1 | /* |
2 | * Double-precision x^y function. |
3 | * |
4 | * Copyright (c) 2018, Arm Limited. |
5 | * SPDX-License-Identifier: MIT |
6 | */ |
7 | |
8 | #include <math.h> |
9 | #include <stdint.h> |
10 | #include "libm.h" |
11 | #include "exp_data.h" |
12 | #include "pow_data.h" |
13 | |
14 | /* |
15 | Worst-case error: 0.54 ULP (~= ulperr_exp + 1024*Ln2*relerr_log*2^53) |
16 | relerr_log: 1.3 * 2^-68 (Relative error of log, 1.5 * 2^-68 without fma) |
17 | ulperr_exp: 0.509 ULP (ULP error of exp, 0.511 ULP without fma) |
18 | */ |
19 | |
20 | #define T __pow_log_data.tab |
21 | #define A __pow_log_data.poly |
22 | #define Ln2hi __pow_log_data.ln2hi |
23 | #define Ln2lo __pow_log_data.ln2lo |
24 | #define N (1 << POW_LOG_TABLE_BITS) |
25 | #define OFF 0x3fe6955500000000 |
26 | |
27 | /* Top 12 bits of a double (sign and exponent bits). */ |
28 | static inline uint32_t top12(double x) |
29 | { |
30 | return asuint64(x) >> 52; |
31 | } |
32 | |
33 | /* Compute y+TAIL = log(x) where the rounded result is y and TAIL has about |
34 | additional 15 bits precision. IX is the bit representation of x, but |
35 | normalized in the subnormal range using the sign bit for the exponent. */ |
36 | static inline double_t log_inline(uint64_t ix, double_t *tail) |
37 | { |
38 | /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */ |
39 | double_t z, r, y, invc, logc, logctail, kd, hi, t1, t2, lo, lo1, lo2, p; |
40 | uint64_t iz, tmp; |
41 | int k, i; |
42 | |
43 | /* x = 2^k z; where z is in range [OFF,2*OFF) and exact. |
44 | The range is split into N subintervals. |
45 | The ith subinterval contains z and c is near its center. */ |
46 | tmp = ix - OFF; |
47 | i = (tmp >> (52 - POW_LOG_TABLE_BITS)) % N; |
48 | k = (int64_t)tmp >> 52; /* arithmetic shift */ |
49 | iz = ix - (tmp & 0xfffULL << 52); |
50 | z = asdouble(iz); |
51 | kd = (double_t)k; |
52 | |
53 | /* log(x) = k*Ln2 + log(c) + log1p(z/c-1). */ |
54 | invc = T[i].invc; |
55 | logc = T[i].logc; |
56 | logctail = T[i].logctail; |
57 | |
58 | /* Note: 1/c is j/N or j/N/2 where j is an integer in [N,2N) and |
59 | |z/c - 1| < 1/N, so r = z/c - 1 is exactly representible. */ |
60 | #if __FP_FAST_FMA |
61 | r = __builtin_fma(z, invc, -1.0); |
62 | #else |
63 | /* Split z such that rhi, rlo and rhi*rhi are exact and |rlo| <= |r|. */ |
64 | double_t zhi = asdouble((iz + (1ULL << 31)) & (-1ULL << 32)); |
65 | double_t zlo = z - zhi; |
66 | double_t rhi = zhi * invc - 1.0; |
67 | double_t rlo = zlo * invc; |
68 | r = rhi + rlo; |
69 | #endif |
70 | |
71 | /* k*Ln2 + log(c) + r. */ |
72 | t1 = kd * Ln2hi + logc; |
73 | t2 = t1 + r; |
74 | lo1 = kd * Ln2lo + logctail; |
75 | lo2 = t1 - t2 + r; |
76 | |
77 | /* Evaluation is optimized assuming superscalar pipelined execution. */ |
78 | double_t ar, ar2, ar3, lo3, lo4; |
79 | ar = A[0] * r; /* A[0] = -0.5. */ |
80 | ar2 = r * ar; |
81 | ar3 = r * ar2; |
82 | /* k*Ln2 + log(c) + r + A[0]*r*r. */ |
83 | #if __FP_FAST_FMA |
84 | hi = t2 + ar2; |
85 | lo3 = __builtin_fma(ar, r, -ar2); |
86 | lo4 = t2 - hi + ar2; |
87 | #else |
88 | double_t arhi = A[0] * rhi; |
89 | double_t arhi2 = rhi * arhi; |
90 | hi = t2 + arhi2; |
91 | lo3 = rlo * (ar + arhi); |
92 | lo4 = t2 - hi + arhi2; |
93 | #endif |
94 | /* p = log1p(r) - r - A[0]*r*r. */ |
95 | p = (ar3 * (A[1] + r * A[2] + |
96 | ar2 * (A[3] + r * A[4] + ar2 * (A[5] + r * A[6])))); |
97 | lo = lo1 + lo2 + lo3 + lo4 + p; |
98 | y = hi + lo; |
99 | *tail = hi - y + lo; |
100 | return y; |
101 | } |
102 | |
103 | #undef N |
104 | #undef T |
105 | #define N (1 << EXP_TABLE_BITS) |
106 | #define InvLn2N __exp_data.invln2N |
107 | #define NegLn2hiN __exp_data.negln2hiN |
108 | #define NegLn2loN __exp_data.negln2loN |
109 | #define Shift __exp_data.shift |
110 | #define T __exp_data.tab |
111 | #define C2 __exp_data.poly[5 - EXP_POLY_ORDER] |
112 | #define C3 __exp_data.poly[6 - EXP_POLY_ORDER] |
113 | #define C4 __exp_data.poly[7 - EXP_POLY_ORDER] |
114 | #define C5 __exp_data.poly[8 - EXP_POLY_ORDER] |
115 | #define C6 __exp_data.poly[9 - EXP_POLY_ORDER] |
116 | |
117 | /* Handle cases that may overflow or underflow when computing the result that |
118 | is scale*(1+TMP) without intermediate rounding. The bit representation of |
119 | scale is in SBITS, however it has a computed exponent that may have |
120 | overflown into the sign bit so that needs to be adjusted before using it as |
121 | a double. (int32_t)KI is the k used in the argument reduction and exponent |
122 | adjustment of scale, positive k here means the result may overflow and |
123 | negative k means the result may underflow. */ |
124 | static inline double specialcase(double_t tmp, uint64_t sbits, uint64_t ki) |
125 | { |
126 | double_t scale, y; |
127 | |
128 | if ((ki & 0x80000000) == 0) { |
129 | /* k > 0, the exponent of scale might have overflowed by <= 460. */ |
130 | sbits -= 1009ull << 52; |
131 | scale = asdouble(sbits); |
132 | y = 0x1p1009 * (scale + scale * tmp); |
133 | return eval_as_double(y); |
134 | } |
135 | /* k < 0, need special care in the subnormal range. */ |
136 | sbits += 1022ull << 52; |
137 | /* Note: sbits is signed scale. */ |
138 | scale = asdouble(sbits); |
139 | y = scale + scale * tmp; |
140 | if (fabs(y) < 1.0) { |
141 | /* Round y to the right precision before scaling it into the subnormal |
142 | range to avoid double rounding that can cause 0.5+E/2 ulp error where |
143 | E is the worst-case ulp error outside the subnormal range. So this |
144 | is only useful if the goal is better than 1 ulp worst-case error. */ |
145 | double_t hi, lo, one = 1.0; |
146 | if (y < 0.0) |
147 | one = -1.0; |
148 | lo = scale - y + scale * tmp; |
149 | hi = one + y; |
150 | lo = one - hi + y + lo; |
151 | y = eval_as_double(hi + lo) - one; |
152 | /* Fix the sign of 0. */ |
153 | if (y == 0.0) |
154 | y = asdouble(sbits & 0x8000000000000000); |
155 | /* The underflow exception needs to be signaled explicitly. */ |
156 | fp_force_eval(fp_barrier(0x1p-1022) * 0x1p-1022); |
157 | } |
158 | y = 0x1p-1022 * y; |
159 | return eval_as_double(y); |
160 | } |
161 | |
162 | #define SIGN_BIAS (0x800 << EXP_TABLE_BITS) |
163 | |
164 | /* Computes sign*exp(x+xtail) where |xtail| < 2^-8/N and |xtail| <= |x|. |
165 | The sign_bias argument is SIGN_BIAS or 0 and sets the sign to -1 or 1. */ |
166 | static inline double exp_inline(double_t x, double_t xtail, uint32_t sign_bias) |
167 | { |
168 | uint32_t abstop; |
169 | uint64_t ki, idx, top, sbits; |
170 | /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */ |
171 | double_t kd, z, r, r2, scale, tail, tmp; |
172 | |
173 | abstop = top12(x) & 0x7ff; |
174 | if (predict_false(abstop - top12(0x1p-54) >= |
175 | top12(512.0) - top12(0x1p-54))) { |
176 | if (abstop - top12(0x1p-54) >= 0x80000000) { |
177 | /* Avoid spurious underflow for tiny x. */ |
178 | /* Note: 0 is common input. */ |
179 | double_t one = WANT_ROUNDING ? 1.0 + x : 1.0; |
180 | return sign_bias ? -one : one; |
181 | } |
182 | if (abstop >= top12(1024.0)) { |
183 | /* Note: inf and nan are already handled. */ |
184 | if (asuint64(x) >> 63) |
185 | return __math_uflow(sign_bias); |
186 | else |
187 | return __math_oflow(sign_bias); |
188 | } |
189 | /* Large x is special cased below. */ |
190 | abstop = 0; |
191 | } |
192 | |
193 | /* exp(x) = 2^(k/N) * exp(r), with exp(r) in [2^(-1/2N),2^(1/2N)]. */ |
194 | /* x = ln2/N*k + r, with int k and r in [-ln2/2N, ln2/2N]. */ |
195 | z = InvLn2N * x; |
196 | #if TOINT_INTRINSICS |
197 | kd = roundtoint(z); |
198 | ki = converttoint(z); |
199 | #elif EXP_USE_TOINT_NARROW |
200 | /* z - kd is in [-0.5-2^-16, 0.5] in all rounding modes. */ |
201 | kd = eval_as_double(z + Shift); |
202 | ki = asuint64(kd) >> 16; |
203 | kd = (double_t)(int32_t)ki; |
204 | #else |
205 | /* z - kd is in [-1, 1] in non-nearest rounding modes. */ |
206 | kd = eval_as_double(z + Shift); |
207 | ki = asuint64(kd); |
208 | kd -= Shift; |
209 | #endif |
210 | r = x + kd * NegLn2hiN + kd * NegLn2loN; |
211 | /* The code assumes 2^-200 < |xtail| < 2^-8/N. */ |
212 | r += xtail; |
213 | /* 2^(k/N) ~= scale * (1 + tail). */ |
214 | idx = 2 * (ki % N); |
215 | top = (ki + sign_bias) << (52 - EXP_TABLE_BITS); |
216 | tail = asdouble(T[idx]); |
217 | /* This is only a valid scale when -1023*N < k < 1024*N. */ |
218 | sbits = T[idx + 1] + top; |
219 | /* exp(x) = 2^(k/N) * exp(r) ~= scale + scale * (tail + exp(r) - 1). */ |
220 | /* Evaluation is optimized assuming superscalar pipelined execution. */ |
221 | r2 = r * r; |
222 | /* Without fma the worst case error is 0.25/N ulp larger. */ |
223 | /* Worst case error is less than 0.5+1.11/N+(abs poly error * 2^53) ulp. */ |
224 | tmp = tail + r + r2 * (C2 + r * C3) + r2 * r2 * (C4 + r * C5); |
225 | if (predict_false(abstop == 0)) |
226 | return specialcase(tmp, sbits, ki); |
227 | scale = asdouble(sbits); |
228 | /* Note: tmp == 0 or |tmp| > 2^-200 and scale > 2^-739, so there |
229 | is no spurious underflow here even without fma. */ |
230 | return eval_as_double(scale + scale * tmp); |
231 | } |
232 | |
233 | /* Returns 0 if not int, 1 if odd int, 2 if even int. The argument is |
234 | the bit representation of a non-zero finite floating-point value. */ |
235 | static inline int checkint(uint64_t iy) |
236 | { |
237 | int e = iy >> 52 & 0x7ff; |
238 | if (e < 0x3ff) |
239 | return 0; |
240 | if (e > 0x3ff + 52) |
241 | return 2; |
242 | if (iy & ((1ULL << (0x3ff + 52 - e)) - 1)) |
243 | return 0; |
244 | if (iy & (1ULL << (0x3ff + 52 - e))) |
245 | return 1; |
246 | return 2; |
247 | } |
248 | |
249 | /* Returns 1 if input is the bit representation of 0, infinity or nan. */ |
250 | static inline int zeroinfnan(uint64_t i) |
251 | { |
252 | return 2 * i - 1 >= 2 * asuint64(INFINITY) - 1; |
253 | } |
254 | |
255 | double pow(double x, double y) |
256 | { |
257 | uint32_t sign_bias = 0; |
258 | uint64_t ix, iy; |
259 | uint32_t topx, topy; |
260 | |
261 | ix = asuint64(x); |
262 | iy = asuint64(y); |
263 | topx = top12(x); |
264 | topy = top12(y); |
265 | if (predict_false(topx - 0x001 >= 0x7ff - 0x001 || |
266 | (topy & 0x7ff) - 0x3be >= 0x43e - 0x3be)) { |
267 | /* Note: if |y| > 1075 * ln2 * 2^53 ~= 0x1.749p62 then pow(x,y) = inf/0 |
268 | and if |y| < 2^-54 / 1075 ~= 0x1.e7b6p-65 then pow(x,y) = +-1. */ |
269 | /* Special cases: (x < 0x1p-126 or inf or nan) or |
270 | (|y| < 0x1p-65 or |y| >= 0x1p63 or nan). */ |
271 | if (predict_false(zeroinfnan(iy))) { |
272 | if (2 * iy == 0) |
273 | return issignaling_inline(x) ? x + y : 1.0; |
274 | if (ix == asuint64(1.0)) |
275 | return issignaling_inline(y) ? x + y : 1.0; |
276 | if (2 * ix > 2 * asuint64(INFINITY) || |
277 | 2 * iy > 2 * asuint64(INFINITY)) |
278 | return x + y; |
279 | if (2 * ix == 2 * asuint64(1.0)) |
280 | return 1.0; |
281 | if ((2 * ix < 2 * asuint64(1.0)) == !(iy >> 63)) |
282 | return 0.0; /* |x|<1 && y==inf or |x|>1 && y==-inf. */ |
283 | return y * y; |
284 | } |
285 | if (predict_false(zeroinfnan(ix))) { |
286 | double_t x2 = x * x; |
287 | if (ix >> 63 && checkint(iy) == 1) |
288 | x2 = -x2; |
289 | /* Without the barrier some versions of clang hoist the 1/x2 and |
290 | thus division by zero exception can be signaled spuriously. */ |
291 | return iy >> 63 ? fp_barrier(1 / x2) : x2; |
292 | } |
293 | /* Here x and y are non-zero finite. */ |
294 | if (ix >> 63) { |
295 | /* Finite x < 0. */ |
296 | int yint = checkint(iy); |
297 | if (yint == 0) |
298 | return __math_invalid(x); |
299 | if (yint == 1) |
300 | sign_bias = SIGN_BIAS; |
301 | ix &= 0x7fffffffffffffff; |
302 | topx &= 0x7ff; |
303 | } |
304 | if ((topy & 0x7ff) - 0x3be >= 0x43e - 0x3be) { |
305 | /* Note: sign_bias == 0 here because y is not odd. */ |
306 | if (ix == asuint64(1.0)) |
307 | return 1.0; |
308 | if ((topy & 0x7ff) < 0x3be) { |
309 | /* |y| < 2^-65, x^y ~= 1 + y*log(x). */ |
310 | if (WANT_ROUNDING) |
311 | return ix > asuint64(1.0) ? 1.0 + y : |
312 | 1.0 - y; |
313 | else |
314 | return 1.0; |
315 | } |
316 | return (ix > asuint64(1.0)) == (topy < 0x800) ? |
317 | __math_oflow(0) : |
318 | __math_uflow(0); |
319 | } |
320 | if (topx == 0) { |
321 | /* Normalize subnormal x so exponent becomes negative. */ |
322 | ix = asuint64(x * 0x1p52); |
323 | ix &= 0x7fffffffffffffff; |
324 | ix -= 52ULL << 52; |
325 | } |
326 | } |
327 | |
328 | double_t lo; |
329 | double_t hi = log_inline(ix, &lo); |
330 | double_t ehi, elo; |
331 | #if __FP_FAST_FMA |
332 | ehi = y * hi; |
333 | elo = y * lo + __builtin_fma(y, hi, -ehi); |
334 | #else |
335 | double_t yhi = asdouble(iy & -1ULL << 27); |
336 | double_t ylo = y - yhi; |
337 | double_t lhi = asdouble(asuint64(hi) & -1ULL << 27); |
338 | double_t llo = hi - lhi + lo; |
339 | ehi = yhi * lhi; |
340 | elo = ylo * lhi + y * llo; /* |elo| < |ehi| * 2^-25. */ |
341 | #endif |
342 | return exp_inline(ehi, elo, sign_bias); |
343 | } |
344 | |