Merge pull request #1656 from xianyi/develop

Update the 0.3 branch from develop

CMakeLists.txt

@@ -6,12 +6,15 @@ cmake_minimum_required(VERSION 2.8.5)
 project(OpenBLAS C ASM)
 set(OpenBLAS_MAJOR_VERSION 0)
 set(OpenBLAS_MINOR_VERSION 3)
-set(OpenBLAS_PATCH_VERSION 0.dev)
+set(OpenBLAS_PATCH_VERSION 1.dev)
 set(OpenBLAS_VERSION "${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.${OpenBLAS_PATCH_VERSION}")
 
 # Adhere to GNU filesystem layout conventions
 include(GNUInstallDirs)
 
+include(CMakePackageConfigHelpers)
+
+
 set(OpenBLAS_LIBNAME openblas)
 
 #######
@@ -20,6 +23,7 @@ option(BUILD_WITHOUT_LAPACK "Without LAPACK and LAPACKE (Only BLAS or CBLAS)" ON
 endif()
 option(BUILD_WITHOUT_CBLAS "Without CBLAS" OFF)
 option(DYNAMIC_ARCH "Build with DYNAMIC_ARCH" OFF)
+option(DYNAMIC_OLDER "Support older cpus with DYNAMIC_ARCH" OFF)
 option(BUILD_RELAPACK "Build with ReLAPACK (recursive LAPACK" OFF)
 #######
 if(BUILD_WITHOUT_LAPACK)
@@ -208,6 +212,7 @@ set_target_properties(${OpenBLAS_LIBNAME} PROPERTIES
 
 # Install libraries
 install(TARGETS ${OpenBLAS_LIBNAME}
+	EXPORT "OpenBLASTargets"
 	RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
 	ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
   LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR} )
@@ -267,3 +272,21 @@ if(PKG_CONFIG_FOUND)
 	configure_file(${PROJECT_SOURCE_DIR}/cmake/openblas.pc.in ${PROJECT_BINARY_DIR}/openblas.pc @ONLY)
 	install (FILES ${PROJECT_BINARY_DIR}/openblas.pc DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig/)
 endif()
+
+
+# GNUInstallDirs "DATADIR" wrong here; CMake search path wants "share".
+set(PN OpenBLAS)
+set(CMAKECONFIG_INSTALL_DIR "share/cmake/${PN}")
+configure_package_config_file(cmake/${PN}Config.cmake.in
+                              "${CMAKE_CURRENT_BINARY_DIR}/${PN}Config.cmake"
+                              INSTALL_DESTINATION ${CMAKECONFIG_INSTALL_DIR})
+write_basic_package_version_file(${CMAKE_CURRENT_BINARY_DIR}/${PN}ConfigVersion.cmake
+                                 VERSION ${${PN}_VERSION}
+                                 COMPATIBILITY AnyNewerVersion)
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${PN}Config.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/${PN}ConfigVersion.cmake
+        DESTINATION ${CMAKECONFIG_INSTALL_DIR})
+install(EXPORT "${PN}Targets"
+        NAMESPACE "${PN}::"
+        DESTINATION ${CMAKECONFIG_INSTALL_DIR})
+

Makefile

@@ -153,6 +153,9 @@ ifeq ($(DYNAMIC_ARCH), 1)
 	do  $(MAKE) GOTOBLAS_MAKEFILE= -C kernel TARGET_CORE=$$d kernel || exit 1 ;\
 	done
 	@echo DYNAMIC_ARCH=1 >> Makefile.conf_last
+ifeq ($(DYNAMIC_OLDER), 1)
+	@echo DYNAMIC_OLDER=1 >> Makefile.conf_last
+endif	
 endif
 ifdef USE_THREAD
 	@echo USE_THREAD=$(USE_THREAD) >>  Makefile.conf_last
@@ -294,9 +297,10 @@ 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
+	$(MAKE) -j 1 -C $(NETLIB_LAPACK_DIR)/TESTING/EIG xeigtstc  xeigtstd  xeigtsts  xeigtstz 
+	$(MAKE) -j 1 -C $(NETLIB_LAPACK_DIR)/TESTING/LIN xlintstc  xlintstd  xlintstds  xlintstrfd  xlintstrfz  xlintsts  xlintstz  xlintstzc xlintstrfs xlintstrfc
 ifneq ($(CROSS), 1)
-	( cd $(NETLIB_LAPACK_DIR)/INSTALL; ./testlsame; ./testslamch; ./testdlamch; \
+	( cd $(NETLIB_LAPACK_DIR)/INSTALL; make all; ./testlsame; ./testslamch; ./testdlamch; \
         ./testsecond; ./testdsecnd; ./testieee; ./testversion )
 	(cd $(NETLIB_LAPACK_DIR); ./lapack_testing.py -r )
 endif
@@ -308,9 +312,9 @@ lapack-runtest:
 
 
 blas-test:
-	(cd $(NETLIB_LAPACK_DIR)/BLAS && rm -f x* *.out)
+	(cd $(NETLIB_LAPACK_DIR)/BLAS/TESTING && rm -f x* *.out)
 	$(MAKE) -j 1 -C $(NETLIB_LAPACK_DIR) blas_testing
-	(cd $(NETLIB_LAPACK_DIR)/BLAS && cat *.out)
+	(cd $(NETLIB_LAPACK_DIR)/BLAS/TESTING && cat *.out)
 
 
 dummy :

Makefile.install

@@ -98,7 +98,7 @@ endif
 	@echo Generating openblas.pc in "$(DESTDIR)$(OPENBLAS_PKGCONFIG_DIR)"
 	@echo 'libdir='$(OPENBLAS_LIBRARY_DIR) > "$(DESTDIR)$(OPENBLAS_PKGCONFIG_DIR)/openblas.pc"
 	@echo 'includedir='$(OPENBLAS_INCLUDE_DIR) >> "$(DESTDIR)$(OPENBLAS_PKGCONFIG_DIR)/openblas.pc"
-	@echo 'openblas_config= USE_64BITINT='$(USE_64BITINT) 'DYNAMIC_ARCH='$(DYNAMIC_ARCH) 'NO_CBLAS='$(NO_CBLAS) 'NO_LAPACK='$(NO_LAPACK) 'NO_LAPACKE='$(NO_LAPACKE) 'NO_AFFINITY='$(NO_AFFINITY) 'USE_OPENMP='$(USE_OPENMP) $(CORE) 'MAX_THREADS='$(NUM_THREADS)>> "$(DESTDIR)$(OPENBLAS_PKGCONFIG_DIR)/openblas.pc"
+	@echo 'openblas_config= USE_64BITINT='$(USE_64BITINT) 'DYNAMIC_ARCH='$(DYNAMIC_ARCH) 'DYNAMIC_OLDER='$(DYNAMIC_OLDER) 'NO_CBLAS='$(NO_CBLAS) 'NO_LAPACK='$(NO_LAPACK) 'NO_LAPACKE='$(NO_LAPACKE) 'NO_AFFINITY='$(NO_AFFINITY) 'USE_OPENMP='$(USE_OPENMP) $(CORE) 'MAX_THREADS='$(NUM_THREADS)>> "$(DESTDIR)$(OPENBLAS_PKGCONFIG_DIR)/openblas.pc"
 	@echo 'version='$(VERSION) >> "$(DESTDIR)$(OPENBLAS_PKGCONFIG_DIR)/openblas.pc"
 	@echo 'extralib='$(EXTRALIB) >> "$(DESTDIR)$(OPENBLAS_PKGCONFIG_DIR)/openblas.pc"
 	@cat openblas.pc.in >> "$(DESTDIR)$(OPENBLAS_PKGCONFIG_DIR)/openblas.pc"

Makefile.rule

@@ -3,7 +3,7 @@
 #
 
 # This library's version
-VERSION = 0.3.0.dev
+VERSION = 0.3.1.dev
 
 # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a
 # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library
@@ -17,6 +17,11 @@ VERSION = 0.3.0.dev
 # If you want to support multiple architecture in one binary
 # DYNAMIC_ARCH = 1
 
+# If you want the full list of x86_64 architectures supported in DYNAMIC_ARCH
+# mode (including individual optimizied codes for PENRYN, DUNNINGTON, OPTERON,
+# OPTERON_SSE3, ATOM and NANO rather than fallbacks to older architectures)
+# DYNAMIC_OLDER = 1
+
 # C compiler including binary type(32bit / 64bit). Default is gcc.
 # Don't use Intel Compiler or PGI, it won't generate right codes as I expect.
 # CC = gcc
@@ -55,6 +60,14 @@ VERSION = 0.3.0.dev
 # This flag is always set for POWER8. Don't modify the flag 
 # USE_OPENMP = 1
 
+# The OpenMP scheduler to use - by default this is "static" and you
+# will normally not want to change this unless you know that your main
+# workload will involve tasks that have highly unbalanced running times
+# for individual threads. Changing away from "static" may also adversely
+# affect memory access locality in NUMA systems. Setting to "runtime" will
+# allow you to select the scheduler from the environment variable OMP_SCHEDULE
+# CCOMMON_OPT += -DOMP_SCHED=dynamic
+
 # You can define maximum number of threads. Basically it should be
 # less than actual number of cores. If you don't specify one, it's
 # automatically detected by the the script.
@@ -151,8 +164,11 @@ NO_AFFINITY = 1
 # CONSISTENT_FPCSR = 1
 
 # If any gemm arguement m, n or k is less or equal this threshold, gemm will be execute
-# with single thread. You can use this flag to avoid the overhead of multi-threading
+# with single thread. (Actually in recent versions this is a factor proportional to the
-# in small matrix sizes. The default value is 4.
+# number of floating point operations necessary for the given problem size, no longer
+# an individual dimension). You can use this setting to avoid the overhead of multi-
+# threading in small matrix sizes. The default value is 4, but values as high as 50 have 
+# been reported to be optimal for certain workloads (50 is the recommended value for Julia).
 # GEMM_MULTITHREAD_THRESHOLD = 4
 
 # If you need santy check by comparing reference BLAS. It'll be very

Makefile.system

@@ -62,6 +62,9 @@ ifeq ($(BINARY), 32)
 ifeq ($(TARGET), HASWELL)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
+ifeq ($(TARGET), SKYLAKEX)
+GETARCH_FLAGS := -DFORCE_NEHALEM
+endif
 ifeq ($(TARGET), SANDYBRIDGE)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
@@ -95,6 +98,9 @@ ifeq ($(BINARY), 32)
 ifeq ($(TARGET_CORE), HASWELL)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
+ifeq ($(TARGET_CORE), SKYLAKEX)
+GETARCH_FLAGS := -DFORCE_NEHALEM
+endif
 ifeq ($(TARGET_CORE), SANDYBRIDGE)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
@@ -141,6 +147,10 @@ ifeq ($(NO_AVX2), 1)
 GETARCH_FLAGS	+= -DNO_AVX2
 endif
 
+ifeq ($(NO_AVX512), 1)
+GETARCH_FLAGS	+= -DNO_AVX512
+endif
+
 ifeq ($(DEBUG), 1)
 GETARCH_FLAGS	+= -g
 endif
@@ -238,7 +248,7 @@ endif
 
 ifeq ($(OSNAME), Darwin)
 ifndef MACOSX_DEPLOYMENT_TARGET
-export MACOSX_DEPLOYMENT_TARGET=10.6
+export MACOSX_DEPLOYMENT_TARGET=10.8
 endif
 MD5SUM = md5 -r
 endif
@@ -462,13 +472,37 @@ DYNAMIC_CORE = KATMAI COPPERMINE NORTHWOOD PRESCOTT BANIAS \
 endif
 
 ifeq ($(ARCH), x86_64)
-DYNAMIC_CORE = PRESCOTT CORE2 PENRYN DUNNINGTON NEHALEM OPTERON OPTERON_SSE3 BARCELONA BOBCAT ATOM NANO
+DYNAMIC_CORE = PRESCOTT CORE2 
+ifeq ($(DYNAMIC_OLDER), 1)
+DYNAMIC_CORE += PENRYN DUNNINGTON 
+endif
+DYNAMIC_CORE += NEHALEM 
+ifeq ($(DYNAMIC_OLDER), 1)
+DYNAMIC_CORE += OPTERON OPTERON_SSE3 
+endif
+DYNAMIC_CORE += BARCELONA 
+ifeq ($(DYNAMIC_OLDER), 1)
+DYNAMIC_CORE += BOBCAT ATOM NANO
+endif
 ifneq ($(NO_AVX), 1)
 DYNAMIC_CORE += SANDYBRIDGE BULLDOZER PILEDRIVER STEAMROLLER EXCAVATOR
 endif
 ifneq ($(NO_AVX2), 1)
 DYNAMIC_CORE += HASWELL ZEN
 endif
+ifneq ($(NO_AVX512), 1)
+ifneq ($(NO_AVX2), 1)
+DYNAMIC_CORE += SKYLAKEX
+endif
+endif
+endif
+
+ifdef DYNAMIC_LIST
+override DYNAMIC_CORE = PRESCOTT $(DYNAMIC_LIST)
+XCCOMMON_OPT = -DDYNAMIC_LIST -DDYN_PRESCOTT
+XCCOMMON_OPT += $(foreach dcore,$(DYNAMIC_LIST),-DDYN_$(dcore))
+CCOMMON_OPT += $(XCCOMMON_OPT)
+#CCOMMON_OPT	+= -DDYNAMIC_LIST='$(DYNAMIC_LIST)'
 endif
 
 # If DYNAMIC_CORE is not set, DYNAMIC_ARCH cannot do anything, so force it to empty
@@ -902,6 +936,10 @@ ifeq ($(DYNAMIC_ARCH), 1)
 CCOMMON_OPT	+= -DDYNAMIC_ARCH
 endif
 
+ifeq ($(DYNAMIC_OLDER), 1)
+CCOMMON_OPT	+= -DDYNAMIC_OLDER
+endif
+
 ifeq ($(NO_LAPACK), 1)
 CCOMMON_OPT	+= -DNO_LAPACK
 #Disable LAPACK C interface
@@ -924,6 +962,10 @@ ifeq ($(NO_AVX2), 1)
 CCOMMON_OPT	+= -DNO_AVX2
 endif
 
+ifeq ($(NO_AVX512), 1)
+CCOMMON_OPT	+= -DNO_AVX512
+endif
+
 ifdef SMP
 CCOMMON_OPT	+= -DSMP_SERVER
 
@@ -1230,6 +1272,7 @@ export MSA_FLAGS
 export KERNELDIR
 export FUNCTION_PROFILE
 export TARGET_CORE
+export NO_AVX512
 
 export SGEMM_UNROLL_M
 export SGEMM_UNROLL_N

Makefile.x86_64

@@ -8,6 +8,13 @@ endif
 endif
 endif
 
+ifeq ($(CORE), SKYLAKEX)
+ifndef NO_AVX512
+CCOMMON_OPT += -march=skylake-avx512
+FCOMMON_OPT += -march=skylake-avx512
+endif
+endif
+
 ifeq ($(OSNAME), Interix)
 ARFLAGS		= -m x64
 endif

TargetList.txt

@@ -20,6 +20,7 @@ DUNNINGTON
 NEHALEM
 SANDYBRIDGE
 HASWELL
+SKYLAKEX
 ATOM
 
 b)AMD CPU:

c_check

@@ -201,6 +201,21 @@ $architecture = zarch  if ($data =~ /ARCH_ZARCH/);
 $binformat    = bin32;
 $binformat    = bin64  if ($data =~ /BINARY_64/);
 
+$no_avx512= 0;
+if (($architecture eq "x86") || ($architecture eq "x86_64")) {
+    $code = '"vbroadcastss -4 * 4(%rsi), %zmm2"';
+    print $tmpf "int main(void){ __asm__ volatile($code); }\n";
+    $args = " -march=skylake-avx512 -o $tmpf.o -x c $tmpf";
+    my @cmd = ("$compiler_name $args >/dev/null 2>/dev/null");
+    system(@cmd) == 0;
+    if ($? != 0) {
+	$no_avx512 = 1;
+    } else {
+	$no_avx512 = 0;
+    }
+    unlink("tmpf.o");
+}
+
 $data = `$compiler_name -S ctest1.c && grep globl ctest1.s | head -n 1 && rm -f ctest1.s`;
 
 $data =~ /globl\s([_\.]*)(.*)/;
@@ -288,6 +303,7 @@ print MAKEFILE "CROSS=1\n" if $cross != 0;
 print MAKEFILE "CEXTRALIB=$linker_L $linker_l $linker_a\n";
 print MAKEFILE "HAVE_MSA=1\n" if $have_msa eq 1;
 print MAKEFILE "MSA_FLAGS=$msa_flags\n" if $have_msa eq 1;
+print MAKEFILE "NO_AVX512=1\n" if $no_avx512 eq 1;
 
 $os           =~ tr/[a-z]/[A-Z]/;
 $architecture =~ tr/[a-z]/[A-Z]/;

cblas.h

@@ -82,6 +82,11 @@ CBLAS_INDEX cblas_idamax(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPE
 CBLAS_INDEX cblas_icamax(OPENBLAS_CONST blasint n, OPENBLAS_CONST void  *x, OPENBLAS_CONST blasint incx);
 CBLAS_INDEX cblas_izamax(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);
 
+CBLAS_INDEX cblas_isamin(OPENBLAS_CONST blasint n, OPENBLAS_CONST float  *x, OPENBLAS_CONST blasint incx);
+CBLAS_INDEX cblas_idamin(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPENBLAS_CONST blasint incx);
+CBLAS_INDEX cblas_icamin(OPENBLAS_CONST blasint n, OPENBLAS_CONST void  *x, OPENBLAS_CONST blasint incx);
+CBLAS_INDEX cblas_izamin(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);
+
 void cblas_saxpy(OPENBLAS_CONST blasint n, OPENBLAS_CONST float alpha, OPENBLAS_CONST float *x, OPENBLAS_CONST blasint incx, float *y, OPENBLAS_CONST blasint incy);
 void cblas_daxpy(OPENBLAS_CONST blasint n, OPENBLAS_CONST double alpha, OPENBLAS_CONST double *x, OPENBLAS_CONST blasint incx, double *y, OPENBLAS_CONST blasint incy);
 void cblas_caxpy(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *alpha, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx, void *y, OPENBLAS_CONST blasint incy);

cmake/OpenBLASConfig.cmake.in

@@ -0,0 +1,79 @@
+# OpenBLASConfig.cmake
+# --------------------
+#
+# OpenBLAS cmake module.
+# This module sets the following variables in your project::
+#
+#   OpenBLAS_FOUND - true if OpenBLAS and all required components found on the system
+#   OpenBLAS_VERSION - OpenBLAS version in format Major.Minor.Release
+#   OpenBLAS_INCLUDE_DIRS - Directory where OpenBLAS header is located.
+#   OpenBLAS_INCLUDE_DIR - same as DIRS
+#   OpenBLAS_LIBRARIES - OpenBLAS library to link against.
+#   OpenBLAS_LIBRARY - same as LIBRARIES
+#
+#
+# Available components::
+#
+##   shared - search for only shared library
+##   static - search for only static library
+#   serial - search for unthreaded library
+#   pthread - search for native pthread threaded library
+#   openmp - search for OpenMP threaded library
+#
+#
+# Exported targets::
+#
+# If OpenBLAS is found, this module defines the following :prop_tgt:`IMPORTED`
+## target. Target is shared _or_ static, so, for both, use separate, not
+## overlapping, installations. ::
+#
+#   OpenBLAS::OpenBLAS - the main OpenBLAS library #with header & defs attached.
+#
+#
+# Suggested usage::
+#
+#   find_package(OpenBLAS)
+#   find_package(OpenBLAS 0.2.20 EXACT CONFIG REQUIRED COMPONENTS pthread)
+#
+#
+# The following variables can be set to guide the search for this package::
+#
+#   OpenBLAS_DIR - CMake variable, set to directory containing this Config file
+#   CMAKE_PREFIX_PATH - CMake variable, set to root directory of this package
+#   PATH - environment variable, set to bin directory of this package
+#   CMAKE_DISABLE_FIND_PACKAGE_OpenBLAS - CMake variable, disables
+#     find_package(OpenBLAS) when not REQUIRED, perhaps to force internal build
+
+@PACKAGE_INIT@
+
+set(PN OpenBLAS)
+
+# need to check that the @USE_*@ evaluate to something cmake can perform boolean logic upon
+if(@USE_OPENMP@)
+    set(${PN}_openmp_FOUND 1)
+elseif(@USE_THREAD@)
+    set(${PN}_pthread_FOUND 1)
+else()
+    set(${PN}_serial_FOUND 1)
+endif()
+
+check_required_components(${PN})
+
+#-----------------------------------------------------------------------------
+# Don't include targets if this file is being picked up by another
+# project which has already built this as a subproject
+#-----------------------------------------------------------------------------
+if(NOT TARGET ${PN}::OpenBLAS)
+    include("${CMAKE_CURRENT_LIST_DIR}/${PN}Targets.cmake")
+
+    get_property(_loc TARGET ${PN}::OpenBLAS PROPERTY LOCATION)
+    set(${PN}_LIBRARY ${_loc})
+    get_property(_ill TARGET ${PN}::OpenBLAS PROPERTY INTERFACE_LINK_LIBRARIES)
+    set(${PN}_LIBRARIES ${_ill})
+
+    get_property(_id TARGET ${PN}::OpenBLAS PROPERTY INCLUDE_DIRECTORIES)
+    set(${PN}_INCLUDE_DIR ${_id})
+    get_property(_iid TARGET ${PN}::OpenBLAS PROPERTY INTERFACE_INCLUDE_DIRECTORIES)
+    set(${PN}_INCLUDE_DIRS ${_iid})
+endif()
+

cmake/arch.cmake

@@ -49,13 +49,27 @@ if (DYNAMIC_ARCH)
   endif ()
 
   if (X86_64)
-    set(DYNAMIC_CORE PRESCOTT CORE2 PENRYN DUNNINGTON NEHALEM OPTERON OPTERON_SSE3 BARCELONA BOBCAT ATOM NANO)
+    set(DYNAMIC_CORE PRESCOTT CORE2)
+    if (DYNAMIC_OLDER)
+	set (DYNAMIC_CORE ${DYNAMIC_CORE} PENRYN DUNNINGTON)
+    endif ()
+    set (DYNAMIC_CORE ${DYNAMIC_CORE} NEHALEM)
+    if (DYNAMIC_OLDER)
+	set (DYNAMIC_CORE ${DYNAMIC_CORE} OPTERON OPTERON_SSE3)
+    endif ()
+    set (DYNAMIC_CORE ${DYNAMIC_CORE} BARCELONA) 
+    if (DYNAMIC_OLDER)
+	set (DYNAMIC_CORE ${DYNAMIC_CORE} BOBCAT ATOM NANO)
+    endif ()
     if (NOT NO_AVX)
       set(DYNAMIC_CORE ${DYNAMIC_CORE} SANDYBRIDGE BULLDOZER PILEDRIVER STEAMROLLER EXCAVATOR)
     endif ()
     if (NOT NO_AVX2)
       set(DYNAMIC_CORE ${DYNAMIC_CORE} HASWELL ZEN)
     endif ()
+    if (NOT NO_AVX512)
+      set(DYNAMIC_CORE ${DYNAMIC_CORE} SKYLAKEX)
+    endif ()
   endif ()
 
   if (NOT DYNAMIC_CORE)

cmake/openblas.pc.in

@@ -1,7 +1,7 @@
 libdir=@CMAKE_INSTALL_FULL_LIBDIR@
 includedir=@CMAKE_INSTALL_FULL_INCLUDEDIR@
 
-openblas_config=USE_64BITINT=@USE_64BITINT@ NO_CBLAS=@NO_CBLAS@ NO_LAPACK=@NO_LAPACK@ NO_LAPACKE=@NO_LAPACKE@ DYNAMIC_ARCH=@DYNAMIC_ARCH@ NO_AFFINITY=@NO_AFFINITY@ USE_OPENMP=@USE_OPENMP@ @CORE@ MAX_THREADS=@NUM_THREADS@ 
+openblas_config=USE_64BITINT=@USE_64BITINT@ NO_CBLAS=@NO_CBLAS@ NO_LAPACK=@NO_LAPACK@ NO_LAPACKE=@NO_LAPACKE@ DYNAMIC_ARCH=@DYNAMIC_ARCH@ DYNAMIC_OLDER=@DYNAMIC_OLDER@ NO_AFFINITY=@NO_AFFINITY@ USE_OPENMP=@USE_OPENMP@ @CORE@ MAX_THREADS=@NUM_THREADS@ 
 Name: OpenBLAS
 Description: OpenBLAS is an optimized BLAS library based on GotoBLAS2 1.13 BSD version
 Version: @OPENBLAS_VERSION@

cmake/system.cmake

@@ -33,7 +33,7 @@ endif ()
 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 ${TARGET} STREQUAL "SANDYBRIDGE")
+  if (${TARGET} STREQUAL "HASWELL" OR ${TARGET} STREQUAL "SANDYBRIDGE" OR ${TARGET} STREQUAL "SKYLAKEX")
     set(TARGET "NEHALEM")
   endif ()
   if (${TARGET} STREQUAL "BULLDOZER" OR ${TARGET} STREQUAL "PILEDRIVER" OR ${TARGET} STREQUAL "ZEN")
@@ -163,6 +163,9 @@ endif ()
 
 if (DYNAMIC_ARCH)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_ARCH")
+  if (DYNAMIC_OLDER)
+    set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_OLDER")
+  endif ()
 endif ()
 
 if (NO_LAPACK)

cmake/system_check.cmake

@@ -66,3 +66,12 @@ else()
   set(BINARY32 1)
 endif()
 
+if (X86_64 OR X86)
+  file(WRITE ${PROJECT_BINARY_DIR}/avx512.tmp "int main(void){ __asm__ volatile(\"vbroadcastss -4 * 4(%rsi), %zmm2\"); }")
+execute_process(COMMAND ${CMAKE_C_COMPILER} -v -o ${PROJECT_BINARY_DIR}/avx512.o -x c ${PROJECT_BINARY_DIR}/avx512.tmp RESULT_VARIABLE NO_AVX512)
+if (NO_AVX512 EQUAL 1)
+set (CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX512")
+endif()
+  file(REMOVE "avx512.tmp" "avx512.o")
+endif()
+

common.h

@@ -642,6 +642,7 @@ void gotoblas_profile_init(void);
 void gotoblas_profile_quit(void);
 
 #ifdef USE_OPENMP
+
 #ifndef C_MSVC
 int omp_in_parallel(void);
 int omp_get_num_procs(void);
@@ -649,12 +650,21 @@ int omp_get_num_procs(void);
 __declspec(dllimport) int __cdecl omp_in_parallel(void);
 __declspec(dllimport) int __cdecl omp_get_num_procs(void);
 #endif
+
 #if (__STDC_VERSION__ >= 201112L)
+#if defined(C_GCC) && ( __GNUC__ < 7) 
+// workaround for GCC bug 65467
 #ifndef _Atomic
 #define _Atomic volatile
 #endif
-#include <stdatomic.h>
 #endif
+#include <stdatomic.h>
+#else
+#ifndef _Atomic
+#define _Atomic volatile
+#endif
+#endif
+
 #else
 #ifdef __ELF__
 int omp_in_parallel  (void) __attribute__ ((weak));

common_stackalloc.h

@@ -51,10 +51,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
   /* make it volatile because some function (ex: dgemv_n.S) */                 \
   /* do not restore all register */                                            \
   volatile int stack_alloc_size = SIZE;                                        \
-  if(stack_alloc_size > MAX_STACK_ALLOC / sizeof(TYPE))                    \
+  if (stack_alloc_size > MAX_STACK_ALLOC / sizeof(TYPE)) stack_alloc_size = 0; \
-    stack_alloc_size = 0;                                                  \
   STACK_ALLOC_PROTECT_SET                                                      \
-  TYPE stack_buffer[stack_alloc_size] __attribute__((aligned(0x20)));    \
+  /* Avoid declaring an array of length 0 */                                   \
+  TYPE stack_buffer[stack_alloc_size ? stack_alloc_size : 1]                   \
+      __attribute__((aligned(0x20)));                                          \
   BUFFER = stack_alloc_size ? stack_buffer : (TYPE *)blas_memory_alloc(1);
 #else
   //Original OpenBLAS/GotoBLAS codes.

common_x86_64.h

@@ -60,8 +60,13 @@
 #endif
 */
 
-#define MB
+#ifdef __GNUC__
-#define WMB
+#define MB do { __asm__ __volatile__("": : :"memory"); } while (0)
+#define WMB do { __asm__ __volatile__("": : :"memory"); } while (0)
+#else
+#define MB do {} while (0)
+#define WMB do {} while (0)
+#endif
 
 static void __inline blas_lock(volatile BLASULONG *address){
 

cpuid.h

@@ -115,6 +115,7 @@
 #define CORE_STEAMROLLER 25
 #define CORE_EXCAVATOR   26
 #define CORE_ZEN         27
+#define CORE_SKYLAKEX    28
 
 #define HAVE_SSE      (1 <<  0)
 #define HAVE_SSE2     (1 <<  1)
@@ -137,6 +138,7 @@
 #define HAVE_AVX      (1 <<  18)
 #define HAVE_FMA4     (1 <<  19)
 #define HAVE_FMA3     (1 <<  20)
+#define HAVE_AVX512VL (1 <<  21)
 
 #define CACHE_INFO_L1_I     1
 #define CACHE_INFO_L1_D     2
@@ -211,5 +213,6 @@ typedef struct {
 #define CPUTYPE_STEAMROLLER 		49
 #define CPUTYPE_EXCAVATOR 		50
 #define CPUTYPE_ZEN 			51
+#define CPUTYPE_SKYLAKEX		52
 
 #endif

cpuid_x86.c

@@ -50,6 +50,8 @@
 #ifdef NO_AVX
 #define CPUTYPE_HASWELL CPUTYPE_NEHALEM
 #define CORE_HASWELL CORE_NEHALEM
+#define CPUTYPE_SKYLAKEX CPUTYPE_NEHALEM
+#define CORE_SKYLAKEX CORE_NEHALEM
 #define CPUTYPE_SANDYBRIDGE CPUTYPE_NEHALEM
 #define CORE_SANDYBRIDGE CORE_NEHALEM
 #define CPUTYPE_BULLDOZER CPUTYPE_BARCELONA
@@ -1299,6 +1301,19 @@ int get_cpuname(void){
           else
 	    return CPUTYPE_NEHALEM;
 	case 5:
+	  // Skylake X
+#ifndef NO_AVX512
+	  return CPUTYPE_SKYLAKEX;
+#else
+	  if(support_avx())
+#ifndef NO_AVX2
+	  return CPUTYPE_HASWELL;
+#else
+	  return CPUTYPE_SANDYBRIDGE;
+#endif
+	  else
+	  return CPUTYPE_NEHALEM;
+#endif			
         case 14:
 	  // Skylake
           if(support_avx())
@@ -1324,6 +1339,23 @@ int get_cpuname(void){
 	    return CPUTYPE_NEHALEM;
 	}
 	break;
+      case 6:
+        switch (model) {
+        case 6: // Cannon Lake
+#ifndef NO_AVX512
+	  return CPUTYPE_SKYLAKEX;
+#else
+	  if(support_avx())
+#ifndef NO_AVX2
+	  return CPUTYPE_HASWELL;
+#else
+	  return CPUTYPE_SANDYBRIDGE;
+#endif
+	  else
+	  return CPUTYPE_NEHALEM;
+#endif			
+        }
+      break;  
       case 9:
       case 8: 
         switch (model) {
@@ -1556,6 +1588,7 @@ static char *cpuname[] = {
   "STEAMROLLER",
   "EXCAVATOR",
   "ZEN",
+  "SKYLAKEX"	
 };
 
 static char *lowercpuname[] = {
@@ -1610,6 +1643,7 @@ static char *lowercpuname[] = {
   "steamroller",
   "excavator",
   "zen",
+  "skylakex"
 };
 
 static char *corename[] = {
@@ -1641,6 +1675,7 @@ static char *corename[] = {
   "STEAMROLLER",
   "EXCAVATOR",
   "ZEN",
+  "SKYLAKEX"	
 };
 
 static char *corename_lower[] = {
@@ -1672,6 +1707,7 @@ static char *corename_lower[] = {
   "steamroller",
   "excavator",
   "zen",
+  "skylakex"	
 };
 
 
@@ -1860,6 +1896,19 @@ int get_coretype(void){
           else
 	    return CORE_NEHALEM;
 	case 5:
+	 // Skylake X
+#ifndef NO_AVX512
+	    return CORE_SKYLAKEX;
+#else
+	  if(support_avx())
+#ifndef NO_AVX2
+	    return CORE_HASWELL;
+#else
+	    return CORE_SANDYBRIDGE;
+#endif
+	  else
+	    return CORE_NEHALEM;
+#endif			
 	case 14:
 	  // Skylake
           if(support_avx())

ctest/Makefile

@@ -102,7 +102,13 @@ clean ::
 	rm -f x*
 
 FLDFLAGS = $(FFLAGS:-fPIC=) $(LDFLAGS)
-CEXTRALIB =
+ifeq ($(USE_OPENMP), 1)
+ifeq ($(F_COMPILER), GFORTRAN)
+ifeq ($(C_COMPILER), CLANG)
+CEXTRALIB = -lomp
+endif
+endif
+endif
 
 # Single real
 xscblat1: $(stestl1o) c_sblat1.o $(TOPDIR)/$(LIBNAME)

driver/level3/Makefile

@@ -362,7 +362,7 @@ cgemm_ct.$(SUFFIX) : gemm.c level3.c  ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 cgemm_cr.$(SUFFIX) : gemm.c level3.c  ../../param.h
-	$(CC) $(CFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 cgemm_cc.$(SUFFIX) : gemm.c level3.c  ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -410,7 +410,7 @@ zgemm_ct.$(SUFFIX) : gemm.c level3.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 zgemm_cr.$(SUFFIX) : gemm.c level3.c ../../param.h
-	$(CC) $(CFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 zgemm_cc.$(SUFFIX) : gemm.c level3.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -458,7 +458,7 @@ xgemm_ct.$(SUFFIX) : gemm.c level3.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 xgemm_cr.$(SUFFIX) : gemm.c level3.c ../../param.h
-	$(CC) $(CFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 xgemm_cc.$(SUFFIX) : gemm.c level3.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -558,7 +558,7 @@ cgemm_thread_ct.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 cgemm_thread_cr.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
-	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 cgemm_thread_cc.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -606,7 +606,7 @@ zgemm_thread_ct.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 zgemm_thread_cr.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
-	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 zgemm_thread_cc.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -654,7 +654,7 @@ xgemm_thread_ct.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 xgemm_thread_cr.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
-	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 xgemm_thread_cc.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -1821,7 +1821,7 @@ cgemm3m_ct.$(SUFFIX) : gemm3m.c gemm3m_level3.c
 	$(CC) $(CFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 cgemm3m_cr.$(SUFFIX) : gemm3m.c gemm3m_level3.c
-	$(CC) $(CFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 cgemm3m_cc.$(SUFFIX) : gemm3m.c gemm3m_level3.c
 	$(CC) $(CFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -1869,7 +1869,7 @@ zgemm3m_ct.$(SUFFIX) : gemm3m.c gemm3m_level3.c
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 zgemm3m_cr.$(SUFFIX) : gemm3m.c gemm3m_level3.c
-	$(CC) $(CFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 zgemm3m_cc.$(SUFFIX) : gemm3m.c gemm3m_level3.c
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -1917,7 +1917,7 @@ xgemm3m_ct.$(SUFFIX) : gemm3m.c gemm3m_level3.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 xgemm3m_cr.$(SUFFIX) : gemm3m.c gemm3m_level3.c ../../param.h
-	$(CC) $(CFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 xgemm3m_cc.$(SUFFIX) : gemm3m.c gemm3m_level3.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -1974,7 +1974,7 @@ cgemm3m_thread_ct.$(SUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 cgemm3m_thread_cr.$(SUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
-	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 cgemm3m_thread_cc.$(SUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -2022,7 +2022,7 @@ zgemm3m_thread_ct.$(SUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 zgemm3m_thread_cr.$(SUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
-	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 zgemm3m_thread_cc.$(SUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -2070,7 +2070,7 @@ xgemm3m_thread_ct.$(SUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 xgemm3m_thread_cr.$(SUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
-	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 xgemm3m_thread_cc.$(SUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -2731,7 +2731,7 @@ cgemm_ct.$(PSUFFIX) : gemm.c level3.c  ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 cgemm_cr.$(PSUFFIX) : gemm.c level3.c  ../../param.h
-	$(CC) $(PFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 cgemm_cc.$(PSUFFIX) : gemm.c level3.c  ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -2779,7 +2779,7 @@ zgemm_ct.$(PSUFFIX) : gemm.c level3.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 zgemm_cr.$(PSUFFIX) : gemm.c level3.c ../../param.h
-	$(CC) $(PFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 zgemm_cc.$(PSUFFIX) : gemm.c level3.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -2827,7 +2827,7 @@ xgemm_ct.$(PSUFFIX) : gemm.c level3.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 xgemm_cr.$(PSUFFIX) : gemm.c level3.c ../../param.h
-	$(CC) $(PFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 xgemm_cc.$(PSUFFIX) : gemm.c level3.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -2927,7 +2927,7 @@ cgemm_thread_ct.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 cgemm_thread_cr.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
-	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 cgemm_thread_cc.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -2975,7 +2975,7 @@ zgemm_thread_ct.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 zgemm_thread_cr.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
-	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 zgemm_thread_cc.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -3023,7 +3023,7 @@ xgemm_thread_ct.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 xgemm_thread_cr.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
-	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 xgemm_thread_cc.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -4190,7 +4190,7 @@ cgemm3m_ct.$(PSUFFIX) : gemm3m.c gemm3m_level3.c
 	$(CC) $(PFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 cgemm3m_cr.$(PSUFFIX) : gemm3m.c gemm3m_level3.c
-	$(CC) $(PFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 cgemm3m_cc.$(PSUFFIX) : gemm3m.c gemm3m_level3.c
 	$(CC) $(PFLAGS) $(BLOCKS) -c -UDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -4238,7 +4238,7 @@ zgemm3m_ct.$(PSUFFIX) : gemm3m.c gemm3m_level3.c
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 zgemm3m_cr.$(PSUFFIX) : gemm3m.c gemm3m_level3.c
-	$(CC) $(PFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 zgemm3m_cc.$(PSUFFIX) : gemm3m.c gemm3m_level3.c
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -4286,7 +4286,7 @@ xgemm3m_ct.$(PSUFFIX) : gemm3m.c gemm3m_level3.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 xgemm3m_cr.$(PSUFFIX) : gemm3m.c gemm3m_level3.c ../../param.h
-	$(CC) $(PFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 xgemm3m_cc.$(PSUFFIX) : gemm3m.c gemm3m_level3.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DXDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -4343,7 +4343,7 @@ cgemm3m_thread_ct.$(PSUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 cgemm3m_thread_cr.$(PSUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
-	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 cgemm3m_thread_cc.$(PSUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -4391,7 +4391,7 @@ zgemm3m_thread_ct.$(PSUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 zgemm3m_thread_cr.$(PSUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
-	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 zgemm3m_thread_cc.$(PSUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DDOUBLE -DCOMPLEX -DCC $< -o $(@F)
@@ -4439,7 +4439,7 @@ xgemm3m_thread_ct.$(PSUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCT $< -o $(@F)
 
 xgemm3m_thread_cr.$(PSUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
-	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCR $< -o $(@F)
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCR=CR $< -o $(@F)
 
 xgemm3m_thread_cc.$(PSUFFIX) : gemm3m.c level3_gemm3m_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DXDOUBLE -DCOMPLEX -DCC $< -o $(@F)

driver/level3/level3_thread.c

@@ -91,11 +91,7 @@
 #endif
 
 typedef struct {
-#if __STDC_VERSION__ >= 201112L
-_Atomic
-#else  
   volatile
-#endif
    BLASLONG working[MAX_CPU_NUMBER][CACHE_LINE_SIZE * DIVIDE_RATE];
 } job_t;
 
@@ -348,12 +344,6 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
     div_n = (n_to - n_from + DIVIDE_RATE - 1) / DIVIDE_RATE;
     for (js = n_from, bufferside = 0; js < n_to; js += div_n, bufferside ++) {
 
-      /* Make sure if no one is using workspace */
-      START_RPCC();
-      for (i = 0; i < args -> nthreads; i++)
-	while (job[mypos].working[i][CACHE_LINE_SIZE * bufferside]) {YIELDING;};
-      STOP_RPCC(waiting1);
-
 #if defined(FUSED_GEMM) && !defined(TIMING)
 
       /* Fused operation to copy region of B into workspace and apply kernel */
@@ -391,11 +381,16 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
       }
 #endif
 
+      for (i = mypos_n * nthreads_m; i < (mypos_n + 1) * nthreads_m; i++) {
+        /* Make sure if no one is using workspace */
+        START_RPCC();
+        while (job[mypos].working[i][CACHE_LINE_SIZE * bufferside]) {YIELDING;MB;};
+        STOP_RPCC(waiting1);
         /* Set flag so other threads can access local region of B */
-      for (i = mypos_n * nthreads_m; i < (mypos_n + 1) * nthreads_m; i++)
         job[mypos].working[i][CACHE_LINE_SIZE * bufferside] = (BLASLONG)buffer[bufferside];
         WMB;
       }
+    }
 
     /* Get regions of B from other threads and apply kernel */
     current = mypos;
@@ -413,7 +408,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 
 	  /* Wait until other region of B is initialized */
 	  START_RPCC();
-	  while(job[current].working[mypos][CACHE_LINE_SIZE * bufferside] == 0) {YIELDING;};
+	  while(job[current].working[mypos][CACHE_LINE_SIZE * bufferside] == 0) {YIELDING;MB;};
 	  STOP_RPCC(waiting2);
 
           /* Apply kernel with local region of A and part of other region of B */
@@ -430,7 +425,8 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 
         /* Clear synchronization flag if this thread is done with other region of B */
 	if (m_to - m_from == min_i) {
-	  job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
+	  job[current].working[mypos][CACHE_LINE_SIZE * bufferside] = 0;
+	  WMB;
 	}
       }
     } while (current != mypos);
@@ -472,7 +468,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 
           /* Clear synchronization flag if this thread is done with region of B */
           if (is + min_i >= m_to) {
-            job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
+            job[current].working[mypos][CACHE_LINE_SIZE * bufferside] = 0;
             WMB;
           }
 	}
@@ -492,7 +488,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
   START_RPCC();
   for (i = 0; i < args -> nthreads; i++) {
     for (js = 0; js < DIVIDE_RATE; js++) {
-      while (job[mypos].working[i][CACHE_LINE_SIZE * js] ) {YIELDING;};
+      while (job[mypos].working[i][CACHE_LINE_SIZE * js] ) {YIELDING;MB;};
     }
   }
   STOP_RPCC(waiting3);
@@ -658,8 +654,8 @@ static int gemm_driver(blas_arg_t *args, BLASLONG *range_m, BLASLONG
     }
 
     /* Clear synchronization flags */
-    for (i = 0; i < MAX_CPU_NUMBER; i++) {
+    for (i = 0; i < nthreads; i++) {
-      for (j = 0; j < MAX_CPU_NUMBER; j++) {
+      for (j = 0; j < nthreads; j++) {
 	for (k = 0; k < DIVIDE_RATE; k++) {
 	  job[i].working[j][CACHE_LINE_SIZE * k] = 0;
 	}

driver/others/blas_server_omp.c

@@ -48,6 +48,10 @@
 
 #else
 
+#ifndef OMP_SCHED
+#define OMP_SCHED static
+#endif
+
 int blas_server_avail = 0;
 
 static void * blas_thread_buffer[MAX_PARALLEL_NUMBER][MAX_CPU_NUMBER];
@@ -331,7 +335,7 @@ int exec_blas(BLASLONG num, blas_queue_t *queue){
       break;
   }
 
-#pragma omp parallel for schedule(static)
+#pragma omp parallel for schedule(OMP_SCHED)
   for (i = 0; i < num; i ++) {
 
 #ifndef USE_SIMPLE_THREADED_LEVEL3

driver/others/dynamic.c

@@ -49,6 +49,167 @@
 #define EXTERN
 #endif
 
+#ifdef DYNAMIC_LIST
+extern gotoblas_t gotoblas_PRESCOTT;
+
+#ifdef DYN_ATHLON
+extern gotoblas_t gotoblas_ATHLON;
+#else
+#define gotoblas_ATHLON gotoblas_PRESCOTT
+#endif
+#ifdef DYN_KATMAI
+extern gotoblas_t gotoblas_KATMAI;
+#else
+#define gotoblas_KATMAI gotoblas_PRESCOTT
+#endif
+#ifdef DYN_BANIAS
+extern gotoblas_t gotoblas_BANIAS;
+#else
+#define gotoblas_BANIAS gotoblas_PRESCOTT
+#endif
+#ifdef DYN_COPPERMINE
+extern gotoblas_t gotoblas_COPPERMINE;
+#else
+#define gotoblas_COPPERMINE gotoblas_PRESCOTT
+#endif
+#ifdef DYN_NORTHWOOD
+extern gotoblas_t gotoblas_NORTHWOOD;
+#else
+#define gotoblas_NORTHWOOD gotoblas_PRESCOTT
+#endif
+#ifdef DYN_CORE2
+extern gotoblas_t gotoblas_CORE2;
+#else
+#define gotoblas_CORE2 gotoblas_PRESCOTT
+#endif
+#ifdef DYN_NEHALEM
+extern gotoblas_t gotoblas_NEHALEM;
+#else
+#define gotoblas_NEHALEM gotoblas_PRESCOTT
+#endif
+#ifdef DYN_BARCELONA
+extern gotoblas_t gotoblas_BARCELONA;
+#elif defined(DYN_NEHALEM)
+#define gotoblas_BARCELONA gotoblas_NEHALEM
+#else
+#define gotoblas_BARCELONA gotoblas_PRESCOTT
+#endif
+#ifdef DYN_ATOM
+extern gotoblas_t gotoblas_ATOM;
+elif defined(DYN_NEHALEM)
+#define gotoblas_ATOM gotoblas_NEHALEM
+#else
+#define gotoblas_ATOM gotoblas_PRESCOTT
+#endif
+#ifdef DYN_NANO
+extern gotoblas_t gotoblas_NANO;
+#else
+#define gotoblas_NANO gotoblas_PRESCOTT
+#endif
+#ifdef DYN_PENRYN
+extern gotoblas_t gotoblas_PENRYN;
+#else
+#define gotoblas_PENRYN gotoblas_PRESCOTT
+#endif
+#ifdef DYN_DUNNINGTON
+extern gotoblas_t gotoblas_DUNNINGTON;
+#else
+#define gotoblas_DUNNINGTON gotoblas_PRESCOTT
+#endif
+#ifdef DYN_OPTERON
+extern gotoblas_t gotoblas_OPTERON;
+#else
+#define gotoblas_OPTERON gotoblas_PRESCOTT
+#endif
+#ifdef DYN_OPTERON_SSE3
+extern gotoblas_t gotoblas_OPTERON_SSE3;
+#else
+#define gotoblas_OPTERON_SSE3 gotoblas_PRESCOTT
+#endif
+#ifdef DYN_BOBCAT
+extern gotoblas_t gotoblas_BOBCAT;
+#elif defined(DYN_NEHALEM)
+#define gotoblas_BOBCAT gotoblas_NEHALEM
+#else
+#define gotoblas_BOBCAT gotoblas_PRESCOTT
+#endif
+#ifdef DYN_SANDYBRIDGE
+extern gotoblas_t gotoblas_SANDYBRIDGE;
+#elif defined(DYN_NEHALEM)
+#define gotoblas_SANDYBRIDGE gotoblas_NEHALEM
+#else
+#define gotoblas_SANDYBRIDGE gotoblas_PRESCOTT
+#endif
+#ifdef DYN_BULLDOZER
+extern gotoblas_t gotoblas_BULLDOZER;
+#elif defined(DYN_SANDYBRIDGE)
+#define gotoblas_BULLDOZER gotoblas_SANDYBRIDGE
+#elif defined(DYN_NEHALEM)
+#define gotoblas_BULLDOZER gotoblas_NEHALEM
+#else
+#define gotoblas_BULLDOZER gotoblas_PRESCOTT
+#endif
+#ifdef DYN_PILEDRIVER
+extern gotoblas_t gotoblas_PILEDRIVER;
+#elif defined(DYN_SANDYBRIDGE)
+#define gotoblas_PILEDRIVER gotoblas_SANDYBRIDGE
+#elif defined(DYN_NEHALEM)
+#define gotoblas_PILEDRIVER gotoblas_NEHALEM
+#else
+#define gotoblas_PILEDRIVER gotoblas_PRESCOTT
+#endif
+#ifdef DYN_STEAMROLLER
+extern gotoblas_t gotoblas_STEAMROLLER;
+#elif defined(DYN_SANDYBRIDGE)
+#define gotoblas_STEAMROLLER gotoblas_SANDYBRIDGE
+#elif defined(DYN_NEHALEM)
+#define gotoblas_STEAMROLLER gotoblas_NEHALEM
+#else
+#define gotoblas_STEAMROLLER gotoblas_PRESCOTT
+#endif
+#ifdef DYN_EXCAVATOR
+extern gotoblas_t gotoblas_EXCAVATOR;
+#elif defined(DYN_SANDYBRIDGE)
+#define gotoblas_EXCAVATOR gotoblas_SANDYBRIDGE
+#elif defined(DYN_NEHALEM)
+#define gotoblas_EXCAVATOR gotoblas_NEHALEM
+#else
+#define gotoblas_EXCAVATOR gotoblas_PRESCOTT
+#endif
+#ifdef DYN_HASWELL
+extern gotoblas_t gotoblas_HASWELL;
+#elif defined(DYN_SANDYBRIDGE)
+#define gotoblas_HASWELL gotoblas_SANDYBRIDGE
+#elif defined(DYN_NEHALEM)
+#define gotoblas_HASWELL gotoblas_NEHALEM
+#else
+#define gotoblas_HASWELL gotoblas_PRESCOTT
+#endif
+#ifdef DYN_ZEN
+extern gotoblas_t gotoblas_ZEN;
+#elif defined(DYN_HASWELL)
+#define gotoblas_ZEN gotoblas_HASWELL
+#elif defined(DYN_SANDYBRIDGE)
+#define gotoblas_ZEN gotoblas_SANDYBRIDGE
+#elif defined(DYN_NEHALEM)
+#define gotoblas_ZEN gotoblas_NEHALEM
+#else
+#define gotoblas_ZEN gotoblas_PRESCOTT
+#endif
+#ifdef DYN_SKYLAKEX
+extern gotoblas_t gotoblas_SKYLAKEX;
+#elif defined(DYN_HASWELL)
+#define gotoblas_SKYLAKEX gotoblas_HASWELL
+#elif defined(DYN_SANDYBRIDGE)
+#define gotoblas_SKYLAKEX gotoblas_SANDYBRIDGE
+#elif defined(DYN_NEHALEM)
+#define gotoblas_SKYLAKEX gotoblas_NEHALEM
+#else
+#define gotoblas_SKYLAKEX gotoblas_PRESCOTT
+#endif
+
+
+#else // not DYNAMIC_LIST
 EXTERN gotoblas_t  gotoblas_KATMAI;
 EXTERN gotoblas_t  gotoblas_COPPERMINE;
 EXTERN gotoblas_t  gotoblas_NORTHWOOD;
@@ -56,16 +217,27 @@ EXTERN gotoblas_t  gotoblas_BANIAS;
 EXTERN gotoblas_t  gotoblas_ATHLON;
 
 extern gotoblas_t  gotoblas_PRESCOTT;
+extern gotoblas_t  gotoblas_CORE2;
+extern gotoblas_t  gotoblas_NEHALEM;
+extern gotoblas_t  gotoblas_BARCELONA;
+#ifdef DYNAMIC_OLDER
 extern gotoblas_t  gotoblas_ATOM;
 extern gotoblas_t  gotoblas_NANO;
-extern gotoblas_t  gotoblas_CORE2;
 extern gotoblas_t  gotoblas_PENRYN;
 extern gotoblas_t  gotoblas_DUNNINGTON;
-extern gotoblas_t  gotoblas_NEHALEM;
 extern gotoblas_t  gotoblas_OPTERON;
 extern gotoblas_t  gotoblas_OPTERON_SSE3;
-extern gotoblas_t  gotoblas_BARCELONA;
 extern gotoblas_t  gotoblas_BOBCAT;
+#else
+#define gotoblas_ATOM gotoblas_NEHALEM
+#define gotoblas_NANO gotoblas_NEHALEM
+#define gotoblas_PENRYN gotoblas_CORE2
+#define gotoblas_DUNNINGTON gotoblas_CORE2
+#define gotoblas_OPTERON gotoblas_CORE2
+#define gotoblas_OPTERON_SSE3 gotoblas_CORE2
+#define gotoblas_BOBCAT gotoblas_CORE2
+#endif
+
 #ifndef NO_AVX
 extern gotoblas_t  gotoblas_SANDYBRIDGE;
 extern gotoblas_t  gotoblas_BULLDOZER;
@@ -74,15 +246,22 @@ extern gotoblas_t  gotoblas_STEAMROLLER;
 extern gotoblas_t  gotoblas_EXCAVATOR;
 #ifdef NO_AVX2
 #define gotoblas_HASWELL gotoblas_SANDYBRIDGE
+#define gotoblas_SKYLAKEX gotoblas_SANDYBRIDGE
 #define gotoblas_ZEN gotoblas_SANDYBRIDGE
 #else
 extern gotoblas_t  gotoblas_HASWELL;
 extern gotoblas_t  gotoblas_ZEN;
+#ifndef NO_AVX512
+extern gotoblas_t  gotoblas_SKYLAKEX;
+#else
+#define gotoblas_SKYLAKEX gotoblas_HASWELL
+#endif
 #endif
 #else
 //Use NEHALEM kernels for sandy bridge
 #define gotoblas_SANDYBRIDGE gotoblas_NEHALEM
 #define gotoblas_HASWELL gotoblas_NEHALEM
+#define gotoblas_SKYLAKEX gotoblas_NEHALEM
 #define gotoblas_BULLDOZER gotoblas_BARCELONA
 #define gotoblas_PILEDRIVER gotoblas_BARCELONA
 #define gotoblas_STEAMROLLER gotoblas_BARCELONA
@@ -90,6 +269,7 @@ extern gotoblas_t  gotoblas_ZEN;
 #define gotoblas_ZEN gotoblas_BARCELONA
 #endif
 
+#endif // DYNAMIC_LIST
 
 #define VENDOR_INTEL      1
 #define VENDOR_AMD        2
@@ -284,8 +464,21 @@ static gotoblas_t *get_coretype(void){
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
 	}
+	if (model == 5) {	
+	// Intel Skylake X
+#ifndef NO_AVX512
+	  return &gotoblas_SKYLAKEX;
+#else		
+	  if(support_avx())
+	    return &gotoblas_HASWELL;
+	  else {
+	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
+	    return &gotoblas_NEHALEM;
+	  }
+#endif		
+	}
 	//Intel Skylake
-	if (model == 14 || model == 5) {
+	if (model == 14) {
 	  if(support_avx())
 	    return &gotoblas_HASWELL;
 	  else{
@@ -307,6 +500,23 @@ static gotoblas_t *get_coretype(void){
 	  return &gotoblas_NEHALEM;
 	}	
 	return NULL;
+      case 6:
+        if (model == 6) {
+          // Cannon Lake
+#ifndef NO_AVX512
+	  return &gotoblas_SKYLAKEX;
+#else
+	  if(support_avx())
+#ifndef NO_AVX2
+	  return &gotoblas_HASWELL;
+#else
+	  return &gotblas_SANDYBRIDGE;
+#endif
+	  else
+	  return &gotoblas_NEHALEM;
+#endif			
+        }
+        return NULL;  
       case 9:
       case 8:
 	if (model == 14 ) { // Kaby Lake
@@ -445,7 +655,8 @@ static char *corename[] = {
     "Haswell",
     "Steamroller",
     "Excavator",
-    "Zen"
+    "Zen",
+    "SkylakeX"	
 };
 
 char *gotoblas_corename(void) {
@@ -473,7 +684,7 @@ char *gotoblas_corename(void) {
   if (gotoblas == &gotoblas_STEAMROLLER)  return corename[21];
   if (gotoblas == &gotoblas_EXCAVATOR)    return corename[22];
   if (gotoblas == &gotoblas_ZEN)          return corename[23];
-
+  if (gotoblas == &gotoblas_SKYLAKEX)     return corename[24];
   return corename[0];
 }
 
@@ -485,7 +696,7 @@ static gotoblas_t *force_coretype(char *coretype){
 	char message[128];
 	//char mname[20];
 
-	for ( i=1 ; i <= 23; i++)
+	for ( i=1 ; i <= 24; i++)
 	{
 		if (!strncasecmp(coretype,corename[i],20))
 		{
@@ -503,6 +714,7 @@ static gotoblas_t *force_coretype(char *coretype){
 
 	switch (found)
 	{
+		case 24: return (&gotoblas_SKYLAKEX);	
 		case 23: return (&gotoblas_ZEN);
 		case 22: return (&gotoblas_EXCAVATOR);
 		case 21: return (&gotoblas_STEAMROLLER);

driver/others/memory.c

@@ -139,6 +139,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define FIXED_PAGESIZE 4096
 #endif
 
+#ifndef BUFFERS_PER_THREAD
+#ifdef USE_OPENMP
+#define BUFFERS_PER_THREAD (MAX_CPU_NUMBER * 2 * MAX_PARALLEL_NUMBER)
+#else
+#define BUFFERS_PER_THREAD NUM_BUFFERS
+#endif
+#endif
+
 #define BITMASK(a, b, c) ((((a) >> (b)) & (c)))
 
 #if defined(_MSC_VER) && !defined(__clang__)
@@ -180,7 +188,7 @@ int get_num_procs(void) {
 cpu_set_t *cpusetp;
 size_t size;
 int ret;
-// int i,n;
+int i,n;
 
   if (!nums) nums = sysconf(_SC_NPROCESSORS_CONF);
 #if !defined(OS_LINUX)
@@ -318,6 +326,8 @@ int  goto_get_num_procs  (void) {
   return blas_cpu_number;
 }
 
+static void blas_memory_init();
+
 void openblas_fork_handler()
 {
   // This handler shuts down the OpenBLAS-managed PTHREAD pool when OpenBLAS is
@@ -329,7 +339,7 @@ void openblas_fork_handler()
   // implementation of OpenMP.
 #if !((defined(OS_WINDOWS) && !defined(OS_CYGWIN_NT)) || defined(OS_ANDROID)) && defined(SMP_SERVER)
   int err;
-  err = pthread_atfork ((void (*)(void)) BLASFUNC(blas_thread_shutdown), NULL, NULL);
+  err = pthread_atfork ((void (*)(void)) BLASFUNC(blas_thread_shutdown), NULL, blas_memory_init);
   if(err != 0)
     openblas_warning(0, "OpenBLAS Warning ... cannot install fork handler. You may meet hang after fork.\n");
 #endif
@@ -407,16 +417,104 @@ int openblas_get_num_threads(void) {
 #endif
 }
 
-struct release_t {
-  void *address;
-  void (*func)(struct release_t *);
-  long attr;
-};
-
 int hugetlb_allocated = 0;
 
-static struct release_t release_info[NUM_BUFFERS];
+#if defined(OS_WINDOWS)
-static int release_pos = 0;
+#define THREAD_LOCAL __declspec(thread)
+#define LIKELY_ONE(x) (x)
+#else
+#define THREAD_LOCAL __thread
+#define LIKELY_ONE(x) (__builtin_expect(x, 1))
+#endif
+
+/* Stores information about the allocation and how to release it */
+struct alloc_t {
+  /* Whether this allocation is being used */
+  int used;
+  /* Any special attributes needed when releasing this allocation */
+  int attr;
+  /* Function that can properly release this memory */
+  void (*release_func)(struct alloc_t *);
+  /* Pad to 64-byte alignment */
+  char pad[64 - 2 * sizeof(int) - sizeof(void(*))];
+};
+
+/* Convenience macros for storing release funcs */
+#define STORE_RELEASE_FUNC(address, func)                   \
+  if (address != (void *)-1) {                              \
+    struct alloc_t *alloc_info = (struct alloc_t *)address; \
+    alloc_info->release_func = func;                        \
+  }
+
+#define STORE_RELEASE_FUNC_WITH_ATTR(address, func, attr)   \
+  if (address != (void *)-1) {                              \
+    struct alloc_t *alloc_info = (struct alloc_t *)address; \
+    alloc_info->release_func = func;                        \
+    alloc_info->attr = attr;                                \
+  }
+
+/* The number of bytes that will be allocated for each buffer. When allocating
+   memory, we store an alloc_t followed by the actual buffer memory. This means
+   that each allocation always has its associated alloc_t, without the need
+   for an auxiliary tracking structure. */
+static const int allocation_block_size = BUFFER_SIZE + sizeof(struct alloc_t);
+
+/* Clang supports TLS from version 2.8 */
+#if defined(__clang__) && __clang_major__ > 2 || \
+    (__clang_minor__ == 2 || __clang_minor__ == 8)
+#define HAS_COMPILER_TLS
+#endif
+
+/* GCC supports TLS from version 4.1 */
+#if !defined(__clang__) && defined(__GNUC__) && \
+    (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 1))
+#define HAS_COMPILER_TLS
+#endif
+
+/* MSVC supports TLS from version 2005 */
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+#define HAS_COMPILER_TLS
+#endif
+
+/* Versions of XCode before 8 did not properly support TLS */
+#if defined(__apple_build_version__) && __apple_build_version__ < 8000042
+#undef HAS_COMPILER_TLS
+#endif
+
+/* Android NDK's before version 12b did not support TLS */
+#if defined(__ANDROID__) && defined(__clang__)
+#if __has_include(<android/ndk-version.h>)
+#include <android/ndk-version.h>
+#endif
+#if defined(__ANDROID__) && defined(__clang__) && defined(__NDK_MAJOR__) && \
+    defined(__NDK_MINOR__) &&                                               \
+    ((__NDK_MAJOR__ < 12) || ((__NDK_MAJOR__ == 12) && (__NDK_MINOR__ < 1)))
+#undef HAS_COMPILER_TLS
+#endif
+#endif
+
+/* Holds pointers to allocated memory */
+#if defined(SMP) && !defined(USE_OPENMP)
+/* This is the number of threads than can be spawned by the server, which is the
+   server plus the number of threads in the thread pool */
+#  define MAX_ALLOCATING_THREADS MAX_CPU_NUMBER * 2 * MAX_PARALLEL_NUMBER +1
+static int next_memory_table_pos = 0;
+#  if defined(HAS_COMPILER_TLS)
+/* Use compiler generated thread-local-storage */
+static int THREAD_LOCAL local_memory_table_pos = 0;
+#  else
+/* Use system-dependent thread-local-storage */
+#    if defined(OS_WINDOWS)
+static DWORD local_storage_key;
+#    else
+static pthread_key_t local_storage_key;
+#    endif /* defined(OS_WINDOWS) */
+#  endif /* defined(HAS_COMPILER_TLS) */
+#else
+/* There is only one allocating thread when in single-threaded mode and when using OpenMP */
+#  define MAX_ALLOCATING_THREADS 1
+#endif /* defined(SMP) && !defined(USE_OPENMP) */
+static struct alloc_t * local_memory_table[MAX_ALLOCATING_THREADS][BUFFERS_PER_THREAD];
 
 #if defined(OS_LINUX) && !defined(NO_WARMUP)
 static int hot_alloc = 0;
@@ -432,11 +530,41 @@ static pthread_spinlock_t alloc_lock = 0;
 static BLASULONG  alloc_lock = 0UL;
 #endif
 
+/* Returns a pointer to the start of the per-thread memory allocation data */
+static __inline struct alloc_t ** get_memory_table() {
+#if defined(SMP) && !defined(USE_OPENMP)
+#  if !defined(HAS_COMPILER_TLS)
+#    if defined(OS_WINDOWS)
+  int local_memory_table_pos = (int)::TlsGetValue(local_storage_key);
+#    else
+  int local_memory_table_pos = (int)pthread_getspecific(local_storage_key);
+#    endif /* defined(OS_WINDOWS) */
+#  endif /* !defined(HAS_COMPILER_TLS) */
+  if (!local_memory_table_pos) {
+    LOCK_COMMAND(&alloc_lock);
+    local_memory_table_pos = next_memory_table_pos++;
+    if (next_memory_table_pos > MAX_ALLOCATING_THREADS)
+      printf("OpenBLAS : Program will terminate because you tried to start too many threads.\n");
+    UNLOCK_COMMAND(&alloc_lock);
+#  if !defined(HAS_COMPILER_TLS)
+#    if defined(OS_WINDOWS)
+    ::TlsSetValue(local_storage_key, (void*)local_memory_table_pos);
+#    else
+    pthread_setspecific(local_storage_key, (void*)local_memory_table_pos);
+#    endif /* defined(OS_WINDOWS) */
+#  endif /* !defined(HAS_COMPILER_TLS) */
+  }
+  return local_memory_table[local_memory_table_pos];
+#else
+  return local_memory_table[0];
+#endif /* defined(SMP) && !defined(USE_OPENMP) */
+}
+
 #ifdef ALLOC_MMAP
 
-static void alloc_mmap_free(struct release_t *release){
+static void alloc_mmap_free(struct alloc_t *alloc_info){
 
-  if (munmap(release -> address, BUFFER_SIZE)) {
+  if (munmap(alloc_info, allocation_block_size)) {
     printf("OpenBLAS : munmap failed\n");
   }
 }
@@ -450,28 +578,18 @@ static void *alloc_mmap(void *address){
 
   if (address){
     map_address = mmap(address,
-		       BUFFER_SIZE,
+		       allocation_block_size,
 		       MMAP_ACCESS, MMAP_POLICY | MAP_FIXED, -1, 0);
   } else {
     map_address = mmap(address,
-		       BUFFER_SIZE,
+		       allocation_block_size,
 		       MMAP_ACCESS, MMAP_POLICY, -1, 0);
   }
 
-  if (map_address != (void *)-1) {
+  STORE_RELEASE_FUNC(map_address, alloc_mmap_free);
-#if defined(SMP) && !defined(USE_OPENMP)
-    LOCK_COMMAND(&alloc_lock);
-#endif    
-    release_info[release_pos].address = map_address;
-    release_info[release_pos].func    = alloc_mmap_free;
-    release_pos ++;
-#if defined(SMP) && !defined(USE_OPENMP)
-    UNLOCK_COMMAND(&alloc_lock);
-#endif    
-  }
 
 #ifdef OS_LINUX
-  my_mbind(map_address, BUFFER_SIZE, MPOL_PREFERRED, NULL, 0, 0);
+  my_mbind(map_address, allocation_block_size, MPOL_PREFERRED, NULL, 0, 0);
 #endif
 
   return map_address;
@@ -524,25 +642,25 @@ static void *alloc_mmap(void *address){
 
   if (address){
     /* Just give up use advanced operation */
-    map_address = mmap(address, BUFFER_SIZE, MMAP_ACCESS, MMAP_POLICY | MAP_FIXED, -1, 0);
+    map_address = mmap(address, allocation_block_size, MMAP_ACCESS, MMAP_POLICY | MAP_FIXED, -1, 0);
 
 #ifdef OS_LINUX
-    my_mbind(map_address, BUFFER_SIZE, MPOL_PREFERRED, NULL, 0, 0);
+    my_mbind(map_address, allocation_block_size, MPOL_PREFERRED, NULL, 0, 0);
 #endif
 
   } else {
 #if defined(OS_LINUX) && !defined(NO_WARMUP)
     if (hot_alloc == 0) {
-      map_address = mmap(NULL, BUFFER_SIZE, MMAP_ACCESS, MMAP_POLICY, -1, 0);
+      map_address = mmap(NULL, allocation_block_size, MMAP_ACCESS, MMAP_POLICY, -1, 0);
 
 #ifdef OS_LINUX
-      my_mbind(map_address, BUFFER_SIZE, MPOL_PREFERRED, NULL, 0, 0);
+      my_mbind(map_address, allocation_block_size, MPOL_PREFERRED, NULL, 0, 0);
 #endif
 
     } else {
 #endif
 
-      map_address = mmap(NULL, BUFFER_SIZE * SCALING,
+      map_address = mmap(NULL, allocation_block_size * SCALING,
 			 MMAP_ACCESS, MMAP_POLICY, -1, 0);
 
       if (map_address != (void *)-1) {
@@ -550,7 +668,7 @@ static void *alloc_mmap(void *address){
 #ifdef OS_LINUX
 #ifdef DEBUG
 		  int ret=0;
-		  ret=my_mbind(map_address, BUFFER_SIZE * SCALING, MPOL_PREFERRED, NULL, 0, 0);
+		  ret=my_mbind(map_address, allocation_block_size * SCALING, MPOL_PREFERRED, NULL, 0, 0);
 		  if(ret==-1){
 			  int errsv=errno;
 			  perror("OpenBLAS alloc_mmap:");
@@ -558,7 +676,7 @@ static void *alloc_mmap(void *address){
 		  }
 
 #else
-		  my_mbind(map_address, BUFFER_SIZE * SCALING, MPOL_PREFERRED, NULL, 0, 0);
+		  my_mbind(map_address, allocation_block_size * SCALING, MPOL_PREFERRED, NULL, 0, 0);
 #endif
 #endif
 
@@ -566,7 +684,7 @@ static void *alloc_mmap(void *address){
 	allocsize = DGEMM_P * DGEMM_Q * sizeof(double);
 
 	start   = (BLASULONG)map_address;
-	current = (SCALING - 1) * BUFFER_SIZE;
+	current = (SCALING - 1) * allocation_block_size;
 
 	while(current > 0) {
 	  *(BLASLONG *)start = (BLASLONG)start + PAGESIZE;
@@ -581,7 +699,7 @@ static void *alloc_mmap(void *address){
 	best = (BLASULONG)-1;
 	best_address = map_address;
 
-	while ((start + allocsize  < (BLASULONG)map_address + (SCALING - 1) * BUFFER_SIZE)) {
+	while ((start + allocsize  < (BLASULONG)map_address + (SCALING - 1) * allocation_block_size)) {
 
 	  current = run_bench(start, allocsize);
 
@@ -597,7 +715,7 @@ static void *alloc_mmap(void *address){
       if ((BLASULONG)best_address > (BLASULONG)map_address)
 	munmap(map_address,  (BLASULONG)best_address - (BLASULONG)map_address);
 
-      munmap((void *)((BLASULONG)best_address + BUFFER_SIZE), (SCALING - 1) * BUFFER_SIZE + (BLASULONG)map_address - (BLASULONG)best_address);
+      munmap((void *)((BLASULONG)best_address + allocation_block_size), (SCALING - 1) * allocation_block_size + (BLASULONG)map_address - (BLASULONG)best_address);
 
       map_address = best_address;
 
@@ -610,17 +728,7 @@ static void *alloc_mmap(void *address){
   }
 #endif
 
-  if (map_address != (void *)-1) {
+  STORE_RELEASE_FUNC(map_address, alloc_mmap_free);
-#if defined(SMP) && !defined(USE_OPENMP)
-    LOCK_COMMAND(&alloc_lock);
-#endif
-    release_info[release_pos].address = map_address;
-    release_info[release_pos].func    = alloc_mmap_free;
-    release_pos ++;
-#if defined(SMP) && !defined(USE_OPENMP)
-    UNLOCK_COMMAND(&alloc_lock);
-#endif
-  }
 
   return map_address;
 }
@@ -632,9 +740,9 @@ static void *alloc_mmap(void *address){
 
 #ifdef ALLOC_MALLOC
 
-static void alloc_malloc_free(struct release_t *release){
+static void alloc_malloc_free(struct alloc_t *alloc_info){
 
-  free(release -> address);
+  free(alloc_info);
 
 }
 
@@ -642,15 +750,11 @@ static void *alloc_malloc(void *address){
 
   void *map_address;
 
-  map_address = (void *)malloc(BUFFER_SIZE + FIXED_PAGESIZE);
+  map_address = (void *)malloc(allocation_block_size + FIXED_PAGESIZE);
 
   if (map_address == (void *)NULL) map_address = (void *)-1;
 
-  if (map_address != (void *)-1) {
+  STORE_RELEASE_FUNC(map_address, alloc_malloc_free);
-    release_info[release_pos].address = map_address;
-    release_info[release_pos].func    = alloc_malloc_free;
-    release_pos ++;
-  }
 
   return map_address;
 
@@ -667,24 +771,20 @@ void *qfree (void *address);
 #define QCOMMS    0x2
 #define QFAST     0x4
 
-static void alloc_qalloc_free(struct release_t *release){
+static void alloc_qalloc_free(struct alloc_t *alloc_info){
 
-  qfree(release -> address);
+  qfree(alloc_info);
 
 }
 
 static void *alloc_qalloc(void *address){
   void *map_address;
 
-  map_address = (void *)qalloc(QCOMMS | QFAST, BUFFER_SIZE + FIXED_PAGESIZE);
+  map_address = (void *)qalloc(QCOMMS | QFAST, allocation_block_size + FIXED_PAGESIZE);
 
   if (map_address == (void *)NULL) map_address = (void *)-1;
 
-  if (map_address != (void *)-1) {
+  STORE_RELEASE_FUNC(map_address, alloc_qalloc_free);
-    release_info[release_pos].address = map_address;
-    release_info[release_pos].func    = alloc_qalloc_free;
-    release_pos ++;
-  }
 
   return (void *)(((BLASULONG)map_address + FIXED_PAGESIZE - 1) & ~(FIXED_PAGESIZE - 1));
 }
@@ -693,9 +793,9 @@ static void *alloc_qalloc(void *address){
 
 #ifdef ALLOC_WINDOWS
 
-static void alloc_windows_free(struct release_t *release){
+static void alloc_windows_free(struct alloc_t *alloc_info){
 
-  VirtualFree(release -> address, BUFFER_SIZE, MEM_DECOMMIT);
+  VirtualFree(alloc_info, allocation_block_size, MEM_DECOMMIT);
 
 }
 
@@ -703,17 +803,13 @@ static void *alloc_windows(void *address){
   void *map_address;
 
   map_address  = VirtualAlloc(address,
-			      BUFFER_SIZE,
+			      allocation_block_size,
 			      MEM_RESERVE | MEM_COMMIT,
 			      PAGE_READWRITE);
 
   if (map_address == (void *)NULL) map_address = (void *)-1;
 
-  if (map_address != (void *)-1) {
+  STORE_RELEASE_FUNC(map_address, alloc_windows_free);
-    release_info[release_pos].address = map_address;
-    release_info[release_pos].func    = alloc_windows_free;
-    release_pos ++;
-  }
 
   return map_address;
 }
@@ -725,13 +821,14 @@ static void *alloc_windows(void *address){
 #define DEVICEDRIVER_NAME "/dev/mapper"
 #endif
 
-static void alloc_devicedirver_free(struct release_t *release){
+static void alloc_devicedirver_free(struct alloc_t *alloc_info){
 
-  if (munmap(release -> address, BUFFER_SIZE)) {
+  int attr = alloc_info -> attr;
+  if (munmap(address, allocation_block_size)) {
     printf("OpenBLAS : Bugphysarea unmap failed.\n");
   }
 
-  if (close(release -> attr)) {
+  if (close(attr)) {
     printf("OpenBLAS : Bugphysarea close failed.\n");
   }
 
@@ -748,17 +845,12 @@ static void *alloc_devicedirver(void *address){
 
   }
 
-  map_address = mmap(address, BUFFER_SIZE,
+  map_address = mmap(address, allocation_block_size,
 		     PROT_READ | PROT_WRITE,
 		     MAP_FILE | MAP_SHARED,
 		     fd, 0);
 
-  if (map_address != (void *)-1) {
+  STORE_RELEASE_FUNC_WITH_ATTR(map_address, alloc_devicedirver_free, fd);
-    release_info[release_pos].address = map_address;
-    release_info[release_pos].attr    = fd;
-    release_info[release_pos].func    = alloc_devicedirver_free;
-    release_pos ++;
-  }
 
   return map_address;
 }
@@ -767,9 +859,9 @@ static void *alloc_devicedirver(void *address){
 
 #ifdef ALLOC_SHM
 
-static void alloc_shm_free(struct release_t *release){
+static void alloc_shm_free(struct alloc_t *alloc_info){
 
-  if (shmdt(release -> address)) {
+  if (shmdt(alloc_info)) {
     printf("OpenBLAS : Shared memory unmap failed.\n");
     }
 }
@@ -778,22 +870,21 @@ static void *alloc_shm(void *address){
   void *map_address;
   int shmid;
 
-  shmid = shmget(IPC_PRIVATE, BUFFER_SIZE,IPC_CREAT | 0600);
+  shmid = shmget(IPC_PRIVATE, allocation_block_size,IPC_CREAT | 0600);
 
   map_address = (void *)shmat(shmid, address, 0);
 
   if (map_address != (void *)-1){
 
 #ifdef OS_LINUX
-    my_mbind(map_address, BUFFER_SIZE, MPOL_PREFERRED, NULL, 0, 0);
+    my_mbind(map_address, allocation_block_size, MPOL_PREFERRED, NULL, 0, 0);
 #endif
 
     shmctl(shmid, IPC_RMID, 0);
 
-    release_info[release_pos].address = map_address;
+    struct alloc_t *alloc_info = (struct alloc_t *)map_address;
-    release_info[release_pos].attr    = shmid;
+    alloc_info->release_func = alloc_shm_free;
-    release_info[release_pos].func    = alloc_shm_free;
+    alloc_info->attr = shmid;
-    release_pos ++;
   }
 
   return map_address;
@@ -801,23 +892,23 @@ static void *alloc_shm(void *address){
 
 #if defined OS_LINUX  || defined OS_AIX  || defined __sun__  || defined OS_WINDOWS
 
-static void alloc_hugetlb_free(struct release_t *release){
+static void alloc_hugetlb_free(struct alloc_t *alloc_info){
 
 #if defined(OS_LINUX) || defined(OS_AIX)
-  if (shmdt(release -> address)) {
+  if (shmdt(alloc_info)) {
     printf("OpenBLAS : Hugepage unmap failed.\n");
   }
 #endif
 
 #ifdef __sun__
 
-  munmap(release -> address, BUFFER_SIZE);
+  munmap(alloc_info, allocation_block_size);
 
 #endif
 
 #ifdef OS_WINDOWS
 
-  VirtualFree(release -> address, BUFFER_SIZE, MEM_LARGE_PAGES | MEM_DECOMMIT);
+  VirtualFree(alloc_info, allocation_block_size, MEM_LARGE_PAGES | MEM_DECOMMIT);
 
 #endif
 
@@ -830,7 +921,7 @@ static void *alloc_hugetlb(void *address){
 #if defined(OS_LINUX) || defined(OS_AIX)
   int shmid;
 
-  shmid = shmget(IPC_PRIVATE, BUFFER_SIZE,
+  shmid = shmget(IPC_PRIVATE, allocation_block_size,
 #ifdef OS_LINUX
 		 SHM_HUGETLB |
 #endif
@@ -843,7 +934,7 @@ static void *alloc_hugetlb(void *address){
     map_address = (void *)shmat(shmid, address, SHM_RND);
 
 #ifdef OS_LINUX
-    my_mbind(map_address, BUFFER_SIZE, MPOL_PREFERRED, NULL, 0, 0);
+    my_mbind(map_address, allocation_block_size, MPOL_PREFERRED, NULL, 0, 0);
 #endif
 
     if (map_address != (void *)-1){
@@ -860,7 +951,7 @@ static void *alloc_hugetlb(void *address){
   mha.mha_pagesize = HUGE_PAGESIZE;
   memcntl(NULL, 0, MC_HAT_ADVISE, (char *)&mha, 0, 0);
 
-  map_address = (BLASULONG)memalign(HUGE_PAGESIZE, BUFFER_SIZE);
+  map_address = (BLASULONG)memalign(HUGE_PAGESIZE, allocation_block_size);
 #endif
 
 #ifdef OS_WINDOWS
@@ -884,7 +975,7 @@ static void *alloc_hugetlb(void *address){
   }
 
   map_address  = (void *)VirtualAlloc(address,
-				      BUFFER_SIZE,
+				      allocation_block_size,
 				      MEM_LARGE_PAGES | MEM_RESERVE | MEM_COMMIT,
 				      PAGE_READWRITE);
 
@@ -895,11 +986,7 @@ static void *alloc_hugetlb(void *address){
 
 #endif
 
-  if (map_address != (void *)-1){
+  STORE_RELEASE_FUNC(map_address, alloc_hugetlb_free);
-    release_info[release_pos].address = map_address;
-    release_info[release_pos].func    = alloc_hugetlb_free;
-    release_pos ++;
-  }
 
   return map_address;
 }
@@ -911,13 +998,14 @@ static void *alloc_hugetlb(void *address){
 
 static int hugetlb_pid = 0;
 
-static void alloc_hugetlbfile_free(struct release_t *release){
+static void alloc_hugetlbfile_free(struct alloc_t *alloc_info){
 
-  if (munmap(release -> address, BUFFER_SIZE)) {
+  int attr = alloc_info -> attr;
+  if (munmap(alloc_info, allocation_block_size)) {
     printf("OpenBLAS : HugeTLBfs unmap failed.\n");
   }
 
-  if (close(release -> attr)) {
+  if (close(attr)) {
     printf("OpenBLAS : HugeTLBfs close failed.\n");
   }
 }
@@ -938,17 +1026,12 @@ static void *alloc_hugetlbfile(void *address){
 
   unlink(filename);
 
-  map_address = mmap(address, BUFFER_SIZE,
+  map_address = mmap(address, allocation_block_size,
 		     PROT_READ | PROT_WRITE,
 		     MAP_SHARED,
 		     fd, 0);
 
-  if (map_address != (void *)-1) {
+  STORE_RELEASE_FUNC_WITH_ATTR(map_address, alloc_hugetlbfile_free, fd);
-    release_info[release_pos].address = map_address;
-    release_info[release_pos].attr    = fd;
-    release_info[release_pos].func    = alloc_hugetlbfile_free;
-    release_pos ++;
-  }
 
   return map_address;
 }
@@ -961,35 +1044,35 @@ static BLASULONG base_address      = 0UL;
 static BLASULONG base_address      = BASE_ADDRESS;
 #endif
 
-static volatile struct {
+#if __STDC_VERSION__ >= 201112L
-  BLASULONG lock;
+static _Atomic int memory_initialized = 0;
-  void *addr;
-#if defined(WHEREAMI) && !defined(USE_OPENMP)
-  int   pos;
-#endif
-  int used;
-#ifndef __64BIT__
-  char dummy[48];
 #else
-  char dummy[40];
+static volatile int memory_initialized = 0;
 #endif
 
-} memory[NUM_BUFFERS];
-
-static int memory_initialized = 0;
-
 /*       Memory allocation routine           */
 /* procpos ... indicates where it comes from */
 /*                0 : Level 3 functions      */
 /*                1 : Level 2 functions      */
 /*                2 : Thread                 */
 
+static void blas_memory_init(){
+#if defined(SMP) && !defined(USE_OPENMP)
+  next_memory_table_pos = 0;
+#  if !defined(HAS_COMPILER_TLS)
+#    if defined(OS_WINDOWS)
+  local_storage_key = ::TlsAlloc();
+#    else
+  pthread_key_create(&local_storage_key, NULL);
+#    endif /* defined(OS_WINDOWS) */
+#  endif /* defined(HAS_COMPILER_TLS) */
+#endif /* defined(SMP) && !defined(USE_OPENMP) */
+  memset(local_memory_table, 0, sizeof(local_memory_table));
+}
+
 void *blas_memory_alloc(int procpos){
 
   int position;
-#if defined(WHEREAMI) && !defined(USE_OPENMP)
-  int mypos;
-#endif
 
   void *map_address;
 
@@ -1019,24 +1102,17 @@ void *blas_memory_alloc(int procpos){
     NULL,
   };
   void *(**func)(void *address);
+  struct alloc_t * alloc_info;
+  struct alloc_t ** alloc_table;
 
-#if defined(USE_OPENMP)
+  if (!LIKELY_ONE(memory_initialized)) {
-  if (!memory_initialized) {
+#if defined(SMP) && !defined(USE_OPENMP)
-#endif
+    /* Only allow a single thread to initialize memory system */
-
     LOCK_COMMAND(&alloc_lock);
 
     if (!memory_initialized) {
-
-#if defined(WHEREAMI) && !defined(USE_OPENMP)
-    for (position = 0; position < NUM_BUFFERS; position ++){
-      memory[position].addr   = (void *)0;
-      memory[position].pos    = -1;
-      memory[position].used   = 0;
-      memory[position].lock   = 0;
-    }
 #endif
-
+      blas_memory_init();
 #ifdef DYNAMIC_ARCH
       gotoblas_dynamic_init();
 #endif
@@ -1057,65 +1133,23 @@ void *blas_memory_alloc(int procpos){
 
       memory_initialized = 1;
 
+#if defined(SMP) && !defined(USE_OPENMP)
     }
     UNLOCK_COMMAND(&alloc_lock);
-#if defined(USE_OPENMP)
-  }
 #endif
+  }
 
 #ifdef DEBUG
   printf("Alloc Start ...\n");
-#endif
-
-#if defined(WHEREAMI) && !defined(USE_OPENMP)
-
-  mypos = WhereAmI();
-
-  position = mypos;
-  while (position >= NUM_BUFFERS) position >>= 1;
-
-  do {
-    if (!memory[position].used && (memory[position].pos == mypos)) {
-#if defined(SMP) && !defined(USE_OPENMP)
-      LOCK_COMMAND(&alloc_lock);
-#else      
-      blas_lock(&memory[position].lock);
-#endif
-      if (!memory[position].used) goto allocation;
-#if defined(SMP) && !defined(USE_OPENMP)
-      UNLOCK_COMMAND(&alloc_lock);
-#else
-      blas_unlock(&memory[position].lock);
-#endif      
-    }
-
-    position ++;
-
-  } while (position < NUM_BUFFERS);
-
-
 #endif
 
   position = 0;
-
+  alloc_table = get_memory_table();
   do {
-#if defined(SMP) && !defined(USE_OPENMP)
+      if (!alloc_table[position] || !alloc_table[position]->used) goto allocation;
-      LOCK_COMMAND(&alloc_lock);
-#else
-    if (!memory[position].used) { 
-      blas_lock(&memory[position].lock);
-#endif
-      if (!memory[position].used) goto allocation;
-#if defined(SMP) && !defined(USE_OPENMP)
-      UNLOCK_COMMAND(&alloc_lock);
-#else      
-      blas_unlock(&memory[position].lock);
-      }
-#endif
-
     position ++;
 
-  } while (position < NUM_BUFFERS);
+  } while (position < BUFFERS_PER_THREAD);
 
   goto error;
 
@@ -1125,14 +1159,8 @@ void *blas_memory_alloc(int procpos){
   printf("  Position -> %d\n", position);
 #endif
 
-  memory[position].used = 1;
+  alloc_info = alloc_table[position];
-#if defined(SMP) && !defined(USE_OPENMP)
+  if (!alloc_info) {
-  UNLOCK_COMMAND(&alloc_lock);
-#else
-  blas_unlock(&memory[position].lock);
-#endif
-
-  if (!memory[position].addr) {
     do {
 #ifdef DEBUG
       printf("Allocation Start : %lx\n", base_address);
@@ -1148,14 +1176,14 @@ void *blas_memory_alloc(int procpos){
 
 #ifdef ALLOC_DEVICEDRIVER
 	if ((*func ==  alloc_devicedirver) && (map_address == (void *)-1)) {
-	    fprintf(stderr, "OpenBLAS Warning ... Physically contigous allocation was failed.\n");
+	    fprintf(stderr, "OpenBLAS Warning ... Physically contiguous allocation failed.\n");
 	}
 #endif
 
 #ifdef ALLOC_HUGETLBFILE
 	if ((*func == alloc_hugetlbfile) && (map_address == (void *)-1)) {
 #ifndef OS_WINDOWS
-	    fprintf(stderr, "OpenBLAS Warning ... HugeTLB(File) allocation was failed.\n");
+	    fprintf(stderr, "OpenBLAS Warning ... HugeTLB(File) allocation failed.\n");
 #endif
 	}
 #endif
@@ -1172,89 +1200,44 @@ void *blas_memory_alloc(int procpos){
 #endif
       if (((BLASLONG) map_address) == -1) base_address = 0UL;
 
-      if (base_address) base_address += BUFFER_SIZE + FIXED_PAGESIZE;
+      if (base_address) base_address += allocation_block_size + FIXED_PAGESIZE;
 
     } while ((BLASLONG)map_address == -1);
 
-#if defined(SMP) && !defined(USE_OPENMP)
+    alloc_table[position] = alloc_info = map_address;
-    LOCK_COMMAND(&alloc_lock);
-#endif    
-    memory[position].addr = map_address;
-#if defined(SMP) && !defined(USE_OPENMP)
-    UNLOCK_COMMAND(&alloc_lock);
-#endif
 
 #ifdef DEBUG
-    printf("  Mapping Succeeded. %p(%d)\n", (void *)memory[position].addr, position);
+    printf("  Mapping Succeeded. %p(%d)\n", (void *)alloc_info, position);
 #endif
   }
 
-#if defined(WHEREAMI) && !defined(USE_OPENMP)
-
-  if (memory[position].pos == -1) memory[position].pos = mypos;
-
-#endif
-
-#ifdef DYNAMIC_ARCH
-
-  if (memory_initialized == 1) {
-
-    LOCK_COMMAND(&alloc_lock);
-
-    if (memory_initialized == 1) {
-
-      if (!gotoblas) gotoblas_dynamic_init();
-
-      memory_initialized = 2;
-    }
-
-    UNLOCK_COMMAND(&alloc_lock);
-
-  }
-#endif
-
-
 #ifdef DEBUG
-  printf("Mapped   : %p  %3d\n\n",
+  printf("Mapped   : %p  %3d\n\n", (void *)alloc_info, position);
-	  (void *)memory[position].addr, position);
 #endif
 
-  return (void *)memory[position].addr;
+  alloc_info->used = 1;
+
+  return (void *)(((char *)alloc_info) + sizeof(struct alloc_t));
 
  error:
-  printf("BLAS : Program is Terminated. Because you tried to allocate too many memory regions.\n");
+  printf("OpenBLAS : Program will terminate because you tried to allocate too many memory regions.\n");
 
   return NULL;
 }
 
-void blas_memory_free(void *free_area){
+void blas_memory_free(void *buffer){
-
+#ifdef DEBUG
   int position;
+  struct alloc_t ** alloc_table;
+#endif
+  /* Since we passed an offset pointer to the caller, get back to the actual allocation */
+  struct alloc_t *alloc_info = (void *)(((char *)buffer) - sizeof(struct alloc_t));
 
 #ifdef DEBUG
-  printf("Unmapped Start : %p ...\n", free_area);
+  printf("Unmapped Start : %p ...\n", alloc_info);
 #endif
 
-  position = 0;
+  alloc_info->used = 0;
-#if defined(SMP) && !defined(USE_OPENMP)
-  LOCK_COMMAND(&alloc_lock);
-#endif
-  while ((position < NUM_BUFFERS) && (memory[position].addr != free_area))
-    position++;
-
-  if (memory[position].addr != free_area) goto error;
-
-#ifdef DEBUG
-  printf("  Position : %d\n", position);
-#endif
-
-  // arm: ensure all writes are finished before other thread takes this memory
-  WMB;
-
-  memory[position].used = 0;
-#if defined(SMP) && !defined(USE_OPENMP)
-  UNLOCK_COMMAND(&alloc_lock);
-#endif
 
 #ifdef DEBUG
   printf("Unmap Succeeded.\n\n");
@@ -1262,15 +1245,13 @@ void blas_memory_free(void *free_area){
 
   return;
 
- error:
-  printf("BLAS : Bad memory unallocation! : %4d  %p\n", position,  free_area);
-
 #ifdef DEBUG
-  for (position = 0; position < NUM_BUFFERS; position++)
+  alloc_table = get_memory_table();
-    printf("%4ld  %p : %d\n", position, memory[position].addr, memory[position].used);
+  for (position = 0; position < BUFFERS_PER_THREAD; position++){
-#endif
+    if (alloc_table[position]) {
-#if defined(SMP) && !defined(USE_OPENMP)
+      printf("%4ld  %p : %d\n", position, alloc_table[position], alloc_table[position]->used);
-  UNLOCK_COMMAND(&alloc_lock);
+    }
+  }
 #endif
   return;
 }
@@ -1287,16 +1268,20 @@ void blas_memory_free_nolock(void * map_address) {
 
 void blas_shutdown(void){
 
-  int pos;
+  int pos, thread;
 
 #ifdef SMP
   BLASFUNC(blas_thread_shutdown)();
 #endif
 
-  LOCK_COMMAND(&alloc_lock);
+  for (thread = 0; thread < MAX_ALLOCATING_THREADS; thread ++){
-
+    for (pos = 0; pos < BUFFERS_PER_THREAD; pos ++){
-  for (pos = 0; pos < release_pos; pos ++) {
+      struct alloc_t *alloc_info = local_memory_table[thread][pos];
-    release_info[pos].func(&release_info[pos]);
+      if (alloc_info) {
+        alloc_info->release_func(alloc_info);
+        alloc_info = (void *)0;
+      }
+    }
   }
 
 #ifdef SEEK_ADDRESS
@@ -1305,17 +1290,6 @@ void blas_shutdown(void){
   base_address      = BASE_ADDRESS;
 #endif
 
-  for (pos = 0; pos < NUM_BUFFERS; pos ++){
-    memory[pos].addr   = (void *)0;
-    memory[pos].used   = 0;
-#if defined(WHEREAMI) && !defined(USE_OPENMP)
-    memory[pos].pos    = -1;
-#endif
-    memory[pos].lock   = 0;
-  }
-
-  UNLOCK_COMMAND(&alloc_lock);
-
   return;
 }
 
@@ -1339,7 +1313,7 @@ static void _touch_memory(blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n,
   size_t size;
   BLASULONG buffer;
 
-  size   = BUFFER_SIZE - PAGESIZE;
+  size   = allocation_block_size - PAGESIZE;
   buffer = (BLASULONG)sa + GEMM_OFFSET_A;
 
 #if defined(OS_LINUX) && !defined(NO_WARMUP)
@@ -1360,7 +1334,7 @@ static void _touch_memory(blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n,
   UNLOCK_COMMAND(&init_lock);
 #endif
 
-  size = MIN((BUFFER_SIZE - PAGESIZE), L2_SIZE);
+  size = MIN((allocation_block_size - PAGESIZE), L2_SIZE);
   buffer = (BLASULONG)sa + GEMM_OFFSET_A;
 
   while (size > 0) {

driver/others/parameter.c

@@ -167,7 +167,7 @@ int get_L2_size(void){
 #if defined(ATHLON) || defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER) || \
     defined(CORE_PRESCOTT) || defined(CORE_CORE2) || defined(PENRYN) || defined(DUNNINGTON) || \
     defined(CORE_NEHALEM) || defined(CORE_SANDYBRIDGE) || defined(ATOM) || defined(GENERIC) || \
-    defined(PILEDRIVER) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR) || defined(ZEN)
+    defined(PILEDRIVER) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR) || defined(ZEN) || defined(SKYLAKEX)
 
   cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
 
@@ -251,7 +251,7 @@ int get_L2_size(void){
 void blas_set_parameter(void){
 
   int factor;
-#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(SANDYBRIDGE) || defined(NEHALEM) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR) || defined(ZEN)
+#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(SANDYBRIDGE) || defined(NEHALEM) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR) || defined(ZEN) || defined(SKYLAKEX)
   int size = 16;
 #else
   int size = get_L2_size();

exports/Makefile

@@ -128,6 +128,8 @@ so : ../$(LIBSONAME)
 
 ifeq ($(OSNAME), Android)
 INTERNALNAME = $(LIBPREFIX).so
+FEXTRALIB += -lm
+EXTRALIB += -lm
 else
 INTERNALNAME = $(LIBPREFIX).so.$(MAJOR_VERSION)
 endif

getarch.c

@@ -326,6 +326,21 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CORENAME  "HASWELL"
 #endif
 
+#ifdef FORCE_SKYLAKEX
+#define FORCE
+#define FORCE_INTEL
+#define ARCHITECTURE    "X86"
+#define SUBARCHITECTURE "SKYLAKEX"
+#define ARCHCONFIG   "-DSKYLAKEX " \
+		     "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
+		     "-DL2_SIZE=262144 -DL2_LINESIZE=64 " \
+		     "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
+		     "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3 -DHAVE_SSE4_1 -DHAVE_SSE4_2 -DHAVE_AVX " \
+                     "-DFMA3 -DHAVE_AVX512VL -march=skylake-avx512"
+#define LIBNAME   "skylakex"
+#define CORENAME  "SKYLAKEX"
+#endif
+
 #ifdef FORCE_ATOM
 #define FORCE
 #define FORCE_INTEL
@@ -1181,9 +1196,7 @@ int main(int argc, char *argv[]){
 #elif NO_PARALLEL_MAKE==1
     printf("MAKE += -j 1\n");
 #else
-#ifndef OS_WINDOWS
     printf("MAKE += -j %d\n", get_num_cores());
-#endif
 #endif
 
     break;

interface/Makefile

@@ -260,7 +260,7 @@ HPLOBJS = dgemm.$(SUFFIX) dtrsm.$(SUFFIX) \
 	  idamax.$(SUFFIX) daxpy.$(SUFFIX) dcopy.$(SUFFIX) dscal.$(SUFFIX)
 
 CSBLAS1OBJS   = \
-	cblas_isamax.$(SUFFIX) cblas_sasum.$(SUFFIX) cblas_saxpy.$(SUFFIX) \
+	cblas_isamax.$(SUFFIX) cblas_isamin.$(SUFFIX) cblas_sasum.$(SUFFIX) cblas_saxpy.$(SUFFIX) \
 	cblas_scopy.$(SUFFIX) cblas_sdot.$(SUFFIX) cblas_sdsdot.$(SUFFIX) cblas_dsdot.$(SUFFIX) \
 	cblas_srot.$(SUFFIX) cblas_srotg.$(SUFFIX) cblas_srotm.$(SUFFIX) cblas_srotmg.$(SUFFIX) \
 	cblas_sscal.$(SUFFIX) cblas_sswap.$(SUFFIX) cblas_snrm2.$(SUFFIX) cblas_saxpby.$(SUFFIX)
@@ -277,7 +277,7 @@ CSBLAS3OBJS   = \
 	cblas_sgeadd.$(SUFFIX)
 
 CDBLAS1OBJS   = \
-	cblas_idamax.$(SUFFIX) cblas_dasum.$(SUFFIX) cblas_daxpy.$(SUFFIX) \
+	cblas_idamax.$(SUFFIX) cblas_idamin.$(SUFFIX) cblas_dasum.$(SUFFIX) cblas_daxpy.$(SUFFIX) \
 	cblas_dcopy.$(SUFFIX) cblas_ddot.$(SUFFIX) \
 	cblas_drot.$(SUFFIX) cblas_drotg.$(SUFFIX) cblas_drotm.$(SUFFIX) cblas_drotmg.$(SUFFIX) \
 	cblas_dscal.$(SUFFIX) cblas_dswap.$(SUFFIX) cblas_dnrm2.$(SUFFIX) cblas_daxpby.$(SUFFIX)
@@ -294,7 +294,7 @@ CDBLAS3OBJS   += \
         cblas_dgeadd.$(SUFFIX) 
 
 CCBLAS1OBJS   = \
-	cblas_icamax.$(SUFFIX) cblas_scasum.$(SUFFIX)  cblas_caxpy.$(SUFFIX) \
+	cblas_icamax.$(SUFFIX) cblas_icamin.$(SUFFIX) cblas_scasum.$(SUFFIX)  cblas_caxpy.$(SUFFIX) \
 	cblas_ccopy.$(SUFFIX) \
 	cblas_cdotc.$(SUFFIX) cblas_cdotu.$(SUFFIX) \
 	cblas_cdotc_sub.$(SUFFIX) cblas_cdotu_sub.$(SUFFIX) \
@@ -320,7 +320,7 @@ CCBLAS3OBJS   = \
 
 
 CZBLAS1OBJS   = \
-	cblas_izamax.$(SUFFIX) cblas_dzasum.$(SUFFIX)  cblas_zaxpy.$(SUFFIX) \
+	cblas_izamax.$(SUFFIX) cblas_izamin.$(SUFFIX) cblas_dzasum.$(SUFFIX)  cblas_zaxpy.$(SUFFIX) \
 	cblas_zcopy.$(SUFFIX) \
 	cblas_zdotc.$(SUFFIX) cblas_zdotu.$(SUFFIX) \
 	cblas_zdotc_sub.$(SUFFIX) cblas_zdotu_sub.$(SUFFIX) \
@@ -1359,6 +1359,18 @@ cblas_icamax.$(SUFFIX) cblas_icamax.$(PSUFFIX) : imax.c
 cblas_izamax.$(SUFFIX) cblas_izamax.$(PSUFFIX) : imax.c
 	$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -UUSE_MIN $< -o $(@F)
 
+cblas_isamin.$(SUFFIX) cblas_isamin.$(PSUFFIX) : imax.c
+	$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -DUSE_MIN $< -o $(@F)
+
+cblas_idamin.$(SUFFIX) cblas_idamin.$(PSUFFIX) : imax.c
+	$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -DUSE_MIN $< -o $(@F)
+
+cblas_icamin.$(SUFFIX) cblas_icamin.$(PSUFFIX) : imax.c
+	$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -DUSE_MIN $< -o $(@F)
+
+cblas_izamin.$(SUFFIX) cblas_izamin.$(PSUFFIX) : imax.c
+	$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -DUSE_MIN $< -o $(@F)
+
 cblas_ismax.$(SUFFIX) cblas_ismax.$(PSUFFIX) : imax.c
 	$(CC) $(CFLAGS) -DCBLAS -c -UUSE_ABS -UUSE_MIN $< -o $(@F)
 

interface/axpy.c

@@ -83,17 +83,15 @@ void CNAME(blasint n, FLOAT alpha, FLOAT *x, blasint incx, FLOAT *y, blasint inc
   if (incy < 0) y -= (n - 1) * incy;
 
 #ifdef SMP
-  nthreads = num_cpu_avail(1);
-
   //disable multi-thread when incx==0 or incy==0
   //In that case, the threads would be dependent.
-  if (incx == 0 || incy == 0)
+  //
-	  nthreads = 1;
-
   //Temporarily work-around the low performance issue with small imput size &
   //multithreads.
-  if (n <= MULTI_THREAD_MINIMAL)
+  if (incx == 0 || incy == 0 || n <= MULTI_THREAD_MINIMAL)
 	  nthreads = 1;
+  else
+	  nthreads = num_cpu_avail(1);
 
   if (nthreads == 1) {
 #endif

interface/gemm.c

@@ -44,6 +44,7 @@
 #endif
 
 #ifndef COMPLEX
+#define SMP_THRESHOLD_MIN 65536.0
 #ifdef XDOUBLE
 #define ERROR_NAME "QGEMM "
 #elif defined(DOUBLE)
@@ -52,6 +53,7 @@
 #define ERROR_NAME "SGEMM "
 #endif
 #else
+#define SMP_THRESHOLD_MIN 8192.0
 #ifndef GEMM3M
 #ifdef XDOUBLE
 #define ERROR_NAME "XGEMM "
@@ -121,8 +123,6 @@ void NAME(char *TRANSA, char *TRANSB,
   FLOAT *sa, *sb;
 
 #ifdef SMP
-  int nthreads_max;
-  int nthreads_avail;
   double MNK;
 #ifndef COMPLEX
 #ifdef XDOUBLE
@@ -245,8 +245,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
   XFLOAT *sa, *sb;
 
 #ifdef SMP
-  int nthreads_max;
-  int nthreads_avail;
   double MNK;
 #ifndef COMPLEX
 #ifdef XDOUBLE
@@ -411,25 +409,12 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
   mode |= (transa << BLAS_TRANSA_SHIFT);
   mode |= (transb << BLAS_TRANSB_SHIFT);
 
-  nthreads_max = num_cpu_avail(3);
-  nthreads_avail = nthreads_max;
-
-#ifndef COMPLEX
   MNK = (double) args.m * (double) args.n * (double) args.k;
-  if ( MNK <= (65536.0  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
+  if ( MNK <= (SMP_THRESHOLD_MIN  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
-	nthreads_max = 1;
+	args.nthreads = 1;
-#else
-  MNK = (double) args.m * (double) args.n * (double) args.k;
-  if ( MNK <= (8192.0  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
-	nthreads_max = 1;
-#endif
-  args.common = NULL;
-
-  if ( nthreads_max > nthreads_avail )
-  	args.nthreads = nthreads_avail;
   else
-  	args.nthreads = nthreads_max;
+	args.nthreads = num_cpu_avail(3);
-
+  args.common = NULL;
 
  if (args.nthreads == 1) {
 #endif

interface/scal.c

@@ -76,10 +76,11 @@ void CNAME(blasint n, FLOAT alpha, FLOAT *x, blasint incx){
 
 
 #ifdef SMP
-  nthreads = num_cpu_avail(1);
-
   if (n <= 1048576 )
 	nthreads = 1;
+  else
+	nthreads = num_cpu_avail(1);
+
 
   if (nthreads == 1) {
 #endif

interface/trsm.c

@@ -366,12 +366,13 @@ void CNAME(enum CBLAS_ORDER order,
   mode |= (trans << BLAS_TRANSA_SHIFT);
   mode |= (side  << BLAS_RSIDE_SHIFT);
 
-  args.nthreads = num_cpu_avail(3);
   if ( args.m < 2*GEMM_MULTITHREAD_THRESHOLD )
 	args.nthreads = 1;
   else
 	if ( args.n < 2*GEMM_MULTITHREAD_THRESHOLD )
 		args.nthreads = 1;
+  else
+	args.nthreads = num_cpu_avail(3);
 		
 
   if (args.nthreads == 1) {

interface/zaxpy.c

@@ -41,7 +41,11 @@
 #ifdef FUNCTION_PROFILE
 #include "functable.h"
 #endif
-
+#if  defined(Z13)
+#define MULTI_THREAD_MINIMAL  200000
+#else
+#define MULTI_THREAD_MINIMAL  10000
+#endif
 #ifndef CBLAS
 
 void NAME(blasint *N, FLOAT *ALPHA, FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY){
@@ -69,7 +73,7 @@ void CNAME(blasint n, FLOAT *ALPHA, FLOAT *x, blasint incx, FLOAT *y, blasint in
 #endif
 
 #ifndef CBLAS
-  PRINT_DEBUG_CNAME;
+  PRINT_DEBUG_NAME;
 #else
   PRINT_DEBUG_CNAME;
 #endif
@@ -86,12 +90,15 @@ void CNAME(blasint n, FLOAT *ALPHA, FLOAT *x, blasint incx, FLOAT *y, blasint in
   if (incy < 0) y -= (n - 1) * incy * 2;
 
 #ifdef SMP
-  nthreads = num_cpu_avail(1);
-
   //disable multi-thread when incx==0 or incy==0
   //In that case, the threads would be dependent.
-  if (incx == 0 || incy == 0)
+  //
+  //Temporarily work-around the low performance issue with small imput size &
+  //multithreads.
+  if (incx == 0 || incy == 0 || n <= MULTI_THREAD_MINIMAL)
 	  nthreads = 1;
+  else
+	  nthreads = num_cpu_avail(1);
 
   if (nthreads == 1) {
 #endif

interface/zscal.c

@@ -90,10 +90,10 @@ void CNAME(blasint n, FLOAT alpha_r, void *vx, blasint incx){
   FUNCTION_PROFILE_START();
 
 #ifdef SMP
-  nthreads = num_cpu_avail(1);
-
   if ( n <= 1048576 )
 	nthreads = 1;
+  else
+	nthreads = num_cpu_avail(1);
 
   if (nthreads == 1) {
 #endif

interface/zswap.c

@@ -79,12 +79,12 @@ FLOAT *y = (FLOAT*)vy;
   if (incy < 0) y -= (n - 1) * incy * 2;
 
 #ifdef SMP
-  nthreads = num_cpu_avail(1);
-
   //disable multi-thread when incx==0 or incy==0
   //In that case, the threads would be dependent.
   if (incx == 0 || incy == 0)
 	  nthreads = 1;
+  else
+	  nthreads = num_cpu_avail(1);
 
   if (nthreads == 1) {
 #endif

kernel/CMakeLists.txt

@@ -121,7 +121,7 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
     # Makefile.L3
     set(USE_TRMM false)
 
-    if (ARM OR ARM64 OR "${TARGET_CORE}" STREQUAL "LONGSOON3B" OR "${TARGET_CORE}" STREQUAL "GENERIC" OR "${CORE}" STREQUAL "generic" OR "${TARGET_CORE}" STREQUAL "HASWELL" OR "${CORE}" STREQUAL "haswell" OR "${CORE}" STREQUAL "zen")
+    if (ARM OR ARM64 OR "${TARGET_CORE}" STREQUAL "LONGSOON3B" OR "${TARGET_CORE}" STREQUAL "GENERIC" OR "${CORE}" STREQUAL "generic" OR "${TARGET_CORE}" STREQUAL "HASWELL" OR "${CORE}" STREQUAL "haswell" OR "${CORE}" STREQUAL "zen" OR "${TARGET_CORE}" STREQUAL "SKYLAKEX" OR "${CORE}" STREQUAL "skylakex")
       set(USE_TRMM true)
     endif ()
 

kernel/Makefile.L3

@@ -32,6 +32,10 @@ ifeq ($(CORE), HASWELL)
 USE_TRMM = 1
 endif
 
+ifeq ($(CORE), SKYLAKEX)
+USE_TRMM = 1
+endif
+
 ifeq ($(CORE), ZEN)
 USE_TRMM = 1
 endif

kernel/arm/cdot_vfp.S

@@ -215,11 +215,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	N, #0
 	ble	cdot_kernel_L999
 
-	cmp	INC_X, #0
+#	cmp	INC_X, #0
-	beq	cdot_kernel_L999
+#	beq	cdot_kernel_L999
 
-	cmp	INC_Y, #0
+#	cmp	INC_Y, #0
-	beq	cdot_kernel_L999
+#	beq	cdot_kernel_L999
 
 	cmp	INC_X, #1
 	bne	cdot_kernel_S_BEGIN

kernel/arm/ddot_vfp.S

@@ -164,11 +164,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	N, #0
 	ble	ddot_kernel_L999
 
-	cmp	INC_X, #0
+#	cmp	INC_X, #0
-	beq	ddot_kernel_L999
+#	beq	ddot_kernel_L999
 
-	cmp	INC_Y, #0
+#	cmp	INC_Y, #0
-	beq	ddot_kernel_L999
+#	beq	ddot_kernel_L999
 
 	cmp	INC_X, #1
 	bne	ddot_kernel_S_BEGIN

kernel/arm/sdot_vfp.S

@@ -253,11 +253,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	N, #0
 	ble	sdot_kernel_L999
 
-	cmp	INC_X, #0
+#	cmp	INC_X, #0
-	beq	sdot_kernel_L999
+#	beq	sdot_kernel_L999
 
-	cmp	INC_Y, #0
+#	cmp	INC_Y, #0
-	beq	sdot_kernel_L999
+#	beq	sdot_kernel_L999
 
 	cmp	INC_X, #1
 	bne	sdot_kernel_S_BEGIN

kernel/arm/zdot_vfp.S

@@ -218,11 +218,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	N, #0
 	ble	zdot_kernel_L999
 
-	cmp	INC_X, #0
+#	cmp	INC_X, #0
-	beq	zdot_kernel_L999
+#	beq	zdot_kernel_L999
 
-	cmp	INC_Y, #0
+#	cmp	INC_Y, #0
-	beq	zdot_kernel_L999
+#	beq	zdot_kernel_L999
 
 	cmp	INC_X, #1
 	bne	zdot_kernel_S_BEGIN

kernel/arm64/casum_thunderx2t99.c

@@ -233,14 +233,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	FLOAT asum = 0.0;
 
 #if defined(SMP)
+	if (inc_x == 0 || n <= 10000)
+		nthreads = 1;
+	else
 		nthreads = num_cpu_avail(1);
 
-	if (inc_x == 0)
-		nthreads = 1;
-
-	if (n <= 10000)
-		nthreads = 1;
-
 	if (nthreads == 1) {
 		asum = casum_compute(n, x, inc_x);
 	} else {

kernel/arm64/copy_thunderx2t99.c

@@ -183,14 +183,11 @@ int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 	if (n <= 0) return 0;
 
 #if defined(SMP)
+	if (inc_x == 0 || n <= 10000)
+		nthreads = 1;
+	else
 		nthreads = num_cpu_avail(1);
 
-	if (inc_x == 0)
-		nthreads = 1;
-
-	if (n <= 10000)
-		nthreads = 1;
-
 	if (nthreads == 1) {
 		do_copy(n, x, inc_x, y, inc_y);
 	} else {

kernel/arm64/dasum_thunderx2t99.c

@@ -228,14 +228,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	FLOAT asum = 0.0;
 
 #if defined(SMP)
+	if (inc_x == 0 || n <= 10000)
+		nthreads = 1;
+	else
 		nthreads = num_cpu_avail(1);
 
-	if (inc_x == 0)
-		nthreads = 1;
-
-	if (n <= 10000)
-		nthreads = 1;
-
 	if (nthreads == 1) {
 		asum = dasum_compute(n, x, inc_x);
 	} else {

kernel/arm64/dot_thunderx2t99.c

@@ -384,14 +384,11 @@ RETURN_TYPE CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y
 	RETURN_TYPE dot = 0.0;
 
 #if defined(SMP)
+	if (inc_x == 0 || inc_y == 0 || n <= 10000)
+		nthreads = 1;
+	else
 		nthreads = num_cpu_avail(1);
 
-	if (inc_x == 0 || inc_y == 0)
-		nthreads = 1;
-
-	if (n <= 10000)
-		nthreads = 1;
-
 	if (nthreads == 1) {
 		dot = dot_compute(n, x, inc_x, y, inc_y);
 	} else {

kernel/arm64/dznrm2_thunderx2t99.c

@@ -328,10 +328,10 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	if (n <= 0 || inc_x <= 0) return 0.0;
 
 #if defined(SMP)
-	nthreads = num_cpu_avail(1);
-
 	if (n <= 10000)
 		nthreads = 1;
+	else
+		nthreads = num_cpu_avail(1);
 
 	if (nthreads == 1) {
 		nrm2_compute(n, x, inc_x, &ssq, &scale);

kernel/arm64/dznrm2_thunderx2t99_fast.c

@@ -235,10 +235,10 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	if (n <= 0 || inc_x <= 0) return 0.0;
 
 #if defined(SMP)
-	nthreads = num_cpu_avail(1);
-
 	if (n <= 10000)
 		nthreads = 1;
+	else
+		nthreads = num_cpu_avail(1);
 
 	if (nthreads == 1) {
 		nrm2 = nrm2_compute(n, x, inc_x);

kernel/arm64/iamax_thunderx2t99.c

@@ -321,14 +321,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG max_index = 0;
 
 #if defined(SMP)
+	if (inc_x == 0 || n <= 10000)
+		nthreads = 1;
+	else
 		nthreads = num_cpu_avail(1);
 
-	if (inc_x == 0)
-		nthreads = 1;
-
-	if (n <= 10000)
-		nthreads = 1;
-
 	if (nthreads == 1) {
 		max_index = iamax_compute(n, x, inc_x);
 	} else {

kernel/arm64/izamax_thunderx2t99.c

@@ -330,14 +330,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG max_index = 0;
 
 #if defined(SMP)
+	if (inc_x == 0 || n <= 10000)
+		nthreads = 1;
+	else
 		nthreads = num_cpu_avail(1);
 
-	if (inc_x == 0)
-		nthreads = 1;
-
-	if (n <= 10000)
-		nthreads = 1;
-
 	if (nthreads == 1) {
 		max_index = izamax_compute(n, x, inc_x);
 	} else {

kernel/arm64/sasum_thunderx2t99.c

@@ -230,14 +230,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	FLOAT asum = 0.0;
 
 #if defined(SMP)
+	if (inc_x == 0 || n <= 10000)
+		nthreads = 1;
+	else
 		nthreads = num_cpu_avail(1);
 
-	if (inc_x == 0)
-		nthreads = 1;
-
-	if (n <= 10000)
-		nthreads = 1;
-
 	if (nthreads == 1) {
 		asum = sasum_compute(n, x, inc_x);
 	} else {

kernel/arm64/scnrm2_thunderx2t99.c

@@ -318,10 +318,10 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	if (n <= 0 || inc_x <= 0) return 0.0;
 
 #if defined(SMP)
-	nthreads = num_cpu_avail(1);
-
 	if (n <= 10000)
 		nthreads = 1;
+	else
+		nthreads = num_cpu_avail(1);
 
 	if (nthreads == 1) {
 		nrm2_double = nrm2_compute(n, x, inc_x);

kernel/arm64/zasum_thunderx2t99.c

@@ -230,14 +230,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	FLOAT asum = 0.0;
 
 #if defined(SMP)
+	if (inc_x == 0 || n <= 10000)
+		nthreads = 1;
+	else
 		nthreads = num_cpu_avail(1);
 
-	if (inc_x == 0)
-		nthreads = 1;
-
-	if (n <= 10000)
-		nthreads = 1;
-
 	if (nthreads == 1) {
 		asum = zasum_compute(n, x, inc_x);
 	} else {

kernel/arm64/zdot_thunderx2t99.c

@@ -317,14 +317,11 @@ OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLA
        CIMAG(zdot) = 0.0;
 
 #if defined(SMP)
+	if (inc_x == 0 || inc_y == 0 || n <= 10000)
+		nthreads = 1;
+	else
 		nthreads = num_cpu_avail(1);
 
-	if (inc_x == 0 || inc_y == 0)
-		nthreads = 1;
-
-	if (n <= 10000)
-		nthreads = 1;
-
 	if (nthreads == 1) {
 		zdot_compute(n, x, inc_x, y, inc_y, &zdot);
 	} else {

kernel/power/KERNEL.POWER8

@@ -133,7 +133,7 @@ ZNRM2KERNEL  = ../arm/znrm2.c
 #
 SROTKERNEL   = srot.c
 DROTKERNEL   = drot.c
-#CROTKERNEL   = ../arm/zrot.c
+CROTKERNEL   = zrot.c
 ZROTKERNEL   = zrot.c
 #
 SSCALKERNEL  = sscal.c

kernel/setparam-ref.c

@@ -647,7 +647,9 @@ static int get_l2_size_old(void){
       return 6144;
     }
   }
-  return 0;
+//  return 0;
+fprintf (stderr,"OpenBLAS WARNING - could not determine the L2 cache size on this system, assuming 256k\n");
+return 256;
 }
 #endif
 
@@ -660,6 +662,10 @@ static __inline__ int get_l2_size(void){
   l2 = BITMASK(ecx, 16, 0xffff);
 
 #ifndef ARCH_X86
+  if (l2 <= 0) {
+     fprintf (stderr,"OpenBLAS WARNING - could not determine the L2 cache size on this system, assuming 256k\n");
+     return 256;
+  }
   return l2;
 
 #else
@@ -871,6 +877,22 @@ static void init_parameter(void) {
 #endif
 #endif
 
+#ifdef SKYLAKEX
+
+#ifdef DEBUG
+  fprintf(stderr, "SkylakeX\n");
+#endif
+
+  TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
+  TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
+  TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
+  TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
+#ifdef EXPRECISION
+  TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
+  TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
+#endif
+#endif
+
 
 #ifdef OPTERON
 

kernel/x86/KERNEL.NEHALEM

@@ -1,3 +1 @@
 include $(KERNELDIR)/KERNEL.PENRYN
-SSWAPKERNEL  = ../arm/swap.c
-DSWAPKERNEL  = ../arm/swap.c

kernel/x86/swap.S

@@ -138,6 +138,14 @@
 /* INCX != 1 or INCY != 1 */
 
 .L14:
+	cmpl	$0, %ebx
+	jne	.L141
+	cmpl	$0, %ecx
+	jne	.L141
+/* INCX == 0 and INCY == 0 */	
+	jmp	.L27
+
+.L141:	
 	movl	%edx, %eax
 	sarl	$2,   %eax
 	jle	.L28

kernel/x86/trsm_kernel_LN_2x4_penryn.S

@@ -62,7 +62,7 @@
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined (SKYLAKEX)
 #define PREFETCH     prefetcht0
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif

kernel/x86/trsm_kernel_LN_4x4_penryn.S

@@ -62,7 +62,7 @@
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined (SKYLAKEX)
 #define PREFETCH     prefetcht0
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif

kernel/x86/trsm_kernel_LT_2x4_penryn.S

@@ -62,7 +62,7 @@
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined (SKYLAKEX)
 #define PREFETCH     prefetcht0
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif

kernel/x86/trsm_kernel_LT_4x4_penryn.S

@@ -62,7 +62,7 @@
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined (SKYLAKEX)
 #define PREFETCH     prefetcht0
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif

kernel/x86/trsm_kernel_RT_2x4_penryn.S

@@ -62,7 +62,7 @@
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined (SKYLAKEX)
 #define PREFETCH     prefetcht0
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif

kernel/x86/trsm_kernel_RT_4x4_penryn.S

@@ -62,7 +62,7 @@
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined (SKYLAKEX)
 #define PREFETCH     prefetcht0
 #define PREFETCHSIZE  (8 * 21 + 4)
 #endif

kernel/x86/ztrsm_kernel_LN_2x2_penryn.S

@@ -61,7 +61,7 @@
 #define PREFETCHSIZE 84
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined (SKYLAKEX)
 #define PREFETCH	prefetcht1
 #define PREFETCHSIZE 84
 #endif

kernel/x86/ztrsm_kernel_LT_1x2_penryn.S

@@ -63,7 +63,7 @@
 #define PREFETCHSIZE 84
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined (SKYLAKEX)
 #define PREFETCH	prefetcht1
 #define PREFETCHSIZE 84
 #endif

kernel/x86/ztrsm_kernel_LT_2x2_penryn.S

@@ -61,7 +61,7 @@
 #define PREFETCHSIZE 84
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined (SKYLAKEX)
 #define PREFETCH	prefetcht1
 #define PREFETCHSIZE 84
 #endif

kernel/x86/ztrsm_kernel_RT_1x2_penryn.S

@@ -63,7 +63,7 @@
 #define PREFETCHSIZE 84
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined (SKYLAKEX)
 #define PREFETCH	prefetcht1
 #define PREFETCHSIZE 84
 #endif

kernel/x86/ztrsm_kernel_RT_2x2_penryn.S

@@ -61,7 +61,7 @@
 #define PREFETCHSIZE 84
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined (SKYLAKEX)
 #define PREFETCH	prefetcht1
 #define PREFETCHSIZE 84
 #endif

kernel/x86_64/KERNEL.SKYLAKEX

@@ -0,0 +1,19 @@
+include $(KERNELDIR)/KERNEL.HASWELL
+
+SGEMMKERNEL    =  sgemm_kernel_16x4_skylakex.S
+
+
+#DTRMMKERNEL    =  ../generic/trmmkernel_16x2.c
+#DGEMMKERNEL    =  dgemm_kernel_16x2_skylakex.S
+#DGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
+#DGEMMITCOPY    =  ../generic/gemm_tcopy_16.c
+#DGEMMONCOPY    =  ../generic/gemm_ncopy_2.c
+#DGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c
+#DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
+#DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+#DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+#DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+
+SGEMM_BETA = ../generic/gemm_beta.c
+DGEMM_BETA = ../generic/gemm_beta.c

kernel/x86_64/caxpy.c

@@ -33,7 +33,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "caxpy_microk_steamroller-2.c"
 #elif defined(BULLDOZER)
 #include "caxpy_microk_bulldozer-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined(SKYLAKEX)
 #include "caxpy_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "caxpy_microk_sandy-2.c"

kernel/x86_64/cdot.c

@@ -34,7 +34,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "cdot_microk_bulldozer-2.c"
 #elif defined(STEAMROLLER) || defined(PILEDRIVER)  || defined(EXCAVATOR)
 #include "cdot_microk_steamroller-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "cdot_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "cdot_microk_sandy-2.c"

kernel/x86_64/cgemv_n_4.c

@@ -29,7 +29,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include <stdio.h>
 #include "common.h"
 
-#if defined(HASWELL) || defined(ZEN)
+#if defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "cgemv_n_microk_haswell-4.c"
 #elif defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
 #include "cgemv_n_microk_bulldozer-4.c"

kernel/x86_64/cgemv_t_4.c

@@ -28,7 +28,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
-#if defined(HASWELL) || defined(ZEN)
+#if defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "cgemv_t_microk_haswell-4.c"
 #elif defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)  || defined(EXCAVATOR)
 #include "cgemv_t_microk_bulldozer-4.c"

kernel/x86_64/cscal.c

@@ -28,7 +28,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "common.h"
 
 
-#if defined(HASWELL) || defined(ZEN)
+#if defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "cscal_microk_haswell-2.c"
 #elif defined(BULLDOZER)  || defined(PILEDRIVER)
 #include "cscal_microk_bulldozer-2.c"

kernel/x86_64/daxpy.c

@@ -37,7 +37,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "daxpy_microk_steamroller-2.c"
 #elif defined(PILEDRIVER)
 #include "daxpy_microk_piledriver-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "daxpy_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "daxpy_microk_sandy-2.c"

kernel/x86_64/ddot.c

@@ -37,7 +37,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "ddot_microk_piledriver-2.c"
 #elif defined(NEHALEM)
 #include "ddot_microk_nehalem-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "ddot_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "ddot_microk_sandy-2.c"
@@ -169,14 +169,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 	FLOAT dot = 0.0;
 
 #if defined(SMP)
+	if (inc_x == 0 || inc_y == 0 || n <= 10000)
+		nthreads = 1;
+	else
 		nthreads = num_cpu_avail(1);
 
-	if (inc_x == 0 || inc_y == 0)
-		nthreads = 1;
-
-	if (n <= 10000)
-		nthreads = 1;
-
 	if (nthreads == 1) {
 		dot = dot_compute(n, x, inc_x, y, inc_y);
 	} else {

kernel/x86_64/dgemv_n_4.c

@@ -31,7 +31,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(NEHALEM)
 #include "dgemv_n_microk_nehalem-4.c"
-#elif defined(HASWELL) || defined(ZEN) || defined(STEAMROLLER) || defined(EXCAVATOR)
+#elif defined(HASWELL) || defined(ZEN) || defined(STEAMROLLER) || defined(EXCAVATOR) || defined (SKYLAKEX)
 #include "dgemv_n_microk_haswell-4.c"
 #endif
 

kernel/x86_64/dgemv_t_4.c

@@ -28,7 +28,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
-#if defined(HASWELL) || defined(ZEN) || defined(STEAMROLLER)  || defined(EXCAVATOR)
+#if defined(HASWELL) || defined(ZEN) || defined(STEAMROLLER)  || defined(EXCAVATOR) || defined (SKYLAKEX)
 #include "dgemv_t_microk_haswell-4.c"
 #endif
 

kernel/x86_64/dscal.c

@@ -31,7 +31,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "dscal_microk_bulldozer-2.c"
 #elif defined(SANDYBRIDGE)
 #include "dscal_microk_sandy-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "dscal_microk_haswell-2.c"
 #endif
 

kernel/x86_64/dsymv_L.c

@@ -30,7 +30,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
 #include "dsymv_L_microk_bulldozer-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "dsymv_L_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "dsymv_L_microk_sandy-2.c"

kernel/x86_64/dsymv_U.c

@@ -31,7 +31,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)  || defined(EXCAVATOR)
 #include "dsymv_U_microk_bulldozer-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "dsymv_U_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "dsymv_U_microk_sandy-2.c"

kernel/x86_64/saxpy.c

@@ -31,7 +31,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(NEHALEM)
 #include "saxpy_microk_nehalem-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "saxpy_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "saxpy_microk_sandy-2.c"

kernel/x86_64/sdot.c

@@ -34,7 +34,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "sdot_microk_steamroller-2.c"
 #elif defined(NEHALEM)
 #include "sdot_microk_nehalem-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "sdot_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "sdot_microk_sandy-2.c"

kernel/x86_64/sgemv_n_4.c

@@ -35,7 +35,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "sgemv_n_microk_nehalem-4.c"
 #elif defined(SANDYBRIDGE)
 #include "sgemv_n_microk_sandy-4.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "sgemv_n_microk_haswell-4.c"
 #endif
 

kernel/x86_64/sgemv_t_4.c

@@ -34,7 +34,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "sgemv_t_microk_bulldozer-4.c"
 #elif defined(SANDYBRIDGE)
 #include "sgemv_t_microk_sandy-4.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "sgemv_t_microk_haswell-4.c"
 #endif
 

kernel/x86_64/ssymv_L.c

@@ -32,7 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "ssymv_L_microk_bulldozer-2.c"
 #elif defined(NEHALEM)
 #include "ssymv_L_microk_nehalem-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "ssymv_L_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "ssymv_L_microk_sandy-2.c"

kernel/x86_64/ssymv_U.c

@@ -33,7 +33,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "ssymv_U_microk_bulldozer-2.c"
 #elif defined(NEHALEM)
 #include "ssymv_U_microk_nehalem-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "ssymv_U_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "ssymv_U_microk_sandy-2.c"

kernel/x86_64/symv_L_sse.S

@@ -57,7 +57,7 @@
 #define PREFETCHSIZE	(16 * 12)
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined(ZEN)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #define PREFETCH	prefetcht0
 #define PREFETCHW	prefetcht0
 #define PREFETCHSIZE	(16 * 12)

kernel/x86_64/symv_L_sse2.S

@@ -57,7 +57,7 @@
 #define PREFETCHSIZE	(16 * 12)
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined(ZEN)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #define PREFETCH	prefetcht0
 #define PREFETCHW	prefetcht0
 #define PREFETCHSIZE	(16 * 12)

kernel/x86_64/symv_U_sse.S

@@ -57,7 +57,7 @@
 #define PREFETCHSIZE	(16 * 12)
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE)  || defined(HASWELL) || defined(ZEN)
+#if defined(NEHALEM) || defined(SANDYBRIDGE)  || defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #define PREFETCH	prefetcht0
 #define PREFETCHW	prefetcht0
 #define PREFETCHSIZE	(16 * 12)

kernel/x86_64/symv_U_sse2.S

@@ -57,7 +57,7 @@
 #define PREFETCHSIZE	(16 * 12)
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE)  || defined(HASWELL) || defined(ZEN)
+#if defined(NEHALEM) || defined(SANDYBRIDGE)  || defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #define PREFETCH	prefetcht0
 #define PREFETCHW	prefetcht0
 #define PREFETCHSIZE	(16 * 24)

kernel/x86_64/zaxpy.c

@@ -33,7 +33,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "zaxpy_microk_bulldozer-2.c"
 #elif defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
 #include "zaxpy_microk_steamroller-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "zaxpy_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "zaxpy_microk_sandy-2.c"

kernel/x86_64/zdot.c

@@ -33,7 +33,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "zdot_microk_bulldozer-2.c"
 #elif defined(STEAMROLLER) || defined(PILEDRIVER) || defined(EXCAVATOR)
 #include "zdot_microk_steamroller-2.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "zdot_microk_haswell-2.c"
 #elif defined(SANDYBRIDGE)
 #include "zdot_microk_sandy-2.c"

kernel/x86_64/zgemv_n_4.c

@@ -30,7 +30,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "common.h"
 
 
-#if defined(HASWELL) || defined(ZEN)
+#if defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "zgemv_n_microk_haswell-4.c"
 #elif defined(SANDYBRIDGE)
 #include "zgemv_n_microk_sandy-4.c"

kernel/x86_64/zgemv_t_4.c

@@ -31,7 +31,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)  || defined(EXCAVATOR)
 #include "zgemv_t_microk_bulldozer-4.c"
-#elif defined(HASWELL) || defined(ZEN)
+#elif defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "zgemv_t_microk_haswell-4.c"
 #endif
 

kernel/x86_64/zscal.c

@@ -28,7 +28,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "common.h"
 
 
-#if defined(HASWELL) || defined(ZEN)
+#if defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #include "zscal_microk_haswell-2.c"
 #elif defined(BULLDOZER)  || defined(PILEDRIVER)
 #include "zscal_microk_bulldozer-2.c"

kernel/x86_64/zsymv_L_sse.S

@@ -57,7 +57,7 @@
 #define PREFETCHSIZE	(16 * 24)
 #endif
 
-#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined(ZEN)
+#if defined(NEHALEM) || defined(SANDYBRIDGE) || defined(HASWELL) || defined(ZEN) || defined (SKYLAKEX)
 #define PREFETCH	prefetcht0
 #define PREFETCHW	prefetcht0
 #define PREFETCHSIZE	(16 * 24)