[Relax] Add FRelaxInferLayout for scatter_nd operator

[guan404ming]

Why

The scatter_nd operator was missing FRelaxInferLayout attribute, which is needed for proper layout transformation during model optimization.

How

• Added InferLayoutScatterND function that uses data tensor's layout for output since scatter_nd maintains input shape
• Registered FRelaxInferLayout attribute

[gemini-code-assist]

Summary of Changes

Hello @guan404ming, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request addresses a critical missing component for the scatter_nd operator within the Relax framework by adding its FRelaxInferLayout attribute. This enhancement ensures that scatter_nd can correctly participate in layout transformations during model optimization, which is essential for maintaining data consistency and achieving optimal performance. The change includes the implementation of the layout inference logic and a new test to verify its functionality.

Highlights

• Layout Inference Function: Implemented the InferLayoutScatterND function, which determines the output layout for the scatter_nd operator by adopting the layout of its primary data input tensor.
• Attribute Registration: Registered the FRelaxInferLayout attribute for the relax.scatter_nd operator, linking it to the newly created InferLayoutScatterND function to enable proper layout transformation.
• New Test Case: Added a new test case, test_conv2d_scatter_nd, to test_transform_convert_layout.py to validate the correct layout transformation behavior of scatter_nd when used in conjunction with a conv2d operation.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request adds the FRelaxInferLayout attribute for the scatter_nd operator, which is a good addition for enabling layout transformations. The implementation is mostly correct, but I've found a potential issue in the layout inference logic for the updates tensor. The current logic doesn't handle the layout of updates, which could lead to compilation errors in certain scenarios. I've provided a suggestion to make the implementation more robust by ensuring layout consistency between the data and updates tensors.

[gemini-code-assist]

The current implementation of InferLayoutScatterND doesn't enforce any layout on the updates tensor. This can lead to shape mismatches after layout conversion if updates has a layout that is incompatible with the new layout of data, as the updates tensor's shape is dependent on the data tensor's shape. For instance, if data is transformed to NHWC, but updates remains in NCHW, InferStructInfo will likely fail.

I suggest a more robust implementation that enforces the same layout for data and updates when they have the same rank. For more complex cases, like different ranks or when data has a sub-indexed layout, it safely falls back to the initial layout for both tensors. This will prevent potential compilation errors.

InferLayoutOutput InferLayoutScatterND(
    const Call& call, const ffi::Map<ffi::String, ffi::Array<ffi::String>>& desired_layouts,
    const VarLayoutMap& var_layout_map) {
  ICHECK(NoDesiredLayout(call, desired_layouts));

  LayoutDecision data_layout = GetLayoutDecision(var_layout_map, call->args[0]);
  LayoutDecision indices_layout = GetLayoutDecision(var_layout_map, call->args[1]);
  LayoutDecision updates_layout = GetLayoutDecision(var_layout_map, call->args[2]);

  const auto* data_sinfo = GetStructInfoAs<TensorStructInfoNode>(call->args[0]);
  ICHECK(data_sinfo != nullptr) << "Invalid Call";
  const auto* updates_sinfo = GetStructInfoAs<TensorStructInfoNode>(call->args[2]);
  ICHECK(updates_sinfo != nullptr) << "Invalid Call";

  bool same_rank = !data_sinfo->IsUnknownNdim() && !updates_sinfo->IsUnknownNdim() &&
                   data_sinfo->ndim == updates_sinfo->ndim;

  if (data_layout->layout.ndim() != data_layout->layout.ndim_primal() || !same_rank) {
    // Fallback to initial layout if data has sub-indexed layout or ranks are different.
    // The case of different ranks is complex to handle, so we fallback for safety.
    ICHECK(!data_sinfo->IsUnknownNdim()) << "Only support static ndim for now";
    data_layout = LayoutDecision(InitialLayout(data_sinfo->ndim));
    if (!updates_sinfo->IsUnknownNdim()) {
      updates_layout = LayoutDecision(InitialLayout(updates_sinfo->ndim));
    } else {
      updates_layout = LayoutDecision::InitUnknownDim();
    }
  } else {
    // Same rank, not sub-indexed. Enforce same layout for data and updates.
    updates_layout = data_layout;
  }

  return InferLayoutOutput({data_layout, indices_layout, updates_layout}, {data_layout},
                           Attrs(call->attrs));
}

[guan404ming]

cc @tlopex @mshr-h

[guan404ming]

Thanks!