ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #1226 from akohlmey/cmake-gpu-enhancements

Browse Source 
Enhancements for using CMake with the GPU package, improved compatibility with cmake 3.x versions, improved handling of shared library building.
This commit is contained in:
Axel Kohlmeyer
2018-11-27 16:05:47 -05:00
committed by GitHub
parent 6e8c537564 ebacd5ca6b
commit f254b8e3a3
5 changed files with 120 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
doc/src/Build_cmake.txt
View File

@ -44,7 +44,7 @@ LAMMPS or need to re-compile LAMMPS repeatedly, installation of the
ccache (= Compiler Cache) software may speed up compilation even more.
After compilation, you can optionally copy the LAMMPS executable and
library into your system folders (by default under /usr/local) with:
library into your system folders (by default under $HOME/.local) with:
make install # optional, copy LAMMPS executable & library elsewhere :pre

32
doc/src/Build_extras.txt
View File

@ -87,22 +87,30 @@ which GPU hardware to build for.
# value = double or mixed (default) or single
-D OCL_TUNE=value # hardware choice for GPU_API=opencl
# generic (default) or intel (Intel CPU) or fermi, kepler, cypress (NVIDIA)
-D GPU_ARCH=value # hardware choice for GPU_API=cuda
-D GPU_ARCH=value # primary GPU hardware choice for GPU_API=cuda
# value = sm_XX, see below
# default is Cuda-compiler dependent, but typically sm_20
-D CUDPP_OPT=value # optimization setting for GPU_API=cudea
-D CUDPP_OPT=value # optimization setting for GPU_API=cuda
# enables CUDA Performance Primitives Optimizations
# yes (default) or no :pre
GPU_ARCH settings for different GPU hardware is as follows:
sm_20 for Fermi (C2050/C2070, deprecated as of CUDA 8.0) or GeForce GTX 580 or similar
sm_30 for Kepler (K10)
sm_35 for Kepler (K40) or GeForce GTX Titan or similar
sm_37 for Kepler (dual K80)
sm_50 for Maxwell
sm_60 for Pascal (P100)
sm_70 for Volta :ul
sm_20 or sm_21 for Fermi (supported by CUDA 3.2 until CUDA 7.5)
sm_30 or sm_35 or sm_37 for Kepler (supported since CUDA 5)
sm_50 or sm_52 for Maxwell (supported since CUDA 6)
sm_60 or sm_61 for Pascal (supported since CUDA 8)
sm_70 for Volta (supported since CUDA 9)
sm_75 for Turing (supported since CUDA 10) :ul
A more detailed list can be found, for example,
at "Wikipedia's CUDA article"_https://en.wikipedia.org/wiki/CUDA#GPUs_supported
CMake can detect which version of the CUDA toolkit is used and thus can
include support for [all] major GPU architectures supported by this toolkit.
Thus the GPU_ARCH setting is merely an optimization, to have code for
the preferred GPU architecture directly included rather than having to wait
for the JIT compiler of the CUDA driver to translate it.
[Traditional make]:
@ -137,6 +145,11 @@ CUDA_ARCH = sm_XX, what GPU hardware you have, same as CMake GPU_ARCH above
CUDA_PRECISION = precision (double, mixed, single)
EXTRAMAKE = which Makefile.lammps.* file to copy to Makefile.lammps :ul
The file Makefile.linux_multi is set up to include support for multiple
GPU architectures as supported by the CUDA toolkit in use. This is done
through using the "--gencode " flag, which can be used multiple times and
thus support all GPU architectures supported by your CUDA compiler.
If the library build is successful, 3 files should be created:
lib/gpu/libgpu.a, lib/gpu/nvc_get_devices, and
lib/gpu/Makefile.lammps. The latter has settings that enable LAMMPS
@ -150,6 +163,7 @@ re-build LAMMPS. This is because the compilation of files in the GPU
package uses the library settings from the lib/gpu/Makefile.machine
used to build the GPU library.
:line
KIM package :h4,link(kim)