#pragma once

#include <c10/core/Device.h>
#include <c10/core/Layout.h>
#include <c10/core/MemoryFormat.h>
#include <c10/core/QScheme.h>
#include <c10/core/Stream.h>
#include <c10/core/Scalar.h>
#include <c10/core/ScalarType.h>
#include <c10/core/ScalarTypeToTypeMeta.h>
#include <c10/core/Storage.h>
#include <ATen/core/TensorAccessor.h>
#include <c10/core/TensorImpl.h>
#include <c10/core/UndefinedTensorImpl.h>
#include <c10/core/WrapDimMinimal.h>
#include <c10/util/Exception.h>
#include <c10/util/Deprecated.h>
#include <c10/util/Optional.h>
#include <c10/util/intrusive_ptr.h>
#include <ATen/core/DeprecatedTypePropertiesRegistry.h>
#include <ATen/core/DeprecatedTypeProperties.h>
#include <ATen/core/NamedTensor.h>
#include <ATen/core/QuantizerBase.h>
#include <torch/csrc/WindowsTorchApiMacro.h>

namespace caffe2 {
class Tensor;
}
namespace c10{
struct TensorOptions;
template<class T> class List;
}
namespace at {
struct Generator;
struct Type;
class DeprecatedTypeProperties;
class Tensor;
} // namespace at
namespace at {
namespace indexing {
struct TensorIndex;
} // namespace indexing
} // namespace at

namespace torch { namespace autograd {

struct Node;

}} // namespace torch::autograd

namespace at {

class Tensor;
using TensorList = ArrayRef<Tensor>;

using Stream = c10::Stream;

namespace impl {
inline bool variable_excluded_from_dispatch() {
#ifdef C10_MOBILE
  // Please read the comment in `VariableFallbackKernel.cpp` about the background of this change.
  return true;
#else
  TORCH_INTERNAL_ASSERT_DEBUG_ONLY(!c10::impl::tls_local_dispatch_key_set().excluded_.has(DispatchKey::Autograd));
  return c10::impl::tls_local_dispatch_key_set().excluded_.isSupersetOf(c10::autograd_dispatch_keyset);
#endif
}
}

// Tensor is a "generic" object holding a pointer to the underlying TensorImpl object, which
// has an embedded reference count. In this way, Tensor is similar to boost::intrusive_ptr.
//
// For example:
//
// void func(Tensor a) {
//   Tensor b = a;
//   ...
// }
//
// In this example, when we say Tensor b = a, we are creating a new object that points to the
// same underlying TensorImpl, and bumps its reference count. When b goes out of scope, the
// destructor decrements the reference count by calling release() on the TensorImpl it points to.
// The existing constructors, operator overloads, etc. take care to implement the correct semantics.
//
// Note that Tensor can also be NULL, i.e. it is not associated with any underlying TensorImpl, and
// special care must be taken to handle this.
class TORCH_API Tensor {
 public:
  Tensor(){};
  // This constructor should not be used by end users and is an implementation
  // detail invoked by autogenerated code.
  explicit Tensor(
      c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> tensor_impl)
      : impl_(std::move(tensor_impl)) {
    if (impl_.get() == nullptr) {
      throw std::runtime_error("TensorImpl with nullptr is not supported");
    }
  }
  Tensor(const Tensor&) = default;
  Tensor(Tensor&&) = default;


 public:
  // Creates a new wrapper from TensorImpl. Intentionally a free method because
  // it should be used with care. Checks necessary invariants
  static Tensor wrap_tensor_impl(
      c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> tensor_impl) {
    Tensor r(std::move(tensor_impl));
    r.enforce_invariants();
    return r;
  }

  int64_t dim() const {
    return impl_->dim();
  }
  int64_t storage_offset() const {
    return impl_->storage_offset();
  }

  Tensor contiguous(MemoryFormat memory_format=MemoryFormat::Contiguous) const {
    if (is_contiguous(memory_format)) {
      return *this;
    } else {
      return __dispatch_contiguous(memory_format);
    }
  }

  int64_t size(int64_t dim) const {
    // false is passed to maybe_wrap_dim so behavior is identical to array access (but with wrapping)
    dim = c10::maybe_wrap_dim(dim, this->dim(), false);
    return sizes()[dim];
  }

  int64_t stride(int64_t dim) const {
    // false is passed to maybe_wrap_dim so behavior is identical to array access (but with wrapping)
    dim = c10::maybe_wrap_dim(dim, this->dim(), false);
    return strides()[dim];
  }

  TensorImpl * unsafeGetTensorImpl() const {
    return impl_.get();
  }
  TensorImpl * unsafeReleaseTensorImpl() {
    return impl_.release();
  }
  const c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl>& getIntrusivePtr() const {
    return impl_;
  }

  c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> unsafeReleaseIntrusivePtr()  {
    return std::move(impl_);
  }

  bool defined() const {
    return impl_;
  }

  void reset() {
    impl_.reset();
  }

  // The following overloads are very intruiging.  Consider the following
  // program:
  //
  //    x[1] = 3;
  //
  // We would expect that the first entry of x is written to 3.  But how can we
  // actually achieve this?  x[1] evaluates to a tensor...
  //
  // The answer is, using a ref-qualifier.  x[1] is an rvalue, which cannot be
  // (profitably) assigned to in the traditional sense, so we overload
  // assignment to mean, "Actually, copy 3 into the tensor data."  This is done
  // with an rvalue-reference ref-qualified overload (the methods with && at the
  // end of their type.)
  //
  // There's one more fly in the ointment: We also want
  //
  //    Tensor x = y;
  //
  // to work, and we want it NOT to copy.  So we need a traditional operator=
  // overload.  But we MUST specify a mutable lvalue ref-qualifier, to
  // disambiguate the traditional overload from the rvalue-reference
  // ref-qualified overload.  Otherwise, it will be ambiguous, because
  // a non ref-qualified method is eligible for all situations.

  // Unfortunately, we have to write these constructors out manually
  // to work around an MSVC bug:
  //    error C2580: 'at::Tensor &at::Tensor::operator =(const at::Tensor &) &':
  //    multiple versions of a defaulted special member functions are not allowed
  // Tensor& operator=(const Tensor&) & = default;
  // Tensor& operator=(Tensor&&) & = default;

  // Also MSVC will wrongly issue the following warning with the aforementioned fix
  //    warning C4522: 'at::Tensor': multiple assignment operators specified
  // Let's just skip the warning.

  #ifdef _MSC_VER
  #pragma warning( push )
  #pragma warning( disable : 4522 )
  #endif

  Tensor& operator=(const Tensor& x) & {
    impl_ = x.impl_;
    return *this;
  }
  Tensor& operator=(Tensor&& x) & {
    impl_ = std::move(x.impl_);
    return *this;
  }

  Tensor& operator=(Scalar v) &&;
  Tensor& operator=(const Tensor&) &&;
  Tensor& operator=(Tensor&&) &&;

  bool is_same(const Tensor& other) const noexcept {
    return impl_ == other.impl_;
  }
  size_t use_count() const noexcept {
    return impl_.use_count();
  }
  size_t weak_use_count() const noexcept {
    return impl_.weak_use_count();
  }

  std::string toString() const;

  IntArrayRef sizes() const {
    return impl_->sizes();
  }
  IntArrayRef strides() const {
    return impl_->strides();
  }
  // See impl::get_opt_names in ATen/NamedTensor.h for docs.
  c10::optional<DimnameList> opt_names() const {
    return impl::get_opt_names(unsafeGetTensorImpl());
  }
  // See impl::get_names in ATen/NamedTensor.h for docs.
  DimnameList names() const {
    return impl::get_names(unsafeGetTensorImpl());
  }
  int64_t ndimension() const {
    return dim();
  }

  bool is_contiguous(at::MemoryFormat memory_format=at::MemoryFormat::Contiguous) const {
    return impl_->is_contiguous(memory_format);
  }

  bool is_non_overlapping_and_dense() const {
    return impl_->is_non_overlapping_and_dense();
  }

  at::MemoryFormat suggest_memory_format(
      bool channels_last_strides_exact_match = false) const {
    // Setting channels_last_strides_exact_match to true forces function to
    // check 0,1 - sized dimension strides.
    if (!is_mkldnn() && !is_sparse()) {
      if (impl_->is_strides_like_channels_last()) {
        if (!channels_last_strides_exact_match ||
            get_channels_last_strides_2d(sizes()) == strides()) {
          return at::MemoryFormat::ChannelsLast;
        }
      }
      else if (impl_->is_strides_like_channels_last_3d()) {
        if (!channels_last_strides_exact_match ||
            get_channels_last_strides_3d(sizes()) == strides()) {
          return at::MemoryFormat::ChannelsLast3d;
        }
      }
    }
    return at::MemoryFormat::Contiguous;
  }

  // Total bytes consumed by the "view" of elements of the array.  Does not
  // include size of metadata.  The number reported here does not necessarily
  // correspond to the true physical memory consumed by a tensor; instead,
  // it reports the memory the tensor would take *if* it were contiguous.
  // Defined to be numel() * itemsize()
  size_t nbytes() const {
    TORCH_CHECK(layout () != at::kSparse,
                "nbytes is not defined for sparse tensors.  If you want the size of the constituent " \
                "tensors, add the nbytes of the indices and values.  If you want the size of the  " \
                "equivalent dense tensor, multiply numel() by element_size()");
    return impl_->numel() * impl_->itemsize();
  }

  int64_t numel() const {
    return impl_->numel();
  }

  // Length of one array element in bytes.  This is the traditional
  // Numpy naming.
  size_t itemsize() const {
    return impl_->itemsize();
  }

  // Same as itemsize().  This is the PyTorch naming.
  int64_t element_size() const {
    return static_cast<int64_t>(impl_->itemsize());
  }

  C10_DEPRECATED_MESSAGE("Tensor.type() is deprecated. Instead use Tensor.options(), which in many cases (e.g. in a constructor) is a drop-in replacement. If you were using data from type(), that is now available from Tensor itself, so instead of tensor.type().scalar_type(), use tensor.scalar_type() instead and instead of tensor.type().backend() use tensor.device().")
  DeprecatedTypeProperties & type() const {
    return globalDeprecatedTypePropertiesRegistry().getDeprecatedTypeProperties(
        dispatchKeyToBackend(legacyExtractDispatchKey(key_set())),
        scalar_type());
  }
  DispatchKeySet key_set() const {
    return impl_->key_set();
  }
  ScalarType scalar_type() const {
    return typeMetaToScalarType(impl_->dtype());
  }
  bool has_storage() const {
    return defined() && impl_->has_storage();
  }
  const Storage& storage() const {
    return impl_->storage();
  }
  bool is_alias_of(const at::Tensor& other) const{
    return impl_->storage().is_alias_of(other.storage());
  }
  Tensor toType(ScalarType t) const;
  Tensor toBackend(Backend b) const;

  C10_DEPRECATED_MESSAGE("Tensor.is_variable() is deprecated; everything is a variable now. (If you want to assert that variable has been appropriately handled already, use at::impl::variable_excluded_from_dispatch())")
  bool is_variable() const noexcept {
    return !at::impl::variable_excluded_from_dispatch();
  }

  /// Returns a `Tensor`'s layout. Defined in Type.h
  Layout layout() const noexcept;

  /// Returns a `Tensor`'s dtype (`TypeMeta`). Defined in TensorMethods.cpp
  caffe2::TypeMeta dtype() const noexcept;

  /// Returns a `Tensor`'s device.
  Device device() const;

  /// Returns a `Tensor`'s device index.
  int64_t get_device() const;

  /// Returns if a `Tensor` has CUDA backend.
  bool is_cuda() const;