def test_local_gpualloc_memset_0():
i = theano.tensor.iscalar()
z = numpy.zeros((1,), dtype='float32')
o = numpy.ones((1,), dtype='float32')
ones = numpy.ones((2,), dtype='float32')
# Test with 0
a = gpu_alloc(z, i)
f = theano.function([i], a, mode=mode_with_gpu)
topo = f.maker.fgraph.toposort()
assert len(topo) == 1
assert isinstance(topo[0].op, GpuAlloc) and topo[0].op.memset_0
assert (numpy.asarray(f(6)) == 0).all()
# Test with 1
a = gpu_alloc(o, i)
f = theano.function([i], a, mode=mode_with_gpu)
topo = f.maker.fgraph.toposort()
assert len(topo) == 1
assert isinstance(topo[0].op, GpuAlloc)
assert not topo[0].op.memset_0
assert (numpy.asarray(f(6)) == 1).all()
# Test with 1, 1
a = gpu_alloc(ones, i)
f = theano.function([i], a, mode=mode_with_gpu)
topo = f.maker.fgraph.toposort()
assert len(topo) == 1
assert isinstance(topo[0].op, GpuAlloc)
assert not topo[0].op.memset_0
assert (numpy.asarray(f(2)) == 1).all()
def test_local_gpualloc_memset_0():
i = theano.tensor.iscalar()
z = numpy.zeros((1, ), dtype='float32')
o = numpy.ones((1, ), dtype='float32')
ones = numpy.ones((2, ), dtype='float32')
# Test with 0
a = gpu_alloc(z, i)
f = theano.function([i], a, mode=mode_with_gpu)
topo = f.maker.fgraph.toposort()
assert len(topo) == 1
assert isinstance(topo[0].op, GpuAlloc) and topo[0].op.memset_0
assert (numpy.asarray(f(6)) == 0).all()
# Test with 1
a = gpu_alloc(o, i)
f = theano.function([i], a, mode=mode_with_gpu)
topo = f.maker.fgraph.toposort()
assert len(topo) == 1
assert isinstance(topo[0].op, GpuAlloc)
assert not topo[0].op.memset_0
assert (numpy.asarray(f(6)) == 1).all()
# Test with 1, 1
a = gpu_alloc(ones, i)
f = theano.function([i], a, mode=mode_with_gpu)
topo = f.maker.fgraph.toposort()
assert len(topo) == 1
assert isinstance(topo[0].op, GpuAlloc)
assert not topo[0].op.memset_0
assert (numpy.asarray(f(2)) == 1).all()
def local_gpuaalloc(node):
new_out = gpu_alloc(*node.inputs)
# We need to hide new broadcastable dimensions because
# ReplaceValidate doesn't like when they change.
if new_out.broadcastable != node.outputs[0].broadcastable:
# but if a dim is suddenly not broadcastable anymore then that's a bug
for b_old, b_new in zip(node.outputs[0].broadcastable, new_out.broadcastable):
assert b_new or (not b_old)
new_out = tensor.patternbroadcast(new_out, node.outputs[0].broadcastable)
return (new_out,)
def local_gpuaalloc2(node):
"""
Join(axis, Alloc, Alloc, ...) -> Join(axis, GpuAlloc, Alloc, ...)
Moves an alloc that is an input to join to the gpu.
"""
if (isinstance(node.op, tensor.Alloc)
and all(c != 'output' and c.op == tensor.join and all(
i.owner and i.owner.op in [host_from_gpu, tensor.alloc]
for i in c.inputs[1:]) for c, idx in node.outputs[0].clients)):
return [host_from_gpu(gpu_alloc(*node.inputs))]
def local_gpuaalloc(node):
new_out = gpu_alloc(*node.inputs)
# We need to hide new broadcastable dimensions because
# ReplaceValidate doesn't like when they change.
if new_out.broadcastable != node.outputs[0].broadcastable:
# but if a dim is suddenly not broadcastable anymore then that's a bug
for b_old, b_new in zip(node.outputs[0].broadcastable,
new_out.broadcastable):
assert b_new or (not b_old)
new_out = tensor.patternbroadcast(new_out,
node.outputs[0].broadcastable)
return (new_out,)
def local_gpuaalloc2(node):
"""
Join(axis, Alloc, Alloc, ...) -> Join(axis, GpuAlloc, Alloc, ...)
Moves an alloc that is an input to join to the gpu.
"""
if isinstance(node.op, tensor.Alloc) and all(
c != "output"
and c.op == tensor.join
and all(i.owner and i.owner.op in [host_from_gpu, tensor.alloc] for i in c.inputs[1:])
for c, idx in node.outputs[0].clients
):
return [host_from_gpu(gpu_alloc(*node.inputs))]