#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/Scalar.h>
#include <c10/core/ScalarType.h>
#include <c10/core/Storage.h>
#include <ATen/core/TensorAccessor.h>
#include <c10/core/TensorImpl.h>
#include <c10/core/UndefinedTensorImpl.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/EnableNamedTensor.h>
#include <ATen/core/NamedTensor.h>

namespace caffe2 {
class Tensor;
}
namespace c10{
struct TensorOptions;
}
namespace at {
struct Generator;
struct Type;
class DeprecatedTypeProperties;
class Tensor;
} // namespace at

namespace at {

class Tensor;
using TensorList = ArrayRef<Tensor>;

struct Quantizer;
// This is temporary typedef to enable Quantizer in aten native function API
// we'll remove them when we are actually exposing Quantizer class
// to frontend
using QuantizerPtr = c10::intrusive_ptr<Quantizer>;
using ConstQuantizerPtr = const c10::intrusive_ptr<Quantizer>&;

// 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 CAFFE2_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();
  }

  TensorImpl * unsafeGetTensorImpl() const {
    return impl_.get();
  }
  TensorImpl * unsafeReleaseTensorImpl() {
    return impl_.release();
  }
  const c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl>& getIntrusivePtr() const {
    return 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;
  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();
  }
#ifdef BUILD_NAMEDTENSOR
  // See impl::get_opt_names in ATen/NamedTensor.h for docs.
  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());
  }
#endif
  int64_t ndimension() const {
    return dim();
  }
  bool is_contiguous(at::MemoryFormat memory_format=at::MemoryFormat::Contiguous) const {
    return impl_->is_contiguous(memory_format);
  }

  at::MemoryFormat suggest_memory_format() const {
    if (impl_->is_strides_like_channels_last()) {
      return at::MemoryFormat::ChannelsLast;
    }
    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 {
    return impl_->numel() * impl_->itemsize();
  }

  // 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.
  size_t element_size() const {
    return impl_->itemsize();
  }

  DeprecatedTypeProperties & type() const {
    return globalDeprecatedTypePropertiesRegistry().getDeprecatedTypeProperties(
        tensorTypeIdToBackend(legacyExtractTypeId(type_set())),
        scalar_type(),
        is_variable());
  }
  TensorTypeSet type_set() const {
    return impl_->type_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;

  /// Returns true if the `Tensor` is actually a `torch::autograd::Variable`.
  /// Defined in Type.h because of include order issues.
  bool is_variable() const noexcept {
    return impl_->is_variable();
  }

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

  /// Returns a `Tensor`'s dtype (`TypeMeta`). Defined in TensorMethods.h
  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;

  /// Returns if a `Tensor` has HIP backend.
  bool is_hip() const;

  /// Returns if a `Tensor` has sparse backend.
  bool is_sparse() const;

  /// Returns if a `Tensor` is mkldnn tensor.
  bool is_mkldnn() const;

  /// Returns if a `Tensor` has quantized backend.
  bool is_quantized() const;

  /// If a tensor is a quantized tensor, returns its quantizer
  /// TODO: it's not in native_functions.yaml yet as it's not exposed to python
  QuantizerPtr quantizer() const;

#ifdef BUILD_NAMEDTENSOR
  /// Returns if a `Tensor` has any dimension names
  bool has_names() const;

  /// Returns a `Tensor`'s dimension names data structure
  const NamedTensorMeta* get_named_tensor_meta() const;
  NamedTensorMeta* get_named_tensor_meta();
#endif

  /// Returns the `TensorOptions` corresponding to this `Tensor`. Defined in
  /// TensorOptions.h.
  TensorOptions options() const;

  void* data_ptr() const {
    return this->unsafeGetTensorImpl()->data();
  }

  template <typename T>
  T * data_ptr() const;

  template<typename T>
  C10_DEPRECATED_MESSAGE("Tensor.data<T>() is deprecated. Please use Tensor.data_ptr<T>() instead.")
  T * data() const {
    return data_ptr<T>();
  }

  template <typename T>
  T item() const;

  // Purposely not defined here to avoid inlining
  void print() const;

  // Return a `TensorAccessor` for CPU `Tensor`s. You have to specify scalar type and
  // dimension.
  template<typename T, size_t N>
  TensorAccessor<T,N> accessor() const& {
    static_assert(N > 0, "accessor is used for indexing tensor, for scalars use *data_ptr<T>()");
    TORCH_CHECK(dim() == N, "expected ", N, " dims but tensor has ", dim());
    return TensorAccessor<T,N>(data_ptr<T>(),sizes().data(),strides().data());
  }
  template<typename T, size_t N>
  TensorAccessor<T,N> accessor() && = delete;

  // Return a `GenericPackedTensorAccessor` for CUDA `Tensor`s. You have to specify scalar type and
  // dimension. You can optionally specify RestrictPtrTraits as a template parameter to
  // cast the data pointer to a __restrict__ pointer.
  // In order to use this, your CUDA kernel has to take a corresponding GenericPackedTensorAccessor
  // as an argument.
  template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits, typename index_t = int64_t>
  GenericPackedTensorAccessor<T,N,PtrTraits,index_t> generic_packed_accessor() const& {
    static_assert(N > 0, "accessor is used for indexing tensor, for scalars use *data_ptr<T>()");
    TORCH_CHECK(dim() == N, "expected ", N, " dims but tensor has ", dim());
    return GenericPackedTensorAccessor<T,N,PtrTraits,index_t>(static_cast<typename PtrTraits<T>::PtrType>(data_ptr<T>()),sizes().data(),strides().data());
  }
  template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits, typename index_t = int64_t>
  GenericPackedTensorAccessor<T,N> generic_packed_accessor() && = delete;

  template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits>
  PackedTensorAccessor32<T,N,PtrTraits> packed_accessor32() const& {
    return generic_packed_accessor<T,N,PtrTraits,int32_t>();
  }
  template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits>
  PackedTensorAccessor32<T,N,PtrTraits> packed_accessor32() && = delete;

  template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits>
  PackedTensorAccessor64<T,N,PtrTraits> packed_accessor64() const& {
    return generic_packed_accessor<T,N,PtrTraits,int64_t>();
  }
  template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits>
  PackedTensorAccessor64<T,N,PtrTraits> packed_accessor64() && = delete;

  template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits, typename index_t = int64_t>
  C10_DEPRECATED_MESSAGE("packed_accessor is deprecated, use packed_accessor32 or packed_accessor64 instead")
  GenericPackedTensorAccessor<T,N,PtrTraits,index_t> packed_accessor() const & {
    return generic_packed_accessor<T,N,PtrTraits,index_t>();
  }
  template<typename T, size_t N, template <typename U> class PtrTraits = DefaultPtrTraits, typename index_t = int64_t>
  C10_DEPRECATED_MESSAGE("packed_accessor is deprecated, use packed_accessor32 or packed_accessor64 instead")
  GenericPackedTensorAccessor<T,N,PtrTraits,index_t> packed_accessor() && = delete;

  Tensor operator-() const;
  Tensor& operator+=(const Tensor & other);
  Tensor& operator+=(Scalar other);
  Tensor& operator-=(const Tensor & other);
  Tensor& operator-=(Scalar other);
  Tensor& operator*=(const Tensor & other);
  Tensor& operator*=(Scalar other);
  Tensor& operator/=(const Tensor & other);
  Tensor& operator/=(Scalar other);
  Tensor operator[](Scalar index) const;
  Tensor operator[](Tensor index) const;
  Tensor operator[](int64_t index) const;

  Tensor cpu() const;
  Tensor cuda() const;
  Tensor hip() const;

  // ~~~~~ Autograd API ~~~~~

  Tensor& set_requires_grad(bool requires_grad) {
    impl_->set_requires_grad(requires_grad);
    return *this;
  }
  bool requires_grad() const {
    return impl_->requires_grad();
  }

  Tensor& grad() {
    return impl_->grad();
  }
  const Tensor& grad() const {
    return impl_->grad();
  }

  // STOP.  Thinking of adding a method here, which only makes use
  // of other ATen methods?  Define it in native_functions.yaml.

  //example
  //Tensor * add(Tensor & b);
  void backward(const Tensor & gradient={}, bool keep_graph=false, bool create_graph=false) const;
  void set_data(const Tensor & new_data) const;
  Tensor data() const;
  bool is_leaf() const;
  int64_t output_nr() const;
  int64_t _version() const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor & rename_(c10::optional<DimnameList> names) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor rename(c10::optional<DimnameList> names) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor align_to(DimnameList names) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor align_as(const Tensor & other) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor refine_names(DimnameList names) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor unflatten(Dimname dim, IntArrayRef sizes, DimnameList names) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor unflatten(int64_t dim, IntArrayRef sizes, DimnameList names) const;
  #endif
  Tensor abs() const;
  Tensor & abs_() const;
  Tensor acos() const;
  Tensor & acos_() const;
  Tensor add(const Tensor & other, Scalar alpha=1) const;
  Tensor & add_(const Tensor & other, Scalar alpha=1) const;
  Tensor add(Scalar other, Scalar alpha=1) const;
  Tensor & add_(Scalar other, Scalar alpha=1) const;
  Tensor addmv(const Tensor & mat, const Tensor & vec, Scalar beta=1, Scalar alpha=1) const;
  Tensor & addmv_(const Tensor & mat, const Tensor & vec, Scalar beta=1, Scalar alpha=1) const;
  Tensor addr(const Tensor & vec1, const Tensor & vec2, Scalar beta=1, Scalar alpha=1) const;
  Tensor & addr_(const Tensor & vec1, const Tensor & vec2, Scalar beta=1, Scalar alpha=1) const;
  Tensor all(int64_t dim, bool keepdim=false) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor all(Dimname dim, bool keepdim=false) const;
  #endif
  bool allclose(const Tensor & other, double rtol=1e-05, double atol=1e-08, bool equal_nan=false) const;
  Tensor any(int64_t dim, bool keepdim=false) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor any(Dimname dim, bool keepdim=false) const;
  #endif
  Tensor argmax(c10::optional<int64_t> dim=c10::nullopt, bool keepdim=false) const;
  Tensor argmin(c10::optional<int64_t> dim=c10::nullopt, bool keepdim=false) const;
  Tensor as_strided(IntArrayRef size, IntArrayRef stride, c10::optional<int64_t> storage_offset=c10::nullopt) const;
  Tensor & as_strided_(IntArrayRef size, IntArrayRef stride, c10::optional<int64_t> storage_offset=c10::nullopt) const;
  Tensor asin() const;
  Tensor & asin_() const;
  Tensor atan() const;
  Tensor & atan_() const;
  Tensor baddbmm(const Tensor & batch1, const Tensor & batch2, Scalar beta=1, Scalar alpha=1) const;
  Tensor & baddbmm_(const Tensor & batch1, const Tensor & batch2, Scalar beta=1, Scalar alpha=1) const;
  Tensor bernoulli(Generator * generator=nullptr) const;
  Tensor & bernoulli_(const Tensor & p, Generator * generator=nullptr) const;
  Tensor & bernoulli_(double p=0.5, Generator * generator=nullptr) const;
  Tensor bernoulli(double p, Generator * generator=nullptr) const;
  Tensor bincount(const Tensor & weights={}, int64_t minlength=0) const;
  Tensor bitwise_not() const;
  Tensor & bitwise_not_() const;
  Tensor logical_not() const;
  Tensor & logical_not_() const;
  Tensor logical_xor(const Tensor & other) const;
  Tensor & logical_xor_(const Tensor & other) const;
  Tensor bmm(const Tensor & mat2) const;
  Tensor ceil() const;
  Tensor & ceil_() const;
  std::vector<Tensor> chunk(int64_t chunks, int64_t dim=0) const;
  Tensor clamp(c10::optional<Scalar> min=c10::nullopt, c10::optional<Scalar> max=c10::nullopt) const;
  Tensor & clamp_(c10::optional<Scalar> min=c10::nullopt, c10::optional<Scalar> max=c10::nullopt) const;
  Tensor clamp_max(Scalar max) const;
  Tensor & clamp_max_(Scalar max) const;
  Tensor clamp_min(Scalar min) const;
  Tensor & clamp_min_(Scalar min) const;
  Tensor contiguous(MemoryFormat memory_format=MemoryFormat::Contiguous) const;
  Tensor & copy_(const Tensor & src, bool non_blocking=false) const;
  Tensor cos() const;
  Tensor & cos_() const;
  Tensor cosh() const;
  Tensor & cosh_() const;
  Tensor cumsum(int64_t dim, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor cumsum(Dimname dim, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #endif
  Tensor cumprod(int64_t dim, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor cumprod(Dimname dim, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #endif
  Tensor det() const;
  Tensor diag_embed(int64_t offset=0, int64_t dim1=-2, int64_t dim2=-1) const;
  Tensor diagflat(int64_t offset=0) const;
  Tensor diagonal(int64_t offset=0, int64_t dim1=0, int64_t dim2=1) const;
  Tensor & fill_diagonal_(Scalar fill_value, bool wrap=false) const;
  Tensor div(const Tensor & other) const;
  Tensor & div_(const Tensor & other) const;
  Tensor div(Scalar other) const;
  Tensor & div_(Scalar other) const;
  Tensor dot(const Tensor & tensor) const;
  Tensor new_empty(IntArrayRef size, const TensorOptions & options={}) const;
  Tensor new_full(IntArrayRef size, Scalar fill_value, const TensorOptions & options={}) const;
  Tensor & resize_(IntArrayRef size) const;
  Tensor erf() const;
  Tensor & erf_() const;
  Tensor erfc() const;
  Tensor & erfc_() const;
  Tensor exp() const;
  Tensor & exp_() const;
  Tensor expm1() const;
  Tensor & expm1_() const;
  Tensor expand(IntArrayRef size, bool implicit=false) const;
  Tensor expand_as(const Tensor & other) const;
  Tensor flatten(int64_t start_dim=0, int64_t end_dim=-1) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor flatten(int64_t start_dim, int64_t end_dim, Dimname out_dim) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor flatten(Dimname start_dim, Dimname end_dim, Dimname out_dim) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor flatten(DimnameList dims, Dimname out_dim) const;
  #endif
  Tensor & fill_(Scalar value) const;
  Tensor & fill_(const Tensor & value) const;
  Tensor floor() const;
  Tensor & floor_() const;
  Tensor frac() const;
  Tensor & frac_() const;
  Tensor ger(const Tensor & vec2) const;
  Tensor fft(int64_t signal_ndim, bool normalized=false) const;
  Tensor ifft(int64_t signal_ndim, bool normalized=false) const;
  Tensor rfft(int64_t signal_ndim, bool normalized=false, bool onesided=true) const;
  Tensor irfft(int64_t signal_ndim, bool normalized=false, bool onesided=true, IntArrayRef signal_sizes={}) const;
  Tensor index(TensorList indices) const;
  Tensor & index_copy_(int64_t dim, const Tensor & index, const Tensor & source) const;
  Tensor index_copy(int64_t dim, const Tensor & index, const Tensor & source) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor & index_copy_(Dimname dim, const Tensor & index, const Tensor & source) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor index_copy(Dimname dim, const Tensor & index, const Tensor & source) const;
  #endif
  Tensor & index_put_(TensorList indices, const Tensor & values, bool accumulate=false) const;
  Tensor index_put(TensorList indices, const Tensor & values, bool accumulate=false) const;
  Tensor inverse() const;
  Tensor isclose(const Tensor & other, double rtol=1e-05, double atol=1e-08, bool equal_nan=false) const;
  bool is_distributed() const;
  bool is_floating_point() const;
  bool is_complex() const;
  bool is_nonzero() const;
  bool is_same_size(const Tensor & other) const;
  bool is_signed() const;
  std::tuple<Tensor,Tensor> kthvalue(int64_t k, int64_t dim=-1, bool keepdim=false) const;
  #ifdef BUILD_NAMEDTENSOR
  std::tuple<Tensor,Tensor> kthvalue(int64_t k, Dimname dim, bool keepdim=false) const;
  #endif
  Tensor log() const;
  Tensor & log_() const;
  Tensor log10() const;
  Tensor & log10_() const;
  Tensor log1p() const;
  Tensor & log1p_() const;
  Tensor log2() const;
  Tensor & log2_() const;
  Tensor logdet() const;
  Tensor log_softmax(int64_t dim, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor log_softmax(Dimname dim, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #endif
  Tensor logsumexp(IntArrayRef dim, bool keepdim=false) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor logsumexp(DimnameList dim, bool keepdim=false) const;
  #endif
  Tensor matmul(const Tensor & other) const;
  Tensor matrix_power(int64_t n) const;
  std::tuple<Tensor,Tensor> max(int64_t dim, bool keepdim=false) const;
  Tensor max_values(IntArrayRef dim, bool keepdim=false) const;
  #ifdef BUILD_NAMEDTENSOR
  std::tuple<Tensor,Tensor> max(Dimname dim, bool keepdim=false) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor max_values(DimnameList dim, bool keepdim=false) const;
  #endif
  Tensor mean(c10::optional<ScalarType> dtype=c10::nullopt) const;
  Tensor mean(IntArrayRef dim, bool keepdim=false, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor mean(DimnameList dim, bool keepdim=false, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #endif
  std::tuple<Tensor,Tensor> median(int64_t dim, bool keepdim=false) const;
  #ifdef BUILD_NAMEDTENSOR
  std::tuple<Tensor,Tensor> median(Dimname dim, bool keepdim=false) const;
  #endif
  std::tuple<Tensor,Tensor> min(int64_t dim, bool keepdim=false) const;
  Tensor min_values(IntArrayRef dim, bool keepdim=false) const;
  #ifdef BUILD_NAMEDTENSOR
  std::tuple<Tensor,Tensor> min(Dimname dim, bool keepdim=false) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor min_values(DimnameList dim, bool keepdim=false) const;
  #endif
  Tensor mm(const Tensor & mat2) const;
  std::tuple<Tensor,Tensor> mode(int64_t dim=-1, bool keepdim=false) const;
  #ifdef BUILD_NAMEDTENSOR
  std::tuple<Tensor,Tensor> mode(Dimname dim, bool keepdim=false) const;
  #endif
  Tensor mul(const Tensor & other) const;
  Tensor & mul_(const Tensor & other) const;
  Tensor mul(Scalar other) const;
  Tensor & mul_(Scalar other) const;
  Tensor mv(const Tensor & vec) const;
  Tensor mvlgamma(int64_t p) const;
  Tensor & mvlgamma_(int64_t p) const;
  Tensor narrow_copy(int64_t dim, int64_t start, int64_t length) const;
  Tensor narrow(int64_t dim, int64_t start, int64_t length) const;
  Tensor permute(IntArrayRef dims) const;
  Tensor numpy_T() const;
  bool is_pinned() const;
  Tensor pin_memory() const;
  Tensor pinverse(double rcond=1e-15) const;
  Tensor reciprocal() const;
  Tensor & reciprocal_() const;
  Tensor neg() const;
  Tensor & neg_() const;
  Tensor repeat(IntArrayRef repeats) const;
  Tensor repeat_interleave(const Tensor & repeats, c10::optional<int64_t> dim=c10::nullopt) const;
  Tensor repeat_interleave(int64_t repeats, c10::optional<int64_t> dim=c10::nullopt) const;
  Tensor reshape(IntArrayRef shape) const;
  Tensor reshape_as(const Tensor & other) const;
  Tensor round() const;
  Tensor & round_() const;
  Tensor relu() const;
  Tensor & relu_() const;
  Tensor prelu(const Tensor & weight) const;
  std::tuple<Tensor,Tensor> prelu_backward(const Tensor & grad_output, const Tensor & weight) const;
  Tensor hardshrink(Scalar lambd=0.5) const;
  Tensor hardshrink_backward(const Tensor & grad_out, Scalar lambd) const;
  Tensor rsqrt() const;
  Tensor & rsqrt_() const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor select(Dimname dim, int64_t index) const;
  #endif
  Tensor select(int64_t dim, int64_t index) const;
  Tensor sigmoid() const;
  Tensor & sigmoid_() const;
  Tensor sin() const;
  Tensor & sin_() const;
  Tensor sinh() const;
  Tensor & sinh_() const;
  Tensor detach() const;
  Tensor & detach_() const;
  int64_t size(int64_t dim) const;
  #ifdef BUILD_NAMEDTENSOR
  int64_t size(Dimname dim) const;
  #endif
  Tensor slice(int64_t dim=0, int64_t start=0, int64_t end=9223372036854775807, int64_t step=1) const;
  std::tuple<Tensor,Tensor> slogdet() const;
  Tensor smm(const Tensor & mat2) const;
  Tensor softmax(int64_t dim, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor softmax(Dimname dim, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #endif
  std::vector<Tensor> split(int64_t split_size, int64_t dim=0) const;
  std::vector<Tensor> split_with_sizes(IntArrayRef split_sizes, int64_t dim=0) const;
  Tensor squeeze() const;
  Tensor squeeze(int64_t dim) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor squeeze(Dimname dim) const;
  #endif
  Tensor & squeeze_() const;
  Tensor & squeeze_(int64_t dim) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor & squeeze_(Dimname dim) const;
  #endif
  Tensor sspaddmm(const Tensor & mat1, const Tensor & mat2, Scalar beta=1, Scalar alpha=1) const;
  Tensor stft(int64_t n_fft, c10::optional<int64_t> hop_length=c10::nullopt, c10::optional<int64_t> win_length=c10::nullopt, const Tensor & window={}, bool normalized=false, bool onesided=true) const;
  int64_t stride(int64_t dim) const;
  #ifdef BUILD_NAMEDTENSOR
  int64_t stride(Dimname dim) const;
  #endif
  Tensor sum(c10::optional<ScalarType> dtype=c10::nullopt) const;
  Tensor sum(IntArrayRef dim, bool keepdim=false, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor sum(DimnameList dim, bool keepdim=false, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #endif
  Tensor sum_to_size(IntArrayRef size) const;
  Tensor sqrt() const;
  Tensor & sqrt_() const;
  Tensor std(bool unbiased=true) const;
  Tensor std(IntArrayRef dim, bool unbiased=true, bool keepdim=false) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor std(DimnameList dim, bool unbiased=true, bool keepdim=false) const;
  #endif
  Tensor prod(c10::optional<ScalarType> dtype=c10::nullopt) const;
  Tensor prod(int64_t dim, bool keepdim=false, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor prod(Dimname dim, bool keepdim=false, c10::optional<ScalarType> dtype=c10::nullopt) const;
  #endif
  Tensor t() const;
  Tensor & t_() const;
  Tensor tan() const;
  Tensor & tan_() const;
  Tensor tanh() const;
  Tensor & tanh_() const;
  Tensor transpose(int64_t dim0, int64_t dim1) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor transpose(Dimname dim0, Dimname dim1) const;
  #endif
  Tensor & transpose_(int64_t dim0, int64_t dim1) const;
  Tensor flip(IntArrayRef dims) const;
  Tensor roll(IntArrayRef shifts, IntArrayRef dims={}) const;
  Tensor rot90(int64_t k=1, IntArrayRef dims={0,1}) const;
  Tensor trunc() const;
  Tensor & trunc_() const;
  Tensor type_as(const Tensor & other) const;
  Tensor unsqueeze(int64_t dim) const;
  Tensor & unsqueeze_(int64_t dim) const;
  Tensor var(bool unbiased=true) const;
  Tensor var(IntArrayRef dim, bool unbiased=true, bool keepdim=false) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor var(DimnameList dim, bool unbiased=true, bool keepdim=false) const;
  #endif
  Tensor view_as(const Tensor & other) const;
  Tensor where(const Tensor & condition, const Tensor & other) const;
  Tensor norm(c10::optional<Scalar> p, ScalarType dtype) const;
  Tensor norm(Scalar p=2) const;
  Tensor norm(c10::optional<Scalar> p, IntArrayRef dim, bool keepdim, ScalarType dtype) const;
  Tensor norm(c10::optional<Scalar> p, IntArrayRef dim, bool keepdim=false) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor norm(c10::optional<Scalar> p, DimnameList dim, bool keepdim, ScalarType dtype) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor norm(c10::optional<Scalar> p, DimnameList dim, bool keepdim=false) const;
  #endif
  Tensor clone() const;
  Tensor & resize_as_(const Tensor & the_template) const;
  Tensor pow(Scalar exponent) const;
  Tensor & zero_() const;
  Tensor sub(const Tensor & other, Scalar alpha=1) const;
  Tensor & sub_(const Tensor & other, Scalar alpha=1) const;
  Tensor sub(Scalar other, Scalar alpha=1) const;
  Tensor & sub_(Scalar other, Scalar alpha=1) const;
  Tensor addmm(const Tensor & mat1, const Tensor & mat2, Scalar beta=1, Scalar alpha=1) const;
  Tensor & addmm_(const Tensor & mat1, const Tensor & mat2, Scalar beta=1, Scalar alpha=1) const;
  Tensor & sparse_resize_(IntArrayRef size, int64_t sparse_dim, int64_t dense_dim) const;
  Tensor & sparse_resize_and_clear_(IntArrayRef size, int64_t sparse_dim, int64_t dense_dim) const;
  Tensor sparse_mask(const Tensor & mask) const;
  Tensor to_dense() const;
  int64_t sparse_dim() const;
  int64_t _dimI() const;
  int64_t dense_dim() const;
  int64_t _dimV() const;
  int64_t _nnz() const;
  Tensor coalesce() const;
  bool is_coalesced() const;
  Tensor _indices() const;
  Tensor _values() const;
  Tensor & _coalesced_(bool coalesced) const;
  Tensor indices() const;
  Tensor values() const;
  int64_t numel() const;
  std::vector<Tensor> unbind(int64_t dim=0) const;
  #ifdef BUILD_NAMEDTENSOR
  std::vector<Tensor> unbind(Dimname dim) const;
  #endif
  Tensor to_sparse(int64_t sparse_dim) const;
  Tensor to_sparse() const;
  Tensor to_mkldnn() const;
  Tensor dequantize() const;
  double q_scale() const;
  int64_t q_zero_point() const;
  Tensor q_per_channel_scales() const;
  Tensor q_per_channel_zero_points() const;
  int64_t q_per_channel_axis() const;
  Tensor int_repr() const;
  QScheme qscheme() const;
  Tensor to(const TensorOptions & options, bool non_blocking=false, bool copy=false) const;
  Tensor to(Device device, ScalarType dtype, bool non_blocking=false, bool copy=false) const;
  Tensor to(ScalarType dtype, bool non_blocking=false, bool copy=false) const;
  Tensor to(const Tensor & other, bool non_blocking=false, bool copy=false) const;
  Scalar item() const;
  Tensor & set_(Storage source) const;
  Tensor & set_(Storage source, int64_t storage_offset, IntArrayRef size, IntArrayRef stride={}) const;
  Tensor & set_(const Tensor & source) const;
  Tensor & set_() const;
  Tensor & set_quantizer_(ConstQuantizerPtr quantizer) const;
  bool is_set_to(const Tensor & tensor) const;
  Tensor & masked_fill_(const Tensor & mask, Scalar value) const;
  Tensor masked_fill(const Tensor & mask, Scalar value) const;
  Tensor & masked_fill_(const Tensor & mask, const Tensor & value) const;
  Tensor masked_fill(const Tensor & mask, const Tensor & value) const;
  Tensor & masked_scatter_(const Tensor & mask, const Tensor & source) const;
  Tensor masked_scatter(const Tensor & mask, const Tensor & source) const;
  Tensor view(IntArrayRef size) const;
  Tensor & put_(const Tensor & index, const Tensor & source, bool accumulate=false) const;
  Tensor & index_add_(int64_t dim, const Tensor & index, const Tensor & source) const;
  Tensor index_add(int64_t dim, const Tensor & index, const Tensor & source) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor index_add(Dimname dim, const Tensor & index, const Tensor & source) const;
  #endif
  Tensor & index_fill_(int64_t dim, const Tensor & index, Scalar value) const;
  Tensor index_fill(int64_t dim, const Tensor & index, Scalar value) const;
  Tensor & index_fill_(int64_t dim, const Tensor & index, const Tensor & value) const;
  Tensor index_fill(int64_t dim, const Tensor & index, const Tensor & value) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor & index_fill_(Dimname dim, const Tensor & index, Scalar value) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor & index_fill_(Dimname dim, const Tensor & index, const Tensor & value) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor index_fill(Dimname dim, const Tensor & index, Scalar value) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor index_fill(Dimname dim, const Tensor & index, const Tensor & value) const;
  #endif
  Tensor & scatter_(int64_t dim, const Tensor & index, const Tensor & src) const;
  Tensor scatter(int64_t dim, const Tensor & index, const Tensor & src) const;
  Tensor & scatter_(int64_t dim, const Tensor & index, Scalar value) const;
  Tensor scatter(int64_t dim, const Tensor & index, Scalar value) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor scatter(Dimname dim, const Tensor & index, const Tensor & src) const;
  #endif
  #ifdef BUILD_NAMEDTENSOR
  Tensor scatter(Dimname dim, const Tensor & index, Scalar value) const;
  #endif
  Tensor & scatter_add_(int64_t dim, const Tensor & index, const Tensor & src) const;
  Tensor scatter_add(int64_t dim, const Tensor & index, const Tensor & src) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor scatter_add(Dimname dim, const Tensor & index, const Tensor & src) const;
  #endif
  Tensor & lt_(Scalar other) const;
  Tensor & lt_(const Tensor & other) const;
  Tensor & gt_(Scalar other) const;
  Tensor & gt_(const Tensor & other) const;
  Tensor & le_(Scalar other) const;
  Tensor & le_(const Tensor & other) const;
  Tensor & ge_(Scalar other) const;
  Tensor & ge_(const Tensor & other) const;
  Tensor & eq_(Scalar other) const;
  Tensor & eq_(const Tensor & other) const;
  Tensor & ne_(Scalar other) const;
  Tensor & ne_(const Tensor & other) const;
  Tensor __and__(Scalar other) const;
  Tensor __and__(const Tensor & other) const;
  Tensor & __iand__(Scalar other) const;
  Tensor & __iand__(const Tensor & other) const;
  Tensor __or__(Scalar other) const;
  Tensor __or__(const Tensor & other) const;
  Tensor & __ior__(Scalar other) const;
  Tensor & __ior__(const Tensor & other) const;
  Tensor __xor__(Scalar other) const;
  Tensor __xor__(const Tensor & other) const;
  Tensor & __ixor__(Scalar other) const;
  Tensor & __ixor__(const Tensor & other) const;
  Tensor __lshift__(Scalar other) const;
  Tensor __lshift__(const Tensor & other) const;
  Tensor & __ilshift__(Scalar other) const;
  Tensor & __ilshift__(const Tensor & other) const;
  Tensor __rshift__(Scalar other) const;
  Tensor __rshift__(const Tensor & other) const;
  Tensor & __irshift__(Scalar other) const;
  Tensor & __irshift__(const Tensor & other) const;
  Tensor & lgamma_() const;
  Tensor & atan2_(const Tensor & other) const;
  Tensor & tril_(int64_t diagonal=0) const;
  Tensor & triu_(int64_t diagonal=0) const;
  Tensor & digamma_() const;
  Tensor & polygamma_(int64_t n) const;
  Tensor & renorm_(Scalar p, int64_t dim, Scalar maxnorm) const;
  Tensor & pow_(Scalar exponent) const;
  Tensor & pow_(const Tensor & exponent) const;
  Tensor & lerp_(const Tensor & end, Scalar weight) const;
  Tensor & lerp_(const Tensor & end, const Tensor & weight) const;
  Tensor & fmod_(Scalar other) const;
  Tensor & fmod_(const Tensor & other) const;
  Tensor & remainder_(Scalar other) const;
  Tensor & remainder_(const Tensor & other) const;
  Tensor & addbmm_(const Tensor & batch1, const Tensor & batch2, Scalar beta=1, Scalar alpha=1) const;
  Tensor addbmm(const Tensor & batch1, const Tensor & batch2, Scalar beta=1, Scalar alpha=1) const;
  Tensor & addcdiv_(const Tensor & tensor1, const Tensor & tensor2, Scalar value=1) const;
  Tensor & random_(int64_t from, int64_t to, Generator * generator=nullptr) const;
  Tensor & random_(int64_t to, Generator * generator=nullptr) const;
  Tensor & random_(Generator * generator=nullptr) const;
  Tensor & uniform_(double from=0, double to=1, Generator * generator=nullptr) const;
  Tensor & normal_(double mean=0, double std=1, Generator * generator=nullptr) const;
  Tensor & cauchy_(double median=0, double sigma=1, Generator * generator=nullptr) const;
  Tensor & log_normal_(double mean=1, double std=2, Generator * generator=nullptr) const;
  Tensor & exponential_(double lambd=1, Generator * generator=nullptr) const;
  Tensor & geometric_(double p, Generator * generator=nullptr) const;
  Tensor diag(int64_t diagonal=0) const;
  Tensor cross(const Tensor & other, c10::optional<int64_t> dim=c10::nullopt) const;
  Tensor triu(int64_t diagonal=0) const;
  Tensor tril(int64_t diagonal=0) const;
  Tensor trace() const;
  Tensor ne(Scalar other) const;
  Tensor ne(const Tensor & other) const;
  Tensor eq(Scalar other) const;
  Tensor eq(const Tensor & other) const;
  Tensor ge(Scalar other) const;
  Tensor ge(const Tensor & other) const;
  Tensor le(Scalar other) const;
  Tensor le(const Tensor & other) const;
  Tensor gt(Scalar other) const;
  Tensor gt(const Tensor & other) const;
  Tensor lt(Scalar other) const;
  Tensor lt(const Tensor & other) const;
  Tensor take(const Tensor & index) const;
  Tensor index_select(int64_t dim, const Tensor & index) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor index_select(Dimname dim, const Tensor & index) const;
  #endif
  Tensor masked_select(const Tensor & mask) const;
  Tensor nonzero() const;
  std::vector<Tensor> nonzero_numpy() const;
  Tensor gather(int64_t dim, const Tensor & index, bool sparse_grad=false) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor gather(Dimname dim, const Tensor & index, bool sparse_grad=false) const;
  #endif
  Tensor addcmul(const Tensor & tensor1, const Tensor & tensor2, Scalar value=1) const;
  Tensor & addcmul_(const Tensor & tensor1, const Tensor & tensor2, Scalar value=1) const;
  Tensor addcdiv(const Tensor & tensor1, const Tensor & tensor2, Scalar value=1) const;
  std::tuple<Tensor,Tensor> lstsq(const Tensor & A) const;
  std::tuple<Tensor,Tensor> triangular_solve(const Tensor & A, bool upper=true, bool transpose=false, bool unitriangular=false) const;
  std::tuple<Tensor,Tensor> symeig(bool eigenvectors=false, bool upper=true) const;
  std::tuple<Tensor,Tensor> eig(bool eigenvectors=false) const;
  std::tuple<Tensor,Tensor,Tensor> svd(bool some=true, bool compute_uv=true) const;
  Tensor cholesky(bool upper=false) const;
  Tensor cholesky_solve(const Tensor & input2, bool upper=false) const;
  std::tuple<Tensor,Tensor> solve(const Tensor & A) const;
  Tensor cholesky_inverse(bool upper=false) const;
  std::tuple<Tensor,Tensor> qr(bool some=true) const;
  std::tuple<Tensor,Tensor> geqrf() const;
  Tensor orgqr(const Tensor & input2) const;
  Tensor ormqr(const Tensor & input2, const Tensor & input3, bool left=true, bool transpose=false) const;
  Tensor lu_solve(const Tensor & LU_data, const Tensor & LU_pivots) const;
  Tensor multinomial(int64_t num_samples, bool replacement=false, Generator * generator=nullptr) const;
  Tensor lgamma() const;
  Tensor digamma() const;
  Tensor polygamma(int64_t n) const;
  Tensor erfinv() const;
  Tensor & erfinv_() const;
  Tensor sign() const;
  Tensor & sign_() const;
  Tensor dist(const Tensor & other, Scalar p=2) const;
  Tensor atan2(const Tensor & other) const;
  Tensor lerp(const Tensor & end, Scalar weight) const;
  Tensor lerp(const Tensor & end, const Tensor & weight) const;
  Tensor histc(int64_t bins=100, Scalar min=0, Scalar max=0) const;
  Tensor fmod(Scalar other) const;
  Tensor fmod(const Tensor & other) const;
  Tensor remainder(Scalar other) const;
  Tensor remainder(const Tensor & other) const;
  Tensor min(const Tensor & other) const;
  Tensor min() const;
  Tensor max(const Tensor & other) const;
  Tensor max() const;
  Tensor median() const;
  std::tuple<Tensor,Tensor> sort(int64_t dim=-1, bool descending=false) const;
  #ifdef BUILD_NAMEDTENSOR
  std::tuple<Tensor,Tensor> sort(Dimname dim, bool descending=false) const;
  #endif
  Tensor argsort(int64_t dim=-1, bool descending=false) const;
  #ifdef BUILD_NAMEDTENSOR
  Tensor argsort(Dimname dim, bool descending=false) const;
  #endif
  std::tuple<Tensor,Tensor> topk(int64_t k, int64_t dim=-1, bool largest=true, bool sorted=true) const;
  Tensor all() const;
  Tensor any() const;
  Tensor renorm(Scalar p, int64_t dim, Scalar maxnorm) const;
  Tensor unfold(int64_t dimension, int64_t size, int64_t step) const;
  bool equal(const Tensor & other) const;
  Tensor pow(const Tensor & exponent) const;
  Tensor alias() const;

  // We changed .dtype() to return a TypeMeta in #12766. Ideally, we want the
  // at::kDouble and its friends to be TypeMeta's, but that hasn't happened yet.
  // Before that change, we make this method to maintain BC for C++ usage like
  // `x.to(y.dtype)`.
  // TODO: remove following two after at::kDouble and its friends are TypeMeta's.
  inline Tensor to(caffe2::TypeMeta type_meta, bool non_blocking=false, bool copy=false) const {
    return this->to(/*scalar_type=*/typeMetaToScalarType(type_meta), non_blocking, copy);
  }
  inline Tensor to(Device device, caffe2::TypeMeta type_meta, bool non_blocking=false, bool copy=false) const {
    return this->to(device, /*scalar_type=*/typeMetaToScalarType(type_meta), non_blocking, copy);
  }

  template <typename F, typename... Args>
  auto m(F func, Args&&... params) const -> decltype(func(*this, std::forward<Args>(params)...)) {
    return func(*this, std::forward<Args>(params)...);
  }

protected:
  friend class ::caffe2::Tensor;

  void enforce_invariants();
  c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> impl_;
};

namespace detail {
// Helper creator for Tensor class which doesn't requires the users to pass
// in an intrusive_ptr instead it just converts the argument passed to
// requested intrusive_ptr type.
template <typename T, typename... Args>
Tensor make_tensor(Args&&... args) {
  return Tensor(c10::make_intrusive<T>(std::forward<Args>(args)...));
}

} // namespace detail

static inline TensorTypeId legacyExtractTypeId(const Tensor& t) {
  return legacyExtractTypeId(t.type_set());
}

} // namespace at
