Account Options

  1. Sign in
The old Google Groups will be going away soon, but your browser is incompatible with the new version.
Google Groups Home
« Groups Home
theano-users
Home
+ new post
About this group
Join this group
Theano + GPU on Windows
Options
There are currently too many topics in this group that display first. To make this topic appear first, remove this option from another topic.
There was an error processing your request. Please try again.
flag
   Messages 1 - 25 of 38 - Collapse all  -  Translate all to Translated (View all originals)   Newer >
The group you are posting to is a Usenet group. Messages posted to this group will make your email address visible to anyone on the Internet.
Your reply message has not been sent.
Your post was successful
 
From:
To:
Cc:
Followup To:
Add Cc | Add Followup-to | Edit Subject
Subject:
Validation:
For verification purposes please type the characters you see in the picture below or the numbers you hear by clicking the accessibility icon. Listen and type the numbers you hear
 
Mark  
View profile  
 More options Jul 20 2012, 2:17 pm
From: Mark <mark.fen...@coveros.com>
Date: Fri, 20 Jul 2012 11:17:37 -0700 (PDT)
Local: Fri, Jul 20 2012 2:17 pm
Subject: Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Hi folks,

I'm attempting to isolate some problems with a Theano/Windows/GPU setup
(window 7/visual studio 2010/nvidia 4.2 cuda downloads/theano latest git) .
 I isolated the theano nvcc call and separated that out.  Standalone, from
a VS2010 command prompt, the nvcc command gives the following output:

nvcc -shared -g
     -IC:\Users\mfenner\Theano\theano\sandbox\cuda
     -IC:\Python27\lib\site-packages\numpy\core\include
     -IC:\Python27\include
     -o out.cuda_ndarray.pyd
     ./cuda_ndarray.cu
     -LC:\Python27\libs
     -L"C:\Program Files\NVDIIA GPU Computing Toolkit\CUDA\v4.2\lib\x64"
     -lpython27
     -lcublas
     -lcudart

which produces the following error:
./cuda_ndarray.cu(480): warning: conversion from pointer to smaller integer

./cuda_ndarray.cu(769): warning: statement is unreachable

./cuda_ndarray.cu(850): error: expression must have a constant value

./cuda_ndarray.cu(865): error: no instance of function template
"CudaNdarray_alloc_contiguous" matches the argument list
            argument types are: (CudaNdarray *, int, <error-type>)

./cuda_ndarray.cu(2258): warning: conversion from pointer to smaller integer

2 errors detected in the compilation of
"C:/Users/mfenner/AppData/Local/Temp/tmp
xft_0000060c_00000000-6_cuda_ndarray.cpp1.ii".

Any thoughts?

Best,
Mark


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Frédéric Bastien  
View profile  
 More options Jul 20 2012, 5:35 pm
From: Frédéric Bastien <no...@nouiz.org>
Date: Fri, 20 Jul 2012 17:35:33 -0400
Local: Fri, Jul 20 2012 5:35 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author
witch verison of theano do you use? Also, can you join your version of
cuda_ndarray.cu file? That way, I can know what the line number mean.

This line you give:

>             argument types are: (CudaNdarray *, int, <error-type>)

the last parameter don't seam good. But this is working on linux. So
it is probably something that is different on linux and windows.

thanks for the report.

Fred


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Mark  
View profile  
 More options Jul 20 2012, 5:54 pm
From: Mark <mark.fen...@coveros.com>
Date: Fri, 20 Jul 2012 14:54:30 -0700 (PDT)
Local: Fri, Jul 20 2012 5:54 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Fred,

Hi.  Thanks for your interest in my problem.

I'm working with the development version of Theano.  The specific git
commit id is:

% git log --oneline | more
b672a3a Merge pull request #758 from goodfeli/doc

The follow should get the version of cuda_ndarray.cu that I'm working
against (from a working directory of Theano/theano/sandbox/cuda):
% git show b672a3a:./cuda_ndarray.cu

Let me know if I can provide anything else.

Best,
Mark


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Frédéric Bastien  
View profile  
 More options Jul 20 2012, 7:14 pm
From: Frédéric Bastien <no...@nouiz.org>
Date: Fri, 20 Jul 2012 19:14:06 -0400
Local: Fri, Jul 20 2012 7:14 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author
I found the reason,

This microsoft compiler don't support variable len array in there c++
compilier. So this line is not accepted:

I made a new PR and David just reviewed it. So if you update your
version of Theano, it should work.

thanks for the report. If there is other problem(i hope not), don't
hesitate to tell us.

Fred


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Frédéric Bastien  
View profile  
 More options Jul 21 2012, 11:00 am
From: Frédéric Bastien <no...@nouiz.org>
Date: Sat, 21 Jul 2012 11:00:21 -0400
Local: Sat, Jul 21 2012 11:00 am
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author
My first merge fix introduced a bug. Now it is fixed. So you should reupdate.

Fred


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Mark  
View profile  
 More options Jul 23 2012, 10:43 am
From: Mark <mark.fen...@coveros.com>
Date: Mon, 23 Jul 2012 07:43:26 -0700 (PDT)
Local: Mon, Jul 23 2012 10:43 am
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Fred,

Thank you very much for working on this.  I will update my repository and
try it again sometime today.

Best,
Mark


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Mark  
View profile  
 More options Jul 23 2012, 11:41 am
From: Mark <mark.fen...@coveros.com>
Date: Mon, 23 Jul 2012 08:41:54 -0700 (PDT)
Local: Mon, Jul 23 2012 11:41 am
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

The patch seems to have fixed the nvcc compilation error.  Thanks again!  

Now, I'm getting a large number of link errors.  I will try to debug and
isolate what's going on there.

Best,
Mark


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Olivier Delalleau  
View profile  
 More options Jul 23 2012, 12:36 pm
From: Olivier Delalleau <delal...@iro.umontreal.ca>
Date: Mon, 23 Jul 2012 12:36:58 -0400
Local: Mon, Jul 23 2012 12:36 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Most likely won't help, but just in case, run "python theano-cache clear"
(theano-cache can be found in the bin subfolder). Sometimes outdated
compiled code can lead to weird linking errors.

-=- Olivier

2012/7/23 Mark <mark.fen...@coveros.com>


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Mark  
View profile  
 More options Jul 24 2012, 10:53 am
From: Mark <mark.fen...@coveros.com>
Date: Tue, 24 Jul 2012 07:53:07 -0700 (PDT)
Local: Tues, Jul 24 2012 10:53 am
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Olivier,

Thanks for that hint.  I wouldn't have tried that otherwise:  however, it
didn't help.  

Thanks to a stackoverflow thread:  
http://stackoverflow.com/questions/5150866/vs2008-not-compiling-cuda
I did have some success.  From a VS2010 command line, the following command
(^ is line continuation in Windows cmd prompt land):

nvcc -shared -g ^
     -IC:\Users\mfenner\Theano\theano\sandbox\cuda ^
     -IC:\Python27\lib\site-packages\numpy\core\include ^
     -IC:\Python27\include ^
     -o out.cuda_ndarray.pyd ^
     ./cuda_ndarray.cu ^
     -LC:\Python27\libs ^
     -L"C:\Program Files\NVDIIA GPU Computing Toolkit\CUDA\v4.2\lib\x64" ^
     -lpython27 ^
     -lcublas ^
     -lcudart ^
     --machine 32

Notice the important addition of "--machine 32" at the tail end.  This
resolved a large number (10s?  100s?) of  LNK2019, unresolved external
symbol errors that look like this:

tmpxft_000005a0_00000000-15_cuda_ndarray.obj : error LNK2019: unresolved
external symbol __imp_PyErr_Format referenced in function "void * __cdecl
device_malloc (unsigned __int64)" (?device_malloc@@YAPEAX_K@Z)

I'm personally a little confused as to what "bit-ness" my toolchain is -- I
initially thought I was aiming for an "all 64-bit" toolchain (from python
to python libs to nvidia components to compiled/linked nvcc code).
 However, it appears I was mistaken and/or I decided/realized I needed to
move to 32-bit.  I'm going to add --machine 32 to my .theanorc and see how
the internal theano/nvcc compilation goes.

Also:  I saw a comment that talked about unrecognized "<" and the need to
register ".cu" files (I think) in Visual Studio.  It's the last comment
(currently) on this SO thread.  Incidentally, that thread deals with some
of the issues in using CUDA 4 with VS2010.
http://stackoverflow.com/questions/3778799/how-do-i-start-a-cuda-app-...

Best,
Mark


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Olivier Delalleau  
View profile  
 More options Jul 24 2012, 6:39 pm
From: Olivier Delalleau <olivier.delall...@gmail.com>
Date: Tue, 24 Jul 2012 18:39:24 -0400
Local: Tues, Jul 24 2012 6:39 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Thanks for the info. Please let us know if adding "--machine 32" is enough to solve your problem.
It would also be interesting to know which part of your toolchain is 32 bit:
- Python?
- Cuda?
- Visual Studio?

-=- Olivier

Le 2012-07-24 à 10:53, Mark <mark.fen...@coveros.com> a écrit :


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Gabe Schwartz  
View profile   Translate to Translated (View Original)
 More options Jul 25 2012, 10:03 am
From: Gabe Schwartz <hapem...@gmail.com>
Date: Wed, 25 Jul 2012 07:03:29 -0700 (PDT)
Local: Wed, Jul 25 2012 10:03 am
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

This might be too obvious and you may be already doing it, but were you
using the x64/Win64 command prompt?

I remember having a number of issues compiling other software because I was
using the "Visual Studio Command Prompt (2010)" instead of the "Visual
Studio x64 Win64 Command Prompt (2010)". If you don't use the latter then
all sorts of wrong libraries will be used.


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Mark  
View profile  
 More options Jul 25 2012, 10:13 am
From: Mark <mark.fen...@coveros.com>
Date: Wed, 25 Jul 2012 07:13:23 -0700 (PDT)
Local: Wed, Jul 25 2012 10:13 am
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Ok, so, here's the latest.  I can't figure out how to pass the "--machine
32" parameter through .theanorc to the internal nvcc command line.  I've
tried several permutations, but the parameter isn't being sent to the
command in a way that makes the command happy.  Here's my latest attempt:

from .theanorc:
[nvcc]
compiler_bindir=C:\Program Files (x86)\Microsoft Visual Studio 10.0\VC\bin
flags=--machine 32

I've also tried:
flags="--machine 32" and flags=--"machine 32" and some others.  How is the
right-hand side of the assignment interpreted and how is that passed along
to (I'm assuming) -- eventually -- a subprocess.Popen?

*********************************

As far as versions go, here's what I found (and the evidence as to why I
think it):

Python (from mingw32/msys):
$ which python
/c/python27/python.exe
$ python

>>> import platform; platform.architecture()

('32bit', 'WindowsPE')

CUDA:
NVIDIA 64-bit driver/toolkit/sdk

PyCUDA (working):
amd64

VS2010:  
the compilations I've done are with the Visual Studio Command Prompt (2010)
[non-x64 Win64]


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Mark  
View profile  
 More options Jul 25 2012, 10:24 am
From: Mark <mark.fen...@coveros.com>
Date: Wed, 25 Jul 2012 07:24:36 -0700 (PDT)
Local: Wed, Jul 25 2012 10:24 am
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Gabe,

I've had success with command line nvcc from VS2010's non-x64 command
prompt (I just posted some other details of my environment).

I'm curious, how do you get around the missing g++ problem when you are
using a VS prompt?  I end up using mingw/msys as my command prompt so I can
access g++.

Thanks,
Mark

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Olivier Delalleau  
View profile  
 More options Jul 25 2012, 1:23 pm
From: Olivier Delalleau <delal...@iro.umontreal.ca>
Date: Wed, 25 Jul 2012 13:23:58 -0400
Local: Wed, Jul 25 2012 1:23 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Looks like you found a bug in the parsing of nvcc flags ;)
Use instead:
flags=-m32
This should work for now (I made an issue for this bug:
https://github.com/Theano/Theano/issues/785)

I'm not sure using 64 bit Cuda with 32 bit Python works, but you can give
it a shot...

-=- Olivier

2012/7/25 Mark <mark.fen...@coveros.com>


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Olivier Delalleau  
View profile  
 More options Jul 25 2012, 1:25 pm
From: Olivier Delalleau <delal...@iro.umontreal.ca>
Date: Wed, 25 Jul 2012 13:25:34 -0400
Local: Wed, Jul 25 2012 1:25 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

It's a good point though, if you wanted a full 64 bit toolchain you'd
probably want to use the x64 Visual Studio command prompt instead.

For g++, try adding to your Windows PATH environment variable the folder
that contains g++.exe. This may be enough.

-=- Olivier

2012/7/25 Mark <mark.fen...@coveros.com>

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Mark  
View profile  
 More options Jul 25 2012, 3:07 pm
From: Mark <mark.fen...@coveros.com>
Date: Wed, 25 Jul 2012 12:07:02 -0700 (PDT)
Local: Wed, Jul 25 2012 3:07 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Regarding 64-bit CUDA + 32-bit python ... I'm on my linux workstation, but
I -think- (please double check if anything important depends on this) that
the 64-bit CUDA libs come with BOTH 32 and 64 bit headers and libs.  Thus,
you can make 32-bit python happy by pointing the build
scripts/configurations at the correct directory, gets you correct behavior.
 Again, that's from sketchy memory.

Alright, so, I reviewed a few notes and part of my preceeding comments
apply directly to installing and testing PyCuda (not theono, directly).
 However, the following seems relevent:

(1)  the python-amd64 that I downloaded and installed actually reports the
platform information that I mentioned earlier (indicating 32-bit).  I don't
understand that and I may have made an error in my installation notes
(however, my download directory only has the python-amd64 installer).
 Perhaps it is an architecture optimized build, but not a 64-bit build.  I
could definitely be wrong; just a guess.

(2)  to make PyCuda happy, I placed these lines in my
PYCUDA_SOURCE/siteconf.py (pycuda's configuration) [the
$CUDA_ROOT/lib/Win32 could have been $CUDA_ROOT/lib/x64 to select 64-bit
libs instead]:
BOOST_INC_DIR = []
BOOST_LIB_DIR = []
BOOST_COMPILER = 'gcc43'
USE_SHIPPED_BOOST = True
BOOST_PYTHON_LIBNAME = ['boost_python']
BOOST_THREAD_LIBNAME = ['boost_thread']
CUDA_TRACE = False
CUDA_ROOT = 'C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v4.2'
CUDA_ENABLE_GL = False
CUDA_ENABLE_CURAND = True
CUDADRV_LIB_DIR = ['${CUDA_ROOT}/lib/Win32']
CUDADRV_LIBNAME = ['cuda']
CUDART_LIB_DIR = ['${CUDA_ROOT}/lib/Win32']
CUDART_LIBNAME = ['cudart']
CURAND_LIB_DIR = ['${CUDA_ROOT}/lib/Win32']
CURAND_LIBNAME = ['curand']
CXXFLAGS = ['/EHsc']
LDFLAGS = ['/FORCE']

Best,
Mark

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Mark  
View profile  
 More options Jul 25 2012, 3:08 pm
From: Mark <mark.fen...@coveros.com>
Date: Wed, 25 Jul 2012 12:08:34 -0700 (PDT)
Local: Wed, Jul 25 2012 3:08 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Olivier,

Thanks.  I was hesitant to try with PATH at the VS2010 prompt b/c I figured
everything else (e.g., the toolchain and linking environments) would fail
miserably.  I'll give that a try also.

Best,
Mark

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Mark  
View profile  
 More options Jul 25 2012, 3:26 pm
From: Mark <mark.fen...@coveros.com>
Date: Wed, 25 Jul 2012 12:26:34 -0700 (PDT)
Local: Wed, Jul 25 2012 3:26 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Olivier,

This seems to fail for me [tail of python a-theano-script.py]:

ERROR (theano.sandbox.cuda): Failed to compile cuda_ndarray.cu: ('nvcc
return st
atus', 2, 'for cmd', 'nvcc -shared -g --compiler-bindir C:\\Program Files
(x86)\
\Microsoft Visual Studio 10.0\\VC\\bin -Xlinker /DEBUG -Xcompiler
-m32,-DCUDA_ND
ARRAY_CUH=d9e2442af5da4979f20fbd43efb84d5f,/Zi
-Ic:\\Users\\mfenner\\theano\\the
ano\\sandbox\\cuda -Ic:\\python27\\lib\\site-packages\\numpy\\core\\include
-Ic:
\\python27\\include -o
C:\\Users\\mfenner\\theano-compile\\compiledir_Windows-7-
6.1.7601-SP1-Intel64_Family_6_Model_37_Stepping_5_GenuineIntel-2.7.2\\cuda_ ndarr
ay\\cuda_ndarray.pyd mod.cu -Lc:\\python27\\libs -LAUTO\\lib -LAUTO\\lib64
-Lc:\
\python27 -lpython27 -lcublas -lcudart')

It looks like the nvcc.flags field is getting mashed up against the -D arg.
 I don't know if that is causing the failures.  

Note, the following *does* work from a VS2010 command line with nvcc
[implying that the -m32 flag is not, itself, causing the failure]:

nvcc -shared -g ^
     -IC:\Users\mfenner\Theano\theano\sandbox\cuda ^
     -IC:\Python27\lib\site-packages\numpy\core\include ^
     -IC:\Python27\include ^
     -o out.cuda_ndarray.pyd ^
     ./cuda_ndarray.cu ^
     -LC:\Python27\libs ^
     -L"C:\Program Files\NVDIIA GPU Computing Toolkit\CUDA\v4.2\lib\x64" ^
     -lpython27 ^
     -lcublas ^
     -lcudart ^
     -m32

Now that's VERY interesting.  I didn't notice this before:  the -L arg is
for CUDA/4.2/lib/x64 and the -m32 seem incompatible.  Yet this compiles
cleanly (two warnings about unreachable code).  Perhaps it would fail at
run time.

Best,
Mark

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Frédéric Bastien  
View profile  
 More options Jul 25 2012, 3:47 pm
From: Frédéric Bastien <no...@nouiz.org>
Date: Wed, 25 Jul 2012 15:47:06 -0400
Local: Wed, Jul 25 2012 3:47 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author
Hi,

for the nvcc compiler, we add the -m32 or -m64 code for all platform
except for win32:

        if sys.platform != 'win32':
            if local_bitwidth() == 64:
                cmd.append('-m64')
                preargs2.append('-m64')
            else:
                cmd.append('-m32')
                preargs2.append('-m32')

Should we do that also on windows? git blame tell it is me that added
the special case for windows to try to make the gpu code work on
windows. But I never succeded. That was a bad try. Olivier, do you
think I should remove the first if? I think that local_bitwidth() will
return the number of bits for the python version and I think that is
what we need to ask nvcc to generate. Do you agree?

Mark, I saw all you email, but i don't know what else I can do to help
make the mixed 32/64 bits environmant work correctly. Do you think of
something? The -L...x64 add a path when searching for library, but on
linux, if it find a library with the right name, but wrong 32/64 bits,
it just ignore it and continue its search. So adding it is not a
problem on linux. I suppose it is the same in your case as when there
is a problem related to this, the error are a compile time, not at run
time.

Fred

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Olivier Delalleau  
View profile  
 More options Jul 25 2012, 4:53 pm
From: Olivier Delalleau <delal...@iro.umontreal.ca>
Date: Wed, 25 Jul 2012 16:53:39 -0400
Local: Wed, Jul 25 2012 4:53 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Ah sorry, I hadn't paid enough attention: it looks like modifying
nvcc.flags actually forwards the option to the compiler (through
-Xcompiler) instead of actually using this option on the command line :(

I guess for now, just to test it, manually edit
theano/sandbox/cuda/nvcc_compiler.py, look for "-shared", and replace that
line with:
cmd = [nvcc_path, '-shared', '-g', '-m32'] + preargs1
to see if that helps. Then later we can figure out how to make it clean :)

-=- Olivier

2012/7/25 Mark <mark.fen...@coveros.com>

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Olivier Delalleau  
View profile  
 More options Jul 25 2012, 4:56 pm
From: Olivier Delalleau <delal...@iro.umontreal.ca>
Date: Wed, 25 Jul 2012 16:56:00 -0400
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Fred,

I can't remember why we didn't need this on Windows in our tests... I
suspect however that adding the "-mX" flag to preargs2 is a bad idea,
because I don't think it's a flag that Visual Studio will understand.
But maybe we indeed need to append it to cmd.

-=- Olivier

2012/7/25 Frédéric Bastien <no...@nouiz.org>

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Olivier Delalleau  
View profile  
 More options Jul 25 2012, 4:58 pm
From: Olivier Delalleau <delal...@iro.umontreal.ca>
Date: Wed, 25 Jul 2012 16:58:05 -0400
Local: Wed, Jul 25 2012 4:58 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

For your Python install, the "bitness" of your architecture should be
available by running:
python -c "import struct ; print struct.calcsize('P') * 8"

-=- Olivier

2012/7/25 Mark <mark.fen...@coveros.com>

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Mark  
View profile  
 More options Jul 26 2012, 9:59 am
From: Mark <mark.fen...@coveros.com>
Date: Thu, 26 Jul 2012 06:59:09 -0700 (PDT)
Local: Thurs, Jul 26 2012 9:59 am
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Right, you can also use:  python -c "import platform;
platform.architecture()".  My uncertainty is about the fact that I
installed:

http://www.python.org/ftp/python/2.7.3/python-2.7.3.amd64.msi

And my architecture is reported as 32-bit:  ('32bit', 'WindowsPE').  There
is always a non-zero possibility that I installed a different python.

Best,
Mark

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Mark  
View profile  
 More options Jul 26 2012, 10:17 am
From: Mark <mark.fen...@coveros.com>
Date: Thu, 26 Jul 2012 07:17:13 -0700 (PDT)
Local: Thurs, Jul 26 2012 10:17 am
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

Success!  Well, partial success:  My theano script (aka, python script
using theano calls) is running happily; however, I'm not sure that it is
GPU-enabled and/or accelerated.  I'm going to double check that.

Olivier, thank you very much for your help.  In return, I'm going to start
from a clean slate (uninstall everything I can find) and re-work my steps,
document them, and submit it so others can follow it.

Best,
Mark

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Olivier Delalleau  
View profile  
 More options Jul 26 2012, 12:25 pm
From: Olivier Delalleau <delal...@iro.umontreal.ca>
Date: Thu, 26 Jul 2012 12:25:37 -0400
Local: Thurs, Jul 26 2012 12:25 pm
Subject: Re: [theano-users] Theano + GPU on Windows
Print | Individual message | Show original | Report this message | Find messages by this author

That sounds wrong, I'm pretty sure the installation file you mention should
be 64 bit (although I can't easily verify right now).

-=- Olivier

2012/7/26 Mark <mark.fen...@coveros.com>

...

read more »


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Messages 1 - 25 of 38   Newer >
« Back to Discussions « Newer topic     Older topic »

Create a group - Google Groups - Google Home - Terms of Service - Privacy Policy
©2013 Google