numo-narray-alt 0.9.12 → 0.9.13

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

@@ -54,6 +54,11 @@ extern VALUE cRT;
 #include "mh/op/div.h"
 #include "mh/op/mod.h"
 #include "mh/divmod.h"
+#include "mh/pow.h"
+#include "mh/minus.h"
+#include "mh/reciprocal.h"
+#include "mh/sign.h"
+#include "mh/square.h"
 #include "mh/round/floor.h"
 #include "mh/round/round.h"
 #include "mh/round/ceil.h"
@@ -95,6 +100,7 @@ extern VALUE cRT;
 #include "mh/logseq.h"
 #include "mh/eye.h"
 #include "mh/rand.h"
+#include "mh/rand_norm.h"
 #include "mh/math/sqrt.h"
 #include "mh/math/cbrt.h"
 #include "mh/math/log.h"
@@ -146,6 +152,11 @@ 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_POW_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_MINUS_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_RECIPROCAL_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_SIGN_METHOD_FUNC(dfloat, numo_cDFloat)
+DEF_NARRAY_SQUARE_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)
@@ -187,6 +198,7 @@ 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)
+DEF_NARRAY_FLT_RAND_NORM_METHOD_FUNC(dfloat)
 #ifdef __SSE2__
 DEF_NARRAY_FLT_SQRT_SSE2_DBL_METHOD_FUNC(dfloat, numo_cDFloat)
 #else
@@ -1622,348 +1634,6 @@ static VALUE dfloat_abs(VALUE self) {
   return na_ndloop(&ndf, 1, self);
 }
 
-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;
-          //
-        }
-      }
-      for (i = 0; i < n; i++) {
-        GET_DATA_STRIDE(p1, s1, dtype, x);
-        x = m_minus(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);
-}
-
 #define check_intdivzero(y)                                                                    \
   {}
 
@@ -2174,79 +1844,6 @@ static VALUE dfloat_kahan_sum(int argc, VALUE* argv, VALUE self) {
   return dfloat_extract(v);
 }
 
-typedef struct {
-  dtype mu;
-  rtype sigma;
-} randn_opt_t;
-
-static void iter_dfloat_rand_norm(na_loop_t* const lp) {
-  size_t i;
-  char* p1;
-  ssize_t s1;
-  size_t* idx1;
-
-  dtype *a0, *a1;
-
-  dtype mu;
-  rtype sigma;
-  randn_opt_t* g;
-
-  INIT_COUNTER(lp, i);
-  INIT_PTR_IDX(lp, 0, p1, s1, idx1);
-  g = (randn_opt_t*)(lp->opt_ptr);
-  mu = g->mu;
-  sigma = g->sigma;
-
-  if (idx1) {
-
-    for (; i > 1; i -= 2) {
-      a0 = (dtype*)(p1 + *idx1);
-      a1 = (dtype*)(p1 + *(idx1 + 1));
-      m_rand_norm(mu, sigma, a0, a1);
-      idx1 += 2;
-    }
-    if (i > 0) {
-      a0 = (dtype*)(p1 + *idx1);
-      m_rand_norm(mu, sigma, a0, 0);
-    }
-
-  } else {
-
-    for (; i > 1; i -= 2) {
-      a0 = (dtype*)(p1);
-      a1 = (dtype*)(p1 + s1);
-      m_rand_norm(mu, sigma, a0, a1);
-      p1 += s1 * 2;
-    }
-    if (i > 0) {
-      a0 = (dtype*)(p1);
-      m_rand_norm(mu, sigma, a0, 0);
-    }
-  }
-}
-
-static VALUE dfloat_rand_norm(int argc, VALUE* args, VALUE self) {
-  int n;
-  randn_opt_t g;
-  VALUE v1 = Qnil, v2 = Qnil;
-  ndfunc_arg_in_t ain[1] = { { OVERWRITE, 0 } };
-  ndfunc_t ndf = { iter_dfloat_rand_norm, FULL_LOOP, 1, 0, ain, 0 };
-
-  n = rb_scan_args(argc, args, "02", &v1, &v2);
-  if (n == 0) {
-    g.mu = m_zero;
-  } else {
-    g.mu = m_num_to_data(v1);
-  }
-  if (n == 2) {
-    g.sigma = NUM2DBL(v2);
-  } else {
-    g.sigma = 1;
-  }
-  na_ndloop3(&ndf, &g, 1, self);
-  return self;
-}
-
 static void iter_dfloat_poly(na_loop_t* const lp) {
   size_t i;
   dtype x, y, a;