def test_consistency_GPUA_parallel():
# Verify that the random numbers generated by GPUA_mrg_uniform, in
# parallel, are the same as the reference (Java) implementation by
# L'Ecuyer et al.
seed = 12345
n_samples = 5
n_streams = 12
n_substreams = 7 # 7 samples will be drawn in parallel
samples = []
curr_rstate = np.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 = np.asarray(rstate)
rstate = gpuarray_shared_constructor(rstate)
new_rstate, sample = GPUA_mrg_uniform.new(rstate,
ndim=None,
dtype="float32",
size=(n_substreams, ))
rstate.default_update = new_rstate
# Not really necessary, just mimicking
# rng_mrg.MRG_RandomStream' behavior
sample.rstate = rstate
sample.update = (rstate, new_rstate)
# We need the sample back in the main memory
cpu_sample = at.as_tensor_variable(sample)
f = aesara.function([], cpu_sample, mode=mode)
for k in range(n_samples):
s = f()
stream_samples.append(s)
samples.append(np.array(stream_samples).T.flatten())
# next stream
curr_rstate = rng_mrg.ff_2p134(curr_rstate)
samples = np.array(samples).flatten()
assert np.allclose(samples, java_samples)
def test_consistency_GPUA_serial():
# Verify that the random numbers generated by GPUA_mrg_uniform, serially,
# are the same as the reference (Java) implementation by L'Ecuyer et al.
seed = 12345
n_samples = 5
n_streams = 12
n_substreams = 7
samples = []
curr_rstate = np.array([seed] * 6, dtype="int32")
for i in range(n_streams):
stream_rstate = curr_rstate.copy()
for j in range(n_substreams):
substream_rstate = np.array([stream_rstate.copy()], dtype="int32")
# Transfer to device
rstate = gpuarray_shared_constructor(substream_rstate)
new_rstate, sample = GPUA_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 = aesara.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 = np.array(samples).flatten()
assert np.allclose(samples, java_samples)
def test_GPUA_full_fill():
# Make sure the whole sample buffer is filled. Also make sure
# large samples are consistent with CPU results.
# This needs to be large to trigger the problem on GPU
size = (10, 1000)
R = MRG_RandomStream(234)
uni = R.uniform(size, nstreams=60 * 256)
f_cpu = aesara.function([], uni)
rstate_gpu = gpuarray_shared_constructor(
R.state_updates[-1][0].get_value())
new_rstate, sample = GPUA_mrg_uniform.new(rstate_gpu,
ndim=None,
dtype="float32",
size=size)
rstate_gpu.default_update = new_rstate
f_gpu = aesara.function([], sample, mode=mode)
utt.assert_allclose(f_cpu(), f_gpu())