mx-inlines.cc

Go to the documentation of this file.

/*

Copyright (C) 1993-2012 John W. Eaton
Copyright (C) 2009 Jaroslav Hajek
Copyright (C) 2009 VZLU Prague

This file is part of Octave.

Octave is free software; you can redistribute it and/or modify it
under the terms of the GNU General Public License as published by the
Free Software Foundation; either version 3 of the License, or (at your
option) any later version.

Octave is distributed in the hope that it will be useful, but WITHOUT
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
for more details.

You should have received a copy of the GNU General Public License
along with Octave; see the file COPYING.  If not, see
<http://www.gnu.org/licenses/>.

*/

#if !defined (octave_mx_inlines_h)
#define octave_mx_inlines_h 1

#include <cstddef>
#include <cmath>
#include <memory>

#include "quit.h"

#include "oct-cmplx.h"
#include "oct-locbuf.h"
#include "oct-inttypes.h"
#include "Array.h"
#include "Array-util.h"

#include "bsxfun.h"

// Provides some commonly repeated, basic loop templates.

template <class R, class S>
inline void mx_inline_fill (size_t n, R *r, S s) throw ()
{ for (size_t i = 0; i < n; i++) r[i] = s; }

#define DEFMXUNOP(F, OP) \
template <class R, class X> \
inline void F (size_t n, R *r, const X *x) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = OP x[i]; }

DEFMXUNOP (mx_inline_uminus, -)

#define DEFMXUNOPEQ(F, OP) \
template <class R> \
inline void F (size_t n, R *r) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = OP r[i]; }

DEFMXUNOPEQ (mx_inline_uminus2, -)

#define DEFMXUNBOOLOP(F, OP) \
template <class X> \
inline void F (size_t n, bool *r, const X *x) throw () \
{ const X zero = X(); for (size_t i = 0; i < n; i++) r[i] = x[i] OP zero; }

DEFMXUNBOOLOP (mx_inline_iszero, ==)
DEFMXUNBOOLOP (mx_inline_notzero, !=)

#define DEFMXBINOP(F, OP) \
template <class R, class X, class Y> \
inline void F (size_t n, R *r, const X *x, const Y *y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = x[i] OP y[i]; } \
template <class R, class X, class Y> \
inline void F (size_t n, R *r, const X *x, Y y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = x[i] OP y; } \
template <class R, class X, class Y> \
inline void F (size_t n, R *r, X x, const Y *y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = x OP y[i]; }

DEFMXBINOP (mx_inline_add, +)
DEFMXBINOP (mx_inline_sub, -)
DEFMXBINOP (mx_inline_mul, *)
DEFMXBINOP (mx_inline_div, /)

#define DEFMXBINOPEQ(F, OP) \
template <class R, class X> \
inline void F (size_t n, R *r, const X *x) throw () \
{ for (size_t i = 0; i < n; i++) r[i] OP x[i]; } \
template <class R, class X> \
inline void F (size_t n, R *r, X x) throw () \
{ for (size_t i = 0; i < n; i++) r[i] OP x; }

DEFMXBINOPEQ (mx_inline_add2, +=)
DEFMXBINOPEQ (mx_inline_sub2, -=)
DEFMXBINOPEQ (mx_inline_mul2, *=)
DEFMXBINOPEQ (mx_inline_div2, /=)

#define DEFMXCMPOP(F, OP) \
template <class X, class Y> \
inline void F (size_t n, bool *r, const X *x, const Y *y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = x[i] OP y[i]; } \
template <class X, class Y> \
inline void F (size_t n, bool *r, const X *x, Y y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = x[i] OP y; } \
template <class X, class Y> \
inline void F (size_t n, bool *r, X x, const Y *y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = x OP y[i]; }

DEFMXCMPOP (mx_inline_lt, <)
DEFMXCMPOP (mx_inline_le, <=)
DEFMXCMPOP (mx_inline_gt, >)
DEFMXCMPOP (mx_inline_ge, >=)
DEFMXCMPOP (mx_inline_eq, ==)
DEFMXCMPOP (mx_inline_ne, !=)

// Convert to logical value, for logical op purposes.
template <class T> inline bool logical_value (T x) { return x; }
template <class T> inline bool logical_value (const std::complex<T>& x)
{ return x.real () != 0 || x.imag () != 0; }
template <class T> inline bool logical_value (const octave_int<T>& x)
{ return x.value (); }

template <class X>
void mx_inline_not (size_t n, bool *r, const X* x) throw ()
{
  for (size_t i = 0; i < n; i++)
    r[i] = ! logical_value (x[i]);
}

inline void mx_inline_not2 (size_t n, bool *r) throw ()
{
  for (size_t i = 0; i < n; i++) r[i] = ! r[i];
}

#define DEFMXBOOLOP(F, NOT1, OP, NOT2) \
template <class X, class Y> \
inline void F (size_t n, bool *r, const X *x, const Y *y) throw () \
{ \
  for (size_t i = 0; i < n; i++) \
    r[i] = (NOT1 logical_value (x[i])) OP (NOT2 logical_value (y[i])); \
} \
template <class X, class Y> \
inline void F (size_t n, bool *r, const X *x, Y y) throw () \
{ \
  const bool yy = (NOT2 logical_value (y)); \
  for (size_t i = 0; i < n; i++) \
    r[i] = (NOT1 logical_value (x[i])) OP yy; \
} \
template <class X, class Y> \
inline void F (size_t n, bool *r, X x, const Y *y) throw () \
{ \
  const bool xx = (NOT1 logical_value (x)); \
  for (size_t i = 0; i < n; i++) \
    r[i] = xx OP (NOT2 logical_value (y[i])); \
}

DEFMXBOOLOP (mx_inline_and, , &, )
DEFMXBOOLOP (mx_inline_or, , |, )
DEFMXBOOLOP (mx_inline_not_and, !, &, )
DEFMXBOOLOP (mx_inline_not_or, !, |, )
DEFMXBOOLOP (mx_inline_and_not, , &, !)
DEFMXBOOLOP (mx_inline_or_not, , |, !)

#define DEFMXBOOLOPEQ(F, OP) \
template <class X> \
inline void F (size_t n, bool *r, const X *x) throw () \
{ \
  for (size_t i = 0; i < n; i++) \
    r[i] OP logical_value (x[i]); \
} \
template <class X> \
inline void F (size_t n, bool *r, X x) throw () \
{ for (size_t i = 0; i < n; i++) r[i] OP x; }

DEFMXBOOLOPEQ (mx_inline_and2, &=)
DEFMXBOOLOPEQ (mx_inline_or2, |=)

template <class T>
inline bool
mx_inline_any_nan (size_t n, const T* x)  throw ()
{
  for (size_t i = 0; i < n; i++)
    {
      if (xisnan (x[i]))
        return true;
    }

  return false;
}

template <class T>
inline bool
mx_inline_all_finite (size_t n, const T* x)  throw ()
{
  for (size_t i = 0; i < n; i++)
    {
      if (! xfinite (x[i]))
        return false;
    }

  return true;
}

template <class T>
inline bool
mx_inline_any_negative (size_t n, const T* x) throw ()
{
  for (size_t i = 0; i < n; i++)
    {
      if (x[i] < 0)
        return true;
    }

  return false;
}

template <class T>
inline bool
mx_inline_any_positive (size_t n, const T* x) throw ()
{
  for (size_t i = 0; i < n; i++)
    {
      if (x[i] > 0)
        return true;
    }

  return false;
}

template<class T>
inline bool
mx_inline_all_real (size_t n, const std::complex<T>* x) throw ()
{
  for (size_t i = 0; i < n; i++)
    {
      if (x[i].imag () != 0)
        return false;
    }

  return true;
}

#define DEFMXMAPPER(F, FUN) \
template <class T> \
inline void F (size_t n, T *r, const T *x) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = FUN (x[i]); }

template<class T>
inline void mx_inline_real (size_t n, T *r, const std::complex<T>* x) throw ()
{ for (size_t i = 0; i < n; i++) r[i] = x[i].real (); }
template<class T>
inline void mx_inline_imag (size_t n, T *r, const std::complex<T>* x) throw ()
{ for (size_t i = 0; i < n; i++) r[i] = x[i].imag (); }

// Pairwise minimums/maximums
#define DEFMXMAPPER2(F, FUN) \
template <class T> \
inline void F (size_t n, T *r, const T *x, const T *y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = FUN (x[i], y[i]); } \
template <class T> \
inline void F (size_t n, T *r, const T *x, T y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = FUN (x[i], y); } \
template <class T> \
inline void F (size_t n, T *r, T x, const T *y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = FUN (x, y[i]); }

DEFMXMAPPER2 (mx_inline_xmin, xmin)
DEFMXMAPPER2 (mx_inline_xmax, xmax)

// Specialize array-scalar max/min
#define DEFMINMAXSPEC(T, F, OP) \
template <> \
inline void F<T> (size_t n, T *r, const T *x, T y) throw () \
{ \
  if (xisnan (y)) \
    std::memcpy (r, x, n * sizeof (T)); \
  else \
    for (size_t i = 0; i < n; i++) r[i] = (x[i] OP y) ? x[i] : y; \
} \
template <> \
inline void F<T> (size_t n, T *r, T x, const T *y) throw () \
{ \
  if (xisnan (x)) \
    std::memcpy (r, y, n * sizeof (T)); \
  else \
    for (size_t i = 0; i < n; i++) r[i] = (y[i] OP x) ? y[i] : x; \
}

DEFMINMAXSPEC (double, mx_inline_xmin, <=)
DEFMINMAXSPEC (double, mx_inline_xmax, >=)
DEFMINMAXSPEC (float, mx_inline_xmin, <=)
DEFMINMAXSPEC (float, mx_inline_xmax, >=)

// Pairwise power
#define DEFMXMAPPER2X(F, FUN) \
template <class R, class X, class Y> \
inline void F (size_t n, R *r, const X *x, const Y *y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = FUN (x[i], y[i]); } \
template <class R, class X, class Y> \
inline void F (size_t n, R *r, const X *x, Y y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = FUN (x[i], y); } \
template <class R, class X, class Y> \
inline void F (size_t n, R *r, X x, const Y *y) throw () \
{ for (size_t i = 0; i < n; i++) r[i] = FUN (x, y[i]); }

// Let the compiler decide which pow to use, whichever best matches the
// arguments provided.
using std::pow;
DEFMXMAPPER2X (mx_inline_pow, pow)

// Arbitrary function appliers. The function is a template parameter to enable
// inlining.
template <class R, class X, R fun (X x)>
inline void mx_inline_map (size_t n, R *r, const X *x) throw ()
{ for (size_t i = 0; i < n; i++) r[i] = fun (x[i]); }

template <class R, class X, R fun (const X& x)>
inline void mx_inline_map (size_t n, R *r, const X *x) throw ()
{ for (size_t i = 0; i < n; i++) r[i] = fun (x[i]); }

// Appliers. Since these call the operation just once, we pass it as
// a pointer, to allow the compiler reduce number of instances.

template <class R, class X>
inline Array<R>
do_mx_unary_op (const Array<X>& x,
                void (*op) (size_t, R *, const X *) throw ())
{
  Array<R> r (x.dims ());
  op (r.numel (), r.fortran_vec (), x.data ());
  return r;
}

// Shortcuts for applying mx_inline_map.

template <class R, class X, R fun (X)>
inline Array<R>
do_mx_unary_map (const Array<X>& x)
{
  return do_mx_unary_op<R, X> (x, mx_inline_map<R, X, fun>);
}

template <class R, class X, R fun (const X&)>
inline Array<R>
do_mx_unary_map (const Array<X>& x)
{
  return do_mx_unary_op<R, X> (x, mx_inline_map<R, X, fun>);
}

template <class R>
inline Array<R>&
do_mx_inplace_op (Array<R>& r,
                  void (*op) (size_t, R *) throw ())
{
  op (r.numel (), r.fortran_vec ());
  return r;
}

template <class R, class X, class Y>
inline Array<R>
do_mm_binary_op (const Array<X>& x, const Array<Y>& y,
                 void (*op) (size_t, R *, const X *, const Y *) throw (),
                 void (*op1) (size_t, R *, X, const Y *) throw (),
                 void (*op2) (size_t, R *, const X *, Y) throw (),
                 const char *opname)
{
  dim_vector dx = x.dims (), dy = y.dims ();
  if (dx == dy)
    {
      Array<R> r (dx);
      op (r.length (), r.fortran_vec (), x.data (), y.data ());
      return r;
    }
  else if (is_valid_bsxfun (opname, dx, dy))
    {
      return do_bsxfun_op (x, y, op, op1, op2);
    }
  else
    {
      gripe_nonconformant (opname, dx, dy);
      return Array<R> ();
    }
}

template <class R, class X, class Y>
inline Array<R>
do_ms_binary_op (const Array<X>& x, const Y& y,
                 void (*op) (size_t, R *, const X *, Y) throw ())
{
  Array<R> r (x.dims ());
  op (r.length (), r.fortran_vec (), x.data (), y);
  return r;
}

template <class R, class X, class Y>
inline Array<R>
do_sm_binary_op (const X& x, const Array<Y>& y,
                 void (*op) (size_t, R *, X, const Y *) throw ())
{
  Array<R> r (y.dims ());
  op (r.length (), r.fortran_vec (), x, y.data ());
  return r;
}

template <class R, class X>
inline Array<R>&
do_mm_inplace_op (Array<R>& r, const Array<X>& x,
                  void (*op) (size_t, R *, const X *) throw (),
                  void (*op1) (size_t, R *, X) throw (),
                  const char *opname)
{
  dim_vector dr = r.dims (), dx = x.dims ();
  if (dr == dx)
    {
      op (r.length (), r.fortran_vec (), x.data ());
    }
  else if (is_valid_inplace_bsxfun (opname, dr, dx))
    {
      do_inplace_bsxfun_op (r, x, op, op1);
    }
  else
    gripe_nonconformant (opname, dr, dx);
  return r;
}

template <class R, class X>
inline Array<R>&
do_ms_inplace_op (Array<R>& r, const X& x,
                  void (*op) (size_t, R *, X) throw ())
{
  op (r.length (), r.fortran_vec (), x);
  return r;
}

template <class T1, class T2>
inline bool
mx_inline_equal (size_t n, const T1 *x, const T2 *y) throw ()
{
  for (size_t i = 0; i < n; i++)
    if (x[i] != y[i])
      return false;
  return true;
}

template <class T>
inline bool
do_mx_check (const Array<T>& a,
             bool (*op) (size_t, const T *) throw ())
{
  return op (a.numel (), a.data ());
}

// NOTE: we don't use std::norm because it typically does some heavyweight
// magic to avoid underflows, which we don't need here.
template <class T>
inline T cabsq (const std::complex<T>& c)
{ return c.real () * c.real () + c.imag () * c.imag (); }

// default. works for integers and bool.
template <class T>
inline bool xis_true (T x) { return x; }
template <class T>
inline bool xis_false (T x) { return ! x; }
// for octave_ints
template <class T>
inline bool xis_true (const octave_int<T>& x) { return x.value (); }
template <class T>
inline bool xis_false (const octave_int<T>& x) { return ! x.value (); }
// for reals, we want to ignore NaNs.
inline bool xis_true (double x) { return ! xisnan (x) && x != 0.0; }
inline bool xis_false (double x) { return x == 0.0; }
inline bool xis_true (float x) { return ! xisnan (x) && x != 0.0f; }
inline bool xis_false (float x) { return x == 0.0f; }
// Ditto for complex.
inline bool xis_true (const Complex& x) { return ! xisnan (x) && x != 0.0; }
inline bool xis_false (const Complex& x) { return x == 0.0; }
inline bool xis_true (const FloatComplex& x) { return ! xisnan (x) && x != 0.0f; }
inline bool xis_false (const FloatComplex& x) { return x == 0.0f; }

#define OP_RED_SUM(ac, el) ac += el
#define OP_RED_PROD(ac, el) ac *= el
#define OP_RED_SUMSQ(ac, el) ac += el*el
#define OP_RED_SUMSQC(ac, el) ac += cabsq (el)

inline void op_dble_sum(double& ac, float el)
{ ac += el; }
inline void op_dble_sum(Complex& ac, const FloatComplex& el)
{ ac += el; } // FIXME: guaranteed?
template <class T>
inline void op_dble_sum(double& ac, const octave_int<T>& el)
{ ac += el.double_value (); }

// The following two implement a simple short-circuiting.
#define OP_RED_ANYC(ac, el) if (xis_true (el)) { ac = true; break; } else continue
#define OP_RED_ALLC(ac, el) if (xis_false (el)) { ac = false; break; } else continue

#define OP_RED_FCN(F, TSRC, TRES, OP, ZERO) \
template <class T> \
inline TRES \
F (const TSRC* v, octave_idx_type n) \
{ \
  TRES ac = ZERO; \
  for (octave_idx_type i = 0; i < n; i++) \
    OP(ac, v[i]); \
  return ac; \
}

#define PROMOTE_DOUBLE(T) typename subst_template_param<std::complex, T, double>::type

OP_RED_FCN (mx_inline_sum, T, T, OP_RED_SUM, 0)
OP_RED_FCN (mx_inline_dsum, T, PROMOTE_DOUBLE(T), op_dble_sum, 0.0)
OP_RED_FCN (mx_inline_count, bool, T, OP_RED_SUM, 0)
OP_RED_FCN (mx_inline_prod, T, T, OP_RED_PROD, 1)
OP_RED_FCN (mx_inline_sumsq, T, T, OP_RED_SUMSQ, 0)
OP_RED_FCN (mx_inline_sumsq, std::complex<T>, T, OP_RED_SUMSQC, 0)
OP_RED_FCN (mx_inline_any, T, bool, OP_RED_ANYC, false)
OP_RED_FCN (mx_inline_all, T, bool, OP_RED_ALLC, true)


#define OP_RED_FCN2(F, TSRC, TRES, OP, ZERO) \
template <class T> \
inline void \
F (const TSRC* v, TRES *r, octave_idx_type m, octave_idx_type n) \
{ \
  for (octave_idx_type i = 0; i < m; i++) \
    r[i] = ZERO; \
  for (octave_idx_type j = 0; j < n; j++) \
    { \
      for (octave_idx_type i = 0; i < m; i++) \
        OP(r[i], v[i]); \
      v += m; \
    } \
}

OP_RED_FCN2 (mx_inline_sum, T, T, OP_RED_SUM, 0)
OP_RED_FCN2 (mx_inline_dsum, T, PROMOTE_DOUBLE(T), op_dble_sum, 0.0)
OP_RED_FCN2 (mx_inline_count, bool, T, OP_RED_SUM, 0)
OP_RED_FCN2 (mx_inline_prod, T, T, OP_RED_PROD, 1)
OP_RED_FCN2 (mx_inline_sumsq, T, T, OP_RED_SUMSQ, 0)
OP_RED_FCN2 (mx_inline_sumsq, std::complex<T>, T, OP_RED_SUMSQC, 0)

#define OP_RED_ANYR(ac, el) ac |= xis_true (el)
#define OP_RED_ALLR(ac, el) ac &= xis_true (el)

OP_RED_FCN2 (mx_inline_any_r, T, bool, OP_RED_ANYR, false)
OP_RED_FCN2 (mx_inline_all_r, T, bool, OP_RED_ALLR, true)

// Using the general code for any/all would sacrifice short-circuiting.
// OTOH, going by rows would sacrifice cache-coherence. The following algorithm
// will achieve both, at the cost of a temporary octave_idx_type array.

#define OP_ROW_SHORT_CIRCUIT(F, PRED, ZERO) \
template <class T> \
inline void \
F (const T* v, bool *r, octave_idx_type m, octave_idx_type n) \
{ \
  if (n <= 8) \
    return F ## _r (v, r, m, n); \
  \
  /* FIXME: it may be sub-optimal to allocate the buffer here. */ \
  OCTAVE_LOCAL_BUFFER (octave_idx_type, iact, m); \
  for (octave_idx_type i = 0; i < m; i++) iact[i] = i; \
  octave_idx_type nact = m; \
  for (octave_idx_type j = 0; j < n; j++) \
    { \
      octave_idx_type k = 0; \
      for (octave_idx_type i = 0; i < nact; i++) \
        { \
          octave_idx_type ia = iact[i]; \
          if (! PRED (v[ia])) \
            iact[k++] = ia; \
        } \
      nact = k; \
      v += m; \
    } \
  for (octave_idx_type i = 0; i < m; i++) r[i] = ! ZERO; \
  for (octave_idx_type i = 0; i < nact; i++) r[iact[i]] = ZERO; \
}

OP_ROW_SHORT_CIRCUIT (mx_inline_any, xis_true, false)
OP_ROW_SHORT_CIRCUIT (mx_inline_all, xis_false, true)

#define OP_RED_FCNN(F, TSRC, TRES) \
template <class T> \
inline void \
F (const TSRC *v, TRES *r, octave_idx_type l, \
   octave_idx_type n, octave_idx_type u) \
{ \
  if (l == 1) \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          r[i] = F<T> (v, n); \
          v += n; \
        } \
    } \
  else \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          F (v, r, l, n); \
          v += l*n; \
          r += l; \
        } \
    } \
}

OP_RED_FCNN (mx_inline_sum, T, T)
OP_RED_FCNN (mx_inline_dsum, T, PROMOTE_DOUBLE(T))
OP_RED_FCNN (mx_inline_count, bool, T)
OP_RED_FCNN (mx_inline_prod, T, T)
OP_RED_FCNN (mx_inline_sumsq, T, T)
OP_RED_FCNN (mx_inline_sumsq, std::complex<T>, T)
OP_RED_FCNN (mx_inline_any, T, bool)
OP_RED_FCNN (mx_inline_all, T, bool)

#define OP_CUM_FCN(F, TSRC, TRES, OP) \
template <class T> \
inline void \
F (const TSRC *v, TRES *r, octave_idx_type n) \
{ \
  if (n) \
    { \
      TRES t = r[0] = v[0]; \
      for (octave_idx_type i = 1; i < n; i++) \
        r[i] = t = t OP v[i]; \
    } \
}

OP_CUM_FCN (mx_inline_cumsum, T, T, +)
OP_CUM_FCN (mx_inline_cumprod, T, T, *)
OP_CUM_FCN (mx_inline_cumcount, bool, T, +)

#define OP_CUM_FCN2(F, TSRC, TRES, OP) \
template <class T> \
inline void \
F (const TSRC *v, TRES *r, octave_idx_type m, octave_idx_type n) \
{ \
  if (n) \
    { \
      for (octave_idx_type i = 0; i < m; i++) \
        r[i] = v[i]; \
      const T *r0 = r; \
      for (octave_idx_type j = 1; j < n; j++) \
        { \
          r += m; v += m; \
          for (octave_idx_type i = 0; i < m; i++) \
            r[i] = r0[i] OP v[i]; \
          r0 += m; \
        } \
    } \
}

OP_CUM_FCN2 (mx_inline_cumsum, T, T, +)
OP_CUM_FCN2 (mx_inline_cumprod, T, T, *)
OP_CUM_FCN2 (mx_inline_cumcount, bool, T, +)

#define OP_CUM_FCNN(F, TSRC, TRES) \
template <class T> \
inline void \
F (const TSRC *v, TRES *r, octave_idx_type l, \
   octave_idx_type n, octave_idx_type u) \
{ \
  if (l == 1) \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          F (v, r, n); \
          v += n; r += n; \
        } \
    } \
  else \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          F (v, r, l, n); \
          v += l*n; \
          r += l*n; \
        } \
    } \
}

OP_CUM_FCNN (mx_inline_cumsum, T, T)
OP_CUM_FCNN (mx_inline_cumprod, T, T)
OP_CUM_FCNN (mx_inline_cumcount, bool, T)

#define OP_MINMAX_FCN(F, OP) \
template <class T> \
void F (const T *v, T *r, octave_idx_type n) \
{ \
  if (! n) return; \
  T tmp = v[0]; \
  octave_idx_type i = 1; \
  if (xisnan (tmp)) \
    { \
      for (; i < n && xisnan (v[i]); i++) ; \
      if (i < n) tmp = v[i]; \
    } \
  for (; i < n; i++) \
    if (v[i] OP tmp) tmp = v[i]; \
  *r = tmp; \
} \
template <class T> \
void F (const T *v, T *r, octave_idx_type *ri, octave_idx_type n) \
{ \
  if (! n) return; \
  T tmp = v[0]; \
  octave_idx_type tmpi = 0; \
  octave_idx_type i = 1; \
  if (xisnan (tmp)) \
    { \
      for (; i < n && xisnan (v[i]); i++) ; \
      if (i < n) { tmp = v[i]; tmpi = i; } \
    } \
  for (; i < n; i++) \
    if (v[i] OP tmp) { tmp = v[i]; tmpi = i; }\
  *r = tmp; \
  *ri = tmpi; \
}

OP_MINMAX_FCN (mx_inline_min, <)
OP_MINMAX_FCN (mx_inline_max, >)

// Row reductions will be slightly complicated.  We will proceed with checks
// for NaNs until we detect that no row will yield a NaN, in which case we
// proceed to a faster code.

#define OP_MINMAX_FCN2(F, OP) \
template <class T> \
inline void \
F (const T *v, T *r, octave_idx_type m, octave_idx_type n) \
{ \
  if (! n) return; \
  bool nan = false; \
  octave_idx_type j = 0; \
  for (octave_idx_type i = 0; i < m; i++) \
    {  \
      r[i] = v[i]; \
      if (xisnan (v[i])) nan = true;  \
    } \
  j++; v += m; \
  while (nan && j < n) \
    { \
      nan = false; \
      for (octave_idx_type i = 0; i < m; i++) \
        {  \
          if (xisnan (v[i])) \
            nan = true;  \
          else if (xisnan (r[i]) || v[i] OP r[i]) \
            r[i] = v[i]; \
        } \
      j++; v += m; \
    } \
  while (j < n) \
    { \
      for (octave_idx_type i = 0; i < m; i++) \
        if (v[i] OP r[i]) r[i] = v[i]; \
      j++; v += m; \
    } \
} \
template <class T> \
inline void \
F (const T *v, T *r, octave_idx_type *ri, \
   octave_idx_type m, octave_idx_type n) \
{ \
  if (! n) return; \
  bool nan = false; \
  octave_idx_type j = 0; \
  for (octave_idx_type i = 0; i < m; i++) \
    {  \
      r[i] = v[i]; ri[i] = j; \
      if (xisnan (v[i])) nan = true;  \
    } \
  j++; v += m; \
  while (nan && j < n) \
    { \
      nan = false; \
      for (octave_idx_type i = 0; i < m; i++) \
        {  \
          if (xisnan (v[i])) \
            nan = true;  \
          else if (xisnan (r[i]) || v[i] OP r[i]) \
            { r[i] = v[i]; ri[i] = j; } \
        } \
      j++; v += m; \
    } \
  while (j < n) \
    { \
      for (octave_idx_type i = 0; i < m; i++) \
        if (v[i] OP r[i]) \
          { r[i] = v[i]; ri[i] = j; } \
      j++; v += m; \
    } \
}

OP_MINMAX_FCN2 (mx_inline_min, <)
OP_MINMAX_FCN2 (mx_inline_max, >)

#define OP_MINMAX_FCNN(F) \
template <class T> \
inline void \
F (const T *v, T *r, octave_idx_type l, \
   octave_idx_type n, octave_idx_type u) \
{ \
  if (! n) return; \
  if (l == 1) \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          F (v, r, n); \
          v += n; r++; \
        } \
    } \
  else \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          F (v, r, l, n); \
          v += l*n; \
          r += l; \
        } \
    } \
} \
template <class T> \
inline void \
F (const T *v, T *r, octave_idx_type *ri, \
   octave_idx_type l, octave_idx_type n, octave_idx_type u) \
{ \
  if (! n) return; \
  if (l == 1) \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          F (v, r, ri, n); \
          v += n; r++; ri++; \
        } \
    } \
  else \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          F (v, r, ri, l, n); \
          v += l*n; \
          r += l; ri += l; \
        } \
    } \
}

OP_MINMAX_FCNN (mx_inline_min)
OP_MINMAX_FCNN (mx_inline_max)

#define OP_CUMMINMAX_FCN(F, OP) \
template <class T> \
void F (const T *v, T *r, octave_idx_type n) \
{ \
  if (! n) return; \
  T tmp = v[0]; \
  octave_idx_type i = 1, j = 0; \
  if (xisnan (tmp)) \
    { \
      for (; i < n && xisnan (v[i]); i++) ; \
      for (; j < i; j++) r[j] = tmp; \
      if (i < n) tmp = v[i]; \
    } \
  for (; i < n; i++) \
    if (v[i] OP tmp) \
      { \
        for (; j < i; j++) r[j] = tmp; \
        tmp = v[i]; \
      } \
  for (; j < i; j++) r[j] = tmp; \
} \
template <class T> \
void F (const T *v, T *r, octave_idx_type *ri, octave_idx_type n) \
{ \
  if (! n) return; \
  T tmp = v[0]; octave_idx_type tmpi = 0; \
  octave_idx_type i = 1, j = 0; \
  if (xisnan (tmp)) \
    { \
      for (; i < n && xisnan (v[i]); i++) ; \
      for (; j < i; j++) { r[j] = tmp; ri[j] = tmpi; } \
      if (i < n) { tmp = v[i]; tmpi = i; } \
    } \
  for (; i < n; i++) \
    if (v[i] OP tmp) \
      { \
        for (; j < i; j++) { r[j] = tmp; ri[j] = tmpi; } \
        tmp = v[i]; tmpi = i; \
      } \
  for (; j < i; j++) { r[j] = tmp; ri[j] = tmpi; } \
}

OP_CUMMINMAX_FCN (mx_inline_cummin, <)
OP_CUMMINMAX_FCN (mx_inline_cummax, >)

// Row reductions will be slightly complicated.  We will proceed with checks
// for NaNs until we detect that no row will yield a NaN, in which case we
// proceed to a faster code.

#define OP_CUMMINMAX_FCN2(F, OP) \
template <class T> \
inline void \
F (const T *v, T *r, octave_idx_type m, octave_idx_type n) \
{ \
  if (! n) return; \
  bool nan = false; \
  const T *r0; \
  octave_idx_type j = 0; \
  for (octave_idx_type i = 0; i < m; i++) \
    {  \
      r[i] = v[i]; \
      if (xisnan (v[i])) nan = true;  \
    } \
  j++; v += m; r0 = r; r += m; \
  while (nan && j < n) \
    { \
      nan = false; \
      for (octave_idx_type i = 0; i < m; i++) \
        {  \
          if (xisnan (v[i])) \
            { r[i] = r0[i]; nan = true; } \
          else if (xisnan (r0[i]) || v[i] OP r0[i]) \
            r[i] = v[i]; \
        } \
      j++; v += m; r0 = r; r += m; \
    } \
  while (j < n) \
    { \
      for (octave_idx_type i = 0; i < m; i++) \
        if (v[i] OP r0[i]) \
          r[i] = v[i]; \
        else \
          r[i] = r0[i]; \
      j++; v += m; r0 = r; r += m; \
    } \
} \
template <class T> \
inline void \
F (const T *v, T *r, octave_idx_type *ri, \
   octave_idx_type m, octave_idx_type n) \
{ \
  if (! n) return; \
  bool nan = false; \
  const T *r0; const octave_idx_type *r0i; \
  octave_idx_type j = 0; \
  for (octave_idx_type i = 0; i < m; i++) \
    {  \
      r[i] = v[i]; ri[i] = 0; \
      if (xisnan (v[i])) nan = true;  \
    } \
  j++; v += m; r0 = r; r += m; r0i = ri; ri += m;  \
  while (nan && j < n) \
    { \
      nan = false; \
      for (octave_idx_type i = 0; i < m; i++) \
        {  \
          if (xisnan (v[i])) \
            { r[i] = r0[i]; ri[i] = r0i[i]; nan = true; } \
          else if (xisnan (r0[i]) || v[i] OP r0[i]) \
            { r[i] = v[i]; ri[i] = j; }\
        } \
      j++; v += m; r0 = r; r += m; r0i = ri; ri += m;  \
    } \
  while (j < n) \
    { \
      for (octave_idx_type i = 0; i < m; i++) \
        if (v[i] OP r0[i]) \
          { r[i] = v[i]; ri[i] = j; } \
        else \
          { r[i] = r0[i]; ri[i] = r0i[i]; } \
      j++; v += m; r0 = r; r += m; r0i = ri; ri += m;  \
    } \
}

OP_CUMMINMAX_FCN2 (mx_inline_cummin, <)
OP_CUMMINMAX_FCN2 (mx_inline_cummax, >)

#define OP_CUMMINMAX_FCNN(F) \
template <class T> \
inline void \
F (const T *v, T *r, octave_idx_type l, \
   octave_idx_type n, octave_idx_type u) \
{ \
  if (! n) return; \
  if (l == 1) \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          F (v, r, n); \
          v += n; r += n; \
        } \
    } \
  else \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          F (v, r, l, n); \
          v += l*n; \
          r += l*n; \
        } \
    } \
} \
template <class T> \
inline void \
F (const T *v, T *r, octave_idx_type *ri, \
   octave_idx_type l, octave_idx_type n, octave_idx_type u) \
{ \
  if (! n) return; \
  if (l == 1) \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          F (v, r, ri, n); \
          v += n; r += n; ri += n; \
        } \
    } \
  else \
    { \
      for (octave_idx_type i = 0; i < u; i++) \
        { \
          F (v, r, ri, l, n); \
          v += l*n; \
          r += l*n; ri += l*n; \
        } \
    } \
}

OP_CUMMINMAX_FCNN (mx_inline_cummin)
OP_CUMMINMAX_FCNN (mx_inline_cummax)

template <class T>
void mx_inline_diff (const T *v, T *r, octave_idx_type n,
                     octave_idx_type order)
{
  switch (order)
    {
    case 1:
      for (octave_idx_type i = 0; i < n-1; i++)
        r[i] = v[i+1] - v[i];
      break;
    case 2:
      if (n > 1)
        {
          T lst = v[1] - v[0];
          for (octave_idx_type i = 0; i < n-2; i++)
            {
              T dif = v[i+2] - v[i+1];
              r[i] = dif - lst;
              lst = dif;
            }
        }
      break;
    default:
        {
          OCTAVE_LOCAL_BUFFER (T, buf, n-1);

          for (octave_idx_type i = 0; i < n-1; i++)
            buf[i] = v[i+1] - v[i];

          for (octave_idx_type o = 2; o <= order; o++)
            {
              for (octave_idx_type i = 0; i < n-o; i++)
                buf[i] = buf[i+1] - buf[i];
            }

          for (octave_idx_type i = 0; i < n-order; i++)
            r[i] = buf[i];
        }
    }
}

template <class T>
void mx_inline_diff (const T *v, T *r,
                     octave_idx_type m, octave_idx_type n,
                     octave_idx_type order)
{
  switch (order)
    {
    case 1:
      for (octave_idx_type i = 0; i < m*(n-1); i++)
        r[i] = v[i+m] - v[i];
      break;
    case 2:
      for (octave_idx_type i = 0; i < n-2; i++)
        {
          for (octave_idx_type j = i*m; j < i*m+m; j++)
            r[j] = (v[j+m+m] - v[j+m]) - (v[j+m] - v[j]);
        }
      break;
    default:
        {
          OCTAVE_LOCAL_BUFFER (T, buf, n-1);

          for (octave_idx_type j = 0; j < m; j++)
            {
              for (octave_idx_type i = 0; i < n-1; i++)
                buf[i] = v[i*m+j+m] - v[i*m+j];

              for (octave_idx_type o = 2; o <= order; o++)
                {
                  for (octave_idx_type i = 0; i < n-o; i++)
                    buf[i] = buf[i+1] - buf[i];
                }

              for (octave_idx_type i = 0; i < n-order; i++)
                r[i*m+j] = buf[i];
            }
        }
    }
}

template <class T>
inline void
mx_inline_diff (const T *v, T *r,
                octave_idx_type l, octave_idx_type n, octave_idx_type u,
                octave_idx_type order)
{
  if (! n) return;
  if (l == 1)
    {
      for (octave_idx_type i = 0; i < u; i++)
        {
          mx_inline_diff (v, r, n, order);
          v += n; r += n-order;
        }
    }
  else
    {
      for (octave_idx_type i = 0; i < u; i++)
        {
          mx_inline_diff (v, r, l, n, order);
          v += l*n;
          r += l*(n-order);
        }
    }
}

// Assistant function

inline void
get_extent_triplet (const dim_vector& dims, int& dim,
                    octave_idx_type& l, octave_idx_type& n,
                    octave_idx_type& u)
{
  octave_idx_type ndims = dims.length ();
  if (dim >= ndims)
    {
      l = dims.numel ();
      n = 1;
      u = 1;
    }
  else
    {
      if (dim < 0)
        dim = dims.first_non_singleton ();

      // calculate extent triplet.
      l = 1, n = dims(dim), u = 1;
      for (octave_idx_type i = 0; i < dim; i++)
        l *= dims (i);
      for (octave_idx_type i = dim + 1; i < ndims; i++)
        u *= dims (i);
    }
}

// Appliers.
// FIXME: is this the best design? C++ gives a lot of options here...
// maybe it can be done without an explicit parameter?

template <class R, class T>
inline Array<R>
do_mx_red_op (const Array<T>& src, int dim,
              void (*mx_red_op) (const T *, R *, octave_idx_type,
                                 octave_idx_type, octave_idx_type))
{
  octave_idx_type l, n, u;
  dim_vector dims = src.dims ();
  // M*b inconsistency: sum([]) = 0 etc.
  if (dims.length () == 2 && dims(0) == 0 && dims(1) == 0)
    dims (1) = 1;

  get_extent_triplet (dims, dim, l, n, u);

  // Reduction operation reduces the array size.
  if (dim < dims.length ()) dims(dim) = 1;
  dims.chop_trailing_singletons ();

  Array<R> ret (dims);
  mx_red_op (src.data (), ret.fortran_vec (), l, n, u);

  return ret;
}

template <class R, class T>
inline Array<R>
do_mx_cum_op (const Array<T>& src, int dim,
              void (*mx_cum_op) (const T *, R *, octave_idx_type,
                                 octave_idx_type, octave_idx_type))
{
  octave_idx_type l, n, u;
  dim_vector dims = src.dims ();
  get_extent_triplet (dims, dim, l, n, u);

  // Cumulative operation doesn't reduce the array size.
  Array<R> ret (dims);
  mx_cum_op (src.data (), ret.fortran_vec (), l, n, u);

  return ret;
}

template <class R>
inline Array<R>
do_mx_minmax_op (const Array<R>& src, int dim,
                 void (*mx_minmax_op) (const R *, R *, octave_idx_type,
                                       octave_idx_type, octave_idx_type))
{
  octave_idx_type l, n, u;
  dim_vector dims = src.dims ();
  get_extent_triplet (dims, dim, l, n, u);

  // If the dimension is zero, we don't do anything.
  if (dim < dims.length () && dims(dim) != 0) dims(dim) = 1;
  dims.chop_trailing_singletons ();

  Array<R> ret (dims);
  mx_minmax_op (src.data (), ret.fortran_vec (), l, n, u);

  return ret;
}

template <class R>
inline Array<R>
do_mx_minmax_op (const Array<R>& src, Array<octave_idx_type>& idx, int dim,
                 void (*mx_minmax_op) (const R *, R *, octave_idx_type *,
                                       octave_idx_type, octave_idx_type, octave_idx_type))
{
  octave_idx_type l, n, u;
  dim_vector dims = src.dims ();
  get_extent_triplet (dims, dim, l, n, u);

  // If the dimension is zero, we don't do anything.
  if (dim < dims.length () && dims(dim) != 0) dims(dim) = 1;
  dims.chop_trailing_singletons ();

  Array<R> ret (dims);
  if (idx.dims () != dims) idx = Array<octave_idx_type> (dims);

  mx_minmax_op (src.data (), ret.fortran_vec (), idx.fortran_vec (),
                l, n, u);

  return ret;
}

template <class R>
inline Array<R>
do_mx_cumminmax_op (const Array<R>& src, int dim,
                    void (*mx_cumminmax_op) (const R *, R *, octave_idx_type,
                                             octave_idx_type, octave_idx_type))
{
  octave_idx_type l, n, u;
  dim_vector dims = src.dims ();
  get_extent_triplet (dims, dim, l, n, u);

  Array<R> ret (dims);
  mx_cumminmax_op (src.data (), ret.fortran_vec (), l, n, u);

  return ret;
}

template <class R>
inline Array<R>
do_mx_cumminmax_op (const Array<R>& src, Array<octave_idx_type>& idx, int dim,
                    void (*mx_cumminmax_op) (const R *, R *, octave_idx_type *,
                                             octave_idx_type, octave_idx_type, octave_idx_type))
{
  octave_idx_type l, n, u;
  dim_vector dims = src.dims ();
  get_extent_triplet (dims, dim, l, n, u);

  Array<R> ret (dims);
  if (idx.dims () != dims) idx = Array<octave_idx_type> (dims);

  mx_cumminmax_op (src.data (), ret.fortran_vec (), idx.fortran_vec (),
                   l, n, u);

  return ret;
}

template <class R>
inline Array<R>
do_mx_diff_op (const Array<R>& src, int dim, octave_idx_type order,
               void (*mx_diff_op) (const R *, R *,
                                   octave_idx_type, octave_idx_type, octave_idx_type,
                                   octave_idx_type))
{
  octave_idx_type l, n, u;
  if (order <= 0)
    return src;

  dim_vector dims = src.dims ();

  get_extent_triplet (dims, dim, l, n, u);
  if (dim >= dims.length ())
    dims.resize (dim+1, 1);

  if (dims(dim) <= order)
    {
      dims (dim) = 0;
      return Array<R> (dims);
    }
  else
    {
      dims(dim) -= order;
    }

  Array<R> ret (dims);
  mx_diff_op (src.data (), ret.fortran_vec (), l, n, u, order);

  return ret;
}

// Fast extra-precise summation. According to
// T. Ogita, S. M. Rump, S. Oishi:
// Accurate Sum And Dot Product,
// SIAM J. Sci. Computing, Vol. 26, 2005

template <class T>
inline void twosum_accum (T& s, T& e,
                          const T& x)
{
  T s1 = s + x, t = s1 - s, e1 = (s - (s1 - t)) + (x - t);
  s = s1;
  e += e1;
}

template <class T>
inline T
mx_inline_xsum (const T *v, octave_idx_type n)
{
  T s = 0, e = 0;
  for (octave_idx_type i = 0; i < n; i++)
    twosum_accum (s, e, v[i]);

  return s + e;
}

template <class T>
inline void
mx_inline_xsum (const T *v, T *r,
                octave_idx_type m, octave_idx_type n)
{
  OCTAVE_LOCAL_BUFFER (T, e, m);
  for (octave_idx_type i = 0; i < m; i++)
    e[i] = r[i] = T ();

  for (octave_idx_type j = 0; j < n; j++)
    {
      for (octave_idx_type i = 0; i < m; i++)
        twosum_accum (r[i], e[i], v[i]);

      v += m;
    }

  for (octave_idx_type i = 0; i < m; i++)
    r[i] += e[i];
}

OP_RED_FCNN (mx_inline_xsum, T, T)

#endif