commit d37b1daf6c05b99d4fe48e487ca8495e84ac5b79
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 01:55:57 2026 +0200

app-admin/rdc: trim 10.0.0 comment prose

commit 6dda03e93d369c86abee6f188f516e61c41c5714
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 3 23:11:21 2026 +0200

app-admin/rdc: add gRPC codegen build dependency

grpc was declared only as a runtime and target dependency
(>=net-libs/grpc-1.78.1:= in RDEPEND, inherited by DEPEND). The build
also runs gRPC's code generators -- protoc together with
grpc_cpp_plugin -- over rdc's .proto files, and those are build-host
executables rather than target libraries.

Under EAPI 8 that is what BDEPEND is for. On a native build the DEPEND
copy lands in / and its binaries happen to be runnable, which is why
nothing ever failed; under cross-compilation DEPEND resolves into the
sysroot, where protoc cannot execute.

commit 7dc2519c7cda30f61651369c01e04d7f46aefe44
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Aug 31 10:44:28 2026 +0200

app-admin/rdc: new package, add 10.0.0

ROCm Data Center Tool: GPU telemetry -- utilisation, power, clocks,
temperatures, ECC counters, per-job statistics -- exposed over a gRPC service
so a fleet can be monitored centrally. Ships the rdcd daemon, the rdci client
and Python bindings for collectd-style integration. ::gentoo carries nothing of
it. app-admin rather than sci-libs: the consumers are monitoring collectors,
not compute code.

net-libs/grpc is a hard requirement with no bundled fallback
(find_package(gRPC 1.78.1 CONFIG REQUIRED)), and it is the transport between
the two binaries rather than an optional extra. libcap is a bare
find_library(cap REQUIRED) -- the daemon drops to CAP_DAC_OVERRIDE -- and
amd_smi supplies the telemetry itself.

BUILD_PROFILER and BUILD_RVS stay off: the first wants the rocprofiler stack,
which this overlay cannot carry yet (rocprofiler-sdk is blocked on unpackaged
gotcha/PTL/perfetto), and the second wants the ROCm Validation Suite, also
unpackaged. rdcd logs a benign "Fail to open librdc_rocp.so" on startup as a
result.

Two install-side notes:

- No service file is installed, and none leaks. Upstream generates
DEBIAN/{postinst,prerm} into the source tree and stages a systemd unit for
CPack, but installs neither -- checked against every install() rule and
confirmed against the image. Its unit also expects a dedicated 'rdc' user
that this package does not create, so pkg_postinst documents running rdcd
directly instead.

- USE=rocr installs 48 PREBUILT GPU code objects that upstream ships in the
source tree rather than building from the .cl sources beside them. They are
GPU ELF images, so portage's strip cannot parse them; dostrip -x excludes
them (STRIP_MASK did not match reliably) and QA_PREBUILT declares them.
Their arch coverage is gfx700 through gfx942 with nothing for gfx11xx or
gfx12xx, so the librdc_rocr.so diagnostics have no image for an RDNA3+ part
-- amd_smi telemetry is unaffected. Recorded in the ebuild.

Build-verified and run-verified end to end on haarmek: rdcd -u comes up
("Server listening on 0.0.0.0:50051"), and rdci against it reports
"1 GPUs found / GPU Index 0 / AMD Radeon 890M Graphics" and enumerates the
supported field IDs. The daemon was stopped afterwards.