1 | /* SPDX-License-Identifier: MIT */ |
2 | /* origin: musl src/math/fmaf.c. Ported to generic Rust algorithm in 2025, TG. */ |
3 | |
4 | use super::support::{FpResult, IntTy, Round, Status}; |
5 | use super::{CastFrom, CastInto, DFloat, Float, HFloat, MinInt}; |
6 | |
7 | // Placeholder so we can have `fmaf16` in the `Float` trait. |
8 | #[allow (unused)] |
9 | #[cfg (f16_enabled)] |
10 | #[cfg_attr (all(test, assert_no_panic), no_panic::no_panic)] |
11 | pub(crate) fn fmaf16(_x: f16, _y: f16, _z: f16) -> f16 { |
12 | unimplemented!() |
13 | } |
14 | |
15 | /// Floating multiply add (f32) |
16 | /// |
17 | /// Computes `(x*y)+z`, rounded as one ternary operation (i.e. calculated with infinite precision). |
18 | #[cfg_attr (all(test, assert_no_panic), no_panic::no_panic)] |
19 | pub fn fmaf(x: f32, y: f32, z: f32) -> f32 { |
20 | fma_wide_round(x, y, z, Round::Nearest).val |
21 | } |
22 | |
23 | /// Fma implementation when a hardware-backed larger float type is available. For `f32` and `f64`, |
24 | /// `f64` has enough precision to represent the `f32` in its entirety, except for double rounding. |
25 | pub fn fma_wide_round<F, B>(x: F, y: F, z: F, round: Round) -> FpResult<F> |
26 | where |
27 | F: Float + HFloat<D = B>, |
28 | B: Float + DFloat<H = F>, |
29 | B::Int: CastInto<i32>, |
30 | i32: CastFrom<i32>, |
31 | { |
32 | let one = IntTy::<B>::ONE; |
33 | |
34 | let xy: B = x.widen() * y.widen(); |
35 | let mut result: B = xy + z.widen(); |
36 | let mut ui: B::Int = result.to_bits(); |
37 | let re = result.ex(); |
38 | let zb: B = z.widen(); |
39 | |
40 | let prec_diff = B::SIG_BITS - F::SIG_BITS; |
41 | let excess_prec = ui & ((one << prec_diff) - one); |
42 | let halfway = one << (prec_diff - 1); |
43 | |
44 | // Common case: the larger precision is fine if... |
45 | // This is not a halfway case |
46 | if excess_prec != halfway |
47 | // Or the result is NaN |
48 | || re == B::EXP_SAT |
49 | // Or the result is exact |
50 | || (result - xy == zb && result - zb == xy) |
51 | // Or the mode is something other than round to nearest |
52 | || round != Round::Nearest |
53 | { |
54 | let min_inexact_exp = (B::EXP_BIAS as i32 + F::EXP_MIN_SUBNORM) as u32; |
55 | let max_inexact_exp = (B::EXP_BIAS as i32 + F::EXP_MIN) as u32; |
56 | |
57 | let mut status = Status::OK; |
58 | |
59 | if (min_inexact_exp..max_inexact_exp).contains(&re) && status.inexact() { |
60 | // This branch is never hit; requires previous operations to set a status |
61 | status.set_inexact(false); |
62 | |
63 | result = xy + z.widen(); |
64 | if status.inexact() { |
65 | status.set_underflow(true); |
66 | } else { |
67 | status.set_inexact(true); |
68 | } |
69 | } |
70 | |
71 | return FpResult { val: result.narrow(), status }; |
72 | } |
73 | |
74 | let neg = ui >> (B::BITS - 1) != IntTy::<B>::ZERO; |
75 | let err = if neg == (zb > xy) { xy - result + zb } else { zb - result + xy }; |
76 | if neg == (err < B::ZERO) { |
77 | ui += one; |
78 | } else { |
79 | ui -= one; |
80 | } |
81 | |
82 | FpResult::ok(B::from_bits(ui).narrow()) |
83 | } |
84 | |
85 | #[cfg (test)] |
86 | mod tests { |
87 | use super::*; |
88 | |
89 | #[test ] |
90 | fn issue_263() { |
91 | let a = f32::from_bits(1266679807); |
92 | let b = f32::from_bits(1300234242); |
93 | let c = f32::from_bits(1115553792); |
94 | let expected = f32::from_bits(1501560833); |
95 | assert_eq!(fmaf(a, b, c), expected); |
96 | } |
97 | } |
98 | |