stelliferous

Search uses release 0.0.2.0 · 2,086 definitions

Releases (2)

def pixel
type PixelBlock
def read_block_pixel
def blockEight lanes of one output row.
type WrapBlockLane 0 is output column `offset`.
def wrap_pixel
def wrap_lane
def read_wrap_pixel
def wrap_block
def Block
type RowStateThe entries of a tile whose lane 0 is output column `offset` of output row `row` of an output `output_width` wide.
def RowTiles
def store_block
def column_step
def next_kernel_row
def kernel_row_step
def next_channel
def channel_step
type CrossingStateThe entries of a tile that crosses an output row, by reduction index.
def Crossings
def store_crossing
def crossing_step
type TileStateThe tiles from output position `position`, eight positions each.
def Tiled
def after_row_tile
def after_crossing
def tileA tile inside one output row, one that ends in the next row, and any other.
def tile_stepEight positions from output column c stay in one output row when c + 8 is at most the output width, and in two when it is at most twice the width.
def finish
def packcount entries, inner per tile, from output position start.
type CallOne packing call shared by raw convolution and certified Buffer entry points.
def run

type Tile8
def filled
def lane0
def lane1
def lane2
def lane3
def lane4
def lane5
def lane6
def lane7
def updateOne ordered multiply-add per lane: eight output positions, one weight.
def from_sequence
type HeadContext
def read_with_head
def pairs_value
def finish_pairs
def read_tail
def view
def element_at

type LayoutEach ordered K coordinate owns eight adjacent weights, one per output row of a block; rows past the last output read zero.
type Context
def eighth
def seventh
def sixth
def fifth
def fourth
def third
def second
def first
def read
def destination
def copy_block
def reordered
def run

type Band
def first_row
def end_row
def bandRows read by output positions [offset, offset + span), kept inside the parent window [parent_first, parent_first + parent_rows).
def coversThe checked facts that let a band stand in for its parent over a nonempty span: rows inside the parent window, a stride-aligned first row that no read needs above, an end no read needs below (or the image end), and room for the kernel.
def copies_bandA fork copies the left child's band only when rows are whole eight-position blocks and the band passes its coverage check; otherwise it clones the parent.
def paddedOutput positions rounded up to whole eight-position blocks.
def startPacking start of a window: output rows before its first row move the start back by first / stride rows.
def copy_planesChannel c copies count values from c * source_stride + source_start to c * target_stride.

type ActivationCustom is a caller-supplied function; only its operation count lives here.
def identity_value
def relu_value
def relu6_value
def leaky
def leaky_relu_valuex for x >= 0, slope * x otherwise.
def sigmoid_value
def silu_value
def tanh_value
def hard_sigmoid_valuePyTorch's order: relu6(x + 3) / 6 and x * relu6(x + 3) / 6.
def swished
def hard_swish_value