-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathINSTALL
134 lines (96 loc) · 4.5 KB
/
INSTALL
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
gputools
=========
contents
=========
+ notes
+ machines w/ older gpus
+ machines w/o a gpu at all
+ your cuda install is not in /usr/local/cuda
+ mac os x systems
+ windows systems
+ cula
========
+ notes
========
The usual installation procedure should go something like this:
1. procure a machine with a cuda capable nvidia gpu
2. install nvidia's cuda capable video driver
3. install nvidia's cuda toolkit
4. install this package using R
The package gputools attempts to follow the standard R package installation
procedure. To install the package, you must first install
nvidia's cuda toolkit (preferably version 2.3) available from
http://www.nvidia.com/object/cuda_get.html
You need not use a gpu, to use the package without a cuda capable gpu and
video driver, please skip to the third section 'machines w/o a gpu at all'.
To use a gpu, you will need a cuda capable gpu and a cuda capable video driver.
You can get the video driver from
http://www.nvidia.com/object/cuda_get.html
also.
For exotic R installs, the following configure arguments may help:
--with-r-lib=<location of R shared libraries>
--with-r-include=<location of R header files>
As far as compatibility goes, Windows is not supported. Neither is OS X 10.6.
Rumor has it that the source package does work with the 32bit build of R under
OS X 10.6. The source package does not support Mac dynamic libraries, so you
may have trouble with certain Mac versions of CULA.
In case you have a Linux distribution with default gcc compiler version > 4.4,
install gcc 4.4 elsewhere and us the configure argument
--with-compiler-bindir=<directory containing gcc 4.4>
or set the environment variable COMPILER_BINDIR.
=========================
+ machines w/ older gpus
=========================
See
http://brainarray.mbni.med.umich.edu/Brainarray/Rgpgpu/#issues
for a table listing compute capabilities of various nVidia cards.
If you have a GPU that has compute capability < 1.3 then
the 'kendall' method of 'gpuCor' and
the functions 'gpuSvmTrain' and 'gpuSvmPredict'
will not function correctly! All other methods should be fine with such a card.
============================
+ machines w/o a gpu at all
============================
If you do not have a CUDA capable GPU or just don't wish to use one read
this section. You need not install a cuda capable video driver. You can
still use the package using one of the following commands.
From a terminal:
R CMD INSTALL --configure-args='--enable-emulation' <path-to-package>gputools_<version-of-package>.tar.gz
From inside an R environment:
install.packages(gputools, configure.args = '--enable-emulation')
Please note that you must still install the CUDA toolkit.
Device emulation may not work with CUDA toolkit version 2.2.
Version 2.3 is strongly recommended for gpu emulation.
The support vector machine functions are disabled in device emulation mode.
They make use of device driver api calls which are not available unless you
are able to install a CUDA capable driver. To install a CUDA capable drive,
you usually need a CUDA capable device.
==============================================
+ your cuda install is not in /usr/local/cuda
==============================================
If the CUDA toolkit is not in '/user/local/cuda' then you can specify the
location of the CUDA toolkit by either setting the environment variable
'CUDA_HOME' or using one of the following commands.
From a terminal:
R CMD INSTALL --configure-args='--with-cuda-home=<path-to-cuda>' <path-to-package>gputools_<version-of-package>.tar.gz
From inside an R environment:
install.packages(gputools, configure.args = '--with-cuda-home=/usr/local/cuda')
====================
+ mac os x systems
====================
All of the above should work on mac os x systems and even in the R gui without
modification. If something doesn't work as expected on the mac please email
the maintainer.
==================
+ windows systems
==================
There is no support for Windows yet. If you would like to write a windows port,
please feel free. You are welcome to email it to the maintainer for merging
with the main package, but only if you want to.
========
+ cula
========
Since version 0.1-4, gputools supports the use of the CULA library from
http://www.culatools.com
If you install CULA to a directory other than '/usr/local/cula' you'll want to adjust the CULA_HOME variable similar to CUDA_HOME mentioned above.
Also, on linux machines, please remove the cuda library directory from your LD_LIBRARY_PATH environment variable. Make sure it has the cula library directory instead!