Merge pull request #1 from xianyi/develop

update

.github/workflows/dynamic_arch.yml

@@ -0,0 +1,81 @@
+name: continuous build
+
+on: [push, pull_request]
+
+jobs:
+  build:
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest]
+        build: [cmake, make]
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v2
+
+      - name: Compilation cache
+        uses: actions/cache@v2
+        with:
+          path: ~/.ccache
+          # We include the commit sha in the cache key, as new cache entries are
+          # only created if there is no existing entry for the key yet.
+          key: ${{ runner.os }}-ccache-${{ github.sha }}
+          # Restore any ccache cache entry, if none for
+          # ${{ runner.os }}-ccache-${{ github.sha }} exists
+          restore-keys: |
+            ${{ runner.os }}-ccache
+
+      - name: Print system information
+        run: |
+          if [ "$RUNNER_OS" == "Linux" ]; then
+            cat /proc/cpuinfo
+          elif [ "$RUNNER_OS" == "macOS" ]; then
+            sysctl -a | grep machdep.cpu
+          else
+            echo "$RUNNER_OS not supported"
+            exit 1
+          fi
+
+      - name: Install Dependencies
+        run: |
+          if [ "$RUNNER_OS" == "Linux" ]; then
+            sudo apt-get install -y gfortran cmake ccache
+          elif [ "$RUNNER_OS" == "macOS" ]; then
+            brew install coreutils cmake ccache
+          else
+            echo "$RUNNER_OS not supported"
+            exit 1
+          fi
+          ccache -M 300M  # Limit the ccache size; Github's overall cache limit is 5GB
+
+      - name: Build
+        if: matrix.build == 'make'
+        run: |
+          if [ "$RUNNER_OS" == "Linux" ]; then
+            export PATH="/usr/lib/ccache:${PATH}"
+          elif [ "$RUNNER_OS" == "macOS" ]; then
+            export PATH="$(brew --prefix)/opt/ccache/libexec:${PATH}"
+          else
+            echo "$RUNNER_OS not supported"
+            exit 1
+          fi
+
+          make -j$(nproc) DYNAMIC_ARCH=1 USE_OPENMP=0
+
+      - name: CMake build
+        if: matrix.build == 'cmake'
+        run: |
+          if [ "$RUNNER_OS" == "Linux" ]; then
+            export PATH="/usr/lib/ccache:${PATH}"
+          elif [ "$RUNNER_OS" == "macOS" ]; then
+            export PATH="$(brew --prefix)/opt/ccache/libexec:${PATH}"
+          else
+            echo "$RUNNER_OS not supported"
+            exit 1
+          fi
+
+          mkdir build
+          cd build
+          cmake -DDYNAMIC_ARCH=1 -DNOFORTRAN=0 -DBUILD_WITHOUT_LAPACK=0  -DCMAKE_VERBOSE_MAKEFILE=ON -DCMAKE_BUILD_TYPE=Release ..
+          make -j$(nproc)

.github/workflows/nightly-Homebrew-build.yml

@@ -21,6 +21,7 @@ jobs:
   build-OpenBLAS-with-Homebrew:
     runs-on: macos-latest
     env:
+      DEVELOPER_DIR: /Applications/Xcode_11.4.1.app/Contents/Developer
       HOMEBREW_DEVELOPER: "ON"
       HOMEBREW_DISPLAY_INSTALL_TIMES: "ON"
       HOMEBREW_NO_ANALYTICS: "ON"

.gitignore

@@ -70,6 +70,7 @@ test/SBLAT2.SUMM
 test/SBLAT3.SUMM
 test/ZBLAT2.SUMM
 test/ZBLAT3.SUMM
+test/SHBLAT3.SUMM
 test/cblat1
 test/cblat2
 test/cblat3
@@ -79,6 +80,7 @@ test/dblat3
 test/sblat1
 test/sblat2
 test/sblat3
+test/test_shgemm
 test/zblat1
 test/zblat2
 test/zblat3

.travis.yml

@@ -16,7 +16,6 @@ matrix:
       before_script: &common-before
         - COMMON_FLAGS="DYNAMIC_ARCH=1 TARGET=NEHALEM NUM_THREADS=32"
       script:
-        - set -e
         - make QUIET_MAKE=1 $COMMON_FLAGS $BTYPE
         - make -C test $COMMON_FLAGS $BTYPE
         - make -C ctest $COMMON_FLAGS $BTYPE
@@ -108,7 +107,6 @@ matrix:
         - sudo sh alpine-chroot-install -p 'build-base gfortran perl linux-headers'
       before_script: *common-before
       script:
-        - set -e
         # XXX: Disable some warnings for now to avoid exceeding Travis limit for log size.
         - alpine make QUIET_MAKE=1 $COMMON_FLAGS $BTYPE
               CFLAGS="-Wno-misleading-indentation -Wno-sign-conversion -Wno-incompatible-pointer-types"
@@ -151,7 +149,6 @@ matrix:
       before_script:
         - COMMON_ARGS="-DTARGET=NEHALEM -DNUM_THREADS=32"
       script:
-        - set -e
         - mkdir build
         - CONFIG=Release
         - cmake -Bbuild -H. $CMAKE_ARGS $COMMON_ARGS -DCMAKE_BUILD_TYPE=$CONFIG

CMakeLists.txt

@@ -6,7 +6,8 @@ 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 9.dev)
+set(OpenBLAS_PATCH_VERSION 10.dev)
+
 set(OpenBLAS_VERSION "${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.${OpenBLAS_PATCH_VERSION}")
 
 # Adhere to GNU filesystem layout conventions
@@ -23,6 +24,7 @@ option(BUILD_WITHOUT_CBLAS "Do not build the C interface (CBLAS) to the BLAS fun
 option(DYNAMIC_ARCH "Include support for multiple CPU targets, with automatic selection at runtime (x86/x86_64, aarch64 or ppc only)" OFF)
 option(DYNAMIC_OLDER "Include specific support for older x86 cpu models (Penryn,Dunnington,Atom,Nano,Opteron) with DYNAMIC_ARCH" OFF)
 option(BUILD_RELAPACK "Build with ReLAPACK (recursive implementation of several LAPACK functions on top of standard LAPACK)" OFF)
+option(USE_LOCKING "Use locks even in single-threaded builds to make them callable from multiple threads" OFF)
 if(${CMAKE_SYSTEM_NAME} MATCHES "Linux")
 option(NO_AFFINITY "Disable support for CPU affinity masks to avoid binding processes from e.g. R or numpy/scipy to a single core" ON)
 else()
@@ -86,9 +88,13 @@ if (NOT NO_LAPACK)
   list(APPEND SUBDIRS lapack)
 endif ()
 
+if (NOT DEFINED BUILD_HALF)
+ set (BUILD_HALF false)
+endif ()
 # set which float types we want to build for
 if (NOT DEFINED BUILD_SINGLE AND NOT DEFINED BUILD_DOUBLE AND NOT DEFINED BUILD_COMPLEX AND NOT DEFINED BUILD_COMPLEX16)
   # if none are defined, build for all
+#  set(BUILD_HALF true)
   set(BUILD_SINGLE true)
   set(BUILD_DOUBLE true)
   set(BUILD_COMPLEX true)
@@ -120,6 +126,11 @@ if (BUILD_COMPLEX16)
   list(APPEND FLOAT_TYPES "ZCOMPLEX") # defines COMPLEX and DOUBLE
 endif ()
 
+if (BUILD_HALF)
+  message(STATUS "Building Half Precision")
+  list(APPEND FLOAT_TYPES "HALF") # defines nothing
+endif ()
+
 if (NOT DEFINED CORE OR "${CORE}" STREQUAL "UNKNOWN")
   message(FATAL_ERROR "Detecting CPU failed. Please set TARGET explicitly, e.g. make TARGET=your_cpu_target. Please read README for details.")
 endif ()
@@ -234,7 +245,7 @@ if (BUILD_SHARED_LIBS AND BUILD_RELAPACK)
   if (NOT MSVC)
     target_link_libraries(${OpenBLAS_LIBNAME} "-Wl,-allow-multiple-definition")
   else()
-    target_link_libraries(${OpenBLAS_LIBNAME} "/FORCE:MULTIPLE")
+   set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} /FORCE:MULTIPLE")
   endif()
 endif()
 

CONTRIBUTORS.md

@@ -180,3 +180,13 @@ In chronological order:
   * [2019-12-23] optimize AVX2 CGEMM and ZGEMM
   * [2019-12-30] AVX2 CGEMM3M & ZGEMM3M kernels
   * [2020-01-07] optimize AVX2 SGEMM and STRMM
+
+* Rajalakshmi Srinivasaraghavan <https://github.com/RajalakshmiSR>
+  * [2020-04-15] Half-precision GEMM for bfloat16
+
+* Marius Hillenbrand <https://github.com/mhillenibm>
+  * [2020-05-12] Revise dynamic architecture detection for IBM z
+  * [2020-05-12] Add new sgemm and strmm kernel for IBM z14
+
+* Danfeng Zhang <https://github.com/craft-zhang>
+  * [2020-05-20] Improve performance of SGEMM and STRMM on Arm Cortex-A53

Changelog.txt

@@ -1,4 +1,77 @@
 OpenBLAS ChangeLog
+====================================================================
+Version 0.3.10
+ 14-Jun-2020
+
+common:
+	* Improved thread locking behaviour in blas_server and parallel getrf                         
+	* Imported bugfix 394 from LAPACK (spurious reference to "XERBL"
+	  due to overlong lines)
+	* Imported bugfix 403 from LAPACK (compile option "recursive" required
+	  for correctness with Intel and PGI)
+	* Imported bugfix 408 from LAPACK (wrong scaling in ZHEEQUB)
+	* Imported bugfix 411 from LAPACK (infinite loop in LARGV/LARTG/LARTGP)
+	* Fixed mismatches between BUFFERSIZE and GEMM_UNROLL parameters that
+	  could lead to crashes at large matrix sizes  
+	* Restored internal soname in dynamic libraries on FreeBSD and Dragonfly 
+	* Added API (openblas_setaffinity) to set the thread affinity on Linux
+	* Added initial infrastructure for half-precision floating point 
+	  (bfloat16) support with a generic implementation of SHGEMM     
+	* Added CMAKE build system support for building the cblas_Xgemm3m
+	  functions
+	* Fixed CMAKE support for building in a path with embedded spaces
+	* Fixed CMAKE (non)handling of NO_EXPRECISION and MAX_STACK_ALLOC
+	* Fixed GCC version detection in the Makefiles
+	* Allowed overriding the names of AR, AS and LD in Makefile builds
+
+POWER:
+	* Fixed big-endian POWER8 ELFv2 builds on FreeBSD
+	* Fixed GCC version checks and DYNAMIC_ARCH builds on POWER9
+	* Fixed CMAKE build support for POWER9
+	* fixed a potential race condition in the thread buffer allocation
+	* Worked around LAPACK test failures on PPC G4
+
+MIPS:
+	* Fixed a potential race condition in the thread buffer allocation
+	* Added support for MIPS 24K/24KE family based on P5600 kernels
+
+MIPS64:
+	* fixed a potential race condition in the thread buffer allocation
+	* Added TARGET=GENERIC
+
+ARMV7:
+	* Fixed a race condition in the thread buffer allocation
+
+ARMV8:
+	* Fixed a race condition in the thread buffer allocation
+	* Fixed zero initialisation in the assembly for SGEMM and DGEMM BETA
+	* Improved performance of the ThunderX2 DAXPY kernel
+	* Added an optimized SGEMM kernel for Cortex A53    
+	* Fixed Makefile support for INTERFACE64 (8-byte integer)
+
+x86_64:
+	* Fixed a syntax error in the CMAKE setup for SkylakeX
+	* Improved performance of STRSM on Haswell, SkylakeX and Ryzen
+	* Improved SGEMM performance on SGEMM for workloads with ldc a
+	  multiple of 1024
+	* Improved DGEMM performance on Skylake X
+	* Fixed unwanted AVX512-dependency of SGEMM in DYNAMIC_ARCH
+	  builds created on SkylakeX
+	* Removed data alignment requirement in the SSE2 copy kernels
+	  that could cause spurious crashes
+	* Added a workaround for an optimizer bug in AppleClang 11.0.3
+	* Fixed LAPACK test failures due to wrong options for Intel Fortran
+	* Fixed compilation and LAPACK test results with recent Flang
+	  and AMD AOCC
+	* Fixed DYNAMIC_ARCH builds with CMAKE on OS X
+	* Fixed missing exports of cblas_i?amin, cblas_i?min, cblas_i?max,
+	  cblas_?sum, cblas_?gemm3m in the shared library on OS 
+	* Fixed reporting of cpu name in DYNAMIC_ARCH builds (would sometimes
+	  show the name of an older generation chip supported by the same kernels)
+
+IBM Z:
+	* Improved performance of SGEMM/STRMM and DGEMM/DTRMM on Z14
+
 ====================================================================
 Version 0.3.9
  1-Mar-2020

Makefile

@@ -264,6 +264,7 @@ lapack_prebuild :
 ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
 	-@echo "FC          = $(FC)" > $(NETLIB_LAPACK_DIR)/make.inc
 	-@echo "FFLAGS      = $(LAPACK_FFLAGS)" >> $(NETLIB_LAPACK_DIR)/make.inc
+	-@echo "FFLAGS_DRV  = $(LAPACK_FFLAGS)" >> $(NETLIB_LAPACK_DIR)/make.inc
 	-@echo "POPTS       = $(LAPACK_FPFLAGS)" >> $(NETLIB_LAPACK_DIR)/make.inc
 	-@echo "FFLAGS_NOOPT       = -O0 $(LAPACK_NOOPT)" >> $(NETLIB_LAPACK_DIR)/make.inc
 	-@echo "PNOOPT      = $(LAPACK_FPFLAGS) -O0" >> $(NETLIB_LAPACK_DIR)/make.inc

Makefile.power

@@ -9,6 +9,16 @@ else
 USE_OPENMP = 1
 endif
 
+ifeq ($(CORE), POWER10)
+ifeq ($(USE_OPENMP), 1)
+COMMON_OPT += -Ofast -mcpu=future -mtune=future -mvsx -malign-power -DUSE_OPENMP -fno-fast-math -fopenmp
+FCOMMON_OPT += -O2 -frecursive -mcpu=future -mtune=future -malign-power -DUSE_OPENMP -fno-fast-math -fopenmp
+else
+COMMON_OPT += -Ofast -mcpu=future -mtune=future -mvsx -malign-power -fno-fast-math
+FCOMMON_OPT += -O2 -frecursive -mcpu=future -mtune=future -malign-power -fno-fast-math
+endif
+endif
+
 ifeq ($(CORE), POWER9)
 ifeq ($(USE_OPENMP), 1)
 COMMON_OPT += -Ofast -mcpu=power9 -mtune=power9 -mvsx -malign-power -DUSE_OPENMP -fno-fast-math -fopenmp

Makefile.prebuild

@@ -17,7 +17,11 @@ ifdef CPUIDEMU
 EXFLAGS = -DCPUIDEMU -DVENDOR=99
 endif
 
-ifeq ($(TARGET), 1004K)
+ifeq ($(TARGET), MIPS24K)
+TARGET_FLAGS = -mips32r2
+endif
+
+ifeq ($(TARGET), MIPS1004K)
 TARGET_FLAGS = -mips32r2
 endif
 

Makefile.rule

@@ -3,7 +3,7 @@
 #
 
 # This library's version
-VERSION = 0.3.9.dev
+VERSION = 0.3.10.dev
 
 # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a
 # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library
@@ -273,6 +273,9 @@ COMMON_PROF = -pg
 #
 # CPP_THREAD_SAFETY_TEST = 1
 
+
+# If you want to enable the experimental BFLOAT16 support
+# BUILD_HALF = 1
 #
 #  End of user configuration
 #

Makefile.system

@@ -21,6 +21,8 @@ ifeq ($(ARCH), amd64)
 override ARCH=x86_64
 else ifeq ($(ARCH), powerpc64)
 override ARCH=power
+else ifeq ($(ARCH), powerpc)
+override ARCH=power
 else ifeq ($(ARCH), i386)
 override ARCH=x86
 else ifeq ($(ARCH), aarch64)
@@ -261,10 +263,10 @@ endif
 
 ARFLAGS	=
 CPP	= $(COMPILER) -E
-AR	= $(CROSS_SUFFIX)ar
+AR	?= $(CROSS_SUFFIX)ar
-AS	= $(CROSS_SUFFIX)as
+AS	?= $(CROSS_SUFFIX)as
-LD	= $(CROSS_SUFFIX)ld
+LD	?= $(CROSS_SUFFIX)ld
-RANLIB	= $(CROSS_SUFFIX)ranlib
+RANLIB	?= $(CROSS_SUFFIX)ranlib
 NM	= $(CROSS_SUFFIX)nm
 DLLWRAP = $(CROSS_SUFFIX)dllwrap
 OBJCOPY = $(CROSS_SUFFIX)objcopy
@@ -277,6 +279,17 @@ NO_LAPACK = 1
 override FEXTRALIB = 
 endif
 
+ifeq ($(C_COMPILER), GCC)
+GCCVERSIONGTEQ4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 4)
+GCCVERSIONGT4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 4)
+GCCVERSIONEQ5 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` = 5)
+GCCVERSIONGT5 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 5)
+GCCVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 7)
+GCCVERSIONGTEQ9 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 9)
+GCCMINORVERSIONGTEQ2 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 2)
+GCCMINORVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 7)
+endif
+
 #
 #  OS dependent settings
 #
@@ -323,13 +336,7 @@ ifeq ($(C_COMPILER), CLANG)
 CCOMMON_OPT	+= -DMS_ABI
 endif
 
-ifeq ($(C_COMPILER), GCC)
 #Version tests for supporting specific features (MS_ABI, POWER9 intrinsics)
-GCCVERSIONGTEQ4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 4)
-GCCVERSIONGT4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 4)
-GCCVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 7)
-GCCVERSIONGTEQ9 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 9)
-GCCMINORVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 7)
 ifeq ($(GCCVERSIONGT4), 1)
 # GCC Major version > 4
 # It is compatible with MSVC ABI.
@@ -343,7 +350,6 @@ ifeq ($(GCCMINORVERSIONGTEQ7), 1)
 CCOMMON_OPT	+= -DMS_ABI
 endif
 endif
-endif
 
 # Ensure the correct stack alignment on Win32
 # http://permalink.gmane.org/gmane.comp.lib.openblas.general/97
@@ -563,8 +569,34 @@ DYNAMIC_CORE += EMAG8180
 endif
 
 ifeq ($(ARCH), zarch)
-DYNAMIC_CORE = Z13
+DYNAMIC_CORE = ZARCH_GENERIC
+
+# Z13 is supported since gcc-5.2, gcc-6, and in RHEL 7.3 and newer
+ifeq ($(GCCVERSIONGT5), 1)
+	ZARCH_SUPPORT_Z13 := 1
+else ifeq ($(GCCVERSIONEQ5), 1)
+ifeq ($(GCCMINORVERSIONGTEQ2), 1)
+	ZARCH_SUPPORT_Z13 := 1
+endif
+endif
+
+ifeq ($(wildcard /etc/redhat-release), /etc/redhat-release)
+ifeq ($(shell source /etc/os-release ; expr $$VERSION_ID \>= "7.3"), 1)
+	ZARCH_SUPPORT_Z13 := 1
+endif
+endif
+
+ifeq ($(ZARCH_SUPPORT_Z13), 1)
+DYNAMIC_CORE += Z13
+else
+$(info OpenBLAS: Not building Z13 kernels because gcc is older than 5.2 or 6.x)
+endif
+
+ifeq ($(GCCVERSIONGTEQ7), 1)
 DYNAMIC_CORE += Z14
+else
+$(info OpenBLAS: Not building Z14 kernels because gcc is older than 7.x)
+endif
 endif
 
 ifeq ($(ARCH), power)
@@ -572,14 +604,20 @@ DYNAMIC_CORE = POWER6
 DYNAMIC_CORE += POWER8
 ifneq ($(C_COMPILER), GCC)
 DYNAMIC_CORE += POWER9
+DYNAMIC_CORE += POWER10
 endif
 ifeq ($(C_COMPILER), GCC)
-GCCVERSIONGT5 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 5)
 ifeq ($(GCCVERSIONGT5), 1)
 DYNAMIC_CORE += POWER9
 else
 $(info, OpenBLAS: Your gcc version is too old to build the POWER9 kernels.)
 endif
+GCCVERSIONGTEQ11 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 11)
+ifeq ($(GCCVERSIONGTEQ11), 1)
+DYNAMIC_CORE += POWER10
+else
+$(info, OpenBLAS: Your gcc version is too old to build the POWER10 kernels.)
+endif
 endif
 endif
 
@@ -690,7 +728,12 @@ CCOMMON_OPT += -march=mips64
 FCOMMON_OPT += -march=mips64
 endif
 
-ifeq ($(CORE), 1004K)
+ifeq ($(CORE), MIPS24K)
+CCOMMON_OPT += -mips32r2 -mtune=24kc  $(MSA_FLAGS)
+FCOMMON_OPT += -mips32r2 -mtune=24kc  $(MSA_FLAGS)
+endif
+
+ifeq ($(CORE), MIPS1004K)
 CCOMMON_OPT += -mips32r2   $(MSA_FLAGS)
 FCOMMON_OPT += -mips32r2   $(MSA_FLAGS)
 endif
@@ -755,6 +798,15 @@ endif
 
 ifeq ($(F_COMPILER), FLANG)
 CCOMMON_OPT += -DF_INTERFACE_FLANG
+FCOMMON_OPT += -Mrecursive -Kieee
+ifeq ($(OSNAME), Linux)
+ifeq ($(ARCH), x86_64)
+FLANG_VENDOR := $(shell expr `$(FC) --version|cut -f 1 -d "."|head -1`)
+ifeq ($(FLANG_VENDOR),AOCC)
+FCOMMON_OPT += -fno-unroll-loops
+endif
+endif
+endif
 ifdef BINARY64
 ifdef INTERFACE64
 ifneq ($(INTERFACE64), 0)
@@ -850,7 +902,7 @@ ifneq ($(INTERFACE64), 0)
 FCOMMON_OPT += -i8
 endif
 endif
-FCOMMON_OPT += -recursive
+FCOMMON_OPT += -recursive -fp-model strict -assume protect-parens
 ifeq ($(USE_OPENMP), 1)
 FCOMMON_OPT += -fopenmp
 endif
@@ -1119,6 +1171,10 @@ ifeq ($(USE_TLS), 1)
 CCOMMON_OPT += -DUSE_TLS
 endif
 
+ifeq ($(BUILD_HALF), 1)
+CCOMMON_OPT += -DBUILD_HALF
+endif
+
 CCOMMON_OPT += -DVERSION=\"$(VERSION)\"
 
 ifndef SYMBOLPREFIX
@@ -1145,6 +1201,7 @@ KERNELDIR	= $(TOPDIR)/kernel/$(ARCH)
 
 include $(TOPDIR)/Makefile.$(ARCH)
 
+CCOMMON_OPT     += -UASMNAME -UASMFNAME -UNAME -UCNAME -UCHAR_NAME -UCHAR_CNAME
 CCOMMON_OPT	+= -DASMNAME=$(FU)$(*F) -DASMFNAME=$(FU)$(*F)$(BU) -DNAME=$(*F)$(BU) -DCNAME=$(*F) -DCHAR_NAME=\"$(*F)$(BU)\" -DCHAR_CNAME=\"$(*F)\"
 
 ifeq ($(CORE), PPC440)
@@ -1237,7 +1294,6 @@ endif
 
 override CFLAGS     += $(COMMON_OPT) $(CCOMMON_OPT) -I$(TOPDIR)
 override PFLAGS     += $(COMMON_OPT) $(CCOMMON_OPT) -I$(TOPDIR) -DPROFILE $(COMMON_PROF)
-
 override FFLAGS     += $(COMMON_OPT) $(FCOMMON_OPT)
 override FPFLAGS    += $(FCOMMON_OPT) $(COMMON_PROF)
 #MAKEOVERRIDES =
@@ -1344,6 +1400,7 @@ export ARCH
 export CORE
 export LIBCORE
 export __BYTE_ORDER__
+export ELF_VERSION
 export PGCPATH
 export CONFIG
 export CC
@@ -1389,7 +1446,10 @@ export KERNELDIR
 export FUNCTION_PROFILE
 export TARGET_CORE
 export NO_AVX512
+export BUILD_HALF
 
+export SHGEMM_UNROLL_M
+export SHGEMM_UNROLL_N
 export SGEMM_UNROLL_M
 export SGEMM_UNROLL_N
 export DGEMM_UNROLL_M

Makefile.tail

@@ -1,3 +1,4 @@
+SHBLASOBJS_P = $(SHBLASOBJS:.$(SUFFIX)=.$(PSUFFIX))
 SBLASOBJS_P = $(SBLASOBJS:.$(SUFFIX)=.$(PSUFFIX))
 DBLASOBJS_P = $(DBLASOBJS:.$(SUFFIX)=.$(PSUFFIX))
 QBLASOBJS_P = $(QBLASOBJS:.$(SUFFIX)=.$(PSUFFIX))
@@ -9,8 +10,8 @@ COMMONOBJS_P = $(COMMONOBJS:.$(SUFFIX)=.$(PSUFFIX))
 
 HPLOBJS_P   = $(HPLOBJS:.$(SUFFIX)=.$(PSUFFIX))
 
-BLASOBJS    = $(SBLASOBJS)   $(DBLASOBJS)   $(CBLASOBJS)   $(ZBLASOBJS)
+BLASOBJS    = $(SHBLASOBJS)  $(SBLASOBJS)   $(DBLASOBJS)   $(CBLASOBJS)   $(ZBLASOBJS)
-BLASOBJS_P  = $(SBLASOBJS_P) $(DBLASOBJS_P) $(CBLASOBJS_P) $(ZBLASOBJS_P)
+BLASOBJS_P  = $(SHBLASOBJS_P) $(SBLASOBJS_P) $(DBLASOBJS_P) $(CBLASOBJS_P) $(ZBLASOBJS_P)
 
 ifdef EXPRECISION
 BLASOBJS   += $(QBLASOBJS)   $(XBLASOBJS)
@@ -22,6 +23,7 @@ BLASOBJS   += $(QBLASOBJS)   $(XBLASOBJS)
 BLASOBJS_P += $(QBLASOBJS_P) $(XBLASOBJS_P)
 endif
 
+$(SHBLASOBJS) $(SHBLASOBJS_P) : override CFLAGS += -DHALF -UDOUBLE  -UCOMPLEX
 $(SBLASOBJS) $(SBLASOBJS_P) : override CFLAGS += -UDOUBLE  -UCOMPLEX
 $(DBLASOBJS) $(DBLASOBJS_P) : override CFLAGS += -DDOUBLE  -UCOMPLEX
 $(QBLASOBJS) $(QBLASOBJS_P) : override CFLAGS += -DXDOUBLE -UCOMPLEX
@@ -29,6 +31,7 @@ $(CBLASOBJS) $(CBLASOBJS_P) : override CFLAGS += -UDOUBLE  -DCOMPLEX
 $(ZBLASOBJS) $(ZBLASOBJS_P) : override CFLAGS += -DDOUBLE  -DCOMPLEX
 $(XBLASOBJS) $(XBLASOBJS_P) : override CFLAGS += -DXDOUBLE -DCOMPLEX
 
+$(SHBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)
 $(SBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)
 $(DBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)
 $(QBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)

Makefile.zarch

@@ -5,6 +5,6 @@ FCOMMON_OPT += -march=z13 -mzvector
 endif
 
 ifeq ($(CORE), Z14)
-CCOMMON_OPT += -march=z14 -mzvector
+CCOMMON_OPT += -march=z14 -mzvector -O3
 FCOMMON_OPT += -march=z14 -mzvector
 endif

README.md

@@ -122,6 +122,11 @@ Please read `GotoBLAS_01Readme.txt` for older CPU models already supported by th
 - **AMD STEAMROLLER**: Uses Bulldozer codes with some optimizations.
 - **AMD ZEN**: Uses Haswell codes with some optimizations.
 
+#### MIPS32
+
+- **MIPS 1004K**: uses P5600 codes
+- **MIPS 24K**: uses P5600 codes
+
 #### MIPS64
 
 - **ICT Loongson 3A**: Optimized Level-3 BLAS and the part of Level-1,2.

TargetList.txt

@@ -49,6 +49,7 @@ POWER6
 POWER7
 POWER8
 POWER9
+POWER10
 PPCG4
 PPC970
 PPC970MP
@@ -58,7 +59,8 @@ CELL
 
 3.MIPS CPU:
 P5600
-1004K
+MIPS1004K
+MIPS24K
 
 4.MIPS64 CPU:
 SICORTEX

azure-pipelines.yml

@@ -49,3 +49,23 @@ jobs:
       # we need a privileged docker run for sde process attachment
       docker run --privileged intel_sde
     displayName: 'Run AVX512 SkylakeX docker build / test'
+
+- job: Windows_cl
+  pool:
+     vmImage: 'windows-latest'
+  steps:   
+  - task: CMake@1
+    inputs:
+      workingDirectory: 'build' # Optional
+      cmakeArgs: '-G "Visual Studio 16 2019" ..'
+  - task: CMake@1
+    inputs:
+      cmakeArgs: '--build . --config Release'
+      workingDirectory: 'build'
+  - script: |
+      cd build
+      cd utest
+      dir
+      openblas_utest.exe
+  
+      

benchmark/Makefile

@@ -49,6 +49,12 @@ else
 GOTO_LAPACK_TARGETS=
 endif
 
+ifeq ($(BUILD_HALF),1)
+GOTO_HALF_TARGETS=shgemm.goto
+else
+GOTO_HALF_TARGETS=
+endif
+
 ifeq ($(OSNAME), WINNT)
 
 goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
@@ -91,7 +97,7 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
        sgetri.goto dgetri.goto cgetri.goto zgetri.goto \
        spotrf.goto dpotrf.goto cpotrf.goto zpotrf.goto \
        ssymm.goto dsymm.goto csymm.goto zsymm.goto \
-       saxpby.goto daxpby.goto caxpby.goto zaxpby.goto
+       saxpby.goto daxpby.goto caxpby.goto zaxpby.goto $(GOTO_HALF_TARGETS)
 
 acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
        scholesky.acml dcholesky.acml ccholesky.acml zcholesky.acml \
@@ -264,7 +270,7 @@ goto :: sgemm.goto dgemm.goto cgemm.goto zgemm.goto \
        samin.goto damin.goto camin.goto zamin.goto \
        smin.goto dmin.goto \
        saxpby.goto daxpby.goto caxpby.goto zaxpby.goto \
-       snrm2.goto dnrm2.goto scnrm2.goto dznrm2.goto $(GOTO_LAPACK_TARGETS)
+       snrm2.goto dnrm2.goto scnrm2.goto dznrm2.goto $(GOTO_LAPACK_TARGETS) $(GOTO_HALF_TARGETS)
 
 acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
        scholesky.acml dcholesky.acml ccholesky.acml zcholesky.acml \
@@ -614,6 +620,11 @@ zcholesky.essl : zcholesky.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBESSL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
 ##################################### Sgemm ####################################################
+ifeq ($(BUILD_HALF),1)
+shgemm.goto : shgemm.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+endif
+
 sgemm.goto : sgemm.$(SUFFIX) ../$(LIBNAME)
 	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
@@ -1814,7 +1825,7 @@ zsymv.veclib : zsymv.$(SUFFIX)
 
 ##################################### Sgeev ####################################################
 sgeev.goto : sgeev.$(SUFFIX) ../$(LIBNAME)
-	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+	$(FC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
 sgeev.acml : sgeev.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
@@ -1830,7 +1841,7 @@ sgeev.veclib : sgeev.$(SUFFIX)
 
 ##################################### Dgeev ####################################################
 dgeev.goto : dgeev.$(SUFFIX) ../$(LIBNAME)
-	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+	$(FC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
 dgeev.acml : dgeev.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
@@ -1847,7 +1858,7 @@ dgeev.veclib : dgeev.$(SUFFIX)
 ##################################### Cgeev ####################################################
 
 cgeev.goto : cgeev.$(SUFFIX) ../$(LIBNAME)
-	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+	$(FC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
 cgeev.acml : cgeev.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
@@ -1864,7 +1875,7 @@ cgeev.veclib : cgeev.$(SUFFIX)
 ##################################### Zgeev ####################################################
 
 zgeev.goto : zgeev.$(SUFFIX) ../$(LIBNAME)
-	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+	$(FC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
 zgeev.acml : zgeev.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
@@ -1880,7 +1891,7 @@ zgeev.veclib : zgeev.$(SUFFIX)
 
 ##################################### Sgetri ####################################################
 sgetri.goto : sgetri.$(SUFFIX) ../$(LIBNAME)
-	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+	$(FC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
 sgetri.acml : sgetri.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
@@ -1896,7 +1907,7 @@ sgetri.veclib : sgetri.$(SUFFIX)
 
 ##################################### Dgetri ####################################################
 dgetri.goto : dgetri.$(SUFFIX) ../$(LIBNAME)
-	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+	$(FC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
 dgetri.acml : dgetri.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
@@ -1913,7 +1924,7 @@ dgetri.veclib : dgetri.$(SUFFIX)
 ##################################### Cgetri ####################################################
 
 cgetri.goto : cgetri.$(SUFFIX) ../$(LIBNAME)
-	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+	$(FC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
 cgetri.acml : cgetri.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
@@ -1930,7 +1941,7 @@ cgetri.veclib : cgetri.$(SUFFIX)
 ##################################### Zgetri ####################################################
 
 zgetri.goto : zgetri.$(SUFFIX) ../$(LIBNAME)
-	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+	$(FC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
 zgetri.acml : zgetri.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
@@ -2916,6 +2927,11 @@ ccholesky.$(SUFFIX) : cholesky.c
 zcholesky.$(SUFFIX) : cholesky.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
+ifeq ($(BUILD_HALF),1)
+shgemm.$(SUFFIX) : gemm.c
+	$(CC) $(CFLAGS) -c -DHALF -UCOMPLEX -UDOUBLE -o $(@F) $^
+endif
+
 sgemm.$(SUFFIX) : gemm.c
 	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
 

benchmark/gemm.c

@@ -39,6 +39,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifdef DOUBLE
 #define GEMM   BLASFUNC(dgemm)
+#elif defined(HALF)
+#define GEMM   BLASFUNC(shgemm)
 #else
 #define GEMM   BLASFUNC(sgemm)
 #endif
@@ -120,7 +122,8 @@ static void *huge_malloc(BLASLONG size){
 
 int main(int argc, char *argv[]){
 
-  FLOAT *a, *b, *c;
+  IFLOAT *a, *b;
+  FLOAT *c;
   FLOAT alpha[] = {1.0, 0.0};
   FLOAT beta [] = {0.0, 0.0};
   char transa = 'N';
@@ -184,10 +187,10 @@ int main(int argc, char *argv[]){
     k = to;
   }
 
-  if (( a = (FLOAT *)malloc(sizeof(FLOAT) * m * k * COMPSIZE)) == NULL) {
+  if (( a = (IFLOAT *)malloc(sizeof(IFLOAT) * m * k * COMPSIZE)) == NULL) {
     fprintf(stderr,"Out of Memory!!\n");exit(1);
   }
-  if (( b = (FLOAT *)malloc(sizeof(FLOAT) * k * n * COMPSIZE)) == NULL) {
+  if (( b = (IFLOAT *)malloc(sizeof(IFLOAT) * k * n * COMPSIZE)) == NULL) {
     fprintf(stderr,"Out of Memory!!\n");exit(1);
   }
   if (( c = (FLOAT *)malloc(sizeof(FLOAT) * m * n * COMPSIZE)) == NULL) {
@@ -199,10 +202,10 @@ int main(int argc, char *argv[]){
 #endif
 
   for (i = 0; i < m * k * COMPSIZE; i++) {
-    a[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+    a[i] = ((IFLOAT) rand() / (IFLOAT) RAND_MAX) - 0.5;
   }
   for (i = 0; i < k * n * COMPSIZE; i++) {
-    b[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+    b[i] = ((IFLOAT) rand() / (IFLOAT) RAND_MAX) - 0.5;
   }
   for (i = 0; i < m * n * COMPSIZE; i++) {
     c[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;

benchmark/zdot.c

@@ -170,9 +170,11 @@ int main(int argc, char *argv[]){
 			y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
    	}
     	gettimeofday( &start, (struct timezone *)0);
-
+#ifdef RETURN_BY_STACK
+    	DOT (&result , &m, x, &inc_x, y, &inc_y );
+#else
     	result = DOT (&m, x, &inc_x, y, &inc_y );
-
+#endif
     	gettimeofday( &stop, (struct timezone *)0);
 
     	time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;

c_check

@@ -310,6 +310,7 @@ $linker_a = "";
 	    && ($flags !~ /advapi32/)
 	    && ($flags !~ /shell32/)
 	    && ($flags !~ /omp/)
+	    && ($flags !~ /[0-9]+/)
 	    ) {
 	    $linker_l .= $flags . " "
 	}

cmake/arch.cmake

@@ -49,7 +49,7 @@ if (DYNAMIC_ARCH)
   endif ()
   
   if (POWER)
-    set(DYNAMIC_CORE POWER6 POWER8 POWER9)
+	  set(DYNAMIC_CORE POWER6 POWER8 POWER9 POWER10)
   endif ()
   
   if (X86)

cmake/fc.cmake

@@ -16,6 +16,7 @@ if (${F_COMPILER} STREQUAL "FLANG")
   if (USE_OPENMP)
     set(FCOMMON_OPT "${FCOMMON_OPT} -fopenmp")
   endif ()
+  set(FCOMMON_OPT "${FCOMMON_OPT} -Mrecursive -Kieee")
 endif ()
 
 if (${F_COMPILER} STREQUAL "G77")

cmake/kernel.cmake

@@ -113,11 +113,31 @@ macro(SetDefaultL1)
   set(ZSUMKERNEL zsum.S)
   set(QSUMKERNEL sum.S)
   set(XSUMKERNEL zsum.S)
+if (BUILD_HALF)
+  set(SHAMINKERNEL ../arm/amin.c)
+  set(SHAMAXKERNEL ../arm/amax.c)
+  set(SHMAXKERNEL ../arm/max.c)
+  set(SHMINKERNEL ../arm/min.c)
+  set(ISHAMAXKERNEL ../arm/iamax.c)
+  set(ISHAMINKERNEL ../arm/iamin.c)
+  set(ISHMAXKERNEL ../arm/imax.c)
+  set(ISHMINKERNEL ../arm/imin.c)
+  set(SHASUMKERNEL ../arm/asum.c)
+  set(SHAXPYKERNEL ../arm/axpy.c)
+  set(SHAXPBYKERNEL ../arm/axpby.c)
+  set(SHCOPYKERNEL ../arm/copy.c)
+  set(SHDOTKERNEL ../arm/dot.c)
+  set(SHROTKERNEL ../arm/rot.c)
+  set(SHSCALKERNEL ../arm/scal.c)
+  set(SHNRM2KERNEL ../arm/nrm2.c)
+  set(SHSUMKERNEL ../arm/sum.c)
+  set(SHSWAPKERNEL ../arm/swap.c)
+endif ()
 endmacro ()
 
 macro(SetDefaultL2)
-  set(SGEMVNKERNEL gemv_n.S)
+  set(SGEMVNKERNEL ../arm/gemv_n.c)
-  set(SGEMVTKERNEL gemv_t.S)
+  set(SGEMVTKERNEL ../arm/gemv_t.c)
   set(DGEMVNKERNEL gemv_n.S)
   set(DGEMVTKERNEL gemv_t.S)
   set(CGEMVNKERNEL zgemv_n.S)
@@ -161,6 +181,11 @@ macro(SetDefaultL2)
   set(XHEMV_L_KERNEL ../generic/zhemv_k.c)
   set(XHEMV_V_KERNEL ../generic/zhemv_k.c)
   set(XHEMV_M_KERNEL ../generic/zhemv_k.c)
+if (BUILD_HALF)
+  set(SHGEMVNKERNEL ../arm/gemv_n.c)
+  set(SHGEMVTKERNEL ../arm/gemv_t.c)
+  set(SHGERKERNEL ../generic/ger.c)
+endif ()
 endmacro ()
 
 macro(SetDefaultL3)
@@ -168,4 +193,18 @@ macro(SetDefaultL3)
   set(DGEADD_KERNEL ../generic/geadd.c)
   set(CGEADD_KERNEL ../generic/zgeadd.c)
   set(ZGEADD_KERNEL ../generic/zgeadd.c)
+if (BUILD_HALF)
+  set(SHGEADD_KERNEL ../generic/geadd.c)
+  set(SHGEMMKERNEL ../generic/gemmkernel_2x2.c)
+  set(SHGEMM_BETA  ../generic/gemm_beta.c)
+  set(SHGEMMINCOPY ../generic/gemm_ncopy_2.c)
+  set(SHGEMMITCOPY ../generic/gemm_tcopy_2.c)
+  set(SHGEMMONCOPY ../generic/gemm_ncopy_2.c)
+  set(SHGEMMOTCOPY ../generic/gemm_tcopy_2.c)
+  set(SHGEMMINCOPYOBJ shgemm_incopy.o)
+  set(SHGEMMITCOPYOBJ shgemm_itcopy.o)
+  set(SHGEMMONCOPYOBJ shgemm_oncopy.o)
+  set(SHGEMMOTCOPYOBJ shgemm_otcopy.o)
+endif ()
+
 endmacro ()

cmake/os.cmake

@@ -8,7 +8,7 @@ if (${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
   set(NO_EXPRECISION 1)
 endif ()
 
-if (${CMAKE_SYSTEM_NAME} MATCHES "FreeBSD|OpenBSD|NetBSD|DragonFly")
+if (${CMAKE_SYSTEM_NAME} MATCHES "FreeBSD|OpenBSD|NetBSD|DragonFly|Darwin")
   set(EXTRALIB "${EXTRALIB} -lm")
   set(NO_EXPRECISION 1)
 endif ()

cmake/prebuild.cmake

@@ -16,6 +16,8 @@
 # HAVE_SSE2
 # HAVE_SSE3
 # MAKE
+# SHGEMM_UNROLL_M
+# SHGEMM_UNROLL_N
 # SGEMM_UNROLL_M
 # SGEMM_UNROLL_N
 # DGEMM_UNROLL_M
@@ -418,7 +420,7 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
     set(ZGEMM_UNROLL_M 8)
     set(ZGEMM_UNROLL_N 2)
     set(SYMV_P 8)
-  elseif ("${TCORE}" STREQUAL "POWER9")
+  elseif ("${TCORE}" STREQUAL "POWER9" OR "${TCORE}" STREQUAL "POWER10")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_DATA_SIZE 32768\n"
       "#define L1_DATA_LINESIZE 128\n"
@@ -437,6 +439,8 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
     set(ZGEMM_UNROLL_N 2)
     set(SYMV_P 8)
   endif()
+  set(SHGEMM_UNROLL_M 8)
+  set(SHGEMM_UNROLL_N 4)
 
   # Or should this actually be NUM_CORES?
   if (${NUM_THREADS} GREATER 0)
@@ -488,7 +492,7 @@ else(NOT CMAKE_CROSSCOMPILING)
   if (NOT "${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
     try_compile(GETARCH_RESULT ${GETARCH_DIR}
       SOURCES ${GETARCH_SRC}
-    COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} -I${GETARCH_DIR} -I"${PROJECT_SOURCE_DIR}" -I"${PROJECT_BINARY_DIR}"
+    COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} -I"${GETARCH_DIR}" -I"${PROJECT_SOURCE_DIR}" -I"${PROJECT_BINARY_DIR}"
       OUTPUT_VARIABLE GETARCH_LOG
       COPY_FILE ${PROJECT_BINARY_DIR}/${GETARCH_BIN}
     )
@@ -516,7 +520,7 @@ execute_process(COMMAND "${PROJECT_BINARY_DIR}/${GETARCH_BIN}" 1 OUTPUT_VARIABLE
   if (NOT "${CMAKE_SYSTEM_NAME}" STREQUAL "WindowsStore")
     try_compile(GETARCH2_RESULT ${GETARCH2_DIR}
       SOURCES ${PROJECT_SOURCE_DIR}/getarch_2nd.c
-    COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} ${GETARCH2_FLAGS} -I${GETARCH2_DIR} -I"${PROJECT_SOURCE_DIR}" -I"${PROJECT_BINARY_DIR}"
+    COMPILE_DEFINITIONS ${EXFLAGS} ${GETARCH_FLAGS} ${GETARCH2_FLAGS} -I"${GETARCH2_DIR}" -I"${PROJECT_SOURCE_DIR}" -I"${PROJECT_BINARY_DIR}"
       OUTPUT_VARIABLE GETARCH2_LOG
       COPY_FILE ${PROJECT_BINARY_DIR}/${GETARCH2_BIN}
     )

cmake/system.cmake

@@ -297,6 +297,16 @@ if (USE_SIMPLE_THREADED_LEVEL3)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_SIMPLE_THREADED_LEVEL3")
 endif ()
 
+if (NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
+if (DEFINED MAX_STACK_ALLOC)
+if (NOT ${MAX_STACK_ALLOC} EQUAL 0)
+set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_STACK_ALLOC=${MAX_STACK_ALLOC}")
+endif ()
+else ()
+set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_STACK_ALLOC=2048")
+endif ()
+endif ()
+
 if (DEFINED LIBNAMESUFFIX)
   set(LIBPREFIX "libopenblas_${LIBNAMESUFFIX}")
 else ()
@@ -407,6 +417,14 @@ if (${CMAKE_C_COMPILER} STREQUAL "LSB" OR ${CMAKE_SYSTEM_NAME} STREQUAL "Windows
   set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DLAPACK_COMPLEX_STRUCTURE")
 endif ()
 
+if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
+if ("${F_COMPILER}" STREQUAL "FLANG")
+if (${CMAKE_Fortran_COMPILER_VERSION} VERSION_LESS_EQUAL 3)
+  set(CMAKE_Fortran_FLAGS_RELEASE "${CMAKE_Fortran_FLAGS_RELEASE} -fno-unroll-loops")
+endif ()
+endif ()
+endif ()
+
 if (NOT DEFINED SUFFIX)
   set(SUFFIX o)
 endif ()
@@ -530,6 +548,8 @@ endif ()
 #export FUNCTION_PROFILE
 #export TARGET_CORE
 #
+#export SHGEMM_UNROLL_M
+#export SHGEMM_UNROLL_N
 #export SGEMM_UNROLL_M
 #export SGEMM_UNROLL_N
 #export DGEMM_UNROLL_M

cmake/utils.cmake

@@ -15,12 +15,36 @@ endfunction ()
 # Reads a Makefile into CMake vars.
 macro(ParseMakefileVars MAKEFILE_IN)
   message(STATUS "Reading vars from ${MAKEFILE_IN}...")
+        set (IfElse 0)
+        set (ElseSeen 0)
   file(STRINGS ${MAKEFILE_IN} makefile_contents)
   foreach (makefile_line ${makefile_contents})
+#message(STATUS "parsing ${makefile_line}")
+    if (${IfElse} GREATER 0)
+      string(REGEX MATCH "endif[ \t]*" line_match "${makefile_line}")
+      if (NOT "${line_match}" STREQUAL "")
+#           message(STATUS "ENDIF ${makefile_line}")
+        set (IfElse 0)
+        set (ElseSeen 0)
+        continue ()
+      endif ()
+      string(REGEX MATCH "else[ \t]*" line_match "${makefile_line}")
+        if (NOT "${line_match}" STREQUAL "") 
+#           message(STATUS "ELSE ${makefile_line}")
+           set (ElseSeen 1)        
+           continue ()  
+        endif()
+      if ( (${IfElse} EQUAL 2 AND ${ElseSeen} EQUAL 0) OR ( ${IfElse} EQUAL 1 AND ${ElseSeen} EQUAL 1))
+#           message(STATUS "skipping ${makefile_line}")
+         continue ()
+      endif ()    
+    endif ()    
     string(REGEX MATCH "([0-9_a-zA-Z]+)[ \t]*=[ \t]*(.+)$" line_match "${makefile_line}")
     if (NOT "${line_match}" STREQUAL "")
+#message(STATUS "match on ${line_match}")
       set(var_name ${CMAKE_MATCH_1})
-      set(var_value ${CMAKE_MATCH_2})
+#      set(var_value ${CMAKE_MATCH_2})
+      string(STRIP ${CMAKE_MATCH_2} var_value)
       # check for Makefile variables in the string, e.g. $(TSUFFIX)
       string(REGEX MATCHALL "\\$\\(([0-9_a-zA-Z]+)\\)" make_var_matches ${var_value})
       foreach (make_var ${make_var_matches})
@@ -33,7 +57,31 @@ macro(ParseMakefileVars MAKEFILE_IN)
     else ()
       string(REGEX MATCH "include \\$\\(KERNELDIR\\)/(.+)$" line_match "${makefile_line}")
       if (NOT "${line_match}" STREQUAL "")
+#message(STATUS "match on include ${line_match}")
         ParseMakefileVars(${KERNELDIR}/${CMAKE_MATCH_1})
+      else ()
+#        message(STATUS "unmatched line ${line_match}")
+        string(REGEX MATCH "ifeq \\(\\$\\(([_A-Z]+)\\),[ \t]*([0-9_A-Z]+)\\)" line_match "${makefile_line}")
+        if (NOT "${line_match}" STREQUAL "")
+#          message(STATUS "IFEQ: ${line_match} first: ${CMAKE_MATCH_1} second: ${CMAKE_MATCH_2}")
+          if (DEFINED ${${CMAKE_MATCH_1}} AND ${${CMAKE_MATCH_1}} STREQUAL ${CMAKE_MATCH_2})
+#            message (STATUS "condition is true")
+            set (IfElse 1)
+          else ()
+            set (IfElse 2)
+          endif ()
+        else ()
+          string(REGEX MATCH "ifneq \\(\\$\\(([_A-Z]+)\\),[ \t]*([0-9_A-Z]+)\\)" line_match "${makefile_line}")
+          if (NOT "${line_match}" STREQUAL "")
+#            message(STATUS "IFNEQ: ${line_match} first: ${CMAKE_MATCH_1} second: ${CMAKE_MATCH_2}")
+            if (NOT ( ${${CMAKE_MATCH_1}} STREQUAL ${CMAKE_MATCH_2}))
+#              message (STATUS "condition is true")
+              set (IfElse 1)
+            else ()
+              set (IfElse 2)
+            endif ()
+          endif ()
+        endif ()
       endif ()
     endif ()
   endforeach ()
@@ -163,6 +211,7 @@ function(GenerateNamedObjects sources_in)
   if (complex_only)
     list(REMOVE_ITEM float_list "SINGLE")
     list(REMOVE_ITEM float_list "DOUBLE")
+    list(REMOVE_ITEM float_list "HALF")
   elseif (real_only)
     list(REMOVE_ITEM float_list "COMPLEX")
     list(REMOVE_ITEM float_list "ZCOMPLEX")
@@ -176,6 +225,9 @@ function(GenerateNamedObjects sources_in)
       if (NOT no_float_type)
         string(SUBSTRING ${float_type} 0 1 float_char)
         string(TOLOWER ${float_char} float_char)
+	if (${float_type} STREQUAL "HALF")
+		set (float_char "sh")
+	endif ()
       endif ()
 
       if (NOT name_in)
@@ -210,6 +262,9 @@ function(GenerateNamedObjects sources_in)
       if (${float_type} STREQUAL "DOUBLE" OR ${float_type} STREQUAL "ZCOMPLEX")
         list(APPEND obj_defines "DOUBLE")
       endif ()
+      if (${float_type} STREQUAL "HALF")
+        list(APPEND obj_defines "HALF")
+      endif ()
       if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
         list(APPEND obj_defines "COMPLEX")
         if (mangle_complex_sources)

common.h

@@ -257,6 +257,11 @@ typedef long BLASLONG;
 typedef unsigned long BLASULONG;
 #endif
 
+#ifndef BFLOAT16
+typedef unsigned short bfloat16;
+#define HALFCONVERSION 1
+#endif
+
 #ifdef USE64BITINT
 typedef BLASLONG blasint;
 #if defined(OS_WINDOWS) && defined(__64BIT__)
@@ -297,6 +302,13 @@ typedef int blasint;
 #define SIZE	8
 #define  BASE_SHIFT 3
 #define ZBASE_SHIFT 4
+#elif defined(HALF)
+#define IFLOAT	bfloat16
+#define XFLOAT IFLOAT
+#define FLOAT	float
+#define SIZE   2
+#define BASE_SHIFT 1
+#define ZBASE_SHIFT 2
 #else
 #define FLOAT	float
 #define SIZE    4
@@ -308,6 +320,10 @@ typedef int blasint;
 #define XFLOAT	FLOAT
 #endif
 
+#ifndef IFLOAT
+#define IFLOAT	FLOAT
+#endif
+
 #ifndef COMPLEX
 #define COMPSIZE  1
 #else
@@ -344,13 +360,8 @@ typedef int blasint;
 #endif
 #endif
 
-#ifdef POWER8
-#ifndef YIELDING
-#define YIELDING        __asm__ __volatile__ ("nop;nop;nop;nop;nop;nop;nop;nop;\n");
-#endif
-#endif
 
-#ifdef POWER9
+#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #ifndef YIELDING
 #define YIELDING        __asm__ __volatile__ ("nop;nop;nop;nop;nop;nop;nop;nop;\n");
 #endif

common_interface.h

@@ -469,6 +469,8 @@ void BLASFUNC(xhbmv)(char *, blasint *, blasint *, xdouble *, xdouble *, blasint
 
 /* Level 3 routines */
 
+void BLASFUNC(shgemm)(char *, char *, blasint *, blasint *, blasint *, float *,
+	   bfloat16 *, blasint *, bfloat16 *, blasint *, float *, float *, blasint *);
 void BLASFUNC(sgemm)(char *, char *, blasint *, blasint *, blasint *, float *,
 	   float  *, blasint *, float  *, blasint *, float  *, float  *, blasint *);
 void BLASFUNC(dgemm)(char *, char *, blasint *, blasint *, blasint *, double *,

common_level3.h

@@ -55,6 +55,8 @@ extern void sgemm_kernel_direct(BLASLONG M, BLASLONG N, BLASLONG K,
 extern int sgemm_kernel_direct_performant(BLASLONG M, BLASLONG N, BLASLONG K);
 
 
+int shgemm_beta(BLASLONG, BLASLONG, BLASLONG, float,
+	       bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
 int sgemm_beta(BLASLONG, BLASLONG, BLASLONG, float,
 	       float  *, BLASLONG, float   *, BLASLONG, float  *, BLASLONG);
 int dgemm_beta(BLASLONG, BLASLONG, BLASLONG, double,
@@ -76,6 +78,10 @@ int xgemm_beta(BLASLONG, BLASLONG, BLASLONG, xdouble *,
 	       xdouble *, BLASLONG, xdouble  *, BLASLONG, xdouble *, BLASLONG);
 #endif
 
+int shgemm_incopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
+int shgemm_itcopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
+int shgemm_oncopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
+int shgemm_otcopy(BLASLONG m, BLASLONG n, bfloat16 *a, BLASLONG lda, bfloat16 *b);
 int sgemm_incopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
 int sgemm_itcopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
 int sgemm_oncopy(BLASLONG m, BLASLONG n, float *a, BLASLONG lda, float *b);
@@ -499,6 +505,7 @@ int xher2k_kernel_UC(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdoubl
 int xher2k_kernel_LN(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble alpha_i, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
 int xher2k_kernel_LC(BLASLONG m, BLASLONG n, BLASLONG k, xdouble alpha_r, xdouble alpha_i, xdouble *a, xdouble *b, xdouble *c, BLASLONG ldc, BLASLONG offset, int flag);
 
+int shgemm_kernel(BLASLONG, BLASLONG, BLASLONG, float,  bfloat16 *, bfloat16 *, float *, BLASLONG);
 int sgemm_kernel(BLASLONG, BLASLONG, BLASLONG, float,  float  *, float  *, float  *, BLASLONG);
 int dgemm_kernel(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG);
 
@@ -527,6 +534,11 @@ int cgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, float,  float,  float  *, float
 int zgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
 int xgemm3m_kernel(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
 
+int shgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
+int shgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
+int shgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
+int shgemm_tt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
+
 int sgemm_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
 int sgemm_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
 int sgemm_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
@@ -619,6 +631,11 @@ int xgemm_cr(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLON
 int xgemm_cc(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
 #endif
 
+int shgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
+int shgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
+int shgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
+int shgemm_thread_tt(blas_arg_t *, BLASLONG *, BLASLONG *, bfloat16 *, bfloat16 *, BLASLONG);
+
 int sgemm_thread_nn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
 int sgemm_thread_nt(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
 int sgemm_thread_tn(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);

common_macro.h

@@ -39,6 +39,7 @@
 #ifndef COMMON_MACRO
 #define COMMON_MACRO
 
+#include "common_sh.h"
 #include "common_s.h"
 #include "common_d.h"
 #include "common_q.h"
@@ -642,6 +643,288 @@
 #define IMATCOPY_K_RT		DIMATCOPY_K_RT
 
 #define GEADD_K                 DGEADD_K
+
+#elif defined(HALF)
+
+#define	AMAX_K			SAMAX_K
+#define	AMIN_K			SAMIN_K
+#define	MAX_K			SMAX_K
+#define	MIN_K			SMIN_K
+#define	IAMAX_K			ISAMAX_K
+#define	IAMIN_K			ISAMIN_K
+#define	IMAX_K			ISMAX_K
+#define	IMIN_K			ISMIN_K
+#define	ASUM_K			SASUM_K
+#define	DOTU_K			SDOTU_K
+#define	DOTC_K			SDOTC_K
+#define	AXPYU_K			SAXPYU_K
+#define	AXPYC_K			SAXPYC_K
+#define AXPBY_K     SAXPBY_K
+#define SCAL_K			SSCAL_K
+#define GEMV_N			SGEMV_N
+#define GEMV_T			SGEMV_T
+#define SYMV_U			SSYMV_U
+#define SYMV_L			SSYMV_L
+#define	GERU_K			SGERU_K
+#define	GERC_K			SGERC_K
+#define	GERV_K			SGERV_K
+#define	GERD_K			SGERD_K
+#define	SUM_K			SSUM_K
+#define	SWAP_K			SSWAP_K
+#define	ROT_K			SROT_K
+#define COPY_K    SCOPY_K
+#define NRM2_K    SNRM2_K
+#define SYMV_THREAD_U		SSYMV_THREAD_U
+#define SYMV_THREAD_L		SSYMV_THREAD_L
+#define GEMM_BETA               SHGEMM_BETA
+#define	GEMM_KERNEL_N		SHGEMM_KERNEL
+#define	GEMM_KERNEL_L		SHGEMM_KERNEL
+#define	GEMM_KERNEL_R		SHGEMM_KERNEL
+#define	GEMM_KERNEL_B		SHGEMM_KERNEL
+
+#define	GEMM_NN			SHGEMM_NN
+#define	GEMM_CN			SHGEMM_TN
+#define	GEMM_TN			SHGEMM_TN
+#define	GEMM_NC			SHGEMM_NT
+#define	GEMM_NT			SHGEMM_NT
+#define	GEMM_CC			SHGEMM_TT
+#define	GEMM_CT			SHGEMM_TT
+#define	GEMM_TC			SHGEMM_TT
+#define	GEMM_TT			SHGEMM_TT
+#define	GEMM_NR			SHGEMM_NN
+#define	GEMM_TR			SHGEMM_TN
+#define	GEMM_CR			SHGEMM_TN
+#define	GEMM_RN			SHGEMM_NN
+#define	GEMM_RT			SHGEMM_NT
+#define	GEMM_RC			SHGEMM_NT
+#define	GEMM_RR			SHGEMM_NN
+#define	GEMM_ONCOPY		SHGEMM_ONCOPY
+#define	GEMM_OTCOPY		SHGEMM_OTCOPY
+#define	GEMM_INCOPY		SHGEMM_INCOPY
+#define	GEMM_ITCOPY		SHGEMM_ITCOPY
+#define SYMM_THREAD_LU          SSYMM_THREAD_LU
+#define SYMM_THREAD_LL          SSYMM_THREAD_LL
+#define SYMM_THREAD_RU          SSYMM_THREAD_RU
+#define SYMM_THREAD_RL          SSYMM_THREAD_RL
+#define SYMM_LU                 SSYMM_LU
+#define SYMM_LL                 SSYMM_LL
+#define SYMM_RU                 SSYMM_RU
+#define SYMM_RL                 SSYMM_RL
+
+
+#define HEMM_THREAD_LU          SHEMM_THREAD_LU
+#define HEMM_THREAD_LL          SHEMM_THREAD_LL
+#define HEMM_THREAD_RU          SHEMM_THREAD_RU
+#define HEMM_THREAD_RL          SHEMM_THREAD_RL
+
+#define	GEMM_THREAD_NN		SHGEMM_THREAD_NN
+#define	GEMM_THREAD_CN		SHGEMM_THREAD_TN
+#define	GEMM_THREAD_TN		SHGEMM_THREAD_TN
+#define	GEMM_THREAD_NC		SHGEMM_THREAD_NT
+#define	GEMM_THREAD_NT		SHGEMM_THREAD_NT
+#define	GEMM_THREAD_CC		SHGEMM_THREAD_TT
+#define	GEMM_THREAD_CT		SHGEMM_THREAD_TT
+#define	GEMM_THREAD_TC		SHGEMM_THREAD_TT
+#define	GEMM_THREAD_TT		SHGEMM_THREAD_TT
+#define	GEMM_THREAD_NR		SHGEMM_THREAD_NN
+#define	GEMM_THREAD_TR		SHGEMM_THREAD_TN
+#define	GEMM_THREAD_CR		SHGEMM_THREAD_TN
+#define	GEMM_THREAD_RN		SHGEMM_THREAD_NN
+#define	GEMM_THREAD_RT		SHGEMM_THREAD_NT
+#define	GEMM_THREAD_RC		SHGEMM_THREAD_NT
+#define	GEMM_THREAD_RR		SHGEMM_THREAD_NN
+
+#ifdef UNIT
+
+#define	TRMM_OUNCOPY		STRMM_OUNUCOPY
+#define	TRMM_OUTCOPY		STRMM_OUTUCOPY
+#define	TRMM_OLNCOPY		STRMM_OLNUCOPY
+#define	TRMM_OLTCOPY		STRMM_OLTUCOPY
+#define	TRSM_OUNCOPY		STRSM_OUNUCOPY
+#define	TRSM_OUTCOPY		STRSM_OUTUCOPY
+#define	TRSM_OLNCOPY		STRSM_OLNUCOPY
+#define	TRSM_OLTCOPY		STRSM_OLTUCOPY
+
+#define	TRMM_IUNCOPY		STRMM_IUNUCOPY
+#define	TRMM_IUTCOPY		STRMM_IUTUCOPY
+#define	TRMM_ILNCOPY		STRMM_ILNUCOPY
+#define	TRMM_ILTCOPY		STRMM_ILTUCOPY
+#define	TRSM_IUNCOPY		STRSM_IUNUCOPY
+#define	TRSM_IUTCOPY		STRSM_IUTUCOPY
+#define	TRSM_ILNCOPY		STRSM_ILNUCOPY
+#define	TRSM_ILTCOPY		STRSM_ILTUCOPY
+
+#else
+
+#define	TRMM_OUNCOPY		STRMM_OUNNCOPY
+#define	TRMM_OUTCOPY		STRMM_OUTNCOPY
+#define	TRMM_OLNCOPY		STRMM_OLNNCOPY
+#define	TRMM_OLTCOPY		STRMM_OLTNCOPY
+#define	TRSM_OUNCOPY		STRSM_OUNNCOPY
+#define	TRSM_OUTCOPY		STRSM_OUTNCOPY
+#define	TRSM_OLNCOPY		STRSM_OLNNCOPY
+#define	TRSM_OLTCOPY		STRSM_OLTNCOPY
+
+#define	TRMM_IUNCOPY		STRMM_IUNNCOPY
+#define	TRMM_IUTCOPY		STRMM_IUTNCOPY
+#define	TRMM_ILNCOPY		STRMM_ILNNCOPY
+#define	TRMM_ILTCOPY		STRMM_ILTNCOPY
+#define	TRSM_IUNCOPY		STRSM_IUNNCOPY
+#define	TRSM_IUTCOPY		STRSM_IUTNCOPY
+#define	TRSM_ILNCOPY		STRSM_ILNNCOPY
+#define	TRSM_ILTCOPY		STRSM_ILTNCOPY
+
+#define	TRMM_KERNEL_LN		STRMM_KERNEL_LN
+#define	TRMM_KERNEL_LT		STRMM_KERNEL_LT
+#define	TRMM_KERNEL_LR		STRMM_KERNEL_LN
+#define	TRMM_KERNEL_LC		STRMM_KERNEL_LT
+#define	TRMM_KERNEL_RN		STRMM_KERNEL_RN
+#define	TRMM_KERNEL_RT		STRMM_KERNEL_RT
+#define	TRMM_KERNEL_RR		STRMM_KERNEL_RN
+#define	TRMM_KERNEL_RC		STRMM_KERNEL_RT
+
+#define	TRSM_KERNEL_LN		STRSM_KERNEL_LN
+#define	TRSM_KERNEL_LT		STRSM_KERNEL_LT
+#define	TRSM_KERNEL_LR		STRSM_KERNEL_LN
+#define	TRSM_KERNEL_LC		STRSM_KERNEL_LT
+#define	TRSM_KERNEL_RN		STRSM_KERNEL_RN
+#define	TRSM_KERNEL_RT		STRSM_KERNEL_RT
+#define	TRSM_KERNEL_RR		STRSM_KERNEL_RN
+#define	TRSM_KERNEL_RC		STRSM_KERNEL_RT
+
+#define SYMM_IUTCOPY		SSYMM_IUTCOPY
+#define SYMM_ILTCOPY		SSYMM_ILTCOPY
+#define SYMM_OUTCOPY		SSYMM_OUTCOPY
+#define SYMM_OLTCOPY		SSYMM_OLTCOPY
+#define	TRMM_LNUU		STRMM_LNUU
+#define	TRMM_LNUN		STRMM_LNUN
+#define	TRMM_LNLU		STRMM_LNLU
+#define	TRMM_LNLN		STRMM_LNLN
+#define	TRMM_LTUU		STRMM_LTUU
+#define	TRMM_LTUN		STRMM_LTUN
+#define	TRMM_LTLU		STRMM_LTLU
+#define	TRMM_LTLN		STRMM_LTLN
+#define	TRMM_LRUU		STRMM_LNUU
+#define	TRMM_LRUN		STRMM_LNUN
+#define	TRMM_LRLU		STRMM_LNLU
+#define	TRMM_LRLN		STRMM_LNLN
+#define	TRMM_LCUU		STRMM_LTUU
+#define	TRMM_LCUN		STRMM_LTUN
+#define	TRMM_LCLU		STRMM_LTLU
+#define	TRMM_LCLN		STRMM_LTLN
+#define	TRMM_RNUU		STRMM_RNUU
+#define	TRMM_RNUN		STRMM_RNUN
+#define	TRMM_RNLU		STRMM_RNLU
+#define	TRMM_RNLN		STRMM_RNLN
+#define	TRMM_RTUU		STRMM_RTUU
+#define	TRMM_RTUN		STRMM_RTUN
+#define	TRMM_RTLU		STRMM_RTLU
+#define	TRMM_RTLN		STRMM_RTLN
+#define	TRMM_RRUU		STRMM_RNUU
+#define	TRMM_RRUN		STRMM_RNUN
+#define	TRMM_RRLU		STRMM_RNLU
+#define	TRMM_RRLN		STRMM_RNLN
+#define	TRMM_RCUU		STRMM_RTUU
+#define	TRMM_RCUN		STRMM_RTUN
+#define	TRMM_RCLU		STRMM_RTLU
+#define	TRMM_RCLN		STRMM_RTLN
+
+#define	TRSM_LNUU		STRSM_LNUU
+#define	TRSM_LNUN		STRSM_LNUN
+#define	TRSM_LNLU		STRSM_LNLU
+#define	TRSM_LNLN		STRSM_LNLN
+#define	TRSM_LTUU		STRSM_LTUU
+#define	TRSM_LTUN		STRSM_LTUN
+#define	TRSM_LTLU		STRSM_LTLU
+#define	TRSM_LTLN		STRSM_LTLN
+#define	TRSM_LRUU		STRSM_LNUU
+#define	TRSM_LRUN		STRSM_LNUN
+#define	TRSM_LRLU		STRSM_LNLU
+#define	TRSM_LRLN		STRSM_LNLN
+#define	TRSM_LCUU		STRSM_LTUU
+#define	TRSM_LCUN		STRSM_LTUN
+#define	TRSM_LCLU		STRSM_LTLU
+#define	TRSM_LCLN		STRSM_LTLN
+#define	TRSM_RNUU		STRSM_RNUU
+#define	TRSM_RNUN		STRSM_RNUN
+#define	TRSM_RNLU		STRSM_RNLU
+#define	TRSM_RNLN		STRSM_RNLN
+#define	TRSM_RTUU		STRSM_RTUU
+#define	TRSM_RTUN		STRSM_RTUN
+#define	TRSM_RTLU		STRSM_RTLU
+#define	TRSM_RTLN		STRSM_RTLN
+#define	TRSM_RRUU		STRSM_RNUU
+#define	TRSM_RRUN		STRSM_RNUN
+#define	TRSM_RRLU		STRSM_RNLU
+#define	TRSM_RRLN		STRSM_RNLN
+#define	TRSM_RCUU		STRSM_RTUU
+#define	TRSM_RCUN		STRSM_RTUN
+#define	TRSM_RCLU		STRSM_RTLU
+#define	TRSM_RCLN		STRSM_RTLN
+#define	SYRK_UN			SSYRK_UN
+#define	SYRK_UT			SSYRK_UT
+#define	SYRK_LN			SSYRK_LN
+#define	SYRK_LT			SSYRK_LT
+#define	SYRK_UR			SSYRK_UN
+#define	SYRK_UC			SSYRK_UT
+#define	SYRK_LR			SSYRK_LN
+#define	SYRK_LC			SSYRK_LT
+
+#define	SYRK_KERNEL_U		SSYRK_KERNEL_U
+#define	SYRK_KERNEL_L		SSYRK_KERNEL_L
+
+#define	HERK_UN			SSYRK_UN
+#define	HERK_LN			SSYRK_LN
+#define	HERK_UC			SSYRK_UT
+#define	HERK_LC			SSYRK_LT
+
+#define	HER2K_UN		SSYR2K_UN
+#define	HER2K_LN		SSYR2K_LN
+#define	HER2K_UC		SSYR2K_UT
+#define	HER2K_LC		SSYR2K_LT
+
+#define	SYR2K_UN		SSYR2K_UN
+#define	SYR2K_UT		SSYR2K_UT
+#define	SYR2K_LN		SSYR2K_LN
+#define	SYR2K_LT		SSYR2K_LT
+#define	SYR2K_UR		SSYR2K_UN
+#define	SYR2K_UC		SSYR2K_UT
+#define	SYR2K_LR		SSYR2K_LN
+#define	SYR2K_LC		SSYR2K_LT
+
+#define	SYR2K_KERNEL_U		SSYR2K_KERNEL_U
+#define	SYR2K_KERNEL_L		SSYR2K_KERNEL_L
+#define	SYRK_THREAD_UN		SSYRK_THREAD_UN
+#define	SYRK_THREAD_UT		SSYRK_THREAD_UT
+#define	SYRK_THREAD_LN		SSYRK_THREAD_LN
+#define	SYRK_THREAD_LT		SSYRK_THREAD_LT
+#define	SYRK_THREAD_UR		SSYRK_THREAD_UR
+#define	SYRK_THREAD_UC		SSYRK_THREAD_UC
+#define	SYRK_THREAD_LR		SSYRK_THREAD_LN
+#define	SYRK_THREAD_LC		SSYRK_THREAD_LT
+
+#define	HERK_THREAD_UN		SSYRK_THREAD_UN
+#define	HERK_THREAD_UT		SSYRK_THREAD_UT
+#define	HERK_THREAD_LN		SSYRK_THREAD_LN
+#define	HERK_THREAD_LT		SSYRK_THREAD_LT
+#define	HERK_THREAD_UR		SSYRK_THREAD_UR
+#define	HERK_THREAD_UC		SSYRK_THREAD_UC
+#define	HERK_THREAD_LR		SSYRK_THREAD_LN
+#define	HERK_THREAD_LC		SSYRK_THREAD_LT
+
+#define OMATCOPY_K_CN		SOMATCOPY_K_CN
+#define OMATCOPY_K_RN		SOMATCOPY_K_RN
+#define OMATCOPY_K_CT		SOMATCOPY_K_CT
+#define OMATCOPY_K_RT		SOMATCOPY_K_RT
+#define IMATCOPY_K_CN		SIMATCOPY_K_CN
+#define IMATCOPY_K_RN		SIMATCOPY_K_RN
+#define IMATCOPY_K_CT		SIMATCOPY_K_CT
+#define IMATCOPY_K_RT		SIMATCOPY_K_RT
+
+#define GEADD_K 		SGEADD_K
+
+#endif
+
 #else
 
 #define	AMAX_K			SAMAX_K
@@ -673,14 +956,14 @@
 #define	GEMV_S			SGEMV_S
 #define	GEMV_D			SGEMV_D
 
+
+#define	SYMV_U			SSYMV_U
+#define	SYMV_L			SSYMV_L
 #define	GERU_K			SGERU_K
 #define	GERC_K			SGERC_K
 #define	GERV_K			SGERV_K
 #define	GERD_K			SGERD_K
 
-#define	SYMV_U			SSYMV_U
-#define	SYMV_L			SSYMV_L
-
 #define	SYMV_THREAD_U		SSYMV_THREAD_U
 #define	SYMV_THREAD_L		SSYMV_THREAD_L
 
@@ -2202,6 +2485,9 @@
 #if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64) || defined(ARCH_ARM64)
 extern BLASLONG gemm_offset_a;
 extern BLASLONG gemm_offset_b;
+extern BLASLONG shgemm_p;
+extern BLASLONG shgemm_q;
+extern BLASLONG shgemm_r;
 extern BLASLONG sgemm_p;
 extern BLASLONG sgemm_q;
 extern BLASLONG sgemm_r;

common_mips.h

@@ -43,6 +43,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifndef ASSEMBLER
 
+#if !defined(MIPS24K)
 static inline unsigned int rpcc(void){
   unsigned long ret;
 
@@ -53,6 +54,7 @@ static inline unsigned int rpcc(void){
   return ret;
 }
 #define RPCC_DEFINED
+#endif
 
 static inline int blas_quickdivide(blasint x, blasint y){
   return x / y;
@@ -92,7 +94,7 @@ REALNAME:
 #endif
 #define HUGE_PAGESIZE   ( 4 << 20)
 
-#define BUFFER_SIZE     (16 << 20)
+#define BUFFER_SIZE     (16 << 21)
 
 
 #define BASE_ADDRESS (START_ADDRESS - BUFFER_SIZE * MAX_CPU_NUMBER)

common_mips64.h

@@ -227,7 +227,7 @@ REALNAME: ;\
 
 #define SEEK_ADDRESS
 
-#define BUFFER_SIZE     ( 32 << 20)
+#define BUFFER_SIZE     ( 32 << 21)
 
 #if defined(LOONGSON3A)
 #define PAGESIZE	(16UL << 10)

common_param.h

@@ -47,6 +47,100 @@ typedef struct {
   int dtb_entries;
   int offsetA, offsetB, align;
 
+#ifdef BUILD_HALF
+  int shgemm_p, shgemm_q, shgemm_r;
+  int shgemm_unroll_m, shgemm_unroll_n, shgemm_unroll_mn;
+
+  float  (*shamax_k) (BLASLONG, float *, BLASLONG);
+  float  (*shamin_k) (BLASLONG, float *, BLASLONG);
+  float  (*shmax_k)  (BLASLONG, float *, BLASLONG);
+  float  (*shmin_k)  (BLASLONG, float *, BLASLONG);
+BLASLONG (*ishamax_k)(BLASLONG, float *, BLASLONG);
+BLASLONG (*ishamin_k)(BLASLONG, float *, BLASLONG);
+BLASLONG (*ishmax_k) (BLASLONG, float *, BLASLONG);
+BLASLONG (*ishmin_k) (BLASLONG, float *, BLASLONG);
+
+  float  (*shnrm2_k) (BLASLONG, float *, BLASLONG);
+  float  (*shasum_k) (BLASLONG, float *, BLASLONG);
+  float  (*shsum_k)  (BLASLONG, float *, BLASLONG);
+  int    (*shcopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
+  float  (*shdot_k)  (BLASLONG, float *, BLASLONG, float *, BLASLONG);
+  double (*dshdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
+
+  int    (*shrot_k)  (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
+
+  int    (*shaxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
+  int    (*shscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
+  int    (*shswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
+
+  int    (*shgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
+  int    (*shgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
+  int    (*shger_k)  (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
+
+  int    (*shsymv_L) (BLASLONG, BLASLONG, float,  float  *, BLASLONG, float  *, BLASLONG, float  *, BLASLONG, float *);
+  int    (*shsymv_U) (BLASLONG, BLASLONG, float,  float  *, BLASLONG, float  *, BLASLONG, float  *, BLASLONG, float *);
+
+  int    (*shgemm_kernel   )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, bfloat16 *, float *, BLASLONG);
+  int    (*shgemm_beta     )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
+
+  int    (*shgemm_incopy   )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
+  int    (*shgemm_itcopy   )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
+  int    (*shgemm_oncopy   )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
+  int    (*shgemm_otcopy   )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
+
+  int    (*shtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
+  int    (*shtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
+  int    (*shtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
+  int    (*shtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
+
+  int    (*shtrsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+  int    (*shtrsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
+
+  int    (*shtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
+  int    (*shtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
+  int    (*shtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
+  int    (*shtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
+
+  int    (*shtrmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shtrmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+
+  int    (*shsymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shsymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shsymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+  int    (*shsymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
+
+  int	 (*shneg_tcopy)   (BLASLONG, BLASLONG, float *, BLASLONG, float *);
+  int    (*shlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
+
+#endif
   int sgemm_p, sgemm_q, sgemm_r;
   int sgemm_unroll_m, sgemm_unroll_n, sgemm_unroll_mn;
 
@@ -84,6 +178,7 @@ BLASLONG (*ismin_k) (BLASLONG, float *, BLASLONG);
   int    (*sgemm_kernel   )(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG);
   int    (*sgemm_beta     )(BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float  *, BLASLONG);
 
+
   int    (*sgemm_incopy   )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
   int    (*sgemm_itcopy   )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
   int    (*sgemm_oncopy   )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
@@ -907,6 +1002,15 @@ extern gotoblas_t *gotoblas;
 
 #define HAVE_EX_L2	gotoblas -> exclusive_cache
 
+#ifdef BUILD_HALF
+#define	SHGEMM_P		gotoblas -> shgemm_p
+#define	SHGEMM_Q		gotoblas -> shgemm_q
+#define	SHGEMM_R		gotoblas -> shgemm_r
+#define	SHGEMM_UNROLL_M	gotoblas -> shgemm_unroll_m
+#define	SHGEMM_UNROLL_N	gotoblas -> shgemm_unroll_n
+#define	SHGEMM_UNROLL_MN	gotoblas -> shgemm_unroll_mn
+#endif
+
 #define	SGEMM_P		gotoblas -> sgemm_p
 #define	SGEMM_Q		gotoblas -> sgemm_q
 #define	SGEMM_R		gotoblas -> sgemm_r
@@ -984,6 +1088,19 @@ extern gotoblas_t *gotoblas;
 #define HAVE_EX_L2	0
 #endif
 
+#ifdef BUILD_HALF
+#define	SHGEMM_P		SHGEMM_DEFAULT_P
+#define	SHGEMM_Q		SHGEMM_DEFAULT_Q
+#define	SHGEMM_R		SHGEMM_DEFAULT_R
+#define SHGEMM_UNROLL_M	SHGEMM_DEFAULT_UNROLL_M
+#define SHGEMM_UNROLL_N	SHGEMM_DEFAULT_UNROLL_N
+#ifdef  SHGEMM_DEFAULT_UNROLL_MN
+#define SHGEMM_UNROLL_MN	SHGEMM_DEFAULT_UNROLL_MN
+#else
+#define SHGEMM_UNROLL_MN	MAX((SHGEMM_UNROLL_M), (SHGEMM_UNROLL_N))
+#endif
+#endif
+
 #define	SGEMM_P		SGEMM_DEFAULT_P
 #define	SGEMM_Q		SGEMM_DEFAULT_Q
 #define	SGEMM_R		SGEMM_DEFAULT_R
@@ -1119,6 +1236,18 @@ extern gotoblas_t *gotoblas;
 #define GEMM_DEFAULT_R		DGEMM_DEFAULT_R
 #define GEMM_DEFAULT_UNROLL_M	DGEMM_DEFAULT_UNROLL_M
 #define GEMM_DEFAULT_UNROLL_N	DGEMM_DEFAULT_UNROLL_N
+#elif defined(HALF)
+#define GEMM_P			SHGEMM_P
+#define GEMM_Q			SHGEMM_Q
+#define GEMM_R			SHGEMM_R
+#define GEMM_UNROLL_M		SHGEMM_UNROLL_M
+#define GEMM_UNROLL_N		SHGEMM_UNROLL_N
+#define GEMM_UNROLL_MN		SHGEMM_UNROLL_MN
+#define GEMM_DEFAULT_P		SHGEMM_DEFAULT_P
+#define GEMM_DEFAULT_Q		SHGEMM_DEFAULT_Q
+#define GEMM_DEFAULT_R		SHGEMM_DEFAULT_R
+#define GEMM_DEFAULT_UNROLL_M	SHGEMM_DEFAULT_UNROLL_M
+#define GEMM_DEFAULT_UNROLL_N	SHGEMM_DEFAULT_UNROLL_N
 #else
 #define GEMM_P			SGEMM_P
 #define GEMM_Q			SGEMM_Q
@@ -1204,28 +1333,32 @@ extern gotoblas_t *gotoblas;
 #define GEMM_THREAD gemm_thread_n
 #endif
 
+#ifndef SHGEMM_DEFAULT_R
+#define SHGEMM_DEFAULT_R (((BUFFER_SIZE - ((SHGEMM_DEFAULT_P * SHGEMM_DEFAULT_Q *  4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SHGEMM_DEFAULT_Q *  4) - 15) & ~15UL)
+#endif
+
 #ifndef SGEMM_DEFAULT_R
-#define SGEMM_DEFAULT_R (((BUFFER_SIZE - ((SGEMM_DEFAULT_P * SGEMM_DEFAULT_Q *  4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SGEMM_DEFAULT_Q *  4) - 15) & ~15)
+#define SGEMM_DEFAULT_R (((BUFFER_SIZE - ((SGEMM_DEFAULT_P * SGEMM_DEFAULT_Q *  4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SGEMM_DEFAULT_Q *  4) - 15) & ~15UL)
 #endif
 
 #ifndef DGEMM_DEFAULT_R
-#define DGEMM_DEFAULT_R (((BUFFER_SIZE - ((DGEMM_DEFAULT_P * DGEMM_DEFAULT_Q *  8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (DGEMM_DEFAULT_Q *  8) - 15) & ~15)
+#define DGEMM_DEFAULT_R (((BUFFER_SIZE - ((DGEMM_DEFAULT_P * DGEMM_DEFAULT_Q *  8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (DGEMM_DEFAULT_Q *  8) - 15) & ~15UL)
 #endif
 
 #ifndef QGEMM_DEFAULT_R
-#define QGEMM_DEFAULT_R (((BUFFER_SIZE - ((QGEMM_DEFAULT_P * QGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (QGEMM_DEFAULT_Q * 16) - 15) & ~15)
+#define QGEMM_DEFAULT_R (((BUFFER_SIZE - ((QGEMM_DEFAULT_P * QGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (QGEMM_DEFAULT_Q * 16) - 15) & ~15UL)
 #endif
 
 #ifndef CGEMM_DEFAULT_R
-#define CGEMM_DEFAULT_R (((BUFFER_SIZE - ((CGEMM_DEFAULT_P * CGEMM_DEFAULT_Q *  8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (CGEMM_DEFAULT_Q *  8) - 15) & ~15)
+#define CGEMM_DEFAULT_R (((BUFFER_SIZE - ((CGEMM_DEFAULT_P * CGEMM_DEFAULT_Q *  8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (CGEMM_DEFAULT_Q *  8) - 15) & ~15UL)
 #endif
 
 #ifndef ZGEMM_DEFAULT_R
-#define ZGEMM_DEFAULT_R (((BUFFER_SIZE - ((ZGEMM_DEFAULT_P * ZGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (ZGEMM_DEFAULT_Q * 16) - 15) & ~15)
+#define ZGEMM_DEFAULT_R (((BUFFER_SIZE - ((ZGEMM_DEFAULT_P * ZGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (ZGEMM_DEFAULT_Q * 16) - 15) & ~15UL)
 #endif
 
 #ifndef XGEMM_DEFAULT_R
-#define XGEMM_DEFAULT_R (((BUFFER_SIZE - ((XGEMM_DEFAULT_P * XGEMM_DEFAULT_Q * 32 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (XGEMM_DEFAULT_Q * 32) - 15) & ~15)
+#define XGEMM_DEFAULT_R (((BUFFER_SIZE - ((XGEMM_DEFAULT_P * XGEMM_DEFAULT_Q * 32 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (XGEMM_DEFAULT_Q * 32) - 15) & ~15UL)
 #endif
 
 #ifndef SNUMOPT

common_power.h

@@ -68,7 +68,7 @@
 #endif
 
 
-#if defined(POWER8) || defined(POWER9)
+#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #define MB		__asm__ __volatile__ ("eieio":::"memory")
 #define WMB		__asm__ __volatile__ ("eieio":::"memory")
 #define RMB		__asm__ __volatile__ ("eieio":::"memory")
@@ -272,7 +272,7 @@ static inline int blas_quickdivide(blasint x, blasint y){
 #define HAVE_PREFETCH
 #endif
 
-#if defined(POWER3) || defined(POWER6) || defined(PPCG4) || defined(CELL) || defined(POWER8) || defined(POWER9) || defined(PPC970)
+#if defined(POWER3) || defined(POWER6) || defined(PPCG4) || defined(CELL) || defined(POWER8) || defined(POWER9) || defined(POWER10) || defined(PPC970)
 #define DCBT_ARG	0
 #else
 #define DCBT_ARG	8
@@ -294,7 +294,7 @@ static inline int blas_quickdivide(blasint x, blasint y){
 #define L1_PREFETCH	dcbtst
 #endif
 
-#if defined(POWER8) || defined(POWER9)
+#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #define L1_DUALFETCH
 #define L1_PREFETCHSIZE (16 + 128 * 100)
 #define L1_PREFETCH	dcbtst
@@ -843,7 +843,7 @@ Lmcount$lazy_ptr:
 #define BUFFER_SIZE     (  2 << 20)
 #elif defined(PPC440FP2)
 #define BUFFER_SIZE     ( 16 << 20)
-#elif defined(POWER8) || defined(POWER9)
+#elif defined(POWER8) || defined(POWER9) || defined(POWER10)
 #define BUFFER_SIZE     ( 64 << 20)
 #else
 #define BUFFER_SIZE     ( 16 << 20)

common_sh.h

@@ -0,0 +1,65 @@
+#ifndef COMMON_SH_H
+#define COMMON_SH_H
+
+#ifndef DYNAMIC_ARCH
+
+#define	SHGEMM_ONCOPY		shgemm_oncopy
+#define	SHGEMM_OTCOPY		shgemm_otcopy
+
+#if SHGEMM_DEFAULT_UNROLL_M == SHGEMM_DEFAULT_UNROLL_N
+#define	SHGEMM_INCOPY		shgemm_oncopy
+#define	SHGEMM_ITCOPY		shgemm_otcopy
+#else
+#define	SHGEMM_INCOPY		shgemm_incopy
+#define	SHGEMM_ITCOPY		shgemm_itcopy
+#endif
+#define	SHGEMM_BETA		shgemm_beta
+#define SHGEMM_KERNEL            shgemm_kernel
+
+#else
+
+#define	SHGEMM_ONCOPY		gotoblas -> shgemm_oncopy
+#define	SHGEMM_OTCOPY		gotoblas -> shgemm_otcopy
+#define	SHGEMM_INCOPY		gotoblas -> shgemm_incopy
+#define	SHGEMM_ITCOPY		gotoblas -> shgemm_itcopy
+#define	SHGEMM_BETA		gotoblas -> shgemm_beta
+#define	SHGEMM_KERNEL		gotoblas -> shgemm_kernel
+
+#endif
+
+#define	SHGEMM_NN		shgemm_nn
+#define	SHGEMM_CN		shgemm_tn
+#define	SHGEMM_TN		shgemm_tn
+#define	SHGEMM_NC		shgemm_nt
+#define	SHGEMM_NT		shgemm_nt
+#define	SHGEMM_CC		shgemm_tt
+#define	SHGEMM_CT		shgemm_tt
+#define	SHGEMM_TC		shgemm_tt
+#define	SHGEMM_TT		shgemm_tt
+#define	SHGEMM_NR		shgemm_nn
+#define	SHGEMM_TR		shgemm_tn
+#define	SHGEMM_CR		shgemm_tn
+#define	SHGEMM_RN		shgemm_nn
+#define	SHGEMM_RT		shgemm_nt
+#define	SHGEMM_RC		shgemm_nt
+#define	SHGEMM_RR		shgemm_nn
+
+#define	SHGEMM_THREAD_NN		shgemm_thread_nn
+#define	SHGEMM_THREAD_CN		shgemm_thread_tn
+#define	SHGEMM_THREAD_TN		shgemm_thread_tn
+#define	SHGEMM_THREAD_NC		shgemm_thread_nt
+#define	SHGEMM_THREAD_NT		shgemm_thread_nt
+#define	SHGEMM_THREAD_CC		shgemm_thread_tt
+#define	SHGEMM_THREAD_CT		shgemm_thread_tt
+#define	SHGEMM_THREAD_TC		shgemm_thread_tt
+#define	SHGEMM_THREAD_TT		shgemm_thread_tt
+#define	SHGEMM_THREAD_NR		shgemm_thread_nn
+#define	SHGEMM_THREAD_TR		shgemm_thread_tn
+#define	SHGEMM_THREAD_CR		shgemm_thread_tn
+#define	SHGEMM_THREAD_RN		shgemm_thread_nn
+#define	SHGEMM_THREAD_RT		shgemm_thread_nt
+#define	SHGEMM_THREAD_RC		shgemm_thread_nt
+#define	SHGEMM_THREAD_RR		shgemm_thread_nn
+
+#endif
+

common_x86_64.h

@@ -80,7 +80,7 @@ static void __inline blas_lock(volatile BLASULONG *address){
 #endif
 
   do {
-    while (*address) {YIELDING;};
+    while (*address) {YIELDING;}
 
 #ifndef C_MSVC
     __asm__ __volatile__(
@@ -199,9 +199,9 @@ static __inline BLASLONG blas_quickdivide(BLASLONG x, BLASLONG y){
 #else
 extern unsigned int blas_quick_divide_table[];
 
-static __inline int blas_quickdivide(unsigned int x, unsigned int y){
+static __inline unsigned int blas_quickdivide(unsigned int x, unsigned int y){
 
-  unsigned int result;
+  volatile unsigned int result;
 
   if (y <= 1) return x;
 
@@ -215,7 +215,6 @@ 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));
-
   return result;
 }
 #endif

cpp_thread_test/cpp_thread_safety_common.h

@@ -5,6 +5,14 @@ inline void pauser(){
     std::getline(std::cin, dummy);
 }
 
+void FailIfThreadsAreZero(uint32_t numConcurrentThreads) {
+	if(numConcurrentThreads == 0) {
+		std::cout<<"ERROR: Invalid parameter 0 for number of concurrent calls into OpenBLAS!"<<std::endl;
+		std::cout<<"CBLAS DGEMV thread safety test FAILED!"<<std::endl;
+		exit(-1);
+	}
+}
+
 void FillMatrices(std::vector<std::vector<double>>& matBlock, std::mt19937_64& PRNG, std::uniform_real_distribution<double>& rngdist, const blasint randomMatSize, const uint32_t numConcurrentThreads, const uint32_t numMat){
 	for(uint32_t i=0; i<numMat; i++){
 		for(uint32_t j = 0; j < static_cast<uint32_t>(randomMatSize*randomMatSize); j++){

cpp_thread_test/dgemm_thread_safety.cpp

@@ -47,6 +47,8 @@ int main(int argc, char* argv[]){
 	std::cout<<"Number of testing rounds : "<<numTestRounds<<'\n';
 	std::cout<<"This test will need "<<(static_cast<uint64_t>(randomMatSize*randomMatSize)*numConcurrentThreads*3*8)/static_cast<double>(1024*1024)<<" MiB of RAM\n"<<std::endl;
 
+	FailIfThreadsAreZero(numConcurrentThreads);
+	
 	std::cout<<"Initializing random number generator..."<<std::flush;
 	std::mt19937_64 PRNG = InitPRNG();
 	std::cout<<"done\n";

cpp_thread_test/dgemv_thread_safety.cpp

@@ -18,7 +18,7 @@ int main(int argc, char* argv[]){
 	uint32_t maxHwThreads = omp_get_max_threads();
 	
 	if (maxHwThreads < 52)
-		numConcurrentThreads = maxHwThreads -4;
+		numConcurrentThreads = maxHwThreads;
 	
 	if (argc > 4){
 		std::cout<<"ERROR: too many arguments for thread safety tester"<<std::endl;
@@ -48,6 +48,8 @@ int main(int argc, char* argv[]){
 	std::cout<<"Number of testing rounds : "<<numTestRounds<<'\n';
 	std::cout<<"This test will need "<<((static_cast<uint64_t>(randomMatSize*randomMatSize)*numConcurrentThreads*8)+(static_cast<uint64_t>(randomMatSize)*numConcurrentThreads*8*2))/static_cast<double>(1024*1024)<<" MiB of RAM\n"<<std::endl;
 
+	FailIfThreadsAreZero(numConcurrentThreads);
+	
 	std::cout<<"Initializing random number generator..."<<std::flush;
 	std::mt19937_64 PRNG = InitPRNG();
 	std::cout<<"done\n";

cpuid_mips.c

@@ -73,11 +73,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CPU_UNKNOWN     0
 #define CPU_P5600       1
 #define CPU_1004K	2
+#define CPU_24K		3
 
 static char *cpuname[] = {
   "UNKNOWN",
   "P5600",
-  "1004K"
+  "MIPS1004K",
+  "MIPS24K"
 };
 
 int detect(void){
@@ -105,6 +107,8 @@ int detect(void){
     return CPU_P5600;
   } else if (strstr(p, "1004K")) {
     return CPU_1004K;
+  } else if (strstr(p, " 24K")) {
+    return CPU_24K;
   } else  
     return CPU_UNKNOWN;
   }
@@ -121,7 +125,7 @@ void get_architecture(void){
 }
 
 void get_subarchitecture(void){
-  if(detect()==CPU_P5600|| detect()==CPU_1004K){
+  if(detect()==CPU_P5600|| detect()==CPU_1004K|| detect()==CPU_24K){
     printf("P5600");
   }else{
     printf("UNKNOWN");
@@ -146,7 +150,15 @@ void get_cpuconfig(void){
     printf("#define MIPS1004K\n");
     printf("#define L1_DATA_SIZE 32768\n");
     printf("#define L1_DATA_LINESIZE 32\n");
-    printf("#define L2_SIZE 26144\n");
+    printf("#define L2_SIZE 262144\n");
+    printf("#define DTB_DEFAULT_ENTRIES 8\n");
+    printf("#define DTB_SIZE 4096\n");
+    printf("#define L2_ASSOCIATIVE 4\n");
+  } else if (detect()==CPU_24K) {
+    printf("#define MIPS24K\n");
+    printf("#define L1_DATA_SIZE 32768\n");
+    printf("#define L1_DATA_LINESIZE 32\n");
+    printf("#define L2_SIZE 32768\n");
     printf("#define DTB_DEFAULT_ENTRIES 8\n");
     printf("#define DTB_SIZE 4096\n");
     printf("#define L2_ASSOCIATIVE 4\n");
@@ -159,7 +171,9 @@ void get_libname(void){
   if(detect()==CPU_P5600) {
     printf("p5600\n");
   } else if (detect()==CPU_1004K) {
-    printf("1004K\n");
+    printf("mips1004K\n");
+  } else if (detect()==CPU_24K) {
+    printf("mips24K\n");
   }else{
     printf("mips\n");
   }

cpuid_power.c

@@ -57,6 +57,7 @@
 #define CPUTYPE_PPCG4	   7
 #define CPUTYPE_POWER8     8
 #define CPUTYPE_POWER9     9
+#define CPUTYPE_POWER10    10
 
 char *cpuname[] = {
   "UNKNOWN",
@@ -68,7 +69,8 @@ char *cpuname[] = {
   "CELL",
   "PPCG4",
   "POWER8",
-  "POWER9"
+  "POWER9",
+  "POWER10"
 };
 
 char *lowercpuname[] = {
@@ -81,7 +83,8 @@ char *lowercpuname[] = {
   "cell",
   "ppcg4",
   "power8",
-  "power9"	
+  "power9",
+  "power10"
 };
 
 char *corename[] = {
@@ -94,7 +97,8 @@ char *corename[] = {
   "CELL",
   "PPCG4",
   "POWER8",
-  "POWER9"   	
+  "POWER9",
+  "POWER10"
 };
 
 int detect(void){
@@ -125,6 +129,7 @@ int detect(void){
   if (!strncasecmp(p, "POWER7", 6)) return CPUTYPE_POWER6;
   if (!strncasecmp(p, "POWER8", 6)) return CPUTYPE_POWER8;
   if (!strncasecmp(p, "POWER9", 6)) return CPUTYPE_POWER9;
+  if (!strncasecmp(p, "POWER10", 7)) return CPUTYPE_POWER10;
   if (!strncasecmp(p, "Cell",   4)) return CPUTYPE_CELL;
   if (!strncasecmp(p, "7447",   4)) return CPUTYPE_PPCG4;
 
@@ -157,6 +162,7 @@ int detect(void){
   if (!strncasecmp(p, "POWER7", 6)) return CPUTYPE_POWER6;
   if (!strncasecmp(p, "POWER8", 6)) return CPUTYPE_POWER8;
   if (!strncasecmp(p, "POWER9", 6)) return CPUTYPE_POWER9;
+  if (!strncasecmp(p, "POWER10", 7)) return CPUTYPE_POWER10;
   if (!strncasecmp(p, "Cell",   4)) return CPUTYPE_CELL;
   if (!strncasecmp(p, "7447",   4)) return CPUTYPE_PPCG4;
   return CPUTYPE_POWER5;
@@ -179,6 +185,9 @@ int detect(void){
 int id;
 __asm __volatile("mfpvr %0" : "=r"(id));
 switch ( id >> 16 ) {
+  case 0x80: // POWER10
+    return CPUTYPE_POWER10;
+    break;
   case 0x4e: // POWER9
     return CPUTYPE_POWER9;
     break;

cpuid_x86.c

@@ -1406,6 +1406,17 @@ int get_cpuname(void){
 	    return CPUTYPE_SANDYBRIDGE;
           else
 	    return CPUTYPE_NEHALEM;
+    }
+      case 10: //family 6 exmodel 10
+        switch (model) {
+    case 5: // Comet Lake H and S
+    case 6: // Comet Lake U
+          if(support_avx2())
+            return CPUTYPE_HASWELL;
+          if(support_avx())
+        return CPUTYPE_SANDYBRIDGE;
+          else
+        return CPUTYPE_NEHALEM;
 	}
 	break;    
       }
@@ -1955,6 +1966,19 @@ int get_coretype(void){
 	    return CORE_NEHALEM;
         }
         break;
+      case 10:
+        switch (model) {
+	  case 5: // Comet Lake H and S
+    	  case 6: // Comet Lake U
+            if(support_avx())
+  #ifndef NO_AVX2
+              return CORE_HASWELL;
+  #else
+              return CORE_SANDYBRIDGE;
+  #endif
+            else
+              return CORE_NEHALEM;
+        }
       case 5:
         switch (model) {
 	case 6:

ctest/CMakeLists.txt

@@ -12,6 +12,9 @@ FILE(WRITE ${CMAKE_CURRENT_BINARY_DIR}/test_cblas_helper.sh
 foreach(float_type ${FLOAT_TYPES})
   string(SUBSTRING ${float_type} 0 1 float_char_upper)
   string(TOLOWER ${float_char_upper} float_char)
+  if (${float_char} STREQUAL "h")
+   continue()
+  endif()
   #level1
   add_executable(x${float_char}cblat1
     c_${float_char}blat1.f

driver/level3/Makefile

@@ -19,6 +19,10 @@ ifeq ($(ARCH), MIPS)
 USE_GEMM3M = 1
 endif
 
+ifeq ($(BUILD_HALF),1)
+SHBLASOBJS       += shgemm_nn.$(SUFFIX) shgemm_nt.$(SUFFIX) shgemm_tn.$(SUFFIX) shgemm_tt.$(SUFFIX)
+endif
+
 SBLASOBJS	+= \
 	sgemm_nn.$(SUFFIX) sgemm_nt.$(SUFFIX) sgemm_tn.$(SUFFIX) sgemm_tt.$(SUFFIX) \
 	strmm_LNUU.$(SUFFIX) strmm_LNUN.$(SUFFIX) strmm_LNLU.$(SUFFIX) strmm_LNLN.$(SUFFIX) \
@@ -203,7 +207,9 @@ COMMONOBJS  += gemm_thread_m.$(SUFFIX) gemm_thread_n.$(SUFFIX) gemm_thread_mn.$(
 COMMONOBJS  += syrk_thread.$(SUFFIX)
 
 ifndef USE_SIMPLE_THREADED_LEVEL3
-
+ifeq ($(BUILD_HALF),1)
+SHBLASOBJS    += shgemm_thread_nn.$(SUFFIX) shgemm_thread_nt.$(SUFFIX) shgemm_thread_tn.$(SUFFIX) shgemm_thread_tt.$(SUFFIX)
+endif
 SBLASOBJS    += sgemm_thread_nn.$(SUFFIX) sgemm_thread_nt.$(SUFFIX) sgemm_thread_tn.$(SUFFIX) sgemm_thread_tt.$(SUFFIX)
 DBLASOBJS    += dgemm_thread_nn.$(SUFFIX) dgemm_thread_nt.$(SUFFIX) dgemm_thread_tn.$(SUFFIX) dgemm_thread_tt.$(SUFFIX)
 QBLASOBJS    += qgemm_thread_nn.$(SUFFIX) qgemm_thread_nt.$(SUFFIX) qgemm_thread_tn.$(SUFFIX) qgemm_thread_tt.$(SUFFIX)
@@ -283,6 +289,18 @@ endif
 
 all ::
 
+shgemm_nn.$(SUFFIX) : gemm.c level3.c ../../param.h
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DHALF -UDOUBLE -UCOMPLEX -DNN $< -o $(@F)
+
+shgemm_nt.$(SUFFIX) : gemm.c level3.c ../../param.h
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DHALF -UDOUBLE -UCOMPLEX -DNT $< -o $(@F)
+
+shgemm_tn.$(SUFFIX) : gemm.c level3.c ../../param.h
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DHALF -UDOUBLE -UCOMPLEX -DTN $< -o $(@F)
+
+shgemm_tt.$(SUFFIX) : gemm.c level3.c ../../param.h
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DHALF -UDOUBLE -UCOMPLEX -DTT $< -o $(@F)
+
 sgemm_nn.$(SUFFIX) : gemm.c level3.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -UDOUBLE -UCOMPLEX -DNN $< -o $(@F)
 
@@ -478,6 +496,17 @@ gemm_thread_variable.$(SUFFIX) : gemm_thread_variable.c ../../common.h
 beta_thread.$(SUFFIX) : beta_thread.c ../../common.h
 	$(CC) -c $(CFLAGS) $< -o $(@F)
 
+shgemm_thread_nn.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DHALF -UDOUBLE -UCOMPLEX -DNN $< -o $(@F)
+
+shgemm_thread_nt.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DHALF -UDOUBLE -UCOMPLEX -DNT $< -o $(@F)
+
+shgemm_thread_tn.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DHALF -UDOUBLE -UCOMPLEX -DTN $< -o $(@F)
+
+shgemm_thread_tt.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
+	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DHALF -UDOUBLE -UCOMPLEX -DTT $< -o $(@F)
 
 sgemm_thread_nn.$(SUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(CFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -UCOMPLEX -DNN $< -o $(@F)
@@ -2652,6 +2681,18 @@ xtrsm_RCLU.$(SUFFIX) : trsm_R.c
 xtrsm_RCLN.$(SUFFIX) : trsm_R.c
 	$(CC) -c $(CFLAGS) -DCOMPLEX -DXDOUBLE -DTRANSA -UUPPER -UUNIT -DCONJ $< -o $(@F)
 
+shgemm_nn.$(PSUFFIX) : gemm.c level3.c ../../param.h
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DHALF -UDOUBLE -UCOMPLEX -DNN $< -o $(@F)
+
+shgemm_nt.$(PSUFFIX) : gemm.c level3.c ../../param.h
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DHALF -UDOUBLE -UCOMPLEX -DNT $< -o $(@F)
+
+shgemm_tn.$(PSUFFIX) : gemm.c level3.c ../../param.h
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DHALF -UDOUBLE -UCOMPLEX -DTN $< -o $(@F)
+
+shgemm_tt.$(PSUFFIX) : gemm.c level3.c ../../param.h
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DHALF -UDOUBLE -UCOMPLEX -DTT $< -o $(@F)
+
 sgemm_nn.$(PSUFFIX) : gemm.c level3.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -UDOUBLE -UCOMPLEX -DNN $< -o $(@F)
 
@@ -2848,6 +2889,18 @@ beta_thread.$(PSUFFIX) : beta_thread.c ../../common.h
 	$(CC) -c $(PFLAGS) $< -o $(@F)
 
 
+shgemm_thread_nn.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DHALF -UDOUBLE -UCOMPLEX -DNN $< -o $(@F)
+
+shgemm_thread_nt.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DHALF -UDOUBLE -UCOMPLEX -DNT $< -o $(@F)
+
+shgemm_thread_tn.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DHALF -UDOUBLE -UCOMPLEX -DTN $< -o $(@F)
+
+shgemm_thread_tt.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
+	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -DHALF -UDOUBLE -UCOMPLEX -DTT $< -o $(@F)
+
 sgemm_thread_nn.$(PSUFFIX) : gemm.c level3_thread.c ../../param.h
 	$(CC) $(PFLAGS) $(BLOCKS) -c -DTHREADED_LEVEL3 -UDOUBLE -UCOMPLEX -DNN $< -o $(@F)
 

driver/level3/level3.c

@@ -62,18 +62,18 @@
 #ifndef ICOPY_OPERATION
 #if defined(NN) || defined(NT) || defined(NC) || defined(NR) || \
     defined(RN) || defined(RT) || defined(RC) || defined(RR)
-#define ICOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_ITCOPY(M, N, (FLOAT *)(A) + ((Y) + (X) * (LDA)) * COMPSIZE, LDA, BUFFER);
+#define ICOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_ITCOPY(M, N, (IFLOAT *)(A) + ((Y) + (X) * (LDA)) * COMPSIZE, LDA, BUFFER);
 #else
-#define ICOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_INCOPY(M, N, (FLOAT *)(A) + ((X) + (Y) * (LDA)) * COMPSIZE, LDA, BUFFER);
+#define ICOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_INCOPY(M, N, (IFLOAT *)(A) + ((X) + (Y) * (LDA)) * COMPSIZE, LDA, BUFFER);
 #endif
 #endif
 
 #ifndef OCOPY_OPERATION
 #if defined(NN) || defined(TN) || defined(CN) || defined(RN) || \
     defined(NR) || defined(TR) || defined(CR) || defined(RR)
-#define OCOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_ONCOPY(M, N, (FLOAT *)(A) + ((X) + (Y) * (LDA)) * COMPSIZE, LDA, BUFFER);
+#define OCOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_ONCOPY(M, N, (IFLOAT *)(A) + ((X) + (Y) * (LDA)) * COMPSIZE, LDA, BUFFER);
 #else
-#define OCOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_OTCOPY(M, N, (FLOAT *)(A) + ((Y) + (X) * (LDA)) * COMPSIZE, LDA, BUFFER);
+#define OCOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_OTCOPY(M, N, (IFLOAT *)(A) + ((Y) + (X) * (LDA)) * COMPSIZE, LDA, BUFFER);
 #endif
 #endif
 
@@ -173,7 +173,8 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 		  XFLOAT *sa, XFLOAT *sb, BLASLONG dummy){
   BLASLONG k, lda, ldb, ldc;
   FLOAT *alpha, *beta;
-  FLOAT *a, *b, *c;
+  IFLOAT *a, *b;
+  FLOAT *c;
   BLASLONG m_from, m_to, n_from, n_to;
 
   BLASLONG ls, is, js;
@@ -198,8 +199,8 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 
   k = K;
 
-  a = (FLOAT *)A;
+  a = (IFLOAT *)A;
-  b = (FLOAT *)B;
+  b = (IFLOAT *)B;
   c = (FLOAT *)C;
 
   lda = LDA;

driver/level3/level3_thread.c

@@ -117,18 +117,18 @@ typedef struct {
 #ifndef ICOPY_OPERATION
 #if defined(NN) || defined(NT) || defined(NC) || defined(NR) || \
   defined(RN) || defined(RT) || defined(RC) || defined(RR)
-#define ICOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_ITCOPY(M, N, (FLOAT *)(A) + ((Y) + (X) * (LDA)) * COMPSIZE, LDA, BUFFER);
+#define ICOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_ITCOPY(M, N, (IFLOAT *)(A) + ((Y) + (X) * (LDA)) * COMPSIZE, LDA, BUFFER);
 #else
-#define ICOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_INCOPY(M, N, (FLOAT *)(A) + ((X) + (Y) * (LDA)) * COMPSIZE, LDA, BUFFER);
+#define ICOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_INCOPY(M, N, (IFLOAT *)(A) + ((X) + (Y) * (LDA)) * COMPSIZE, LDA, BUFFER);
 #endif
 #endif
 
 #ifndef OCOPY_OPERATION
 #if defined(NN) || defined(TN) || defined(CN) || defined(RN) || \
   defined(NR) || defined(TR) || defined(CR) || defined(RR)
-#define OCOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_ONCOPY(M, N, (FLOAT *)(A) + ((X) + (Y) * (LDA)) * COMPSIZE, LDA, BUFFER);
+#define OCOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_ONCOPY(M, N, (IFLOAT *)(A) + ((X) + (Y) * (LDA)) * COMPSIZE, LDA, BUFFER);
 #else
-#define OCOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_OTCOPY(M, N, (FLOAT *)(A) + ((Y) + (X) * (LDA)) * COMPSIZE, LDA, BUFFER);
+#define OCOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_OTCOPY(M, N, (IFLOAT *)(A) + ((Y) + (X) * (LDA)) * COMPSIZE, LDA, BUFFER);
 #endif
 #endif
 
@@ -219,15 +219,16 @@ typedef struct {
 #define STOP_RPCC(COUNTER)
 #endif
 
-static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLOAT *sb, BLASLONG mypos){
+static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, IFLOAT *sa, IFLOAT *sb, BLASLONG mypos){
 
-  FLOAT *buffer[DIVIDE_RATE];
+  IFLOAT *buffer[DIVIDE_RATE];
 
   BLASLONG k, lda, ldb, ldc;
   BLASLONG m_from, m_to, n_from, n_to;
 
   FLOAT *alpha, *beta;
-  FLOAT *a, *b, *c;
+  IFLOAT *a, *b;
+  FLOAT *c;
   job_t *job = (job_t *)args -> common;
 
   BLASLONG nthreads_m;
@@ -255,8 +256,8 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 
   k = K;
 
-  a = (FLOAT *)A;
+  a = (IFLOAT *)A;
-  b = (FLOAT *)B;
+  b = (IFLOAT *)B;
   c = (FLOAT *)C;
 
   lda = LDA;
@@ -425,7 +426,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
           /* Apply kernel with local region of A and part of other region of B */
 	  START_RPCC();
 	  KERNEL_OPERATION(min_i, MIN(range_n[current + 1]  - js,  div_n), min_l, alpha,
-			   sa, (FLOAT *)job[current].working[mypos][CACHE_LINE_SIZE * bufferside],
+			   sa, (IFLOAT *)job[current].working[mypos][CACHE_LINE_SIZE * bufferside],
 			   c, ldc, m_from, js);
           STOP_RPCC(kernel);
 
@@ -469,7 +470,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
           /* Apply kernel with local region of A and part of region of B */
 	  START_RPCC();
 	  KERNEL_OPERATION(min_i, MIN(range_n[current + 1] - js, div_n), min_l, alpha,
-			   sa, (FLOAT *)job[current].working[mypos][CACHE_LINE_SIZE * bufferside],
+			   sa, (IFLOAT *)job[current].working[mypos][CACHE_LINE_SIZE * bufferside],
 			   c, ldc, is, js);
           STOP_RPCC(kernel);
           
@@ -532,7 +533,7 @@ static int round_up(int remainder, int width, int multiple)
 
 
 static int gemm_driver(blas_arg_t *args, BLASLONG *range_m, BLASLONG
-		       *range_n, FLOAT *sa, FLOAT *sb,
+		       *range_n, IFLOAT *sa, IFLOAT *sb,
                        BLASLONG nthreads_m, BLASLONG nthreads_n) {
 
 #ifndef USE_OPENMP
@@ -728,7 +729,7 @@ EnterCriticalSection((PCRITICAL_SECTION)&level3_lock);
   return 0;
 }
 
-int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLOAT *sb, BLASLONG mypos){
+int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, IFLOAT *sa, IFLOAT *sb, BLASLONG mypos){
 
   BLASLONG m = args -> m;
   BLASLONG n = args -> n;

driver/others/blas_server.c

@@ -281,6 +281,8 @@ int get_node(void);
 static int increased_threads = 0;
 
 #ifdef OS_LINUX
+extern int openblas_get_num_threads(void);  
+
 int openblas_setaffinity(int thread_idx, size_t cpusetsize, cpu_set_t* cpu_set) {
   const int active_threads = openblas_get_num_threads();
 
@@ -602,7 +604,7 @@ int blas_thread_init(void){
       if(ret!=0){
 	struct rlimit rlim;
         const char *msg = strerror(ret);
-        fprintf(STDERR, "OpenBLAS blas_thread_init: pthread_create failed for thread %ld of %ld: %s\n", i+1,blas_num_threads,msg);
+        fprintf(STDERR, "OpenBLAS blas_thread_init: pthread_create failed for thread %ld of %d: %s\n", i+1,blas_num_threads,msg);
 #ifdef RLIMIT_NPROC
         if(0 == getrlimit(RLIMIT_NPROC, &rlim)) {
           fprintf(STDERR, "OpenBLAS blas_thread_init: RLIMIT_NPROC "

driver/others/dynamic.c

@@ -332,7 +332,7 @@ int support_avx512(){
   if((ebx & (1<<7)) == 0){
       ret=0;  //OS does not even support AVX2
   }
-  if((ebx & (1<<31)) != 0){
+  if((ebx & (1u<<31)) != 0){
     xgetbv(0, &eax, &edx);
     if((eax & 0xe0) == 0xe0)
       ret=1;  //OS supports AVX512VL
@@ -617,6 +617,18 @@ static gotoblas_t *get_coretype(void){
 	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
+	}
+      case 10:
+    if (model == 5 || model == 6) {
+	  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; //OS doesn't support AVX. Use old kernels.
+	  }
     }
 	return NULL;
       }
@@ -632,7 +644,7 @@ static gotoblas_t *get_coretype(void){
         cpuid(0x80000000, &eax, &ebx, &ecx, &edx);
         if ( (eax & 0xffff)  >= 0x01) {
             cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
-            if ((edx & (1 << 30)) == 0 || (edx & (1 << 31)) == 0)
+            if ((edx & (1 << 30)) == 0 || (edx & (1u << 31)) == 0)
               return NULL;
           }
         else
@@ -764,18 +776,53 @@ char *gotoblas_corename(void) {
   if (gotoblas == &gotoblas_NORTHWOOD)    return corename[ 3];
   if (gotoblas == &gotoblas_PRESCOTT)     return corename[ 4];
   if (gotoblas == &gotoblas_BANIAS)       return corename[ 5];
-  if (gotoblas == &gotoblas_ATOM)         return corename[ 6];
+  if (gotoblas == &gotoblas_ATOM)
+#ifdef DYNAMIC_OLDER
+           return corename[ 6];
+#else
+           return corename[10];
+#endif
   if (gotoblas == &gotoblas_CORE2)        return corename[ 7];
-  if (gotoblas == &gotoblas_PENRYN)       return corename[ 8];
+  if (gotoblas == &gotoblas_PENRYN)
-  if (gotoblas == &gotoblas_DUNNINGTON)   return corename[ 9];
+#ifdef DYNAMIC_OLDER
+           return corename[ 8];
+#else
+           return corename[7];
+#endif
+  if (gotoblas == &gotoblas_DUNNINGTON)
+#ifdef DYNAMIC_OLDER
+           return corename[ 9];
+#else
+           return corename[7];
+#endif
   if (gotoblas == &gotoblas_NEHALEM)      return corename[10];
   if (gotoblas == &gotoblas_ATHLON)       return corename[11];
-  if (gotoblas == &gotoblas_OPTERON_SSE3) return corename[12];
+  if (gotoblas == &gotoblas_OPTERON_SSE3)
-  if (gotoblas == &gotoblas_OPTERON)      return corename[13];
+#ifdef DYNAMIC_OLDER
+           return corename[12];
+#else
+           return corename[7];
+#endif
+  if (gotoblas == &gotoblas_OPTERON)
+#ifdef DYNAMIC_OLDER
+           return corename[13];
+#else
+           return corename[7];
+#endif
   if (gotoblas == &gotoblas_BARCELONA)    return corename[14];
-  if (gotoblas == &gotoblas_NANO)         return corename[15];
+  if (gotoblas == &gotoblas_NANO)
+#ifdef DYNAMIC_OLDER
+           return corename[15];
+#else
+           return corename[10];
+#endif
   if (gotoblas == &gotoblas_SANDYBRIDGE)  return corename[16];
-  if (gotoblas == &gotoblas_BOBCAT)       return corename[17];
+  if (gotoblas == &gotoblas_BOBCAT)
+#ifdef DYNAMIC_OLDER
+           return corename[17];
+#else
+           return corename[7];
+#endif
   if (gotoblas == &gotoblas_BULLDOZER)    return corename[18];
   if (gotoblas == &gotoblas_PILEDRIVER)   return corename[19];
   if (gotoblas == &gotoblas_HASWELL)      return corename[20];
@@ -787,6 +834,7 @@ char *gotoblas_corename(void) {
 }
 
 
+
 static gotoblas_t *force_coretype(char *coretype){
 
 	int i ;

driver/others/dynamic_power.c

@@ -6,6 +6,9 @@ extern gotoblas_t gotoblas_POWER8;
 #if (!defined __GNUC__) || ( __GNUC__ >= 6)
 extern gotoblas_t gotoblas_POWER9;
 #endif
+#if (!defined __GNUC__) || ( __GNUC__ >= 11)
+extern gotoblas_t gotoblas_POWER10;
+#endif
 
 extern void openblas_warning(int verbose, const char *msg);
 
@@ -13,7 +16,8 @@ static char *corename[] = {
 	"unknown",
 	"POWER6",
 	"POWER8",
-	"POWER9"
+	"POWER9",
+	"POWER10"
 };
 
 #define NUM_CORETYPES 4
@@ -23,6 +27,9 @@ char *gotoblas_corename(void) {
 	if (gotoblas == &gotoblas_POWER8)	return corename[2];
 #if (!defined __GNUC__) || ( __GNUC__ >= 6)
 	if (gotoblas == &gotoblas_POWER9)	return corename[3];
+#endif
+#if (!defined __GNUC__) || ( __GNUC__ >= 11)
+	if (gotoblas == &gotoblas_POWER10)	return corename[4];
 #endif
 	return corename[0];
 }
@@ -36,6 +43,10 @@ static gotoblas_t *get_coretype(void) {
 #if (!defined __GNUC__) || ( __GNUC__ >= 6)
 	if (__builtin_cpu_is("power9"))
 		return &gotoblas_POWER9;
+#endif
+#if (!defined __GNUC__) || ( __GNUC__ >= 11)
+	if (__builtin_cpu_is("isa_3_1") && __builtin_cpu_supports ("mma"))
+		return &gotoblas_POWER10;
 #endif
 	return NULL;
 }
@@ -61,6 +72,9 @@ static gotoblas_t *force_coretype(char * coretype) {
 	case  2: return (&gotoblas_POWER8);
 #if (!defined __GNUC__) || ( __GNUC__ >= 6)
 	case  3: return (&gotoblas_POWER9);
+#endif
+#if (!defined __GNUC__) || ( __GNUC__ >= 11)
+	case  4: return (&gotoblas_POWER10);
 #endif
 	default: return NULL;
 	}

driver/others/dynamic_zarch.c

@@ -1,12 +1,58 @@
-
 #include "common.h"
+#include <stdbool.h>
 
+// Gate kernels for z13 and z14 on gcc version
+#if (__GNUC__ == 5 && __GNUC_MINOR__ >= 2) || __GNUC__ >= 6 ||           \
+    /* RHEL 7 since 7.3: */                                              \
+    (__GNUC__ == 4 && __GNUC_MINOR__ == 8 && __GNUC_PATCHLEVEL__ == 5 && \
+     __GNUC_RH_RELEASE__ >= 11)
+#define HAVE_Z13_SUPPORT
+#endif
+
+#if __GNUC__ >= 7
+#define HAVE_Z14_SUPPORT
+#endif
+
+// Guard the use of getauxval() on glibc version >= 2.16
+#ifdef __GLIBC__
+#include <features.h>
+#if __GLIBC_PREREQ(2, 16)
+#include <sys/auxv.h>
+#define HAVE_GETAUXVAL 1
+
+static unsigned long get_hwcap(void)
+{
+	unsigned long hwcap = getauxval(AT_HWCAP);
+	char *maskenv;
+
+	// honor requests for not using specific CPU features in LD_HWCAP_MASK
+	maskenv = getenv("LD_HWCAP_MASK");
+	if (maskenv)
+		hwcap &= strtoul(maskenv, NULL, 0);
+
+	return hwcap;
+	// note that a missing auxval is interpreted as no capabilities
+	// available, which is safe.
+}
+
+#else // __GLIBC_PREREQ(2, 16)
+#warn "Cannot detect SIMD support in Z13 or newer architectures since glibc is older than 2.16"
+
+static unsigned long get_hwcap(void) {
+	// treat missing support for getauxval() as no capabilities available,
+	// which is safe.
+	return 0;
+}
+#endif // __GLIBC_PREREQ(2, 16)
+#endif // __GLIBC
+
+extern gotoblas_t gotoblas_ZARCH_GENERIC;
+#ifdef HAVE_Z13_SUPPORT
 extern gotoblas_t gotoblas_Z13;
+#endif
+#ifdef HAVE_Z14_SUPPORT
 extern gotoblas_t gotoblas_Z14;
-//extern gotoblas_t gotoblas_Z15;
+#endif
-//#if (!defined C_GCC) || (GCC_VERSION >= 60000)
-//extern gotoblas_t gotoblas_Z14;
-//#endif
 
 #define NUM_CORETYPES 4
 
@@ -16,47 +62,50 @@ static char* corename[] = {
 	"unknown",
 	"Z13",
 	"Z14",
-//	"Z15",
 	"ZARCH_GENERIC",
 };
 
 char* gotoblas_corename(void) {
+#ifdef HAVE_Z13_SUPPORT
 	if (gotoblas == &gotoblas_Z13)	return corename[1];
+#endif
+#ifdef HAVE_Z14_SUPPORT
 	if (gotoblas == &gotoblas_Z14)	return corename[2];
-//	if (gotoblas == &gotoblas_Z15)	return corename[3];
+#endif
-//#if (!defined C_GCC) || (GCC_VERSION >= 60000)
+	if (gotoblas == &gotoblas_ZARCH_GENERIC) return corename[3];
-//	if (gotoblas == &gotoblas_POWER9)	return corename[3];
+
-//#endif
+	return corename[0];
-	return corename[0]; // try generic?
 }
 
-// __builtin_cpu_is is not supported by zarch
+/**
+ * Detect the fitting set of kernels by retrieving the CPU features supported by
+ * OS from the auxiliary value AT_HWCAP and choosing the set of kernels
+ * ("coretype") that exploits most of the features and can be compiled with the
+ * available gcc version.
+ * Note that we cannot use vector registers on a z13 or newer unless supported
+ * by the OS kernel (which needs to handle them properly during context switch).
+ */
 static gotoblas_t* get_coretype(void) {
-	FILE* infile;
-	char buffer[512], * p;
 
-	p = (char*)NULL;
+	unsigned long hwcap __attribute__((unused)) = get_hwcap();
-	infile = fopen("/proc/sysinfo", "r");
+
-	while (fgets(buffer, sizeof(buffer), infile)) {
+	// z14 and z15 systems: exploit Vector Facility (SIMD) and
-		if (!strncmp("Type", buffer, 4)) {
+	// Vector-Enhancements Facility 1 (float SIMD instructions), if present.
-			p = strchr(buffer, ':') + 2;
+#ifdef HAVE_Z14_SUPPORT
-#if 0
+	if ((hwcap & HWCAP_S390_VX) && (hwcap & HWCAP_S390_VXE))
-			fprintf(stderr, "%s\n", p);
+		return &gotoblas_Z14;
 #endif
-			break;
-		}
-	}
 
-	fclose(infile);
+	// z13: Vector Facility (SIMD for double)
+#ifdef HAVE_Z13_SUPPORT
+	if (hwcap & HWCAP_S390_VX)
+		return &gotoblas_Z13;
+#endif
 
-	if (strstr(p, "2964")) return &gotoblas_Z13;
+	// fallback in case of missing compiler support, systems before z13, or
-	if (strstr(p, "2965")) return &gotoblas_Z13;
+	// when the OS does not advertise support for the Vector Facility (e.g.,
-	if (strstr(p, "3906")) return &gotoblas_Z14;
+	// missing support in the OS kernel)
-	if (strstr(p, "3907")) return &gotoblas_Z14;
+	return &gotoblas_ZARCH_GENERIC;
-	if (strstr(p, "8561")) return &gotoblas_Z14;        // fallback z15 to z14
-	if (strstr(p, "8562")) return &gotoblas_Z14;        // fallback z15 to z14
-
-	return NULL; // should be ZARCH_GENERIC
 }
 
 static gotoblas_t* force_coretype(char* coretype) {
@@ -76,12 +125,13 @@ static gotoblas_t* force_coretype(char* coretype) {
 
 	switch (found)
 	{
+#ifdef HAVE_Z13_SUPPORT
 	case  1: return (&gotoblas_Z13);
+#endif
+#ifdef HAVE_Z14_SUPPORT
 	case  2: return (&gotoblas_Z14);
-//	case  3: return (&gotoblas_Z15);
+#endif
-//#if (!defined C_GCC) || (GCC_VERSION >= 60000)
+	case  3: return (&gotoblas_ZARCH_GENERIC);
-//	case  3: return (&gotoblas_POWER9);
-//#endif
 	default: return NULL;
 	}
 	snprintf(message, 128, "Core not found: %s\n", coretype);
@@ -109,9 +159,9 @@ void gotoblas_dynamic_init(void) {
 
 	if (gotoblas == NULL)
 	{
-		snprintf(coremsg, 128, "Falling back to Z14 core\n");
+		snprintf(coremsg, 128, "Failed to detect system, falling back to generic z support.\n");
 		openblas_warning(1, coremsg);
-		gotoblas = &gotoblas_Z14;
+		gotoblas = &gotoblas_ZARCH_GENERIC;
 	}
 
 	if (gotoblas && gotoblas->init) {

driver/others/memory.c

@@ -2070,7 +2070,7 @@ if (!release->address) return;
   if (munmap(release -> address, BUFFER_SIZE)) {
       int errsv=errno;
        perror("OpenBLAS : munmap failed:");
-       printf("error code=%d,\trelease->address=%lx\n",errsv,release->address);
+       printf("error code=%d,\trelease->address=%p\n",errsv,release->address);
   }
 }
 

driver/others/parameter.c

@@ -62,6 +62,11 @@ BLASLONG gemm_offset_b = DEFAULT_GEMM_OFFSET_B;
 BLASLONG gemm_offset_b = GEMM_OFFSET_B;
 #endif
 
+#if SHGEMM_P == shgemm_p
+BLASLONG shgemm_p = DEFAULT_GEMM_P;
+#else
+BLASLONG shgemm_p = SHGEMM_P;
+#endif
 #if SGEMM_P == sgemm_p
 BLASLONG sgemm_p = DEFAULT_GEMM_P;
 #else
@@ -83,6 +88,11 @@ BLASLONG zgemm_p = DEFAULT_GEMM_P;
 BLASLONG zgemm_p = ZGEMM_P;
 #endif
 
+#if SHGEMM_Q == shgemm_q
+BLASLONG shgemm_q = DEFAULT_GEMM_Q;
+#else
+BLASLONG shgemm_q = SHGEMM_Q;
+#endif
 #if SGEMM_Q == sgemm_q
 BLASLONG sgemm_q = DEFAULT_GEMM_Q;
 #else
@@ -104,6 +114,11 @@ BLASLONG zgemm_q = DEFAULT_GEMM_Q;
 BLASLONG zgemm_q = ZGEMM_Q;
 #endif
 
+#if SHGEMM_R == shgemm_r
+BLASLONG shgemm_r = DEFAULT_GEMM_R;
+#else
+BLASLONG shgemm_r = SHGEMM_R;
+#endif
 #if SGEMM_R == sgemm_r
 BLASLONG sgemm_r = DEFAULT_GEMM_R;
 #else
@@ -597,6 +612,7 @@ void blas_set_parameter(void){
 
   size = BITMASK(cpuid3, 16, 0xff);
 
+  shgemm_p = 192 * (size + 1);
   sgemm_p = 192 * (size + 1);
   dgemm_p =  96 * (size + 1);
   cgemm_p =  96 * (size + 1);
@@ -610,6 +626,7 @@ void blas_set_parameter(void){
   xgemm_p =  16 * (size + 1);
 #endif
 
+  shgemm_r = (((BUFFER_SIZE - ((SHGEMM_P * SHGEMM_Q *  4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SHGEMM_Q *  4)) - 15) & ~15;
   sgemm_r = (((BUFFER_SIZE - ((SGEMM_P * SGEMM_Q *  4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SGEMM_Q *  4)) - 15) & ~15;
   dgemm_r = (((BUFFER_SIZE - ((DGEMM_P * DGEMM_Q *  8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (DGEMM_Q *  8)) - 15) & ~15;
   cgemm_r = (((BUFFER_SIZE - ((CGEMM_P * CGEMM_Q *  8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (CGEMM_Q *  8)) - 15) & ~15;

exports/Makefile

@@ -30,6 +30,10 @@ ifndef BUILD_LAPACK_DEPRECATED
 BUILD_LAPACK_DEPRECATED = 0
 endif
 
+ifndef BUILD_HALF
+BUILD_HALF = 0
+endif
+
 ifeq ($(OSNAME), WINNT)
 ifeq ($(F_COMPILER), GFORTRAN)
 ifndef ONLY_CBLAS
@@ -151,8 +155,12 @@ ifeq ($(F_COMPILER), INTEL)
 	-Wl,--whole-archive $< -Wl,--no-whole-archive \
 	-Wl,-soname,$(INTERNALNAME) $(EXTRALIB)
 	$(CC) $(CFLAGS) $(LDFLAGS) -w -o linktest linktest.c ../$(LIBSONAME) $(FEXTRALIB) && echo OK.
+else ifeq ($(F_COMPILER), FLANG)
+	$(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 $< -Wl,--no-whole-archive \
@@ -234,23 +242,23 @@ static : ../$(LIBNAME)
 	rm -f goto.$(SUFFIX)
 
 osx.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol osx $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
+	perl ./gensymbol osx $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) $(BUILD_HALF) > $(@F)
 
 aix.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol aix $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
+	perl ./gensymbol aix $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) $(BUILD_HALF) > $(@F)
 
 objcopy.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol objcopy $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
+	perl ./gensymbol objcopy $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) $(BUILD_HALF) > $(@F)
 
 objconv.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol objconv $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
+	perl ./gensymbol objconv $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) $(BUILD_HALF) > $(@F)
 
 test : linktest.c
 	$(CC) $(CFLAGS) $(LDFLAGS) -w -o linktest linktest.c ../$(LIBSONAME) -lm && echo OK.
 	rm -f linktest
 
 linktest.c : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol linktest  $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > linktest.c
+	perl ./gensymbol linktest  $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) $(BUILD_HALF) > linktest.c
 
 clean ::
 	@rm -f *.def *.dylib __.SYMDEF* *.renamed

exports/gensymbol

@@ -40,17 +40,13 @@
     ztbsv,ztpmv,ztpsv,ztrmm,ztrmv,ztrsm,ztrsv,
     xerbla,
     saxpby,daxpby,caxpby,zaxpby,
+    somatcopy, domatcopy, comatcopy, zomatcopy,
+    simatcopy, dimatcopy, cimatcopy, zimatcopy,
     sgeadd,dgeadd,cgeadd,zgeadd,
-    somatcopy,
+    ssum, dsum, scsum, dzsum
-    simatcopy,
-    domatcopy,
-    dimatcopy,
-    comatcopy,
-    cimatcopy,
-    zomatcopy,
-    zimatcopy,
 );
 
+@halfblasobjs = (shgemm);
 @cblasobjs = (
     cblas_caxpy, cblas_ccopy, cblas_cdotc, cblas_cdotu, cblas_cgbmv, cblas_cgemm, cblas_cgemv,
     cblas_cgerc, cblas_cgeru, cblas_chbmv, cblas_chemm, cblas_chemv, cblas_cher2, cblas_cher2k,
@@ -80,9 +76,16 @@
     cblas_saxpby,cblas_daxpby,cblas_caxpby,cblas_zaxpby,
     cblas_somatcopy, cblas_domatcopy, cblas_comatcopy, cblas_zomatcopy,
     cblas_simatcopy, cblas_dimatcopy, cblas_cimatcopy, cblas_zimatcopy,
-    cblas_sgeadd, cblas_dgeadd,cblas_cgeadd, cblas_zgeadd
+    cblas_sgeadd, cblas_dgeadd,cblas_cgeadd, cblas_zgeadd,
+    cblas_isamin, cblas_idamin, cblas_icamin, cblas_izamin,
+    cblas_ismin, cblas_idmin, cblas_icmin, cblas_izmin,
+    cblas_ismax, cblas_idmax, cblas_icmax, cblas_izmax,
+    cblas_ssum, cblas_dsum, cblas_scsum, cblas_dzsum,
+    cblas_xerbla
 );
 
+@halfcblasobjs = (cblas_shgemm);
+
 @exblasobjs = (
     qamax,qamin,qasum,qaxpy,qcabs1,qcopy,qdot,qgbmv,qgemm,
     qgemv,qger,qmax,qmin,
@@ -3454,6 +3457,10 @@ use File::Spec;
 use File::Basename;
 my $dirname = File::Spec->catfile(dirname(dirname(File::Spec->rel2abs(__FILE__))), "lapack-netlib");
 
+if ($ARGV[12] == 1) {
+	@blasobjs = (@blasobjs, @halfblasobjs);
+	@cblasobjs = (@cblasobjs, @halfcblasobjs);
+}
 if ($ARGV[8] == 1) {
     #ONLY_CBLAS=1
     @underscore_objs = (@misc_underscore_objs);
@@ -3494,9 +3501,12 @@ if ($ARGV[1] eq "x86")    { @underscore_objs = (@underscore_objs, @gemm3mobjs);
 if ($ARGV[1] eq "ia64")   { @underscore_objs = (@underscore_objs, @gemm3mobjs); };
 if ($ARGV[1] eq "MIPS")   { @underscore_objs = (@underscore_objs, @gemm3mobjs); };
 
-
 if ($ARGV[4] == 0) {
     @no_underscore_objs = (@cblasobjs, @misc_no_underscore_objs);
+    if ($ARGV[1] eq "x86_64") { @no_underscore_objs = (@no_underscore_objs, @cblasgemm3mobjs); };
+    if ($ARGV[1] eq "x86")    { @no_underscore_objs = (@no_underscore_objs, @cblasgemm3mobjs); };
+    if ($ARGV[1] eq "ia64")   { @no_underscore_objs = (@no_underscore_objs, @cblasgemm3mobjs); };
+    if ($ARGV[1] eq "MIPS")   { @no_underscore_objs = (@no_underscore_objs, @cblasgemm3mobjs); };
 }else{
     #NO_CBLAS=1
     @no_underscore_objs = (@misc_no_underscore_objs);

f_check

@@ -334,7 +334,8 @@ if ($link ne "") {
 	    && ($flags !~ /kernel32/)
 	    && ($flags !~ /advapi32/)
 	    && ($flags !~ /shell32/)
-	    && ($flags !~ /omp/)
+	    && ($flags !~ /omp/ || ($vendor !~ /PGI/ && $flags =~ /omp/))
+	    && ($flags !~ /[0-9]+/)
 		&& ($flags !~ /^\-l$/)
 	    ) {
 	    $linker_l .= $flags . " ";

getarch.c

@@ -650,6 +650,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CORENAME  "POWER9"
 #endif
 
+#if defined(FORCE_POWER10)
+#define FORCE
+#define ARCHITECTURE    "POWER"
+#define SUBARCHITECTURE "POWER10"
+#define SUBDIRNAME      "power"
+#define ARCHCONFIG   "-DPOWER10 " \
+		     "-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   "power10"
+#define CORENAME  "POWER10"
+#endif
+
 #ifdef FORCE_PPCG4
 #define FORCE
 #define ARCHITECTURE    "POWER"
@@ -812,6 +825,34 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #else
 #endif
 
+#ifdef FORCE_MIPS1004K
+#define FORCE
+#define ARCHITECTURE    "MIPS"
+#define SUBARCHITECTURE "MIPS1004K"
+#define SUBDIRNAME      "mips"
+#define ARCHCONFIG   "-DMIPS1004K " \
+       "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=32 " \
+       "-DL2_SIZE=262144 -DL2_LINESIZE=32 " \
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
+#define LIBNAME   "mips1004K"
+#define CORENAME  "MIPS1004K"
+#else
+#endif
+
+#ifdef FORCE_MIPS24K
+#define FORCE
+#define ARCHITECTURE    "MIPS"
+#define SUBARCHITECTURE "MIPS24K"
+#define SUBDIRNAME      "mips"
+#define ARCHCONFIG   "-DMIPS24K " \
+       "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=32 " \
+       "-DL2_SIZE=32768 -DL2_LINESIZE=32 " \
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
+#define LIBNAME   "mips24K"
+#define CORENAME  "MIPS24K"
+#else
+#endif
+
 #ifdef FORCE_I6500
 #define FORCE
 #define ARCHITECTURE    "MIPS"
@@ -1334,10 +1375,12 @@ int main(int argc, char *argv[]){
 
 #if defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
 printf("__BYTE_ORDER__=__ORDER_BIG_ENDIAN__\n");
-#endif
+#elif defined(__BIG_ENDIAN__) && __BIG_ENDIAN__ > 0
-#if defined(__BIG_ENDIAN__) && __BIG_ENDIAN__ > 0
 printf("__BYTE_ORDER__=__ORDER_BIG_ENDIAN__\n");
 #endif
+#if defined(_CALL_ELF) && (_CALL_ELF == 2)
+printf("ELF_VERSION=2\n");
+#endif
 
 #ifdef MAKE_NB_JOBS
   #if MAKE_NB_JOBS > 0

getarch_2nd.c

@@ -9,6 +9,8 @@
 int main(int argc, char **argv) {
 
   if ( (argc <= 1) || ((argc >= 2) && (*argv[1] == '0'))) {
+    printf("SHGEMM_UNROLL_M=%d\n", SHGEMM_DEFAULT_UNROLL_M);
+    printf("SHGEMM_UNROLL_N=%d\n", SHGEMM_DEFAULT_UNROLL_N);
     printf("SGEMM_UNROLL_M=%d\n", SGEMM_DEFAULT_UNROLL_M);
     printf("SGEMM_UNROLL_N=%d\n", SGEMM_DEFAULT_UNROLL_N);
     printf("DGEMM_UNROLL_M=%d\n", DGEMM_DEFAULT_UNROLL_M);

interface/CMakeLists.txt

@@ -115,7 +115,7 @@ foreach (float_type ${FLOAT_TYPES})
     GenerateNamedObjects("syr2k.c" "HEMM" "her2k" ${CBLAS_FLAG} "" "" false ${float_type})
 
     if (USE_GEMM3M)
-      GenerateNamedObjects("gemm.c" "GEMM3M" "gemm3m" false "" "" false ${float_type})
+      GenerateNamedObjects("gemm.c" "GEMM3M" "gemm3m" ${CBLAS_FLAG} "" "" false ${float_type})
     endif()
   endif ()
   if (${float_type} STREQUAL "COMPLEX")

interface/Makefile

@@ -46,6 +46,9 @@ SBLAS3OBJS    = \
 		somatcopy.$(SUFFIX) simatcopy.$(SUFFIX)\
 		sgeadd.$(SUFFIX)
 
+ifeq ($(BUILD_HALF),1)
+SHBLAS3OBJS    = shgemm.$(SUFFIX)
+endif
 
 DBLAS1OBJS    = \
 		daxpy.$(SUFFIX) dswap.$(SUFFIX) \
@@ -277,6 +280,10 @@ CSBLAS3OBJS   = \
 	cblas_ssyrk.$(SUFFIX) cblas_ssyr2k.$(SUFFIX) cblas_somatcopy.$(SUFFIX)  cblas_simatcopy.$(SUFFIX)\
 	cblas_sgeadd.$(SUFFIX)
 
+ifeq ($(BUILD_HALF),1)
+CSHBLAS3OBJS = cblas_shgemm.$(SUFFIX)
+endif
+
 CDBLAS1OBJS   = \
 	cblas_idamax.$(SUFFIX) cblas_idamin.$(SUFFIX) cblas_dasum.$(SUFFIX) cblas_daxpy.$(SUFFIX) \
 	cblas_dcopy.$(SUFFIX) cblas_ddot.$(SUFFIX) \
@@ -367,6 +374,7 @@ override CFLAGS += -I.
 SBLAS1OBJS   += $(CSBLAS1OBJS)
 SBLAS2OBJS   += $(CSBLAS2OBJS)
 SBLAS3OBJS   += $(CSBLAS3OBJS)
+SHBLAS3OBJS  += $(CSHBLAS3OBJS)
 DBLAS1OBJS   += $(CDBLAS1OBJS)
 DBLAS2OBJS   += $(CDBLAS2OBJS)
 DBLAS3OBJS   += $(CDBLAS3OBJS)
@@ -380,6 +388,7 @@ ZBLAS3OBJS   += $(CZBLAS3OBJS)
 endif
 
 SBLASOBJS    = $(SBLAS1OBJS) $(SBLAS2OBJS) $(SBLAS3OBJS)
+SHBLASOBJS   = $(SHBLAS3OBJS)
 DBLASOBJS    = $(DBLAS1OBJS) $(DBLAS2OBJS) $(DBLAS3OBJS)
 QBLASOBJS    = $(QBLAS1OBJS) $(QBLAS2OBJS) $(QBLAS3OBJS)
 CBLASOBJS    = $(CBLAS1OBJS) $(CBLAS2OBJS) $(CBLAS3OBJS)
@@ -454,7 +463,7 @@ ZBLASOBJS += $(ZLAPACKOBJS)
 
 endif
 
-FUNCOBJS    = $(SBLASOBJS) $(DBLASOBJS) $(CBLASOBJS) $(ZBLASOBJS)
+FUNCOBJS    = $(SHBLASOBJS) $(SBLASOBJS) $(DBLASOBJS) $(CBLASOBJS) $(ZBLASOBJS)
 
 ifdef EXPRECISION
 FUNCOBJS   += $(QBLASOBJS) $(XBLASOBJS)
@@ -488,10 +497,10 @@ level1 : $(SBLAS1OBJS) $(DBLAS1OBJS) $(QBLAS1OBJS) $(CBLAS1OBJS) $(ZBLAS1OBJS) $
 level2 : $(SBLAS2OBJS) $(DBLAS2OBJS) $(QBLAS2OBJS) $(CBLAS2OBJS) $(ZBLAS2OBJS) $(XBLAS2OBJS)
 	$(AR) $(ARFLAGS) -ru $(TOPDIR)/$(LIBNAME) $^
 
-level3 : $(SBLAS3OBJS) $(DBLAS3OBJS) $(QBLAS3OBJS) $(CBLAS3OBJS) $(ZBLAS3OBJS) $(XBLAS3OBJS)
+level3 : $(SHBLAS3OBJS) $(SBLAS3OBJS) $(DBLAS3OBJS) $(QBLAS3OBJS) $(CBLAS3OBJS) $(ZBLAS3OBJS) $(XBLAS3OBJS)
 	$(AR) $(ARFLAGS) -ru $(TOPDIR)/$(LIBNAME) $^
 
-$(CSBLASOBJS) $(CSBLASOBJS_P) $(CDBLASOBJS) $(CDBLASOBJS_P) $(CQBLASOBJS) $(CQBLASOBJS_P) \
+$(CSHBLASOBJS) $(CSHBLASOBJS_P) $(CSBLASOBJS) $(CSBLASOBJS_P) $(CDBLASOBJS) $(CDBLASOBJS_P) $(CQBLASOBJS) $(CQBLASOBJS_P) \
 $(CCBLASOBJS) $(CCBLASOBJS_P) $(CZBLASOBJS) $(CZBLASOBJS_P) $(CXBLASOBJS) $(CXBLASOBJS_P) : override CFLAGS += -DCBLAS
 
 srot.$(SUFFIX) srot.$(PSUFFIX) : rot.c
@@ -1209,6 +1218,11 @@ zhpr2.$(SUFFIX) zhpr2.$(PSUFFIX) : zhpr2.c
 xhpr2.$(SUFFIX) xhpr2.$(PSUFFIX) : zhpr2.c
 	$(CC) -c $(CFLAGS) $< -o $(@F)
 
+ifeq ($(BUILD_HALF),1)
+shgemm.$(SUFFIX) shgemm.$(PSUFFIX) : gemm.c ../param.h
+	$(CC) -c $(CFLAGS) $< -o $(@F)
+endif
+
 sgemm.$(SUFFIX) sgemm.$(PSUFFIX) : gemm.c ../param.h
 	$(CC) -c $(CFLAGS) $< -o $(@F)
 
@@ -1770,6 +1784,11 @@ cblas_zhemv.$(SUFFIX) cblas_zhemv.$(PSUFFIX) : zhemv.c
 cblas_sgemm.$(SUFFIX) cblas_sgemm.$(PSUFFIX) : gemm.c ../param.h
 	$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
 
+ifeq ($(BUILD_HALF),1)
+cblas_shgemm.$(SUFFIX) cblas_shgemm.$(PSUFFIX) : gemm.c ../param.h
+	$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
+endif
+
 cblas_dgemm.$(SUFFIX) cblas_dgemm.$(PSUFFIX) : gemm.c ../param.h
 	$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
 

interface/gemm.c

@@ -77,7 +77,7 @@
 #define GEMM_MULTITHREAD_THRESHOLD 4
 #endif
 
-static int (*gemm[])(blas_arg_t *, BLASLONG *, BLASLONG *, FLOAT *, FLOAT *, BLASLONG) = {
+static int (*gemm[])(blas_arg_t *, BLASLONG *, BLASLONG *, IFLOAT *, IFLOAT *, BLASLONG) = {
 #ifndef GEMM3M
   GEMM_NN, GEMM_TN, GEMM_RN, GEMM_CN,
   GEMM_NT, GEMM_TT, GEMM_RT, GEMM_CT,
@@ -108,8 +108,8 @@ static int (*gemm[])(blas_arg_t *, BLASLONG *, BLASLONG *, FLOAT *, FLOAT *, BLA
 void NAME(char *TRANSA, char *TRANSB,
 	  blasint *M, blasint *N, blasint *K,
 	  FLOAT *alpha,
-	  FLOAT *a, blasint *ldA,
+	  IFLOAT *a, blasint *ldA,
-	  FLOAT *b, blasint *ldB,
+	  IFLOAT *b, blasint *ldB,
 	  FLOAT *beta,
 	  FLOAT *c, blasint *ldC){
 
@@ -119,8 +119,8 @@ void NAME(char *TRANSA, char *TRANSB,
   blasint info;
 
   char transA, transB;
-  FLOAT *buffer;
+  IFLOAT *buffer;
-  FLOAT *sa, *sb;
+  IFLOAT *sa, *sb;
 
 #ifdef SMP
   double MNK;

kernel/CMakeLists.txt

@@ -41,6 +41,9 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
     foreach (float_type ${FLOAT_TYPES})
       # a bit of metaprogramming here to pull out the appropriate KERNEL var
       string(SUBSTRING ${float_type} 0 1 float_char)
+      if (${float_type} STREQUAL "HALF")
+	set (float_char "SH")
+      endif ()
       GenerateNamedObjects("${KERNELDIR}/${${float_char}AMAXKERNEL}" "USE_ABS" "amax_k" false "" "" false ${float_type})
       GenerateNamedObjects("${KERNELDIR}/${${float_char}AMINKERNEL}" "USE_ABS;USE_MIN" "amin_k" false "" "" false ${float_type})
       if (DEFINED ${float_char}MAXKERNEL)
@@ -93,6 +96,9 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
     GenerateNamedObjects("generic/ger.c" "" "ger_k" false "" "" "" 3)
     foreach (float_type ${FLOAT_TYPES})
       string(SUBSTRING ${float_type} 0 1 float_char)
+      if (${float_type} STREQUAL "HALF")
+	set (float_char "SH")
+      endif ()
       if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
         GenerateNamedObjects("${KERNELDIR}/${${float_char}GERUKERNEL}" "" "geru_k" false "" "" false ${float_type})
         GenerateNamedObjects("${KERNELDIR}/${${float_char}GERCKERNEL}" "CONJ" "gerc_k" false "" "" false ${float_type})
@@ -124,17 +130,27 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
     if (ARM OR ARM64 OR (TARGET_CORE MATCHES LONGSOON3B) OR (TARGET_CORE MATCHES GENERIC) OR (TARGET_CORE MATCHES HASWELL) OR (TARGET_CORE MATCHES ZEN) OR (TARGET_CORE MATCHES SKYLAKEX) )
       set(USE_TRMM true)
     endif ()
-    if (ZARCH OR (TARGET_CORE MATCHES POWER8) OR (TARGET_CORE MATCHES POWER9)) 
+    if (ZARCH OR (TARGET_CORE MATCHES POWER8) OR (TARGET_CORE MATCHES POWER9) OR (TARGET_CORE MATCHES POWER10))
       set(USE_TRMM true)
     endif ()
 
-    foreach (float_type SINGLE DOUBLE)
+    foreach (float_type SINGLE DOUBLE HALF)
       string(SUBSTRING ${float_type} 0 1 float_char)
+      if (${float_type} STREQUAL "HALF")
+        if (NOT ${BUILD_HALF})
+	  continue ()
+        else ()
+	  set (float_char "SH")
+      endif ()
+      endif ()
       GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMKERNEL}" "" "gemm_kernel" false "" "" false ${float_type})
     endforeach()
 
     foreach (float_type ${FLOAT_TYPES})
       string(SUBSTRING ${float_type} 0 1 float_char)
+      if (${float_type} STREQUAL "HALF")
+	set (float_char "SH")
+      endif ()
       if (${float_char}GEMMINCOPY)
         GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMINCOPY}" "${float_type}" "${${float_char}GEMMINCOPYOBJ}" false "" "" true ${float_type})
       endif ()
@@ -470,9 +486,13 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
       GenerateNamedObjects("${KERNELDIR}/${${float_char}GEADD_KERNEL}" "" "geadd_k" false "" "" false ${float_type})
     endforeach ()
 
+
     # Makefile.LA
     if(NOT NO_LAPACK)
       foreach (float_type ${FLOAT_TYPES})
+      if (${float_type} STREQUAL "HALF")
+	set (float_char "SH")
+      endif ()
         if (NOT DEFINED ${float_char}NEG_TCOPY)
           if (${float_char} STREQUAL "Z" OR ${float_char} STREQUAL "C" OR ${float_char} STREQUAL "X")
             set(${float_char}NEG_TCOPY ../generic/zneg_tcopy.c)
@@ -516,6 +536,9 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
     foreach (float_type ${FLOAT_TYPES})
       # a bit of metaprogramming here to pull out the appropriate KERNEL var
       string(SUBSTRING ${float_type} 0 1 float_char)
+      if (${float_type} STREQUAL "HALF")
+	set (float_char "SH")
+      endif ()
       GenerateNamedObjects("generic/neg_tcopy_${${float_char}GEMM_UNROLL_M}.c" "" "neg_tcopy" false "" ${TSUFFIX}  false ${float_type})
       GenerateNamedObjects("generic/laswp_ncopy_${${float_char}GEMM_UNROLL_N}.c" "" "laswp_ncopy" false "" ${TSUFFIX}  false ${float_type})
     endforeach ()

kernel/Makefile.L3

@@ -51,6 +51,10 @@ ifeq ($(CORE), POWER9)
 USE_TRMM = 1
 endif
 
+ifeq ($(CORE), POWER10)
+USE_TRMM = 1
+endif
+
 ifeq ($(ARCH), zarch)
 USE_TRMM = 1
 endif
@@ -59,6 +63,25 @@ ifeq ($(CORE), Z14)
 USE_TRMM = 1
 endif
 
+ifeq ($(BUILD_HALF), 1)
+ifndef SHGEMMKERNEL
+SHGEMM_BETA = ../generic/gemm_beta.c
+SHGEMMKERNEL    = ../generic/gemmkernel_2x2.c
+SHGEMMINCOPY    = ../generic/gemm_ncopy_2.c
+SHGEMMITCOPY    = ../generic/gemm_tcopy_2.c
+SHGEMMONCOPY    = ../generic/gemm_ncopy_2.c
+SHGEMMOTCOPY    = ../generic/gemm_tcopy_2.c
+SHGEMMINCOPYOBJ =  shgemm_incopy$(TSUFFIX).$(SUFFIX)
+SHGEMMITCOPYOBJ =  shgemm_itcopy$(TSUFFIX).$(SUFFIX)
+SHGEMMONCOPYOBJ =  shgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SHGEMMOTCOPYOBJ =  shgemm_otcopy$(TSUFFIX).$(SUFFIX)
+endif
+
+SHKERNELOBJS	+= \
+	shgemm_kernel$(TSUFFIX).$(SUFFIX) \
+	$(SHGEMMINCOPYOBJ) $(SHGEMMITCOPYOBJ) \
+	$(SHGEMMONCOPYOBJ) $(SHGEMMOTCOPYOBJ)
+endif
 
 SKERNELOBJS	+= \
 	sgemm_kernel$(TSUFFIX).$(SUFFIX) \
@@ -93,6 +116,9 @@ XKERNELOBJS	+= \
 	$(XGEMMINCOPYOBJ) $(XGEMMITCOPYOBJ) \
 	$(XGEMMONCOPYOBJ) $(XGEMMOTCOPYOBJ)
 
+ifeq ($(BUILD_HALF),1)
+SHBLASOBJS      += $(SHKERNELOBJS)
+endif
 SBLASOBJS	+= $(SKERNELOBJS)
 DBLASOBJS	+= $(DKERNELOBJS)
 QBLASOBJS	+= $(QKERNELOBJS)
@@ -100,6 +126,10 @@ CBLASOBJS	+= $(CKERNELOBJS)
 ZBLASOBJS	+= $(ZKERNELOBJS)
 XBLASOBJS	+= $(XKERNELOBJS)
 
+ifeq ($(BUILD_HALF),1)
+SHBLASOBJS += shgemm_beta$(TSUFFIX).$(SUFFIX)
+endif
+
 SBLASOBJS	+= \
 	sgemm_beta$(TSUFFIX).$(SUFFIX) \
 	strmm_kernel_LN$(TSUFFIX).$(SUFFIX) strmm_kernel_LT$(TSUFFIX).$(SUFFIX) \
@@ -389,6 +419,12 @@ ZBLASOBJS += \
 	zimatcopy_k_ctc$(TSUFFIX).$(SUFFIX) zimatcopy_k_rtc$(TSUFFIX).$(SUFFIX) \
 	zgeadd_k$(TSUFFIX).$(SUFFIX) 
 
+ifeq ($(BUILD_HALF), 1)
+SHGEMMINCOPYOBJ_P = $(SHGEMMINCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
+SHGEMMITCOPYOBJ_P = $(SHGEMMITCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
+SHGEMMONCOPYOBJ_P = $(SHGEMMONCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
+SHGEMMOTCOPYOBJ_P = $(SHGEMMOTCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
+endif
 
 SGEMMINCOPYOBJ_P = $(SGEMMINCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 SGEMMITCOPYOBJ_P = $(SGEMMITCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
@@ -415,6 +451,11 @@ XGEMMITCOPYOBJ_P = $(XGEMMITCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 XGEMMONCOPYOBJ_P = $(XGEMMONCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 XGEMMOTCOPYOBJ_P = $(XGEMMOTCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 
+ifeq ($(BUILD_HALF),1)
+$(KDIR)shgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SHGEMM_BETA)
+	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+endif
+
 $(KDIR)sgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_BETA)
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 
@@ -433,12 +474,47 @@ $(KDIR)zgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMM_BETA)
 $(KDIR)xgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(XGEMM_BETA)
 	$(CC) $(CFLAGS) -c -DXDOUBLE -DCOMPLEX $< -o $@
 
+
+ifeq ($(BUILD_HALF), 1)
+
+$(KDIR)$(SHGEMMONCOPYOBJ) : $(KERNELDIR)/$(SHGEMMONCOPY)
+	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+
+$(KDIR)$(SHGEMMOTCOPYOBJ) : $(KERNELDIR)/$(SHGEMMOTCOPY)
+
+ifeq ($(OS), AIX)
+	$(CC) $(CFLAGS) -S -DHALF -UDOUBLE -UCOMPLEX $< -o - > shgemmotcopy.s
+	m4 shgemmotcopy.s > shgemmotcopy_nomacros.s
+	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX shgemmotcopy_nomacros.s -o $@
+	rm shgemmotcopy.s shgemmotcopy_nomacros.s
+else
+	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+endif
+
+ifneq ($(SHGEMM_UNROLL_M), $(SHGEMM_UNROLL_N))
+
+$(KDIR)$(SHGEMMINCOPYOBJ) : $(KERNELDIR)/$(SHGEMMINCOPY)
+	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+
+$(KDIR)$(SHGEMMITCOPYOBJ) : $(KERNELDIR)/$(SHGEMMITCOPY)
+ifeq ($(OS), AIX)
+	$(CC) $(CFLAGS) -S -DHALF -UDOUBLE -UCOMPLEX $< -o - > shgemmitcopy.s
+	m4 shgemmitcopy.s > shgemmitcopy_nomacros.s
+	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX shgemmitcopy_nomacros.s -o $@
+	rm shgemmitcopy.s shgemmitcopy_nomacros.s
+else
+	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+endif
+
+endif
+endif
+
 $(KDIR)$(SGEMMONCOPYOBJ) : $(KERNELDIR)/$(SGEMMONCOPY)
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 
 $(KDIR)$(SGEMMOTCOPYOBJ) : $(KERNELDIR)/$(SGEMMOTCOPY)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -UDOUBLE -UCOMPLEX $< -o sgemmotcopy.s	
+	$(CC) $(CFLAGS) -S -UDOUBLE -UCOMPLEX $< -o - > sgemmotcopy.s	
 	m4 sgemmotcopy.s > sgemmotcopy_nomacros.s
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX sgemmotcopy_nomacros.s -o $@
 	rm sgemmotcopy.s sgemmotcopy_nomacros.s
@@ -454,7 +530,7 @@ $(KDIR)$(SGEMMINCOPYOBJ) : $(KERNELDIR)/$(SGEMMINCOPY)
 
 $(KDIR)$(SGEMMITCOPYOBJ) : $(KERNELDIR)/$(SGEMMITCOPY)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -UDOUBLE -UCOMPLEX $< -o sgemmitcopy.s	
+	$(CC) $(CFLAGS) -S -UDOUBLE -UCOMPLEX $< -o - > sgemmitcopy.s	
 	m4 sgemmitcopy.s > sgemmitcopy_nomacros.s
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX sgemmitcopy_nomacros.s -o $@
 	rm sgemmitcopy.s sgemmitcopy_nomacros.s
@@ -466,7 +542,7 @@ endif
 
 $(KDIR)$(DGEMMONCOPYOBJ) : $(KERNELDIR)/$(DGEMMONCOPY)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DDOUBLE -UCOMPLEX $< -o dgemm_ncopy.s
+	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > dgemm_ncopy.s
 	m4 dgemm_ncopy.s > dgemm_ncopy_nomacros.s
 	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX dgemm_ncopy_nomacros.s -o $@
 	rm dgemm_ncopy.s dgemm_ncopy_nomacros.s
@@ -484,7 +560,7 @@ $(KDIR)$(DGEMMINCOPYOBJ) : $(KERNELDIR)/$(DGEMMINCOPY)
 
 $(KDIR)$(DGEMMITCOPYOBJ) : $(KERNELDIR)/$(DGEMMITCOPY)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DDOUBLE -UCOMPLEX $< -o dgemm_itcopy.s
+	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > dgemm_itcopy.s
 	m4 dgemm_itcopy.s > dgemm_itcopy_nomacros.s
 	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX dgemm_itcopy_nomacros.s -o $@
 	rm dgemm_itcopy.s dgemm_itcopy_nomacros.s
@@ -527,7 +603,7 @@ $(KDIR)$(CGEMMINCOPYOBJ) : $(KERNELDIR)/$(CGEMMINCOPY)
 
 $(KDIR)$(CGEMMITCOPYOBJ) : $(KERNELDIR)/$(CGEMMITCOPY)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -UDOUBLE -UCOMPLEX -E $< -o cgemm_itcopy.s
+	$(CC) $(CFLAGS) -UDOUBLE -UCOMPLEX -S $< -o - > cgemm_itcopy.s
 	m4 cgemm_itcopy.s > cgemm_itcopy_nomacros.s
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX cgemm_itcopy_nomacros.s -o $@
 	rm cgemm_itcopy.s cgemm_itcopy_nomacros.s
@@ -550,7 +626,7 @@ $(KDIR)$(ZGEMMINCOPYOBJ) : $(KERNELDIR)/$(ZGEMMINCOPY)
 
 $(KDIR)$(ZGEMMITCOPYOBJ) : $(KERNELDIR)/$(ZGEMMITCOPY)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DDOUBLE -UCOMPLEX $< -o zgemm_itcopy.s
+	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > zgemm_itcopy.s
 	m4 zgemm_itcopy.s > zgemm_itcopy_nomacros.s
 	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX zgemm_itcopy_nomacros.s -o $@
 	rm zgemm_itcopy.s zgemm_itcopy_nomacros.s
@@ -582,7 +658,7 @@ endif
 
 $(KDIR)sgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL) $(SGEMMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -UDOUBLE -UCOMPLEX  $< -o sgemm_kernel$(TSUFFIX).s
+	$(CC) $(CFLAGS) -S -UDOUBLE -UCOMPLEX  $< -o - > sgemm_kernel$(TSUFFIX).s
 	m4 sgemm_kernel$(TSUFFIX).s > sgemm_kernel$(TSUFFIX)_nomacros.s
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX sgemm_kernel$(TSUFFIX)_nomacros.s -o $@
 	rm sgemm_kernel$(TSUFFIX).s sgemm_kernel$(TSUFFIX)_nomacros.s
@@ -590,9 +666,22 @@ else
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 endif
 
+ifeq ($(BUILD_HALF), 1)
+
+$(KDIR)shgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SHGEMMKERNEL) $(SHGEMMDEPEND)
+ifeq ($(OS), AIX)
+	$(CC) $(CFLAGS) -S -DHALF -UDOUBLE -UCOMPLEX  $< -o - > shgemm_kernel$(TSUFFIX).s
+	m4 shgemm_kernel$(TSUFFIX).s > shgemm_kernel$(TSUFFIX)_nomacros.s
+	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX shgemm_kernel$(TSUFFIX)_nomacros.s -o $@
+	rm shgemm_kernel$(TSUFFIX).s shgemm_kernel$(TSUFFIX)_nomacros.s
+else
+	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+endif
+endif
+
 $(KDIR)dgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMMKERNEL) $(DGEMMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DDOUBLE -UCOMPLEX $< -o dgemm_kernel$(TSUFFIX).s
+	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > dgemm_kernel$(TSUFFIX).s
 	m4 dgemm_kernel$(TSUFFIX).s > dgemm_kernel$(TSUFFIX)_nomacros.s
 	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX dgemm_kernel$(TSUFFIX)_nomacros.s -o $@
 	rm dgemm_kernel$(TSUFFIX).s dgemm_kernel$(TSUFFIX)_nomacros.s
@@ -605,7 +694,7 @@ $(KDIR)qgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(QGEMMKERNEL) $(QGEMMDEP
 
 $(KDIR)cgemm_kernel_n$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -UDOUBLE -DCOMPLEX -DNN $< -o cgemm_kernel_n.s
+	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DNN $< -o - > cgemm_kernel_n.s
 	m4 cgemm_kernel_n.s > cgemm_kernel_n_nomacros.s
 	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DNN cgemm_kernel_n_nomacros.s -o $@
 	rm cgemm_kernel_n.s cgemm_kernel_n_nomacros.s
@@ -615,7 +704,7 @@ endif
 
 $(KDIR)cgemm_kernel_l$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -UDOUBLE -DCOMPLEX -DCN $< -o cgemm_kernel_l.s
+	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DCN $< -o - > cgemm_kernel_l.s
 	m4 cgemm_kernel_l.s > cgemm_kernel_l_nomacros.s
 	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DCN cgemm_kernel_l_nomacros.s -o $@
 	rm cgemm_kernel_l.s cgemm_kernel_l_nomacros.s
@@ -625,7 +714,7 @@ endif
 
 $(KDIR)cgemm_kernel_r$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -UDOUBLE -DCOMPLEX -DNC  $< -o cgemm_kernel_r.s
+	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DNC  $< -o - > cgemm_kernel_r.s
 	m4 cgemm_kernel_r.s > cgemm_kernel_r_nomacros.s
 	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DNC cgemm_kernel_r_nomacros.s -o $@
 	rm cgemm_kernel_r.s cgemm_kernel_r_nomacros.s
@@ -635,7 +724,7 @@ endif
 
 $(KDIR)cgemm_kernel_b$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -UDOUBLE -DCOMPLEX -DCC $< -o cgemm_kernel_b.s
+	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DCC $< -o - > cgemm_kernel_b.s
 	m4 cgemm_kernel_b.s > cgemm_kernel_b_nomacros.s
 	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DCC cgemm_kernel_b_nomacros.s -o $@
 	rm cgemm_kernel_b.s cgemm_kernel_b_nomacros.s
@@ -645,7 +734,7 @@ endif
 
 $(KDIR)zgemm_kernel_n$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DDOUBLE -DCOMPLEX -DNN $< -o zgemm_kernel_n.s
+	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DNN $< -o - > zgemm_kernel_n.s
 	m4 zgemm_kernel_n.s > zgemm_kernel_n_nomacros.s
 	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DNN zgemm_kernel_n_nomacros.s -o $@
 	rm zgemm_kernel_n.s zgemm_kernel_n_nomacros.s
@@ -655,7 +744,7 @@ endif
 
 $(KDIR)zgemm_kernel_l$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DDOUBLE -DCOMPLEX -DCN $< -o zgemm_kernel_l.s
+	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DCN $< -o - > zgemm_kernel_l.s
 	m4 zgemm_kernel_l.s > zgemm_kernel_l_nomacros.s
 	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DCN zgemm_kernel_l_nomacros.s -o $@
 	rm zgemm_kernel_l.s zgemm_kernel_l_nomacros.s
@@ -665,7 +754,7 @@ endif
 
 $(KDIR)zgemm_kernel_r$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DDOUBLE -DCOMPLEX -DNC $< -o zgemm_kernel_r.s
+	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DNC $< -o - > zgemm_kernel_r.s
 	m4 zgemm_kernel_r.s > zgemm_kernel_r_nomacros.s
 	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DNC zgemm_kernel_r_nomacros.s -o $@
 	rm zgemm_kernel_r.s zgemm_kernel_r_nomacros.s
@@ -675,7 +764,7 @@ endif
 
 $(KDIR)zgemm_kernel_b$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DDOUBLE -DCOMPLEX -DCC $< -o zgemm_kernel_b.s
+	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DCC $< -o - > zgemm_kernel_b.s
 	m4 zgemm_kernel_b.s > zgemm_kernel_b_nomacros.s
 	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DCC zgemm_kernel_b_nomacros.s -o $@
 	rm zgemm_kernel_b.s zgemm_kernel_b_nomacros.s
@@ -699,7 +788,7 @@ $(KDIR)xgemm_kernel_b$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(XGEMMKERNEL) $(XGEMMD
 ifdef USE_TRMM
 $(KDIR)strmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o strmmkernel_ln.s	
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o - > strmmkernel_ln.s	
 	m4 strmmkernel_ln.s > strmmkernel_ln_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA strmmkernel_ln_nomacros.s -o $@
 	rm strmmkernel_ln.s strmmkernel_ln_nomacros.s
@@ -709,7 +798,7 @@ endif
 
 $(KDIR)strmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o strmmkernel_lt.s	
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o - > strmmkernel_lt.s	
 	m4 strmmkernel_lt.s > strmmkernel_lt_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA strmmkernel_lt_nomacros.s -o $@
 	rm strmmkernel_lt.s strmmkernel_lt_nomacros.s
@@ -719,7 +808,7 @@ endif
 
 $(KDIR)strmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o strmmkernel_rn.s	
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o - > strmmkernel_rn.s	
 	m4 strmmkernel_rn.s > strmmkernel_rn_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA strmmkernel_rn_nomacros.s -o $@
 	rm strmmkernel_rn.s strmmkernel_rn_nomacros.s
@@ -729,7 +818,7 @@ endif
 
 $(KDIR)strmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o strmm_kernel_rt.s	
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
 	m4 strmm_kernel_rt.s > strmm_kernel_rt_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
 	rm strmm_kernel_rt.s strmm_kernel_rt_nomacros.s
@@ -739,7 +828,7 @@ endif
 
 $(KDIR)dtrmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o dtrmm_kernel_ln.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o - > dtrmm_kernel_ln.s
 	m4 dtrmm_kernel_ln.s > dtrmm_kernel_ln_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA dtrmm_kernel_ln_nomacros.s -o $@
 	rm dtrmm_kernel_ln.s dtrmm_kernel_ln_nomacros.s
@@ -749,7 +838,7 @@ endif
 
 $(KDIR)dtrmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o dtrmm_kernel_lt.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o - > dtrmm_kernel_lt.s
 	m4 dtrmm_kernel_lt.s > dtrmm_kernel_lt_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA dtrmm_kernel_lt_nomacros.s -o $@
 	rm dtrmm_kernel_lt.s dtrmm_kernel_lt_nomacros.s
@@ -759,7 +848,7 @@ endif
 
 $(KDIR)dtrmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o dtrmm_kernel_rn.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o - > dtrmm_kernel_rn.s
 	m4 dtrmm_kernel_rn.s > dtrmm_kernel_rn_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA dtrmm_kernel_rn_nomacros.s -o $@
 	rm dtrmm_kernel_rn.s dtrmm_kernel_rn_nomacros.s
@@ -769,7 +858,7 @@ endif
 
 $(KDIR)dtrmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o dtrmm_kernel_rt.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > dtrmm_kernel_rt.s
 	m4 dtrmm_kernel_rt.s > dtrmm_kernel_rt_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA dtrmm_kernel_rt_nomacros.s -o $@
 	rm dtrmm_kernel_rt.s dtrmm_kernel_rt_nomacros.s
@@ -791,7 +880,7 @@ $(KDIR)qtrmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(QGEMMKERNEL)
 
 $(KDIR)ctrmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN  $< -o ctrmm_kernel_ln.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN  $< -o - > ctrmm_kernel_ln.s
 	m4 ctrmm_kernel_ln.s > ctrmm_kernel_ln_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN ctrmm_kernel_ln_nomacros.s -o $@
 	rm ctrmm_kernel_ln.s ctrmm_kernel_ln_nomacros.s
@@ -801,7 +890,7 @@ endif
 
 $(KDIR)ctrmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o ctrmm_kernel_lt.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_lt.s
 	m4 ctrmm_kernel_lt.s > ctrmm_kernel_lt_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN ctrmm_kernel_lt_nomacros.s -o $@
 	rm ctrmm_kernel_lt.s ctrmm_kernel_lt_nomacros.s
@@ -811,7 +900,7 @@ endif
 
 $(KDIR)ctrmm_kernel_LR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o ctrmm_kernel_lr.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o - > ctrmm_kernel_lr.s
 	m4 ctrmm_kernel_lr.s > ctrmm_kernel_lr_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN  ctrmm_kernel_lr_nomacros.s -o $@
 	rm ctrmm_kernel_lr.s ctrmm_kernel_lr_nomacros.s
@@ -821,7 +910,7 @@ endif
 
 $(KDIR)ctrmm_kernel_LC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o ctrmm_kernel_lc.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o - > ctrmm_kernel_lc.s
 	m4 ctrmm_kernel_lc.s > ctrmm_kernel_lc_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN ctrmm_kernel_lc_nomacros.s -o $@
 	rm ctrmm_kernel_lc_nomacros.s ctrmm_kernel_lc.s
@@ -831,7 +920,7 @@ endif
 
 $(KDIR)ctrmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o ctrmm_kernel_rn.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_rn.s
 	m4 ctrmm_kernel_rn.s > ctrmm_kernel_rn_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN ctrmm_kernel_rn_nomacros.s -o $@
 	rm ctrmm_kernel_rn.s ctrmm_kernel_rn_nomacros.s
@@ -841,7 +930,7 @@ endif
 
 $(KDIR)ctrmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o ctrmm_kernel_rt.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_rt.s
 	m4 ctrmm_kernel_rt.s > ctrmm_kernel_rt_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN ctrmm_kernel_rt_nomacros.s -o $@
 	rm ctrmm_kernel_rt.s ctrmm_kernel_rt_nomacros.s
@@ -851,7 +940,7 @@ endif
 
 $(KDIR)ctrmm_kernel_RR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o ctrmm_kernel_rr.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o - > ctrmm_kernel_rr.s
 	m4 ctrmm_kernel_rr.s > ctrmm_kernel_rr_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC ctrmm_kernel_rr_nomacros.s -o $@
 	rm ctrmm_kernel_rr.s ctrmm_kernel_rr_nomacros.s
@@ -861,7 +950,7 @@ endif
 
 $(KDIR)ctrmm_kernel_RC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o ctrmm_kernel_RC.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o - > ctrmm_kernel_RC.s
 	m4 ctrmm_kernel_RC.s > ctrmm_kernel_RC_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC ctrmm_kernel_RC_nomacros.s -o $@
 	rm ctrmm_kernel_RC.s ctrmm_kernel_RC_nomacros.s
@@ -871,7 +960,7 @@ endif
 
 $(KDIR)ztrmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o ztrmm_kernel_ln.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_ln.s
 	m4 ztrmm_kernel_ln.s > ztrmm_kernel_ln_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN ztrmm_kernel_ln_nomacros.s -o $@
 	rm ztrmm_kernel_ln.s ztrmm_kernel_ln_nomacros.s
@@ -881,7 +970,7 @@ endif
 
 $(KDIR)ztrmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o ztrmm_kernel_lt.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_lt.s
 	m4 ztrmm_kernel_lt.s > ztrmm_kernel_lt_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN ztrmm_kernel_lt_nomacros.s -o $@
 	rm ztrmm_kernel_lt.s ztrmm_kernel_lt_nomacros.s
@@ -891,7 +980,7 @@ endif
 
 $(KDIR)ztrmm_kernel_LR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o ztrmm_kernel_lr.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o - > ztrmm_kernel_lr.s
 	m4 ztrmm_kernel_lr.s > ztrmm_kernel_lr_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN ztrmm_kernel_lr_nomacros.s -o $@
 	rm ztrmm_kernel_lr.s ztrmm_kernel_lr_nomacros.s
@@ -901,7 +990,7 @@ endif
 
 $(KDIR)ztrmm_kernel_LC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o ztrmm_kernel_lc.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o - > ztrmm_kernel_lc.s
 	m4 ztrmm_kernel_lc.s >ztrmm_kernel_lc_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN ztrmm_kernel_lc_nomacros.s -o $@
 	rm ztrmm_kernel_lc.s ztrmm_kernel_lc_nomacros.s 
@@ -911,7 +1000,7 @@ endif
 
 $(KDIR)ztrmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o ztrmm_kernel_rn.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_rn.s
 	m4 ztrmm_kernel_rn.s > ztrmm_kernel_rn_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN ztrmm_kernel_rn_nomacros.s -o $@
 	rm ztrmm_kernel_rn.s ztrmm_kernel_rn_nomacros.s
@@ -921,7 +1010,7 @@ endif
 
 $(KDIR)ztrmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o ztrmm_kernel_rt.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_rt.s
 	m4 ztrmm_kernel_rt.s > ztrmm_kernel_rt_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN ztrmm_kernel_rt_nomacros.s -o $@
 	rm ztrmm_kernel_rt.s ztrmm_kernel_rt_nomacros.s
@@ -931,7 +1020,7 @@ endif
 
 $(KDIR)ztrmm_kernel_RR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o ztrmm_kernel_rr.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o - > ztrmm_kernel_rr.s
 	m4 ztrmm_kernel_rr.s > ztrmm_kernel_rr_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC ztrmm_kernel_rr_nomacros.s -o $@
 	rm ztrmm_kernel_rr.s ztrmm_kernel_rr_nomacros.s
@@ -941,7 +1030,7 @@ endif
 
 $(KDIR)ztrmm_kernel_RC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o ztrmm_kernel_rc.s
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o - > ztrmm_kernel_rc.s
 	m4 ztrmm_kernel_rc.s > ztrmm_kernel_rc_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC ztrmm_kernel_rc_nomacros.s -o $@
 	rm ztrmm_kernel_rc.s ztrmm_kernel_rc_nomacros.s
@@ -961,7 +1050,7 @@ $(KDIR)strmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL)
 
 $(KDIR)strmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o strmm_kernel_rt.s	
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
 	m4 strmm_kernel_rt.s > strmm_kernel_rt_nomacros.s
 	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
 	rm strmm_kernel_rt.s strmm_kernel_rt_nomacros.s
@@ -1095,7 +1184,7 @@ $(KDIR)dtrsm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRSMKERNEL_LN) $(DT
 
 $(KDIR)dtrsm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRSMKERNEL_LT) $(DTRSMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ $< -o dtrsm_kernel_lt.s
+	$(CC) $(CFLAGS) -S -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ $< -o - > dtrsm_kernel_lt.s
 	m4 dtrsm_kernel_lt.s > dtrsm_kernel_lt_nomacros.s
 	$(CC) -c $(CFLAGS) -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ dtrsm_kernel_lt_nomacros.s -o $@
 	rm dtrsm_kernel_lt.s dtrsm_kernel_lt_nomacros.s
@@ -2206,6 +2295,11 @@ $(KDIR)xtrsm_oltncopy$(TSUFFIX).$(SUFFIX) : generic/ztrsm_ltcopy_$(XGEMM_UNROLL_
 $(KDIR)sgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMM_BETA)
 	$(CC) $(PFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 
+ifeq ($(BUILD_HALF),1)
+$(KDIR)shgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SHGEMM_BETA)
+	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+endif
+
 $(KDIR)dgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(DGEMM_BETA)
 	$(CC) $(PFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
 
@@ -2221,6 +2315,24 @@ $(KDIR)zgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(ZGEMM_BETA)
 $(KDIR)xgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(XGEMM_BETA)
 	$(CC) $(PFLAGS) -c -DXDOUBLE -DCOMPLEX $< -o $@
 
+
+ifeq ($(BUILD_HALF), 1)
+$(SHGEMMONCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMONCOPY)
+	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+
+$(SHGEMMOTCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMOTCOPY)
+	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+
+ifneq ($(SHGEMM_UNROLL_M), $(SHGEMM_UNROLL_N))
+$(SHGEMMINCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMINCOPY)
+	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+
+$(SHGEMMITCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMITCOPY)
+	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+
+endif
+endif
+
 $(SGEMMONCOPYOBJ_P) : $(KERNELDIR)/$(SGEMMONCOPY)
 	$(CC) $(PFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 
@@ -2325,6 +2437,12 @@ endif
 
 endif
 
+
+ifeq ($(BUILD_HALF), 1)
+$(KDIR)shgemm_kernel$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SHGEMMKERNEL) $(SHGEMMDEPEND)
+	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+endif
+
 $(KDIR)sgemm_kernel$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL) $(SGEMMDEPEND)
 	$(CC) $(PFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 
@@ -2342,7 +2460,7 @@ $(KDIR)cgemm_kernel_l$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMM
 
 $(KDIR)cgemm_kernel_r$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
 ifeq ($(OS), AIX)
-	$(CC) $(PFLAGS) -E -UDOUBLE -DCOMPLEX -DNC $< -o cgemm_kernel_r.s
+	$(CC) $(PFLAGS) -S -UDOUBLE -DCOMPLEX -DNC $< -o - > cgemm_kernel_r.s
 	m4 cgemm_kernel_r.s > cgemm_kernel_r_nomacros.s
 	$(CC) $(PFLAGS) -c -UDOUBLE -DCOMPLEX -DNC cgemm_kernel_r_nomacros.s -o $@
 	rm cgemm_kernel_r.s cgemm_kernel_r_nomacros.s 
@@ -2388,7 +2506,7 @@ $(KDIR)strmm_kernel_RN$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL)
 
 $(KDIR)strmm_kernel_RT$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL)
 ifeq ($(OS), AIX)
-	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o strmm_kernel_rt.s	
+	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
 	m4 strmmkernel_rn.s > strmm_kernel_rt_nomacros.s
 	$(CC) $(PFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
 	rm strmm_kernel_rt.s strmm_kernel_rt_nomacros.s

kernel/arm64/KERNEL.CORTEXA53

@@ -1,3 +1,187 @@
-include $(KERNELDIR)/KERNEL.ARMV8
+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
+
+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
+
+SAXPYKERNEL  = axpy.S
+DAXPYKERNEL  = axpy.S
+CAXPYKERNEL  = zaxpy.S
+ZAXPYKERNEL  = zaxpy.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
+
+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
 
 
+SASUMKERNEL    = asum.S
+DASUMKERNEL    = asum.S
+CASUMKERNEL    = casum.S
+ZASUMKERNEL    = zasum.S
+
+SCOPYKERNEL    = copy.S
+DCOPYKERNEL    = copy.S
+CCOPYKERNEL    = copy.S
+ZCOPYKERNEL    = copy.S
+
+SSWAPKERNEL    = swap.S
+DSWAPKERNEL    = swap.S
+CSWAPKERNEL    = swap.S
+ZSWAPKERNEL    = swap.S
+
+ISAMAXKERNEL   = iamax.S
+IDAMAXKERNEL   = iamax.S
+ICAMAXKERNEL   = izamax.S
+IZAMAXKERNEL   = izamax.S
+
+SNRM2KERNEL    = nrm2.S
+DNRM2KERNEL    = nrm2.S
+CNRM2KERNEL    = znrm2.S
+ZNRM2KERNEL    = znrm2.S
+
+DDOTKERNEL     = dot.S
+SDOTKERNEL     = dot.S
+CDOTKERNEL     = zdot.S
+ZDOTKERNEL     = zdot.S
+DSDOTKERNEL    = dot.S
+
+DGEMM_BETA     = dgemm_beta.S
+SGEMM_BETA     = sgemm_beta.S
+
+ifeq ($(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N), 8x8)
+SGEMMKERNEL    =  sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N)_cortexa53.S
+STRMMKERNEL    =  strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N)_cortexa53.S
+else
+SGEMMKERNEL    =  sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
+STRMMKERNEL    =  strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
+endif
+ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
+ifeq ($(SGEMM_UNROLL_M), 16)
+SGEMMITCOPY    =  sgemm_tcopy_$(SGEMM_UNROLL_M).S
+else
+SGEMMITCOPY    =  ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
+endif
+ifeq ($(SGEMM_UNROLL_M), 4)
+SGEMMINCOPY    =  sgemm_ncopy_$(SGEMM_UNROLL_M).S
+else
+SGEMMINCOPY    =  ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
+endif
+SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
+SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+
+SGEMMOTCOPY    =  sgemm_tcopy_$(SGEMM_UNROLL_N).S
+SGEMMONCOPY    =  sgemm_ncopy_$(SGEMM_UNROLL_N).S
+SGEMMONCOPYOBJ =  sgemm_oncopy$(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/daxpy_thunderx2t99.S

@@ -98,11 +98,58 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	X, X, #128
 .endm
 
+/*
+ * No need to do software prefetches if the vector fits
+ * into L1 cache
+ */
+.macro KERNEL_F16_L1CACHE
+	ldp	q4, q5, [X]
+	ldp	q16, q17, [Y]
+
+	ldp	q6, q7, [X, #32]
+	ldp	q18, q19, [Y, #32]
+
+	fmla	v16.2d, v4.2d, v0.d[0]
+	fmla	v17.2d, v5.2d, v0.d[0]
+
+	stp	q16, q17, [Y]
+
+	ldp	q20, q21, [X, #64]
+	ldp	q24, q25, [Y, #64]
+
+	fmla	v18.2d, v6.2d, v0.d[0]
+	fmla	v19.2d, v7.2d, v0.d[0]
+
+	stp	q18, q19, [Y, #32]
+
+	ldp	q22, q23, [X, #96]
+	ldp	q26, q27, [Y, #96]
+
+	fmla	v24.2d, v20.2d, v0.d[0]
+	fmla	v25.2d, v21.2d, v0.d[0]
+
+	stp	q24, q25, [Y, #64]
+
+	fmla	v26.2d, v22.2d, v0.d[0]
+	fmla	v27.2d, v23.2d, v0.d[0]
+
+	stp	q26, q27, [Y, #96]
+
+	add	Y, Y, #128
+	add	X, X, #128
+.endm
+
 .macro KERNEL_F32
 	KERNEL_F16 
 	KERNEL_F16 
 .endm
 
+
+.macro KERNEL_F32_L1CACHE
+	KERNEL_F16_L1CACHE
+	KERNEL_F16_L1CACHE
+.endm
+
 .macro INIT_S
 	lsl	INC_X, INC_X, #3
 	lsl	INC_Y, INC_Y, #3
@@ -138,6 +185,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	I, xzr
 	beq	.Ldaxpy_kernel_F1
 
+	cmp	N, #2048
+	ble	.Ldaxpy_kernel_F32_L1CACHE
+
 	.align 5
 .Ldaxpy_kernel_F32:
 
@@ -145,6 +195,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	subs	I, I, #1
 	bne	.Ldaxpy_kernel_F32
+	b	.Ldaxpy_kernel_F1
+
+	.align 5
+.Ldaxpy_kernel_F32_L1CACHE:
+
+	KERNEL_F32_L1CACHE
+
+	subs	I, I, #1
+	bne	.Ldaxpy_kernel_F32_L1CACHE
 
 .Ldaxpy_kernel_F1:
 

kernel/arm64/sgemm_ncopy_8.S

@@ -0,0 +1,562 @@
+/***************************************************************************
+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 A00 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	M	x0
+#define	N	x1
+#define	A00	x2
+#define	LDA	x3
+#define	B00	x4
+
+#define	A01	x5
+#define	A02	x6
+#define	A03	x7
+#define	A04	x8
+#define	A05	x9
+#define	A06	x10
+#define	A07	x11
+#define	A08	x12
+
+#define I	x13
+#define	J	x14
+#define	K	x15
+
+#define	TEMP1	x16
+#define	TEMP2	x17
+
+/**************************************************************************************
+* Macro definitions
+**************************************************************************************/
+
+.macro SAVE_REGS
+	add	sp, sp, #-(11 * 16)
+	stp	d8, d9, [sp, #(0 * 16)]
+	stp	d10, d11, [sp, #(1 * 16)]
+	stp	d12, d13, [sp, #(2 * 16)]
+	stp	d14, d15, [sp, #(3 * 16)]
+	stp	d16, d17, [sp, #(4 * 16)]
+	stp	x18, x19, [sp, #(5 * 16)]
+	stp	x20, x21, [sp, #(6 * 16)]
+	stp	x22, x23, [sp, #(7 * 16)]
+	stp	x24, x25, [sp, #(8 * 16)]
+	stp	x26, x27, [sp, #(9 * 16)]
+	str	x28, [sp, #(10 * 16)]
+.endm
+
+.macro RESTORE_REGS
+	ldp	d8, d9, [sp, #(0 * 16)]
+	ldp	d10, d11, [sp, #(1 * 16)]
+	ldp	d12, d13, [sp, #(2 * 16)]
+	ldp	d14, d15, [sp, #(3 * 16)]
+	ldp	d16, d17, [sp, #(4 * 16)]
+	ldp	x18, x19, [sp, #(5 * 16)]
+	ldp	x20, x21, [sp, #(6 * 16)]
+	ldp	x22, x23, [sp, #(7 * 16)]
+	ldp	x24, x25, [sp, #(8 * 16)]
+	ldp	x26, x27, [sp, #(9 * 16)]
+	ldr	x28, [sp, #(10 * 16)]
+	add	sp, sp, #(11*16)
+.endm
+
+.macro COPY4x8
+	ldr	q0, [A01], #16
+	ldr	q1, [A02], #16
+	ins	v8.s[0], v0.s[0]
+	ins	v10.s[0], v0.s[1]
+	ins	v12.s[0], v0.s[2]
+	ins	v14.s[0], v0.s[3]
+	ins	v8.s[1], v1.s[0]
+	ins	v10.s[1], v1.s[1]
+	ins	v12.s[1], v1.s[2]
+	ins	v14.s[1], v1.s[3]
+
+	ldr	q2, [A03], #16
+	ldr	q3, [A04], #16
+	ins	v8.s[2], v2.s[0]
+	ins	v10.s[2], v2.s[1]
+	ins	v12.s[2], v2.s[2]
+	ins	v14.s[2], v2.s[3]
+	ins	v8.s[3], v3.s[0]
+	ins	v10.s[3], v3.s[1]
+	ins	v12.s[3], v3.s[2]
+	ins	v14.s[3], v3.s[3]
+
+	ldr	q4, [A05], #16
+	ldr	q5, [A06], #16
+	ins	v9.s[0], v4.s[0]
+	ins	v11.s[0], v4.s[1]
+	ins	v13.s[0], v4.s[2]
+	ins	v15.s[0], v4.s[3]
+	ins	v9.s[1], v5.s[0]
+	ins	v11.s[1], v5.s[1]
+	ins	v13.s[1], v5.s[2]
+	ins	v15.s[1], v5.s[3]
+
+	ldr	q6, [A07], #16
+	ldr	q7, [A08], #16
+	ins	v9.s[2], v6.s[0]
+	ins	v11.s[2], v6.s[1]
+	ins	v13.s[2], v6.s[2]
+	ins	v15.s[2], v6.s[3]
+	ins	v9.s[3], v7.s[0]
+	ins	v11.s[3], v7.s[1]
+	ins	v13.s[3], v7.s[2]
+	ins	v15.s[3], v7.s[3]
+
+	st1	{v8.4s, v9.4s, v10.4s, v11.4s}, [B00], #64
+	st1	{v12.4s, v13.4s, v14.4s, v15.4s}, [B00], #64
+.endm
+
+.macro COPY2x8
+	ldr	d0, [A01], #8
+	ldr	d1, [A02], #8
+	ins	v8.s[0], v0.s[0]
+	ins	v10.s[0], v0.s[1]
+	ins	v8.s[1], v1.s[0]
+	ins	v10.s[1], v1.s[1]
+
+	ldr	d2, [A03], #8
+	ldr	d3, [A04], #8
+	ins	v8.s[2], v2.s[0]
+	ins	v10.s[2], v2.s[1]
+	ins	v8.s[3], v3.s[0]
+	ins	v10.s[3], v3.s[1]
+
+	ldr	d4, [A05], #8
+	ldr	d5, [A06], #8
+	ins	v9.s[0], v4.s[0]
+	ins	v11.s[0], v4.s[1]
+	ins	v9.s[1], v5.s[0]
+	ins	v11.s[1], v5.s[1]
+
+	ldr	d6, [A07], #8
+	ldr	d7, [A08], #8
+	ins	v9.s[2], v6.s[0]
+	ins	v11.s[2], v6.s[1]
+	ins	v9.s[3], v7.s[0]
+	ins	v11.s[3], v7.s[1]
+
+	st1	{v8.4s, v9.4s, v10.4s, v11.4s}, [B00], #64
+.endm
+
+.macro COPY1x8
+	ldr	s0, [A01], #4
+	ldr	s1, [A02], #4
+	ins	v8.s[0], v0.s[0]
+	ins	v8.s[1], v1.s[0]
+
+	ldr	s2, [A03], #4
+	ldr	s3, [A04], #4
+	ins	v8.s[2], v2.s[0]
+	ins	v8.s[3], v3.s[0]
+
+	ldr	s4, [A05], #4
+	ldr	s5, [A06], #4
+	ins	v9.s[0], v4.s[0]
+	ins	v9.s[1], v5.s[0]
+
+	ldr	s6, [A07], #4
+	ldr	s7, [A08], #4
+	ins	v9.s[2], v6.s[0]
+	ins	v9.s[3], v7.s[0]
+
+	st1	{v8.4s, v9.4s}, [B00], #32
+.endm
+
+.macro COPY4x4
+	ldr	q0, [A01], #16
+	ldr	q1, [A02], #16
+	ins	v8.s[0], v0.s[0]
+	ins	v9.s[0], v0.s[1]
+	ins	v10.s[0], v0.s[2]
+	ins	v11.s[0], v0.s[3]
+	ins	v8.s[1], v1.s[0]
+	ins	v9.s[1], v1.s[1]
+	ins	v10.s[1], v1.s[2]
+	ins	v11.s[1], v1.s[3]
+
+	ldr	q2, [A03], #16
+	ldr	q3, [A04], #16
+	ins	v8.s[2], v2.s[0]
+	ins	v9.s[2], v2.s[1]
+	ins	v10.s[2], v2.s[2]
+	ins	v11.s[2], v2.s[3]
+	ins	v8.s[3], v3.s[0]
+	ins	v9.s[3], v3.s[1]
+	ins	v10.s[3], v3.s[2]
+	ins	v11.s[3], v3.s[3]
+
+	st1	{v8.4s, v9.4s, v10.4s, v11.4s}, [B00], #64
+.endm
+
+.macro COPY2x4
+	ldr	d0, [A01], #8
+	ldr	d1, [A02], #8
+	ins	v8.s[0], v0.s[0]
+	ins	v9.s[0], v0.s[1]
+	ins	v8.s[1], v1.s[0]
+	ins	v9.s[1], v1.s[1]
+
+	ldr	d2, [A03], #8
+	ldr	d3, [A04], #8
+	ins	v8.s[2], v2.s[0]
+	ins	v9.s[2], v2.s[1]
+	ins	v8.s[3], v3.s[0]
+	ins	v9.s[3], v3.s[1]
+
+	st1	{v8.4s, v9.4s}, [B00], #32
+.endm
+
+.macro COPY1x4
+	ldr	s0, [A01], #4
+	ldr	s1, [A02], #4
+	ins	v8.s[0], v0.s[0]
+	ins	v8.s[1], v1.s[0]
+
+	ldr	s2, [A03], #4
+	ldr	s3, [A04], #4
+	ins	v8.s[2], v2.s[0]
+	ins	v8.s[3], v3.s[0]
+
+	st1	{v8.4s}, [B00], #16
+.endm
+
+.macro COPY4x2
+	ldr	q0, [A01], #16
+	ldr	q1, [A02], #16
+	ins	v8.s[0], v0.s[0]
+	ins	v9.s[0], v0.s[1]
+	ins	v10.s[0], v0.s[2]
+	ins	v11.s[0], v0.s[3]
+	ins	v8.s[1], v1.s[0]
+	ins	v9.s[1], v1.s[1]
+	ins	v10.s[1], v1.s[2]
+	ins	v11.s[1], v1.s[3]
+
+	st1	{v8.2s, v9.2s, v10.2s, v11.2s}, [B00], #32
+.endm
+
+.macro COPY2x2
+	ldr	d0, [A01], #8
+	ldr	d1, [A02], #8
+	ins	v8.s[0], v0.s[0]
+	ins	v9.s[0], v0.s[1]
+	ins	v8.s[1], v1.s[0]
+	ins	v9.s[1], v1.s[1]
+
+	st1	{v8.2s, v9.2s}, [B00], #16
+.endm
+
+.macro COPY1x2
+	ldr	s0, [A01], #4
+	ldr	s1, [A02], #4
+	ins	v8.s[0], v0.s[0]
+	ins	v8.s[1], v1.s[0]
+
+	st1	{v8.2s}, [B00], #8
+.endm
+
+.macro COPY1x1
+	ldr	s0, [A01], #4
+	str	s0, [B00], #4
+.endm
+
+/**************************************************************************************
+* End of macro definitions
+**************************************************************************************/
+
+	PROLOGUE
+
+	.align 5
+
+	SAVE_REGS
+
+	lsl	LDA, LDA, #2					// LDA = LDA * SIZE
+
+.Lsgemm_ncopy_L8_BEGIN:
+
+	asr	J, N, #3					// J = N / 8
+	cmp 	J, #0
+	ble	.Lsgemm_ncopy_L4_BEGIN
+
+	.align	5
+.Lsgemm_ncopy_L8_M4_BEGIN:
+
+	mov	A01, A00
+	add	A02, A01, LDA
+	add	A03, A02, LDA
+	add	A04, A03, LDA
+	add	A05, A04, LDA
+	add	A06, A05, LDA
+	add	A07, A06, LDA
+	add	A08, A07, LDA
+	add	A00, A08, LDA
+
+	asr	I, M, #2					// I = M / 4
+	cmp	I, #0
+	ble	.Lsgemm_ncopy_L8_M4_40
+
+	asr	K, M, #4					// K = M / 16(cacheline)
+	mov	TEMP1, A01
+
+	.align	5
+.Lsgemm_tcopy_L8_warnup_1:
+
+	ldr	s0, [TEMP1], #64
+
+	subs	K, K, #1
+	bgt	.Lsgemm_tcopy_L8_warnup_1
+
+	asr	K, M, #4					// K = M / 16(cacheline)
+	mov	TEMP1, A02
+
+	.align	5
+.Lsgemm_tcopy_L8_warnup_2:
+
+	ldr	s0, [TEMP1], #64
+
+	subs	K, K, #1
+	bgt	.Lsgemm_tcopy_L8_warnup_2
+
+	asr	K, M, #4					// K = M / 16(cacheline)
+	mov	TEMP1, A03
+
+	.align	5
+.Lsgemm_tcopy_L8_warnup_3:
+
+	ldr	s0, [TEMP1], #64
+
+	subs	K, K, #1
+	bgt	.Lsgemm_tcopy_L8_warnup_3
+
+	asr	K, M, #4					// K = M / 16(cacheline)
+	mov	TEMP1, A04
+
+	.align	5
+.Lsgemm_tcopy_L8_warnup_4:
+
+	ldr	s0, [TEMP1], #64
+
+	subs	K, K, #1
+	bgt	.Lsgemm_tcopy_L8_warnup_4
+
+	asr	K, M, #4					// K = M / 16(cacheline)
+	mov	TEMP1, A05
+
+	.align	5
+.Lsgemm_tcopy_L8_warnup_5:
+
+	ldr	s0, [TEMP1], #64
+
+	subs	K, K, #1
+	bgt	.Lsgemm_tcopy_L8_warnup_5
+
+	asr	K, M, #4					// K = M / 16(cacheline)
+	mov	TEMP1, A06
+
+	.align	5
+.Lsgemm_tcopy_L8_warnup_6:
+
+	ldr	s0, [TEMP1], #64
+
+	subs	K, K, #1
+	bgt	.Lsgemm_tcopy_L8_warnup_6
+
+	asr	K, M, #4					// K = M / 16(cacheline)
+	mov	TEMP1, A07
+
+	.align	5
+.Lsgemm_tcopy_L8_warnup_7:
+
+	ldr	s0, [TEMP1], #64
+
+	subs	K, K, #1
+	bgt	.Lsgemm_tcopy_L8_warnup_7
+
+	asr	K, M, #4					// K = M / 16(cacheline)
+	mov	TEMP1, A08
+
+	.align	5
+.Lsgemm_tcopy_L8_warnup_8:
+
+	ldr	s0, [TEMP1], #64
+
+	subs	K, K, #1
+	bgt	.Lsgemm_tcopy_L8_warnup_8
+
+	.align	5
+.Lsgemm_ncopy_L8_M4_20:
+
+	COPY4x8
+
+	subs	I, I, #1
+	bne	.Lsgemm_ncopy_L8_M4_20
+
+.Lsgemm_ncopy_L8_M4_40:
+
+	and	I, M, #2
+	cmp	I, #0
+	ble	.Lsgemm_ncopy_L8_M4_60
+
+	COPY2x8
+
+.Lsgemm_ncopy_L8_M4_60:
+
+	and	I, M, #1
+	cmp	I, #0
+	ble	.Lsgemm_ncopy_L8_M4_END
+
+	COPY1x8
+
+.Lsgemm_ncopy_L8_M4_END:
+
+	subs	J , J, #1						// j--
+	bne	.Lsgemm_ncopy_L8_M4_BEGIN
+
+/*********************************************************************************************/
+
+.Lsgemm_ncopy_L4_BEGIN:
+
+	tst	N, #7
+	ble	.Lsgemm_ncopy_L999
+
+	tst	N, #4
+	ble	.Lsgemm_ncopy_L2_BEGIN
+
+.Lsgemm_ncopy_L4_M4_BEGIN:
+	mov	A01, A00
+	add	A02, A01, LDA
+	add	A03, A02, LDA
+	add	A04, A03, LDA
+	add	A00, A04, LDA
+
+	asr	I, M, #2					// I = M / 4
+	cmp 	I, #0
+	ble	.Lsgemm_ncopy_L4_M4_40
+
+	.align	5
+.Lsgemm_ncopy_L4_M4_20:
+
+	COPY4x4
+
+	subs	I, I, #1
+	bne	.Lsgemm_ncopy_L4_M4_20
+
+.Lsgemm_ncopy_L4_M4_40:
+
+	and	I, M, #2
+	cmp	I, #0
+	ble	.Lsgemm_ncopy_L4_M4_60
+
+	COPY2x4
+
+.Lsgemm_ncopy_L4_M4_60:
+
+	and	I, M, #1
+	cmp	I, #0
+	ble	.Lsgemm_ncopy_L4_M4_END
+
+	COPY1x4
+
+.Lsgemm_ncopy_L4_M4_END:
+
+
+/*********************************************************************************************/
+
+.Lsgemm_ncopy_L2_BEGIN:
+
+	tst	N, #2
+	ble	.Lsgemm_ncopy_L1_BEGIN
+
+.Lsgemm_ncopy_L2_M4_BEGIN:
+
+	mov	A01, A00
+	add	A02, A01, LDA
+	add	A00, A02, LDA
+
+	asr	I, M, #2					// I = M / 4
+	cmp	I, #0
+	ble	.Lsgemm_ncopy_L2_M4_40
+
+	.align	5
+.Lsgemm_ncopy_L2_M4_20:
+
+	COPY4x2
+
+	subs	I , I , #1
+	bne	.Lsgemm_ncopy_L2_M4_20
+
+
+.Lsgemm_ncopy_L2_M4_40:
+
+	and	I, M, #2
+	cmp	I, #0
+	ble	.Lsgemm_ncopy_L2_M4_60
+
+	COPY2x2
+
+.Lsgemm_ncopy_L2_M4_60:
+
+	and	I, M, #1
+	cmp	I, #0
+	ble	.Lsgemm_ncopy_L2_M4_END
+
+	COPY1x2
+
+.Lsgemm_ncopy_L2_M4_END:
+
+.Lsgemm_ncopy_L1_BEGIN:
+
+	tst	N, #1
+	ble	.Lsgemm_ncopy_L999
+
+.Lsgemm_ncopy_L1_M1_BEGIN:
+
+	mov	A01, A00
+
+	mov	I, M
+	cmp	I, #0
+	ble	.Lsgemm_ncopy_L1_M1_END
+
+	.align	5
+.Lsgemm_ncopy_L1_M1_20:
+
+	COPY1x1
+
+	subs	I, I, #1
+	bne	.Lsgemm_ncopy_L1_M1_20
+
+.Lsgemm_ncopy_L1_M1_END:
+
+.Lsgemm_ncopy_L999:
+
+	mov	x0, #0
+	RESTORE_REGS
+	ret
+
+	EPILOGUE

kernel/arm64/sgemm_tcopy_8.S

@@ -0,0 +1,707 @@
+/***************************************************************************
+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 A00 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	M		x0
+#define	N		x1
+#define	A		x2
+#define	LDA		x3
+#define	B		x4
+
+#define M8		x5
+
+#define	A01		x6
+#define	A02		x7
+#define	A03		x8
+#define	A04		x9
+#define	A05		x10
+#define	A06		x11
+#define	A07		x12
+#define	A08		x13
+
+#define	B01		x14
+#define	B02		x15
+#define	B03		x16
+#define	B04		x17
+#define	B00		x22
+
+
+#define I		x18
+#define	J		x19
+
+#define TEMP1		x20
+
+#define A_PREFETCH	256
+
+/**************************************************************************************
+* Macro definitions
+**************************************************************************************/
+.macro SAVE_REGS
+	add	sp, sp, #-(11 * 16)
+	stp	d8, d9, [sp, #(0 * 16)]
+	stp	d10, d11, [sp, #(1 * 16)]
+	stp	d12, d13, [sp, #(2 * 16)]
+	stp	d14, d15, [sp, #(3 * 16)]
+	stp	d16, d17, [sp, #(4 * 16)]
+	stp	x18, x19, [sp, #(5 * 16)]
+	stp	x20, x21, [sp, #(6 * 16)]
+	stp	x22, x23, [sp, #(7 * 16)]
+	stp	x24, x25, [sp, #(8 * 16)]
+	stp	x26, x27, [sp, #(9 * 16)]
+	str	x28, [sp, #(10 * 16)]
+.endm
+
+.macro RESTORE_REGS
+	ldp	d8, d9, [sp, #(0 * 16)]
+	ldp	d10, d11, [sp, #(1 * 16)]
+	ldp	d12, d13, [sp, #(2 * 16)]
+	ldp	d14, d15, [sp, #(3 * 16)]
+	ldp	d16, d17, [sp, #(4 * 16)]
+	ldp	x18, x19, [sp, #(5 * 16)]
+	ldp	x20, x21, [sp, #(6 * 16)]
+	ldp	x22, x23, [sp, #(7 * 16)]
+	ldp	x24, x25, [sp, #(8 * 16)]
+	ldp	x26, x27, [sp, #(9 * 16)]
+	ldr	x28, [sp, #(10 * 16)]
+	add	sp, sp, #(11*16)
+.endm
+
+/*************************************************************************************************************************/
+
+.macro COPY8x8
+	prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+	prfm	PLDL1KEEP, [A03, #A_PREFETCH]
+	prfm	PLDL1KEEP, [A04, #A_PREFETCH]
+	prfm	PLDL1KEEP, [A05, #A_PREFETCH]
+	prfm	PLDL1KEEP, [A06, #A_PREFETCH]
+	prfm	PLDL1KEEP, [A07, #A_PREFETCH]
+	prfm	PLDL1KEEP, [A08, #A_PREFETCH]
+
+	ldp	q0, q1, [A01]
+	ldp	q2, q3, [A02]
+	add	A01, A01, #32
+	add	A02, A02, #32
+
+	st1	{v0.4s, v1.4s, v2.4s, v3.4s}, [B00]
+	add	TEMP1, B00, #64
+
+	ldp	q4, q5, [A03]
+	ldp	q6, q7, [A04]
+	add	A03, A03, #32
+	add	A04, A04, #32
+
+	st1	{v4.4s, v5.4s, v6.4s, v7.4s}, [TEMP1]
+	add	TEMP1, TEMP1, #64
+
+	ldp	q8, q9, [A05]
+	ldp	q10, q11, [A06]
+	add	A05, A05, #32
+	add	A06, A06, #32
+
+	st1	{v8.4s, v9.4s, v10.4s, v11.4s}, [TEMP1]
+	add	TEMP1, TEMP1, #64
+
+	ldp	q12, q13, [A07]
+	ldp	q14, q15, [A08]
+	add	A07, A07, #32
+	add	A08, A08, #32
+
+	st1	{v12.4s, v13.4s, v14.4s, v15.4s}, [TEMP1]
+	add	TEMP1, TEMP1, #64
+
+	add	B00, B00, M8
+.endm
+
+.macro COPY4x8
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A03, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A04, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A05, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A06, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A07, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A08, #A_PREFETCH]
+
+	ldr	q0, [A01]
+	ldr	q1, [A02]
+	ldr	q2, [A03]
+	ldr	q3, [A04]
+	add	A01, A01, #16
+	add	A02, A02, #16
+	add	A03, A03, #16
+	add	A04, A04, #16
+
+	st1	{v0.4s, v1.4s, v2.4s, v3.4s}, [B01]
+	add	B01, B01, #64
+
+	ldr	q4, [A05]
+	ldr	q5, [A06]
+	ldr	q6, [A07]
+	ldr	q7, [A08]
+
+	add	A05, A05, #16
+	add	A06, A06, #16
+	add	A07, A07, #16
+	add	A08, A08, #16
+
+	st1	{v4.4s, v5.4s, v6.4s, v7.4s}, [B01]
+	add	B01, B01, #64
+.endm
+
+.macro COPY2x8
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A03, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A04, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A05, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A06, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A07, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A08, #A_PREFETCH]
+
+	ldr	d0, [A01]
+	ldr	d1, [A02]
+	ldr	d2, [A03]
+	ldr	d3, [A04]
+
+	add	A01, A01, #8
+	add	A02, A02, #8
+	add	A03, A03, #8
+	add	A04, A04, #8
+
+	stp	d0, d1, [B02]
+	add	B02, B02, #16
+	stp	d2, d3, [B02]
+	add	B02, B02, #16
+
+	ldr	d4, [A05]
+	ldr	d5, [A06]
+	ldr	d6, [A07]
+	ldr	d7, [A08]
+
+	add	A05, A05, #8
+	add	A06, A06, #8
+	add	A07, A07, #8
+	add	A08, A08, #8
+
+	stp	d4, d5, [B02]
+	add	B02, B02, #16
+	stp	d6, d7, [B02]
+	add	B02, B02, #16
+
+.endm
+
+.macro COPY1x8
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A03, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A04, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A05, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A06, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A07, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A08, #A_PREFETCH]
+
+	ldr	s0, [A01]
+	ldr	s1, [A02]
+	ldr	s2, [A03]
+	ldr	s3, [A04]
+
+	add	A01, A01, #4
+	add	A02, A02, #4
+	add	A03, A03, #4
+	add	A04, A04, #4
+
+	stp	s0, s1, [B03]
+	add	B03, B03, #8
+	stp	s2, s3, [B03]
+	add	B03, B03, #8
+
+	ldr	s4, [A05]
+	ldr	s5, [A06]
+	ldr	s6, [A07]
+	ldr	s7, [A08]
+
+	ldr	d4, [A05], #8
+	ldr	d5, [A06], #8
+	ldr	d6, [A07], #8
+	ldr	d7, [A08], #8
+
+	stp	s4, s5, [B03]
+	add	B03, B03, #8
+	stp	s6, s7, [B03]
+	add	B03, B03, #8
+
+.endm
+
+/*************************************************************************************************************************/
+
+.macro COPY8x4
+	prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+	prfm	PLDL1KEEP, [A03, #A_PREFETCH]
+	prfm	PLDL1KEEP, [A04, #A_PREFETCH]
+
+	ldp	q0, q1, [A01]
+	ldp	q2, q3, [A02]
+	add	A01, A01, #32
+	add	A02, A02, #32
+
+	st1	{v0.4s, v1.4s, v2.4s, v3.4s}, [B00]
+	add	TEMP1, B00, #64
+
+	ldp	q4, q5, [A03]
+	ldp	q6, q7, [A04]
+	add	A03, A03, #32
+	add	A04, A04, #32
+
+	st1	{v4.4s, v5.4s, v6.4s, v7.4s}, [TEMP1]
+	add	TEMP1, TEMP1, #64
+
+	add	B00, B00, M8
+.endm
+
+.macro COPY4x4
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A03, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A04, #A_PREFETCH]
+
+	ldr	q0, [A01]
+	ldr	q1, [A02]
+	ldr	q2, [A03]
+	ldr	q3, [A04]
+	add	A01, A01, #16
+	add	A02, A02, #16
+	add	A03, A03, #16
+	add	A04, A04, #16
+
+	st1	{v0.4s, v1.4s, v2.4s, v3.4s}, [B01]
+
+	add	B01, B01, #64
+.endm
+
+.macro COPY2x4
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A03, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A04, #A_PREFETCH]
+
+	ldr	d0, [A01]
+	ldr	d1, [A02]
+	ldr	d2, [A03]
+	ldr	d3, [A04]
+
+	add	A01, A01, #8
+	add	A02, A02, #8
+	add	A03, A03, #8
+	add	A04, A04, #8
+
+	stp	d0, d1, [B02]
+	add	B02, B02, #16
+	stp	d2, d3, [B02]
+
+	add	B02, B02, #16
+.endm
+
+.macro COPY1x4
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A03, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A04, #A_PREFETCH]
+
+	ldr	s0, [A01]
+	ldr	s1, [A02]
+	ldr	s2, [A03]
+	ldr	s3, [A04]
+
+	add	A01, A01, #4
+	add	A02, A02, #4
+	add	A03, A03, #4
+	add	A04, A04, #4
+
+	stp	s0, s1, [B03]
+	add	B03, B03, #8
+	stp	s2, s3, [B03]
+	add	B03, B03, #8
+
+.endm
+
+/*************************************************************************************************************************/
+
+.macro COPY8x2
+	prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+
+	ld1	{v0.4s, v1.4s}, [A01]
+	ld1	{v2.4s, v3.4s}, [A02]
+	add	A01, A01, #32
+	add	A02, A02, #32
+
+	st1	{v0.4s, v1.4s, v2.4s, v3.4s}, [B00]
+	add	B00, B00, M8
+.endm
+
+.macro COPY4x2
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+
+	ldr	q0, [A01]
+	ldr	q1, [A02]
+	add	A01, A01, #16
+	add	A02, A02, #16
+
+	stp	q0, q1, [B01]
+	add	B01, B01, #32
+.endm
+
+.macro COPY2x2
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+
+	ldr	d0, [A01]
+	ldr	d1, [A02]
+
+	add	A01, A01, #8
+	add	A02, A02, #8
+
+	stp	d0, d1, [B02]
+	add	B02, B02, #16
+.endm
+
+.macro COPY1x2
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+	//prfm	PLDL1KEEP, [A02, #A_PREFETCH]
+
+	ldr	s0, [A01]
+	ldr	s1, [A02]
+
+	add	A01, A01, #4
+	add	A02, A02, #4
+
+	stp	s0, s1, [B03]
+
+	add	B03, B03, #8
+.endm
+
+/*************************************************************************************************************************/
+
+.macro COPY8x1
+	prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+
+	ldp	q0, q1, [A01]
+	add	A01, A01, #32
+	stp	q0, q1, [B00]
+
+	add	B00, B00, M8
+.endm
+
+.macro COPY4x1
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+
+	ldr q0, [A01]
+	add	A01, A01, #16
+	str q0, [B01]
+
+	add	B01, B01, #16
+.endm
+
+.macro COPY2x1
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+
+	ldr	d0, [A01]
+	add	A01, A01, #8
+	str d0, [B02]
+
+	add	B02, B02, #8
+.endm
+
+.macro COPY1x1
+	//prfm	PLDL1KEEP, [A01, #A_PREFETCH]
+
+	ldr	s0, [A01]
+	add	A01, A01, #4
+	str	s0, [B03]
+
+	add	B03, B03, #4
+.endm
+
+/**************************************************************************************
+* End of macro definitions
+**************************************************************************************/
+
+	PROLOGUE
+
+	.align 5
+
+	SAVE_REGS
+
+	lsl	LDA, LDA, #2					// LDA = LDA * SIZE
+
+	lsl	TEMP1, M, #2					// TEMP1 = M * SIZE
+
+	and	B01 , N , #-8
+	and	B02 , N , #-4
+	and	B03 , N , #-2
+
+	mul	B01, B01, TEMP1
+	mul	B02, B02, TEMP1
+	mul	B03, B03, TEMP1
+
+	add	B01 , B01, B
+	add	B02 , B02, B
+	add	B03 , B03, B
+
+	lsl	M8, M, #5					// M8 = M * 8 * SIZE
+
+.Lsgemm_tcopy_L8_BEGIN:
+
+	asr 	J, M, #3					// J = M / 8
+	cmp 	J, #0
+	ble	.Lsgemm_tcopy_L4_BEGIN
+
+	.align	5
+.Lsgemm_tcopy_L8_M8_BEGIN:
+
+	mov	A01, A
+	add	A02, A01, LDA
+	add	A03, A02, LDA
+	add	A04, A03, LDA
+	add	A05, A04, LDA
+	add	A06, A05, LDA
+	add	A07, A06, LDA
+	add	A08, A07, LDA
+	add	A, A08, LDA
+
+	mov	B00, B
+	add	B, B00, #256					// B = B + 8 * 8 * SIZE
+
+	asr	I, N, #3					// I = N / 8
+	cmp 	I, #0
+	ble	.Lsgemm_tcopy_L8_M8_40
+
+	.align	5
+.Lsgemm_tcopy_L8_M8_20:
+
+	COPY8x8
+
+	subs	I , I , #1
+	bne	.Lsgemm_tcopy_L8_M8_20
+
+.Lsgemm_tcopy_L8_M8_40:
+
+	tst	N , #4
+	ble	.Lsgemm_tcopy_L8_M8_60
+
+	COPY4x8
+
+.Lsgemm_tcopy_L8_M8_60:
+
+	tst	N , #2
+	ble	.Lsgemm_tcopy_L8_M8_80
+
+	COPY2x8
+
+.Lsgemm_tcopy_L8_M8_80:
+
+	tst	N, #1
+	ble	.Lsgemm_tcopy_L8_M8_END
+
+	COPY1x8
+
+.Lsgemm_tcopy_L8_M8_END:
+
+	subs	J, J, #1						// j--
+	bne	.Lsgemm_tcopy_L8_M8_BEGIN
+
+/*********************************************************************************************/
+
+.Lsgemm_tcopy_L4_BEGIN:
+
+	tst	M, #7
+	ble	.Lsgemm_tcopy_L999
+
+	tst	M, #4
+	ble	.Lsgemm_tcopy_L2_BEGIN
+
+.Lsgemm_tcopy_L4_M8_BEGIN:
+
+	mov	A01, A
+	add	A02, A01, LDA
+	add	A03, A02, LDA
+	add	A04, A03, LDA
+	add	A, A04, LDA
+
+	mov	B00, B
+	add	B, B00, #128					// B = B + 4 * 8 * SIZE
+
+	asr	I, N, #3					// I = N / 8
+	cmp 	I, #0
+	ble	.Lsgemm_tcopy_L4_M8_40
+
+	.align	5
+.Lsgemm_tcopy_L4_M8_20:
+
+	COPY8x4
+
+	subs	I , I , #1
+	bne	.Lsgemm_tcopy_L4_M8_20
+
+.Lsgemm_tcopy_L4_M8_40:
+
+	tst	N , #4
+	ble	.Lsgemm_tcopy_L4_M8_60
+
+	COPY4x4
+
+.Lsgemm_tcopy_L4_M8_60:
+
+	tst	N , #2
+	ble	.Lsgemm_tcopy_L4_M8_80
+
+	COPY2x4
+
+.Lsgemm_tcopy_L4_M8_80:
+
+	tst	N , #1
+	ble	.Lsgemm_tcopy_L4_M8_END
+
+	COPY1x4
+
+
+.Lsgemm_tcopy_L4_M8_END:
+
+/*********************************************************************************************/
+
+.Lsgemm_tcopy_L2_BEGIN:
+
+	tst	M, #3
+	ble	.Lsgemm_tcopy_L999
+
+	tst	M, #2
+	ble	.Lsgemm_tcopy_L1_BEGIN
+
+.Lsgemm_tcopy_L2_M16_BEGIN:
+
+	mov	A01, A
+	add	A02, A01, LDA
+	add	A, A02, LDA
+
+	mov	B00, B
+	add	B, B00, #64					// B = B + 2 * 8 * SIZE
+
+	asr	I, N, #3					// I = N / 8
+	cmp 	I, #0
+	ble	.Lsgemm_tcopy_L2_M8_40
+
+	.align	5
+.Lsgemm_tcopy_L2_M8_20:
+
+	COPY8x2
+
+	subs	I , I , #1
+	bne	.Lsgemm_tcopy_L2_M8_20
+
+.Lsgemm_tcopy_L2_M8_40:
+
+	tst	N , #4
+	ble	.Lsgemm_tcopy_L2_M8_60
+
+	COPY4x2
+
+.Lsgemm_tcopy_L2_M8_60:
+
+	tst	N , #2
+	ble	.Lsgemm_tcopy_L2_M8_80
+
+	COPY2x2
+
+.Lsgemm_tcopy_L2_M8_80:
+
+	tst	N , #1
+	ble	.Lsgemm_tcopy_L2_M8_END
+
+	COPY1x2
+
+.Lsgemm_tcopy_L2_M8_END:
+
+/*********************************************************************************************/
+
+.Lsgemm_tcopy_L1_BEGIN:
+
+	tst	M, #1
+	ble	.Lsgemm_tcopy_L999
+
+
+.Lsgemm_tcopy_L1_M16_BEGIN:
+
+	mov	A01, A						// A01 = A
+	mov	B00, B
+
+	asr	I, N, #3					// I = M / 8
+	cmp 	I, #0
+	ble	.Lsgemm_tcopy_L1_M8_40
+
+	.align	5
+.Lsgemm_tcopy_L1_M8_20:
+
+	COPY8x1
+
+	subs	I , I , #1
+	bne	.Lsgemm_tcopy_L1_M8_20
+
+.Lsgemm_tcopy_L1_M8_40:
+
+	tst	N , #4
+	ble	.Lsgemm_tcopy_L1_M8_60
+
+	COPY4x1
+
+.Lsgemm_tcopy_L1_M8_60:
+
+	tst	N , #2
+	ble	.Lsgemm_tcopy_L1_M8_80
+
+	COPY2x1
+
+.Lsgemm_tcopy_L1_M8_80:
+
+	tst	N , #1
+	ble	.Lsgemm_tcopy_L1_M8_END
+
+	COPY1x1
+
+
+.Lsgemm_tcopy_L1_M8_END:
+
+.Lsgemm_tcopy_L999:
+
+	mov	x0, #0						// set return value
+	RESTORE_REGS
+	ret
+
+	EPILOGUE

kernel/generic/gemm_beta.c

@@ -39,7 +39,7 @@
 #include "common.h"
 
 int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
-	  FLOAT *dummy2, BLASLONG dummy3, FLOAT *dummy4, BLASLONG dummy5,
+	  IFLOAT *dummy2, BLASLONG dummy3, IFLOAT *dummy4, BLASLONG dummy5,
 	  FLOAT *c, BLASLONG ldc){
 
 

kernel/generic/gemm_ncopy_16.c

@@ -39,24 +39,24 @@
 #include <stdio.h>
 #include "common.h"
 
-int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
+int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
   BLASLONG i, j;
 
-  FLOAT *aoffset;
+  IFLOAT *aoffset;
-  FLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
+  IFLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
-  FLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
+  IFLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
-  FLOAT *aoffset9, *aoffset10, *aoffset11, *aoffset12;
+  IFLOAT *aoffset9, *aoffset10, *aoffset11, *aoffset12;
-  FLOAT *aoffset13, *aoffset14, *aoffset15, *aoffset16;
+  IFLOAT *aoffset13, *aoffset14, *aoffset15, *aoffset16;
 
-  FLOAT *boffset;
+  IFLOAT *boffset;
-  FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
+  IFLOAT ctemp01, ctemp02, ctemp03, ctemp04;
-  FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
+  IFLOAT ctemp05, ctemp06, ctemp07, ctemp08;
-  FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
+  IFLOAT ctemp09, ctemp10, ctemp11, ctemp12;
-  FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
+  IFLOAT ctemp13, ctemp14, ctemp15, ctemp16;
-  FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
+  IFLOAT ctemp17, ctemp18, ctemp19, ctemp20;
-  FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
+  IFLOAT ctemp21, ctemp22, ctemp23, ctemp24;
-  FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
+  IFLOAT ctemp25, ctemp26, ctemp27, ctemp28;
-  FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
+  IFLOAT ctemp29, ctemp30, ctemp31, ctemp32;
 
   aoffset = a;
   boffset = b;

kernel/generic/gemm_ncopy_2.c

@@ -39,10 +39,10 @@
 #include <stdio.h>
 #include "common.h"
 
-int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
+int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
   BLASLONG i, j;
-  FLOAT *a_offset, *a_offset1, *a_offset2;
+  IFLOAT *a_offset, *a_offset1, *a_offset2;
-  FLOAT *b_offset;
+  IFLOAT *b_offset;
 
   a_offset = a;
   b_offset = b;

kernel/generic/gemm_ncopy_8.c

@@ -39,30 +39,30 @@
 #include <stdio.h>
 #include "common.h"
 
-int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
+int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
   BLASLONG i, j;
 
-  FLOAT *aoffset;
+  IFLOAT *aoffset;
-  FLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
+  IFLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
-  FLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
+  IFLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
 
-  FLOAT *boffset;
+  IFLOAT *boffset;
-  FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
+  IFLOAT ctemp01, ctemp02, ctemp03, ctemp04;
-  FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
+  IFLOAT ctemp05, ctemp06, ctemp07, ctemp08;
-  FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
+  IFLOAT ctemp09, ctemp10, ctemp11, ctemp12;
-  FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
+  IFLOAT ctemp13, ctemp14, ctemp15, ctemp16;
-  FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
+  IFLOAT ctemp17, ctemp18, ctemp19, ctemp20;
-  FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
+  IFLOAT ctemp21, ctemp22, ctemp23, ctemp24;
-  FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
+  IFLOAT ctemp25, ctemp26, ctemp27, ctemp28;
-  FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
+  IFLOAT ctemp29, ctemp30, ctemp31, ctemp32;
-  FLOAT ctemp33, ctemp34, ctemp35, ctemp36;
+  IFLOAT ctemp33, ctemp34, ctemp35, ctemp36;
-  FLOAT ctemp37, ctemp38, ctemp39, ctemp40;
+  IFLOAT ctemp37, ctemp38, ctemp39, ctemp40;
-  FLOAT ctemp41, ctemp42, ctemp43, ctemp44;
+  IFLOAT ctemp41, ctemp42, ctemp43, ctemp44;
-  FLOAT ctemp45, ctemp46, ctemp47, ctemp48;
+  IFLOAT ctemp45, ctemp46, ctemp47, ctemp48;
-  FLOAT ctemp49, ctemp50, ctemp51, ctemp52;
+  IFLOAT ctemp49, ctemp50, ctemp51, ctemp52;
-  FLOAT ctemp53, ctemp54, ctemp55, ctemp56;
+  IFLOAT ctemp53, ctemp54, ctemp55, ctemp56;
-  FLOAT ctemp57, ctemp58, ctemp59, ctemp60;
+  IFLOAT ctemp57, ctemp58, ctemp59, ctemp60;
-  FLOAT ctemp61, ctemp62, ctemp63, ctemp64;
+  IFLOAT ctemp61, ctemp62, ctemp63, ctemp64;
 
 
   aoffset = a;

kernel/generic/gemm_tcopy_16.c

@@ -39,22 +39,22 @@
 #include <stdio.h>
 #include "common.h"
 
-int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
+int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
 
   BLASLONG i, j;
 
-  FLOAT *aoffset;
+  IFLOAT *aoffset;
-  FLOAT *aoffset1, *aoffset2;
+  IFLOAT *aoffset1, *aoffset2;
-  FLOAT *boffset;
+  IFLOAT *boffset;
 
-  FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
+  IFLOAT ctemp01, ctemp02, ctemp03, ctemp04;
-  FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
+  IFLOAT ctemp05, ctemp06, ctemp07, ctemp08;
-  FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
+  IFLOAT ctemp09, ctemp10, ctemp11, ctemp12;
-  FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
+  IFLOAT ctemp13, ctemp14, ctemp15, ctemp16;
-  FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
+  IFLOAT ctemp17, ctemp18, ctemp19, ctemp20;
-  FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
+  IFLOAT ctemp21, ctemp22, ctemp23, ctemp24;
-  FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
+  IFLOAT ctemp25, ctemp26, ctemp27, ctemp28;
-  FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
+  IFLOAT ctemp29, ctemp30, ctemp31, ctemp32;
 
   aoffset   = a;
   boffset   = b;

kernel/generic/gemm_tcopy_2.c

@@ -39,11 +39,11 @@
 #include <stdio.h>
 #include "common.h"
 
-int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
+int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
   BLASLONG i, j;
 
-  FLOAT *a_offset, *a_offset1, *a_offset2;
+  IFLOAT *a_offset, *a_offset1, *a_offset2;
-  FLOAT *b_offset, *b_offset1, *b_offset2;
+  IFLOAT *b_offset, *b_offset1, *b_offset2;
 
   a_offset = a;
   b_offset = b;

kernel/generic/gemm_tcopy_8.c

@@ -39,32 +39,32 @@
 #include <stdio.h>
 #include "common.h"
 
-int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
+int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
 
   BLASLONG i, j;
 
-  FLOAT *aoffset;
+  IFLOAT *aoffset;
-  FLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
+  IFLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
-  FLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
+  IFLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
 
-  FLOAT *boffset,  *boffset1, *boffset2, *boffset3, *boffset4;
+  IFLOAT *boffset,  *boffset1, *boffset2, *boffset3, *boffset4;
 
-  FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
+  IFLOAT ctemp01, ctemp02, ctemp03, ctemp04;
-  FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
+  IFLOAT ctemp05, ctemp06, ctemp07, ctemp08;
-  FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
+  IFLOAT ctemp09, ctemp10, ctemp11, ctemp12;
-  FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
+  IFLOAT ctemp13, ctemp14, ctemp15, ctemp16;
-  FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
+  IFLOAT ctemp17, ctemp18, ctemp19, ctemp20;
-  FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
+  IFLOAT ctemp21, ctemp22, ctemp23, ctemp24;
-  FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
+  IFLOAT ctemp25, ctemp26, ctemp27, ctemp28;
-  FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
+  IFLOAT ctemp29, ctemp30, ctemp31, ctemp32;
-  FLOAT ctemp33, ctemp34, ctemp35, ctemp36;
+  IFLOAT ctemp33, ctemp34, ctemp35, ctemp36;
-  FLOAT ctemp37, ctemp38, ctemp39, ctemp40;
+  IFLOAT ctemp37, ctemp38, ctemp39, ctemp40;
-  FLOAT ctemp41, ctemp42, ctemp43, ctemp44;
+  IFLOAT ctemp41, ctemp42, ctemp43, ctemp44;
-  FLOAT ctemp45, ctemp46, ctemp47, ctemp48;
+  IFLOAT ctemp45, ctemp46, ctemp47, ctemp48;
-  FLOAT ctemp49, ctemp50, ctemp51, ctemp52;
+  IFLOAT ctemp49, ctemp50, ctemp51, ctemp52;
-  FLOAT ctemp53, ctemp54, ctemp55, ctemp56;
+  IFLOAT ctemp53, ctemp54, ctemp55, ctemp56;
-  FLOAT ctemp57, ctemp58, ctemp59, ctemp60;
+  IFLOAT ctemp57, ctemp58, ctemp59, ctemp60;
-  FLOAT ctemp61, ctemp62, ctemp63, ctemp64;
+  IFLOAT ctemp61, ctemp62, ctemp63, ctemp64;
 
   aoffset   = a;
   boffset   = b;

kernel/generic/gemmkernel_2x2.c

@@ -1,13 +1,32 @@
 #include "common.h"
-int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FLOAT* C,BLASLONG ldc
+#if defined(HALF) && defined(HALFCONVERSION)
+static float
+bfloat16tof32 (bfloat16 f16)
+{
+  float result = 0;
+  unsigned short* q = (unsigned short*)(&result);
+#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+  q[0] = f16;
+#else
+  q[1] = f16;
+#endif
+  return result;
+}
+#define BF16TOF32(x) (bfloat16tof32(x))
+#else
+#define BF16TOF32(x) x
+#endif
+int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,IFLOAT* ba,IFLOAT* bb,FLOAT* C,BLASLONG ldc
 #ifdef TRMMKERNEL
 		,BLASLONG offset
 #endif
 		)
 {
    BLASLONG i,j,k;
-   FLOAT *C0,*C1,*ptrba,*ptrbb;
+   FLOAT *C0,*C1;
-   FLOAT res0,res1,res2,res3,load0,load1,load2,load3,load4,load5,load6,load7;
+   IFLOAT *ptrba,*ptrbb;
+   FLOAT res0,res1,res2,res3;
+   IFLOAT load0,load1,load2,load3,load4,load5,load6,load7;
    for (j=0; j<bn/2; j+=1)
      {
         C0 = C;
@@ -24,36 +43,36 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
                {
                   load0 = ptrba[2*0+0];
                   load1 = ptrbb[2*0+0];
-                  res0 = res0+load0*load1;
+                  res0 = res0+BF16TOF32(load0)*BF16TOF32(load1);
                   load2 = ptrba[2*0+1];
-                  res1 = res1+load2*load1;
+                  res1 = res1+BF16TOF32(load2)*BF16TOF32(load1);
                   load3 = ptrbb[2*0+1];
-                  res2 = res2+load0*load3;
+                  res2 = res2+BF16TOF32(load0)*BF16TOF32(load3);
-                  res3 = res3+load2*load3;
+                  res3 = res3+BF16TOF32(load2)*BF16TOF32(load3);
                   load4 = ptrba[2*1+0];
                   load5 = ptrbb[2*1+0];
-                  res0 = res0+load4*load5;
+                  res0 = res0+BF16TOF32(load4)*BF16TOF32(load5);
                   load6 = ptrba[2*1+1];
-                  res1 = res1+load6*load5;
+                  res1 = res1+BF16TOF32(load6)*BF16TOF32(load5);
                   load7 = ptrbb[2*1+1];
-                  res2 = res2+load4*load7;
+                  res2 = res2+BF16TOF32(load4)*BF16TOF32(load7);
-                  res3 = res3+load6*load7;
+                  res3 = res3+BF16TOF32(load6)*BF16TOF32(load7);
                   load0 = ptrba[2*2+0];
                   load1 = ptrbb[2*2+0];
-                  res0 = res0+load0*load1;
+                  res0 = res0+BF16TOF32(load0)*BF16TOF32(load1);
                   load2 = ptrba[2*2+1];
-                  res1 = res1+load2*load1;
+                  res1 = res1+BF16TOF32(load2)*BF16TOF32(load1);
                   load3 = ptrbb[2*2+1];
-                  res2 = res2+load0*load3;
+                  res2 = res2+BF16TOF32(load0)*BF16TOF32(load3);
-                  res3 = res3+load2*load3;
+                  res3 = res3+BF16TOF32(load2)*BF16TOF32(load3);
                   load4 = ptrba[2*3+0];
                   load5 = ptrbb[2*3+0];
-                  res0 = res0+load4*load5;
+                  res0 = res0+BF16TOF32(load4)*BF16TOF32(load5);
                   load6 = ptrba[2*3+1];
-                  res1 = res1+load6*load5;
+                  res1 = res1+BF16TOF32(load6)*BF16TOF32(load5);
                   load7 = ptrbb[2*3+1];
-                  res2 = res2+load4*load7;
+                  res2 = res2+BF16TOF32(load4)*BF16TOF32(load7);
-                  res3 = res3+load6*load7;
+                  res3 = res3+BF16TOF32(load6)*BF16TOF32(load7);
                   ptrba = ptrba+8;
                   ptrbb = ptrbb+8;
                }
@@ -61,12 +80,12 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
                {
                   load0 = ptrba[2*0+0];
                   load1 = ptrbb[2*0+0];
-                  res0 = res0+load0*load1;
+                  res0 = res0+BF16TOF32(load0)*BF16TOF32(load1);
                   load2 = ptrba[2*0+1];
-                  res1 = res1+load2*load1;
+                  res1 = res1+BF16TOF32(load2)*BF16TOF32(load1);
                   load3 = ptrbb[2*0+1];
-                  res2 = res2+load0*load3;
+                  res2 = res2+BF16TOF32(load0)*BF16TOF32(load3);
-                  res3 = res3+load2*load3;
+                  res3 = res3+BF16TOF32(load2)*BF16TOF32(load3);
                   ptrba = ptrba+2;
                   ptrbb = ptrbb+2;
                }
@@ -90,9 +109,9 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
                {
                   load0 = ptrba[0+0];
                   load1 = ptrbb[2*0+0];
-                  res0 = res0+load0*load1;
+                  res0 = res0+BF16TOF32(load0)*BF16TOF32(load1);
                   load2 = ptrbb[2*0+1];
-                  res1 = res1+load0*load2;
+                  res1 = res1+BF16TOF32(load0)*BF16TOF32(load2);
                   ptrba = ptrba+1;
                   ptrbb = ptrbb+2;
                }
@@ -121,9 +140,9 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
                {
                   load0 = ptrba[2*0+0];
                   load1 = ptrbb[0+0];
-                  res0 = res0+load0*load1;
+                  res0 = res0+BF16TOF32(load0)*BF16TOF32(load1);
                   load2 = ptrba[2*0+1];
-                  res1 = res1+load2*load1;
+                  res1 = res1+BF16TOF32(load2)*BF16TOF32(load1);
                   ptrba = ptrba+2;
                   ptrbb = ptrbb+1;
                }
@@ -141,7 +160,7 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
                {
                   load0 = ptrba[0+0];
                   load1 = ptrbb[0+0];
-                  res0 = res0+load0*load1;
+                  res0 = res0+BF16TOF32(load0)*BF16TOF32(load1);
                   ptrba = ptrba+1;
                   ptrbb = ptrbb+1;
                }

kernel/mips/KERNEL.MIPS24K

@@ -0,0 +1 @@
+include $(KERNELDIR)/KERNEL.P5600

kernel/power/KERNEL.POWER10

@@ -0,0 +1,225 @@
+ifeq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__)
+include $(KERNELDIR)/KERNEL.POWER8
+else
+
+#SGEMM_BETA = ../generic/gemm_beta.c
+#DGEMM_BETA = ../generic/gemm_beta.c
+#CGEMM_BETA = ../generic/zgemm_beta.c
+#ZGEMM_BETA = ../generic/zgemm_beta.c
+
+SHGEMM_BETA = ../generic/gemm_beta.c
+SHGEMMKERNEL    = shgemm_kernel_power10.c
+SHGEMMINCOPY    = ../generic/gemm_ncopy_16.c
+SHGEMMITCOPY    = ../generic/gemm_tcopy_16.c
+SHGEMMONCOPY    = ../generic/gemm_ncopy_8.c
+SHGEMMOTCOPY    = ../generic/gemm_tcopy_8.c
+SHGEMMINCOPYOBJ =  shgemm_incopy$(TSUFFIX).$(SUFFIX)
+SHGEMMITCOPYOBJ =  shgemm_itcopy$(TSUFFIX).$(SUFFIX)
+SHGEMMONCOPYOBJ =  shgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SHGEMMOTCOPYOBJ =  shgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+STRMMKERNEL	= sgemm_kernel_power10.c
+DTRMMKERNEL	= dgemm_kernel_power10.c
+CTRMMKERNEL	= cgemm_kernel_power10.S
+ZTRMMKERNEL	= zgemm_kernel_power10.S
+
+SGEMMKERNEL    =  sgemm_kernel_power10.c
+SGEMMINCOPY    = ../generic/gemm_ncopy_16.c
+SGEMMITCOPY    = sgemm_tcopy_16_power8.S
+SGEMMONCOPY    =  ../generic/gemm_ncopy_8.c
+SGEMMOTCOPY    = sgemm_tcopy_8_power8.S
+SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
+SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+DGEMMKERNEL    =  dgemm_kernel_power10.c
+DGEMMINCOPY    = ../generic/gemm_ncopy_16.c
+DGEMMITCOPY    =  dgemm_tcopy_16_power8.S
+DGEMMONCOPY    =  dgemm_ncopy_4_power8.S
+DGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
+DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+CGEMMKERNEL    = cgemm_kernel_power10.S
+CGEMMINCOPY    = ../generic/zgemm_ncopy_8.c
+CGEMMITCOPY    = ../generic/zgemm_tcopy_8.c
+CGEMMONCOPY    = ../generic/zgemm_ncopy_4.c
+CGEMMOTCOPY    = ../generic/zgemm_tcopy_4.c
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
+CGEMMINCOPYOBJ =  cgemm_incopy$(TSUFFIX).$(SUFFIX)
+CGEMMITCOPYOBJ =  cgemm_itcopy$(TSUFFIX).$(SUFFIX)
+
+ZGEMMKERNEL    = zgemm_kernel_power10.S
+ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
+ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
+ZGEMMINCOPY    = ../generic/zgemm_ncopy_8.c
+ZGEMMITCOPY    = zgemm_tcopy_8_power8.S
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
+ZGEMMINCOPYOBJ =  zgemm_incopy$(TSUFFIX).$(SUFFIX)
+ZGEMMITCOPYOBJ =  zgemm_itcopy$(TSUFFIX).$(SUFFIX)
+
+STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
+STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c
+STRSMKERNEL_RN	=  ../generic/trsm_kernel_RN.c
+STRSMKERNEL_RT	=  ../generic/trsm_kernel_RT.c
+
+DTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LT	= dtrsm_kernel_LT_16x4_power8.S
+DTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+DTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+CTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+CTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+CTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+CTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+ZTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+ZTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+ZTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+#Todo: CGEMM3MKERNEL should be 4x4 blocksizes.
+#CGEMM3MKERNEL    =  zgemm3m_kernel_8x4_sse3.S
+#ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_sse3.S
+
+#Pure C for other kernels
+#SAMAXKERNEL  = ../arm/amax.c
+#DAMAXKERNEL  = ../arm/amax.c
+#CAMAXKERNEL  = ../arm/zamax.c
+#ZAMAXKERNEL  = ../arm/zamax.c
+#
+#SAMINKERNEL  = ../arm/amin.c
+#DAMINKERNEL  = ../arm/amin.c
+#CAMINKERNEL  = ../arm/zamin.c
+#ZAMINKERNEL  = ../arm/zamin.c
+#
+#SMAXKERNEL   = ../arm/max.c
+#DMAXKERNEL   = ../arm/max.c
+#
+#SMINKERNEL   = ../arm/min.c
+#DMINKERNEL   = ../arm/min.c
+#
+ifneq ($(GCCVERSIONGTEQ9),1)
+ISAMAXKERNEL = isamax_power9.S
+else
+ISAMAXKERNEL = isamax.c
+endif
+IDAMAXKERNEL = idamax.c
+ifneq ($(GCCVERSIONGTEQ9),1)
+ICAMAXKERNEL = icamax_power9.S
+else
+ICAMAXKERNEL = icamax.c
+endif
+IZAMAXKERNEL = izamax.c
+#
+ifneq ($(GCCVERSIONGTEQ9),1)
+ISAMINKERNEL = isamin_power9.S
+else
+ISAMINKERNEL = isamin.c
+endif
+IDAMINKERNEL = idamin.c
+ifneq ($(GCCVERSIONGTEQ9),1)
+ICAMINKERNEL = icamin_power9.S
+else
+ICAMINKERNEL = icamin.c
+endif
+IZAMINKERNEL = izamin.c
+#
+#ISMAXKERNEL  = ../arm/imax.c
+#IDMAXKERNEL  = ../arm/imax.c
+#
+#ISMINKERNEL  = ../arm/imin.c
+#IDMINKERNEL  = ../arm/imin.c
+#
+SASUMKERNEL  = sasum.c
+DASUMKERNEL  = dasum.c
+CASUMKERNEL  = casum.c
+ZASUMKERNEL  = zasum.c
+#
+SAXPYKERNEL  = saxpy.c
+DAXPYKERNEL  = daxpy.c
+ifneq ($(GCCVERSIONGTEQ9),1)
+CAXPYKERNEL  = caxpy_power9.S
+else
+CAXPYKERNEL  = caxpy.c
+endif
+ZAXPYKERNEL  = zaxpy.c
+#
+SCOPYKERNEL  = scopy.c
+DCOPYKERNEL  = dcopy.c
+CCOPYKERNEL  = ccopy.c
+ZCOPYKERNEL  = zcopy.c
+#
+SDOTKERNEL   =  sdot.c
+DDOTKERNEL   =  ddot.c
+DSDOTKERNEL  =  sdot.c
+ifneq ($(GCCVERSIONGTEQ9),1)
+CDOTKERNEL   =  cdot_power9.S
+else
+CDOTKERNEL   =  cdot.c
+endif
+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
+
+endif

kernel/power/KERNEL.POWER8

@@ -232,3 +232,11 @@ QCABS_KERNEL	= ../generic/cabs.c
 #Dump kernel
 CGEMM3MKERNEL    = ../generic/zgemm3mkernel_dump.c
 ZGEMM3MKERNEL    = ../generic/zgemm3mkernel_dump.c
+
+ifeq ($(__BYTE_ORDER__)$(ELF_VERSION),__ORDER_BIG_ENDIAN__2)
+IDAMAXKERNEL  = ../arm/iamax.c
+IDAMINKERNEL  = ../arm/iamin.c
+IZAMAXKERNEL  = ../arm/izamax.c
+IZAMINKERNEL  = ../arm/izamin.c
+endif
+

kernel/power/KERNEL.PPCG4

@@ -20,8 +20,10 @@ ZAXPYKERNEL	= zaxpy_ppc440.S
 
 SDOTKERNEL	=  dot_ppc440.S
 DDOTKERNEL	=  dot_ppc440.S
-CDOTKERNEL	= zdot_ppc440.S
+#CDOTKERNEL	= zdot_ppc440.S
-ZDOTKERNEL	= zdot_ppc440.S
+#ZDOTKERNEL	= zdot_ppc440.S
+CDOTKERNEL	= ../arm/zdot.c
+ZDOTKERNEL	= ../arm/zdot.c
 
 ISAMAXKERNEL	=  iamax_ppc440.S
 IDAMAXKERNEL	=  iamax_ppc440.S
@@ -52,8 +54,11 @@ ZNRM2KERNEL	= znrm2_ppc440.S
 
 SROTKERNEL	=  rot_ppc440.S
 DROTKERNEL	=  rot_ppc440.S
-CROTKERNEL	= zrot_ppc440.S
+#CROTKERNEL	= zrot_ppc440.S
-ZROTKERNEL	= zrot_ppc440.S
+#ZROTKERNEL	= zrot_ppc440.S
+CROTKERNEL	= ../arm/zrot.c
+ZROTKERNEL	= ../arm/zrot.c
+
 
 SSCALKERNEL	=  scal_ppc440.S
 DSCALKERNEL	=  scal_ppc440.S
@@ -78,13 +83,18 @@ DGEMMINCOPYOBJ =
 DGEMMITCOPYOBJ =
 DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
 DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
-CGEMMKERNEL    =  zgemm_kernel_altivec_g4.S
+#CGEMMKERNEL    =  zgemm_kernel_altivec_g4.S
-CGEMMINCOPY    =  ../generic/zgemm_ncopy_8.c
+#CGEMMINCOPY    =  ../generic/zgemm_ncopy_8.c
-CGEMMITCOPY    =  ../generic/zgemm_tcopy_8.c
+#CGEMMITCOPY    =  ../generic/zgemm_tcopy_8.c
+CGEMMKERNEL    =  zgemm_kernel.S
+CGEMMINCOPY    =
+CGEMMONCOPY    =
 CGEMMONCOPY    =  ../generic/zgemm_ncopy_2.c
 CGEMMOTCOPY    =  ../generic/zgemm_tcopy_2.c
-CGEMMINCOPYOBJ =  cgemm_incopy$(TSUFFIX).$(SUFFIX)
+CGEMMINCOPYOBJ =  
-CGEMMITCOPYOBJ =  cgemm_itcopy$(TSUFFIX).$(SUFFIX)
+#cgemm_incopy$(TSUFFIX).$(SUFFIX)
+CGEMMITCOPYOBJ =  
+#cgemm_itcopy$(TSUFFIX).$(SUFFIX)
 CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
 CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
 ZGEMMKERNEL    =  zgemm_kernel_g4.S

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)  || defined(POWER9)
+#if defined(POWER8)  || defined(POWER9) || defined(POWER10)
 #include "casum_microk_power8.c"
 #endif
 

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) || defined(POWER9)
+#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #include "ccopy_microk_power8.c"
 #endif
 

kernel/power/cgemm_kernel_8x4_power8.S

@@ -424,7 +424,7 @@ L999:
 	lwz	r16,  204(SP)
 	lwz	r15,  208(SP)
 	lwz	r14,  212(SP)
-        addi    r11, 224
+        addi    r11, SP, 224
 #endif
         lvx     v20, r11, r0
         addi    r11, r11, 16
@@ -459,4 +459,4 @@ L999:
 	blr
 
 	EPILOGUE
-#endif^
+#endif

kernel/power/cgemm_kernel_power10.S

@@ -0,0 +1,286 @@
+/***************************************************************************
+Copyright (c) 2013-2020, 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 FLINK_SAVE (STACKSIZE+16) /* 16($r12) */  
+#define	M	r3
+#define	N	r4
+#define	K	r5
+
+
+#define A	r8
+#define	B	r9
+#define	C	r10
+#define	LDC	r6
+#define OFFSET	r7
+
+
+#define alpha_r vs51
+#define alpha_i vs55
+#define save_permute_1 vs59
+#define permute_mask vs63
+#define o0	0
+ 
+
+#define T1	r11
+#define T2	r12
+#define T3	r14
+#define T4	r15
+#define T5	r16
+#define T6	r17
+#define L	r18
+#define T7	r19
+#define T8	r20
+#define TEMP_REG	r21
+#define	I	r22
+#define J	r23
+#define AO	r24
+#define	BO	r25
+#define	CO 	r26
+#define T9	r27
+#define	T10	r28
+#define	PRE	r29
+
+#define T12	r30
+#define T13	r31
+
+#include "cgemm_macros_power10.S"
+
+.equ    perm_const1, 0x0405060700010203
+.equ    perm_const2, 0x0c0d0e0f08090a0b
+.equ save_permute_12, 0x0c0d0e0f1c1d1e1f
+.equ save_permute_11, 0x0405060714151617
+
+
+
+#ifndef NEEDPARAM
+
+	PROLOGUE
+	PROFCODE
+
+
+	addi	SP, SP, -STACKSIZE
+	mflr r0
+
+
+	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    vs52,  288(SP)
+  stxv    vs53,  304(SP)
+  stxv    vs54,  320(SP)
+  stxv    vs55,  336(SP)
+  stxv    vs56,  352(SP)
+  stxv    vs57,  368(SP)
+  stxv    vs58,  384(SP)
+  stxv    vs59,  400(SP)
+  stxv    vs60,  416(SP)
+  stxv    vs61,  432(SP)
+  stxv    vs62,  448(SP)
+  stxv    vs63,  464(SP)
+  std     r0,   FLINK_SAVE(SP)
+ 
+
+
+	ld	LDC, FRAMESLOT(0) + STACKSIZE(SP)
+
+
+
+#ifdef TRMMKERNEL
+	ld	OFFSET,  FRAMESLOT(1) + STACKSIZE(SP)
+#endif
+   slwi    LDC, LDC, ZBASE_SHIFT
+
+ 
+ 
+	/*alpha is stored in f1. convert to single and splat*/
+    xscvdpspn alpha_r,vs1 
+    xscvdpspn alpha_i,vs2 
+	xxspltw   alpha_r,alpha_r,0 
+	xxspltw   alpha_i,alpha_i,0 
+/*load reverse permute mask for big endian
+  uint128 = 0xc0d0e0f08090a0b0405060700010203
+*/ 
+		
+	lis T2, perm_const2@highest
+	lis T1, perm_const1@highest
+	lis T3, save_permute_12@highest
+	lis T4, save_permute_11@highest
+
+	
+	ori T2, T2, perm_const2@higher
+	ori T1, T1, perm_const1@higher
+	ori T3, T3, save_permute_12@higher
+	ori T4, T4, save_permute_11@higher
+
+	
+	rldicr T2, T2, 32, 31
+	rldicr T1, T1, 32, 31
+	rldicr T3, T3, 32, 31
+	rldicr T4, T4, 32, 31 
+
+	oris T2, T2, perm_const2@h
+	oris T1, T1, perm_const1@h
+	oris T3, T3, save_permute_12@h
+	oris T4, T4, save_permute_11@h
+
+	
+	ori T2, T2, perm_const2@l  
+	ori T1, T1, perm_const1@l
+	ori T3, T3, save_permute_12@l  
+	ori T4, T4, save_permute_11@l
+
+	
+  li r0,0
+  li PRE,512
+
+#if defined(CC) || defined(CR) || defined(RC) || defined(RR) 
+/*negate for this case as we will use addition -1*(a+b) */
+  xvnegsp alpha_r,alpha_r
+  xvnegsp alpha_i,alpha_i
+#endif
+
+	mtvsrdd permute_mask,T2,T1
+	mtvsrdd save_permute_1,T3,T4 	
+
+     /*mask is reverse permute so we have to make it inner permute */
+ 	xxpermdi	permute_mask,	permute_mask,	permute_mask,2 
+
+#include "cgemm_logic_power10.S"
+
+.L999: 
+	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)
+
+	ld    r0, 	 FLINK_SAVE(SP)	
+ 
+    lxv    vs52,  288(SP)
+    lxv    vs53,  304(SP)
+    lxv    vs54,  320(SP)
+    lxv    vs55,  336(SP)
+    lxv    vs56,  352(SP)
+    lxv    vs57,  368(SP)
+    lxv    vs58,  384(SP) 
+    lxv    vs59,  400(SP)
+	mtlr r0
+    lxv    vs60,  416(SP)
+    lxv    vs61,  432(SP) 
+    lxv    vs62,  448(SP)
+    lxv    vs63,  464(SP)
+
+	addi	SP, SP, STACKSIZE 
+	blr
+
+
+	EPILOGUE
+#endif

kernel/power/crot.c

@@ -27,7 +27,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  
 #include "common.h"
  
-#if defined(POWER8) || defined(POWER9)
+#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 
 static void crot_kernel_8 (long n, float *x, float *y, float c, float s)
 {

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)  || defined(POWER9)
+#if defined(POWER8)  || defined(POWER9) || defined(POWER10)
 #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) || defined(POWER9)
+#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #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) || defined(POWER9)
+#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #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) || defined(POWER9)
+#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #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)  || defined(POWER9)
+#if defined(POWER8)  || defined(POWER9) || defined(POWER10)
 #include "ddot_microk_power8.c"
 #endif
 

kernel/power/dgemm_kernel_power10.c

@@ -0,0 +1,864 @@
+/*********************************************************************************
+Copyright (c) 2020, 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 <altivec.h>
+
+typedef unsigned char vec_t __attribute__ ((vector_size (16)));
+typedef FLOAT v4sf_t __attribute__ ((vector_size (16)));
+typedef FLOAT v2sf_t __attribute__ ((vector_size (8)));
+
+#ifdef TRMMKERNEL
+#define SAVE_ACC(ACC, J)  \
+          __builtin_mma_disassemble_acc (result, ACC); \
+          rowC = (v4sf_t *) &CO[0* ldc+J]; \
+          rowC[0] = result[3] * alpha; \
+          rowC = (v4sf_t *) &CO[1*ldc+J]; \
+          rowC[0] = result[2] * alpha; \
+          rowC = (v4sf_t *) &CO[2*ldc+J]; \
+          rowC[0] = result[1] * alpha; \
+          rowC = (v4sf_t *) &CO[3*ldc+J]; \
+          rowC[0] = result[0] * alpha;
+#define SAVE_ACC1(ACC, J)  \
+          __builtin_mma_disassemble_acc (result, ACC); \
+          rowC = (v4sf_t *) &CO[4* ldc+J]; \
+          rowC[0] = result[3] * alpha; \
+          rowC = (v4sf_t *) &CO[5*ldc+J]; \
+          rowC[0] = result[2] * alpha; \
+          rowC = (v4sf_t *) &CO[6*ldc+J]; \
+          rowC[0] = result[1] * alpha; \
+          rowC = (v4sf_t *) &CO[7*ldc+J]; \
+          rowC[0] = result[0] * alpha;
+#define  SAVE2x4_ACC(ACC, J)  \
+	  __builtin_mma_disassemble_acc (result, ACC); \
+	  rowC = (v4sf_t *) &CO[0* ldc+J]; \
+          rowC[0] = result[3] * alpha; \
+	  rowC = (v4sf_t *) &CO[1* ldc+J]; \
+          rowC[0] = result[2] * alpha;
+#else
+#define SAVE_ACC(ACC, J)  \
+          __builtin_mma_disassemble_acc (result, ACC); \
+          rowC = (v4sf_t *) &CO[0* ldc+J]; \
+          rowC[0] += result[3] * alpha; \
+          rowC = (v4sf_t *) &CO[1*ldc+J]; \
+          rowC[0] += result[2] * alpha; \
+          rowC = (v4sf_t *) &CO[2*ldc+J]; \
+          rowC[0] += result[1] * alpha; \
+          rowC = (v4sf_t *) &CO[3*ldc+J]; \
+          rowC[0] += result[0] * alpha;
+#define SAVE_ACC1(ACC, J)  \
+          __builtin_mma_disassemble_acc (result, ACC); \
+          rowC = (v4sf_t *) &CO[4* ldc+J]; \
+          rowC[0] += result[3] * alpha; \
+          rowC = (v4sf_t *) &CO[5*ldc+J]; \
+          rowC[0] += result[2] * alpha; \
+          rowC = (v4sf_t *) &CO[6*ldc+J]; \
+          rowC[0] += result[1] * alpha; \
+          rowC = (v4sf_t *) &CO[7*ldc+J]; \
+          rowC[0] += result[0] * alpha;
+#define  SAVE2x4_ACC(ACC, J)  \
+	  __builtin_mma_disassemble_acc (result, ACC); \
+	  rowC = (v4sf_t *) &CO[0* ldc+J]; \
+          rowC[0] += result[3] * alpha; \
+	  rowC = (v4sf_t *) &CO[1* ldc+J]; \
+          rowC[0] += result[2] * alpha;
+#endif
+
+#define SET_ACC_ZERO4() \
+          __builtin_mma_xxsetaccz (&acc0); \
+          __builtin_mma_xxsetaccz (&acc1); \
+          __builtin_mma_xxsetaccz (&acc2); \
+          __builtin_mma_xxsetaccz (&acc3);
+
+#define SET_ACC_ZERO8() \
+          __builtin_mma_xxsetaccz (&acc0); \
+          __builtin_mma_xxsetaccz (&acc1); \
+          __builtin_mma_xxsetaccz (&acc2); \
+          __builtin_mma_xxsetaccz (&acc3); \
+          __builtin_mma_xxsetaccz (&acc4); \
+          __builtin_mma_xxsetaccz (&acc5); \
+          __builtin_mma_xxsetaccz (&acc6); \
+          __builtin_mma_xxsetaccz (&acc7);
+
+#define PREFETCH1(x, y) asm volatile ("dcbt %0, %1" : : "r" (x), "b" (y) : "memory");
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+#define REFRESH_TEMP_BK(x, y) \
+            temp = k - off;
+#elif defined(LEFT)
+#define REFRESH_TEMP_BK(x, y) \
+            temp = off + x;
+#else
+#define REFRESH_TEMP_BK(x, y) \
+            temp = off + y;
+#endif
+#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+#define REFRESH_POINTERS(x, y) \
+          BO = B; \
+          REFRESH_TEMP_BK(x, y)
+#else
+#define REFRESH_POINTERS(x, y) \
+          AO += off * x; \
+          BO = B + off * y; \
+          REFRESH_TEMP_BK(x, y)
+#endif
+
+#ifdef LEFT
+#define REFRESH_OFF(x) \
+            off += x;
+#else
+#define REFRESH_OFF(x)
+#endif
+
+#ifdef LEFT
+#define UPDATE_TEMP(x, y) \
+            temp -= x;
+#else
+#define UPDATE_TEMP(x, y) \
+            temp -= y;
+#endif
+
+#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+#define REFRESH_TMP_AFTER_SAVE(x, y) \
+            temp = k - off; \
+            UPDATE_TEMP(x, y) \
+            AO += temp * x; \
+            BO += temp * y;
+#else
+#define REFRESH_TMP_AFTER_SAVE(x, y)
+#endif
+
+#define REFRESH_AFTER_SAVE(x,y) \
+        REFRESH_TMP_AFTER_SAVE(x, y) \
+        REFRESH_OFF(x)
+/*************************************************************************************
+* GEMM Kernel
+*************************************************************************************/
+int
+CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
+       FLOAT * C, BLASLONG ldc
+#ifdef TRMMKERNEL
+       , BLASLONG offset
+#endif
+  )
+{
+  BLASLONG N = n;
+  BLASLONG i1;
+#if defined(TRMMKERNEL)
+  BLASLONG off;
+#endif
+#if defined(TRMMKERNEL) && !defined(LEFT)
+  off = -offset;
+#endif
+  v4sf_t valpha = { alpha, alpha };
+  N = n >> 2;
+  for (i1 = 0; i1 < N; i1++)
+    {
+      BLASLONG i, j, temp;
+      FLOAT *CO;
+      FLOAT *AO;
+#if defined(TRMMKERNEL) && defined(LEFT)
+      off = offset;
+#endif
+      CO = C;
+      C += ldc << 2;
+      AO = A;
+      PREFETCH1 (A, 128);
+      PREFETCH1 (A, 256);
+      i = m >> 4;
+      for (j = 0; j < i; j++)
+	{
+          FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (16, 4);
+#else
+          BO = B;
+          temp = k;
+#endif
+	  v4sf_t *rowC;
+	  v4sf_t result[4];
+	  BLASLONG l = 0;
+	  PREFETCH1 (CO, 0);
+	  PREFETCH1 (CO + ldc, 0);
+	  PREFETCH1 (CO + ldc + ldc, 0);
+	  PREFETCH1 (CO + ldc + ldc + ldc, 0);
+	  PREFETCH1 (CO, 128);
+	  PREFETCH1 (CO + ldc, 128);
+	  PREFETCH1 (CO + ldc + ldc, 128);
+	  PREFETCH1 (CO + ldc + ldc + ldc, 128);
+	  __vector_quad acc0, acc1, acc2, acc3, acc4, acc5, acc6, acc7;
+	  SET_ACC_ZERO8 ();
+	  for (l = 0; l < temp; l++)
+	    {
+	      vec_t *rowA = (vec_t *) & AO[l << 4];
+	      __vector_pair rowB;
+	      vec_t *rb = (vec_t *) & BO[l << 2];
+	      __builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
+	      __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
+	      __builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
+	      __builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]);
+	      __builtin_mma_xvf64gerpp (&acc3, rowB, rowA[3]);
+	      __builtin_mma_xvf64gerpp (&acc4, rowB, rowA[4]);
+	      __builtin_mma_xvf64gerpp (&acc5, rowB, rowA[5]);
+	      __builtin_mma_xvf64gerpp (&acc6, rowB, rowA[6]);
+	      __builtin_mma_xvf64gerpp (&acc7, rowB, rowA[7]);
+	    }
+	  SAVE_ACC (&acc0, 0);
+	  SAVE_ACC (&acc2, 4);
+	  SAVE_ACC (&acc1, 2);
+	  SAVE_ACC (&acc3, 6);
+	  SAVE_ACC (&acc4, 8);
+	  SAVE_ACC (&acc6, 12);
+	  SAVE_ACC (&acc5, 10);
+	  SAVE_ACC (&acc7, 14);
+	  AO += temp << 4;
+	  BO += temp << 2;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (16, 4)
+#endif
+	  CO += 16;
+	}
+      i = (m & 15) >> 3;
+      for (j = 0; j < i; j++)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (8, 4);
+#else
+          BO = B;
+          temp = k;
+#endif
+	  v4sf_t *rowC;
+	  v4sf_t result[4];
+	  __vector_quad acc0, acc1, acc2, acc3;
+	  SET_ACC_ZERO4 ();
+	  BLASLONG l = 0;
+	  for (l = 0; l < temp; l++)
+	    {
+	      vec_t *rowA = (vec_t *) & AO[l << 3];
+	      __vector_pair rowB;
+	      vec_t *rb = (vec_t *) & BO[l << 2];
+	      __builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
+	      __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
+	      __builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
+	      __builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]);
+	      __builtin_mma_xvf64gerpp (&acc3, rowB, rowA[3]);
+	    }
+	  SAVE_ACC (&acc0, 0);
+	  SAVE_ACC (&acc2, 4);
+	  SAVE_ACC (&acc1, 2);
+	  SAVE_ACC (&acc3, 6);
+	  CO += 8;
+	  AO += temp << 3;
+	  BO += temp << 2;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (8, 4)
+#endif
+	}
+      i = (m & 7) >> 2;
+      for (j = 0; j < i; j++)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (4, 4);
+#else
+          BO = B;
+          temp = k;
+#endif
+	  v4sf_t *rowC;
+	  v4sf_t result[4];
+	  __vector_quad acc0, acc1;
+	  __builtin_mma_xxsetaccz (&acc0);
+	  __builtin_mma_xxsetaccz (&acc1);
+	  BLASLONG l = 0;
+	  for (l = 0; l < temp; l++)
+	    {
+	      vec_t *rowA = (vec_t *) & AO[l << 2];
+	      __vector_pair rowB;
+	      vec_t *rb = (vec_t *) & BO[l << 2];
+	      __builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
+	      __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
+	      __builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
+	    }
+	  SAVE_ACC (&acc0, 0);
+	  SAVE_ACC (&acc1, 2);
+	  CO += 4;
+	  AO += temp << 2;
+	  BO += temp << 2;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (4, 4)
+#endif
+	}
+      i = (m & 3) >> 1;
+      for (j = 0; j < i; j++)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (2, 4);
+#else
+          BO = B;
+          temp = k;
+#endif
+	  v4sf_t *rowC;
+	  v4sf_t result[4];
+	  __vector_quad acc0;
+	  __builtin_mma_xxsetaccz (&acc0);
+	  BLASLONG l = 0;
+	  for (l = 0; l < temp; l++)
+	    {
+	      vec_t *rowA = (vec_t *) & AO[l << 1];
+	      __vector_pair rowB;
+	      vec_t *rb = (vec_t *) & BO[l << 2];
+	      __builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
+	      __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
+	    }
+	  SAVE_ACC (&acc0, 0);
+	  CO += 2;
+	  AO += temp << 1;
+	  BO += temp << 2;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (2, 4)
+#endif
+	}
+      i = (m & 1) >> 0;
+      for (j = 0; j < i; j++)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (1, 4);
+#else
+          BO = B;
+          temp = k;
+#endif
+	  BLASLONG l = 0;
+	  v4sf_t t = { 0, 0 };
+	  v4sf_t t1 = { 0, 0 };
+	  for (l = 0; l < temp; l++)
+	    {
+	      v4sf_t rowA = { AO[l], AO[l] };
+	      v4sf_t rowB = { BO[l << 2], BO[(l << 2) + 1] };
+	      v4sf_t rowB1 = { BO[(l << 2) + 2], BO[(l << 2) + 3] };
+	      t += rowA * rowB;
+	      t1 += rowA * rowB1;
+	    }
+	  t = t * valpha;
+	  t1 = t1 * valpha;
+#if defined(TRMMKERNEL)
+	  CO[0 * ldc] = t[0];
+	  CO[1 * ldc] = t[1];
+	  CO[2 * ldc] = t1[0];
+	  CO[3 * ldc] = t1[1];
+#else
+	  CO[0 * ldc] += t[0];
+	  CO[1 * ldc] += t[1];
+	  CO[2 * ldc] += t1[0];
+	  CO[3 * ldc] += t1[1];
+#endif
+	  CO += 1;
+	  AO += temp;
+	  BO += temp << 2;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (1, 4)
+#endif
+	}
+#if defined(TRMMKERNEL) && !defined(LEFT)
+      off += 4;                 // number of values in A
+#endif
+      B += k << 2;
+    }
+  N = (n & 3) >> 1;
+  for (i1 = 0; i1 < N; i1++)
+    {
+      BLASLONG i, j, temp;
+#if defined(TRMMKERNEL) && defined(LEFT)
+      off = offset;
+#endif
+      FLOAT *CO;
+      FLOAT *AO;
+      CO = C;
+      C += ldc << 1;
+      AO = A;
+      i = m >> 4;
+      for (j = 0; j < i; j++)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (16, 2);
+#else
+          BO = B;
+          temp = k;
+#endif
+	  v4sf_t *rowC;
+	  v4sf_t result[4];
+	  __vector_quad acc0, acc1, acc2, acc3, acc4, acc5, acc6, acc7;
+	  SET_ACC_ZERO8 ();
+	  BLASLONG l = 0;
+	  for (l = 0; l < temp; l++)
+	    {
+	      FLOAT t[4] = { 0, 0, 0, 0 };
+	      t[0] = BO[l << 1], t[1] = BO[(l << 1) + 1];
+	      __vector_pair rowB;
+	      vec_t *rb = (vec_t *) & t[0];
+	      __builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
+	      vec_t *rowA = (vec_t *) & AO[l << 4];
+	      __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
+	      __builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
+	      __builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]);
+	      __builtin_mma_xvf64gerpp (&acc3, rowB, rowA[3]);
+	      __builtin_mma_xvf64gerpp (&acc4, rowB, rowA[4]);
+	      __builtin_mma_xvf64gerpp (&acc5, rowB, rowA[5]);
+	      __builtin_mma_xvf64gerpp (&acc6, rowB, rowA[6]);
+	      __builtin_mma_xvf64gerpp (&acc7, rowB, rowA[7]);
+	    }
+	  SAVE2x4_ACC (&acc0, 0);
+	  SAVE2x4_ACC (&acc1, 2);
+	  SAVE2x4_ACC (&acc2, 4);
+	  SAVE2x4_ACC (&acc3, 6);
+	  SAVE2x4_ACC (&acc4, 8);
+	  SAVE2x4_ACC (&acc5, 10);
+	  SAVE2x4_ACC (&acc6, 12);
+	  SAVE2x4_ACC (&acc7, 14);
+	  CO += 16;
+	  AO += temp << 4;
+	  BO += temp << 1;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (16, 2)
+#endif
+	}
+      i = (m & 15) >> 3;
+      for (j = 0; j < i; j++)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (8, 2);
+#else
+          BO = B;
+          temp = k;
+#endif
+	  v4sf_t *rowC;
+	  v4sf_t result[4];
+	  __vector_quad acc0, acc1, acc2, acc3;
+	  SET_ACC_ZERO4 ();
+	  BLASLONG l = 0;
+	  for (l = 0; l < temp; l++)
+	    {
+	      FLOAT t[4] = { 0, 0, 0, 0 };
+	      t[0] = BO[l << 1], t[1] = BO[(l << 1) + 1];
+	      __vector_pair rowB;
+	      vec_t *rb = (vec_t *) & t[0];
+	      __builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
+	      vec_t *rowA = (vec_t *) & AO[l << 3];
+	      __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
+	      __builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
+	      __builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]);
+	      __builtin_mma_xvf64gerpp (&acc3, rowB, rowA[3]);
+	    }
+	  SAVE2x4_ACC (&acc0, 0);
+	  SAVE2x4_ACC (&acc1, 2);
+	  SAVE2x4_ACC (&acc2, 4);
+	  SAVE2x4_ACC (&acc3, 6);
+	  CO += 8;
+	  AO += temp << 3;
+	  BO += temp << 1;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (8, 2)
+#endif
+	}
+      i = (m & 7) >> 2;
+      for (j = 0; j < i; j++)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (4, 2);
+#else
+          BO = B;
+          temp = k;
+#endif
+	  v4sf_t *rowC;
+	  v4sf_t result[4];
+	  __vector_quad acc0, acc1;
+	  __builtin_mma_xxsetaccz (&acc0);
+	  __builtin_mma_xxsetaccz (&acc1);
+	  BLASLONG l = 0;
+	  for (l = 0; l < temp; l++)
+	    {
+	      FLOAT t[4] = { 0, 0, 0, 0 };
+	      t[0] = BO[l << 1], t[1] = BO[(l << 1) + 1];
+	      __vector_pair rowB;
+	      vec_t *rb = (vec_t *) & t[0];
+	      __builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
+	      vec_t *rowA = (vec_t *) & AO[l << 2];
+	      __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
+	      __builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
+	    }
+	  SAVE2x4_ACC (&acc0, 0);
+	  SAVE2x4_ACC (&acc1, 2);
+	  CO += 4;
+	  AO += temp << 2;
+	  BO += temp << 1;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (4, 2)
+#endif
+	}
+      i = (m & 3) >> 1;
+      for (j = 0; j < i; j++)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (2, 2);
+#else
+          BO = B;
+          temp = k;
+#endif
+	  v4sf_t *rowC;
+	  v4sf_t result[4];
+	  __vector_quad acc0;
+	  __builtin_mma_xxsetaccz (&acc0);
+	  BLASLONG l = 0;
+	  for (l = 0; l < temp; l++)
+	    {
+	      FLOAT t[4] = { 0, 0, 0, 0 };
+	      t[0] = BO[l << 1], t[1] = BO[(l << 1) + 1];
+	      __vector_pair rowB;
+	      vec_t *rb = (vec_t *) & t[0];
+	      __builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
+	      vec_t *rowA = (vec_t *) & AO[l << 1];
+	      __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
+	    }
+	  SAVE2x4_ACC (&acc0, 0);
+	  CO += 2;
+	  AO += temp << 1;
+	  BO += temp << 1;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (2, 2)
+#endif
+	}
+      i = (m & 1) >> 0;
+      for (j = 0; j < i; j++)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (1, 2);
+#else
+          BO = B;
+          temp = k;
+#endif
+	  BLASLONG l = 0;
+	  v4sf_t t = { 0, 0 };
+	  for (l = 0; l < temp; l++)
+	    {
+	      v4sf_t rowA = { AO[l], AO[l] };
+	      v4sf_t rowB = { BO[l << 1], BO[(l << 1) + 1] };
+	      t += rowA * rowB;
+	    }
+	  t = t * valpha;
+#if defined(TRMMKERNEL)
+	  CO[0 * ldc] = t[0];
+	  CO[1 * ldc] = t[1];
+#else
+	  CO[0 * ldc] += t[0];
+	  CO[1 * ldc] += t[1];
+#endif
+	  CO += 1;
+	  AO += temp;
+	  BO += temp << 1;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (1, 2)
+#endif
+	}
+#if defined(TRMMKERNEL) && !defined(LEFT)
+      off += 2;                 // number of values in A
+#endif
+      B += k << 1;
+    }
+  N = (n & 1) >> 0;
+  for (i1 = 0; i1 < N; i1++)
+    {
+      BLASLONG i, temp;
+#if defined(TRMMKERNEL) && defined(LEFT)
+      off = offset;
+#endif
+      FLOAT *CO;
+      FLOAT *AO;
+      CO = C;
+      C += ldc;
+      AO = A;
+      i = m;
+      while (i >= 16)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (16, 1)
+#else
+          BO = B;
+          temp = k;
+#endif
+	  BLASLONG l = 0;
+	  v4sf_t t = { 0, 0 };
+	  v4sf_t t1 = { 0, 0 };
+	  v4sf_t t2 = { 0, 0 };
+	  v4sf_t t3 = { 0, 0 };
+	  v4sf_t t4 = { 0, 0 };
+	  v4sf_t t5 = { 0, 0 };
+	  v4sf_t t6 = { 0, 0 };
+	  v4sf_t t7 = { 0, 0 };
+	  for (l = 0; l < temp; l++)
+	    {
+	      v4sf_t rowB = { BO[l], BO[l] };
+	      v4sf_t rowA = { AO[l << 4], AO[(l << 4) + 1] };
+	      v4sf_t rowA1 = { AO[(l << 4) + 2], AO[(l << 4) + 3] };
+	      v4sf_t rowA2 = { AO[(l << 4) + 4], AO[(l << 4) + 5] };
+	      v4sf_t rowA3 = { AO[(l << 4) + 6], AO[(l << 4) + 7] };
+	      v4sf_t rowA4 = { AO[(l << 4) + 8], AO[(l << 4) + 9] };
+	      v4sf_t rowA5 = { AO[(l << 4) + 10], AO[(l << 4) + 11] };
+	      v4sf_t rowA6 = { AO[(l << 4) + 12], AO[(l << 4) + 13] };
+	      v4sf_t rowA7 = { AO[(l << 4) + 14], AO[(l << 4) + 15] };
+	      t += rowA * rowB;
+	      t1 += rowA1 * rowB;
+	      t2 += rowA2 * rowB;
+	      t3 += rowA3 * rowB;
+	      t4 += rowA4 * rowB;
+	      t5 += rowA5 * rowB;
+	      t6 += rowA6 * rowB;
+	      t7 += rowA7 * rowB;
+	    }
+	  t = t * valpha;
+	  t1 = t1 * valpha;
+	  t2 = t2 * valpha;
+	  t3 = t3 * valpha;
+	  t4 = t4 * valpha;
+	  t5 = t5 * valpha;
+	  t6 = t6 * valpha;
+	  t7 = t7 * valpha;
+#if defined(TRMMKERNEL)
+	  CO[0] = t[0];
+	  CO[1] = t[1];
+	  CO[2] = t1[0];
+	  CO[3] = t1[1];
+	  CO[4] = t2[0];
+	  CO[5] = t2[1];
+	  CO[6] = t3[0];
+	  CO[7] = t3[1];
+	  CO[8] = t4[0];
+	  CO[9] = t4[1];
+	  CO[10] = t5[0];
+	  CO[11] = t5[1];
+	  CO[12] = t6[0];
+	  CO[13] = t6[1];
+	  CO[14] = t7[0];
+	  CO[15] = t7[1];
+#else
+	  CO[0] += t[0];
+	  CO[1] += t[1];
+	  CO[2] += t1[0];
+	  CO[3] += t1[1];
+	  CO[4] += t2[0];
+	  CO[5] += t2[1];
+	  CO[6] += t3[0];
+	  CO[7] += t3[1];
+	  CO[8] += t4[0];
+	  CO[9] += t4[1];
+	  CO[10] += t5[0];
+	  CO[11] += t5[1];
+	  CO[12] += t6[0];
+	  CO[13] += t6[1];
+	  CO[14] += t7[0];
+	  CO[15] += t7[1];
+#endif
+	  AO += temp << 4;
+	  BO += temp;
+	  CO += 16;
+	  i -= 16;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (16, 1)
+#endif
+	}
+      while (i >= 8)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (8, 1)
+#else
+          BO = B;
+          temp = k;
+#endif
+	  BLASLONG l = 0;
+	  v4sf_t t = { 0, 0 };
+	  v4sf_t t1 = { 0, 0 };
+	  v4sf_t t2 = { 0, 0 };
+	  v4sf_t t3 = { 0, 0 };
+	  for (l = 0; l < temp; l++)
+	    {
+	      v4sf_t rowB = { BO[l], BO[l] };
+	      v4sf_t rowA = { AO[l << 3], AO[(l << 3) + 1] };
+	      v4sf_t rowA1 = { AO[(l << 3) + 2], AO[(l << 3) + 3] };
+	      v4sf_t rowA2 = { AO[(l << 3) + 4], AO[(l << 3) + 5] };
+	      v4sf_t rowA3 = { AO[(l << 3) + 6], AO[(l << 3) + 7] };
+	      t += rowA * rowB;
+	      t1 += rowA1 * rowB;
+	      t2 += rowA2 * rowB;
+	      t3 += rowA3 * rowB;
+	    }
+	  t = t * valpha;
+	  t1 = t1 * valpha;
+	  t2 = t2 * valpha;
+	  t3 = t3 * valpha;
+#if defined(TRMMKERNEL)
+	  CO[0] = t[0];
+	  CO[1] = t[1];
+	  CO[2] = t1[0];
+	  CO[3] = t1[1];
+	  CO[4] = t2[0];
+	  CO[5] = t2[1];
+	  CO[6] = t3[0];
+	  CO[7] = t3[1];
+#else
+	  CO[0] += t[0];
+	  CO[1] += t[1];
+	  CO[2] += t1[0];
+	  CO[3] += t1[1];
+	  CO[4] += t2[0];
+	  CO[5] += t2[1];
+	  CO[6] += t3[0];
+	  CO[7] += t3[1];
+#endif
+	  AO += temp << 3;
+	  BO += temp;
+	  CO += 8;
+	  i -= 8;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (8, 1)
+#endif
+	}
+      while (i >= 4)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (4, 1)
+#else
+          BO = B;
+          temp = k;
+#endif
+	  BLASLONG l = 0;
+	  v4sf_t t = { 0, 0 };
+	  v4sf_t t1 = { 0, 0 };
+	  for (l = 0; l < temp; l++)
+	    {
+	      v4sf_t rowB = { BO[l], BO[l] };
+	      v4sf_t rowA = { AO[l << 2], AO[(l << 2) + 1] };
+	      v4sf_t rowA1 = { AO[(l << 2) + 2], AO[(l << 2) + 3] };
+	      t += rowA * rowB;
+	      t1 += rowA1 * rowB;
+	    }
+	  t = t * valpha;
+	  t1 = t1 * valpha;
+#if defined(TRMMKERNEL)
+	  CO[0] = t[0];
+	  CO[1] = t[1];
+	  CO[2] = t1[0];
+	  CO[3] = t1[1];
+#else
+	  CO[0] += t[0];
+	  CO[1] += t[1];
+	  CO[2] += t1[0];
+	  CO[3] += t1[1];
+#endif
+	  AO += temp << 2;
+	  BO += temp;
+	  CO += 4;
+	  i -= 4;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (4, 1)
+#endif
+	}
+      while (i >= 2)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (2, 1)
+#else
+          BO = B;
+          temp = k;
+#endif
+	  BLASLONG l = 0;
+	  v4sf_t t = { 0, 0 };
+	  for (l = 0; l < temp; l++)
+	    {
+	      v4sf_t rowB = { BO[l], BO[l] };
+	      v4sf_t rowA = { AO[l << 1], AO[(l << 1) + 1] };
+	      t += rowA * rowB;
+	    }
+	  t = t * valpha;
+#if defined(TRMMKERNEL)
+	  CO[0] = t[0];
+	  CO[1] = t[1];
+#else
+	  CO[0] += t[0];
+	  CO[1] += t[1];
+#endif
+	  AO += temp << 1;
+	  BO += temp;
+	  CO += 2;
+	  i -= 2;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (2, 1)
+#endif
+	}
+      while (i >= 1)
+	{
+	  FLOAT *BO;
+#if defined(TRMMKERNEL)
+          REFRESH_POINTERS (1, 1)
+#else
+          BO = B;
+          temp = k;
+#endif
+	  BLASLONG l = 0;
+	  FLOAT t = 0;
+	  for (l = 0; l < temp; l++)
+	    {
+	      t += AO[l] * BO[l];
+	    }
+	  AO += temp;
+	  BO += temp;
+#if defined(TRMMKERNEL)
+	  CO[0] = t * alpha;
+#else
+	  CO[0] += t * alpha;
+#endif
+	  CO += 1;
+	  i -= 1;
+#if defined(TRMMKERNEL)
+          REFRESH_AFTER_SAVE (1, 1)
+#endif
+	}
+#if defined(TRMMKERNEL) && !defined(LEFT)
+      off += 1;                 // number of values in A
+#endif
+      B += k;
+    }
+  return 0;
+}

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) || defined(POWER9)
+#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #include "dgemv_n_microk_power8.c"
 #endif
 

kernel/power/drot.c

@@ -39,7 +39,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #pragma GCC optimize "O1"
 
-#if defined(POWER8) || defined(POWER9)
+#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #include "drot_microk_power8.c"
 #endif
 

kernel/power/dscal.c

@@ -35,7 +35,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
-#if defined(POWER8) || defined(POWER9)
+#if defined(POWER8) || defined(POWER9) || defined(POWER10)
 #include "dscal_microk_power8.c"
 #endif