[Yash Dayani (YashDayani)]

Description

ConvAddFusion folds a constant per-channel Add into the Conv bias, but it only looked for that constant at Add input 1. As a result, Add(bias, Conv(X, W)) was left unfused while Add(Conv(X, W), bias) was fused. Add is commutative, so both forms should fuse.

• onnxruntime/core/optimizer/conv_add_fusion.cc: a new helper, GetAddBiasInputIndex, returns the Add input that is not the Conv output. SatisfyCondition and Apply use it instead of a hard-coded index 1.
  • SatisfyCondition already requires the Add to have exactly one input edge, and that edge comes from the Conv. So exactly one Add input is the Conv output, and the other must be a constant. Add(conv_out, conv_out) has two input edges and is still rejected.
  • The existing rank and shape checks limit the bias to [1, C, 1, 1] or [C, 1, 1], so the operand order does not change the output shape or values.
  • FinalizeNodeFusion does not depend on which Add input the Conv output uses.
• onnxruntime/test/optimizer/graph_transform_test.cc:
  • FuseConvAddWithBiasOnEitherAddInput checks both operand orders, each with and without a Conv bias. It asserts that the Add is fused away and compares outputs against the unoptimized graph. Without this change, the two bias-first cases fail.
  • NegativeFuseConvAddWithNonConstantOtherInput checks that Add(Y, Conv(X, W)), where Y is a graph input, and Add(conv_out, conv_out) are still not fused.

Tested on macOS arm64 (Release, CPU): all ConvAdd tests pass, as do 934 tests matching GraphTransformation*:NchwcOptimizer*:QDQTransformer*:TransposeOptimizer*:*Fusion*:*Optimizer*. The changed files are formatted with the repo's pinned clang-format.

Motivation and Context

Fixes #33818. Models whose exporter emits the bias as the first Add operand kept a separate Add after every such Conv, at all optimization levels.

ConvMulFusion (conv_mul_fusion.cc) has the same limitation for Mul(scale, Conv(X, W)). I left it out of this PR to keep the change focused, and can follow up separately.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

[Yash Dayani (YashDayani)]

@microsoft-github-policy-service agree