stelliferous
Releases (2)
Definitions (2,086)
kernel_packing_buffer.bend
- def run
kernel_packing_samples.bend
- def masked
- def mask_pixel
- def pixel
- type PixelBlock
- def pixel_block
- type PixelAddress
- def pixel_address
- def read_pixel_address
- def read_block_pixel
kernel_pipeline.bend
- type Work
- type TaskA partition node's work and the activation its leaves apply.
- def Outputs
- def activated_value
- def mapped
- def leaky_relu_value
- def unit_mapped
- def unit_lanes
- def activatedOne tight pass over each finished leaf while it is still in cache; applying the activation inside the store measured about twice as slow.
- type InputWindowRows [first, first + rows) of every channel of the image, stored contiguously.
- def copy_rows
- def windows
- def cloned
- def copied
- def split_when
- def splitThe right child keeps the parent's rows; the left child copies only its band.
- def split_task
- def block_tilesOne tile is 32 bytes; 8192 tiles are 256 KB.
- def block_positions
- type BlockState
- def block_kept
- def block_packed
- def block_stepPacking addresses the input window; output positions are local to the leaf.
- def blocks
- def blocks_finished
- def blockedExactly one reusable tile buffer and one output buffer per leaf.
- def leaf
- def partition
- def gather
- def finish_leaf
- def finish
- def depth_limitThe deepest split that leaves every leaf at least one eight-position block.
- def run_partitioned
- def run_selected
- def run_packed
- def run
kernel_planning.bend
- type Owners
- type GeometryImage and stencil extents that decide which input rows a fork copies.
- type Work
- type InputRowsThe input rows a partition node holds and the capacity of their owner.
- type ForkInputAllocation and moved/initialized payload are distinct for a copied band.
- type Movementcritical_ns is the span (the costliest root-to-leaf path) and work_ns the sum over every leaf and fork; together they bound time on a given worker count.
- type Plan
- def capacity_loop
- def capacity
- def float_parts
- def payload_float
- def weight_bytes
- def reorder_nsThe weight reorder writes complete eight-channel tiles, including zero lanes.
- def fork_bytesEach fork clones the reordered weights and bias, and splits the input rows.
- def clone_bytes
- def band_rows
- def band_bytes
- def input_transfercopied initializes the new owner and copies every logical band value.
- def input_bytesAs in the pipelines: a checked band over aligned rows is copied and the second half keeps the parent; otherwise the parent owner is cloned.
- def left_rows
- def fork_window
- def fork_input
- def forked_bytes
- def forked_transfer
- def forked_left
- def root_rows
- def activation_nsThe activation pass reads and writes each logical output in place; the identity skips the pass.
- def leaf_cost
- def merge
- def movement
- def workspace
- def scheduledEstimate the partition on workers by (work - span) / workers + span, Graham's bound for greedy list scheduling.
- def estimate
- def fits
- def choose
- def candidate
- def candidates
- def selectworkers is the runtime's pool size.
kernel_rows.bend
- type RowsFour independent output rows share one input tile: half of an eight-row weight block.
- def filled
- def first
- def second
- def third
- def fourth
- def halfThe lower or upper four rows of an eight-row weight tile.
- def map
kernel_shape.bend
- type Dimensions
- def axis_size
- def valid
- def direct_bounds
- def packed_bounds
- def supports_directNever duplicate the Limits owner: its Facts field depends on its values.
- def supports_packed
kernel_spatial_packing.bend
- def masked