Fixed rare issue with GPUs. Only for systems with partially empty DD cells, sometimes forces from interactions handled on GPUs on other domains were added to the forces for the correct step, and a subsequent step. Or if there were no interactions computed for this DD cell, then none would be added from neighbouring cells either. Neither condition will arise in normal condensed-phase simulations. #1721
Fixed rare issue where some forms of replica exchange could have written to unallocated memory when permutations of all replicas formed a single cycle (only likely for very low numbers of replicas) #1697
Fixed pbc=xy with 2 walls under the Verlet scheme, which was broken #1660
Fixed free-energy calculations with GPUs and 2D or 3D domain decomposition, which was broken. 1D DD is used much more frequently, and is unaffected. #1661
Fixed Andersen thermostats, where the logic for which type of Andersen thermostat has which implementation got swapped #1695
Re-enabled FMA support on AMD, accidentally removed before 5.0. Simulations were correct, but not as fast as they should have been. #1719
Corrected -ntmpi auto for Intel Nehalem CPUs (and older). Thread-MPI mdrun used to automatically choose 1 MPI rank with up to 16 OpenMP threads on all Intel family-6 CPUs, whereas this was only intended for Sandy Bridge and later.
Changed configure-time FFTW warning against --enable-avx to one against using --enable-avx without --enable-sse (for AVX builds). With FFTW 3.3.4, the performance of FFTW with both SSE and AVX enabled is often a bit better and never much worse than SSE alone. You may get better performance on all kinds of AVX hardware with ftp://ftp.gromacs.org/pub/beta/fftw-...tersimd.tar.gz. These improvements may appear in future releases of FFTW.
Re-fixed PME bug with high OpenMP thread count #1572.
Correct CUDA kernel energy flag The CUDA kernels calculated energies based on the GMX_FORCE_VIRIAL flag. This did not cause wrong results, since (currently) GMX_FORCE_ENERGY is always set when the virial flag is set. But using the latter flag gives a small performance improvement when using pressure coupling.
Fix segmentation fault in gmx covar. With fewer trajectory frames than degrees of freedom in the analysis group, gmx covar would print a warning to a NULL pointer instead of stderr. #1733
Fixed trjconv -box -1. When using -1 in the -box specification, box vectors were instead set to zero rather than being preserved. A fatal error is triggered if the user is trying to preserve a non-existent box to avoid repeating the undesired behavior.
Fixed some error handling code, which might solve some long-standing unknown issues with (e.g.) DSSP mysteriously not working even when the user had done everything right. #1717
gmx hbond no longer removes PBC when none is present #1662
Avoided near-silent generation of zero energies during reruns. There is note printed about this case for Verlet+GPU+energy groups, and this is OK for normal simulations because the decomposition of the energy into groups is unimportant. However, a user doing such a rerun is probably seeking the energy decomposition, and in that case getting zeroes is unexpected behaviour. Instead, default to not using GPUs with mdrun -rerun -nb auto and more than one energy group. In that case with mdrun -rerun -nb gpu, give a fatal error and suggest the probable solutions.
Fixed incorrect output of gb_algorithm as sa_algorithm.
Fixed make uninstall not being able to remove of symlinks created during installation.
Fixed memory-allocation issue with vv and vv-ak integrators. #1669
Corrected RDTSCP warning text.
Fixed convenience feature for developers doing auto-downloaded of the regression test tarball version matching the current source.
Made various minor fixes to test cases and error messages
Facilitated easier linking of fully static binaries. The user has to manually set both -DBUILD_SHARED_EXE=no and perhaps CFLAGS=CXXFLAGS=-static, perhaps manage their own toolchain, and certainly make static libraries available for all dependencies. See install guide. #911 Related to #1641
Fixed linking with static libxml2, which can dependend on libz.
Fixed bugs in FindCUDA.cmake when cross compiling (the fix has been passed on to CMake)
Added a note about sc-coul being automatically turned on. The documentation didn't mention that the soft-core potential is automatically switched on in the case of a single-component lambda. #1421
Clarified documentation for time-dependent electric fields.
Corrected force-switch manual section There was a factor alpha missing in the formulas for parameter A and B for the force-switch function (the code was correct). Also removed the introduction and formulas for switching the Coulomb function, since that's a bad thing to do.
Fixed default optimization level for xlc on BG/Q.
作者Author: wangxinyu 时间: 2015-6-11 15:22
"Changed configure-time FFTW warning against --enable-avx to one against using --enable-avx without --enable-sse (for AVX builds). With FFTW 3.3.4, the performance of FFTW with both SSE and AVX enabled is often a bit better and never much worse than SSE alone. You may get better performance on all kinds of AVX hardware with ftp://ftp.gromacs.org/pub/beta/fftw-...tersimd.tar.gz. These improvements may appear in future releases of FFTW."这句里面
"Changed configure-time FFTW warning against --enable-avx to one against using --enable-avx without --enable-sse (for AVX builds)."是什么意思啊？
谢谢！作者Author: sobereva 时间: 2015-6-11 15:32
wangxinyu 发表于 2015-6-11 15:22
"Changed configure-time FFTW warning against --enable-avx to one against using --enable-avx without ...