1 | //===-- Double-precision e^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/exp.h" |
10 | #include "common_constants.h" // Lookup tables EXP_M1 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 | // log2(e) |
38 | constexpr double LOG2_E = 0x1.71547652b82fep+0; |
39 | |
40 | // Error bounds: |
41 | // Errors when using double precision. |
42 | constexpr double ERR_D = 0x1.8p-63; |
43 | // Errors when using double-double precision. |
44 | constexpr double ERR_DD = 0x1.0p-99; |
45 | |
46 | // -2^-12 * log(2) |
47 | // > a = -2^-12 * log(2); |
48 | // > b = round(a, 30, RN); |
49 | // > c = round(a - b, 30, RN); |
50 | // > d = round(a - b - c, D, RN); |
51 | // Errors < 1.5 * 2^-133 |
52 | constexpr double MLOG_2_EXP2_M12_HI = -0x1.62e42ffp-13; |
53 | constexpr double MLOG_2_EXP2_M12_MID = 0x1.718432a1b0e26p-47; |
54 | constexpr double MLOG_2_EXP2_M12_MID_30 = 0x1.718432ap-47; |
55 | constexpr double MLOG_2_EXP2_M12_LO = 0x1.b0e2633fe0685p-79; |
56 | |
57 | namespace { |
58 | |
59 | // Polynomial approximations with double precision: |
60 | // Return expm1(dx) / x ~ 1 + dx / 2 + dx^2 / 6 + dx^3 / 24. |
61 | // For |dx| < 2^-13 + 2^-30: |
62 | // | output - expm1(dx) / dx | < 2^-51. |
63 | LIBC_INLINE double poly_approx_d(double dx) { |
64 | // dx^2 |
65 | double dx2 = dx * dx; |
66 | // c0 = 1 + dx / 2 |
67 | double c0 = fputil::multiply_add(x: dx, y: 0.5, z: 1.0); |
68 | // c1 = 1/6 + dx / 24 |
69 | double c1 = |
70 | fputil::multiply_add(x: dx, y: 0x1.5555555555555p-5, z: 0x1.5555555555555p-3); |
71 | // p = dx^2 * c1 + c0 = 1 + dx / 2 + dx^2 / 6 + dx^3 / 24 |
72 | double p = fputil::multiply_add(x: dx2, y: c1, z: c0); |
73 | return p; |
74 | } |
75 | |
76 | // Polynomial approximation with double-double precision: |
77 | // Return exp(dx) ~ 1 + dx + dx^2 / 2 + ... + dx^6 / 720 |
78 | // For |dx| < 2^-13 + 2^-30: |
79 | // | output - exp(dx) | < 2^-101 |
80 | DoubleDouble poly_approx_dd(const DoubleDouble &dx) { |
81 | // Taylor polynomial. |
82 | constexpr DoubleDouble COEFFS[] = { |
83 | {.lo: 0, .hi: 0x1p0}, // 1 |
84 | {.lo: 0, .hi: 0x1p0}, // 1 |
85 | {.lo: 0, .hi: 0x1p-1}, // 1/2 |
86 | {.lo: 0x1.5555555555555p-57, .hi: 0x1.5555555555555p-3}, // 1/6 |
87 | {.lo: 0x1.5555555555555p-59, .hi: 0x1.5555555555555p-5}, // 1/24 |
88 | {.lo: 0x1.1111111111111p-63, .hi: 0x1.1111111111111p-7}, // 1/120 |
89 | {.lo: -0x1.f49f49f49f49fp-65, .hi: 0x1.6c16c16c16c17p-10}, // 1/720 |
90 | }; |
91 | |
92 | DoubleDouble p = fputil::polyeval(x: dx, a0: COEFFS[0], a: COEFFS[1], a: COEFFS[2], |
93 | a: COEFFS[3], a: COEFFS[4], a: COEFFS[5], a: COEFFS[6]); |
94 | return p; |
95 | } |
96 | |
97 | // Polynomial approximation with 128-bit precision: |
98 | // Return exp(dx) ~ 1 + dx + dx^2 / 2 + ... + dx^7 / 5040 |
99 | // For |dx| < 2^-13 + 2^-30: |
100 | // | output - exp(dx) | < 2^-126. |
101 | Float128 poly_approx_f128(const Float128 &dx) { |
102 | constexpr Float128 COEFFS_128[]{ |
103 | {Sign::POS, -127, 0x80000000'00000000'00000000'00000000_u128}, // 1.0 |
104 | {Sign::POS, -127, 0x80000000'00000000'00000000'00000000_u128}, // 1.0 |
105 | {Sign::POS, -128, 0x80000000'00000000'00000000'00000000_u128}, // 0.5 |
106 | {Sign::POS, -130, 0xaaaaaaaa'aaaaaaaa'aaaaaaaa'aaaaaaab_u128}, // 1/6 |
107 | {Sign::POS, -132, 0xaaaaaaaa'aaaaaaaa'aaaaaaaa'aaaaaaab_u128}, // 1/24 |
108 | {Sign::POS, -134, 0x88888888'88888888'88888888'88888889_u128}, // 1/120 |
109 | {Sign::POS, -137, 0xb60b60b6'0b60b60b'60b60b60'b60b60b6_u128}, // 1/720 |
110 | {Sign::POS, -140, 0xd00d00d0'0d00d00d'00d00d00'd00d00d0_u128}, // 1/5040 |
111 | }; |
112 | |
113 | Float128 p = fputil::polyeval(x: dx, a0: COEFFS_128[0], a: COEFFS_128[1], a: COEFFS_128[2], |
114 | a: COEFFS_128[3], a: COEFFS_128[4], a: COEFFS_128[5], |
115 | a: COEFFS_128[6], a: COEFFS_128[7]); |
116 | return p; |
117 | } |
118 | |
119 | // Compute exp(x) using 128-bit precision. |
120 | // TODO(lntue): investigate triple-double precision implementation for this |
121 | // step. |
122 | Float128 exp_f128(double x, double kd, int idx1, int idx2) { |
123 | // Recalculate dx: |
124 | |
125 | double t1 = fputil::multiply_add(x: kd, y: MLOG_2_EXP2_M12_HI, z: x); // exact |
126 | double t2 = kd * MLOG_2_EXP2_M12_MID_30; // exact |
127 | double t3 = kd * MLOG_2_EXP2_M12_LO; // Error < 2^-133 |
128 | |
129 | Float128 dx = fputil::quick_add( |
130 | a: Float128(t1), b: fputil::quick_add(a: Float128(t2), b: Float128(t3))); |
131 | |
132 | // TODO: Skip recalculating exp_mid1 and exp_mid2. |
133 | Float128 exp_mid1 = |
134 | fputil::quick_add(a: Float128(EXP2_MID1[idx1].hi), |
135 | b: fputil::quick_add(a: Float128(EXP2_MID1[idx1].mid), |
136 | b: Float128(EXP2_MID1[idx1].lo))); |
137 | |
138 | Float128 exp_mid2 = |
139 | fputil::quick_add(a: Float128(EXP2_MID2[idx2].hi), |
140 | b: fputil::quick_add(a: Float128(EXP2_MID2[idx2].mid), |
141 | b: Float128(EXP2_MID2[idx2].lo))); |
142 | |
143 | Float128 exp_mid = fputil::quick_mul(a: exp_mid1, b: exp_mid2); |
144 | |
145 | Float128 p = poly_approx_f128(dx); |
146 | |
147 | Float128 r = fputil::quick_mul(a: exp_mid, b: p); |
148 | |
149 | r.exponent += static_cast<int>(kd) >> 12; |
150 | |
151 | return r; |
152 | } |
153 | |
154 | // Compute exp(x) with double-double precision. |
155 | DoubleDouble exp_double_double(double x, double kd, |
156 | const DoubleDouble &exp_mid) { |
157 | // Recalculate dx: |
158 | // dx = x - k * 2^-12 * log(2) |
159 | double t1 = fputil::multiply_add(x: kd, y: MLOG_2_EXP2_M12_HI, z: x); // exact |
160 | double t2 = kd * MLOG_2_EXP2_M12_MID_30; // exact |
161 | double t3 = kd * MLOG_2_EXP2_M12_LO; // Error < 2^-130 |
162 | |
163 | DoubleDouble dx = fputil::exact_add(a: t1, b: t2); |
164 | dx.lo += t3; |
165 | |
166 | // Degree-6 Taylor polynomial approximation in double-double precision. |
167 | // | p - exp(x) | < 2^-100. |
168 | DoubleDouble p = poly_approx_dd(dx); |
169 | |
170 | // Error bounds: 2^-99. |
171 | DoubleDouble r = fputil::quick_mult(a: exp_mid, b: p); |
172 | |
173 | return r; |
174 | } |
175 | |
176 | // Check for exceptional cases when |
177 | // |x| <= 2^-53 or x < log(2^-1075) or x >= 0x1.6232bdd7abcd3p+9 |
178 | double set_exceptional(double x) { |
179 | using FPBits = typename fputil::FPBits<double>; |
180 | FPBits xbits(x); |
181 | |
182 | uint64_t x_u = xbits.uintval(); |
183 | uint64_t x_abs = xbits.abs().uintval(); |
184 | |
185 | // |x| <= 2^-53 |
186 | if (x_abs <= 0x3ca0'0000'0000'0000ULL) { |
187 | // exp(x) ~ 1 + x |
188 | return 1 + x; |
189 | } |
190 | |
191 | // x <= log(2^-1075) || x >= 0x1.6232bdd7abcd3p+9 or inf/nan. |
192 | |
193 | // x <= log(2^-1075) or -inf/nan |
194 | if (x_u >= 0xc087'4910'd52d'3052ULL) { |
195 | // exp(-Inf) = 0 |
196 | if (xbits.is_inf()) |
197 | return 0.0; |
198 | |
199 | // exp(nan) = nan |
200 | if (xbits.is_nan()) |
201 | return x; |
202 | |
203 | if (fputil::quick_get_round() == FE_UPWARD) |
204 | return FPBits::min_subnormal().get_val(); |
205 | fputil::set_errno_if_required(ERANGE); |
206 | fputil::raise_except_if_required(FE_UNDERFLOW); |
207 | return 0.0; |
208 | } |
209 | |
210 | // x >= round(log(MAX_NORMAL), D, RU) = 0x1.62e42fefa39fp+9 or +inf/nan |
211 | // x is finite |
212 | if (x_u < 0x7ff0'0000'0000'0000ULL) { |
213 | int rounding = fputil::quick_get_round(); |
214 | if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) |
215 | return FPBits::max_normal().get_val(); |
216 | |
217 | fputil::set_errno_if_required(ERANGE); |
218 | fputil::raise_except_if_required(FE_OVERFLOW); |
219 | } |
220 | // x is +inf or nan |
221 | return x + FPBits::inf().get_val(); |
222 | } |
223 | |
224 | } // namespace |
225 | |
226 | LLVM_LIBC_FUNCTION(double, exp, (double x)) { |
227 | using FPBits = typename fputil::FPBits<double>; |
228 | FPBits xbits(x); |
229 | |
230 | uint64_t x_u = xbits.uintval(); |
231 | |
232 | // Upper bound: max normal number = 2^1023 * (2 - 2^-52) |
233 | // > round(log (2^1023 ( 2 - 2^-52 )), D, RU) = 0x1.62e42fefa39fp+9 |
234 | // > round(log (2^1023 ( 2 - 2^-52 )), D, RD) = 0x1.62e42fefa39efp+9 |
235 | // > round(log (2^1023 ( 2 - 2^-52 )), D, RN) = 0x1.62e42fefa39efp+9 |
236 | // > round(exp(0x1.62e42fefa39fp+9), D, RN) = infty |
237 | |
238 | // Lower bound: min denormal number / 2 = 2^-1075 |
239 | // > round(log(2^-1075), D, RN) = -0x1.74910d52d3052p9 |
240 | |
241 | // Another lower bound: min normal number = 2^-1022 |
242 | // > round(log(2^-1022), D, RN) = -0x1.6232bdd7abcd2p9 |
243 | |
244 | // x < log(2^-1075) or x >= 0x1.6232bdd7abcd3p+9 or |x| < 2^-53. |
245 | if (LIBC_UNLIKELY(x_u >= 0xc0874910d52d3052 || |
246 | (x_u < 0xbca0000000000000 && x_u >= 0x40862e42fefa39f0) || |
247 | x_u < 0x3ca0000000000000)) { |
248 | return set_exceptional(x); |
249 | } |
250 | |
251 | // Now log(2^-1075) <= x <= -2^-53 or 2^-53 <= x < log(2^1023 * (2 - 2^-52)) |
252 | |
253 | // Range reduction: |
254 | // Let x = log(2) * (hi + mid1 + mid2) + lo |
255 | // in which: |
256 | // hi is an integer |
257 | // mid1 * 2^6 is an integer |
258 | // mid2 * 2^12 is an integer |
259 | // then: |
260 | // exp(x) = 2^hi * 2^(mid1) * 2^(mid2) * exp(lo). |
261 | // With this formula: |
262 | // - multiplying by 2^hi is exact and cheap, simply by adding the exponent |
263 | // field. |
264 | // - 2^(mid1) and 2^(mid2) are stored in 2 x 64-element tables. |
265 | // - exp(lo) ~ 1 + lo + a0 * lo^2 + ... |
266 | // |
267 | // They can be defined by: |
268 | // hi + mid1 + mid2 = 2^(-12) * round(2^12 * log_2(e) * x) |
269 | // If we store L2E = round(log2(e), D, RN), then: |
270 | // log2(e) - L2E ~ 1.5 * 2^(-56) |
271 | // So the errors when computing in double precision is: |
272 | // | x * 2^12 * log_2(e) - D(x * 2^12 * L2E) | <= |
273 | // <= | x * 2^12 * log_2(e) - x * 2^12 * L2E | + |
274 | // + | x * 2^12 * L2E - D(x * 2^12 * L2E) | |
275 | // <= 2^12 * ( |x| * 1.5 * 2^-56 + eps(x)) for RN |
276 | // 2^12 * ( |x| * 1.5 * 2^-56 + 2*eps(x)) for other rounding modes. |
277 | // So if: |
278 | // hi + mid1 + mid2 = 2^(-12) * round(x * 2^12 * L2E) is computed entirely |
279 | // in double precision, the reduced argument: |
280 | // lo = x - log(2) * (hi + mid1 + mid2) is bounded by: |
281 | // |lo| <= 2^-13 + (|x| * 1.5 * 2^-56 + 2*eps(x)) |
282 | // < 2^-13 + (1.5 * 2^9 * 1.5 * 2^-56 + 2*2^(9 - 52)) |
283 | // < 2^-13 + 2^-41 |
284 | // |
285 | |
286 | // The following trick computes the round(x * L2E) more efficiently |
287 | // than using the rounding instructions, with the tradeoff for less accuracy, |
288 | // and hence a slightly larger range for the reduced argument `lo`. |
289 | // |
290 | // To be precise, since |x| < |log(2^-1075)| < 1.5 * 2^9, |
291 | // |x * 2^12 * L2E| < 1.5 * 2^9 * 1.5 < 2^23, |
292 | // So we can fit the rounded result round(x * 2^12 * L2E) in int32_t. |
293 | // Thus, the goal is to be able to use an additional addition and fixed width |
294 | // shift to get an int32_t representing round(x * 2^12 * L2E). |
295 | // |
296 | // Assuming int32_t using 2-complement representation, since the mantissa part |
297 | // of a double precision is unsigned with the leading bit hidden, if we add an |
298 | // extra constant C = 2^e1 + 2^e2 with e1 > e2 >= 2^25 to the product, the |
299 | // part that are < 2^e2 in resulted mantissa of (x*2^12*L2E + C) can be |
300 | // considered as a proper 2-complement representations of x*2^12*L2E. |
301 | // |
302 | // One small problem with this approach is that the sum (x*2^12*L2E + C) in |
303 | // double precision is rounded to the least significant bit of the dorminant |
304 | // factor C. In order to minimize the rounding errors from this addition, we |
305 | // want to minimize e1. Another constraint that we want is that after |
306 | // shifting the mantissa so that the least significant bit of int32_t |
307 | // corresponds to the unit bit of (x*2^12*L2E), the sign is correct without |
308 | // any adjustment. So combining these 2 requirements, we can choose |
309 | // C = 2^33 + 2^32, so that the sign bit corresponds to 2^31 bit, and hence |
310 | // after right shifting the mantissa, the resulting int32_t has correct sign. |
311 | // With this choice of C, the number of mantissa bits we need to shift to the |
312 | // right is: 52 - 33 = 19. |
313 | // |
314 | // Moreover, since the integer right shifts are equivalent to rounding down, |
315 | // we can add an extra 0.5 so that it will become round-to-nearest, tie-to- |
316 | // +infinity. So in particular, we can compute: |
317 | // hmm = x * 2^12 * L2E + C, |
318 | // where C = 2^33 + 2^32 + 2^-1, then if |
319 | // k = int32_t(lower 51 bits of double(x * 2^12 * L2E + C) >> 19), |
320 | // the reduced argument: |
321 | // lo = x - log(2) * 2^-12 * k is bounded by: |
322 | // |lo| <= 2^-13 + 2^-41 + 2^-12*2^-19 |
323 | // = 2^-13 + 2^-31 + 2^-41. |
324 | // |
325 | // Finally, notice that k only uses the mantissa of x * 2^12 * L2E, so the |
326 | // exponent 2^12 is not needed. So we can simply define |
327 | // C = 2^(33 - 12) + 2^(32 - 12) + 2^(-13 - 12), and |
328 | // k = int32_t(lower 51 bits of double(x * L2E + C) >> 19). |
329 | |
330 | // Rounding errors <= 2^-31 + 2^-41. |
331 | double tmp = fputil::multiply_add(x, y: LOG2_E, z: 0x1.8000'0000'4p21); |
332 | int k = static_cast<int>(cpp::bit_cast<uint64_t>(from: tmp) >> 19); |
333 | double kd = static_cast<double>(k); |
334 | |
335 | uint32_t idx1 = (k >> 6) & 0x3f; |
336 | uint32_t idx2 = k & 0x3f; |
337 | int hi = k >> 12; |
338 | |
339 | bool denorm = (hi <= -1022); |
340 | |
341 | DoubleDouble exp_mid1{.lo: EXP2_MID1[idx1].mid, .hi: EXP2_MID1[idx1].hi}; |
342 | DoubleDouble exp_mid2{.lo: EXP2_MID2[idx2].mid, .hi: EXP2_MID2[idx2].hi}; |
343 | |
344 | DoubleDouble exp_mid = fputil::quick_mult(a: exp_mid1, b: exp_mid2); |
345 | |
346 | // |x - (hi + mid1 + mid2) * log(2) - dx| < 2^11 * eps(M_LOG_2_EXP2_M12.lo) |
347 | // = 2^11 * 2^-13 * 2^-52 |
348 | // = 2^-54. |
349 | // |dx| < 2^-13 + 2^-30. |
350 | double lo_h = fputil::multiply_add(x: kd, y: MLOG_2_EXP2_M12_HI, z: x); // exact |
351 | double dx = fputil::multiply_add(x: kd, y: MLOG_2_EXP2_M12_MID, z: lo_h); |
352 | |
353 | // We use the degree-4 Taylor polynomial to approximate exp(lo): |
354 | // exp(lo) ~ 1 + lo + lo^2 / 2 + lo^3 / 6 + lo^4 / 24 = 1 + lo * P(lo) |
355 | // So that the errors are bounded by: |
356 | // |P(lo) - expm1(lo)/lo| < |lo|^4 / 64 < 2^(-13 * 4) / 64 = 2^-58 |
357 | // Let P_ be an evaluation of P where all intermediate computations are in |
358 | // double precision. Using either Horner's or Estrin's schemes, the evaluated |
359 | // errors can be bounded by: |
360 | // |P_(dx) - P(dx)| < 2^-51 |
361 | // => |dx * P_(dx) - expm1(lo) | < 1.5 * 2^-64 |
362 | // => 2^(mid1 + mid2) * |dx * P_(dx) - expm1(lo)| < 1.5 * 2^-63. |
363 | // Since we approximate |
364 | // 2^(mid1 + mid2) ~ exp_mid.hi + exp_mid.lo, |
365 | // We use the expression: |
366 | // (exp_mid.hi + exp_mid.lo) * (1 + dx * P_(dx)) ~ |
367 | // ~ exp_mid.hi + (exp_mid.hi * dx * P_(dx) + exp_mid.lo) |
368 | // with errors bounded by 1.5 * 2^-63. |
369 | |
370 | double mid_lo = dx * exp_mid.hi; |
371 | |
372 | // Approximate expm1(dx)/dx ~ 1 + dx / 2 + dx^2 / 6 + dx^3 / 24. |
373 | double p = poly_approx_d(dx); |
374 | |
375 | double lo = fputil::multiply_add(x: p, y: mid_lo, z: exp_mid.lo); |
376 | |
377 | if (LIBC_UNLIKELY(denorm)) { |
378 | if (auto r = ziv_test_denorm(hi, mid: exp_mid.hi, lo, err: ERR_D); |
379 | LIBC_LIKELY(r.has_value())) |
380 | return r.value(); |
381 | } else { |
382 | double upper = exp_mid.hi + (lo + ERR_D); |
383 | double lower = exp_mid.hi + (lo - ERR_D); |
384 | |
385 | if (LIBC_LIKELY(upper == lower)) { |
386 | // to multiply by 2^hi, a fast way is to simply add hi to the exponent |
387 | // field. |
388 | int64_t exp_hi = static_cast<int64_t>(hi) << FPBits::FRACTION_LEN; |
389 | double r = cpp::bit_cast<double>(from: exp_hi + cpp::bit_cast<int64_t>(from: upper)); |
390 | return r; |
391 | } |
392 | } |
393 | |
394 | // Use double-double |
395 | DoubleDouble r_dd = exp_double_double(x, kd, exp_mid); |
396 | |
397 | if (LIBC_UNLIKELY(denorm)) { |
398 | if (auto r = ziv_test_denorm(hi, mid: r_dd.hi, lo: r_dd.lo, err: ERR_DD); |
399 | LIBC_LIKELY(r.has_value())) |
400 | return r.value(); |
401 | } else { |
402 | double upper_dd = r_dd.hi + (r_dd.lo + ERR_DD); |
403 | double lower_dd = r_dd.hi + (r_dd.lo - ERR_DD); |
404 | |
405 | if (LIBC_LIKELY(upper_dd == lower_dd)) { |
406 | int64_t exp_hi = static_cast<int64_t>(hi) << FPBits::FRACTION_LEN; |
407 | double r = |
408 | cpp::bit_cast<double>(from: exp_hi + cpp::bit_cast<int64_t>(from: upper_dd)); |
409 | return r; |
410 | } |
411 | } |
412 | |
413 | // Use 128-bit precision |
414 | Float128 r_f128 = exp_f128(x, kd, idx1, idx2); |
415 | |
416 | return static_cast<double>(r_f128); |
417 | } |
418 | |
419 | } // namespace LIBC_NAMESPACE |
420 | |