* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
* \file tensor.h
* \brief
*/
#pragma once
#include <string>
#include <memory>
#include <vector>
#include <initializer_list>
#include "tilefwk/data_type.h"
#include "tilefwk/symbolic_scalar.h"
namespace npu::tile_fwk {
class LogicalTensor;
class SymbolicScalar;
using BinDataPtr = uint8_t*;
using Shape = std::vector<int64_t>;
using Offset = std::vector<int64_t>;
using Stride = std::vector<int64_t>;
class Tensor {
public:
* \brief Constructs a new default Tensor object
*/
Tensor();
* \brief Destroy the Tensor object
*/
~Tensor();
* \brief Constructs a new Tensor object with one input parameter
*
* \param s : a shared pointer to a LogicalTensor object
*/
Tensor(std::shared_ptr<LogicalTensor> s);
* \brief Construct a new Tensor object with 5 input parameters
*
* \param dataType : Data type of the tensor.
* \param shape : A vector that stores the shape of the tensor.
* \param name : Name of the tensor. The default value is "".
* \param format : Format of the tensor. The default value is TileOpFormat::TILEOP_ND.
* \attention : The parameters dataType and shape are required parameters.
*/
Tensor(DataType dataType, const Shape& shape, std::string name = "", TileOpFormat format = TileOpFormat::TILEOP_ND);
* \brief Construct a new Tensor object with 6 input parameters
*
* \param dataType : Data type of the tensor.
* \param shape : A vector that stores the shape of the tensor.
* \param dataPtr : Pointer to the dataPtr of the tensor.
* \param name : Name of the tensor.
* \param format : Format of the tensor. The default value is TileOpFormat::TILEOP_ND.
* \attention : The parameters dataType,shape,dataPtr and name are required parameters.
*/
Tensor(
DataType dataType, const Shape& shape, uint8_t* dataPtr, std::string name,
TileOpFormat format = TileOpFormat::TILEOP_ND)
: Tensor(dataType, shape, name, format)
{
SetData(dataPtr);
}
* \brief Construct a new Tensor object
*
* \param dataType : Datatype
* \param shape : Shape of the tensor
* \param name : Name of the tensor.
* \param format : Format of the tensor. The default value is TileOpFormat::TILEOP_ND.
*/
Tensor(
DataType dataType, std::vector<SymbolicScalar> shape, std::string name = "",
TileOpFormat format = TileOpFormat::TILEOP_ND);
* \brief Construct a new Tensor object
*
* \param t : Datatype
* \param shape : Shape of the tensor
* \param name : Name of the tensor.
* \param format : Format of the tensor. The default value is TileOpFormat::TILEOP_ND.
* \code {.cpp}
* Tensor t0(DT_FP32, {32, 32}) // shape with fixed type
* Tensor t1(DT_FP32, {?, 32}) // first axis use dynamic shape
* Tensor t2(DT_FP32, {GetInputShapeDim(t1, 0), 32}) // shape same as t1
* \endcode
*/
Tensor(
DataType t, std::initializer_list<SymbolicScalar> shape, std::string name = "",
TileOpFormat format = TileOpFormat::TILEOP_ND)
: Tensor(t, std::vector<SymbolicScalar>(shape), name, format)
{}
* \brief Overload the assignment operator to assign the value of another Tensor object to the current Tensor
* object.
*
* \param rhs : A constant reference to another Tensor object.
* \return Tensor& : A reference to the current Tensor object.
*/
Tensor& operator=(const Tensor& rhs);
* \brief Move assignment operator.
*
* \param rhs : Rvalue reference to another Tensor object.
* \return Tensor& : A reference to the current Tensor object.
* \attention : The noexcept declaration indicates that the function will not throw exceptions.
*/
Tensor& operator=(Tensor&& rhs) noexcept;
* \brief Construct a new Tensor object by copying another Tensor object.
*
* \param rhs : A constant reference to another Tensor object.
*/
Tensor(const Tensor& rhs);
* \brief Construct a new Tensor object by moving another Tensor object.
*
* \param rhs : Rvalue reference to another Tensor object.
*/
Tensor(Tensor&& rhs);
* \brief Overload the * operator to access the LogicalTensor object.
*
* \return const LogicalTensor& : A reference to the LogicalTensor object.
* \attention : The const keyword indicates that the function does not modify the object.
*/
const LogicalTensor& operator*() const;
* \brief Overload the * operator to access the LogicalTensor object.
*
* \return LogicalTensor& : A reference to the LogicalTensor object.
*/
LogicalTensor& operator*();
* \brief Get the const Storage object.
*
* \param readSlot : This parameter indicates whether slot reading is required. The default value is true.
* \return const std::shared_ptr<LogicalTensor>& : A constant reference to the storage_ object.
* \attention : The const keyword indicates that the function does not modify the object.
*/
const std::shared_ptr<LogicalTensor>& GetStorage(bool readSlot = true) const;
* \brief Get the Storage object.
*
* \param readSlot : This parameter indicates whether slot reading is required. The default value is true.
* \return std::shared_ptr<LogicalTensor>& : A reference to the storage_ object.
*/
std::shared_ptr<LogicalTensor>& GetStorage(bool readSlot = true);
* \brief Set tensor cache policy.
*
* \param policy : PREFETCH mark this tensor will prefetch to cache before calculate.(Max num is 4.)
* NO_CACHEABLE mark this tensor will not get into cache.
* value : true mean enable, default is false.
* \attention : NONE_CACHEABLE will only effect function input.
* NO_CACHEABLE will only effect function input and output.
* Two policy can not apply in one tensor.
*/
void SetCachePolicy(CachePolicy policy, bool value);
* \brief Get tensor cache policy.
*
* \param policy : CachePolicy enum.
* \return bool : policy value.
*/
bool GetCachePolicy(CachePolicy policy) const;
* \brief Get the Data Type object
*
* \return DataType : The data type of the tensor.
*/
DataType GetDataType() const;
* \brief Get the shape of a tensor.
*
* \return const Shape & : A constant reference to the shape of the tensor.
*/
const Shape& GetShape() const;
* \brief Get the shape information of the specified axis of Tensor.
*
* \param axis : The axis of the shape to be obtained.
* \return int : The shape of the specified axis.
*/
int32_t GetShape(int axis) const;
* \brief Get the valid shape of the tensor.
*
* \return std::vector<SymbolicScalar> : The valid shape of the tensor.
*/
std::vector<SymbolicScalar>& GetValidShape() const;
* \brief Get the format of the tensor.
*
* \return TileOpFormat : The format of the tensor.
*/
TileOpFormat Format() const;
* \brief Get the Id information of the Tensor.
*
* \return int : The Id information of the Tensor.
*/
int Id() const { return index_; }
* \brief Set the data of Tensor.
*
* \param data : Pointer to the data of the tensor. The data type is uint8_t.
*/
void SetData(BinDataPtr data);
* \brief Get the data of Tensor.
*
* \return auto : A pointer to the data of the tensor.
*/
auto GetData() const { return data_; }
* \brief Set the name of Tensor.
*
* \param name : The name of the tensor.
*/
void SetName(const std::string& name) const;
* \brief Get the name of Tensor.
*
* \return const std::string& : The name of the tensor.
*/
std::string GetName() const;
* \brief Get the Shape Dim Size of Tensor.
*/
uint64_t Dim() const;
* \brief Check if the tensor is empty.
*
* \return true : If the tensor is empty.
* \return false : Otherwise.
*/
bool IsEmpty() const;
private:
std::shared_ptr<LogicalTensor> storage_;
int index_{-1};
BinDataPtr data_{};
};
* @brief Get the size of a special dimension of input tensor
*
* @param t input tensor
* @param n dimension index
* @return SymbolicScalar : size of a special dimension of tensor
*/
SymbolicScalar GetInputShape(const Tensor& t, int n);
const std::vector<SymbolicScalar>& GetInputShape(const Tensor& tensor);
* @brief Get the Input Data of a tensor
*
* @param t input tensor
* @param offset positional shift applied to the each axis of the tensor
* @return SymbolicScalar
*/
SymbolicScalar GetTensorData(const Tensor& t, const std::vector<SymbolicScalar>& offset);
* @brief Determines if the current iteration is the start of loop
*
* @param symbol current loop index
* @param begin begin loop index
* @return SymbolicScalar : expression to determine if currently at loop start
*/
SymbolicScalar IsLoopBegin(const SymbolicScalar& symbol, const SymbolicScalar& begin);
* @brief Determines if the current iteration is the end of loop
*
* @param symbol current loop index
* @param end end loop index
* @return SymbolicScalar : expression to determine if currently at loop start
*/
SymbolicScalar IsLoopEnd(const SymbolicScalar& symbol, const SymbolicScalar& end);
void SetTensorData(const SymbolicScalar& v, const std::vector<SymbolicScalar>& off, Tensor& dst);
}
