The root cause is not in this package.

rocblas, rocrand, and rocsparse all set GPU_TARGETS at build time from
`rocm-target-arch` (source package: pkg-rocm-tools), and resolute's
target list — data/build-targets/resolute — carries gfx1151 but not
gfx1150. The history explains the one-arch hole. gfx11-generic, which
covers gfx1150, was deliberately dropped from resolute's list in pkg-
rocm-tools 0.9.6ubuntu2 because upstream doesn't support generic ISAs
without rocblas patches; the specific gfx1150 entry was never added,
while its sibling gfx1151 was. Stonking has gfx11-generic back in its
list, so 26.10 is covered once its libraries rebuild. Resolute needs the
specific ISA.

I've added a pkg-rocm-tools task to this bug and attached a debdiff: one
line adding gfx1150 to resolute's list, plus changelog, based on
0.9.6ubuntu4~26.04.1 from resolute-proposed so it carries the gfx950
change forward. It follows the exact precedent of that gfx950 addition
(LP: #2159601) — same file, same shape of change, five weeks ago. After
acceptance, rocblas, rocrand, and rocsparse need no-change rebuilds to
ship gfx1150 code objects; the proposed changelog references all three
bugs (LP: #2162809, LP: #2162810, LP: #2162811). The SRU team may prefer
different version numbering.

Verified: `rocm-target-arch --dist resolute` from the patched tree emits
gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1150;gfx1151;gfx1200;gfx1201,
and the source package rebuilds cleanly with dpkg-source. I have not
build-tested the library rebuilds themselves — rocblas is a multi-hour
Tensile build. Tensile already emits gfx1151 logic in the current
resolute rocblas and upstream ROCm 7.1 supports gfx1150, so a gfx1150
compile failure is unlikely, and it would surface at build time rather
than as a runtime regression.

SRU justification:

[Impact]
gfx1150 (Strix Point — the Ryzen AI 300 laptop iGPU) gets no device code from 
any ROCm math library in 26.04: rocRAND and rocSPARSE segfault and rocBLAS 
aborts, each on first use, reproducible with the ~10-line C programs attached 
to these three bugs.

[Test Plan]
Rebuild rocrand, rocblas, and rocsparse against the updated pkg-rocm-tools. On 
gfx1150 hardware, run the three attached reproducers: all currently crash 
(SIGSEGV, SIGABRT, SIGSEGV); all should exit 0. roc-obj-ls on the rebuilt 
libraries should list gfx1150 code objects, and rocblas's library directory 
should gain TensileLibrary_lazy_gfx1150.dat.

[Where problems could occur]
The change appends one target to a build-time list; code objects for existing 
architectures are unaffected. Risk concentrates in build time and package size 
growth for the rebuilt libraries, and in the possibility that some kernel fails 
to compile for gfx1150 — a build failure caught before publication, not a 
runtime regression.


** Patch added: "debdiff adding gfx1150 to pkg-rocm-tools resolute build 
targets"
   
https://bugs.launchpad.net/bugs/2162810/+attachment/5991705/+files/pkg-rocm-tools_0.9.6ubuntu5.debdiff

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2162810

Title:
  librocblas5 ships Tensile GEMM kernels for nine gfx targets including
  gfx1151 but omits gfx1150; sgemm aborts on Strix Point

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/pkg-rocm-tools/+bug/2162810/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to