1 /******************************************************************************
2 *
3 * Copyright 2022 Google LLC
4 *
5 * Licensed under the Apache License, Version 2.0 (the "License");
6 * you may not use this file except in compliance with the License.
7 * You may obtain a copy of the License at:
8 *
9 * http://www.apache.org/licenses/LICENSE-2.0
10 *
11 * Unless required by applicable law or agreed to in writing, software
12 * distributed under the License is distributed on an "AS IS" BASIS,
13 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 * See the License for the specific language governing permissions and
15 * limitations under the License.
16 *
17 ******************************************************************************/
18
19 #ifndef __LC3_FASTMATH_H
20 #define __LC3_FASTMATH_H
21
22 #include <stdint.h>
23 #include <float.h>
24 #include <math.h>
25
26
27 /**
28 * IEEE 754 Floating point representation
29 */
30
31 #define LC3_IEEE754_SIGN_SHL (31)
32 #define LC3_IEEE754_SIGN_MASK (1 << LC3_IEEE754_SIGN_SHL)
33
34 #define LC3_IEEE754_EXP_SHL (23)
35 #define LC3_IEEE754_EXP_MASK (0xff << LC3_IEEE754_EXP_SHL)
36 #define LC3_IEEE754_EXP_BIAS (127)
37
38
39 /**
40 * Fast multiply floating-point number by integral power of 2
41 * x Operand, finite number
42 * exp Exponent such that 2^x is finite
43 * return 2^exp
44 */
lc3_ldexpf(float _x,int exp)45 static inline float lc3_ldexpf(float _x, int exp) {
46 union { float f; int32_t s; } x = { .f = _x };
47
48 if (x.s & LC3_IEEE754_EXP_MASK)
49 x.s += exp << LC3_IEEE754_EXP_SHL;
50
51 return x.f;
52 }
53
54 /**
55 * Fast convert floating-point number to fractional and integral components
56 * x Operand, finite number
57 * exp Return the exponent part
58 * return The normalized fraction in [0.5:1[
59 */
lc3_frexpf(float _x,int * exp)60 static inline float lc3_frexpf(float _x, int *exp) {
61 union { float f; uint32_t u; } x = { .f = _x };
62
63 int e = (x.u & LC3_IEEE754_EXP_MASK) >> LC3_IEEE754_EXP_SHL;
64 *exp = e - (LC3_IEEE754_EXP_BIAS - 1);
65
66 x.u = (x.u & ~LC3_IEEE754_EXP_MASK) |
67 ((LC3_IEEE754_EXP_BIAS - 1) << LC3_IEEE754_EXP_SHL);
68
69 return x.f;
70 }
71
72 /**
73 * Fast 2^n approximation
74 * x Operand, range -100 to 100
75 * return 2^x approximation (max relative error ~ 4e-7)
76 */
lc3_exp2f(float x)77 static inline float lc3_exp2f(float x)
78 {
79 /* --- 2^(i/8) for i from 0 to 7 --- */
80
81 static const float e[] = {
82 1.00000000e+00, 1.09050773e+00, 1.18920712e+00, 1.29683955e+00,
83 1.41421356e+00, 1.54221083e+00, 1.68179283e+00, 1.83400809e+00 };
84
85 /* --- Polynomial approx in range 0 to 1/8 --- */
86
87 static const float p[] = {
88 1.00448128e-02, 5.54563260e-02, 2.40228756e-01, 6.93147140e-01 };
89
90 /* --- Split the operand ---
91 *
92 * Such as x = k/8 + y, with k an integer, and |y| < 0.5/8
93 *
94 * Note that `fast-math` compiler option leads to rounding error,
95 * disable optimisation with `volatile`. */
96
97 volatile union { float f; int32_t s; } v;
98
99 v.f = x + 0x1.8p20f;
100 int k = v.s;
101 x -= v.f - 0x1.8p20f;
102
103 /* --- Compute 2^x, with |x| < 1 ---
104 * Perform polynomial approximation in range -0.5/8 to 0.5/8,
105 * and muplity by precomputed value of 2^(i/8), i in [0:7] */
106
107 union { float f; int32_t s; } y;
108
109 y.f = ( p[0]) * x;
110 y.f = (y.f + p[1]) * x;
111 y.f = (y.f + p[2]) * x;
112 y.f = (y.f + p[3]) * x;
113 y.f = (y.f + 1.f) * e[k & 7];
114
115 /* --- Add the exponent --- */
116
117 y.s += (k >> 3) << LC3_IEEE754_EXP_SHL;
118
119 return y.f;
120 }
121
122 /**
123 * Fast log2(x) approximation
124 * x Operand, greater than 0
125 * return log2(x) approximation (max absolute error ~ 1e-4)
126 */
lc3_log2f(float x)127 static inline float lc3_log2f(float x)
128 {
129 float y;
130 int e;
131
132 /* --- Polynomial approx in range 0.5 to 1 --- */
133
134 static const float c[] = {
135 -1.29479677, 5.11769018, -8.42295281, 8.10557963, -3.50567360 };
136
137 x = lc3_frexpf(x, &e);
138
139 y = ( c[0]) * x;
140 y = (y + c[1]) * x;
141 y = (y + c[2]) * x;
142 y = (y + c[3]) * x;
143 y = (y + c[4]);
144
145 /* --- Add log2f(2^e) and return --- */
146
147 return e + y;
148 }
149
150 /**
151 * Fast log10(x) approximation
152 * x Operand, greater than 0
153 * return log10(x) approximation (max absolute error ~ 1e-4)
154 */
lc3_log10f(float x)155 static inline float lc3_log10f(float x)
156 {
157 return log10f(2) * lc3_log2f(x);
158 }
159
160 /**
161 * Fast `10 * log10(x)` (or dB) approximation in fixed Q16
162 * x Operand, in range 2^-63 to 2^63 (1e-19 to 1e19)
163 * return 10 * log10(x) in fixed Q16 (-190 to 192 dB)
164 *
165 * - The 0 value is accepted and return the minimum value ~ -191dB
166 * - This function assumed that float 32 bits is coded IEEE 754
167 */
lc3_db_q16(float x)168 static inline int32_t lc3_db_q16(float x)
169 {
170 /* --- Table in Q15 --- */
171
172 static const uint16_t t[][2] = {
173
174 /* [n][0] = 10 * log10(2) * log2(1 + n/32), with n = [0..15] */
175 /* [n][1] = [n+1][0] - [n][0] (while defining [16][0]) */
176
177 { 0, 4379 }, { 4379, 4248 }, { 8627, 4125 }, { 12753, 4009 },
178 { 16762, 3899 }, { 20661, 3795 }, { 24456, 3697 }, { 28153, 3603 },
179 { 31755, 3514 }, { 35269, 3429 }, { 38699, 3349 }, { 42047, 3272 },
180 { 45319, 3198 }, { 48517, 3128 }, { 51645, 3061 }, { 54705, 2996 },
181
182 /* [n][0] = 10 * log10(2) * log2(1 + n/32) - 10 * log10(2) / 2, */
183 /* with n = [16..31] */
184 /* [n][1] = [n+1][0] - [n][0] (while defining [32][0]) */
185
186 { 8381, 2934 }, { 11315, 2875 }, { 14190, 2818 }, { 17008, 2763 },
187 { 19772, 2711 }, { 22482, 2660 }, { 25142, 2611 }, { 27754, 2564 },
188 { 30318, 2519 }, { 32837, 2475 }, { 35312, 2433 }, { 37744, 2392 },
189 { 40136, 2352 }, { 42489, 2314 }, { 44803, 2277 }, { 47080, 2241 },
190
191 };
192
193 /* --- Approximation ---
194 *
195 * 10 * log10(x^2) = 10 * log10(2) * log2(x^2)
196 *
197 * And log2(x^2) = 2 * log2( (1 + m) * 2^e )
198 * = 2 * (e + log2(1 + m)) , with m in range [0..1]
199 *
200 * Split the float values in :
201 * e2 Double value of the exponent (2 * e + k)
202 * hi High 5 bits of mantissa, for precalculated result `t[hi][0]`
203 * lo Low 16 bits of mantissa, for linear interpolation `t[hi][1]`
204 *
205 * Two cases, from the range of the mantissa :
206 * 0 to 0.5 `k = 0`, use 1st part of the table
207 * 0.5 to 1 `k = 1`, use 2nd part of the table */
208
209 union { float f; uint32_t u; } x2 = { .f = x*x };
210
211 int e2 = (int)(x2.u >> 22) - 2*127;
212 int hi = (x2.u >> 18) & 0x1f;
213 int lo = (x2.u >> 2) & 0xffff;
214
215 return e2 * 49321 + t[hi][0] + ((t[hi][1] * lo) >> 16);
216 }
217
218
219 #endif /* __LC3_FASTMATH_H */
220