[WebNN EP] Support GroupNormalization, GroupNorm and SkipGroupNorm

[Wanming Lin (Honry)]

Support GroupNormalization and com.microsoft GroupNorm/SkipGroupNorm. WebNN has no native group-norm ops, so each is lowered to:

reshape[N,G,(C/G)*spatial,1] -> instanceNormalization -> reshape back -> mul(gamma) -> add(beta)

This is the InstanceNormalization pattern that backends can re-fuse into a native group normalization. Handles NHWC/NCHW layout, optional SiLU activation, the SkipGroupNorm residual sum (+ optional S output), per-group scale/bias (opset 18), and any input rank >= 3.

[azure-pipelines]

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

[Wanming Lin (Honry)]

Dwayne Robinson (@fdwr), PTAL, thanks!

[Copilot]

🟡 Changes recommended

Input validation, mixed-precision support, and WebNN test coverage need correction.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Adds WebNN lowering for standard and Microsoft group-normalization operators using InstanceNormalization.

Changes:

• Registers GroupNormalization, GroupNorm, and SkipGroupNorm.
• Implements layout handling, affine parameters, residuals, optional SiLU, and sum output.
• Updates decomposition metadata and operator documentation.

File summaries

File	Description
op_builder_factory.h	Declares the builder factory.
op_builder_factory.cc	Registers supported operators.
map_info.h	Lists decomposed WebNN operations.
group_norm_op_builder.cc	Implements lowering and capability checks.
webnn-operators.md	Documents operator support.

Review details

Suppressed comments (2)

onnxruntime/core/providers/webnn/builders/impl/group_norm_op_builder.cc:316

• This only validates the skip tensor's rank, so shapes such as [4, 4] for an input [2, 3, 4, 8] are claimed as supported and then reinterpreted by the reshape as [2, 1, 1, 8], producing incorrect per-batch broadcasting. Require [N,C], the exact input shape, or a rank-matched shape with only the spatial dimensions equal to 1 before assigning the node to WebNN.

    // Skip is either (N, C) or the full input shape.
    if (skip_shape.size() != 2 && skip_shape.size() != input_shape.size()) {
      LOGS(logger, VERBOSE) << op_type << " skip must be 2D or match the input rank.";
      return false;

onnxruntime/core/providers/webnn/builders/impl/group_norm_op_builder.cc:344

• Requiring every input to have the same type rejects valid contrib GroupNorm/SkipGroupNorm nodes because their schemas intentionally separate T (X/skip/output) from M (gamma/beta). In particular, the repository's FP16 conversion keeps inputs 1 and 2 as float32 (python/tools/transformers/float16.py:147-149), so the common FP16 diffusion models targeted by these operators will still fall back. Support the T/M combination by inserting the required WebNN casts around the affine stage and include Cast in the decomposition capability checks.

  if (!AreDataTypesSame(op_type, input_types, logger)) {
    return false;
  }

• Files reviewed: 5/5 changed files
• Comments generated: 2
• Review effort level: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.