How to use Theano? (with OpenCL)

I have a MacBook Pro with AMD display card, so I want Theano use OpenCL. First, create virtualenv and activate it:

python3 -m venv venv
source venv/bin/activate

Install some dependencies for Theano and OpenCL:

pip install cython nose

those dependencies isn’t in setup.py of Theano (or pygpu).

Install Theano (latest development version, so far, 0.8.2 doesn’t support OpenCL)：

pip install git+https://github.com/Theano/Theano.git

If github is temporary unavailable, try my personal csdn mirror(CSDN CODE can sync github repo conveniently):

pip install git+https://code.csdn.net/u010096836/theano.git

Next step is installing gpuarray for supporting OpenCL:

git clone https://github.com/Theano/libgpuarray.git
cd libgpuarray

build native part

cmake . -DCMAKE_INSTALL_PREFIX=../venv/ -DCMAKE_BUILD_TYPE=Release
make
make install

export some env for Theano’s dynamic compilation

export LIBRARY_PATH=$LIBRARY_PATH:$PWD/../venv/lib
export CPATH=$CPATH:$PWD/../venv/

install pygpu/gpuarray

python setup.py build
python setup.py install

Now, Theano can use OpenCL to accelerate computing.

---

We can use Theano’s check1.py to check that OpenCL is available:

from theano import function, config, shared, tensor, sandbox
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')

Performance:

• Only use CPU (1.662902 s)

$ THEANO_FLAGS=mode=FAST_RUN,device=cpu,floatX=float32 python check1.py 
[Elemwise{exp,no_inplace}(<TensorType(float32, vector)>)]
Looping 1000 times took 1.662902 seconds
Result is [ 1.23178029  1.61879337  1.52278066 ...,  2.20771813  2.29967761
  1.62323284]

Using the cpu

• OpenCL over CPU (1.057008 s)

$ THEANO_FLAGS=mode=FAST_RUN,device=opencl0:0,floatX=float32 python check1.py 
Mapped name None to device opencl0:0: Intel(R) Core(TM) i7-4870HQ CPU @ 2.50GHz
[GpuElemwise{exp,no_inplace}(<GpuArrayType<None>(float32, (False,))>), HostFromGpu(gpuarray)(GpuElemwise{exp,no_inplace}.0)]
Looping 1000 times took 1.057008 seconds
Result is [ 1.23178029  1.61879325  1.52278078 ...,  2.20771813  2.29967737
  1.62323272]

Using the cpu

• OpenCL over Intel GPU (0.554572 s)

$ THEANO_FLAGS=mode=FAST_RUN,device=opencl0:1,floatX=float32 python check1.py 
Mapped name None to device opencl0:1: Iris Pro
[GpuElemwise{exp,no_inplace}(<GpuArrayType<None>(float32, (False,))>), HostFromGpu(gpuarray)(GpuElemwise{exp,no_inplace}.0)]
Looping 1000 times took 0.554572 seconds
Result is [ 1.23178029  1.61879337  1.52278066 ...,  2.20771813  2.29967761
  1.62323284]

Using the Intel gpu

• OpenCL over AMD GPU (0.470640 s)

$ THEANO_FLAGS=mode=FAST_RUN,device=opencl0:2,floatX=float32 python check1.py 
Mapped name None to device opencl0:2: AMD Radeon R9 M370X Compute Engine
[GpuElemwise{exp,no_inplace}(<GpuArrayType<None>(float32, (False,))>), HostFromGpu(gpuarray)(GpuElemwise{exp,no\_inplace}.0)\]
Looping 1000 times took 0.470640 seconds
Result is [ 1.23178029  1.61879337  1.52278066 ...,  2.20771813  2.29967761
  1.62323284]

Using the AMD gpu

---

If you using stable(0.8.2) Theano, you will meet this error when you try to use OpenCL:

RuntimeError: ('Wrong major API version for gpuarray:', -9998, 'Make sure Theano and libgpuarray/pygpu are in sync.')

so, you should install development version of Theano.