numo-narray-alt 0.9.10 → 0.9.12

data/ext/numo/narray/src/t_dfloat.c

@@ -42,10 +42,59 @@ static ID id_to_a;
 VALUE cT;
 extern VALUE cRT;
 
+#include "mh/coerce_cast.h"
+#include "mh/to_a.h"
+#include "mh/fill.h"
+#include "mh/format.h"
+#include "mh/format_to_a.h"
+#include "mh/inspect.h"
+#include "mh/op/add.h"
+#include "mh/op/sub.h"
+#include "mh/op/mul.h"
+#include "mh/op/div.h"
+#include "mh/op/mod.h"
+#include "mh/divmod.h"
+#include "mh/round/floor.h"
+#include "mh/round/round.h"
+#include "mh/round/ceil.h"
+#include "mh/round/trunc.h"
+#include "mh/round/rint.h"
+#include "mh/comp/eq.h"
+#include "mh/comp/ne.h"
+#include "mh/comp/nearly_eq.h"
+#include "mh/comp/gt.h"
+#include "mh/comp/ge.h"
+#include "mh/comp/lt.h"
+#include "mh/comp/le.h"
+#include "mh/clip.h"
+#include "mh/isnan.h"
+#include "mh/isinf.h"
+#include "mh/isposinf.h"
+#include "mh/isneginf.h"
+#include "mh/isfinite.h"
+#include "mh/sum.h"
+#include "mh/prod.h"
 #include "mh/mean.h"
 #include "mh/var.h"
 #include "mh/stddev.h"
 #include "mh/rms.h"
+#include "mh/min.h"
+#include "mh/max.h"
+#include "mh/ptp.h"
+#include "mh/max_index.h"
+#include "mh/min_index.h"
+#include "mh/argmax.h"
+#include "mh/argmin.h"
+#include "mh/maximum.h"
+#include "mh/minimum.h"
+#include "mh/minmax.h"
+#include "mh/cumsum.h"
+#include "mh/cumprod.h"
+#include "mh/mulsum.h"
+#include "mh/seq.h"
+#include "mh/logseq.h"
+#include "mh/eye.h"
+#include "mh/rand.h"
 #include "mh/math/sqrt.h"
 #include "mh/math/cbrt.h"
 #include "mh/math/log.h"
@@ -78,10 +127,66 @@ extern VALUE cRT;
 
 typedef double dfloat; // Type aliases for shorter notation
                        // following the codebase naming convention.
+DEF_NARRAY_COERCE_CAST_METHOD_FUNC(dfloat)
+DEF_NARRAY_TO_A_METHOD_FUNC(dfloat)
+DEF_NARRAY_FILL_METHOD_FUNC(dfloat)
+DEF_NARRAY_FORMAT_METHOD_FUNC(dfloat)
+DEF_NARRAY_FORMAT_TO_A_METHOD_FUNC(dfloat)
+DEF_NARRAY_INSPECT_METHOD_FUNC(dfloat)
+#ifdef __SSE2__
+DEF_NARRAY_DFLT_ADD_SSE2_METHOD_FUNC()
+DEF_NARRAY_DFLT_SUB_SSE2_METHOD_FUNC()
+DEF_NARRAY_DFLT_MUL_SSE2_METHOD_FUNC()
+DEF_NARRAY_DFLT_DIV_SSE2_METHOD_FUNC()
+#else
+DEF_NARRAY_ADD_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_SUB_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_MUL_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_DIV_METHOD_FUNC(dfloat, numo_cDFloat)
+#endif
+DEF_NARRAY_FLT_MOD_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_DIVMOD_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_FLOOR_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_ROUND_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_CEIL_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_TRUNC_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_RINT_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_EQ_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_NE_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_NEARLY_EQ_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_GT_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_GE_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_LT_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_LE_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_CLIP_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_ISNAN_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_ISINF_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_ISPOSINF_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_ISNEGINF_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_ISFINITE_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_SUM_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_PROD_METHOD_FUNC(dfloat, numo_cDFloat)
 DEF_NARRAY_FLT_MEAN_METHOD_FUNC(dfloat, numo_cDFloat, double, numo_cDFloat)
 DEF_NARRAY_FLT_VAR_METHOD_FUNC(dfloat, numo_cDFloat, double, numo_cDFloat)
 DEF_NARRAY_FLT_STDDEV_METHOD_FUNC(dfloat, numo_cDFloat, double, numo_cDFloat)
 DEF_NARRAY_FLT_RMS_METHOD_FUNC(dfloat, numo_cDFloat, double, numo_cDFloat)
+DEF_NARRAY_FLT_MIN_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_MAX_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_PTP_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_MAX_INDEX_METHOD_FUNC(dfloat)
+DEF_NARRAY_FLT_MIN_INDEX_METHOD_FUNC(dfloat)
+DEF_NARRAY_FLT_ARGMAX_METHOD_FUNC(dfloat)
+DEF_NARRAY_FLT_ARGMIN_METHOD_FUNC(dfloat)
+DEF_NARRAY_FLT_MAXIMUM_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_MINIMUM_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_MINMAX_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_CUMSUM_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_CUMPROD_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_MULSUM_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_FLT_SEQ_METHOD_FUNC(dfloat)
+DEF_NARRAY_FLT_LOGSEQ_METHOD_FUNC(dfloat)
+DEF_NARRAY_EYE_METHOD_FUNC(dfloat)
+DEF_NARRAY_FLT_RAND_METHOD_FUNC(dfloat)
 #ifdef __SSE2__
 DEF_NARRAY_FLT_SQRT_SSE2_DBL_METHOD_FUNC(dfloat, numo_cDFloat)
 #else
@@ -1235,171 +1340,6 @@ static VALUE dfloat_aset(int argc, VALUE* argv, VALUE self) {
   return argv[argc];
 }
 
-static VALUE dfloat_coerce_cast(VALUE self, VALUE type) {
-  return Qnil;
-}
-
-static void iter_dfloat_to_a(na_loop_t* const lp) {
-  size_t i, s1;
-  char* p1;
-  size_t* idx1;
-  dtype x;
-  volatile VALUE a, y;
-
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  a = rb_ary_new2(i);
-  rb_ary_push(lp->args[1].value, a);
-  if (idx1) {
-    for (; i--;) {
-      GET_DATA_INDEX(p1, idx1, dtype, x);
-      y = m_data_to_num(x);
-      rb_ary_push(a, y);
-    }
-  } else {
-    for (; i--;) {
-      GET_DATA_STRIDE(p1, s1, dtype, x);
-      y = m_data_to_num(x);
-      rb_ary_push(a, y);
-    }
-  }
-}
-
-static VALUE dfloat_to_a(VALUE self) {
-  ndfunc_arg_in_t ain[3] = { { Qnil, 0 }, { sym_loop_opt }, { sym_option } };
-  ndfunc_arg_out_t aout[1] = { { rb_cArray, 0 } }; // dummy?
-  ndfunc_t ndf = { iter_dfloat_to_a, FULL_LOOP_NIP, 3, 1, ain, aout };
-  return na_ndloop_cast_narray_to_rarray(&ndf, self, Qnil);
-}
-
-static void iter_dfloat_fill(na_loop_t* const lp) {
-  size_t i;
-  char* p1;
-  ssize_t s1;
-  size_t* idx1;
-  VALUE x = lp->option;
-  dtype y;
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  y = m_num_to_data(x);
-  if (idx1) {
-    for (; i--;) {
-      SET_DATA_INDEX(p1, idx1, dtype, y);
-    }
-  } else {
-    for (; i--;) {
-      SET_DATA_STRIDE(p1, s1, dtype, y);
-    }
-  }
-}
-
-static VALUE dfloat_fill(VALUE self, VALUE val) {
-  ndfunc_arg_in_t ain[2] = { { OVERWRITE, 0 }, { sym_option } };
-  ndfunc_t ndf = { iter_dfloat_fill, FULL_LOOP, 2, 0, ain, 0 };
-
-  na_ndloop(&ndf, 2, self, val);
-  return self;
-}
-
-static VALUE format_dfloat(VALUE fmt, dtype* x) {
-  // fix-me
-  char s[48];
-  int n;
-
-  if (NIL_P(fmt)) {
-    n = m_sprintf(s, *x);
-    return rb_str_new(s, n);
-  }
-  return rb_funcall(fmt, '%', 1, m_data_to_num(*x));
-}
-
-static void iter_dfloat_format(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  size_t* idx1;
-  dtype* x;
-  VALUE y;
-  VALUE fmt = lp->option;
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR(lp, 1, p2, s2);
-  if (idx1) {
-    for (; i--;) {
-      x = (dtype*)(p1 + *idx1);
-      idx1++;
-      y = format_dfloat(fmt, x);
-      SET_DATA_STRIDE(p2, s2, VALUE, y);
-    }
-  } else {
-    for (; i--;) {
-      x = (dtype*)p1;
-      p1 += s1;
-      y = format_dfloat(fmt, x);
-      SET_DATA_STRIDE(p2, s2, VALUE, y);
-    }
-  }
-}
-
-static VALUE dfloat_format(int argc, VALUE* argv, VALUE self) {
-  VALUE fmt = Qnil;
-
-  ndfunc_arg_in_t ain[2] = { { Qnil, 0 }, { sym_option } };
-  ndfunc_arg_out_t aout[1] = { { numo_cRObject, 0 } };
-  ndfunc_t ndf = { iter_dfloat_format, FULL_LOOP_NIP, 2, 1, ain, aout };
-
-  rb_scan_args(argc, argv, "01", &fmt);
-  return na_ndloop(&ndf, 2, self, fmt);
-}
-
-static void iter_dfloat_format_to_a(na_loop_t* const lp) {
-  size_t i;
-  char* p1;
-  ssize_t s1;
-  size_t* idx1;
-  dtype* x;
-  VALUE y;
-  volatile VALUE a;
-  VALUE fmt = lp->option;
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  a = rb_ary_new2(i);
-  rb_ary_push(lp->args[1].value, a);
-  if (idx1) {
-    for (; i--;) {
-      x = (dtype*)(p1 + *idx1);
-      idx1++;
-      y = format_dfloat(fmt, x);
-      rb_ary_push(a, y);
-    }
-  } else {
-    for (; i--;) {
-      x = (dtype*)p1;
-      p1 += s1;
-      y = format_dfloat(fmt, x);
-      rb_ary_push(a, y);
-    }
-  }
-}
-
-static VALUE dfloat_format_to_a(int argc, VALUE* argv, VALUE self) {
-  VALUE fmt = Qnil;
-  ndfunc_arg_in_t ain[3] = { { Qnil, 0 }, { sym_loop_opt }, { sym_option } };
-  ndfunc_arg_out_t aout[1] = { { rb_cArray, 0 } }; // dummy?
-  ndfunc_t ndf = { iter_dfloat_format_to_a, FULL_LOOP_NIP, 3, 1, ain, aout };
-
-  rb_scan_args(argc, argv, "01", &fmt);
-  return na_ndloop_cast_narray_to_rarray(&ndf, self, fmt);
-}
-
-static VALUE iter_dfloat_inspect(char* ptr, size_t pos, VALUE fmt) {
-  return format_dfloat(fmt, (dtype*)(ptr + pos));
-}
-
-static VALUE dfloat_inspect(VALUE ary) {
-  return na_ndloop_inspect(ary, iter_dfloat_inspect, Qnil);
-}
-
 static void iter_dfloat_each(na_loop_t* const lp) {
   size_t i, s1;
   char* p1;
@@ -1682,3369 +1622,354 @@ static VALUE dfloat_abs(VALUE self) {
   return na_ndloop(&ndf, 1, self);
 }
 
-#define check_intdivzero(y)                                                                    \
-  {}
-
-static void iter_dfloat_add(na_loop_t* const lp) {
-  size_t i = 0;
-  size_t n;
+static void iter_dfloat_pow(na_loop_t* const lp) {
+  size_t i;
   char *p1, *p2, *p3;
   ssize_t s1, s2, s3;
-
-#ifdef __SSE2__
-  size_t cnt;
-  size_t cnt_simd_loop = -1;
-
-  __m128d a;
-  __m128d b;
-
-  size_t num_pack; // Number of elements packed for SIMD.
-  num_pack = SIMD_ALIGNMENT_SIZE / sizeof(dtype);
-#endif
-  INIT_COUNTER(lp, n);
+  dtype x, y;
+  INIT_COUNTER(lp, i);
   INIT_PTR(lp, 0, p1, s1);
   INIT_PTR(lp, 1, p2, s2);
   INIT_PTR(lp, 2, p3, s3);
-
-  //
-  if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype)) &&
-      is_aligned(p3, sizeof(dtype))) {
-
-    if (s1 == sizeof(dtype) && s2 == sizeof(dtype) && s3 == sizeof(dtype)) {
-#ifdef __SSE2__
-      // Check number of elements. & Check same alignment.
-      if ((n >= num_pack) &&
-          is_same_aligned3(
-            &((dtype*)p1)[i], &((dtype*)p2)[i], &((dtype*)p3)[i], SIMD_ALIGNMENT_SIZE
-          )) {
-        // Calculate up to the position just before the start of SIMD computation.
-        cnt = get_count_of_elements_not_aligned_to_simd_size(
-          &((dtype*)p1)[i], SIMD_ALIGNMENT_SIZE, sizeof(dtype)
-        );
-#endif
-        if (p1 == p3) { // inplace case
-#ifdef __SSE2__
-          for (; i < cnt; i++) {
-#else
-        for (; i < n; i++) {
-          check_intdivzero(((dtype*)p2)[i]);
-#endif
-            ((dtype*)p1)[i] = m_add(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        } else {
-#ifdef __SSE2__
-          for (; i < cnt; i++) {
-#else
-        for (; i < n; i++) {
-          check_intdivzero(((dtype*)p2)[i]);
-#endif
-            ((dtype*)p3)[i] = m_add(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        }
-#ifdef __SSE2__
-        // Get the count of SIMD computation loops.
-        cnt_simd_loop = (n - i) % num_pack;
-
-        // SIMD computation.
-        if (p1 == p3) { // inplace case
-          for (; i < n - cnt_simd_loop; i += num_pack) {
-            a = _mm_load_pd(&((dtype*)p1)[i]);
-            b = _mm_load_pd(&((dtype*)p2)[i]);
-            a = _mm_add_pd(a, b);
-            _mm_store_pd(&((dtype*)p1)[i], a);
-          }
-        } else {
-          for (; i < n - cnt_simd_loop; i += num_pack) {
-            a = _mm_load_pd(&((dtype*)p1)[i]);
-            b = _mm_load_pd(&((dtype*)p2)[i]);
-            a = _mm_add_pd(a, b);
-            _mm_stream_pd(&((dtype*)p3)[i], a);
-          }
-        }
-      }
-
-      // Compute the remainder of the SIMD operation.
-      if (cnt_simd_loop != 0) {
-        if (p1 == p3) { // inplace case
-          for (; i < n; i++) {
-            check_intdivzero(((dtype*)p2)[i]);
-            ((dtype*)p1)[i] = m_add(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        } else {
-          for (; i < n; i++) {
-            check_intdivzero(((dtype*)p2)[i]);
-            ((dtype*)p3)[i] = m_add(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        }
-      }
-#endif
-      return;
-    }
-
-    if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype)) &&
-        is_aligned_step(s3, sizeof(dtype))) {
-      //
-
-      if (s2 == 0) { // Broadcasting from scalar value.
-        check_intdivzero(*(dtype*)p2);
-        if (s1 == sizeof(dtype) && s3 == sizeof(dtype)) {
-#ifdef __SSE2__
-          // Broadcast a scalar value and use it for SIMD computation.
-          b = _mm_load1_pd(&((dtype*)p2)[0]);
-
-          // Check number of elements. & Check same alignment.
-          if ((n >= num_pack) &&
-              is_same_aligned2(&((dtype*)p1)[i], &((dtype*)p3)[i], SIMD_ALIGNMENT_SIZE)) {
-            // Calculate up to the position just before the start of SIMD computation.
-            cnt = get_count_of_elements_not_aligned_to_simd_size(
-              &((dtype*)p1)[i], SIMD_ALIGNMENT_SIZE, sizeof(dtype)
-            );
-#endif
-            if (p1 == p3) { // inplace case
-#ifdef __SSE2__
-              for (; i < cnt; i++) {
-#else
-            for (; i < n; i++) {
-#endif
-                ((dtype*)p1)[i] = m_add(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            } else {
-#ifdef __SSE2__
-              for (; i < cnt; i++) {
-#else
-            for (; i < n; i++) {
-#endif
-                ((dtype*)p3)[i] = m_add(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            }
-#ifdef __SSE2__
-            // Get the count of SIMD computation loops.
-            cnt_simd_loop = (n - i) % num_pack;
-
-            // SIMD computation.
-            if (p1 == p3) { // inplace case
-              for (; i < n - cnt_simd_loop; i += num_pack) {
-                a = _mm_load_pd(&((dtype*)p1)[i]);
-                a = _mm_add_pd(a, b);
-                _mm_store_pd(&((dtype*)p1)[i], a);
-              }
-            } else {
-              for (; i < n - cnt_simd_loop; i += num_pack) {
-                a = _mm_load_pd(&((dtype*)p1)[i]);
-                a = _mm_add_pd(a, b);
-                _mm_stream_pd(&((dtype*)p3)[i], a);
-              }
-            }
-          }
-
-          // Compute the remainder of the SIMD operation.
-          if (cnt_simd_loop != 0) {
-            if (p1 == p3) { // inplace case
-              for (; i < n; i++) {
-                ((dtype*)p1)[i] = m_add(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            } else {
-              for (; i < n; i++) {
-                ((dtype*)p3)[i] = m_add(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            }
-          }
-#endif
-        } else {
-          for (i = 0; i < n; i++) {
-            *(dtype*)p3 = m_add(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p3 += s3;
-          }
-        }
-      } else {
-        if (p1 == p3) { // inplace case
-          for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p1 = m_add(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p2 += s2;
-          }
-        } else {
-          for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p3 = m_add(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p2 += s2;
-            p3 += s3;
-          }
-        }
-      }
-
-      return;
-      //
-    }
-  }
-  for (i = 0; i < n; i++) {
-    dtype x, y, z;
+  for (; i--;) {
     GET_DATA_STRIDE(p1, s1, dtype, x);
     GET_DATA_STRIDE(p2, s2, dtype, y);
-    check_intdivzero(y);
-    z = m_add(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, z);
+    x = m_pow(x, y);
+    SET_DATA_STRIDE(p3, s3, dtype, x);
+  }
+}
+
+static void iter_dfloat_pow_int32(na_loop_t* const lp) {
+  size_t i;
+  char *p1, *p2, *p3;
+  ssize_t s1, s2, s3;
+  dtype x;
+  int32_t y;
+  INIT_COUNTER(lp, i);
+  INIT_PTR(lp, 0, p1, s1);
+  INIT_PTR(lp, 1, p2, s2);
+  INIT_PTR(lp, 2, p3, s3);
+  for (; i--;) {
+    GET_DATA_STRIDE(p1, s1, dtype, x);
+    GET_DATA_STRIDE(p2, s2, int32_t, y);
+    x = m_pow_int(x, y);
+    SET_DATA_STRIDE(p3, s3, dtype, x);
   }
-  //
 }
-#undef check_intdivzero
 
-static VALUE dfloat_add_self(VALUE self, VALUE other) {
+static VALUE dfloat_pow_self(VALUE self, VALUE other) {
   ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
+  ndfunc_arg_in_t ain_i[2] = { { cT, 0 }, { numo_cInt32, 0 } };
   ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_add, STRIDE_LOOP, 2, 1, ain, aout };
+  ndfunc_t ndf = { iter_dfloat_pow, STRIDE_LOOP, 2, 1, ain, aout };
+  ndfunc_t ndf_i = { iter_dfloat_pow_int32, STRIDE_LOOP, 2, 1, ain_i, aout };
 
-  return na_ndloop(&ndf, 2, self, other);
+  // fixme : use na.integer?
+  if (FIXNUM_P(other) || rb_obj_is_kind_of(other, numo_cInt32)) {
+    return na_ndloop(&ndf_i, 2, self, other);
+  } else {
+    return na_ndloop(&ndf, 2, self, other);
+  }
 }
 
-static VALUE dfloat_add(VALUE self, VALUE other) {
+static VALUE dfloat_pow(VALUE self, VALUE other) {
 
   VALUE klass, v;
-
   klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
   if (klass == cT) {
-    return dfloat_add_self(self, other);
+    return dfloat_pow_self(self, other);
   } else {
     v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, '+', 1, other);
+    return rb_funcall(v, id_pow, 1, other);
   }
 }
 
-#define check_intdivzero(y)                                                                    \
-  {}
+static void iter_dfloat_minus(na_loop_t* const lp) {
+  size_t i, n;
+  char *p1, *p2;
+  ssize_t s1, s2;
+  size_t *idx1, *idx2;
+  dtype x;
 
-static void iter_dfloat_sub(na_loop_t* const lp) {
-  size_t i = 0;
-  size_t n;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-
-#ifdef __SSE2__
-  size_t cnt;
-  size_t cnt_simd_loop = -1;
-
-  __m128d a;
-  __m128d b;
-
-  size_t num_pack; // Number of elements packed for SIMD.
-  num_pack = SIMD_ALIGNMENT_SIZE / sizeof(dtype);
-#endif
   INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-
-  //
-  if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype)) &&
-      is_aligned(p3, sizeof(dtype))) {
-
-    if (s1 == sizeof(dtype) && s2 == sizeof(dtype) && s3 == sizeof(dtype)) {
-#ifdef __SSE2__
-      // Check number of elements. & Check same alignment.
-      if ((n >= num_pack) &&
-          is_same_aligned3(
-            &((dtype*)p1)[i], &((dtype*)p2)[i], &((dtype*)p3)[i], SIMD_ALIGNMENT_SIZE
-          )) {
-        // Calculate up to the position just before the start of SIMD computation.
-        cnt = get_count_of_elements_not_aligned_to_simd_size(
-          &((dtype*)p1)[i], SIMD_ALIGNMENT_SIZE, sizeof(dtype)
-        );
-#endif
-        if (p1 == p3) { // inplace case
-#ifdef __SSE2__
-          for (; i < cnt; i++) {
-#else
-        for (; i < n; i++) {
-          check_intdivzero(((dtype*)p2)[i]);
-#endif
-            ((dtype*)p1)[i] = m_sub(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        } else {
-#ifdef __SSE2__
-          for (; i < cnt; i++) {
-#else
-        for (; i < n; i++) {
-          check_intdivzero(((dtype*)p2)[i]);
-#endif
-            ((dtype*)p3)[i] = m_sub(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        }
-
-#ifdef __SSE2__
-        // Get the count of SIMD computation loops.
-        cnt_simd_loop = (n - i) % num_pack;
+  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
+  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
 
-        // SIMD computation.
-        if (p1 == p3) { // inplace case
-          for (; i < n - cnt_simd_loop; i += num_pack) {
-            a = _mm_load_pd(&((dtype*)p1)[i]);
-            b = _mm_load_pd(&((dtype*)p2)[i]);
-            a = _mm_sub_pd(a, b);
-            _mm_store_pd(&((dtype*)p1)[i], a);
-          }
-        } else {
-          for (; i < n - cnt_simd_loop; i += num_pack) {
-            a = _mm_load_pd(&((dtype*)p1)[i]);
-            b = _mm_load_pd(&((dtype*)p2)[i]);
-            a = _mm_sub_pd(a, b);
-            _mm_stream_pd(&((dtype*)p3)[i], a);
-          }
-        }
+  if (idx1) {
+    if (idx2) {
+      for (i = 0; i < n; i++) {
+        GET_DATA_INDEX(p1, idx1, dtype, x);
+        x = m_minus(x);
+        SET_DATA_INDEX(p2, idx2, dtype, x);
       }
-
-      // Compute the remainder of the SIMD operation.
-      if (cnt_simd_loop != 0) {
-        if (p1 == p3) { // inplace case
-          for (; i < n; i++) {
-            check_intdivzero(((dtype*)p2)[i]);
-            ((dtype*)p1)[i] = m_sub(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        } else {
-          for (; i < n; i++) {
-            check_intdivzero(((dtype*)p2)[i]);
-            ((dtype*)p3)[i] = m_sub(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        }
+    } else {
+      for (i = 0; i < n; i++) {
+        GET_DATA_INDEX(p1, idx1, dtype, x);
+        x = m_minus(x);
+        SET_DATA_STRIDE(p2, s2, dtype, x);
       }
-#endif
-      return;
     }
-
-    if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype)) &&
-        is_aligned_step(s3, sizeof(dtype))) {
+  } else {
+    if (idx2) {
+      for (i = 0; i < n; i++) {
+        GET_DATA_STRIDE(p1, s1, dtype, x);
+        x = m_minus(x);
+        SET_DATA_INDEX(p2, idx2, dtype, x);
+      }
+    } else {
       //
-
-      if (s2 == 0) { // Broadcasting from scalar value.
-        check_intdivzero(*(dtype*)p2);
-        if (s1 == sizeof(dtype) && s3 == sizeof(dtype)) {
-#ifdef __SSE2__
-          // Broadcast a scalar value and use it for SIMD computation.
-          b = _mm_load1_pd(&((dtype*)p2)[0]);
-
-          // Check number of elements. & Check same alignment.
-          if ((n >= num_pack) &&
-              is_same_aligned2(&((dtype*)p1)[i], &((dtype*)p3)[i], SIMD_ALIGNMENT_SIZE)) {
-            // Calculate up to the position just before the start of SIMD computation.
-            cnt = get_count_of_elements_not_aligned_to_simd_size(
-              &((dtype*)p1)[i], SIMD_ALIGNMENT_SIZE, sizeof(dtype)
-            );
-#endif
-            if (p1 == p3) { // inplace case
-#ifdef __SSE2__
-              for (; i < cnt; i++) {
-#else
-            for (; i < n; i++) {
-#endif
-                ((dtype*)p1)[i] = m_sub(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            } else {
-#ifdef __SSE2__
-              for (; i < cnt; i++) {
-#else
-            for (; i < n; i++) {
-#endif
-                ((dtype*)p3)[i] = m_sub(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            }
-
-#ifdef __SSE2__
-            // Get the count of SIMD computation loops.
-            cnt_simd_loop = (n - i) % num_pack;
-
-            // SIMD computation.
-            if (p1 == p3) { // inplace case
-              for (; i < n - cnt_simd_loop; i += num_pack) {
-                a = _mm_load_pd(&((dtype*)p1)[i]);
-                a = _mm_sub_pd(a, b);
-                _mm_store_pd(&((dtype*)p1)[i], a);
-              }
-            } else {
-              for (; i < n - cnt_simd_loop; i += num_pack) {
-                a = _mm_load_pd(&((dtype*)p1)[i]);
-                a = _mm_sub_pd(a, b);
-                _mm_stream_pd(&((dtype*)p3)[i], a);
-              }
-            }
-          }
-
-          // Compute the remainder of the SIMD operation.
-          if (cnt_simd_loop != 0) {
-            if (p1 == p3) { // inplace case
-              for (; i < n; i++) {
-                ((dtype*)p1)[i] = m_sub(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            } else {
-              for (; i < n; i++) {
-                ((dtype*)p3)[i] = m_sub(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            }
-          }
-#endif
-        } else {
+      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
+        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
           for (i = 0; i < n; i++) {
-            *(dtype*)p3 = m_sub(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p3 += s3;
+            ((dtype*)p2)[i] = m_minus(((dtype*)p1)[i]);
           }
+          return;
         }
-      } else {
-        if (p1 == p3) { // inplace case
-          for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p1 = m_sub(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p2 += s2;
-          }
-        } else {
+        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
+          //
           for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p3 = m_sub(*(dtype*)p1, *(dtype*)p2);
+            *(dtype*)p2 = m_minus(*(dtype*)p1);
             p1 += s1;
             p2 += s2;
-            p3 += s3;
           }
+          return;
+          //
         }
       }
-
-      return;
+      for (i = 0; i < n; i++) {
+        GET_DATA_STRIDE(p1, s1, dtype, x);
+        x = m_minus(x);
+        SET_DATA_STRIDE(p2, s2, dtype, x);
+      }
       //
     }
   }
-  for (i = 0; i < n; i++) {
-    dtype x, y, z;
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    check_intdivzero(y);
-    z = m_sub(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, z);
-  }
-  //
 }
-#undef check_intdivzero
 
-static VALUE dfloat_sub_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
+static VALUE dfloat_minus(VALUE self) {
+  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
   ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_sub, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_sub(VALUE self, VALUE other) {
-
-  VALUE klass, v;
+  ndfunc_t ndf = { iter_dfloat_minus, FULL_LOOP, 1, 1, ain, aout };
 
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_sub_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, '-', 1, other);
-  }
+  return na_ndloop(&ndf, 1, self);
 }
 
-#define check_intdivzero(y)                                                                    \
-  {}
-
-static void iter_dfloat_mul(na_loop_t* const lp) {
-  size_t i = 0;
-  size_t n;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-
-#ifdef __SSE2__
-  size_t cnt;
-  size_t cnt_simd_loop = -1;
-
-  __m128d a;
-  __m128d b;
+static void iter_dfloat_reciprocal(na_loop_t* const lp) {
+  size_t i, n;
+  char *p1, *p2;
+  ssize_t s1, s2;
+  size_t *idx1, *idx2;
+  dtype x;
 
-  size_t num_pack; // Number of elements packed for SIMD.
-  num_pack = SIMD_ALIGNMENT_SIZE / sizeof(dtype);
-#endif
   INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-
-  //
-  if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype)) &&
-      is_aligned(p3, sizeof(dtype))) {
+  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
+  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
 
-    if (s1 == sizeof(dtype) && s2 == sizeof(dtype) && s3 == sizeof(dtype)) {
-#ifdef __SSE2__
-      // Check number of elements. & Check same alignment.
-      if ((n >= num_pack) &&
-          is_same_aligned3(
-            &((dtype*)p1)[i], &((dtype*)p2)[i], &((dtype*)p3)[i], SIMD_ALIGNMENT_SIZE
-          )) {
-        // Calculate up to the position just before the start of SIMD computation.
-        cnt = get_count_of_elements_not_aligned_to_simd_size(
-          &((dtype*)p1)[i], SIMD_ALIGNMENT_SIZE, sizeof(dtype)
-        );
-#endif
-        if (p1 == p3) { // inplace case
-#ifdef __SSE2__
-          for (; i < cnt; i++) {
-#else
-        for (; i < n; i++) {
-          check_intdivzero(((dtype*)p2)[i]);
-#endif
-            ((dtype*)p1)[i] = m_mul(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        } else {
-#ifdef __SSE2__
-          for (; i < cnt; i++) {
-#else
-        for (; i < n; i++) {
-          check_intdivzero(((dtype*)p2)[i]);
-#endif
-            ((dtype*)p3)[i] = m_mul(((dtype*)p1)[i], ((dtype*)p2)[i]);
+  if (idx1) {
+    if (idx2) {
+      for (i = 0; i < n; i++) {
+        GET_DATA_INDEX(p1, idx1, dtype, x);
+        x = m_reciprocal(x);
+        SET_DATA_INDEX(p2, idx2, dtype, x);
+      }
+    } else {
+      for (i = 0; i < n; i++) {
+        GET_DATA_INDEX(p1, idx1, dtype, x);
+        x = m_reciprocal(x);
+        SET_DATA_STRIDE(p2, s2, dtype, x);
+      }
+    }
+  } else {
+    if (idx2) {
+      for (i = 0; i < n; i++) {
+        GET_DATA_STRIDE(p1, s1, dtype, x);
+        x = m_reciprocal(x);
+        SET_DATA_INDEX(p2, idx2, dtype, x);
+      }
+    } else {
+      //
+      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
+        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
+          for (i = 0; i < n; i++) {
+            ((dtype*)p2)[i] = m_reciprocal(((dtype*)p1)[i]);
           }
+          return;
         }
-
-#ifdef __SSE2__
-        // Get the count of SIMD computation loops.
-        cnt_simd_loop = (n - i) % num_pack;
-
-        // SIMD computation.
-        if (p1 == p3) { // inplace case
-          for (; i < n - cnt_simd_loop; i += num_pack) {
-            a = _mm_load_pd(&((dtype*)p1)[i]);
-            b = _mm_load_pd(&((dtype*)p2)[i]);
-            a = _mm_mul_pd(a, b);
-            _mm_store_pd(&((dtype*)p1)[i], a);
-          }
-        } else {
-          for (; i < n - cnt_simd_loop; i += num_pack) {
-            a = _mm_load_pd(&((dtype*)p1)[i]);
-            b = _mm_load_pd(&((dtype*)p2)[i]);
-            a = _mm_mul_pd(a, b);
-            _mm_stream_pd(&((dtype*)p3)[i], a);
+        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
+          //
+          for (i = 0; i < n; i++) {
+            *(dtype*)p2 = m_reciprocal(*(dtype*)p1);
+            p1 += s1;
+            p2 += s2;
           }
+          return;
+          //
         }
       }
-
-      // Compute the remainder of the SIMD operation.
-      if (cnt_simd_loop != 0) {
-        if (p1 == p3) { // inplace case
-          for (; i < n; i++) {
-            check_intdivzero(((dtype*)p2)[i]);
-            ((dtype*)p1)[i] = m_mul(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        } else {
-          for (; i < n; i++) {
-            check_intdivzero(((dtype*)p2)[i]);
-            ((dtype*)p3)[i] = m_mul(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        }
+      for (i = 0; i < n; i++) {
+        GET_DATA_STRIDE(p1, s1, dtype, x);
+        x = m_reciprocal(x);
+        SET_DATA_STRIDE(p2, s2, dtype, x);
       }
-#endif
-      return;
-    }
-
-    if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype)) &&
-        is_aligned_step(s3, sizeof(dtype))) {
       //
+    }
+  }
+}
 
-      if (s2 == 0) { // Broadcasting from scalar value.
-        check_intdivzero(*(dtype*)p2);
-        if (s1 == sizeof(dtype) && s3 == sizeof(dtype)) {
-#ifdef __SSE2__
-          // Broadcast a scalar value and use it for SIMD computation.
-          b = _mm_load1_pd(&((dtype*)p2)[0]);
-
-          // Check number of elements. & Check same alignment.
-          if ((n >= num_pack) &&
-              is_same_aligned2(&((dtype*)p1)[i], &((dtype*)p3)[i], SIMD_ALIGNMENT_SIZE)) {
-            // Calculate up to the position just before the start of SIMD computation.
-            cnt = get_count_of_elements_not_aligned_to_simd_size(
-              &((dtype*)p1)[i], SIMD_ALIGNMENT_SIZE, sizeof(dtype)
-            );
-#endif
-            if (p1 == p3) { // inplace case
-#ifdef __SSE2__
-              for (; i < cnt; i++) {
-#else
-            for (; i < n; i++) {
-#endif
-                ((dtype*)p1)[i] = m_mul(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            } else {
-#ifdef __SSE2__
-              for (; i < cnt; i++) {
-#else
-            for (; i < n; i++) {
-#endif
-                ((dtype*)p3)[i] = m_mul(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            }
+static VALUE dfloat_reciprocal(VALUE self) {
+  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
+  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
+  ndfunc_t ndf = { iter_dfloat_reciprocal, FULL_LOOP, 1, 1, ain, aout };
 
-#ifdef __SSE2__
-            // Get the count of SIMD computation loops.
-            cnt_simd_loop = (n - i) % num_pack;
-
-            // SIMD computation.
-            if (p1 == p3) { // inplace case
-              for (; i < n - cnt_simd_loop; i += num_pack) {
-                a = _mm_load_pd(&((dtype*)p1)[i]);
-                a = _mm_mul_pd(a, b);
-                _mm_store_pd(&((dtype*)p1)[i], a);
-              }
-            } else {
-              for (; i < n - cnt_simd_loop; i += num_pack) {
-                a = _mm_load_pd(&((dtype*)p1)[i]);
-                a = _mm_mul_pd(a, b);
-                _mm_stream_pd(&((dtype*)p3)[i], a);
-              }
-            }
-          }
+  return na_ndloop(&ndf, 1, self);
+}
 
-          // Compute the remainder of the SIMD operation.
-          if (cnt_simd_loop != 0) {
-            if (p1 == p3) { // inplace case
-              for (; i < n; i++) {
-                ((dtype*)p1)[i] = m_mul(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            } else {
-              for (; i < n; i++) {
-                ((dtype*)p3)[i] = m_mul(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            }
-          }
-#endif
-        } else {
+static void iter_dfloat_sign(na_loop_t* const lp) {
+  size_t i, n;
+  char *p1, *p2;
+  ssize_t s1, s2;
+  size_t *idx1, *idx2;
+  dtype x;
+
+  INIT_COUNTER(lp, n);
+  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
+  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
+
+  if (idx1) {
+    if (idx2) {
+      for (i = 0; i < n; i++) {
+        GET_DATA_INDEX(p1, idx1, dtype, x);
+        x = m_sign(x);
+        SET_DATA_INDEX(p2, idx2, dtype, x);
+      }
+    } else {
+      for (i = 0; i < n; i++) {
+        GET_DATA_INDEX(p1, idx1, dtype, x);
+        x = m_sign(x);
+        SET_DATA_STRIDE(p2, s2, dtype, x);
+      }
+    }
+  } else {
+    if (idx2) {
+      for (i = 0; i < n; i++) {
+        GET_DATA_STRIDE(p1, s1, dtype, x);
+        x = m_sign(x);
+        SET_DATA_INDEX(p2, idx2, dtype, x);
+      }
+    } else {
+      //
+      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
+        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
           for (i = 0; i < n; i++) {
-            *(dtype*)p3 = m_mul(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p3 += s3;
+            ((dtype*)p2)[i] = m_sign(((dtype*)p1)[i]);
           }
+          return;
         }
-      } else {
-        if (p1 == p3) { // inplace case
-          for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p1 = m_mul(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p2 += s2;
-          }
-        } else {
+        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
+          //
           for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p3 = m_mul(*(dtype*)p1, *(dtype*)p2);
+            *(dtype*)p2 = m_sign(*(dtype*)p1);
             p1 += s1;
             p2 += s2;
-            p3 += s3;
           }
+          return;
+          //
         }
       }
-
-      return;
+      for (i = 0; i < n; i++) {
+        GET_DATA_STRIDE(p1, s1, dtype, x);
+        x = m_sign(x);
+        SET_DATA_STRIDE(p2, s2, dtype, x);
+      }
       //
     }
   }
-  for (i = 0; i < n; i++) {
-    dtype x, y, z;
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    check_intdivzero(y);
-    z = m_mul(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, z);
-  }
-  //
 }
-#undef check_intdivzero
 
-static VALUE dfloat_mul_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
+static VALUE dfloat_sign(VALUE self) {
+  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
   ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_mul, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_mul(VALUE self, VALUE other) {
-
-  VALUE klass, v;
+  ndfunc_t ndf = { iter_dfloat_sign, FULL_LOOP, 1, 1, ain, aout };
 
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_mul_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, '*', 1, other);
-  }
+  return na_ndloop(&ndf, 1, self);
 }
 
-#define check_intdivzero(y)                                                                    \
-  {}
-
-static void iter_dfloat_div(na_loop_t* const lp) {
-  size_t i = 0;
-  size_t n;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-
-#ifdef __SSE2__
-  size_t cnt;
-  size_t cnt_simd_loop = -1;
-
-  __m128d a;
-  __m128d b;
+static void iter_dfloat_square(na_loop_t* const lp) {
+  size_t i, n;
+  char *p1, *p2;
+  ssize_t s1, s2;
+  size_t *idx1, *idx2;
+  dtype x;
 
-  size_t num_pack; // Number of elements packed for SIMD.
-  num_pack = SIMD_ALIGNMENT_SIZE / sizeof(dtype);
-#endif
   INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-
-  //
-  if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype)) &&
-      is_aligned(p3, sizeof(dtype))) {
-
-    if (s1 == sizeof(dtype) && s2 == sizeof(dtype) && s3 == sizeof(dtype)) {
-#ifdef __SSE2__
-      // Check number of elements. & Check same alignment.
-      if ((n >= num_pack) &&
-          is_same_aligned3(
-            &((dtype*)p1)[i], &((dtype*)p2)[i], &((dtype*)p3)[i], SIMD_ALIGNMENT_SIZE
-          )) {
-        // Calculate up to the position just before the start of SIMD computation.
-        cnt = get_count_of_elements_not_aligned_to_simd_size(
-          &((dtype*)p1)[i], SIMD_ALIGNMENT_SIZE, sizeof(dtype)
-        );
-#endif
-        if (p1 == p3) { // inplace case
-#ifdef __SSE2__
-          for (; i < cnt; i++) {
-#else
-        for (; i < n; i++) {
-#endif
-            ((dtype*)p1)[i] = m_div(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        } else {
-#ifdef __SSE2__
-          for (; i < cnt; i++) {
-#else
-        for (; i < n; i++) {
-#endif
-            ((dtype*)p3)[i] = m_div(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        }
-
-#ifdef __SSE2__
-        // Get the count of SIMD computation loops.
-        cnt_simd_loop = (n - i) % num_pack;
+  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
+  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
 
-        // SIMD computation.
-        if (p1 == p3) { // inplace case
-          for (; i < n - cnt_simd_loop; i += num_pack) {
-            a = _mm_load_pd(&((dtype*)p1)[i]);
-            b = _mm_load_pd(&((dtype*)p2)[i]);
-            a = _mm_div_pd(a, b);
-            _mm_store_pd(&((dtype*)p1)[i], a);
-          }
-        } else {
-          for (; i < n - cnt_simd_loop; i += num_pack) {
-            a = _mm_load_pd(&((dtype*)p1)[i]);
-            b = _mm_load_pd(&((dtype*)p2)[i]);
-            a = _mm_div_pd(a, b);
-            _mm_stream_pd(&((dtype*)p3)[i], a);
-          }
-        }
+  if (idx1) {
+    if (idx2) {
+      for (i = 0; i < n; i++) {
+        GET_DATA_INDEX(p1, idx1, dtype, x);
+        x = m_square(x);
+        SET_DATA_INDEX(p2, idx2, dtype, x);
       }
-
-      // Compute the remainder of the SIMD operation.
-      if (cnt_simd_loop != 0) {
-        if (p1 == p3) { // inplace case
-          for (; i < n; i++) {
-            check_intdivzero(((dtype*)p2)[i]);
-            ((dtype*)p1)[i] = m_div(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        } else {
-          for (; i < n; i++) {
-            check_intdivzero(((dtype*)p2)[i]);
-            ((dtype*)p3)[i] = m_div(((dtype*)p1)[i], ((dtype*)p2)[i]);
-          }
-        }
+    } else {
+      for (i = 0; i < n; i++) {
+        GET_DATA_INDEX(p1, idx1, dtype, x);
+        x = m_square(x);
+        SET_DATA_STRIDE(p2, s2, dtype, x);
       }
-#endif
-      return;
     }
-
-    if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype)) &&
-        is_aligned_step(s3, sizeof(dtype))) {
+  } else {
+    if (idx2) {
+      for (i = 0; i < n; i++) {
+        GET_DATA_STRIDE(p1, s1, dtype, x);
+        x = m_square(x);
+        SET_DATA_INDEX(p2, idx2, dtype, x);
+      }
+    } else {
       //
-
-      if (s2 == 0) { // Broadcasting from scalar value.
-        check_intdivzero(*(dtype*)p2);
-        if (s1 == sizeof(dtype) && s3 == sizeof(dtype)) {
-#ifdef __SSE2__
-          // Broadcast a scalar value and use it for SIMD computation.
-          b = _mm_load1_pd(&((dtype*)p2)[0]);
-
-          // Check number of elements. & Check same alignment.
-          if ((n >= num_pack) &&
-              is_same_aligned2(&((dtype*)p1)[i], &((dtype*)p3)[i], SIMD_ALIGNMENT_SIZE)) {
-            // Calculate up to the position just before the start of SIMD computation.
-            cnt = get_count_of_elements_not_aligned_to_simd_size(
-              &((dtype*)p1)[i], SIMD_ALIGNMENT_SIZE, sizeof(dtype)
-            );
-#endif
-            if (p1 == p3) { // inplace case
-#ifdef __SSE2__
-              for (; i < cnt; i++) {
-#else
-            for (; i < n; i++) {
-              check_intdivzero(((dtype*)p2)[i]);
-#endif
-                ((dtype*)p1)[i] = m_div(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            } else {
-#ifdef __SSE2__
-              for (; i < cnt; i++) {
-#else
-            for (; i < n; i++) {
-              check_intdivzero(((dtype*)p2)[i]);
-#endif
-                ((dtype*)p3)[i] = m_div(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            }
-
-#ifdef __SSE2__
-            // Get the count of SIMD computation loops.
-            cnt_simd_loop = (n - i) % num_pack;
-
-            // SIMD computation.
-            if (p1 == p3) { // inplace case
-              for (; i < n - cnt_simd_loop; i += num_pack) {
-                a = _mm_load_pd(&((dtype*)p1)[i]);
-                a = _mm_div_pd(a, b);
-                _mm_store_pd(&((dtype*)p1)[i], a);
-              }
-            } else {
-              for (; i < n - cnt_simd_loop; i += num_pack) {
-                a = _mm_load_pd(&((dtype*)p1)[i]);
-                a = _mm_div_pd(a, b);
-                _mm_stream_pd(&((dtype*)p3)[i], a);
-              }
-            }
-          }
-
-          // Compute the remainder of the SIMD operation.
-          if (cnt_simd_loop != 0) {
-            if (p1 == p3) { // inplace case
-              for (; i < n; i++) {
-                ((dtype*)p1)[i] = m_div(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            } else {
-              for (; i < n; i++) {
-                ((dtype*)p3)[i] = m_div(((dtype*)p1)[i], *(dtype*)p2);
-              }
-            }
-          }
-#endif
-        } else {
+      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
+        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
           for (i = 0; i < n; i++) {
-            *(dtype*)p3 = m_div(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p3 += s3;
+            ((dtype*)p2)[i] = m_square(((dtype*)p1)[i]);
           }
+          return;
         }
-      } else {
-        if (p1 == p3) { // inplace case
-          for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p1 = m_div(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p2 += s2;
-          }
-        } else {
+        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
+          //
           for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p3 = m_div(*(dtype*)p1, *(dtype*)p2);
+            *(dtype*)p2 = m_square(*(dtype*)p1);
             p1 += s1;
             p2 += s2;
-            p3 += s3;
           }
-        }
-      }
-
-      return;
-      //
-    }
-  }
-  for (i = 0; i < n; i++) {
-    dtype x, y, z;
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    check_intdivzero(y);
-    z = m_div(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, z);
-  }
-  //
-}
-#undef check_intdivzero
-
-static VALUE dfloat_div_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_div, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_div(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_div_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, '/', 1, other);
-  }
-}
-
-#define check_intdivzero(y)                                                                    \
-  {}
-
-static void iter_dfloat_mod(na_loop_t* const lp) {
-  size_t i = 0;
-  size_t n;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-
-  //
-  if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype)) &&
-      is_aligned(p3, sizeof(dtype))) {
-
-    if (s1 == sizeof(dtype) && s2 == sizeof(dtype) && s3 == sizeof(dtype)) {
-      if (p1 == p3) { // inplace case
-        for (; i < n; i++) {
-          check_intdivzero(((dtype*)p2)[i]);
-          ((dtype*)p1)[i] = m_mod(((dtype*)p1)[i], ((dtype*)p2)[i]);
-        }
-      } else {
-        for (; i < n; i++) {
-          check_intdivzero(((dtype*)p2)[i]);
-          ((dtype*)p3)[i] = m_mod(((dtype*)p1)[i], ((dtype*)p2)[i]);
-        }
-      }
-      return;
-    }
-
-    if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype)) &&
-        is_aligned_step(s3, sizeof(dtype))) {
-      //
-
-      if (s2 == 0) { // Broadcasting from scalar value.
-        check_intdivzero(*(dtype*)p2);
-        if (s1 == sizeof(dtype) && s3 == sizeof(dtype)) {
-          if (p1 == p3) { // inplace case
-            for (; i < n; i++) {
-              ((dtype*)p1)[i] = m_mod(((dtype*)p1)[i], *(dtype*)p2);
-            }
-          } else {
-            for (; i < n; i++) {
-              ((dtype*)p3)[i] = m_mod(((dtype*)p1)[i], *(dtype*)p2);
-            }
-          }
-        } else {
-          for (i = 0; i < n; i++) {
-            *(dtype*)p3 = m_mod(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p3 += s3;
-          }
-        }
-      } else {
-        if (p1 == p3) { // inplace case
-          for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p1 = m_mod(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p2 += s2;
-          }
-        } else {
-          for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p3 = m_mod(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p2 += s2;
-            p3 += s3;
-          }
-        }
-      }
-
-      return;
-      //
-    }
-  }
-  for (i = 0; i < n; i++) {
-    dtype x, y, z;
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    check_intdivzero(y);
-    z = m_mod(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, z);
-  }
-  //
-}
-#undef check_intdivzero
-
-static VALUE dfloat_mod_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_mod, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_mod(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_mod_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, '%', 1, other);
-  }
-}
-
-static void iter_dfloat_divmod(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2, *p3, *p4;
-  ssize_t s1, s2, s3, s4;
-  dtype x, y, a, b;
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-  INIT_PTR(lp, 3, p4, s4);
-  for (i = n; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    m_divmod(x, y, a, b);
-    SET_DATA_STRIDE(p3, s3, dtype, a);
-    SET_DATA_STRIDE(p4, s4, dtype, b);
-  }
-}
-
-static VALUE dfloat_divmod_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_divmod, STRIDE_LOOP, 2, 2, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_divmod(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_divmod_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, id_divmod, 1, other);
-  }
-}
-
-static void iter_dfloat_pow(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-  dtype x, y;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    x = m_pow(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, x);
-  }
-}
-
-static void iter_dfloat_pow_int32(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-  dtype x;
-  int32_t y;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, int32_t, y);
-    x = m_pow_int(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, x);
-  }
-}
-
-static VALUE dfloat_pow_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_in_t ain_i[2] = { { cT, 0 }, { numo_cInt32, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_pow, STRIDE_LOOP, 2, 1, ain, aout };
-  ndfunc_t ndf_i = { iter_dfloat_pow_int32, STRIDE_LOOP, 2, 1, ain_i, aout };
-
-  // fixme : use na.integer?
-  if (FIXNUM_P(other) || rb_obj_is_kind_of(other, numo_cInt32)) {
-    return na_ndloop(&ndf_i, 2, self, other);
-  } else {
-    return na_ndloop(&ndf, 2, self, other);
-  }
-}
-
-static VALUE dfloat_pow(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_pow_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, id_pow, 1, other);
-  }
-}
-
-static void iter_dfloat_minus(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  size_t *idx1, *idx2;
-  dtype x;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
-
-  if (idx1) {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_minus(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_minus(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-    }
-  } else {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_minus(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      //
-      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
-        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
-          for (i = 0; i < n; i++) {
-            ((dtype*)p2)[i] = m_minus(((dtype*)p1)[i]);
-          }
-          return;
-        }
-        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
-          //
-          for (i = 0; i < n; i++) {
-            *(dtype*)p2 = m_minus(*(dtype*)p1);
-            p1 += s1;
-            p2 += s2;
-          }
-          return;
-          //
+          return;
+          //
         }
       }
       for (i = 0; i < n; i++) {
         GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_minus(x);
+        x = m_square(x);
         SET_DATA_STRIDE(p2, s2, dtype, x);
       }
-      //
-    }
-  }
-}
-
-static VALUE dfloat_minus(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_minus, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_reciprocal(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  size_t *idx1, *idx2;
-  dtype x;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
-
-  if (idx1) {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_reciprocal(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_reciprocal(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-    }
-  } else {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_reciprocal(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      //
-      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
-        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
-          for (i = 0; i < n; i++) {
-            ((dtype*)p2)[i] = m_reciprocal(((dtype*)p1)[i]);
-          }
-          return;
-        }
-        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
-          //
-          for (i = 0; i < n; i++) {
-            *(dtype*)p2 = m_reciprocal(*(dtype*)p1);
-            p1 += s1;
-            p2 += s2;
-          }
-          return;
-          //
-        }
-      }
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_reciprocal(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-      //
-    }
-  }
-}
-
-static VALUE dfloat_reciprocal(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_reciprocal, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_sign(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  size_t *idx1, *idx2;
-  dtype x;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
-
-  if (idx1) {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_sign(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_sign(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-    }
-  } else {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_sign(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      //
-      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
-        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
-          for (i = 0; i < n; i++) {
-            ((dtype*)p2)[i] = m_sign(((dtype*)p1)[i]);
-          }
-          return;
-        }
-        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
-          //
-          for (i = 0; i < n; i++) {
-            *(dtype*)p2 = m_sign(*(dtype*)p1);
-            p1 += s1;
-            p2 += s2;
-          }
-          return;
-          //
-        }
-      }
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_sign(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-      //
-    }
-  }
-}
-
-static VALUE dfloat_sign(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_sign, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_square(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  size_t *idx1, *idx2;
-  dtype x;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
-
-  if (idx1) {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_square(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_square(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-    }
-  } else {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_square(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      //
-      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
-        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
-          for (i = 0; i < n; i++) {
-            ((dtype*)p2)[i] = m_square(((dtype*)p1)[i]);
-          }
-          return;
-        }
-        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
-          //
-          for (i = 0; i < n; i++) {
-            *(dtype*)p2 = m_square(*(dtype*)p1);
-            p1 += s1;
-            p2 += s2;
-          }
-          return;
-          //
-        }
-      }
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_square(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-      //
-    }
-  }
-}
-
-static VALUE dfloat_square(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_square, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_eq(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  BIT_DIGIT* a3;
-  size_t p3;
-  ssize_t s1, s2, s3;
-  dtype x, y;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR_BIT(lp, 2, a3, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    b = (m_eq(x, y)) ? 1 : 0;
-    STORE_BIT(a3, p3, b);
-    p3 += s3;
-  }
-}
-
-static VALUE dfloat_eq_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_eq, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_eq(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_eq_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, id_eq, 1, other);
-  }
-}
-
-static void iter_dfloat_ne(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  BIT_DIGIT* a3;
-  size_t p3;
-  ssize_t s1, s2, s3;
-  dtype x, y;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR_BIT(lp, 2, a3, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    b = (m_ne(x, y)) ? 1 : 0;
-    STORE_BIT(a3, p3, b);
-    p3 += s3;
-  }
-}
-
-static VALUE dfloat_ne_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_ne, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_ne(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_ne_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, id_ne, 1, other);
-  }
-}
-
-static void iter_dfloat_nearly_eq(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  BIT_DIGIT* a3;
-  size_t p3;
-  ssize_t s1, s2, s3;
-  dtype x, y;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR_BIT(lp, 2, a3, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    b = (m_nearly_eq(x, y)) ? 1 : 0;
-    STORE_BIT(a3, p3, b);
-    p3 += s3;
-  }
-}
-
-static VALUE dfloat_nearly_eq_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_nearly_eq, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_nearly_eq(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_nearly_eq_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, id_nearly_eq, 1, other);
-  }
-}
-
-static void iter_dfloat_floor(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  size_t *idx1, *idx2;
-  dtype x;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
-
-  if (idx1) {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_floor(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_floor(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-    }
-  } else {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_floor(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      //
-      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
-        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
-          for (i = 0; i < n; i++) {
-            ((dtype*)p2)[i] = m_floor(((dtype*)p1)[i]);
-          }
-          return;
-        }
-        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
-          //
-          for (i = 0; i < n; i++) {
-            *(dtype*)p2 = m_floor(*(dtype*)p1);
-            p1 += s1;
-            p2 += s2;
-          }
-          return;
-          //
-        }
-      }
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_floor(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-      //
-    }
-  }
-}
-
-static VALUE dfloat_floor(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_floor, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_round(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  size_t *idx1, *idx2;
-  dtype x;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
-
-  if (idx1) {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_round(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_round(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-    }
-  } else {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_round(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      //
-      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
-        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
-          for (i = 0; i < n; i++) {
-            ((dtype*)p2)[i] = m_round(((dtype*)p1)[i]);
-          }
-          return;
-        }
-        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
-          //
-          for (i = 0; i < n; i++) {
-            *(dtype*)p2 = m_round(*(dtype*)p1);
-            p1 += s1;
-            p2 += s2;
-          }
-          return;
-          //
-        }
-      }
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_round(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-      //
-    }
-  }
-}
-
-static VALUE dfloat_round(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_round, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_ceil(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  size_t *idx1, *idx2;
-  dtype x;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
-
-  if (idx1) {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_ceil(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_ceil(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-    }
-  } else {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_ceil(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      //
-      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
-        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
-          for (i = 0; i < n; i++) {
-            ((dtype*)p2)[i] = m_ceil(((dtype*)p1)[i]);
-          }
-          return;
-        }
-        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
-          //
-          for (i = 0; i < n; i++) {
-            *(dtype*)p2 = m_ceil(*(dtype*)p1);
-            p1 += s1;
-            p2 += s2;
-          }
-          return;
-          //
-        }
-      }
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_ceil(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-      //
-    }
-  }
-}
-
-static VALUE dfloat_ceil(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_ceil, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_trunc(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  size_t *idx1, *idx2;
-  dtype x;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
-
-  if (idx1) {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_trunc(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_trunc(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-    }
-  } else {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_trunc(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      //
-      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
-        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
-          for (i = 0; i < n; i++) {
-            ((dtype*)p2)[i] = m_trunc(((dtype*)p1)[i]);
-          }
-          return;
-        }
-        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
-          //
-          for (i = 0; i < n; i++) {
-            *(dtype*)p2 = m_trunc(*(dtype*)p1);
-            p1 += s1;
-            p2 += s2;
-          }
-          return;
-          //
-        }
-      }
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_trunc(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-      //
-    }
-  }
-}
-
-static VALUE dfloat_trunc(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_trunc, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_rint(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  size_t *idx1, *idx2;
-  dtype x;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_IDX(lp, 1, p2, s2, idx2);
-
-  if (idx1) {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_rint(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      for (i = 0; i < n; i++) {
-        GET_DATA_INDEX(p1, idx1, dtype, x);
-        x = m_rint(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-    }
-  } else {
-    if (idx2) {
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_rint(x);
-        SET_DATA_INDEX(p2, idx2, dtype, x);
-      }
-    } else {
-      //
-      if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype))) {
-        if (s1 == sizeof(dtype) && s2 == sizeof(dtype)) {
-          for (i = 0; i < n; i++) {
-            ((dtype*)p2)[i] = m_rint(((dtype*)p1)[i]);
-          }
-          return;
-        }
-        if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype))) {
-          //
-          for (i = 0; i < n; i++) {
-            *(dtype*)p2 = m_rint(*(dtype*)p1);
-            p1 += s1;
-            p2 += s2;
-          }
-          return;
-          //
-        }
-      }
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_rint(x);
-        SET_DATA_STRIDE(p2, s2, dtype, x);
-      }
-      //
-    }
-  }
-}
-
-static VALUE dfloat_rint(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_rint, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-#define check_intdivzero(y)                                                                    \
-  {}
-
-static void iter_dfloat_copysign(na_loop_t* const lp) {
-  size_t i = 0;
-  size_t n;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-
-  //
-  if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype)) &&
-      is_aligned(p3, sizeof(dtype))) {
-
-    if (s1 == sizeof(dtype) && s2 == sizeof(dtype) && s3 == sizeof(dtype)) {
-      if (p1 == p3) { // inplace case
-        for (; i < n; i++) {
-          check_intdivzero(((dtype*)p2)[i]);
-          ((dtype*)p1)[i] = m_copysign(((dtype*)p1)[i], ((dtype*)p2)[i]);
-        }
-      } else {
-        for (; i < n; i++) {
-          check_intdivzero(((dtype*)p2)[i]);
-          ((dtype*)p3)[i] = m_copysign(((dtype*)p1)[i], ((dtype*)p2)[i]);
-        }
-      }
-      return;
-    }
-
-    if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype)) &&
-        is_aligned_step(s3, sizeof(dtype))) {
-      //
-
-      if (s2 == 0) { // Broadcasting from scalar value.
-        check_intdivzero(*(dtype*)p2);
-        if (s1 == sizeof(dtype) && s3 == sizeof(dtype)) {
-          if (p1 == p3) { // inplace case
-            for (; i < n; i++) {
-              ((dtype*)p1)[i] = m_copysign(((dtype*)p1)[i], *(dtype*)p2);
-            }
-          } else {
-            for (; i < n; i++) {
-              ((dtype*)p3)[i] = m_copysign(((dtype*)p1)[i], *(dtype*)p2);
-            }
-          }
-        } else {
-          for (i = 0; i < n; i++) {
-            *(dtype*)p3 = m_copysign(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p3 += s3;
-          }
-        }
-      } else {
-        if (p1 == p3) { // inplace case
-          for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p1 = m_copysign(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p2 += s2;
-          }
-        } else {
-          for (i = 0; i < n; i++) {
-            check_intdivzero(*(dtype*)p2);
-            *(dtype*)p3 = m_copysign(*(dtype*)p1, *(dtype*)p2);
-            p1 += s1;
-            p2 += s2;
-            p3 += s3;
-          }
-        }
-      }
-
-      return;
-      //
-    }
-  }
-  for (i = 0; i < n; i++) {
-    dtype x, y, z;
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    check_intdivzero(y);
-    z = m_copysign(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, z);
-  }
-  //
-}
-#undef check_intdivzero
-
-static VALUE dfloat_copysign_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_copysign, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_copysign(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_copysign_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, id_copysign, 1, other);
-  }
-}
-
-static void iter_dfloat_signbit(na_loop_t* const lp) {
-  size_t i;
-  char* p1;
-  BIT_DIGIT* a2;
-  size_t p2;
-  ssize_t s1, s2;
-  size_t* idx1;
-  dtype x;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_BIT(lp, 1, a2, p2, s2);
-  if (idx1) {
-    for (; i--;) {
-      GET_DATA_INDEX(p1, idx1, dtype, x);
-      b = (m_signbit(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  } else {
-    for (; i--;) {
-      GET_DATA_STRIDE(p1, s1, dtype, x);
-      b = (m_signbit(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  }
-}
-
-static VALUE dfloat_signbit(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_signbit, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_modf(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-  dtype x, y, z;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    m_modf(x, y, z);
-    SET_DATA_STRIDE(p2, s2, dtype, y);
-    SET_DATA_STRIDE(p3, s3, dtype, z);
-  }
-}
-
-static VALUE dfloat_modf(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_modf, STRIDE_LOOP, 1, 2, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_gt(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  BIT_DIGIT* a3;
-  size_t p3;
-  ssize_t s1, s2, s3;
-  dtype x, y;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR_BIT(lp, 2, a3, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    b = (m_gt(x, y)) ? 1 : 0;
-    STORE_BIT(a3, p3, b);
-    p3 += s3;
-  }
-}
-
-static VALUE dfloat_gt_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_gt, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_gt(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_gt_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, id_gt, 1, other);
-  }
-}
-
-static void iter_dfloat_ge(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  BIT_DIGIT* a3;
-  size_t p3;
-  ssize_t s1, s2, s3;
-  dtype x, y;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR_BIT(lp, 2, a3, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    b = (m_ge(x, y)) ? 1 : 0;
-    STORE_BIT(a3, p3, b);
-    p3 += s3;
-  }
-}
-
-static VALUE dfloat_ge_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_ge, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_ge(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_ge_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, id_ge, 1, other);
-  }
-}
-
-static void iter_dfloat_lt(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  BIT_DIGIT* a3;
-  size_t p3;
-  ssize_t s1, s2, s3;
-  dtype x, y;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR_BIT(lp, 2, a3, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    b = (m_lt(x, y)) ? 1 : 0;
-    STORE_BIT(a3, p3, b);
-    p3 += s3;
-  }
-}
-
-static VALUE dfloat_lt_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_lt, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_lt(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_lt_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, id_lt, 1, other);
-  }
-}
-
-static void iter_dfloat_le(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  BIT_DIGIT* a3;
-  size_t p3;
-  ssize_t s1, s2, s3;
-  dtype x, y;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR_BIT(lp, 2, a3, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    b = (m_le(x, y)) ? 1 : 0;
-    STORE_BIT(a3, p3, b);
-    p3 += s3;
-  }
-}
-
-static VALUE dfloat_le_self(VALUE self, VALUE other) {
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_le, STRIDE_LOOP, 2, 1, ain, aout };
-
-  return na_ndloop(&ndf, 2, self, other);
-}
-
-static VALUE dfloat_le(VALUE self, VALUE other) {
-
-  VALUE klass, v;
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
-  if (klass == cT) {
-    return dfloat_le_self(self, other);
-  } else {
-    v = rb_funcall(klass, id_cast, 1, self);
-    return rb_funcall(v, id_le, 1, other);
-  }
-}
-
-static void iter_dfloat_clip(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2, *p3, *p4;
-  ssize_t s1, s2, s3, s4;
-  dtype x, min, max;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-  INIT_PTR(lp, 3, p4, s4);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, min);
-    GET_DATA_STRIDE(p3, s3, dtype, max);
-    if (m_gt(min, max)) {
-      rb_raise(nary_eOperationError, "min is greater than max");
-    }
-    if (m_lt(x, min)) {
-      x = min;
-    }
-    if (m_gt(x, max)) {
-      x = max;
-    }
-    SET_DATA_STRIDE(p4, s4, dtype, x);
-  }
-}
-
-static void iter_dfloat_clip_min(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-  dtype x, min;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, min);
-    if (m_lt(x, min)) {
-      x = min;
-    }
-    SET_DATA_STRIDE(p3, s3, dtype, x);
-  }
-}
-
-static void iter_dfloat_clip_max(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-  dtype x, max;
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-  for (; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, max);
-    if (m_gt(x, max)) {
-      x = max;
-    }
-    SET_DATA_STRIDE(p3, s3, dtype, x);
-  }
-}
-
-static VALUE dfloat_clip(VALUE self, VALUE min, VALUE max) {
-  ndfunc_arg_in_t ain[3] = { { Qnil, 0 }, { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf_min = { iter_dfloat_clip_min, STRIDE_LOOP, 2, 1, ain, aout };
-  ndfunc_t ndf_max = { iter_dfloat_clip_max, STRIDE_LOOP, 2, 1, ain, aout };
-  ndfunc_t ndf_both = { iter_dfloat_clip, STRIDE_LOOP, 3, 1, ain, aout };
-
-  if (RTEST(min)) {
-    if (RTEST(max)) {
-      return na_ndloop(&ndf_both, 3, self, min, max);
-    } else {
-      return na_ndloop(&ndf_min, 2, self, min);
-    }
-  } else {
-    if (RTEST(max)) {
-      return na_ndloop(&ndf_max, 2, self, max);
-    }
-  }
-  rb_raise(rb_eArgError, "min and max are not given");
-  return Qnil;
-}
-
-static void iter_dfloat_isnan(na_loop_t* const lp) {
-  size_t i;
-  char* p1;
-  BIT_DIGIT* a2;
-  size_t p2;
-  ssize_t s1, s2;
-  size_t* idx1;
-  dtype x;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_BIT(lp, 1, a2, p2, s2);
-  if (idx1) {
-    for (; i--;) {
-      GET_DATA_INDEX(p1, idx1, dtype, x);
-      b = (m_isnan(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  } else {
-    for (; i--;) {
-      GET_DATA_STRIDE(p1, s1, dtype, x);
-      b = (m_isnan(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  }
-}
-
-static VALUE dfloat_isnan(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_isnan, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_isinf(na_loop_t* const lp) {
-  size_t i;
-  char* p1;
-  BIT_DIGIT* a2;
-  size_t p2;
-  ssize_t s1, s2;
-  size_t* idx1;
-  dtype x;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_BIT(lp, 1, a2, p2, s2);
-  if (idx1) {
-    for (; i--;) {
-      GET_DATA_INDEX(p1, idx1, dtype, x);
-      b = (m_isinf(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  } else {
-    for (; i--;) {
-      GET_DATA_STRIDE(p1, s1, dtype, x);
-      b = (m_isinf(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  }
-}
-
-static VALUE dfloat_isinf(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_isinf, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_isposinf(na_loop_t* const lp) {
-  size_t i;
-  char* p1;
-  BIT_DIGIT* a2;
-  size_t p2;
-  ssize_t s1, s2;
-  size_t* idx1;
-  dtype x;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_BIT(lp, 1, a2, p2, s2);
-  if (idx1) {
-    for (; i--;) {
-      GET_DATA_INDEX(p1, idx1, dtype, x);
-      b = (m_isposinf(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  } else {
-    for (; i--;) {
-      GET_DATA_STRIDE(p1, s1, dtype, x);
-      b = (m_isposinf(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  }
-}
-
-static VALUE dfloat_isposinf(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_isposinf, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_isneginf(na_loop_t* const lp) {
-  size_t i;
-  char* p1;
-  BIT_DIGIT* a2;
-  size_t p2;
-  ssize_t s1, s2;
-  size_t* idx1;
-  dtype x;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_BIT(lp, 1, a2, p2, s2);
-  if (idx1) {
-    for (; i--;) {
-      GET_DATA_INDEX(p1, idx1, dtype, x);
-      b = (m_isneginf(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  } else {
-    for (; i--;) {
-      GET_DATA_STRIDE(p1, s1, dtype, x);
-      b = (m_isneginf(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  }
-}
-
-static VALUE dfloat_isneginf(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_isneginf, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_isfinite(na_loop_t* const lp) {
-  size_t i;
-  char* p1;
-  BIT_DIGIT* a2;
-  size_t p2;
-  ssize_t s1, s2;
-  size_t* idx1;
-  dtype x;
-  BIT_DIGIT b;
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  INIT_PTR_BIT(lp, 1, a2, p2, s2);
-  if (idx1) {
-    for (; i--;) {
-      GET_DATA_INDEX(p1, idx1, dtype, x);
-      b = (m_isfinite(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  } else {
-    for (; i--;) {
-      GET_DATA_STRIDE(p1, s1, dtype, x);
-      b = (m_isfinite(x)) ? 1 : 0;
-      STORE_BIT(a2, p2, b);
-      p2 += s2;
-    }
-  }
-}
-
-static VALUE dfloat_isfinite(VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
-  ndfunc_t ndf = { iter_dfloat_isfinite, FULL_LOOP, 1, 1, ain, aout };
-
-  return na_ndloop(&ndf, 1, self);
-}
-
-static void iter_dfloat_sum(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_sum(n, p1, s1);
-}
-static void iter_dfloat_sum_nan(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_sum_nan(n, p1, s1);
-}
-
-static VALUE dfloat_sum(int argc, VALUE* argv, VALUE self) {
-  VALUE v, reduce;
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_sum, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE, 2, 1, ain, aout };
-
-  reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_sum_nan);
-
-  v = na_ndloop(&ndf, 2, self, reduce);
-
-  return dfloat_extract(v);
-}
-
-static void iter_dfloat_prod(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_prod(n, p1, s1);
-}
-static void iter_dfloat_prod_nan(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_prod_nan(n, p1, s1);
-}
-
-static VALUE dfloat_prod(int argc, VALUE* argv, VALUE self) {
-  VALUE v, reduce;
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_prod, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE, 2, 1, ain, aout };
-
-  reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_prod_nan);
-
-  v = na_ndloop(&ndf, 2, self, reduce);
-
-  return dfloat_extract(v);
-}
-
-static void iter_dfloat_kahan_sum(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_kahan_sum(n, p1, s1);
-}
-static void iter_dfloat_kahan_sum_nan(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_kahan_sum_nan(n, p1, s1);
-}
-
-static VALUE dfloat_kahan_sum(int argc, VALUE* argv, VALUE self) {
-  VALUE v, reduce;
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_kahan_sum, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE, 2, 1, ain, aout };
-
-  reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_kahan_sum_nan);
-
-  v = na_ndloop(&ndf, 2, self, reduce);
-
-  return dfloat_extract(v);
-}
-
-static void iter_dfloat_min(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_min(n, p1, s1);
-}
-static void iter_dfloat_min_nan(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_min_nan(n, p1, s1);
-}
-
-static VALUE dfloat_min(int argc, VALUE* argv, VALUE self) {
-  VALUE v, reduce;
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_min, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE, 2, 1, ain, aout };
-
-  reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_min_nan);
-
-  v = na_ndloop(&ndf, 2, self, reduce);
-
-  return dfloat_extract(v);
-}
-
-static void iter_dfloat_max(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_max(n, p1, s1);
-}
-static void iter_dfloat_max_nan(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_max_nan(n, p1, s1);
-}
-
-static VALUE dfloat_max(int argc, VALUE* argv, VALUE self) {
-  VALUE v, reduce;
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_max, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE, 2, 1, ain, aout };
-
-  reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_max_nan);
-
-  v = na_ndloop(&ndf, 2, self, reduce);
-
-  return dfloat_extract(v);
-}
-
-static void iter_dfloat_ptp(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_ptp(n, p1, s1);
-}
-static void iter_dfloat_ptp_nan(na_loop_t* const lp) {
-  size_t n;
-  char *p1, *p2;
-  ssize_t s1;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
-
-  *(dtype*)p2 = f_ptp_nan(n, p1, s1);
-}
-
-static VALUE dfloat_ptp(int argc, VALUE* argv, VALUE self) {
-  VALUE v, reduce;
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_ptp, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE, 2, 1, ain, aout };
-
-  reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_ptp_nan);
-
-  v = na_ndloop(&ndf, 2, self, reduce);
-
-  return dfloat_extract(v);
-}
-
-#define idx_t int64_t
-static void iter_dfloat_max_index_index64(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *i_ptr, *o_ptr;
-  ssize_t d_step, i_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_max_index(n, d_ptr, d_step);
-
-  INIT_PTR(lp, 1, i_ptr, i_step);
-  o_ptr = NDL_PTR(lp, 2);
-  *(idx_t*)o_ptr = *(idx_t*)(i_ptr + i_step * idx);
-}
-#undef idx_t
-
-#define idx_t int32_t
-static void iter_dfloat_max_index_index32(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *i_ptr, *o_ptr;
-  ssize_t d_step, i_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_max_index(n, d_ptr, d_step);
-
-  INIT_PTR(lp, 1, i_ptr, i_step);
-  o_ptr = NDL_PTR(lp, 2);
-  *(idx_t*)o_ptr = *(idx_t*)(i_ptr + i_step * idx);
-}
-#undef idx_t
-
-#define idx_t int64_t
-static void iter_dfloat_max_index_index64_nan(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *i_ptr, *o_ptr;
-  ssize_t d_step, i_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_max_index_nan(n, d_ptr, d_step);
-
-  INIT_PTR(lp, 1, i_ptr, i_step);
-  o_ptr = NDL_PTR(lp, 2);
-  *(idx_t*)o_ptr = *(idx_t*)(i_ptr + i_step * idx);
-}
-#undef idx_t
-
-#define idx_t int32_t
-static void iter_dfloat_max_index_index32_nan(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *i_ptr, *o_ptr;
-  ssize_t d_step, i_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_max_index_nan(n, d_ptr, d_step);
-
-  INIT_PTR(lp, 1, i_ptr, i_step);
-  o_ptr = NDL_PTR(lp, 2);
-  *(idx_t*)o_ptr = *(idx_t*)(i_ptr + i_step * idx);
-}
-#undef idx_t
-
-static VALUE dfloat_max_index(int argc, VALUE* argv, VALUE self) {
-  narray_t* na;
-  VALUE idx, reduce;
-  ndfunc_arg_in_t ain[3] = { { Qnil, 0 }, { Qnil, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[1] = { { 0, 0, 0 } };
-  ndfunc_t ndf = { 0, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE | NDF_EXTRACT, 3, 1, ain, aout };
-
-  GetNArray(self, na);
-  if (na->ndim == 0) {
-    return INT2FIX(0);
-  }
-  if (na->size > (~(u_int32_t)0)) {
-    aout[0].type = numo_cInt64;
-    idx = nary_new(numo_cInt64, na->ndim, na->shape);
-    ndf.func = iter_dfloat_max_index_index64;
-
-    reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_max_index_index64_nan);
-
-  } else {
-    aout[0].type = numo_cInt32;
-    idx = nary_new(numo_cInt32, na->ndim, na->shape);
-    ndf.func = iter_dfloat_max_index_index32;
-
-    reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_max_index_index32_nan);
-  }
-  rb_funcall(idx, rb_intern("seq"), 0);
-
-  return na_ndloop(&ndf, 3, self, idx, reduce);
-}
-
-#define idx_t int64_t
-static void iter_dfloat_min_index_index64(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *i_ptr, *o_ptr;
-  ssize_t d_step, i_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_min_index(n, d_ptr, d_step);
-
-  INIT_PTR(lp, 1, i_ptr, i_step);
-  o_ptr = NDL_PTR(lp, 2);
-  *(idx_t*)o_ptr = *(idx_t*)(i_ptr + i_step * idx);
-}
-#undef idx_t
-
-#define idx_t int32_t
-static void iter_dfloat_min_index_index32(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *i_ptr, *o_ptr;
-  ssize_t d_step, i_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_min_index(n, d_ptr, d_step);
-
-  INIT_PTR(lp, 1, i_ptr, i_step);
-  o_ptr = NDL_PTR(lp, 2);
-  *(idx_t*)o_ptr = *(idx_t*)(i_ptr + i_step * idx);
-}
-#undef idx_t
-
-#define idx_t int64_t
-static void iter_dfloat_min_index_index64_nan(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *i_ptr, *o_ptr;
-  ssize_t d_step, i_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_min_index_nan(n, d_ptr, d_step);
-
-  INIT_PTR(lp, 1, i_ptr, i_step);
-  o_ptr = NDL_PTR(lp, 2);
-  *(idx_t*)o_ptr = *(idx_t*)(i_ptr + i_step * idx);
-}
-#undef idx_t
-
-#define idx_t int32_t
-static void iter_dfloat_min_index_index32_nan(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *i_ptr, *o_ptr;
-  ssize_t d_step, i_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_min_index_nan(n, d_ptr, d_step);
-
-  INIT_PTR(lp, 1, i_ptr, i_step);
-  o_ptr = NDL_PTR(lp, 2);
-  *(idx_t*)o_ptr = *(idx_t*)(i_ptr + i_step * idx);
-}
-#undef idx_t
-
-static VALUE dfloat_min_index(int argc, VALUE* argv, VALUE self) {
-  narray_t* na;
-  VALUE idx, reduce;
-  ndfunc_arg_in_t ain[3] = { { Qnil, 0 }, { Qnil, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[1] = { { 0, 0, 0 } };
-  ndfunc_t ndf = { 0, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE | NDF_EXTRACT, 3, 1, ain, aout };
-
-  GetNArray(self, na);
-  if (na->ndim == 0) {
-    return INT2FIX(0);
-  }
-  if (na->size > (~(u_int32_t)0)) {
-    aout[0].type = numo_cInt64;
-    idx = nary_new(numo_cInt64, na->ndim, na->shape);
-    ndf.func = iter_dfloat_min_index_index64;
-
-    reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_min_index_index64_nan);
-
-  } else {
-    aout[0].type = numo_cInt32;
-    idx = nary_new(numo_cInt32, na->ndim, na->shape);
-    ndf.func = iter_dfloat_min_index_index32;
-
-    reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_min_index_index32_nan);
-  }
-  rb_funcall(idx, rb_intern("seq"), 0);
-
-  return na_ndloop(&ndf, 3, self, idx, reduce);
-}
-
-#define idx_t int64_t
-static void iter_dfloat_argmax_arg64(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *o_ptr;
-  ssize_t d_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_max_index(n, d_ptr, d_step);
-
-  o_ptr = NDL_PTR(lp, 1);
-  *(idx_t*)o_ptr = (idx_t)idx;
-}
-#undef idx_t
-
-#define idx_t int32_t
-static void iter_dfloat_argmax_arg32(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *o_ptr;
-  ssize_t d_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_max_index(n, d_ptr, d_step);
-
-  o_ptr = NDL_PTR(lp, 1);
-  *(idx_t*)o_ptr = (idx_t)idx;
-}
-#undef idx_t
-
-#define idx_t int64_t
-static void iter_dfloat_argmax_arg64_nan(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *o_ptr;
-  ssize_t d_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_max_index_nan(n, d_ptr, d_step);
-
-  o_ptr = NDL_PTR(lp, 1);
-  *(idx_t*)o_ptr = (idx_t)idx;
-}
-#undef idx_t
-
-#define idx_t int32_t
-static void iter_dfloat_argmax_arg32_nan(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *o_ptr;
-  ssize_t d_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_max_index_nan(n, d_ptr, d_step);
-
-  o_ptr = NDL_PTR(lp, 1);
-  *(idx_t*)o_ptr = (idx_t)idx;
-}
-#undef idx_t
-
-static VALUE dfloat_argmax(int argc, VALUE* argv, VALUE self) {
-  narray_t* na;
-  VALUE reduce;
-  ndfunc_arg_in_t ain[2] = { { Qnil, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[1] = { { 0, 0, 0 } };
-  ndfunc_t ndf = { 0, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE | NDF_EXTRACT, 2, 1, ain, aout };
-
-  GetNArray(self, na);
-  if (na->ndim == 0) {
-    return INT2FIX(0);
-  }
-  if (na->size > (~(u_int32_t)0)) {
-    aout[0].type = numo_cInt64;
-    ndf.func = iter_dfloat_argmax_arg64;
-
-    reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_argmax_arg64_nan);
-
-  } else {
-    aout[0].type = numo_cInt32;
-    ndf.func = iter_dfloat_argmax_arg32;
-
-    reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_argmax_arg32_nan);
-  }
-
-  return na_ndloop(&ndf, 2, self, reduce);
-}
-
-#define idx_t int64_t
-static void iter_dfloat_argmin_arg64(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *o_ptr;
-  ssize_t d_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_min_index(n, d_ptr, d_step);
-
-  o_ptr = NDL_PTR(lp, 1);
-  *(idx_t*)o_ptr = (idx_t)idx;
-}
-#undef idx_t
-
-#define idx_t int32_t
-static void iter_dfloat_argmin_arg32(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *o_ptr;
-  ssize_t d_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_min_index(n, d_ptr, d_step);
-
-  o_ptr = NDL_PTR(lp, 1);
-  *(idx_t*)o_ptr = (idx_t)idx;
-}
-#undef idx_t
-
-#define idx_t int64_t
-static void iter_dfloat_argmin_arg64_nan(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *o_ptr;
-  ssize_t d_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_min_index_nan(n, d_ptr, d_step);
-
-  o_ptr = NDL_PTR(lp, 1);
-  *(idx_t*)o_ptr = (idx_t)idx;
-}
-#undef idx_t
-
-#define idx_t int32_t
-static void iter_dfloat_argmin_arg32_nan(na_loop_t* const lp) {
-  size_t n, idx;
-  char *d_ptr, *o_ptr;
-  ssize_t d_step;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, d_ptr, d_step);
-
-  idx = f_min_index_nan(n, d_ptr, d_step);
-
-  o_ptr = NDL_PTR(lp, 1);
-  *(idx_t*)o_ptr = (idx_t)idx;
-}
-#undef idx_t
-
-static VALUE dfloat_argmin(int argc, VALUE* argv, VALUE self) {
-  narray_t* na;
-  VALUE reduce;
-  ndfunc_arg_in_t ain[2] = { { Qnil, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[1] = { { 0, 0, 0 } };
-  ndfunc_t ndf = { 0, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE | NDF_EXTRACT, 2, 1, ain, aout };
-
-  GetNArray(self, na);
-  if (na->ndim == 0) {
-    return INT2FIX(0);
-  }
-  if (na->size > (~(u_int32_t)0)) {
-    aout[0].type = numo_cInt64;
-    ndf.func = iter_dfloat_argmin_arg64;
-
-    reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_argmin_arg64_nan);
-
-  } else {
-    aout[0].type = numo_cInt32;
-    ndf.func = iter_dfloat_argmin_arg32;
-
-    reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_argmin_arg32_nan);
-  }
-
-  return na_ndloop(&ndf, 2, self, reduce);
-}
-
-static void iter_dfloat_minmax(na_loop_t* const lp) {
-  size_t n;
-  char* p1;
-  ssize_t s1;
-  dtype xmin, xmax;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-
-  f_minmax(n, p1, s1, &xmin, &xmax);
-
-  *(dtype*)(lp->args[1].ptr + lp->args[1].iter[0].pos) = xmin;
-  *(dtype*)(lp->args[2].ptr + lp->args[2].iter[0].pos) = xmax;
-}
-static void iter_dfloat_minmax_nan(na_loop_t* const lp) {
-  size_t n;
-  char* p1;
-  ssize_t s1;
-  dtype xmin, xmax;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-
-  f_minmax_nan(n, p1, s1, &xmin, &xmax);
-
-  *(dtype*)(lp->args[1].ptr + lp->args[1].iter[0].pos) = xmin;
-  *(dtype*)(lp->args[2].ptr + lp->args[2].iter[0].pos) = xmax;
-}
-
-static VALUE dfloat_minmax(int argc, VALUE* argv, VALUE self) {
-  VALUE reduce;
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_t ndf = {
-    iter_dfloat_minmax, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE | NDF_EXTRACT, 2, 2, ain, aout
-  };
-
-  reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_minmax_nan);
-
-  return na_ndloop(&ndf, 2, self, reduce);
-}
-
-static void iter_dfloat_s_maximum(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-
-  for (i = 0; i < n; i++) {
-    dtype x, y, z;
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    GET_DATA(p3, dtype, z);
-    z = f_maximum(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, z);
-  }
-}
-static void iter_dfloat_s_maximum_nan(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-
-  for (i = 0; i < n; i++) {
-    dtype x, y, z;
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    GET_DATA(p3, dtype, z);
-    z = f_maximum_nan(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, z);
-  }
-}
-
-static VALUE dfloat_s_maximum(int argc, VALUE* argv, VALUE mod) {
-  VALUE a1 = Qnil;
-  VALUE a2 = Qnil;
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_s_maximum, STRIDE_LOOP_NIP, 2, 1, ain, aout };
-
-  VALUE kw_hash = Qnil;
-  ID kw_table[1] = { id_nan };
-  VALUE opts[1] = { Qundef };
-
-  rb_scan_args(argc, argv, "20:", &a1, &a2, &kw_hash);
-  rb_get_kwargs(kw_hash, kw_table, 0, 1, opts);
-  if (opts[0] != Qundef) {
-    ndf.func = iter_dfloat_s_maximum_nan;
-  }
-
-  return na_ndloop(&ndf, 2, a1, a2);
-}
-
-static void iter_dfloat_s_minimum(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-
-  for (i = 0; i < n; i++) {
-    dtype x, y, z;
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    GET_DATA(p3, dtype, z);
-    z = f_minimum(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, z);
-  }
-}
-static void iter_dfloat_s_minimum_nan(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
-
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
-
-  for (i = 0; i < n; i++) {
-    dtype x, y, z;
-    GET_DATA_STRIDE(p1, s1, dtype, x);
-    GET_DATA_STRIDE(p2, s2, dtype, y);
-    GET_DATA(p3, dtype, z);
-    z = f_minimum_nan(x, y);
-    SET_DATA_STRIDE(p3, s3, dtype, z);
-  }
-}
-
-static VALUE dfloat_s_minimum(int argc, VALUE* argv, VALUE mod) {
-  VALUE a1 = Qnil;
-  VALUE a2 = Qnil;
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_s_minimum, STRIDE_LOOP_NIP, 2, 1, ain, aout };
-
-  VALUE kw_hash = Qnil;
-  ID kw_table[1] = { id_nan };
-  VALUE opts[1] = { Qundef };
-
-  rb_scan_args(argc, argv, "20:", &a1, &a2, &kw_hash);
-  rb_get_kwargs(kw_hash, kw_table, 0, 1, opts);
-  if (opts[0] != Qundef) {
-    ndf.func = iter_dfloat_s_minimum_nan;
-  }
-
-  return na_ndloop(&ndf, 2, a1, a2);
-}
-
-static void iter_dfloat_cumsum(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  dtype x, y;
-
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-
-  GET_DATA_STRIDE(p1, s1, dtype, x);
-  SET_DATA_STRIDE(p2, s2, dtype, x);
-  for (i--; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, y);
-    m_cumsum(x, y);
-    SET_DATA_STRIDE(p2, s2, dtype, x);
-  }
-}
-static void iter_dfloat_cumsum_nan(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  dtype x, y;
-
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-
-  GET_DATA_STRIDE(p1, s1, dtype, x);
-  SET_DATA_STRIDE(p2, s2, dtype, x);
-  for (i--; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, y);
-    m_cumsum_nan(x, y);
-    SET_DATA_STRIDE(p2, s2, dtype, x);
-  }
-}
-
-static VALUE dfloat_cumsum(int argc, VALUE* argv, VALUE self) {
-  VALUE reduce;
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { sym_reduce, 0 } };
-  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = {
-    iter_dfloat_cumsum, STRIDE_LOOP | NDF_FLAT_REDUCE | NDF_CUM, 2, 1, ain, aout
-  };
-
-  reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_cumsum_nan);
-
-  return na_ndloop(&ndf, 2, self, reduce);
-}
-
-static void iter_dfloat_cumprod(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  dtype x, y;
-
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-
-  GET_DATA_STRIDE(p1, s1, dtype, x);
-  SET_DATA_STRIDE(p2, s2, dtype, x);
-  for (i--; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, y);
-    m_cumprod(x, y);
-    SET_DATA_STRIDE(p2, s2, dtype, x);
-  }
-}
-static void iter_dfloat_cumprod_nan(na_loop_t* const lp) {
-  size_t i;
-  char *p1, *p2;
-  ssize_t s1, s2;
-  dtype x, y;
-
-  INIT_COUNTER(lp, i);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-
-  GET_DATA_STRIDE(p1, s1, dtype, x);
-  SET_DATA_STRIDE(p2, s2, dtype, x);
-  for (i--; i--;) {
-    GET_DATA_STRIDE(p1, s1, dtype, y);
-    m_cumprod_nan(x, y);
-    SET_DATA_STRIDE(p2, s2, dtype, x);
+      //
+    }
   }
 }
 
-static VALUE dfloat_cumprod(int argc, VALUE* argv, VALUE self) {
-  VALUE reduce;
-  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { sym_reduce, 0 } };
+static VALUE dfloat_square(VALUE self) {
+  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
   ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = {
-    iter_dfloat_cumprod, STRIDE_LOOP | NDF_FLAT_REDUCE | NDF_CUM, 2, 1, ain, aout
-  };
-
-  reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_cumprod_nan);
+  ndfunc_t ndf = { iter_dfloat_square, FULL_LOOP, 1, 1, ain, aout };
 
-  return na_ndloop(&ndf, 2, self, reduce);
+  return na_ndloop(&ndf, 1, self);
 }
 
-//
-static void iter_dfloat_mulsum(na_loop_t* const lp) {
-  size_t i, n;
+#define check_intdivzero(y)                                                                    \
+  {}
+
+static void iter_dfloat_copysign(na_loop_t* const lp) {
+  size_t i = 0;
+  size_t n;
   char *p1, *p2, *p3;
   ssize_t s1, s2, s3;
 
@@ -5053,376 +1978,200 @@ static void iter_dfloat_mulsum(na_loop_t* const lp) {
   INIT_PTR(lp, 1, p2, s2);
   INIT_PTR(lp, 2, p3, s3);
 
-  if (s3 == 0) {
-    dtype z;
-    // Reduce loop
-    GET_DATA(p3, dtype, z);
-    for (i = 0; i < n; i++) {
-      dtype x, y;
-      GET_DATA_STRIDE(p1, s1, dtype, x);
-      GET_DATA_STRIDE(p2, s2, dtype, y);
-      m_mulsum(x, y, z);
-    }
-    SET_DATA(p3, dtype, z);
-    return;
-  } else {
-    for (i = 0; i < n; i++) {
-      dtype x, y, z;
-      GET_DATA_STRIDE(p1, s1, dtype, x);
-      GET_DATA_STRIDE(p2, s2, dtype, y);
-      GET_DATA(p3, dtype, z);
-      m_mulsum(x, y, z);
-      SET_DATA_STRIDE(p3, s3, dtype, z);
+  //
+  if (is_aligned(p1, sizeof(dtype)) && is_aligned(p2, sizeof(dtype)) &&
+      is_aligned(p3, sizeof(dtype))) {
+
+    if (s1 == sizeof(dtype) && s2 == sizeof(dtype) && s3 == sizeof(dtype)) {
+      if (p1 == p3) { // inplace case
+        for (; i < n; i++) {
+          check_intdivzero(((dtype*)p2)[i]);
+          ((dtype*)p1)[i] = m_copysign(((dtype*)p1)[i], ((dtype*)p2)[i]);
+        }
+      } else {
+        for (; i < n; i++) {
+          check_intdivzero(((dtype*)p2)[i]);
+          ((dtype*)p3)[i] = m_copysign(((dtype*)p1)[i], ((dtype*)p2)[i]);
+        }
+      }
+      return;
     }
-  }
-}
-//
-static void iter_dfloat_mulsum_nan(na_loop_t* const lp) {
-  size_t i, n;
-  char *p1, *p2, *p3;
-  ssize_t s1, s2, s3;
 
-  INIT_COUNTER(lp, n);
-  INIT_PTR(lp, 0, p1, s1);
-  INIT_PTR(lp, 1, p2, s2);
-  INIT_PTR(lp, 2, p3, s3);
+    if (is_aligned_step(s1, sizeof(dtype)) && is_aligned_step(s2, sizeof(dtype)) &&
+        is_aligned_step(s3, sizeof(dtype))) {
+      //
 
-  if (s3 == 0) {
-    dtype z;
-    // Reduce loop
-    GET_DATA(p3, dtype, z);
-    for (i = 0; i < n; i++) {
-      dtype x, y;
-      GET_DATA_STRIDE(p1, s1, dtype, x);
-      GET_DATA_STRIDE(p2, s2, dtype, y);
-      m_mulsum_nan(x, y, z);
-    }
-    SET_DATA(p3, dtype, z);
-    return;
-  } else {
-    for (i = 0; i < n; i++) {
-      dtype x, y, z;
-      GET_DATA_STRIDE(p1, s1, dtype, x);
-      GET_DATA_STRIDE(p2, s2, dtype, y);
-      GET_DATA(p3, dtype, z);
-      m_mulsum_nan(x, y, z);
-      SET_DATA_STRIDE(p3, s3, dtype, z);
+      if (s2 == 0) { // Broadcasting from scalar value.
+        check_intdivzero(*(dtype*)p2);
+        if (s1 == sizeof(dtype) && s3 == sizeof(dtype)) {
+          if (p1 == p3) { // inplace case
+            for (; i < n; i++) {
+              ((dtype*)p1)[i] = m_copysign(((dtype*)p1)[i], *(dtype*)p2);
+            }
+          } else {
+            for (; i < n; i++) {
+              ((dtype*)p3)[i] = m_copysign(((dtype*)p1)[i], *(dtype*)p2);
+            }
+          }
+        } else {
+          for (i = 0; i < n; i++) {
+            *(dtype*)p3 = m_copysign(*(dtype*)p1, *(dtype*)p2);
+            p1 += s1;
+            p3 += s3;
+          }
+        }
+      } else {
+        if (p1 == p3) { // inplace case
+          for (i = 0; i < n; i++) {
+            check_intdivzero(*(dtype*)p2);
+            *(dtype*)p1 = m_copysign(*(dtype*)p1, *(dtype*)p2);
+            p1 += s1;
+            p2 += s2;
+          }
+        } else {
+          for (i = 0; i < n; i++) {
+            check_intdivzero(*(dtype*)p2);
+            *(dtype*)p3 = m_copysign(*(dtype*)p1, *(dtype*)p2);
+            p1 += s1;
+            p2 += s2;
+            p3 += s3;
+          }
+        }
+      }
+
+      return;
+      //
     }
   }
+  for (i = 0; i < n; i++) {
+    dtype x, y, z;
+    GET_DATA_STRIDE(p1, s1, dtype, x);
+    GET_DATA_STRIDE(p2, s2, dtype, y);
+    check_intdivzero(y);
+    z = m_copysign(x, y);
+    SET_DATA_STRIDE(p3, s3, dtype, z);
+  }
+  //
 }
-//
+#undef check_intdivzero
 
-static VALUE dfloat_mulsum_self(int argc, VALUE* argv, VALUE self) {
-  VALUE v, reduce;
-  VALUE naryv[2];
-  ndfunc_arg_in_t ain[4] = { { cT, 0 }, { cT, 0 }, { sym_reduce, 0 }, { sym_init, 0 } };
+static VALUE dfloat_copysign_self(VALUE self, VALUE other) {
+  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { cT, 0 } };
   ndfunc_arg_out_t aout[1] = { { cT, 0 } };
-  ndfunc_t ndf = { iter_dfloat_mulsum, STRIDE_LOOP_NIP, 4, 1, ain, aout };
-
-  if (argc < 1) {
-    rb_raise(rb_eArgError, "wrong number of arguments (%d for >=1)", argc);
-  }
-  // should fix below: [self.ndim,other.ndim].max or?
-  naryv[0] = self;
-  naryv[1] = argv[0];
-  //
-  reduce = na_reduce_dimension(argc - 1, argv + 1, 2, naryv, &ndf, iter_dfloat_mulsum_nan);
-  //
+  ndfunc_t ndf = { iter_dfloat_copysign, STRIDE_LOOP, 2, 1, ain, aout };
 
-  v = na_ndloop(&ndf, 4, self, argv[0], reduce, m_mulsum_init);
-  return dfloat_extract(v);
+  return na_ndloop(&ndf, 2, self, other);
 }
 
-static VALUE dfloat_mulsum(int argc, VALUE* argv, VALUE self) {
-  //
+static VALUE dfloat_copysign(VALUE self, VALUE other) {
+
   VALUE klass, v;
-  //
-  if (argc < 1) {
-    rb_raise(rb_eArgError, "wrong number of arguments (%d for >=1)", argc);
-  }
-  //
-  klass = na_upcast(rb_obj_class(self), rb_obj_class(argv[0]));
+
+  klass = na_upcast(rb_obj_class(self), rb_obj_class(other));
   if (klass == cT) {
-    return dfloat_mulsum_self(argc, argv, self);
+    return dfloat_copysign_self(self, other);
   } else {
     v = rb_funcall(klass, id_cast, 1, self);
-    //
-    return rb_funcallv_kw(v, rb_intern("mulsum"), argc, argv, RB_PASS_CALLED_KEYWORDS);
-    //
+    return rb_funcall(v, id_copysign, 1, other);
   }
-  //
 }
 
-typedef dtype seq_data_t;
-
-typedef double seq_count_t;
-
-typedef struct {
-  seq_data_t beg;
-  seq_data_t step;
-  seq_count_t count;
-} seq_opt_t;
-
-static void iter_dfloat_seq(na_loop_t* const lp) {
+static void iter_dfloat_signbit(na_loop_t* const lp) {
   size_t i;
   char* p1;
-  ssize_t s1;
+  BIT_DIGIT* a2;
+  size_t p2;
+  ssize_t s1, s2;
   size_t* idx1;
   dtype x;
-  seq_data_t beg, step;
-  seq_count_t c;
-  seq_opt_t* g;
-
+  BIT_DIGIT b;
   INIT_COUNTER(lp, i);
   INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  g = (seq_opt_t*)(lp->opt_ptr);
-  beg = g->beg;
-  step = g->step;
-  c = g->count;
+  INIT_PTR_BIT(lp, 1, a2, p2, s2);
   if (idx1) {
     for (; i--;) {
-      x = f_seq(beg, step, c++);
-      *(dtype*)(p1 + *idx1) = x;
-      idx1++;
+      GET_DATA_INDEX(p1, idx1, dtype, x);
+      b = (m_signbit(x)) ? 1 : 0;
+      STORE_BIT(a2, p2, b);
+      p2 += s2;
     }
   } else {
     for (; i--;) {
-      x = f_seq(beg, step, c++);
-      *(dtype*)(p1) = x;
-      p1 += s1;
+      GET_DATA_STRIDE(p1, s1, dtype, x);
+      b = (m_signbit(x)) ? 1 : 0;
+      STORE_BIT(a2, p2, b);
+      p2 += s2;
     }
   }
-  g->count = c;
 }
 
-static VALUE dfloat_seq(int argc, VALUE* args, VALUE self) {
-  seq_opt_t* g;
-  VALUE vbeg = Qnil, vstep = Qnil;
-  ndfunc_arg_in_t ain[1] = { { OVERWRITE, 0 } };
-  ndfunc_t ndf = { iter_dfloat_seq, FULL_LOOP, 1, 0, ain, 0 };
-
-  g = ALLOCA_N(seq_opt_t, 1);
-  g->beg = m_zero;
-  g->step = m_one;
-  g->count = 0;
-  rb_scan_args(argc, args, "02", &vbeg, &vstep);
-  if (vbeg != Qnil) {
-    g->beg = m_num_to_data(vbeg);
-  }
-  if (vstep != Qnil) {
-    g->step = m_num_to_data(vstep);
-  }
+static VALUE dfloat_signbit(VALUE self) {
+  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
+  ndfunc_arg_out_t aout[1] = { { numo_cBit, 0 } };
+  ndfunc_t ndf = { iter_dfloat_signbit, FULL_LOOP, 1, 1, ain, aout };
 
-  na_ndloop3(&ndf, g, 1, self);
-  return self;
+  return na_ndloop(&ndf, 1, self);
 }
 
-typedef struct {
-  seq_data_t beg;
-  seq_data_t step;
-  seq_data_t base;
-  seq_count_t count;
-} logseq_opt_t;
-
-static void iter_dfloat_logseq(na_loop_t* const lp) {
+static void iter_dfloat_modf(na_loop_t* const lp) {
   size_t i;
-  char* p1;
-  ssize_t s1;
-  size_t* idx1;
-  dtype x;
-  seq_data_t beg, step, base;
-  seq_count_t c;
-  logseq_opt_t* g;
-
+  char *p1, *p2, *p3;
+  ssize_t s1, s2, s3;
+  dtype x, y, z;
   INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  g = (logseq_opt_t*)(lp->opt_ptr);
-  beg = g->beg;
-  step = g->step;
-  base = g->base;
-  c = g->count;
-  if (idx1) {
-    for (; i--;) {
-      x = f_seq(beg, step, c++);
-      *(dtype*)(p1 + *idx1) = m_pow(base, x);
-      idx1++;
-    }
-  } else {
-    for (; i--;) {
-      x = f_seq(beg, step, c++);
-      *(dtype*)(p1) = m_pow(base, x);
-      p1 += s1;
-    }
+  INIT_PTR(lp, 0, p1, s1);
+  INIT_PTR(lp, 1, p2, s2);
+  INIT_PTR(lp, 2, p3, s3);
+  for (; i--;) {
+    GET_DATA_STRIDE(p1, s1, dtype, x);
+    m_modf(x, y, z);
+    SET_DATA_STRIDE(p2, s2, dtype, y);
+    SET_DATA_STRIDE(p3, s3, dtype, z);
   }
-  g->count = c;
 }
 
-static VALUE dfloat_logseq(int argc, VALUE* args, VALUE self) {
-  logseq_opt_t* g;
-  VALUE vbeg, vstep, vbase;
-  ndfunc_arg_in_t ain[1] = { { OVERWRITE, 0 } };
-  ndfunc_t ndf = { iter_dfloat_logseq, FULL_LOOP, 1, 0, ain, 0 };
-
-  g = ALLOCA_N(logseq_opt_t, 1);
-  rb_scan_args(argc, args, "21", &vbeg, &vstep, &vbase);
-  g->beg = m_num_to_data(vbeg);
-  g->step = m_num_to_data(vstep);
-  if (vbase == Qnil) {
-    g->base = m_from_real(10);
-  } else {
-    g->base = m_num_to_data(vbase);
-  }
-  na_ndloop3(&ndf, g, 1, self);
-  return self;
-}
+static VALUE dfloat_modf(VALUE self) {
+  ndfunc_arg_in_t ain[1] = { { cT, 0 } };
+  ndfunc_arg_out_t aout[2] = { { cT, 0 }, { cT, 0 } };
+  ndfunc_t ndf = { iter_dfloat_modf, STRIDE_LOOP, 1, 2, ain, aout };
 
-static void iter_dfloat_eye(na_loop_t* const lp) {
-  size_t n0, n1;
-  size_t i0, i1;
-  ssize_t s0, s1;
-  char *p0, *p1;
-  char* g;
-  ssize_t kofs;
-  dtype data;
-
-  g = (char*)(lp->opt_ptr);
-  kofs = *(ssize_t*)g;
-  data = *(dtype*)(g + sizeof(ssize_t));
-
-  n0 = lp->args[0].shape[0];
-  n1 = lp->args[0].shape[1];
-  s0 = lp->args[0].iter[0].step;
-  s1 = lp->args[0].iter[1].step;
-  p0 = NDL_PTR(lp, 0);
-
-  for (i0 = 0; i0 < n0; i0++) {
-    p1 = p0;
-    for (i1 = 0; i1 < n1; i1++) {
-      *(dtype*)p1 = (i0 + kofs == i1) ? data : m_zero;
-      p1 += s1;
-    }
-    p0 += s0;
-  }
+  return na_ndloop(&ndf, 1, self);
 }
 
-static VALUE dfloat_eye(int argc, VALUE* argv, VALUE self) {
-  ndfunc_arg_in_t ain[1] = { { OVERWRITE, 2 } };
-  ndfunc_t ndf = { iter_dfloat_eye, NO_LOOP, 1, 0, ain, 0 };
-  ssize_t kofs;
-  dtype data;
-  char* g;
-  int nd;
-  narray_t* na;
-
-  // check arguments
-  if (argc > 2) {
-    rb_raise(rb_eArgError, "too many arguments (%d for 0..2)", argc);
-  } else if (argc == 2) {
-    data = m_num_to_data(argv[0]);
-    kofs = NUM2SSIZET(argv[1]);
-  } else if (argc == 1) {
-    data = m_num_to_data(argv[0]);
-    kofs = 0;
-  } else {
-    data = m_one;
-    kofs = 0;
-  }
-
-  GetNArray(self, na);
-  nd = na->ndim;
-  if (nd < 2) {
-    rb_raise(nary_eDimensionError, "less than 2-d array");
-  }
-
-  // Diagonal offset from the main diagonal.
-  if (kofs >= 0) {
-    if ((size_t)(kofs) >= na->shape[nd - 1]) {
-      rb_raise(
-        rb_eArgError,
-        "invalid diagonal offset(%" SZF "d) for "
-        "last dimension size(%" SZF "d)",
-        kofs, na->shape[nd - 1]
-      );
-    }
-  } else {
-    if ((size_t)(-kofs) >= na->shape[nd - 2]) {
-      rb_raise(
-        rb_eArgError,
-        "invalid diagonal offset(%" SZF "d) for "
-        "last-1 dimension size(%" SZF "d)",
-        kofs, na->shape[nd - 2]
-      );
-    }
-  }
+static void iter_dfloat_kahan_sum(na_loop_t* const lp) {
+  size_t n;
+  char *p1, *p2;
+  ssize_t s1;
 
-  g = ALLOCA_N(char, sizeof(ssize_t) + sizeof(dtype));
-  *(ssize_t*)g = kofs;
-  *(dtype*)(g + sizeof(ssize_t)) = data;
+  INIT_COUNTER(lp, n);
+  INIT_PTR(lp, 0, p1, s1);
+  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
 
-  na_ndloop3(&ndf, g, 1, self);
-  return self;
+  *(dtype*)p2 = f_kahan_sum(n, p1, s1);
 }
-
-typedef struct {
-  dtype low;
-  dtype max;
-} rand_opt_t;
-
-static void iter_dfloat_rand(na_loop_t* const lp) {
-  size_t i;
-  char* p1;
+static void iter_dfloat_kahan_sum_nan(na_loop_t* const lp) {
+  size_t n;
+  char *p1, *p2;
   ssize_t s1;
-  size_t* idx1;
-  dtype x;
-  rand_opt_t* g;
-  dtype low;
-  dtype max;
 
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  g = (rand_opt_t*)(lp->opt_ptr);
-  low = g->low;
-  max = g->max;
+  INIT_COUNTER(lp, n);
+  INIT_PTR(lp, 0, p1, s1);
+  p2 = lp->args[1].ptr + lp->args[1].iter[0].pos;
 
-  if (idx1) {
-    for (; i--;) {
-      x = m_add(m_rand(max), low);
-      SET_DATA_INDEX(p1, idx1, dtype, x);
-    }
-  } else {
-    for (; i--;) {
-      x = m_add(m_rand(max), low);
-      SET_DATA_STRIDE(p1, s1, dtype, x);
-    }
-  }
+  *(dtype*)p2 = f_kahan_sum_nan(n, p1, s1);
 }
 
-static VALUE dfloat_rand(int argc, VALUE* args, VALUE self) {
-  rand_opt_t g;
-  VALUE v1 = Qnil, v2 = Qnil;
-  dtype high;
-  ndfunc_arg_in_t ain[1] = { { OVERWRITE, 0 } };
-  ndfunc_t ndf = { iter_dfloat_rand, FULL_LOOP, 1, 0, ain, 0 };
-
-  rb_scan_args(argc, args, "02", &v1, &v2);
-  if (v2 == Qnil) {
-    g.low = m_zero;
-    if (v1 == Qnil) {
-
-      g.max = high = m_one;
+static VALUE dfloat_kahan_sum(int argc, VALUE* argv, VALUE self) {
+  VALUE v, reduce;
+  ndfunc_arg_in_t ain[2] = { { cT, 0 }, { sym_reduce, 0 } };
+  ndfunc_arg_out_t aout[1] = { { cT, 0 } };
+  ndfunc_t ndf = { iter_dfloat_kahan_sum, STRIDE_LOOP_NIP | NDF_FLAT_REDUCE, 2, 1, ain, aout };
 
-    } else {
-      g.max = high = m_num_to_data(v1);
-    }
+  reduce = na_reduce_dimension(argc, argv, 1, &self, &ndf, iter_dfloat_kahan_sum_nan);
 
-  } else {
-    g.low = m_num_to_data(v1);
-    high = m_num_to_data(v2);
-    g.max = m_sub(high, g.low);
-  }
+  v = na_ndloop(&ndf, 2, self, reduce);
 
-  na_ndloop3(&ndf, &g, 1, self);
-  return self;
+  return dfloat_extract(v);
 }
 
 typedef struct {