/* * // Copyright (c) Radzivon Bartoshyk 4/2025. All rights reserved. * // * // Redistribution and use in source and binary forms, with or without modification, * // are permitted provided that the following conditions are met: * // * // 1. Redistributions of source code must retain the above copyright notice, this * // list of conditions and the following disclaimer. * // * // 2. Redistributions in binary form must reproduce the above copyright notice, * // this list of conditions and the following disclaimer in the documentation * // and/or other materials provided with the distribution. * // * // 3. Neither the name of the copyright holder nor the names of its * // contributors may be used to endorse or promote products derived from * // this software without specific prior written permission. * // * // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * // DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE * // FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * // DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * // SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER * // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * // OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #![allow(clippy::too_many_arguments)] use crate::common::{dd_fmla, dyad_fmla, f_fmla, f_fmlaf, fmlaf, pow2if, rintfk}; use crate::double_double::DoubleDouble; use crate::rounding::{CpuFloor, CpuRound}; const L2U_F: f32 = 0.693_145_751_953_125; const L2L_F: f32 = 1.428_606_765_330_187_045_e-6; const R_LN2_F: f32 = std::f32::consts::LOG2_E; /// Exp for given value for const context. /// This is simplified version just to make a good approximation on const context. #[inline] pub const fn expf(d: f32) -> f32 { const EXP_POLY_1_S: f32 = 2f32; const EXP_POLY_2_S: f32 = 0.16666707f32; const EXP_POLY_3_S: f32 = -0.002775669f32; let qf = rintfk(d * R_LN2_F); let q = qf as i32; let r = fmlaf(qf, -L2U_F, d); let r = fmlaf(qf, -L2L_F, r); let f = r * r; // Poly for u = r*(exp(r)+1)/(exp(r)-1) let mut u = EXP_POLY_3_S; u = fmlaf(u, f, EXP_POLY_2_S); u = fmlaf(u, f, EXP_POLY_1_S); let u = 1f32 + 2f32 * r / (u - r); let i2 = pow2if(q); u * i2 // if d < -87f32 { // r = 0f32; // } // if d > 88f32 { // r = f32::INFINITY; // } } // Lookup table for exp(m) with m = -104, ..., 102. // -104 = floor(log(single precision's min denormal)) // 103 = ceil(log(single precision's max bessel K(n) that will be used)) // Table is generated with SageMath as follows: // for r in range(-104, 103): // print(double_to_hex(RealField(180)(r).exp()) + ",") pub(crate) static EXP_M1: [u64; 207] = [ 0x368f1e6b68529e33, 0x36a525be4e4e601d, 0x36bcbe0a45f75eb1, 0x36d3884e838aea68, 0x36ea8c1f14e2af5d, 0x37020a717e64a9bd, 0x3718851d84118908, 0x3730a9bdfb02d240, 0x3746a5bea046b42e, 0x375ec7f3b269efa8, 0x3774eafb87eab0f2, 0x378c6e2d05bbc000, 0x37a35208867c2683, 0x37ba425b317eeacd, 0x37d1d8508fa8246a, 0x37e840fbc08fdc8a, 0x38007b7112bc1ffe, 0x381666d0dad2961d, 0x382e726c3f64d0fe, 0x3844b0dc07cabf98, 0x385c1f2daf3b6a46, 0x38731c5957a47de2, 0x3889f96445648b9f, 0x38a1a6baeadb4fd1, 0x38b7fd974d372e45, 0x38d04da4d1452919, 0x38e62891f06b3450, 0x38fe1dd273aa8a4a, 0x3914775e0840bfdd, 0x392bd109d9d94bda, 0x3942e73f53fba844, 0x3959b138170d6bfe, 0x397175af0cf60ec5, 0x3987baee1bffa80b, 0x39a02057d1245ceb, 0x39b5eafffb34ba31, 0x39cdca23bae16424, 0x39e43e7fc88b8056, 0x39fb83bf23a9a9eb, 0x3a12b2b8dd05b318, 0x3a2969d47321e4cc, 0x3a41452b7723aed2, 0x3a5778fe2497184c, 0x3a6fe7116182e9cc, 0x3a85ae191a99585a, 0x3a9d775d87da854d, 0x3ab4063f8cc8bb98, 0x3acb374b315f87c1, 0x3ae27ec458c65e3c, 0x3af923372c67a074, 0x3b11152eaeb73c08, 0x3b2737c5645114b5, 0x3b3f8e6c24b5592e, 0x3b5571db733a9d61, 0x3b6d257d547e083f, 0x3b83ce9b9de78f85, 0x3b9aebabae3a41b5, 0x3bb24b6031b49bda, 0x3bc8dd5e1bb09d7e, 0x3be0e5b73d1ff53d, 0x3bf6f741de1748ec, 0x3c0f36bd37f42f3e, 0x3c2536452ee2f75c, 0x3c3cd480a1b74820, 0x3c539792499b1a24, 0x3c6aa0de4bf35b38, 0x3c82188ad6ae3303, 0x3c9898471fca6055, 0x3cb0b6c3afdde064, 0x3cc6b7719a59f0e0, 0x3cdee001eed62aa0, 0x3cf4fb547c775da8, 0x3d0c8464f7616468, 0x3d236121e24d3bba, 0x3d3a56e0c2ac7f75, 0x3d51e642baeb84a0, 0x3d6853f01d6d53ba, 0x3d80885298767e9a, 0x3d967852a7007e42, 0x3dae8a37a45fc32e, 0x3dc4c1078fe9228a, 0x3ddc3527e433fab1, 0x3df32b48bf117da2, 0x3e0a0db0d0ddb3ec, 0x3e21b48655f37267, 0x3e381056ff2c5772, 0x3e505a628c699fa1, 0x3e6639e3175a689d, 0x3e7e355bbaee85cb, 0x3e94875ca227ec38, 0x3eabe6c6fdb01612, 0x3ec2f6053b981d98, 0x3ed9c54c3b43bc8b, 0x3ef18354238f6764, 0x3f07cd79b5647c9b, 0x3f202cf22526545a, 0x3f35fc21041027ad, 0x3f4de16b9c24a98f, 0x3f644e51f113d4d6, 0x3f7b993fe00d5376, 0x3f92c155b8213cf4, 0x3fa97db0ccceb0af, 0x3fc152aaa3bf81cc, 0x3fd78b56362cef38, 0x3ff0000000000000, 0x4005bf0a8b145769, 0x401d8e64b8d4ddae, 0x403415e5bf6fb106, 0x404b4c902e273a58, 0x40628d389970338f, 0x407936dc5690c08f, 0x409122885aaeddaa, 0x40a749ea7d470c6e, 0x40bfa7157c470f82, 0x40d5829dcf950560, 0x40ed3c4488ee4f7f, 0x4103de1654d37c9a, 0x411b00b5916ac955, 0x413259ac48bf05d7, 0x4148f0ccafad2a87, 0x4160f2ebd0a80020, 0x417709348c0ea4f9, 0x418f4f22091940bd, 0x41a546d8f9ed26e1, 0x41bceb088b68e804, 0x41d3a6e1fd9eecfd, 0x41eab5adb9c43600, 0x420226af33b1fdc1, 0x4218ab7fb5475fb7, 0x4230c3d3920962c9, 0x4246c932696a6b5d, 0x425ef822f7f6731d, 0x42750bba3796379a, 0x428c9aae4631c056, 0x42a370470aec28ed, 0x42ba6b765d8cdf6d, 0x42d1f43fcc4b662c, 0x42e866f34a725782, 0x4300953e2f3a1ef7, 0x431689e221bc8d5b, 0x432ea215a1d20d76, 0x4344d13fbb1a001a, 0x435c4b334617cc67, 0x43733a43d282a519, 0x438a220d397972eb, 0x43a1c25c88df6862, 0x43b8232558201159, 0x43d0672a3c9eb871, 0x43e64b41c6d37832, 0x43fe4cf766fe49be, 0x44149767bc0483e3, 0x442bfc951eb8bb76, 0x444304d6aeca254b, 0x4459d97010884251, 0x44719103e4080b45, 0x4487e013cd114461, 0x44a03996528e074c, 0x44b60d4f6fdac731, 0x44cdf8c5af17ba3b, 0x44e45e3076d61699, 0x44fbaed16a6e0da7, 0x4512cffdfebde1a1, 0x4529919cabefcb69, 0x454160345c9953e3, 0x45579dbc9dc53c66, 0x45700c810d464097, 0x4585d009394c5c27, 0x459da57de8f107a8, 0x45b425982cf597cd, 0x45cb61e5ca3a5e31, 0x45e29bb825dfcf87, 0x45f94a90db0d6fe2, 0x46112fec759586fd, 0x46275c1dc469e3af, 0x463fbfd219c43b04, 0x4655936d44e1a146, 0x466d531d8a7ee79c, 0x4683ed9d24a2d51b, 0x469b15cfe5b6e17b, 0x46b268038c2c0e00, 0x46c9044a73545d48, 0x46e1002ab6218b38, 0x46f71b3540cbf921, 0x470f6799ea9c414a, 0x47255779b984f3eb, 0x473d01a210c44aa4, 0x4753b63da8e91210, 0x476aca8d6b0116b8, 0x478234de9e0c74e9, 0x4798bec7503ca477, 0x47b0d0eda9796b90, 0x47c6db0118477245, 0x47df1056dc7bf22d, 0x47f51c2cc3433801, 0x480cb108ffbec164, 0x48237f780991b584, 0x483a801c0ea8ac4d, 0x48520247cc4c46c1, 0x48687a0553328015, 0x4880a233dee4f9bb, 0x48969b7f55b808ba, 0x48aeba064644060a, 0x48c4e184933d9364, 0x48dc614fe2531841, 0x48f3494a9b171bf5, 0x490a36798b9d969b, 0x4921d03d8c0c04af, ]; // Lookup table for exp(m * 2^(-7)) with m = 0, ..., 127. // Table is generated with Sollya as follows: // > display = hexadecimal; // > for i from 0 to 127 do { D(exp(i / 128)); }; pub(crate) static EXP_M2: [u64; 128] = [ 0x3ff0000000000000, 0x3ff0202015600446, 0x3ff04080ab55de39, 0x3ff06122436410dd, 0x3ff08205601127ed, 0x3ff0a32a84e9c1f6, 0x3ff0c49236829e8c, 0x3ff0e63cfa7ab09d, 0x3ff1082b577d34ed, 0x3ff12a5dd543ccc5, 0x3ff14cd4fc989cd6, 0x3ff16f9157587069, 0x3ff192937074e0cd, 0x3ff1b5dbd3f68122, 0x3ff1d96b0eff0e79, 0x3ff1fd41afcba45e, 0x3ff2216045b6f5cd, 0x3ff245c7613b8a9b, 0x3ff26a7793f60164, 0x3ff28f7170a755fd, 0x3ff2b4b58b372c79, 0x3ff2da4478b620c7, 0x3ff3001ecf601af7, 0x3ff32645269ea829, 0x3ff34cb8170b5835, 0x3ff373783a722012, 0x3ff39a862bd3c106, 0x3ff3c1e2876834aa, 0x3ff3e98deaa11dcc, 0x3ff41188f42c3e32, 0x3ff439d443f5f159, 0x3ff462707b2bac21, 0x3ff48b5e3c3e8186, 0x3ff4b49e2ae5ac67, 0x3ff4de30ec211e60, 0x3ff50817263c13cd, 0x3ff5325180cfacf7, 0x3ff55ce0a4c58c7c, 0x3ff587c53c5a7af0, 0x3ff5b2fff3210fd9, 0x3ff5de9176045ff5, 0x3ff60a7a734ab0e8, 0x3ff636bb9a983258, 0x3ff663559cf1bc7c, 0x3ff690492cbf9433, 0x3ff6bd96fdd034a2, 0x3ff6eb3fc55b1e76, 0x3ff719443a03acb9, 0x3ff747a513dbef6a, 0x3ff776630c678bc1, 0x3ff7a57ede9ea23e, 0x3ff7d4f946f0ba8d, 0x3ff804d30347b546, 0x3ff8350cd30ac390, 0x3ff865a7772164c5, 0x3ff896a3b1f66a0e, 0x3ff8c802477b0010, 0x3ff8f9c3fd29beaf, 0x3ff92be99a09bf00, 0x3ff95e73e6b1b75e, 0x3ff99163ad4b1dcc, 0x3ff9c4b9b995509b, 0x3ff9f876d8e8c566, 0x3ffa2c9bda3a3e78, 0x3ffa61298e1e069c, 0x3ffa9620c6cb3374, 0x3ffacb82581eee54, 0x3ffb014f179fc3b8, 0x3ffb3787dc80f95f, 0x3ffb6e2d7fa5eb18, 0x3ffba540dba56e56, 0x3ffbdcc2cccd3c85, 0x3ffc14b431256446, 0x3ffc4d15e873c193, 0x3ffc85e8d43f7cd0, 0x3ffcbf2dd7d490f2, 0x3ffcf8e5d84758a9, 0x3ffd3311bc7822b4, 0x3ffd6db26d16cd67, 0x3ffda8c8d4a66969, 0x3ffde455df80e3c0, 0x3ffe205a7bdab73e, 0x3ffe5cd799c6a54e, 0x3ffe99ce2b397649, 0x3ffed73f240dc142, 0x3fff152b7a07bb76, 0x3fff539424d90f5e, 0x3fff927a1e24bb76, 0x3fffd1de6182f8c9, 0x400008e0f64294ab, 0x40002912df5ce72a, 0x400049856cd84339, 0x40006a39207f0a09, 0x40008b2e7d2035cf, 0x4000ac6606916501, 0x4000cde041b0e9ae, 0x4000ef9db467dcf8, 0x4001119ee5ac36b6, 0x400133e45d82e952, 0x4001566ea50201d7, 0x4001793e4652cc50, 0x40019c53ccb3fc6b, 0x4001bfafc47bda73, 0x4001e352bb1a74ad, 0x4002073d3f1bd518, 0x40022b6fe02a3b9c, 0x40024feb2f105cb8, 0x400274afbdbba4a6, 0x400299be1f3e7f1c, 0x4002bf16e7d2a38c, 0x4002e4baacdb6614, 0x40030aaa04e80d05, 0x400330e587b62b28, 0x4003576dce33fead, 0x40037e437282d4ee, 0x4003a5670ff972ed, 0x4003ccd9432682b4, 0x4003f49aa9d30590, 0x40041cabe304cb34, 0x4004450d8f00edd4, 0x40046dc04f4e5338, 0x400496c4c6b832da, 0x4004c01b9950a111, 0x4004e9c56c731f5d, 0x400513c2e6c731d7, 0x40053e14b042f9ca, 0x400568bb722dd593, 0x400593b7d72305bb, ]; pub(crate) trait ExpfBackend { fn fmaf(&self, x: f32, y: f32, z: f32) -> f32; fn fma(&self, x: f64, y: f64, z: f64) -> f64; fn dd_fma(&self, x: f64, y: f64, z: f64) -> f64; fn dyad_fma(&self, x: f64, y: f64, z: f64) -> f64; fn polyeval3(&self, x: f64, a0: f64, a1: f64, a2: f64) -> f64; fn polyeval5(&self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64, a4: f64) -> f64; fn polyeval6(&self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64, a4: f64, a5: f64) -> f64; fn polyeval7( &self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64, a4: f64, a5: f64, a6: f64, ) -> f64; fn roundf(&self, x: f32) -> f32; fn round(&self, x: f64) -> f64; fn floor(&self, x: f64) -> f64; fn round_ties_even(&self, x: f64) -> f64; fn quick_mult(&self, x: DoubleDouble, y: DoubleDouble) -> DoubleDouble; fn quick_mult_f64(&self, x: DoubleDouble, y: f64) -> DoubleDouble; fn quick_f64_mult(&self, x: f64, y: DoubleDouble) -> DoubleDouble; fn exact_mult(&self, x: f64, y: f64) -> DoubleDouble; } pub(crate) struct GenericExpfBackend {} impl ExpfBackend for GenericExpfBackend { #[inline(always)] fn fmaf(&self, x: f32, y: f32, z: f32) -> f32 { f_fmlaf(x, y, z) } #[inline(always)] fn fma(&self, x: f64, y: f64, z: f64) -> f64 { use crate::common::f_fmla; f_fmla(x, y, z) } #[inline(always)] fn dd_fma(&self, x: f64, y: f64, z: f64) -> f64 { dd_fmla(x, y, z) } #[inline(always)] fn dyad_fma(&self, x: f64, y: f64, z: f64) -> f64 { dyad_fmla(x, y, z) } #[inline(always)] fn polyeval3(&self, x: f64, a0: f64, a1: f64, a2: f64) -> f64 { use crate::polyeval::f_polyeval3; f_polyeval3(x, a0, a1, a2) } #[inline(always)] fn polyeval5(&self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64, a4: f64) -> f64 { use crate::polyeval::f_polyeval5; f_polyeval5(x, a0, a1, a2, a3, a4) } #[inline(always)] fn polyeval6(&self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64, a4: f64, a5: f64) -> f64 { use crate::polyeval::f_polyeval6; f_polyeval6(x, a0, a1, a2, a3, a4, a5) } #[inline(always)] fn polyeval7( &self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64, a4: f64, a5: f64, a6: f64, ) -> f64 { use crate::polyeval::f_polyeval7; f_polyeval7(x, a0, a1, a2, a3, a4, a5, a6) } #[inline(always)] fn roundf(&self, x: f32) -> f32 { x.cpu_round() } #[inline(always)] fn round(&self, x: f64) -> f64 { x.cpu_round() } #[inline(always)] fn floor(&self, x: f64) -> f64 { x.cpu_floor() } #[inline(always)] fn round_ties_even(&self, x: f64) -> f64 { use crate::rounding::CpuRoundTiesEven; x.cpu_round_ties_even() } #[inline(always)] fn quick_mult(&self, x: DoubleDouble, y: DoubleDouble) -> DoubleDouble { DoubleDouble::quick_mult(x, y) } #[inline(always)] fn quick_mult_f64(&self, x: DoubleDouble, y: f64) -> DoubleDouble { DoubleDouble::quick_mult_f64(x, y) } #[inline(always)] fn quick_f64_mult(&self, x: f64, y: DoubleDouble) -> DoubleDouble { DoubleDouble::quick_mult_f64(y, x) } #[inline(always)] fn exact_mult(&self, x: f64, y: f64) -> DoubleDouble { DoubleDouble::from_exact_mult(x, y) } } #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] pub(crate) struct FmaBackend {} #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] impl ExpfBackend for FmaBackend { #[inline(always)] fn fmaf(&self, x: f32, y: f32, z: f32) -> f32 { f32::mul_add(x, y, z) } #[inline(always)] fn fma(&self, x: f64, y: f64, z: f64) -> f64 { f64::mul_add(x, y, z) } #[inline(always)] fn dd_fma(&self, x: f64, y: f64, z: f64) -> f64 { f64::mul_add(x, y, z) } #[inline(always)] fn dyad_fma(&self, x: f64, y: f64, z: f64) -> f64 { f64::mul_add(x, y, z) } #[inline(always)] fn polyeval3(&self, x: f64, a0: f64, a1: f64, a2: f64) -> f64 { use crate::polyeval::d_polyeval3; d_polyeval3(x, a0, a1, a2) } #[inline(always)] fn polyeval5(&self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64, a4: f64) -> f64 { use crate::polyeval::d_polyeval5; d_polyeval5(x, a0, a1, a2, a3, a4) } #[inline(always)] fn polyeval6(&self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64, a4: f64, a5: f64) -> f64 { use crate::polyeval::d_polyeval6; d_polyeval6(x, a0, a1, a2, a3, a4, a5) } #[inline(always)] fn polyeval7( &self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64, a4: f64, a5: f64, a6: f64, ) -> f64 { use crate::polyeval::d_polyeval7; d_polyeval7(x, a0, a1, a2, a3, a4, a5, a6) } #[inline(always)] fn roundf(&self, x: f32) -> f32 { x.round() } #[inline(always)] fn round(&self, x: f64) -> f64 { x.round() } #[inline(always)] fn floor(&self, x: f64) -> f64 { x.floor() } #[inline(always)] fn round_ties_even(&self, x: f64) -> f64 { x.round_ties_even() } #[inline(always)] fn quick_mult(&self, x: DoubleDouble, y: DoubleDouble) -> DoubleDouble { DoubleDouble::quick_mult_fma(x, y) } #[inline(always)] fn quick_mult_f64(&self, x: DoubleDouble, y: f64) -> DoubleDouble { DoubleDouble::quick_mult_f64_fma(x, y) } #[inline(always)] fn quick_f64_mult(&self, x: f64, y: DoubleDouble) -> DoubleDouble { DoubleDouble::quick_mult_f64_fma(y, x) } #[inline(always)] fn exact_mult(&self, x: f64, y: f64) -> DoubleDouble { DoubleDouble::from_exact_mult_fma(x, y) } } #[inline(always)] fn expf_gen_impl(x: f32, backend: B) -> f32 { let x_u = x.to_bits(); let x_abs = x_u & 0x7fff_ffffu32; if x_abs >= 0x42b2_0000u32 || x_abs <= 0x3280_0000u32 { let exp = ((x_u >> 23) & 0xFF) as i32; // |x| < 2^-25 if exp <= 101i32 { return 1.0 + x; } // When x < log(2^-150) or nan if x_u >= 0xc2cf_f1b5u32 { // exp(-Inf) = 0 if x.is_infinite() { return 0.0; } // exp(nan) = nan if x.is_nan() { return x; } return 0.0; } // x >= 89 or nan if x.is_sign_positive() && (x_u >= 0x42b2_0000) { // x is +inf or nan return x + f32::INFINITY; } } // For -104 < x < 89, to compute exp(x), we perform the following range // reduction: find hi, mid, lo such that: // x = hi + mid + lo, in which // hi is an integer, // mid * 2^7 is an integer // -2^(-8) <= lo < 2^-8. // In particular, // hi + mid = round(x * 2^7) * 2^(-7). // Then, // exp(x) = exp(hi + mid + lo) = exp(hi) * exp(mid) * exp(lo). // We store exp(hi) and exp(mid) in the lookup tables EXP_M1 and EXP_M2 // respectively. exp(lo) is computed using a degree-4 minimax polynomial // generated by Sollya. // x_hi = (hi + mid) * 2^7 = round(x * 2^7). let kf = backend.roundf(x * 128.); // Subtract (hi + mid) from x to get lo. let xd = backend.fmaf(kf, -0.0078125 /* - 1/128 */, x) as f64; let mut x_hi = unsafe { kf.to_int_unchecked::() }; // it's already not indeterminate. x_hi += 104 << 7; // hi = x_hi >> 7 let exp_hi = f64::from_bits(EXP_M1[(x_hi >> 7) as usize]); // mid * 2^7 = x_hi & 0x0000'007fU; let exp_mid = f64::from_bits(EXP_M2[(x_hi & 0x7f) as usize]); // Degree-4 minimax polynomial generated by Sollya with the following // commands: // d = [-2^-8, 2^-8]; // f_exp = expm1(x)/x; // Q = fpminimax(f_exp, 3, [|D...|], [-2^-8, 2^-8]); let p = backend.polyeval5( xd, 1., f64::from_bits(0x3feffffffffff777), f64::from_bits(0x3fe000000000071c), f64::from_bits(0x3fc555566668e5e7), f64::from_bits(0x3fa55555555ef243), ); (p * exp_hi * exp_mid) as f32 } #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] #[target_feature(enable = "avx", enable = "fma")] unsafe fn expf_fma_impl(x: f32) -> f32 { expf_gen_impl(x, FmaBackend {}) } /// Computes exp /// /// Max found ULP 0.5 #[inline] pub fn f_expf(x: f32) -> f32 { #[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))] { expf_gen_impl(x, GenericExpfBackend {}) } #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] { use std::sync::OnceLock; static EXECUTOR: OnceLock f32> = OnceLock::new(); let q = EXECUTOR.get_or_init(|| { if std::arch::is_x86_feature_detected!("avx") && std::arch::is_x86_feature_detected!("fma") { expf_fma_impl } else { fn def_expf(x: f32) -> f32 { expf_gen_impl(x, GenericExpfBackend {}) } def_expf } }); unsafe { q(x) } } } #[inline] pub(crate) fn core_expf(x: f32) -> f64 { // x_hi = (hi + mid) * 2^7 = round(x * 2^7). let kf = (x * 128.).cpu_round(); // Subtract (hi + mid) from x to get lo. let xd = f_fmlaf(kf, -0.0078125 /* - 1/128 */, x) as f64; let mut x_hi = unsafe { kf.to_int_unchecked::() }; // it's already not indeterminate. x_hi += 104 << 7; // hi = x_hi >> 7 let exp_hi = f64::from_bits(EXP_M1[(x_hi >> 7) as usize]); // mid * 2^7 = x_hi & 0x0000'007fU; let exp_mid = f64::from_bits(EXP_M2[(x_hi & 0x7f) as usize]); // Degree-4 minimax polynomial generated by Sollya with the following // commands: // d = [-2^-8, 2^-8]; // f_exp = expm1(x)/x; // Q = fpminimax(f_exp, 3, [|D...|], [-2^-8, 2^-8]); use crate::polyeval::f_polyeval5; let p = f_polyeval5( xd, 1., f64::from_bits(0x3feffffffffff777), f64::from_bits(0x3fe000000000071c), f64::from_bits(0x3fc555566668e5e7), f64::from_bits(0x3fa55555555ef243), ); p * exp_hi * exp_mid } #[inline] pub(crate) fn core_expdf(x: f64) -> f64 { // x_hi = (hi + mid) * 2^7 = round(x * 2^7). let kf = (x * 128.).cpu_round(); // Subtract (hi + mid) from x to get lo. let xd = f_fmla(kf, -0.0078125 /* - 1/128 */, x); let mut x_hi = unsafe { kf.to_int_unchecked::() }; // it's already not indeterminate. x_hi += 104 << 7; // hi = x_hi >> 7 let exp_hi = f64::from_bits(EXP_M1[(x_hi >> 7) as usize]); // mid * 2^7 = x_hi & 0x0000'007fU; let exp_mid = f64::from_bits(EXP_M2[(x_hi & 0x7f) as usize]); // Degree-4 minimax polynomial generated by Sollya with the following // commands: // d = [-2^-8, 2^-8]; // f_exp = expm1(x)/x; // Q = fpminimax(f_exp, 3, [|D...|], [-2^-8, 2^-8]); use crate::polyeval::f_polyeval5; let p = f_polyeval5( xd, 1., f64::from_bits(0x3feffffffffff777), f64::from_bits(0x3fe000000000071c), f64::from_bits(0x3fc555566668e5e7), f64::from_bits(0x3fa55555555ef243), ); p * exp_hi * exp_mid } #[cfg(test)] mod tests { use super::*; #[test] fn expf_test() { assert!( (expf(0f32) - 1f32).abs() < 1e-6, "Invalid result {}", expf(0f32) ); assert!( (expf(5f32) - 148.4131591025766f32).abs() < 1e-6, "Invalid result {}", expf(5f32) ); } #[test] fn f_expf_test() { assert_eq!(f_expf(-103.971596), 1e-45); assert!( (f_expf(0f32) - 1f32).abs() < 1e-6, "Invalid result {}", f_expf(0f32) ); assert!( (f_expf(5f32) - 148.4131591025766f32).abs() < 1e-6, "Invalid result {}", f_expf(5f32) ); assert_eq!(f_expf(f32::INFINITY), f32::INFINITY); assert_eq!(f_expf(f32::NEG_INFINITY), 0.); assert!(f_expf(f32::NAN).is_nan()); } }