carray 2.0.1 → 3.0.1

data/lib/carray/slab_iterator.rb

@@ -0,0 +1,305 @@
+# CASlabIterator.
+#
+# Created by the C indexer dispatch when an index uses the `:>` slab-axis
+# sigil (e.g. `ca[nil, :>, :>]`).  It is a thin sugar over each_slab /
+# map_slab / reduce_slab: `:>` axes become the slab handed to the block;
+# the remaining (sliced) axes are the outer iteration space.
+#
+#   ca[nil, :>].each { |row| ... }       # = ca.each_slab(axis: 1)
+#   ca[2..5, :>, :>].map { |slab| ... }  # = ca[2..5,nil,nil].map_slab(axis: [-2,-1])
+#
+# Surface is intentionally minimal (each / map / reduce); to_a / count /
+# size and other Enumerable extras are deferred.
+#
+# Naming note: distinct from the C-defined `CArray::SlabIterator` engine
+# in ext/carray_slab.c -- this is the Ruby-level sigil iterator that
+# delegates to each_slab / map_slab / reduce_slab.  Loaded lazily via
+# autoload from lib/carray/autoload/autoload_base.rb on first
+# `ca[..., :>]` evaluation.
+
+class CASlabIterator < CAIterator
+  # Built from C via `CASlabIterator.new(reference, slab_axes)`: `reference`
+  # is the sliced base view (with `:>` axes as full range), `slab_axes` are
+  # the positions marked with `:>`.
+  def initialize (reference, slab_axes)
+    @reference = reference
+    @slab_axes = slab_axes
+    rdim = reference.shape
+    @outer_positions = (0...reference.ndim).reject { |k| slab_axes.include?(k) }.freeze
+    @shape             = @outer_positions.map { |k| rdim[k] }
+    @ndim            = @shape.size
+    self
+  end
+
+  # The array being iterated (slab exposes it as `reference`; block / window
+  # iterators expose theirs as `source`).  The base has no common accessor.
+  attr_reader :reference, :slab_axes
+
+  # Slab view at outer index `idx` (Array, length = self.ndim).  Returns
+  # `@reference[*full_idx]` where slab axes are nil (full range) and
+  # outer axes take their values from idx.
+  def kernel_at_index (idx)
+    full = Array.new(@reference.ndim)        # nil at every position
+    @outer_positions.each_with_index { |k, i| full[k] = idx[i] }
+    @reference[*full]
+  end
+
+  # @overload each { |slab| ... }
+  #   Yields each slab as an inner CArray. Without a block, returns
+  #   an `Enumerator` from `each_slab`.
+  #   @yieldparam slab [CArray]
+  #   @return [Enumerator, self]
+  def each (&block)
+    return @reference.each_slab(axis: @slab_axes) unless block
+    @reference.each_slab(axis: @slab_axes, &block)
+  end
+
+  # @overload map { |slab| ... }
+  #   Returns a new CArray built by applying the block to each slab,
+  #   delegating to `map_slab`.
+  #   @yieldparam slab [CArray]
+  #   @yieldreturn [Object] per-slab result.
+  #   @return [CArray]
+  def map (&block)
+    @reference.map_slab(axis: @slab_axes, &block)
+  end
+
+  # @overload reduce { |acc, slab| ... }
+  #   Reduces the slabs with `reduce_slab`. The block-only form uses
+  #   the first slab as the seed; the `init` form starts from `init`.
+  #   @yieldparam acc [Object] running accumulator.
+  #   @yieldparam slab [CArray] next slab.
+  #   @yieldreturn [Object] updated accumulator.
+  #   @return [Object]
+  # @overload reduce(init) { |acc, slab| ... }
+  #   @param init [Object] initial accumulator value.
+  #   @yieldparam acc [Object]
+  #   @yieldparam slab [CArray]
+  #   @yieldreturn [Object]
+  #   @return [Object]
+  def reduce (*args, &block)
+    if args.empty?
+      @reference.reduce_slab(axis: @slab_axes, &block)
+    else
+      @reference.reduce_slab(axis: @slab_axes, init: args[0], &block)
+    end
+  end
+
+  # ---- named reductions -------------------------------------------------
+  #
+  # A per-slab reduction that folds each slab to one value is exactly the core
+  # per-axis reduction over the slab axes, so every reduction is a direct
+  # delegation to `reference.<op>(axis: slab_axes)`. This inherits the core
+  # data type, mask, empty/all-masked (identity vs UNDEF) and epsilon-close
+  # contracts unchanged -- there is no separate slab reduction kernel. (Unlike
+  # the map / reduce block surface, these are mask-aware: they route through the
+  # core reduction, which handles masked sources.)
+
+  # @overload sum
+  #   Per-slab sum, delegating to `reference.sum(axis: slab_axes)`.
+  #   @return [CArray] one value per slab (shape = self.dim minus the slab axes)
+  # @overload accumulate
+  #   Per-slab sum kept in the source's own data type, wrapping at its width,
+  #   as the core `accumulate` does -- `sum` answers in the type the core
+  #   promotes to (float64 for integers).
+  #   @return [CArray] one value per slab
+  # The rest are analogous: prod / mean / min / max, sample and population
+  # variance / stddev, all / any, fused minmax ([min, max] pair), the axis-local
+  # position min_index / max_index (index within the slab axes), and the flat
+  # source address min_addr / max_addr (which source cell holds the extremum).
+  [:sum, :accumulate, :prod, :mean, :min, :max, :variance, :stddev, :all, :any,
+   :variancep, :stddevp, :minmax, :min_index, :max_index,
+   :min_addr, :max_addr].each do |op|
+    class_eval <<~RUBY, __FILE__, __LINE__ + 1
+      def #{op}
+        @reference.#{op}(axis: @slab_axes)
+      end
+    RUBY
+  end
+
+  # @overload count(v = <none>)
+  #   Per-slab count, delegating to `reference.count(..., axis: slab_axes)`.
+  #   No argument counts present (non-masked) cells; `count(UNDEF)` counts
+  #   masked cells; `count(v)` counts cells equal to `v`.
+  #   @return [CArray] one count per slab
+  def count (*args)
+    return count_not_masked if args.empty?
+    # CABlock / CAWindow shadow #count with a geometry accessor, and @reference
+    # may be such a view, so dispatch CArray#count explicitly for count(v).
+    CArray.instance_method(:count).bind_call(@reference, *args, axis: @slab_axes)
+  end
+
+  # @overload count_not_masked
+  #   Per-slab count of present (non-masked) cells.
+  #   @return [CArray]
+  def count_not_masked
+    @reference.count_not_masked(axis: @slab_axes)
+  end
+
+  # @overload count_masked
+  #   Per-slab count of masked cells.
+  #   @return [CArray]
+  def count_masked
+    @reference.count_masked(axis: @slab_axes)
+  end
+
+  # @overload elements
+  #   Per-slab cell count (structural, mask-independent). Every slab has the
+  #   same shape, so this is a constant array shaped like the outer iteration
+  #   space.
+  #   @return [CArray]
+  def elements
+    sz  = @slab_axes.inject(1) { |p, ax| p * @reference.shape[ax] }
+    # count_not_masked gives the correct output shape (and, unlike bare #count,
+    # is not shadowed by CABlock / CAWindow); overwrite with the constant size.
+    out = @reference.count_not_masked(axis: @slab_axes)
+    out[] = sz
+    out
+  end
+
+  # ---- order statistics (tier 3) ----------------------------------------
+  #
+  # Core order statistics take a single integer axis, so a single slab axis is
+  # delegated directly (fast C path) and a multi-axis slab is folded per slab
+  # with reduce_slab (each slab flattened). As with `CArray#median(axis:)`, a
+  # masked source raises (the known per-axis masked limitation); strip the mask
+  # with `ca.value` first if needed.
+
+  # @overload median
+  #   Per-slab median.
+  #   @return [CArray]
+  def median
+    if @slab_axes.size == 1
+      @reference.median(axis: @slab_axes[0])
+    else
+      @reference.reduce_slab(axis: @slab_axes) { |s| s.median }
+    end
+  end
+
+  # @overload percentile(*pers)
+  #   Per-slab percentile(s). One argument returns one CArray; several return an
+  #   array of CArrays (as `CArray#percentile`).
+  #   @return [CArray, Array<CArray>]
+  def percentile (*pers)
+    if @slab_axes.size == 1
+      @reference.percentile(*pers, axis: @slab_axes[0])
+    else
+      rs = pers.map { |p| @reference.reduce_slab(axis: @slab_axes) { |s| s.percentile(p) } }
+      pers.size == 1 ? rs[0] : rs
+    end
+  end
+
+  # @overload quantile
+  #   Per-slab five-number summary `[min, Q1, median, Q3, max]` (five CArrays),
+  #   as `CArray#quantile`.
+  #   @return [Array<CArray>]
+  def quantile
+    if @slab_axes.size == 1
+      @reference.quantile(axis: @slab_axes[0])
+    else
+      [0, 25, 50, 75, 100].map { |p|
+        @reference.reduce_slab(axis: @slab_axes) { |s| s.percentile(p) }
+      }
+    end
+  end
+
+  # ---- order surface ----------------------------------------------------
+  #
+  # sort_addr / sort_index lift the core order surface per slab. Core sort takes
+  # a single axis, so a one-axis slab delegates straight to the C path (unlike
+  # the reductions above, sort does not accept an axis array). A multi-axis slab
+  # has no single sort axis, so it raises; sort one axis at a time.
+
+  # @overload sort_addr
+  #   Per-slab sort by flat source address. Delegates to
+  #   `reference.sort_addr(axis: slab_axis)`: reference-shaped, each slab's cells
+  #   carry the flat source addresses that sort that slab ascending.
+  #   @return [CArray] reference-shaped int64
+  def sort_addr
+    @reference.sort_addr(axis: single_sort_axis(:sort_addr))
+  end
+
+  # @overload sort_index
+  #   Per-slab sort by axis-local index (usable with take_along_axis).
+  #   Delegates to `reference.sort_index(axis: slab_axis)`: reference-shaped, the
+  #   axis-local rank order within the slab axis.
+  #   @return [CArray] reference-shaped int64
+  def sort_index
+    @reference.sort_index(axis: single_sort_axis(:sort_index))
+  end
+
+  # ---- weighted (tier 2) ------------------------------------------------
+  #
+  # Delegated to the core weighted reduction over the slab axes; `weights` has
+  # the same shape as the reference (one weight per cell). Core wsum / wmean
+  # accept multi-axis, so this covers multi-axis slabs directly.
+
+  # @overload wsum(weights)
+  #   Per-slab weighted sum, `weights` shaped like the reference.
+  #   @return [CArray]
+  def wsum (weights)
+    @reference.wsum(weights, axis: @slab_axes)
+  end
+
+  # @overload wmean(weights)
+  #   Per-slab weighted mean, `weights` shaped like the reference.
+  #   @return [CArray]
+  def wmean (weights)
+    @reference.wmean(weights, axis: @slab_axes)
+  end
+
+  # ---- segment scan: within-slab running statistics ---------------------
+  #
+  # A per-slab running accumulation along the slab axis is exactly the core
+  # per-axis cumulative over that axis, so every scan delegates to
+  # `reference.<op>(axis: slab_axis)`, inheriting the core data type / mask (masked
+  # cells hold the running total, output unmasked) contracts unchanged. Each
+  # cell is in exactly one slab (a partition), so the running value is
+  # well-defined; the surface is uniform with the family even though it
+  # coincides with a plain per-axis cumulative. Core scan takes a single axis (a
+  # multi-axis running accumulation is ambiguous), so a multi-axis slab raises --
+  # scan one axis at a time, like the order surface.
+
+  # @overload cumsum
+  #   Per-slab inclusive running sum (float64), reference-shaped.
+  #   @return [CArray]
+  # @overload cumprod
+  #   Per-slab inclusive running product (float64), reference-shaped.
+  #   @return [CArray]
+  # @overload cummax
+  #   Per-slab inclusive running maximum (reference data type), reference-shaped.
+  #   @return [CArray]
+  # @overload cummin
+  #   Per-slab inclusive running minimum (reference data type), reference-shaped.
+  #   @return [CArray]
+  # @overload cumcount
+  #   Per-slab running count of present cells (int64), reference-shaped.
+  #   @return [CArray]
+  [:cumsum, :cumprod, :cummax, :cummin, :cumcount].each do |op|
+    class_eval <<~RUBY, __FILE__, __LINE__ + 1
+      def #{op}
+        @reference.#{op}(axis: single_scan_axis(:#{op}))
+      end
+    RUBY
+  end
+
+  private
+
+  # The single slab axis for the order surface (core sort takes one integer
+  # axis). A multi-axis slab has no single sort axis, so raise.
+  def single_sort_axis (op)
+    return @slab_axes[0] if @slab_axes.size == 1
+    raise ArgumentError,
+          "#{op}: a multi-axis slab (axes #{@slab_axes.inspect}) has no single " \
+          "sort axis; sort one axis at a time"
+  end
+
+  # The single slab axis for a segment scan (core scan takes one integer axis).
+  # A multi-axis slab has no single scan axis, so raise.
+  def single_scan_axis (op)
+    return @slab_axes[0] if @slab_axes.size == 1
+    raise ArgumentError,
+          "#{op}: a multi-axis slab (axes #{@slab_axes.inspect}) has no single " \
+          "scan axis; scan one axis at a time"
+  end
+end

data/lib/carray/stack.rb

@@ -0,0 +1,291 @@
+# CAStack view-default composition surface
+#
+# This file holds the composition methods that build (directly or via
+# reshape/transpose) a CAStack view from storage-uniform pieces:
+#
+#   stack   = K parents pushed onto a new K axis
+#   meld    = melt + weld; pieces dissolve their boundaries along the
+#             named axis and are regarded as one (uniform)
+#   montage = uniform tiles arranged in a regular grid (ImageMagick)
+#
+# Their ragged eager counterparts (concatenate / mosaic / tabulate) live in
+# carray/methods/composition.rb.  See that file's header for the full
+# composition cheat sheet.
+#
+# Examples (a, b, c are each shape [3, 4]):
+#
+#   CArray.stack([a, b, c])                       #=> view, shape (3, 3, 4)   (K outermost)
+#   CArray.stack([a, b, c], axis: -1)             #=> view, shape (3, 4, 3)   (RGB pattern)
+#   CArray.meld([a, b, c])                        #=> view, shape (9, 4)
+#   CArray.meld([a, b, c], axis: 1)               #=> view, shape (3, 12)
+#   CArray.montage([a, b, c, d], [2, 2], axis: 0) #=> view, shape (6, 8) (2x2 image tile)
+#
+# Known limitation: scalar indexing on a CAStack-rooted view drops ndim
+# and triggers a full
+# materialise fallback.  Use range indexing (e.g. `meld(...)[k..k, nil]`
+# instead of `meld(...)[k, nil]`) for partial-use perf, or `.to_ca` upfront
+# to materialise eagerly.
+
+class CArray
+
+  # ---------------------------------------------------------------- view-default
+
+  # Stack `list` of CArrays along a new axis inserted at position
+  # `axis:` (default 0 = outermost).  Returns a view
+  # (CAStack with k_axis = axis) when inputs are storage-uniform, or a
+  # Face-lifted view (= CATime, CATimedelta, ...) when inputs are
+  # homogeneous Face instances.  Call `.to_ca` to materialise eagerly.
+  #
+  # `data_type:` kwarg (optional, primitive Symbol only) forces primitive
+  # promotion; cannot be used when the list contains Face elements.  Class
+  # / Module targets are rejected (= `data_type: CATime` is invalid;
+  # use auto-detect for Face round-trip).
+  #
+  # 3.0 (post-K_AXIS, F.S1-stack landed): replaces `CArray.merge`.  The
+  # low-level raw constructor is `CAStack.new(list, axis:)`; this method
+  # is the high-level surface that performs promote_list + CAStack.new +
+  # (face_lift when homogeneous Face).
+  # @overload stack(list, axis: 0, data_type: nil)
+  #   Returns a view stacking uniform-shape arrays along a new K
+  #   axis at position `axis`. Runs `promote_list` for a common
+  #   `data_type` and re-wraps homogeneous Face inputs via
+  #   `face_lift`. Output `ndim` is one greater than each piece.
+  #   @param list [Array<CArray>] pieces to stack; must not be empty.
+  #   @param axis [Integer] position of the new K axis.
+  #   @param data_type [Symbol, Integer, nil] result `data_type`;
+  #     inferred when `nil`.
+  #   @return [CArray] CAStack view.
+  #   @raise [ArgumentError] when `list` is empty.
+  def self.stack (list, axis: 0, data_type: nil)
+    raise ArgumentError, "stack: list must not be empty" if list.empty?
+    list = CArray.promote_list(list, data_type: data_type)
+    axis = CArray.normalize_axis(axis, list[0].ndim + 1, "stack")
+    CAStack.new(list, axis: axis)                # CAStack.new does Face lift internally
+  end
+
+  # Returns a {CAMeld} view of the arrays welded along an existing axis.
+  # No data is copied; reads gather from parents on demand and writes flow
+  # back to them (chain composability preserved).
+  #
+  # Pieces must agree on `ndim`, `data_type`, byte width, and every axis
+  # length except `axis` (the "meld axis").  Mismatched `data_type` raises:
+  # cast the pieces yourself (`.to_type(:float64)`) or use
+  # {CArray.concatenate} (eager, auto-casts).
+  #
+  # "meld" = melt + weld — pieces dissolve their boundaries along the named
+  # axis and are regarded as one.
+  #
+  # @overload meld(*arrays, axis: 0)
+  # @overload meld(list, axis: 0)
+  #   Convenience form: a single Array argument is treated as the list.
+  #   @param arrays [Array<CArray>] pieces to weld.  A single Array
+  #     argument is accepted for compatibility with older callers.
+  #   @param axis [Integer] existing axis to extend (normalises negative
+  #     values against the reference ndim).
+  #   @return [CAMeld] view over the welded pieces.
+  #   @raise [ArgumentError] when the list is empty, ndim mismatch,
+  #     data_type mismatch, or non-axis dim mismatch across pieces
+  #     (surfaced by CAMeld.new).
+  def self.meld (*arrays, axis: 0)
+    if arrays.length == 1 && arrays[0].is_a?(Array)
+      arrays = arrays[0]
+    end
+    raise ArgumentError, "meld: list must not be empty" if arrays.empty?
+    first = arrays[0]
+    unless first.is_a?(CArray)
+      raise ArgumentError, "meld: entries must be CArray (got #{first.class})"
+    end
+    axis_norm = CArray.normalize_axis(axis, first.ndim, "meld")
+    # Flatten nested CAMeld inputs that share our meld axis: they already
+    # describe a segment sequence, so absorbing their parents keeps chain
+    # depth at 1 (avoids 2-level xfer_all / reduce chains through the
+    # intermediate CAMeld).  A CAMeld with a different meld_axis is left
+    # intact — its segment structure is orthogonal.
+    if arrays.any? { |a| a.is_a?(CAMeld) && a.meld_axis == axis_norm }
+      arrays = arrays.flat_map { |a|
+        a.is_a?(CAMeld) && a.meld_axis == axis_norm ? a.parents : [a]
+      }
+    end
+    CAMeld.new(arrays, axis: axis_norm)
+  end
+end
+
+class CArray
+  # Returns a {CAMeld} view of `[self, *others]` welded along `axis`.
+  # Instance form of {CArray.meld}; non-destructive, see the class method
+  # for full semantics.
+  #
+  # @overload meld(*others, axis: 0)
+  #   @param others [Array<CArray>] additional pieces.
+  #   @param axis [Integer] existing axis to extend.
+  #   @return [CAMeld]
+  #   @raise [ArgumentError] when no `others` are given.
+  def meld (*others, axis: 0)
+    raise ArgumentError, "meld: at least one other array required" if others.empty?
+    CArray.meld(self, *others, axis: axis)
+  end
+
+  # Arrange `list` of uniform-shape pieces in a `tdim`-shape grid that
+  # extends parent axes `axis..axis+tdim.size-1` by the corresponding
+  # `tdim[i]` factor (= ImageMagick `montage` analog).  Output ndim equals
+  # each piece's ndim; the tile axes occupy positions
+  # `axis..axis+tdim.size-1`.  Returns a view; call `.to_ca` to materialise.
+  #
+  # `tdim.product` must equal `list.size`.  For non-uniform pieces along
+  # tile axes, use `CArray.mosaic`.
+  #
+  # Example (parent shape (3, 4), 6-element list, tdim=[2, 3], axis: 0):
+  #
+  #   CArray.montage([a, b, c, d, e, f], [2, 3], axis: 0)
+  #   #=> shape (6, 12) -- 2 rows x 3 cols grid of (3, 4) blocks
+  #   #   +-----+-----+-----+
+  #   #   |  a  |  b  |  c  |  rows 0..2
+  #   #   +-----+-----+-----+
+  #   #   |  d  |  e  |  f  |  rows 3..5
+  #   #   +-----+-----+-----+
+  #
+  # 3.0 (post K_AXIS / promote_list / stack rename): renamed from `combine`
+  # (= 20-year vocabulary that didn't describe the action).  Positional
+  # `at` replaced with `axis:` kwarg for consistency with bind / stack.
+  # Parameter order changed from `(tdim, list, at)` to `(list, tdim, axis:)`
+  # to align with bind / stack (`list` first).
+  # @overload montage(list, tdim, axis: 0, data_type: nil)
+  #   Returns a view arranging uniform-shape pieces in a `tdim`-shape
+  #   grid that extends parent axes `axis..axis+tdim.size-1` by the
+  #   corresponding `tdim[i]` factors. Output `ndim` equals each
+  #   piece's `ndim`. `tdim.product` must equal `list.size`. For
+  #   non-uniform pieces along tile axes, use `CArray.mosaic`.
+  #   @param list [Array<CArray>] pieces to arrange.
+  #   @param tdim [Array<Integer>] tile grid shape.
+  #   @param axis [Integer] first tile axis in the result.
+  #   @param data_type [Symbol, Integer, nil] result `data_type`.
+  #   @return [CArray] tiled view.
+  #   @raise [ArgumentError] when `list` is empty or `tdim.product !=
+  #     list.size`.
+  def self.montage (list, tdim, axis: 0, data_type: nil)
+    raise ArgumentError, "montage: list must not be empty" if list.empty?
+    unless tdim.is_a?(Array) && tdim.size > 0
+      raise ArgumentError, "montage: tdim must be a non-empty Array of Integer"
+    end
+    expected = tdim.inject(1) { |acc, n| acc * n }
+    unless expected == list.size
+      raise ArgumentError,
+            "montage: tdim product (#{expected}) must equal list size (#{list.size})"
+    end
+
+    list = CArray.promote_list(list, data_type: data_type)
+    parent_shape = list[0].shape
+    ntile = tdim.size
+    nparent = parent_shape.size
+    axis = CArray.normalize_axis(axis, nparent - ntile + 1, "montage")
+
+    s = CArray.stack(list).reshape(*tdim, *parent_shape)   # (K, *) → (*tdim, *)
+
+    # Interleave: tile axis i (= s axis i, i ∈ [0, ntile)) is moved to
+    # just before parent axis (axis + i) in s coordinates (= s axis
+    # ntile + axis + i).
+    perm = []
+    nparent.times do |j|
+      if j.between?(axis, axis + ntile - 1)
+        perm << (j - axis)           # tile axis
+      end
+      perm << ntile + j              # parent axis
+    end
+    s = s.transpose(*perm)
+
+    # Merge each (tile[i], parent[axis+i]) pair via reshape.
+    new_shape = parent_shape.dup
+    ntile.times { |i| new_shape[axis + i] *= tdim[i] }
+    s.reshape(*new_shape)
+  end
+
+  # Instance-side stack: build a new K-stack from `[self] + others`
+  # along the new K axis at position `axis:`.  Always treats `self` as a
+  # parent (= even when self is a CAStack, the resulting stack has self
+  # as one of its parents, NOT flat-appended into self's parents).
+  #
+  # For flat-appending into an existing CAStack (= same k_axis, parents
+  # extended), use `CAStack#append`.
+  #
+  # 3.0: high-level Face-aware surface, mirrors `CArray.stack(list, axis:)`.
+  # @overload stack(*others, axis: 0, data_type: nil)
+  #   Returns a K-stack view built from `[self] + others` along a
+  #   new K axis at position `axis`. Always treats `self` as one
+  #   parent, even when `self` is a CAStack; use {CAStack#append} to
+  #   flat-append into an existing CAStack.
+  #   @param others [Array<CArray>] additional parents.
+  #   @param axis [Integer] position of the new K axis.
+  #   @param data_type [Symbol, Integer, nil] result `data_type`.
+  #   @return [CArray] CAStack view.
+  #   @raise [ArgumentError] when no `others` are given.
+  def stack (*others, axis: 0, data_type: nil)
+    raise ArgumentError, "stack: at least one other parent required" if others.empty?
+    CArray.stack([self] + others, axis: axis, data_type: data_type)
+  end
+
+  # Split self along a single axis into an Array of (ndim-1)-D slices, each
+  # a writable CABlock view.  The exact inverse of CArray.stack -- split's
+  # slices are all the same shape, so they round-trip back through stack:
+  #
+  #   CArray.stack(a.split(axis: k), axis: k) == a
+  #
+  #   a = CA_INT([[1,2,3], [4,5,6]])
+  #   a.split(axis: 0)   #=> [ <[1,2,3]>, <[4,5,6]> ]      (row views)
+  #   a.split(axis: 1)   #=> [ <[1,4]>, <[2,5]>, <[3,6]> ] (column views)
+  #
+  # 3.0 breaking:
+  #   - returns a Ruby Array of views (was an object CArray), so it
+  #     round-trips with CArray.stack (which takes an Array)
+  #   - +axis:+ takes a single Integer (the multi-axis Array form, which
+  #     returned an N-D object grid, is no longer accepted)
+  #   - pieces are CABlock views, NOT copies; writing through a piece
+  #     mutates +self+.  Chain +.copy+ / +.to_ca+ for independent entities.
+  # @overload split(axis:)
+  #   Returns an Array of `(ndim-1)`-D writable CABlock views obtained
+  #   by splitting `self` along `axis`. Inverse of `CArray.stack`.
+  #   Slices share storage with `self`; chain `.copy` for independent
+  #   entities.
+  #   @param axis [Integer] axis to split along.
+  #   @return [Array<CArray>] one slice per index along `axis`.
+  #   @raise [ArgumentError] when `axis` is not a single Integer.
+  def split (axis:)
+    if axis.is_a?(Array)
+      raise ArgumentError, "split: axis must be a single Integer"
+    end
+    k = normalize_axis(axis, "split")
+    (0...shape[k]).map do |i|
+      idx = [nil] * ndim
+      idx[k] = i
+      self[*idx]
+    end
+  end
+
+end
+
+# CAStack.new(list, axis: 0) is implemented in C (ext/ca_obj_stack.c
+# rb_ca_stack_s_new): it overrides Class#new to inspect Face homogeneity
+# and (when applicable) re-wrap the raw CAStack via ca_face_lift.  The
+# Ruby class only adds the #append instance method below.
+class CAStack
+  # Flat-append `others` into self's parents, preserving the receiver's
+  # k_axis.  Returns a new CAStack with the combined parents list.
+  #
+  # Distinct from `CArray#stack(other)` which always treats self as a
+  # single parent of a new K-stack -- here, the existing K-stack is
+  # extended in place.
+  #
+  # Example:
+  #   s = CArray.stack([a, b], axis: 1)   # 2-parent stack, k_axis=1
+  #   s.append(c, d)                       # 4-parent stack, k_axis=1
+  # @overload append(*others)
+  #   Returns a new CAStack whose parents are `self.parents + others`,
+  #   preserving `self.k_axis`. Distinct from `CArray#stack`, which
+  #   always treats `self` as a single parent of a new K-stack.
+  #   @param others [Array<CArray>] extra parents to append.
+  #   @return [CAStack]
+  #   @raise [ArgumentError] when no `others` are given.
+  def append (*others)
+    raise ArgumentError, "append: at least one parent required" if others.empty?
+    CArray.stack(self.parents + others, axis: self.k_axis)
+  end
+end