def test_gpujoin_preserves_broadcasting():
_a = numpy.asarray([[1, 2], [3, 4]], dtype="float32")
_b = numpy.asarray([[5, 6, 7], [8, 9, 10]], dtype="float32")
a = tcn.shared_constructor(_a)
b = tcn.shared_constructor(_b)
# [0,0] : the two original dims were non-broadcastable
# [1,x,0]: new order and broadcastability
gpu_dimshuffle = GpuDimShuffle([0, 0], [1, "x", 0])
a_shuffled = gpu_dimshuffle(a)
b_shuffled = gpu_dimshuffle(b)
c = gpu_join(0, a_shuffled, b_shuffled)
assert c.type.broadcastable == (False, True, False)
f = theano.function([], c, mode=mode_with_gpu)
res = f()
a_reshaped = numpy.asarray([[[1, 3]], [[2, 4]]], dtype="float32")
b_reshaped = numpy.asarray([[[5, 8]], [[6, 9]], [[7, 10]]], dtype="float32")
concat = numpy.concatenate([a_reshaped, b_reshaped], axis=0)
assert numpy.all(res == concat)
def local_gpu_join(node):
"""
Inspired by the opt for convop.
Very loose notation follows.
Subgraphs concerned first look like
[array of HostTensor] -> HostToGpu -> GpuToHost
-> Join -> HostToGpu -> GpuToHost
First we apply this Opt:
join(host_from_gpu) -> host_from_gpu(gpu_join)
then, as an intermediate result, there should be
host_from_gpu(gpu_join) -> HostToGpu -> GpuToHost
this unnecessary GpuToHost -> HostToGpu should be removed
by other opts, leaving us with
host_from_gpu(gpu_join)
For intermediate places in the graph not covered by the first opt, the following could be useful:
gpu_from_host(join) -> gpu_join(gpu_from_host)
not implemented yet.
"""
if isinstance(node.op, tensor.Join):
# optimizing this case:
# join(host_from_gpu) -> host_from_gpu(gpu_join)
# print "OPT: we've got a Join instance"
axis_and_tensors = node.inputs
# print "OPT: axis_and_tensors=", axis_and_tensors
matches = [not t.owner is None and t.owner.op == host_from_gpu for t in axis_and_tensors[1:]]
# print "OPT: matches =", matches
# if all input tensors are host_from_gpu'ified
if numpy.all(matches):
# the extra gpu_from_host introduced here will
# be removed by further optimizations
new_tensors = [gpu_from_host(t) for t in axis_and_tensors[1:]]
new_a_and_t = [axis_and_tensors[0]] + new_tensors
replacement_node = host_from_gpu(gpu_join(*new_a_and_t))
# print "OPT: replacement_node", replacement_node
return [replacement_node]
def test_gpujoin_twomatrices_joincolumns():
_a = numpy.asarray([[1, 2], [3, 4]], dtype="float32")
_b = numpy.asarray([[5, 6, 7], [8, 9, 10]], dtype="float32")
a = tcn.shared_constructor(_a)
b = tcn.shared_constructor(_b)
c = gpu_join(1, a, b)
f = theano.function([], c)
assert numpy.all(f() == numpy.concatenate([_a, _b], axis=1))
def test_gpujoin_assert_cndas():
# this will end up being an ndarray, as it's float64
_a = numpy.asarray([[1, 2], [3, 4]], dtype="float64")
a = theano.shared(_a)
try:
c = gpu_join(1, a)
# can't "assert False" here, as we want the assertion
# error from gpu_join
except AssertionError:
assert True
return
assert False
def test_gpujoin_twomatrices_badshapes():
_a = numpy.asarray([[1, 2], [3, 4]], dtype="float32")
_b = numpy.asarray([[5, 6, 7], [8, 9, 10]], dtype="float32")
a = tcn.shared_constructor(_a)
b = tcn.shared_constructor(_b)
# try to join on dimension 0 where they don't agree (2!=3)
c = gpu_join(0, a, b)
f = theano.function([], c)
try:
f()
assert False
except ValueError:
assert True
