Plan bounded observation-by-feature blocks

plan_blocks() uses source chunk hints and an explicit byte budget to build a serializable, metadata-only execution plan. "imagewise" blocks retain complete feature rows, "featurewise" blocks retain complete observation columns, and "balanced" blocks scale both axes in the source chunk ratio.

Usage

plan_blocks(
  x,
  assay = active_assay(x),
  layout = c("balanced", "imagewise", "featurewise"),
  memory_budget = getOption("fmridataset.block_budget", 512 * 1024^2),
  target_block_bytes = getOption("fmridataset.target_block_bytes", 4 * 1024^2)
)

Arguments

x

An fmri_frame or lazy view.

assay

Assay name.

layout

One of "balanced", "imagewise", or "featurewise".

memory_budget

Hard maximum estimated peak bytes for one input block.

target_block_bytes

Preferred block size, capped by memory_budget.

Value

A serializable frame_block_plan.

Examples

frame <- fmri_frame(
  assays = list(signal = memory_source(matrix(seq_len(20), 5, 4))),
  observations = data.frame(.obs_id = sprintf("obs-%d", 1:5)),
  active_assay = "signal"
)
plan <- plan_blocks(frame, memory_budget = 10 * 1024^2)
plan
#> <frame_block_plan> balanced 
#>   shape: 5 x 4 
#>   block shape: 5 x 4 
#>   blocks: 1 
#>   maximum block bytes: 160