Split builds for core library and tensile modules

Build and package core library and arch dependent
tensile modules separately to parallelize the build.
Add:
Modify-CMakeLists.txt-files-to-allow-to-build-modules-independently.patch

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

Modify-CMakeLists.txt-files-to-allow-to-build-modules-independently.patch

@@ -0,0 +1,80 @@
+From: Egbert Eich <eich@suse.com>
+Date: Wed Apr 30 16:18:49 2025 +0200
+Subject: Modify CMakeLists.txt files to allow to build modules independently
+Patch-mainline: Not yet
+Git-commit: f4724507a2770b2ed5ecc633aa406ad70a675e6f
+References: 
+
+Signed-off-by: Egbert Eich <eich@suse.com>
+Signed-off-by: Egbert Eich <eich@suse.de>
+---
+ library/src/CMakeLists.txt                | 24 +++++-------------------
+ library/src/TensileInstall/CMakeLists.txt | 19 +++++++++++++++++++
+ 2 files changed, 24 insertions(+), 19 deletions(-)
+diff --git a/library/src/CMakeLists.txt b/library/src/CMakeLists.txt
+index 35342e1..efa732c 100644
+--- a/library/src/CMakeLists.txt
++++ b/library/src/CMakeLists.txt
+@@ -97,7 +97,9 @@ if( BUILD_WITH_TENSILE )
+   set_target_properties( TensileHost PROPERTIES OUTPUT_NAME rocblas-tensile CXX_EXTENSIONS NO )
+ 
+   # Tensile host depends on libs build target
+-  add_dependencies( TensileHost TENSILE_LIBRARY_TARGET )
++  if(NOT DEFINED ENV{TENSILE_SKIP_LIBRARY} OR NOT $ENV{TENSILE_SKIP_LIBRARY})
++    add_dependencies( TensileHost TENSILE_LIBRARY_TARGET )
++  endif()
+ 
+   if( ROCBLAS_SHARED_LIBS )
+     set( BUILD_SHARED_LIBS ON )
+@@ -823,24 +825,8 @@ rocm_install_targets(
+     ${CMAKE_BINARY_DIR}/${CMAKE_INSTALL_INCLUDEDIR}
+ )
+ 
+-if( BUILD_WITH_TENSILE )
+-  if (WIN32)
+-    set( ROCBLAS_TENSILE_LIBRARY_DIR "\${CPACK_PACKAGING_INSTALL_PREFIX}/bin/rocblas" CACHE PATH "path to tensile library" )
+-  else()
+-    set( ROCBLAS_TENSILE_LIBRARY_DIR "${CMAKE_INSTALL_LIBDIR}/rocblas" CACHE PATH "path to tensile library" )
+-  endif()
+-  # For ASAN package, Tensile library files(which are not shared libraries) are not required
+-  if( NOT ENABLE_ASAN_PACKAGING )
+-    if( BUILD_SHARED_LIBS )
+-      set( TENSILE_DATA_COMPONENT_NAME ${CMAKE_INSTALL_DEFAULT_COMPONENT_NAME} )
+-    else()
+-      set( TENSILE_DATA_COMPONENT_NAME devel )
+-    endif()
+-    rocm_install(
+-      DIRECTORY ${CMAKE_BINARY_DIR}/Tensile/library
+-      DESTINATION ${ROCBLAS_TENSILE_LIBRARY_DIR}
+-      COMPONENT ${TENSILE_DATA_COMPONENT_NAME}) # Use this cmake variable to be compatible with rocm-cmake 0.6 and 0.7
+-  endif()
++if(NOT DEFINED ENV{TENSILE_SKIP_LIBRARY} OR NOT $ENV{TENSILE_SKIP_LIBRARY})
++  add_subdirectory( TensileInstall )
+ endif()
+ 
+ if(NOT WIN32)
+diff --git a/library/src/TensileInstall/CMakeLists.txt b/library/src/TensileInstall/CMakeLists.txt
+new file mode 100644
+index 0000000..fa39e9f
+--- /dev/null
++++ b/library/src/TensileInstall/CMakeLists.txt
+@@ -0,0 +1,19 @@
++if( BUILD_WITH_TENSILE )
++  if (WIN32)
++    set( ROCBLAS_TENSILE_LIBRARY_DIR "\${CPACK_PACKAGING_INSTALL_PREFIX}/bin/rocblas" CACHE PATH "path to tensile library" )
++  else()
++    set( ROCBLAS_TENSILE_LIBRARY_DIR "${CMAKE_INSTALL_LIBDIR}/rocblas" CACHE PATH "path to tensile library" )
++  endif()
++  # For ASAN package, Tensile library files(which are not shared libraries) are not required
++  if( NOT ENABLE_ASAN_PACKAGING )
++    if( BUILD_SHARED_LIBS )
++      set( TENSILE_DATA_COMPONENT_NAME ${CMAKE_INSTALL_DEFAULT_COMPONENT_NAME} )
++    else()
++      set( TENSILE_DATA_COMPONENT_NAME devel )
++    endif()
++    rocm_install(
++      DIRECTORY ${CMAKE_BINARY_DIR}/Tensile/library
++      DESTINATION ${ROCBLAS_TENSILE_LIBRARY_DIR}
++      COMPONENT ${TENSILE_DATA_COMPONENT_NAME}) # Use this cmake variable to be compatible with rocm-cmake 0.6 and 0.7
++  endif()
++endif()

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 2
 %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/' -e 's/-mtls-dialect=gnu2//')
@@ -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 +58,19 @@
 %global build_tensile OFF
 %endif
 
+%if 0%{?rhel}
+%bcond_with msgpack
+%else
+%bcond_without msgpack
+%endif
+
+# Use ninja if it is available
+%if 0%{?fedora} || 0%{?suse_version}
+%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 +89,6 @@
 # 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
@@ -119,16 +133,29 @@
 
 Name:           %{rocblas_name}
 Version:        %{rocm_version}
-Release:        4%{?dist}
+Release:        5%{?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
@@ -141,26 +168,24 @@ BuildRequires:  rocm-rpm-macros-modules
 
 %if %{with tensile}
 %if 0%{?suse_version}
-%if %{suse_version} < 1699
-BuildRequires:  python3-tensile-devel
-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
+BuildRequires: %{python_module tensile-devel}
+BuildRequires: %{python_module joblib}
+%{?with_msgpack:BuildRequires:  msgpack-cxx-devel}
 %else
 BuildRequires:  python3dist(tensile)
-%if 0%{?rhel}
-%global tensile_verbose 2
-%global tensile_library_format yaml
-%else
-BuildRequires:  msgpack-devel
-%global tensile_verbose 1
-%global tensile_library_format msgpack
-%endif
+%{?with_msgpack:BuildRequires:  msgpack-devel}
 %endif # suse_version
+# OBS vm times out without console output
+%if 0%{?rhel} || 0%{?suse_version}
+%global tensile_verbose 2
+%else
+%global tensile_verbose 1
+%endif
+%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}
@@ -171,10 +196,13 @@ 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
@@ -221,9 +249,17 @@ Provides:       rocblas-devel = %{version}-%{release}
 %description devel
 %{summary}
 
-%if %{with test}
+%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} && %{without tensile_package}
 %package test
 Summary:        Tests for %{name}
+Requires:       diffutils
 Requires:       %{name}%{?_isa} = %{version}-%{release}
 
 %description test
@@ -276,21 +312,30 @@ if [ ${CORES} = 1 ]; then
     fi
 fi
 
+%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} \
     -DCMAKE_INSTALL_LIBDIR=%_libdir \
 
-%cmake_build
+%cmake_build %{?with_tensile_package:TENSILE_LIBRARY_TARGET}
 
 %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}
@@ -302,16 +347,20 @@ export LD_LIBRARY_PATH=%{_vpath_builddir}/library/src:$LD_LIBRARY_PATH
 %endif
 %endif
 %endif
+%endif
 
+%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
@@ -326,7 +375,31 @@ 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
+
 %changelog
+* Fri Aug 1 2025 Egbert Eich <eich@suse.com> - 6.4.2-5
+- Fix build and runtime dependencies of test package.
+- Enable Ninja on SUSE and Move bcond_with* settings to the top of spec file.
+- Add rpmlitrc to suppress erroneous error messages when
+  building tensile module package.
+- Enable msgpack on SUSE, restructure build requires and build settings.
+- Consoldiate Python module BuildRequires for SUSE.
+- Restructure tensile verbosity option.
+- Build and package core library and arch dependent
+  tensile modules separately to parallelize the build.
+  Add: Modify-CMakeLists.txt-files-to-allow-to-build-modules-independently.patch
+
 * Tue Jul 29 2025 Tom Rix <Tom.Rix@amd.com> - 6.4.2-4
 - Remove -mtls-dialect cflag