sparse tensor data structures

COOTensor init
CSRTensor init
added SparseTensorAllocator
moved SparseTensor to arm_compute/core
fixed to_dense method for COOTensor
fixed to_sparse and to_dense methods for CSRTensor
changed CSR indices; print function moved to the bae class
implemented get_value to CSR and COO tensors
added SparseIterator

Change-Id: Id5ed095587662b750b89cbabd600a286ab9bf53b

Author: aestriplex

arm_compute/core/Coordinates.h

@@ -49,6 +49,16 @@ class Coordinates : public Dimensions<int>
     constexpr Coordinates(Ts... coords) : Dimensions{coords...}
     {
     }
+
+    /** Constructor to initialize the coordinates from a vector.
+     *
+     * @param[in] coords Vector containing the values to initialize the dimensions.
+     */
+    template <typename T>
+    constexpr Coordinates(std::vector<T> coords) : Dimensions(coords)
+    {
+    }
+
     /** Allow instances of this class to be copy constructed */
     constexpr Coordinates(const Coordinates &) = default;
     /** Allow instances of this class to be copied */

arm_compute/core/Dimensions.h

@@ -58,6 +58,15 @@ class Dimensions
     {
     }
 
+    /** Constructor to initialize the tensor shape with a list of dim values.
+     *
+     * @param[in] dims Vector of values to initialize the dimensions.
+     */
+    explicit Dimensions(std::vector<T> dims) : _id(), _num_dimensions{dims.size()}
+    {
+        std::copy_n(dims.begin(), std::min(num_max_dimensions, dims.size()), _id.begin());
+    }
+
     /** Allow instances of this class to be copy constructed */
     Dimensions(const Dimensions &) = default;
 

arm_compute/core/Helpers.h

@@ -34,6 +34,7 @@
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
+#include "arm_compute/core/SparseTensor.h"
 
 #include <array>
 #include <cstddef>
@@ -126,6 +127,42 @@ class Iterator
     std::array<Dimension, Coordinates::num_max_dimensions> _dims;
 };
 
+class SparseIterator
+{
+public:
+    /** Create an iterator for the metadata and allocation contained in the SparseTensor
+     *
+     * @param[in] tensor A reference to the tensor to associate to the iterator.
+     */
+    SparseIterator(const SparseTensor &tensor);
+
+    /** Returns true if there is at least one more element to iterate */
+    bool has_next() const;
+
+    /** Move to the next non-zero element */
+    void next();
+
+    /** Get the coordinates of the current non-zero element */
+    Coordinates coordinates() const;
+
+    /** Get the value of the current non-zero element */
+    const uint8_t *value() const;
+
+    /** Reset iterator to the beginning */
+    void reset();
+
+    /** Get current index (nth non-zero) */
+    size_t index() const;
+
+    /** Get the number of non-zero elements that are left to iterate */
+    size_t num_left() const;
+
+private:
+    const SparseTensor &_tensor;
+    size_t              _index;
+    size_t              _nnz;
+};
+
 /** Iterate through the passed window, automatically adjusting the iterators and calling the lambda_functino for each element.
  *  It passes the x and y positions to the lambda_function for each iteration
  *

arm_compute/core/Helpers.inl

@@ -101,6 +101,7 @@ inline Iterator::Iterator(const ITensor *tensor, const Window &win) : Iterator()
 {
     ARM_COMPUTE_ERROR_ON(tensor == nullptr);
     ARM_COMPUTE_ERROR_ON(tensor->info() == nullptr);
+    ARM_COMPUTE_ERROR_ON_MSG(tensor->info()->is_sparse(), "Sparse tensors are not supported by Iterators. Use a SparseIterator instead");
 
     initialize(tensor->info()->num_dimensions(), tensor->info()->strides_in_bytes(), tensor->buffer(),
                tensor->info()->offset_first_element_in_bytes(), win);
@@ -169,6 +170,45 @@ inline void Iterator::reset(const size_t dimension)
     }
 }
 
+inline SparseIterator::SparseIterator(const SparseTensor &tensor) : _tensor(tensor), _index(0), _nnz(tensor.nnz())
+{
+}
+
+inline bool SparseIterator::has_next() const
+{
+    return _index < _nnz;
+}
+
+inline void SparseIterator::next()
+{
+    ++_index;
+}
+
+inline Coordinates SparseIterator::coordinates() const
+{
+    return _tensor.get_coordinates(_index);
+}
+
+inline const uint8_t *SparseIterator::value() const
+{
+    return _tensor.get_value(coordinates());
+}
+
+inline void SparseIterator::reset()
+{
+    _index = 0;
+}
+
+inline size_t SparseIterator::index() const
+{
+    return _index;
+}
+
+inline size_t SparseIterator::num_left() const
+{
+    return _nnz - _index;
+}
+
 inline Coordinates index2coords(const TensorShape &shape, int index)
 {
     int num_elements = shape.total_size();

arm_compute/core/IReducibleTensor.h

@@ -0,0 +1,76 @@
+/*
+ * Copyright (c) 2025 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef ACL_ARM_COMPUTE_CORE_IREDUCIBLETENSOR_H
+#define ACL_ARM_COMPUTE_CORE_IREDUCIBLETENSOR_H
+
+#include "arm_compute/core/SparseTensor.h"
+#include "arm_compute/runtime/COOTensor.h"
+#include "arm_compute/runtime/CSRTensor.h"
+
+namespace arm_compute
+{
+/** */
+class IReducibleTensor
+{
+public:
+    virtual ~IReducibleTensor() = default;
+    /** Convert a dense tensor to sparse tensor with specified sparse dimensions using the default
+     *  sparse tensor representation: COO format.
+     *
+     * @param[in] dim sparse dimension
+     *
+     * @return A unique pointer to a SparseTensor object.
+     */
+    virtual std::unique_ptr<SparseTensor> to_sparse(size_t dim) const = 0;
+    /** Convert a dense tensor to COO sparse tensor with specified sparse dimensions.
+     *
+     * @param[in] dim sparse dimension
+     *
+     * @return A unique pointer to a COOTensor object.
+     */
+    virtual std::unique_ptr<COOTensor> to_coo_sparse(size_t dim) const = 0;
+    /** Convert a dense tensor to COO sparse tensor with the default sparse dimension.
+     *  For COO format, the number of sparse dimensions is equal to the total number of dimensions.
+     *
+     * @return A unique pointer to a COOTensor object.
+     */
+    virtual std::unique_ptr<COOTensor> to_coo_sparse() const = 0;
+    /** Convert a dense tensor to CSR sparse tensor with specified sparse dimensions.
+     *
+     * @param[in] dim sparse dimension
+     *
+     * @return A unique pointer to a CSRTensor object.
+     */
+    virtual std::unique_ptr<CSRTensor> to_csr_sparse(size_t dim) const = 0;
+    /** Convert a dense tensor to CSR sparse tensor with the default sparse dimension.
+     *  For CSR format, the number of sparse dimensions is equal to 2.
+     *
+     * @return A unique pointer to a CSRTensor object.
+     */
+    virtual std::unique_ptr<CSRTensor> to_csr_sparse() const = 0;
+};
+}
+
+#endif // ACL_ARM_COMPUTE_CORE_IREDUCIBLETENSOR_H

arm_compute/core/ITensorInfo.h

@@ -31,6 +31,7 @@
 #include "arm_compute/core/Coordinates.h"
 #include "arm_compute/core/Strides.h"
 #include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/TensorFormat.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/utils/misc/Utility.h"
 
@@ -105,6 +106,13 @@ class ITensorInfo : public misc::ICloneable<ITensorInfo>
      * @return Reference to this ITensorInfo object
      */
     virtual ITensorInfo &set_format(Format format) = 0;
+    /** Set the tensor format of an already initialized tensor.
+     *
+     * @param[in] tensor format to distinguish between dense and sparse tensors.
+     *
+     * @return Reference to this ITensorInfo object
+     */
+    virtual ITensorInfo &set_tensor_format(TensorFormat tf) = 0;
     /** Set the shape of an already initialized tensor.
      *
      * @warning Changing the shape requires to recompute the strides and is
@@ -261,6 +269,16 @@ class ITensorInfo : public misc::ICloneable<ITensorInfo>
      * @return True if the tensor size can be changed.
      */
     virtual bool is_resizable() const = 0;
+    /** Flag indicating whether the tensor is sparse.
+     *
+     * @return True if the tensor format different from TensorFormat::Dense.
+     */
+    virtual bool is_sparse() const = 0;
+    /** Returns the format of the Tensor
+     *
+     * @return True if the tensor is an instance of SparseTensor.
+     */
+    virtual TensorFormat tensor_format() const = 0;
     /** Set the tensor as dynamic/static
      *
      * @param[in] dynamic True if tensor is dynamic

arm_compute/core/SparseTensor.h

@@ -0,0 +1,102 @@
+/*
+ * Copyright (c) 2025 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef ACL_ARM_COMPUTE_RUNTIME_SPARSETENSOR_H
+#define ACL_ARM_COMPUTE_RUNTIME_SPARSETENSOR_H
+
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+
+#include <functional>
+#include <stdexcept>
+
+typedef std::function<bool(const void *)> predicate_t;
+
+namespace arm_compute
+{
+/** Common base class for all sparse tensors */
+class SparseTensor : public ITensor
+{
+public:
+    /** Prevent instances of this class to be constructed by the default constructor */
+    SparseTensor() = delete;
+    /** Prevent instances of this class to be copy constructed */
+    SparseTensor(const SparseTensor&) = delete;
+    /** Prevent instances of this class to be copied */
+    SparseTensor& operator=(const SparseTensor&) = delete;
+    ~SparseTensor() = default;
+
+    /** Returns the number of sparse dimensions */
+    size_t sparse_dim() const;
+    /** Returns the number of dense dimensions */
+    size_t dense_dim() const;
+    /** Returns the (total) number of dimensions */
+    size_t dim() const;
+    /** Returns true if the tensor is hybrid (contains both sparse and dense dimensions)
+     *  
+     *  @note A sparse tensor is hybrid if it has at least one dense dimension.
+     */
+    bool is_hybrid() const;
+    /** Returns the ratio of zero-valued elements to the total number of elements */
+    float sparsity() const;
+    /** Returns the ratio of non-zero elements to the total number of elements */
+    float density() const;
+    /** Returns the dense volume */
+    uint32_t dense_volume(size_t sparse_dim) const;
+    /** Returns the number of non zero elements */
+    virtual size_t nnz() const = 0;
+    /** Converts the sparse tensor to a dense tensor */
+    virtual std::unique_ptr<ITensor> to_dense() = 0;
+    /** Returns the coordinates of the n-th (non-zero) element.
+     *
+     *  @param nth The *zero-base* index of the element
+     *
+     *  @return The coordinates of the element
+     */
+    virtual Coordinates get_coordinates(size_t nth) const = 0;
+    /** Returns a pointer to the n-th (non-zero) element. If the element specified by
+     *  the coordinates is zero, nullptr is returned.
+     *
+     *  @param nth The *zero-base* index of the element
+     *
+     *  @return The value of the element
+     *
+     *  @note The value has size dense_volume(sparse_dim()).
+     */
+    virtual const uint8_t *get_value(Coordinates coords) const = 0;
+
+protected:
+    SparseTensor(size_t dim, size_t sparse_dim);
+
+    std::function<bool(const void *)> make_is_nonzero_predicate(DataType dt) const;
+    bool has_non_zero_elements(uint8_t *arr, size_t len, size_t element_size, predicate_t is_non_zero) const;
+    void print_values(std::ostream &os, const uint8_t *data, size_t offset, size_t count) const;
+
+private:
+    size_t _total_dim;
+    size_t _sparse_dim;
+};
+}
+
+#endif // ACL_ARM_COMPUTE_RUNTIME_SPARSETENSOR_H

arm_compute/core/SubTensorInfo.h

@@ -34,6 +34,7 @@
 #include "arm_compute/core/Strides.h"
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/TensorFormat.h"
 
 #include <cstddef>
 #include <memory>
@@ -200,6 +201,17 @@ class SubTensorInfo final : public ITensorInfo
         ARM_COMPUTE_ERROR_ON(_parent == nullptr);
         return _parent->is_resizable();
     }
+    bool is_sparse() const override
+    {
+        ARM_COMPUTE_ERROR_ON(_parent == nullptr);
+        return _parent->is_sparse();
+    }
+    ITensorInfo &set_tensor_format(TensorFormat tf) override;
+    TensorFormat tensor_format() const override
+    {
+        ARM_COMPUTE_ERROR_ON(_parent == nullptr);
+        return _parent->tensor_format();
+    }
     ITensorInfo &set_dynamic(bool dynamic) override;
     bool         is_dynamic() const override
     {