diff --git a/.drone.yml b/.drone.yml index 779912954..696c5a99d 100644 --- a/.drone.yml +++ b/.drone.yml @@ -92,7 +92,7 @@ steps: - mkdir build && cd build - cmake $CMAKE_FLAGS .. - make -j - - ctest + - ctest -V --- kind: pipeline @@ -116,7 +116,7 @@ steps: - mkdir build && cd build - cmake $CMAKE_FLAGS .. - make -j - - ctest + - ctest -V --- kind: pipeline @@ -140,4 +140,4 @@ steps: - mkdir build && cd build - cmake $CMAKE_FLAGS .. - make -j - - ctest + - ctest -V diff --git a/.github/workflows/nightly-Homebrew-build.yml b/.github/workflows/nightly-Homebrew-build.yml new file mode 100644 index 000000000..f55e73d23 --- /dev/null +++ b/.github/workflows/nightly-Homebrew-build.yml @@ -0,0 +1,78 @@ +# Only the "head" branch of the OpenBLAS package is tested + +on: + push: + paths: + - '**/nightly-Homebrew-build.yml' + pull_request: + branches: + - develop + paths: + - '**/nightly-Homebrew-build.yml' + schedule: + - cron: 45 7 * * * +# This is 7:45 AM UTC daily, late at night in the USA + +# Since push and pull_request will still always be building and testing the `develop` branch, +# it only makes sense to test if this file has been changed + +name: Nightly-Homebrew-Build +jobs: + build-OpenBLAS-with-Homebrew: + runs-on: macos-latest + env: + HOMEBREW_DEVELOPER: "ON" + HOMEBREW_DISPLAY_INSTALL_TIMES: "ON" + HOMEBREW_NO_ANALYTICS: "ON" + HOMEBREW_NO_AUTO_UPDATE: "ON" + HOMEBREW_NO_BOTTLE_SOURCE_FALLBACK: "ON" + HOMEBREW_NO_INSTALL_CLEANUP: "ON" + + steps: + - name: Random delay for cron job + run: | + delay=$(( RANDOM % 600 )) + printf 'Delaying for %s seconds on event %s' ${delay} "${{ github.event_name }}" + sleep ${delay} + if: github.event_name == 'schedule' + + - uses: actions/checkout@v2 + # This isn't even needed, technically. Homebrew will get `develop` via git + + - name: Update Homebrew + if: github.event_name != 'pull_request' + run: brew update || true + + - name: Install prerequisites + run: brew install --fetch-HEAD --HEAD --only-dependencies --keep-tmp openblas + + - name: Install and bottle OpenBLAS + run: brew install --fetch-HEAD --HEAD --build-bottle --keep-tmp openblas + # the HEAD flags tell Homebrew to build the develop branch fetch via git + + - name: Create bottle + run: | + brew bottle -v openblas + mkdir bottles + mv *.bottle.tar.gz bottles + + - name: Upload bottle + uses: actions/upload-artifact@v1 + with: + name: openblas--HEAD.catalina.bottle.tar.gz + path: bottles + + - name: Show linkage + run: brew linkage -v openblas + + - name: Test openblas + run: brew test --HEAD --verbose openblas + + - name: Audit openblas formula + run: | + brew audit --strict openblas + brew cat openblas + + - name: Post logs on failure + if: failure() + run: brew gist-logs --with-hostname -v openblas diff --git a/.gitignore b/.gitignore index e9d08ca7e..6803a919e 100644 --- a/.gitignore +++ b/.gitignore @@ -87,4 +87,5 @@ build.* *.swp benchmark/*.goto benchmark/smallscaling - +CMakeCache.txt +CMakeFiles/* diff --git a/.travis.yml b/.travis.yml index 9e18412e8..0f20aef5c 100644 --- a/.travis.yml +++ b/.travis.yml @@ -176,7 +176,7 @@ matrix: - <<: *test-macos osx_image: xcode10.1 env: - - CC="/Applications/Xcode-10.1.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang -isysroot /Applications/Xcode-10.1.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS12.1.sdk" + - CC="/Applications/Xcode-10.1.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang" - CFLAGS="-O2 -Wno-macro-redefined -isysroot /Applications/Xcode-10.1.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS12.1.sdk -arch arm64 -miphoneos-version-min=10.0" - BTYPE="TARGET=ARMV8 BINARY=64 HOSTCC=clang NOFORTRAN=1" diff --git a/CMakeLists.txt b/CMakeLists.txt index 74db77135..951271717 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -6,7 +6,7 @@ cmake_minimum_required(VERSION 2.8.5) project(OpenBLAS C ASM) set(OpenBLAS_MAJOR_VERSION 0) set(OpenBLAS_MINOR_VERSION 3) -set(OpenBLAS_PATCH_VERSION 8.dev) +set(OpenBLAS_PATCH_VERSION 9.dev) set(OpenBLAS_VERSION "${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.${OpenBLAS_PATCH_VERSION}") # Adhere to GNU filesystem layout conventions diff --git a/Changelog.txt b/Changelog.txt index f160a4e13..5f924629b 100644 --- a/Changelog.txt +++ b/Changelog.txt @@ -1,45 +1,144 @@ OpenBLAS ChangeLog +==================================================================== +Version 0.3.9 + 1-Mar-2020 + + common: + * Fixed a miscompilation of the GETRF functions with CMAKE + * Imported bugfix 390 from LAPACK (missing NaN propagation in xCOMBSSQ) + * The size of the memory buffer used for splitting GEMM tasks across + multiple threads can now be configured in the build system. + +POWER: + * Fixed several compilation problems related to endianness + and ELF version on POWER8 and POWER9 + * Fixed use of the absolute value IAMIN/IAMAX instead of IMIN/IMAX + * Fixed a race condition in the level3 blas code + +MIPS64: + * Fixed use of the absoltute value IAMIN/IAMAX instead of IMIN/IMAX + +ARMV7: + * Fixed a race condition in the level3 blas code + * Fixed compilation on Android +ARMV8: + * Added support for Ampere EMAG8180 + * Added support for Neoverse N1 + * Improved performance of the blas_lock function + * Fixed a race condition in the level3 blas code + * Fixed a performance regression on TSV110-based servers + +x86_64: + * Fixed a long-standing error with undeclared register overwrites + in the DSCAL microkernel for HASWELL,SKYLAKEX and ZEN + * Fixed a long-standing bug in the SSE implementation of IAMAX + * Fixed a CMAKE build failure with DYNAMIC_ARCH + * Fixed cpu autodetection of Goldmont+, Cannon Lake and Ice Lake + * Fixed a compilation failure on OSX with compiler name containing dash + * Fixed compilation with MinGW on SkylakeX + * Improved speed of the AVX512 GEMM3M kernel on SkylakeX + * Added an AVX512 STRMM kernel for SkylakeX + * Improved GEMM performance on Haswell and Zen + +zarch: + * fixed compilation of the DYNAMIC_ARCH code + +==================================================================== +Version 0.3.8 + 9-Feb-2020 + +common: +` * LAPACK has been updated to 3.9.0 (plus patches up to + January 2nd, 2020) + * CMAKE support has been improved in several areas including + cross-compilation + * a thread race condition in the GEMM3M kernels was resolved + * the "generic" (plain C) gemm beta kernel used by many targets + has been sped up + * an optimized version of the LAPACK trtrs functions has been added + * an incompatibilty between the LAPACK tests and the OpenBLAS + implementation of XERBLA was resolved, removing the numerous + warnings about wrong error exits in the former + * support for NetBSD has been added + * support for compilation with g95 and non-GNU versions of ld + has been improved + * support for compilation with (upcoming) gcc 10 has been added + +POWER: + * worked around miscompilation of several POWER8 and POWER9 + kernels by older versions of gcc + * added support for big-endian POWER8 and for compilation on AIX + * corrected bugs in the big-endian support for PPC440 and PPC970 + * DYNAMIC_ARCH support is now available in CMAKE builds as well + +ARMV8: + * performance of DGEMM_BETA and SGEMM_NCOPY has been improved + * compilation for 32bit works again + * performance of the RPCC function has been improved + * improved performance on small systems + * DYNAMIC_ARCH support is now available in CMAKE builds as well + * cross-compilation from OSX to IOS was simplified + +x86_64: + * a new AVX512 DGEMM kernel was added and the AVX512 SGEMM kernel + was significantly improved + * optimized AVX512 kernels for CGEMM and ZGEMM have been added + * AVX2 kernels for STRMM, SGEMM, and CGEMM have been significantly + sped up and optimized CGEMM3M and ZGEMM3M kernels have been added + * added support for QEMU virtual cpus + * a compilation problem with PGI and SUN compilers was fixed + * Intel "Goldmont plus" is now autodetected + * a potential crash on program exit on MS Windows has been fixed + +x86: + * an unwanted case sensitivity in the implementation of LSAME + on older 32bit AMD cpus was fixed + +zarch: + * Z15 is now supported as Z14 + * DYNAMIC_ARCH is now available on ZARCH as well + ==================================================================== Version 0.3.7 11-Aug 2019 common: - * having the gmake special variables TARGET_ARCH or TARGET_MACH - defined no longer causes build failures in ctest or utest - * defining NO_AFFINITY or USE_TLS to 0 in gmake builds no longer - has the same effect as setting them to 1 - * a new test program was added to allow checking the library for - thread safety - * a new option USE_LOCKING was added to ensure thread safety when - OpenBLAS itself is built without multithreading but will be - called from multiple threads. - * a build failure on Linux with glibc versions earlier than 2.5 - was fixed - * a runtime error with CPU enumeration (and NO_AFFINITY not set) - on glibc 2.6 was fixed - * NO_AFFINITY was added to the CMAKE options (and defaults to being - active on Linux, as in the gmake builds) + * having the gmake special variables TARGET_ARCH or TARGET_MACH + defined no longer causes build failures in ctest or utest + * defining NO_AFFINITY or USE_TLS to 0 in gmake builds no longer + has the same effect as setting them to 1 + * a new test program was added to allow checking the library for + thread safety + * a new option USE_LOCKING was added to ensure thread safety when + OpenBLAS itself is built without multithreading but will be + called from multiple threads. + * a build failure on Linux with glibc versions earlier than 2.5 + was fixed + * a runtime error with CPU enumeration (and NO_AFFINITY not set) + on glibc 2.6 was fixed + * NO_AFFINITY was added to the CMAKE options (and defaults to being + active on Linux, as in the gmake builds) x86_64: - * the build-time logic for detection of AVX512 availability in - the processor and compiler was fixed - * gmake builds on OSX now set the internal name of the library to - libopenblas.0.dylib (consistent with CMAKE) - * the Haswell DGEMM kernel received a significant speedup through - improved prefetch and load instructions - * performance of DGEMM, DTRMM, DTRSM and ZDOT on Zen/Zen2 was markedly - increased by avoiding vpermpd instructions - * the SKYLAKEX (AVX512) DGEMM helper functions have now been disabled - to fix remaining errors in DGEMM, DSYMM and DTRMM + * the build-time logic for detection of AVX512 availability in + the processor and compiler was fixed + * gmake builds on OSX now set the internal name of the library to + libopenblas.0.dylib (consistent with CMAKE) + * the Haswell DGEMM kernel received a significant speedup through + improved prefetch and load instructions + * performance of DGEMM, DTRMM, DTRSM and ZDOT on Zen/Zen2 was markedly + increased by avoiding vpermpd instructions + * the SKYLAKEX (AVX512) DGEMM helper functions have now been disabled + to fix remaining errors in DGEMM, DSYMM and DTRMM -## POWER: - * added support for building on FreeBSD/powerpc64 and FreeBSD/ppc970 - * added optimized kernels for POWER9 SGEMM and STRMM +POWER: + * added support for building on FreeBSD/powerpc64 and FreeBSD/ppc970 + * added optimized kernels for POWER9 SGEMM and STRMM -## ARMV7: - * fixed the softfp implementations of xAMAX and IxAMAX - * removed the predefined -march= flags on both ARMV5 and ARMV6 as - they were appropriate for only a subset of platforms +ARMV7: + * fixed the softfp implementations of xAMAX and IxAMAX + * removed the predefined -march= flags on both ARMV5 and ARMV6 as + they were appropriate for only a subset of platforms ==================================================================== Version 0.3.6 diff --git a/Makefile b/Makefile index a22e16bab..a87ffdf2a 100644 --- a/Makefile +++ b/Makefile @@ -56,10 +56,21 @@ ifneq ($(INTERFACE64), 0) @echo " Use 64 bits int (equivalent to \"-i8\" in Fortran) " endif endif - - @echo " C compiler ... $(C_COMPILER) (command line : $(CC))" + @$(CC) --version > /dev/null 2>&1;\ + if [ $$? -eq 0 ]; then \ + cverinfo=`$(CC) --version | sed -n '1p'`; \ + echo " C compiler ... $(C_COMPILER) (cmd & version : $${cverinfo})";\ + else \ + echo " C compiler ... $(C_COMPILER) (command line : $(CC))";\ + fi ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN))) - @echo " Fortran compiler ... $(F_COMPILER) (command line : $(FC))" + @$(FC) --version > /dev/null 2>&1;\ + if [ $$? -eq 0 ]; then \ + fverinfo=`$(FC) --version | sed -n '1p'`; \ + echo " Fortran compiler ... $(F_COMPILER) (cmd & version : $${fverinfo})";\ + else \ + echo " Fortran compiler ... $(F_COMPILER) (command line : $(FC))";\ + fi endif ifneq ($(OSNAME), AIX) @echo -n " Library Name ... $(LIBNAME)" @@ -68,9 +79,9 @@ else endif ifndef SMP - @echo " (Single threaded) " + @echo " (Single-threading) " else - @echo " (Multi threaded; Max num-threads is $(NUM_THREADS))" + @echo " (Multi-threading; Max num-threads is $(NUM_THREADS))" endif ifeq ($(USE_OPENMP), 1) @@ -317,7 +328,7 @@ lapack-test : $(MAKE) -j 1 -C $(NETLIB_LAPACK_DIR)/TESTING/EIG xeigtstc xeigtstd xeigtsts xeigtstz $(MAKE) -j 1 -C $(NETLIB_LAPACK_DIR)/TESTING/LIN xlintstc xlintstd xlintstds xlintstrfd xlintstrfz xlintsts xlintstz xlintstzc xlintstrfs xlintstrfc ifneq ($(CROSS), 1) - ( cd $(NETLIB_LAPACK_DIR)/INSTALL; make all; ./testlsame; ./testslamch; ./testdlamch; \ + ( cd $(NETLIB_LAPACK_DIR)/INSTALL; $(MAKE) all; ./testlsame; ./testslamch; ./testdlamch; \ ./testsecond; ./testdsecnd; ./testieee; ./testversion ) (cd $(NETLIB_LAPACK_DIR); ./lapack_testing.py -r -b TESTING) endif diff --git a/Makefile.arm b/Makefile.arm index b5d80f8e6..fac6b56824 100644 --- a/Makefile.arm +++ b/Makefile.arm @@ -1,7 +1,7 @@ ifeq ($(CORE), $(filter $(CORE),ARMV7 CORTEXA9 CORTEXA15)) ifeq ($(OSNAME), Android) -CCOMMON_OPT += -mfpu=neon -FCOMMON_OPT += -mfpu=neon +CCOMMON_OPT += -mfpu=neon -march=armv7-a +FCOMMON_OPT += -mfpu=neon -march=armv7-a else CCOMMON_OPT += -mfpu=vfpv3 -march=armv7-a FCOMMON_OPT += -mfpu=vfpv3 -march=armv7-a diff --git a/Makefile.arm64 b/Makefile.arm64 index c17ea7938..a7cd82e3a 100644 --- a/Makefile.arm64 +++ b/Makefile.arm64 @@ -24,6 +24,23 @@ CCOMMON_OPT += -march=armv8-a -mtune=cortex-a73 FCOMMON_OPT += -march=armv8-a -mtune=cortex-a73 endif +# Use a72 tunings because Neoverse-N1 is only available +# in GCC>=9 +ifeq ($(CORE), NEOVERSEN1) +ifeq ($(GCCVERSIONGTEQ7), 1) +ifeq ($(GCCVERSIONGTEQ9), 1) +CCOMMON_OPT += -march=armv8.2-a -mtune=neoverse-n1 +FCOMMON_OPT += -march=armv8.2-a -mtune=neoverse-n1 +else +CCOMMON_OPT += -march=armv8.2-a -mtune=cortex-a72 +FCOMMON_OPT += -march=armv8.2-a -mtune=cortex-a72 +endif +else +CCOMMON_OPT += -march=armv8-a -mtune=cortex-a72 +FCOMMON_OPT += -march=armv8-a -mtune=cortex-a72 +endif +endif + ifeq ($(CORE), THUNDERX) CCOMMON_OPT += -march=armv8-a -mtune=thunderx FCOMMON_OPT += -march=armv8-a -mtune=thunderx diff --git a/Makefile.install b/Makefile.install index e01d866c9..2dc32c3d9 100644 --- a/Makefile.install +++ b/Makefile.install @@ -82,7 +82,7 @@ ifeq ($(OSNAME), $(filter $(OSNAME),FreeBSD OpenBSD NetBSD DragonFly)) endif ifeq ($(OSNAME), Darwin) @-cp $(LIBDYNNAME) "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)" - @-install_name_tool -id "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)/$(LIBDYNNAME)" "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)/$(LIBDYNNAME)" + @-install_name_tool -id "$(OPENBLAS_LIBRARY_DIR)/$(LIBPREFIX).$(MAJOR_VERSION).dylib" "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)/$(LIBDYNNAME)" @cd "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)" ; \ ln -fs $(LIBDYNNAME) $(LIBPREFIX).dylib ; \ ln -fs $(LIBDYNNAME) $(LIBPREFIX).$(MAJOR_VERSION).dylib diff --git a/Makefile.prebuild b/Makefile.prebuild index a366004a1..b00f13368 100644 --- a/Makefile.prebuild +++ b/Makefile.prebuild @@ -42,7 +42,7 @@ all: getarch_2nd ./getarch_2nd 1 >> $(TARGET_CONF) config.h : c_check f_check getarch - perl ./c_check $(TARGET_MAKE) $(TARGET_CONF) $(CC) $(TARGET_FLAGS) + perl ./c_check $(TARGET_MAKE) $(TARGET_CONF) $(CC) $(TARGET_FLAGS) $(CFLAGS) ifneq ($(ONLY_CBLAS), 1) perl ./f_check $(TARGET_MAKE) $(TARGET_CONF) $(FC) $(TARGET_FLAGS) else @@ -59,13 +59,13 @@ endif getarch : getarch.c cpuid.S dummy $(CPUIDEMU) - $(HOSTCC) $(CFLAGS) $(EXFLAGS) -o $(@F) getarch.c cpuid.S $(CPUIDEMU) + $(HOSTCC) $(HOST_CFLAGS) $(EXFLAGS) -o $(@F) getarch.c cpuid.S $(CPUIDEMU) getarch_2nd : getarch_2nd.c config.h dummy ifndef TARGET_CORE - $(HOSTCC) -I. $(CFLAGS) -o $(@F) getarch_2nd.c + $(HOSTCC) -I. $(HOST_CFLAGS) -o $(@F) getarch_2nd.c else - $(HOSTCC) -I. $(CFLAGS) -DBUILD_KERNEL -o $(@F) getarch_2nd.c + $(HOSTCC) -I. $(HOST_CFLAGS) -DBUILD_KERNEL -o $(@F) getarch_2nd.c endif dummy: diff --git a/Makefile.rule b/Makefile.rule index c0941e488..724a60ec4 100644 --- a/Makefile.rule +++ b/Makefile.rule @@ -3,7 +3,7 @@ # # This library's version -VERSION = 0.3.8.dev +VERSION = 0.3.9.dev # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library @@ -97,6 +97,15 @@ VERSION = 0.3.8.dev # they need to wait for the preceding API calls to finish or risk data corruption. # NUM_PARALLEL = 2 +# When multithreading, OpenBLAS needs to use a memory buffer for communicating +# and collating results for individual subranges of the original matrix. Since +# the original GotoBLAS of the early 2000s, the default size of this buffer has +# been set at a value of 32<<20 (which is 32MB) on x86_64 , twice that on PPC. +# If you expect to handle large problem sizes (beyond about 30000x30000) uncomment +# this line and adjust the (32<= 4) GCCVERSIONGT4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 4) -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) GCCMINORVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 7) ifeq ($(GCCVERSIONGT4), 1) @@ -554,10 +554,12 @@ DYNAMIC_CORE += CORTEXA53 DYNAMIC_CORE += CORTEXA57 DYNAMIC_CORE += CORTEXA72 DYNAMIC_CORE += CORTEXA73 +DYNAMIC_CORE += NEOVERSEN1 DYNAMIC_CORE += FALKOR DYNAMIC_CORE += THUNDERX DYNAMIC_CORE += THUNDERX2T99 DYNAMIC_CORE += TSV110 +DYNAMIC_CORE += EMAG8180 endif ifeq ($(ARCH), zarch) @@ -572,6 +574,7 @@ ifneq ($(C_COMPILER), GCC) DYNAMIC_CORE += POWER9 endif ifeq ($(C_COMPILER), GCC) +GCCVERSIONGT5 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 5) ifeq ($(GCCVERSIONGT5), 1) DYNAMIC_CORE += POWER9 else @@ -1328,6 +1331,7 @@ export OSNAME export ARCH export CORE export LIBCORE +export __BYTE_ORDER__ export PGCPATH export CONFIG export CC diff --git a/Makefile.x86_64 b/Makefile.x86_64 index 99364752f..f2de51ef4 100644 --- a/Makefile.x86_64 +++ b/Makefile.x86_64 @@ -15,10 +15,12 @@ CCOMMON_OPT += -march=skylake-avx512 FCOMMON_OPT += -march=skylake-avx512 ifeq ($(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 diff --git a/README.md b/README.md index 14815ff00..04f43f4c7 100644 --- a/README.md +++ b/README.md @@ -26,6 +26,8 @@ You can download them from [file hosting on sourceforge.net](https://sourceforge Download from project homepage, https://xianyi.github.com/OpenBLAS/, or check out the code using Git from https://github.com/xianyi/OpenBLAS.git. +Buildtime parameters can be chosen in Makefile.rule, see there for a short description of each option. +Most can also be given directly on the make or cmake command line. ### Dependencies @@ -101,7 +103,7 @@ The default installation directory is `/opt/OpenBLAS`. ## Supported CPUs and Operating Systems -Please read `GotoBLAS_01Readme.txt`. +Please read `GotoBLAS_01Readme.txt` for older CPU models already supported by the 2010 GotoBLAS. ### Additional supported CPUs @@ -109,8 +111,8 @@ Please read `GotoBLAS_01Readme.txt`. - **Intel Xeon 56xx (Westmere)**: Used GotoBLAS2 Nehalem codes. - **Intel Sandy Bridge**: Optimized Level-3 and Level-2 BLAS with AVX on x86-64. -- **Intel Haswell**: Optimized Level-3 and Level-2 BLAS with AVX2 and FMA on x86-64. -- **Intel Skylake**: Optimized Level-3 and Level-2 BLAS with AVX512 and FMA on x86-64. +- **Intel Haswell**: Optimized Level-3 and Level-2 BLAS with AVX2 and FMA on x86-64. +- **Intel Skylake-X**: Optimized Level-3 and Level-2 BLAS with AVX512 and FMA on x86-64. - **AMD Bobcat**: Used GotoBLAS2 Barcelona codes. - **AMD Bulldozer**: x86-64 ?GEMM FMA4 kernels. (Thanks to Werner Saar) - **AMD PILEDRIVER**: Uses Bulldozer codes with some optimizations. @@ -129,8 +131,15 @@ Please read `GotoBLAS_01Readme.txt`. #### ARM64 -- **ARMv8**: Experimental -- **ARM Cortex-A57**: Experimental +- **ARMv8**: Basic ARMV8 with small caches, optimized Level-3 and Level-2 BLAS +- **Cortex-A53**: same as ARMV8 (different cpu specifications) +- **Cortex A57**: Optimized Level-3 and Level-2 functions +- **Cortex A72**: same as A57 ( different cpu specifications) +- **Cortex A73**: same as A57 (different cpu specifications) +- **Falkor**: same as A57 (different cpu specifications) +- **ThunderX**: Optimized some Level-1 functions +- **ThunderX2T99**: Optimized Level-3 BLAS and parts of Levels 1 and 2 +- **TSV110**: Optimized some Level-3 helper functions #### PPC/PPC64 @@ -139,18 +148,34 @@ Please read `GotoBLAS_01Readme.txt`. #### IBM zEnterprise System -- **Z13**: Optimized Level-3 BLAS and Level-1,2 (double precision) -- **Z14**: Optimized Level-3 BLAS and Level-1,2 (single precision) +- **Z13**: Optimized Level-3 BLAS and Level-1,2 +- **Z14**: Optimized Level-3 BLAS and (single precision) Level-1,2 + +### Support for multiple targets in a single library + +OpenBLAS can be built for multiple targets with runtime detection of the target cpu by specifiying DYNAMIC_ARCH=1 in Makefile.rule, on the gmake command line or as -DDYNAMIC_ARCH=TRUE in cmake. +For **x86_64**, the list of targets this activates contains Prescott, Core2, Nehalem, Barcelona, Sandybridge, Bulldozer, Piledriver, Steamroller, Excavator, Haswell, Zen, SkylakeX. For cpu generations not included in this list, the corresponding older model is used. If you also specify DYNAMIC_OLDER=1, specific support for Penryn, Dunnington, Opteron, Opteron/SSE3, Bobcat, Atom and Nano is added. Finally there is an option DYNAMIC_LIST that allows to specify an individual list of targets to include instead of the default. +DYNAMIC_ARCH is also supported on **x86**, where it translates to Katmai, Coppermine, Northwood, Prescott, Banias, +Core2, Penryn, Dunnington, Nehalem, Athlon, Opteron, Opteron_SSE3, Barcelona, Bobcat, Atom and Nano. +On **ARMV8**, it enables support for CortexA53, CortexA57, CortexA72, CortexA73, Falkor, ThunderX, ThunderX2T99, TSV110 as well as generic ARMV8 cpus. +For **POWER**, the list encompasses POWER6, POWER8 and POWER9, on **ZARCH** it comprises Z13 and Z14. +The TARGET option can be used in conjunction with DYNAMIC_ARCH=1 to specify which cpu model should be assumed for all the +common code in the library, usually you will want to set this to the oldest model you expect to encounter. +Please note that it is not possible to combine support for different architectures, so no combined 32 and 64 bit or x86_64 and arm64 in the same library. ### Supported OS - **GNU/Linux** - **MinGW or Visual Studio (CMake)/Windows**: Please read . -- **Darwin/macOS**: Experimental. Although GotoBLAS2 supports Darwin, we are not macOS experts. +- **Darwin/macOS/OSX/iOS**: Experimental. Although GotoBLAS2 already supports Darwin, we are not OSX/iOS experts. - **FreeBSD**: Supported by the community. We don't actively test the library on this OS. - **OpenBSD**: Supported by the community. We don't actively test the library on this OS. +- **NetBSD**: Supported by the community. We don't actively test the library on this OS. - **DragonFly BSD**: Supported by the community. We don't actively test the library on this OS. - **Android**: Supported by the community. Please read . +- **AIX**: Supported on PPC up to POWER8 +- **Haiku**: Supported by the community. We don't actively test the library on this OS. +- **SunOS**: Supported by the community. We don't actively test the library on this OS: ## Usage @@ -205,7 +230,7 @@ Please see Changelog.txt to view the differences between OpenBLAS and GotoBLAS2 * Please use Clang version 3.1 and above to compile the library on Sandy Bridge microarchitecture. Clang 3.0 will generate the wrong AVX binary code. * Please use GCC version 6 or LLVM version 6 and above to compile Skylake AVX512 kernels. -* The number of CPUs/cores should less than or equal to 256. On Linux `x86_64` (`amd64`), +* The number of CPUs/cores should be less than or equal to 256. On Linux `x86_64` (`amd64`), there is experimental support for up to 1024 CPUs/cores and 128 numa nodes if you build the library with `BIGNUMA=1`. * OpenBLAS does not set processor affinity by default. diff --git a/TargetList.txt b/TargetList.txt index 6a57bf1af..f4a40ed02 100644 --- a/TargetList.txt +++ b/TargetList.txt @@ -88,6 +88,8 @@ CORTEXA53 CORTEXA57 CORTEXA72 CORTEXA73 +NEOVERSEN1 +EMAG8180 FALKOR THUNDERX THUNDERX2T99 diff --git a/benchmark/Makefile b/benchmark/Makefile index 51e9c64aa..139496085 100644 --- a/benchmark/Makefile +++ b/benchmark/Makefile @@ -12,9 +12,9 @@ include $(TOPDIR)/Makefile.system # ACML 6.1 custom ACML=/home/saar/acml6.1/gfortran64_mp/lib LIBACML = -fopenmp $(ACML)/libacml_mp.so -lgfortran -lm - -# Atlas Ubuntu + +# Atlas Ubuntu #ATLAS=/usr/lib/atlas-base #LIBATLAS = -fopenmp $(ATLAS)/liblapack_atlas.a $(ATLAS)/libptcblas.a $(ATLAS)/libptf77blas.a $(ATLAS)/libatlas.a -lgfortran -lm @@ -56,11 +56,16 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \ sgemm.goto dgemm.goto cgemm.goto zgemm.goto \ strmm.goto dtrmm.goto ctrmm.goto ztrmm.goto \ strsm.goto dtrsm.goto ctrsm.goto ztrsm.goto \ + sspr.goto dspr.goto \ + sspr2.goto dspr2.goto \ + ssyr.goto dsyr.goto \ + ssyr2.goto dsyr2.goto \ ssyrk.goto dsyrk.goto csyrk.goto zsyrk.goto \ ssyr2k.goto dsyr2k.goto csyr2k.goto zsyr2k.goto \ sger.goto dger.goto cger.goto zger.goto \ sdot.goto ddot.goto \ - srot.goto drot.goto \ + srot.goto drot.goto csrot.goto zdrot.goto \ + srotm.goto drotm.goto \ saxpy.goto daxpy.goto caxpy.goto zaxpy.goto \ scopy.goto dcopy.goto ccopy.goto zcopy.goto \ sswap.goto dswap.goto cswap.goto zswap.goto \ @@ -68,25 +73,36 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \ sasum.goto dasum.goto casum.goto zasum.goto \ ssymv.goto dsymv.goto csymv.goto zsymv.goto \ chemv.goto zhemv.goto \ + chbmv.goto zhbmv.goto \ + chpmv.goto zhpmv.goto \ chemm.goto zhemm.goto \ cherk.goto zherk.goto \ cher2k.goto zher2k.goto \ sgemv.goto dgemv.goto cgemv.goto zgemv.goto \ + strmv.goto dtrmv.goto ctrmv.goto ztrmv.goto \ + strsv.goto dtrsv.goto ctrsv.goto ztrsv.goto \ sgeev.goto dgeev.goto cgeev.goto zgeev.goto \ sgesv.goto dgesv.goto cgesv.goto zgesv.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 + ssymm.goto dsymm.goto csymm.goto zsymm.goto \ + saxpby.goto daxpby.goto caxpby.goto zaxpby.goto acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \ scholesky.acml dcholesky.acml ccholesky.acml zcholesky.acml \ sgemm.acml dgemm.acml cgemm.acml zgemm.acml \ strmm.acml dtrmm.acml ctrmm.acml ztrmm.acml \ strsm.acml dtrsm.acml ctrsm.acml ztrsm.acml \ + sspr.acml dspr.acml \ + sspr2.acml dspr2.acml \ + ssyr.acml dsyr.acml \ + ssyr2.acml dsyr2.acml \ ssyrk.acml dsyrk.acml csyrk.acml zsyrk.acml \ ssyr2k.acml dsyr2k.acml csyr2k.acml zsyr2k.acml \ sger.acml dger.acml cger.acml zger.acml \ sdot.acml ddot.acml \ + srot.acml drot.acml csrot.acml zdrot.acml \ + srotm.acml drotm.acml \ saxpy.acml daxpy.acml caxpy.acml zaxpy.acml \ scopy.acml dcopy.acml ccopy.acml zcopy.acml \ sswap.acml dswap.acml cswap.acml zswap.acml \ @@ -94,25 +110,36 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \ sasum.acml dasum.acml casum.acml zasum.acml \ ssymv.acml dsymv.acml csymv.acml zsymv.acml \ chemv.acml zhemv.acml \ + chbmv.acml zhbmv.acml \ + chpmv.acml zhpmv.acml \ chemm.acml zhemm.acml \ cherk.acml zherk.acml \ cher2k.acml zher2k.acml \ sgemv.acml dgemv.acml cgemv.acml zgemv.acml \ + strmv.acml dtrmv.acml ctrmv.acml ztrmv.acml \ + strsv.acml dtrsv.acml ctrsv.acml ztrsv.acml \ sgeev.acml dgeev.acml cgeev.acml zgeev.acml \ sgesv.acml dgesv.acml cgesv.acml zgesv.acml \ sgetri.acml dgetri.acml cgetri.acml zgetri.acml \ spotrf.acml dpotrf.acml cpotrf.acml zpotrf.acml \ - ssymm.acml dsymm.acml csymm.acml zsymm.acml + ssymm.acml dsymm.acml csymm.acml zsymm.acml \ + saxpby.goto daxpby.goto caxpby.goto zaxpby.goto atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \ scholesky.atlas dcholesky.atlas ccholesky.atlas zcholesky.atlas \ sgemm.atlas dgemm.atlas cgemm.atlas zgemm.atlas \ strmm.atlas dtrmm.atlas ctrmm.atlas ztrmm.atlas \ strsm.atlas dtrsm.atlas ctrsm.atlas ztrsm.atlas \ + sspr.atlas dspr.atlas \ + sspr2.atlas dspr2.atlas \ + ssyr.atlas dsyr.atlas \ + ssyr2.atlas dsyr2.atlas \ ssyrk.atlas dsyrk.atlas csyrk.atlas zsyrk.atlas \ ssyr2k.atlas dsyr2k.atlas csyr2k.atlas zsyr2k.atlas \ sger.atlas dger.atlas cger.atlas zger.atlas\ sdot.atlas ddot.atlas \ + srot.atlas drot.atlas csrot.atlas zdrot.atlas \ + srotm.atlas drotm.atlas \ saxpy.atlas daxpy.atlas caxpy.atlas zaxpy.atlas \ scopy.atlas dcopy.atlas ccopy.atlas zcopy.atlas \ sswap.atlas dswap.atlas cswap.atlas zswap.atlas \ @@ -120,26 +147,37 @@ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \ sasum.atlas dasum.atlas casum.atlas zasum.atlas \ ssymv.atlas dsymv.atlas csymv.atlas zsymv.atlas \ chemv.atlas zhemv.atlas \ + chbmv.atlas zhbmv.atlas \ + chpmv.atlas zhpmv.atlas \ chemm.acml zhemm.acml \ chemm.atlas zhemm.atlas \ cherk.atlas zherk.atlas \ cher2k.atlas zher2k.atlas \ sgemv.atlas dgemv.atlas cgemv.atlas zgemv.atlas \ + strmv.atlas dtrmv.atlas ctrmv.atlas ztrmv.atlas \ + strsv.atlas dtrsv.atlas ctrsv.atlas ztrsv.atlas \ sgeev.atlas dgeev.atlas cgeev.atlas zgeev.atlas \ sgesv.atlas dgesv.atlas cgesv.atlas zgesv.atlas \ sgetri.atlas dgetri.atlas cgetri.atlas zgetri.atlas \ spotrf.atlas dpotrf.atlas cpotrf.atlas zpotrf.atlas \ - ssymm.atlas dsymm.atlas csymm.atlas zsymm.atlas + ssymm.atlas dsymm.atlas csymm.atlas zsymm.atlas \ + saxpby.goto daxpby.goto caxpby.goto zaxpby.goto mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \ scholesky.mkl dcholesky.mkl ccholesky.mkl zcholesky.mkl \ sgemm.mkl dgemm.mkl cgemm.mkl zgemm.mkl \ strmm.mkl dtrmm.mkl ctrmm.mkl ztrmm.mkl \ strsm.mkl dtrsm.mkl ctrsm.mkl ztrsm.mkl \ + sspr.mkl dspr.mkl \ + sspr2.mkl dspr2.mkl \ + ssyr.mkl dsyr.mkl \ + ssyr2.mkl dsyr2.mkl \ ssyrk.mkl dsyrk.mkl csyrk.mkl zsyrk.mkl \ ssyr2k.mkl dsyr2k.mkl csyr2k.mkl zsyr2k.mkl \ sger.mkl dger.mkl cger.mkl zger.mkl \ sdot.mkl ddot.mkl \ + srot.mkl drot.mkl csrot.mkl zdrot.mkl \ + srotm.mkl drotm.mkl \ saxpy.mkl daxpy.mkl caxpy.mkl zaxpy.mkl \ scopy.mkl dcopy.mkl ccopy.mkl zcopy.mkl \ sswap.mkl dswap.mkl cswap.mkl zswap.mkl \ @@ -147,26 +185,36 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \ sasum.mkl dasum.mkl casum.mkl zasum.mkl \ ssymv.mkl dsymv.mkl csymv.mkl zsymv.mkl \ chemv.mkl zhemv.mkl \ + chbmv.mkl zhbmv.mkl \ + chpmv.mkl zhpmv.mkl \ chemm.mkl zhemm.mkl \ cherk.mkl zherk.mkl \ cher2k.mkl zher2k.mkl \ sgemv.mkl dgemv.mkl cgemv.mkl zgemv.mkl \ + strmv.mkl dtrmv.mkl ctrmv.mkl ztrmv.mkl \ + strsv.mkl dtrsv.mkl ctrsv.mkl ztrsv.mkl \ sgeev.mkl dgeev.mkl cgeev.mkl zgeev.mkl \ sgesv.mkl dgesv.mkl cgesv.mkl zgesv.mkl \ sgetri.mkl dgetri.mkl cgetri.mkl zgetri.mkl \ spotrf.mkl dpotrf.mkl cpotrf.mkl zpotrf.mkl \ - ssymm.mkl dsymm.mkl csymm.mkl zsymm.mkl + ssymm.mkl dsymm.mkl csymm.mkl zsymm.mkl \ + saxpby.goto daxpby.goto caxpby.goto zaxpby.goto else goto :: sgemm.goto dgemm.goto cgemm.goto zgemm.goto \ strmm.goto dtrmm.goto ctrmm.goto ztrmm.goto \ strsm.goto dtrsm.goto ctrsm.goto ztrsm.goto \ + sspr.goto dspr.goto \ + sspr2.goto dspr2.goto \ + ssyr.goto dsyr.goto \ + ssyr2.goto dsyr2.goto \ ssyrk.goto dsyrk.goto csyrk.goto zsyrk.goto \ ssyr2k.goto dsyr2k.goto csyr2k.goto zsyr2k.goto \ sger.goto dger.goto cger.goto zger.goto \ sdot.goto ddot.goto cdot.goto zdot.goto \ - srot.goto drot.goto \ + srot.goto drot.goto csrot.goto zdrot.goto \ + srotm.goto drotm.goto \ saxpy.goto daxpy.goto caxpy.goto zaxpy.goto \ scopy.goto dcopy.goto ccopy.goto zcopy.goto \ sswap.goto dswap.goto cswap.goto zswap.goto \ @@ -174,13 +222,18 @@ goto :: sgemm.goto dgemm.goto cgemm.goto zgemm.goto \ sasum.goto dasum.goto casum.goto zasum.goto \ ssymv.goto dsymv.goto \ chemv.goto zhemv.goto \ + chbmv.goto zhbmv.goto \ + chpmv.goto zhpmv.goto \ chemm.goto zhemm.goto \ cherk.goto zherk.goto \ cher2k.goto zher2k.goto \ sgemv.goto dgemv.goto cgemv.goto zgemv.goto \ + strmv.goto dtrmv.goto ctrmv.goto ztrmv.goto \ + strsv.goto dtrsv.goto ctrsv.goto ztrsv.goto \ ssymm.goto dsymm.goto csymm.goto zsymm.goto \ smallscaling \ isamax.goto idamax.goto icamax.goto izamax.goto \ + saxpby.goto daxpby.goto caxpby.goto zaxpby.goto \ snrm2.goto dnrm2.goto scnrm2.goto dznrm2.goto $(GOTO_LAPACK_TARGETS) acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \ @@ -188,10 +241,16 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \ sgemm.acml dgemm.acml cgemm.acml zgemm.acml \ strmm.acml dtrmm.acml ctrmm.acml ztrmm.acml \ strsm.acml dtrsm.acml ctrsm.acml ztrsm.acml \ + sspr.acml dspr.acml \ + sspr2.acml dspr2.acml \ + ssyr.acml dsyr.acml \ + ssyr2.acml dsyr2.acml \ ssyrk.acml dsyrk.acml csyrk.acml zsyrk.acml \ ssyr2k.acml dsyr2k.acml csyr2k.acml zsyr2k.acml \ sger.acml dger.acml cger.acml zger.acml \ sdot.acml ddot.acml \ + srot.acml drot.acml csrot.acml zdrot.acml \ + srotm.acml drotm.acml \ saxpy.acml daxpy.acml caxpy.acml zaxpy.acml \ scopy.acml dcopy.acml ccopy.acml zcopy.acml \ sswap.acml dswap.acml cswap.acml zswap.acml \ @@ -199,25 +258,36 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \ sasum.acml dasum.acml casum.acml zasum.acml \ ssymv.acml dsymv.acml csymv.acml zsymv.acml \ chemv.acml zhemv.acml \ + chbmv.acml zhbmv.acml \ + chpmv.acml zhpmv.acml \ chemm.acml zhemm.acml \ cherk.acml zherk.acml \ cher2k.acml zher2k.acml \ sgemv.acml dgemv.acml cgemv.acml zgemv.acml \ + strmv.acml dtrmv.acml ctrmv.acml ztrmv.acml \ + strsv.acml dtrsv.acml ctrsv.acml ztrsv.acml \ sgeev.acml dgeev.acml cgeev.acml zgeev.acml \ sgesv.acml dgesv.acml cgesv.acml zgesv.acml \ sgetri.acml dgetri.acml cgetri.acml zgetri.acml \ spotrf.acml dpotrf.acml cpotrf.acml zpotrf.acml \ - ssymm.acml dsymm.acml csymm.acml zsymm.acml + ssymm.acml dsymm.acml csymm.acml zsymm.acml \ + saxpby.goto daxpby.goto caxpby.goto zaxpby.goto atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \ scholesky.atlas dcholesky.atlas ccholesky.atlas zcholesky.atlas \ sgemm.atlas dgemm.atlas cgemm.atlas zgemm.atlas \ strmm.atlas dtrmm.atlas ctrmm.atlas ztrmm.atlas \ strsm.atlas dtrsm.atlas ctrsm.atlas ztrsm.atlas \ + sspr.atlas dspr.atlas \ + sspr2.atlas dspr2.atlas \ + ssyr.atlas dsyr.atlas \ + ssyr2.atlas dsyr2.atlas \ ssyrk.atlas dsyrk.atlas csyrk.atlas zsyrk.atlas \ ssyr2k.atlas dsyr2k.atlas csyr2k.atlas zsyr2k.atlas \ sger.atlas dger.atlas cger.atlas zger.atlas\ sdot.atlas ddot.atlas \ + srot.atlas drot.atlas csrot.atlas zdrot.atlas \ + srotm.atlas drotm.atlas \ saxpy.atlas daxpy.atlas caxpy.atlas zaxpy.atlas \ scopy.atlas dcopy.atlas ccopy.atlas zcopy.atlas \ sswap.atlas dswap.atlas cswap.atlas zswap.atlas \ @@ -225,28 +295,39 @@ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \ sasum.atlas dasum.atlas casum.atlas zasum.atlas \ ssymv.atlas dsymv.atlas csymv.atlas zsymv.atlas \ chemv.atlas zhemv.atlas \ + chbmv.atlas zhbmv.atlas \ + chpmv.atlas zhpmv.atlas \ chemm.acml zhemm.acml \ chemm.atlas zhemm.atlas \ cherk.atlas zherk.atlas \ cher2k.atlas zher2k.atlas \ sgemv.atlas dgemv.atlas cgemv.atlas zgemv.atlas \ + strmv.atlas dtrmv.atlas ctrmv.atlas ztrmv.atlas \ + strsv.atlas dtrsv.atlas ctrsv.atlas ztrsv.atlas \ sgeev.atlas dgeev.atlas cgeev.atlas zgeev.atlas \ sgesv.atlas dgesv.atlas cgesv.atlas zgesv.atlas \ sgetri.atlas dgetri.atlas cgetri.atlas zgetri.atlas \ spotrf.atlas dpotrf.atlas cpotrf.atlas zpotrf.atlas \ ssymm.atlas dsymm.atlas csymm.atlas zsymm.atlas \ isamax.atlas idamax.atlas icamax.atlas izamax.atlas \ - snrm2.goto dnrm2.goto scnrm2.goto dznrm2.goto + snrm2.goto dnrm2.goto scnrm2.goto dznrm2.goto \ + saxpby.goto daxpby.goto caxpby.goto zaxpby.goto mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \ scholesky.mkl dcholesky.mkl ccholesky.mkl zcholesky.mkl \ sgemm.mkl dgemm.mkl cgemm.mkl zgemm.mkl \ strmm.mkl dtrmm.mkl ctrmm.mkl ztrmm.mkl \ strsm.mkl dtrsm.mkl ctrsm.mkl ztrsm.mkl \ + sspr.mkl dspr.mkl \ + sspr2.mkl dspr2.mkl \ + ssyr.mkl dsyr.mkl \ + ssyr2.mkl dsyr2.mkl \ ssyrk.mkl dsyrk.mkl csyrk.mkl zsyrk.mkl \ ssyr2k.mkl dsyr2k.mkl csyr2k.mkl zsyr2k.mkl \ sger.mkl dger.mkl cger.mkl zger.mkl \ sdot.mkl ddot.mkl cdot.mkl zdot.mkl \ + srot.atlas drot.atlas csrot.atlas zdrot.atlas \ + srotm.atlas drotm.atlas \ saxpy.mkl daxpy.mkl caxpy.mkl zaxpy.mkl \ scopy.mkl dcopy.mkl ccopy.mkl zcopy.mkl \ sswap.mkl dswap.mkl cswap.mkl zswap.mkl \ @@ -254,15 +335,20 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \ sasum.mkl dasum.mkl casum.mkl zasum.mkl \ ssymv.mkl dsymv.mkl csymv.mkl zsymv.mkl \ chemv.mkl zhemv.mkl \ + chbmv.mkl zhbmv.mkl \ + chpmv.mkl zhpmv.mkl \ chemm.mkl zhemm.mkl \ cherk.mkl zherk.mkl \ cher2k.mkl zher2k.mkl \ sgemv.mkl dgemv.mkl cgemv.mkl zgemv.mkl \ + strmv.mkl dtrmv.mkl ctrmv.mkl ztrmv.mkl \ + strsv.mkl dtrsv.mkl ctrsv.mkl ztrsv.mkl \ sgeev.mkl dgeev.mkl cgeev.mkl zgeev.mkl \ sgesv.mkl dgesv.mkl cgesv.mkl zgesv.mkl \ sgetri.mkl dgetri.mkl cgetri.mkl zgetri.mkl \ spotrf.mkl dpotrf.mkl cpotrf.mkl zpotrf.mkl \ - ssymm.mkl dsymm.mkl csymm.mkl zsymm.mkl + ssymm.mkl dsymm.mkl csymm.mkl zsymm.mkl \ + saxpby.goto daxpby.goto caxpby.goto zaxpby.goto @@ -280,10 +366,16 @@ veclib :: slinpack.veclib dlinpack.veclib clinpack.veclib zlinpack.veclib \ sgemm.veclib dgemm.veclib cgemm.veclib zgemm.veclib \ strmm.veclib dtrmm.veclib ctrmm.veclib ztrmm.veclib \ strsm.veclib dtrsm.veclib ctrsm.veclib ztrsm.veclib \ + sspr.veclib dspr.veclib \ + sspr2.veclib dspr2.veclib \ + ssyr.veclib dsyr.veclib \ + ssyr2.veclib dsyr2.veclib \ ssyrk.veclib dsyrk.veclib csyrk.veclib zsyrk.veclib \ ssyr2k.veclib dsyr2k.veclib csyr2k.veclib zsyr2k.veclib \ sger.veclib dger.veclib cger.veclib zger.veclib \ sdot.veclib ddot.veclib cdot.veclib zdot.veclib \ + srot.veclib drot.veclib csrot.veclib zdrot.veclib \ + srotm.veclib drotm.veclib \ saxpy.veclib daxpy.veclib caxpy.veclib zaxpy.veclib \ scopy.veclib dcopy.veclib ccopy.veclib zcopy.veclib \ sswap.veclib dswap.veclib cswap.veclib zswap.veclib \ @@ -291,15 +383,20 @@ veclib :: slinpack.veclib dlinpack.veclib clinpack.veclib zlinpack.veclib \ sasum.veclib dasum.veclib casum.veclib zasum.veclib \ ssymv.veclib dsymv.veclib csymv.veclib zsymv.veclib \ chemv.veclib zhemv.veclib \ + chbmv.veclib zhbmv.veclib \ + chpmv.veclib zhpmv.veclib \ chemm.veclib zhemm.veclib \ cherk.veclib zherk.veclib \ cher2k.veclib zher2k.veclib \ sgemv.veclib dgemv.veclib cgemv.veclib zgemv.veclib \ + strmv.veclib dtrmv.veclib ctrmv.veclib ztrmv.veclib \ + strsv.veclib dtrsv.veclib ctrsv.veclib ztrsv.veclib \ sgeev.veclib dgeev.veclib cgeev.veclib zgeev.veclib \ sgesv.veclib dgesv.veclib cgesv.veclib zgesv.veclib \ sgetri.veclib dgetri.veclib cgetri.veclib zgetri.veclib \ spotrf.veclib dpotrf.veclib cpotrf.veclib zpotrf.veclib \ - ssymm.veclib dsymm.veclib csymm.veclib zsymm.veclib + ssymm.veclib dsymm.veclib csymm.veclib zsymm.veclib \ + saxpby.goto daxpby.goto caxpby.goto zaxpby.goto goto_3m :: cgemm3m.goto zgemm3m.goto @@ -768,6 +865,130 @@ ztrsm.veclib : ztrsm.$(SUFFIX) ztrsm.essl : ztrsm.$(SUFFIX) -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBESSL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) +##################################### Ssyr #################################################### +ssyr.goto : ssyr.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +ssyr.acml : ssyr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ssyr.atlas : ssyr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ssyr.mkl : ssyr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ssyr.veclib : ssyr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) +##################################### Dsyr #################################################### +dsyr.goto : dsyr.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +dsyr.acml : dsyr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dsyr.atlas : dsyr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dsyr.mkl : dsyr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dsyr.veclib : dsyr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Sspr #################################################### +sspr.goto : sspr.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +sspr.acml : sspr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +sspr.atlas : sspr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +sspr.mkl : sspr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +sspr.veclib : sspr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Dspr #################################################### +dspr.goto : dspr.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +dspr.acml : dspr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dspr.atlas : dspr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dspr.mkl : dspr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dspr.veclib : dspr.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Sspr2 #################################################### +sspr2.goto : sspr2.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +sspr2.acml : sspr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +sspr2.atlas : sspr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +sspr2.mkl : sspr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +sspr2.veclib : sspr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Dspr2 #################################################### +dspr2.goto : dspr2.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +dspr2.acml : dspr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dspr2.atlas : dspr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dspr2.mkl : dspr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dspr2.veclib : dspr2.$(SUFFIX) + +##################################### Ssyr2 #################################################### +ssyr2.goto : ssyr2.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +ssyr2.acml : ssyr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ssyr2.atlas : ssyr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ssyr2.mkl : ssyr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ssyr2.veclib : ssyr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) +##################################### Dsyr2 #################################################### +dsyr2.goto : dsyr2.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +dsyr2.acml : dsyr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dsyr2.atlas : dsyr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dsyr2.mkl : dsyr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dsyr2.veclib : dsyr2.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) ##################################### Ssyrk #################################################### ssyrk.goto : ssyrk.$(SUFFIX) ../$(LIBNAME) @@ -1069,6 +1290,138 @@ zgemv.mkl : zgemv.$(SUFFIX) zgemv.veclib : zgemv.$(SUFFIX) -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) +##################################### Strmv #################################################### +strmv.goto : strmv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +strmv.acml : strmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +strmv.atlas : strmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +strmv.mkl : strmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +strmv.veclib : strmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Dtrmv #################################################### +dtrmv.goto : dtrmv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +dtrmv.acml : dtrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dtrmv.atlas : dtrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dtrmv.mkl : dtrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dtrmv.veclib : dtrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Ctrmv #################################################### + +ctrmv.goto : ctrmv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +ctrmv.acml : ctrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ctrmv.atlas : ctrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ctrmv.mkl : ctrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ctrmv.veclib : ctrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Ztrmv #################################################### + +ztrmv.goto : ztrmv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +ztrmv.acml : ztrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ztrmv.atlas : ztrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ztrmv.mkl : ztrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ztrmv.veclib : ztrmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Strsv #################################################### +strsv.goto : strsv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +strsv.acml : strsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +strsv.atlas : strsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +strsv.mkl : strsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +strsv.veclib : strsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Dtrsv #################################################### +dtrsv.goto : dtrsv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +dtrsv.acml : dtrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dtrsv.atlas : dtrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dtrsv.mkl : dtrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +dtrsv.veclib : dtrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Ctrsv #################################################### + +ctrsv.goto : ctrsv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +ctrsv.acml : ctrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ctrsv.atlas : ctrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ctrsv.mkl : ctrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ctrsv.veclib : ctrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Ztrsv #################################################### + +ztrsv.goto : ztrsv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +ztrsv.acml : ztrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ztrsv.atlas : ztrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ztrsv.mkl : ztrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +ztrsv.veclib : ztrsv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + ##################################### Sger #################################################### sger.goto : sger.$(SUFFIX) ../$(LIBNAME) $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm @@ -1428,7 +1781,70 @@ zhemv.mkl : zhemv.$(SUFFIX) zhemv.veclib : zhemv.$(SUFFIX) -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) +##################################### Chbmv #################################################### +chbmv.goto : chbmv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +chbmv.acml : chbmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +chbmv.atlas : chbmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +chbmv.mkl : chbmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +chbmv.veclib : chbmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) +##################################### Zhbmv #################################################### + +zhbmv.goto : zhbmv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +zhbmv.acml : zhbmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zhbmv.atlas : zhbmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zhbmv.mkl : zhbmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zhbmv.veclib : zhbmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) +##################################### Chpmv #################################################### + +chpmv.goto : chpmv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +chpmv.acml : chpmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +chpmv.atlas : chpmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +chpmv.mkl : chpmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +chpmv.veclib : chpmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) +##################################### Zhpmv #################################################### + +zhpmv.goto : zhpmv.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +zhpmv.acml : zhpmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zhpmv.atlas : zhpmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zhpmv.mkl : zhpmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zhpmv.veclib : zhpmv.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) ##################################### Sdot #################################################### sdot.goto : sdot.$(SUFFIX) ../$(LIBNAME) $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm @@ -1525,6 +1941,69 @@ drot.mkl : drot.$(SUFFIX) drot.veclib : drot.$(SUFFIX) $(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) +##################################### csrot #################################################### +csrot.goto : csrot.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +csrot.acml : csrot.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +csrot.atlas : csrot.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +csrot.mkl : csrot.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +csrot.veclib : csrot.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### zdrot #################################################### +zdrot.goto : zdrot.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +zdrot.acml : zdrot.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zdrot.atlas : zdrot.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zdrot.mkl : zdrot.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zdrot.veclib : zdrot.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### srotm #################################################### +srotm.goto : srotm.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +srotm.acml : srotm.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +srotm.atlas : srotm.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +srotm.mkl : srotm.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +srotm.veclib : srotm.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### drotm #################################################### +drotm.goto : drotm.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +drotm.acml : drotm.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +drotm.atlas : drotm.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +drotm.mkl : drotm.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +drotm.veclib : drotm.$(SUFFIX) + $(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) ##################################### Saxpy #################################################### saxpy.goto : saxpy.$(SUFFIX) ../$(LIBNAME) @@ -1592,7 +2071,72 @@ zaxpy.mkl : zaxpy.$(SUFFIX) zaxpy.veclib : zaxpy.$(SUFFIX) -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) +##################################### Saxpby #################################################### +saxpby.goto : saxpby.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm +saxpby.acml : saxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +saxpby.atlas : saxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +saxpby.mkl : saxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +saxpby.veclib : saxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Daxpby #################################################### +daxpby.goto : daxpby.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +daxpby.acml : daxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +daxpby.atlas : daxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +daxpby.mkl : daxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +daxpby.veclib : daxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Caxpby #################################################### + +caxpby.goto : caxpby.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +caxpby.acml : caxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +caxpby.atlas : caxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +caxpby.mkl : caxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +caxpby.veclib : caxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +##################################### Zaxpby #################################################### + +zaxpby.goto : zaxpby.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +zaxpby.acml : zaxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zaxpby.atlas : zaxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zaxpby.mkl : zaxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + +zaxpby.veclib : zaxpby.$(SUFFIX) + -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) + ##################################### Scopy #################################################### scopy.goto : scopy.$(SUFFIX) ../$(LIBNAME) $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm @@ -2078,6 +2622,30 @@ ctrsm.$(SUFFIX) : trsm.c ztrsm.$(SUFFIX) : trsm.c $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ +ssyr.$(SUFFIX) : syr.c + $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ + +dsyr.$(SUFFIX) : syr.c + $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^ + +sspr.$(SUFFIX) : spr.c + $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ + +dspr.$(SUFFIX) : spr.c + $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^ + +sspr2.$(SUFFIX) : spr2.c + $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ + +dspr2.$(SUFFIX) : spr2.c + $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^ + +ssyr2.$(SUFFIX) : syr2.c + $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ + +dsyr2.$(SUFFIX) : syr2.c + $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^ + ssyrk.$(SUFFIX) : syrk.c $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ @@ -2132,6 +2700,30 @@ cgemv.$(SUFFIX) : gemv.c zgemv.$(SUFFIX) : gemv.c $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ +strmv.$(SUFFIX) : trmv.c + $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ + +dtrmv.$(SUFFIX) : trmv.c + $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^ + +ctrmv.$(SUFFIX) : trmv.c + $(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^ + +ztrmv.$(SUFFIX) : trmv.c + $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ + +strsv.$(SUFFIX) : trsv.c + $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ + +dtrsv.$(SUFFIX) : trsv.c + $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^ + +ctrsv.$(SUFFIX) : trsv.c + $(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^ + +ztrsv.$(SUFFIX) : trsv.c + $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ + sger.$(SUFFIX) : ger.c $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ @@ -2199,6 +2791,18 @@ chemv.$(SUFFIX) : hemv.c zhemv.$(SUFFIX) : hemv.c $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ +chbmv.$(SUFFIX) : hbmv.c + $(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^ + +zhbmv.$(SUFFIX) : hbmv.c + $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ + +chpmv.$(SUFFIX) : hpmv.c + $(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^ + +zhpmv.$(SUFFIX) : hpmv.c + $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ + sdot.$(SUFFIX) : dot.c $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ @@ -2231,6 +2835,18 @@ caxpy.$(SUFFIX) : axpy.c zaxpy.$(SUFFIX) : axpy.c $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ +saxpby.$(SUFFIX) : axpby.c + $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ + +daxpby.$(SUFFIX) : axpby.c + $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^ + +caxpby.$(SUFFIX) : axpby.c + $(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^ + +zaxpby.$(SUFFIX) : axpby.c + $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ + scopy.$(SUFFIX) : copy.c $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ @@ -2300,7 +2916,17 @@ srot.$(SUFFIX) : rot.c drot.$(SUFFIX) : rot.c $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^ +csrot.$(SUFFIX) : rot.c + $(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^ +zdrot.$(SUFFIX) : rot.c + $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ + +srotm.$(SUFFIX) : rotm.c + $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ + +drotm.$(SUFFIX) : rotm.c + $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^ diff --git a/benchmark/axpby.c b/benchmark/axpby.c new file mode 100644 index 000000000..3b3dd9979 --- /dev/null +++ b/benchmark/axpby.c @@ -0,0 +1,202 @@ +/*************************************************************************** +Copyright (c) 2014, 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 +#include +#ifdef __CYGWIN32__ +#include +#endif +#include "common.h" + + +#undef AXPBY + +#ifdef COMPLEX +#ifdef DOUBLE +#define AXPBY BLASFUNC(zaxpby) +#else +#define AXPBY BLASFUNC(caxpby) +#endif +#else +#ifdef DOUBLE +#define AXPBY BLASFUNC(daxpby) +#else +#define AXPBY BLASFUNC(saxpby) +#endif +#endif + +#if defined(__WIN32__) || defined(__WIN64__) + +#ifndef DELTA_EPOCH_IN_MICROSECS +#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL +#endif + +int gettimeofday(struct timeval *tv, void *tz){ + + FILETIME ft; + unsigned __int64 tmpres = 0; + static int tzflag; + + if (NULL != tv) + { + GetSystemTimeAsFileTime(&ft); + + tmpres |= ft.dwHighDateTime; + tmpres <<= 32; + tmpres |= ft.dwLowDateTime; + + /*converting file time to unix epoch*/ + tmpres /= 10; /*convert into microseconds*/ + tmpres -= DELTA_EPOCH_IN_MICROSECS; + tv->tv_sec = (long)(tmpres / 1000000UL); + tv->tv_usec = (long)(tmpres % 1000000UL); + } + + return 0; +} + +#endif + +#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0 + +static void *huge_malloc(BLASLONG size){ + int shmid; + void *address; + +#ifndef SHM_HUGETLB +#define SHM_HUGETLB 04000 +#endif + + if ((shmid =shmget(IPC_PRIVATE, + (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1), + SHM_HUGETLB | IPC_CREAT |0600)) < 0) { + printf( "Memory allocation failed(shmget).\n"); + exit(1); + } + + address = shmat(shmid, NULL, SHM_RND); + + if ((BLASLONG)address == -1){ + printf( "Memory allocation failed(shmat).\n"); + exit(1); + } + + shmctl(shmid, IPC_RMID, 0); + + return address; +} + +#define malloc huge_malloc + +#endif + +int main(int argc, char *argv[]){ + + FLOAT *x, *y; + FLOAT alpha[2] = { 2.0, 2.0 }; + FLOAT beta[2] = {2.0, 2.0}; + blasint m, i; + blasint inc_x=1,inc_y=1; + int loops = 1; + int l; + char *p; + + int from = 1; + int to = 200; + int step = 1; + + struct timeval start, stop; + double time1,timeg; + + argc--;argv++; + + if (argc > 0) { from = atol(*argv); argc--; argv++;} + if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++;} + if (argc > 0) { step = atol(*argv); argc--; argv++;} + + if ((p = getenv("OPENBLAS_LOOPS"))) loops = atoi(p); + if ((p = getenv("OPENBLAS_INCX"))) inc_x = atoi(p); + if ((p = getenv("OPENBLAS_INCY"))) inc_y = atoi(p); + + fprintf(stderr, "From : %3d To : %3d Step = %3d Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,inc_x,inc_y,loops); + + if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + + if (( y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + +#ifdef linux + srandom(getpid()); +#endif + + fprintf(stderr, " SIZE Flops\n"); + + for(m = from; m <= to; m += step) + { + + timeg=0; + + fprintf(stderr, " %6d : ", (int)m); + + + for(i = 0; i < m * COMPSIZE * abs(inc_x); i++){ + x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + for(i = 0; i < m * COMPSIZE * abs(inc_y); i++){ + y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + for (l=0; l +#include +#ifdef __CYGWIN32__ +#include +#endif +#include "common.h" + + +#undef HBMV + + +#ifdef DOUBLE +#define HBMV BLASFUNC(zhbmv) +#else +#define HBMV BLASFUNC(chbmv) +#endif + + +#if defined(__WIN32__) || defined(__WIN64__) + +#ifndef DELTA_EPOCH_IN_MICROSECS +#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL +#endif + +int gettimeofday(struct timeval *tv, void *tz) { + + FILETIME ft; + unsigned __int64 tmpres = 0; + static int tzflag; + + if (NULL != tv) + { + GetSystemTimeAsFileTime(&ft); + + tmpres |= ft.dwHighDateTime; + tmpres <<= 32; + tmpres |= ft.dwLowDateTime; + + /*converting file time to unix epoch*/ + tmpres /= 10; /*convert into microseconds*/ + tmpres -= DELTA_EPOCH_IN_MICROSECS; + tv->tv_sec = (long)(tmpres / 1000000UL); + tv->tv_usec = (long)(tmpres % 1000000UL); + } + + return 0; +} + +#endif + +#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0 + +static void *huge_malloc(BLASLONG size) { + int shmid; + void *address; + +#ifndef SHM_HUGETLB +#define SHM_HUGETLB 04000 +#endif + + if ((shmid =shmget(IPC_PRIVATE, + (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1), + SHM_HUGETLB | IPC_CREAT |0600)) < 0) { + printf( "Memory allocation failed(shmget).\n"); + exit(1); + } + + address = shmat(shmid, NULL, SHM_RND); + + if ((BLASLONG)address == -1){ + printf( "Memory allocation failed(shmat).\n"); + exit(1); + } + + shmctl(shmid, IPC_RMID, 0); + + return address; +} + +#define malloc huge_malloc + +#endif + +int main(int argc, char *argv[]){ + + FLOAT *a, *x, *y; + FLOAT alpha[] = {1.0, 1.0}; + FLOAT beta [] = {0.0, 0.0}; + blasint k = 1; + char uplo='L'; + blasint m, i, j; + blasint inc_x=1, inc_y=1; + int loops = 1; + int l; + char *p; + + int from = 1; + int to = 200; + int step = 1; + + struct timeval start, stop; + double time1,timeg; + + argc--;argv++; + + if (argc > 0) { from = atol(*argv); argc--; argv++;} + if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++;} + if (argc > 0) { step = atol(*argv); argc--; argv++;} + + if ((p = getenv("OPENBLAS_LOOPS"))) loops = atoi(p); + if ((p = getenv("OPENBLAS_INCX"))) inc_x = atoi(p); + if ((p = getenv("OPENBLAS_INCY"))) inc_y = atoi(p); + if ((p = getenv("OPENBLAS_UPLO"))) uplo=*p; + if ((p = getenv("OPENBLAS_K"))) k = atoi(p); + + fprintf(stderr, "From : %3d To : %3d Step = %3d Uplo = '%c' k = %d Inc_x = %d Inc_y = %d Loops = %d\n", + from, to, step, uplo, k, inc_x, inc_y, loops); + + if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL) { + fprintf(stderr,"Out of Memory!!\n"); + exit(1); + } + + if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL) { + fprintf(stderr,"Out of Memory!!\n"); + exit(1); + } + + if (( y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL) { + fprintf(stderr,"Out of Memory!!\n"); + exit(1); + } + +#ifdef linux + srandom(getpid()); +#endif + + fprintf(stderr, " SIZE Flops\n"); + + for(m = from; m <= to; m += step) { + + timeg=0; + + fprintf(stderr, " %6dx%d : ", (int)m, (int)m); + + for(j = 0; j < m; j++) { + for(i = 0; i < m * COMPSIZE; i++) { + a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + } + + for (l = 0; l < loops; l++) { + + for (i = 0; i < m * COMPSIZE * abs(inc_x); i++) { + x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + for (i = 0; i < m * COMPSIZE * abs(inc_y); i++) { + y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + gettimeofday( &start, (struct timezone *)0); + + HBMV (&uplo, &m, &k, alpha, a, &m, x, &inc_x, beta, y, &inc_y ); + + gettimeofday( &stop, (struct timezone *)0); + + time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; + + timeg += time1; + + } + + timeg /= loops; + + fprintf(stderr, " %10.2f MFlops\n", + COMPSIZE * COMPSIZE * 2. * (double)(2 * k + 1) * (double)m / timeg * 1.e-6); + } + + return 0; +} + +// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__"))); diff --git a/benchmark/hemm.c b/benchmark/hemm.c index 318c407ba..2fe0f5c5f 100644 --- a/benchmark/hemm.c +++ b/benchmark/hemm.c @@ -164,9 +164,9 @@ int main(int argc, char *argv[]){ for(j = 0; j < m; j++){ for(i = 0; i < m * COMPSIZE; i++){ - a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; - b[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; - c[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + b[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + c[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; } } @@ -178,8 +178,6 @@ int main(int argc, char *argv[]){ time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; - gettimeofday( &start, (struct timezone *)0); - fprintf(stderr, " %10.2f MFlops\n", COMPSIZE * COMPSIZE * 2. * (double)m * (double)m * (double)m / time1 * 1.e-6); diff --git a/benchmark/hemv.c b/benchmark/hemv.c index 05028e3cf..b6ff512ce 100644 --- a/benchmark/hemv.c +++ b/benchmark/hemv.c @@ -167,7 +167,7 @@ int main(int argc, char *argv[]){ for(j = 0; j < m; j++){ for(i = 0; i < m * COMPSIZE; i++){ - a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; } } diff --git a/benchmark/her2k.c b/benchmark/her2k.c index 028e2718f..a0772feff 100644 --- a/benchmark/her2k.c +++ b/benchmark/her2k.c @@ -163,9 +163,9 @@ int main(int argc, char *argv[]){ for(j = 0; j < m; j++){ for(i = 0; i < m * COMPSIZE; i++){ - a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; - b[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; - c[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + b[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + c[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; } } @@ -177,8 +177,6 @@ int main(int argc, char *argv[]){ time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; - gettimeofday( &start, (struct timezone *)0); - fprintf(stderr, " %10.2f MFlops\n", COMPSIZE * COMPSIZE * 2. * (double)m * (double)m * (double)m / time1 * 1.e-6); diff --git a/benchmark/herk.c b/benchmark/herk.c index d2e25ff46..eed8ed738 100644 --- a/benchmark/herk.c +++ b/benchmark/herk.c @@ -162,8 +162,8 @@ int main(int argc, char *argv[]){ for(j = 0; j < m; j++){ for(i = 0; i < m * COMPSIZE; i++){ - a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; - c[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + c[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; } } @@ -175,8 +175,6 @@ int main(int argc, char *argv[]){ time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; - gettimeofday( &start, (struct timezone *)0); - fprintf(stderr, " %10.2f MFlops\n", COMPSIZE * COMPSIZE * 1. * (double)m * (double)m * (double)m / time1 * 1.e-6); diff --git a/benchmark/hpmv.c b/benchmark/hpmv.c new file mode 100644 index 000000000..6e6634fcf --- /dev/null +++ b/benchmark/hpmv.c @@ -0,0 +1,207 @@ +/*************************************************************************** +Copyright (c) 2014, 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 +#include +#ifdef __CYGWIN32__ +#include +#endif +#include "common.h" + + +#undef HPMV + + +#ifdef DOUBLE +#define HPMV BLASFUNC(zhpmv) +#else +#define HPMV BLASFUNC(chpmv) +#endif + + +#if defined(__WIN32__) || defined(__WIN64__) + +#ifndef DELTA_EPOCH_IN_MICROSECS +#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL +#endif + +int gettimeofday(struct timeval *tv, void *tz) { + + FILETIME ft; + unsigned __int64 tmpres = 0; + static int tzflag; + + if (NULL != tv) + { + GetSystemTimeAsFileTime(&ft); + + tmpres |= ft.dwHighDateTime; + tmpres <<= 32; + tmpres |= ft.dwLowDateTime; + + /*converting file time to unix epoch*/ + tmpres /= 10; /*convert into microseconds*/ + tmpres -= DELTA_EPOCH_IN_MICROSECS; + tv->tv_sec = (long)(tmpres / 1000000UL); + tv->tv_usec = (long)(tmpres % 1000000UL); + } + + return 0; +} + +#endif + +#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0 + +static void *huge_malloc(BLASLONG size) { + int shmid; + void *address; + +#ifndef SHM_HUGETLB +#define SHM_HUGETLB 04000 +#endif + + if ((shmid =shmget(IPC_PRIVATE, + (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1), + SHM_HUGETLB | IPC_CREAT |0600)) < 0) { + printf( "Memory allocation failed(shmget).\n"); + exit(1); + } + + address = shmat(shmid, NULL, SHM_RND); + + if ((BLASLONG)address == -1){ + printf( "Memory allocation failed(shmat).\n"); + exit(1); + } + + shmctl(shmid, IPC_RMID, 0); + + return address; +} + +#define malloc huge_malloc + +#endif + +int main(int argc, char *argv[]){ + + FLOAT *a, *x, *y; + FLOAT alpha[] = {1.0, 1.0}; + FLOAT beta [] = {1.0, 1.0}; + char uplo='L'; + blasint m, i, j; + blasint inc_x=1, inc_y=1; + int loops = 1; + int l; + char *p; + + int from = 1; + int to = 200; + int step = 1; + + struct timeval start, stop; + double time1,timeg; + + argc--;argv++; + + if (argc > 0) { from = atol(*argv); argc--; argv++;} + if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++;} + if (argc > 0) { step = atol(*argv); argc--; argv++;} + + if ((p = getenv("OPENBLAS_LOOPS"))) loops = atoi(p); + if ((p = getenv("OPENBLAS_INCX"))) inc_x = atoi(p); + if ((p = getenv("OPENBLAS_INCY"))) inc_y = atoi(p); + if ((p = getenv("OPENBLAS_UPLO"))) uplo=*p; + + fprintf(stderr, "From : %3d To : %3d Step = %3d Uplo = '%c' Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,uplo,inc_x,inc_y,loops); + + if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL) { + fprintf(stderr,"Out of Memory!!\n"); + exit(1); + } + + if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL) { + fprintf(stderr,"Out of Memory!!\n"); + exit(1); + } + + if (( y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL) { + fprintf(stderr,"Out of Memory!!\n"); + exit(1); + } + +#ifdef linux + srandom(getpid()); +#endif + + fprintf(stderr, " SIZE Flops\n"); + + for(m = from; m <= to; m += step) { + + timeg=0; + + fprintf(stderr, " %6dx%d : ", (int)m, (int)m); + + for(j = 0; j < m; j++) { + for(i = 0; i < m * COMPSIZE; i++) { + a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + } + + for (l = 0; l < loops; l++) { + + for (i = 0; i < m * COMPSIZE * abs(inc_x); i++) { + x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + for (i = 0; i < m * COMPSIZE * abs(inc_y); i++) { + y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + gettimeofday( &start, (struct timezone *)0); + + HPMV (&uplo, &m, alpha, a, x, &inc_x, beta, y, &inc_y ); + + gettimeofday( &stop, (struct timezone *)0); + + time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; + + timeg += time1; + + } + + timeg /= loops; + + fprintf(stderr, " %10.2f MFlops\n", + COMPSIZE * COMPSIZE * 2. * (double)m * (double)m / timeg * 1.e-6); + } + + return 0; +} + +// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__"))); diff --git a/benchmark/linpack.c b/benchmark/linpack.c index 7d5c87163..e4b20e99d 100644 --- a/benchmark/linpack.c +++ b/benchmark/linpack.c @@ -186,7 +186,7 @@ int main(int argc, char *argv[]){ for(j = 0; j < m; j++){ for(i = 0; i < m * COMPSIZE; i++){ - a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; } } @@ -194,7 +194,7 @@ int main(int argc, char *argv[]){ for (j = 0; j < m; ++j) { for (i = 0; i < m * COMPSIZE; ++i) { - b[i] += a[i + j * m * COMPSIZE]; + b[i] += a[(long)i + (long)j * (long)m * COMPSIZE]; } } diff --git a/benchmark/potrf.c b/benchmark/potrf.c index 1d714549b..580e46072 100644 --- a/benchmark/potrf.c +++ b/benchmark/potrf.c @@ -170,46 +170,46 @@ int main(int argc, char *argv[]){ #ifndef COMPLEX if (uplos & 1) { for (j = 0; j < m; j++) { - for(i = 0; i < j; i++) a[i + j * m] = 0.; - a[j + j * m] = ((double) rand() / (double) RAND_MAX) + 8.; - for(i = j + 1; i < m; i++) a[i + j * m] = ((double) rand() / (double) RAND_MAX) - 0.5; + for(i = 0; i < j; i++) a[(long)i + (long)j * (long)m] = 0.; + a[(long)j + (long)j * (long)m] = ((double) rand() / (double) RAND_MAX) + 8.; + for(i = j + 1; i < m; i++) a[(long)i + (long)j * (long)m] = ((double) rand() / (double) RAND_MAX) - 0.5; } } else { for (j = 0; j < m; j++) { - for(i = 0; i < j; i++) a[i + j * m] = ((double) rand() / (double) RAND_MAX) - 0.5; - a[j + j * m] = ((double) rand() / (double) RAND_MAX) + 8.; - for(i = j + 1; i < m; i++) a[i + j * m] = 0.; + for(i = 0; i < j; i++) a[(long)i + (long)j * (long)m] = ((double) rand() / (double) RAND_MAX) - 0.5; + a[(long)j + (long)j * (long)m] = ((double) rand() / (double) RAND_MAX) + 8.; + for(i = j + 1; i < m; i++) a[(long)i + (long)j * (long)m] = 0.; } } #else if (uplos & 1) { for (j = 0; j < m; j++) { for(i = 0; i < j; i++) { - a[(i + j * m) * 2 + 0] = 0.; - a[(i + j * m) * 2 + 1] = 0.; + a[((long)i + (long)j * (long)m) * 2 + 0] = 0.; + a[((long)i + (long)j * (long)m) * 2 + 1] = 0.; } - a[(j + j * m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) + 8.; - a[(j + j * m) * 2 + 1] = 0.; + a[((long)j + (long)j * (long)m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) + 8.; + a[((long)j + (long)j * (long)m) * 2 + 1] = 0.; for(i = j + 1; i < m; i++) { - a[(i + j * m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) - 0.5; - a[(i + j * m) * 2 + 1] = ((double) rand() / (double) RAND_MAX) - 0.5; + a[((long)i + (long)j * (long)m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) - 0.5; + a[((long)i + (long)j * (long)m) * 2 + 1] = ((double) rand() / (double) RAND_MAX) - 0.5; } } } else { for (j = 0; j < m; j++) { for(i = 0; i < j; i++) { - a[(i + j * m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) - 0.5; - a[(i + j * m) * 2 + 1] = ((double) rand() / (double) RAND_MAX) - 0.5; + a[((long)i + (long)j * (long)m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) - 0.5; + a[((long)i + (long)j * (long)m) * 2 + 1] = ((double) rand() / (double) RAND_MAX) - 0.5; } - a[(j + j * m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) + 8.; - a[(j + j * m) * 2 + 1] = 0.; + a[((long)j + (long)j * (long)m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) + 8.; + a[((long)j + (long)j * (long)m) * 2 + 1] = 0.; for(i = j + 1; i < m; i++) { - a[(i + j * m) * 2 + 0] = 0.; - a[(i + j * m) * 2 + 1] = 0.; + a[((long)i + (long)j * (long)m) * 2 + 0] = 0.; + a[((long)i + (long)j * (long)m) * 2 + 1] = 0.; } } } diff --git a/benchmark/rot.c b/benchmark/rot.c index 3ff783cc6..8ec8b1d97 100644 --- a/benchmark/rot.c +++ b/benchmark/rot.c @@ -32,9 +32,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #endif #include "common.h" +#undef ROT -#undef DOT - +#ifndef COMPLEX #ifdef DOUBLE #define ROT BLASFUNC(drot) @@ -42,6 +42,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define ROT BLASFUNC(srot) #endif +#else + +#ifdef DOUBLE +#define ROT BLASFUNC(zdrot) +#else +#define ROT BLASFUNC(csrot) +#endif + +#endif #if defined(__WIN32__) || defined(__WIN64__) @@ -160,17 +169,16 @@ int main(int argc, char *argv[]){ fprintf(stderr, " %6d : ", (int)m); + for(i = 0; i < m * COMPSIZE * abs(inc_x); i++){ + x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + for(i = 0; i < m * COMPSIZE * abs(inc_y); i++){ + y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } for (l=0; l +#include +#ifdef __CYGWIN32__ +#include +#endif +#include "common.h" + +#undef ROTM + +#ifdef DOUBLE +#define ROTM BLASFUNC(drotm) +#else +#define ROTM BLASFUNC(srotm) +#endif + +#if defined(__WIN32__) || defined(__WIN64__) + +#ifndef DELTA_EPOCH_IN_MICROSECS +#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL +#endif + +int gettimeofday(struct timeval *tv, void *tz) +{ + + FILETIME ft; + unsigned __int64 tmpres = 0; + static int tzflag; + + if (NULL != tv) { + GetSystemTimeAsFileTime(&ft); + + tmpres |= ft.dwHighDateTime; + tmpres <<= 32; + tmpres |= ft.dwLowDateTime; + + /*converting file time to unix epoch*/ + tmpres /= 10; /*convert into microseconds*/ + tmpres -= DELTA_EPOCH_IN_MICROSECS; + tv->tv_sec = (long)(tmpres / 1000000UL); + tv->tv_usec = (long)(tmpres % 1000000UL); + } + + return 0; +} + +#endif + +#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0 + +static void *huge_malloc(BLASLONG size) +{ + int shmid; + void *address; + +#ifndef SHM_HUGETLB +#define SHM_HUGETLB 04000 +#endif + + if ((shmid = + shmget(IPC_PRIVATE, (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1), + SHM_HUGETLB | IPC_CREAT | 0600)) < 0) { + printf("Memory allocation failed(shmget).\n"); + exit(1); + } + + address = shmat(shmid, NULL, SHM_RND); + + if ((BLASLONG)address == -1) { + printf("Memory allocation failed(shmat).\n"); + exit(1); + } + + shmctl(shmid, IPC_RMID, 0); + + return address; +} + +#define malloc huge_malloc + +#endif + +int main(int argc, char *argv[]) +{ + + FLOAT *x, *y; + // FLOAT result; + blasint m, i; + blasint inc_x = 1, inc_y = 1; + FLOAT param[5] = {1, 2.0, 3.0, 4.0, 5.0}; + int loops = 1; + int l; + char *p; + + int from = 1; + int to = 200; + int step = 1; + + struct timeval start, stop; + double time1, timeg; + + argc--; + argv++; + + if (argc > 0) { + from = atol(*argv); + argc--; + argv++; + } + if (argc > 0) { + to = MAX(atol(*argv), from); + argc--; + argv++; + } + if (argc > 0) { + step = atol(*argv); + argc--; + argv++; + } + + if ((p = getenv("OPENBLAS_LOOPS"))) + loops = atoi(p); + if ((p = getenv("OPENBLAS_INCX"))) + inc_x = atoi(p); + if ((p = getenv("OPENBLAS_INCY"))) + inc_y = atoi(p); + + fprintf( + stderr, + "From : %3d To : %3d Step = %3d Inc_x = %d Inc_y = %d Loops = %d\n", + from, to, step, inc_x, inc_y, loops); + + if ((x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == + NULL) { + fprintf(stderr, "Out of Memory!!\n"); + exit(1); + } + + if ((y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == + NULL) { + fprintf(stderr, "Out of Memory!!\n"); + exit(1); + } + +#ifdef linux + srandom(getpid()); +#endif + + fprintf(stderr, " SIZE Flops\n"); + + for (m = from; m <= to; m += step) { + + timeg = 0; + + fprintf(stderr, " %6d : ", (int)m); + for (i = 0; i < m * COMPSIZE * abs(inc_x); i++) { + x[i] = ((FLOAT)rand() / (FLOAT)RAND_MAX) - 0.5; + } + + for (i = 0; i < m * COMPSIZE * abs(inc_y); i++) { + y[i] = ((FLOAT)rand() / (FLOAT)RAND_MAX) - 0.5; + } + + for (l = 0; l < loops; l++) { + gettimeofday(&start, (struct timezone *)0); + + ROTM(&m, x, &inc_x, y, &inc_y, param); + + gettimeofday(&stop, (struct timezone *)0); + + time1 = (double)(stop.tv_sec - start.tv_sec) + + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; + + timeg += time1; + } + + timeg /= loops; + + fprintf(stderr, " %10.2f MFlops %10.6f sec\n", + COMPSIZE * COMPSIZE * 6. * (double)m / timeg * 1.e-6, timeg); + } + + return 0; +} diff --git a/benchmark/spr.c b/benchmark/spr.c new file mode 100755 index 000000000..61a972c08 --- /dev/null +++ b/benchmark/spr.c @@ -0,0 +1,198 @@ +/*************************************************************************** +Copyright (c) 2014, 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 +#include +#ifdef __CYGWIN32__ +#include +#endif +#include "common.h" + + +#undef SPR + +#ifdef DOUBLE +#define SPR BLASFUNC(dspr) +#else +#define SPR BLASFUNC(sspr) +#endif + + + +#if defined(__WIN32__) || defined(__WIN64__) + +#ifndef DELTA_EPOCH_IN_MICROSECS +#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL +#endif + +int gettimeofday(struct timeval *tv, void *tz){ + + FILETIME ft; + unsigned __int64 tmpres = 0; + static int tzflag; + + if (NULL != tv) + { + GetSystemTimeAsFileTime(&ft); + + tmpres |= ft.dwHighDateTime; + tmpres <<= 32; + tmpres |= ft.dwLowDateTime; + + /*converting file time to unix epoch*/ + tmpres /= 10; /*convert into microseconds*/ + tmpres -= DELTA_EPOCH_IN_MICROSECS; + tv->tv_sec = (long)(tmpres / 1000000UL); + tv->tv_usec = (long)(tmpres % 1000000UL); + } + + return 0; +} + +#endif + +#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0 + +static void *huge_malloc(BLASLONG size){ + int shmid; + void *address; + +#ifndef SHM_HUGETLB +#define SHM_HUGETLB 04000 +#endif + + if ((shmid =shmget(IPC_PRIVATE, + (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1), + SHM_HUGETLB | IPC_CREAT |0600)) < 0) { + printf( "Memory allocation failed(shmget).\n"); + exit(1); + } + + address = shmat(shmid, NULL, SHM_RND); + + if ((BLASLONG)address == -1){ + printf( "Memory allocation failed(shmat).\n"); + exit(1); + } + + shmctl(shmid, IPC_RMID, 0); + + return address; +} + +#define malloc huge_malloc + +#endif + +int main(int argc, char *argv[]){ + + FLOAT *a,*c; + FLOAT alpha[] = {1.0, 1.0}; + blasint inc_x=1; + int loops = 1; + int l; + char *p; + + char uplo='U'; + + if ((p = getenv("OPENBLAS_UPLO"))) uplo=*p; + if ((p = getenv("OPENBLAS_LOOPS"))) loops = atoi(p); + if ((p = getenv("OPENBLAS_INCX"))) inc_x = atoi(p); + + blasint m, i, j; + + int from = 1; + int to = 200; + int step = 1; + + struct timeval start, stop; + double time1,timeg; + + argc--;argv++; + + if (argc > 0) { from = atol(*argv); argc--; argv++;} + if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++;} + if (argc > 0) { step = atol(*argv); argc--; argv++;} + + fprintf(stderr, "From : %3d To : %3d Step = %3d Uplo = %c Inc_x = %d\n", from, to, step,uplo,inc_x); + + + if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + + if (( c = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + +#ifdef linux + srandom(getpid()); +#endif + + fprintf(stderr, " SIZE Flops Time\n"); + + for(m = from; m <= to; m += step) + { + timeg=0; + + fprintf(stderr, " %6d : ", (int)m); + + for (l=0; l +#include +#ifdef __CYGWIN32__ +#include +#endif +#include "common.h" + + +#undef SPR2 + +#ifdef DOUBLE +#define SPR2 BLASFUNC(dspr2) +#else +#define SPR2 BLASFUNC(sspr2) +#endif + + + +#if defined(__WIN32__) || defined(__WIN64__) + +#ifndef DELTA_EPOCH_IN_MICROSECS +#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL +#endif + +int gettimeofday(struct timeval *tv, void *tz){ + + FILETIME ft; + unsigned __int64 tmpres = 0; + static int tzflag; + + if (NULL != tv) + { + GetSystemTimeAsFileTime(&ft); + + tmpres |= ft.dwHighDateTime; + tmpres <<= 32; + tmpres |= ft.dwLowDateTime; + + /*converting file time to unix epoch*/ + tmpres /= 10; /*convert into microseconds*/ + tmpres -= DELTA_EPOCH_IN_MICROSECS; + tv->tv_sec = (long)(tmpres / 1000000UL); + tv->tv_usec = (long)(tmpres % 1000000UL); + } + + return 0; +} + +#endif + +#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0 + +static void *huge_malloc(BLASLONG size){ + int shmid; + void *address; + +#ifndef SHM_HUGETLB +#define SHM_HUGETLB 04000 +#endif + + if ((shmid =shmget(IPC_PRIVATE, + (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1), + SHM_HUGETLB | IPC_CREAT |0600)) < 0) { + printf( "Memory allocation failed(shmget).\n"); + exit(1); + } + + address = shmat(shmid, NULL, SHM_RND); + + if ((BLASLONG)address == -1){ + printf( "Memory allocation failed(shmat).\n"); + exit(1); + } + + shmctl(shmid, IPC_RMID, 0); + + return address; +} + +#define malloc huge_malloc + +#endif + +int main(int argc, char *argv[]){ + + FLOAT *a,*b,*c; + FLOAT alpha[] = {1.0, 1.0}; + blasint inc_x=1,inc_y=1; + int loops = 1; + int l; + char *p; + + char uplo='U'; + + if ((p = getenv("OPENBLAS_UPLO"))) uplo=*p; + if ((p = getenv("OPENBLAS_LOOPS"))) loops = atoi(p); + if ((p = getenv("OPENBLAS_INCX"))) inc_x = atoi(p); + + blasint m, i, j; + + int from = 1; + int to = 200; + int step = 1; + + struct timeval start, stop; + double time1,timeg; + + argc--;argv++; + + if (argc > 0) { from = atol(*argv); argc--; argv++;} + if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++;} + if (argc > 0) { step = atol(*argv); argc--; argv++;} + + fprintf(stderr, "From : %3d To : %3d Step = %3d Uplo = %c Inc_x = %d Inc_y = %d\n", from, to, step,uplo,inc_x,inc_y); + + + if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + + if (( b = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + + if (( c = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + +#ifdef linux + srandom(getpid()); +#endif + + fprintf(stderr, " SIZE Flops Time\n"); + + for(m = from; m <= to; m += step) + { + timeg=0; + + fprintf(stderr, " %6d : ", (int)m); + + for (l=0; l +#include +#ifdef __CYGWIN32__ +#include +#endif +#include "common.h" + + +#undef SYR + +#ifdef DOUBLE +#define SYR BLASFUNC(dsyr) +#else +#define SYR BLASFUNC(ssyr) +#endif + + + +#if defined(__WIN32__) || defined(__WIN64__) + +#ifndef DELTA_EPOCH_IN_MICROSECS +#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL +#endif + +int gettimeofday(struct timeval *tv, void *tz){ + + FILETIME ft; + unsigned __int64 tmpres = 0; + static int tzflag; + + if (NULL != tv) + { + GetSystemTimeAsFileTime(&ft); + + tmpres |= ft.dwHighDateTime; + tmpres <<= 32; + tmpres |= ft.dwLowDateTime; + + /*converting file time to unix epoch*/ + tmpres /= 10; /*convert into microseconds*/ + tmpres -= DELTA_EPOCH_IN_MICROSECS; + tv->tv_sec = (long)(tmpres / 1000000UL); + tv->tv_usec = (long)(tmpres % 1000000UL); + } + + return 0; +} + +#endif + +#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0 + +static void *huge_malloc(BLASLONG size){ + int shmid; + void *address; + +#ifndef SHM_HUGETLB +#define SHM_HUGETLB 04000 +#endif + + if ((shmid =shmget(IPC_PRIVATE, + (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1), + SHM_HUGETLB | IPC_CREAT |0600)) < 0) { + printf( "Memory allocation failed(shmget).\n"); + exit(1); + } + + address = shmat(shmid, NULL, SHM_RND); + + if ((BLASLONG)address == -1){ + printf( "Memory allocation failed(shmat).\n"); + exit(1); + } + + shmctl(shmid, IPC_RMID, 0); + + return address; +} + +#define malloc huge_malloc + +#endif + +int main(int argc, char *argv[]){ + + FLOAT *x,*a; + FLOAT alpha[] = {1.0, 1.0}; + char *p; + + char uplo='U'; + + if ((p = getenv("OPENBLAS_UPLO"))) uplo=*p; + + blasint m, i, j; + blasint inc_x= 1; + int from = 1; + int to = 200; + int step = 1; + + struct timeval start, stop; + double time1; + + argc--;argv++; + + if (argc > 0) { from = atol(*argv); argc--; argv++;} + if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++;} + if (argc > 0) { step = atol(*argv); argc--; argv++;} + + fprintf(stderr, "From : %3d To : %3d Step = %3d Uplo = %c Inc_x = %d\n", from, to, step,uplo,inc_x); + + + if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + + if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + +#ifdef linux + srandom(getpid()); +#endif + + fprintf(stderr, " SIZE Flops\n"); + + for(m = from; m <= to; m += step) + { + + fprintf(stderr, " %6d : ", (int)m); + + for(i = 0; i < m * COMPSIZE * abs(inc_x); i++){ + x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + for(j = 0; j < m; j++){ + for(i = 0; i < m * COMPSIZE; i++){ + a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + } + + gettimeofday( &start, (struct timezone *)0); + + SYR (&uplo, &m, alpha, x, &inc_x, a, &m ); + + gettimeofday( &stop, (struct timezone *)0); + + time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; + + fprintf(stderr, + " %10.2f MFlops\n", + COMPSIZE * COMPSIZE * 1. * (double)m * (double)m / time1 * 1.e-6); + + } + + return 0; +} + +// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__"))); diff --git a/benchmark/syr2.c b/benchmark/syr2.c new file mode 100644 index 000000000..0129dd09a --- /dev/null +++ b/benchmark/syr2.c @@ -0,0 +1,194 @@ +/*************************************************************************** +Copyright (c) 2014, 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 +#include +#ifdef __CYGWIN32__ +#include +#endif +#include "common.h" + + +#undef SYR2 + + +#ifdef DOUBLE +#define SYR2 BLASFUNC(dsyr2) +#else +#define SYR2 BLASFUNC(ssyr2) +#endif + + +#if defined(__WIN32__) || defined(__WIN64__) + +#ifndef DELTA_EPOCH_IN_MICROSECS +#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL +#endif + +int gettimeofday(struct timeval *tv, void *tz){ + + FILETIME ft; + unsigned __int64 tmpres = 0; + static int tzflag; + + if (NULL != tv) + { + GetSystemTimeAsFileTime(&ft); + + tmpres |= ft.dwHighDateTime; + tmpres <<= 32; + tmpres |= ft.dwLowDateTime; + + /*converting file time to unix epoch*/ + tmpres /= 10; /*convert into microseconds*/ + tmpres -= DELTA_EPOCH_IN_MICROSECS; + tv->tv_sec = (long)(tmpres / 1000000UL); + tv->tv_usec = (long)(tmpres % 1000000UL); + } + + return 0; +} + +#endif + +#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0 + +static void *huge_malloc(BLASLONG size){ + int shmid; + void *address; + +#ifndef SHM_HUGETLB +#define SHM_HUGETLB 04000 +#endif + + if ((shmid =shmget(IPC_PRIVATE, + (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1), + SHM_HUGETLB | IPC_CREAT |0600)) < 0) { + printf( "Memory allocation failed(shmget).\n"); + exit(1); + } + + address = shmat(shmid, NULL, SHM_RND); + + if ((BLASLONG)address == -1){ + printf( "Memory allocation failed(shmat).\n"); + exit(1); + } + + shmctl(shmid, IPC_RMID, 0); + + return address; +} + +#define malloc huge_malloc + +#endif + +int main(int argc, char *argv[]){ + + FLOAT *x, *y, *a; + FLOAT alpha[] = {1.0, 1.0}; + char *p; + + char uplo='U'; + + if ((p = getenv("OPENBLAS_UPLO"))) uplo=*p; + + blasint m, i, j; + blasint inc_x= 1; + blasint inc_y= 1; + int from = 1; + int to = 200; + int step = 1; + + struct timeval start, stop; + double time1; + + argc--;argv++; + + if (argc > 0) { from = atol(*argv); argc--; argv++;} + if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++;} + if (argc > 0) { step = atol(*argv); argc--; argv++;} + + fprintf(stderr, "From : %3d To : %3d Step = %3d Uplo = %c Inc_x = %d Inc_y = %d\n", from, to, step,uplo,inc_x,inc_y); + + if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + + if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + + if (( y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + + + +#ifdef linux + srandom(getpid()); +#endif + + fprintf(stderr, " SIZE Flops\n"); + + for(m = from; m <= to; m += step) + { + + fprintf(stderr, " %6d : ", (int)m); + for(i = 0; i < m * COMPSIZE * abs(inc_x); i++){ + x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + for(i = 0; i < m * COMPSIZE * abs(inc_y); i++){ + y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + for(j = 0; j < m; j++){ + for(i = 0; i < m * COMPSIZE; i++){ + a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + } + + gettimeofday( &start, (struct timezone *)0); + + SYR2 (&uplo, &m, alpha, x, &inc_x, y, &inc_y, a, &m ); + + gettimeofday( &stop, (struct timezone *)0); + + time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; + + fprintf(stderr, + " %10.2f MFlops\n", + COMPSIZE * COMPSIZE * 2. * (double)m * (double)m / time1 * 1.e-6); + + } + + return 0; +} + +// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__"))); diff --git a/benchmark/syr2k.c b/benchmark/syr2k.c index 9840b5f3e..b1fcd8a18 100644 --- a/benchmark/syr2k.c +++ b/benchmark/syr2k.c @@ -175,9 +175,9 @@ int main(int argc, char *argv[]){ for(j = 0; j < m; j++){ for(i = 0; i < m * COMPSIZE; i++){ - a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; - b[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; - c[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + b[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + c[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; } } @@ -189,8 +189,6 @@ int main(int argc, char *argv[]){ time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; - gettimeofday( &start, (struct timezone *)0); - fprintf(stderr, " %10.2f MFlops\n", COMPSIZE * COMPSIZE * 2. * (double)m * (double)m * (double)m / time1 * 1.e-6); diff --git a/benchmark/syrk.c b/benchmark/syrk.c index 34817f2bb..95625a6c4 100644 --- a/benchmark/syrk.c +++ b/benchmark/syrk.c @@ -172,8 +172,8 @@ int main(int argc, char *argv[]){ for(j = 0; j < m; j++){ for(i = 0; i < m * COMPSIZE; i++){ - a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; - c[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + c[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; } } @@ -185,8 +185,6 @@ int main(int argc, char *argv[]){ time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; - gettimeofday( &start, (struct timezone *)0); - fprintf(stderr, " %10.2f MFlops\n", COMPSIZE * COMPSIZE * 1. * (double)m * (double)m * (double)m / time1 * 1.e-6); diff --git a/benchmark/trmm.c b/benchmark/trmm.c index 54c7972db..e095b85ee 100644 --- a/benchmark/trmm.c +++ b/benchmark/trmm.c @@ -175,8 +175,8 @@ int main(int argc, char *argv[]){ for(j = 0; j < m; j++){ for(i = 0; i < m * COMPSIZE; i++){ - a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; - b[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + b[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; } } @@ -188,8 +188,6 @@ int main(int argc, char *argv[]){ time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; - gettimeofday( &start, (struct timezone *)0); - fprintf(stderr, " %10.2f MFlops %10.6f sec\n", COMPSIZE * COMPSIZE * 1. * (double)m * (double)m * (double)m / time1 * 1.e-6, time1); diff --git a/benchmark/trmv.c b/benchmark/trmv.c new file mode 100644 index 000000000..f5a5fe31a --- /dev/null +++ b/benchmark/trmv.c @@ -0,0 +1,172 @@ +/*************************************************************************** +Copyright (c) 2014, 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 +#include +#ifdef __CYGWIN32__ +#include +#endif +#include "common.h" + +#undef TRMV + +#ifndef COMPLEX + +#ifdef DOUBLE +#define TRMV BLASFUNC(dtrmv) +#else +#define TRMV BLASFUNC(strmv) +#endif + +#else + +#ifdef DOUBLE +#define TRMV BLASFUNC(ztrmv) +#else +#define TRMV BLASFUNC(ctrmv) +#endif + +#endif + +#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0 + +static void *huge_malloc(BLASLONG size) +{ + int shmid; + void *address; + +#ifndef SHM_HUGETLB +#define SHM_HUGETLB 04000 +#endif + + if ((shmid =shmget(IPC_PRIVATE, + (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1), + SHM_HUGETLB | IPC_CREAT |0600)) < 0) { + printf( "Memory allocation failed(shmget).\n"); + exit(1); + } + + address = shmat(shmid, NULL, SHM_RND); + + if ((BLASLONG)address == -1) { + printf( "Memory allocation failed(shmat).\n"); + exit(1); + } + + shmctl(shmid, IPC_RMID, 0); + + return address; +} + +#define malloc huge_malloc + +#endif + +int main(int argc, char *argv[]) +{ + + FLOAT *a, *x; + char *p; + + char uplo ='U'; + char trans='N'; + char diag ='U'; + + int loops = 1; + int l; + blasint inc_x=1; + + if ((p = getenv("OPENBLAS_UPLO"))) uplo=*p; + if ((p = getenv("OPENBLAS_TRANS"))) trans=*p; + if ((p = getenv("OPENBLAS_DIAG"))) diag=*p; + if ((p = getenv("OPENBLAS_LOOPS"))) loops = atoi(p); + if ((p = getenv("OPENBLAS_INCX"))) inc_x = atoi(p); + + blasint n, i, j; + + int from = 1; + int to = 200; + int step = 1; + + struct timespec start = { 0, 0 }, stop = { 0, 0 }; + double time1, timeg; + + argc--;argv++; + + if (argc > 0) { from = atol(*argv); argc--; argv++;} + if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++;} + if (argc > 0) { step = atol(*argv); argc--; argv++;} + + fprintf(stderr, "From : %3d To : %3d Step = %3d Uplo = %c Trans = %c Diag = %c Loops=%d Inc_x=%d\n", from, + to, step, uplo, trans, diag, loops, inc_x); + + if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL) { + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + + if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + +#ifdef linux + srandom(getpid()); +#endif + + fprintf(stderr, " SIZE Flops\n"); + + for(n = from; n <= to; n += step) { + timeg=0; + + fprintf(stderr, " %6d : ", (int)n); + for(j = 0; j < n; j++) { + for(i = 0; i < n * COMPSIZE; i++) { + a[(long)i + (long)j * (long)n * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + } + + for (i = 0; i < n * COMPSIZE * abs(inc_x); i++) { + x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + for (l = 0; l < loops; l++) { + clock_gettime(CLOCK_REALTIME, &start); + TRMV (&uplo, &trans, &diag, &n, a, &n, x, &inc_x); + clock_gettime(CLOCK_REALTIME, &stop); + + time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_nsec - start.tv_nsec)) / 1.e9; + timeg += time1; + } + + timeg /= loops; + fprintf(stderr, " %10.2f MFlops %12.9f sec\n", + COMPSIZE * COMPSIZE * 1. * (double)n * (double)n / timeg / 1.e6, timeg); + } + + return 0; +} + +// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__"))); diff --git a/benchmark/trsm.c b/benchmark/trsm.c index 9eae3380c..6ce1d532c 100644 --- a/benchmark/trsm.c +++ b/benchmark/trsm.c @@ -191,8 +191,8 @@ int main(int argc, char *argv[]){ for(j = 0; j < m; j++){ for(i = 0; i < m * COMPSIZE; i++){ - a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; - b[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + b[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; } } diff --git a/benchmark/trsv.c b/benchmark/trsv.c new file mode 100644 index 000000000..c60890de4 --- /dev/null +++ b/benchmark/trsv.c @@ -0,0 +1,222 @@ +/*************************************************************************** +Copyright (c) 2014, 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 +#include +#ifdef __CYGWIN32__ +#include +#endif +#include +#include "common.h" + + +#undef GEMV +#undef TRSV + +#ifndef COMPLEX + +#ifdef DOUBLE +#define TRSV BLASFUNC(dtrsv) +#else +#define TRSV BLASFUNC(strsv) +#endif + +#else + +#ifdef DOUBLE +#define TRSV BLASFUNC(ztrsv) +#else +#define TRSV BLASFUNC(ctrsv) +#endif + +#endif + +#if defined(__WIN32__) || defined(__WIN64__) + +#ifndef DELTA_EPOCH_IN_MICROSECS +#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL +#endif + +int gettimeofday(struct timeval *tv, void *tz){ + + FILETIME ft; + unsigned __int64 tmpres = 0; + static int tzflag; + + if (NULL != tv) + { + GetSystemTimeAsFileTime(&ft); + + tmpres |= ft.dwHighDateTime; + tmpres <<= 32; + tmpres |= ft.dwLowDateTime; + + /*converting file time to unix epoch*/ + tmpres /= 10; /*convert into microseconds*/ + tmpres -= DELTA_EPOCH_IN_MICROSECS; + tv->tv_sec = (long)(tmpres / 1000000UL); + tv->tv_usec = (long)(tmpres % 1000000UL); + } + + return 0; +} + +#endif + +#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0 + +static void *huge_malloc(BLASLONG size){ + int shmid; + void *address; + +#ifndef SHM_HUGETLB +#define SHM_HUGETLB 04000 +#endif + + if ((shmid =shmget(IPC_PRIVATE, + (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1), + SHM_HUGETLB | IPC_CREAT |0600)) < 0) { + printf( "Memory allocation failed(shmget).\n"); + exit(1); + } + + address = shmat(shmid, NULL, SHM_RND); + + if ((BLASLONG)address == -1){ + printf( "Memory allocation failed(shmat).\n"); + exit(1); + } + + shmctl(shmid, IPC_RMID, 0); + + return address; +} + +#define malloc huge_malloc + +#endif + +int main(int argc, char *argv[]){ + + FLOAT *a, *x; + blasint n = 0, i, j; + blasint inc_x=1; + int loops = 1; + int l; + char *p; + + int from = 1; + int to = 200; + int step = 1; + + struct timespec time_start, time_end; + time_t seconds = 0; + + double time1,timeg; + long long nanos = 0; + + argc--;argv++; + + if (argc > 0) { from = atol(*argv); argc--; argv++;} + if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++;} + if (argc > 0) { step = atol(*argv); argc--; argv++;} + + char uplo ='L'; + char transa = 'N'; + char diag ='U'; + + if ((p = getenv("OPENBLAS_LOOPS"))) loops = atoi(p); + if ((p = getenv("OPENBLAS_INCX"))) inc_x = atoi(p); + if ((p = getenv("OPENBLAS_TRANSA"))) transa=*p; + if ((p = getenv("OPENBLAS_DIAG"))) diag=*p; + if ((p = getenv("OPENBLAS_UPLO"))) uplo=*p; + + fprintf(stderr, "From : %3d To : %3d Step = %3d Transa = '%c' Inc_x = %d uplo=%c diag=%c loop = %d\n", from, to, step,transa,inc_x, + uplo,diag,loops); + + +#ifdef linux + srandom(getpid()); +#endif + + fprintf(stderr, " SIZE Flops\n"); + fprintf(stderr, "============================================\n"); + + for(n = from; n <= to; n += step) + { + timeg=0; + if (( a = (FLOAT *)malloc(sizeof(FLOAT) * n * n * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + + if (( x = (FLOAT *)malloc(sizeof(FLOAT) * n * abs(inc_x) * COMPSIZE)) == NULL){ + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + + for(j = 0; j < n; j++){ + for(i = 0; i < n * COMPSIZE; i++){ + a[i + j * n * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + } + + for(i = 0; i < n * COMPSIZE * abs(inc_x); i++){ + x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + for(l =0;l< loops;l++){ + + clock_gettime(CLOCK_PROCESS_CPUTIME_ID,&time_start); + + TRSV(&uplo,&transa,&diag,&n,a,&n,x,&inc_x); + + clock_gettime(CLOCK_PROCESS_CPUTIME_ID,&time_end); + nanos = time_end.tv_nsec - time_start.tv_nsec; + seconds = time_end.tv_sec - time_start.tv_sec; + + time1 = seconds + nanos /1.e9; + timeg += time1; + } + + + timeg /= loops; + long long muls = n*(n+1)/2.0; + long long adds = (n - 1.0)*n/2.0; + + fprintf(stderr, "%10d %10.2f MFlops %10.6f sec\n", n,(muls+adds) / timeg * 1.e-6, timeg); + if(a != NULL){ + free(a); + } + + if( x != NULL){ + free(x); + } + + } + + return 0; +} + diff --git a/c_check b/c_check index 3d82aa73c..c7899c84f 100644 --- a/c_check +++ b/c_check @@ -18,11 +18,12 @@ $binary = $ENV{"BINARY"}; $makefile = shift(@ARGV); $config = shift(@ARGV); -$compiler_name = join(" ", @ARGV); +$compiler_name = shift(@ARGV); +$flags = join(" ", @ARGV); # First, we need to know the target OS and compiler name -$data = `$compiler_name -E ctest.c`; +$data = `$compiler_name $flags -E ctest.c`; if ($?) { printf STDERR "C Compiler ($compiler_name) is something wrong.\n"; @@ -175,7 +176,7 @@ if ($defined == 0) { # Do again -$data = `$compiler_name -E ctest.c`; +$data = `$compiler_name $flags -E ctest.c`; if ($?) { printf STDERR "C Compiler ($compiler_name) is something wrong.\n"; @@ -188,14 +189,14 @@ if (($architecture eq "mips") || ($architecture eq "mips64")) { if ($@){ warn "could not load PERL module File::Temp, so could not check MSA capatibility"; } else { - $tmpf = new File::Temp( UNLINK => 1 ); + $tmpf = new File::Temp( SUFFIX => '.c' , UNLINK => 1 ); $code = '"addvi.b $w0, $w1, 1"'; $msa_flags = "-mmsa -mfp64 -msched-weight -mload-store-pairs"; print $tmpf "#include \n\n"; print $tmpf "void main(void){ __asm__ volatile($code); }\n"; - $args = "$msa_flags -o $tmpf.o -x c $tmpf"; - my @cmd = ("$compiler_name $args"); + $args = "$msa_flags -o $tmpf.o $tmpf"; + my @cmd = ("$compiler_name $flags $args >/dev/null 2>/dev/null"); system(@cmd) == 0; if ($? != 0) { $have_msa = 0; @@ -229,11 +230,14 @@ if (($architecture eq "x86") || ($architecture eq "x86_64")) { $no_avx512 = 0; } else { # $tmpf = new File::Temp( UNLINK => 1 ); - ($fh,$tmpf) = tempfile( UNLINK => 1 ); + ($fh,$tmpf) = tempfile( SUFFIX => '.c' , UNLINK => 1 ); $code = '"vbroadcastss -4 * 4(%rsi), %zmm2"'; print $tmpf "#include \n\nint main(void){ __asm__ volatile($code); }\n"; - $args = " -march=skylake-avx512 -c -o $tmpf.o -x c $tmpf"; - my @cmd = ("$compiler_name $args >/dev/null 2>/dev/null"); + $args = " -march=skylake-avx512 -c -o $tmpf.o $tmpf"; + if ($compiler eq "PGI") { + $args = " -tp skylake -c -o $tmpf.o $tmpf"; + } + my @cmd = ("$compiler_name $flags $args >/dev/null 2>/dev/null"); system(@cmd) == 0; if ($? != 0) { $no_avx512 = 1; @@ -244,7 +248,7 @@ if (($architecture eq "x86") || ($architecture eq "x86_64")) { } } -$data = `$compiler_name -S ctest1.c && grep globl ctest1.s | head -n 1 && rm -f ctest1.s`; +$data = `$compiler_name $flags -S ctest1.c && grep globl ctest1.s | head -n 1 && rm -f ctest1.s`; $data =~ /globl\s([_\.]*)(.*)/; @@ -260,19 +264,6 @@ if ($architecture ne $hostarch) { $cross = 1 if ($os ne $hostos); -# rework cross suffix and architecture if we are on OSX cross-compiling for ARMV8-based IOS -# the initial autodetection will have been confused by the command-line arguments to clang -# and the cross-compiler apparently still claims to build for x86_64 in its CC -E output -if (($os eq "Darwin") && ($cross_suffix ne "")) { - my $tmpnam = `xcrun --sdk iphoneos --find clang`; - $cross_suffix = substr($tmpnam, 0, rindex($tmpnam, "/")+1 ); -# this should produce something like $cross_suffix="/Applications/Xcode-10.1.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/"; - $cross =1; - $architecture = arm64; -} - - - $openmp = "" if $ENV{USE_OPENMP} != 1; $linker_L = ""; @@ -280,7 +271,7 @@ $linker_l = ""; $linker_a = ""; { - $link = `$compiler_name -c ctest2.c -o ctest2.o 2>&1 && $compiler_name $openmp -v ctest2.o -o ctest2 2>&1 && rm -f ctest2.o ctest2 ctest2.exe`; + $link = `$compiler_name $flags -c ctest2.c -o ctest2.o 2>&1 && $compiler_name $flags $openmp -v ctest2.o -o ctest2 2>&1 && rm -f ctest2.o ctest2 ctest2.exe`; $link =~ s/\-Y\sP\,/\-Y/g; @@ -318,6 +309,7 @@ $linker_a = ""; && ($flags !~ /kernel32/) && ($flags !~ /advapi32/) && ($flags !~ /shell32/) + && ($flags !~ /omp/) ) { $linker_l .= $flags . " " } diff --git a/cmake/arch.cmake b/cmake/arch.cmake index 8280d6274..9d51f777c 100644 --- a/cmake/arch.cmake +++ b/cmake/arch.cmake @@ -45,7 +45,7 @@ endif () if (DYNAMIC_ARCH) if (ARM64) - set(DYNAMIC_CORE ARMV8 CORTEXA53 CORTEXA57 CORTEXA72 CORTEXA73 FALKOR THUNDERX THUNDERX2T99 TSV110) + set(DYNAMIC_CORE ARMV8 CORTEXA53 CORTEXA57 CORTEXA72 CORTEXA73 FALKOR THUNDERX THUNDERX2T99 TSV110 EMAG8180 NEOVERSEN1) endif () if (POWER) diff --git a/cmake/cc.cmake b/cmake/cc.cmake index 37da0d6ed..d5551147c 100644 --- a/cmake/cc.cmake +++ b/cmake/cc.cmake @@ -96,3 +96,10 @@ if (${CMAKE_C_COMPILER_ID} STREQUAL "SUN") endif () endif () +if (${CORE} STREQUAL "SKYLAKEX") + if (NOT DYNAMIC_ARCH) + if (NOT NO_AVX512) + set (CCOMMON_OPT "${CCOMMON_OPT} -march=skylake-avx512") + endif () + endif () +endif () diff --git a/cmake/prebuild.cmake b/cmake/prebuild.cmake index c6d109356..44e1473d1 100644 --- a/cmake/prebuild.cmake +++ b/cmake/prebuild.cmake @@ -229,6 +229,33 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS set(ZGEMM_UNROLL_M 4) set(ZGEMM_UNROLL_N 4) set(SYMV_P 16) + elseif ("${TCORE}" STREQUAL "NEOVERSEN1") + file(APPEND ${TARGET_CONF_TEMP} + "#define L1_CODE_SIZE\t65536\n" + "#define L1_CODE_LINESIZE\t64\n" + "#define L1_CODE_ASSOCIATIVE\t4\n" + "#define L1_DATA_SIZE\t65536\n" + "#define L1_DATA_LINESIZE\t64\n" + "#define L1_DATA_ASSOCIATIVE\t2\n" + "#define L2_SIZE\t1048576\n\n" + "#define L2_LINESIZE\t64\n" + "#define L2_ASSOCIATIVE\t16\n" + "#define DTB_DEFAULT_ENTRIES\t64\n" + "#define DTB_SIZE\t4096\n" + "#define HAVE_VFPV4\n" + "#define HAVE_VFPV3\n" + "#define HAVE_VFP\n" + "#define HAVE_NEON\n" + "#define ARMV8\n") + set(SGEMM_UNROLL_M 16) + set(SGEMM_UNROLL_N 4) + set(DGEMM_UNROLL_M 8) + set(DGEMM_UNROLL_N 4) + set(CGEMM_UNROLL_M 8) + set(CGEMM_UNROLL_N 4) + set(ZGEMM_UNROLL_M 4) + set(ZGEMM_UNROLL_N 4) + set(SYMV_P 16) elseif ("${TCORE}" STREQUAL "FALKOR") file(APPEND ${TARGET_CONF_TEMP} "#define L1_CODE_SIZE\t65536\n" @@ -332,6 +359,29 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS set(ZGEMM_UNROLL_M 4) set(ZGEMM_UNROLL_N 4) set(SYMV_P 16) + elseif ("${TCORE}" STREQUAL "EMAG8180") + file(APPEND ${TARGET_CONF_TEMP} + "#define ARMV8\n" + "#define L1_CODE_SIZE\t32768\n" + "#define L1_CODE_LINESIZE\t64\n" + "#define L1_CODE_ASSOCIATIVE\t4\n" + "#define L1_DATA_SIZE\t32768\n" + "#define L1_DATA_LINESIZE\t64\n" + "#define L1_DATA_ASSOCIATIVE\t4\n" + "#define L2_SIZE\t5262144\n" + "#define L2_LINESIZE\t64\n" + "#define L2_ASSOCIATIVE\t8\n" + "#define DTB_DEFAULT_ENTRIES\t64\n" + "#define DTB_SIZE\t4096\n") + set(SGEMM_UNROLL_M 16) + set(SGEMM_UNROLL_N 4) + set(DGEMM_UNROLL_M 8) + set(DGEMM_UNROLL_N 4) + set(CGEMM_UNROLL_M 8) + set(CGEMM_UNROLL_N 4) + set(ZGEMM_UNROLL_M 4) + set(ZGEMM_UNROLL_N 4) + set(SYMV_P 16) elseif ("${TCORE}" STREQUAL "POWER6") file(APPEND ${TARGET_CONF_TEMP} "#define L1_DATA_SIZE 32768\n" diff --git a/cmake/system.cmake b/cmake/system.cmake index 4f8011603..ce980a7b9 100644 --- a/cmake/system.cmake +++ b/cmake/system.cmake @@ -289,6 +289,10 @@ set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_CPU_NUMBER=${NUM_THREADS}") set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_PARALLEL_NUMBER=${NUM_PARALLEL}") +if (BUFFERSIZE) +set(CCOMMON_OPT "${CCOMMON_OPT} -DBUFFERSIZE=${BUFFERSIZE}") +endif () + if (USE_SIMPLE_THREADED_LEVEL3) set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_SIMPLE_THREADED_LEVEL3") endif () diff --git a/common_arm64.h b/common_arm64.h index 5951e1ee5..a928dbe7b 100644 --- a/common_arm64.h +++ b/common_arm64.h @@ -53,16 +53,16 @@ static void __inline blas_lock(volatile BLASULONG *address){ BLASULONG ret; do { - while (*address) {YIELDING;}; - __asm__ __volatile__( "mov x4, #1 \n\t" + "sevl \n\t" "1: \n\t" + "wfe \n\t" + "2: \n\t" "ldaxr x2, [%1] \n\t" "cbnz x2, 1b \n\t" - "2: \n\t" "stxr w3, x4, [%1] \n\t" - "cbnz w3, 1b \n\t" + "cbnz w3, 2b \n\t" "mov %0, #0 \n\t" : "=r"(ret), "=r"(address) : "1"(address) @@ -81,10 +81,12 @@ static void __inline blas_lock(volatile BLASULONG *address){ #if !defined(OS_DARWIN) && !defined (OS_ANDROID) static __inline BLASULONG rpcc(void){ BLASULONG ret = 0; + blasint shift; __asm__ __volatile__ ("isb; mrs %0,cntvct_el0":"=r"(ret)); + __asm__ __volatile__ ("mrs %0,cntfrq_el0; clz %w0, %w0":"=&r"(shift)); - return ret; + return ret << shift; } #define RPCC_DEFINED diff --git a/common_x86_64.h b/common_x86_64.h index c05998d58..fe5539abe 100644 --- a/common_x86_64.h +++ b/common_x86_64.h @@ -225,7 +225,11 @@ static __inline int blas_quickdivide(unsigned int x, unsigned int y){ #endif #define HUGE_PAGESIZE ( 2 << 20) +#ifndef BUFFERSIZE #define BUFFER_SIZE (32 << 20) +#else +#define BUFFER_SIZE (32 << BUFFERSIZE) +#endif #define SEEK_ADDRESS diff --git a/cpuid_arm64.c b/cpuid_arm64.c index 9e019fe3e..4103216e6 100644 --- a/cpuid_arm64.c +++ b/cpuid_arm64.c @@ -34,6 +34,7 @@ #define CPU_CORTEXA57 3 #define CPU_CORTEXA72 4 #define CPU_CORTEXA73 5 +#define CPU_NEOVERSEN1 11 // Qualcomm #define CPU_FALKOR 6 // Cavium @@ -41,6 +42,8 @@ #define CPU_THUNDERX2T99 8 //Hisilicon #define CPU_TSV110 9 +// Ampere +#define CPU_EMAG8180 10 static char *cpuname[] = { "UNKNOWN", @@ -52,7 +55,9 @@ static char *cpuname[] = { "FALKOR", "THUNDERX", "THUNDERX2T99", - "TSV110" + "TSV110", + "EMAG8180", + "NEOVERSEN1" }; static char *cpuname_lower[] = { @@ -65,7 +70,9 @@ static char *cpuname_lower[] = { "falkor", "thunderx", "thunderx2t99", - "tsv110" + "tsv110", + "emag8180", + "neoversen1" }; int get_feature(char *search) @@ -140,6 +147,8 @@ int detect(void) return CPU_CORTEXA72; else if (strstr(cpu_part, "0xd09")) return CPU_CORTEXA73; + else if (strstr(cpu_part, "0xd0c")) + return CPU_NEOVERSEN1; } // Qualcomm else if (strstr(cpu_implementer, "0x51") && strstr(cpu_part, "0xc00")) @@ -152,6 +161,9 @@ int detect(void) // HiSilicon else if (strstr(cpu_implementer, "0x48") && strstr(cpu_part, "0xd01")) return CPU_TSV110; + // Ampere + else if (strstr(cpu_implementer, "0x50") && strstr(cpu_part, "0x000")) + return CPU_EMAG8180; } p = (char *) NULL ; @@ -278,6 +290,20 @@ void get_cpuconfig(void) printf("#define DTB_DEFAULT_ENTRIES 64\n"); printf("#define DTB_SIZE 4096\n"); break; + case CPU_NEOVERSEN1: + printf("#define %s\n", cpuname[d]); + printf("#define L1_CODE_SIZE 65536\n"); + printf("#define L1_CODE_LINESIZE 64\n"); + printf("#define L1_CODE_ASSOCIATIVE 4\n"); + printf("#define L1_DATA_SIZE 65536\n"); + printf("#define L1_DATA_LINESIZE 64\n"); + printf("#define L1_DATA_ASSOCIATIVE 4\n"); + printf("#define L2_SIZE 1048576\n"); + printf("#define L2_LINESIZE 64\n"); + printf("#define L2_ASSOCIATIVE 16\n"); + printf("#define DTB_DEFAULT_ENTRIES 64\n"); + printf("#define DTB_SIZE 4096\n"); + break; case CPU_FALKOR: printf("#define FALKOR\n"); @@ -335,6 +361,18 @@ void get_cpuconfig(void) printf("#define DTB_DEFAULT_ENTRIES 64 \n"); printf("#define DTB_SIZE 4096 \n"); break; + + case CPU_EMAG8180: + // Minimum parameters for ARMv8 (based on A53) + printf("#define EMAG8180\n"); + printf("#define L1_CODE_SIZE 32768\n"); + printf("#define L1_DATA_SIZE 32768\n"); + printf("#define L1_DATA_LINESIZE 64\n"); + printf("#define L2_SIZE 262144\n"); + printf("#define L2_LINESIZE 64\n"); + printf("#define DTB_DEFAULT_ENTRIES 64\n"); + printf("#define DTB_SIZE 4096\n"); + } get_cpucount(); } diff --git a/cpuid_x86.c b/cpuid_x86.c index 9e1c8e752..e29adecae 100644 --- a/cpuid_x86.c +++ b/cpuid_x86.c @@ -2006,6 +2006,38 @@ int get_coretype(void){ return CORE_NEHALEM; } break; + case 6: + if (model == 6) +#ifndef NO_AVX512 + return CORE_SKYLAKEX; +#else + if(support_avx()) +#ifndef NO_AVX2 + return CORE_HASWELL; +#else + return CORE_SANDYBRIDGE; +#endif + else + return CORE_NEHALEM; +#endif + break; + case 7: + if (model == 10) + return CORE_NEHALEM; + if (model == 14) +#ifndef NO_AVX512 + return CORE_SKYLAKEX; +#else + if(support_avx()) +#ifndef NO_AVX2 + return CORE_HASWELL; +#else + return CORE_SANDYBRIDGE; +#endif + else + return CORE_NEHALEM; +#endif + break; case 9: case 8: if (model == 14) { // Kaby Lake diff --git a/driver/level3/level3.c b/driver/level3/level3.c index 1ab7a740e..9aa67286f 100644 --- a/driver/level3/level3.c +++ b/driver/level3/level3.c @@ -332,13 +332,16 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, #else for(jjs = js; jjs < js + min_j; jjs += min_jj){ min_jj = min_j + js - jjs; - +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj >= 3*GEMM_UNROLL_N) min_jj = 3*GEMM_UNROLL_N; else if (min_jj >= 2*GEMM_UNROLL_N) min_jj = 2*GEMM_UNROLL_N; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif START_RPCC(); diff --git a/driver/level3/level3_gemm3m_thread.c b/driver/level3/level3_gemm3m_thread.c index 21d431b60..9216daaed 100644 --- a/driver/level3/level3_gemm3m_thread.c +++ b/driver/level3/level3_gemm3m_thread.c @@ -104,7 +104,7 @@ typedef struct { #define BETA_OPERATION(M_FROM, M_TO, N_FROM, N_TO, BETA, C, LDC) \ GEMM_BETA((M_TO) - (M_FROM), (N_TO - N_FROM), 0, \ BETA[0], BETA[1], NULL, 0, NULL, 0, \ - (FLOAT *)(C) + (M_FROM) + (N_FROM) * (LDC) * COMPSIZE, LDC) + (FLOAT *)(C) + ((M_FROM) + (N_FROM) * (LDC)) * COMPSIZE, LDC) #endif #ifndef ICOPYB_OPERATION @@ -414,7 +414,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, for(jjs = xxx; jjs < MIN(n_to, xxx + div_n); jjs += min_jj){ min_jj = MIN(n_to, xxx + div_n) - jjs; - if (min_jj > GEMM3M_UNROLL_N) min_jj = GEMM3M_UNROLL_N; + if (min_jj > GEMM3M_UNROLL_N*3) min_jj = GEMM3M_UNROLL_N*3; START_RPCC(); @@ -550,7 +550,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, for(jjs = xxx; jjs < MIN(n_to, xxx + div_n); jjs += min_jj){ min_jj = MIN(n_to, xxx + div_n) - jjs; - if (min_jj > GEMM3M_UNROLL_N) min_jj = GEMM3M_UNROLL_N; + if (min_jj > GEMM3M_UNROLL_N*3) min_jj = GEMM3M_UNROLL_N*3; START_RPCC(); @@ -687,7 +687,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, for(jjs = xxx; jjs < MIN(n_to, xxx + div_n); jjs += min_jj){ min_jj = MIN(n_to, xxx + div_n) - jjs; - if (min_jj > GEMM3M_UNROLL_N) min_jj = GEMM3M_UNROLL_N; + if (min_jj > GEMM3M_UNROLL_N*3) min_jj = GEMM3M_UNROLL_N*3; START_RPCC(); diff --git a/driver/level3/level3_thread.c b/driver/level3/level3_thread.c index cfbff7554..ca0085e71 100644 --- a/driver/level3/level3_thread.c +++ b/driver/level3/level3_thread.c @@ -351,8 +351,9 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, /* Make sure if no one is using workspace */ START_RPCC(); for (i = 0; i < args -> nthreads; i++) - while (job[mypos].working[i][CACHE_LINE_SIZE * bufferside]) {YIELDING;MB;}; + while (job[mypos].working[i][CACHE_LINE_SIZE * bufferside]) {YIELDING;}; STOP_RPCC(waiting1); + MB; #if defined(FUSED_GEMM) && !defined(TIMING) @@ -365,12 +366,16 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, /* Split local region of B into parts */ for(jjs = js; jjs < MIN(n_to, js + div_n); jjs += min_jj){ min_jj = MIN(n_to, js + div_n) - jjs; +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj >= 3*GEMM_UNROLL_N) min_jj = 3*GEMM_UNROLL_N; else if (min_jj >= 2*GEMM_UNROLL_N) min_jj = 2*GEMM_UNROLL_N; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif /* Copy part of local region of B into workspace */ START_RPCC(); OCOPY_OPERATION(min_l, min_jj, b, ldb, ls, jjs, @@ -391,10 +396,10 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, } #endif + WMB; /* Set flag so other threads can access local region of B */ for (i = mypos_n * nthreads_m; i < (mypos_n + 1) * nthreads_m; i++) job[mypos].working[i][CACHE_LINE_SIZE * bufferside] = (BLASLONG)buffer[bufferside]; - WMB; } /* Get regions of B from other threads and apply kernel */ @@ -413,8 +418,9 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, /* Wait until other region of B is initialized */ START_RPCC(); - while(job[current].working[mypos][CACHE_LINE_SIZE * bufferside] == 0) {YIELDING;MB;}; + while(job[current].working[mypos][CACHE_LINE_SIZE * bufferside] == 0) {YIELDING;}; STOP_RPCC(waiting2); + MB; /* Apply kernel with local region of A and part of other region of B */ START_RPCC(); @@ -430,8 +436,8 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, /* Clear synchronization flag if this thread is done with other region of B */ if (m_to - m_from == min_i) { - job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0; WMB; + job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0; } } } while (current != mypos); @@ -473,8 +479,8 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, /* Clear synchronization flag if this thread is done with region of B */ if (is + min_i >= m_to) { - job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0; WMB; + job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0; } } @@ -493,10 +499,11 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, START_RPCC(); for (i = 0; i < args -> nthreads; i++) { for (js = 0; js < DIVIDE_RATE; js++) { - while (job[mypos].working[i][CACHE_LINE_SIZE * js] ) {YIELDING;MB;}; + while (job[mypos].working[i][CACHE_LINE_SIZE * js] ) {YIELDING;}; } } STOP_RPCC(waiting3); + MB; #ifdef TIMING BLASLONG waiting = waiting1 + waiting2 + waiting3; @@ -701,7 +708,7 @@ EnterCriticalSection((PCRITICAL_SECTION)&level3_lock); } } } - + WMB; /* Execute parallel computation */ exec_blas(nthreads, queue); } diff --git a/driver/level3/trmm_L.c b/driver/level3/trmm_L.c index 8a81d31a0..9117090b5 100644 --- a/driver/level3/trmm_L.c +++ b/driver/level3/trmm_L.c @@ -135,10 +135,14 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO for(jjs = js; jjs < js + min_j; jjs += min_jj){ min_jj = min_j + js - jjs; +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif START_RPCC(); GEMM_ONCOPY(min_l, min_jj, b + (jjs * ldb) * COMPSIZE, ldb, sb + min_l * (jjs - js) * COMPSIZE); @@ -201,10 +205,14 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO for(jjs = js; jjs < js + min_j; jjs += min_jj){ min_jj = min_j + js - jjs; +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif START_RPCC(); GEMM_ONCOPY(min_l, min_jj, b + (ls + jjs * ldb) * COMPSIZE, ldb, sb + min_l * (jjs - js) * COMPSIZE); @@ -292,10 +300,14 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO for(jjs = js; jjs < js + min_j; jjs += min_jj){ min_jj = min_j + js - jjs; +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif START_RPCC(); GEMM_ONCOPY(min_l, min_jj, b + (m - min_l + jjs * ldb) * COMPSIZE, ldb, @@ -358,10 +370,14 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO for(jjs = js; jjs < js + min_j; jjs += min_jj){ min_jj = min_j + js - jjs; +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif START_RPCC(); GEMM_ONCOPY(min_l, min_jj, b + (ls - min_l + jjs * ldb) * COMPSIZE, ldb, diff --git a/driver/level3/trmm_R.c b/driver/level3/trmm_R.c index 0882aa496..62c6a2442 100644 --- a/driver/level3/trmm_R.c +++ b/driver/level3/trmm_R.c @@ -122,10 +122,14 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO for(jjs = 0; jjs < ls - js; jjs += min_jj){ min_jj = ls - js - jjs; +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif #ifndef TRANSA GEMM_ONCOPY(min_l, min_jj, a + (ls + (js + jjs) * lda) * COMPSIZE, lda, sb + min_l * jjs * COMPSIZE); #else @@ -142,10 +146,14 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO for(jjs = 0; jjs < min_l; jjs += min_jj){ min_jj = min_l - jjs; +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif #ifndef TRANSA TRMM_OLNCOPY(min_l, min_jj, a, lda, ls, ls + jjs, sb + min_l * (ls - js + jjs) * COMPSIZE); #else @@ -195,10 +203,14 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO for(jjs = js; jjs < js + min_j; jjs += min_jj){ min_jj = min_j + js - jjs; +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif #ifndef TRANSA GEMM_ONCOPY(min_l, min_jj, a + (ls + jjs * lda) * COMPSIZE, lda, sb + min_l * (jjs - js) * COMPSIZE); #else @@ -246,10 +258,14 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO for(jjs = 0; jjs < min_l; jjs += min_jj){ min_jj = min_l - jjs; +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif #ifndef TRANSA TRMM_OUNCOPY(min_l, min_jj, a, lda, ls, ls + jjs, sb + min_l * jjs * COMPSIZE); #else @@ -267,10 +283,14 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO for(jjs = 0; jjs < js - ls - min_l; jjs += min_jj){ min_jj = js - ls - min_l - jjs; +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif #ifndef TRANSA GEMM_ONCOPY(min_l, min_jj, a + (ls + (ls + min_l + jjs) * lda) * COMPSIZE, lda, sb + min_l * (min_l + jjs) * COMPSIZE); @@ -324,10 +344,14 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO for(jjs = js; jjs < js + min_j; jjs += min_jj){ min_jj = min_j + js - jjs; +#ifdef SKYLAKEX + /* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */ + if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N; +#else if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3; else if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N; - +#endif #ifndef TRANSA GEMM_ONCOPY(min_l, min_jj, a + (ls + (jjs - min_j) * lda) * COMPSIZE, lda, sb + min_l * (jjs - js) * COMPSIZE); #else diff --git a/driver/others/blas_server.c b/driver/others/blas_server.c index 6f4e20610..aa0644845 100644 --- a/driver/others/blas_server.c +++ b/driver/others/blas_server.c @@ -140,6 +140,16 @@ typedef struct { } thread_status_t; +#if (__STDC_VERSION__ >= 201112L) +#define atomic_load_queue(p) __atomic_load_n(p, __ATOMIC_RELAXED) +#define atomic_store_queue(p, v) __atomic_store_n(p, v, __ATOMIC_RELAXED) +#else +#define atomic_load_queue(p) (blas_queue_t*)(*(volatile blas_queue_t**)(p)) +#define atomic_store_queue(p, v) (*(volatile blas_queue_t* volatile*)(p) = (v)) +#endif + + + static thread_status_t thread_status[MAX_CPU_NUMBER] __attribute__((aligned(ATTRIBUTE_SIZE))); #ifndef THREAD_TIMEOUT @@ -312,20 +322,19 @@ blas_queue_t *tscq; last_tick = (unsigned int)rpcc(); - pthread_mutex_lock (&thread_status[cpu].lock); - tscq=thread_status[cpu].queue; - pthread_mutex_unlock (&thread_status[cpu].lock); + tscq = atomic_load_queue(&thread_status[cpu].queue); while(!tscq) { YIELDING; if ((unsigned int)rpcc() - last_tick > thread_timeout) { - pthread_mutex_lock (&thread_status[cpu].lock); - if (!thread_status[cpu].queue) { + if (!atomic_load_queue(&thread_status[cpu].queue)) { + pthread_mutex_lock (&thread_status[cpu].lock); thread_status[cpu].status = THREAD_STATUS_SLEEP; - while (thread_status[cpu].status == THREAD_STATUS_SLEEP) { + while (thread_status[cpu].status == THREAD_STATUS_SLEEP && + !atomic_load_queue(&thread_status[cpu].queue)) { #ifdef MONITOR main_status[cpu] = MAIN_SLEEPING; @@ -333,19 +342,18 @@ blas_queue_t *tscq; pthread_cond_wait(&thread_status[cpu].wakeup, &thread_status[cpu].lock); } + pthread_mutex_unlock(&thread_status[cpu].lock); } - pthread_mutex_unlock(&thread_status[cpu].lock); - last_tick = (unsigned int)rpcc(); } - pthread_mutex_lock (&thread_status[cpu].lock); - tscq=thread_status[cpu].queue; - pthread_mutex_unlock (&thread_status[cpu].lock); + + tscq = atomic_load_queue(&thread_status[cpu].queue); } - queue = thread_status[cpu].queue; + queue = atomic_load_queue(&thread_status[cpu].queue); + MB; if ((long)queue == -1) break; @@ -360,9 +368,7 @@ blas_queue_t *tscq; if (queue) { int (*routine)(blas_arg_t *, void *, void *, void *, void *, BLASLONG) = queue -> routine; - pthread_mutex_lock (&thread_status[cpu].lock); - thread_status[cpu].queue = (blas_queue_t *)1; - pthread_mutex_unlock (&thread_status[cpu].lock); + atomic_store_queue(&thread_status[cpu].queue, (blas_queue_t *)1); sa = queue -> sa; sb = queue -> sb; @@ -442,13 +448,9 @@ blas_queue_t *tscq; // arm: make sure all results are written out _before_ // thread is marked as done and other threads use them - WMB; + MB; + atomic_store_queue(&thread_status[cpu].queue, (blas_queue_t *)0); - pthread_mutex_lock (&thread_status[cpu].lock); - thread_status[cpu].queue = (blas_queue_t * volatile) ((long)thread_status[cpu].queue & 0); /* Need a trick */ - pthread_mutex_unlock (&thread_status[cpu].lock); - - WMB; } @@ -566,7 +568,7 @@ int blas_thread_init(void){ for(i = 0; i < blas_num_threads - 1; i++){ - thread_status[i].queue = (blas_queue_t *)NULL; + atomic_store_queue(&thread_status[i].queue, (blas_queue_t *)0); thread_status[i].status = THREAD_STATUS_WAKEUP; pthread_mutex_init(&thread_status[i].lock, NULL); @@ -655,7 +657,8 @@ int exec_blas_async(BLASLONG pos, blas_queue_t *queue){ if (queue -> mode & BLAS_NODE) { do { - while((thread_status[i].node != node || thread_status[i].queue) && (i < blas_num_threads - 1)) i ++; + + while((thread_status[i].node != node || atomic_load_queue(&thread_status[i].queue)) && (i < blas_num_threads - 1)) i ++; if (i < blas_num_threads - 1) break; @@ -669,36 +672,26 @@ int exec_blas_async(BLASLONG pos, blas_queue_t *queue){ } while (1); } else { - pthread_mutex_lock (&thread_status[i].lock); - tsiq = thread_status[i].queue; - pthread_mutex_unlock (&thread_status[i].lock); + tsiq = atomic_load_queue(&thread_status[i].queue); while(tsiq) { i ++; if (i >= blas_num_threads - 1) i = 0; - pthread_mutex_lock (&thread_status[i].lock); - tsiq = thread_status[i].queue; - pthread_mutex_unlock (&thread_status[i].lock); + tsiq = atomic_load_queue(&thread_status[i].queue); } } #else - pthread_mutex_lock (&thread_status[i].lock); - tsiq=thread_status[i].queue ; - pthread_mutex_unlock (&thread_status[i].lock); + tsiq = atomic_load_queue(&thread_status[i].queue); while(tsiq) { i ++; if (i >= blas_num_threads - 1) i = 0; - pthread_mutex_lock (&thread_status[i].lock); - tsiq=thread_status[i].queue ; - pthread_mutex_unlock (&thread_status[i].lock); + tsiq = atomic_load_queue(&thread_status[i].queue); } #endif queue -> assigned = i; - WMB; - pthread_mutex_lock (&thread_status[i].lock); - thread_status[i].queue = queue; - pthread_mutex_unlock (&thread_status[i].lock); - WMB; + MB; + + atomic_store_queue(&thread_status[i].queue, queue); queue = queue -> next; pos ++; @@ -718,9 +711,7 @@ int exec_blas_async(BLASLONG pos, blas_queue_t *queue){ pos = current -> assigned; - pthread_mutex_lock (&thread_status[pos].lock); - tspq=thread_status[pos].queue; - pthread_mutex_unlock (&thread_status[pos].lock); + tspq = atomic_load_queue(&thread_status[pos].queue); if ((BLASULONG)tspq > 1) { pthread_mutex_lock (&thread_status[pos].lock); @@ -752,24 +743,20 @@ int exec_blas_async_wait(BLASLONG num, blas_queue_t *queue){ while ((num > 0) && queue) { - pthread_mutex_lock(&thread_status[queue->assigned].lock); - tsqq=thread_status[queue -> assigned].queue; - pthread_mutex_unlock(&thread_status[queue->assigned].lock); + tsqq = atomic_load_queue(&thread_status[queue->assigned].queue); while(tsqq) { YIELDING; - pthread_mutex_lock(&thread_status[queue->assigned].lock); - tsqq=thread_status[queue -> assigned].queue; - pthread_mutex_unlock(&thread_status[queue->assigned].lock); - - + tsqq = atomic_load_queue(&thread_status[queue->assigned].queue); }; queue = queue -> next; num --; } + MB; + #ifdef SMP_DEBUG fprintf(STDERR, "Done.\n\n"); #endif @@ -880,7 +867,7 @@ void goto_set_num_threads(int num_threads) { for(i = blas_num_threads - 1; i < num_threads - 1; i++){ - thread_status[i].queue = (blas_queue_t *)NULL; + atomic_store_queue(&thread_status[i].queue, (blas_queue_t *)0); thread_status[i].status = THREAD_STATUS_WAKEUP; pthread_mutex_init(&thread_status[i].lock, NULL); @@ -971,12 +958,11 @@ int BLASFUNC(blas_thread_shutdown)(void){ for (i = 0; i < blas_num_threads - 1; i++) { + pthread_mutex_lock (&thread_status[i].lock); - thread_status[i].queue = (blas_queue_t *)-1; - + atomic_store_queue(&thread_status[i].queue, (blas_queue_t *)-1); thread_status[i].status = THREAD_STATUS_WAKEUP; - pthread_cond_signal (&thread_status[i].wakeup); pthread_mutex_unlock(&thread_status[i].lock); diff --git a/driver/others/dynamic_arm64.c b/driver/others/dynamic_arm64.c index 72f5fcca2..11ef2725c 100644 --- a/driver/others/dynamic_arm64.c +++ b/driver/others/dynamic_arm64.c @@ -51,10 +51,12 @@ extern gotoblas_t gotoblas_FALKOR; extern gotoblas_t gotoblas_THUNDERX; extern gotoblas_t gotoblas_THUNDERX2T99; extern gotoblas_t gotoblas_TSV110; +extern gotoblas_t gotoblas_EMAG8180; +extern gotoblas_t gotoblas_NEOVERSEN1; extern void openblas_warning(int verbose, const char * msg); -#define NUM_CORETYPES 9 +#define NUM_CORETYPES 11 /* * In case asm/hwcap.h is outdated on the build system, make sure @@ -78,6 +80,8 @@ static char *corename[] = { "thunderx", "thunderx2t99", "tsv110", + "emag8180", + "neoversen1", "unknown" }; @@ -91,6 +95,8 @@ char *gotoblas_corename(void) { if (gotoblas == &gotoblas_THUNDERX) return corename[ 6]; if (gotoblas == &gotoblas_THUNDERX2T99) return corename[ 7]; if (gotoblas == &gotoblas_TSV110) return corename[ 8]; + if (gotoblas == &gotoblas_EMAG8180) return corename[ 9]; + if (gotoblas == &gotoblas_NEOVERSEN1) return corename[10]; return corename[NUM_CORETYPES]; } @@ -119,6 +125,8 @@ static gotoblas_t *force_coretype(char *coretype) { case 6: return (&gotoblas_THUNDERX); case 7: return (&gotoblas_THUNDERX2T99); case 8: return (&gotoblas_TSV110); + case 9: return (&gotoblas_EMAG8180); + case 10: return (&gotoblas_NEOVERSEN1); } snprintf(message, 128, "Core not found: %s\n", coretype); openblas_warning(1, message); @@ -164,6 +172,8 @@ static gotoblas_t *get_coretype(void) { return &gotoblas_CORTEXA72; case 0xd09: // Cortex A73 return &gotoblas_CORTEXA73; + case 0xd0c: // Neoverse N1 + return &gotoblas_NEOVERSEN1; } break; case 0x42: // Broadcom @@ -189,6 +199,13 @@ static gotoblas_t *get_coretype(void) { return &gotoblas_TSV110; } break; + case 0x50: // Ampere + switch (part) + { + case 0x000: // Skylark/EMAG8180 + return &gotoblas_EMAG8180; + } + break; case 0x51: // Qualcomm switch (part) { diff --git a/driver/others/dynamic_power.c b/driver/others/dynamic_power.c index 1dec5f4b3..8c831b998 100644 --- a/driver/others/dynamic_power.c +++ b/driver/others/dynamic_power.c @@ -3,7 +3,7 @@ extern gotoblas_t gotoblas_POWER6; extern gotoblas_t gotoblas_POWER8; -#if (!defined C_GCC) || (GCC_VERSION >= 60000) +#if (!defined __GNUC__) || ( __GNUC__ >= 6) extern gotoblas_t gotoblas_POWER9; #endif @@ -21,7 +21,7 @@ static char *corename[] = { char *gotoblas_corename(void) { if (gotoblas == &gotoblas_POWER6) return corename[1]; if (gotoblas == &gotoblas_POWER8) return corename[2]; -#if (!defined C_GCC) || (GCC_VERSION >= 60000) +#if (!defined __GNUC__) || ( __GNUC__ >= 6) if (gotoblas == &gotoblas_POWER9) return corename[3]; #endif return corename[0]; @@ -33,7 +33,7 @@ static gotoblas_t *get_coretype(void) { return &gotoblas_POWER6; if (__builtin_cpu_is("power8")) return &gotoblas_POWER8; -#if (!defined C_GCC) || (GCC_VERSION >= 60000) +#if (!defined __GNUC__) || ( __GNUC__ >= 6) if (__builtin_cpu_is("power9")) return &gotoblas_POWER9; #endif @@ -59,7 +59,7 @@ static gotoblas_t *force_coretype(char * coretype) { { case 1: return (&gotoblas_POWER6); case 2: return (&gotoblas_POWER8); -#if (!defined C_GCC) || (GCC_VERSION >= 60000) +#if (!defined __GNUC__) || ( __GNUC__ >= 6) case 3: return (&gotoblas_POWER9); #endif default: return NULL; diff --git a/driver/others/dynamic_zarch.c b/driver/others/dynamic_zarch.c index 1206bf870..90d3051b1 100644 --- a/driver/others/dynamic_zarch.c +++ b/driver/others/dynamic_zarch.c @@ -3,12 +3,12 @@ extern gotoblas_t gotoblas_Z13; extern gotoblas_t gotoblas_Z14; -extern gotoblas_t gotoblas_Z15; +//extern gotoblas_t gotoblas_Z15; //#if (!defined C_GCC) || (GCC_VERSION >= 60000) //extern gotoblas_t gotoblas_Z14; //#endif -#define NUM_CORETYPES 5 +#define NUM_CORETYPES 4 extern void openblas_warning(int verbose, const char* msg); @@ -16,14 +16,14 @@ static char* corename[] = { "unknown", "Z13", "Z14", - "Z15", +// "Z15", "ZARCH_GENERIC", }; char* gotoblas_corename(void) { if (gotoblas == &gotoblas_Z13) return corename[1]; if (gotoblas == &gotoblas_Z14) return corename[2]; - if (gotoblas == &gotoblas_Z15) return corename[3]; +// if (gotoblas == &gotoblas_Z15) return corename[3]; //#if (!defined C_GCC) || (GCC_VERSION >= 60000) // if (gotoblas == &gotoblas_POWER9) return corename[3]; //#endif @@ -31,7 +31,7 @@ char* gotoblas_corename(void) { } // __builtin_cpu_is is not supported by zarch -static gotolabs_t* get_coretype(void) { +static gotoblas_t* get_coretype(void) { FILE* infile; char buffer[512], * p; @@ -78,7 +78,7 @@ static gotoblas_t* force_coretype(char* coretype) { { case 1: return (&gotoblas_Z13); case 2: return (&gotoblas_Z14); - case 3: return (&gotoblas_Z15); +// case 3: return (&gotoblas_Z15); //#if (!defined C_GCC) || (GCC_VERSION >= 60000) // case 3: return (&gotoblas_POWER9); //#endif diff --git a/driver/others/memory.c b/driver/others/memory.c index 55dce72b8..62a5a0214 100644 --- a/driver/others/memory.c +++ b/driver/others/memory.c @@ -822,7 +822,7 @@ static void *alloc_qalloc(void *address){ static void alloc_windows_free(struct alloc_t *alloc_info){ - VirtualFree(alloc_info, allocation_block_size, MEM_DECOMMIT); + VirtualFree(alloc_info, 0, MEM_RELEASE); } @@ -935,7 +935,7 @@ static void alloc_hugetlb_free(struct alloc_t *alloc_info){ #ifdef OS_WINDOWS - VirtualFree(alloc_info, allocation_block_size, MEM_LARGE_PAGES | MEM_DECOMMIT); + VirtualFree(alloc_info, 0, MEM_LARGE_PAGES | MEM_RELEASE); #endif @@ -2310,7 +2310,7 @@ static void *alloc_qalloc(void *address){ static void alloc_windows_free(struct release_t *release){ - VirtualFree(release -> address, BUFFER_SIZE, MEM_DECOMMIT); + VirtualFree(release -> address, 0, MEM_RELEASE); } @@ -2432,7 +2432,7 @@ static void alloc_hugetlb_free(struct release_t *release){ #ifdef OS_WINDOWS - VirtualFree(release -> address, BUFFER_SIZE, MEM_LARGE_PAGES | MEM_DECOMMIT); + VirtualFree(release -> address, 0, MEM_LARGE_PAGES | MEM_RELEASE); #endif diff --git a/exports/dllinit.c b/exports/dllinit.c index 4a05c0e14..88f9af658 100644 --- a/exports/dllinit.c +++ b/exports/dllinit.c @@ -50,7 +50,10 @@ BOOL APIENTRY DllMain(HINSTANCE hInst, DWORD reason, LPVOID reserved) { gotoblas_init(); break; case DLL_PROCESS_DETACH: - gotoblas_quit(); + // If the process is about to exit, don't bother releasing any resources + // The kernel is much better at bulk releasing then. + if (!reserved) + gotoblas_quit(); break; case DLL_THREAD_ATTACH: break; diff --git a/f_check b/f_check index 79b24e2dc..fac8fc707 100644 --- a/f_check +++ b/f_check @@ -334,6 +334,7 @@ if ($link ne "") { && ($flags !~ /kernel32/) && ($flags !~ /advapi32/) && ($flags !~ /shell32/) + && ($flags !~ /omp/) && ($flags !~ /^\-l$/) ) { $linker_l .= $flags . " "; diff --git a/getarch.c b/getarch.c index 1f590390a..30ca290e3 100644 --- a/getarch.c +++ b/getarch.c @@ -1028,6 +1028,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #else #endif +#ifdef FORCE_NEOVERSEN1 +#define FORCE +#define ARCHITECTURE "ARM64" +#define SUBARCHITECTURE "NEOVERSEN1" +#define SUBDIRNAME "arm64" +#define ARCHCONFIG "-DNEOVERSEN1 " \ + "-DL1_CODE_SIZE=65536 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=4 " \ + "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 -DL1_DATA_ASSOCIATIVE=4 " \ + "-DL2_SIZE=1048576 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=16 " \ + "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \ + "-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8" \ + "-march=armv8.2-a -mtune=cortex-a72" +#define LIBNAME "neoversen1" +#define CORENAME "NEOVERSEN1" +#else +#endif + + #ifdef FORCE_FALKOR #define FORCE #define ARCHITECTURE "ARM64" @@ -1093,6 +1111,22 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #else #endif +#ifdef FORCE_EMAG8180 +#define ARMV8 +#define FORCE +#define ARCHITECTURE "ARM64" +#define SUBARCHITECTURE "EMAG8180" +#define SUBDIRNAME "arm64" +#define ARCHCONFIG "-DEMAG8180 " \ + "-DL1_CODE_SIZE=32768 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=8 " \ + "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 -DL1_DATA_ASSOCIATIVE=8 " \ + "-DL2_SIZE=262144 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=8 " \ + "-DL3_SIZE=33554432 -DL3_LINESIZE=64 -DL3_ASSOCIATIVE=32 " \ + "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \ + "-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8" +#define LIBNAME "emag8180" +#define CORENAME "EMAG8180" +#endif #ifdef FORCE_ZARCH_GENERIC #define FORCE @@ -1298,6 +1332,13 @@ int main(int argc, char *argv[]){ #endif #endif +#if defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ +printf("__BYTE_ORDER__=__ORDER_BIG_ENDIAN__\n"); +#endif +#if defined(__BIG_ENDIAN__) && __BIG_ENDIAN__ > 0 +printf("__BYTE_ORDER__=__ORDER_BIG_ENDIAN__\n"); +#endif + #ifdef MAKE_NB_JOBS #if MAKE_NB_JOBS > 0 printf("MAKE += -j %d\n", MAKE_NB_JOBS); diff --git a/interface/lapack/gesv.c b/interface/lapack/gesv.c index 721da970d..175350329 100644 --- a/interface/lapack/gesv.c +++ b/interface/lapack/gesv.c @@ -44,19 +44,19 @@ #ifndef COMPLEX #ifdef XDOUBLE -#define ERROR_NAME "QGESV " +#define ERROR_NAME "QGESV" #elif defined(DOUBLE) -#define ERROR_NAME "DGESV " +#define ERROR_NAME "DGESV" #else -#define ERROR_NAME "SGESV " +#define ERROR_NAME "SGESV" #endif #else #ifdef XDOUBLE -#define ERROR_NAME "XGESV " +#define ERROR_NAME "XGESV" #elif defined(DOUBLE) -#define ERROR_NAME "ZGESV " +#define ERROR_NAME "ZGESV" #else -#define ERROR_NAME "CGESV " +#define ERROR_NAME "CGESV" #endif #endif @@ -89,7 +89,7 @@ int NAME(blasint *N, blasint *NRHS, FLOAT *a, blasint *ldA, blasint *ipiv, if (args.m < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/getf2.c b/interface/lapack/getf2.c index 3e66c0403..8506feca9 100644 --- a/interface/lapack/getf2.c +++ b/interface/lapack/getf2.c @@ -74,7 +74,7 @@ int NAME(blasint *M, blasint *N, FLOAT *a, blasint *ldA, blasint *ipiv, blasint if (args.n < 0) info = 2; if (args.m < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/getrf.c b/interface/lapack/getrf.c index 44a92ddc4..02bb124b3 100644 --- a/interface/lapack/getrf.c +++ b/interface/lapack/getrf.c @@ -74,7 +74,7 @@ int NAME(blasint *M, blasint *N, FLOAT *a, blasint *ldA, blasint *ipiv, blasint if (args.n < 0) info = 2; if (args.m < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/getrs.c b/interface/lapack/getrs.c index 1b8c83aca..c2a9eb882 100644 --- a/interface/lapack/getrs.c +++ b/interface/lapack/getrs.c @@ -102,7 +102,7 @@ int NAME(char *TRANS, blasint *N, blasint *NRHS, FLOAT *a, blasint *ldA, if (trans < 0) info = 1; if (info != 0) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); return 0; } diff --git a/interface/lapack/lauu2.c b/interface/lapack/lauu2.c index 3599a4791..e581e3c15 100644 --- a/interface/lapack/lauu2.c +++ b/interface/lapack/lauu2.c @@ -90,7 +90,7 @@ int NAME(char *UPLO, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){ if (args.n < 0) info = 2; if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/lauum.c b/interface/lapack/lauum.c index 2c49eb0b0..70f6a0ec5 100644 --- a/interface/lapack/lauum.c +++ b/interface/lapack/lauum.c @@ -90,7 +90,7 @@ int NAME(char *UPLO, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){ if (args.n < 0) info = 2; if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/potf2.c b/interface/lapack/potf2.c index 837192265..1537b6ee4 100644 --- a/interface/lapack/potf2.c +++ b/interface/lapack/potf2.c @@ -90,7 +90,7 @@ int NAME(char *UPLO, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){ if (args.n < 0) info = 2; if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/potrf.c b/interface/lapack/potrf.c index 092272225..dbd55f62f 100644 --- a/interface/lapack/potrf.c +++ b/interface/lapack/potrf.c @@ -90,7 +90,7 @@ int NAME(char *UPLO, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){ if (args.n < 0) info = 2; if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/potri.c b/interface/lapack/potri.c index d6230621f..2c0c64b6f 100644 --- a/interface/lapack/potri.c +++ b/interface/lapack/potri.c @@ -99,7 +99,7 @@ int NAME(char *UPLO, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){ if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/trti2.c b/interface/lapack/trti2.c index 42c4c4815..47f04f06f 100644 --- a/interface/lapack/trti2.c +++ b/interface/lapack/trti2.c @@ -96,7 +96,7 @@ int NAME(char *UPLO, char *DIAG, blasint *N, FLOAT *a, blasint *ldA, blasint *In if (diag < 0) info = 2; if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/trtri.c b/interface/lapack/trtri.c index 6724a678a..028529389 100644 --- a/interface/lapack/trtri.c +++ b/interface/lapack/trtri.c @@ -99,7 +99,7 @@ int NAME(char *UPLO, char *DIAG, blasint *N, FLOAT *a, blasint *ldA, blasint *In if (diag < 0) info = 2; if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/zgetf2.c b/interface/lapack/zgetf2.c index 59ec4874e..68b9a7e4b 100644 --- a/interface/lapack/zgetf2.c +++ b/interface/lapack/zgetf2.c @@ -74,7 +74,7 @@ int NAME(blasint *M, blasint *N, FLOAT *a, blasint *ldA, blasint *ipiv, blasint if (args.n < 0) info = 2; if (args.m < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/zgetrf.c b/interface/lapack/zgetrf.c index 5031f587b..7f8db94f6 100644 --- a/interface/lapack/zgetrf.c +++ b/interface/lapack/zgetrf.c @@ -74,7 +74,7 @@ int NAME(blasint *M, blasint *N, FLOAT *a, blasint *ldA, blasint *ipiv, blasint if (args.n < 0) info = 2; if (args.m < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/zgetrs.c b/interface/lapack/zgetrs.c index 54d4b0905..0add909ca 100644 --- a/interface/lapack/zgetrs.c +++ b/interface/lapack/zgetrs.c @@ -102,7 +102,7 @@ int NAME(char *TRANS, blasint *N, blasint *NRHS, FLOAT *a, blasint *ldA, if (trans < 0) info = 1; if (info != 0) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); return 0; } diff --git a/interface/lapack/zlauu2.c b/interface/lapack/zlauu2.c index b0698ef2e..ae972543c 100644 --- a/interface/lapack/zlauu2.c +++ b/interface/lapack/zlauu2.c @@ -91,7 +91,7 @@ int NAME(char *UPLO, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){ if (args.n < 0) info = 2; if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/zpotf2.c b/interface/lapack/zpotf2.c index 27ee0891a..c74b66728 100644 --- a/interface/lapack/zpotf2.c +++ b/interface/lapack/zpotf2.c @@ -91,7 +91,7 @@ int NAME(char *UPLO, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){ if (args.n < 0) info = 2; if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/zpotrf.c b/interface/lapack/zpotrf.c index 8cd3980d5..c4cd99bf6 100644 --- a/interface/lapack/zpotrf.c +++ b/interface/lapack/zpotrf.c @@ -90,7 +90,7 @@ int NAME(char *UPLO, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){ if (args.n < 0) info = 2; if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/zpotri.c b/interface/lapack/zpotri.c index 7c72a7e62..8da211683 100644 --- a/interface/lapack/zpotri.c +++ b/interface/lapack/zpotri.c @@ -99,7 +99,7 @@ int NAME(char *UPLO, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){ if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/ztrti2.c b/interface/lapack/ztrti2.c index a25476677..cb9c0d557 100644 --- a/interface/lapack/ztrti2.c +++ b/interface/lapack/ztrti2.c @@ -96,7 +96,7 @@ int NAME(char *UPLO, char *DIAG, blasint *N, FLOAT *a, blasint *ldA, blasint *In if (diag < 0) info = 2; if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/interface/lapack/ztrtri.c b/interface/lapack/ztrtri.c index b3ce85b9f..dda4a9e4b 100644 --- a/interface/lapack/ztrtri.c +++ b/interface/lapack/ztrtri.c @@ -96,7 +96,7 @@ int NAME(char *UPLO, char *DIAG, blasint *N, FLOAT *a, blasint *ldA, blasint *In if (diag < 0) info = 2; if (uplo < 0) info = 1; if (info) { - BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME)); + BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME) - 1); *Info = - info; return 0; } diff --git a/kernel/CMakeLists.txt b/kernel/CMakeLists.txt index ad15b8f25..35e0fff25 100644 --- a/kernel/CMakeLists.txt +++ b/kernel/CMakeLists.txt @@ -47,7 +47,7 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS) GenerateNamedObjects("${KERNELDIR}/${${float_char}MAXKERNEL}" "" "max_k" false "" "" false ${float_type}) endif () if (DEFINED ${float_char}MINKERNEL) - GenerateNamedObjects("${KERNELDIR}/${${float_char}MINKERNEL}" "" "min_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}MINKERNEL}" "USE_MIN" "min_k" false "" "" false ${float_type}) endif () GenerateNamedObjects("${KERNELDIR}/${I${float_char}AMAXKERNEL}" "USE_ABS" "i*amax_k" false "" "" false ${float_type}) GenerateNamedObjects("${KERNELDIR}/${I${float_char}AMINKERNEL}" "USE_ABS;USE_MIN" "i*amin_k" false "" "" false ${float_type}) @@ -55,7 +55,7 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS) GenerateNamedObjects("${KERNELDIR}/${I${float_char}MAXKERNEL}" "" "i*max_k" false "" "" false ${float_type}) endif () if (DEFINED I${float_char}MINKERNEL) - GenerateNamedObjects("${KERNELDIR}/${I${float_char}MINKERNEL}" "" "i*min_k" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${I${float_char}MINKERNEL}" "USE_MIN" "i*min_k" false "" "" false ${float_type}) endif () GenerateNamedObjects("${KERNELDIR}/${${float_char}ASUMKERNEL}" "" "asum_k" false "" "" false ${float_type}) GenerateNamedObjects("${KERNELDIR}/${${float_char}AXPYKERNEL}" "" "axpy_k" false "" "" false ${float_type}) @@ -121,8 +121,10 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS) # Makefile.L3 set(USE_TRMM false) - - if (ARM OR ARM64 OR "${TARGET_CORE}" STREQUAL "LONGSOON3B" OR "${TARGET_CORE}" STREQUAL "GENERIC" OR "${CORE}" STREQUAL "generic" OR "${TARGET_CORE}" STREQUAL "HASWELL" OR "${CORE}" STREQUAL "haswell" OR "${CORE}" STREQUAL "zen" OR "${TARGET_CORE}" STREQUAL "SKYLAKEX" OR "${CORE}" STREQUAL "skylakex") + 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)) set(USE_TRMM true) endif () diff --git a/kernel/arm64/KERNEL.EMAG8180 b/kernel/arm64/KERNEL.EMAG8180 new file mode 100644 index 000000000..007b2ce26 --- /dev/null +++ b/kernel/arm64/KERNEL.EMAG8180 @@ -0,0 +1,3 @@ +include $(KERNELDIR)/KERNEL.CORTEXA57 + + diff --git a/kernel/arm64/KERNEL.NEOVERSEN1 b/kernel/arm64/KERNEL.NEOVERSEN1 new file mode 100644 index 000000000..ea010db42 --- /dev/null +++ b/kernel/arm64/KERNEL.NEOVERSEN1 @@ -0,0 +1,189 @@ +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 = daxpy_thunderx2t99.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 = sasum_thunderx2t99.c +DASUMKERNEL = dasum_thunderx2t99.c +CASUMKERNEL = casum_thunderx2t99.c +ZASUMKERNEL = zasum_thunderx2t99.c + +SCOPYKERNEL = copy_thunderx2t99.c +DCOPYKERNEL = copy_thunderx2t99.c +CCOPYKERNEL = copy_thunderx2t99.c +ZCOPYKERNEL = copy_thunderx2t99.c + +SSWAPKERNEL = swap_thunderx2t99.S +DSWAPKERNEL = swap_thunderx2t99.S +CSWAPKERNEL = swap_thunderx2t99.S +ZSWAPKERNEL = swap_thunderx2t99.S + +ISAMAXKERNEL = iamax_thunderx2t99.c +IDAMAXKERNEL = iamax_thunderx2t99.c +ICAMAXKERNEL = izamax_thunderx2t99.c +IZAMAXKERNEL = izamax_thunderx2t99.c + +SNRM2KERNEL = scnrm2_thunderx2t99.c +DNRM2KERNEL = dznrm2_thunderx2t99.c +CNRM2KERNEL = scnrm2_thunderx2t99.c +ZNRM2KERNEL = dznrm2_thunderx2t99.c + +DDOTKERNEL = dot_thunderx2t99.c +SDOTKERNEL = dot_thunderx2t99.c +CDOTKERNEL = zdot_thunderx2t99.c +ZDOTKERNEL = zdot_thunderx2t99.c +DSDOTKERNEL = dot.S + +DGEMM_BETA = dgemm_beta.S +SGEMM_BETA = sgemm_beta.S + +SGEMMKERNEL = sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S +STRMMKERNEL = strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S +ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N)) +ifeq ($(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).$(SUFFIX) +endif +ifeq ($(SGEMM_UNROLL_N), 16) +SGEMMOTCOPY = sgemm_tcopy_$(SGEMM_UNROLL_N).S +else +SGEMMOTCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c +endif +ifeq ($(SGEMM_UNROLL_N), 4) +SGEMMONCOPY = sgemm_ncopy_$(SGEMM_UNROLL_N).S +else +SGEMMONCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c +endif +SGEMMONCOPYOBJ = sgemm_oncopy$(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) diff --git a/kernel/generic/gemm_beta.c b/kernel/generic/gemm_beta.c index c4e4f7abe..fa9d7680d 100644 --- a/kernel/generic/gemm_beta.c +++ b/kernel/generic/gemm_beta.c @@ -42,101 +42,53 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta, FLOAT *dummy2, BLASLONG dummy3, FLOAT *dummy4, BLASLONG dummy5, FLOAT *c, BLASLONG ldc){ + BLASLONG i, j; + BLASLONG chunk, remain; FLOAT *c_offset1, *c_offset; - FLOAT ctemp1, ctemp2, ctemp3, ctemp4; - FLOAT ctemp5, ctemp6, ctemp7, ctemp8; - c_offset = c; - + chunk = m >> 3; + remain = m & 7; if (beta == ZERO){ - - j = n; - do { - c_offset1 = c_offset; - c_offset += ldc; - - i = (m >> 3); - if (i > 0){ - do { - *(c_offset1 + 0) = ZERO; - *(c_offset1 + 1) = ZERO; - *(c_offset1 + 2) = ZERO; - *(c_offset1 + 3) = ZERO; - *(c_offset1 + 4) = ZERO; - *(c_offset1 + 5) = ZERO; - *(c_offset1 + 6) = ZERO; - *(c_offset1 + 7) = ZERO; - c_offset1 += 8; - i --; - } while (i > 0); - } - - i = (m & 7); - if (i > 0){ - do { - *c_offset1 = ZERO; - c_offset1 ++; - i --; - } while (i > 0); - } - j --; - } while (j > 0); - + for(j=n; j>0; j--){ + c_offset1 = c_offset; + c_offset += ldc; + for(i=chunk; i>0; i--){ + *(c_offset1 + 0) = ZERO; + *(c_offset1 + 1) = ZERO; + *(c_offset1 + 2) = ZERO; + *(c_offset1 + 3) = ZERO; + *(c_offset1 + 4) = ZERO; + *(c_offset1 + 5) = ZERO; + *(c_offset1 + 6) = ZERO; + *(c_offset1 + 7) = ZERO; + c_offset1 += 8; + } + for(i=remain; i>0; i--){ + *c_offset1 = ZERO; + c_offset1 ++; + } + } } else { - - j = n; - do { - c_offset1 = c_offset; - c_offset += ldc; - - i = (m >> 3); - if (i > 0){ - do { - ctemp1 = *(c_offset1 + 0); - ctemp2 = *(c_offset1 + 1); - ctemp3 = *(c_offset1 + 2); - ctemp4 = *(c_offset1 + 3); - ctemp5 = *(c_offset1 + 4); - ctemp6 = *(c_offset1 + 5); - ctemp7 = *(c_offset1 + 6); - ctemp8 = *(c_offset1 + 7); - - ctemp1 *= beta; - ctemp2 *= beta; - ctemp3 *= beta; - ctemp4 *= beta; - ctemp5 *= beta; - ctemp6 *= beta; - ctemp7 *= beta; - ctemp8 *= beta; - - *(c_offset1 + 0) = ctemp1; - *(c_offset1 + 1) = ctemp2; - *(c_offset1 + 2) = ctemp3; - *(c_offset1 + 3) = ctemp4; - *(c_offset1 + 4) = ctemp5; - *(c_offset1 + 5) = ctemp6; - *(c_offset1 + 6) = ctemp7; - *(c_offset1 + 7) = ctemp8; - c_offset1 += 8; - i --; - } while (i > 0); - } - - i = (m & 7); - if (i > 0){ - do { - ctemp1 = *c_offset1; - ctemp1 *= beta; - *c_offset1 = ctemp1; - c_offset1 ++; - i --; - } while (i > 0); - } - j --; - } while (j > 0); - + for(j=n; j>0; j--){ + c_offset1 = c_offset; + c_offset += ldc; + for(i=chunk; i>0; i--){ + *(c_offset1 + 0) *= beta; + *(c_offset1 + 1) *= beta; + *(c_offset1 + 2) *= beta; + *(c_offset1 + 3) *= beta; + *(c_offset1 + 4) *= beta; + *(c_offset1 + 5) *= beta; + *(c_offset1 + 6) *= beta; + *(c_offset1 + 7) *= beta; + c_offset1 += 8; + } + for(i=remain; i>0; i--){ + *c_offset1 *= beta; + c_offset1 ++; + } + } } return 0; }; diff --git a/kernel/mips64/KERNEL b/kernel/mips64/KERNEL index 61da7445f..97ef3692c 100644 --- a/kernel/mips64/KERNEL +++ b/kernel/mips64/KERNEL @@ -167,3 +167,27 @@ endif CGEMM3MKERNEL = zgemm3m_kernel.S ZGEMM3MKERNEL = zgemm3m_kernel.S + +ifndef ISMINKERNEL +ISMINKERNEL = imin.S +endif + +ifndef IDMINKERNEL +IDMINKERNEL = imin.S +endif + +ifndef IQMINKERNEL +IQMINKERNEL = imin.S +endif + +ifndef ISMAXKERNEL +ISMAXKERNEL = imax.S +endif + +ifndef IDMAXKERNEL +IDMAXKERNEL = imax.S +endif + +ifndef IQMAXKERNEL +IQMAXKERNEL = imax.S +endif diff --git a/kernel/power/KERNEL b/kernel/power/KERNEL index c3c86b310..9070450f4 100644 --- a/kernel/power/KERNEL +++ b/kernel/power/KERNEL @@ -50,3 +50,26 @@ ifndef DSDOTKERNEL DSDOTKERNEL = ../generic/dot.c endif +ifndef ISMINKERNEL +ISMINKERNEL = imin.S +endif + +ifndef IDMINKERNEL +IDMINKERNEL = imin.S +endif + +ifndef IQMINKERNEL +IQMINKERNEL = imin.S +endif + +ifndef ISMAXKERNEL +ISMAXKERNEL = imax.S +endif + +ifndef IDMAXKERNEL +IDMAXKERNEL = imax.S +endif + +ifndef IQMAXKERNEL +IQMAXKERNEL = imax.S +endif diff --git a/kernel/power/KERNEL.POWER8 b/kernel/power/KERNEL.POWER8 index fb9452a35..c7867012b 100644 --- a/kernel/power/KERNEL.POWER8 +++ b/kernel/power/KERNEL.POWER8 @@ -88,31 +88,56 @@ ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c # #SMINKERNEL = ../arm/min.c #DMINKERNEL = ../arm/min.c -# -ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) + +ISMINKERNEL = imin.S +ISMAXKERNEL = imax.S + +ifneq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) +ifneq ($(GCCVERSIONGTEQ9),1) ISAMAXKERNEL = isamax_power8.S else ISAMAXKERNEL = isamax.c endif +else +ISAMAXKERNEL = isamax.c +endif +# IDAMAXKERNEL = idamax.c -ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) +# +ifneq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) +ifneq ($(GCCVERSIONGTEQ9),1) ICAMAXKERNEL = icamax_power8.S else ICAMAXKERNEL = icamax.c endif +else +ICAMAXKERNEL = icamax.c +endif +# IZAMAXKERNEL = izamax.c # -ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) +ifneq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) +ifneq ($(GCCVERSIONGTEQ9),1) ISAMINKERNEL = isamin_power8.S else ISAMINKERNEL = isamin.c endif +else +ISAMINKERNEL = isamin.c +endif +# IDAMINKERNEL = idamin.c -ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) +# +ifneq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) +ifneq ($(GCCVERSIONGTEQ9),1) ICAMINKERNEL = icamin_power8.S else ICAMINKERNEL = icamin.c endif +else +ICAMINKERNEL = icamin.c +endif +# IZAMINKERNEL = izamin.c # #ISMAXKERNEL = ../arm/imax.c @@ -128,11 +153,16 @@ ZASUMKERNEL = zasum.c # SAXPYKERNEL = saxpy.c DAXPYKERNEL = daxpy.c -ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) +ifneq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) +ifneq ($(GCCVERSIONGTEQ9),1) CAXPYKERNEL = caxpy_power8.S else CAXPYKERNEL = caxpy.c endif +else +CAXPYKERNEL = caxpy.c +endif +# ZAXPYKERNEL = zaxpy.c # SCOPYKERNEL = scopy.c diff --git a/kernel/power/KERNEL.POWER9 b/kernel/power/KERNEL.POWER9 index 4bfa017e1..aabb5d976 100644 --- a/kernel/power/KERNEL.POWER9 +++ b/kernel/power/KERNEL.POWER9 @@ -1,3 +1,7 @@ +ifeq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) +include $(KERNELDIR)/KERNEL.POWER8 +else + #SGEMM_BETA = ../generic/gemm_beta.c #DGEMM_BETA = ../generic/gemm_beta.c #CGEMM_BETA = ../generic/zgemm_beta.c @@ -206,3 +210,5 @@ QCABS_KERNEL = ../generic/cabs.c #Dump kernel CGEMM3MKERNEL = ../generic/zgemm3mkernel_dump.c ZGEMM3MKERNEL = ../generic/zgemm3mkernel_dump.c + +endif diff --git a/kernel/power/KERNEL.PPC440 b/kernel/power/KERNEL.PPC440 index a0696b548..677af5f21 100644 --- a/kernel/power/KERNEL.PPC440 +++ b/kernel/power/KERNEL.PPC440 @@ -15,7 +15,7 @@ ZASUMKERNEL = zasum_ppc440.S SAXPYKERNEL = axpy_ppc440.S DAXPYKERNEL = axpy_ppc440.S -ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) +ifneq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) CAXPYKERNEL = ../arm/zaxpy.c ZAXPYKERNEL = ../arm/zaxpy.c else @@ -25,7 +25,7 @@ endif SDOTKERNEL = dot_ppc440.S DDOTKERNEL = dot_ppc440.S -ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) +ifneq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) CDOTKERNEL = zdot_ppc440.S ZDOTKERNEL = zdot_ppc440.S else @@ -62,7 +62,7 @@ ZNRM2KERNEL = znrm2_ppc440.S SROTKERNEL = rot_ppc440.S DROTKERNEL = rot_ppc440.S -ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) +ifneq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) CROTKERNEL = zrot_ppc440.S ZROTKERNEL = zrot_ppc440.S else @@ -132,7 +132,7 @@ ZTRSMKERNEL_LT = ztrsm_kernel_ppc440_LT.S ZTRSMKERNEL_RN = ztrsm_kernel_ppc440_LT.S ZTRSMKERNEL_RT = ztrsm_kernel_ppc440_RT.S -ifeq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) +ifeq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) SGEMVNKERNEL = ../arm/gemv_n.c DGEMVNKERNEL = ../arm/gemv_n.c SGEMVTKERNEL = ../arm/gemv_t.c diff --git a/kernel/power/KERNEL.PPC970 b/kernel/power/KERNEL.PPC970 index de30977de..a99fb7d96 100644 --- a/kernel/power/KERNEL.PPC970 +++ b/kernel/power/KERNEL.PPC970 @@ -1,4 +1,4 @@ -ifeq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) +ifeq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) SGEMMKERNEL = gemm_kernel.S SGEMMINCOPY = SGEMMITCOPY = @@ -30,7 +30,7 @@ DGEMMITCOPYOBJ = DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX) DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX) -ifeq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) +ifeq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) CGEMMKERNEL = zgemm_kernel.S CGEMMINCOPY = CGEMMITCOPY = @@ -72,7 +72,7 @@ ZTRSMKERNEL_LT = ztrsm_kernel_LT.S ZTRSMKERNEL_RN = ztrsm_kernel_LT.S ZTRSMKERNEL_RT = ztrsm_kernel_RT.S -ifeq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__)) +ifeq ($(__BYTE_ORDER__),__ORDER_BIG_ENDIAN__) STRSMKERNEL_LN = trsm_kernel_LN.S STRSMKERNEL_LT = trsm_kernel_LT.S STRSMKERNEL_RN = trsm_kernel_LT.S diff --git a/kernel/power/caxpy_power8.S b/kernel/power/caxpy_power8.S index 294a1d24d..10a372832 100644 --- a/kernel/power/caxpy_power8.S +++ b/kernel/power/caxpy_power8.S @@ -12,6 +12,13 @@ PROLOGUE +#if _CALL_ELF ==2 +#ifdef CONJ +caxpyc_k: +#else +caxpy_k: +#endif +#endif .LCF0: 0: addis 2,12,.TOC.-.LCF0@ha addi 2,2,.TOC.-.LCF0@l diff --git a/kernel/power/caxpy_power9.S b/kernel/power/caxpy_power9.S index 844cacd50..b4733ff9f 100644 --- a/kernel/power/caxpy_power9.S +++ b/kernel/power/caxpy_power9.S @@ -13,7 +13,11 @@ PROLOGUE +#ifdef CONJ +caxpyc_k: +#else caxpy_k: +#endif .LCF0: 0: addis 2,12,.TOC.-.LCF0@ha addi 2,2,.TOC.-.LCF0@l diff --git a/kernel/power/icamax_power9.S b/kernel/power/icamax_power9.S index 2968b3f8b..bf6ab6e82 100644 --- a/kernel/power/icamax_power9.S +++ b/kernel/power/icamax_power9.S @@ -1,3 +1,4 @@ +/* .file "icamax.c" .abiversion 2 .section ".text" @@ -5,6 +6,12 @@ .p2align 4,,15 .globl icamax_k .type icamax_k, @function +*/ +#define ASSEMBLER +#include "common.h" + + PROLOGUE + icamax_k: .LCF0: 0: addis 2,12,.TOC.-.LCF0@ha diff --git a/kernel/power/icamin_power8.S b/kernel/power/icamin_power8.S index f2993e83e..e4469eb64 100644 --- a/kernel/power/icamin_power8.S +++ b/kernel/power/icamin_power8.S @@ -10,7 +10,9 @@ #include "common.h" PROLOGUE - +#if _CALL_ELF ==2 +icamin_k: +#endif .LCF0: 0: addis 2,12,.TOC.-.LCF0@ha addi 2,2,.TOC.-.LCF0@l diff --git a/kernel/power/icamin_power9.S b/kernel/power/icamin_power9.S index 8eaa79f33..58a3c53a3 100644 --- a/kernel/power/icamin_power9.S +++ b/kernel/power/icamin_power9.S @@ -1,3 +1,4 @@ +/* .file "icamin.c" .abiversion 2 .section ".text" @@ -5,6 +6,12 @@ .p2align 4,,15 .globl icamin_k .type icamin_k, @function +*/ +#define ASSEMBLER +#include "common.h" + + PROLOGUE + icamin_k: .LCF0: 0: addis 2,12,.TOC.-.LCF0@ha diff --git a/kernel/power/isamax_power8.S b/kernel/power/isamax_power8.S index fa5433333..ec1c283f3 100644 --- a/kernel/power/isamax_power8.S +++ b/kernel/power/isamax_power8.S @@ -11,7 +11,10 @@ #include "common.h" PROLOGUE - + +#if _CALL_ELF == 2 +isamax_k: +#endif .LCF0: 0: addis 2,12,.TOC.-.LCF0@ha addi 2,2,.TOC.-.LCF0@l diff --git a/kernel/power/isamax_power9.S b/kernel/power/isamax_power9.S index 9df1e773c..259c996fc 100644 --- a/kernel/power/isamax_power9.S +++ b/kernel/power/isamax_power9.S @@ -1,3 +1,4 @@ +/* .file "isamax.c" .abiversion 2 .section ".text" @@ -5,6 +6,12 @@ .p2align 4,,15 .globl isamax_k .type isamax_k, @function +*/ +#define ASSEMBLER +#include "common.h" + + PROLOGUE + isamax_k: .LCF0: 0: addis 2,12,.TOC.-.LCF0@ha diff --git a/kernel/power/isamin_power8.S b/kernel/power/isamin_power8.S index c9b6acb85..1978af880 100644 --- a/kernel/power/isamin_power8.S +++ b/kernel/power/isamin_power8.S @@ -11,6 +11,9 @@ PROLOGUE +#if _CALL_ELF ==2 +isamin_k: +#endif .LCF0: 0: addis 2,12,.TOC.-.LCF0@ha addi 2,2,.TOC.-.LCF0@l diff --git a/kernel/power/isamin_power9.S b/kernel/power/isamin_power9.S index 0475edf46..36486ff02 100644 --- a/kernel/power/isamin_power9.S +++ b/kernel/power/isamin_power9.S @@ -1,3 +1,4 @@ +/* .file "isamin.c" .abiversion 2 .section ".text" @@ -5,6 +6,12 @@ .p2align 4,,15 .globl isamin_k .type isamin_k, @function +*/ +#define ASSEMBLER +#include "common.h" + + PROLOGUE + isamin_k: .LCF0: 0: addis 2,12,.TOC.-.LCF0@ha diff --git a/kernel/x86_64/KERNEL b/kernel/x86_64/KERNEL index 92d121ab2..4874711bb 100644 --- a/kernel/x86_64/KERNEL +++ b/kernel/x86_64/KERNEL @@ -171,7 +171,7 @@ IXAMAXKERNEL = izamax.S endif ifndef ISAMINKERNEL -ISAMINKERNEL = iamax.S +ISAMINKERNEL = iamax_sse.S endif ifndef IDAMINKERNEL @@ -207,7 +207,7 @@ IQMAXKERNEL = iamax.S endif ifndef ISMINKERNEL -ISMINKERNEL = iamax.S +ISMINKERNEL = iamax_sse.S endif ifndef IDMINKERNEL diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL index d24b7f3b3..f6ca5c2d5 100644 --- a/kernel/x86_64/KERNEL.HASWELL +++ b/kernel/x86_64/KERNEL.HASWELL @@ -32,7 +32,7 @@ CAXPYKERNEL = caxpy.c ZAXPYKERNEL = zaxpy.c STRMMKERNEL = sgemm_kernel_8x4_haswell.c -SGEMMKERNEL = sgemm_kernel_8x4_haswell.c +SGEMMKERNEL = sgemm_kernel_8x4_haswell_2.c SGEMM_BETA = sgemm_beta_skylakex.c SGEMMINCOPY = ../generic/gemm_ncopy_8.c SGEMMITCOPY = ../generic/gemm_tcopy_8.c diff --git a/kernel/x86_64/KERNEL.SKYLAKEX b/kernel/x86_64/KERNEL.SKYLAKEX index 0e6275748..9b3c83e42 100644 --- a/kernel/x86_64/KERNEL.SKYLAKEX +++ b/kernel/x86_64/KERNEL.SKYLAKEX @@ -1,16 +1,19 @@ include $(KERNELDIR)/KERNEL.HASWELL SGEMMKERNEL = sgemm_kernel_16x4_skylakex_2.c -STRMMKERNEL = sgemm_kernel_16x4_haswell.S +STRMMKERNEL = sgemm_kernel_16x4_skylakex_2.c SGEMMINCOPY = ../generic/gemm_ncopy_16.c SGEMMITCOPY = sgemm_tcopy_16_skylakex.c SGEMMONCOPY = sgemm_ncopy_4_skylakex.c SGEMMOTCOPY = ../generic/gemm_tcopy_4.c -DGEMMKERNEL = dgemm_kernel_4x8_skylakex_2.c - -DGEMMONCOPY = dgemm_ncopy_8_skylakex.c -DGEMMOTCOPY = dgemm_tcopy_8_skylakex.c +DGEMMKERNEL = dgemm_kernel_16x2_skylakex.c +DTRMMKERNEL = dgemm_kernel_16x2_skylakex.c +DGEMMINCOPY = ../generic/gemm_ncopy_16.c +DGEMMITCOPY = ../generic/gemm_tcopy_16.c +DGEMMONCOPY = ../generic/gemm_ncopy_2.c +DGEMMOTCOPY = ../generic/gemm_tcopy_2.c +DTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c SGEMM_BETA = sgemm_beta_skylakex.c DGEMM_BETA = dgemm_beta_skylakex.c diff --git a/kernel/x86_64/KERNEL.ZEN b/kernel/x86_64/KERNEL.ZEN index 7cec2e5ed..1cd02db74 100644 --- a/kernel/x86_64/KERNEL.ZEN +++ b/kernel/x86_64/KERNEL.ZEN @@ -31,7 +31,7 @@ CAXPYKERNEL = caxpy.c ZAXPYKERNEL = zaxpy.c STRMMKERNEL = sgemm_kernel_8x4_haswell.c -SGEMMKERNEL = sgemm_kernel_8x4_haswell.c +SGEMMKERNEL = sgemm_kernel_8x4_haswell_2.c SGEMMINCOPY = ../generic/gemm_ncopy_8.c SGEMMITCOPY = ../generic/gemm_tcopy_8.c SGEMMONCOPY = ../generic/gemm_ncopy_4.c diff --git a/kernel/x86_64/cgemm_kernel_8x2_haswell.c b/kernel/x86_64/cgemm_kernel_8x2_haswell.c index eab8c9ea5..08882346d 100644 --- a/kernel/x86_64/cgemm_kernel_8x2_haswell.c +++ b/kernel/x86_64/cgemm_kernel_8x2_haswell.c @@ -50,7 +50,7 @@ "vmovups (%0),%%ymm0; vmovups 32(%0),%%ymm1; prefetcht0 512(%0); addq $64,%0;"\ acc_m8n2_con(0,1,4,5,6,7,0,8,%1) acc_m8n2_con(0,1,8,9,10,11,0,8,%1,%%r12,1) #define KERNEL_2_k1m8n4 \ - "vpermilps $177,%%ymm0,%%ymm0; vpermilps $177,%%ymm1,%%ymm1;"\ + "vpermilps $177,-64(%0),%%ymm0; vpermilps $177,-32(%0),%%ymm1;"\ acc_m8n2_con(0,1,4,5,6,7,4,12,%1) acc_m8n2_con(0,1,8,9,10,11,4,12,%1,%%r12,1) #define KERNEL_1_k1m8n6 KERNEL_1_k1m8n4 acc_m8n2_con(0,1,12,13,14,15,0,8,%1,%%r12,2) #define KERNEL_2_k1m8n6 KERNEL_2_k1m8n4 acc_m8n2_con(0,1,12,13,14,15,4,12,%1,%%r12,2) @@ -93,7 +93,6 @@ "movq $10,%5; movq $84,%%r15;"\ #ndim"8881:\n\t"\ "prefetcht1 (%3); subq $63,%3; addq %%r15,%3;"\ - "prefetcht0 64(%1); prefetcht0 64(%1,%%r12,1); prefetcht0 64(%1,%%r12,2);"\ KERNEL_k1m8n##ndim KERNEL_k1m8n##ndim\ "testq $12,%5; movq $84,%%r15; cmovz %4,%%r15; prefetcht1 (%8); addq $16,%8;"\ KERNEL_k1m8n##ndim KERNEL_k1m8n##ndim\ diff --git a/kernel/x86_64/dgemm_kernel_16x2_skylakex.c b/kernel/x86_64/dgemm_kernel_16x2_skylakex.c new file mode 100644 index 000000000..416ace59b --- /dev/null +++ b/kernel/x86_64/dgemm_kernel_16x2_skylakex.c @@ -0,0 +1,492 @@ +/* %0 = a_ptr, %1 = b_ptr, %2 = c_ptr, %3 = c_tmp, %4 = ldc(bytes), %5 = k_counter, %6 = b_pref */ +/* r10 = tmp, r11 = m_counter, r12 = size_of_1_tile_in_b, r13 = k, r14 = b_head, r15 = %1+3*r12 */ + +#if (defined (LEFT) && !defined(TRANSA)) || (!defined (LEFT) && defined(TRANSA)) + #define BACKWARDS 1 +#else + #define BACKWARDS 0 +#endif +#define GEMM_SET_PB "movq %%r14,%1; leaq (%%r14,%%r12,2),%%r15; addq %%r12,%%r15;" +#define set_p_copy1(ptr) "sarq $1,%%r12; addq %%r12,"#ptr"; salq $1,%%r12; salq $3,%%r13; subq %%r13,"#ptr"; sarq $3,%%r13;" +#define set_p_copy2(ptr) "addq %%r12,"#ptr"; salq $4,%%r13; subq %%r13,"#ptr"; sarq $4,%%r13;" +#define set_p_copy4(ptr) "leaq ("#ptr",%%r12,2),"#ptr"; salq $5,%%r13; subq %%r13,"#ptr"; sarq $5,%%r13;" +#define set_p_copy8(ptr) "leaq ("#ptr",%%r12,4),"#ptr"; salq $6,%%r13; subq %%r13,"#ptr"; sarq $6,%%r13;" +#define set_p_copy16(ptr) "leaq ("#ptr",%%r12,8),"#ptr"; salq $7,%%r13; subq %%r13,"#ptr"; sarq $7,%%r13;" +#define set_p_b_dim1(ptr) set_p_copy1(ptr) +#define set_p_b_dim2(ptr) set_p_copy2(ptr) +#define set_p_b_dim4(ptr) set_p_copy2(ptr) +#define set_p_b_dim6(ptr) set_p_copy2(ptr) +#define set_p_b_dim8(ptr) set_p_copy2(ptr) +#define set_p_b_dim10(ptr) set_p_copy2(ptr) +#define set_p_b_dim12(ptr) set_p_copy2(ptr) +#ifdef TRMMKERNEL + #if BACKWARDS == 1 + #define INIT_set_papb(mdim,ndim) GEMM_SET_PB set_p_copy##mdim(%0) set_p_b_dim##ndim(%1) set_p_b_dim##ndim(%%r15) + #define SAVE_set_pa(mdim) "" + #else + #define INIT_set_papb(mdim,ndim) GEMM_SET_PB + #define SAVE_set_pa(mdim) set_p_copy##mdim(%0) + #endif +#else + #define INIT_set_papb(mdim,ndim) GEMM_SET_PB + #define SAVE_set_pa(mdim) "" +#endif +#if defined(TRMMKERNEL) && !defined(LEFT) + #if BACKWARDS == 1 + #define HEAD_SET_OFF(ndim) {} + #define TAIL_SET_OFF(ndim) {off += ndim;} + #define kernel_kstart_n4(mdim,updk) KERNEL_k1m##mdim##n2 KERNEL_k1m##mdim##n2 "addq $32,%%r15; "#updk" $2,%5;" + #define kernel_kstart_n6(mdim,updk) kernel_kstart_n4(mdim,updk) KERNEL_k1m##mdim##n4 KERNEL_k1m##mdim##n4 "addq $32,%%r15; "#updk" $2,%5;" + #define kernel_kstart_n8(mdim,updk) kernel_kstart_n6(mdim,updk) KERNEL_k1m##mdim##n6 KERNEL_k1m##mdim##n6 "addq $32,%%r15; "#updk" $2,%5;" + #define kernel_kstart_n10(mdim,updk) kernel_kstart_n8(mdim,updk) KERNEL_k1m##mdim##n8 KERNEL_k1m##mdim##n8 #updk" $2,%5;" + #define kernel_kstart_n12(mdim,updk) kernel_kstart_n10(mdim,updk) KERNEL_k1m##mdim##n10 KERNEL_k1m##mdim##n10 #updk" $2,%5;" + #define kernel_kend_n4(mdim) "" + #define kernel_kend_n6(mdim) "" + #define kernel_kend_n8(mdim) "" + #define kernel_kend_n10(mdim) "" + #define kernel_kend_n12(mdim) "" + #else + #define HEAD_SET_OFF(ndim) {off += (ndim > 2 ? 2 : ndim);} + #define TAIL_SET_OFF(ndim) {off += (ndim > 2 ? (ndim-2) : 0);} + #define kernel_kstart_n4(mdim,updk) "" + #define kernel_kstart_n6(mdim,updk) "" + #define kernel_kstart_n8(mdim,updk) "" + #define kernel_kstart_n10(mdim,updk) "" + #define kernel_kstart_n12(mdim,updk) "" + #define kernel_kend_n4(mdim) "xorq %3,%3;"\ + loada_kend_k1m##mdim acc_kend_nc2_k1m##mdim(0,8)\ + loada_kend_k1m##mdim acc_kend_nc2_k1m##mdim(16,24) + #define kernel_kend_n6(mdim) "xorq %3,%3;"\ + loada_kend_k1m##mdim acc_kend_nc2_k1m##mdim(0,8) acc_kend_nc3_k1m##mdim(0,8)\ + loada_kend_k1m##mdim acc_kend_nc2_k1m##mdim(16,24) acc_kend_nc3_k1m##mdim(16,24)\ + loada_kend_k1m##mdim acc_kend_nc3_k1m##mdim(32,40)\ + loada_kend_k1m##mdim acc_kend_nc3_k1m##mdim(48,56) + #define kernel_kend_n8(mdim) "xorq %3,%3;"\ + loada_kend_k1m##mdim acc_kend_nc2_k1m##mdim(0,8) acc_kend_nc3_k1m##mdim(0,8) acc_kend_nc4_k1m##mdim(0,8)\ + loada_kend_k1m##mdim acc_kend_nc2_k1m##mdim(16,24) acc_kend_nc3_k1m##mdim(16,24) acc_kend_nc4_k1m##mdim(16,24)\ + loada_kend_k1m##mdim acc_kend_nc3_k1m##mdim(32,40) acc_kend_nc4_k1m##mdim(32,40)\ + loada_kend_k1m##mdim acc_kend_nc3_k1m##mdim(48,56) acc_kend_nc4_k1m##mdim(48,56)\ + loada_kend_k1m##mdim acc_kend_nc4_k1m##mdim(64,72)\ + loada_kend_k1m##mdim acc_kend_nc4_k1m##mdim(80,88) + #define kernel_kend_n10(mdim) "xorq %3,%3;"\ + loada_kend_k1m##mdim acc_kend_nc2_k1m##mdim(0,8) acc_kend_nc3_k1m##mdim(0,8) acc_kend_nc4_k1m##mdim(0,8) acc_kend_nc5_k1m##mdim(0,8)\ + loada_kend_k1m##mdim acc_kend_nc2_k1m##mdim(16,24) acc_kend_nc3_k1m##mdim(16,24) acc_kend_nc4_k1m##mdim(16,24) acc_kend_nc5_k1m##mdim(16,24)\ + loada_kend_k1m##mdim acc_kend_nc3_k1m##mdim(32,40) acc_kend_nc4_k1m##mdim(32,40) acc_kend_nc5_k1m##mdim(32,40)\ + loada_kend_k1m##mdim acc_kend_nc3_k1m##mdim(48,56) acc_kend_nc4_k1m##mdim(48,56) acc_kend_nc5_k1m##mdim(48,56)\ + loada_kend_k1m##mdim acc_kend_nc4_k1m##mdim(64,72) acc_kend_nc5_k1m##mdim(64,72)\ + loada_kend_k1m##mdim acc_kend_nc4_k1m##mdim(80,88) acc_kend_nc5_k1m##mdim(80,88)\ + loada_kend_k1m##mdim acc_kend_nc5_k1m##mdim(96,104)\ + loada_kend_k1m##mdim acc_kend_nc5_k1m##mdim(112,120) + #define kernel_kend_n12(mdim) "xorq %3,%3;"\ + loada_kend_k1m##mdim acc_kend_nc2_k1m##mdim(0,8) acc_kend_nc3_k1m##mdim(0,8) acc_kend_nc4_k1m##mdim(0,8) acc_kend_nc5_k1m##mdim(0,8) acc_kend_nc6_k1m##mdim(0,8)\ + loada_kend_k1m##mdim acc_kend_nc2_k1m##mdim(16,24) acc_kend_nc3_k1m##mdim(16,24) acc_kend_nc4_k1m##mdim(16,24) acc_kend_nc5_k1m##mdim(16,24) acc_kend_nc6_k1m##mdim(16,24)\ + loada_kend_k1m##mdim acc_kend_nc3_k1m##mdim(32,40) acc_kend_nc4_k1m##mdim(32,40) acc_kend_nc5_k1m##mdim(32,40) acc_kend_nc6_k1m##mdim(32,40)\ + loada_kend_k1m##mdim acc_kend_nc3_k1m##mdim(48,56) acc_kend_nc4_k1m##mdim(48,56) acc_kend_nc5_k1m##mdim(48,56) acc_kend_nc6_k1m##mdim(48,56)\ + loada_kend_k1m##mdim acc_kend_nc4_k1m##mdim(64,72) acc_kend_nc5_k1m##mdim(64,72) acc_kend_nc6_k1m##mdim(64,72)\ + loada_kend_k1m##mdim acc_kend_nc4_k1m##mdim(80,88) acc_kend_nc5_k1m##mdim(80,88) acc_kend_nc6_k1m##mdim(80,88)\ + loada_kend_k1m##mdim acc_kend_nc5_k1m##mdim(96,104) acc_kend_nc6_k1m##mdim(96,104)\ + loada_kend_k1m##mdim acc_kend_nc5_k1m##mdim(112,120) acc_kend_nc6_k1m##mdim(112,120)\ + loada_kend_k1m##mdim acc_kend_nc6_k1m##mdim(128,136)\ + loada_kend_k1m##mdim acc_kend_nc6_k1m##mdim(144,152) + #endif +#else + #define HEAD_SET_OFF(ndim) {} + #define TAIL_SET_OFF(ndim) {} + #define kernel_kstart_n4(mdim,updk) "" + #define kernel_kstart_n6(mdim,updk) "" + #define kernel_kstart_n8(mdim,updk) "" + #define kernel_kstart_n10(mdim,updk) "" + #define kernel_kstart_n12(mdim,updk) "" + #define kernel_kend_n4(mdim) "" + #define kernel_kend_n6(mdim) "" + #define kernel_kend_n8(mdim) "" + #define kernel_kend_n10(mdim) "" + #define kernel_kend_n12(mdim) "" +#endif +#define kernel_kstart_n1(mdim,updk) "" +#define kernel_kstart_n2(mdim,updk) "" +#define kernel_kend_n1(mdim) "" +#define kernel_kend_n2(mdim) "" + +#ifdef TRMMKERNEL + #if BACKWARDS == 1 + #define INITASM_SET_K "movq %10,%%r13; subq %9,%%r13;" + #else + #define INITASM_SET_K "movq %9,%%r13;" + #endif +#else + #define INITASM_SET_K "movq %10,%%r13;" +#endif +#if defined(TRMMKERNEL) && defined(LEFT) + #if BACKWARDS==1 + #define init_update_k(mdim) "" + #define save_update_k(mdim) "subq $"#mdim",%%r13;" + #else + #define init_update_k(mdim) "addq $"#mdim",%%r13;" + #define save_update_k(mdim) "" + #endif +#else + #define init_update_k(mdim) "" + #define save_update_k(mdim) "" +#endif + +#define KERNEL_h_k1m16n1 \ + "vmovupd (%0),%%zmm1; vmovupd 64(%0),%%zmm2; addq $128,%0;"\ + "vbroadcastsd (%1),%%zmm3; vfmadd231pd %%zmm1,%%zmm3,%%zmm8; vfmadd231pd %%zmm2,%%zmm3,%%zmm9;" +#define KERNEL_k1m16n1 KERNEL_h_k1m16n1 "addq $8,%1;" +#define KERNEL_h_k1m16n2 KERNEL_h_k1m16n1\ + "vbroadcastsd 8(%1),%%zmm4; vfmadd231pd %%zmm1,%%zmm4,%%zmm10; vfmadd231pd %%zmm2,%%zmm4,%%zmm11;" +#define KERNEL_k1m16n2 KERNEL_h_k1m16n2 "addq $16,%1;" +#define unit_acc_gen_m16n2(c1_no,c2_no,c3_no,c4_no,boff1,boff2,...)\ + "vbroadcastsd "#boff1"("#__VA_ARGS__"),%%zmm3; vfmadd231pd %%zmm1,%%zmm3,%%zmm"#c1_no"; vfmadd231pd %%zmm2,%%zmm3,%%zmm"#c2_no";"\ + "vbroadcastsd "#boff2"("#__VA_ARGS__"),%%zmm4; vfmadd231pd %%zmm1,%%zmm4,%%zmm"#c3_no"; vfmadd231pd %%zmm2,%%zmm4,%%zmm"#c4_no";" +#define unit_acc_m16n2(c1_no,c2_no,c3_no,c4_no,...) unit_acc_gen_m16n2(c1_no,c2_no,c3_no,c4_no,0,8,__VA_ARGS__) +#define KERNEL_h_k1m16n4 KERNEL_h_k1m16n2 "prefetcht0 384(%0);" unit_acc_m16n2(12,13,14,15,%1,%%r12,1) +#define KERNEL_k1m16n4 KERNEL_h_k1m16n4 "addq $16,%1;" +#define KERNEL_k1m16n6 KERNEL_h_k1m16n4 unit_acc_m16n2(16,17,18,19,%1,%%r12,2) "addq $16,%1;" +#define KERNEL_h_k1m16n8 KERNEL_k1m16n6 "prefetcht0 448(%0);" unit_acc_m16n2(20,21,22,23,%%r15) +#define KERNEL_k1m16n8 KERNEL_h_k1m16n8 "addq $16,%%r15;" +#define KERNEL_h_k1m16n10 KERNEL_h_k1m16n8 unit_acc_m16n2(24,25,26,27,%%r15,%%r12,1) +#define KERNEL_k1m16n10 KERNEL_h_k1m16n10 "addq $16,%%r15;" +#define KERNEL_h_k1m16n12 KERNEL_h_k1m16n10 unit_acc_m16n2(28,29,30,31,%%r15,%%r12,2) +#define KERNEL_k1m16n12 KERNEL_h_k1m16n12 "addq $16,%%r15;" +#if defined(TRMMKERNEL) && !defined(LEFT) && (BACKWARDS == 0) + #define loada_kend_k1m16 "vmovupd (%0,%3,1),%%zmm1; vmovupd 64(%0,%3,1),%%zmm2; addq $128,%3;" + #define acc_kend_nc2_k1m16(boff1,boff2) unit_acc_gen_m16n2(12,13,14,15,boff1,boff2,%1,%%r12,1) + #define acc_kend_nc3_k1m16(boff1,boff2) unit_acc_gen_m16n2(16,17,18,19,boff1,boff2,%1,%%r12,2) + #define acc_kend_nc4_k1m16(boff1,boff2) unit_acc_gen_m16n2(20,21,22,23,boff1,boff2,%%r15) + #define acc_kend_nc5_k1m16(boff1,boff2) unit_acc_gen_m16n2(24,25,26,27,boff1,boff2,%%r15,%%r12,1) + #define acc_kend_nc6_k1m16(boff1,boff2) unit_acc_gen_m16n2(28,29,30,31,boff1,boff2,%%r15,%%r12,2) +#endif +#define save_init_m16 "movq %2,%3; addq $128,%2;" +#ifdef TRMMKERNEL + #define SAVE_m16n1 "vmulpd %%zmm8,%%zmm0,%%zmm8; vmovupd %%zmm8,(%2); vmulpd %%zmm9,%%zmm0,%%zmm9; vmovupd %%zmm9,64(%2); addq $128,%2;" + #define unit_save_m16n2(c1_no,c2_no,c3_no,c4_no)\ + "vmulpd %%zmm"#c1_no",%%zmm0,%%zmm"#c1_no"; vmovupd %%zmm"#c1_no",(%3); vmulpd %%zmm"#c2_no",%%zmm0,%%zmm"#c2_no"; vmovupd %%zmm"#c2_no",64(%3);"\ + "vmulpd %%zmm"#c3_no",%%zmm0,%%zmm"#c3_no"; vmovupd %%zmm"#c3_no",(%3,%4,1); vmulpd %%zmm"#c4_no",%%zmm0,%%zmm"#c4_no"; vmovupd %%zmm"#c4_no",64(%3,%4,1); leaq (%3,%4,2),%3;" +#else + #define SAVE_m16n1 "vfmadd213pd (%2),%%zmm0,%%zmm8; vmovupd %%zmm8,(%2); vfmadd213pd 64(%2),%%zmm0,%%zmm9; vmovupd %%zmm9,64(%2); addq $128,%2;" + #define unit_save_m16n2(c1_no,c2_no,c3_no,c4_no)\ + "vfmadd213pd (%3),%%zmm0,%%zmm"#c1_no"; vmovupd %%zmm"#c1_no",(%3); vfmadd213pd 64(%3),%%zmm0,%%zmm"#c2_no"; vmovupd %%zmm"#c2_no",64(%3);"\ + "vfmadd213pd (%3,%4,1),%%zmm0,%%zmm"#c3_no"; vmovupd %%zmm"#c3_no",(%3,%4,1); vfmadd213pd 64(%3,%4,1),%%zmm0,%%zmm"#c4_no"; vmovupd %%zmm"#c4_no",64(%3,%4,1); leaq (%3,%4,2),%3;" +#endif +#define SAVE_m16n2 save_init_m16 unit_save_m16n2(8,9,10,11) +#define SAVE_m16n4 SAVE_m16n2 unit_save_m16n2(12,13,14,15) +#define SAVE_m16n6 SAVE_m16n4 unit_save_m16n2(16,17,18,19) +#define SAVE_m16n8 SAVE_m16n6 unit_save_m16n2(20,21,22,23) +#define SAVE_m16n10 SAVE_m16n8 unit_save_m16n2(24,25,26,27) +#define SAVE_m16n12 SAVE_m16n10 unit_save_m16n2(28,29,30,31) +#define unit_init_2zmm(c1_no,c2_no) "vpxorq %%zmm"#c1_no",%%zmm"#c1_no",%%zmm"#c1_no"; vpxorq %%zmm"#c2_no",%%zmm"#c2_no",%%zmm"#c2_no";" +#define unit_init_4zmm(c1_no,c2_no,c3_no,c4_no) unit_init_2zmm(c1_no,c2_no) unit_init_2zmm(c3_no,c4_no) +#define INIT_m16n1 unit_init_2zmm(8,9) +#define INIT_m16n2 unit_init_4zmm(8,9,10,11) +#define INIT_m16n4 INIT_m16n2 unit_init_4zmm(12,13,14,15) +#define INIT_m16n6 INIT_m16n4 unit_init_4zmm(16,17,18,19) +#define INIT_m16n8 INIT_m16n6 unit_init_4zmm(20,21,22,23) +#define INIT_m16n10 INIT_m16n8 unit_init_4zmm(24,25,26,27) +#define INIT_m16n12 INIT_m16n10 unit_init_4zmm(28,29,30,31) + +#define KERNEL_k1m8n1 \ + "vbroadcastsd (%1),%%zmm1; addq $8,%1;"\ + "vfmadd231pd (%0),%%zmm1,%%zmm8; addq $64,%0;" +#define unit_acc_gen_m8n2(c1_no,c2_no,boff,...)\ + "vbroadcastf32x4 "#boff"("#__VA_ARGS__"),%%zmm3; vfmadd231pd %%zmm1,%%zmm3,%%zmm"#c1_no"; vfmadd231pd %%zmm2,%%zmm3,%%zmm"#c2_no";" +#define unit_acc_m8n2(c1_no,c2_no,...) unit_acc_gen_m8n2(c1_no,c2_no,0,__VA_ARGS__) +#define KERNEL_h_k1m8n2 \ + "vmovddup (%0),%%zmm1; vmovddup 8(%0),%%zmm2; addq $64,%0;" unit_acc_m8n2(8,9,%1) +#define KERNEL_k1m8n2 KERNEL_h_k1m8n2 "addq $16,%1;" +#define KERNEL_h_k1m8n4 KERNEL_h_k1m8n2 unit_acc_m8n2(10,11,%1,%%r12,1) +#define KERNEL_k1m8n4 KERNEL_h_k1m8n4 "addq $16,%1;" +#define KERNEL_k1m8n6 KERNEL_h_k1m8n4 unit_acc_m8n2(12,13,%1,%%r12,2) "addq $16,%1;" +#define KERNEL_h_k1m8n8 KERNEL_k1m8n6 unit_acc_m8n2(14,15,%%r15) +#define KERNEL_k1m8n8 KERNEL_h_k1m8n8 "addq $16,%%r15;" +#define KERNEL_h_k1m8n10 KERNEL_h_k1m8n8 unit_acc_m8n2(16,17,%%r15,%%r12,1) +#define KERNEL_k1m8n10 KERNEL_h_k1m8n10 "addq $16,%%r15;" +#define KERNEL_h_k1m8n12 KERNEL_h_k1m8n10 unit_acc_m8n2(18,19,%%r15,%%r12,2) +#define KERNEL_k1m8n12 KERNEL_h_k1m8n12 "addq $16,%%r15;" +#if defined(TRMMKERNEL) && !defined(LEFT) && (BACKWARDS == 0) + #define loada_kend_k1m8 "vmovddup (%0,%3,1),%%zmm1; vmovddup 8(%0,%3,1),%%zmm2; addq $64,%3;" + #define acc_kend_nc2_k1m8(boff1,boff2) unit_acc_gen_m8n2(10,11,boff1,%1,%%r12,1) + #define acc_kend_nc3_k1m8(boff1,boff2) unit_acc_gen_m8n2(12,13,boff1,%1,%%r12,2) + #define acc_kend_nc4_k1m8(boff1,boff2) unit_acc_gen_m8n2(14,15,boff1,%%r15) + #define acc_kend_nc5_k1m8(boff1,boff2) unit_acc_gen_m8n2(16,17,boff1,%%r15,%%r12,1) + #define acc_kend_nc6_k1m8(boff1,boff2) unit_acc_gen_m8n2(18,19,boff1,%%r15,%%r12,2) +#endif +#define save_init_m8 "movq %2,%3; addq $64,%2;" +#ifdef TRMMKERNEL + #define SAVE_m8n1 "vmulpd %%zmm8,%%zmm0,%%zmm8; vmovupd %%zmm8,(%2); addq $64,%2;" + #define unit_save_m8n2(c1_no,c2_no)\ + "vunpcklpd %%zmm"#c2_no",%%zmm"#c1_no",%%zmm1; vmulpd %%zmm1,%%zmm0,%%zmm1; vmovupd %%zmm1,(%3);"\ + "vunpckhpd %%zmm"#c2_no",%%zmm"#c1_no",%%zmm2; vmulpd %%zmm2,%%zmm0,%%zmm2; vmovupd %%zmm2,(%3,%4,1); leaq (%3,%4,2),%3;" +#else + #define SAVE_m8n1 "vfmadd213pd (%2),%%zmm0,%%zmm8; vmovupd %%zmm8,(%2); addq $64,%2;" + #define unit_save_m8n2(c1_no,c2_no)\ + "vunpcklpd %%zmm"#c2_no",%%zmm"#c1_no",%%zmm1; vfmadd213pd (%3),%%zmm0,%%zmm1; vmovupd %%zmm1,(%3);"\ + "vunpckhpd %%zmm"#c2_no",%%zmm"#c1_no",%%zmm2; vfmadd213pd (%3,%4,1),%%zmm0,%%zmm2; vmovupd %%zmm2,(%3,%4,1); leaq (%3,%4,2),%3;" +#endif +#define SAVE_m8n2 save_init_m8 unit_save_m8n2(8,9) +#define SAVE_m8n4 SAVE_m8n2 unit_save_m8n2(10,11) +#define SAVE_m8n6 SAVE_m8n4 unit_save_m8n2(12,13) +#define SAVE_m8n8 SAVE_m8n6 unit_save_m8n2(14,15) +#define SAVE_m8n10 SAVE_m8n8 unit_save_m8n2(16,17) +#define SAVE_m8n12 SAVE_m8n10 unit_save_m8n2(18,19) +#define INIT_m8n1 "vpxorq %%zmm8,%%zmm8,%%zmm8;" +#define INIT_m8n2 unit_init_2zmm(8,9) +#define INIT_m8n4 INIT_m8n2 unit_init_2zmm(10,11) +#define INIT_m8n6 INIT_m8n4 unit_init_2zmm(12,13) +#define INIT_m8n8 INIT_m8n6 unit_init_2zmm(14,15) +#define INIT_m8n10 INIT_m8n8 unit_init_2zmm(16,17) +#define INIT_m8n12 INIT_m8n10 unit_init_2zmm(18,19) + +#define KERNEL_k1m4n1 \ + "vbroadcastsd (%1),%%ymm1; addq $8,%1;"\ + "vfmadd231pd (%0),%%ymm1,%%ymm4; addq $32,%0;" +#define unit_acc_gen_m4n2(c1_no,c2_no,boff,...)\ + "vbroadcastf128 "#boff"("#__VA_ARGS__"),%%ymm3; vfmadd231pd %%ymm1,%%ymm3,%%ymm"#c1_no"; vfmadd231pd %%ymm2,%%ymm3,%%ymm"#c2_no";" +#define unit_acc_m4n2(c1_no,c2_no,...) unit_acc_gen_m4n2(c1_no,c2_no,0,__VA_ARGS__) +#define KERNEL_h_k1m4n2 \ + "vmovddup (%0),%%ymm1; vmovddup 8(%0),%%ymm2; addq $32,%0;" unit_acc_m4n2(4,5,%1) +#define KERNEL_k1m4n2 KERNEL_h_k1m4n2 "addq $16,%1;" +#define KERNEL_h_k1m4n4 KERNEL_h_k1m4n2 unit_acc_m4n2(6,7,%1,%%r12,1) +#define KERNEL_k1m4n4 KERNEL_h_k1m4n4 "addq $16,%1;" +#define KERNEL_k1m4n6 KERNEL_h_k1m4n4 unit_acc_m4n2(8,9,%1,%%r12,2) "addq $16,%1;" +#define KERNEL_h_k1m4n8 KERNEL_k1m4n6 unit_acc_m4n2(10,11,%%r15) +#define KERNEL_k1m4n8 KERNEL_h_k1m4n8 "addq $16,%%r15;" +#define KERNEL_h_k1m4n10 KERNEL_h_k1m4n8 unit_acc_m4n2(12,13,%%r15,%%r12,1) +#define KERNEL_k1m4n10 KERNEL_h_k1m4n10 "addq $16,%%r15;" +#define KERNEL_h_k1m4n12 KERNEL_h_k1m4n10 unit_acc_m4n2(14,15,%%r15,%%r12,2) +#define KERNEL_k1m4n12 KERNEL_h_k1m4n12 "addq $16,%%r15;" +#if defined(TRMMKERNEL) && !defined(LEFT) && (BACKWARDS == 0) + #define loada_kend_k1m4 "vmovddup (%0,%3,1),%%ymm1; vmovddup 8(%0,%3,1),%%ymm2; addq $32,%3;" + #define acc_kend_nc2_k1m4(boff1,boff2) unit_acc_gen_m4n2(6,7,boff1,%1,%%r12,1) + #define acc_kend_nc3_k1m4(boff1,boff2) unit_acc_gen_m4n2(8,9,boff1,%1,%%r12,2) + #define acc_kend_nc4_k1m4(boff1,boff2) unit_acc_gen_m4n2(10,11,boff1,%%r15) + #define acc_kend_nc5_k1m4(boff1,boff2) unit_acc_gen_m4n2(12,13,boff1,%%r15,%%r12,1) + #define acc_kend_nc6_k1m4(boff1,boff2) unit_acc_gen_m4n2(14,15,boff1,%%r15,%%r12,2) +#endif +#define save_init_m4 "movq %2,%3; addq $32,%2;" +#ifdef TRMMKERNEL + #define SAVE_m4n1 "vmulpd %%ymm4,%%ymm0,%%ymm4; vmovupd %%ymm4,(%2); addq $32,%2;" + #define unit_save_m4n2(c1_no,c2_no)\ + "vunpcklpd %%ymm"#c2_no",%%ymm"#c1_no",%%ymm1; vmulpd %%ymm1,%%ymm0,%%ymm1; vmovupd %%ymm1,(%3);"\ + "vunpckhpd %%ymm"#c2_no",%%ymm"#c1_no",%%ymm2; vmulpd %%ymm2,%%ymm0,%%ymm2; vmovupd %%ymm2,(%3,%4,1); leaq (%3,%4,2),%3;" +#else + #define SAVE_m4n1 "vfmadd213pd (%2),%%ymm0,%%ymm4; vmovupd %%ymm4,(%2); addq $32,%2;" + #define unit_save_m4n2(c1_no,c2_no)\ + "vunpcklpd %%ymm"#c2_no",%%ymm"#c1_no",%%ymm1; vfmadd213pd (%3),%%ymm0,%%ymm1; vmovupd %%ymm1,(%3);"\ + "vunpckhpd %%ymm"#c2_no",%%ymm"#c1_no",%%ymm2; vfmadd213pd (%3,%4,1),%%ymm0,%%ymm2; vmovupd %%ymm2,(%3,%4,1); leaq (%3,%4,2),%3;" +#endif +#define SAVE_m4n2 save_init_m4 unit_save_m4n2(4,5) +#define SAVE_m4n4 SAVE_m4n2 unit_save_m4n2(6,7) +#define SAVE_m4n6 SAVE_m4n4 unit_save_m4n2(8,9) +#define SAVE_m4n8 SAVE_m4n6 unit_save_m4n2(10,11) +#define SAVE_m4n10 SAVE_m4n8 unit_save_m4n2(12,13) +#define SAVE_m4n12 SAVE_m4n10 unit_save_m4n2(14,15) +#define INIT_m4n1 "vpxor %%ymm4,%%ymm4,%%ymm4;" +#define unit_init_2ymm(c1_no,c2_no) "vpxor %%ymm"#c1_no",%%ymm"#c1_no",%%ymm"#c1_no"; vpxor %%ymm"#c2_no",%%ymm"#c2_no",%%ymm"#c2_no";" +#define INIT_m4n2 unit_init_2ymm(4,5) +#define INIT_m4n4 INIT_m4n2 unit_init_2ymm(6,7) +#define INIT_m4n6 INIT_m4n4 unit_init_2ymm(8,9) +#define INIT_m4n8 INIT_m4n6 unit_init_2ymm(10,11) +#define INIT_m4n10 INIT_m4n8 unit_init_2ymm(12,13) +#define INIT_m4n12 INIT_m4n10 unit_init_2ymm(14,15) + +#define KERNEL_k1m2n1 \ + "vmovddup (%1),%%xmm1; addq $8,%1;"\ + "vfmadd231pd (%0),%%xmm1,%%xmm4; addq $16,%0;" +#define unit_acc_gen_m2n2(c1_no,c2_no,boff,...)\ + "vmovupd "#boff"("#__VA_ARGS__"),%%xmm3; vfmadd231pd %%xmm1,%%xmm3,%%xmm"#c1_no"; vfmadd231pd %%xmm2,%%xmm3,%%xmm"#c2_no";" +#define unit_acc_m2n2(c1_no,c2_no,...) unit_acc_gen_m2n2(c1_no,c2_no,0,__VA_ARGS__) +#define KERNEL_h_k1m2n2 \ + "vmovddup (%0),%%xmm1; vmovddup 8(%0),%%xmm2; addq $16,%0;" unit_acc_m2n2(4,5,%1) +#define KERNEL_k1m2n2 KERNEL_h_k1m2n2 "addq $16,%1;" +#define KERNEL_h_k1m2n4 KERNEL_h_k1m2n2 unit_acc_m2n2(6,7,%1,%%r12,1) +#define KERNEL_k1m2n4 KERNEL_h_k1m2n4 "addq $16,%1;" +#define KERNEL_k1m2n6 KERNEL_h_k1m2n4 unit_acc_m2n2(8,9,%1,%%r12,2) "addq $16,%1;" +#define KERNEL_h_k1m2n8 KERNEL_k1m2n6 unit_acc_m2n2(10,11,%%r15) +#define KERNEL_k1m2n8 KERNEL_h_k1m2n8 "addq $16,%%r15;" +#define KERNEL_h_k1m2n10 KERNEL_h_k1m2n8 unit_acc_m2n2(12,13,%%r15,%%r12,1) +#define KERNEL_k1m2n10 KERNEL_h_k1m2n10 "addq $16,%%r15;" +#define KERNEL_h_k1m2n12 KERNEL_h_k1m2n10 unit_acc_m2n2(14,15,%%r15,%%r12,2) +#define KERNEL_k1m2n12 KERNEL_h_k1m2n12 "addq $16,%%r15;" +#if defined(TRMMKERNEL) && !defined(LEFT) && (BACKWARDS == 0) + #define loada_kend_k1m2 "vmovddup (%0,%3,1),%%xmm1; vmovddup 8(%0,%3,1),%%xmm2; addq $16,%3;" + #define acc_kend_nc2_k1m2(boff1,boff2) unit_acc_gen_m2n2(6,7,boff1,%1,%%r12,1) + #define acc_kend_nc3_k1m2(boff1,boff2) unit_acc_gen_m2n2(8,9,boff1,%1,%%r12,2) + #define acc_kend_nc4_k1m2(boff1,boff2) unit_acc_gen_m2n2(10,11,boff1,%%r15) + #define acc_kend_nc5_k1m2(boff1,boff2) unit_acc_gen_m2n2(12,13,boff1,%%r15,%%r12,1) + #define acc_kend_nc6_k1m2(boff1,boff2) unit_acc_gen_m2n2(14,15,boff1,%%r15,%%r12,2) +#endif +#define save_init_m2 "movq %2,%3; addq $16,%2;" +#ifdef TRMMKERNEL + #define SAVE_m2n1 "vmulpd %%xmm4,%%xmm0,%%xmm4; vmovupd %%xmm4,(%2); addq $16,%2;" + #define unit_save_m2n2(c1_no,c2_no)\ + "vunpcklpd %%xmm"#c2_no",%%xmm"#c1_no",%%xmm1; vmulpd %%xmm1,%%xmm0,%%xmm1; vmovupd %%xmm1,(%3);"\ + "vunpckhpd %%xmm"#c2_no",%%xmm"#c1_no",%%xmm2; vmulpd %%xmm2,%%xmm0,%%xmm2; vmovupd %%xmm2,(%3,%4,1); leaq (%3,%4,2),%3;" +#else + #define SAVE_m2n1 "vfmadd213pd (%2),%%xmm0,%%xmm4; vmovupd %%xmm4,(%2); addq $16,%2;" + #define unit_save_m2n2(c1_no,c2_no)\ + "vunpcklpd %%xmm"#c2_no",%%xmm"#c1_no",%%xmm1; vfmadd213pd (%3),%%xmm0,%%xmm1; vmovupd %%xmm1,(%3);"\ + "vunpckhpd %%xmm"#c2_no",%%xmm"#c1_no",%%xmm2; vfmadd213pd (%3,%4,1),%%xmm0,%%xmm2; vmovupd %%xmm2,(%3,%4,1); leaq (%3,%4,2),%3;" +#endif +#define SAVE_m2n2 save_init_m2 unit_save_m2n2(4,5) +#define SAVE_m2n4 SAVE_m2n2 unit_save_m2n2(6,7) +#define SAVE_m2n6 SAVE_m2n4 unit_save_m2n2(8,9) +#define SAVE_m2n8 SAVE_m2n6 unit_save_m2n2(10,11) +#define SAVE_m2n10 SAVE_m2n8 unit_save_m2n2(12,13) +#define SAVE_m2n12 SAVE_m2n10 unit_save_m2n2(14,15) +#define INIT_m2n1 "vpxor %%xmm4,%%xmm4,%%xmm4;" +#define unit_init_2xmm(c1_no,c2_no) "vpxor %%xmm"#c1_no",%%xmm"#c1_no",%%xmm"#c1_no"; vpxor %%xmm"#c2_no",%%xmm"#c2_no",%%xmm"#c2_no";" +#define INIT_m2n2 unit_init_2xmm(4,5) +#define INIT_m2n4 INIT_m2n2 unit_init_2xmm(6,7) +#define INIT_m2n6 INIT_m2n4 unit_init_2xmm(8,9) +#define INIT_m2n8 INIT_m2n6 unit_init_2xmm(10,11) +#define INIT_m2n10 INIT_m2n8 unit_init_2xmm(12,13) +#define INIT_m2n12 INIT_m2n10 unit_init_2xmm(14,15) + +#define KERNEL_k1m1n1 \ + "vmovsd (%1),%%xmm1; addq $8,%1;"\ + "vfmadd231sd (%0),%%xmm1,%%xmm4; addq $8,%0;" +#define KERNEL_h_k1m1n2 \ + "vmovddup (%0),%%xmm1; addq $8,%0;"\ + "vfmadd231pd (%1),%%xmm1,%%xmm4;" +#define KERNEL_k1m1n2 KERNEL_h_k1m1n2 "addq $16,%1;" +#define KERNEL_h_k1m1n4 KERNEL_h_k1m1n2 "vfmadd231pd (%1,%%r12,1),%%xmm1,%%xmm5;" +#define KERNEL_k1m1n4 KERNEL_h_k1m1n4 "addq $16,%1;" +#define KERNEL_k1m1n6 KERNEL_h_k1m1n4 "vfmadd231pd (%1,%%r12,2),%%xmm1,%%xmm6; addq $16,%1;" +#define KERNEL_h_k1m1n8 KERNEL_k1m1n6 "vfmadd231pd (%%r15),%%xmm1,%%xmm7;" +#define KERNEL_k1m1n8 KERNEL_h_k1m1n8 "addq $16,%%r15;" +#define KERNEL_h_k1m1n10 KERNEL_h_k1m1n8 "vfmadd231pd (%%r15,%%r12,1),%%xmm1,%%xmm8;" +#define KERNEL_k1m1n10 KERNEL_h_k1m1n10 "addq $16,%%r15;" +#define KERNEL_h_k1m1n12 KERNEL_h_k1m1n10 "vfmadd231pd (%%r15,%%r12,2),%%xmm1,%%xmm9;" +#define KERNEL_k1m1n12 KERNEL_h_k1m1n12 "addq $16,%%r15;" +#if defined(TRMMKERNEL) && !defined(LEFT) && (BACKWARDS == 0) + #define loada_kend_k1m1 "vmovddup (%0,%3,1),%%xmm1; addq $8,%3;" + #define acc_kend_nc2_k1m1(boff1,boff2) "vfmadd231pd "#boff1"(%1,%%r12,1),%%xmm1,%%xmm5;" + #define acc_kend_nc3_k1m1(boff1,boff2) "vfmadd231pd "#boff1"(%1,%%r12,2),%%xmm1,%%xmm6;" + #define acc_kend_nc4_k1m1(boff1,boff2) "vfmadd231pd "#boff1"(%%r15),%%xmm1,%%xmm7;" + #define acc_kend_nc5_k1m1(boff1,boff2) "vfmadd231pd "#boff1"(%%r15,%%r12,1),%%xmm1,%%xmm8;" + #define acc_kend_nc6_k1m1(boff1,boff2) "vfmadd231pd "#boff1"(%%r15,%%r12,2),%%xmm1,%%xmm9;" +#endif +#define save_init_m1 "movq %2,%3; addq $8,%2;" +#ifdef TRMMKERNEL + #define SAVE_m1n1 "vmulsd %%xmm4,%%xmm0,%%xmm4; vmovsd %%xmm4,(%2); addq $8,%2;" + #define unit_save_m1n2(c1_no)\ + "vmulpd %%xmm"#c1_no",%%xmm0,%%xmm2; vmovsd %%xmm2,(%3); vmovhpd %%xmm2,(%3,%4,1); leaq (%3,%4,2),%3;" +#else + #define SAVE_m1n1 "vfmadd213sd (%2),%%xmm0,%%xmm4; vmovsd %%xmm4,(%2); addq $8,%2;" + #define unit_save_m1n2(c1_no)\ + "vmovsd (%3),%%xmm2; vmovhpd (%3,%4,1),%%xmm2,%%xmm2; vfmadd231pd %%xmm"#c1_no",%%xmm0,%%xmm2; vmovsd %%xmm2,(%3); vmovhpd %%xmm2,(%3,%4,1); leaq (%3,%4,2),%3;" +#endif +#define SAVE_m1n2 save_init_m1 unit_save_m1n2(4) +#define SAVE_m1n4 SAVE_m1n2 unit_save_m1n2(5) +#define SAVE_m1n6 SAVE_m1n4 unit_save_m1n2(6) +#define SAVE_m1n8 SAVE_m1n6 unit_save_m1n2(7) +#define SAVE_m1n10 SAVE_m1n8 unit_save_m1n2(8) +#define SAVE_m1n12 SAVE_m1n10 unit_save_m1n2(9) +#define INIT_m1n1 "vpxor %%xmm4,%%xmm4,%%xmm4;" +#define INIT_m1n2 INIT_m1n1 +#define INIT_m1n4 INIT_m1n2 "vpxor %%xmm5,%%xmm5,%%xmm5;" +#define INIT_m1n6 INIT_m1n4 "vpxor %%xmm6,%%xmm6,%%xmm6;" +#define INIT_m1n8 INIT_m1n6 "vpxor %%xmm7,%%xmm7,%%xmm7;" +#define INIT_m1n10 INIT_m1n8 "vpxor %%xmm8,%%xmm8,%%xmm8;" +#define INIT_m1n12 INIT_m1n10 "vpxor %%xmm9,%%xmm9,%%xmm9;" + +#define COMPUTE_SIMPLE(mdim,ndim)\ + init_update_k(mdim) INIT_m##mdim##n##ndim\ + "movq %%r13,%5;" INIT_set_papb(mdim,ndim)\ + kernel_kstart_n##ndim(mdim,subq)\ + "testq %5,%5; jz 7"#mdim"7"#ndim"9f;"\ + "7"#mdim"7"#ndim"1:\n\t"\ + KERNEL_k1m##mdim##n##ndim "decq %5; jnz 7"#mdim"7"#ndim"1b;"\ + "7"#mdim"7"#ndim"9:\n\t"\ + kernel_kend_n##ndim(mdim)\ + SAVE_set_pa(mdim) SAVE_m##mdim##n##ndim save_update_k(mdim) +#define COMPUTE_m16n1 COMPUTE_SIMPLE(16,1) +#define COMPUTE_m16n2 COMPUTE_SIMPLE(16,2) +#define COMPUTE_m16n4 COMPUTE_SIMPLE(16,4) +#define COMPUTE_m16n6 COMPUTE_SIMPLE(16,6) +#define COMPUTE_m16n8 COMPUTE_SIMPLE(16,8) +#define COMPUTE_m16n10 COMPUTE_SIMPLE(16,10) +#if defined(TRMMKERNEL) && !defined(LEFT) && defined(TRANSA) + #define INVERSE_K_MID "negq %5; leaq 6(%%r13,%5,1),%5;" +#else + #define INVERSE_K_MID "negq %5; leaq 16(%%r13,%5,1),%5;" +#endif +#define COMPUTE_m16n12 \ + init_update_k(16) INIT_m16n12 "movq %%r13,%5;" INIT_set_papb(16,12) "movq %2,%3;"\ + kernel_kstart_n12(16,subq)\ + "cmpq $16,%5; jb 7167123f; movq $16,%5;"\ + "7167121:\n\t"\ + KERNEL_k1m16n12 "addq $4,%5; testq $12,%5; movq $172,%%r10; cmovz %4,%%r10;"\ + KERNEL_k1m16n12 "prefetcht1 (%3); subq $129,%3; addq %%r10,%3;"\ + KERNEL_k1m16n12 "prefetcht1 (%6); addq $32,%6; cmpq $208,%5; cmoveq %2,%3;"\ + KERNEL_k1m16n12 "cmpq %5,%%r13; jnb 7167121b;"\ + "movq %2,%3;" INVERSE_K_MID\ + "7167123:\n\t"\ + "testq %5,%5; jz 7167129f;"\ + "7167125:\n\t"\ + "prefetcht0 (%3); prefetcht0 64(%3); prefetcht0 127(%3);"\ + KERNEL_k1m16n12 "addq %4,%3; decq %5;jnz 7167125b;"\ + "7167129:\n\t"\ + kernel_kend_n12(16)\ + "prefetcht0 (%%r14);" SAVE_set_pa(16) SAVE_m16n12 save_update_k(16) +#define COMPUTE(ndim) {\ + b_pref = b_ptr + ndim * K; HEAD_SET_OFF(ndim)\ + __asm__ __volatile__(\ + "vbroadcastsd %8,%%zmm0; movq %7,%%r11; movq %1,%%r14; movq %10,%%r12; salq $4,%%r12;" INITASM_SET_K\ + "cmpq $16,%%r11; jb "#ndim"33102f;"\ + #ndim"33101:\n\t"\ + COMPUTE_m16n##ndim "subq $16,%%r11; cmpq $16,%%r11; jnb "#ndim"33101b;"\ + #ndim"33102:\n\t"\ + "cmpq $8,%%r11; jb "#ndim"33103f;"\ + COMPUTE_SIMPLE(8,ndim) "subq $8,%%r11;"\ + #ndim"33103:\n\t"\ + "cmpq $4,%%r11; jb "#ndim"33104f;"\ + COMPUTE_SIMPLE(4,ndim) "subq $4,%%r11;"\ + #ndim"33104:\n\t"\ + "cmpq $2,%%r11; jb "#ndim"33105f;"\ + COMPUTE_SIMPLE(2,ndim) "subq $2,%%r11;"\ + #ndim"33105:\n\t"\ + "testq %%r11,%%r11; jz "#ndim"33106f;"\ + COMPUTE_SIMPLE(1,ndim) "subq $1,%%r11;"\ + #ndim"33106:\n\t"\ + "movq %%r14,%1;"\ + :"+r"(a_ptr),"+r"(b_ptr),"+r"(c_ptr),"+r"(c_tmp),"+r"(ldc_in_bytes),"+r"(k_count),"+r"(b_pref):"m"(M),"m"(ALPHA),"m"(off),"m"(K):"r10","r11","r12","r13","r14","r15","cc","memory",\ + "zmm0","zmm1","zmm2","zmm3","zmm4","zmm5","zmm6","zmm7","zmm8","zmm9","zmm10","zmm11","zmm12","zmm13","zmm14","zmm15",\ + "zmm16","zmm17","zmm18","zmm19","zmm20","zmm21","zmm22","zmm23","zmm24","zmm25","zmm26","zmm27","zmm28","zmm29","zmm30","zmm31");\ + a_ptr -= M * K; b_ptr += ndim * K; c_ptr += ndim * ldc - M; TAIL_SET_OFF(ndim)\ +} + +#include "common.h" +#include + +int __attribute__ ((noinline)) +CNAME(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double * __restrict__ A, double * __restrict__ B, double * __restrict__ C, BLASLONG ldc +#ifdef TRMMKERNEL + , BLASLONG offset +#endif +) +{ + if(m==0||n==0) return 0; + int64_t ldc_in_bytes = (int64_t)ldc * sizeof(double); double ALPHA = alpha; + int64_t M = (int64_t)m, K = (int64_t)k, k_count = 0; + BLASLONG n_count = n, off = 0; + double *a_ptr = A,*b_ptr = B,*c_ptr = C,*c_tmp = C,*b_pref = B; +#ifdef TRMMKERNEL + #ifdef LEFT + off = offset; + #else + off = -offset; + #endif +#endif + for(;n_count>11;n_count-=12) COMPUTE(12) + for(;n_count>9;n_count-=10) COMPUTE(10) + for(;n_count>7;n_count-=8) COMPUTE(8) + for(;n_count>5;n_count-=6) COMPUTE(6) + for(;n_count>3;n_count-=4) COMPUTE(4) + for(;n_count>1;n_count-=2) COMPUTE(2) + if(n_count>0) COMPUTE(1) + return 0; +} + diff --git a/kernel/x86_64/dscal.c b/kernel/x86_64/dscal.c index d0d7801fd..e2436f789 100644 --- a/kernel/x86_64/dscal.c +++ b/kernel/x86_64/dscal.c @@ -136,10 +136,10 @@ static void dscal_kernel_inc_8(BLASLONG n, FLOAT *alpha, FLOAT *x, BLASLONG inc_ "jnz 1b \n\t" : - "+r" (n) // 0 + "+r" (n), // 0 + "+r" (x), // 1 + "+r" (x1) // 2 : - "r" (x), // 1 - "r" (x1), // 2 "r" (alpha), // 3 "r" (inc_x), // 4 "r" (inc_x3) // 5 diff --git a/kernel/x86_64/iamax_sse.S b/kernel/x86_64/iamax_sse.S index d50c1699c..4f62b9be2 100644 --- a/kernel/x86_64/iamax_sse.S +++ b/kernel/x86_64/iamax_sse.S @@ -36,10 +36,6 @@ /* or implied, of The University of Texas at Austin. */ /*********************************************************************/ -/* This kernel was found to give wrong results when used for ISMIN/ISAMIN - with increment != 1, although it appears to be correct for corresponding - MAX operations. See issue 2116 */ - #define ASSEMBLER #include "common.h" @@ -59,6 +55,15 @@ #define MAXSS minss #endif +.macro LOAD_AND_COMPARE_TO_MXX REG + movss 0 * SIZE(X), \REG + addq INCX, X +#ifdef USE_ABS + andps %xmm15, \REG +#endif + cmpeqss %xmm0, \REG +.endm + #include "l1param.h" PROLOGUE @@ -830,61 +835,14 @@ ALIGN_4 .L93: - movss 0 * SIZE(X), %xmm1 - addq INCX, X -#ifdef USE_ABS - andps %xmm15, %xmm1 -#endif - cmpeqss %xmm0, %xmm1 - - movss 0 * SIZE(X), %xmm2 - addq INCX, X -#ifdef USE_ABS - andps %xmm15, %xmm2 -#endif - cmpeqss %xmm0, %xmm2 - - movss 0 * SIZE(X), %xmm3 - addq INCX, X -#ifdef USE_ABS - andps %xmm15, %xmm3 -#endif - cmpeqss %xmm0, %xmm3 - - movss 0 * SIZE(X), %xmm4 - addq INCX, X -#ifdef USE_ABS - andps %xmm15, %xmm4 -#endif - cmpeqss %xmm0, %xmm4 - - movss 0 * SIZE(X), %xmm5 - addq INCX, X -#ifdef USE_ABS - andps %xmm15, %xmm5 -#endif - cmpeqps %xmm0, %xmm5 - - movss 0 * SIZE(X), %xmm6 - addq INCX, X -#ifdef USE_ABS - andps %xmm15, %xmm6 -#endif - cmpeqss %xmm0, %xmm6 - - movss 0 * SIZE(X), %xmm7 - addq INCX, X -#ifdef USE_ABS - andps %xmm15, %xmm7 -#endif - cmpeqss %xmm0, %xmm7 - - movss 0 * SIZE(X), %xmm8 - addq INCX, X -#ifdef USE_ABS - andps %xmm15, %xmm8 -#endif - cmpeqss %xmm0, %xmm8 + LOAD_AND_COMPARE_TO_MXX %xmm1 + LOAD_AND_COMPARE_TO_MXX %xmm2 + LOAD_AND_COMPARE_TO_MXX %xmm3 + LOAD_AND_COMPARE_TO_MXX %xmm4 + LOAD_AND_COMPARE_TO_MXX %xmm5 + LOAD_AND_COMPARE_TO_MXX %xmm6 + LOAD_AND_COMPARE_TO_MXX %xmm7 + LOAD_AND_COMPARE_TO_MXX %xmm8 orps %xmm2, %xmm1 orps %xmm4, %xmm3 diff --git a/kernel/x86_64/sgemm_direct_skylakex.c b/kernel/x86_64/sgemm_direct_skylakex.c new file mode 100644 index 000000000..0e8f1318f --- /dev/null +++ b/kernel/x86_64/sgemm_direct_skylakex.c @@ -0,0 +1,467 @@ + +/* the direct sgemm code written by Arjan van der Ven */ +//#include + +/* + * "Direct sgemm" code. This code operates directly on the inputs and outputs + * of the sgemm call, avoiding the copies, memory realignments and threading, + * and only supports alpha = 1 and beta = 0. + * This is a common case and provides value for relatively small matrixes. + * For larger matrixes the "regular" sgemm code is superior, there the cost of + * copying/shuffling the B matrix really pays off. + */ + + + +#define DECLARE_RESULT_512(N,M) __m512 result##N##M = _mm512_setzero_ps() +#define BROADCAST_LOAD_A_512(N,M) __m512 Aval##M = _mm512_broadcastss_ps(_mm_load_ss(&A[k + strideA * (i+M)])) +#define LOAD_B_512(N,M) __m512 Bval##N = _mm512_loadu_ps(&B[strideB * k + j + (N*16)]) +#define MATMUL_512(N,M) result##N##M = _mm512_fmadd_ps(Aval##M, Bval##N , result##N##M) +#define STORE_512(N,M) _mm512_storeu_ps(&R[(i+M) * strideR + j+(N*16)], result##N##M) + + +#define DECLARE_RESULT_256(N,M) __m256 result##N##M = _mm256_setzero_ps() +#define BROADCAST_LOAD_A_256(N,M) __m256 Aval##M = _mm256_broadcastss_ps(_mm_load_ss(&A[k + strideA * (i+M)])) +#define LOAD_B_256(N,M) __m256 Bval##N = _mm256_loadu_ps(&B[strideB * k + j + (N*8)]) +#define MATMUL_256(N,M) result##N##M = _mm256_fmadd_ps(Aval##M, Bval##N , result##N##M) +#define STORE_256(N,M) _mm256_storeu_ps(&R[(i+M) * strideR + j+(N*8)], result##N##M) + +#define DECLARE_RESULT_128(N,M) __m128 result##N##M = _mm_setzero_ps() +#define BROADCAST_LOAD_A_128(N,M) __m128 Aval##M = _mm_broadcastss_ps(_mm_load_ss(&A[k + strideA * (i+M)])) +#define LOAD_B_128(N,M) __m128 Bval##N = _mm_loadu_ps(&B[strideB * k + j + (N*4)]) +#define MATMUL_128(N,M) result##N##M = _mm_fmadd_ps(Aval##M, Bval##N , result##N##M) +#define STORE_128(N,M) _mm_storeu_ps(&R[(i+M) * strideR + j+(N*4)], result##N##M) + +#define DECLARE_RESULT_SCALAR(N,M) float result##N##M = 0; +#define BROADCAST_LOAD_A_SCALAR(N,M) float Aval##M = A[k + strideA * (i + M)]; +#define LOAD_B_SCALAR(N,M) float Bval##N = B[k * strideB + j + N]; +#define MATMUL_SCALAR(N,M) result##N##M += Aval##M * Bval##N; +#define STORE_SCALAR(N,M) R[(i+M) * strideR + j + N] = result##N##M; + +int sgemm_kernel_direct_performant(BLASLONG M, BLASLONG N, BLASLONG K) +{ + unsigned long long mnk = M * N * K; + /* large matrixes -> not performant */ + if (mnk >= 28 * 512 * 512) + return 0; + + /* + * if the B matrix is not a nice multiple if 4 we get many unaligned accesses, + * and the regular sgemm copy/realignment of data pays off much quicker + */ + if ((N & 3) != 0 && (mnk >= 8 * 512 * 512)) + return 0; + +#ifdef SMP + /* if we can run multithreaded, the threading changes the based threshold */ + if (mnk > 2 * 350 * 512 && num_cpu_avail(3)> 1) + return 0; +#endif + + return 1; +} + + + +void sgemm_kernel_direct (BLASLONG M, BLASLONG N, BLASLONG K, float * __restrict A, BLASLONG strideA, float * __restrict B, BLASLONG strideB , float * __restrict R, BLASLONG strideR) +{ + int i, j, k; + + int m4 = M & ~3; + int m2 = M & ~1; + + int n64 = N & ~63; + int n32 = N & ~31; + int n16 = N & ~15; + int n8 = N & ~7; + int n4 = N & ~3; + int n2 = N & ~1; + + i = 0; + + for (i = 0; i < m4; i+=4) { + + for (j = 0; j < n64; j+= 64) { + k = 0; + DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0); + DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1); + DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2); DECLARE_RESULT_512(2, 2); DECLARE_RESULT_512(3, 2); + DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3); DECLARE_RESULT_512(2, 3); DECLARE_RESULT_512(3, 3); + + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_512(x, 0); + BROADCAST_LOAD_A_512(x, 1); + BROADCAST_LOAD_A_512(x, 2); + BROADCAST_LOAD_A_512(x, 3); + + LOAD_B_512(0, x); LOAD_B_512(1, x); LOAD_B_512(2, x); LOAD_B_512(3, x); + + MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0); + MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1); + MATMUL_512(0, 2); MATMUL_512(1, 2); MATMUL_512(2, 2); MATMUL_512(3, 2); + MATMUL_512(0, 3); MATMUL_512(1, 3); MATMUL_512(2, 3); MATMUL_512(3, 3); + } + STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0); + STORE_512(0, 1); STORE_512(1, 1); STORE_512(2, 1); STORE_512(3, 1); + STORE_512(0, 2); STORE_512(1, 2); STORE_512(2, 2); STORE_512(3, 2); + STORE_512(0, 3); STORE_512(1, 3); STORE_512(2, 3); STORE_512(3, 3); + } + + for (; j < n32; j+= 32) { + DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); + DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); + DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2); + DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_512(x, 0); + BROADCAST_LOAD_A_512(x, 1); + BROADCAST_LOAD_A_512(x, 2); + BROADCAST_LOAD_A_512(x, 3); + + LOAD_B_512(0, x); LOAD_B_512(1, x); + + MATMUL_512(0, 0); MATMUL_512(1, 0); + MATMUL_512(0, 1); MATMUL_512(1, 1); + MATMUL_512(0, 2); MATMUL_512(1, 2); + MATMUL_512(0, 3); MATMUL_512(1, 3); + } + STORE_512(0, 0); STORE_512(1, 0); + STORE_512(0, 1); STORE_512(1, 1); + STORE_512(0, 2); STORE_512(1, 2); + STORE_512(0, 3); STORE_512(1, 3); + } + + for (; j < n16; j+= 16) { + DECLARE_RESULT_512(0, 0); + DECLARE_RESULT_512(0, 1); + DECLARE_RESULT_512(0, 2); + DECLARE_RESULT_512(0, 3); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_512(x, 0); + BROADCAST_LOAD_A_512(x, 1); + BROADCAST_LOAD_A_512(x, 2); + BROADCAST_LOAD_A_512(x, 3); + + LOAD_B_512(0, x); + + MATMUL_512(0, 0); + MATMUL_512(0, 1); + MATMUL_512(0, 2); + MATMUL_512(0, 3); + } + STORE_512(0, 0); + STORE_512(0, 1); + STORE_512(0, 2); + STORE_512(0, 3); + } + + for (; j < n8; j+= 8) { + DECLARE_RESULT_256(0, 0); + DECLARE_RESULT_256(0, 1); + DECLARE_RESULT_256(0, 2); + DECLARE_RESULT_256(0, 3); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_256(x, 0); + BROADCAST_LOAD_A_256(x, 1); + BROADCAST_LOAD_A_256(x, 2); + BROADCAST_LOAD_A_256(x, 3); + + LOAD_B_256(0, x); + + MATMUL_256(0, 0); + MATMUL_256(0, 1); + MATMUL_256(0, 2); + MATMUL_256(0, 3); + } + STORE_256(0, 0); + STORE_256(0, 1); + STORE_256(0, 2); + STORE_256(0, 3); + } + + for (; j < n4; j+= 4) { + DECLARE_RESULT_128(0, 0); + DECLARE_RESULT_128(0, 1); + DECLARE_RESULT_128(0, 2); + DECLARE_RESULT_128(0, 3); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_128(x, 0); + BROADCAST_LOAD_A_128(x, 1); + BROADCAST_LOAD_A_128(x, 2); + BROADCAST_LOAD_A_128(x, 3); + + LOAD_B_128(0, x); + + MATMUL_128(0, 0); + MATMUL_128(0, 1); + MATMUL_128(0, 2); + MATMUL_128(0, 3); + } + STORE_128(0, 0); + STORE_128(0, 1); + STORE_128(0, 2); + STORE_128(0, 3); + } + + for (; j < n2; j+= 2) { + DECLARE_RESULT_SCALAR(0, 0); DECLARE_RESULT_SCALAR(1, 0); + DECLARE_RESULT_SCALAR(0, 1); DECLARE_RESULT_SCALAR(1, 1); + DECLARE_RESULT_SCALAR(0, 2); DECLARE_RESULT_SCALAR(1, 2); + DECLARE_RESULT_SCALAR(0, 3); DECLARE_RESULT_SCALAR(1, 3); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_SCALAR(x, 0); + BROADCAST_LOAD_A_SCALAR(x, 1); + BROADCAST_LOAD_A_SCALAR(x, 2); + BROADCAST_LOAD_A_SCALAR(x, 3); + + LOAD_B_SCALAR(0, x); LOAD_B_SCALAR(1, x); + + MATMUL_SCALAR(0, 0); MATMUL_SCALAR(1, 0); + MATMUL_SCALAR(0, 1); MATMUL_SCALAR(1, 1); + MATMUL_SCALAR(0, 2); MATMUL_SCALAR(1, 2); + MATMUL_SCALAR(0, 3); MATMUL_SCALAR(1, 3); + } + STORE_SCALAR(0, 0); STORE_SCALAR(1, 0); + STORE_SCALAR(0, 1); STORE_SCALAR(1, 1); + STORE_SCALAR(0, 2); STORE_SCALAR(1, 2); + STORE_SCALAR(0, 3); STORE_SCALAR(1, 3); + } + + for (; j < N; j++) { + DECLARE_RESULT_SCALAR(0, 0) + DECLARE_RESULT_SCALAR(0, 1) + DECLARE_RESULT_SCALAR(0, 2) + DECLARE_RESULT_SCALAR(0, 3) + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_SCALAR(0, 0); + BROADCAST_LOAD_A_SCALAR(0, 1); + BROADCAST_LOAD_A_SCALAR(0, 2); + BROADCAST_LOAD_A_SCALAR(0, 3); + + LOAD_B_SCALAR(0, 0); + + MATMUL_SCALAR(0, 0); + MATMUL_SCALAR(0, 1); + MATMUL_SCALAR(0, 2); + MATMUL_SCALAR(0, 3); + } + STORE_SCALAR(0, 0); + STORE_SCALAR(0, 1); + STORE_SCALAR(0, 2); + STORE_SCALAR(0, 3); + } + } + + for (; i < m2; i+=2) { + j = 0; + + for (; j < n64; j+= 64) { + DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0); + DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1); + + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_512(x, 0); + BROADCAST_LOAD_A_512(x, 1); + + LOAD_B_512(0, x); LOAD_B_512(1, x); LOAD_B_512(2, x); LOAD_B_512(3, x); + + MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0); + MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1); + } + STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0); + STORE_512(0, 1); STORE_512(1, 1); STORE_512(2, 1); STORE_512(3, 1); + } + + for (; j < n32; j+= 32) { + DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); + DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_512(x, 0); + BROADCAST_LOAD_A_512(x, 1); + + LOAD_B_512(0, x); LOAD_B_512(1, x); + + MATMUL_512(0, 0); MATMUL_512(1, 0); + MATMUL_512(0, 1); MATMUL_512(1, 1); + } + STORE_512(0, 0); STORE_512(1, 0); + STORE_512(0, 1); STORE_512(1, 1); + } + + + for (; j < n16; j+= 16) { + DECLARE_RESULT_512(0, 0); + DECLARE_RESULT_512(0, 1); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_512(x, 0); + BROADCAST_LOAD_A_512(x, 1); + + LOAD_B_512(0, x); + + MATMUL_512(0, 0); + MATMUL_512(0, 1); + } + STORE_512(0, 0); + STORE_512(0, 1); + } + + for (; j < n8; j+= 8) { + DECLARE_RESULT_256(0, 0); + DECLARE_RESULT_256(0, 1); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_256(x, 0); + BROADCAST_LOAD_A_256(x, 1); + + LOAD_B_256(0, x); + + MATMUL_256(0, 0); + MATMUL_256(0, 1); + } + STORE_256(0, 0); + STORE_256(0, 1); + } + + for (; j < n4; j+= 4) { + DECLARE_RESULT_128(0, 0); + DECLARE_RESULT_128(0, 1); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_128(x, 0); + BROADCAST_LOAD_A_128(x, 1); + + LOAD_B_128(0, x); + + MATMUL_128(0, 0); + MATMUL_128(0, 1); + } + STORE_128(0, 0); + STORE_128(0, 1); + } + for (; j < n2; j+= 2) { + DECLARE_RESULT_SCALAR(0, 0); DECLARE_RESULT_SCALAR(1, 0); + DECLARE_RESULT_SCALAR(0, 1); DECLARE_RESULT_SCALAR(1, 1); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_SCALAR(x, 0); + BROADCAST_LOAD_A_SCALAR(x, 1); + + LOAD_B_SCALAR(0, x); LOAD_B_SCALAR(1, x); + + MATMUL_SCALAR(0, 0); MATMUL_SCALAR(1, 0); + MATMUL_SCALAR(0, 1); MATMUL_SCALAR(1, 1); + } + STORE_SCALAR(0, 0); STORE_SCALAR(1, 0); + STORE_SCALAR(0, 1); STORE_SCALAR(1, 1); + } + + for (; j < N; j++) { + DECLARE_RESULT_SCALAR(0, 0); + DECLARE_RESULT_SCALAR(0, 1); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_SCALAR(0, 0); + BROADCAST_LOAD_A_SCALAR(0, 1); + + LOAD_B_SCALAR(0, 0); + + MATMUL_SCALAR(0, 0); + MATMUL_SCALAR(0, 1); + } + STORE_SCALAR(0, 0); + STORE_SCALAR(0, 1); + } + } + + for (; i < M; i+=1) { + j = 0; + for (; j < n64; j+= 64) { + DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_512(x, 0); + LOAD_B_512(0, x); LOAD_B_512(1, x); LOAD_B_512(2, x); LOAD_B_512(3, x); + MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0); + } + STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0); + } + for (; j < n32; j+= 32) { + DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_512(x, 0); + LOAD_B_512(0, x); LOAD_B_512(1, x); + MATMUL_512(0, 0); MATMUL_512(1, 0); + } + STORE_512(0, 0); STORE_512(1, 0); + } + + + for (; j < n16; j+= 16) { + DECLARE_RESULT_512(0, 0); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_512(x, 0); + + LOAD_B_512(0, x); + + MATMUL_512(0, 0); + } + STORE_512(0, 0); + } + + for (; j < n8; j+= 8) { + DECLARE_RESULT_256(0, 0); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_256(x, 0); + LOAD_B_256(0, x); + MATMUL_256(0, 0); + } + STORE_256(0, 0); + } + + for (; j < n4; j+= 4) { + DECLARE_RESULT_128(0, 0); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_128(x, 0); + LOAD_B_128(0, x); + MATMUL_128(0, 0); + } + STORE_128(0, 0); + } + + for (; j < n2; j+= 2) { + DECLARE_RESULT_SCALAR(0, 0); DECLARE_RESULT_SCALAR(1, 0); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_SCALAR(x, 0); + LOAD_B_SCALAR(0, 0); LOAD_B_SCALAR(1, 0); + MATMUL_SCALAR(0, 0); MATMUL_SCALAR(1, 0); + } + STORE_SCALAR(0, 0); STORE_SCALAR(1, 0); + } + + for (; j < N; j++) { + DECLARE_RESULT_SCALAR(0, 0); + + for (k = 0; k < K; k++) { + BROADCAST_LOAD_A_SCALAR(0, 0); + LOAD_B_SCALAR(0, 0); + MATMUL_SCALAR(0, 0); + } + STORE_SCALAR(0, 0); + } + } +} diff --git a/kernel/x86_64/sgemm_kernel_16x4_skylakex.c b/kernel/x86_64/sgemm_kernel_16x4_skylakex.c index 76b82e65b..d174bbcc3 100644 --- a/kernel/x86_64/sgemm_kernel_16x4_skylakex.c +++ b/kernel/x86_64/sgemm_kernel_16x4_skylakex.c @@ -1176,467 +1176,4 @@ CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict A, flo return 0; } - -/* - * "Direct sgemm" code. This code operates directly on the inputs and outputs - * of the sgemm call, avoiding the copies, memory realignments and threading, - * and only supports alpha = 1 and beta = 0. - * This is a common case and provides value for relatively small matrixes. - * For larger matrixes the "regular" sgemm code is superior, there the cost of - * copying/shuffling the B matrix really pays off. - */ - - - -#define DECLARE_RESULT_512(N,M) __m512 result##N##M = _mm512_setzero_ps() -#define BROADCAST_LOAD_A_512(N,M) __m512 Aval##M = _mm512_broadcastss_ps(_mm_load_ss(&A[k + strideA * (i+M)])) -#define LOAD_B_512(N,M) __m512 Bval##N = _mm512_loadu_ps(&B[strideB * k + j + (N*16)]) -#define MATMUL_512(N,M) result##N##M = _mm512_fmadd_ps(Aval##M, Bval##N , result##N##M) -#define STORE_512(N,M) _mm512_storeu_ps(&R[(i+M) * strideR + j+(N*16)], result##N##M) - - -#define DECLARE_RESULT_256(N,M) __m256 result##N##M = _mm256_setzero_ps() -#define BROADCAST_LOAD_A_256(N,M) __m256 Aval##M = _mm256_broadcastss_ps(_mm_load_ss(&A[k + strideA * (i+M)])) -#define LOAD_B_256(N,M) __m256 Bval##N = _mm256_loadu_ps(&B[strideB * k + j + (N*8)]) -#define MATMUL_256(N,M) result##N##M = _mm256_fmadd_ps(Aval##M, Bval##N , result##N##M) -#define STORE_256(N,M) _mm256_storeu_ps(&R[(i+M) * strideR + j+(N*8)], result##N##M) - -#define DECLARE_RESULT_128(N,M) __m128 result##N##M = _mm_setzero_ps() -#define BROADCAST_LOAD_A_128(N,M) __m128 Aval##M = _mm_broadcastss_ps(_mm_load_ss(&A[k + strideA * (i+M)])) -#define LOAD_B_128(N,M) __m128 Bval##N = _mm_loadu_ps(&B[strideB * k + j + (N*4)]) -#define MATMUL_128(N,M) result##N##M = _mm_fmadd_ps(Aval##M, Bval##N , result##N##M) -#define STORE_128(N,M) _mm_storeu_ps(&R[(i+M) * strideR + j+(N*4)], result##N##M) - -#define DECLARE_RESULT_SCALAR(N,M) float result##N##M = 0; -#define BROADCAST_LOAD_A_SCALAR(N,M) float Aval##M = A[k + strideA * (i + M)]; -#define LOAD_B_SCALAR(N,M) float Bval##N = B[k * strideB + j + N]; -#define MATMUL_SCALAR(N,M) result##N##M += Aval##M * Bval##N; -#define STORE_SCALAR(N,M) R[(i+M) * strideR + j + N] = result##N##M; - -int sgemm_kernel_direct_performant(BLASLONG M, BLASLONG N, BLASLONG K) -{ - unsigned long long mnk = M * N * K; - /* large matrixes -> not performant */ - if (mnk >= 28 * 512 * 512) - return 0; - - /* - * if the B matrix is not a nice multiple if 4 we get many unaligned accesses, - * and the regular sgemm copy/realignment of data pays off much quicker - */ - if ((N & 3) != 0 && (mnk >= 8 * 512 * 512)) - return 0; - -#ifdef SMP - /* if we can run multithreaded, the threading changes the based threshold */ - if (mnk > 2 * 350 * 512 && num_cpu_avail(3)> 1) - return 0; -#endif - - return 1; -} - - - -void sgemm_kernel_direct (BLASLONG M, BLASLONG N, BLASLONG K, float * __restrict A, BLASLONG strideA, float * __restrict B, BLASLONG strideB , float * __restrict R, BLASLONG strideR) -{ - int i, j, k; - - int m4 = M & ~3; - int m2 = M & ~1; - - int n64 = N & ~63; - int n32 = N & ~31; - int n16 = N & ~15; - int n8 = N & ~7; - int n4 = N & ~3; - int n2 = N & ~1; - - i = 0; - - for (i = 0; i < m4; i+=4) { - - for (j = 0; j < n64; j+= 64) { - k = 0; - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0); - DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1); - DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2); DECLARE_RESULT_512(2, 2); DECLARE_RESULT_512(3, 2); - DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3); DECLARE_RESULT_512(2, 3); DECLARE_RESULT_512(3, 3); - - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - BROADCAST_LOAD_A_512(x, 2); - BROADCAST_LOAD_A_512(x, 3); - - LOAD_B_512(0, x); LOAD_B_512(1, x); LOAD_B_512(2, x); LOAD_B_512(3, x); - - MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0); - MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1); - MATMUL_512(0, 2); MATMUL_512(1, 2); MATMUL_512(2, 2); MATMUL_512(3, 2); - MATMUL_512(0, 3); MATMUL_512(1, 3); MATMUL_512(2, 3); MATMUL_512(3, 3); - } - STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0); - STORE_512(0, 1); STORE_512(1, 1); STORE_512(2, 1); STORE_512(3, 1); - STORE_512(0, 2); STORE_512(1, 2); STORE_512(2, 2); STORE_512(3, 2); - STORE_512(0, 3); STORE_512(1, 3); STORE_512(2, 3); STORE_512(3, 3); - } - - for (; j < n32; j+= 32) { - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); - DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); - DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2); - DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - BROADCAST_LOAD_A_512(x, 2); - BROADCAST_LOAD_A_512(x, 3); - - LOAD_B_512(0, x); LOAD_B_512(1, x); - - MATMUL_512(0, 0); MATMUL_512(1, 0); - MATMUL_512(0, 1); MATMUL_512(1, 1); - MATMUL_512(0, 2); MATMUL_512(1, 2); - MATMUL_512(0, 3); MATMUL_512(1, 3); - } - STORE_512(0, 0); STORE_512(1, 0); - STORE_512(0, 1); STORE_512(1, 1); - STORE_512(0, 2); STORE_512(1, 2); - STORE_512(0, 3); STORE_512(1, 3); - } - - for (; j < n16; j+= 16) { - DECLARE_RESULT_512(0, 0); - DECLARE_RESULT_512(0, 1); - DECLARE_RESULT_512(0, 2); - DECLARE_RESULT_512(0, 3); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - BROADCAST_LOAD_A_512(x, 2); - BROADCAST_LOAD_A_512(x, 3); - - LOAD_B_512(0, x); - - MATMUL_512(0, 0); - MATMUL_512(0, 1); - MATMUL_512(0, 2); - MATMUL_512(0, 3); - } - STORE_512(0, 0); - STORE_512(0, 1); - STORE_512(0, 2); - STORE_512(0, 3); - } - - for (; j < n8; j+= 8) { - DECLARE_RESULT_256(0, 0); - DECLARE_RESULT_256(0, 1); - DECLARE_RESULT_256(0, 2); - DECLARE_RESULT_256(0, 3); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_256(x, 0); - BROADCAST_LOAD_A_256(x, 1); - BROADCAST_LOAD_A_256(x, 2); - BROADCAST_LOAD_A_256(x, 3); - - LOAD_B_256(0, x); - - MATMUL_256(0, 0); - MATMUL_256(0, 1); - MATMUL_256(0, 2); - MATMUL_256(0, 3); - } - STORE_256(0, 0); - STORE_256(0, 1); - STORE_256(0, 2); - STORE_256(0, 3); - } - - for (; j < n4; j+= 4) { - DECLARE_RESULT_128(0, 0); - DECLARE_RESULT_128(0, 1); - DECLARE_RESULT_128(0, 2); - DECLARE_RESULT_128(0, 3); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_128(x, 0); - BROADCAST_LOAD_A_128(x, 1); - BROADCAST_LOAD_A_128(x, 2); - BROADCAST_LOAD_A_128(x, 3); - - LOAD_B_128(0, x); - - MATMUL_128(0, 0); - MATMUL_128(0, 1); - MATMUL_128(0, 2); - MATMUL_128(0, 3); - } - STORE_128(0, 0); - STORE_128(0, 1); - STORE_128(0, 2); - STORE_128(0, 3); - } - - for (; j < n2; j+= 2) { - DECLARE_RESULT_SCALAR(0, 0); DECLARE_RESULT_SCALAR(1, 0); - DECLARE_RESULT_SCALAR(0, 1); DECLARE_RESULT_SCALAR(1, 1); - DECLARE_RESULT_SCALAR(0, 2); DECLARE_RESULT_SCALAR(1, 2); - DECLARE_RESULT_SCALAR(0, 3); DECLARE_RESULT_SCALAR(1, 3); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(x, 0); - BROADCAST_LOAD_A_SCALAR(x, 1); - BROADCAST_LOAD_A_SCALAR(x, 2); - BROADCAST_LOAD_A_SCALAR(x, 3); - - LOAD_B_SCALAR(0, x); LOAD_B_SCALAR(1, x); - - MATMUL_SCALAR(0, 0); MATMUL_SCALAR(1, 0); - MATMUL_SCALAR(0, 1); MATMUL_SCALAR(1, 1); - MATMUL_SCALAR(0, 2); MATMUL_SCALAR(1, 2); - MATMUL_SCALAR(0, 3); MATMUL_SCALAR(1, 3); - } - STORE_SCALAR(0, 0); STORE_SCALAR(1, 0); - STORE_SCALAR(0, 1); STORE_SCALAR(1, 1); - STORE_SCALAR(0, 2); STORE_SCALAR(1, 2); - STORE_SCALAR(0, 3); STORE_SCALAR(1, 3); - } - - for (; j < N; j++) { - DECLARE_RESULT_SCALAR(0, 0) - DECLARE_RESULT_SCALAR(0, 1) - DECLARE_RESULT_SCALAR(0, 2) - DECLARE_RESULT_SCALAR(0, 3) - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(0, 0); - BROADCAST_LOAD_A_SCALAR(0, 1); - BROADCAST_LOAD_A_SCALAR(0, 2); - BROADCAST_LOAD_A_SCALAR(0, 3); - - LOAD_B_SCALAR(0, 0); - - MATMUL_SCALAR(0, 0); - MATMUL_SCALAR(0, 1); - MATMUL_SCALAR(0, 2); - MATMUL_SCALAR(0, 3); - } - STORE_SCALAR(0, 0); - STORE_SCALAR(0, 1); - STORE_SCALAR(0, 2); - STORE_SCALAR(0, 3); - } - } - - for (; i < m2; i+=2) { - j = 0; - - for (; j < n64; j+= 64) { - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0); - DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1); - - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - - LOAD_B_512(0, x); LOAD_B_512(1, x); LOAD_B_512(2, x); LOAD_B_512(3, x); - - MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0); - MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1); - } - STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0); - STORE_512(0, 1); STORE_512(1, 1); STORE_512(2, 1); STORE_512(3, 1); - } - - for (; j < n32; j+= 32) { - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); - DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - - LOAD_B_512(0, x); LOAD_B_512(1, x); - - MATMUL_512(0, 0); MATMUL_512(1, 0); - MATMUL_512(0, 1); MATMUL_512(1, 1); - } - STORE_512(0, 0); STORE_512(1, 0); - STORE_512(0, 1); STORE_512(1, 1); - } - - - for (; j < n16; j+= 16) { - DECLARE_RESULT_512(0, 0); - DECLARE_RESULT_512(0, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - - LOAD_B_512(0, x); - - MATMUL_512(0, 0); - MATMUL_512(0, 1); - } - STORE_512(0, 0); - STORE_512(0, 1); - } - - for (; j < n8; j+= 8) { - DECLARE_RESULT_256(0, 0); - DECLARE_RESULT_256(0, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_256(x, 0); - BROADCAST_LOAD_A_256(x, 1); - - LOAD_B_256(0, x); - - MATMUL_256(0, 0); - MATMUL_256(0, 1); - } - STORE_256(0, 0); - STORE_256(0, 1); - } - - for (; j < n4; j+= 4) { - DECLARE_RESULT_128(0, 0); - DECLARE_RESULT_128(0, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_128(x, 0); - BROADCAST_LOAD_A_128(x, 1); - - LOAD_B_128(0, x); - - MATMUL_128(0, 0); - MATMUL_128(0, 1); - } - STORE_128(0, 0); - STORE_128(0, 1); - } - for (; j < n2; j+= 2) { - DECLARE_RESULT_SCALAR(0, 0); DECLARE_RESULT_SCALAR(1, 0); - DECLARE_RESULT_SCALAR(0, 1); DECLARE_RESULT_SCALAR(1, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(x, 0); - BROADCAST_LOAD_A_SCALAR(x, 1); - - LOAD_B_SCALAR(0, x); LOAD_B_SCALAR(1, x); - - MATMUL_SCALAR(0, 0); MATMUL_SCALAR(1, 0); - MATMUL_SCALAR(0, 1); MATMUL_SCALAR(1, 1); - } - STORE_SCALAR(0, 0); STORE_SCALAR(1, 0); - STORE_SCALAR(0, 1); STORE_SCALAR(1, 1); - } - - for (; j < N; j++) { - DECLARE_RESULT_SCALAR(0, 0); - DECLARE_RESULT_SCALAR(0, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(0, 0); - BROADCAST_LOAD_A_SCALAR(0, 1); - - LOAD_B_SCALAR(0, 0); - - MATMUL_SCALAR(0, 0); - MATMUL_SCALAR(0, 1); - } - STORE_SCALAR(0, 0); - STORE_SCALAR(0, 1); - } - } - - for (; i < M; i+=1) { - j = 0; - for (; j < n64; j+= 64) { - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - LOAD_B_512(0, x); LOAD_B_512(1, x); LOAD_B_512(2, x); LOAD_B_512(3, x); - MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0); - } - STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0); - } - for (; j < n32; j+= 32) { - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - LOAD_B_512(0, x); LOAD_B_512(1, x); - MATMUL_512(0, 0); MATMUL_512(1, 0); - } - STORE_512(0, 0); STORE_512(1, 0); - } - - - for (; j < n16; j+= 16) { - DECLARE_RESULT_512(0, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - - LOAD_B_512(0, x); - - MATMUL_512(0, 0); - } - STORE_512(0, 0); - } - - for (; j < n8; j+= 8) { - DECLARE_RESULT_256(0, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_256(x, 0); - LOAD_B_256(0, x); - MATMUL_256(0, 0); - } - STORE_256(0, 0); - } - - for (; j < n4; j+= 4) { - DECLARE_RESULT_128(0, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_128(x, 0); - LOAD_B_128(0, x); - MATMUL_128(0, 0); - } - STORE_128(0, 0); - } - - for (; j < n2; j+= 2) { - DECLARE_RESULT_SCALAR(0, 0); DECLARE_RESULT_SCALAR(1, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(x, 0); - LOAD_B_SCALAR(0, 0); LOAD_B_SCALAR(1, 0); - MATMUL_SCALAR(0, 0); MATMUL_SCALAR(1, 0); - } - STORE_SCALAR(0, 0); STORE_SCALAR(1, 0); - } - - for (; j < N; j++) { - DECLARE_RESULT_SCALAR(0, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(0, 0); - LOAD_B_SCALAR(0, 0); - MATMUL_SCALAR(0, 0); - } - STORE_SCALAR(0, 0); - } - } -} +#include "sgemm_direct_skylakex.c" diff --git a/kernel/x86_64/sgemm_kernel_16x4_skylakex_2.c b/kernel/x86_64/sgemm_kernel_16x4_skylakex_2.c index ee3417505..ec7570179 100644 --- a/kernel/x86_64/sgemm_kernel_16x4_skylakex_2.c +++ b/kernel/x86_64/sgemm_kernel_16x4_skylakex_2.c @@ -1,8 +1,152 @@ -/* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 for k_count, %5 for c_store */ -/* r12 = k << 4(const), r13 = k(const), r14 = b_head_pos(const), r15 = %1 + 3r12 */ +/* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 for k_count, %5 for c_store, %6 = b_pref */ +/* r10 to assist prefetch, r11 = m_counter, r12 = k << 4(const), r13 = k_todo, r14 = b_head_pos(const), r15 = %1 + 3r12 */ #include "common.h" #include +#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + #define BACKWARDS 1 +#else + #define BACKWARDS 0 +#endif +#define REC_POINTER_1(ptr) "salq $2,%%r13; subq %%r13,"#ptr"; sarq $2,%%r13;" +#define REC_POINTER_2(ptr) "salq $3,%%r13; subq %%r13,"#ptr"; sarq $3,%%r13;" +#define REC_POINTER_4(ptr) "salq $4,%%r13; subq %%r13,"#ptr"; sarq $4,%%r13;" +#define REC_POINTER_8(ptr) "salq $5,%%r13; subq %%r13,"#ptr"; sarq $5,%%r13;" +#define REC_POINTER_16(ptr) "salq $6,%%r13; subq %%r13,"#ptr"; sarq $6,%%r13;" +#define INC_POINTER_1(ptr) "sarq $2,%%r12; addq %%r12,"#ptr"; salq $2,%%r12;" +#define INC_POINTER_2(ptr) "sarq $1,%%r12; addq %%r12,"#ptr"; salq $1,%%r12;" +#define INC_POINTER_4(ptr) "addq %%r12,"#ptr";" +#define INC_POINTER_8(ptr) "leaq ("#ptr",%%r12,2),"#ptr";" +#define INC_POINTER_16(ptr) "leaq ("#ptr",%%r12,4),"#ptr";" +#define SET_POINTER(ptr,dim) REC_POINTER_##dim(ptr) INC_POINTER_##dim(ptr) +#define SET_PB_1 SET_POINTER(%1,1) +#define SET_PB_2 SET_POINTER(%1,2) +#define SET_PB_4 SET_POINTER(%1,4) +#define SET_PB_8 SET_POINTER(%1,4) +#define SET_PB_12 SET_POINTER(%1,4) +#define SET_PB_16 SET_POINTER(%1,4) +#define SET_PB_20 SET_POINTER(%1,4) +#define SET_PB_24 SET_POINTER(%1,4) +#ifdef TRMMKERNEL + #if BACKWARDS == 1 + #define START_SET_PAPB(mdim,ndim) SET_POINTER(%0,mdim) "movq %%r14,%1;" SET_PB_##ndim "leaq (%1,%%r12,2),%%r15; addq %%r12,%%r15;" + #define END_SET_PA(mdim) "" + #else + #define START_SET_PAPB(mdim,ndim) "movq %%r14,%1; leaq (%1,%%r12,2),%%r15; addq %%r12,%%r15;" + #define END_SET_PA(mdim) SET_POINTER(%0,mdim) + #endif +#else + #define START_SET_PAPB(mdim,ndim) "movq %%r14,%1; leaq (%1,%%r12,2),%%r15; addq %%r12,%%r15;" + #define END_SET_PA(mdim) "" +#endif +#define RECOVER_PA(mdim) REC_POINTER_##mdim(%0) + +#if defined(TRMMKERNEL) && !defined(LEFT) + #if BACKWARDS == 1 + #define KERNEL_HEAD_C_n8(mdim) \ + KERNEL_k1m##mdim##n4 KERNEL_k1m##mdim##n4 KERNEL_k1m##mdim##n4 KERNEL_k1m##mdim##n4 "subq $4,%4; addq $64,%%r15;" + #define KERNEL_HEAD_C_n12(mdim) KERNEL_HEAD_C_n8(mdim)\ + KERNEL_k1m##mdim##n8 KERNEL_k1m##mdim##n8 KERNEL_k1m##mdim##n8 KERNEL_k1m##mdim##n8 "subq $4,%4; addq $64,%%r15;" + #define KERNEL_HEAD_C_n16(mdim) KERNEL_HEAD_C_n12(mdim)\ + KERNEL_k1m##mdim##n12 KERNEL_k1m##mdim##n12 KERNEL_k1m##mdim##n12 KERNEL_k1m##mdim##n12 "subq $4,%4; addq $64,%%r15;" + #define KERNEL_HEAD_C_n20(mdim) KERNEL_HEAD_C_n16(mdim)\ + KERNEL_k1m##mdim##n16 KERNEL_k1m##mdim##n16 KERNEL_k1m##mdim##n16 KERNEL_k1m##mdim##n16 "subq $4,%4;" + #define KERNEL_HEAD_C_n24(mdim) KERNEL_HEAD_C_n20(mdim)\ + KERNEL_k1m##mdim##n20 KERNEL_k1m##mdim##n20 KERNEL_k1m##mdim##n20 KERNEL_k1m##mdim##n20 "subq $4,%4;" + #define KERNEL_HEAD_R_n4(mdim) "subq $12,%4; addq $64,%%r15; addq $"#mdim"*48,%0;" + #define KERNEL_HEAD_R_n8(mdim) KERNEL_HEAD_R_n4(mdim)\ + kernel_k1m##mdim##n4(%%r15) kernel_k1m##mdim##n4(%%r15) kernel_k1m##mdim##n4(%%r15) kernel_k1m##mdim##n4(%%r15) "subq $4,%4;" + #define KERNEL_HEAD_R_n12(mdim) KERNEL_HEAD_R_n8(mdim)\ + kernel_k1m##mdim##n8(%%r15) kernel_k1m##mdim##n8(%%r15) kernel_k1m##mdim##n8(%%r15) kernel_k1m##mdim##n8(%%r15) "subq $4,%4;" + #define KERNEL_TAIL_C_n8(mdim) "" + #define KERNEL_TAIL_C_n12(mdim) "" + #define KERNEL_TAIL_C_n16(mdim) "" + #define KERNEL_TAIL_C_n20(mdim) "" + #define KERNEL_TAIL_C_n24(mdim) "" + #define KERNEL_TAIL_R_n4(mdim) "" + #define KERNEL_TAIL_R_n8(mdim) "" + #define KERNEL_TAIL_R_n12(mdim) "" + #else + #define KERNEL_HEAD_C_n8(mdim) "" + #define KERNEL_HEAD_C_n12(mdim) "" + #define KERNEL_HEAD_C_n16(mdim) "" + #define KERNEL_HEAD_C_n20(mdim) "" + #define KERNEL_HEAD_C_n24(mdim) "" + #define KERNEL_HEAD_R_n4(mdim) "" + #define KERNEL_HEAD_R_n8(mdim) "" + #define KERNEL_HEAD_R_n12(mdim) "" + #define end_kernel_k4_ncx1(k_0,k_1,k_2,k_3,n1,mdim) \ + end_load_a_k1m##mdim(k_0) end_acc_nc##n1##_k1m##mdim(k_0)\ + end_load_a_k1m##mdim(k_1) end_acc_nc##n1##_k1m##mdim(k_1)\ + end_load_a_k1m##mdim(k_2) end_acc_nc##n1##_k1m##mdim(k_2)\ + end_load_a_k1m##mdim(k_3) end_acc_nc##n1##_k1m##mdim(k_3) + #define end_kernel_k4_ncx2(k_0,k_1,k_2,k_3,n1,n2,mdim) \ + end_load_a_k1m##mdim(k_0) end_acc_nc##n1##_k1m##mdim(k_0) end_acc_nc##n2##_k1m##mdim(k_0)\ + end_load_a_k1m##mdim(k_1) end_acc_nc##n1##_k1m##mdim(k_1) end_acc_nc##n2##_k1m##mdim(k_1)\ + end_load_a_k1m##mdim(k_2) end_acc_nc##n1##_k1m##mdim(k_2) end_acc_nc##n2##_k1m##mdim(k_2)\ + end_load_a_k1m##mdim(k_3) end_acc_nc##n1##_k1m##mdim(k_3) end_acc_nc##n2##_k1m##mdim(k_3) + #define end_kernel_k4_ncx3(k_0,k_1,k_2,k_3,n1,n2,n3,mdim) \ + end_load_a_k1m##mdim(k_0) end_acc_nc##n1##_k1m##mdim(k_0) end_acc_nc##n2##_k1m##mdim(k_0) end_acc_nc##n3##_k1m##mdim(k_0)\ + end_load_a_k1m##mdim(k_1) end_acc_nc##n1##_k1m##mdim(k_1) end_acc_nc##n2##_k1m##mdim(k_1) end_acc_nc##n3##_k1m##mdim(k_1)\ + end_load_a_k1m##mdim(k_2) end_acc_nc##n1##_k1m##mdim(k_2) end_acc_nc##n2##_k1m##mdim(k_2) end_acc_nc##n3##_k1m##mdim(k_2)\ + end_load_a_k1m##mdim(k_3) end_acc_nc##n1##_k1m##mdim(k_3) end_acc_nc##n2##_k1m##mdim(k_3) end_acc_nc##n3##_k1m##mdim(k_3) + #define end_kernel_k4_ncx4(k_0,k_1,k_2,k_3,n1,n2,n3,n4,mdim) \ + end_load_a_k1m##mdim(k_0) end_acc_nc##n1##_k1m##mdim(k_0) end_acc_nc##n2##_k1m##mdim(k_0) end_acc_nc##n3##_k1m##mdim(k_0) end_acc_nc##n4##_k1m##mdim(k_0)\ + end_load_a_k1m##mdim(k_1) end_acc_nc##n1##_k1m##mdim(k_1) end_acc_nc##n2##_k1m##mdim(k_1) end_acc_nc##n3##_k1m##mdim(k_1) end_acc_nc##n4##_k1m##mdim(k_1)\ + end_load_a_k1m##mdim(k_2) end_acc_nc##n1##_k1m##mdim(k_2) end_acc_nc##n2##_k1m##mdim(k_2) end_acc_nc##n3##_k1m##mdim(k_2) end_acc_nc##n4##_k1m##mdim(k_2)\ + end_load_a_k1m##mdim(k_3) end_acc_nc##n1##_k1m##mdim(k_3) end_acc_nc##n2##_k1m##mdim(k_3) end_acc_nc##n3##_k1m##mdim(k_3) end_acc_nc##n4##_k1m##mdim(k_3) + #define end_kernel_k4_ncx5(k_0,k_1,k_2,k_3,n1,n2,n3,n4,n5,mdim) \ + end_load_a_k1m##mdim(k_0) end_acc_nc##n1##_k1m##mdim(k_0) end_acc_nc##n2##_k1m##mdim(k_0)\ + end_acc_nc##n3##_k1m##mdim(k_0) end_acc_nc##n4##_k1m##mdim(k_0) end_acc_nc##n5##_k1m##mdim(k_0)\ + end_load_a_k1m##mdim(k_1) end_acc_nc##n1##_k1m##mdim(k_1) end_acc_nc##n2##_k1m##mdim(k_1)\ + end_acc_nc##n3##_k1m##mdim(k_1) end_acc_nc##n4##_k1m##mdim(k_1) end_acc_nc##n5##_k1m##mdim(k_1)\ + end_load_a_k1m##mdim(k_2) end_acc_nc##n1##_k1m##mdim(k_2) end_acc_nc##n2##_k1m##mdim(k_2)\ + end_acc_nc##n3##_k1m##mdim(k_2) end_acc_nc##n4##_k1m##mdim(k_2) end_acc_nc##n5##_k1m##mdim(k_2)\ + end_load_a_k1m##mdim(k_3) end_acc_nc##n1##_k1m##mdim(k_3) end_acc_nc##n2##_k1m##mdim(k_3)\ + end_acc_nc##n3##_k1m##mdim(k_3) end_acc_nc##n4##_k1m##mdim(k_3) end_acc_nc##n5##_k1m##mdim(k_3) + #define KERNEL_TAIL_C_n8(mdim) end_kernel_k4_ncx1(0,1,2,3,2,mdim) + #define KERNEL_TAIL_C_n12(mdim) \ + end_kernel_k4_ncx2(0,1,2,3,2,3,mdim) end_kernel_k4_ncx1(4,5,6,7,3,mdim) + #define KERNEL_TAIL_C_n16(mdim) \ + end_kernel_k4_ncx3(0,1,2,3,2,3,4,mdim) end_kernel_k4_ncx2(4,5,6,7,3,4,mdim) end_kernel_k4_ncx1(8,9,10,11,4,mdim) + #define KERNEL_TAIL_C_n20(mdim) \ + end_kernel_k4_ncx4(0,1,2,3,2,3,4,5,mdim) end_kernel_k4_ncx3(4,5,6,7,3,4,5,mdim)\ + end_kernel_k4_ncx2(8,9,10,11,4,5,mdim) end_kernel_k4_ncx1(12,13,14,15,5,mdim) + #define KERNEL_TAIL_C_n24(mdim) \ + end_kernel_k4_ncx5(0,1,2,3,2,3,4,5,6,mdim) end_kernel_k4_ncx4(4,5,6,7,3,4,5,6,mdim) end_kernel_k4_ncx3(8,9,10,11,4,5,6,mdim)\ + end_kernel_k4_ncx2(12,13,14,15,5,6,mdim) end_kernel_k4_ncx1(16,17,18,19,6,mdim) + #define KERNEL_TAIL_R_n4(mdim) \ + end_kernel_k4_ncx1(0,1,2,3,4,mdim) end_kernel_k4_ncx1(4,5,6,7,4,mdim) end_kernel_k4_ncx1(8,9,10,11,4,mdim) + #define KERNEL_TAIL_R_n8(mdim) \ + end_kernel_k4_ncx2(0,1,2,3,4,5,mdim) end_kernel_k4_ncx2(4,5,6,7,4,5,mdim) end_kernel_k4_ncx2(8,9,10,11,4,5,mdim) end_kernel_k4_ncx1(12,13,14,15,5,mdim) + #define KERNEL_TAIL_R_n12(mdim) \ + end_kernel_k4_ncx3(0,1,2,3,4,5,6,mdim) end_kernel_k4_ncx3(4,5,6,7,4,5,6,mdim) end_kernel_k4_ncx3(8,9,10,11,4,5,6,mdim)\ + end_kernel_k4_ncx2(12,13,14,15,5,6,mdim) end_kernel_k4_ncx1(16,17,18,19,6,mdim) + #endif +#else + #define KERNEL_HEAD_C_n8(mdim) "" + #define KERNEL_HEAD_C_n12(mdim) "" + #define KERNEL_HEAD_C_n16(mdim) "" + #define KERNEL_HEAD_C_n20(mdim) "" + #define KERNEL_HEAD_C_n24(mdim) "" + #define KERNEL_HEAD_R_n4(mdim) "" + #define KERNEL_HEAD_R_n8(mdim) "" + #define KERNEL_HEAD_R_n12(mdim) "" + #define KERNEL_TAIL_C_n8(mdim) "" + #define KERNEL_TAIL_C_n12(mdim) "" + #define KERNEL_TAIL_C_n16(mdim) "" + #define KERNEL_TAIL_C_n20(mdim) "" + #define KERNEL_TAIL_C_n24(mdim) "" + #define KERNEL_TAIL_R_n4(mdim) "" + #define KERNEL_TAIL_R_n8(mdim) "" + #define KERNEL_TAIL_R_n12(mdim) "" +#endif +#define KERNEL_HEAD_C_n1(mdim) "" +#define KERNEL_HEAD_C_n2(mdim) "" +#define KERNEL_HEAD_C_n4(mdim) "" +#define KERNEL_TAIL_C_n1(mdim) "" +#define KERNEL_TAIL_C_n2(mdim) "" +#define KERNEL_TAIL_C_n4(mdim) "" /* m = 16 */ /* zmm8-zmm31 for accumulators, zmm1-zmm7 for temporary use, zmm0 for alpha */ #define KERNEL_k1m16n1 \ @@ -15,9 +159,10 @@ #define KERNEL_k1m16n2 KERNEL_h_k1m16n2 "addq $8,%1;" #define KERNEL_h_k1m16n4 KERNEL_h_k1m16n2 "vbroadcastsd 8(%1),%%zmm7; vfmadd231ps %%zmm4,%%zmm7,%%zmm10; vfmadd231ps %%zmm5,%%zmm7,%%zmm11;" #define KERNEL_k1m16n4 KERNEL_h_k1m16n4 "addq $16,%1;" -#define unit_kernel_k1m16n4(c1,c2,c3,c4, ...) \ - "vbroadcastsd ("#__VA_ARGS__"),%%zmm6; vfmadd231ps %%zmm4,%%zmm6,"#c1"; vfmadd231ps %%zmm5,%%zmm6,"#c2";"\ - "vbroadcastsd 8("#__VA_ARGS__"),%%zmm7; vfmadd231ps %%zmm4,%%zmm7,"#c3"; vfmadd231ps %%zmm5,%%zmm7,"#c4";" +#define unit_gen_kernel_k1m16n4(c1,c2,c3,c4,k_no,...) \ + "vbroadcastsd "#k_no"*16 ("#__VA_ARGS__"),%%zmm6; vfmadd231ps %%zmm4,%%zmm6,"#c1"; vfmadd231ps %%zmm5,%%zmm6,"#c2";"\ + "vbroadcastsd "#k_no"*16+8("#__VA_ARGS__"),%%zmm7; vfmadd231ps %%zmm4,%%zmm7,"#c3"; vfmadd231ps %%zmm5,%%zmm7,"#c4";" +#define unit_kernel_k1m16n4(c1,c2,c3,c4, ...) unit_gen_kernel_k1m16n4(c1,c2,c3,c4,0,__VA_ARGS__) #define KERNEL_h_k1m16n8 KERNEL_h_k1m16n4 unit_kernel_k1m16n4(%%zmm12,%%zmm13,%%zmm14,%%zmm15,%1,%%r12,1) #define KERNEL_k1m16n8 KERNEL_h_k1m16n8 "addq $16,%1;" #define KERNEL_h_k1m16n12 KERNEL_h_k1m16n8 unit_kernel_k1m16n4(%%zmm16,%%zmm17,%%zmm18,%%zmm19,%1,%%r12,2) @@ -28,6 +173,12 @@ #define KERNEL_k1m16n20 KERNEL_h_k1m16n20 "addq $16,%%r15;" #define KERNEL_h_k1m16n24 KERNEL_h_k1m16n20 unit_kernel_k1m16n4(%%zmm28,%%zmm29,%%zmm30,%%zmm31,%%r15,%%r12,2) #define KERNEL_k1m16n24 KERNEL_h_k1m16n24 "addq $16,%%r15;" +#define end_load_a_k1m16(k_no) "vmovsldup "#k_no"*64(%0),%%zmm4; vmovshdup "#k_no"*64(%0),%%zmm5;" +#define end_acc_nc2_k1m16(k_no) unit_gen_kernel_k1m16n4(%%zmm12,%%zmm13,%%zmm14,%%zmm15,k_no,%1,%%r12,1) +#define end_acc_nc3_k1m16(k_no) unit_gen_kernel_k1m16n4(%%zmm16,%%zmm17,%%zmm18,%%zmm19,k_no,%1,%%r12,2) +#define end_acc_nc4_k1m16(k_no) unit_gen_kernel_k1m16n4(%%zmm20,%%zmm21,%%zmm22,%%zmm23,k_no,%%r15) +#define end_acc_nc5_k1m16(k_no) unit_gen_kernel_k1m16n4(%%zmm24,%%zmm25,%%zmm26,%%zmm27,k_no,%%r15,%%r12,1) +#define end_acc_nc6_k1m16(k_no) unit_gen_kernel_k1m16n4(%%zmm28,%%zmm29,%%zmm30,%%zmm31,k_no,%%r15,%%r12,2) #define INIT_m16n1 "vpxorq %%zmm8,%%zmm8,%%zmm8;" #define INIT_m16n2 INIT_m16n1 "vpxorq %%zmm9,%%zmm9,%%zmm9;" #define INIT_m16n4 INIT_m16n2 "vpxorq %%zmm10,%%zmm10,%%zmm10;vpxorq %%zmm11,%%zmm11,%%zmm11;" @@ -38,11 +189,19 @@ #define INIT_m16n16 INIT_m16n12 unit_init_m16n4(%%zmm20,%%zmm21,%%zmm22,%%zmm23) #define INIT_m16n20 INIT_m16n16 unit_init_m16n4(%%zmm24,%%zmm25,%%zmm26,%%zmm27) #define INIT_m16n24 INIT_m16n20 unit_init_m16n4(%%zmm28,%%zmm29,%%zmm30,%%zmm31) -#define SAVE_h_m16n1 "vfmadd213ps (%2),%%zmm0,%%zmm8; vmovups %%zmm8,(%2);" -#define unit_save_m16n2(c1,c2) \ +#ifdef TRMMKERNEL + #define SAVE_h_m16n1 "vmulps %%zmm8,%%zmm0,%%zmm8; vmovups %%zmm8,(%2);" + #define unit_save_m16n2(c1,c2) \ + "vunpcklps "#c2","#c1",%%zmm6; vunpckhps "#c2","#c1",%%zmm7; vunpcklpd %%zmm7,%%zmm6,%%zmm4; vunpckhpd %%zmm7,%%zmm6,%%zmm5;"\ + "vmulps %%zmm4,%%zmm0,%%zmm4; vmulps %%zmm5,%%zmm0,%%zmm5;"\ + "vmovups %%zmm4,(%5); vmovups %%zmm5,(%5,%3,1); leaq (%5,%3,2),%5;" +#else + #define SAVE_h_m16n1 "vfmadd213ps (%2),%%zmm0,%%zmm8; vmovups %%zmm8,(%2);" + #define unit_save_m16n2(c1,c2) \ "vunpcklps "#c2","#c1",%%zmm6; vunpckhps "#c2","#c1",%%zmm7; vunpcklpd %%zmm7,%%zmm6,%%zmm4; vunpckhpd %%zmm7,%%zmm6,%%zmm5;"\ "vfmadd213ps (%5),%%zmm0,%%zmm4; vfmadd213ps (%5,%3,1),%%zmm0,%%zmm5;"\ "vmovups %%zmm4,(%5); vmovups %%zmm5,(%5,%3,1); leaq (%5,%3,2),%5;" +#endif #define SAVE_h_m16n2 "movq %2,%5;" unit_save_m16n2(%%zmm8,%%zmm9) #define SAVE_h_m16n4 SAVE_h_m16n2 unit_save_m16n2(%%zmm10,%%zmm11) #define SAVE_h_m16n8 SAVE_h_m16n4 unit_save_m16n2(%%zmm12,%%zmm13) unit_save_m16n2(%%zmm14,%%zmm15) @@ -52,49 +211,62 @@ #define SAVE_h_m16n24 SAVE_h_m16n20 unit_save_m16n2(%%zmm28,%%zmm29) unit_save_m16n2(%%zmm30,%%zmm31) #define SAVE_m16(ndim) SAVE_h_m16n##ndim "addq $64,%2;" #define COMPUTE_m16(ndim) \ - INIT_m16n##ndim\ - "movq %%r13,%4; movq %%r14,%1; leaq (%1,%%r12,2),%%r15; addq %%r12,%%r15; movq %2,%5;"\ - "cmpq $18,%4; jb "#ndim"016162f;"\ + INIT_m16n##ndim START_SET_PAPB(16,ndim)\ + "movq %%r13,%4; movq %2,%5; xorq %%r10,%%r10;"\ + KERNEL_HEAD_C_n##ndim(16)\ + "cmpq $16,%4; jb "#ndim"016162f;"\ #ndim"016161:\n\t"\ + "cmpq $126,%%r10; movq $126,%%r10; cmoveq %3,%%r10;"\ KERNEL_k1m16n##ndim\ KERNEL_k1m16n##ndim\ - KERNEL_k1m16n##ndim\ - "prefetcht1 (%5); prefetcht1 63(%5); addq %3,%5;"\ + "prefetcht1 (%5); subq $63,%5; addq %%r10,%5;"\ KERNEL_k1m16n##ndim\ KERNEL_k1m16n##ndim\ - KERNEL_k1m16n##ndim\ - "prefetcht1 (%8); addq $32,%8;"\ - "subq $6,%4; cmpq $18,%4; jnb "#ndim"016161b;"\ + "prefetcht1 (%6); addq $32,%6;"\ + "subq $4,%4; cmpq $16,%4; jnb "#ndim"016161b;"\ "movq %2,%5;"\ #ndim"016162:\n\t"\ - "testq %4,%4; jz "#ndim"016163f;"\ + "testq %4,%4; jz "#ndim"016164f;"\ + #ndim"016163:\n\t"\ "prefetcht0 (%5); prefetcht0 63(%5); prefetcht0 (%5,%3,1); prefetcht0 63(%5,%3,1);"\ KERNEL_k1m16n##ndim\ - "leaq (%5,%3,2),%5;"\ - "decq %4; jmp "#ndim"016162b;"\ - #ndim"016163:\n\t"\ + "leaq (%5,%3,2),%5; decq %4; jnz "#ndim"016163b;"\ + #ndim"016164:\n\t"\ + KERNEL_TAIL_C_n##ndim(16)\ "prefetcht0 (%%r14); prefetcht0 64(%%r14);"\ - SAVE_m16(ndim) + SAVE_m16(ndim) END_SET_PA(16) /* m = 8 *//* ymm0 for alpha, ymm1-ymm3 for temporary use, ymm4-ymm15 for accumulators */ -#define KERNEL_k1m8n1(b_addr) \ +#define kernel_k1m8n1(b_addr) \ "vmovups (%0),%%ymm1; addq $32,%0;"\ "vbroadcastss ("#b_addr"),%%ymm2; vfmadd231ps %%ymm1,%%ymm2,%%ymm4;"\ "addq $4,"#b_addr";" -#define KERNEL_h_k1m8n2(b_addr) \ +#define kernel_h_k1m8n2(b_addr) \ "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; addq $32,%0;"\ "vbroadcastsd ("#b_addr"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;" -#define KERNEL_k1m8n2(b_addr) KERNEL_h_k1m8n2(b_addr) "addq $8,"#b_addr";" -#define KERNEL_h_k1m8n4(b_addr) \ - KERNEL_h_k1m8n2(b_addr) "vbroadcastsd 8("#b_addr"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;" -#define KERNEL_k1m8n4(b_addr) KERNEL_h_k1m8n4(b_addr) "addq $16,"#b_addr";" -#define unit_kernel_k1m8n4(c1,c2,c3,c4,...) \ - "vbroadcastsd ("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c1"; vfmadd231ps %%ymm2,%%ymm3,"#c2";"\ - "vbroadcastsd 8("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c3"; vfmadd231ps %%ymm2,%%ymm3,"#c4";" -#define KERNEL_h_k1m8n8(b_addr) KERNEL_h_k1m8n4(b_addr) unit_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,b_addr,%%r12,1) -#define KERNEL_k1m8n8(b_addr) KERNEL_h_k1m8n8(b_addr) "addq $16,"#b_addr";" -#define KERNEL_h_k1m8n12(b_addr) KERNEL_h_k1m8n8(b_addr) unit_kernel_k1m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15,b_addr,%%r12,2) -#define KERNEL_k1m8n12(b_addr) KERNEL_h_k1m8n12(b_addr) "addq $16,"#b_addr";" +#define kernel_k1m8n2(b_addr) kernel_h_k1m8n2(b_addr) "addq $8,"#b_addr";" +#define kernel_h_k1m8n4(b_addr) \ + kernel_h_k1m8n2(b_addr) "vbroadcastsd 8("#b_addr"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;" +#define kernel_k1m8n4(b_addr) kernel_h_k1m8n4(b_addr) "addq $16,"#b_addr";" +#define unit_gen_kernel_k1m8n4(c1,c2,c3,c4,k_no,...) \ + "vbroadcastsd "#k_no"*16 ("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c1"; vfmadd231ps %%ymm2,%%ymm3,"#c2";"\ + "vbroadcastsd "#k_no"*16+8("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c3"; vfmadd231ps %%ymm2,%%ymm3,"#c4";" +#define unit_kernel_k1m8n4(c1,c2,c3,c4,...) unit_gen_kernel_k1m8n4(c1,c2,c3,c4,0,__VA_ARGS__) +#define kernel_h_k1m8n8(b_addr) kernel_h_k1m8n4(b_addr) unit_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,b_addr,%%r12,1) +#define kernel_k1m8n8(b_addr) kernel_h_k1m8n8(b_addr) "addq $16,"#b_addr";" +#define kernel_h_k1m8n12(b_addr) kernel_h_k1m8n8(b_addr) unit_kernel_k1m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15,b_addr,%%r12,2) +#define kernel_k1m8n12(b_addr) kernel_h_k1m8n12(b_addr) "addq $16,"#b_addr";" +#define KERNEL_k1m8n1 kernel_k1m8n1(%1) +#define KERNEL_k1m8n2 kernel_k1m8n2(%1) +#define KERNEL_k1m8n4 kernel_k1m8n4(%1) +#define KERNEL_k1m8n8 kernel_k1m8n8(%1) +#define KERNEL_k1m8n12 kernel_k1m8n12(%1) +#define end_load_a_k1m8(k_no) "vmovsldup "#k_no"*32(%0),%%ymm1; vmovshdup "#k_no"*32(%0),%%ymm2;" +#define end_acc_nc2_k1m8(k_no) unit_gen_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,k_no,%1,%%r12,1) +#define end_acc_nc3_k1m8(k_no) unit_gen_kernel_k1m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15,k_no,%1,%%r12,2) +#define end_acc_nc4_k1m8(k_no) unit_gen_kernel_k1m8n4(%%ymm4,%%ymm5,%%ymm6,%%ymm7,k_no,%%r15) +#define end_acc_nc5_k1m8(k_no) unit_gen_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,k_no,%%r15,%%r12,1) +#define end_acc_nc6_k1m8(k_no) unit_gen_kernel_k1m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15,k_no,%%r15,%%r12,2) #define INIT_m8n1 "vpxor %%ymm4,%%ymm4,%%ymm4;" #define INIT_m8n2 INIT_m8n1 "vpxor %%ymm5,%%ymm5,%%ymm5;" #define INIT_m8n4 INIT_m8n2 "vpxor %%ymm6,%%ymm6,%%ymm6;vpxor %%ymm7,%%ymm7,%%ymm7;" @@ -102,12 +274,21 @@ "vpxor "#c1","#c1","#c1";vpxor "#c2","#c2","#c2";vpxor "#c3","#c3","#c3";vpxor "#c4","#c4","#c4";" #define INIT_m8n8 INIT_m8n4 unit_init_m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11) #define INIT_m8n12 INIT_m8n8 unit_init_m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15) -#define SAVE_L_m8n1 "vfmadd213ps (%2),%%ymm0,%%ymm4; vmovups %%ymm4,(%2);" -#define unit_save_m8n2(c1,c2) \ +#ifdef TRMMKERNEL + #define SAVE_L_m8n1 "vmulps %%ymm4,%%ymm0,%%ymm4; vmovups %%ymm4,(%2);" + #define unit_save_m8n2(c1,c2) \ "vunpcklps "#c2","#c1",%%ymm2; vunpckhps "#c2","#c1",%%ymm3;"\ - "vunpcklpd %%ymm3,%%ymm2,%%ymm1;vfmadd213ps (%5), %%ymm0,%%ymm1;vmovups %%ymm1,(%5);"\ - "vunpckhpd %%ymm3,%%ymm2,%%ymm1;vfmadd213ps (%5,%3,1),%%ymm0,%%ymm1;vmovups %%ymm1,(%5,%3,1);"\ + "vunpcklpd %%ymm3,%%ymm2,%%ymm1; vmulps %%ymm1,%%ymm0,%%ymm1; vmovups %%ymm1,(%5);"\ + "vunpckhpd %%ymm3,%%ymm2,%%ymm1; vmulps %%ymm1,%%ymm0,%%ymm1; vmovups %%ymm1,(%5,%3,1);"\ "leaq (%5,%3,2),%5;" +#else + #define SAVE_L_m8n1 "vfmadd213ps (%2),%%ymm0,%%ymm4; vmovups %%ymm4,(%2);" + #define unit_save_m8n2(c1,c2) \ + "vunpcklps "#c2","#c1",%%ymm2; vunpckhps "#c2","#c1",%%ymm3;"\ + "vunpcklpd %%ymm3,%%ymm2,%%ymm1; vfmadd213ps (%5), %%ymm0,%%ymm1; vmovups %%ymm1,(%5);"\ + "vunpckhpd %%ymm3,%%ymm2,%%ymm1; vfmadd213ps (%5,%3,1),%%ymm0,%%ymm1; vmovups %%ymm1,(%5,%3,1);"\ + "leaq (%5,%3,2),%5;" +#endif #define SAVE_L_m8n2 "movq %2,%5;" unit_save_m8n2(%%ymm4,%%ymm5) #define SAVE_L_m8n4 SAVE_L_m8n2 unit_save_m8n2(%%ymm6,%%ymm7) #define SAVE_L_m8n8 SAVE_L_m8n4 unit_save_m8n2(%%ymm8,%%ymm9) unit_save_m8n2(%%ymm10,%%ymm11) @@ -116,53 +297,68 @@ #define SAVE_R_m8n8 SAVE_R_m8n4 unit_save_m8n2(%%ymm8,%%ymm9) unit_save_m8n2(%%ymm10,%%ymm11) #define SAVE_R_m8n12 SAVE_R_m8n8 unit_save_m8n2(%%ymm12,%%ymm13) unit_save_m8n2(%%ymm14,%%ymm15) #define COMPUTE_L_m8(ndim,sim) \ - INIT_m8n##ndim\ - "movq %%r13,%4; movq %%r14,%1;"\ - #ndim""#sim"882:\n\t"\ + INIT_m8n##ndim START_SET_PAPB(8,ndim)\ + "movq %%r13,%4;"\ + KERNEL_HEAD_C_n##ndim(8)\ "testq %4,%4; jz "#ndim""#sim"883f;"\ - KERNEL_k1m8n##ndim(%1)\ - "decq %4; jmp "#ndim""#sim"882b;"\ + #ndim""#sim"882:\n\t"\ + kernel_k1m8n##ndim(%1)\ + "decq %4; jnz "#ndim""#sim"882b;"\ #ndim""#sim"883:\n\t"\ + KERNEL_TAIL_C_n##ndim(8)\ SAVE_L_m8n##ndim "addq $32,%2;" #define COMPUTE_R_m8(ndim,sim) \ - "subq %%r12,%0; subq %%r12,%0;"\ - INIT_m8n##ndim\ - "movq %%r13,%4; leaq (%%r14,%%r12,2),%%r15; addq %%r12,%%r15;"\ - #ndim""#sim"882:\n\t"\ + INIT_m8n##ndim RECOVER_PA(8)\ + "movq %%r13,%4;"\ + KERNEL_HEAD_R_n##ndim(8)\ "testq %4,%4; jz "#ndim""#sim"883f;"\ - KERNEL_k1m8n##ndim(%%r15)\ - "decq %4; jmp "#ndim""#sim"882b;"\ + #ndim""#sim"882:\n\t"\ + kernel_k1m8n##ndim(%%r15)\ + "decq %4; jnz "#ndim""#sim"882b;"\ #ndim""#sim"883:\n\t"\ - SAVE_R_m8n##ndim -#define COMPUTE_m8_n1 COMPUTE_L_m8(1,33833) -#define COMPUTE_m8_n2 COMPUTE_L_m8(2,33833) -#define COMPUTE_m8_n4 COMPUTE_L_m8(4,33833) -#define COMPUTE_m8_n8 COMPUTE_L_m8(8,33833) -#define COMPUTE_m8_n12 COMPUTE_L_m8(12,33833) + KERNEL_TAIL_R_n##ndim(8)\ + SAVE_R_m8n##ndim END_SET_PA(8) +#define COMPUTE_m8_n1 COMPUTE_L_m8(1,33833) END_SET_PA(8) +#define COMPUTE_m8_n2 COMPUTE_L_m8(2,33833) END_SET_PA(8) +#define COMPUTE_m8_n4 COMPUTE_L_m8(4,33833) END_SET_PA(8) +#define COMPUTE_m8_n8 COMPUTE_L_m8(8,33833) END_SET_PA(8) +#define COMPUTE_m8_n12 COMPUTE_L_m8(12,33833) END_SET_PA(8) #define COMPUTE_m8_n16 COMPUTE_L_m8(12,33733) COMPUTE_R_m8(4,33933) #define COMPUTE_m8_n20 COMPUTE_L_m8(12,33633) COMPUTE_R_m8(8,33933) #define COMPUTE_m8_n24 COMPUTE_L_m8(12,33533) COMPUTE_R_m8(12,33933) #define COMPUTE_m8(ndim) COMPUTE_m8_n##ndim /* m = 4 *//* xmm0 for alpha, xmm1-xmm3 for temporary use, xmm4-xmm15 for accumulators */ -#define KERNEL_k1m4n1(b_addr) \ +#define kernel_k1m4n1(b_addr) \ "vmovups (%0),%%xmm1; addq $16,%0;"\ "vbroadcastss ("#b_addr"),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\ "addq $4,"#b_addr";" -#define KERNEL_h_k1m4n2(b_addr) \ +#define kernel_h_k1m4n2(b_addr) \ "vmovsldup (%0),%%xmm1; vmovshdup (%0),%%xmm2; addq $16,%0;"\ "vmovddup ("#b_addr"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm4; vfmadd231ps %%xmm2,%%xmm3,%%xmm5;" -#define KERNEL_k1m4n2(b_addr) KERNEL_h_k1m4n2(b_addr) "addq $8,"#b_addr";" -#define KERNEL_h_k1m4n4(b_addr) \ - KERNEL_h_k1m4n2(b_addr) "vmovddup 8("#b_addr"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm6; vfmadd231ps %%xmm2,%%xmm3,%%xmm7;" -#define KERNEL_k1m4n4(b_addr) KERNEL_h_k1m4n4(b_addr) "addq $16,"#b_addr";" -#define unit_kernel_k1m4n4(c1,c2,c3,c4,...) \ - "vmovddup ("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c1"; vfmadd231ps %%xmm2,%%xmm3,"#c2";"\ - "vmovddup 8("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c3"; vfmadd231ps %%xmm2,%%xmm3,"#c4";" -#define KERNEL_h_k1m4n8(b_addr) KERNEL_h_k1m4n4(b_addr) unit_kernel_k1m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11,b_addr,%%r12,1) -#define KERNEL_k1m4n8(b_addr) KERNEL_h_k1m4n8(b_addr) "addq $16,"#b_addr";" -#define KERNEL_h_k1m4n12(b_addr) KERNEL_h_k1m4n8(b_addr) unit_kernel_k1m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15,b_addr,%%r12,2) -#define KERNEL_k1m4n12(b_addr) KERNEL_h_k1m4n12(b_addr) "addq $16,"#b_addr";" +#define kernel_k1m4n2(b_addr) kernel_h_k1m4n2(b_addr) "addq $8,"#b_addr";" +#define kernel_h_k1m4n4(b_addr) \ + kernel_h_k1m4n2(b_addr) "vmovddup 8("#b_addr"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm6; vfmadd231ps %%xmm2,%%xmm3,%%xmm7;" +#define kernel_k1m4n4(b_addr) kernel_h_k1m4n4(b_addr) "addq $16,"#b_addr";" +#define unit_gen_kernel_k1m4n4(c1,c2,c3,c4,k_no,...) \ + "vmovddup "#k_no"*16 ("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c1"; vfmadd231ps %%xmm2,%%xmm3,"#c2";"\ + "vmovddup "#k_no"*16+8("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c3"; vfmadd231ps %%xmm2,%%xmm3,"#c4";" +#define unit_kernel_k1m4n4(c1,c2,c3,c4,...) unit_gen_kernel_k1m4n4(c1,c2,c3,c4,0,__VA_ARGS__) +#define kernel_h_k1m4n8(b_addr) kernel_h_k1m4n4(b_addr) unit_kernel_k1m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11,b_addr,%%r12,1) +#define kernel_k1m4n8(b_addr) kernel_h_k1m4n8(b_addr) "addq $16,"#b_addr";" +#define kernel_h_k1m4n12(b_addr) kernel_h_k1m4n8(b_addr) unit_kernel_k1m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15,b_addr,%%r12,2) +#define kernel_k1m4n12(b_addr) kernel_h_k1m4n12(b_addr) "addq $16,"#b_addr";" +#define KERNEL_k1m4n1 kernel_k1m4n1(%1) +#define KERNEL_k1m4n2 kernel_k1m4n2(%1) +#define KERNEL_k1m4n4 kernel_k1m4n4(%1) +#define KERNEL_k1m4n8 kernel_k1m4n8(%1) +#define KERNEL_k1m4n12 kernel_k1m4n12(%1) +#define end_load_a_k1m4(k_no) "vmovsldup "#k_no"*16(%0),%%xmm1; vmovshdup "#k_no"*16(%0),%%xmm2;" +#define end_acc_nc2_k1m4(k_no) unit_gen_kernel_k1m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11,k_no,%1,%%r12,1) +#define end_acc_nc3_k1m4(k_no) unit_gen_kernel_k1m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15,k_no,%1,%%r12,2) +#define end_acc_nc4_k1m4(k_no) unit_gen_kernel_k1m4n4(%%xmm4,%%xmm5,%%xmm6,%%xmm7,k_no,%%r15) +#define end_acc_nc5_k1m4(k_no) unit_gen_kernel_k1m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11,k_no,%%r15,%%r12,1) +#define end_acc_nc6_k1m4(k_no) unit_gen_kernel_k1m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15,k_no,%%r15,%%r12,2) #define INIT_m4n1 "vpxor %%xmm4,%%xmm4,%%xmm4;" #define INIT_m4n2 INIT_m4n1 "vpxor %%xmm5,%%xmm5,%%xmm5;" #define INIT_m4n4 INIT_m4n2 "vpxor %%xmm6,%%xmm6,%%xmm6;vpxor %%xmm7,%%xmm7,%%xmm7;" @@ -170,12 +366,21 @@ "vpxor "#c1","#c1","#c1";vpxor "#c2","#c2","#c2";vpxor "#c3","#c3","#c3";vpxor "#c4","#c4","#c4";" #define INIT_m4n8 INIT_m4n4 unit_init_m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11) #define INIT_m4n12 INIT_m4n8 unit_init_m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15) -#define SAVE_L_m4n1 "vfmadd213ps (%2),%%xmm0,%%xmm4; vmovups %%xmm4,(%2);" -#define unit_save_m4n2(c1,c2) \ +#ifdef TRMMKERNEL + #define SAVE_L_m4n1 "vmulps %%xmm4,%%xmm0,%%xmm4; vmovups %%xmm4,(%2);" + #define unit_save_m4n2(c1,c2) \ + "vunpcklps "#c2","#c1",%%xmm2; vunpckhps "#c2","#c1",%%xmm3;"\ + "vunpcklpd %%xmm3,%%xmm2,%%xmm1;vmulps %%xmm1,%%xmm0,%%xmm1;vmovups %%xmm1,(%5);"\ + "vunpckhpd %%xmm3,%%xmm2,%%xmm1;vmulps %%xmm1,%%xmm0,%%xmm1;vmovups %%xmm1,(%5,%3,1);"\ + "leaq (%5,%3,2),%5;" +#else + #define SAVE_L_m4n1 "vfmadd213ps (%2),%%xmm0,%%xmm4; vmovups %%xmm4,(%2);" + #define unit_save_m4n2(c1,c2) \ "vunpcklps "#c2","#c1",%%xmm2; vunpckhps "#c2","#c1",%%xmm3;"\ "vunpcklpd %%xmm3,%%xmm2,%%xmm1;vfmadd213ps (%5), %%xmm0,%%xmm1;vmovups %%xmm1,(%5);"\ "vunpckhpd %%xmm3,%%xmm2,%%xmm1;vfmadd213ps (%5,%3,1),%%xmm0,%%xmm1;vmovups %%xmm1,(%5,%3,1);"\ "leaq (%5,%3,2),%5;" +#endif #define SAVE_L_m4n2 "movq %2,%5;" unit_save_m4n2(%%xmm4,%%xmm5) #define SAVE_L_m4n4 SAVE_L_m4n2 unit_save_m4n2(%%xmm6,%%xmm7) #define SAVE_L_m4n8 SAVE_L_m4n4 unit_save_m4n2(%%xmm8,%%xmm9) unit_save_m4n2(%%xmm10,%%xmm11) @@ -184,222 +389,281 @@ #define SAVE_R_m4n8 SAVE_R_m4n4 unit_save_m4n2(%%xmm8,%%xmm9) unit_save_m4n2(%%xmm10,%%xmm11) #define SAVE_R_m4n12 SAVE_R_m4n8 unit_save_m4n2(%%xmm12,%%xmm13) unit_save_m4n2(%%xmm14,%%xmm15) #define COMPUTE_L_m4(ndim,sim) \ - INIT_m4n##ndim\ - "movq %%r13,%4; movq %%r14,%1;"\ - #ndim""#sim"442:\n\t"\ + INIT_m4n##ndim START_SET_PAPB(4,ndim)\ + "movq %%r13,%4;"\ + KERNEL_HEAD_C_n##ndim(4)\ "testq %4,%4; jz "#ndim""#sim"443f;"\ - KERNEL_k1m4n##ndim(%1)\ - "decq %4; jmp "#ndim""#sim"442b;"\ + #ndim""#sim"442:\n\t"\ + kernel_k1m4n##ndim(%1)\ + "decq %4; jnz "#ndim""#sim"442b;"\ #ndim""#sim"443:\n\t"\ + KERNEL_TAIL_C_n##ndim(4)\ SAVE_L_m4n##ndim "addq $16,%2;" #define COMPUTE_R_m4(ndim,sim) \ - "subq %%r12,%0;"\ - INIT_m4n##ndim\ - "movq %%r13,%4; leaq (%%r14,%%r12,2),%%r15; addq %%r12,%%r15;"\ - #ndim""#sim"442:\n\t"\ + INIT_m4n##ndim RECOVER_PA(4)\ + "movq %%r13,%4;"\ + KERNEL_HEAD_R_n##ndim(4)\ "testq %4,%4; jz "#ndim""#sim"443f;"\ - KERNEL_k1m4n##ndim(%%r15)\ - "decq %4; jmp "#ndim""#sim"442b;"\ + #ndim""#sim"442:\n\t"\ + kernel_k1m4n##ndim(%%r15)\ + "decq %4; jnz "#ndim""#sim"442b;"\ #ndim""#sim"443:\n\t"\ - SAVE_R_m4n##ndim -#define COMPUTE_m4_n1 COMPUTE_L_m4(1,55855) -#define COMPUTE_m4_n2 COMPUTE_L_m4(2,55855) -#define COMPUTE_m4_n4 COMPUTE_L_m4(4,55855) -#define COMPUTE_m4_n8 COMPUTE_L_m4(8,55855) -#define COMPUTE_m4_n12 COMPUTE_L_m4(12,55855) + KERNEL_TAIL_R_n##ndim(4)\ + SAVE_R_m4n##ndim END_SET_PA(4) +#define COMPUTE_m4_n1 COMPUTE_L_m4(1,55855) END_SET_PA(4) +#define COMPUTE_m4_n2 COMPUTE_L_m4(2,55855) END_SET_PA(4) +#define COMPUTE_m4_n4 COMPUTE_L_m4(4,55855) END_SET_PA(4) +#define COMPUTE_m4_n8 COMPUTE_L_m4(8,55855) END_SET_PA(4) +#define COMPUTE_m4_n12 COMPUTE_L_m4(12,55855) END_SET_PA(4) #define COMPUTE_m4_n16 COMPUTE_L_m4(12,55755) COMPUTE_R_m4(4,55955) #define COMPUTE_m4_n20 COMPUTE_L_m4(12,55655) COMPUTE_R_m4(8,55955) #define COMPUTE_m4_n24 COMPUTE_L_m4(12,55555) COMPUTE_R_m4(12,55955) #define COMPUTE_m4(ndim) COMPUTE_m4_n##ndim -/* m = 2 *//* xmm0 for alpha, xmm1-xmm3 and xmm10 for temporary use, xmm4-xmm9 for accumulators */ +/* m = 2 *//* xmm0 for alpha, xmm1-xmm3 for temporary use, xmm4-xmm15 for accumulators */ #define INIT_m2n1 "vpxor %%xmm4,%%xmm4,%%xmm4;" -#define KERNEL_k1m2n1(b_addr) \ +#define KERNEL_k1m2n1 \ "vmovsd (%0),%%xmm1; addq $8,%0;"\ - "vbroadcastss ("#b_addr"),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\ - "addq $4,"#b_addr";" -#define SAVE_L_m2n1 "vmovsd (%2),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm4; vmovsd %%xmm4,(%2);" + "vbroadcastss (%1),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\ + "addq $4,%1;" #define INIT_m2n2 INIT_m2n1 "vpxor %%xmm5,%%xmm5,%%xmm5;" -#define KERNEL_k1m2n2(b_addr) \ +#define KERNEL_k1m2n2 \ "vmovsd (%0),%%xmm1; addq $8,%0;"\ - "vbroadcastss ("#b_addr"),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\ - "vbroadcastss 4("#b_addr"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm5;"\ - "addq $8,"#b_addr";" -#define SAVE_L_m2n2 SAVE_L_m2n1 "vmovsd (%2,%3,1),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm5; vmovsd %%xmm5,(%2,%3,1);" + "vbroadcastss (%1),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\ + "vbroadcastss 4(%1),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm5;"\ + "addq $8,%1;" +#ifdef TRMMKERNEL + #define SAVE_h_m2n1 "vmulps %%xmm4,%%xmm0,%%xmm4; vmovsd %%xmm4,(%2);" + #define SAVE_h_m2n2 SAVE_h_m2n1 "vmulps %%xmm5,%%xmm0,%%xmm5; vmovsd %%xmm5,(%2,%3,1);" +#else + #define SAVE_h_m2n1 "vmovsd (%2),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm4; vmovsd %%xmm4,(%2);" + #define SAVE_h_m2n2 SAVE_h_m2n1 "vmovsd (%2,%3,1),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm5; vmovsd %%xmm5,(%2,%3,1);" +#endif #define INIT_m2n4 INIT_m2n2 #define INIT_m2n8 INIT_m2n4 "vpxor %%xmm6,%%xmm6,%%xmm6; vpxor %%xmm7,%%xmm7,%%xmm7;" #define INIT_m2n12 INIT_m2n8 "vpxor %%xmm8,%%xmm8,%%xmm8; vpxor %%xmm9,%%xmm9,%%xmm9;" -#define KERNEL_k1m2n4(b_addr) \ - "vmovups ("#b_addr"),%%xmm3; addq $16,"#b_addr";"\ - "vbroadcastss (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4;"\ - "vbroadcastss 4(%0),%%xmm2; vfmadd231ps %%xmm3,%%xmm2,%%xmm5;"\ - "addq $8,%0;" -#define KERNEL_k1m2n8(b_addr) \ - "vmovups ("#b_addr"),%%xmm3; vmovups ("#b_addr",%%r12,1),%%xmm2; addq $16,"#b_addr";"\ - "vbroadcastss (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4; vfmadd231ps %%xmm2,%%xmm1,%%xmm6;"\ - "vbroadcastss 4(%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm5; vfmadd231ps %%xmm2,%%xmm1,%%xmm7;"\ - "addq $8,%0;" -#define KERNEL_k1m2n12(b_addr) \ - "vmovups ("#b_addr"),%%xmm3; vmovups ("#b_addr",%%r12,1),%%xmm2; vmovups ("#b_addr",%%r12,2),%%xmm1; addq $16,"#b_addr";"\ - "vbroadcastss (%0),%%xmm10; vfmadd231ps %%xmm3,%%xmm10,%%xmm4; vfmadd231ps %%xmm2,%%xmm10,%%xmm6; vfmadd231ps %%xmm1,%%xmm10,%%xmm8;"\ - "vbroadcastss 4(%0),%%xmm10; vfmadd231ps %%xmm3,%%xmm10,%%xmm5; vfmadd231ps %%xmm2,%%xmm10,%%xmm7; vfmadd231ps %%xmm1,%%xmm10,%%xmm9;"\ - "addq $8,%0;" -#define unit_save_m2n4(c1,c2) \ +#define INIT_m2n16 INIT_m2n12 "vpxor %%xmm10,%%xmm10,%%xmm10; vpxor %%xmm11,%%xmm11,%%xmm11;" +#define INIT_m2n20 INIT_m2n16 "vpxor %%xmm12,%%xmm12,%%xmm12; vpxor %%xmm13,%%xmm13,%%xmm13;" +#define INIT_m2n24 INIT_m2n20 "vpxor %%xmm14,%%xmm14,%%xmm14; vpxor %%xmm15,%%xmm15,%%xmm15;" +#define unit_gen_kernel_k1m2n4(c1,c2,k_no,...) \ + "vmovups "#k_no"*16("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c1"; vfmadd231ps %%xmm2,%%xmm3,"#c2";" +#define KERNEL_h_k1m2n4 \ + "vbroadcastss (%0),%%xmm1; vbroadcastss 4(%0),%%xmm2; addq $8,%0;" unit_gen_kernel_k1m2n4(%%xmm4,%%xmm5,0,%1) +#define KERNEL_k1m2n4 KERNEL_h_k1m2n4 "addq $16,%1;" +#define KERNEL_h_k1m2n8 KERNEL_h_k1m2n4 unit_gen_kernel_k1m2n4(%%xmm6,%%xmm7,0,%1,%%r12,1) +#define KERNEL_k1m2n8 KERNEL_h_k1m2n8 "addq $16,%1;" +#define KERNEL_k1m2n12 KERNEL_h_k1m2n8 unit_gen_kernel_k1m2n4(%%xmm8,%%xmm9,0,%1,%%r12,2) "addq $16,%1;" +#define KERNEL_h_k1m2n16 KERNEL_k1m2n12 unit_gen_kernel_k1m2n4(%%xmm10,%%xmm11,0,%%r15) +#define KERNEL_k1m2n16 KERNEL_h_k1m2n16 "addq $16,%%r15;" +#define KERNEL_h_k1m2n20 KERNEL_h_k1m2n16 unit_gen_kernel_k1m2n4(%%xmm12,%%xmm13,0,%%r15,%%r12,1) +#define KERNEL_k1m2n20 KERNEL_h_k1m2n20 "addq $16,%%r15;" +#define KERNEL_h_k1m2n24 KERNEL_h_k1m2n20 unit_gen_kernel_k1m2n4(%%xmm14,%%xmm15,0,%%r15,%%r12,2) +#define KERNEL_k1m2n24 KERNEL_h_k1m2n24 "addq $16,%%r15;" +#define end_load_a_k1m2(k_no) "vbroadcastss "#k_no"*8(%0),%%xmm1; vbroadcastss "#k_no"*8+4(%0),%%xmm2;" +#define end_acc_nc2_k1m2(k_no) unit_gen_kernel_k1m2n4(%%xmm6,%%xmm7,k_no,%1,%%r12,1) +#define end_acc_nc3_k1m2(k_no) unit_gen_kernel_k1m2n4(%%xmm8,%%xmm9,k_no,%1,%%r12,2) +#define end_acc_nc4_k1m2(k_no) unit_gen_kernel_k1m2n4(%%xmm10,%%xmm11,k_no,%%r15) +#define end_acc_nc5_k1m2(k_no) unit_gen_kernel_k1m2n4(%%xmm12,%%xmm13,k_no,%%r15,%%r12,1) +#define end_acc_nc6_k1m2(k_no) unit_gen_kernel_k1m2n4(%%xmm14,%%xmm15,k_no,%%r15,%%r12,2) +#ifdef TRMMKERNEL + #define unit_save_m2n4(c1,c2) \ + "vunpcklps "#c2","#c1",%%xmm1; vunpckhps "#c2","#c1",%%xmm2;"\ + "vmulps %%xmm1,%%xmm0,%%xmm1; vmovsd %%xmm1,(%5); vmovhpd %%xmm1,(%5,%3,1);"\ + "leaq (%5,%3,2),%5;"\ + "vmulps %%xmm2,%%xmm0,%%xmm2; vmovsd %%xmm2,(%5); vmovhpd %%xmm2,(%5,%3,1);"\ + "leaq (%5,%3,2),%5;" +#else + #define unit_save_m2n4(c1,c2) \ "vunpcklps "#c2","#c1",%%xmm1; vunpckhps "#c2","#c1",%%xmm2;"\ "vmovsd (%5),%%xmm3; vmovhpd (%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm1; vmovsd %%xmm1,(%5); vmovhpd %%xmm1,(%5,%3,1);"\ "leaq (%5,%3,2),%5;"\ "vmovsd (%5),%%xmm3; vmovhpd (%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm2; vmovsd %%xmm2,(%5); vmovhpd %%xmm2,(%5,%3,1);"\ "leaq (%5,%3,2),%5;" -#define SAVE_L_m2n4 "movq %2,%5;" unit_save_m2n4(%%xmm4,%%xmm5) -#define SAVE_L_m2n8 SAVE_L_m2n4 unit_save_m2n4(%%xmm6,%%xmm7) -#define SAVE_L_m2n12 SAVE_L_m2n8 unit_save_m2n4(%%xmm8,%%xmm9) -#define SAVE_R_m2n4 unit_save_m2n4(%%xmm4,%%xmm5) -#define SAVE_R_m2n8 SAVE_R_m2n4 unit_save_m2n4(%%xmm6,%%xmm7) -#define SAVE_R_m2n12 SAVE_R_m2n8 unit_save_m2n4(%%xmm8,%%xmm9) -#define COMPUTE_L_m2(ndim,sim) \ - INIT_m2n##ndim\ - "movq %%r13,%4; movq %%r14,%1;"\ - #ndim""#sim"222:\n\t"\ - "testq %4,%4; jz "#ndim""#sim"223f;"\ - KERNEL_k1m2n##ndim(%1)\ - "decq %4; jmp "#ndim""#sim"222b;"\ - #ndim""#sim"223:\n\t"\ - SAVE_L_m2n##ndim "addq $8,%2;" -#define COMPUTE_R_m2(ndim,sim) \ - "salq $3,%%r13;subq %%r13,%0;sarq $3,%%r13;"\ - INIT_m2n##ndim\ - "movq %%r13,%4; leaq (%%r14,%%r12,2),%%r15; addq %%r12,%%r15;"\ - #ndim""#sim"222:\n\t"\ - "testq %4,%4; jz "#ndim""#sim"223f;"\ - KERNEL_k1m2n##ndim(%%r15)\ - "decq %4; jmp "#ndim""#sim"222b;"\ - #ndim""#sim"223:\n\t"\ - SAVE_R_m2n##ndim -#define COMPUTE_m2_n1 COMPUTE_L_m2(1,77877) -#define COMPUTE_m2_n2 COMPUTE_L_m2(2,77877) -#define COMPUTE_m2_n4 COMPUTE_L_m2(4,77877) -#define COMPUTE_m2_n8 COMPUTE_L_m2(8,77877) -#define COMPUTE_m2_n12 COMPUTE_L_m2(12,77877) -#define COMPUTE_m2_n16 COMPUTE_L_m2(12,77777) COMPUTE_R_m2(4,77977) -#define COMPUTE_m2_n20 COMPUTE_L_m2(12,77677) COMPUTE_R_m2(8,77977) -#define COMPUTE_m2_n24 COMPUTE_L_m2(12,77577) COMPUTE_R_m2(12,77977) -#define COMPUTE_m2(ndim) COMPUTE_m2_n##ndim +#endif +#define SAVE_h_m2n4 "movq %2,%5;" unit_save_m2n4(%%xmm4,%%xmm5) +#define SAVE_h_m2n8 SAVE_h_m2n4 unit_save_m2n4(%%xmm6,%%xmm7) +#define SAVE_h_m2n12 SAVE_h_m2n8 unit_save_m2n4(%%xmm8,%%xmm9) +#define SAVE_h_m2n16 SAVE_h_m2n12 unit_save_m2n4(%%xmm10,%%xmm11) +#define SAVE_h_m2n20 SAVE_h_m2n16 unit_save_m2n4(%%xmm12,%%xmm13) +#define SAVE_h_m2n24 SAVE_h_m2n20 unit_save_m2n4(%%xmm14,%%xmm15) +#define SAVE_m2(ndim) SAVE_h_m2n##ndim "addq $8,%2;" +#define COMPUTE_m2(ndim) \ + INIT_m2n##ndim START_SET_PAPB(2,ndim)\ + "movq %%r13,%4;"\ + KERNEL_HEAD_C_n##ndim(2)\ + "testq %4,%4; jz "#ndim"002022f;"\ + #ndim"002021:\n\t"\ + KERNEL_k1m2n##ndim "decq %4; jnz "#ndim"002021b;"\ + #ndim"002022:\n\t"\ + KERNEL_TAIL_C_n##ndim(2)\ + SAVE_m2(ndim) END_SET_PA(2) -/* m = 1 *//* xmm0 for alpha, xmm1-xmm3 and xmm10 for temporary use, xmm4-xmm6 for accumulators */ +/* m = 1 *//* xmm0 for alpha, xmm1-xmm3 and xmm10 for temporary use, xmm4-xmm9 for accumulators */ #define INIT_m1n1 "vpxor %%xmm4,%%xmm4,%%xmm4;" -#define KERNEL_k1m1n1(b_addr) \ - "vmovss ("#b_addr"),%%xmm3; addq $4,"#b_addr";"\ +#define KERNEL_k1m1n1 \ + "vmovss (%1),%%xmm3; addq $4,%1;"\ "vmovss (%0),%%xmm1; vfmadd231ss %%xmm3,%%xmm1,%%xmm4;"\ "addq $4,%0;" -#define SAVE_L_m1n1 "vfmadd213ss (%2),%%xmm0,%%xmm4; vmovss %%xmm4,(%2);" +#ifdef TRMMKERNEL + #define SAVE_h_m1n1 "vmulss %%xmm4,%%xmm0,%%xmm4; vmovss %%xmm4,(%2);" +#else + #define SAVE_h_m1n1 "vfmadd213ss (%2),%%xmm0,%%xmm4; vmovss %%xmm4,(%2);" +#endif #define INIT_m1n2 INIT_m1n1 -#define KERNEL_k1m1n2(b_addr) \ - "vmovsd ("#b_addr"),%%xmm3; addq $8,"#b_addr";"\ +#define KERNEL_k1m1n2 \ + "vmovsd (%1),%%xmm3; addq $8,%1;"\ "vbroadcastss (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4;"\ "addq $4,%0;" -#define SAVE_L_m1n2 \ +#ifdef TRMMKERNEL + #define SAVE_h_m1n2 \ + "vmulps %%xmm4,%%xmm0,%%xmm4;"\ + "vmovss %%xmm4,(%2); vextractps $1,%%xmm4,(%2,%3,1);" +#else + #define SAVE_h_m1n2 \ "vmovss (%2),%%xmm3; vinsertps $16,(%2,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm4;"\ "vmovss %%xmm4,(%2); vextractps $1,%%xmm4,(%2,%3,1);" +#endif #define INIT_m1n4 INIT_m1n2 #define INIT_m1n8 INIT_m1n4 "vpxor %%xmm5,%%xmm5,%%xmm5;" #define INIT_m1n12 INIT_m1n8 "vpxor %%xmm6,%%xmm6,%%xmm6;" -#define KERNEL_k1m1n4(b_addr) \ - "vmovups ("#b_addr"),%%xmm3; addq $16,"#b_addr";"\ - "vbroadcastss (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4;"\ - "addq $4,%0;" -#define KERNEL_k1m1n8(b_addr) \ - "vmovups ("#b_addr"),%%xmm3; vmovups ("#b_addr",%%r12,1),%%xmm2; addq $16,"#b_addr";"\ - "vbroadcastss (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4; vfmadd231ps %%xmm2,%%xmm1,%%xmm5;"\ - "addq $4,%0;" -#define KERNEL_k1m1n12(b_addr) \ - "vmovups ("#b_addr"),%%xmm3; vmovups ("#b_addr",%%r12,1),%%xmm2; vmovups ("#b_addr",%%r12,2),%%xmm1; addq $16,"#b_addr";"\ - "vbroadcastss (%0),%%xmm10; vfmadd231ps %%xmm3,%%xmm10,%%xmm4; vfmadd231ps %%xmm2,%%xmm10,%%xmm5; vfmadd231ps %%xmm1,%%xmm10,%%xmm6;"\ - "addq $4,%0;" -#define unit_save_m1n4(c1) \ +#define INIT_m1n16 INIT_m1n12 "vpxor %%xmm7,%%xmm7,%%xmm7;" +#define INIT_m1n20 INIT_m1n16 "vpxor %%xmm8,%%xmm8,%%xmm8;" +#define INIT_m1n24 INIT_m1n20 "vpxor %%xmm9,%%xmm9,%%xmm9;" +#define KERNEL_h_k1m1n4 \ + "vbroadcastss (%0),%%xmm1; addq $4,%0; vfmadd231ps (%1),%%xmm1,%%xmm4;" +#define KERNEL_k1m1n4 KERNEL_h_k1m1n4 "addq $16,%1;" +#define KERNEL_h_k1m1n8 KERNEL_h_k1m1n4 "vfmadd231ps (%1,%%r12,1),%%xmm1,%%xmm5;" +#define KERNEL_k1m1n8 KERNEL_h_k1m1n8 "addq $16,%1;" +#define KERNEL_k1m1n12 KERNEL_h_k1m1n8 "vfmadd231ps (%1,%%r12,2),%%xmm1,%%xmm6; addq $16,%1;" +#define KERNEL_h_k1m1n16 KERNEL_k1m1n12 "vfmadd231ps (%%r15),%%xmm1,%%xmm7;" +#define KERNEL_k1m1n16 KERNEL_h_k1m1n16 "addq $16,%%r15;" +#define KERNEL_h_k1m1n20 KERNEL_h_k1m1n16 "vfmadd231ps (%%r15,%%r12,1),%%xmm1,%%xmm8;" +#define KERNEL_k1m1n20 KERNEL_h_k1m1n20 "addq $16,%%r15;" +#define KERNEL_h_k1m1n24 KERNEL_h_k1m1n20 "vfmadd231ps (%%r15,%%r12,2),%%xmm1,%%xmm9;" +#define KERNEL_k1m1n24 KERNEL_h_k1m1n24 "addq $16,%%r15;" +#define end_load_a_k1m1(k_no) "vbroadcastss "#k_no"*4(%0),%%xmm1;" +#define end_acc_nc2_k1m1(k_no) "vfmadd231ps "#k_no"*16(%1,%%r12,1),%%xmm1,%%xmm5;" +#define end_acc_nc3_k1m1(k_no) "vfmadd231ps "#k_no"*16(%1,%%r12,2),%%xmm1,%%xmm6;" +#define end_acc_nc4_k1m1(k_no) "vfmadd231ps "#k_no"*16(%%r15),%%xmm1,%%xmm7;" +#define end_acc_nc5_k1m1(k_no) "vfmadd231ps "#k_no"*16(%%r15,%%r12,1),%%xmm1,%%xmm8;" +#define end_acc_nc6_k1m1(k_no) "vfmadd231ps "#k_no"*16(%%r15,%%r12,2),%%xmm1,%%xmm9;" +#ifdef TRMMKERNEL + #define unit_save_m1n4(c1) \ + "vmulps "#c1",%%xmm0,"#c1"; vpxor %%xmm10,%%xmm10,%%xmm10; vmovsd "#c1",%%xmm10,%%xmm2; vmovhlps "#c1",%%xmm10,%%xmm1;"\ + "vmovss %%xmm2,(%5); vextractps $1,%%xmm2,(%5,%3,1); leaq (%5,%3,2),%5;"\ + "vmovss %%xmm1,(%5); vextractps $1,%%xmm1,(%5,%3,1); leaq (%5,%3,2),%5;" +#else + #define unit_save_m1n4(c1) \ "vpxor %%xmm10,%%xmm10,%%xmm10; vmovsd "#c1",%%xmm10,%%xmm2; vmovhlps "#c1",%%xmm10,%%xmm1;"\ "vmovss (%5),%%xmm3; vinsertps $16,(%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm2;"\ "vmovss %%xmm2,(%5); vextractps $1,%%xmm2,(%5,%3,1); leaq (%5,%3,2),%5;"\ "vmovss (%5),%%xmm3; vinsertps $16,(%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm1;"\ "vmovss %%xmm1,(%5); vextractps $1,%%xmm1,(%5,%3,1); leaq (%5,%3,2),%5;" -#define SAVE_L_m1n4 "movq %2,%5;" unit_save_m1n4(%%xmm4) -#define SAVE_L_m1n8 SAVE_L_m1n4 unit_save_m1n4(%%xmm5) -#define SAVE_L_m1n12 SAVE_L_m1n8 unit_save_m1n4(%%xmm6) -#define SAVE_R_m1n4 unit_save_m1n4(%%xmm4) -#define SAVE_R_m1n8 SAVE_R_m1n4 unit_save_m1n4(%%xmm5) -#define SAVE_R_m1n12 SAVE_R_m1n8 unit_save_m1n4(%%xmm6) -#define COMPUTE_L_m1(ndim,sim) \ - INIT_m1n##ndim\ - "movq %%r13,%4; movq %%r14,%1;"\ - #ndim""#sim"112:\n\t"\ - "testq %4,%4; jz "#ndim""#sim"113f;"\ - KERNEL_k1m1n##ndim(%1)\ - "decq %4; jmp "#ndim""#sim"112b;"\ - #ndim""#sim"113:\n\t"\ - SAVE_L_m1n##ndim "addq $4,%2;" -#define COMPUTE_R_m1(ndim,sim) \ - "salq $2,%%r13;subq %%r13,%0;sarq $2,%%r13;"\ - INIT_m1n##ndim\ - "movq %%r13,%4; leaq (%%r14,%%r12,2),%%r15; addq %%r12,%%r15;"\ - #ndim""#sim"112:\n\t"\ - "testq %4,%4; jz "#ndim""#sim"113f;"\ - KERNEL_k1m1n##ndim(%%r15)\ - "decq %4; jmp "#ndim""#sim"112b;"\ - #ndim""#sim"113:\n\t"\ - SAVE_R_m1n##ndim -#define COMPUTE_m1_n1 COMPUTE_L_m1(1,99899) -#define COMPUTE_m1_n2 COMPUTE_L_m1(2,99899) -#define COMPUTE_m1_n4 COMPUTE_L_m1(4,99899) -#define COMPUTE_m1_n8 COMPUTE_L_m1(8,99899) -#define COMPUTE_m1_n12 COMPUTE_L_m1(12,99899) -#define COMPUTE_m1_n16 COMPUTE_L_m1(12,99799) COMPUTE_R_m1(4,99999) -#define COMPUTE_m1_n20 COMPUTE_L_m1(12,99699) COMPUTE_R_m1(8,99999) -#define COMPUTE_m1_n24 COMPUTE_L_m1(12,99599) COMPUTE_R_m1(12,99999) -#define COMPUTE_m1(ndim) COMPUTE_m1_n##ndim +#endif +#define SAVE_h_m1n4 "movq %2,%5;" unit_save_m1n4(%%xmm4) +#define SAVE_h_m1n8 SAVE_h_m1n4 unit_save_m1n4(%%xmm5) +#define SAVE_h_m1n12 SAVE_h_m1n8 unit_save_m1n4(%%xmm6) +#define SAVE_h_m1n16 SAVE_h_m1n12 unit_save_m1n4(%%xmm7) +#define SAVE_h_m1n20 SAVE_h_m1n16 unit_save_m1n4(%%xmm8) +#define SAVE_h_m1n24 SAVE_h_m1n20 unit_save_m1n4(%%xmm9) +#define SAVE_m1(ndim) SAVE_h_m1n##ndim "addq $4,%2;" +#define COMPUTE_m1(ndim) \ + INIT_m1n##ndim START_SET_PAPB(1,ndim)\ + "movq %%r13,%4;"\ + KERNEL_HEAD_C_n##ndim(1)\ + "testq %4,%4; jz "#ndim"001012f;"\ + #ndim"001011:\n\t"\ + KERNEL_k1m1n##ndim "decq %4; jnz "#ndim"001011b;"\ + #ndim"001012:\n\t"\ + KERNEL_TAIL_C_n##ndim(1)\ + SAVE_m1(ndim) END_SET_PA(1) -/* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 = "+r"(K), %5 = "+r"(ctemp) */ -/* %6 = "+r"(&alpha), %7 = "+r"(M), %8 = "+r"(next_b) */ -/* r11 = m(const), r12 = k << 4(const), r13 = k(const), r14 = b_head_pos(const), r15 = %1 + 3r12 */ +/* %7 = "m"(ALPHA), %8 = "m"(M), %9 = "m"(K), %10 = "m"(off) */ +#ifdef TRMMKERNEL + #if BACKWARDS == 1 + #define OFFSET_TO_K "movq %9,%%r13; subq %10,%%r13;" + #else + #define OFFSET_TO_K "movq %10,%%r13;" + #endif +#else + #define OFFSET_TO_K "movq %9,%%r13;" +#endif +#if defined(TRMMKERNEL) && !defined(LEFT) + #if BACKWARDS == 1 + #define START_UPDATE_OFFSET(ndim) {} + #define END_UPDATE_OFFSET(ndim) {off += (ndim);} + #else + #define START_UPDATE_OFFSET(ndim) {off += (ndim)>4 ? 4:(ndim);} + #define END_UPDATE_OFFSET(ndim) {off += (ndim)>4 ? ((ndim)-4):0;} + #endif +#else + #define START_UPDATE_OFFSET(ndim) {} + #define END_UPDATE_OFFSET(ndim) {} +#endif +#if defined(TRMMKERNEL) && defined(LEFT) + #if BACKWARDS == 1 + #define START_UPDATE_K(mdim) "" + #define END_UPDATE_K(mdim) "subq $"#mdim",%%r13;" + #else + #define START_UPDATE_K(mdim) "addq $"#mdim",%%r13;" + #define END_UPDATE_K(mdim) "" + #endif +#else + #define START_UPDATE_K(mdim) "" + #define END_UPDATE_K(mdim) "" +#endif #define COMPUTE(ndim) {\ - next_b = b_pointer + ndim * K;\ - __asm__ __volatile__(\ - "vbroadcastss (%6),%%zmm0;"\ - "movq %4,%%r13; movq %4,%%r12; salq $4,%%r12; movq %1,%%r14; movq %7,%%r11;"\ - "cmpq $16,%7;jb 33101"#ndim"f;"\ + next_b = b_pointer + ndim * K; START_UPDATE_OFFSET(ndim)\ + __asm__ __volatile__(\ + "vbroadcastss %7,%%zmm0;"\ + OFFSET_TO_K "movq %9,%%r12; salq $4,%%r12; movq %1,%%r14; movq %8,%%r11;"\ + "cmpq $16,%%r11;jb 33101"#ndim"f;"\ "33109"#ndim":\n\t"\ - COMPUTE_m16(ndim)\ - "subq $16,%7;cmpq $16,%7;jnb 33109"#ndim"b;"\ + START_UPDATE_K(16) COMPUTE_m16(ndim) END_UPDATE_K(16)\ + "subq $16,%%r11;cmpq $16,%%r11;jnb 33109"#ndim"b;"\ "33101"#ndim":\n\t"\ - "cmpq $8,%7;jb 33102"#ndim"f;"\ - COMPUTE_m8(ndim)\ - "subq $8,%7;"\ + "cmpq $8,%%r11;jb 33102"#ndim"f;"\ + START_UPDATE_K(8) COMPUTE_m8(ndim) END_UPDATE_K(8)\ + "subq $8,%%r11;"\ "33102"#ndim":\n\t"\ - "cmpq $4,%7;jb 33103"#ndim"f;"\ - COMPUTE_m4(ndim)\ - "subq $4,%7;"\ + "cmpq $4,%%r11;jb 33103"#ndim"f;"\ + START_UPDATE_K(4) COMPUTE_m4(ndim) END_UPDATE_K(4)\ + "subq $4,%%r11;"\ "33103"#ndim":\n\t"\ - "cmpq $2,%7;jb 33104"#ndim"f;"\ - COMPUTE_m2(ndim)\ - "subq $2,%7;"\ + "cmpq $2,%%r11;jb 33104"#ndim"f;"\ + START_UPDATE_K(2) COMPUTE_m2(ndim) END_UPDATE_K(2)\ + "subq $2,%%r11;"\ "33104"#ndim":\n\t"\ - "testq %7,%7;jz 33105"#ndim"f;"\ - COMPUTE_m1(ndim)\ + "testq %%r11,%%r11;jz 33105"#ndim"f;"\ + START_UPDATE_K(1) COMPUTE_m1(ndim) END_UPDATE_K(1)\ "33105"#ndim":\n\t"\ - "movq %%r13,%4; movq %%r14,%1; movq %%r11,%7;"\ - :"+r"(a_pointer),"+r"(b_pointer),"+r"(c_pointer),"+r"(ldc_in_bytes),"+r"(K),"+r"(ctemp),"+r"(alp),"+r"(M),"+r"(next_b)\ - ::"r11","r12","r13","r14","r15","zmm0","zmm1","zmm2","zmm3","zmm4","zmm5","zmm6","zmm7","zmm8","zmm9","zmm10","zmm11","zmm12","zmm13","zmm14",\ - "zmm15","zmm16","zmm17","zmm18","zmm19","zmm20","zmm21","zmm22","zmm23","zmm24","zmm25","zmm26","zmm27","zmm28","zmm29","zmm30","zmm31",\ - "cc","memory");\ - a_pointer -= M * K; b_pointer += ndim * K;c_pointer += LDC * ndim - M;\ + "movq %%r14,%1; vzeroupper;"\ + :"+r"(a_pointer),"+r"(b_pointer),"+r"(c_pointer),"+r"(ldc_in_bytes),"+r"(k_counter),"+r"(ctemp),"+r"(next_b)\ + :"m"(ALPHA),"m"(M),"m"(K),"m"(off):"r10","r11","r12","r13","r14","r15","cc","memory",\ + "zmm0","zmm1","zmm2","zmm3","zmm4","zmm5","zmm6","zmm7","zmm8","zmm9","zmm10","zmm11","zmm12","zmm13","zmm14","zmm15",\ + "zmm16","zmm17","zmm18","zmm19","zmm20","zmm21","zmm22","zmm23","zmm24","zmm25","zmm26","zmm27","zmm28","zmm29","zmm30","zmm31");\ + a_pointer -= M * K; b_pointer += ndim * K; c_pointer += LDC * ndim - M; END_UPDATE_OFFSET(ndim)\ } int __attribute__ ((noinline)) -CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, float * __restrict__ B, float * __restrict__ C, BLASLONG LDC) +CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, float * __restrict__ B, float * __restrict__ C, BLASLONG LDC +#ifdef TRMMKERNEL +,BLASLONG offset +#endif +) { - if(m==0||n==0||k==0||alpha==(float)0.0) return 0; + if(m==0||n==0) return 0; int64_t ldc_in_bytes = (int64_t)LDC * sizeof(float);float ALPHA = alpha; - int64_t M = (int64_t)m, K = (int64_t)k; + int64_t M = (int64_t)m, K = (int64_t)k, k_counter = K, off = 0; BLASLONG n_count = n; - float *a_pointer = A,*b_pointer = B,*c_pointer = C,*ctemp = C,*alp = &ALPHA,*next_b = B; + float *a_pointer = A,*b_pointer = B,*c_pointer = C,*ctemp = C,*next_b = B; +#ifdef TRMMKERNEL + #ifdef LEFT + off = offset; + #else + off = -offset; + #endif +#endif for(;n_count>23;n_count-=24) COMPUTE(24) for(;n_count>19;n_count-=20) COMPUTE(20) for(;n_count>15;n_count-=16) COMPUTE(16) @@ -410,470 +674,7 @@ CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, f if(n_count>0) COMPUTE(1) return 0; } - -#include -/* codes below are copied from the sgemm kernel written by Arjan van der Ven */ - -/* - * "Direct sgemm" code. This code operates directly on the inputs and outputs - * of the sgemm call, avoiding the copies, memory realignments and threading, - * and only supports alpha = 1 and beta = 0. - * This is a common case and provides value for relatively small matrixes. - * For larger matrixes the "regular" sgemm code is superior, there the cost of - * copying/shuffling the B matrix really pays off. - */ - - - -#define DECLARE_RESULT_512(N,M) __m512 result##N##M = _mm512_setzero_ps() -#define BROADCAST_LOAD_A_512(N,M) __m512 Aval##M = _mm512_broadcastss_ps(_mm_load_ss(&A[k + strideA * (i+M)])) -#define LOAD_B_512(N,M) __m512 Bval##N = _mm512_loadu_ps(&B[strideB * k + j + (N*16)]) -#define MATMUL_512(N,M) result##N##M = _mm512_fmadd_ps(Aval##M, Bval##N , result##N##M) -#define STORE_512(N,M) _mm512_storeu_ps(&R[(i+M) * strideR + j+(N*16)], result##N##M) - - -#define DECLARE_RESULT_256(N,M) __m256 result##N##M = _mm256_setzero_ps() -#define BROADCAST_LOAD_A_256(N,M) __m256 Aval##M = _mm256_broadcastss_ps(_mm_load_ss(&A[k + strideA * (i+M)])) -#define LOAD_B_256(N,M) __m256 Bval##N = _mm256_loadu_ps(&B[strideB * k + j + (N*8)]) -#define MATMUL_256(N,M) result##N##M = _mm256_fmadd_ps(Aval##M, Bval##N , result##N##M) -#define STORE_256(N,M) _mm256_storeu_ps(&R[(i+M) * strideR + j+(N*8)], result##N##M) - -#define DECLARE_RESULT_128(N,M) __m128 result##N##M = _mm_setzero_ps() -#define BROADCAST_LOAD_A_128(N,M) __m128 Aval##M = _mm_broadcastss_ps(_mm_load_ss(&A[k + strideA * (i+M)])) -#define LOAD_B_128(N,M) __m128 Bval##N = _mm_loadu_ps(&B[strideB * k + j + (N*4)]) -#define MATMUL_128(N,M) result##N##M = _mm_fmadd_ps(Aval##M, Bval##N , result##N##M) -#define STORE_128(N,M) _mm_storeu_ps(&R[(i+M) * strideR + j+(N*4)], result##N##M) - -#define DECLARE_RESULT_SCALAR(N,M) float result##N##M = 0; -#define BROADCAST_LOAD_A_SCALAR(N,M) float Aval##M = A[k + strideA * (i + M)]; -#define LOAD_B_SCALAR(N,M) float Bval##N = B[k * strideB + j + N]; -#define MATMUL_SCALAR(N,M) result##N##M += Aval##M * Bval##N; -#define STORE_SCALAR(N,M) R[(i+M) * strideR + j + N] = result##N##M; - -int sgemm_kernel_direct_performant(BLASLONG M, BLASLONG N, BLASLONG K) -{ - unsigned long long mnk = M * N * K; - /* large matrixes -> not performant */ - if (mnk >= 28 * 512 * 512) - return 0; - - /* - * if the B matrix is not a nice multiple if 4 we get many unaligned accesses, - * and the regular sgemm copy/realignment of data pays off much quicker - */ - if ((N & 3) != 0 && (mnk >= 8 * 512 * 512)) - return 0; - -#ifdef SMP - /* if we can run multithreaded, the threading changes the based threshold */ - if (mnk > 2 * 350 * 512 && num_cpu_avail(3)> 1) - return 0; +#ifndef TRMMKERNEL + #include + #include "sgemm_direct_skylakex.c" #endif - - return 1; -} - - - -void sgemm_kernel_direct (BLASLONG M, BLASLONG N, BLASLONG K, float * __restrict A, BLASLONG strideA, float * __restrict B, BLASLONG strideB , float * __restrict R, BLASLONG strideR) -{ - int i, j, k; - - int m4 = M & ~3; - int m2 = M & ~1; - - int n64 = N & ~63; - int n32 = N & ~31; - int n16 = N & ~15; - int n8 = N & ~7; - int n4 = N & ~3; - int n2 = N & ~1; - - i = 0; - - for (i = 0; i < m4; i+=4) { - - for (j = 0; j < n64; j+= 64) { - k = 0; - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0); - DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1); - DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2); DECLARE_RESULT_512(2, 2); DECLARE_RESULT_512(3, 2); - DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3); DECLARE_RESULT_512(2, 3); DECLARE_RESULT_512(3, 3); - - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - BROADCAST_LOAD_A_512(x, 2); - BROADCAST_LOAD_A_512(x, 3); - - LOAD_B_512(0, x); LOAD_B_512(1, x); LOAD_B_512(2, x); LOAD_B_512(3, x); - - MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0); - MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1); - MATMUL_512(0, 2); MATMUL_512(1, 2); MATMUL_512(2, 2); MATMUL_512(3, 2); - MATMUL_512(0, 3); MATMUL_512(1, 3); MATMUL_512(2, 3); MATMUL_512(3, 3); - } - STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0); - STORE_512(0, 1); STORE_512(1, 1); STORE_512(2, 1); STORE_512(3, 1); - STORE_512(0, 2); STORE_512(1, 2); STORE_512(2, 2); STORE_512(3, 2); - STORE_512(0, 3); STORE_512(1, 3); STORE_512(2, 3); STORE_512(3, 3); - } - - for (; j < n32; j+= 32) { - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); - DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); - DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2); - DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - BROADCAST_LOAD_A_512(x, 2); - BROADCAST_LOAD_A_512(x, 3); - - LOAD_B_512(0, x); LOAD_B_512(1, x); - - MATMUL_512(0, 0); MATMUL_512(1, 0); - MATMUL_512(0, 1); MATMUL_512(1, 1); - MATMUL_512(0, 2); MATMUL_512(1, 2); - MATMUL_512(0, 3); MATMUL_512(1, 3); - } - STORE_512(0, 0); STORE_512(1, 0); - STORE_512(0, 1); STORE_512(1, 1); - STORE_512(0, 2); STORE_512(1, 2); - STORE_512(0, 3); STORE_512(1, 3); - } - - for (; j < n16; j+= 16) { - DECLARE_RESULT_512(0, 0); - DECLARE_RESULT_512(0, 1); - DECLARE_RESULT_512(0, 2); - DECLARE_RESULT_512(0, 3); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - BROADCAST_LOAD_A_512(x, 2); - BROADCAST_LOAD_A_512(x, 3); - - LOAD_B_512(0, x); - - MATMUL_512(0, 0); - MATMUL_512(0, 1); - MATMUL_512(0, 2); - MATMUL_512(0, 3); - } - STORE_512(0, 0); - STORE_512(0, 1); - STORE_512(0, 2); - STORE_512(0, 3); - } - - for (; j < n8; j+= 8) { - DECLARE_RESULT_256(0, 0); - DECLARE_RESULT_256(0, 1); - DECLARE_RESULT_256(0, 2); - DECLARE_RESULT_256(0, 3); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_256(x, 0); - BROADCAST_LOAD_A_256(x, 1); - BROADCAST_LOAD_A_256(x, 2); - BROADCAST_LOAD_A_256(x, 3); - - LOAD_B_256(0, x); - - MATMUL_256(0, 0); - MATMUL_256(0, 1); - MATMUL_256(0, 2); - MATMUL_256(0, 3); - } - STORE_256(0, 0); - STORE_256(0, 1); - STORE_256(0, 2); - STORE_256(0, 3); - } - - for (; j < n4; j+= 4) { - DECLARE_RESULT_128(0, 0); - DECLARE_RESULT_128(0, 1); - DECLARE_RESULT_128(0, 2); - DECLARE_RESULT_128(0, 3); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_128(x, 0); - BROADCAST_LOAD_A_128(x, 1); - BROADCAST_LOAD_A_128(x, 2); - BROADCAST_LOAD_A_128(x, 3); - - LOAD_B_128(0, x); - - MATMUL_128(0, 0); - MATMUL_128(0, 1); - MATMUL_128(0, 2); - MATMUL_128(0, 3); - } - STORE_128(0, 0); - STORE_128(0, 1); - STORE_128(0, 2); - STORE_128(0, 3); - } - - for (; j < n2; j+= 2) { - DECLARE_RESULT_SCALAR(0, 0); DECLARE_RESULT_SCALAR(1, 0); - DECLARE_RESULT_SCALAR(0, 1); DECLARE_RESULT_SCALAR(1, 1); - DECLARE_RESULT_SCALAR(0, 2); DECLARE_RESULT_SCALAR(1, 2); - DECLARE_RESULT_SCALAR(0, 3); DECLARE_RESULT_SCALAR(1, 3); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(x, 0); - BROADCAST_LOAD_A_SCALAR(x, 1); - BROADCAST_LOAD_A_SCALAR(x, 2); - BROADCAST_LOAD_A_SCALAR(x, 3); - - LOAD_B_SCALAR(0, x); LOAD_B_SCALAR(1, x); - - MATMUL_SCALAR(0, 0); MATMUL_SCALAR(1, 0); - MATMUL_SCALAR(0, 1); MATMUL_SCALAR(1, 1); - MATMUL_SCALAR(0, 2); MATMUL_SCALAR(1, 2); - MATMUL_SCALAR(0, 3); MATMUL_SCALAR(1, 3); - } - STORE_SCALAR(0, 0); STORE_SCALAR(1, 0); - STORE_SCALAR(0, 1); STORE_SCALAR(1, 1); - STORE_SCALAR(0, 2); STORE_SCALAR(1, 2); - STORE_SCALAR(0, 3); STORE_SCALAR(1, 3); - } - - for (; j < N; j++) { - DECLARE_RESULT_SCALAR(0, 0) - DECLARE_RESULT_SCALAR(0, 1) - DECLARE_RESULT_SCALAR(0, 2) - DECLARE_RESULT_SCALAR(0, 3) - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(0, 0); - BROADCAST_LOAD_A_SCALAR(0, 1); - BROADCAST_LOAD_A_SCALAR(0, 2); - BROADCAST_LOAD_A_SCALAR(0, 3); - - LOAD_B_SCALAR(0, 0); - - MATMUL_SCALAR(0, 0); - MATMUL_SCALAR(0, 1); - MATMUL_SCALAR(0, 2); - MATMUL_SCALAR(0, 3); - } - STORE_SCALAR(0, 0); - STORE_SCALAR(0, 1); - STORE_SCALAR(0, 2); - STORE_SCALAR(0, 3); - } - } - - for (; i < m2; i+=2) { - j = 0; - - for (; j < n64; j+= 64) { - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0); - DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1); - - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - - LOAD_B_512(0, x); LOAD_B_512(1, x); LOAD_B_512(2, x); LOAD_B_512(3, x); - - MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0); - MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1); - } - STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0); - STORE_512(0, 1); STORE_512(1, 1); STORE_512(2, 1); STORE_512(3, 1); - } - - for (; j < n32; j+= 32) { - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); - DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - - LOAD_B_512(0, x); LOAD_B_512(1, x); - - MATMUL_512(0, 0); MATMUL_512(1, 0); - MATMUL_512(0, 1); MATMUL_512(1, 1); - } - STORE_512(0, 0); STORE_512(1, 0); - STORE_512(0, 1); STORE_512(1, 1); - } - - - for (; j < n16; j+= 16) { - DECLARE_RESULT_512(0, 0); - DECLARE_RESULT_512(0, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - BROADCAST_LOAD_A_512(x, 1); - - LOAD_B_512(0, x); - - MATMUL_512(0, 0); - MATMUL_512(0, 1); - } - STORE_512(0, 0); - STORE_512(0, 1); - } - - for (; j < n8; j+= 8) { - DECLARE_RESULT_256(0, 0); - DECLARE_RESULT_256(0, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_256(x, 0); - BROADCAST_LOAD_A_256(x, 1); - - LOAD_B_256(0, x); - - MATMUL_256(0, 0); - MATMUL_256(0, 1); - } - STORE_256(0, 0); - STORE_256(0, 1); - } - - for (; j < n4; j+= 4) { - DECLARE_RESULT_128(0, 0); - DECLARE_RESULT_128(0, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_128(x, 0); - BROADCAST_LOAD_A_128(x, 1); - - LOAD_B_128(0, x); - - MATMUL_128(0, 0); - MATMUL_128(0, 1); - } - STORE_128(0, 0); - STORE_128(0, 1); - } - for (; j < n2; j+= 2) { - DECLARE_RESULT_SCALAR(0, 0); DECLARE_RESULT_SCALAR(1, 0); - DECLARE_RESULT_SCALAR(0, 1); DECLARE_RESULT_SCALAR(1, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(x, 0); - BROADCAST_LOAD_A_SCALAR(x, 1); - - LOAD_B_SCALAR(0, x); LOAD_B_SCALAR(1, x); - - MATMUL_SCALAR(0, 0); MATMUL_SCALAR(1, 0); - MATMUL_SCALAR(0, 1); MATMUL_SCALAR(1, 1); - } - STORE_SCALAR(0, 0); STORE_SCALAR(1, 0); - STORE_SCALAR(0, 1); STORE_SCALAR(1, 1); - } - - for (; j < N; j++) { - DECLARE_RESULT_SCALAR(0, 0); - DECLARE_RESULT_SCALAR(0, 1); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(0, 0); - BROADCAST_LOAD_A_SCALAR(0, 1); - - LOAD_B_SCALAR(0, 0); - - MATMUL_SCALAR(0, 0); - MATMUL_SCALAR(0, 1); - } - STORE_SCALAR(0, 0); - STORE_SCALAR(0, 1); - } - } - - for (; i < M; i+=1) { - j = 0; - for (; j < n64; j+= 64) { - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - LOAD_B_512(0, x); LOAD_B_512(1, x); LOAD_B_512(2, x); LOAD_B_512(3, x); - MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0); - } - STORE_512(0, 0); STORE_512(1, 0); STORE_512(2, 0); STORE_512(3, 0); - } - for (; j < n32; j+= 32) { - DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - LOAD_B_512(0, x); LOAD_B_512(1, x); - MATMUL_512(0, 0); MATMUL_512(1, 0); - } - STORE_512(0, 0); STORE_512(1, 0); - } - - - for (; j < n16; j+= 16) { - DECLARE_RESULT_512(0, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_512(x, 0); - - LOAD_B_512(0, x); - - MATMUL_512(0, 0); - } - STORE_512(0, 0); - } - - for (; j < n8; j+= 8) { - DECLARE_RESULT_256(0, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_256(x, 0); - LOAD_B_256(0, x); - MATMUL_256(0, 0); - } - STORE_256(0, 0); - } - - for (; j < n4; j+= 4) { - DECLARE_RESULT_128(0, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_128(x, 0); - LOAD_B_128(0, x); - MATMUL_128(0, 0); - } - STORE_128(0, 0); - } - - for (; j < n2; j+= 2) { - DECLARE_RESULT_SCALAR(0, 0); DECLARE_RESULT_SCALAR(1, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(x, 0); - LOAD_B_SCALAR(0, 0); LOAD_B_SCALAR(1, 0); - MATMUL_SCALAR(0, 0); MATMUL_SCALAR(1, 0); - } - STORE_SCALAR(0, 0); STORE_SCALAR(1, 0); - } - - for (; j < N; j++) { - DECLARE_RESULT_SCALAR(0, 0); - - for (k = 0; k < K; k++) { - BROADCAST_LOAD_A_SCALAR(0, 0); - LOAD_B_SCALAR(0, 0); - MATMUL_SCALAR(0, 0); - } - STORE_SCALAR(0, 0); - } - } -} diff --git a/kernel/x86_64/sgemm_kernel_8x4_haswell.c b/kernel/x86_64/sgemm_kernel_8x4_haswell.c index 2b8aa9862..2250e4b97 100644 --- a/kernel/x86_64/sgemm_kernel_8x4_haswell.c +++ b/kernel/x86_64/sgemm_kernel_8x4_haswell.c @@ -467,7 +467,7 @@ CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, f ,BLASLONG offset #endif ){ - if(m==0||n==0||k==0||alpha==0.0) return 0; + if(m==0||n==0) return 0; int64_t ldc_in_bytes = (int64_t)LDC * sizeof(float); float constval = alpha; float *const_val=&constval; diff --git a/kernel/x86_64/sgemm_kernel_8x4_haswell_2.c b/kernel/x86_64/sgemm_kernel_8x4_haswell_2.c new file mode 100644 index 000000000..5ab3e6d1f --- /dev/null +++ b/kernel/x86_64/sgemm_kernel_8x4_haswell_2.c @@ -0,0 +1,424 @@ +/* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 for k_count, %5 for c_store, %6 = b_pref */ +/* r10 = tmp, r11 = m_counter, r12 = k << 2(const), r13 = tmp, r14 = b_head_pos(const), r15 = tmp */ + +/* m = 8 *//* ymm0 for alpha, ymm1-ymm3 for temporary use, ymm4-ymm15 for accumulators */ +#define KERNEL_k1m8n1 \ + "vmovups (%0),%%ymm1; addq $32,%0;"\ + "vbroadcastss (%1),%%ymm2; vfmadd231ps %%ymm1,%%ymm2,%%ymm4;"\ + "addq $4,%1;" +#define KERNEL_h_k1m8n2 \ + "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; addq $32,%0;"\ + "vbroadcastsd (%1),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;" +#define KERNEL_k1m8n2 KERNEL_h_k1m8n2 "addq $8,%1;" +#define KERNEL_h_k1m8n4 \ + KERNEL_h_k1m8n2 "vbroadcastsd 8(%1),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;" +#define KERNEL_k1m8n4 KERNEL_h_k1m8n4 "addq $16,%1;" +#define unit_kernel_k1m8n4(c1,c2,c3,c4,boff,...) \ + "vbroadcastsd "#boff"("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c1"; vfmadd231ps %%ymm2,%%ymm3,"#c2";"\ + "vbroadcastsd "#boff"+8("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c3"; vfmadd231ps %%ymm2,%%ymm3,"#c4";" +#define KERNEL_h_k1m8n8 KERNEL_h_k1m8n4 unit_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,0,%1,%%r12,4) +#define KERNEL_k1m8n8 KERNEL_h_k1m8n8 "addq $16,%1;" +#define KERNEL_h_k1m8n12 KERNEL_h_k1m8n8 unit_kernel_k1m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15,0,%1,%%r12,8) +#define KERNEL_k1m8n12 KERNEL_h_k1m8n12 "addq $16,%1;" +#define KERNEL_k2m8n4 \ + "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; prefetcht0 512(%0);"\ + unit_kernel_k1m8n4(%%ymm4,%%ymm5,%%ymm6,%%ymm7,0,%1)\ + "vmovsldup 32(%0),%%ymm1; vmovshdup 32(%0),%%ymm2; addq $64,%0;"\ + unit_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,16,%1)\ + "addq $32,%1;" +#define KERNEL_L_k1m8n6 \ + "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; prefetcht0 512(%0); addq $32,%0;"\ + "vbroadcastsd (%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\ + "vbroadcastsd 8(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\ + "vbroadcastsd (%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8; vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\ + "addq $16,%1;" +#define KERNEL_L_k2m8n6 \ + "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; prefetcht0 512(%0);"\ + "vbroadcastsd (%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\ + "vbroadcastsd 8(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\ + "vbroadcastsd (%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8; vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\ + "vmovsldup 32(%0),%%ymm1; vmovshdup 32(%0),%%ymm2; addq $64,%0;"\ + "vbroadcastsd 16(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\ + "vbroadcastsd 24(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\ + "vbroadcastsd 16(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\ + "addq $32,%1;" +#define KERNEL_L_k1m16n6 \ + "vmovups (%0),%%ymm1; vmovups (%0,%%r12,8),%%ymm2; prefetcht0 512(%0,%%r12,8); addq $32,%0;"\ + "vbroadcastss (%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\ + "vbroadcastss 4(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\ + "vbroadcastss 8(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8; vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\ + "vbroadcastss 12(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\ + "vbroadcastss (%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\ + "vbroadcastss 4(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\ + "addq $16,%1;" +#define KERNEL_L_k2m16n6 \ + "vmovups (%0),%%ymm1; vmovups (%0,%%r12,8),%%ymm2; prefetcht0 512(%0,%%r12,8);"\ + "vbroadcastss (%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\ + "vbroadcastss 4(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\ + "vbroadcastss 8(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8; vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\ + "vbroadcastss 12(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\ + "vbroadcastss (%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\ + "vbroadcastss 4(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\ + "vmovups 32(%0),%%ymm1; vmovups 32(%0,%%r12,8),%%ymm2; addq $64,%0;"\ + "vbroadcastss 16(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\ + "vbroadcastss 20(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\ + "vbroadcastss 24(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8; vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\ + "vbroadcastss 28(%1) ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\ + "vbroadcastss 16(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\ + "vbroadcastss 20(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\ + "addq $32,%1;" +#define KERNEL_R_k1m16n6 \ + "vmovups (%0),%%ymm1; vmovups (%0,%%r12,8),%%ymm2; prefetcht0 512(%0,%%r12,8); addq $32,%0;"\ + "vbroadcastss 8(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\ + "vbroadcastss 12(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\ + "vbroadcastss (%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8; vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\ + "vbroadcastss 4(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\ + "vbroadcastss 8(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\ + "vbroadcastss 12(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\ + "addq $16,%1;" +#define KERNEL_R_k2m16n6 \ + "vmovups (%0),%%ymm1; vmovups (%0,%%r12,8),%%ymm2; prefetcht0 512(%0,%%r12,8);"\ + "vbroadcastss 8(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\ + "vbroadcastss 12(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\ + "vbroadcastss (%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8; vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\ + "vbroadcastss 4(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\ + "vbroadcastss 8(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\ + "vbroadcastss 12(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\ + "vmovups 32(%0),%%ymm1; vmovups 32(%0,%%r12,8),%%ymm2; addq $64,%0;"\ + "vbroadcastss 24(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\ + "vbroadcastss 28(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\ + "vbroadcastss 16(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8; vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\ + "vbroadcastss 20(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\ + "vbroadcastss 24(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\ + "vbroadcastss 28(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\ + "addq $32,%1;" +#define KERNEL_R_k1m8n6 \ + "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; prefetcht0 512(%0); addq $32,%0;"\ + "vbroadcastsd 8(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\ + "vbroadcastsd (%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\ + "vbroadcastsd 8(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8; vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\ + "addq $16,%1;" +#define KERNEL_R_k2m8n6 \ + "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; prefetcht0 512(%0);"\ + "vbroadcastsd 8(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\ + "vbroadcastsd (%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\ + "vbroadcastsd 8(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8; vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\ + "vmovsldup 32(%0),%%ymm1; vmovshdup 32(%0),%%ymm2; addq $64,%0;"\ + "vbroadcastsd 24(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\ + "vbroadcastsd 16(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\ + "vbroadcastsd 24(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\ + "addq $32,%1;" +#define INIT_m8n1 "vpxor %%ymm4,%%ymm4,%%ymm4;" +#define INIT_m8n2 INIT_m8n1 "vpxor %%ymm5,%%ymm5,%%ymm5;" +#define unit_init_m8n4(c1,c2,c3,c4) \ + "vpxor "#c1","#c1","#c1";vpxor "#c2","#c2","#c2";vpxor "#c3","#c3","#c3";vpxor "#c4","#c4","#c4";" +#define INIT_m8n8 unit_init_m8n4(%%ymm4,%%ymm5,%%ymm6,%%ymm7) unit_init_m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11) +#define INIT_m8n4 INIT_m8n8 +#define INIT_m8n12 INIT_m8n8 unit_init_m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15) +#define INIT_m8n6 INIT_m8n12 +#define INIT_m16n6 INIT_m8n12 +#define SAVE_m8n1 "vfmadd213ps (%2),%%ymm0,%%ymm4; vmovups %%ymm4,(%2);" +#define unit_save_m8n2(c1,c2) \ + "vunpcklps "#c2","#c1",%%ymm2; vunpckhps "#c2","#c1",%%ymm3; vunpcklpd %%ymm3,%%ymm2,"#c1"; vunpckhpd %%ymm3,%%ymm2,"#c2";"\ + "vfmadd213ps (%5),%%ymm0,"#c1"; vfmadd213ps (%5,%3,1),%%ymm0,"#c2"; vmovups "#c1",(%5); vmovups "#c2",(%5,%3,1); leaq (%5,%3,2),%5;" +#define SAVE_m8n2 "movq %2,%5;" unit_save_m8n2(%%ymm4,%%ymm5) +#define SAVE_m8n4 "movq %2,%5;"\ + "vaddps %%ymm4,%%ymm8,%%ymm4; vaddps %%ymm5,%%ymm9,%%ymm5; vaddps %%ymm6,%%ymm10,%%ymm6; vaddps %%ymm7,%%ymm11,%%ymm7;"\ + unit_save_m8n2(%%ymm4,%%ymm5) unit_save_m8n2(%%ymm6,%%ymm7) +#define SAVE_m8n8 "movq %2,%5;"\ + unit_save_m8n2(%%ymm4,%%ymm5) unit_save_m8n2(%%ymm6,%%ymm7) unit_save_m8n2(%%ymm8,%%ymm9) unit_save_m8n2(%%ymm10,%%ymm11) +#define SAVE_m8n12 SAVE_m8n8 unit_save_m8n2(%%ymm12,%%ymm13) unit_save_m8n2(%%ymm14,%%ymm15) +#define unit_save_m16n2(c1,c2,c3,c4) \ + "vfmadd213ps (%5),%%ymm0,"#c1"; vfmadd213ps 32(%5),%%ymm0,"#c2"; vmovups "#c1",(%5); vmovups "#c2",32(%5);"\ + "vfmadd213ps (%5,%3,1),%%ymm0,"#c3"; vfmadd213ps 32(%5,%3,1),%%ymm0,"#c4"; vmovups "#c3",(%5,%3,1); vmovups "#c4",32(%5,%3,1); leaq (%5,%3,2),%5;" +#define SAVE_L_m16n6 "movq %2,%5;"\ + unit_save_m16n2(%%ymm4,%%ymm5,%%ymm6,%%ymm7) unit_save_m16n2(%%ymm8,%%ymm9,%%ymm10,%%ymm11) unit_save_m16n2(%%ymm12,%%ymm13,%%ymm14,%%ymm15) +#define SAVE_R_m16n6 "leaq (%2,%3,4),%5; leaq (%5,%3,2),%5;"\ + unit_save_m16n2(%%ymm4,%%ymm5,%%ymm6,%%ymm7) unit_save_m16n2(%%ymm8,%%ymm9,%%ymm10,%%ymm11) unit_save_m16n2(%%ymm12,%%ymm13,%%ymm14,%%ymm15) +#define SAVE_L_m8n6 "movq %2,%5;"\ + "vaddps %%ymm4,%%ymm10,%%ymm4; vaddps %%ymm5,%%ymm11,%%ymm5; vaddps %%ymm6,%%ymm12,%%ymm6;"\ + "vaddps %%ymm7,%%ymm13,%%ymm7; vaddps %%ymm8,%%ymm14,%%ymm8; vaddps %%ymm9,%%ymm15,%%ymm9;"\ + unit_save_m8n2(%%ymm4,%%ymm5) unit_save_m8n2(%%ymm6,%%ymm7) unit_save_m8n2(%%ymm8,%%ymm9) +#define SAVE_R_m8n6 "leaq (%2,%3,4),%5; leaq (%5,%3,2),%5;"\ + "vaddps %%ymm4,%%ymm10,%%ymm4; vaddps %%ymm5,%%ymm11,%%ymm5; vaddps %%ymm6,%%ymm12,%%ymm6;"\ + "vaddps %%ymm7,%%ymm13,%%ymm7; vaddps %%ymm8,%%ymm14,%%ymm8; vaddps %%ymm9,%%ymm15,%%ymm9;"\ + unit_save_m8n2(%%ymm4,%%ymm5) unit_save_m8n2(%%ymm6,%%ymm7) unit_save_m8n2(%%ymm8,%%ymm9) + +/* m = 4 *//* xmm0 for alpha, xmm1-xmm3 for temporary use, xmm4-xmm15 for accumulators */ +#define KERNEL_k1m4n1 \ + "vmovups (%0),%%xmm1; addq $16,%0;"\ + "vbroadcastss (%1),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\ + "addq $4,%1;" +#define KERNEL_h_k1m4n2 \ + "vmovsldup (%0),%%xmm1; vmovshdup (%0),%%xmm2; addq $16,%0;"\ + "vmovddup (%1),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm4; vfmadd231ps %%xmm2,%%xmm3,%%xmm5;" +#define KERNEL_k1m4n2 KERNEL_h_k1m4n2 "addq $8,%1;" +#define KERNEL_h_k1m4n4 \ + KERNEL_h_k1m4n2 "vmovddup 8(%1),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm6; vfmadd231ps %%xmm2,%%xmm3,%%xmm7;" +#define KERNEL_k1m4n4 KERNEL_h_k1m4n4 "addq $16,%1;" +#define unit_kernel_k1m4n4(c1,c2,c3,c4,...) \ + "vmovddup ("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c1"; vfmadd231ps %%xmm2,%%xmm3,"#c2";"\ + "vmovddup 8("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c3"; vfmadd231ps %%xmm2,%%xmm3,"#c4";" +#define KERNEL_h_k1m4n8 KERNEL_h_k1m4n4 unit_kernel_k1m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11,%1,%%r12,4) +#define KERNEL_k1m4n8 KERNEL_h_k1m4n8 "addq $16,%1;" +#define KERNEL_h_k1m4n12 KERNEL_h_k1m4n8 unit_kernel_k1m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15,%1,%%r12,8) +#define KERNEL_k1m4n12 KERNEL_h_k1m4n12 "addq $16,%1;" +#define INIT_m4n1 "vpxor %%xmm4,%%xmm4,%%xmm4;" +#define INIT_m4n2 INIT_m4n1 "vpxor %%xmm5,%%xmm5,%%xmm5;" +#define INIT_m4n4 INIT_m4n2 "vpxor %%xmm6,%%xmm6,%%xmm6;vpxor %%xmm7,%%xmm7,%%xmm7;" +#define unit_init_m4n4(c1,c2,c3,c4) \ + "vpxor "#c1","#c1","#c1";vpxor "#c2","#c2","#c2";vpxor "#c3","#c3","#c3";vpxor "#c4","#c4","#c4";" +#define INIT_m4n8 INIT_m4n4 unit_init_m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11) +#define INIT_m4n12 INIT_m4n8 unit_init_m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15) +#define SAVE_m4n1 "vfmadd213ps (%2),%%xmm0,%%xmm4; vmovups %%xmm4,(%2);" +#define unit_save_m4n2(c1,c2) \ + "vunpcklps "#c2","#c1",%%xmm2; vunpckhps "#c2","#c1",%%xmm3; vunpcklpd %%xmm3,%%xmm2,"#c1"; vunpckhpd %%xmm3,%%xmm2,"#c2";"\ + "vfmadd213ps (%5),%%xmm0,"#c1"; vmovups "#c1",(%5);"\ + "vfmadd213ps (%5,%3,1),%%xmm0,"#c2"; vmovups "#c2",(%5,%3,1);"\ + "leaq (%5,%3,2),%5;" +#define SAVE_m4n2 "movq %2,%5;" unit_save_m4n2(%%xmm4,%%xmm5) +#define SAVE_m4n4 SAVE_m4n2 unit_save_m4n2(%%xmm6,%%xmm7) +#define SAVE_m4n8 SAVE_m4n4 unit_save_m4n2(%%xmm8,%%xmm9) unit_save_m4n2(%%xmm10,%%xmm11) +#define SAVE_m4n12 SAVE_m4n8 unit_save_m4n2(%%xmm12,%%xmm13) unit_save_m4n2(%%xmm14,%%xmm15) + +/* m = 2 *//* xmm0 for alpha, xmm1-xmm3 and xmm10 for temporary use, xmm4-xmm9 for accumulators */ +#define INIT_m2n1 "vpxor %%xmm4,%%xmm4,%%xmm4;" +#define KERNEL_k1m2n1 \ + "vmovsd (%0),%%xmm1; addq $8,%0;"\ + "vbroadcastss (%1),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\ + "addq $4,%1;" +#define SAVE_m2n1 "vmovsd (%2),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm4; vmovsd %%xmm4,(%2);" +#define INIT_m2n2 INIT_m2n1 "vpxor %%xmm5,%%xmm5,%%xmm5;" +#define KERNEL_k1m2n2 \ + "vmovsd (%0),%%xmm1; addq $8,%0;"\ + "vbroadcastss (%1),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\ + "vbroadcastss 4(%1),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm5;"\ + "addq $8,%1;" +#define SAVE_m2n2 SAVE_m2n1 "vmovsd (%2,%3,1),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm5; vmovsd %%xmm5,(%2,%3,1);" +#define INIT_m2n4 INIT_m2n2 +#define INIT_m2n8 INIT_m2n4 "vpxor %%xmm6,%%xmm6,%%xmm6; vpxor %%xmm7,%%xmm7,%%xmm7;" +#define INIT_m2n12 INIT_m2n8 "vpxor %%xmm8,%%xmm8,%%xmm8; vpxor %%xmm9,%%xmm9,%%xmm9;" +#define KERNEL_k1m2n4 \ + "vmovups (%1),%%xmm3; addq $16,%1;"\ + "vbroadcastss (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4;"\ + "vbroadcastss 4(%0),%%xmm2; vfmadd231ps %%xmm3,%%xmm2,%%xmm5;"\ + "addq $8,%0;" +#define KERNEL_k1m2n8 \ + "vmovups (%1),%%xmm3; vmovups (%1,%%r12,4),%%xmm2; addq $16,%1;"\ + "vbroadcastss (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4; vfmadd231ps %%xmm2,%%xmm1,%%xmm6;"\ + "vbroadcastss 4(%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm5; vfmadd231ps %%xmm2,%%xmm1,%%xmm7;"\ + "addq $8,%0;" +#define KERNEL_k1m2n12 \ + "vmovups (%1),%%xmm3; vmovups (%1,%%r12,4),%%xmm2; vmovups (%1,%%r12,8),%%xmm1; addq $16,%1;"\ + "vbroadcastss (%0),%%xmm10; vfmadd231ps %%xmm3,%%xmm10,%%xmm4; vfmadd231ps %%xmm2,%%xmm10,%%xmm6; vfmadd231ps %%xmm1,%%xmm10,%%xmm8;"\ + "vbroadcastss 4(%0),%%xmm10; vfmadd231ps %%xmm3,%%xmm10,%%xmm5; vfmadd231ps %%xmm2,%%xmm10,%%xmm7; vfmadd231ps %%xmm1,%%xmm10,%%xmm9;"\ + "addq $8,%0;" +#define unit_save_m2n4(c1,c2) \ + "vunpcklps "#c2","#c1",%%xmm1; vunpckhps "#c2","#c1",%%xmm2;"\ + "vmovsd (%5),%%xmm3; vmovhpd (%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm1;"\ + "vmovsd %%xmm1,(%5); vmovhpd %%xmm1,(%5,%3,1); leaq (%5,%3,2),%5;"\ + "vmovsd (%5),%%xmm3; vmovhpd (%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm2;"\ + "vmovsd %%xmm2,(%5); vmovhpd %%xmm2,(%5,%3,1); leaq (%5,%3,2),%5;" +#define SAVE_m2n4 "movq %2,%5;" unit_save_m2n4(%%xmm4,%%xmm5) +#define SAVE_m2n8 SAVE_m2n4 unit_save_m2n4(%%xmm6,%%xmm7) +#define SAVE_m2n12 SAVE_m2n8 unit_save_m2n4(%%xmm8,%%xmm9) + +/* m = 1 *//* xmm0 for alpha, xmm1-xmm3 and xmm10 for temporary use, xmm4-xmm6 for accumulators */ +#define INIT_m1n1 "vpxor %%xmm4,%%xmm4,%%xmm4;" +#define KERNEL_k1m1n1 \ + "vmovss (%1),%%xmm3; addq $4,%1;"\ + "vmovss (%0),%%xmm1; vfmadd231ss %%xmm3,%%xmm1,%%xmm4;"\ + "addq $4,%0;" +#define SAVE_m1n1 "vfmadd213ss (%2),%%xmm0,%%xmm4; vmovss %%xmm4,(%2);" +#define INIT_m1n2 INIT_m1n1 +#define KERNEL_k1m1n2 \ + "vmovsd (%1),%%xmm3; addq $8,%1;"\ + "vbroadcastss (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4;"\ + "addq $4,%0;" +#define SAVE_m1n2 \ + "vmovss (%2),%%xmm3; vinsertps $16,(%2,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm4;"\ + "vmovss %%xmm4,(%2); vextractps $1,%%xmm4,(%2,%3,1);" +#define INIT_m1n4 INIT_m1n2 +#define INIT_m1n8 INIT_m1n4 "vpxor %%xmm5,%%xmm5,%%xmm5;" +#define INIT_m1n12 INIT_m1n8 "vpxor %%xmm6,%%xmm6,%%xmm6;" +#define KERNEL_k1m1n4 \ + "vmovups (%1),%%xmm3; addq $16,%1;"\ + "vbroadcastss (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4;"\ + "addq $4,%0;" +#define KERNEL_k1m1n8 \ + "vmovups (%1),%%xmm3; vmovups (%1,%%r12,4),%%xmm2; addq $16,%1;"\ + "vbroadcastss (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4; vfmadd231ps %%xmm2,%%xmm1,%%xmm5;"\ + "addq $4,%0;" +#define KERNEL_k1m1n12 \ + "vmovups (%1),%%xmm3; vmovups (%1,%%r12,4),%%xmm2; vmovups (%1,%%r12,8),%%xmm1; addq $16,%1;"\ + "vbroadcastss (%0),%%xmm10; vfmadd231ps %%xmm3,%%xmm10,%%xmm4; vfmadd231ps %%xmm2,%%xmm10,%%xmm5; vfmadd231ps %%xmm1,%%xmm10,%%xmm6;"\ + "addq $4,%0;" +#define unit_save_m1n4(c1) \ + "vpxor %%xmm10,%%xmm10,%%xmm10; vmovsd "#c1",%%xmm10,%%xmm2; vmovhlps "#c1",%%xmm10,%%xmm1;"\ + "vmovss (%5),%%xmm3; vinsertps $16,(%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm2;"\ + "vmovss %%xmm2,(%5); vextractps $1,%%xmm2,(%5,%3,1); leaq (%5,%3,2),%5;"\ + "vmovss (%5),%%xmm3; vinsertps $16,(%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm1;"\ + "vmovss %%xmm1,(%5); vextractps $1,%%xmm1,(%5,%3,1); leaq (%5,%3,2),%5;" +#define SAVE_m1n4 "movq %2,%5;" unit_save_m1n4(%%xmm4) +#define SAVE_m1n8 SAVE_m1n4 unit_save_m1n4(%%xmm5) +#define SAVE_m1n12 SAVE_m1n8 unit_save_m1n4(%%xmm6) + +/* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 for k_count, %5 for c_store, %6 = b_pref */ +/* r10 = tmp, r11 = m_counter, r12 = k << 2(const), r13 = tmp, r14 = b_head_pos(const), r15 = tmp */ + +#define COMPUTE_SIMPLE(mdim,ndim) \ + "movq %%r12,%4; sarq $2,%4; movq %%r14,%1;" INIT_m##mdim##n##ndim\ + "testq %4,%4; jz 7"#mdim"7"#ndim"2f;"\ + "7"#mdim"7"#ndim"1:\n\t"\ + KERNEL_k1m##mdim##n##ndim "decq %4; jnz 7"#mdim"7"#ndim"1b;"\ + "7"#mdim"7"#ndim"2:\n\t"\ + SAVE_m##mdim##n##ndim "addq $"#mdim"*4,%2;" +#define COMPUTE_m8n1 COMPUTE_SIMPLE(8,1) +#define COMPUTE_m8n2 COMPUTE_SIMPLE(8,2) +#define COMPUTE_m8n8 COMPUTE_SIMPLE(8,8) +#define COMPUTE_m8n12 COMPUTE_SIMPLE(8,12) +#define COMPUTE_m8n4 \ + "movq %%r12,%4; sarq $2,%4; movq %%r14,%1;" INIT_m8n4\ + "cmpq $8,%4; jb 78740f;"\ + "78749:\n\t"\ + KERNEL_k2m8n4 KERNEL_k2m8n4 KERNEL_k2m8n4 KERNEL_k2m8n4\ + "subq $8,%4; cmpq $8,%4; jnb 78749b;"\ + "78740:\n\t"\ + "testq %4,%4; jz 78742f;"\ + "78741:\n\t"\ + KERNEL_k1m8n4 "decq %4; jnz 78741b;"\ + "78742:\n\t"\ + SAVE_m8n4 "addq $32,%2;" +#define COMPUTE_L_m16n6 \ + "movq %%r12,%%r13; sarq $2,%%r13; movq %%r14,%1;" INIT_m16n6\ + "movq %%r13,%4; movq %2,%5; cmpq $16,%%r13; jb 7116762f; movq $14,%4;"\ + "7116761:\n\t"\ + KERNEL_L_k2m16n6 "prefetcht0 128(%1); testq $24,%4; movq $84,%%r15; cmovz %3,%%r15;"\ + KERNEL_L_k2m16n6 "prefetcht1 (%5); subq $63,%5; addq %%r15,%5;"\ + KERNEL_L_k2m16n6 "prefetcht0 128(%1); prefetcht1 (%6); cmpq $198,%4; cmoveq %2,%5;"\ + KERNEL_L_k2m16n6 "addq $16,%6; addq $8,%4; cmpq %4,%%r13; jnb 7116761b;"\ + "movq %2,%5; negq %4; leaq 14(%%r13,%4,1),%4;"\ + "7116762:\n\t"\ + "xorq %%r15,%%r15; testq %4,%4; jz 7116764f;"\ + "7116763:\n\t"\ + "prefetcht0 (%5); prefetcht0 63(%5); addq %3,%5; incq %%r15;"\ + KERNEL_L_k1m16n6 "cmpq $6,%%r15; cmoveq %2,%5; decq %4; jnz 7116763b;"\ + "7116764:\n\t"\ + SAVE_L_m16n6 "addq $32,%2;" +#define COMPUTE_R_m16n6 \ + "movq %%r12,%%r13; sarq $2,%%r13; movq %%r14,%1;" INIT_m16n6\ + "movq %%r13,%4; leaq (%2,%3,4),%5; leaq (%5,%3,2),%5; movq %5,%%r10; cmpq $16,%%r13; jb 7216762f; movq $14,%4;"\ + "7216761:\n\t"\ + KERNEL_R_k2m16n6 "prefetcht0 128(%1,%%r12,8); testq $24,%4; movq $84,%%r15; cmovz %3,%%r15;"\ + KERNEL_R_k2m16n6 "prefetcht1 (%5); subq $63,%5; addq %%r15,%5;"\ + KERNEL_R_k2m16n6 "prefetcht0 128(%1,%%r12,8); prefetcht1 (%6); cmpq $198,%4; cmoveq %%r10,%5;"\ + KERNEL_R_k2m16n6 "addq $16,%6; addq $8,%4; cmpq %4,%%r13; jnb 7216761b;"\ + "movq %%r10,%5; negq %4; leaq 14(%%r13,%4,1),%4;"\ + "7216762:\n\t"\ + "xorq %%r15,%%r15; testq %4,%4; jz 7216764f;"\ + "7216763:\n\t"\ + "prefetcht0 (%5); prefetcht0 63(%5); addq %3,%5; incq %%r15;"\ + KERNEL_R_k1m16n6 "cmpq $6,%%r15; cmoveq %%r10,%5; decq %4; jnz 7216763b;"\ + "7216764:\n\t"\ + "prefetcht0 (%%r14); prefetcht0 64(%%r14);" SAVE_R_m16n6 "addq $32,%2;" +#define COMPUTE_H_m8n6 \ + "movq %%r12,%4; sarq $2,%4; movq %%r14,%1;" INIT_m8n6\ + "cmpq $8,%4; jb 718760f; movq %2,%5; xorq %%r15,%%r15;"\ + "718769:\n\t"\ + KERNEL_L_k2m8n6 KERNEL_L_k2m8n6 "cmpq $62,%%r15; movq $62,%%r15; cmoveq %3,%%r15;"\ + KERNEL_L_k2m8n6 KERNEL_L_k2m8n6 "prefetcht2 (%5); leaq -31(%5,%%r15,1),%5;"\ + "subq $8,%4; cmpq $8,%4; jnb 718769b;"\ + "718760:\n\t"\ + "testq %4,%4; jz 718762f;"\ + "718761:\n\t"\ + KERNEL_L_k1m8n6 "decq %4; jnz 718761b;"\ + "718762:\n\t"\ + SAVE_L_m8n6 "negq %%r12; leaq (%0,%%r12,8),%0; negq %%r12;" +#define COMPUTE_T_m8n6(side,sim) \ + "movq %%r12,%4; sarq $2,%4; movq %%r14,%1;" INIT_m8n6\ + "cmpq $8,%4; jb 72"#sim"8760f;"\ + "72"#sim"8769:\n\t"\ + KERNEL_##side##_k2m8n6 KERNEL_##side##_k2m8n6 KERNEL_##side##_k2m8n6 KERNEL_##side##_k2m8n6\ + "subq $8,%4; cmpq $8,%4; jnb 72"#sim"8769b;"\ + "72"#sim"8760:\n\t"\ + "testq %4,%4; jz 72"#sim"8762f;"\ + "72"#sim"8761:\n\t"\ + KERNEL_##side##_k1m8n6 "decq %4; jnz 72"#sim"8761b;"\ + "72"#sim"8762:\n\t"\ + SAVE_##side##_m8n6 "addq $32,%2;" +#define COMPUTE_NORMAL(ndim) {\ + next_b = b_pointer + ndim * K;\ + __asm__ __volatile__(\ + "vbroadcastss %9,%%ymm0;"\ + "movq %8,%%r12; salq $2,%%r12; movq %1,%%r14; movq %7,%%r11;"\ + "cmpq $8,%%r11;jb 33101"#ndim"f;"\ + "33109"#ndim":\n\t"\ + COMPUTE_m8n##ndim\ + "subq $8,%%r11;cmpq $8,%%r11;jnb 33109"#ndim"b;"\ + "33101"#ndim":\n\t"\ + "cmpq $4,%%r11;jb 33103"#ndim"f;"\ + COMPUTE_SIMPLE(4,ndim) "subq $4,%%r11;"\ + "33103"#ndim":\n\t"\ + "cmpq $2,%%r11;jb 33104"#ndim"f;"\ + COMPUTE_SIMPLE(2,ndim) "subq $2,%%r11;"\ + "33104"#ndim":\n\t"\ + "testq %%r11,%%r11;jz 33105"#ndim"f;"\ + COMPUTE_SIMPLE(1,ndim)\ + "33105"#ndim":\n\t"\ + "movq %%r14,%1; vzeroupper;"\ + :"+r"(a_pointer),"+r"(b_pointer),"+r"(c_pointer),"+r"(ldc_in_bytes),"+r"(k_count),"+r"(ctemp),"+r"(next_b)\ + :"m"(M),"m"(K),"m"(ALPHA):"r10","r11","r12","r13","r14","r15",\ + "xmm0","xmm1","xmm2","xmm3","xmm4","xmm5","xmm6","xmm7","xmm8","xmm9","xmm10","xmm11","xmm12","xmm13","xmm14","xmm15","cc","memory");\ + a_pointer -= M * K; b_pointer += ndim * K; c_pointer += (LDC * ndim - M);\ +} +#define COMPUTE_n12 {\ + next_b = b_pointer + 12 * K;\ + __asm__ __volatile__(\ + "vbroadcastss %9,%%ymm0;"\ + "movq %8,%%r12; salq $2,%%r12; movq %1,%%r14; movq %7,%%r11;"\ + "cmpq $16,%%r11;jb 3310112f;"\ + COMPUTE_H_m8n6\ + "3310612:\n\t"\ + COMPUTE_R_m16n6 "subq $8,%%r11; cmpq $16,%%r11;jb 3310712f;"\ + COMPUTE_L_m16n6 "subq $8,%%r11; cmpq $16,%%r11;jnb 3310612b;"\ + COMPUTE_T_m8n6(R,5) "subq $8,%%r11; jmp 3310212f;"\ + "3310712:\n\t"\ + COMPUTE_T_m8n6(L,7) "subq $8,%%r11; jmp 3310212f;"\ + "3310112:\n\t"\ + "cmpq $8,%%r11;jb 3310212f;"\ + COMPUTE_SIMPLE(8,12) "subq $8,%%r11;"\ + "3310212:\n\t"\ + "cmpq $4,%%r11;jb 3310312f;"\ + COMPUTE_SIMPLE(4,12) "subq $4,%%r11;"\ + "3310312:\n\t"\ + "cmpq $2,%%r11;jb 3310412f;"\ + COMPUTE_SIMPLE(2,12) "subq $2,%%r11;"\ + "3310412:\n\t"\ + "testq %%r11,%%r11;jz 3310512f;"\ + COMPUTE_SIMPLE(1,12)\ + "3310512:\n\t"\ + "movq %%r14,%1; vzeroupper;"\ + :"+r"(a_pointer),"+r"(b_pointer),"+r"(c_pointer),"+r"(ldc_in_bytes),"+r"(k_count),"+r"(ctemp),"+r"(next_b)\ + :"m"(M),"m"(K),"m"(ALPHA):"r10","r11","r12","r13","r14","r15",\ + "xmm0","xmm1","xmm2","xmm3","xmm4","xmm5","xmm6","xmm7","xmm8","xmm9","xmm10","xmm11","xmm12","xmm13","xmm14","xmm15","cc","memory");\ + a_pointer -= M * K; b_pointer += 12 * K; c_pointer += (LDC * 12 - M);\ +} + +#include "common.h" +#include +int __attribute__ ((noinline)) +CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, float * __restrict__ B, float * __restrict__ C, BLASLONG LDC){ + if(m==0||n==0||k==0||alpha==(float)0.0) return 0; + int64_t ldc_in_bytes = (int64_t)LDC * sizeof(float); + float ALPHA = alpha; + int64_t M = (int64_t)m, K = (int64_t)k, k_count = 0; + BLASLONG n_count = n; + float *a_pointer = A,*b_pointer = B,*c_pointer = C,*ctemp = C,*next_b = B; + for(;n_count>11;n_count-=12) COMPUTE_n12 + for(;n_count>7;n_count-=8) COMPUTE_NORMAL(8) + for(;n_count>3;n_count-=4) COMPUTE_NORMAL(4) + for(;n_count>1;n_count-=2) COMPUTE_NORMAL(2) + if(n_count>0) COMPUTE_NORMAL(1) + return 0; +} + diff --git a/kernel/x86_64/zgemm_kernel_4x2_haswell.c b/kernel/x86_64/zgemm_kernel_4x2_haswell.c index 3279b8b8c..917a3fd48 100644 --- a/kernel/x86_64/zgemm_kernel_4x2_haswell.c +++ b/kernel/x86_64/zgemm_kernel_4x2_haswell.c @@ -50,7 +50,7 @@ "vmovupd (%0),%%ymm0; vmovupd 32(%0),%%ymm1; prefetcht0 512(%0); addq $64,%0;"\ acc_m4n2_con(0,1,4,5,6,7,0,16,%1) acc_m4n2_con(0,1,8,9,10,11,0,16,%1,%%r12,1) #define KERNEL_2_k1m4n4 \ - "vpermilpd $5,%%ymm0,%%ymm0; vpermilpd $5,%%ymm1,%%ymm1;"\ + "vpermilpd $5,-64(%0),%%ymm0; vpermilpd $5,-32(%0),%%ymm1;"\ acc_m4n2_con(0,1,4,5,6,7,8,24,%1) acc_m4n2_con(0,1,8,9,10,11,8,24,%1,%%r12,1) #define KERNEL_1_k1m4n6 KERNEL_1_k1m4n4 acc_m4n2_con(0,1,12,13,14,15,0,16,%1,%%r12,2) #define KERNEL_2_k1m4n6 KERNEL_2_k1m4n4 acc_m4n2_con(0,1,12,13,14,15,8,24,%1,%%r12,2) @@ -93,9 +93,9 @@ "movq $10,%5; movq $84,%%r15;"\ #ndim"4441:\n\t"\ "prefetcht1 (%3); subq $63,%3; addq %%r15,%3;"\ - "prefetcht0 96(%1); prefetcht0 96(%1,%%r12,1); prefetcht0 96(%1,%%r12,2);" KERNEL_k1m4n##ndim KERNEL_k1m4n##ndim\ + KERNEL_k1m4n##ndim KERNEL_k1m4n##ndim\ "testq $12,%5; movq $84,%%r15; cmovz %4,%%r15; prefetcht1 (%8); addq $16,%8;"\ - "prefetcht0 96(%1); prefetcht0 96(%1,%%r12,1); prefetcht0 96(%1,%%r12,2);" KERNEL_k1m4n##ndim KERNEL_k1m4n##ndim\ + KERNEL_k1m4n##ndim KERNEL_k1m4n##ndim\ "addq $4,%5; cmpq %5,%%r13; jnb "#ndim"4441b;"\ "movq %2,%3; negq %5; leaq 10(%%r13,%5,1),%5; prefetcht0 (%6); prefetcht0 15(%6);"\ #ndim"4442:\n\t"\ diff --git a/lapack-netlib/SRC/dcombssq.f b/lapack-netlib/SRC/dcombssq.f index 79f6d95c9..7a1ddd1af 100644 --- a/lapack-netlib/SRC/dcombssq.f +++ b/lapack-netlib/SRC/dcombssq.f @@ -80,6 +80,8 @@ IF( V1( 1 ).GE.V2( 1 ) ) THEN IF( V1( 1 ).NE.ZERO ) THEN V1( 2 ) = V1( 2 ) + ( V2( 1 ) / V1( 1 ) )**2 * V2( 2 ) + ELSE + V1( 2 ) = V1( 2 ) + V2( 2 ) END IF ELSE V1( 2 ) = V2( 2 ) + ( V1( 1 ) / V2( 1 ) )**2 * V1( 2 ) diff --git a/lapack-netlib/SRC/scombssq.f b/lapack-netlib/SRC/scombssq.f index 76bc0e320..cc51a324b 100644 --- a/lapack-netlib/SRC/scombssq.f +++ b/lapack-netlib/SRC/scombssq.f @@ -80,6 +80,8 @@ IF( V1( 1 ).GE.V2( 1 ) ) THEN IF( V1( 1 ).NE.ZERO ) THEN V1( 2 ) = V1( 2 ) + ( V2( 1 ) / V1( 1 ) )**2 * V2( 2 ) + ELSE + V1( 2 ) = V1( 2 ) + V2( 2 ) END IF ELSE V1( 2 ) = V2( 2 ) + ( V1( 1 ) / V2( 1 ) )**2 * V1( 2 ) diff --git a/lapack/CMakeLists.txt b/lapack/CMakeLists.txt index d48a270ab..e21a9aabb 100644 --- a/lapack/CMakeLists.txt +++ b/lapack/CMakeLists.txt @@ -4,7 +4,6 @@ include_directories(${PROJECT_BINARY_DIR}) set(LAPACK_SOURCES - getrf/getrf_single.c potrf/potrf_U_single.c potrf/potrf_L_single.c lauum/lauum_U_single.c @@ -45,6 +44,10 @@ GenerateNamedObjects("${LAPACK_MANGLED_SOURCES}" "" "" false "" "" false 3) GenerateNamedObjects("laswp/generic/laswp_k_4.c" "" "laswp_plus" false "" "" false 3) GenerateNamedObjects("laswp/generic/laswp_k_4.c" "MINUS" "laswp_minus" false "" "" false 3) +foreach (float_type ${FLOAT_TYPES}) +GenerateNamedObjects("getrf/getrf_single.c" "UNIT" "getrf_single" false "" "" false ${float_type}) +endforeach () + # dynamic_arch laswp needs arch specific code ? #foreach(TARGET_CORE ${DYNAMIC_CORE}) # set(TSUFFIX "_${TARGET_CORE}") @@ -81,7 +84,7 @@ if (USE_THREAD) ) foreach (float_type ${FLOAT_TYPES}) - GenerateNamedObjects("${GETRF_SRC}" "" "getrf_parallel" false "" "" false ${float_type}) + GenerateNamedObjects("${GETRF_SRC}" "UNIT" "getrf_parallel" false "" "" false ${float_type}) endforeach() GenerateNamedObjects("${PARALLEL_SOURCES}") diff --git a/lapack/getrf/getrf_parallel.c b/lapack/getrf/getrf_parallel.c index c82defcab..c602822a8 100644 --- a/lapack/getrf/getrf_parallel.c +++ b/lapack/getrf/getrf_parallel.c @@ -68,23 +68,14 @@ double sqrt(double); #define GETRF_FACTOR 1.00 -#if defined(USE_PTHREAD_LOCK) -static pthread_mutex_t getrf_lock = PTHREAD_MUTEX_INITIALIZER; -#elif defined(USE_PTHREAD_SPINLOCK) -static pthread_spinlock_t getrf_lock = 0; +#if (__STDC_VERSION__ >= 201112L) +#define atomic_load_long(p) __atomic_load_n(p, __ATOMIC_RELAXED) +#define atomic_store_long(p, v) __atomic_store_n(p, v, __ATOMIC_RELAXED) #else -static BLASULONG getrf_lock = 0UL; +#define atomic_load_long(p) (BLASLONG)(*(volatile BLASLONG*)(p)) +#define atomic_store_long(p, v) (*(volatile BLASLONG *)(p)) = (v) #endif -#if defined(USE_PTHREAD_LOCK) -static pthread_mutex_t getrf_flag_lock = PTHREAD_MUTEX_INITIALIZER; -#elif defined(USE_PTHREAD_SPINLOCK) -static pthread_spinlock_t getrf_flag_lock = 0; -#else -static BLASULONG getrf_flag_lock = 0UL; -#endif - - static __inline BLASLONG FORMULA1(BLASLONG M, BLASLONG N, BLASLONG IS, BLASLONG BK, BLASLONG T) { @@ -119,11 +110,7 @@ static void inner_basic_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *ra FLOAT *d = (FLOAT *)args -> b + (k + k * lda) * COMPSIZE; FLOAT *sbb = sb; -#if __STDC_VERSION__ >= 201112L - _Atomic BLASLONG *flag = (_Atomic BLASLONG *)args -> d; -#else volatile BLASLONG *flag = (volatile BLASLONG *)args -> d; -#endif blasint *ipiv = (blasint *)args -> c; @@ -180,7 +167,10 @@ static void inner_basic_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *ra } } - if ((js + REAL_GEMM_R >= n) && (mypos >= 0)) flag[mypos * CACHE_LINE_SIZE] = 0; + if ((js + REAL_GEMM_R >= n) && (mypos >= 0)) { + MB; + atomic_store_long(&flag[mypos * CACHE_LINE_SIZE], 0); + } for (is = 0; is < m; is += GEMM_P){ min_i = m - is; @@ -201,14 +191,10 @@ static void inner_basic_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *ra /* Non blocking implementation */ typedef struct { -#if __STDC_VERSION__ >= 201112L - _Atomic -#else - volatile -#endif - BLASLONG working[MAX_CPU_NUMBER][CACHE_LINE_SIZE * DIVIDE_RATE]; + volatile BLASLONG working[MAX_CPU_NUMBER][CACHE_LINE_SIZE * DIVIDE_RATE]; } job_t; + #define ICOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_ITCOPY(M, N, (FLOAT *)(A) + ((Y) + (X) * (LDA)) * COMPSIZE, LDA, BUFFER); #define OCOPY_OPERATION(M, N, A, LDA, X, Y, BUFFER) GEMM_ONCOPY(M, N, (FLOAT *)(A) + ((X) + (Y) * (LDA)) * COMPSIZE, LDA, BUFFER); @@ -246,11 +232,8 @@ static int inner_advanced_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG * blasint *ipiv = (blasint *)args -> c; BLASLONG jw; -#if __STDC_VERSION__ >= 201112L - _Atomic BLASLONG *flag = (_Atomic BLASLONG *)args -> d; -#else volatile BLASLONG *flag = (volatile BLASLONG *)args -> d; -#endif + if (args -> a == NULL) { TRSM_ILTCOPY(k, k, (FLOAT *)args -> b, lda, 0, sb); sbb = (FLOAT *)((((BLASULONG)(sb + k * k * COMPSIZE) + GEMM_ALIGN) & ~GEMM_ALIGN) + GEMM_OFFSET_B); @@ -280,10 +263,9 @@ static int inner_advanced_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG * #if 1 { do { - LOCK_COMMAND(&getrf_lock); - jw = job[mypos].working[i][CACHE_LINE_SIZE * bufferside]; - UNLOCK_COMMAND(&getrf_lock); + jw = atomic_load_long(&job[mypos].working[i][CACHE_LINE_SIZE * bufferside]); } while (jw); + MB; } #else while (job[mypos].working[i][CACHE_LINE_SIZE * bufferside]) {}; @@ -326,21 +308,17 @@ static int inner_advanced_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG * } MB; for (i = 0; i < args -> nthreads; i++) { - LOCK_COMMAND(&getrf_lock); - job[mypos].working[i][CACHE_LINE_SIZE * bufferside] = (BLASLONG)buffer[bufferside]; - UNLOCK_COMMAND(&getrf_lock); + atomic_store_long(&job[mypos].working[i][CACHE_LINE_SIZE * bufferside], (BLASLONG)buffer[bufferside]); } } - LOCK_COMMAND(&getrf_flag_lock); - flag[mypos * CACHE_LINE_SIZE] = 0; - UNLOCK_COMMAND(&getrf_flag_lock); + MB; + atomic_store_long(&flag[mypos * CACHE_LINE_SIZE], 0); if (m == 0) { + MB; for (xxx = 0; xxx < DIVIDE_RATE; xxx++) { - LOCK_COMMAND(&getrf_lock); - job[mypos].working[mypos][CACHE_LINE_SIZE * xxx] = 0; - UNLOCK_COMMAND(&getrf_lock); + atomic_store_long(&job[mypos].working[mypos][CACHE_LINE_SIZE * xxx], 0); } } @@ -366,10 +344,9 @@ static int inner_advanced_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG * if ((current != mypos) && (!is)) { #if 1 do { - LOCK_COMMAND(&getrf_lock); - jw = job[current].working[mypos][CACHE_LINE_SIZE * bufferside]; - UNLOCK_COMMAND(&getrf_lock); - } while (jw == 0); + jw = atomic_load_long(&job[current].working[mypos][CACHE_LINE_SIZE * bufferside]); + } while (jw == 0); + MB; #else while(job[current].working[mypos][CACHE_LINE_SIZE * bufferside] == 0) {}; #endif @@ -381,9 +358,7 @@ static int inner_advanced_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG * MB; if (is + min_i >= m) { - LOCK_COMMAND(&getrf_lock); - job[current].working[mypos][CACHE_LINE_SIZE * bufferside] = 0; - UNLOCK_COMMAND(&getrf_lock); + atomic_store_long(&job[current].working[mypos][CACHE_LINE_SIZE * bufferside], 0); } } @@ -397,10 +372,9 @@ static int inner_advanced_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG * for (xxx = 0; xxx < DIVIDE_RATE; xxx++) { #if 1 do { - LOCK_COMMAND(&getrf_lock); - jw = job[mypos].working[i][CACHE_LINE_SIZE *xxx]; - UNLOCK_COMMAND(&getrf_lock); + jw = atomic_load_long(&job[mypos].working[i][CACHE_LINE_SIZE *xxx]); } while(jw != 0); + MB; #else while (job[mypos].working[i][CACHE_LINE_SIZE * xxx] ) {}; #endif @@ -443,12 +417,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, #ifdef _MSC_VER BLASLONG flag[MAX_CPU_NUMBER * CACHE_LINE_SIZE]; #else -#if __STDC_VERSION__ >= 201112L - _Atomic -#else - volatile -#endif - BLASLONG flag[MAX_CPU_NUMBER * CACHE_LINE_SIZE] __attribute__((aligned(128))); + volatile BLASLONG flag[MAX_CPU_NUMBER * CACHE_LINE_SIZE] __attribute__((aligned(128))); #endif #ifndef COMPLEX @@ -543,7 +512,11 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, if (width > mn - is - bk) width = mn - is - bk; } - if (num_cpu > 0) exec_blas_async_wait(num_cpu, &queue[0]); + + if (num_cpu > 0) { + WMB; + exec_blas_async_wait(num_cpu, &queue[0]); + } mm = m - bk - is; nn = n - bk - is; @@ -608,7 +581,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, queue[num_cpu].sa = NULL; queue[num_cpu].sb = NULL; queue[num_cpu].next = &queue[num_cpu + 1]; - flag[num_cpu * CACHE_LINE_SIZE] = 1; + atomic_store_long(&flag[num_cpu * CACHE_LINE_SIZE], 1); num_cpu ++; @@ -637,6 +610,8 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, if (num_cpu > 0) { queue[num_cpu - 1].next = NULL; + WMB; + exec_blas_async(0, &queue[0]); inner_basic_thread(&newarg, NULL, range_n_mine, sa, sbb, -1); @@ -647,14 +622,10 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, for (i = 0; i < num_cpu; i ++) { #if 1 - LOCK_COMMAND(&getrf_flag_lock); - f=flag[i*CACHE_LINE_SIZE]; - UNLOCK_COMMAND(&getrf_flag_lock); - while (f!=0) { - LOCK_COMMAND(&getrf_flag_lock); - f=flag[i*CACHE_LINE_SIZE]; - UNLOCK_COMMAND(&getrf_flag_lock); - }; + do { + f = atomic_load_long(&flag[i*CACHE_LINE_SIZE]); + } while (f != 0); + MB; #else while (flag[i*CACHE_LINE_SIZE]) {}; #endif @@ -719,12 +690,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, BLASLONG range[MAX_CPU_NUMBER + 1]; BLASLONG width, nn, num_cpu; -#if __STDC_VERSION__ >= 201112L - _Atomic -#else - volatile -#endif - BLASLONG flag[MAX_CPU_NUMBER * CACHE_LINE_SIZE] __attribute__((aligned(128))); + volatile BLASLONG flag[MAX_CPU_NUMBER * CACHE_LINE_SIZE] __attribute__((aligned(128))); #ifndef COMPLEX #ifdef XDOUBLE @@ -833,6 +799,8 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, nn = n - bk - is; if (width > nn) width = nn; + WMB; + if (num_cpu > 1) exec_blas_async_wait(num_cpu - 1, &queue[1]); range[0] = 0; @@ -867,7 +835,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, queue[num_cpu].sa = NULL; queue[num_cpu].sb = NULL; queue[num_cpu].next = &queue[num_cpu + 1]; - flag[num_cpu * CACHE_LINE_SIZE] = 1; + atomic_store_long(&flag[num_cpu * CACHE_LINE_SIZE], 1); num_cpu ++; } @@ -882,6 +850,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, range_n_new[0] = offset + is; range_n_new[1] = offset + is + bk; + WMB; if (num_cpu > 1) { exec_blas_async(1, &queue[1]); @@ -917,7 +886,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, #endif - for (i = 1; i < num_cpu; i ++) while (flag[i * CACHE_LINE_SIZE]) {}; + for (i = 1; i < num_cpu; i ++) while (atomic_load_long(&flag[i * CACHE_LINE_SIZE])) {}; TRSM_ILTCOPY(bk, bk, a + (is + is * lda) * COMPSIZE, lda, 0, sb); diff --git a/param.h b/param.h index 70c5945ae..68e429734 100644 --- a/param.h +++ b/param.h @@ -1507,8 +1507,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define SYMV_P 8 -#define SWITCH_RATIO 32 -#define GEMM_PREFERED_SIZE 16 +#if defined(XDOUBLE) || defined(DOUBLE) +#define SWITCH_RATIO 4 +#define GEMM_PREFERED_SIZE 4 +#else +#define SWITCH_RATIO 8 +#define GEMM_PREFERED_SIZE 8 +#endif #ifdef ARCH_X86 @@ -1627,8 +1632,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define SYMV_P 8 -#define SWITCH_RATIO 32 -#define GEMM_PREFERED_SIZE 32 +#if defined(XDOUBLE) || defined(DOUBLE) +#define SWITCH_RATIO 8 +#define GEMM_PREFERED_SIZE 8 +#else +#define SWITCH_RATIO 16 +#define GEMM_PREFERED_SIZE 16 +#endif #define USE_SGEMM_KERNEL_DIRECT 1 #ifdef ARCH_X86 @@ -1650,14 +1660,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #else #define SGEMM_DEFAULT_UNROLL_M 16 -#define DGEMM_DEFAULT_UNROLL_M 4 +#define DGEMM_DEFAULT_UNROLL_M 16 #define QGEMM_DEFAULT_UNROLL_M 2 #define CGEMM_DEFAULT_UNROLL_M 8 #define ZGEMM_DEFAULT_UNROLL_M 4 #define XGEMM_DEFAULT_UNROLL_M 1 #define SGEMM_DEFAULT_UNROLL_N 4 -#define DGEMM_DEFAULT_UNROLL_N 8 +#define DGEMM_DEFAULT_UNROLL_N 2 #define QGEMM_DEFAULT_UNROLL_N 2 #define CGEMM_DEFAULT_UNROLL_N 2 #define ZGEMM_DEFAULT_UNROLL_N 2 @@ -1690,23 +1700,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #else -#define SGEMM_DEFAULT_P 768 -#define DGEMM_DEFAULT_P 384 +#define SGEMM_DEFAULT_P 640 +#define DGEMM_DEFAULT_P 192 #define CGEMM_DEFAULT_P 384 #define ZGEMM_DEFAULT_P 256 -#ifdef WINDOWS_ABI -#define SGEMM_DEFAULT_Q 192 -#define DGEMM_DEFAULT_Q 168 -#else -#define SGEMM_DEFAULT_Q 192 -#define DGEMM_DEFAULT_Q 168 -#endif +#define SGEMM_DEFAULT_Q 320 +#define DGEMM_DEFAULT_Q 384 #define CGEMM_DEFAULT_Q 192 #define ZGEMM_DEFAULT_Q 128 #define SGEMM_DEFAULT_R sgemm_r -#define DGEMM_DEFAULT_R 13824 +#define DGEMM_DEFAULT_R 8640 #define CGEMM_DEFAULT_R cgemm_r #define ZGEMM_DEFAULT_R zgemm_r @@ -1717,16 +1722,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define XGEMM_DEFAULT_R xgemm_r #define XGEMM_DEFAULT_Q 128 -#define CGEMM3M_DEFAULT_UNROLL_N 8 -#define CGEMM3M_DEFAULT_UNROLL_M 4 -#define ZGEMM3M_DEFAULT_UNROLL_N 8 -#define ZGEMM3M_DEFAULT_UNROLL_M 2 +#define CGEMM3M_DEFAULT_UNROLL_N 4 +#define CGEMM3M_DEFAULT_UNROLL_M 8 +#define ZGEMM3M_DEFAULT_UNROLL_N 4 +#define ZGEMM3M_DEFAULT_UNROLL_M 4 -#define CGEMM3M_DEFAULT_P 448 -#define ZGEMM3M_DEFAULT_P 224 +#define CGEMM3M_DEFAULT_P 320 +#define ZGEMM3M_DEFAULT_P 256 #define XGEMM3M_DEFAULT_P 112 -#define CGEMM3M_DEFAULT_Q 224 -#define ZGEMM3M_DEFAULT_Q 224 +#define CGEMM3M_DEFAULT_Q 320 +#define ZGEMM3M_DEFAULT_Q 256 #define XGEMM3M_DEFAULT_Q 224 #define CGEMM3M_DEFAULT_R 12288 #define ZGEMM3M_DEFAULT_R 12288 @@ -2598,7 +2603,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #if defined(CORTEXA53) || defined(CORTEXA57) || \ defined(CORTEXA72) || defined(CORTEXA73) || \ - defined(FALKOR) || defined(TSV110) + defined(FALKOR) || defined(TSV110) || defined(EMAG8180) #define SGEMM_DEFAULT_UNROLL_M 16 #define SGEMM_DEFAULT_UNROLL_N 4 @@ -2615,7 +2620,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. /*FIXME: this should be using the cache size, but there is currently no easy way to query that on ARM. So if getarch counted more than 8 cores we simply assume the host is a big desktop or server with abundant cache rather than a phone or embedded device */ -#if NUM_CORES > 8 +#if NUM_CORES > 8 || defined(TSV110) || defined(EMAG8180) #define SGEMM_DEFAULT_P 512 #define DGEMM_DEFAULT_P 256 #define CGEMM_DEFAULT_P 256 @@ -2700,6 +2705,35 @@ is a big desktop or server with abundant cache rather than a phone or embedded d #define CGEMM_DEFAULT_R 4096 #define ZGEMM_DEFAULT_R 4096 +#elif defined(NEOVERSEN1) + +#define SGEMM_DEFAULT_UNROLL_M 16 +#define SGEMM_DEFAULT_UNROLL_N 4 + +#define DGEMM_DEFAULT_UNROLL_M 8 +#define DGEMM_DEFAULT_UNROLL_N 4 + +#define CGEMM_DEFAULT_UNROLL_M 8 +#define CGEMM_DEFAULT_UNROLL_N 4 + +#define ZGEMM_DEFAULT_UNROLL_M 4 +#define ZGEMM_DEFAULT_UNROLL_N 4 + +#define SGEMM_DEFAULT_P 128 +#define DGEMM_DEFAULT_P 160 +#define CGEMM_DEFAULT_P 128 +#define ZGEMM_DEFAULT_P 128 + +#define SGEMM_DEFAULT_Q 352 +#define DGEMM_DEFAULT_Q 128 +#define CGEMM_DEFAULT_Q 224 +#define ZGEMM_DEFAULT_Q 112 + +#define SGEMM_DEFAULT_R 4096 +#define DGEMM_DEFAULT_R 4096 +#define CGEMM_DEFAULT_R 4096 +#define ZGEMM_DEFAULT_R 4096 + #else // Other/undetected ARMv8 cores #define SGEMM_DEFAULT_UNROLL_M 16 diff --git a/utest/CMakeLists.txt b/utest/CMakeLists.txt index 1e3051a8f..dc5175fc5 100644 --- a/utest/CMakeLists.txt +++ b/utest/CMakeLists.txt @@ -6,7 +6,9 @@ if (MSVC AND "${CMAKE_C_COMPILER_ID}" MATCHES Clang) else () set(OpenBLAS_utest_src utest_main.c + test_min.c test_amax.c + test_ismin.c test_rotmg.c test_rot.c test_axpy.c diff --git a/utest/Makefile b/utest/Makefile index 8c7e6b9f8..0b9892411 100644 --- a/utest/Makefile +++ b/utest/Makefile @@ -11,7 +11,7 @@ UTESTBIN=openblas_utest include $(TOPDIR)/Makefile.system -OBJS=utest_main.o test_amax.o test_rotmg.o test_axpy.o test_dotu.o test_dsdot.o test_swap.o test_rot.o +OBJS=utest_main.o test_min.o test_amax.o test_ismin.o test_rotmg.o test_axpy.o test_dotu.o test_dsdot.o test_swap.o test_rot.o #test_rot.o test_swap.o test_axpy.o test_dotu.o test_dsdot.o test_fork.o ifneq ($(NO_LAPACK), 1) @@ -31,6 +31,10 @@ OBJS += test_fork.o endif endif +ifeq ($(C_COMPILER), PGI) +OBJS = utest_main2.o +endif + all : run_test $(UTESTBIN): $(OBJS) diff --git a/utest/test_amax.c b/utest/test_amax.c index 411598410..831804027 100644 --- a/utest/test_amax.c +++ b/utest/test_amax.c @@ -40,6 +40,17 @@ CTEST(amax, samax){ te_max=BLASFUNC(samax)(&N, x, &inc); tr_max=3.3; - + ASSERT_DBL_NEAR_TOL((double)(tr_max), (double)(te_max), SINGLE_EPS); } + +CTEST(amax, damax){ + blasint N=3, inc=1; + double te_max=0.0, tr_max=0.0; + double x[]={-1.1, 2.2, -3.3}; + + te_max=BLASFUNC(damax)(&N, x, &inc); + tr_max=3.3; + + ASSERT_DBL_NEAR_TOL((double)(tr_max), (double)(te_max), DOUBLE_EPS); +} diff --git a/utest/test_ismin.c b/utest/test_ismin.c new file mode 100644 index 000000000..f23d6b545 --- /dev/null +++ b/utest/test_ismin.c @@ -0,0 +1,89 @@ +/***************************************************************************** +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 COPYRIGHT OWNER 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 "openblas_utest.h" + +#define ELEMENTS 50 +#define INCREMENT 2 + +CTEST(ismin, positive_step_2){ + blasint i; + blasint N = ELEMENTS, inc = INCREMENT; + float x[ELEMENTS * INCREMENT]; + for (i = 0; i < N * inc; i ++) { + x[i] = i + 1000; + } + + x[8 * inc] = 0; + blasint index = BLASFUNC(ismin)(&N, x, &inc); + ASSERT_EQUAL(9, index); +} + +CTEST(ismin, negative_step_2){ + blasint i; + blasint N = ELEMENTS, inc = INCREMENT; + float x[ELEMENTS * INCREMENT]; + for (i = 0; i < N * inc; i ++) { + x[i] = - i - 1000; + } + + x[8 * inc] = -123456.0f; + blasint index = BLASFUNC(ismin)(&N, x, &inc); + ASSERT_EQUAL(9, index); +} + +CTEST(ismax, positive_step_2){ + blasint i; + blasint N = ELEMENTS, inc = INCREMENT; + float x[ELEMENTS * INCREMENT]; + for (i = 0; i < N * inc; i ++) { + x[i] = i + 1000; + } + + x[8 * inc] = 123456.0f; + blasint index = BLASFUNC(ismax)(&N, x, &inc); + ASSERT_EQUAL(9, index); +} + +CTEST(ismax, negative_step_2){ + blasint i; + blasint N = ELEMENTS, inc = INCREMENT; + float x[ELEMENTS * INCREMENT]; + for (i = 0; i < N * inc; i ++) { + x[i] = - i - 1000; + } + + x[8 * inc] = 0; + blasint index = BLASFUNC(ismax)(&N, x, &inc); + ASSERT_EQUAL(9, index); +} diff --git a/utest/test_min.c b/utest/test_min.c new file mode 100644 index 000000000..fd31b5982 --- /dev/null +++ b/utest/test_min.c @@ -0,0 +1,100 @@ +/***************************************************************************** +Copyright (c) 2011-2016, The OpenBLAS Project +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are +met: + + 1. Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + + 2. Redistributions in binary form must reproduce the above copyright + notice, this list of conditions and the following disclaimer in + the documentation and/or other materials provided with the + distribution. + 3. Neither the name of the OpenBLAS project nor the names of + its contributors may be used to endorse or promote products + derived from this software without specific prior written + permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER 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 "openblas_utest.h" + +CTEST(min, smin_negative){ + blasint N=3, inc=1; + float te_min=0.0, tr_min=0.0; + float x[]={-1.1, -2.2, -3.3}; + + te_min=BLASFUNC(smin)(&N, x, &inc); + tr_min=-3.3; + + ASSERT_DBL_NEAR_TOL((double)(tr_min), (double)(te_min), SINGLE_EPS); +} + +CTEST(min, dmin_positive){ + blasint N=3, inc=1; + double te_min=0.0, tr_min=0.0; + double x[]={1.1, 0.0, 3.3}; + + te_min=BLASFUNC(dmin)(&N, x, &inc); + tr_min=0.0; + + ASSERT_DBL_NEAR_TOL((double)(tr_min), (double)(te_min), DOUBLE_EPS); +} + +CTEST(min, smin_zero){ + blasint N=3, inc=1; + float te_min=0.0, tr_min=0.0; + float x[]={1.1, 2.2, 0.0}; + + te_min=BLASFUNC(smin)(&N, x, &inc); + tr_min=0.0; + + ASSERT_DBL_NEAR_TOL((double)(tr_min), (double)(te_min), SINGLE_EPS); +} + +CTEST(max, smax_negative){ + blasint N=3, inc=1; + float te_max=0.0, tr_max=0.0; + float x[]={-1.1, -2.2, -3.3}; + + te_max=BLASFUNC(smax)(&N, x, &inc); + tr_max=-1.1; + + ASSERT_DBL_NEAR_TOL((double)(tr_max), (double)(te_max), SINGLE_EPS); +} + +CTEST(max, dmax_positive){ + blasint N=3, inc=1; + double te_max=0.0, tr_max=0.0; + double x[]={1.1, 0.0, 3.3}; + + te_max=BLASFUNC(dmax)(&N, x, &inc); + tr_max=3.3; + + ASSERT_DBL_NEAR_TOL((double)(tr_max), (double)(te_max), DOUBLE_EPS); +} + +CTEST(max, smax_zero){ + blasint N=3, inc=1; + float te_max=0.0, tr_max=0.0; + float x[]={-1.1, -2.2, 0.0}; + + te_max=BLASFUNC(smax)(&N, x, &inc); + tr_max=0.0; + + ASSERT_DBL_NEAR_TOL((double)(tr_max), (double)(te_max), SINGLE_EPS); +} diff --git a/utest/utest_main2.c b/utest/utest_main2.c index aa95a5a3f..6b252863a 100644 --- a/utest/utest_main2.c +++ b/utest/utest_main2.c @@ -50,6 +50,17 @@ CTEST(amax, samax){ ASSERT_DBL_NEAR_TOL((double)(tr_max), (double)(te_max), SINGLE_EPS); } +CTEST(amax, damax){ + blasint N=3, inc=1; + double te_max=0.0, tr_max=0.0; + double x[]={-1.1, 2.2, -3.3}; + + te_max=BLASFUNC(damax)(&N, x, &inc); + tr_max=3.3; + + ASSERT_DBL_NEAR_TOL((double)(tr_max), (double)(te_max), DOUBLE_EPS); +} + CTEST (drotmg,rotmg) { double te_d1, tr_d1; @@ -508,9 +519,82 @@ CTEST(swap,cswap_inc_0) } } +CTEST(min, smin_negative){ + blasint N=3, inc=1; + float te_min=0.0, tr_min=0.0; + float x[]={-1.1, -2.2, -3.3}; + + te_min=BLASFUNC(smin)(&N, x, &inc); + tr_min=-3.3; + + ASSERT_DBL_NEAR_TOL((double)(tr_min), (double)(te_min), SINGLE_EPS); +} + +CTEST(min, dmin_positive){ + blasint N=3, inc=1; + double te_min=0.0, tr_min=0.0; + double x[]={1.1, 0.0, 3.3}; + + te_min=BLASFUNC(dmin)(&N, x, &inc); + tr_min=0.0; + + ASSERT_DBL_NEAR_TOL((double)(tr_min), (double)(te_min), DOUBLE_EPS); +} + +CTEST(min, smin_zero){ + blasint N=3, inc=1; + float te_min=0.0, tr_min=0.0; + float x[]={1.1, 2.2, 0.0}; + + te_min=BLASFUNC(smin)(&N, x, &inc); + tr_min=0.0; + + ASSERT_DBL_NEAR_TOL((double)(tr_min), (double)(te_min), SINGLE_EPS); +} + +CTEST(max, smax_negative){ + blasint N=3, inc=1; + float te_max=0.0, tr_max=0.0; + float x[]={-1.1, -2.2, -3.3}; + + te_max=BLASFUNC(smax)(&N, x, &inc); + tr_max=-1.1; + + ASSERT_DBL_NEAR_TOL((double)(tr_max), (double)(te_max), SINGLE_EPS); +} + +CTEST(max, dmax_positive){ + blasint N=3, inc=1; + double te_max=0.0, tr_max=0.0; + double x[]={1.1, 0.0, 3.3}; + + te_max=BLASFUNC(dmax)(&N, x, &inc); + tr_max=3.3; + + ASSERT_DBL_NEAR_TOL((double)(tr_max), (double)(te_max), DOUBLE_EPS); +} + +CTEST(max, smax_zero){ + blasint N=3, inc=1; + float te_max=0.0, tr_max=0.0; + float x[]={-1.1, -2.2, 0.0}; + + te_max=BLASFUNC(smax)(&N, x, &inc); + tr_max=0.0; + + ASSERT_DBL_NEAR_TOL((double)(tr_max), (double)(te_max), SINGLE_EPS); +} + int main(int argc, const char ** argv){ - CTEST_ADD(amax, samax); + CTEST_ADD (amax, samax); + CTEST_ADD (amax, damax); + CTEST_ADD (min, smin_negative); + CTEST_ADD (min, dmin_positive); + CTEST_ADD (min, smin_zero); + CTEST_ADD (max, smax_negative); + CTEST_ADD (max, dmax_positive); + CTEST_ADD (max, smax_zero); CTEST_ADD (drotmg,rotmg); CTEST_ADD (drotmg,rotmg_issue1452); CTEST_ADD (drotmg,rotmg_D1eqD2_X1eqX2);