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>
Fix buildinfo and compiler flags handling for SUSE
2025-08-13 11:48:10 +02:00 · 2025-08-11 16:51:42 +02:00
2 changed files with 150 additions and 5 deletions
--- a/Modify-CMakeLists.txt-files-to-allow-to-build-modules-independently.patch
+++ b/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()
--- a/rocblas.spec
+++ b/rocblas.spec
@@ -9,16 +9,32 @@
 %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//')

 %bcond_with debug
+%if 0%{?suse_version}
+%if %{without debug}
+%global build_type RELEASE
+%global build_cxxflags %(echo %{optflags} | sed -e 's/-g\\( \\|$\\)/ /')
+%endif
+%else
 %if %{with debug}
 %global build_type DEBUG
 %else
 %global build_type RELEASE
 %endif
+%endif

 %bcond_without compress
 %if %{with compress}
@@ -28,7 +44,7 @@
 %endif

 %bcond_with test
-%if %{with test}
+%if %{with test} && %{without tensile_package}
 %global build_test ON
 %global __brp_check_rpaths %{nil}
 %else
@@ -92,7 +108,7 @@
  -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld \\\
  -DCMAKE_AR=%rocmllvm_bindir/llvm-ar \\\
  -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib \\\
-  -DCMAKE_BUILD_TYPE=%{build_type} \\\
+  %{?build_type:-DCMAKE_BUILD_TYPE=%{build_type}} \\\
  -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/.. \\\
  -DCMAKE_SKIP_RPATH=ON \\\
  -DCMAKE_VERBOSE_MAKEFILE=ON \\\
@@ -124,7 +140,7 @@

 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
@@ -132,9 +148,21 @@ 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
@@ -228,7 +256,14 @@ 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
@@ -284,21 +319,31 @@ if [ ${CORES} = 1 ]; then
    fi
 fi

+%if %{build_tensile_separately} && %{without tensile_package}
+export TENSILE_SKIP_LIBRARY=true
+%endif
+
+%{?suse_version:%{?build_cxxflags:CXXFLAGS="%{build_cxxflags}"}}
 %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}
@@ -310,16 +355,20 @@ 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
@@ -334,6 +383,18 @@ 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.
@@ -343,6 +404,10 @@ export LD_LIBRARY_PATH=%{_vpath_builddir}/library/src:$LD_LIBRARY_PATH
 - Enable msgpack on SUSE, restructure build requires and build settings.
 - Consoldiate Python module BuildRequires for SUSE.
 - Restructure tensile verbosity option.
+- Fix buildinfo and compiler flags handling for SUSE.
+- 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