fd_tensor.h

// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once

#include <iostream>
#include <numeric>
#include <string>
#include <vector>

#include "fastdeploy/core/allocate.h"
#include "fastdeploy/core/fd_scalar.h"
#include "fastdeploy/core/fd_type.h"
#include "fastdeploy/runtime/enum_variables.h"

namespace fastdeploy {

struct FASTDEPLOY_DECL FDTensor {
  void SetData(const std::vector<int64_t>& tensor_shape, const FDDataType& data_type, void* data_buffer, bool copy = false, const Device& data_device = Device::CPU, int data_device_id = -1) {
    SetExternalData(tensor_shape, data_type, data_buffer, data_device, data_device_id);
    if (copy) {
      StopSharing();
    }
  }

  void* GetData() {
    return MutableData();
  }
  const void* GetData() const {
    return Data();
  }

  void ExpandDim(int64_t axis = 0);

  void Squeeze(int64_t axis = 0);

  bool Reshape(const std::vector<int64_t>& new_shape);

  int Nbytes() const;

  int Numel() const;

  std::vector<int64_t> Shape() const { return shape; }

  FDDataType Dtype() const { return dtype; }

  void Allocate(const FDDataType& data_type, const std::vector<int64_t>& data_shape) {
    Allocate(data_shape, data_type, name);
  }

  void PrintInfo(const std::string& prefix = "Debug TensorInfo: ") const;

  std::string name = "";

  bool IsShared() { return external_data_ptr != nullptr; }
  void StopSharing();


  // ******************************************************
  // The following member and function only used by inside FastDeploy, maybe removed in next version

  void* buffer_ = nullptr;
  std::vector<int64_t> shape = {0};
  FDDataType dtype = FDDataType::INT8;

  // This use to skip memory copy step
  // the external_data_ptr will point to the user allocated memory
  // user has to maintain the memory, allocate and release
  void* external_data_ptr = nullptr;
  // The internal data will be on CPU
  // Some times, the external data is on the GPU, and we are going to use
  // GPU to inference the model
  // so we can skip data transfer, which may improve the efficience
  Device device = Device::CPU;
  // By default the device id of FDTensor is -1, which means this value is
  // invalid, and FDTensor is using the same device id as Runtime.
  int device_id = -1;

  // Whether the data buffer is in pinned memory, which is allocated
  // with cudaMallocHost()
  bool is_pinned_memory = false;

  // if the external data is not on CPU, we use this temporary buffer
  // to transfer data to CPU at some cases we need to visit the
  // other devices' data
  std::vector<int8_t> temporary_cpu_buffer;

  // The number of bytes allocated so far.
  // When resizing GPU memory, we will free and realloc the memory only if the
  // required size is larger than this value.
  size_t nbytes_allocated = 0;

  // Get data buffer pointer
  void* MutableData();

  void* Data();

  const void* Data() const;

  // Use this data to get the tensor data to process
  // Since the most senario is process data in CPU
  // this function will return a pointer to cpu memory
  // buffer.
  // If the original data is on other device, the data
  // will copy to cpu store in `temporary_cpu_buffer`
  const void* CpuData() const;

  // void SetDataBuffer(const std::vector<int64_t>& new_shape, const FDDataType& data_type, void* data_buffer, bool copy = false, const Device& new_device = Device::CPU, int new_device_id = -1);
  // Set user memory buffer for Tensor, the memory is managed by
  // the user it self, but the Tensor will share the memory with user
  // So take care with the user buffer
  void SetExternalData(const std::vector<int64_t>& new_shape,
                       const FDDataType& data_type, void* data_buffer,
                       const Device& new_device = Device::CPU,
                       int new_device_id = -1);
  // Initialize Tensor
  // Include setting attribute for tensor
  // and allocate cpu memory buffer
  void Allocate(const std::vector<int64_t>& new_shape,
                const FDDataType& data_type,
                const std::string& tensor_name = "",
                const Device& new_device = Device::CPU);

  void Resize(size_t nbytes);

  void Resize(const std::vector<int64_t>& new_shape);

  void Resize(const std::vector<int64_t>& new_shape,
              const FDDataType& data_type, const std::string& tensor_name = "",
              const Device& new_device = Device::CPU);

  bool ReallocFn(size_t nbytes);

  void FreeFn();

  FDTensor() {}
  explicit FDTensor(const std::string& tensor_name);
  explicit FDTensor(const char* tensor_name);

  // Deep copy
  FDTensor(const FDTensor& other);
  // Move constructor
  FDTensor(FDTensor&& other);

  // Deep copy assignment
  FDTensor& operator=(const FDTensor& other);
  // Move assignment
  FDTensor& operator=(FDTensor&& other);

  // Scalar to FDTensor
  explicit FDTensor(const Scalar& scalar);

  ~FDTensor() { FreeFn(); }

  static void CopyBuffer(void* dst, const void* src, size_t nbytes,
                         const Device& device = Device::CPU,
                         bool is_pinned_memory = false);
};

}  // namespace fastdeploy