carray 2.0.1 → 3.0.0

data/ext/ca_compare.c

@@ -0,0 +1,94 @@
+/* ---------------------------------------------------------------------------
+
+  Shared element comparators by data_type (see ca_compare.h).  The per-type
+  comparators take the (const void *, const void *) prototype directly, so
+  the dispatch table needs no function-pointer casts.
+
+  Internal C API (no Ruby surface).  Indexed by data_type id via
+  `ca_elem_cmp[data_type]`.  Callers include sort_addr_cmp
+  (carray_sort.c) and the comparator-based ca_quickselect_bytes path
+  (carray_partition.c).
+
+--------------------------------------------------------------------------- */
+
+#include "ruby.h"
+#include "carray.h"
+#include "ca_compare.h"
+#include <math.h>
+
+#define DEFINE_ELEM_CMP(type)                          \
+static int                                             \
+elem_cmp_## type (const void *va, const void *vb)      \
+{                                                      \
+  const type *a = (const type *) va;                   \
+  const type *b = (const type *) vb;                   \
+  if ( *a > *b ) return  1;                            \
+  if ( *a < *b ) return -1;                            \
+  return 0;                                            \
+}
+
+/* Float order: NaN sorts last (matches the sort kernels' NaN-at-end). */
+#define DEFINE_ELEM_CMP_FLOAT(type)                    \
+static int                                             \
+elem_cmp_## type (const void *va, const void *vb)      \
+{                                                      \
+  const type *a = (const type *) va;                   \
+  const type *b = (const type *) vb;                   \
+  if ( isnan(*a) && ( ! isnan(*b) ) ) return  1;       \
+  if ( isnan(*b) && ( ! isnan(*a) ) ) return -1;       \
+  if ( *a > *b ) return  1;                            \
+  if ( *a < *b ) return -1;                            \
+  return 0;                                            \
+}
+
+DEFINE_ELEM_CMP(boolean8_t)
+DEFINE_ELEM_CMP(int8_t)
+DEFINE_ELEM_CMP(uint8_t)
+DEFINE_ELEM_CMP(int16_t)
+DEFINE_ELEM_CMP(uint16_t)
+DEFINE_ELEM_CMP(int32_t)
+DEFINE_ELEM_CMP(uint32_t)
+DEFINE_ELEM_CMP(int64_t)
+DEFINE_ELEM_CMP(uint64_t)
+DEFINE_ELEM_CMP_FLOAT(float32_t)
+DEFINE_ELEM_CMP_FLOAT(float64_t)
+
+/* CA_OBJECT: compare via Ruby Comparable (`<=>`). */
+static int
+elem_cmp_VALUE (const void *va, const void *vb)
+{
+  const VALUE *a = (const VALUE *) va;
+  const VALUE *b = (const VALUE *) vb;
+  return NUM2INT(rb_funcall(*a, rb_intern("<=>"), 1, *b));
+}
+
+/* FIXLEN / retired / otherwise unorderable data_types: callers must handle
+   these themselves; reaching the table for them is a bug. */
+static int
+elem_cmp_unsupported (const void *va, const void *vb)
+{
+  (void) va; (void) vb;
+  rb_raise(rb_eNotImpError,
+           "element comparison is not implemented for the data type");
+}
+
+ca_cmp_func
+ca_elem_cmp[CA_NTYPE] = {
+  elem_cmp_unsupported,  /* CA_FIXLEN -- caller compares packed blobs itself */
+  elem_cmp_boolean8_t,
+  elem_cmp_int8_t,
+  elem_cmp_uint8_t,
+  elem_cmp_int16_t,
+  elem_cmp_uint16_t,
+  elem_cmp_int32_t,
+  elem_cmp_uint32_t,
+  elem_cmp_int64_t,
+  elem_cmp_uint64_t,
+  elem_cmp_float32_t,
+  elem_cmp_float64_t,
+  elem_cmp_unsupported,  /* float128_t -- not supported */
+  elem_cmp_unsupported,
+  elem_cmp_unsupported,
+  elem_cmp_unsupported,  /* cmplx256_t -- not supported */
+  elem_cmp_VALUE,
+};

data/ext/ca_compare.h

@@ -0,0 +1,26 @@
+/* ---------------------------------------------------------------------------
+
+  Element comparators by data_type: a 3-way compare (-1 / 0 / +1) of two
+  same-typed cells, with the (const void *, const void *) prototype that
+  libc qsort / bsearch expect.  This is a shared *ordering primitive*, not
+  tied to any one algorithm: the index sort (carray_sort.c) and the binary
+  search (carray_order.c) both dispatch through ca_elem_cmp[data_type].
+
+  CA_FIXLEN (packed byte blobs) is intentionally `unsupported` in the table:
+  a 2-arg comparator cannot carry the element byte width, so each caller
+  compares fixlen cells itself (carray_order.c via a cmp_data wrapper for
+  bsearch; carray_sort.c via a direct memcmp).  Float slots order NaN last.
+
+--------------------------------------------------------------------------- */
+
+#ifndef CA_COMPARE_H
+#define CA_COMPARE_H
+
+/* 3-way element comparator, qsort/bsearch-compatible signature. */
+typedef int (*ca_cmp_func)(const void *, const void *);
+
+/* Per-data_type element comparators, indexed by data_type (CA_NTYPE wide).
+   FIXLEN / retired / unused slots hold an unsupported stub that raises. */
+extern ca_cmp_func ca_elem_cmp[CA_NTYPE];
+
+#endif /* CA_COMPARE_H */

data/ext/ca_composite_dispatch.c

@@ -0,0 +1,414 @@
+/* ---------------------------------------------------------------------------
+
+  Geometric primitives for region-mapped views (CAWindow / CATile /
+  CARoll): per-region strided gather, scatter, and broadcast fill.
+  See ca_composite_dispatch.h for the full per-function API contract.
+
+  Routines here are pure C, no Ruby VALUE handling.  They operate on
+  raw pointers + byte stride arrays.  Caller is responsible for parent
+  attach state and output buffer allocation.
+
+---------------------------------------------------------------------------- */
+
+#include "ca_composite_dispatch.h"
+#include <string.h>
+
+/* --------------------------------------------------------------------- */
+/* ca_fill_typed                                                          */
+/*                                                                        */
+/* Per-size typed-store loops let the compiler autovectorize (NEON /      */
+/* SSE), converting N tight memcpy(bytes) calls into one O(N) SIMD'd      */
+/* typed-store loop.  Generic memcpy fallback covers CA_FIXLEN and        */
+/* unusual widths.                                                        */
+/* --------------------------------------------------------------------- */
+
+void
+ca_fill_typed (char *dst, const char *val, ca_size_t bytes, ca_size_t n)
+{
+  ca_size_t i;
+  switch ( bytes ) {
+    case 1: {
+      memset(dst, *(const unsigned char *)val, n);
+      break;
+    }
+    case 2: {
+      uint16_t  v = *(const uint16_t *)val;
+      uint16_t *p = (uint16_t *)dst;
+      for ( i = 0; i < n; i++ ) p[i] = v;
+      break;
+    }
+    case 4: {
+      uint32_t  v = *(const uint32_t *)val;
+      uint32_t *p = (uint32_t *)dst;
+      for ( i = 0; i < n; i++ ) p[i] = v;
+      break;
+    }
+    case 8: {
+      uint64_t  v = *(const uint64_t *)val;
+      uint64_t *p = (uint64_t *)dst;
+      for ( i = 0; i < n; i++ ) p[i] = v;
+      break;
+    }
+    case 16: {
+      /* complex128 etc.  Two uint64_t stores per element.  Compiler
+         pairs them into a NEON 128-bit or SSE2 store. */
+      uint64_t v0 = ((const uint64_t *)val)[0];
+      uint64_t v1 = ((const uint64_t *)val)[1];
+      uint64_t *p = (uint64_t *)dst;
+      for ( i = 0; i < n; i++ ) {
+        p[2*i]     = v0;
+        p[2*i + 1] = v1;
+      }
+      break;
+    }
+    default: {
+      /* Generic fallback: CA_FIXLEN (struct types), unusual widths. */
+      for ( i = 0; i < n; i++ ) memcpy(dst + i * bytes, val, bytes);
+      break;
+    }
+  }
+}
+
+/* --------------------------------------------------------------------- */
+/* ca_composite_region_gather                                             */
+/*                                                                        */
+/* Per-region strided memcpy from a parent sub-rectangle to an output    */
+/* sub-rectangle, with 1D / 2D / N-D fast paths.                         */
+/* --------------------------------------------------------------------- */
+
+void
+ca_composite_region_gather (char *parent_ptr,
+                            const ca_size_t *parent_strides,
+                            const ca_size_t *parent_start,
+                            char *out_ptr,
+                            const ca_size_t *output_strides,
+                            const ca_size_t *output_offset,
+                            const ca_size_t *count,
+                            int8_t ndim,
+                            ca_size_t bytes)
+{
+  ca_size_t idx[CA_RANK_MAX];
+  ca_size_t parent_off, output_off, run_bytes, total_slabs, slab;
+  int8_t    k;
+  int8_t    outer_axes;
+
+  /* Initial byte offsets into parent and output. */
+  parent_off = 0;
+  output_off = 0;
+  for ( k = 0; k < ndim; k++ ) {
+    parent_off += parent_start[k]  * parent_strides[k];
+    output_off += output_offset[k] * output_strides[k];
+    idx[k] = 0;
+  }
+
+  /* Innermost axis contributes the contig run; outer axes are walked. */
+  run_bytes = count[ndim - 1] * bytes;
+  outer_axes = ndim - 1;
+
+  if ( outer_axes <= 0 ) {
+    /* 1-D shortcut: single memcpy. */
+    memcpy(out_ptr + output_off, parent_ptr + parent_off, run_bytes);
+    return;
+  }
+
+  if ( outer_axes == 1 ) {
+    /* 2-D fast path: single outer axis, pointer-advancing loop.  No
+       multi-dim cursor, no idx[] array — keeps the per-iter body tight
+       so SIMD-friendly memcpy can dominate. */
+    char     *src = parent_ptr + parent_off;
+    char     *dst = out_ptr + output_off;
+    ca_size_t step_p = parent_strides[0];
+    ca_size_t step_o = output_strides[0];
+    ca_size_t count0 = count[0];
+    ca_size_t i;
+    for ( i = 0; i < count0; i++ ) {
+      memcpy(dst, src, run_bytes);
+      dst += step_o;
+      src += step_p;
+    }
+    return;
+  }
+
+  /* General N-D path (ndim >= 3): multi-dim cursor walk over outer axes. */
+  total_slabs = 1;
+  for ( k = 0; k < outer_axes; k++ ) {
+    total_slabs *= count[k];
+  }
+
+  for ( slab = 0; slab < total_slabs; slab++ ) {
+    memcpy(out_ptr + output_off, parent_ptr + parent_off, run_bytes);
+
+    /* Increment the multi-dim cursor (row-major, outer axes only). */
+    for ( k = outer_axes - 1; k >= 0; k-- ) {
+      idx[k]++;
+      parent_off += parent_strides[k];
+      output_off += output_strides[k];
+      if ( idx[k] < count[k] ) break;
+      parent_off -= parent_strides[k] * count[k];
+      output_off -= output_strides[k] * count[k];
+      idx[k] = 0;
+    }
+  }
+}
+
+/* --------------------------------------------------------------------- */
+/* ca_composite_region_scatter                                            */
+/*                                                                        */
+/* Reverse direction of _gather: copies an input sub-rectangle back into */
+/* a parent sub-rectangle (in_ptr -> parent_ptr).                        */
+/* --------------------------------------------------------------------- */
+
+void
+ca_composite_region_scatter (char *parent_ptr,
+                             const ca_size_t *parent_strides,
+                             const ca_size_t *parent_start,
+                             const char *in_ptr,
+                             const ca_size_t *input_strides,
+                             const ca_size_t *input_offset,
+                             const ca_size_t *count,
+                             int8_t ndim,
+                             ca_size_t bytes)
+{
+  ca_size_t idx[CA_RANK_MAX];
+  ca_size_t parent_off, input_off, run_bytes, total_slabs, slab;
+  int8_t    k;
+  int8_t    outer_axes;
+
+  parent_off = 0;
+  input_off  = 0;
+  for ( k = 0; k < ndim; k++ ) {
+    parent_off += parent_start[k] * parent_strides[k];
+    input_off  += input_offset[k] * input_strides[k];
+    idx[k] = 0;
+  }
+
+  run_bytes = count[ndim - 1] * bytes;
+  outer_axes = ndim - 1;
+
+  if ( outer_axes <= 0 ) {
+    memcpy(parent_ptr + parent_off, in_ptr + input_off, run_bytes);
+    return;
+  }
+
+  if ( outer_axes == 1 ) {
+    char     *dst = parent_ptr + parent_off;
+    const char *src = in_ptr + input_off;
+    ca_size_t step_p = parent_strides[0];
+    ca_size_t step_i = input_strides[0];
+    ca_size_t count0 = count[0];
+    ca_size_t i;
+    for ( i = 0; i < count0; i++ ) {
+      memcpy(dst, src, run_bytes);
+      dst += step_p;
+      src += step_i;
+    }
+    return;
+  }
+
+  total_slabs = 1;
+  for ( k = 0; k < outer_axes; k++ ) {
+    total_slabs *= count[k];
+  }
+
+  for ( slab = 0; slab < total_slabs; slab++ ) {
+    memcpy(parent_ptr + parent_off, in_ptr + input_off, run_bytes);
+
+    for ( k = outer_axes - 1; k >= 0; k-- ) {
+      idx[k]++;
+      parent_off += parent_strides[k];
+      input_off  += input_strides[k];
+      if ( idx[k] < count[k] ) break;
+      parent_off -= parent_strides[k] * count[k];
+      input_off  -= input_strides[k] * count[k];
+      idx[k] = 0;
+    }
+  }
+}
+
+/* --------------------------------------------------------------------- */
+/* ca_composite_fill_rect                                                 */
+/*                                                                        */
+/* Fill a rectangle sub-region of out_ptr with `fill_value` (broadcast). */
+/* Mirrors _region_gather but with a fixed source value (no source       */
+/* pointer/strides).  1D / 2D / N-D fast paths.                          */
+/* --------------------------------------------------------------------- */
+
+void
+ca_composite_fill_rect (char *out_ptr,
+                        const ca_size_t *output_strides,
+                        const ca_size_t *offset,
+                        const ca_size_t *count,
+                        const char *fill_value,
+                        ca_size_t bytes,
+                        int8_t ndim)
+{
+  ca_size_t idx[CA_RANK_MAX];
+  ca_size_t output_off, total_slabs, slab;
+  int8_t    k;
+  int8_t    outer_axes;
+  ca_size_t inner_count;
+
+  /* Initial byte offset into output. */
+  output_off = 0;
+  for ( k = 0; k < ndim; k++ ) {
+    output_off += offset[k] * output_strides[k];
+    idx[k] = 0;
+  }
+
+  /* Innermost axis = SIMD typed fill of `count[ndim-1]` cells per slab. */
+  inner_count = count[ndim - 1];
+  outer_axes = ndim - 1;
+
+  /* Early exit on empty inner. */
+  if ( inner_count <= 0 ) return;
+
+  if ( outer_axes <= 0 ) {
+    /* 1-D: single typed-fill. */
+    ca_fill_typed(out_ptr + output_off, fill_value, bytes, inner_count);
+    return;
+  }
+
+  if ( outer_axes == 1 ) {
+    /* 2-D fast path. */
+    ca_size_t step_o = output_strides[0];
+    ca_size_t count0 = count[0];
+    if ( count0 <= 0 ) return;
+    /* Contig-in-output detection: if successive "rows" of the strip
+       are adjacent in output memory (= step_o == inner_count * bytes),
+       merge into a single big typed-fill so SIMD bandwidth is fully used.
+       For complement-fill axis-0 strips this case fires (= the strip spans
+       the full inner extent of output along the rows it covers). */
+    if ( step_o == inner_count * bytes ) {
+      ca_fill_typed(out_ptr + output_off, fill_value, bytes,
+                    count0 * inner_count);
+      return;
+    }
+    /* Generic per-row fill. */
+    {
+      char     *dst = out_ptr + output_off;
+      ca_size_t i;
+      for ( i = 0; i < count0; i++ ) {
+        ca_fill_typed(dst, fill_value, bytes, inner_count);
+        dst += step_o;
+      }
+    }
+    return;
+  }
+
+  /* General N-D path (ndim >= 3): collapse all trailing axes whose
+     count == output_dim (= step matches contig stride) into the inner
+     run, walk only the truly outer axes via multi-dim cursor. */
+  {
+    int8_t   walk_axes = outer_axes;
+    ca_size_t merged_inner = inner_count;
+    while ( walk_axes > 0
+         && output_strides[walk_axes - 1] == merged_inner * bytes
+         && count[walk_axes - 1] > 0 ) {
+      merged_inner *= count[walk_axes - 1];
+      walk_axes--;
+    }
+
+    if ( walk_axes == 0 ) {
+      /* Whole strip is contig in output. */
+      ca_fill_typed(out_ptr + output_off, fill_value, bytes, merged_inner);
+      return;
+    }
+
+    total_slabs = 1;
+    for ( k = 0; k < walk_axes; k++ ) {
+      if ( count[k] <= 0 ) return;
+      total_slabs *= count[k];
+    }
+
+    for ( slab = 0; slab < total_slabs; slab++ ) {
+      ca_fill_typed(out_ptr + output_off, fill_value, bytes, merged_inner);
+
+      for ( k = walk_axes - 1; k >= 0; k-- ) {
+        idx[k]++;
+        output_off += output_strides[k];
+        if ( idx[k] < count[k] ) break;
+        output_off -= output_strides[k] * count[k];
+        idx[k] = 0;
+      }
+    }
+  }
+}
+
+/* --------------------------------------------------------------------- */
+/* ca_composite_fill_complement                                           */
+/*                                                                        */
+/* Fill the complement of an alias rectangle.  Enumerates up to 2*ndim    */
+/* disjoint strips and calls _fill_rect for each non-empty strip.        */
+/*                                                                        */
+/* Strip k "before": alias range on dims 0..k-1,                          */
+/*                   [0..alias_offset[k]) on dim k,                       */
+/*                   full [0..output_dim[i]) on dims k+1..ndim-1          */
+/* Strip k "after":  alias range on dims 0..k-1,                          */
+/*                   [alias_offset[k]+alias_count[k]..output_dim[k])      */
+/*                   on dim k,                                            */
+/*                   full on dims k+1..ndim-1                             */
+/*                                                                        */
+/* Strips are pairwise disjoint and together cover output minus alias.    */
+/* Caller must ensure alias rectangle is within output bounds.            */
+/* --------------------------------------------------------------------- */
+
+void
+ca_composite_fill_complement (char *out_ptr,
+                              const ca_size_t *output_strides,
+                              const ca_size_t *output_dim,
+                              const ca_size_t *alias_offset,
+                              const ca_size_t *alias_count,
+                              const char *fill_value,
+                              ca_size_t bytes,
+                              int8_t ndim)
+{
+  ca_size_t strip_offset[CA_RANK_MAX];
+  ca_size_t strip_count[CA_RANK_MAX];
+  int8_t k, m;
+
+  for ( k = 0; k < ndim; k++ ) {
+    ca_size_t a_off = alias_offset[k];
+    ca_size_t a_cnt = alias_count[k];
+    ca_size_t a_end = a_off + a_cnt;
+
+    /* "Before" strip on axis k:
+         dims 0..k-1:    alias range
+         dim k:          [0..a_off)
+         dims k+1..N-1:  full output extent */
+    if ( a_off > 0 ) {
+      for ( m = 0; m < k; m++ ) {
+        strip_offset[m] = alias_offset[m];
+        strip_count[m]  = alias_count[m];
+      }
+      strip_offset[k] = 0;
+      strip_count[k]  = a_off;
+      for ( m = k + 1; m < ndim; m++ ) {
+        strip_offset[m] = 0;
+        strip_count[m]  = output_dim[m];
+      }
+      ca_composite_fill_rect(out_ptr, output_strides,
+                             strip_offset, strip_count,
+                             fill_value, bytes, ndim);
+    }
+
+    /* "After" strip on axis k:
+         dims 0..k-1:    alias range
+         dim k:          [a_end..output_dim[k])
+         dims k+1..N-1:  full output extent */
+    if ( a_end < output_dim[k] ) {
+      for ( m = 0; m < k; m++ ) {
+        strip_offset[m] = alias_offset[m];
+        strip_count[m]  = alias_count[m];
+      }
+      strip_offset[k] = a_end;
+      strip_count[k]  = output_dim[k] - a_end;
+      for ( m = k + 1; m < ndim; m++ ) {
+        strip_offset[m] = 0;
+        strip_count[m]  = output_dim[m];
+      }
+      ca_composite_fill_rect(out_ptr, output_strides,
+                             strip_offset, strip_count,
+                             fill_value, bytes, ndim);
+    }
+  }
+}

data/ext/ca_composite_dispatch.h

@@ -0,0 +1,116 @@
+/* ---------------------------------------------------------------------------
+
+  Geometric primitives for region-mapped views (CAWindow / CATile /
+  CARoll).  These helpers are the "per-region routine" called by each
+  region-mapped view's attach/sync path.
+
+  Design pattern (= split routines, not a tagged union over source kind):
+  - the region descriptor carries only geometric metadata
+    (output_offset, count); source info (parent, parent_start, fill_value)
+    is passed as a per-call parameter.
+  - ALIAS regions use _region_gather / _region_scatter; broadcast fill
+    is ca_fill_typed (full output) or _fill_rect / _fill_complement for
+    sub-regions.
+
+---------------------------------------------------------------------------- */
+
+#ifndef CA_COMPOSITE_DISPATCH_H
+#define CA_COMPOSITE_DISPATCH_H
+
+#include "carray.h"
+
+/* Fill `n` cells of `dst`, each `bytes` wide, with the value at `val`.
+
+   SIMD-friendly typed-store dispatch on bytes ∈ {1, 2, 4, 8, 16}, generic
+   memcpy fallback for other widths (CA_FIXLEN, unusual struct sizes).
+   Compiler vectorises the uint16/32/64 stores into NEON / SSE2 stores.
+
+   Called from ca_obj_window.c and ca_axis_dispatch.c for the
+   broadcast-fill step. */
+void ca_fill_typed (char *dst, const char *val,
+                    ca_size_t bytes, ca_size_t n);
+
+/* Strided memcpy from a sub-rectangle of `parent_ptr` to a sub-rectangle
+   of `out_ptr`.  Both rectangles have the same shape (`count`).
+
+   parent_ptr     : base pointer of parent's row-major contig buffer
+                    (caller has attached parent; parent->ptr is valid)
+   parent_strides : per-axis byte stride within parent (row-major)
+   parent_start   : per-axis start index in parent for the source rectangle
+   out_ptr        : base pointer of the output buffer (row-major contig
+                    over the view's own shape)
+   output_strides : per-axis byte stride within output
+   output_offset  : per-axis start index in output for the destination rectangle
+   count          : per-axis size of the rectangle (same for both)
+   ndim, bytes    : view ndim and element bytes
+
+   Internal fast paths:
+   - ndim == 1: single memcpy
+   - ndim == 2 (outer_axes == 1): pointer-advancing 2D loop, no idx[]
+   - ndim >= 3: multi-dim cursor walk over outer axes
+
+   Called from the attach paths of ca_obj_window.c, ca_obj_tile.c and
+   ca_obj_roll.c. */
+void ca_composite_region_gather (char *parent_ptr,
+                                 const ca_size_t *parent_strides,
+                                 const ca_size_t *parent_start,
+                                 char *out_ptr,
+                                 const ca_size_t *output_strides,
+                                 const ca_size_t *output_offset,
+                                 const ca_size_t *count,
+                                 int8_t ndim,
+                                 ca_size_t bytes);
+
+/* Reverse direction: copy from out_ptr rectangle to parent rectangle.
+   Called from the sync (write-back) paths of ca_obj_window.c,
+   ca_obj_tile.c and ca_obj_roll.c.  Same parameter geometry as
+   _region_gather. */
+void ca_composite_region_scatter (char *parent_ptr,
+                                  const ca_size_t *parent_strides,
+                                  const ca_size_t *parent_start,
+                                  const char *in_ptr,
+                                  const ca_size_t *input_strides,
+                                  const ca_size_t *input_offset,
+                                  const ca_size_t *count,
+                                  int8_t ndim,
+                                  ca_size_t bytes);
+
+/* Complement-aware fill.
+
+   ca_composite_fill_rect: fill a rectangle sub-region of out_ptr with
+   typed value `fill_value` (broadcast).  Same rectangle geometry as
+   _region_gather but the source is a single value, not a parent
+   rectangle.  1D / 2D / N-D internal fast paths.
+
+   ca_composite_fill_complement: fill the complement of an alias
+   rectangle (= output minus alias).  Enumerates up to 2*ndim disjoint
+   strip regions and calls _fill_rect for each non-empty strip.  Each
+   strip is a rectangle:
+     axis k "before" strip: alias range on dims 0..k-1, [0..alias_offset[k])
+                            on dim k, full on dims k+1..ndim-1
+     axis k "after"  strip: alias range on dims 0..k-1,
+                            [alias_offset[k]+alias_count[k]..output_dim[k])
+                            on dim k, full on dims k+1..ndim-1
+   Strips are pairwise disjoint and together cover output minus alias.
+
+   Called from CAWindow's attach path: rather than fill_typed over the
+   entire output and then overwrite the alias rectangle, fill only the
+   complement strips (= no wasted writes on the alias region). */
+void ca_composite_fill_rect       (char *out_ptr,
+                                   const ca_size_t *output_strides,
+                                   const ca_size_t *offset,
+                                   const ca_size_t *count,
+                                   const char *fill_value,
+                                   ca_size_t bytes,
+                                   int8_t ndim);
+
+void ca_composite_fill_complement (char *out_ptr,
+                                   const ca_size_t *output_strides,
+                                   const ca_size_t *output_dim,
+                                   const ca_size_t *alias_offset,
+                                   const ca_size_t *alias_count,
+                                   const char *fill_value,
+                                   ca_size_t bytes,
+                                   int8_t ndim);
+
+#endif /* CA_COMPOSITE_DISPATCH_H */