def opt(node):
if (isinstance(node.op, GpuElemwise) and
node.op.scalar_op == scal.add and
node.nin == 2):
targ = find_node(node.inputs[0], cls)
W = node.inputs[1]
if targ is None:
targ = find_node(node.inputs[1], cls)
W = node.inputs[0]
if targ is None:
return None
if W.dtype != targ.outputs[0].dtype:
return None
if not is_equal(targ.inputs[beta_in], 0.0):
# other cases are too complex for now
return None
if W.broadcastable != targ.inputs[out_in].broadcastable:
# Would need to explicitly tile the output to fill
# the full shape here. Disable for now.
return None
inputs = list(targ.inputs)
inputs[out_in] = W
dtype = inputs[beta_in].dtype
one = scal.constant(np.asarray(1.0, dtype=dtype))
inputs[beta_in] = one
with inherit_stack_trace(node.outputs):
return maker(targ, *inputs)
def opt(node):
if (isinstance(node.op, GpuElemwise) and
node.op.scalar_op == scal.mul and
node.nin == 2):
targ = find_node(node.inputs[0], cls)
if targ is None:
targ = find_node(node.inputs[1], cls)
if targ is None:
return
lr = grab_cpu_scalar(node.inputs[0],
nd=targ.outputs[0].ndim)
else:
lr = grab_cpu_scalar(node.inputs[1],
nd=targ.outputs[0].ndim)
if lr is None or lr.dtype != targ.outputs[0].dtype:
return None
inputs = list(targ.inputs)
try:
c = get_scalar_constant_value(lr)
if c == 0:
inputs[alpha_in] = lr
inputs[beta_in] = lr
elif c == 1:
inputs[alpha_in] = targ.inputs[alpha_in]
inputs[beta_in] = targ.inputs[beta_in]
else:
inputs[alpha_in] = lr * targ.inputs[alpha_in]
inputs[beta_in] = lr * targ.inputs[beta_in]
except NotScalarConstantError:
inputs[alpha_in] = lr * targ.inputs[alpha_in]
inputs[beta_in] = lr * targ.inputs[beta_in]
with inherit_stack_trace(node.outputs):
return maker(targ, *inputs)
def opt(node):
if (isinstance(node.op, GpuElemwise)
and node.op.scalar_op == scal.add and node.nin == 2):
targ = find_node(node.inputs[0], cls)
W = node.inputs[1]
if targ is None:
targ = find_node(node.inputs[1], cls)
W = node.inputs[0]
if targ is None:
return None
if W.dtype != targ.outputs[0].dtype:
return None
if not is_equal(targ.inputs[beta_in], 0.0):
# other cases are too complex for now
return None
if W.broadcastable != targ.inputs[out_in].broadcastable:
# Would need to explicitly tile the output to fill
# the full shape here. Disable for now.
return None
inputs = list(targ.inputs)
inputs[out_in] = W
dtype = inputs[beta_in].dtype
one = scal.constant(np.asarray(1.0, dtype=dtype))
inputs[beta_in] = one
with inherit_stack_trace(node.outputs):
return maker(targ, *inputs)
def opt(node):
if (isinstance(node.op, GpuElemwise)
and node.op.scalar_op == scal.mul and node.nin == 2):
targ = find_node(node.inputs[0], cls)
if targ is None:
targ = find_node(node.inputs[1], cls)
if targ is None:
return
lr = grab_cpu_scalar(node.inputs[0],
nd=targ.outputs[0].ndim)
else:
lr = grab_cpu_scalar(node.inputs[1],
nd=targ.outputs[0].ndim)
if lr is None or lr.dtype != targ.outputs[0].dtype:
return None
inputs = list(targ.inputs)
try:
c = get_scalar_constant_value(lr)
if c == 0:
inputs[alpha_in] = lr
inputs[beta_in] = lr
elif c == 1:
inputs[alpha_in] = targ.inputs[alpha_in]
inputs[beta_in] = targ.inputs[beta_in]
else:
inputs[alpha_in] = lr * targ.inputs[alpha_in]
inputs[beta_in] = lr * targ.inputs[beta_in]
except NotScalarConstantError:
inputs[alpha_in] = lr * targ.inputs[alpha_in]
inputs[beta_in] = lr * targ.inputs[beta_in]
with inherit_stack_trace(node.outputs):
return maker(targ, *inputs)
def local_abstractconv_cudnn(node):
ctx = infer_context_name(*node.inputs)
if not isinstance(node.inputs[0].type, GpuArrayType):
return
if node.op.unshared:
return None
if isinstance(node.op.border_mode, tuple) and any(
isinstance(p, tuple) for p in node.op.border_mode):
# Asymmetric padding not yet supported
return None
if isinstance(node.op, AbstractConv2d):
with inherit_stack_trace(node.outputs):
return local_abstractconv_cudnn_graph(node.op, ctx, node.inputs,
node.outputs)
elif isinstance(node.op, AbstractConv3d):
with inherit_stack_trace(node.outputs):
return local_abstractconv3d_cudnn_graph(node.op, ctx, node.inputs,
node.outputs)
def opt(node):
if type(node.op) != op or node.op.inplace:
return
inputs = list(node.inputs)
alloc = inputs[idx]
if (alloc.owner and
isinstance(alloc.owner.op, GpuAllocEmpty) and
len(alloc.clients) > 1):
alloc_op = GpuAllocEmpty(alloc.owner.op.dtype, alloc.owner.op.context_name)
inputs[idx] = alloc_op(*alloc.owner.inputs)
with inherit_stack_trace(node.outputs):
return maker(node, inputs)
def opt(node):
if type(node.op) != op or node.op.inplace:
return
inputs = list(node.inputs)
alloc = inputs[idx]
if (alloc.owner and isinstance(alloc.owner.op, GpuAllocEmpty)
and len(alloc.clients) > 1):
alloc_op = GpuAllocEmpty(alloc.owner.op.dtype,
alloc.owner.op.context_name)
inputs[idx] = alloc_op(*alloc.owner.inputs)
with inherit_stack_trace(node.outputs):
return maker(node, inputs)
def local_dnn_reduction(node):
if not isinstance(node.op, GpuCAReduceCuda):
return
if not dnn_available(node.inputs[0].type.context_name):
return
if version(raises=False) < 6000:
return
if node.inputs[0].ndim > 8:
return
acc_dtype = node.op._acc_dtype(node.inputs[0].dtype)
if node.inputs[0].dtype != node.outputs[0].dtype:
# We can mix float16 and float32, but not float64.
if node.inputs[0].dtype == "float64" or node.outputs[
0].dtype == "float64":
return
if acc_dtype != "float32":
return
if node.inputs[0].dtype not in ["float16", "float32", "float64"]:
return
if node.inputs[0].dtype == "float64" and acc_dtype != "float64":
return
if node.inputs[0].dtype == "float32" and acc_dtype != "float32":
return
if node.inputs[0].dtype == "float16" and acc_dtype == "float64":
return
def _identity(a):
return a
def _square(a):
return GpuElemwise(theano.scalar.basic.sqr)(a)
scal = node.op.scalar_op.name
post = _identity
if node.op.pre_scalar_op is not None:
if isinstance(node.op.scalar_op, theano.scalar.basic.Add):
if isinstance(node.op.pre_scalar_op, theano.scalar.basic.Sqr):
scal = "norm2"
post = _square
elif isinstance(node.op.pre_scalar_op, theano.scalar.basic.Abs):
scal = "norm1"
else:
return
elif isinstance(node.op.scalar_op,
theano.scalar.basic.Maximum) and isinstance(
node.op.pre_scalar_op, theano.scalar.basic.Abs):
scal = "absmax"
else:
return
if not cudnn.cudnnReduceTensorOp_t.has_alias(scal):
return
with inherit_stack_trace(node.outputs):
ret = GpuDnnReduction(scal, node.op.axis, acc_dtype, node.op.dtype,
False)(node.inputs[0])
return [post(ret)]
