Variable.h

// Copyright 2024, UChicago Argonne, LLC
// All Rights Reserved
// Software Name: NEML2 -- the New Engineering material Model Library, version 2
// By: Argonne National Laboratory
// OPEN SOURCE LICENSE (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.
 
#pragma once
 
#include "neml2/models/LabeledAxisAccessor.h"
#include "neml2/base/DependencyResolver.h"
#include "neml2/tensors/Tensor.h"
#include "neml2/models/map_types.h"
 
namespace neml2
{
// Forward declarations
class Model;
class Derivative;
enum class TensorType : int8_t;
 
class VariableBase
{
public:
  VariableBase() = default;
 
  VariableBase(const VariableBase &) = delete;
  VariableBase(VariableBase &&) = delete;
  VariableBase & operator=(const VariableBase &) = delete;
  VariableBase & operator=(VariableBase &&) = delete;
  virtual ~VariableBase() = default;
 
  VariableBase(VariableName name_in, Model * owner, TensorShapeRef list_shape);
 
  const VariableName & name() const { return _name; }
 
  const Model & owner() const;
  Model & owner();
 
  virtual TensorType type() const = 0;
 
  bool is_state() const;
  bool is_old_state() const;
  bool is_force() const;
  bool is_old_force() const;
  bool is_residual() const;
  bool is_parameter() const;
  bool is_solve_dependent() const;
  // Note that the check depends on whether we are currently solving nonlinear system
  bool is_dependent() const;
 
  // These methods mirror TensorBase
  torch::TensorOptions options() const { return tensor().options(); }
  torch::Dtype scalar_type() const { return tensor().scalar_type(); }
  torch::Device device() const { return tensor().device(); }
  Size dim() const { return tensor().dim(); }
  TensorShapeRef sizes() const { return tensor().sizes(); }
  Size size(Size dim) const { return tensor().size(dim); }
  bool batched() const { return tensor().batched(); }
  Size batch_dim() const { return tensor().batch_dim(); }
  Size list_dim() const { return Size(list_sizes().size()); }
  Size base_dim() const { return Size(base_sizes().size()); }
  TraceableTensorShape batch_sizes() const { return tensor().batch_sizes(); }
  TensorShapeRef list_sizes() const { return _list_sizes; }
  virtual TensorShapeRef base_sizes() const = 0;
  TraceableSize batch_size(Size dim) const { return tensor().batch_size(dim); }
  Size base_size(Size dim) const { return base_sizes()[dim]; }
  Size list_size(Size dim) const { return list_sizes()[dim]; }
  Size base_storage() const { return utils::storage_size(base_sizes()); }
  Size assembly_storage() const
  {
    return utils::storage_size(list_sizes()) * utils::storage_size(base_sizes());
  }
 
  virtual std::unique_ptr<VariableBase> clone(const VariableName & name = {},
                                              Model * owner = nullptr) const = 0;
 
  virtual void ref(const VariableBase & other, bool ref_is_mutable = false) = 0;
 
  virtual const VariableBase * ref() const = 0;
 
  virtual bool owning() const = 0;
 
  virtual void zero(const torch::TensorOptions & options) = 0;
 
  virtual void set(const Tensor & val) = 0;
 
  virtual Tensor get() const = 0;
 
  virtual Tensor tensor() const = 0;
 
  bool requires_grad() const { return tensor().requires_grad(); }
 
  virtual void requires_grad_(bool req = true) = 0;
 
  virtual void operator=(const Tensor & val) = 0;
 
  Derivative d(const VariableBase & var);
 
  Derivative d(const VariableBase & var1, const VariableBase & var2);
 
  void request_AD(const VariableBase & u);
  void request_AD(const std::vector<const VariableBase *> & us);
 
  void request_AD(const VariableBase & u1, const VariableBase & u2);
  void request_AD(const std::vector<const VariableBase *> & u1s,
                  const std::vector<const VariableBase *> & u2s);
 
  const ValueMap & derivatives() const { return _derivs; }
  ValueMap & derivatives() { return _derivs; }
 
  const DerivMap & second_derivatives() const { return _sec_derivs; }
  DerivMap & second_derivatives() { return _sec_derivs; }
 
  virtual void clear();
 
  void apply_chain_rule(const DependencyResolver<Model, VariableName> &);
 
  void apply_second_order_chain_rule(const DependencyResolver<Model, VariableName> &);
 
  const VariableName _name = {};
 
  Model * const _owner = nullptr;
 
private:
  ValueMap total_derivatives(const DependencyResolver<Model, VariableName> & dep,
                             Model * model,
                             const VariableName & yvar) const;
 
  DerivMap total_second_derivatives(const DependencyResolver<Model, VariableName> & dep,
                                    Model * model,
                                    const VariableName & yvar) const;
 
  const TensorShape _list_sizes = {};
 
  ValueMap _derivs;
 
  DerivMap _sec_derivs;
};
 
template <typename T>
class Variable : public VariableBase
{
public:
  template <typename T2 = T, typename = typename std::enable_if_t<!std::is_same_v<Tensor, T2>>>
  Variable(VariableName name_in, Model * owner, TensorShapeRef list_shape)
    : VariableBase(std::move(name_in), owner, list_shape),
      _base_sizes(T::const_base_sizes),
      _ref(nullptr),
      _ref_is_mutable(false)
  {
  }
 
  template <typename T2 = T, typename = typename std::enable_if_t<std::is_same_v<Tensor, T2>>>
  Variable(VariableName name_in,
           Model * owner,
           TensorShapeRef list_shape,
           TensorShapeRef base_shape)
    : VariableBase(std::move(name_in), owner, list_shape),
      _base_sizes(base_shape),
      _ref(nullptr),
      _ref_is_mutable(false)
  {
  }
 
  TensorType type() const override;
 
  TensorShapeRef base_sizes() const override { return _base_sizes; }
 
  std::unique_ptr<VariableBase> clone(const VariableName & name = {},
                                      Model * owner = nullptr) const override;
 
  void ref(const VariableBase & var, bool ref_is_mutable = false) override;
 
  const VariableBase * ref() const override { return _ref ? _ref->ref() : this; }
 
  bool owning() const override { return !_ref; }
 
  void zero(const torch::TensorOptions & options) override;
 
  void set(const Tensor & val) override;
 
  Tensor get() const override { return tensor().base_flatten(); }
 
  Tensor tensor() const override;
 
  void requires_grad_(bool req = true) override;
 
  void operator=(const Tensor & val) override;
 
  const T & value() const { return owning() ? _value : _ref->value(); }
 
  T operator-() const { return -value(); }
 
  operator T() const { return value(); }
 
  template <typename T2 = T, typename = typename std::enable_if_t<!std::is_same_v<T2, Tensor>>>
  operator Tensor() const
  {
    return value();
  }
 
  void clear() override;
 
protected:
  const TensorShape _base_sizes;
 
  const Variable<T> * _ref;
 
  bool _ref_is_mutable;
 
  T _value;
};
 
class Derivative
{
public:
  Derivative()
    : _base_sizes({}),
      _deriv(nullptr)
  {
  }
 
  Derivative(TensorShapeRef base_sizes, Tensor * deriv)
    : _base_sizes(base_sizes),
      _deriv(deriv)
  {
  }
 
  void operator=(const Tensor & val);
 
private:
  const TensorShape _base_sizes;
 
  Tensor * const _deriv;
};
 
// Everything below is just for convenience: We just forward operations to the the variable values
// so that we can do
//
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
//   var4 = (var1 - var2) * var3
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
//
// instead of the (ugly?) expression below
//
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
//   var4 = (var1.v - var2.v) * var3.v
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
#define FWD_VARIABLE_BINARY_OP(op)                                                                 \
  template <typename T1,                                                                           \
            typename T2,                                                                           \
            typename = typename std::enable_if_t<std::is_base_of_v<VariableBase, T1> ||            \
                                                 std::is_base_of_v<VariableBase, T2>>>             \
  auto op(const T1 & a, const T2 & b)                                                              \
  {                                                                                                \
    if constexpr (std::is_base_of_v<VariableBase, T1> && std::is_base_of_v<VariableBase, T2>)      \
      return op(a.value(), b.value());                                                             \
                                                                                                   \
    if constexpr (std::is_base_of_v<VariableBase, T1> && !std::is_base_of_v<VariableBase, T2>)     \
      return op(a.value(), b);                                                                     \
                                                                                                   \
    if constexpr (!std::is_base_of_v<VariableBase, T1> && std::is_base_of_v<VariableBase, T2>)     \
      return op(a, b.value());                                                                     \
  }                                                                                                \
  static_assert(true)
FWD_VARIABLE_BINARY_OP(operator+);
FWD_VARIABLE_BINARY_OP(operator-);
FWD_VARIABLE_BINARY_OP(operator*);
FWD_VARIABLE_BINARY_OP(operator/);
}