//===- TFUtils.h - utilities for tensorflow C API ---------------*- C++ -*-===//
 
//
 
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 
// See https://llvm.org/LICENSE.txt for license information.
 
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
//
 
//===----------------------------------------------------------------------===//
 
//
 
#ifndef LLVM_ANALYSIS_UTILS_TFUTILS_H
 
#define LLVM_ANALYSIS_UTILS_TFUTILS_H
 
 
 
#include "llvm/Config/llvm-config.h"
 
 
 
#ifdef LLVM_HAVE_TFLITE
 
#include "llvm/ADT/StringMap.h"
 
#include "llvm/Analysis/TensorSpec.h"
 
#include "llvm/IR/LLVMContext.h"
 
#include "llvm/Support/JSON.h"
 
 
 
#include <memory>
 
#include <vector>
 
 
 
namespace llvm {
 
 
 
/// Load a SavedModel, find the given inputs and outputs, and setup storage
 
/// for input tensors. The user is responsible for correctly dimensioning the
 
/// input tensors and setting their values before calling evaluate().
 
/// To initialize:
 
/// - construct the object
 
/// - initialize the input tensors using initInput. Indices must correspond to
 
///   indices in the InputNames used at construction.
 
/// To use:
 
/// - set input values by using getInput to get each input tensor, and then
 
///   setting internal scalars, for all dimensions (tensors are row-major:
 
///   https://github.com/tensorflow/tensorflow/blob/r1.5/tensorflow/c/c_api.h#L205)
 
/// - call evaluate. The input tensors' values are not consumed after this, and
 
///   may still be read.
 
/// - use the outputs in the output vector
 
class TFModelEvaluatorImpl;
 
class EvaluationResultImpl;
 
 
 
class TFModelEvaluator final {
 
public:
 
  /// The result of a model evaluation. Handles the lifetime of the output
 
  /// tensors, which means that their values need to be used before
 
  /// the EvaluationResult's dtor is called.
 
  class EvaluationResult {
 
  public:
 
    EvaluationResult(const EvaluationResult &) = delete;
 
    EvaluationResult &operator=(const EvaluationResult &Other) = delete;
 
 
 
    EvaluationResult(EvaluationResult &&Other);
 
    EvaluationResult &operator=(EvaluationResult &&Other);
 
 
 
    ~EvaluationResult();
 
 
 
    /// Get a (const) pointer to the first element of the tensor at Index.
 
    template <typename T> T *getTensorValue(size_t Index) {
 
      return static_cast<T *>(getUntypedTensorValue(Index));
 
    }
 
 
 
    template <typename T> const T *getTensorValue(size_t Index) const {
 
      return static_cast<T *>(getUntypedTensorValue(Index));
 
    }
 
 
 
    /// Get a (const) pointer to the untyped data of the tensor.
 
    void *getUntypedTensorValue(size_t Index);
 
    const void *getUntypedTensorValue(size_t Index) const;
 
 
 
  private:
 
    friend class TFModelEvaluator;
 
    EvaluationResult(std::unique_ptr<EvaluationResultImpl> Impl);
 
    std::unique_ptr<EvaluationResultImpl> Impl;
 
  };
 
 
 
  TFModelEvaluator(StringRef SavedModelPath,
 
                   const std::vector<TensorSpec> &InputSpecs,
 
                   const std::vector<TensorSpec> &OutputSpecs,
 
                   const char *Tags = "serve");
 
 
 
  ~TFModelEvaluator();
 
  TFModelEvaluator(const TFModelEvaluator &) = delete;
 
  TFModelEvaluator(TFModelEvaluator &&) = delete;
 
 
 
  /// Evaluate the model, assuming it is valid. Returns std::nullopt if the
 
  /// evaluation fails or the model is invalid, or an EvaluationResult
 
  /// otherwise. The inputs are assumed to have been already provided via
 
  /// getInput(). When returning std::nullopt, it also invalidates this object.
 
  std::optional<EvaluationResult> evaluate();
 
 
 
  /// Provides access to the input vector.
 
  template <typename T> T *getInput(size_t Index) {
 
    return static_cast<T *>(getUntypedInput(Index));
 
  }
 
 
 
  /// Returns true if the tensorflow model was loaded successfully, false
 
  /// otherwise.
 
  bool isValid() const { return !!Impl; }
 
 
 
  /// Untyped access to input.
 
  void *getUntypedInput(size_t Index);
 
 
 
private:
 
  std::unique_ptr<TFModelEvaluatorImpl> Impl;
 
};
 
 
 
} // namespace llvm
 
 
 
#endif // LLVM_HAVE_TFLITE
 
#endif // LLVM_ANALYSIS_UTILS_TFUTILS_H