Re: CMake musings for ROCm 4.5.2 (Was Re: ROCm packaging session memo)
Hello,
About finding the device libs at compile-time, what about switching to a
pure find_package?
Here is such an example of a switch from heuristics for ROCr-Runtime [1].
For the compiler, rocr currently passes the libs itself [2], hipamd
passes the compiler flag [3],
and the backend has its fallback environment variables and paths as
already mentioned:
what about, everywhere the device-libs are needed, passing to the
backend the flag
--hip-device-lib-path [3], possibly with a function in the cmake export?
About finding the device libs at runtime, I am not sure what to do. (Is
there even a need?)
For reference, right now my involvement in ROCm is very high level,
focused mostly on licensing,
installation, and packaging related to runtime
(rocr/compiler/rocclr/ocl/hip/hipamd).
Other than device libs, as far as I know, the open points remaining are:
- have the hipamd shared library support all target GPU archs in one
shared object / Debian package.
- install hip in GNUInstallDirs (I am fiddling about it here [4]), with
the use of the clang flag
-isystem at stake for the reverse dependencies of hipamd.
- have openCL 2.2 working (had no help from the Debian OpenCL team so far,
will possibly open an issue at AMD's github).
Do you know why libamdhip64.so does not build with debug symbols?
I would like to focus next days on running debug sessions on HIP,
and possibly inside compute kernels too.
Best regards! Maxime
[1]
https://salsa.debian.org/rocm-team/rocr-runtime/-/blob/master/debian/patches/0001-find-devlibs-in-image-blit-src-cmakelist.patch
[2]
https://salsa.debian.org/rocm-team/rocr-runtime/-/blob/master/src/image/blit_src/CMakeLists.txt#L112
[3]
https://salsa.debian.org/rocm-team/rocm-hipamd/-/blob/master/hip-config.cmake.in#L235
[4] https://salsa.debian.org/rocm-team/rocm-hipamd/-/tree/install_wip
Reply to: