1 | //===-- Implementation of hypotf 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 | #ifndef LLVM_LIBC_SRC___SUPPORT_FPUTIL_HYPOT_H |
10 | #define LLVM_LIBC_SRC___SUPPORT_FPUTIL_HYPOT_H |
11 | |
12 | #include "BasicOperations.h" |
13 | #include "FEnvImpl.h" |
14 | #include "FPBits.h" |
15 | #include "rounding_mode.h" |
16 | #include "src/__support/CPP/bit.h" |
17 | #include "src/__support/CPP/type_traits.h" |
18 | #include "src/__support/common.h" |
19 | #include "src/__support/uint128.h" |
20 | |
21 | namespace LIBC_NAMESPACE { |
22 | namespace fputil { |
23 | |
24 | namespace internal { |
25 | |
26 | template <typename T> |
27 | LIBC_INLINE T find_leading_one(T mant, int &shift_length) { |
28 | shift_length = 0; |
29 | if (mant > 0) { |
30 | shift_length = (sizeof(mant) * 8) - 1 - cpp::countl_zero(mant); |
31 | } |
32 | return T(1) << shift_length; |
33 | } |
34 | |
35 | } // namespace internal |
36 | |
37 | template <typename T> struct DoubleLength; |
38 | |
39 | template <> struct DoubleLength<uint16_t> { |
40 | using Type = uint32_t; |
41 | }; |
42 | |
43 | template <> struct DoubleLength<uint32_t> { |
44 | using Type = uint64_t; |
45 | }; |
46 | |
47 | template <> struct DoubleLength<uint64_t> { |
48 | using Type = UInt128; |
49 | }; |
50 | |
51 | // Correctly rounded IEEE 754 HYPOT(x, y) with round to nearest, ties to even. |
52 | // |
53 | // Algorithm: |
54 | // - Let a = max(|x|, |y|), b = min(|x|, |y|), then we have that: |
55 | // a <= sqrt(a^2 + b^2) <= min(a + b, a*sqrt(2)) |
56 | // 1. So if b < eps(a)/2, then HYPOT(x, y) = a. |
57 | // |
58 | // - Moreover, the exponent part of HYPOT(x, y) is either the same or 1 more |
59 | // than the exponent part of a. |
60 | // |
61 | // 2. For the remaining cases, we will use the digit-by-digit (shift-and-add) |
62 | // algorithm to compute SQRT(Z): |
63 | // |
64 | // - For Y = y0.y1...yn... = SQRT(Z), |
65 | // let Y(n) = y0.y1...yn be the first n fractional digits of Y. |
66 | // |
67 | // - The nth scaled residual R(n) is defined to be: |
68 | // R(n) = 2^n * (Z - Y(n)^2) |
69 | // |
70 | // - Since Y(n) = Y(n - 1) + yn * 2^(-n), the scaled residual |
71 | // satisfies the following recurrence formula: |
72 | // R(n) = 2*R(n - 1) - yn*(2*Y(n - 1) + 2^(-n)), |
73 | // with the initial conditions: |
74 | // Y(0) = y0, and R(0) = Z - y0. |
75 | // |
76 | // - So the nth fractional digit of Y = SQRT(Z) can be decided by: |
77 | // yn = 1 if 2*R(n - 1) >= 2*Y(n - 1) + 2^(-n), |
78 | // 0 otherwise. |
79 | // |
80 | // 3. Precision analysis: |
81 | // |
82 | // - Notice that in the decision function: |
83 | // 2*R(n - 1) >= 2*Y(n - 1) + 2^(-n), |
84 | // the right hand side only uses up to the 2^(-n)-bit, and both sides are |
85 | // non-negative, so R(n - 1) can be truncated at the 2^(-(n + 1))-bit, so |
86 | // that 2*R(n - 1) is corrected up to the 2^(-n)-bit. |
87 | // |
88 | // - Thus, in order to round SQRT(a^2 + b^2) correctly up to n-fractional |
89 | // bits, we need to perform the summation (a^2 + b^2) correctly up to (2n + |
90 | // 2)-fractional bits, and the remaining bits are sticky bits (i.e. we only |
91 | // care if they are 0 or > 0), and the comparisons, additions/subtractions |
92 | // can be done in n-fractional bits precision. |
93 | // |
94 | // - For single precision (float), we can use uint64_t to store the sum a^2 + |
95 | // b^2 exact up to (2n + 2)-fractional bits. |
96 | // |
97 | // - Then we can feed this sum into the digit-by-digit algorithm for SQRT(Z) |
98 | // described above. |
99 | // |
100 | // |
101 | // Special cases: |
102 | // - HYPOT(x, y) is +Inf if x or y is +Inf or -Inf; else |
103 | // - HYPOT(x, y) is NaN if x or y is NaN. |
104 | // |
105 | template <typename T, cpp::enable_if_t<cpp::is_floating_point_v<T>, int> = 0> |
106 | LIBC_INLINE T hypot(T x, T y) { |
107 | using FPBits_t = FPBits<T>; |
108 | using StorageType = typename FPBits<T>::StorageType; |
109 | using DStorageType = typename DoubleLength<StorageType>::Type; |
110 | |
111 | FPBits_t x_bits(x), y_bits(y); |
112 | |
113 | if (x_bits.is_inf() || y_bits.is_inf()) { |
114 | return FPBits_t::inf().get_val(); |
115 | } |
116 | if (x_bits.is_nan()) { |
117 | return x; |
118 | } |
119 | if (y_bits.is_nan()) { |
120 | return y; |
121 | } |
122 | |
123 | uint16_t x_exp = x_bits.get_biased_exponent(); |
124 | uint16_t y_exp = y_bits.get_biased_exponent(); |
125 | uint16_t exp_diff = (x_exp > y_exp) ? (x_exp - y_exp) : (y_exp - x_exp); |
126 | |
127 | if ((exp_diff >= FPBits_t::FRACTION_LEN + 2) || (x == 0) || (y == 0)) { |
128 | return abs(x) + abs(y); |
129 | } |
130 | |
131 | uint16_t a_exp, b_exp, out_exp; |
132 | StorageType a_mant, b_mant; |
133 | DStorageType a_mant_sq, b_mant_sq; |
134 | bool sticky_bits; |
135 | |
136 | if (abs(x) >= abs(y)) { |
137 | a_exp = x_exp; |
138 | a_mant = x_bits.get_mantissa(); |
139 | b_exp = y_exp; |
140 | b_mant = y_bits.get_mantissa(); |
141 | } else { |
142 | a_exp = y_exp; |
143 | a_mant = y_bits.get_mantissa(); |
144 | b_exp = x_exp; |
145 | b_mant = x_bits.get_mantissa(); |
146 | } |
147 | |
148 | out_exp = a_exp; |
149 | |
150 | // Add an extra bit to simplify the final rounding bit computation. |
151 | constexpr StorageType ONE = StorageType(1) << (FPBits_t::FRACTION_LEN + 1); |
152 | |
153 | a_mant <<= 1; |
154 | b_mant <<= 1; |
155 | |
156 | StorageType leading_one; |
157 | int y_mant_width; |
158 | if (a_exp != 0) { |
159 | leading_one = ONE; |
160 | a_mant |= ONE; |
161 | y_mant_width = FPBits_t::FRACTION_LEN + 1; |
162 | } else { |
163 | leading_one = internal::find_leading_one(a_mant, y_mant_width); |
164 | a_exp = 1; |
165 | } |
166 | |
167 | if (b_exp != 0) { |
168 | b_mant |= ONE; |
169 | } else { |
170 | b_exp = 1; |
171 | } |
172 | |
173 | a_mant_sq = static_cast<DStorageType>(a_mant) * a_mant; |
174 | b_mant_sq = static_cast<DStorageType>(b_mant) * b_mant; |
175 | |
176 | // At this point, a_exp >= b_exp > a_exp - 25, so in order to line up aSqMant |
177 | // and bSqMant, we need to shift bSqMant to the right by (a_exp - b_exp) bits. |
178 | // But before that, remember to store the losing bits to sticky. |
179 | // The shift length is for a^2 and b^2, so it's double of the exponent |
180 | // difference between a and b. |
181 | uint16_t shift_length = static_cast<uint16_t>(2 * (a_exp - b_exp)); |
182 | sticky_bits = |
183 | ((b_mant_sq & ((DStorageType(1) << shift_length) - DStorageType(1))) != |
184 | DStorageType(0)); |
185 | b_mant_sq >>= shift_length; |
186 | |
187 | DStorageType sum = a_mant_sq + b_mant_sq; |
188 | if (sum >= (DStorageType(1) << (2 * y_mant_width + 2))) { |
189 | // a^2 + b^2 >= 4* leading_one^2, so we will need an extra bit to the left. |
190 | if (leading_one == ONE) { |
191 | // For normal result, we discard the last 2 bits of the sum and increase |
192 | // the exponent. |
193 | sticky_bits = sticky_bits || ((sum & 0x3U) != 0); |
194 | sum >>= 2; |
195 | ++out_exp; |
196 | if (out_exp >= FPBits_t::MAX_BIASED_EXPONENT) { |
197 | if (int round_mode = quick_get_round(); |
198 | round_mode == FE_TONEAREST || round_mode == FE_UPWARD) |
199 | return FPBits_t::inf().get_val(); |
200 | return FPBits_t::max_normal().get_val(); |
201 | } |
202 | } else { |
203 | // For denormal result, we simply move the leading bit of the result to |
204 | // the left by 1. |
205 | leading_one <<= 1; |
206 | ++y_mant_width; |
207 | } |
208 | } |
209 | |
210 | StorageType y_new = leading_one; |
211 | StorageType r = static_cast<StorageType>(sum >> y_mant_width) - leading_one; |
212 | StorageType tail_bits = static_cast<StorageType>(sum) & (leading_one - 1); |
213 | |
214 | for (StorageType current_bit = leading_one >> 1; current_bit; |
215 | current_bit >>= 1) { |
216 | r = (r << 1) + ((tail_bits & current_bit) ? 1 : 0); |
217 | StorageType tmp = (y_new << 1) + current_bit; // 2*y_new(n - 1) + 2^(-n) |
218 | if (r >= tmp) { |
219 | r -= tmp; |
220 | y_new += current_bit; |
221 | } |
222 | } |
223 | |
224 | bool round_bit = y_new & StorageType(1); |
225 | bool lsb = y_new & StorageType(2); |
226 | |
227 | if (y_new >= ONE) { |
228 | y_new -= ONE; |
229 | |
230 | if (out_exp == 0) { |
231 | out_exp = 1; |
232 | } |
233 | } |
234 | |
235 | y_new >>= 1; |
236 | |
237 | // Round to the nearest, tie to even. |
238 | int round_mode = quick_get_round(); |
239 | switch (round_mode) { |
240 | case FE_TONEAREST: |
241 | // Round to nearest, ties to even |
242 | if (round_bit && (lsb || sticky_bits || (r != 0))) |
243 | ++y_new; |
244 | break; |
245 | case FE_UPWARD: |
246 | if (round_bit || sticky_bits || (r != 0)) |
247 | ++y_new; |
248 | break; |
249 | } |
250 | |
251 | if (y_new >= (ONE >> 1)) { |
252 | y_new -= ONE >> 1; |
253 | ++out_exp; |
254 | if (out_exp >= FPBits_t::MAX_BIASED_EXPONENT) { |
255 | if (round_mode == FE_TONEAREST || round_mode == FE_UPWARD) |
256 | return FPBits_t::inf().get_val(); |
257 | return FPBits_t::max_normal().get_val(); |
258 | } |
259 | } |
260 | |
261 | y_new |= static_cast<StorageType>(out_exp) << FPBits_t::FRACTION_LEN; |
262 | return cpp::bit_cast<T>(y_new); |
263 | } |
264 | |
265 | } // namespace fputil |
266 | } // namespace LIBC_NAMESPACE |
267 | |
268 | #endif // LLVM_LIBC_SRC___SUPPORT_FPUTIL_HYPOT_H |
269 | |