PyOpenCL June 2010

pyopencl@tiker.net

7 participants
12 discussions

by mt

Hi, I'm trying to access GL textures from pyopencl. Here is my test program: import sys, os, pygame from OpenGL.GL import * sys.path.append("extern/pyopencl/build/lib.linux-x86_64-2.6") import pyopencl pygame.init() screen = pygame.display.set_mode((1024, 768), pygame.HWSURFACE | pygame.OPENGL | pygame.DOUBLEBUF) if pyopencl.have_gl(): context = pyopencl.create_some_context() tex = glGenTextures(1) glBindTexture(GL_TEXTURE_2D, tex) glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, 256, 256, 0, GL_RGBA, GL_UNSIGNED_BYTE, None) cltex = pyopencl.GLTexture(context, pyopencl.mem_flags.READ_ONLY, GL_TEXTURE_2D, 0, tex, 2) It fails with error: Traceback (most recent call last): File "cl.py", line 14, in <module> cltex = pyopencl.GLTexture(context, pyopencl.mem_flags.READ_ONLY, GL_TEXTURE_2D, 0, tex, 2) pyopencl.LogicError: clCreateFromGLTexture2D failed: invalid context I thought that the problem might be in pyopencl's context creation, which doesn't take the GL context into account. I tried to fix it by adding appropriate CL_GL_CONTEXT_KHR, CL_GLX_DISPLAY_KHR and CL_CONTEXT_PLATFORM props to the context, but then I got another error "pyopencl.LogicError: clCreateFromGLTexture2D failed: invalid value". I can run kernels just fine with my setup, but this GL stuff won't work. What am I doing wrong?

8 years, 11 months

get_info fix for cl_nv_device_attribute_query extension

by Paolo Simone Gasparello

This patch make the cl_nv_device_attribute_query extension work with the latest nVidia Cuda toolkit 3.1. Tested on Ubuntu 10.4. Actually, the inclusion of the extension header is commented. Anyway I have used the cl_ext.h header provided by the official nVidia installer labeled "CUDA Toolkit for Ubuntu Linux 9.10" All the best, Paolo -- Paolo Simone Gasparello

10 years

Arrays, CL 1.1

by Andreas Kloeckner

Hi all, PyOpenCL finally has its own numpy.ndarray work-alikes. This has just landed in the git tree. The code is a port of PyCUDA's equivalent facilities, and like that code is built on machinery to evaluate arbitrary element-wise expressions. (See [1] for a description.) I will add documentation as soon as I have time, for now you may read the source in pyopencl/array.py. Tests are available in test/test_array.py. Unrelatedly, I'm not sure if I mentioned this: Since last Friday, PyOpenCL git supports and wraps OpenCL 1.1 if it is available at compile time. (It will of course still compile against OpenCL 1.0 implementations.) Looks like we'll have a pretty packed 0.92 release. :) I'll cook up a release candidate soon. Andreas [1] http://documen.tician.de/pycuda/array.html#module-pycuda.elementwise

10 years

Re: [PyOpenCL] API evolution (Important!)

by Andreas Kloeckner

On Sat, 26 Jun 2010 12:30:17 -0600, Cyrus Omar <cyology(a)gmail.com> wrote: > I think it makes more sense to force both global and local size to be > keyword arguments, as folks are used to the order of positional > arguments in a call being the same as those in a definition and it's > not clear on first glance why e.g. a.shape is being passed instead. > Similarly with queue, these are sort of meta-arguments and should > probably be distinguished as such. This seems more clear: > > sum(a, b, dest, queue=queue, global_size=a.shape, local_size=(256,)) > > Perhaps shorter names could be used (gsize, lsize, q) to minimize > typing? Or I wrote some extensions that implicitly pass a default > queue around and allow kernels to specify how to calculate sizes in > their definitions to mostly eliminate the issue as well. I'm -1 on your proposal, since it's a lot of typing. Here's a counterproposal, partially based on how the C++ wrapper does things: bound_sum = sum.bind(queue, None, (256,)) bound_sum(a, b, dest, global_size=a.shape) I.e. create an instance of a 'BoundKernel' type, which represents a binding of a kernel to a queue, and which may also hold defaults for local and global size, all of which *can* be overriden by kwargs at invocation time. This provides the clear separation of meta- and actual arguments that you're after. Opinions? (Just to be clear--whatever gets decided here will be backward-compatible, although it might deprecate current behavior.) Andreas PS: It'd be good if we had a consensus on this by Tuesday, because that's when I'll teach the pyopencl tutorial at SciPy'10. :)

10 years

Snow leopard woahs.

by Ryan Haynes

I was trying to build pyopencl on my default snow leopard. I booted into 64 bit mode. Built boost as in the how to. Built pyopencl as in the how to. Everything seems to run and compile fine. But upon importing pyopencl in my python interpreter I get the familiar. Python 2.6.1 (r261:67515, Feb 11 2010, 00:51:29) [GCC 4.2.1 (Apple Inc. build 5646)] on darwin Type "help", "copyright", "credits" or "license" for more information. >>> import pyopencl Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/Library/Python/2.6/site-packages/pyopencl/__init__.py", line 3, in <module> import pyopencl._cl as _cl ImportError: dlopen(/Library/Python/2.6/site-packages/pyopencl/_cl.so, 2): Symbol not found: __ZN5boost6python17error_already_setD1Ev Referenced from: /Library/Python/2.6/site-packages/pyopencl/_cl.so Expected in: flat namespace in /Library/Python/2.6/site-packages/pyopencl/_cl.so >>> exit() I think I've seen some of these to help. rhaynes$ otool -L /Library/Python/2.6/site-packages/pyopencl/_cl.so /Library/Python/2.6/site-packages/pyopencl/_cl.so: /System/Library/Frameworks/OpenCL.framework/Versions/A/OpenCL (compatibility version 1.0.0, current version 1.0.0) /usr/lib/libstdc++.6.dylib (compatibility version 7.0.0, current version 7.9.0) /usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 123.0.0) rhaynes$ otool -L /Users/rhaynes/pool/lib/libboost_python.dylib /Users/rhaynes/pool/lib/libboost_python.dylib: libboost_python.dylib (compatibility version 0.0.0, current version 0.0.0) /System/Library/Frameworks/Python.framework/Versions/2.6/Python (compatibility version 2.6.0, current version 2.6.1) /usr/lib/libstdc++.6.dylib (compatibility version 7.0.0, current version 7.9.0) /usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 125.2.0) rhaynes$ lipo -info `which python` Architectures in the fat file: /usr/bin/python are: x86_64 i386 ppc7400 I was using the Default install of python because the install wiki didn't advise against it. For some reason I cannot get the link to the complete build of python, numpy, scipy etc. I'm not opposed to building it all from scratch. But I'd prefer be lazy... I'm doing the tutorials at scipy conf this coming week. Any help I could get to have this running would be much appreciated. Cheers, Ryan

10 years

API evolution (Important!)

by Andreas Kloeckner

Hi all, I've made a (compatible) change to a fairly central pyopencl interface, in the hopes of making everybody's life easier--but this also means that nearly every pyopencl program in existence now contains code that is considered deprecated and will result in warnings from pyopencl 0.92 onward and will stop working in 0.94. The change involves Kernel.__call__, where I've moved 'local_size' from keyword argument to third positional argument. Luckily, old and new usage can be safely disambiguated. Before: kernel(queue, h_c.shape, d_c_buf, d_a_buf, d_b_buf, local_size=(block_size, block_size)) After: kernel(queue, h_c.shape, (block_size, block_size), d_c_buf, d_a_buf, d_b_buf) I find the new interface better, because a) it groups related arguments together and b) it more easily deals with *args-style argument passing (At least as long as we're targeting Python 2.x.). The reason for the change is that local_size will be passed to nearly every call to this interface, so the keyword-arg position was inconvenient. This change is painful now, but keeping this annoying interface around would have continued to be painful forever. That's why I decided to change. The change just landed in git. Verbose warnings are given for each case of deprecated use. All code in pyopencl itself was updated to the new usage. If you have comments, questions, or suggestions, please do speak up now, *before* this change becomes part of a release. Andreas

10 years

No module named _cl

by Rick Gerkin

In Windows 7x64 on a 64-bit machine (Mac Pro running Boot Camp), I am trying to install pyOpenCl. I can successfully compile Boost 1.43 from source, an run the setup.py in pyOpenCl with appropriate (I think) options in site-config, using Visual Studio 9. I get no errors. After installing pyOpenCl in site-packages, I get the following error on trying to import it: >>> import pyopencl Traceback (most recent call last): File "<stdin>", line 1, in <module> File "pyopencl\__init__.py", line 3, in <module> import pyopencl._cl as _cl ImportError: No module named _cl I tried running Dependency Walker on the pyopencl\_cl.pyd file and found it looking for both boost_python-vc90-mt-1_43.dll and IEShims.dll, so I threw these in the pyopencl directory and it still does not import (same error). Dependency Walker lists two kinds of errors: Error: At least one module has an unresolved import due to a missing export function in an implicitly dependent module. Error: Modules with different CPU types were found. I don't understand the former. For the latter, all dependencies are 64-bit libraries except for MSVCR90.dll, MSVCP90.dll, the boost_python dll mentioned above, python26.dll, and the _cl.pyd I am examining. For the next step in getting this to work, I don't what of all this is relevant. Should I abandon trying to make it work in a 64-bit OS? Has anyone ever made it work in 64-bit Windows? Thanks, Rick -- Richard C. Gerkin Postdoctoral Fellow, Dept. of Biological Sciences Carnegie Mellon University

10 years

Re: [PyOpenCL] [Andreas Kloeckner] Re: OpenGL Interoperability

by Andreas Kloeckner

Hi Paolo, On Wed, 23 Jun 2010 12:31:13 +0200, Paolo Simone Gasparello <djgaspa(a)gmail.com> wrote: > Hello everybody, > I make my code doing ctypes casting in C++ work, so there is no more the > need to do that casting explicitly in python code. (patch attached) > > I also wrote a simple example working on linux (I have used glX APIs and > flags) showing the usage of GL interop. It draws the sin function using > coordinates that are generated directly by the video hardware. > > I hope this will be useful :) Thanks for your hard work in getting this far! I've added your fixes and your example to pyopencl git. Andreas

10 years

Re: [PyOpenCL] [Andreas Kloeckner] Re: OpenGL Interoperability

by Paolo Simone Gasparello

Il giorno sab, 19/06/2010 alle 15.04 -0400, Andreas Kloeckner ha scritto: > Hi Paolo, > > On Tue, 15 Jun 2010 12:02:07 +0200, Paolo Simone Gasparello <djgaspa(a)gmail.= > com> wrote: > > Hi guys, anyone got OpenGL interoperability work? > > With the following patch I'm able to create a context sharing OpenGL > > buffers. > > Applied, thanks. (And I'm kind of embarrassed that you had to fix these > two...) :) > > > Where fun_double is a kernel included in the file program.cl that simply > > does each coord *=3D 2. > >=20 > > There are better ways to do this? I tryed to perform the cast to > > c_void_p in C++ code using boost::python, but I was not able to make the > > code work. > > I took a look, and it doesn't seem like ctypes has much in the way of > C-accessible interface. One can of course emulate the Python code in > C++, but I'm not sure that's necessary. If you'd like guidance on how to > do that, let me know. > > Lastly, would you mind posting a fully functional example so that I can > add it to the examples/ directory shipped with PyOpenCL? I think that > could be rather helpful for people trying to follow in your footsteps. > > Thanks for your work! > Andreas Hello everybody, I make my code doing ctypes casting in C++ work, so there is no more the need to do that casting explicitly in python code. (patch attached) I also wrote a simple example working on linux (I have used glX APIs and flags) showing the usage of GL interop. It draws the sin function using coordinates that are generated directly by the video hardware. I hope this will be useful :) -- Paolo Simone Gasparello <djgaspa(a)gmail.com>

10 years

Fixed: curious slowness of PyOpenCL matrix-multiply example

by Andreas Kloeckner

Hi all, if you have been wondering why the matrix-multiply example shipped with PyOpenCL shows sub-standard performance on Nvidia hardware, wonder no longer. In anticipation of next week's SciPy conference, I've finally fixed that, and it turned out to be (d'oh!) bank conflicts. Which is odd, since the example was (at some point) derived from Nvidia's own SDK example. Anyway, for me, matmul performance on the same hardware is now comparable between CL and CUDA. Just thought I'd let you know. Happy hacking, Andreas

10 years

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

PyOpenCL June 2010