octave-nkf: liboctave/mx-inlines.cc annotate

annotate liboctave/mx-inlines.cc @ 9827:c15a5ed0da58

optimize bsxfun (@power, ...)

author	Jaroslav Hajek <highegg@gmail.com>
date	Wed, 18 Nov 2009 12:03:07 +0100
parents	2b29f3472e20
children	9597eea7fa36

rev	line source
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	1 /*
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	2
7017 a1dbe9d80eee [project @ 2007-10-12 21:27:11 by jwe] jwe parents: 7016 diff changeset	3 Copyright (C) 1993, 1994, 1995, 1996, 1997, 1999, 2000, 2001, 2002,
8920 eb63fbe60fab update copyright notices John W. Eaton <jwe@octave.org> parents: 8899 diff changeset	4 2003, 2004, 2005, 2006, 2007, 2008 John W. Eaton
8777 724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	5 Copyright (C) 2009 Jaroslav Hajek
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	6 Copyright (C) 2009 VZLU Prague
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	7
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	8 This file is part of Octave.
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	9
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	10 Octave is free software; you can redistribute it and/or modify it
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	11 under the terms of the GNU General Public License as published by the
7016 93c65f2a5668 [project @ 2007-10-12 06:40:56 by jwe] jwe parents: 5972 diff changeset	12 Free Software Foundation; either version 3 of the License, or (at your
93c65f2a5668 [project @ 2007-10-12 06:40:56 by jwe] jwe parents: 5972 diff changeset	13 option) any later version.
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	14
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	15 Octave is distributed in the hope that it will be useful, but WITHOUT
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	16 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	17 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	18 for more details.
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	19
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	20 You should have received a copy of the GNU General Public License
7016 93c65f2a5668 [project @ 2007-10-12 06:40:56 by jwe] jwe parents: 5972 diff changeset	21 along with Octave; see the file COPYING. If not, see
93c65f2a5668 [project @ 2007-10-12 06:40:56 by jwe] jwe parents: 5972 diff changeset	22 <http://www.gnu.org/licenses/>.
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	23
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	24 */
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	25
2828 92826d6e8bd9 [project @ 1997-03-25 23:41:41 by jwe] jwe parents: 2811 diff changeset	26 #if !defined (octave_mx_inlines_h)
92826d6e8bd9 [project @ 1997-03-25 23:41:41 by jwe] jwe parents: 2811 diff changeset	27 #define octave_mx_inlines_h 1
2804 eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	28
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	29 #include <cstddef>
8650 a1ae2aae903e abs,real,imag,conj: use code from mx-inlines rather than the generic map Jaroslav Hajek <highegg@gmail.com> parents: 8380 diff changeset	30 #include <cmath>
2804 eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	31
5525 558c69736dc0 [project @ 2005-11-01 00:49:50 by jwe] jwe parents: 5524 diff changeset	32 #include "quit.h"
558c69736dc0 [project @ 2005-11-01 00:49:50 by jwe] jwe parents: 5524 diff changeset	33
1650 23aa282707e8 [project @ 1995-12-20 06:53:12 by jwe] jwe parents: 1574 diff changeset	34 #include "oct-cmplx.h"
8758 83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	35 #include "oct-locbuf.h"
8897 fd83cc87f55e add missing include Jaroslav Hajek <highegg@gmail.com> parents: 8780 diff changeset	36 #include "oct-inttypes.h"
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	37 #include "Array-util.h"
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	38
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	39 // Provides some commonly repeated, basic loop templates.
461 00f8b2242a18 [project @ 1994-06-06 00:45:57 by jwe] jwe parents: 238 diff changeset	40
8380 dbe67764e628 fix & improve speed of diagonal matrix multiplication Jaroslav Hajek <highegg@gmail.com> parents: 7789 diff changeset	41 template <class R, class S>
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	42 inline void mx_inline_fill (size_t n, R *r, S s)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	43 { for (size_t i = 0; i < n; i++) r[i] = s; }
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	44
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	45 #define DEFMXUNOP(F, OP) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	46 template <class R, class X> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	47 inline void F (size_t n, R r, const X x) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	48 { for (size_t i = 0; i < n; i++) r[i] = OP x[i]; }
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	49
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	50 DEFMXUNOP (mx_inline_uminus, -)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	51
9607 1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	52 #define DEFMXUNOPEQ(F, OP) \
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	53 template <class R> \
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	54 inline void F (size_t n, R *r) \
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	55 { for (size_t i = 0; i < n; i++) r[i] = OP r[i]; }
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	56
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	57 DEFMXUNOPEQ (mx_inline_uminus2, -)
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	58
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	59 #define DEFMXUNBOOLOP(F, OP) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	60 template <class X> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	61 inline void F (size_t n, bool r, const X x) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	62 { const X zero = X(); for (size_t i = 0; i < n; i++) r[i] = x[i] OP zero; }
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	63
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	64 DEFMXUNBOOLOP (mx_inline_iszero, ==)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	65 DEFMXUNBOOLOP (mx_inline_notzero, !=)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	66
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	67 #define DEFMXBINOP(F, OP) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	68 template <class R, class X, class Y> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	69 inline void F (size_t n, R r, const X x, const Y *y) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	70 { for (size_t i = 0; i < n; i++) r[i] = x[i] OP y[i]; } \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	71 template <class R, class X, class Y> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	72 inline void F (size_t n, R r, const X x, Y y) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	73 { for (size_t i = 0; i < n; i++) r[i] = x[i] OP y; } \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	74 template <class R, class X, class Y> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	75 inline void F (size_t n, R r, X x, const Y y) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	76 { for (size_t i = 0; i < n; i++) r[i] = x OP y[i]; }
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	77
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	78 DEFMXBINOP (mx_inline_add, +)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	79 DEFMXBINOP (mx_inline_sub, -)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	80 DEFMXBINOP (mx_inline_mul, *)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	81 DEFMXBINOP (mx_inline_div, /)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	82
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	83 #define DEFMXBINOPEQ(F, OP) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	84 template <class R, class X> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	85 inline void F (size_t n, R r, const X x) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	86 { for (size_t i = 0; i < n; i++) r[i] OP x[i]; } \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	87 template <class R, class X> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	88 inline void F (size_t n, R *r, X x) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	89 { for (size_t i = 0; i < n; i++) r[i] OP x; }
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	90
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	91 DEFMXBINOPEQ (mx_inline_add2, +=)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	92 DEFMXBINOPEQ (mx_inline_sub2, -=)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	93 DEFMXBINOPEQ (mx_inline_mul2, *=)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	94 DEFMXBINOPEQ (mx_inline_div2, /=)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	95
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	96 #define DEFMXCMPOP(F, OP) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	97 template <class X, class Y> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	98 inline void F (size_t n, bool r, const X x, const Y *y) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	99 { for (size_t i = 0; i < n; i++) r[i] = x[i] OP y[i]; } \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	100 template <class X, class Y> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	101 inline void F (size_t n, bool r, const X x, Y y) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	102 { for (size_t i = 0; i < n; i++) r[i] = x[i] OP y; } \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	103 template <class X, class Y> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	104 inline void F (size_t n, bool r, X x, const Y y) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	105 { for (size_t i = 0; i < n; i++) r[i] = x OP y[i]; }
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	106
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	107 DEFMXCMPOP (mx_inline_lt, <)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	108 DEFMXCMPOP (mx_inline_le, <=)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	109 DEFMXCMPOP (mx_inline_gt, >)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	110 DEFMXCMPOP (mx_inline_ge, >=)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	111 DEFMXCMPOP (mx_inline_eq, ==)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	112 DEFMXCMPOP (mx_inline_ne, !=)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	113
9553 0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	114 // Convert to logical value, for logical op purposes.
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	115 template <class T> inline bool logical_value (T x) { return x; }
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	116 template <class T> inline bool logical_value (const octave_int<T>& x)
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	117 { return x.value (); }
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	118
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	119 // NaNs in real data should generate an error. Doing it on-the-fly is faster.
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	120
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	121 #define DEFLOGCHKNAN(ARG, ZERO) \
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	122 inline bool logical_value (ARG x) \
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	123 { if (xisnan (x)) gripe_nan_to_logical_conversion (); return x != ZERO; }
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	124
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	125 DEFLOGCHKNAN (double, 0.0)
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	126 DEFLOGCHKNAN (const Complex&, 0.0)
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	127 DEFLOGCHKNAN (float, 0.0f)
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	128 DEFLOGCHKNAN (const FloatComplex&, 0.0f)
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	129
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	130 template <class X>
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	131 void mx_inline_not (size_t n, bool r, const X x)
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	132 {
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	133 for (size_t i = 0; i < n; i++)
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	134 r[i] = ! logical_value (x[i]);
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	135 }
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	136
9607 1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	137 inline void mx_inline_not2 (size_t n, bool *r)
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	138 {
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	139 for (size_t i = 0; i < n; i++) r[i] = ! r[i];
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	140 }
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	141
9553 0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	142 #define DEFMXBOOLOP(F, NOT1, OP, NOT2) \
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	143 template <class X, class Y> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	144 inline void F (size_t n, bool r, const X x, const Y *y) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	145 { \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	146 for (size_t i = 0; i < n; i++) \
9553 0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	147 r[i] = (NOT1 logical_value (x[i])) OP (NOT2 logical_value (y[i])); \
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	148 } \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	149 template <class X, class Y> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	150 inline void F (size_t n, bool r, const X x, Y y) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	151 { \
9553 0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	152 const bool yy = (NOT2 logical_value (y)); \
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	153 for (size_t i = 0; i < n; i++) \
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	154 r[i] = (NOT1 logical_value (x[i])) OP yy; \
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	155 } \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	156 template <class X, class Y> \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	157 inline void F (size_t n, bool r, X x, const Y y) \
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	158 { \
9553 0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	159 const bool xx = (NOT1 logical_value (x)); \
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	160 for (size_t i = 0; i < n; i++) \
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	161 r[i] = xx OP (NOT2 logical_value (y[i])); \
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	162 }
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	163
9553 0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	164 DEFMXBOOLOP (mx_inline_and, , &, )
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	165 DEFMXBOOLOP (mx_inline_or, , \|, )
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	166 DEFMXBOOLOP (mx_inline_not_and, !, &, )
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	167 DEFMXBOOLOP (mx_inline_not_or, !, \|, )
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	168 DEFMXBOOLOP (mx_inline_and_not, , &, !)
0c72d9284087 further bool ops tweaks Jaroslav Hajek <highegg@gmail.com> parents: 9550 diff changeset	169 DEFMXBOOLOP (mx_inline_or_not, , \|, !)
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	170
9612 66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	171 #define DEFMXBOOLOPEQ(F, OP) \
66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	172 template <class X> \
66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	173 inline void F (size_t n, bool r, const X x) \
66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	174 { \
66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	175 for (size_t i = 0; i < n; i++) \
66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	176 r[i] OP logical_value (x[i]); \
66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	177 } \
66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	178
66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	179 DEFMXBOOLOPEQ (mx_inline_and2, &=)
66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	180 DEFMXBOOLOPEQ (mx_inline_or2, \|=)
66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	181
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	182 template <class T>
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	183 inline bool
9814 2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	184 mx_inline_any_nan (size_t n, const T* x)
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	185 {
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	186 for (size_t i = 0; i < n; i++)
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	187 {
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	188 if (xisnan (x[i]))
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	189 return true;
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	190 }
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	191
9814 2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	192 return false;
8380 dbe67764e628 fix & improve speed of diagonal matrix multiplication Jaroslav Hajek <highegg@gmail.com> parents: 7789 diff changeset	193 }
dbe67764e628 fix & improve speed of diagonal matrix multiplication Jaroslav Hajek <highegg@gmail.com> parents: 7789 diff changeset	194
9827 c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	195 template <class T>
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	196 inline bool
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	197 mx_inline_any_negative (size_t n, const T* x)
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	198 {
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	199 for (size_t i = 0; i < n; i++)
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	200 {
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	201 if (x[i] < 0)
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	202 return true;
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	203 }
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	204
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	205 return false;
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	206 }
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	207
9814 2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	208 template<class T>
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	209 inline bool
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	210 mx_inline_all_real (size_t n, const std::complex<T>* x)
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	211 {
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	212 for (size_t i = 0; i < n; i++)
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	213 {
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	214 if (x[i].imag () != 0)
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	215 return false;
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	216 }
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	217
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	218 return true;
2b29f3472e20 add a couple of useful loops Jaroslav Hajek <highegg@gmail.com> parents: 9800 diff changeset	219 }
2811 1dd37f97364a [project @ 1997-03-12 23:06:13 by jwe] jwe parents: 2804 diff changeset	220
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	221 #define DEFMXMAPPER(F, FUN) \
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	222 template <class T> \
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	223 inline void F (size_t n, T r, const T x) \
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	224 { for (size_t i = 0; i < n; i++) r[i] = FUN (x[i]); }
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	225
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	226 template<class T>
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	227 inline void mx_inline_real (size_t n, T r, const std::complex<T> x)
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	228 { for (size_t i = 0; i < n; i++) r[i] = x[i].real (); }
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	229 template<class T>
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	230 inline void mx_inline_imag (size_t n, T r, const std::complex<T> x)
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	231 { for (size_t i = 0; i < n; i++) r[i] = x[i].imag (); }
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	232
9743 26abff55f6fe optimize bsxfun for common built-in operations Jaroslav Hajek <highegg@gmail.com> parents: 9721 diff changeset	233 // Pairwise minimums/maximums
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	234 #define DEFMXMAPPER2(F, FUN) \
9743 26abff55f6fe optimize bsxfun for common built-in operations Jaroslav Hajek <highegg@gmail.com> parents: 9721 diff changeset	235 template <class T> \
26abff55f6fe optimize bsxfun for common built-in operations Jaroslav Hajek <highegg@gmail.com> parents: 9721 diff changeset	236 inline void F (size_t n, T r, const T x, const T *y) \
26abff55f6fe optimize bsxfun for common built-in operations Jaroslav Hajek <highegg@gmail.com> parents: 9721 diff changeset	237 { for (size_t i = 0; i < n; i++) r[i] = FUN (x[i], y[i]); } \
26abff55f6fe optimize bsxfun for common built-in operations Jaroslav Hajek <highegg@gmail.com> parents: 9721 diff changeset	238 template <class T> \
26abff55f6fe optimize bsxfun for common built-in operations Jaroslav Hajek <highegg@gmail.com> parents: 9721 diff changeset	239 inline void F (size_t n, T r, const T x, T y) \
26abff55f6fe optimize bsxfun for common built-in operations Jaroslav Hajek <highegg@gmail.com> parents: 9721 diff changeset	240 { for (size_t i = 0; i < n; i++) r[i] = FUN (x[i], y); } \
26abff55f6fe optimize bsxfun for common built-in operations Jaroslav Hajek <highegg@gmail.com> parents: 9721 diff changeset	241 template <class T> \
26abff55f6fe optimize bsxfun for common built-in operations Jaroslav Hajek <highegg@gmail.com> parents: 9721 diff changeset	242 inline void F (size_t n, T r, T x, const T y) \
26abff55f6fe optimize bsxfun for common built-in operations Jaroslav Hajek <highegg@gmail.com> parents: 9721 diff changeset	243 { for (size_t i = 0; i < n; i++) r[i] = FUN (x, y[i]); }
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	244
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	245 DEFMXMAPPER2 (mx_inline_xmin, xmin)
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	246 DEFMXMAPPER2 (mx_inline_xmax, xmax)
2811 1dd37f97364a [project @ 1997-03-12 23:06:13 by jwe] jwe parents: 2804 diff changeset	247
9827 c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	248 // Pairwise power
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	249 #define DEFMXMAPPER2X(F, FUN) \
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	250 template <class R, class X, class Y> \
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	251 inline void F (size_t n, R r, const X x, const Y *y) \
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	252 { for (size_t i = 0; i < n; i++) r[i] = FUN (x[i], y[i]); } \
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	253 template <class R, class X, class Y> \
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	254 inline void F (size_t n, R r, const X x, Y y) \
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	255 { for (size_t i = 0; i < n; i++) r[i] = FUN (x[i], y); } \
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	256 template <class R, class X, class Y> \
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	257 inline void F (size_t n, R r, X x, const Y y) \
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	258 { for (size_t i = 0; i < n; i++) r[i] = FUN (x, y[i]); }
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	259
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	260 DEFMXMAPPER2X (mx_inline_pow, std::pow)
c15a5ed0da58 optimize bsxfun (@power, ...) Jaroslav Hajek <highegg@gmail.com> parents: 9814 diff changeset	261
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	262 // Arbitrary function appliers. The function is a template parameter to enable
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	263 // inlining.
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	264 template <class R, class X, R fun (X x)>
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	265 inline void mx_inline_map (size_t n, R r, const X x)
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	266 { for (size_t i = 0; i < n; i++) r[i] = fun (x[i]); }
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	267
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	268 template <class R, class X, R fun (const X& x)>
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	269 inline void mx_inline_map (size_t n, R r, const X x)
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	270 { for (size_t i = 0; i < n; i++) r[i] = fun (x[i]); }
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	271
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	272 // Appliers. Since these call the operation just once, we pass it as
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	273 // a pointer, to allow the compiler reduce number of instances.
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	274
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	275 #define AELEMT(ARRAY) typename ARRAY::element_type
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	276 template <class RNDA, class XNDA>
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	277 inline RNDA
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	278 do_mx_unary_op (const XNDA& x,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	279 void (op) (size_t, AELEMT(RNDA) ,
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	280 const AELEMT(XNDA) *))
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	281 {
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	282 RNDA r (x.dims ());
9612 66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	283 op (r.length (), r.fortran_vec (), x.data ());
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	284 return r;
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	285 }
2811 1dd37f97364a [project @ 1997-03-12 23:06:13 by jwe] jwe parents: 2804 diff changeset	286
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	287 // Shortcuts for applying mx_inline_map.
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	288
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	289 template <class RNDA, class XNDA, AELEMT(RNDA) fun (AELEMT(XNDA))>
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	290 inline RNDA
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	291 do_mx_unary_map (const XNDA& x)
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	292 {
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	293 return do_mx_unary_op<RNDA, XNDA> (x, mx_inline_map<AELEMT(RNDA), AELEMT(XNDA), fun>);
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	294 }
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	295
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	296 template <class RNDA, class XNDA, AELEMT(RNDA) fun (const AELEMT(XNDA)&)>
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	297 inline RNDA
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	298 do_mx_unary_map (const XNDA& x)
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	299 {
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	300 return do_mx_unary_op<RNDA, XNDA> (x, mx_inline_map<AELEMT(RNDA), AELEMT(XNDA), fun>);
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	301 }
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	302
9607 1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	303 template <class RNDA>
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	304 inline RNDA&
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	305 do_mx_inplace_op (RNDA& r,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	306 void (op) (size_t, AELEMT(RNDA) ))
9607 1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	307 {
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	308 op (r.numel (), r.fortran_vec ());
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	309 return r;
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	310 }
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	311
1be3c73ed7b5 reuse temporary arrays in nested expressions Jaroslav Hajek <highegg@gmail.com> parents: 9578 diff changeset	312
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	313 template <class RNDA, class XNDA, class YNDA>
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	314 inline RNDA
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	315 do_mm_binary_op (const XNDA& x, const YNDA& y,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	316 void (op) (size_t, AELEMT(RNDA) ,
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	317 const AELEMT(XNDA) *,
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	318 const AELEMT(YNDA) *),
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	319 const char *opname)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	320 {
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	321 dim_vector dx = x.dims (), dy = y.dims ();
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	322 if (dx == dy)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	323 {
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	324 RNDA r (dx);
9612 66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	325 op (r.length (), r.fortran_vec (), x.data (), y.data ());
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	326 return r;
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	327 }
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	328 else
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	329 {
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	330 gripe_nonconformant (opname, dx, dy);
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	331 return RNDA ();
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	332 }
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	333 }
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	334
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	335 template <class RNDA, class XNDA, class YS>
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	336 inline RNDA
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	337 do_ms_binary_op (const XNDA& x, const YS& y,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	338 void (op) (size_t, AELEMT(RNDA) ,
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	339 const AELEMT(XNDA) *, YS))
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	340 {
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	341 RNDA r (x.dims ());
9612 66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	342 op (r.length (), r.fortran_vec (), x.data (), y);
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	343 return r;
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	344 }
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	345
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	346 template <class RNDA, class XS, class YNDA>
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	347 inline RNDA
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	348 do_sm_binary_op (const XS& x, const YNDA& y,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	349 void (op) (size_t, AELEMT(RNDA) , XS,
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	350 const AELEMT(YNDA) *))
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	351 {
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	352 RNDA r (y.dims ());
9612 66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	353 op (r.length (), r.fortran_vec (), x, y.data ());
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	354 return r;
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	355 }
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	356
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	357 template <class RNDA, class XNDA>
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	358 inline RNDA&
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	359 do_mm_inplace_op (RNDA& r, const XNDA& x,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	360 void (op) (size_t, AELEMT(RNDA) ,
ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	361 const AELEMT(XNDA) *),
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	362 const char *opname)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	363 {
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	364 dim_vector dr = r.dims (), dx = x.dims ();
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	365 if (dr == dx)
9612 66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	366 op (r.length (), r.fortran_vec (), x.data ());
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	367 else
9612 66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	368 gripe_nonconformant (opname, dr, dx);
66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	369 return r;
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	370 }
7789 82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	371
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	372 template <class RNDA, class XS>
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	373 inline RNDA&
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	374 do_ms_inplace_op (RNDA& r, const XS& x,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	375 void (op) (size_t, AELEMT(RNDA) , XS))
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	376 {
9612 66970dd627f6 further liboctave design improvements Jaroslav Hajek <highegg@gmail.com> parents: 9607 diff changeset	377 op (r.length (), r.fortran_vec (), x);
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	378 return r;
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	379 }
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	380
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	381 template <class T1, class T2>
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	382 inline bool
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	383 mx_inline_equal (size_t n, const T1 x, const T2 y)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	384 {
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	385 for (size_t i = 0; i < n; i++)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	386 if (x[i] != y[i])
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	387 return false;
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	388 return true;
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	389 }
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	390
2804 eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	391 #define OP_DUP_FCN(OP, F, R, T) \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	392 static inline R * \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	393 F (const T *x, size_t n) \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	394 { \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	395 R *r = 0; \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	396 if (n > 0) \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	397 { \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	398 r = new R [n]; \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	399 for (size_t i = 0; i < n; i++) \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	400 r[i] = OP (x[i]); \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	401 } \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	402 return r; \
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	403 }
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	404
3769 7c8e3c42ed04 [project @ 2001-01-31 22:15:51 by jwe] jwe parents: 3504 diff changeset	405 OP_DUP_FCN (, mx_inline_dup, double, double)
7c8e3c42ed04 [project @ 2001-01-31 22:15:51 by jwe] jwe parents: 3504 diff changeset	406 OP_DUP_FCN (, mx_inline_dup, Complex, Complex)
7789 82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	407 OP_DUP_FCN (, mx_inline_dup, float, float)
82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	408 OP_DUP_FCN (, mx_inline_dup, FloatComplex, FloatComplex)
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	409
2804 eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	410 // These should really return a bool *. Also, they should probably be
eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	411 // in with a collection of other element-by-element boolean ops.
3769 7c8e3c42ed04 [project @ 2001-01-31 22:15:51 by jwe] jwe parents: 3504 diff changeset	412 OP_DUP_FCN (0.0 ==, mx_inline_not, double, double)
7c8e3c42ed04 [project @ 2001-01-31 22:15:51 by jwe] jwe parents: 3504 diff changeset	413 OP_DUP_FCN (0.0 ==, mx_inline_not, double, Complex)
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	414
3769 7c8e3c42ed04 [project @ 2001-01-31 22:15:51 by jwe] jwe parents: 3504 diff changeset	415 OP_DUP_FCN (, mx_inline_make_complex, Complex, double)
2804 eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	416
3769 7c8e3c42ed04 [project @ 2001-01-31 22:15:51 by jwe] jwe parents: 3504 diff changeset	417 OP_DUP_FCN (-, mx_inline_change_sign, double, double)
7c8e3c42ed04 [project @ 2001-01-31 22:15:51 by jwe] jwe parents: 3504 diff changeset	418 OP_DUP_FCN (-, mx_inline_change_sign, Complex, Complex)
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	419
8650 a1ae2aae903e abs,real,imag,conj: use code from mx-inlines rather than the generic map Jaroslav Hajek <highegg@gmail.com> parents: 8380 diff changeset	420 OP_DUP_FCN (std::abs, mx_inline_fabs_dup, double, double)
a1ae2aae903e abs,real,imag,conj: use code from mx-inlines rather than the generic map Jaroslav Hajek <highegg@gmail.com> parents: 8380 diff changeset	421 OP_DUP_FCN (std::abs, mx_inline_cabs_dup, double, Complex)
3769 7c8e3c42ed04 [project @ 2001-01-31 22:15:51 by jwe] jwe parents: 3504 diff changeset	422 OP_DUP_FCN (real, mx_inline_real_dup, double, Complex)
7c8e3c42ed04 [project @ 2001-01-31 22:15:51 by jwe] jwe parents: 3504 diff changeset	423 OP_DUP_FCN (imag, mx_inline_imag_dup, double, Complex)
7c8e3c42ed04 [project @ 2001-01-31 22:15:51 by jwe] jwe parents: 3504 diff changeset	424 OP_DUP_FCN (conj, mx_inline_conj_dup, Complex, Complex)
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	425
7789 82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	426 OP_DUP_FCN (0.0 ==, mx_inline_not, float, float)
82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	427 OP_DUP_FCN (static_cast<float>(0.0) ==, mx_inline_not, float, FloatComplex)
82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	428
82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	429 OP_DUP_FCN (, mx_inline_make_complex, FloatComplex, float)
82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	430
82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	431 OP_DUP_FCN (-, mx_inline_change_sign, float, float)
82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	432 OP_DUP_FCN (-, mx_inline_change_sign, FloatComplex, FloatComplex)
82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	433
8650 a1ae2aae903e abs,real,imag,conj: use code from mx-inlines rather than the generic map Jaroslav Hajek <highegg@gmail.com> parents: 8380 diff changeset	434 OP_DUP_FCN (std::abs, mx_inline_fabs_dup, float, float)
a1ae2aae903e abs,real,imag,conj: use code from mx-inlines rather than the generic map Jaroslav Hajek <highegg@gmail.com> parents: 8380 diff changeset	435 OP_DUP_FCN (std::abs, mx_inline_cabs_dup, float, FloatComplex)
7789 82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	436 OP_DUP_FCN (real, mx_inline_real_dup, float, FloatComplex)
82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	437 OP_DUP_FCN (imag, mx_inline_imag_dup, float, FloatComplex)
82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	438 OP_DUP_FCN (conj, mx_inline_conj_dup, FloatComplex, FloatComplex)
82be108cc558 First attempt at single precision tyeps David Bateman <dbateman@free.fr> parents: 7017 diff changeset	439
8899 4ab9654eceac update comment Jaroslav Hajek <highegg@gmail.com> parents: 8897 diff changeset	440 // FIXME: Due to a performance defect in g++ (<= 4.3), std::norm is slow unless
4ab9654eceac update comment Jaroslav Hajek <highegg@gmail.com> parents: 8897 diff changeset	441 // ffast-math is on (not by default even with -O3). The following helper function
4ab9654eceac update comment Jaroslav Hajek <highegg@gmail.com> parents: 8897 diff changeset	442 // gives the expected straightforward implementation of std::norm.
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	443 template <class T>
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	444 inline T cabsq (const std::complex<T>& c)
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	445 { return c.real () * c.real () + c.imag () * c.imag (); }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	446
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	447 // default. works for integers and bool.
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	448 template <class T>
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	449 inline bool xis_true (T x) { return x; }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	450 template <class T>
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	451 inline bool xis_false (T x) { return ! x; }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	452 // for octave_ints
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	453 template <class T>
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	454 inline bool xis_true (const octave_int<T>& x) { return x.value (); }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	455 template <class T>
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	456 inline bool xis_false (const octave_int<T>& x) { return ! x.value (); }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	457 // for reals, we want to ignore NaNs.
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	458 inline bool xis_true (double x) { return ! xisnan (x) && x != 0.0; }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	459 inline bool xis_false (double x) { return x == 0.0; }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	460 inline bool xis_true (float x) { return ! xisnan (x) && x != 0.0f; }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	461 inline bool xis_false (float x) { return x == 0.0f; }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	462 // Ditto for complex.
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	463 inline bool xis_true (const Complex& x) { return ! xisnan (x) && x != 0.0; }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	464 inline bool xis_false (const Complex& x) { return x == 0.0; }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	465 inline bool xis_true (const FloatComplex& x) { return ! xisnan (x) && x != 0.0f; }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	466 inline bool xis_false (const FloatComplex& x) { return x == 0.0f; }
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	467
9550 3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	468 #define OP_RED_SUM(ac, el) ac += el
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	469 #define OP_RED_PROD(ac, el) ac *= el
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	470 #define OP_RED_SUMSQ(ac, el) ac += el*el
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	471 #define OP_RED_SUMSQC(ac, el) ac += cabsq (el)
3d6a9aea2aea refactor binary & bool ops in liboctave Jaroslav Hajek <highegg@gmail.com> parents: 9513 diff changeset	472
9721 192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	473 inline void op_dble_sum(double& ac, float el)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	474 { ac += el; }
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	475 inline void op_dble_sum(Complex& ac, const FloatComplex& el)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	476 { ac += el; } // FIXME: guaranteed?
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	477 template <class T>
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	478 inline void op_dble_sum(double& ac, const octave_int<T>& el)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	479 { ac += el.double_value (); }
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	480
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	481 // The following two implement a simple short-circuiting.
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	482 #define OP_RED_ANYC(ac, el) if (xis_true (el)) { ac = true; break; } else continue
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	483 #define OP_RED_ALLC(ac, el) if (xis_false (el)) { ac = false; break; } else continue
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	484
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	485 #define OP_RED_FCN(F, TSRC, TRES, OP, ZERO) \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	486 template <class T> \
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	487 inline TRES \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	488 F (const TSRC* v, octave_idx_type n) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	489 { \
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	490 TRES ac = ZERO; \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	491 for (octave_idx_type i = 0; i < n; i++) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	492 OP(ac, v[i]); \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	493 return ac; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	494 }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	495
9721 192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	496 #define PROMOTE_DOUBLE(T) typename subst_template_param<std::complex, T, double>::type
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	497
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	498 OP_RED_FCN (mx_inline_sum, T, T, OP_RED_SUM, 0)
9721 192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	499 OP_RED_FCN (mx_inline_dsum, T, PROMOTE_DOUBLE(T), op_dble_sum, 0.0)
8756 d0755c9db5ed implement fast logical sum (counting) Jaroslav Hajek <highegg@gmail.com> parents: 8751 diff changeset	500 OP_RED_FCN (mx_inline_count, bool, T, OP_RED_SUM, 0)
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	501 OP_RED_FCN (mx_inline_prod, T, T, OP_RED_PROD, 1)
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	502 OP_RED_FCN (mx_inline_sumsq, T, T, OP_RED_SUMSQ, 0)
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	503 OP_RED_FCN (mx_inline_sumsq, std::complex<T>, T, OP_RED_SUMSQC, 0)
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	504 OP_RED_FCN (mx_inline_any, T, bool, OP_RED_ANYC, false)
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	505 OP_RED_FCN (mx_inline_all, T, bool, OP_RED_ALLC, true)
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	506
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	507
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	508 #define OP_RED_FCN2(F, TSRC, TRES, OP, ZERO) \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	509 template <class T> \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	510 inline void \
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	511 F (const TSRC* v, TRES *r, octave_idx_type m, octave_idx_type n) \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	512 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	513 for (octave_idx_type i = 0; i < m; i++) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	514 r[i] = ZERO; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	515 for (octave_idx_type j = 0; j < n; j++) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	516 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	517 for (octave_idx_type i = 0; i < m; i++) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	518 OP(r[i], v[i]); \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	519 v += m; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	520 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	521 }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	522
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	523 OP_RED_FCN2 (mx_inline_sum, T, T, OP_RED_SUM, 0)
9721 192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	524 OP_RED_FCN2 (mx_inline_dsum, T, PROMOTE_DOUBLE(T), op_dble_sum, 0.0)
8756 d0755c9db5ed implement fast logical sum (counting) Jaroslav Hajek <highegg@gmail.com> parents: 8751 diff changeset	525 OP_RED_FCN2 (mx_inline_count, bool, T, OP_RED_SUM, 0)
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	526 OP_RED_FCN2 (mx_inline_prod, T, T, OP_RED_PROD, 1)
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	527 OP_RED_FCN2 (mx_inline_sumsq, T, T, OP_RED_SUMSQ, 0)
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	528 OP_RED_FCN2 (mx_inline_sumsq, std::complex<T>, T, OP_RED_SUMSQC, 0)
8758 83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	529
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	530 // Using the general code for any/all would sacrifice short-circuiting.
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	531 // OTOH, going by rows would sacrifice cache-coherence. The following algorithm
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	532 // will achieve both, at the cost of a temporary octave_idx_type array.
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	533
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	534 #define OP_ROW_SHORT_CIRCUIT(F, PRED, ZERO) \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	535 template <class T> \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	536 inline void \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	537 F (const T* v, bool *r, octave_idx_type m, octave_idx_type n) \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	538 { \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	539 /* FIXME: it may be sub-optimal to allocate the buffer here. */ \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	540 OCTAVE_LOCAL_BUFFER (octave_idx_type, iact, m); \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	541 for (octave_idx_type i = 0; i < m; i++) iact[i] = i; \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	542 octave_idx_type nact = m; \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	543 for (octave_idx_type j = 0; j < n; j++) \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	544 { \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	545 octave_idx_type k = 0; \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	546 for (octave_idx_type i = 0; i < nact; i++) \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	547 { \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	548 octave_idx_type ia = iact[i]; \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	549 if (! PRED (v[ia])) \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	550 iact[k++] = ia; \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	551 } \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	552 nact = k; \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	553 v += m; \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	554 } \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	555 for (octave_idx_type i = 0; i < m; i++) r[i] = ! ZERO; \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	556 for (octave_idx_type i = 0; i < nact; i++) r[iact[i]] = ZERO; \
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	557 }
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	558
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	559 OP_ROW_SHORT_CIRCUIT (mx_inline_any, xis_true, false)
83c9d60c3c47 implement short-circuiting row-reduction any/all algorithm Jaroslav Hajek <highegg@gmail.com> parents: 8756 diff changeset	560 OP_ROW_SHORT_CIRCUIT (mx_inline_all, xis_false, true)
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	561
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	562 #define OP_RED_FCNN(F, TSRC, TRES) \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	563 template <class T> \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	564 inline void \
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	565 F (const TSRC v, TRES r, octave_idx_type l, \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	566 octave_idx_type n, octave_idx_type u) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	567 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	568 if (l == 1) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	569 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	570 for (octave_idx_type i = 0; i < u; i++) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	571 { \
8756 d0755c9db5ed implement fast logical sum (counting) Jaroslav Hajek <highegg@gmail.com> parents: 8751 diff changeset	572 r[i] = F<T> (v, n); \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	573 v += n; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	574 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	575 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	576 else \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	577 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	578 for (octave_idx_type i = 0; i < u; i++) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	579 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	580 F (v, r, l, n); \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	581 v += l*n; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	582 r += l; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	583 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	584 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	585 }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	586
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	587 OP_RED_FCNN (mx_inline_sum, T, T)
9721 192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	588 OP_RED_FCNN (mx_inline_dsum, T, PROMOTE_DOUBLE(T))
8756 d0755c9db5ed implement fast logical sum (counting) Jaroslav Hajek <highegg@gmail.com> parents: 8751 diff changeset	589 OP_RED_FCNN (mx_inline_count, bool, T)
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	590 OP_RED_FCNN (mx_inline_prod, T, T)
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	591 OP_RED_FCNN (mx_inline_sumsq, T, T)
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	592 OP_RED_FCNN (mx_inline_sumsq, std::complex<T>, T)
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	593 OP_RED_FCNN (mx_inline_any, T, bool)
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	594 OP_RED_FCNN (mx_inline_all, T, bool)
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	595
8780 ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	596 #define OP_CUM_FCN(F, TSRC, TRES, OP) \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	597 template <class T> \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	598 inline void \
8780 ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	599 F (const TSRC v, TRES r, octave_idx_type n) \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	600 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	601 if (n) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	602 { \
8780 ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	603 TRES t = r[0] = v[0]; \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	604 for (octave_idx_type i = 1; i < n; i++) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	605 r[i] = t = t OP v[i]; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	606 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	607 }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	608
8780 ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	609 OP_CUM_FCN (mx_inline_cumsum, T, T, +)
ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	610 OP_CUM_FCN (mx_inline_cumprod, T, T, *)
ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	611 OP_CUM_FCN (mx_inline_cumcount, bool, T, +)
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	612
8780 ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	613 #define OP_CUM_FCN2(F, TSRC, TRES, OP) \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	614 template <class T> \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	615 inline void \
8780 ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	616 F (const TSRC v, TRES r, octave_idx_type m, octave_idx_type n) \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	617 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	618 if (n) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	619 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	620 for (octave_idx_type i = 0; i < m; i++) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	621 r[i] = v[i]; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	622 const T *r0 = r; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	623 for (octave_idx_type j = 1; j < n; j++) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	624 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	625 r += m; v += m; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	626 for (octave_idx_type i = 0; i < m; i++) \
8780 ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	627 r[i] = r0[i] OP v[i]; \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	628 r0 += m; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	629 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	630 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	631 }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	632
8780 ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	633 OP_CUM_FCN2 (mx_inline_cumsum, T, T, +)
ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	634 OP_CUM_FCN2 (mx_inline_cumprod, T, T, *)
ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	635 OP_CUM_FCN2 (mx_inline_cumcount, bool, T, *)
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	636
8780 ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	637 #define OP_CUM_FCNN(F, TSRC, TRES) \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	638 template <class T> \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	639 inline void \
8780 ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	640 F (const TSRC v, TRES r, octave_idx_type l, \
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	641 octave_idx_type n, octave_idx_type u) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	642 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	643 if (l == 1) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	644 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	645 for (octave_idx_type i = 0; i < u; i++) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	646 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	647 F (v, r, n); \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	648 v += n; r += n; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	649 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	650 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	651 else \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	652 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	653 for (octave_idx_type i = 0; i < u; i++) \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	654 { \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	655 F (v, r, l, n); \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	656 v += l*n; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	657 r += l*n; \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	658 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	659 } \
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	660 }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	661
8780 ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	662 OP_CUM_FCNN (mx_inline_cumsum, T, T)
ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	663 OP_CUM_FCNN (mx_inline_cumprod, T, T)
ea76466605ba support native cumsum, gripe on overflow in sum/cumsum Jaroslav Hajek <highegg@gmail.com> parents: 8777 diff changeset	664 OP_CUM_FCNN (mx_inline_cumcount, bool, T)
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	665
8751 9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	666 #define OP_MINMAX_FCN(F, OP) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	667 template <class T> \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	668 void F (const T v, T r, octave_idx_type n) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	669 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	670 if (! n) return; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	671 T tmp = v[0]; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	672 octave_idx_type i = 1; \
8776 d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	673 if (xisnan (tmp)) \
d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	674 { \
d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	675 for (; i < n && xisnan (v[i]); i++) ; \
d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	676 if (i < n) tmp = v[i]; \
d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	677 } \
d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	678 for (; i < n; i++) \
8751 9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	679 if (v[i] OP tmp) tmp = v[i]; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	680 *r = tmp; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	681 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	682 template <class T> \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	683 void F (const T v, T r, octave_idx_type *ri, octave_idx_type n) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	684 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	685 if (! n) return; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	686 T tmp = v[0]; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	687 octave_idx_type tmpi = 0; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	688 octave_idx_type i = 1; \
8776 d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	689 if (xisnan (tmp)) \
d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	690 { \
d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	691 for (; i < n && xisnan (v[i]); i++) ; \
d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	692 if (i < n) { tmp = v[i]; tmpi = i; } \
d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	693 } \
d23c33ec6bd3 fix min/max behaviour with NaNs Jaroslav Hajek <highegg@gmail.com> parents: 8759 diff changeset	694 for (; i < n; i++) \
8751 9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	695 if (v[i] OP tmp) { tmp = v[i]; tmpi = i; }\
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	696 *r = tmp; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	697 *ri = tmpi; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	698 }
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	699
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	700 OP_MINMAX_FCN (mx_inline_min, <)
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	701 OP_MINMAX_FCN (mx_inline_max, >)
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	702
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	703 // Row reductions will be slightly complicated. We will proceed with checks
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	704 // for NaNs until we detect that no row will yield a NaN, in which case we
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	705 // proceed to a faster code.
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	706
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	707 #define OP_MINMAX_FCN2(F, OP) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	708 template <class T> \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	709 inline void \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	710 F (const T v, T r, octave_idx_type m, octave_idx_type n) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	711 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	712 if (! n) return; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	713 bool nan = false; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	714 octave_idx_type j = 0; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	715 for (octave_idx_type i = 0; i < m; i++) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	716 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	717 r[i] = v[i]; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	718 if (xisnan (v[i])) nan = true; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	719 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	720 j++; v += m; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	721 while (nan && j < n) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	722 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	723 nan = false; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	724 for (octave_idx_type i = 0; i < m; i++) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	725 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	726 if (xisnan (v[i])) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	727 nan = true; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	728 else if (xisnan (r[i]) \|\| v[i] OP r[i]) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	729 r[i] = v[i]; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	730 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	731 j++; v += m; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	732 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	733 while (j < n) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	734 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	735 for (octave_idx_type i = 0; i < m; i++) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	736 if (v[i] OP r[i]) r[i] = v[i]; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	737 j++; v += m; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	738 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	739 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	740 template <class T> \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	741 inline void \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	742 F (const T v, T r, octave_idx_type *ri, \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	743 octave_idx_type m, octave_idx_type n) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	744 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	745 if (! n) return; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	746 bool nan = false; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	747 octave_idx_type j = 0; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	748 for (octave_idx_type i = 0; i < m; i++) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	749 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	750 r[i] = v[i]; ri[i] = j; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	751 if (xisnan (v[i])) nan = true; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	752 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	753 j++; v += m; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	754 while (nan && j < n) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	755 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	756 nan = false; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	757 for (octave_idx_type i = 0; i < m; i++) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	758 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	759 if (xisnan (v[i])) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	760 nan = true; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	761 else if (xisnan (r[i]) \|\| v[i] OP r[i]) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	762 { r[i] = v[i]; ri[i] = j; } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	763 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	764 j++; v += m; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	765 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	766 while (j < n) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	767 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	768 for (octave_idx_type i = 0; i < m; i++) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	769 if (v[i] OP r[i]) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	770 { r[i] = v[i]; ri[i] = j; } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	771 j++; v += m; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	772 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	773 }
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	774
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	775 OP_MINMAX_FCN2 (mx_inline_min, <)
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	776 OP_MINMAX_FCN2 (mx_inline_max, >)
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	777
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	778 #define OP_MINMAX_FCNN(F) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	779 template <class T> \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	780 inline void \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	781 F (const T v, T r, octave_idx_type l, \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	782 octave_idx_type n, octave_idx_type u) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	783 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	784 if (! n) return; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	785 if (l == 1) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	786 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	787 for (octave_idx_type i = 0; i < u; i++) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	788 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	789 F (v, r, n); \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	790 v += n; r++; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	791 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	792 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	793 else \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	794 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	795 for (octave_idx_type i = 0; i < u; i++) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	796 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	797 F (v, r, l, n); \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	798 v += l*n; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	799 r += l; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	800 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	801 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	802 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	803 template <class T> \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	804 inline void \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	805 F (const T v, T r, octave_idx_type *ri, \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	806 octave_idx_type l, octave_idx_type n, octave_idx_type u) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	807 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	808 if (! n) return; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	809 if (l == 1) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	810 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	811 for (octave_idx_type i = 0; i < u; i++) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	812 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	813 F (v, r, ri, n); \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	814 v += n; r++; ri++; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	815 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	816 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	817 else \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	818 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	819 for (octave_idx_type i = 0; i < u; i++) \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	820 { \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	821 F (v, r, ri, l, n); \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	822 v += l*n; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	823 r += l; ri += l; \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	824 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	825 } \
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	826 }
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	827
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	828 OP_MINMAX_FCNN (mx_inline_min)
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	829 OP_MINMAX_FCNN (mx_inline_max)
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	830
8777 724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	831 #define OP_CUMMINMAX_FCN(F, OP) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	832 template <class T> \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	833 void F (const T v, T r, octave_idx_type n) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	834 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	835 if (! n) return; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	836 T tmp = v[0]; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	837 octave_idx_type i = 1, j = 0; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	838 if (xisnan (tmp)) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	839 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	840 for (; i < n && xisnan (v[i]); i++) ; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	841 for (; j < i; j++) r[j] = tmp; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	842 if (i < n) tmp = v[i]; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	843 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	844 for (; i < n; i++) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	845 if (v[i] OP tmp) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	846 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	847 for (; j < i; j++) r[j] = tmp; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	848 tmp = v[i]; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	849 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	850 for (; j < i; j++) r[j] = tmp; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	851 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	852 template <class T> \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	853 void F (const T v, T r, octave_idx_type *ri, octave_idx_type n) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	854 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	855 if (! n) return; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	856 T tmp = v[0]; octave_idx_type tmpi = 0; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	857 octave_idx_type i = 1, j = 0; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	858 if (xisnan (tmp)) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	859 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	860 for (; i < n && xisnan (v[i]); i++) ; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	861 for (; j < i; j++) { r[j] = tmp; ri[j] = tmpi; } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	862 if (i < n) { tmp = v[i]; tmpi = i; } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	863 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	864 for (; i < n; i++) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	865 if (v[i] OP tmp) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	866 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	867 for (; j < i; j++) { r[j] = tmp; ri[j] = tmpi; } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	868 tmp = v[i]; tmpi = i; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	869 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	870 for (; j < i; j++) { r[j] = tmp; ri[j] = tmpi; } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	871 }
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	872
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	873 OP_CUMMINMAX_FCN (mx_inline_cummin, <)
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	874 OP_CUMMINMAX_FCN (mx_inline_cummax, >)
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	875
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	876 // Row reductions will be slightly complicated. We will proceed with checks
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	877 // for NaNs until we detect that no row will yield a NaN, in which case we
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	878 // proceed to a faster code.
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	879
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	880 #define OP_CUMMINMAX_FCN2(F, OP) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	881 template <class T> \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	882 inline void \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	883 F (const T v, T r, octave_idx_type m, octave_idx_type n) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	884 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	885 if (! n) return; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	886 bool nan = false; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	887 const T *r0; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	888 octave_idx_type j = 0; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	889 for (octave_idx_type i = 0; i < m; i++) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	890 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	891 r[i] = v[i]; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	892 if (xisnan (v[i])) nan = true; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	893 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	894 j++; v += m; r0 = r; r += m; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	895 while (nan && j < n) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	896 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	897 nan = false; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	898 for (octave_idx_type i = 0; i < m; i++) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	899 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	900 if (xisnan (v[i])) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	901 { r[i] = r0[i]; nan = true; } \
8949 e31d47f2c9bb fixes to row-reducing cummin/cummax Jaroslav Hajek <highegg@gmail.com> parents: 8920 diff changeset	902 else if (xisnan (r0[i]) \|\| v[i] OP r0[i]) \
8777 724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	903 r[i] = v[i]; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	904 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	905 j++; v += m; r0 = r; r += m; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	906 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	907 while (j < n) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	908 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	909 for (octave_idx_type i = 0; i < m; i++) \
8949 e31d47f2c9bb fixes to row-reducing cummin/cummax Jaroslav Hajek <highegg@gmail.com> parents: 8920 diff changeset	910 if (v[i] OP r0[i]) \
8777 724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	911 r[i] = v[i]; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	912 else \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	913 r[i] = r0[i]; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	914 j++; v += m; r0 = r; r += m; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	915 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	916 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	917 template <class T> \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	918 inline void \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	919 F (const T v, T r, octave_idx_type *ri, \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	920 octave_idx_type m, octave_idx_type n) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	921 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	922 if (! n) return; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	923 bool nan = false; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	924 const T r0; const octave_idx_type r0i; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	925 octave_idx_type j = 0; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	926 for (octave_idx_type i = 0; i < m; i++) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	927 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	928 r[i] = v[i]; ri[i] = 0; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	929 if (xisnan (v[i])) nan = true; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	930 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	931 j++; v += m; r0 = r; r += m; r0i = ri; ri += m; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	932 while (nan && j < n) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	933 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	934 nan = false; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	935 for (octave_idx_type i = 0; i < m; i++) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	936 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	937 if (xisnan (v[i])) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	938 { r[i] = r0[i]; ri[i] = r0i[i]; nan = true; } \
8949 e31d47f2c9bb fixes to row-reducing cummin/cummax Jaroslav Hajek <highegg@gmail.com> parents: 8920 diff changeset	939 else if (xisnan (r0[i]) \|\| v[i] OP r0[i]) \
8777 724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	940 { r[i] = v[i]; ri[i] = j; }\
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	941 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	942 j++; v += m; r0 = r; r += m; r0i = ri; ri += m; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	943 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	944 while (j < n) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	945 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	946 for (octave_idx_type i = 0; i < m; i++) \
8949 e31d47f2c9bb fixes to row-reducing cummin/cummax Jaroslav Hajek <highegg@gmail.com> parents: 8920 diff changeset	947 if (v[i] OP r0[i]) \
8777 724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	948 { r[i] = v[i]; ri[i] = j; } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	949 else \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	950 { r[i] = r0[i]; ri[i] = r0i[i]; } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	951 j++; v += m; r0 = r; r += m; r0i = ri; ri += m; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	952 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	953 }
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	954
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	955 OP_CUMMINMAX_FCN2 (mx_inline_cummin, <)
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	956 OP_CUMMINMAX_FCN2 (mx_inline_cummax, >)
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	957
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	958 #define OP_CUMMINMAX_FCNN(F) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	959 template <class T> \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	960 inline void \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	961 F (const T v, T r, octave_idx_type l, \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	962 octave_idx_type n, octave_idx_type u) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	963 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	964 if (! n) return; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	965 if (l == 1) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	966 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	967 for (octave_idx_type i = 0; i < u; i++) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	968 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	969 F (v, r, n); \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	970 v += n; r += n; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	971 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	972 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	973 else \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	974 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	975 for (octave_idx_type i = 0; i < u; i++) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	976 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	977 F (v, r, l, n); \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	978 v += l*n; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	979 r += l*n; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	980 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	981 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	982 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	983 template <class T> \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	984 inline void \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	985 F (const T v, T r, octave_idx_type *ri, \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	986 octave_idx_type l, octave_idx_type n, octave_idx_type u) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	987 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	988 if (! n) return; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	989 if (l == 1) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	990 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	991 for (octave_idx_type i = 0; i < u; i++) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	992 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	993 F (v, r, ri, n); \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	994 v += n; r += n; ri += n; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	995 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	996 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	997 else \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	998 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	999 for (octave_idx_type i = 0; i < u; i++) \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1000 { \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1001 F (v, r, ri, l, n); \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1002 v += l*n; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1003 r += ln; ri += ln; \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1004 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1005 } \
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1006 }
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1007
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1008 OP_CUMMINMAX_FCNN (mx_inline_cummin)
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1009 OP_CUMMINMAX_FCNN (mx_inline_cummax)
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1010
9513 9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1011 template <class T>
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1012 void mx_inline_diff (const T v, T r, octave_idx_type n,
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1013 octave_idx_type order)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1014 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1015 switch (order)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1016 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1017 case 1:
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1018 for (octave_idx_type i = 0; i < n-1; i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1019 r[i] = v[i+1] - v[i];
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1020 break;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1021 case 2:
9702 9ecd35a606e3 avoid some warnings from g++ John W. Eaton <jwe@octave.org> parents: 9612 diff changeset	1022 if (n > 1)
9ecd35a606e3 avoid some warnings from g++ John W. Eaton <jwe@octave.org> parents: 9612 diff changeset	1023 {
9ecd35a606e3 avoid some warnings from g++ John W. Eaton <jwe@octave.org> parents: 9612 diff changeset	1024 T lst = v[1] - v[0];
9513 9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1025 for (octave_idx_type i = 0; i < n-2; i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1026 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1027 T dif = v[i+2] - v[i+1];
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1028 r[i] = dif - lst;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1029 lst = dif;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1030 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1031 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1032 break;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1033 default:
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1034 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1035 OCTAVE_LOCAL_BUFFER (T, buf, n-1);
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1036
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1037 for (octave_idx_type i = 0; i < n-1; i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1038 buf[i] = v[i+1] - v[i];
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1039
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1040 for (octave_idx_type o = 2; o <= order; o++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1041 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1042 for (octave_idx_type i = 0; i < n-o; i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1043 buf[i] = buf[i+1] - buf[i];
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1044 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1045
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1046 for (octave_idx_type i = 0; i < n-order; i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1047 r[i] = buf[i];
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1048 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1049 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1050 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1051
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1052 template <class T>
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1053 void mx_inline_diff (const T v, T r,
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1054 octave_idx_type m, octave_idx_type n,
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1055 octave_idx_type order)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1056 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1057 switch (order)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1058 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1059 case 1:
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1060 for (octave_idx_type i = 0; i < m*(n-1); i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1061 r[i] = v[i+m] - v[i];
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1062 break;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1063 case 2:
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1064 for (octave_idx_type i = 0; i < n-2; i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1065 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1066 for (octave_idx_type j = im; j < im+m; j++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1067 r[j] = (v[j+m+m] - v[j+m]) + (v[j+m] - v[j]);
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1068 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1069 break;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1070 default:
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1071 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1072 OCTAVE_LOCAL_BUFFER (T, buf, n-1);
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1073
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1074 for (octave_idx_type j = 0; j < m; j++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1075 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1076 for (octave_idx_type i = 0; i < n-1; i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1077 buf[i] = v[im+j+m] - v[im+j];
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1078
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1079 for (octave_idx_type o = 2; o <= order; o++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1080 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1081 for (octave_idx_type i = 0; i < n-o; i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1082 buf[i] = buf[i+1] - buf[i];
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1083 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1084
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1085 for (octave_idx_type i = 0; i < n-order; i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1086 r[i*m+j] = buf[i];
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1087 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1088 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1089 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1090 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1091
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1092 template <class T>
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1093 inline void
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1094 mx_inline_diff (const T v, T r,
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1095 octave_idx_type l, octave_idx_type n, octave_idx_type u,
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1096 octave_idx_type order)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1097 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1098 if (! n) return;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1099 if (l == 1)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1100 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1101 for (octave_idx_type i = 0; i < u; i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1102 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1103 mx_inline_diff (v, r, n, order);
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1104 v += n; r += n-order;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1105 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1106 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1107 else
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1108 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1109 for (octave_idx_type i = 0; i < u; i++)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1110 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1111 mx_inline_diff (v, r, l, n, order);
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1112 v += l*n;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1113 r += l*(n-order);
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1114 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1115 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1116 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1117
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1118 // Assistant function
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1119
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1120 inline void
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1121 get_extent_triplet (const dim_vector& dims, int& dim,
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1122 octave_idx_type& l, octave_idx_type& n,
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1123 octave_idx_type& u)
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1124 {
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1125 octave_idx_type ndims = dims.length ();
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1126 if (dim >= ndims)
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1127 {
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1128 l = dims.numel ();
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1129 n = 1;
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1130 u = 1;
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1131 }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1132 else
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1133 {
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1134 if (dim < 0)
9513 9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1135 dim = dims.first_non_singleton ();
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1136
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1137 // calculate extent triplet.
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1138 l = 1, n = dims(dim), u = 1;
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1139 for (octave_idx_type i = 0; i < dim; i++)
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1140 l *= dims (i);
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1141 for (octave_idx_type i = dim + 1; i < ndims; i++)
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1142 u *= dims (i);
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1143 }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1144 }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1145
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1146 // Appliers.
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1147 // FIXME: is this the best design? C++ gives a lot of options here...
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1148 // maybe it can be done without an explicit parameter?
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1149
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1150 template <class ArrayType, class T>
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1151 inline ArrayType
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1152 do_mx_red_op (const Array<T>& src, int dim,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	1153 void (mx_red_op) (const T , AELEMT(ArrayType) *,
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1154 octave_idx_type, octave_idx_type, octave_idx_type))
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1155 {
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1156 octave_idx_type l, n, u;
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1157 dim_vector dims = src.dims ();
8743 1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	1158 // M*b inconsistency: sum([]) = 0 etc.
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	1159 if (dims.length () == 2 && dims(0) == 0 && dims(1) == 0)
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	1160 dims (1) = 1;
1bd918cfb6e2 reimplement any & all using the new reduction code Jaroslav Hajek <highegg@gmail.com> parents: 8736 diff changeset	1161
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1162 get_extent_triplet (dims, dim, l, n, u);
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1163
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1164 // Reduction operation reduces the array size.
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1165 if (dim < dims.length ()) dims(dim) = 1;
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1166 dims.chop_trailing_singletons ();
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1167
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1168 ArrayType ret (dims);
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1169 mx_red_op (src.data (), ret.fortran_vec (), l, n, u);
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1170
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1171 return ret;
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1172 }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1173
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1174 template <class ArrayType, class T>
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1175 inline ArrayType
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1176 do_mx_cum_op (const Array<T>& src, int dim,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	1177 void (mx_cum_op) (const T , AELEMT(ArrayType) *,
8736 53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1178 octave_idx_type, octave_idx_type, octave_idx_type))
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1179 {
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1180 octave_idx_type l, n, u;
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1181 dim_vector dims = src.dims ();
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1182 get_extent_triplet (dims, dim, l, n, u);
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1183
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1184 // Cumulative operation doesn't reduce the array size.
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1185 ArrayType ret (dims);
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1186 mx_cum_op (src.data (), ret.fortran_vec (), l, n, u);
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1187
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1188 return ret;
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1189 }
53b4fdeacc2e improve reduction functions Jaroslav Hajek <highegg@gmail.com> parents: 8650 diff changeset	1190
8751 9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1191 template <class ArrayType>
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1192 inline ArrayType
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1193 do_mx_minmax_op (const ArrayType& src, int dim,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	1194 void (mx_minmax_op) (const AELEMT(ArrayType) , AELEMT(ArrayType) *,
8751 9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1195 octave_idx_type, octave_idx_type, octave_idx_type))
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1196 {
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1197 octave_idx_type l, n, u;
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1198 dim_vector dims = src.dims ();
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1199 get_extent_triplet (dims, dim, l, n, u);
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1200
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1201 // If the dimension is zero, we don't do anything.
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1202 if (dim < dims.length () && dims(dim) != 0) dims(dim) = 1;
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1203 dims.chop_trailing_singletons ();
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1204
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1205 ArrayType ret (dims);
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1206 mx_minmax_op (src.data (), ret.fortran_vec (), l, n, u);
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1207
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1208 return ret;
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1209 }
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1210
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1211 template <class ArrayType>
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1212 inline ArrayType
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1213 do_mx_minmax_op (const ArrayType& src, Array<octave_idx_type>& idx, int dim,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	1214 void (mx_minmax_op) (const AELEMT(ArrayType) , AELEMT(ArrayType) *,
8751 9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1215 octave_idx_type *,
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1216 octave_idx_type, octave_idx_type, octave_idx_type))
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1217 {
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1218 octave_idx_type l, n, u;
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1219 dim_vector dims = src.dims ();
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1220 get_extent_triplet (dims, dim, l, n, u);
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1221
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1222 // If the dimension is zero, we don't do anything.
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1223 if (dim < dims.length () && dims(dim) != 0) dims(dim) = 1;
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1224 dims.chop_trailing_singletons ();
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1225
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1226 ArrayType ret (dims);
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1227 if (idx.dims () != dims) idx = Array<octave_idx_type> (dims);
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1228
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1229 mx_minmax_op (src.data (), ret.fortran_vec (), idx.fortran_vec (),
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1230 l, n, u);
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1231
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1232 return ret;
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1233 }
9f7ce4bf7650 optimize min/max functions Jaroslav Hajek <highegg@gmail.com> parents: 8743 diff changeset	1234
8777 724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1235 template <class ArrayType>
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1236 inline ArrayType
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1237 do_mx_cumminmax_op (const ArrayType& src, int dim,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	1238 void (mx_cumminmax_op) (const AELEMT(ArrayType) , AELEMT(ArrayType) *,
8777 724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1239 octave_idx_type, octave_idx_type, octave_idx_type))
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1240 {
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1241 octave_idx_type l, n, u;
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1242 dim_vector dims = src.dims ();
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1243 get_extent_triplet (dims, dim, l, n, u);
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1244
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1245 ArrayType ret (dims);
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1246 mx_cumminmax_op (src.data (), ret.fortran_vec (), l, n, u);
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1247
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1248 return ret;
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1249 }
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1250
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1251 template <class ArrayType>
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1252 inline ArrayType
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1253 do_mx_cumminmax_op (const ArrayType& src, Array<octave_idx_type>& idx, int dim,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	1254 void (mx_cumminmax_op) (const AELEMT(ArrayType) , AELEMT(ArrayType) *,
8777 724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1255 octave_idx_type *,
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1256 octave_idx_type, octave_idx_type, octave_idx_type))
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1257 {
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1258 octave_idx_type l, n, u;
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1259 dim_vector dims = src.dims ();
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1260 get_extent_triplet (dims, dim, l, n, u);
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1261
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1262 ArrayType ret (dims);
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1263 if (idx.dims () != dims) idx = Array<octave_idx_type> (dims);
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1264
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1265 mx_cumminmax_op (src.data (), ret.fortran_vec (), idx.fortran_vec (),
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1266 l, n, u);
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1267
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1268 return ret;
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1269 }
724c0f46d9d4 implement cummin/cummax functions Jaroslav Hajek <highegg@gmail.com> parents: 8776 diff changeset	1270
9513 9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1271 template <class ArrayType>
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1272 inline ArrayType
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1273 do_mx_diff_op (const ArrayType& src, int dim, octave_idx_type order,
9800 ef4c4186cb47 improve some mx_inline loops Jaroslav Hajek <highegg@gmail.com> parents: 9766 diff changeset	1274 void (mx_diff_op) (const AELEMT(ArrayType) , AELEMT(ArrayType) *,
9513 9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1275 octave_idx_type, octave_idx_type, octave_idx_type,
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1276 octave_idx_type))
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1277 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1278 octave_idx_type l, n, u;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1279 if (order <= 0)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1280 return src;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1281
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1282 dim_vector dims = src.dims ();
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1283
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1284 get_extent_triplet (dims, dim, l, n, u);
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1285 if (dim >= dims.length ())
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1286 dims.resize (dim+1, 1);
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1287
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1288 if (dims(dim) <= order)
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1289 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1290 dims (dim) = 0;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1291 return ArrayType (dims);
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1292 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1293 else
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1294 {
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1295 dims(dim) -= order;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1296 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1297
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1298 ArrayType ret (dims);
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1299 mx_diff_op (src.data (), ret.fortran_vec (), l, n, u, order);
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1300
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1301 return ret;
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1302 }
9f870f73ab7d implement built-in diff Jaroslav Hajek <highegg@gmail.com> parents: 8949 diff changeset	1303
9721 192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1304 // Fast extra-precise summation. According to
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1305 // T. Ogita, S. M. Rump, S. Oishi:
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1306 // Accurate Sum And Dot Product,
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1307 // SIAM J. Sci. Computing, Vol. 26, 2005
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1308
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1309 template <class T>
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1310 inline void twosum_accum (T& s, T& e,
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1311 const T& x)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1312 {
9766 82fe4db20dec fix bad FLOAT_TRUNCATE in mx-inlines.cc Jaroslav Hajek <highegg@gmail.com> parents: 9743 diff changeset	1313 T s1 = s + x, t = s1 - s, e1 = (s - (s1 - t)) + (x - t);
9721 192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1314 s = s1;
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1315 e += e1;
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1316 }
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1317
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1318 template <class T>
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1319 inline T
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1320 mx_inline_xsum (const T *v, octave_idx_type n)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1321 {
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1322 T s = 0, e = 0;
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1323 for (octave_idx_type i = 0; i < n; i++)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1324 twosum_accum (s, e, v[i]);
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1325
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1326 return s + e;
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1327 }
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1328
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1329 template <class T>
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1330 inline void
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1331 mx_inline_xsum (const T v, T r,
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1332 octave_idx_type m, octave_idx_type n)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1333 {
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1334 OCTAVE_LOCAL_BUFFER (T, e, m);
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1335 for (octave_idx_type i = 0; i < m; i++)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1336 e[i] = r[i] = T ();
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1337
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1338 for (octave_idx_type j = 0; j < n; j++)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1339 {
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1340 for (octave_idx_type i = 0; i < m; i++)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1341 twosum_accum (r[i], e[i], v[i]);
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1342
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1343 v += m;
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1344 }
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1345
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1346 for (octave_idx_type i = 0; i < m; i++)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1347 r[i] += e[i];
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1348 }
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1349
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1350 OP_RED_FCNN (mx_inline_xsum, T, T)
192d94cff6c1 improve sum & implement the 'extra' option, refactor some code Jaroslav Hajek <highegg@gmail.com> parents: 9702 diff changeset	1351
2804 eedc2f3f61f7 [project @ 1997-03-11 04:34:50 by jwe] jwe parents: 2386 diff changeset	1352 #endif
3 9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	1353
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	1354 /*
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	1355 ;;; Local Variables: ***
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	1356 ;;; mode: C++ ***
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	1357 ;;; End: ***
9a4c07481e61 [project @ 1993-08-08 01:20:23 by jwe] jwe parents: diff changeset	1358 */

Mercurial > octave-nkf

annotate liboctave/mx-inlines.cc @ 9827:c15a5ed0da58