All host (shim) libraries are downloaded unconditionally for each
ROCm version.
It looks like a Gentoo way to solve this would be introduction of some
USE expand like rocm_release_X_X. But at this moment this package is
one of its kind; also binary for each release is only 3-4 mb, so for
now let's download all of them.
Another simple alternative would be to compile host (shim) part and use
prebuilt device kernels, but host part consists of 20000 autogenerated
C++ files, which barely can be named as "aotriton-bin".
Regarding gfx1103: it is not supported now, but will be supported in the
next release, see https://github.com/ROCm/aotriton/pull/138.
Closes: https://bugs.gentoo.org/965907
Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
Part-of: https://github.com/gentoo/gentoo/pull/44595
Closes: https://github.com/gentoo/gentoo/pull/44595
Signed-off-by: Sam James <sam@gentoo.org>
rocm_bandwidth_test cmake files reinvents variables for everything:
installation paths, flags, generator, compiler, linker selection, etc.
Then cmake calls bash, which calls another cmake, which ignores
niceness, verbose logs, CXX, etc.
That code is objectively bad, but at least it can be patched.
New ebuild erases nested build shell script in favor of cmake.eclass
functions and removes automagic adding of ld.lld.
Closes: https://bugs.gentoo.org/965916
Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
Part-of: https://github.com/gentoo/gentoo/pull/44633
Signed-off-by: Sam James <sam@gentoo.org>
Otherwise, they affect global settings and can bleed into other,
subsequently opened buffers without warning.
The fileencoding option is also set via a similar pattern, but it is
only buffer local, so avoid the :setlocal there (it would be harmless).
Signed-off-by: D. Ben Knoble <ben.knoble+github@gmail.com>
Part-of: https://github.com/gentoo/gentoo/pull/44504
Signed-off-by: Sam James <sam@gentoo.org>
It contains some problems or otherwise unnecessary settings for users
with highly customized personal configurations.
Here's taste of what it takes to disable the system vimrc's effects
otherwise:
" Undo /etc/vim/vimrc settings first
set backspace& autoindent& history& ruler& viminfo& suffixes& numberwidth&
set modeline& fileencodings& guifontset& hlsearch&
silent! unmap Q
" - Skip turning off syntax, since my vimrc is smart enough to already check
" t_Co, has('gui_running'), _and_ !exists("g:syntax_on") before calling
" `syntax enabled`.
" - Skip undoing term settings because they don't apply to me
" - Ignore filetype plugin on, even though it comes way too early; I do it again
" later.
" - &shell should not be empty for me
" - I set is_bash set later, fine
" Hmph. These autocommand set options like tabstop, shiftwidth, expandtab
" globally insetad of locally. Broken.
silent! autocmd! gentoo
" Don't need this variable: I have a vimrc
unlet! g:skip_defaults_vim
" Can't revert the Filetype * autocommand for omnifunc because it's not in a
" group, so I'll just have to live with duplicates.
Obviously this doesn't cover all the effects, some of which are
next-to-impossible to reliably revert, and some of the changes or
omissions are relevant only to certain users. But for folks who just
want a blank slate, we need to provide an escape hatch. Unlike
defaults.vim, we cannot provide a simple mechanism to skip the whole
file: by the time we would set the variable, in a local or user vimrc,
it's too late!
Instead, we first source any files in runtimepath called
"skip_etc_vimrc.vim": they have an opportunity to go early and perform
tasks like
let skip_etc_vimrc = 1
to skip the remainder of the script. It's a complicated workaround, but
it avoids a new USE flag.
Signed-off-by: D. Ben Knoble <ben.knoble+github@gmail.com>
Part-of: https://github.com/gentoo/gentoo/pull/44504
Signed-off-by: Sam James <sam@gentoo.org>
Support -flto-incremental=/path/to/cache being used with GCC. Personally,
I am using this hook in /etc/portage/bashrc:
```
pre_src_configure() {
if [[ -n ${PORTAGE_LTO_CACHE} ]] ; then
local lto_path=/var/tmp/portage/lto/${CATEGORY}/${PN}
mkdir -p "${lto_path}"
export CFLAGS="${CFLAGS} -flto-incremental=${lto_path} -flto-compression-level=19"
export CXXFLAGS="${CXXFLAGS} -flto-incremental=${lto_path} -flto-compression-level=19"
export LDFLAGS="${LDFLAGS} -flto-incremental=${lto_path} -flto-compression-level=19"
fi
}
```
... and then `PORTAGE_LTO_CACHE=1` in my development environment.
Signed-off-by: Sam James <sam@gentoo.org>