Workplan task ID: P4-2
Files: tile config table
Description: gfx950 inherits gfx9 non-TrLoad tiles. Add separate arch tag to distinguish gfx950 from gfx9 for TrLoad eligibility.
Example to follow: P3-1 tile config table
CK Tile reference: include/ck_tile/ops/fmha/ (gfx950 path)
CK Tile feature check (required): Before implementing this task, verify the feature is present in the CK Tile source tree in BOTH locations:
- fmha_bwd example — especially codegen:
projects/composablekernel/example/ck_tile/01_fmha/codegen/ops/fmha_bwd.py (and supporting host files: fmha_bwd.hpp, fmha_bwd_runner.hpp, example_fmha_bwd.cpp).
- kernels:
projects/composablekernel/include/ck_tile/ops/fmha/{kernel,pipeline,block}/.
If the feature is missing or only partially wired in either location, raise it as a blocker before writing the rocm_ck wrapper — the wrapper cannot expose what CK Tile does not implement.
Tests: experimental/rocm_ck/tests/test_fmha_bwd_consteval.cpp
Blocked by: P4-1, P3-1
Source: plans/builder_kpack_fmha/fmha_bwd_feature_gap_workplan.md and progress.md (status as of 2026-05-04).
Workplan task ID: P4-2
Files: tile config table
Description: gfx950 inherits gfx9 non-TrLoad tiles. Add separate arch tag to distinguish gfx950 from gfx9 for TrLoad eligibility.
Example to follow: P3-1 tile config table
CK Tile reference: include/ck_tile/ops/fmha/ (gfx950 path)
CK Tile feature check (required): Before implementing this task, verify the feature is present in the CK Tile source tree in BOTH locations:
projects/composablekernel/example/ck_tile/01_fmha/codegen/ops/fmha_bwd.py(and supporting host files:fmha_bwd.hpp,fmha_bwd_runner.hpp,example_fmha_bwd.cpp).projects/composablekernel/include/ck_tile/ops/fmha/{kernel,pipeline,block}/.If the feature is missing or only partially wired in either location, raise it as a blocker before writing the
rocm_ckwrapper — the wrapper cannot expose what CK Tile does not implement.Tests: experimental/rocm_ck/tests/test_fmha_bwd_consteval.cpp
Blocked by: P4-1, P3-1
Source:
plans/builder_kpack_fmha/fmha_bwd_feature_gap_workplan.mdandprogress.md(status as of 2026-05-04).