1 | /* |
2 | * Single-precision pow function. |
3 | * |
4 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
5 | * See https://llvm.org/LICENSE.txt for license information. |
6 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
7 | */ |
8 | |
9 | #include <math.h> |
10 | #include <stdint.h> |
11 | #include "math_config.h" |
12 | |
13 | /* |
14 | POWF_LOG2_POLY_ORDER = 5 |
15 | EXP2F_TABLE_BITS = 5 |
16 | |
17 | ULP error: 0.82 (~ 0.5 + relerr*2^24) |
18 | relerr: 1.27 * 2^-26 (Relative error ~= 128*Ln2*relerr_log2 + relerr_exp2) |
19 | relerr_log2: 1.83 * 2^-33 (Relative error of logx.) |
20 | relerr_exp2: 1.69 * 2^-34 (Relative error of exp2(ylogx).) |
21 | */ |
22 | |
23 | #define N (1 << POWF_LOG2_TABLE_BITS) |
24 | #define T __powf_log2_data.tab |
25 | #define A __powf_log2_data.poly |
26 | #define OFF 0x3f330000 |
27 | |
28 | /* Subnormal input is normalized so ix has negative biased exponent. |
29 | Output is multiplied by N (POWF_SCALE) if TOINT_INTRINSICS is set. */ |
30 | static inline double_t |
31 | log2_inline (uint32_t ix) |
32 | { |
33 | /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */ |
34 | double_t z, r, r2, r4, p, q, y, y0, invc, logc; |
35 | uint32_t iz, top, tmp; |
36 | int k, i; |
37 | |
38 | /* x = 2^k z; where z is in range [OFF,2*OFF] and exact. |
39 | The range is split into N subintervals. |
40 | The ith subinterval contains z and c is near its center. */ |
41 | tmp = ix - OFF; |
42 | i = (tmp >> (23 - POWF_LOG2_TABLE_BITS)) % N; |
43 | top = tmp & 0xff800000; |
44 | iz = ix - top; |
45 | k = (int32_t) top >> (23 - POWF_SCALE_BITS); /* arithmetic shift */ |
46 | invc = T[i].invc; |
47 | logc = T[i].logc; |
48 | z = (double_t) asfloat (i: iz); |
49 | |
50 | /* log2(x) = log1p(z/c-1)/ln2 + log2(c) + k */ |
51 | r = z * invc - 1; |
52 | y0 = logc + (double_t) k; |
53 | |
54 | /* Pipelined polynomial evaluation to approximate log1p(r)/ln2. */ |
55 | r2 = r * r; |
56 | y = A[0] * r + A[1]; |
57 | p = A[2] * r + A[3]; |
58 | r4 = r2 * r2; |
59 | q = A[4] * r + y0; |
60 | q = p * r2 + q; |
61 | y = y * r4 + q; |
62 | return y; |
63 | } |
64 | |
65 | #undef N |
66 | #undef T |
67 | #define N (1 << EXP2F_TABLE_BITS) |
68 | #define T __exp2f_data.tab |
69 | #define SIGN_BIAS (1 << (EXP2F_TABLE_BITS + 11)) |
70 | |
71 | /* The output of log2 and thus the input of exp2 is either scaled by N |
72 | (in case of fast toint intrinsics) or not. The unscaled xd must be |
73 | in [-1021,1023], sign_bias sets the sign of the result. */ |
74 | static inline float |
75 | exp2_inline (double_t xd, uint32_t sign_bias) |
76 | { |
77 | uint64_t ki, ski, t; |
78 | /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */ |
79 | double_t kd, z, r, r2, y, s; |
80 | |
81 | #if TOINT_INTRINSICS |
82 | # define C __exp2f_data.poly_scaled |
83 | /* N*x = k + r with r in [-1/2, 1/2] */ |
84 | kd = roundtoint (xd); /* k */ |
85 | ki = converttoint (xd); |
86 | #else |
87 | # define C __exp2f_data.poly |
88 | # define SHIFT __exp2f_data.shift_scaled |
89 | /* x = k/N + r with r in [-1/(2N), 1/(2N)] */ |
90 | kd = eval_as_double (x: xd + SHIFT); |
91 | ki = asuint64 (f: kd); |
92 | kd -= SHIFT; /* k/N */ |
93 | #endif |
94 | r = xd - kd; |
95 | |
96 | /* exp2(x) = 2^(k/N) * 2^r ~= s * (C0*r^3 + C1*r^2 + C2*r + 1) */ |
97 | t = T[ki % N]; |
98 | ski = ki + sign_bias; |
99 | t += ski << (52 - EXP2F_TABLE_BITS); |
100 | s = asdouble (i: t); |
101 | z = C[0] * r + C[1]; |
102 | r2 = r * r; |
103 | y = C[2] * r + 1; |
104 | y = z * r2 + y; |
105 | y = y * s; |
106 | return eval_as_float (x: y); |
107 | } |
108 | |
109 | /* Returns 0 if not int, 1 if odd int, 2 if even int. The argument is |
110 | the bit representation of a non-zero finite floating-point value. */ |
111 | static inline int |
112 | checkint (uint32_t iy) |
113 | { |
114 | int e = iy >> 23 & 0xff; |
115 | if (e < 0x7f) |
116 | return 0; |
117 | if (e > 0x7f + 23) |
118 | return 2; |
119 | if (iy & ((1 << (0x7f + 23 - e)) - 1)) |
120 | return 0; |
121 | if (iy & (1 << (0x7f + 23 - e))) |
122 | return 1; |
123 | return 2; |
124 | } |
125 | |
126 | static inline int |
127 | zeroinfnan (uint32_t ix) |
128 | { |
129 | return 2 * ix - 1 >= 2u * 0x7f800000 - 1; |
130 | } |
131 | |
132 | float |
133 | powf (float x, float y) |
134 | { |
135 | uint32_t sign_bias = 0; |
136 | uint32_t ix, iy; |
137 | |
138 | ix = asuint (f: x); |
139 | iy = asuint (f: y); |
140 | if (unlikely (ix - 0x00800000 >= 0x7f800000 - 0x00800000 || zeroinfnan (iy))) |
141 | { |
142 | /* Either (x < 0x1p-126 or inf or nan) or (y is 0 or inf or nan). */ |
143 | if (unlikely (zeroinfnan (iy))) |
144 | { |
145 | if (2 * iy == 0) |
146 | return issignalingf_inline (x) ? x + y : 1.0f; |
147 | if (ix == 0x3f800000) |
148 | return issignalingf_inline (x: y) ? x + y : 1.0f; |
149 | if (2 * ix > 2u * 0x7f800000 || 2 * iy > 2u * 0x7f800000) |
150 | return x + y; |
151 | if (2 * ix == 2 * 0x3f800000) |
152 | return 1.0f; |
153 | if ((2 * ix < 2 * 0x3f800000) == !(iy & 0x80000000)) |
154 | return 0.0f; /* |x|<1 && y==inf or |x|>1 && y==-inf. */ |
155 | return y * y; |
156 | } |
157 | if (unlikely (zeroinfnan (ix))) |
158 | { |
159 | float_t x2 = x * x; |
160 | if (ix & 0x80000000 && checkint (iy) == 1) |
161 | { |
162 | x2 = -x2; |
163 | sign_bias = 1; |
164 | } |
165 | #if WANT_ERRNO |
166 | if (2 * ix == 0 && iy & 0x80000000) |
167 | return __math_divzerof (sign_bias); |
168 | #endif |
169 | /* Without the barrier some versions of clang hoist the 1/x2 and |
170 | thus division by zero exception can be signaled spuriously. */ |
171 | return iy & 0x80000000 ? opt_barrier_float (x: 1 / x2) : x2; |
172 | } |
173 | /* x and y are non-zero finite. */ |
174 | if (ix & 0x80000000) |
175 | { |
176 | /* Finite x < 0. */ |
177 | int yint = checkint (iy); |
178 | if (yint == 0) |
179 | return __math_invalidf (x); |
180 | if (yint == 1) |
181 | sign_bias = SIGN_BIAS; |
182 | ix &= 0x7fffffff; |
183 | } |
184 | if (ix < 0x00800000) |
185 | { |
186 | /* Normalize subnormal x so exponent becomes negative. */ |
187 | ix = asuint (f: x * 0x1p23f); |
188 | ix &= 0x7fffffff; |
189 | ix -= 23 << 23; |
190 | } |
191 | } |
192 | double_t logx = log2_inline (ix); |
193 | double_t ylogx = y * logx; /* Note: cannot overflow, y is single prec. */ |
194 | if (unlikely ((asuint64 (ylogx) >> 47 & 0xffff) |
195 | >= asuint64 (126.0 * POWF_SCALE) >> 47)) |
196 | { |
197 | /* |y*log(x)| >= 126. */ |
198 | if (ylogx > 0x1.fffffffd1d571p+6 * POWF_SCALE) |
199 | /* |x^y| > 0x1.ffffffp127. */ |
200 | return __math_oflowf (sign_bias); |
201 | if (WANT_ROUNDING && WANT_ERRNO |
202 | && ylogx > 0x1.fffffffa3aae2p+6 * POWF_SCALE) |
203 | /* |x^y| > 0x1.fffffep127, check if we round away from 0. */ |
204 | if ((!sign_bias |
205 | && eval_as_float (x: 1.0f + opt_barrier_float (x: 0x1p-25f)) != 1.0f) |
206 | || (sign_bias |
207 | && eval_as_float (x: -1.0f - opt_barrier_float (x: 0x1p-25f)) |
208 | != -1.0f)) |
209 | return __math_oflowf (sign_bias); |
210 | if (ylogx <= -150.0 * POWF_SCALE) |
211 | return __math_uflowf (sign_bias); |
212 | #if WANT_ERRNO_UFLOW |
213 | if (ylogx < -149.0 * POWF_SCALE) |
214 | return __math_may_uflowf (sign_bias); |
215 | #endif |
216 | } |
217 | return exp2_inline (xd: ylogx, sign_bias); |
218 | } |
219 | #if USE_GLIBC_ABI |
220 | strong_alias (powf, __powf_finite) |
221 | hidden_alias (powf, __ieee754_powf) |
222 | #endif |
223 | |