Arch Linux User Repository

EDIT 3: A solution here: https://bbs.archlinux.org/viewtopic.php?pid=1870411#p1870411

I have the same error as @mefistofeles.

EDIT 2: And @malinke too. It's partly because my CPU doesn't support AVX2, I think. There's a double "-march=core-avx2 -march=native" when compiling fftw. It fails because my "-march=native" doesn't actually support AVX2.

#error "compiling simd-avx2.h without avx2 support"

I've also changed my /etc/makepkg.conf as @hseara suggested.

It's a Thinkpad X220 with Sandy Bridge, so I used -DGMX_SIMD=AVX_256 on my PKGBUILD. At least in my case, computer doesn't support avx2, so it failed to compile fftw. Not sure what to do here, it seems fftw is automatically compiled with "--enable-avx2". I'll try again later without DGMX_BUILD_OWN_FFTW=ON on the PKGBUILD.

EDIT 1: So it builds successfully when I replaced DGMX_BUILD_OWN_FFTW=ON with -DGMX_FFT_LIBRARY=fftw3 on the PKGBUILD. Not sure if using pacman-installed fftw with GROMACS will lead to problems later on.

For what it's worth, like @mefistofeles, I also tried to manually build the package using cmake and DGMX_BUILD_OWN_FFTW=ON, and it worked. I wonder what the problem is. From what I can tell, gromacs-2019.4/src/external/build-fftw/CMakeLists.txt says it will build fftw with only either just "--enable-sse2" or the whole "--enable-sse2;--enable-avx;--enable-avx2". The "--enable-avx" always comes with "--enable-avx2". But it worked this time, despite my CPU not supporting AVX 2. When running cmake, it also outputted:

The GROMACS-managed build of FFTW 3 will configure with the following optimizations: --enable-sse2;--enable-avx;--enable-avx2;--enable-avx512

Just for comparison. This line from manual install succeeds:

libtool: compile:  gcc -DHAVE_CONFIG_H -I. -I/home/user/download/gromacs-2019.4/build/src/external/build-fftw/fftwBuild-prefix/src/fftwBuild/dft/simd/avx2 -I../../.. -I /home/user/download/gromacs-2019.4/build/src/external/build-fftw/fftwBuild-prefix/src/fftwBuild -march=core-avx2 -mfma -O3 -fomit-frame-pointer -mtune=native -malign-double -fstrict-aliasing -fno-schedule-insns -MT n1fv_2.lo -MD -MP -MF .deps/n1fv_2.Tpo -c /home/user/download/gromacs-2019.4/build/src/external/build-fftw/fftwBuild-prefix/src/fftwBuild/dft/simd/avx2/n1fv_2.c  -fPIC -DPIC -o n1fv_2.o

This line from aur using pikaur fails:

libtool: compile:  gcc-8 -DHAVE_CONFIG_H -I. -I/home/user/.cache/pikaur/build/gromacs/src/double/src/external/build-fftw/fftwBuild-prefix/src/fftwBuild/dft/simd/avx2 -I../../.. -I /home/user/.cache/pikaur/build/gromacs/src/double/src/external/build-fftw/fftwBuild-prefix/src/fftwBuild -D_FORTIFY_SOURCE=2 -march=core-avx2 -mfma -march=native -O2 -pipe -fno-plt -MT n1fv_2.lo -MD -MP -MF .deps/n1fv_2.Tpo -c /home/user/.cache/pikaur/build/gromacs/src/double/src/external/build-fftw/fftwBuild-prefix/src/fftwBuild/dft/simd/avx2/n1fv_2.c  -fPIC -DPIC -o n1fv_2.o

So possibly the double "-march" thing makes it fail? Just "-march=core-avx2" vs "-march=core-avx2 -march=native". Because my Sandy Bridge ("-march=native" part) doesn't actually support AVX2.

@mefistofeles, either your cpu does not support AVX2 (check in /proc/cpuinfo) or your /etc/makepkg.conf is not set to native architecture:

CFLAGS="-march=native -O2 -pipe -fno-plt"
CXXFLAGS="${CFLAGS}"

PS: -fstack-protector-strong in makepkg.conf is not needed

For some reason using the following PKGBUILD fails compiling, check https://paste.rs/IqF

The error is in https://paste.rs/6W0

Specifically the part that says error: #error "compiling simd-avx2.h without avx2 support" when compiling FFTW.

The funny thing is that when I manually compile using "cmake .. -DGMX_GPU=ON -DGMX_SIMD=AVX2_256 -DGMX_BUILD_OWN_FFTW=ON" it works fine. Don't really know what is causing the error. Help appreciated.

BTW, using the -fstack-protector-strong in makepkg.conf didn't work. as suggested by @hseara

No longer needed in gromacs 2019.4

Installation notes

With cmake update to v3.14.x, gromacs is no longer capable of identifying correctly the CPU and therefore the appropriated SIMD level for compilation. If you want to compile gromacs with SIMD while the problem is not solved add the following option to PKGBUILD:

-DGMX_SIMD={AVX_128_FMA, AVX_256, AVX2_256, AVX2_128, AVX_512}

The KEYWORD depends on the SIMD support provided by your CPU.

Usage notes

source /etc/profile.d/GMXRC.bash

hseara, the current (2016.4 released September 15, 2017) doesn't works with CUDA in old systems. But new gromacs (2018-beta2 released December 12 2017) works perfectly with current PKGBUILD (gromacs 2016.4-2) and CUDA 9.1. So don't wait and update the version.

The current PKGBUILD works perfectly in new systems.

@manlinke: I have the feeling that at least part of your problem is Cuda related. You seem to have an "NVIDIA GeForce GTX 550 Ti" which is no longer supported by Cuda 9. That means that your code will compile well, but tests will fail. The only possible workaround which I'm using on an old system myself is to use Cuda 8 in your system instead of Cuda 9. It is not possible for me to deal with this issue as it is graphics card dependent. You can always disable CUDA "-DGMX_GPU=OFF".

@manlike: The following lines overriding the /etc/makepkg.conf defaults should take care of the avx2 issues:

export CFLAGS="-march=native -O2 -pipe -fstack-protector-strong"

export CXXFLAGS="${CFLAGS}"

Once you fix the Cuda issues, please check again if you have really avx2 issues.

@malinke I was also getting avx2 errors. Specifically this: error: #error "compiling simd-avx2.h without avx2 support", which some guys at #archlinux think it's an error from upstream. The complete error output is at: https://bpaste.net/show/3bbef4217034 (it may have a short lifetime, sorry)

There one can see that the problem is building the fftw, so I tried removing the -DGMX_BUILD_OWN_FFTW=ON \ lines in the PKGBUILD (lines 51 and 62) and got better results but still errors on testing it. The output I'm getting doing what I said is here: https://bpaste.net/show/01d9d68140bd

BTW, when compiling directly the source code by using cmake and make it builds, with not special options, just as it's shown in: http://web.archive.org/web/20170714101612/http://www.gromacs.org/Developer_Zone/Building_and_Testing/Cmake in the "Configuring with default settings" section. Hope this helps to give ideas on how to make gromacs build using this PKGBUILD/AUR?

@malink "(I didn't care if I had a avx build or not)"

You should care. mdrun will be 2x to 4x faster if compiled with avx (depending on the version) activated.

My fix will activate the highest level of avx by default. My fix will allow also CUDA support if available. This is why we need gcc5. CUDA 9 does not seem to solve this issue.

In any case, thanks for your comment as the fix simplifies the installation.

I read the avx instructions. My build problem was that the default pkgbuild would fail since gromacs cmake config tried to enable avx and the global flags in `/etc/makepkg.conf` deactivated it later resulting in a build error (I didn't care if I had a avx build or not). The changed cflags in the current version fix that.

@mailkinke please read the content of the PKGBUILD file. There it is clearly stated:
###### CMAKE OPTIONS DISABLE BY DEFAULT ###########
# If you are using an AVX2 capable CPU, you will #
# not have AVX2 binaries unless you set -march to #
# 'native', your respective architecture flag: #
# https://gcc.gnu.org/onlinedocs/gcc-5.3.0/gcc/x86-Options.html#x86-Options #
# or just include '-mavx2' to the default compiler#
# flags in the /etc/makepkg.conf: #
# https://wiki.archlinux.org/index.php/Makepkg#Architecture.2C_compile_flagsAdd #
###################################################