Merge pull request #8 from pv/blas64_

ENH: build also openblas64_ 64-bit symbol-suffixed library

Author: tylerjereddy

.travis.yml

@@ -30,6 +30,13 @@ matrix:
       language: objective-c
       env:
         - PLAT=i686
+    - os: linux
+      env:
+        - INTERFACE64=1
+    - os: osx
+      language: objective-c
+      env:
+        - INTERFACE64=1
 
 before_install:
     - source travis-ci/build_steps.sh
@@ -41,7 +48,7 @@ install:
 
 script:
     # Build library and collect into libs subdirectory
-    - build_lib $PLAT
+    - build_lib "$PLAT" "$INTERFACE64"
 
 after_success:
     # Upload libraries to Rackspace container

appveyor.yml

@@ -17,6 +17,10 @@ environment:
   # Downloads via https://sourceforge.net/projects/mingw-w64/files
   # Note spaces instead of %20 in URL, to avoid Windows env var expansion
   matrix:
+    - BUILD_BITS: 64
+      INTERFACE64: 1
+      MINGW_URL: "https://sourceforge.net/projects/mingw-w64/files/Toolchains targetting Win64/Personal Builds/mingw-builds/7.1.0/threads-posix/seh/x86_64-7.1.0-release-posix-seh-rt_v5-rev0.7z/download"
+
     - BUILD_BITS: 32
       MINGW_URL: "https://sourceforge.net/projects/mingw-w64/files/Toolchains targetting Win32/Personal Builds/mingw-builds/7.1.0/threads-posix/dwarf/i686-7.1.0-release-posix-dwarf-rt_v5-rev0.7z/download"
 
@@ -43,14 +47,17 @@ build_script:
   - set SCRIPT_DIR=%START_DIR%appveyor\
   # Need --login to process MSYSTEM variable
   - bash --login %SCRIPT_DIR%build_openblas.sh
-  - bash --login %SCRIPT_DIR%build_gfortran.sh
 
 test_script:
-  - bash --login -c "$(cygpath $START_DIR)test.exe"
+  - bash --login %SCRIPT_DIR%build_gfortran.sh
   - copy test.exe builds
+  - copy test_dyn.exe builds
+  - bash --login -c "$(cygpath $START_DIR)test.exe"
+  - copy %OPENBLAS_ROOT%\%BUILD_BITS%\bin\*.dll .
+  - bash --login -c "$(cygpath $START_DIR)test_dyn.exe"
 
-artifacts:
-  - path: builds\*.*
+on_finish:
+  - ps: Get-ChildItem builds\*.* | % { Push-AppveyorArtifact $_.FullName -FileName $_.Name }
 
 on_success:
   # Upload the generated package to Rackspace

appveyor/build_gfortran.sh

@@ -1,6 +1,14 @@
 # Build gfortran binary against OpenBLAS
 cd $(cygpath "$START_DIR")
 OBP=$(cygpath $OPENBLAS_ROOT\\$BUILD_BITS)
-gfortran -I $OBP/include -o test.exe test.f90 \
-    $OBP/lib/libopenblas_*.a
-./test
+
+static_libname=`find $OBP/lib -maxdepth 1 -type f -name '*.a' \! -name '*.dll.a'`
+dynamic_libname=`find $OBP/lib -maxdepth 1 -type f -name '*.dll.a'`
+
+if [ "$INTERFACE64" == "1" ]; then
+  gfortran -I $OBP/include -fdefault-integer-8 -o test.exe test64_.f90 $static_libname
+  gfortran -I $OBP/include -fdefault-integer-8 -o test_dyn.exe test64_.f90 $dynamic_libname
+else
+  gfortran -I $OBP/include -o test.exe test.f90 $static_libname
+  gfortran -I $OBP/include -o test_dyn.exe test.f90 $dynamic_libname
+fi

appveyor/build_openblas.sh

@@ -46,6 +46,16 @@ fi
 cflags="-O2 -march=$march -mtune=generic $extra"
 fflags="$cflags -frecursive -ffpe-summary=invalid,zero"
 
+# Set suffixed-ILP64 flags
+if [ "$INTERFACE64" == "1" ]; then
+    interface64_flags="INTERFACE64=1 SYMBOLSUFFIX=64_"
+    SYMBOLSUFFIX=64_
+    # We override FCOMMON_OPT, so we need to set default integer manually
+    fflags="$fflags -fdefault-integer-8"
+else
+    interface64_flags=""
+fi
+
 # Build name for output library from gcc version and OpenBLAS commit.
 GCC_TAG="gcc_$(gcc -dumpversion | tr .- _)"
 OPENBLAS_VERSION=$(git describe --tags)
@@ -57,25 +67,41 @@ make BINARY=$BUILD_BITS DYNAMIC_ARCH=1 USE_THREAD=1 USE_OPENMP=0 \
      BUILD_LAPACK_DEPRECATED=1 \
      COMMON_OPT="$cflags" \
      FCOMMON_OPT="$fflags" \
-     MAX_STACK_ALLOC=2048
-make PREFIX=$OPENBLAS_ROOT/$BUILD_BITS install
-DLL_BASENAME=libopenblas_${LIBNAMESUFFIX}
+     MAX_STACK_ALLOC=2048 \
+     $interface64_flags
+make PREFIX=$OPENBLAS_ROOT/$BUILD_BITS $interface64_flags install
+DLL_BASENAME=libopenblas${SYMBOLSUFFIX}_${LIBNAMESUFFIX}
+if [ "$INTERFACE64" == "1" ]; then
+    # OpenBLAS does not build a symbol-suffixed static library on Windows:
+    # do it ourselves
+    static_libname=`find . -maxdepth 1 -type f -name '*.a' \! -name '*.dll.a' -printf '%P\n'`
+    set -x  # echo commands
+    make -C exports $interface64_flags objcopy.def
+    objcopy --redefine-syms exports/objcopy.def "${static_libname}" "${static_libname}.renamed"
+    cp -f "${static_libname}.renamed" "$OPENBLAS_ROOT/$BUILD_BITS/lib/${static_libname}"
+    cp -f "${static_libname}.renamed" "$OPENBLAS_ROOT/$BUILD_BITS/lib/${DLL_BASENAME}.a"
+    set +x
+fi
 cd $OPENBLAS_ROOT
 # Copy library link file for custom name
 cd $BUILD_BITS/lib
 # At least for the mingwpy wheel, we have to use the VC tools to build the
 # export library. Maybe fixed in later binutils by patch referred to in
 # https://sourceware.org/ml/binutils/2016-02/msg00002.html
-cp ${our_wd}/OpenBLAS/exports/libopenblas.def ${DLL_BASENAME}.def
+if [ "$INTERFACE64" == "1" ]; then
+    cp ${our_wd}/OpenBLAS/exports/${DLL_BASENAME}.def ${DLL_BASENAME}.def
+else
+    cp ${our_wd}/OpenBLAS/exports/libopenblas.def ${DLL_BASENAME}.def
+fi
 "$VC9_ROOT/bin/lib.exe" /machine:${vc_arch} /def:${DLL_BASENAME}.def
 cd ../..
 # Build template site.cfg for using this build
 cat > ${BUILD_BITS}/site.cfg.template << EOF
-[openblas]
+[openblas${SYMBOLSUFFIX}]
 libraries = $DLL_BASENAME
 library_dirs = {openblas_root}\\${BUILD_BITS}\\lib
 include_dirs = {openblas_root}\\${BUILD_BITS}\\include
 EOF
-ZIP_NAME="openblas-${OPENBLAS_VERSION}-${plat_tag}-${GCC_TAG}.zip"
+ZIP_NAME="openblas${SYMBOLSUFFIX}-${OPENBLAS_VERSION}-${plat_tag}-${GCC_TAG}.zip"
 zip -r $ZIP_NAME $BUILD_BITS
 cp $ZIP_NAME $our_wd/builds

objconv/LICENSE.txt

@@ -0,0 +1,31 @@
+A utility for cross-platform development of function libraries, for converting
+and modifying object files and for dumping and disassembling object and
+executable files for all x86 and x86-64 platforms.
+
+https://www.agner.org/optimize/#objconv
+
+***
+
+Version 2.49. By Agner Fog © 2018.
+GNU General Public License.
+
+Legal notice
+
+Objconv is an open source program published under the conditions of the GNU General
+Public License, as defined in www.gnu.org/licenses/. The program is provided without any
+warranty or support.
+It may in some cases be illegal to modify, convert or disassemble copyright protected
+software files without permission from the copyright owner. It is an open question whether it
+is legal to modify or convert a copyright protected function library and use it for other
+purposes than presupposed in the license conditions. It is recommended to ask the vendor
+for permission before developing and publishing any software that is built with the use of a
+converted copyright protected function library.
+Copyright law does not generally permit disassembly of copyright protected software for the
+purpose of circumventing a copy protection mechanism, for using part of the code in other
+contexts, or for extracting the algorithms behind the code.
+European, Australian and US copyright law does, however, under certain conditions permit
+reverse engineering of copyright protected software when the purpose is to extract the
+information necessary for establishing interoperability with other software, and only to the
+extent necessary for this purpose. However, I am not a legal expert. The user should seek
+legal advise before deciding whether it is legal to use objconv on copyrighted software for a
+specific purpose.

objconv/changelog.txt

@@ -0,0 +1,71 @@
+2018-10-08 version 2.51
+* fix bug in movq mm0,rax
+
+2018-08-15 version 2.50
+* minor change in omf2cof
+
+2018-01-29 version 2.49
+* fixed minor problem with disassembly of DLL
+
+2017-10-18 version 2.48
+* minor changes
+
+2017-07-27 version 2.47
+* fixed bug in library manager for long module names
+
+2017-07-15 version 2.46
+* dump of ELF files improved
+* a new unnecessary warnings removed
+
+2017-04-10 version 2.45
+* support disassembly of forthcoming Control-flow Enforcement instructions
+
+2016-11-27 version 2.44
+* fixed bug when disassembling file generated by Tiny C compiler
+
+2016-11-09 version 2.43
+* fixed "string table corrupt" message for ELF files
+* support for disassembly of AVX512_4VNNIW, AVX512_4FMAPS instructions
+
+2016-01-16 version 2.42
+* fixed bugs with disassembly of some AVX512 and BMI2 instructions
+
+2015-10-15 version 2.41
+* fixed bugs with disassembly of some AVX512 instructions
+* fixed missing "ptr" in masm syntax disassembly
+
+2015-10-15 version 2.40
+* fixed bugs with disassembly of some AVX512 instructions
+
+2015-09-14 version 2.39
+* fixed bugs with disassembly of some AVX512 instructions
+
+2014-12-15 version 2.38
+* fixed bugs with disassembly of Knights Corner instruction set
+* exe file made compatible with Windows XP
+
+2014-12-06 version 2.37
+* support for disassembly of AVX512BW/DQ, etc.
+* fixed disassembly bug of vpgatherdd on knights corner
+* fixed disassembly bug for FMA4 instructions
+* improved disassembly of switch/case tables for gcc and clang compilers
+
+2014-07-21 version 2.36
+* improved execution times for disassembling very big files
+
+2013-11-27 version 2.32
+* fix bug with .bss section in elf2elf.cpp and elf2coff.cpp
+
+2013-08-23 version 2.31
+* support for disassembly of AVX512F
+* library operations were extremely slow on big libraries because of rebuilding and search for unique names. This problem is fixed. 
+* long member names allowed in libraries
+* option -ls to make member names short. Perhaps good for compatibility with BSD. Member names are made unique just by adding a running hex number, not by comparing with all previous names, which was the cause of slow operations in previous versions.
+* option -ns to change symbol suffixes
+* options -ap and -as to change prefix or suffix and keep old name as alias
+* added missing relocation type in ELF to COFF conversion
+* fixed error that often occurred when disassembling .exe files
+* fixed minor errors
+
+2012-08-31 version 2.16
+* support for disassembly of Knights Corner instruction set

objconv/sort-zipfile.py

@@ -0,0 +1,42 @@
+#!/usr/bin/env python
+"""
+sort-zipfile.py FILENAME
+
+Convert .zip file in-place to zero compression and sorted file names,
+suitable for storing in Git.
+"""
+import zipfile
+import os
+import sys
+import argparse
+import tempfile
+import shutil
+
+def main():
+    p = argparse.ArgumentParser(usage=__doc__)
+    p.add_argument("filename", action="store", type=str)
+    options = p.parse_args()
+
+    mode = zipfile.ZIP_STORED
+
+    fd, target_name = tempfile.mkstemp()
+    try:
+        os.close(fd)
+
+        with zipfile.ZipFile(options.filename) as source:
+            with zipfile.ZipFile(target_name, "w", mode) as target:
+                copy_zip_contents(source, target, mode)
+
+        shutil.move(target_name, options.filename)
+        target_name = None
+    finally:
+        if target_name is not None:
+            os.unlink(target_name)
+
+def copy_zip_contents(source, target, mode):
+    infolist = sorted(source.infolist(), key=lambda info: info.filename)
+    for info in infolist:
+        target.writestr(info, source.read(info.filename), mode)
+
+if __name__ == "__main__":
+    main()

objconv/source.zip

@@ -0,0 +1,80 @@
+!testdpotr_test_gh_2691.f90 with _64 suffix
+subroutine garbage(okflag)
+     implicit none
+     integer, intent(out) :: okflag
+     double precision, dimension(3,3) :: a, a2
+     integer, parameter :: lwork = 100
+     double precision, dimension(lwork) :: work(lwork)
+     integer, dimension(3) :: ipiv
+     integer :: info, i, j
+
+     okflag = 1
+
+     ! condition number =~ 19, well invertible
+     a(1,1) = 0.68534241
+     a(1,2) = 0.63723771
+     a(2,1) = 0.63723771
+     a(1,3) = 0.37423535
+     a(3,1) = 0.37423535
+     a(2,2) = 2.42926786
+     a(2,3) = 2.33541214
+     a(3,2) = 2.33541214
+     a(3,3) = 3.30327538
+
+     a2 = a
+
+     call dpotrf_64('L', 3, a, 3, info)
+     if (info.ne.0) then
+         okflag = 0
+         write(*,*) 'DPOTRF failed'
+         return
+     end if
+     do i = 1, 3
+         do j = 1, i
+             write(*,*) 'DPOTRF result', i, j, a(i,j)
+         end do
+     end do
+     do i = 1, 3
+         do j = i+1, 3
+             a(i,j) = 0
+         end do
+     end do
+     call dpotri_64('L', 3, a, 3, info)
+     if (info.ne.0) then
+         okflag = 0
+         write(*,*) 'DPOTRI failed'
+         return
+     end if
+
+     call dgetrf_64(3, 3, a2, 3, ipiv, info)
+     if (info.ne.0) then
+         okflag = 0
+         write(*,*) 'DGETRF failed'
+         return
+     end if
+     call dgetri_64(3, a2, 3, ipiv, work, lwork, info)
+     if (info.ne.0) then
+         okflag = 0
+         write(*,*) 'DGETRI failed'
+         return
+     end if
+     do i = 1, 3
+         do j = 1, i
+             write(*,*) i, j, a(i,j), a2(i,j)
+             if (abs(a(i,j) - a2(i,j)) > 1e-3 + 1e-3*abs(a(i,j))) then
+                okflag = 0
+                write(*,*) '# po/ge mismatch!'
+             end if
+         end do
+     end do
+end subroutine
+
+program main
+     integer :: okflag
+     call garbage(okflag)
+     if (okflag.eq.1) then
+         write(*,*) 'OK'
+     else
+         write(*,*) 'FAIL'
+     end if
+end program

test64_.f90

@@ -0,0 +1,4 @@
+echo "Building objconv..."
+set -e -x
+unzip source.zip
+g++ -O1 -o objconv *.cpp