public inbox for gentoo-dev@lists.gentoo.org
 help / color / mirror / Atom feed
From: "Sv. Lockal" <lockalsash@gmail.com>
To: gentoo-dev@lists.gentoo.org
Subject: [gentoo-dev] Re: [PATCH] rocm.eclass: add rocm_use_hipcc function and update example accordingly
Date: Sat, 17 Aug 2024 20:10:39 +0800	[thread overview]
Message-ID: <CACSrLEqgSuDwB1GPqfwkOJH_gR8K5acyNomenpV7gmDoYvWOEQ@mail.gmail.com> (raw)
In-Reply-To: <CACSrLErNt6g+tzaytsKDqRkO3P=kKGGn5pfUgVZAAxEzFjpFBw@mail.gmail.com>

Following the discussion in https://github.com/gentoo/gentoo/pull/37639,
rocm_use_hipcc now creates ROCM_TARGET_LST to prevent GPU access
during src_configure phase. Interdiff is below:

diff -u b/eclass/rocm.eclass b/eclass/rocm.eclass
--- b/eclass/rocm.eclass
+++ b/eclass/rocm.eclass
@@ -20,7 +20,7 @@
 # in /etc/portage/make.conf). Function check_amdgpu can help ebuild ensure
 # read and write permissions to GPU device in src_test phase, throwing friendly
 # error message if unavailable. However src_configure in general should not
-# access any GPU devices. If it does, it usually means that CMakeLists.txt
+# access any AMDGPU devices. If it does, it usually means that CMakeLists.txt
 # ignores AMDGPU_TARGETS in favor of autodetected GPU, which is not desired.
 #
 # @EXAMPLE:
@@ -239,8 +239,23 @@
 # @FUNCTION: rocm_use_hipcc
 # @USAGE: rocm_use_hipcc
 # @DESCRIPTION:
-# switch active C and C++ compilers to hipcc and clean unsupported flags.
+# switch active C and C++ compilers to hipcc, clean unsupported flags
and setup ROCM_TARGET_LST file.
 rocm_use_hipcc() {
+       # During the configuration stage, CMake tests whether the
compiler is able to compile a simple program.
+       # Since CMake checker does not specify --offload-arch=, hipcc
enumerates devices using four methods
+       # until it finds at least one device. Last way is by accessing
them (via rocminfo).
+       # To prevent potential sandbox violations, we set the
ROCM_TARGET_LST variable (which is checked first).
+       local target_lst="${T}"/gentoo_rocm_target.lst
+       if [[ "${AMDGPU_TARGETS[@]}" = "" ]]; then
+               # Expected no GPU code; still need to calm down sandbox
+               echo "gfx000" > "${target_lst}" || die
+       else
+               printf "%s\n" ${AMDGPU_TARGETS[@]} > "${target_lst}" || die
+       fi
+       export ROCM_TARGET_LST="${target_lst}"
+
+       # Export updated CC and CXX. Note that CC is needed even if no
C code used,
+       # as CMake checks that C compiler can compile a simple test program.
        export CC=hipcc CXX=hipcc
        strip-unsupported-flags
 }

On Sun, Aug 4, 2024 at 5:09 AM Sv. Lockal <lockalsash@gmail.com> wrote:
>
> This adds a new function rocm_use_hipcc in rocm.eclass to be used in
> every src_configure, where the compiler is switched to hipcc (a. k. a.
> clang with extra flags).
>
> Github pull request: https://github.com/gentoo/gentoo/pull/37639
> Bug: https://bugs.gentoo.org/936099
> Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
> ---
>  eclass/rocm.eclass | 25 +++++++++++++++++++------
>  1 file changed, 19 insertions(+), 6 deletions(-)
>
> diff --git a/eclass/rocm.eclass b/eclass/rocm.eclass
> index 7039455dec6b5..44ddf4fd09a87 100644
> --- a/eclass/rocm.eclass
> +++ b/eclass/rocm.eclass
> @@ -15,9 +15,13 @@
>  # edit USE flag to control which GPU architecture to compile. Using
>  # ${ROCM_USEDEP} can ensure coherence among dependencies. Ebuilds can call the
>  # function get_amdgpu_flag to translate activated target to GPU compile flags,
> -# passing it to configuration. Function check_amdgpu can help ebuild ensure
> +# passing it to configuration. Function rocm_use_hipcc switches active compiler
> +# to hipcc and cleans incompatible flags (useful for users with gcc-only flags
> +# in /etc/portage/make.conf). Function check_amdgpu can help ebuild ensure
>  # read and write permissions to GPU device in src_test phase, throwing friendly
> -# error message if unavailable.
> +# error message if unavailable. However src_configure in general should not
> +# access any GPU devices. If it does, it usually means that CMakeLists.txt
> +# ignores AMDGPU_TARGETS in favor of autodetected GPU, which is not desired.
>  #
>  # @EXAMPLE:
>  # Example ebuild for ROCm library in https://github.com/ROCmSoftwarePlatform
> @@ -39,14 +43,12 @@
>  # "
>  #
>  # src_configure() {
> -#     # avoid sandbox violation
> -#     addpredict /dev/kfd
> -#     addpredict /dev/dri/
> +#     rocm_use_hipcc
>  #     local mycmakeargs=(
>  #         -DAMDGPU_TARGETS="$(get_amdgpu_flags)"
>  #         -DBUILD_CLIENTS_TESTS=$(usex test ON OFF)
>  #     )
> -#     CXX=hipcc cmake_src_configure
> +#     cmake_src_configure
>  # }
>  #
>  # src_test() {
> @@ -90,6 +92,8 @@ esac
>  if [[ ! ${_ROCM_ECLASS} ]]; then
>  _ROCM_ECLASS=1
>
> +inherit flag-o-matic
> +
>  # @ECLASS_VARIABLE: ROCM_VERSION
>  # @REQUIRED
>  # @PRE_INHERIT
> @@ -231,3 +235,12 @@ check_amdgpu() {
>  }
>
>  fi
> +
> +# @FUNCTION: rocm_use_hipcc
> +# @USAGE: rocm_use_hipcc
> +# @DESCRIPTION:
> +# switch active C and C++ compilers to hipcc and clean unsupported flags.
> +rocm_use_hipcc() {
> + export CC=hipcc CXX=hipcc
> + strip-unsupported-flags
> +}


      reply	other threads:[~2024-08-17 12:10 UTC|newest]

Thread overview: 2+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2024-08-03 21:09 [gentoo-dev] [PATCH] rocm.eclass: add rocm_use_hipcc function and update example accordingly Sv. Lockal
2024-08-17 12:10 ` Sv. Lockal [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CACSrLEqgSuDwB1GPqfwkOJH_gR8K5acyNomenpV7gmDoYvWOEQ@mail.gmail.com \
    --to=lockalsash@gmail.com \
    --cc=gentoo-dev@lists.gentoo.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox