carray 2.0.0 → 3.0.0

data/ext/carray_call_cfunc.c

@@ -1,819 +1,1424 @@
 /* ---------------------------------------------------------------------------
-
-  carray_math_call.c
-
-  This file is part of Ruby/CArray extension library.
-
-  Copyright (C) 2005-2020 Hiroki Motoyoshi
-
----------------------------------------------------------------------------- */
+ *
+ *  carray_call_cfunc.c -- ext-author math-call wrapper API
+ *
+ *  GENERATED by ext/mk_call_cfunc.rb -- DO NOT EDIT
+ *  To modify the template or add a new arity, edit mk_call_cfunc.rb
+ *  and regenerate (= make / extconf re-runs the generator).
+ *
+ *  Provides `ca_call_cfunc_N` (raw arity 1..7) and `ca_call_cfunc_M_N`
+ *  (typed dispatcher) used by external CArray gems (carray-gsl, etc.)
+ *  to wrap C math functions over CArray operations.
+ *
+ *  Reentrant variants `ca_call_cfunc_N_r` and `ca_call_cfunc_M_N_r`
+ *  (= POSIX `_r` convention, cf. qsort_r / bsearch_r / strtok_r) take
+ *  a trailing `void *userdata` argument that is passed through to every
+ *  per-cell callback invocation as its last argument.  Use when the
+ *  callback needs to share state with the caller (accumulators,
+ *  configuration flags, library plan handles, ...) without resorting to
+ *  file-static / global plumbing.
+ *
+ *  Attach-safe: INPUT-only operands (= fsync[k] == '0') never call
+ *  ca_attach on the operand or its mask (= alias check + ALLOCV +
+ *  ca_xfer_all for non-alias).  OUTPUT operands (= fsync[k] == '1')
+ *  use ca_attach + ca_sync (= legitimate per the refined input-only-
+ *  operand invariant established in PROPOSAL_EAGER_ELEMENTWISE_NO_ATTACH
+ *  and extended in PROPOSAL_EAGER_SLOWPATH_CHUNKING_ARENA).
+ *
+ *  L0.1 (PROPOSAL_L0_AUTHOR_SURFACE, 2026-06-11): the per-operand acquire
+ *  + broadcast check + mask OR + release lifecycle is now factored out
+ *  into ext/ca_sweep_engine.{c,h} (ca_sweep_acquire / ca_sweep_release).  This
+ *  file keeps the arity-dependent inner loop only.  ABI unchanged.
+ *
+ *  --------------------------------------------------------------------------- */
 
 #include "carray.h"
+#include "ca_sweep_engine.h"
+#include <string.h>
 
 VALUE
 ca_call_cfunc_1 (void (*func)(void *p0), const char *fsync,
-                          VALUE rcx0)
+                            VALUE rcx0)
 {
-  CArray *cx0;
-  ca_size_t n;
-
-  if ( strlen(fsync) != 1 ) {
-    rb_raise(rb_eRuntimeError,
-             "[BUG] invalid length of fsync arg in rb_ca_call_mathfunc");
-  }
-
-  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx0);
-
-  ca_attach(cx0);
-
+  CArray   *cx[1];
+  char     *base[1];
+  ca_size_t stride[1];
+  char     *owned_buf[1];
+  int       attached[1];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+
+  /* sweep engine: per-operand acquire (alias / xmalloc + ca_xfer_all),
+     broadcast shape check, mask OR across INPUTs, mask propagate to
+     OUTPUTs.  Lifecycle template lives in ext/ca_sweep_engine.{c,h}. */
+  state.n_ops     = 1;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_1";
+
+  ca_sweep_acquire(&state);
+
+  /* inner loop: advance per-cell ptrs and invoke user kernel func */
   {
-    char      *p0;
-    char      *q0;
-    ca_size_t    s0;
-    ca_size_t    k;
-    n = ca_set_iterator(1, cx0, &q0, &s0);
-    s0 *= cx0->bytes;
-    #ifdef _OPENMP
-    #pragma omp parallel for private(p0)
-    #endif
-    for (k=0; k<n; k++) {
-      p0 = q0 + k*s0;
-      func(p0);
+    char *p[1];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 1; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0]);
+        }
+      }
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 1; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0]);
+      }
     }
   }
 
-  ca_sync(cx0);
-  ca_detach(cx0);
+  ca_sweep_release(&state);
 
   return rcx0;
 }
 
-
 VALUE
 ca_call_cfunc_2 (void (*func)(void *p0, void *p1), const char *fsync,
-                          VALUE rcx0, VALUE rcx1)
+                            VALUE rcx0, VALUE rcx1)
 {
-  CArray *cx0, *cx1;
-  boolean8_t *m0 = NULL, *m;
-  ca_size_t n;
-
-  if ( strlen(fsync) != 2 ) {
-    rb_raise(rb_eRuntimeError,
-             "[BUG] invalid length of fsync arg in rb_ca_call_mathfunc");
+  CArray   *cx[2];
+  char     *base[2];
+  ca_size_t stride[2];
+  char     *owned_buf[2];
+  int       attached[2];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+
+  /* sweep engine: per-operand acquire (alias / xmalloc + ca_xfer_all),
+     broadcast shape check, mask OR across INPUTs, mask propagate to
+     OUTPUTs.  Lifecycle template lives in ext/ca_sweep_engine.{c,h}. */
+  state.n_ops     = 2;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_2";
+
+  ca_sweep_acquire(&state);
+
+  /* inner loop: advance per-cell ptrs and invoke user kernel func */
+  {
+    char *p[2];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 2; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1]);
+        }
+      }
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 2; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1]);
+      }
+    }
   }
 
-  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx0);
-  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx1);
+  ca_sweep_release(&state);
 
-  ca_attach_n(2, cx0, cx1);
+  return rcx0;
+}
 
+VALUE
+ca_call_cfunc_3 (void (*func)(void *p0, void *p1, void *p2), const char *fsync,
+                            VALUE rcx0, VALUE rcx1, VALUE rcx2)
+{
+  CArray   *cx[3];
+  char     *base[3];
+  ca_size_t stride[3];
+  char     *owned_buf[3];
+  int       attached[3];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx[2]);
+
+  /* sweep engine: per-operand acquire (alias / xmalloc + ca_xfer_all),
+     broadcast shape check, mask OR across INPUTs, mask propagate to
+     OUTPUTs.  Lifecycle template lives in ext/ca_sweep_engine.{c,h}. */
+  state.n_ops     = 3;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_3";
+
+  ca_sweep_acquire(&state);
+
+  /* inner loop: advance per-cell ptrs and invoke user kernel func */
   {
-    CArray *cx[2];
-    int i = 0;
-    if ( fsync[0] == '0' ) cx[i++] = cx0;
-    if ( fsync[1] == '0' ) cx[i++] = cx1;
-    m = m0 = ca_allocate_mask_iterator_n(i, cx);
-    if ( fsync[0] == '1' ) ca_copy_mask_overwrite_n(cx0, cx0->elements, i, cx);
-    if ( fsync[1] == '1' ) ca_copy_mask_overwrite_n(cx1, cx1->elements, i, cx);
+    char *p[3];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 3; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1], p[2]);
+        }
+      }
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 3; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1], p[2]);
+      }
+    }
   }
 
+  ca_sweep_release(&state);
+
+  return rcx0;
+}
+
+VALUE
+ca_call_cfunc_4 (void (*func)(void *p0, void *p1, void *p2, void *p3), const char *fsync,
+                            VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3)
+{
+  CArray   *cx[4];
+  char     *base[4];
+  ca_size_t stride[4];
+  char     *owned_buf[4];
+  int       attached[4];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx[2]);
+  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx[3]);
+
+  /* sweep engine: per-operand acquire (alias / xmalloc + ca_xfer_all),
+     broadcast shape check, mask OR across INPUTs, mask propagate to
+     OUTPUTs.  Lifecycle template lives in ext/ca_sweep_engine.{c,h}. */
+  state.n_ops     = 4;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_4";
+
+  ca_sweep_acquire(&state);
+
+  /* inner loop: advance per-cell ptrs and invoke user kernel func */
   {
-    char      *p0, *p1;
-    char      *q0, *q1;
-    ca_size_t    s0,  s1;
-    ca_size_t    k;
-
-    n = ca_set_iterator(2, cx0, &q0, &s0,
-                           cx1, &q1, &s1);
-    s0 *= cx0->bytes;
-    s1 *= cx1->bytes;
-
-    if ( m0 ) {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1)
-      #endif
-      for (k=0; k<n; k++) {
-        m = m0 + k;
-        if ( ! *m ) { 
-          p0 = q0 + k*s0; 
-          p1 = q1 + k*s1;
-          func(p0, p1);        
+    char *p[4];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 4; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1], p[2], p[3]);
         }
       }
-    }
-    else {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1)
-      #endif
-      for (k=0; k<n; k++) {
-        p0 = q0 + k*s0; 
-        p1 = q1 + k*s1;
-        func(p0, p1);        
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 4; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1], p[2], p[3]);
       }
     }
   }
-  if ( fsync[0] == '1' ) ca_sync(cx0);
-  if ( fsync[1] == '1' ) ca_sync(cx1);
-  ca_detach_n(2, cx0, cx1);
 
-  free(m0);
+  ca_sweep_release(&state);
 
   return rcx0;
 }
 
 VALUE
-ca_call_cfunc_3 (void (*func)(void *p0, void *p1, void *p2), const char *fsync,
-                          VALUE rcx0, VALUE rcx1, VALUE rcx2)
+ca_call_cfunc_5 (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4), const char *fsync,
+                            VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3, VALUE rcx4)
 {
-  CArray *cx0, *cx1, *cx2;
-  boolean8_t *m0 = NULL, *m;
-  ca_size_t n;
-
-  if ( strlen(fsync) != 3 ) {
-    rb_raise(rb_eRuntimeError,
-             "[BUG] invalid length of fsync arg in rb_ca_call_mathfunc");
+  CArray   *cx[5];
+  char     *base[5];
+  ca_size_t stride[5];
+  char     *owned_buf[5];
+  int       attached[5];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx[2]);
+  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx[3]);
+  TypedData_Get_Struct(rcx4, CArray, &carray_data_type, cx[4]);
+
+  /* sweep engine: per-operand acquire (alias / xmalloc + ca_xfer_all),
+     broadcast shape check, mask OR across INPUTs, mask propagate to
+     OUTPUTs.  Lifecycle template lives in ext/ca_sweep_engine.{c,h}. */
+  state.n_ops     = 5;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_5";
+
+  ca_sweep_acquire(&state);
+
+  /* inner loop: advance per-cell ptrs and invoke user kernel func */
+  {
+    char *p[5];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 5; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1], p[2], p[3], p[4]);
+        }
+      }
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 5; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1], p[2], p[3], p[4]);
+      }
+    }
   }
 
-  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx0);
-  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx1);
-  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx2);
+  ca_sweep_release(&state);
 
-  ca_attach_n(3, cx0, cx1, cx2);
+  return rcx0;
+}
 
+VALUE
+ca_call_cfunc_6 (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4, void *p5), const char *fsync,
+                            VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3, VALUE rcx4, VALUE rcx5)
+{
+  CArray   *cx[6];
+  char     *base[6];
+  ca_size_t stride[6];
+  char     *owned_buf[6];
+  int       attached[6];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx[2]);
+  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx[3]);
+  TypedData_Get_Struct(rcx4, CArray, &carray_data_type, cx[4]);
+  TypedData_Get_Struct(rcx5, CArray, &carray_data_type, cx[5]);
+
+  /* sweep engine: per-operand acquire (alias / xmalloc + ca_xfer_all),
+     broadcast shape check, mask OR across INPUTs, mask propagate to
+     OUTPUTs.  Lifecycle template lives in ext/ca_sweep_engine.{c,h}. */
+  state.n_ops     = 6;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_6";
+
+  ca_sweep_acquire(&state);
+
+  /* inner loop: advance per-cell ptrs and invoke user kernel func */
   {
-    CArray *cx[3];
-    int i = 0;
-    if ( fsync[0] == '0' ) cx[i++] = cx0;
-    if ( fsync[1] == '0' ) cx[i++] = cx1;
-    if ( fsync[2] == '0' ) cx[i++] = cx2;
-    m = m0 = ca_allocate_mask_iterator_n(i, cx);
-    if ( fsync[0] == '1' ) ca_copy_mask_overwrite_n(cx0, cx0->elements, i, cx);
-    if ( fsync[1] == '1' ) ca_copy_mask_overwrite_n(cx1, cx1->elements, i, cx);
-    if ( fsync[2] == '1' ) ca_copy_mask_overwrite_n(cx2, cx2->elements, i, cx);
+    char *p[6];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 6; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1], p[2], p[3], p[4], p[5]);
+        }
+      }
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 6; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1], p[2], p[3], p[4], p[5]);
+      }
+    }
   }
 
+  ca_sweep_release(&state);
+
+  return rcx0;
+}
+
+VALUE
+ca_call_cfunc_7 (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4, void *p5, void *p6), const char *fsync,
+                            VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3, VALUE rcx4, VALUE rcx5, VALUE rcx6)
+{
+  CArray   *cx[7];
+  char     *base[7];
+  ca_size_t stride[7];
+  char     *owned_buf[7];
+  int       attached[7];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx[2]);
+  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx[3]);
+  TypedData_Get_Struct(rcx4, CArray, &carray_data_type, cx[4]);
+  TypedData_Get_Struct(rcx5, CArray, &carray_data_type, cx[5]);
+  TypedData_Get_Struct(rcx6, CArray, &carray_data_type, cx[6]);
+
+  /* sweep engine: per-operand acquire (alias / xmalloc + ca_xfer_all),
+     broadcast shape check, mask OR across INPUTs, mask propagate to
+     OUTPUTs.  Lifecycle template lives in ext/ca_sweep_engine.{c,h}. */
+  state.n_ops     = 7;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_7";
+
+  ca_sweep_acquire(&state);
+
+  /* inner loop: advance per-cell ptrs and invoke user kernel func */
   {
-    char      *p0, *p1, *p2;
-    char      *q0, *q1, *q2;
-    ca_size_t    s0,  s1,  s2;
-    ca_size_t    k;
-
-    n = ca_set_iterator(3, cx0, &q0, &s0,
-                           cx1, &q1, &s1,
-                           cx2, &q2, &s2);
-    s0 *= cx0->bytes;
-    s1 *= cx1->bytes;
-    s2 *= cx2->bytes;
-
-    if ( m0 ) {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1,p2)
-      #endif
-      for (k=0; k<n; k++) {
-        m = m0 + k;
-        if ( ! *m ) { 
-          p0 = q0 + k*s0; 
-          p1 = q1 + k*s1;
-          p2 = q2 + k*s2;
-          func(p0, p1, p2);      
+    char *p[7];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 7; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1], p[2], p[3], p[4], p[5], p[6]);
         }
       }
-    }
-    else {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1,p2)
-      #endif
-      for (k=0; k<n; k++) {
-        p0 = q0 + k*s0; 
-        p1 = q1 + k*s1;
-        p2 = q2 + k*s2;
-        func(p0, p1, p2);      
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 7; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1], p[2], p[3], p[4], p[5], p[6]);
       }
     }
   }
-  if ( fsync[0] == '1' ) ca_sync(cx0);
-  if ( fsync[1] == '1' ) ca_sync(cx1);
-  if ( fsync[2] == '1' ) ca_sync(cx2);
-  ca_detach_n(3, cx0, cx1, cx2);
 
-  free(m0);
+  ca_sweep_release(&state);
 
   return rcx0;
 }
 
+/* -------------------------------------------------------------------- */
+/* Typed dispatchers: M outputs, N inputs.  Wrap each input readonly to */
+/* its declared data_type, allocate output template(s), delegate to raw.    */
+/* -------------------------------------------------------------------- */
+
 VALUE
-ca_call_cfunc_4 (void (*func)(void *p0, void *p1, void *p2, void *p3), const char *fsync,
-                                            VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3)
+ca_call_cfunc_1_1 (int8_t dty, int8_t dtx1, void (*mathfunc)(void*, void*), volatile VALUE rx1)
 {
-  CArray *cx0, *cx1, *cx2, *cx3;
-  boolean8_t *m0 = NULL, *m;
-  ca_size_t n;
+  volatile VALUE ry = Qnil;
 
-  if ( strlen(fsync) != 4 ) {
-    rb_raise(rb_eRuntimeError,
-             "[BUG] invalid length of fsync arg in rb_ca_call_mathfunc");
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+
+  if ( dty != dtx1 ) {
+    ry = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(1, rx1);
   }
 
-  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx0);
-  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx1);
-  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx2);
-  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx3);
+  ca_call_cfunc_2(mathfunc, "10", ry, rx1);
 
-  ca_attach_n(4, cx0, cx1, cx2, cx3);
+  if ( rb_ca_is_scalar(ry) ) {
+    ry = rb_ca_fetch_addr(ry, 0);
+  }
+  return ry;
+}
 
-  {
-    CArray *cx[4];
-    int i = 0;
-    if ( fsync[0] == '0' ) cx[i++] = cx0;
-    if ( fsync[1] == '0' ) cx[i++] = cx1;
-    if ( fsync[2] == '0' ) cx[i++] = cx2;
-    if ( fsync[3] == '0' ) cx[i++] = cx3;
-    m = m0 = ca_allocate_mask_iterator_n(i, cx);
-    if ( fsync[0] == '1' ) ca_copy_mask_overwrite_n(cx0, cx0->elements, i, cx);
-    if ( fsync[1] == '1' ) ca_copy_mask_overwrite_n(cx1, cx1->elements, i, cx);
-    if ( fsync[2] == '1' ) ca_copy_mask_overwrite_n(cx2, cx2->elements, i, cx);
-    if ( fsync[3] == '1' ) ca_copy_mask_overwrite_n(cx3, cx3->elements, i, cx);
+VALUE
+ca_call_cfunc_1_2 (int8_t dty, int8_t dtx1, int8_t dtx2, void (*mathfunc)(void*, void*, void*), volatile VALUE rx1, volatile VALUE rx2)
+{
+  volatile VALUE ry = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+
+  if ( dty != dtx1 || dty != dtx2 ) {
+    ry = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty)), rb_ca_wrap_readonly(rx2, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(2, rx1, rx2);
+  }
+
+  ca_call_cfunc_3(mathfunc, "100", ry, rx1, rx2);
+
+  if ( rb_ca_is_scalar(ry) ) {
+    ry = rb_ca_fetch_addr(ry, 0);
+  }
+  return ry;
+}
+
+VALUE
+ca_call_cfunc_1_3 (int8_t dty, int8_t dtx1, int8_t dtx2, int8_t dtx3, void (*mathfunc)(void*, void*, void*, void*), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3)
+{
+  volatile VALUE ry = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+
+  if ( dty != dtx1 || dty != dtx2 || dty != dtx3 ) {
+    ry = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty)), rb_ca_wrap_readonly(rx2, INT2NUM(dty)), rb_ca_wrap_readonly(rx3, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(3, rx1, rx2, rx3);
+  }
+
+  ca_call_cfunc_4(mathfunc, "1000", ry, rx1, rx2, rx3);
+
+  if ( rb_ca_is_scalar(ry) ) {
+    ry = rb_ca_fetch_addr(ry, 0);
+  }
+  return ry;
+}
+
+VALUE
+ca_call_cfunc_1_4 (int8_t dty, int8_t dtx1, int8_t dtx2, int8_t dtx3, int8_t dtx4, void (*mathfunc)(void*, void*, void*, void*, void*), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3, volatile VALUE rx4)
+{
+  volatile VALUE ry = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+  rx4 = rb_ca_wrap_readonly(rx4, INT2NUM(dtx4));
+
+  if ( dty != dtx1 || dty != dtx2 || dty != dtx3 || dty != dtx4 ) {
+    ry = rb_ca_template_n(4, rb_ca_wrap_readonly(rx1, INT2NUM(dty)), rb_ca_wrap_readonly(rx2, INT2NUM(dty)), rb_ca_wrap_readonly(rx3, INT2NUM(dty)), rb_ca_wrap_readonly(rx4, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(4, rx1, rx2, rx3, rx4);
+  }
+
+  ca_call_cfunc_5(mathfunc, "10000", ry, rx1, rx2, rx3, rx4);
+
+  if ( rb_ca_is_scalar(ry) ) {
+    ry = rb_ca_fetch_addr(ry, 0);
+  }
+  return ry;
+}
+
+VALUE
+ca_call_cfunc_1_5 (int8_t dty, int8_t dtx1, int8_t dtx2, int8_t dtx3, int8_t dtx4, int8_t dtx5, void (*mathfunc)(void*, void*, void*, void*, void*, void*), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3, volatile VALUE rx4, volatile VALUE rx5)
+{
+  volatile VALUE ry = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+  rx4 = rb_ca_wrap_readonly(rx4, INT2NUM(dtx4));
+  rx5 = rb_ca_wrap_readonly(rx5, INT2NUM(dtx5));
+
+  if ( dty != dtx1 || dty != dtx2 || dty != dtx3 || dty != dtx4 || dty != dtx5 ) {
+    ry = rb_ca_template_n(5, rb_ca_wrap_readonly(rx1, INT2NUM(dty)), rb_ca_wrap_readonly(rx2, INT2NUM(dty)), rb_ca_wrap_readonly(rx3, INT2NUM(dty)), rb_ca_wrap_readonly(rx4, INT2NUM(dty)), rb_ca_wrap_readonly(rx5, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(5, rx1, rx2, rx3, rx4, rx5);
+  }
+
+  ca_call_cfunc_6(mathfunc, "100000", ry, rx1, rx2, rx3, rx4, rx5);
+
+  if ( rb_ca_is_scalar(ry) ) {
+    ry = rb_ca_fetch_addr(ry, 0);
+  }
+  return ry;
+}
+
+VALUE
+ca_call_cfunc_1_6 (int8_t dty, int8_t dtx1, int8_t dtx2, int8_t dtx3, int8_t dtx4, int8_t dtx5, int8_t dtx6, void (*mathfunc)(void*, void*, void*, void*, void*, void*, void*), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3, volatile VALUE rx4, volatile VALUE rx5, volatile VALUE rx6)
+{
+  volatile VALUE ry = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+  rx4 = rb_ca_wrap_readonly(rx4, INT2NUM(dtx4));
+  rx5 = rb_ca_wrap_readonly(rx5, INT2NUM(dtx5));
+  rx6 = rb_ca_wrap_readonly(rx6, INT2NUM(dtx6));
+
+  if ( dty != dtx1 || dty != dtx2 || dty != dtx3 || dty != dtx4 || dty != dtx5 || dty != dtx6 ) {
+    ry = rb_ca_template_n(6, rb_ca_wrap_readonly(rx1, INT2NUM(dty)), rb_ca_wrap_readonly(rx2, INT2NUM(dty)), rb_ca_wrap_readonly(rx3, INT2NUM(dty)), rb_ca_wrap_readonly(rx4, INT2NUM(dty)), rb_ca_wrap_readonly(rx5, INT2NUM(dty)), rb_ca_wrap_readonly(rx6, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(6, rx1, rx2, rx3, rx4, rx5, rx6);
+  }
+
+  ca_call_cfunc_7(mathfunc, "1000000", ry, rx1, rx2, rx3, rx4, rx5, rx6);
+
+  if ( rb_ca_is_scalar(ry) ) {
+    ry = rb_ca_fetch_addr(ry, 0);
+  }
+  return ry;
+}
+
+VALUE
+ca_call_cfunc_2_1 (int8_t dty1, int8_t dty2, int8_t dtx1, void (*mathfunc)(void*, void*, void*), volatile VALUE rx1)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+
+  if ( dty1 != dtx1 ) {
+    ry1 = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(1, rx1);
+  }
+  if ( dty2 != dtx1 ) {
+    ry2 = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(1, rx1);
+  }
+
+  ca_call_cfunc_3(mathfunc, "110", ry1, ry2, rx1);
+
+  if ( rb_ca_is_scalar(ry1) ) {
+    ry1 = rb_ca_fetch_addr(ry1, 0);
+  }
+  if ( rb_ca_is_scalar(ry2) ) {
+    ry2 = rb_ca_fetch_addr(ry2, 0);
+  }
+  return rb_ary_new3(2, ry1, ry2);
+}
+
+VALUE
+ca_call_cfunc_2_2 (int8_t dty1, int8_t dty2, int8_t dtx1, int8_t dtx2, void (*mathfunc)(void*, void*, void*, void*), volatile VALUE rx1, volatile VALUE rx2)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+
+  if ( dty1 != dtx1 || dty1 != dtx2 ) {
+    ry1 = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), rb_ca_wrap_readonly(rx2, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(2, rx1, rx2);
+  }
+  if ( dty2 != dtx1 || dty2 != dtx2 ) {
+    ry2 = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), rb_ca_wrap_readonly(rx2, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(2, rx1, rx2);
+  }
+
+  ca_call_cfunc_4(mathfunc, "1100", ry1, ry2, rx1, rx2);
+
+  if ( rb_ca_is_scalar(ry1) ) {
+    ry1 = rb_ca_fetch_addr(ry1, 0);
+  }
+  if ( rb_ca_is_scalar(ry2) ) {
+    ry2 = rb_ca_fetch_addr(ry2, 0);
+  }
+  return rb_ary_new3(2, ry1, ry2);
+}
+
+VALUE
+ca_call_cfunc_2_3 (int8_t dty1, int8_t dty2, int8_t dtx1, int8_t dtx2, int8_t dtx3, void (*mathfunc)(void*, void*, void*, void*, void*), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+
+  if ( dty1 != dtx1 || dty1 != dtx2 || dty1 != dtx3 ) {
+    ry1 = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), rb_ca_wrap_readonly(rx2, INT2NUM(dty1)), rb_ca_wrap_readonly(rx3, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(3, rx1, rx2, rx3);
+  }
+  if ( dty2 != dtx1 || dty2 != dtx2 || dty2 != dtx3 ) {
+    ry2 = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), rb_ca_wrap_readonly(rx2, INT2NUM(dty2)), rb_ca_wrap_readonly(rx3, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(3, rx1, rx2, rx3);
+  }
+
+  ca_call_cfunc_5(mathfunc, "11000", ry1, ry2, rx1, rx2, rx3);
+
+  if ( rb_ca_is_scalar(ry1) ) {
+    ry1 = rb_ca_fetch_addr(ry1, 0);
+  }
+  if ( rb_ca_is_scalar(ry2) ) {
+    ry2 = rb_ca_fetch_addr(ry2, 0);
+  }
+  return rb_ary_new3(2, ry1, ry2);
+}
+
+VALUE
+ca_call_cfunc_2_4 (int8_t dty1, int8_t dty2, int8_t dtx1, int8_t dtx2, int8_t dtx3, int8_t dtx4, void (*mathfunc)(void*, void*, void*, void*, void*, void*), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3, volatile VALUE rx4)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+  rx4 = rb_ca_wrap_readonly(rx4, INT2NUM(dtx4));
+
+  if ( dty1 != dtx1 || dty1 != dtx2 || dty1 != dtx3 || dty1 != dtx4 ) {
+    ry1 = rb_ca_template_n(4, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), rb_ca_wrap_readonly(rx2, INT2NUM(dty1)), rb_ca_wrap_readonly(rx3, INT2NUM(dty1)), rb_ca_wrap_readonly(rx4, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(4, rx1, rx2, rx3, rx4);
+  }
+  if ( dty2 != dtx1 || dty2 != dtx2 || dty2 != dtx3 || dty2 != dtx4 ) {
+    ry2 = rb_ca_template_n(4, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), rb_ca_wrap_readonly(rx2, INT2NUM(dty2)), rb_ca_wrap_readonly(rx3, INT2NUM(dty2)), rb_ca_wrap_readonly(rx4, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(4, rx1, rx2, rx3, rx4);
+  }
+
+  ca_call_cfunc_6(mathfunc, "110000", ry1, ry2, rx1, rx2, rx3, rx4);
+
+  if ( rb_ca_is_scalar(ry1) ) {
+    ry1 = rb_ca_fetch_addr(ry1, 0);
+  }
+  if ( rb_ca_is_scalar(ry2) ) {
+    ry2 = rb_ca_fetch_addr(ry2, 0);
+  }
+  return rb_ary_new3(2, ry1, ry2);
+}
+
+VALUE
+ca_call_cfunc_3_1 (int8_t dty1, int8_t dty2, int8_t dty3, int8_t dtx1, void (*mathfunc)(void*, void*, void*, void*), volatile VALUE rx1)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil, ry3 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+
+  if ( dty1 != dtx1 ) {
+    ry1 = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(1, rx1);
+  }
+  if ( dty2 != dtx1 ) {
+    ry2 = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(1, rx1);
+  }
+  if ( dty3 != dtx1 ) {
+    ry3 = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty3)));
+  } else {
+    ry3 = rb_ca_template_n(1, rx1);
+  }
+
+  ca_call_cfunc_4(mathfunc, "1110", ry1, ry2, ry3, rx1);
+
+  if ( rb_ca_is_scalar(ry1) ) {
+    ry1 = rb_ca_fetch_addr(ry1, 0);
+  }
+  if ( rb_ca_is_scalar(ry2) ) {
+    ry2 = rb_ca_fetch_addr(ry2, 0);
+  }
+  if ( rb_ca_is_scalar(ry3) ) {
+    ry3 = rb_ca_fetch_addr(ry3, 0);
+  }
+  return rb_ary_new3(3, ry1, ry2, ry3);
+}
+
+VALUE
+ca_call_cfunc_3_2 (int8_t dty1, int8_t dty2, int8_t dty3, int8_t dtx1, int8_t dtx2, void (*mathfunc)(void*, void*, void*, void*, void*), volatile VALUE rx1, volatile VALUE rx2)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil, ry3 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+
+  if ( dty1 != dtx1 || dty1 != dtx2 ) {
+    ry1 = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), rb_ca_wrap_readonly(rx2, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(2, rx1, rx2);
+  }
+  if ( dty2 != dtx1 || dty2 != dtx2 ) {
+    ry2 = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), rb_ca_wrap_readonly(rx2, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(2, rx1, rx2);
+  }
+  if ( dty3 != dtx1 || dty3 != dtx2 ) {
+    ry3 = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty3)), rb_ca_wrap_readonly(rx2, INT2NUM(dty3)));
+  } else {
+    ry3 = rb_ca_template_n(2, rx1, rx2);
+  }
+
+  ca_call_cfunc_5(mathfunc, "11100", ry1, ry2, ry3, rx1, rx2);
+
+  if ( rb_ca_is_scalar(ry1) ) {
+    ry1 = rb_ca_fetch_addr(ry1, 0);
+  }
+  if ( rb_ca_is_scalar(ry2) ) {
+    ry2 = rb_ca_fetch_addr(ry2, 0);
+  }
+  if ( rb_ca_is_scalar(ry3) ) {
+    ry3 = rb_ca_fetch_addr(ry3, 0);
+  }
+  return rb_ary_new3(3, ry1, ry2, ry3);
+}
+
+VALUE
+ca_call_cfunc_3_3 (int8_t dty1, int8_t dty2, int8_t dty3, int8_t dtx1, int8_t dtx2, int8_t dtx3, void (*mathfunc)(void*, void*, void*, void*, void*, void*), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil, ry3 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+
+  if ( dty1 != dtx1 || dty1 != dtx2 || dty1 != dtx3 ) {
+    ry1 = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), rb_ca_wrap_readonly(rx2, INT2NUM(dty1)), rb_ca_wrap_readonly(rx3, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(3, rx1, rx2, rx3);
+  }
+  if ( dty2 != dtx1 || dty2 != dtx2 || dty2 != dtx3 ) {
+    ry2 = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), rb_ca_wrap_readonly(rx2, INT2NUM(dty2)), rb_ca_wrap_readonly(rx3, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(3, rx1, rx2, rx3);
+  }
+  if ( dty3 != dtx1 || dty3 != dtx2 || dty3 != dtx3 ) {
+    ry3 = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty3)), rb_ca_wrap_readonly(rx2, INT2NUM(dty3)), rb_ca_wrap_readonly(rx3, INT2NUM(dty3)));
+  } else {
+    ry3 = rb_ca_template_n(3, rx1, rx2, rx3);
+  }
+
+  ca_call_cfunc_6(mathfunc, "111000", ry1, ry2, ry3, rx1, rx2, rx3);
+
+  if ( rb_ca_is_scalar(ry1) ) {
+    ry1 = rb_ca_fetch_addr(ry1, 0);
+  }
+  if ( rb_ca_is_scalar(ry2) ) {
+    ry2 = rb_ca_fetch_addr(ry2, 0);
+  }
+  if ( rb_ca_is_scalar(ry3) ) {
+    ry3 = rb_ca_fetch_addr(ry3, 0);
   }
+  return rb_ary_new3(3, ry1, ry2, ry3);
+}
+
+/* -------------------------------------------------------------------- */
+/* Reentrant variants (`_r` suffix, POSIX convention): same as the    */
+/* corresponding non-`_r` variant but the callback takes a trailing   */
+/* `void *userdata` argument and the outer function takes a trailing  */
+/* `void *userdata` parameter forwarded to every per-cell invocation. */
+/* -------------------------------------------------------------------- */
+
+VALUE
+ca_call_cfunc_1_r (void (*func)(void *p0, void *userdata), const char *fsync,
+                              VALUE rcx0,
+                              void *userdata)
+{
+  CArray   *cx[1];
+  char     *base[1];
+  ca_size_t stride[1];
+  char     *owned_buf[1];
+  int       attached[1];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+
+  /* sweep engine: same lifecycle as ca_call_cfunc_1; the difference is
+     the per-cell `func(...)` call has `userdata` as its last argument. */
+  state.n_ops     = 1;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_1_r";
+
+  ca_sweep_acquire(&state);
 
   {
-    char      *p0, *p1, *p2, *p3;
-    char      *q0, *q1, *q2, *q3;
-    ca_size_t    s0,  s1,  s2,  s3;
-    ca_size_t    k;
-
-    n = ca_set_iterator(4, cx0, &q0, &s0,
-                           cx1, &q1, &s1,
-                           cx2, &q2, &s2,
-                           cx3, &q3, &s3);
-
-    s0 *= cx0->bytes;
-    s1 *= cx1->bytes;
-    s2 *= cx2->bytes;
-    s3 *= cx3->bytes;
-
-    if ( m0 ) {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1,p2,p3)
-      #endif
-      for (k=0; k<n; k++) {
-        m = m0 + k;
-        if ( ! *m ) { 
-          p0 = q0 + k*s0; 
-          p1 = q1 + k*s1;
-          p2 = q2 + k*s2;
-          p3 = q3 + k*s3;
-          func(p0, p1, p2, p3);      
+    char *p[1];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 1; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], userdata);
         }
       }
-    }
-    else {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1,p2,p3)
-      #endif
-      for (k=0; k<n; k++) {
-        p0 = q0 + k*s0; 
-        p1 = q1 + k*s1;
-        p2 = q2 + k*s2;
-        p3 = q3 + k*s3;
-        func(p0, p1, p2, p3);      
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 1; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], userdata);
       }
     }
   }
-  if ( fsync[0] == '1' ) ca_sync(cx0);
-  if ( fsync[1] == '1' ) ca_sync(cx1);
-  if ( fsync[2] == '1' ) ca_sync(cx2);
-  if ( fsync[3] == '1' ) ca_sync(cx3);
-  ca_detach_n(4, cx0, cx1, cx2, cx3);
 
-  free(m0);
+  ca_sweep_release(&state);
 
   return rcx0;
 }
 
 VALUE
-ca_call_cfunc_5 (void (*func)(void*,void*,void*,void*,void*), const char *fsync,
-                                            VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3, VALUE rcx4)
+ca_call_cfunc_2_r (void (*func)(void *p0, void *p1, void *userdata), const char *fsync,
+                              VALUE rcx0, VALUE rcx1,
+                              void *userdata)
 {
-  CArray *cx0, *cx1, *cx2, *cx3, *cx4;
-  boolean8_t *m0 = NULL, *m;
-  ca_size_t n;
+  CArray   *cx[2];
+  char     *base[2];
+  ca_size_t stride[2];
+  char     *owned_buf[2];
+  int       attached[2];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+
+  /* sweep engine: same lifecycle as ca_call_cfunc_2; the difference is
+     the per-cell `func(...)` call has `userdata` as its last argument. */
+  state.n_ops     = 2;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_2_r";
+
+  ca_sweep_acquire(&state);
 
-  if ( strlen(fsync) != 5 ) {
-    rb_raise(rb_eRuntimeError,
-             "[BUG] invalid length of fsync arg in rb_ca_call_mathfunc");
+  {
+    char *p[2];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 2; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1], userdata);
+        }
+      }
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 2; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1], userdata);
+      }
+    }
   }
 
-  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx0);
-  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx1);
-  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx2);
-  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx3);
-  TypedData_Get_Struct(rcx4, CArray, &carray_data_type, cx4);
+  ca_sweep_release(&state);
 
-  ca_attach_n(5, cx0, cx1, cx2, cx3, cx4);
+  return rcx0;
+}
 
-  {
-    CArray *cx[5];
-    int i = 0;
-    if ( fsync[0] == '0' ) cx[i++] = cx0;
-    if ( fsync[1] == '0' ) cx[i++] = cx1;
-    if ( fsync[2] == '0' ) cx[i++] = cx2;
-    if ( fsync[3] == '0' ) cx[i++] = cx3;
-    if ( fsync[4] == '0' ) cx[i++] = cx4;
-    m = m0 = ca_allocate_mask_iterator_n(i, cx);
-    if ( fsync[0] == '1' ) ca_copy_mask_overwrite_n(cx0, cx0->elements, i, cx);
-    if ( fsync[1] == '1' ) ca_copy_mask_overwrite_n(cx1, cx1->elements, i, cx);
-    if ( fsync[2] == '1' ) ca_copy_mask_overwrite_n(cx2, cx2->elements, i, cx);
-    if ( fsync[3] == '1' ) ca_copy_mask_overwrite_n(cx3, cx3->elements, i, cx);
-    if ( fsync[4] == '1' ) ca_copy_mask_overwrite_n(cx4, cx4->elements, i, cx);
-  }
+VALUE
+ca_call_cfunc_3_r (void (*func)(void *p0, void *p1, void *p2, void *userdata), const char *fsync,
+                              VALUE rcx0, VALUE rcx1, VALUE rcx2,
+                              void *userdata)
+{
+  CArray   *cx[3];
+  char     *base[3];
+  ca_size_t stride[3];
+  char     *owned_buf[3];
+  int       attached[3];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx[2]);
+
+  /* sweep engine: same lifecycle as ca_call_cfunc_3; the difference is
+     the per-cell `func(...)` call has `userdata` as its last argument. */
+  state.n_ops     = 3;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_3_r";
+
+  ca_sweep_acquire(&state);
 
   {
-    char      *p0, *p1, *p2, *p3, *p4;
-    char      *q0, *q1, *q2, *q3, *q4;
-    ca_size_t    s0,  s1,  s2,  s3,  s4;
-    ca_size_t    k;
-
-    n = ca_set_iterator(5, cx0, &q0, &s0,
-                           cx1, &q1, &s1,
-                           cx2, &q2, &s2,
-                           cx3, &q3, &s3,
-                           cx4, &q4, &s4);
-
-    s0 *= cx0->bytes;
-    s1 *= cx1->bytes;
-    s2 *= cx2->bytes;
-    s3 *= cx3->bytes;
-    s4 *= cx4->bytes;
-
-    if ( m0 ) {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1,p2,p3,p4)
-      #endif
-      for (k=0; k<n; k++) {
-        m = m0 + k;
-        if ( ! *m ) { 
-          p0 = q0 + k*s0; 
-          p1 = q1 + k*s1;
-          p2 = q2 + k*s2;
-          p3 = q3 + k*s3;
-          p4 = q4 + k*s4;
-          func(p0, p1, p2, p3, p4);      
+    char *p[3];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 3; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1], p[2], userdata);
         }
       }
-    }
-    else {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1,p2,p3,p4)
-      #endif
-      for (k=0; k<n; k++) {
-        p0 = q0 + k*s0; 
-        p1 = q1 + k*s1;
-        p2 = q2 + k*s2;
-        p3 = q3 + k*s3;
-        p4 = q4 + k*s4;
-        func(p0, p1, p2, p3, p4);      
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 3; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1], p[2], userdata);
       }
     }
   }
-  if ( fsync[0] == '1' ) ca_sync(cx0);
-  if ( fsync[1] == '1' ) ca_sync(cx1);
-  if ( fsync[2] == '1' ) ca_sync(cx2);
-  if ( fsync[3] == '1' ) ca_sync(cx3);
-  if ( fsync[4] == '1' ) ca_sync(cx4);
-  ca_detach_n(5, cx0, cx1, cx2, cx3, cx4);
 
-  free(m0);
+  ca_sweep_release(&state);
 
   return rcx0;
 }
 
 VALUE
-ca_call_cfunc_6 (void (*func)(void*,void*,void*,void*,void*,void*), const char *fsync,
-                                            VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3, VALUE rcx4, VALUE rcx5)
+ca_call_cfunc_4_r (void (*func)(void *p0, void *p1, void *p2, void *p3, void *userdata), const char *fsync,
+                              VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3,
+                              void *userdata)
 {
-  CArray *cx0, *cx1, *cx2, *cx3, *cx4, *cx5;
-  boolean8_t *m0 = NULL, *m;
-  ca_size_t n;
+  CArray   *cx[4];
+  char     *base[4];
+  ca_size_t stride[4];
+  char     *owned_buf[4];
+  int       attached[4];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx[2]);
+  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx[3]);
+
+  /* sweep engine: same lifecycle as ca_call_cfunc_4; the difference is
+     the per-cell `func(...)` call has `userdata` as its last argument. */
+  state.n_ops     = 4;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_4_r";
+
+  ca_sweep_acquire(&state);
 
-  if ( strlen(fsync) != 6 ) {
-    rb_raise(rb_eRuntimeError,
-             "[BUG] invalid length of fsync arg in rb_ca_call_mathfunc");
+  {
+    char *p[4];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 4; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1], p[2], p[3], userdata);
+        }
+      }
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 4; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1], p[2], p[3], userdata);
+      }
+    }
   }
 
-  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx0);
-  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx1);
-  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx2);
-  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx3);
-  TypedData_Get_Struct(rcx4, CArray, &carray_data_type, cx4);
-  TypedData_Get_Struct(rcx5, CArray, &carray_data_type, cx5);
+  ca_sweep_release(&state);
 
-  ca_attach_n(6, cx0, cx1, cx2, cx3, cx4, cx5);
+  return rcx0;
+}
 
-  {
-    CArray *cx[6];
-    int i = 0;
-    if ( fsync[0] == '0' ) cx[i++] = cx0;
-    if ( fsync[1] == '0' ) cx[i++] = cx1;
-    if ( fsync[2] == '0' ) cx[i++] = cx2;
-    if ( fsync[3] == '0' ) cx[i++] = cx3;
-    if ( fsync[4] == '0' ) cx[i++] = cx4;
-    if ( fsync[5] == '0' ) cx[i++] = cx5;
-    m = m0 = ca_allocate_mask_iterator_n(i, cx);
-    if ( fsync[0] == '1' ) ca_copy_mask_overwrite_n(cx0, cx0->elements, i, cx);
-    if ( fsync[1] == '1' ) ca_copy_mask_overwrite_n(cx1, cx1->elements, i, cx);
-    if ( fsync[2] == '1' ) ca_copy_mask_overwrite_n(cx2, cx2->elements, i, cx);
-    if ( fsync[3] == '1' ) ca_copy_mask_overwrite_n(cx3, cx3->elements, i, cx);
-    if ( fsync[4] == '1' ) ca_copy_mask_overwrite_n(cx4, cx4->elements, i, cx);
-    if ( fsync[5] == '1' ) ca_copy_mask_overwrite_n(cx5, cx5->elements, i, cx);
-  }
+VALUE
+ca_call_cfunc_5_r (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4, void *userdata), const char *fsync,
+                              VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3, VALUE rcx4,
+                              void *userdata)
+{
+  CArray   *cx[5];
+  char     *base[5];
+  ca_size_t stride[5];
+  char     *owned_buf[5];
+  int       attached[5];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx[2]);
+  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx[3]);
+  TypedData_Get_Struct(rcx4, CArray, &carray_data_type, cx[4]);
+
+  /* sweep engine: same lifecycle as ca_call_cfunc_5; the difference is
+     the per-cell `func(...)` call has `userdata` as its last argument. */
+  state.n_ops     = 5;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_5_r";
+
+  ca_sweep_acquire(&state);
 
   {
-    char      *p0, *p1, *p2, *p3, *p4, *p5;
-    char      *q0, *q1, *q2, *q3, *q4, *q5;
-    ca_size_t    s0,  s1,  s2,  s3,  s4,  s5;
-    ca_size_t    k;
-
-    n = ca_set_iterator(6, cx0, &q0, &s0,
-                           cx1, &q1, &s1,
-                           cx2, &q2, &s2,
-                           cx3, &q3, &s3,
-                           cx4, &q4, &s4,
-                           cx5, &q5, &s5);
-
-    s0 *= cx0->bytes;
-    s1 *= cx1->bytes;
-    s2 *= cx2->bytes;
-    s3 *= cx3->bytes;
-    s4 *= cx4->bytes;
-    s5 *= cx5->bytes;
-
-    if ( m0 ) {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1,p2,p3,p4,p5)
-      #endif
-      for (k=0; k<n; k++) {
-        m = m0 + k;
-        if ( ! *m ) { 
-          p0 = q0 + k*s0; 
-          p1 = q1 + k*s1;
-          p2 = q2 + k*s2;
-          p3 = q3 + k*s3;
-          p4 = q4 + k*s4;
-          p5 = q5 + k*s5;
-          func(p0, p1, p2, p3, p4, p5);      
+    char *p[5];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 5; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1], p[2], p[3], p[4], userdata);
         }
       }
-    }
-    else {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1,p2,p3,p4,p5)
-      #endif
-      for (k=0; k<n; k++) {
-        p0 = q0 + k*s0; 
-        p1 = q1 + k*s1;
-        p2 = q2 + k*s2;
-        p3 = q3 + k*s3;
-        p4 = q4 + k*s4;
-        p5 = q5 + k*s5;
-        func(p0, p1, p2, p3, p4, p5);      
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 5; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1], p[2], p[3], p[4], userdata);
       }
     }
   }
-  if ( fsync[0] == '1' ) ca_sync(cx0);
-  if ( fsync[1] == '1' ) ca_sync(cx1);
-  if ( fsync[2] == '1' ) ca_sync(cx2);
-  if ( fsync[3] == '1' ) ca_sync(cx3);
-  if ( fsync[4] == '1' ) ca_sync(cx4);
-  if ( fsync[5] == '1' ) ca_sync(cx5);
-  ca_detach_n(6, cx0, cx1, cx2, cx3, cx4, cx5);
 
-  free(m0);
+  ca_sweep_release(&state);
 
   return rcx0;
 }
 
 VALUE
-ca_call_cfunc_7 (void (*func)(void*,void*,void*,void*,void*,void*,void*), const char *fsync,
-                                            VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3, VALUE rcx4, VALUE rcx5, VALUE rcx6)
+ca_call_cfunc_6_r (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4, void *p5, void *userdata), const char *fsync,
+                              VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3, VALUE rcx4, VALUE rcx5,
+                              void *userdata)
 {
-  CArray *cx0, *cx1, *cx2, *cx3, *cx4, *cx5, *cx6;
-  boolean8_t *m0 = NULL, *m;
-  ca_size_t n;
+  CArray   *cx[6];
+  char     *base[6];
+  ca_size_t stride[6];
+  char     *owned_buf[6];
+  int       attached[6];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx[2]);
+  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx[3]);
+  TypedData_Get_Struct(rcx4, CArray, &carray_data_type, cx[4]);
+  TypedData_Get_Struct(rcx5, CArray, &carray_data_type, cx[5]);
+
+  /* sweep engine: same lifecycle as ca_call_cfunc_6; the difference is
+     the per-cell `func(...)` call has `userdata` as its last argument. */
+  state.n_ops     = 6;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_6_r";
+
+  ca_sweep_acquire(&state);
 
-  if ( strlen(fsync) != 7 ) {
-    rb_raise(rb_eRuntimeError,
-             "[BUG] invalid length of fsync arg in rb_ca_call_mathfunc");
+  {
+    char *p[6];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 6; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1], p[2], p[3], p[4], p[5], userdata);
+        }
+      }
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 6; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1], p[2], p[3], p[4], p[5], userdata);
+      }
+    }
   }
 
-  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx0);
-  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx1);
-  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx2);
-  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx3);
-  TypedData_Get_Struct(rcx4, CArray, &carray_data_type, cx4);
-  TypedData_Get_Struct(rcx5, CArray, &carray_data_type, cx5);
-  TypedData_Get_Struct(rcx6, CArray, &carray_data_type, cx6);
+  ca_sweep_release(&state);
 
-  ca_attach_n(7, cx0, cx1, cx2, cx3, cx4, cx5, cx6);
+  return rcx0;
+}
 
-  {
-    CArray *cx[7];
-    int i = 0;
-    if ( fsync[0] == '0' ) cx[i++] = cx0;
-    if ( fsync[1] == '0' ) cx[i++] = cx1;
-    if ( fsync[2] == '0' ) cx[i++] = cx2;
-    if ( fsync[3] == '0' ) cx[i++] = cx3;
-    if ( fsync[4] == '0' ) cx[i++] = cx4;
-    if ( fsync[5] == '0' ) cx[i++] = cx5;
-    if ( fsync[6] == '0' ) cx[i++] = cx6;
-    m = m0 = ca_allocate_mask_iterator_n(i, cx);
-    if ( fsync[0] == '1' ) ca_copy_mask_overwrite_n(cx0, cx0->elements, i, cx);
-    if ( fsync[1] == '1' ) ca_copy_mask_overwrite_n(cx1, cx1->elements, i, cx);
-    if ( fsync[2] == '1' ) ca_copy_mask_overwrite_n(cx2, cx2->elements, i, cx);
-    if ( fsync[3] == '1' ) ca_copy_mask_overwrite_n(cx3, cx3->elements, i, cx);
-    if ( fsync[4] == '1' ) ca_copy_mask_overwrite_n(cx4, cx4->elements, i, cx);
-    if ( fsync[5] == '1' ) ca_copy_mask_overwrite_n(cx5, cx5->elements, i, cx);
-    if ( fsync[6] == '1' ) ca_copy_mask_overwrite_n(cx6, cx6->elements, i, cx);
-  }
+VALUE
+ca_call_cfunc_7_r (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4, void *p5, void *p6, void *userdata), const char *fsync,
+                              VALUE rcx0, VALUE rcx1, VALUE rcx2, VALUE rcx3, VALUE rcx4, VALUE rcx5, VALUE rcx6,
+                              void *userdata)
+{
+  CArray   *cx[7];
+  char     *base[7];
+  ca_size_t stride[7];
+  char     *owned_buf[7];
+  int       attached[7];
+  ca_sweep_state_t state;
+  int k_op;
+
+  TypedData_Get_Struct(rcx0, CArray, &carray_data_type, cx[0]);
+  TypedData_Get_Struct(rcx1, CArray, &carray_data_type, cx[1]);
+  TypedData_Get_Struct(rcx2, CArray, &carray_data_type, cx[2]);
+  TypedData_Get_Struct(rcx3, CArray, &carray_data_type, cx[3]);
+  TypedData_Get_Struct(rcx4, CArray, &carray_data_type, cx[4]);
+  TypedData_Get_Struct(rcx5, CArray, &carray_data_type, cx[5]);
+  TypedData_Get_Struct(rcx6, CArray, &carray_data_type, cx[6]);
+
+  /* sweep engine: same lifecycle as ca_call_cfunc_7; the difference is
+     the per-cell `func(...)` call has `userdata` as its last argument. */
+  state.n_ops     = 7;
+  state.fsync     = fsync;
+  state.cx        = cx;
+  state.base      = base;
+  state.stride    = stride;
+  state.owned_buf = owned_buf;
+  state.attached  = attached;
+  state.no_mask   = 0;
+  state.src_label = "ca_call_cfunc_7_r";
+
+  ca_sweep_acquire(&state);
 
   {
-    char      *p0, *p1, *p2, *p3, *p4, *p5, *p6;
-    char      *q0, *q1, *q2, *q3, *q4, *q5, *q6;
-    ca_size_t    s0,  s1,  s2,  s3,  s4,  s5,  s6;
-    ca_size_t    k;
-
-    n = ca_set_iterator(7, cx0, &q0, &s0,
-                           cx1, &q1, &s1,
-                           cx2, &q2, &s2,
-                           cx3, &q3, &s3,
-                           cx4, &q4, &s4,
-                           cx5, &q5, &s5,
-                           cx6, &q6, &s6);
-
-    s0 *= cx0->bytes;
-    s1 *= cx1->bytes;
-    s2 *= cx2->bytes;
-    s3 *= cx3->bytes;
-    s4 *= cx4->bytes;
-    s5 *= cx5->bytes;
-    s6 *= cx6->bytes;
-
-    if ( m0 ) {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1,p2,p3,p4,p5,p6)
-      #endif
-      for (k=0; k<n; k++) {
-        m = m0 + k;
-        if ( ! *m ) { 
-          p0 = q0 + k*s0; 
-          p1 = q1 + k*s1;
-          p2 = q2 + k*s2;
-          p3 = q3 + k*s3;
-          p4 = q4 + k*s4;
-          p5 = q5 + k*s5;
-          p6 = q6 + k*s6;
-          func(p0, p1, p2, p3, p4, p5, p6);      
+    char *p[7];
+    ca_size_t k;
+    if ( state.m0 ) {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        if ( ! state.m0[k] ) {
+          for ( k_op = 0; k_op < 7; k_op++ ) {
+            p[k_op] = base[k_op] + k * stride[k_op];
+          }
+          func(p[0], p[1], p[2], p[3], p[4], p[5], p[6], userdata);
         }
       }
-    }
-    else {
-      #ifdef _OPENMP
-      #pragma omp parallel for private(p0,p1,p2,p3,p4,p5,p6)
-      #endif
-      for (k=0; k<n; k++) {
-        p0 = q0 + k*s0; 
-        p1 = q1 + k*s1;
-        p2 = q2 + k*s2;
-        p3 = q3 + k*s3;
-        p4 = q4 + k*s4;
-        p5 = q5 + k*s5;
-        p6 = q6 + k*s6;
-        func(p0, p1, p2, p3, p4, p5, p6);      
+    } else {
+      for ( k = 0; k < state.n_kernel; k++ ) {
+        for ( k_op = 0; k_op < 7; k_op++ ) {
+          p[k_op] = base[k_op] + k * stride[k_op];
+        }
+        func(p[0], p[1], p[2], p[3], p[4], p[5], p[6], userdata);
       }
     }
-
   }
-  if ( fsync[0] == '1' ) ca_sync(cx0);
-  if ( fsync[1] == '1' ) ca_sync(cx1);
-  if ( fsync[2] == '1' ) ca_sync(cx2);
-  if ( fsync[3] == '1' ) ca_sync(cx3);
-  if ( fsync[4] == '1' ) ca_sync(cx4);
-  if ( fsync[5] == '1' ) ca_sync(cx5);
-  if ( fsync[6] == '1' ) ca_sync(cx6);
-  ca_detach_n(7, cx0, cx1, cx2, cx3, cx4, cx5, cx6);
 
-  free(m0);
+  ca_sweep_release(&state);
 
   return rcx0;
 }
 
-/* -------------------------------------------------------------------- */
-
 VALUE
-ca_call_cfunc_1_1 (int8_t dty, int8_t dtx, 
-                     void (*mathfunc)(void*,void*), VALUE rx) 
-{ 
-  volatile VALUE ry; 
-  rx = rb_ca_wrap_readonly(rx, INT2NUM(dtx)); 
-  if ( dty != dtx ) {
-    ry = rb_ca_template(rb_ca_wrap_readonly(rx, INT2NUM(dty))); 
-  }
-  else {
-    ry = rb_ca_template(rx);
-  }
-  ca_call_cfunc_2(mathfunc, "10", ry, rx); 
-  if ( rb_ca_is_scalar(ry) ) {
-    return rb_ca_fetch_addr(ry, 0);
+ca_call_cfunc_1_1_r (int8_t dty, int8_t dtx1, void (*mathfunc)(void*, void*, void *userdata), volatile VALUE rx1, void *userdata)
+{
+  volatile VALUE ry = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+
+  if ( dty != dtx1 ) {
+    ry = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(1, rx1);
   }
-  else {
-    return ry;
+
+  ca_call_cfunc_2_r(mathfunc, "10", ry, rx1, userdata);
+
+  if ( rb_ca_is_scalar(ry) ) {
+    ry = rb_ca_fetch_addr(ry, 0);
   }
+  return ry;
 }
 
 VALUE
-ca_call_cfunc_1_2 (int8_t dty, 
-                     int8_t dtx1, 
-                     int8_t dtx2, 
-                     void (*mathfunc)(void*,void*,void*), 
-                     volatile VALUE rx1, 
-                     volatile VALUE rx2) 
-{ 
-  volatile VALUE ry = Qnil; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
-  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2)); 
+ca_call_cfunc_1_2_r (int8_t dty, int8_t dtx1, int8_t dtx2, void (*mathfunc)(void*, void*, void*, void *userdata), volatile VALUE rx1, volatile VALUE rx2, void *userdata)
+{
+  volatile VALUE ry = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+
   if ( dty != dtx1 || dty != dtx2 ) {
-    ry  = rb_ca_template_n(2,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty))); 
-  }
-  else {
-    ry  = rb_ca_template_n(2, rx1, rx2);
+    ry = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty)), rb_ca_wrap_readonly(rx2, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(2, rx1, rx2);
   }
-  ca_call_cfunc_3(mathfunc, "100", ry, rx1, rx2); 
+
+  ca_call_cfunc_3_r(mathfunc, "100", ry, rx1, rx2, userdata);
+
   if ( rb_ca_is_scalar(ry) ) {
-    return rb_ca_fetch_addr(ry, 0);
-  }
-  else {
-    return ry;
+    ry = rb_ca_fetch_addr(ry, 0);
   }
+  return ry;
 }
 
-
 VALUE
-ca_call_cfunc_1_3 (int8_t dty, 
-                     int8_t dtx1, 
-                     int8_t dtx2, 
-                     int8_t dtx3, 
-                     void (*mathfunc)(void*,void*,void*,void*), 
-                     volatile VALUE rx1, 
-                     volatile VALUE rx2, 
-                     volatile VALUE rx3) 
-{ 
-  volatile VALUE ry; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
-  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2)); 
-  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3)); 
-  ry  = rb_ca_template_n(3,
-                         rb_ca_wrap_readonly(rx1, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx2, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx3, INT2NUM(dty))); 
-  ca_call_cfunc_4(mathfunc, "1000", ry, rx1, rx2, rx3); 
-  if ( rb_ca_is_scalar(ry) ) {
-    return rb_ca_fetch_addr(ry, 0);
+ca_call_cfunc_1_3_r (int8_t dty, int8_t dtx1, int8_t dtx2, int8_t dtx3, void (*mathfunc)(void*, void*, void*, void*, void *userdata), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3, void *userdata)
+{
+  volatile VALUE ry = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+
+  if ( dty != dtx1 || dty != dtx2 || dty != dtx3 ) {
+    ry = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty)), rb_ca_wrap_readonly(rx2, INT2NUM(dty)), rb_ca_wrap_readonly(rx3, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(3, rx1, rx2, rx3);
   }
-  else {
-    return ry;
+
+  ca_call_cfunc_4_r(mathfunc, "1000", ry, rx1, rx2, rx3, userdata);
+
+  if ( rb_ca_is_scalar(ry) ) {
+    ry = rb_ca_fetch_addr(ry, 0);
   }
+  return ry;
 }
 
 VALUE
-ca_call_cfunc_1_4 (int8_t dty, 
-                     int8_t dtx1, 
-                     int8_t dtx2, 
-                     int8_t dtx3, 
-                     int8_t dtx4, 
-                     void (*mathfunc)(void*,void*,void*,void*,void*), 
-                     volatile VALUE rx1, 
-                     volatile VALUE rx2, 
-                     volatile VALUE rx3,
-                     volatile VALUE rx4) 
-{ 
-  volatile VALUE ry; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
-  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2)); 
-  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3)); 
-  rx4 = rb_ca_wrap_readonly(rx4, INT2NUM(dtx4)); 
-  ry  = rb_ca_template_n(4,
-                         rb_ca_wrap_readonly(rx1, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx2, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx3, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx4, INT2NUM(dty))); 
-  ca_call_cfunc_5(mathfunc, "10000", ry, rx1, rx2, rx3, rx4); 
-  if ( rb_ca_is_scalar(ry) ) {
-    return rb_ca_fetch_addr(ry, 0);
+ca_call_cfunc_1_4_r (int8_t dty, int8_t dtx1, int8_t dtx2, int8_t dtx3, int8_t dtx4, void (*mathfunc)(void*, void*, void*, void*, void*, void *userdata), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3, volatile VALUE rx4, void *userdata)
+{
+  volatile VALUE ry = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+  rx4 = rb_ca_wrap_readonly(rx4, INT2NUM(dtx4));
+
+  if ( dty != dtx1 || dty != dtx2 || dty != dtx3 || dty != dtx4 ) {
+    ry = rb_ca_template_n(4, rb_ca_wrap_readonly(rx1, INT2NUM(dty)), rb_ca_wrap_readonly(rx2, INT2NUM(dty)), rb_ca_wrap_readonly(rx3, INT2NUM(dty)), rb_ca_wrap_readonly(rx4, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(4, rx1, rx2, rx3, rx4);
   }
-  else {
-    return ry;
+
+  ca_call_cfunc_5_r(mathfunc, "10000", ry, rx1, rx2, rx3, rx4, userdata);
+
+  if ( rb_ca_is_scalar(ry) ) {
+    ry = rb_ca_fetch_addr(ry, 0);
   }
+  return ry;
 }
 
 VALUE
-ca_call_cfunc_1_5 (int8_t dty, 
-                     int8_t dtx1, 
-                     int8_t dtx2, 
-                     int8_t dtx3, 
-                     int8_t dtx4, 
-                     int8_t dtx5, 
-                     void (*mathfunc)(void*,void*,void*,void*,void*,void*), 
-                     volatile VALUE rx1, 
-                     volatile VALUE rx2, 
-                     volatile VALUE rx3,
-                     volatile VALUE rx4,
-                     volatile VALUE rx5) 
-{ 
-  volatile VALUE ry; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
-  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2)); 
-  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3)); 
-  rx4 = rb_ca_wrap_readonly(rx4, INT2NUM(dtx4)); 
-  rx5 = rb_ca_wrap_readonly(rx5, INT2NUM(dtx5)); 
-  ry  = rb_ca_template_n(5,
-                         rb_ca_wrap_readonly(rx1, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx2, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx3, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx4, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx5, INT2NUM(dty))); 
-  ca_call_cfunc_6(mathfunc, "10000", ry, rx1, rx2, rx3, rx4, rx5); 
-  if ( rb_ca_is_scalar(ry) ) {
-    return rb_ca_fetch_addr(ry, 0);
+ca_call_cfunc_1_5_r (int8_t dty, int8_t dtx1, int8_t dtx2, int8_t dtx3, int8_t dtx4, int8_t dtx5, void (*mathfunc)(void*, void*, void*, void*, void*, void*, void *userdata), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3, volatile VALUE rx4, volatile VALUE rx5, void *userdata)
+{
+  volatile VALUE ry = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+  rx4 = rb_ca_wrap_readonly(rx4, INT2NUM(dtx4));
+  rx5 = rb_ca_wrap_readonly(rx5, INT2NUM(dtx5));
+
+  if ( dty != dtx1 || dty != dtx2 || dty != dtx3 || dty != dtx4 || dty != dtx5 ) {
+    ry = rb_ca_template_n(5, rb_ca_wrap_readonly(rx1, INT2NUM(dty)), rb_ca_wrap_readonly(rx2, INT2NUM(dty)), rb_ca_wrap_readonly(rx3, INT2NUM(dty)), rb_ca_wrap_readonly(rx4, INT2NUM(dty)), rb_ca_wrap_readonly(rx5, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(5, rx1, rx2, rx3, rx4, rx5);
   }
-  else {
-    return ry;
+
+  ca_call_cfunc_6_r(mathfunc, "100000", ry, rx1, rx2, rx3, rx4, rx5, userdata);
+
+  if ( rb_ca_is_scalar(ry) ) {
+    ry = rb_ca_fetch_addr(ry, 0);
   }
+  return ry;
 }
 
 VALUE
-ca_call_cfunc_1_6 (int8_t dty, 
-                     int8_t dtx1, 
-                     int8_t dtx2, 
-                     int8_t dtx3, 
-                     int8_t dtx4, 
-                     int8_t dtx5, 
-                     int8_t dtx6, 
-                     void (*mathfunc)(void*,void*,void*,void*,void*,void*,void*), 
-                     volatile VALUE rx1, 
-                     volatile VALUE rx2, 
-                     volatile VALUE rx3,
-                     volatile VALUE rx4,
-                     volatile VALUE rx5,
-                     volatile VALUE rx6) 
-{ 
-  volatile VALUE ry; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
-  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2)); 
-  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3)); 
-  rx4 = rb_ca_wrap_readonly(rx4, INT2NUM(dtx4)); 
-  rx5 = rb_ca_wrap_readonly(rx5, INT2NUM(dtx5)); 
-  rx6 = rb_ca_wrap_readonly(rx5, INT2NUM(dtx6)); 
-  ry  = rb_ca_template_n(5,
-                         rb_ca_wrap_readonly(rx1, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx2, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx3, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx4, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx5, INT2NUM(dty)), 
-                         rb_ca_wrap_readonly(rx6, INT2NUM(dty))); 
-  ca_call_cfunc_7(mathfunc, "100000", ry, rx1, rx2, rx3, rx4, rx5, rx6); 
-  if ( rb_ca_is_scalar(ry) ) {
-    return rb_ca_fetch_addr(ry, 0);
+ca_call_cfunc_1_6_r (int8_t dty, int8_t dtx1, int8_t dtx2, int8_t dtx3, int8_t dtx4, int8_t dtx5, int8_t dtx6, void (*mathfunc)(void*, void*, void*, void*, void*, void*, void*, void *userdata), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3, volatile VALUE rx4, volatile VALUE rx5, volatile VALUE rx6, void *userdata)
+{
+  volatile VALUE ry = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+  rx4 = rb_ca_wrap_readonly(rx4, INT2NUM(dtx4));
+  rx5 = rb_ca_wrap_readonly(rx5, INT2NUM(dtx5));
+  rx6 = rb_ca_wrap_readonly(rx6, INT2NUM(dtx6));
+
+  if ( dty != dtx1 || dty != dtx2 || dty != dtx3 || dty != dtx4 || dty != dtx5 || dty != dtx6 ) {
+    ry = rb_ca_template_n(6, rb_ca_wrap_readonly(rx1, INT2NUM(dty)), rb_ca_wrap_readonly(rx2, INT2NUM(dty)), rb_ca_wrap_readonly(rx3, INT2NUM(dty)), rb_ca_wrap_readonly(rx4, INT2NUM(dty)), rb_ca_wrap_readonly(rx5, INT2NUM(dty)), rb_ca_wrap_readonly(rx6, INT2NUM(dty)));
+  } else {
+    ry = rb_ca_template_n(6, rx1, rx2, rx3, rx4, rx5, rx6);
   }
-  else {
-    return ry;
+
+  ca_call_cfunc_7_r(mathfunc, "1000000", ry, rx1, rx2, rx3, rx4, rx5, rx6, userdata);
+
+  if ( rb_ca_is_scalar(ry) ) {
+    ry = rb_ca_fetch_addr(ry, 0);
   }
+  return ry;
 }
 
 VALUE
-ca_call_cfunc_2_1 (int8_t dty1, 
-                    int8_t dty2, 
-                    int8_t dtx1, 
-                    void (*mathfunc)(void*,void*,void*), 
-                    volatile VALUE rx1) 
-{ 
-  volatile VALUE ry1 = Qnil, ry2 = Qnil; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
+ca_call_cfunc_2_1_r (int8_t dty1, int8_t dty2, int8_t dtx1, void (*mathfunc)(void*, void*, void*, void *userdata), volatile VALUE rx1, void *userdata)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+
   if ( dty1 != dtx1 ) {
-    ry1  = rb_ca_template_n(1,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty1))); 
-  }
-  else {
-    ry1  = rb_ca_template_n(1, rx1);
+    ry1 = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(1, rx1);
   }
   if ( dty2 != dtx1 ) {
-    ry2  = rb_ca_template_n(1,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty2))); 
+    ry2 = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(1, rx1);
   }
-  else {
-    ry2  = rb_ca_template_n(1, rx1);
-  }
-  ca_call_cfunc_3(mathfunc, "110", ry1, ry2, rx1); 
+
+  ca_call_cfunc_3_r(mathfunc, "110", ry1, ry2, rx1, userdata);
+
   if ( rb_ca_is_scalar(ry1) ) {
     ry1 = rb_ca_fetch_addr(ry1, 0);
   }
@@ -823,36 +1428,27 @@ ca_call_cfunc_2_1 (int8_t dty1,
   return rb_ary_new3(2, ry1, ry2);
 }
 
-
 VALUE
-ca_call_cfunc_2_2 (int8_t dty1, 
-                      int8_t dty2, 
-                      int8_t dtx1, 
-                      int8_t dtx2, 
-                      void (*mathfunc)(void*,void*,void*,void*), 
-                      volatile VALUE rx1, 
-                      volatile VALUE rx2) 
-{ 
-  volatile VALUE ry1 = Qnil, ry2 = Qnil; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
-  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2)); 
+ca_call_cfunc_2_2_r (int8_t dty1, int8_t dty2, int8_t dtx1, int8_t dtx2, void (*mathfunc)(void*, void*, void*, void*, void *userdata), volatile VALUE rx1, volatile VALUE rx2, void *userdata)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+
   if ( dty1 != dtx1 || dty1 != dtx2 ) {
-    ry1  = rb_ca_template_n(2,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty1))); 
-  }
-  else {
-    ry1  = rb_ca_template_n(2, rx1, rx2);
+    ry1 = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), rb_ca_wrap_readonly(rx2, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(2, rx1, rx2);
   }
   if ( dty2 != dtx1 || dty2 != dtx2 ) {
-    ry2  = rb_ca_template_n(2,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty2))); 
-  }
-  else {
-    ry2  = rb_ca_template_n(2, rx1, rx2);
+    ry2 = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), rb_ca_wrap_readonly(rx2, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(2, rx1, rx2);
   }
-  ca_call_cfunc_4(mathfunc, "1100", ry1, ry2, rx1, rx2); 
+
+  ca_call_cfunc_4_r(mathfunc, "1100", ry1, ry2, rx1, rx2, userdata);
+
   if ( rb_ca_is_scalar(ry1) ) {
     ry1 = rb_ca_fetch_addr(ry1, 0);
   }
@@ -863,39 +1459,27 @@ ca_call_cfunc_2_2 (int8_t dty1,
 }
 
 VALUE
-ca_call_cfunc_2_3 (int8_t dty1, 
-                      int8_t dty2, 
-                      int8_t dtx1, 
-                      int8_t dtx2, 
-                      int8_t dtx3, 
-                      void (*mathfunc)(void*,void*,void*,void*,void*), 
-                      volatile VALUE rx1, 
-                      volatile VALUE rx2, 
-                      volatile VALUE rx3) 
-{ 
-  volatile VALUE ry1 = Qnil, ry2 = Qnil; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
-  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2)); 
-  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3)); 
+ca_call_cfunc_2_3_r (int8_t dty1, int8_t dty2, int8_t dtx1, int8_t dtx2, int8_t dtx3, void (*mathfunc)(void*, void*, void*, void*, void*, void *userdata), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3, void *userdata)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+
   if ( dty1 != dtx1 || dty1 != dtx2 || dty1 != dtx3 ) {
-    ry1  = rb_ca_template_n(3,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty1)), 
-                           rb_ca_wrap_readonly(rx3, INT2NUM(dty1))); 
-  }
-  else {
-    ry1  = rb_ca_template_n(3, rx1, rx2, rx3);
+    ry1 = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), rb_ca_wrap_readonly(rx2, INT2NUM(dty1)), rb_ca_wrap_readonly(rx3, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(3, rx1, rx2, rx3);
   }
   if ( dty2 != dtx1 || dty2 != dtx2 || dty2 != dtx3 ) {
-    ry2  = rb_ca_template_n(3,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty2)), 
-                           rb_ca_wrap_readonly(rx3, INT2NUM(dty2))); 
-  }
-  else {
-    ry2  = rb_ca_template_n(3, rx1, rx2, rx3);
+    ry2 = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), rb_ca_wrap_readonly(rx2, INT2NUM(dty2)), rb_ca_wrap_readonly(rx3, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(3, rx1, rx2, rx3);
   }
-  ca_call_cfunc_5(mathfunc, "11000", ry1, ry2, rx1, rx2, rx3); 
+
+  ca_call_cfunc_5_r(mathfunc, "11000", ry1, ry2, rx1, rx2, rx3, userdata);
+
   if ( rb_ca_is_scalar(ry1) ) {
     ry1 = rb_ca_fetch_addr(ry1, 0);
   }
@@ -906,44 +1490,28 @@ ca_call_cfunc_2_3 (int8_t dty1,
 }
 
 VALUE
-ca_call_cfunc_2_4 (int8_t dty1, 
-                      int8_t dty2, 
-                      int8_t dtx1, 
-                      int8_t dtx2, 
-                      int8_t dtx3, 
-                      int8_t dtx4, 
-                      void (*mathfunc)(void*,void*,void*,void*,void*,void*), 
-                      volatile VALUE rx1, 
-                      volatile VALUE rx2, 
-                      volatile VALUE rx3,
-                      volatile VALUE rx4) 
-{ 
-  volatile VALUE ry1 = Qnil, ry2 = Qnil; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
-  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2)); 
-  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3)); 
-  rx4 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx4)); 
+ca_call_cfunc_2_4_r (int8_t dty1, int8_t dty2, int8_t dtx1, int8_t dtx2, int8_t dtx3, int8_t dtx4, void (*mathfunc)(void*, void*, void*, void*, void*, void*, void *userdata), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3, volatile VALUE rx4, void *userdata)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+  rx4 = rb_ca_wrap_readonly(rx4, INT2NUM(dtx4));
+
   if ( dty1 != dtx1 || dty1 != dtx2 || dty1 != dtx3 || dty1 != dtx4 ) {
-    ry1  = rb_ca_template_n(4,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty1)), 
-                           rb_ca_wrap_readonly(rx3, INT2NUM(dty1)),
-                           rb_ca_wrap_readonly(rx4, INT2NUM(dty1))); 
-  }
-  else {
-    ry1  = rb_ca_template_n(4, rx1, rx2, rx3, rx4);
+    ry1 = rb_ca_template_n(4, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), rb_ca_wrap_readonly(rx2, INT2NUM(dty1)), rb_ca_wrap_readonly(rx3, INT2NUM(dty1)), rb_ca_wrap_readonly(rx4, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(4, rx1, rx2, rx3, rx4);
   }
   if ( dty2 != dtx1 || dty2 != dtx2 || dty2 != dtx3 || dty2 != dtx4 ) {
-    ry2  = rb_ca_template_n(4,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty2)), 
-                           rb_ca_wrap_readonly(rx3, INT2NUM(dty2)),
-                           rb_ca_wrap_readonly(rx4, INT2NUM(dty2))); 
-  }
-  else {
-    ry2  = rb_ca_template_n(4, rx1, rx2, rx3, rx4);
+    ry2 = rb_ca_template_n(4, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), rb_ca_wrap_readonly(rx2, INT2NUM(dty2)), rb_ca_wrap_readonly(rx3, INT2NUM(dty2)), rb_ca_wrap_readonly(rx4, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(4, rx1, rx2, rx3, rx4);
   }
-  ca_call_cfunc_6(mathfunc, "110000", ry1, ry2, rx1, rx2, rx3, rx4); 
+
+  ca_call_cfunc_6_r(mathfunc, "110000", ry1, ry2, rx1, rx2, rx3, rx4, userdata);
+
   if ( rb_ca_is_scalar(ry1) ) {
     ry1 = rb_ca_fetch_addr(ry1, 0);
   }
@@ -954,37 +1522,30 @@ ca_call_cfunc_2_4 (int8_t dty1,
 }
 
 VALUE
-ca_call_cfunc_3_1 (int8_t dty1, 
-                      int8_t dty2, 
-                      int8_t dty3, 
-                      int8_t dtx1, 
-                      void (*mathfunc)(void*,void*,void*,void*), 
-                      volatile VALUE rx1) 
-{ 
-  volatile VALUE ry1 = Qnil, ry2 = Qnil, ry3 = Qnil; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
+ca_call_cfunc_3_1_r (int8_t dty1, int8_t dty2, int8_t dty3, int8_t dtx1, void (*mathfunc)(void*, void*, void*, void*, void *userdata), volatile VALUE rx1, void *userdata)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil, ry3 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+
   if ( dty1 != dtx1 ) {
-    ry1  = rb_ca_template_n(1,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty1))); 
-  }
-  else {
-    ry1  = rb_ca_template_n(1, rx1);
+    ry1 = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(1, rx1);
   }
   if ( dty2 != dtx1 ) {
-    ry2  = rb_ca_template_n(1,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty2))); 
-  }
-  else {
-    ry2  = rb_ca_template_n(1, rx1);
+    ry2 = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(1, rx1);
   }
-  if ( dty3 != dtx1  ) {
-    ry3  = rb_ca_template_n(1,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty3))); 
+  if ( dty3 != dtx1 ) {
+    ry3 = rb_ca_template_n(1, rb_ca_wrap_readonly(rx1, INT2NUM(dty3)));
+  } else {
+    ry3 = rb_ca_template_n(1, rx1);
   }
-  else {
-    ry3  = rb_ca_template_n(1, rx1);
-  }
-  ca_call_cfunc_4(mathfunc, "1110", ry1, ry2, ry3, rx1); 
+
+  ca_call_cfunc_4_r(mathfunc, "1110", ry1, ry2, ry3, rx1, userdata);
+
   if ( rb_ca_is_scalar(ry1) ) {
     ry1 = rb_ca_fetch_addr(ry1, 0);
   }
@@ -998,43 +1559,31 @@ ca_call_cfunc_3_1 (int8_t dty1,
 }
 
 VALUE
-ca_call_cfunc_3_2 (int8_t dty1, 
-                      int8_t dty2, 
-                      int8_t dty3, 
-                      int8_t dtx1, 
-                      int8_t dtx2, 
-                      void (*mathfunc)(void*,void*,void*,void*,void*), 
-                      volatile VALUE rx1, 
-                      volatile VALUE rx2) 
-{ 
-  volatile VALUE ry1 = Qnil, ry2 = Qnil, ry3 = Qnil; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
-  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2)); 
+ca_call_cfunc_3_2_r (int8_t dty1, int8_t dty2, int8_t dty3, int8_t dtx1, int8_t dtx2, void (*mathfunc)(void*, void*, void*, void*, void*, void *userdata), volatile VALUE rx1, volatile VALUE rx2, void *userdata)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil, ry3 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+
   if ( dty1 != dtx1 || dty1 != dtx2 ) {
-    ry1  = rb_ca_template_n(2,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty1))); 
-  }
-  else {
-    ry1  = rb_ca_template_n(2, rx1, rx2);
+    ry1 = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), rb_ca_wrap_readonly(rx2, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(2, rx1, rx2);
   }
   if ( dty2 != dtx1 || dty2 != dtx2 ) {
-    ry2  = rb_ca_template_n(2,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty2))); 
-  }
-  else {
-    ry2  = rb_ca_template_n(2, rx1, rx2);
+    ry2 = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), rb_ca_wrap_readonly(rx2, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(2, rx1, rx2);
   }
   if ( dty3 != dtx1 || dty3 != dtx2 ) {
-    ry3  = rb_ca_template_n(2,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty3)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty3))); 
+    ry3 = rb_ca_template_n(2, rb_ca_wrap_readonly(rx1, INT2NUM(dty3)), rb_ca_wrap_readonly(rx2, INT2NUM(dty3)));
+  } else {
+    ry3 = rb_ca_template_n(2, rx1, rx2);
   }
-  else {
-    ry3  = rb_ca_template_n(2, rx1, rx2);
-  }
-  ca_call_cfunc_5(mathfunc, "11100", ry1, ry2, ry3, rx1, rx2); 
+
+  ca_call_cfunc_5_r(mathfunc, "11100", ry1, ry2, ry3, rx1, rx2, userdata);
+
   if ( rb_ca_is_scalar(ry1) ) {
     ry1 = rb_ca_fetch_addr(ry1, 0);
   }
@@ -1048,50 +1597,32 @@ ca_call_cfunc_3_2 (int8_t dty1,
 }
 
 VALUE
-ca_call_cfunc_3_3 (int8_t dty1, 
-                      int8_t dty2, 
-                      int8_t dty3, 
-                      int8_t dtx1, 
-                      int8_t dtx2, 
-                      int8_t dtx3, 
-                      void (*mathfunc)(void*,void*,void*,void*,void*,void*), 
-                      volatile VALUE rx1, 
-                      volatile VALUE rx2,                       
-                      volatile VALUE rx3) 
-{ 
-  volatile VALUE ry1 = Qnil, ry2 = Qnil, ry3 = Qnil; 
-  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1)); 
-  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2)); 
-  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3)); 
+ca_call_cfunc_3_3_r (int8_t dty1, int8_t dty2, int8_t dty3, int8_t dtx1, int8_t dtx2, int8_t dtx3, void (*mathfunc)(void*, void*, void*, void*, void*, void*, void *userdata), volatile VALUE rx1, volatile VALUE rx2, volatile VALUE rx3, void *userdata)
+{
+  volatile VALUE ry1 = Qnil, ry2 = Qnil, ry3 = Qnil;
+
+  rx1 = rb_ca_wrap_readonly(rx1, INT2NUM(dtx1));
+  rx2 = rb_ca_wrap_readonly(rx2, INT2NUM(dtx2));
+  rx3 = rb_ca_wrap_readonly(rx3, INT2NUM(dtx3));
+
   if ( dty1 != dtx1 || dty1 != dtx2 || dty1 != dtx3 ) {
-    ry1  = rb_ca_template_n(3,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty1)), 
-                           rb_ca_wrap_readonly(rx3, INT2NUM(dty1))); 
-  }
-  else {
-    ry1  = rb_ca_template_n(3, rx1, rx2, rx3);
+    ry1 = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty1)), rb_ca_wrap_readonly(rx2, INT2NUM(dty1)), rb_ca_wrap_readonly(rx3, INT2NUM(dty1)));
+  } else {
+    ry1 = rb_ca_template_n(3, rx1, rx2, rx3);
   }
   if ( dty2 != dtx1 || dty2 != dtx2 || dty2 != dtx3 ) {
-    ry2  = rb_ca_template_n(3,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty2)), 
-                           rb_ca_wrap_readonly(rx3, INT2NUM(dty2))); 
-  }
-  else {
-    ry2  = rb_ca_template_n(3, rx1, rx2, rx3);
+    ry2 = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty2)), rb_ca_wrap_readonly(rx2, INT2NUM(dty2)), rb_ca_wrap_readonly(rx3, INT2NUM(dty2)));
+  } else {
+    ry2 = rb_ca_template_n(3, rx1, rx2, rx3);
   }
   if ( dty3 != dtx1 || dty3 != dtx2 || dty3 != dtx3 ) {
-    ry3  = rb_ca_template_n(3,
-                           rb_ca_wrap_readonly(rx1, INT2NUM(dty3)), 
-                           rb_ca_wrap_readonly(rx2, INT2NUM(dty3)), 
-                           rb_ca_wrap_readonly(rx3, INT2NUM(dty3))); 
-  }
-  else {
-    ry3  = rb_ca_template_n(3, rx1, rx2, rx3);
+    ry3 = rb_ca_template_n(3, rb_ca_wrap_readonly(rx1, INT2NUM(dty3)), rb_ca_wrap_readonly(rx2, INT2NUM(dty3)), rb_ca_wrap_readonly(rx3, INT2NUM(dty3)));
+  } else {
+    ry3 = rb_ca_template_n(3, rx1, rx2, rx3);
   }
 
-  ca_call_cfunc_6(mathfunc, "111000", ry1, ry2, ry3, rx1, rx2, rx3); 
+  ca_call_cfunc_6_r(mathfunc, "111000", ry1, ry2, ry3, rx1, rx2, rx3, userdata);
+
   if ( rb_ca_is_scalar(ry1) ) {
     ry1 = rb_ca_fetch_addr(ry1, 0);
   }
@@ -1103,3 +1634,4 @@ ca_call_cfunc_3_3 (int8_t dty1,
   }
   return rb_ary_new3(3, ry1, ry2, ry3);
 }
+