Skip to content

perf(rocke): measured per-mode defaults for gfx950 GDN/KDA decode #14353

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
AviralGoelAMD merged 3 commits into from
Oct 29, 2026
Merged

perf(rocke): measured per-mode defaults for gfx950 GDN/KDA decode #14353

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
3e5da5c
perf(rocke): measured per-mode defaults for gfx950 GDN/KDA decode
AviralGoelAMD Oct 12, 2026
0b2d37a
fix(rocke): run auto's waves_per_eu only at the measured head dims
AviralGoelAMD Oct 28, 2026
436c6d0
test(rocke): label the tune test's stub timings as synthetic
AviralGoelAMD Oct 28, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 14 additions & 11 deletions dnn-providers/hip-kernel-provider/rocke/docs/gpu-attention-test-coverage.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,11 +20,14 @@ cases and three negative checks passed; the wider library GPU suite reported
84 passes and 134 skips.
Published artifact packaging was separately verified at run `36897764926`.
A source-declared shape is not necessarily selected, executed, or passed in CI.
One edit postdates that inventory: the `expected_spec_id` values of the
`test_kda_dispatch_band_launches_selected_kernel` rows use the tile-named KDA
spec ids (`kda_nw<N>_wtk<K>_bpv<B>`) that later replaced the work-band names;
the tests and shapes are unchanged. GPU tests added after the inventory are not
listed here.
Two edits postdate that inventory, both to the KDA decode dispatch rows. The
`expected_spec_id` values use the tile-named KDA spec ids
(`kda_nw<N>_wtk<K>_bpv<B>`) that replaced the work-band names. The work-keyed
tile table was later replaced by one static `auto` configuration per case: the
band test is now `test_kda_dispatch_launches_the_auto_kernel` and the tuned-tile
test `test_every_kda_auto_tile_is_correct`, with the `auto` tiles as their
shapes. Both also run with an `f32` state; those state-width rows are not
listed. Other GPU tests added after the inventory are not listed here.

## Reading this map

Expand Down Expand Up @@ -425,12 +428,12 @@ not a rocKE kernel execution; it remains here to expose that distinction.
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_kda_warp_tiled_matches_reference | gfx950 | 3 | 1 | 16 | 32 | 128 | 128 | bf16 | {"batch": 3}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_kda_warp_tiled_matches_reference | gfx950 | 16 | 1 | 16 | 32 | 128 | 128 | bf16 | {"batch": 16}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_kda_warp_tiled_matches_reference | gfx950 | 64 | 1 | 16 | 32 | 128 | 128 | bf16 | {"batch": 64}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_every_kda_tuned_tile_is_correct | gfx950 | 16 | 1 | 16 | 32 | 128 | 128 | bf16 | {"batch": 16, "tile": [4, 16, 4]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_every_kda_tuned_tile_is_correct | gfx950 | 16 | 1 | 16 | 32 | 128 | 128 | bf16 | {"batch": 16, "tile": [1, 16, 4]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_every_kda_tuned_tile_is_correct | gfx950 | 16 | 1 | 16 | 32 | 128 | 128 | bf16 | {"batch": 16, "tile": [2, 16, 1]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_kda_dispatch_band_launches_selected_kernel | gfx950 | 1 | 1 | 32 | 32 | 128 | 128 | bf16 | {"batch": 1, "expected_spec_id": "kda_nw4_wtk16_bpv4", "expected_tile": [4, 16, 4]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_kda_dispatch_band_launches_selected_kernel | gfx950 | 8 | 1 | 32 | 32 | 128 | 128 | bf16 | {"batch": 8, "expected_spec_id": "kda_nw1_wtk16_bpv4", "expected_tile": [1, 16, 4]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_kda_dispatch_band_launches_selected_kernel | gfx950 | 32 | 1 | 32 | 32 | 128 | 128 | bf16 | {"batch": 32, "expected_spec_id": "kda_nw2_wtk16_bpv1", "expected_tile": [2, 16, 1]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_every_kda_auto_tile_is_correct | gfx950 | 16 | 1 | 16 | 32 | 128 | 128 | bf16 | {"batch": 16, "tile": [1, 16, 8]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_every_kda_auto_tile_is_correct | gfx950 | 16 | 1 | 16 | 32 | 128 | 128 | bf16 | {"batch": 16, "tile": [1, 16, 16]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_every_kda_auto_tile_is_correct | gfx950 | 16 | 1 | 16 | 32 | 128 | 128 | bf16 | {"batch": 16, "tile": [4, 8, 1]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_kda_dispatch_launches_the_auto_kernel | gfx950 | 1 | 1 | 32 | 32 | 128 | 128 | bf16 | {"batch": 1, "expected_spec_id": "kda_nw1_wtk16_bpv8", "expected_tile": [1, 16, 8]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_kda_dispatch_launches_the_auto_kernel | gfx950 | 8 | 1 | 32 | 32 | 128 | 128 | bf16 | {"batch": 8, "expected_spec_id": "kda_nw1_wtk16_bpv8", "expected_tile": [1, 16, 8]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_kda_dispatch_launches_the_auto_kernel | gfx950 | 32 | 1 | 32 | 32 | 128 | 128 | bf16 | {"batch": 32, "expected_spec_id": "kda_nw1_wtk16_bpv8", "expected_tile": [1, 16, 8]}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_kda_mha_shape_matches_reference | gfx950 | 8 | 1 | 32 | 32 | 128 | 128 | bf16 | {}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_precomputed_decay_matches_reference | gfx950 | 1 | 1 | 16 | 32 | 128 | 128 | bf16 | {"batch": 1}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
| [test_kda_decode_gfx950_numeric.py](../library/tests/test_kda_decode_gfx950_numeric.py) :: test_precomputed_decay_matches_reference | gfx950 | 16 | 1 | 16 | 32 | 128 | 128 | bf16 | {"batch": 16}; output and updated state checked; state [pool,Hv,DV,DK] | Selected by library CTest file list |
Expand Down
Loading
Loading