IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Avoid excess tensor copy for Reshape/Squeeze/Unsqueeze folding (#2834)

Browse Source 
* Updated Reshape ConstantFolding to avoid excess tensor copies

* Updated Squeeze/Unsqueeze CF to avoid excess tensor copies

* Fixed typo
This commit is contained in:
Gleb Kazantaev 2020-10-27 12:05:02 +03:00 committed by GitHub
parent 314ec2df72
commit d36bd8c87b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
8 changed files with 109 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
ngraph/core/include/ngraph/op/constant.hpp
View File

@ -285,6 +285,12 @@ namespace ngraph
/// Repeated values are allowed.
AxisSet get_axis_set_val() const;
/// \brief Update Constant shape. New shape size must equal to the data elements
/// count
///
/// \param shape The shape of the tensor constant.
void set_data_shape(const Shape& shape);
/// \brief Wrapper around constructing a shared_ptr of a Constant
///
/// \param type The element type of the tensor constant.

2
ngraph/core/include/ngraph/op/reshape.hpp
View File

@ -153,6 +153,8 @@ namespace ngraph
void set_special_zero(bool special_zero) { m_special_zero = special_zero; }
bool evaluate(const HostTensorVector& outputs,
const HostTensorVector& inputs) const override;
bool constant_fold(OutputVector& output_values,
const OutputVector& inputs_values) override;
protected:
bool m_special_zero;

2
ngraph/core/include/ngraph/op/squeeze.hpp
View File

@ -44,6 +44,8 @@ namespace ngraph
virtual void pre_validate_and_infer_types() override;
bool evaluate(const HostTensorVector& outputs,
const HostTensorVector& inputs) const override;
bool constant_fold(OutputVector& output_values,
const OutputVector& inputs_values) override;
virtual std::shared_ptr<Node>
clone_with_new_inputs(const OutputVector& new_args) const override;

2
ngraph/core/include/ngraph/op/unsqueeze.hpp
View File

@ -45,6 +45,8 @@ namespace ngraph
bool visit_attributes(AttributeVisitor& visitor) override;
bool evaluate(const HostTensorVector& outputs,
const HostTensorVector& inputs) const override;
bool constant_fold(OutputVector& output_values,
const OutputVector& inputs_values) override;
virtual std::shared_ptr<Node>
clone_with_new_inputs(const OutputVector& new_args) const override;

6
ngraph/core/src/op/constant.cpp
View File

@ -540,6 +540,12 @@ AxisSet op::Constant::get_axis_set_val() const
return output_axis_set;
}
void op::Constant::set_data_shape(const Shape& shape)
{
NGRAPH_CHECK(shape_size(shape) == shape_size(m_shape));
m_shape = shape;
}
shared_ptr<Node> op::Constant::clone_with_new_inputs(const OutputVector& new_args) const
{
check_new_args_count(this, new_args);

30
ngraph/core/src/op/reshape.cpp
View File

@ -453,3 +453,33 @@ bool op::v1::Reshape::evaluate(const HostTensorVector& outputs,
const AxisVector order = get_default_order(inputs[0]->get_shape());
return evaluate_reshape(inputs[0], outputs[0], order);
}
bool op::v1::Reshape::constant_fold(OutputVector& output_values, const OutputVector& inputs_values)
{
if (get_output_partial_shape(0).is_dynamic())
{
return false;
}
const auto& shape = get_output_shape(0);
if (auto data_const =
std::dynamic_pointer_cast<op::Constant>(inputs_values[0].get_node_shared_ptr()))
{
// In case if data constant has single consumer we can change it shape without making a copy
// Otherwise we create Constant copy with shape from reshape node
if (data_const->output(0).get_target_inputs().size() == 1)
{
data_const->set_data_shape(shape);
data_const->validate_and_infer_types();
output_values[0] = data_const;
}
else
{
output_values[0] = std::make_shared<op::Constant>(
data_const->get_element_type(), shape, data_const->get_data_ptr());
}
return true;
}
return false;
}

30
ngraph/core/src/op/squeeze.cpp
View File

@ -212,3 +212,33 @@ bool op::v0::Squeeze::evaluate(const HostTensorVector& outputs,
OV_ITT_SCOPED_TASK(itt::domains::nGraphOp, "op::v0::Squeeze::evaluate");
return squeeze::evaluate_squeeze(inputs[0], inputs[1], outputs[0]);
}
bool op::v0::Squeeze::constant_fold(OutputVector& output_values, const OutputVector& inputs_values)
{
if (get_output_partial_shape(0).is_dynamic())
{
return false;
}
const auto& shape = get_output_shape(0);
if (auto data_const =
std::dynamic_pointer_cast<op::Constant>(inputs_values[0].get_node_shared_ptr()))
{
// In case if data constant has single consumer we can change it shape without making a copy
// Otherwise we create Constant copy with shape from squeeze node
if (data_const->output(0).get_target_inputs().size() == 1)
{
data_const->set_data_shape(shape);
data_const->validate_and_infer_types();
output_values[0] = data_const;
}
else
{
output_values[0] = std::make_shared<op::Constant>(
data_const->get_element_type(), shape, data_const->get_data_ptr());
}
return true;
}
return false;
}

31
ngraph/core/src/op/unsqueeze.cpp
View File

@ -173,3 +173,34 @@ bool op::v0::Unsqueeze::evaluate(const HostTensorVector& outputs,
OV_ITT_SCOPED_TASK(itt::domains::nGraphOp, "op::v0::Unsqueeze::evaluate");
return unsqueeze::evaluate_unsqueeze(inputs[0], inputs[1], outputs[0]);
}
bool op::v0::Unsqueeze::constant_fold(OutputVector& output_values,
const OutputVector& inputs_values)
{
if (get_output_partial_shape(0).is_dynamic())
{
return false;
}
const auto& shape = get_output_shape(0);
if (auto data_const =
std::dynamic_pointer_cast<op::Constant>(inputs_values[0].get_node_shared_ptr()))
{
// In case if data constant has single consumer we can change it shape without making a copy
// Otherwise we create Constant copy with shape from unsqueeze node
if (data_const->output(0).get_target_inputs().size() == 1)
{
data_const->set_data_shape(shape);
data_const->validate_and_infer_types();
output_values[0] = data_const;
}
else
{
output_values[0] = std::make_shared<op::Constant>(
data_const->get_element_type(), shape, data_const->get_data_ptr());
}
return true;
}
return false;
}