::errs() << " adding to malloc " << *malloc << "\n";

if (isa<UndefValue>(malloc)) {
addedMallocs.push_back(malloc);
return malloc;
Expand Down Expand Up @@ -1274,16 +1277,16 @@ class DiffeGradientUtils : public GradientUtils {

//! optimize fadd of select to select of fadd
if (SelectInst* select = dyn_cast<SelectInst>(dif)) {
if (ConstantFP* ci = dyn_cast<ConstantFP>(select->getTrueValue())) {
if (ci->isZero()) {
if (Constant* ci = dyn_cast<Constant>(select->getTrueValue())) {
if (ci->isZeroValue()) {
cast<Instruction>(res)->eraseFromParent();
res = BuilderM.CreateSelect(select->getCondition(), old, BuilderM.CreateFAdd(old, select->getFalseValue()));
addedSelects.emplace_back(cast<SelectInst>(res));
goto endselect;
}
}
if (ConstantFP* ci = dyn_cast<ConstantFP>(select->getFalseValue())) {
if (ci->isZero()) {
if (Constant* ci = dyn_cast<Constant>(select->getFalseValue())) {
if (ci->isZeroValue()) {
cast<Instruction>(res)->eraseFromParent();
res = BuilderM.CreateSelect(select->getCondition(), BuilderM.CreateFAdd(old, select->getTrueValue()), old);
addedSelects.emplace_back(cast<SelectInst>(res));
Expand All @@ -1306,7 +1309,7 @@ class DiffeGradientUtils : public GradientUtils {
}
return addedSelects;
} else {
assert(0 && "lol");
llvm_unreachable("unknown type to add to diffe");
exit(1);
}
}
Expand Down
68 changes: 68 additions & 0 deletions enzyme/test/Enzyme/vecrelu.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,68 @@
; RUN: opt < %s %loadEnzyme -enzyme -enzyme_preopt=false -mem2reg -inline -early-cse -instcombine -simplifycfg -S | FileCheck %s

; __attribute__((noinline))
; double f(double x) {
; return x;
; }
;
; double relu(double x) {
; return (x > 0) ? f(x) : 0;
; }
;
; double drelu(double x) {
; return __builtin_autodiff(relu, x);
; }

define dso_local <4 x double> @f(<4 x double> %x) #1 {
entry:
ret <4 x double> %x
}

define dso_local i1 @cmp() #1 {
entry:
ret i1 true
}

define dso_local <4 x double> @relu(<4 x double> %x) {
entry:
%cmp = call i1 @cmp()
br i1 %cmp, label %cond.true, label %cond.end

cond.true: ; preds = %entry
%call = tail call fast <4 x double> @f(<4 x double> %x)
br label %cond.end

cond.end: ; preds = %entry, %cond.true
%cond = phi <4 x double> [ %call, %cond.true ], [ zeroinitializer, %entry ]
ret <4 x double> %cond
}

define dso_local <4 x double> @drelu(<4 x double> %x) {
entry:
%0 = tail call <4 x double> (<4 x double> (<4 x double>)*, ...) @__enzyme_autodiff(<4 x double> (<4 x double>)* nonnull @relu, <4 x double> %x)
ret <4 x double> %0
}

declare <4 x double> @__enzyme_autodiff(<4 x double> (<4 x double>)*, ...) #0

attributes #0 = { nounwind }
attributes #1 = { nounwind readnone noinline }

; CHECK: define dso_local <4 x double> @drelu(<4 x double> %x)
; CHECK-NEXT: entry:
; CHECK-NEXT: %cmp.i = call i1 @cmp()
; CHECK-NEXT: br i1 %cmp.i, label %invertcond.true.i, label %differelu.exit
; CHECK: invertcond.true.i: ; preds = %entry
; CHECK-NEXT: %[[diffef:.+]] = call { <4 x double> } @diffef(<4 x double> %x, <4 x double> <double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00>)
; CHECK-NEXT: %[[result:.+]] = extractvalue { <4 x double> } %[[diffef]], 0
; CHECK-NEXT: br label %differelu.exit
; CHECK: differelu.exit: ; preds = %entry, %invertcond.true.i
; CHECK-NEXT: %"x'de.0.i" = phi <4 x double> [ %[[result]], %invertcond.true.i ], [ zeroinitializer, %entry ]
; CHECK-NEXT: ret <4 x double> %"x'de.0.i"
; CHECK-NEXT: }

; CHECK: define internal {{(dso_local )?}}{ <4 x double> } @diffef(<4 x double> %x, <4 x double> %[[differet:.+]])
; CHECK-NEXT: entry:
; CHECK-NEXT: %[[result:.+]] = insertvalue { <4 x double> } undef, <4 x double> %[[differet]], 0
; CHECK-NEXT: ret { <4 x double> } %[[result]]
; CHECK-NEXT: }