1 /* Copyright (c) 2002-2008 Jean-Marc Valin
2 Copyright (c) 2007-2008 CSIRO
3 Copyright (c) 2007-2009 Xiph.Org Foundation
4 Written by Jean-Marc Valin */
7 @brief Various math functions
10 Redistribution and use in source and binary forms, with or without
11 modification, are permitted provided that the following conditions
14 - Redistributions of source code must retain the above copyright
15 notice, this list of conditions and the following disclaimer.
17 - Redistributions in binary form must reproduce the above copyright
18 notice, this list of conditions and the following disclaimer in the
19 documentation and/or other materials provided with the distribution.
21 - Neither the name of the Xiph.org Foundation nor the names of its
22 contributors may be used to endorse or promote products derived from
23 this software without specific prior written permission.
25 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
28 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR
29 CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
30 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
31 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
32 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
33 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
34 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
35 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
44 /*Compute floor(sqrt(_val)) with exact arithmetic.
45 This has been tested on all possible 32-bit inputs.*/
46 unsigned isqrt32(celt_uint32 _val){
50 /*Uses the second method from
51 http://www.azillionmonkeys.com/qed/sqroot.html
52 The main idea is to search for the largest binary digit b such that
53 (g+b)*(g+b) <= _val, and add it to the solution g.*/
55 bshift=EC_ILOG(_val)-1>>1;
59 t=((celt_uint32)g<<1)+b<<bshift;
73 celt_word32 frac_div32(celt_word32 a, celt_word32 b)
76 celt_word32 result, rem;
77 int shift = 30-celt_ilog2(b);
81 /* 16-bit reciprocal */
82 rcp = ROUND16(celt_rcp(ROUND16(b,16)),2);
83 result = SHL32(MULT16_32_Q15(rcp, a),1);
84 rem = a-MULT32_32_Q31(result, b);
85 result += SHL32(MULT16_32_Q15(rcp, rem),1);
89 /** Reciprocal sqrt approximation in the range [0.25,1) (Q16 in, Q14 out) */
90 celt_word16 celt_rsqrt_norm(celt_word32 x)
96 /* Range of n is [-16384,32767] ([-0.5,1) in Q15). */
98 /* Get a rough initial guess for the root.
99 The optimal minimax quadratic approximation (using relative error) is
100 r = 1.437799046117536+n*(-0.823394375837328+n*0.4096419668459485).
101 Coefficients here, and the final result r, are Q14.*/
102 r = ADD16(23557, MULT16_16_Q15(n, ADD16(-13490, MULT16_16_Q15(n, 6713))));
103 /* We want y = x*r*r-1 in Q15, but x is 32-bit Q16 and r is Q14.
104 We can compute the result from n and r using Q15 multiplies with some
105 adjustment, carefully done to avoid overflow.
106 Range of y is [-1564,1594]. */
107 r2 = MULT16_16_Q15(r, r);
108 y = SHL16(SUB16(ADD16(MULT16_16_Q15(r2, n), r2), 16384), 1);
109 /* Apply a 2nd-order Householder iteration: r += r*y*(y*0.375-0.5).
110 This yields the Q14 reciprocal square root of the Q16 x, with a maximum
111 relative error of 1.04956E-4, a (relative) RMSE of 2.80979E-5, and a
112 peak absolute error of 2.26591/16384. */
113 return ADD16(r, MULT16_16_Q15(r, MULT16_16_Q15(y,
114 SUB16(MULT16_16_Q15(y, 12288), 16384))));
117 /** Sqrt approximation (QX input, QX/2 output) */
118 celt_word32 celt_sqrt(celt_word32 x)
123 static const celt_word16 C[5] = {23175, 11561, -3011, 1699, -664};
126 k = (celt_ilog2(x)>>1)-7;
127 x = VSHR32(x, (k<<1));
129 rt = ADD16(C[0], MULT16_16_Q15(n, ADD16(C[1], MULT16_16_Q15(n, ADD16(C[2],
130 MULT16_16_Q15(n, ADD16(C[3], MULT16_16_Q15(n, (C[4])))))))));
140 static inline celt_word16 _celt_cos_pi_2(celt_word16 x)
144 x2 = MULT16_16_P15(x,x);
145 return ADD16(1,MIN16(32766,ADD32(SUB16(L1,x2), MULT16_16_P15(x2, ADD32(L2, MULT16_16_P15(x2, ADD32(L3, MULT16_16_P15(L4, x2
154 celt_word16 celt_cos_norm(celt_word32 x)
157 if (x>SHL32(EXTEND32(1), 16))
158 x = SUB32(SHL32(EXTEND32(1), 17),x);
161 if (x<SHL32(EXTEND32(1), 15))
163 return _celt_cos_pi_2(EXTRACT16(x));
165 return NEG32(_celt_cos_pi_2(EXTRACT16(65536-x)));
170 else if (x&0x0001ffff)
177 /** Reciprocal approximation (Q15 input, Q16 output) */
178 celt_word32 celt_rcp(celt_word32 x)
183 celt_assert2(x>0, "celt_rcp() only defined for positive values");
185 /* n is Q15 with range [0,1). */
186 n = VSHR32(x,i-15)-32768;
187 /* Start with a linear approximation:
188 r = 1.8823529411764706-0.9411764705882353*n.
189 The coefficients and the result are Q14 in the range [15420,30840].*/
190 r = ADD16(30840, MULT16_16_Q15(-15420, n));
191 /* Perform two Newton iterations:
193 = r*((r*n)+(r-1.Q15)). */
194 r = SUB16(r, MULT16_16_Q15(r,
195 ADD16(MULT16_16_Q15(r, n), ADD16(r, -32768))));
196 /* We subtract an extra 1 in the second iteration to avoid overflow; it also
197 neatly compensates for truncation error in the rest of the process. */
198 r = SUB16(r, ADD16(1, MULT16_16_Q15(r,
199 ADD16(MULT16_16_Q15(r, n), ADD16(r, -32768)))));
200 /* r is now the Q15 solution to 2/(n+1), with a maximum relative error
201 of 7.05346E-5, a (relative) RMSE of 2.14418E-5, and a peak absolute
202 error of 1.24665/32768. */
203 return VSHR32(EXTEND32(r),i-16);