gpo.zugaina.org

Search Portage & Overlays:

dev-libs/rocprofiler-register

Registration library that brokers profiler attachment to the ROCm runtimes

Screenshots

  • rocprofiler-register-10.0.0
    ~amd64
    test

    View      Download      Browse     License: MIT   
    Overlay: stuff

ChangeLog

commit 92dd81781f12641953b47ac9e97f15638b40fbc5
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 02:45:19 2026 +0200

dev-libs/rocprofiler-register: trim 10.0.0 comment prose

commit 5526a78c5228cf0728d24ca4866735253ab33f25
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 23:52:05 2026 +0200

dev-libs/rocprofiler-register: new package, add 10.0.0

First of the rocprofiler family, none of which ::gentoo carries. This is the
brokering library the ROCm runtimes link against so a profiler can attach to
them at runtime; dev-util/rocprofiler-sdk hard-depends on it, so it lands
first.

Versioned by the ROCm release rather than upstream's own VERSION file (0.6.0
here), matching the rest of the stack -- the subslot is what keys rebuilds
across a ROCm bump, and the soname still carries the real 0.6.0.

Two things the release tarball forces:

- external/glog and external/fmt are git submodules, and the tarball ships
them as EMPTY directories. Upstream defaults
ROCPROFILER_REGISTER_BUILD_ to ON, which makes CMake clone them
at configure time -- impossible under network-sandbox. Both options have a
find_package branch, so both are turned OFF and dev-cpp/glog and
dev-libs/libfmt become real dependencies. Confirmed against the built
artifact rather than assumed: objdump -p on librocprofiler-register.so.0.6.0
reports NEEDED libfmt.so.12 and libglog.so.1. Both get := since both
packages subslot on their full version.

- CMakeLists.txt hardcodes CMAKE_INSTALL_LIBDIR to "lib" after
include(GNUInstallDirs), which also defeats -DCMAKE_INSTALL_LIBDIR, so it
has to be sed'd in src_prepare. Guarded with grep-or-die: sed exits 0 on
no-match, and the failure mode here is a silent install into /usr/lib on a
multilib profile rather than a build error.

src_install drops the modulefile and setup-env.sh -- both assume upstream's
self-contained /opt/rocm prefix, and once merged into /usr the modulefile
resolves ROOT to /usr and then prepends /usr/lib64 to LD_LIBRARY_PATH and
/usr/ to PYTHONPATH. It also drops the 300K of test sources upstream installs
unconditionally (a separate install rule, not gated on the tests option) when
USE=-test.

Build-verified and merged on haarmek; install tree is the library, its cmake
config, two headers and docs, with libdir correctly resolved to lib64.