Skip to content

refactor(hipblaslt): convert gfx950 F4 library logic to dict format - #10390

Open
dsaffars wants to merge 1 commit into
users/dsaffars/Lib_conv_gfx950_miscfrom
users/dsaffars/Lib_conv_gfx950_F4
Open

refactor(hipblaslt): convert gfx950 F4 library logic to dict format#10390
dsaffars wants to merge 1 commit into
users/dsaffars/Lib_conv_gfx950_miscfrom
users/dsaffars/Lib_conv_gfx950_F4

Conversation

@dsaffars

@dsaffars dsaffars commented Aug 5, 2026

Copy link
Copy Markdown
Contributor

JIRA ID: AIHPBLAS-4216

Convert 9 gfx950 F4 library logic YAML files from list to dict format.

Motivation

TensileLite is migrating gfx950 library logic YAML from the legacy list format to the canonical dict format. The dict format uses explicit root keys (ArchitectureName, CUCount, ProblemType, Solutions, etc.) and is the format TensileCreateLibrary now emits by default.
The F4 datatype family continues the stacked gfx950 library logic conversion. Migrating these files improves maintainability, aligns shipped logic with the current serialization contract, and reduces YAML size without changing kernel selection behavior.

Technical Details

This PR converts 9 gfx950 F4 library logic YAML files under:
projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/gfx950/Origami/
projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/gfx950/Origami/Origami_nta4/
projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/gfx950/Origami/Origami_ntb4/
Format changes (no semantic changes intended):

  • Root metadata moves from positional list entries to named keys (MinimumRequiredVersion, ScheduleName, ArchitectureName, CUCount, DeviceNames, ProblemType, …).
  • ProblemType fields are nested under a ProblemType: mapping instead of inline list items.
  • Solution entries move from list-of-flow-mappings to a Solutions: dict keyed by solution name.
  • Boolean literals are normalized (True/Falsetrue/false).
    Scope summary:
Metric Before After
Files 9 9
Total lines 150,678 112,668

This PR is YAML-only; no runtime, generator, or test-code changes. It builds on users/dsaffars/Lib_conv_gfx950_misc in the stacked PR series.

Test Plan

For each of the 9 converted files, pre-conversion list-format YAML is compared against the new dict-format YAML using a script (same parser path as TensileCreateLibrary):

Test Result

For all 9/9 F4 files parsed solutions, problem types, and matching tables are semantically identical between list and dict formats.
File integrity: No files added or removed; only format conversion within the existing 9 F4 logic files (Origami, Origami/Origami_nta4, Origami/Origami_ntb4).
Line count: Total YAML size reduced from 150,678 to 112,668 lines (~25% reduction) with no intended change to kernel selection behavior.

Convert 9 gfx950 F4 library logic YAML files from list
to dict format.

# Batch: F4
# Per-file lines (before -> after):
  gfx950/Origami/Origami_nta4/gfx950_Cijk_Alik_Bljk_F4BS_MXA32_MXB32_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml: 16742 -> 12536
  gfx950/Origami/Origami_nta4/gfx950_Cijk_Alik_Bljk_F4HS_MXA32_MXB32_BH_BiasSH_HAS_SAB_SAV_UserArgs.yaml: 16742 -> 12536
  gfx950/Origami/Origami_nta4/gfx950_Cijk_Alik_Bljk_F4SS_MXA32_MXB32_BH_BiasS_HAS_SAB_SAV_UserArgs.yaml: 16742 -> 12536
  gfx950/Origami/Origami_ntb4/gfx950_Cijk_Alik_Bljk_F4BS_MXA32_MXB32_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml: 16742 -> 12536
  gfx950/Origami/Origami_ntb4/gfx950_Cijk_Alik_Bljk_F4HS_MXA32_MXB32_BH_BiasSH_HAS_SAB_SAV_UserArgs.yaml: 16742 -> 12536
  gfx950/Origami/Origami_ntb4/gfx950_Cijk_Alik_Bljk_F4SS_MXA32_MXB32_BH_BiasS_HAS_SAB_SAV_UserArgs.yaml: 16742 -> 12536
  gfx950/Origami/gfx950_Cijk_Alik_Bljk_F4BS_MXA32_MXB32_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml: 16742 -> 12484
  gfx950/Origami/gfx950_Cijk_Alik_Bljk_F4HS_MXA32_MXB32_BH_BiasSH_HAS_SAB_SAV_UserArgs.yaml: 16742 -> 12484
  gfx950/Origami/gfx950_Cijk_Alik_Bljk_F4SS_MXA32_MXB32_BH_BiasS_HAS_SAB_SAV_UserArgs.yaml: 16742 -> 12484
Total: 150678 -> 112668 lines across 9 files.
@dsaffars
dsaffars requested a review from a team as a code owner August 5, 2026 09:17
@therock-pr-bot

therock-pr-bot Bot commented Aug 5, 2026

Copy link
Copy Markdown

✅ All Checks Passed — Ready for Review

Check Status Details
📝 PR Description ✅ Pass
Forbidden Files ✅ Pass
🧪 Unit Test ✅ Pass PR does not contain code files — Unit Test auto-passed
🔎 pre-commit ✅ Pass
🚫 Draft PR 🔜 To Be Enabled
🚩 Feature Flag 🔜 To Be Enabled
📊 Code Coverage 🔜 To Be Enabled
🤖 therock-pr-bot ✅ Pass

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

@therock-pr-bot

therock-pr-bot Bot commented Aug 5, 2026

Copy link
Copy Markdown

🎉 All checks passed! This PR is ready for review.

@AmosLewis

Copy link
Copy Markdown
Contributor

Heads-up from the rocm-libraries gardener rotation: #10016
(perf(hipblaslt): Updating Origami libraries for MXFP4 TN for gfx950) has just landed on develop
as 934fa79, so this PR is unblocked and ready to rebase.

The overlap is exact — #10016 updated 3 of the 9 files this PR touches:

  • .../gfx950/Origami/gfx950_Cijk_Alik_Bljk_F4BS_MXA32_MXB32_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml
  • .../gfx950/Origami/Origami_nta4/gfx950_Cijk_Alik_Bljk_F4BS_MXA32_MXB32_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml
  • .../gfx950/Origami/Origami_ntb4/gfx950_Cijk_Alik_Bljk_F4BS_MXA32_MXB32_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml

#10016 added 23 new MTs to those three F4BS files, so when you rebase please make sure the
dict-format conversion picks up the new entries rather than reverting them — that is the one place
where a clean textual rebase could silently drop the tuning update.

One thing to be aware of before you push: GitHub Actions is in an active major outage right now
(Actions: major_outage, see https://www.githubstatus.com), so CI on a fresh push will likely fail
in Getting action download info or sit queued for a long time. It may be worth waiting for the
incident to clear before rebasing, otherwise you will be triaging platform noise rather than real
signal.

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.

2 participants