def local_gpualloc(node):
replace = False
if node.op == tensor.alloc:
if node.inputs[0].owner and node.inputs[0].owner.op == host_from_gpu:
replace = True
elif all([c != 'output' and c.op == gpu_from_host
for c, idx in node.outputs[0].clients]):
replace = True
elif 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]):
replace = True
if replace:
val = node.inputs[0]
shp = node.inputs[1:]
old_out = node.outputs[0]
val2 = tensor.shape_padleft(val, len(shp) - val.ndim)
new_out = host_from_gpu(gpu_alloc(val, *shp))
if new_out.type != old_out.type:
assert new_out.type.ndim == old_out.type.ndim
assert new_out.type.dtype == old_out.type.dtype
for b_old, b_new in zip(old_out.type.broadcastable,
new_out.type.broadcastable):
assert b_new or (not b_old)
new_out = tensor.patternbroadcast(new_out. old_out.broadcastable)
return [new_out]
def local_gpu_elemwise(node):
do_replace = False
gpu_out = False
# check for gpu_from_host(Elemwise)) and extract the Elemwise node
if node.op == gpu_from_host:
host_i, = node.inputs
if (host_i.owner and
isinstance(host_i.owner.op, tensor.Elemwise) and
len(host_i.clients) == 1):
node = host_i.owner
do_replace = True
gpu_out = True
# check for elemwise(..., host_from_gpu, ...)
if isinstance(node.op, tensor.Elemwise):
if numpy.any([i.owner and
i.owner.op == host_from_gpu
for i in node.inputs]):
do_replace = True
if numpy.all([_is_scalar(i)
for i in node.inputs]):
do_replace = False
if do_replace:
new_op = GpuElemwise(node.op.scalar_op)
gpu_elemwise = new_op(*(gpu_from_host(i) for i in node.inputs))
if gpu_out:
return [gpu_elemwise]
else:
return [host_from_gpu(gpu_elemwise)]
else:
return False
def local_gpualloc(node):
replace = False
if node.op == tensor.alloc:
if node.inputs[0].owner and node.inputs[0].owner.op == host_from_gpu:
replace = True
elif all([
c != 'output' and c.op == gpu_from_host
for c, idx in node.outputs[0].clients
]):
replace = True
elif 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
]):
replace = True
if replace:
val = node.inputs[0]
shp = node.inputs[1:]
old_out = node.outputs[0]
val2 = tensor.shape_padleft(val, len(shp) - val.ndim)
new_out = host_from_gpu(gpu_alloc(val, *shp))
if new_out.type != old_out.type:
assert new_out.type.ndim == old_out.type.ndim
assert new_out.type.dtype == old_out.type.dtype
for b_old, b_new in zip(old_out.type.broadcastable,
new_out.type.broadcastable):
assert b_new or (not b_old)
new_out = tensor.patternbroadcast(new_out.old_out.broadcastable)
return [new_out]
def local_gpu_elemwise(node):
do_replace = False
gpu_out = False
# check for gpu_from_host(Elemwise)) and extract the Elemwise node
if node.op == gpu_from_host:
host_i, = node.inputs
if (host_i.owner and isinstance(host_i.owner.op, tensor.Elemwise)
and len(host_i.clients) == 1):
node = host_i.owner
do_replace = True
gpu_out = True
# check for elemwise(..., host_from_gpu, ...)
if isinstance(node.op, tensor.Elemwise):
if numpy.any(
[i.owner and i.owner.op == host_from_gpu for i in node.inputs]):
do_replace = True
if numpy.all([_is_scalar(i) for i in node.inputs]):
do_replace = False
if do_replace:
new_op = GpuElemwise(node.op.scalar_op)
gpu_elemwise = new_op(*(gpu_from_host(i) for i in node.inputs))
if gpu_out:
return [gpu_elemwise]
else:
return [host_from_gpu(gpu_elemwise)]
else:
return False
def apply(self, fgraph):
for input in fgraph.inputs:
if isinstance(input.type, GpuArrayType):
continue
if (len(input.clients) == 1 and
(input.clients[0][0] == 'output' or
input.clients[0][0].op == gpu_from_host)):
continue
try:
new_input = host_from_gpu(gpu_from_host(input))
fgraph.replace_validate(input, new_input,
"InputToGpuOptimizer")
except TypeError, e:
# This could fail if the inputs are not TensorTypes
pass
def apply(self, fgraph):
for input in fgraph.inputs:
if isinstance(input.type, GpuArrayType):
continue
if (len(input.clients) == 1
and (input.clients[0][0] == 'output'
or input.clients[0][0].op == gpu_from_host)):
continue
try:
new_input = host_from_gpu(gpu_from_host(input))
fgraph.replace_validate(input, new_input,
"InputToGpuOptimizer")
except TypeError, e:
# This could fail if the inputs are not TensorTypes
pass
def _as_TensorVariable(self):
from basic_ops import host_from_gpu
return host_from_gpu(self)
def _as_TensorVariable(self):
from basic_ops import host_from_gpu
return host_from_gpu(self)
