dbt_reshape_ops::dbt_copy_indexed Interface Reference

Public Member Functions
subroutine	dbt_reshape (tensor_in, tensor_out, summation, move_data, bounds, index_map, group_mask, group_dim)
	copy data (involves reshape) tensor_out = tensor_out + tensor_in move_data memory optimization: transfer data from tensor_in to tensor_out s.t. tensor_in is empty on return
subroutine	dbt_gather (tensor_in, tensor_out, images, index_map, stack_dim)
	Gather whole blocks from selected tensors, optionally relabelling each block index. Zero images or mapped indices are skipped. Blocks retain their sizes. Each image occupies one source tensor extent along stack_dim, including skipped images. Retained index maps must be injective; the stack-dimension map stays within one source extent.

Detailed Description

Definition at line 44 of file dbt_reshape_ops.F.

Member Function/Subroutine Documentation

dbt_reshape()

subroutine dbt_reshape_ops::dbt_copy_indexed::dbt_reshape	(	type(dbt_type), intent(inout), target	tensor_in,
		type(dbt_type), intent(inout)	tensor_out,
		logical, intent(in), optional	summation,
		logical, intent(in), optional	move_data,
		integer, dimension(:, :), intent(in), optional	bounds,
		integer, dimension(:, :), intent(in), optional	index_map,
		logical, dimension(:, :), intent(in), optional	group_mask,
		integer, intent(in), optional	group_dim )

copy data (involves reshape) tensor_out = tensor_out + tensor_in move_data memory optimization: transfer data from tensor_in to tensor_out s.t. tensor_in is empty on return

Parameters

bounds	crop tensor data while copying (as dbt_crop): start and end index per dimension
tensor_in	source tensor on the parent communicator
tensor_out	destination tensor; smaller communicators denote equal, contiguous subgroups
summation	add to destination instead of replacing it
move_data	clear the source after copying
index_map	optional source-to-destination block indices; zero skips a block
group_mask	optional subgroup selection by source block index
group_dim	source dimension indexing group_mask

Author

Ole Schuett

Definition at line 68 of file dbt_reshape_ops.F.

dbt_gather()

subroutine dbt_reshape_ops::dbt_copy_indexed::dbt_gather	(	type(dbt_type), dimension(:), intent(inout), target	tensor_in,
		type(dbt_type), intent(inout)	tensor_out,
		integer, dimension(:), intent(in)	images,
		integer, dimension(:, :), intent(in)	index_map,
		integer, intent(in)	stack_dim )

Gather whole blocks from selected tensors, optionally relabelling each block index. Zero images or mapped indices are skipped. Blocks retain their sizes. Each image occupies one source tensor extent along stack_dim, including skipped images. Retained index maps must be injective; the stack-dimension map stays within one source extent.

Parameters

tensor_in	source tensors on one communicator, with identical shapes
tensor_out	destination tensor
images	source tensor indices; zero skips an image
index_map	source-to-destination block indices, padded to the largest source dimension
stack_dim	dimension along which images are stacked

Definition at line 94 of file dbt_reshape_ops.F.

Here is the call graph for this function:

---

The documentation for this interface was generated from the following file:

• dbt/dbt_reshape_ops.F