Skip to content

[mlir][SCF] Avoid generating unnecessary div/rem operations during coalescing #91562

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
May 9, 2024
Merged

[mlir][SCF] Avoid generating unnecessary div/rem operations during coalescing #91562

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
57 changes: 47 additions & 10 deletions mlir/lib/Dialect/SCF/Utils/Utils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -544,11 +544,24 @@ static void denormalizeInductionVariable(RewriterBase &rewriter, Location loc,
static Value getProductOfIntsOrIndexes(RewriterBase &rewriter, Location loc,
ArrayRef<Value> values) {
assert(!values.empty() && "unexpected empty list");
Value productOf = values.front();
for (auto v : values.drop_front()) {
productOf = rewriter.create<arith::MulIOp>(loc, productOf, v);
std::optional<Value> productOf;
for (auto v : values) {
auto vOne = getConstantIntValue(v);
if (vOne && vOne.value() == 1)
continue;
if (productOf)
productOf =
rewriter.create<arith::MulIOp>(loc, productOf.value(), v).getResult();
else
productOf = v;
}
return productOf;
if (!productOf) {
productOf = rewriter
.create<arith::ConstantOp>(
loc, rewriter.getOneAttr(values.front().getType()))
.getResult();
}
return productOf.value();
}

/// For each original loop, the value of the
Expand All @@ -562,19 +575,43 @@ static Value getProductOfIntsOrIndexes(RewriterBase &rewriter, Location loc,
static std::pair<SmallVector<Value>, SmallPtrSet<Operation *, 2>>
delinearizeInductionVariable(RewriterBase &rewriter, Location loc,
Value linearizedIv, ArrayRef<Value> ubs) {
Value previous = linearizedIv;
SmallVector<Value> delinearizedIvs(ubs.size());
SmallPtrSet<Operation *, 2> preservedUsers;
for (unsigned i = 0, e = ubs.size(); i < e; ++i) {
unsigned idx = ubs.size() - i - 1;
if (i != 0) {

llvm::BitVector isUbOne(ubs.size());
for (auto [index, ub] : llvm::enumerate(ubs)) {
auto ubCst = getConstantIntValue(ub);
if (ubCst && ubCst.value() == 1)
isUbOne.set(index);
}
Copy link
Contributor

Choose a reason for hiding this comment

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

Are the upper bounds already normalized to unit step and zero lower bound by this point?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Yup...


// Prune the lead ubs that are all ones.
unsigned numLeadingOneUbs = 0;
for (auto [index, ub] : llvm::enumerate(ubs)) {
if (!isUbOne.test(index)) {
break;
}
delinearizedIvs[index] = rewriter.create<arith::ConstantOp>(
loc, rewriter.getZeroAttr(ub.getType()));
numLeadingOneUbs++;
}

Value previous = linearizedIv;
for (unsigned i = numLeadingOneUbs, e = ubs.size(); i < e; ++i) {
unsigned idx = ubs.size() - (i - numLeadingOneUbs) - 1;
if (i != numLeadingOneUbs && !isUbOne.test(idx + 1)) {
previous = rewriter.create<arith::DivSIOp>(loc, previous, ubs[idx + 1]);
preservedUsers.insert(previous.getDefiningOp());
}
Value iv = previous;
if (i != e - 1) {
iv = rewriter.create<arith::RemSIOp>(loc, previous, ubs[idx]);
preservedUsers.insert(iv.getDefiningOp());
if (!isUbOne.test(idx)) {
iv = rewriter.create<arith::RemSIOp>(loc, previous, ubs[idx]);
preservedUsers.insert(iv.getDefiningOp());
} else {
iv = rewriter.create<arith::ConstantOp>(
loc, rewriter.getZeroAttr(ubs[idx].getType()));
}
}
delinearizedIvs[idx] = iv;
}
Expand Down
77 changes: 77 additions & 0 deletions mlir/test/Dialect/SCF/transform-op-coalesce.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -299,3 +299,80 @@ module attributes {transform.with_named_sequence} {
// CHECK-NOT: scf.for
// CHECK: transform.named_sequence

// -----

// Check avoiding generating unnecessary operations while collapsing trip-1 loops.
func.func @trip_one_loops(%arg0 : tensor<?x?xf32>, %arg1 : index, %arg2 : index) -> tensor<?x?xf32> {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%0 = scf.for %iv0 = %c0 to %c1 step %c1 iter_args(%iter0 = %arg0) -> tensor<?x?xf32> {
%1 = scf.for %iv1 = %c0 to %c1 step %c1 iter_args(%iter1 = %iter0) -> tensor<?x?xf32> {
%2 = scf.for %iv2 = %c0 to %arg1 step %c1 iter_args(%iter2 = %iter1) -> tensor<?x?xf32> {
%3 = scf.for %iv3 = %c0 to %c1 step %c1 iter_args(%iter3 = %iter2) -> tensor<?x?xf32> {
%4 = scf.for %iv4 = %c0 to %arg2 step %c1 iter_args(%iter4 = %iter3) -> tensor<?x?xf32> {
%5 = "some_use"(%iter4, %iv0, %iv1, %iv2, %iv3, %iv4)
: (tensor<?x?xf32>, index, index, index, index, index) -> (tensor<?x?xf32>)
scf.yield %5 : tensor<?x?xf32>
}
scf.yield %4 : tensor<?x?xf32>
}
scf.yield %3 : tensor<?x?xf32>
}
scf.yield %2 : tensor<?x?xf32>
}
scf.yield %1 : tensor<?x?xf32>
} {coalesce}
return %0 : tensor<?x?xf32>
}
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["scf.for"]} attributes {coalesce} in %arg1 : (!transform.any_op) -> !transform.any_op
%1 = transform.cast %0 : !transform.any_op to !transform.op<"scf.for">
%2 = transform.loop.coalesce %1 : (!transform.op<"scf.for">) -> (!transform.op<"scf.for">)
transform.yield
}
}
// CHECK-LABEL: func @trip_one_loops
// CHECK-SAME: , %[[ARG1:.+]]: index,
// CHECK-SAME: %[[ARG2:.+]]: index)
// CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
// CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index
// CHECK: %[[UB:.+]] = arith.muli %[[ARG1]], %[[ARG2]]
// CHECK: scf.for %[[IV:.+]] = %[[C0]] to %[[UB]] step %[[C1]]
// CHECK: %[[IV1:.+]] = arith.remsi %[[IV]], %[[ARG2]]
// CHECK: %[[IV2:.+]] = arith.divsi %[[IV]], %[[ARG2]]
// CHECK: "some_use"(%{{[a-zA-Z0-9]+}}, %[[C0]], %[[C0]], %[[IV2]], %[[C0]], %[[IV1]])

// -----

// Check generating no instructions when all except one loops is non unit-trip.
func.func @all_outer_trip_one(%arg0 : tensor<?x?xf32>, %arg1 : index) -> tensor<?x?xf32> {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%0 = scf.for %iv0 = %c0 to %c1 step %c1 iter_args(%iter0 = %arg0) -> tensor<?x?xf32> {
%1 = scf.for %iv1 = %c0 to %c1 step %c1 iter_args(%iter1 = %iter0) -> tensor<?x?xf32> {
%2 = scf.for %iv2 = %c0 to %arg1 step %c1 iter_args(%iter2 = %iter1) -> tensor<?x?xf32> {
%3 = "some_use"(%iter2, %iv0, %iv1, %iv2)
: (tensor<?x?xf32>, index, index, index) -> (tensor<?x?xf32>)
scf.yield %3 : tensor<?x?xf32>
}
scf.yield %2 : tensor<?x?xf32>
}
scf.yield %1 : tensor<?x?xf32>
} {coalesce}
return %0 : tensor<?x?xf32>
}
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["scf.for"]} attributes {coalesce} in %arg1 : (!transform.any_op) -> !transform.any_op
%1 = transform.cast %0 : !transform.any_op to !transform.op<"scf.for">
%2 = transform.loop.coalesce %1 : (!transform.op<"scf.for">) -> (!transform.op<"scf.for">)
transform.yield
}
}
// CHECK-LABEL: func @all_outer_trip_one
// CHECK-SAME: , %[[ARG1:.+]]: index)
// CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
// CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index
// CHECK: scf.for %[[IV:.+]] = %[[C0]] to %[[ARG1]] step %[[C1]]
// CHECK: "some_use"(%{{[a-zA-Z0-9]+}}, %[[C0]], %[[C0]], %[[IV]])
Copy link
Contributor

Choose a reason for hiding this comment

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

Can you add a check with UB = 1 and non-zero lower bound?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

I tried that... it basically just generates the divsi/remsi instructions... I think I am going to call bankrupt on making this be "more smart" . Best way forward is to just move affine.delinearize out of affine dialect into arith dialect and then just use that instead of what we are doing here.

Copy link
Contributor

Choose a reason for hiding this comment

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

SGTM, I was mainly asking to check if the pattern is correct in this case (tied with the earlier question)

Loading