Install the platform » Historique » Révision 61
« Précédent |
Révision 61/62
(diff)
| Suivant »
Arnaud Sevin, 15/02/2017 11:21
Install Anaconda with python2¶
more info: https://www.continuum.io/downloads#linux
Download and installation¶
- wget https://repo.continuum.io/archive/Anaconda2-4.2.0-Linux-x86_64.sh
- bash ./Anaconda2-4.2.0-Linux-x86_64.sh
- update your .batchrc
add anaconda2/bin into the $PATH
add more packets¶
To avoid any incompatibility this python modules, it's highly recommended to use the gcc provided with anaconda:
- conda install -c compass compass
Install MAGMA¶
Why MAGMA ?¶
The MAGMA project aims to develop a dense linear algebra library similar to LAPACK but for heterogeneous/hybrid architectures, starting with current "Multicore+GPU" systems.
Unlike CULA, MAGMA propose a dense linear algebra library handling double for free.
But MAGMA needs a LAPACK and a BLAS implementation. Actually, we try two options : openBLAS (free, easy to install) and MKL (free, need a registration but better optimized on Intel processors)
Configure MAGMA with openBLAS¶
Dependencies : openblas (http://www.openblas.net)¶
First, clone the GIT repository:
git clone https://github.com/xianyi/OpenBLAS.git
compile it:
cd OpenBLAS/ make
install it:
make install PREFIX=$HOME/local/openblas
add to you .bashrc:
export OPENBLAS_ROOT=$HOME/local/openblas
extraction¶
MAGMA is available here : http://icl.cs.utk.edu/magma/software/index.html
extract the tgz file and go into the new directory
$ wget http://icl.cs.utk.edu/projectsfiles/magma/downloads/magma-2.2.0.tar.gz
$ tar xf magma-2.2.0.tar.gz
~$ cd magma-2.2.0
configuration¶
You have to create your own make.inc based on make.inc.openblas:
~$ cp make.inc-examples/make.inc.openblas make.inc
example : please verify GPU_TARGET, OPENBLASDIR, CUDADIR
#//////////////////////////////////////////////////////////////////////////////
# -- MAGMA (version 2.2.0) --
# Univ. of Tennessee, Knoxville
# Univ. of California, Berkeley
# Univ. of Colorado, Denver
# @date November 2016
#//////////////////////////////////////////////////////////////////////////////
# GPU_TARGET contains one or more of Fermi, Kepler, or Maxwell,
# to specify for which GPUs you want to compile MAGMA:
# Fermi - NVIDIA compute capability 2.x cards
# Kepler - NVIDIA compute capability 3.x cards
# Maxwell - NVIDIA compute capability 5.x cards
# Pascal - NVIDIA compute capability 6.x cards
# The default is "Fermi Kepler".
# Note that NVIDIA no longer supports 1.x cards, as of CUDA 6.5.
# See http://developer.nvidia.com/cuda-gpus
#
GPU_TARGET ?= Pascal
# --------------------
# programs
CC = gcc
CXX = g++
NVCC = nvcc
FORT = gfortran
ARCH = ar
ARCHFLAGS = cr
RANLIB = ranlib
# --------------------
# flags
# Use -fPIC to make shared (.so) and static (.a) library;
# can be commented out if making only static library.
FPIC = -fPIC
CFLAGS = -O3 $(FPIC) -DNDEBUG -DADD_ -Wall -fopenmp
FFLAGS = -O3 $(FPIC) -DNDEBUG -DADD_ -Wall -Wno-unused-dummy-argument
F90FLAGS = -O3 $(FPIC) -DNDEBUG -DADD_ -Wall -Wno-unused-dummy-argument -x f95-cpp-input
NVCCFLAGS = -O3 -DNDEBUG -DADD_ -Xcompiler "$(FPIC)"
LDFLAGS = $(FPIC) -fopenmp
# C++11 (gcc >= 4.7) is not required, but has benefits like atomic operations
CXXFLAGS := $(CFLAGS) -std=c++11
CFLAGS += -std=c99
# --------------------
# libraries
# gcc with OpenBLAS (includes LAPACK)
LIB = -lopenblas
LIB += -lcublas -lcusparse -lcudart -lcudadevrt
# --------------------
# directories
# define library directories preferably in your environment, or here.
OPENBLASDIR ?= $(HOME)/local/openblas
CUDADIR ?= /usr/local/cuda
-include make.check-openblas
-include make.check-cuda
LIBDIR = -L$(CUDADIR)/lib64 \
-L$(OPENBLASDIR)/lib
INC = -I$(CUDADIR)/include \
-I$(OPENBLASDIR)/include
Configure MAGMA with MKL¶
extraction¶
To download MKL, you have to create a account here : https://registrationcenter.intel.com/RegCenter/NComForm.aspx?ProductID=1517
extract l_ccompxe_2013_sp1.1.106.tgz and go into l_ccompxe_2013_sp1.1.106
install it with ./install_GUI.sh and add IPP stuff to default choices
configuration¶
You have to create your own make.inc based on make.inc.mkl-gcc-ilp64:
example: please verify GPU_TARGET, MKLROOT, CUDADIR
#//////////////////////////////////////////////////////////////////////////////
# -- MAGMA (version 2.1.0) --
# Univ. of Tennessee, Knoxville
# Univ. of California, Berkeley
# Univ. of Colorado, Denver
# @date August 2016
#//////////////////////////////////////////////////////////////////////////////
# GPU_TARGET contains one or more of Fermi, Kepler, or Maxwell,
# to specify for which GPUs you want to compile MAGMA:
# Fermi - NVIDIA compute capability 2.x cards
# Kepler - NVIDIA compute capability 3.x cards
# Maxwell - NVIDIA compute capability 5.x cards
# Pascal - NVIDIA compute capability 6.x cards
# The default is "Fermi Kepler".
# Note that NVIDIA no longer supports 1.x cards, as of CUDA 6.5.
# See http://developer.nvidia.com/cuda-gpus
#
#GPU_TARGET ?= Fermi Kepler
# --------------------
# programs
CC = icc
CXX = icpc
NVCC = nvcc
FORT = ifort
ARCH = ar
ARCHFLAGS = cr
RANLIB = ranlib
# --------------------
# flags
# Use -fPIC to make shared (.so) and static (.a) library;
# can be commented out if making only static library.
FPIC = -fPIC
CFLAGS = -O3 $(FPIC) -openmp -DADD_ -Wall -Wshadow -DMAGMA_WITH_MKL
FFLAGS = -O3 $(FPIC) -DADD_ -warn all -warn nounused -nogen-interfaces
F90FLAGS = -O3 $(FPIC) -DADD_ -warn all -warn nounused
NVCCFLAGS = -O3 -DADD_ -Xcompiler "$(FPIC) -Wall -Wno-unused-function"
LDFLAGS = $(FPIC) -openmp
# Defining MAGMA_ILP64 or MKL_ILP64 changes magma_int_t to int64_t in include/magma_types.h
CFLAGS += -DMKL_ILP64
FFLAGS += -integer-size 64
F90FLAGS += -integer-size 64
NVCCFLAGS += -DMKL_ILP64
# Options to do extra checks for non-standard things like variable length arrays;
# it is safe to disable all these
CFLAGS += -pedantic -Wno-long-long
#CFLAGS += -Werror # uncomment to ensure all warnings are dealt with
# C++11 (icc >= 13) is not required, but has benefits like atomic operations
CXXFLAGS := $(CFLAGS) -std=c++11
CFLAGS += -std=c99
# --------------------
# libraries
# IMPORTANT: these link lines are for 64-bit int !!!!
# For regular 64-bit builds using 64-bit pointers and 32-bit int,
# use the lp64 library, not the ilp64 library. See make.inc.mkl-gcc or make.inc.mkl-icc.
# see MKL Link Advisor at http://software.intel.com/sites/products/mkl/
# icc with MKL 10.3, Intel OpenMP threads, 64-bit int
# note -DMAGMA_ILP64 or -DMKL_ILP64, and -integer-size 64 in FLAGS above
LIB = -lmkl_intel_ilp64 -lmkl_intel_thread -lmkl_core -lpthread -lstdc++ -lm
LIB += -lcublas -lcusparse -lcudart
# --------------------
# directories
# define library directories preferably in your environment, or here.
# for MKL run, e.g.: source /opt/intel/composerxe/mkl/bin/mklvars.sh intel64
#MKLROOT ?= /opt/intel/composerxe/mkl
#CUDADIR ?= /usr/local/cuda
-include make.check-mkl
-include make.check-cuda
LIBDIR = -L$(CUDADIR)/lib64 \
-L$(MKLROOT)/lib/intel64
INC = -I$(CUDADIR)/include \
-I$(MKLROOT)/include
In this example, I use gcc but with MKL, you can use icc instead of gcc. In this case, you have to compile yorick with icc. For this, you have to change the CC flag in Make.cfg
compilation and installation¶
compilation¶
just compile the shared target (and test if you want)
~$ make -j 8 shared sparse-shared
installation¶
To install libraries and include files in a given prefix, run:
~$ make install prefix=$HOME/local/magma
The default prefix is /usr/local/magma. You can also set prefix in make.inc.
tuning (not tested)¶
For multi-GPU functions, set $MAGMA_NUM_GPUS to set the number of GPUs to use.
For multi-core BLAS libraries, set $OMP_NUM_THREADS or $MKL_NUM_THREADS or $VECLIB_MAXIMUM_THREADS to set the number of CPU threads, depending on your BLAS library.
Install the platform¶
The COMPASS platform is distributed as a single bundle of CArMA and SuTrA C++ / Cuda libraries and their Python extensions NAGA & SHESHA.
Hardware requirements¶
The system must contain at least an x86 CPU and a CUDA capable GPU. list of compatible GPUs can be found here http://www.nvidia.com/object/cuda_gpus.html. Specific requirements apply to clusters (to be updated).
Environment requirements¶
The system must be running a 64 bit distribution of Linux with the latest NVIDIA drivers and CUDA toolkit. The following installation instructions are valid if the default installation paths have been selected for these components.
Additionally, to benefit from the user-oriented features of the platform, Anaconda2 should be installed (https://www.continuum.io/downloads#_unix).
In the last versions of compass (r608+), Yorick is no more supported.
Installation process¶
First check out the latest version from the svn repository :
svn co https://version-lesia.obspm.fr/repos/compass/trunk compass
then go in the newly created directory and then trunk:
cd compass
once there, you need to modify system variables in our .bashrc :
# CUDA default definitions
export CUDA_ROOT=$CUDA_ROOT #/usr/local/cuda
export CUDA_INC_PATH=$CUDA_ROOT/include
export CUDA_LIB_PATH=$CUDA_ROOT/lib
export CUDA_LIB_PATH_64=$CUDA_ROOT/lib64
export PATH=$CUDA_ROOT/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_LIB_PATH_64:$CUDA_LIB_PATH:$LD_LIBRARY_PATH
in this file, you also have to indicate the proper architecture of your GPU so as the compiler will generate the appropriate code.
export GENCODE="arch=compute_52,code=sm_52"
and change both 52 to your architecture : for instance a Tesla Fermi will have 2.0 computing capabilities so change 52 to 20, a Kepler GPU will have 3.0 or 3.5 (K20) computing capabilities, change 52 to 30 (or 35), a Maxwell GPU have 5.2 (M6000), a Pascal have 6.0 (P100).
(more informations here: https://developer.nvidia.com/cuda-gpus)
If you are using CULA, you have to specify it:
# CULA default definitions
export CULA_ROOT= /usr/local/cula
export CULA_INC_PATH= $CULA_ROOT/include
export CULA_LIB_PATH= $CULA_ROOT/lib
export CULA_LIB_PATH_64= $CULA_ROOT/lib64
export LD_LIBRARY_PATH=$CULA_LIB_PATH_64:$CULA_LIB_PATH:$LD_LIBRARY_PATH
If you are using MAGMA, you have to specify it:
# MAGMA definitions (uncomment this line if MAGMA is installed)
export MAGMA_ROOT=$HOME/local/magma
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$MAGMA_ROOT/lib
export PKG_CONFIG_PATH=$MAGMA_ROOT/lib/pkgconfig
Last variables to define:
export COMPASS_ROOT=/path/to/compass/trunk
export NAGA_ROOT=$COMPASS_ROOT/naga
export SHESHA_ROOT=$COMPASS_ROOT/shesha
export LD_LIBRARY_PATH=$COMPASS_ROOT/libcarma:$COMPASS_ROOT/libsutra:$LD_LIBRARY_PATH
At the end, you .bashrc shoud containts all those informations:
# conda default definitions
export CONDA_ROOT=/your/path/anaconda2
export PATH=$CONDA_ROOT/bin:$PATH
# CUDA default definitions
export CUDA_INC_PATH=$CUDA_ROOT/include
export CUDA_LIB_PATH=$CUDA_ROOT/lib
export CUDA_LIB_PATH_64=$CUDA_ROOT/lib64
export PATH=$CUDA_ROOT/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_LIB_PATH_64:$CUDA_LIB_PATH:$LD_LIBRARY_PATH
export GENCODE="arch=compute_52,code=sm_52"
# OPENBLAS definitions
export OPENBLAS_ROOT=$HOME/local/openblas
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$OPENBLAS_ROOT/lib
export PKG_CONFIG_PATH=$PKG_CONFIG_PATH:$OPENBLAS_ROOT/lib/pkgconfig
# MAGMA definitions
export MAGMA_ROOT=$HOME/local/magma
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$MAGMA_ROOT/lib
export PKG_CONFIG_PATH=$PKG_CONFIG_PATH:$MAGMA_ROOT/lib/pkgconfig
# COMPASS default definitions
export COMPASS_ROOT=/your/path/compass
export NAGA_ROOT=$COMPASS_ROOT/naga
export SHESHA_ROOT=$COMPASS_ROOT/shesha
export LD_LIBRARY_PATH=$COMPASS_ROOT/libcarma:$COMPASS_ROOT/libsutra:$LD_LIBRARY_PATH
Once this is done, you're ready to compile the whole library:
make clean all
If you did not get any error, CArMA, SuTrA, NAGA and SHESHA are now installed on your machine. You can check that everything is working by launching a GUI to test a simulation:
ipython -i $SHESHA_ROOT/widgets/widget_ao.py
Mis à jour par Arnaud Sevin il y a presque 8 ans · 61 révisions