AMDGPU backend on a consumer iGPU (Strix Point, gfx1150): 249/249 pass — because JACC's portable paths bypass three broken vendor libraries

[aurascoper]

Filling the AMD column's consumer-hardware gap the way #402 did for Metal. Versions throughout: JACC 0dbf990a (main), AMDGPU.jl 2.7.1, Julia 1.12.6, Ubuntu 26.04's archive ROCm 7.1 (7.1.0/7.1.1 Debian packaging, no /opt/rocm), kernel 7.0.0-28-generic, AMD Ryzen AI 9 HX 370 / Radeon 890M (gfx1150, RDNA 3.5 iGPU). Your AMD CI runs on an MI100; as far as I can find, nobody has published a consumer-iGPU result for this backend.

Result

Full suite via the canonical invocation (julia --project in the clone → Pkg.instantiate() → JACC.set_backend("AMDGPU") → Pkg.test() in a fresh process): 249/249, zero failures, including AXPY, reduce, reduce-ND, shared, JACC.BLAS, Add-2D/3D/ND, CG, LBM, rand-Float32/Float64, UnitRange/StepRange, LaunchSpec. Since AMDGPU has no default_float override, the whole suite ran in Float64 — notable because Float64 was exactly what broke on Metal (#402), and on this consumer part it did not (correctness only; I measured no throughput).

Two setup facts worth recording for anyone reproducing on Strix Point:

• gfx1150 enumerates natively under this ROCm — HSA_OVERRIDE_GFX_VERSION is unnecessary (=11.5.0 is identical to unset; =11.0.0 actively masquerades the device as gfx1100).
• AMDGPU.jl found the Debian layout unaided — no ROCM_PATH, no LocalPreferences.

The finding I'd actually flag: the green run is the abstraction working, not the platform working

Three of the distro's ROCm math libraries are broken for gfx1150, each differently:

• rocRAND: SIGSEGV in librocrand.so during rocrand_initialize_generator, type-independent (AMDGPU.rand(Float32, …) and Float64 both crash). The binary embeds gfx1151 kernels but no gfx1150.
• rocBLAS: clean failure — Cannot read /usr/lib/x86_64-linux-gnu/rocblas/5.1.0/library/TensileLibrary.dat: Illegal seek for GPU arch : gfx1150; the only Tensile file shipped is TensileLibrary_lazy_gfx942.dat.
• rocSPARSE: crashes AMDGPU.jl's isolated version-query subprocess (the warning cites Trying install AMDGPU on ubuntu 26.04 with no success. AMDGPU.jl#920).

The suite passes anyway because JACC's portable API never reaches them: JACC.BLAS is JACC's own parallel_for kernels (src/blas.jl), and the rand testsets call device-side rand() inside kernels (unittests.jl:1228), not the rocRAND host API. So the accurate statement is narrow: JACC insulated a user from three independent vendor gaps — which seems like the strongest concrete evidence for the portability thesis I've seen, and a better result than the checkmark. The edge cuts the other way too: a JACC user who steps outside the API (AMDGPU.rand, anything rocBLAS-backed) hits the breakage immediately, so "JACC passes on gfx1150" must not be read as "gfx1150 is fine."

These are Ubuntu packaging gaps (vendor-stack ≠ distro-package; AMD's 24.04 repo may differ on every axis) — I'll report them there, not here. They're listed because they define exactly what the 249/249 does and does not mean.

Full environment pins, verbatim logs, and the reproduction command sequence are retained; glad to attach any of it.

[aurascoper]

One correction to the issue body, caught while preparing distro bug reports, plus a strengthening.

Correction. The body says the only Tensile file shipped is TensileLibrary_lazy_gfx942.dat. That's wrong — my own captured log listed lazy .dat files for gfx942/90a/908/1030/1200/1151 all along, and a direct ls of /usr/lib/x86_64-linux-gnu/rocblas/5.1.0/library/ shows nine targets: gfx908/90a/942/1030/1100/1101/1151/1200/1201. The claim was built from a truncated display of the log, not from the log. The accurate statement is narrower and more interesting: Ubuntu's rocBLAS builds GEMM kernels for every current target including gfx1151 (Strix Halo) and omits exactly gfx1150 (Strix Point) — the same one-arch omission as rocRAND's code objects and rocBLAS's own Kernels.so-*.hsaco set. The three-libraries-broken conclusion and everything downstream of it stands; the "datacenter-only" characterization does not.

Strengthening. All three breakages now reproduce without Julia, in ~10-line C programs compiled with the distro's own hipcc (rocRAND: SIGSEGV in generator create/generate, exit 139, before any output; rocBLAS: rocblas_sgemm → the TensileLibrary error, then SIGABRT, exit 134; rocSPARSE: SIGSEGV inside rocsparse_create_handle, exit 139). So none of this is a JuliaGPU-stack interaction — it's the packaged libraries themselves. Filing these against Ubuntu (rocrand 7.1.1-0ubuntu1, rocblas 7.1.0-1ubuntu4, rocsparse 7.1.0-0ubuntu2) and will link the Launchpad bugs here when they exist.

[aurascoper]

Launchpad bugs now filed for the three library gaps, each with the ~10-line C reproducer: rocrand → https://bugs.launchpad.net/bugs/2162809, rocblas → https://bugs.launchpad.net/bugs/2162810, rocsparse → https://bugs.launchpad.net/bugs/2162811.