onert::shape_inference Namespace Reference

Namespaces
namespace	bcq

Data Structures
struct	StridedSliceParams

Typedefs
using	Shapes = std::vector< ir::Shape >

Functions
ir::Shape	inferArgMinMaxShape (const ir::Shape &input_shape, int axis, int rank)
ir::Shape	inferBatchMatMulShape (const ir::Shape &lhs_shape, const ir::Shape &rhs_shape, const ir::operation::BatchMatMul::Param &param)
ir::Shape	inferBCQFullyConnectedShape (const ir::Shape &in_shape, const ir::Shape &cluster_shape, const int32_t *cluster_buf)
ir::Shape	inferBCQGatherShape (const ir::Shape &indices_shape, const ir::Shape &cluster_shape, const int32_t *cluster_buf, int rank, const ir::operation::BCQGather::Param &param)
ir::Shape	inferBCQUnembeddingShape (const ir::Shape &in_shape)
ir::Shape	inferBroadcastToShape (const ir::Shape shp_shape, const int32_t *shp_buf)
ir::Shape	inferConcatShape (const Shapes &in_shapes, const ir::operation::Concat::Param &param)
ir::Shape	inferConv2DShape (const ir::Shape &in_shape, const ir::Shape &ker_shape, const ir::operation::Conv2D::Param &param)
ir::Shape	inferDepthwiseConv2DShape (const ir::Shape &in_shape, const ir::Shape &ker_shape, const ir::operation::DepthwiseConv2D::Param &param)
ir::Shape	inferEltwiseShape (const ir::Shape &lhs_shape, const ir::Shape &rhs_shape)
ir::Shape	inferExpandDimsShape (const ir::Shape &in_shape, int32_t axis)
template<typename T >
ir::Shape	inferFillShape (const ir::Shape &fill_shape, const T *shape_buf)
ir::Shape	inferFullyConnectedShape (const ir::Shape &in_shape, const ir::Shape &ker_shape, bool keep_num_dims)
ir::Shape	inferGatherShape (const ir::Shape &input_shape, const ir::Shape &indices_shape, int axis, int rank)
ir::Shape	inferOnehotShape (const ir::Shape &input_shape, const int depth, int axis)
ir::Shape	inferPackShape (const ir::Shape &input_shape, int axis, int rank, int num)
ir::Shape	inferPadShape (const ir::Shape &in_shape, const int32_t *pad_buf, const size_t num_pads)
ir::Shape	inferPoolShape (const ir::Shape &in_shape, const ir::operation::Pool2D::Param &param)
template<typename T >
ir::Shape	inferRangeShape (T start_val, T limit_val, T delta_val)
ir::Shape	inferReshapeShape (const ir::Shape &input_shape, const int32_t *shape_buf, const int32_t shape_num_elements)
ir::Shape	inferReduceShape (const ir::Shape &input_shape, const std::vector< int > &axes, bool keep_dims)
template<float * >
ir::Shape	inferRangeShape (float *start_val, float *limit_val, float *delta_val)
ir::Shape	inferResizeBilinearShape (const ir::Shape &in_shape, const int32_t output_height, const int32_t output_width)
ir::Shape	inferSelectShape (const ir::Shape &input_cond_shape, const ir::Shape &input_true_shape, const ir::Shape &input_false_shape)
template<typename T >
ir::Shape	inferSliceShape (const ir::Shape &input_shape, const T *begins_buf, const T *sizes_buf)
ir::Shape	inferSpaceToBatchNDShape (const ir::Shape &input_shape, const ir::Shape &block_shape_shape, const ir::Shape &padding_shape, const int32_t *block_shape_buf, const int32_t *padding_buf)
ir::Shape	inferSplitShape (const ir::Shape input_shape, int axis_value, int num_splits)
ir::Shape	inferSqueezeShape (const ir::Shape &in_shape, const ir::operation::Squeeze::Param &param)
template<typename T >
StridedSliceParams	buildStridedSliceParams (const T *begin, const T *end, const T *strides, const uint32_t begin_mask, const uint32_t end_mask, const uint32_t shrink_axis_mask, const uint8_t rank)
ir::Shape	inferStridedSliceShape (const ir::Shape &input_shape, const StridedSliceParams &op_params, uint32_t rank)
ir::Shape	inferTileShape (const ir::Shape &in_shape, const int32_t *multiplier_buf, const int32_t multiplier_size)
ir::Shape	inferTransposeShape (const ir::Shape &in_shape, const int32_t *perm_buf, const int32_t rank)
ir::Shape	inferUnpackShape (const ir::Shape &input_shape, int axis, int rank)
std::pair< int, int >	calcConvLikeHeightAndWidth (const int in_h, const int in_w, const int ker_h, const int ker_w, const ir::Padding pad, const ir::Stride stride, const ir::Dilation dilation={1, 1})
template ir::Shape	inferFillShape (const ir::Shape &fill_shape, const int32_t *shape_buf)
template ir::Shape	inferFillShape (const ir::Shape &fill_shape, const int64_t *shape_buf)
template ir::Shape	inferRangeShape (int start_val, int limit_val, int delta_val)
template ir::Shape	inferRangeShape (float start_val, float limit_val, float delta_val)
template ir::Shape	inferSliceShape (const ir::Shape &input_shape, const int32_t *begins_buf, const int32_t *sizes_buf)
template ir::Shape	inferSliceShape (const ir::Shape &input_shape, const int64_t *begins_buf, const int64_t *sizes_buf)
template StridedSliceParams	buildStridedSliceParams (const uint32_t *begin, const uint32_t *end, const uint32_t *strides, const uint32_t begin_mask, const uint32_t end_mask, const uint32_t shrink_axis_mask, const uint8_t rank)
int	Clamp (const int v, const int lo, const int hi)
int	StartForAxis (const StridedSliceParams &params, const ir::Shape &input_shape, int axis)
int	StopForAxis (const StridedSliceParams &params, const ir::Shape &input_shape, int axis, int start_for_axis)

Typedef Documentation

Shapes

using onert::shape_inference::Shapes = typedef std::vector<ir::Shape>

Definition at line 37 of file ShapeInference.h.

Function Documentation

buildStridedSliceParams() [1/2]

template<typename T >

StridedSliceParams onert::shape_inference::buildStridedSliceParams	(	const T *	begin,
		const T *	end,
		const T *	strides,
		const uint32_t	begin_mask,
		const uint32_t	end_mask,
		const uint32_t	shrink_axis_mask,
		const uint8_t	rank
	)

Definition at line 879 of file ShapeInference.cc.

{
  StridedSliceParams op_params;
  op_params.start_indices_count = rank;
  op_params.stop_indices_count = rank;
  op_params.strides_count = rank;
 
  for (int i = 0; i < op_params.strides_count; ++i)
  {
    op_params.start_indices[i] = begin[i];
    op_params.stop_indices[i] = end[i];
    op_params.strides[i] = strides[i];
 
    assert(op_params.strides[i] != 0);
  }
 
  op_params.begin_mask = begin_mask;
  op_params.ellipsis_mask = 0; // NYI
  op_params.end_mask = end_mask;
  op_params.new_axis_mask = 0; // NYI
  op_params.shrink_axis_mask = shrink_axis_mask;
 
  assert(sizeof(op_params.begin_mask) * 4 >= rank);
 
  return op_params;
}

References begin, onert::shape_inference::StridedSliceParams::begin_mask, onert::shape_inference::StridedSliceParams::ellipsis_mask, onert::shape_inference::StridedSliceParams::end_mask, onert::shape_inference::StridedSliceParams::new_axis_mask, onert::shape_inference::StridedSliceParams::shrink_axis_mask, onert::shape_inference::StridedSliceParams::start_indices, onert::shape_inference::StridedSliceParams::start_indices_count, onert::shape_inference::StridedSliceParams::stop_indices, onert::shape_inference::StridedSliceParams::stop_indices_count, onert::shape_inference::StridedSliceParams::strides, and onert::shape_inference::StridedSliceParams::strides_count.

Referenced by onert::exec::DynamicShapeInferer::visit().

buildStridedSliceParams() [2/2]

template StridedSliceParams onert::shape_inference::buildStridedSliceParams	(	const uint32_t *	begin,
		const uint32_t *	end,
		const uint32_t *	strides,
		const uint32_t	begin_mask,
		const uint32_t	end_mask,
		const uint32_t	shrink_axis_mask,
		const uint8_t	rank
	)

calcConvLikeHeightAndWidth()

std::pair< int, int > onert::shape_inference::calcConvLikeHeightAndWidth	(	const int	in_h,
		const int	in_w,
		const int	ker_h,
		const int	ker_w,
		const ir::Padding	pad,
		const ir::Stride	stride,
		const ir::Dilation	dilation = {1, 1}
	)

Definition at line 90 of file ShapeInference.cc.

                                                                           {1, 1})
{
  int32_t out_h = 0, out_w = 0;
  int32_t effective_filter_w_size = (ker_w - 1) * dilation.width_factor + 1;
  int32_t effective_filter_h_size = (ker_h - 1) * dilation.height_factor + 1;
  switch (pad.type)
  {
    case ir::PaddingType::SAME:
      out_h = ceil_div(in_h, stride.vertical);
      out_w = ceil_div(in_w, stride.horizontal);
      break;
    case ir::PaddingType::VALID:
      out_h = ceil_div(in_h - effective_filter_h_size + 1, stride.vertical);
      out_w = ceil_div(in_w - effective_filter_w_size + 1, stride.horizontal);
      break;
    case ir::PaddingType::EXPLICIT:
      out_h =
        (in_h + pad.param.top + pad.param.bottom - effective_filter_h_size) / stride.vertical + 1;
      out_w =
        (in_w + pad.param.left + pad.param.right - effective_filter_w_size) / stride.horizontal + 1;
      break;
    default:
      assert(false);
  }
 
  return {out_h, out_w};
}

Referenced by inferConv2DShape(), inferDepthwiseConv2DShape(), and inferPoolShape().

Clamp()

int onert::shape_inference::Clamp	(	const int	v,
		const int	lo,
		const int	hi
	)

Definition at line 914 of file ShapeInference.cc.

{
  assert(!(hi < lo));
  if (hi < v)
    return hi;
  if (v < lo)
    return lo;
  return v;
}

Referenced by StartForAxis(), and StopForAxis().

inferArgMinMaxShape()

ir::Shape onert::shape_inference::inferArgMinMaxShape	(	const ir::Shape &	input_shape,
		int	axis,
		int	rank
	)

Definition at line 126 of file ShapeInference.cc.

{
  if (axis < 0 || axis >= rank)
  {
    throw std::runtime_error("ArgMinMax shape inference: Wrong axis value " + std::to_string(axis));
  }
 
  ir::Shape out_shape;
  for (int idx = 0; idx < rank; ++idx)
  {
    if (idx != axis)
    {
      int32_t input_dim = input_shape.dim(idx);
      out_shape.append(input_dim);
    }
  }
 
  return out_shape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferBatchMatMulShape()

ir::Shape onert::shape_inference::inferBatchMatMulShape	(	const ir::Shape &	lhs_shape,
		const ir::Shape &	rhs_shape,
		const ir::operation::BatchMatMul::Param &	param
	)

Definition at line 228 of file ShapeInference.cc.

{
  bool adj_x = param.adj_x;
  bool adj_y = param.adj_y;
  ir::Shape output_shape;
 
  int output_rank = std::max(lhs_shape.rank(), rhs_shape.rank());
 
  // Extend lhs and rhs shape
  ir::Shape extended_lhs_shape(lhs_shape);
  ir::Shape extended_rhs_shape(rhs_shape);
  extended_lhs_shape.extendRank(output_rank);
  extended_rhs_shape.extendRank(output_rank);
 
  for (int i = 0; i < output_rank - 2; i++)
  {
    const int lhs_dim = extended_lhs_shape.dim(i);
    const int rhs_dim = extended_rhs_shape.dim(i);
    int broadcast_dim = lhs_dim;
    if (lhs_dim != rhs_dim)
    {
      if (lhs_dim == 1)
      {
        broadcast_dim = rhs_dim;
      }
      else if (rhs_dim != 1)
      {
        throw std::runtime_error{"BatchMatMul shape inference: invalid brodcasting input shape"};
      }
    }
 
    output_shape.append(broadcast_dim);
  }
 
  // Fill in the matmul dimensions.
  int lhs_rows_index = adj_x ? output_rank - 1 : output_rank - 2;
  int rhs_cols_index = adj_y ? output_rank - 2 : output_rank - 1;
 
  output_shape.append(extended_lhs_shape.dim(lhs_rows_index));
  output_shape.append(extended_rhs_shape.dim(rhs_cols_index));
 
  return output_shape;
}

References onert::ir::operation::BatchMatMul::Param::adj_x, onert::ir::operation::BatchMatMul::Param::adj_y, and output_shape.

Referenced by onert::exec::DynamicShapeInferer::visit().

inferBCQFullyConnectedShape()

ir::Shape onert::shape_inference::inferBCQFullyConnectedShape	(	const ir::Shape &	in_shape,
		const ir::Shape &	cluster_shape,
		const int32_t *	cluster_buf
	)

Definition at line 416 of file ShapeInference.cc.

{
  assert(cluster_shape.rank() == 2);
  assert(cluster_shape.dim(1) == 2);
 
  const auto input_size = in_shape.dim(1);
  const auto output_size = bcq::getOutputSize(cluster_shape, cluster_buf);
 
  return {ir::Shape({output_size, input_size})};
}

References onert::shape_inference::bcq::getOutputSize().

Referenced by onert::exec::DynamicShapeInferer::visit().

inferBCQGatherShape()

ir::Shape onert::shape_inference::inferBCQGatherShape	(	const ir::Shape &	indices_shape,
		const ir::Shape &	cluster_shape,
		const int32_t *	cluster_buf,
		int	rank,
		const ir::operation::BCQGather::Param &	param
	)

Definition at line 428 of file ShapeInference.cc.

{
  ir::Shape out_shape;
  ir::Shape in_original_shape;
 
  assert(cluster_shape.rank() == 2);
  assert(cluster_shape.dim(1) == 2);
 
  auto hidden_size = param.input_hidden_size;
  auto axis = param.axis;
 
  in_original_shape.append(bcq::getOutputSize(cluster_shape, cluster_buf));
  in_original_shape.append(hidden_size);
 
  const int indices_rank = indices_shape.rank();
  for (int idx = 0; idx < rank; ++idx)
  {
    if (idx == (int)axis)
    {
      for (int indices_idx = 0; indices_idx < indices_rank; indices_idx++)
      {
        out_shape.append(indices_shape.dim(indices_idx));
      }
    }
    else
    {
      out_shape.append(in_original_shape.dim(idx));
    }
  }
 
  return out_shape;
}

References onert::ir::operation::BCQGather::Param::axis, onert::shape_inference::bcq::getOutputSize(), and onert::ir::operation::BCQGather::Param::input_hidden_size.

Referenced by onert::exec::DynamicShapeInferer::visit().

inferBCQUnembeddingShape()

ir::Shape onert::shape_inference::inferBCQUnembeddingShape

(

const ir::Shape &

in_shape

)

Definition at line 463 of file ShapeInference.cc.

{
  return {ir::Shape({1, in_shape.dim(1)})};
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferBroadcastToShape()

ir::Shape onert::shape_inference::inferBroadcastToShape	(	const ir::Shape	shp_shape,
		const int32_t *	shp_buf
	)

Definition at line 277 of file ShapeInference.cc.

{
 
  const int num_elements = shp_shape.num_elements();
 
  assert(num_elements != 0);
  assert(shp_buf);
 
  ir::Shape new_shape(num_elements);
 
  for (int i = 0; i < num_elements; ++i)
  {
    assert(shp_buf[i] != 0); // It shouldn't be 0.
    new_shape.dim(i) = shp_buf[i];
  }
 
  return new_shape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferConcatShape()

ir::Shape onert::shape_inference::inferConcatShape	(	const Shapes &	in_shapes,
		const ir::operation::Concat::Param &	param
	)

Definition at line 296 of file ShapeInference.cc.

{
  const int32_t concat_axis = param.axis >= 0 ? param.axis : in_shapes[0].rank() + param.axis;
  const auto &first_in_shape = in_shapes[0];
 
  // Check that all shapes are equal except for concat axis dimension
  for (const auto &in_shape : in_shapes)
  {
    if (in_shape.rank() != first_in_shape.rank())
      throw std::runtime_error("Rank in all input tensors should be same");
 
    for (int64_t dim_idx = 0; dim_idx < in_shape.rank(); ++dim_idx)
      if (!(dim_idx == concat_axis || in_shape.dim(dim_idx) == first_in_shape.dim(dim_idx)))
        throw std::runtime_error("All tensor should have same dimension "
                                 "except dimension on passed axis");
  }
 
  // Calculate output shape
  ir::Shape out_shape(first_in_shape);
  out_shape.dim(concat_axis) = 0;
  for (const auto &in_shape : in_shapes)
    out_shape.dim(concat_axis) += in_shape.dim(concat_axis);
  return out_shape;
}

References onert::ir::operation::Concat::Param::axis.

Referenced by onert::exec::DynamicShapeInferer::visit().

inferConv2DShape()

ir::Shape onert::shape_inference::inferConv2DShape	(	const ir::Shape &	in_shape,
		const ir::Shape &	ker_shape,
		const ir::operation::Conv2D::Param &	param
	)

Definition at line 321 of file ShapeInference.cc.

{
  if (param.stride.horizontal == 0 || param.stride.vertical == 0)
    throw std::runtime_error{"Conv2D: stride values must be positive"};
 
  auto ifm_shape = in_shape.asFeature();
 
  // Kernel format is [depth_out, kernel_height, kernel_width, depth_in]
  auto kf_shape = ker_shape.asFeature();
  assert(ifm_shape.C == kf_shape.C);
 
  const auto [out_h, out_w] = calcConvLikeHeightAndWidth(
    ifm_shape.H, ifm_shape.W, kf_shape.H, kf_shape.W, param.padding, param.stride, param.dilation);
 
  return ir::Shape{ifm_shape.N, out_h, out_w, kf_shape.N};
}

References calcConvLikeHeightAndWidth(), onert::ir::operation::Conv2D::Param::dilation, onert::ir::Stride::horizontal, onert::ir::operation::Conv2D::Param::padding, onert::ir::operation::Conv2D::Param::stride, and onert::ir::Stride::vertical.

Referenced by onert::exec::DynamicShapeInferer::visit().

inferDepthwiseConv2DShape()

ir::Shape onert::shape_inference::inferDepthwiseConv2DShape	(	const ir::Shape &	in_shape,
		const ir::Shape &	ker_shape,
		const ir::operation::DepthwiseConv2D::Param &	param
	)

Definition at line 339 of file ShapeInference.cc.

{
  if (param.stride.horizontal == 0 || param.stride.vertical == 0)
    throw std::runtime_error{"DepthwiseConv2D: stride values must be positive"};
 
  auto ifm_shape = in_shape.asFeature();
 
  // Kernel format is [1, kernel_height, kernel_width, depth_out]
  auto kf_shape = ker_shape.asFeature();
  assert(kf_shape.C == static_cast<int32_t>(ifm_shape.C * param.multiplier));
  assert(kf_shape.N == 1);
 
  const auto [out_h, out_w] = calcConvLikeHeightAndWidth(
    ifm_shape.H, ifm_shape.W, kf_shape.H, kf_shape.W, param.padding, param.stride, param.dilation);
 
  return ir::Shape{ifm_shape.N, out_h, out_w, kf_shape.C};
}

References calcConvLikeHeightAndWidth(), onert::ir::operation::DepthwiseConv2D::Param::dilation, onert::ir::Stride::horizontal, onert::ir::operation::DepthwiseConv2D::Param::multiplier, onert::ir::operation::DepthwiseConv2D::Param::padding, onert::ir::operation::DepthwiseConv2D::Param::stride, and onert::ir::Stride::vertical.

Referenced by onert::exec::DynamicShapeInferer::visit().

inferEltwiseShape()

ir::Shape onert::shape_inference::inferEltwiseShape	(	const ir::Shape &	lhs_shape,
		const ir::Shape &	rhs_shape
	)

Definition at line 121 of file ShapeInference.cc.

{
  return broadcastShapes(lhs_shape, rhs_shape);
}

inferExpandDimsShape()

ir::Shape onert::shape_inference::inferExpandDimsShape	(	const ir::Shape &	in_shape,
		int32_t	axis
	)

Definition at line 358 of file ShapeInference.cc.

{
  ir::Shape out_shape(in_shape.rank() + 1);
 
  axis = ((axis >= 0) ? axis : /* when axis < 0 */ (out_shape.rank() + axis));
  if (!(0 <= axis && axis <= in_shape.rank()))
    throw std::runtime_error("axis of dim is out of range");
 
  for (int x = 0, out_x = 0; out_x < out_shape.rank(); ++out_x)
  {
    if (out_x == axis)
      out_shape.dim(out_x) = 1;
    else
      out_shape.dim(out_x) = in_shape.dim(x++);
  }
 
  return out_shape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferFillShape() [1/3]

template ir::Shape onert::shape_inference::inferFillShape	(	const ir::Shape &	fill_shape,
		const int32_t *	shape_buf
	)

inferFillShape() [2/3]

template ir::Shape onert::shape_inference::inferFillShape	(	const ir::Shape &	fill_shape,
		const int64_t *	shape_buf
	)

inferFillShape() [3/3]

template<typename T >

ir::Shape onert::shape_inference::inferFillShape	(	const ir::Shape &	fill_shape,
		const T *	shape_buf
	)

Definition at line 377 of file ShapeInference.cc.

{
  ir::Shape out_shape(fill_shape.dim(0));
 
  for (int out_x = 0; out_x < out_shape.rank(); ++out_x)
  {
    out_shape.dim(out_x) = static_cast<int32_t>(shape_buf[out_x]);
  }
 
  return out_shape;
}

inferFullyConnectedShape()

ir::Shape onert::shape_inference::inferFullyConnectedShape	(	const ir::Shape &	in_shape,
		const ir::Shape &	ker_shape,
		bool	keep_num_dims
	)

Definition at line 393 of file ShapeInference.cc.

{
  assert(in_shape.rank() >= 2);
  assert(ker_shape.rank() == 2);
 
  const auto input_size_with_batch = in_shape.num_elements();
  const auto num_units = ker_shape.dim(0);
  const auto input_size = ker_shape.dim(1);
  const auto batch_size = input_size_with_batch / input_size;
  assert(input_size_with_batch % input_size == 0);
 
  if (keep_num_dims)
  {
    assert(in_shape.dim(in_shape.rank() - 1) == input_size);
    auto output_shape = ir::Shape(in_shape);
    output_shape.dim(output_shape.rank() - 1) = num_units;
    return output_shape;
  }
 
  return {ir::Shape({static_cast<int32_t>(batch_size), num_units})};
}

References output_shape.

Referenced by onert::exec::DynamicShapeInferer::visit().

inferGatherShape()

ir::Shape onert::shape_inference::inferGatherShape	(	const ir::Shape &	input_shape,
		const ir::Shape &	indices_shape,
		int	axis,
		int	rank
	)

Definition at line 468 of file ShapeInference.cc.

{
  ir::Shape out_shape;
 
  const int indices_rank = indices_shape.rank();
 
  for (int idx = 0; idx < rank; ++idx)
  {
    if (idx == axis)
    {
      for (int indices_idx = 0; indices_idx < indices_rank; indices_idx++)
      {
        out_shape.append(indices_shape.dim(indices_idx));
      }
    }
    else
    {
      out_shape.append(input_shape.dim(idx));
    }
  }
 
  return out_shape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferOnehotShape()

ir::Shape onert::shape_inference::inferOnehotShape	(	const ir::Shape &	input_shape,
		const int	depth,
		int	axis
	)

Definition at line 493 of file ShapeInference.cc.

{
  assert(depth >= 0);
  const auto rank = input_shape.rank() + 1;
  ir::Shape newShape(rank);
 
  axis = (axis == -1) ? (rank - 1) : axis;
 
  for (int i = 0; i < rank; ++i)
  {
    if (i < axis)
    {
      newShape.dim(i) = input_shape.dim(i);
    }
    else if (i == axis)
    {
      newShape.dim(i) = depth;
    }
    else
    {
      newShape.dim(i) = input_shape.dim(i - 1);
    }
  }
 
  return newShape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferPackShape()

ir::Shape onert::shape_inference::inferPackShape	(	const ir::Shape &	input_shape,
		int	axis,
		int	rank,
		int	num
	)

Definition at line 520 of file ShapeInference.cc.

{
  ir::Shape out_shape;
  int in_idx = 0;
 
  for (int out_idx = 0; out_idx < rank; ++out_idx)
  {
    if (out_idx == axis)
    {
      out_shape.append(num);
    }
    else
    {
      out_shape.append(input_shape.dim(in_idx++));
    }
  }
 
  return out_shape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferPadShape()

ir::Shape onert::shape_inference::inferPadShape	(	const ir::Shape &	in_shape,
		const int32_t *	pad_buf,
		const size_t	num_pads
	)

Definition at line 540 of file ShapeInference.cc.

{
  assert(num_pads % 2 == 0);
  const int32_t rank = num_pads / 2;
 
  ir::Shape ret(rank);
  for (int32_t i = 0; i < rank; ++i)
  {
    const auto before_padding = pad_buf[i * 2];
    const auto after_padding = pad_buf[i * 2 + 1];
 
    ret.dim(i) = in_shape.dim(i) + before_padding + after_padding;
  }
 
  return ret;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferPoolShape()

ir::Shape onert::shape_inference::inferPoolShape	(	const ir::Shape &	in_shape,
		const ir::operation::Pool2D::Param &	param
	)

Definition at line 557 of file ShapeInference.cc.

{
  if (param.stride.horizontal == 0 || param.stride.vertical == 0)
    throw std::runtime_error{"Pool2D: stride values must be positive"};
 
  auto ifm_shape = in_shape.asFeature();
  const auto [out_h, out_w] = calcConvLikeHeightAndWidth(ifm_shape.H, ifm_shape.W, param.kh,
                                                         param.kw, param.padding, param.stride);
  // Pooling don't change number of channels and batch size
  return ir::Shape{ifm_shape.N, out_h, out_w, ifm_shape.C};
}

References calcConvLikeHeightAndWidth(), onert::ir::Stride::horizontal, onert::ir::operation::Pool2D::Param::kh, onert::ir::operation::Pool2D::Param::kw, onert::ir::operation::Pool2D::Param::padding, onert::ir::operation::Pool2D::Param::stride, and onert::ir::Stride::vertical.

Referenced by onert::exec::DynamicShapeInferer::visit().

inferRangeShape() [1/4]

template<float * >

ir::Shape onert::shape_inference::inferRangeShape	(	float *	start_val,
		float *	limit_val,
		float *	delta_val
	)

inferRangeShape() [2/4]

template ir::Shape onert::shape_inference::inferRangeShape	(	float	start_val,
		float	limit_val,
		float	delta_val
	)

inferRangeShape() [3/4]

template ir::Shape onert::shape_inference::inferRangeShape	(	int	start_val,
		int	limit_val,
		int	delta_val
	)

inferRangeShape() [4/4]

template<typename T >

ir::Shape onert::shape_inference::inferRangeShape	(	T	start_val,
		T	limit_val,
		T	delta_val
	)

Definition at line 594 of file ShapeInference.cc.

{
  ir::Shape out_shape(static_cast<int>(1));
 
  out_shape.dim(0) =
    (std::is_integral<T>::value
       ? ((std::abs(start_val - limit_val) + std::abs(delta_val) - 1) / std::abs(delta_val))
       : std::ceil(std::abs((start_val - limit_val) / delta_val)));
  return out_shape;
}

inferReduceShape()

ir::Shape onert::shape_inference::inferReduceShape	(	const ir::Shape &	input_shape,
		const std::vector< int > &	axes,
		bool	keep_dims
	)

Definition at line 146 of file ShapeInference.cc.

{
  int num_axis = axes.size();
  int input_num_dims = input_shape.rank();
  if (input_num_dims == 0)
  {
    ir::Shape out_shape(0);
    return out_shape;
  }
  if (keep_dims)
  {
    ir::Shape out_shape;
    for (int idx = 0; idx < input_num_dims; ++idx)
    {
      bool is_axis = false;
      for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx)
      {
        if (axes[axis_idx] == idx || axes[axis_idx] + input_num_dims == idx)
        {
          is_axis = true;
          break;
        }
      }
      if (is_axis)
      {
        out_shape.append(1);
      }
      else
      {
        out_shape.append(input_shape.dim(idx));
      }
    }
    return out_shape;
  }
  else
  {
    // Calculates size of reducing axis.
    for (int i = 0; i < num_axis; ++i)
    {
      int current = axes[i];
      if (!(-input_num_dims <= current && current < input_num_dims))
        throw std::runtime_error{"Invalid dim value " + std::to_string(current)};
      if (current < 0)
      {
        current += input_num_dims;
      }
      for (int j = 0; j < i; ++j)
      {
        int previous = axes[j];
        if (previous < 0)
        {
          previous += input_num_dims;
        }
        if (current == previous)
        {
          break;
        }
      }
    }
    // Determines output dimensions.
    ir::Shape out_shape;
    for (int idx = 0; idx < input_num_dims; ++idx)
    {
      bool is_axis = false;
      for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx)
      {
        if (axes[axis_idx] == idx || axes[axis_idx] + input_num_dims == idx)
        {
          is_axis = true;
          break;
        }
      }
      if (!is_axis)
      {
        out_shape.append(input_shape.dim(idx));
      }
    }
    return out_shape;
  }
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferReshapeShape()

ir::Shape onert::shape_inference::inferReshapeShape	(	const ir::Shape &	input_shape,
		const int32_t *	shape_buf,
		const int32_t	shape_num_elements
	)

Definition at line 609 of file ShapeInference.cc.

{
  ir::Shape ret(shape_num_elements);
  int32_t flatten_dim = ir::Shape::kUnspecifiedDim;
  auto total_num_elements = input_shape.num_elements();
  for (int32_t i = 0; i < shape_num_elements; ++i)
  {
    if (shape_buf[i] < 0)
    {
      if (flatten_dim != ir::Shape::kUnspecifiedDim)
        throw std::runtime_error("Reshape: 2nd param has special dim(for flatten) more than twice");
      flatten_dim = i;
      ret.dim(i) = 1;
    }
    else
    {
      ret.dim(i) = shape_buf[i];
    }
  }
  if (flatten_dim != ir::Shape::kUnspecifiedDim)
    ret.dim(flatten_dim) = total_num_elements / ret.num_elements();
 
  // Check reshapable
  if (total_num_elements != static_cast<size_t>(ret.num_elements()))
  {
    // Multi batch case
    // TODO Handle multi batch case more precisely on runtime level
    if ((ret.dim(0) == 1) &&
        (total_num_elements == static_cast<size_t>(ret.num_elements() * input_shape.dim(0))))
      ret.dim(0) = input_shape.dim(0);
    else
      throw std::runtime_error("Reshape: 2nd param is not compatible with the shape of input");
  }
 
  return ret;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferResizeBilinearShape()

ir::Shape onert::shape_inference::inferResizeBilinearShape	(	const ir::Shape &	in_shape,
		const int32_t	output_height,
		const int32_t	output_width
	)

Definition at line 569 of file ShapeInference.cc.

{
  assert(in_shape.rank() == 4);
  if (output_height < 0)
  {
    throw std::runtime_error{"ResizeBilinear: size value must be positive value, output_height = " +
                             std::to_string(output_height)};
  }
  if (output_width < 0)
  {
    throw std::runtime_error{"ResizeBilinear: size value must be positive value, output_width = " +
                             std::to_string(output_width)};
  }
 
  ir::Shape ret(in_shape.rank());
 
  ret.dim(0) = in_shape.dim(0);
  ret.dim(1) = output_height;
  ret.dim(2) = output_width;
  ret.dim(3) = in_shape.dim(3);
 
  return ret;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferSelectShape()

ir::Shape onert::shape_inference::inferSelectShape	(	const ir::Shape &	input_cond_shape,
		const ir::Shape &	input_true_shape,
		const ir::Shape &	input_false_shape
	)

Definition at line 647 of file ShapeInference.cc.

{
  auto haveSameShapes = [](const ir::Shape &input_cond_shape, const ir::Shape &input_true_shape,
                           const ir::Shape &input_false_shape) {
    if ((input_cond_shape.rank() != input_true_shape.rank()) ||
        input_cond_shape.rank() != input_false_shape.rank())
    {
      return false;
    }
 
    int rank = input_cond_shape.rank();
    for (int i = 0; i < rank; ++i)
    {
      if (input_cond_shape.dim(i) != input_true_shape.dim(i) ||
          input_cond_shape.dim(i) != input_false_shape.dim(i))
      {
        return false;
      }
    }
 
    return true;
  };
 
  auto calculateShape = [](const ir::Shape &input_cond_shape, const ir::Shape &input_true_shape,
                           const ir::Shape &input_false_shape, ir::Shape &new_shape) {
    ir::Shape cond_shape = input_cond_shape;
    ir::Shape true_shape = input_true_shape;
    ir::Shape false_shape = input_false_shape;
    int most_rank =
      (cond_shape.rank() >= true_shape.rank()) && (cond_shape.rank() >= false_shape.rank())
        ? cond_shape.rank()
        : (false_shape.rank() >= true_shape.rank() ? false_shape.rank() : true_shape.rank());
 
    ir::Shape calculate_shape(most_rank);
 
    cond_shape.extendRank(most_rank);
    true_shape.extendRank(most_rank);
    false_shape.extendRank(most_rank);
 
    for (int i = 0; i < most_rank; ++i)
    {
      calculate_shape.dim(i) =
        (cond_shape.dim(i) >= true_shape.dim(i)) && (cond_shape.dim(i) >= false_shape.dim(i))
          ? cond_shape.dim(i)
          : (false_shape.dim(i) >= true_shape.dim(i) ? false_shape.dim(i) : true_shape.dim(i));
 
      if ((cond_shape.dim(i) != calculate_shape.dim(i) && cond_shape.dim(i) != 1) ||
          (true_shape.dim(i) != calculate_shape.dim(i) && true_shape.dim(i) != 1) ||
          (false_shape.dim(i) != calculate_shape.dim(i) && false_shape.dim(i) != 1))
      {
        return false;
      }
    }
 
    new_shape = calculate_shape;
 
    return true;
  };
 
  bool havesame = haveSameShapes(input_cond_shape, input_true_shape, input_false_shape);
  if (havesame)
  {
    return input_cond_shape;
  }
 
  ir::Shape new_shape;
  bool possible = calculateShape(input_cond_shape, input_true_shape, input_false_shape, new_shape);
 
  if (!possible)
  {
    throw std::runtime_error("Broadcasting is not possible.");
  }
 
  return new_shape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferSliceShape() [1/3]

template ir::Shape onert::shape_inference::inferSliceShape	(	const ir::Shape &	input_shape,
		const int32_t *	begins_buf,
		const int32_t *	sizes_buf
	)

inferSliceShape() [2/3]

template ir::Shape onert::shape_inference::inferSliceShape	(	const ir::Shape &	input_shape,
		const int64_t *	begins_buf,
		const int64_t *	sizes_buf
	)

inferSliceShape() [3/3]

template<typename T >

ir::Shape onert::shape_inference::inferSliceShape	(	const ir::Shape &	input_shape,
		const T *	begins_buf,
		const T *	sizes_buf
	)

Definition at line 725 of file ShapeInference.cc.

{
  const uint32_t rank = input_shape.rank();
  ir::Shape out_shape(rank);
 
  for (uint32_t idx = 0; idx < rank; ++idx)
  {
    const auto input_dim = input_shape.dim(idx);
 
    // begin is zero-based
    auto begin = begins_buf[idx];
    if (begin < 0)
      throw std::runtime_error("shape inference Slice: Invalid begin.");
 
    // size is one-based
    auto size = sizes_buf[idx];
    if (size < -1)
      throw std::runtime_error("shape inference Slice: Invalid size.");
 
    if (size == -1)
    {
      size = input_dim - begin;
    }
    else
    {
      if (input_dim < static_cast<int32_t>(begin + size))
        throw std::runtime_error("shape inference Slice: Invalid begin and size.");
    }
    out_shape.dim(idx) = static_cast<int32_t>(size);
  }
 
  return out_shape;
}

References begin, and size.

Referenced by onert::exec::DynamicShapeInferer::visit().

inferSpaceToBatchNDShape()

ir::Shape onert::shape_inference::inferSpaceToBatchNDShape	(	const ir::Shape &	input_shape,
		const ir::Shape &	block_shape_shape,
		const ir::Shape &	padding_shape,
		const int32_t *	block_shape_buf,
		const int32_t *	padding_buf
	)

Definition at line 764 of file ShapeInference.cc.

{
  const uint32_t rank = input_shape.rank();
  ir::Shape out_shape(rank);
 
  // Currently, only 4D NHWC input/output op_context are supported.
  // The 4D array need to have exactly 2 spatial dimensions.
  // TODO(nupurgarg): Support arbitrary dimension in SpaceToBatchND.
  [[maybe_unused]] const int32_t kInputDimensionNum = 4;
  [[maybe_unused]] const int32_t kBlockSizeDimensionNum = 1;
  const int32_t kSpatialDimensionNum = 2;
 
  assert(block_shape_shape.rank() == kBlockSizeDimensionNum);
  assert(block_shape_shape.dim(0) == kSpatialDimensionNum);
  assert(padding_shape.dim(0) == kSpatialDimensionNum);
  assert(padding_shape.dim(1) == 2); // fixed, meaning left/right padding for each element
  assert(padding_shape.rank() == 2); // fixed, meaning dimension(dim 0) and padding length(dim 1)
 
  // Ensures the input height and width (with padding) is a multiple of block
  // shape height and width.
  for (int dim = 0; dim < kSpatialDimensionNum; ++dim)
  {
    int final_dim_size =
      (input_shape.dim(dim + 1) + padding_buf[dim * 2] + padding_buf[dim * 2 + 1]);
 
    assert(final_dim_size % block_shape_buf[dim] == 0);
 
    out_shape.dim(dim + 1) = final_dim_size / block_shape_buf[dim];
  }
 
  const int output_batch_size = input_shape.dim(0) * block_shape_buf[0] * block_shape_buf[1];
  const int output_channel_size = input_shape.dim(3);
 
  out_shape.dim(0) = output_batch_size;
  out_shape.dim(3) = output_channel_size;
 
  return out_shape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferSplitShape()

ir::Shape onert::shape_inference::inferSplitShape	(	const ir::Shape	input_shape,
		int	axis_value,
		int	num_splits
	)

Definition at line 806 of file ShapeInference.cc.

{
  ir::Shape newShape(input_shape);
 
  assert(axis_value >= 0);
  assert(axis_value < input_shape.rank());
 
  const int input_size = input_shape.dim(axis_value);
  assert(input_size % num_splits == 0);
  const int slice_size = input_size / num_splits;
 
  newShape.dim(axis_value) = slice_size;
 
  return newShape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferSqueezeShape()

ir::Shape onert::shape_inference::inferSqueezeShape	(	const ir::Shape &	in_shape,
		const ir::operation::Squeeze::Param &	param
	)

Definition at line 822 of file ShapeInference.cc.

{
  const int ndims = param.ndim;
  const int *squeeze_dims = param.dims;
  bool should_squeeze[8] = {false};
  int num_squeezed_dims = 0;
  int shape_rank = in_shape.rank();
  if (ndims == 0)
  {
    for (int idx = 0; idx < shape_rank; ++idx)
    {
      if (in_shape.dim(idx) == 1)
      {
        should_squeeze[idx] = true;
        ++num_squeezed_dims;
      }
    }
  }
  else
  {
    for (int idx = 0; idx < ndims; ++idx)
    {
      int current = squeeze_dims[idx];
      if (current < 0)
      {
        current += shape_rank;
      }
 
      if (!(current >= 0 && current < shape_rank && in_shape.dim(current) == 1))
      {
        throw std::runtime_error(
          "The following conditions must be met: 0 <= dim < Shape rank, dim == 1");
      }
 
      if (!should_squeeze[current])
      {
        ++num_squeezed_dims;
      }
      should_squeeze[current] = true;
    }
  }
 
  // Set output shape.
  ir::Shape out_shape(shape_rank - num_squeezed_dims);
  for (int in_idx = 0, out_idx = 0; in_idx < shape_rank; ++in_idx)
  {
    if (!should_squeeze[in_idx])
    {
      out_shape.dim(out_idx++) = in_shape.dim(in_idx);
    }
  }
 
  return out_shape;
}

References onert::ir::operation::Squeeze::Param::dims, and onert::ir::operation::Squeeze::Param::ndim.

Referenced by onert::exec::DynamicShapeInferer::visit().

inferStridedSliceShape()

ir::Shape onert::shape_inference::inferStridedSliceShape	(	const ir::Shape &	input_shape,
		const StridedSliceParams &	op_params,
		uint32_t	rank
	)

Definition at line 1029 of file ShapeInference.cc.

{
  ir::Shape out_shape;
 
  for (uint32_t idx = 0; idx < rank; ++idx)
  {
    int32_t stride = op_params.strides[idx];
    int32_t begin = StartForAxis(op_params, input_shape, idx);
    int32_t end = StopForAxis(op_params, input_shape, idx, begin);
 
    // When shrinking an axis, the end position does not matter (and can be
    // incorrect when negative indexing is used, see Issue #19260). Always use
    // begin + 1 to generate a length 1 slice, since begin has
    // already been adjusted for negative indices by StartForAxis.
    const bool shrink_axis = op_params.shrink_axis_mask & (1 << idx);
    if (shrink_axis)
    {
      end = begin + 1;
    }
 
    int32_t dim_shape = std::ceil((end - begin) / static_cast<float>(stride));
    dim_shape = dim_shape < 0 ? 0 : dim_shape;
    if (!shrink_axis)
    {
      out_shape.append(dim_shape);
    }
  }
 
  return out_shape;
}

References begin, onert::shape_inference::StridedSliceParams::shrink_axis_mask, StartForAxis(), StopForAxis(), and onert::shape_inference::StridedSliceParams::strides.

Referenced by onert::exec::DynamicShapeInferer::visit().

inferTileShape()

ir::Shape onert::shape_inference::inferTileShape	(	const ir::Shape &	in_shape,
		const int32_t *	multiplier_buf,
		const int32_t	multiplier_size
	)

Definition at line 1061 of file ShapeInference.cc.

{
  if (multiplier_size != in_shape.rank())
  {
    throw std::runtime_error(
      "inferTileShape failed, input rank: " + std::to_string(in_shape.rank()) +
      ", bad multipliers size: " + std::to_string(multiplier_size) + "");
  }
  ir::Shape new_Shape(in_shape.rank());
 
  for (int i = 0; i < in_shape.rank(); ++i)
  {
    assert(multiplier_buf[i]); // multiplier_buf[i] shuld not be 0.
    new_Shape.dim(i) = in_shape.dim(i) * multiplier_buf[i];
  }
  return new_Shape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferTransposeShape()

ir::Shape onert::shape_inference::inferTransposeShape	(	const ir::Shape &	in_shape,
		const int32_t *	perm_buf,
		const int32_t	rank
	)

Definition at line 1080 of file ShapeInference.cc.

{
  const auto rank = in_shape.rank();
  if (perm_size > rank)
  {
    throw std::runtime_error("inferTransposeShape failed, bad permutation size: " +
                             std::to_string(perm_size));
  }
 
  const int32_t *perm_data = perm_buf;
  std::vector<int32_t> regular_perm_vec;
  if (perm_size == 0)
  {
    // perm_data will be set to (n-1...0)
    regular_perm_vec.resize(rank);
    std::iota(regular_perm_vec.begin(), regular_perm_vec.end(), 0);
    std::reverse(regular_perm_vec.begin(), regular_perm_vec.end());
    perm_data = regular_perm_vec.data();
  }
  else
  {
    assert(rank == perm_size);
  }
 
  ir::Shape out_shape(rank);
  std::vector<bool> visit_perms(rank, false);
  for (int idx = 0; idx < rank; idx++)
  {
    const auto perm_val = perm_data[idx];
    // Check invalid permutation value
    if (perm_val < 0 || perm_val >= rank)
    {
      throw std::runtime_error("inferTransposeShape failed, bad permutation value: " +
                               std::to_string(perm_val));
    }
 
    // Check duplicated permutation value
    if (visit_perms.at(perm_val))
    {
      throw std::runtime_error("inferTransposeShape failed, duplicated permutation value: " +
                               std::to_string(perm_val));
    }
    visit_perms.at(perm_val) = true;
 
    out_shape.dim(idx) = in_shape.dim(perm_val);
  }
  return out_shape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

inferUnpackShape()

ir::Shape onert::shape_inference::inferUnpackShape	(	const ir::Shape &	input_shape,
		int	axis,
		int	rank
	)

Definition at line 1130 of file ShapeInference.cc.

{
  ir::Shape out_shape;
 
  for (int out_idx = 0; out_idx < rank; out_idx++)
  {
    if (out_idx != axis)
    {
      out_shape.append(input_shape.dim(out_idx));
    }
  }
 
  return out_shape;
}

Referenced by onert::exec::DynamicShapeInferer::visit().

StartForAxis()

int onert::shape_inference::StartForAxis	(	const StridedSliceParams &	params,
		const ir::Shape &	input_shape,
		int	axis
	)

Definition at line 924 of file ShapeInference.cc.

{
  const auto begin_mask = params.begin_mask;
  const auto *start_indices = params.start_indices;
  const auto *strides = params.strides;
  // Begin with the specified index.
  int start = start_indices[axis];
 
  // begin_mask override
  if (begin_mask & 1 << axis)
  {
    if (strides[axis] > 0)
    {
      // Forward iteration - use the first element. These values will get
      // clamped below (Note: We could have set them to 0 and axis_size-1, but
      // use lowest() and max() to maintain symmetry with StopForAxis())
      start = std::numeric_limits<int>::lowest();
    }
    else
    {
      // Backward iteration - use the last element.
      start = std::numeric_limits<int>::max();
    }
  }
 
  // Handle negative indices
  int axis_size = input_shape.dim(axis);
  if (start < 0)
  {
    start += axis_size;
  }
 
  // Clamping
  start = Clamp(start, 0, axis_size - 1);
 
  return start;
}

References onert::shape_inference::StridedSliceParams::begin_mask, Clamp(), onert::shape_inference::StridedSliceParams::start_indices, and onert::shape_inference::StridedSliceParams::strides.

Referenced by inferStridedSliceShape().

StopForAxis()

int onert::shape_inference::StopForAxis	(	const StridedSliceParams &	params,
		const ir::Shape &	input_shape,
		int	axis,
		int	start_for_axis
	)

Definition at line 967 of file ShapeInference.cc.

{
  const auto end_mask = params.end_mask;
  const auto shrink_axis_mask = params.shrink_axis_mask;
  const auto *stop_indices = params.stop_indices;
  const auto *strides = params.strides;
 
  // Begin with the specified index
  const bool shrink_axis = shrink_axis_mask & (1 << axis);
  int stop = stop_indices[axis];
 
  // When shrinking an axis, the end position does not matter (and can be
  // incorrect when negative indexing is used, see Issue #19260). Always use
  // start_for_axis + 1 to generate a length 1 slice, since start_for_axis has
  // already been adjusted for negative indices.
  if (shrink_axis)
  {
    stop = start_for_axis + 1;
  }
 
  // end_mask override
  if (end_mask & (1 << axis))
  {
    if (strides[axis] > 0)
    {
      // Forward iteration - use the last element. These values will get
      // clamped below
      stop = std::numeric_limits<int>::max();
    }
    else
    {
      // Backward iteration - use the first element.
      stop = std::numeric_limits<int>::lowest();
    }
  }
 
  // Handle negative indices
 
  const int axis_size = input_shape.dim(axis);
  if (stop < 0)
  {
    stop += axis_size;
  }
 
  // Clamping
  // Because the end index points one past the last element, we need slightly
  // different clamping ranges depending on the direction.
  if (strides[axis] > 0)
  {
    // Forward iteration
    stop = Clamp(stop, 0, axis_size);
  }
  else
  {
    // Backward iteration
    stop = Clamp(stop, -1, axis_size - 1);
  }
 
  return stop;
}

References Clamp(), onert::shape_inference::StridedSliceParams::end_mask, onert::shape_inference::StridedSliceParams::shrink_axis_mask, onert::shape_inference::StridedSliceParams::stop_indices, and onert::shape_inference::StridedSliceParams::strides.

Referenced by inferStridedSliceShape().