3:21 p.m.
Hi,
I'm new to pyopencl as well as to OpenCL in general. Not surprisingly, I
got to a point where I'm stuck.
Here's my setup:
Hardware: To start learning and play around with GPU computing I bought
a cheap Nvidia 8400GS card for a PCI (non-Express!) slot and put it in
an oldish computer (Pentium M - based, laptop CPU on a desktop board).
The computer also has an old AGP graphics card, but no display attached
right now to that one, only to the Nvidia PCI card. The 840GS has 256MB.
Software: OS is Ubuntu 9.10. I managed to build and install pyopencl ok
with the usual amount of fiddling. (import pyopencl works, as does
running the example file dump-properties.py) First I used 0.91.4, then
today the git version, hoping that the error would go away (but it didn't).
The error: Whenever I run any of the example files, I get:
Traceback (most recent call last):
File "examples/demo.py", line 8, in <module>
ctx = cl.create_some_context()
File "/usr/local/lib/python2.6/dist-packages/pyopencl/__init__.py",
line 275, in create_some_context
return Context(devices)
pyopencl.RuntimeError: Context failed: out of resources
As I said, this is where I'm stuck. Any ideas?
And while I'm at it, another question: Does pyopencl have something like
the gpuarray thing that I read about in pycuda?
Thanks for your help,
sven
3:51 p.m.
Hi Sven,
On Freitag 05 Februar 2010, Sven Schreiber wrote:
I'm new to pyopencl as well as to OpenCL in general. Not
surprisingly, I
got to a point where I'm stuck.
The error: Whenever I run any of the example files, I get:
Traceback (most recent call last):
File "examples/demo.py", line 8, in <module>
ctx = cl.create_some_context()
File "/usr/local/lib/python2.6/dist-packages/pyopencl/__init__.py",
line 275, in create_some_context
return Context(devices)
pyopencl.RuntimeError: Context failed: out of resources
As I said, this is where I'm stuck. Any ideas?
Weird--never seen such an error. Can you try and run a pure-C CL
example (such as the ones shipped by Nvidia) and see if that works?
And while I'm at it, another question: Does pyopencl have
something like
the gpuarray thing that I read about in pycuda?
Not yet--but if you're just getting your feet wet, porting that code to
CL would make a nice introductory exercise. Patches welcome. :)
Andreas
9:58 p.m.
Andreas Klöckner schrieb:
Hi Sven,
thanks for the quick reply!
On Freitag 05 Februar 2010, Sven Schreiber wrote:
> I'm new to pyopencl as well as to OpenCL in general. Not surprisingly, I
> got to a point where I'm stuck.
> The error: Whenever I run any of the example files, I get:
>
> Traceback (most recent call last):
> File "examples/demo.py", line 8, in <module>
> ctx = cl.create_some_context()
> File "/usr/local/lib/python2.6/dist-packages/pyopencl/__init__.py",
> line 275, in create_some_context
> return Context(devices)
> pyopencl.RuntimeError: Context failed: out of resources
>
> As I said, this is where I'm stuck. Any ideas?
Weird--never seen such an error. Can you try and run a pure-C CL
example (such as the ones shipped by Nvidia) and see if that works?
I'll try -- but it'll probably take a while as I'm facing some other
deadlines right now. But I will report back eventually ( != German
"eventuell" which it occurs to me otherwise you might interpret it in
this false-friend kind of way).
> And while I'm at it, another question: Does pyopencl have something like
> the gpuarray thing that I read about in pycuda?
Not yet--but if you're just getting your feet wet, porting that code to
CL would make a nice introductory exercise. Patches welcome. :)
well I'm sorry but I don't really speak C which I suppose is what you
mean. That's why pyopencl (and also pycuda) look so great to me.
thanks,
sven
11:07 p.m.
That is interesting. This is only a hunch and probably wrong.
What brand is the AGP card that doesn't have a display? Chances are that the
OpenCL driver detects that there are two NVidia cards on the system and when
you try to create a context it realizes that one of them cannot possibly
support OpenCL because it's too old.
Possibly we could modify pyopencl so that if there are multiple devices in a
platform and the context creation fails it could try creating the context on
a different device (in the same platform).
Cheers,
David
On Fri, Feb 5, 2010 at 4:58 PM, Sven Schreiber <svetosch(a)gmx.net> wrote:
Andreas Klöckner schrieb:
> Hi Sven,
thanks for the quick reply!
>
> On Freitag 05 Februar 2010, Sven Schreiber wrote:
>> I'm new to pyopencl as well as to OpenCL in general. Not surprisingly, I
>> got to a point where I'm stuck.
>> The error: Whenever I run any of the example files, I get:
>>
>> Traceback (most recent call last):
>> File "examples/demo.py", line 8, in <module>
>> ctx = cl.create_some_context()
>> File "/usr/local/lib/python2.6/dist-packages/pyopencl/__init__.py",
>> line 275, in create_some_context
>> return Context(devices)
>> pyopencl.RuntimeError: Context failed: out of resources
>>
>> As I said, this is where I'm stuck. Any ideas?
>
> Weird--never seen such an error. Can you try and run a pure-C CL
> example (such as the ones shipped by Nvidia) and see if that works?
I'll try -- but it'll probably take a while as I'm facing some other
deadlines right now. But I will report back eventually ( != German
"eventuell" which it occurs to me otherwise you might interpret it in
this false-friend kind of way).
>
>> And while I'm at it, another question: Does pyopencl have something like
>> the gpuarray thing that I read about in pycuda?
>
> Not yet--but if you're just getting your feet wet, porting that code to
> CL would make a nice introductory exercise. Patches welcome. :)
>
well I'm sorry but I don't really speak C which I suppose is what you
mean. That's why pyopencl (and also pycuda) look so great to me.
thanks,
sven
_______________________________________________
PyOpenCL mailing list
PyOpenCL(a)host304.hostmonster.com
http://host304.hostmonster.com/mailman/listinfo/pyopencl_tiker.net
2:03 p.m.
Hi David,
the AGP card is an Ati Radeon 9200.
Actually, the whole driver and hardware situation in my case seems to be
a bit flaky. There is also an Intel onboard graphics which however is
disabled in the BIOS, apart from the fact that the mainboard is not
designed to use the onboard and AGP graphics in parallel, I think. So
when an AGP card is plugged in, onboard graphics wouldn't work even if
it's not explicitly disabled in the BIOS.
First I wanted to have the display on the AGP card and use the Nvidia
PCI card only for (py)opencl. But then it seemed that the Nvidia driver
didn't get loaded, and pyopencl complained that it didn't find any
devices. Still that would be my ideal setup for now, so if you have any
advice on how to get that to work I would also be grateful.
Removing the Radeon AGP card caused big problems although in principle
it shouldn't affect anything (no display attached after all...), but
then I didn't get a usable output even via the Nvidia PCI card. Weird.
(Maybe related to the xorg Intel driver mess?)
So -- when I have the time I'll try to check some plain C/CL examples as
Andreas suggested.
thanks,
sven
David Garcia schrieb:
That is interesting. This is only a hunch and probably wrong.
What brand is the AGP card that doesn't have a display? Chances are that
the OpenCL driver detects that there are two NVidia cards on the system
and when you try to create a context it realizes that one of them cannot
possibly support OpenCL because it's too old.
Possibly we could modify pyopencl so that if there are multiple devices
in a platform and the context creation fails it could try creating the
context on a different device (in the same platform).
Cheers,
David
2:05 p.m.
Sven Schreiber schrieb:
So -- when I have the time I'll try to check some plain C/CL
examples as
Andreas suggested.
Turns out those aren't working either. Seems that gcc 4.4 in Ubuntu 9.10
is not compatible with Nvidia's current SDK? I tried to remove gcc 4.4
and install 4.3 (and adding a g++ symlink to g++-4.3 in /usr/bin I
think), which got me a little farther, but no success either.
Is anybody here running on Ubuntu 9.10 and/or gcc 4.4?
thanks,
sven
2:09 p.m.
On Mon, Feb 8, 2010 at 7:35 PM, Sven Schreiber <svetosch(a)gmx.net> wrote:
Sven Schreiber schrieb:
> So -- when I have the time I'll try to check some plain C/CL examples as
> Andreas suggested.
>
Turns out those aren't working either. Seems that gcc 4.4 in Ubuntu 9.10
is not compatible with Nvidia's current SDK? I tried to remove gcc 4.4
and install 4.3 (and adding a g++ symlink to g++-4.3 in /usr/bin I
think), which got me a little farther, but no success either.
Is anybody here running on Ubuntu 9.10 and/or gcc 4.4?
I am using this combination, but I am also using AMD's opencl
implementation for CPU's. It's working fine for me, so I think it is a
nv sdk issue.
thanks,
sven
_______________________________________________
PyOpenCL mailing list
PyOpenCL(a)host304.hostmonster.com
http://host304.hostmonster.com/mailman/listinfo/pyopencl_tiker.net
--
Rohit Garg
http://rpg-314.blogspot.com/
Senior Undergraduate
Department of Physics
Indian Institute of Technology
Bombay
3:07 p.m.
Rohit Garg schrieb:
>>
> Turns out those aren't working either. Seems that gcc 4.4 in Ubuntu 9.10
> is not compatible with Nvidia's current SDK? I tried to remove gcc 4.4
> and install 4.3 (and adding a g++ symlink to g++-4.3 in /usr/bin I
> think), which got me a little farther, but no success either.
>
> Is anybody here running on Ubuntu 9.10 and/or gcc 4.4?
I am using this combination, but I am also using AMD's opencl
implementation for CPU's. It's working fine for me, so I think it is a
nv sdk issue.
Yes, absolutely! So really my question was meant as:
Ubuntu 9.10 and Nvidia SDK howto?
-sven
6:25 p.m.
Sven Schreiber wrote:
Rohit Garg schrieb:
>> Turns out those aren't working either. Seems that gcc 4.4 in Ubuntu 9.10
>> is not compatible with Nvidia's current SDK? I tried to remove gcc 4.4
>> and install 4.3 (and adding a g++ symlink to g++-4.3 in /usr/bin I
>> think), which got me a little farther, but no success either.
>>
>> Is anybody here running on Ubuntu 9.10 and/or gcc 4.4?
>>
> I am using this combination, but I am also using AMD's opencl
> implementation for CPU's. It's working fine for me, so I think it is a
> nv sdk issue.
>
Yes, absolutely! So really my question was meant as:
Ubuntu 9.10 and Nvidia SDK howto?
The examples are working for me on Karmic with the attached siteconf.py,
but I haven't gone any further. I'm using amd64 arch, the 195.30 beta
drivers, and a GeForce GTX 260.
-Andrew
12:35 p.m.
Andrew Straw schrieb:
Sven Schreiber wrote:
> Rohit Garg schrieb:
>
>
>>> Turns out those aren't working either. Seems that gcc 4.4 in Ubuntu 9.10
>>> is not compatible with Nvidia's current SDK? I tried to remove gcc 4.4
>>> and install 4.3 (and adding a g++ symlink to g++-4.3 in /usr/bin I
>>> think), which got me a little farther, but no success either.
>>>
>>> Is anybody here running on Ubuntu 9.10 and/or gcc 4.4?
>>>
>> I am using this combination, but I am also using AMD's opencl
>> implementation for CPU's. It's working fine for me, so I think it is a
>> nv sdk issue.
>>
> Yes, absolutely! So really my question was meant as:
>
> Ubuntu 9.10 and Nvidia SDK howto?
>
The examples are working for me on Karmic with the attached siteconf.py,
but I haven't gone any further. I'm using amd64 arch, the 195.30 beta
drivers, and a GeForce GTX 260.
Thanks, I will keep the possibility in mind to upgrade to the beta
drivers. However, searching for "nvidia sdk gcc 4.4" I found some
instructions how to get the cuda sdk up and running on ubuntu 9.10. I'll
try these soon and probably report back here to leave some hints for
future readers with the same problem.
-sven
1:41 p.m.
Sven Schreiber schrieb:
Andrew Straw schrieb:
> Sven Schreiber wrote:
>>>
>> Yes, absolutely! So really my question was meant as:
>>
>> Ubuntu 9.10 and Nvidia SDK howto?
>>
> The examples are working for me on Karmic with the attached siteconf.py,
> but I haven't gone any further. I'm using amd64 arch, the 195.30 beta
> drivers, and a GeForce GTX 260.
>
Thanks, I will keep the possibility in mind to upgrade to the beta
drivers. However, searching for "nvidia sdk gcc 4.4" I found some
instructions how to get the cuda sdk up and running on ubuntu 9.10. I'll
try these soon and probably report back here to leave some hints for
future readers with the same problem.
Ok, so I have pyopencl-0.91.4 as well as pycuda-0.93 now up and running
here.
The combination is (short version):
* Driver 195.30 beta
* gcc symlink pointing to gcc-4.4 for compiling the driver kernel
module, but pointing to gcc-4.3 for the rest
* cudatoolkit_2.3_linux_32_ubuntu9.04.run
* gpucomputingsdk_2.3b_linux.run
* (for the Cuda stuff, following the advice in
http://moelhave.dk/2009/12/nvidia-cuda-on-ubuntu-karmic-koala/; and for
Nvidia's OpenCL examples I also changed the CXX, CC, and LINK lines in
OpenCL/common/common_opencl.mk)
I had problems with the 190.29 drivers, and while pycuda worked with the
190.53 drivers, (py)opencl didn't -- I guess the latter is expected. So
for me indeed only the 195.30 beta drivers seem to work with both.
BTW, a remark about the benchmark-all.py example file. I think the speed
comparison there is a little biased in favor of pyopencl. It compares
(almost) pure Python with pyopencl, but IMHO the more meaningful
comparison would be between Numpy vectorized code and pyopencl. AFAICS
the numpy equivalent of the pure Python code would be:
for j in range(1000): # number of iterations, just for comparability
n_result = (a+b)**2 * (a/2.0)
At least the results seem to agree when checked afterwards. On my test
system I get the following timings:
* pure Python: 20.85s
* vectorized Numpy on CPU: 0.044s
* pyopencl on GPU: 0.034s
Of course I'm *not* saying that the pyopencl approach isn't fast and
useful. (My test graphics card is very low end and is on the slow PCI
bus.) But the first one or two orders of magnitude can be achieved
already without any GPU magic.
thank you for these very cool tools,
sven
5:13 p.m.
The raw computing power of any modern GPU is vastly superior to a modern
CPU. However, you need to keep in mind that GPUs are devices that run
asynchronously from the CPU that is running the Python interpreter and also
they have their own physically separate memory (this is true at least on
desktop PCs).
This two restrictions put together mean that there's a significant overhead
associated with doing any brief computation on the GPU. You need to consider
the amount of data that is being transferred from the CPU's RAM into the
GPU's RAM and compare it with the time that the computation itself is going
to take. If all you are doing is doing a component-wise vector addition, the
cost of moving data around is going to be greater than the cost of the
actual ALU instructions, which is why you are seeing some disappointing
performance.
Reduce the amount of communication and synchronization between the CPU and
the GPU and then you will see what these devices can do.
Cheers,
David
On Sat, Feb 13, 2010 at 8:41 AM, Sven Schreiber <svetosch(a)gmx.net> wrote:
Sven Schreiber schrieb:
> Andrew Straw schrieb:
>> Sven Schreiber wrote:
>>>>
>>> Yes, absolutely! So really my question was meant as:
>>>
>>> Ubuntu 9.10 and Nvidia SDK howto?
>>>
>> The examples are working for me on Karmic with the attached siteconf.py,
>> but I haven't gone any further. I'm using amd64 arch, the 195.30 beta
>> drivers, and a GeForce GTX 260.
>>
>
> Thanks, I will keep the possibility in mind to upgrade to the beta
> drivers. However, searching for "nvidia sdk gcc 4.4" I found some
> instructions how to get the cuda sdk up and running on ubuntu 9.10. I'll
> try these soon and probably report back here to leave some hints for
> future readers with the same problem.
>
Ok, so I have pyopencl-0.91.4 as well as pycuda-0.93 now up and running
here.
The combination is (short version):
* Driver 195.30 beta
* gcc symlink pointing to gcc-4.4 for compiling the driver kernel
module, but pointing to gcc-4.3 for the rest
* cudatoolkit_2.3_linux_32_ubuntu9.04.run
* gpucomputingsdk_2.3b_linux.run
* (for the Cuda stuff, following the advice in
http://moelhave.dk/2009/12/nvidia-cuda-on-ubuntu-karmic-koala/; and for
Nvidia's OpenCL examples I also changed the CXX, CC, and LINK lines in
OpenCL/common/common_opencl.mk)
I had problems with the 190.29 drivers, and while pycuda worked with the
190.53 drivers, (py)opencl didn't -- I guess the latter is expected. So
for me indeed only the 195.30 beta drivers seem to work with both.
BTW, a remark about the benchmark-all.py example file. I think the speed
comparison there is a little biased in favor of pyopencl. It compares
(almost) pure Python with pyopencl, but IMHO the more meaningful
comparison would be between Numpy vectorized code and pyopencl. AFAICS
the numpy equivalent of the pure Python code would be:
for j in range(1000): # number of iterations, just for comparability
n_result = (a+b)**2 * (a/2.0)
At least the results seem to agree when checked afterwards. On my test
system I get the following timings:
* pure Python: 20.85s
* vectorized Numpy on CPU: 0.044s
* pyopencl on GPU: 0.034s
Of course I'm *not* saying that the pyopencl approach isn't fast and
useful. (My test graphics card is very low end and is on the slow PCI
bus.) But the first one or two orders of magnitude can be achieved
already without any GPU magic.
thank you for these very cool tools,
sven
_______________________________________________
PyOpenCL mailing list
PyOpenCL(a)host304.hostmonster.com
http://host304.hostmonster.com/mailman/listinfo/pyopencl_tiker.net
8:33 p.m.
David Garcia schrieb:
This two restrictions put together mean that there's a significant
overhead associated with doing any brief computation on the GPU. You
need to consider the amount of data that is being transferred from the
CPU's RAM into the GPU's RAM and compare it with the time that the
computation itself is going to take. If all you are doing is doing a
component-wise vector addition, the cost of moving data around is going
to be greater than the cost of the actual ALU instructions, which is why
you are seeing some disappointing performance.
David, I'm aware of the issues you mention, and I wasn't disappointed
about the timings. I just took the benchmark case as given which is
distributed with pyopencl; I didn't cook it up myself. Your comments
seem to imply that using another benchmark case may be more informative.
thanks for your feedback,
sven
3628
days inactive
3637
days old
12 comments
5 participants
participants (5)
-
Andreas Klöckner -
Andrew Straw -
David Garcia -
Rohit Garg -
Sven Schreiber