aten/src/ATen/FunctionalInverses.cpp

#include <ATen/FunctionalInverses.h>

#include <ATen/ATen.h>
#include <ATen/ExpandUtils.h>
#include <ATen/WrapDimUtilsMulti.h>

#include <utility>
namespace at {
namespace functionalization {

// This logic is similar to autograd code for view backwards calls.
// We can't easily share it though, because (eventually) these functions
// will all call `permute/unsqueeze_copy()` instead of `permute/unsqueeze`.

static Tensor permute_copy_inverse(const Tensor& self, IntArrayRef dims, bool reapply_views) {
  // invert the permutation
  auto ndims = dims.size();
  std::vector<int64_t> dims_(ndims);
  for(const auto i : c10::irange(ndims)) {
    dims_[at::maybe_wrap_dim(dims[i], ndims)] = i;
  }
  if (reapply_views) {
    return at::permute(self, dims_);
  } else {
    return at::permute_copy(self, dims_);
  }
}

static Tensor unsqueeze_copy_to(const Tensor & self, c10::SymIntArrayRef sizes, bool reapply_views) {
  auto result = self;

  int64_t nDims = sizes.size();
  for(const auto dim : c10::irange(nDims)) {
    if (sizes[dim] == 1) {
      if (reapply_views) {
        result = at::unsqueeze(result, dim);
      } else {
        result = at::unsqueeze_copy(result, dim);
      }
    }
  }
  return result;
}

static Tensor unsqueeze_copy_to(const Tensor & self, IntArrayRef dim, c10::SymIntArrayRef sizes, bool reapply_views) {
  const auto ndim = sizes.size();
  const auto mask = at::dim_list_to_bitset(dim, ndim);
  // in NumPy it's not an error to unsqueeze a scalar, but we still need to avoided
  // unsqueezing in the backward.
  if (ndim == 0) {
    return self;
  }

  Tensor result = self;
  for (const auto d : c10::irange(ndim)) {
    if (mask.test(d) && sizes[d] == 1) {
      if (reapply_views) {
        result = at::unsqueeze(result, d);
      } else {
        result = at::unsqueeze_copy(result, d);
      }
    }
  }
  return result;
}

// Note [Functionalization Pass: View Inverses].
// This file contains the implementation of each "view inverse".
// These aren't really true inverses in the mathematically sense: each view inverse describes how to undo
// the original view (although it takes in different arguments).
//
// E.g. Below is an example of a program that has alias operations removed, and the role that view inverses play:
//
// normal program with views and mutations:
// view1 = input1.view_op(args...)
// view1.add_(1) (perform a mutation on the view, which should also modify input)

// version of the program with no aliasing, that instead uses view_inverse functions:
// view_copy1 = input1.view_copy_op(args...)
// view_copy1.add_(1) (perform a mutation on view_copy1. At this point, input1 is NOT modified)
// x = view_op_inverse(input1, view_copy1, args...)
//
// at this point, input1 and x should be equal
//
// Note that input1 is also passed as an argument to view_op_inverse in the above example.
// This isn't actually required for most view operators: it's only required for view ops
// where you can't figure out what the size of the base tensor is given just the view tensor and arguments.
// Examples are slice/select/scatter/squeeze/as_strided.
// We happen to be passing in the base tensor in all cases, mostly to make the codegen simpler.
// But you'll see below that the "base" argument is ignored by most view_inverse implementations.

// ----------------------------------------------------------
// Implementations of each view_inverse() function are below.
// One of these needs to be implemented for every existing non-composite view operator.
// The codegen automatically generates the corresponding function declaration.
// ----------------------------------------------------------

Tensor FunctionalInverses::_fw_primal_copy_inverse(const at::Tensor& base, const at::Tensor& mutated_view, bool reapply_views, int64_t level) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call _fw_primal() during the functionalization pass. For now, this is not supported.");
    return Tensor();
}

Tensor FunctionalInverses::_make_dual_copy_inverse(const at::Tensor& base, const at::Tensor& mutated_view, bool reapply_views, const at::Tensor& tangent, int64_t level) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call _make_dual() during the functionalization pass. For now, this is not supported.");
    return Tensor();
}

Tensor FunctionalInverses::view_as_real_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    if (reapply_views) {
      return at::view_as_complex(mutated_view);
    } else {
      return at::view_as_complex_copy(mutated_view);
    }
}

Tensor FunctionalInverses::view_as_complex_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    if (reapply_views) {
      return at::view_as_real(mutated_view.resolve_conj());
    } else {
      return at::view_as_real_copy(mutated_view.resolve_conj());
    }
}

Tensor FunctionalInverses::_conj_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    if (reapply_views) {
      return at::_conj(mutated_view);
    } else {
      return at::_conj_copy(mutated_view);
    }
}

Tensor FunctionalInverses::_neg_view_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    if (reapply_views) {
      return at::_neg_view(mutated_view);
    } else {
      return at::_neg_view_copy(mutated_view);
    }
}

Tensor FunctionalInverses::as_strided_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, at::SymIntArrayRef size, at::SymIntArrayRef stride, c10::optional<c10::SymInt> storage_offset) {
    // Pessimism: we can't reapply views for as_strided_scatter.
    return base.as_strided_scatter_symint(mutated_view, size, stride, std::move(storage_offset));
}

Tensor FunctionalInverses::diagonal_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, int64_t offset, int64_t dim1, int64_t dim2) {
    // Pessimism: we can't reapply views for slice_scatter.
    return base.diagonal_scatter(mutated_view, offset, dim1, dim2);
}

Tensor FunctionalInverses::expand_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, at::SymIntArrayRef size, bool implicit) {
    return at::sum_to(mutated_view, base.sym_sizes(),/*always_return_non_view=*/!reapply_views);
}

Tensor FunctionalInverses::permute_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, at::IntArrayRef dims) {
    return at::functionalization::permute_copy_inverse(mutated_view, dims, reapply_views);
}

Tensor FunctionalInverses::_reshape_alias_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, at::SymIntArrayRef size, at::SymIntArrayRef stride) {
    // Note that I'm directly calling reshape(), and ignoring the strides.
    // _reshape_alias() isn't available from user code, and is an implementation detail of reshape().
    // Specifically, passing in the strides directly can get us into trouble in cases like:
    // b = a[0]; c = b.reshape(...); c.add_(1); print(a)
    // When we eventually run the _reshape_alias_inverse() call here, if we were to pass in both sizes and strides,
    // The call would fail because `mutated_view` doesn't have enough bytes of storage.
    if (reapply_views) {
      return at::_reshape_alias_symint(mutated_view, base.sym_sizes(), base.sym_strides());
    } else {
      return at::_reshape_alias_copy_symint(mutated_view, base.sym_sizes(), base.sym_strides());
    }
}

Tensor FunctionalInverses::select_copy_int_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, int64_t dim, c10::SymInt index) {
    // Pessimism: we can't reapply views for slice_scatter.
    return base.select_scatter_symint(mutated_view, dim, std::move(index));
}

Tensor FunctionalInverses::detach_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    // the functionalization pass doesn't care about autograd metadata - as a view, I think detach() is just an identity function
    return mutated_view;
}

Tensor FunctionalInverses::lift_fresh_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    return mutated_view;
}

Tensor FunctionalInverses::slice_copy_Tensor_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, int64_t dim, c10::optional<c10::SymInt> start, c10::optional<c10::SymInt> end, c10::SymInt step) {
    // Pessimism: we can't reapply views for slice_scatter.
    return base.slice_scatter_symint(mutated_view, dim, std::move(start), std::move(end), std::move(step));
}

Tensor FunctionalInverses::split_copy_Tensor_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, int64_t mutated_view_idx, c10::SymInt split_size, int64_t dim) {
    // It would be nice if this logic could be re-used from autograd's split_backward(), but I don't think it can.
    // For functionalization, we have only have one of the tensors from the TensorList outputed by split(), and we want to layer i
    // on top of the base tensor.
    // For autograd, we have all of the tensors outputted by split() and we just want to stack them.
    dim = at::maybe_wrap_dim(dim, base.dim());
    auto dim_size = base.sym_size(dim);
    auto start = split_size * mutated_view_idx;
    auto end = split_size + start;
    if (end > dim_size) end = dim_size;
    // Pessimism: we can't reapply views for slice_scatter.
    return base.slice_scatter_symint(mutated_view, dim, start, end, 1);
}

Tensor FunctionalInverses::split_with_sizes_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, int64_t mutated_view_idx, c10::SymIntArrayRef split_sizes, int64_t dim) {
    dim = at::maybe_wrap_dim(dim, base.dim());
    auto dim_size = base.sym_size(dim);
    c10::SymInt start = 0;
    for (auto i = 0; i < mutated_view_idx; ++i) {
        start += split_sizes[i];
    }
    auto end = start + split_sizes[mutated_view_idx];
    if (end > dim_size) end = dim_size;
    // Pessimism: we can't reapply views for slice_scatter.
    return base.slice_scatter_symint(mutated_view, dim, start, end, 1);
}

Tensor FunctionalInverses::squeeze_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    return unsqueeze_copy_to(mutated_view, base.sym_sizes(), reapply_views);
}

Tensor FunctionalInverses::squeeze_copy_dim_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, int64_t dim) {
    return unsqueeze_copy_to(mutated_view, dim, base.sym_sizes(), reapply_views);
}

Tensor FunctionalInverses::squeeze_copy_dims_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, IntArrayRef dim) {
    return unsqueeze_copy_to(mutated_view, dim, base.sym_sizes(), reapply_views);
}

Tensor FunctionalInverses::t_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    if (reapply_views) {
      return at::t(mutated_view);
    } else {
      return at::t_copy(mutated_view);
    }
}

Tensor FunctionalInverses::transpose_copy_int_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, int64_t dim0, int64_t dim1) {
    if (reapply_views) {
      return transpose(mutated_view, dim0, dim1);
    } else {
      return transpose_copy(mutated_view, dim0, dim1);
    }
}

Tensor FunctionalInverses::_nested_view_from_buffer_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, const Tensor& nested_sizes, const Tensor& nested_strides, const Tensor& storage_offsets) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call _nested_view_from_buffer() during the functionalization pass. For now, nested tensors aren't supported during functionalization");
    return Tensor();
}

Tensor FunctionalInverses::unsqueeze_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, int64_t dim) {
    if (reapply_views) {
      return at::squeeze(mutated_view, dim);
    } else {
      return at::squeeze_copy(mutated_view, dim);
    }
}

Tensor FunctionalInverses::_indices_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call _indices() during the functionalization pass. For now, sparse tensors aren't supported during functionalization");
    return Tensor();
}

Tensor FunctionalInverses::_values_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call _values() during the functionalization pass. For now, sparse tensors aren't supported during functionalization");
    return Tensor();
}

Tensor FunctionalInverses::indices_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call indices() during the functionalization pass. For now, sparse tensors aren't supported during functionalization");
    return Tensor();
}

Tensor FunctionalInverses::values_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call values() during the functionalization pass. For now, sparse tensors aren't supported during functionalization");
    return Tensor();
}

Tensor FunctionalInverses::_sparse_broadcast_to_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, at::IntArrayRef size) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call _sparse_broadcast_to() during the functionalization pass. For now, sparse tensors aren't supported during functionalization");
    return Tensor();
}

Tensor FunctionalInverses::crow_indices_copy_inverse(const at::Tensor& base, const at::Tensor& mutated_view, bool reapply_views) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call crow_indices() during the functionalization pass. For now, sparse tensors aren't supported during functionalization");
    return Tensor();
}

Tensor FunctionalInverses::col_indices_copy_inverse(const at::Tensor& base, const at::Tensor& mutated_view, bool reapply_views) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call col_indices() during the functionalization pass. For now, sparse tensors aren't supported during functionalization");
    return Tensor();
}

Tensor FunctionalInverses::ccol_indices_copy_inverse(const at::Tensor& base, const at::Tensor& mutated_view, bool reapply_views) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call ccol_indices() during the functionalization pass. For now, sparse tensors aren't supported during functionalization");
    return Tensor();
}

Tensor FunctionalInverses::row_indices_copy_inverse(const at::Tensor& base, const at::Tensor& mutated_view, bool reapply_views) {
    TORCH_INTERNAL_ASSERT(false, "Attempted to call row_indices() during the functionalization pass. For now, sparse tensors aren't supported during functionalization");
    return Tensor();
}

Tensor FunctionalInverses::unbind_copy_int_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, int64_t mutated_view_idx, int64_t dim) {
    dim = at::maybe_wrap_dim(dim, base.sizes().size());
    // Pessimism: we can't reapply views for select_scatter.
    return base.select_scatter(mutated_view, dim, mutated_view_idx);
}

Tensor FunctionalInverses::view_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, at::SymIntArrayRef size) {
    if (reapply_views) {
      return mutated_view.view_symint(base.sym_sizes());
    } else {
      return at::view_copy_symint(mutated_view, base.sym_sizes());
    }
}


Tensor FunctionalInverses::view_copy_dtype_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, at::ScalarType dtype) {
    if (reapply_views) {
      return mutated_view.view(base.scalar_type());
    } else {
      return at::view_copy(mutated_view, base.scalar_type());
    }
}

Tensor FunctionalInverses::unfold_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views, int64_t dimension, int64_t size, int64_t step) {
    // I think autograd and the functionalization pass want the exact same thing here, but need to test to confirm.
    // unfold_backward() is safe to use here because it is NOT a view op.
    // (note: technically, "reapply_views" won't do anything here and we'll have an extra memory copy.
    // We'd need to add an aliasing version of unfold_backward to fix that though).
    TORCH_CHECK(
      !(reapply_views && size > step),
      "While executing unfold, functionalization encountered a tensor being mutated that has internal overlap. \
When using torch.compile (or running functionalization directly), this is banned \
as the behavior is not well defined. Consider cloning the tensor before mutating it, \
or removing the mutation from your model."
        );
    return unfold_backward(mutated_view, base.sizes(), dimension, size, step);
}

Tensor FunctionalInverses::alias_copy_inverse(const Tensor& base, const Tensor& mutated_view, bool reapply_views) {
    if (reapply_views) {
      return at::alias(mutated_view);
    } else {
      return at::alias_copy(mutated_view);
    }
}

} // functionalization
} // at