Skip to content

[P3-2] Remove dqdkdv_dev.hpp:105-107 d128 static_assert; dispatch via if constexpr - #7638

Merged
aosewski merged 2 commits into
users/chris-tsiaousis-hpc/ck/rocm-ck-p3-1-tile-configfrom
users/chris-tsiaousis-hpc/ck/rocm-ck-p3-2-remove-static-assert
Jun 1, 2026
Merged

aosewski merged 2 commits into
users/chris-tsiaousis-hpc/ck/rocm-ck-p3-1-tile-configfrom
users/chris-tsiaousis-hpc/ck/rocm-ck-p3-2-remove-static-assert

Conversation

@chris-tsiaousis-hpc

Copy link
Copy Markdown
Contributor

Closes #7560

Note: Part of it was done in
#7506

@chris-tsiaousis-hpc chris-tsiaousis-hpc changed the title Users/chris tsiaousis hpc/ck/rocm ck p3 2 remove static assert [P3-2] Remove dqdkdv_dev.hpp:105-107 d128 static_assert; dispatch via if constexpr May 22, 2026
@chris-tsiaousis-hpc
chris-tsiaousis-hpc marked this pull request as ready for review May 22, 2026 15:37
@chris-tsiaousis-hpc
chris-tsiaousis-hpc requested a review from a team as a code owner May 22, 2026 15:37
@chris-tsiaousis-hpc
chris-tsiaousis-hpc force-pushed the users/chris-tsiaousis-hpc/ck/rocm-ck-p3-2-remove-static-assert branch from a281e20 to e6e0f78 Compare May 28, 2026 07:53
@chris-tsiaousis-hpc
chris-tsiaousis-hpc changed the base branch from users/shumway/ck/exp-kpack to users/chris-tsiaousis-hpc/ck/rocm-ck-p3-1-tile-config May 28, 2026 08:00
@chris-tsiaousis-hpc
chris-tsiaousis-hpc force-pushed the users/chris-tsiaousis-hpc/ck/rocm-ck-p3-2-remove-static-assert branch from e6e0f78 to 6123540 Compare May 28, 2026 08:01
chris-tsiaousis-hpc and others added 2 commits June 1, 2026 07:38
Signed-off-by: Chris Tsiaousis <chris.tsiaousis@streamhpc.com>
Address review findings on the P3-2 frozen-baseline / consteval canary
layer. Comment-and-test-only; no production logic changes.

- Fix NOTE citing non-existent GFX9_FP16_DQDKDV_TILES (correct symbol
  is GFX9_FP16_DQDKDV_BASE_TILES).
- Clarify that the OGradDotO/ConvertDQ "multi-hdim" tests assert
  hdim-invariant constants, not per-hdim geometry coverage.
- Reword NOTEs/canary comments that over-claimed: baselines are
  compile-time spec-factory checks (not runtime numerical validation),
  and the canary guards block_n0 only (not full tile geometry).
- Add asymmetric-hdim DqDkDv coverage (D64Q_D256V, D256Q_D64V) to
  exercise the max(hdim_q, hdim_v) lookup; (64,256) is discriminating.
- Add the missing d128 block_n0 static_assert canary so the block now
  covers all five hdims.

Verified: standalone host compile passes; compat suite 65/65 PASS;
clang-format clean.

Signed-off-by: Adam Osewski <Adam.Osewski@amd.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@aosewski
aosewski force-pushed the users/chris-tsiaousis-hpc/ck/rocm-ck-p3-2-remove-static-assert branch from 0f1f86f to 9c9391d Compare June 1, 2026 12:00
@aosewski
aosewski merged commit 47c39a9 into users/chris-tsiaousis-hpc/ck/rocm-ck-p3-1-tile-config Jun 1, 2026
2 of 5 checks passed
@aosewski
aosewski deleted the users/chris-tsiaousis-hpc/ck/rocm-ck-p3-2-remove-static-assert branch June 1, 2026 12:01
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[P3-2] Remove dqdkdv_dev.hpp:105-107 d128 static_assert; dispatch via if constexpr

2 participants