[jaopaulolc]

JIRA ID : AIHPBLAS-4801

Motivation

hipBLASLt runs each matrix multiplication (GEMM) with a kernel from a library that is generated and tuned at build time, so a problem the library serves poorly has no better kernel at run time. Just-in-time (JIT) generation produces a kernel for a problem when it is needed. The previous PR builds generated kernel sources with comgr and loads the code object into a TensileLite library of one or more solutions. This PR runs those solutions through hipblasLtMatmul and hipblaslt_ext::Gemm like a prebuilt one, and it proves the path end to end on a GPU with the committed kernel.

Technical Details

The feature is behind HIPBLASLT_ENABLE_JIT, which is off by default. The commits each build on their own:

• Pipeline. Jit runs one request for one device through four stages, each an interface: Generate (a backend returns library entries and their kernel sources), Build, Support (each solution's predicates and workspace) and Load. It returns up to the requested count of supported solutions, best first. A failure in one entry skips it and records its stage.
• In-process GEMM. The internal entry points makeGemmRequest, getJitAlgo and getGemmAlgo return a hipblasLtMatmulAlgo_t that carries a process-local token and the solution's local index. The TensileLite loader in tensile_host.cpp loads an entry once and wraps each supported solution, with the shared library and adapter, in a TensileGemmBundle. get_library_and_adapter resolves the token, and the solution then runs through the normal launch path, including Stream-K's per-stream synchronization. A token from another process or device is rejected before any work is submitted. No installed header declares these entry points.
• Replay backend. Every JIT build compiles it. It implements Backend and returns pre-generated source bundles instead of running a generator, and Jit builds and loads those entries with the same builder and loader as any other backend.
• Docs in JIT.md and the JIT test guide. The workflow watches the GEMM host files.

With JIT off, the shared host code still changes in small ways. The GEMM data keeps the selected algorithm and the scale layouts. makeArgument stores its kernels only after they are ready, and it restores the handle's synchronizer on reinitialization. run before initialize returns NOT_INITIALIZED. Kernel-name lookups for an unknown index return an empty name. heuristicResult_copy copies the whole algorithm. Two internal status codes are added.

Not in this PR: a generator backend and a solution store. Later PRs add them.

Test Plan

Host-only builds with GPU_TARGETS=gfx950: JIT on with JIT testing, and JIT off. A third host build uses GPU_TARGETS=gfx942 with JIT on. In each build, CTest runs the JIT tests in clients/tests/jit. The previous PR adds jit-bundles, jit-source-bundle, jit-builder and jit-loader. This PR adds:

• jit-end-to-end replays the two-solution plain-pair bundle that jit-bundles writes from the committed kernel assembly, builds and loads it, and runs M=256, N=128 through hipblasLtMatmul and hipblaslt_ext::Gemm twice, checking D against a host reference: K=512 runs the first solution and K=256 the second. It also checks that a transposed-A problem is not supported. It replays the bundle of device 0's architecture, so the gfx90a, gfx942 and gfx950 lanes each run their own kernel. CMake registers it in builds with JIT testing when GPU_TARGETS include an architecture with committed bundles.
• jit-disabled, in the JIT-off build, checks that the JIT headers are not in the public include tree.

A device-library build for gfx950, restricted to one FP16 logic file, ran the JIT tests and the hipblaslt-test quick suite. The suite was compared with the same build of develop. actionlint checked the workflow.

Test Result

Build	Tests	Result
gfx950, JIT on, on one gfx950 GPU	jit-bundles, jit-source-bundle, jit-builder, jit-loader, jit-end-to-end	5 of 5 passed
gfx950, JIT on, every GPU hidden	jit-bundles, jit-source-bundle, jit-builder	3 of 3 passed
gfx950, JIT off, every GPU hidden	jit-source-bundle, jit-disabled	2 of 2 passed
gfx942, JIT on, every GPU hidden	jit-bundles, jit-source-bundle, jit-builder (-L jit-cpu; the build also registers jit-loader and jit-end-to-end)	3 of 3 passed
gfx950 device library, JIT on	JIT tests	5 of 5 passed
gfx950 device library	hipblaslt-test quick suite	139 passed, 785 failed on both develop and this branch, with identical test names

The 785 failures report "NO solution found", because the build keeps one logic file. Each commit built with JIT on and passed the JIT tests it registers; the last commit also built with JIT off. There is no gfx90a or gfx942 GPU on the test host, so the workflow's lanes are the first native runs of jit-loader and jit-end-to-end on those architectures. actionlint 1.7.12 reported no errors. The workflow's three lanes and the other automated GitHub checks are pending.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests.

Risk level

Low (2/5): the JIT sources build only with HIPBLASLT_ENABLE_JIT=ON, which is off by default. With JIT off, the changes are the small shared host-code edits listed in Technical Details, and the quick suite matched develop.

[codecov]

Codecov Report

❌ Patch coverage is 59.55056% with 36 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...rary/src/amd_detail/rocblaslt/src/tensile_host.cpp	64.47%	16 Missing and 11 partials ⚠️
...cts/hipblaslt/library/src/amd_detail/hipblaslt.cpp	0.00%	4 Missing and 4 partials ⚠️
...ary/src/amd_detail/rocblaslt/src/rocblaslt_mat.cpp	75.00%	0 Missing and 1 partial ⚠️

❌ Your project check has failed because the head coverage (49.03%) is below the target coverage (80.00%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files

@@                                Coverage Diff                                @@
##           users/jolabega/jit-upstream-01-jit-interfaces   #13131      +/-   ##
=================================================================================
- Coverage                                          72.13%   72.13%   -0.00%     
=================================================================================
  Files                                               2861     2861              
  Lines                                             472127   472165      +38     
  Branches                                           69811    69823      +12     
=================================================================================
+ Hits                                              340551   340570      +19     
- Misses                                            107831   107839       +8     
- Partials                                           23745    23756      +11     

Flag	Coverage Δ		*Carryforward flag
TensileLite-CPP	49.03% <ø> (ø)
TensileLite-Unit	83.28% <ø> (ø)
hipBLAS	90.62% <ø> (ø)		Carriedforward from a6c8eb0
hipBLASLt	39.06% <59.55%> (+0.02%)	⬆️
hipCUB	82.68% <ø> (ø)		Carriedforward from a6c8eb0
hipDNN	86.64% <ø> (ø)		Carriedforward from a6c8eb0
hipFFT	54.32% <ø> (ø)		Carriedforward from a6c8eb0
hipRAND	76.12% <ø> (ø)		Carriedforward from a6c8eb0
hipSOLVER	68.96% <ø> (ø)		Carriedforward from a6c8eb0
hipSPARSE	86.99% <ø> (ø)		Carriedforward from a6c8eb0
rocBLAS	48.25% <ø> (ø)		Carriedforward from a6c8eb0
rocFFT	48.78% <ø> (ø)		Carriedforward from a6c8eb0
rocRAND	57.45% <ø> (ø)		Carriedforward from a6c8eb0
rocSOLVER	76.83% <ø> (ø)		Carriedforward from a6c8eb0
rocSPARSE	74.75% <ø> (ø)		Carriedforward from a6c8eb0
rocThrust	93.15% <ø> (ø)		Carriedforward from a6c8eb0

*This pull request uses carry forward flags. Click here to find out more.

Files with missing lines	Coverage Δ
...hipblaslt/library/src/amd_detail/hipblaslt-ext.cpp	60.10% <ø> (ø)
...c/amd_detail/rocblaslt/src/rocblaslt_auxiliary.cpp	74.56% <100.00%> (-0.02%)	⬇️
...ary/src/amd_detail/rocblaslt/src/rocblaslt_mat.cpp	83.82% <75.00%> (-0.08%)	⬇️
...cts/hipblaslt/library/src/amd_detail/hipblaslt.cpp	66.96% <0.00%> (-0.94%)	⬇️
...rary/src/amd_detail/rocblaslt/src/tensile_host.cpp	43.67% <64.47%> (+0.19%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🔎 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]

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

[jaopaulolc]

Closing this as a review checkpoint. The branch stays, because #13202 is based on it.

Solutions run only as JIT library indices. A query looks up the library first and publishes on a miss. The process-local in-memory algorithm path is not the intended way to run a JIT solution and will be removed on #13202. The pipeline and the replay backend in this branch remain, because a miss still has to build a code object before it can publish.