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/config.h" |
26 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
27 | |
28 | namespace LIBC_NAMESPACE_DECL { |
29 | |
30 | using fputil::DoubleDouble; |
31 | using fputil::TripleDouble; |
32 | using Float128 = typename fputil::DyadicFloat<128>; |
33 | |
34 | using LIBC_NAMESPACE::operator""_u128 ; |
35 | |
36 | // Error bounds: |
37 | // Errors when using double precision. |
38 | #ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
39 | constexpr double ERR_D = 0x1.0p-63; |
40 | #else |
41 | constexpr double ERR_D = 0x1.8p-63; |
42 | #endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
43 | |
44 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
45 | // Errors when using double-double precision. |
46 | constexpr double ERR_DD = 0x1.0p-100; |
47 | #endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
48 | |
49 | namespace { |
50 | |
51 | // Polynomial approximations with double precision. Generated by Sollya with: |
52 | // > P = fpminimax((2^x - 1)/x, 3, [|D...|], [-2^-13 - 2^-30, 2^-13 + 2^-30]); |
53 | // > P; |
54 | // Error bounds: |
55 | // | output - (2^dx - 1) / dx | < 1.5 * 2^-52. |
56 | LIBC_INLINE double poly_approx_d(double dx) { |
57 | // dx^2 |
58 | double dx2 = dx * dx; |
59 | double c0 = |
60 | fputil::multiply_add(dx, 0x1.ebfbdff82c58ep-3, 0x1.62e42fefa39efp-1); |
61 | double c1 = |
62 | fputil::multiply_add(dx, 0x1.3b2aba7a95a89p-7, 0x1.c6b08e8fc0c0ep-5); |
63 | double p = fputil::multiply_add(dx2, c1, c0); |
64 | return p; |
65 | } |
66 | |
67 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
68 | // Polynomial approximation with double-double precision. Generated by Solya |
69 | // with: |
70 | // > P = fpminimax((2^x - 1)/x, 5, [|DD...|], [-2^-13 - 2^-30, 2^-13 + 2^-30]); |
71 | // Error bounds: |
72 | // | output - 2^(dx) | < 2^-101 |
73 | DoubleDouble poly_approx_dd(const DoubleDouble &dx) { |
74 | // Taylor polynomial. |
75 | constexpr DoubleDouble COEFFS[] = { |
76 | {0, 0x1p0}, |
77 | {0x1.abc9e3b39824p-56, 0x1.62e42fefa39efp-1}, |
78 | {-0x1.5e43a53e4527bp-57, 0x1.ebfbdff82c58fp-3}, |
79 | {-0x1.d37963a9444eep-59, 0x1.c6b08d704a0cp-5}, |
80 | {0x1.4eda1a81133dap-62, 0x1.3b2ab6fba4e77p-7}, |
81 | {-0x1.c53fd1ba85d14p-64, 0x1.5d87fe7a265a5p-10}, |
82 | {0x1.d89250b013eb8p-70, 0x1.430912f86cb8ep-13}, |
83 | }; |
84 | |
85 | DoubleDouble p = fputil::polyeval(dx, COEFFS[0], COEFFS[1], COEFFS[2], |
86 | COEFFS[3], COEFFS[4], COEFFS[5], COEFFS[6]); |
87 | return p; |
88 | } |
89 | |
90 | // Polynomial approximation with 128-bit precision: |
91 | // Return exp(dx) ~ 1 + a0 * dx + a1 * dx^2 + ... + a6 * dx^7 |
92 | // For |dx| < 2^-13 + 2^-30: |
93 | // | output - exp(dx) | < 2^-126. |
94 | Float128 poly_approx_f128(const Float128 &dx) { |
95 | constexpr Float128 COEFFS_128[]{ |
96 | {Sign::POS, -127, 0x80000000'00000000'00000000'00000000_u128}, // 1.0 |
97 | {Sign::POS, -128, 0xb17217f7'd1cf79ab'c9e3b398'03f2f6af_u128}, |
98 | {Sign::POS, -128, 0x3d7f7bff'058b1d50'de2d60dd'9c9a1d9f_u128}, |
99 | {Sign::POS, -132, 0xe35846b8'2505fc59'9d3b15d9'e7fb6897_u128}, |
100 | {Sign::POS, -134, 0x9d955b7d'd273b94e'184462f6'bcd2b9e7_u128}, |
101 | {Sign::POS, -137, 0xaec3ff3c'53398883'39ea1bb9'64c51a89_u128}, |
102 | {Sign::POS, -138, 0x2861225f'345c396a'842c5341'8fa8ae61_u128}, |
103 | {Sign::POS, -144, 0xffe5fe2d'109a319d'7abeb5ab'd5ad2079_u128}, |
104 | }; |
105 | |
106 | Float128 p = fputil::polyeval(dx, COEFFS_128[0], COEFFS_128[1], COEFFS_128[2], |
107 | COEFFS_128[3], COEFFS_128[4], COEFFS_128[5], |
108 | COEFFS_128[6], COEFFS_128[7]); |
109 | return p; |
110 | } |
111 | |
112 | // Compute 2^(x) using 128-bit precision. |
113 | // TODO(lntue): investigate triple-double precision implementation for this |
114 | // step. |
115 | Float128 exp2_f128(double x, int hi, int idx1, int idx2) { |
116 | Float128 dx = Float128(x); |
117 | |
118 | // TODO: Skip recalculating exp_mid1 and exp_mid2. |
119 | Float128 exp_mid1 = |
120 | fputil::quick_add(Float128(EXP2_MID1[idx1].hi), |
121 | fputil::quick_add(Float128(EXP2_MID1[idx1].mid), |
122 | Float128(EXP2_MID1[idx1].lo))); |
123 | |
124 | Float128 exp_mid2 = |
125 | fputil::quick_add(Float128(EXP2_MID2[idx2].hi), |
126 | fputil::quick_add(Float128(EXP2_MID2[idx2].mid), |
127 | Float128(EXP2_MID2[idx2].lo))); |
128 | |
129 | Float128 exp_mid = fputil::quick_mul(exp_mid1, exp_mid2); |
130 | |
131 | Float128 p = poly_approx_f128(dx); |
132 | |
133 | Float128 r = fputil::quick_mul(exp_mid, p); |
134 | |
135 | r.exponent += hi; |
136 | |
137 | return r; |
138 | } |
139 | |
140 | // Compute 2^x with double-double precision. |
141 | DoubleDouble exp2_double_double(double x, const DoubleDouble &exp_mid) { |
142 | DoubleDouble dx({0, x}); |
143 | |
144 | // Degree-6 polynomial approximation in double-double precision. |
145 | // | p - 2^x | < 2^-103. |
146 | DoubleDouble p = poly_approx_dd(dx); |
147 | |
148 | // Error bounds: 2^-102. |
149 | DoubleDouble r = fputil::quick_mult(exp_mid, p); |
150 | |
151 | return r; |
152 | } |
153 | #endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
154 | |
155 | // When output is denormal. |
156 | double exp2_denorm(double x) { |
157 | // Range reduction. |
158 | int k = |
159 | static_cast<int>(cpp::bit_cast<uint64_t>(x + 0x1.8000'0000'4p21) >> 19); |
160 | double kd = static_cast<double>(k); |
161 | |
162 | uint32_t idx1 = (k >> 6) & 0x3f; |
163 | uint32_t idx2 = k & 0x3f; |
164 | |
165 | int hi = k >> 12; |
166 | |
167 | DoubleDouble exp_mid1{EXP2_MID1[idx1].mid, EXP2_MID1[idx1].hi}; |
168 | DoubleDouble exp_mid2{EXP2_MID2[idx2].mid, EXP2_MID2[idx2].hi}; |
169 | DoubleDouble exp_mid = fputil::quick_mult(exp_mid1, exp_mid2); |
170 | |
171 | // |dx| < 2^-13 + 2^-30. |
172 | double dx = fputil::multiply_add(kd, -0x1.0p-12, x); // exact |
173 | |
174 | double mid_lo = dx * exp_mid.hi; |
175 | |
176 | // Approximate (2^dx - 1)/dx ~ 1 + a0*dx + a1*dx^2 + a2*dx^3 + a3*dx^4. |
177 | double p = poly_approx_d(dx); |
178 | |
179 | double lo = fputil::multiply_add(p, mid_lo, exp_mid.lo); |
180 | |
181 | #ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
182 | return ziv_test_denorm</*SKIP_ZIV_TEST=*/true>(hi, exp_mid.hi, lo, ERR_D) |
183 | .value(); |
184 | #else |
185 | if (auto r = ziv_test_denorm(hi, exp_mid.hi, lo, ERR_D); |
186 | LIBC_LIKELY(r.has_value())) |
187 | return r.value(); |
188 | |
189 | // Use double-double |
190 | DoubleDouble r_dd = exp2_double_double(dx, exp_mid); |
191 | |
192 | if (auto r = ziv_test_denorm(hi, r_dd.hi, r_dd.lo, ERR_DD); |
193 | LIBC_LIKELY(r.has_value())) |
194 | return r.value(); |
195 | |
196 | // Use 128-bit precision |
197 | Float128 r_f128 = exp2_f128(dx, hi, idx1, idx2); |
198 | |
199 | return static_cast<double>(r_f128); |
200 | #endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
201 | } |
202 | |
203 | // Check for exceptional cases when: |
204 | // * log2(1 - 2^-54) < x < log2(1 + 2^-53) |
205 | // * x >= 1024 |
206 | // * x <= -1022 |
207 | // * x is inf or nan |
208 | double set_exceptional(double x) { |
209 | using FPBits = typename fputil::FPBits<double>; |
210 | FPBits xbits(x); |
211 | |
212 | uint64_t x_u = xbits.uintval(); |
213 | uint64_t x_abs = xbits.abs().uintval(); |
214 | |
215 | // |x| < log2(1 + 2^-53) |
216 | if (x_abs <= 0x3ca71547652b82fd) { |
217 | // 2^(x) ~ 1 + x/2 |
218 | return fputil::multiply_add(x, 0.5, 1.0); |
219 | } |
220 | |
221 | // x <= -1022 || x >= 1024 or inf/nan. |
222 | if (x_u > 0xc08ff00000000000) { |
223 | // x <= -1075 or -inf/nan |
224 | if (x_u >= 0xc090cc0000000000) { |
225 | // exp(-Inf) = 0 |
226 | if (xbits.is_inf()) |
227 | return 0.0; |
228 | |
229 | // exp(nan) = nan |
230 | if (xbits.is_nan()) |
231 | return x; |
232 | |
233 | if (fputil::quick_get_round() == FE_UPWARD) |
234 | return FPBits::min_subnormal().get_val(); |
235 | fputil::set_errno_if_required(ERANGE); |
236 | fputil::raise_except_if_required(FE_UNDERFLOW); |
237 | return 0.0; |
238 | } |
239 | |
240 | return exp2_denorm(x); |
241 | } |
242 | |
243 | // x >= 1024 or +inf/nan |
244 | // x is finite |
245 | if (x_u < 0x7ff0'0000'0000'0000ULL) { |
246 | int rounding = fputil::quick_get_round(); |
247 | if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) |
248 | return FPBits::max_normal().get_val(); |
249 | |
250 | fputil::set_errno_if_required(ERANGE); |
251 | fputil::raise_except_if_required(FE_OVERFLOW); |
252 | } |
253 | // x is +inf or nan |
254 | return x + FPBits::inf().get_val(); |
255 | } |
256 | |
257 | } // namespace |
258 | |
259 | LLVM_LIBC_FUNCTION(double, exp2, (double x)) { |
260 | using FPBits = typename fputil::FPBits<double>; |
261 | FPBits xbits(x); |
262 | |
263 | uint64_t x_u = xbits.uintval(); |
264 | |
265 | // x < -1022 or x >= 1024 or log2(1 - 2^-54) < x < log2(1 + 2^-53). |
266 | if (LIBC_UNLIKELY(x_u > 0xc08ff00000000000 || |
267 | (x_u <= 0xbc971547652b82fe && x_u >= 0x4090000000000000) || |
268 | x_u <= 0x3ca71547652b82fd)) { |
269 | return set_exceptional(x); |
270 | } |
271 | |
272 | // Now -1075 < x <= log2(1 - 2^-54) or log2(1 + 2^-53) < x < 1024 |
273 | |
274 | // Range reduction: |
275 | // Let x = (hi + mid1 + mid2) + lo |
276 | // in which: |
277 | // hi is an integer |
278 | // mid1 * 2^6 is an integer |
279 | // mid2 * 2^12 is an integer |
280 | // then: |
281 | // 2^(x) = 2^hi * 2^(mid1) * 2^(mid2) * 2^(lo). |
282 | // With this formula: |
283 | // - multiplying by 2^hi is exact and cheap, simply by adding the exponent |
284 | // field. |
285 | // - 2^(mid1) and 2^(mid2) are stored in 2 x 64-element tables. |
286 | // - 2^(lo) ~ 1 + a0*lo + a1 * lo^2 + ... |
287 | // |
288 | // We compute (hi + mid1 + mid2) together by perform the rounding on x * 2^12. |
289 | // Since |x| < |-1075)| < 2^11, |
290 | // |x * 2^12| < 2^11 * 2^12 < 2^23, |
291 | // So we can fit the rounded result round(x * 2^12) in int32_t. |
292 | // Thus, the goal is to be able to use an additional addition and fixed width |
293 | // shift to get an int32_t representing round(x * 2^12). |
294 | // |
295 | // Assuming int32_t using 2-complement representation, since the mantissa part |
296 | // of a double precision is unsigned with the leading bit hidden, if we add an |
297 | // extra constant C = 2^e1 + 2^e2 with e1 > e2 >= 2^25 to the product, the |
298 | // part that are < 2^e2 in resulted mantissa of (x*2^12*L2E + C) can be |
299 | // considered as a proper 2-complement representations of x*2^12. |
300 | // |
301 | // One small problem with this approach is that the sum (x*2^12 + C) in |
302 | // double precision is rounded to the least significant bit of the dorminant |
303 | // factor C. In order to minimize the rounding errors from this addition, we |
304 | // want to minimize e1. Another constraint that we want is that after |
305 | // shifting the mantissa so that the least significant bit of int32_t |
306 | // corresponds to the unit bit of (x*2^12*L2E), the sign is correct without |
307 | // any adjustment. So combining these 2 requirements, we can choose |
308 | // C = 2^33 + 2^32, so that the sign bit corresponds to 2^31 bit, and hence |
309 | // after right shifting the mantissa, the resulting int32_t has correct sign. |
310 | // With this choice of C, the number of mantissa bits we need to shift to the |
311 | // right is: 52 - 33 = 19. |
312 | // |
313 | // Moreover, since the integer right shifts are equivalent to rounding down, |
314 | // we can add an extra 0.5 so that it will become round-to-nearest, tie-to- |
315 | // +infinity. So in particular, we can compute: |
316 | // hmm = x * 2^12 + C, |
317 | // where C = 2^33 + 2^32 + 2^-1, then if |
318 | // k = int32_t(lower 51 bits of double(x * 2^12 + C) >> 19), |
319 | // the reduced argument: |
320 | // lo = x - 2^-12 * k is bounded by: |
321 | // |lo| <= 2^-13 + 2^-12*2^-19 |
322 | // = 2^-13 + 2^-31. |
323 | // |
324 | // Finally, notice that k only uses the mantissa of x * 2^12, so the |
325 | // exponent 2^12 is not needed. So we can simply define |
326 | // C = 2^(33 - 12) + 2^(32 - 12) + 2^(-13 - 12), and |
327 | // k = int32_t(lower 51 bits of double(x + C) >> 19). |
328 | |
329 | // Rounding errors <= 2^-31. |
330 | int k = |
331 | static_cast<int>(cpp::bit_cast<uint64_t>(x + 0x1.8000'0000'4p21) >> 19); |
332 | double kd = static_cast<double>(k); |
333 | |
334 | uint32_t idx1 = (k >> 6) & 0x3f; |
335 | uint32_t idx2 = k & 0x3f; |
336 | |
337 | int hi = k >> 12; |
338 | |
339 | DoubleDouble exp_mid1{EXP2_MID1[idx1].mid, EXP2_MID1[idx1].hi}; |
340 | DoubleDouble exp_mid2{EXP2_MID2[idx2].mid, EXP2_MID2[idx2].hi}; |
341 | DoubleDouble exp_mid = fputil::quick_mult(exp_mid1, exp_mid2); |
342 | |
343 | // |dx| < 2^-13 + 2^-30. |
344 | double dx = fputil::multiply_add(kd, -0x1.0p-12, x); // exact |
345 | |
346 | // We use the degree-4 polynomial to approximate 2^(lo): |
347 | // 2^(lo) ~ 1 + a0 * lo + a1 * lo^2 + a2 * lo^3 + a3 * lo^4 = 1 + lo * P(lo) |
348 | // So that the errors are bounded by: |
349 | // |P(lo) - (2^lo - 1)/lo| < |lo|^4 / 64 < 2^(-13 * 4) / 64 = 2^-58 |
350 | // Let P_ be an evaluation of P where all intermediate computations are in |
351 | // double precision. Using either Horner's or Estrin's schemes, the evaluated |
352 | // errors can be bounded by: |
353 | // |P_(lo) - P(lo)| < 2^-51 |
354 | // => |lo * P_(lo) - (2^lo - 1) | < 2^-64 |
355 | // => 2^(mid1 + mid2) * |lo * P_(lo) - expm1(lo)| < 2^-63. |
356 | // Since we approximate |
357 | // 2^(mid1 + mid2) ~ exp_mid.hi + exp_mid.lo, |
358 | // We use the expression: |
359 | // (exp_mid.hi + exp_mid.lo) * (1 + dx * P_(dx)) ~ |
360 | // ~ exp_mid.hi + (exp_mid.hi * dx * P_(dx) + exp_mid.lo) |
361 | // with errors bounded by 2^-63. |
362 | |
363 | double mid_lo = dx * exp_mid.hi; |
364 | |
365 | // Approximate (2^dx - 1)/dx ~ 1 + a0*dx + a1*dx^2 + a2*dx^3 + a3*dx^4. |
366 | double p = poly_approx_d(dx); |
367 | |
368 | double lo = fputil::multiply_add(p, mid_lo, exp_mid.lo); |
369 | |
370 | #ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
371 | // To multiply by 2^hi, a fast way is to simply add hi to the exponent |
372 | // field. |
373 | int64_t exp_hi = static_cast<int64_t>(hi) << FPBits::FRACTION_LEN; |
374 | double r = |
375 | cpp::bit_cast<double>(exp_hi + cpp::bit_cast<int64_t>(exp_mid.hi + lo)); |
376 | return r; |
377 | #else |
378 | double upper = exp_mid.hi + (lo + ERR_D); |
379 | double lower = exp_mid.hi + (lo - ERR_D); |
380 | |
381 | if (LIBC_LIKELY(upper == lower)) { |
382 | // To multiply by 2^hi, a fast way is to simply add hi to the exponent |
383 | // field. |
384 | int64_t exp_hi = static_cast<int64_t>(hi) << FPBits::FRACTION_LEN; |
385 | double r = cpp::bit_cast<double>(exp_hi + cpp::bit_cast<int64_t>(upper)); |
386 | return r; |
387 | } |
388 | |
389 | // Use double-double |
390 | DoubleDouble r_dd = exp2_double_double(dx, exp_mid); |
391 | |
392 | double upper_dd = r_dd.hi + (r_dd.lo + ERR_DD); |
393 | double lower_dd = r_dd.hi + (r_dd.lo - ERR_DD); |
394 | |
395 | if (LIBC_LIKELY(upper_dd == lower_dd)) { |
396 | // To multiply by 2^hi, a fast way is to simply add hi to the exponent |
397 | // field. |
398 | int64_t exp_hi = static_cast<int64_t>(hi) << FPBits::FRACTION_LEN; |
399 | double r = cpp::bit_cast<double>(exp_hi + cpp::bit_cast<int64_t>(upper_dd)); |
400 | return r; |
401 | } |
402 | |
403 | // Use 128-bit precision |
404 | Float128 r_f128 = exp2_f128(dx, hi, idx1, idx2); |
405 | |
406 | return static_cast<double>(r_f128); |
407 | #endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
408 | } |
409 | |
410 | } // namespace LIBC_NAMESPACE_DECL |
411 | |