multiply.py

input = numpy.array([data[0] for data in input_data])

mf = cl.mem_flags
temp = cl.Buffer(self.opencl_context, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=input) 

a_g = temp

work_size = len(input_data)
res = None

for layer in layers:
    params = [1,layer.shape[0],layer.shape[1]]

    b_g = cl.Buffer(self.opencl_context, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=layer) #layer.view(dtype=numpy.float32)

    res = numpy.empty([work_size, params[2]], dtype=numpy.float64) #dtype=numpy.float32
    c_g = cl.Buffer(self.opencl_context, mf.COPY_HOST_PTR, hostbuf=res)
    params_buffer = cl.Buffer(self.opencl_context, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=numpy.array(params))

    self.opencl_program.multiply(self.opencl_queue, (work_size,), None, params_buffer, a_g, b_g, c_g)
    a_g = c_g

cl.enqueue_copy(opencl_queue, res, a_g)