3 * Copyright (C) 2005-2008 Team XBMC
6 * This Program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2, or (at your option)
11 * This Program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with XBMC; see the file COPYING. If not, write to
18 * the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
19 * http://www.gnu.org/copyleft/gpl.html
29 #include <emmintrin.h>
32 /*! \brief Math utility class.
33 Note that the test() routine should return true for all implementations
35 See http://ldesoras.free.fr/doc/articles/rounding_en.pdf for an explanation
36 of the technique used on x86.
40 // GCC does something stupid with optimization on release builds if we try
41 // to assert in these functions
43 /*! \brief Round to nearest integer.
44 This routine does fast rounding to the nearest integer.
45 In the case (k + 0.5 for any integer k) we round up to k+1, and in all other
46 instances we should return the nearest integer.
47 Thus, { -1.5, -0.5, 0.5, 1.5 } is rounded to { -1, 0, 1, 2 }.
48 It preserves the property that round(k) - round(k-1) = 1 for all doubles k.
50 Make sure MathUtils::test() returns true for each implementation.
51 \sa truncate_int, test
53 inline int round_int (double x)
55 assert(x > static_cast<double>(INT_MIN / 2) - 1.0);
56 assert(x < static_cast <double>(INT_MAX / 2) + 1.0);
57 const float round_to_nearest = 0.5f;
61 const float round_dn_to_nearest = 0.4999999f;
62 i = (x > 0) ? _mm_cvttsd_si32(_mm_set_sd(x + round_to_nearest)) : _mm_cvttsd_si32(_mm_set_sd(x - round_dn_to_nearest));
63 #elif !defined(_LINUX)
73 #if defined(__powerpc__) || defined(__ppc__) || defined(__ARM_PCS_VFP) || defined(TARGET_DARWIN_OSX)
74 i = floor(x + round_to_nearest);
75 #elif defined(__arm__)
76 // From 'ARMĀ®v7-M Architecture Reference Manual' page A7-569:
77 // "The floating-point to integer operation (vcvt) [normally] uses the Round towards Zero rounding mode"
78 // Because of this...we must use some less-than-straightforward logic to perform this operation without
79 // changing the rounding mode flags
81 /* The assembly below implements the following logic:
86 int_val = trunc(x+inc);
93 __asm__ __volatile__ (
94 "vmov.F64 d1,%[rnd_val] \n\t" // Copy round_to_nearest into a working register (d1 = 0.5)
95 "fcmpezd %P[value] \n\t" // Check value against zero (value == 0?)
96 "fmstat \n\t" // Copy the floating-point status flags into the general-purpose status flags
98 "vnegmi.F64 d1, d1 \n\t" // if N-flag is set, negate round_to_nearest (if (value < 0) d1 = -1 * d1)
99 "vadd.F64 d1,%P[value],d1 \n\t" // Add round_to_nearest to value, store result in working register (d1 += value)
100 "vcvt.S32.F64 s3,d1 \n\t" // Truncate(round towards zero) (s3 = (int)d1)
101 "vmov %[result],s3 \n\t" // Store the integer result in a general-purpose register (result = s3)
102 "vcvt.F64.S32 d1,s3 \n\t" // Convert back to floating-point (d1 = (double)s3)
103 "vsub.F64 d1,%P[value],d1 \n\t" // Calculate the error (d1 = value - d1)
104 "vmov.F64 d2,%[rnd_val] \n\t" // d2 = 0.5;
105 "fcmped d1, d2 \n\t" // (d1 == 0.5?)
106 "fmstat \n\t" // Copy the floating-point status flags into the general-purpose status flags
108 "addeq %[result],#1 \n\t" // (if (d1 == d2) result++;)
109 : [result] "=r"(i) // Outputs
110 : [rnd_val] "Dv" (round_to_nearest), [value] "w"(x) // Inputs
111 : "d1", "d2", "s3" // Clobbers
114 __asm__ __volatile__ (
119 : "=m"(i) : "u"(round_to_nearest), "t"(x) : "st"
126 /*! \brief Truncate to nearest integer.
127 This routine does fast truncation to an integer.
128 It should simply drop the fractional portion of the floating point number.
130 Make sure MathUtils::test() returns true for each implementation.
133 inline int truncate_int(double x)
135 assert(x > static_cast<double>(INT_MIN / 2) - 1.0);
136 assert(x < static_cast <double>(INT_MAX / 2) + 1.0);
138 #if !defined(__powerpc__) && !defined(__ppc__) && !defined(__arm__) && !defined(TARGET_DARWIN_OSX)
139 const float round_towards_m_i = -0.5f;
149 fadd round_towards_m_i
154 #if defined(__powerpc__) || defined(__ppc__) || defined(TARGET_DARWIN_OSX)
156 #elif defined(__arm__)
157 __asm__ __volatile__ (
158 "vcvt.S32.F64 %[result],%P[value] \n\t" // Truncate(round towards zero) and store the result
159 : [result] "=w"(i) // Outputs
160 : [value] "w"(x) // Inputs
164 __asm__ __volatile__ (
170 : "=m"(i) : "u"(round_towards_m_i), "t"(x) : "st"
179 inline int64_t abs(int64_t a)
181 return (a < 0) ? -a : a;
186 // stupid hack to keep compiler from dropping these
187 // functions as unused
188 MathUtils::round_int(0.0);
189 MathUtils::truncate_int(0.0);
194 /*! \brief test routine for round_int and truncate_int
195 Must return true on all platforms.
199 for (int i = -8; i < 8; ++i)
202 int r = (i < 0) ? (i - 1) / 4 : (i + 2) / 4;
204 if (round_int(d) != r || truncate_int(d) != t)
210 } // namespace MathUtils