Sorry for the confusion in my earlier answer. At the time you asked there was 
no PPA.
What was attached to this bug was a debdiff 
(pkg-rocm-tools_0.9.6ubuntu5.debdiff), not
an upload. There are two now:

  https://launchpad.net/~aurascoper/+archive/ubuntu/rocm-gfx1150
    pkg-rocm-tools 0.9.6ubuntu4~26.04.1ppa1, the patched build-target list.

  https://launchpad.net/~aurascoper/+archive/ubuntu/rocm-gfx1150-rebuild
    that same package copied in with its binaries, plus no-change rebuilds of 
the
    libraries against it.

The split is not by design, and the reason may be worth knowing if you rebuild 
these
yourselves. My first attempt copied the three libraries into the first PPA 
about four
minutes before its pkg-rocm-tools binary finished publishing. All three built 
against
resolute's pkg-rocm-tools 0.9.6ubuntu3, produced no gfx1150 at all, and reported
"Successfully built" while doing it. Launchpad will not re-copy a version that 
has
already published in a given archive, so the corrected rebuilds needed a second 
one.
The gate to wait on is the binary publication, not the source.

rocRAND is rebuilt and verified:

  https://launchpad.net/~aurascoper/+archive/ubuntu/rocm-
gfx1150-rebuild/+build/33518650

Its chroot installed the patched tool:

    Setting up pkg-rocm-tools (0.9.6ubuntu4~26.04.1ppa1) ...

and cmake received the eleven-target list:

-DGPU_TARGETS="gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1150;gfx1151;gfx1200;gfx1201"

gfx1150 occurs 138 times in that build log, against 0 in the archive and
ppa:bullwinkle-team/rocm builds I reported on LP #2162809.

The shipped library carries the code objects, not just the build flags. From
librocrand1 7.1.1-0ubuntu1 in that PPA, llvm-objdump --offloading on
/usr/lib/x86_64-linux-gnu/librocrand.so.1.1 lists eleven bundles:

    gfx1030 gfx1100 gfx1101 gfx1150 gfx1151 gfx1200 gfx1201 gfx908
gfx90a gfx942 gfx950

with the same set from
strings -a librocrand.so.1.1 | grep -oE 
'hipv4-amdgcn-amd-amdhsa--gfx[0-9a-z:+-]*' | sort -u.

Scope of that check: I downloaded and extracted the .deb rather than installing 
it, so
this is a static check of the shipped artifact. I have not re-run the rocRAND
reproducer against it on the gfx1150 machine yet. Happy to do that and post the 
result
if it is useful to you.

rocSPARSE and rocBLAS are now rebuilt and verified the same way:

  
https://launchpad.net/~aurascoper/+archive/ubuntu/rocm-gfx1150-rebuild/+build/33519189
  
https://launchpad.net/~aurascoper/+archive/ubuntu/rocm-gfx1150-rebuild/+build/33519190

Both chroots installed pkg-rocm-tools 0.9.6ubuntu4~26.04.1ppa1 and both 
received the
same eleven-target -DGPU_TARGETS= line. gfx1150 occurs 1125 times in the 
rocSPARSE log
and 2011 times in the rocBLAS log, against 0 in the archive builds.

rocBLAS is worth showing separately because its device code does not live in 
the shared
object. It ships Tensile code objects as files. In librocblas5 7.1.0-1ubuntu4 
from this
PPA:

    92 gfx1150 files in /usr/lib/x86_64-linux-gnu/rocblas/5.1.0/library/

gfx1151, gfx1101 and gfx1100 each get the same 92. The archive build of the 
identical
version string, resolute/universe librocblas5 7.1.0-1ubuntu4 as installed on 
this
machine, has 0. The names are not the evidence on their own, so from the ELF
header of one of them:

    $ llvm-readelf -h Kernels.so-000-gfx1150.hsaco
      OS/ABI:   AMDGPU - HSA
      Machine:  EM_AMDGPU
      Flags:    0x43, gfx1150

One difference in that comparison is not mine: the rebuild also gains 152 
gfx950 files
the archive build does not have. That is 0.9.6ubuntu4~26.04.1 from LP #2159601, 
already
in resolute-proposed, doing exactly what it should once a library is rebuilt 
against it.
It is a useful control. The same rebuild that picks up your gfx950 addition 
picks up
gfx1150.

librocsparse1 builds with -DBUILD_WITH_OFFLOAD_COMPRESS=ON, and llvm-objdump 
21.1.8
cannot read the compressed bundles ("Unknown frame descriptor"), so its 
evidence is the
build log rather than the shipped artifact. Same caveat I noted on LP #2162809.

Two things worth settling separately.

Version string. I have moved the debdiff off 0.9.6ubuntu5. Resolute's stream 
for this
package uses the backport-style suffix, and 0.9.6ubuntu4~26.04.1 is what is 
sitting in
resolute-proposed, so 0.9.6ubuntu4~26.04.2 is the version an SRU would actually 
take.
That is what the updated debdiff targets. The content is unchanged: one line 
added to
data/build-targets/resolute.

The devel-series half. Confirmed against the shipped package rather than from 
memory:
pkg-rocm-tools 0.9.8ubuntu2 in stonking has data/build-targets/stonking 
containing
gfx11-generic alongside gfx10-3-generic. That covers gfx1150, so 26.10 needs 
nothing
from this bug beyond its libraries rebuilding. This is resolute-only.

One catch. The resolute list inside that same 0.9.8ubuntu2 is still gfx908 
gfx90a
gfx942 gfx950 gfx1030 gfx1100 gfx1101 gfx1151 gfx1200 gfx1201, with no gfx1150. 
A later
sync will not pick this up on its own. Resolute needs the explicit ISA added.

-- 
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