Initial import

Signed-off-by: Egbert Eich <eich@suse.com>
2025-07-03 21:02:37 +02:00
11 changed files with 600751 additions and 0 deletions
--- a/0001-Hardcode-cblas-as-the-blas-library.patch
+++ b/0001-Hardcode-cblas-as-the-blas-library.patch
@@ -0,0 +1,26 @@
+From e2a44fe6ad3bcf3c6df84b80d413ed09e1428e72 Mon Sep 17 00:00:00 2001
+From: Tom Rix <trix@redhat.com>
+Date: Sat, 13 Jan 2024 07:48:37 -0500
+Subject: [PATCH] Hardcode cblas as the blas library
+
+Signed-off-by: Tom Rix <trix@redhat.com>
+---
+ clients/gtest/CMakeLists.txt | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/clients/gtest/CMakeLists.txt b/clients/gtest/CMakeLists.txt
+index 9f17fb28..3df9eab4 100644
+--- a/clients/gtest/CMakeLists.txt
+++ b/clients/gtest/CMakeLists.txt
+@@ -153,7 +153,7 @@ target_include_directories( rocblas-test
+ if( BUILD_FORTRAN_CLIENTS )
+   target_link_libraries( rocblas-test PRIVATE rocblas_fortran_client )
+ endif( )
+-target_link_libraries( rocblas-test PRIVATE ${BLAS_LIBRARY} ${GTEST_BOTH_LIBRARIES} roc::rocblas )
+target_link_libraries( rocblas-test PRIVATE cblas ${GTEST_BOTH_LIBRARIES} roc::rocblas )
+ 
+ if( CUDA_FOUND )
+   target_include_directories( rocblas-test
+-- 
+2.43.0
+
--- a/0001-add-gfx1103-support-for-rocBLAS.patch
+++ b/0001-add-gfx1103-support-for-rocBLAS.patch
--- a/0001-fixup-install-of-tensile-output.patch
+++ b/0001-fixup-install-of-tensile-output.patch
@@ -0,0 +1,25 @@
+From 3d82251d51a9804c28cb84d598084fc12ca0418f Mon Sep 17 00:00:00 2001
+From: Tom Rix <trix@redhat.com>
+Date: Sat, 13 Jan 2024 14:36:01 -0500
+Subject: [PATCH] fixup install of tensile output
+
+---
+ library/src/CMakeLists.txt | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/library/src/CMakeLists.txt b/library/src/CMakeLists.txt
+index 6acedfb2..2877c6e1 100755
+--- a/library/src/CMakeLists.txt
+++ b/library/src/CMakeLists.txt
+@@ -591,7 +591,7 @@ if( BUILD_WITH_TENSILE )
+   if (WIN32)
+     set( ROCBLAS_TENSILE_LIBRARY_DIR "\${CPACK_PACKAGING_INSTALL_PREFIX}/bin/rocblas" CACHE PATH "path to tensile library" )
+   else()
+-    set( ROCBLAS_TENSILE_LIBRARY_DIR "\${CPACK_PACKAGING_INSTALL_PREFIX}${CMAKE_INSTALL_LIBDIR}/rocblas" CACHE PATH "path to tensile library" )
+    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 )
+-- 
+2.43.0
+
--- a/0001-offload-compress-option.patch
+++ b/0001-offload-compress-option.patch
@@ -0,0 +1,44 @@
+From 2966285dc09ca9c7e6b95c5212a2d5bd46ab8376 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Fri, 27 Sep 2024 05:40:14 -0700
+Subject: [PATCH] offload compress option
+
+Try out --offload-compress
+
+Signed-off-by: Tom Rix <Tom.Rix@amd.com>
+---
+ cmake/build-options.cmake | 2 ++
+ library/CMakeLists.txt    | 4 ++++
+ 2 files changed, 6 insertions(+)
+
+diff --git a/cmake/build-options.cmake b/cmake/build-options.cmake
+index cb35e72fb157..26d04c0aefdb 100755
+--- a/cmake/build-options.cmake
+++ b/cmake/build-options.cmake
+@@ -36,6 +36,8 @@ option( BUILD_SHARED_LIBS "Build rocBLAS as a shared library" ON )
+ # library without tensile to allow for rapid iteration without GEMM functionality
+ option( BUILD_WITH_TENSILE "Build full functionality which requires tensile?" ON )
+ 
+option( BUILD_OFFLOAD_COMPRESS "Build with offload compress?" OFF )
+
+ include(clients/cmake/client-build-options.cmake)
+ 
+ if (WIN32)
+diff --git a/library/CMakeLists.txt b/library/CMakeLists.txt
+index 90a75dd394d2..0386a3058d45 100755
+--- a/library/CMakeLists.txt
+++ b/library/CMakeLists.txt
+@@ -66,6 +66,10 @@ function( rocblas_library_settings lib_target_ )
+   # Do not allow Variable Length Arrays (use unique_ptr instead)
+   target_compile_options( ${lib_target_} PRIVATE -Werror=vla )
+ 
+  if ( BUILD_OFFLOAD_COMPRESS )
+    target_compile_options( ${lib_target_} PRIVATE --offload-compress )
+  endif ()
+  
+   target_compile_definitions( ${lib_target_} PRIVATE ROCM_USE_FLOAT16 ROCBLAS_INTERNAL_API ROCBLAS_BETA_FEATURES_API )
+ 
+   # both libraries will use rocblas_EXPORTS
+-- 
+2.46.0
+
--- a/0001-option-to-disable-roctracer-logging.patch
+++ b/0001-option-to-disable-roctracer-logging.patch
@@ -0,0 +1,82 @@
+diff --git a/library/CMakeLists.txt b/library/CMakeLists.txt
+index 762580d7..f02d62e6 100644
+--- a/library/CMakeLists.txt
+++ b/library/CMakeLists.txt
+@@ -80,7 +80,7 @@ function( rocblas_library_settings lib_target_ )
+     target_link_libraries( ${lib_target_} PRIVATE hip::device )
+   else()
+     target_link_libraries( ${lib_target_} PRIVATE hip::device -lstdc++fs --rtlib=compiler-rt --unwindlib=libgcc)
+-    if (BUILD_SHARED_LIBS)
+    if (BUILD_SHARED_LIBS AND NOT DISABLE_ROCTRACER)
+       target_link_libraries(${lib_target_} PRIVATE -lroctx64)
+     endif()
+   endif()
+diff --git a/library/src/CMakeLists.txt b/library/src/CMakeLists.txt
+index 5deab888..f96c7e3e 100644
+--- a/library/src/CMakeLists.txt
+++ b/library/src/CMakeLists.txt
+@@ -28,6 +28,10 @@
+ # package_targets is used as a list of install target
+ set( package_targets rocblas )
+ 
+# we want to decrease the number of build deps for EPEL packages so patching in an option
+# to disable its use
+
+option( DISABLE_ROCTRACER "Disable use of rocTRACER for logging" ON )
+ 
+ # Set up Tensile Dependency
+ if( BUILD_WITH_TENSILE )
+@@ -688,7 +692,12 @@ if(BUILD_WITH_HIPBLASLT)
+   find_package( hipblaslt ${HIPBLASLT_VERSION} REQUIRED CONFIG PATHS ${hipblaslt_path} ${ROCM_PATH})
+ endif()
+ 
+-if( NOT BUILD_SHARED_LIBS )
+if ( DISABLE_ROCTRACER )
+    message( "roctracer is disabled for this build" )
+    target_compile_definitions( rocblas PRIVATE DISABLE_ROCTRACER )
+endif()
+
+if( NOT DISABLE_ROCTRACER AND NOT BUILD_SHARED_LIBS )
+   target_compile_definitions( rocblas PRIVATE ROCBLAS_STATIC_LIB )
+ endif()
+ 
+diff --git a/library/src/include/logging.hpp b/library/src/include/logging.hpp
+index 352e4e39..d222879a 100644
+--- a/library/src/include/logging.hpp
+++ b/library/src/include/logging.hpp
+@@ -40,7 +40,7 @@
+ #include <unordered_map>
+ #include <utility>
+ 
+-#if !defined(ROCBLAS_STATIC_LIB) && !defined(WIN32)
+#if !defined(DISABLE_ROCTRACER) && !defined(ROCBLAS_STATIC_LIB) && !defined(WIN32)
+ #include <roctracer/roctx.h>
+ #endif
+ 
+@@ -453,7 +453,7 @@ class Logger
+ public:
+     Logger() = default;
+ 
+-#if !defined(ROCBLAS_STATIC_LIB) && !defined(WIN32)
+#if !defined(DISABLE_ROCTRACER) && !defined(ROCBLAS_STATIC_LIB) && !defined(WIN32)
+     void log_range(const std::string& name)
+     {
+         if(!m_active)
+@@ -472,7 +472,7 @@ public:
+         // ((os << sep << std::forward<Ts>(xs)), ...);
+         (void)(int[]){(os << sep << std::forward<Ts>(xs), 0)...};
+ 
+-#if !defined(ROCBLAS_STATIC_LIB) && !defined(WIN32)
+#if !defined(DISABLE_ROCTRACER) && !defined(ROCBLAS_STATIC_LIB) && !defined(WIN32)
+         log_range(os.str());
+ #endif
+         os << std::endl;
+@@ -527,7 +527,7 @@ public:
+ 
+     ~Logger()
+     {
+-#if !defined(ROCBLAS_STATIC_LIB) && !defined(WIN32)
+#if !defined(DISABLE_ROCTRACER) && !defined(ROCBLAS_STATIC_LIB) && !defined(WIN32)
+         if(m_active)
+         {
+             roctxRangePop();
--- a/0001-prepare-rocblas-cmake-for-fedora.patch
+++ b/0001-prepare-rocblas-cmake-for-fedora.patch
@@ -0,0 +1,26 @@
+From aba3a118fd32f415e63b4e24555b8df98e89292d Mon Sep 17 00:00:00 2001
+From: Tom Rix <trix@redhat.com>
+Date: Tue, 3 Oct 2023 10:37:12 -0700
+Subject: [PATCH] prepare rocblas cmake for fedora
+
+Signed-off-by: Tom Rix <trix@redhat.com>
+---
+ clients/CMakeLists.txt | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/clients/CMakeLists.txt b/clients/CMakeLists.txt
+index da44cef0..6d39c1f3 100755
+--- a/clients/CMakeLists.txt
+++ b/clients/CMakeLists.txt
+@@ -131,7 +131,7 @@ if( BUILD_CLIENTS_BENCHMARKS OR BUILD_CLIENTS_TESTS)
+         message(STATUS "Linking BLIS LIB: ${BLAS_LIBRARY}")
+       endif()
+     else()
+-      set( BLAS_LIBRARY "blas" )
+      set( BLAS_LIBRARY "cblas" )
+     endif()
+   else() # WIN32
+     set( BLAS_INCLUDE_DIR ${OPENBLAS_DIR}/include CACHE PATH "OpenBLAS library include path" )
+-- 
+2.41.0
+
--- 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/14
+++ b/14
@@ -0,0 +1,14 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<constraints>
+  <hardware>
+    <disk>
+      <size unit="G">60</size>
+    </disk>
+    <physicalmemory>
+      <size unit="G">16</size>
+    </physicalmemory>
+    <processors>4</processors>
+    <jobs>4</jobs>
+  </hardware>
+  <hostlabel exclude="true">SLOW_CPU</hostlabel>
+</constraints>
--- a/rocBLAS-6.4.0.tar.gz
+++ b/rocBLAS-6.4.0.tar.gz
--- a/rocblas.rpmlintrc
+++ b/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.")
--- a/rocblas.spec
+++ b/rocblas.spec
@@ -0,0 +1,551 @@
+%if 0%{?suse_version}
+%global rocblas_name librocblas4
+%else
+%global rocblas_name rocblas
+%endif
+
+%global upstreamname rocBLAS
+%global rocm_release 6.4
+%global rocm_patch 0
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+%if 0%{?suse_version}
+# On SUSE build tensile modules - the .so module is built in a separate package
+%bcond_without 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/')
+
+%bcond_with debug
+%if %{with debug}
+%global build_type DEBUG
+%else
+%global build_type RELEASE
+%endif
+
+%bcond_without compress
+%if %{with compress}
+%global build_compress ON
+%else
+%global build_compress OFF
+%endif
+
+%bcond_with test
+%if %{with test} && %{without tensile_package}
+%global build_test ON
+%global __brp_check_rpaths %{nil}
+%else
+%global build_test OFF
+%endif
+
+# Option to test suite for testing on real HW:
+# May have to set gpu under test with
+# export HIP_VISIBLE_DEVICES=<num> - 0, 1 etc.
+%bcond_with check
+
+# Tensile in 6.4 does not support generics
+# https://github.com/ROCm/Tensile/issues/2124
+%bcond_without tensile
+%if %{with tensile}
+%global build_tensile ON
+%else
+%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}
+%endif
+
+# Compression type and level for source/binary package payloads.
+#  "w7T0.xzdio"	xz level 7 using %%{getncpus} threads
+%global _source_payload w7T0.xzdio
+%global _binary_payload w7T0.xzdio
+
+# SUSE/OSB times out because -O is added to the make args
+# This accumulates all the output from the long running tensile
+# jobs.
+%global _make_output_sync %{nil}
+
+# OracleLinux 9 has a problem with it's strip not recognizing *.co's
+%global __strip %rocmllvm_bindir/llvm-strip
+
+%if %{with ninja}
+%global cmake_generator -G Ninja
+%else
+%global cmake_generator %{nil}
+%endif
+
+%global cmake_config \\\
+  -DCMAKE_CXX_COMPILER=hipcc \\\
+  -DCMAKE_C_COMPILER=hipcc \\\
+  -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld \\\
+  -DCMAKE_AR=%rocmllvm_bindir/llvm-ar \\\
+  -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib \\\
+  -DCMAKE_BUILD_TYPE=%{build_type} \\\
+  -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/.. \\\
+  -DCMAKE_SKIP_RPATH=ON \\\
+  -DCMAKE_VERBOSE_MAKEFILE=ON \\\
+  -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF \\\
+  -DROCM_SYMLINK_LIBS=OFF \\\
+  -DHIP_PLATFORM=amd \\\
+  -DBUILD_CLIENTS_BENCHMARKS=%{build_test} \\\
+  -DBUILD_CLIENTS_TESTS=%{build_test} \\\
+  -DBUILD_CLIENTS_TESTS_OPENMP=OFF \\\
+  -DBUILD_FORTRAN_CLIENTS=OFF \\\
+  -DBLAS_LIBRARY=cblas \\\
+  -DBUILD_OFFLOAD_COMPRESS=%{build_compress} \\\
+  -DBUILD_WITH_HIPBLASLT=OFF \\\
+  -DTensile_COMPILER=hipcc \\\
+  -DTensile_CPU_THREADS=${CORES} \\\
+  -DTensile_LIBRARY_FORMAT=%{tensile_library_format} \\\
+  -DTensile_VERBOSE=%{tensile_verbose} \\\
+  -DTensile_DIR=${TP}/cmake \\\
+  -DDISABLE_ROCTRACER=ON \\\
+  -DBUILD_WITH_PIP=OFF
+
+%bcond_with generic
+%global rocm_gpu_list_generic "gfx9-generic;gfx9-4-generic;gfx10-1-generic;gfx10-3-generic;gfx11-generic;gfx12-generic"
+%if %{with generic}
+%global gpu_list %{rocm_gpu_list_generic}
+%else
+%global gpu_list %{rocm_gpu_list_default}
+%endif
+
+# gfx950 is an experimental target
+# Enabling will short circuit the normal build.
+# There is no check support.
+# To use do
+# $ module load rocm/gfx950
+#     <do stuff>
+# $ module purge
+%bcond_with gfx950
+
+%if %{with gfx950} && %{with tensile_package}
+ExclusiveArch:  do_not_build
+%endif
+
+Name:           %{rocblas_name}
+Version:        %{rocm_version}
+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
+BuildRequires:  rocm-comgr-devel
+BuildRequires:  rocm-compilersupport-macros
+BuildRequires:  rocm-hip-devel
+BuildRequires:  rocm-runtime-devel
+BuildRequires:  rocm-rpm-macros
+BuildRequires:  rocm-rpm-macros-modules
+
+%if %{with tensile}
+%if 0%{?suse_version}
+BuildRequires:  %{python_module tensile-devel}
+%if %{suse_version} < 1699
+BuildRequires:  %{python_module joblib}
+%endif # suse_version < 1699
+# OBS vm times out without console output
+%global tensile_verbose 2
+%{?with_msgpack:BuildRequires:  msgpack-cxx-devel}
+%else # ?suse_version
+BuildRequires:  python3dist(tensile)
+%if 0%{?rhel}
+%global tensile_verbose 2
+%else
+%{?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}
+%endif # tensile
+
+%if %{with compress}
+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:  rocminfo
+BuildRequires:  rocm-smi-devel
+BuildRequires:  roctracer-devel
+
+%if 0%{?suse_version}
+BuildRequires:  cblas-devel
+BuildRequires:  gcc-fortran
+BuildRequires:  gtest
+%else
+BuildRequires:  gtest-devel
+%endif
+
+%endif
+
+%if %{with ninja}
+%if 0%{?fedora} || 0%{?rhel}
+BuildRequires:  ninja-build
+%endif
+%if 0%{?suse_version}
+BuildRequires:  ninja
+%define __builder ninja
+%endif
+%endif
+
+Provides:       rocblas = %{version}-%{release}
+
+# Only x86_64 works right now:
+ExclusiveArch:  x86_64
+
+%description
+rocBLAS is the AMD library for Basic Linear Algebra Subprograms
+(BLAS) on the ROCm platform. It is implemented in the HIP
+programming language and optimized for AMD GPUs.
+
+%post -p /sbin/ldconfig
+%postun -p /sbin/ldconfig
+
+%package devel
+Summary:        Libraries and headers for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+Requires:       cmake(hip)
+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
+%{summary}
+%endif
+
+%if %{with gfx950}
+
+%package gfx950
+Summary:        The gfx950 rocBLAS package
+Provides:       rocblas-gfx950 = %{version}-%{release}
+Conflicts:      %{name}
+
+%description gfx950
+%{summary}
+
+%package gfx950-devel
+Summary:        The gfx950 rocBLAS development package
+Requires:       %{name}-gfx950%{?_isa} = %{version}-%{release}
+Provides:       rocblas-gfx950-devel = %{version}-%{release}
+Conflicts:      %{name}-devel
+
+%description gfx950-devel
+%{summary}
+
+%if %{with test} && %{without tensile_package}
+%package gfx950-test
+Summary:        The gfx950 rocBLAS test package
+Requires:       %{name}-gfx950%{?_isa} = %{version}-%{release}
+Conflicts:      %{name}-test
+
+%description gfx950-test
+%{summary}
+
+%endif # gfx950-test
+%endif # gfx950
+
+%prep
+%autosetup -p1 -n %{upstreamname}-rocm-%{version}
+sed -i -e 's@set( BLAS_LIBRARY "blas" )@set( BLAS_LIBRARY "cblas" )@' clients/CMakeLists.txt
+sed -i -e 's@target_link_libraries( rocblas-test PRIVATE ${BLAS_LIBRARY} ${GTEST_BOTH_LIBRARIES} roc::rocblas )@target_link_libraries( rocblas-test PRIVATE cblas ${GTEST_BOTH_LIBRARIES} roc::rocblas )@' clients/gtest/CMakeLists.txt
+
+# no git in this build
+sed -i -e 's@find_package(Git REQUIRED)@find_package(Git)@' library/CMakeLists.txt
+
+# On Tumbleweed Q2,2025
+# /usr/include/gtest/internal/gtest-port.h:279:2: error: C++ versions less than C++14 are not supported.
+#   279 | #error C++ versions less than C++14 are not supported.
+# Convert the c++11's to c++14
+sed -i -e 's@CXX_STANDARD 11@CXX_STANDARD 14@' clients/samples/CMakeLists.txt
+
+%if 0%{?suse_version}
+# Suse's libgfortran.so for gcc 14 is here
+# /usr/lib64/gcc/x86_64-suse-linux/14/libgfortran.so
+# Without adding this path with -L, it isn't found, but thankfully it isn't really needed
+sed -i -e 's@list( APPEND COMMON_LINK_LIBS "-lgfortran")@#list( APPEND COMMON_LINK_LIBS "-lgfortran")@' clients/{benchmarks,gtest}/CMakeLists.txt
+%endif
+
+%build
+
+# With compat llvm the system clang is wrong
+CLANG_PATH=`hipconfig --hipclangpath`
+export TENSILE_ROCM_ASSEMBLER_PATH=${CLANG_PATH}/clang++
+export TENSILE_ROCM_OFFLOAD_BUNDLER_PATH=${CLANG_PATH}/clang-offload-bundler
+# Work around problem with koji's ld
+export HIPCC_LINK_FLAGS_APPEND=-fuse-ld=lld
+
+%if %{with tensile}
+TP=`/usr/bin/TensileGetPath`
+%endif
+
+CORES=`lscpu | grep 'Core(s)' | awk '{ print $4 }'`
+if [ ${CORES}x = x ]; then
+    CORES=1
+fi
+# Try again..
+if [ ${CORES} = 1 ]; then
+    CORES=`lscpu | grep '^CPU(s)' | awk '{ print $2 }'`
+    if [ ${CORES}x = x ]; then
+        CORES=4
+    fi
+fi
+
+%if %{with gfx950}
+
+module load rocm/gfx950
+
+%cmake %{cmake_generator} %{cmake_config} \
+    -DGPU_TARGETS=${ROCM_GPUS} \
+    -DBUILD_WITH_TENSILE=OFF \
+    -DCMAKE_INSTALL_BINDIR=${ROCM_BIN} \
+    -DCMAKE_INSTALL_INCLUDEDIR=${ROCM_INCLUDE} \
+    -DCMAKE_INSTALL_LIBDIR=${ROCM_LIB}
+
+%else
+
+%if %{build_tensile_separately} && %{without tensile_package}
+export TENSILE_SKIP_LIBRARY=true
+%endif
+
+%cmake %{cmake_generator} %{cmake_config} \
+    -DGPU_TARGETS=%{gpu_list} \
+    -DBUILD_WITH_TENSILE=%{build_tensile} \
+    -DCMAKE_INSTALL_LIBDIR=%_libdir \
+
+%endif
+
+%cmake_build %{?with_tensile_package:TENSILE_LIBRARY_TARGET}
+%if %{with gfx950}
+module purge
+%endif
+
+%install
+%if %{with tensile_package}
+DESTDIR=%{buildroot} /usr/bin/cmake -P build/library/src/TensileInstall/cmake_install.cmake
+%else
+%cmake_install
+
+if [ -f %{buildroot}%{_prefix}/share/doc/rocblas/LICENSE.md ]; then
+    rm %{buildroot}%{_prefix}/share/doc/rocblas/LICENSE.md
+fi
+%endif
+
+%check
+%if %{without tensile_package}
+%if %{with test}
+%if %{with check}
+%if 0%{?suse_version}
+export LD_LIBRARY_PATH=%{__builddir}/library/src:$LD_LIBRARY_PATH
+%{__builddir}/clients/staging/rocblas-test --gtest_brief=1
+%else
+export LD_LIBRARY_PATH=%{_vpath_builddir}/library/src:$LD_LIBRARY_PATH
+%{_vpath_builddir}/clients/staging/rocblas-test --gtest_brief=1
+%endif
+%endif
+%endif
+%endif
+
+%if %{with gfx950}
+%files gfx950
+%license LICENSE.md
+%{_libdir}/rocm/gfx950/lib/librocblas.so.4{,.*}
+
+%files gfx950-devel
+%dir %{_libdir}/rocm/gfx950/include/rocblas
+%dir %{_libdir}/rocm/gfx950/lib/cmake/rocblas
+%{_libdir}/rocm/gfx950/include/rocblas/rocblas_module.f90
+%{_libdir}/rocm/gfx950/lib/librocblas.so
+%{_libdir}/rocm/gfx950/lib/cmake/rocblas/*.cmake
+
+%if %{with test}
+%files gfx950-test
+%{_libdir}/rocm/gfx950/bin/rocblas*
+%endif
+
+%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
+%dir %{_libdir}/cmake/rocblas
+%dir %{_includedir}/rocblas
+%{_includedir}/rocblas/*
+%{_libdir}/cmake/rocblas/*.cmake
+%{_libdir}/librocblas.so
+
+%if %{with test}
+%files test
+%{_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
+
+* Sun May 11 2025 Tom Rix <Tom.Rix@amd.com> - 6.4.0-7
+- Add experimental gfx950
+
+* Tue May 6 2025 Tom Rix <Tom.Rix@amd.com> - 6.4.0-6
+- disable roctracer for everyone
+
+* Tue Apr 29 2025 Tim Flink <tflink@fedoraproject.org> - 6.4.0-5
+- add patch for option to disable roctracer logging
+- disable roctracer logging for rhel builds
+- allow for builds on rhel with ninja
+
+* Tue Apr 29 2025 Tom Rix <Tom.Rix@amd.com> - 6.4.0-4
+- Improve testing for suse
+
+* Sat Apr 26 2025 Tom Rix <Tom.Rix@amd.com> - 6.4.0-3
+- Add generic gpus
+
+* Wed Apr 23 2025 Tom Rix <Tom.Rix@amd.com> - 6.4.0-2
+- Use joblib on sle 15.6 and 16.0
+
+* Fri Apr 18 2025 Tom Rix <Tom.Rix@amd.com> - 6.4.0-1
+- Update to 6.4.0
+
+* Thu Apr 10 2025 Tom Rix <Tom.Rix@amd.com> - 6.3.0-12
+- Reenable ninja
+
+* Fri Apr 4 2025 Tom Rix <Tom.Rix@amd.com> - 6.3.0-11
+- Use rocm-llvm strip
+
+* Thu Feb 27 2025 Cristian Le <git@lecris.dev> - 6.3.0-10
+- Add hip requirement to devel package
+
+* Thu Feb 27 2025 Tom Rix <Tom.Rix@amd.com> - 6.3.0-9
+- Enable tensile for RHEL
+
+* Wed Feb 26 2025 Tom Rix <Tom.Rix@amd.com> - 6.3.0-8
+- Enable tensile for SUSE
+
+* Sun Feb 23 2025 Tom Rix <Tom.Rix@amd.com> - 6.3.0-7
+- Use tensile verbosity to avoid OSB timeout
+
+* Wed Feb 19 2025 Tom Rix <Tom.Rix@amd.com> - 6.3.0-6
+- Use tensile cmake from the python location
+
+* Tue Feb 11 2025 Tom Rix <Tom.Rix@amd.com> - 6.3.0-5
+- Remove multibuild
+- Fix SLE 15.6
+
+* Sat Jan 18 2025 Tom Rix <Tom.Rix@amd.com> - 6.3.0-4
+- multithread rpm compress
+
+* Tue Jan 14 2025 Tom Rix <Tom.Rix@amd.com> - 6.3.0-3
+- build requires gcc-c++
+
+* Fri Dec 20 2024 Tom Rix <Tom.Rix@amd.com> - 6.3.0-2
+- Build type should be release
+
+* Fri Dec 6 2024 Tom Rix <Tom.Rix@amd.com> - 6.3.0-1
+- Update to 6.3
+
+* Sun Nov 10 2024 Tom Rix <Tom.Rix@amd.com> - 6.2.1-1
+- Stub for tumbleweed
+
+