webgpu: Refactor SubgroupMatrixMatMulNBits to vendor-agnostic config …

[qjia7]

…+ add NVIDIA 16x16x16

Refactor subgroup matrix MatMulNBits support from vendor-specific (Apple/Intel) to a vendor-agnostic config-based approach. Any GPU reporting a matching subgroup matrix config from Dawn is now automatically supported.

Key changes:

• Replace vendor-specific config table with SupportedSubgroupMatrixConfig struct containing {componentType, resultComponentType, M, N, K, subgroupMinSize, subgroupMaxSize, needsPrepack}. No architecture or backendType required.
• Remove vendor_ member from SubgroupMatrixMatMulNBitsProgram. Shader selection is now driven by config dimensions (8x8x8, 8x16x16, 16x16x16).
• Remove vendor gate in matmul_nbits.cc call site.
• Rename shader templates: _apple -> _8x8x8, _intel -> _8x16x16.
• Add new 16x16x16 shader template for NVIDIA Blackwell (RTX 5080).
  • 4 subgroups x 32 lanes = 128 threads per workgroup
  • 64x64 tile with 16x16 subgroup matrices
  • Bounds-checked output via scratch buffer for partial M tiles
• Fix prepack shader OOB reads: add scalar fallback with zero-fill for partial blocks where M is not a multiple of kSgMatM.
• Prioritize larger configs (16x16x16 > 8x16x16 > 8x8x8) when multiple match.

Verified on NVIDIA RTX 5080 (Blackwell, Vulkan backend):

• Correctness: model-qa.py with phi4-graph-prune produces identical output to D3D12 baseline
• Prefill (phi4, l=1024):

phi4-graph-prune	D3D12 DP4A	Vulkan DP4A	Vulkan TC (16x16x16)	Vulkan TC (16x16x16_128)
Prefill (tps)	3,134	6,389	7,089	10,744

• NVIDIA reports ChromiumExperimentalSubgroupMatrix with F16/F16 16x16x16 config

Description

Motivation and Context

[Copilot]

Pull request overview

Refactors WebGPU SubgroupMatrix MatMulNBits support from vendor-gated (Apple/Intel) logic to a vendor-agnostic, adapter-reported subgroup-matrix-config matching approach, and adds a new 16x16x16 (128x128 tile) shader variant targeting NVIDIA Blackwell-class GPUs.

Changes:

• Introduces a vendor-agnostic SupportedSubgroupMatrixConfig table and matching logic driven by Dawn-reported subgroup matrix configs.
• Renames/rewrites subgroup-matrix WGSL templates into dimension-keyed variants (8x8x8, 8x16x16) and adds a new 16x16x16_128 shader.
• Updates prepack WGSL to avoid OOB reads via a scalar fallback for partial M blocks, and removes vendor gating at the MatMulNBits call site.

Reviewed changes

Copilot reviewed 5 out of 7 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
onnxruntime/contrib_ops/webgpu/quantization/subgroup_matrix_matmul_nbits_prepack.wgsl.template	Adds partial-block handling in prepack to avoid OOB reads and updates documentation about padding behavior.
onnxruntime/contrib_ops/webgpu/quantization/subgroup_matrix_matmul_nbits_8x8x8.wgsl.template	Adds the 8x8x8 subgroup-matrix matmul shader template (formerly vendor-specific).
onnxruntime/contrib_ops/webgpu/quantization/subgroup_matrix_matmul_nbits_8x16x16.wgsl.template	Adds the 8x16x16 subgroup-matrix matmul shader template (formerly Intel-specific).
onnxruntime/contrib_ops/webgpu/quantization/subgroup_matrix_matmul_nbits_16x16x16_128.wgsl.template	Adds the 16x16x16 subgroup-matrix matmul shader template using 128x128 tiles and edge-tile bounds-checked stores.
onnxruntime/contrib_ops/webgpu/quantization/subgroup_matrix_matmul_nbits.h	Removes vendor string plumbing from the program API.
onnxruntime/contrib_ops/webgpu/quantization/subgroup_matrix_matmul_nbits.cc	Replaces vendor-specific config matching with a vendor-agnostic config table; selects shaders by config dimensions; updates prepack padding logic.
onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc	Removes vendor gate so subgroup-matrix path is enabled for any adapter reporting a supported config.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[qjia7]

cc @jchen10