simtk.org

Hi,

I'm having trouble getting the Gromacs preview working on an AMD64 Debian Lenny system with CUDA. Are the libraries/binaries 32-bit? If so, can you provide 64-bit versions?

I have also tried to compile a modified Gromacs 4.0.3 without luck. It might be useful to provide modifications to the Gromacs source as a well-defined patch that can be installed in an automated way. It appears that dependencies from the OpenMM-Source tarball are not being included correctly.

I am looking forward to getting this code running and contributing to the testing process.

If you type 'file <name of binary>' you can see that they're 32-bit. They should still work though, if you install the 32-bit libraries.

I think it's still only CUDA for Linux, so if you want to use the graphics acceleration you'll need to install the 32-bit CUDA toolkit and sdk as well (the drivers for 64-bit linux work fine).

You'll have to make sure the binaries use the right libraries once you've installed them as well.

I managed to get it working on our 64-bit Linux system this way at least. Post the error message if you're still having issues.

Still working on getting it running, but it seems that ia32 - capabilities are misdetected by configure. I get lots of errors like nb_kernel_ia32_3dnow/nb_kernel101_ia32_3dnow.s:776: Error: suffix or operands invalid for `pop'.

./configure --disable-ia32-3dnow --disable-ia32-sse seems to help.

Where are you seeing those error messages? At compile time? When you run it in gdb? Somewhere else?

If you're seeing them in gdb, that's actually normal. When Gromacs starts up, it determines what features your CPU has by trying out a lot of different instructions and seeing which ones work. This confuses the debugger, but actually the program is working normally.

Peter

No, I got those compiling Gromacs4.3pre_for_OpenMMZephyr0.2_Source.tar.gz

I might be on a completely wrong path. I first tried a simple ./configure make make install and got the errors above. I tried the --disable-ia32-.. options for configure and it compiled without problems, but it didn't create a mdrun_openmm anywhere (but a complete gromacs in /usr/local/gromacs).

I then tried a cmake . make mdrun_openmm, make install mdrun_openmm and here again had to disable those two options in CMakeCache.txt:

//Use x86 3DNow! instruction set?
GMX_IA32_3DNOW:BOOL=OFF

//Use x86 SSE instruction set?
GMX_IA32_SSE:BOOL=OFF

And the build (and make install mdrun_openmm) completed without errors. But again, there is no mdrun_openmm anywhere. /usr/local/gromacs_openmm/share and and include/ are created, but nothing else.

I might be doing something completely stupid. Are there instructions for installing from source anywhere? Is it even possible to compile a 64bit version, considering that there is no 64bit CUDA?

I think the build script may just call the executable mdrun, not mdrun_openmm. Try running it. If it prints out the name of the OpenMM platform it's using, you know it's really the OpenMM version.

Peter

I was right, I _was_ doing something stupid. It did just call it mdrun as you said, thank you.

It's not using cuda, but that's probably just because I'm not using AMBER, will have to try that tomorrow.

If it's not using CUDA, that's most likely because you don't have the necessary libraries (openmm/lib and cuda/lib) in your library path (PATH on Windows, LD_LIBRARY_PATH on Linux, DYLD_LIBRARY_PATH on Mac). Or because you didn't copy the OpenMMCuda library to gromacs/lib/openmm.

Peter

how to build 64bit openmm libs? please help me,thank you

Read the forum please, it is already here.