[webgpu] Remove global_id and workgroup_id in gemm_utils.cc
#26662
Conversation
This patch replaces `global_id` and `workgroup_id` with `logical_global_id` and `logical_workgroup_id` which are computed from `workgroup_idx` and the dispatch workgroup sizes set in `ProgramBase::SetDispatchGroupSize()` because the dispatch workgroup sizes may be normalized in `ProgramManager::NormalizeDispatchGroupSize()`.
There was a problem hiding this comment.
Pull request overview
This PR replaces direct usage of global_id and workgroup_id WebGPU built-in variables with computed logical_global_id and logical_workgroup_id values in GEMM-related shader generation code. This change is necessary because the dispatch workgroup sizes set via ProgramBase::SetDispatchGroupSize() may be normalized by ProgramManager::NormalizeDispatchGroupSize() to fit within WebGPU device limits.
Key changes:
- Added
logical_dispatch_x/y/zuniform variables to multiple program classes (MatMulProgram, GemmProgram, Conv2dMMProgram) - Implemented logical workgroup ID computation from
workgroup_idxusing the logical dispatch dimensions - Updated all references from
global_idandworkgroup_idto their logical counterparts in shader code generation - Removed outdated TODO comment in matmul.cc that referenced avoiding
global_id/workgroup_id
Reviewed changes
Copilot reviewed 7 out of 7 changed files in this pull request and generated 2 comments.
Show a summary per file
| File | Description |
|---|---|
| onnxruntime/core/providers/webgpu/nn/conv2d_mm.h | Added logical_dispatch_x/y/z uniform variable declarations to Conv2dMMProgram |
| onnxruntime/core/providers/webgpu/nn/conv2d_mm.cc | Pass dispatch dimensions as uniform variables to Conv2dMMProgram |
| onnxruntime/core/providers/webgpu/math/matmul_packed.h | Added logical_dispatch_x/y/z uniform variable declarations to MatMulProgram |
| onnxruntime/core/providers/webgpu/math/matmul.cc | Removed outdated TODO comment and pass dispatch dimensions as uniform variables to MatMulProgram |
| onnxruntime/core/providers/webgpu/math/gemm_utils.cc | Implemented logical workgroup/global ID computation and replaced all usage of global_id/workgroup_id with logical equivalents in shader generation |
| onnxruntime/core/providers/webgpu/math/gemm_packed.h | Added logical_dispatch_x/y/z uniform variable declarations to GemmProgram |
| onnxruntime/core/providers/webgpu/math/gemm_packed.cc | Pass dispatch dimensions as uniform variables to GemmProgram |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 7 out of 7 changed files in this pull request and generated 2 comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 7 out of 7 changed files in this pull request and generated no new comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
|
Hi reviewers, The failed test is PTAL, thanks! |
3 participants
Description
This patch replaces
global_idandworkgroup_idwithlogical_global_idandlogical_workgroup_idwhich are computed fromworkgroup_idxand the dispatch workgroup sizes set inProgramBase::SetDispatchGroupSize().Motivation and Context
We shouldn't use
global_idorworkgroup_iddirectly because the dispatch workgroup sizes may be normalized inProgramManager::NormalizeDispatchGroupSize().