carray 3.0.3 → 3.0.4

data/ext/carray_attribute.c

@@ -437,6 +437,7 @@ rb_ca_is_object_type (VALUE self)
 /* ------------------------------------------------------------------- */
 
 static ID id_parent;
+static ID id_lifted_from;
 static ID id_attr;
 static ID id_attrs;
 
@@ -446,6 +447,47 @@ rb_ca_parent (VALUE self)
   return rb_ivar_get(self, id_parent);
 }
 
+/* A view of a Face is built on the Face's storage and wrapped in a new Face
+   (ca_face_lift), so the Face it was taken from is not on its parent chain.
+   The lift records that Face here, and the attribute walk goes to it in
+   place of the parent: the attributes set on a Face show on its views. */
+void
+rb_ca_set_lifted_from (VALUE lifted, VALUE face)
+{
+  rb_ivar_set(lifted, id_lifted_from, face);
+}
+
+/* The next array of the attribute walk: the Face a lifted view was taken
+   from, or else the parent. */
+static VALUE
+ca_attr_parent (VALUE obj)
+{
+  if ( rb_ivar_defined(obj, id_lifted_from) ) {
+    return rb_ivar_get(obj, id_lifted_from);
+  }
+  return rb_ivar_get(obj, id_parent);
+}
+
+static VALUE
+rb_ca_attribute_parent (VALUE self)
+{
+  return ca_attr_parent(self);
+}
+
+/* dup and clone copy the instance variables before initialize_copy, so the
+   copy would share the receiver's attribute Hash and a set_attr on it would
+   write into the receiver.  Give the copy a Hash of its own. */
+void
+rb_ca_own_attr (VALUE self)
+{
+  if ( rb_ivar_defined(self, id_attr) ) {
+    VALUE h = rb_ivar_get(self, id_attr);
+    if ( RB_TYPE_P(h, T_HASH) ) {
+      rb_ivar_set(self, id_attr, rb_hash_dup(h));
+    }
+  }
+}
+
 /* Give dst the attributes (lib/carray/attribute.rb) that src shows, as its
    own Hash.  Used by the methods that produce a new entity holding the
    same values -- copy, to_type.  The parent walk here only decides
@@ -458,13 +500,14 @@ rb_ca_inherit_attr (VALUE dst, VALUE src)
   VALUE obj = src, attrs;
   int depth;
 
+  /* A chain longer than this is left to #attrs, which walks it all. */
   for (depth = 0; depth < 64 && ! NIL_P(obj); depth++) {
     if ( rb_ivar_defined(obj, id_attr) ) {
       break;
     }
-    obj = rb_ivar_get(obj, id_parent);
+    obj = ca_attr_parent(obj);
   }
-  if ( depth == 64 || NIL_P(obj) ) {
+  if ( NIL_P(obj) ) {
     return;
   }
   attrs = rb_funcall(src, id_attrs, 0);
@@ -553,12 +596,50 @@ ca_root_array (void *ap)
 }
 
 /* Whether two arrays may share memory: their cells live in the same
-   root.  An operation that writes one while reading the other must read
-   from a copy taken before it writes. */
+   root, or in two entities whose memory overlaps (one wraps the other's
+   memory, as wrap_memory_view does).  A view with several parents (a
+   stack, a meld, a lazy expression) is the end of its own chain, so its
+   parents are asked in turn.  An operation that writes one while reading
+   the other must read from a copy taken before it writes. */
 int
 ca_shares_root (void *ap, void *aq)
 {
-  return ca_root_array(ap) == ca_root_array(aq);
+  CArray *ra = ca_root_array(ap);
+  CArray *rb = ca_root_array(aq);
+  if ( ra == rb ) {
+    return 1;
+  }
+  if ( ca_test_flag(ra, CA_FLAG_MULTI_PARENTS) ) {
+    CAMultiParent *mp = (CAMultiParent *) ra;
+    int32_t k;
+    for ( k = 0; k < mp->n_parents; k++ ) {
+      if ( mp->parents[k] && ca_shares_root(mp->parents[k], rb) ) {
+        return 1;
+      }
+    }
+    return 0;
+  }
+  if ( ca_test_flag(rb, CA_FLAG_MULTI_PARENTS) ) {
+    return ca_shares_root(rb, ra);
+  }
+  if ( ca_is_entity(ra) && ca_is_entity(rb) && ra->ptr && rb->ptr ) {
+    char *a0 = ra->ptr, *a1 = ra->ptr + ca_length(ra);
+    char *b0 = rb->ptr, *b1 = rb->ptr + ca_length(rb);
+    return a0 < b1 && b0 < a1;
+  }
+  return 0;
+}
+
+/* Whether self and other may share memory (ca_shares_root), for the plan
+   reader that keeps an evaluator off a store reading what it writes. */
+static VALUE
+rb_ca_storage_overlaps (VALUE self, VALUE other)
+{
+  CArray *ca, *cb;
+  rb_check_carray_object(other);
+  TypedData_Get_Struct(self,  CArray, &carray_data_type, ca);
+  TypedData_Get_Struct(other, CArray, &carray_data_type, cb);
+  return ca_shares_root(ca, cb) ? Qtrue : Qfalse;
 }
 
 static VALUE
@@ -624,6 +705,7 @@ void
 Init_carray_attribute (void)
 {
   id_parent     = rb_intern("parent");
+  id_lifted_from = rb_intern("lifted_from");
   id_attr       = rb_intern("@attr");
   id_attrs      = rb_intern("attrs");
 
@@ -632,7 +714,6 @@ Init_carray_attribute (void)
   rb_define_method(rb_cCArray, "bytes", rb_ca_bytes, 0);
   rb_define_method(rb_cCArray, "flags", rb_ca_flags, 0);
   rb_define_method(rb_cCArray, "ndim", rb_ca_ndim, 0);
-  rb_define_method(rb_cCArray, "rank", rb_ca_ndim, 0);
   rb_define_method(rb_cCArray, "shape", rb_ca_dim, 0);
   rb_define_method(rb_cCArray, "dim", rb_ca_dim, 0);
   rb_define_method(rb_cCArray, "dim0", rb_ca_dim0, 0);
@@ -646,6 +727,7 @@ Init_carray_attribute (void)
   rb_define_method(rb_cCArray, "data_type_name", rb_ca_data_type_name, 0);
 
   rb_define_method(rb_cCArray, "parent", rb_ca_parent, 0);
+  rb_define_private_method(rb_cCArray, "attribute_parent", rb_ca_attribute_parent, 0);
 
   rb_define_method(rb_cCArray, "data_class", rb_ca_data_class, 0);
   rb_define_method(rb_cCArray, "data_class=", rb_ca_set_data_class, 1);
@@ -673,6 +755,7 @@ Init_carray_attribute (void)
   rb_define_method(rb_cCArray, "object?",   rb_ca_is_object_type, 0);
 
   rb_define_method(rb_cCArray, "root_array", rb_ca_root_array, 0);
+  rb_define_private_method(rb_cCArray, "storage_overlaps?", rb_ca_storage_overlaps, 1);
   rb_define_method(rb_cCArray, "ancestors", rb_ca_ancestors, 0);
 }
 

data/ext/carray_bincount.c

@@ -67,8 +67,8 @@
   case CA_UINT64:  COUNT_KERNEL(uint64_t, OUT_T); break; \
   default: \
     rb_raise(rb_eCADataTypeError, \
-             "bincount: integer label array required (got %d)", \
-             cl->data_type); \
+             "bincount: integer label array required (got %s)", \
+             ca_type_name[cl->data_type]); \
   } \
 } while (0)
 
@@ -83,8 +83,8 @@ bincount_check_types (CArray *cl, CArray *cw)
     break;
   default:
     rb_raise(rb_eCADataTypeError,
-             "bincount: integer label array required (got %d)",
-             cl->data_type);
+             "bincount: integer label array required (got %s)",
+             ca_type_name[cl->data_type]);
   }
   if ( ! cw ) {
     return;
@@ -96,7 +96,8 @@ bincount_check_types (CArray *cl, CArray *cw)
     break;
   default:
     rb_raise(rb_eCADataTypeError,
-             "bincount: weights must be numeric (got %d)", cw->data_type);
+             "bincount: weights must be numeric (got %s)",
+             ca_type_name[cw->data_type]);
   }
 }
 
@@ -199,8 +200,8 @@ rb_ca_bincount_count_kernel (VALUE self, VALUE rlength)
   case CA_UINT64:  WEIGHTED_KERNEL(uint64_t, W_T); break; \
   default: \
     rb_raise(rb_eCADataTypeError, \
-             "bincount: integer label array required (got %d)", \
-             cl->data_type); \
+             "bincount: integer label array required (got %s)", \
+             ca_type_name[cl->data_type]); \
   } \
 } while (0)
 
@@ -262,7 +263,8 @@ rb_ca_bincount_weighted_kernel (VALUE self, VALUE rweights, VALUE rlength)
   default:
     ca_detach_n(2, cl, cw);
     rb_raise(rb_eCADataTypeError,
-             "bincount: weights must be numeric (got %d)", w_type);
+             "bincount: weights must be numeric (got %s)",
+             ca_type_name[w_type]);
   }
 
   ca_detach_n(2, cl, cw);

data/ext/carray_broadcast.c

@@ -30,6 +30,7 @@
 ---------------------------------------------------------------------------- */
 
 #include "carray.h"
+#include "carray_internal.h"   /* ca_check_rank_fits */
 #include "ca_obj_face.h"  /* CA_WRAPPER_LIFT */
 
 NORETURN(static void ca_broadcast_refuse_write (VALUE dst, VALUE src));
@@ -41,10 +42,11 @@ ca_broadcast_refuse_write (VALUE dst, VALUE src)
   volatile VALUE dst_s = rb_inspect(rb_funcall(dst, rb_intern("shape"), 0));
   volatile VALUE src_s = rb_inspect(rb_funcall(src, rb_intern("shape"), 0));
   rb_raise(rb_eArgError,
-           "shape mismatch writing to carray (%s <- %s); shapes must agree "
+           "%s: shape mismatch writing to carray (%s <- %s); shapes must agree "
            "once size-1 axes are dropped, or one side must be 1-D, or the "
            "source must be smaller and broadcastable -- use .flatten to "
            "write the values in the order they lie",
+           ca_calling_method_name(),
            StringValueCStr(dst_s), StringValueCStr(src_s));
 }
 
@@ -87,7 +89,11 @@ ca_broadcast_to_destination (VALUE dst, volatile VALUE *src)
     return;
   }
 
-  if ( cs->elements < cd->elements && cs->ndim == cd->ndim ) {
+  /* A size-1 axis repeated onto an axis of length 0 leaves the source
+     larger than an empty destination; the axes say it is a repeat all
+     the same. */
+  if ( ( cs->elements < cd->elements || cd->elements == 0 )
+       && cs->ndim == cd->ndim ) {
     int i;
     for ( i = 0; i < cd->ndim; i++ ) {
       if ( cs->dim[i] != cd->dim[i] && cs->dim[i] != 1 ) break;
@@ -196,10 +202,11 @@ ca_broadcast_refuse (VALUE self, VALUE other)
 
   if (ca->ndim != cb->ndim) {
     rb_raise(rb_eArgError,
-             "shape mismatch between operands (%s and %s): the operands "
+             "%s: shape mismatch between operands (%s and %s): the operands "
              "have different numbers of axes, and axes are not added "
              "implicitly -- add size-1 axes with :_ so both have the same "
              "number (x[:_, nil] makes a 1-D x a row, x[nil, :_] a column)%s",
+             ca_calling_method_name(),
              StringValueCStr(self_s), StringValueCStr(other_s),
              flatten_hint);
   }
@@ -210,10 +217,10 @@ ca_broadcast_refuse (VALUE self, VALUE other)
     }
   }
   rb_raise(rb_eArgError,
-           "shape mismatch between operands (%s and %s): axis %d has "
+           "%s: shape mismatch between operands (%s and %s): axis %d has "
            "sizes %lld and %lld; each axis must have the same size, or "
            "size 1 on one side%s",
-           StringValueCStr(self_s), StringValueCStr(other_s), i,
+           ca_calling_method_name(), StringValueCStr(self_s), StringValueCStr(other_s), i,
            (long long) ca->dim[i], (long long) cb->dim[i], flatten_hint);
 }
 
@@ -295,9 +302,7 @@ rb_ca_broadcast_to (int argc, VALUE *argv, VALUE self)
 
   TypedData_Get_Struct(self, CArray, &carray_data_type, cs);
 
-  if (argc < 0 || argc > CA_RANK_MAX) {
-    rb_raise(rb_eArgError, "broadcast_to: invalid number of dims (%d)", argc);
-  }
+  ca_check_rank_fits(argc, "broadcast_to");
   target_ndim = (int8_t) argc;
 
   if (target_ndim < cs->ndim) {

data/ext/carray_build_flags.h

@@ -0,0 +1,3 @@
+/* GENERATED by extconf.rb: the flags this build's arithmetic
+   was compiled with.  See CArray::BUILD_FLAGS. */
+#define CA_BUILD_FLAGS "-O3 -fno-fast-math -fwrapv -fopenmp-simd -march=native"

data/ext/carray_call_cfunc.c

@@ -226,6 +226,8 @@ ca_call_cfunc_2 (void (*func)(void *p0, void *p1), const char *fsync,
   ctx.userdata = NULL;
   ca_sweep_run(&state, ca_call_cfunc_2_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+
   return rcx0;
 }
 
@@ -298,6 +300,9 @@ ca_call_cfunc_3 (void (*func)(void *p0, void *p1, void *p2), const char *fsync,
   ctx.userdata = NULL;
   ca_sweep_run(&state, ca_call_cfunc_3_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+
   return rcx0;
 }
 
@@ -371,6 +376,10 @@ ca_call_cfunc_4 (void (*func)(void *p0, void *p1, void *p2, void *p3), const cha
   ctx.userdata = NULL;
   ca_sweep_run(&state, ca_call_cfunc_4_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+
   return rcx0;
 }
 
@@ -445,6 +454,11 @@ ca_call_cfunc_5 (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4),
   ctx.userdata = NULL;
   ca_sweep_run(&state, ca_call_cfunc_5_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+
   return rcx0;
 }
 
@@ -520,6 +534,12 @@ ca_call_cfunc_6 (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4,
   ctx.userdata = NULL;
   ca_sweep_run(&state, ca_call_cfunc_6_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+  RB_GC_GUARD(rcx5);
+
   return rcx0;
 }
 
@@ -596,6 +616,13 @@ ca_call_cfunc_7 (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4,
   ctx.userdata = NULL;
   ca_sweep_run(&state, ca_call_cfunc_7_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+  RB_GC_GUARD(rcx5);
+  RB_GC_GUARD(rcx6);
+
   return rcx0;
 }
 
@@ -1128,6 +1155,8 @@ ca_call_cfunc_2_r (void (*func)(void *p0, void *p1, void *userdata), const char
   ctx.userdata = userdata;
   ca_sweep_run(&state, ca_call_cfunc_2_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+
   return rcx0;
 }
 
@@ -1201,6 +1230,9 @@ ca_call_cfunc_3_r (void (*func)(void *p0, void *p1, void *p2, void *userdata), c
   ctx.userdata = userdata;
   ca_sweep_run(&state, ca_call_cfunc_3_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+
   return rcx0;
 }
 
@@ -1275,6 +1307,10 @@ ca_call_cfunc_4_r (void (*func)(void *p0, void *p1, void *p2, void *p3, void *us
   ctx.userdata = userdata;
   ca_sweep_run(&state, ca_call_cfunc_4_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+
   return rcx0;
 }
 
@@ -1350,6 +1386,11 @@ ca_call_cfunc_5_r (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4
   ctx.userdata = userdata;
   ca_sweep_run(&state, ca_call_cfunc_5_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+
   return rcx0;
 }
 
@@ -1426,6 +1467,12 @@ ca_call_cfunc_6_r (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4
   ctx.userdata = userdata;
   ca_sweep_run(&state, ca_call_cfunc_6_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+  RB_GC_GUARD(rcx5);
+
   return rcx0;
 }
 
@@ -1503,6 +1550,13 @@ ca_call_cfunc_7_r (void (*func)(void *p0, void *p1, void *p2, void *p3, void *p4
   ctx.userdata = userdata;
   ca_sweep_run(&state, ca_call_cfunc_7_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+  RB_GC_GUARD(rcx5);
+  RB_GC_GUARD(rcx6);
+
   return rcx0;
 }
 
@@ -1973,6 +2027,8 @@ ca_call_cslab_2 (ca_cslab_t func, const char *fsync,
   ctx.userdata = NULL;
   ca_sweep_run_chunked(&state, ca_cslab_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+
   return rcx0;
 }
 
@@ -2017,6 +2073,9 @@ ca_call_cslab_3 (ca_cslab_t func, const char *fsync,
   ctx.userdata = NULL;
   ca_sweep_run_chunked(&state, ca_cslab_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+
   return rcx0;
 }
 
@@ -2062,6 +2121,10 @@ ca_call_cslab_4 (ca_cslab_t func, const char *fsync,
   ctx.userdata = NULL;
   ca_sweep_run_chunked(&state, ca_cslab_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+
   return rcx0;
 }
 
@@ -2108,6 +2171,11 @@ ca_call_cslab_5 (ca_cslab_t func, const char *fsync,
   ctx.userdata = NULL;
   ca_sweep_run_chunked(&state, ca_cslab_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+
   return rcx0;
 }
 
@@ -2155,6 +2223,12 @@ ca_call_cslab_6 (ca_cslab_t func, const char *fsync,
   ctx.userdata = NULL;
   ca_sweep_run_chunked(&state, ca_cslab_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+  RB_GC_GUARD(rcx5);
+
   return rcx0;
 }
 
@@ -2203,6 +2277,13 @@ ca_call_cslab_7 (ca_cslab_t func, const char *fsync,
   ctx.userdata = NULL;
   ca_sweep_run_chunked(&state, ca_cslab_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+  RB_GC_GUARD(rcx5);
+  RB_GC_GUARD(rcx6);
+
   return rcx0;
 }
 
@@ -2290,6 +2371,8 @@ ca_call_cslab_2_r (ca_cslab_r_t func, const char *fsync,
   ctx.userdata = userdata;
   ca_sweep_run_chunked(&state, ca_cslab_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+
   return rcx0;
 }
 
@@ -2335,6 +2418,9 @@ ca_call_cslab_3_r (ca_cslab_r_t func, const char *fsync,
   ctx.userdata = userdata;
   ca_sweep_run_chunked(&state, ca_cslab_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+
   return rcx0;
 }
 
@@ -2381,6 +2467,10 @@ ca_call_cslab_4_r (ca_cslab_r_t func, const char *fsync,
   ctx.userdata = userdata;
   ca_sweep_run_chunked(&state, ca_cslab_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+
   return rcx0;
 }
 
@@ -2428,6 +2518,11 @@ ca_call_cslab_5_r (ca_cslab_r_t func, const char *fsync,
   ctx.userdata = userdata;
   ca_sweep_run_chunked(&state, ca_cslab_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+
   return rcx0;
 }
 
@@ -2476,6 +2571,12 @@ ca_call_cslab_6_r (ca_cslab_r_t func, const char *fsync,
   ctx.userdata = userdata;
   ca_sweep_run_chunked(&state, ca_cslab_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+  RB_GC_GUARD(rcx5);
+
   return rcx0;
 }
 
@@ -2525,6 +2626,13 @@ ca_call_cslab_7_r (ca_cslab_r_t func, const char *fsync,
   ctx.userdata = userdata;
   ca_sweep_run_chunked(&state, ca_cslab_r_walk, (VALUE) &ctx);
 
+  RB_GC_GUARD(rcx1);
+  RB_GC_GUARD(rcx2);
+  RB_GC_GUARD(rcx3);
+  RB_GC_GUARD(rcx4);
+  RB_GC_GUARD(rcx5);
+  RB_GC_GUARD(rcx6);
+
   return rcx0;
 }