1/* 2 * Copyright (c) 2014,2015 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a copy 5 * of this software and associated documentation files (the "Software"), to deal 6 * in the Software without restriction, including without limitation the rights 7 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 8 * copies of the Software, and to permit persons to whom the Software is 9 * furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 17 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 19 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 20 * THE SOFTWARE. 21 */ 22 23#include <clc/clc.h> 24 25#include "math.h" 26#include "ep_log.h" 27#include "../clcmacro.h" 28 29_CLC_OVERLOAD _CLC_DEF float asinh(float x) { 30 uint ux = as_uint(x); 31 uint ax = ux & EXSIGNBIT_SP32; 32 uint xsgn = ax ^ ux; 33 34 // |x| <= 2 35 float t = x * x; 36 float a = mad(t, 37 mad(t, 38 mad(t, 39 mad(t, -1.177198915954942694e-4f, -4.162727710583425360e-2f), 40 -5.063201055468483248e-1f), 41 -1.480204186473758321f), 42 -1.152965835871758072f); 43 float b = mad(t, 44 mad(t, 45 mad(t, 46 mad(t, 6.284381367285534560e-2f, 1.260024978680227945f), 47 6.582362487198468066f), 48 11.99423176003939087f), 49 6.917795026025976739f); 50 51 float q = MATH_DIVIDE(a, b); 52 float z1 = mad(x*t, q, x); 53 54 // |x| > 2 55 56 // Arguments greater than 1/sqrt(epsilon) in magnitude are 57 // approximated by asinh(x) = ln(2) + ln(abs(x)), with sign of x 58 // Arguments such that 4.0 <= abs(x) <= 1/sqrt(epsilon) are 59 // approximated by asinhf(x) = ln(abs(x) + sqrt(x*x+1)) 60 // with the sign of x (see Abramowitz and Stegun 4.6.20) 61 62 float absx = as_float(ax); 63 int hi = ax > 0x46000000U; 64 float y = MATH_SQRT(absx * absx + 1.0f) + absx; 65 y = hi ? absx : y; 66 float r = log(y) + (hi ? 0x1.62e430p-1f : 0.0f); 67 float z2 = as_float(xsgn | as_uint(r)); 68 69 float z = ax <= 0x40000000 ? z1 : z2; 70 z = ax < 0x39800000U | ax >= PINFBITPATT_SP32 ? x : z; 71 72 return z; 73} 74 75_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, float, asinh, float) 76 77#ifdef cl_khr_fp64 78#pragma OPENCL EXTENSION cl_khr_fp64 : enable 79 80#define NA0 -0.12845379283524906084997e0 81#define NA1 -0.21060688498409799700819e0 82#define NA2 -0.10188951822578188309186e0 83#define NA3 -0.13891765817243625541799e-1 84#define NA4 -0.10324604871728082428024e-3 85 86#define DA0 0.77072275701149440164511e0 87#define DA1 0.16104665505597338100747e1 88#define DA2 0.11296034614816689554875e1 89#define DA3 0.30079351943799465092429e0 90#define DA4 0.235224464765951442265117e-1 91 92#define NB0 -0.12186605129448852495563e0 93#define NB1 -0.19777978436593069928318e0 94#define NB2 -0.94379072395062374824320e-1 95#define NB3 -0.12620141363821680162036e-1 96#define NB4 -0.903396794842691998748349e-4 97 98#define DB0 0.73119630776696495279434e0 99#define DB1 0.15157170446881616648338e1 100#define DB2 0.10524909506981282725413e1 101#define DB3 0.27663713103600182193817e0 102#define DB4 0.21263492900663656707646e-1 103 104#define NC0 -0.81210026327726247622500e-1 105#define NC1 -0.12327355080668808750232e0 106#define NC2 -0.53704925162784720405664e-1 107#define NC3 -0.63106739048128554465450e-2 108#define NC4 -0.35326896180771371053534e-4 109 110#define DC0 0.48726015805581794231182e0 111#define DC1 0.95890837357081041150936e0 112#define DC2 0.62322223426940387752480e0 113#define DC3 0.15028684818508081155141e0 114#define DC4 0.10302171620320141529445e-1 115 116#define ND0 -0.4638179204422665073e-1 117#define ND1 -0.7162729496035415183e-1 118#define ND2 -0.3247795155696775148e-1 119#define ND3 -0.4225785421291932164e-2 120#define ND4 -0.3808984717603160127e-4 121#define ND5 0.8023464184964125826e-6 122 123#define DD0 0.2782907534642231184e0 124#define DD1 0.5549945896829343308e0 125#define DD2 0.3700732511330698879e0 126#define DD3 0.9395783438240780722e-1 127#define DD4 0.7200057974217143034e-2 128 129#define NE0 -0.121224194072430701e-4 130#define NE1 -0.273145455834305218e-3 131#define NE2 -0.152866982560895737e-2 132#define NE3 -0.292231744584913045e-2 133#define NE4 -0.174670900236060220e-2 134#define NE5 -0.891754209521081538e-12 135 136#define DE0 0.499426632161317606e-4 137#define DE1 0.139591210395547054e-2 138#define DE2 0.107665231109108629e-1 139#define DE3 0.325809818749873406e-1 140#define DE4 0.415222526655158363e-1 141#define DE5 0.186315628774716763e-1 142 143#define NF0 -0.195436610112717345e-4 144#define NF1 -0.233315515113382977e-3 145#define NF2 -0.645380957611087587e-3 146#define NF3 -0.478948863920281252e-3 147#define NF4 -0.805234112224091742e-12 148#define NF5 0.246428598194879283e-13 149 150#define DF0 0.822166621698664729e-4 151#define DF1 0.135346265620413852e-2 152#define DF2 0.602739242861830658e-2 153#define DF3 0.972227795510722956e-2 154#define DF4 0.510878800983771167e-2 155 156#define NG0 -0.209689451648100728e-6 157#define NG1 -0.219252358028695992e-5 158#define NG2 -0.551641756327550939e-5 159#define NG3 -0.382300259826830258e-5 160#define NG4 -0.421182121910667329e-17 161#define NG5 0.492236019998237684e-19 162 163#define DG0 0.889178444424237735e-6 164#define DG1 0.131152171690011152e-4 165#define DG2 0.537955850185616847e-4 166#define DG3 0.814966175170941864e-4 167#define DG4 0.407786943832260752e-4 168 169#define NH0 -0.178284193496441400e-6 170#define NH1 -0.928734186616614974e-6 171#define NH2 -0.923318925566302615e-6 172#define NH3 -0.776417026702577552e-19 173#define NH4 0.290845644810826014e-21 174 175#define DH0 0.786694697277890964e-6 176#define DH1 0.685435665630965488e-5 177#define DH2 0.153780175436788329e-4 178#define DH3 0.984873520613417917e-5 179 180#define NI0 -0.538003743384069117e-10 181#define NI1 -0.273698654196756169e-9 182#define NI2 -0.268129826956403568e-9 183#define NI3 -0.804163374628432850e-29 184 185#define DI0 0.238083376363471960e-9 186#define DI1 0.203579344621125934e-8 187#define DI2 0.450836980450693209e-8 188#define DI3 0.286005148753497156e-8 189 190_CLC_OVERLOAD _CLC_DEF double asinh(double x) { 191 const double rteps = 0x1.6a09e667f3bcdp-27; 192 const double recrteps = 0x1.6a09e667f3bcdp+26; 193 194 // log2_lead and log2_tail sum to an extra-precise version of log(2) 195 const double log2_lead = 0x1.62e42ep-1; 196 const double log2_tail = 0x1.efa39ef35793cp-25; 197 198 ulong ux = as_ulong(x); 199 ulong ax = ux & ~SIGNBIT_DP64; 200 double absx = as_double(ax); 201 202 double t = x * x; 203 double pn, tn, pd, td; 204 205 // XXX we are betting here that we can evaluate 8 pairs of 206 // polys faster than we can grab 12 coefficients from a table 207 // This also uses fewer registers 208 209 // |x| >= 8 210 pn = fma(t, fma(t, fma(t, NI3, NI2), NI1), NI0); 211 pd = fma(t, fma(t, fma(t, DI3, DI2), DI1), DI0); 212 213 tn = fma(t, fma(t, fma(t, fma(t, NH4, NH3), NH2), NH1), NH0); 214 td = fma(t, fma(t, fma(t, DH3, DH2), DH1), DH0); 215 pn = absx < 8.0 ? tn : pn; 216 pd = absx < 8.0 ? td : pd; 217 218 tn = fma(t, fma(t, fma(t, fma(t, fma(t, NG5, NG4), NG3), NG2), NG1), NG0); 219 td = fma(t, fma(t, fma(t, fma(t, DG4, DG3), DG2), DG1), DG0); 220 pn = absx < 4.0 ? tn : pn; 221 pd = absx < 4.0 ? td : pd; 222 223 tn = fma(t, fma(t, fma(t, fma(t, fma(t, NF5, NF4), NF3), NF2), NF1), NF0); 224 td = fma(t, fma(t, fma(t, fma(t, DF4, DF3), DF2), DF1), DF0); 225 pn = absx < 2.0 ? tn : pn; 226 pd = absx < 2.0 ? td : pd; 227 228 tn = fma(t, fma(t, fma(t, fma(t, fma(t, NE5, NE4), NE3), NE2), NE1), NE0); 229 td = fma(t, fma(t, fma(t, fma(t, fma(t, DE5, DE4), DE3), DE2), DE1), DE0); 230 pn = absx < 1.5 ? tn : pn; 231 pd = absx < 1.5 ? td : pd; 232 233 tn = fma(t, fma(t, fma(t, fma(t, fma(t, ND5, ND4), ND3), ND2), ND1), ND0); 234 td = fma(t, fma(t, fma(t, fma(t, DD4, DD3), DD2), DD1), DD0); 235 pn = absx <= 1.0 ? tn : pn; 236 pd = absx <= 1.0 ? td : pd; 237 238 tn = fma(t, fma(t, fma(t, fma(t, NC4, NC3), NC2), NC1), NC0); 239 td = fma(t, fma(t, fma(t, fma(t, DC4, DC3), DC2), DC1), DC0); 240 pn = absx < 0.75 ? tn : pn; 241 pd = absx < 0.75 ? td : pd; 242 243 tn = fma(t, fma(t, fma(t, fma(t, NB4, NB3), NB2), NB1), NB0); 244 td = fma(t, fma(t, fma(t, fma(t, DB4, DB3), DB2), DB1), DB0); 245 pn = absx < 0.5 ? tn : pn; 246 pd = absx < 0.5 ? td : pd; 247 248 tn = fma(t, fma(t, fma(t, fma(t, NA4, NA3), NA2), NA1), NA0); 249 td = fma(t, fma(t, fma(t, fma(t, DA4, DA3), DA2), DA1), DA0); 250 pn = absx < 0.25 ? tn : pn; 251 pd = absx < 0.25 ? td : pd; 252 253 double pq = MATH_DIVIDE(pn, pd); 254 255 // |x| <= 1 256 double result1 = fma(absx*t, pq, absx); 257 258 // Other ranges 259 int xout = absx <= 32.0 | absx > recrteps; 260 double y = absx + sqrt(fma(absx, absx, 1.0)); 261 y = xout ? absx : y; 262 263 double r1, r2; 264 int xexp; 265 __clc_ep_log(y, &xexp, &r1, &r2); 266 267 double dxexp = (double)(xexp + xout); 268 r1 = fma(dxexp, log2_lead, r1); 269 r2 = fma(dxexp, log2_tail, r2); 270 271 // 1 < x <= 32 272 double v2 = (pq + 0.25) / t; 273 double r = v2 + r1; 274 double s = ((r1 - r) + v2) + r2; 275 double v1 = r + s; 276 v2 = (r - v1) + s; 277 double result2 = v1 + v2; 278 279 // x > 32 280 double result3 = r1 + r2; 281 282 double ret = absx > 1.0 ? result2 : result1; 283 ret = absx > 32.0 ? result3 : ret; 284 ret = x < 0.0 ? -ret : ret; 285 286 // NaN, +-Inf, or x small enough that asinh(x) = x 287 ret = ax >= PINFBITPATT_DP64 | absx < rteps ? x : ret; 288 return ret; 289} 290 291_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, double, asinh, double) 292 293#endif 294