One schedulable unit of a Plan.

Usage

Stage(
  id = integer(0),
  kind = character(0),
  members = integer(0),
  fn = function() NULL,
  halo = integer(0),
  grid = GridSpec(),
  chunks = ChunkGrid(),
  device = character(0),
  inputs = integer(0),
  input_nodes = integer(0),
  exports = integer(0),
  out_pad = 0L,
  export_pads = integer(0)
)

Arguments

id: Stage id (dense, 1-based).
kind: One of "source_read", "compute", "reduce_partial", "reduce_combine", "warp".
members: IR node ids composed into this stage (ascending = topo).
fn: The composed stage closure. For "compute" stages it is called as fn(inputs) with a list of chunk arrays in input_nodes order, each padded to the stage halo, and returns the chunk-core array; a "reduce_partial" stage returns a named list of per-chunk partials and "reduce_combine" combines those lists into the final value; for "source_read" and "warp" stages fn is the identity (the executor and the GDAL warper supply the values).
halo: Halo radius in pixels the stage requires on its inputs.
grid: Output GridSpec of the stage.
chunks: ChunkGrid partitioning the stage's output.
device: Device tag ("cpu" or "cuda").
inputs: Stage ids feeding this stage.
input_nodes: IR node ids whose values fn receives, in order.
exports: Member node ids fn returns, ascending (consumed by other stages, plus the stage tail).
out_pad: Spatial padding rings the stage's chunks are computed with: consumers needing a halo on this stage's exports receive it as a ring of recomputed cells rather than requiring the stage to materialise first. Inputs arrive padded to halo + out_pad.
export_pads: Integer vector parallel to exports: the padding each export value carries (post-focal exports carry less than pre-focal ones). Empty means all zero.

Value

A Stage.