For gemm multi-threading, simply split M.

e.g.
layer 1: A (1600k, 576), B(576, 64)

B is very small. We split M.

.gitignore

@@ -15,6 +15,7 @@ lapack-netlib/make.inc
 lapack-netlib/lapacke/include/lapacke_mangling.h
 lapack-netlib/TESTING/testing_results.txt
 *.so
+*.so.*
 *.a
 .svn
 *~
@@ -65,3 +66,6 @@ test/sblat3
 test/zblat1
 test/zblat2
 test/zblat3
+build
+build.*
+*.swp

.travis.yml

@@ -1,4 +1,13 @@
 language: c
+
+notifications:
+  webhooks:
+    urls:
+      - https://webhooks.gitter.im/e/8a6e4470a0cebd090344
+    on_success: change  # options: [always|never|change] default: always
+    on_failure: always  # options: [always|never|change] default: always
+    on_start: never     # options: [always|never|change] default: always
+
 compiler:
   - gcc
 

CMakeLists.txt

@@ -0,0 +1,190 @@
+##
+## Author: Hank Anderson <hank@statease.com>
+##
+
+cmake_minimum_required(VERSION 2.8.4)
+project(OpenBLAS)
+set(OpenBLAS_MAJOR_VERSION 0)
+set(OpenBLAS_MINOR_VERSION 2)
+set(OpenBLAS_PATCH_VERSION 16.dev)
+set(OpenBLAS_VERSION "${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.${OpenBLAS_PATCH_VERSION}")
+
+enable_language(ASM)
+enable_language(C)
+
+if(MSVC)
+set(OpenBLAS_LIBNAME libopenblas)
+else()
+set(OpenBLAS_LIBNAME openblas)
+endif()
+
+#######
+if(MSVC)
+option(BUILD_WITHOUT_LAPACK "Without LAPACK and LAPACKE (Only BLAS or CBLAS)" ON)
+endif()
+option(BUILD_WITHOUT_CBLAS "Without CBLAS" OFF)
+option(BUILD_DEBUG "Build Debug Version" OFF)
+#######
+if(BUILD_WITHOUT_LAPACK)
+set(NO_LAPACK 1)
+set(NO_LAPACKE 1)
+endif()
+
+if(BUILD_DEBUG)
+set(CMAKE_BUILD_TYPE Debug)
+else()
+set(CMAKE_BUILD_TYPE Release)
+endif()
+
+if(BUILD_WITHOUT_CBLAS)
+set(NO_CBLAS 1)
+endif()
+
+#######
+
+
+message(WARNING "CMake support is experimental. This will not produce the same Makefiles that OpenBLAS ships with. Only x86 support is currently available.")
+
+include("${CMAKE_SOURCE_DIR}/cmake/utils.cmake")
+include("${CMAKE_SOURCE_DIR}/cmake/system.cmake")
+
+set(BLASDIRS interface driver/level2 driver/level3 driver/others)
+
+if (NOT DYNAMIC_ARCH)
+  list(APPEND BLASDIRS kernel)
+endif ()
+
+if (DEFINED UTEST_CHECK)
+  set(SANITY_CHECK 1)
+endif ()
+
+if (DEFINED SANITY_CHECK)
+  list(APPEND BLASDIRS reference)
+endif ()
+
+set(SUBDIRS	${BLASDIRS})
+if (NOT NO_LAPACK)
+  list(APPEND SUBDIRS lapack)
+endif ()
+
+# set which float types we want to build for
+if (NOT DEFINED BUILD_SINGLE AND NOT DEFINED BUILD_DOUBLE AND NOT DEFINED BUILD_COMPLEX AND NOT DEFINED BUILD_COMPLEX16)
+  # if none are defined, build for all
+  set(BUILD_SINGLE true)
+  set(BUILD_DOUBLE true)
+  set(BUILD_COMPLEX true)
+  set(BUILD_COMPLEX16 true)
+endif ()
+
+set(FLOAT_TYPES "")
+if (BUILD_SINGLE)
+  message(STATUS "Building Single Precision")
+  list(APPEND FLOAT_TYPES "SINGLE") # defines nothing
+endif ()
+
+if (BUILD_DOUBLE)
+  message(STATUS "Building Double Precision")
+  list(APPEND FLOAT_TYPES "DOUBLE") # defines DOUBLE
+endif ()
+
+if (BUILD_COMPLEX)
+  message(STATUS "Building Complex Precision")
+  list(APPEND FLOAT_TYPES "COMPLEX") # defines COMPLEX
+endif ()
+
+if (BUILD_COMPLEX16)
+  message(STATUS "Building Double Complex Precision")
+  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 ()
+
+# get obj vars into format that add_library likes: $<TARGET_OBJS:objlib> (see http://www.cmake.org/cmake/help/v3.0/command/add_library.html)
+set(TARGET_OBJS "")
+foreach (SUBDIR ${SUBDIRS})
+  add_subdirectory(${SUBDIR})
+  string(REPLACE "/" "_" subdir_obj ${SUBDIR})
+  list(APPEND TARGET_OBJS "$<TARGET_OBJECTS:${subdir_obj}>")
+endforeach ()
+
+# netlib:
+
+# Can't just use lapack-netlib's CMake files, since they are set up to search for BLAS, build and install a binary. We just want to build a couple of lib files out of lapack and lapacke.
+# Not using add_subdirectory here because lapack-netlib already has its own CMakeLists.txt. Instead include a cmake script with the sources we want.
+if (NOT NOFORTRAN AND NOT NO_LAPACK)
+  include("${CMAKE_SOURCE_DIR}/cmake/lapack.cmake")
+if (NOT NO_LAPACKE)
+  include("${CMAKE_SOURCE_DIR}/cmake/lapacke.cmake")
+endif ()
+endif ()
+
+#Only generate .def for dll on MSVC
+if(MSVC)
+set(OpenBLAS_DEF_FILE "${PROJECT_BINARY_DIR}/openblas.def")
+endif()
+
+# add objects to the openblas lib
+add_library(${OpenBLAS_LIBNAME} SHARED ${LA_SOURCES} ${LAPACKE_SOURCES} ${TARGET_OBJS} ${OpenBLAS_DEF_FILE})
+
+include("${CMAKE_SOURCE_DIR}/cmake/export.cmake")
+
+
+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)
+endif()
+
+#build test and ctest
+enable_testing()
+add_subdirectory(test)
+if(NOT NO_CBLAS)
+add_subdirectory(ctest)
+endif()
+endif()
+
+set_target_properties(${OpenBLAS_LIBNAME} PROPERTIES 
+  VERSION ${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}
+  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
+###
+#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
+#ifdef USE_THREAD
+#	@echo USE_THREAD=$(USE_THREAD) >>  Makefile.conf_last
+#endif
+#	@touch lib.grd
+

CONTRIBUTORS.md

@@ -127,5 +127,8 @@ In chronological order:
 * Ton van den Heuvel <https://github.com/ton>
   * [2015-03-18] Fix race condition during shutdown causing a crash in gotoblas_set_affinity().
 
+* Martin Koehler <https://github.com/grisuthedragon/>
+  * [2015-09-07] Improved imatcopy
+
 * [Your name or handle] <[email or website]>
   * [Date] [Brief summary of your changes]

Changelog.txt

@@ -1,4 +1,57 @@
 OpenBLAS ChangeLog
+====================================================================
+Version 0.2.15
+27-Oct-2015
+common:
+	* Support cmake on x86/x86-64. Natively compiling on MS Visual Studio.
+	  (experimental. Thank Hank Anderson for the initial cmake porting work.)
+	  
+	  On Linux and Mac OSX, OpenBLAS cmake supports assembly kernels.
+	  e.g. cmake .
+	       make
+	       make test (Optional)
+
+	  On Windows MS Visual Studio, OpenBLAS cmake only support C kernels.
+	  (OpenBLAS uses AT&T style assembly, which is not supported by MSVC.)
+	  e.g. cmake -G "Visual Studio 12 Win64" .
+	       Open OpenBLAS.sln and build.
+	  
+	* Enable MAX_STACK_ALLOC flags by default.
+	  Improve ger and gemv for small matrices.
+	* Improve gemv parallel with small m and large n case.
+	* Improve ?imatcopy when lda==ldb (#633. Thanks, Martin Koehler)
+	* Add vecLib benchmarks (#565. Thanks, Andreas Noack.)
+	* Fix LAPACK lantr for row major matrices (#634. Thanks, Dan Kortschak)
+	* Fix LAPACKE lansy (#640. Thanks, Dan Kortschak)
+	* Import bug fixes for LAPACKE s/dormlq, c/zunmlq 
+	* Raise the signal when pthread_create fails (#668. Thanks, James K. Lowden)
+	* Remove g77 from compiler list.
+	* Enable AppVeyor Windows CI.
+
+x86/x86-64:
+	* Support pure C generic kernels for x86/x86-64.
+	* Support Intel Boardwell and Skylake by Haswell kernels.
+	* Support AMD Excavator by Steamroller kernels.
+	* Optimize s/d/c/zdot for Intel SandyBridge and Haswell.
+	* Optimize s/d/c/zdot for AMD Piledriver and Steamroller.
+	* Optimize s/d/c/zapxy for Intel SandyBridge and Haswell.
+	* Optimize s/d/c/zapxy for AMD Piledriver and Steamroller.
+	* Optimize d/c/zscal for Intel Haswell, dscal for Intel SandyBridge.
+	* Optimize d/c/zscal for AMD Bulldozer, Piledriver and Steamroller.
+	* Optimize s/dger for Intel SandyBridge.
+	* Optimize s/dsymv for Intel SandyBridge.
+	* Optimize ssymv for Intel Haswell.
+	* Optimize dgemv for Intel Nehalem and Haswell.
+	* Optimize dtrmm for Intel Haswell.
+
+ARM:
+	* Support Android NDK armeabi-v7a-hard ABI (-mfloat-abi=hard)
+	  e.g. make HOSTCC=gcc CC=arm-linux-androideabi-gcc NO_LAPACK=1 TARGET=ARMV7
+	* Fix lock, rpcc bugs (#616, #617. Thanks, Grazvydas Ignotas)
+POWER:
+	* Support ppc64le platform (ELF ABI v2. #612. Thanks, Matthew Brandyberry.)
+	* Support POWER7/8 by POWER6 kernels. (#612. Thanks, Fábio Perez.)
+
 ====================================================================
 Version 0.2.14
 24-Mar-2015

Makefile

@@ -249,10 +249,14 @@ ifndef NOFORTRAN
 	-@echo "SUFFIX      = $(SUFFIX)" >> $(NETLIB_LAPACK_DIR)/make.inc
 	-@echo "PSUFFIX     = $(PSUFFIX)" >> $(NETLIB_LAPACK_DIR)/make.inc
 	-@echo "CEXTRALIB   = $(EXTRALIB)" >> $(NETLIB_LAPACK_DIR)/make.inc
-ifeq ($(FC), gfortran)
+ifeq ($(F_COMPILER), GFORTRAN)
 	-@echo "TIMER       = INT_ETIME" >> $(NETLIB_LAPACK_DIR)/make.inc
 ifdef SMP
+ifeq ($(OSNAME), WINNT)
+	-@echo "LOADER      = $(FC)" >> $(NETLIB_LAPACK_DIR)/make.inc
+else
 	-@echo "LOADER      = $(FC) -pthread" >> $(NETLIB_LAPACK_DIR)/make.inc
+endif
 else
 	-@echo "LOADER      = $(FC)" >> $(NETLIB_LAPACK_DIR)/make.inc
 endif
@@ -288,7 +292,17 @@ endif
 lapack-test :
 	(cd $(NETLIB_LAPACK_DIR)/TESTING && rm -f x* *.out)
 	make -j 1 -C $(NETLIB_LAPACK_DIR)/TESTING xeigtstc  xeigtstd  xeigtsts  xeigtstz  xlintstc  xlintstd  xlintstds  xlintstrfd  xlintstrfz  xlintsts  xlintstz  xlintstzc xlintstrfs xlintstrfc
+ifneq ($(CROSS), 1)
+	( cd $(NETLIB_LAPACK_DIR)/INSTALL; ./testlsame; ./testslamch; ./testdlamch; \
+        ./testsecond; ./testdsecnd; ./testieee; ./testversion )
 	(cd $(NETLIB_LAPACK_DIR); ./lapack_testing.py -r )
+endif
+
+lapack-runtest:
+	( cd $(NETLIB_LAPACK_DIR)/INSTALL; ./testlsame; ./testslamch; ./testdlamch; \
+        ./testsecond; ./testdsecnd; ./testieee; ./testversion )
+	(cd $(NETLIB_LAPACK_DIR); ./lapack_testing.py -r )
+
 
 blas-test:
 	(cd $(NETLIB_LAPACK_DIR)/BLAS && rm -f x* *.out)

Makefile.arm

@@ -26,8 +26,8 @@ endif
 
 
 ifeq ($(CORE), ARMV5)
-CCOMMON_OPT += -marm -mfpu=vfp -mfloat-abi=hard  -march=armv6
-FCOMMON_OPT += -marm -mfpu=vfp -mfloat-abi=hard  -march=armv6
+CCOMMON_OPT += -marm -march=armv5
+FCOMMON_OPT += -marm -march=armv5
 endif
 
 

Makefile.arm64

@@ -4,4 +4,8 @@ CCOMMON_OPT += -march=armv8-a
 FCOMMON_OPT += -march=armv8-a
 endif
 
+ifeq ($(CORE), CORTEXA57)
+CCOMMON_OPT += -march=armv8-a+crc+crypto+fp+simd -mtune=cortex-a57
+FCOMMON_OPT += -march=armv8-a+crc+crypto+fp+simd -mtune=cortex-a57
+endif
 

Makefile.install

@@ -11,6 +11,7 @@ OPENBLAS_BINARY_DIR := $(PREFIX)/bin
 OPENBLAS_BUILD_DIR := $(CURDIR)
 OPENBLAS_CMAKE_DIR := $(OPENBLAS_LIBRARY_DIR)/cmake/openblas
 OPENBLAS_CMAKE_CONFIG := OpenBLASConfig.cmake
+OPENBLAS_CMAKE_CONFIG_VERSION := OpenBLASConfigVersion.cmake
 
 .PHONY : install
 .NOTPARALLEL : install
@@ -97,6 +98,7 @@ endif
 	@echo Generating $(OPENBLAS_CMAKE_CONFIG) in $(DESTDIR)$(OPENBLAS_CMAKE_DIR)
 	@echo "SET(OpenBLAS_VERSION \"${VERSION}\")" > $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG)	
 	@echo "SET(OpenBLAS_INCLUDE_DIRS ${OPENBLAS_INCLUDE_DIR})" >> $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG)
+
 ifndef NO_SHARED
 #ifeq logical or
 ifeq ($(OSNAME), $(filter $(OSNAME),Linux FreeBSD NetBSD))
@@ -112,5 +114,16 @@ else
 #only static
 	@echo "SET(OpenBLAS_LIBRARIES ${OPENBLAS_LIBRARY_DIR}/$(LIBPREFIX).$(LIBSUFFIX))" >> $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG)
 endif
+#Generating OpenBLASConfigVersion.cmake
+	@echo Generating $(OPENBLAS_CMAKE_CONFIG_VERSION) in $(DESTDIR)$(OPENBLAS_CMAKE_DIR)
+	@echo "set (PACKAGE_VERSION \"${VERSION}\")" > $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)
+	@echo "if (PACKAGE_VERSION VERSION_LESS PACKAGE_FIND_VERSION)" >> $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)
+	@echo "  set (PACKAGE_VERSION_COMPATIBLE FALSE)" >> $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)
+	@echo "else ()" >> $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)
+	@echo "  set (PACKAGE_VERSION_COMPATIBLE TRUE)" >> $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)
+	@echo "  if (PACKAGE_FIND_VERSION STREQUAL PACKAGE_VERSION)" >> $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)
+	@echo "    set (PACKAGE_VERSION_EXACT TRUE)" >> $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)
+	@echo "  endif ()" >> $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)
+	@echo "endif ()" >> $(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)
 	@echo Install OK!
 

Makefile.rule

@@ -3,7 +3,7 @@
 #
 
 # This library's version
-VERSION = 0.2.14
+VERSION = 0.2.16.dev
 
 # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a
 # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library
@@ -80,7 +80,7 @@ VERSION = 0.2.14
 # NO_LAPACKE = 1
 
 # If you want to use legacy threaded Level 3 implementation.
-# USE_SIMPLE_THREADED_LEVEL3 = 1
+ USE_SIMPLE_THREADED_LEVEL3 = 1
 
 # If you want to drive whole 64bit region by BLAS. Not all Fortran
 # compiler supports this. It's safe to keep comment it out if you
@@ -169,6 +169,9 @@ COMMON_PROF = -pg
 # 64 bit integer interfaces in OpenBLAS.
 # For details, https://github.com/xianyi/OpenBLAS/pull/459
 #
+# The same prefix and suffix are also added to the library name,
+# i.e. you get lib$(SYMBOLPREFIX)openblas$(SYMBOLSUFFIX) rather than libopenblas
+#
 # SYMBOLPREFIX=
 # SYMBOLSUFFIX=
 

Makefile.system

@@ -336,6 +336,11 @@ ifeq ($(ARCH), x86)
 ifndef BINARY
 NO_BINARY_MODE	= 1
 endif
+
+ifeq ($(CORE), generic)
+NO_EXPRECISION = 1
+endif
+
 ifndef NO_EXPRECISION
 ifeq ($(F_COMPILER), GFORTRAN)
 # ifeq logical or. GCC or LSB
@@ -354,6 +359,11 @@ endif
 endif
 
 ifeq ($(ARCH), x86_64)
+
+ifeq ($(CORE), generic)
+NO_EXPRECISION = 1
+endif
+
 ifndef NO_EXPRECISION
 ifeq ($(F_COMPILER), GFORTRAN)
 # ifeq logical or. GCC or LSB
@@ -870,12 +880,6 @@ ifdef USE_SIMPLE_THREADED_LEVEL3
 CCOMMON_OPT	+= -DUSE_SIMPLE_THREADED_LEVEL3
 endif
 
-ifndef LIBNAMESUFFIX
-LIBPREFIX = libopenblas
-else
-LIBPREFIX = libopenblas_$(LIBNAMESUFFIX)
-endif
-
 ifndef SYMBOLPREFIX
 SYMBOLPREFIX =
 endif
@@ -884,6 +888,12 @@ ifndef SYMBOLSUFFIX
 SYMBOLSUFFIX =
 endif
 
+ifndef LIBNAMESUFFIX
+LIBPREFIX = lib$(SYMBOLPREFIX)openblas$(SYMBOLSUFFIX)
+else
+LIBPREFIX = lib$(SYMBOLPREFIX)openblas$(SYMBOLSUFFIX)_$(LIBNAMESUFFIX)
+endif
+
 KERNELDIR	= $(TOPDIR)/kernel/$(ARCH)
 
 include $(TOPDIR)/Makefile.$(ARCH)
@@ -961,16 +971,29 @@ ifeq ($(DEBUG), 1)
 COMMON_OPT += -g
 endif
 
+ifeq ($(DEBUG), 1)
+FCOMMON_OPT += -g
+endif
+
 ifndef COMMON_OPT
 COMMON_OPT = -O2
 endif
 
+ifndef FCOMMON_OPT
+ifeq ($(OSNAME), WINNT)
+FCOMMON_OPT = -O0 
+else
+FCOMMON_OPT = -O2 -frecursive
+endif
+endif
+
+
 
 override CFLAGS     += $(COMMON_OPT) $(CCOMMON_OPT) -I$(TOPDIR)
 override PFLAGS     += $(COMMON_OPT) $(CCOMMON_OPT) -I$(TOPDIR) -DPROFILE $(COMMON_PROF)
 
-override FFLAGS     += $(COMMON_OPT) $(FCOMMON_OPT)
-override FPFLAGS    += $(COMMON_OPT) $(FCOMMON_OPT) $(COMMON_PROF)
+override FFLAGS     += $(FCOMMON_OPT)
+override FPFLAGS    += $(FCOMMON_OPT) $(COMMON_PROF)
 #MAKEOVERRIDES =
 
 #For LAPACK Fortran codes.

README.md

@@ -1,7 +1,10 @@
 # OpenBLAS
 
-[![Build Status](https://travis-ci.org/xianyi/OpenBLAS.png?branch=develop)](https://travis-ci.org/xianyi/OpenBLAS)
+[![Join the chat at https://gitter.im/xianyi/OpenBLAS](https://badges.gitter.im/Join%20Chat.svg)](https://gitter.im/xianyi/OpenBLAS?utm_source=badge&utm_medium=badge&utm_campaign=pr-badge&utm_content=badge)
 
+Travis CI: [![Build Status](https://travis-ci.org/xianyi/OpenBLAS.png?branch=develop)](https://travis-ci.org/xianyi/OpenBLAS)
+
+AppVeyor: [![Build status](https://ci.appveyor.com/api/projects/status/09sohd35n8nkkx64/branch/develop?svg=true)](https://ci.appveyor.com/project/xianyi/openblas/branch/develop)
 ## Introduction
 OpenBLAS is an optimized BLAS library based on GotoBLAS2 1.13 BSD version.
 

TargetList.txt

@@ -44,6 +44,8 @@ NANO
 POWER4
 POWER5
 POWER6
+POWER7
+POWER8
 PPCG4
 PPC970
 PPC970MP
@@ -72,3 +74,5 @@ ARMV5
 
 7.ARM 64-bit CPU:
 ARMV8
+CORTEXA57
+

appveyor.yml

@@ -0,0 +1,42 @@
+version: 0.2.15.{build}
+
+#environment:
+
+platform: 
+  - x64
+
+configuration: Release
+
+clone_folder: c:\projects\OpenBLAS
+
+init:
+  - git config --global core.autocrlf input
+
+build:
+  project: OpenBLAS.sln
+
+clone_depth: 5
+
+#branches to build
+branches:
+  only:
+    - master
+    - develop
+    - cmake
+
+skip_tags: true
+
+matrix:
+  fast_finish: true
+
+skip_commits:
+# Add [av skip] to commit messages
+  message: /\[av skip\]/
+
+before_build:
+  - echo Running cmake...
+  - cd c:\projects\OpenBLAS
+  - cmake -G "Visual Studio 12 Win64" .
+
+test_script:
+  - echo Build OK!

benchmark/gemm.c

@@ -172,7 +172,7 @@ int main(int argc, char *argv[]){
   srandom(getpid());
 #endif
  
-	for(j = 0; j < m; j++){
+	for(j = 0; j < to; j++){
       		for(i = 0; i < to * COMPSIZE; i++){
 			a[i + j * to * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
 			b[i + j * to * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;

c_check

@@ -6,6 +6,7 @@ $hostarch = `uname -m | sed -e s/i.86/x86/`;chop($hostarch);
 $hostarch = "x86_64" if ($hostarch eq "amd64");
 $hostarch = "arm" if ($hostarch =~ /^arm.*/);
 $hostarch = "arm64" if ($hostarch eq "aarch64");
+$hostarch = "power" if ($hostarch =~ /^(powerpc|ppc).*/);
 
 $binary = $ENV{"BINARY"};
 
@@ -30,7 +31,7 @@ if ($ARGV[0] =~ /(.*)(-[.\d]+)/) {
 	$cross_suffix = $1;
     }
 } else {
-    if ($ARGV[0] =~ /(.*-)(.*)/) {
+    if ($ARGV[0] =~ /([^\/]*-)([^\/]*$)/) {
 	$cross_suffix = $1;
     }
 }

cblas_noconst.h

@@ -1,350 +0,0 @@
-#ifndef CBLAS_H
-#define CBLAS_H
-
-#include <stddef.h>
-#include "common.h"
-
-#ifdef __cplusplus
-extern "C" {
-	/* Assume C declarations for C++ */
-#endif  /* __cplusplus */
-
-/*Set the number of threads on runtime.*/
-void openblas_set_num_threads(int num_threads);
-void goto_set_num_threads(int num_threads);
-
-/*Get the number of threads on runtime.*/
-int openblas_get_num_threads(void);
-
-/*Get the number of physical processors (cores).*/
-int openblas_get_num_procs(void);
-
-/*Get the build configure on runtime.*/
-char* openblas_get_config(void);
-
-/* Get the parallelization type which is used by OpenBLAS */
-int openblas_get_parallel(void);
-/* OpenBLAS is compiled for sequential use  */
-#define OPENBLAS_SEQUENTIAL  0
-/* OpenBLAS is compiled using normal threading model */
-#define OPENBLAS_THREAD  1
-/* OpenBLAS is compiled using OpenMP threading model */
-#define OPENBLAS_OPENMP 2
-
-
-#define CBLAS_INDEX size_t
-
-typedef enum CBLAS_ORDER     {CblasRowMajor=101, CblasColMajor=102} CBLAS_ORDER;
-typedef enum CBLAS_TRANSPOSE {CblasNoTrans=111, CblasTrans=112, CblasConjTrans=113, CblasConjNoTrans=114} CBLAS_TRANSPOSE;
-typedef enum CBLAS_UPLO      {CblasUpper=121, CblasLower=122} CBLAS_UPLO;
-typedef enum CBLAS_DIAG      {CblasNonUnit=131, CblasUnit=132} CBLAS_DIAG;
-typedef enum CBLAS_SIDE      {CblasLeft=141, CblasRight=142} CBLAS_SIDE;
-
-float  cblas_sdsdot(blasint n, float alpha, float *x, blasint incx, float *y, blasint incy);
-double cblas_dsdot (blasint n, float *x, blasint incx, float *y, blasint incy);
-float  cblas_sdot(blasint n, float  *x, blasint incx, float  *y, blasint incy);
-double cblas_ddot(blasint n, double *x, blasint incx, double *y, blasint incy);
-
-openblas_complex_float  cblas_cdotu(blasint n, float  *x, blasint incx, float  *y, blasint incy);
-openblas_complex_float  cblas_cdotc(blasint n, float  *x, blasint incx, float  *y, blasint incy);
-openblas_complex_double cblas_zdotu(blasint n, double *x, blasint incx, double *y, blasint incy);
-openblas_complex_double cblas_zdotc(blasint n, double *x, blasint incx, double *y, blasint incy);
-
-void  cblas_cdotu_sub(blasint n, float  *x, blasint incx, float  *y, blasint incy, openblas_complex_float  *ret);
-void  cblas_cdotc_sub(blasint n, float  *x, blasint incx, float  *y, blasint incy, openblas_complex_float  *ret);
-void  cblas_zdotu_sub(blasint n, double *x, blasint incx, double *y, blasint incy, openblas_complex_double *ret);
-void  cblas_zdotc_sub(blasint n, double *x, blasint incx, double *y, blasint incy, openblas_complex_double *ret);
-
-float  cblas_sasum (blasint n, float  *x, blasint incx);
-double cblas_dasum (blasint n, double *x, blasint incx);
-float  cblas_scasum(blasint n, float  *x, blasint incx);
-double cblas_dzasum(blasint n, double *x, blasint incx);
-
-float  cblas_snrm2 (blasint N, float  *X, blasint incX);
-double cblas_dnrm2 (blasint N, double *X, blasint incX);
-float  cblas_scnrm2(blasint N, float  *X, blasint incX);
-double cblas_dznrm2(blasint N, double *X, blasint incX);
-
-CBLAS_INDEX cblas_isamax(blasint n, float  *x, blasint incx);
-CBLAS_INDEX cblas_idamax(blasint n, double *x, blasint incx);
-CBLAS_INDEX cblas_icamax(blasint n, float  *x, blasint incx);
-CBLAS_INDEX cblas_izamax(blasint n, double *x, blasint incx);
-
-void cblas_saxpy(blasint n, float alpha, float *x, blasint incx, float *y, blasint incy);
-void cblas_daxpy(blasint n, double alpha, double *x, blasint incx, double *y, blasint incy);
-void cblas_caxpy(blasint n, float *alpha, float *x, blasint incx, float *y, blasint incy);
-void cblas_zaxpy(blasint n, double *alpha, double *x, blasint incx, double *y, blasint incy);
-
-void cblas_scopy(blasint n, float *x, blasint incx, float *y, blasint incy);
-void cblas_dcopy(blasint n, double *x, blasint incx, double *y, blasint incy);
-void cblas_ccopy(blasint n, float *x, blasint incx, float *y, blasint incy);
-void cblas_zcopy(blasint n, double *x, blasint incx, double *y, blasint incy);
-
-void cblas_sswap(blasint n, float *x, blasint incx, float *y, blasint incy);
-void cblas_dswap(blasint n, double *x, blasint incx, double *y, blasint incy);
-void cblas_cswap(blasint n, float *x, blasint incx, float *y, blasint incy);
-void cblas_zswap(blasint n, double *x, blasint incx, double *y, blasint incy);
-
-void cblas_srot(blasint N, float *X, blasint incX, float *Y, blasint incY, float c, float s);
-void cblas_drot(blasint N, double *X, blasint incX, double *Y, blasint incY, double c, double  s);
-
-void cblas_srotg(float *a, float *b, float *c, float *s);
-void cblas_drotg(double *a, double *b, double *c, double *s);
-
-void cblas_srotm(blasint N, float *X, blasint incX, float *Y, blasint incY, float *P);
-void cblas_drotm(blasint N, double *X, blasint incX, double *Y, blasint incY, double *P);
-
-void cblas_srotmg(float *d1, float *d2, float *b1, float b2, float *P);
-void cblas_drotmg(double *d1, double *d2, double *b1, double b2, double *P);
-
-void cblas_sscal(blasint N, float alpha, float *X, blasint incX);
-void cblas_dscal(blasint N, double alpha, double *X, blasint incX);
-void cblas_cscal(blasint N, float *alpha, float *X, blasint incX);
-void cblas_zscal(blasint N, double *alpha, double *X, blasint incX);
-void cblas_csscal(blasint N, float alpha, float *X, blasint incX);
-void cblas_zdscal(blasint N, double alpha, double *X, blasint incX);
-
-void cblas_sgemv(enum CBLAS_ORDER order,  enum CBLAS_TRANSPOSE trans,  blasint m, blasint n,
-		 float alpha, float  *a, blasint lda,  float  *x, blasint incx,  float beta,  float  *y, blasint incy);
-void cblas_dgemv(enum CBLAS_ORDER order,  enum CBLAS_TRANSPOSE trans,  blasint m, blasint n,
-		 double alpha, double  *a, blasint lda,  double  *x, blasint incx,  double beta,  double  *y, blasint incy);
-void cblas_cgemv(enum CBLAS_ORDER order,  enum CBLAS_TRANSPOSE trans,  blasint m, blasint n,
-		 float *alpha, float  *a, blasint lda,  float  *x, blasint incx,  float *beta,  float  *y, blasint incy);
-void cblas_zgemv(enum CBLAS_ORDER order,  enum CBLAS_TRANSPOSE trans,  blasint m, blasint n,
-		 double *alpha, double  *a, blasint lda,  double  *x, blasint incx,  double *beta,  double  *y, blasint incy);
-
-void cblas_sger (enum CBLAS_ORDER order, blasint M, blasint N, float   alpha, float  *X, blasint incX, float  *Y, blasint incY, float  *A, blasint lda);
-void cblas_dger (enum CBLAS_ORDER order, blasint M, blasint N, double  alpha, double *X, blasint incX, double *Y, blasint incY, double *A, blasint lda);
-void cblas_cgeru(enum CBLAS_ORDER order, blasint M, blasint N, float  *alpha, float  *X, blasint incX, float  *Y, blasint incY, float  *A, blasint lda);
-void cblas_cgerc(enum CBLAS_ORDER order, blasint M, blasint N, float  *alpha, float  *X, blasint incX, float  *Y, blasint incY, float  *A, blasint lda);
-void cblas_zgeru(enum CBLAS_ORDER order, blasint M, blasint N, double *alpha, double *X, blasint incX, double *Y, blasint incY, double *A, blasint lda);
-void cblas_zgerc(enum CBLAS_ORDER order, blasint M, blasint N, double *alpha, double *X, blasint incX, double *Y, blasint incY, double *A, blasint lda);
-
-void cblas_strsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag, blasint N, float *A, blasint lda, float *X, blasint incX);
-void cblas_dtrsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag, blasint N, double *A, blasint lda, double *X, blasint incX);
-void cblas_ctrsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag, blasint N, float *A, blasint lda, float *X, blasint incX);
-void cblas_ztrsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag, blasint N, double *A, blasint lda, double *X, blasint incX);
-
-void cblas_strmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag, blasint N, float *A, blasint lda, float *X, blasint incX);
-void cblas_dtrmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag, blasint N, double *A, blasint lda, double *X, blasint incX);
-void cblas_ctrmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag, blasint N, float *A, blasint lda, float *X, blasint incX);
-void cblas_ztrmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag, blasint N, double *A, blasint lda, double *X, blasint incX);
-
-void cblas_ssyr(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, float alpha, float *X, blasint incX, float *A, blasint lda);
-void cblas_dsyr(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, double alpha, double *X, blasint incX, double *A, blasint lda);
-void cblas_cher(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, float alpha, float *X, blasint incX, float *A, blasint lda);
-void cblas_zher(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, double alpha, double *X, blasint incX, double *A, blasint lda);
-
-void cblas_ssyr2(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,blasint N, float alpha, float *X,
-                blasint incX, float *Y, blasint incY, float *A, blasint lda);
-void cblas_dsyr2(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, double alpha, double *X,
-                blasint incX, double *Y, blasint incY, double *A, blasint lda);
-void cblas_cher2(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, float *alpha, float *X, blasint incX,
-                float *Y, blasint incY, float *A, blasint lda);
-void cblas_zher2(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, double *alpha, double *X, blasint incX,
-                double *Y, blasint incY, double *A, blasint lda);
-
-void cblas_sgbmv(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, blasint M, blasint N,
-                 blasint KL, blasint KU, float alpha, float *A, blasint lda, float *X, blasint incX, float beta, float *Y, blasint incY);
-void cblas_dgbmv(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, blasint M, blasint N,
-                 blasint KL, blasint KU, double alpha, double *A, blasint lda, double *X, blasint incX, double beta, double *Y, blasint incY);
-void cblas_cgbmv(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, blasint M, blasint N,
-                 blasint KL, blasint KU, float *alpha, float *A, blasint lda, float *X, blasint incX, float *beta, float *Y, blasint incY);
-void cblas_zgbmv(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, blasint M, blasint N,
-                 blasint KL, blasint KU, double *alpha, double *A, blasint lda, double *X, blasint incX, double *beta, double *Y, blasint incY);
-
-void cblas_ssbmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, blasint K, float alpha, float *A,
-                 blasint lda, float *X, blasint incX, float beta, float *Y, blasint incY);
-void cblas_dsbmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, blasint K, double alpha, double *A,
-                 blasint lda, double *X, blasint incX, double beta, double *Y, blasint incY);
-
-
-void cblas_stbmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, blasint K, float *A, blasint lda, float *X, blasint incX);
-void cblas_dtbmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, blasint K, double *A, blasint lda, double *X, blasint incX);
-void cblas_ctbmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, blasint K, float *A, blasint lda, float *X, blasint incX);
-void cblas_ztbmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, blasint K, double *A, blasint lda, double *X, blasint incX);
-
-void cblas_stbsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, blasint K, float *A, blasint lda, float *X, blasint incX);
-void cblas_dtbsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, blasint K, double *A, blasint lda, double *X, blasint incX);
-void cblas_ctbsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, blasint K, float *A, blasint lda, float *X, blasint incX);
-void cblas_ztbsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, blasint K, double *A, blasint lda, double *X, blasint incX);
-
-void cblas_stpmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, float *Ap, float *X, blasint incX);
-void cblas_dtpmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, double *Ap, double *X, blasint incX);
-void cblas_ctpmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, float *Ap, float *X, blasint incX);
-void cblas_ztpmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, double *Ap, double *X, blasint incX);
-
-void cblas_stpsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, float *Ap, float *X, blasint incX);
-void cblas_dtpsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, double *Ap, double *X, blasint incX);
-void cblas_ctpsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, float *Ap, float *X, blasint incX);
-void cblas_ztpsv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
-                 blasint N, double *Ap, double *X, blasint incX);
-
-void cblas_ssymv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, float alpha, float *A,
-                 blasint lda, float *X, blasint incX, float beta, float *Y, blasint incY);
-void cblas_dsymv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, double alpha, double *A,
-                 blasint lda, double *X, blasint incX, double beta, double *Y, blasint incY);
-void cblas_chemv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, float *alpha, float *A,
-                 blasint lda, float *X, blasint incX, float *beta, float *Y, blasint incY);
-void cblas_zhemv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, double *alpha, double *A,
-                 blasint lda, double *X, blasint incX, double *beta, double *Y, blasint incY);
-
-
-void cblas_sspmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, float alpha, float *Ap,
-                 float *X, blasint incX, float beta, float *Y, blasint incY);
-void cblas_dspmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, double alpha, double *Ap,
-                 double *X, blasint incX, double beta, double *Y, blasint incY);
-
-void cblas_sspr(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, float alpha, float *X, blasint incX, float *Ap);
-void cblas_dspr(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, double alpha, double *X, blasint incX, double *Ap);
-
-void cblas_chpr(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, float alpha, float *X, blasint incX, float *A);
-void cblas_zhpr(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, double alpha, double *X,blasint incX, double *A);
-
-void cblas_sspr2(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, float alpha, float *X, blasint incX, float *Y, blasint incY, float *A);
-void cblas_dspr2(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, double alpha, double *X, blasint incX, double *Y, blasint incY, double *A);
-void cblas_chpr2(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, float *alpha, float *X, blasint incX, float *Y, blasint incY, float *Ap);
-void cblas_zhpr2(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, double *alpha, double *X, blasint incX, double *Y, blasint incY, double *Ap);
-
-void cblas_chbmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, blasint K,
-		 float *alpha, float *A, blasint lda, float *X, blasint incX, float *beta, float *Y, blasint incY);
-void cblas_zhbmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N, blasint K,
-		 double *alpha, double *A, blasint lda, double *X, blasint incX, double *beta, double *Y, blasint incY);
-
-void cblas_chpmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N,
-		 float *alpha, float *Ap, float *X, blasint incX, float *beta, float *Y, blasint incY);
-void cblas_zhpmv(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint N,
-		 double *alpha, double *Ap, double *X, blasint incX, double *beta, double *Y, blasint incY);
-
-void cblas_sgemm(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, blasint M, blasint N, blasint K,
-		 float alpha, float *A, blasint lda, float *B, blasint ldb, float beta, float *C, blasint ldc);
-void cblas_dgemm(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, blasint M, blasint N, blasint K,
-		 double alpha, double *A, blasint lda, double *B, blasint ldb, double beta, double *C, blasint ldc);
-void cblas_cgemm(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, blasint M, blasint N, blasint K,
-		 float *alpha, float *A, blasint lda, float *B, blasint ldb, float *beta, float *C, blasint ldc);
-void cblas_cgemm3m(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, blasint M, blasint N, blasint K,
-		 float *alpha, float *A, blasint lda, float *B, blasint ldb, float *beta, float *C, blasint ldc);
-void cblas_zgemm(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, blasint M, blasint N, blasint K,
-		 double *alpha, double *A, blasint lda, double *B, blasint ldb, double *beta, double *C, blasint ldc);
-void cblas_zgemm3m(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, blasint M, blasint N, blasint K,
-		 double *alpha, double *A, blasint lda, double *B, blasint ldb, double *beta, double *C, blasint ldc);
-
-void cblas_ssymm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, blasint M, blasint N,
-                 float alpha, float *A, blasint lda, float *B, blasint ldb, float beta, float *C, blasint ldc);
-void cblas_dsymm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, blasint M, blasint N,
-                 double alpha, double *A, blasint lda, double *B, blasint ldb, double beta, double *C, blasint ldc);
-void cblas_csymm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, blasint M, blasint N,
-                 float *alpha, float *A, blasint lda, float *B, blasint ldb, float *beta, float *C, blasint ldc);
-void cblas_zsymm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, blasint M, blasint N,
-                 double *alpha, double *A, blasint lda, double *B, blasint ldb, double *beta, double *C, blasint ldc);
-
-void cblas_ssyrk(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans,
-		 blasint N, blasint K, float alpha, float *A, blasint lda, float beta, float *C, blasint ldc);
-void cblas_dsyrk(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans,
-		 blasint N, blasint K, double alpha, double *A, blasint lda, double beta, double *C, blasint ldc);
-void cblas_csyrk(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans,
-		 blasint N, blasint K, float *alpha, float *A, blasint lda, float *beta, float *C, blasint ldc);
-void cblas_zsyrk(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans,
-		 blasint N, blasint K, double *alpha, double *A, blasint lda, double *beta, double *C, blasint ldc);
-
-void cblas_ssyr2k(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans,
-		  blasint N, blasint K, float alpha, float *A, blasint lda, float *B, blasint ldb, float beta, float *C, blasint ldc);
-void cblas_dsyr2k(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans,
-		  blasint N, blasint K, double alpha, double *A, blasint lda, double *B, blasint ldb, double beta, double *C, blasint ldc);
-void cblas_csyr2k(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans,
-		  blasint N, blasint K, float *alpha, float *A, blasint lda, float *B, blasint ldb, float *beta, float *C, blasint ldc);
-void cblas_zsyr2k(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans,
-		  blasint N, blasint K, double *alpha, double *A, blasint lda, double *B, blasint ldb, double *beta, double *C, blasint ldc);
-
-void cblas_strmm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA,
-                 enum CBLAS_DIAG Diag, blasint M, blasint N, float alpha, float *A, blasint lda, float *B, blasint ldb);
-void cblas_dtrmm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA,
-                 enum CBLAS_DIAG Diag, blasint M, blasint N, double alpha, double *A, blasint lda, double *B, blasint ldb);
-void cblas_ctrmm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA,
-                 enum CBLAS_DIAG Diag, blasint M, blasint N, float *alpha, float *A, blasint lda, float *B, blasint ldb);
-void cblas_ztrmm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA,
-                 enum CBLAS_DIAG Diag, blasint M, blasint N, double *alpha, double *A, blasint lda, double *B, blasint ldb);
-
-void cblas_strsm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA,
-                 enum CBLAS_DIAG Diag, blasint M, blasint N, float alpha, float *A, blasint lda, float *B, blasint ldb);
-void cblas_dtrsm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA,
-                 enum CBLAS_DIAG Diag, blasint M, blasint N, double alpha, double *A, blasint lda, double *B, blasint ldb);
-void cblas_ctrsm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA,
-                 enum CBLAS_DIAG Diag, blasint M, blasint N, float *alpha, float *A, blasint lda, float *B, blasint ldb);
-void cblas_ztrsm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE TransA,
-                 enum CBLAS_DIAG Diag, blasint M, blasint N, double *alpha, double *A, blasint lda, double *B, blasint ldb);
-
-void cblas_chemm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, blasint M, blasint N,
-                 float *alpha, float *A, blasint lda, float *B, blasint ldb, float *beta, float *C, blasint ldc);
-void cblas_zhemm(enum CBLAS_ORDER Order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo, blasint M, blasint N,
-                 double *alpha, double *A, blasint lda, double *B, blasint ldb, double *beta, double *C, blasint ldc);
-
-void cblas_cherk(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans, blasint N, blasint K,
-                 float alpha, float *A, blasint lda, float beta, float *C, blasint ldc);
-void cblas_zherk(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans, blasint N, blasint K,
-                 double alpha, double *A, blasint lda, double beta, double *C, blasint ldc);
-
-void cblas_cher2k(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans, blasint N, blasint K,
-                  float *alpha, float *A, blasint lda, float *B, blasint ldb, float beta, float *C, blasint ldc);
-void cblas_zher2k(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, enum CBLAS_TRANSPOSE Trans, blasint N, blasint K,
-                  double *alpha, double *A, blasint lda, double *B, blasint ldb, double beta, double *C, blasint ldc);
-
-void cblas_xerbla(blasint p, char *rout, char *form, ...);
-
-/*** BLAS extensions ***/
-
-void cblas_saxpby(blasint n, float alpha, float *x, blasint incx,float beta, float *y, blasint incy);
-
-void cblas_daxpby(blasint n, double alpha, double *x, blasint incx,double beta, double *y, blasint incy);
-
-void cblas_caxpby(blasint n, float *alpha, float *x, blasint incx,float *beta, float *y, blasint incy);
-
-void cblas_zaxpby(blasint n, double *alpha, double *x, blasint incx,double *beta, double *y, blasint incy);
-
-void cblas_somatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  float calpha,  float *a, 
-		      blasint clda, float *b,  blasint cldb); 
-void cblas_domatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  double calpha,  double *a,
-		      blasint clda, double *b,  blasint cldb); 
-void cblas_comatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  void* calpha,  void* a, 
-		      blasint clda, void *b,  blasint cldb); 
-void cblas_zomatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  void* calpha,  void* a, 
-		      blasint clda,  void *b,  blasint cldb); 
-
-void cblas_simatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  float calpha, float *a, 
-		      blasint clda,  blasint cldb); 
-void cblas_dimatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  double calpha, double *a,
-		      blasint clda,  blasint cldb); 
-void cblas_cimatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  float* calpha, float* a, 
-		      blasint clda,  blasint cldb); 
-void cblas_zimatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  double* calpha, double* a, 
-		      blasint clda,  blasint cldb); 
-
-void cblas_sgeadd( enum CBLAS_ORDER CORDER, blasint crows,  blasint ccols,  float calpha, float *a,  blasint clda,  float cbeta, 
-		  float *c,  blasint cldc); 
-void cblas_dgeadd( enum CBLAS_ORDER CORDER, blasint crows,  blasint ccols,  double calpha, double *a,  blasint clda,  double cbeta, 
-		  double *c,  blasint cldc); 
-void cblas_cgeadd( enum CBLAS_ORDER CORDER, blasint crows,  blasint ccols,  float *calpha, float *a,  blasint clda,  float *cbeta, 
-		  float *c,  blasint cldc); 
-void cblas_zgeadd( enum CBLAS_ORDER CORDER, blasint crows,  blasint ccols,  double *calpha, double *a,  blasint clda,  double *cbeta, 
-		  double *c,  blasint cldc); 
-
-#ifdef __cplusplus
-}
-#endif  /* __cplusplus */
-
-#endif

cmake/arch.cmake

@@ -0,0 +1,115 @@
+##
+## Author: Hank Anderson <hank@statease.com>
+## Description: Ported from portion of OpenBLAS/Makefile.system
+##              Sets various variables based on architecture.
+
+if (${ARCH} STREQUAL "x86" OR ${ARCH} STREQUAL "x86_64")
+
+  if (${ARCH} STREQUAL "x86")
+    if (NOT BINARY)
+      set(NO_BINARY_MODE 1)
+    endif ()
+  endif ()
+
+  if (NOT NO_EXPRECISION)
+    if (${F_COMPILER} MATCHES "GFORTRAN")
+      # N.B. I'm not sure if CMake differentiates between GCC and LSB -hpa
+      if (${CMAKE_C_COMPILER} STREQUAL "GNU" OR ${CMAKE_C_COMPILER} STREQUAL "LSB")
+        set(EXPRECISION	1)
+        set(CCOMMON_OPT "${CCOMMON_OPT} -DEXPRECISION -m128bit-long-double")
+        set(FCOMMON_OPT	"${FCOMMON_OPT} -m128bit-long-double")
+      endif ()
+      if (${CMAKE_C_COMPILER} STREQUAL "Clang")
+        set(EXPRECISION	1)
+        set(CCOMMON_OPT "${CCOMMON_OPT} -DEXPRECISION")
+        set(FCOMMON_OPT	"${FCOMMON_OPT} -m128bit-long-double")
+      endif ()
+    endif ()
+  endif ()
+endif ()
+
+if (${CMAKE_C_COMPILER} STREQUAL "Intel")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -wd981")
+endif ()
+
+if (USE_OPENMP)
+
+  if (${CMAKE_C_COMPILER} STREQUAL "GNU" OR ${CMAKE_C_COMPILER} STREQUAL "LSB")
+    set(CCOMMON_OPT "${CCOMMON_OPT} -fopenmp")
+  endif ()
+
+  if (${CMAKE_C_COMPILER} STREQUAL "Clang")
+    message(WARNING "Clang doesn't support OpenMP yet.")
+    set(CCOMMON_OPT "${CCOMMON_OPT} -fopenmp")
+  endif ()
+
+  if (${CMAKE_C_COMPILER} STREQUAL "Intel")
+    set(CCOMMON_OPT "${CCOMMON_OPT} -openmp")
+  endif ()
+
+  if (${CMAKE_C_COMPILER} STREQUAL "PGI")
+    set(CCOMMON_OPT "${CCOMMON_OPT} -mp")
+  endif ()
+
+  if (${CMAKE_C_COMPILER} STREQUAL "OPEN64")
+    set(CCOMMON_OPT "${CCOMMON_OPT} -mp")
+    set(CEXTRALIB "${CEXTRALIB} -lstdc++")
+  endif ()
+
+  if (${CMAKE_C_COMPILER} STREQUAL "PATHSCALE")
+    set(CCOMMON_OPT "${CCOMMON_OPT} -mp")
+  endif ()
+endif ()
+
+
+if (DYNAMIC_ARCH)
+  if (${ARCH} STREQUAL "x86")
+    set(DYNAMIC_CORE "KATMAI COPPERMINE NORTHWOOD PRESCOTT BANIAS CORE2 PENRYN DUNNINGTON NEHALEM ATHLON OPTERON OPTERON_SSE3 BARCELONA BOBCAT ATOM NANO")
+  endif ()
+
+  if (${ARCH} STREQUAL "x86_64")
+    set(DYNAMIC_CORE "PRESCOTT CORE2 PENRYN DUNNINGTON NEHALEM OPTERON OPTERON_SSE3 BARCELONA BOBCAT ATOM NANO")
+    if (NOT NO_AVX)
+      set(DYNAMIC_CORE "${DYNAMIC_CORE} SANDYBRIDGE BULLDOZER PILEDRIVER STEAMROLLER")
+    endif ()
+    if (NOT NO_AVX2)
+      set(DYNAMIC_CORE "${DYNAMIC_CORE} HASWELL")
+    endif ()
+  endif ()
+
+  if (NOT DYNAMIC_CORE)
+    unset(DYNAMIC_ARCH)
+  endif ()
+endif ()
+
+if (${ARCH} STREQUAL "ia64")
+  set(NO_BINARY_MODE 1)
+  set(BINARY_DEFINED 1)
+
+  if (${F_COMPILER} MATCHES "GFORTRAN")
+    if (${CMAKE_C_COMPILER} STREQUAL "GNU")
+      # EXPRECISION	= 1
+      # CCOMMON_OPT	+= -DEXPRECISION
+    endif ()
+  endif ()
+endif ()
+
+if (${ARCH} STREQUAL "mips64")
+  set(NO_BINARY_MODE 1)
+endif ()
+
+if (${ARCH} STREQUAL "alpha")
+  set(NO_BINARY_MODE 1)
+  set(BINARY_DEFINED 1)
+endif ()
+
+if (${ARCH} STREQUAL "arm")
+  set(NO_BINARY_MODE 1)
+  set(BINARY_DEFINED 1)
+endif ()
+
+if (${ARCH} STREQUAL "arm64")
+  set(NO_BINARY_MODE 1)
+  set(BINARY_DEFINED 1)
+endif ()
+

cmake/c_check.cmake

@@ -0,0 +1,89 @@
+##
+## 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)
+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 ${CMAKE_SYSTEM_PROCESSOR})
+if (${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 ()
+
+set(COMPILER_ID ${CMAKE_CXX_COMPILER_ID})
+if (${COMPILER_ID} STREQUAL "GNU")
+  set(COMPILER_ID "GCC")
+endif ()
+
+string(TOUPPER ${ARCH} UC_ARCH)
+
+file(WRITE ${TARGET_CONF}
+  "#define OS_${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")
+

cmake/cc.cmake

@@ -0,0 +1,103 @@
+##
+## Author: Hank Anderson <hank@statease.com>
+## Description: Ported from portion of OpenBLAS/Makefile.system
+##              Sets C related variables.
+
+if (${CMAKE_C_COMPILER} STREQUAL "GNU" OR ${CMAKE_C_COMPILER} STREQUAL "LSB" OR ${CMAKE_C_COMPILER} STREQUAL "Clang")
+
+  set(CCOMMON_OPT "${CCOMMON_OPT} -Wall")
+  set(COMMON_PROF "${COMMON_PROF} -fno-inline")
+  set(NO_UNINITIALIZED_WARN "-Wno-uninitialized")
+
+  if (QUIET_MAKE)
+    set(CCOMMON_OPT "${CCOMMON_OPT} ${NO_UNINITIALIZED_WARN} -Wno-unused")
+  endif ()
+
+  if (NO_BINARY_MODE)
+
+    if (${ARCH} STREQUAL "mips64")
+      if (BINARY64)
+        set(CCOMMON_OPT "${CCOMMON_OPT} -mabi=64")
+      else ()
+        set(CCOMMON_OPT "${CCOMMON_OPT} -mabi=n32")
+      endif ()
+      set(BINARY_DEFINED 1)
+    endif ()
+
+    if (${CORE} STREQUAL "LOONGSON3A")
+      set(CCOMMON_OPT "${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 (${OSNAME} STREQUAL "AIX")
+      set(BINARY_DEFINED 1)
+    endif ()
+  endif ()
+
+  if (NOT BINARY_DEFINED)
+    if (BINARY64)
+      set(CCOMMON_OPT "${CCOMMON_OPT} -m64")
+    else ()
+      set(CCOMMON_OPT "${CCOMMON_OPT} -m32")
+    endif ()
+  endif ()
+endif ()
+
+if (${CMAKE_C_COMPILER} STREQUAL "PGI")
+  if (BINARY64)
+    set(CCOMMON_OPT "${CCOMMON_OPT} -tp p7-64")
+  else ()
+    set(CCOMMON_OPT "${CCOMMON_OPT} -tp p7")
+  endif ()
+endif ()
+
+if (${CMAKE_C_COMPILER} STREQUAL "PATHSCALE")
+  if (BINARY64)
+    set(CCOMMON_OPT "${CCOMMON_OPT} -m64")
+  else ()
+    set(CCOMMON_OPT "${CCOMMON_OPT} -m32")
+  endif ()
+endif ()
+
+if (${CMAKE_C_COMPILER} STREQUAL "OPEN64")
+
+  if (${ARCH} STREQUAL "mips64")
+
+    if (NOT BINARY64)
+      set(CCOMMON_OPT "${CCOMMON_OPT} -n32")
+    else ()
+      set(CCOMMON_OPT "${CCOMMON_OPT} -n64")
+    endif ()
+
+    if (${CORE} STREQUAL "LOONGSON3A")
+      set(CCOMMON_OPT "${CCOMMON_OPT} -loongson3 -static")
+    endif ()
+
+    if (${CORE} STREQUAL "LOONGSON3B")
+      set(CCOMMON_OPT "${CCOMMON_OPT} -loongson3 -static")
+    endif ()
+
+  else ()
+
+    if (BINARY64)
+      set(CCOMMON_OPT "${CCOMMON_OPT} -m32")
+    else ()
+      set(CCOMMON_OPT "${CCOMMON_OPT} -m64")
+    endif ()
+  endif ()
+endif ()
+
+if (${CMAKE_C_COMPILER} STREQUAL "SUN")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -w")
+  if (${ARCH} STREQUAL "x86")
+    set(CCOMMON_OPT "${CCOMMON_OPT} -m32")
+  else ()
+    set(FCOMMON_OPT "${FCOMMON_OPT} -m64")
+  endif ()
+endif ()
+

cmake/export.cmake

@@ -0,0 +1,60 @@
+
+#Only generate .def for dll on MSVC
+if(MSVC)
+
+set_source_files_properties(${OpenBLAS_DEF_FILE} PROPERTIES GENERATED 1)
+
+if (NOT DEFINED ARCH)
+  set(ARCH_IN "x86_64")
+else()
+  set(ARCH_IN ${ARCH})
+endif()
+
+if (${CORE} STREQUAL "generic")
+  set(ARCH_IN "GENERIC")
+endif ()
+
+if (NOT DEFINED EXPRECISION)
+  set(EXPRECISION_IN 0)
+else()
+  set(EXPRECISION_IN ${EXPRECISION})
+endif()
+
+if (NOT DEFINED NO_CBLAS)
+  set(NO_CBLAS_IN 0)
+else()
+  set(NO_CBLAS_IN ${NO_CBLAS})
+endif()
+
+if (NOT DEFINED NO_LAPACK)
+  set(NO_LAPACK_IN 0)
+else()
+  set(NO_LAPACK_IN ${NO_LAPACK})
+endif()
+
+if (NOT DEFINED NO_LAPACKE)
+  set(NO_LAPACKE_IN 0)
+else()
+  set(NO_LAPACKE_IN ${NO_LAPACKE})
+endif()
+
+if (NOT DEFINED NEED2UNDERSCORES)
+  set(NEED2UNDERSCORES_IN 0)
+else()
+  set(NEED2UNDERSCORES_IN ${NEED2UNDERSCORES})
+endif()
+
+if (NOT DEFINED ONLY_CBLAS)
+  set(ONLY_CBLAS_IN 0)
+else()
+  set(ONLY_CBLAS_IN ${ONLY_CBLAS})
+endif()
+
+add_custom_command(
+  TARGET ${OpenBLAS_LIBNAME} PRE_LINK 
+  COMMAND perl 
+  ARGS "${CMAKE_SOURCE_DIR}/exports/gensymbol" "win2k" "${ARCH_IN}" "dummy" "${EXPRECISION_IN}" "${NO_CBLAS_IN}" "${NO_LAPACK_IN}" "${NO_LAPACKE_IN}" "${NEED2UNDERSCORES_IN}" "${ONLY_CBLAS_IN}" "${SYMBOLPREFIX}" "${SYMBOLSUFFIX}" > "${PROJECT_BINARY_DIR}/openblas.def"
+  COMMENT "Create openblas.def file"
+  VERBATIM)
+
+endif()

cmake/f_check.cmake

@@ -0,0 +1,66 @@
+##
+## Author: Hank Anderson <hank@statease.com>
+## Copyright: (c) Stat-Ease, Inc.
+## Created: 12/29/14
+## Last Modified: 12/29/14
+## Description: Ported from the OpenBLAS/f_check perl script.
+##              This is triggered by prebuild.cmake and runs before any of the code is built.
+##              Appends Fortran information to config.h and Makefile.conf.
+
+# CMake vars set by this file:
+# F_COMPILER
+# FC
+# BU
+# NOFORTRAN
+# NEED2UNDERSCORES
+# FEXTRALIB
+
+# Defines set by this file:
+# BUNDERSCORE
+# NEEDBUNDERSCORE
+# NEED2UNDERSCORES
+
+if (MSVC)
+  # had to do this for MSVC, else CMake automatically assumes I have ifort... -hpa
+  include(CMakeForceCompiler)
+  CMAKE_FORCE_Fortran_COMPILER(gfortran GNU)
+endif ()
+
+if (NOT NO_LAPACK)
+  enable_language(Fortran)
+else()
+  include(CMakeForceCompiler)
+  CMAKE_FORCE_Fortran_COMPILER(gfortran GNU)
+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?
+
+  set(BU "_")
+  file(APPEND ${TARGET_CONF}
+    "#define BUNDERSCORE _\n"
+    "#define NEEDBUNDERSCORE 1\n"
+    "#define NEED2UNDERSCORES 0\n")
+
+else ()
+
+  #When we only build CBLAS, we set NOFORTRAN=2
+  set(NOFORTRAN 2)
+  set(NO_FBLAS 1)
+  #set(F_COMPILER GFORTRAN) # CMake handles the fortran compiler
+  set(BU "_")
+  file(APPEND ${TARGET_CONF}
+    "#define BUNDERSCORE _\n"
+    "#define NEEDBUNDERSCORE 1\n")
+endif()
+
+get_filename_component(F_COMPILER ${CMAKE_Fortran_COMPILER} NAME_WE)
+string(TOUPPER ${F_COMPILER} F_COMPILER)
+

cmake/fc.cmake

@@ -0,0 +1,200 @@
+##
+## Author: Hank Anderson <hank@statease.com>
+## Description: Ported from portion of OpenBLAS/Makefile.system
+##              Sets Fortran related variables.
+
+if (${F_COMPILER} STREQUAL "G77")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_G77")
+  set(FCOMMON_OPT "${FCOMMON_OPT} -Wall")
+  if (NOT NO_BINARY_MODE)
+    if (BINARY64)
+      set(FCOMMON_OPT "${FCOMMON_OPT} -m64")
+    else ()
+      set(FCOMMON_OPT "${FCOMMON_OPT} -m32")
+    endif ()
+  endif ()
+endif ()
+
+if (${F_COMPILER} STREQUAL "G95")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_G95")
+  set(FCOMMON_OPT "${FCOMMON_OPT} -Wall")
+  if (NOT NO_BINARY_MODE)
+    if (BINARY64)
+      set(FCOMMON_OPT "${FCOMMON_OPT} -m64")
+    else ()
+      set(FCOMMON_OPT "${FCOMMON_OPT} -m32")
+    endif ()
+  endif ()
+endif ()
+
+if (${F_COMPILER} STREQUAL "GFORTRAN")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_GFORT")
+  set(FCOMMON_OPT "${FCOMMON_OPT} -Wall")
+  #Don't include -lgfortran, when NO_LAPACK=1 or lsbcc
+  if (NOT NO_LAPACK)
+    set(EXTRALIB "{EXTRALIB} -lgfortran")
+  endif ()
+  if (NO_BINARY_MODE)
+    if (${ARCH} STREQUAL "mips64")
+      if (BINARY64)
+        set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=64")
+      else ()
+        set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=n32")
+      endif ()
+    endif ()
+  else ()
+    if (BINARY64)
+      set(FCOMMON_OPT "${FCOMMON_OPT} -m64")
+      if (INTERFACE64)
+        set(FCOMMON_OPT "${FCOMMON_OPT} -fdefault-integer-8")
+      endif ()
+    else ()
+      set(FCOMMON_OPT "${FCOMMON_OPT} -m32")
+    endif ()
+  endif ()
+
+  if (USE_OPENMP)
+    set(FCOMMON_OPT "${FCOMMON_OPT} -fopenmp")
+  endif ()
+endif ()
+
+if (${F_COMPILER} STREQUAL "INTEL")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_INTEL")
+  if (INTERFACE64)
+    set(FCOMMON_OPT "${FCOMMON_OPT} -i8")
+  endif ()
+  if (USE_OPENMP)
+    set(FCOMMON_OPT "${FCOMMON_OPT} -openmp")
+  endif ()
+endif ()
+
+if (${F_COMPILER} STREQUAL "FUJITSU")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_FUJITSU")
+  if (USE_OPENMP)
+    set(FCOMMON_OPT "${FCOMMON_OPT} -openmp")
+  endif ()
+endif ()
+
+if (${F_COMPILER} STREQUAL "IBM")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_IBM")
+  # FCOMMON_OPT	+= -qarch=440
+  if (BINARY64)
+    set(FCOMMON_OPT "${FCOMMON_OPT} -q64")
+    if (INTERFACE64)
+      set(FCOMMON_OPT "${FCOMMON_OPT} -qintsize=8")
+    endif ()
+  else ()
+    set(FCOMMON_OPT "${FCOMMON_OPT} -q32")
+  endif ()
+  if (USE_OPENMP)
+    set(FCOMMON_OPT "${FCOMMON_OPT} -openmp")
+  endif ()
+endif ()
+
+if (${F_COMPILER} STREQUAL "PGI")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_PGI")
+  set(COMMON_PROF "${COMMON_PROF} -DPGICOMPILER")
+  if (BINARY64)
+    if (INTERFACE64)
+      set(FCOMMON_OPT "${FCOMMON_OPT} -i8")
+    endif ()
+    set(FCOMMON_OPT "${FCOMMON_OPT} -tp p7-64")
+  else ()
+    set(FCOMMON_OPT "${FCOMMON_OPT} -tp p7")
+  endif ()
+  if (USE_OPENMP)
+    set(FCOMMON_OPT "${FCOMMON_OPT} -mp")
+  endif ()
+endif ()
+
+if (${F_COMPILER} STREQUAL "PATHSCALE")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_PATHSCALE")
+  if (BINARY64)
+    if (INTERFACE64)
+      set(FCOMMON_OPT "${FCOMMON_OPT} -i8")
+    endif ()
+  endif ()
+
+  if (NOT ${ARCH} STREQUAL "mips64")
+    if (NOT BINARY64)
+      set(FCOMMON_OPT "${FCOMMON_OPT} -m32")
+    else ()
+      set(FCOMMON_OPT "${FCOMMON_OPT} -m64")
+    endif ()
+  else ()
+    if (BINARY64)
+      set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=64")
+    else ()
+    set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=n32")
+    endif ()
+  endif ()
+
+  if (USE_OPENMP)
+    set(FCOMMON_OPT "${FCOMMON_OPT} -mp")
+  endif ()
+endif ()
+
+if (${F_COMPILER} STREQUAL "OPEN64")
+
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_OPEN64")
+  if (BINARY64)
+    if (INTERFACE64)
+      set(FCOMMON_OPT "${FCOMMON_OPT} -i8")
+    endif ()
+  endif ()
+
+  if (${ARCH} STREQUAL "mips64")
+
+    if (NOT BINARY64)
+      set(FCOMMON_OPT "${FCOMMON_OPT} -n32")
+    else ()
+      set(FCOMMON_OPT "${FCOMMON_OPT} -n64")
+    endif ()
+
+    if (${CORE} STREQUAL "LOONGSON3A")
+      set(FCOMMON_OPT "${FCOMMON_OPT} -loongson3 -static")
+    endif ()
+
+    if (${CORE} STREQUAL "LOONGSON3B")
+    set(FCOMMON_OPT "${FCOMMON_OPT} -loongson3 -static")
+    endif ()
+  else ()
+    if (NOT BINARY64)
+      set(FCOMMON_OPT "${FCOMMON_OPT} -m32")
+    else ()
+      set(FCOMMON_OPT "${FCOMMON_OPT} -m64")
+    endif ()
+  endif ()
+
+  if (USE_OPENMP)
+    set(FEXTRALIB "${FEXTRALIB} -lstdc++")
+    set(FCOMMON_OPT "${FCOMMON_OPT} -mp")
+  endif ()
+endif ()
+
+if (${F_COMPILER} STREQUAL "SUN")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_SUN")
+  if (${ARCH} STREQUAL "x86")
+    set(FCOMMON_OPT "${FCOMMON_OPT} -m32")
+  else ()
+    set(FCOMMON_OPT "${FCOMMON_OPT} -m64")
+  endif ()
+  if (USE_OPENMP)
+    set(FCOMMON_OPT "${FCOMMON_OPT} -xopenmp=parallel")
+  endif ()
+endif ()
+
+if (${F_COMPILER} STREQUAL "COMPAQ")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_COMPAQ")
+  if (USE_OPENMP)
+    set(FCOMMON_OPT "${FCOMMON_OPT} -openmp")
+  endif ()
+endif ()
+
+# from the root Makefile - this is for lapack-netlib to compile the correct secnd file.
+if (${F_COMPILER} STREQUAL "GFORTRAN")
+  set(TIMER "INT_ETIME")
+else ()
+  set(TIMER "NONE")
+endif ()
+

cmake/kernel.cmake

@@ -0,0 +1,165 @@
+# helper functions for the kernel CMakeLists.txt
+
+
+# Set the default filenames for L1 objects. Most of these will be overriden by the appropriate KERNEL file.
+macro(SetDefaultL1)
+  set(SAMAXKERNEL amax.S)
+  set(DAMAXKERNEL amax.S)
+  set(QAMAXKERNEL amax.S)
+  set(CAMAXKERNEL zamax.S)
+  set(ZAMAXKERNEL zamax.S)
+  set(XAMAXKERNEL zamax.S)
+  set(SAMINKERNEL amin.S)
+  set(DAMINKERNEL amin.S)
+  set(QAMINKERNEL amin.S)
+  set(CAMINKERNEL zamin.S)
+  set(ZAMINKERNEL zamin.S)
+  set(XAMINKERNEL zamin.S)
+  set(SMAXKERNEL max.S)
+  set(DMAXKERNEL max.S)
+  set(QMAXKERNEL max.S)
+  set(SMINKERNEL min.S)
+  set(DMINKERNEL min.S)
+  set(QMINKERNEL min.S)
+  set(ISAMAXKERNEL iamax.S)
+  set(IDAMAXKERNEL iamax.S)
+  set(IQAMAXKERNEL iamax.S)
+  set(ICAMAXKERNEL izamax.S)
+  set(IZAMAXKERNEL izamax.S)
+  set(IXAMAXKERNEL izamax.S)
+  set(ISAMINKERNEL iamin.S)
+  set(IDAMINKERNEL iamin.S)
+  set(IQAMINKERNEL iamin.S)
+  set(ICAMINKERNEL izamin.S)
+  set(IZAMINKERNEL izamin.S)
+  set(IXAMINKERNEL izamin.S)
+  set(ISMAXKERNEL iamax.S)
+  set(IDMAXKERNEL iamax.S)
+  set(IQMAXKERNEL iamax.S)
+  set(ISMINKERNEL iamin.S)
+  set(IDMINKERNEL iamin.S)
+  set(IQMINKERNEL iamin.S)
+  set(SASUMKERNEL asum.S)
+  set(DASUMKERNEL asum.S)
+  set(CASUMKERNEL zasum.S)
+  set(ZASUMKERNEL zasum.S)
+  set(QASUMKERNEL asum.S)
+  set(XASUMKERNEL zasum.S)
+  set(SAXPYKERNEL axpy.S)
+  set(DAXPYKERNEL axpy.S)
+  set(CAXPYKERNEL zaxpy.S)
+  set(ZAXPYKERNEL zaxpy.S)
+  set(QAXPYKERNEL axpy.S)
+  set(XAXPYKERNEL zaxpy.S)
+  set(SCOPYKERNEL copy.S)
+  set(DCOPYKERNEL copy.S)
+  set(CCOPYKERNEL zcopy.S)
+  set(ZCOPYKERNEL zcopy.S)
+  set(QCOPYKERNEL copy.S)
+  set(XCOPYKERNEL zcopy.S)
+  set(SDOTKERNEL dot.S)
+  set(DDOTKERNEL dot.S)
+  set(CDOTKERNEL zdot.S)
+  set(ZDOTKERNEL zdot.S)
+  set(QDOTKERNEL dot.S)
+  set(XDOTKERNEL zdot.S)
+  set(SNRM2KERNEL nrm2.S)
+  set(DNRM2KERNEL nrm2.S)
+  set(QNRM2KERNEL nrm2.S)
+  set(CNRM2KERNEL znrm2.S)
+  set(ZNRM2KERNEL znrm2.S)
+  set(XNRM2KERNEL znrm2.S)
+  set(SROTKERNEL rot.S)
+  set(DROTKERNEL rot.S)
+  set(QROTKERNEL rot.S)
+  set(CROTKERNEL zrot.S)
+  set(ZROTKERNEL zrot.S)
+  set(XROTKERNEL zrot.S)
+  set(SSCALKERNEL scal.S)
+  set(DSCALKERNEL scal.S)
+  set(CSCALKERNEL zscal.S)
+  set(ZSCALKERNEL zscal.S)
+  set(QSCALKERNEL scal.S)
+  set(XSCALKERNEL zscal.S)
+  set(SSWAPKERNEL swap.S)
+  set(DSWAPKERNEL swap.S)
+  set(CSWAPKERNEL zswap.S)
+  set(ZSWAPKERNEL zswap.S)
+  set(QSWAPKERNEL swap.S)
+  set(XSWAPKERNEL zswap.S)
+  set(SGEMVNKERNEL gemv_n.S)
+  set(SGEMVTKERNEL gemv_t.S)
+  set(DGEMVNKERNEL gemv_n.S)
+  set(DGEMVTKERNEL gemv_t.S)
+  set(CGEMVNKERNEL zgemv_n.S)
+  set(CGEMVTKERNEL zgemv_t.S)
+  set(ZGEMVNKERNEL zgemv_n.S)
+  set(ZGEMVTKERNEL zgemv_t.S)
+  set(QGEMVNKERNEL gemv_n.S)
+  set(QGEMVTKERNEL gemv_t.S)
+  set(XGEMVNKERNEL zgemv_n.S)
+  set(XGEMVTKERNEL zgemv_t.S)
+  set(SCABS_KERNEL ../generic/cabs.c)
+  set(DCABS_KERNEL ../generic/cabs.c)
+  set(QCABS_KERNEL ../generic/cabs.c)
+  set(LSAME_KERNEL ../generic/lsame.c)
+  set(SAXPBYKERNEL ../arm/axpby.c)
+  set(DAXPBYKERNEL ../arm/axpby.c)
+  set(CAXPBYKERNEL ../arm/zaxpby.c)
+  set(ZAXPBYKERNEL ../arm/zaxpby.c)
+endmacro ()
+
+macro(SetDefaultL2)
+  set(SGEMVNKERNEL gemv_n.S)
+  set(SGEMVTKERNEL gemv_t.S)
+  set(DGEMVNKERNEL gemv_n.S)
+  set(DGEMVTKERNEL gemv_t.S)
+  set(CGEMVNKERNEL zgemv_n.S)
+  set(CGEMVTKERNEL zgemv_t.S)
+  set(ZGEMVNKERNEL zgemv_n.S)
+  set(ZGEMVTKERNEL zgemv_t.S)
+  set(QGEMVNKERNEL gemv_n.S)
+  set(QGEMVTKERNEL gemv_t.S)
+  set(XGEMVNKERNEL zgemv_n.S)
+  set(XGEMVTKERNEL zgemv_t.S)
+  set(SGERKERNEL ../generic/ger.c)
+  set(DGERKERNEL ../generic/ger.c)
+  set(QGERKERNEL ../generic/ger.c)
+  set(CGERUKERNEL ../generic/zger.c)
+  set(CGERCKERNEL ../generic/zger.c)
+  set(ZGERUKERNEL ../generic/zger.c)
+  set(ZGERCKERNEL ../generic/zger.c)
+  set(XGERUKERNEL ../generic/zger.c)
+  set(XGERCKERNEL ../generic/zger.c)
+  set(SSYMV_U_KERNEL ../generic/symv_k.c)
+  set(SSYMV_L_KERNEL ../generic/symv_k.c)
+  set(DSYMV_U_KERNEL ../generic/symv_k.c)
+  set(DSYMV_L_KERNEL ../generic/symv_k.c)
+  set(QSYMV_U_KERNEL ../generic/symv_k.c)
+  set(QSYMV_L_KERNEL ../generic/symv_k.c)
+  set(CSYMV_U_KERNEL ../generic/zsymv_k.c)
+  set(CSYMV_L_KERNEL ../generic/zsymv_k.c)
+  set(ZSYMV_U_KERNEL ../generic/zsymv_k.c)
+  set(ZSYMV_L_KERNEL ../generic/zsymv_k.c)
+  set(XSYMV_U_KERNEL ../generic/zsymv_k.c)
+  set(XSYMV_L_KERNEL ../generic/zsymv_k.c)
+  set(CHEMV_U_KERNEL ../generic/zhemv_k.c)
+  set(CHEMV_L_KERNEL ../generic/zhemv_k.c)
+  set(CHEMV_V_KERNEL ../generic/zhemv_k.c)
+  set(CHEMV_M_KERNEL ../generic/zhemv_k.c)
+  set(ZHEMV_U_KERNEL ../generic/zhemv_k.c)
+  set(ZHEMV_L_KERNEL ../generic/zhemv_k.c)
+  set(ZHEMV_V_KERNEL ../generic/zhemv_k.c)
+  set(ZHEMV_M_KERNEL ../generic/zhemv_k.c)
+  set(XHEMV_U_KERNEL ../generic/zhemv_k.c)
+  set(XHEMV_L_KERNEL ../generic/zhemv_k.c)
+  set(XHEMV_V_KERNEL ../generic/zhemv_k.c)
+  set(XHEMV_M_KERNEL ../generic/zhemv_k.c)
+endmacro ()
+
+macro(SetDefaultL3)
+  set(SGEADD_KERNEL ../generic/geadd.c)
+  set(DGEADD_KERNEL ../generic/geadd.c)
+  set(CGEADD_KERNEL ../generic/zgeadd.c)
+  set(ZGEADD_KERNEL ../generic/zgeadd.c)
+endmacro ()

cmake/lapack.cmake

@@ -0,0 +1,347 @@
+# Sources for compiling lapack-netlib. Can't use CMakeLists.txt because lapack-netlib already has its own cmake files.
+
+set(ALLAUX
+  ilaenv.f ieeeck.f lsamen.f xerbla_array.f iparmq.f
+  ilaprec.f ilatrans.f ilauplo.f iladiag.f chla_transtype.f
+  ../INSTALL/ilaver.f ../INSTALL/slamch.f
+)
+
+set(SCLAUX
+  sbdsdc.f
+  sbdsqr.f sdisna.f slabad.f slacpy.f sladiv.f slae2.f  slaebz.f
+  slaed0.f slaed1.f slaed2.f slaed3.f slaed4.f slaed5.f slaed6.f
+  slaed7.f slaed8.f slaed9.f slaeda.f slaev2.f slagtf.f
+  slagts.f slamrg.f slanst.f
+  slapy2.f slapy3.f slarnv.f
+  slarra.f slarrb.f slarrc.f slarrd.f slarre.f slarrf.f slarrj.f
+  slarrk.f slarrr.f slaneg.f
+  slartg.f slaruv.f slas2.f  slascl.f
+  slasd0.f slasd1.f slasd2.f slasd3.f slasd4.f slasd5.f slasd6.f
+  slasd7.f slasd8.f slasda.f slasdq.f slasdt.f
+  slaset.f slasq1.f slasq2.f slasq3.f slasq4.f slasq5.f slasq6.f
+  slasr.f  slasrt.f slassq.f slasv2.f spttrf.f sstebz.f sstedc.f
+  ssteqr.f ssterf.f slaisnan.f sisnan.f
+  slartgp.f slartgs.f
+  ../INSTALL/second_${TIMER}.f
+)
+
+set(DZLAUX
+  dbdsdc.f
+  dbdsqr.f ddisna.f dlabad.f dlacpy.f dladiv.f dlae2.f  dlaebz.f
+  dlaed0.f dlaed1.f dlaed2.f dlaed3.f dlaed4.f dlaed5.f dlaed6.f
+  dlaed7.f dlaed8.f dlaed9.f dlaeda.f dlaev2.f dlagtf.f
+  dlagts.f dlamrg.f dlanst.f
+  dlapy2.f dlapy3.f dlarnv.f
+  dlarra.f dlarrb.f dlarrc.f dlarrd.f dlarre.f dlarrf.f dlarrj.f
+  dlarrk.f dlarrr.f dlaneg.f
+  dlartg.f dlaruv.f dlas2.f  dlascl.f
+  dlasd0.f dlasd1.f dlasd2.f dlasd3.f dlasd4.f dlasd5.f dlasd6.f
+  dlasd7.f dlasd8.f dlasda.f dlasdq.f dlasdt.f
+  dlaset.f dlasq1.f dlasq2.f dlasq3.f dlasq4.f dlasq5.f dlasq6.f
+  dlasr.f  dlasrt.f dlassq.f dlasv2.f dpttrf.f dstebz.f dstedc.f
+  dsteqr.f dsterf.f dlaisnan.f disnan.f
+  dlartgp.f dlartgs.f
+  ../INSTALL/dlamch.f ../INSTALL/dsecnd_${TIMER}.f
+)
+
+set(SLASRC
+  sgbbrd.f sgbcon.f sgbequ.f sgbrfs.f sgbsv.f
+  sgbsvx.f sgbtf2.f sgbtrf.f sgbtrs.f sgebak.f sgebal.f sgebd2.f
+  sgebrd.f sgecon.f sgeequ.f sgees.f  sgeesx.f sgeev.f  sgeevx.f
+  sgegs.f  sgegv.f  sgehd2.f sgehrd.f sgelq2.f sgelqf.f
+  sgels.f  sgelsd.f sgelss.f sgelsx.f sgelsy.f sgeql2.f sgeqlf.f
+  sgeqp3.f sgeqpf.f sgeqr2.f sgeqr2p.f sgeqrf.f sgeqrfp.f sgerfs.f
+  sgerq2.f sgerqf.f sgesc2.f sgesdd.f  sgesvd.f sgesvx.f
+  sgetc2.f sgetri.f
+  sggbak.f sggbal.f sgges.f  sggesx.f sggev.f  sggevx.f
+  sggglm.f sgghrd.f sgglse.f sggqrf.f
+  sggrqf.f sggsvd.f sggsvp.f sgtcon.f sgtrfs.f sgtsv.f
+  sgtsvx.f sgttrf.f sgttrs.f sgtts2.f shgeqz.f
+  shsein.f shseqr.f slabrd.f slacon.f slacn2.f
+  slaein.f slaexc.f slag2.f  slags2.f slagtm.f slagv2.f slahqr.f
+  slahrd.f slahr2.f slaic1.f slaln2.f slals0.f slalsa.f slalsd.f
+  slangb.f slange.f slangt.f slanhs.f slansb.f slansp.f
+  slansy.f slantb.f slantp.f slantr.f slanv2.f
+  slapll.f slapmt.f
+  slaqgb.f slaqge.f slaqp2.f slaqps.f slaqsb.f slaqsp.f slaqsy.f
+  slaqr0.f slaqr1.f slaqr2.f slaqr3.f slaqr4.f slaqr5.f
+  slaqtr.f slar1v.f slar2v.f ilaslr.f ilaslc.f
+  slarf.f  slarfb.f slarfg.f slarfgp.f slarft.f slarfx.f slargv.f
+  slarrv.f slartv.f
+  slarz.f  slarzb.f slarzt.f slasy2.f slasyf.f slasyf_rook.f
+  slatbs.f slatdf.f slatps.f slatrd.f slatrs.f slatrz.f slatzm.f
+  sopgtr.f sopmtr.f sorg2l.f sorg2r.f
+  sorgbr.f sorghr.f sorgl2.f sorglq.f sorgql.f sorgqr.f sorgr2.f
+  sorgrq.f sorgtr.f sorm2l.f sorm2r.f
+  sormbr.f sormhr.f sorml2.f sormlq.f sormql.f sormqr.f sormr2.f
+  sormr3.f sormrq.f sormrz.f sormtr.f spbcon.f spbequ.f spbrfs.f
+  spbstf.f spbsv.f  spbsvx.f
+  spbtf2.f spbtrf.f spbtrs.f spocon.f spoequ.f sporfs.f sposv.f
+  sposvx.f spstrf.f spstf2.f
+  sppcon.f sppequ.f
+  spprfs.f sppsv.f  sppsvx.f spptrf.f spptri.f spptrs.f sptcon.f
+  spteqr.f sptrfs.f sptsv.f  sptsvx.f spttrs.f sptts2.f srscl.f
+  ssbev.f  ssbevd.f ssbevx.f ssbgst.f ssbgv.f  ssbgvd.f ssbgvx.f
+  ssbtrd.f sspcon.f sspev.f  sspevd.f sspevx.f sspgst.f
+  sspgv.f  sspgvd.f sspgvx.f ssprfs.f sspsv.f  sspsvx.f ssptrd.f
+  ssptrf.f ssptri.f ssptrs.f sstegr.f sstein.f sstev.f  sstevd.f sstevr.f
+  sstevx.f
+  ssycon.f ssyev.f  ssyevd.f ssyevr.f ssyevx.f ssygs2.f
+  ssygst.f ssygv.f  ssygvd.f ssygvx.f ssyrfs.f ssysv.f  ssysvx.f
+  ssytd2.f ssytf2.f ssytrd.f ssytrf.f ssytri.f ssytri2.f ssytri2x.f
+  ssyswapr.f ssytrs.f ssytrs2.f ssyconv.f
+  ssytf2_rook.f ssytrf_rook.f ssytrs_rook.f
+  ssytri_rook.f ssycon_rook.f ssysv_rook.f
+  stbcon.f
+  stbrfs.f stbtrs.f stgevc.f stgex2.f stgexc.f stgsen.f
+  stgsja.f stgsna.f stgsy2.f stgsyl.f stpcon.f stprfs.f stptri.f
+  stptrs.f
+  strcon.f strevc.f strexc.f strrfs.f strsen.f strsna.f strsyl.f
+  strtrs.f stzrqf.f stzrzf.f sstemr.f
+  slansf.f spftrf.f spftri.f spftrs.f ssfrk.f stfsm.f stftri.f stfttp.f
+  stfttr.f stpttf.f stpttr.f strttf.f strttp.f
+  sgejsv.f  sgesvj.f  sgsvj0.f  sgsvj1.f
+  sgeequb.f ssyequb.f spoequb.f sgbequb.f
+  sbbcsd.f slapmr.f sorbdb.f sorbdb1.f sorbdb2.f sorbdb3.f sorbdb4.f
+  sorbdb5.f sorbdb6.f sorcsd.f sorcsd2by1.f
+  sgeqrt.f sgeqrt2.f sgeqrt3.f sgemqrt.f
+  stpqrt.f stpqrt2.f stpmqrt.f stprfb.f spotri.f
+)
+
+set(DSLASRC spotrs.f)
+
+set(CLASRC
+  cbdsqr.f cgbbrd.f cgbcon.f cgbequ.f cgbrfs.f cgbsv.f  cgbsvx.f
+  cgbtf2.f cgbtrf.f cgbtrs.f cgebak.f cgebal.f cgebd2.f cgebrd.f
+  cgecon.f cgeequ.f cgees.f  cgeesx.f cgeev.f  cgeevx.f
+  cgegs.f  cgegv.f  cgehd2.f cgehrd.f cgelq2.f cgelqf.f
+  cgels.f  cgelsd.f cgelss.f cgelsx.f cgelsy.f cgeql2.f cgeqlf.f cgeqp3.f
+  cgeqpf.f cgeqr2.f cgeqr2p.f cgeqrf.f cgeqrfp.f cgerfs.f
+  cgerq2.f cgerqf.f cgesc2.f cgesdd.f  cgesvd.f
+  cgesvx.f cgetc2.f cgetri.f
+  cggbak.f cggbal.f cgges.f  cggesx.f cggev.f  cggevx.f cggglm.f
+  cgghrd.f cgglse.f cggqrf.f cggrqf.f
+  cggsvd.f cggsvp.f
+  cgtcon.f cgtrfs.f cgtsv.f  cgtsvx.f cgttrf.f cgttrs.f cgtts2.f chbev.f
+  chbevd.f chbevx.f chbgst.f chbgv.f  chbgvd.f chbgvx.f chbtrd.f
+  checon.f cheev.f  cheevd.f cheevr.f cheevx.f chegs2.f chegst.f
+  chegv.f  chegvd.f chegvx.f cherfs.f chesv.f  chesvx.f chetd2.f
+  chetf2.f chetrd.f
+  chetrf.f chetri.f chetri2.f chetri2x.f cheswapr.f
+  chetrs.f chetrs2.f
+  chetf2_rook.f chetrf_rook.f chetri_rook.f chetrs_rook.f checon_rook.f chesv_rook.f
+  chgeqz.f chpcon.f chpev.f  chpevd.f
+  chpevx.f chpgst.f chpgv.f  chpgvd.f chpgvx.f chprfs.f chpsv.f
+  chpsvx.f
+  chptrd.f chptrf.f chptri.f chptrs.f chsein.f chseqr.f clabrd.f
+  clacgv.f clacon.f clacn2.f clacp2.f clacpy.f clacrm.f clacrt.f cladiv.f
+  claed0.f claed7.f claed8.f
+  claein.f claesy.f claev2.f clags2.f clagtm.f
+  clahef.f clahef_rook.f clahqr.f
+  clahrd.f clahr2.f claic1.f clals0.f clalsa.f clalsd.f clangb.f clange.f clangt.f
+  clanhb.f clanhe.f
+  clanhp.f clanhs.f clanht.f clansb.f clansp.f clansy.f clantb.f
+  clantp.f clantr.f clapll.f clapmt.f clarcm.f claqgb.f claqge.f
+  claqhb.f claqhe.f claqhp.f claqp2.f claqps.f claqsb.f
+  claqr0.f claqr1.f claqr2.f claqr3.f claqr4.f claqr5.f
+  claqsp.f claqsy.f clar1v.f clar2v.f ilaclr.f ilaclc.f
+  clarf.f  clarfb.f clarfg.f clarft.f clarfgp.f
+  clarfx.f clargv.f clarnv.f clarrv.f clartg.f clartv.f
+  clarz.f  clarzb.f clarzt.f clascl.f claset.f clasr.f  classq.f
+  clasyf.f clasyf_rook.f clatbs.f clatdf.f clatps.f clatrd.f clatrs.f clatrz.f
+  clatzm.f cpbcon.f cpbequ.f cpbrfs.f cpbstf.f cpbsv.f
+  cpbsvx.f cpbtf2.f cpbtrf.f cpbtrs.f cpocon.f cpoequ.f cporfs.f
+  cposv.f  cposvx.f cpstrf.f cpstf2.f
+  cppcon.f cppequ.f cpprfs.f cppsv.f  cppsvx.f cpptrf.f cpptri.f cpptrs.f
+  cptcon.f cpteqr.f cptrfs.f cptsv.f  cptsvx.f cpttrf.f cpttrs.f cptts2.f
+  crot.f   cspcon.f csprfs.f cspsv.f
+  cspsvx.f csptrf.f csptri.f csptrs.f csrscl.f cstedc.f
+  cstegr.f cstein.f csteqr.f
+  csycon.f
+  csyrfs.f csysv.f  csysvx.f csytf2.f csytrf.f csytri.f csytri2.f csytri2x.f
+  csyswapr.f csytrs.f csytrs2.f csyconv.f
+  csytf2_rook.f csytrf_rook.f csytrs_rook.f
+  csytri_rook.f csycon_rook.f csysv_rook.f
+  ctbcon.f ctbrfs.f ctbtrs.f ctgevc.f ctgex2.f
+  ctgexc.f ctgsen.f ctgsja.f ctgsna.f ctgsy2.f ctgsyl.f ctpcon.f
+  ctprfs.f ctptri.f
+  ctptrs.f ctrcon.f ctrevc.f ctrexc.f ctrrfs.f ctrsen.f ctrsna.f
+  ctrsyl.f ctrtrs.f ctzrqf.f ctzrzf.f cung2l.f cung2r.f
+  cungbr.f cunghr.f cungl2.f cunglq.f cungql.f cungqr.f cungr2.f
+  cungrq.f cungtr.f cunm2l.f cunm2r.f cunmbr.f cunmhr.f cunml2.f
+  cunmlq.f cunmql.f cunmqr.f cunmr2.f cunmr3.f cunmrq.f cunmrz.f
+  cunmtr.f cupgtr.f cupmtr.f icmax1.f scsum1.f cstemr.f
+  chfrk.f ctfttp.f clanhf.f cpftrf.f cpftri.f cpftrs.f ctfsm.f ctftri.f
+  ctfttr.f ctpttf.f ctpttr.f ctrttf.f ctrttp.f
+  cgeequb.f cgbequb.f csyequb.f cpoequb.f cheequb.f
+  cbbcsd.f clapmr.f cunbdb.f cunbdb1.f cunbdb2.f cunbdb3.f cunbdb4.f
+  cunbdb5.f cunbdb6.f cuncsd.f cuncsd2by1.f
+  cgeqrt.f cgeqrt2.f cgeqrt3.f cgemqrt.f
+  ctpqrt.f ctpqrt2.f ctpmqrt.f ctprfb.f cpotri.f
+)
+
+set(ZCLASRC cpotrs.f)
+
+set(DLASRC
+  dgbbrd.f dgbcon.f dgbequ.f dgbrfs.f dgbsv.f
+  dgbsvx.f dgbtf2.f dgbtrf.f dgbtrs.f dgebak.f dgebal.f dgebd2.f
+  dgebrd.f dgecon.f dgeequ.f dgees.f  dgeesx.f dgeev.f  dgeevx.f
+  dgegs.f  dgegv.f  dgehd2.f dgehrd.f dgelq2.f dgelqf.f
+  dgels.f  dgelsd.f dgelss.f dgelsx.f dgelsy.f dgeql2.f dgeqlf.f
+  dgeqp3.f dgeqpf.f dgeqr2.f dgeqr2p.f dgeqrf.f dgeqrfp.f dgerfs.f
+  dgerq2.f dgerqf.f dgesc2.f dgesdd.f  dgesvd.f dgesvx.f
+  dgetc2.f dgetri.f
+  dggbak.f dggbal.f dgges.f  dggesx.f dggev.f  dggevx.f
+  dggglm.f dgghrd.f dgglse.f dggqrf.f
+  dggrqf.f dggsvd.f dggsvp.f dgtcon.f dgtrfs.f dgtsv.f
+  dgtsvx.f dgttrf.f dgttrs.f dgtts2.f dhgeqz.f
+  dhsein.f dhseqr.f dlabrd.f dlacon.f dlacn2.f
+  dlaein.f dlaexc.f dlag2.f  dlags2.f dlagtm.f dlagv2.f dlahqr.f
+  dlahrd.f dlahr2.f dlaic1.f dlaln2.f dlals0.f dlalsa.f dlalsd.f
+  dlangb.f dlange.f dlangt.f dlanhs.f dlansb.f dlansp.f
+  dlansy.f dlantb.f dlantp.f dlantr.f dlanv2.f
+  dlapll.f dlapmt.f
+  dlaqgb.f dlaqge.f dlaqp2.f dlaqps.f dlaqsb.f dlaqsp.f dlaqsy.f
+  dlaqr0.f dlaqr1.f dlaqr2.f dlaqr3.f dlaqr4.f dlaqr5.f
+  dlaqtr.f dlar1v.f dlar2v.f iladlr.f iladlc.f
+  dlarf.f  dlarfb.f dlarfg.f dlarfgp.f dlarft.f dlarfx.f
+  dlargv.f dlarrv.f dlartv.f
+  dlarz.f  dlarzb.f dlarzt.f dlasy2.f dlasyf.f dlasyf_rook.f
+  dlatbs.f dlatdf.f dlatps.f dlatrd.f dlatrs.f dlatrz.f dlatzm.f
+  dopgtr.f dopmtr.f dorg2l.f dorg2r.f
+  dorgbr.f dorghr.f dorgl2.f dorglq.f dorgql.f dorgqr.f dorgr2.f
+  dorgrq.f dorgtr.f dorm2l.f dorm2r.f
+  dormbr.f dormhr.f dorml2.f dormlq.f dormql.f dormqr.f dormr2.f
+  dormr3.f dormrq.f dormrz.f dormtr.f dpbcon.f dpbequ.f dpbrfs.f
+  dpbstf.f dpbsv.f  dpbsvx.f
+  dpbtf2.f dpbtrf.f dpbtrs.f dpocon.f dpoequ.f dporfs.f dposv.f
+  dposvx.f dpotrs.f dpstrf.f dpstf2.f
+  dppcon.f dppequ.f
+  dpprfs.f dppsv.f  dppsvx.f dpptrf.f dpptri.f dpptrs.f dptcon.f
+  dpteqr.f dptrfs.f dptsv.f  dptsvx.f dpttrs.f dptts2.f drscl.f
+  dsbev.f  dsbevd.f dsbevx.f dsbgst.f dsbgv.f  dsbgvd.f dsbgvx.f
+  dsbtrd.f  dspcon.f dspev.f  dspevd.f dspevx.f dspgst.f
+  dspgv.f  dspgvd.f dspgvx.f dsprfs.f dspsv.f  dspsvx.f dsptrd.f
+  dsptrf.f dsptri.f dsptrs.f dstegr.f dstein.f dstev.f  dstevd.f dstevr.f
+  dstevx.f
+  dsycon.f dsyev.f  dsyevd.f dsyevr.f
+  dsyevx.f dsygs2.f dsygst.f dsygv.f  dsygvd.f dsygvx.f dsyrfs.f
+  dsysv.f  dsysvx.f
+  dsytd2.f dsytf2.f dsytrd.f dsytrf.f dsytri.f dsytri2.f dsytri2x.f
+  dsyswapr.f dsytrs.f dsytrs2.f dsyconv.f
+  dsytf2_rook.f dsytrf_rook.f dsytrs_rook.f
+  dsytri_rook.f dsycon_rook.f dsysv_rook.f
+  dtbcon.f dtbrfs.f dtbtrs.f dtgevc.f dtgex2.f dtgexc.f dtgsen.f
+  dtgsja.f dtgsna.f dtgsy2.f dtgsyl.f dtpcon.f dtprfs.f dtptri.f
+  dtptrs.f
+  dtrcon.f dtrevc.f dtrexc.f dtrrfs.f dtrsen.f dtrsna.f dtrsyl.f
+  dtrtrs.f dtzrqf.f dtzrzf.f dstemr.f
+  dsgesv.f dsposv.f dlag2s.f slag2d.f dlat2s.f
+  dlansf.f dpftrf.f dpftri.f dpftrs.f dsfrk.f dtfsm.f dtftri.f dtfttp.f
+  dtfttr.f dtpttf.f dtpttr.f dtrttf.f dtrttp.f
+  dgejsv.f  dgesvj.f  dgsvj0.f  dgsvj1.f
+  dgeequb.f dsyequb.f dpoequb.f dgbequb.f
+  dbbcsd.f dlapmr.f dorbdb.f dorbdb1.f dorbdb2.f dorbdb3.f dorbdb4.f
+  dorbdb5.f dorbdb6.f dorcsd.f dorcsd2by1.f
+  dgeqrt.f dgeqrt2.f dgeqrt3.f dgemqrt.f
+  dtpqrt.f dtpqrt2.f dtpmqrt.f dtprfb.f dpotri.f
+)
+
+set(ZLASRC
+  zbdsqr.f zgbbrd.f zgbcon.f zgbequ.f zgbrfs.f zgbsv.f  zgbsvx.f
+  zgbtf2.f zgbtrf.f zgbtrs.f zgebak.f zgebal.f zgebd2.f zgebrd.f
+  zgecon.f zgeequ.f zgees.f  zgeesx.f zgeev.f  zgeevx.f
+  zgegs.f  zgegv.f  zgehd2.f zgehrd.f zgelq2.f zgelqf.f
+  zgels.f  zgelsd.f zgelss.f zgelsx.f zgelsy.f zgeql2.f zgeqlf.f zgeqp3.f
+  zgeqpf.f zgeqr2.f zgeqr2p.f zgeqrf.f zgeqrfp.f zgerfs.f zgerq2.f zgerqf.f
+  zgesc2.f zgesdd.f zgesvd.f zgesvx.f zgetc2.f
+  zgetri.f
+  zggbak.f zggbal.f zgges.f  zggesx.f zggev.f  zggevx.f zggglm.f
+  zgghrd.f zgglse.f zggqrf.f zggrqf.f
+  zggsvd.f zggsvp.f
+  zgtcon.f zgtrfs.f zgtsv.f  zgtsvx.f zgttrf.f zgttrs.f zgtts2.f zhbev.f
+  zhbevd.f zhbevx.f zhbgst.f zhbgv.f  zhbgvd.f zhbgvx.f zhbtrd.f
+  zhecon.f zheev.f  zheevd.f zheevr.f zheevx.f zhegs2.f zhegst.f
+  zhegv.f  zhegvd.f zhegvx.f zherfs.f zhesv.f  zhesvx.f zhetd2.f
+  zhetf2.f zhetrd.f
+  zhetrf.f zhetri.f zhetri2.f zhetri2x.f zheswapr.f
+  zhetrs.f zhetrs2.f
+  zhetf2_rook.f zhetrf_rook.f zhetri_rook.f zhetrs_rook.f zhecon_rook.f zhesv_rook.f
+  zhgeqz.f zhpcon.f zhpev.f  zhpevd.f
+  zhpevx.f zhpgst.f zhpgv.f  zhpgvd.f zhpgvx.f zhprfs.f zhpsv.f
+  zhpsvx.f
+  zhptrd.f zhptrf.f zhptri.f zhptrs.f zhsein.f zhseqr.f zlabrd.f
+  zlacgv.f zlacon.f zlacn2.f zlacp2.f zlacpy.f zlacrm.f zlacrt.f zladiv.f
+  zlaed0.f zlaed7.f zlaed8.f
+  zlaein.f zlaesy.f zlaev2.f zlags2.f zlagtm.f
+  zlahef.f zlahef_rook.f zlahqr.f
+  zlahrd.f zlahr2.f zlaic1.f zlals0.f zlalsa.f zlalsd.f zlangb.f zlange.f
+  zlangt.f zlanhb.f
+  zlanhe.f
+  zlanhp.f zlanhs.f zlanht.f zlansb.f zlansp.f zlansy.f zlantb.f
+  zlantp.f zlantr.f zlapll.f zlapmt.f zlaqgb.f zlaqge.f
+  zlaqhb.f zlaqhe.f zlaqhp.f zlaqp2.f zlaqps.f zlaqsb.f
+  zlaqr0.f zlaqr1.f zlaqr2.f zlaqr3.f zlaqr4.f zlaqr5.f
+  zlaqsp.f zlaqsy.f zlar1v.f zlar2v.f ilazlr.f ilazlc.f
+  zlarcm.f zlarf.f  zlarfb.f
+  zlarfg.f zlarft.f zlarfgp.f
+  zlarfx.f zlargv.f zlarnv.f zlarrv.f zlartg.f zlartv.f
+  zlarz.f  zlarzb.f zlarzt.f zlascl.f zlaset.f zlasr.f
+  zlassq.f zlasyf.f zlasyf_rook.f
+  zlatbs.f zlatdf.f zlatps.f zlatrd.f zlatrs.f zlatrz.f zlatzm.f
+  zpbcon.f zpbequ.f zpbrfs.f zpbstf.f zpbsv.f
+  zpbsvx.f zpbtf2.f zpbtrf.f zpbtrs.f zpocon.f zpoequ.f zporfs.f
+  zposv.f  zposvx.f zpotrs.f zpstrf.f zpstf2.f
+  zppcon.f zppequ.f zpprfs.f zppsv.f  zppsvx.f zpptrf.f zpptri.f zpptrs.f
+  zptcon.f zpteqr.f zptrfs.f zptsv.f  zptsvx.f zpttrf.f zpttrs.f zptts2.f
+  zrot.f   zspcon.f zsprfs.f zspsv.f
+  zspsvx.f zsptrf.f zsptri.f zsptrs.f zdrscl.f zstedc.f
+  zstegr.f zstein.f zsteqr.f
+  zsycon.f
+  zsyrfs.f zsysv.f  zsysvx.f zsytf2.f zsytrf.f zsytri.f zsytri2.f zsytri2x.f
+  zsyswapr.f zsytrs.f zsytrs2.f zsyconv.f
+  zsytf2_rook.f zsytrf_rook.f zsytrs_rook.f
+  zsytri_rook.f zsycon_rook.f zsysv_rook.f
+  ztbcon.f ztbrfs.f ztbtrs.f ztgevc.f ztgex2.f
+  ztgexc.f ztgsen.f ztgsja.f ztgsna.f ztgsy2.f ztgsyl.f ztpcon.f
+  ztprfs.f ztptri.f
+  ztptrs.f ztrcon.f ztrevc.f ztrexc.f ztrrfs.f ztrsen.f ztrsna.f
+  ztrsyl.f ztrtrs.f ztzrqf.f ztzrzf.f zung2l.f
+  zung2r.f zungbr.f zunghr.f zungl2.f zunglq.f zungql.f zungqr.f zungr2.f
+  zungrq.f zungtr.f zunm2l.f zunm2r.f zunmbr.f zunmhr.f zunml2.f
+  zunmlq.f zunmql.f zunmqr.f zunmr2.f zunmr3.f zunmrq.f zunmrz.f
+  zunmtr.f zupgtr.f
+  zupmtr.f izmax1.f dzsum1.f zstemr.f
+  zcgesv.f zcposv.f zlag2c.f clag2z.f zlat2c.f
+  zhfrk.f ztfttp.f zlanhf.f zpftrf.f zpftri.f zpftrs.f ztfsm.f ztftri.f
+  ztfttr.f ztpttf.f ztpttr.f ztrttf.f ztrttp.f
+  zgeequb.f zgbequb.f zsyequb.f zpoequb.f zheequb.f
+  zbbcsd.f zlapmr.f zunbdb.f zunbdb1.f zunbdb2.f zunbdb3.f zunbdb4.f
+  zunbdb5.f zunbdb6.f zuncsd.f zuncsd2by1.f
+  zgeqrt.f zgeqrt2.f zgeqrt3.f zgemqrt.f
+  ztpqrt.f ztpqrt2.f ztpmqrt.f ztprfb.f zpotri.f
+)
+
+set(LA_REL_SRC ${ALLAUX})
+if (BUILD_SINGLE)
+  list(APPEND LA_REL_SRC ${SLASRC} ${DSLASRC} ${SCLAUX})
+endif ()
+
+if (BUILD_DOUBLE)
+  list(APPEND LA_REL_SRC ${DLASRC} ${DSLASRC} ${DZLAUX})
+endif ()
+
+if (BUILD_COMPLEX)
+  list(APPEND LA_REL_SRC ${CLASRC} ${ZCLASRC} ${SCLAUX})
+endif ()
+
+if (BUILD_COMPLEX16)
+  list(APPEND LA_REL_SRC ${ZLASRC} ${ZCLASRC} ${DZLAUX})
+endif ()
+
+# add lapack-netlib folder to the sources
+set(LA_SOURCES "")
+foreach (LA_FILE ${LA_REL_SRC})
+  list(APPEND LA_SOURCES "${NETLIB_LAPACK_DIR}/SRC/${LA_FILE}")
+endforeach ()
+set_source_files_properties(${LA_SOURCES} PROPERTIES COMPILE_FLAGS "${LAPACK_FFLAGS}")

cmake/os.cmake

@@ -0,0 +1,104 @@
+##
+## Author: Hank Anderson <hank@statease.com>
+## 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 "${EXTRALIB} -lm")
+  set(NO_EXPRECISION 1)
+endif ()
+
+if (${CMAKE_SYSTEM_NAME} STREQUAL "AIX")
+  set(EXTRALIB "${EXTRALIB} -lm")
+endif ()
+
+# TODO: this is probably meant for mingw, not other windows compilers
+if (${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
+
+  set(NEED_PIC 0)
+  set(NO_EXPRECISION 1)
+
+  set(EXTRALIB "${EXTRALIB} -defaultlib:advapi32")
+
+  # probably not going to use these
+  set(SUFFIX "obj")
+  set(PSUFFIX "pobj")
+  set(LIBSUFFIX "a")
+
+  if (${CMAKE_C_COMPILER_ID} STREQUAL "Clang")
+    set(CCOMMON_OPT	"${CCOMMON_OPT} -DMS_ABI")
+  endif ()
+
+  if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
+
+    # Test for supporting MS_ABI
+    # removed string parsing in favor of CMake's version comparison -hpa
+    execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
+    if (${GCC_VERSION} VERSION_GREATER 4.7 OR ${GCC_VERSION} VERSION_EQUAL 4.7)
+      # GCC Version >=4.7
+      # It is compatible with MSVC ABI.
+      set(CCOMMON_OPT "${CCOMMON_OPT} -DMS_ABI")
+    endif ()
+  endif ()
+
+  # Ensure the correct stack alignment on Win32
+  # http://permalink.gmane.org/gmane.comp.lib.openblas.general/97
+  if (${ARCH} STREQUAL "x86")
+    if (NOT MSVC AND NOT ${CMAKE_C_COMPILER_ID} STREQUAL "Clang")
+      set(CCOMMON_OPT "${CCOMMON_OPT} -mincoming-stack-boundary=2")
+    endif ()
+    set(FCOMMON_OPT "${FCOMMON_OPT} -mincoming-stack-boundary=2")
+  endif ()
+  
+endif ()
+
+if (${CMAKE_SYSTEM_NAME} STREQUAL "Interix")
+  set(NEED_PIC 0)
+  set(NO_EXPRECISION 1)
+  
+  set(INTERIX_TOOL_DIR STREQUAL "/opt/gcc.3.3/i586-pc-interix3/bin")
+endif ()
+
+if (CYGWIN)
+  set(NEED_PIC 0)
+  set(NO_EXPRECISION 1)
+endif ()
+
+if (NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Windows" AND NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Interix")
+  if (SMP)
+    set(EXTRALIB "${EXTRALIB} -lpthread")
+  endif ()
+endif ()
+
+if (QUAD_PRECISION)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DQUAD_PRECISION")
+  set(NO_EXPRECISION 1)
+endif ()
+
+if (${ARCH} STREQUAL "x86")
+  set(NO_EXPRECISION 1)
+endif ()
+
+if (UTEST_CHECK)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DUTEST_CHECK")
+  set(SANITY_CHECK 1)
+endif ()
+
+if (SANITY_CHECK)
+  # TODO: need some way to get $(*F) (target filename)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DSANITY_CHECK -DREFNAME=$(*F)f${BU}")
+endif ()
+

cmake/prebuild.cmake

@@ -0,0 +1,113 @@
+##
+## Author: Hank Anderson <hank@statease.com>
+## Description: Ported from OpenBLAS/Makefile.prebuild
+##              This is triggered by system.cmake and runs before any of the code is built.
+##              Creates config.h and Makefile.conf by first running the c_check perl script (which creates those files).
+##              Next it runs f_check and appends some fortran information to the files.
+##              Finally it runs getarch and getarch_2nd for even more environment information.
+
+# CMake vars set by this file:
+# CORE
+# LIBCORE
+# NUM_CORES
+# HAVE_MMX
+# HAVE_SSE
+# HAVE_SSE2
+# HAVE_SSE3
+# MAKE
+# SGEMM_UNROLL_M
+# SGEMM_UNROLL_N
+# DGEMM_UNROLL_M
+# DGEMM_UNROLL_M
+# QGEMM_UNROLL_N
+# QGEMM_UNROLL_N
+# CGEMM_UNROLL_M
+# CGEMM_UNROLL_M
+# ZGEMM_UNROLL_N
+# ZGEMM_UNROLL_N
+# XGEMM_UNROLL_M
+# XGEMM_UNROLL_N
+# CGEMM3M_UNROLL_M
+# CGEMM3M_UNROLL_N
+# ZGEMM3M_UNROLL_M
+# ZGEMM3M_UNROLL_M
+# XGEMM3M_UNROLL_N
+# XGEMM3M_UNROLL_N
+
+# CPUIDEMU = ../../cpuid/table.o
+
+if (DEFINED CPUIDEMU)
+  set(EXFLAGS "-DCPUIDEMU -DVENDOR=99")
+endif ()
+
+if (DEFINED TARGET_CORE)
+  # 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")
+else()
+  set(TARGET_MAKE "Makefile.conf")
+  set(TARGET_CONF "config.h")
+endif ()
+
+include("${CMAKE_SOURCE_DIR}/cmake/c_check.cmake")
+
+if (NOT NOFORTRAN)
+  include("${CMAKE_SOURCE_DIR}/cmake/f_check.cmake")
+endif ()
+
+# compile getarch
+set(GETARCH_SRC
+  ${CMAKE_SOURCE_DIR}/getarch.c
+  ${CPUIDEMO}
+)
+
+if (NOT MSVC)
+  list(APPEND GETARCH_SRC ${CMAKE_SOURCE_DIR}/cpuid.S)
+endif ()
+
+if (MSVC)
+#Use generic for MSVC now
+set(GETARCH_FLAGS ${GETARCH_FLAGS} -DFORCE_GENERIC)
+endif()
+
+set(GETARCH_DIR "${PROJECT_BINARY_DIR}/getarch_build")
+set(GETARCH_BIN "getarch${CMAKE_EXECUTABLE_SUFFIX}")
+file(MAKE_DIRECTORY ${GETARCH_DIR})
+try_compile(GETARCH_RESULT ${GETARCH_DIR}
+  SOURCES ${GETARCH_SRC}
+  COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} -I${CMAKE_SOURCE_DIR}
+  OUTPUT_VARIABLE GETARCH_LOG
+  COPY_FILE ${PROJECT_BINARY_DIR}/${GETARCH_BIN}
+)
+
+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} ${GETARCH_CONF_OUT})
+ParseGetArchVars(${GETARCH_MAKE_OUT})
+
+set(GETARCH2_DIR "${PROJECT_BINARY_DIR}/getarch2_build")
+set(GETARCH2_BIN "getarch_2nd${CMAKE_EXECUTABLE_SUFFIX}")
+file(MAKE_DIRECTORY ${GETARCH2_DIR})
+try_compile(GETARCH2_RESULT ${GETARCH2_DIR}
+  SOURCES ${CMAKE_SOURCE_DIR}/getarch_2nd.c
+  COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} ${GETARCH2_FLAGS} -I${CMAKE_SOURCE_DIR}
+  OUTPUT_VARIABLE GETARCH2_LOG
+  COPY_FILE ${PROJECT_BINARY_DIR}/${GETARCH2_BIN}
+)
+
+# 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)
+
+# append config data from getarch_2nd to the TARGET file and read in CMake vars
+file(APPEND ${TARGET_CONF} ${GETARCH2_CONF_OUT})
+ParseGetArchVars(${GETARCH2_MAKE_OUT})
+

cmake/system.cmake

@@ -0,0 +1,552 @@
+##
+## Author: Hank Anderson <hank@statease.com>
+## Description: Ported from OpenBLAS/Makefile.system
+##
+
+set(NETLIB_LAPACK_DIR "${CMAKE_SOURCE_DIR}/lapack-netlib")
+
+# TODO: Makefile.system detects Darwin (mac) and switches to clang here -hpa
+# http://stackoverflow.com/questions/714100/os-detecting-makefile
+
+# TODO: Makefile.system sets HOSTCC = $(CC) here if not already set -hpa
+
+# TARGET_CORE will override TARGET which is used in DYNAMIC_ARCH=1.
+if (DEFINED TARGET_CORE)
+  set(TARGET ${TARGET_CORE})
+endif ()
+
+# Force fallbacks for 32bit
+if (DEFINED BINARY AND DEFINED TARGET AND BINARY EQUAL 32)
+  message(STATUS "Compiling a ${BINARY}-bit binary.")
+  set(NO_AVX 1)
+  if (${TARGET} STREQUAL "HASWELL" OR ${TARGET} STREQUAL "SANDYBRIDGE")
+    set(TARGET "NEHALEM")
+  endif ()
+  if (${TARGET} STREQUAL "BULLDOZER" OR ${TARGET} STREQUAL "PILEDRIVER")
+    set(TARGET "BARCELONA")
+  endif ()
+endif ()
+
+if (DEFINED TARGET)
+  message(STATUS "Targetting the ${TARGET} architecture.")
+  set(GETARCH_FLAGS "-DFORCE_${TARGET}")
+endif ()
+
+if (INTERFACE64)
+  message(STATUS "Using 64-bit integers.")
+  set(GETARCH_FLAGS	"${GETARCH_FLAGS} -DUSE64BITINT")
+endif ()
+
+if (NOT DEFINED GEMM_MULTITHREAD_THRESHOLD)
+  set(GEMM_MULTITHREAD_THRESHOLD 4)
+endif ()
+message(STATUS "GEMM multithread threshold set to ${GEMM_MULTITHREAD_THRESHOLD}.")
+set(GETARCH_FLAGS	"${GETARCH_FLAGS} -DGEMM_MULTITHREAD_THRESHOLD=${GEMM_MULTITHREAD_THRESHOLD}")
+
+if (NO_AVX)
+  message(STATUS "Disabling Advanced Vector Extensions (AVX).")
+  set(GETARCH_FLAGS "${GETARCH_FLAGS} -DNO_AVX")
+endif ()
+
+if (NO_AVX2)
+  message(STATUS "Disabling Advanced Vector Extensions 2 (AVX2).")
+  set(GETARCH_FLAGS "${GETARCH_FLAGS} -DNO_AVX2")
+endif ()
+
+if (CMAKE_BUILD_TYPE STREQUAL Debug)
+  set(GETARCH_FLAGS "${GETARCH_FLAGS} -g")
+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)
+  set(GETARCH_FLAGS	"${GETARCH_FLAGS} -static")
+endif ()
+
+#if don't use Fortran, it will only compile CBLAS.
+if (ONLY_CBLAS)
+  set(NO_LAPACK 1)
+else ()
+  set(ONLY_CBLAS 0)
+endif ()
+
+include("${CMAKE_SOURCE_DIR}/cmake/prebuild.cmake")
+
+if (NOT DEFINED NUM_THREADS)
+  set(NUM_THREADS ${NUM_CORES})
+endif ()
+
+if (${NUM_THREADS} EQUAL 1)
+  set(USE_THREAD 0)
+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 ()
+endif ()
+
+if (${SMP})
+  message(STATUS "SMP enabled.")
+endif ()
+
+if (NOT DEFINED NEED_PIC)
+  set(NEED_PIC 1)
+endif ()
+
+# TODO: I think CMake should be handling all this stuff -hpa
+unset(ARFLAGS)
+set(CPP "${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("${CMAKE_SOURCE_DIR}/cmake/os.cmake")
+
+# Architecture dependent settings
+include("${CMAKE_SOURCE_DIR}/cmake/arch.cmake")
+
+# C Compiler dependent settings
+include("${CMAKE_SOURCE_DIR}/cmake/cc.cmake")
+
+if (NOT NOFORTRAN)
+  # Fortran Compiler dependent settings
+  include("${CMAKE_SOURCE_DIR}/cmake/fc.cmake")
+endif ()
+
+if (BINARY64)
+  if (INTERFACE64)
+    # CCOMMON_OPT += -DUSE64BITINT
+  endif ()
+endif ()
+
+if (NEED_PIC)
+  if (${CMAKE_C_COMPILER} STREQUAL "IBM")
+    set(CCOMMON_OPT "${CCOMMON_OPT} -qpic=large")
+  else ()
+    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 ()
+endif ()
+
+if (DYNAMIC_ARCH)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_ARCH")
+endif ()
+
+if (NO_LAPACK)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_LAPACK")
+  #Disable LAPACK C interface
+  set(NO_LAPACKE 1)
+endif ()
+
+if (NO_LAPACKE)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_LAPACKE")
+endif ()
+
+if (NO_AVX)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX")
+endif ()
+
+if (${ARCH} STREQUAL "x86")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX")
+endif ()
+
+if (NO_AVX2)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX2")
+endif ()
+
+if (SMP)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DSMP_SERVER")
+
+  if (${ARCH} STREQUAL "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 "${CCOMMON_OPT} -DBIGNUMA")
+  endif ()
+
+endif ()
+
+if (NO_WARMUP)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_WARMUP")
+endif ()
+
+if (CONSISTENT_FPCSR)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DCONSISTENT_FPCSR")
+endif ()
+
+# Only for development
+# set(CCOMMON_OPT "${CCOMMON_OPT} -DPARAMTEST")
+# set(CCOMMON_OPT "${CCOMMON_OPT} -DPREFETCHTEST")
+# set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_SWITCHING")
+# set(USE_PAPI 1)
+
+if (USE_PAPI)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_PAPI")
+  set(EXTRALIB "${EXTRALIB} -lpapi -lperfctr")
+endif ()
+
+if (DYNAMIC_THREADS)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_THREADS")
+endif ()
+
+set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_CPU_NUMBER=${NUM_THREADS}")
+
+if (USE_SIMPLE_THREADED_LEVEL3)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_SIMPLE_THREADED_LEVEL3")
+endif ()
+
+if (DEFINED LIBNAMESUFFIX)
+  set(LIBPREFIX "libopenblas_${LIBNAMESUFFIX}")
+else ()
+  set(LIBPREFIX "libopenblas")
+endif ()
+
+if (NOT DEFINED SYMBOLPREFIX)
+  set(SYMBOLPREFIX "")
+endif ()
+
+if (NOT DEFINED SYMBOLSUFFIX)
+  set(SYMBOLSUFFIX "")
+endif ()
+
+set(KERNELDIR	"${CMAKE_SOURCE_DIR}/kernel/${ARCH}")
+
+# TODO: nead to convert these Makefiles
+# include ${CMAKE_SOURCE_DIR}/cmake/${ARCH}.cmake
+
+if (${CORE} STREQUAL "PPC440")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_QALLOC")
+endif ()
+
+if (${CORE} STREQUAL "PPC440FP2")
+  set(STATIC_ALLOCATION 1)
+endif ()
+
+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 ${CORE} STREQUAL "LOONGSON3B")
+  set(NO_AFFINITY 1)
+endif ()
+
+if (NO_AFFINITY)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AFFINITY")
+endif ()
+
+if (FUNCTION_PROFILE)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DFUNCTION_PROFILE")
+endif ()
+
+if (HUGETLB_ALLOCATION)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_HUGETLB")
+endif ()
+
+if (DEFINED HUGETLBFILE_ALLOCATION)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_HUGETLBFILE -DHUGETLB_FILE_NAME=${HUGETLBFILE_ALLOCATION})")
+endif ()
+
+if (STATIC_ALLOCATION)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_STATIC")
+endif ()
+
+if (DEVICEDRIVER_ALLOCATION)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_DEVICEDRIVER -DDEVICEDRIVER_NAME=\"/dev/mapper\"")
+endif ()
+
+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(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} ${CCOMMON_OPT}")
+if(NOT MSVC)
+set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${COMMON_OPT} ${CCOMMON_OPT}")
+endif()
+# TODO: not sure what PFLAGS is -hpa
+set(PFLAGS "${PFLAGS} ${COMMON_OPT} ${CCOMMON_OPT} -I${TOPDIR} -DPROFILE ${COMMON_PROF}")
+
+set(CMAKE_Fortran_FLAGS "${CMAKE_Fortran_FLAGS} ${COMMON_OPT} ${FCOMMON_OPT}")
+# TODO: not sure what FPFLAGS is -hpa
+set(FPFLAGS "${FPFLAGS} ${COMMON_OPT} ${FCOMMON_OPT} ${COMMON_PROF}")
+
+#For LAPACK Fortran codes.
+set(LAPACK_FFLAGS "${LAPACK_FFLAGS} ${CMAKE_Fortran_FLAGS}")
+set(LAPACK_FPFLAGS "${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")
+  foreach (FILTER_FLAG ${FILTER_FLAGS})
+    string(REPLACE ${FILTER_FLAG} "" LAPACK_FFLAGS ${LAPACK_FFLAGS})
+    string(REPLACE ${FILTER_FLAG} "" LAPACK_FPFLAGS ${LAPACK_FPFLAGS})
+  endforeach ()
+endif ()
+
+if ("${F_COMPILER}" STREQUAL "GFORTRAN")
+  # lapack-netlib is rife with uninitialized warnings -hpa
+  set(LAPACK_FFLAGS "${LAPACK_FFLAGS} -Wno-maybe-uninitialized")
+endif ()
+
+set(LAPACK_CFLAGS "${CMAKE_C_CFLAGS} -DHAVE_LAPACK_CONFIG_H")
+if (INTERFACE64)
+  set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DLAPACK_ILP64")
+endif ()
+
+if (${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
+  set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DOPENBLAS_OS_WINDOWS")
+endif ()
+
+if (${CMAKE_C_COMPILER} STREQUAL "LSB" OR ${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
+  set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DLAPACK_COMPLEX_STRUCTURE")
+endif ()
+
+if (NOT DEFINED SUFFIX)
+  set(SUFFIX o)
+endif ()
+
+if (NOT DEFINED PSUFFIX)
+  set(PSUFFIX po)
+endif ()
+
+if (NOT DEFINED LIBSUFFIX)
+  set(LIBSUFFIX a)
+endif ()
+
+if (DYNAMIC_ARCH)
+  if (DEFINED SMP)
+    set(LIBNAME "${LIBPREFIX}p${REVISION}.${LIBSUFFIX}")
+    set(LIBNAME_P	"${LIBPREFIX}p${REVISION}_p.${LIBSUFFIX}")
+  else ()
+    set(LIBNAME "${LIBPREFIX}${REVISION}.${LIBSUFFIX}")
+    set(LIBNAME_P	"${LIBPREFIX}${REVISION}_p.${LIBSUFFIX}")
+  endif ()
+else ()
+  if (DEFINED SMP)
+    set(LIBNAME "${LIBPREFIX}_${LIBCORE}p${REVISION}.${LIBSUFFIX}")
+    set(LIBNAME_P	"${LIBPREFIX}_${LIBCORE}p${REVISION}_p.${LIBSUFFIX}")
+  else ()
+    set(LIBNAME	"${LIBPREFIX}_${LIBCORE}${REVISION}.${LIBSUFFIX}")
+    set(LIBNAME_P	"${LIBPREFIX}_${LIBCORE}${REVISION}_p.${LIBSUFFIX}")
+  endif ()
+endif ()
+
+
+set(LIBDLLNAME "${LIBPREFIX}.dll")
+set(LIBSONAME "${LIBNAME}.${LIBSUFFIX}.so")
+set(LIBDYNNAME "${LIBNAME}.${LIBSUFFIX}.dylib")
+set(LIBDEFNAME "${LIBNAME}.${LIBSUFFIX}.def")
+set(LIBEXPNAME "${LIBNAME}.${LIBSUFFIX}.exp")
+set(LIBZIPNAME "${LIBNAME}.${LIBSUFFIX}.zip")
+
+set(LIBS "${CMAKE_SOURCE_DIR}/${LIBNAME}")
+set(LIBS_P "${CMAKE_SOURCE_DIR}/${LIBNAME_P}")
+
+
+set(LIB_COMPONENTS BLAS)
+if (NOT NO_CBLAS)
+  set(LIB_COMPONENTS "${LIB_COMPONENTS} CBLAS")
+endif ()
+
+if (NOT NO_LAPACK)
+  set(LIB_COMPONENTS "${LIB_COMPONENTS} LAPACK")
+  if (NOT NO_LAPACKE)
+    set(LIB_COMPONENTS "${LIB_COMPONENTS} LAPACKE")
+  endif ()
+endif ()
+
+if (ONLY_CBLAS)
+  set(LIB_COMPONENTS CBLAS)
+endif ()
+
+
+# For GEMM3M
+set(USE_GEMM3M 0)
+
+if (DEFINED ARCH)
+  if (${ARCH} STREQUAL "x86" OR ${ARCH} STREQUAL "x86_64" OR ${ARCH} STREQUAL "ia64" OR ${ARCH} STREQUAL "MIPS")
+    set(USE_GEMM3M 1)
+  endif ()
+
+  if (${CORE} STREQUAL "generic")
+    set(USE_GEMM3M 0)
+  endif ()
+endif ()
+
+
+#export OSNAME
+#export ARCH
+#export CORE
+#export LIBCORE
+#export PGCPATH
+#export CONFIG
+#export CC
+#export FC
+#export BU
+#export FU
+#export NEED2UNDERSCORES
+#export USE_THREAD
+#export NUM_THREADS
+#export NUM_CORES
+#export SMP
+#export MAKEFILE_RULE
+#export NEED_PIC
+#export BINARY
+#export BINARY32
+#export BINARY64
+#export F_COMPILER
+#export C_COMPILER
+#export USE_OPENMP
+#export CROSS
+#export CROSS_SUFFIX
+#export NOFORTRAN
+#export NO_FBLAS
+#export EXTRALIB
+#export CEXTRALIB
+#export FEXTRALIB
+#export HAVE_SSE
+#export HAVE_SSE2
+#export HAVE_SSE3
+#export HAVE_SSSE3
+#export HAVE_SSE4_1
+#export HAVE_SSE4_2
+#export HAVE_SSE4A
+#export HAVE_SSE5
+#export HAVE_AVX
+#export HAVE_VFP
+#export HAVE_VFPV3
+#export HAVE_VFPV4
+#export HAVE_NEON
+#export KERNELDIR
+#export FUNCTION_PROFILE
+#export TARGET_CORE
+#
+#export SGEMM_UNROLL_M
+#export SGEMM_UNROLL_N
+#export DGEMM_UNROLL_M
+#export DGEMM_UNROLL_N
+#export QGEMM_UNROLL_M
+#export QGEMM_UNROLL_N
+#export CGEMM_UNROLL_M
+#export CGEMM_UNROLL_N
+#export ZGEMM_UNROLL_M
+#export ZGEMM_UNROLL_N
+#export XGEMM_UNROLL_M
+#export XGEMM_UNROLL_N
+#export CGEMM3M_UNROLL_M
+#export CGEMM3M_UNROLL_N
+#export ZGEMM3M_UNROLL_M
+#export ZGEMM3M_UNROLL_N
+#export XGEMM3M_UNROLL_M
+#export XGEMM3M_UNROLL_N
+
+
+#if (USE_CUDA)
+#  export CUDADIR
+#  export CUCC
+#  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/utils.cmake

@@ -0,0 +1,346 @@
+# Functions to help with the OpenBLAS build
+
+# Reads string from getarch into CMake vars. Format of getarch vars is VARNAME=VALUE
+function(ParseGetArchVars GETARCH_IN)
+  string(REGEX MATCHALL "[0-9_a-zA-Z]+=[0-9_a-zA-Z]+" GETARCH_RESULT_LIST "${GETARCH_IN}")
+  foreach (GETARCH_LINE ${GETARCH_RESULT_LIST})
+    # split the line into var and value, then assign the value to a CMake var
+    string(REGEX MATCHALL "[0-9_a-zA-Z]+" SPLIT_VAR "${GETARCH_LINE}")
+    list(GET SPLIT_VAR 0 VAR_NAME)
+    list(GET SPLIT_VAR 1 VAR_VALUE)
+    set(${VAR_NAME} ${VAR_VALUE} PARENT_SCOPE)
+  endforeach ()
+endfunction ()
+
+# Reads a Makefile into CMake vars.
+macro(ParseMakefileVars MAKEFILE_IN)
+  message(STATUS "Reading vars from ${MAKEFILE_IN}...")
+  file(STRINGS ${MAKEFILE_IN} makefile_contents)
+  foreach (makefile_line ${makefile_contents})
+    string(REGEX MATCH "([0-9_a-zA-Z]+)[ \t]*=[ \t]*(.+)$" line_match "${makefile_line}")
+    if (NOT "${line_match}" STREQUAL "")
+      set(var_name ${CMAKE_MATCH_1})
+      set(var_value ${CMAKE_MATCH_2})
+      # check for Makefile variables in the string, e.g. $(TSUFFIX)
+      string(REGEX MATCHALL "\\$\\(([0-9_a-zA-Z]+)\\)" make_var_matches ${var_value})
+      foreach (make_var ${make_var_matches})
+        # strip out Makefile $() markup
+        string(REGEX REPLACE "\\$\\(([0-9_a-zA-Z]+)\\)" "\\1" make_var ${make_var})
+        # now replace the instance of the Makefile variable with the value of the CMake variable (note the double quote)
+        string(REPLACE "$(${make_var})" "${${make_var}}" var_value ${var_value})
+      endforeach ()
+      set(${var_name} ${var_value})
+    else ()
+      string(REGEX MATCH "include \\$\\(KERNELDIR\\)/(.+)$" line_match "${makefile_line}")
+      if (NOT "${line_match}" STREQUAL "")
+        ParseMakefileVars(${KERNELDIR}/${CMAKE_MATCH_1})
+      endif ()
+    endif ()
+  endforeach ()
+endmacro ()
+
+# Returns all combinations of the input list, as a list with colon-separated combinations
+# E.g. input of A B C returns A B C A:B A:C B:C
+# N.B. The input is meant to be a list, and to past a list to a function in CMake you must quote it (e.g. AllCombinations("${LIST_VAR}")).
+# #param absent_codes codes to use when an element is absent from a combination. For example, if you have TRANS;UNIT;UPPER you may want the code to be NNL when nothing is present.
+# @returns LIST_OUT a list of combinations
+#          CODES_OUT a list of codes corresponding to each combination, with N meaning the item is not present, and the first letter of the list item meaning it is presen
+function(AllCombinations list_in absent_codes_in)
+  list(LENGTH list_in list_count)
+  set(num_combos 1)
+  # subtract 1 since we will iterate from 0 to num_combos
+  math(EXPR num_combos "(${num_combos} << ${list_count}) - 1")
+  set(LIST_OUT "")
+  set(CODES_OUT "")
+  foreach (c RANGE 0 ${num_combos})
+
+    set(current_combo "")
+    set(current_code "")
+
+    # this is a little ridiculous just to iterate through a list w/ indices
+    math(EXPR last_list_index "${list_count} - 1")
+    foreach (list_index RANGE 0 ${last_list_index})
+      math(EXPR bit "1 << ${list_index}")
+      math(EXPR combo_has_bit "${c} & ${bit}")
+      list(GET list_in ${list_index} list_elem)
+      if (combo_has_bit)
+        if (current_combo)
+          set(current_combo "${current_combo}:${list_elem}")
+        else ()
+          set(current_combo ${list_elem})
+        endif ()
+        string(SUBSTRING ${list_elem} 0 1 code_char)
+      else ()
+        list(GET absent_codes_in ${list_index} code_char)
+      endif ()
+      set(current_code "${current_code}${code_char}")
+    endforeach ()
+
+    if (current_combo STREQUAL "")
+      list(APPEND LIST_OUT " ") # Empty set is a valid combination, but CMake isn't appending the empty string for some reason, use a space
+    else ()
+      list(APPEND LIST_OUT ${current_combo})
+    endif ()
+    list(APPEND CODES_OUT ${current_code})
+
+  endforeach ()
+
+  set(LIST_OUT ${LIST_OUT} PARENT_SCOPE)
+  set(CODES_OUT ${CODES_OUT} PARENT_SCOPE)
+endfunction ()
+
+# generates object files for each of the sources, using the BLAS naming scheme to pass the funciton name as a preprocessor definition
+# @param sources_in the source files to build from
+# @param defines_in (optional) preprocessor definitions that will be applied to all objects
+# @param name_in (optional) if this is set this name will be used instead of the filename. Use a * to indicate where the float character should go, if no star the character will be prepended.
+#                           e.g. with DOUBLE set, "i*max" will generate the name "idmax", and "max" will be "dmax"
+# @param replace_last_with replaces the last character in the filename with this string (e.g. symm_k should be symm_TU)
+# @param append_with appends the filename with this string (e.g. trmm_R should be trmm_RTUU or some other combination of characters)
+# @param no_float_type turns off the float type define for this build (e.g. SINGLE/DOUBLE/etc)
+# @param complex_filename_scheme some routines have separate source files for complex and non-complex float types.
+#                               0 - compiles for all types
+#                               1 - compiles the sources for non-complex types only (SINGLE/DOUBLE)
+#                               2 - compiles for complex types only (COMPLEX/DOUBLE COMPLEX)
+#                               3 - compiles for all types, but changes source names for complex by prepending z (e.g. axpy.c becomes zaxpy.c)
+#                               4 - compiles for complex types only, but changes source names for complex by prepending z (e.g. hemv.c becomes zhemv.c)
+#                               STRING - compiles only the given type (e.g. DOUBLE)
+function(GenerateNamedObjects sources_in)
+
+  if (DEFINED ARGV1)
+    set(defines_in ${ARGV1})
+  endif ()
+
+  if (DEFINED ARGV2 AND NOT "${ARGV2}" STREQUAL "")
+    set(name_in ${ARGV2})
+    # strip off extension for kernel files that pass in the object name.
+    get_filename_component(name_in ${name_in} NAME_WE)
+  endif ()
+
+  if (DEFINED ARGV3)
+    set(use_cblas ${ARGV3})
+  else ()
+    set(use_cblas false)
+  endif ()
+
+  if (DEFINED ARGV4)
+    set(replace_last_with ${ARGV4})
+  endif ()
+
+  if (DEFINED ARGV5)
+    set(append_with ${ARGV5})
+  endif ()
+
+  if (DEFINED ARGV6)
+    set(no_float_type ${ARGV6})
+  else ()
+    set(no_float_type false)
+  endif ()
+
+  if (no_float_type)
+    set(float_list "DUMMY") # still need to loop once
+  else ()
+    set(float_list "${FLOAT_TYPES}")
+  endif ()
+
+  set(real_only false)
+  set(complex_only false)
+  set(mangle_complex_sources false)
+  if (DEFINED ARGV7 AND NOT "${ARGV7}" STREQUAL "")
+    if (${ARGV7} EQUAL 1)
+      set(real_only true)
+    elseif (${ARGV7} EQUAL 2)
+      set(complex_only true)
+    elseif (${ARGV7} EQUAL 3)
+      set(mangle_complex_sources true)
+    elseif (${ARGV7} EQUAL 4)
+      set(mangle_complex_sources true)
+      set(complex_only true)
+    elseif (NOT ${ARGV7} EQUAL 0)
+      set(float_list ${ARGV7})
+    endif ()
+  endif ()
+
+  if (complex_only)
+    list(REMOVE_ITEM float_list "SINGLE")
+    list(REMOVE_ITEM float_list "DOUBLE")
+  elseif (real_only)
+    list(REMOVE_ITEM float_list "COMPLEX")
+    list(REMOVE_ITEM float_list "ZCOMPLEX")
+  endif ()
+
+  set(float_char "")
+  set(OBJ_LIST_OUT "")
+  foreach (float_type ${float_list})
+    foreach (source_file ${sources_in})
+
+      if (NOT no_float_type)
+        string(SUBSTRING ${float_type} 0 1 float_char)
+        string(TOLOWER ${float_char} float_char)
+      endif ()
+
+      if (NOT name_in)
+        get_filename_component(source_name ${source_file} NAME_WE)
+        set(obj_name "${float_char}${source_name}")
+      else ()
+        # replace * with float_char
+        if (${name_in} MATCHES "\\*")
+          string(REPLACE "*" ${float_char} obj_name ${name_in})
+        else ()
+          set(obj_name "${float_char}${name_in}")
+        endif ()
+      endif ()
+
+      if (replace_last_with)
+        string(REGEX REPLACE ".$" ${replace_last_with} obj_name ${obj_name})
+      else ()
+        set(obj_name "${obj_name}${append_with}")
+      endif ()
+
+      # now add the object and set the defines
+      set(obj_defines ${defines_in})
+
+      if (use_cblas)
+        set(obj_name "cblas_${obj_name}")
+        list(APPEND obj_defines "CBLAS")
+      endif ()
+
+      list(APPEND obj_defines "ASMNAME=${FU}${obj_name};ASMFNAME=${FU}${obj_name}${BU};NAME=${obj_name}${BU};CNAME=${obj_name};CHAR_NAME=\"${obj_name}${BU}\";CHAR_CNAME=\"${obj_name}\"")
+      if (${float_type} STREQUAL "DOUBLE" OR ${float_type} STREQUAL "ZCOMPLEX")
+        list(APPEND obj_defines "DOUBLE")
+      endif ()
+      if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
+        list(APPEND obj_defines "COMPLEX")
+        if (mangle_complex_sources)
+          # add a z to the filename
+          get_filename_component(source_name ${source_file} NAME)
+          get_filename_component(source_dir ${source_file} DIRECTORY)
+          string(REPLACE ${source_name} "z${source_name}" source_file ${source_file})
+        endif ()
+      endif ()
+
+      if (VERBOSE_GEN)
+        message(STATUS "${obj_name}:${source_file}")
+        message(STATUS "${obj_defines}")
+      endif ()
+
+      # create a copy of the source to avoid duplicate obj filename problem with ar.exe
+      get_filename_component(source_extension ${source_file} EXT)
+      set(new_source_file "${CMAKE_CURRENT_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/${obj_name}${source_extension}")
+      if (IS_ABSOLUTE ${source_file})
+        set(old_source_file ${source_file})
+      else ()
+        set(old_source_file "${CMAKE_CURRENT_LIST_DIR}/${source_file}")
+      endif ()
+
+      string(REPLACE ";" "\n#define " define_source "${obj_defines}")
+      string(REPLACE "=" " " define_source "${define_source}")
+      file(WRITE ${new_source_file} "#define ${define_source}\n#include \"${old_source_file}\"")
+      list(APPEND SRC_LIST_OUT ${new_source_file})
+
+    endforeach ()
+  endforeach ()
+
+  list(APPEND OPENBLAS_SRC ${SRC_LIST_OUT})
+  set(OPENBLAS_SRC ${OPENBLAS_SRC} PARENT_SCOPE)
+endfunction ()
+
+# generates object files for each of the sources for each of the combinations of the preprocessor definitions passed in
+# @param sources_in the source files to build from
+# @param defines_in the preprocessor definitions that will be combined to create the object files
+# @param all_defines_in (optional) preprocessor definitions that will be applied to all objects
+# @param replace_scheme If 1, replace the "k" in the filename with the define combo letters. E.g. symm_k.c with TRANS and UNIT defined will be symm_TU.
+#                  If 0, it will simply append the code, e.g. symm_L.c with TRANS and UNIT will be symm_LTU.
+#                  If 2, it will append the code with an underscore, e.g. symm.c with TRANS and UNIT will be symm_TU.
+#                  If 3, it will insert the code *around* the last character with an underscore, e.g. symm_L.c with TRANS and UNIT will be symm_TLU (required by BLAS level2 objects).
+#                  If 4, it will insert the code before the last underscore. E.g. trtri_U_parallel with TRANS will be trtri_UT_parallel
+# @param alternate_name replaces the source name as the object name (define codes are still appended)
+# @param no_float_type turns off the float type define for this build (e.g. SINGLE/DOUBLE/etc)
+# @param complex_filename_scheme see GenerateNamedObjects
+function(GenerateCombinationObjects sources_in defines_in absent_codes_in all_defines_in replace_scheme)
+
+  set(alternate_name_in "")
+  if (DEFINED ARGV5)
+    set(alternate_name_in ${ARGV5})
+  endif ()
+
+  set(no_float_type false)
+  if (DEFINED ARGV6)
+    set(no_float_type ${ARGV6})
+  endif ()
+
+  set(complex_filename_scheme "")
+  if (DEFINED ARGV7)
+    set(complex_filename_scheme ${ARGV7})
+  endif ()
+
+  AllCombinations("${defines_in}" "${absent_codes_in}")
+  set(define_combos ${LIST_OUT})
+  set(define_codes ${CODES_OUT})
+
+  list(LENGTH define_combos num_combos)
+  math(EXPR num_combos "${num_combos} - 1")
+
+  foreach (c RANGE 0 ${num_combos})
+
+    list(GET define_combos ${c} define_combo)
+    list(GET define_codes ${c} define_code)
+
+    foreach (source_file ${sources_in})
+
+      set(alternate_name ${alternate_name_in})
+
+      # replace colon separated list with semicolons, this turns it into a CMake list that we can use foreach with
+      string(REPLACE ":" ";" define_combo ${define_combo})
+
+      # now add the object and set the defines
+      set(cur_defines ${define_combo})
+      if ("${cur_defines}" STREQUAL " ")
+        set(cur_defines ${all_defines_in})
+      else ()
+        list(APPEND cur_defines ${all_defines_in})
+      endif ()
+
+      set(replace_code "")
+      set(append_code "")
+      if (replace_scheme EQUAL 1)
+        set(replace_code ${define_code})
+      else ()
+        if (replace_scheme EQUAL 2)
+          set(append_code "_${define_code}")
+        elseif (replace_scheme EQUAL 3)
+          if ("${alternate_name}" STREQUAL "")
+            string(REGEX MATCH "[a-zA-Z]\\." last_letter ${source_file})
+          else ()
+            string(REGEX MATCH "[a-zA-Z]$" last_letter ${alternate_name})
+          endif ()
+          # first extract the last letter
+          string(SUBSTRING ${last_letter} 0 1 last_letter) # remove period from match
+          # break the code up into the first letter and the remaining (should only be 2 anyway)
+          string(SUBSTRING ${define_code} 0 1 define_code_first)
+          string(SUBSTRING ${define_code} 1 -1 define_code_second)
+          set(replace_code "${define_code_first}${last_letter}${define_code_second}")
+        elseif (replace_scheme EQUAL 4)
+          # insert code before the last underscore and pass that in as the alternate_name
+          if ("${alternate_name}" STREQUAL "")
+            get_filename_component(alternate_name ${source_file} NAME_WE)
+          endif ()
+          set(extra_underscore "")
+          # check if filename has two underscores, insert another if not (e.g. getrs_parallel needs to become getrs_U_parallel not getrsU_parallel)
+          string(REGEX MATCH "_[a-zA-Z]+_" underscores ${alternate_name})
+          string(LENGTH "${underscores}" underscores)
+          if (underscores EQUAL 0)
+            set(extra_underscore "_")
+          endif ()
+          string(REGEX REPLACE "(.+)(_[^_]+)$" "\\1${extra_underscore}${define_code}\\2" alternate_name ${alternate_name})
+        else()
+          set(append_code ${define_code}) # replace_scheme should be 0
+        endif ()
+      endif ()
+
+      GenerateNamedObjects("${source_file}" "${cur_defines}" "${alternate_name}" false "${replace_code}" "${append_code}" "${no_float_type}" "${complex_filename_scheme}")
+    endforeach ()
+  endforeach ()
+
+  set(OPENBLAS_SRC ${OPENBLAS_SRC} PARENT_SCOPE)
+endfunction ()
+

common.h

@@ -82,7 +82,11 @@ extern "C" {
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+
+#if !defined(_MSC_VER)
 #include <unistd.h>
+#endif
+#include <time.h>
 
 #ifdef OS_LINUX
 #include <malloc.h>
@@ -95,6 +99,10 @@ extern "C" {
 
 #ifdef OS_ANDROID
 #define NO_SYSV_IPC
+//Android NDK only supports complex.h since Android 5.0
+#if __ANDROID_API__ < 21
+#define FORCE_OPENBLAS_COMPLEX_STRUCT
+#endif
 #endif
 
 #ifdef OS_WINDOWS
@@ -114,6 +122,7 @@ extern "C" {
 #include <sys/shm.h>
 #endif
 #include <sys/time.h>
+#include <time.h>
 #include <unistd.h>
 #include <math.h>
 #ifdef SMP
@@ -293,13 +302,6 @@ typedef int blasint;
 #define COMPSIZE  2
 #endif
 
-#if defined(C_PGI) || defined(C_SUN)
-#define CREAL(X)	(*((FLOAT *)&X + 0))
-#define CIMAG(X)	(*((FLOAT *)&X + 1))
-#else
-#define CREAL	__real__
-#define CIMAG	__imag__
-#endif
 
 #define Address_H(x) (((x)+(1<<15))>>16)
 #define Address_L(x) ((x)-((Address_H(x))<<16))
@@ -313,8 +315,12 @@ typedef int blasint;
 #endif
 
 #if defined(OS_WINDOWS)
+#if defined(_MSC_VER) && !defined(__clang__)
+#define YIELDING    YieldProcessor()
+#else
 #define YIELDING	SwitchToThread()
 #endif
+#endif
 
 #if defined(ARMV7) || defined(ARMV6) || defined(ARMV8) || defined(ARMV5)
 #define YIELDING        asm volatile ("nop;nop;nop;nop;nop;nop;nop;nop; \n");
@@ -410,7 +416,51 @@ typedef char env_var_t[MAX_PATH];
 typedef char* env_var_t;
 #define readenv(p, n) ((p)=getenv(n))
 #endif
+
+#if !defined(RPCC_DEFINED) && !defined(OS_WINDOWS)
+#ifdef _POSIX_MONOTONIC_CLOCK
+#if defined(__GLIBC_PREREQ) // cut the if condition if two lines, otherwise will fail at __GLIBC_PREREQ(2, 17)
+#if __GLIBC_PREREQ(2, 17) // don't require -lrt
+#define USE_MONOTONIC
 #endif
+#elif defined(OS_ANDROID)
+#define USE_MONOTONIC
+#endif
+#endif
+/* use similar scale as x86 rdtsc for timeouts to work correctly */
+static inline unsigned long long rpcc(void){
+#ifdef USE_MONOTONIC
+  struct timespec ts;
+  clock_gettime(CLOCK_MONOTONIC, &ts);
+  return (unsigned long long)ts.tv_sec * 1000000000ull + ts.tv_nsec;
+#else
+  struct timeval tv;
+  gettimeofday(&tv,NULL);
+  return (unsigned long long)tv.tv_sec * 1000000000ull + tv.tv_usec * 1000;
+#endif
+}
+#define RPCC_DEFINED
+#define RPCC64BIT
+#endif // !RPCC_DEFINED
+
+#if !defined(BLAS_LOCK_DEFINED) && defined(__GNUC__)
+static void __inline blas_lock(volatile BLASULONG *address){
+
+  do {
+    while (*address) {YIELDING;};
+
+  } while (!__sync_bool_compare_and_swap(address, 0, 1));
+}
+#define BLAS_LOCK_DEFINED
+#endif
+
+#ifndef RPCC_DEFINED
+#error "rpcc() implementation is missing for your platform"
+#endif
+#ifndef BLAS_LOCK_DEFINED
+#error "blas_lock() implementation is missing for your platform"
+#endif
+#endif // !ASSEMBLER
 
 #ifdef OS_LINUX
 #include "common_linux.h"
@@ -456,18 +506,52 @@ typedef char* env_var_t;
 /* C99 supports complex floating numbers natively, which GCC also offers as an
    extension since version 3.0.  If neither are available, use a compatible
    structure as fallback (see Clause 6.2.5.13 of the C99 standard). */
-#if (defined(__STDC_IEC_559_COMPLEX__) || __STDC_VERSION__ >= 199901L || \
-     (__GNUC__ >= 3 && !defined(__cplusplus)))
+#if ((defined(__STDC_IEC_559_COMPLEX__) || __STDC_VERSION__ >= 199901L || \
+      (__GNUC__ >= 3 && !defined(__cplusplus))) && !(defined(FORCE_OPENBLAS_COMPLEX_STRUCT)))
   #define OPENBLAS_COMPLEX_C99
+  #ifndef __cplusplus
+    #include <complex.h>
+  #endif
   typedef float _Complex openblas_complex_float;
   typedef double _Complex openblas_complex_double;
   typedef xdouble _Complex openblas_complex_xdouble;
+  #define openblas_make_complex_float(real, imag)    ((real) + ((imag) * _Complex_I))
+  #define openblas_make_complex_double(real, imag)   ((real) + ((imag) * _Complex_I))
+  #define openblas_make_complex_xdouble(real, imag)  ((real) + ((imag) * _Complex_I))
 #else
   #define OPENBLAS_COMPLEX_STRUCT
   typedef struct { float real, imag; } openblas_complex_float;
   typedef struct { double real, imag; } openblas_complex_double;
   typedef struct { xdouble real, imag; } openblas_complex_xdouble;
+  #define openblas_make_complex_float(real, imag)    {(real), (imag)}
+  #define openblas_make_complex_double(real, imag)   {(real), (imag)}
+  #define openblas_make_complex_xdouble(real, imag)  {(real), (imag)}
 #endif
+
+#ifdef XDOUBLE
+#define OPENBLAS_COMPLEX_FLOAT openblas_complex_xdouble
+#define OPENBLAS_MAKE_COMPLEX_FLOAT(r,i) openblas_make_complex_xdouble(r,i)
+#elif defined(DOUBLE)
+#define OPENBLAS_COMPLEX_FLOAT openblas_complex_double
+#define OPENBLAS_MAKE_COMPLEX_FLOAT(r,i) openblas_make_complex_double(r,i)
+#else
+#define OPENBLAS_COMPLEX_FLOAT openblas_complex_float
+#define OPENBLAS_MAKE_COMPLEX_FLOAT(r,i) openblas_make_complex_float(r,i)
+#endif
+
+#if defined(C_PGI) || defined(C_SUN)
+#define CREAL(X)	(*((FLOAT *)&X + 0))
+#define CIMAG(X)	(*((FLOAT *)&X + 1))
+#else
+#ifdef OPENBLAS_COMPLEX_STRUCT
+#define CREAL(Z)	((Z).real)
+#define CIMAG(Z)	((Z).imag)
+#else
+#define CREAL	__real__
+#define CIMAG	__imag__
+#endif
+#endif
+
 #endif  // ASSEMBLER
 
 #ifndef IFLUSH
@@ -484,6 +568,10 @@ typedef char* env_var_t;
 #endif
 #endif
 
+#if defined(C_MSVC)
+#define inline __inline
+#endif
+
 #ifndef ASSEMBLER
 
 #ifndef MIN

common_alpha.h

@@ -76,6 +76,7 @@ static void __inline blas_lock(unsigned long *address){
     "30:", address);
 #endif
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline unsigned int rpcc(void){
 
@@ -89,6 +90,7 @@ static __inline unsigned int rpcc(void){
 
   return r0;
 }
+#define RPCC_DEFINED
 
 
 #define HALT 	ldq	$0, 0($0)

common_arm.h

@@ -51,6 +51,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifndef ASSEMBLER
 
+#if defined(ARMV6) || defined(ARMV7) || defined(ARMV8)
+
 static void __inline blas_lock(volatile BLASULONG *address){
 
   int register ret;
@@ -59,40 +61,29 @@ static void __inline blas_lock(volatile BLASULONG *address){
     while (*address) {YIELDING;};
 
     __asm__ __volatile__(
-			 "1:								\n\t"
-                         "ldrex r2, [%1]                                                \n\t"
-                         "mov   r2, #0                                                  \n\t"
-                         "strex r3, r2, [%1]                                            \n\t"
-			 "cmp	r3, #0							\n\t"
-			 "bne	1b							\n\t"
-			 "mov	%0 , r3							\n\t"
-                         : "=r"(ret), "=r"(address)
-                         : "1"(address)
-                         : "memory", "r2" , "r3"
-
-
+                         "ldrex r2, [%1]      \n\t"
+                         "strex %0, %2, [%1]  \n\t"
+                         "orr   %0, r2        \n\t"
+                         : "=&r"(ret)
+                         : "r"(address), "r"(1)
+                         : "memory", "r2"
     );
 
   } while (ret);
-
+  MB;
 }
 
-
-static inline unsigned long long rpcc(void){
-  unsigned long long ret=0;
-  double v;
-  struct timeval tv;
-  gettimeofday(&tv,NULL);
-  v=(double) tv.tv_sec + (double) tv.tv_usec * 1e-6;
-  ret = (unsigned long long) ( v * 1000.0d );
-  return ret;
-}
+#define BLAS_LOCK_DEFINED
+#endif
 
 static inline int blas_quickdivide(blasint x, blasint y){
   return x / y;
 }
 
-#if defined(DOUBLE)
+#if !defined(HAVE_VFP)
+/* no FPU, soft float */
+#define GET_IMAGE(res)
+#elif defined(DOUBLE)
 #define GET_IMAGE(res)  __asm__ __volatile__("vstr.f64 d1, %0" : "=m"(res) : : "memory")
 #else
 #define GET_IMAGE(res)  __asm__ __volatile__("vstr.f32 s1, %0" : "=m"(res) : : "memory")
@@ -140,4 +131,8 @@ REALNAME:
 #define MAP_ANONYMOUS MAP_ANON
 #endif
 
+#if !defined(ARMV5) && !defined(ARMV6) && !defined(ARMV7) && !defined(ARMV8)
+#error "you must define ARMV5, ARMV6, ARMV7 or ARMV8"
+#endif
+
 #endif

common_arm64.h

@@ -43,43 +43,38 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifndef ASSEMBLER
 
+
 static void __inline blas_lock(volatile BLASULONG *address){
 
-  int register ret;
-  int register tmp;
+  BLASULONG ret;
 
   do {
     while (*address) {YIELDING;};
 
     __asm__ __volatile__(
+			 "mov	x4, #1							\n\t"
                          "1:                                                            \n\t"
-                         "ldaxr %2, [%1]                                                \n\t"
-                         "mov   %2, #0                                                  \n\t"
-                         "stlxr %w0, %2, [%1]                                            \n\t"
-                         "cbnz  %w0, 1b                                                  \n\t"
-			 "mov   %0 , #0                                                  \n\t"
-                         : "=r"(ret), "=r"(address), "=r"(tmp)
+                         "ldaxr x2, [%1]                                                \n\t"
+                         "cbnz  x2, 1b                                                  \n\t"
+			 "2:								\n\t"
+                         "stxr  w3, x4, [%1]                                            \n\t"
+                         "cbnz  w3, 1b                                                  \n\t"
+                         "mov   %0, #0                                                  \n\t"
+                         : "=r"(ret), "=r"(address)
                          : "1"(address)
-                         : "memory", "%w0"
-			   //, "%r2" , "%r3"
+                         : "memory", "x2" , "x3", "x4"
 
 
     );
 
+
   } while (ret);
 
 }
 
+#define BLAS_LOCK_DEFINED
+
 
-static inline unsigned long long rpcc(void){
-  unsigned long long ret=0;
-  double v;
-  struct timeval tv;
-  gettimeofday(&tv,NULL);
-  v=(double) tv.tv_sec + (double) tv.tv_usec * 1e-6;
-  ret = (unsigned long long) ( v * 1000.0d );
-  return ret;
-}
 
 static inline int blas_quickdivide(blasint x, blasint y){
   return x / y;
@@ -105,8 +100,10 @@ static inline int blas_quickdivide(blasint x, blasint y){
 #if defined(ASSEMBLER) && !defined(NEEDPARAM)
 
 #define PROLOGUE \
+	.text ;\
+	.align	4 ;\
 	.global	REALNAME ;\
-	.func	REALNAME  ;\
+	.type	REALNAME, %function ;\
 REALNAME:
 
 #define EPILOGUE
@@ -123,7 +120,11 @@ REALNAME:
 #endif
 #define HUGE_PAGESIZE   ( 4 << 20)
 
+#if defined(CORTEXA57)
+#define BUFFER_SIZE     (20 << 20)
+#else
 #define BUFFER_SIZE     (16 << 20)
+#endif
 
 
 #define BASE_ADDRESS (START_ADDRESS - BUFFER_SIZE * MAX_CPU_NUMBER)

common_c.h

@@ -220,6 +220,15 @@
 #define COMATCOPY_K_CTC         comatcopy_k_ctc
 #define COMATCOPY_K_RTC         comatcopy_k_rtc
 
+#define CIMATCOPY_K_CN          cimatcopy_k_cn
+#define CIMATCOPY_K_RN          cimatcopy_k_rn
+#define CIMATCOPY_K_CT          cimatcopy_k_ct
+#define CIMATCOPY_K_RT          cimatcopy_k_rt
+#define CIMATCOPY_K_CNC         cimatcopy_k_cnc
+#define CIMATCOPY_K_RNC         cimatcopy_k_rnc
+#define CIMATCOPY_K_CTC         cimatcopy_k_ctc
+#define CIMATCOPY_K_RTC         cimatcopy_k_rtc
+
 #define CGEADD_K                cgeadd_k 
 
 #else
@@ -403,6 +412,16 @@
 #define COMATCOPY_K_RNC         gotoblas -> comatcopy_k_rnc
 #define COMATCOPY_K_CTC         gotoblas -> comatcopy_k_ctc
 #define COMATCOPY_K_RTC         gotoblas -> comatcopy_k_rtc
+
+#define CIMATCOPY_K_CN          gotoblas -> cimatcopy_k_cn
+#define CIMATCOPY_K_RN          gotoblas -> cimatcopy_k_rn
+#define CIMATCOPY_K_CT          gotoblas -> cimatcopy_k_ct
+#define CIMATCOPY_K_RT          gotoblas -> cimatcopy_k_rt
+#define CIMATCOPY_K_CNC         gotoblas -> cimatcopy_k_cnc
+#define CIMATCOPY_K_RNC         gotoblas -> cimatcopy_k_rnc
+#define CIMATCOPY_K_CTC         gotoblas -> cimatcopy_k_ctc
+#define CIMATCOPY_K_RTC         gotoblas -> cimatcopy_k_rtc
+
 #define CGEADD_K                gotoblas -> cgeadd_k 
 
 #endif

common_d.h

@@ -149,6 +149,11 @@
 #define DOMATCOPY_K_RN		domatcopy_k_rn
 #define DOMATCOPY_K_CT		domatcopy_k_ct
 #define DOMATCOPY_K_RT		domatcopy_k_rt
+
+#define DIMATCOPY_K_CN		dimatcopy_k_cn
+#define DIMATCOPY_K_RN		dimatcopy_k_rn
+#define DIMATCOPY_K_CT      dimatcopy_k_ct
+#define DIMATCOPY_K_RT      dimatcopy_k_rt
 #define DGEADD_K                dgeadd_k 
 
 #else
@@ -267,6 +272,10 @@
 #define DOMATCOPY_K_RN		gotoblas -> domatcopy_k_rn
 #define DOMATCOPY_K_CT		gotoblas -> domatcopy_k_ct
 #define DOMATCOPY_K_RT		gotoblas -> domatcopy_k_rt
+#define DIMATCOPY_K_CN		gotoblas -> dimatcopy_k_cn
+#define DIMATCOPY_K_RN		gotoblas -> dimatcopy_k_rn
+#define DIMATCOPY_K_CT		gotoblas -> dimatcopy_k_ct
+#define DIMATCOPY_K_RT		gotoblas -> dimatcopy_k_rt
 
 #define DGEADD_K                gotoblas -> dgeadd_k 
 

common_ia64.h

@@ -68,6 +68,7 @@ static __inline void blas_lock(volatile unsigned long *address){
 			  : "ar.ccv", "memory");
   } while (ret);
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline unsigned long rpcc(void) {
   unsigned long clocks;
@@ -75,6 +76,7 @@ static __inline unsigned long rpcc(void) {
   __asm__ __volatile__ ("mov %0=ar.itc" : "=r"(clocks));
   return clocks;
 }
+#define RPCC_DEFINED
 
 
 static __inline unsigned long stmxcsr(void){
@@ -99,10 +101,12 @@ static __inline void blas_lock(volatile unsigned long *address){
   while (*address || _InterlockedCompareExchange((volatile int *) address,1,0))
     ;
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline unsigned int rpcc(void) {
   return __getReg(_IA64_REG_AR_ITC);
 }
+#define RPCC_DEFINED
 
 static __inline unsigned int stmxcsr(void) {
   return __getReg(_IA64_REG_AR_FPSR);

common_level1.h

@@ -47,12 +47,12 @@ double dsdot_k(BLASLONG, float   *, BLASLONG, float *, BLASLONG);
 double  ddot_k(BLASLONG, double  *, BLASLONG, double  *, BLASLONG);
 xdouble qdot_k(BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
 
-float   _Complex cdotc_k (BLASLONG, float  *, BLASLONG, float  *, BLASLONG);
-float   _Complex cdotu_k (BLASLONG, float  *, BLASLONG, float  *, BLASLONG);
-double  _Complex zdotc_k (BLASLONG, double *, BLASLONG, double *, BLASLONG);
-double  _Complex zdotu_k (BLASLONG, double *, BLASLONG, double *, BLASLONG);
-xdouble _Complex xdotc_k (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
-xdouble _Complex xdotu_k (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
+openblas_complex_float cdotc_k (BLASLONG, float  *, BLASLONG, float  *, BLASLONG);
+openblas_complex_float cdotu_k (BLASLONG, float  *, BLASLONG, float  *, BLASLONG);
+openblas_complex_double zdotc_k (BLASLONG, double *, BLASLONG, double *, BLASLONG);
+openblas_complex_double zdotu_k (BLASLONG, double *, BLASLONG, double *, BLASLONG);
+openblas_complex_xdouble xdotc_k (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
+openblas_complex_xdouble xdotu_k (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
 
 int    saxpy_k (BLASLONG, BLASLONG, BLASLONG, float,
 	       float  *, BLASLONG, float  *, BLASLONG, float  *, BLASLONG);

common_level3.h

@@ -1736,31 +1736,55 @@ int somatcopy_k_cn(BLASLONG, BLASLONG,  float, float *, BLASLONG, float  *, BLAS
 int somatcopy_k_rn(BLASLONG, BLASLONG,  float, float *, BLASLONG, float  *, BLASLONG);
 int somatcopy_k_ct(BLASLONG, BLASLONG,  float, float *, BLASLONG, float  *, BLASLONG);
 int somatcopy_k_rt(BLASLONG, BLASLONG,  float, float *, BLASLONG, float  *, BLASLONG);
+int simatcopy_k_cn(BLASLONG, BLASLONG,  float, float *, BLASLONG);
+int simatcopy_k_rn(BLASLONG, BLASLONG,  float, float *, BLASLONG);
+int simatcopy_k_ct(BLASLONG, BLASLONG,  float, float *, BLASLONG);
+int simatcopy_k_rt(BLASLONG, BLASLONG,  float, float *, BLASLONG);
 
 int domatcopy_k_cn(BLASLONG, BLASLONG,  double, double *, BLASLONG, double  *, BLASLONG);
 int domatcopy_k_rn(BLASLONG, BLASLONG,  double, double *, BLASLONG, double  *, BLASLONG);
 int domatcopy_k_ct(BLASLONG, BLASLONG,  double, double *, BLASLONG, double  *, BLASLONG);
 int domatcopy_k_rt(BLASLONG, BLASLONG,  double, double *, BLASLONG, double  *, BLASLONG);
+int dimatcopy_k_cn(BLASLONG, BLASLONG,  double, double *, BLASLONG);
+int dimatcopy_k_rn(BLASLONG, BLASLONG,  double, double *, BLASLONG);
+int dimatcopy_k_ct(BLASLONG, BLASLONG,  double, double *, BLASLONG);
+int dimatcopy_k_rt(BLASLONG, BLASLONG,  double, double *, BLASLONG);
 
 int comatcopy_k_cn(BLASLONG, BLASLONG,  float, float, float *, BLASLONG, float  *, BLASLONG);
 int comatcopy_k_rn(BLASLONG, BLASLONG,  float, float, float *, BLASLONG, float  *, BLASLONG);
 int comatcopy_k_ct(BLASLONG, BLASLONG,  float, float, float *, BLASLONG, float  *, BLASLONG);
 int comatcopy_k_rt(BLASLONG, BLASLONG,  float, float, float *, BLASLONG, float  *, BLASLONG);
+int cimatcopy_k_cn(BLASLONG, BLASLONG,  float, float, float *, BLASLONG);
+int cimatcopy_k_rn(BLASLONG, BLASLONG,  float, float, float *, BLASLONG);
+int cimatcopy_k_ct(BLASLONG, BLASLONG,  float, float, float *, BLASLONG);
+int cimatcopy_k_rt(BLASLONG, BLASLONG,  float, float, float *, BLASLONG);
 
 int comatcopy_k_cnc(BLASLONG, BLASLONG,  float, float, float *, BLASLONG, float  *, BLASLONG);
 int comatcopy_k_rnc(BLASLONG, BLASLONG,  float, float, float *, BLASLONG, float  *, BLASLONG);
 int comatcopy_k_ctc(BLASLONG, BLASLONG,  float, float, float *, BLASLONG, float  *, BLASLONG);
 int comatcopy_k_rtc(BLASLONG, BLASLONG,  float, float, float *, BLASLONG, float  *, BLASLONG);
+int cimatcopy_k_cnc(BLASLONG, BLASLONG,  float, float, float *, BLASLONG);
+int cimatcopy_k_rnc(BLASLONG, BLASLONG,  float, float, float *, BLASLONG);
+int cimatcopy_k_ctc(BLASLONG, BLASLONG,  float, float, float *, BLASLONG);
+int cimatcopy_k_rtc(BLASLONG, BLASLONG,  float, float, float *, BLASLONG);
 
 int zomatcopy_k_cn(BLASLONG, BLASLONG,  double, double, double *, BLASLONG, double  *, BLASLONG);
 int zomatcopy_k_rn(BLASLONG, BLASLONG,  double, double, double *, BLASLONG, double  *, BLASLONG);
 int zomatcopy_k_ct(BLASLONG, BLASLONG,  double, double, double *, BLASLONG, double  *, BLASLONG);
 int zomatcopy_k_rt(BLASLONG, BLASLONG,  double, double, double *, BLASLONG, double  *, BLASLONG);
+int zimatcopy_k_cn(BLASLONG, BLASLONG,  double, double, double *, BLASLONG);
+int zimatcopy_k_rn(BLASLONG, BLASLONG,  double, double, double *, BLASLONG);
+int zimatcopy_k_ct(BLASLONG, BLASLONG,  double, double, double *, BLASLONG);
+int zimatcopy_k_rt(BLASLONG, BLASLONG,  double, double, double *, BLASLONG);
 
 int zomatcopy_k_cnc(BLASLONG, BLASLONG,  double, double, double *, BLASLONG, double  *, BLASLONG);
 int zomatcopy_k_rnc(BLASLONG, BLASLONG,  double, double, double *, BLASLONG, double  *, BLASLONG);
 int zomatcopy_k_ctc(BLASLONG, BLASLONG,  double, double, double *, BLASLONG, double  *, BLASLONG);
 int zomatcopy_k_rtc(BLASLONG, BLASLONG,  double, double, double *, BLASLONG, double  *, BLASLONG);
+int zimatcopy_k_cnc(BLASLONG, BLASLONG,  double, double, double *, BLASLONG);
+int zimatcopy_k_rnc(BLASLONG, BLASLONG,  double, double, double *, BLASLONG);
+int zimatcopy_k_ctc(BLASLONG, BLASLONG,  double, double, double *, BLASLONG);
+int zimatcopy_k_rtc(BLASLONG, BLASLONG,  double, double, double *, BLASLONG);
 
 int sgeadd_k(BLASLONG, BLASLONG, float, float*, BLASLONG, float, float *, BLASLONG); 
 int dgeadd_k(BLASLONG, BLASLONG, double, double*, BLASLONG, double, double *, BLASLONG); 

common_macro.h

@@ -634,6 +634,11 @@
 #define OMATCOPY_K_RN		DOMATCOPY_K_RN
 #define OMATCOPY_K_CT		DOMATCOPY_K_CT
 #define OMATCOPY_K_RT		DOMATCOPY_K_RT
+#define IMATCOPY_K_CN		DIMATCOPY_K_CN
+#define IMATCOPY_K_RN		DIMATCOPY_K_RN
+#define IMATCOPY_K_CT		DIMATCOPY_K_CT
+#define IMATCOPY_K_RT		DIMATCOPY_K_RT
+
 #define GEADD_K                 DGEADD_K 
 #else
 
@@ -931,6 +936,10 @@
 #define OMATCOPY_K_RN		SOMATCOPY_K_RN
 #define OMATCOPY_K_CT		SOMATCOPY_K_CT
 #define OMATCOPY_K_RT		SOMATCOPY_K_RT
+#define IMATCOPY_K_CN		SIMATCOPY_K_CN
+#define IMATCOPY_K_RN		SIMATCOPY_K_RN
+#define IMATCOPY_K_CT		SIMATCOPY_K_CT
+#define IMATCOPY_K_RT		SIMATCOPY_K_RT
 
 #define GEADD_K 		SGEADD_K 
 #endif
@@ -1747,6 +1756,15 @@
 #define OMATCOPY_K_RNC		ZOMATCOPY_K_RNC
 #define OMATCOPY_K_CTC		ZOMATCOPY_K_CTC
 #define OMATCOPY_K_RTC		ZOMATCOPY_K_RTC
+#define IMATCOPY_K_CN		ZIMATCOPY_K_CN
+#define IMATCOPY_K_RN		ZIMATCOPY_K_RN
+#define IMATCOPY_K_CT		ZIMATCOPY_K_CT
+#define IMATCOPY_K_RT		ZIMATCOPY_K_RT
+#define IMATCOPY_K_CNC		ZIMATCOPY_K_CNC
+#define IMATCOPY_K_RNC		ZIMATCOPY_K_RNC
+#define IMATCOPY_K_CTC		ZIMATCOPY_K_CTC
+#define IMATCOPY_K_RTC		ZIMATCOPY_K_RTC
+
 #define GEADD_K                 ZGEADD_K 
 
 #else
@@ -2160,6 +2178,14 @@
 #define OMATCOPY_K_RNC		COMATCOPY_K_RNC
 #define OMATCOPY_K_CTC		COMATCOPY_K_CTC
 #define OMATCOPY_K_RTC		COMATCOPY_K_RTC
+#define IMATCOPY_K_CN		CIMATCOPY_K_CN
+#define IMATCOPY_K_RN		CIMATCOPY_K_RN
+#define IMATCOPY_K_CT		CIMATCOPY_K_CT
+#define IMATCOPY_K_RT		CIMATCOPY_K_RT
+#define IMATCOPY_K_CNC		CIMATCOPY_K_CNC
+#define IMATCOPY_K_RNC		CIMATCOPY_K_RNC
+#define IMATCOPY_K_CTC		CIMATCOPY_K_CTC
+#define IMATCOPY_K_RTC		CIMATCOPY_K_RTC
 
 #define GEADD_K                 CGEADD_K 
 

common_mips64.h

@@ -98,6 +98,7 @@ static void INLINE blas_lock(volatile unsigned long *address){
 
   } while (ret);
 }
+#define BLAS_LOCK_DEFINED
 
 static inline unsigned int rpcc(void){
   unsigned long ret;
@@ -118,6 +119,7 @@ static inline unsigned int rpcc(void){
 #endif
   return ret;
 }
+#define RPCC_DEFINED
 
 #if defined(LOONGSON3A) || defined(LOONGSON3B)
 #ifndef NO_AFFINITY

common_param.h

@@ -855,6 +855,36 @@ BLASLONG (*ixamin_k)(BLASLONG, xdouble *, BLASLONG);
   int    (*zomatcopy_k_rnc)	(BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
   int    (*zomatcopy_k_rtc)	(BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
 
+  int    (*simatcopy_k_cn)	(BLASLONG, BLASLONG, float, float*, BLASLONG);
+  int    (*simatcopy_k_ct)	(BLASLONG, BLASLONG, float, float*, BLASLONG);
+  int    (*simatcopy_k_rn)	(BLASLONG, BLASLONG, float, float*, BLASLONG);
+  int    (*simatcopy_k_rt)	(BLASLONG, BLASLONG, float, float*, BLASLONG);
+
+  int    (*dimatcopy_k_cn)	(BLASLONG, BLASLONG, double, double*, BLASLONG);
+  int    (*dimatcopy_k_ct)	(BLASLONG, BLASLONG, double, double*, BLASLONG);
+  int    (*dimatcopy_k_rn)	(BLASLONG, BLASLONG, double, double*, BLASLONG);
+  int    (*dimatcopy_k_rt)	(BLASLONG, BLASLONG, double, double*, BLASLONG);
+
+  int    (*cimatcopy_k_cn)	(BLASLONG, BLASLONG, float, float, float*, BLASLONG);
+  int    (*cimatcopy_k_ct)	(BLASLONG, BLASLONG, float, float, float*, BLASLONG);
+  int    (*cimatcopy_k_rn)	(BLASLONG, BLASLONG, float, float, float*, BLASLONG);
+  int    (*cimatcopy_k_rt)	(BLASLONG, BLASLONG, float, float, float*, BLASLONG);
+
+  int    (*cimatcopy_k_cnc)	(BLASLONG, BLASLONG, float, float, float*, BLASLONG);
+  int    (*cimatcopy_k_ctc)	(BLASLONG, BLASLONG, float, float, float*, BLASLONG);
+  int    (*cimatcopy_k_rnc)	(BLASLONG, BLASLONG, float, float, float*, BLASLONG);
+  int    (*cimatcopy_k_rtc)	(BLASLONG, BLASLONG, float, float, float*, BLASLONG);
+
+  int    (*zimatcopy_k_cn)	(BLASLONG, BLASLONG, double, double, double*, BLASLONG);
+  int    (*zimatcopy_k_ct)	(BLASLONG, BLASLONG, double, double, double*, BLASLONG);
+  int    (*zimatcopy_k_rn)	(BLASLONG, BLASLONG, double, double, double*, BLASLONG);
+  int    (*zimatcopy_k_rt)	(BLASLONG, BLASLONG, double, double, double*, BLASLONG);
+
+  int    (*zimatcopy_k_cnc)	(BLASLONG, BLASLONG, double, double, double*, BLASLONG);
+  int    (*zimatcopy_k_ctc)	(BLASLONG, BLASLONG, double, double, double*, BLASLONG);
+  int    (*zimatcopy_k_rnc)	(BLASLONG, BLASLONG, double, double, double*, BLASLONG);
+  int    (*zimatcopy_k_rtc)	(BLASLONG, BLASLONG, double, double, double*, BLASLONG);
+
   int    (*sgeadd_k) (BLASLONG, BLASLONG, float, float *, BLASLONG, float, float *, BLASLONG); 
   int    (*dgeadd_k) (BLASLONG, BLASLONG, double, double *, BLASLONG, double, double *, BLASLONG); 
   int    (*cgeadd_k) (BLASLONG, BLASLONG, float, float,  float *,  BLASLONG, float, float, float *, BLASLONG); 
@@ -1164,6 +1194,8 @@ extern gotoblas_t *gotoblas;
 #define XGEMM_DEFAULT_UNROLL_N 2
 #endif
 
+#define GEMM_THREAD gemm_thread_m
+
 #ifndef GEMM_THREAD
 #define GEMM_THREAD gemm_thread_n
 #endif

common_power.h

@@ -87,6 +87,7 @@ static void INLINE blas_lock(volatile unsigned long *address){
 #endif
   } while (ret);
 }
+#define BLAS_LOCK_DEFINED
 
 static inline unsigned long rpcc(void){
   unsigned long ret;
@@ -103,6 +104,7 @@ static inline unsigned long rpcc(void){
 #endif
 
 }
+#define RPCC_DEFINED
 
 #ifdef __64BIT__
 #define RPCC64BIT
@@ -495,6 +497,15 @@ static inline int blas_quickdivide(blasint x, blasint y){
 REALNAME:
 #define EPILOGUE	.size	REALNAME, .-REALNAME
 #else
+#if _CALL_ELF == 2
+#define PROLOGUE \
+	.section .text;\
+	.align 6;\
+	.globl	REALNAME;\
+	.type	REALNAME, @function;\
+REALNAME:
+#define EPILOGUE	.size	REALNAME, .-REALNAME
+#else
 #define PROLOGUE \
 	.section .text;\
 	.align 5;\
@@ -514,6 +525,7 @@ REALNAME:;\
 	.size	.REALNAME, .-.REALNAME; \
 	.section	.note.GNU-stack,"",@progbits
 #endif
+#endif
 
 #ifdef PROFILE
 #ifndef __64BIT__
@@ -792,4 +804,25 @@ Lmcount$lazy_ptr:
 #ifndef MAP_ANONYMOUS
 #define MAP_ANONYMOUS MAP_ANON
 #endif
+
+#ifdef OS_LINUX
+#ifndef __64BIT__
+#define FRAMESLOT(X) (((X) * 4) + 8)
+#else
+#if _CALL_ELF == 2
+#define FRAMESLOT(X) (((X) * 8) + 96)
+#else
+#define FRAMESLOT(X) (((X) * 8) + 112)
+#endif
+#endif
+#endif
+
+#if defined(OS_AIX) || defined(OS_DARWIN)
+#ifndef __64BIT__
+#define FRAMESLOT(X) (((X) * 4) + 56)
+#else
+#define FRAMESLOT(X) (((X) * 8) + 112)
+#endif
+#endif
+
 #endif

common_s.h

@@ -152,6 +152,10 @@
 #define SOMATCOPY_K_RN          somatcopy_k_rn
 #define SOMATCOPY_K_CT          somatcopy_k_ct
 #define SOMATCOPY_K_RT          somatcopy_k_rt
+#define SIMATCOPY_K_CN          simatcopy_k_cn
+#define SIMATCOPY_K_RN          simatcopy_k_rn
+#define SIMATCOPY_K_CT          simatcopy_k_ct
+#define SIMATCOPY_K_RT          simatcopy_k_rt
 
 #define SGEADD_K                sgeadd_k 
 
@@ -274,6 +278,10 @@
 #define SOMATCOPY_K_RN          gotoblas -> somatcopy_k_rn
 #define SOMATCOPY_K_CT          gotoblas -> somatcopy_k_ct
 #define SOMATCOPY_K_RT          gotoblas -> somatcopy_k_rt
+#define SIMATCOPY_K_CN          gotoblas -> simatcopy_k_cn
+#define SIMATCOPY_K_RN          gotoblas -> simatcopy_k_rn
+#define SIMATCOPY_K_CT          gotoblas -> simatcopy_k_ct
+#define SIMATCOPY_K_RT          gotoblas -> simatcopy_k_rt
 
 #define SGEADD_K                gotoblas -> sgeadd_k 
 

common_sparc.h

@@ -58,6 +58,7 @@ static void __inline blas_lock(volatile unsigned long *address){
 			 : "memory");
   } while (ret);
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline unsigned long rpcc(void){
   unsigned long clocks;
@@ -66,6 +67,7 @@ static __inline unsigned long rpcc(void){
 
   return clocks;
 };
+#define RPCC_DEFINED
 
 #ifdef __64BIT__
 #define RPCC64BIT

common_x86.h

@@ -56,41 +56,67 @@ static void __inline blas_lock(volatile BLASULONG *address){
   do {
     while (*address) {YIELDING;};
 
+#if defined(_MSC_VER) && !defined(__clang__)
+	// use intrinsic instead of inline assembly
+	ret = _InterlockedExchange(address, 1);
+	// inline assembly
+	/*__asm {
+		mov eax, address
+		mov ebx, 1
+		xchg [eax], ebx
+		mov ret, ebx
+	}*/
+#else
     __asm__ __volatile__(
 			 "xchgl %0, %1\n"
 			 : "=r"(ret), "=m"(*address)
 			 : "0"(1), "m"(*address)
 			 : "memory");
+#endif
 
   } while (ret);
 
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline unsigned long long rpcc(void){
+#if defined(_MSC_VER) && !defined(__clang__)
+  return __rdtsc(); // use MSVC intrinsic
+#else
   unsigned int a, d;
 
   __asm__ __volatile__ ("rdtsc" : "=a" (a), "=d" (d));
 
   return ((unsigned long long)a + ((unsigned long long)d << 32));
+#endif
 };
+#define RPCC_DEFINED
 
 static __inline unsigned long getstackaddr(void){
+#if defined(_MSC_VER) && !defined(__clang__)
+  return (unsigned long)_ReturnAddress(); // use MSVC intrinsic
+#else
   unsigned long addr;
 
   __asm__ __volatile__ ("mov %%esp, %0"
 			 : "=r"(addr) : : "memory");
 
   return addr;
+#endif
 };
 
 
 static __inline long double sqrt_long(long double val) {
+#if defined(_MSC_VER) && !defined(__clang__)
+  return sqrt(val); // not sure if this will use fsqrt
+#else
   long double result;
 
   __asm__ __volatile__ ("fldt %1\n"
 		    "fsqrt\n"
 		    "fstpt %0\n" : "=m" (result) : "m"(val));
   return result;
+#endif
 }
 
 #define SQRT(a)  sqrt_long(a)
@@ -100,7 +126,7 @@ void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx);
 
 #define WHEREAMI
 
-static inline int WhereAmI(void){
+static __inline int WhereAmI(void){
   int eax, ebx, ecx, edx;
   int apicid;
 
@@ -146,9 +172,14 @@ static __inline int blas_quickdivide(unsigned int x, unsigned int y){
 
   y = blas_quick_divide_table[y];
 
+#if defined(_MSC_VER) && !defined(__clang__)
+  (void*)result;
+  return x*y;
+#else
   __asm__ __volatile__  ("mull %0" :"=d" (result) :"a"(x), "0" (y));
 
   return result;
+#endif
 }
 #endif
 
@@ -284,8 +315,12 @@ REALNAME:
 
 #define PROFCODE
 
+#ifdef __clang__
+#define EPILOGUE .end
+#else
 #define EPILOGUE .end	 REALNAME
 #endif
+#endif
 
 #if defined(OS_LINUX) || defined(OS_FREEBSD) || defined(OS_NETBSD) || defined(__ELF__)
 #define PROLOGUE \

common_x86_64.h

@@ -41,6 +41,10 @@
 
 #ifndef ASSEMBLER
 
+#ifdef C_MSVC
+#include <intrin.h>
+#endif
+
 #ifdef C_SUN
 #define	__asm__ __asm
 #define	__volatile__
@@ -61,30 +65,45 @@
 
 static void __inline blas_lock(volatile BLASULONG *address){
 
+#ifndef C_MSVC
   int ret;
+#else
+  BLASULONG ret;
+#endif
 
   do {
     while (*address) {YIELDING;};
 
+#ifndef C_MSVC
     __asm__ __volatile__(
 			 "xchgl %0, %1\n"
 			 : "=r"(ret), "=m"(*address)
 			 : "0"(1), "m"(*address)
 			 : "memory");
-
+#else
+    ret=InterlockedExchange64((volatile LONG64 *)(address), 1);
+#endif
   } while (ret);
+
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline BLASULONG rpcc(void){
+#ifdef C_MSVC
+  return __rdtsc();
+#else
   BLASULONG a, d;
 
   __asm__ __volatile__ ("rdtsc" : "=a" (a), "=d" (d));
 
   return ((BLASULONG)a + ((BLASULONG)d << 32));
+#endif
 }
+#define RPCC_DEFINED
 
 #define RPCC64BIT
 
+#ifndef C_MSVC
 static __inline BLASULONG getstackaddr(void){
   BLASULONG addr;
 
@@ -93,22 +112,32 @@ static __inline BLASULONG getstackaddr(void){
 
   return addr;
 }
+#endif
 
 static __inline void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx){
 
+#ifdef C_MSVC
+  int cpuinfo[4];
+  __cpuid(cpuinfo, op);
+  *eax=cpuinfo[0];
+  *ebx=cpuinfo[1];
+  *ecx=cpuinfo[2];
+  *edx=cpuinfo[3];
+#else
         __asm__ __volatile__("cpuid"
 			     : "=a" (*eax),
 			     "=b" (*ebx),
 			     "=c" (*ecx),
 			     "=d" (*edx)
 			     : "0" (op));
+#endif
 }
 
 /*
 #define WHEREAMI
 */
 
-static inline int WhereAmI(void){
+static __inline int WhereAmI(void){
   int eax, ebx, ecx, edx;
   int apicid;
 
@@ -150,10 +179,14 @@ static inline int WhereAmI(void){
 #define GET_IMAGE_CANCEL
 
 #ifdef SMP
-#ifdef USE64BITINT
+#if defined(USE64BITINT)
 static __inline blasint blas_quickdivide(blasint x, blasint y){
   return x / y;
 }
+#elif defined (C_MSVC)
+static __inline BLASLONG blas_quickdivide(BLASLONG x, BLASLONG y){
+  return x / y;
+}
 #else
 extern unsigned int blas_quick_divide_table[];
 

common_z.h

@@ -220,6 +220,15 @@
 #define ZOMATCOPY_K_CTC         zomatcopy_k_ctc
 #define ZOMATCOPY_K_RTC         zomatcopy_k_rtc
 
+#define ZIMATCOPY_K_CN          zimatcopy_k_cn
+#define ZIMATCOPY_K_RN          zimatcopy_k_rn
+#define ZIMATCOPY_K_CT          zimatcopy_k_ct
+#define ZIMATCOPY_K_RT          zimatcopy_k_rt
+#define ZIMATCOPY_K_CNC         zimatcopy_k_cnc
+#define ZIMATCOPY_K_RNC         zimatcopy_k_rnc
+#define ZIMATCOPY_K_CTC         zimatcopy_k_ctc
+#define ZIMATCOPY_K_RTC         zimatcopy_k_rtc
+
 #define ZGEADD_K                zgeadd_k 
 
 #else
@@ -404,6 +413,15 @@
 #define ZOMATCOPY_K_CTC         gotoblas -> zomatcopy_k_ctc
 #define ZOMATCOPY_K_RTC         gotoblas -> zomatcopy_k_rtc
 
+#define ZIMATCOPY_K_CN          gotoblas -> zimatcopy_k_cn
+#define ZIMATCOPY_K_RN          gotoblas -> zimatcopy_k_rn
+#define ZIMATCOPY_K_CT          gotoblas -> zimatcopy_k_ct
+#define ZIMATCOPY_K_RT          gotoblas -> zimatcopy_k_rt
+#define ZIMATCOPY_K_CNC         gotoblas -> zimatcopy_k_cnc
+#define ZIMATCOPY_K_RNC         gotoblas -> zimatcopy_k_rnc
+#define ZIMATCOPY_K_CTC         gotoblas -> zimatcopy_k_ctc
+#define ZIMATCOPY_K_RTC         gotoblas -> zimatcopy_k_rtc
+
 #define ZGEADD_K                gotoblas -> zgeadd_k
 
 #endif

cpuid.h

@@ -39,6 +39,10 @@
 #ifndef CPUID_H
 #define CPUID_H
 
+#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
+#define INTEL_AMD
+#endif
+
 #define VENDOR_INTEL      1
 #define VENDOR_UMC        2
 #define VENDOR_AMD        3
@@ -59,7 +63,7 @@
 #define FAMILY_PM     7
 #define FAMILY_IA64   8
 
-#if defined(__i386__) || defined(__x86_64__)
+#ifdef INTEL_AMD
 #define GET_EXFAMILY  1
 #define GET_EXMODEL   2
 #define GET_TYPE      3

cpuid_arm.c

@@ -192,6 +192,7 @@ void get_cpuconfig(void)
 	{
 	       case CPU_CORTEXA9:
     			printf("#define CORTEXA9\n");
+    			printf("#define ARMV7\n");
     			printf("#define HAVE_VFP\n");
     			printf("#define HAVE_VFPV3\n");
 			if ( get_feature("neon"))	printf("#define HAVE_NEON\n");
@@ -207,6 +208,7 @@ void get_cpuconfig(void)
 
 	       case CPU_CORTEXA15:
     			printf("#define CORTEXA15\n");
+    			printf("#define ARMV7\n");
     			printf("#define HAVE_VFP\n");
     			printf("#define HAVE_VFPV3\n");
 			if ( get_feature("neon"))	printf("#define HAVE_NEON\n");

cpuid_arm64.c

@@ -29,12 +29,19 @@
 
 #define CPU_UNKNOWN     	0
 #define CPU_ARMV8       	1
+#define CPU_CORTEXA57       	2
 
 static char *cpuname[] = {
-  "UNKOWN",
-  "ARMV8"
+  "UNKNOWN",
+  "ARMV8" ,
+  "CORTEXA57"
 };
 
+static char *cpuname_lower[] = {
+  "unknown",
+  "armv8" ,
+  "cortexa57"
+};
 
 int get_feature(char *search)
 {
@@ -53,13 +60,13 @@ int get_feature(char *search)
 		{
 			p = strchr(buffer, ':') + 2;
 			break;
-      		}
-  	}
+		}
+	}
 
-  	fclose(infile);
+	fclose(infile);
 
 
-	if( p == NULL ) return;
+	if( p == NULL ) return 0;
 
 	t = strtok(p," ");
 	while( t = strtok(NULL," "))
@@ -82,11 +89,30 @@ int detect(void)
   	p = (char *) NULL ;
 
   	infile = fopen("/proc/cpuinfo", "r");
-
 	while (fgets(buffer, sizeof(buffer), infile))
 	{
 
-		if ((!strncmp("model name", buffer, 10)) || (!strncmp("Processor", buffer, 9)))
+		if (!strncmp("CPU part", buffer, 8))
+		{
+			p = strchr(buffer, ':') + 2;
+			break;
+		}
+	}
+
+	fclose(infile);
+	if(p != NULL) {
+	  if (strstr(p, "0xd07")) {
+	    return CPU_CORTEXA57;
+	  }
+	}
+
+	p = (char *) NULL ;
+	infile = fopen("/proc/cpuinfo", "r");
+	while (fgets(buffer, sizeof(buffer), infile))
+	{
+
+		if ((!strncmp("model name", buffer, 10)) || (!strncmp("Processor", buffer, 9)) ||
+		    (!strncmp("CPU architecture", buffer, 16)))
 		{
 			p = strchr(buffer, ':') + 2;
 			break;
@@ -100,7 +126,7 @@ int detect(void)
 
 		if (strstr(p, "AArch64"))
 		{
-			 	return CPU_ARMV8;
+			return CPU_ARMV8;
 
 		}
 
@@ -118,23 +144,13 @@ char *get_corename(void)
 
 void get_architecture(void)
 {
-	printf("ARM");
+	printf("ARM64");
 }
 
 void get_subarchitecture(void)
 {
 	int d = detect();
-	switch (d)
-	{
-
-		case CPU_ARMV8:
-			printf("ARMV8");
-			break;
-
-		default:
-			printf("UNKNOWN");
-			break;
-	}
+	printf("%s", cpuname[d]);
 }
 
 void get_subdirname(void)
@@ -160,26 +176,32 @@ void get_cpuconfig(void)
     			printf("#define L2_ASSOCIATIVE 4\n");
 			break;
 
-
+		case CPU_CORTEXA57:
+			printf("#define CORTEXA57\n");
+			printf("#define HAVE_VFP\n");
+			printf("#define HAVE_VFPV3\n");
+			printf("#define HAVE_NEON\n");
+			printf("#define HAVE_VFPV4\n");
+			printf("#define L1_CODE_SIZE 49152\n");
+			printf("#define L1_CODE_LINESIZE 64\n");
+			printf("#define L1_CODE_ASSOCIATIVE 3\n");
+			printf("#define L1_DATA_SIZE 32768\n");
+			printf("#define L1_DATA_LINESIZE 64\n");
+			printf("#define L1_DATA_ASSOCIATIVE 2\n");
+			printf("#define L2_SIZE 2097152\n");
+			printf("#define L2_LINESIZE 64\n");
+			printf("#define L2_ASSOCIATIVE 16\n");
+			break;
 	}
 }
 
 
 void get_libname(void)
 {
-
 	int d = detect();
-	switch (d)
-	{
-
-		case CPU_ARMV8:
-    			printf("armv8\n");
-			break;
-
-	}
+	printf("%s", cpuname_lower[d]);
 }
 
-
 void get_features(void)
 {
 

cpuid_power.c

@@ -115,6 +115,7 @@ int detect(void){
   if (!strncasecmp(p, "POWER5", 6)) return CPUTYPE_POWER5;
   if (!strncasecmp(p, "POWER6", 6)) return CPUTYPE_POWER6;
   if (!strncasecmp(p, "POWER7", 6)) return CPUTYPE_POWER6;
+  if (!strncasecmp(p, "POWER8", 6)) return CPUTYPE_POWER6;
   if (!strncasecmp(p, "Cell",   4)) return CPUTYPE_CELL;
   if (!strncasecmp(p, "7447",   4)) return CPUTYPE_PPCG4;
 

cpuid_x86.c

@@ -40,6 +40,12 @@
 #include <string.h>
 #include "cpuid.h"
 
+#if defined(_MSC_VER) && !defined(__clang__)
+#define C_INLINE __inline
+#else
+#define C_INLINE inline
+#endif
+
 /*
 #ifdef NO_AVX
 #define CPUTYPE_HASWELL CPUTYPE_NEHALEM
@@ -53,12 +59,26 @@
 #endif
 */
 
+#if defined(_MSC_VER) && !defined(__clang__)
+
+void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx)
+{
+  int cpuInfo[4] = {-1};
+  __cpuid(cpuInfo, op);
+  *eax = cpuInfo[0];
+  *ebx = cpuInfo[1];
+  *ecx = cpuInfo[2];
+  *edx = cpuInfo[3];
+}
+
+#else
+
 #ifndef CPUIDEMU
 
 #if defined(__APPLE__) && defined(__i386__)
 void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx);
 #else
-static inline void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx){
+static C_INLINE void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx){
 #if defined(__i386__) && defined(__PIC__)
   __asm__ __volatile__
     ("mov %%ebx, %%edi;"
@@ -115,14 +135,16 @@ void cpuid(unsigned int op, unsigned int *eax, unsigned int *ebx, unsigned int *
 
 #endif
 
-static inline int have_cpuid(void){
+#endif // _MSC_VER
+
+static C_INLINE int have_cpuid(void){
   int eax, ebx, ecx, edx;
 
   cpuid(0, &eax, &ebx, &ecx, &edx);
   return eax;
 }
 
-static inline int have_excpuid(void){
+static C_INLINE int have_excpuid(void){
   int eax, ebx, ecx, edx;
 
   cpuid(0x80000000, &eax, &ebx, &ecx, &edx);
@@ -130,10 +152,14 @@ static inline int have_excpuid(void){
 }
 
 #ifndef NO_AVX
-static inline void xgetbv(int op, int * eax, int * edx){
+static C_INLINE void xgetbv(int op, int * eax, int * edx){
   //Use binary code for xgetbv
+#if defined(_MSC_VER) && !defined(__clang__)
+  *eax = __xgetbv(op);
+#else
   __asm__ __volatile__
     (".byte 0x0f, 0x01, 0xd0": "=a" (*eax), "=d" (*edx) : "c" (op) : "cc");
+#endif
 }
 #endif
 
@@ -1133,6 +1159,16 @@ int get_cpuname(void){
             return CPUTYPE_HASWELL;
 #else
 	    return CPUTYPE_SANDYBRIDGE;
+#endif
+          else
+	    return CPUTYPE_NEHALEM;
+	case 14:
+	  //Skylake
+          if(support_avx())
+#ifndef NO_AVX2
+            return CPUTYPE_HASWELL;
+#else
+	    return CPUTYPE_SANDYBRIDGE;
 #endif
           else
 	    return CPUTYPE_NEHALEM;
@@ -1147,6 +1183,17 @@ int get_cpuname(void){
             return CPUTYPE_HASWELL;
 #else
 	    return CPUTYPE_SANDYBRIDGE;
+#endif
+          else
+	    return CPUTYPE_NEHALEM;
+	case 5:
+        case 14:
+	  // Skylake
+          if(support_avx())
+#ifndef NO_AVX2
+            return CPUTYPE_HASWELL;
+#else
+	    return CPUTYPE_SANDYBRIDGE;
 #endif
           else
 	    return CPUTYPE_NEHALEM;
@@ -1192,13 +1239,19 @@ int get_cpuname(void){
 	    return CPUTYPE_BULLDOZER;
 	  else
 	    return CPUTYPE_BARCELONA; //OS don't support AVX.
-	case 2:
+	case 2: //AMD Piledriver
+	case 3: //AMD Richland
 	  if(support_avx())
 	    return CPUTYPE_PILEDRIVER;
 	  else
 	    return CPUTYPE_BARCELONA; //OS don't support AVX.
 	case 0:
 	  switch(exmodel){
+	  case 1: //AMD Trinity
+	    if(support_avx())
+	      return CPUTYPE_PILEDRIVER;
+	    else
+	      return CPUTYPE_BARCELONA; //OS don't support AVX.
 	  case 3:
 	    if(support_avx())
 	      return CPUTYPE_STEAMROLLER;
@@ -1608,6 +1661,16 @@ int get_coretype(void){
             return CORE_HASWELL;
 #else
 	    return CORE_SANDYBRIDGE;
+#endif
+          else
+	    return CORE_NEHALEM;
+	case 14:
+	  //Skylake
+          if(support_avx())
+#ifndef NO_AVX2
+            return CORE_HASWELL;
+#else
+	    return CORE_SANDYBRIDGE;
 #endif
           else
 	    return CORE_NEHALEM;
@@ -1622,6 +1685,17 @@ int get_coretype(void){
             return CORE_HASWELL;
 #else
 	    return CORE_SANDYBRIDGE;
+#endif
+          else
+	    return CORE_NEHALEM;
+	case 5:
+	case 14:
+	  // Skylake
+          if(support_avx())
+#ifndef NO_AVX2
+            return CORE_HASWELL;
+#else
+	    return CORE_SANDYBRIDGE;
 #endif
           else
 	    return CORE_NEHALEM;
@@ -1650,7 +1724,8 @@ int get_coretype(void){
 	    return CORE_BULLDOZER;
 	  else
 	    return CORE_BARCELONA; //OS don't support AVX.
-	case 2:
+	case 2: //AMD Piledriver
+	case 3: //AMD Richland
 	  if(support_avx())
 	    return CORE_PILEDRIVER;
 	  else
@@ -1658,6 +1733,12 @@ int get_coretype(void){
 	
 	case 0:
 	  switch(exmodel){
+	  case 1: //AMD Trinity
+	    if(support_avx())
+	      return CORE_PILEDRIVER;
+	    else
+	      return CORE_BARCELONA; //OS don't support AVX.
+
 	  case 3:
 	    if(support_avx())
 	      return CORE_STEAMROLLER;

ctest/CMakeLists.txt

@@ -0,0 +1,46 @@
+include_directories(${CMAKE_SOURCE_DIR})
+
+enable_language(Fortran)
+
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DADD${BU} -DCBLAS")
+
+FILE(WRITE ${CMAKE_CURRENT_BINARY_DIR}/test_cblas_helper.sh
+"$1 < $2\n"
+)
+
+foreach(float_type ${FLOAT_TYPES})
+  string(SUBSTRING ${float_type} 0 1 float_char_upper)
+  string(TOLOWER ${float_char_upper} float_char)
+  #level1
+  add_executable(x${float_char}cblat1
+    c_${float_char}blat1.f
+    c_${float_char}blas1.c)
+  target_link_libraries(x${float_char}cblat1 ${OpenBLAS_LIBNAME}_static)
+  add_test(NAME "x${float_char}cblat1"
+    COMMAND "${CMAKE_CURRENT_BINARY_DIR}/x${float_char}cblat1")
+
+  #level2
+  add_executable(x${float_char}cblat2
+    c_${float_char}blat2.f
+    c_${float_char}blas2.c
+    c_${float_char}2chke.c
+    auxiliary.c
+    c_xerbla.c
+    constant.c)
+  target_link_libraries(x${float_char}cblat2 ${OpenBLAS_LIBNAME}_static)
+  add_test(NAME "x${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/${float_char}in2")
+
+  #level3
+  add_executable(x${float_char}cblat3
+    c_${float_char}blat3.f
+    c_${float_char}blas3.c
+    c_${float_char}3chke.c
+    auxiliary.c
+    c_xerbla.c
+    constant.c)
+  target_link_libraries(x${float_char}cblat3 ${OpenBLAS_LIBNAME}_static)
+  add_test(NAME "x${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")
+
+endforeach()

driver/level2/CMakeLists.txt

@@ -0,0 +1,203 @@
+
+include_directories(${CMAKE_SOURCE_DIR})
+
+# sources that need to be compiled twice, once with no flags and once with LOWER
+set(UL_SOURCES
+  sbmv_k.c
+  spmv_k.c
+  spr_k.c
+  spr2_k.c
+  syr_k.c
+  syr2_k.c
+)
+
+# sources that need to be compiled several times, for UNIT, TRANSA
+set(U_SOURCES
+  trmv_U.c
+  tbmv_U.c
+  tbsv_U.c
+  tpmv_U.c
+  tpsv_U.c
+  trsv_U.c
+)
+
+set(L_SOURCES
+  trmv_L.c
+  tbmv_L.c
+  tbsv_L.c
+  tpmv_L.c
+  tpsv_L.c
+  trsv_L.c
+)
+
+set(UL_SMP_SOURCES
+  symv_thread.c
+  syr_thread.c
+  syr2_thread.c
+  spr_thread.c
+  spr2_thread.c
+  spmv_thread.c
+  sbmv_thread.c
+)
+
+set(NU_SMP_SOURCES
+  trmv_thread.c
+  tpmv_thread.c
+  tbmv_thread.c
+)
+
+set(ULVM_COMPLEX_SOURCES
+  hbmv_k.c
+  hpmv_k.c
+  hpr_k.c
+  hpr2_k.c
+  her_k.c
+  her2_k.c
+)
+
+# objects that need LOWER set
+GenerateCombinationObjects("${UL_SOURCES}" "LOWER" "U" "" 1 "" "" 3)
+
+# gbmv uses a lowercase n and t
+GenerateNamedObjects("gbmv_k.c" "" "gbmv_n" false "" "" "" 3)
+GenerateNamedObjects("gbmv_k.c" "TRANS" "gbmv_t" false "" "" "" 3)
+# c/zgbmv 
+GenerateNamedObjects("zgbmv_k.c" "CONJ" "gbmv_r" false "" "" "" 2)
+GenerateNamedObjects("zgbmv_k.c" "TRANS;CONJ" "gbmv_c" false "" "" "" 2)
+GenerateNamedObjects("zgbmv_k.c" "XCONJ" "gbmv_o" false "" "" "" 2)
+GenerateNamedObjects("zgbmv_k.c" "TRANS;XCONJ" "gbmv_u" false "" "" "" 2)
+GenerateNamedObjects("zgbmv_k.c" "CONJ;XCONJ" "gbmv_s" false "" "" "" 2)
+GenerateNamedObjects("zgbmv_k.c" "TRANS;CONJ;XCONJ" "gbmv_d" false "" "" "" 2)
+
+# special defines for complex
+foreach (float_type ${FLOAT_TYPES})
+
+  if (SMP)
+    GenerateNamedObjects("gemv_thread.c" "" "gemv_thread_n" false "" "" false ${float_type})
+    GenerateNamedObjects("gemv_thread.c" "TRANSA" "gemv_thread_t" false "" "" false ${float_type})
+
+    GenerateNamedObjects("gbmv_thread.c" "" "gbmv_thread_n" false "" "" false ${float_type})
+    GenerateNamedObjects("gbmv_thread.c" "TRANSA" "gbmv_thread_t" false "" "" false ${float_type})
+  endif ()
+
+  if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
+
+    foreach (u_source ${U_SOURCES})
+      string(REGEX MATCH "[a-z]+" op_name ${u_source})
+      GenerateCombinationObjects("z${u_source}" "UNIT" "N" "TRANSA=1" 0 "${op_name}_NU" false ${float_type})
+      GenerateCombinationObjects("z${u_source}" "UNIT" "N" "TRANSA=2" 0 "${op_name}_TL" false ${float_type})
+      GenerateCombinationObjects("z${u_source}" "UNIT" "N" "TRANSA=3" 0 "${op_name}_RU" false ${float_type})
+      GenerateCombinationObjects("z${u_source}" "UNIT" "N" "TRANSA=4" 0 "${op_name}_CL" false ${float_type})
+    endforeach ()
+
+    foreach (l_source ${L_SOURCES})
+      string(REGEX MATCH "[a-z]+" op_name ${l_source})
+      GenerateCombinationObjects("z${l_source}" "UNIT" "N" "TRANSA=1" 0 "${op_name}_NL" false ${float_type})
+      GenerateCombinationObjects("z${l_source}" "UNIT" "N" "TRANSA=2" 0 "${op_name}_TU" false ${float_type})
+      GenerateCombinationObjects("z${l_source}" "UNIT" "N" "TRANSA=3" 0 "${op_name}_RL" false ${float_type})
+      GenerateCombinationObjects("z${l_source}" "UNIT" "N" "TRANSA=4" 0 "${op_name}_CU" false ${float_type})
+    endforeach ()
+
+    foreach (ulvm_source ${ULVM_COMPLEX_SOURCES})
+      string(REGEX MATCH "[a-z0-9]+" op_name ${ulvm_source})
+      GenerateNamedObjects("z${ulvm_source}" "" "${op_name}_U" false "" "" false ${float_type})
+      GenerateNamedObjects("z${ulvm_source}" "LOWER" "${op_name}_L" false "" "" false ${float_type})
+      GenerateNamedObjects("z${ulvm_source}" "HEMVREV" "${op_name}_V" false "" "" false ${float_type})
+      GenerateNamedObjects("z${ulvm_source}" "LOWER;HEMVREV" "${op_name}_M" false "" "" false ${float_type})
+    endforeach()
+
+    if (SMP)
+
+      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})
+      GenerateNamedObjects("gemv_thread.c" "XCONJ" "gemv_thread_o" false "" "" false ${float_type})
+      GenerateNamedObjects("gemv_thread.c" "XCONJ;TRANSA" "gemv_thread_u" false "" "" false ${float_type})
+      GenerateNamedObjects("gemv_thread.c" "XCONJ;CONJ" "gemv_thread_s" false "" "" false ${float_type})
+      GenerateNamedObjects("gemv_thread.c" "XCONJ;CONJ;TRANSA" "gemv_thread_d" false "" "" false ${float_type})
+
+      GenerateNamedObjects("gbmv_thread.c" "CONJ" "gbmv_thread_r" false "" "" false ${float_type})
+      GenerateNamedObjects("gbmv_thread.c" "CONJ;TRANSA" "gbmv_thread_c" false "" "" false ${float_type})
+      GenerateNamedObjects("gbmv_thread.c" "XCONJ" "gbmv_thread_o" false "" "" false ${float_type})
+      GenerateNamedObjects("gbmv_thread.c" "XCONJ;TRANSA" "gbmv_thread_u" false "" "" false ${float_type})
+      GenerateNamedObjects("gbmv_thread.c" "XCONJ;CONJ" "gbmv_thread_s" false "" "" false ${float_type})
+      GenerateNamedObjects("gbmv_thread.c" "XCONJ;CONJ;TRANSA" "gbmv_thread_d" false "" "" false ${float_type})
+
+      GenerateNamedObjects("ger_thread.c" "" "ger_thread_U" false "" "" false ${float_type})
+      GenerateNamedObjects("ger_thread.c" "CONJ" "ger_thread_C" false "" "" false ${float_type})
+      GenerateNamedObjects("ger_thread.c" "XCONJ" "ger_thread_V" false "" "" false ${float_type})
+      GenerateNamedObjects("ger_thread.c" "XCONJ;CONJ" "ger_thread_D" false "" "" false ${float_type})
+
+      GenerateNamedObjects("sbmv_thread.c" "HEMV" "hbmv_thread_U" false "" "" false ${float_type})
+      GenerateNamedObjects("sbmv_thread.c" "HEMV;LOWER" "hbmv_thread_L" false "" "" false ${float_type})
+      GenerateNamedObjects("sbmv_thread.c" "HEMVREV" "hbmv_thread_V" false "" "" false ${float_type})
+      GenerateNamedObjects("sbmv_thread.c" "LOWER;HEMVREV" "hbmv_thread_M" false "" "" false ${float_type})
+
+      GenerateNamedObjects("spmv_thread.c" "HEMV" "hpmv_thread_U" false "" "" false ${float_type})
+      GenerateNamedObjects("spmv_thread.c" "HEMV;LOWER" "hpmv_thread_L" false "" "" false ${float_type})
+      GenerateNamedObjects("spmv_thread.c" "HEMVREV" "hpmv_thread_V" false "" "" false ${float_type})
+      GenerateNamedObjects("spmv_thread.c" "LOWER;HEMVREV" "hpmv_thread_M" false "" "" false ${float_type})
+
+      GenerateNamedObjects("spr_thread.c" "HEMV" "hpr_thread_U" false "" "" false ${float_type})
+      GenerateNamedObjects("spr_thread.c" "HEMV;LOWER" "hpr_thread_L" false "" "" false ${float_type})
+      GenerateNamedObjects("spr_thread.c" "HEMVREV" "hpr_thread_V" false "" "" false ${float_type})
+      GenerateNamedObjects("spr_thread.c" "LOWER;HEMVREV" "hpr_thread_M" false "" "" false ${float_type})
+
+      GenerateNamedObjects("spr2_thread.c" "HEMV" "hpr2_thread_U" false "" "" false ${float_type})
+      GenerateNamedObjects("spr2_thread.c" "HEMV;LOWER" "hpr2_thread_L" false "" "" false ${float_type})
+      GenerateNamedObjects("spr2_thread.c" "HEMVREV" "hpr2_thread_V" false "" "" false ${float_type})
+      GenerateNamedObjects("spr2_thread.c" "LOWER;HEMVREV" "hpr2_thread_M" false "" "" false ${float_type})
+
+      GenerateNamedObjects("symv_thread.c" "HEMV" "hemv_thread_U" false "" "" false ${float_type})
+      GenerateNamedObjects("symv_thread.c" "HEMV;LOWER" "hemv_thread_L" false "" "" false ${float_type})
+      GenerateNamedObjects("symv_thread.c" "HEMVREV" "hemv_thread_V" false "" "" false ${float_type})
+      GenerateNamedObjects("symv_thread.c" "LOWER;HEMVREV" "hemv_thread_M" false "" "" false ${float_type})
+
+      GenerateNamedObjects("syr_thread.c" "HER" "her_thread_U" false "" "" false ${float_type})
+      GenerateNamedObjects("syr_thread.c" "HER;LOWER" "her_thread_L" false "" "" false ${float_type})
+      GenerateNamedObjects("syr_thread.c" "HERREV" "her_thread_V" false "" "" false ${float_type})
+      GenerateNamedObjects("syr_thread.c" "LOWER;HERREV" "her_thread_M" false "" "" false ${float_type})
+
+      GenerateNamedObjects("syr2_thread.c" "HER" "her2_thread_U" false "" "" false ${float_type})
+      GenerateNamedObjects("syr2_thread.c" "HER;LOWER" "her2_thread_L" false "" "" false ${float_type})
+      GenerateNamedObjects("syr2_thread.c" "HERREV" "her2_thread_V" false "" "" false ${float_type})
+      GenerateNamedObjects("syr2_thread.c" "LOWER;HERREV" "her2_thread_M" false "" "" false ${float_type})
+
+      foreach (nu_smp_src ${NU_SMP_SOURCES})
+        string(REGEX MATCH "[a-z]+_[a-z]+" op_name ${nu_smp_src})
+        GenerateCombinationObjects("${nu_smp_src}" "LOWER;UNIT" "U;N" "TRANSA=1" 0 "${op_name}_N" false ${float_type})
+        GenerateCombinationObjects("${nu_smp_src}" "LOWER;UNIT" "U;N" "TRANSA=2" 0 "${op_name}_T" false ${float_type})
+        GenerateCombinationObjects("${nu_smp_src}" "LOWER;UNIT" "U;N" "TRANSA=3" 0 "${op_name}_R" false ${float_type})
+        GenerateCombinationObjects("${nu_smp_src}" "LOWER;UNIT" "U;N" "TRANSA=4" 0 "${op_name}_C" false ${float_type})
+      endforeach ()
+    endif ()
+
+  else ()
+    # For real number functions
+    foreach (u_source ${U_SOURCES})
+      string(REGEX MATCH "[a-z]+" op_name ${u_source})
+      GenerateCombinationObjects("${u_source}" "UNIT" "N" "" 0 "${op_name}_NU" false ${float_type})
+      GenerateCombinationObjects("${u_source}" "UNIT" "N" "TRANSA" 0 "${op_name}_TL" false ${float_type})
+    endforeach ()
+
+    foreach (l_source ${L_SOURCES})
+      string(REGEX MATCH "[a-z]+" op_name ${l_source})
+      GenerateCombinationObjects("${l_source}" "UNIT" "N" "" 0 "${op_name}_NL" false ${float_type})
+      GenerateCombinationObjects("${l_source}" "UNIT" "N" "TRANSA" 0 "${op_name}_TU" false ${float_type})
+    endforeach ()
+
+    if (SMP)
+      GenerateNamedObjects("ger_thread.c" "" "" false "" "" false ${float_type})
+      foreach(nu_smp_source ${NU_SMP_SOURCES})
+        string(REGEX MATCH "[a-z]+_[a-z]+" op_name ${nu_smp_source})
+        GenerateCombinationObjects("${nu_smp_source}" "LOWER;UNIT" "U;N" "" 0 "${op_name}_N" false ${float_type})
+        GenerateCombinationObjects("${nu_smp_source}" "LOWER;UNIT" "U;N" "TRANSA" 0 "${op_name}_T" false ${float_type})
+      endforeach()
+    endif ()
+  endif ()
+endforeach ()
+
+if (SMP)
+  GenerateCombinationObjects("${UL_SMP_SOURCES}" "LOWER" "U" "" 2)
+endif ()
+
+add_library(driver_level2 OBJECT ${OPENBLAS_SRC})

driver/level2/gbmv_thread.c

@@ -64,7 +64,7 @@ static int gbmv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, F
 #ifndef COMPLEX
   FLOAT result;
 #else
-  FLOAT _Complex result;
+  OPENBLAS_COMPLEX_FLOAT result;
 #endif
 #endif
 

driver/level2/sbmv_thread.c

@@ -60,7 +60,7 @@ static int sbmv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, F
 #ifndef COMPLEX
   FLOAT result;
 #else
-  FLOAT _Complex result;
+  OPENBLAS_COMPLEX_FLOAT result;
 #endif
 
   a = (FLOAT *)args -> a;

driver/level2/spmv_thread.c

@@ -55,12 +55,12 @@
 static int spmv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *dummy1, FLOAT *buffer, BLASLONG pos){
 
   FLOAT *a, *x, *y;
-  BLASLONG incx, incy;
+  BLASLONG incx;
   BLASLONG m_from, m_to, i;
 #ifndef COMPLEX
   FLOAT result;
 #else
-  FLOAT _Complex result;
+  OPENBLAS_COMPLEX_FLOAT result;
 #endif
 
   a = (FLOAT *)args -> a;
@@ -68,7 +68,6 @@ static int spmv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, F
   y = (FLOAT *)args -> c;
 
   incx = args -> ldb;
-  incy = args -> ldc;
 
   m_from = 0;
   m_to   = args -> m;

driver/level2/spr2_thread.c

@@ -43,7 +43,7 @@
 static int syr_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *dummy1, FLOAT *buffer, BLASLONG pos){
 
   FLOAT *a, *x, *y;
-  BLASLONG lda, incx, incy;
+  BLASLONG incx, incy;
   BLASLONG i, m_from, m_to;
   FLOAT alpha_r;
 #ifdef COMPLEX
@@ -56,7 +56,6 @@ static int syr_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FL
 
   incx = args -> lda;
   incy = args -> ldb;
-  lda  = args -> ldc;
 
   alpha_r = *((FLOAT *)args -> alpha + 0);
 #ifdef COMPLEX

driver/level2/spr_thread.c

@@ -46,7 +46,7 @@ static int syr_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FL
   BLASLONG incx;
   BLASLONG i, m_from, m_to;
   FLOAT alpha_r;
-#if defined(COMPLEX) && !defined(HER) && !defined(HERREV)
+#if defined(COMPLEX) && !defined(HEMV) && !defined(HEMVREV)
   FLOAT alpha_i;
 #endif
 
@@ -56,7 +56,7 @@ static int syr_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FL
   incx = args -> lda;
 
   alpha_r = *((FLOAT *)args -> alpha + 0);
-#if defined(COMPLEX) && !defined(HER) && !defined(HERREV)
+#if defined(COMPLEX) && !defined(HEMV) && !defined(HEMVREV)
   alpha_i = *((FLOAT *)args -> alpha + 1);
 #endif
 

driver/level2/symv_thread.c

@@ -55,7 +55,7 @@
 static int symv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *dummy1, FLOAT *buffer, BLASLONG pos){
 
   FLOAT *a, *x, *y;
-  BLASLONG lda, incx, incy;
+  BLASLONG lda, incx;
   BLASLONG m_from, m_to;
 
   a = (FLOAT *)args -> a;
@@ -64,7 +64,6 @@ static int symv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, F
 
   lda  = args -> lda;
   incx = args -> ldb;
-  incy = args -> ldc;
 
   m_from = 0;
   m_to   = args -> m;

driver/level2/tbmv_L.c

@@ -45,13 +45,11 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/tbmv_U.c

@@ -45,13 +45,11 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/tbmv_thread.c

@@ -76,7 +76,7 @@ static int trmv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, F
 #ifndef COMPLEX
   FLOAT          result;
 #else
-  FLOAT _Complex result;
+  OPENBLAS_COMPLEX_FLOAT result;
 #endif
 #endif
 

driver/level2/tbsv_L.c

@@ -45,13 +45,11 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/tbsv_U.c

@@ -45,13 +45,11 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/tpmv_thread.c

@@ -81,7 +81,7 @@ static int tpmv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, F
 #ifndef COMPLEX
   FLOAT          result;
 #else
-  FLOAT _Complex result;
+  OPENBLAS_COMPLEX_FLOAT result;
 #endif
 #endif
 

driver/level2/tpsv_L.c

@@ -43,12 +43,10 @@
 int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level2/tpsv_U.c

@@ -43,12 +43,10 @@
 int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level2/trmv_thread.c

@@ -87,7 +87,7 @@ static int trmv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, F
 #ifndef COMPLEX
   FLOAT          result;
 #else
-  FLOAT _Complex result;
+  OPENBLAS_COMPLEX_FLOAT result;
 #endif
 #endif
 

driver/level2/zgbmv_k.c

@@ -77,7 +77,7 @@ void CNAME(BLASLONG m, BLASLONG n, BLASLONG ku, BLASLONG kl, FLOAT alpha_r, FLOA
   FLOAT *bufferY    = gemvbuffer;
   FLOAT *bufferX    = gemvbuffer;
 #ifdef TRANS
-  FLOAT _Complex temp;
+  OPENBLAS_COMPLEX_FLOAT temp;
 #endif
 
   if (incy != 1) {

driver/level2/zhbmv_k.c

@@ -56,6 +56,8 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i,
   FLOAT *bufferX    = sbmvbuffer;
   FLOAT temp[2];
 
+  OPENBLAS_COMPLEX_FLOAT result;
+
   if (incy != 1) {
     Y = bufferY;
     bufferX    = (FLOAT *)(((BLASLONG)bufferY + n * sizeof(FLOAT) * COMPSIZE + 4095) & ~4095);
@@ -93,7 +95,7 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i,
     Y[i * 2 + 1] +=  alpha_r * temp[1] + alpha_i * temp[0];
 
     if (length > 0) {
-      FLOAT _Complex result = DOTC_K(length, a + offset * COMPSIZE, 1, X + (i - length) * COMPSIZE, 1);
+      result = DOTC_K(length, a + offset * COMPSIZE, 1, X + (i - length) * COMPSIZE, 1);
 
       Y[i * 2 + 0] += alpha_r * CREAL(result) - alpha_i * CIMAG(result);
       Y[i * 2 + 1] += alpha_r * CIMAG(result) + alpha_i * CREAL(result);
@@ -118,7 +120,7 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i,
     Y[i * 2 + 1] +=  alpha_r * temp[1] + alpha_i * temp[0];
 
     if (length > 0) {
-      FLOAT _Complex result = DOTC_K(length, a + COMPSIZE, 1, X + (i + 1) * COMPSIZE, 1);
+      result = DOTC_K(length, a + COMPSIZE, 1, X + (i + 1) * COMPSIZE, 1);
 
       Y[i * 2 + 0] += alpha_r * CREAL(result) - alpha_i * CIMAG(result);
       Y[i * 2 + 1] += alpha_r * CIMAG(result) + alpha_i * CREAL(result);
@@ -143,7 +145,7 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i,
     Y[i * 2 + 1] +=  alpha_r * temp[1] + alpha_i * temp[0];
 
     if (length > 0) {
-      FLOAT _Complex result = DOTU_K(length, a + offset * COMPSIZE, 1, X + (i - length) * COMPSIZE, 1);
+      result = DOTU_K(length, a + offset * COMPSIZE, 1, X + (i - length) * COMPSIZE, 1);
 
       Y[i * 2 + 0] +=   alpha_r * CREAL(result) - alpha_i * CIMAG(result);
       Y[i * 2 + 1] +=  alpha_r * CIMAG(result) + alpha_i * CREAL(result);
@@ -168,7 +170,7 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i,
     Y[i * 2 + 1] +=  alpha_r * temp[1] + alpha_i * temp[0];
 
     if (length > 0) {
-      FLOAT _Complex result = DOTU_K(length, a + COMPSIZE, 1, X + (i + 1) * COMPSIZE, 1);
+      result = DOTU_K(length, a + COMPSIZE, 1, X + (i + 1) * COMPSIZE, 1);
 
       Y[i * 2 + 0] += alpha_r * CREAL(result) - alpha_i * CIMAG(result);
       Y[i * 2 + 1] += alpha_r * CIMAG(result) + alpha_i * CREAL(result);

driver/level2/zhpmv_k.c

@@ -51,6 +51,8 @@ int CNAME(BLASLONG m, FLOAT alpha_r, FLOAT alpha_i,
   FLOAT *bufferX    = gemvbuffer;
   FLOAT temp[2];
 
+  OPENBLAS_COMPLEX_FLOAT result;
+
   if (incy != 1) {
     Y = bufferY;
     bufferX    = (FLOAT *)(((BLASLONG)bufferY + m * sizeof(FLOAT) * 2 + 4095) & ~4095);
@@ -69,7 +71,7 @@ int CNAME(BLASLONG m, FLOAT alpha_r, FLOAT alpha_i,
 #ifndef HEMVREV
 #ifndef LOWER
     if (i > 0) {
-      FLOAT _Complex result = DOTC_K(i, a, 1, X, 1);
+      result = DOTC_K(i, a, 1, X, 1);
 
       Y[i * 2 + 0] += alpha_r * CREAL(result) - alpha_i * CIMAG(result);
       Y[i * 2 + 1] += alpha_r * CIMAG(result) + alpha_i * CREAL(result);
@@ -93,7 +95,7 @@ int CNAME(BLASLONG m, FLOAT alpha_r, FLOAT alpha_i,
 #else
 
     if (m - i > 1) {
-      FLOAT _Complex result = DOTC_K(m - i - 1, a + (i + 1) * 2, 1, X + (i + 1) * 2, 1);
+      result = DOTC_K(m - i - 1, a + (i + 1) * 2, 1, X + (i + 1) * 2, 1);
 
       Y[i * 2 + 0] += alpha_r * CREAL(result) - alpha_i * CIMAG(result);
       Y[i * 2 + 1] += alpha_r * CIMAG(result) + alpha_i * CREAL(result);
@@ -118,7 +120,7 @@ int CNAME(BLASLONG m, FLOAT alpha_r, FLOAT alpha_i,
 #else
 #ifndef LOWER
     if (i > 0) {
-      FLOAT _Complex result = DOTU_K(i, a, 1, X, 1);
+      result = DOTU_K(i, a, 1, X, 1);
 
       Y[i * 2 + 0] += alpha_r * CREAL(result) - alpha_i * CIMAG(result);
       Y[i * 2 + 1] += alpha_r * CIMAG(result) + alpha_i * CREAL(result);
@@ -142,7 +144,7 @@ int CNAME(BLASLONG m, FLOAT alpha_r, FLOAT alpha_i,
 #else
 
     if (m - i > 1) {
-      FLOAT _Complex result = DOTU_K(m - i - 1, a + (i + 1) * 2, 1, X + (i + 1) * 2, 1);
+      result = DOTU_K(m - i - 1, a + (i + 1) * 2, 1, X + (i + 1) * 2, 1);
 
       Y[i * 2 + 0] += alpha_r * CREAL(result) - alpha_i * CIMAG(result);
       Y[i * 2 + 1] += alpha_r * CIMAG(result) + alpha_i * CREAL(result);

driver/level2/zsbmv_k.c

@@ -55,6 +55,8 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i,
   FLOAT *bufferY    = sbmvbuffer;
   FLOAT *bufferX    = sbmvbuffer;
 
+  OPENBLAS_COMPLEX_FLOAT result;
+
   if (incy != 1) {
     Y = bufferY;
     bufferX    = (FLOAT *)(((BLASLONG)bufferY + n * sizeof(FLOAT) * COMPSIZE + 4095) & ~4095);
@@ -83,7 +85,7 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i,
 	    a + offset * COMPSIZE, 1, Y + (i - length) * COMPSIZE, 1, NULL, 0);
 
     if (length > 0) {
-      FLOAT _Complex result = DOTU_K(length, a + offset * COMPSIZE, 1, X + (i - length) * COMPSIZE, 1);
+      result = DOTU_K(length, a + offset * COMPSIZE, 1, X + (i - length) * COMPSIZE, 1);
 
       Y[i * 2 + 0] += alpha_r * CREAL(result) - alpha_i * CIMAG(result);
       Y[i * 2 + 1] += alpha_r * CIMAG(result) + alpha_i * CREAL(result);
@@ -100,7 +102,7 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT alpha_r, FLOAT alpha_i,
 	    a, 1, Y + i * COMPSIZE, 1, NULL, 0);
 
     if (length > 0) {
-      FLOAT _Complex result = DOTU_K(length, a + COMPSIZE, 1, X + (i + 1) * COMPSIZE, 1);
+      result = DOTU_K(length, a + COMPSIZE, 1, X + (i + 1) * COMPSIZE, 1);
 
       Y[i * 2 + 0] += alpha_r * CREAL(result) - alpha_i * CIMAG(result);
       Y[i * 2 + 1] += alpha_r * CIMAG(result) + alpha_i * CREAL(result);

driver/level2/zspmv_k.c

@@ -49,7 +49,8 @@ int CNAME(BLASLONG m, FLOAT alpha_r, FLOAT alpha_i,
   FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *bufferY    = gemvbuffer;
   FLOAT *bufferX    = gemvbuffer;
-  FLOAT _Complex result;
+
+  OPENBLAS_COMPLEX_FLOAT result;
 
   if (incy != 1) {
     Y = bufferY;

driver/level2/ztbmv_L.c

@@ -45,11 +45,10 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex temp;
+  OPENBLAS_COMPLEX_FLOAT temp;
 #endif
 #ifndef UNIT
   FLOAT atemp1, atemp2, btemp1, btemp2;
@@ -57,7 +56,6 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG inc
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) * COMPSIZE+ 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/ztbmv_U.c

@@ -45,11 +45,10 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex temp;
+  OPENBLAS_COMPLEX_FLOAT temp;
 #endif
 #ifndef UNIT
   FLOAT atemp1, atemp2, btemp1, btemp2;
@@ -57,7 +56,6 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG inc
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) * COMPSIZE + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/ztbsv_L.c

@@ -45,11 +45,10 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex temp;
+  OPENBLAS_COMPLEX_FLOAT temp;
 #endif
 #ifndef UNIT
   FLOAT ar, ai, br, bi, ratio, den;
@@ -57,7 +56,6 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG inc
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) * COMPSIZE + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/ztbsv_U.c

@@ -45,11 +45,10 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex temp;
+  OPENBLAS_COMPLEX_FLOAT temp;
 #endif
 #ifndef UNIT
   FLOAT ar, ai, br, bi, ratio, den;
@@ -57,7 +56,6 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG inc
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) * COMPSIZE+ 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/ztpmv_L.c

@@ -44,17 +44,15 @@ int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex temp;
+  OPENBLAS_COMPLEX_FLOAT temp;
 #endif
 #ifndef UNIT
   FLOAT atemp1, atemp2, btemp1, btemp2;
 #endif
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) * 2 + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level2/ztpmv_U.c

@@ -44,17 +44,15 @@ int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex temp;
+  OPENBLAS_COMPLEX_FLOAT temp;
 #endif
 #ifndef UNIT
   FLOAT atemp1, atemp2, btemp1, btemp2;
 #endif
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) * 2 + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level2/ztpsv_L.c

@@ -46,17 +46,15 @@ int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex result;
+  OPENBLAS_COMPLEX_FLOAT result;
 #endif
 #ifndef UNIT
   FLOAT ar, ai, br, bi, ratio, den;
 #endif
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) * 2 + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level2/ztpsv_U.c

@@ -44,17 +44,15 @@ int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex result;
+  OPENBLAS_COMPLEX_FLOAT result;
 #endif
 #ifndef UNIT
   FLOAT ar, ai, br, bi, ratio, den;
 #endif
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) * 2 + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level2/ztrmv_L.c

@@ -46,7 +46,7 @@ int CNAME(BLASLONG m, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, FLOAT *bu
 
   BLASLONG i, is, min_i;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex temp;
+  OPENBLAS_COMPLEX_FLOAT temp;
 #endif
 #ifndef UNIT
   FLOAT atemp1, atemp2, btemp1, btemp2;

driver/level2/ztrmv_U.c

@@ -46,7 +46,7 @@ int CNAME(BLASLONG m, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, FLOAT *bu
 
   BLASLONG i, is, min_i;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex temp;
+  OPENBLAS_COMPLEX_FLOAT temp;
 #endif
 #ifndef UNIT
   FLOAT atemp1, atemp2, btemp1, btemp2;

driver/level2/ztrsv_L.c

@@ -46,7 +46,7 @@ int CNAME(BLASLONG m, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buf
 
   BLASLONG i, is, min_i;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex result;
+  OPENBLAS_COMPLEX_FLOAT result;
 #endif
 #ifndef UNIT
   FLOAT ar, ai, br, bi, ratio, den;

driver/level2/ztrsv_U.c

@@ -46,7 +46,7 @@ int CNAME(BLASLONG m, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buf
 
   BLASLONG i, is, min_i;
 #if (TRANSA == 2) || (TRANSA == 4)
-  FLOAT _Complex result;
+  OPENBLAS_COMPLEX_FLOAT result;
 #endif
 #ifndef UNIT
   FLOAT ar, ai, br, bi, ratio, den;

driver/level3/CMakeLists.txt

@@ -0,0 +1,115 @@
+include_directories(${CMAKE_SOURCE_DIR})
+
+# N.B. In the original makefile there was a BLOCKS define used in the compilation of these files but I don't see any evidence of it being set anywhere. -hpa
+
+# loop through gemm.c defines
+set(GEMM_DEFINES NN NT TN TT)
+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_${GEMM_DEFINE_LC}" 0)
+  if (SMP AND NOT USE_SIMPLE_THREADED_LEVEL3)
+    GenerateNamedObjects("gemm.c" "${GEMM_DEFINE};THREADED_LEVEL3" "gemm_thread_${GEMM_DEFINE_LC}" 0)
+  endif ()
+endforeach ()
+
+
+set(TRMM_TRSM_SOURCES
+  trmm_L.c
+  trmm_R.c
+  trsm_L.c
+  trsm_R.c)
+
+foreach(trmm_trsm_source ${TRMM_TRSM_SOURCES})
+  string(REGEX MATCH "[a-z]+_[A-Z]+" op_name ${trmm_trsm_source})
+  GenerateCombinationObjects("${trmm_trsm_source}" "UPPER;UNIT" "L;N" "" 0 "${op_name}N")
+  GenerateCombinationObjects("${trmm_trsm_source}" "UPPER;UNIT" "L;N" "TRANSA" 0 "${op_name}T")
+endforeach()
+
+GenerateCombinationObjects("symm_k.c" "RSIDE;LOWER" "L;U" "NN" 1)
+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)
+
+  # 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)
+
+  if (NOT USE_SIMPLE_THREADED_LEVEL3)
+    GenerateCombinationObjects("syrk_k.c" "LOWER;TRANS" "U;N" "THREADED_LEVEL3" 2 "syrk_thread")
+    GenerateCombinationObjects("symm_k.c" "RSIDE;LOWER" "L;U" "THREADED_LEVEL3;NN" 2 "symm_thread")
+  endif ()
+endif ()
+
+foreach (float_type ${FLOAT_TYPES})
+  if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
+    GenerateCombinationObjects("zherk_kernel.c" "LOWER;CONJ" "U;N" "HERK" 2 "herk_kernel" false ${float_type})
+    # TRANS needs to be set/unset when CONJ is set/unset, so can't use it as a combination
+    GenerateCombinationObjects("zherk_k.c" "LOWER" "U" "HERK" 3 "herk_N" false ${float_type})
+    GenerateCombinationObjects("zherk_k.c" "LOWER" "U" "HERK;TRANS;CONJ" 3 "herk_C" false ${float_type})
+    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})
+    # Need to set CONJ for trmm and trsm
+    GenerateCombinationObjects("trmm_L.c" "UPPER;UNIT" "L;N" "CONJ" 0 "trmm_LR" false ${float_type})
+    GenerateCombinationObjects("trmm_L.c" "UPPER;UNIT" "L;N" "TRANSA;CONJ" 0 "trmm_LC" false ${float_type})
+    GenerateCombinationObjects("trmm_R.c" "UPPER;UNIT" "L;N" "CONJ" 0 "trmm_RR" false ${float_type})
+    GenerateCombinationObjects("trmm_R.c" "UPPER;UNIT" "L;N" "TRANSA;CONJ" 0 "trmm_RC" false ${float_type})
+    GenerateCombinationObjects("trsm_L.c" "UPPER;UNIT" "L;N" "CONJ" 0 "trsm_LR" false ${float_type})
+    GenerateCombinationObjects("trsm_L.c" "UPPER;UNIT" "L;N" "TRANSA;CONJ" 0 "trsm_LC" false ${float_type})
+    GenerateCombinationObjects("trsm_R.c" "UPPER;UNIT" "L;N" "CONJ" 0 "trsm_RR" false ${float_type})
+    GenerateCombinationObjects("trsm_R.c" "UPPER;UNIT" "L;N" "TRANSA;CONJ" 0 "trsm_RC" false ${float_type})
+
+    #hemm
+    GenerateCombinationObjects("zhemm_k.c" "LOWER" "U" "NN" 0 "hemm_L" false ${float_type})
+    GenerateCombinationObjects("zhemm_k.c" "LOWER" "U" "NC;RSIDE" 0 "hemm_R" false ${float_type})
+
+    #her2k
+    GenerateCombinationObjects("zher2k_kernel.c" "LOWER;CONJ" "U;N" "" 2 "her2k_kernel" false ${float_type})
+    GenerateNamedObjects("zher2k_k.c" "HER2K" "her2k_UN" false "" "" false ${float_type})
+    GenerateNamedObjects("zher2k_k.c" "HER2K;TRANS;CONJ" "her2k_UC" false "" "" false ${float_type})
+    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)
+      #hemm
+      GenerateCombinationObjects("zhemm_k.c" "LOWER" "U" "NN;THREADED_LEVEL3" 0 "hemm_thread_L" false ${float_type})
+      GenerateCombinationObjects("zhemm_k.c" "LOWER" "U" "NC;RSIDE;THREADED_LEVEL3" 0 "hemm_thread_R" false ${float_type})
+      #her2k
+      GenerateNamedObjects("zher2k_k.c" "HER2K" "her2k_UN" false "" "" false ${float_type})
+      GenerateNamedObjects("zher2k_k.c" "HER2K;TRANS;CONJ" "her2k_UC" false "" "" false ${float_type})
+      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})
+    endif()
+
+    # special gemm defines for complex
+    foreach (gemm_define ${GEMM_COMPLEX_DEFINES})
+      string(TOLOWER ${gemm_define} gemm_define_LC)
+      GenerateNamedObjects("gemm.c" "${gemm_define}" "gemm_${gemm_define_LC}" false "" "" false ${float_type})
+      if(USE_GEMM3M)
+	GenerateNamedObjects("gemm3m.c" "${gemm_define}" "gemm3m_${gemm_define_LC}" false "" "" false ${float_type})
+      endif()
+      if (SMP AND NOT USE_SIMPLE_THREADED_LEVEL3)
+        GenerateNamedObjects("gemm.c" "${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})
+	endif()
+      endif ()
+    endforeach ()
+  endif ()
+endforeach ()
+
+#HPLOBJS =
+#	dgemm_nn.c dgemm_nt.c dgemm_tn.c dgemm_tt.c
+#	dtrsm_LNUU.c dtrsm_LNUN.c dtrsm_LNLU.c dtrsm_LNLN.c
+#	dtrsm_LTUU.c dtrsm_LTUN.c dtrsm_LTLU.c dtrsm_LTLN.c
+#	dtrsm_RNUU.c dtrsm_RNUN.c dtrsm_RNLU.c dtrsm_RNLN.c
+#	dtrsm_RTUU.c dtrsm_RTUN.c dtrsm_RTLU.c dtrsm_RTLN.c
+#
+#if (USE_SIMPLE_THREADED_LEVEL3)
+#  HPLOBJS += dgemm_thread_nn.c dgemm_thread_nt.c
+#	   dgemm_thread_tn.c dgemm_thread_tt.c
+#endif
+#
+
+add_library(driver_level3 OBJECT ${OPENBLAS_SRC})

driver/level3/gemm_thread_mn.c

@@ -65,7 +65,7 @@ int CNAME(int mode, blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n, int (
   blas_queue_t queue[MAX_CPU_NUMBER];
 
   BLASLONG range_M[MAX_CPU_NUMBER + 1], range_N[MAX_CPU_NUMBER + 1];
-  BLASLONG procs, total_procs, num_cpu_m, num_cpu_n;
+  BLASLONG procs, num_cpu_m, num_cpu_n;
 
   BLASLONG width, i, j;
   BLASLONG divM, divN;

driver/level3/level3_thread.c

@@ -230,7 +230,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
   BLASLONG is, min_i, div_n;
 
   BLASLONG i, current;
-  BLASLONG l1stride, l2size;
+  BLASLONG l1stride;
 
 #ifdef TIMING
   BLASULONG rpcc_counter;
@@ -298,8 +298,6 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 #endif
       ) return 0;
 
-  l2size = GEMM_P * GEMM_Q;
-
 #if 0
   fprintf(stderr, "Thread[%ld]  m_from : %ld m_to : %ld n_from : %ld n_to : %ld N_from : %ld N_to : %ld\n",
 	  mypos, m_from, m_to, n_from, n_to, N_from, N_to);
@@ -706,7 +704,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
     n = n_to - n_from;
   }
 
-  if ((args -> m < nthreads * SWITCH_RATIO) || (args -> n < nthreads * SWITCH_RATIO)) {
+  if ((m < nthreads * SWITCH_RATIO) || (n < nthreads * SWITCH_RATIO)) {
     GEMM_LOCAL(args, range_m, range_n, sa, sb, 0);
     return 0;
   }

driver/level3/syr2k_k.c

@@ -47,7 +47,7 @@
 #endif
 #endif
 
-static inline int syrk_beta(BLASLONG m_from, BLASLONG m_to, BLASLONG n_from, BLASLONG n_to, FLOAT *alpha, FLOAT *c, BLASLONG ldc) {
+static __inline int syrk_beta(BLASLONG m_from, BLASLONG m_to, BLASLONG n_from, BLASLONG n_to, FLOAT *alpha, FLOAT *c, BLASLONG ldc) {
 
   BLASLONG i;
 

driver/level3/syrk_k.c

@@ -49,7 +49,7 @@
 #endif
 #endif
 
-static inline int syrk_beta(BLASLONG m_from, BLASLONG m_to, BLASLONG n_from, BLASLONG n_to, FLOAT *alpha, FLOAT *c, BLASLONG ldc) {
+static __inline int syrk_beta(BLASLONG m_from, BLASLONG m_to, BLASLONG n_from, BLASLONG n_to, FLOAT *alpha, FLOAT *c, BLASLONG ldc) {
 
   BLASLONG i;
 

driver/level3/trmm_R.c

@@ -70,6 +70,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
   BLASLONG ls, is, js;
   BLASLONG min_l, min_i, min_j;
   BLASLONG jjs, min_jj;
+#if !((!defined(UPPER) && !defined(TRANSA)) || (defined(UPPER) && defined(TRANSA)))
+  BLASLONG start_ls;
+#endif
 
   m = args -> m;
   n = args -> n;
@@ -226,8 +229,6 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
   }
 
 #else
-  BLASLONG start_ls;
-
   for(js = n; js > 0; js -= GEMM_R){
     min_j = js;
     if (min_j > GEMM_R) min_j = GEMM_R;

driver/level3/trsm_L.c

@@ -76,6 +76,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
   BLASLONG ls, is, js;
   BLASLONG min_l, min_i, min_j;
   BLASLONG jjs, min_jj;
+#if !((!defined(UPPER) && !defined(TRANSA)) || (defined(UPPER) && defined(TRANSA)))
+  BLASLONG start_is;
+#endif
 
   m = args -> m;
   n = args -> n;
@@ -178,8 +181,6 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
       }
     }
 #else
-    BLASLONG start_is;
-
     for(ls = m; ls > 0; ls -= GEMM_Q){
       min_l = ls;
       if (min_l > GEMM_Q) min_l = GEMM_Q;

driver/level3/trsm_R.c

@@ -75,6 +75,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
   BLASLONG ls, is, js;
   BLASLONG min_l, min_i, min_j;
   BLASLONG jjs, min_jj;
+#if !((defined(UPPER) && !defined(TRANSA)) || (!defined(UPPER) && defined(TRANSA)))
+  BLASLONG start_ls;
+#endif
 
   m = args -> m;
   n = args -> n;
@@ -226,8 +229,6 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
   }
 
 #else
-  BLASLONG start_ls;
-
   for(js = n; js > 0; js -= GEMM_R){
     min_j = js;
     if (min_j > GEMM_R) min_j = GEMM_R;

driver/others/CMakeLists.txt

@@ -0,0 +1,75 @@
+include_directories(${CMAKE_SOURCE_DIR})
+
+if (${CORE} STREQUAL "PPC440")
+  set(MEMORY memory_qalloc.c)
+else ()
+  set(MEMORY memory.c)
+endif ()
+
+if (SMP)
+
+  if (USE_OPENMP)
+    set(BLAS_SERVER blas_server_omp.c)
+  elseif (${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
+    set(BLAS_SERVER blas_server_win32.c)
+  endif ()
+
+  if (NOT DEFINED BLAS_SERVER)
+    set(BLAS_SERVER blas_server.c)
+  endif ()
+
+  set(SMP_SOURCES
+    ${BLAS_SERVER}
+    divtable.c # TODO: Makefile has -UDOUBLE
+    blas_l1_thread.c
+  )
+
+  if (NOT NO_AFFINITY)
+    list(APPEND SMP_SOURCES init.c)
+  endif ()
+endif ()
+
+set(COMMON_SOURCES
+  xerbla.c
+  openblas_set_num_threads.c
+  openblas_error_handle.c
+  openblas_get_num_procs.c
+  openblas_get_num_threads.c
+)
+
+# these need to have NAME/CNAME set, so use GenerateNamedObjects, but don't use standard name mangling
+GenerateNamedObjects("abs.c" "" "c_abs" 0 "" "" 1 )
+GenerateNamedObjects("abs.c" "DOUBLE" "z_abs" 0 "" "" 1)
+GenerateNamedObjects("openblas_get_config.c;openblas_get_parallel.c" "" "" 0 "" "" 1)
+
+if (DYNAMIC_ARCH)
+  list(APPEND COMMON_SOURCES dynamic.c)
+else ()
+  list(APPEND COMMON_SOURCES parameter.c)
+endif ()
+
+#ifdef EXPRECISION
+#COMMONOBJS	+= x_abs.$(SUFFIX) qlamch.$(SUFFIX) qlamc3.$(SUFFIX)
+#endif
+#
+#ifdef QUAD_PRECISION
+#COMMONOBJS	+= addx.$(SUFFIX) mulx.$(SUFFIX)
+#endif
+#
+#ifdef USE_CUDA
+#COMMONOBJS	+= cuda_init.$(SUFFIX)
+#endif
+#
+#ifdef FUNCTION_PROFILE
+#COMMONOBJS	+= profile.$(SUFFIX)
+#endif
+
+#LIBOTHERS = libothers.$(LIBSUFFIX)
+
+#ifeq ($(DYNAMIC_ARCH), 1)
+#HPLOBJS = memory.$(SUFFIX) xerbla.$(SUFFIX) dynamic.$(SUFFIX)
+#else
+#HPLOBJS = memory.$(SUFFIX) xerbla.$(SUFFIX) parameter.$(SUFFIX)
+#endif
+
+add_library(driver_others OBJECT ${OPENBLAS_SRC} ${MEMORY} ${SMP_SOURCES} ${COMMON_SOURCES})

driver/others/blas_server.c

@@ -70,9 +70,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /*********************************************************************/
 
 #include "common.h"
-#ifdef OS_LINUX
+#if defined(OS_LINUX) || defined(OS_NETBSD) || defined(OS_DARWIN) || defined(OS_ANDROID)
 #include <dlfcn.h>
+#include <signal.h>
 #include <sys/resource.h>
+#include <sys/time.h>
 #endif
 
 #ifndef likely
@@ -265,7 +267,7 @@ int get_node(void);
 
 static int increased_threads = 0;
 
-static int blas_thread_server(void *arg){
+static void* blas_thread_server(void *arg){
 
   /* Thread identifier */
   BLASLONG  cpu = (BLASLONG)arg;
@@ -425,6 +427,10 @@ static int blas_thread_server(void *arg){
       main_status[cpu] = MAIN_FINISH;
 #endif
 
+      // arm: make sure all results are written out _before_
+      // thread is marked as done and other threads use them
+      WMB;
+
       thread_status[cpu].queue = (blas_queue_t * volatile) ((long)thread_status[cpu].queue & 0);  /* Need a trick */
       WMB;
 
@@ -454,7 +460,7 @@ static int blas_thread_server(void *arg){
 
   //pthread_exit(NULL);
 
-  return 0;
+  return NULL;
 }
 
 #ifdef MONITOR
@@ -561,14 +567,23 @@ int blas_thread_init(void){
 
 #ifdef NEED_STACKATTR
       ret=pthread_create(&blas_threads[i], &attr,
-		     (void *)&blas_thread_server, (void *)i);
+		     &blas_thread_server, (void *)i);
 #else
       ret=pthread_create(&blas_threads[i], NULL,
-		     (void *)&blas_thread_server, (void *)i);
+		     &blas_thread_server, (void *)i);
 #endif
       if(ret!=0){
-	fprintf(STDERR,"OpenBLAS: pthread_creat error in blas_thread_init function. Error code:%d\n",ret);
-	exit(1);
+	struct rlimit rlim;
+        const char *msg = strerror(ret);
+        fprintf(STDERR, "OpenBLAS blas_thread_init: pthread_create: %s\n", msg);
+        if(0 == getrlimit(RLIMIT_NPROC, &rlim)) {
+          fprintf(STDERR, "OpenBLAS blas_thread_init: RLIMIT_NPROC "
+                  "%ld current, %ld max\n", (long)(rlim.rlim_cur), (long)(rlim.rlim_max));
+        }
+        if(0 != raise(SIGINT)) {
+          fprintf(STDERR, "OpenBLAS blas_thread_init: calling exit(3)\n");
+          exit(EXIT_FAILURE);
+        }
       }
     }
 
@@ -775,7 +790,12 @@ int exec_blas(BLASLONG num, blas_queue_t *queue){
   stop = rpcc();
 #endif
 
-  if ((num > 1) && queue -> next) exec_blas_async_wait(num - 1, queue -> next);
+  if ((num > 1) && queue -> next) {
+    exec_blas_async_wait(num - 1, queue -> next);
+
+    // arm: make sure results from other threads are visible
+    MB;
+  }
 
 #ifdef TIMING_DEBUG
   fprintf(STDERR, "Thread[0] : %16lu %16lu (%8lu cycles)\n",
@@ -823,10 +843,10 @@ void goto_set_num_threads(int num_threads) {
 
 #ifdef NEED_STACKATTR
       pthread_create(&blas_threads[i], &attr,
-		     (void *)&blas_thread_server, (void *)i);
+		     &blas_thread_server, (void *)i);
 #else
       pthread_create(&blas_threads[i], NULL,
-		     (void *)&blas_thread_server, (void *)i);
+		     &blas_thread_server, (void *)i);
 #endif
     }
 

driver/others/dynamic.c

@@ -252,6 +252,15 @@ static gotoblas_t *get_coretype(void){
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
 	}
+	//Intel Skylake
+	if (model == 14) {
+	  if(support_avx())
+	    return &gotoblas_HASWELL;
+	  else{
+	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
+	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
+	  }
+	}
 	return NULL;
       case 5:
 	//Intel Broadwell
@@ -263,6 +272,15 @@ static gotoblas_t *get_coretype(void){
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
 	}
+	//Intel Skylake
+	if (model == 14 || model == 5) {
+	  if(support_avx())
+	    return &gotoblas_HASWELL;
+	  else{
+	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
+	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
+	  }
+	}
 	return NULL;
       }
       case 0xf:
@@ -300,7 +318,7 @@ static gotoblas_t *get_coretype(void){
 	    openblas_warning(FALLBACK_VERBOSE, BARCELONA_FALLBACK);
 	    return &gotoblas_BARCELONA; //OS doesn't support AVX. Use old kernels.
 	  }
-	}else if(model == 2){
+	}else if(model == 2 || model == 3){
 	  //AMD Bulldozer Opteron 6300 / Opteron 4300 / Opteron 3300
 	  if(support_avx())
 	    return &gotoblas_PILEDRIVER;
@@ -309,7 +327,15 @@ static gotoblas_t *get_coretype(void){
 	    return &gotoblas_BARCELONA; //OS doesn't support AVX. Use old kernels.
 	  }
 	}else if(model == 0){
-	  if (exmodel == 3) {
+	  if (exmodel == 1) {
+	    //AMD Trinity
+	    if(support_avx())
+	      return &gotoblas_PILEDRIVER;
+	    else{
+	      openblas_warning(FALLBACK_VERBOSE, BARCELONA_FALLBACK);
+	      return &gotoblas_BARCELONA; //OS doesn't support AVX. Use old kernels.
+	    }
+	   }else if (exmodel == 3) {
 	    //AMD STEAMROLLER
 	    if(support_avx())
 	      return &gotoblas_STEAMROLLER;

driver/others/memory.c

@@ -104,6 +104,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include <errno.h>
 #include <linux/unistd.h>
 #include <sys/syscall.h>
+#include <sys/time.h>
+#include <sys/resource.h>
 #endif
 
 #if defined(OS_FREEBSD) || defined(OS_DARWIN)
@@ -139,8 +141,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define BITMASK(a, b, c) ((((a) >> (b)) & (c)))
 
+#if defined(_MSC_VER) && !defined(__clang__)
+#define CONSTRUCTOR __cdecl
+#define DESTRUCTOR __cdecl
+#elif defined(OS_DARWIN) && defined(C_GCC)
 #define CONSTRUCTOR	__attribute__ ((constructor))
 #define DESTRUCTOR	__attribute__ ((destructor))
+#else
+#define CONSTRUCTOR	__attribute__ ((constructor(101)))
+#define DESTRUCTOR	__attribute__ ((destructor(101)))
+#endif
 
 #ifdef DYNAMIC_ARCH
 gotoblas_t *gotoblas = NULL;
@@ -795,12 +805,12 @@ static void *alloc_hugetlb(void *address){
   
   if (LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME, &tp.Privileges[0].Luid) != TRUE) {
       CloseHandle(hToken);
-      return -1;
+      return (void*)-1;
   }
 
   if (AdjustTokenPrivileges(hToken, FALSE, &tp, 0, NULL, NULL) != TRUE) {
       CloseHandle(hToken);
-      return -1;
+      return (void*)-1;
   }
 
   map_address  = (void *)VirtualAlloc(address,
@@ -906,7 +916,6 @@ static volatile struct {
 } memory[NUM_BUFFERS];
 
 static int memory_initialized = 0;
-static void gotoblas_memory_init(void);
 
 /*       Memory allocation routine           */
 /* procpos ... indicates where it comes from */
@@ -1153,6 +1162,9 @@ void blas_memory_free(void *free_area){
   printf("  Position : %d\n", position);
 #endif
 
+  // arm: ensure all writes are finished before other thread takes this memory
+  WMB;
+
   memory[position].used = 0;
 
 #ifdef DEBUG
@@ -1351,6 +1363,18 @@ void CONSTRUCTOR gotoblas_init(void) {
    gotoblas_memory_init();
 #endif
 
+#if defined(OS_LINUX)
+   struct rlimit curlimit;
+   if ( getrlimit(RLIMIT_STACK, &curlimit ) == 0 )
+   {
+	if ( curlimit.rlim_cur != curlimit.rlim_max )
+	{
+		curlimit.rlim_cur = curlimit.rlim_max;
+		setrlimit(RLIMIT_STACK, &curlimit);
+	}
+   }
+#endif
+
 #ifdef SMP
   if (blas_cpu_number == 0) blas_get_cpu_number();
 #ifdef SMP_SERVER
@@ -1399,6 +1423,28 @@ void DESTRUCTOR gotoblas_quit(void) {
 #endif
 }
 
+#if defined(_MSC_VER) && !defined(__clang__)
+BOOL APIENTRY DllMain(HMODULE hModule, DWORD  ul_reason_for_call, LPVOID lpReserved)
+{
+  switch (ul_reason_for_call)
+  {
+    case DLL_PROCESS_ATTACH:
+      gotoblas_init();
+      break;
+    case DLL_THREAD_ATTACH:
+      break;
+    case DLL_THREAD_DETACH:
+      break;
+    case DLL_PROCESS_DETACH:
+      gotoblas_quit();
+      break;
+    default:
+      break;
+  }
+  return TRUE;
+}
+#endif
+
 #if (defined(C_PGI) || (!defined(C_SUN) && defined(F_INTERFACE_SUN))) && (defined(ARCH_X86) || defined(ARCH_X86_64))
 /* Don't call me; this is just work around for PGI / Sun bug */
 void gotoblas_dummy_for_PGI(void) {

exports/Makefile

@@ -28,8 +28,10 @@ endif
 
 ifeq ($(OSNAME), WINNT)
 ifeq ($(F_COMPILER), GFORTRAN)
+ifndef ONLY_CBLAS
 EXTRALIB += -lgfortran
 endif
+endif
 ifeq ($(USE_OPENMP), 1)
 ifeq ($(C_COMPILER), GCC)
 EXTRALIB += -lgomp
@@ -39,9 +41,11 @@ endif
 
 ifeq ($(OSNAME), CYGWIN_NT)
 ifeq ($(F_COMPILER), GFORTRAN)
+ifndef ONLY_CBLAS
 EXTRALIB += -lgfortran
 endif
 endif
+endif
 
 all::