Arch Linux

Aphid_ARCH

Member

Registered: 2015-08-08

Posts: 7

Arch version of ROCm is broken.

So far, I've found that the version of rocm (6.0.2) in the official repositories is broken and out of date.

Try to run stable diffusion --> runs into memory corruption and GPU resets.

So I've taken it upon myself (foolishly) to try to compile it. I found a useful template over at https://www.ixip.net/rocm/ which is recent-ish rocm compile for slackware collection of scripts.

So using that as a basis, I tried compiling piece by piece, executing the scripts line by line (though using the latest git version rather than 6.0 or 6.1, as I wanted to see if I can somehow bisect), but I got stuck on HIP compilation.

First, there's the hip_prof_gen.py script, which is badly documented. Anyone know how to use that thing? This is what I do now, but it's wrong somewhere.

OPT_PROF_API=""
PROF_API_STR="$BASEDIR/09_hip/clr/hipamd/include/hip/amd_detail/hip_prof_str.h"
PROF_API_HDR="$BASEDIR/09_hip/hip/include/hip/hip_runtime_api.h"
PROF_API_SRC="$BASEDIR/09_hip/clr/hipamd/src" 
PROF_API_LOG="$BASEDIR/09_hip/build/hip_prof_gen_log.txt"
PROF_API_STR_NEW="$BASEDIR/09_hip/clr/hipamd/include/hip/amd_detail/hip_prof_str.h.new"

# set(PROF_API_CMD "${PROF_API_GEN} -v ${OPT_PROF_API} ${PROF_API_HDR} ${PROF_API_SRC} ${PROF_API_STR} >${PROF_API_LOG}")
python $BASEDIR/09_hip/clr/hipamd/src/hip_prof_gen.py -v ${OPT_PROF_API} ${PROF_API_HDR} ${PROF_API_SRC} ${PROF_API_STR} ${PROF_API_STR_NEW}  >${PROF_API_LOG}

This doesn't produce anything, so I'm still stuck with circular reference in hip_prof_str.h. I get various errors depending on how I call this thing.

Warning: bad args: args_str: 'amd::Os::FileDesc fdesc,size_t fsize,const void** image,const std::vector<std::string>& device_names,std::vector<std::pair<const void* ,size_t>>& code_objs' arg_pair: 'std::vector<std::pair<const void* ', file '/home/aphid/src/rocm/09_hip/clr/hipamd/src/hip_code_object.cpp', line (498)

or this one

/home/aphid/src/rocm/09_hip/clr/hipamd/src/hip_prof_gen.py Warning: implementation not found: hipGetErrorString
/home/aphid/src/rocm/09_hip/clr/hipamd/src/hip_prof_gen.py Warning: 1 API calls missing in interception layer

Here, it does seem to output a new file. But trying the new file gets me the same problem as the old one, the reason why I set out on all this trouble in the first place;

ninja: error: dependency cycle: hipamd/include/hip/amd_detail/hip_prof_str.h -> hipamd/include/hip/amd_detail/hip_prof_str.h

Just by default, not modifying it, it gets stuck trying to double-include/redefine a bunch of stuff from cmake, really basic math functions like 'min'  and 'max' as well. I've made a couple patches that modify stuff at https://github.com/AphidGit/rocm_compile/tree/main to get around the redefinitions.

Last edited by Aphid_ARCH (2024-07-20 05:31:29)

Luciddream

Member

From: Greece

Registered: 2014-12-08

Posts: 31

Re: Arch version of ROCm is broken.

So far, I've found that the version of rocm (6.0.2) in the official repositories is broken and out of date.

You can always try opencl-amd as an alternative ROCm library.

Aphid_ARCH

Member

Registered: 2015-08-08

Posts: 7

Re: Arch version of ROCm is broken.

Oh, also, this:

pacman -Sy rocminfo
PATH=$PATH:/opt/rocm:/opt/rocm/bin
rocminfo 

#> rocminfo: error while loading shared libraries: librocprofiler-register.so.0: cannot open shared object file: No such file or directory
# Okay, let's try installing that. 

pkgfile librocprofiler-register.so.0 

#>
# Crickets...

Out of the box it can't even work without hacky workarounds (compiling missing libraries yourself).

Luciddream

Member

From: Greece

Registered: 2014-12-08

Posts: 31

Re: Arch version of ROCm is broken.

Oh, also, this:

You probably have extra-staging repository enabled which has updated rocminfo to the latest version.

Last edited by Luciddream (2024-08-16 06:45:00)