• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===-- Double-precision 2^x function -------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include "src/math/exp2.h"
10 #include "common_constants.h" // Lookup tables EXP2_MID1 and EXP_M2.
11 #include "explogxf.h"         // ziv_test_denorm.
12 #include "src/__support/CPP/bit.h"
13 #include "src/__support/CPP/optional.h"
14 #include "src/__support/FPUtil/FEnvImpl.h"
15 #include "src/__support/FPUtil/FPBits.h"
16 #include "src/__support/FPUtil/PolyEval.h"
17 #include "src/__support/FPUtil/double_double.h"
18 #include "src/__support/FPUtil/dyadic_float.h"
19 #include "src/__support/FPUtil/multiply_add.h"
20 #include "src/__support/FPUtil/nearest_integer.h"
21 #include "src/__support/FPUtil/rounding_mode.h"
22 #include "src/__support/FPUtil/triple_double.h"
23 #include "src/__support/common.h"
24 #include "src/__support/integer_literals.h"
25 #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY
26 
27 #include <errno.h>
28 
29 namespace LIBC_NAMESPACE {
30 
31 using fputil::DoubleDouble;
32 using fputil::TripleDouble;
33 using Float128 = typename fputil::DyadicFloat<128>;
34 
35 using LIBC_NAMESPACE::operator""_u128;
36 
37 // Error bounds:
38 // Errors when using double precision.
39 #ifdef LIBC_TARGET_CPU_HAS_FMA
40 constexpr double ERR_D = 0x1.0p-63;
41 #else
42 constexpr double ERR_D = 0x1.8p-63;
43 #endif // LIBC_TARGET_CPU_HAS_FMA
44 
45 // Errors when using double-double precision.
46 constexpr double ERR_DD = 0x1.0p-100;
47 
48 namespace {
49 
50 // Polynomial approximations with double precision.  Generated by Sollya with:
51 // > P = fpminimax((2^x - 1)/x, 3, [|D...|], [-2^-13 - 2^-30, 2^-13 + 2^-30]);
52 // > P;
53 // Error bounds:
54 //   | output - (2^dx - 1) / dx | < 1.5 * 2^-52.
poly_approx_d(double dx)55 LIBC_INLINE double poly_approx_d(double dx) {
56   // dx^2
57   double dx2 = dx * dx;
58   double c0 =
59       fputil::multiply_add(dx, 0x1.ebfbdff82c58ep-3, 0x1.62e42fefa39efp-1);
60   double c1 =
61       fputil::multiply_add(dx, 0x1.3b2aba7a95a89p-7, 0x1.c6b08e8fc0c0ep-5);
62   double p = fputil::multiply_add(dx2, c1, c0);
63   return p;
64 }
65 
66 // Polynomial approximation with double-double precision.  Generated by Solya
67 // with:
68 // > P = fpminimax((2^x - 1)/x, 5, [|DD...|], [-2^-13 - 2^-30, 2^-13 + 2^-30]);
69 // Error bounds:
70 //   | output - 2^(dx) | < 2^-101
poly_approx_dd(const DoubleDouble & dx)71 DoubleDouble poly_approx_dd(const DoubleDouble &dx) {
72   // Taylor polynomial.
73   constexpr DoubleDouble COEFFS[] = {
74       {0, 0x1p0},
75       {0x1.abc9e3b39824p-56, 0x1.62e42fefa39efp-1},
76       {-0x1.5e43a53e4527bp-57, 0x1.ebfbdff82c58fp-3},
77       {-0x1.d37963a9444eep-59, 0x1.c6b08d704a0cp-5},
78       {0x1.4eda1a81133dap-62, 0x1.3b2ab6fba4e77p-7},
79       {-0x1.c53fd1ba85d14p-64, 0x1.5d87fe7a265a5p-10},
80       {0x1.d89250b013eb8p-70, 0x1.430912f86cb8ep-13},
81   };
82 
83   DoubleDouble p = fputil::polyeval(dx, COEFFS[0], COEFFS[1], COEFFS[2],
84                                     COEFFS[3], COEFFS[4], COEFFS[5], COEFFS[6]);
85   return p;
86 }
87 
88 // Polynomial approximation with 128-bit precision:
89 // Return exp(dx) ~ 1 + a0 * dx + a1 * dx^2 + ... + a6 * dx^7
90 // For |dx| < 2^-13 + 2^-30:
91 //   | output - exp(dx) | < 2^-126.
poly_approx_f128(const Float128 & dx)92 Float128 poly_approx_f128(const Float128 &dx) {
93   constexpr Float128 COEFFS_128[]{
94       {Sign::POS, -127, 0x80000000'00000000'00000000'00000000_u128}, // 1.0
95       {Sign::POS, -128, 0xb17217f7'd1cf79ab'c9e3b398'03f2f6af_u128},
96       {Sign::POS, -128, 0x3d7f7bff'058b1d50'de2d60dd'9c9a1d9f_u128},
97       {Sign::POS, -132, 0xe35846b8'2505fc59'9d3b15d9'e7fb6897_u128},
98       {Sign::POS, -134, 0x9d955b7d'd273b94e'184462f6'bcd2b9e7_u128},
99       {Sign::POS, -137, 0xaec3ff3c'53398883'39ea1bb9'64c51a89_u128},
100       {Sign::POS, -138, 0x2861225f'345c396a'842c5341'8fa8ae61_u128},
101       {Sign::POS, -144, 0xffe5fe2d'109a319d'7abeb5ab'd5ad2079_u128},
102   };
103 
104   Float128 p = fputil::polyeval(dx, COEFFS_128[0], COEFFS_128[1], COEFFS_128[2],
105                                 COEFFS_128[3], COEFFS_128[4], COEFFS_128[5],
106                                 COEFFS_128[6], COEFFS_128[7]);
107   return p;
108 }
109 
110 // Compute 2^(x) using 128-bit precision.
111 // TODO(lntue): investigate triple-double precision implementation for this
112 // step.
exp2_f128(double x,int hi,int idx1,int idx2)113 Float128 exp2_f128(double x, int hi, int idx1, int idx2) {
114   Float128 dx = Float128(x);
115 
116   // TODO: Skip recalculating exp_mid1 and exp_mid2.
117   Float128 exp_mid1 =
118       fputil::quick_add(Float128(EXP2_MID1[idx1].hi),
119                         fputil::quick_add(Float128(EXP2_MID1[idx1].mid),
120                                           Float128(EXP2_MID1[idx1].lo)));
121 
122   Float128 exp_mid2 =
123       fputil::quick_add(Float128(EXP2_MID2[idx2].hi),
124                         fputil::quick_add(Float128(EXP2_MID2[idx2].mid),
125                                           Float128(EXP2_MID2[idx2].lo)));
126 
127   Float128 exp_mid = fputil::quick_mul(exp_mid1, exp_mid2);
128 
129   Float128 p = poly_approx_f128(dx);
130 
131   Float128 r = fputil::quick_mul(exp_mid, p);
132 
133   r.exponent += hi;
134 
135   return r;
136 }
137 
138 // Compute 2^x with double-double precision.
exp2_double_double(double x,const DoubleDouble & exp_mid)139 DoubleDouble exp2_double_double(double x, const DoubleDouble &exp_mid) {
140   DoubleDouble dx({0, x});
141 
142   // Degree-6 polynomial approximation in double-double precision.
143   // | p - 2^x | < 2^-103.
144   DoubleDouble p = poly_approx_dd(dx);
145 
146   // Error bounds: 2^-102.
147   DoubleDouble r = fputil::quick_mult(exp_mid, p);
148 
149   return r;
150 }
151 
152 // When output is denormal.
exp2_denorm(double x)153 double exp2_denorm(double x) {
154   // Range reduction.
155   int k =
156       static_cast<int>(cpp::bit_cast<uint64_t>(x + 0x1.8000'0000'4p21) >> 19);
157   double kd = static_cast<double>(k);
158 
159   uint32_t idx1 = (k >> 6) & 0x3f;
160   uint32_t idx2 = k & 0x3f;
161 
162   int hi = k >> 12;
163 
164   DoubleDouble exp_mid1{EXP2_MID1[idx1].mid, EXP2_MID1[idx1].hi};
165   DoubleDouble exp_mid2{EXP2_MID2[idx2].mid, EXP2_MID2[idx2].hi};
166   DoubleDouble exp_mid = fputil::quick_mult(exp_mid1, exp_mid2);
167 
168   // |dx| < 2^-13 + 2^-30.
169   double dx = fputil::multiply_add(kd, -0x1.0p-12, x); // exact
170 
171   double mid_lo = dx * exp_mid.hi;
172 
173   // Approximate (2^dx - 1)/dx ~ 1 + a0*dx + a1*dx^2 + a2*dx^3 + a3*dx^4.
174   double p = poly_approx_d(dx);
175 
176   double lo = fputil::multiply_add(p, mid_lo, exp_mid.lo);
177 
178   if (auto r = ziv_test_denorm(hi, exp_mid.hi, lo, ERR_D);
179       LIBC_LIKELY(r.has_value()))
180     return r.value();
181 
182   // Use double-double
183   DoubleDouble r_dd = exp2_double_double(dx, exp_mid);
184 
185   if (auto r = ziv_test_denorm(hi, r_dd.hi, r_dd.lo, ERR_DD);
186       LIBC_LIKELY(r.has_value()))
187     return r.value();
188 
189   // Use 128-bit precision
190   Float128 r_f128 = exp2_f128(dx, hi, idx1, idx2);
191 
192   return static_cast<double>(r_f128);
193 }
194 
195 // Check for exceptional cases when:
196 //  * log2(1 - 2^-54) < x < log2(1 + 2^-53)
197 //  * x >= 1024
198 //  * x <= -1022
199 //  * x is inf or nan
set_exceptional(double x)200 double set_exceptional(double x) {
201   using FPBits = typename fputil::FPBits<double>;
202   FPBits xbits(x);
203 
204   uint64_t x_u = xbits.uintval();
205   uint64_t x_abs = xbits.abs().uintval();
206 
207   // |x| < log2(1 + 2^-53)
208   if (x_abs <= 0x3ca71547652b82fd) {
209     // 2^(x) ~ 1 + x/2
210     return fputil::multiply_add(x, 0.5, 1.0);
211   }
212 
213   // x <= -1022 || x >= 1024 or inf/nan.
214   if (x_u > 0xc08ff00000000000) {
215     // x <= -1075 or -inf/nan
216     if (x_u >= 0xc090cc0000000000) {
217       // exp(-Inf) = 0
218       if (xbits.is_inf())
219         return 0.0;
220 
221       // exp(nan) = nan
222       if (xbits.is_nan())
223         return x;
224 
225       if (fputil::quick_get_round() == FE_UPWARD)
226         return FPBits::min_subnormal().get_val();
227       fputil::set_errno_if_required(ERANGE);
228       fputil::raise_except_if_required(FE_UNDERFLOW);
229       return 0.0;
230     }
231 
232     return exp2_denorm(x);
233   }
234 
235   // x >= 1024 or +inf/nan
236   // x is finite
237   if (x_u < 0x7ff0'0000'0000'0000ULL) {
238     int rounding = fputil::quick_get_round();
239     if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO)
240       return FPBits::max_normal().get_val();
241 
242     fputil::set_errno_if_required(ERANGE);
243     fputil::raise_except_if_required(FE_OVERFLOW);
244   }
245   // x is +inf or nan
246   return x + FPBits::inf().get_val();
247 }
248 
249 } // namespace
250 
251 LLVM_LIBC_FUNCTION(double, exp2, (double x)) {
252   using FPBits = typename fputil::FPBits<double>;
253   FPBits xbits(x);
254 
255   uint64_t x_u = xbits.uintval();
256 
257   // x < -1022 or x >= 1024 or log2(1 - 2^-54) < x < log2(1 + 2^-53).
258   if (LIBC_UNLIKELY(x_u > 0xc08ff00000000000 ||
259                     (x_u <= 0xbc971547652b82fe && x_u >= 0x4090000000000000) ||
260                     x_u <= 0x3ca71547652b82fd)) {
261     return set_exceptional(x);
262   }
263 
264   // Now -1075 < x <= log2(1 - 2^-54) or log2(1 + 2^-53) < x < 1024
265 
266   // Range reduction:
267   // Let x = (hi + mid1 + mid2) + lo
268   // in which:
269   //   hi is an integer
270   //   mid1 * 2^6 is an integer
271   //   mid2 * 2^12 is an integer
272   // then:
273   //   2^(x) = 2^hi * 2^(mid1) * 2^(mid2) * 2^(lo).
274   // With this formula:
275   //   - multiplying by 2^hi is exact and cheap, simply by adding the exponent
276   //     field.
277   //   - 2^(mid1) and 2^(mid2) are stored in 2 x 64-element tables.
278   //   - 2^(lo) ~ 1 + a0*lo + a1 * lo^2 + ...
279   //
280   // We compute (hi + mid1 + mid2) together by perform the rounding on x * 2^12.
281   // Since |x| < |-1075)| < 2^11,
282   //   |x * 2^12| < 2^11 * 2^12 < 2^23,
283   // So we can fit the rounded result round(x * 2^12) in int32_t.
284   // Thus, the goal is to be able to use an additional addition and fixed width
285   // shift to get an int32_t representing round(x * 2^12).
286   //
287   // Assuming int32_t using 2-complement representation, since the mantissa part
288   // of a double precision is unsigned with the leading bit hidden, if we add an
289   // extra constant C = 2^e1 + 2^e2 with e1 > e2 >= 2^25 to the product, the
290   // part that are < 2^e2 in resulted mantissa of (x*2^12*L2E + C) can be
291   // considered as a proper 2-complement representations of x*2^12.
292   //
293   // One small problem with this approach is that the sum (x*2^12 + C) in
294   // double precision is rounded to the least significant bit of the dorminant
295   // factor C.  In order to minimize the rounding errors from this addition, we
296   // want to minimize e1.  Another constraint that we want is that after
297   // shifting the mantissa so that the least significant bit of int32_t
298   // corresponds to the unit bit of (x*2^12*L2E), the sign is correct without
299   // any adjustment.  So combining these 2 requirements, we can choose
300   //   C = 2^33 + 2^32, so that the sign bit corresponds to 2^31 bit, and hence
301   // after right shifting the mantissa, the resulting int32_t has correct sign.
302   // With this choice of C, the number of mantissa bits we need to shift to the
303   // right is: 52 - 33 = 19.
304   //
305   // Moreover, since the integer right shifts are equivalent to rounding down,
306   // we can add an extra 0.5 so that it will become round-to-nearest, tie-to-
307   // +infinity.  So in particular, we can compute:
308   //   hmm = x * 2^12 + C,
309   // where C = 2^33 + 2^32 + 2^-1, then if
310   //   k = int32_t(lower 51 bits of double(x * 2^12 + C) >> 19),
311   // the reduced argument:
312   //   lo = x - 2^-12 * k is bounded by:
313   //   |lo| <= 2^-13 + 2^-12*2^-19
314   //         = 2^-13 + 2^-31.
315   //
316   // Finally, notice that k only uses the mantissa of x * 2^12, so the
317   // exponent 2^12 is not needed.  So we can simply define
318   //   C = 2^(33 - 12) + 2^(32 - 12) + 2^(-13 - 12), and
319   //   k = int32_t(lower 51 bits of double(x + C) >> 19).
320 
321   // Rounding errors <= 2^-31.
322   int k =
323       static_cast<int>(cpp::bit_cast<uint64_t>(x + 0x1.8000'0000'4p21) >> 19);
324   double kd = static_cast<double>(k);
325 
326   uint32_t idx1 = (k >> 6) & 0x3f;
327   uint32_t idx2 = k & 0x3f;
328 
329   int hi = k >> 12;
330 
331   DoubleDouble exp_mid1{EXP2_MID1[idx1].mid, EXP2_MID1[idx1].hi};
332   DoubleDouble exp_mid2{EXP2_MID2[idx2].mid, EXP2_MID2[idx2].hi};
333   DoubleDouble exp_mid = fputil::quick_mult(exp_mid1, exp_mid2);
334 
335   // |dx| < 2^-13 + 2^-30.
336   double dx = fputil::multiply_add(kd, -0x1.0p-12, x); // exact
337 
338   // We use the degree-4 polynomial to approximate 2^(lo):
339   //   2^(lo) ~ 1 + a0 * lo + a1 * lo^2 + a2 * lo^3 + a3 * lo^4 = 1 + lo * P(lo)
340   // So that the errors are bounded by:
341   //   |P(lo) - (2^lo - 1)/lo| < |lo|^4 / 64 < 2^(-13 * 4) / 64 = 2^-58
342   // Let P_ be an evaluation of P where all intermediate computations are in
343   // double precision.  Using either Horner's or Estrin's schemes, the evaluated
344   // errors can be bounded by:
345   //      |P_(lo) - P(lo)| < 2^-51
346   //   => |lo * P_(lo) - (2^lo - 1) | < 2^-64
347   //   => 2^(mid1 + mid2) * |lo * P_(lo) - expm1(lo)| < 2^-63.
348   // Since we approximate
349   //   2^(mid1 + mid2) ~ exp_mid.hi + exp_mid.lo,
350   // We use the expression:
351   //    (exp_mid.hi + exp_mid.lo) * (1 + dx * P_(dx)) ~
352   //  ~ exp_mid.hi + (exp_mid.hi * dx * P_(dx) + exp_mid.lo)
353   // with errors bounded by 2^-63.
354 
355   double mid_lo = dx * exp_mid.hi;
356 
357   // Approximate (2^dx - 1)/dx ~ 1 + a0*dx + a1*dx^2 + a2*dx^3 + a3*dx^4.
358   double p = poly_approx_d(dx);
359 
360   double lo = fputil::multiply_add(p, mid_lo, exp_mid.lo);
361 
362   double upper = exp_mid.hi + (lo + ERR_D);
363   double lower = exp_mid.hi + (lo - ERR_D);
364 
365   if (LIBC_LIKELY(upper == lower)) {
366     // To multiply by 2^hi, a fast way is to simply add hi to the exponent
367     // field.
368     int64_t exp_hi = static_cast<int64_t>(hi) << FPBits::FRACTION_LEN;
369     double r = cpp::bit_cast<double>(exp_hi + cpp::bit_cast<int64_t>(upper));
370     return r;
371   }
372 
373   // Use double-double
374   DoubleDouble r_dd = exp2_double_double(dx, exp_mid);
375 
376   double upper_dd = r_dd.hi + (r_dd.lo + ERR_DD);
377   double lower_dd = r_dd.hi + (r_dd.lo - ERR_DD);
378 
379   if (LIBC_LIKELY(upper_dd == lower_dd)) {
380     // To multiply by 2^hi, a fast way is to simply add hi to the exponent
381     // field.
382     int64_t exp_hi = static_cast<int64_t>(hi) << FPBits::FRACTION_LEN;
383     double r = cpp::bit_cast<double>(exp_hi + cpp::bit_cast<int64_t>(upper_dd));
384     return r;
385   }
386 
387   // Use 128-bit precision
388   Float128 r_f128 = exp2_f128(dx, hi, idx1, idx2);
389 
390   return static_cast<double>(r_f128);
391 }
392 
393 } // namespace LIBC_NAMESPACE
394