aten/src/ATen/native/Distance.cpp

#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
#include <ATen/core/Tensor.h>
#include <ATen/core/grad_mode.h>
#include <ATen/ExpandUtils.h>
#include <ATen/NamedTensorUtils.h>
#include <ATen/TensorOperators.h>
#include <ATen/native/Distance.h>
#include <c10/util/accumulate.h>

#ifndef AT_PER_OPERATOR_HEADERS
#include <ATen/Functions.h>
#include <ATen/NativeFunctions.h>
#else
#include <ATen/ops/_cdist_backward_native.h>
#include <ATen/ops/_cdist_forward.h>
#include <ATen/ops/_cdist_forward_native.h>
#include <ATen/ops/_euclidean_dist.h>
#include <ATen/ops/_euclidean_dist_native.h>
#include <ATen/ops/_pdist_backward_native.h>
#include <ATen/ops/_pdist_forward.h>
#include <ATen/ops/_pdist_forward_native.h>
#include <ATen/ops/cat.h>
#include <ATen/ops/cdist_native.h>
#include <ATen/ops/cosine_similarity_native.h>
#include <ATen/ops/empty.h>
#include <ATen/ops/empty_like.h>
#include <ATen/ops/linalg_vector_norm.h>
#include <ATen/ops/norm.h>
#include <ATen/ops/ones_like.h>
#include <ATen/ops/pairwise_distance_native.h>
#include <ATen/ops/pdist_native.h>
#include <ATen/ops/pow.h>
#include <ATen/ops/result_type.h>
#include <ATen/ops/sum.h>
#include <ATen/ops/zeros.h>
#include <ATen/ops/zeros_like.h>

#include <utility>
#endif

namespace at::native {

DEFINE_DISPATCH(pdist_forward_stub);
DEFINE_DISPATCH(pdist_backward_stub);
DEFINE_DISPATCH(cdist_stub);
DEFINE_DISPATCH(cdist_backward_stub);

Tensor pairwise_distance(const Tensor& x1, const Tensor& x2, double p, double eps, bool keepdim) {
  // Since either x1 or x2 could be broadcasted
  auto x1_dim = x1.dim();
  auto x2_dim = x2.dim();
  auto output_dim = x1_dim > x2_dim ? x1_dim : x2_dim;
  auto innermost_dim = output_dim - 1;
  return at::norm(x1 - x2 + eps, p, innermost_dim, keepdim);
}

// This is to guarantee that the contiguous memory is passed to the backward pass
Tensor pdist(const Tensor& self, const double p) {
  TORCH_CHECK(self.dim() == 2,
      "pdist only supports 2D tensors, got: ", self.dim(), "D");
  TORCH_CHECK(at::isFloatingType(self.scalar_type()), "pdist only supports floating-point dtypes");
  TORCH_CHECK(p >= 0, "pdist only supports non-negative p values");
  return at::_pdist_forward(self.contiguous(), p);
}

Tensor _euclidean_dist(const Tensor& x1, const Tensor& x2) {
  /** This function does the fist part of the euclidean distance calculation
   * We divide it in two steps to simplify dealing with subgradients in the
   * backward step */
  Tensor x1_norm = x1.pow(2).sum(-1, true);
  Tensor x1_pad = at::ones_like(x1_norm, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
  Tensor x2_norm = x2.pow(2).sum(-1, true);
  Tensor x2_pad = at::ones_like(x2_norm, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
  Tensor x1_ = at::cat({x1.mul(-2), std::move(x1_norm), std::move(x1_pad)}, -1);
  Tensor x2_ = at::cat({x2, std::move(x2_pad), std::move(x2_norm)}, -1);
  Tensor result = x1_.matmul(x2_.mT());
  result.clamp_min_(0).sqrt_();
  return result;
}

static Tensor cdist_impl(const Tensor& x1, const Tensor& x2, const double p, c10::optional<int64_t> compute_mode) {
  TORCH_CHECK(at::isFloatingType(x1.scalar_type()), "cdist only supports floating-point dtypes, X1 got: ", x1.scalar_type());
  auto device1 = x1.device().type();
  TORCH_CHECK(at::isFloatingType(x2.scalar_type()), "cdist only supports floating-point dtypes, X2 got: ", x2.scalar_type());
  auto device2 = x2.device().type();
  TORCH_CHECK(p >= 0, "cdist only supports non-negative p values");
  TORCH_CHECK(device1 == device2, "X1 and X2 must have the same device type. X1: ", device1, " X2: ", device2);
  // TODO: This is bad; this test should apply universally
  TORCH_CHECK(!x1.is_cuda() || x1.get_device() == x2.get_device(), "device of X1 (", x1.get_device(), ") must match device of X2 (", x2.get_device(), ")");
  SymInt c1 = x1.sym_size(-1);
  SymInt c2 = x2.sym_size(-1);
  // 0 - default value. If p = 2 and r1 > 25 or r2 > 25 (these values are based on performance metrics),
  // it will try to compute distance using matrix multiplication approach
  // 1 - force to use matrix multiplication for p = 2
  // 2 - do not use matrix multiplication for p = 2
  int64_t mode = compute_mode.value_or(0);
  TORCH_CHECK(mode >= 0 && mode <= 2, "possible modes: 0, 1, 2, but was: ", mode);

  SymInt r1 = x1.sym_size(-2);
  SymInt r2 = x2.sym_size(-2);

  // See Note [cdist relies on cdist_impl redispatching]
  // Keep this condition in sync with the condition at the Note
  if (!(p == 2 && (mode == 1 || (mode == 0 && (r1 > 25 || r2 > 25))))) {
    TORCH_CHECK(device1 == kCPU || device1 == kCUDA, "cdist only supports CPU and CUDA devices, X1 got: ", device1);
    TORCH_CHECK(device2 == kCPU || device2 == kCUDA, "cdist only supports CPU and CUDA devices, X2 got: ", device2);
  }

  auto dim1 = x1.dim();
  auto dim2 = x2.dim();

  //For batch calculation we expand all dimensions(except the last two) to one, with size that equals to product of them.
  //The last two dimensions will stay the same
  SymIntArrayRef batch_tensor1(x1.sym_sizes().data(), dim1 - 2);
  SymIntArrayRef batch_tensor2(x2.sym_sizes().data(), dim2 - 2);
  std::vector<SymInt> expand_batch_portion = infer_size_symint(batch_tensor1, batch_tensor2);
  std::vector<SymInt> tensor1_expand_size(expand_batch_portion);
  tensor1_expand_size.insert(tensor1_expand_size.end(), {r1, c1});
  std::vector<SymInt> tensor2_expand_size(expand_batch_portion);
  tensor2_expand_size.insert(tensor2_expand_size.end(), {r2, c2});

  const SymInt expand_batch_product = c10::multiply_integers(expand_batch_portion);
  std::vector<SymInt> tensor1_view{expand_batch_product, r1, c1};
  std::vector<SymInt> tensor2_view{expand_batch_product, r2, c2};

  Tensor tensor1_expanded = x1.expand_symint(tensor1_expand_size).contiguous().view_symint(tensor1_view);
  Tensor tensor2_expanded = x2.expand_symint(tensor2_expand_size).contiguous().view_symint(tensor2_view);

  std::vector<SymInt> output_shape(std::move(expand_batch_portion));
  output_shape.insert(output_shape.end(), {r1, r2});

  Tensor result;
  if (r1 == 0 || r2 == 0 || expand_batch_product == 0) {
    result = at::empty_symint(output_shape, x1.options());
  } else if (c1 == 0) {
    result = at::zeros_symint(output_shape, x1.options());
  } else if (p == 2 && (mode == 1 || (mode == 0 && (r1 > 25 || r2 > 25)))) {
    // See Note [cdist relies on cdist_impl redispatching]
    // Keep the condition above in sync with the condition at the Note
    Tensor dist = (expand_batch_product == 1) ? at::_euclidean_dist(x1, x2) :
                  at::_euclidean_dist(tensor1_expanded, tensor2_expanded);
    result = dist.view_symint(output_shape);
  } else {
    result = at::empty_symint(output_shape, x1.options());
    cdist_stub(device1, result, tensor1_expanded, tensor2_expanded, p);
  }
  return result;
}

Tensor cdist(const Tensor& x1, const Tensor& x2, const double p, c10::optional<int64_t> compute_mode) {
  TORCH_CHECK(x1.dim() >= 2, "cdist only supports at least 2D tensors, X1 got: ", x1.dim(), "D");
  TORCH_CHECK(x2.dim() >= 2, "cdist only supports at least 2D tensors, X2 got: ", x2.dim(), "D");
  TORCH_CHECK(x1.sym_size(-1) == x2.sym_size(-1), "X1 and X2 must have the same number of columns. X1: ", x1.sym_size(-1), " X2: ", x2.sym_size(-1));
  auto maybe_outnames = namedinference::compute_cdist_outnames(x1, x2);
  auto result = [&]() {
    NoNamesGuard guard;
    SymInt r1 = x1.sym_size(-2);
    SymInt r2 = x2.sym_size(-2);
    // Special case for empty input: always call the version with explicit autograd to ensure the graph is properly connected
    if (x1.sym_numel() == 0 || x2.sym_numel() == 0) {
        return at::_cdist_forward(x1, x2, p, compute_mode);
    }
    int64_t mode = compute_mode.value_or(0);
    // Note [cdist relies on cdist_impl redispatching]
    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    // This is for pytorch to figure the backward pass itself
    // when p=2.  Keep this condition in sync with the See Note reference
    if (p == 2 && (mode == 1 || (mode == 0 && (r1 > 25 || r2 > 25)))) {
        return cdist_impl(x1, x2, p, compute_mode);
    } else {
        return at::_cdist_forward(x1, x2, p, compute_mode);
    }
  }();
  namedinference::propagate_names_if_nonempty(result, maybe_outnames);
  return result;
}

Tensor _cdist_forward(const Tensor& x1, const Tensor& x2, const double p, c10::optional<int64_t> compute_mode) {
  TORCH_CHECK(x1.dim() >= 2, "cdist only supports at least 2D tensors, X1 got: ", x1.dim(), "D");
  TORCH_CHECK(x2.dim() >= 2, "cdist only supports at least 2D tensors, X2 got: ", x2.dim(), "D");
  TORCH_CHECK(x1.size(-1) == x2.size(-1), "X1 and X2 must have the same number of columns. X1: ", x1.size(-1), " X2: ", x2.size(-1));
  auto maybe_outnames = namedinference::compute_cdist_outnames(x1, x2);
  auto result = [&]() {
    NoNamesGuard guard;
    return cdist_impl(x1, x2, p, compute_mode);
  }();
  namedinference::propagate_names_if_nonempty(result, maybe_outnames);
  return result;
}

Tensor _cdist_backward(const Tensor& _grad, const Tensor& _x1, const Tensor& _x2, const double p, const Tensor& _cdist) {
  // Broadcasting might generate non-contiguous Tensors, so handle it before doing checks
  int64_t c1 = _x1.size(-1);
  int64_t c2 = _x2.size(-1);
  int64_t r1 = _x1.size(-2);
  int64_t r2 = _x2.size(-2);
  auto dim1 = _x1.dim();
  auto dim2 = _x2.dim();
  IntArrayRef batch_tensor1(_x1.sizes().data(), dim1 - 2);
  IntArrayRef batch_tensor2(_x2.sizes().data(), dim2 - 2);
  std::vector<int64_t> expand_batch_portion = infer_size(batch_tensor1, batch_tensor2);
  std::vector<int64_t> tensor1_expand_size(expand_batch_portion);
  tensor1_expand_size.insert(tensor1_expand_size.end(), {r1, c1});
  std::vector<int64_t> tensor2_expand_size(expand_batch_portion);
  tensor2_expand_size.insert(tensor2_expand_size.end(), {r2, c2});

  // Compute the linearized batch size
  const int64_t batch_product = c10::multiply_integers(expand_batch_portion);

  // Gracefully handle empty Tensors
  if (r1 == 0 || r2 == 0 || c1 == 0 || batch_product == 0) {
    return at::zeros_like(_x1, _x1.options());
  }

  Tensor x1 = _x1;
  if (tensor1_expand_size != x1.sizes()) {
    x1 = x1.expand(tensor1_expand_size);
  }
  Tensor x2 = _x2;
  if (tensor2_expand_size != x2.sizes()) {
    x2 = x2.expand(tensor2_expand_size);
  }

  x1 = x1.contiguous();
  x2 = x2.contiguous();
  auto cdist = _cdist.contiguous();
  auto grad = _grad.contiguous();
  int64_t n = x1.size(-2);
  int64_t m = x1.size(-1);
  auto device1 = x1.device().type();
  TORCH_CHECK(device1 == kCPU || device1 == kCUDA, "_cdist_backward only supports CPU and CUDA devices, X1 got: ", device1);
  auto device2 = x2.device().type();
  TORCH_CHECK(device2 == kCPU || device2 == kCUDA, "_cdist_backward only supports CPU and CUDA devices, X2 got: ", device2);

  Tensor grad_x1 =
      at::empty({batch_product, n, m}, x1.options(), LEGACY_CONTIGUOUS_MEMORY_FORMAT);
  cdist_backward_stub(device1, grad_x1, grad, x1, x2, p, cdist);

  // Use x1.size() here and not the original size of _x1.size() as this gradient is not taking broadcasting into account
  // Broadcasting will be handled automatically by the autograd engine
  return grad_x1.view(x1.sizes());
}

Tensor _pdist_forward(const Tensor& self, const double p) {
  TORCH_CHECK(self.is_contiguous(), "_pdist_forward requires contiguous input");
  auto device = self.device().type();
  TORCH_CHECK(device == kCPU || device == kCUDA, "_pdist_forward only supports CPU and CUDA devices, got: ", device);
  Tensor result = at::empty({0}, self.options(), LEGACY_CONTIGUOUS_MEMORY_FORMAT);
  if (self.size(0) <= 1) {
    result.resize_({0});
  } else {
    int64_t n = self.size(0);
    int64_t c = n * (n - 1) / 2;
    result.resize_({c});
    if (self.size(1) == 0) {
      result.fill_(0);
    } else {
      pdist_forward_stub(device, result, self, p);
    }
  }
  return result;
}

Tensor _pdist_backward(const Tensor& grad, const Tensor& self, const double p, const Tensor& pdist) {
  TORCH_CHECK(self.is_contiguous(), "_pdist_backward requires self to be contiguous");
  TORCH_CHECK(pdist.is_contiguous(), "_pdist_backward requires pdist to be contiguous");
  auto device = self.device().type();
  TORCH_CHECK(device == kCPU || device == kCUDA, "_pdist_backward only supports CPU and CUDA devices, got: ", device);
  Tensor result = at::empty_like(self, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
  pdist_backward_stub(device, result, grad, self, p, pdist);
  return result;
}

Tensor cosine_similarity(const Tensor& x1_, const Tensor& x2_, int64_t dim, double eps) {
    /*
   * cosine_similarity(x1, x2) = <x1, x2> / (||x1|| * ||x2||)
   *
   * The current implementation is an improvement over the previous version.
   *
   * Previous implementation:
   * 1. Compute num = <x1, x2>,
   * 2. Compute denom = ||x1|| * ||x2||,
   * 3. Compute denom = max(denom, eps) to avoid division by zero,
   * 4. Return num / denom.
   *
   * Previous implementation has the following issues:
   * 1. Chance of losing precision in <x1, x2> when ||x1|| and ||x2|| are large.
   * 2. Chance of losing precision in ||x1|| * ||x2|| when ||x1|| and ||x2|| are large.
   * 3. Losing precision may cause |cosing_similarity(x1, x2)| > 1.0.
   *
   * Current implementation:
   * 1. Compute x1_normalized = x1 / max(||x1||, eps),
   *            x2_normalized = x2 / max(||x2||, eps),
   * 2. Return <x1_normalized, x2_normalized>.
   *
   * The current implementation improves over the previous one by:
   * 1. Making sure that <x1, x2> and ||x1|| * ||x2|| are not computed explicitly,
   *    hence avoiding floating point overflows.
   * 2. Both methods might have issues with computing ||x1|| and ||x2||, but for
   *    the current method this is the only source of the floating point imprecision.
   * 3. Makes sure |cosing_similarity(x1, x2)| <= 1.0.
   *
   */

  auto commonDtype = at::result_type(x1_, x2_);
  TORCH_CHECK(at::isFloatingType(commonDtype), "expected common dtype to be floating point, yet common dtype is ", commonDtype);

  // We accept integral types (and bools lol) but vector_norm does not
  auto x1_is_int = c10::isIntegralType(x1_.scalar_type(), /*încludeBool=*/true);
  auto x2_is_int = c10::isIntegralType(x2_.scalar_type(), /*încludeBool=*/true);
  auto x1_t = x1_is_int ? x1_.to(commonDtype) : x1_;
  auto x2_t = x2_is_int ? x2_.to(commonDtype) : x2_;
  auto [x1, x2] = expand_outplace(x1_t, x2_t);


  // We want to divide each tensor by its norm first, as it's more numerically stable.
  // This keeps the result between -1.0 and 1.0
  // We clone them, as we're going to modify them in-place
  // This allows the gradients to propagate properly all the way to x1 and x2
  auto x1_norm = at::linalg_vector_norm(*x1, 2, /*dim=*/dim, /*keepdim=*/true).clone();
  auto x2_norm = at::linalg_vector_norm(*x2, 2, /*dim=*/dim, /*keepdim=*/true).clone();

  {
    at::NoGradGuard guard;
    x1_norm.clamp_min_(eps);
    x2_norm.clamp_min_(eps);
  }

  return ((*x1 / x1_norm) * (*x2 / x2_norm)).sum(dim);
}

}  // namespace at::native