carray 3.0.2 → 3.0.4

data/ext/carray_kernels_reduce_boolean.c

@@ -24,6 +24,13 @@
    probed by operation-table function pointer (obj_type is runtime-
    assigned, so no CA_OBJ_STACK compile-time constant exists). */
 extern ca_operation_function_t ca_stack_func;
+/* The square root of a variance: an ε-negative from rounding becomes 0,
+   a NaN stays NaN (fmax(NaN, 0.0) is 0.0). */
+#ifndef CARRAY_SQRT_CLAMP_DEFINED
+#define CARRAY_SQRT_CLAMP_DEFINED
+static inline double ca_sqrt_clamp (double v) { return sqrt(v < 0.0 ? 0.0 : v); }
+#endif
+
 /* BOOL2VAL: bool -> Ruby (Qtrue/Qfalse).  Used as ruby_scalar wrapper
    for the all/any flat-reduction Ruby surface so `a.all` / `a.any`
    return real true/false, not Integer 0/1.  (BOOL2OBJ already exists
@@ -38,13 +45,16 @@ static inline VALUE BOOL2VAL (boolean8_t x) { return x ? Qtrue : Qfalse; }
 
 /* ===== all_ki ============================================ */
 
+typedef struct {
+  void      *chunk;
+  CArray    *ca;
+} all_ki_stream_bool_t;
+
 static VALUE
-all_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis)
+all_ki_stream_bool (VALUE arg)
 {
-    /* P.4.5.3b/c streaming chunked reduce: full reduction over lazy unmasked.  */
-    /* P.4.5.3c (N-D): outer-axis chunking, 1-D is the natural sub-case.        */
-    if ( naxes == ca->ndim && ca->ndim >= 1 && !keep_axis &&
-         ca_is_lazy_view(ca) && ! ca_has_mask(ca) ) {
+      all_ki_stream_bool_t *__ctx = (all_ki_stream_bool_t *) arg;
+      CArray *ca = __ctx->ca;
       ca_size_t __inner = 1;
       int       __k;
       for ( __k = 1; __k < ca->ndim; __k++ ) __inner *= ca->dim[__k];
@@ -65,8 +75,8 @@ all_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
       ca_size_t __outer_off = 0;
       boolean8_t acc;
       acc = (1);
-      ca_lazy_arena_enter();
       boolean8_t *__chunk = (boolean8_t *) ca_lazy_arena_acquire(__chunk_elems * sizeof(boolean8_t));
+      __ctx->chunk = __chunk;
       while ( __outer_off < __outer ) {
         ca_size_t __r = (__outer - __outer_off < __rows) ? (__outer - __outer_off) : __rows;
         ca_size_t __n = __r * __inner;
@@ -80,9 +90,21 @@ all_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
         }
         __outer_off += __r;
       }
-      ca_lazy_arena_release(__chunk);
-      ca_lazy_arena_exit();
       return BOOL2VAL((boolean8_t) (acc));
+}
+
+static VALUE
+all_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis)
+{
+    if ( naxes == ca->ndim && ca->ndim >= 1 && !keep_axis &&
+         ca->elements > 0 &&
+         ca_is_lazy_view(ca) && ! ca_has_mask(ca) &&
+         ca_fast_paths_enabled() ) {
+      all_ki_stream_bool_t __ctx;
+      __ctx.chunk = NULL;
+      __ctx.ca    = ca;
+      return ca_lazy_arena_protect(all_ki_stream_bool, ca_lazy_arena_release_held,
+                                   (VALUE) &__ctx);
     }
 VALUE   vout  = rb_ca_new_reduced(self, slab_axes, naxes, CA_BOOLEAN, keep_axis);
 CArray *co;
@@ -93,7 +115,7 @@ int rc = ca_iter_state_init_l2(&st, ca, CA_SLAB_AXES,
                                slab_axes, naxes, 0);
 if ( rc != CA_ITER_OK ) {
   rb_raise(rb_eRuntimeError,
-           "all_ki: kernel_iterator init failed rc=%d", rc);
+           "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
 }
 char       *p;
 boolean8_t *m;
@@ -119,26 +141,37 @@ rb_ca_all_ki (int argc, VALUE *argv, VALUE self)
   int keep_axis = RTEST(rkeep_axis);
   int8_t slab_axes[CA_RANK_MAX];
   int8_t naxes = rb_ca_parse_reduce_axes_kw(raxis, src, slab_axes);
+  if ( ca_is_lazy_view(src) && ! ( naxes == src->ndim && src->ndim >= 1 && ! keep_axis && ! ca_has_mask(src) ) ) {
+    volatile VALUE evaluated = ca_lazy_evaluated(self);
+    if ( ! NIL_P(evaluated) ) {
+      self = evaluated;
+      GetCArray(self, src);
+    }
+  }
   if ( argc > 0 ) {
-    rb_raise(rb_eArgError, "all_ki: positional axis arguments are no longer accepted (got %d); use axis: kwarg, e.g. a.all(axis: 0) or a.all(axis: [0, 1])", argc);
+    rb_raise(rb_eArgError, "%s: positional axis arguments are no longer accepted (got %d); use the axis: keyword, e.g. a.%s(axis: 0) or a.%s(axis: [0, 1])", ca_calling_method_name(), argc, ca_calling_method_name(), ca_calling_method_name());
   }
   switch ( src->data_type ) {
-    case CA_BOOLEAN: return all_ki_native_bool(self, src, slab_axes, naxes, keep_axis);
+    case CA_BOOLEAN: CA_RETURN_KEEPING(self, all_ki_native_bool(self, src, slab_axes, naxes, keep_axis));
     default:
-      rb_raise(rb_eCADataTypeError, "all_ki: source data_type :%s not supported (expected one of: bool)", ca_type_name[src->data_type]);
+      ca_face_refuse_numeric(self, ca_calling_method_name(), rb_eCADataTypeError);
+      rb_raise(rb_eCADataTypeError, "%s: source data_type :%s not supported (expected one of: boolean)", ca_calling_method_name(), ca_type_name[src->data_type]);
   }
   return Qnil;  /* unreachable */
 }
 
 /* ===== any_ki ============================================ */
 
+typedef struct {
+  void      *chunk;
+  CArray    *ca;
+} any_ki_stream_bool_t;
+
 static VALUE
-any_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis)
+any_ki_stream_bool (VALUE arg)
 {
-    /* P.4.5.3b/c streaming chunked reduce: full reduction over lazy unmasked.  */
-    /* P.4.5.3c (N-D): outer-axis chunking, 1-D is the natural sub-case.        */
-    if ( naxes == ca->ndim && ca->ndim >= 1 && !keep_axis &&
-         ca_is_lazy_view(ca) && ! ca_has_mask(ca) ) {
+      any_ki_stream_bool_t *__ctx = (any_ki_stream_bool_t *) arg;
+      CArray *ca = __ctx->ca;
       ca_size_t __inner = 1;
       int       __k;
       for ( __k = 1; __k < ca->ndim; __k++ ) __inner *= ca->dim[__k];
@@ -159,8 +192,8 @@ any_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
       ca_size_t __outer_off = 0;
       boolean8_t acc;
       acc = (0);
-      ca_lazy_arena_enter();
       boolean8_t *__chunk = (boolean8_t *) ca_lazy_arena_acquire(__chunk_elems * sizeof(boolean8_t));
+      __ctx->chunk = __chunk;
       while ( __outer_off < __outer ) {
         ca_size_t __r = (__outer - __outer_off < __rows) ? (__outer - __outer_off) : __rows;
         ca_size_t __n = __r * __inner;
@@ -174,9 +207,21 @@ any_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
         }
         __outer_off += __r;
       }
-      ca_lazy_arena_release(__chunk);
-      ca_lazy_arena_exit();
       return BOOL2VAL((boolean8_t) (acc));
+}
+
+static VALUE
+any_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis)
+{
+    if ( naxes == ca->ndim && ca->ndim >= 1 && !keep_axis &&
+         ca->elements > 0 &&
+         ca_is_lazy_view(ca) && ! ca_has_mask(ca) &&
+         ca_fast_paths_enabled() ) {
+      any_ki_stream_bool_t __ctx;
+      __ctx.chunk = NULL;
+      __ctx.ca    = ca;
+      return ca_lazy_arena_protect(any_ki_stream_bool, ca_lazy_arena_release_held,
+                                   (VALUE) &__ctx);
     }
 VALUE   vout  = rb_ca_new_reduced(self, slab_axes, naxes, CA_BOOLEAN, keep_axis);
 CArray *co;
@@ -187,7 +232,7 @@ int rc = ca_iter_state_init_l2(&st, ca, CA_SLAB_AXES,
                                slab_axes, naxes, 0);
 if ( rc != CA_ITER_OK ) {
   rb_raise(rb_eRuntimeError,
-           "any_ki: kernel_iterator init failed rc=%d", rc);
+           "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
 }
 char       *p;
 boolean8_t *m;
@@ -213,26 +258,37 @@ rb_ca_any_ki (int argc, VALUE *argv, VALUE self)
   int keep_axis = RTEST(rkeep_axis);
   int8_t slab_axes[CA_RANK_MAX];
   int8_t naxes = rb_ca_parse_reduce_axes_kw(raxis, src, slab_axes);
+  if ( ca_is_lazy_view(src) && ! ( naxes == src->ndim && src->ndim >= 1 && ! keep_axis && ! ca_has_mask(src) ) ) {
+    volatile VALUE evaluated = ca_lazy_evaluated(self);
+    if ( ! NIL_P(evaluated) ) {
+      self = evaluated;
+      GetCArray(self, src);
+    }
+  }
   if ( argc > 0 ) {
-    rb_raise(rb_eArgError, "any_ki: positional axis arguments are no longer accepted (got %d); use axis: kwarg, e.g. a.any(axis: 0) or a.any(axis: [0, 1])", argc);
+    rb_raise(rb_eArgError, "%s: positional axis arguments are no longer accepted (got %d); use the axis: keyword, e.g. a.%s(axis: 0) or a.%s(axis: [0, 1])", ca_calling_method_name(), argc, ca_calling_method_name(), ca_calling_method_name());
   }
   switch ( src->data_type ) {
-    case CA_BOOLEAN: return any_ki_native_bool(self, src, slab_axes, naxes, keep_axis);
+    case CA_BOOLEAN: CA_RETURN_KEEPING(self, any_ki_native_bool(self, src, slab_axes, naxes, keep_axis));
     default:
-      rb_raise(rb_eCADataTypeError, "any_ki: source data_type :%s not supported (expected one of: bool)", ca_type_name[src->data_type]);
+      ca_face_refuse_numeric(self, ca_calling_method_name(), rb_eCADataTypeError);
+      rb_raise(rb_eCADataTypeError, "%s: source data_type :%s not supported (expected one of: boolean)", ca_calling_method_name(), ca_type_name[src->data_type]);
   }
   return Qnil;  /* unreachable */
 }
 
 /* ===== none_ki ============================================ */
 
+typedef struct {
+  void      *chunk;
+  CArray    *ca;
+} none_ki_stream_bool_t;
+
 static VALUE
-none_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis)
+none_ki_stream_bool (VALUE arg)
 {
-    /* P.4.5.3b/c streaming chunked reduce: full reduction over lazy unmasked.  */
-    /* P.4.5.3c (N-D): outer-axis chunking, 1-D is the natural sub-case.        */
-    if ( naxes == ca->ndim && ca->ndim >= 1 && !keep_axis &&
-         ca_is_lazy_view(ca) && ! ca_has_mask(ca) ) {
+      none_ki_stream_bool_t *__ctx = (none_ki_stream_bool_t *) arg;
+      CArray *ca = __ctx->ca;
       ca_size_t __inner = 1;
       int       __k;
       for ( __k = 1; __k < ca->ndim; __k++ ) __inner *= ca->dim[__k];
@@ -253,8 +309,8 @@ none_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, in
       ca_size_t __outer_off = 0;
       boolean8_t acc;
       acc = (1);
-      ca_lazy_arena_enter();
       boolean8_t *__chunk = (boolean8_t *) ca_lazy_arena_acquire(__chunk_elems * sizeof(boolean8_t));
+      __ctx->chunk = __chunk;
       while ( __outer_off < __outer ) {
         ca_size_t __r = (__outer - __outer_off < __rows) ? (__outer - __outer_off) : __rows;
         ca_size_t __n = __r * __inner;
@@ -268,9 +324,21 @@ none_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, in
         }
         __outer_off += __r;
       }
-      ca_lazy_arena_release(__chunk);
-      ca_lazy_arena_exit();
       return BOOL2VAL((boolean8_t) (acc));
+}
+
+static VALUE
+none_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis)
+{
+    if ( naxes == ca->ndim && ca->ndim >= 1 && !keep_axis &&
+         ca->elements > 0 &&
+         ca_is_lazy_view(ca) && ! ca_has_mask(ca) &&
+         ca_fast_paths_enabled() ) {
+      none_ki_stream_bool_t __ctx;
+      __ctx.chunk = NULL;
+      __ctx.ca    = ca;
+      return ca_lazy_arena_protect(none_ki_stream_bool, ca_lazy_arena_release_held,
+                                   (VALUE) &__ctx);
     }
 VALUE   vout  = rb_ca_new_reduced(self, slab_axes, naxes, CA_BOOLEAN, keep_axis);
 CArray *co;
@@ -281,7 +349,7 @@ int rc = ca_iter_state_init_l2(&st, ca, CA_SLAB_AXES,
                                slab_axes, naxes, 0);
 if ( rc != CA_ITER_OK ) {
   rb_raise(rb_eRuntimeError,
-           "none_ki: kernel_iterator init failed rc=%d", rc);
+           "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
 }
 char       *p;
 boolean8_t *m;
@@ -307,13 +375,21 @@ rb_ca_none_ki (int argc, VALUE *argv, VALUE self)
   int keep_axis = RTEST(rkeep_axis);
   int8_t slab_axes[CA_RANK_MAX];
   int8_t naxes = rb_ca_parse_reduce_axes_kw(raxis, src, slab_axes);
+  if ( ca_is_lazy_view(src) && ! ( naxes == src->ndim && src->ndim >= 1 && ! keep_axis && ! ca_has_mask(src) ) ) {
+    volatile VALUE evaluated = ca_lazy_evaluated(self);
+    if ( ! NIL_P(evaluated) ) {
+      self = evaluated;
+      GetCArray(self, src);
+    }
+  }
   if ( argc > 0 ) {
-    rb_raise(rb_eArgError, "none_ki: positional axis arguments are no longer accepted (got %d); use axis: kwarg, e.g. a.none(axis: 0) or a.none(axis: [0, 1])", argc);
+    rb_raise(rb_eArgError, "%s: positional axis arguments are no longer accepted (got %d); use the axis: keyword, e.g. a.%s(axis: 0) or a.%s(axis: [0, 1])", ca_calling_method_name(), argc, ca_calling_method_name(), ca_calling_method_name());
   }
   switch ( src->data_type ) {
-    case CA_BOOLEAN: return none_ki_native_bool(self, src, slab_axes, naxes, keep_axis);
+    case CA_BOOLEAN: CA_RETURN_KEEPING(self, none_ki_native_bool(self, src, slab_axes, naxes, keep_axis));
     default:
-      rb_raise(rb_eCADataTypeError, "none_ki: source data_type :%s not supported (expected one of: bool)", ca_type_name[src->data_type]);
+      ca_face_refuse_numeric(self, ca_calling_method_name(), rb_eCADataTypeError);
+      rb_raise(rb_eCADataTypeError, "%s: source data_type :%s not supported (expected one of: boolean)", ca_calling_method_name(), ca_type_name[src->data_type]);
   }
   return Qnil;  /* unreachable */
 }