xref: /freebsd/contrib/arm-optimized-routines/pl/math/finite_pow.h (revision 6580f5c38dd5b01aeeaed16b370f1a12423437f0)
1 /*
2  * Double-precision x^y function.
3  *
4  * Copyright (c) 2018-2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "math_config.h"
9 
10 /* Scalar version of pow used for fallbacks in vector implementations.  */
11 
12 /* Data is defined in v_pow_log_data.c.  */
13 #define N_LOG (1 << V_POW_LOG_TABLE_BITS)
14 #define Off 0x3fe6955500000000
15 #define As __v_pow_log_data.poly
16 
17 /* Data is defined in v_pow_exp_data.c.  */
18 #define N_EXP (1 << V_POW_EXP_TABLE_BITS)
19 #define SignBias (0x800 << V_POW_EXP_TABLE_BITS)
20 #define SmallExp 0x3c9 /* top12(0x1p-54).  */
21 #define BigExp 0x408   /* top12(512.0).  */
22 #define ThresExp 0x03f /* BigExp - SmallExp.  */
23 #define InvLn2N __v_pow_exp_data.n_over_ln2
24 #define Ln2HiN __v_pow_exp_data.ln2_over_n_hi
25 #define Ln2LoN __v_pow_exp_data.ln2_over_n_lo
26 #define SBits __v_pow_exp_data.sbits
27 #define Cs __v_pow_exp_data.poly
28 
29 /* Constants associated with pow.  */
30 #define SmallPowX 0x001 /* top12(0x1p-126).  */
31 #define BigPowX 0x7ff	/* top12(INFINITY).  */
32 #define ThresPowX 0x7fe /* BigPowX - SmallPowX.  */
33 #define SmallPowY 0x3be /* top12(0x1.e7b6p-65).  */
34 #define BigPowY 0x43e	/* top12(0x1.749p62).  */
35 #define ThresPowY 0x080 /* BigPowY - SmallPowY.  */
36 
37 /* Top 12 bits of a double (sign and exponent bits).  */
38 static inline uint32_t
39 top12 (double x)
40 {
41   return asuint64 (x) >> 52;
42 }
43 
44 /* Compute y+TAIL = log(x) where the rounded result is y and TAIL has about
45    additional 15 bits precision.  IX is the bit representation of x, but
46    normalized in the subnormal range using the sign bit for the exponent.  */
47 static inline double
48 log_inline (uint64_t ix, double *tail)
49 {
50   /* x = 2^k z; where z is in range [Off,2*Off) and exact.
51      The range is split into N subintervals.
52      The ith subinterval contains z and c is near its center.  */
53   uint64_t tmp = ix - Off;
54   int i = (tmp >> (52 - V_POW_LOG_TABLE_BITS)) & (N_LOG - 1);
55   int k = (int64_t) tmp >> 52; /* arithmetic shift.  */
56   uint64_t iz = ix - (tmp & 0xfffULL << 52);
57   double z = asdouble (iz);
58   double kd = (double) k;
59 
60   /* log(x) = k*Ln2 + log(c) + log1p(z/c-1).  */
61   double invc = __v_pow_log_data.invc[i];
62   double logc = __v_pow_log_data.logc[i];
63   double logctail = __v_pow_log_data.logctail[i];
64 
65   /* Note: 1/c is j/N or j/N/2 where j is an integer in [N,2N) and
66      |z/c - 1| < 1/N, so r = z/c - 1 is exactly representible.  */
67   double r = fma (z, invc, -1.0);
68 
69   /* k*Ln2 + log(c) + r.  */
70   double t1 = kd * __v_pow_log_data.ln2_hi + logc;
71   double t2 = t1 + r;
72   double lo1 = kd * __v_pow_log_data.ln2_lo + logctail;
73   double lo2 = t1 - t2 + r;
74 
75   /* Evaluation is optimized assuming superscalar pipelined execution.  */
76   double ar = As[0] * r;
77   double ar2 = r * ar;
78   double ar3 = r * ar2;
79   /* k*Ln2 + log(c) + r + A[0]*r*r.  */
80   double hi = t2 + ar2;
81   double lo3 = fma (ar, r, -ar2);
82   double lo4 = t2 - hi + ar2;
83   /* p = log1p(r) - r - A[0]*r*r.  */
84   double p = (ar3
85 	      * (As[1] + r * As[2]
86 		 + ar2 * (As[3] + r * As[4] + ar2 * (As[5] + r * As[6]))));
87   double lo = lo1 + lo2 + lo3 + lo4 + p;
88   double y = hi + lo;
89   *tail = hi - y + lo;
90   return y;
91 }
92 
93 /* Handle cases that may overflow or underflow when computing the result that
94    is scale*(1+TMP) without intermediate rounding.  The bit representation of
95    scale is in SBITS, however it has a computed exponent that may have
96    overflown into the sign bit so that needs to be adjusted before using it as
97    a double.  (int32_t)KI is the k used in the argument reduction and exponent
98    adjustment of scale, positive k here means the result may overflow and
99    negative k means the result may underflow.  */
100 static inline double
101 special_case (double tmp, uint64_t sbits, uint64_t ki)
102 {
103   double scale, y;
104 
105   if ((ki & 0x80000000) == 0)
106     {
107       /* k > 0, the exponent of scale might have overflowed by <= 460.  */
108       sbits -= 1009ull << 52;
109       scale = asdouble (sbits);
110       y = 0x1p1009 * (scale + scale * tmp);
111       return check_oflow (eval_as_double (y));
112     }
113   /* k < 0, need special care in the subnormal range.  */
114   sbits += 1022ull << 52;
115   /* Note: sbits is signed scale.  */
116   scale = asdouble (sbits);
117   y = scale + scale * tmp;
118 #if WANT_SIMD_EXCEPT
119   if (fabs (y) < 1.0)
120     {
121       /* Round y to the right precision before scaling it into the subnormal
122 	 range to avoid double rounding that can cause 0.5+E/2 ulp error where
123 	 E is the worst-case ulp error outside the subnormal range.  So this
124 	 is only useful if the goal is better than 1 ulp worst-case error.  */
125       double hi, lo, one = 1.0;
126       if (y < 0.0)
127 	one = -1.0;
128       lo = scale - y + scale * tmp;
129       hi = one + y;
130       lo = one - hi + y + lo;
131       y = eval_as_double (hi + lo) - one;
132       /* Fix the sign of 0.  */
133       if (y == 0.0)
134 	y = asdouble (sbits & 0x8000000000000000);
135       /* The underflow exception needs to be signaled explicitly.  */
136       force_eval_double (opt_barrier_double (0x1p-1022) * 0x1p-1022);
137     }
138 #endif
139   y = 0x1p-1022 * y;
140   return check_uflow (eval_as_double (y));
141 }
142 
143 /* Computes sign*exp(x+xtail) where |xtail| < 2^-8/N and |xtail| <= |x|.
144    The sign_bias argument is SignBias or 0 and sets the sign to -1 or 1.  */
145 static inline double
146 exp_inline (double x, double xtail, uint32_t sign_bias)
147 {
148   uint32_t abstop = top12 (x) & 0x7ff;
149   if (unlikely (abstop - SmallExp >= ThresExp))
150     {
151       if (abstop - SmallExp >= 0x80000000)
152 	{
153 	  /* Avoid spurious underflow for tiny x.  */
154 	  /* Note: 0 is common input.  */
155 	  return sign_bias ? -1.0 : 1.0;
156 	}
157       if (abstop >= top12 (1024.0))
158 	{
159 	  /* Note: inf and nan are already handled.  */
160 	  /* Skip errno handling.  */
161 #if WANT_SIMD_EXCEPT
162 	  return asuint64 (x) >> 63 ? __math_uflow (sign_bias)
163 				    : __math_oflow (sign_bias);
164 #else
165 	  double res_uoflow = asuint64 (x) >> 63 ? 0.0 : INFINITY;
166 	  return sign_bias ? -res_uoflow : res_uoflow;
167 #endif
168 	}
169       /* Large x is special cased below.  */
170       abstop = 0;
171     }
172 
173   /* exp(x) = 2^(k/N) * exp(r), with exp(r) in [2^(-1/2N),2^(1/2N)].  */
174   /* x = ln2/N*k + r, with int k and r in [-ln2/2N, ln2/2N].  */
175   double z = InvLn2N * x;
176   double kd = round (z);
177   uint64_t ki = lround (z);
178   double r = x - kd * Ln2HiN - kd * Ln2LoN;
179   /* The code assumes 2^-200 < |xtail| < 2^-8/N.  */
180   r += xtail;
181   /* 2^(k/N) ~= scale.  */
182   uint64_t idx = ki & (N_EXP - 1);
183   uint64_t top = (ki + sign_bias) << (52 - V_POW_EXP_TABLE_BITS);
184   /* This is only a valid scale when -1023*N < k < 1024*N.  */
185   uint64_t sbits = SBits[idx] + top;
186   /* exp(x) = 2^(k/N) * exp(r) ~= scale + scale * (exp(r) - 1).  */
187   /* Evaluation is optimized assuming superscalar pipelined execution.  */
188   double r2 = r * r;
189   double tmp = r + r2 * Cs[0] + r * r2 * (Cs[1] + r * Cs[2]);
190   if (unlikely (abstop == 0))
191     return special_case (tmp, sbits, ki);
192   double scale = asdouble (sbits);
193   /* Note: tmp == 0 or |tmp| > 2^-200 and scale > 2^-739, so there
194      is no spurious underflow here even without fma.  */
195   return eval_as_double (scale + scale * tmp);
196 }
197 
198 /* Computes exp(x+xtail) where |xtail| < 2^-8/N and |xtail| <= |x|.
199    A version of exp_inline that is not inlined and for which sign_bias is
200    equal to 0.  */
201 static double NOINLINE
202 exp_nosignbias (double x, double xtail)
203 {
204   uint32_t abstop = top12 (x) & 0x7ff;
205   if (unlikely (abstop - SmallExp >= ThresExp))
206     {
207       /* Avoid spurious underflow for tiny x.  */
208       if (abstop - SmallExp >= 0x80000000)
209 	return 1.0;
210       /* Note: inf and nan are already handled.  */
211       if (abstop >= top12 (1024.0))
212 #if WANT_SIMD_EXCEPT
213 	return asuint64 (x) >> 63 ? __math_uflow (0) : __math_oflow (0);
214 #else
215 	return asuint64 (x) >> 63 ? 0.0 : INFINITY;
216 #endif
217       /* Large x is special cased below.  */
218       abstop = 0;
219     }
220 
221   /* exp(x) = 2^(k/N) * exp(r), with exp(r) in [2^(-1/2N),2^(1/2N)].  */
222   /* x = ln2/N*k + r, with k integer and r in [-ln2/2N, ln2/2N].  */
223   double z = InvLn2N * x;
224   double kd = round (z);
225   uint64_t ki = lround (z);
226   double r = x - kd * Ln2HiN - kd * Ln2LoN;
227   /* The code assumes 2^-200 < |xtail| < 2^-8/N.  */
228   r += xtail;
229   /* 2^(k/N) ~= scale.  */
230   uint64_t idx = ki & (N_EXP - 1);
231   uint64_t top = ki << (52 - V_POW_EXP_TABLE_BITS);
232   /* This is only a valid scale when -1023*N < k < 1024*N.  */
233   uint64_t sbits = SBits[idx] + top;
234   /* exp(x) = 2^(k/N) * exp(r) ~= scale + scale * (tail + exp(r) - 1).  */
235   double r2 = r * r;
236   double tmp = r + r2 * Cs[0] + r * r2 * (Cs[1] + r * Cs[2]);
237   if (unlikely (abstop == 0))
238     return special_case (tmp, sbits, ki);
239   double scale = asdouble (sbits);
240   /* Note: tmp == 0 or |tmp| > 2^-200 and scale > 2^-739, so there
241      is no spurious underflow here even without fma.  */
242   return eval_as_double (scale + scale * tmp);
243 }
244 
245 /* Returns 0 if not int, 1 if odd int, 2 if even int.  The argument is
246    the bit representation of a non-zero finite floating-point value.  */
247 static inline int
248 checkint (uint64_t iy)
249 {
250   int e = iy >> 52 & 0x7ff;
251   if (e < 0x3ff)
252     return 0;
253   if (e > 0x3ff + 52)
254     return 2;
255   if (iy & ((1ULL << (0x3ff + 52 - e)) - 1))
256     return 0;
257   if (iy & (1ULL << (0x3ff + 52 - e)))
258     return 1;
259   return 2;
260 }
261 
262 /* Returns 1 if input is the bit representation of 0, infinity or nan.  */
263 static inline int
264 zeroinfnan (uint64_t i)
265 {
266   return 2 * i - 1 >= 2 * asuint64 (INFINITY) - 1;
267 }
268 
269 static double NOINLINE
270 __pl_finite_pow (double x, double y)
271 {
272   uint32_t sign_bias = 0;
273   uint64_t ix, iy;
274   uint32_t topx, topy;
275 
276   ix = asuint64 (x);
277   iy = asuint64 (y);
278   topx = top12 (x);
279   topy = top12 (y);
280   if (unlikely (topx - SmallPowX >= ThresPowX
281 		|| (topy & 0x7ff) - SmallPowY >= ThresPowY))
282     {
283       /* Note: if |y| > 1075 * ln2 * 2^53 ~= 0x1.749p62 then pow(x,y) = inf/0
284 	 and if |y| < 2^-54 / 1075 ~= 0x1.e7b6p-65 then pow(x,y) = +-1.  */
285       /* Special cases: (x < 0x1p-126 or inf or nan) or
286 	 (|y| < 0x1p-65 or |y| >= 0x1p63 or nan).  */
287       if (unlikely (zeroinfnan (iy)))
288 	{
289 	  if (2 * iy == 0)
290 	    return issignaling_inline (x) ? x + y : 1.0;
291 	  if (ix == asuint64 (1.0))
292 	    return issignaling_inline (y) ? x + y : 1.0;
293 	  if (2 * ix > 2 * asuint64 (INFINITY)
294 	      || 2 * iy > 2 * asuint64 (INFINITY))
295 	    return x + y;
296 	  if (2 * ix == 2 * asuint64 (1.0))
297 	    return 1.0;
298 	  if ((2 * ix < 2 * asuint64 (1.0)) == !(iy >> 63))
299 	    return 0.0; /* |x|<1 && y==inf or |x|>1 && y==-inf.  */
300 	  return y * y;
301 	}
302       if (unlikely (zeroinfnan (ix)))
303 	{
304 	  double x2 = x * x;
305 	  if (ix >> 63 && checkint (iy) == 1)
306 	    {
307 	      x2 = -x2;
308 	      sign_bias = 1;
309 	    }
310 #if WANT_SIMD_EXCEPT
311 	  if (2 * ix == 0 && iy >> 63)
312 	    return __math_divzero (sign_bias);
313 #endif
314 	  /* Without the barrier some versions of clang hoist the 1/x2 and
315 	     thus division by zero exception can be signaled spuriously.  */
316 	  return iy >> 63 ? opt_barrier_double (1 / x2) : x2;
317 	}
318       /* Here x and y are non-zero finite.  */
319       if (ix >> 63)
320 	{
321 	  /* Finite x < 0.  */
322 	  int yint = checkint (iy);
323 	  if (yint == 0)
324 #if WANT_SIMD_EXCEPT
325 	    return __math_invalid (x);
326 #else
327 	    return __builtin_nan ("");
328 #endif
329 	  if (yint == 1)
330 	    sign_bias = SignBias;
331 	  ix &= 0x7fffffffffffffff;
332 	  topx &= 0x7ff;
333 	}
334       if ((topy & 0x7ff) - SmallPowY >= ThresPowY)
335 	{
336 	  /* Note: sign_bias == 0 here because y is not odd.  */
337 	  if (ix == asuint64 (1.0))
338 	    return 1.0;
339 	  /* |y| < 2^-65, x^y ~= 1 + y*log(x).  */
340 	  if ((topy & 0x7ff) < SmallPowY)
341 	    return 1.0;
342 #if WANT_SIMD_EXCEPT
343 	  return (ix > asuint64 (1.0)) == (topy < 0x800) ? __math_oflow (0)
344 							 : __math_uflow (0);
345 #else
346 	  return (ix > asuint64 (1.0)) == (topy < 0x800) ? INFINITY : 0;
347 #endif
348 	}
349       if (topx == 0)
350 	{
351 	  /* Normalize subnormal x so exponent becomes negative.  */
352 	  /* Without the barrier some versions of clang evalutate the mul
353 	     unconditionally causing spurious overflow exceptions.  */
354 	  ix = asuint64 (opt_barrier_double (x) * 0x1p52);
355 	  ix &= 0x7fffffffffffffff;
356 	  ix -= 52ULL << 52;
357 	}
358     }
359 
360   double lo;
361   double hi = log_inline (ix, &lo);
362   double ehi = y * hi;
363   double elo = y * lo + fma (y, hi, -ehi);
364   return exp_inline (ehi, elo, sign_bias);
365 }
366