Skip to content

DotGeneralReshape: keep the dot's own result element type #2986

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鈥檒l occasionally send you account related emails.

Already on GitHub? Sign in to your account

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

DotGeneralReshape: keep the dot's own result element type #2986

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
9 changes: 5 additions & 4 deletions src/enzyme_ad/jax/Passes/EnzymeHLOOpt.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28940,12 +28940,13 @@ struct DotGeneralReshape final
newShape.push_back(dim);
}

// The result element type may differ from the operands' (e.g. a bf16 x bf16
// -> f32 accumulating dot), so take it from the op being replaced: the
// trailing reshape has to be element-type preserving.
auto newDotGeneral = stablehlo::DotGeneralOp::create(
rewriter, op.getLoc(),
RankedTensorType::get(
newShape,
cast<RankedTensorType>(newLhs.getType()).getElementType()),
newLhs, newRhs,
RankedTensorType::get(newShape, op.getType().getElementType()), newLhs,
newRhs,
stablehlo::DotDimensionNumbersAttr::get(
rewriter.getContext(), adjustedLhsBatchingDims,
adjustedRhsBatchingDims, adjustedLhsContractingDims,
Expand Down
26 changes: 26 additions & 0 deletions test/lit_tests/dotgeneralreshape_mixed_element_types.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,26 @@
// RUN: enzymexlamlir-opt --transform-interpreter --enzyme-hlo-remove-transform %s | FileCheck %s

// A dot_general whose result element type differs from its operands' (a bf16 x
// bf16 -> f32 accumulating dot, what JAX emits with preferred_element_type=f32)
// must keep that result type when the feeding reshape is hoisted past it,
// otherwise the trailing reshape changes the element type and fails to verify.
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg0: !transform.any_op) {
%0 = transform.structured.match ops{["func.func"]} in %arg0 : (!transform.any_op) -> !transform.any_op
transform.apply_patterns to %0 {
transform.apply_patterns.enzyme_hlo.dot_general_reshape
} : !transform.any_op
transform.yield
}
func.func @mixed(%arg0: tensor<4x8xbf16>, %arg1: tensor<8x16xbf16>) -> tensor<1x4x16xf32> {
%0 = stablehlo.reshape %arg0 : (tensor<4x8xbf16>) -> tensor<1x4x8xbf16>
%1 = stablehlo.dot_general %0, %arg1, contracting_dims = [2] x [0] : (tensor<1x4x8xbf16>, tensor<8x16xbf16>) -> tensor<1x4x16xf32>
return %1 : tensor<1x4x16xf32>
}
}

// CHECK: func.func @mixed(%arg0: tensor<4x8xbf16>, %arg1: tensor<8x16xbf16>) -> tensor<1x4x16xf32> {
// CHECK-NEXT: %[[DOT:.+]] = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0]{{.*}} : (tensor<4x8xbf16>, tensor<8x16xbf16>) -> tensor<4x16xf32>
// CHECK-NEXT: %[[RES:.+]] = stablehlo.reshape %[[DOT]] : (tensor<4x16xf32>) -> tensor<1x4x16xf32>
// CHECK-NEXT: return %[[RES]] : tensor<1x4x16xf32>
// CHECK-NEXT: }