🔥 Jadwal & Hasil: microsoft/onnxruntime/pull/

Yang Gu (gyagp) · 2026-09-17T06:13:40Z

Description

Update Dawn to v20260828.013844 and Agility SDK to 1.721.3-preview.
Support subgroup-matrix kernels on adapters that expose a variable subgroup-size range.
Require subgroup-size-control for variable-size adapters and request the selected configuration's subgroup size for prepack and multiplication shaders.
Add coverage for valid and invalid subgroup-size configurations.

Motivation and Context

AMD RDNA 3+ supports the FP16 16x16x16 subgroup-matrix configuration with a required subgroup size of 32, while the adapter exposes a variable subgroup-size range. The previous ORT check required the adapter's minimum and maximum subgroup sizes to exactly match the kernel size, so the optimized path was not selected.

The Dawn update is required in addition to the ORT changes. The previous Dawn revision:

depended solely on the global D3D12 LinAlg tier to expose subgroup matrices, although preview drivers can report valid operation-specific configurations when that tier is absent;
did not reject reported type combinations whose D3D12 operation SupportFlags lack the supported bit; and
did not forward enumerated subgroup-matrix configurations to Tint's HLSL lowering.

The newer Dawn revision evaluates the authoritative D3D12 capability data and supplies the configurations to Tint, allowing ORT to expose and compile the WebGPU subgroup-matrix path correctly.

Related Dawn issue: https://issues.chromium.org/issues/549226780

Testing

Windows D3D12 on AMD Radeon RX 9070 XT:

AMD preview driver 26.10.07.02 (32.0.31007.2064)
Agility SDK 721
SubgroupMatrixConfigTest.RequiredSubgroupSizeCompatibility: passed
MatMulNBits.Float16_4b_Accuracy4: passed
Confirmed FP16 16x16x16 shaders using subgroupMatrixLoad, subgroupMatrixMultiplyAccumulate, and subgroupMatrixStore.

azure-pipelines · 2026-09-17T06:13:49Z

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

Allow subgroup matrix kernels to select their required subgroup size on adapters that expose a variable subgroup-size range. Require subgroup-size control in that case and request the selected configuration's size for prepack and multiplication shaders. Update Dawn to v20260828.013844 and Agility SDK to 1.721.3-preview. The newer Dawn is required because the previous revision depended solely on the global D3D12 LinAlg tier, did not reject type combinations whose operation SupportFlags were unsupported, and did not forward the enumerated matrix configurations to Tint's HLSL lowering. AMD preview drivers can report valid operation-specific configurations even when the global tier is absent, so those fixes are needed to expose and compile the WebGPU subgroup-matrix feature correctly. Tested on Windows D3D12 with an AMD Radeon RX 9070 XT, AMD driver 26.10.07.02 (32.0.31007.2064), and Agility SDK 721. MatMulNBits.Float16_4b_Accuracy4 passed while generating FP16 16x16x16 subgroup-matrix shaders. SubgroupMatrixConfigTest.RequiredSubgroupSizeCompatibility also passed.

Yang Gu (gyagp) · 2026-09-17T06:21:32Z

@microsoft-github-policy-service agree company="Microsoft"

azure-pipelines · 2026-09-17T06:23:20Z

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

Yang Gu (gyagp) · 2026-09-17T06:23:50Z

/azp

azure-pipelines · 2026-09-17T06:23:57Z

Supported commands help: Get descriptions, examples and documentation about supported commands Example: help "command_name" list: List all pipelines for this repository using a comment. Example: "list" run: Run all pipelines or specific pipelines for this repository using a comment. Use this command by itself to trigger all related pipelines, or specify specific pipelines to run. Example: "run" or "run pipeline_name, pipeline_name, pipeline_name" where: Report back the Azure DevOps orgs that are related to this repository and org Example: "where" See additional documentation.

Copilot

🔵 Needs a closer look

The vendor-specific GPU behavior and foundational Dawn dependency update require final human and hardware-backed validation.

Pull request overview

Enables WebGPU subgroup-matrix kernels on AMD GPUs with variable subgroup sizes.

Changes:

Selects and requests each kernel configuration’s required subgroup size.
Adds subgroup-size compatibility tests.
Updates Dawn and Direct3D Agility SDK dependencies.

File summaries

File	Description
`onnxruntime/test/providers/cpu/math/matmul_test.cc`	Tests fixed, variable, and invalid subgroup-size ranges.
`onnxruntime/core/providers/webgpu/math/subgroup_matrix_matmul.cc`	Configures MatMul workgroups with the selected subgroup size.
`onnxruntime/core/providers/webgpu/math/subgroup_matrix_gemm.cc`	Configures Gemm workgroups with the selected subgroup size.
`onnxruntime/core/providers/webgpu/math/subgroup_matrix_config.h`	Defines required subgroup sizes and compatibility logic.
`onnxruntime/core/providers/webgpu/math/subgroup_matrix_config.cc`	Validates adapter ranges and size-control support.
`onnxruntime/contrib_ops/webgpu/quantization/subgroup_matrix_matmul_nbits.cc`	Pins prepack and multiplication shaders to the selected size.
`cmake/patches/dawn/dawn_binskim.patch`	Rebases the warning-policy patch onto the new Dawn revision.
`cmake/deps.txt`	Updates Dawn and Agility SDK pins.

Review details

Files reviewed: 8/8 changed files
Comments generated: 0
Review effort level: Balanced

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

Yang Gu (gyagp) · 2026-09-17T06:26:53Z

Jiajia Qin (@qjia7), please help to review this PR to support subgroup matrix on AMD.
Jie Chen (@jchen10), please help to check if this will break Intel.

Yang Gu (gyagp) · 2026-09-17T06:50:07Z

/run bots

Jiajia Qin (qjia7) · 2026-09-17T07:01:55Z

The /run bots request was received from Yang Gu (@gyagp), but no CI run was started. No pending action-required workflow runs matched the current PR head.

Jie Chen (jchen10) · 2026-09-17T08:38:48Z

Yang Gu (@gyagp) Just checked, no regression.

Yang Gu (gyagp) · 2026-09-17T08:53:51Z

Yang Gu (Yang Gu (@gyagp)) Just checked, no regression.

Thanks, Jie Chen (@jchen10)!

### Description - Update Dawn to `v20260828.013844` and Agility SDK to `1.721.3-preview`. - Support subgroup-matrix kernels on adapters that expose a variable subgroup-size range. - Require `subgroup-size-control` for variable-size adapters and request the selected configuration's subgroup size for prepack and multiplication shaders. - Add coverage for valid and invalid subgroup-size configurations. ### Motivation and Context AMD RDNA 3+ supports the FP16 16x16x16 subgroup-matrix configuration with a required subgroup size of 32, while the adapter exposes a variable subgroup-size range. The previous ORT check required the adapter's minimum and maximum subgroup sizes to exactly match the kernel size, so the optimized path was not selected. The Dawn update is required in addition to the ORT changes. The previous Dawn revision: - depended solely on the global D3D12 LinAlg tier to expose subgroup matrices, although preview drivers can report valid operation-specific configurations when that tier is absent; - did not reject reported type combinations whose D3D12 operation `SupportFlags` lack the supported bit; and - did not forward enumerated subgroup-matrix configurations to Tint's HLSL lowering. The newer Dawn revision evaluates the authoritative D3D12 capability data and supplies the configurations to Tint, allowing ORT to expose and compile the WebGPU subgroup-matrix path correctly. Related Dawn issue: https://issues.chromium.org/issues/549226780 ### Testing Windows D3D12 on AMD Radeon RX 9070 XT: - AMD preview driver `26.10.07.02` (`32.0.31007.2064`) - Agility SDK 721 - `SubgroupMatrixConfigTest.RequiredSubgroupSizeCompatibility`: passed - `MatMulNBits.Float16_4b_Accuracy4`: passed - Confirmed FP16 16x16x16 shaders using `subgroupMatrixLoad`, `subgroupMatrixMultiplyAccumulate`, and `subgroupMatrixStore`.

### Description - Replace the transitive `compute_context.h` dependency in `subgroup_matrix_config.h` with the Dawn type header and a forward declaration of `ComputeContextBase`. - Include `compute_context.h` in the implementation file where the complete type is required. ### Motivation and Context #32667 added `subgroup_matrix_config.h` to `matmul_test.cc`. In WebGPU plugin builds, the old transitive include chain reached `webgpu_execution_provider.h` without the EP adapter precompiled header, so `onnxruntime::ep::adapter::Logger` was undeclared. This broke `webgpu_plugin_build_x64_RelWithDebInfo` both on #32667 and on `main`. Keeping the implementation dependency out of the public configuration header fixes the plugin test build and avoids exposing the execution-provider header to callers that only need subgroup matrix configuration types. Failing main run: https://github.com/microsoft/onnxruntime/actions/runs/35204008069 ### Testing - `git diff --check` Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Yang Gu (gyagp) force-pushed the webgpu-amd-subgroup-matrix branch from a0688c7 to 40f3518 Compare September 17, 2026 06:14

Yang Gu (gyagp) marked this pull request as ready for review September 17, 2026 06:23

Copilot AI balanced review requested due to automatic review settings September 17, 2026 06:23

Copilot started reviewing on behalf of Yang Gu (gyagp) September 17, 2026 06:23

View session

Copilot AI reviewed Sep 17, 2026

View reviewed changes

Jiajia Qin (qjia7) approved these changes Sep 17, 2026

View reviewed changes

Jiajia Qin (qjia7) enabled auto-merge (squash) September 17, 2026 08:55

Jiajia Qin (qjia7) merged commit bff4fcf into microsoft:main Sep 17, 2026
88 of 90 checks passed

This was referenced Sep 17, 2026

Fix WebGPU subgroup config header dependency #32672

Merged

Load CUDA external data through pinned buffers #32437

Merged

Rithik Sharma (SharmaRithik) mentioned this pull request Sep 29, 2026

[WebGPU] Support 16x16x16 subgroup matrix MatMul #32937

Merged

BrewTestBot mentioned this pull request Oct 9, 2026

onnxruntime 1.31.0 Homebrew/homebrew-core#316686

Merged

Navigation Menu

Repository navigation

[WebGPU] Enable subgroup matrix kernels on AMD GPUs - #32667

[WebGPU] Enable subgroup matrix kernels on AMD GPUs#32667
Jiajia Qin (qjia7) merged 1 commit into
microsoft:mainfrom
gyagp:webgpu-amd-subgroup-matrix

Yang Gu (gyagp) commented Sep 17, 2026

Uh oh!

azure-pipelines Bot commented Sep 17, 2026

Uh oh!

Yang Gu (gyagp) commented Sep 17, 2026 •

edited

Loading

Uh oh!

azure-pipelines Bot commented Sep 17, 2026

Uh oh!

Yang Gu (gyagp) commented Sep 17, 2026

Uh oh!

azure-pipelines Bot commented Sep 17, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Yang Gu (gyagp) commented Sep 17, 2026

Uh oh!

Yang Gu (gyagp) commented Sep 17, 2026

Uh oh!

Jiajia Qin (qjia7) commented Sep 17, 2026

Uh oh!

Jie Chen (jchen10) commented Sep 17, 2026

Uh oh!

Yang Gu (gyagp) commented Sep 17, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

Yang Gu (gyagp) commented Sep 17, 2026

Description

Motivation and Context

Testing

Uh oh!

azure-pipelines Bot commented Sep 17, 2026

Uh oh!

Yang Gu (gyagp) commented Sep 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

azure-pipelines Bot commented Sep 17, 2026

Uh oh!

Yang Gu (gyagp) commented Sep 17, 2026

Uh oh!

azure-pipelines Bot commented Sep 17, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🔵 Needs a closer look

Uh oh!

Yang Gu (gyagp) commented Sep 17, 2026

Uh oh!

Yang Gu (gyagp) commented Sep 17, 2026

Uh oh!

Jiajia Qin (qjia7) commented Sep 17, 2026

Uh oh!

Jie Chen (jchen10) commented Sep 17, 2026

Uh oh!

Yang Gu (gyagp) commented Sep 17, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Yang Gu (gyagp) commented Sep 17, 2026 •

edited

Loading