1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
use crate::{
add::{add_in_place, add_word_in_place, sub_in_place, sub_one_in_place},
arch::word::{DoubleWord, Word},
div,
fast_div::FastDivideNormalized2,
memory::{self, Memory},
mul::add_mul_word_in_place,
primitive::{double_word, extend_word, highest_dword, split_dword, WORD_BITS},
shift::shr_in_place_with_carry,
sqr,
};
use alloc::alloc::Layout;
use dashu_base::{DivRem, SquareRootRem};
// n is the size of the output, or half the size of the input
pub fn memory_requirement_sqrt_rem(n: usize) -> Layout {
if n == 2 {
memory::zero_layout()
} else {
// We need to perform a squaring with n words and an n by n/2 division
memory::max_layout(
sqr::memory_requirement_exact(n),
div::memory_requirement_exact(n, n - n / 2),
)
}
}
// Requires a is normalized to 2n words (length must be even)
// Returns the carry of the remainder
pub fn sqrt_rem(b: &mut [Word], a: &mut [Word], memory: &mut Memory) -> bool {
debug_assert!(a.len() % 2 == 0);
debug_assert!(a.len() >= 4, "use native sqrt when a has less than 2 words");
debug_assert!(a.len() == b.len() * 2);
// shortcut when a has exactly 4 words
if a.len() == 4 {
return sqrt_rem_42(b, a);
}
/*
* the "Karatsuba Square Root" algorithm:
* assume n = a*B^2 + b1*B + b0, B=2^k, a has 2k bits and
* is normalized (the top two bits of a are not all zeros)
* 1. calculate sqrt on high part:
* s1, r1 = sqrt_rem(a) (r1 <= 2*s1)
* 2. estimate the root with low part
* q, u = div_rem(r1*B + b1, 2*s1)
* s = s1*B + q
* r = u*B + b0 - q^2
* at this step, since a is normalized, we have s1 >= B/2,
* therefore q <= floor((r1*B + b1) / B) <= r1 <= 2*s1
* also notice b1 < B <= 2*s1, so q <= B
*
* 3. if a3 is normalized, then s is either correct or 1 too big.
* r is negative in the latter case, needs adjustment
* if r < 0 {
* r += 2*s - 1
* s -= 1
* }
*
* Reference: Zimmermann, P. (1999). Karatsuba square root (Doctoral dissertation, INRIA).
* https://hal.inria.fr/inria-00072854/en/
*/
let n = a.len() / 2; // the length of a
let split = n / 2; // the length of b0
// step1: sqrt on the higher half
// afterwards, s1 = b[split..], r1 = a[2*split..split + n]
let r1_top = sqrt_rem(&mut b[split..], &mut a[2 * split..], memory);
if r1_top {
// if the remainder `r1` has a carry, subtract `s1` from it so that the carry is removed
// so later when calculate 2*q = (r1*B + b1) / s1, the result is actually one less
let carry = sub_in_place(&mut a[2 * split..split + n], &b[split..]);
debug_assert!(carry);
}
// step2: estimate the result with lower half
let fast_div_top = FastDivideNormalized2::new(highest_dword(b));
let carry = div::div_rem_in_place(&mut a[split..split + n], &b[split..], fast_div_top, memory);
let (a_lo, a_hi) = a.split_at_mut(n);
b[..split].copy_from_slice(&a_hi[..split]);
// by now 2*q = b[..split], u = a[split..n], carry is true only if r1 >= s1.
// also notice that r1 <= 2 * s1, if r1 was subtracted by s1, then r1 <= s1.
// so r_top and carry are both true only if r1 == 2 * s1 at the beginning.
// the top bit of q is true if either r_top or carry is true, but not both
let _ =
shr_in_place_with_carry(&mut b[..split], 1, ((r1_top ^ carry) as Word) << (WORD_BITS - 1));
let q_top = r1_top && carry; // true only when q = B, and then b[..split] = 0
let mut c = 0i8; // stores final carry (top bit) of the remainder
if a_hi[0] & 1 != 0 {
// this step fixes the error in u caused by using s1 as divisor instead of 2*s1
c = add_in_place(&mut a_lo[split..], &b[split..]) as i8;
}
// store q^2 in high part of a, ignoring q_top.
// afterwards, the q_top flag will be considered in the subtraction,
a_hi.fill(0);
if !q_top {
// if q_top is True, then q^2 = B^2, so we don't need to do squaring
if split == 1 {
let (b2_lo, b2_hi) = split_dword(extend_word(b[0]) * extend_word(b[0]));
a_hi[0] = b2_lo;
a_hi[1] = b2_hi;
} else {
sqr::square(&mut a_hi[..2 * split], &b[..split], memory);
}
}
if 2 * split < n {
a_hi[2 * split] = q_top as Word;
} else {
c -= q_top as i8;
}
c -= sub_in_place(a_lo, a_hi) as i8;
// step3: fix the estimation error if necessary
if c < 0 {
// r += 2*s - 1; s -= 1;
// apply the q_top to s first, and then adjust s and r
let overflow = add_word_in_place(&mut b[split..], q_top as _);
c += add_mul_word_in_place(a_lo, 2, b) as i8 + 2 * overflow as i8;
c -= sub_one_in_place(a_lo) as i8;
let borrow = sub_one_in_place(b);
debug_assert!(!(overflow ^ borrow)); // borrow should happen if and only if when overflow is true
}
c > 0
}
// Special case when a has exactly 4 Words
fn sqrt_rem_42(b: &mut [Word], a: &mut [Word]) -> bool {
debug_assert!(a.len() == 4 && b.len() == 2);
// see sqrt_rem() for algorithm explanation
// step1: sqrt on the higher half
let (s1, r1) = highest_dword(a).sqrt_rem();
let s1 = s1 as Word;
// step2: estimate the result with lower half
// here r0 = (r1*B + b1) / 2
let (r1_lo, r1_hi) = split_dword(r1);
let r0_hi = r1_hi << (WORD_BITS - 1) | r1_lo >> 1;
let r0_lo = r1_lo << (WORD_BITS - 1) | a[1] >> 1;
let (mut q, mut u) = double_word(r0_lo, r0_hi).div_rem(s1 as DoubleWord);
if q >> WORD_BITS > 0 {
// if q >= B (then q = B), reduce the overestimate
q -= 1;
u += s1 as DoubleWord;
}
u = u << 1 | (a[1] & 1) as DoubleWord;
let q = q as Word; // now q must fit in a Word
let (u_lo, u_hi) = split_dword(u);
let mut s = double_word(q, s1);
let q2 = extend_word(q) * extend_word(q);
let (mut r, borrow) = double_word(a[0], u_lo).overflowing_sub(q2);
let mut c: i8 = u_hi as i8 - borrow as i8;
// step3: fix the estimation error if necessary
if c < 0 {
let (new_r, c1) = r.overflowing_add(s);
s -= 1;
let (new_r, c2) = new_r.overflowing_add(s);
c += c1 as i8 + c2 as i8;
r = new_r;
}
let (r_lo, r_hi) = split_dword(r);
let (s_lo, s_hi) = split_dword(s);
a[0] = r_lo;
a[1] = r_hi;
b[0] = s_lo;
b[1] = s_hi;
c > 0
}