Skip to content

[RISCV] Adjust select shuffle cost to reflect mask creation cost #77963

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 1 commit into from
Jan 18, 2024
Merged

[RISCV] Adjust select shuffle cost to reflect mask creation cost #77963

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 6 additions & 5 deletions llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,8 +84,6 @@ RISCVTTIImpl::getRISCVInstructionCost(ArrayRef<unsigned> OpCodes, MVT VT,
Cost += VL;
break;
}
case RISCV::VMV_S_X:
// FIXME: VMV_S_X doesn't use LMUL, the cost should be 1
default:
Cost += LMULCost;
}
Expand Down Expand Up @@ -443,10 +441,13 @@ InstructionCost RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
// vsetivli zero, 8, e8, mf2, ta, ma (ignored)
// vmv.s.x v0, a0
// vmerge.vvm v8, v9, v8, v0
// We use 2 for the cost of the mask materialization as this is the true
// cost for small masks and most shuffles are small. At worst, this cost
// should be a very small constant for the constant pool load. As such,
// we may bias towards large selects slightly more than truely warranted.
return LT.first *
(TLI->getLMULCost(LT.second) + // FIXME: should be 1 for li
getRISCVInstructionCost({RISCV::VMV_S_X, RISCV::VMERGE_VVM},
LT.second, CostKind));
(2 + getRISCVInstructionCost({RISCV::VMERGE_VVM},
Copy link
Contributor

@arcbbb arcbbb Jan 13, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we query the cost of VMV_S_*/VMV_*_S instead? So vendors can provide their own number in getRISCVInstructionCost.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That makes sense. We would also need to address this FIXME in getRISCVInstructionCost too:

llvm-project/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp

Lines 87 to 90 in e040252

case RISCV::VMV_S_X:
// FIXME: VMV_S_X doesn't use LMUL, the cost should be 1
default:
Cost += LMULCost;

Probably needs to be done in a separate patch.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we query the cost of VMV_S_/VMV__S instead? So vendors can provide their own number in getRISCVInstructionCost.

I would actively prefer not to add complexity unless needed. To you have knowledge of a particular processor which would need this override?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@lukel97 Agree. Is is not blocking issue.
@preames I would like to override VMV_X_S for sifive-x280, and VMV_S_X is for consistency.

LT.second, CostKind));
}
case TTI::SK_Broadcast: {
bool HasScalar = (Args.size() > 0) && (Operator::getOpcode(Args[0]) ==
Expand Down
Loading