PyCUDA April 2010

pycuda@tiker.net

23 participants
21 discussions

by Riccardo Taormina

Hi everybody, I am new on PyCuda. I just installed everything on Windows XP and, from the installation log, I think that I did it properly. However, I tried to run the test files provided with pycuda and I get this error Traceback (most recent call last): File "C:\PyCuda\test\test_gpuarray.py", line 2, in <module> import pycuda.autoinit File "C:\PyCuda\pycuda\autoinit.py", line 1, in <module> import pycuda.driver as cuda File "C:\PyCuda\pycuda\driver.py", line 1, in <module> from _driver import * ImportError: No module named _driver how can I solve it? thanks and sorry for the newbieness of this post den3b _________________________________________________________________ I tuoi amici sempre a portata di clic, sul nuovo Web Messenger http://www.windowslive.it/foto.aspx

9 years, 11 months

How to manually free GPUarray to avoid leak?

by gerald wrong

I think I may be running into a memory leak using GPUarray. I have a function using GPUarrays that is working stable on single calls. If I loop this function within python from another script like this: for i in xrange(m): do_some_gpuarray_stuff() I can watch the memory pointers of the gpuarrays increase until I get a launch error... presumably due to lack of memory. ie I need gpu mem to free upon exit of do_some_gpuarray_stuff(), so I can repeat same gpu calculation many times on new data sets. Can I manually free GPUarray instances? If not, can I somehow manually remove all PyCUDA stuff from memory? like... for i in xrange(m): do_some_gpuarray_stuff() de_init_pycuda_mem I could not find this in the docs, and I understand everything is supposed to be automagically handled by PyCUDA, but manually freeing will be an easy confirmation/workaround for my problem. I know this can be done with pycuda.driver completely manually, but gpu_array is already working nicely and cleanly.... except for this leak. Any input from the experts would be much appreciated. Thanks much :) Garrett Wright

10 years, 4 months

creating a new view of a GPUArray

by Amir

I would like to create different views of a GPUArray. I have no idea how to do it. Let A be a 2D float32 c-continuous array. How do I create a 1D view of one of its rows? I am not sure how to set gpudata in the view. An example would be great. Is there a pointer arithmetic trick? Is it possible to create a float32 view of a complex64 array? I am trying to use gpuarray.multi_take_put to copy slices between float and complex arrays but it requires dtypes to match. Is using gpuarray.multi_take_put the best way to copy a slice of an linear array? Thanks, Amir.

10 years, 4 months

Re: [PyCUDA] test_gpuarray.py errors

by Chethan Pandarinath

Hi Bryan, thanks for the quick reply. I understand the issue now. So it looks like any devices of compute capability 1.x would not have support for double precision (from section 5.1.1.1 of the NVIDIA CUDA Programming Guide). Does this seem like a reasonable way of checking for double precision support in the test script? e.g., : import pycuda.driver as cuda dev=cuda.Device(0); if dev.compute_capability() < (2,0): no double precision support, don't run those tests... On Fri, Apr 23, 2010 at 9:07 PM, Bryan Catanzaro <bryan.catanzaro(a)gmail.com>wrote: > The GPU in your MacBook doesn't support double precision, which is why > these tests are failing. > You're of course welcome to change the tests so that they check for double > precision support before running them - that would probably help out others > who end up in this situation. > > - bryan > > On Apr 23, 2010, at 5:41 PM, Chethan Pandarinath < > chethan.pandarinath(a)gmail.com> wrote: > > Hi everybody, > > I've been working on installing PyCUDA on my MacBook (Snow Leopard). I > think I'm close to having it working, it's been quite a long road: getting > compatible versions of boost, setuptools, pytools... anyway, I think I'm > past all that. > > I can now successfully run test_math.py and test_driver.py with no errors. > > I am having trouble, however, with test_gpuarray.py > > I'm attaching the output from running the test script, with the standard > error appended as well. > $ arch -i386 python test_gpuarray.py >test_gpuarray_output.txt > > > There are a couple types of errors that I can see here: > > E AssertionError: (array(-4.2129004090721698e+36), > -1.2509687918788644e+303, <type 'numpy.float64'>, 'min') > test_gpuarray.py:328: AssertionError > > E assert array(-6.0786212321272663e+144) == > -2.6357594520767543e+301 > test_gpuarray.py:357: AssertionError > > E LaunchError: cuCtxPopCurrent failed: launch failed > /Library/Python/2.6/site-packages/pycuda/tools.py:504: LaunchError > > E RuntimeError: make_default_context() wasn't able to create a > context on any of the 1 detected devices > /Library/Python/2.6/site-packages/pycuda/tools.py:216: RuntimeError > > > To tell you the truth I don't know how to begin debugging this. If anyone > can point me in the right direction, I'd greatly appreciate it. Willing to > provide any info that would help... > > I'm running this on a MacBook, Snow Leopard (10.6.3), with an NVIDIA > GeForce 9400M. Using Python 2.6.1. > > Thanks. > Chethan > > > > -- > Chethan Pandarinath > <chethan.pandarinath(a)gmail.com>chethan.pandarinath(a)gmail.com > > <test_gpuarray_output.txt> > > _______________________________________________ > PyCUDA mailing list > PyCUDA(a)host304.hostmonster.com > http://host304.hostmonster.com/mailman/listinfo/pycuda_tiker.net > > -- Chethan Pandarinath chethan.pandarinath(a)gmail.com

10 years, 5 months

test_gpuarray.py errors

by Chethan Pandarinath

Hi everybody, I've been working on installing PyCUDA on my MacBook (Snow Leopard). I think I'm close to having it working, it's been quite a long road: getting compatible versions of boost, setuptools, pytools... anyway, I think I'm past all that. I can now successfully run test_math.py and test_driver.py with no errors. I am having trouble, however, with test_gpuarray.py I'm attaching the output from running the test script, with the standard error appended as well. $ arch -i386 python test_gpuarray.py >test_gpuarray_output.txt There are a couple types of errors that I can see here: E AssertionError: (array(-4.2129004090721698e+36), -1.2509687918788644e+303, <type 'numpy.float64'>, 'min') test_gpuarray.py:328: AssertionError E assert array(-6.0786212321272663e+144) == -2.6357594520767543e+301 test_gpuarray.py:357: AssertionError E LaunchError: cuCtxPopCurrent failed: launch failed /Library/Python/2.6/site-packages/pycuda/tools.py:504: LaunchError E RuntimeError: make_default_context() wasn't able to create a context on any of the 1 detected devices /Library/Python/2.6/site-packages/pycuda/tools.py:216: RuntimeError To tell you the truth I don't know how to begin debugging this. If anyone can point me in the right direction, I'd greatly appreciate it. Willing to provide any info that would help... I'm running this on a MacBook, Snow Leopard (10.6.3), with an NVIDIA GeForce 9400M. Using Python 2.6.1. Thanks. Chethan -- Chethan Pandarinath chethan.pandarinath(a)gmail.com

10 years, 5 months

[ANN] 0.94rc -- please test

by Andreas Klöckner

Hi all, PyCUDA's present release version (0.93) is starting to show its age, and so I've just rolled a release candidate for 0.94, after tying up a few loose ends--such as complete CUDA 3.0 support. Please help make sure 0.94 is solid. Go to http://pypi.python.org/pypi/pycuda/0.94rc to download the package, see if it works for you, and report back. The change log for 0.94 is here: http://documen.tician.de/pycuda/misc.html#version-0-94 but the big-ticket things in this release are: - Support for CUDA 3.0 - Sparse matrices - Complex numbers Let's make this another another rockin' release! Thanks very much for your help, Andreas

10 years, 5 months

32-bit PyCUDA on Snow Leopard

by Cyrus Omar

So I've been spending the last several hours trying to install PyCUDA on Snow Leopard. Initially, seeing that CUDA 3.0 was 64-bit on OSX, I attempted to install everything 64-bit (boost, python, pycuda, etc.). I got as far as running test_driver.py which froze my machine completely, needing a hard reboot. No error message or anything (not the ones others have been reporting). Ok, take 2, try to install everything 32-bit. I got boost compiled as 32-bit, a fresh install of Sage 32-bit for Python. PyCUDA 0.93 after patching out setuptools for distutils makes and test_driver.py runs to success with the following strange warnings: /Users/cyrus/sage/local/lib/python2.6/site-packages/pycuda-0.93-py2.6-macosx-10.5-x86_64.egg/pycuda/compiler.py:11: UserWarning: call_capture_stdout is deprecated: use call_capture_output instead return call_capture_stdout([nvcc, "--version"]) /Users/cyrus/sage/local/lib/python2.6/site-packages/pycuda-0.93-py2.6-macosx-10.5-x86_64.egg/pycuda/compiler.py:192: UserWarning: Reading 'lmem' from cubin failed--SourceModule metadata may be unavailable. warn("Reading '%s' from cubin failed--SourceModule metadata may be unavailable." % key) /Users/cyrus/sage/local/lib/python2.6/site-packages/pycuda-0.93-py2.6-macosx-10.5-x86_64.egg/pycuda/compiler.py:192: UserWarning: Reading 'smem' from cubin failed--SourceModule metadata may be unavailable. warn("Reading '%s' from cubin failed--SourceModule metadata may be unavailable." % key) /Users/cyrus/sage/local/lib/python2.6/site-packages/pycuda-0.93-py2.6-macosx-10.5-x86_64.egg/pycuda/compiler.py:192: UserWarning: Reading 'reg' from cubin failed--SourceModule metadata may be unavailable. warn("Reading '%s' from cubin failed--SourceModule metadata may be unavailable." % key) Which seems to mean mod.smem and all that return None instead of the proper values. The kernels still seem to work however, so this is vaguely acceptable at the moment. Trying to install PyCUDA 0.94rc (from PyPI) also finishes make but many of the tests now fail with errors like this and nothing works: E ImportError: dlopen(/Users/cyrus/sage/local/lib/python2.6/site-packages/pycuda-0.94rc-py2.6-macosx-10.5-x86_64.egg/pycuda/_pvt_struct.so, 2): no suitable image found. Did find: E /Users/cyrus/sage/local/lib/python2.6/site-packages/pycuda-0.94rc-py2.6-macosx-10.5-x86_64.egg/pycuda/_pvt_struct.so: mach-o, but wrong architecture I have no idea why both 0.93 and 0.94 call the directory macosx-10.5-x86_64 instead of 10.6-i386. The latest git version has strange issues during make install where it ends with an error about configure.py not having been run the first time I do it (even though it had been) and then running it again goes to completion. The test_driver.py errors are the same as 0.94rc. For reference, PyOpenCL works fine in 64-bit mode on this machine. Any ideas? Cyrus

10 years, 5 months

Hiring a GIS Systems Optimizer in the Modi Research Group of the Earth Institute at Columbia University

by Roy Hyunjin Han

Dear NYC Pythonistas, We at the Modi Research Group of the Earth Institute at Columbia University are hiring another 100% Python commando to join our team of consultants. The Modi Research Group plans energy services and infrastructure in developing countries. This is a full-time position with benefits, based on the university campus. Please apply through the Columbia University website at http://jobs.columbia.edu/applicants/Central?quickFind=118575 We will be optimizing, web-enabling and scaling our remote sensing system for finding houses in satellite images and integrating it with our existing web-based infrastructure planning system. Both systems are written almost entirely using Python. The web framework is build on Pylons using jQuery and OpenLayers, scaled locally using RabbitMQ and AMQP but we are considering Amazon EC2 for the public deployment. Despite the serious nature of our work and the hectic pace of the lab around us, our software team maintains a fun and relaxed environment to maximize creativity and learning. If you are a skilled and disciplined software engineer, you will appreciate the level of creativity that this position requires as well as the freedom to incorporate cutting-edge techniques to solve challenging engineering and mathematical problems. Our software buildout cycle is rapid and you will be able to see your work deployed quickly depending on the quality of your code. Though we work as a team, you will largely be expected to be self-motivated, self-directed and able to find and fix problems with minimal supervision. Please be warned -- You must love mathematics as much as you love writing code. If you are not a citizen, be sure to include the status of your visa. The people in the lab are diverse in technical expertise and background. You can read more about us and our work at http://modi.mech.columbia.edu Thank you for considering a position with us. You're welcome to forward this email to anyone who is qualified and interested. The full job description is below. Job Title GIS Systems Optimizer Job Requisition Number 058260 Department 664-EARTH INSTITUTE Location Morningside Job Type Officer Full-Time Regular Job Family Engineering Salary Grade 12 Salary Range Commensurate with experience Advertised Summary Job Description A GIS Systems Optimizer is needed to document, optimize and deploy a remote sensing system on satellite images from twelve sites of the Millennium Villages Project, as well as prepare papers and presentations on the system. To document the performance of the system, the incumbent will design experiments and run the system on satellite images from twelve Millennium Village sites. To optimize the system to scan images faster, s/he will lead a team of developers in rewriting the image recognition code to run on an NVIDIA GPU (graphics processing unit) instead of the standard CPU, a step that has been proven in industry and the literature to accelerate image recognition by several orders of magnitude. To make the system more robust, s/he will design unit tests to ensure that future enhancements to the system will not break existing features. To make the system accessible, s/he will work with our developers on web-enabling and deploying the system on the Amazon EC2 cloud computing platform. The publishing component of the position will consist of submitting papers to journals and giving demonstrations at conferences. The incumbent will also train individuals on how to use and improve the system, record screen casts of the system in use and apply to grants for possible funding of the next stage of the project. Minimum Qualifications for Grade Applicant MUST meet these minimum qualifications to be considered an applicant Bachelor's degree required, preferably in engineering. Minimum of 3 years related experience in research or industry required. Additional Position-Specific Minimum Qualifications Applicant MUST meet these minimum qualifications to be considered an applicant Experience in building a Python-based software system is essential, as is prior work with GIS data and satellite imagery. Proven attention to detail and the ability to prioritize and manage multiple projects simultaneously is a must. Excellent oral and written communication skills are required. Must be able to work in a team and be willing to learn new technologies quickly. Preferred Qualifications Preference will be given to candidates who possess a master's degree in computer science, mathematics or electrical engineering and who have experience with image recognition and cloud deployment. Requisition Open Date 04-16-2010 Quick Link jobs.columbia.edu/applicants/Central?quickFind=118575 EEO Statement Columbia University is an Equal Opportunity/Affirmative Action employer.

10 years, 5 months

Are sin/log supported for complex number (0.94rc)? Odd results...

by Ian Ozsvald

I find myself out of my depth again. I'm playing with complex numbers using 0.94rc (on Windows XP with CUDA 2.3). I've successfully used simple operations (addition, multiplication) on complex numbers, that resulted in the Mandelbrot example in the wiki. Now I'm trying sin and log and I'm getting errors and one positive result (see the end). I'm not sure if these functions are supported yet? If anyone can point me in the right direction (and perhaps suggest what needs implementing) then I'll have a go at fixing the problem. For reference, you can do the following using numpy on the CPU for verification: In [117]: numpy.sin(numpy.array(numpy.complex64(1-1j))) Out[117]: (1.2984576-0.63496387j) Here are two pieces of test code. At first I used multiplication (rather than sin()) to confirm that the code ran as expected. Next I tried creating a simple complex64 array, passing it to the GPU and then asking for pycuda.cumath.sin() - this results in "Error: External calls are not supported". Here's the code and error: ======== import pycuda.driver as drv import pycuda.autoinit import pycuda.gpuarray as gpuarray import pycuda.cumath import numpy a = numpy.array(numpy.complex64(1-1j)) # make list of 1 complex element a_gpu = gpuarray.to_gpu(a) pycuda.cumath.sin(a_gpu) # should produce (1.29-0.63j) print a_gpu.get() ======== In [62]: %run complex_test.py kernel.cu tmpxft_00000840_00000000-3_kernel.cudafe1.gpu tmpxft_00000840_00000000-8_kernel.cudafe2.gpu ./kernel.cu(19): Error: External calls are not supported (found non-inlined call to _ZN6pycuda3sinERKNS_7complexIfEE) --------------------------------------------------------------------------- CompileError Traceback (most recent call last) 16 a = numpy.array(numpy.complex64(1-1j)) # make list of 1 complex element 17 a_gpu = gpuarray.to_gpu(a) ---> 18 pycuda.cumath.sin(a_gpu) # should produce (1.29-0.63j) ======== I replaced: pycuda.cumath.sin(a_gpu) # should produce (1.29-0.63j) with: pycuda.cumath.log(a_gpu) # should produce (0.34-0.78j) and the code ran without an error...but produced the wrong result. It generates (1-1j) which looks like a no-op. Next I tried similar functionality using sin() in an ElementwiseKernel, this generates the same "Error: External calls are not supported" problem: ======== import pycuda.driver as drv import pycuda.autoinit import pycuda.gpuarray as gpuarray import pycuda.cumath import numpy from pycuda.elementwise import ElementwiseKernel complex_gpu = ElementwiseKernel ( "pycuda::complex<float> *z", "z[i] = sin(z[i]);", "complex_fn", "#include <pycuda-complex.hpp>," ) a = numpy.array(numpy.complex64(1-1j)) # make list of 1 complex element a_gpu = gpuarray.to_gpu(a) #sin((1-1j)) should produce (1.29-0.63j) #log((1-1j)) should produce (0.34-0.78j) complex_gpu(a_gpu) print a_gpu.get() ======== In [56]: %run complex_test.py *** compiler output in c:\docume~1\parc\locals~1\temp\tmpjqvxpv kernel.cu kernel.cudafe1.gpu kernel.cudafe2.gpu ./kernel.cu(19): Error: External calls are not supported (found non-inlined call to _ZN6pycuda3sinERKNS_7complexIfEE) --------------------------------------------------------------------------- CompileError Traceback (most recent call last) C:\Panalytical\pycuda_git\pycuda0.94\pycuda\examples\complex_test.py in <module>() 26 "z[i] = sin(z[i]);", 27 "complex_fn", ---> 28 "#include <pycuda-complex.hpp>," 29 ) .... CompileError: nvcc compilation of c:\docume~1\parc\locals~1\temp\tmpjqvxpv\kernel.cu failed [command: nvcc --cubin -arch sm_11 -IC:\Python26\lib\site-packages\pycuda-0.94rc-py2.6-win32.egg\pycuda\..\include\pycud a --keep kernel.cu] WARNING: Failure executing file: <complex_test.py> ======== *However* if I replace sin() with log(): from pycuda.elementwise import ElementwiseKernel complex_gpu = ElementwiseKernel ( "pycuda::complex<float> *z", "z[i] = log(z[i]);", "complex_fn", "#include <pycuda-complex.hpp>," ) then I get the correct result! ======== In [126]: %run complex_test.py (0.346573650837-0.785398185253j) ======== Does anyone know why sin() doesn't work in both cases and log() works in an ElementwiseKernel but not correctly as a pycuda.cumath call? Ian. -- Ian Ozsvald (A.I. researcher, screencaster) ian(a)IanOzsvald.com http://IanOzsvald.com http://morconsulting.com/ http://TheScreencastingHandbook.com http://ProCasts.co.uk/examples.html http://twitter.com/ianozsvald

10 years, 5 months

Re: [PyCUDA] Correct context handling when mixing runtime and (PyCUDA) code?

by Louis Theran

Hi, Brian --- Thanks for the pointer. Just to make sure I understand what's going on, from the CUDA documentation, the correct sequence seems to be: (1) Initialize the runtime, which creates a context. (2) Attach PyCUDA to this context from the runtime (3) Run kernel code (4) Detach Is this right? If so, is there something about PyCUDA that makes it difficult to do this, or is it just a matter of setting up and tearing down the driver instance in the right order? (Another thing I am not totally clear on is what PyCUDA's lifecycle is for managing contexts. Does it leave a single context up as long as the python process runs, or does it tear them down as soon as there aren't any references to gpuarray instances?) I admit to being something of a novice with PyCUDA and CUDA generally, so sorry for pestering the list. ^L On Sun, Apr 18, 2010 at 10:17 AM, Bryan Catanzaro <bryan.catanzaro(a)gmail.com > wrote: > It's not a solution, but the workaround I've been using is to use > context.detach() rather than context.pop() at the end of the > computation. If you look at pycuda.autoinit, you can see what needs > to be done to initialize a CUDA context. Do the same thing manually, > just change the function you register with atexit to be > context.detach, and the errors should go away. > > This is just a workaround, though, and doesn't solve the underlying > problem. > > - bryan > > On Apr 18, 2010, at 6:09 AM, Louis Theran <theran(a)temple.edu> wrote: > > > > >> I'm trying to mix some ctypes wrapped CUDA runtime C libraries with > >> PyCUDA. > >> > >> Even using CUDA 3.0, I am getting errors because PyCUDA seems to be > >> trying to push/pop the current context. > >> > >> From the docs, it's not quite clear what I have to do in this > >> regard. Any help from those of you successfully doing this would be > >> much appreciated. > >> > >> Thanks! > >> > >> ^L > > > > _______________________________________________ > > PyCUDA mailing list > > PyCUDA(a)host304.hostmonster.com > > http://host304.hostmonster.com/mailman/listinfo/pycuda_tiker.net >

10 years, 5 months

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

PyCUDA April 2010