| 1 | //! Ported from musl, which is MIT licensed: |
| 2 | //! https://git.musl-libc.org/cgit/musl/tree/COPYRIGHT |
| 3 | //! |
| 4 | //! https://git.musl-libc.org/cgit/musl/tree/src/math/fmal.c |
| 5 | //! https://git.musl-libc.org/cgit/musl/tree/src/math/fmaf.c |
| 6 | //! https://git.musl-libc.org/cgit/musl/tree/src/math/fma.c |
| 7 | |
| 8 | const std = @import("std"); |
| 9 | const math = std.math; |
| 10 | const expect = std.testing.expect; |
| 11 | const compiler_rt = @import("../compiler_rt.zig"); |
| 12 | const symbol = compiler_rt.symbol; |
| 13 | |
| 14 | comptime { |
| 15 | symbol(&__fmah, "__fmah"); |
| 16 | symbol(&fmaf, "fmaf"); |
| 17 | symbol(&fma, "fma"); |
| 18 | symbol(&__fmax, "__fmax"); |
| 19 | symbol(&fmaq, "fmaf128"); |
| 20 | symbol(&fmal, "fmal"); |
| 21 | } |
| 22 | |
| 23 | fn __fmah(x: compiler_rt.f16.Abi, y: compiler_rt.f16.Abi, z: compiler_rt.f16.Abi) callconv(.c) compiler_rt.f16.Abi { |
| 24 | return compiler_rt.f16.toAbi(fma_f16(compiler_rt.f16.fromAbi(x), compiler_rt.f16.fromAbi(y), compiler_rt.f16.fromAbi(z))); |
| 25 | } |
| 26 | pub fn fma_f16(x: f16, y: f16, z: f16) f16 { |
| 27 | // TODO: more efficient implementation |
| 28 | return @floatCast(fma_f32(x, y, z)); |
| 29 | } |
| 30 | |
| 31 | fn fmaf(x: compiler_rt.f32.Abi, y: compiler_rt.f32.Abi, z: compiler_rt.f32.Abi) callconv(.c) compiler_rt.f32.Abi { |
| 32 | return compiler_rt.f32.toAbi(fma_f32(compiler_rt.f32.fromAbi(x), compiler_rt.f32.fromAbi(y), compiler_rt.f32.fromAbi(z))); |
| 33 | } |
| 34 | pub fn fma_f32(x: f32, y: f32, z: f32) f32 { |
| 35 | const xy = @as(f64, x) * y; |
| 36 | const xy_z = xy + z; |
| 37 | const u = @as(u64, @bitCast(xy_z)); |
| 38 | const e = (u >> 52) & 0x7FF; |
| 39 | |
| 40 | if ((u & 0x1FFFFFFF) != 0x10000000 or e == 0x7FF or (xy_z - xy == z and xy_z - z == xy)) { |
| 41 | return @floatCast(xy_z); |
| 42 | } else { |
| 43 | // TODO: Handle inexact case with double-rounding |
| 44 | return @floatCast(xy_z); |
| 45 | } |
| 46 | } |
| 47 | |
| 48 | fn fma(x: compiler_rt.f64.Abi, y: compiler_rt.f64.Abi, z: compiler_rt.f64.Abi) callconv(.c) compiler_rt.f64.Abi { |
| 49 | return compiler_rt.f64.toAbi(fma_f64(compiler_rt.f64.fromAbi(x), compiler_rt.f64.fromAbi(y), compiler_rt.f64.fromAbi(z))); |
| 50 | } |
| 51 | /// NOTE: Upstream fma.c has been rewritten completely to raise fp exceptions more accurately. |
| 52 | pub fn fma_f64(x: f64, y: f64, z: f64) f64 { |
| 53 | if (!math.isFinite(x) or !math.isFinite(y)) { |
| 54 | return x * y + z; |
| 55 | } |
| 56 | if (!math.isFinite(z)) { |
| 57 | return z; |
| 58 | } |
| 59 | if (x == 0.0 or y == 0.0) { |
| 60 | return x * y + z; |
| 61 | } |
| 62 | if (z == 0.0) { |
| 63 | return x * y; |
| 64 | } |
| 65 | |
| 66 | const x1 = math.frexp(x); |
| 67 | const ex = x1.exponent; |
| 68 | const xs = x1.significand; |
| 69 | const x2 = math.frexp(y); |
| 70 | const ey = x2.exponent; |
| 71 | const ys = x2.significand; |
| 72 | const x3 = math.frexp(z); |
| 73 | const ez = x3.exponent; |
| 74 | var zs = x3.significand; |
| 75 | |
| 76 | var spread = ex + ey - ez; |
| 77 | if (spread <= 53 * 2) { |
| 78 | zs = math.scalbn(zs, -spread); |
| 79 | } else { |
| 80 | zs = math.copysign(math.floatMin(f64), zs); |
| 81 | } |
| 82 | |
| 83 | const xy = dd_mul(xs, ys); |
| 84 | const r = dd_add(xy.hi, zs); |
| 85 | spread = ex + ey; |
| 86 | |
| 87 | if (r.hi == 0.0) { |
| 88 | return xy.hi + zs + math.scalbn(xy.lo, spread); |
| 89 | } |
| 90 | |
| 91 | const adj = add_adjusted(r.lo, xy.lo); |
| 92 | if (spread + math.ilogb(r.hi) > -1023) { |
| 93 | return math.scalbn(r.hi + adj, spread); |
| 94 | } else { |
| 95 | return add_and_denorm(r.hi, adj, spread); |
| 96 | } |
| 97 | } |
| 98 | |
| 99 | fn __fmax(a: compiler_rt.f80.Abi, b: compiler_rt.f80.Abi, c: compiler_rt.f80.Abi) callconv(.c) compiler_rt.f80.Abi { |
| 100 | return compiler_rt.f80.toAbi(fma_f80(compiler_rt.f80.fromAbi(a), compiler_rt.f80.fromAbi(b), compiler_rt.f80.fromAbi(c))); |
| 101 | } |
| 102 | pub fn fma_f80(a: f80, b: f80, c: f80) f80 { |
| 103 | // TODO: more efficient implementation |
| 104 | return @floatCast(fma_f128(a, b, c)); |
| 105 | } |
| 106 | |
| 107 | fn fmaq(x: compiler_rt.f128.Abi, y: compiler_rt.f128.Abi, z: compiler_rt.f128.Abi) callconv(.c) compiler_rt.f128.Abi { |
| 108 | return compiler_rt.f128.toAbi(fma_f128(compiler_rt.f128.fromAbi(x), compiler_rt.f128.fromAbi(y), compiler_rt.f128.fromAbi(z))); |
| 109 | } |
| 110 | /// Fused multiply-add: Compute x * y + z with a single rounding error. |
| 111 | /// |
| 112 | /// We use scaling to avoid overflow/underflow, along with the |
| 113 | /// canonical precision-doubling technique adapted from: |
| 114 | /// |
| 115 | /// Dekker, T. A Floating-Point Technique for Extending the |
| 116 | /// Available Precision. Numer. Math. 18, 224-242 (1971). |
| 117 | pub fn fma_f128(x: f128, y: f128, z: f128) f128 { |
| 118 | if (!math.isFinite(x) or !math.isFinite(y)) { |
| 119 | return x * y + z; |
| 120 | } |
| 121 | if (!math.isFinite(z)) { |
| 122 | return z; |
| 123 | } |
| 124 | if (x == 0.0 or y == 0.0) { |
| 125 | return x * y + z; |
| 126 | } |
| 127 | if (z == 0.0) { |
| 128 | return x * y; |
| 129 | } |
| 130 | |
| 131 | const x1 = math.frexp(x); |
| 132 | const ex = x1.exponent; |
| 133 | const xs = x1.significand; |
| 134 | const x2 = math.frexp(y); |
| 135 | const ey = x2.exponent; |
| 136 | const ys = x2.significand; |
| 137 | const x3 = math.frexp(z); |
| 138 | const ez = x3.exponent; |
| 139 | var zs = x3.significand; |
| 140 | |
| 141 | var spread = ex + ey - ez; |
| 142 | if (spread <= 113 * 2) { |
| 143 | zs = math.scalbn(zs, -spread); |
| 144 | } else { |
| 145 | zs = math.copysign(math.floatMin(f128), zs); |
| 146 | } |
| 147 | |
| 148 | const xy = dd_mul128(xs, ys); |
| 149 | const r = dd_add128(xy.hi, zs); |
| 150 | spread = ex + ey; |
| 151 | |
| 152 | if (r.hi == 0.0) { |
| 153 | return xy.hi + zs + math.scalbn(xy.lo, spread); |
| 154 | } |
| 155 | |
| 156 | const adj = add_adjusted128(r.lo, xy.lo); |
| 157 | if (spread + math.ilogb(r.hi) > -16383) { |
| 158 | return math.scalbn(r.hi + adj, spread); |
| 159 | } else { |
| 160 | return add_and_denorm128(r.hi, adj, spread); |
| 161 | } |
| 162 | } |
| 163 | |
| 164 | pub fn fmal(x: c_longdouble, y: c_longdouble, z: c_longdouble) callconv(.c) c_longdouble { |
| 165 | switch (@typeInfo(c_longdouble).float.bits) { |
| 166 | 64 => return fma_f64(x, y, z), |
| 167 | 80 => return fma_f80(x, y, z), |
| 168 | 128 => return fma_f128(x, y, z), |
| 169 | else => comptime unreachable, |
| 170 | } |
| 171 | } |
| 172 | |
| 173 | const dd = struct { |
| 174 | hi: f64, |
| 175 | lo: f64, |
| 176 | }; |
| 177 | |
| 178 | fn dd_add(a: f64, b: f64) dd { |
| 179 | var ret: dd = undefined; |
| 180 | ret.hi = a + b; |
| 181 | const s = ret.hi - a; |
| 182 | ret.lo = (a - (ret.hi - s)) + (b - s); |
| 183 | return ret; |
| 184 | } |
| 185 | |
| 186 | fn dd_mul(a: f64, b: f64) dd { |
| 187 | var ret: dd = undefined; |
| 188 | const split: f64 = 0x1.0p27 + 1.0; |
| 189 | |
| 190 | var p = a * split; |
| 191 | var ha = a - p; |
| 192 | ha += p; |
| 193 | const la = a - ha; |
| 194 | |
| 195 | p = b * split; |
| 196 | var hb = b - p; |
| 197 | hb += p; |
| 198 | const lb = b - hb; |
| 199 | |
| 200 | p = ha * hb; |
| 201 | const q = ha * lb + la * hb; |
| 202 | |
| 203 | ret.hi = p + q; |
| 204 | ret.lo = p - ret.hi + q + la * lb; |
| 205 | return ret; |
| 206 | } |
| 207 | |
| 208 | fn add_adjusted(a: f64, b: f64) f64 { |
| 209 | var sum = dd_add(a, b); |
| 210 | if (sum.lo != 0) { |
| 211 | var uhii: u64 = @bitCast(sum.hi); |
| 212 | if (uhii & 1 == 0) { |
| 213 | // hibits += copysign(1.0, sum.hi, sum.lo) |
| 214 | const uloi: u64 = @bitCast(sum.lo); |
| 215 | uhii = uhii + 1 - ((uhii ^ uloi) >> 62); |
| 216 | sum.hi = @bitCast(uhii); |
| 217 | } |
| 218 | } |
| 219 | return sum.hi; |
| 220 | } |
| 221 | |
| 222 | fn add_and_denorm(a: f64, b: f64, scale: i32) f64 { |
| 223 | var sum = dd_add(a, b); |
| 224 | if (sum.lo != 0) { |
| 225 | var uhii: u64 = @bitCast(sum.hi); |
| 226 | const bits_lost = -@as(i32, @intCast((uhii >> 52) & 0x7FF)) - scale + 1; |
| 227 | if ((bits_lost != 1) == (uhii & 1 != 0)) { |
| 228 | const uloi: u64 = @bitCast(sum.lo); |
| 229 | uhii = uhii + 1 - (((uhii ^ uloi) >> 62) & 2); |
| 230 | sum.hi = @bitCast(uhii); |
| 231 | } |
| 232 | } |
| 233 | return math.scalbn(sum.hi, scale); |
| 234 | } |
| 235 | |
| 236 | /// A struct that represents a floating-point number with twice the precision |
| 237 | /// of f128. We maintain the invariant that "hi" stores the high-order |
| 238 | /// bits of the result. |
| 239 | const dd128 = struct { |
| 240 | hi: f128, |
| 241 | lo: f128, |
| 242 | }; |
| 243 | |
| 244 | /// Compute a+b exactly, returning the exact result in a struct dd. We assume |
| 245 | /// that both a and b are finite, but make no assumptions about their relative |
| 246 | /// magnitudes. |
| 247 | fn dd_add128(a: f128, b: f128) dd128 { |
| 248 | var ret: dd128 = undefined; |
| 249 | ret.hi = a + b; |
| 250 | const s = ret.hi - a; |
| 251 | ret.lo = (a - (ret.hi - s)) + (b - s); |
| 252 | return ret; |
| 253 | } |
| 254 | |
| 255 | /// Compute a+b, with a small tweak: The least significant bit of the |
| 256 | /// result is adjusted into a sticky bit summarizing all the bits that |
| 257 | /// were lost to rounding. This adjustment negates the effects of double |
| 258 | /// rounding when the result is added to another number with a higher |
| 259 | /// exponent. For an explanation of round and sticky bits, see any reference |
| 260 | /// on FPU design, e.g., |
| 261 | /// |
| 262 | /// J. Coonen. An Implementation Guide to a Proposed Standard for |
| 263 | /// Floating-Point Arithmetic. Computer, vol. 13, no. 1, Jan 1980. |
| 264 | fn add_adjusted128(a: f128, b: f128) f128 { |
| 265 | var sum = dd_add128(a, b); |
| 266 | if (sum.lo != 0) { |
| 267 | var uhii: u128 = @bitCast(sum.hi); |
| 268 | if (uhii & 1 == 0) { |
| 269 | // hibits += copysign(1.0, sum.hi, sum.lo) |
| 270 | const uloi: u128 = @bitCast(sum.lo); |
| 271 | uhii = uhii + 1 - ((uhii ^ uloi) >> 126); |
| 272 | sum.hi = @bitCast(uhii); |
| 273 | } |
| 274 | } |
| 275 | return sum.hi; |
| 276 | } |
| 277 | |
| 278 | /// Compute ldexp(a+b, scale) with a single rounding error. It is assumed |
| 279 | /// that the result will be subnormal, and care is taken to ensure that |
| 280 | /// double rounding does not occur. |
| 281 | fn add_and_denorm128(a: f128, b: f128, scale: i32) f128 { |
| 282 | var sum = dd_add128(a, b); |
| 283 | // If we are losing at least two bits of accuracy to denormalization, |
| 284 | // then the first lost bit becomes a round bit, and we adjust the |
| 285 | // lowest bit of sum.hi to make it a sticky bit summarizing all the |
| 286 | // bits in sum.lo. With the sticky bit adjusted, the hardware will |
| 287 | // break any ties in the correct direction. |
| 288 | // |
| 289 | // If we are losing only one bit to denormalization, however, we must |
| 290 | // break the ties manually. |
| 291 | if (sum.lo != 0) { |
| 292 | var uhii: u128 = @bitCast(sum.hi); |
| 293 | const bits_lost = -@as(i32, @intCast((uhii >> 112) & 0x7FFF)) - scale + 1; |
| 294 | if ((bits_lost != 1) == (uhii & 1 != 0)) { |
| 295 | const uloi: u128 = @bitCast(sum.lo); |
| 296 | uhii = uhii + 1 - (((uhii ^ uloi) >> 126) & 2); |
| 297 | sum.hi = @bitCast(uhii); |
| 298 | } |
| 299 | } |
| 300 | return math.scalbn(sum.hi, scale); |
| 301 | } |
| 302 | |
| 303 | /// Compute a*b exactly, returning the exact result in a struct dd. We assume |
| 304 | /// that both a and b are normalized, so no underflow or overflow will occur. |
| 305 | /// The current rounding mode must be round-to-nearest. |
| 306 | fn dd_mul128(a: f128, b: f128) dd128 { |
| 307 | var ret: dd128 = undefined; |
| 308 | const split: f128 = 0x1.0p57 + 1.0; |
| 309 | |
| 310 | var p = a * split; |
| 311 | var ha = a - p; |
| 312 | ha += p; |
| 313 | const la = a - ha; |
| 314 | |
| 315 | p = b * split; |
| 316 | var hb = b - p; |
| 317 | hb += p; |
| 318 | const lb = b - hb; |
| 319 | |
| 320 | p = ha * hb; |
| 321 | const q = ha * lb + la * hb; |
| 322 | |
| 323 | ret.hi = p + q; |
| 324 | ret.lo = p - ret.hi + q + la * lb; |
| 325 | return ret; |
| 326 | } |
| 327 | |
| 328 | test "32" { |
| 329 | const epsilon = 0.000001; |
| 330 | |
| 331 | try expect(math.approxEqAbs(f32, fma_f32(0.0, 5.0, 9.124), 9.124, epsilon)); |
| 332 | try expect(math.approxEqAbs(f32, fma_f32(0.2, 5.0, 9.124), 10.124, epsilon)); |
| 333 | try expect(math.approxEqAbs(f32, fma_f32(0.8923, 5.0, 9.124), 13.5855, epsilon)); |
| 334 | try expect(math.approxEqAbs(f32, fma_f32(1.5, 5.0, 9.124), 16.624, epsilon)); |
| 335 | try expect(math.approxEqAbs(f32, fma_f32(37.45, 5.0, 9.124), 196.374004, epsilon)); |
| 336 | try expect(math.approxEqAbs(f32, fma_f32(89.123, 5.0, 9.124), 454.739005, epsilon)); |
| 337 | try expect(math.approxEqAbs(f32, fma_f32(123123.234375, 5.0, 9.124), 615625.295875, epsilon)); |
| 338 | } |
| 339 | |
| 340 | test "64" { |
| 341 | const epsilon = 0.000001; |
| 342 | |
| 343 | try expect(math.approxEqAbs(f64, fma_f64(0.0, 5.0, 9.124), 9.124, epsilon)); |
| 344 | try expect(math.approxEqAbs(f64, fma_f64(0.2, 5.0, 9.124), 10.124, epsilon)); |
| 345 | try expect(math.approxEqAbs(f64, fma_f64(0.8923, 5.0, 9.124), 13.5855, epsilon)); |
| 346 | try expect(math.approxEqAbs(f64, fma_f64(1.5, 5.0, 9.124), 16.624, epsilon)); |
| 347 | try expect(math.approxEqAbs(f64, fma_f64(37.45, 5.0, 9.124), 196.374, epsilon)); |
| 348 | try expect(math.approxEqAbs(f64, fma_f64(89.123, 5.0, 9.124), 454.739, epsilon)); |
| 349 | try expect(math.approxEqAbs(f64, fma_f64(123123.234375, 5.0, 9.124), 615625.295875, epsilon)); |
| 350 | } |
| 351 | |
| 352 | test "128" { |
| 353 | const epsilon = 0.000001; |
| 354 | |
| 355 | try expect(math.approxEqAbs(f128, fma_f128(0.0, 5.0, 9.124), 9.124, epsilon)); |
| 356 | try expect(math.approxEqAbs(f128, fma_f128(0.2, 5.0, 9.124), 10.124, epsilon)); |
| 357 | try expect(math.approxEqAbs(f128, fma_f128(0.8923, 5.0, 9.124), 13.5855, epsilon)); |
| 358 | try expect(math.approxEqAbs(f128, fma_f128(1.5, 5.0, 9.124), 16.624, epsilon)); |
| 359 | try expect(math.approxEqAbs(f128, fma_f128(37.45, 5.0, 9.124), 196.374, epsilon)); |
| 360 | try expect(math.approxEqAbs(f128, fma_f128(89.123, 5.0, 9.124), 454.739, epsilon)); |
| 361 | try expect(math.approxEqAbs(f128, fma_f128(123123.234375, 5.0, 9.124), 615625.295875, epsilon)); |
| 362 | } |