sci-ml/gloo: add support of AMD GPUs (USE=rocm) and fix cuda

There is a chicken-egg build dependency between gloo and pytorch.
To build with GLOO_USE_TORCH_DTYPES, few pytorch headers are required.
So instead of this:
1) USE=-gloo emerge caffe2
2) emerge gloo
3) USE=gloo emerge caffe2
gloo ebuild now includes some (2.9.0) version of caffe2 just for headers

Closes: https://bugs.gentoo.org/965091
Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
Part-of: https://github.com/gentoo/gentoo/pull/44550
Signed-off-by: Alfredo Tupone <tupone@gentoo.org>

sci-ml/gloo/Manifest

@@ -1,2 +1,3 @@
 DIST gloo-2023.12.03.tar.gz 259857 BLAKE2B b56cbbb0986b70ed327a8d019f140099c5d1d725576120089cd5e640dc42daa6c6bb67de877d540fd2160f5b8d89cf54af7c11a1b81039e62e4de97daa4f63d6 SHA512 597679dd96394e7d0009be34ef9bd6f8446d22e2faa006c0853e055caf0d6d821c7532ff867b4d924197e92ab53c4c7ee9b74f897d6c5a469ee797c11ce0d28d
 DIST gloo-2025.06.04.tar.gz 272417 BLAKE2B f3cf3326276bb839166f3094d49fcd537e5a446bda65fb34adad5540df4c505bf7f7f56071d78d14cbc49db6d02f190198e45dd25ae307cfbc5d27db4c2c3703 SHA512 87264b940f2a9f50f3054d3853ad3c0aff33c29253e27c4d0e5fc18a674ebb362719a811b01847daaadf0b9a151a51eabd9dae694ac7d1ad300e24842bea1241
+DIST pytorch-2.9.0.tar.gz 55750268 BLAKE2B 943459ec60a4e1f5e36766defc7018fbf9722fb1564b723c2a7ebcb2a5d8b1735f0b1542dc67a77f788af3e2454ea6261dbdee5beb2bcfa4af2e58ca566edc93 SHA512 2ecdc0eac39ecee68b0f4c98e498424cde00c45bbeeff576c8778046f97119cd02885498b072352dd3cdd9aecd02baf61cdc5554bce8d757b30c673053a0cc80

sci-ml/gloo/files/gloo-2025.06.04-gentoo.patch

@@ -1,6 +1,6 @@
---- a/CMakeLists.txt	2022-06-09 08:29:13.172119535 +0200
-+++ b/CMakeLists.txt	2022-06-09 08:29:27.474929097 +0200
-@@ -2,6 +2,9 @@
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -2,6 +2,9 @@ cmake_minimum_required(VERSION 3.18 FATAL_ERROR)
  
  project(gloo CXX C)
  
@@ -10,27 +10,9 @@
  set(GLOO_VERSION_MAJOR 0)
  set(GLOO_VERSION_MINOR 5)
  set(GLOO_VERSION_PATCH 0)
---- a/gloo/CMakeLists.txt	2022-06-09 08:41:09.883773600 +0200
-+++ b/gloo/CMakeLists.txt	2022-06-09 08:42:39.045637103 +0200
-@@ -191,7 +191,7 @@
- # want to statically link with Gloo and not install any artifacts.
- if(GLOO_INSTALL)
-   install(TARGETS gloo EXPORT GlooTargets
--      DESTINATION ${CMAKE_INSTALL_PREFIX}/lib)
-+          DESTINATION ${CMAKE_INSTALL_LIBDIR})
-   if(USE_CUDA)
-     install(TARGETS gloo_cuda EXPORT GlooTargets
-         DESTINATION ${CMAKE_INSTALL_PREFIX}/lib)
---- a/gloo/test/CMakeLists.txt	2022-06-10 22:17:03.682856314 +0200
-+++ b/gloo/test/CMakeLists.txt	2022-06-10 22:17:49.671242697 +0200
-@@ -80,3 +80,4 @@
-   gloo_hip_add_executable(gloo_test_hip ${GLOO_TEST_HIP_SRCS})
-   target_link_libraries(gloo_test_hip gloo_hip gtest OpenSSL::SSL OpenSSL::Crypto)
- endif()
-+add_test(NAME gloo_test COMMAND gloo_test)
---- a/cmake/Dependencies.cmake	2022-06-11 23:38:57.367089917 +0200
-+++ b/cmake/Dependencies.cmake	2022-06-11 23:48:57.690866797 +0200
-@@ -81,7 +81,7 @@
+--- a/cmake/Dependencies.cmake
++++ b/cmake/Dependencies.cmake
+@@ -81,7 +81,7 @@ if(USE_LIBUV)
        pkg_search_module(libuv REQUIRED libuv>=1.26)
        find_file(
          libuv_LIBRARY
@@ -39,7 +21,7 @@
          PATHS ${libuv_LIBDIR}
          NO_DEFAULT_PATH)
        if(NOT EXISTS ${libuv_LIBRARY})
-@@ -90,10 +90,7 @@
+@@ -90,10 +90,7 @@ if(USE_LIBUV)
      endif()
  
      add_library(uv_a INTERFACE IMPORTED)
@@ -51,19 +33,42 @@
    endif()
  endif()
  
---- a/gloo/transport/tcp/tls/pair.cc	2022-10-04 21:14:04.099150060 +0200
-+++ b/gloo/transport/tcp/tls/pair.cc	2022-10-04 21:14:22.330886610 +0200
-@@ -17,6 +17,7 @@
+--- a/gloo/CMakeLists.txt
++++ b/gloo/CMakeLists.txt
+@@ -191,14 +191,14 @@ endif()
+ # want to statically link with Gloo and not install any artifacts.
+ if(GLOO_INSTALL)
+   install(TARGETS gloo EXPORT GlooTargets
+-      DESTINATION ${CMAKE_INSTALL_PREFIX}/lib)
++      DESTINATION ${CMAKE_INSTALL_LIBDIR})
+   if(USE_CUDA)
+     install(TARGETS gloo_cuda EXPORT GlooTargets
+-        DESTINATION ${CMAKE_INSTALL_PREFIX}/lib)
++        DESTINATION ${CMAKE_INSTALL_LIBDIR})
+   endif()
+   if(USE_ROCM)
+     install(TARGETS gloo_hip EXPORT GlooTargets
+-        DESTINATION ${CMAKE_INSTALL_PREFIX}/lib)
++        DESTINATION ${CMAKE_INSTALL_LIBDIR})
+   endif()
+   install(FILES ${CMAKE_CURRENT_BINARY_DIR}/config.h
+     DESTINATION ${CMAKE_INSTALL_PREFIX}/include/gloo)
+@@ -215,11 +215,7 @@ if(GLOO_INSTALL)
+     endforeach()
+   endif()
+   if(USE_ROCM)
+-    foreach(HEADER ${GLOO_HIP_HDRS})
+-      string(REGEX MATCH "(.*)[/\\]" DIR ${HEADER})
+-      string(REGEX REPLACE "${CMAKE_CURRENT_SOURCE_DIR}" "gloo" DIR ${DIR})
+-      install(FILES ${HEADER} DESTINATION ${CMAKE_INSTALL_PREFIX}/include/${DIR})
+-    endforeach()
++    install(FILES ${GLOO_HIP_HDRS} DESTINATION ${CMAKE_INSTALL_PREFIX}/include/gloo)
+   endif()
+ endif()
  
- #include <poll.h>
- #include <cstring>
-+#include <array>
- 
- namespace gloo {
- namespace transport {
---- a/gloo/common/linux.cc	2022-10-04 21:20:21.401691929 +0200
-+++ b/gloo/common/linux.cc	2022-10-04 21:23:13.924192100 +0200
-@@ -198,7 +198,7 @@
+--- a/gloo/common/linux.cc
++++ b/gloo/common/linux.cc
+@@ -198,7 +198,7 @@ static int getInterfaceSpeedGLinkSettings(int sock, struct ifreq* ifr) {
    } ecmd;
    int rv;
  
@@ -72,7 +77,7 @@
    memset(&ecmd, 0, sizeof(ecmd));
    ecmd.req.cmd = ETHTOOL_GLINKSETTINGS;
  
-@@ -226,7 +226,7 @@
+@@ -226,7 +226,7 @@ static int getInterfaceSpeedGSet(int sock, struct ifreq* ifr) {
    struct ethtool_cmd edata;
    int rv;
  
@@ -81,3 +86,24 @@
    memset(&edata, 0, sizeof(edata));
    edata.cmd = ETHTOOL_GSET;
  
+--- a/gloo/transport/tcp/tls/pair.cc
++++ b/gloo/transport/tcp/tls/pair.cc
+@@ -17,6 +17,7 @@
+ 
+ #include <poll.h>
+ #include <cstring>
++#include <array>
+ 
+ namespace gloo {
+ namespace transport {
+--- a/cmake/Hip.cmake
++++ b/cmake/Hip.cmake
+@@ -13,7 +13,7 @@ ELSE()
+ ENDIF()
+ 
+ # Add HIP to the CMAKE Module Path
+-set(CMAKE_MODULE_PATH ${ROCM_PATH}/lib/cmake/hip ${CMAKE_MODULE_PATH})
++set(CMAKE_MODULE_PATH ${ROCM_PATH}/${CMAKE_INSTALL_LIBDIR}/cmake/hip ${CMAKE_MODULE_PATH})
+ 
+ # Disable Asserts In Code (Can't use asserts on HIP stack.)
+ ADD_DEFINITIONS(-DNDEBUG)

sci-ml/gloo/gloo-2025.06.04-r2.ebuild

@@ -2,21 +2,31 @@
 # Distributed under the terms of the GNU General Public License v2
 
 EAPI=8
-inherit cmake cuda
+ROCM_VERSION=6.3
+inherit cmake cuda flag-o-matic rocm
 
 CommitId=c7b7b022c124d9643957d9bd55f57ac59fce8fa2
 
+# Need half/bf16 headers from any pytorch
+PYTORCH_PV=2.9.0
+PYTORCH_P=pytorch-${PYTORCH_PV}
+
 DESCRIPTION="library of floating-point neural network inference operators"
 HOMEPAGE="https://github.com/facebookincubator/gloo/"
-SRC_URI="https://github.com/facebookincubator/${PN}/archive/${CommitId}.tar.gz
-	-> ${P}.tar.gz"
+SRC_URI="
+	https://github.com/facebookincubator/${PN}/archive/${CommitId}.tar.gz -> ${P}.tar.gz
+
+	cuda? ( https://github.com/pytorch/pytorch/archive/refs/tags/v${PYTORCH_PV}.tar.gz -> ${PYTORCH_P}.tar.gz )
+	rocm? ( https://github.com/pytorch/pytorch/archive/refs/tags/v${PYTORCH_PV}.tar.gz -> ${PYTORCH_P}.tar.gz )
+"
 
 S="${WORKDIR}"/${PN}-${CommitId}
+PYTORCH_S="${WORKDIR}"/${PYTORCH_P}
 
 LICENSE="MIT"
 SLOT="0"
 KEYWORDS="~amd64"
-IUSE="cuda libuv mpi redis ssl test"
+IUSE="cuda libuv mpi redis rocm ssl test"
 
 RDEPEND="
 	cuda? ( dev-util/nvidia-cuda-toolkit:= )
@@ -26,6 +36,7 @@ RDEPEND="
 		dev-db/redis
 		dev-libs/hiredis
 	)
+	rocm? ( dev-util/hip:= )
 	ssl? ( dev-libs/openssl:= )
 "
 DEPEND="${RDEPEND}
@@ -37,22 +48,21 @@ RESTRICT="test" # For some test the network is needed
 
 PATCHES=(
 	"${FILESDIR}"/${P}-gentoo.patch
-	"${FILESDIR}"/${PN}-2023.01.17-cuda.patch
 	"${FILESDIR}"/${P}-ssl3.patch
 	"${FILESDIR}"/${PN}-2023.12.03-gcc15.patch
 )
 
 src_prepare() {
-	eapply_user
+	if use cuda || use rocm; then
+		touch "${PYTORCH_S}"/torch/headeronly/macros/cmake_macros.h || die
+	fi
 	cmake_src_prepare
 	use cuda && cuda_add_sandbox
 }
 
 src_configure() {
 	local mycmakeargs=(
-		-DGLOO_USE_TORCH_DTYPES=$(usex cuda ON OFF)
 		-DBUILD_TEST=$(usex test ON OFF)
-		-DUSE_CUDA=$(usex cuda ON OFF)
 		-DGLOO_USE_CUDA_TOOLKIT=$(usex cuda ON OFF)
 		-DUSE_LIBUV=$(usex libuv ON OFF)
 		-DUSE_MPI=$(usex mpi ON OFF)
@@ -64,6 +74,21 @@ src_configure() {
 
 		mycmakeargs+=(
 			-DCMAKE_CUDA_FLAGS="$(cuda_gccdir -f | tr -d \")"
+			-DGLOO_TORCH_DIR="${PYTORCH_S}"
+			-DGLOO_USE_TORCH_DTYPES=ON
+			-DUSE_CUDA=ON
+		)
+	fi
+	if use rocm; then
+		export ROCM_PATH="${EPREFIX}/usr"
+		export GLOO_ROCM_ARCH="$(get_amdgpu_flags)"
+		mycmakeargs+=(
+			-DCMAKE_HIP_ARCHITECTURES="${GLOO_ROCM_ARCH}"
+			-DCMAKE_REQUIRE_FIND_PACKAGE_HIP=$(usex rocm ON OFF)
+			-DGLOO_TORCH_DIR="${PYTORCH_S}"
+			-DHIP_CXX_FLAGS="$(test-flags-HIPCXX ${CXXFLAGS})"
+			-DGLOO_USE_TORCH_DTYPES=ON
+			-DUSE_ROCM=ON
 		)
 	fi
 	cmake_src_configure

sci-ml/gloo/metadata.xml

@@ -11,5 +11,6 @@
 	<use>
 		<flag name="libuv">Enable libuv support</flag>
 		<flag name="redis">Enable Redis backend for storage via <pkg>dev-libs/hiredis</pkg></flag>
+		<flag name="rocm">Enable ROCm gpu computing support</flag>
 	</use>
 </pkgmetadata>