Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add SegmentMax-16 reference implementation #28788

Open
wants to merge 24 commits into
base: master
Choose a base branch
from
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions src/core/include/openvino/op/ops.hpp
Original file line number Diff line number Diff line change
Expand Up @@ -171,6 +171,7 @@
#include "openvino/op/scatter_nd_update.hpp"
#include "openvino/op/scatter_update.hpp"
#include "openvino/op/search_sorted.hpp"
#include "openvino/op/segment_max.hpp"
#include "openvino/op/select.hpp"
#include "openvino/op/selu.hpp"
#include "openvino/op/shape_of.hpp"
Expand Down
50 changes: 50 additions & 0 deletions src/core/include/openvino/op/segment_max.hpp
Original file line number Diff line number Diff line change
@@ -0,0 +1,50 @@
// Copyright (C) 2018-2025 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//

#pragma once

#include "openvino/op/op.hpp"

namespace ov {
namespace op {
namespace v16 {
/// \brief An operation which computes the maximum values along segments of a tensor.
/// \ingroup ov_ops_cpp_api
class OPENVINO_API SegmentMax : public ov::op::Op {
public:
OPENVINO_OP("SegmentMax", "opset16", ov::op::Op);

SegmentMax() = default;

/// \brief Constructs a SegmentMax operation.
///
/// \param data Input tensor with data
/// \param segment_ids Indices of segments in the data input tensor
/// \param fill_mode The value assigned to segments which are empty
SegmentMax(const Output<Node>& data, const Output<Node>& segment_ids, const op::FillMode fill_mode);

/// \brief Constructs a SegmentMax operation.
///
/// \param data Input tensor with data
/// \param segment_ids Indices of segments in the data input tensor
/// \param num_segments The segments count
/// \param fill_mode The value assigned to segments which are empty
SegmentMax(const Output<Node>& data,
const Output<Node>& segment_ids,
const Output<Node>& num_segments,
const op::FillMode fill_mode);

bool visit_attributes(AttributeVisitor& visitor) override;
void validate_and_infer_types() override;
std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& new_args) const override;

const op::FillMode get_fill_mode() const;

private:
op::FillMode m_fill_mode;
};

} // namespace v16
} // namespace op
} // namespace ov
14 changes: 14 additions & 0 deletions src/core/include/openvino/op/util/attr_types.hpp
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,12 @@ enum class PadMode { CONSTANT = 0, EDGE, REFLECT, SYMMETRIC };
OPENVINO_API
std::ostream& operator<<(std::ostream& s, const PadMode& type);

/// \brief Fill modes for the `SegmentMax` operator.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If this is specific for SegmentMax op only, then it should be placed in the op dedicated namespace and file.
Otherwise generalize the comment.

Copy link
Contributor

@mitruska mitruska Feb 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There is a chance to extend it and reuse it with other ops in the future like SegmentMin

Suggested change
/// \brief Fill modes for the `SegmentMax` operator.
/// \brief Fill modes to set default value for operators like `SegmentMax`.

But if it's preferred to use a separate enums for such ops, it should be inside the SegmentMax class/namespace/file.

enum class FillMode { ZERO = 0, LOWEST };

OPENVINO_API
std::ostream& operator<<(std::ostream& s, const FillMode& type);

/// \brief Padding Type used for `Convolution` and `Pooling`
///
/// Follows ONNX padding type definitions
Expand Down Expand Up @@ -219,6 +225,14 @@ class OPENVINO_API AttributeAdapter<op::PadMode> : public EnumAttributeAdapterBa
OPENVINO_RTTI("AttributeAdapter<PadMode>");
};

template <>
class OPENVINO_API AttributeAdapter<op::FillMode> : public EnumAttributeAdapterBase<op::FillMode> {
public:
AttributeAdapter(op::FillMode& value) : EnumAttributeAdapterBase<op::FillMode>(value) {}

OPENVINO_RTTI("AttributeAdapter<FillMode>");
};

template <>
class OPENVINO_API AttributeAdapter<op::PadType> : public EnumAttributeAdapterBase<op::PadType> {
public:
Expand Down
1 change: 1 addition & 0 deletions src/core/include/openvino/opsets/opset16_tbl.hpp
Original file line number Diff line number Diff line change
Expand Up @@ -15,3 +15,4 @@ _OPENVINO_OP_REG(ShapeOf, ov::op::v3)

// New operations added in opset16
_OPENVINO_OP_REG(Identity, ov::op::v16)
_OPENVINO_OP_REG(SegmentMax, ov::op::v16)
56 changes: 56 additions & 0 deletions src/core/reference/include/openvino/reference/segment_max.hpp
Original file line number Diff line number Diff line change
@@ -0,0 +1,56 @@
// Copyright (C) 2018-2025 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//

#pragma once

#include <algorithm>
#include <limits>
#include <vector>

#include "openvino/core/shape.hpp"

namespace ov {
namespace reference {

template <typename T, typename T_idx>
void segment_max(const T* data,
const Shape& data_shape,
const T_idx* segment_ids,
Comment on lines +16 to +19
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

To avoid additional type specialization for int32 and int64 (in favour of smaller binary size) there is a common approach to upcast the i32 to i64, and use only i64 in the ref instead of T_idx.
But as this reference implementation is not used in the evaluate from the op class, it's not crucial.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

An example here

T* out,
const Shape& output_shape,
const T empty_segment_value) {
const T_idx num_segments = output_shape[0];
const size_t inner_dim_size =
std::accumulate(data_shape.begin() + 1, data_shape.end(), 1, std::multiplies<size_t>());
Comment on lines +24 to +25
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
const size_t inner_dim_size =
std::accumulate(data_shape.begin() + 1, data_shape.end(), 1, std::multiplies<size_t>());
const auto inner_dim_size = shape_size(data_shape.begin() +1, data_shape.end());

std::vector<std::vector<T>> max_values(num_segments,
std::vector<T>(inner_dim_size, std::numeric_limits<T>::lowest()));
Comment on lines +26 to +27
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can be input data copy avoided and set max values in the result tensor instead?

std::vector<bool> segment_has_values(num_segments, false);

// Iterate over each element in the first dimension
T_idx base_idx;
for (size_t i = 0; i < data_shape[0]; ++i) {
const T_idx segment_id = segment_ids[i];
if (segment_id >= num_segments) {
break;
}
segment_has_values[segment_id] = true;
// Iterate over each element in the inner dimensions
base_idx = i * inner_dim_size;
for (size_t j = 0; j < inner_dim_size; ++j) {
// Update the maximum value for the current segment and inner dimension
max_values[segment_id][j] = std::max(max_values[segment_id][j], data[base_idx + j]);
}
}

// Populate the output array with the maximum values for each segment
for (T_idx segment_id = 0; segment_id < num_segments; ++segment_id) {
base_idx = segment_id * inner_dim_size;
for (size_t j = 0; j < inner_dim_size; ++j) {
out[base_idx + j] = segment_has_values[segment_id] ? max_values[segment_id][j] : empty_segment_value;
}
}
}

} // namespace reference
} // namespace ov
Original file line number Diff line number Diff line change
@@ -0,0 +1,83 @@
// Copyright (C) 2018-2025 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//

#pragma once

#include <cmath>

#include "openvino/op/segment_max.hpp"
#include "utils.hpp"

namespace ov {
namespace op {
namespace v16 {
template <class TShape, class TRShape = result_shape_t<TShape>>
std::vector<TRShape> shape_infer(const SegmentMax* op,
const std::vector<TShape>& input_shapes,
const ITensorAccessor& tensor_accessor = make_tensor_accessor()) {
NODE_VALIDATION_CHECK(op, input_shapes.size() == 2 || input_shapes.size() == 3);

// validate data input
const auto& data_shape = input_shapes[0];
const auto is_data_shape_rank_static = data_shape.rank().is_static();
if (is_data_shape_rank_static) {
NODE_SHAPE_INFER_CHECK(op, input_shapes, data_shape.size() > 0, "The data input cannot be a scalar.");
}

// validate segment_ids input
const auto& segment_ids_shape = input_shapes[1];
const auto is_segment_ids_rank_static = segment_ids_shape.rank().is_static();
if (is_segment_ids_rank_static) {
NODE_SHAPE_INFER_CHECK(op,
input_shapes,
segment_ids_shape.size() == 1,
"segment_ids must be a 1D input. Got: ",
segment_ids_shape);
if (is_data_shape_rank_static) {
NODE_SHAPE_INFER_CHECK(op,
input_shapes,
data_shape[0].compatible(segment_ids_shape[0]),
"The number of elements in segment_ids must match the first dimension of data.");
}
}
const auto segment_ids = ov::op::get_input_const_data_as<TRShape, int64_t>(op, 1, tensor_accessor);
if (segment_ids) {
NODE_VALIDATION_CHECK(op,
std::is_sorted(segment_ids->begin(), segment_ids->end()),
"segment_ids must be sorted.");
}

// validate num_segments input
const auto num_segments_available = op->inputs().size() == 3;
const auto num_segments = num_segments_available ? get_input_const_data_as_shape<TRShape>(op, 2, tensor_accessor)
: ov::optional<TRShape>{};
if (num_segments_available) {
const auto& num_segments_shape = input_shapes[2];
NODE_SHAPE_INFER_CHECK(op,
input_shapes,
num_segments_shape.rank().compatible(0),
"num_segments must be a scalar input. Got: ",
num_segments_shape);
}

if (!is_data_shape_rank_static) {
return {PartialShape::dynamic()};
}
using TDim = typename TShape::value_type;
auto output_shapes = std::vector<TRShape>{data_shape};
auto& output_shape = output_shapes[0];
if (num_segments) {
output_shape[0] = TDim((*num_segments)[0]);
} else if (segment_ids &&
op->inputs().size() ==
2) { // if num_segments is an input but not provided, the first dimension should still be dynamic
output_shape[0] = TDim(*std::max_element(segment_ids->begin(), segment_ids->end()) + 1);
} else {
output_shape[0] = Dimension::dynamic();
}
return output_shapes;
}
} // namespace v16
} // namespace op
} // namespace ov
72 changes: 72 additions & 0 deletions src/core/src/op/segment_max.cpp
Original file line number Diff line number Diff line change
@@ -0,0 +1,72 @@
// Copyright (C) 2018-2025 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//

#include "openvino/op/segment_max.hpp"

#include "itt.hpp"
#include "openvino/core/validation_util.hpp"
#include "openvino/op/op.hpp"
#include "segment_max_shape_inference.hpp"

namespace ov {
namespace op {
namespace v16 {

SegmentMax::SegmentMax(const Output<Node>& data, const Output<Node>& segment_ids, const op::FillMode fill_mode)
: Op({data, segment_ids}),
m_fill_mode(fill_mode) {
constructor_validate_and_infer_types();
}

SegmentMax::SegmentMax(const Output<Node>& data,
const Output<Node>& segment_ids,
const Output<Node>& num_segments,
const op::FillMode fill_mode)
: Op({data, segment_ids, num_segments}),
m_fill_mode(fill_mode) {
constructor_validate_and_infer_types();
}

bool SegmentMax::visit_attributes(ov::AttributeVisitor& visitor) {
OV_OP_SCOPE(v16_SegmentMax_visit_attributes);
visitor.on_attribute("fill_mode", m_fill_mode);
return true;
}

void SegmentMax::validate_and_infer_types() {
OV_OP_SCOPE(v16_SegmentMax_validate_and_infer_types);
const auto& segment_ids_element_type = get_input_element_type(1);
NODE_VALIDATION_CHECK(this,
segment_ids_element_type == element::i32 || segment_ids_element_type == element::i64,
"The element type of the segment_ids input be i32 or i64. Got: ",
segment_ids_element_type);
if (inputs().size() == 3) {
const auto& num_segments_element_type = get_input_element_type(2);
NODE_VALIDATION_CHECK(this,
num_segments_element_type == element::i32 || num_segments_element_type == element::i64,
"The element type of the num_segments input be i32 or i64. Got: ",
num_segments_element_type);
}

const auto output_shapes = shape_infer(this, ov::util::get_node_input_partial_shapes(*this));
set_output_type(0, get_input_element_type(0), output_shapes[0]);
}

std::shared_ptr<Node> SegmentMax::clone_with_new_inputs(const ov::OutputVector& new_args) const {
OV_OP_SCOPE(v16_SegmentMax_clone_with_new_inputs);
check_new_args_count(this, new_args);
if (new_args.size() == 3) {
return std::make_shared<SegmentMax>(new_args.at(0), new_args.at(1), new_args.at(2), m_fill_mode);
} else {
return std::make_shared<SegmentMax>(new_args.at(0), new_args.at(1), m_fill_mode);
}
}

const op::FillMode SegmentMax::get_fill_mode() const {
return m_fill_mode;
}

} // namespace v16
} // namespace op
} // namespace ov
12 changes: 12 additions & 0 deletions src/core/src/op/util/attr_types.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,14 @@ OPENVINO_API EnumNames<ov::op::PadMode>& EnumNames<ov::op::PadMode>::get() {
return enum_names;
}

template <>
OPENVINO_API EnumNames<ov::op::FillMode>& EnumNames<ov::op::FillMode>::get() {
static auto enum_names =
EnumNames<ov::op::FillMode>("ov::op::FillMode",
{{"zero", ov::op::FillMode::ZERO}, {"lowest", ov::op::FillMode::LOWEST}});
return enum_names;
}

template <>
OPENVINO_API EnumNames<ov::op::PadType>& EnumNames<ov::op::PadType>::get() {
static auto enum_names = EnumNames<ov::op::PadType>("ov::op::PadType",
Expand Down Expand Up @@ -132,6 +140,10 @@ std::ostream& op::operator<<(std::ostream& s, const ov::op::PadMode& type) {
return s << as_string(type);
}

std::ostream& op::operator<<(std::ostream& s, const ov::op::FillMode& type) {
return s << as_string(type);
}

std::ostream& op::operator<<(std::ostream& s, const ov::op::PadType& type) {
return s << as_string(type);
}
Expand Down
Loading
Loading