Merge pull request #1279 from xsacha/develop

CMake improvements

.travis.yml

@@ -119,6 +119,34 @@ jobs:
         - TARGET_BOX=LINUX64_MUSL
         - BTYPE="BINARY=64 NO_AFFINITY=1 USE_OPENMP=0 NO_LAPACK=0 TARGET=core2"
 
+    - &test-cmake
+      stage: test
+      compiler: clang
+      addons:
+        apt:
+          packages:
+            - gfortran
+            - cmake
+      dist: trusty
+      sudo: true
+      before_script:
+        - COMMON_ARGS="-DTARGET=NEHALEM -DNUM_THREADS=32"
+      script:
+        - set -e
+        - mkdir build
+        - CONFIG=Release
+        - cmake -Bbuild -H. $CMAKE_ARGS $COMMON_ARGS -DCMAKE_BUILD_TYPE=$CONFIG
+        - cmake --build build --config $CONFIG -- -j2
+      env:
+        - CMAKE=1
+    - <<: *test-cmake
+      env:
+        - CMAKE=1 CMAKE_ARGS="-DNOFORTRAN=1"
+    - <<: *test-cmake
+      compiler: gcc
+      env:
+        - CMAKE=1
+
 # whitelist
 branches:
   only:

CMakeLists.txt

@@ -3,15 +3,12 @@
 ##
 
 cmake_minimum_required(VERSION 2.8.5)
-project(OpenBLAS)
+project(OpenBLAS C ASM)
 set(OpenBLAS_MAJOR_VERSION 0)
 set(OpenBLAS_MINOR_VERSION 3)
 set(OpenBLAS_PATCH_VERSION 0.dev)
 set(OpenBLAS_VERSION "${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.${OpenBLAS_PATCH_VERSION}")
 
-enable_language(ASM)
-enable_language(C)
-
 # Adhere to GNU filesystem layout conventions
 include(GNUInstallDirs)
 
@@ -26,7 +23,6 @@ if(MSVC)
 option(BUILD_WITHOUT_LAPACK "Without LAPACK and LAPACKE (Only BLAS or CBLAS)" ON)
 endif()
 option(BUILD_WITHOUT_CBLAS "Without CBLAS" OFF)
-option(BUILD_DEBUG "Build Debug Version" OFF)
 option(DYNAMIC_ARCH "Build with DYNAMIC_ARCH" OFF)
 #######
 if(BUILD_WITHOUT_LAPACK)
@@ -34,22 +30,6 @@ set(NO_LAPACK 1)
 set(NO_LAPACKE 1)
 endif()
 
-if(CMAKE_CONFIGURATION_TYPES) # multiconfig generator?
-        set(CMAKE_CONFIGURATION_TYPES "Debug;Release" CACHE STRING "" FORCE)
-        set(CMAKE_BUILD_TYPE
-	        Debug Debug
-		Release Release
-	)
-else()
-	if( NOT CMAKE_BUILD_TYPE )
-		if(BUILD_DEBUG)
-			set(CMAKE_BUILD_TYPE Debug)
-		else()
-			set(CMAKE_BUILD_TYPE Release)
-		endif()
-	endif()
-endif()
-
 if(BUILD_WITHOUT_CBLAS)
 set(NO_CBLAS 1)
 endif()
@@ -107,19 +87,10 @@ if (BUILD_COMPLEX16)
   list(APPEND FLOAT_TYPES "ZCOMPLEX") # defines COMPLEX and DOUBLE
 endif ()
 
-set(SUBDIRS_ALL ${SUBDIRS} test ctest utest exports benchmark ../laswp ../bench)
-
-# all :: libs netlib tests shared
-
-# libs :
 if (NOT DEFINED CORE OR "${CORE}" STREQUAL "UNKNOWN")
   message(FATAL_ERROR "Detecting CPU failed. Please set TARGET explicitly, e.g. make TARGET=your_cpu_target. Please read README for details.")
 endif ()
 
-if (${NO_STATIC} AND ${NO_SHARED})
-  message(FATAL_ERROR "Neither static nor shared are enabled.")
-endif ()
-
 #Set default output directory
 set( CMAKE_LIBRARY_OUTPUT_DIRECTORY ${PROJECT_BINARY_DIR}/lib)
 set( CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${PROJECT_BINARY_DIR}/lib)
@@ -141,26 +112,20 @@ endforeach ()
 # Not using add_subdirectory here because lapack-netlib already has its own CMakeLists.txt. Instead include a cmake script with the sources we want.
 if (NOT NOFORTRAN AND NOT NO_LAPACK)
   include("${PROJECT_SOURCE_DIR}/cmake/lapack.cmake")
-if (NOT NO_LAPACKE)
-  include("${PROJECT_SOURCE_DIR}/cmake/lapacke.cmake")
-endif ()
+  if (NOT NO_LAPACKE)
+    include("${PROJECT_SOURCE_DIR}/cmake/lapacke.cmake")
+  endif ()
 endif ()
 
 # Only generate .def for dll on MSVC and always produce pdb files for debug and release
 if(MSVC)
-set(OpenBLAS_DEF_FILE "${PROJECT_BINARY_DIR}/openblas.def")
-set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /Zi")
-set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /Zi")
-set(CMAKE_SHARED_LINKER_FLAGS_RELEASE "${CMAKE_SHARED_LINKER_FLAGS_RELEASE} /DEBUG /OPT:REF /OPT:ICF")
+  if (${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION} LESS 3.4)
+    set(OpenBLAS_DEF_FILE "${PROJECT_BINARY_DIR}/openblas.def")
+  endif()
+  set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /Zi")
+  set(CMAKE_SHARED_LINKER_FLAGS_RELEASE "${CMAKE_SHARED_LINKER_FLAGS_RELEASE} /DEBUG /OPT:REF /OPT:ICF")
 endif()
 
-#ifeq ($(DYNAMIC_ARCH), 1)
-#	@$(MAKE) -C kernel commonlibs || exit 1
-#	@for d in $(DYNAMIC_CORE) ; \
-#	do  $(MAKE) GOTOBLAS_MAKEFILE= -C kernel TARGET_CORE=$$d kernel || exit 1 ;\
-#	done
-#	@echo DYNAMIC_ARCH=1 >> Makefile.conf_last
-#endif
 if (${DYNAMIC_ARCH})
   add_subdirectory(kernel)
   foreach(TARGET_CORE ${DYNAMIC_CORE})
@@ -169,11 +134,24 @@ if (${DYNAMIC_ARCH})
   endforeach()
 endif ()
 
+# Only build shared libs for MSVC
+if (MSVC)
+  set(BUILD_SHARED_LIBS ON)
+endif()
 
-# add objects to the openblas lib
-add_library(${OpenBLAS_LIBNAME} SHARED ${LA_SOURCES} ${LAPACKE_SOURCES} ${TARGET_OBJS} ${OpenBLAS_DEF_FILE})
 
-include("${PROJECT_SOURCE_DIR}/cmake/export.cmake")
+# add objects to the openblas lib
+add_library(${OpenBLAS_LIBNAME} ${LA_SOURCES} ${LAPACKE_SOURCES} ${TARGET_OBJS} ${OpenBLAS_DEF_FILE})
+
+# Handle MSVC exports
+if(MSVC AND BUILD_SHARED_LIBS)
+  if (${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION} LESS 3.4)
+    include("${PROJECT_SOURCE_DIR}/cmake/export.cmake")
+  else()
+    # Creates verbose .def file (51KB vs 18KB)
+    set_target_properties(${OpenBLAS_LIBNAME} PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS true)
+  endif()
+endif()
 
 # Set output for libopenblas
 set_target_properties( ${OpenBLAS_LIBNAME} PROPERTIES RUNTIME_OUTPUT_DIRECTORY ${PROJECT_BINARY_DIR}/lib)
@@ -188,25 +166,28 @@ foreach (OUTPUTCONFIG ${CMAKE_CONFIGURATION_TYPES})
 endforeach()
 
 enable_testing()
-add_subdirectory(utest)
-
-if (NOT MSVC)
-	#only build shared library for MSVC
-
-	add_library(${OpenBLAS_LIBNAME}_static STATIC ${LA_SOURCES} ${LAPACKE_SOURCES} ${TARGET_OBJS})
-	set_target_properties(${OpenBLAS_LIBNAME}_static PROPERTIES OUTPUT_NAME ${OpenBLAS_LIBNAME})
-	set_target_properties(${OpenBLAS_LIBNAME}_static PROPERTIES CLEAN_DIRECT_OUTPUT 1)
 
-	if(SMP)
-	target_link_libraries(${OpenBLAS_LIBNAME} pthread)
-	target_link_libraries(${OpenBLAS_LIBNAME}_static pthread)
+if (USE_THREAD)
+  # Add threading library to linker
+  find_package(Threads)
+  if (THREADS_HAVE_PTHREAD_ARG)
+    set_property(TARGET ${OpenBLAS_LIBNAME} PROPERTY COMPILE_OPTIONS "-pthread")
+    set_property(TARGET ${OpenBLAS_LIBNAME} PROPERTY INTERFACE_COMPILE_OPTIONS "-pthread")
+  endif()
+  target_link_libraries(${OpenBLAS_LIBNAME} ${CMAKE_THREAD_LIBS_INIT})
 endif()
 
-#build test and ctest
-add_subdirectory(test)
-if(NOT NO_CBLAS)
-add_subdirectory(ctest)
+if (MSVC OR NOT NOFORTRAN)
+  # Broken without fortran on unix
+  add_subdirectory(utest)
 endif()
+
+if (NOT MSVC AND NOT NOFORTRAN)
+  # Build test and ctest
+  add_subdirectory(test)
+  if(NOT NO_CBLAS)
+    add_subdirectory(ctest)
+  endif()
 endif()
 
 set_target_properties(${OpenBLAS_LIBNAME} PROPERTIES 
@@ -214,53 +195,42 @@ set_target_properties(${OpenBLAS_LIBNAME} PROPERTIES
   SOVERSION ${OpenBLAS_MAJOR_VERSION}
 )
 
-
-# TODO: Why is the config saved here? Is this necessary with CMake?
-#Save the config files for installation
-#	@cp Makefile.conf Makefile.conf_last
-#	@cp config.h config_last.h
-#ifdef QUAD_PRECISION
-#	@echo "#define QUAD_PRECISION">> config_last.h
-#endif
-#ifeq ($(EXPRECISION), 1)
-#	@echo "#define EXPRECISION">> config_last.h
-#endif
-###
-
-#ifdef USE_THREAD
-#	@echo USE_THREAD=$(USE_THREAD) >>  Makefile.conf_last
-#endif
-#	@touch lib.grd
-
 # Install project
 
 # Install libraries
 install(TARGETS ${OpenBLAS_LIBNAME}
 	RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
 	ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
-	LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR} )
-
-# Install include files
-	set (GENCONFIG_BIN ${CMAKE_BINARY_DIR}/gen_config_h${CMAKE_EXECUTABLE_SUFFIX})
-
-	execute_process(COMMAND ${GENCONFIG_BIN}
-	                        ${CMAKE_CURRENT_SOURCE_DIR}/config.h
-	                        ${CMAKE_CURRENT_SOURCE_DIR}/openblas_config_template.h
-	                OUTPUT_VARIABLE OPENBLAS_CONFIG_H_CONTENTS)
-
-	file(WRITE ${CMAKE_BINARY_DIR}/openblas_config.tmp "${OPENBLAS_CONFIG_H_CONTENTS}")
-	configure_file(${CMAKE_BINARY_DIR}/openblas_config.tmp ${CMAKE_BINARY_DIR}/openblas_config.h COPYONLY)
-
-	install (FILES ${CMAKE_BINARY_DIR}/openblas_config.h DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
-
-	message(STATUS "Generating f77blas.h in ${CMAKE_INSTALL_INCLUDEDIR}")
-
-	file(WRITE ${CMAKE_BINARY_DIR}/f77blas.h "")
-	file(APPEND ${CMAKE_BINARY_DIR}/f77blas.h "#ifndef OPENBLAS_F77BLAS_H\n#define OPENBLAS_F77BLAS_H\n#include \"openblas_config.h\"\n")
-	file(READ ${CMAKE_CURRENT_SOURCE_DIR}/common_interface.h COMMON_INTERFACE_H_CONTENTS)
-	file(APPEND ${CMAKE_BINARY_DIR}/f77blas.h "${COMMON_INTERFACE_H_CONTENTS}")
-	file(APPEND ${CMAKE_BINARY_DIR}/f77blas.h "#endif")
-	install (FILES ${CMAKE_BINARY_DIR}/f77blas.h DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
+  LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR} )
+  
+message(STATUS "Generating openblas_config.h in ${CMAKE_INSTALL_INCLUDEDIR}")
+
+set(OPENBLAS_CONFIG_H ${CMAKE_BINARY_DIR}/openblas_config.h)
+file(WRITE  ${OPENBLAS_CONFIG_H} "#ifndef OPENBLAS_CONFIG_H\n")
+file(APPEND ${OPENBLAS_CONFIG_H} "#define OPENBLAS_CONFIG_H\n")
+file(STRINGS ${PROJECT_BINARY_DIR}/config.h __lines)
+foreach(line ${__lines})
+  string(REPLACE "#define " "" line ${line})
+  file(APPEND ${OPENBLAS_CONFIG_H} "#define OPENBLAS_${line}\n")
+endforeach()
+file(APPEND ${OPENBLAS_CONFIG_H} "#define OPENBLAS_VERSION \"OpenBLAS ${OpenBLAS_VERSION}\"\n")
+file(READ   ${CMAKE_CURRENT_SOURCE_DIR}/openblas_config_template.h OPENBLAS_CONFIG_TEMPLATE_H_CONTENTS)
+file(APPEND ${OPENBLAS_CONFIG_H} "${OPENBLAS_CONFIG_TEMPLATE_H_CONTENTS}\n")
+file(APPEND ${OPENBLAS_CONFIG_H} "#endif /* OPENBLAS_CONFIG_H */\n")
+install (FILES ${OPENBLAS_CONFIG_H} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
+
+if(NOT NOFORTRAN)
+  message(STATUS "Generating f77blas.h in ${CMAKE_INSTALL_INCLUDEDIR}")
+
+  set(F77BLAS_H ${CMAKE_BINARY_DIR}/f77blas.h)
+  file(WRITE  ${F77BLAS_H} "#ifndef OPENBLAS_F77BLAS_H\n")
+  file(APPEND ${F77BLAS_H} "#define OPENBLAS_F77BLAS_H\n")
+  file(APPEND ${F77BLAS_H} "#include \"openblas_config.h\"\n")
+  file(READ ${CMAKE_CURRENT_SOURCE_DIR}/common_interface.h COMMON_INTERFACE_H_CONTENTS)
+  file(APPEND ${F77BLAS_H} "${COMMON_INTERFACE_H_CONTENTS}\n")
+  file(APPEND ${F77BLAS_H} "#endif")
+  install (FILES ${F77BLAS_H} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
+endif()
 
 if(NOT NO_CBLAS)
 	message (STATUS "Generating cblas.h in ${CMAKE_INSTALL_INCLUDEDIR}")
@@ -282,10 +252,6 @@ if(NOT NO_LAPACKE)
 	)
 	install (FILES ${CMAKE_BINARY_DIR}/lapacke_mangling.h DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
 endif()
- 
-if(NOT MSVC)
-	install (TARGETS ${OpenBLAS_LIBNAME}_static DESTINATION ${CMAKE_INSTALL_LIBDIR})
-endif()
 
 include(FindPkgConfig QUIET)
 if(PKG_CONFIG_FOUND)

cmake/arch.cmake

@@ -3,9 +3,9 @@
 ## Description: Ported from portion of OpenBLAS/Makefile.system
 ##              Sets various variables based on architecture.
 
-if (${ARCH} STREQUAL "x86" OR ${ARCH} STREQUAL "x86_64")
+if (X86 OR X86_64)
 
-  if (${ARCH} STREQUAL "x86")
+  if (X86)
     if (NOT BINARY)
       set(NO_BINARY_MODE 1)
     endif ()
@@ -33,41 +33,22 @@ if (${CMAKE_C_COMPILER_ID} STREQUAL "Intel")
 endif ()
 
 if (USE_OPENMP)
-
-  if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU" OR ${CMAKE_C_COMPILER_ID} STREQUAL "LSB")
-    set(CCOMMON_OPT "${CCOMMON_OPT} -fopenmp")
-  endif ()
-
-  if (${CMAKE_C_COMPILER_ID} STREQUAL "Clang")
-    message(WARNING "Clang doesn't support OpenMP yet.")
-    set(CCOMMON_OPT "${CCOMMON_OPT} -fopenmp")
-  endif ()
-
-  if (${CMAKE_C_COMPILER_ID} STREQUAL "Intel")
-    set(CCOMMON_OPT "${CCOMMON_OPT} -openmp")
-  endif ()
-
-  if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI")
-    set(CCOMMON_OPT "${CCOMMON_OPT} -mp")
-  endif ()
-
-  if (${CMAKE_C_COMPILER_ID} STREQUAL "OPEN64")
-    set(CCOMMON_OPT "${CCOMMON_OPT} -mp")
-    set(CEXTRALIB "${CEXTRALIB} -lstdc++")
-  endif ()
-
-  if (${CMAKE_C_COMPILER_ID} STREQUAL "PATHSCALE")
-    set(CCOMMON_OPT "${CCOMMON_OPT} -mp")
-  endif ()
+  # USE_SIMPLE_THREADED_LEVEL3 = 1
+  # NO_AFFINITY = 1
+  find_package(OpenMP REQUIRED)
+  if (OpenMP_FOUND)
+    set(CCOMMON_OPT "${CCOMMON_OPT} ${OpenMP_C_FLAGS} -DUSE_OPENMP")
+    set(FCOMMON_OPT "${FCOMMON_OPT} ${OpenMP_Fortran_FLAGS}")
+  endif()
 endif ()
 
 
 if (DYNAMIC_ARCH)
-  if (${ARCH} STREQUAL "x86")
+  if (X86)
     set(DYNAMIC_CORE KATMAI COPPERMINE NORTHWOOD PRESCOTT BANIAS CORE2 PENRYN DUNNINGTON NEHALEM ATHLON OPTERON OPTERON_SSE3 BARCELONA BOBCAT ATOM NANO)
   endif ()
 
-  if (${ARCH} STREQUAL "x86_64")
+  if (X86_64)
     set(DYNAMIC_CORE PRESCOTT CORE2 PENRYN DUNNINGTON NEHALEM OPTERON OPTERON_SSE3 BARCELONA BOBCAT ATOM NANO)
     if (NOT NO_AVX)
       set(DYNAMIC_CORE ${DYNAMIC_CORE} SANDYBRIDGE BULLDOZER PILEDRIVER STEAMROLLER EXCAVATOR)
@@ -94,7 +75,7 @@ if (${ARCH} STREQUAL "ia64")
   endif ()
 endif ()
 
-if (${ARCH} STREQUAL "mips64")
+if (MIPS64)
   set(NO_BINARY_MODE 1)
 endif ()
 
@@ -103,12 +84,12 @@ if (${ARCH} STREQUAL "alpha")
   set(BINARY_DEFINED 1)
 endif ()
 
-if (${ARCH} STREQUAL "arm")
+if (ARM)
   set(NO_BINARY_MODE 1)
   set(BINARY_DEFINED 1)
 endif ()
 
-if (${ARCH} STREQUAL "arm64")
+if (ARM64)
   set(NO_BINARY_MODE 1)
   set(BINARY_DEFINED 1)
 endif ()

cmake/c_check.cmake

@@ -1,101 +0,0 @@
-##
-## Author: Hank Anderson <hank@statease.com>
-## Description: Ported from the OpenBLAS/c_check perl script.
-##              This is triggered by prebuild.cmake and runs before any of the code is built.
-##              Creates config.h and Makefile.conf.
-
-# CMake vars set by this file:
-# OSNAME (use CMAKE_SYSTEM_NAME)
-# ARCH
-# C_COMPILER (use CMAKE_C_COMPILER)
-# BINARY32
-# BINARY64
-# FU
-# CROSS_SUFFIX
-# CROSS
-# CEXTRALIB
-
-# Defines set by this file:
-# OS_
-# ARCH_
-# C_
-# __32BIT__
-# __64BIT__
-# FUNDERSCORE
-# PTHREAD_CREATE_FUNC
-
-# N.B. c_check (and ctest.c) is not cross-platform, so instead try to use CMake variables.
-set(FU "")
-if(APPLE)
-set(FU "_")
-elseif(MSVC AND ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
-set(FU "")
-elseif(MSVC)
-set(FU "_")
-elseif(UNIX)
-set(FU "")
-endif()
-
-# Convert CMake vars into the format that OpenBLAS expects
-string(TOUPPER ${CMAKE_SYSTEM_NAME} HOST_OS)
-if (${HOST_OS} STREQUAL "WINDOWS")
-  set(HOST_OS WINNT)
-endif ()
-
-# added by hpa - check size of void ptr to detect 64-bit compile
-if (NOT DEFINED BINARY)
-  set(BINARY 32)
-  if (CMAKE_SIZEOF_VOID_P EQUAL 8)
-    set(BINARY 64)
-  endif ()
-endif ()
-
-if (BINARY EQUAL 64)
-  set(BINARY64 1)
-else ()
-  set(BINARY32 1)
-endif ()
-
-# CMake docs define these:
-# CMAKE_SYSTEM_PROCESSOR - The name of the CPU CMake is building for.
-# CMAKE_HOST_SYSTEM_PROCESSOR - The name of the CPU CMake is running on.
-#
-# TODO: CMAKE_SYSTEM_PROCESSOR doesn't seem to be correct - instead get it from the compiler a la c_check
-set(ARCH ${CMAKE_SYSTEM_PROCESSOR} CACHE STRING "Target Architecture")
-
-if (${ARCH} STREQUAL "AMD64")
-  set(ARCH "x86_64")
-endif ()
-
-# If you are using a 32-bit compiler on a 64-bit system CMAKE_SYSTEM_PROCESSOR will be wrong
-if (${ARCH} STREQUAL "x86_64" AND BINARY EQUAL 32)
-  set(ARCH x86)
-endif ()
-
-if (${ARCH} STREQUAL "X86")
-  set(ARCH x86)
-endif ()
-
-if (${ARCH} MATCHES "ppc")
-  set(ARCH power)
-endif ()
-
-set(COMPILER_ID ${CMAKE_CXX_COMPILER_ID})
-if (${COMPILER_ID} STREQUAL "GNU")
-  set(COMPILER_ID "GCC")
-endif ()
-
-string(TOUPPER ${ARCH} UC_ARCH)
-
-file(WRITE ${TARGET_CONF_TEMP}
-  "#define OS_${HOST_OS}\t1\n"
-  "#define ARCH_${UC_ARCH}\t1\n"
-  "#define C_${COMPILER_ID}\t1\n"
-  "#define __${BINARY}BIT__\t1\n"
-  "#define FUNDERSCORE\t${FU}\n")
-
-if (${HOST_OS} STREQUAL "WINDOWSSTORE")
-  file(APPEND ${TARGET_CONF_TEMP}
-    "#define OS_WINNT\t1\n")
-endif ()
-

cmake/cc.cmake

@@ -15,7 +15,7 @@ if (${CMAKE_C_COMPILER} STREQUAL "GNU" OR ${CMAKE_C_COMPILER} STREQUAL "LSB" OR
 
   if (NO_BINARY_MODE)
 
-    if (${ARCH} STREQUAL "mips64")
+    if (MIPS64)
       if (BINARY64)
         set(CCOMMON_OPT "${CCOMMON_OPT} -mabi=64")
       else ()
@@ -24,17 +24,12 @@ if (${CMAKE_C_COMPILER} STREQUAL "GNU" OR ${CMAKE_C_COMPILER} STREQUAL "LSB" OR
       set(BINARY_DEFINED 1)
     endif ()
 
-    if (${CORE} STREQUAL "LOONGSON3A")
+    if (${CORE} STREQUAL "LOONGSON3A" OR ${CORE} STREQUAL "LOONGSON3B")
       set(CCOMMON_OPT "${CCOMMON_OPT} -march=mips64")
       set(FCOMMON_OPT "${FCOMMON_OPT} -march=mips64")
     endif ()
 
-    if (${CORE} STREQUAL "LOONGSON3B")
-      set(CCOMMON_OPT "${CCOMMON_OPT} -march=mips64")
-      set(FCOMMON_OPT "${FCOMMON_OPT} -march=mips64")
-    endif ()
-
-    if (${OSNAME} STREQUAL "AIX")
+    if (CMAKE_SYSTEM_NAME STREQUAL "AIX")
       set(BINARY_DEFINED 1)
     endif ()
   endif ()
@@ -66,7 +61,7 @@ endif ()
 
 if (${CMAKE_C_COMPILER} STREQUAL "OPEN64")
 
-  if (${ARCH} STREQUAL "mips64")
+  if (MIPS64)
 
     if (NOT BINARY64)
       set(CCOMMON_OPT "${CCOMMON_OPT} -n32")
@@ -94,10 +89,10 @@ endif ()
 
 if (${CMAKE_C_COMPILER} STREQUAL "SUN")
   set(CCOMMON_OPT "${CCOMMON_OPT} -w")
-  if (${ARCH} STREQUAL "x86")
+  if (X86)
     set(CCOMMON_OPT "${CCOMMON_OPT} -m32")
   else ()
-    set(FCOMMON_OPT "${FCOMMON_OPT} -m64")
+    set(CCOMMON_OPT "${CCOMMON_OPT} -m64")
   endif ()
 endif ()
 

cmake/f_check.cmake

@@ -34,11 +34,7 @@ else()
 endif()
 
 if (NOT ONLY_CBLAS)
-  # N.B. f_check is not cross-platform, so instead try to use CMake variables
   # run f_check (appends to TARGET files)
-#  message(STATUS "Running f_check...")
-#  execute_process(COMMAND perl f_check ${TARGET_MAKE} ${TARGET_CONF} ${CMAKE_Fortran_COMPILER}
-#    WORKING_DIRECTORY ${PROJECT_SOURCE_DIR})
 
   # TODO: detect whether underscore needed, set #defines and BU appropriately - use try_compile
   # TODO: set FEXTRALIB flags a la f_check?

cmake/fc.cmake

@@ -50,7 +50,7 @@ if (${F_COMPILER} STREQUAL "GFORTRAN")
     set(EXTRALIB "{EXTRALIB} -lgfortran")
   endif ()
   if (NO_BINARY_MODE)
-    if (${ARCH} STREQUAL "mips64")
+    if (MIPS64)
       if (BINARY64)
         set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=64")
       else ()
@@ -130,7 +130,7 @@ if (${F_COMPILER} STREQUAL "PATHSCALE")
     endif ()
   endif ()
 
-  if (NOT ${ARCH} STREQUAL "mips64")
+  if (NOT MIPS64)
     if (NOT BINARY64)
       set(FCOMMON_OPT "${FCOMMON_OPT} -m32")
     else ()
@@ -158,7 +158,7 @@ if (${F_COMPILER} STREQUAL "OPEN64")
     endif ()
   endif ()
 
-  if (${ARCH} STREQUAL "mips64")
+  if (MIPS64)
 
     if (NOT BINARY64)
       set(FCOMMON_OPT "${FCOMMON_OPT} -n32")
@@ -189,7 +189,7 @@ endif ()
 
 if (${F_COMPILER} STREQUAL "SUN")
   set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_SUN")
-  if (${ARCH} STREQUAL "x86")
+  if (X86)
     set(FCOMMON_OPT "${FCOMMON_OPT} -m32")
   else ()
     set(FCOMMON_OPT "${FCOMMON_OPT} -m64")

cmake/os.cmake

@@ -3,19 +3,6 @@
 ## Description: Ported from portion of OpenBLAS/Makefile.system
 ##              Detects the OS and sets appropriate variables.
 
-if (${CMAKE_SYSTEM_NAME} STREQUAL "Darwin")
-  set(ENV{MACOSX_DEPLOYMENT_TARGET} "10.2") # TODO: should be exported as an env var
-  set(MD5SUM "md5 -r")
-endif ()
-
-if (${CMAKE_SYSTEM_NAME} STREQUAL "FreeBSD")
-  set(MD5SUM "md5 -r")
-endif ()
-
-if (${CMAKE_SYSTEM_NAME} STREQUAL "NetBSD")
-  set(MD5SUM "md5 -n")
-endif ()
-
 if (${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
   set(EXTRALIB "${EXTRALIB} -lm")
   set(NO_EXPRECISION 1)
@@ -56,7 +43,7 @@ if (${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
 
   # Ensure the correct stack alignment on Win32
   # http://permalink.gmane.org/gmane.comp.lib.openblas.general/97
-  if (${ARCH} STREQUAL "x86")
+  if (X86)
     if (NOT MSVC AND NOT ${CMAKE_C_COMPILER_ID} STREQUAL "Clang")
       set(CCOMMON_OPT "${CCOMMON_OPT} -mincoming-stack-boundary=2")
     endif ()
@@ -78,7 +65,7 @@ if (CYGWIN)
 endif ()
 
 if (NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Windows" AND NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Interix" AND NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Android")
-  if (SMP)
+  if (USE_THREAD)
     set(EXTRALIB "${EXTRALIB} -lpthread")
   endif ()
 endif ()
@@ -88,7 +75,7 @@ if (QUAD_PRECISION)
   set(NO_EXPRECISION 1)
 endif ()
 
-if (${ARCH} STREQUAL "x86")
+if (X86)
   set(NO_EXPRECISION 1)
 endif ()
 

cmake/prebuild.cmake

@@ -37,123 +37,208 @@
 
 # CPUIDEMU = ../../cpuid/table.o
 
+
 if (DEFINED CPUIDEMU)
   set(EXFLAGS "-DCPUIDEMU -DVENDOR=99")
 endif ()
 
-if (DEFINED TARGET_CORE)
+if (BUILD_KERNEL)
   # set the C flags for just this file
   set(GETARCH2_FLAGS "-DBUILD_KERNEL")
-  set(TARGET_MAKE "Makefile_kernel.conf")
   set(TARGET_CONF "config_kernel.h")
   set(TARGET_CONF_DIR ${PROJECT_BINARY_DIR}/kernel_config/${TARGET_CORE})
 else()
-  set(TARGET_MAKE "Makefile.conf")
   set(TARGET_CONF "config.h")
   set(TARGET_CONF_DIR ${PROJECT_BINARY_DIR})
 endif ()
 
 set(TARGET_CONF_TEMP "${PROJECT_BINARY_DIR}/${TARGET_CONF}.tmp")
-include("${PROJECT_SOURCE_DIR}/cmake/c_check.cmake")
 
-if (NOT NOFORTRAN)
-  include("${PROJECT_SOURCE_DIR}/cmake/f_check.cmake")
+# c_check
+set(FU "")
+if (APPLE OR (MSVC AND NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang"))
+  set(FU "_")
+endif()
+
+set(COMPILER_ID ${CMAKE_C_COMPILER_ID})
+if (${COMPILER_ID} STREQUAL "GNU")
+  set(COMPILER_ID "GCC")
 endif ()
 
-# compile getarch
-set(GETARCH_SRC
-  ${PROJECT_SOURCE_DIR}/getarch.c
-  ${CPUIDEMO}
-)
+string(TOUPPER ${ARCH} UC_ARCH)
 
-if ("${CMAKE_C_COMPILER_ID}" STREQUAL "MSVC")
-  #Use generic for MSVC now
-  message("MSVC")
-  set(GETARCH_FLAGS ${GETARCH_FLAGS} -DFORCE_GENERIC)
-else()
-  list(APPEND GETARCH_SRC ${PROJECT_SOURCE_DIR}/cpuid.S)
+file(WRITE ${TARGET_CONF_TEMP}
+  "#define OS_${HOST_OS}\t1\n"
+  "#define ARCH_${UC_ARCH}\t1\n"
+  "#define C_${COMPILER_ID}\t1\n"
+  "#define __${BINARY}BIT__\t1\n"
+  "#define FUNDERSCORE\t${FU}\n")
+
+if (${HOST_OS} STREQUAL "WINDOWSSTORE")
+  file(APPEND ${TARGET_CONF_TEMP}
+    "#define OS_WINNT\t1\n")
 endif ()
 
-if ("${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
-  # disable WindowsStore strict CRT checks
-  set(GETARCH_FLAGS ${GETARCH_FLAGS} -D_CRT_SECURE_NO_WARNINGS)
+# f_check
+if (NOT NOFORTRAN)
+  include("${PROJECT_SOURCE_DIR}/cmake/f_check.cmake")
 endif ()
 
-set(GETARCH_DIR "${PROJECT_BINARY_DIR}/getarch_build")
-set(GETARCH_BIN "getarch${CMAKE_EXECUTABLE_SUFFIX}")
-file(MAKE_DIRECTORY ${GETARCH_DIR})
-configure_file(${TARGET_CONF_TEMP} ${GETARCH_DIR}/${TARGET_CONF} COPYONLY)
-if (NOT "${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
-  try_compile(GETARCH_RESULT ${GETARCH_DIR}
-    SOURCES ${GETARCH_SRC}
-    COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} -I${GETARCH_DIR} -I${PROJECT_SOURCE_DIR} -I${PROJECT_BINARY_DIR}
-    OUTPUT_VARIABLE GETARCH_LOG
-    COPY_FILE ${PROJECT_BINARY_DIR}/${GETARCH_BIN}
+# Cannot run getarch on target if we are cross-compiling
+if (DEFINED CORE AND CMAKE_CROSSCOMPILING)
+  # Write to config as getarch would
+
+  # TODO: Set up defines that getarch sets up based on every other target
+  # Perhaps this should be inside a different file as it grows larger
+  file(APPEND ${TARGET_CONF_TEMP}
+    "#define ${CORE}\n"
+    "#define CHAR_CORENAME \"${CORE}\"\n")
+  if ("${CORE}" STREQUAL "ARMV7")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_DATA_SIZE\t65536\n"
+      "#define L1_DATA_LINESIZE\t32\n"
+      "#define L2_SIZE\t512488\n"
+      "#define L2_LINESIZE\t32\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define L2_ASSOCIATIVE\t4\n"
+      "#define HAVE_VFPV3\n"
+      "#define HAVE_VFP\n")
+    set(SGEMM_UNROLL_M 4)
+    set(SGEMM_UNROLL_N 4)
+    set(DGEMM_UNROLL_M 4)
+    set(DGEMM_UNROLL_N 4)
+  elseif ("${CORE}" STREQUAL "ARMV8")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_DATA_SIZE\t32768\n"
+      "#define L1_DATA_LINESIZE\t64\n"
+      "#define L2_SIZE\t262144\n"
+      "#define L2_LINESIZE\t64\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define L2_ASSOCIATIVE\t32\n")
+    set(SGEMM_UNROLL_M 4)
+    set(SGEMM_UNROLL_N 4)
+  elseif ("${CORE}" STREQUAL "CORTEXA57")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_CODE_SIZE\t49152\n"
+      "#define L1_CODE_LINESIZE\t64\n"
+      "#define L1_CODE_ASSOCIATIVE\t3\n"
+      "#define L1_DATA_SIZE\t32768\n"
+      "#define L1_DATA_LINESIZE\t64\n"
+      "#define L1_DATA_ASSOCIATIVE\t2\n"
+      "#define L2_SIZE\t2097152\n"
+      "#define L2_LINESIZE\t64\n"
+      "#define L2_ASSOCIATIVE\t16\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define HAVE_VFPV4\n"
+      "#define HAVE_VFPV3\n"
+      "#define HAVE_VFP\n"
+      "#define HAVE_NEON\n")
+    set(SGEMM_UNROLL_M 16)
+    set(SGEMM_UNROLL_N 4)
+    set(DGEMM_UNROLL_M 8)
+    set(DGEMM_UNROLL_N 4)
+    set(CGEMM_UNROLL_M 8)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 8)
+    set(ZGEMM_UNROLL_N 4)
+  endif()
+
+  # Or should this actually be NUM_CORES?
+  if (${NUM_THREADS} GREATER 0)
+    file(APPEND ${TARGET_CONF_TEMP} "#define NUM_CORES\t${NUM_THREADS}\n")
+  endif()
+
+  # GetArch_2nd
+  foreach(float_char S;D;Q;C;Z;X)
+    if (NOT DEFINED ${float_char}GEMM_UNROLL_M)
+      set(${float_char}GEMM_UNROLL_M 2)
+    endif()
+    if (NOT DEFINED ${float_char}GEMM_UNROLL_N)
+      set(${float_char}GEMM_UNROLL_N 2)
+    endif()
+  endforeach()
+  file(APPEND ${TARGET_CONF_TEMP}
+    "#define GEMM_MULTITHREAD_THRESHOLD\t${GEMM_MULTITHREAD_THRESHOLD}\n")
+  # Move to where gen_config_h would place it
+  file(RENAME ${TARGET_CONF_TEMP} "${TARGET_CONF_DIR}/${TARGET_CONF}")  
+
+else(NOT CMAKE_CROSSCOMPILING)
+  # compile getarch
+  set(GETARCH_SRC
+    ${PROJECT_SOURCE_DIR}/getarch.c
+    ${CPUIDEMU}
   )
 
-  if (NOT ${GETARCH_RESULT})
-    MESSAGE(FATAL_ERROR "Compiling getarch failed ${GETARCH_LOG}")
+  if ("${CMAKE_C_COMPILER_ID}" STREQUAL "MSVC")
+    #Use generic for MSVC now
+    message("MSVC")
+    set(GETARCH_FLAGS ${GETARCH_FLAGS} -DFORCE_GENERIC)
+  else()
+    list(APPEND GETARCH_SRC ${PROJECT_SOURCE_DIR}/cpuid.S)
   endif ()
-endif ()
-message(STATUS "Running getarch")
-
-# use the cmake binary w/ the -E param to run a shell command in a cross-platform way
-execute_process(COMMAND ${PROJECT_BINARY_DIR}/${GETARCH_BIN} 0 OUTPUT_VARIABLE GETARCH_MAKE_OUT)
-execute_process(COMMAND ${PROJECT_BINARY_DIR}/${GETARCH_BIN} 1 OUTPUT_VARIABLE GETARCH_CONF_OUT)
-
-message(STATUS "GETARCH results:\n${GETARCH_MAKE_OUT}")
-
-# append config data from getarch to the TARGET file and read in CMake vars
-file(APPEND ${TARGET_CONF_TEMP} ${GETARCH_CONF_OUT})
-ParseGetArchVars(${GETARCH_MAKE_OUT})
-
-set(GETARCH2_DIR "${PROJECT_BINARY_DIR}/getarch2_build")
-set(GETARCH2_BIN "getarch_2nd${CMAKE_EXECUTABLE_SUFFIX}")
-file(MAKE_DIRECTORY ${GETARCH2_DIR})
-configure_file(${TARGET_CONF_TEMP} ${GETARCH2_DIR}/${TARGET_CONF} COPYONLY)
-if (NOT "${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
-  try_compile(GETARCH2_RESULT ${GETARCH2_DIR}
-    SOURCES ${PROJECT_SOURCE_DIR}/getarch_2nd.c
-    COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} ${GETARCH2_FLAGS} -I${GETARCH2_DIR} -I${PROJECT_SOURCE_DIR} -I${PROJECT_BINARY_DIR}
-    OUTPUT_VARIABLE GETARCH2_LOG
-    COPY_FILE ${PROJECT_BINARY_DIR}/${GETARCH2_BIN}
-  )
 
-  if (NOT ${GETARCH2_RESULT})
-    MESSAGE(FATAL_ERROR "Compiling getarch_2nd failed ${GETARCH2_LOG}")
+  if ("${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
+    # disable WindowsStore strict CRT checks
+    set(GETARCH_FLAGS ${GETARCH_FLAGS} -D_CRT_SECURE_NO_WARNINGS)
   endif ()
-endif ()
 
-# use the cmake binary w/ the -E param to run a shell command in a cross-platform way
-execute_process(COMMAND ${PROJECT_BINARY_DIR}/${GETARCH2_BIN} 0 OUTPUT_VARIABLE GETARCH2_MAKE_OUT)
-execute_process(COMMAND ${PROJECT_BINARY_DIR}/${GETARCH2_BIN} 1 OUTPUT_VARIABLE GETARCH2_CONF_OUT)
+  set(GETARCH_DIR "${PROJECT_BINARY_DIR}/getarch_build")
+  set(GETARCH_BIN "getarch${CMAKE_EXECUTABLE_SUFFIX}")
+  file(MAKE_DIRECTORY ${GETARCH_DIR})
+  configure_file(${TARGET_CONF_TEMP} ${GETARCH_DIR}/${TARGET_CONF} COPYONLY)
+  if (NOT "${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
+    try_compile(GETARCH_RESULT ${GETARCH_DIR}
+      SOURCES ${GETARCH_SRC}
+      COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} -I${GETARCH_DIR} -I${PROJECT_SOURCE_DIR} -I${PROJECT_BINARY_DIR}
+      OUTPUT_VARIABLE GETARCH_LOG
+      COPY_FILE ${PROJECT_BINARY_DIR}/${GETARCH_BIN}
+    )
+  
+    if (NOT ${GETARCH_RESULT})
+      MESSAGE(FATAL_ERROR "Compiling getarch failed ${GETARCH_LOG}")
+    endif ()
+  endif ()
+  message(STATUS "Running getarch")
+
+  # use the cmake binary w/ the -E param to run a shell command in a cross-platform way
+  execute_process(COMMAND ${PROJECT_BINARY_DIR}/${GETARCH_BIN} 0 OUTPUT_VARIABLE GETARCH_MAKE_OUT)
+  execute_process(COMMAND ${PROJECT_BINARY_DIR}/${GETARCH_BIN} 1 OUTPUT_VARIABLE GETARCH_CONF_OUT)
+
+  message(STATUS "GETARCH results:\n${GETARCH_MAKE_OUT}")
+
+  # append config data from getarch to the TARGET file and read in CMake vars
+  file(APPEND ${TARGET_CONF_TEMP} ${GETARCH_CONF_OUT})
+  ParseGetArchVars(${GETARCH_MAKE_OUT})
+
+  set(GETARCH2_DIR "${PROJECT_BINARY_DIR}/getarch2_build")
+  set(GETARCH2_BIN "getarch_2nd${CMAKE_EXECUTABLE_SUFFIX}")
+  file(MAKE_DIRECTORY ${GETARCH2_DIR})
+  configure_file(${TARGET_CONF_TEMP} ${GETARCH2_DIR}/${TARGET_CONF} COPYONLY)
+  if (NOT "${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
+    try_compile(GETARCH2_RESULT ${GETARCH2_DIR}
+      SOURCES ${PROJECT_SOURCE_DIR}/getarch_2nd.c
+      COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} ${GETARCH2_FLAGS} -I${GETARCH2_DIR} -I${PROJECT_SOURCE_DIR} -I${PROJECT_BINARY_DIR}
+      OUTPUT_VARIABLE GETARCH2_LOG
+      COPY_FILE ${PROJECT_BINARY_DIR}/${GETARCH2_BIN}
+    )
+
+    if (NOT ${GETARCH2_RESULT})
+      MESSAGE(FATAL_ERROR "Compiling getarch_2nd failed ${GETARCH2_LOG}")
+    endif ()
+  endif ()
 
-# append config data from getarch_2nd to the TARGET file and read in CMake vars
-file(APPEND ${TARGET_CONF_TEMP} ${GETARCH2_CONF_OUT})
+  # use the cmake binary w/ the -E param to run a shell command in a cross-platform way
+  execute_process(COMMAND ${PROJECT_BINARY_DIR}/${GETARCH2_BIN} 0 OUTPUT_VARIABLE GETARCH2_MAKE_OUT)
+  execute_process(COMMAND ${PROJECT_BINARY_DIR}/${GETARCH2_BIN} 1 OUTPUT_VARIABLE GETARCH2_CONF_OUT)
 
-if (${BUILD_KERNEL})
-    configure_file(${TARGET_CONF_TEMP} ${PROJECT_BINARY_DIR}/kernel_config/${TARGET_CORE}/${TARGET_CONF} COPYONLY)
-else ()
-    configure_file(${TARGET_CONF_TEMP} ${PROJECT_BINARY_DIR}/${TARGET_CONF} COPYONLY)
-endif ()
+  # append config data from getarch_2nd to the TARGET file and read in CMake vars
+  file(APPEND ${TARGET_CONF_TEMP} ${GETARCH2_CONF_OUT})
 
-ParseGetArchVars(${GETARCH2_MAKE_OUT})
+  configure_file(${TARGET_CONF_TEMP} ${TARGET_CONF_DIR}/${TARGET_CONF} COPYONLY)
 
-# compile get_config_h
-set(GEN_CONFIG_H_DIR "${PROJECT_BINARY_DIR}/genconfig_h_build")
-set(GEN_CONFIG_H_BIN "gen_config_h${CMAKE_EXECUTABLE_SUFFIX}")
-set(GEN_CONFIG_H_FLAGS "-DVERSION=\"${OpenBLAS_VERSION}\"")
-file(MAKE_DIRECTORY ${GEN_CONFIG_H_DIR})
+  ParseGetArchVars(${GETARCH2_MAKE_OUT})
 
-if (NOT "${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
-  try_compile(GEN_CONFIG_H_RESULT ${GEN_CONFIG_H_DIR}
-    SOURCES ${PROJECT_SOURCE_DIR}/gen_config_h.c
-    COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} ${GEN_CONFIG_H_FLAGS} -I${PROJECT_SOURCE_DIR}
-    OUTPUT_VARIABLE GEN_CONFIG_H_LOG
-    COPY_FILE ${PROJECT_BINARY_DIR}/${GEN_CONFIG_H_BIN}
-  )
-
-  if (NOT ${GEN_CONFIG_H_RESULT})
-    MESSAGE(FATAL_ERROR "Compiling gen_config_h failed ${GEN_CONFIG_H_LOG}")
-  endif ()
-endif ()
+endif()

cmake/system.cmake

@@ -4,10 +4,25 @@
 ##
 set(NETLIB_LAPACK_DIR "${PROJECT_SOURCE_DIR}/lapack-netlib")
 
-# TODO: Makefile.system detects Darwin (mac) and switches to clang here -hpa
-# http://stackoverflow.com/questions/714100/os-detecting-makefile
+# System detection, via CMake.
+include("${PROJECT_SOURCE_DIR}/cmake/system_check.cmake")
+
+if(CMAKE_CROSSCOMPILING AND NOT DEFINED TARGET)
+  # Detect target without running getarch
+  if (ARM64)
+    set(TARGET "ARMV8")
+  elseif(ARM)
+    set(TARGET "ARMV7") # TODO: Ask compiler which arch this is
+  else()
+    message(FATAL_ERROR "When cross compiling, a TARGET is required.")
+  endif()
+endif()
 
-# TODO: Makefile.system sets HOSTCC = $(CC) here if not already set -hpa
+# Other files expect CORE, which is actually TARGET and will become TARGET_CORE for kernel build. Confused yet?
+# It seems we are meant to use TARGET as input and CORE internally as kernel.
+if(NOT DEFINED CORE AND DEFINED TARGET)
+  set(CORE ${TARGET})
+endif()
 
 # TARGET_CORE will override TARGET which is used in DYNAMIC_ARCH=1.
 if (DEFINED TARGET_CORE)
@@ -27,7 +42,7 @@ if (DEFINED BINARY AND DEFINED TARGET AND BINARY EQUAL 32)
 endif ()
 
 if (DEFINED TARGET)
-  message(STATUS "Targetting the ${TARGET} architecture.")
+  message(STATUS "Targeting the ${TARGET} architecture.")
   set(GETARCH_FLAGS "-DFORCE_${TARGET}")
 endif ()
 
@@ -52,21 +67,16 @@ if (NO_AVX2)
   set(GETARCH_FLAGS "${GETARCH_FLAGS} -DNO_AVX2")
 endif ()
 
-if (CMAKE_BUILD_TYPE STREQUAL Debug)
-  set(GETARCH_FLAGS "${GETARCH_FLAGS} -g")
+if (CMAKE_BUILD_TYPE STREQUAL "Debug")
+  set(GETARCH_FLAGS "${GETARCH_FLAGS} ${CMAKE_C_FLAGS_DEBUG}")
 endif ()
 
-# TODO: let CMake handle this? -hpa
-#if (${QUIET_MAKE})
-#  set(MAKE "${MAKE} -s")
-#endif()
-
 if (NOT DEFINED NO_PARALLEL_MAKE)
   set(NO_PARALLEL_MAKE 0)
 endif ()
 set(GETARCH_FLAGS	"${GETARCH_FLAGS} -DNO_PARALLEL_MAKE=${NO_PARALLEL_MAKE}")
 
-if (CMAKE_CXX_COMPILER STREQUAL loongcc)
+if (CMAKE_C_COMPILER STREQUAL loongcc)
   set(GETARCH_FLAGS	"${GETARCH_FLAGS} -static")
 endif ()
 
@@ -77,51 +87,39 @@ else ()
   set(ONLY_CBLAS 0)
 endif ()
 
-include("${PROJECT_SOURCE_DIR}/cmake/prebuild.cmake")
+# N.B. this is NUM_THREAD in Makefile.system which is probably a bug -hpa
+if (NOT CMAKE_CROSSCOMPILING)
+  if (NOT DEFINED NUM_CORES)
+    include(ProcessorCount)
+    ProcessorCount(NUM_CORES)
+  endif()
+
+  if (NOT NUM_CORES EQUAL 0)
+    # HT?
+    set(NUM_THREADS ${NUM_CORES})
+  endif ()
+endif()
 
 if (NOT DEFINED NUM_THREADS)
-  set(NUM_THREADS ${NUM_CORES})
-endif ()
+  set(NUM_THREADS 0)
+endif()
 
-if (${NUM_THREADS} EQUAL 1)
+if (${NUM_THREADS} LESS 2)
   set(USE_THREAD 0)
+elseif(NOT DEFINED USE_THREAD)
+  set(USE_THREAD 1)
 endif ()
 
-if (DEFINED USE_THREAD)
-  if (NOT ${USE_THREAD})
-    unset(SMP)
-  else ()
-    set(SMP 1)
-  endif ()
-else ()
-  # N.B. this is NUM_THREAD in Makefile.system which is probably a bug -hpa
-  if (${NUM_THREADS} EQUAL 1)
-    unset(SMP)
-  else ()
-    set(SMP 1)
-  endif ()
+if (USE_THREAD)
+  message(STATUS "Multi-threading enabled with ${NUM_THREADS} threads.")
 endif ()
 
-if (${SMP})
-  message(STATUS "SMP enabled.")
-endif ()
+include("${PROJECT_SOURCE_DIR}/cmake/prebuild.cmake")
 
 if (NOT DEFINED NEED_PIC)
   set(NEED_PIC 1)
 endif ()
 
-# TODO: I think CMake should be handling all this stuff -hpa
-unset(ARFLAGS)
-set(CPP "${COMPILER} -E")
-set(AR "${CROSS_SUFFIX}ar")
-set(AS "${CROSS_SUFFIX}as")
-set(LD "${CROSS_SUFFIX}ld")
-set(RANLIB "${CROSS_SUFFIX}ranlib")
-set(NM "${CROSS_SUFFIX}nm")
-set(DLLWRAP "${CROSS_SUFFIX}dllwrap")
-set(OBJCOPY "${CROSS_SUFFIX}objcopy")
-set(OBJCONV "${CROSS_SUFFIX}objconv")
-
 # OS dependent settings
 include("${PROJECT_SOURCE_DIR}/cmake/os.cmake")
 
@@ -149,11 +147,13 @@ if (NEED_PIC)
     set(CCOMMON_OPT "${CCOMMON_OPT} -fPIC")
   endif ()
 
-  if (${F_COMPILER} STREQUAL "SUN")
-    set(FCOMMON_OPT "${FCOMMON_OPT} -pic")
-  else ()
-    set(FCOMMON_OPT "${FCOMMON_OPT} -fPIC")
-  endif ()
+  if (NOT NOFORTRAN)
+    if (${F_COMPILER} STREQUAL "SUN")
+      set(FCOMMON_OPT "${FCOMMON_OPT} -pic")
+    else ()
+      set(FCOMMON_OPT "${FCOMMON_OPT} -fPIC")
+    endif ()
+  endif()
 endif ()
 
 if (DYNAMIC_ARCH)
@@ -174,7 +174,7 @@ if (NO_AVX)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX")
 endif ()
 
-if (${ARCH} STREQUAL "x86")
+if (X86)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX")
 endif ()
 
@@ -182,25 +182,20 @@ if (NO_AVX2)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX2")
 endif ()
 
-if (SMP)
+if (USE_THREAD)
+  # USE_SIMPLE_THREADED_LEVEL3 = 1
+  # NO_AFFINITY = 1
   set(CCOMMON_OPT "${CCOMMON_OPT} -DSMP_SERVER")
 
-  if (${ARCH} STREQUAL "mips64")
+  if (MIPS64)
     if (NOT ${CORE} STREQUAL "LOONGSON3B")
       set(USE_SIMPLE_THREADED_LEVEL3 1)
     endif ()
   endif ()
 
-  if (USE_OPENMP)
-    # USE_SIMPLE_THREADED_LEVEL3 = 1
-    # NO_AFFINITY = 1
-    set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_OPENMP")
-  endif ()
-
   if (BIGNUMA)
     set(CCOMMON_OPT "${CCOMMON_OPT} -DBIGNUMA")
   endif ()
-
 endif ()
 
 if (NO_WARMUP)
@@ -263,7 +258,7 @@ if (NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
   set(NO_AFFINITY 1)
 endif ()
 
-if (NOT ${ARCH} STREQUAL "x86_64" AND NOT ${ARCH} STREQUAL "x86" AND NOT ${CORE} STREQUAL "LOONGSON3B")
+if (NOT X86_64 AND NOT X86 AND NOT ${CORE} STREQUAL "LOONGSON3B")
   set(NO_AFFINITY 1)
 endif ()
 
@@ -295,52 +290,19 @@ if (MIXED_MEMORY_ALLOCATION)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DMIXED_MEMORY_ALLOCATION")
 endif ()
 
-if (${CMAKE_SYSTEM_NAME} STREQUAL "SunOS")
-  set(TAR	gtar)
-  set(PATCH	gpatch)
-  set(GREP ggrep)
-else ()
-  set(TAR tar)
-  set(PATCH patch)
-  set(GREP grep)
-endif ()
-
-if (NOT DEFINED MD5SUM)
-  set(MD5SUM md5sum)
-endif ()
-
-set(AWK awk)
-
-set(SED sed)
-
 set(REVISION "-r${OpenBLAS_VERSION}")
 set(MAJOR_VERSION ${OpenBLAS_MAJOR_VERSION})
 
-if (DEBUG)
-  set(COMMON_OPT "${COMMON_OPT} -g")
-endif ()
-
-if (NOT DEFINED COMMON_OPT)
-  set(COMMON_OPT "-O2")
-endif ()
-
-#For x86 32-bit
-if (DEFINED BINARY AND BINARY EQUAL 32)
-if (NOT MSVC)
-  set(COMMON_OPT "${COMMON_OPT} -m32")
-endif()
-endif()
-
-set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${COMMON_OPT} ${CCOMMON_OPT}")
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${CCOMMON_OPT}")
 if(NOT MSVC)
-set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${COMMON_OPT} ${CCOMMON_OPT}")
+set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${CCOMMON_OPT}")
 endif()
 # TODO: not sure what PFLAGS is -hpa
-set(PFLAGS "${PFLAGS} ${COMMON_OPT} ${CCOMMON_OPT} -I${TOPDIR} -DPROFILE ${COMMON_PROF}")
+set(PFLAGS "${PFLAGS} ${CCOMMON_OPT} -I${TOPDIR} -DPROFILE ${COMMON_PROF}")
 
-set(CMAKE_Fortran_FLAGS "${CMAKE_Fortran_FLAGS} ${COMMON_OPT} ${FCOMMON_OPT}")
+set(CMAKE_Fortran_FLAGS "${CMAKE_Fortran_FLAGS} ${FCOMMON_OPT}")
 # TODO: not sure what FPFLAGS is -hpa
-set(FPFLAGS "${FPFLAGS} ${COMMON_OPT} ${FCOMMON_OPT} ${COMMON_PROF}")
+set(FPFLAGS "${FPFLAGS} ${FCOMMON_OPT} ${COMMON_PROF}")
 
 #For LAPACK Fortran codes.
 set(LAPACK_FFLAGS "${LAPACK_FFLAGS} ${CMAKE_Fortran_FLAGS}")
@@ -348,7 +310,7 @@ set(LAPACK_FPFLAGS "${LAPACK_FPFLAGS} ${FPFLAGS}")
 
 #Disable -fopenmp for LAPACK Fortran codes on Windows.
 if (${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
-  set(FILTER_FLAGS "-fopenmp;-mp;-openmp;-xopenmp=parralel")
+  set(FILTER_FLAGS "-fopenmp;-mp;-openmp;-xopenmp=parallel")
   foreach (FILTER_FLAG ${FILTER_FLAGS})
     string(REPLACE ${FILTER_FLAG} "" LAPACK_FFLAGS ${LAPACK_FFLAGS})
     string(REPLACE ${FILTER_FLAG} "" LAPACK_FPFLAGS ${LAPACK_FPFLAGS})
@@ -386,7 +348,7 @@ if (NOT DEFINED LIBSUFFIX)
 endif ()
 
 if (DYNAMIC_ARCH)
-  if (DEFINED SMP)
+  if (USE_THREAD)
     set(LIBNAME "${LIBPREFIX}p${REVISION}.${LIBSUFFIX}")
     set(LIBNAME_P	"${LIBPREFIX}p${REVISION}_p.${LIBSUFFIX}")
   else ()
@@ -394,7 +356,7 @@ if (DYNAMIC_ARCH)
     set(LIBNAME_P	"${LIBPREFIX}${REVISION}_p.${LIBSUFFIX}")
   endif ()
 else ()
-  if (DEFINED SMP)
+  if (USE_THREAD)
     set(LIBNAME "${LIBPREFIX}_${LIBCORE}p${REVISION}.${LIBSUFFIX}")
     set(LIBNAME_P	"${LIBPREFIX}_${LIBCORE}p${REVISION}_p.${LIBSUFFIX}")
   else ()
@@ -436,7 +398,7 @@ endif ()
 set(USE_GEMM3M 0)
 
 if (DEFINED ARCH)
-  if (${ARCH} STREQUAL "x86" OR ${ARCH} STREQUAL "x86_64" OR ${ARCH} STREQUAL "ia64" OR ${ARCH} STREQUAL "MIPS")
+  if (X86 OR X86_64 OR ${ARCH} STREQUAL "ia64" OR MIPS64)
     set(USE_GEMM3M 1)
   endif ()
 
@@ -519,35 +481,3 @@ endif ()
 #  export CUFLAGS
 #  export CULIB
 #endif
-
-#.SUFFIXES: .$(PSUFFIX) .$(SUFFIX) .f
-#
-#.f.$(SUFFIX):
-#	$(FC) $(FFLAGS) -c $<  -o $(@F)
-#
-#.f.$(PSUFFIX):
-#	$(FC) $(FPFLAGS) -pg -c $<  -o $(@F)
-
-# these are not cross-platform
-#ifdef BINARY64
-#PATHSCALEPATH	= /opt/pathscale/lib/3.1
-#PGIPATH		= /opt/pgi/linux86-64/7.1-5/lib
-#else
-#PATHSCALEPATH	= /opt/pathscale/lib/3.1/32
-#PGIPATH		= /opt/pgi/linux86/7.1-5/lib
-#endif
-
-#ACMLPATH	= /opt/acml/4.3.0
-#ifneq ($(OSNAME), Darwin)
-#MKLPATH         = /opt/intel/mkl/10.2.2.025/lib
-#else
-#MKLPATH         = /Library/Frameworks/Intel_MKL.framework/Versions/10.0.1.014/lib
-#endif
-#ATLASPATH	= /opt/atlas/3.9.17/opteron
-#FLAMEPATH	= $(HOME)/flame/lib
-#ifneq ($(OSNAME), SunOS)
-#SUNPATH		= /opt/sunstudio12.1
-#else
-#SUNPATH		= /opt/SUNWspro
-#endif
-

cmake/system_check.cmake

@@ -0,0 +1,68 @@
+##
+## Author: Hank Anderson <hank@statease.com>
+## Description: Ported from the OpenBLAS/c_check perl script.
+##              This is triggered by prebuild.cmake and runs before any of the code is built.
+##              Creates config.h and Makefile.conf.
+
+# Convert CMake vars into the format that OpenBLAS expects
+string(TOUPPER ${CMAKE_SYSTEM_NAME} HOST_OS)
+if (${HOST_OS} STREQUAL "WINDOWS")
+  set(HOST_OS WINNT)
+endif ()
+
+if(CMAKE_COMPILER_IS_GNUCC AND WIN32)
+    execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpmachine
+              OUTPUT_VARIABLE OPENBLAS_GCC_TARGET_MACHINE
+              OUTPUT_STRIP_TRAILING_WHITESPACE)
+    if(OPENBLAS_GCC_TARGET_MACHINE MATCHES "amd64|x86_64|AMD64")
+      set(MINGW64 1)
+    endif()
+endif()
+
+# Pretty thorough determination of arch. Add more if needed
+if(CMAKE_CL_64 OR MINGW64)
+  set(X86_64 1)
+elseif(MINGW OR (MSVC AND NOT CMAKE_CROSSCOMPILING))
+  set(X86 1)
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "ppc.*|power.*|Power.*")
+  set(PPC 1)
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "mips64.*")
+  set(MIPS64 1)
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "amd64.*|x86_64.*|AMD64.*")
+  set(X86_64 1)
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "i686.*|i386.*|x86.*|amd64.*|AMD64.*")
+  set(X86 1)
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(arm.*|ARM.*)")
+  set(ARM 1)
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64.*|AARCH64.*)")
+  set(ARM64 1)
+endif()
+
+if (X86_64)
+  set(ARCH "x86_64")
+elseif(X86)
+  set(ARCH "x86")
+elseif(PPC)
+  set(ARCH "power")
+elseif(ARM)
+  set(ARCH "arm")
+elseif(ARM64)
+  set(ARCH "arm64")
+else()
+  set(ARCH ${CMAKE_SYSTEM_PROCESSOR} CACHE STRING "Target Architecture")
+endif ()
+
+if (NOT BINARY)
+  if (X86_64 OR ARM64 OR PPC OR MIPS64)
+    set(BINARY 64)
+  else ()
+    set(BINARY 32)
+  endif ()
+endif()
+
+if(BINARY EQUAL 64)
+  set(BINARY64 1)
+else()
+  set(BINARY32 1)
+endif()
+

ctest/CMakeLists.txt

@@ -16,7 +16,7 @@ foreach(float_type ${FLOAT_TYPES})
   add_executable(x${float_char}cblat1
     c_${float_char}blat1.f
     c_${float_char}blas1.c)
-  target_link_libraries(x${float_char}cblat1 ${OpenBLAS_LIBNAME}_static)
+  target_link_libraries(x${float_char}cblat1 ${OpenBLAS_LIBNAME})
   add_test(NAME "x${float_char}cblat1"
     COMMAND "${CMAKE_CURRENT_BINARY_DIR}/x${float_char}cblat1")
 
@@ -28,7 +28,7 @@ foreach(float_type ${FLOAT_TYPES})
     auxiliary.c
     c_xerbla.c
     constant.c)
-  target_link_libraries(x${float_char}cblat2 ${OpenBLAS_LIBNAME}_static)
+  target_link_libraries(x${float_char}cblat2 ${OpenBLAS_LIBNAME})
   add_test(NAME "x${float_char}cblat2"
     COMMAND sh "${CMAKE_CURRENT_BINARY_DIR}/test_cblas_helper.sh" "${CMAKE_CURRENT_BINARY_DIR}/x${float_char}cblat2" "${PROJECT_SOURCE_DIR}/ctest/${float_char}in2")
 
@@ -40,7 +40,7 @@ foreach(float_type ${FLOAT_TYPES})
     auxiliary.c
     c_xerbla.c
     constant.c)
-  target_link_libraries(x${float_char}cblat3 ${OpenBLAS_LIBNAME}_static)
+  target_link_libraries(x${float_char}cblat3 ${OpenBLAS_LIBNAME})
   add_test(NAME "x${float_char}cblat3"
     COMMAND sh "${CMAKE_CURRENT_BINARY_DIR}/test_cblas_helper.sh" "${CMAKE_CURRENT_BINARY_DIR}/x${float_char}cblat3" "${PROJECT_SOURCE_DIR}/ctest/${float_char}in3")
 

driver/level2/CMakeLists.txt

@@ -73,7 +73,7 @@ GenerateNamedObjects("zgbmv_k.c" "TRANS;CONJ;XCONJ" "gbmv_d" false "" "" "" 2)
 # special defines for complex
 foreach (float_type ${FLOAT_TYPES})
 
-  if (SMP)
+  if (USE_THREAD)
     GenerateNamedObjects("gemv_thread.c" "" "gemv_thread_n" false "" "" false ${float_type})
     GenerateNamedObjects("gemv_thread.c" "TRANSA" "gemv_thread_t" false "" "" false ${float_type})
 
@@ -107,7 +107,7 @@ foreach (float_type ${FLOAT_TYPES})
       GenerateNamedObjects("z${ulvm_source}" "LOWER;HEMVREV" "${op_name}_M" false "" "" false ${float_type})
     endforeach()
 
-    if (SMP)
+    if (USE_THREAD)
 
       GenerateNamedObjects("gemv_thread.c" "CONJ" "gemv_thread_r" false "" "" false ${float_type})
       GenerateNamedObjects("gemv_thread.c" "CONJ;TRANSA" "gemv_thread_c" false "" "" false ${float_type})
@@ -186,7 +186,7 @@ foreach (float_type ${FLOAT_TYPES})
       GenerateCombinationObjects("${l_source}" "UNIT" "N" "TRANSA" 0 "${op_name}_TU" false ${float_type})
     endforeach ()
 
-    if (SMP)
+    if (USE_THREAD)
       GenerateNamedObjects("ger_thread.c" "" "" false "" "" false ${float_type})
       foreach(nu_smp_source ${NU_SMP_SOURCES})
         string(REGEX MATCH "[a-z]+_[a-z]+" op_name ${nu_smp_source})
@@ -197,7 +197,7 @@ foreach (float_type ${FLOAT_TYPES})
   endif ()
 endforeach ()
 
-if (SMP)
+if (USE_THREAD)
   GenerateCombinationObjects("${UL_SMP_SOURCES}" "LOWER" "U" "" 2)
 endif ()
 

driver/level3/CMakeLists.txt

@@ -9,7 +9,7 @@ set(GEMM_COMPLEX_DEFINES RN CN RT CT NR TR RR CR NC TC RC CC)
 foreach (GEMM_DEFINE ${GEMM_DEFINES})
   string(TOLOWER ${GEMM_DEFINE} GEMM_DEFINE_LC)
   GenerateNamedObjects("gemm.c" "${GEMM_DEFINE}" "gemm_${GEMM_DEFINE_LC}" 0)
-  if (SMP AND NOT USE_SIMPLE_THREADED_LEVEL3)
+  if (USE_THREAD AND NOT USE_SIMPLE_THREADED_LEVEL3)
     GenerateNamedObjects("gemm.c" "${GEMM_DEFINE};THREADED_LEVEL3" "gemm_thread_${GEMM_DEFINE_LC}" 0)
   endif ()
 endforeach ()
@@ -32,7 +32,7 @@ GenerateCombinationObjects("syrk_k.c" "LOWER;TRANS" "U;N" "" 1)
 GenerateCombinationObjects("syr2k_k.c" "LOWER;TRANS" "U;N" "" 1)
 GenerateCombinationObjects("syrk_kernel.c" "LOWER" "U" "" 2)
 GenerateCombinationObjects("syr2k_kernel.c" "LOWER" "U" "" 2)
-if (SMP)
+if (USE_THREAD)
 
   # N.B. these do NOT have a float type (e.g. DOUBLE) defined!
   GenerateNamedObjects("gemm_thread_m.c;gemm_thread_n.c;gemm_thread_mn.c;gemm_thread_variable.c;syrk_thread.c" "" "" 0 "" "" 1)
@@ -71,7 +71,7 @@ foreach (float_type ${FLOAT_TYPES})
     GenerateNamedObjects("zher2k_k.c" "HER2K;LOWER" "her2k_LN" false "" "" false ${float_type})
     GenerateNamedObjects("zher2k_k.c" "HER2K;LOWER;TRANS;CONJ" "her2k_LC" false "" "" false ${float_type})
 
-    if (SMP AND NOT USE_SIMPLE_THREADED_LEVEL3)
+    if (USE_THREAD AND NOT USE_SIMPLE_THREADED_LEVEL3)
       #herk
       GenerateCombinationObjects("zherk_k.c" "LOWER" "U" "HERK;THREADED_LEVEL3" 3 "herk_thread_N" false ${float_type})
       GenerateCombinationObjects("zherk_k.c" "LOWER" "U" "HERK;THREADED_LEVEL3;TRANS;CONJ" 3 "herk_thread_C" false ${float_type})
@@ -93,7 +93,7 @@ foreach (float_type ${FLOAT_TYPES})
       if(USE_GEMM3M)
 	GenerateNamedObjects("gemm3m.c" "${gemm_define}" "gemm3m_${gemm_define_LC}" false "" "" false ${float_type})
       endif()
-      if (SMP AND NOT USE_SIMPLE_THREADED_LEVEL3)
+      if (USE_THREAD AND NOT USE_SIMPLE_THREADED_LEVEL3)
         GenerateNamedObjects("gemm.c" "${gemm_define};THREADED_LEVEL3" "gemm_thread_${gemm_define_LC}" false "" "" false ${float_type})
 	if(USE_GEMM3M)
 	  GenerateNamedObjects("gemm3m.c" "${gemm_define};THREADED_LEVEL3" "gemm3m_thread_${gemm_define_LC}" false "" "" false ${float_type})
@@ -106,7 +106,7 @@ foreach (float_type ${FLOAT_TYPES})
       foreach (GEMM_DEFINE ${GEMM_DEFINES})
 	string(TOLOWER ${GEMM_DEFINE} GEMM_DEFINE_LC)
 	GenerateNamedObjects("gemm3m.c" "${GEMM_DEFINE}" "gemm3m_${GEMM_DEFINE_LC}" false "" "" false ${float_type})
-	if (SMP AND NOT USE_SIMPLE_THREADED_LEVEL3)
+	if (USE_THREAD AND NOT USE_SIMPLE_THREADED_LEVEL3)
 	  GenerateNamedObjects("gemm3m.c" "${GEMM_DEFINE};THREADED_LEVEL3" "gemm3m_thread_${GEMM_DEFINE_LC}" false "" "" false ${float_type})
 	endif ()
       endforeach ()

driver/others/CMakeLists.txt

@@ -7,7 +7,7 @@ else ()
   set(MEMORY memory.c)
 endif ()
 
-if (SMP)
+if (USE_THREAD)
 
   if (USE_OPENMP)
     set(BLAS_SERVER blas_server_omp.c)

kernel/CMakeLists.txt

@@ -21,8 +21,8 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
     endif ()
 
     if (${ADD_COMMONOBJS})
-        if (${ARCH} STREQUAL "x86")
-        if (NOT "${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
+        if (X86)
+        if (NOT "${CMAKE_C_COMPILER_ID}" STREQUAL "MSVC")
           GenerateNamedObjects("${KERNELDIR}/cpuid.S" "" "" false "" "" true)
         else()
           GenerateNamedObjects("${KERNELDIR}/cpuid_win.c" "" "" false "" "" true)
@@ -121,7 +121,7 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
     # Makefile.L3
     set(USE_TRMM false)
 
-    if (${ARCH} STREQUAL "arm" OR ${ARCH} STREQUAL "arm64" OR "${TARGET}" STREQUAL "LONGSOON3B" OR "${TARGET}" STREQUAL "GENERIC" OR "${CORE}" STREQUAL "generic" OR "${TARGET}" STREQUAL "HASWELL" OR "${CORE}" STREQUAL "haswell" OR "{CORE}" STREQUAL "zen")
+    if (ARM OR ARM64 OR "${TARGET}" STREQUAL "LONGSOON3B" OR "${TARGET}" STREQUAL "GENERIC" OR "${CORE}" STREQUAL "generic" OR "${TARGET}" STREQUAL "HASWELL" OR "${CORE}" STREQUAL "haswell" OR "{CORE}" STREQUAL "zen")
       set(USE_TRMM true)
     endif ()
 
@@ -464,6 +464,30 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
       GenerateNamedObjects("${KERNELDIR}/${${float_char}GEADD_KERNEL}" "" "geadd_k" false "" "" false ${float_type})
     endforeach ()
 
+    # Makefile.LA
+    if(NOT NO_LAPACK)
+      foreach (float_type ${FLOAT_TYPES})
+        if (NOT DEFINED ${float_char}NEG_TCOPY)
+          if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C" OR ${float_char} STREQUAL "X")
+            set(${float_char}NEG_TCOPY ../generic/zneg_tcopy.c)
+          else ()
+            set(${float_char}NEG_TCOPY ../generic/neg_tcopy.c)
+          endif ()
+        endif ()
+
+        if (NOT DEFINED ${float_char}LASWP_NCOPY)
+          if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C" OR ${float_char} STREQUAL "X")
+            set(${float_char}LASWP_NCOPY ../generic/zlaswp_ncopy.c)
+          else ()
+            set(${float_char}LASWP_NCOPY ../generic/laswp_ncopy.c)
+          endif ()
+        endif ()
+        string(SUBSTRING ${float_type} 0 1 float_char)
+        GenerateNamedObjects("${KERNELDIR}/${${float_char}NEG_TCOPY}_${${float_char}GEMM_UNROLL_M}" "" "neg_tcopy" false "" "" false ${float_type})
+        GenerateNamedObjects("${KERNELDIR}/${${float_char}LASWP_NCOPY}_${${float_char}GEMM_UNROLL_N}" "" "laswp_ncopy" false "" "" false ${float_type})
+      endforeach()
+    endif()
+
     if (${DYNAMIC_ARCH})
         set(SETPARAM_TARGET_DIR ${CMAKE_CURRENT_BINARY_DIR}${CMAKE_FILES_DIRECTORY})
         file(READ ${CMAKE_CURRENT_SOURCE_DIR}/setparam-ref.c SETPARAM_REF_CONTENTS)
@@ -484,9 +508,6 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
         file(REMOVE ${SETPARAM_TARGET_DIR}/kernel${TSUFFIX}.tmp)
     endif ()
 
-    # Makefile.LA
-    #DBLASOBJS	+= dneg_tcopy$(TSUFFIX).$(SUFFIX) dlaswp_ncopy$(TSUFFIX).$(SUFFIX)
-
     add_library(kernel${TSUFFIX} OBJECT ${OPENBLAS_SRC})
     set_target_properties(kernel${TSUFFIX} PROPERTIES COMPILE_FLAGS "${KERNEL_DEFINITIONS}")
     get_target_property(KERNEL_INCLUDE_DIRECTORIES kernel${TSUFFIX} INCLUDE_DIRECTORIES)

lapack/CMakeLists.txt

@@ -46,7 +46,7 @@ GenerateNamedObjects("${LAPACK_MANGLED_SOURCES}" "" "" false "" "" false 3)
 GenerateNamedObjects("laswp/generic/laswp_k.c" "" "laswp_plus" false "" "" false 3)
 GenerateNamedObjects("laswp/generic/laswp_k.c" "MINUS" "laswp_minus" false "" "" false 3)
 
-if (SMP)
+if (USE_THREAD)
 
   if (USE_OPENMP)
     set(GETRF_SRC getrf/getrf_parallel_omp.c)

test/CMakeLists.txt

@@ -11,7 +11,7 @@ set(OpenBLAS_Tests
 
 foreach(test_bin ${OpenBLAS_Tests})
 add_executable(${test_bin} ${test_bin}.f)
-target_link_libraries(${test_bin} ${OpenBLAS_LIBNAME}_static)
+target_link_libraries(${test_bin} ${OpenBLAS_LIBNAME})
 endforeach()
 
 # $1 exec, $2 input, $3 output_result