PyCUDA January 2011

pycuda@tiker.net

13 participants
16 discussions

Cannot import both pycuda and pyopencl in the same program

by Bogdan Opanchuk

Hi all, I'm observing the following behavior with latest (git-fetched today) pycuda and opencl versions on Snow Leopard 10.6.4: $ python >>> import pycuda.driver >>> import pyopencl Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/Library/Python/2.6/site-packages/pyopencl-0.92beta-py2.6-macosx-10.6-i386.egg/pyopencl/__init__.py", line 3, in <module> import pyopencl._cl as _cl AttributeError: 'NoneType' object has no attribute '__dict__' $ python >>> import pyopencl >>> import pycuda.driver Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/Library/Python/2.6/site-packages/pycuda-0.94rc-py2.6-macosx-10.6-i386.egg/pycuda/driver.py", line 1, in <module> from _driver import * AttributeError: 'NoneType' object has no attribute '__dict__' This worked with stable versions. Does anyone know why is this happening? (One may ask why do I need both libraries in the same program. I have the set of tests for my module, which can use both Cuda and OpenCL, and it is convenient to run all the tests using the single file. Although it is not a critical issue, I'm just curious). Best regards, Bogdan

10 years

pycuda install help

by nhieu duong

hi, ive been endlessly trying to install pycuda on a red hat dist. machine, but to no avail. it would be much appreciated if i could get some help. i am able to get past the configure part of the installation, but the when i "make" , the problem occurs. here is my siteconf.py file BOOST_INC_DIR = ['/usr/local/include/boost/'] BOOST_LIB_DIR = ['/usr/lib'] BOOST_COMPILER = 'gcc4.1.2' BOOST_PYTHON_LIBNAME = ['boost_python'] BOOST_THREAD_LIBNAME = ['boost_thread'] CUDA_TRACE = False CUDA_ROOT = '/usr/local/cuda/' CUDA_ENABLE_GL = False CUDADRV_LIB_DIR = ['/usr/lib'] CUDADRV_LIBNAME = ['cuda'] CXXFLAGS = ['-DBOOST_PYTHON_NO_PY_SIGNATURES'] LDFLAGS = [] i beleive i built boost with gcc version 4.1.2 the error im getting is..... /usr/local/include/boost/type_traits/remove_const.hpp:61: instantiated from ‘boost::remove_const<<unnamed>::pooled_host_allocation>’ /usr/local/include/boost/python/object/pointer_holder.hpp:127: instantiated from ‘void* boost::python::objects::pointer_holder<Pointer, Value>::holds(boost::python::type_info, bool) [with Pointer = std::auto_ptr<<unnamed>::pooled_host_allocation>, Value = <unnamed>::pooled_host_allocation]’ src/wrapper/mempool.cpp:278: instantiated from here /usr/local/include/boost/type_traits/detail/cv_traits_impl.hpp:38: internal compiler error: in make_rtl_for_nonlocal_decl, at cp/decl.c:5067 i only included the ends. if you want the entire thing let me know. but the error seems to point to a gcc problem. ive read thru your archives but doesnt seem to solve this problem if someone could shed some light on this issue, i would very appreciate it. thanks -nhieu

10 years, 3 months

SparseSolve.py example

by elafrit

Hello - I'm trying to run the SparseSolve.py example. I installed PyMetis package after fixing the configuration like here : ./configure --python-exe=python2.6 --boost-inc-dir=/usr/include/boost --boost-lib-dir=/usr/lib/ --boost-python-libname=boost_python-mt-py26 But when running the SparseSolve.py example I encountred this error : ImportError: /usr/local/lib/python2.6/dist-packages/PyMetis-0.91-py2.6-linux-x86_64.egg/pymetis/_internal.so: undefined symbol: regerrorA What does this error means? thanks for any suggestions. -- View this message in context: http://pycuda.2962900.n2.nabble.com/PyCUDA-SparseSolve-py-example-tp5969743… Sent from the PyCuda mailing list archive at Nabble.com.

10 years, 4 months

compiler.get_nvcc_version() problems with mpi4py/mpich2 1.3

by avidday

I have hit a wall moving some existing pycuda code to a distributed memory cluster and am hoping someone cleverer than I can suggest a work around. In a nutshell, I have found that get_nvcc_version doesn't work in the way the pyCUDA.compiler module expects with the MPI flavours I use under certain circumstances, which makes behind the scenes JIT compilation inside pyCUDA fail if all the MPI processes don't share a common /tmp filesystem. The simplest repro case is this snippet: ---------- import sys from pycuda import compiler from mpi4py import MPI rank = MPI.COMM_WORLD.Get_rank() sys.stdout.write("[%d] %s\n" %(rank, compiler.get_nvcc_version("nvcc"))) ---------- which will do this when run on a single node $ mpiexec -n 4 python ./pycudachk.py [2] None [3] None [1] None [0] nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2010 NVIDIA Corporation Built on Wed_Nov__3_16:16:57_PDT_2010 Cuda compilation tools, release 3.2, V0.2.1221 I presume this is because the pytools stdout/stderr capture routines don't get anything on MPI processes where stdout/stderr are being managed by the MPI runtime. If all the MPI processes share a common /tmp filesystem, it doesn't seem to matter because the pycuda compiler cache is visible to each process and it all just works. But if they don't (and the cluster I am now using has local /tmp on every node), the compiler module will need to JIT compile stuff locally on each node, and it winds up failing because get_nvcc_version returns None, which in turn makes the md5 hashing calls fail. Something like this: [avidday@n0005 fim]$ mpiexec -n 2 -hosts n0008,n0005 ./fimMPItest.py [1] mpi dims = (2,1), coords = (1,0) [0] mpi dims = (2,1), coords = (0,0) [1]{n0005} CUDA driver GeForce GTX 275, 1791Mb ram, using fim_sm13.cubin [0]{n0008} CUDA driver GeForce GTX 275, 1791Mb ram, using fim_sm13.cubin Traceback (most recent call last): File "./fimMPItest.py", line 20, in ? phi,its = dotest() File "./fimMPItest.py", line 16, in dotest return fim.fimMPIScatter2(gs,a,1.,h,maxiters=1000,tol=1e-6,CC=fimcuda.fimCC) File "/scratch/fim/fim.py", line 464, in fimMPIScatter2 its = mpiCC.Iterate(f, h, tol=tol, maxiters=maxiters) File "/scratch/fim/fimcuda.py", line 187, in Iterate self.active_.fill(np.int32(0)) File "/usr/lib64/python2.4/site-packages/pycuda/gpuarray.py", line 336, in fill File "<string>", line 1, in <lambda> File "/usr/lib64/python2.4/site-packages/pycuda/tools.py", line 485, in context_dependent_memoize File "/usr/lib64/python2.4/site-packages/pycuda/elementwise.py", line 384, in get_fill_kernel File "/usr/lib64/python2.4/site-packages/pycuda/elementwise.py", line 98, in get_elwise_kernel File "/usr/lib64/python2.4/site-packages/pycuda/elementwise.py", line 85, in get_elwise_kernel_and_types File "/usr/lib64/python2.4/site-packages/pycuda/elementwise.py", line 74, in get_elwise_module File "/usr/lib64/python2.4/site-packages/pycuda/compiler.py", line 238, in __init__ File "/usr/lib64/python2.4/site-packages/pycuda/compiler.py", line 228, in compile File "/usr/lib64/python2.4/site-packages/pycuda/compiler.py", line 47, in compile_plain TypeError: update() argument 1 must be string or read-only buffer, not None which I interpret as meaning that an internal compile to satisfy a gpuarray fill() call is failing. Running on only a single node works fine. If I am reading things correcly, it looks like this checksum code in compile_plain fails because of what get_nvcc_version returns 41 if cache_dir: 42 checksum = _new_md5() 43 44 checksum.update(source) 45 for option in options: 46 checksum.update(option) 47 checksum.update(get_nvcc_version(nvcc)) The question then becomes how to fix it? It is important to note that nvcc is available to all process and works, so I assume that the fork itself is fine (my reading of the code says that an OSError exception would be raised otherwise). So I am guessing the problem is only that get_nvcc_version can return None even when the fork worked. Would it be too hackish to have get_nvcc_version return something generic like "nvcc unknown version" or something that would still hash ok in the case where the fork worked but the captured output from the fork is not available? Suggestions or corrections to my analysis are greatly welcomed (I know almost nothing about python, pyCUDA or MPI so be gentle if I am completely off target here.....)

10 years, 5 months

Problems compling

by Bryan Absher

hello all, I have heard great things about the pycuda, if only it would run. Here is the output to the test_driver.py $ python ./test_driver.py Traceback (most recent call last): File "./test_driver.py", line 4, in <module> from pycuda.tools import mark_cuda_test File "/usr/local/lib/python2.6/dist-packages/pycuda-0.94.1-py2.6-linux-x86_64.egg/pycuda/tools.py", line 30, in <module> import pycuda.driver as cuda File "/usr/local/lib/python2.6/dist-packages/pycuda-0.94.1-py2.6-linux-x86_64.egg/pycuda/driver.py", line 1, in <module> from pycuda._driver import * ImportError: /usr/local/lib/python2.6/dist-packages/pycuda-0.94.1-py2.6-linux-x86_64.egg/pycuda/_driver.so: undefined symbol: cuMemAllocPitch_v2 $ ldd ./_driver.so linux-vdso.so.1 => (0x00007fff80f56000) libcuda.so.1 => /usr/lib/nvidia-current/libcuda.so.1 (0x00007fc378457000) libstdc++.so.6 => /usr/lib/libstdc++.so.6 (0x00007fc378143000) libm.so.6 => /lib/libm.so.6 (0x00007fc377ebf000) libgcc_s.so.1 => /lib/libgcc_s.so.1 (0x00007fc377ca8000) libpthread.so.0 => /lib/libpthread.so.0 (0x00007fc377a8b000) libc.so.6 => /lib/libc.so.6 (0x00007fc377707000) libz.so.1 => /lib/libz.so.1 (0x00007fc3774f0000) libdl.so.2 => /lib/libdl.so.2 (0x00007fc3772ec000) /lib64/ld-linux-x86-64.so.2 (0x00007fc379030000) There were a couple of warnings when compiling about undefined functions similar to this: gcc -pthread -fno-strict-aliasing -fwrapv -Wall -O3 -DNDEBUG -fPIC -DBOOST_MULTI_INDEX_DISABLE_SERIALIZATION=1 -Isrc/cpp -Ibpl-subset/bpl_subset -I/usr/local/cuda/include -I/usr/lib/python2.6/dist-packages/numpy/core/include -I/usr/include/python2.6 -c bpl-subset/bpl_subset/libs/python/src/converter/from_python.cpp -o build/temp.linux-x86_64-2.6/bpl-subset/bpl_subset/libs/python/src/converter/from_python.o bpl-subset/bpl_subset/libs/python/src/converter/from_python.cpp: In function ‘boost::python::converter::rvalue_from_python_stage1_data boost::python::converter::rvalue_from_python_stage1(PyObject*, const boost::python::converter::registration&)’: bpl-subset/bpl_subset/libs/python/src/converter/from_python.cpp:42: warning: ‘data.boost::python::converter::rvalue_from_python_stage1_data::construct’ may be used uninitialized in this function I did make sure that the boost libraries are installed but it kept giving me an error about failing to search for the boost libraries. I think there may be a bug in the way the setup.py search to check for the boost libraries. Any help would be much appreciated! I am looking forward to playing around with pycuda.

10 years, 5 months

question re complex pow

by Lev Givon

I recently attempted to run the following code with CUDA 3.2 and Pycuda 0.94.2 on a Quadro NVS 290 installed on a Linux x86_64 system: import pycuda.gpuarray as gpuarray import pycuda.driver as drv import pycuda.autoinit import numpy as np from pycuda.compiler import SourceModule func_mod = SourceModule(""" #include <pycuda/pycuda-complex.hpp> #define TYPE pycuda::complex<float> __global__ void func(TYPE *a, TYPE *b, int N) { int idx = threadIdx.x; if (idx < N) b[idx] = pow(a[idx], 2); } """) func = func_mod.get_function("func") N = 10 a = np.complex64(np.random.rand(N)+np.random.rand(N)*1j) b = np.complex64(np.zeros(N)) func(drv.In(a), drv.Out(b), np.uint32(N), block=(512,1,1)) print 'in: ', a print 'out (cuda): ', b print 'out (np): ', a**2 When I did so, I observed the following error: pytools.prefork.ExecError: error invoking 'nvcc --cubin -arch sm_11 -I/usr/lib64/python2.6/site-packages/pycuda/../../../../include/pycuda kernel.cu': status 2 invoking 'nvcc --cubin -arch sm_11 -I/usr/lib64/python2.6/site-packages/pycuda/../../../../include/pycuda kernel.cu': ./kernel.cu(10): Error: External calls are not supported (found non-inlined call to _ZN6pycuda3powERKNS_7complexIfEEi) Casting the exponent to a float or pycuda::complex<float> prevents the error from occurring, but casting it to int does not. Is this expected? L.G.

10 years, 5 months

Building PyOpenCL for CPU only

by Riaan van den Dool

Andreas Would it be possible at all to create a version of PyOpenCL that does not link to and does not need any proprietary drivers? Maybe via a make option. I think it will help adoption of pyOpenCL if code that is written using pyopencl will 'always run' even if no GPU is available. If there is already such a version, please excuse my ignorance and point me in the right direction. Riaan

10 years, 5 months

CURAND wrappers - Long pause after generation of the a random matrix

by Martin Laprise

Hi, I just made some experiments with the CURAND wrappers. It seem to work very nicely except for a little detail that I can't figure out. The initialization of the generator and the actual random number generation seem very fast. But for what ever reason, PyCUDA take a long time to "recover" after the number generation. This pause is significantly longer than the actual computation and the delay increase with N. Here is an example: import numpy as np import pycuda.autoinit import pycuda.gpuarray from pycuda.curandom import PseudoRandomNumberGenerator, QuasiRandomNumberGenerator import cProfile import time as clock def curand_prof(): N = 100000000 t1 = clock.time() # GPU rr = PseudoRandomNumberGenerator(0, np.random.random(128).astype(np.int32)) data = pycuda.gpuarray.zeros([N], np.float32) rr.fill_normal_float(data.gpudata, N) t2 = clock.time() print "Bench 1: " + str(t2-t1) + " sec" if __name__ == "__main__": t1 = clock.time() curand_prof() t2 = clock.time() print "Bench 2: " + str(t2-t1) + " sec" Here is the actual output with a GTX 260 gpu: Bench 1: 0.0117599964142 sec Bench 2: 4.40562295914 sec In the example, the pause have no consequence, but if I want to use the random matrix in an other kernel ... it's quite a delay. I've made some research and my guess is that the problem is linked to this already reported problem here: http://forums.nvidia.com/index.php?showtopic=185740 Anyone knows how we can implement the solution to the wrapper ? Martin

10 years, 6 months

CURAND wrappers

by Andreas Kloeckner

On Mon, 20 Dec 2010 21:41:16 +0100, Tomasz Rybak <bogomips(a)post.pl> wrote: > At the same time - could you look into CURAND patch I have sent > to the list (attached here)? Last email I have sent on 2010-12-15 22:06 > I would like to finish it and then finish prefix scan. I've taken a look at your CURAND code, here are a few comments: - The user should not *have* to specify generator_count. Instead, we should supply a reasonable default based on the device's compute capability, as you describe in the docs. (Likewise, the docs don't need to be redundant.) - I don't like the name "Randomizer". "RandomNumberGenerator" is long, but IMO a better name. - What's the difference between the quasi- and non-quasi versions? It looks like there's a ton of duplicated code between the two. This should be eliminated, perhaps by inheritance or through another way. - Tests should go in tests/test_gpuarray.py. - Rename fill_in_* to fill_*. Thanks for your contribution! Looking forward to your comments. Andreas

10 years, 6 months

pyfft on large 3d arrays

by Saigopal Nelaturi

Hello all, I am implementing a simple 3d convolution on the gpu using pyfft. The basic idea is straightforward - obtain the 3d Fourier transform for each array, multiply and take the inverse transform of the product. I am using pyfft for the implementation. The code below works correctly when my input array is 256^3 but fails (executes but gives garbage results) for a 512^3 voxel grid. # w,h,k are the array dimensions in a power of 2 # im1, im2 are the input 3d arrays of dtype complex64 plan = Plan((w,h,k), normalize=True) # forward transform on device im1_gpu = gpuarray.to_gpu(im1) plan.execute(im1_gpu) im1_ft = im1_gpu.get() del im1_gpu im2_gpu = gpuarray.to_gpu(im2) plan.execute(im2_gpu) im2_ft = im2_gpu.get() del im2_gpu # do multiplication on host - can be done on device. conv = im1_ft * im2_ft #inverse transform on device conv_gpu = gpuarray.to_gpu(conv) plan.execute(conv_gpu, inverse='True') corr = conv_gpu.get() I don't think there's anything wrong with the code (it works for smaller array sizes) as such but I am perplexed as to why the failure occurs. I am running the code on a Tesla C2050 (2.8GB available memory) and so there's enough space to hold the 512^3 array with complex64 dtype. Does anyone have an explanation? -Saigopal

10 years, 6 months

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

PyCUDA January 2011