Merge pull request #3488 from xianyi/develop

Update from develop branch for 0.3.19 release

Author: martin-frbg

CMakeLists.txt

@@ -197,3 +197,7 @@ In chronological order:
 
 * River Dillon <oss@outerpassage.net>
   * [2021-07-10] fix compilation with musl libc
+
+* Bine Brank <https://github.com/binebrank>
+  * [2021-10-27] Add vector-length-agnostic DGEMM kernels for Arm SVE
+  * [2021-11-20] Vector-length-agnostic Arm SVE copy routines for DGEMM, DTRMM, DSYMM

CONTRIBUTORS.md

@@ -1,4 +1,51 @@
 OpenBLAS ChangeLog
+====================================================================
+Version 0.3.19
+ 19-Dec-2021
+ 
+ general:
+ - reverted unsafe TRSV/ZRSV optimizations introduced in 0.3.16
+ - fixed a potential thread race in the thread buffer reallocation routines
+   that were introduced in 0.3.18
+ - fixed miscounting of thread pool size on Linux with OMP_PROC_BIND=TRUE
+ - fixed CBLAS interfaces for CSROT/ZSROT and CROTG/ZROTG
+ - made automatic library suffix for CMAKE builds with INTERFACE64 available
+   to CBLAS-only builds
+
+x86_64:
+ - DYNAMIC_ARCH builds now fall back to the cpu with most similar capabilities
+   when an unknown CPUID is encountered, instead of defaulting to Prescott
+ - added cpu detection for Intel Alder Lake
+ - added cpu detection for Intel Sapphire Rapids
+ - added an optimized SBGEMM kernel for Sapphire Rapids
+ - fixed DYNAMIC_ARCH builds on OSX with CMAKE
+ - worked around DYNAMIC_ARCH builds made on Sandybridge failing on SkylakeX
+ - fixed missing thread initialization for static builds on Windows/MSVC
+ - fixed an excessive read in ZSYMV
+
+POWER:
+ - added support for POWER10 in big-endian mode
+ - added support for building with CMAKE
+ - added optimized SGEMM and DGEMM kernels for small matrix sizes
+
+ARMV8:
+ - added basic support and cputype detection for Fujitsu A64FX
+ - added a generic ARMV8SVE target  
+ - added SVE-enabled SGEMM and DGEMM kernels for ARMV8SVE and A64FX
+ - added optimized CGEMM and ZGEMM kernels for Cortex A53 and A55 cpus
+ - fixed cpuid detection for Apple M1 and improved performance
+ - improved compiler flag setting in CMAKE builds
+
+RISCV64:
+ - fixed improper initialization in CSCAL/ZSCAL for strided access patterns
+
+MIPS:
+ - added a GENERIC target for MIPS32
+ - added support for cross-compiling to MIPS32 on x86_64 using CMAKE
+
+MIPS64:
+ - fixed misdetection of MSA capability
+
 ====================================================================
 Version 0.3.18
  02-Oct-2021

Changelog.txt

@@ -32,7 +32,7 @@ export NOFORTRAN
 export NO_LAPACK
 endif
 
-LAPACK_NOOPT := $(filter-out -O0 -O1 -O2 -O3 -Ofast,$(LAPACK_FFLAGS))
+LAPACK_NOOPT := $(filter-out -O0 -O1 -O2 -O3 -Ofast -O -Og -Os,$(LAPACK_FFLAGS))
 
 SUBDIRS_ALL = $(SUBDIRS) test ctest utest exports benchmark ../laswp ../bench cpp_thread_test
 

Makefile

@@ -1,6 +1,9 @@
 ifneq ($(C_COMPILER), PGI)
 
-ifneq ($(GCCVERSIONGT4), 1)
+ifeq ($(C_COMPILER), CLANG)
+ISCLANG=1
+endif
+ifneq (1, $(filter 1,$(GCCVERSIONGT4) $(ISCLANG)))
 CCOMMON_OPT += -march=armv8-a
 ifneq ($(F_COMPILER), NAG)
 FCOMMON_OPT += -march=armv8-a
@@ -17,6 +20,13 @@ FCOMMON_OPT += -march=armv8-a
 endif
 endif
 
+ifeq ($(CORE), ARMV8SVE)
+CCOMMON_OPT += -march=armv8-a+sve
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=armv8-a+sve
+endif
+endif
+
 ifeq ($(CORE), CORTEXA53)
 CCOMMON_OPT += -march=armv8-a -mtune=cortex-a53
 ifneq ($(F_COMPILER), NAG)
@@ -48,7 +58,7 @@ endif
 # Use a72 tunings because Neoverse-N1 is only available
 # in GCC>=9
 ifeq ($(CORE), NEOVERSEN1)
-ifeq ($(GCCVERSIONGTEQ7), 1)
+ifeq (1, $(filter 1,$(GCCVERSIONGTEQ7) $(ISCLANG)))
 ifeq ($(GCCVERSIONGTEQ9), 1)
 CCOMMON_OPT += -march=armv8.2-a -mtune=neoverse-n1
 ifneq ($(F_COMPILER), NAG)
@@ -70,7 +80,7 @@ endif
 
 # Use a53 tunings because a55 is only available in GCC>=8.1
 ifeq ($(CORE), CORTEXA55)
-ifeq ($(GCCVERSIONGTEQ7), 1)
+ifeq (1, $(filter 1,$(GCCVERSIONGTEQ7) $(ISCLANG)))
 ifeq ($(GCCVERSIONGTEQ8), 1)
 CCOMMON_OPT += -march=armv8.2-a -mtune=cortex-a55
 ifneq ($(F_COMPILER), NAG)
@@ -132,7 +142,7 @@ FCOMMON_OPT += -march=armv8.3-a
 endif
 endif
 
-ifeq ($(GCCVERSIONGTEQ9), 1)
+ifeq (1, $(filter 1,$(GCCVERSIONGTEQ9) $(ISCLANG)))
 ifeq ($(CORE), TSV110)
 CCOMMON_OPT += -march=armv8.2-a -mtune=tsv110
 ifneq ($(F_COMPILER), NAG)
@@ -150,6 +160,15 @@ endif
 endif
 endif
 
+ifeq (1, $(filter 1,$(GCCVERSIONGTEQ11) $(ISCLANG)))
+ifeq ($(CORE), A64FX)
+CCOMMON_OPT += -march=armv8.2-a+sve -mtune=a64fx
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=armv8.2-a+sve -mtune=a64fx
+endif
+endif
 endif
 
-endif
+endif
+
+endif

Makefile.arm64

@@ -3,7 +3,7 @@
 #
 
 # This library's version
-VERSION = 0.3.18
+VERSION = 0.3.18.dev
 
 # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a
 # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library

Makefile.rule

@@ -9,11 +9,10 @@ ifndef TOPDIR
 TOPDIR = . 
 endif
 
- # If ARCH is not set, we use the host system's architecture for getarch compile options.
-ifndef ARCH
+# we need to use the host system's architecture for getarch compile options even especially when cross-compiling
 HOSTARCH := $(shell uname -m)
-else
-HOSTARCH = $(ARCH)
+ifeq ($(HOSTARCH), amd64)
+HOSTARCH=x86_64
 endif
 
 # Catch conflicting usage of ARCH in some BSD environments
@@ -102,7 +101,7 @@ GETARCH_FLAGS += -DUSER_TARGET
 ifeq ($(TARGET), GENERIC)
 ifeq ($(DYNAMIC_ARCH), 1)
 override NO_EXPRECISION=1
-export NO_EXPRECiSION
+export NO_EXPRECISION
 endif
 endif
 endif
@@ -119,6 +118,9 @@ endif
 ifeq ($(TARGET), COOPERLAKE)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
+ifeq ($(TARGET), SAPPHIRERAPIDS)
+GETARCH_FLAGS := -DFORCE_NEHALEM
+endif
 ifeq ($(TARGET), SANDYBRIDGE)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
@@ -143,8 +145,13 @@ endif
 ifeq ($(TARGET), POWER8)
 GETARCH_FLAGS := -DFORCE_POWER6
 endif
+ifeq ($(TARGET), POWER9)
+GETARCH_FLAGS := -DFORCE_POWER6
+endif
+ifeq ($(TARGET), POWER10)
+GETARCH_FLAGS := -DFORCE_POWER6
+endif
 endif
-
 
 #TARGET_CORE will override TARGET which is used in DYNAMIC_ARCH=1.
 #
@@ -164,6 +171,9 @@ endif
 ifeq ($(TARGET_CORE), COOPERLAKE)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
+ifeq ($(TARGET_CORE), SAPPHIRERAPIDS)
+GETARCH_FLAGS := -DFORCE_NEHALEM
+endif
 ifeq ($(TARGET_CORE), SANDYBRIDGE)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
@@ -251,6 +261,8 @@ endif
 #For small matrix optimization
 ifeq ($(ARCH), x86_64)
 SMALL_MATRIX_OPT = 1
+else ifeq ($(CORE), POWER10)
+SMALL_MATRIX_OPT = 1
 endif
 ifeq ($(SMALL_MATRIX_OPT), 1)
 CCOMMON_OPT += -DSMALL_MATRIX_OPT
@@ -260,6 +272,10 @@ endif
 ifndef GOTOBLAS_MAKEFILE
 export GOTOBLAS_MAKEFILE = 1
 
+# Determine if the assembler is GNU Assembler
+HAVE_GAS := $(shell $(AS) -v < /dev/null 2>&1 | grep GNU 2>&1 >/dev/null ; echo $$?)
+GETARCH_FLAGS += -DHAVE_GAS=$(HAVE_GAS)
+
 # Generating Makefile.conf and config.h
 DUMMY := $(shell $(MAKE) -C $(TOPDIR) -f Makefile.prebuild CC="$(CC)" FC="$(FC)" HOSTCC="$(HOSTCC)" HOST_CFLAGS="$(GETARCH_FLAGS)" CFLAGS="$(CFLAGS)" BINARY=$(BINARY) USE_OPENMP=$(USE_OPENMP) TARGET_CORE=$(TARGET_CORE) ONLY_CBLAS=$(ONLY_CBLAS) TARGET=$(TARGET) all)
 
@@ -307,7 +323,7 @@ else
 SMP = 1
 endif
 else
-ifeq ($(NUM_THREAD), 1)
+ifeq ($(NUM_THREADS), 1)
 SMP =
 else
 SMP = 1
@@ -892,15 +908,25 @@ endif
 
 ifeq ($(C_COMPILER), PGI)
 PGCVERSIONGT20 := $(shell expr `$(CC) --version|sed -n "2p" |sed -e "s/[^0-9.]//g" |cut -d "." -f 1` \> 20)
-PGCVERSIONGTEQ20 := $(shell expr `$(CC) --version|sed -n "2p" |sed -e "s/[^0-9.]//g" |cut -d "." -f 1` \>= 20)
-PGCMINORVERSIONGE11 := $(shell expr `$(CC) --version|sed -n "2p" |sed -e "s/[^0-9.]//g" |cut -c 4-5` == 11)
+PGCVERSIONEQ20 := $(shell expr `$(CC) --version|sed -n "2p" |sed -e "s/[^0-9.]//g" |cut -d "." -f 1` == 20)
+PGCMINORVERSIONGE11 := $(shell expr `$(CC) --version|sed -n "2p" |cut -d "-" -f 1 |sed -e "s/[^0-9.]//g" |cut -c 4-5` \>= 11)
 PGCVERSIONCHECK := $(PGCVERSIONGT20)$(PGCVERSIONEQ20)$(PGCMINORVERSIONGE11)
-ifeq ($(PGCVERSIONCHECK), $(filter $(PGCVERSIONCHECK), 110 111 011))
+ifeq ($(PGCVERSIONCHECK), $(filter $(PGCVERSIONCHECK), 100 101 011))
 NEWPGI := 1
+PGCVERSIONGT21 := $(shell expr `$(CC) --version|sed -n "2p" |sed -e "s/[^0-9.]//g" |cut -d "." -f 1` \> 21)
+PGCVERSIONEQ21 := $(shell expr `$(CC) --version|sed -n "2p" |sed -e "s/[^0-9.]//g" |cut -d "." -f 1` == 21)
+PGCVERSIONCHECK2 := $(PGCVERSIONGT21)$(PGCVERSIONEQ21)$(PGCMINORVERSIONGE11)
+ifeq ($(PGCVERSIONCHECK2), $(filter $(PGCVERSIONCHECK2), 100 101 011))
+NEWPGI2 := 1
+endif
 endif
 ifdef BINARY64
 ifeq ($(ARCH), x86_64)
+ifneq ($(NEWPGI2),1)
 CCOMMON_OPT += -tp p7-64
+else
+CCOMMON_OPT += -tp px
+endif
 ifneq ($(NEWPGI),1)
 CCOMMON_OPT +=  -D__MMX__ -Mnollvm
 endif
@@ -915,7 +941,11 @@ endif
 endif
 endif
 else
+ifneq ($(NEWPGI2),1)
 CCOMMON_OPT += -tp p7
+else
+CCOMMON_OPT += -tp px
+endif
 endif
 endif
 
@@ -1092,8 +1122,12 @@ FCOMMON_OPT += -i8
 endif
 endif
 ifeq ($(ARCH), x86_64)
+ifneq ($(NEWPGI2),1)
 FCOMMON_OPT += -tp p7-64
 else
+FCOMMON_OPT += -tp px
+endif
+else
 ifeq ($(ARCH), power)
 ifeq ($(CORE), POWER6)
 $(warning NVIDIA HPC compilers do not support POWER6.)
@@ -1643,8 +1677,10 @@ export HAVE_VFP
 export HAVE_VFPV3
 export HAVE_VFPV4
 export HAVE_NEON
-export HAVE_MSA
-export MSA_FLAGS
+ifndef NO_MSA
+  export HAVE_MSA
+  export MSA_FLAGS
+endif
 export KERNELDIR
 export FUNCTION_PROFILE
 export TARGET_CORE

Makefile.system

@@ -81,6 +81,40 @@ CCOMMON_OPT += -march=cooperlake
 ifneq ($(F_COMPILER), NAG)
 FCOMMON_OPT += -march=cooperlake
 endif
+else  # gcc not support, fallback to avx512
+CCOMMON_OPT += -march=skylake-avx512
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=skylake-avx512
+endif
+endif
+endif
+ifeq ($(OSNAME), CYGWIN_NT)
+CCOMMON_OPT += -fno-asynchronous-unwind-tables
+FCOMMON_OPT += -fno-asynchronous-unwind-tables
+endif
+ifeq ($(OSNAME), WINNT)
+ifeq ($(C_COMPILER), GCC)
+CCOMMON_OPT += -fno-asynchronous-unwind-tables
+FCOMMON_OPT += -fno-asynchronous-unwind-tables
+endif
+endif
+endif
+endif
+
+ifeq ($(CORE), SAPPHIRERAPIDS)
+ifndef NO_AVX512
+ifeq ($(C_COMPILER), GCC)
+# sapphire rapids support was added in 11
+ifeq ($(GCCVERSIONGTEQ11), 1)
+CCOMMON_OPT += -march=sapphirerapids
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=sapphirerapids
+endif
+else  # gcc not support, fallback to avx512
+CCOMMON_OPT += -march=skylake-avx512
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=skylake-avx512
+endif
 endif
 endif
 ifeq ($(OSNAME), CYGWIN_NT)

Makefile.x86_64

@@ -23,6 +23,7 @@ HASWELL
 SKYLAKEX
 ATOM
 COOPERLAKE
+SAPPHIRERAPIDS
 
 b)AMD CPU:
 ATHLON

TargetList.txt

@@ -29,15 +29,15 @@ environment:
   global:
     CONDA_INSTALL_LOCN: C:\\Miniconda36-x64
   matrix:
-    - COMPILER: clang-cl
-      WITH_FORTRAN: ON
-    - COMPILER: clang-cl
-      DYNAMIC_ARCH: ON
-      WITH_FORTRAN: OFF
-    - COMPILER: cl
-    - COMPILER: MinGW64-gcc-7.2.0-mingw
-      DYNAMIC_ARCH: OFF
-      WITH_FORTRAN: ignore
+#    - COMPILER: clang-cl
+#      WITH_FORTRAN: ON
+#    - COMPILER: clang-cl
+#      DYNAMIC_ARCH: ON
+#      WITH_FORTRAN: OFF
+#    - COMPILER: cl
+#    - COMPILER: MinGW64-gcc-7.2.0-mingw
+#      DYNAMIC_ARCH: OFF
+#      WITH_FORTRAN: ignore
     - APPVEYOR_BUILD_WORKER_IMAGE: Visual Studio 2015
       COMPILER: MinGW-gcc-6.3.0-32   
     - APPVEYOR_BUILD_WORKER_IMAGE: Visual Studio 2015
@@ -46,6 +46,7 @@ environment:
 
 install:
   - if [%COMPILER%]==[clang-cl] call %CONDA_INSTALL_LOCN%\Scripts\activate.bat
+  - if [%COMPILER%]==[clang-cl] conda update --yes -n base conda
   - if [%COMPILER%]==[clang-cl] conda config --add channels conda-forge --force
   - if [%COMPILER%]==[clang-cl] conda config --set auto_update_conda false
   - if [%COMPILER%]==[clang-cl] conda install --yes --quiet clangdev cmake ninja flang=11.0.1
@@ -64,8 +65,8 @@ before_build:
   - if [%COMPILER%]==[MinGW64-gcc-7.2.0-mingw] cmake -G "MinGW Makefiles" -DNOFORTRAN=1 ..
   - if [%COMPILER%]==[MinGW-gcc-6.3.0-32] cmake -G "MSYS Makefiles" -DNOFORTRAN=1 ..
   - if [%COMPILER%]==[MinGW-gcc-5.3.0] cmake -G "MSYS Makefiles" -DNOFORTRAN=1 ..
-  - if [%WITH_FORTRAN%]==[OFF] cmake -G "Ninja" -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_C_COMPILER=clang-cl -DMSVC_STATIC_CRT=ON ..
-  - if [%WITH_FORTRAN%]==[ON] cmake -G "Ninja" -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_C_COMPILER=clang-cl -DCMAKE_Fortran_COMPILER=flang -DBUILD_WITHOUT_LAPACK=no -DNOFORTRAN=0 ..
+  - if [%WITH_FORTRAN%]==[OFF] cmake -G "Ninja" -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_C_COMPILER=clang-cl -DCMAKE_MT=mt -DMSVC_STATIC_CRT=ON ..
+  - if [%WITH_FORTRAN%]==[ON] cmake -G "Ninja" -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_C_COMPILER=clang-cl -DCMAKE_Fortran_COMPILER=flang -DCMAKE_MT=mt -DBUILD_WITHOUT_LAPACK=no -DNOFORTRAN=0 ..
   - if [%USE_OPENMP%]==[ON] cmake -DUSE_OPENMP=ON ..
   - if [%DYNAMIC_ARCH%]==[ON] cmake -DDYNAMIC_ARCH=ON -DDYNAMIC_LIST='CORE2;NEHALEM;SANDYBRIDGE;BULLDOZER;HASWELL' ..