Yes — the PPA has the same gap. It is not a rocrand-side difference; the PPA 
carries the
same target list, so it produces the same one-arch hole.

The PPA's own pkg-rocm-tools is the shortest way to see it.
`pkg-rocm-tools 0.9.6ubuntu4~26.04.1~git202607052315.cf5207a` ships
`data/build-targets/resolute` as:

    gfx908 gfx90a gfx942 gfx950 gfx1030 gfx1100 gfx1101 gfx1151 gfx1200
gfx1201

That is the resolute-proposed list — gfx950 present, gfx1151 present, gfx1150 
still
absent. Every library built in that PPA inherits it.

The build logs confirm it reached cmake unchanged. All three of rocrand, 
rocblas and
rocsparse configured with:

-DGPU_TARGETS="gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1151;gfx1200;gfx1201"

`gfx1150` occurs 0 times in any of the three build logs. For contrast, 
`gfx1151` occurs
165 times (rocrand), 1170 (rocsparse) and 1712 (rocblas).

And in the shipped binaries:

* librocrand1 7.2.4 — `llvm-objdump --offloading` lists exactly those ten 
targets
  and no gfx1150 bundle. (Same result via
  `strings -a librocrand.so.1.1 | grep -oE 
'hipv4-amdgcn-amd-amdhsa--gfx[0-9a-z:+-]*' | sort -u`.)
* librocblas5 7.2.4 — zero files matching `*gfx1150*` anywhere in the package,
  against 96 matching `*gfx1151*`. The `Kernels.so-000-<arch>.hsaco` set gained
  gfx950 relative to the archive build but still has no gfx1150 entry.
* librocsparse1 7.2.4 — this one builds with 
`-DBUILD_WITH_OFFLOAD_COMPRESS=ON`, and
  llvm-objdump 21.1.8 (Ubuntu llvm-21 1:21.1.8-6ubuntu1) cannot walk its 
compressed
  bundles: `while extracting offload FatBin bundles: Failed to decompress 
input: Could
  not decompress embedded file contents: Unknown frame descriptor`. So for 
rocsparse the
  build log above is my evidence rather than the shipped `.so`; I did not 
enumerate its
  code objects directly.

Scope of what I checked: I inspected the PPA's amd64 packages by download and 
extraction
rather than installing them over the archive versions, so the above is a static 
check of
the shipped artifacts and the build logs, not a runtime re-run of the 
reproducer against
PPA packages. Happy to install them on this gfx1150 machine and re-run all three
reproducers if that is more useful to you — say the word and I will post the 
results.

Root cause is unchanged and shared across LP #2162809, #2162810 and #2162811: 
the
gfx1150 omission lives in pkg-rocm-tools' resolute target list, which the PPA 
has too.
The debdiff is on LP #2162810.

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2162809

Title:
  librocrand1 ships no gfx1150 code objects; any generator use segfaults
  on Strix Point (Radeon 890M)

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/rocrand/+bug/2162809/+subscriptions


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

Reply via email to