Markos Chandras | e24c3be | 2015-08-13 09:56:31 +0200 | [diff] [blame] | 1 | /* |
| 2 | * IEEE754 floating point arithmetic |
| 3 | * double precision: MADDF.f (Fused Multiply Add) |
| 4 | * MADDF.fmt: FPR[fd] = FPR[fd] + (FPR[fs] x FPR[ft]) |
| 5 | * |
| 6 | * MIPS floating point support |
| 7 | * Copyright (C) 2015 Imagination Technologies, Ltd. |
| 8 | * Author: Markos Chandras <markos.chandras@imgtec.com> |
| 9 | * |
| 10 | * This program is free software; you can distribute it and/or modify it |
| 11 | * under the terms of the GNU General Public License as published by the |
| 12 | * Free Software Foundation; version 2 of the License. |
| 13 | */ |
| 14 | |
| 15 | #include "ieee754dp.h" |
| 16 | |
Paul Burton | d728f67 | 2016-04-21 14:04:50 +0100 | [diff] [blame] | 17 | enum maddf_flags { |
| 18 | maddf_negate_product = 1 << 0, |
| 19 | }; |
| 20 | |
| 21 | static union ieee754dp _dp_maddf(union ieee754dp z, union ieee754dp x, |
| 22 | union ieee754dp y, enum maddf_flags flags) |
Markos Chandras | e24c3be | 2015-08-13 09:56:31 +0200 | [diff] [blame] | 23 | { |
| 24 | int re; |
| 25 | int rs; |
| 26 | u64 rm; |
| 27 | unsigned lxm; |
| 28 | unsigned hxm; |
| 29 | unsigned lym; |
| 30 | unsigned hym; |
| 31 | u64 lrm; |
| 32 | u64 hrm; |
| 33 | u64 t; |
| 34 | u64 at; |
| 35 | int s; |
| 36 | |
| 37 | COMPXDP; |
| 38 | COMPYDP; |
Paul Burton | e2d11e1 | 2016-04-21 14:04:51 +0100 | [diff] [blame] | 39 | COMPZDP; |
Markos Chandras | e24c3be | 2015-08-13 09:56:31 +0200 | [diff] [blame] | 40 | |
| 41 | EXPLODEXDP; |
| 42 | EXPLODEYDP; |
Paul Burton | e2d11e1 | 2016-04-21 14:04:51 +0100 | [diff] [blame] | 43 | EXPLODEZDP; |
Markos Chandras | e24c3be | 2015-08-13 09:56:31 +0200 | [diff] [blame] | 44 | |
| 45 | FLUSHXDP; |
| 46 | FLUSHYDP; |
Paul Burton | e2d11e1 | 2016-04-21 14:04:51 +0100 | [diff] [blame] | 47 | FLUSHZDP; |
Markos Chandras | e24c3be | 2015-08-13 09:56:31 +0200 | [diff] [blame] | 48 | |
| 49 | ieee754_clearcx(); |
| 50 | |
| 51 | switch (zc) { |
| 52 | case IEEE754_CLASS_SNAN: |
| 53 | ieee754_setcx(IEEE754_INVALID_OPERATION); |
| 54 | return ieee754dp_nanxcpt(z); |
| 55 | case IEEE754_CLASS_DNORM: |
Paul Burton | e2d11e1 | 2016-04-21 14:04:51 +0100 | [diff] [blame] | 56 | DPDNORMZ; |
Aleksandar Markovic | ddbfff7 | 2017-06-19 17:50:12 +0200 | [diff] [blame] | 57 | /* QNAN and ZERO cases are handled separately below */ |
Markos Chandras | e24c3be | 2015-08-13 09:56:31 +0200 | [diff] [blame] | 58 | } |
| 59 | |
| 60 | switch (CLPAIR(xc, yc)) { |
| 61 | case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_SNAN): |
| 62 | case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_SNAN): |
| 63 | case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_SNAN): |
| 64 | case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_SNAN): |
| 65 | case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_SNAN): |
| 66 | return ieee754dp_nanxcpt(y); |
| 67 | |
| 68 | case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_SNAN): |
| 69 | case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_QNAN): |
| 70 | case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_ZERO): |
| 71 | case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_NORM): |
| 72 | case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_DNORM): |
| 73 | case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_INF): |
| 74 | return ieee754dp_nanxcpt(x); |
| 75 | |
| 76 | case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_QNAN): |
| 77 | case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_QNAN): |
| 78 | case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_QNAN): |
| 79 | case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_QNAN): |
| 80 | return y; |
| 81 | |
| 82 | case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_QNAN): |
| 83 | case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_ZERO): |
| 84 | case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_NORM): |
| 85 | case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_DNORM): |
| 86 | case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_INF): |
| 87 | return x; |
| 88 | |
| 89 | |
| 90 | /* |
| 91 | * Infinity handling |
| 92 | */ |
| 93 | case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_ZERO): |
| 94 | case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_INF): |
| 95 | if (zc == IEEE754_CLASS_QNAN) |
| 96 | return z; |
| 97 | ieee754_setcx(IEEE754_INVALID_OPERATION); |
| 98 | return ieee754dp_indef(); |
| 99 | |
| 100 | case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_INF): |
| 101 | case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_INF): |
| 102 | case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_NORM): |
| 103 | case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_DNORM): |
| 104 | case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_INF): |
| 105 | if (zc == IEEE754_CLASS_QNAN) |
| 106 | return z; |
| 107 | return ieee754dp_inf(xs ^ ys); |
| 108 | |
| 109 | case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_ZERO): |
| 110 | case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_NORM): |
| 111 | case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_DNORM): |
| 112 | case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_ZERO): |
| 113 | case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_ZERO): |
| 114 | if (zc == IEEE754_CLASS_INF) |
| 115 | return ieee754dp_inf(zs); |
| 116 | /* Multiplication is 0 so just return z */ |
| 117 | return z; |
| 118 | |
| 119 | case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_DNORM): |
| 120 | DPDNORMX; |
| 121 | |
| 122 | case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_DNORM): |
| 123 | if (zc == IEEE754_CLASS_QNAN) |
| 124 | return z; |
| 125 | else if (zc == IEEE754_CLASS_INF) |
| 126 | return ieee754dp_inf(zs); |
| 127 | DPDNORMY; |
| 128 | break; |
| 129 | |
| 130 | case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_NORM): |
| 131 | if (zc == IEEE754_CLASS_QNAN) |
| 132 | return z; |
| 133 | else if (zc == IEEE754_CLASS_INF) |
| 134 | return ieee754dp_inf(zs); |
| 135 | DPDNORMX; |
| 136 | break; |
| 137 | |
| 138 | case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_NORM): |
| 139 | if (zc == IEEE754_CLASS_QNAN) |
| 140 | return z; |
| 141 | else if (zc == IEEE754_CLASS_INF) |
| 142 | return ieee754dp_inf(zs); |
| 143 | /* fall through to real computations */ |
| 144 | } |
| 145 | |
| 146 | /* Finally get to do some computation */ |
| 147 | |
| 148 | /* |
| 149 | * Do the multiplication bit first |
| 150 | * |
| 151 | * rm = xm * ym, re = xe + ye basically |
| 152 | * |
| 153 | * At this point xm and ym should have been normalized. |
| 154 | */ |
| 155 | assert(xm & DP_HIDDEN_BIT); |
| 156 | assert(ym & DP_HIDDEN_BIT); |
| 157 | |
| 158 | re = xe + ye; |
| 159 | rs = xs ^ ys; |
Paul Burton | d728f67 | 2016-04-21 14:04:50 +0100 | [diff] [blame] | 160 | if (flags & maddf_negate_product) |
| 161 | rs ^= 1; |
Markos Chandras | e24c3be | 2015-08-13 09:56:31 +0200 | [diff] [blame] | 162 | |
| 163 | /* shunt to top of word */ |
| 164 | xm <<= 64 - (DP_FBITS + 1); |
| 165 | ym <<= 64 - (DP_FBITS + 1); |
| 166 | |
| 167 | /* |
Paul Burton | 95bff24 | 2016-04-21 14:04:52 +0100 | [diff] [blame] | 168 | * Multiply 64 bits xm, ym to give high 64 bits rm with stickness. |
Markos Chandras | e24c3be | 2015-08-13 09:56:31 +0200 | [diff] [blame] | 169 | */ |
| 170 | |
| 171 | /* 32 * 32 => 64 */ |
| 172 | #define DPXMULT(x, y) ((u64)(x) * (u64)y) |
| 173 | |
| 174 | lxm = xm; |
| 175 | hxm = xm >> 32; |
| 176 | lym = ym; |
| 177 | hym = ym >> 32; |
| 178 | |
| 179 | lrm = DPXMULT(lxm, lym); |
| 180 | hrm = DPXMULT(hxm, hym); |
| 181 | |
| 182 | t = DPXMULT(lxm, hym); |
| 183 | |
| 184 | at = lrm + (t << 32); |
| 185 | hrm += at < lrm; |
| 186 | lrm = at; |
| 187 | |
| 188 | hrm = hrm + (t >> 32); |
| 189 | |
| 190 | t = DPXMULT(hxm, lym); |
| 191 | |
| 192 | at = lrm + (t << 32); |
| 193 | hrm += at < lrm; |
| 194 | lrm = at; |
| 195 | |
| 196 | hrm = hrm + (t >> 32); |
| 197 | |
| 198 | rm = hrm | (lrm != 0); |
| 199 | |
| 200 | /* |
| 201 | * Sticky shift down to normal rounding precision. |
| 202 | */ |
| 203 | if ((s64) rm < 0) { |
| 204 | rm = (rm >> (64 - (DP_FBITS + 1 + 3))) | |
| 205 | ((rm << (DP_FBITS + 1 + 3)) != 0); |
Paul Burton | 5c18c936 | 2016-04-21 14:04:53 +0100 | [diff] [blame] | 206 | re++; |
Markos Chandras | e24c3be | 2015-08-13 09:56:31 +0200 | [diff] [blame] | 207 | } else { |
| 208 | rm = (rm >> (64 - (DP_FBITS + 1 + 3 + 1))) | |
| 209 | ((rm << (DP_FBITS + 1 + 3 + 1)) != 0); |
| 210 | } |
| 211 | assert(rm & (DP_HIDDEN_BIT << 3)); |
| 212 | |
Aleksandar Markovic | ddbfff7 | 2017-06-19 17:50:12 +0200 | [diff] [blame] | 213 | if (zc == IEEE754_CLASS_ZERO) |
| 214 | return ieee754dp_format(rs, re, rm); |
| 215 | |
Markos Chandras | e24c3be | 2015-08-13 09:56:31 +0200 | [diff] [blame] | 216 | /* And now the addition */ |
| 217 | assert(zm & DP_HIDDEN_BIT); |
| 218 | |
| 219 | /* |
| 220 | * Provide guard,round and stick bit space. |
| 221 | */ |
| 222 | zm <<= 3; |
| 223 | |
| 224 | if (ze > re) { |
| 225 | /* |
| 226 | * Have to shift y fraction right to align. |
| 227 | */ |
| 228 | s = ze - re; |
| 229 | rm = XDPSRS(rm, s); |
| 230 | re += s; |
| 231 | } else if (re > ze) { |
| 232 | /* |
| 233 | * Have to shift x fraction right to align. |
| 234 | */ |
| 235 | s = re - ze; |
| 236 | zm = XDPSRS(zm, s); |
| 237 | ze += s; |
| 238 | } |
| 239 | assert(ze == re); |
| 240 | assert(ze <= DP_EMAX); |
| 241 | |
| 242 | if (zs == rs) { |
| 243 | /* |
| 244 | * Generate 28 bit result of adding two 27 bit numbers |
| 245 | * leaving result in xm, xs and xe. |
| 246 | */ |
| 247 | zm = zm + rm; |
| 248 | |
| 249 | if (zm >> (DP_FBITS + 1 + 3)) { /* carry out */ |
| 250 | zm = XDPSRS1(zm); |
| 251 | ze++; |
| 252 | } |
| 253 | } else { |
| 254 | if (zm >= rm) { |
| 255 | zm = zm - rm; |
| 256 | } else { |
| 257 | zm = rm - zm; |
| 258 | zs = rs; |
| 259 | } |
| 260 | if (zm == 0) |
| 261 | return ieee754dp_zero(ieee754_csr.rm == FPU_CSR_RD); |
| 262 | |
| 263 | /* |
| 264 | * Normalize to rounding precision. |
| 265 | */ |
| 266 | while ((zm >> (DP_FBITS + 3)) == 0) { |
| 267 | zm <<= 1; |
| 268 | ze--; |
| 269 | } |
| 270 | } |
| 271 | |
| 272 | return ieee754dp_format(zs, ze, zm); |
| 273 | } |
Paul Burton | d728f67 | 2016-04-21 14:04:50 +0100 | [diff] [blame] | 274 | |
| 275 | union ieee754dp ieee754dp_maddf(union ieee754dp z, union ieee754dp x, |
| 276 | union ieee754dp y) |
| 277 | { |
| 278 | return _dp_maddf(z, x, y, 0); |
| 279 | } |
| 280 | |
| 281 | union ieee754dp ieee754dp_msubf(union ieee754dp z, union ieee754dp x, |
| 282 | union ieee754dp y) |
| 283 | { |
| 284 | return _dp_maddf(z, x, y, maddf_negate_product); |
| 285 | } |