[gmx-users] Link to Intel MKL (fftw) via cmake options

[Szilárd Páll]

Please keep all discussions on the mailing list! Also, I'm also CC-ing
to the gmx-devel list, maybe somebody over there has an better what
causes your CMake issue.

>>> //Flags used by the compiler during all build types
>>> CMAKE_C_FLAGS:STRING=' -msse2 -ip -funroll-all-loops -std=gnu99  '
>>>
>>> //Flags used by the compiler during release builds.
>>> CMAKE_C_FLAGS_RELEASE:STRING=-mtune=itanium2 -mtune=core2  -O3 -DNDEBUG
>>> ...
>>> These are obviously the wrong flags for the detected architecture,
>>> sse2 is no longer available and so are the the mtune architectures.
>>
>> Indeed, some of those flags are wrong (CXX -std=gnu99), others simply
>> outdated (-mtune=itanium2). There are plans to correct these issues
>> for 4.6, but let's agree on the fact that they are not show-stoppers
>> (FYI you won't get more than a few percent speedup with the right
>> compiler options, the most I've seen is ~5%.)
>
> OK, I was under the impression it would be simple to correct that -
> independend of gains if any.

Theoretically it is, but in practice not so much. Keeping up with the
ever changing compiler options - especially the architecture-related
optimization flags - is not a fun job. We'll try to improve these
aspects of the CMkake build system the next release.

>> If you want to get rid of them, the solution is pretty simple: just
>> remove the offending options and pick new ones if you think they work
>> better for you.
>
> That involves hand-editing the CMakeCache which may be no less
> bothersome than enduring the warnings ;-)

You can use ccmake (or cmake-gui), they will make your life easier.
Otherwise, editing the cache is nowhere near as difficult as it
sounds. ;)

>> However, I'm not sure what do you mean by "sse2 is no longer
>> available"? Of course it is, both the Intel Compiler and the majority
>> of current processors support it. Also, -mtune=core2 is perfectly
>> valid as is -msse2.
>
> I checked that with the documentation on the Intel site and
> the code generation options written out by the compiler (as
> help info). On the Intel site, -msse2 and -mtune=core2 are
> still listed as valid, whereas they aren't mentioned in the
> help provided by the compiler I installed. Maybe they are
> redundant with the Intel64-compiler and set by default
> (not entirely sure). Bottom line: you are right here, I
> missed some information.

i) Indded the "-msse2" is redundant. From the icc man page:

"On  Linux  systems, if neither -x nor -m is specified, the default is -msse2."

ii) Nope, the default as usually is -mtune=generic on "IA-32" and
"Intel 64" - as Intel likes to call them.

Btw, core2 is essentially equivalent with -xSSSE3. Depending on the
processors you intend to run on you can try -xSSE4.1 or -xSSE4.2, but
I doubt you'll see a noticeable difference.

>>> CMAKE_CXX_FLAGS:STRING=' -msse3 -ip -funroll-all-loops -std=gnu99 '
>>
>> As Mark said, SSE3 won't help you in any noticeable way.
>
> The only 'help' I envisioned here was to avoid screen cluttering
> by loads of warnings. mMy suggested 'fixes' weren't really fixes,
> as I can see now. In my own development work, I can't stand
> pages of recurring warning messages ;-) So I probably
> overreacted here ...

Sure, warnings are not nice and we dislike them as well. That's why
these issues will be corrected in the future.

>>> But the install is broken. On `make install-mdrun`, the scripts would
>>> remove any library from "src/gmxlib/CMakeFiles/CMakeRelink.dir"
>>> and bail out with the error below. Even if you copy the libraries
>>> by hand to CMakeRelink.dir/, the'll get removed by make install-mdrun
>>> before trying to link with them.
>>
>> Now that is a bug. I'd appreciate if you could enter it in
>> redmine.gromacs.org withe all details on how to reproduce it!
>
> I didn't do that until now but examined this further. There is already
> an identical error posted by some other researcher (as you know):
> http://www.mail-archive.com/gmx-users@gromacs.org/msg39228.html
>
> I was able to replicate the error somehow on a different machine without
> any intel compiler around (when building with -DGMX_MPI=ON and cmake).
>
> Gromacs 4.5.5: autoconf/configure on SuSE 11.4/x64 on i7
> ----------------------------------------------------------------
>                 icc/64-12.1.0    gcc/x64/4.5.1
> intThread           OK                OK
> MPI(1.3.2)          OK                OK
>
>
> Gromacs 4.5.5: cmake 2.8.3 on SuSE 11.4/x64 on i7
> ----------------------------------------------------------------
>                 icc/64-12.1.0    gcc/x64/4.5.1
> intThread          error              OK
> MPI(1.3.2)         error             error
>
>
> Gromacs 4.5.5:  autoconf/configure on SuSE 11.4/x64 on Phenom/X4
> ----------------------------------------------------------------
>                 gcc/x64/4.5.1    gcc/x64/4.6.1
> intThread           OK                OK
> MPI(1.3.2)          OK                OK
>
>
> Gromacs 4.5.5:  cmake 2.8.3 on SuSE 11.4/x64 on Phenom/X4
> ----------------------------------------------------------------
>                 gcc/x64/4.5.1    gcc/x64/4.6.1
> intThread           OK                OK
> MPI(1.3.2)         error             error
>
>
> Gromacs/git (2011-10-18): cmake 2.8.3 on SuSE 11.4/x64 on Phenom/X4
> -------------------------------------------------------------------
>                gcc/x64/4.5.1    gcc/x64/4.6.1
> intThread         error              error
> MPI(1.3.2)        error              error
>
>
> Notes: - the MPI is openmpi out-out-the-box w/SuSE 11.4
>       - Gromacs/git - w/autoconf was not tested
>       - Intel-Compiler on AMD Phenom N/A
>       - Gromacs/git failed w/compile error
>         on Boost(?) -> src/gromacs/analysisdata/modules/plot.cpp:60
>            boost::mpl::aux::template_arity<mpl_::na>::value’ is not a
>            valid template argument for type ‘int’ because it is a
>            non-constant expression (many following)
>       - Boost version is 1.4.6 (actual version)
>
> Resumé:
>  - autoconf succeeds in almost all cases with any compiler
>  - cmake fails in almost all cases except the simplest ones (cmake
> ../gromacs-src)
>
> I really don't know what to make out of this.

Me neither, but let's try to figure it out.

First of all, your OpenMPI is quite outdated; as the
build-generation-time message warns OpenMPI <v1.4.1 has known issues.

However, I've just tried gcc 4.6.0 + OpenMPI 1.3 with cmake 2.6.4 as
well as 2.8.2 and in both cases my installation went just fine.
Although I didn't remember seeing you full command line call to CMake,
I tried to replicate a quite common scenario: build & install full
Gromacs with MPI off and "threads" on, then build & install mdrun with
MPI. Here are my command lines:

CMAKE_PREFIX_PATH=/my/path/to/fftw3 cmake ../ -DGMX_MPI=OFF
-DCMAKE_INSTALL_PREFIX=/tmp/gromacs-4.5 && make -j6 && make install

CMAKE_PREFIX_PATH=/my/path/to/fftw3 cmake ../ -DGMX_MPI=ON
-DCMAKE_INSTALL_PREFIX=/tmp/gromacs-4.5 && make mdrun -j6 && make
install-mdrun

Could you please run the above commands and see what you get. If make
install-mdrun still fails, please attach your CMakeCache.txt so we can
take a look at it.

Cheers,
--
Szilárd