[gmx-developers] Re: GROMACS 4.6 pre-release: new algorithms, parallelization schemes, & GPU acceleration

[Mirco Wahab]

Am 01.03.2012 18:55, schrieb Szilárd Páll:
> Not that many have complained, but the one-liner is broken. A correct
> alternative:
> git clone git://git.gromacs.org/gromacs.git -b nbnxn_hybrid_acc&&  cd
> gromacs&&  mkdir buid&&  cd build&&  cmake ../&&  make install -j4

I did a "working" Win64 build (VERSION 4.6-GPU-dev-20120301-d16fc-dirty)
and couldn't find out how do I "select" the gpu?

"mdrun -v -nb gpu" doesn't do anything. Is there something
that allows mdrun to detect a device?

Some small problems encountered during the adventure: You used
at some places a 'erf[c]()' function which is nowhere defined,
and at some other places a 'gmx_erf[c]()'. Can I replace all
erf[c]() calls by gmx_erf[c]()? (This is what I did.)

There are several places where C99 constructs (variable declaration
not on the start of a block) had to be compensated for. One
macro in mdlib/nbnxn_kernels/nbnxn_kernel_sse_inner.h:715
did cause error on expansion, I had to rewrite:

  #ifndef ENERGY_GROUPS
        VvdwtotSSE = gmx_add_pr(VvdwtotSSE,
    #ifndef HALF_LJ
 
gmx_sum4_pr(VLJ_SSE0,VLJ_SSE1,VLJ_SSE2,VLJ_SSE3)
    #else
                                gmx_add_pr(VLJ_SSE0,VLJ_SSE1)
    #endif
                               );
  #else

and hope it's still correct ...


On the cuda side - this seem to compile and link fine so far
(I had to patch one cmake-1.8.7 FindCUDA macro), but there are
some warnings I don't understand. The generated cl options
include -fPIC which doesn't seem to belong to Windows. Maybe
this comes from cmake itself, dunno.

The "default build" for Win64 seems to be "static libs"
and huge binaries. OK, they work fine.

But I can't find out how to get mdrun to select a gpu.

Thanks & regards

M.