Grouped Conv Bwd Weight Direct Load

[bartekxk]

Proposed changes

Add Grouped Conv Bwd Weight Direct Load implementation and instances

Checklist

Please put an x into the boxes that apply. You can also fill these out after creating the PR. If you're not sure, please don't hesitate to ask.

• I have added tests relevant to the introduced functionality, and the unit tests are passing locally
• I have added the test to REGRESSION_TESTS list defined at the top of CMakeLists.txt in tests/CMakeLists.txt, IF the test takes more than 30 seconds to run.
• I have added inline documentation which enables the maintainers with understanding the motivation
• I have removed the stale documentation which is no longer relevant after this pull request
• (If this change is user-facing) I have added release notes which provide the end users with a brief summary of the improvement from this pull request
• I have run clang-format on all changed files
• Any dependent changes have been merged

Discussion

If this is a relatively large or complex change, feel free to start a discussion by explaining why you chose the solution you did and what alternatives you considered

[jakpiase]

I believe that BK0, BK1Number and such are proper, since it's B layout

[bartekxk]

You are right, thanks

[Copilot]

Pull request overview

This PR adds support for direct load implementation in grouped convolution backward weight operations, introducing a hardware-optimized memory transfer path for gfx950 devices. The implementation uses the ThreadGroupTensorSliceTransfer_DirectLoad mechanism with specific handling for F16 and BF16 data types.

Changes:

• Added direct load instances for F16 and BF16 grouped conv backward weight operations
• Introduced DirectLoad and LdsScalarLoad template parameters throughout the pipeline
• Implemented device-specific validation to restrict direct load to gfx950 architecture

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_direct_load.cpp	Instantiates F16 direct load device operations for grouped conv backward weight
device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_direct_load.cpp	Instantiates BF16 direct load device operations for grouped conv backward weight
CMakeLists.txt	Registers new F16 and BF16 direct load source files in build system
grouped_convolution_backward_weight_xdl.inc	Declares F16 and BF16 direct load instance functions
grouped_convolution_backward_weight.hpp	Integrates direct load instances into factory pattern
device_grouped_conv_bwd_weight_v3_xdl_instance.hpp	Defines F16 and BF16 direct load instance configurations with true flag
gridwise_gemm_xdl_cshuffle_conv_v3.hpp	Adds direct load conditional logic with specialized block descriptors and transfer handling
device_grouped_conv_bwd_weight_xdl_cshuffle_v3.hpp	Implements direct load support with gfx950 validation and group merging logic
thread_group_tensor_slice_transfer_direct_load.hpp	Removes destination vector dimension constraint for direct load
blockwise_gemm_pipeline_xdlops_v1.hpp	Adds LdsScalarLoad parameter to direct load pipeline
blockwise_gemm_pipeline_xdlops_selector.hpp	Adds LdsScalarLoad selection logic with validation
blockwise_gemm_pipeline_xdlops_base.hpp	Implements scalar load logic for LDS transfers when enabled

---

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

[aosewski]

What's the difference ? Both invocations seem to pass identical parameter set.

[bartekxk]

I need to disable this on other archs