rochplmxp_dtk26.patch

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 91afcc4..6331291 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -88,7 +88,7 @@ foreach(i ${rochplmxp_device_source})
 endforeach()
 
 # HIP flags workaround while target_compile_options does not work
-list(APPEND HIP_HIPCC_FLAGS "-Wno-unused-command-line-argument -fPIE")
+list(APPEND HIP_HIPCC_FLAGS "-Wno-unused-command-line-argument -fPIE --gpu-max-threads-per-block=1024")
 list(APPEND CMAKE_HOST_FLAGS "")
 
 if (CMAKE_BUILD_TYPE STREQUAL "Debug")
@@ -100,21 +100,35 @@ else()
 endif()
 
 # GPU arch targets
-set(TARGETS "gfx900;gfx906")
-if(HIP_VERSION VERSION_GREATER_EQUAL "3.7")
-  set(TARGETS "${TARGETS};gfx908")
-endif()
-if(HIP_VERSION VERSION_GREATER_EQUAL "4.3")
-  set(TARGETS "${TARGETS};gfx90a")
-endif()
-if (HIP_VERSION VERSION_GREATER_EQUAL "5.7")
-  set(TARGETS "${TARGETS};gfx942")
+set(ARCHS "")
+if(DEFINED HPL_BUILD_ARCH AND NOT HPL_BUILD_ARCH STREQUAL "")
+  string(REPLACE "," ";" ARCHS "${HPL_BUILD_ARCH}")
+  list(TRANSFORM ARCHS STRIP)
+  list(REMOVE_DUPLICATES ARCHS)
+  message(STATUS "Using manually specified GPU targets: ${ARCHS}")
+else()
+  message(STATUS "Detecting available architecture")
+  find_program(ROCMINFO_EXECUTABLE rocminfo)
+  if(ROCMINFO_EXECUTABLE)
+    execute_process(
+      COMMAND ${ROCMINFO_EXECUTABLE}
+      OUTPUT_VARIABLE ROCMINFO_OUTPUT
+      ERROR_QUIET
+      OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+    string(REGEX MATCHALL "Name:[ \t]+gfx[0-9a-z]+" ARCH_MATCHES "${ROCMINFO_OUTPUT}")
+    string(REGEX REPLACE "Name:[ \t]+" "" ARCHS "${ARCH_MATCHES}")
+    list(REMOVE_DUPLICATES ARCHS)
+  endif()
 endif()
-if (HIP_VERSION VERSION_GREATER_EQUAL "6.5")
-  set(TARGETS "${TARGETS};gfx950")
+
+if(ARCHS STREQUAL "")
+  message(FATAL_ERROR "No GPU architectures detected via rocminfo and no BUILD_ARCH specified. Use ./install.sh --arch=gfxXXX")
 endif()
 
-foreach(target ${TARGETS})
+message(STATUS "Building for GPU architecture: ${ARCHS}")
+
+foreach(target ${ARCHS})
   list(APPEND HIP_HIPCC_FLAGS "--offload-arch=${target}")
 endforeach()
 
@@ -173,7 +187,7 @@ set_target_properties(rochplmxp PROPERTIES RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BIN
 
 set_target_properties(rochplmxp PROPERTIES LINKER_LANGUAGE CXX)
 
-set_target_properties(rochplmxp PROPERTIES HIP_ARCHITECTURES "${DEFAULT_AMDGPU_TARGETS}")
+set_target_properties(rochplmxp PROPERTIES HIP_ARCHITECTURES "${ARCHS}")
 
 # # Configure a header file to pass the rocHPL-MxP version
 configure_file("${CMAKE_CURRENT_SOURCE_DIR}/include/hplmxp_version.hpp.in"
diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
index 164d06d..041a8e2 100644
--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
@@ -109,7 +109,7 @@ if(NOT ROCM_FOUND)
   execute_process(COMMAND ${CMAKE_COMMAND} -E tar xzf ${PROJECT_EXTERN_DIR}/rocm-cmake-${rocm_cmake_tag}.zip
                   WORKING_DIRECTORY ${PROJECT_EXTERN_DIR})
 
-  find_package(ROCmCMakeBuildTools REQUIRED CONFIG PATHS ${PROJECT_EXTERN_DIR}/rocm-cmake-${rocm_cmake_tag})
+  set(CMAKE_MODULE_PATH "${PROJECT_EXTERN_DIR}/rocm-cmake-${rocm_cmake_tag}/share/rocm/cmake;${CMAKE_MODULE_PATH}")
 endif()
 
 include(ROCMSetupVersion)
diff --git a/install.sh b/install.sh
index de72a20..6f2ef05 100755
--- a/install.sh
+++ b/install.sh
@@ -18,6 +18,7 @@ function display_help()
   echo "    [--with-rocblas=<dir>] Path to rocBLAS library (Default: /opt/rocm/rocblas)"
   echo "    [--with-rocsolver=<dir>] Path to rocSOLVER library (Default: /opt/rocm/rocsolver)"
   echo "    [--with-mpi=<dir>] Path to external MPI install (Default: clone+build OpenMPI)"
+  echo "    [--arch=<archs>] Specify comma separated architecture list to build (Default: detect from rocminfo)"
   echo "    [--verbose-print] Verbose output during HPL setup (Default: true)"
   echo "    [--enable-tracing] Annotate profiler traces with rocTX markers (Default: false)"
   echo "    [--progress-report] Print progress report to terminal during HPL run (Default: true)"
@@ -33,10 +34,10 @@ supported_distro( )
   fi
 
   case "${ID}" in
-    ubuntu|centos|rhel|fedora|sles)
+    ubuntu|centos|rhel|fedora|sles|kylin|rocky)
         true
         ;;
-    *)  printf "This script is currently supported on Ubuntu, CentOS, RHEL, Fedora and SLES\n"
+    *)  printf "This script is currently supported on Ubuntu, CentOS, RHEL, Fedora, SLES, Kylin and Rocky\n"
         exit 2
         ;;
   esac
@@ -68,11 +69,11 @@ exit_with_error( )
         printf "sudo apt install -y ${library_dependencies_ubuntu[*]}\n"
         ;;
 
-      centos|rhel)
+      centos|rhel|kylin)
         printf "sudo yum -y --nogpgcheck install ${library_dependencies_centos[*]}\n"
         ;;
 
-      fedora)
+      fedora|rocky)
         printf "sudo dnf install -y ${library_dependencies_fedora[*]}\n"
         ;;
 
@@ -217,6 +218,7 @@ verbose_print=true
 enable_tracing=false
 progress_report=true
 detailed_timing=true
+arch=
 
 # #################################################
 # Parameter parsing
@@ -225,7 +227,7 @@ detailed_timing=true
 # check if we have a modern version of getopt that can handle whitespace and long parameters
 getopt -T
 if [[ $? -eq 4 ]]; then
-  GETOPT_PARSE=$(getopt --name "${0}" --longoptions help,debug,prefix:,with-rocm:,with-mpi:,with-rocblas:,with-rocsolver:,verbose-print:,enable-tracing:,progress-report:,detailed-timing: --options hg -- "$@")
+  GETOPT_PARSE=$(getopt --name "${0}" --longoptions help,debug,prefix:,with-rocm:,with-mpi:,with-rocblas:,with-rocsolver:,verbose-print:,enable-tracing:,progress-report:,detailed-timing:,arch: --options hg -- "$@")
 else
   echo "Need a new version of getopt"
   exit_with_error 1
@@ -262,6 +264,9 @@ while true; do
     --with-rocsolver)
         with_rocsolver=${2}
         shift 2 ;;
+    --arch)
+        arch=${2}
+        shift 2 ;;
     --verbose-print)
         verbose_print=${2}
         shift 2 ;;
@@ -335,11 +340,14 @@ pushd .
   if [[ "${enable_tracing}" == on || "${enable_tracing}" == true || "${enable_tracing}" == 1 || "${enable_tracing}" == enabled ]]; then
     cmake_common_options="${cmake_common_options} -DHPLMXP_TRACING=ON"
   fi
+  if [[ -n "${arch}" ]]; then
+    cmake_common_options="${cmake_common_options} -DHPL_BUILD_ARCH=${arch}"
+  fi
   shopt -u nocasematch
 
   # Build library with AMD toolchain because of existence of device kernels
   mkdir -p ${build_dir} && cd ${build_dir}
-  ${cmake_executable} ${cmake_common_options} ..
+  ${cmake_executable} --fresh ${cmake_common_options} ..
   check_exit_code 2
 
   make -j$(nproc) install
diff --git a/src/hplmxp_ptest.cpp b/src/hplmxp_ptest.cpp
index 11d0f44..e8b1eee 100644
--- a/src/hplmxp_ptest.cpp
+++ b/src/hplmxp_ptest.cpp
@@ -211,7 +211,7 @@ void HPLMXP_ptest(HPLMXP_T_test& test,
                        ctime(&current_time_end));
       }
 #ifdef HPLMXP_PROGRESS_REPORT
-      printf("Final Score:    %7.4e GFLOPS \n", Gflops);
+      printf("Final Score:    %7.9e GFLOPS \n", Gflops);
 #endif
     }
 #ifdef HPLMXP_DETAILED_TIMING
diff --git a/src/pgesv/hplmxp_pgetrf.cpp b/src/pgesv/hplmxp_pgetrf.cpp
index ccbd4c0..0230b44 100644
--- a/src/pgesv/hplmxp_pgetrf.cpp
+++ b/src/pgesv/hplmxp_pgetrf.cpp
@@ -420,7 +420,7 @@ void HPLMXP_pgetrf(HPLMXP_T_grid&         grid,
       printf("  %9.3e  |", step_gflops);
 #endif
 
-      printf("    %9.3e   \n", gflops);
+      printf("    %9.9e   \n", gflops);
     }
 #endif