Disable AVX-512 on Windows

[ararslan]

There is apparently a GCC bug that prevents building AVX-512 code on Windows. We observed this when building OpenBLAS on the 64-bit Windows Julia buildbot (which I'm told has a Skylake-X processor). Is there a way to disable OpenBLAS building AVX-512 on Windows? Or perhaps a patch that disables it automatically?

[brada4]

But you do not use -O3 , so the bug is not relevant, nor pertains windows globally.

Most likely your 'as' is much older than your gcc, so it does not understand assembly emitted by gcc... It does not choke on openblas code.

[martin-frbg]

Ideally the compile test in c_check (and system_check.cmake) should catch this (and define NO_AVX512 automatically), but from the GCC bug it appears to depend on some particular use of registers that is not reflected in the current test code. (Not entirely sure if we can just borrow their test from the end of https://gcc.gnu.org/viewcvs/gcc/trunk/libgfortran/acinclude.m4?view=markup&pathrev=244636 (GCC PR79127) as that would move it from GPL to three-clause BSD, but it is probably easy to come up with an equivalent test now that the underlying problem is understood.)

[brada4]

Red hat6 could use avx2 test of a class, no gpl does not remove that easily unless you convince author to give it to you without gpl

[martin-frbg]

Perhaps you could try if simply changing the zmm2 register now used in the test at
https://github.com/xianyi/OpenBLAS/blob/9e654305049caf80bf53369304d1fc4e3662ba7e/cmake/system_check.cmake#L70
or
https://github.com/xianyi/OpenBLAS/blob/9e654305049caf80bf53369304d1fc4e3662ba7e/c_check#L207
to zmm16 is sufficient to catch this ? (If not, adding a clobber line similar to what they used in the gfortran PR would hopefully do the trick without causing any licensing complications)

[brada4]

Also worth experimenting in direction of planting avx512 capable binutils to compilers.
Indication one needs 2.24 https://gitlab.com/ultr/glibc/commit/f43cb35c9b3c35addc6dc0f1427caf51786ca1d2
@ararslan - gcc uses libexec/gcc-as in absence of executable as, probably that does not stand true for cross-build.
For inspiration : https://github.com/xianyi/OpenBLAS/wiki/faq#binutils

Can you show as --version and x86_64-w64-mingw32-as --version
Your gcc should have 2.25.1, with AVX512 support, but maybe "system" as lags behind and gets in the way?

[ararslan]

My colleague @staticfloat knows more about the toolchain we're using, perhaps he can chime in here.

[staticfloat]

$ as --version
GNU assembler (GNU Binutils) 2.29.1.20171006
Copyright (C) 2017 Free Software Foundation, Inc.
This program is free software; you may redistribute it under the terms of
the GNU General Public License version 3 or later.
This program has absolutely no warranty.
This assembler was configured for a target of `x86_64-pc-cygwin'.

$ x86_64-w64-mingw32-as --version
GNU assembler (GNU Binutils) 2.29.1.20171006
Copyright (C) 2017 Free Software Foundation, Inc.
This program is free software; you may redistribute it under the terms of
the GNU General Public License version 3 or later.
This program has absolutely no warranty.
This assembler was configured for a target of `x86_64-w64-mingw32'.

We are using a Cygwin-based environment, within which we install a mingw32-based GCC toolchain. Looks like we have pretty recent assemblers, so that's most likely not the issue here. Just to make sure, I ran the failing compiler command again, but with -v to see the toolchain invocations:

$ x86_64-w64-mingw32-gcc -march=x86-64 -m64 -c -O2 -O2 -DMS_ABI -DMAX_STACK_ALLOC=2048 -Wall -m64 -DF_INTERFACE_GFORT  -DDYNAMIC_ARCH -DSMP_SERVER -DNO_WARMUP -DMAX_CPU_NUMBER=16 -DMAX_PARALLEL_NUMBER=1 -march=skylake-avx512 -DASMNAME=ssymv_L_SKYLAKEX -DASMFNAME=ssymv_L_SKYLAKEX_ -DNAME=ssymv_L_SKYLAKEX_ -DCNAME=ssymv_L_SKYLAKEX -DCHAR_NAME=\"ssymv_L_SKYLAKEX_\" -DCHAR_CNAME=\"ssymv_L_SKYLAKEX\" -DNO_AFFINITY -DTS=_SKYLAKEX -I.. -DBUILD_KERNEL -DTABLE_NAME=gotoblas_SKYLAKEX -UDOUBLE  -UCOMPLEX -UCOMPLEX -UDOUBLE -DLOWER ../kernel/x86_64/ssymv_L.c -o ssymv_L_SKYLAKEX.obj -v
Using built-in specs.
COLLECT_GCC=x86_64-w64-mingw32-gcc
Target: x86_64-w64-mingw32
Configured with: /cygdrive/i/szsz/tmpp/cygwin64/mingw64-x86_64/mingw64-x86_64-gcc-6.4.0-2.x86_64/src/gcc-6.4.0/configure --srcdir=/cygdrive/i/szsz/tmpp/cygwin64/mingw64-x86_64/mingw64-x86_64-gcc-6.4.0-2.x86_64/src/gcc-6.4.0 --prefix=/usr --exec-prefix=/usr --localstatedir=/var --sysconfdir=/etc --docdir=/usr/share/doc/mingw64-x86_64-gcc --htmldir=/usr/share/doc/mingw64-x86_64-gcc/html -C --build=x86_64-pc-cygwin --host=x86_64-pc-cygwin --target=x86_64-w64-mingw32 --without-libiconv-prefix --without-libintl-prefix --with-sysroot=/usr/x86_64-w64-mingw32/sys-root --with-build-sysroot=/usr/x86_64-w64-mingw32/sys-root --disable-multilib --disable-win32-registry --enable-languages=c,c++,fortran,lto,objc,obj-c++ --enable-fully-dynamic-string --enable-graphite --enable-libgomp --enable-libquadmath --enable-libquadmath-support --enable-libssp --enable-version-specific-runtime-libs --enable-libgomp --enable-libada --with-dwarf2 --with-gnu-ld --with-gnu-as --with-tune=generic --with-cloog-include=/usr/include/cloog-isl --with-system-zlib --enable-threads=posix --libexecdir=/usr/lib
Thread model: posix
gcc version 6.4.0 (GCC)
COLLECT_GCC_OPTIONS='-march=x86-64' '-c' '-O2' '-O2' '-D' 'MS_ABI' '-D' 'MAX_STACK_ALLOC=2048' '-Wall' '-m64' '-D' 'F_INTERFACE_GFORT' '-D' 'DYNAMIC_ARCH' '-D' 'SMP_SERVER' '-D' 'NO_WARMUP' '-D' 'MAX_CPU_NUMBER=16' '-D' 'MAX_PARALLEL_NUMBER=1' '-march=skylake-avx512' '-D' 'ASMNAME=ssymv_L_SKYLAKEX' '-D' 'ASMFNAME=ssymv_L_SKYLAKEX_' '-D' 'NAME=ssymv_L_SKYLAKEX_' '-D' 'CNAME=ssymv_L_SKYLAKEX' '-D' 'CHAR_NAME="ssymv_L_SKYLAKEX_"' '-D' 'CHAR_CNAME="ssymv_L_SKYLAKEX"' '-D' 'NO_AFFINITY' '-D' 'TS=_SKYLAKEX' '-I' '..' '-D' 'BUILD_KERNEL' '-D' 'TABLE_NAME=gotoblas_SKYLAKEX' '-U' 'DOUBLE' '-U' 'COMPLEX' '-U' 'COMPLEX' '-U' 'DOUBLE' '-D' 'LOWER' '-o' 'ssymv_L_SKYLAKEX.obj' '-v'
 /usr/lib/gcc/x86_64-w64-mingw32/6.4.0/cc1.exe -quiet -v -I .. -D_REENTRANT -D MS_ABI -D MAX_STACK_ALLOC=2048 -D F_INTERFACE_GFORT -D DYNAMIC_ARCH -D SMP_SERVER -D NO_WARMUP -D MAX_CPU_NUMBER=16 -D MAX_PARALLEL_NUMBER=1 -D ASMNAME=ssymv_L_SKYLAKEX -D ASMFNAME=ssymv_L_SKYLAKEX_ -D NAME=ssymv_L_SKYLAKEX_ -D CNAME=ssymv_L_SKYLAKEX -D CHAR_NAME="ssymv_L_SKYLAKEX_" -D CHAR_CNAME="ssymv_L_SKYLAKEX" -D NO_AFFINITY -D TS=_SKYLAKEX -D BUILD_KERNEL -D TABLE_NAME=gotoblas_SKYLAKEX -U DOUBLE -U COMPLEX -U COMPLEX -U DOUBLE -D LOWER ../kernel/x86_64/ssymv_L.c -quiet -dumpbase ssymv_L.c -march=x86-64 -m64 -march=skylake-avx512 -auxbase-strip ssymv_L_SKYLAKEX.obj -O2 -O2 -Wall -version -o /tmp/ccL3UFlg.s
GNU C11 (GCC) version 6.4.0 (x86_64-w64-mingw32)
	compiled by GNU C version 6.4.0, GMP version 6.1.2, MPFR version 3.1.5-p10, MPC version 1.0.3, isl version 0.14 or 0.13
warning: MPFR header version 3.1.5-p10 differs from library version 3.1.6-p1.
warning: MPC header version 1.0.3 differs from library version 1.1.0.
GGC heuristics: --param ggc-min-expand=100 --param ggc-min-heapsize=131072
ignoring nonexistent directory "/usr/x86_64-w64-mingw32/sys-root/usr/local/include"
ignoring nonexistent directory "/usr/lib/gcc/x86_64-w64-mingw32/6.4.0/../../../../x86_64-w64-mingw32/include"
#include "..." search starts here:
#include <...> search starts here:
 ..
 /usr/lib/gcc/x86_64-w64-mingw32/6.4.0/include
 /usr/lib/gcc/x86_64-w64-mingw32/6.4.0/include-fixed
 /usr/x86_64-w64-mingw32/sys-root/mingw/include
End of search list.
GNU C11 (GCC) version 6.4.0 (x86_64-w64-mingw32)
	compiled by GNU C version 6.4.0, GMP version 6.1.2, MPFR version 3.1.5-p10, MPC version 1.0.3, isl version 0.14 or 0.13
warning: MPFR header version 3.1.5-p10 differs from library version 3.1.6-p1.
warning: MPC header version 1.0.3 differs from library version 1.1.0.
GGC heuristics: --param ggc-min-expand=100 --param ggc-min-heapsize=131072
Compiler executable checksum: 5406d774635579063101db625e337f18
COLLECT_GCC_OPTIONS='-march=x86-64' '-c' '-O2' '-O2' '-D' 'MS_ABI' '-D' 'MAX_STACK_ALLOC=2048' '-Wall' '-m64' '-D' 'F_INTERFACE_GFORT' '-D' 'DYNAMIC_ARCH' '-D' 'SMP_SERVER' '-D' 'NO_WARMUP' '-D' 'MAX_CPU_NUMBER=16' '-D' 'MAX_PARALLEL_NUMBER=1' '-march=skylake-avx512' '-D' 'ASMNAME=ssymv_L_SKYLAKEX' '-D' 'ASMFNAME=ssymv_L_SKYLAKEX_' '-D' 'NAME=ssymv_L_SKYLAKEX_' '-D' 'CNAME=ssymv_L_SKYLAKEX' '-D' 'CHAR_NAME="ssymv_L_SKYLAKEX_"' '-D' 'CHAR_CNAME="ssymv_L_SKYLAKEX"' '-D' 'NO_AFFINITY' '-D' 'TS=_SKYLAKEX' '-I' '..' '-D' 'BUILD_KERNEL' '-D' 'TABLE_NAME=gotoblas_SKYLAKEX' '-U' 'DOUBLE' '-U' 'COMPLEX' '-U' 'COMPLEX' '-U' 'DOUBLE' '-D' 'LOWER' '-o' 'ssymv_L_SKYLAKEX.obj' '-v'
 /usr/lib/gcc/x86_64-w64-mingw32/6.4.0/../../../../x86_64-w64-mingw32/bin/as.exe -v -I .. --64 -o ssymv_L_SKYLAKEX.obj /tmp/ccL3UFlg.s
GNU assembler version 2.29.1 (x86_64-w64-mingw32) using BFD version (GNU Binutils) 2.29.1.20171006
/tmp/ccL3UFlg.s: Assembler messages:
/tmp/ccL3UFlg.s:146: Error: invalid register for .seh_savexmm

[martin-frbg]

So could you test please if modifying the c_check code is sufficient to work around lack of avx512 support in mingw ?

[carlkl]

Mayby compiling with -fno-asynchronous-unwind-tables could help. See https://stackoverflow.com/questions/43152633/invalid-register-for-seh-savexmm-in-cygwin

[staticfloat]

Modifying the c_check code to say zmm16 instead of zmm2 is not sufficient; I get the same problem.

Adding -fno-asynchronous-unwind-tables worked! We'll do that for now, perhaps OpenBLAS should do that by default if running on cygwin? (You can check via $($CC -dumpmachine) == *-cygwin)

[martin-frbg]

Thanks. From your description I assume your environment would have been detected as OSNAME=CYGWIN_NT (in the first line of Makefile.conf) already, so Makefile.x86_64 could append the option (where it adds -march=skylake-avx512) based on this ?

[staticfloat]

Yes, that seems correct.