Skip to content
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

Jump to bottom

[SLP]Add cost estimation for gather node reshuffling #115201

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
+530 −320
Open

[SLP]Add cost estimation for gather node reshuffling #115201

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9e240ac
[𝘀𝗽𝗿] initial version
alexey-bataev Nov 6, 2024
7420027
Address comments
alexey-bataev Nov 15, 2024
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
173 changes: 156 additions & 17 deletions llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4846,8 +4846,21 @@ getShuffleCost(const TargetTransformInfo &TTI, TTI::ShuffleKind Kind,
TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput,
int Index = 0, VectorType *SubTp = nullptr,
ArrayRef<const Value *> Args = {}) {
if (Kind != TTI::SK_PermuteTwoSrc)
if (Kind != TTI::SK_PermuteTwoSrc) {
int SplatIdx = PoisonMaskElem;
if (!Mask.empty() && all_of(Mask, [&](int Idx) {
if (Idx == PoisonMaskElem)
return true;
if (SplatIdx == PoisonMaskElem) {
SplatIdx = Idx;
return true;
}
return SplatIdx == Idx;
}))
return TTI.getShuffleCost(TTI::SK_Broadcast, Tp, Mask, CostKind, Index,
SubTp, Args);
return TTI.getShuffleCost(Kind, Tp, Mask, CostKind, Index, SubTp, Args);
}
alexey-bataev marked this conversation as resolved.
Show resolved Hide resolved
int NumSrcElts = Tp->getElementCount().getKnownMinValue();
int NumSubElts;
if (Mask.size() > 2 && ShuffleVectorInst::isInsertSubvectorMask(
Expand Down Expand Up @@ -10257,10 +10270,10 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
Idx = EMask[Idx];
}
CommonVF = E->Scalars.size();
} else if (std::optional<unsigned> Factor = E->getInterleaveFactor();
Factor && E->Scalars.size() != Mask.size() &&
} else if (unsigned Factor = E->getInterleaveFactor();
Factor > 0 && E->Scalars.size() != Mask.size() &&
ShuffleVectorInst::isDeInterleaveMaskOfFactor(CommonMask,
*Factor)) {
Factor)) {
// Deinterleaved nodes are free.
std::iota(CommonMask.begin(), CommonMask.end(), 0);
}
Expand Down Expand Up @@ -12935,6 +12948,7 @@ BoUpSLP::isGatherShuffledSingleRegisterEntry(
// No perfect match, just shuffle, so choose the first tree node from the
// tree.
Entries.push_back(FirstEntries.front());
VF = FirstEntries.front()->getVectorFactor();
} else {
// Try to find nodes with the same vector factor.
assert(UsedTEs.size() == 2 && "Expected at max 2 permuted entries.");
Expand Down Expand Up @@ -12975,6 +12989,8 @@ BoUpSLP::isGatherShuffledSingleRegisterEntry(
Entries.push_back(SecondEntries.front());
VF = std::max(Entries.front()->getVectorFactor(),
Entries.back()->getVectorFactor());
} else {
VF = Entries.front()->getVectorFactor();
}
}

Expand Down Expand Up @@ -13077,26 +13093,149 @@ BoUpSLP::isGatherShuffledSingleRegisterEntry(
// Pair.first is the offset to the vector, while Pair.second is the index of
// scalar in the list.
for (const std::pair<unsigned, int> &Pair : EntryLanes) {
unsigned Idx = Part * VL.size() + Pair.second;
int Idx = Part * VL.size() + Pair.second;
Mask[Idx] =
Pair.first * VF +
(ForOrder ? std::distance(
Entries[Pair.first]->Scalars.begin(),
find(Entries[Pair.first]->Scalars, VL[Pair.second]))
: Entries[Pair.first]->findLaneForValue(VL[Pair.second]));
IsIdentity &= Mask[Idx] == Pair.second;
IsIdentity &= Mask[Idx] % VL.size() == Idx % VL.size();
}
switch (Entries.size()) {
case 1:
if (IsIdentity || EntryLanes.size() > 1 || VL.size() <= 2)
return TargetTransformInfo::SK_PermuteSingleSrc;
break;
case 2:
if (EntryLanes.size() > 2 || VL.size() <= 2)
return TargetTransformInfo::SK_PermuteTwoSrc;
break;
default:
break;
if (ForOrder || IsIdentity || Entries.empty()) {
switch (Entries.size()) {
case 1:
if (IsIdentity || EntryLanes.size() > 1 || VL.size() <= 2)
return TargetTransformInfo::SK_PermuteSingleSrc;
break;
case 2:
if (EntryLanes.size() > 2 || VL.size() <= 2)
return TargetTransformInfo::SK_PermuteTwoSrc;
break;
default:
break;
}
} else if (!isa<VectorType>(VL.front()->getType()) &&
(EntryLanes.size() > Entries.size() || VL.size() <= 2)) {
// Do the cost estimation if shuffle beneficial than buildvector.
SmallVector<int> SubMask(std::next(Mask.begin(), Part * VL.size()),
std::next(Mask.begin(), (Part + 1) * VL.size()));
int MinElement = SubMask.front(), MaxElement = SubMask.front();
for (int Idx : SubMask) {
if (Idx == PoisonMaskElem)
continue;
if (MinElement == PoisonMaskElem || MinElement % VF > Idx % VF)
MinElement = Idx;
if (MaxElement == PoisonMaskElem || MaxElement % VF < Idx % VF)
MaxElement = Idx;
}
alexey-bataev marked this conversation as resolved.
Show resolved Hide resolved
assert(MaxElement >= 0 && MinElement >= 0 &&
"Expected at least single element.");
alexey-bataev marked this conversation as resolved.
Show resolved Hide resolved
unsigned NewVF = std::max<unsigned>(
VL.size(), getFullVectorNumberOfElements(*TTI, VL.front()->getType(),
(MaxElement % VF) -
(MinElement % VF) + 1));
if (NewVF < VF) {
for_each(SubMask, [&](int &Idx) {
if (Idx == PoisonMaskElem)
return;
Idx = (Idx % VF) - (MinElement % VF) +
(Idx >= static_cast<int>(VF) ? NewVF : 0);
});
VF = NewVF;
}

constexpr TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
auto *VecTy = getWidenedType(VL.front()->getType(), VF);
auto *MaskVecTy = getWidenedType(VL.front()->getType(), SubMask.size());
auto GetShuffleCost = [&,
&TTI = *TTI](ArrayRef<int> Mask,
ArrayRef<const TreeEntry *> Entries,
VectorType *VecTy) -> InstructionCost {
if (Entries.size() == 1 && Entries.front()->getInterleaveFactor() > 0 &&
ShuffleVectorInst::isDeInterleaveMaskOfFactor(
Mask, Entries.front()->getInterleaveFactor()))
return TTI::TCC_Free;
return ::getShuffleCost(TTI,
Entries.size() > 1 ? TTI::SK_PermuteTwoSrc
: TTI::SK_PermuteSingleSrc,
VecTy, Mask, CostKind);
};
InstructionCost ShuffleCost = GetShuffleCost(SubMask, Entries, VecTy);
InstructionCost FirstShuffleCost = 0;
SmallVector<int> FirstMask(SubMask.begin(), SubMask.end());
if (Entries.size() == 1 || !Entries[0]->isGather()) {
FirstShuffleCost = ShuffleCost;
} else {
// Transform mask to include only first entry.
APInt DemandedElts = APInt::getAllOnes(SubMask.size());
bool IsIdentity = true;
for (auto [I, Idx] : enumerate(FirstMask)) {
if (Idx >= static_cast<int>(VF)) {
Idx = PoisonMaskElem;
} else {
DemandedElts.clearBit(I);
if (Idx != PoisonMaskElem)
IsIdentity &= static_cast<int>(I) == Idx;
}
}
if (!IsIdentity)
FirstShuffleCost = GetShuffleCost(FirstMask, Entries.front(), VecTy);
FirstShuffleCost += TTI->getScalarizationOverhead(
MaskVecTy, DemandedElts, /*Insert=*/true,
/*Extract=*/false, CostKind);
}
InstructionCost SecondShuffleCost = 0;
SmallVector<int> SecondMask(SubMask.begin(), SubMask.end());
if (Entries.size() == 1 || !Entries[1]->isGather()) {
SecondShuffleCost = ShuffleCost;
} else {
// Transform mask to include only first entry.
APInt DemandedElts = APInt::getAllOnes(SubMask.size());
bool IsIdentity = true;
for (auto [I, Idx] : enumerate(SecondMask)) {
if (Idx < static_cast<int>(VF) && Idx >= 0) {
Idx = PoisonMaskElem;
} else {
DemandedElts.clearBit(I);
if (Idx != PoisonMaskElem) {
Idx -= VF;
IsIdentity &= static_cast<int>(I) == Idx;
}
}
}
if (!IsIdentity)
SecondShuffleCost = GetShuffleCost(SecondMask, Entries[1], VecTy);
SecondShuffleCost += TTI->getScalarizationOverhead(
MaskVecTy, DemandedElts, /*Insert=*/true,
/*Extract=*/false, CostKind);
}
APInt DemandedElts = APInt::getAllOnes(SubMask.size());
for (auto [I, Idx] : enumerate(SubMask))
if (Idx == PoisonMaskElem)
DemandedElts.clearBit(I);
InstructionCost BuildVectorCost =
TTI->getScalarizationOverhead(MaskVecTy, DemandedElts, /*Insert=*/true,
/*Extract=*/false, CostKind);
const TreeEntry *BestEntry = nullptr;
if (FirstShuffleCost < ShuffleCost) {
copy(FirstMask, std::next(Mask.begin(), Part * VL.size()));
BestEntry = Entries.front();
ShuffleCost = FirstShuffleCost;
}
if (SecondShuffleCost < ShuffleCost) {
copy(SecondMask, std::next(Mask.begin(), Part * VL.size()));
BestEntry = Entries[1];
ShuffleCost = SecondShuffleCost;
}
if (BuildVectorCost >= ShuffleCost) {
if (BestEntry) {
Entries.clear();
Entries.push_back(BestEntry);
}
return Entries.size() > 1 ? TargetTransformInfo::SK_PermuteTwoSrc
: TargetTransformInfo::SK_PermuteSingleSrc;
}
}
Entries.clear();
// Clear the corresponding mask elements.
Expand Down
8 changes: 4 additions & 4 deletions llvm/test/Transforms/SLPVectorizer/AArch64/reused-scalar-repeated-in-node.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,12 +46,12 @@ define void @test() {
; CHECK-NEXT: [[TMP16:%.*]] = phi <2 x float> [ poison, %[[BB77]] ], [ [[TMP31:%.*]], %[[BB78]] ]
; CHECK-NEXT: [[TMP17:%.*]] = shufflevector <8 x float> [[TMP15]], <8 x float> poison, <16 x i32> <i32 0, i32 3, i32 1, i32 2, i32 3, i32 0, i32 2, i32 3, i32 2, i32 6, i32 2, i32 3, i32 0, i32 7, i32 6, i32 6>
; CHECK-NEXT: [[TMP18:%.*]] = fmul fast <16 x float> [[TMP17]], [[TMP13]]
; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <8 x float> [[TMP15]], <8 x float> poison, <16 x i32> <i32 1, i32 poison, i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 1, i32 6, i32 7, i32 7>
; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <8 x float> [[TMP15]], <8 x float> poison, <16 x i32> <i32 1, i32 poison, i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP20:%.*]] = shufflevector <2 x float> [[TMP16]], <2 x float> poison, <16 x i32> <i32 0, i32 1, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP21:%.*]] = shufflevector <16 x float> [[TMP19]], <16 x float> [[TMP20]], <16 x i32> <i32 0, i32 17, i32 2, i32 16, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 12, i32 13, i32 14, i32 15>
; CHECK-NEXT: [[TMP21:%.*]] = shufflevector <16 x float> [[TMP19]], <16 x float> [[TMP20]], <16 x i32> <i32 0, i32 17, i32 2, i32 16, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP22:%.*]] = shufflevector <8 x float> [[TMP15]], <8 x float> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP23:%.*]] = shufflevector <16 x float> [[TMP21]], <16 x float> [[TMP22]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 17, i32 6, i32 7, i32 8, i32 23, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <16 x float> [[TMP23]], <16 x float> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 1, i32 5, i32 3, i32 1, i32 3, i32 9, i32 3, i32 1, i32 12, i32 13, i32 14, i32 15>
; CHECK-NEXT: [[TMP23:%.*]] = shufflevector <16 x float> [[TMP21]], <16 x float> [[TMP22]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 17, i32 6, i32 7, i32 8, i32 23, i32 10, i32 11, i32 12, i32 22, i32 14, i32 15>
; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <16 x float> [[TMP23]], <16 x float> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 1, i32 5, i32 3, i32 1, i32 3, i32 9, i32 3, i32 1, i32 5, i32 13, i32 9, i32 9>
; CHECK-NEXT: [[TMP25:%.*]] = call <16 x float> @llvm.vector.insert.v16f32.v2f32(<16 x float> [[TMP14]], <2 x float> [[TMP0]], i64 2)
; CHECK-NEXT: [[TMP26:%.*]] = fmul fast <16 x float> [[TMP24]], [[TMP25]]
; CHECK-NEXT: [[TMP27:%.*]] = fadd fast <16 x float> [[TMP26]], [[TMP18]]
Expand Down
Loading
Loading