def test_consistency_GPU_parallel():
    """
    Verify that the random numbers generated by GPU_mrg_uniform, in
    parallel, are the same as the reference (Java) implementation by
    L'Ecuyer et al.

    """
    if not cuda_available:
        raise SkipTest('Optional package cuda not available')
    if config.mode == 'FAST_COMPILE':
        mode = 'FAST_RUN'
    else:
        mode = config.mode

    seed = 12345
    n_samples = 5
    n_streams = 12
    n_substreams = 7  # 7 samples will be drawn in parallel

    samples = []
    curr_rstate = numpy.array([seed] * 6, dtype='int32')

    for i in range(n_streams):
        stream_samples = []
        rstate = [curr_rstate.copy()]
        for j in range(1, n_substreams):
            rstate.append(rng_mrg.ff_2p72(rstate[-1]))
        rstate = numpy.asarray(rstate).flatten()
        # HACK - transfer these int32 to the GPU memory as float32
        # (reinterpret_cast)
        tmp_float_buf = numpy.frombuffer(rstate.data, dtype='float32')
        # Transfer to device
        rstate = float32_shared_constructor(tmp_float_buf)

        new_rstate, sample = rng_mrg.GPU_mrg_uniform.new(rstate, ndim=None,
                                                         dtype='float32',
                                                         size=(n_substreams,))
        rstate.default_update = new_rstate

        # Not really necessary, just mimicking
        # rng_mrg.MRG_RandomStreams' behavior
        sample.rstate = rstate
        sample.update = (rstate, new_rstate)

        # We need the sample back in the main memory
        cpu_sample = tensor.as_tensor_variable(sample)
        f = theano.function([], cpu_sample, mode=mode)

        for k in range(n_samples):
            s = f()
            stream_samples.append(s)

        samples.append(numpy.array(stream_samples).T.flatten())

        # next stream
        curr_rstate = rng_mrg.ff_2p134(curr_rstate)

    samples = numpy.array(samples).flatten()
    assert(numpy.allclose(samples, java_samples))

def test_consistency_GPU_serial():
    """Verify that the random numbers generated by GPU_mrg_uniform, serially,
    are the same as the reference (Java) implementation by L'Ecuyer et al.
    """
    if not cuda_available:
        raise SkipTest("Optional package cuda not available")
    if config.mode == "FAST_COMPILE":
        mode = "FAST_RUN"
    else:
        mode = config.mode

    seed = 12345
    n_samples = 5
    n_streams = 12
    n_substreams = 7

    samples = []
    curr_rstate = numpy.array([seed] * 6, dtype="int32")

    for i in range(n_streams):
        stream_rstate = curr_rstate.copy()
        for j in range(n_substreams):
            substream_rstate = numpy.array(stream_rstate.copy(), dtype="int32")
            # HACK - we transfer these int32 to the GPU memory as float32
            # (reinterpret_cast)
            tmp_float_buf = numpy.frombuffer(substream_rstate.data, dtype="float32")
            rstate = float32_shared_constructor(tmp_float_buf)  # Transfer to device

            new_rstate, sample = rng_mrg.GPU_mrg_uniform.new(rstate, ndim=None, dtype="float32", size=(1,))
            rstate.default_update = new_rstate

            # Not really necessary, just mimicking rng_mrg.MRG_RandomStreams' behavior
            sample.rstate = rstate
            sample.update = (rstate, new_rstate)

            # We need the sample back in the main memory
            cpu_sample = tensor.as_tensor_variable(sample)
            f = theano.function([], cpu_sample, mode=mode)
            for k in range(n_samples):
                s = f()
                samples.append(s)

            # next substream
            stream_rstate = rng_mrg.ff_2p72(stream_rstate)

        # next stream
        curr_rstate = rng_mrg.ff_2p134(curr_rstate)

    samples = numpy.array(samples).flatten()
    assert numpy.allclose(samples, java_samples)