xref: /aosp_15_r20/external/llvm-libc/src/__support/FPUtil/Hypot.h (revision 71db0c75aadcf003ffe3238005f61d7618a3fead)
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/macros/config.h"
20 #include "src/__support/uint128.h"
21 
22 namespace LIBC_NAMESPACE_DECL {
23 namespace fputil {
24 
25 namespace internal {
26 
27 template <typename T>
find_leading_one(T mant,int & shift_length)28 LIBC_INLINE T find_leading_one(T mant, int &shift_length) {
29   shift_length = 0;
30   if (mant > 0) {
31     shift_length = (sizeof(mant) * 8) - 1 - cpp::countl_zero(mant);
32   }
33   return T(1) << shift_length;
34 }
35 
36 } // namespace internal
37 
38 template <typename T> struct DoubleLength;
39 
40 template <> struct DoubleLength<uint16_t> {
41   using Type = uint32_t;
42 };
43 
44 template <> struct DoubleLength<uint32_t> {
45   using Type = uint64_t;
46 };
47 
48 template <> struct DoubleLength<uint64_t> {
49   using Type = UInt128;
50 };
51 
52 // Correctly rounded IEEE 754 HYPOT(x, y) with round to nearest, ties to even.
53 //
54 // Algorithm:
55 //   -  Let a = max(|x|, |y|), b = min(|x|, |y|), then we have that:
56 //          a <= sqrt(a^2 + b^2) <= min(a + b, a*sqrt(2))
57 //   1. So if b < eps(a)/2, then HYPOT(x, y) = a.
58 //
59 //   -  Moreover, the exponent part of HYPOT(x, y) is either the same or 1 more
60 //      than the exponent part of a.
61 //
62 //   2. For the remaining cases, we will use the digit-by-digit (shift-and-add)
63 //      algorithm to compute SQRT(Z):
64 //
65 //   -  For Y = y0.y1...yn... = SQRT(Z),
66 //      let Y(n) = y0.y1...yn be the first n fractional digits of Y.
67 //
68 //   -  The nth scaled residual R(n) is defined to be:
69 //          R(n) = 2^n * (Z - Y(n)^2)
70 //
71 //   -  Since Y(n) = Y(n - 1) + yn * 2^(-n), the scaled residual
72 //      satisfies the following recurrence formula:
73 //          R(n) = 2*R(n - 1) - yn*(2*Y(n - 1) + 2^(-n)),
74 //      with the initial conditions:
75 //          Y(0) = y0, and R(0) = Z - y0.
76 //
77 //   -  So the nth fractional digit of Y = SQRT(Z) can be decided by:
78 //          yn = 1  if 2*R(n - 1) >= 2*Y(n - 1) + 2^(-n),
79 //               0  otherwise.
80 //
81 //   3. Precision analysis:
82 //
83 //   -  Notice that in the decision function:
84 //          2*R(n - 1) >= 2*Y(n - 1) + 2^(-n),
85 //      the right hand side only uses up to the 2^(-n)-bit, and both sides are
86 //      non-negative, so R(n - 1) can be truncated at the 2^(-(n + 1))-bit, so
87 //      that 2*R(n - 1) is corrected up to the 2^(-n)-bit.
88 //
89 //   -  Thus, in order to round SQRT(a^2 + b^2) correctly up to n-fractional
90 //      bits, we need to perform the summation (a^2 + b^2) correctly up to (2n +
91 //      2)-fractional bits, and the remaining bits are sticky bits (i.e. we only
92 //      care if they are 0 or > 0), and the comparisons, additions/subtractions
93 //      can be done in n-fractional bits precision.
94 //
95 //   -  For single precision (float), we can use uint64_t to store the sum a^2 +
96 //      b^2 exact up to (2n + 2)-fractional bits.
97 //
98 //   -  Then we can feed this sum into the digit-by-digit algorithm for SQRT(Z)
99 //      described above.
100 //
101 //
102 // Special cases:
103 //   - HYPOT(x, y) is +Inf if x or y is +Inf or -Inf; else
104 //   - HYPOT(x, y) is NaN if x or y is NaN.
105 //
106 template <typename T, cpp::enable_if_t<cpp::is_floating_point_v<T>, int> = 0>
107 LIBC_INLINE T hypot(T x, T y) {
108   using FPBits_t = FPBits<T>;
109   using StorageType = typename FPBits<T>::StorageType;
110   using DStorageType = typename DoubleLength<StorageType>::Type;
111 
112   FPBits_t x_abs = FPBits_t(x).abs();
113   FPBits_t y_abs = FPBits_t(y).abs();
114 
115   bool x_abs_larger = x_abs.uintval() >= y_abs.uintval();
116 
117   FPBits_t a_bits = x_abs_larger ? x_abs : y_abs;
118   FPBits_t b_bits = x_abs_larger ? y_abs : x_abs;
119 
120   if (LIBC_UNLIKELY(a_bits.is_inf_or_nan())) {
121     if (x_abs.is_signaling_nan() || y_abs.is_signaling_nan()) {
122       fputil::raise_except_if_required(FE_INVALID);
123       return FPBits_t::quiet_nan().get_val();
124     }
125     if (x_abs.is_inf() || y_abs.is_inf())
126       return FPBits_t::inf().get_val();
127     if (x_abs.is_nan())
128       return x;
129     // y is nan
130     return y;
131   }
132 
133   uint16_t a_exp = a_bits.get_biased_exponent();
134   uint16_t b_exp = b_bits.get_biased_exponent();
135 
136   if ((a_exp - b_exp >= FPBits_t::FRACTION_LEN + 2) || (x == 0) || (y == 0))
137     return x_abs.get_val() + y_abs.get_val();
138 
139   uint64_t out_exp = a_exp;
140   StorageType a_mant = a_bits.get_mantissa();
141   StorageType b_mant = b_bits.get_mantissa();
142   DStorageType a_mant_sq, b_mant_sq;
143   bool sticky_bits;
144 
145   // Add an extra bit to simplify the final rounding bit computation.
146   constexpr StorageType ONE = StorageType(1) << (FPBits_t::FRACTION_LEN + 1);
147 
148   a_mant <<= 1;
149   b_mant <<= 1;
150 
151   StorageType leading_one;
152   int y_mant_width;
153   if (a_exp != 0) {
154     leading_one = ONE;
155     a_mant |= ONE;
156     y_mant_width = FPBits_t::FRACTION_LEN + 1;
157   } else {
158     leading_one = internal::find_leading_one(a_mant, y_mant_width);
159     a_exp = 1;
160   }
161 
162   if (b_exp != 0)
163     b_mant |= ONE;
164   else
165     b_exp = 1;
166 
167   a_mant_sq = static_cast<DStorageType>(a_mant) * a_mant;
168   b_mant_sq = static_cast<DStorageType>(b_mant) * b_mant;
169 
170   // At this point, a_exp >= b_exp > a_exp - 25, so in order to line up aSqMant
171   // and bSqMant, we need to shift bSqMant to the right by (a_exp - b_exp) bits.
172   // But before that, remember to store the losing bits to sticky.
173   // The shift length is for a^2 and b^2, so it's double of the exponent
174   // difference between a and b.
175   uint16_t shift_length = static_cast<uint16_t>(2 * (a_exp - b_exp));
176   sticky_bits =
177       ((b_mant_sq & ((DStorageType(1) << shift_length) - DStorageType(1))) !=
178        DStorageType(0));
179   b_mant_sq >>= shift_length;
180 
181   DStorageType sum = a_mant_sq + b_mant_sq;
182   if (sum >= (DStorageType(1) << (2 * y_mant_width + 2))) {
183     // a^2 + b^2 >= 4* leading_one^2, so we will need an extra bit to the left.
184     if (leading_one == ONE) {
185       // For normal result, we discard the last 2 bits of the sum and increase
186       // the exponent.
187       sticky_bits = sticky_bits || ((sum & 0x3U) != 0);
188       sum >>= 2;
189       ++out_exp;
190       if (out_exp >= FPBits_t::MAX_BIASED_EXPONENT) {
191         if (int round_mode = quick_get_round();
192             round_mode == FE_TONEAREST || round_mode == FE_UPWARD)
193           return FPBits_t::inf().get_val();
194         return FPBits_t::max_normal().get_val();
195       }
196     } else {
197       // For denormal result, we simply move the leading bit of the result to
198       // the left by 1.
199       leading_one <<= 1;
200       ++y_mant_width;
201     }
202   }
203 
204   StorageType y_new = leading_one;
205   StorageType r = static_cast<StorageType>(sum >> y_mant_width) - leading_one;
206   StorageType tail_bits = static_cast<StorageType>(sum) & (leading_one - 1);
207 
208   for (StorageType current_bit = leading_one >> 1; current_bit;
209        current_bit >>= 1) {
210     r = (r << 1) + ((tail_bits & current_bit) ? 1 : 0);
211     StorageType tmp = (y_new << 1) + current_bit; // 2*y_new(n - 1) + 2^(-n)
212     if (r >= tmp) {
213       r -= tmp;
214       y_new += current_bit;
215     }
216   }
217 
218   bool round_bit = y_new & StorageType(1);
219   bool lsb = y_new & StorageType(2);
220 
221   if (y_new >= ONE) {
222     y_new -= ONE;
223 
224     if (out_exp == 0) {
225       out_exp = 1;
226     }
227   }
228 
229   y_new >>= 1;
230 
231   // Round to the nearest, tie to even.
232   int round_mode = quick_get_round();
233   switch (round_mode) {
234   case FE_TONEAREST:
235     // Round to nearest, ties to even
236     if (round_bit && (lsb || sticky_bits || (r != 0)))
237       ++y_new;
238     break;
239   case FE_UPWARD:
240     if (round_bit || sticky_bits || (r != 0))
241       ++y_new;
242     break;
243   }
244 
245   if (y_new >= (ONE >> 1)) {
246     y_new -= ONE >> 1;
247     ++out_exp;
248     if (out_exp >= FPBits_t::MAX_BIASED_EXPONENT) {
249       if (round_mode == FE_TONEAREST || round_mode == FE_UPWARD)
250         return FPBits_t::inf().get_val();
251       return FPBits_t::max_normal().get_val();
252     }
253   }
254 
255   y_new |= static_cast<StorageType>(out_exp) << FPBits_t::FRACTION_LEN;
256 
257   if (!(round_bit || sticky_bits || (r != 0)))
258     fputil::clear_except_if_required(FE_INEXACT);
259 
260   return cpp::bit_cast<T>(y_new);
261 }
262 
263 } // namespace fputil
264 } // namespace LIBC_NAMESPACE_DECL
265 
266 #endif // LLVM_LIBC_SRC___SUPPORT_FPUTIL_HYPOT_H
267