xref: /freebsd/contrib/arm-optimized-routines/math/powf.c (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57)
1 /*
2  * Single-precision pow function.
3  *
4  * Copyright (c) 2017-2019, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include <math.h>
9 #include <stdint.h>
10 #include "math_config.h"
11 
12 /*
13 POWF_LOG2_POLY_ORDER = 5
14 EXP2F_TABLE_BITS = 5
15 
16 ULP error: 0.82 (~ 0.5 + relerr*2^24)
17 relerr: 1.27 * 2^-26 (Relative error ~= 128*Ln2*relerr_log2 + relerr_exp2)
18 relerr_log2: 1.83 * 2^-33 (Relative error of logx.)
19 relerr_exp2: 1.69 * 2^-34 (Relative error of exp2(ylogx).)
20 */
21 
22 #define N (1 << POWF_LOG2_TABLE_BITS)
23 #define T __powf_log2_data.tab
24 #define A __powf_log2_data.poly
25 #define OFF 0x3f330000
26 
27 /* Subnormal input is normalized so ix has negative biased exponent.
28    Output is multiplied by N (POWF_SCALE) if TOINT_INTRINICS is set.  */
29 static inline double_t
30 log2_inline (uint32_t ix)
31 {
32   /* double_t for better performance on targets with FLT_EVAL_METHOD==2.  */
33   double_t z, r, r2, r4, p, q, y, y0, invc, logc;
34   uint32_t iz, top, tmp;
35   int k, i;
36 
37   /* x = 2^k z; where z is in range [OFF,2*OFF] and exact.
38      The range is split into N subintervals.
39      The ith subinterval contains z and c is near its center.  */
40   tmp = ix - OFF;
41   i = (tmp >> (23 - POWF_LOG2_TABLE_BITS)) % N;
42   top = tmp & 0xff800000;
43   iz = ix - top;
44   k = (int32_t) top >> (23 - POWF_SCALE_BITS); /* arithmetic shift */
45   invc = T[i].invc;
46   logc = T[i].logc;
47   z = (double_t) asfloat (iz);
48 
49   /* log2(x) = log1p(z/c-1)/ln2 + log2(c) + k */
50   r = z * invc - 1;
51   y0 = logc + (double_t) k;
52 
53   /* Pipelined polynomial evaluation to approximate log1p(r)/ln2.  */
54   r2 = r * r;
55   y = A[0] * r + A[1];
56   p = A[2] * r + A[3];
57   r4 = r2 * r2;
58   q = A[4] * r + y0;
59   q = p * r2 + q;
60   y = y * r4 + q;
61   return y;
62 }
63 
64 #undef N
65 #undef T
66 #define N (1 << EXP2F_TABLE_BITS)
67 #define T __exp2f_data.tab
68 #define SIGN_BIAS (1 << (EXP2F_TABLE_BITS + 11))
69 
70 /* The output of log2 and thus the input of exp2 is either scaled by N
71    (in case of fast toint intrinsics) or not.  The unscaled xd must be
72    in [-1021,1023], sign_bias sets the sign of the result.  */
73 static inline float
74 exp2_inline (double_t xd, uint32_t sign_bias)
75 {
76   uint64_t ki, ski, t;
77   /* double_t for better performance on targets with FLT_EVAL_METHOD==2.  */
78   double_t kd, z, r, r2, y, s;
79 
80 #if TOINT_INTRINSICS
81 # define C __exp2f_data.poly_scaled
82   /* N*x = k + r with r in [-1/2, 1/2] */
83   kd = roundtoint (xd); /* k */
84   ki = converttoint (xd);
85 #else
86 # define C __exp2f_data.poly
87 # define SHIFT __exp2f_data.shift_scaled
88   /* x = k/N + r with r in [-1/(2N), 1/(2N)] */
89   kd = eval_as_double (xd + SHIFT);
90   ki = asuint64 (kd);
91   kd -= SHIFT; /* k/N */
92 #endif
93   r = xd - kd;
94 
95   /* exp2(x) = 2^(k/N) * 2^r ~= s * (C0*r^3 + C1*r^2 + C2*r + 1) */
96   t = T[ki % N];
97   ski = ki + sign_bias;
98   t += ski << (52 - EXP2F_TABLE_BITS);
99   s = asdouble (t);
100   z = C[0] * r + C[1];
101   r2 = r * r;
102   y = C[2] * r + 1;
103   y = z * r2 + y;
104   y = y * s;
105   return eval_as_float (y);
106 }
107 
108 /* Returns 0 if not int, 1 if odd int, 2 if even int.  The argument is
109    the bit representation of a non-zero finite floating-point value.  */
110 static inline int
111 checkint (uint32_t iy)
112 {
113   int e = iy >> 23 & 0xff;
114   if (e < 0x7f)
115     return 0;
116   if (e > 0x7f + 23)
117     return 2;
118   if (iy & ((1 << (0x7f + 23 - e)) - 1))
119     return 0;
120   if (iy & (1 << (0x7f + 23 - e)))
121     return 1;
122   return 2;
123 }
124 
125 static inline int
126 zeroinfnan (uint32_t ix)
127 {
128   return 2 * ix - 1 >= 2u * 0x7f800000 - 1;
129 }
130 
131 float
132 powf (float x, float y)
133 {
134   uint32_t sign_bias = 0;
135   uint32_t ix, iy;
136 
137   ix = asuint (x);
138   iy = asuint (y);
139   if (unlikely (ix - 0x00800000 >= 0x7f800000 - 0x00800000 || zeroinfnan (iy)))
140     {
141       /* Either (x < 0x1p-126 or inf or nan) or (y is 0 or inf or nan).  */
142       if (unlikely (zeroinfnan (iy)))
143 	{
144 	  if (2 * iy == 0)
145 	    return issignalingf_inline (x) ? x + y : 1.0f;
146 	  if (ix == 0x3f800000)
147 	    return issignalingf_inline (y) ? x + y : 1.0f;
148 	  if (2 * ix > 2u * 0x7f800000 || 2 * iy > 2u * 0x7f800000)
149 	    return x + y;
150 	  if (2 * ix == 2 * 0x3f800000)
151 	    return 1.0f;
152 	  if ((2 * ix < 2 * 0x3f800000) == !(iy & 0x80000000))
153 	    return 0.0f; /* |x|<1 && y==inf or |x|>1 && y==-inf.  */
154 	  return y * y;
155 	}
156       if (unlikely (zeroinfnan (ix)))
157 	{
158 	  float_t x2 = x * x;
159 	  if (ix & 0x80000000 && checkint (iy) == 1)
160 	    {
161 	      x2 = -x2;
162 	      sign_bias = 1;
163 	    }
164 #if WANT_ERRNO
165 	  if (2 * ix == 0 && iy & 0x80000000)
166 	    return __math_divzerof (sign_bias);
167 #endif
168 	  /* Without the barrier some versions of clang hoist the 1/x2 and
169 	     thus division by zero exception can be signaled spuriously.  */
170 	  return iy & 0x80000000 ? opt_barrier_float (1 / x2) : x2;
171 	}
172       /* x and y are non-zero finite.  */
173       if (ix & 0x80000000)
174 	{
175 	  /* Finite x < 0.  */
176 	  int yint = checkint (iy);
177 	  if (yint == 0)
178 	    return __math_invalidf (x);
179 	  if (yint == 1)
180 	    sign_bias = SIGN_BIAS;
181 	  ix &= 0x7fffffff;
182 	}
183       if (ix < 0x00800000)
184 	{
185 	  /* Normalize subnormal x so exponent becomes negative.  */
186 	  ix = asuint (x * 0x1p23f);
187 	  ix &= 0x7fffffff;
188 	  ix -= 23 << 23;
189 	}
190     }
191   double_t logx = log2_inline (ix);
192   double_t ylogx = y * logx; /* Note: cannot overflow, y is single prec.  */
193   if (unlikely ((asuint64 (ylogx) >> 47 & 0xffff)
194 		 >= asuint64 (126.0 * POWF_SCALE) >> 47))
195     {
196       /* |y*log(x)| >= 126.  */
197       if (ylogx > 0x1.fffffffd1d571p+6 * POWF_SCALE)
198 	/* |x^y| > 0x1.ffffffp127.  */
199 	return __math_oflowf (sign_bias);
200       if (WANT_ROUNDING && WANT_ERRNO
201 	  && ylogx > 0x1.fffffffa3aae2p+6 * POWF_SCALE)
202 	/* |x^y| > 0x1.fffffep127, check if we round away from 0.  */
203 	if ((!sign_bias
204 	     && eval_as_float (1.0f + opt_barrier_float (0x1p-25f)) != 1.0f)
205 	    || (sign_bias
206 		&& eval_as_float (-1.0f - opt_barrier_float (0x1p-25f))
207 		     != -1.0f))
208 	  return __math_oflowf (sign_bias);
209       if (ylogx <= -150.0 * POWF_SCALE)
210 	return __math_uflowf (sign_bias);
211 #if WANT_ERRNO_UFLOW
212       if (ylogx < -149.0 * POWF_SCALE)
213 	return __math_may_uflowf (sign_bias);
214 #endif
215     }
216   return exp2_inline (ylogx, sign_bias);
217 }
218 #if USE_GLIBC_ABI
219 strong_alias (powf, __powf_finite)
220 hidden_alias (powf, __ieee754_powf)
221 #endif
222