The no-change rebuild is done and this package now ships gfx1150 code
objects.

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

Built against pkg-rocm-tools 0.9.6ubuntu4~26.04.1ppa1 in
ppa:aurascoper/rocm-gfx1150-rebuild. The chroot log has:

    Setting up pkg-rocm-tools (0.9.6ubuntu4~26.04.1ppa1) ...
    
-DGPU_TARGETS="gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1150;gfx1151;gfx1200;gfx1201"

gfx1150 occurs 1125 times in that build log, against 0 in the archive
build.

I owe this report a correction on method. My original comment said 
librocsparse.so.1
carried code only for gfx1100/gfx1102, from `strings`. That was wrong, and so 
is the
caveat I posted on LP #2162809 saying the code objects are not enumerable 
locally.
rocSPARSE builds with -DBUILD_WITH_OFFLOAD_COMPRESS=ON, so both `strings` and
llvm-objdump --offloading come up empty or fail, but the bundles are readable. 
The
.hip_fatbin section holds a run of compressed bundles, each a 24-byte CCOB v2 
header in
front of a zstd frame wrapping a normal __CLANG_OFFLOAD_BUNDLE__. Ubuntu's
clang-offload-bundler 21.1.8 stops at the first one and reports "Failed to 
decompress
input: Could not decompress embedded file contents: Src size is incorrect", 
because the
section is 319 concatenated bundles and not one. Walking them by hand gives the 
real
answer.

Archive librocsparse1 7.1.0-0ubuntu2, as installed on this machine from
resolute/universe, 319 bundles, entries and total bytes per target:

    319   153244184  gfx1030
    319   156722704  gfx1100
    319   156997672  gfx1101
    319   156867216  gfx1151
    319   157998808  gfx1200
    319   157984024  gfx1201
    319   150201552  gfx908
    319   148222520  gfx90a
    319   144168800  gfx942

Nine targets. No gfx1150. Also no gfx1102, so my earlier `strings` reading was 
noise
rather than a code object list.

The rebuild, same version string 7.1.0-0ubuntu2, from the PPA above:

    319   153231728  gfx1030
    319   156710256  gfx1100
    319   156985200  gfx1101
    319   157075728  gfx1150
    319   156854824  gfx1151
    319   157986328  gfx1200
    319   157971536  gfx1201
    319   150189096  gfx908
    319   148210000  gfx90a
    319   144156384  gfx942
    319   144307568  gfx950

gfx1150 present in all 319 bundles, 157 MB of device code, in line with its 
siblings
rather than a stub.

gfx950 is the other new row and it is not from my patch. That is
pkg-rocm-tools 0.9.6ubuntu4~26.04.1, LP #2159601, already in resolute-proposed. 
It makes
a useful control: the same no-change rebuild that picks up Canonical's gfx950 
addition
picks up gfx1150, so the mechanism is doing the work here, not anything 
specific to my
debdiff.

Scope: 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 
rocsparse_create_handle
reproducer against the rebuilt library yet. Happy to do that and post the 
result.

The debdiff and the SRU discussion are on LP #2162810, now targeting
0.9.6ubuntu4~26.04.2 rather than the 0.9.6ubuntu5 I first proposed.

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2162811

Title:
  librocsparse1 segfaults in rocsparse_create_handle on gfx1150 (no
  gfx11.5 code objects)

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/rocsparse/+bug/2162811/+subscriptions


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

Reply via email to