1 | //! Floating point division routines. |
2 | //! |
3 | //! This module documentation gives an overview of the method used. More documentation is inline. |
4 | //! |
5 | //! # Relevant notation |
6 | //! |
7 | //! - `m_a`: the mantissa of `a`, in base 2 |
8 | //! - `p_a`: the exponent of `a`, in base 2. I.e. `a = m_a * 2^p_a` |
9 | //! - `uqN` (e.g. `uq1`): this refers to Q notation for fixed-point numbers. UQ1.31 is an unsigned |
10 | //! fixed-point number with 1 integral bit, and 31 decimal bits. A `uqN` variable of type `uM` |
11 | //! will have N bits of integer and M-N bits of fraction. |
12 | //! - `hw`: half width, i.e. for `f64` this will be a `u32`. |
13 | //! - `x` is the best estimate of `1/m_b` |
14 | //! |
15 | //! # Method Overview |
16 | //! |
17 | //! Division routines must solve for `a / b`, which is `res = m_a*2^p_a / m_b*2^p_b`. The basic |
18 | //! process is as follows: |
19 | //! |
20 | //! - Rearange the exponent and significand to simplify the operations: |
21 | //! `res = (m_a / m_b) * 2^{p_a - p_b}`. |
22 | //! - Check for early exits (infinity, zero, etc). |
23 | //! - If `a` or `b` are subnormal, normalize by shifting the mantissa and adjusting the exponent. |
24 | //! - Set the implicit bit so math is correct. |
25 | //! - Shift mantissa significant digits (with implicit bit) fully left such that fixed-point UQ1 |
26 | //! or UQ0 numbers can be used for mantissa math. These will have greater precision than the |
27 | //! actual mantissa, which is important for correct rounding. |
28 | //! - Calculate the reciprocal of `m_b`, `x`. |
29 | //! - Use the reciprocal to multiply rather than divide: `res = m_a * x_b * 2^{p_a - p_b}`. |
30 | //! - Reapply rounding. |
31 | //! |
32 | //! # Reciprocal calculation |
33 | //! |
34 | //! Calculating the reciprocal is the most complicated part of this process. It uses the |
35 | //! [Newton-Raphson method], which picks an initial estimation (of the reciprocal) and performs |
36 | //! a number of iterations to increase its precision. |
37 | //! |
38 | //! In general, Newton's method takes the following form: |
39 | //! |
40 | //! ```text |
41 | //! `x_n` is a guess or the result of a previous iteration. Increasing `n` converges to the |
42 | //! desired result. |
43 | //! |
44 | //! The result approaches a zero of `f(x)` by applying a correction to the previous gues. |
45 | //! |
46 | //! x_{n+1} = x_n - f(x_n) / f'(x_n) |
47 | //! ``` |
48 | //! |
49 | //! Applying this to find the reciprocal: |
50 | //! |
51 | //! ```text |
52 | //! 1 / x = b |
53 | //! |
54 | //! Rearrange so we can solve by finding a zero |
55 | //! 0 = (1 / x) - b = f(x) |
56 | //! |
57 | //! f'(x) = -x^{-2} |
58 | //! |
59 | //! x_{n+1} = 2*x_n - b*x_n^2 |
60 | //! ``` |
61 | //! |
62 | //! This is a process that can be repeated to calculate the reciprocal with enough precision to |
63 | //! achieve a correctly rounded result for the overall division operation. The maximum required |
64 | //! number of iterations is known since precision doubles with each iteration. |
65 | //! |
66 | //! # Half-width operations |
67 | //! |
68 | //! Calculating the reciprocal requires widening multiplication and performing arithmetic on the |
69 | //! results, meaning that emulated integer arithmetic on `u128` (for `f64`) and `u256` (for `f128`) |
70 | //! gets used instead of native math. |
71 | //! |
72 | //! To make this more efficient, all but the final operation can be computed using half-width |
73 | //! integers. For example, rather than computing four iterations using 128-bit integers for `f64`, |
74 | //! we can instead perform three iterations using native 64-bit integers and only one final |
75 | //! iteration using the full 128 bits. |
76 | //! |
77 | //! This works because of precision doubling. Some leeway is allowed here because the fixed-point |
78 | //! number has more bits than the final mantissa will. |
79 | //! |
80 | //! [Newton-Raphson method]: https://en.wikipedia.org/wiki/Newton%27s_method |
81 | |
82 | use super::HalfRep; |
83 | use crate::float::Float; |
84 | use crate::int::{CastFrom, CastInto, DInt, HInt, Int, MinInt}; |
85 | use core::mem::size_of; |
86 | use core::ops; |
87 | |
88 | fn div<F: Float>(a: F, b: F) -> F |
89 | where |
90 | F::Int: CastInto<i32>, |
91 | F::Int: From<HalfRep<F>>, |
92 | F::Int: From<u8>, |
93 | F::Int: HInt + DInt, |
94 | <F::Int as HInt>::D: ops::Shr<u32, Output = <F::Int as HInt>::D>, |
95 | F::Int: From<u32>, |
96 | u16: CastInto<F::Int>, |
97 | i32: CastInto<F::Int>, |
98 | u32: CastInto<F::Int>, |
99 | u128: CastInto<HalfRep<F>>, |
100 | { |
101 | let one = F::Int::ONE; |
102 | let zero = F::Int::ZERO; |
103 | let one_hw = HalfRep::<F>::ONE; |
104 | let zero_hw = HalfRep::<F>::ZERO; |
105 | let hw = F::BITS / 2; |
106 | let lo_mask = F::Int::MAX >> hw; |
107 | |
108 | let significand_bits = F::SIG_BITS; |
109 | // Saturated exponent, representing infinity |
110 | let exponent_sat: F::Int = F::EXP_SAT.cast(); |
111 | |
112 | let exponent_bias = F::EXP_BIAS; |
113 | let implicit_bit = F::IMPLICIT_BIT; |
114 | let significand_mask = F::SIG_MASK; |
115 | let sign_bit = F::SIGN_MASK; |
116 | let abs_mask = sign_bit - one; |
117 | let exponent_mask = F::EXP_MASK; |
118 | let inf_rep = exponent_mask; |
119 | let quiet_bit = implicit_bit >> 1; |
120 | let qnan_rep = exponent_mask | quiet_bit; |
121 | let (mut half_iterations, full_iterations) = get_iterations::<F>(); |
122 | let recip_precision = reciprocal_precision::<F>(); |
123 | |
124 | if F::BITS == 128 { |
125 | // FIXME(tgross35): f128 seems to require one more half iteration than expected |
126 | half_iterations += 1; |
127 | } |
128 | |
129 | let a_rep = a.to_bits(); |
130 | let b_rep = b.to_bits(); |
131 | |
132 | // Exponent numeric representationm not accounting for bias |
133 | let a_exponent = (a_rep >> significand_bits) & exponent_sat; |
134 | let b_exponent = (b_rep >> significand_bits) & exponent_sat; |
135 | let quotient_sign = (a_rep ^ b_rep) & sign_bit; |
136 | |
137 | let mut a_significand = a_rep & significand_mask; |
138 | let mut b_significand = b_rep & significand_mask; |
139 | |
140 | // The exponent of our final result in its encoded form |
141 | let mut res_exponent: i32 = |
142 | i32::cast_from(a_exponent) - i32::cast_from(b_exponent) + (exponent_bias as i32); |
143 | |
144 | // Detect if a or b is zero, denormal, infinity, or NaN. |
145 | if a_exponent.wrapping_sub(one) >= (exponent_sat - one) |
146 | || b_exponent.wrapping_sub(one) >= (exponent_sat - one) |
147 | { |
148 | let a_abs = a_rep & abs_mask; |
149 | let b_abs = b_rep & abs_mask; |
150 | |
151 | // NaN / anything = qNaN |
152 | if a_abs > inf_rep { |
153 | return F::from_bits(a_rep | quiet_bit); |
154 | } |
155 | |
156 | // anything / NaN = qNaN |
157 | if b_abs > inf_rep { |
158 | return F::from_bits(b_rep | quiet_bit); |
159 | } |
160 | |
161 | if a_abs == inf_rep { |
162 | if b_abs == inf_rep { |
163 | // infinity / infinity = NaN |
164 | return F::from_bits(qnan_rep); |
165 | } else { |
166 | // infinity / anything else = +/- infinity |
167 | return F::from_bits(a_abs | quotient_sign); |
168 | } |
169 | } |
170 | |
171 | // anything else / infinity = +/- 0 |
172 | if b_abs == inf_rep { |
173 | return F::from_bits(quotient_sign); |
174 | } |
175 | |
176 | if a_abs == zero { |
177 | if b_abs == zero { |
178 | // zero / zero = NaN |
179 | return F::from_bits(qnan_rep); |
180 | } else { |
181 | // zero / anything else = +/- zero |
182 | return F::from_bits(quotient_sign); |
183 | } |
184 | } |
185 | |
186 | // anything else / zero = +/- infinity |
187 | if b_abs == zero { |
188 | return F::from_bits(inf_rep | quotient_sign); |
189 | } |
190 | |
191 | // a is denormal. Renormalize it and set the scale to include the necessary exponent |
192 | // adjustment. |
193 | if a_abs < implicit_bit { |
194 | let (exponent, significand) = F::normalize(a_significand); |
195 | res_exponent += exponent; |
196 | a_significand = significand; |
197 | } |
198 | |
199 | // b is denormal. Renormalize it and set the scale to include the necessary exponent |
200 | // adjustment. |
201 | if b_abs < implicit_bit { |
202 | let (exponent, significand) = F::normalize(b_significand); |
203 | res_exponent -= exponent; |
204 | b_significand = significand; |
205 | } |
206 | } |
207 | |
208 | // Set the implicit significand bit. If we fell through from the |
209 | // denormal path it was already set by normalize( ), but setting it twice |
210 | // won't hurt anything. |
211 | a_significand |= implicit_bit; |
212 | b_significand |= implicit_bit; |
213 | |
214 | // Transform to a fixed-point representation by shifting the significand to the high bits. We |
215 | // know this is in the range [1.0, 2.0] since the implicit bit is set to 1 above. |
216 | let b_uq1 = b_significand << (F::BITS - significand_bits - 1); |
217 | |
218 | // Align the significand of b as a UQ1.(n-1) fixed-point number in the range |
219 | // [1.0, 2.0) and get a UQ0.n approximate reciprocal using a small minimax |
220 | // polynomial approximation: x0 = 3/4 + 1/sqrt(2) - b/2. |
221 | // The max error for this approximation is achieved at endpoints, so |
222 | // abs(x0(b) - 1/b) <= abs(x0(1) - 1/1) = 3/4 - 1/sqrt(2) = 0.04289..., |
223 | // which is about 4.5 bits. |
224 | // The initial approximation is between x0(1.0) = 0.9571... and x0(2.0) = 0.4571... |
225 | // |
226 | // Then, refine the reciprocal estimate using a quadratically converging |
227 | // Newton-Raphson iteration: |
228 | // x_{n+1} = x_n * (2 - x_n * b) |
229 | // |
230 | // Let b be the original divisor considered "in infinite precision" and |
231 | // obtained from IEEE754 representation of function argument (with the |
232 | // implicit bit set). Corresponds to rep_t-sized b_UQ1 represented in |
233 | // UQ1.(W-1). |
234 | // |
235 | // Let b_hw be an infinitely precise number obtained from the highest (HW-1) |
236 | // bits of divisor significand (with the implicit bit set). Corresponds to |
237 | // half_rep_t-sized b_UQ1_hw represented in UQ1.(HW-1) that is a **truncated** |
238 | // version of b_UQ1. |
239 | // |
240 | // Let e_n := x_n - 1/b_hw |
241 | // E_n := x_n - 1/b |
242 | // abs(E_n) <= abs(e_n) + (1/b_hw - 1/b) |
243 | // = abs(e_n) + (b - b_hw) / (b*b_hw) |
244 | // <= abs(e_n) + 2 * 2^-HW |
245 | // |
246 | // rep_t-sized iterations may be slower than the corresponding half-width |
247 | // variant depending on the handware and whether single/double/quad precision |
248 | // is selected. |
249 | // |
250 | // NB: Using half-width iterations increases computation errors due to |
251 | // rounding, so error estimations have to be computed taking the selected |
252 | // mode into account! |
253 | let mut x_uq0 = if half_iterations > 0 { |
254 | // Starting with (n-1) half-width iterations |
255 | let b_uq1_hw: HalfRep<F> = b_uq1.hi(); |
256 | |
257 | // C is (3/4 + 1/sqrt(2)) - 1 truncated to W0 fractional bits as UQ0.HW |
258 | // with W0 being either 16 or 32 and W0 <= HW. |
259 | // That is, C is the aforementioned 3/4 + 1/sqrt(2) constant (from which |
260 | // b/2 is subtracted to obtain x0) wrapped to [0, 1) range. |
261 | let c_hw = c_hw::<F>(); |
262 | |
263 | // Check that the top bit is set, i.e. value is within `[1, 2)`. |
264 | debug_assert!(b_uq1_hw & (one_hw << (HalfRep::<F>::BITS - 1)) > zero_hw); |
265 | |
266 | // b >= 1, thus an upper bound for 3/4 + 1/sqrt(2) - b/2 is about 0.9572, |
267 | // so x0 fits to UQ0.HW without wrapping. |
268 | let mut x_uq0_hw: HalfRep<F> = |
269 | c_hw.wrapping_sub(b_uq1_hw /* exact b_hw/2 as UQ0.HW */); |
270 | |
271 | // An e_0 error is comprised of errors due to |
272 | // * x0 being an inherently imprecise first approximation of 1/b_hw |
273 | // * C_hw being some (irrational) number **truncated** to W0 bits |
274 | // Please note that e_0 is calculated against the infinitely precise |
275 | // reciprocal of b_hw (that is, **truncated** version of b). |
276 | // |
277 | // e_0 <= 3/4 - 1/sqrt(2) + 2^-W0 |
278 | // |
279 | // By construction, 1 <= b < 2 |
280 | // f(x) = x * (2 - b*x) = 2*x - b*x^2 |
281 | // f'(x) = 2 * (1 - b*x) |
282 | // |
283 | // On the [0, 1] interval, f(0) = 0, |
284 | // then it increses until f(1/b) = 1 / b, maximum on (0, 1), |
285 | // then it decreses to f(1) = 2 - b |
286 | // |
287 | // Let g(x) = x - f(x) = b*x^2 - x. |
288 | // On (0, 1/b), g(x) < 0 <=> f(x) > x |
289 | // On (1/b, 1], g(x) > 0 <=> f(x) < x |
290 | // |
291 | // For half-width iterations, b_hw is used instead of b. |
292 | for _ in 0..half_iterations { |
293 | // corr_UQ1_hw can be **larger** than 2 - b_hw*x by at most 1*Ulp |
294 | // of corr_UQ1_hw. |
295 | // "0.0 - (...)" is equivalent to "2.0 - (...)" in UQ1.(HW-1). |
296 | // On the other hand, corr_UQ1_hw should not overflow from 2.0 to 0.0 provided |
297 | // no overflow occurred earlier: ((rep_t)x_UQ0_hw * b_UQ1_hw >> HW) is |
298 | // expected to be strictly positive because b_UQ1_hw has its highest bit set |
299 | // and x_UQ0_hw should be rather large (it converges to 1/2 < 1/b_hw <= 1). |
300 | // |
301 | // Now, we should multiply UQ0.HW and UQ1.(HW-1) numbers, naturally |
302 | // obtaining an UQ1.(HW-1) number and proving its highest bit could be |
303 | // considered to be 0 to be able to represent it in UQ0.HW. |
304 | // From the above analysis of f(x), if corr_UQ1_hw would be represented |
305 | // without any intermediate loss of precision (that is, in twice_rep_t) |
306 | // x_UQ0_hw could be at most [1.]000... if b_hw is exactly 1.0 and strictly |
307 | // less otherwise. On the other hand, to obtain [1.]000..., one have to pass |
308 | // 1/b_hw == 1.0 to f(x), so this cannot occur at all without overflow (due |
309 | // to 1.0 being not representable as UQ0.HW). |
310 | // The fact corr_UQ1_hw was virtually round up (due to result of |
311 | // multiplication being **first** truncated, then negated - to improve |
312 | // error estimations) can increase x_UQ0_hw by up to 2*Ulp of x_UQ0_hw. |
313 | // |
314 | // Now, either no overflow occurred or x_UQ0_hw is 0 or 1 in its half_rep_t |
315 | // representation. In the latter case, x_UQ0_hw will be either 0 or 1 after |
316 | // any number of iterations, so just subtract 2 from the reciprocal |
317 | // approximation after last iteration. |
318 | // |
319 | // In infinite precision, with 0 <= eps1, eps2 <= U = 2^-HW: |
320 | // corr_UQ1_hw = 2 - (1/b_hw + e_n) * b_hw + 2*eps1 |
321 | // = 1 - e_n * b_hw + 2*eps1 |
322 | // x_UQ0_hw = (1/b_hw + e_n) * (1 - e_n*b_hw + 2*eps1) - eps2 |
323 | // = 1/b_hw - e_n + 2*eps1/b_hw + e_n - e_n^2*b_hw + 2*e_n*eps1 - eps2 |
324 | // = 1/b_hw + 2*eps1/b_hw - e_n^2*b_hw + 2*e_n*eps1 - eps2 |
325 | // e_{n+1} = -e_n^2*b_hw + 2*eps1/b_hw + 2*e_n*eps1 - eps2 |
326 | // = 2*e_n*eps1 - (e_n^2*b_hw + eps2) + 2*eps1/b_hw |
327 | // \------ >0 -------/ \-- >0 ---/ |
328 | // abs(e_{n+1}) <= 2*abs(e_n)*U + max(2*e_n^2 + U, 2 * U) |
329 | x_uq0_hw = next_guess(x_uq0_hw, b_uq1_hw); |
330 | } |
331 | |
332 | // For initial half-width iterations, U = 2^-HW |
333 | // Let abs(e_n) <= u_n * U, |
334 | // then abs(e_{n+1}) <= 2 * u_n * U^2 + max(2 * u_n^2 * U^2 + U, 2 * U) |
335 | // u_{n+1} <= 2 * u_n * U + max(2 * u_n^2 * U + 1, 2) |
336 | // |
337 | // Account for possible overflow (see above). For an overflow to occur for the |
338 | // first time, for "ideal" corr_UQ1_hw (that is, without intermediate |
339 | // truncation), the result of x_UQ0_hw * corr_UQ1_hw should be either maximum |
340 | // value representable in UQ0.HW or less by 1. This means that 1/b_hw have to |
341 | // be not below that value (see g(x) above), so it is safe to decrement just |
342 | // once after the final iteration. On the other hand, an effective value of |
343 | // divisor changes after this point (from b_hw to b), so adjust here. |
344 | x_uq0_hw = x_uq0_hw.wrapping_sub(one_hw); |
345 | |
346 | // Error estimations for full-precision iterations are calculated just |
347 | // as above, but with U := 2^-W and taking extra decrementing into account. |
348 | // We need at least one such iteration. |
349 | // |
350 | // Simulating operations on a twice_rep_t to perform a single final full-width |
351 | // iteration. Using ad-hoc multiplication implementations to take advantage |
352 | // of particular structure of operands. |
353 | let blo: F::Int = b_uq1 & lo_mask; |
354 | |
355 | // x_UQ0 = x_UQ0_hw * 2^HW - 1 |
356 | // x_UQ0 * b_UQ1 = (x_UQ0_hw * 2^HW) * (b_UQ1_hw * 2^HW + blo) - b_UQ1 |
357 | // |
358 | // <--- higher half ---><--- lower half ---> |
359 | // [x_UQ0_hw * b_UQ1_hw] |
360 | // + [ x_UQ0_hw * blo ] |
361 | // - [ b_UQ1 ] |
362 | // = [ result ][.... discarded ...] |
363 | let corr_uq1: F::Int = (F::Int::from(x_uq0_hw) * F::Int::from(b_uq1_hw) |
364 | + ((F::Int::from(x_uq0_hw) * blo) >> hw)) |
365 | .wrapping_sub(one) |
366 | .wrapping_neg(); // account for *possible* carry |
367 | |
368 | let lo_corr: F::Int = corr_uq1 & lo_mask; |
369 | let hi_corr: F::Int = corr_uq1 >> hw; |
370 | |
371 | // x_UQ0 * corr_UQ1 = (x_UQ0_hw * 2^HW) * (hi_corr * 2^HW + lo_corr) - corr_UQ1 |
372 | let mut x_uq0: F::Int = ((F::Int::from(x_uq0_hw) * hi_corr) << 1) |
373 | .wrapping_add((F::Int::from(x_uq0_hw) * lo_corr) >> (hw - 1)) |
374 | // 1 to account for the highest bit of corr_UQ1 can be 1 |
375 | // 1 to account for possible carry |
376 | // Just like the case of half-width iterations but with possibility |
377 | // of overflowing by one extra Ulp of x_UQ0. |
378 | .wrapping_sub(F::Int::from(2u8)); |
379 | |
380 | x_uq0 -= one; |
381 | // ... and then traditional fixup by 2 should work |
382 | |
383 | // On error estimation: |
384 | // abs(E_{N-1}) <= (u_{N-1} + 2 /* due to conversion e_n -> E_n */) * 2^-HW |
385 | // + (2^-HW + 2^-W)) |
386 | // abs(E_{N-1}) <= (u_{N-1} + 3.01) * 2^-HW |
387 | // |
388 | // Then like for the half-width iterations: |
389 | // With 0 <= eps1, eps2 < 2^-W |
390 | // E_N = 4 * E_{N-1} * eps1 - (E_{N-1}^2 * b + 4 * eps2) + 4 * eps1 / b |
391 | // abs(E_N) <= 2^-W * [ 4 * abs(E_{N-1}) + max(2 * abs(E_{N-1})^2 * 2^W + 4, 8)) ] |
392 | // abs(E_N) <= 2^-W * [ 4 * (u_{N-1} + 3.01) * 2^-HW + max(4 + 2 * (u_{N-1} + 3.01)^2, 8) ] |
393 | x_uq0 |
394 | } else { |
395 | // C is (3/4 + 1/sqrt(2)) - 1 truncated to 64 fractional bits as UQ0.n |
396 | let c: F::Int = F::Int::from(0x7504F333u32) << (F::BITS - 32); |
397 | let mut x_uq0: F::Int = c.wrapping_sub(b_uq1); |
398 | |
399 | // E_0 <= 3/4 - 1/sqrt(2) + 2 * 2^-64 |
400 | // x_uq0 |
401 | for _ in 0..full_iterations { |
402 | x_uq0 = next_guess(x_uq0, b_uq1); |
403 | } |
404 | |
405 | x_uq0 |
406 | }; |
407 | |
408 | // Finally, account for possible overflow, as explained above. |
409 | x_uq0 = x_uq0.wrapping_sub(2.cast()); |
410 | |
411 | // Suppose 1/b - P * 2^-W < x < 1/b + P * 2^-W |
412 | x_uq0 -= recip_precision.cast(); |
413 | |
414 | // Now 1/b - (2*P) * 2^-W < x < 1/b |
415 | // FIXME Is x_UQ0 still >= 0.5? |
416 | |
417 | let mut quotient_uq1: F::Int = x_uq0.widen_mul(a_significand << 1).hi(); |
418 | // Now, a/b - 4*P * 2^-W < q < a/b for q=<quotient_UQ1:dummy> in UQ1.(SB+1+W). |
419 | |
420 | // quotient_UQ1 is in [0.5, 2.0) as UQ1.(SB+1), |
421 | // adjust it to be in [1.0, 2.0) as UQ1.SB. |
422 | let mut residual_lo = if quotient_uq1 < (implicit_bit << 1) { |
423 | // Highest bit is 0, so just reinterpret quotient_UQ1 as UQ1.SB, |
424 | // effectively doubling its value as well as its error estimation. |
425 | let residual_lo = (a_significand << (significand_bits + 1)) |
426 | .wrapping_sub(quotient_uq1.wrapping_mul(b_significand)); |
427 | res_exponent -= 1; |
428 | a_significand <<= 1; |
429 | residual_lo |
430 | } else { |
431 | // Highest bit is 1 (the UQ1.(SB+1) value is in [1, 2)), convert it |
432 | // to UQ1.SB by right shifting by 1. Least significant bit is omitted. |
433 | quotient_uq1 >>= 1; |
434 | (a_significand << significand_bits).wrapping_sub(quotient_uq1.wrapping_mul(b_significand)) |
435 | }; |
436 | |
437 | // drop mutability |
438 | let quotient = quotient_uq1; |
439 | |
440 | // NB: residualLo is calculated above for the normal result case. |
441 | // It is re-computed on denormal path that is expected to be not so |
442 | // performance-sensitive. |
443 | // |
444 | // Now, q cannot be greater than a/b and can differ by at most 8*P * 2^-W + 2^-SB |
445 | // Each NextAfter() increments the floating point value by at least 2^-SB |
446 | // (more, if exponent was incremented). |
447 | // Different cases (<---> is of 2^-SB length, * = a/b that is shown as a midpoint): |
448 | // q |
449 | // | | * | | | | | |
450 | // <---> 2^t |
451 | // | | | | | * | | |
452 | // q |
453 | // To require at most one NextAfter(), an error should be less than 1.5 * 2^-SB. |
454 | // (8*P) * 2^-W + 2^-SB < 1.5 * 2^-SB |
455 | // (8*P) * 2^-W < 0.5 * 2^-SB |
456 | // P < 2^(W-4-SB) |
457 | // Generally, for at most R NextAfter() to be enough, |
458 | // P < (2*R - 1) * 2^(W-4-SB) |
459 | // For f32 (0+3): 10 < 32 (OK) |
460 | // For f32 (2+1): 32 < 74 < 32 * 3, so two NextAfter() are required |
461 | // For f64: 220 < 256 (OK) |
462 | // For f128: 4096 * 3 < 13922 < 4096 * 5 (three NextAfter() are required) |
463 | // |
464 | // If we have overflowed the exponent, return infinity |
465 | if res_exponent >= i32::cast_from(exponent_sat) { |
466 | return F::from_bits(inf_rep | quotient_sign); |
467 | } |
468 | |
469 | // Now, quotient <= the correctly-rounded result |
470 | // and may need taking NextAfter() up to 3 times (see error estimates above) |
471 | // r = a - b * q |
472 | let mut abs_result = if res_exponent > 0 { |
473 | let mut ret = quotient & significand_mask; |
474 | ret |= F::Int::from(res_exponent as u32) << significand_bits; |
475 | residual_lo <<= 1; |
476 | ret |
477 | } else { |
478 | if ((significand_bits as i32) + res_exponent) < 0 { |
479 | return F::from_bits(quotient_sign); |
480 | } |
481 | |
482 | let ret = quotient.wrapping_shr(u32::cast_from(res_exponent.wrapping_neg()) + 1); |
483 | residual_lo = a_significand |
484 | .wrapping_shl(significand_bits.wrapping_add(CastInto::<u32>::cast(res_exponent))) |
485 | .wrapping_sub(ret.wrapping_mul(b_significand) << 1); |
486 | ret |
487 | }; |
488 | |
489 | residual_lo += abs_result & one; // tie to even |
490 | // conditionally turns the below LT comparison into LTE |
491 | abs_result += u8::from(residual_lo > b_significand).into(); |
492 | |
493 | if F::BITS == 128 || (F::BITS == 32 && half_iterations > 0) { |
494 | // Do not round Infinity to NaN |
495 | abs_result += |
496 | u8::from(abs_result < inf_rep && residual_lo > (2 + 1).cast() * b_significand).into(); |
497 | } |
498 | |
499 | if F::BITS == 128 { |
500 | abs_result += |
501 | u8::from(abs_result < inf_rep && residual_lo > (4 + 1).cast() * b_significand).into(); |
502 | } |
503 | |
504 | F::from_bits(abs_result | quotient_sign) |
505 | } |
506 | |
507 | /// Calculate the number of iterations required for a float type's precision. |
508 | /// |
509 | /// This returns `(h, f)` where `h` is the number of iterations to be done using integers at half |
510 | /// the float's bit width, and `f` is the number of iterations done using integers of the float's |
511 | /// full width. This is further explained in the module documentation. |
512 | /// |
513 | /// # Requirements |
514 | /// |
515 | /// The initial estimate should have at least 8 bits of precision. If this is not true, results |
516 | /// will be inaccurate. |
517 | const fn get_iterations<F: Float>() -> (usize, usize) { |
518 | // Precision doubles with each iteration. Assume we start with 8 bits of precision. |
519 | let total_iterations: usize = F::BITS.ilog2() as usize - 2; |
520 | |
521 | if 2 * size_of::<F>() <= size_of::<*const ()>() { |
522 | // If widening multiplication will be efficient (uses word-sized integers), there is no |
523 | // reason to use half-sized iterations. |
524 | (0, total_iterations) |
525 | } else { |
526 | // Otherwise, do as many iterations as possible at half width. |
527 | (total_iterations - 1, 1) |
528 | } |
529 | } |
530 | |
531 | /// `u_n` for different precisions (with N-1 half-width iterations). |
532 | /// |
533 | /// W0 is the precision of C |
534 | /// u_0 = (3/4 - 1/sqrt(2) + 2^-W0) * 2^HW |
535 | /// |
536 | /// Estimated with bc: |
537 | /// |
538 | /// ```text |
539 | /// define half1(un) { return 2.0 * (un + un^2) / 2.0^hw + 1.0; } |
540 | /// define half2(un) { return 2.0 * un / 2.0^hw + 2.0; } |
541 | /// define full1(un) { return 4.0 * (un + 3.01) / 2.0^hw + 2.0 * (un + 3.01)^2 + 4.0; } |
542 | /// define full2(un) { return 4.0 * (un + 3.01) / 2.0^hw + 8.0; } |
543 | /// |
544 | /// | f32 (0 + 3) | f32 (2 + 1) | f64 (3 + 1) | f128 (4 + 1) |
545 | /// u_0 | < 184224974 | < 2812.1 | < 184224974 | < 791240234244348797 |
546 | /// u_1 | < 15804007 | < 242.7 | < 15804007 | < 67877681371350440 |
547 | /// u_2 | < 116308 | < 2.81 | < 116308 | < 499533100252317 |
548 | /// u_3 | < 7.31 | | < 7.31 | < 27054456580 |
549 | /// u_4 | | | | < 80.4 |
550 | /// Final (U_N) | same as u_3 | < 72 | < 218 | < 13920 |
551 | /// ```` |
552 | /// |
553 | /// Add 2 to `U_N` due to final decrement. |
554 | const fn reciprocal_precision<F: Float>() -> u16 { |
555 | let (half_iterations: usize, full_iterations: usize) = get_iterations::<F>(); |
556 | |
557 | if full_iterations < 1 { |
558 | panic!("Must have at least one full iteration" ); |
559 | } |
560 | |
561 | // FIXME(tgross35): calculate this programmatically |
562 | if F::BITS == 32 && half_iterations == 2 && full_iterations == 1 { |
563 | 74u16 |
564 | } else if F::BITS == 32 && half_iterations == 0 && full_iterations == 3 { |
565 | 10 |
566 | } else if F::BITS == 64 && half_iterations == 3 && full_iterations == 1 { |
567 | 220 |
568 | } else if F::BITS == 128 && half_iterations == 4 && full_iterations == 1 { |
569 | 13922 |
570 | } else { |
571 | panic!("Invalid number of iterations" ) |
572 | } |
573 | } |
574 | |
575 | /// The value of `C` adjusted to half width. |
576 | /// |
577 | /// C is (3/4 + 1/sqrt(2)) - 1 truncated to W0 fractional bits as UQ0.HW with W0 being either |
578 | /// 16 or 32 and W0 <= HW. That is, C is the aforementioned 3/4 + 1/sqrt(2) constant (from |
579 | /// which b/2 is subtracted to obtain x0) wrapped to [0, 1) range. |
580 | fn c_hw<F: Float>() -> HalfRep<F> |
581 | where |
582 | F::Int: DInt, |
583 | u128: CastInto<HalfRep<F>>, |
584 | { |
585 | const C_U128: u128 = 0x7504f333f9de6108b2fb1366eaa6a542; |
586 | const { C_U128 >> (u128::BITS - <HalfRep<F>>::BITS) }.cast() |
587 | } |
588 | |
589 | /// Perform one iteration at any width to approach `1/b`, given previous guess `x`. Returns |
590 | /// the next `x` as a UQ0 number. |
591 | /// |
592 | /// This is the `x_{n+1} = 2*x_n - b*x_n^2` algorithm, implemented as `x_n * (2 - b*x_n)`. It |
593 | /// uses widening multiplication to calculate the result with necessary precision. |
594 | fn next_guess<I>(x_uq0: I, b_uq1: I) -> I |
595 | where |
596 | I: Int + HInt, |
597 | <I as HInt>::D: ops::Shr<u32, Output = <I as HInt>::D>, |
598 | { |
599 | // `corr = 2 - b*x_n` |
600 | // |
601 | // This looks like `0 - b*x_n`. However, this works - in `UQ1`, `0.0 - x = 2.0 - x`. |
602 | let corr_uq1: I = I::ZERO.wrapping_sub(x_uq0.widen_mul(b_uq1).hi()); |
603 | |
604 | // `x_n * corr = x_n * (2 - b*x_n)` |
605 | (x_uq0.widen_mul(corr_uq1) >> (I::BITS - 1)).lo() |
606 | } |
607 | |
608 | intrinsics! { |
609 | #[avr_skip] |
610 | #[arm_aeabi_alias = __aeabi_fdiv] |
611 | pub extern "C" fn __divsf3(a: f32, b: f32) -> f32 { |
612 | div(a, b) |
613 | } |
614 | |
615 | #[avr_skip] |
616 | #[arm_aeabi_alias = __aeabi_ddiv] |
617 | pub extern "C" fn __divdf3(a: f64, b: f64) -> f64 { |
618 | div(a, b) |
619 | } |
620 | |
621 | #[avr_skip] |
622 | #[ppc_alias = __divkf3] |
623 | #[cfg (f128_enabled)] |
624 | pub extern "C" fn __divtf3(a: f128, b: f128) -> f128 { |
625 | div(a, b) |
626 | } |
627 | |
628 | #[cfg (target_arch = "arm" )] |
629 | pub extern "C" fn __divsf3vfp(a: f32, b: f32) -> f32 { |
630 | a / b |
631 | } |
632 | |
633 | #[cfg (target_arch = "arm" )] |
634 | pub extern "C" fn __divdf3vfp(a: f64, b: f64) -> f64 { |
635 | a / b |
636 | } |
637 | } |
638 | |