Skip to content

[GlobalOptimization]fixed constexpr to avoid hoisting dequant operations #22416

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

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[GlobalOptimization]fixed constexpr to avoid hoisting dequant operations #22416

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
24 changes: 24 additions & 0 deletions compiler/src/iree/compiler/Dialect/Util/Analysis/Constant/ConstExpr.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,8 @@

#include "iree/compiler/Dialect/Util/Analysis/Constant/ConstExpr.h"

#include "iree/compiler/Dialect/LinalgExt/IR/LinalgExtDialect.h"
#include "iree/compiler/Dialect/LinalgExt/Utils/Utils.h"
#include "iree/compiler/Dialect/Util/Analysis/Constant/OpOracle.h"
#include "iree/compiler/Dialect/Util/Analysis/Explorer.h"
#include "iree/compiler/Dialect/Util/IR/UtilOps.h"
Expand Down Expand Up @@ -502,6 +504,28 @@ void ConstExprHoistingPolicy::makeDecision(
return;
}

// Check if this operation or any of its producers is a bit-extend operation
// (e.g., i8→f32 dequantization). Don't hoist these to keep quantized weights
// compact and avoid pulling dequantization into global initializers.
if (IREE::LinalgExt::isBitExtendOp(info->getOperation())) {
decision->disableHoist();
return;
}

// Also check if any producer is a bit-extend op - this catches the case
// where we're trying to hoist the result of: i8 -> sitofp -> mulf(scale)
for (auto *producer : info->producers) {
if (IREE::LinalgExt::isBitExtendOp(producer->getOperation())) {
LLVM_DEBUG({
llvm::dbgs()
<< "[ConstExprPolicy] Detected bit-extend producer, "
<< "disabling hoisting to keep quantized weights compact\n";
});
decision->disableHoist();
return;
}
}

// Otherwise, we can conditionally enable hoisting (based on cost model, etc).
// TODO: Implement further conditions.
decision->enableHoist();
Expand Down