Exception: pygpu/gpuarray_api.h:23:90: error: 'gpucontext_props' has not been declared.

[Paperone80]

I am trying to get Theano 1.0.0 to work with pygpu. Any pygpu v0.7+ results in the following error when using the test script (previous Theano==0.9.0 and pygpu==0.6.1 were working fine):

import os
os.environ['DEVICE'] = 'cuda0'
os.environ['LD_LIBRARY_PATH'] = '/usr/lib:/usr/bin:/usr/local/cuda-8.0:/usr/local/lib:/usr/lib64:/opt/CNTK/cntk/cntk/lib:/opt/CNTK/cntk/cntk/dependencies/lib:/usr/lib64/openmpi/lib:/usr/local/lib:/usr/local/cuda/lib64:/usr/local/cudnn/lib64:/opt/opencv-3.2.0/release/build/lib'
os.environ['THEANO_FLAGS'] = 'cuda.root=/usr/local/cuda-8.0, optimizer=fast_compile, device=cuda0, gpuarray.preallocate=0.1, floatX=float32'
os.environ['THEANO_FLAGS_GPU'] = 'cuda.root=/usr/local/cuda-8.0, optimizer=fast_compile, device=cuda0, gpuarray.preallocate=0.1, floatX=float32'
os.environ['CPATH'] = /usr/local/cudnn/include:/usr/local/cuda/include:/usr/local/cuda/lib64:/usr/local/cudnn/include'`
import theano
import pygpu
from theano import function, config, shared, tensor
import numpy
import time
vlen = 10 * 30 * 768  # 10 x #cores x # threads per core
iters = 1000
rng = numpy.random.RandomState(22)
x = shared(numpy.asarray(rng.rand(vlen), config.floatX))
f = function([], tensor.exp(x))
print(f.maker.fgraph.toposort())
t0 = time.time()
for i in range(iters):
    r = f()
t1 = time.time()
print("Looping %d times took %f seconds" % (iters, t1 - t0))
print("Result is %s" % (r,))
if numpy.any([isinstance(x.op, tensor.Elemwise) and
              ('Gpu' not in type(x.op).__name__)
              for x in f.maker.fgraph.toposort()]):
    print('Used the cpu')
else:
    print('Used the gpu')

Using cuDNN version 6021 on context None Preallocating 1218/12189 Mb (0.100000) on cuda0 Mapped name None to device cuda0: TITAN X (Pascal) (0000:41:00.0)

You can find the C code in this temporary file: /tmp/theano_compilation_error_7o_6szvz library inux-x86_64.egg/pygpu/gpuarray_api.h:23:90: is not found.

Exception Traceback (most recent call last) <ipython-input-2-8efec2400760> in <module>()

     21 rng = numpy.random.RandomState(22)
     22 x = shared(numpy.asarray(rng.rand(vlen), config.floatX))`

---> 23 f = function([], tensor.exp(x))
     24 print(f.maker.fgraph.toposort())
     25 t0 = time.time()

...

Exception: ('The following error happened while compiling the node', GpuElemwise{exp,no_inplace}(<GpuArrayType(float32, vector)>), '\n', "Compilation failed (return status=1): In file included from ~/.theano/compiledir_Linux-3.10-el7.x86_64-x86_64-with-redhat-7.4-Maipo-x86_64-3.5.2-64/tmpzokb0gyu/mod.cpp:10:0:. /opt/intel/intelpython35/lib/python3.5/site-packages/pygpu-0.7.1-py3.5-linux-x86_64.egg/pygpu/gpuarray_api.h:23:90: error: 'gpucontext_props' has not been declared. static struct PyGpuContextObject (__pyx_api_f_5pygpu_8gpuarray_pygpu_init)(PyObject , gpucontext_props ) = 0;. ^. ", '[GpuElemwise{exp,no_inplace}(<GpuArrayType(float32, vector)>)]')

import os
os.environ['DEVICE'] = 'cuda0'
import pygpu
from pygpu import gpuarray
from pygpu.elemwise import compare, elemwise2

pygpu.test()

pygpu is installed in /opt/intel/intelpython35/lib/python3.5/site-packages/pygpu-0.7.1-py3.5-linux-x86_64.egg/pygpu NumPy version 1.13.3 NumPy relaxed strides checking option: True NumPy is installed in /opt/intel/intelpython35/lib/python3.5/site-packages/numpy Python version 3.5.2 |Intel Corporation| (default, Oct 20 2016, 03:10:33) [GCC 4.8.2 20140120 (Red Hat 4.8.2-15)] nose version 1.3.7

4 errors due 'Could not load "libnccl.so": libnccl.so: cannot open shared object file: No such file or directory'

Any ideas why this error keeps happening and how to resolve? Thanks a lot!

[abergeron]

You need to keep the installs of pygpu and libgpuarray in sync. This looks like it's using the old 0.6.X headers with a newer Theano.

As for the nccl error, you can probably ignore that one unless you are trying to use multi-gpu.

[Paperone80]

What's the best way of keeping the headers in sync? I used to compile libgpuarry myself and install theano via pip... Did work for theano 0.9 and pygpu 0.6.1. Didn't work when I used the same approach for pygpu 0.75 (later 0.71) and theano 1.0.0.

Tried deleting older files under /usr/local/lib and in the python directory but not sure where else older files/headers are floating around. Any help is more than appreciated. Thanks.

[abergeron]

What you describe should be ok. How are you compiling libgpuarray/pygpu? Can you post the exact commands you ran?

If you want an easy way out, you can use anaconda/miniconda. Then you can install Theano/pygpu/libgpuarray with conda install -c mila-udem theano pygpu, or just install pygpu with it and Theano via pip.

[nouiz]

closing as no news.