Merge pull request #2100 from xianyi/develop

Merge develop in preparation of 0.3.6 release

.travis.yml

@@ -149,7 +149,7 @@ matrix:
 
     - &test-macos
       os: osx
-      osx_image: xcode8
+      osx_image: xcode10.1
       before_script:
         - COMMON_FLAGS="DYNAMIC_ARCH=1 TARGET=NEHALEM NUM_THREADS=32"
         - brew update
@@ -160,6 +160,7 @@ matrix:
         - BTYPE="BINARY=64 INTERFACE64=1"
 
     - <<: *test-macos
+      osx_image: xcode8.3
       env:
         - BTYPE="BINARY=32"
 

CMakeLists.txt

@@ -6,7 +6,7 @@ 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 5)
+set(OpenBLAS_PATCH_VERSION 6)
 set(OpenBLAS_VERSION "${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.${OpenBLAS_PATCH_VERSION}")
 
 # Adhere to GNU filesystem layout conventions
@@ -42,6 +42,19 @@ endif()
 
 #######
 
+if(MSVC AND MSVC_STATIC_CRT)
+    set(CompilerFlags
+            CMAKE_CXX_FLAGS
+            CMAKE_CXX_FLAGS_DEBUG
+            CMAKE_CXX_FLAGS_RELEASE
+            CMAKE_C_FLAGS
+            CMAKE_C_FLAGS_DEBUG
+            CMAKE_C_FLAGS_RELEASE
+            )
+    foreach(CompilerFlag ${CompilerFlags})
+      string(REPLACE "/MD" "/MT" ${CompilerFlag} "${${CompilerFlag}}")
+    endforeach()
+endif()
 
 message(WARNING "CMake support is experimental. It does not yet support all build options and may not produce the same Makefiles that OpenBLAS ships with.")
 
@@ -62,10 +75,10 @@ endif ()
 
 set(SUBDIRS	${BLASDIRS})
 if (NOT NO_LAPACK)
-  list(APPEND SUBDIRS lapack)
   if(BUILD_RELAPACK)
     list(APPEND SUBDIRS relapack/src)
   endif()
+  list(APPEND SUBDIRS lapack)
 endif ()
 
 # set which float types we want to build for
@@ -134,7 +147,7 @@ endif ()
 
 # Only generate .def for dll on MSVC and always produce pdb files for debug and release
 if(MSVC)
-  if (${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION} LESS 3.4)
+  if (${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION} VERSION_LESS 3.4)
     set(OpenBLAS_DEF_FILE "${PROJECT_BINARY_DIR}/openblas.def")
   endif()
   set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /Zi")
@@ -149,15 +162,9 @@ if (${DYNAMIC_ARCH})
   endforeach()
 endif ()
 
-# Only build shared libs for MSVC
-if (MSVC)
-  set(BUILD_SHARED_LIBS ON)
-endif()
-
-
 # add objects to the openblas lib
 add_library(${OpenBLAS_LIBNAME} ${LA_SOURCES} ${LAPACKE_SOURCES} ${RELA_SOURCES} ${TARGET_OBJS} ${OpenBLAS_DEF_FILE})
-target_include_directories(${OpenBLAS_LIBNAME} INTERFACE $<INSTALL_INTERFACE:include>)
+target_include_directories(${OpenBLAS_LIBNAME} INTERFACE $<INSTALL_INTERFACE:include/openblas${SUFFIX64}>)
 
 # Android needs to explicitly link against libm
 if(ANDROID)
@@ -166,7 +173,7 @@ endif()
 
 # Handle MSVC exports
 if(MSVC AND BUILD_SHARED_LIBS)
-  if (${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION} LESS 3.4)
+  if (${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION} VERSION_LESS 3.4)
     include("${PROJECT_SOURCE_DIR}/cmake/export.cmake")
   else()
     # Creates verbose .def file (51KB vs 18KB)
@@ -217,6 +224,14 @@ set_target_properties(${OpenBLAS_LIBNAME} PROPERTIES
   SOVERSION ${OpenBLAS_MAJOR_VERSION}
 )
 
+if (BUILD_SHARED_LIBS AND BUILD_RELAPACK)
+  if (NOT MSVC)
+    target_link_libraries(${OpenBLAS_LIBNAME} "-Wl,-allow-multiple-definition")
+  else()
+    target_link_libraries(${OpenBLAS_LIBNAME} "/FORCE:MULTIPLE")
+  endif()
+endif()
+
 if (BUILD_SHARED_LIBS AND NOT ${SYMBOLPREFIX}${SYMBOLSUFIX} STREQUAL "")
 if (NOT DEFINED ARCH)
   set(ARCH_IN "x86_64")
@@ -314,7 +329,7 @@ install (FILES ${OPENBLAS_CONFIG_H} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
 if(NOT NOFORTRAN)
   message(STATUS "Generating f77blas.h in ${CMAKE_INSTALL_INCLUDEDIR}")
 
-  set(F77BLAS_H ${CMAKE_BINARY_DIR}/f77blas.h)
+  set(F77BLAS_H ${CMAKE_BINARY_DIR}/generated/f77blas.h)
   file(WRITE  ${F77BLAS_H} "#ifndef OPENBLAS_F77BLAS_H\n")
   file(APPEND ${F77BLAS_H} "#define OPENBLAS_F77BLAS_H\n")
   file(APPEND ${F77BLAS_H} "#include \"openblas_config.h\"\n")
@@ -327,10 +342,11 @@ endif()
 if(NOT NO_CBLAS)
 	message (STATUS "Generating cblas.h in ${CMAKE_INSTALL_INCLUDEDIR}")
 
+	set(CBLAS_H ${CMAKE_BINARY_DIR}/generated/cblas.h)
 	file(READ ${CMAKE_CURRENT_SOURCE_DIR}/cblas.h CBLAS_H_CONTENTS)
 	string(REPLACE "common" "openblas_config" CBLAS_H_CONTENTS_NEW "${CBLAS_H_CONTENTS}")
-	file(WRITE ${CMAKE_BINARY_DIR}/cblas.tmp "${CBLAS_H_CONTENTS_NEW}")
-	install (FILES ${CMAKE_BINARY_DIR}/cblas.tmp DESTINATION ${CMAKE_INSTALL_INCLUDEDIR} RENAME cblas.h)
+	file(WRITE ${CBLAS_H} "${CBLAS_H_CONTENTS_NEW}")
+	install (FILES ${CBLAS_H} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
 endif()
 
 if(NOT NO_LAPACKE)

Changelog.txt

@@ -1,4 +1,82 @@
 OpenBLAS ChangeLog
+====================================================================
+Version 0.3.6
+29-Apr-2019
+
+common:
+	* the build tools now check that a given cpu TARGET is actually valid
+	* the build-time check of system features (c_check) has been made
+  	  less dependent on particular perl features (this should mainly
+  	  benefit building on Windows)
+	* several problem with the ReLAPACK integration were fixed,
+	  including INTERFACE64 support and building a shared library
+	* building with CMAKE on BSD systems was improved
+	* a non-absolute SUM function was added based on the
+  	  existing optimized code for ASUM
+	* CBLAS interfaces to the IxMIN and IxMAX functions were added
+	* a name clash between LAPACKE and BOOST headers was resolved
+	* CMAKE builds with OpenMP failed to include the appropriate getrf_parallel
+	  kernels
+	* a crash on thread (key) deletion with the USE_TLS=1 memory management
+	  option was fixed
+	* restored several earlier fixes, in particular for OpenMP performance,
+  	  building on BSD, and calling fork on CYGWIN, which had inadvertently
+  	  been dropped in the 0.3.3 rewrite of the memory management code.
+
+x86_64:
+	* the AVX512 DGEMM kernel has been disabled again due to unsolved problems
+	* building with old versions of MSVC was fixed
+	* it is now possible to build a static library on Windows with CMAKE
+	* accessing environment variables on CYGWIN at run time was fixed
+	* the CMAKE build system now recognizes 32bit userspace on 64bit hardware
+	* Intel "Denverton" atom and Hygon "Dhyana" zen CPUs are now autodetected
+	* building for DYNAMIC_ARCH with a DYNAMIC_LIST of targets is now supported
+  	  with CMAKE as well
+	* building for DYNAMIC_ARCH with GENERIC as the default target is now supported
+	* a buffer overflow in the SSE GEMM kernel for Intel Nano targets was fixed
+	* assembly bugs involving undeclared modification of input operands were fixed
+  	  in the AXPY, DOT, GEMV, GER, SCAL, SYMV and TRSM microkernels for Nehalem,
+	  Sandybridge, Haswell, Bulldozer and Piledriver. These would typically cause
+	  test failures or segfaults when compiled with recent versions of gcc from 8 onward.
+	* a similar bug was fixed in the blas_quickdivide code used to split workloads
+	  in most functions
+	* a bug in the IxMIN implementation for the GENERIC target made it return the result of IxMAX
+	* fixed building on SkylakeX systems when either the compiler or the (emulated) operating
+	  environment does not support AVX512
+	* improved GEMM performance on ZEN targets
+
+x86:
+	* build failures caused by the recently added checks for AVX512 were fixed
+	* an inline assembly bug involving undeclared modification of an input argument was
+  	  fixed in the blas_quickdivide code used to split workloads in most functions
+	* a bug in the IMIN implementation for the GENERIC target made it return the result of IMAX
+
+MIPS32:
+	* a bug in the IMIN implementation made it return the result of IMAX
+
+POWER:
+	* single precision BLAS1/2 functions have received optimized POWER8 kernels
+	* POWER9 is now a separate target, with an optimized DGEMM/DTRMM kernel
+	* building on PPC970 systems under OSX Leopard or Tiger is now supported
+	* out-of-bounds memory accesses in the gemm_beta microkernels were fixed
+	* building a shared library on AIX is now supported for POWER6
+	* DYNAMIC_ARCH support has been added for POWER6 and newer
+
+ARMv7:
+	* corrected xDOT behaviour with zero INC_X or INC_Y
+	* a bug in the IMIN implementation made it return the result of IMAX
+
+ARMv8:
+	* added support for HiSilicon TSV110 cpus
+	* the CMAKE build system now recognizes 32bit userspace on 64bit hardware
+	* cross-compilation with CMAKE now works again
+	* a bug in the IMIN implementation made it return the result of IMAX
+	* ARMV8 builds with the BINARY=32 option are now automatically handled as ARMV7
+
+IBM Z:
+	* optimized microkernels for single precicion BLAS1/2 functions have been added
+	  for both Z13 and Z14
+
 ====================================================================
 Version 0.3.5
 31-Dec-2018

Makefile

@@ -96,7 +96,7 @@ endif
 	@echo
 
 shared :
-ifndef NO_SHARED
+ifneq ($(NO_SHARED), 1)
 ifeq ($(OSNAME), $(filter $(OSNAME),Linux SunOS Android Haiku))
 	@$(MAKE) -C exports so
 	@ln -fs $(LIBSONAME) $(LIBPREFIX).so

Makefile.arm64

@@ -38,3 +38,8 @@ ifeq ($(CORE), THUNDERX2T99)
 CCOMMON_OPT += -march=armv8.1-a -mtune=thunderx2t99
 FCOMMON_OPT += -march=armv8.1-a -mtune=thunderx2t99
 endif
+
+ifeq ($(CORE), TSV110)
+CCOMMON_OPT += -march=armv8.2-a -mtune=tsv110
+FCOMMON_OPT += -march=armv8.2-a -mtune=tsv110
+endif

Makefile.install

@@ -58,14 +58,14 @@ ifndef NO_LAPACKE
 endif
 
 #for install static library
-ifndef NO_STATIC
+ifneq ($(NO_STATIC),1)
 	@echo Copying the static library to $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)
 	@install -pm644 $(LIBNAME) "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)"
 	@cd "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)" ; \
 	ln -fs $(LIBNAME) $(LIBPREFIX).$(LIBSUFFIX)
 endif
 #for install shared library
-ifndef NO_SHARED
+ifneq ($(NO_SHARED),1)
 	@echo Copying the shared library to $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)
 ifeq ($(OSNAME), $(filter $(OSNAME),Linux SunOS Android Haiku))
 	@install -pm755 $(LIBSONAME) "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)"
@@ -106,14 +106,14 @@ ifndef NO_LAPACKE
 endif
 
 #for install static library
-ifndef NO_STATIC
+ifneq ($(NO_STATIC),1)
 	@echo Copying the static library to $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)
 	@installbsd -c -m 644 $(LIBNAME) "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)"
 	@cd "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)" ; \
 	ln -fs $(LIBNAME) $(LIBPREFIX).$(LIBSUFFIX)
 endif
 #for install shared library
-ifndef NO_SHARED
+ifneq ($(NO_SHARED),1)
 	@echo Copying the shared library to $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)
 	@installbsd -c -m 755 $(LIBSONAME) "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)"
 	@cd "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)" ; \
@@ -138,7 +138,7 @@ endif
 	@echo "SET(OpenBLAS_VERSION \"${VERSION}\")" > "$(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG)"
 	@echo "SET(OpenBLAS_INCLUDE_DIRS ${OPENBLAS_INCLUDE_DIR})" >> "$(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG)"
 
-ifndef NO_SHARED
+ifneq ($(NO_SHARED),1)
 #ifeq logical or
 ifeq ($(OSNAME), $(filter $(OSNAME),Linux FreeBSD NetBSD OpenBSD DragonFly))
 	@echo "SET(OpenBLAS_LIBRARIES ${OPENBLAS_LIBRARY_DIR}/$(LIBPREFIX).so)" >> "$(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG)"

Makefile.power

@@ -9,7 +9,15 @@ else
 USE_OPENMP = 1
 endif
 
-
+ifeq ($(CORE), POWER9)
+ifeq ($(USE_OPENMP), 1)
+COMMON_OPT += -Ofast -mcpu=power9 -mtune=power9 -mvsx -malign-power -DUSE_OPENMP -fno-fast-math -fopenmp
+FCOMMON_OPT += -O2 -frecursive -mcpu=power9 -mtune=power9 -malign-power -DUSE_OPENMP -fno-fast-math -fopenmp
+else
+COMMON_OPT += -Ofast -mcpu=power9 -mtune=power9 -mvsx -malign-power -fno-fast-math
+FCOMMON_OPT += -O2 -frecursive -mcpu=power9 -mtune=power9 -malign-power -fno-fast-math
+endif
+endif
 
 ifeq ($(CORE), POWER8)
 ifeq ($(USE_OPENMP), 1)

Makefile.rule

@@ -3,7 +3,7 @@
 #
 
 # This library's version
-VERSION = 0.3.5
+VERSION = 0.3.6
 
 # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a
 # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library
@@ -48,6 +48,8 @@ VERSION = 0.3.5
 # HOSTCC = gcc
 
 # If you need 32bit binary, define BINARY=32, otherwise define BINARY=64
+# Please note that AVX is not available on 32-bit.
+# Setting BINARY=32 disables AVX/AVX2/AVX-512.
 # BINARY=64
 
 # About threaded BLAS. It will be automatically detected if you don't
@@ -57,7 +59,7 @@ VERSION = 0.3.5
 # USE_THREAD = 0
 
 # If you're going to use this library with OpenMP, please comment it in.
-# This flag is always set for POWER8. Don't modify the flag 
+# This flag is always set for POWER8. Don't set USE_OPENMP = 0 if you're targeting POWER8.
 # USE_OPENMP = 1
 
 # The OpenMP scheduler to use - by default this is "static" and you
@@ -68,36 +70,45 @@ VERSION = 0.3.5
 # 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.
+# You can define the maximum number of threads. Basically it should be less
+# than or equal to the number of CPU threads. If you don't specify one, it's
+# automatically detected by the build system.
+# If SMT (aka. HT) is enabled on the system, it may or may not be beneficial to 
+# restrict NUM_THREADS to the number of physical cores. By default, the automatic 
+# detection includes logical CPUs, thus allowing the use of SMT.
+# Users may opt at runtime to use less than NUM_THREADS threads.
+#
+# Note for package maintainers: you can build OpenBLAS with a large NUM_THREADS
+# value (eg. 32-256) if you expect your users to use that many threads. Due to the way
+# some internal structures are allocated, using a large NUM_THREADS value has a RAM
+# footprint penalty, even if users reduce the actual number of threads at runtime.
 # NUM_THREADS = 24
 
 # If you have enabled USE_OPENMP and your application would call
-# OpenBLAS's calculation API from multi threads, please comment it in.
-# This flag defines how many instances of OpenBLAS's calculation API can 
-# actually run in parallel. If more threads call OpenBLAS's calculation API,
+# OpenBLAS's calculation API from multiple threads, please comment this in.
+# This flag defines how many instances of OpenBLAS's calculation API can actually
+# run in parallel. If more than NUM_PARALLEL threads call OpenBLAS's calculation API,
 # they need to wait for the preceding API calls to finish or risk data corruption.
 # NUM_PARALLEL = 2
 
-# if you don't need to install the static library, please comment it in.
+# If you don't need to install the static library, please comment this in.
 # NO_STATIC = 1
 
-# if you don't need generate the shared library, please comment it in.
+# If you don't need to generate the shared library, please comment this in.
 # NO_SHARED = 1
 
-# If you don't need CBLAS interface, please comment it in.
+# If you don't need the CBLAS interface, please comment this in.
 # NO_CBLAS = 1
 
-# If you only want CBLAS interface without installing Fortran compiler,
-# please comment it in.
+# If you only want the CBLAS interface without installing a Fortran compiler,
+# please comment this in.
 # ONLY_CBLAS = 1
 
-# If you don't need LAPACK, please comment it in.
-# If you set NO_LAPACK=1, the library automatically sets NO_LAPACKE=1.
+# If you don't need LAPACK, please comment this in.
+# If you set NO_LAPACK=1, the build system automatically sets NO_LAPACKE=1.
 # NO_LAPACK = 1
 
-# If you don't need LAPACKE (C Interface to LAPACK), please comment it in.
+# If you don't need LAPACKE (C Interface to LAPACK), please comment this in.
 # NO_LAPACKE = 1
 
 # Build LAPACK Deprecated functions since LAPACK 3.6.0
@@ -106,7 +117,7 @@ BUILD_LAPACK_DEPRECATED = 1
 # Build RecursiveLAPACK on top of LAPACK
 # BUILD_RELAPACK = 1
 
-# If you want to use legacy threaded Level 3 implementation.
+# If you want to use the legacy threaded Level 3 implementation.
 # USE_SIMPLE_THREADED_LEVEL3 = 1
 
 # If you want to use the new, still somewhat experimental code that uses
@@ -116,8 +127,8 @@ BUILD_LAPACK_DEPRECATED = 1
 # USE_TLS = 1
 
 # If you want to drive whole 64bit region by BLAS. Not all Fortran
-# compiler supports this. It's safe to keep comment it out if you
-# are not sure(equivalent to "-i8" option).
+# compilers support this. It's safe to keep this commented out if you
+# are not sure. (This is equivalent to the "-i8" ifort option).
 # INTERFACE64 = 1
 
 # Unfortunately most of kernel won't give us high quality buffer.
@@ -125,10 +136,18 @@ BUILD_LAPACK_DEPRECATED = 1
 # but it will consume time. If you don't like it, you can disable one.
 NO_WARMUP = 1
 
-# If you want to disable CPU/Memory affinity on Linux.
+# Comment this in if you want to disable OpenBLAS's CPU/Memory affinity handling.
+# This feature is only implemented on Linux, and is always disabled on other platforms.
+# Enabling affinity handling may improve performance, especially on NUMA systems, but 
+# it may conflict with certain applications that also try to manage affinity.
+# This conflict can result in threads of the application calling OpenBLAS ending up locked
+# to the same core(s) as OpenBLAS, possibly binding all threads to a single core.
+# For this reason, affinity handling is disabled by default. Can be safely enabled if nothing
+# else modifies affinity settings.
+# Note: enabling affinity has been known to cause problems with NumPy and R
 NO_AFFINITY = 1
 
-# if you are compiling for Linux and you have more than 16 numa nodes or more than 256 cpus
+# If you are compiling for Linux and you have more than 16 numa nodes or more than 256 cpus
 # BIGNUMA = 1
 
 # Don't use AVX kernel on Sandy Bridge. It is compatible with old compilers
@@ -180,7 +199,7 @@ NO_AFFINITY = 1
 # 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
+# If you need sanity check by comparing results to reference BLAS. It'll be very
 # slow (Not implemented yet).
 # SANITY_CHECK = 1
 

Makefile.system

@@ -65,6 +65,7 @@ endif
 
 ifdef TARGET
 GETARCH_FLAGS := -DFORCE_$(TARGET)
+GETARCH_FLAGS += -DUSER_TARGET
 endif
 
 # Force fallbacks for 32bit
@@ -94,6 +95,9 @@ endif
 ifeq ($(TARGET), ZEN)
 GETARCH_FLAGS := -DFORCE_BARCELONA
 endif
+ifeq ($(TARGET), ARMV8)
+GETARCH_FLAGS := -DFORCE_ARMV7
+endif
 endif
 
 
@@ -151,7 +155,8 @@ GETARCH_FLAGS	+= -DNO_AVX
 endif
 
 ifeq ($(BINARY), 32)
-GETARCH_FLAGS	+= -DNO_AVX
+GETARCH_FLAGS	+= -DNO_AVX -DNO_AVX2 -DNO_AVX512
+NO_AVX512 = 1
 endif
 
 ifeq ($(NO_AVX2), 1)
@@ -523,6 +528,12 @@ DYNAMIC_CORE += THUNDERX
 DYNAMIC_CORE += THUNDERX2T99
 endif
 
+ifeq ($(ARCH), power)
+DYNAMIC_CORE = POWER6
+DYNAMIC_CORE += POWER8
+DYNAMIC_CORE += POWER9
+endif
+
 # If DYNAMIC_CORE is not set, DYNAMIC_ARCH cannot do anything, so force it to empty
 ifndef DYNAMIC_CORE
 override DYNAMIC_ARCH=

Makefile.zarch

@@ -4,3 +4,7 @@ CCOMMON_OPT += -march=z13 -mzvector
 FCOMMON_OPT += -march=z13 -mzvector
 endif
 
+ifeq ($(CORE), Z14)
+CCOMMON_OPT += -march=z14 -mzvector
+FCOMMON_OPT += -march=z14 -mzvector
+endif

TargetList.txt

@@ -48,6 +48,7 @@ POWER5
 POWER6
 POWER7
 POWER8
+POWER9
 PPCG4
 PPC970
 PPC970MP
@@ -90,7 +91,9 @@ CORTEXA73
 FALKOR
 THUNDERX
 THUNDERX2T99
+TSV110
 
 9.System Z:
 ZARCH_GENERIC
 Z13
+Z14

appveyor.yml

@@ -53,9 +53,9 @@ before_build:
   - ps: if (-Not (Test-Path .\build)) { mkdir build }
   - cd build
   - if [%COMPILER%]==[cl] cmake -G "Visual Studio 15 2017 Win64" ..
-  - if [%WITH_FORTRAN%]==[no] cmake -G "Ninja" -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_C_COMPILER=clang-cl ..
+  - if [%WITH_FORTRAN%]==[no] cmake -G "Ninja" -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_C_COMPILER=clang-cl -DMSVC_STATIC_CRT=ON ..
   - if [%WITH_FORTRAN%]==[yes] cmake -G "Ninja" -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_C_COMPILER=clang-cl -DCMAKE_Fortran_COMPILER=flang -DBUILD_WITHOUT_LAPACK=no -DNOFORTRAN=0 ..
-  - if [%DYNAMIC_ARCH%]==[ON] cmake -DDYNAMIC_ARCH=ON ..
+  - if [%DYNAMIC_ARCH%]==[ON] cmake -DDYNAMIC_ARCH=ON -DDYNAMIC_LIST='CORE2;NEHALEM;SANDYBRIDGE;BULLDOZER;HASWELL' ..
 
 build_script:
   - cmake --build .

benchmark/scripts/R/deig.R

@@ -2,6 +2,8 @@
 
 argv <- commandArgs(trailingOnly = TRUE)
 
+if (!is.null(options("matprod")[[1]])) options(matprod = "blas")
+
 nfrom <- 128
 nto <- 2048
 nstep <- 128
@@ -19,7 +21,6 @@ if (length(argv) > 0) {
       loops <- as.numeric(argv[z])
     }
   }
-
 }
 
 p <- Sys.getenv("OPENBLAS_LOOPS")
@@ -27,29 +28,21 @@ if (p != "") {
   loops <- as.numeric(p)
 }
 
-
-cat(sprintf(
-  "From %.0f To %.0f Step=%.0f Loops=%.0f\n",
-  nfrom,
-  nto,
-  nstep,
-  loops
-))
+cat(sprintf("From %.0f To %.0f Step=%.0f Loops=%.0f\n", nfrom, nto, nstep, loops))
 cat(sprintf("      SIZE             Flops                   Time\n"))
 
 n <- nfrom
 while (n <= nto) {
-  A <- matrix(rnorm(n * n), ncol = n, nrow = n)
+  A <- matrix(rnorm(n * n), nrow = n)
   ev <- 0
   z <- system.time(for (l in 1:loops) {
     ev <- eigen(A)
   })
 
-  mflops <- (26.66 * n * n * n) * loops / (z[3] * 1.0e6)
+  mflops <- (26.66 * n * n * n) * loops / (z[3] * 1e+06)
 
   st <- sprintf("%.0fx%.0f :", n, n)
   cat(sprintf("%20s %10.2f MFlops %10.6f sec\n", st, mflops, z[3]))
 
   n <- n + nstep
-
 }

benchmark/scripts/R/dgemm.R

@@ -2,6 +2,8 @@
 
 argv <- commandArgs(trailingOnly = TRUE)
 
+if (!is.null(options("matprod")[[1]])) options(matprod = "blas")
+
 nfrom <- 128
 nto <- 2048
 nstep <- 128
@@ -19,7 +21,6 @@ if (length(argv) > 0) {
       loops <- as.numeric(argv[z])
     }
   }
-
 }
 
 p <- Sys.getenv("OPENBLAS_LOOPS")
@@ -27,26 +28,13 @@ if (p != "") {
   loops <- as.numeric(p)
 }
 
-
-cat(sprintf(
-  "From %.0f To %.0f Step=%.0f Loops=%.0f\n",
-  nfrom,
-  nto,
-  nstep,
-  loops
-))
+cat(sprintf("From %.0f To %.0f Step=%.0f Loops=%.0f\n", nfrom, nto, nstep, loops))
 cat(sprintf("      SIZE             Flops                   Time\n"))
 
 n <- nfrom
 while (n <= nto) {
-  A <- matrix(runif(n * n),
-              ncol = n,
-              nrow = n,
-              byrow = TRUE)
-  B <- matrix(runif(n * n),
-              ncol = n,
-              nrow = n,
-              byrow = TRUE)
+  A <- matrix(runif(n * n), nrow = n)
+  B <- matrix(runif(n * n), nrow = n)
   C <- 1
 
   z <- system.time(for (l in 1:loops) {
@@ -54,11 +42,10 @@ while (n <= nto) {
     l <- l + 1
   })
 
-  mflops <- (2.0 * n * n * n) * loops / (z[3] * 1.0e6)
+  mflops <- (2.0 * n * n * n) * loops / (z[3] * 1e+06)
 
   st <- sprintf("%.0fx%.0f :", n, n)
   cat(sprintf("%20s %10.2f MFlops %10.6f sec\n", st, mflops, z[3]))
 
   n <- n + nstep
-
 }

benchmark/scripts/R/dsolve.R

@@ -2,6 +2,8 @@
 
 argv <- commandArgs(trailingOnly = TRUE)
 
+if (!is.null(options("matprod")[[1]])) options(matprod = "blas")
+
 nfrom <- 128
 nto <- 2048
 nstep <- 128
@@ -19,7 +21,6 @@ if (length(argv) > 0) {
       loops <- as.numeric(argv[z])
     }
   }
-
 }
 
 p <- Sys.getenv("OPENBLAS_LOOPS")
@@ -27,31 +28,22 @@ if (p != "") {
   loops <- as.numeric(p)
 }
 
-
-cat(sprintf(
-  "From %.0f To %.0f Step=%.0f Loops=%.0f\n",
-  nfrom,
-  nto,
-  nstep,
-  loops
-))
+cat(sprintf("From %.0f To %.0f Step=%.0f Loops=%.0f\n", nfrom, nto, nstep, loops))
 cat(sprintf("      SIZE             Flops                   Time\n"))
 
 n <- nfrom
 while (n <= nto) {
-  A <- matrix(rnorm(n * n), ncol = n, nrow = n)
-  B <- matrix(rnorm(n * n), ncol = n, nrow = n)
+  A <- matrix(rnorm(n * n), nrow = n)
+  B <- matrix(rnorm(n * n), nrow = n)
 
   z <- system.time(for (l in 1:loops) {
     solve(A, B)
   })
 
-  mflops <-
-    (2.0 / 3.0 * n * n * n + 2.0 * n * n * n) * loops / (z[3] * 1.0e6)
+  mflops <- (8.0 / 3 * n * n * n) * loops / (z[3] * 1e+06)
 
   st <- sprintf("%.0fx%.0f :", n, n)
   cat(sprintf("%20s %10.2f MFlops %10.6f sec\n", st, mflops, z[3]))
 
   n <- n + nstep
-
 }

c_check

@@ -1,7 +1,7 @@
 #!/usr/bin/perl
 
-use File::Basename;
-use File::Temp qw(tempfile);
+#use File::Basename;
+# use File::Temp qw(tempfile);
 
 # Checking cross compile
 $hostos   = `uname -s | sed -e s/\-.*//`;    chop($hostos);
@@ -12,7 +12,7 @@ $hostarch = "arm64" if ($hostarch eq "aarch64");
 $hostarch = "power" if ($hostarch =~ /^(powerpc|ppc).*/);
 $hostarch = "zarch" if ($hostarch eq "s390x");
 
-$tmpf = new File::Temp( UNLINK => 1 );
+#$tmpf = new File::Temp( UNLINK => 1 );
 $binary = $ENV{"BINARY"};
 
 $makefile = shift(@ARGV);
@@ -31,12 +31,25 @@ if ($?) {
 
 $cross_suffix = "";
 
-if (dirname($compiler_name) ne ".") {
-    $cross_suffix .= dirname($compiler_name) . "/";
-}
+eval "use File::Basename";
+if ($@){ 
+    warn "could not load PERL module File::Basename, emulating its functionality";
+    my $dirnam = substr($compiler_name, 0, rindex($compiler_name, "/")-1 );
+    if ($dirnam ne ".") {
+	$cross_suffix .= $dirnam . "/";
+    }
+    my $basnam = substr($compiler_name, rindex($compiler_name,"/")+1, length($compiler_name)-rindex($compiler_name,"/")-1);
+	if ($basnam =~ /([^\s]*-)(.*)/) {
+	$cross_suffix .= $1;
+    }
+} else {
+    if (dirname($compiler_name) ne ".") {
+	$cross_suffix .= dirname($compiler_name) . "/";
+    }
 
-if (basename($compiler_name) =~ /([^\s]*-)(.*)/) {
-    $cross_suffix .= $1;
+    if (basename($compiler_name) =~ /([^\s]*-)(.*)/) {
+	$cross_suffix .= $1;
+    }
 }
 
 $compiler = "";
@@ -171,20 +184,26 @@ if ($?) {
 
 $have_msa = 0;
 if (($architecture eq "mips") || ($architecture eq "mips64")) {
-    $code = '"addvi.b $w0, $w1, 1"';
-    $msa_flags = "-mmsa -mfp64 -msched-weight -mload-store-pairs";
-    print $tmpf "#include <msa.h>\n\n";
-    print $tmpf "void main(void){ __asm__ volatile($code); }\n";
-
-    $args = "$msa_flags -o $tmpf.o -x c $tmpf";
-    my @cmd = ("$compiler_name $args");
-    system(@cmd) == 0;
-    if ($? != 0) {
-        $have_msa = 0;
+    eval "use File::Temp qw(tempfile)";
+    if ($@){ 
+	warn "could not load PERL module File::Temp, so could not check MSA capatibility";
     } else {
-        $have_msa = 1;
+	$tmpf = new File::Temp( UNLINK => 1 );
+	$code = '"addvi.b $w0, $w1, 1"';
+	$msa_flags = "-mmsa -mfp64 -msched-weight -mload-store-pairs";
+	print $tmpf "#include <msa.h>\n\n";
+	print $tmpf "void main(void){ __asm__ volatile($code); }\n";
+
+	$args = "$msa_flags -o $tmpf.o -x c $tmpf";
+	my @cmd = ("$compiler_name $args");
+	system(@cmd) == 0;
+	if ($? != 0) {
+	    $have_msa = 0;
+	} else {
+	    $have_msa = 1;
+	}
+	unlink("$tmpf.o");
     }
-    unlink("$tmpf.o");
 }
 
 $architecture = x86    if ($data =~ /ARCH_X86/);
@@ -204,17 +223,25 @@ $binformat    = bin64  if ($data =~ /BINARY_64/);
 
 $no_avx512= 0;
 if (($architecture eq "x86") || ($architecture eq "x86_64")) {
-    $code = '"vbroadcastss -4 * 4(%rsi), %zmm2"';
-    print $tmpf "#include <immintrin.h>\n\nint main(void){ __asm__ volatile($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) {
-	$no_avx512 = 1;
-    } else {
+    eval "use File::Temp qw(tempfile)";
+    if ($@){ 
+	warn "could not load PERL module File::Temp, so could not check compiler compatibility with AVX512";
 	$no_avx512 = 0;
+    } else {
+#	$tmpf = new File::Temp( UNLINK => 1 );
+	($fh,$tmpf) = tempfile( UNLINK => 1 );
+	$code = '"vbroadcastss -4 * 4(%rsi), %zmm2"';
+	print $tmpf "#include <immintrin.h>\n\nint main(void){ __asm__ volatile($code); }\n";
+	$args = " -march=skylake-avx512 -c -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");
     }
-    unlink("tmpf.o");
 }
 
 $data = `$compiler_name -S ctest1.c && grep globl ctest1.s | head -n 1 && rm -f ctest1.s`;

cblas.h

@@ -73,6 +73,11 @@ double cblas_dasum (OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPENBLAS
 float  cblas_scasum(OPENBLAS_CONST blasint n, OPENBLAS_CONST void  *x, OPENBLAS_CONST blasint incx);
 double cblas_dzasum(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);
 
+float  cblas_ssum (OPENBLAS_CONST blasint n, OPENBLAS_CONST float  *x, OPENBLAS_CONST blasint incx);
+double cblas_dsum (OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPENBLAS_CONST blasint incx);
+float  cblas_scsum(OPENBLAS_CONST blasint n, OPENBLAS_CONST void  *x, OPENBLAS_CONST blasint incx);
+double cblas_dzsum(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);
+
 float  cblas_snrm2 (OPENBLAS_CONST blasint N, OPENBLAS_CONST float  *X, OPENBLAS_CONST blasint incX);
 double cblas_dnrm2 (OPENBLAS_CONST blasint N, OPENBLAS_CONST double *X, OPENBLAS_CONST blasint incX);
 float  cblas_scnrm2(OPENBLAS_CONST blasint N, OPENBLAS_CONST void  *X, OPENBLAS_CONST blasint incX);
@@ -88,6 +93,16 @@ CBLAS_INDEX cblas_idamin(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPE
 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);
 
+CBLAS_INDEX cblas_ismax(OPENBLAS_CONST blasint n, OPENBLAS_CONST float  *x, OPENBLAS_CONST blasint incx);
+CBLAS_INDEX cblas_idmax(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPENBLAS_CONST blasint incx);
+CBLAS_INDEX cblas_icmax(OPENBLAS_CONST blasint n, OPENBLAS_CONST void  *x, OPENBLAS_CONST blasint incx);
+CBLAS_INDEX cblas_izmax(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);
+
+CBLAS_INDEX cblas_ismin(OPENBLAS_CONST blasint n, OPENBLAS_CONST float  *x, OPENBLAS_CONST blasint incx);
+CBLAS_INDEX cblas_idmin(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPENBLAS_CONST blasint incx);
+CBLAS_INDEX cblas_icmin(OPENBLAS_CONST blasint n, OPENBLAS_CONST void  *x, OPENBLAS_CONST blasint incx);
+CBLAS_INDEX cblas_izmin(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/arch.cmake

@@ -74,6 +74,9 @@ if (DYNAMIC_ARCH)
     if (NOT NO_AVX512)
       set(DYNAMIC_CORE ${DYNAMIC_CORE} SKYLAKEX)
     endif ()
+    if (DYNAMIC_LIST)
+	set(DYNAMIC_CORE PRESCOTT ${DYNAMIC_LIST})
+    endif ()
   endif ()
 
   if (NOT DYNAMIC_CORE)

cmake/kernel.cmake

@@ -107,6 +107,12 @@ macro(SetDefaultL1)
   set(DAXPBYKERNEL ../arm/axpby.c)
   set(CAXPBYKERNEL ../arm/zaxpby.c)
   set(ZAXPBYKERNEL ../arm/zaxpby.c)
+  set(SSUMKERNEL sum.S)
+  set(DSUMKERNEL sum.S)
+  set(CSUMKERNEL zsum.S)
+  set(ZSUMKERNEL zsum.S)
+  set(QSUMKERNEL sum.S)
+  set(XSUMKERNEL zsum.S)
 endmacro ()
 
 macro(SetDefaultL2)
@@ -162,4 +168,4 @@ macro(SetDefaultL3)
   set(DGEADD_KERNEL ../generic/geadd.c)
   set(CGEADD_KERNEL ../generic/zgeadd.c)
   set(ZGEADD_KERNEL ../generic/zgeadd.c)
-endmacro ()
+endmacro ()

cmake/os.cmake

@@ -8,6 +8,11 @@ if (${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
   set(NO_EXPRECISION 1)
 endif ()
 
+if (${CMAKE_SYSTEM_NAME} MATCHES "FreeBSD|OpenBSD|NetBSD|DragonFly")
+  set(EXTRALIB "${EXTRALIB} -lm")
+  set(NO_EXPRECISION 1)
+endif ()
+
 if (${CMAKE_SYSTEM_NAME} STREQUAL "AIX")
   set(EXTRALIB "${EXTRALIB} -lm")
 endif ()

cmake/prebuild.cmake

@@ -87,13 +87,18 @@ endif ()
 # Cannot run getarch on target if we are cross-compiling
 if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSSTORE"))
   # Write to config as getarch would
+  if (DEFINED TARGET_CORE)
+  set(TCORE ${TARGET_CORE})
+  else()
+  set(TCORE ${CORE})
+  endif()
 
   # TODO: Set up defines that getarch sets up based on every other target
   # Perhaps this should be inside a different file as it grows larger
   file(APPEND ${TARGET_CONF_TEMP}
-    "#define ${CORE}\n"
-    "#define CHAR_CORENAME \"${CORE}\"\n")
-  if ("${CORE}" STREQUAL "ARMV7")
+    "#define ${TCORE}\n"
+    "#define CHAR_CORENAME \"${TCORE}\"\n")
+  if ("${TCORE}" STREQUAL "ARMV7")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_DATA_SIZE\t65536\n"
       "#define L1_DATA_LINESIZE\t32\n"
@@ -108,7 +113,7 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
     set(SGEMM_UNROLL_N 4)
     set(DGEMM_UNROLL_M 4)
     set(DGEMM_UNROLL_N 4)
-  elseif ("${CORE}" STREQUAL "ARMV8")
+  elseif ("${TCORE}" STREQUAL "ARMV8")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_DATA_SIZE\t32768\n"
       "#define L1_DATA_LINESIZE\t64\n"
@@ -118,9 +123,16 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
       "#define DTB_SIZE\t4096\n"
       "#define L2_ASSOCIATIVE\t32\n"
       "#define ARMV8\n")
-    set(SGEMM_UNROLL_M 4)
+    set(SGEMM_UNROLL_M 16)
     set(SGEMM_UNROLL_N 4)
-  elseif ("${CORE}" STREQUAL "CORTEXA57" OR "${CORE}" STREQUAL "CORTEXA53")
+    set(DGEMM_UNROLL_M 8)
+    set(DGEMM_UNROLL_N 4)
+    set(CGEMM_UNROLL_M 8)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 4)
+    set(ZGEMM_UNROLL_N 4)
+    set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "CORTEXA57" OR "${TCORE}" STREQUAL "CORTEXA53")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_CODE_SIZE\t32768\n"
       "#define L1_CODE_LINESIZE\t64\n"
@@ -144,9 +156,10 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
     set(DGEMM_UNROLL_N 4)
     set(CGEMM_UNROLL_M 8)
     set(CGEMM_UNROLL_N 4)
-    set(ZGEMM_UNROLL_M 8)
+    set(ZGEMM_UNROLL_M 4)
     set(ZGEMM_UNROLL_N 4)
-  elseif ("${CORE}" STREQUAL "CORTEXA72" OR "${CORE}" STREQUAL "CORTEXA73")
+    set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "CORTEXA72" OR "${TCORE}" STREQUAL "CORTEXA73")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_CODE_SIZE\t49152\n"
       "#define L1_CODE_LINESIZE\t64\n"
@@ -170,9 +183,10 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
     set(DGEMM_UNROLL_N 4)
     set(CGEMM_UNROLL_M 8)
     set(CGEMM_UNROLL_N 4)
-    set(ZGEMM_UNROLL_M 8)
+    set(ZGEMM_UNROLL_M 4)
     set(ZGEMM_UNROLL_N 4)
-  elseif ("${CORE}" STREQUAL "FALKOR")
+    set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "FALKOR")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_CODE_SIZE\t65536\n"
       "#define L1_CODE_LINESIZE\t64\n"
@@ -196,9 +210,10 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
     set(DGEMM_UNROLL_N 4)
     set(CGEMM_UNROLL_M 8)
     set(CGEMM_UNROLL_N 4)
-    set(ZGEMM_UNROLL_M 8)
+    set(ZGEMM_UNROLL_M 4)
     set(ZGEMM_UNROLL_N 4)
-  elseif ("${CORE}" STREQUAL "THUNDERX)
+    set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "THUNDERX")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_CODE_SIZE\t32768\n"
       "#define L1_CODE_LINESIZE\t64\n"
@@ -224,7 +239,8 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
     set(CGEMM_UNROLL_N 2)
     set(ZGEMM_UNROLL_M 2)
     set(ZGEMM_UNROLL_N 2)
-  elseif ("${CORE}" STREQUAL "THUNDERX2T99)
+    set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "THUNDERX2T99")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_CODE_SIZE\t32768\n"
       "#define L1_CODE_LINESIZE\t64\n"
@@ -240,7 +256,7 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
       "#define L3_ASSOCIATIVE\t32\n"
       "#define DTB_DEFAULT_ENTRIES\t64\n"
       "#define DTB_SIZE\t4096\n"
-      "#define VULCAN\n")
+      "#define ARMV8\n")
     set(SGEMM_UNROLL_M 16)
     set(SGEMM_UNROLL_N 4)
     set(DGEMM_UNROLL_M 8)
@@ -249,6 +265,7 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
     set(CGEMM_UNROLL_N 4)
     set(ZGEMM_UNROLL_M 4)
     set(ZGEMM_UNROLL_N 4)
+    set(SYMV_P 16)
   endif()
 
   # Or should this actually be NUM_CORES?

cmake/system.cmake

@@ -39,6 +39,9 @@ if (DEFINED BINARY AND DEFINED TARGET AND BINARY EQUAL 32)
   if (${TARGET} STREQUAL "BULLDOZER" OR ${TARGET} STREQUAL "PILEDRIVER" OR ${TARGET} STREQUAL "ZEN")
     set(TARGET "BARCELONA")
   endif ()
+  if (${TARGET} STREQUAL "ARMV8" OR ${TARGET} STREQUAL "CORTEXA57" OR ${TARGET} STREQUAL "CORTEXA53")
+    set(TARGET "ARMV7")
+  endif ()
 endif ()
 
 if (DEFINED TARGET)
@@ -184,6 +187,13 @@ if (DYNAMIC_ARCH)
   endif ()
 endif ()
 
+if (DYNAMIC_LIST)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_LIST")
+  foreach(DCORE ${DYNAMIC_LIST})
+    set(CCOMMON_OPT "${CCOMMON_OPT} -DDYN_${DCORE}")
+  endforeach ()
+endif ()
+
 if (NO_LAPACK)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_LAPACK")
   #Disable LAPACK C interface

cmake/system_check.cmake

@@ -39,13 +39,21 @@ elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "ppc.*|power.*|Power.*")
 elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "mips64.*")
   set(MIPS64 1)
 elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "amd64.*|x86_64.*|AMD64.*")
-  set(X86_64 1)
+  if("${CMAKE_SIZEOF_VOID_P}" EQUAL "8")
+    set(X86_64 1)
+  else()
+    set(X86 1)
+  endif()
 elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "i686.*|i386.*|x86.*|amd64.*|AMD64.*")
   set(X86 1)
 elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(arm.*|ARM.*)")
   set(ARM 1)
 elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64.*|AARCH64.*)")
-  set(ARM64 1)
+  if("${CMAKE_SIZEOF_VOID_P}" EQUAL "8")
+    set(ARM64 1)
+  else()
+    set(ARM 1)
+  endif()
 endif()
 
 if (X86_64)
@@ -78,7 +86,7 @@ endif()
 
 if (X86_64 OR X86)
   file(WRITE ${PROJECT_BINARY_DIR}/avx512.tmp "#include <immintrin.h>\n\nint main(void){ __asm__ volatile(\"vbroadcastss -4 * 4(%rsi), %zmm2\"); }")
-execute_process(COMMAND ${CMAKE_C_COMPILER} -march=skylake-avx512 -v -o ${PROJECT_BINARY_DIR}/avx512.o -x c ${PROJECT_BINARY_DIR}/avx512.tmp OUTPUT_QUIET ERROR_QUIET RESULT_VARIABLE NO_AVX512)
+execute_process(COMMAND ${CMAKE_C_COMPILER} -march=skylake-avx512 -c -v -o ${PROJECT_BINARY_DIR}/avx512.o -x c ${PROJECT_BINARY_DIR}/avx512.tmp OUTPUT_QUIET ERROR_QUIET RESULT_VARIABLE NO_AVX512)
 if (NO_AVX512 EQUAL 1)
 set (CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX512")
 endif()

common.h

@@ -85,6 +85,8 @@ extern "C" {
 
 #if !defined(_MSC_VER)
 #include <unistd.h>
+#elif _MSC_VER < 1900
+#define snprintf _snprintf
 #endif
 #include <time.h>
 
@@ -348,6 +350,11 @@ typedef int blasint;
 #endif
 #endif
 
+#ifdef POWER9
+#ifndef YIELDING
+#define YIELDING        __asm__ __volatile__ ("nop;nop;nop;nop;nop;nop;nop;nop;\n");
+#endif
+#endif
 
 /*
 #ifdef PILEDRIVER
@@ -439,7 +446,7 @@ please https://github.com/xianyi/OpenBLAS/issues/246
 typedef char env_var_t[MAX_PATH];
 #define readenv(p, n) 0
 #else
-#ifdef OS_WINDOWS
+#if defined(OS_WINDOWS) && !defined(OS_CYGWIN_NT)
 typedef char env_var_t[MAX_PATH];
 #define readenv(p, n) GetEnvironmentVariable((LPCTSTR)(n), (LPTSTR)(p), sizeof(p))
 #else

common_c.h

@@ -19,6 +19,7 @@
 #define	CDOTC_K			cdotc_k
 #define	CNRM2_K			cnrm2_k
 #define	CSCAL_K			cscal_k
+#define	CSUM_K			csum_k
 #define	CSWAP_K			cswap_k
 #define	CROT_K			csrot_k
 
@@ -249,6 +250,7 @@
 #define	CDOTC_K			gotoblas -> cdotc_k
 #define	CNRM2_K			gotoblas -> cnrm2_k
 #define	CSCAL_K			gotoblas -> cscal_k
+#define	CSUM_K			gotoblas -> csum_k
 #define	CSWAP_K			gotoblas -> cswap_k
 #define	CROT_K			gotoblas -> csrot_k
 

common_d.h

@@ -19,6 +19,7 @@
 #define	DDOTC_K			ddot_k
 #define	DNRM2_K			dnrm2_k
 #define	DSCAL_K			dscal_k
+#define	DSUM_K			dsum_k
 #define	DSWAP_K			dswap_k
 #define	DROT_K			drot_k
 
@@ -174,6 +175,7 @@
 #define	DDOTC_K			gotoblas -> ddot_k
 #define	DNRM2_K			gotoblas -> dnrm2_k
 #define	DSCAL_K			gotoblas -> dscal_k
+#define	DSUM_K			gotoblas -> dsum_k
 #define	DSWAP_K			gotoblas -> dswap_k
 #define	DROT_K			gotoblas -> drot_k
 

common_interface.h

@@ -122,6 +122,13 @@ xdouble BLASFUNC(qasum) (blasint *, xdouble *, blasint *);
 double BLASFUNC(dzasum)(blasint *, double *, blasint *);
 xdouble BLASFUNC(qxasum)(blasint *, xdouble *, blasint *);
 
+FLOATRET  BLASFUNC(ssum) (blasint *, float  *, blasint *);
+FLOATRET  BLASFUNC(scsum)(blasint *, float  *, blasint *);
+double BLASFUNC(dsum) (blasint *, double *, blasint *);
+xdouble BLASFUNC(qsum) (blasint *, xdouble *, blasint *);
+double BLASFUNC(dzsum)(blasint *, double *, blasint *);
+xdouble BLASFUNC(qxsum)(blasint *, xdouble *, blasint *);
+
 blasint    BLASFUNC(isamax)(blasint *, float  *, blasint *);
 blasint    BLASFUNC(idamax)(blasint *, double *, blasint *);
 blasint    BLASFUNC(iqamax)(blasint *, xdouble *, blasint *);

common_level1.h

@@ -100,6 +100,13 @@ float   casum_k (BLASLONG, float  *, BLASLONG);
 double  zasum_k (BLASLONG, double *, BLASLONG);
 xdouble xasum_k (BLASLONG, xdouble *, BLASLONG);
 
+float   ssum_k (BLASLONG, float  *, BLASLONG);
+double  dsum_k (BLASLONG, double *, BLASLONG);
+xdouble qsum_k (BLASLONG, xdouble *, BLASLONG);
+float   csum_k (BLASLONG, float  *, BLASLONG);
+double  zsum_k (BLASLONG, double *, BLASLONG);
+xdouble xsum_k (BLASLONG, xdouble *, BLASLONG);
+
 float   samax_k (BLASLONG, float  *, BLASLONG);
 double  damax_k (BLASLONG, double *, BLASLONG);
 xdouble qamax_k (BLASLONG, xdouble *, BLASLONG);

common_macro.h

@@ -66,6 +66,7 @@
 #define	DOTC_K			QDOTC_K
 #define	NRM2_K			QNRM2_K
 #define	SCAL_K			QSCAL_K
+#define	SUM_K			QSUM_K
 #define	SWAP_K			QSWAP_K
 #define	ROT_K			QROT_K
 
@@ -356,6 +357,7 @@
 #define	DOTC_K			DDOTC_K
 #define	NRM2_K			DNRM2_K
 #define	SCAL_K			DSCAL_K
+#define	SUM_K			DSUM_K
 #define	SWAP_K			DSWAP_K
 #define	ROT_K			DROT_K
 
@@ -658,6 +660,7 @@
 #define	DOTC_K			SDOTC_K
 #define	NRM2_K			SNRM2_K
 #define	SCAL_K			SSCAL_K
+#define	SUM_K			SSUM_K
 #define	SWAP_K			SSWAP_K
 #define	ROT_K			SROT_K
 
@@ -962,6 +965,7 @@
 #define	DOTC_K			XDOTC_K
 #define	NRM2_K			XNRM2_K
 #define	SCAL_K			XSCAL_K
+#define	SUM_K			XSUM_K
 #define	SWAP_K			XSWAP_K
 #define	ROT_K			XROT_K
 
@@ -1363,6 +1367,7 @@
 #define	DOTC_K			ZDOTC_K
 #define	NRM2_K			ZNRM2_K
 #define	SCAL_K			ZSCAL_K
+#define	SUM_K			ZSUM_K
 #define	SWAP_K			ZSWAP_K
 #define	ROT_K			ZROT_K
 
@@ -1785,6 +1790,7 @@
 #define	DOTC_K			CDOTC_K
 #define	NRM2_K			CNRM2_K
 #define	SCAL_K			CSCAL_K
+#define	SUM_K			CSUM_K
 #define	SWAP_K			CSWAP_K
 #define	ROT_K			CROT_K
 

common_param.h

@@ -63,6 +63,7 @@ BLASLONG (*ismin_k) (BLASLONG, float *, BLASLONG);
 
   float  (*snrm2_k) (BLASLONG, float *, BLASLONG);
   float  (*sasum_k) (BLASLONG, float *, BLASLONG);
+  float  (*ssum_k)  (BLASLONG, float *, BLASLONG);
   int    (*scopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
   float  (*sdot_k)  (BLASLONG, float *, BLASLONG, float *, BLASLONG);
   double (*dsdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
@@ -154,6 +155,7 @@ BLASLONG (*idmin_k) (BLASLONG, double *, BLASLONG);
 
   double (*dnrm2_k) (BLASLONG, double *, BLASLONG);
   double (*dasum_k) (BLASLONG, double *, BLASLONG);
+  double (*dsum_k)  (BLASLONG, double *, BLASLONG);
   int    (*dcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
   double (*ddot_k)  (BLASLONG, double *, BLASLONG, double *, BLASLONG);
   int    (*drot_k)  (BLASLONG, double *, BLASLONG, double *, BLASLONG, double, double);
@@ -245,6 +247,7 @@ BLASLONG (*iqmin_k) (BLASLONG, xdouble *, BLASLONG);
 
  xdouble (*qnrm2_k) (BLASLONG, xdouble *, BLASLONG);
  xdouble (*qasum_k) (BLASLONG, xdouble *, BLASLONG);
+ xdouble (*qsum_k)  (BLASLONG, xdouble *, BLASLONG);
   int    (*qcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
  xdouble (*qdot_k)  (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
   int    (*qrot_k)  (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble);
@@ -332,6 +335,7 @@ BLASLONG (*icamin_k)(BLASLONG, float *, BLASLONG);
 
   float (*cnrm2_k) (BLASLONG, float *, BLASLONG);
   float (*casum_k) (BLASLONG, float *, BLASLONG);
+  float (*csum_k)  (BLASLONG, float *, BLASLONG);
   int    (*ccopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
   openblas_complex_float (*cdotu_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
   openblas_complex_float (*cdotc_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
@@ -495,6 +499,7 @@ BLASLONG (*izamin_k)(BLASLONG, double *, BLASLONG);
 
   double (*znrm2_k) (BLASLONG, double *, BLASLONG);
   double (*zasum_k) (BLASLONG, double *, BLASLONG);
+  double (*zsum_k)  (BLASLONG, double *, BLASLONG);
   int    (*zcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
   openblas_complex_double (*zdotu_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
   openblas_complex_double (*zdotc_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
@@ -660,6 +665,7 @@ BLASLONG (*ixamin_k)(BLASLONG, xdouble *, BLASLONG);
 
   xdouble (*xnrm2_k) (BLASLONG, xdouble *, BLASLONG);
   xdouble (*xasum_k) (BLASLONG, xdouble *, BLASLONG);
+  xdouble (*xsum_k) (BLASLONG, xdouble *, BLASLONG);
   int    (*xcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
   openblas_complex_xdouble (*xdotu_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
   openblas_complex_xdouble (*xdotc_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);

common_power.h

@@ -39,7 +39,7 @@
 #ifndef COMMON_POWER
 #define COMMON_POWER
 
-#if defined(POWER8)
+#if defined(POWER8) || defined(POWER9)
 #define MB		__asm__ __volatile__ ("eieio":::"memory")
 #define WMB		__asm__ __volatile__ ("eieio":::"memory")
 #else
@@ -241,7 +241,7 @@ static inline int blas_quickdivide(blasint x, blasint y){
 #define HAVE_PREFETCH
 #endif
 
-#if defined(POWER3) || defined(POWER6) || defined(PPCG4) || defined(CELL) || defined(POWER8)
+#if defined(POWER3) || defined(POWER6) || defined(PPCG4) || defined(CELL) || defined(POWER8) || defined(POWER9) || ( defined(PPC970) && defined(OS_DARWIN) )
 #define DCBT_ARG	0
 #else
 #define DCBT_ARG	8
@@ -263,7 +263,7 @@ static inline int blas_quickdivide(blasint x, blasint y){
 #define L1_PREFETCH	dcbtst
 #endif
 
-#if defined(POWER8)
+#if defined(POWER8) || defined(POWER9)
 #define L1_DUALFETCH
 #define L1_PREFETCHSIZE (16 + 128 * 100)
 #define L1_PREFETCH	dcbtst
@@ -598,9 +598,14 @@ REALNAME:;\
 #ifndef __64BIT__
 #define PROLOGUE \
 	.machine "any";\
+	.toc;\
 	.globl .REALNAME;\
+	.globl REALNAME;\
+	.csect REALNAME[DS],3;\
+REALNAME:;\
+	.long .REALNAME, TOC[tc0], 0;\
 	.csect .text[PR],5;\
-.REALNAME:;
+.REALNAME:
 
 #define EPILOGUE \
 _section_.text:;\
@@ -611,9 +616,14 @@ _section_.text:;\
 
 #define PROLOGUE \
 	.machine "any";\
+	.toc;\
 	.globl .REALNAME;\
+	.globl REALNAME;\
+	.csect REALNAME[DS],3;\
+REALNAME:;\
+	.llong .REALNAME, TOC[tc0], 0;\
 	.csect .text[PR], 5;\
-.REALNAME:;
+.REALNAME:
 
 #define EPILOGUE \
 _section_.text:;\
@@ -802,7 +812,7 @@ Lmcount$lazy_ptr:
 #define BUFFER_SIZE     (  2 << 20)
 #elif defined(PPC440FP2)
 #define BUFFER_SIZE     ( 16 << 20)
-#elif defined(POWER8)
+#elif defined(POWER8) || defined(POWER9)
 #define BUFFER_SIZE     ( 64 << 20)
 #else
 #define BUFFER_SIZE     ( 16 << 20)

common_q.h

@@ -19,6 +19,7 @@
 #define	QDOTC_K			qdot_k
 #define	QNRM2_K			qnrm2_k
 #define	QSCAL_K			qscal_k
+#define	QSUM_K			qsum_k
 #define	QSWAP_K			qswap_k
 #define	QROT_K			qrot_k
 
@@ -161,6 +162,7 @@
 #define	QDOTC_K			gotoblas -> qdot_k
 #define	QNRM2_K			gotoblas -> qnrm2_k
 #define	QSCAL_K			gotoblas -> qscal_k
+#define	QSUM_K			gotoblas -> qsum_k
 #define	QSWAP_K			gotoblas -> qswap_k
 #define	QROT_K			gotoblas -> qrot_k
 

common_s.h

@@ -12,6 +12,7 @@
 #define	ISMAX_K			ismax_k
 #define	ISMIN_K			ismin_k
 #define	SASUM_K			sasum_k
+#define	SSUM_K			ssum_k
 #define	SAXPYU_K		saxpy_k
 #define	SAXPYC_K		saxpy_k
 #define	SCOPY_K			scopy_k
@@ -170,6 +171,7 @@
 #define	ISMAX_K			gotoblas -> ismax_k
 #define	ISMIN_K			gotoblas -> ismin_k
 #define	SASUM_K			gotoblas -> sasum_k
+#define	SSUM_K			gotoblas -> ssum_k
 #define	SAXPYU_K		gotoblas -> saxpy_k
 #define	SAXPYC_K		gotoblas -> saxpy_k
 #define	SCOPY_K			gotoblas -> scopy_k

common_x.h

@@ -19,6 +19,7 @@
 #define	XDOTC_K			xdotc_k
 #define	XNRM2_K			xnrm2_k
 #define	XSCAL_K			xscal_k
+#define	XSUM_K			xsum_k
 #define	XSWAP_K			xswap_k
 #define	XROT_K			xqrot_k
 
@@ -227,6 +228,7 @@
 #define	XDOTC_K			gotoblas -> xdotc_k
 #define	XNRM2_K			gotoblas -> xnrm2_k
 #define	XSCAL_K			gotoblas -> xscal_k
+#define	XSUM_K			gotoblas -> xsum_k
 #define	XSWAP_K			gotoblas -> xswap_k
 #define	XROT_K			gotoblas -> xqrot_k
 

common_x86.h

@@ -187,7 +187,7 @@ static __inline int blas_quickdivide(unsigned int x, unsigned int y){
 	
   y = blas_quick_divide_table[y];
 
-  __asm__ __volatile__  ("mull %0" :"=d" (result) :"a"(x), "0" (y));
+  __asm__ __volatile__  ("mull %0" :"=d" (result), "+a"(x): "0" (y));
 
   return result;
 #endif

common_x86_64.h

@@ -134,7 +134,7 @@ static __inline void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx){
 			     "=b" (*ebx),
 			     "=c" (*ecx),
 			     "=d" (*edx)
-			     : "0" (op));
+			     : "0" (op), "c"(0));
 #endif
 }
 
@@ -210,7 +210,7 @@ static __inline int blas_quickdivide(unsigned int x, unsigned int y){
 	
   y = blas_quick_divide_table[y];
 
-  __asm__ __volatile__  ("mull %0" :"=d" (result) :"a"(x), "0" (y));
+  __asm__ __volatile__  ("mull %0" :"=d" (result), "+a"(x) : "0" (y));
 
   return result;
 }

common_z.h

@@ -19,6 +19,7 @@
 #define	ZDOTC_K			zdotc_k
 #define	ZNRM2_K			znrm2_k
 #define	ZSCAL_K			zscal_k
+#define	ZSUM_K			zsum_k
 #define	ZSWAP_K			zswap_k
 #define	ZROT_K			zdrot_k
 
@@ -249,6 +250,7 @@
 #define	ZDOTC_K			gotoblas -> zdotc_k
 #define	ZNRM2_K			gotoblas -> znrm2_k
 #define	ZSCAL_K			gotoblas -> zscal_k
+#define	ZSUM_K			gotoblas -> zsum_k
 #define	ZSWAP_K			gotoblas -> zswap_k
 #define	ZROT_K			gotoblas -> zdrot_k
 

cpuid.h

@@ -53,6 +53,7 @@
 #define VENDOR_SIS	  8
 #define VENDOR_TRANSMETA  9
 #define VENDOR_NSC	 10
+#define VENDOR_HYGON	 11
 #define VENDOR_UNKNOWN   99
 
 #define BITMASK(a, b, c) ((((a) >> (b)) & (c)))
@@ -116,6 +117,7 @@
 #define CORE_EXCAVATOR   26
 #define CORE_ZEN         27
 #define CORE_SKYLAKEX    28
+#define CORE_DHYANA	 29
 
 #define HAVE_SSE      (1 <<  0)
 #define HAVE_SSE2     (1 <<  1)
@@ -139,6 +141,7 @@
 #define HAVE_FMA4     (1 <<  19)
 #define HAVE_FMA3     (1 <<  20)
 #define HAVE_AVX512VL (1 <<  21)
+#define HAVE_AVX2     (1 <<  22)
 
 #define CACHE_INFO_L1_I     1
 #define CACHE_INFO_L1_D     2
@@ -214,5 +217,8 @@ typedef struct {
 #define CPUTYPE_EXCAVATOR 		50
 #define CPUTYPE_ZEN 			51
 #define CPUTYPE_SKYLAKEX		52
+#define CPUTYPE_DHYANA			53
+
+#define CPUTYPE_HYGON_UNKNOWN		54
 
 #endif

cpuid_arm64.c

@@ -39,6 +39,8 @@
 // Cavium
 #define CPU_THUNDERX      7
 #define CPU_THUNDERX2T99  8
+//Hisilicon
+#define CPU_TSV110        9
 
 static char *cpuname[] = {
   "UNKNOWN",
@@ -49,7 +51,8 @@ static char *cpuname[] = {
   "CORTEXA73",
   "FALKOR",
   "THUNDERX",
-  "THUNDERX2T99"
+  "THUNDERX2T99",
+  "TSV110"
 };
 
 static char *cpuname_lower[] = {
@@ -61,7 +64,8 @@ static char *cpuname_lower[] = {
   "cortexa73",
   "falkor",
   "thunderx",
-  "thunderx2t99"
+  "thunderx2t99",
+  "tsv110"
 };
 
 int get_feature(char *search)
@@ -145,6 +149,9 @@ int detect(void)
 			return CPU_THUNDERX;
     else if (strstr(cpu_implementer, "0x43") && strstr(cpu_part, "0x0af"))
 			return CPU_THUNDERX2T99;
+    // HiSilicon
+    else if (strstr(cpu_implementer, "0x48") && strstr(cpu_part, "0xd01"))
+                        return CPU_TSV110;
 	}
 
 	p = (char *) NULL ;
@@ -286,6 +293,21 @@ void get_cpuconfig(void)
 			printf("#define DTB_DEFAULT_ENTRIES  64       \n");
 			printf("#define DTB_SIZE             4096     \n");
 			break;
+			
+		case CPU_TSV110:
+			printf("#define TSV110                        \n");
+			printf("#define L1_CODE_SIZE         65536    \n");
+			printf("#define L1_CODE_LINESIZE     64       \n");
+			printf("#define L1_CODE_ASSOCIATIVE  4        \n");
+			printf("#define L1_DATA_SIZE         65536    \n");
+			printf("#define L1_DATA_LINESIZE     64       \n");
+			printf("#define L1_DATA_ASSOCIATIVE  4        \n");
+			printf("#define L2_SIZE              524228   \n");
+			printf("#define L2_LINESIZE          64       \n");
+			printf("#define L2_ASSOCIATIVE       8        \n");
+			printf("#define DTB_DEFAULT_ENTRIES  64       \n");
+			printf("#define DTB_SIZE             4096     \n");
+			break;	
 	}
 }
 

cpuid_power.c

@@ -94,7 +94,7 @@ char *corename[] = {
   "CELL",
   "PPCG4",
   "POWER8",
-  "POWER8"   	
+  "POWER9"   	
 };
 
 int detect(void){
@@ -124,7 +124,7 @@ int detect(void){
   if (!strncasecmp(p, "POWER6", 6)) return CPUTYPE_POWER6;
   if (!strncasecmp(p, "POWER7", 6)) return CPUTYPE_POWER6;
   if (!strncasecmp(p, "POWER8", 6)) return CPUTYPE_POWER8;
-  if (!strncasecmp(p, "POWER9", 6)) return CPUTYPE_POWER8;
+  if (!strncasecmp(p, "POWER9", 6)) return CPUTYPE_POWER9;
   if (!strncasecmp(p, "Cell",   4)) return CPUTYPE_CELL;
   if (!strncasecmp(p, "7447",   4)) return CPUTYPE_PPCG4;
 
@@ -156,7 +156,7 @@ int detect(void){
   if (!strncasecmp(p, "POWER6", 6)) return CPUTYPE_POWER6;
   if (!strncasecmp(p, "POWER7", 6)) return CPUTYPE_POWER6;
   if (!strncasecmp(p, "POWER8", 6)) return CPUTYPE_POWER8;
-  if (!strncasecmp(p, "POWER9", 6)) return CPUTYPE_POWER8;
+  if (!strncasecmp(p, "POWER9", 6)) return CPUTYPE_POWER9;
   if (!strncasecmp(p, "Cell",   4)) return CPUTYPE_CELL;
   if (!strncasecmp(p, "7447",   4)) return CPUTYPE_PPCG4;
   return CPUTYPE_POWER5;
@@ -180,7 +180,7 @@ int id;
 __asm __volatile("mfpvr %0" : "=r"(id));
 switch ( id >> 16 ) {
   case 0x4e: // POWER9
-    return CPUTYPE_POWER8;
+    return CPUTYPE_POWER9;
     break;
   case 0x4d:
   case 0x4b: // POWER8/8E 

cpuid_x86.c

@@ -97,10 +97,10 @@ static C_INLINE void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx){
     ("mov %%ebx, %%edi;"
      "cpuid;"
      "xchgl %%ebx, %%edi;"
-     : "=a" (*eax), "=D" (*ebx), "=c" (*ecx), "=d" (*edx) : "a" (op) : "cc");
+     : "=a" (*eax), "=D" (*ebx), "=c" (*ecx), "=d" (*edx) : "a" (op), "c" (0) : "cc");
 #else
   __asm__ __volatile__
-    ("cpuid": "=a" (*eax), "=b" (*ebx), "=c" (*ecx), "=d" (*edx) : "a" (op) : "cc");
+    ("cpuid": "=a" (*eax), "=b" (*ebx), "=c" (*ecx), "=d" (*edx) : "a" (op) , "c" (0) : "cc");
 #endif
 }
 
@@ -211,6 +211,44 @@ int support_avx(){
 #endif
 }
 
+int support_avx2(){
+#ifndef NO_AVX2
+  int eax, ebx, ecx=0, edx;
+  int ret=0;
+
+  if (!support_avx()) 
+    return 0;
+  cpuid(7, &eax, &ebx, &ecx, &edx);
+  if((ebx & (1<<7)) != 0)
+      ret=1;  //OS supports AVX2
+  return ret;
+#else
+  return 0;
+#endif
+}
+
+int support_avx512(){
+#if !defined(NO_AVX) && !defined(NO_AVX512)
+  int eax, ebx, ecx, edx;
+  int ret=0;
+
+  if (!support_avx()) 
+    return 0;
+  cpuid(7, &eax, &ebx, &ecx, &edx);
+  if((ebx & 32) != 32){
+      ret=0;  //OS does not even support AVX2
+  }
+  if((ebx & (1<<31)) != 0){
+    xgetbv(0, &eax, &edx); 
+    if((eax & 0xe0) == 0xe0)
+      ret=1;  //OS supports AVX512VL
+  }
+  return ret;
+#else
+  return 0;
+#endif
+}
+
 
 int get_vendor(void){
   int eax, ebx, ecx, edx;
@@ -233,6 +271,7 @@ int get_vendor(void){
   if (!strcmp(vendor, " SiS SiS SiS")) return VENDOR_SIS;
   if (!strcmp(vendor, "GenuineTMx86")) return VENDOR_TRANSMETA;
   if (!strcmp(vendor, "Geode by NSC")) return VENDOR_NSC;
+  if (!strcmp(vendor, "HygonGenuine")) return VENDOR_HYGON;
 
   if ((eax == 0) || ((eax & 0x500) != 0)) return VENDOR_INTEL;
 
@@ -294,6 +333,8 @@ int get_cputype(int gettype){
     if ((ecx & (1 << 20)) != 0) feature |= HAVE_SSE4_2;
 #ifndef NO_AVX
     if (support_avx()) feature |= HAVE_AVX;
+    if (support_avx2()) feature |= HAVE_AVX2;
+    if (support_avx512()) feature |= HAVE_AVX512VL;
     if ((ecx & (1 << 12)) != 0) feature |= HAVE_FMA3;
 #endif
 
@@ -1006,7 +1047,9 @@ int get_cacheinfo(int type, cache_info_t *cacheinfo){
     }
   }
 
-  if ((get_vendor() == VENDOR_AMD) || (get_vendor() == VENDOR_CENTAUR)) {
+  if ((get_vendor() == VENDOR_AMD) ||
+      (get_vendor() == VENDOR_HYGON) ||
+      (get_vendor() == VENDOR_CENTAUR)) {
     cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
 
     LDTB.size        = 4096;
@@ -1228,22 +1271,18 @@ int get_cpuname(void){
 	    return CPUTYPE_NEHALEM;
         case 12:
 	case 15:
-          if(support_avx())
-#ifndef NO_AVX2
+          if(support_avx2())
             return CPUTYPE_HASWELL;
-#else
+          if(support_avx())
 	    return CPUTYPE_SANDYBRIDGE;
-#endif
           else
 	    return CPUTYPE_NEHALEM;
 	case 13:
 	  //Broadwell
-          if(support_avx())
-#ifndef NO_AVX2
+          if(support_avx2())
             return CPUTYPE_HASWELL;
-#else
+          if(support_avx())
 	    return CPUTYPE_SANDYBRIDGE;
-#endif
           else
 	    return CPUTYPE_NEHALEM;
 	}
@@ -1252,33 +1291,27 @@ int get_cpuname(void){
         switch (model) {
         case 5:
 	case 6:
-          if(support_avx())
-#ifndef NO_AVX2
+          if(support_avx2())
             return CPUTYPE_HASWELL;
-#else
+          if(support_avx())
 	    return CPUTYPE_SANDYBRIDGE;
-#endif
           else
 	    return CPUTYPE_NEHALEM;
 	case 7:
 	case 15:
 	  //Broadwell
-          if(support_avx())
-#ifndef NO_AVX2
+          if(support_avx2())
             return CPUTYPE_HASWELL;
-#else
+          if(support_avx())
 	    return CPUTYPE_SANDYBRIDGE;
-#endif
           else
 	    return CPUTYPE_NEHALEM;
 	case 14:
 	  //Skylake
-          if(support_avx())
-#ifndef NO_AVX2
+          if(support_avx2())
             return CPUTYPE_HASWELL;
-#else
+          if(support_avx())
 	    return CPUTYPE_SANDYBRIDGE;
-#endif
           else
 	    return CPUTYPE_NEHALEM;
 	case 12:
@@ -1292,80 +1325,66 @@ int get_cpuname(void){
         switch (model) {
 	case 6:
 	  //Broadwell
-          if(support_avx())
-#ifndef NO_AVX2
+          if(support_avx2())
             return CPUTYPE_HASWELL;
-#else
+          if(support_avx())
 	    return CPUTYPE_SANDYBRIDGE;
-#endif
           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
+          if(support_avx512())
+            return CPUTYPE_SKYLAKEX;
+          if(support_avx2())
+            return CPUTYPE_HASWELL;
+          if(support_avx())
+	    return CPUTYPE_SANDYBRIDGE;
 	  else
 	  return CPUTYPE_NEHALEM;
-#endif			
         case 14:
 	  // Skylake
-          if(support_avx())
-#ifndef NO_AVX2
+          if(support_avx2())
             return CPUTYPE_HASWELL;
-#else
+          if(support_avx())
 	    return CPUTYPE_SANDYBRIDGE;
-#endif
           else
 	    return CPUTYPE_NEHALEM;
 	case 7:
 	    // Xeon Phi Knights Landing
-          if(support_avx())
-#ifndef NO_AVX2
+          if(support_avx2())
             return CPUTYPE_HASWELL;
-#else
+          if(support_avx())
 	    return CPUTYPE_SANDYBRIDGE;
-#endif
           else
 	    return CPUTYPE_NEHALEM;
 	case 12:
 	    // Apollo Lake
+	case 15:
+	    // Denverton		
 	    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
+          if(support_avx512())
+            return CPUTYPE_SKYLAKEX;
+          if(support_avx2())
+            return CPUTYPE_HASWELL;
+          if(support_avx())
+	    return CPUTYPE_SANDYBRIDGE;
 	  else
 	  return CPUTYPE_NEHALEM;
-#endif			
         }
       break;  
       case 9:
-      case 8: 
+      case 8:      
         switch (model) {
-	case 14: // Kaby Lake
-          if(support_avx())
-#ifndef NO_AVX2
+	case 14: // Kaby Lake and refreshes
+          if(support_avx2())
             return CPUTYPE_HASWELL;
-#else
+          if(support_avx())
 	    return CPUTYPE_SANDYBRIDGE;
-#endif
           else
 	    return CPUTYPE_NEHALEM;
 	}
@@ -1469,6 +1488,26 @@ int get_cpuname(void){
     return CPUTYPE_AMD_UNKNOWN;
   }
 
+  if (vendor == VENDOR_HYGON){
+    switch (family) {
+    case 0xf:
+      switch (exfamily) {
+      case 9:
+          //Hygon Dhyana
+	  if(support_avx())
+#ifndef NO_AVX2
+	    return CPUTYPE_ZEN;
+#else
+	    return CPUTYPE_SANDYBRIDGE; // closer in architecture to Sandy Bridge than to Excavator
+#endif
+	  else
+	    return CPUTYPE_BARCELONA;
+        }
+      break;
+    }
+    return CPUTYPE_HYGON_UNKNOWN;
+  }
+
   if (vendor == VENDOR_CYRIX){
     switch (family) {
     case 0x4:
@@ -1590,7 +1629,8 @@ static char *cpuname[] = {
   "STEAMROLLER",
   "EXCAVATOR",
   "ZEN",
-  "SKYLAKEX"	
+  "SKYLAKEX",
+  "DHYANA"
 };
 
 static char *lowercpuname[] = {
@@ -1645,7 +1685,8 @@ static char *lowercpuname[] = {
   "steamroller",
   "excavator",
   "zen",
-  "skylakex"
+  "skylakex",
+  "dhyana"
 };
 
 static char *corename[] = {
@@ -1677,7 +1718,8 @@ static char *corename[] = {
   "STEAMROLLER",
   "EXCAVATOR",
   "ZEN",
-  "SKYLAKEX"	
+  "SKYLAKEX",
+  "DHYANA"
 };
 
 static char *corename_lower[] = {
@@ -1709,7 +1751,8 @@ static char *corename_lower[] = {
   "steamroller",
   "excavator",
   "zen",
-  "skylakex"	
+  "skylakex",
+  "dhyana"
 };
 
 
@@ -2026,6 +2069,23 @@ int get_coretype(void){
     }
   }
 
+  if (vendor == VENDOR_HYGON){
+    if (family == 0xf){
+        if (exfamily == 9) {
+	  if(support_avx())
+#ifndef NO_AVX2
+	    return CORE_ZEN;
+#else
+	    return CORE_SANDYBRIDGE; // closer in architecture to Sandy Bridge than to Excavator
+#endif
+	  else
+	    return CORE_BARCELONA;
+	} else {
+		return CORE_BARCELONA;
+	}
+    }
+  }
+
   if (vendor == VENDOR_CENTAUR) {
     switch (family) {
     case 0x6:
@@ -2112,6 +2172,8 @@ void get_cpuconfig(void){
     if (features & HAVE_SSE4A)   printf("#define HAVE_SSE4A\n");
     if (features & HAVE_SSE5 )   printf("#define HAVE_SSSE5\n");
     if (features & HAVE_AVX )    printf("#define HAVE_AVX\n");
+    if (features & HAVE_AVX2 )    printf("#define HAVE_AVX2\n");
+    if (features & HAVE_AVX512VL )    printf("#define HAVE_AVX512VL\n");
     if (features & HAVE_3DNOWEX) printf("#define HAVE_3DNOWEX\n");
     if (features & HAVE_3DNOW)   printf("#define HAVE_3DNOW\n");
     if (features & HAVE_FMA4 )    printf("#define HAVE_FMA4\n");
@@ -2180,6 +2242,8 @@ void get_sse(void){
   if (features & HAVE_SSE4A)   printf("HAVE_SSE4A=1\n");
   if (features & HAVE_SSE5 )   printf("HAVE_SSSE5=1\n");
   if (features & HAVE_AVX )    printf("HAVE_AVX=1\n");
+  if (features & HAVE_AVX2 )    printf("HAVE_AVX2=1\n");
+  if (features & HAVE_AVX512VL )    printf("HAVE_AVX512VL=1\n");
   if (features & HAVE_3DNOWEX) printf("HAVE_3DNOWEX=1\n");
   if (features & HAVE_3DNOW)   printf("HAVE_3DNOW=1\n");
   if (features & HAVE_FMA4 )    printf("HAVE_FMA4=1\n");

cpuid_zarch.c

@@ -27,9 +27,9 @@
 
 #include <string.h>
 
-#define CPU_GENERIC    	0
-#define CPU_Z13       	1
-#define CPU_Z14       	2
+#define CPU_GENERIC     0
+#define CPU_Z13         1
+#define CPU_Z14         2
 
 static char *cpuname[] = {
   "ZARCH_GENERIC",
@@ -64,10 +64,8 @@ int detect(void)
 
   if (strstr(p, "2964")) return CPU_Z13;
   if (strstr(p, "2965")) return CPU_Z13;
-
-  /* detect z14, but fall back to z13 */
-  if (strstr(p, "3906")) return CPU_Z13;
-  if (strstr(p, "3907")) return CPU_Z13;
+  if (strstr(p, "3906")) return CPU_Z14;
+  if (strstr(p, "3907")) return CPU_Z14;
 
   return CPU_GENERIC;
 }
@@ -116,7 +114,14 @@ void get_cpuconfig(void)
 	  break;
 	case CPU_Z14:
 	  printf("#define Z14\n");
+	  printf("#define L1_DATA_SIZE 131072\n");
+	  printf("#define L1_DATA_LINESIZE 256\n");
+	  printf("#define L1_DATA_ASSOCIATIVE 8\n");
+	  printf("#define L2_SIZE 4194304\n");
+	  printf("#define L2_LINESIZE 256\n");
+	  printf("#define L2_ASSOCIATIVE 8\n");
 	  printf("#define DTB_DEFAULT_ENTRIES 64\n");
+	  printf("#define DTB_SIZE 4096\n");
 	  break;
 	}
 }

ctest.c

@@ -113,7 +113,7 @@ ARCH_X86
 ARCH_X86_64
 #endif
 
-#if defined(__powerpc___) || defined(__PPC__) || defined(_POWER)
+#if defined(__powerpc___) || defined(__PPC__) || defined(_POWER) || defined(__POWERPC__)
 ARCH_POWER
 #endif
 

driver/level2/trmv_thread.c

@@ -346,7 +346,7 @@ int CNAME(BLASLONG m, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG incx, FLOAT *bu
 
     range_m[MAX_CPU_NUMBER - num_cpu - 1] = range_m[MAX_CPU_NUMBER - num_cpu] - width;
     range_n[num_cpu] = num_cpu * (((m + 15) & ~15) + 16);
-    if (range_n[num_cpu] > m) range_n[num_cpu] = m;
+    if (range_n[num_cpu] > m * num_cpu) range_n[num_cpu] = m * num_cpu;
 
     queue[num_cpu].mode    = mode;
     queue[num_cpu].routine = trmv_kernel;
@@ -386,7 +386,7 @@ int CNAME(BLASLONG m, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG incx, FLOAT *bu
 
     range_m[num_cpu + 1] = range_m[num_cpu] + width;
     range_n[num_cpu] = num_cpu * (((m + 15) & ~15) + 16);
-    if (range_n[num_cpu] > m) range_n[num_cpu] = m;
+    if (range_n[num_cpu] > m * num_cpu) range_n[num_cpu] = m * num_cpu;
 
     queue[num_cpu].mode    = mode;
     queue[num_cpu].routine = trmv_kernel;

driver/others/Makefile

@@ -18,8 +18,12 @@ ifeq ($(DYNAMIC_ARCH), 1)
 ifeq ($(ARCH),arm64)
 COMMONOBJS	+=  dynamic_arm64.$(SUFFIX)
 else
+ifeq ($(ARCH),power)
+COMMONOBJS	+=  dynamic_power.$(SUFFIX)
+else
 COMMONOBJS	+=  dynamic.$(SUFFIX)
 endif
+endif
 else
 COMMONOBJS	+=  parameter.$(SUFFIX)
 endif
@@ -78,8 +82,12 @@ ifeq ($(DYNAMIC_ARCH), 1)
 ifeq ($(ARCH),arm64)
 HPLOBJS = memory.$(SUFFIX) xerbla.$(SUFFIX) dynamic_arm64.$(SUFFIX)
 else
+ifeq ($(ARCH),power)
+HPLOBJS = memory.$(SUFFIX) xerbla.$(SUFFIX) dynamic_power.$(SUFFIX)
+else
 HPLOBJS = memory.$(SUFFIX) xerbla.$(SUFFIX) dynamic.$(SUFFIX)
 endif
+endif
 else
 HPLOBJS = memory.$(SUFFIX) xerbla.$(SUFFIX) parameter.$(SUFFIX)
 endif

driver/others/blas_server_win32.c

@@ -461,13 +461,18 @@ int BLASFUNC(blas_thread_shutdown)(void){
     SetEvent(pool.killed);
 
     for(i = 0; i < blas_num_threads - 1; i++){
+      // Could also just use WaitForMultipleObjects
       WaitForSingleObject(blas_threads[i], 5);  //INFINITE);
 #ifndef OS_WINDOWSSTORE
 // TerminateThread is only available with WINAPI_DESKTOP and WINAPI_SYSTEM not WINAPI_APP in UWP
       TerminateThread(blas_threads[i],0);
 #endif
+      CloseHandle(blas_threads[i]);
     }
 
+    CloseHandle(pool.filled);
+    CloseHandle(pool.killed);
+
     blas_server_avail = 0;
   }
 

driver/others/dynamic.c

@@ -274,6 +274,7 @@ extern gotoblas_t  gotoblas_SKYLAKEX;
 #define VENDOR_INTEL      1
 #define VENDOR_AMD        2
 #define VENDOR_CENTAUR    3
+#define VENDOR_HYGON	  4
 #define VENDOR_UNKNOWN   99
 
 #define BITMASK(a, b, c) ((((a) >> (b)) & (c)))
@@ -304,9 +305,49 @@ int support_avx(){
 #endif
 }
 
+int support_avx2(){
+#ifndef NO_AVX2
+  int eax, ebx, ecx=0, edx;
+  int ret=0;
+
+  if (!support_avx())
+    return 0;
+  cpuid(7, &eax, &ebx, &ecx, &edx);
+  if((ebx & (1<<7)) != 0)
+      ret=1;  //OS supports AVX2
+  return ret;
+#else
+  return 0;
+#endif
+}
+
+int support_avx512(){
+#if !defined(NO_AVX) && !defined(NO_AVX512)
+  int eax, ebx, ecx, edx;
+  int ret=0;
+
+  if (!support_avx())
+    return 0;
+  cpuid(7, &eax, &ebx, &ecx, &edx);
+  if((ebx & (1<<7)) != 1){
+      ret=0;  //OS does not even support AVX2
+  }
+  if((ebx & (1<<31)) != 0){
+    xgetbv(0, &eax, &edx);
+    if((eax & 0xe0) == 0xe0)
+      ret=1;  //OS supports AVX512VL
+  }
+  return ret;
+#else
+  return 0;
+#endif
+}
+
 extern void openblas_warning(int verbose, const char * msg);
 #define FALLBACK_VERBOSE 1
 #define NEHALEM_FALLBACK "OpenBLAS : Your OS does not support AVX instructions. OpenBLAS is using Nehalem kernels as a fallback, which may give poorer performance.\n"
+#define SANDYBRIDGE_FALLBACK "OpenBLAS : Your OS does not support AVX2 instructions. OpenBLAS is using Sandybridge kernels as a fallback, which may give poorer performance.\n"
+#define HASWELL_FALLBACK "OpenBLAS : Your OS does not support AVX512VL instructions. OpenBLAS is using Haswell kernels as a fallback, which may give poorer performance.\n"
 #define BARCELONA_FALLBACK "OpenBLAS : Your OS does not support AVX instructions. OpenBLAS is using Barcelona kernels as a fallback, which may give poorer performance.\n"
 
 static int get_vendor(void){
@@ -329,6 +370,7 @@ static int get_vendor(void){
   if (!strcmp(vendor.vchar, "GenuineIntel")) return VENDOR_INTEL;
   if (!strcmp(vendor.vchar, "AuthenticAMD")) return VENDOR_AMD;
   if (!strcmp(vendor.vchar, "CentaurHauls")) return VENDOR_CENTAUR;
+  if (!strcmp(vendor.vchar, "HygonGenuine")) return VENDOR_HYGON;
 
   if ((eax == 0) || ((eax & 0x500) != 0)) return VENDOR_INTEL;
 
@@ -403,18 +445,24 @@ static gotoblas_t *get_coretype(void){
 	}
 	//Intel Haswell
 	if (model == 12 || model == 15) {
-	  if(support_avx())
+	  if(support_avx2())
 	    return &gotoblas_HASWELL;
-	  else{
+	  if(support_avx()) {
+	    openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+	    return &gotoblas_SANDYBRIDGE;
+	  } else {
 	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
 	}
 	//Intel Broadwell
 	if (model == 13) {
-	  if(support_avx())
+	  if(support_avx2())
 	    return &gotoblas_HASWELL;
-	  else{
+	  if(support_avx()) {
+	    openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+	    return &gotoblas_SANDYBRIDGE;
+	  } else {
 	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
@@ -424,27 +472,36 @@ static gotoblas_t *get_coretype(void){
       case 4:
 		//Intel Haswell
 	if (model == 5 || model == 6) {
-	  if(support_avx())
+	  if(support_avx2())
 	    return &gotoblas_HASWELL;
-	  else{
+	  if(support_avx()) {
+	    openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+	    return &gotoblas_SANDYBRIDGE;
+	  } else {
 	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
 	}
 	//Intel Broadwell
 	if (model == 7 || model == 15) {
-	  if(support_avx())
+	  if(support_avx2())
 	    return &gotoblas_HASWELL;
-	  else{
+	  if(support_avx()) {
+	    openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+	    return &gotoblas_SANDYBRIDGE;
+	  } else {
 	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
 	}
 	//Intel Skylake
 	if (model == 14) {
-	  if(support_avx())
+	  if(support_avx2())
 	    return &gotoblas_HASWELL;
-	  else{
+	  if(support_avx()) {
+	    openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+	    return &gotoblas_SANDYBRIDGE;
+	  } else {
 	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
@@ -457,72 +514,86 @@ static gotoblas_t *get_coretype(void){
       case 5:
 	//Intel Broadwell
 	if (model == 6) {
-	  if(support_avx())
+	  if(support_avx2())
 	    return &gotoblas_HASWELL;
-	  else{
+	  if(support_avx()) {
+	    openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+	    return &gotoblas_SANDYBRIDGE;
+	  } else {
 	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
 	    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())
+          if (support_avx512()) 
+	    return &gotoblas_SKYLAKEX;
+	  if(support_avx2()){
+	    openblas_warning(FALLBACK_VERBOSE, HASWELL_FALLBACK);
 	    return &gotoblas_HASWELL;
-	  else {
-	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
-	    return &gotoblas_NEHALEM;
-	  }
-#endif		
+          }
+	  if(support_avx()) {
+	    openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+	    return &gotoblas_SANDYBRIDGE;
+	  } else {
+          openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
+          return &gotoblas_NEHALEM;
+          }
 	}
 	//Intel Skylake
 	if (model == 14) {
-	  if(support_avx())
+	  if(support_avx2())
 	    return &gotoblas_HASWELL;
-	  else{
+	  if(support_avx()) {
+	    openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+	    return &gotoblas_SANDYBRIDGE;
+	  } else {
 	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
 	}
 	//Intel Phi Knights Landing
 	if (model == 7) {
-	  if(support_avx())
+	  if(support_avx2()){
+	    openblas_warning(FALLBACK_VERBOSE, HASWELL_FALLBACK);
 	    return &gotoblas_HASWELL;
-	  else{
+	  }  
+	  if(support_avx()) {
+	    openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+	    return &gotoblas_SANDYBRIDGE;
+	  } else {
 	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
 	}
-	//Apollo Lake
-	if (model == 12) { 
+	//Apollo Lake or Denverton
+	if (model == 12 || model == 15) { 
 	  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 &gotoblas_SANDYBRIDGE;
-#endif
-	  else
-	  return &gotoblas_NEHALEM;
-#endif			
+	  if(support_avx2())
+	    return &gotoblas_HASWELL;
+	  if(support_avx()) {
+	    openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+	    return &gotoblas_SANDYBRIDGE;
+	  } else {
+	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
+	    return &gotoblas_NEHALEM;
+	  }
         }
         return NULL;  
       case 9:
       case 8:
 	if (model == 14 ) { // Kaby Lake
-	  if(support_avx())
+	  if(support_avx2())
 	    return &gotoblas_HASWELL;
-	  else{
+	  if(support_avx()) {
+	    openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+	    return &gotoblas_SANDYBRIDGE;
+	  } else {
 	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
@@ -535,7 +606,7 @@ static gotoblas_t *get_coretype(void){
     }
   }
 
-  if (vendor == VENDOR_AMD){
+  if (vendor == VENDOR_AMD || vendor == VENDOR_HYGON){
     if (family <= 0xe) {
         // Verify that CPU has 3dnow and 3dnowext before claiming it is Athlon
         cpuid(0x80000000, &eax, &ebx, &ecx, &edx);
@@ -615,6 +686,13 @@ static gotoblas_t *get_coretype(void){
 	    return &gotoblas_BARCELONA; //OS doesn't support AVX. Use old kernels.
 	  }
 	}
+      } else if (exfamily == 9) {
+	  if(support_avx())
+	    return &gotoblas_ZEN;
+	  else{
+	    openblas_warning(FALLBACK_VERBOSE, BARCELONA_FALLBACK);
+	    return &gotoblas_BARCELONA; //OS doesn't support AVX. Use old kernels.
+        }
       }else {
 	return &gotoblas_BARCELONA;
       }

driver/others/dynamic_power.c

@@ -0,0 +1,102 @@
+
+#include "common.h"
+
+extern gotoblas_t gotoblas_POWER6;
+extern gotoblas_t gotoblas_POWER8;
+extern gotoblas_t gotoblas_POWER9;
+
+extern void openblas_warning(int verbose, const char *msg);
+
+static char *corename[] = {
+	"unknown",
+	"POWER6",
+	"POWER8",
+	"POWER9"
+};
+
+#define NUM_CORETYPES 4
+
+char *gotoblas_corename(void) {
+	if (gotoblas == &gotoblas_POWER6)	return corename[1];
+	if (gotoblas == &gotoblas_POWER8)	return corename[2];
+	if (gotoblas == &gotoblas_POWER9)	return corename[3];
+	return corename[0];
+}
+
+static gotoblas_t *get_coretype(void) {
+
+	if (__builtin_cpu_is("power6") || __builtin_cpu_is("power6x"))
+		return &gotoblas_POWER6;
+	if (__builtin_cpu_is("power8"))
+		return &gotoblas_POWER8;
+	if (__builtin_cpu_is("power9"))
+		return &gotoblas_POWER9;
+	return NULL;
+}
+
+static gotoblas_t *force_coretype(char * coretype) {
+
+	int i ;
+	int found = -1;
+	char message[128];
+
+	for ( i = 0 ; i < NUM_CORETYPES; i++)
+	{
+		if (!strncasecmp(coretype, corename[i], 20))
+		{
+			found = i;
+			break;
+		}
+	}
+
+	switch (found)
+	{
+	case  1: return (&gotoblas_POWER6);
+	case  2: return (&gotoblas_POWER8);
+	case  3: return (&gotoblas_POWER9);
+	default: return NULL;
+	}
+	snprintf(message, 128, "Core not found: %s\n", coretype);
+	openblas_warning(1, message);
+}
+
+void gotoblas_dynamic_init(void) {
+
+	char coremsg[128];
+	char coren[22];
+	char *p;
+
+
+	if (gotoblas) return;
+
+	p = getenv("OPENBLAS_CORETYPE");
+	if ( p )
+	{
+		gotoblas = force_coretype(p);
+	}
+	else
+	{
+		gotoblas = get_coretype();
+	}
+
+	if (gotoblas == NULL)
+	{
+		snprintf(coremsg, 128, "Falling back to POWER8 core\n");
+		openblas_warning(1, coremsg);
+		gotoblas = &gotoblas_POWER8;
+	}
+
+	if (gotoblas && gotoblas -> init) {
+		strncpy(coren,gotoblas_corename(),20);
+		sprintf(coremsg, "Core: %s\n",coren);
+		openblas_warning(2, coremsg);
+		gotoblas -> init();
+	} else {
+		openblas_warning(0, "OpenBLAS : Architecture Initialization failed. No initialization function found.\n");
+		exit(1);
+	}
+}
+
+void gotoblas_dynamic_quit(void) {
+	gotoblas = NULL;
+}

driver/others/memory.c

@@ -198,45 +198,68 @@ int get_num_procs(void);
 #else
 int get_num_procs(void) {
   static int nums = 0;
-cpu_set_t *cpusetp;
-size_t size;
-int ret;
-int i,n;
+  cpu_set_t cpuset,*cpusetp;
+  size_t size;
+  int ret;
+
+#if defined(__GLIBC_PREREQ)
+#if !__GLIBC_PREREQ(2, 7)
+  int i;
+#if !__GLIBC_PREREQ(2, 6)
+  int n;
+#endif
+#endif
+#endif
 
   if (!nums) nums = sysconf(_SC_NPROCESSORS_CONF);
 #if !defined(OS_LINUX)
-     return nums;
+  return nums;
 #endif
 
 #if !defined(__GLIBC_PREREQ)
-   return nums;
+  return nums;
 #else
  #if !__GLIBC_PREREQ(2, 3)
-   return nums;
+  return nums;
  #endif
 
  #if !__GLIBC_PREREQ(2, 7)
-  ret = sched_getaffinity(0,sizeof(cpu_set_t), cpusetp);
+  ret = sched_getaffinity(0,sizeof(cpuset), &cpuset);
   if (ret!=0) return nums;
   n=0;
   #if !__GLIBC_PREREQ(2, 6)
   for (i=0;i<nums;i++)
-     if (CPU_ISSET(i,cpusetp)) n++;
+     if (CPU_ISSET(i,cpuset)) n++;
   nums=n;
   #else
-  nums = CPU_COUNT(sizeof(cpu_set_t),cpusetp);
+  nums = CPU_COUNT(sizeof(cpuset),&cpuset);
   #endif
   return nums;
  #else
-  cpusetp = CPU_ALLOC(nums);
-  if (cpusetp == NULL) return nums;
-  size = CPU_ALLOC_SIZE(nums);
-  ret = sched_getaffinity(0,size,cpusetp);
-  if (ret!=0) return nums;
-  ret = CPU_COUNT_S(size,cpusetp);
-  if (ret > 0 && ret < nums) nums = ret;
-  CPU_FREE(cpusetp);
-  return nums;
+  if (nums >= CPU_SETSIZE) {
+    cpusetp = CPU_ALLOC(nums);
+      if (cpusetp == NULL) {
+        return nums;
+      }
+    size = CPU_ALLOC_SIZE(nums);
+    ret = sched_getaffinity(0,size,cpusetp);
+    if (ret!=0) {
+      CPU_FREE(cpusetp);
+      return nums;
+    }
+    ret = CPU_COUNT_S(size,cpusetp);
+    if (ret > 0 && ret < nums) nums = ret;	
+    CPU_FREE(cpusetp);
+    return nums;
+  } else {
+    ret = sched_getaffinity(0,sizeof(cpuset),&cpuset);
+    if (ret!=0) {
+      return nums;
+    }
+    ret = CPU_COUNT(&cpuset);
+    if (ret > 0 && ret < nums) nums = ret;	
+    return nums;
+  }
  #endif
 #endif
 }
@@ -1073,11 +1096,6 @@ static volatile int memory_initialized = 0;
     }
     free(table);
   }
-#if defined(OS_WINDOWS)
-  TlsFree(local_storage_key);
-#else
-  pthread_key_delete(local_storage_key);
-#endif		
 }
 
 static void blas_memory_init(){
@@ -1295,6 +1313,13 @@ void blas_memory_free_nolock(void * map_address) {
   free(map_address);
 }
 
+#ifdef SMP
+void blas_thread_memory_cleanup(void) {
+    blas_memory_cleanup((void*)get_memory_table());
+}
+#endif
+
+
 void blas_shutdown(void){
 #ifdef SMP
   BLASFUNC(blas_thread_shutdown)();
@@ -1304,7 +1329,7 @@ void blas_shutdown(void){
   /* Only cleanupIf we were built for threading and TLS was initialized */
   if (local_storage_key)
 #endif
-    blas_memory_cleanup((void*)get_memory_table());
+    blas_thread_memory_cleanup();
 
 #ifdef SEEK_ADDRESS
   base_address      = 0UL;
@@ -1491,6 +1516,14 @@ void DESTRUCTOR gotoblas_quit(void) {
 
   blas_shutdown();
 
+#if defined(SMP)
+#if defined(OS_WINDOWS)
+  TlsFree(local_storage_key);
+#else
+  pthread_key_delete(local_storage_key);
+#endif		
+#endif
+
 #ifdef PROFILE
    moncontrol (0);
 #endif
@@ -1526,7 +1559,7 @@ BOOL APIENTRY DllMain(HMODULE hModule, DWORD  ul_reason_for_call, LPVOID lpReser
       break;
     case DLL_THREAD_DETACH:
 #if defined(SMP)
-      blas_memory_cleanup((void*)get_memory_table());
+      blas_thread_memory_cleanup();
 #endif
       break;
     case DLL_PROCESS_DETACH:
@@ -1600,9 +1633,11 @@ void gotoblas_dummy_for_PGI(void) {
 #endif
 
 #else
+/* USE_TLS / COMPILE_TLS not set */
+
 #include <errno.h>
 
-#ifdef OS_WINDOWS
+#if defined(OS_WINDOWS) && !defined(OS_CYGWIN_NT)
 #define ALLOC_WINDOWS
 #ifndef MEM_LARGE_PAGES
 #define MEM_LARGE_PAGES  0x20000000
@@ -1616,7 +1651,7 @@ void gotoblas_dummy_for_PGI(void) {
 #include <stdio.h>
 #include <fcntl.h>
 
-#ifndef OS_WINDOWS
+#if !defined(OS_WINDOWS) || defined(OS_CYGWIN_NT)
 #include <sys/mman.h>
 #ifndef NO_SYSV_IPC
 #include <sys/shm.h>
@@ -1636,7 +1671,7 @@ void gotoblas_dummy_for_PGI(void) {
 #include <sys/resource.h>
 #endif
 
-#if defined(OS_FREEBSD) || defined(OS_DARWIN)
+#if defined(OS_FREEBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY) || defined(OS_DARWIN)
 #include <sys/sysctl.h>
 #include <sys/resource.h>
 #endif
@@ -1675,9 +1710,12 @@ void gotoblas_dummy_for_PGI(void) {
 #elif (defined(OS_DARWIN) || defined(OS_SUNOS)) && defined(C_GCC)
 #define CONSTRUCTOR	__attribute__ ((constructor))
 #define DESTRUCTOR	__attribute__ ((destructor))
-#else
+#elif __GNUC__ && INIT_PRIORITY && ((GCC_VERSION >= 40300) || (CLANG_VERSION >= 20900))
 #define CONSTRUCTOR	__attribute__ ((constructor(101)))
 #define DESTRUCTOR	__attribute__ ((destructor(101)))
+#else
+#define CONSTRUCTOR	__attribute__ ((constructor))
+#define DESTRUCTOR	__attribute__ ((destructor))
 #endif
 
 #ifdef DYNAMIC_ARCH
@@ -1701,45 +1739,70 @@ void goto_set_num_threads(int num_threads) {};
 int get_num_procs(void);
 #else
 int get_num_procs(void) {
+
   static int nums = 0;
-cpu_set_t *cpusetp;
-size_t size;
-int ret;
-int i,n;
+  cpu_set_t cpuset,*cpusetp;
+  size_t size;
+  int ret;
+
+#if defined(__GLIBC_PREREQ)
+#if !__GLIBC_PREREQ(2, 7)
+  int i;
+#if !__GLIBC_PREREQ(2, 6)
+  int n;
+#endif
+#endif
+#endif
 
   if (!nums) nums = sysconf(_SC_NPROCESSORS_CONF);
 #if !defined(OS_LINUX)
-     return nums;
+  return nums;
 #endif
 
 #if !defined(__GLIBC_PREREQ)
-   return nums;
+  return nums;
 #else
  #if !__GLIBC_PREREQ(2, 3)
-   return nums;
+  return nums;
  #endif
 
  #if !__GLIBC_PREREQ(2, 7)
-  ret = sched_getaffinity(0,sizeof(cpu_set_t), cpusetp);
+  ret = sched_getaffinity(0,sizeof(cpuset), &cpuset);
   if (ret!=0) return nums;
   n=0;
   #if !__GLIBC_PREREQ(2, 6)
   for (i=0;i<nums;i++)
-     if (CPU_ISSET(i,cpusetp)) n++;
+     if (CPU_ISSET(i,cpuset)) n++;
   nums=n;
   #else
-  nums = CPU_COUNT(sizeof(cpu_set_t),cpusetp);
+  nums = CPU_COUNT(sizeof(cpuset),&cpuset);
   #endif
   return nums;
  #else
-  cpusetp = CPU_ALLOC(nums);
-  if (cpusetp == NULL) return nums;
-  size = CPU_ALLOC_SIZE(nums);
-  ret = sched_getaffinity(0,size,cpusetp);
-  if (ret!=0) return nums;
-  nums = CPU_COUNT_S(size,cpusetp);
-  CPU_FREE(cpusetp);
-  return nums;
+  if (nums >= CPU_SETSIZE) {
+    cpusetp = CPU_ALLOC(nums);
+      if (cpusetp == NULL) {
+        return nums;
+      }
+    size = CPU_ALLOC_SIZE(nums);
+    ret = sched_getaffinity(0,size,cpusetp);
+    if (ret!=0) {
+      CPU_FREE(cpusetp);
+      return nums;
+    }
+    ret = CPU_COUNT_S(size,cpusetp);
+    if (ret > 0 && ret < nums) nums = ret;	
+    CPU_FREE(cpusetp);
+    return nums;
+  } else {
+    ret = sched_getaffinity(0,sizeof(cpuset),&cpuset);
+    if (ret!=0) {
+      return nums;
+    }
+    ret = CPU_COUNT(&cpuset);
+    if (ret > 0 && ret < nums) nums = ret;	
+    return nums;
+  }
  #endif
 #endif
 }
@@ -1753,7 +1816,7 @@ int get_num_procs(void) {
   return nums;
 }
 #endif
-	
+
 #ifdef OS_HAIKU
 int get_num_procs(void) {
   static int nums = 0;
@@ -1790,7 +1853,7 @@ int get_num_procs(void) {
 
 #endif
 
-#if defined(OS_FREEBSD)
+#if defined(OS_FREEBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY)
 
 int get_num_procs(void) {
 
@@ -1867,7 +1930,7 @@ void openblas_fork_handler()
   //   http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60035
   // In the mean time build with USE_OPENMP=0 or link against another
   // implementation of OpenMP.
-#if !(defined(OS_WINDOWS) || defined(OS_ANDROID)) && defined(SMP_SERVER)
+#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);
   if(err != 0)
@@ -1880,7 +1943,7 @@ extern int openblas_goto_num_threads_env();
 extern int openblas_omp_num_threads_env();
 
 int blas_get_cpu_number(void){
-#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_DARWIN) || defined(OS_ANDROID)
+#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY) || defined(OS_DARWIN) || defined(OS_ANDROID)
   int max_num;
 #endif
   int blas_goto_num   = 0;
@@ -1888,11 +1951,11 @@ int blas_get_cpu_number(void){
 
   if (blas_num_threads) return blas_num_threads;
 
-#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_DARWIN) || defined(OS_ANDROID)
+#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY) || defined(OS_DARWIN) || defined(OS_ANDROID)
   max_num = get_num_procs();
 #endif
 
-  blas_goto_num = 0;
+  // blas_goto_num = 0;
 #ifndef USE_OPENMP
   blas_goto_num=openblas_num_threads_env();
   if (blas_goto_num < 0) blas_goto_num = 0;
@@ -1904,7 +1967,7 @@ int blas_get_cpu_number(void){
 
 #endif
 
-  blas_omp_num = 0;
+  // blas_omp_num = 0;
   blas_omp_num=openblas_omp_num_threads_env();
   if (blas_omp_num < 0) blas_omp_num = 0;
 
@@ -1912,7 +1975,7 @@ int blas_get_cpu_number(void){
   else if (blas_omp_num > 0) blas_num_threads = blas_omp_num;
   else blas_num_threads = MAX_CPU_NUMBER;
 
-#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_DARWIN) || defined(OS_ANDROID)
+#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY) || defined(OS_DARWIN) || defined(OS_ANDROID)
   if (blas_num_threads > max_num) blas_num_threads = max_num;
 #endif
 
@@ -1999,11 +2062,15 @@ static void *alloc_mmap(void *address){
   }
 
   if (map_address != (void *)-1) {
+#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
@@ -2145,14 +2212,18 @@ static void *alloc_mmap(void *address){
 #if defined(OS_LINUX) && !defined(NO_WARMUP)
   }
 #endif
-  LOCK_COMMAND(&alloc_lock);
 
   if (map_address != (void *)-1) {
+#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
   }
-  UNLOCK_COMMAND(&alloc_lock);
 
   return map_address;
 }
@@ -2520,7 +2591,7 @@ void *blas_memory_alloc(int procpos){
 
   int position;
 #if defined(WHEREAMI) && !defined(USE_OPENMP)
-  int mypos;
+  int mypos = 0;
 #endif
 
   void *map_address;
@@ -2551,6 +2622,11 @@ void *blas_memory_alloc(int procpos){
     NULL,
   };
   void *(**func)(void *address);
+
+#if defined(USE_OPENMP)
+  if (!memory_initialized) {
+#endif
+
   LOCK_COMMAND(&alloc_lock);
 
   if (!memory_initialized) {
@@ -2586,6 +2662,9 @@ void *blas_memory_alloc(int procpos){
 
   }
   UNLOCK_COMMAND(&alloc_lock);
+#if defined(USE_OPENMP)
+  }
+#endif
 
 #ifdef DEBUG
   printf("Alloc Start ...\n");
@@ -2600,13 +2679,17 @@ void *blas_memory_alloc(int procpos){
 
   do {
     if (!memory[position].used && (memory[position].pos == mypos)) {
+#if defined(SMP) && !defined(USE_OPENMP)
       LOCK_COMMAND(&alloc_lock);
-//      blas_lock(&memory[position].lock);
-
+#else      
+      blas_lock(&memory[position].lock);
+#endif
       if (!memory[position].used) goto allocation;
-
+#if defined(SMP) && !defined(USE_OPENMP)
       UNLOCK_COMMAND(&alloc_lock);
-//      blas_unlock(&memory[position].lock);
+#else
+      blas_unlock(&memory[position].lock);
+#endif      
     }
 
     position ++;
@@ -2618,21 +2701,26 @@ void *blas_memory_alloc(int procpos){
 
   position = 0;
 
+#if defined(SMP) && !defined(USE_OPENMP)
   LOCK_COMMAND(&alloc_lock);
+#endif
   do {
-/*    if (!memory[position].used) { */
-/*      blas_lock(&memory[position].lock);*/
-
+#if defined(USE_OPENMP)	  
+    if (!memory[position].used) { 
+      blas_lock(&memory[position].lock);
+#endif
       if (!memory[position].used) goto allocation;
       
-/*      blas_unlock(&memory[position].lock);*/
-/*    } */
-
+#if defined(USE_OPENMP)
+      blas_unlock(&memory[position].lock);      
+    }
+#endif
     position ++;
 
   } while (position < NUM_BUFFERS);
-  UNLOCK_COMMAND(&alloc_lock);
-
+#if defined(SMP) && !defined(USE_OPENMP)
+  UNLOCK_COMMAND(&alloc_lock);	
+#endif
   goto error;
 
   allocation :
@@ -2642,10 +2730,11 @@ void *blas_memory_alloc(int procpos){
 #endif
 
   memory[position].used = 1;
-
+#if defined(SMP) && !defined(USE_OPENMP)
   UNLOCK_COMMAND(&alloc_lock);
-/*  blas_unlock(&memory[position].lock);*/
-
+#else
+  blas_unlock(&memory[position].lock);	
+#endif
   if (!memory[position].addr) {
     do {
 #ifdef DEBUG
@@ -2690,9 +2779,13 @@ void *blas_memory_alloc(int procpos){
 
     } while ((BLASLONG)map_address == -1);
 
+#if defined(SMP) && !defined(USE_OPENMP)
     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);
@@ -2746,8 +2839,9 @@ void blas_memory_free(void *free_area){
 #endif
 
   position = 0;
+#if defined(SMP) && !defined(USE_OPENMP)
   LOCK_COMMAND(&alloc_lock);
-
+#endif
   while ((position < NUM_BUFFERS) && (memory[position].addr != free_area))
     position++;
 
@@ -2761,7 +2855,9 @@ void blas_memory_free(void *free_area){
   WMB;
 
   memory[position].used = 0;
+#if defined(SMP) && !defined(USE_OPENMP)
   UNLOCK_COMMAND(&alloc_lock);
+#endif
 
 #ifdef DEBUG
   printf("Unmap Succeeded.\n\n");
@@ -2776,8 +2872,9 @@ void blas_memory_free(void *free_area){
   for (position = 0; position < NUM_BUFFERS; position++)
     printf("%4ld  %p : %d\n", position, memory[position].addr, memory[position].used);
 #endif
+#if defined(SMP) && !defined(USE_OPENMP)
   UNLOCK_COMMAND(&alloc_lock);
-
+#endif
   return;
 }
 

driver/others/openblas_get_config.c

@@ -35,12 +35,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include <string.h>
 
-#if defined(_WIN32) && defined(_MSC_VER)
-#if _MSC_VER < 1900
-#define snprintf _snprintf
-#endif
-#endif
-
 static char* openblas_config_str=""
 "OpenBLAS "
  VERSION

exports/Makefile

@@ -141,6 +141,14 @@ else
 	$(OBJCOPY) --redefine-syms objcopy.def ../$(LIBNAME) ../$(LIBNAME).renamed
 ../$(LIBSONAME) : ../$(LIBNAME).renamed linktest.c
 endif
+
+ifeq ($(F_COMPILER), INTEL)
+	$(FC) $(FFLAGS) $(LDFLAGS) -shared -o ../$(LIBSONAME) \
+	-Wl,--whole-archive $< -Wl,--no-whole-archive \
+	-Wl,-soname,$(INTERNALNAME) $(EXTRALIB)
+	$(CC) $(CFLAGS) $(LDFLAGS) -w -o linktest linktest.c ../$(LIBSONAME) $(FEXTRALIB) && echo OK.
+else
+
 ifneq ($(C_COMPILER), LSB)
 	$(CC) $(CFLAGS) $(LDFLAGS) -shared -o ../$(LIBSONAME) \
 	-Wl,--whole-archive $< -Wl,--no-whole-archive \
@@ -152,6 +160,7 @@ else
 	-Wl,--whole-archive $< -Wl,--no-whole-archive \
 	-Wl,-soname,$(INTERNALNAME) $(EXTRALIB)
 	$(FC) $(CFLAGS) $(LDFLAGS) -w -o linktest linktest.c ../$(LIBSONAME) $(FEXTRALIB) && echo OK.
+endif
 endif
 	rm -f linktest
 

exports/dllinit.c

@@ -40,15 +40,25 @@
 
 void gotoblas_init(void);
 void gotoblas_quit(void);
+#if defined(SMP) && defined(USE_TLS)
+void blas_thread_memory_cleanup(void);
+#endif
 
 BOOL APIENTRY DllMain(HINSTANCE hInst, DWORD reason, LPVOID reserved) {
-
-  if (reason == DLL_PROCESS_ATTACH) {
-    gotoblas_init();
-  }
-
-  if (reason == DLL_PROCESS_DETACH) {
-    gotoblas_quit();
+  switch(reason) {
+      case DLL_PROCESS_ATTACH:
+        gotoblas_init();
+        break;
+      case DLL_PROCESS_DETACH:
+        gotoblas_quit();
+        break;
+      case DLL_THREAD_ATTACH:
+        break;
+      case DLL_THREAD_DETACH:
+#if defined(SMP) && defined(USE_TLS)
+        blas_thread_memory_cleanup();
+#endif
+        break;
   }
 
   return TRUE;

getarch.c

@@ -91,6 +91,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include <unistd.h>
 #endif
 
+#if (( defined(__GNUC__)  && __GNUC__   > 6 && defined(__AVX2__)) || (defined(__clang__) && __clang_major__ >= 6))
+#else
+#define NO_AVX512
+#endif
 /* #define FORCE_P2		*/
 /* #define FORCE_KATMAI		*/
 /* #define FORCE_COPPERMINE	*/
@@ -327,6 +331,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #endif
 
 #ifdef FORCE_SKYLAKEX
+#ifdef NO_AVX512
+#define FORCE
+#define FORCE_INTEL
+#define ARCHITECTURE    "X86"
+#define SUBARCHITECTURE "HASWELL"
+#define ARCHCONFIG   "-DHASWELL " \
+		     "-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"
+#define LIBNAME   "haswell"
+#define CORENAME  "HASWELL"
+#else
 #define FORCE
 #define FORCE_INTEL
 #define ARCHITECTURE    "X86"
@@ -340,6 +358,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define LIBNAME   "skylakex"
 #define CORENAME  "SKYLAKEX"
 #endif
+#endif
 
 #ifdef FORCE_ATOM
 #define FORCE
@@ -618,6 +637,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CORENAME  "POWER8"
 #endif
 
+#if defined(FORCE_POWER9) 
+#define FORCE
+#define ARCHITECTURE    "POWER"
+#define SUBARCHITECTURE "POWER9"
+#define SUBDIRNAME      "power"
+#define ARCHCONFIG   "-DPOWER9 " \
+		     "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \
+		     "-DL2_SIZE=4194304 -DL2_LINESIZE=128 " \
+		     "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
+#define LIBNAME   "power9"
+#define CORENAME  "POWER9"
+#endif
 
 #ifdef FORCE_PPCG4
 #define FORCE
@@ -1046,6 +1077,23 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #else
 #endif
 
+#ifdef FORCE_TSV110
+#define FORCE
+#define ARCHITECTURE    "ARM64"
+#define SUBARCHITECTURE "TSV110"
+#define SUBDIRNAME      "arm64"
+#define ARCHCONFIG   "-DTSV110 " \
+       "-DL1_CODE_SIZE=65536  -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=4 " \
+       "-DL1_DATA_SIZE=65536  -DL1_DATA_LINESIZE=64 -DL1_DATA_ASSOCIATIVE=4 " \
+       "-DL2_SIZE=524288 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=8 " \
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
+       "-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
+#define LIBNAME   "tsv110"
+#define CORENAME  "TSV110"
+#else
+#endif
+
+
 #ifdef FORCE_ZARCH_GENERIC
 #define FORCE
 #define ARCHITECTURE    "ZARCH"
@@ -1066,8 +1114,22 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CORENAME  "Z13"
 #endif
 
+#ifdef FORCE_Z14
+#define FORCE
+#define ARCHITECTURE    "ZARCH"
+#define SUBARCHITECTURE "Z14"
+#define ARCHCONFIG   "-DZ14 " \
+       "-DDTB_DEFAULT_ENTRIES=64"
+#define LIBNAME   "z14"
+#define CORENAME  "Z14"
+#endif
+
 #ifndef FORCE
 
+#ifdef USER_TARGET
+#error "The TARGET specified on the command line or in Makefile.rule is not supported. Please choose a target from TargetList.txt"
+#endif
+
 #if defined(__powerpc__) || defined(__powerpc) || defined(powerpc) || \
     defined(__PPC__) || defined(PPC) || defined(_POWER) || defined(__POWERPC__)
 #ifndef POWER

interface/CMakeLists.txt

@@ -12,6 +12,7 @@ set(BLAS1_REAL_ONLY_SOURCES
   rotm.c rotmg.c # N.B. these do not have complex counterparts
   rot.c
   asum.c
+  sum.c
 )
 
 # these will have 'z' prepended for the complex version
@@ -124,6 +125,7 @@ foreach (float_type ${FLOAT_TYPES})
     GenerateNamedObjects("max.c" "USE_ABS;USE_MIN" "scamin" ${CBLAS_FLAG} "" "" true "COMPLEX")
     GenerateNamedObjects("max.c" "USE_ABS" "scamax" ${CBLAS_FLAG} "" "" true "COMPLEX")
     GenerateNamedObjects("asum.c" "" "scasum" ${CBLAS_FLAG} "" "" true "COMPLEX")
+    GenerateNamedObjects("sum.c" "" "scsum" ${CBLAS_FLAG} "" "" true "COMPLEX")
   endif ()
   if (${float_type} STREQUAL "ZCOMPLEX")
     GenerateNamedObjects("zscal.c" "SSCAL" "dscal" ${CBLAS_FLAG} "" "" false "ZCOMPLEX")
@@ -132,6 +134,7 @@ foreach (float_type ${FLOAT_TYPES})
     GenerateNamedObjects("max.c" "USE_ABS;USE_MIN" "dzamin" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
     GenerateNamedObjects("max.c" "USE_ABS" "dzamax" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
     GenerateNamedObjects("asum.c" "" "dzasum" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
+    GenerateNamedObjects("sum.c" "" "dzsum" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
   endif ()
 endforeach ()
 

interface/Makefile

@@ -25,7 +25,7 @@ SBLAS1OBJS    = \
 		saxpy.$(SUFFIX) sswap.$(SUFFIX) \
 		scopy.$(SUFFIX) sscal.$(SUFFIX) \
 		sdot.$(SUFFIX) sdsdot.$(SUFFIX) dsdot.$(SUFFIX) \
-		sasum.$(SUFFIX) snrm2.$(SUFFIX) \
+		sasum.$(SUFFIX) ssum.$(SUFFIX) snrm2.$(SUFFIX) \
 		smax.$(SUFFIX) samax.$(SUFFIX) ismax.$(SUFFIX) isamax.$(SUFFIX) \
 		smin.$(SUFFIX) samin.$(SUFFIX) ismin.$(SUFFIX) isamin.$(SUFFIX) \
 		srot.$(SUFFIX) srotg.$(SUFFIX) srotm.$(SUFFIX) srotmg.$(SUFFIX) \
@@ -51,7 +51,7 @@ DBLAS1OBJS    = \
 		daxpy.$(SUFFIX) dswap.$(SUFFIX) \
 		dcopy.$(SUFFIX) dscal.$(SUFFIX) \
 		ddot.$(SUFFIX) \
-		dasum.$(SUFFIX) dnrm2.$(SUFFIX) \
+		dasum.$(SUFFIX) dsum.$(SUFFIX) dnrm2.$(SUFFIX) \
 		dmax.$(SUFFIX) damax.$(SUFFIX) idmax.$(SUFFIX) idamax.$(SUFFIX) \
 		dmin.$(SUFFIX) damin.$(SUFFIX) idmin.$(SUFFIX) idamin.$(SUFFIX) \
 		drot.$(SUFFIX) drotg.$(SUFFIX) drotm.$(SUFFIX) drotmg.$(SUFFIX) \
@@ -76,7 +76,7 @@ CBLAS1OBJS    = \
 		caxpy.$(SUFFIX) caxpyc.$(SUFFIX) cswap.$(SUFFIX) \
 		ccopy.$(SUFFIX) cscal.$(SUFFIX) csscal.$(SUFFIX) \
 		cdotc.$(SUFFIX)  cdotu.$(SUFFIX) \
-		scasum.$(SUFFIX) scnrm2.$(SUFFIX) \
+		scasum.$(SUFFIX) scsum.$(SUFFIX) scnrm2.$(SUFFIX) \
 		scamax.$(SUFFIX) icamax.$(SUFFIX) \
 		scamin.$(SUFFIX) icamin.$(SUFFIX) \
 		csrot.$(SUFFIX) crotg.$(SUFFIX) \
@@ -105,7 +105,7 @@ ZBLAS1OBJS    = \
 		zaxpy.$(SUFFIX) zaxpyc.$(SUFFIX) zswap.$(SUFFIX) \
 		zcopy.$(SUFFIX) zscal.$(SUFFIX) zdscal.$(SUFFIX) \
 		zdotc.$(SUFFIX)  zdotu.$(SUFFIX) \
-		dzasum.$(SUFFIX)  dznrm2.$(SUFFIX) \
+		dzasum.$(SUFFIX) dzsum.$(SUFFIX) dznrm2.$(SUFFIX) \
 		dzamax.$(SUFFIX) izamax.$(SUFFIX) \
 		dzamin.$(SUFFIX) izamin.$(SUFFIX) \
 		zdrot.$(SUFFIX) zrotg.$(SUFFIX) \
@@ -146,7 +146,7 @@ QBLAS1OBJS    = \
 		qaxpy.$(SUFFIX) qswap.$(SUFFIX) \
 		qcopy.$(SUFFIX) qscal.$(SUFFIX) \
 		qdot.$(SUFFIX) \
-		qasum.$(SUFFIX) qnrm2.$(SUFFIX) \
+		qasum.$(SUFFIX) qsum.$(SUFFIX) qnrm2.$(SUFFIX) \
 		qmax.$(SUFFIX) qamax.$(SUFFIX) iqmax.$(SUFFIX) iqamax.$(SUFFIX) \
 		qmin.$(SUFFIX) qamin.$(SUFFIX) iqmin.$(SUFFIX) iqamin.$(SUFFIX) \
 		qrot.$(SUFFIX) qrotg.$(SUFFIX) qrotm.$(SUFFIX) qrotmg.$(SUFFIX) \
@@ -168,7 +168,7 @@ XBLAS1OBJS    = \
 		xaxpy.$(SUFFIX) xaxpyc.$(SUFFIX) xswap.$(SUFFIX) \
 		xcopy.$(SUFFIX) xscal.$(SUFFIX) xqscal.$(SUFFIX) \
 		xdotc.$(SUFFIX)  xdotu.$(SUFFIX) \
-		qxasum.$(SUFFIX) qxnrm2.$(SUFFIX) \
+		qxasum.$(SUFFIX) qxsum.$(SUFFIX) qxnrm2.$(SUFFIX) \
 		qxamax.$(SUFFIX) ixamax.$(SUFFIX) \
 		qxamin.$(SUFFIX) ixamin.$(SUFFIX) \
 		xqrot.$(SUFFIX) xrotg.$(SUFFIX) \
@@ -203,7 +203,7 @@ ifdef QUAD_PRECISION
 QBLAS1OBJS    = \
 		qaxpy.$(SUFFIX) qswap.$(SUFFIX) \
 		qcopy.$(SUFFIX) qscal.$(SUFFIX) \
-		qasum.$(SUFFIX) qnrm2.$(SUFFIX) \
+		qasum.$(SUFFIX) qsum.$(SUFFIX) qnrm2.$(SUFFIX) \
 		qmax.$(SUFFIX) qamax.$(SUFFIX) iqmax.$(SUFFIX) iqamax.$(SUFFIX) \
 		qmin.$(SUFFIX) qamin.$(SUFFIX) iqmin.$(SUFFIX) iqamin.$(SUFFIX) \
 		qrot.$(SUFFIX) qrotg.$(SUFFIX) qrotm.$(SUFFIX) qrotmg.$(SUFFIX) \
@@ -224,7 +224,7 @@ QBLAS3OBJS    = \
 XBLAS1OBJS    = \
 		xaxpy.$(SUFFIX) xaxpyc.$(SUFFIX) xswap.$(SUFFIX) \
 		xcopy.$(SUFFIX) xscal.$(SUFFIX) xqscal.$(SUFFIX) \
-		qxasum.$(SUFFIX) qxnrm2.$(SUFFIX) \
+		qxasum.$(SUFFIX) qxsum.$(SUFFIX) qxnrm2.$(SUFFIX) \
 		qxamax.$(SUFFIX) ixamax.$(SUFFIX) \
 		qxamin.$(SUFFIX) ixamin.$(SUFFIX) \
 		xqrot.$(SUFFIX) xrotg.$(SUFFIX) \
@@ -263,7 +263,8 @@ CSBLAS1OBJS   = \
 	cblas_isamax.$(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.$(SUFFIX)
+	cblas_sscal.$(SUFFIX) cblas_sswap.$(SUFFIX) cblas_snrm2.$(SUFFIX) cblas_saxpby.$(SUFFIX) \
+	cblas_ismin.$(SUFFIX) cblas_ismax.$(SUFFIX) cblas_ssum.$(SUFFIX)
 
 CSBLAS2OBJS   = \
 	cblas_sgemv.$(SUFFIX) cblas_sger.$(SUFFIX) cblas_ssymv.$(SUFFIX) cblas_strmv.$(SUFFIX) \
@@ -280,7 +281,8 @@ CDBLAS1OBJS   = \
 	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)
+	cblas_dscal.$(SUFFIX) cblas_dswap.$(SUFFIX) cblas_dnrm2.$(SUFFIX) cblas_daxpby.$(SUFFIX) \
+	cblas_idmin.$(SUFFIX) cblas_idmax.$(SUFFIX) cblas_dsum.$(SUFFIX)
 
 CDBLAS2OBJS   = \
 	cblas_dgemv.$(SUFFIX) cblas_dger.$(SUFFIX) cblas_dsymv.$(SUFFIX) cblas_dtrmv.$(SUFFIX) \
@@ -300,7 +302,8 @@ CCBLAS1OBJS   = \
 	cblas_cdotc_sub.$(SUFFIX) cblas_cdotu_sub.$(SUFFIX) \
 	cblas_cscal.$(SUFFIX) cblas_csscal.$(SUFFIX) \
 	cblas_cswap.$(SUFFIX) cblas_scnrm2.$(SUFFIX) \
-	cblas_caxpby.$(SUFFIX)
+	cblas_caxpby.$(SUFFIX) \
+	cblas_icmin.$(SUFFIX) cblas_icmax.$(SUFFIX) cblas_scsum.$(SUFFIX)
 
 CCBLAS2OBJS   = \
 	cblas_cgemv.$(SUFFIX) cblas_cgerc.$(SUFFIX) cblas_cgeru.$(SUFFIX) \
@@ -326,7 +329,9 @@ CZBLAS1OBJS   = \
 	cblas_zdotc_sub.$(SUFFIX) cblas_zdotu_sub.$(SUFFIX) \
 	cblas_zscal.$(SUFFIX) cblas_zdscal.$(SUFFIX) \
 	cblas_zswap.$(SUFFIX) cblas_dznrm2.$(SUFFIX) \
-	cblas_zaxpby.$(SUFFIX)
+	cblas_zaxpby.$(SUFFIX) \
+	cblas_izmin.$(SUFFIX) cblas_izmax.$(SUFFIX) cblas_dzsum.$(SUFFIX)
+
 
 CZBLAS2OBJS   = \
 	cblas_zgemv.$(SUFFIX) cblas_zgerc.$(SUFFIX) cblas_zgeru.$(SUFFIX) \
@@ -560,6 +565,24 @@ dzasum.$(SUFFIX) dzasum.$(PSUFFIX) : asum.c
 qxasum.$(SUFFIX) qxasum.$(PSUFFIX) : asum.c
 	$(CC) $(CFLAGS) -c $< -o $(@F)
 
+ssum.$(SUFFIX) ssum.$(PSUFFIX) : sum.c
+	$(CC) $(CFLAGS) -c $< -o $(@F)
+
+dsum.$(SUFFIX) dsum.$(PSUFFIX) : sum.c
+	$(CC) $(CFLAGS) -c $< -o $(@F)
+
+qsum.$(SUFFIX) qsum.$(PSUFFIX) : sum.c
+	$(CC) $(CFLAGS) -c $< -o $(@F)
+
+scsum.$(SUFFIX) scsum.$(PSUFFIX) : sum.c
+	$(CC) $(CFLAGS) -c $< -o $(@F)
+
+dzsum.$(SUFFIX) dzsum.$(PSUFFIX) : sum.c
+	$(CC) $(CFLAGS) -c $< -o $(@F)
+
+qxsum.$(SUFFIX) qxsum.$(PSUFFIX) : sum.c
+	$(CC) $(CFLAGS) -c $< -o $(@F)
+
 snrm2.$(SUFFIX) snrm2.$(PSUFFIX) : nrm2.c
 	$(CC) $(CFLAGS) -c $< -o $(@F)
 
@@ -1383,6 +1406,18 @@ cblas_ismin.$(SUFFIX) cblas_ismin.$(PSUFFIX) : imax.c
 cblas_idmin.$(SUFFIX) cblas_idmin.$(PSUFFIX) : imax.c
 	$(CC) $(CFLAGS) -DCBLAS -c -UUSE_ABS -DUSE_MIN $< -o $(@F)
 
+cblas_icmax.$(SUFFIX) cblas_icmax.$(PSUFFIX) : imax.c
+	$(CC) $(CFLAGS) -DCBLAS -c -UUSE_ABS -UUSE_MIN $< -o $(@F)
+
+cblas_izmax.$(SUFFIX) cblas_izmax.$(PSUFFIX) : imax.c
+	$(CC) $(CFLAGS) -DCBLAS -c -UUSE_ABS -UUSE_MIN $< -o $(@F)
+
+cblas_icmin.$(SUFFIX) cblas_icmin.$(PSUFFIX) : imax.c
+	$(CC) $(CFLAGS) -DCBLAS -c -UUSE_ABS -DUSE_MIN $< -o $(@F)
+
+cblas_izmin.$(SUFFIX) cblas_izmin.$(PSUFFIX) : imax.c
+	$(CC) $(CFLAGS) -DCBLAS -c -UUSE_ABS -DUSE_MIN $< -o $(@F)
+
 cblas_sasum.$(SUFFIX) cblas_sasum.$(PSUFFIX) : asum.c
 	$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)
 
@@ -1395,6 +1430,18 @@ cblas_scasum.$(SUFFIX) cblas_scasum.$(PSUFFIX) : asum.c
 cblas_dzasum.$(SUFFIX) cblas_dzasum.$(PSUFFIX) : asum.c
 	$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)
 
+cblas_ssum.$(SUFFIX) cblas_ssum.$(PSUFFIX) : sum.c
+	$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)
+
+cblas_dsum.$(SUFFIX) cblas_dsum.$(PSUFFIX) : sum.c
+	$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)
+
+cblas_scsum.$(SUFFIX) cblas_scsum.$(PSUFFIX) : sum.c
+	$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)
+
+cblas_dzsum.$(SUFFIX) cblas_dzsum.$(PSUFFIX) : sum.c
+	$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)
+
 cblas_sdsdot.$(SUFFIX) cblas_sdsdot.$(PSUFFIX) : sdsdot.c
 	$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)
 
@@ -1402,7 +1449,7 @@ cblas_dsdot.$(SUFFIX) cblas_dsdot.$(PSUFFIX) : dsdot.c
 	$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)
 
 cblas_sdot.$(SUFFIX) cblas_sdot.$(PSUFFIX) : dot.c
-	$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)
+		$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)
 
 cblas_ddot.$(SUFFIX) cblas_ddot.$(PSUFFIX) : dot.c
 	$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)

interface/sum.c

@@ -0,0 +1,97 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* All rights reserved.                                              */
+/*                                                                   */
+/* Redistribution and use in source and binary forms, with or        */
+/* without modification, are permitted provided that the following   */
+/* conditions are met:                                               */
+/*                                                                   */
+/*   1. Redistributions of source code must retain the above         */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer.                                                  */
+/*                                                                   */
+/*   2. Redistributions in binary form must reproduce the above      */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer in the documentation and/or other materials       */
+/*      provided with the distribution.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
+/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
+/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
+/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
+/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
+/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
+/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
+/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
+/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
+/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
+/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
+/*    POSSIBILITY OF SUCH DAMAGE.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include <stdio.h>
+#include "common.h"
+#ifdef FUNCTION_PROFILE
+#include "functable.h"
+#endif
+
+#ifndef CBLAS
+
+FLOATRET NAME(blasint *N, FLOAT *x, blasint *INCX){
+
+  BLASLONG n    = *N;
+  BLASLONG incx = *INCX;
+  FLOATRET ret;
+
+  PRINT_DEBUG_NAME;
+
+  if (n <= 0) return 0;
+
+  IDEBUG_START;
+
+  FUNCTION_PROFILE_START();
+
+  ret = (FLOATRET)SUM_K(n, x, incx);
+
+  FUNCTION_PROFILE_END(COMPSIZE, n, n);
+
+  IDEBUG_END;
+
+  return ret;
+}
+
+#else
+#ifdef COMPLEX
+FLOAT CNAME(blasint n, void *vx, blasint incx){
+  FLOAT *x = (FLOAT*) vx;
+#else
+FLOAT CNAME(blasint n, FLOAT *x, blasint incx){
+#endif
+
+  FLOAT ret;
+
+  PRINT_DEBUG_CNAME;
+
+  if (n <= 0) return 0;
+
+  IDEBUG_START;
+
+  FUNCTION_PROFILE_START();
+
+  ret = SUM_K(n, x, incx);
+
+  FUNCTION_PROFILE_END(COMPSIZE, n, n);
+
+  IDEBUG_END;
+
+  return ret;
+}
+
+#endif

interface/trmv.c

@@ -218,11 +218,8 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
   buffer = (FLOAT *)blas_memory_alloc(1);
 
 #ifdef SMP
-/*  nthreads = num_cpu_avail(2);
+  nthreads = num_cpu_avail(2);
 
-FIXME trmv_thread was found to be broken, see issue 1332 */
-  nthreads = 1;
-  
   if (nthreads == 1) {
 #endif
 

interface/trsm.c

@@ -81,6 +81,12 @@
 #endif
 #endif
 
+#ifndef COMPLEX
+#define SMP_FACTOR 256
+#else
+#define SMP_FACTOR 128
+#endif
+
 static int (*trsm[])(blas_arg_t *, BLASLONG *, BLASLONG *, FLOAT *, FLOAT *, BLASLONG) = {
 #ifndef TRMM
   TRSM_LNUU, TRSM_LNUN, TRSM_LNLU, TRSM_LNLN,
@@ -198,7 +204,7 @@ void NAME(char *SIDE, char *UPLO, char *TRANS, char *DIAG,
   if (side  < 0)                info =  1;
 
   if (info != 0) {
-    BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
+    BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)-1);
     return;
   }
 
@@ -366,11 +372,15 @@ void CNAME(enum CBLAS_ORDER order,
   mode |= (trans << BLAS_TRANSA_SHIFT);
   mode |= (side  << BLAS_RSIDE_SHIFT);
 
-  if ( args.m < 2*GEMM_MULTITHREAD_THRESHOLD )
+/*
+  if ( args.m < 2 * GEMM_MULTITHREAD_THRESHOLD )
 	args.nthreads = 1;
   else
-	if ( args.n < 2*GEMM_MULTITHREAD_THRESHOLD )
+	if ( args.n < 2 * GEMM_MULTITHREAD_THRESHOLD )
 		args.nthreads = 1;
+*/
+  if ( args.m * args.n < SMP_FACTOR * GEMM_MULTITHREAD_THRESHOLD)
+	args.nthreads = 1;
   else
 	args.nthreads = num_cpu_avail(3);
 		

interface/ztrmv.c

@@ -239,9 +239,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
   } else
       nthreads = 1;
 
-/* FIXME TRMV multithreading appears to be broken, see issue 1332*/
-  nthreads = 1;
-
   if(nthreads > 1) {
     buffer_size = n > 16 ? 0 : n * 4 + 40;
   }

kernel/CMakeLists.txt

@@ -65,6 +65,7 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
       GenerateNamedObjects("${KERNELDIR}/${${float_char}SCALKERNEL}" "" "scal_k" false "" "" false ${float_type})
       GenerateNamedObjects("${KERNELDIR}/${${float_char}SWAPKERNEL}" "" "swap_k" false "" "" false ${float_type})
       GenerateNamedObjects("${KERNELDIR}/${${float_char}AXPBYKERNEL}" "" "axpby_k" false "" "" false ${float_type})
+      GenerateNamedObjects("${KERNELDIR}/${${float_char}SUMKERNEL}" "" "sum_k" false "" "" false ${float_type})
 
       if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
         GenerateNamedObjects("${KERNELDIR}/${${float_char}AXPYKERNEL}" "CONJ" "axpyc_k" false "" "" false ${float_type})

kernel/Makefile.L1

@@ -340,6 +340,32 @@ ifndef XSCALKERNEL
 XSCALKERNEL = zscal.S
 endif
 
+### SUM ###
+
+ifndef SSUMKERNEL
+SSUMKERNEL =  sum.S
+endif
+
+ifndef DSUMKERNEL
+DSUMKERNEL =  sum.S
+endif
+
+ifndef CSUMKERNEL
+CSUMKERNEL = zsum.S
+endif
+
+ifndef ZSUMKERNEL
+ZSUMKERNEL = zsum.S
+endif
+
+ifndef QSUMKERNEL
+QSUMKERNEL =  sum.S
+endif
+
+ifndef XSUMKERNEL
+XSUMKERNEL = zsum.S
+endif
+
 ### SWAP ###
 
 ifndef SSWAPKERNEL
@@ -453,7 +479,7 @@ endif
 SBLASOBJS	+= \
 	 samax_k$(TSUFFIX).$(SUFFIX)  samin_k$(TSUFFIX).$(SUFFIX)  smax_k$(TSUFFIX).$(SUFFIX)  smin_k$(TSUFFIX).$(SUFFIX) \
 	isamax_k$(TSUFFIX).$(SUFFIX) isamin_k$(TSUFFIX).$(SUFFIX) ismax_k$(TSUFFIX).$(SUFFIX) ismin_k$(TSUFFIX).$(SUFFIX) \
-	sasum_k$(TSUFFIX).$(SUFFIX) saxpy_k$(TSUFFIX).$(SUFFIX) scopy_k$(TSUFFIX).$(SUFFIX) \
+	sasum_k$(TSUFFIX).$(SUFFIX) ssum_k$(TSUFFIX).$(SUFFIX) saxpy_k$(TSUFFIX).$(SUFFIX) scopy_k$(TSUFFIX).$(SUFFIX) \
 	sdot_k$(TSUFFIX).$(SUFFIX) sdsdot_k$(TSUFFIX).$(SUFFIX) dsdot_k$(TSUFFIX).$(SUFFIX) \
 	snrm2_k$(TSUFFIX).$(SUFFIX) srot_k$(TSUFFIX).$(SUFFIX) sscal_k$(TSUFFIX).$(SUFFIX) sswap_k$(TSUFFIX).$(SUFFIX) \
 	saxpby_k$(TSUFFIX).$(SUFFIX)
@@ -463,31 +489,32 @@ DBLASOBJS	+= \
 	idamax_k$(TSUFFIX).$(SUFFIX) idamin_k$(TSUFFIX).$(SUFFIX) idmax_k$(TSUFFIX).$(SUFFIX) idmin_k$(TSUFFIX).$(SUFFIX) \
 	dasum_k$(TSUFFIX).$(SUFFIX) daxpy_k$(TSUFFIX).$(SUFFIX) dcopy_k$(TSUFFIX).$(SUFFIX) ddot_k$(TSUFFIX).$(SUFFIX) \
 	dnrm2_k$(TSUFFIX).$(SUFFIX) drot_k$(TSUFFIX).$(SUFFIX) dscal_k$(TSUFFIX).$(SUFFIX) dswap_k$(TSUFFIX).$(SUFFIX) \
-	daxpby_k$(TSUFFIX).$(SUFFIX)
+	daxpby_k$(TSUFFIX).$(SUFFIX) dsum_k$(TSUFFIX).$(SUFFIX)
 
 QBLASOBJS	+= \
 	 qamax_k$(TSUFFIX).$(SUFFIX)  qamin_k$(TSUFFIX).$(SUFFIX)  qmax_k$(TSUFFIX).$(SUFFIX)  qmin_k$(TSUFFIX).$(SUFFIX) \
 	iqamax_k$(TSUFFIX).$(SUFFIX) iqamin_k$(TSUFFIX).$(SUFFIX) iqmax_k$(TSUFFIX).$(SUFFIX) iqmin_k$(TSUFFIX).$(SUFFIX) \
 	qasum_k$(TSUFFIX).$(SUFFIX) qaxpy_k$(TSUFFIX).$(SUFFIX) qcopy_k$(TSUFFIX).$(SUFFIX) qdot_k$(TSUFFIX).$(SUFFIX) \
-	qnrm2_k$(TSUFFIX).$(SUFFIX) qrot_k$(TSUFFIX).$(SUFFIX) qscal_k$(TSUFFIX).$(SUFFIX) qswap_k$(TSUFFIX).$(SUFFIX)
+	qnrm2_k$(TSUFFIX).$(SUFFIX) qrot_k$(TSUFFIX).$(SUFFIX) qscal_k$(TSUFFIX).$(SUFFIX) qswap_k$(TSUFFIX).$(SUFFIX) \
+	qsum_k$(TSUFFIX).$(SUFFIX)
 
 CBLASOBJS	+= \
 	camax_k$(TSUFFIX).$(SUFFIX)  camin_k$(TSUFFIX).$(SUFFIX) icamax_k$(TSUFFIX).$(SUFFIX) icamin_k$(TSUFFIX).$(SUFFIX) \
 	casum_k$(TSUFFIX).$(SUFFIX) caxpy_k$(TSUFFIX).$(SUFFIX) caxpyc_k$(TSUFFIX).$(SUFFIX) ccopy_k$(TSUFFIX).$(SUFFIX) \
 	cdotc_k$(TSUFFIX).$(SUFFIX) cdotu_k$(TSUFFIX).$(SUFFIX) cnrm2_k$(TSUFFIX).$(SUFFIX) csrot_k$(TSUFFIX).$(SUFFIX) \
-	cscal_k$(TSUFFIX).$(SUFFIX) cswap_k$(TSUFFIX).$(SUFFIX) caxpby_k$(TSUFFIX).$(SUFFIX)
+	cscal_k$(TSUFFIX).$(SUFFIX) cswap_k$(TSUFFIX).$(SUFFIX) caxpby_k$(TSUFFIX).$(SUFFIX) csum_k$(TSUFFIX).$(SUFFIX)
 
 ZBLASOBJS	+= \
 	 zamax_k$(TSUFFIX).$(SUFFIX)  zamin_k$(TSUFFIX).$(SUFFIX) izamax_k$(TSUFFIX).$(SUFFIX) izamin_k$(TSUFFIX).$(SUFFIX) \
 	zasum_k$(TSUFFIX).$(SUFFIX) zaxpy_k$(TSUFFIX).$(SUFFIX) zaxpyc_k$(TSUFFIX).$(SUFFIX) zcopy_k$(TSUFFIX).$(SUFFIX) \
 	zdotc_k$(TSUFFIX).$(SUFFIX) zdotu_k$(TSUFFIX).$(SUFFIX) znrm2_k$(TSUFFIX).$(SUFFIX) zdrot_k$(TSUFFIX).$(SUFFIX) \
-	zscal_k$(TSUFFIX).$(SUFFIX) zswap_k$(TSUFFIX).$(SUFFIX) zaxpby_k$(TSUFFIX).$(SUFFIX)
+	zscal_k$(TSUFFIX).$(SUFFIX) zswap_k$(TSUFFIX).$(SUFFIX) zaxpby_k$(TSUFFIX).$(SUFFIX) zsum_k$(TSUFFIX).$(SUFFIX)
 
 XBLASOBJS	+= \
 	 xamax_k$(TSUFFIX).$(SUFFIX)  xamin_k$(TSUFFIX).$(SUFFIX) ixamax_k$(TSUFFIX).$(SUFFIX) ixamin_k$(TSUFFIX).$(SUFFIX) \
 	xasum_k$(TSUFFIX).$(SUFFIX) xaxpy_k$(TSUFFIX).$(SUFFIX) xaxpyc_k$(TSUFFIX).$(SUFFIX) xcopy_k$(TSUFFIX).$(SUFFIX) \
 	xdotc_k$(TSUFFIX).$(SUFFIX) xdotu_k$(TSUFFIX).$(SUFFIX) xnrm2_k$(TSUFFIX).$(SUFFIX) xqrot_k$(TSUFFIX).$(SUFFIX) \
-	xscal_k$(TSUFFIX).$(SUFFIX) xswap_k$(TSUFFIX).$(SUFFIX)
+	xscal_k$(TSUFFIX).$(SUFFIX) xswap_k$(TSUFFIX).$(SUFFIX) xsum_k$(TSUFFIX).$(SUFFIX)
 
 ### AMAX ###
 
@@ -617,7 +644,7 @@ $(KDIR)idmin_k$(TSUFFIX).$(SUFFIX)  $(KDIR)idmin_k$(TPSUFFIX).$(PSUFFIX)  : $(KE
 $(KDIR)iqmin_k$(TSUFFIX).$(SUFFIX)  $(KDIR)iqmin_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(IQMINKERNEL)
 	$(CC) -c $(CFLAGS) -UCOMPLEX -DXDOUBLE -UUSE_ABS -DUSE_MIN $< -o $@
 
-
+### ASUM ###
 $(KDIR)sasum_k$(TSUFFIX).$(SUFFIX)   $(KDIR)sasum_k$(TPSUFFIX).$(PSUFFIX)   : $(KERNELDIR)/$(SASUMKERNEL)
 	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE $< -o $@
 
@@ -636,6 +663,26 @@ $(KDIR)zasum_k$(TSUFFIX).$(SUFFIX)  $(KDIR)zasum_k$(TPSUFFIX).$(PSUFFIX)  : $(KE
 $(KDIR)xasum_k$(TSUFFIX).$(SUFFIX)  $(KDIR)xasum_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(XASUMKERNEL)
 	$(CC) -c $(CFLAGS) -DCOMPLEX -DXDOUBLE $< -o $@
 
+### SUM ###
+$(KDIR)ssum_k$(TSUFFIX).$(SUFFIX)   $(KDIR)ssum_k$(TPSUFFIX).$(PSUFFIX)   : $(KERNELDIR)/$(SSUMKERNEL)
+	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE $< -o $@
+
+$(KDIR)dsum_k$(TSUFFIX).$(SUFFIX)   $(KDIR)dsum_k$(TPSUFFIX).$(PSUFFIX)   : $(KERNELDIR)/$(DSUMKERNEL)
+	$(CC) -c $(CFLAGS) -UCOMPLEX -DDOUBLE $< -o $@
+
+$(KDIR)qsum_k$(TSUFFIX).$(SUFFIX)   $(KDIR)qsum_k$(TPSUFFIX).$(PSUFFIX)   : $(KERNELDIR)/$(QSUMKERNEL)
+	$(CC) -c $(CFLAGS) -UCOMPLEX -DXDOUBLE $< -o $@
+
+$(KDIR)csum_k$(TSUFFIX).$(SUFFIX)  $(KDIR)csum_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(CSUMKERNEL)
+	$(CC) -c $(CFLAGS) -DCOMPLEX -UDOUBLE $< -o $@
+
+$(KDIR)zsum_k$(TSUFFIX).$(SUFFIX)  $(KDIR)zsum_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(ZSUMKERNEL)
+	$(CC) -c $(CFLAGS) -DCOMPLEX -DDOUBLE $< -o $@
+
+$(KDIR)xsum_k$(TSUFFIX).$(SUFFIX)  $(KDIR)xsum_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(XSUMKERNEL)
+	$(CC) -c $(CFLAGS) -DCOMPLEX -DXDOUBLE $< -o $@
+
+### AXPY ###
 $(KDIR)saxpy_k$(TSUFFIX).$(SUFFIX)  $(KDIR)saxpy_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(SAXPYKERNEL)
 	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE $< -o $@
 

kernel/Makefile.L3

@@ -24,7 +24,7 @@ ifeq ($(TARGET), LOONGSON3B)
 USE_TRMM = 1
 endif
 
-ifeq ($(TARGET), GENERIC)
+ifeq ($(CORE), GENERIC)
 USE_TRMM = 1
 endif
 
@@ -44,10 +44,18 @@ ifeq ($(CORE), POWER8)
 USE_TRMM = 1
 endif
 
+ifeq ($(CORE), POWER9)
+USE_TRMM = 1
+endif
+
 ifeq ($(ARCH), zarch)
 USE_TRMM = 1
 endif
 
+ifeq ($(CORE), Z14)
+USE_TRMM = 1
+endif
+
 
 
 

kernel/alpha/sum.S

@@ -0,0 +1,206 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* All rights reserved.                                              */
+/*                                                                   */
+/* Redistribution and use in source and binary forms, with or        */
+/* without modification, are permitted provided that the following   */
+/* conditions are met:                                               */
+/*                                                                   */
+/*   1. Redistributions of source code must retain the above         */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer.                                                  */
+/*                                                                   */
+/*   2. Redistributions in binary form must reproduce the above      */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer in the documentation and/or other materials       */
+/*      provided with the distribution.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
+/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
+/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
+/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
+/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
+/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
+/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
+/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
+/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
+/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
+/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
+/*    POSSIBILITY OF SUCH DAMAGE.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+#include "version.h"
+
+#define PREFETCHSIZE	88
+
+#define N	$16
+#define X	$17
+#define INCX	$18
+#define I	$19
+
+#define s0	$f0
+#define s1	$f1
+#define s2	$f10
+#define s3	$f11
+
+#define a0	$f12
+#define a1	$f13
+#define a2	$f14
+#define a3	$f15
+#define a4	$f16
+#define a5	$f17
+#define a6	$f18
+#define a7	$f19
+
+#define t0	$f20
+#define t1	$f21
+#define t2	$f22
+#define t3	$f23
+
+	PROLOGUE
+	PROFCODE
+
+	fclr	s0
+	unop
+	fclr	t0
+	ble	N,  $L999
+
+	sra	N, 3, I
+	fclr	s1
+	fclr	s2
+	ble	I, $L15
+
+	LD	a0,  0 * SIZE(X)
+	fclr	t1
+	SXADDQ	INCX, X, X
+	fclr	t2
+
+	LD	a1,  0 * SIZE(X)
+	fclr	t3
+	SXADDQ	INCX, X, X
+	fclr	s3
+
+	LD	a2,  0 * SIZE(X)
+	SXADDQ	INCX, X, X
+	LD	a3,  0 * SIZE(X)
+	SXADDQ	INCX, X, X
+
+	LD	a4,  0 * SIZE(X)
+	SXADDQ	INCX, X, X
+	LD	a5,  0 * SIZE(X)
+	SXADDQ	INCX, X, X
+
+	lda	I,  -1(I)
+	ble	I, $L13
+	.align 4
+
+$L12:
+	ADD	s0, t0, s0
+	ldl	$31, PREFETCHSIZE * 2 * SIZE(X)
+	fmov	a0, t0
+	lda	I,  -1(I)
+
+	ADD	s1, t1, s1
+	LD	a6,  0 * SIZE(X)
+	fmov	a1, t1
+	SXADDQ	INCX, X, X
+
+	ADD	s2, t2, s2
+	LD	a7,  0 * SIZE(X)
+	fmov	a2, t2
+	SXADDQ	INCX, X, X
+
+	ADD	s3, t3, s3
+	LD	a0,  0 * SIZE(X)
+	fmov	a3, t3
+	SXADDQ	INCX, X, X
+
+	ADD	s0, t0, s0
+	LD	a1,  0 * SIZE(X)
+	fmov	a4, t0
+	SXADDQ	INCX, X, X
+
+	ADD	s1, t1, s1
+	LD	a2,  0 * SIZE(X)
+	fmov	a5, t1
+	SXADDQ	INCX, X, X
+
+	ADD	s2, t2, s2
+	LD	a3,  0 * SIZE(X)
+	fmov	a6, t2
+	SXADDQ	INCX, X, X
+
+	ADD	s3, t3, s3
+	LD	a4,  0 * SIZE(X)
+	fmov	a7, t3
+	SXADDQ	INCX, X, X
+
+	LD	a5,  0 * SIZE(X)
+	unop
+	SXADDQ	INCX, X, X
+	bne	I, $L12
+	.align 4
+
+$L13:
+	ADD	s0, t0, s0
+	LD	a6,  0 * SIZE(X)
+	fmov	a0, t0
+	SXADDQ	INCX, X, X
+
+	ADD	s1, t1, s1
+	LD	a7,  0 * SIZE(X)
+	fmov	a1, t1
+	SXADDQ	INCX, X, X
+
+	ADD	s2, t2, s2
+	fmov	a2, t2
+	ADD	s3, t3, s3
+	fmov	a3, t3
+
+	ADD	s0, t0, s0
+	fmov	a4, t0
+	ADD	s1, t1, s1
+	fmov	a5, t1
+	ADD	s2, t2, s2
+	fmov	a6, t2
+	ADD	s3, t3, s3
+	fmov	a7, t3
+
+	ADD	s1, t1, s1
+	ADD	s2, t2, s2
+	ADD	s3, t3, s3
+
+	ADD	s0, s1, s0
+	ADD	s2, s3, s2
+	.align 4
+
+$L15:
+	and	N, 7, I
+	ADD	s0, s2, s0
+	unop
+	ble	I, $L999
+	.align 4
+
+$L17:
+	ADD	s0, t0, s0
+	LD	a0,  0 * SIZE(X)
+	SXADDQ	INCX, X, X
+	fmov	a0, t0
+
+	lda	I,  -1(I)
+	bne	I, $L17
+	.align 4
+
+$L999:
+	ADD	s0, t0, s0
+	ret
+	EPILOGUE

kernel/alpha/zsum.S

@@ -0,0 +1,208 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* All rights reserved.                                              */
+/*                                                                   */
+/* Redistribution and use in source and binary forms, with or        */
+/* without modification, are permitted provided that the following   */
+/* conditions are met:                                               */
+/*                                                                   */
+/*   1. Redistributions of source code must retain the above         */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer.                                                  */
+/*                                                                   */
+/*   2. Redistributions in binary form must reproduce the above      */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer in the documentation and/or other materials       */
+/*      provided with the distribution.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
+/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
+/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
+/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
+/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
+/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
+/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
+/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
+/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
+/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
+/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
+/*    POSSIBILITY OF SUCH DAMAGE.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+#include "version.h"
+
+#define PREFETCHSIZE	88
+
+#define N	$16
+#define X	$17
+#define INCX	$18
+#define I	$19
+
+#define s0	$f0
+#define s1	$f1
+#define s2	$f10
+#define s3	$f11
+
+#define a0	$f12
+#define a1	$f13
+#define a2	$f14
+#define a3	$f15
+#define a4	$f16
+#define a5	$f17
+#define a6	$f18
+#define a7	$f19
+
+#define t0	$f20
+#define t1	$f21
+#define t2	$f22
+#define t3	$f23
+
+	PROLOGUE
+	PROFCODE
+
+	fclr	s0
+	unop
+	fclr	t0
+	addq	INCX, INCX, INCX
+
+	fclr	s1
+	unop
+	fclr	t1
+	ble	N,  $L999
+
+	fclr	s2
+	sra	N, 2, I
+	fclr	s3
+	ble	I, $L15
+
+	LD	a0,  0 * SIZE(X)
+	fclr	t2
+	LD	a1,  1 * SIZE(X)
+	SXADDQ	INCX, X, X
+
+	LD	a2,  0 * SIZE(X)
+	fclr	t3
+	LD	a3,  1 * SIZE(X)
+	SXADDQ	INCX, X, X
+
+	LD	a4,  0 * SIZE(X)
+	LD	a5,  1 * SIZE(X)
+	SXADDQ	INCX, X, X
+	lda	I,  -1(I)
+
+	ble	I, $L13
+	.align 4
+
+$L12:
+	ADD	s0, t0, s0
+	ldl	$31, PREFETCHSIZE * SIZE(X)
+	fmov	a0, t0
+	lda	I,  -1(I)
+
+	ADD	s1, t1, s1
+	LD	a6,  0 * SIZE(X)
+	fmov	a1, t1
+	unop
+
+	ADD	s2, t2, s2
+	LD	a7,  1 * SIZE(X)
+	fmov	a2, t2
+	SXADDQ	INCX, X, X
+
+	ADD	s3, t3, s3
+	LD	a0,  0 * SIZE(X)
+	fmov	a3, t3
+	unop
+
+	ADD	s0, t0, s0
+	LD	a1,  1 * SIZE(X)
+	fmov	a4, t0
+	SXADDQ	INCX, X, X
+
+	ADD	s1, t1, s1
+	LD	a2,  0 * SIZE(X)
+	fmov	a5, t1
+	unop
+
+	ADD	s2, t2, s2
+	LD	a3,  1 * SIZE(X)
+	fmov	a6, t2
+	SXADDQ	INCX, X, X
+
+	ADD	s3, t3, s3
+	LD	a4,  0 * SIZE(X)
+	fmov	a7, t3
+	unop
+
+	LD	a5,  1 * SIZE(X)
+	unop
+	SXADDQ	INCX, X, X
+	bne	I, $L12
+	.align 4
+
+$L13:
+	ADD	s0, t0, s0
+	LD	a6,  0 * SIZE(X)
+	fmov	a0, t0
+
+	ADD	s1, t1, s1
+	LD	a7,  1 * SIZE(X)
+	fmov	a1, t1
+	SXADDQ	INCX, X, X
+
+	ADD	s2, t2, s2
+	fmov	a2, t2
+	ADD	s3, t3, s3
+	fmov	a3, t3
+
+	ADD	s0, t0, s0
+	fmov	a4, t0
+	ADD	s1, t1, s1
+	fmov	a5, t1
+	ADD	s2, t2, s2
+	fmov	a6, t2
+	ADD	s3, t3, s3
+	fmov	a7, t3
+
+	ADD	s2, t2, s2
+	ADD	s3, t3, s3
+
+	.align 4
+
+$L15:
+	ADD	s0, s2, s0
+	and	N, 3, I
+	ADD	s1, s3, s1
+	ble	I, $L999
+	.align 4
+
+$L17:
+	ADD	s0, t0, s0
+	LD	a0,  0 * SIZE(X)
+	fmov	a0, t0
+	lda	I,  -1(I)
+
+	ADD	s1, t1, s1
+	LD	a1,  1 * SIZE(X)
+	fmov	a1, t1
+	SXADDQ	INCX, X, X
+
+	bne	I, $L17
+	.align 4
+
+$L999:
+	ADD	s0, t0, s0
+	ADD	s1, t1, s1
+
+	ADD	s0, s1, s0
+	ret
+	EPILOGUE

kernel/arm/KERNEL.ARMV5

@@ -35,6 +35,11 @@ DASUMKERNEL  = ../arm/asum.c
 CASUMKERNEL  = ../arm/zasum.c
 ZASUMKERNEL  = ../arm/zasum.c
 
+SSUMKERNEL  = ../arm/sum.c
+DSUMKERNEL  = ../arm/sum.c
+CSUMKERNEL  = ../arm/zsum.c
+ZSUMKERNEL  = ../arm/zsum.c
+
 SAXPYKERNEL  = ../arm/axpy.c
 DAXPYKERNEL  = ../arm/axpy.c
 CAXPYKERNEL  = ../arm/zaxpy.c

kernel/arm/KERNEL.ARMV6

@@ -37,6 +37,9 @@ DASUMKERNEL  = asum_vfp.S
 CASUMKERNEL  = asum_vfp.S
 ZASUMKERNEL  = asum_vfp.S
 
+SSUMKERNEL  = sum_vfp.S
+DSUMKERNEL  = sum_vfp.S
+
 SAXPYKERNEL  = axpy_vfp.S
 DAXPYKERNEL  = axpy_vfp.S
 CAXPYKERNEL  = axpy_vfp.S

kernel/arm/imin.c

@@ -53,7 +53,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 
 	while(i < n)
 	{
-		if( x[ix] > minf )
+		if( x[ix] < minf )
 		{
 			min = i;
 			minf = x[ix];

kernel/arm/sum.c

@@ -0,0 +1,51 @@
+/***************************************************************************
+Copyright (c) 2013, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+/**************************************************************************************
+* trivial copy of asum.c with the ABS() removed                                       *
+**************************************************************************************/
+
+
+#include "common.h"
+#include <math.h>
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+	BLASLONG i=0;
+	FLOAT sumf = 0.0;
+	if (n <= 0 || inc_x <= 0) return(sumf);
+
+	n *= inc_x;
+	while(i < n)
+	{
+		sumf += x[i];
+		i += inc_x;
+	}
+	return(sumf);
+}
+
+

kernel/arm/sum_vfp.S

@@ -0,0 +1,425 @@
+/***************************************************************************
+Copyright (c) 2013, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+/**************************************************************************************
+* trivial copy of asum_vfp.S with the in-place vabs.f64 calls removed                                    *
+**************************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define STACKSIZE 256
+
+#define	N	r0
+#define	X	r1
+#define	INC_X	r2
+
+
+#define I	r12
+
+#define X_PRE	512
+
+/**************************************************************************************
+* Macro definitions
+**************************************************************************************/
+
+#if	!defined(COMPLEX)
+
+#if	defined(DOUBLE)
+
+.macro KERNEL_F4
+
+	pld	[ X, #X_PRE  ]
+	vldmia.f64	X!, { d4 - d5 }
+	vadd.f64   d0  , d0,  d4
+	vldmia.f64	X!, { d6 - d7 }
+	vadd.f64   d1  , d1,  d5
+	vadd.f64   d0  , d0,  d6
+	vadd.f64   d1  , d1,  d7
+
+.endm
+
+.macro KERNEL_F1
+
+	vldmia.f64	X!, { d4 }
+	vadd.f64   d0  , d0,  d4
+
+.endm
+
+
+.macro KERNEL_S4
+
+	vldmia.f64	X, { d4 }
+	vadd.f64   d0  , d0,  d4
+	add	X, X, INC_X
+
+	vldmia.f64	X, { d4 }
+	vadd.f64   d0  , d0,  d4
+	add	X, X, INC_X
+
+	vldmia.f64	X, { d4 }
+	vadd.f64   d0  , d0,  d4
+	add	X, X, INC_X
+
+	vldmia.f64	X, { d4 }
+	vadd.f64   d0  , d0,  d4
+	add	X, X, INC_X
+
+.endm
+
+
+.macro KERNEL_S1
+
+	vldmia.f64	X, { d4 }
+	vadd.f64   d0  , d0,  d4
+	add	X, X, INC_X
+
+.endm
+
+#else
+
+.macro KERNEL_F4
+
+	vldmia.f32	X!, { s4 - s5 }
+	vadd.f32   s0  , s0,  s4
+	vldmia.f32	X!, { s6 - s7 }
+	vadd.f32   s1  , s1,  s5
+	vadd.f32   s0  , s0,  s6
+	vadd.f32   s1  , s1,  s7
+
+.endm
+
+.macro KERNEL_F1
+
+	vldmia.f32	X!, { s4 }
+	vadd.f32   s0  , s0,  s4
+
+.endm
+
+
+.macro KERNEL_S4
+
+	vldmia.f32	X, { s4 }
+	vadd.f32   s0  , s0,  s4
+	add	X, X, INC_X
+
+	vldmia.f32	X, { s4 }
+	vadd.f32   s0  , s0,  s4
+	add	X, X, INC_X
+
+	vldmia.f32	X, { s4 }
+	vadd.f32   s0  , s0,  s4
+	add	X, X, INC_X
+
+	vldmia.f32	X, { s4 }
+	vadd.f32   s0  , s0,  s4
+	add	X, X, INC_X
+
+.endm
+
+
+.macro KERNEL_S1
+
+	vldmia.f32	X, { s4 }
+	vadd.f32   s0  , s0,  s4
+	add	X, X, INC_X
+
+.endm
+
+
+#endif
+
+#else
+
+#if	defined(DOUBLE)
+
+.macro KERNEL_F4
+
+	pld	[ X, #X_PRE  ]
+	vldmia.f64	X!, { d4 - d5 }
+	vadd.f64   d0  , d0,  d4
+	vldmia.f64	X!, { d6 - d7 }
+	vadd.f64   d1  , d1,  d5
+	vadd.f64   d0  , d0,  d6
+	vadd.f64   d1  , d1,  d7
+
+	pld	[ X, #X_PRE  ]
+	vldmia.f64	X!, { d4 - d5 }
+	vadd.f64   d0  , d0,  d4
+	vldmia.f64	X!, { d6 - d7 }
+	vadd.f64   d1  , d1,  d5
+	vadd.f64   d0  , d0,  d6
+	vadd.f64   d1  , d1,  d7
+
+
+.endm
+
+.macro KERNEL_F1
+
+	vldmia.f64	X!, { d4 }
+	vadd.f64   d0  , d0,  d4
+
+	vldmia.f64	X!, { d4 }
+	vadd.f64   d0  , d0,  d4
+
+
+.endm
+
+
+.macro KERNEL_S4
+
+	vldmia.f64	X, { d4 -d5 }
+	vadd.f64   d0  , d0,  d4
+	vadd.f64   d0  , d0,  d5
+	add	X, X, INC_X
+
+	vldmia.f64	X, { d4 -d5 }
+	vadd.f64   d0  , d0,  d4
+	vadd.f64   d0  , d0,  d5
+	add	X, X, INC_X
+
+	vldmia.f64	X, { d4 -d5 }
+	vadd.f64   d0  , d0,  d4
+	vadd.f64   d0  , d0,  d5
+	add	X, X, INC_X
+
+	vldmia.f64	X, { d4 -d5 }
+	vadd.f64   d0  , d0,  d4
+	vadd.f64   d0  , d0,  d5
+	add	X, X, INC_X
+
+.endm
+
+
+.macro KERNEL_S1
+
+	vldmia.f64	X, { d4 -d5 }
+	vadd.f64   d0  , d0,  d4
+	vadd.f64   d0  , d0,  d5
+	add	X, X, INC_X
+
+.endm
+
+#else
+
+.macro KERNEL_F4
+
+	pld	[ X, #X_PRE  ]
+	vldmia.f32	X!, { s4 - s5 }
+	vadd.f32   s0  , s0,  s4
+	vldmia.f32	X!, { s6 - s7 }
+	vadd.f32   s1  , s1,  s5
+	vadd.f32   s0  , s0,  s6
+	vadd.f32   s1  , s1,  s7
+
+	vldmia.f32	X!, { s4 - s5 }
+	vadd.f32   s0  , s0,  s4
+	vldmia.f32	X!, { s6 - s7 }
+	vadd.f32   s1  , s1,  s5
+	vadd.f32   s0  , s0,  s6
+	vadd.f32   s1  , s1,  s7
+
+
+.endm
+
+.macro KERNEL_F1
+
+	vldmia.f32	X!, { s4 }
+	vadd.f32   s0  , s0,  s4
+
+	vldmia.f32	X!, { s4 }
+	vadd.f32   s0  , s0,  s4
+
+.endm
+
+
+.macro KERNEL_S4
+
+	vldmia.f32	X, { s4 -s5 }
+	vadd.f32   s0  , s0,  s4
+	vadd.f32   s0  , s0,  s5
+	add	X, X, INC_X
+
+	vldmia.f32	X, { s4 -s5 }
+	vadd.f32   s0  , s0,  s4
+	vadd.f32   s0  , s0,  s5
+	add	X, X, INC_X
+
+	vldmia.f32	X, { s4 -s5 }
+	vadd.f32   s0  , s0,  s4
+	vadd.f32   s0  , s0,  s5
+	add	X, X, INC_X
+
+	vldmia.f32	X, { s4 -s5 }
+	vadd.f32   s0  , s0,  s4
+	vadd.f32   s0  , s0,  s5
+	add	X, X, INC_X
+
+.endm
+
+
+.macro KERNEL_S1
+
+	vldmia.f32	X, { s4 -s5 }
+	vadd.f32   s0  , s0,  s4
+	vadd.f32   s0  , s0,  s5
+	add	X, X, INC_X
+
+.endm
+
+#endif
+
+#endif
+
+/**************************************************************************************
+* End of macro definitions
+**************************************************************************************/
+
+	PROLOGUE
+
+	.align 5
+
+        movs    r12, #0                                          // clear floating point register
+        vmov    s0, r12
+        vmov    s1, r12
+#if     defined(DOUBLE)
+        vcvt.f64.f32    d0, s0
+        vcvt.f64.f32    d1, s1
+#endif
+
+	cmp	N, #0
+	ble	asum_kernel_L999
+
+	cmp	INC_X, #0
+	beq	asum_kernel_L999
+
+	cmp	INC_X, #1
+	bne	asum_kernel_S_BEGIN
+
+
+asum_kernel_F_BEGIN:
+
+	asrs	I, N, #2					// I = N / 4
+	ble	asum_kernel_F1
+
+	.align 5
+
+asum_kernel_F4:
+
+#if !defined(DOUBLE) && !defined(COMPLEX)
+	pld	[ X, #X_PRE  ]
+#endif
+	KERNEL_F4
+
+	subs	I, I, #1
+	ble	asum_kernel_F1
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	asum_kernel_F4
+
+asum_kernel_F1:
+
+	ands	I, N, #3
+	ble	asum_kernel_L999
+
+asum_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     asum_kernel_F10
+
+	b	asum_kernel_L999
+
+asum_kernel_S_BEGIN:
+
+#if defined(COMPLEX)
+
+#if defined(DOUBLE)
+	lsl	INC_X, INC_X, #4				// INC_X * SIZE * 2
+#else
+	lsl	INC_X, INC_X, #3				// INC_X * SIZE * 2
+#endif
+
+#else
+
+#if defined(DOUBLE)
+	lsl	INC_X, INC_X, #3				// INC_X * SIZE
+#else
+	lsl	INC_X, INC_X, #2				// INC_X * SIZE
+#endif
+
+#endif
+
+	asrs	I, N, #2					// I = N / 4
+	ble	asum_kernel_S1
+
+	.align 5
+
+asum_kernel_S4:
+
+	KERNEL_S4
+
+	subs	I, I, #1
+	bne	asum_kernel_S4
+
+asum_kernel_S1:
+
+	ands	I, N, #3
+	ble	asum_kernel_L999
+
+asum_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     asum_kernel_S10
+
+
+asum_kernel_L999:
+
+
+#if defined(DOUBLE)
+	vadd.f64	d0 , d0, d1				// set return value
+#else
+	vadd.f32	s0 , s0, s1				// set return value
+#endif
+
+#if !defined(__ARM_PCS_VFP)
+#if !defined(DOUBLE)
+	vmov	r0, s0
+#else
+	vmov	r0, r1, d0
+#endif
+#endif
+
+	bx	lr
+
+	EPILOGUE
+

kernel/arm/zsum.c

@@ -0,0 +1,57 @@
+/***************************************************************************
+Copyright (c) 2013, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+/**************************************************************************************
+* trivial copy of zasum.c with the ABS() removed                                      *
+**************************************************************************************/
+
+
+#include "common.h"
+#include <math.h>
+
+#define CSUM1(x,i)	x[i]+x[i+1]
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+	BLASLONG i=0;
+	FLOAT sumf = 0.0;
+	BLASLONG inc_x2;
+
+	if (n <= 0 || inc_x <= 0) return(sumf);
+
+	inc_x2 = 2 * inc_x;
+
+	n *= inc_x2;
+	while(i < n)
+	{
+		sumf += CSUM1(x,i);
+		i += inc_x2;
+	}
+	return(sumf);
+}
+
+

kernel/arm64/KERNEL.TSV110

@@ -0,0 +1,175 @@
+SAMINKERNEL  = ../arm/amin.c
+DAMINKERNEL  = ../arm/amin.c
+CAMINKERNEL  = ../arm/zamin.c
+ZAMINKERNEL  = ../arm/zamin.c
+
+SMAXKERNEL   = ../arm/max.c
+DMAXKERNEL   = ../arm/max.c
+
+SMINKERNEL   = ../arm/min.c
+DMINKERNEL   = ../arm/min.c
+
+ISAMINKERNEL = ../arm/iamin.c
+IDAMINKERNEL = ../arm/iamin.c
+ICAMINKERNEL = ../arm/izamin.c
+IZAMINKERNEL = ../arm/izamin.c
+
+ISMAXKERNEL  = ../arm/imax.c
+IDMAXKERNEL  = ../arm/imax.c
+
+ISMINKERNEL  = ../arm/imin.c
+IDMINKERNEL  = ../arm/imin.c
+
+STRMMKERNEL	= ../generic/trmmkernel_4x4.c
+DTRMMKERNEL	= ../generic/trmmkernel_2x2.c
+CTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
+ZTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
+
+STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
+STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c
+STRSMKERNEL_RN	=  ../generic/trsm_kernel_RN.c
+STRSMKERNEL_RT	=  ../generic/trsm_kernel_RT.c
+
+DTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+DTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+DTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+CTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+CTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+CTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+CTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+ZTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+ZTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+ZTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+SAMAXKERNEL  = amax.S
+DAMAXKERNEL  = amax.S
+CAMAXKERNEL  = zamax.S
+ZAMAXKERNEL  = zamax.S
+
+ISAMAXKERNEL = iamax.S
+IDAMAXKERNEL = iamax.S
+ICAMAXKERNEL = izamax.S
+IZAMAXKERNEL = izamax.S
+
+SASUMKERNEL  = asum.S
+DASUMKERNEL  = asum.S
+CASUMKERNEL  = casum.S
+ZASUMKERNEL  = zasum.S
+
+SAXPYKERNEL  = axpy.S
+DAXPYKERNEL  = axpy.S
+CAXPYKERNEL  = zaxpy.S
+ZAXPYKERNEL  = zaxpy.S
+
+SCOPYKERNEL  = copy.S
+DCOPYKERNEL  = copy.S
+CCOPYKERNEL  = copy.S
+ZCOPYKERNEL  = copy.S
+
+SDOTKERNEL   = dot.S
+DDOTKERNEL   = dot.S
+CDOTKERNEL   = zdot.S
+ZDOTKERNEL   = zdot.S
+DSDOTKERNEL  = dot.S
+
+SNRM2KERNEL  = nrm2.S
+DNRM2KERNEL  = nrm2.S
+CNRM2KERNEL  = znrm2.S
+ZNRM2KERNEL  = znrm2.S
+
+SROTKERNEL   = rot.S
+DROTKERNEL   = rot.S
+CROTKERNEL   = zrot.S
+ZROTKERNEL   = zrot.S
+
+SSCALKERNEL  = scal.S
+DSCALKERNEL  = scal.S
+CSCALKERNEL  = zscal.S
+ZSCALKERNEL  = zscal.S
+
+SSWAPKERNEL  = swap.S
+DSWAPKERNEL  = swap.S
+CSWAPKERNEL  = swap.S
+ZSWAPKERNEL  = swap.S
+
+SGEMVNKERNEL = gemv_n.S
+DGEMVNKERNEL = gemv_n.S
+CGEMVNKERNEL = zgemv_n.S
+ZGEMVNKERNEL = zgemv_n.S
+
+SGEMVTKERNEL = gemv_t.S
+DGEMVTKERNEL = gemv_t.S
+CGEMVTKERNEL = zgemv_t.S
+ZGEMVTKERNEL = zgemv_t.S
+
+SGEMMKERNEL    =  sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
+STRMMKERNEL    =  strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
+ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
+SGEMMINCOPY    =  ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
+SGEMMITCOPY    =  ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
+SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
+SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+SGEMMONCOPY    =  ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
+SGEMMOTCOPY    =  ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+DGEMMKERNEL    =  dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S
+DTRMMKERNEL    =  dtrmm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S
+
+ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))
+
+ifeq ($(DGEMM_UNROLL_M), 8)
+DGEMMINCOPY    =  dgemm_ncopy_$(DGEMM_UNROLL_M).S
+DGEMMITCOPY    =  dgemm_tcopy_$(DGEMM_UNROLL_M).S
+else
+DGEMMINCOPY    =  ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c
+DGEMMITCOPY    =  ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c
+endif
+
+DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+
+ifeq ($(DGEMM_UNROLL_N), 4)
+DGEMMONCOPY    =  dgemm_ncopy_$(DGEMM_UNROLL_N).S
+DGEMMOTCOPY    =  dgemm_tcopy_$(DGEMM_UNROLL_N).S
+else
+DGEMMONCOPY    =  ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c
+DGEMMOTCOPY    =  ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c
+endif
+
+DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+CGEMMKERNEL    =  cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
+CTRMMKERNEL    =  ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
+ifneq ($(CGEMM_UNROLL_M), $(CGEMM_UNROLL_N))
+CGEMMINCOPY    =  ../generic/zgemm_ncopy_$(CGEMM_UNROLL_M).c
+CGEMMITCOPY    =  ../generic/zgemm_tcopy_$(CGEMM_UNROLL_M).c
+CGEMMINCOPYOBJ =  cgemm_incopy$(TSUFFIX).$(SUFFIX)
+CGEMMITCOPYOBJ =  cgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+CGEMMONCOPY    =  ../generic/zgemm_ncopy_$(CGEMM_UNROLL_N).c
+CGEMMOTCOPY    =  ../generic/zgemm_tcopy_$(CGEMM_UNROLL_N).c
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+ZGEMMKERNEL    =  zgemm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
+ZTRMMKERNEL    =  ztrmm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
+ifneq ($(ZGEMM_UNROLL_M), $(ZGEMM_UNROLL_N))
+ZGEMMINCOPY    =  ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_M).c
+ZGEMMITCOPY    =  ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_M).c
+ZGEMMINCOPYOBJ =  zgemm_incopy$(TSUFFIX).$(SUFFIX)
+ZGEMMITCOPYOBJ =  zgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+ZGEMMONCOPY    =  ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_N).c
+ZGEMMOTCOPY    =  ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
+

kernel/arm64/csum.S

@@ -0,0 +1,164 @@
+/*******************************************************************************
+Copyright (c) 2019, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#define REG0	wzr
+#define SUMF	s0
+#define TMPF	s1
+#define TMPVF	{v1.s}[0]
+#define SZ	4
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+	ld1	{v1.2s}, [X], #8
+	ext	v2.8b, v1.8b, v1.8b, #4
+	fadd	TMPF, TMPF, s2
+	fadd	SUMF, SUMF, TMPF
+.endm
+
+.macro KERNEL_F8
+	ld1	{v1.4s, v2.4s, v3.4s, v4.4s}, [X]
+	add	X, X, #64
+
+	PRFM	PLDL1KEEP, [X, #1024]
+
+	fadd	v1.4s, v1.4s, v2.4s
+	fadd	v3.4s, v3.4s, v4.4s
+	fadd	v0.4s, v0.4s, v1.4s
+	fadd	v0.4s, v0.4s, v3.4s
+.endm
+
+.macro KERNEL_F8_FINALIZE
+	ext	v1.16b, v0.16b, v0.16b, #8
+	fadd	v0.2s, v0.2s, v1.2s
+	faddp	SUMF, v0.2s
+.endm
+
+.macro INIT_S
+	lsl	INC_X, INC_X, #3
+.endm
+
+.macro KERNEL_S1
+	ld1	{v1.2s}, [X], INC_X
+	ext	v2.8b, v1.8b, v1.8b, #4
+	fadd	TMPF, TMPF, s2
+	fadd	SUMF, SUMF, TMPF
+
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	fmov	SUMF, REG0
+	fmov	s1, SUMF
+
+	cmp	N, xzr
+	ble	.Lcsum_kernel_L999
+	cmp	INC_X, xzr
+	ble	.Lcsum_kernel_L999
+
+	cmp	INC_X, #1
+	bne	.Lcsum_kernel_S_BEGIN
+
+.Lcsum_kernel_F_BEGIN:
+
+	asr	I, N, #3
+	cmp	I, xzr
+	beq	.Lcsum_kernel_F1
+
+.Lcsum_kernel_F8:
+
+	KERNEL_F8
+
+	subs	I, I, #1
+	bne	.Lcsum_kernel_F8
+
+	KERNEL_F8_FINALIZE
+
+.Lcsum_kernel_F1:
+
+	ands	I, N, #7
+	ble	.Lcsum_kernel_L999
+
+.Lcsum_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     .Lcsum_kernel_F10
+
+.Lcsum_kernel_L999:
+	ret
+
+.Lcsum_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	.Lcsum_kernel_S1
+
+.Lcsum_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	.Lcsum_kernel_S4
+
+.Lcsum_kernel_S1:
+
+	ands	I, N, #3
+	ble	.Lcsum_kernel_L999
+
+.Lcsum_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     .Lcsum_kernel_S10
+
+	ret
+
+	EPILOGUE

kernel/arm64/sum.S

@@ -0,0 +1,186 @@
+/*******************************************************************************
+Copyright (c) 2019, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define REG0	wzr
+#define SUMF	s0
+#define TMPF	s1
+#define TMPVF	{v1.s}[0]
+#define SZ	4
+#else
+#define REG0	xzr
+#define SUMF	d0
+#define TMPF	d1
+#define TMPVF	{v1.d}[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+	ldr	TMPF, [X], #SZ
+	fadd	SUMF, SUMF, TMPF
+.endm
+
+.macro KERNEL_F8
+#if !defined(DOUBLE)
+	ld1	{v1.4s, v2.4s}, [X], #32	// Load [X3, X2, X1, X0]
+	fadd	v1.4s, v1.4s, v2.4s		// [X3+X1, X2+X0]
+	fadd	v0.4s, v0.4s, v1.4s		// [X3+X1, X2+X0]
+	PRFM	PLDL1KEEP, [X, #1024]
+#else // DOUBLE
+	ld1	{v2.2d, v3.2d, v4.2d, v5.2d}, [X]
+	add	X, X, #64
+
+	PRFM	PLDL1KEEP, [X, #1024]
+
+	fadd	v2.2d, v2.2d, v3.2d
+	fadd	v4.2d, v4.2d, v5.2d
+	fadd	v0.2d, v0.2d, v2.2d
+	fadd	v0.2d, v0.2d, v4.2d
+#endif
+.endm
+
+.macro KERNEL_F8_FINALIZE
+#if !defined(DOUBLE)
+	ext	v1.16b, v0.16b, v0.16b, #8
+	fadd	v0.2s, v0.2s, v1.2s
+	faddp	SUMF, v0.2s
+#else
+	faddp	SUMF, v0.2d
+#endif
+.endm
+
+.macro INIT_S
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #2
+#else
+	lsl	INC_X, INC_X, #3
+#endif
+.endm
+
+.macro KERNEL_S1
+	ld1	TMPVF, [X], INC_X
+	fadd	SUMF, SUMF, TMPF
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	fmov	SUMF, REG0
+#if !defined(DOUBLE)
+	fmov	s1, SUMF
+#else
+	fmov	d1, SUMF
+#endif
+
+	cmp	N, xzr
+	ble	.Lsum_kernel_L999
+	cmp	INC_X, xzr
+	ble	.Lsum_kernel_L999
+
+	cmp	INC_X, #1
+	bne	.Lsum_kernel_S_BEGIN
+
+.Lsum_kernel_F_BEGIN:
+
+	asr	I, N, #3
+	cmp	I, xzr
+	beq	.Lsum_kernel_F1
+
+.Lsum_kernel_F8:
+
+	KERNEL_F8
+
+	subs	I, I, #1
+	bne	.Lsum_kernel_F8
+
+	KERNEL_F8_FINALIZE
+
+.Lsum_kernel_F1:
+
+	ands	I, N, #7
+	ble	.Lsum_kernel_L999
+
+.Lsum_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     .Lsum_kernel_F10
+
+.Lsum_kernel_L999:
+	ret
+
+.Lsum_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	.Lsum_kernel_S1
+
+.Lsum_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	.Lsum_kernel_S4
+
+.Lsum_kernel_S1:
+
+	ands	I, N, #3
+	ble	.Lsum_kernel_L999
+
+.Lsum_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     .Lsum_kernel_S10
+
+	ret
+
+	EPILOGUE

kernel/arm64/zsum.S

@@ -0,0 +1,158 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#define REG0	xzr
+#define SUMF	d0
+#define TMPF	d1
+#define TMPVF	{v1.d}[0]
+#define SZ	8
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+	ld1	{v1.2d}, [X], #16
+	faddp	TMPF, v1.2d
+	fadd	SUMF, SUMF, TMPF
+.endm
+
+.macro KERNEL_F4
+	ld1	{v1.2d, v2.2d, v3.2d, v4.2d}, [X], #64
+
+	fadd	v1.2d, v1.2d, v2.2d
+	fadd	v3.2d, v3.2d, v4.2d
+
+	fadd	v0.2d, v0.2d, v1.2d
+	fadd	v0.2d, v0.2d, v3.2d
+
+	PRFM	PLDL1KEEP, [X, #1024]
+.endm
+
+.macro KERNEL_F4_FINALIZE
+	faddp	SUMF, v0.2d
+.endm
+
+.macro INIT_S
+	lsl	INC_X, INC_X, #4
+.endm
+
+.macro KERNEL_S1
+	ld1	{v1.2d}, [X], INC_X
+	faddp	TMPF, v1.2d
+	fadd	SUMF, SUMF, TMPF
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	fmov	SUMF, REG0
+
+	cmp	N, xzr
+	ble	.Lzsum_kernel_L999
+	cmp	INC_X, xzr
+	ble	.Lzsum_kernel_L999
+
+	cmp	INC_X, #1
+	bne	.Lzsum_kernel_S_BEGIN
+
+.Lzsum_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	.Lzsum_kernel_F1
+
+.Lzsum_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	.Lzsum_kernel_F4
+
+	KERNEL_F4_FINALIZE
+
+.Lzsum_kernel_F1:
+
+	ands	I, N, #3
+	ble	.Lzsum_kernel_L999
+
+.Lzsum_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     .Lzsum_kernel_F10
+
+.Lzsum_kernel_L999:
+	ret
+
+.Lzsum_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	.Lzsum_kernel_S1
+
+.Lzsum_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	.Lzsum_kernel_S4
+
+.Lzsum_kernel_S1:
+
+	ands	I, N, #3
+	ble	.Lzsum_kernel_L999
+
+.Lzsum_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     .Lzsum_kernel_S10
+
+	ret
+
+	EPILOGUE

kernel/ia64/KERNEL

@@ -60,6 +60,10 @@ CASUMKERNEL  = asum.S
 ZASUMKERNEL  = asum.S
 XASUMKERNEL  = asum.S
 
+CSUMKERNEL  = sum.S
+ZSUMKERNEL  = sum.S
+XSUMKERNEL  = sum.S
+
 CNRM2KERNEL  = nrm2.S
 ZNRM2KERNEL  = nrm2.S
 XNRM2KERNEL  = nrm2.S

kernel/ia64/sum.S

@@ -0,0 +1,358 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* Copyright 2019, The OpenBLAS project                              */
+/* All rights reserved.                                              */
+/*                                                                   */
+/* Redistribution and use in source and binary forms, with or        */
+/* without modification, are permitted provided that the following   */
+/* conditions are met:                                               */
+/*                                                                   */
+/*   1. Redistributions of source code must retain the above         */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer.                                                  */
+/*                                                                   */
+/*   2. Redistributions in binary form must reproduce the above      */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer in the documentation and/or other materials       */
+/*      provided with the distribution.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
+/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
+/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
+/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
+/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
+/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
+/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
+/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
+/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
+/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
+/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
+/*    POSSIBILITY OF SUCH DAMAGE.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#ifdef XDOUBLE
+#define PREFETCH_SIZE ( 8 * 16 +  4)
+#elif defined(DOUBLE)
+#define PREFETCH_SIZE (16 * 16 +  8)
+#else
+#define PREFETCH_SIZE (32 * 16 + 16)
+#endif
+
+#ifndef COMPLEX
+#define COMPADD	0
+#define STRIDE INCX
+#else
+#define COMPADD	1
+#define STRIDE SIZE
+#endif
+
+#define PRE1	r2
+
+#define I	r17
+#define J	r18
+#define INCX16	r21
+
+#define PR	r30
+#define ARLC	r31
+
+#define N	r32
+#define X	r33
+#define INCX	r34
+
+
+	PROLOGUE
+	.prologue
+	PROFCODE
+	{ .mfi
+	adds	PRE1 = PREFETCH_SIZE * SIZE, X
+	mov	f8   = f0
+	.save ar.lc, ARLC
+	mov	ARLC = ar.lc
+	}
+	;;
+	.body
+#ifdef F_INTERFACE
+	{ .mmi
+	LDINT	N    = [N]
+	LDINT	INCX = [INCX]
+	nop.i	0
+	}
+	;;
+#ifndef USE64BITINT
+	{ .mii
+	nop.m	0
+	sxt4	N = N
+	sxt4	INCX = INCX
+	}
+	;;
+#endif
+#endif
+	{ .mmi
+	cmp.lt	p0, p6 = r0, INCX
+	cmp.lt	p0, p7 = r0, N
+	shr	I =  N, (4 - COMPADD)
+	}
+	{ .mbb
+	and	J = ((1 << (4 - COMPADD)) - 1), N
+	(p6) 	br.ret.sptk.many b0
+	(p7) 	br.ret.sptk.many b0
+	}
+	;;
+	{ .mfi
+	adds	I = -1, I
+	mov	f10 = f0
+	mov	PR = pr
+	}
+	{ .mfi
+	cmp.eq	p9, p0  =   r0, J
+	mov	f9  = f0
+	tbit.z	p0, p12 = N, 3 - COMPADD
+	}
+	;;
+	{ .mmi
+	cmp.eq	p16, p0 = r0, r0
+	cmp.ne	p17, p0 = r0, r0
+	mov	ar.ec= 3
+	}
+	{ .mfi
+	cmp.ne	p18, p0 = r0, r0
+	mov	f11 = f0
+	shl	INCX = INCX, BASE_SHIFT + COMPADD
+	}
+	;;
+	{ .mmi
+#ifdef XDOUBLE
+	shladd	INCX16  = INCX, (3 - COMPADD), r0
+#else
+	shladd	INCX16  = INCX, (4 - COMPADD), r0
+#endif
+	cmp.ne	p19, p0 = r0, r0
+	mov	ar.lc = I
+	}
+	{ .mmb
+	cmp.gt	p8 ,p0  =   r0, I
+#ifdef COMPLEX
+       adds	INCX = - SIZE, INCX
+#else
+	nop.m	0
+#endif
+	(p8) br.cond.dpnt  .L55
+	}
+	;;
+	.align 32
+
+.L52:
+	{ .mmf
+	(p16) lfetch.nt1 [PRE1], INCX16
+	(p16) LDFD	f32  = [X], STRIDE
+	}
+	{ .mfb
+	(p19) FADD	f8  = f8,  f71
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f35  = [X], INCX
+	}
+	{ .mfb
+	(p19) FADD	f9  = f9,  f74
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f38  = [X], STRIDE
+	}
+	{ .mfb
+	(p19) FADD	f10 = f10, f77
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f41  = [X], INCX
+	}
+	{ .mfb
+	(p19) FADD	f11 = f11, f80
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f44  = [X], STRIDE
+	}
+	{ .mfb
+	(p18) FADD	f8  = f8,  f34
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f47  = [X], INCX
+	}
+	{ .mfb
+	(p18) FADD	f9  = f9,  f37
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f50  = [X], STRIDE
+	}
+	{ .mfb
+	(p18) FADD	f10 = f10, f40
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f53  = [X], INCX
+	}
+	{ .mfb
+	(p18) FADD	f11 = f11, f43
+	}
+	;;
+	{ .mmf
+#ifdef XDOUBLE
+	(p16) lfetch.nt1 [PRE1], INCX16
+#endif
+	(p16) LDFD	f56  = [X], STRIDE
+	}
+	{ .mfb
+	(p18) FADD	f8  = f8,  f46
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f59  = [X], INCX
+	}
+	{ .mfb
+	(p18) FADD	f9  = f9,  f49
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f62  = [X], STRIDE
+	}
+	{ .mfb
+	(p18) FADD	f10 = f10, f52
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f65  = [X], INCX
+	}
+	{ .mfb
+	(p18) FADD	f11 = f11, f55
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f68  = [X], STRIDE
+	}
+	{ .mfb
+	(p18) FADD	f8  = f8,  f58
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f71  = [X], INCX
+	}
+	{ .mfb
+	(p18) FADD	f9  = f9,  f61
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f74  = [X], STRIDE
+	}
+	{ .mfb
+	(p18) FADD	f10 = f10, f64
+	}
+	;;
+	{ .mmf
+	(p16) LDFD	f77  = [X], INCX
+	}
+	{ .mfb
+	(p18) FADD	f11 = f11, f67
+	br.ctop.sptk.few .L52
+	}
+	;;
+	FADD	f8  = f8,  f71
+	FADD	f9  = f9,  f74
+	FADD	f10 = f10, f77
+	FADD	f11 = f11, f80
+	.align 32
+	;;
+.L55:
+	(p12) LDFD	f32  = [X], STRIDE
+	(p9) br.cond.dptk .L998
+	;;
+	(p12) LDFD	f33  = [X], INCX
+	;;
+	(p12) LDFD	f34  = [X], STRIDE
+	;;
+	(p12) LDFD	f35  = [X], INCX
+	tbit.z	p0, p13 = N, (2 - COMPADD)
+	;;
+	(p12) LDFD	f36  = [X], STRIDE
+	tbit.z	p0, p14 = N, (1 - COMPADD)
+	;;
+	(p12) LDFD	f37  = [X], INCX
+#ifndef COMPLEX
+	tbit.z	p0, p15 = N, 0
+#endif
+	;;
+	(p12) LDFD	f38  = [X], STRIDE
+	;;
+	(p12) LDFD	f39  = [X], INCX
+	;;
+	(p13) LDFD	f40  = [X], STRIDE
+	;;
+	(p13) LDFD	f41  = [X], INCX
+	;;
+	(p13) LDFD	f42  = [X], STRIDE
+	(p12) FADD	f8  = f8,  f32
+	;;
+	(p13) LDFD	f43  = [X], INCX
+	(p12) FADD	f9  = f9,  f33
+	;;
+	(p14) LDFD	f44  = [X], STRIDE
+	(p12) FADD	f10 = f10, f34
+	;;
+	(p14) LDFD	f45  = [X], INCX
+	(p12) FADD	f11 = f11, f35
+	;;
+#ifndef COMPLEX
+	(p15) LDFD	f46  = [X]
+#endif
+	(p12) FADD	f8  = f8,  f36
+	;;
+	(p12) FADD	f9  = f9,  f37
+	(p12) FADD	f10 = f10, f38
+	(p12) FADD	f11 = f11, f39
+	;;
+	(p13) FADD	f8  = f8,  f40
+	(p13) FADD	f9  = f9,  f41
+#ifndef COMPLEX
+#endif
+	(p13) FADD	f10 = f10, f42
+	;;
+	(p13) FADD	f11 = f11, f43
+	(p14) FADD	f8  = f8,  f44
+	(p14) FADD	f9  = f9,  f45
+#ifndef COMPLEX
+	(p15) FADD	f10 = f10, f46
+#endif
+	;;
+	.align 32
+
+.L998:
+	{ .mfi
+	FADD	f8  = f8,  f9
+	mov	ar.lc  = ARLC
+	}
+	{ .mmf
+	FADD	f10 = f10, f11
+	}
+	;;
+	{ .mii
+	mov	pr = PR, -65474
+	}
+	;;
+	{ .mfb
+	FADD	f8  = f8,  f10
+	br.ret.sptk.many b0
+	}
+	EPILOGUE

kernel/mips/KERNEL.P5600

@@ -30,6 +30,11 @@ IDMAXKERNEL  = ../mips/imax.c
 ISMINKERNEL  = ../mips/imin.c
 IDMINKERNEL  = ../mips/imin.c
 
+SSUMKERNEL  = ../mips/sum.c
+DSUMKERNEL  = ../mips/sum.c
+CSUMKERNEL  = ../mips/zsum.c
+ZSUMKERNEL  = ../mips/zsum.c
+
 ifdef HAVE_MSA
 SASUMKERNEL  = ../mips/sasum_msa.c
 DASUMKERNEL  = ../mips/dasum_msa.c

kernel/mips/imin.c

@@ -45,7 +45,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 
 	while(i < n)
 	{
-		if( x[ix] > minf )
+		if( x[ix] < minf )
 		{
 			min = i;
 			minf = x[ix];

kernel/mips/sum.c

@@ -0,0 +1,47 @@
+/***************************************************************************
+Copyright (c) 2016, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <math.h>
+
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+	BLASLONG i=0;
+	FLOAT sumf = 0.0;
+	if (n <= 0 || inc_x <= 0) return(sumf);
+
+	n *= inc_x;
+	while(i < n)
+	{
+		sumf += x[i];
+		i += inc_x;
+	}
+	return(sumf);
+}
+
+

kernel/mips/zsum.c

@@ -0,0 +1,52 @@
+/***************************************************************************
+Copyright (c) 2016, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <math.h>
+
+#define CSUM1(x,i)	x[i]+x[i+1]
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+	BLASLONG i=0;
+	FLOAT sumf = 0.0;
+	BLASLONG inc_x2;
+
+	if (n <= 0 || inc_x <= 0) return(sumf);
+
+	inc_x2 = 2 * inc_x;
+
+	n *= inc_x2;
+	while(i < n)
+	{
+		sumf += CSUM1(x,i);
+		i += inc_x2;
+	}
+	return(sumf);
+}
+
+

kernel/mips64/sum.S

@@ -0,0 +1,332 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* All rights reserved.                                              */
+/*                                                                   */
+/* Redistribution and use in source and binary forms, with or        */
+/* without modification, are permitted provided that the following   */
+/* conditions are met:                                               */
+/*                                                                   */
+/*   1. Redistributions of source code must retain the above         */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer.                                                  */
+/*                                                                   */
+/*   2. Redistributions in binary form must reproduce the above      */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer in the documentation and/or other materials       */
+/*      provided with the distribution.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
+/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
+/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
+/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
+/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
+/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
+/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
+/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
+/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
+/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
+/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
+/*    POSSIBILITY OF SUCH DAMAGE.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define N	$4
+#define	X	$5
+#define INCX	$6
+
+#define I	$2
+#define TEMP	$3
+
+#define a1	$f2
+#define a2	$f3
+#define a3	$f4
+#define a4	$f5
+#define a5	$f6
+#define a6	$f7
+#define a7	$f8
+#define a8	$f9
+
+#define t1	$f10
+#define t2	$f11
+#define t3	$f12
+#define t4	$f13
+
+#define s1	$f0
+#define s2	$f1
+
+	PROLOGUE
+
+#ifdef F_INTERFACE
+	LDINT	N,     0(N)
+	LDINT	INCX,  0(INCX)
+#endif
+
+	MTC	$0,  s1
+
+	MTC	$0,  s2
+	dsll	INCX, INCX, BASE_SHIFT
+
+	blez	N, .L999
+	li	TEMP, SIZE
+
+	bne	INCX, TEMP, .L20
+	dsra	I, N, 3
+
+	blez	I, .L15
+	NOP
+
+	LD	a1,  0 * SIZE(X)
+	LD	a2,  1 * SIZE(X)
+	LD	a3,  2 * SIZE(X)
+	LD	a4,  3 * SIZE(X)
+
+	LD	a5,  4 * SIZE(X)
+	MOV	t1, a1
+	LD	a6,  5 * SIZE(X)
+	MOV	t2, a2
+	LD	a7,  6 * SIZE(X)
+	MOV	t3, a3
+
+	MOV	t4, a4
+	daddiu	I, I, -1
+
+	blez	I, .L13
+	LD	a8,  7 * SIZE(X)
+	.align 3
+
+.L12:
+	ADD	s1, s1, t1
+	LD	a1,  8 * SIZE(X)
+
+	MOV	t1, a5
+	daddiu	I, I, -1
+
+	ADD	s2, s2, t2
+	LD	a2,  9 * SIZE(X)
+
+	MOV	t2, a6
+	NOP
+
+	ADD	s1, s1, t3
+	LD	a3, 10 * SIZE(X)
+
+	MOV	t3, a7
+	NOP
+
+	ADD	s2, s2, t4
+	LD	a4, 11 * SIZE(X)
+
+	MOV	t4, a8
+	daddiu	X, X, 8 * SIZE
+
+	ADD	s1, s1, t1
+	LD	a5,  4 * SIZE(X)
+
+	MOV	t1, a1
+	NOP
+
+	ADD	s2, s2, t2
+	LD	a6,  5 * SIZE(X)
+
+	MOV	t2, a2
+	NOP
+
+	ADD	s1, s1, t3
+	LD	a7,  6 * SIZE(X)
+
+	MOV	t3, a3
+	NOP
+
+	ADD	s2, s2, t4
+	LD	a8,  7 * SIZE(X)
+
+	bgtz	I, .L12
+	MOV	t4, a4
+	.align 3
+
+.L13:
+	ADD	s1, s1, t1
+	daddiu	X, X, 8 * SIZE
+
+	MOV	t1, a5
+	NOP
+
+	ADD	s2, s2, t2
+	MOV	t2, a6
+
+	ADD	s1, s1, t3
+	MOV	t3, a7
+
+	ADD	s2, s2, t4
+	MOV	t4, a8
+
+	ADD	s1, s1, t1
+	ADD	s2, s2, t2
+	ADD	s1, s1, t3
+	ADD	s2, s2, t4
+	.align 3
+
+.L15:
+	andi	I,  N, 7
+
+	blez	I, .L999
+	NOP
+	.align	3
+
+.L16:
+	LD	a1,  0 * SIZE(X)
+	daddiu	I, I, -1
+
+	MOV	t1, a1
+
+	ADD	s1, s1, t1
+
+	bgtz	I, .L16
+	daddiu	X, X, SIZE
+
+	j	.L999
+	NOP
+	.align 3
+
+.L20:
+	blez	I, .L25
+	NOP
+
+	LD	a1,  0 * SIZE(X)
+	daddu	X, X, INCX
+
+	LD	a2,  0 * SIZE(X)
+	daddu	X, X, INCX
+
+	LD	a3,  0 * SIZE(X)
+	daddu	X, X, INCX
+
+	LD	a4,  0 * SIZE(X)
+	daddu	X, X, INCX
+
+	LD	a5,  0 * SIZE(X)
+	daddu	X, X, INCX
+
+	LD	a6,  0 * SIZE(X)
+	daddu	X, X, INCX
+
+	MOV	t1, a1
+	LD	a7,  0 * SIZE(X)
+
+	MOV	t2, a2
+	daddu	X, X, INCX
+
+	MOV	t3, a3
+	LD	a8,  0 * SIZE(X)
+
+	MOV	t4, a4
+	daddiu	I, I, -1
+
+	blez	I, .L24
+	daddu	X, X, INCX
+	.align 3
+
+.L23:
+	ADD	s1, s1, t1
+	LD	a1,  0 * SIZE(X)
+
+	MOV	t1, a5
+	daddu	X, X, INCX
+
+	ADD	s2, s2, t2
+	LD	a2,  0 * SIZE(X)
+
+	MOV	t2, a6
+	daddu	X, X, INCX
+
+	ADD	s1, s1, t3
+	LD	a3,  0 * SIZE(X)
+
+	MOV	t3, a7
+	daddu	X, X, INCX
+
+	ADD	s2, s2, t4
+	LD	a4,  0 * SIZE(X)
+
+	MOV	t4, a8
+	daddu	X, X, INCX
+
+	ADD	s1, s1, t1
+	LD	a5,  0 * SIZE(X)
+
+	MOV	t1, a1
+	daddu	X, X, INCX
+
+	ADD	s2, s2, t2
+	LD	a6,  0 * SIZE(X)
+
+	MOV	t2, a2
+	daddu	X, X, INCX
+
+	ADD	s1, s1, t3
+	LD	a7,  0 * SIZE(X)
+
+	MOV	t3, a3
+	daddu	X, X, INCX
+
+	ADD	s2, s2, t4
+	LD	a8,  0 * SIZE(X)
+
+	MOV	t4, a4
+	daddiu	I, I, -1
+
+	bgtz	I, .L23
+	daddu	X, X, INCX
+	.align 3
+
+.L24:
+	ADD	s1, s1, t1
+	MOV	t1, a5
+
+	ADD	s2, s2, t2
+	MOV	t2, a6
+
+	ADD	s1, s1, t3
+	MOV	t3, a7
+
+	ADD	s2, s2, t4
+	MOV	t4, a8
+
+	ADD	s1, s1, t1
+	ADD	s2, s2, t2
+	ADD	s1, s1, t3
+	ADD	s2, s2, t4
+	.align 3
+
+.L25:
+	andi	I,  N, 7
+
+	blez	I, .L999
+	NOP
+	.align	3
+
+.L26:
+	LD	a1,  0 * SIZE(X)
+	daddiu	I, I, -1
+
+	MOV	t1, a1
+	daddu	X, X, INCX
+
+	bgtz	I, .L26
+	ADD	s1, s1, t1
+	.align 3
+
+.L999:
+	j	$31
+	ADD	s1, s1, s2
+
+	EPILOGUE

kernel/mips64/zsum.S

@@ -0,0 +1,204 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* All rights reserved.                                              */
+/*                                                                   */
+/* Redistribution and use in source and binary forms, with or        */
+/* without modification, are permitted provided that the following   */
+/* conditions are met:                                               */
+/*                                                                   */
+/*   1. Redistributions of source code must retain the above         */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer.                                                  */
+/*                                                                   */
+/*   2. Redistributions in binary form must reproduce the above      */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer in the documentation and/or other materials       */
+/*      provided with the distribution.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
+/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
+/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
+/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
+/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
+/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
+/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
+/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
+/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
+/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
+/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
+/*    POSSIBILITY OF SUCH DAMAGE.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define N	$4
+#define	X	$5
+#define INCX	$6
+
+#define I	$2
+#define TEMP	$3
+
+#define a1	$f2
+#define a2	$f3
+#define a3	$f4
+#define a4	$f5
+#define a5	$f6
+#define a6	$f7
+#define a7	$f8
+#define a8	$f9
+
+#define t1	$f10
+#define t2	$f11
+#define t3	$f12
+#define t4	$f13
+
+#define s1	$f0
+#define s2	$f1
+
+	PROLOGUE
+
+#ifdef F_INTERFACE
+	LDINT	N,     0(N)
+	LDINT	INCX,  0(INCX)
+#endif
+
+	MTC	$0,  s1
+
+	MTC	$0,  s2
+	dsll	INCX, INCX, ZBASE_SHIFT
+
+	blez	N, .L999
+	dsra	I, N, 2
+
+	blez	I, .L25
+	NOP
+
+	LD	a1,  0 * SIZE(X)
+	LD	a2,  1 * SIZE(X)
+	daddu	X, X, INCX
+
+	LD	a3,  0 * SIZE(X)
+	LD	a4,  1 * SIZE(X)
+	daddu	X, X, INCX
+
+	LD	a5,  0 * SIZE(X)
+	LD	a6,  1 * SIZE(X)
+	daddu	X, X, INCX
+
+	MOV	t1, a1
+	MOV	t2, a2
+
+	LD	a7,  0 * SIZE(X)
+	LD	a8,  1 * SIZE(X)
+
+	MOV	t3, a3
+	MOV	t4, a4
+	daddiu	I, I, -1
+
+	blez	I, .L24
+	daddu	X, X, INCX
+	.align 3
+
+.L23:
+	ADD	s1, s1, t1
+	LD	a1,  0 * SIZE(X)
+
+	MOV	t1, a5
+	daddiu	I, I, -1
+
+	ADD	s2, s2, t2
+	LD	a2,  1 * SIZE(X)
+
+	MOV	t2, a6
+	daddu	X, X, INCX
+
+	ADD	s1, s1, t3
+	LD	a3,  0 * SIZE(X)
+
+	MOV	t3, a7
+	NOP
+
+	ADD	s2, s2, t4
+	LD	a4,  1 * SIZE(X)
+
+	MOV	t4, a8
+	daddu	X, X, INCX
+
+	ADD	s1, s1, t1
+	LD	a5,  0 * SIZE(X)
+
+	MOV	t1, a1
+	NOP
+
+	ADD	s2, s2, t2
+	LD	a6,  1 * SIZE(X)
+
+	MOV	t2, a2
+	daddu	X, X, INCX
+
+	ADD	s1, s1, t3
+	LD	a7,  0 * SIZE(X)
+
+	MOV	t3, a3
+	LD	a8,  1 * SIZE(X)
+
+	ADD	s2, s2, t4
+	daddu	X, X, INCX
+
+	bgtz	I, .L23
+	MOV	t4, a4
+	.align 3
+
+.L24:
+	ADD	s1, s1, t1
+	MOV	t1, a5
+
+	ADD	s2, s2, t2
+	MOV	t2, a6
+
+	ADD	s1, s1, t3
+	MOV	t3, a7
+
+	ADD	s2, s2, t4
+	MOV	t4, a8
+
+	ADD	s1, s1, t1
+	ADD	s2, s2, t2
+	ADD	s1, s1, t3
+	ADD	s2, s2, t4
+	.align 3
+
+.L25:
+	andi	I,  N, 3
+
+	blez	I, .L999
+	NOP
+	.align	3
+
+.L26:
+	LD	a1,  0 * SIZE(X)
+	LD	a2,  1 * SIZE(X)
+
+	MOV	t1, a1
+	daddiu	I, I, -1
+	MOV	t2, a2
+	daddu	X, X, INCX
+
+	ADD	s1, s1, t1
+	bgtz	I, .L26
+	ADD	s2, s2, t2
+	.align 3
+
+.L999:
+	j	$31
+	ADD	s1, s1, s2
+
+	EPILOGUE

kernel/power/KERNEL.POWER8

@@ -13,40 +13,40 @@ SGEMMINCOPY    = ../generic/gemm_ncopy_16.c
 SGEMMITCOPY    = sgemm_tcopy_16_power8.S
 SGEMMONCOPY    =  ../generic/gemm_ncopy_8.c
 SGEMMOTCOPY    = sgemm_tcopy_8_power8.S 
-SGEMMINCOPYOBJ =  sgemm_incopy.o
-SGEMMITCOPYOBJ =  sgemm_itcopy.o
-SGEMMONCOPYOBJ =  sgemm_oncopy.o
-SGEMMOTCOPYOBJ =  sgemm_otcopy.o
+SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
+SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
 DGEMMKERNEL    =  dgemm_kernel_16x4_power8.S
 DGEMMINCOPY    = ../generic/gemm_ncopy_16.c
 DGEMMITCOPY    =  dgemm_tcopy_16_power8.S
 DGEMMONCOPY    =  dgemm_ncopy_4_power8.S
 DGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
-DGEMMINCOPYOBJ =  dgemm_incopy.o
-DGEMMITCOPYOBJ =  dgemm_itcopy.o
-DGEMMONCOPYOBJ =  dgemm_oncopy.o
-DGEMMOTCOPYOBJ =  dgemm_otcopy.o
+DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
 CGEMMKERNEL    = cgemm_kernel_8x4_power8.S
 CGEMMINCOPY    = ../generic/zgemm_ncopy_8.c
 CGEMMITCOPY    = cgemm_tcopy_8_power8.S
 CGEMMONCOPY    = ../generic/zgemm_ncopy_4.c
 CGEMMOTCOPY    = ../generic/zgemm_tcopy_4.c
-CGEMMONCOPYOBJ =  cgemm_oncopy.o
-CGEMMOTCOPYOBJ =  cgemm_otcopy.o
-CGEMMINCOPYOBJ =  cgemm_incopy.o
-CGEMMITCOPYOBJ =  cgemm_itcopy.o
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
+CGEMMINCOPYOBJ =  cgemm_incopy$(TSUFFIX).$(SUFFIX)
+CGEMMITCOPYOBJ =  cgemm_itcopy$(TSUFFIX).$(SUFFIX)
 
 ZGEMMKERNEL    = zgemm_kernel_8x2_power8.S
 ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
 ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
 ZGEMMINCOPY    = ../generic/zgemm_ncopy_8.c
 ZGEMMITCOPY    = zgemm_tcopy_8_power8.S
-ZGEMMONCOPYOBJ =  zgemm_oncopy.o
-ZGEMMOTCOPYOBJ =  zgemm_otcopy.o
-ZGEMMINCOPYOBJ =  zgemm_incopy.o
-ZGEMMITCOPYOBJ =  zgemm_itcopy.o
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
+ZGEMMINCOPYOBJ =  zgemm_incopy$(TSUFFIX).$(SUFFIX)
+ZGEMMITCOPYOBJ =  zgemm_itcopy$(TSUFFIX).$(SUFFIX)
 
 STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
 STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c
@@ -89,14 +89,14 @@ ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
 #SMINKERNEL   = ../arm/min.c
 #DMINKERNEL   = ../arm/min.c
 #
-#ISAMAXKERNEL = ../arm/iamax.c
+ISAMAXKERNEL = isamax.c
 IDAMAXKERNEL = idamax.c
-#ICAMAXKERNEL = ../arm/izamax.c
-IZAMAXKERNEL =  izamax.c
+ICAMAXKERNEL = icamax.c
+IZAMAXKERNEL = izamax.c
 #
-#ISAMINKERNEL = ../arm/iamin.c
-IDAMINKERNEL =  idamin.c
-#ICAMINKERNEL = ../arm/izamin.c
+ISAMINKERNEL = isamin.c
+IDAMINKERNEL = idamin.c
+ICAMINKERNEL = icamin.c
 IZAMINKERNEL = izamin.c
 #
 #ISMAXKERNEL  = ../arm/imax.c
@@ -110,9 +110,9 @@ DASUMKERNEL  = dasum.c
 CASUMKERNEL  = casum.c
 ZASUMKERNEL  = zasum.c
 #
-#SAXPYKERNEL  = ../arm/axpy.c
+SAXPYKERNEL  = saxpy.c
 DAXPYKERNEL  = daxpy.c
-#CAXPYKERNEL  = ../arm/zaxpy.c
+CAXPYKERNEL  = caxpy.c
 ZAXPYKERNEL  = zaxpy.c
 #
 SCOPYKERNEL  = scopy.c
@@ -123,7 +123,7 @@ ZCOPYKERNEL  = zcopy.c
 SDOTKERNEL   =  sdot.c
 DDOTKERNEL   =  ddot.c
 DSDOTKERNEL  =  sdot.c
-#CDOTKERNEL   = ../arm/zdot.c
+CDOTKERNEL   =  cdot.c
 ZDOTKERNEL   =  zdot.c
 #
 SNRM2KERNEL  = ../arm/nrm2.c
@@ -133,7 +133,7 @@ ZNRM2KERNEL  = ../arm/znrm2.c
 #
 SROTKERNEL   = srot.c
 DROTKERNEL   = drot.c
-CROTKERNEL   = zrot.c
+CROTKERNEL   = crot.c
 ZROTKERNEL   = zrot.c
 #
 SSCALKERNEL  = sscal.c
@@ -147,14 +147,14 @@ CSWAPKERNEL  = cswap.c
 ZSWAPKERNEL  = zswap.c
 #
 
-#SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVNKERNEL = sgemv_n.c
 DGEMVNKERNEL = dgemv_n.c
-#CGEMVNKERNEL = ../arm/zgemv_n.c
+CGEMVNKERNEL = cgemv_n.c
 ZGEMVNKERNEL = zgemv_n_4.c
 #
-#SGEMVTKERNEL = ../arm/gemv_t.c
+SGEMVTKERNEL = sgemv_t.c
 DGEMVTKERNEL = dgemv_t.c
-#CGEMVTKERNEL = ../arm/zgemv_t.c
+CGEMVTKERNEL = cgemv_t.c
 ZGEMVTKERNEL = zgemv_t_4.c
 
 

kernel/power/KERNEL.POWER9

@@ -0,0 +1,184 @@
+#SGEMM_BETA = ../generic/gemm_beta.c
+#DGEMM_BETA = ../generic/gemm_beta.c
+#CGEMM_BETA = ../generic/zgemm_beta.c
+#ZGEMM_BETA = ../generic/zgemm_beta.c
+
+STRMMKERNEL	= strmm_kernel_16x8_power8.S
+DTRMMKERNEL	= dgemm_kernel_power9.S
+CTRMMKERNEL	= ctrmm_kernel_8x4_power8.S
+ZTRMMKERNEL	= ztrmm_kernel_8x2_power8.S
+
+SGEMMKERNEL    =  sgemm_kernel_16x8_power8.S
+SGEMMINCOPY    = ../generic/gemm_ncopy_16.c
+SGEMMITCOPY    = sgemm_tcopy_16_power8.S
+SGEMMONCOPY    =  ../generic/gemm_ncopy_8.c
+SGEMMOTCOPY    = sgemm_tcopy_8_power8.S 
+SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
+SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+DGEMMKERNEL    =  dgemm_kernel_power9.S
+DGEMMINCOPY    = ../generic/gemm_ncopy_16.c
+DGEMMITCOPY    =  dgemm_tcopy_16_power8.S
+DGEMMONCOPY    =  dgemm_ncopy_4_power8.S
+DGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
+DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+CGEMMKERNEL    = cgemm_kernel_8x4_power8.S
+CGEMMINCOPY    = ../generic/zgemm_ncopy_8.c
+CGEMMITCOPY    = cgemm_tcopy_8_power8.S
+CGEMMONCOPY    = ../generic/zgemm_ncopy_4.c
+CGEMMOTCOPY    = ../generic/zgemm_tcopy_4.c
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
+CGEMMINCOPYOBJ =  cgemm_incopy$(TSUFFIX).$(SUFFIX)
+CGEMMITCOPYOBJ =  cgemm_itcopy$(TSUFFIX).$(SUFFIX)
+
+ZGEMMKERNEL    = zgemm_kernel_8x2_power8.S
+ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
+ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
+ZGEMMINCOPY    = ../generic/zgemm_ncopy_8.c
+ZGEMMITCOPY    = zgemm_tcopy_8_power8.S
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
+ZGEMMINCOPYOBJ =  zgemm_incopy$(TSUFFIX).$(SUFFIX)
+ZGEMMITCOPYOBJ =  zgemm_itcopy$(TSUFFIX).$(SUFFIX)
+
+STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
+STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c
+STRSMKERNEL_RN	=  ../generic/trsm_kernel_RN.c
+STRSMKERNEL_RT	=  ../generic/trsm_kernel_RT.c
+
+DTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LT	= dtrsm_kernel_LT_16x4_power8.S
+DTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+DTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+CTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+CTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+CTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+CTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+ZTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+ZTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+ZTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+#Todo: CGEMM3MKERNEL should be 4x4 blocksizes.
+#CGEMM3MKERNEL    =  zgemm3m_kernel_8x4_sse3.S
+#ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_sse3.S
+
+#Pure C for other kernels
+#SAMAXKERNEL  = ../arm/amax.c
+#DAMAXKERNEL  = ../arm/amax.c
+#CAMAXKERNEL  = ../arm/zamax.c
+#ZAMAXKERNEL  = ../arm/zamax.c
+#
+#SAMINKERNEL  = ../arm/amin.c
+#DAMINKERNEL  = ../arm/amin.c
+#CAMINKERNEL  = ../arm/zamin.c
+#ZAMINKERNEL  = ../arm/zamin.c
+#
+#SMAXKERNEL   = ../arm/max.c
+#DMAXKERNEL   = ../arm/max.c
+#
+#SMINKERNEL   = ../arm/min.c
+#DMINKERNEL   = ../arm/min.c
+#
+ISAMAXKERNEL = isamax.c
+IDAMAXKERNEL = idamax.c
+ICAMAXKERNEL = icamax.c
+IZAMAXKERNEL = izamax.c
+#
+ISAMINKERNEL = isamin.c
+IDAMINKERNEL = idamin.c
+ICAMINKERNEL = icamin.c
+IZAMINKERNEL = izamin.c
+#
+#ISMAXKERNEL  = ../arm/imax.c
+#IDMAXKERNEL  = ../arm/imax.c
+#
+#ISMINKERNEL  = ../arm/imin.c
+#IDMINKERNEL  = ../arm/imin.c
+#
+SASUMKERNEL  = sasum.c
+DASUMKERNEL  = dasum.c
+CASUMKERNEL  = casum.c
+ZASUMKERNEL  = zasum.c
+#
+SAXPYKERNEL  = saxpy.c
+DAXPYKERNEL  = daxpy.c
+CAXPYKERNEL  = caxpy.c
+ZAXPYKERNEL  = zaxpy.c
+#
+SCOPYKERNEL  = scopy.c
+DCOPYKERNEL  = dcopy.c
+CCOPYKERNEL  = ccopy.c
+ZCOPYKERNEL  = zcopy.c
+#
+SDOTKERNEL   =  sdot.c
+DDOTKERNEL   =  ddot.c
+DSDOTKERNEL  =  sdot.c
+CDOTKERNEL   =  cdot.c
+ZDOTKERNEL   =  zdot.c
+#
+SNRM2KERNEL  = ../arm/nrm2.c
+DNRM2KERNEL  = ../arm/nrm2.c
+CNRM2KERNEL  = ../arm/znrm2.c
+ZNRM2KERNEL  = ../arm/znrm2.c
+#
+SROTKERNEL   = srot.c
+DROTKERNEL   = drot.c
+CROTKERNEL   = crot.c
+ZROTKERNEL   = zrot.c
+#
+SSCALKERNEL  = sscal.c
+DSCALKERNEL  = dscal.c
+CSCALKERNEL  = zscal.c
+ZSCALKERNEL  = zscal.c
+#
+SSWAPKERNEL  = sswap.c
+DSWAPKERNEL  = dswap.c
+CSWAPKERNEL  = cswap.c
+ZSWAPKERNEL  = zswap.c
+#
+
+SGEMVNKERNEL = sgemv_n.c
+DGEMVNKERNEL = dgemv_n.c
+CGEMVNKERNEL = cgemv_n.c
+ZGEMVNKERNEL = zgemv_n_4.c
+#
+SGEMVTKERNEL = sgemv_t.c
+DGEMVTKERNEL = dgemv_t.c
+CGEMVTKERNEL = cgemv_t.c
+ZGEMVTKERNEL = zgemv_t_4.c
+
+
+#SSYMV_U_KERNEL =  ../generic/symv_k.c
+#SSYMV_L_KERNEL =  ../generic/symv_k.c
+#DSYMV_U_KERNEL =  ../generic/symv_k.c
+#DSYMV_L_KERNEL =  ../generic/symv_k.c
+#QSYMV_U_KERNEL =  ../generic/symv_k.c
+#QSYMV_L_KERNEL =  ../generic/symv_k.c
+#CSYMV_U_KERNEL =  ../generic/zsymv_k.c
+#CSYMV_L_KERNEL =  ../generic/zsymv_k.c
+#ZSYMV_U_KERNEL =  ../generic/zsymv_k.c
+#ZSYMV_L_KERNEL =  ../generic/zsymv_k.c
+#XSYMV_U_KERNEL =  ../generic/zsymv_k.c
+#XSYMV_L_KERNEL =  ../generic/zsymv_k.c
+
+#ZHEMV_U_KERNEL =  ../generic/zhemv_k.c
+#ZHEMV_L_KERNEL =  ../generic/zhemv_k.c
+
+LSAME_KERNEL = ../generic/lsame.c
+SCABS_KERNEL	= ../generic/cabs.c
+DCABS_KERNEL	= ../generic/cabs.c
+QCABS_KERNEL	= ../generic/cabs.c
+
+#Dump kernel
+CGEMM3MKERNEL    = ../generic/zgemm3mkernel_dump.c
+ZGEMM3MKERNEL    = ../generic/zgemm3mkernel_dump.c

kernel/power/casum.c

@@ -46,7 +46,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #endif
 
-#if defined(POWER8)
+#if defined(POWER8)  || defined(POWER9)
 #include "casum_microk_power8.c"
 #endif
 

kernel/power/caxpy.c

@@ -0,0 +1,145 @@
+/*
+Copyright (c) 2013-2018, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+ 
+
+#ifndef HAVE_ASM_KERNEL
+#include <altivec.h> 
+static void caxpy_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i)
+{
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+
+    register __vector float valpha_r = {alpha_r, alpha_r,alpha_r, alpha_r};
+    register __vector float valpha_i = {-alpha_i, alpha_i,-alpha_i, alpha_i};
+
+#else
+    register __vector float valpha_r = {alpha_r, -alpha_r,alpha_r, -alpha_r};
+    register __vector float valpha_i = {alpha_i, alpha_i,alpha_i, alpha_i};
+#endif
+
+    __vector unsigned char swap_mask = { 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
+    register __vector float *vy = (__vector float *) y;
+    register __vector float *vx = (__vector float *) x;
+    BLASLONG i=0;
+    for (; i < n/2; i += 8) {
+
+        register __vector float vy_0 = vy[i];
+        register __vector float vy_1 = vy[i + 1];
+        register __vector float vy_2 = vy[i + 2];
+        register __vector float vy_3 = vy[i + 3];
+        register __vector float vy_4 = vy[i + 4];
+        register __vector float vy_5 = vy[i + 5];
+        register __vector float vy_6 = vy[i + 6];
+        register __vector float vy_7 = vy[i + 7];
+        register __vector float vx_0 = vx[i];
+        register __vector float vx_1 = vx[i + 1];
+        register __vector float vx_2 = vx[i + 2];
+        register __vector float vx_3 = vx[i + 3];
+        register __vector float vx_4 = vx[i + 4];
+        register __vector float vx_5 = vx[i + 5];
+        register __vector float vx_6 = vx[i + 6];
+        register __vector float vx_7 = vx[i + 7];
+        vy_0 += vx_0*valpha_r;
+        vy_1 += vx_1*valpha_r;
+        vy_2 += vx_2*valpha_r;
+        vy_3 += vx_3*valpha_r;
+        vy_4 += vx_4*valpha_r;
+        vy_5 += vx_5*valpha_r;
+        vy_6 += vx_6*valpha_r;
+        vy_7 += vx_7*valpha_r;
+        vx_0 = vec_perm(vx_0, vx_0, swap_mask);
+        vx_1 = vec_perm(vx_1, vx_1, swap_mask);
+        vx_2 = vec_perm(vx_2, vx_2, swap_mask);
+        vx_3 = vec_perm(vx_3, vx_3, swap_mask);
+        vx_4 = vec_perm(vx_4, vx_4, swap_mask);
+        vx_5 = vec_perm(vx_5, vx_5, swap_mask);
+        vx_6 = vec_perm(vx_6, vx_6, swap_mask);
+        vx_7 = vec_perm(vx_7, vx_7, swap_mask);
+        vy_0 += vx_0*valpha_i;
+        vy_1 += vx_1*valpha_i;
+        vy_2 += vx_2*valpha_i;
+        vy_3 += vx_3*valpha_i;
+        vy_4 += vx_4*valpha_i;
+        vy_5 += vx_5*valpha_i;
+        vy_6 += vx_6*valpha_i;
+        vy_7 += vx_7*valpha_i;
+        vy[i] = vy_0;
+        vy[i + 1] = vy_1;
+        vy[i + 2] = vy_2;
+        vy[i + 3] = vy_3;
+        vy[i + 4] = vy_4;
+        vy[i + 5] = vy_5 ;
+        vy[i + 6] = vy_6 ;
+        vy[i + 7] = vy_7 ;        
+
+    }
+}
+#endif
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2) {
+    BLASLONG i = 0;
+    BLASLONG ix = 0, iy = 0;
+    if (n <= 0) return (0);
+    if ((inc_x == 1) && (inc_y == 1)) {
+        BLASLONG n1 = n & -16;
+        if (n1) { 
+            caxpy_kernel_16(n1, x, y, da_r,da_i);
+            ix = 2 * n1;
+        }
+        i = n1;
+        while (i < n) {
+#if !defined(CONJ)
+            y[ix] += (da_r * x[ix] - da_i * x[ix + 1]);
+            y[ix + 1] += (da_r * x[ix + 1] + da_i * x[ix]);
+#else
+            y[ix] += (da_r * x[ix] + da_i * x[ix + 1]);
+            y[ix + 1] -= (da_r * x[ix + 1] - da_i * x[ix]);
+#endif
+            i++;
+            ix += 2;
+        }
+        return (0);
+
+    }
+    inc_x *= 2;
+    inc_y *= 2;
+    while (i < n) {
+#if !defined(CONJ)
+        y[iy] += (da_r * x[ix] - da_i * x[ix + 1]);
+        y[iy + 1] += (da_r * x[ix + 1] + da_i * x[ix]);
+#else
+        y[iy] += (da_r * x[ix] + da_i * x[ix + 1]);
+        y[iy + 1] -= (da_r * x[ix + 1] - da_i * x[ix]);
+#endif
+        ix += inc_x;
+        iy += inc_y;
+        i++;
+    }
+    return (0);
+}
+

kernel/power/ccopy.c

@@ -35,7 +35,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
-#if defined(POWER8)
+#if defined(POWER8) || defined(POWER9)
 #include "ccopy_microk_power8.c"
 #endif
 

kernel/power/cdot.c

@@ -0,0 +1,164 @@
+/*Copyright (c) 2013-201\n8, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *****************************************************************************/
+
+#include "common.h"
+
+#ifndef HAVE_KERNEL_8
+#include <altivec.h> 
+static void cdot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, float *dot)
+{
+    __vector unsigned char swap_mask = { 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
+    register __vector float *vy = (__vector float *) y;
+    register __vector float *vx = (__vector float *) x;
+    BLASLONG i = 0;
+    register __vector float vd_0  = { 0 };
+    register __vector float vd_1  = { 0 };
+    register __vector float vd_2  = { 0 };
+    register __vector float vd_3  = { 0 };
+    register __vector float vdd_0 = { 0 };
+    register __vector float vdd_1 = { 0 };
+    register __vector float vdd_2 = { 0 };
+    register __vector float vdd_3 = { 0 };
+    for (; i < n/2; i += 4) {
+
+        register __vector float vyy_0 ;
+        register __vector float vyy_1 ;
+        register __vector float vyy_2 ;
+        register __vector float vyy_3 ;
+
+        register __vector float vy_0 = vy[i];
+        register __vector float vy_1 = vy[i + 1];
+        register __vector float vy_2 = vy[i + 2];
+        register __vector float vy_3 = vy[i + 3]; 
+        register __vector float vx_0= vx[i];
+        register __vector float vx_1 = vx[i + 1];
+        register __vector float vx_2 = vx[i + 2];
+        register __vector float vx_3 = vx[i + 3]; 
+        vyy_0 = vec_perm(vy_0, vy_0, swap_mask);
+        vyy_1 = vec_perm(vy_1, vy_1, swap_mask);
+        vyy_2 = vec_perm(vy_2, vy_2, swap_mask);
+        vyy_3 = vec_perm(vy_3, vy_3, swap_mask);  
+
+        vd_0 += vx_0 * vy_0;
+        vd_1 += vx_1 * vy_1;
+        vd_2 += vx_2 * vy_2;
+        vd_3 += vx_3 * vy_3;
+
+        vdd_0 += vx_0 * vyy_0;
+        vdd_1 += vx_1 * vyy_1;
+        vdd_2 += vx_2 * vyy_2;
+        vdd_3 += vx_3 * vyy_3;       
+       
+
+    }
+    //aggregate
+    vd_0 = vd_0 + vd_1 +vd_2 +vd_3;
+    vdd_0= vdd_0 + vdd_1 +vdd_2 +vdd_3; 
+     //reverse and aggregate 
+    vd_1=vec_xxpermdi(vd_0,vd_0,2)  ;
+    vdd_1=vec_xxpermdi(vdd_0,vdd_0,2);
+    vd_2=vd_0+vd_1;
+    vdd_2=vdd_0+vdd_1;
+
+    dot[0]=vd_2[0];
+    dot[1]=vd_2[1];
+    dot[2]=vdd_2[0];
+    dot[3]=vdd_2[1];
+ 
+}
+#endif
+ 
+
+OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y) {
+    BLASLONG i = 0;
+    BLASLONG ix=0, iy=0;
+    OPENBLAS_COMPLEX_FLOAT result;
+    FLOAT dot[4] __attribute__ ((aligned(16))) = {0.0, 0.0, 0.0, 0.0};
+
+    if (n <= 0) {
+        CREAL(result) = 0.0;
+        CIMAG(result) = 0.0;
+        return (result);
+
+    }
+
+    if ((inc_x == 1) && (inc_y == 1)) {
+
+        BLASLONG n1 = n & -8;
+        BLASLONG j=0; 
+
+        if (n1){
+            cdot_kernel_8(n1, x, y, dot);
+            i = n1;
+            j = n1 <<1;
+        }
+ 
+
+        while (i < n) {
+
+            dot[0] += x[j] * y[j];
+            dot[1] += x[j + 1] * y[j + 1];
+            dot[2] += x[j] * y[j + 1];
+            dot[3] += x[j + 1] * y[j];
+
+            j += 2;
+            i++;
+
+        }
+
+
+    } else {
+        i = 0;
+        ix = 0;
+        iy = 0;
+        inc_x <<= 1;
+        inc_y <<= 1;
+        while (i < n) {
+
+            dot[0] += x[ix] * y[iy];
+            dot[1] += x[ix + 1] * y[iy + 1];
+            dot[2] += x[ix] * y[iy + 1];
+            dot[3] += x[ix + 1] * y[iy];
+
+            ix += inc_x;
+            iy += inc_y;
+            i++;
+
+        }
+    }
+
+#if !defined(CONJ)
+    CREAL(result) = dot[0] - dot[1];
+    CIMAG(result) = dot[2] + dot[3];
+#else
+    CREAL(result) = dot[0] + dot[1];
+    CIMAG(result) = dot[2] - dot[3];
+
+#endif
+
+    return (result);
+
+}

kernel/power/cgemv_n.c

@@ -0,0 +1,585 @@
+/***************************************************************************
+Copyright (c) 2019, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *****************************************************************************/
+
+#include <stdlib.h>
+#include <stdio.h>
+#include "common.h" 
+#include <altivec.h>   
+#define NBMAX 1024
+
+
+static const unsigned char swap_mask_arr[]={ 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
+
+ 
+static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y) {
+  
+ FLOAT *a0, *a1, *a2, *a3;
+    a0 = ap;
+    a1 = ap + lda;
+    a2 = a1 + lda;
+    a3 = a2 + lda;
+    __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+    register __vector float vx0_r = {x[0], x[0],x[0], x[0]};
+    register __vector float vx0_i = {-x[1], x[1],-x[1], x[1]};
+    register __vector float vx1_r = {x[2], x[2],x[2], x[2]};
+    register __vector float vx1_i = {-x[3], x[3],-x[3], x[3]};
+    register __vector float vx2_r = {x[4], x[4],x[4], x[4]};
+    register __vector float vx2_i = {-x[5], x[5],-x[5], x[5]};
+    register __vector float vx3_r = {x[6], x[6],x[6], x[6]};
+    register __vector float vx3_i = {-x[7], x[7],-x[7], x[7]};
+#else
+    register __vector float vx0_r = {x[0], -x[0],x[0], -x[0]};
+    register __vector float vx0_i = {x[1], x[1],x[1], x[1]};
+    register __vector float vx1_r = {x[2], -x[2],x[2], -x[2]};
+    register __vector float vx1_i = {x[3], x[3],x[3], x[3]};
+    register __vector float vx2_r = {x[4], -x[4],x[4], -x[4]};
+    register __vector float vx2_i = {x[5], x[5],x[5], x[5]};
+    register __vector float vx3_r = {x[6], -x[6],x[6], -x[6]};
+    register __vector float vx3_i = {x[7], x[7],x[7], x[7]};
+#endif
+    register __vector float *vy = (__vector float *) y;
+    register __vector float *vptr_a0 = (__vector float *) a0;
+    register __vector float *vptr_a1 = (__vector float *) a1;
+    register __vector float *vptr_a2 = (__vector float *) a2;
+    register __vector float *vptr_a3 = (__vector float *) a3; 
+    BLASLONG  i = 0; 
+    for (;i< n / 2; i+=2) {
+        register __vector float vy_0  = vy[i];
+        register __vector float vy_1  = vy[i + 1];
+        register __vector float va0   = vptr_a0[i];
+        register __vector float va1   = vptr_a1[i];
+        register __vector float va2   = vptr_a2[i];
+        register __vector float va3   = vptr_a3[i];
+        register __vector float va0_1 = vptr_a0[i + 1];
+        register __vector float va1_1 = vptr_a1[i + 1];
+        register __vector float va2_1 = vptr_a2[i + 1];
+        register __vector float va3_1 = vptr_a3[i + 1];
+
+        vy_0 += va0*vx0_r + va1*vx1_r + va2*vx2_r + va3*vx3_r;
+        vy_1 += va0_1*vx0_r + va1_1*vx1_r + va2_1*vx2_r + va3_1*vx3_r;
+        va0   = vec_perm(va0, va0,swap_mask);
+        va0_1 = vec_perm(va0_1, va0_1,swap_mask);
+        va1   = vec_perm(va1, va1,swap_mask);
+        va1_1 = vec_perm(va1_1, va1_1,swap_mask);
+        va2   = vec_perm(va2, va2,swap_mask);
+        va2_1 = vec_perm(va2_1, va2_1,swap_mask);
+        va3   = vec_perm(va3, va3,swap_mask);
+        va3_1 = vec_perm(va3_1, va3_1,swap_mask);
+        vy_0 += va0*vx0_i + va1*vx1_i + va2*vx2_i + va3*vx3_i;
+        vy_1 += va0_1*vx0_i + va1_1*vx1_i + va2_1*vx2_i + va3_1*vx3_i;
+
+        vy[i] = vy_0;
+        vy[i + 1] = vy_1;
+    }
+
+}	
+ 
+
+
+static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y) {
+ 
+    FLOAT *a0, *a1;
+    a0 = ap;
+    a1 = ap + lda; 
+    __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+    register __vector float vx0_r = {x[0], x[0],x[0], x[0]};
+    register __vector float vx0_i = {-x[1], x[1],-x[1], x[1]};
+    register __vector float vx1_r = {x[2], x[2],x[2], x[2]};
+    register __vector float vx1_i = {-x[3], x[3],-x[3], x[3]}; 
+#else
+    register __vector float vx0_r = {x[0], -x[0],x[0], -x[0]};
+    register __vector float vx0_i = {x[1], x[1],x[1], x[1]};
+    register __vector float vx1_r = {x[2], -x[2],x[2], -x[2]};
+    register __vector float vx1_i = {x[3], x[3],x[3], x[3]}; 
+#endif
+    register __vector float *vy = (__vector float *) y;
+    register __vector float *vptr_a0 = (__vector float *) a0;
+    register __vector float *vptr_a1 = (__vector float *) a1; 
+    BLASLONG  i = 0; 
+    for (;i< n / 2; i+=2) {
+        register __vector float vy_0  = vy[i];
+        register __vector float vy_1  = vy[i + 1];
+        register __vector float va0   = vptr_a0[i];
+        register __vector float va1   = vptr_a1[i]; 
+        register __vector float va0_1 = vptr_a0[i + 1];
+        register __vector float va1_1 = vptr_a1[i + 1]; 
+        register __vector float va0x   = vec_perm(va0, va0,swap_mask);
+        register __vector float va0x_1 = vec_perm(va0_1, va0_1,swap_mask);
+        register __vector float va1x   = vec_perm(va1, va1,swap_mask);
+        register __vector float va1x_1 = vec_perm(va1_1, va1_1,swap_mask);
+        vy_0 += va0*vx0_r + va1*vx1_r + va0x*vx0_i + va1x*vx1_i;
+        vy_1 += va0_1*vx0_r + va1_1*vx1_r + va0x_1*vx0_i + va1x_1*vx1_i; 
+
+        vy[i] = vy_0;
+        vy[i + 1] = vy_1;
+    }
+
+}
+
+ 
+
+static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) {
+
+    __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+    register __vector float vx0_r = {x[0], x[0],x[0], x[0]};
+    register __vector float vx0_i = {-x[1], x[1],-x[1], x[1]}; 
+#else
+    register __vector float vx0_r = {x[0], -x[0],x[0], -x[0]};
+    register __vector float vx0_i = {x[1], x[1],x[1], x[1]}; 
+#endif
+    register __vector float *vy = (__vector float *) y;
+    register __vector float *vptr_a0 = (__vector float *) ap; 
+    BLASLONG  i = 0; 
+    for (;i< n / 2; i+=2) {
+        register __vector float vy_0  = vy[i];
+        register __vector float vy_1  = vy[i + 1];
+        register __vector float va0   = vptr_a0[i];
+        register __vector float va0_1 = vptr_a0[i + 1]; 
+        register __vector float va0x   = vec_perm(va0, va0,swap_mask);
+        register __vector float va0x_1 = vec_perm(va0_1, va0_1,swap_mask);
+        vy_0 += va0*vx0_r + va0x*vx0_i;
+        vy_1 += va0_1*vx0_r + va0x_1*vx0_i; 
+
+        vy[i] = vy_0;
+        vy[i + 1] = vy_1;
+    }
+}
+
+
+
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest, FLOAT alpha_r, FLOAT alpha_i) {
+    BLASLONG i;
+
+
+    if (inc_dest != 2) {
+ 		FLOAT temp_r;
+		FLOAT temp_i;
+		for ( i=0; i<n; i++ )
+		{
+#if !defined(XCONJ) 
+			temp_r = alpha_r * src[0] - alpha_i * src[1];
+			temp_i = alpha_r * src[1] + alpha_i * src[0];
+#else
+			temp_r =  alpha_r * src[0] + alpha_i * src[1];
+			temp_i = -alpha_r * src[1] + alpha_i * src[0];
+#endif
+
+			*dest += temp_r;
+			*(dest+1) += temp_i;
+
+			src+=2;
+			dest += inc_dest;
+		}
+        return;
+    } else {
+        __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);
+#if   !defined(XCONJ) 
+
+        register __vector float valpha_r = {alpha_r, alpha_r, alpha_r, alpha_r};
+        register __vector float valpha_i = {-alpha_i, alpha_i, -alpha_i, alpha_i};
+
+#else
+        register __vector float valpha_r = {alpha_r, -alpha_r, alpha_r, -alpha_r};
+        register __vector float valpha_i = {alpha_i, alpha_i, alpha_i, alpha_i};
+#endif
+
+        register __vector float *vptr_src = (__vector float *) src;
+        register __vector float *vptr_y = (__vector float *) dest; 
+        for (i = 0; i < n/2; i += 2 ){
+
+            register __vector float vy_0 = vptr_y[i];
+            register __vector float vy_1 = vptr_y[i +1]; 
+
+            register __vector float vsrc = vptr_src[i];
+            register __vector float vsrc_1 = vptr_src[i + 1]; 
+            register __vector float vsrcx = vec_perm(vsrc, vsrc, swap_mask);
+            register __vector float vsrcx_1 = vec_perm(vsrc_1, vsrc_1, swap_mask);
+
+            vy_0 += vsrc*valpha_r + vsrcx*valpha_i;
+            vy_1 += vsrc_1*valpha_r +  vsrcx_1*valpha_i;  
+            vptr_y[i] = vy_0;
+            vptr_y[i+1 ] = vy_1;  
+
+        }
+ 
+    }
+    return;
+}
+
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT * buffer) {
+    BLASLONG i;
+    FLOAT *a_ptr;
+    FLOAT *x_ptr;
+    FLOAT *y_ptr;
+
+    BLASLONG n1;
+    BLASLONG m1;
+    BLASLONG m2;
+    BLASLONG m3;
+    BLASLONG n2;
+
+    FLOAT xbuffer[8], *ybuffer;
+
+    if (m < 1) return (0);
+    if (n < 1) return (0);
+
+    ybuffer = buffer;
+
+    inc_x *= 2;
+    inc_y *= 2;
+    lda *= 2;
+
+    n1 = n / 4;
+    n2 = n % 4;
+
+    m3 = m % 4;
+    m1 = m - (m % 4);
+    m2 = (m % NBMAX) - (m % 4);
+
+    y_ptr = y;
+
+    BLASLONG NB = NBMAX;
+
+    while (NB == NBMAX) {
+
+        m1 -= NB;
+        if (m1 < 0) {
+            if (m2 == 0) break;
+            NB = m2;
+        }
+
+        a_ptr = a;
+
+        x_ptr = x; 
+
+        memset(ybuffer, 0, NB * 2*sizeof(FLOAT));  
+
+        if (inc_x == 2) {
+
+            for (i = 0; i < n1; i++) {
+                cgemv_kernel_4x4(NB, lda, a_ptr, x_ptr, ybuffer);
+
+                a_ptr += lda << 2;
+                x_ptr += 8;
+            }
+
+            if (n2 & 2) {
+                cgemv_kernel_4x2(NB, lda, a_ptr, x_ptr, ybuffer);
+                x_ptr += 4;
+                a_ptr += 2 * lda;
+
+            }
+
+            if (n2 & 1) {
+                cgemv_kernel_4x1(NB, a_ptr, x_ptr, ybuffer);
+                x_ptr += 2;
+                a_ptr += lda;
+
+            }
+        } else {
+
+            for (i = 0; i < n1; i++) {
+
+                xbuffer[0] = x_ptr[0];
+                xbuffer[1] = x_ptr[1];
+                x_ptr += inc_x;
+                xbuffer[2] = x_ptr[0];
+                xbuffer[3] = x_ptr[1];
+                x_ptr += inc_x;
+                xbuffer[4] = x_ptr[0];
+                xbuffer[5] = x_ptr[1];
+                x_ptr += inc_x;
+                xbuffer[6] = x_ptr[0];
+                xbuffer[7] = x_ptr[1];
+                x_ptr += inc_x;
+
+                cgemv_kernel_4x4(NB, lda, a_ptr, xbuffer, ybuffer);
+
+                a_ptr += lda << 2;
+            }
+
+            for (i = 0; i < n2; i++) {
+                xbuffer[0] = x_ptr[0];
+                xbuffer[1] = x_ptr[1];
+                x_ptr += inc_x;
+                cgemv_kernel_4x1(NB, a_ptr, xbuffer, ybuffer);
+                a_ptr += lda;
+
+            }
+
+        }
+
+        add_y(NB, ybuffer, y_ptr, inc_y, alpha_r, alpha_i);
+        a += 2 * NB;
+        y_ptr += NB * inc_y;
+    }
+
+    if (m3 == 0) return (0);
+
+    if (m3 == 1) {
+        a_ptr = a;
+        x_ptr = x;
+        FLOAT temp_r = 0.0;
+        FLOAT temp_i = 0.0;
+
+        if (lda == 2 && inc_x == 2) {
+
+            for (i = 0; i < (n & -2); i += 2) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r += a_ptr[2] * x_ptr[2] - a_ptr[3] * x_ptr[3];
+                temp_i += a_ptr[2] * x_ptr[3] + a_ptr[3] * x_ptr[2];
+#else
+                temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r += a_ptr[2] * x_ptr[2] + a_ptr[3] * x_ptr[3];
+                temp_i += a_ptr[2] * x_ptr[3] - a_ptr[3] * x_ptr[2];
+#endif
+
+                a_ptr += 4;
+                x_ptr += 4;
+            }
+
+            for (; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+#else
+                temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+#endif
+
+                a_ptr += 2;
+                x_ptr += 2;
+            }
+
+        } else {
+
+            for (i = 0; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+#else
+                temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+#endif
+
+                a_ptr += lda;
+                x_ptr += inc_x;
+            }
+
+        }
+#if !defined(XCONJ) 
+        y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+        y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+        y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+        y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+        return (0);
+    }
+
+    if (m3 == 2) {
+        a_ptr = a;
+        x_ptr = x;
+        FLOAT temp_r0 = 0.0;
+        FLOAT temp_i0 = 0.0;
+        FLOAT temp_r1 = 0.0;
+        FLOAT temp_i1 = 0.0;
+
+        if (lda == 4 && inc_x == 2) {
+
+            for (i = 0; i < (n & -2); i += 2) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+
+                temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
+
+                temp_r0 += a_ptr[4] * x_ptr[2] - a_ptr[5] * x_ptr[3];
+                temp_i0 += a_ptr[4] * x_ptr[3] + a_ptr[5] * x_ptr[2];
+                temp_r1 += a_ptr[6] * x_ptr[2] - a_ptr[7] * x_ptr[3];
+                temp_i1 += a_ptr[6] * x_ptr[3] + a_ptr[7] * x_ptr[2];
+#else
+                temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
+
+                temp_r0 += a_ptr[4] * x_ptr[2] + a_ptr[5] * x_ptr[3];
+                temp_i0 += a_ptr[4] * x_ptr[3] - a_ptr[5] * x_ptr[2];
+                temp_r1 += a_ptr[6] * x_ptr[2] + a_ptr[7] * x_ptr[3];
+                temp_i1 += a_ptr[6] * x_ptr[3] - a_ptr[7] * x_ptr[2];
+#endif
+
+                a_ptr += 8;
+                x_ptr += 4;
+            }
+
+            for (; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
+#else
+                temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
+#endif
+
+                a_ptr += 4;
+                x_ptr += 2;
+            }
+
+        } else {
+
+            for (i = 0; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
+#else
+                temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
+#endif
+
+                a_ptr += lda;
+                x_ptr += inc_x;
+            }
+
+        }
+#if !defined(XCONJ) 
+        y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+        y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
+        y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
+#else
+        y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+        y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
+        y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+#endif
+        return (0);
+    }
+
+    if (m3 == 3) {
+        a_ptr = a;
+        x_ptr = x;
+        FLOAT temp_r0 = 0.0;
+        FLOAT temp_i0 = 0.0;
+        FLOAT temp_r1 = 0.0;
+        FLOAT temp_i1 = 0.0;
+        FLOAT temp_r2 = 0.0;
+        FLOAT temp_i2 = 0.0;
+
+        if (lda == 6 && inc_x == 2) {
+
+            for (i = 0; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
+                temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
+                temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
+#else
+                temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
+                temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
+                temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
+#endif
+
+                a_ptr += 6;
+                x_ptr += 2;
+            }
+
+        } else {
+
+            for (i = 0; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
+                temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
+                temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
+#else
+                temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
+                temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
+                temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
+#endif
+
+                a_ptr += lda;
+                x_ptr += inc_x;
+            }
+
+        }
+#if !defined(XCONJ) 
+        y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+        y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
+        y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r2 - alpha_i * temp_i2;
+        y_ptr[1] += alpha_r * temp_i2 + alpha_i * temp_r2;
+#else
+        y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+        y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
+        y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r2 + alpha_i * temp_i2;
+        y_ptr[1] -= alpha_r * temp_i2 - alpha_i * temp_r2;
+#endif
+        return (0);
+    }
+
+    return (0);
+}
+

kernel/power/cgemv_t.c

@@ -0,0 +1,571 @@
+/***************************************************************************
+Copyright (c) 2019, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *****************************************************************************/
+
+#include "common.h"
+
+#define NBMAX 1024 
+#include <altivec.h> 
+static const unsigned char swap_mask_arr[]={ 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
+
+static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
+    BLASLONG i;
+    FLOAT *a0, *a1, *a2, *a3;
+    a0 = ap;
+    a1 = ap + lda;
+    a2 = a1 + lda;
+    a3 = a2 + lda;
+    __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);
+    //p for positive(real*real,image*image,real*real,image*image) r for image (real*image,image*real,real*image,image*real)
+    register __vector float vtemp0_p = {0.0, 0.0,0.0,0.0};
+    register __vector float vtemp0_r = {0.0, 0.0,0.0,0.0};
+    register __vector float vtemp1_p = {0.0, 0.0,0.0,0.0};
+    register __vector float vtemp1_r = {0.0, 0.0,0.0,0.0};
+    register __vector float vtemp2_p = {0.0, 0.0,0.0,0.0};
+    register __vector float vtemp2_r = {0.0, 0.0,0.0,0.0};
+    register __vector float vtemp3_p = {0.0, 0.0,0.0,0.0};
+    register __vector float vtemp3_r = {0.0, 0.0,0.0,0.0};
+    __vector float* va0 = (__vector float*) a0;
+    __vector float* va1 = (__vector float*) a1;
+    __vector float* va2 = (__vector float*) a2;
+    __vector float* va3 = (__vector float*) a3;
+    __vector float* v_x = (__vector float*) x;
+
+    for (i = 0; i < n / 2; i+=2) {
+        register __vector float vx_0  = v_x[i]; 
+        register __vector float vx_1  = v_x[i+1];         
+        register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask);
+        register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask);
+
+        vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ;
+        vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1]; 
+        vtemp1_p += vx_0*va1[i] + vx_1*va1[i+1];
+        vtemp1_r += vxr_0*va1[i] + vxr_1*va1[i+1]; 
+        vtemp2_p += vx_0*va2[i] + vx_1*va2[i+1];
+        vtemp2_r += vxr_0*va2[i] + vxr_1*va2[i+1]; 
+        vtemp3_p += vx_0*va3[i] + vx_1*va3[i+1];
+        vtemp3_r += vxr_0*va3[i] + vxr_1*va3[i+1]; 
+
+    }
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+
+    register FLOAT temp_r0 = vtemp0_p[0] - vtemp0_p[1] + vtemp0_p[2] - vtemp0_p[3];
+    register FLOAT temp_i0 = vtemp0_r[0] + vtemp0_r[1] + vtemp0_r[2] + vtemp0_r[3];
+
+    register FLOAT temp_r1 = vtemp1_p[0] - vtemp1_p[1] + vtemp1_p[2] - vtemp1_p[3];
+    register FLOAT temp_i1 = vtemp1_r[0] + vtemp1_r[1] + vtemp1_r[2] + vtemp1_r[3];
+
+    register FLOAT temp_r2 = vtemp2_p[0] - vtemp2_p[1] + vtemp2_p[2] - vtemp2_p[3];
+    register FLOAT temp_i2 = vtemp2_r[0] + vtemp2_r[1] + vtemp2_r[2] + vtemp2_r[3];
+
+    register FLOAT temp_r3 = vtemp3_p[0] - vtemp3_p[1] + vtemp3_p[2] - vtemp3_p[3];
+    register FLOAT temp_i3 = vtemp3_r[0] + vtemp3_r[1] + vtemp3_r[2] + vtemp3_r[3];
+
+#else
+    register FLOAT temp_r0 = vtemp0_p[0] + vtemp0_p[1] + vtemp0_p[2] + vtemp0_p[3];
+    register FLOAT temp_i0 = vtemp0_r[0] - vtemp0_r[1] + vtemp0_r[2] - vtemp0_r[3];
+
+    register FLOAT temp_r1 = vtemp1_p[0] + vtemp1_p[1] + vtemp1_p[2] + vtemp1_p[3];
+    register FLOAT temp_i1 = vtemp1_r[0] - vtemp1_r[1] + vtemp1_r[2] - vtemp1_r[3];
+
+    register FLOAT temp_r2 = vtemp2_p[0] + vtemp2_p[1] + vtemp2_p[2] + vtemp2_p[3];
+    register FLOAT temp_i2 = vtemp2_r[0] - vtemp2_r[1] + vtemp2_r[2] - vtemp2_r[3];
+
+    register FLOAT temp_r3 = vtemp3_p[0] + vtemp3_p[1] + vtemp3_p[2] + vtemp3_p[3];
+    register FLOAT temp_i3 = vtemp3_r[0] - vtemp3_r[1] + vtemp3_r[2] - vtemp3_r[3];
+
+#endif    
+
+#if !defined(XCONJ)
+
+    y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+    y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;
+    y[3] += alpha_r * temp_i1 + alpha_i * temp_r1;
+    y[4] += alpha_r * temp_r2 - alpha_i * temp_i2;
+    y[5] += alpha_r * temp_i2 + alpha_i * temp_r2;
+    y[6] += alpha_r * temp_r3 - alpha_i * temp_i3;
+    y[7] += alpha_r * temp_i3 + alpha_i * temp_r3;
+
+#else
+
+    y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+    y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;
+    y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+    y[4] += alpha_r * temp_r2 + alpha_i * temp_i2;
+    y[5] -= alpha_r * temp_i2 - alpha_i * temp_r2;
+    y[6] += alpha_r * temp_r3 + alpha_i * temp_i3;
+    y[7] -= alpha_r * temp_i3 - alpha_i * temp_r3;
+
+#endif
+
+}
+ 
+
+static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
+    BLASLONG i;
+    FLOAT *a0, *a1;
+    a0 = ap;
+    a1 = ap + lda; 
+    __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);
+    //p for positive(real*real,image*image,real*real,image*image) r for image (real*image,image*real,real*image,image*real)
+    register __vector float vtemp0_p = {0.0, 0.0,0.0,0.0};
+    register __vector float vtemp0_r = {0.0, 0.0,0.0,0.0};
+    register __vector float vtemp1_p = {0.0, 0.0,0.0,0.0};
+    register __vector float vtemp1_r = {0.0, 0.0,0.0,0.0}; 
+    __vector float* va0 = (__vector float*) a0;
+    __vector float* va1 = (__vector float*) a1; 
+    __vector float* v_x = (__vector float*) x;
+
+    for (i = 0; i < n / 2; i+=2) {
+        register __vector float vx_0  = v_x[i]; 
+        register __vector float vx_1  = v_x[i+1];         
+        register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask);
+        register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask);
+
+        vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ;
+        vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1]; 
+        vtemp1_p += vx_0*va1[i] + vx_1*va1[i+1];
+        vtemp1_r += vxr_0*va1[i] + vxr_1*va1[i+1];  
+
+    }
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+
+    register FLOAT temp_r0 = vtemp0_p[0] - vtemp0_p[1] + vtemp0_p[2] - vtemp0_p[3];
+    register FLOAT temp_i0 = vtemp0_r[0] + vtemp0_r[1] + vtemp0_r[2] + vtemp0_r[3];
+
+    register FLOAT temp_r1 = vtemp1_p[0] - vtemp1_p[1] + vtemp1_p[2] - vtemp1_p[3];
+    register FLOAT temp_i1 = vtemp1_r[0] + vtemp1_r[1] + vtemp1_r[2] + vtemp1_r[3];
+ 
+
+#else
+    register FLOAT temp_r0 = vtemp0_p[0] + vtemp0_p[1] + vtemp0_p[2] + vtemp0_p[3];
+    register FLOAT temp_i0 = vtemp0_r[0] - vtemp0_r[1] + vtemp0_r[2] - vtemp0_r[3];
+
+    register FLOAT temp_r1 = vtemp1_p[0] + vtemp1_p[1] + vtemp1_p[2] + vtemp1_p[3];
+    register FLOAT temp_i1 = vtemp1_r[0] - vtemp1_r[1] + vtemp1_r[2] - vtemp1_r[3]; 
+
+#endif    
+
+#if !defined(XCONJ)
+
+    y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+    y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;
+    y[3] += alpha_r * temp_i1 + alpha_i * temp_r1; 
+
+#else
+
+    y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+    y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;
+    y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1; 
+
+#endif
+  
+}
+ 
+
+static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
+   BLASLONG i;  
+    __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);
+    //p for positive(real*real,image*image,real*real,image*image) r for image (real*image,image*real,real*image,image*real)
+    register __vector float vtemp0_p = {0.0, 0.0,0.0,0.0};
+    register __vector float vtemp0_r = {0.0, 0.0,0.0,0.0}; 
+    __vector float* va0 = (__vector float*) ap; 
+    __vector float* v_x = (__vector float*) x;
+
+    for (i = 0; i < n / 2; i+=2) {
+        register __vector float vx_0  = v_x[i]; 
+        register __vector float vx_1  = v_x[i+1];         
+        register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask);
+        register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask);
+
+        vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ;
+        vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1];  
+
+    }
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+
+    register FLOAT temp_r0 = vtemp0_p[0] - vtemp0_p[1] + vtemp0_p[2] - vtemp0_p[3];
+    register FLOAT temp_i0 = vtemp0_r[0] + vtemp0_r[1] + vtemp0_r[2] + vtemp0_r[3]; 
+
+#else
+    register FLOAT temp_r0 = vtemp0_p[0] + vtemp0_p[1] + vtemp0_p[2] + vtemp0_p[3];
+    register FLOAT temp_i0 = vtemp0_r[0] - vtemp0_r[1] + vtemp0_r[2] - vtemp0_r[3]; 
+
+#endif    
+
+#if !defined(XCONJ)
+
+    y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+    y[1] += alpha_r * temp_i0 + alpha_i * temp_r0; 
+
+#else
+
+    y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+    y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0; 
+
+#endif
+
+
+}
+ 
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src) {
+    BLASLONG i;
+    for (i = 0; i < n; i++) {
+        *dest = *src;
+        *(dest + 1) = *(src + 1);
+        dest += 2;
+        src += inc_src;
+    }
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer) {
+    BLASLONG i;
+    BLASLONG j;
+    FLOAT *a_ptr;
+    FLOAT *x_ptr;
+    FLOAT *y_ptr;
+
+    BLASLONG n1;
+    BLASLONG m1;
+    BLASLONG m2;
+    BLASLONG m3;
+    BLASLONG n2;
+
+    FLOAT ybuffer[8], *xbuffer;
+
+    if (m < 1) return (0);
+    if (n < 1) return (0);
+
+    inc_x <<= 1;
+    inc_y <<= 1;
+    lda <<= 1;
+
+    xbuffer = buffer;
+
+    n1 = n >> 2;
+    n2 = n & 3;
+
+    m3 = m & 3;
+    m1 = m - m3;
+    m2 = (m & (NBMAX - 1)) - m3;
+
+    BLASLONG NB = NBMAX;
+
+    while (NB == NBMAX) {
+
+        m1 -= NB;
+        if (m1 < 0) {
+            if (m2 == 0) break;
+            NB = m2;
+        }
+
+        y_ptr = y;
+        a_ptr = a;
+        x_ptr = x;
+
+        if (inc_x != 2)
+            copy_x(NB, x_ptr, xbuffer, inc_x);
+        else
+            xbuffer = x_ptr;
+
+        if (inc_y == 2) {
+
+            for (i = 0; i < n1; i++) {
+                cgemv_kernel_4x4(NB, lda, a_ptr, xbuffer, y_ptr, alpha_r, alpha_i);
+                a_ptr += lda << 2;
+                y_ptr += 8;
+
+            }
+
+            if (n2 & 2) {
+                cgemv_kernel_4x2(NB, lda, a_ptr, xbuffer, y_ptr, alpha_r, alpha_i);
+                a_ptr += lda << 1;
+                y_ptr += 4;
+
+            }
+
+            if (n2 & 1) {
+                cgemv_kernel_4x1(NB, a_ptr, xbuffer, y_ptr, alpha_r, alpha_i);
+                a_ptr += lda;
+                y_ptr += 2;
+
+            }
+
+        } else {
+
+            for (i = 0; i < n1; i++) {
+                memset(ybuffer, 0, sizeof (ybuffer));
+                cgemv_kernel_4x4(NB, lda, a_ptr, xbuffer, ybuffer, alpha_r, alpha_i);
+
+                a_ptr += lda << 2;
+
+                y_ptr[0] += ybuffer[0];
+                y_ptr[1] += ybuffer[1];
+                y_ptr += inc_y;
+                y_ptr[0] += ybuffer[2];
+                y_ptr[1] += ybuffer[3];
+                y_ptr += inc_y;
+                y_ptr[0] += ybuffer[4];
+                y_ptr[1] += ybuffer[5];
+                y_ptr += inc_y;
+                y_ptr[0] += ybuffer[6];
+                y_ptr[1] += ybuffer[7];
+                y_ptr += inc_y;
+
+            }
+
+            for (i = 0; i < n2; i++) {
+                memset(ybuffer, 0, sizeof (ybuffer));
+                cgemv_kernel_4x1(NB, a_ptr, xbuffer, ybuffer, alpha_r, alpha_i);
+                a_ptr += lda;
+                y_ptr[0] += ybuffer[0];
+                y_ptr[1] += ybuffer[1];
+                y_ptr += inc_y;
+
+            }
+
+        }
+        a += 2 * NB;
+        x += NB * inc_x;
+    }
+
+    if (m3 == 0) return (0);
+
+    x_ptr = x;
+    j = 0;
+    a_ptr = a;
+    y_ptr = y;
+
+    if (m3 == 3) {
+
+        FLOAT temp_r;
+        FLOAT temp_i;
+        FLOAT x0 = x_ptr[0];
+        FLOAT x1 = x_ptr[1];
+        x_ptr += inc_x;
+        FLOAT x2 = x_ptr[0];
+        FLOAT x3 = x_ptr[1];
+        x_ptr += inc_x;
+        FLOAT x4 = x_ptr[0];
+        FLOAT x5 = x_ptr[1];
+        while (j < n) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+            temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
+            temp_r += a_ptr[4] * x4 - a_ptr[5] * x5;
+            temp_i += a_ptr[4] * x5 + a_ptr[5] * x4;
+#else
+
+            temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
+            temp_r += a_ptr[4] * x4 + a_ptr[5] * x5;
+            temp_i += a_ptr[4] * x5 - a_ptr[5] * x4;
+#endif
+
+#if !defined(XCONJ) 
+            y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+            y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+            y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+            y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+
+            a_ptr += lda;
+            y_ptr += inc_y;
+            j++;
+        }
+        return (0);
+    }
+
+    if (m3 == 2) {
+
+        FLOAT temp_r;
+        FLOAT temp_i;
+        FLOAT temp_r1;
+        FLOAT temp_i1;
+        FLOAT x0 = x_ptr[0];
+        FLOAT x1 = x_ptr[1];
+        x_ptr += inc_x;
+        FLOAT x2 = x_ptr[0];
+        FLOAT x3 = x_ptr[1];
+
+        while (j < (n & -2)) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+            temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
+            a_ptr += lda;
+            temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;
+            temp_r1 += a_ptr[2] * x2 - a_ptr[3] * x3;
+            temp_i1 += a_ptr[2] * x3 + a_ptr[3] * x2;
+#else
+
+            temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
+            a_ptr += lda;
+            temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;
+            temp_r1 += a_ptr[2] * x2 + a_ptr[3] * x3;
+            temp_i1 += a_ptr[2] * x3 - a_ptr[3] * x2;
+#endif
+
+#if !defined(XCONJ) 
+            y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+            y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+            y_ptr += inc_y;
+            y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
+            y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
+#else
+            y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+            y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+            y_ptr += inc_y;
+            y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
+            y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+#endif
+
+            a_ptr += lda;
+            y_ptr += inc_y;
+            j += 2;
+        }
+
+        while (j < n) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+            temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
+#else
+
+            temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
+#endif
+
+#if !defined(XCONJ) 
+            y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+            y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+            y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+            y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+
+            a_ptr += lda;
+            y_ptr += inc_y;
+            j++;
+        }
+
+        return (0);
+    }
+
+    if (m3 == 1) {
+
+        FLOAT temp_r;
+        FLOAT temp_i;
+        FLOAT temp_r1;
+        FLOAT temp_i1;
+        FLOAT x0 = x_ptr[0];
+        FLOAT x1 = x_ptr[1];
+
+        while (j < (n & -2)) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+            temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
+            a_ptr += lda;
+            temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;
+#else
+
+            temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
+            a_ptr += lda;
+            temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;
+#endif
+
+#if !defined(XCONJ) 
+            y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+            y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+            y_ptr += inc_y;
+            y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
+            y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
+#else
+            y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+            y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+            y_ptr += inc_y;
+            y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
+            y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+#endif
+
+            a_ptr += lda;
+            y_ptr += inc_y;
+            j += 2;
+        }
+
+        while (j < n) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+            temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
+#else
+
+            temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
+#endif
+
+#if !defined(XCONJ) 
+            y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+            y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+            y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+            y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+
+            a_ptr += lda;
+            y_ptr += inc_y;
+            j++;
+        }
+        return (0);
+    }
+
+    return (0);
+
+}
+

kernel/power/crot.c

@@ -0,0 +1,231 @@
+/***************************************************************************
+Copyright (c) 2013-2018, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+ 
+#include "common.h"
+ 
+#if defined(POWER8) || defined(POWER9)
+
+static void crot_kernel_8 (long n, float *x, float *y, float c, float s)
+{
+  __vector float t0;
+  __vector float t1;
+  __vector float t2;
+  __vector float t3;
+  __vector float t4;
+  __vector float t5;
+  __vector float t6;
+  __vector float t7;
+  __asm__
+    (
+       "xscvdpspn   36, %x[cos]               \n\t" // load c to all words
+       "xxspltw     36, 36, 0                 \n\t" 
+       "xscvdpspn   37, %x[sin]               \n\t" // load s to all words
+       "xxspltw     37, 37, 0                 \n\t" 
+       "lxvd2x      32, 0, %[x_ptr]           \n\t" // load x
+       "lxvd2x      33, %[i16], %[x_ptr]      \n\t" 
+       "lxvd2x      34, %[i32], %[x_ptr]      \n\t" 
+       "lxvd2x      35, %[i48], %[x_ptr]      \n\t" 
+       "lxvd2x      48, 0, %[y_ptr]           \n\t" // load y
+       "lxvd2x      49, %[i16], %[y_ptr]      \n\t" 
+       "lxvd2x      50, %[i32], %[y_ptr]      \n\t" 
+       "lxvd2x      51, %[i48], %[y_ptr]      \n\t" 
+       "addi        %[x_ptr], %[x_ptr], 64    \n\t" 
+       "addi        %[y_ptr], %[y_ptr], 64    \n\t" 
+       "addic.      %[temp_n], %[temp_n], -8  \n\t" 
+       "ble         2f                        \n\t" 
+       ".p2align    5                         \n\t" 
+       "1:                                    \n\t" 
+       "xvmulsp     40, 32, 36                \n\t" // c * x
+       "xvmulsp     41, 33, 36                \n\t" 
+       "xvmulsp     42, 34, 36                \n\t" 
+       "xvmulsp     43, 35, 36                \n\t" 
+       "xvmulsp     %x[x0], 48, 36            \n\t" // c * y
+       "xvmulsp     %x[x2], 49, 36            \n\t" 
+       "xvmulsp     %x[x1], 50, 36            \n\t" 
+       "xvmulsp     %x[x3], 51, 36            \n\t" 
+       "xvmulsp     44, 32, 37                \n\t" // s * x
+       "xvmulsp     45, 33, 37                \n\t" 
+       "lxvd2x      32, 0, %[x_ptr]           \n\t" // load x
+       "lxvd2x      33, %[i16], %[x_ptr]      \n\t" 
+       "xvmulsp     46, 34, 37                \n\t" 
+       "xvmulsp     47, 35, 37                \n\t" 
+       "lxvd2x      34, %[i32], %[x_ptr]      \n\t" 
+       "lxvd2x      35, %[i48], %[x_ptr]      \n\t" 
+       "xvmulsp     %x[x4], 48, 37            \n\t" // s * y
+       "xvmulsp     %x[x5], 49, 37            \n\t" 
+       "lxvd2x      48, 0, %[y_ptr]           \n\t" // load y
+       "lxvd2x      49, %[i16], %[y_ptr]      \n\t" 
+       "xvmulsp     %x[x6], 50, 37            \n\t" 
+       "xvmulsp     %x[x7], 51, 37            \n\t" 
+       "lxvd2x      50, %[i32], %[y_ptr]      \n\t" 
+       "lxvd2x      51, %[i48], %[y_ptr]      \n\t" 
+       "xvaddsp     40, 40, %x[x4]            \n\t" // c * x + s * y
+       "xvaddsp     41, 41, %x[x5]            \n\t" // c * x + s * y
+       "addi        %[x_ptr], %[x_ptr], -64   \n\t" 
+       "addi        %[y_ptr], %[y_ptr], -64   \n\t" 
+       "xvaddsp     42, 42, %x[x6]            \n\t" // c * x + s * y
+       "xvaddsp     43, 43, %x[x7]            \n\t" // c * x + s * y
+       "xvsubsp     %x[x0], %x[x0], 44        \n\t" // c * y - s * x
+       "xvsubsp     %x[x2], %x[x2], 45        \n\t" // c * y - s * x
+       "xvsubsp     %x[x1], %x[x1], 46        \n\t" // c * y - s * x
+       "xvsubsp     %x[x3], %x[x3], 47        \n\t" // c * y - s * x
+       "stxvd2x     40, 0, %[x_ptr]           \n\t" // store x
+       "stxvd2x     41, %[i16], %[x_ptr]      \n\t" 
+       "stxvd2x     42, %[i32], %[x_ptr]      \n\t" 
+       "stxvd2x     43, %[i48], %[x_ptr]      \n\t" 
+       "stxvd2x     %x[x0], 0, %[y_ptr]       \n\t" // store y
+       "stxvd2x     %x[x2], %[i16], %[y_ptr]  \n\t" 
+       "stxvd2x     %x[x1], %[i32], %[y_ptr]  \n\t" 
+       "stxvd2x     %x[x3], %[i48], %[y_ptr]  \n\t" 
+       "addi        %[x_ptr], %[x_ptr], 128   \n\t" 
+       "addi        %[y_ptr], %[y_ptr], 128   \n\t" 
+       "addic.      %[temp_n], %[temp_n], -8  \n\t" 
+       "bgt         1b                        \n\t" 
+       "2:                                    \n\t" 
+       "xvmulsp     40, 32, 36                \n\t" // c * x
+       "xvmulsp     41, 33, 36                \n\t" 
+       "xvmulsp     42, 34, 36                \n\t" 
+       "xvmulsp     43, 35, 36                \n\t" 
+       "xvmulsp     %x[x0], 48, 36            \n\t" // c * y
+       "xvmulsp     %x[x2], 49, 36            \n\t" 
+       "xvmulsp     %x[x1], 50, 36            \n\t" 
+       "xvmulsp     %x[x3], 51, 36            \n\t" 
+       "xvmulsp     44, 32, 37                \n\t" // s * x
+       "xvmulsp     45, 33, 37                \n\t" 
+       "xvmulsp     46, 34, 37                \n\t" 
+       "xvmulsp     47, 35, 37                \n\t" 
+       "xvmulsp     %x[x4], 48, 37            \n\t" // s * y
+       "xvmulsp     %x[x5], 49, 37            \n\t" 
+       "xvmulsp     %x[x6], 50, 37            \n\t" 
+       "xvmulsp     %x[x7], 51, 37            \n\t" 
+       "addi        %[x_ptr], %[x_ptr], -64   \n\t" 
+       "addi        %[y_ptr], %[y_ptr], -64   \n\t" 
+       "xvaddsp     40, 40, %x[x4]            \n\t" // c * x + s * y
+       "xvaddsp     41, 41, %x[x5]            \n\t" // c * x + s * y
+       "xvaddsp     42, 42, %x[x6]            \n\t" // c * x + s * y
+       "xvaddsp     43, 43, %x[x7]            \n\t" // c * x + s * y
+       "xvsubsp     %x[x0], %x[x0], 44        \n\t" // c * y - s * x
+       "xvsubsp     %x[x2], %x[x2], 45        \n\t" // c * y - s * x
+       "xvsubsp     %x[x1], %x[x1], 46        \n\t" // c * y - s * x
+       "xvsubsp     %x[x3], %x[x3], 47        \n\t" // c * y - s * x
+       "stxvd2x     40, 0, %[x_ptr]           \n\t" // store x
+       "stxvd2x     41, %[i16], %[x_ptr]      \n\t" 
+       "stxvd2x     42, %[i32], %[x_ptr]      \n\t" 
+       "stxvd2x     43, %[i48], %[x_ptr]      \n\t" 
+       "stxvd2x     %x[x0], 0, %[y_ptr]       \n\t" // store y
+       "stxvd2x     %x[x2], %[i16], %[y_ptr]  \n\t" 
+       "stxvd2x     %x[x1], %[i32], %[y_ptr]  \n\t" 
+       "stxvd2x     %x[x3], %[i48], %[y_ptr]  "
+     :
+       [mem_x]  "+m"  (*(float (*)[2*n])x),
+       [mem_y]  "+m"  (*(float (*)[2*n])y),
+       [temp_n] "+r"  (n),
+       [x_ptr]  "+&b" (x),
+       [y_ptr]  "+&b" (y),
+       [x0]     "=wa" (t0),
+       [x1]     "=wa" (t2),
+       [x2]     "=wa" (t1),
+       [x3]     "=wa" (t3),
+       [x4]     "=wa" (t4),
+       [x5]     "=wa" (t5),
+       [x6]     "=wa" (t6),
+       [x7]     "=wa" (t7)     
+     : 
+       [cos]    "f"   (c),
+       [sin]    "f"   (s),
+       [i16]    "b"   (16),
+       [i32]    "b"   (32),
+       [i48]    "b"   (48)     
+     :
+       "cr0",
+       "vs32","vs33","vs34","vs35","vs36","vs37",
+       "vs40","vs41","vs42","vs43","vs44","vs45","vs46","vs47",
+       "vs48","vs49","vs50","vs51"
+     );
+}
+ 
+#endif
+
+
+int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT c, FLOAT s)
+{
+   BLASLONG i=0;
+    BLASLONG ix=0,iy=0;
+    FLOAT temp[2];
+    BLASLONG inc_x2;
+    BLASLONG inc_y2;
+
+    if ( n <= 0     )  return(0); 
+
+    if ( (inc_x == 1) && (inc_y == 1) )
+    {
+
+        BLASLONG n1 = n & -8; 
+        if ( n1 > 0 )
+        { 
+            crot_kernel_8(n1, x, y, c, s);
+            i=n1; 
+            ix=2*n1; 
+        }
+
+         while(i < n)
+           {
+                temp[0]   = c*x[ix]   + s*y[ix] ;
+                temp[1]   = c*x[ix+1] + s*y[ix+1] ;
+                y[ix]     = c*y[ix]   - s*x[ix] ;
+                y[ix+1]   = c*y[ix+1] - s*x[ix+1] ;
+                x[ix]     = temp[0] ;
+                x[ix+1]   = temp[1] ;
+
+                ix += 2 ; 
+                i++ ;
+
+            }
+
+    }
+    else
+    {
+        inc_x2 = 2 * inc_x ;
+        inc_y2 = 2 * inc_y ;
+        while(i < n)
+        {
+            temp[0]   = c*x[ix]   + s*y[iy] ;
+            temp[1]   = c*x[ix+1] + s*y[iy+1] ;
+            y[iy]     = c*y[iy]   - s*x[ix] ;
+            y[iy+1]   = c*y[iy+1] - s*x[ix+1] ;
+            x[ix]     = temp[0] ;
+            x[ix+1]   = temp[1] ;
+
+            ix += inc_x2 ;
+            iy += inc_y2 ;
+            i++ ;
+
+        }
+    }
+	return(0);
+}
+

kernel/power/cswap.c

@@ -36,7 +36,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "common.h"
 
 
-#if defined(POWER8)
+#if defined(POWER8)  || defined(POWER9)
 #include "cswap_microk_power8.c"
 #endif
 

kernel/power/dasum.c

@@ -46,7 +46,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #endif
 
-#if defined(POWER8)
+#if defined(POWER8) || defined(POWER9)
 #include "dasum_microk_power8.c"
 #endif
 

kernel/power/daxpy.c

@@ -36,7 +36,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "common.h"
 
 
-#if defined(POWER8)
+#if defined(POWER8) || defined(POWER9)
 #include "daxpy_microk_power8.c"
 #endif
 

kernel/power/dcopy.c

@@ -35,7 +35,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
-#if defined(POWER8)
+#if defined(POWER8) || defined(POWER9)
 #include "dcopy_microk_power8.c"
 #endif
 

kernel/power/ddot.c

@@ -36,7 +36,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "common.h"
 
 
-#if defined(POWER8) 
+#if defined(POWER8)  || defined(POWER9)
 #include "ddot_microk_power8.c"
 #endif
 

kernel/power/dgemm_kernel_power9.S

@@ -0,0 +1,249 @@
+/***************************************************************************
+Copyright (c) 2013-2019, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+ 
+#define ASSEMBLER
+#include "common.h"
+#include "def_vsx.h"
+
+ 
+#define LOAD	ld
+ 
+ 
+
+ 
+#define STACKSIZE  (512 )
+#define ALPHA_SP   (296+192)(SP)
+#define FZERO	(304+192)(SP)
+ 
+
+ 
+#define	M	r3
+#define	N	r4
+#define	K	r5
+ 
+#define A	r7
+#define	B	r8
+#define	C	r9
+#define	LDC	r10
+#define OFFSET	r6
+ 
+ 
+
+#define alpha_r vs18
+
+#define o0	0
+
+
+#define T4	r12
+#define T3	r11
+#define C4	r14
+#define o8	r15
+#define o24	r16
+#define C2	r17
+#define L	r18
+#define T1	r19
+#define C3	r20
+#define TEMP_REG	r21
+#define	I	r22
+#define J	r23
+#define AO	r24
+#define	BO	r25
+#define	CO 	r26
+#define o16	r27
+#define	o32	r28
+#define	o48	r29
+
+#define PRE	r30
+#define T2	r31
+
+#include "dgemm_macros_power9.S"
+
+
+#ifndef NEEDPARAM
+
+	PROLOGUE
+	PROFCODE
+
+	addi	SP, SP, -STACKSIZE
+	li	r0, 0
+
+	stfd	f14,    0(SP)
+	stfd	f15,    8(SP)
+	stfd	f16,   16(SP)
+	stfd	f17,   24(SP)
+
+	stfd	f18,   32(SP)
+	stfd	f19,   40(SP)
+	stfd	f20,   48(SP)
+	stfd	f21,   56(SP)
+
+	stfd	f22,   64(SP)
+	stfd	f23,   72(SP)
+	stfd	f24,   80(SP)
+	stfd	f25,   88(SP)
+
+	stfd	f26,   96(SP)
+	stfd	f27,  104(SP)
+	stfd	f28,  112(SP)
+	stfd	f29,  120(SP)
+
+	stfd	f30,  128(SP)
+	stfd	f31,  136(SP)
+
+ 
+	std	r31,  144(SP)
+	std	r30,  152(SP)
+	std	r29,  160(SP)
+	std	r28,  168(SP)
+	std	r27,  176(SP)
+	std	r26,  184(SP)
+	std	r25,  192(SP)
+	std	r24,  200(SP)
+	std	r23,  208(SP)
+	std	r22,  216(SP)
+	std	r21,  224(SP)
+	std	r20,  232(SP)
+	std	r19,  240(SP)
+	std	r18,  248(SP)
+	std	r17,  256(SP)
+	std	r16,  264(SP)
+	std	r15,  272(SP)
+	std	r14,  280(SP)
+ 
+ 
+    stxv    v20,  288(SP)
+    stxv    v21,  304(SP)
+    stxv    v22,  320(SP)
+    stxv    v23,  336(SP)
+    stxv    v24,  352(SP)
+    stxv    v25,  368(SP)
+    stxv    v26,  384(SP)
+    stxv    v27,  400(SP)
+    stxv    v28,  416(SP)
+    stxv    v29,  432(SP)
+    stxv    v30,  448(SP)
+    stxv    v31,  464(SP)
+
+
+	stfd	f1,  ALPHA_SP
+	stw	r0,  FZERO 
+
+	slwi	LDC, LDC, BASE_SHIFT
+
+#if defined(TRMMKERNEL)
+	ld	OFFSET,  FRAMESLOT(0) + STACKSIZE(SP)
+#endif
+
+
+	cmpwi	cr0, M, 0
+	ble	.L999_H1
+	cmpwi	cr0, N, 0
+	ble	.L999_H1
+	cmpwi	cr0, K, 0
+	ble	.L999_H1
+
+ 
+ 
+   	addi	T1, SP, 296+192
+ 
+
+	li	PRE, 384
+	li	o8 , 8
+	li	o16, 16
+	li	o24, 24
+	li	o32, 32
+	li	o48, 48
+
+
+	lxvdsx	alpha_r, 0, T1
+
+#include "dgemm_logic_power9.S"
+
+.L999:
+	addi	r3, 0, 0
+
+	lfd	f14,    0(SP)
+	lfd	f15,    8(SP)
+	lfd	f16,   16(SP)
+	lfd	f17,   24(SP)
+
+	lfd	f18,   32(SP)
+	lfd	f19,   40(SP)
+	lfd	f20,   48(SP)
+	lfd	f21,   56(SP)
+
+	lfd	f22,   64(SP)
+	lfd	f23,   72(SP)
+	lfd	f24,   80(SP)
+	lfd	f25,   88(SP)
+
+	lfd	f26,   96(SP)
+	lfd	f27,  104(SP)
+	lfd	f28,  112(SP)
+	lfd	f29,  120(SP)
+
+	lfd	f30,  128(SP)
+	lfd	f31,  136(SP)
+
+ 
+	ld	r31,  144(SP)
+	ld	r30,  152(SP)
+	ld	r29,  160(SP)
+	ld	r28,  168(SP)
+	ld	r27,  176(SP)
+	ld	r26,  184(SP)
+	ld	r25,  192(SP)
+	ld	r24,  200(SP)
+	ld	r23,  208(SP)
+	ld	r22,  216(SP)
+	ld	r21,  224(SP)
+	ld	r20,  232(SP)
+	ld	r19,  240(SP)
+	ld	r18,  248(SP)
+	ld	r17,  256(SP)
+	ld	r16,  264(SP)
+	ld	r15,  272(SP)
+	ld	r14,  280(SP)
+ 
+    lxv    v20,  288(SP)
+    lxv    v21,  304(SP)
+    lxv    v22,  320(SP)
+    lxv    v23,  336(SP)
+    lxv    v24,  352(SP)
+    lxv    v25,  368(SP)
+    lxv    v26,  384(SP)
+    lxv    v27,  400(SP)
+    lxv    v28,  416(SP)
+    lxv    v29,  432(SP)
+    lxv    v30,  448(SP)
+    lxv    v31,  464(SP)
+
+	addi	SP, SP, STACKSIZE
+	blr
+
+	EPILOGUE
+#endif

kernel/power/dgemv_n.c

@@ -38,7 +38,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "common.h"
 
 
-#if defined(POWER8)
+#if defined(POWER8) || defined(POWER9)
 #include "dgemv_n_microk_power8.c"
 #endif