carray 3.0.2 → 3.0.4

data/ext/carray_kernels_triop.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
@@ -330,7 +337,7 @@ ca_triop_fma_float32_t (ca_size_t n, boolean8_t *m,
         p2 = q2 + k*i2;
         p3 = q3 + k*i3;
         p4 = q4 + k*i4;
-        { (*p4) = fma((*p1), (*p2), (*p3)); }
+        { (*p4) = fmaf((*p1), (*p2), (*p3)); }
       }
     }
   } else {
@@ -339,7 +346,7 @@ ca_triop_fma_float32_t (ca_size_t n, boolean8_t *m,
       p2 = q2 + k*i2;
       p3 = q3 + k*i3;
       p4 = q4 + k*i4;
-      { (*p4) = fma((*p1), (*p2), (*p3)); }
+      { (*p4) = fmaf((*p1), (*p2), (*p3)); }
     }
   }
 }
@@ -801,7 +808,7 @@ ca_triop_fms_float32_t (ca_size_t n, boolean8_t *m,
         p2 = q2 + k*i2;
         p3 = q3 + k*i3;
         p4 = q4 + k*i4;
-        { (*p4) = fma((*p1), (*p2), -(*p3)); }
+        { (*p4) = fmaf((*p1), (*p2), -(*p3)); }
       }
     }
   } else {
@@ -810,7 +817,7 @@ ca_triop_fms_float32_t (ca_size_t n, boolean8_t *m,
       p2 = q2 + k*i2;
       p3 = q3 + k*i3;
       p4 = q4 + k*i4;
-      { (*p4) = fma((*p1), (*p2), -(*p3)); }
+      { (*p4) = fmaf((*p1), (*p2), -(*p3)); }
     }
   }
 }
@@ -1378,8 +1385,473 @@ ca_triop_clip[CA_NTYPE] = {
 VALUE rb_ca_clip (VALUE self, VALUE other2, VALUE other3)
 { return rb_ca_call_triop(self, other2, other3, ca_triop_clip); }
 
-VALUE rb_ca_clip_bang (VALUE self, VALUE other2, VALUE other3)
-{ return rb_ca_call_triop_bang(self, other2, other3, ca_triop_clip); }
+/* ===== triop select ============================================ */
+
+static void
+ca_triop_select_int8_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  int8_t *q1 = (int8_t *) ptr1, *q2 = (int8_t *) ptr2,
+            *q3 = (int8_t *) ptr3, *q4 = (int8_t *) ptr4;
+  int8_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_uint8_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  uint8_t *q1 = (uint8_t *) ptr1, *q2 = (uint8_t *) ptr2,
+            *q3 = (uint8_t *) ptr3, *q4 = (uint8_t *) ptr4;
+  uint8_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_int16_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  int16_t *q1 = (int16_t *) ptr1, *q2 = (int16_t *) ptr2,
+            *q3 = (int16_t *) ptr3, *q4 = (int16_t *) ptr4;
+  int16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_uint16_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  uint16_t *q1 = (uint16_t *) ptr1, *q2 = (uint16_t *) ptr2,
+            *q3 = (uint16_t *) ptr3, *q4 = (uint16_t *) ptr4;
+  uint16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_int32_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  int32_t *q1 = (int32_t *) ptr1, *q2 = (int32_t *) ptr2,
+            *q3 = (int32_t *) ptr3, *q4 = (int32_t *) ptr4;
+  int32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_uint32_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  uint32_t *q1 = (uint32_t *) ptr1, *q2 = (uint32_t *) ptr2,
+            *q3 = (uint32_t *) ptr3, *q4 = (uint32_t *) ptr4;
+  uint32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_int64_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  int64_t *q1 = (int64_t *) ptr1, *q2 = (int64_t *) ptr2,
+            *q3 = (int64_t *) ptr3, *q4 = (int64_t *) ptr4;
+  int64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_uint64_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  uint64_t *q1 = (uint64_t *) ptr1, *q2 = (uint64_t *) ptr2,
+            *q3 = (uint64_t *) ptr3, *q4 = (uint64_t *) ptr4;
+  uint64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_float32_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  float *q1 = (float *) ptr1, *q2 = (float *) ptr2,
+            *q3 = (float *) ptr3, *q4 = (float *) ptr4;
+  float *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_float64_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  double *q1 = (double *) ptr1, *q2 = (double *) ptr2,
+            *q3 = (double *) ptr3, *q4 = (double *) ptr4;
+  double *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_cmplx64_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  cmplx64_t *q1 = (cmplx64_t *) ptr1, *q2 = (cmplx64_t *) ptr2,
+            *q3 = (cmplx64_t *) ptr3, *q4 = (cmplx64_t *) ptr4;
+  cmplx64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_cmplx128_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  cmplx128_t *q1 = (cmplx128_t *) ptr1, *q2 = (cmplx128_t *) ptr2,
+            *q3 = (cmplx128_t *) ptr3, *q4 = (cmplx128_t *) ptr4;
+  cmplx128_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+static void
+ca_triop_select_boolean8_t (ca_size_t n, boolean8_t *m,
+                                 char *ptr1, ca_size_t i1,
+                                 char *ptr2, ca_size_t i2,
+                                 char *ptr3, ca_size_t i3,
+                                 char *ptr4, ca_size_t i4)
+{
+  boolean8_t *q1 = (boolean8_t *) ptr1, *q2 = (boolean8_t *) ptr2,
+            *q3 = (boolean8_t *) ptr3, *q4 = (boolean8_t *) ptr4;
+  boolean8_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_size_t k;
+  if ( m ) {
+    boolean8_t *pm;
+    for (k=0; k<n; k++) {
+      pm = m + k;
+      if ( ! *pm ) {
+        p1 = q1 + k*i1;
+        p2 = q2 + k*i2;
+        p3 = q3 + k*i3;
+        p4 = q4 + k*i4;
+        { (*p4) = (*p1) ? (*p2) : (*p3); }
+      }
+    }
+  } else {
+    for (k=0; k<n; k++) {
+      p1 = q1 + k*i1;
+      p2 = q2 + k*i2;
+      p3 = q3 + k*i3;
+      p4 = q4 + k*i4;
+      { (*p4) = (*p1) ? (*p2) : (*p3); }
+    }
+  }
+}
+
+ca_triop_func_t
+ca_triop_select[CA_NTYPE] = {
+  ca_triop_not_implement,
+  ca_triop_select_boolean8_t,
+  ca_triop_select_int8_t,
+  ca_triop_select_uint8_t,
+  ca_triop_select_int16_t,
+  ca_triop_select_uint16_t,
+  ca_triop_select_int32_t,
+  ca_triop_select_uint32_t,
+  ca_triop_select_int64_t,
+  ca_triop_select_uint64_t,
+  ca_triop_select_float32_t,
+  ca_triop_select_float64_t,
+  ca_triop_not_implement,
+  ca_triop_select_cmplx64_t,
+  ca_triop_select_cmplx128_t,
+  ca_triop_not_implement,
+  ca_triop_not_implement,
+};
+
+VALUE rb_ca_select (VALUE self, VALUE other2, VALUE other3)
+{ return rb_ca_call_triop(self, other2, other3, ca_triop_select); }
 
 void
 Init_carray_kernels_triop (void)
@@ -1388,5 +1860,5 @@ Init_carray_kernels_triop (void)
   rb_define_method(rb_cCArray, "fma!", rb_ca_fma_bang, 2);
   rb_define_method(rb_cCArray, "fms", rb_ca_fms, 2);
   rb_define_method(rb_cCArray, "fms!", rb_ca_fms_bang, 2);
-  rb_define_method(rb_cCArray, "__clip_ki__", rb_ca_clip, 2);
+  rb_define_method(rb_cCArray, "clamp_between", rb_ca_clip, 2);
 }