Skip to content

[TorchToLinalg] Fix the lowering of AtenIndexTensorHackedTwinOp #4280

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

[TorchToLinalg] Fix the lowering of AtenIndexTensorHackedTwinOp #4280

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
8 changes: 8 additions & 0 deletions lib/Conversion/TorchToLinalg/IndirectDataMovement.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -507,6 +507,14 @@ class ConvertAtenIndexSelectOp : public OpConversionPattern<AtenIndexSelectOp> {
static Value makeIndexValuePositive(OpBuilder &b, Location loc, Value index,
Value input, int64_t dim) {
Value cstZero = b.create<arith::ConstantOp>(loc, b.getI64IntegerAttr(0));
auto indexType = mlir::dyn_cast<IntegerType>(index.getType());
unsigned maxBitWidth = 64;
assert((indexType && indexType.isSignless()) &&
"The index operand must be a signless integer");
assert((indexType.getWidth() <= maxBitWidth) &&
"Maximum supported bitwidth of the index operand is 64");
if (indexType.getWidth() < maxBitWidth)
index = b.create<arith::ExtSIOp>(loc, b.getIntegerType(maxBitWidth), index);
Value isIndexNegative =
b.create<arith::CmpIOp>(loc, arith::CmpIPredicate::slt, index, cstZero);
Value inputShape = castIndexToInt64(b, loc, getDimOp(b, loc, input, dim));
Expand Down
18 changes: 18 additions & 0 deletions test/Conversion/TorchToLinalg/basic.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -459,3 +459,21 @@ func.func @torch.ops.aten.replication_pad3d$basic(%arg0: !torch.vtensor<[4,3,5],
%0 = torch.aten.replication_pad3d %arg0, %padding : !torch.vtensor<[4,3,5],f32>, !torch.list<int> -> !torch.vtensor<[7,7,6],f32>
return %0 : !torch.vtensor<[7,7,6],f32>
}

// -----

// This test verifies that the index argument is properly sign-extended,
// when torch.aten.index.Tensor_hacked_twin is lowered into a linalg.generic
// operation.
//
// CHECK-LABEL: func.func @torch.aten.index.Tensor_hacked_twin(
// CHECK: linalg.generic
// CHECK-NEXT: ^bb0(%[[IN:.*]]: i32, %[[OUT:.*]]: f32):
// CHECK-NEXT: %[[C0:.*]] = arith.constant 0 : i64
// CHECK-NEXT: %[[IN_SIGN_EXT:.*]] = arith.extsi %[[IN]] : i32 to i64
// CHECK-NEXT: arith.cmpi slt, %[[IN_SIGN_EXT]], %[[C0]] : i64
func.func @torch.aten.index.Tensor_hacked_twin(%arg0: !torch.vtensor<[1,1,8],si32>, %arg1: !torch.vtensor<[16], f32>) -> !torch.vtensor<[1,1,8],f32> {
%0 = torch.prim.ListConstruct %arg0 : (!torch.vtensor<[1,1,8],si32>) -> !torch.list<vtensor>
%1 = torch.aten.index.Tensor_hacked_twin %arg1, %0 : !torch.vtensor<[16],f32>, !torch.list<vtensor> -> !torch.vtensor<[1,1,8],f32>
return %1 : !torch.vtensor<[1,1,8],f32>
}