| 1 | use crate::support::{ |
| 2 | CastFrom, CastInto, DFloat, Float, FpResult, HFloat, IntTy, MinInt, Round, Status, |
| 3 | }; |
| 4 | |
| 5 | /// Fma implementation when a hardware-backed larger float type is available. For `f32` and `f64`, |
| 6 | /// `f64` has enough precision to represent the `f32` in its entirety, except for double rounding. |
| 7 | #[inline ] |
| 8 | pub fn fma_wide_round<F, B>(x: F, y: F, z: F, round: Round) -> FpResult<F> |
| 9 | where |
| 10 | F: Float + HFloat<D = B>, |
| 11 | B: Float + DFloat<H = F>, |
| 12 | B::Int: CastInto<i32>, |
| 13 | i32: CastFrom<i32>, |
| 14 | { |
| 15 | let one = IntTy::<B>::ONE; |
| 16 | |
| 17 | let xy: B = x.widen() * y.widen(); |
| 18 | let mut result: B = xy + z.widen(); |
| 19 | let mut ui: B::Int = result.to_bits(); |
| 20 | let re = result.ex(); |
| 21 | let zb: B = z.widen(); |
| 22 | |
| 23 | let prec_diff = B::SIG_BITS - F::SIG_BITS; |
| 24 | let excess_prec = ui & ((one << prec_diff) - one); |
| 25 | let halfway = one << (prec_diff - 1); |
| 26 | |
| 27 | // Common case: the larger precision is fine if... |
| 28 | // This is not a halfway case |
| 29 | if excess_prec != halfway |
| 30 | // Or the result is NaN |
| 31 | || re == B::EXP_SAT |
| 32 | // Or the result is exact |
| 33 | || (result - xy == zb && result - zb == xy) |
| 34 | // Or the mode is something other than round to nearest |
| 35 | || round != Round::Nearest |
| 36 | { |
| 37 | let min_inexact_exp = (B::EXP_BIAS as i32 + F::EXP_MIN_SUBNORM) as u32; |
| 38 | let max_inexact_exp = (B::EXP_BIAS as i32 + F::EXP_MIN) as u32; |
| 39 | |
| 40 | let mut status = Status::OK; |
| 41 | |
| 42 | if (min_inexact_exp..max_inexact_exp).contains(&re) && status.inexact() { |
| 43 | // This branch is never hit; requires previous operations to set a status |
| 44 | status.set_inexact(false); |
| 45 | |
| 46 | result = xy + z.widen(); |
| 47 | if status.inexact() { |
| 48 | status.set_underflow(true); |
| 49 | } else { |
| 50 | status.set_inexact(true); |
| 51 | } |
| 52 | } |
| 53 | |
| 54 | return FpResult { |
| 55 | val: result.narrow(), |
| 56 | status, |
| 57 | }; |
| 58 | } |
| 59 | |
| 60 | let neg = ui >> (B::BITS - 1) != IntTy::<B>::ZERO; |
| 61 | let err = if neg == (zb > xy) { |
| 62 | xy - result + zb |
| 63 | } else { |
| 64 | zb - result + xy |
| 65 | }; |
| 66 | if neg == (err < B::ZERO) { |
| 67 | ui += one; |
| 68 | } else { |
| 69 | ui -= one; |
| 70 | } |
| 71 | |
| 72 | FpResult::ok(B::from_bits(ui).narrow()) |
| 73 | } |
| 74 | |