Consolidate python dependencies for SUSE

Signed-off-by: Egbert Eich <eich@suse.com>

rocblas.rpmlintrc

@@ -0,0 +1,3 @@
+# librocblas4-arch-.* are no shared libraries but modules loaded onto the GPU.
+addFilter("rocblas-arch-.*: W: shared-lib-without-dependency-information")
+addFilter("rocblas-arch-.*: E: shlib-fixed-dependency.")

rocblas.spec

@@ -9,6 +9,15 @@
 %global rocm_patch 0
 %global rocm_version %{rocm_release}.%{rocm_patch}
 
+%if 0%{?suse_version}
+# On SUSE build the .so module only - tensile modules are built in a separate package
+%bcond_with tensile_package
+
+%define build_tensile_separately 1
+%else
+%define build_tensile_separately 0
+%endif
+
 %global toolchain rocm
 # hipcc does not support some clang flags
 %global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/')
@@ -28,7 +37,7 @@
 %endif
 
 %bcond_with test
-%if %{with test}
+%if %{with test} && %{without tensile_package}
 %global build_test ON
 %global __brp_check_rpaths %{nil}
 %else
@@ -49,6 +58,20 @@
 %global build_tensile OFF
 %endif
 
+%if 0%{?rhel} || 0%{?sle_version} > 160000
+%bcond_with msgpack
+%else
+%bcond_without msgpack
+%endif
+
+# Use ninja if it is available
+# Ninja is available on suse but obs times out with ninja build, make doesn't
+%if 0%{?fedora}
+%bcond_without ninja
+%else
+%bcond_with ninja
+%endif
+
 %if 0%{?rhel} && 0%{?rhel} < 10
 # On CS9: /usr/bin/debugedit: Cannot handle 8-byte build ID
 %global debug_package %{nil}
@@ -67,14 +90,6 @@
 # OracleLinux 9 has a problem with it's strip not recognizing *.co's
 %global __strip %rocmllvm_bindir/llvm-strip
 
-# Use ninja if it is available
-# Ninja is available on suse but obs times out with ninja build, make doesn't
-%if 0%{?fedora}
-%bcond_without ninja
-%else
-%bcond_with ninja
-%endif
-
 %if %{with ninja}
 %global cmake_generator -G Ninja
 %else
@@ -126,18 +141,34 @@
 # $ module purge
 %bcond_with gfx950
 
+%if %{with gfx950} && %{with tensile_package}
+ExclusiveArch:  do_not_build
+%endif
+
 Name:           %{rocblas_name}
 Version:        %{rocm_version}
-Release:        8%{?dist}
+Release:        9%{?dist}
 Summary:        BLAS implementation for ROCm
 Url:            https://github.com/ROCmSoftwarePlatform/%{upstreamname}
 License:        MIT AND BSD-3-Clause
 
 Source0:        %{url}/archive/refs/tags/rocm-%{rocm_version}.tar.gz#/%{upstreamname}-%{rocm_version}.tar.gz
+Source1:        rocblas.rpmlintrc
 Patch2:         0001-fixup-install-of-tensile-output.patch
+Patch3:         Modify-CMakeLists.txt-files-to-allow-to-build-modules-independently.patch
 Patch4:         0001-offload-compress-option.patch
 Patch6:         0001-option-to-disable-roctracer-logging.patch
 
+%if 0%{build_tensile_separately}
+Requires: rocblas-tensile = %version
+%if %{with check}
+# If %check is enabled, we need to serialize the builds which will introduce
+# a circular dependency. The flag below causes OBS to ignore this.
+#
+#!BuildIgnore: %name
+BuildRequires: rocblas-tensile = %version
+%endif
+%endif
 BuildRequires:  cmake
 BuildRequires:  gcc-c++
 BuildRequires:  rocm-cmake
@@ -150,26 +181,28 @@ BuildRequires:  rocm-rpm-macros-modules
 
 %if %{with tensile}
 %if 0%{?suse_version}
+BuildRequires:  %{python_module tensile-devel}
 %if %{suse_version} < 1699
-BuildRequires:  python3-tensile-devel
+BuildRequires:  %{python_module joblib}
-BuildRequires:  python3-joblib
-%else
-BuildRequires:  python311-tensile-devel
 %endif # suse_version < 1699
 # OBS vm times out without console output
 %global tensile_verbose 2
-%global tensile_library_format yaml
+%{?with_msgpack:BuildRequires:  msgpack-cxx-devel}
-%else
+%else # ?suse_version
 BuildRequires:  python3dist(tensile)
 %if 0%{?rhel}
 %global tensile_verbose 2
-%global tensile_library_format yaml
 %else
-BuildRequires:  msgpack-devel
+%{?with_msgpack:BuildRequires:  msgpack-devel}
 %global tensile_verbose 1
 %global tensile_library_format msgpack
 %endif
 %endif # suse_version
+%if %{with msgpack}
+%global tensile_library_format msgpack
+%else
+%global tensile_library_format yaml
+%endif
 %else
 %global tensile_verbose %{nil}
 %global tensile_library_format %{nil}
@@ -180,10 +213,13 @@ BuildRequires:  pkgconfig(libzstd)
 %endif
 
 %if %{with test}
-
+%if 0%{?suse_version}
+BuildRequires: %{python_module PyYAML}
+%else
+BuildRequires:  python3dist(pyyaml)
+%endif
 BuildRequires:  blas-devel
 BuildRequires:  libomp-devel
-BuildRequires:  python3dist(pyyaml)
 BuildRequires:  rocminfo
 BuildRequires:  rocm-smi-devel
 BuildRequires:  roctracer-devel
@@ -230,9 +266,17 @@ Provides:       rocblas-devel = %{version}-%{release}
 %description devel
 %{summary}
 
+%package -n rocblas-tensile
+Summary:       ROCBlas Tensile Modules
+Requires:      %{name} = %version
+
+%description -n rocblas-tensile
+BLAS architecture modules for all AMDGPU architectures
+
 %if %{with test}
 %package test
 Summary:        Tests for %{name}
+Requires:       diffutils
 Requires:       %{name}%{?_isa} = %{version}-%{release}
 
 %description test
@@ -258,7 +302,7 @@ Conflicts:      %{name}-devel
 %description gfx950-devel
 %{summary}
 
-%if %{with test}
+%if %{with test} && %{without tensile_package}
 %package gfx950-test
 Summary:        The gfx950 rocBLAS test package
 Requires:       %{name}-gfx950%{?_isa} = %{version}-%{release}
@@ -328,6 +372,11 @@ module load rocm/gfx950
     -DCMAKE_INSTALL_LIBDIR=${ROCM_LIB}
 
 %else
+
+%if %{build_tensile_separately} && %{without tensile_package}
+export TENSILE_SKIP_LIBRARY=true
+%endif
+
 %cmake %{cmake_generator} %{cmake_config} \
     -DGPU_TARGETS=%{gpu_list} \
     -DBUILD_WITH_TENSILE=%{build_tensile} \
@@ -335,19 +384,24 @@ module load rocm/gfx950
 
 %endif
 
-%cmake_build
+%cmake_build %{?with_tensile_package:TENSILE_LIBRARY_TARGET}
 %if %{with gfx950}
 module purge
 %endif
 
 %install
+%if %{with tensile_package}
+DESTDIR=%{buildroot} /usr/bin/cmake -P build/library/src/TensileInstall/cmake_install.cmake
+%else
 %cmake_install
 
 if [ -f %{buildroot}%{_prefix}/share/doc/rocblas/LICENSE.md ]; then
     rm %{buildroot}%{_prefix}/share/doc/rocblas/LICENSE.md
 fi
+%endif
 
 %check
+%if %{without tensile_package}
 %if %{with test}
 %if %{with check}
 %if 0%{?suse_version}
@@ -359,6 +413,7 @@ export LD_LIBRARY_PATH=%{_vpath_builddir}/library/src:$LD_LIBRARY_PATH
 %endif
 %endif
 %endif
+%endif
 
 %if %{with gfx950}
 %files gfx950
@@ -379,15 +434,18 @@ export LD_LIBRARY_PATH=%{_vpath_builddir}/library/src:$LD_LIBRARY_PATH
 
 %else
 
+%if %{without tensile_package}
 %files
 %license LICENSE.md
 %{_libdir}/librocblas.so.4{,.*}
 %if %{with tensile}
+%if ! %{build_tensile_separately}
 %dir %{_libdir}/rocblas
 %dir %{_libdir}/rocblas/library
 %{_libdir}/rocblas/library/Kernels*
 %{_libdir}/rocblas/library/Tensile*
 %endif
+%endif # with tensile
 
 %files devel 
 %doc README.md
@@ -402,9 +460,28 @@ export LD_LIBRARY_PATH=%{_vpath_builddir}/library/src:$LD_LIBRARY_PATH
 %{_bindir}/rocblas*
 %endif
 
+%else # ?tensile_package
+
+%if %{with tensile}
+%files -n rocblas-tensile
+%dir %{_libdir}/rocblas
+%dir %{_libdir}/rocblas/library
+%{_libdir}/rocblas/library/Kernels*
+%{_libdir}/rocblas/library/Tensile*
+%endif
+
+%endif # ?tensile_package
 %endif # gfx950
 
 %changelog
+* Thu Jun 12 2025 Egbert Eich <eich@suse.com> - 6.4.0-9
+- Build and package core library and arch dependent
+  tensile modules separately to parallelize the build.
+- Fix build and runtime dependencies of test package.
+- Restructure spec file (move bcond_with* settings to
+  the top).
+- Add rpmlintrc for SUSE.
+
 * Wed Jun 11 2025 Tom Rix <Tom.Rix@amd.com> - 6.4.0-8
 - Remove suse check for using ldconfig