zig

fork of https://codeberg.org/ziglang/zig
Log | Files | Refs | README | LICENSE

exp.zig (11677B) - Raw


      1 // Ported from musl, which is licensed under the MIT license:
      2 // https://git.musl-libc.org/cgit/musl/tree/COPYRIGHT
      3 //
      4 // https://git.musl-libc.org/cgit/musl/tree/src/math/expf.c
      5 // https://git.musl-libc.org/cgit/musl/tree/src/math/exp.c
      6 
      7 const std = @import("std");
      8 const builtin = @import("builtin");
      9 const arch = builtin.cpu.arch;
     10 const math = std.math;
     11 const mem = std.mem;
     12 const expect = std.testing.expect;
     13 const expectEqual = std.testing.expectEqual;
     14 const common = @import("common.zig");
     15 
     16 pub const panic = common.panic;
     17 
     18 comptime {
     19     @export(&__exph, .{ .name = "__exph", .linkage = common.linkage, .visibility = common.visibility });
     20     @export(&expf, .{ .name = "expf", .linkage = common.linkage, .visibility = common.visibility });
     21     @export(&exp, .{ .name = "exp", .linkage = common.linkage, .visibility = common.visibility });
     22     @export(&__expx, .{ .name = "__expx", .linkage = common.linkage, .visibility = common.visibility });
     23     if (common.want_ppc_abi) {
     24         @export(&expq, .{ .name = "expf128", .linkage = common.linkage, .visibility = common.visibility });
     25     }
     26     @export(&expq, .{ .name = "expq", .linkage = common.linkage, .visibility = common.visibility });
     27     @export(&expl, .{ .name = "expl", .linkage = common.linkage, .visibility = common.visibility });
     28 }
     29 
     30 pub fn __exph(a: f16) callconv(.c) f16 {
     31     // TODO: more efficient implementation
     32     return @floatCast(expf(a));
     33 }
     34 
     35 pub fn expf(x_: f32) callconv(.c) f32 {
     36     const half = [_]f32{ 0.5, -0.5 };
     37     const ln2hi = 6.9314575195e-1;
     38     const ln2lo = 1.4286067653e-6;
     39     const invln2 = 1.4426950216e+0;
     40     const P1 = 1.6666625440e-1;
     41     const P2 = -2.7667332906e-3;
     42 
     43     var x = x_;
     44     var hx: u32 = @bitCast(x);
     45     const sign: i32 = @intCast(hx >> 31);
     46     hx &= 0x7FFFFFFF;
     47 
     48     if (math.isNan(x)) {
     49         return x;
     50     }
     51 
     52     // |x| >= -87.33655 or nan
     53     if (hx >= 0x42AEAC50) {
     54         // nan
     55         if (hx > 0x7F800000) {
     56             return x;
     57         }
     58         // x >= 88.722839
     59         if (hx >= 0x42b17218 and sign == 0) {
     60             return x * 0x1.0p127;
     61         }
     62         if (sign != 0) {
     63             if (common.want_float_exceptions) mem.doNotOptimizeAway(-0x1.0p-149 / x); // overflow
     64             // x <= -103.972084
     65             if (hx >= 0x42CFF1B5) {
     66                 return 0;
     67             }
     68         }
     69     }
     70 
     71     var k: i32 = undefined;
     72     var hi: f32 = undefined;
     73     var lo: f32 = undefined;
     74 
     75     // |x| > 0.5 * ln2
     76     if (hx > 0x3EB17218) {
     77         // |x| > 1.5 * ln2
     78         if (hx > 0x3F851592) {
     79             k = @intFromFloat(invln2 * x + half[@intCast(sign)]);
     80         } else {
     81             k = 1 - sign - sign;
     82         }
     83 
     84         const fk: f32 = @floatFromInt(k);
     85         hi = x - fk * ln2hi;
     86         lo = fk * ln2lo;
     87         x = hi - lo;
     88     }
     89     // |x| > 2^(-14)
     90     else if (hx > 0x39000000) {
     91         k = 0;
     92         hi = x;
     93         lo = 0;
     94     } else {
     95         if (common.want_float_exceptions) mem.doNotOptimizeAway(0x1.0p127 + x); // inexact
     96         return 1 + x;
     97     }
     98 
     99     const xx = x * x;
    100     const c = x - xx * (P1 + xx * P2);
    101     const y = 1 + (x * c / (2 - c) - lo + hi);
    102 
    103     if (k == 0) {
    104         return y;
    105     } else {
    106         return math.scalbn(y, k);
    107     }
    108 }
    109 
    110 pub fn exp(x_: f64) callconv(.c) f64 {
    111     const half = [_]f64{ 0.5, -0.5 };
    112     const ln2hi: f64 = 6.93147180369123816490e-01;
    113     const ln2lo: f64 = 1.90821492927058770002e-10;
    114     const invln2: f64 = 1.44269504088896338700e+00;
    115     const P1: f64 = 1.66666666666666019037e-01;
    116     const P2: f64 = -2.77777777770155933842e-03;
    117     const P3: f64 = 6.61375632143793436117e-05;
    118     const P4: f64 = -1.65339022054652515390e-06;
    119     const P5: f64 = 4.13813679705723846039e-08;
    120 
    121     var x = x_;
    122     const ux: u64 = @bitCast(x);
    123     var hx = ux >> 32;
    124     const sign: i32 = @intCast(hx >> 31);
    125     hx &= 0x7FFFFFFF;
    126 
    127     if (math.isNan(x)) {
    128         return x;
    129     }
    130 
    131     // |x| >= 708.39 or nan
    132     if (hx >= 0x4086232B) {
    133         // nan
    134         if (hx > 0x7FF00000) {
    135             return x;
    136         }
    137         if (x > 709.782712893383973096) {
    138             // overflow if x != inf
    139             if (!math.isInf(x)) {
    140                 math.raiseOverflow();
    141             }
    142             return math.inf(f64);
    143         }
    144         if (x < -708.39641853226410622) {
    145             // underflow if x != -inf
    146             // if (common.want_float_exceptions) mem.doNotOptimizeAway(@as(f32, -0x1.0p-149 / x));
    147             if (x < -745.13321910194110842) {
    148                 return 0;
    149             }
    150         }
    151     }
    152 
    153     // argument reduction
    154     var k: i32 = undefined;
    155     var hi: f64 = undefined;
    156     var lo: f64 = undefined;
    157 
    158     // |x| > 0.5 * ln2
    159     if (hx > 0x3FD62E42) {
    160         // |x| >= 1.5 * ln2
    161         if (hx > 0x3FF0A2B2) {
    162             k = @intFromFloat(invln2 * x + half[@intCast(sign)]);
    163         } else {
    164             k = 1 - sign - sign;
    165         }
    166 
    167         const dk: f64 = @floatFromInt(k);
    168         hi = x - dk * ln2hi;
    169         lo = dk * ln2lo;
    170         x = hi - lo;
    171     }
    172     // |x| > 2^(-28)
    173     else if (hx > 0x3E300000) {
    174         k = 0;
    175         hi = x;
    176         lo = 0;
    177     } else {
    178         // inexact if x != 0
    179         // if (common.want_float_exceptions) mem.doNotOptimizeAway(0x1.0p1023 + x);
    180         return 1 + x;
    181     }
    182 
    183     const xx = x * x;
    184     const c = x - xx * (P1 + xx * (P2 + xx * (P3 + xx * (P4 + xx * P5))));
    185     const y = 1 + (x * c / (2 - c) - lo + hi);
    186 
    187     if (k == 0) {
    188         return y;
    189     } else {
    190         return math.scalbn(y, k);
    191     }
    192 }
    193 
    194 pub fn __expx(a: f80) callconv(.c) f80 {
    195     // TODO: more efficient implementation
    196     return @floatCast(expq(a));
    197 }
    198 
    199 pub fn expq(a: f128) callconv(.c) f128 {
    200     // TODO: more correct implementation
    201     return exp(@floatCast(a));
    202 }
    203 
    204 pub fn expl(x: c_longdouble) callconv(.c) c_longdouble {
    205     switch (@typeInfo(c_longdouble).float.bits) {
    206         16 => return __exph(x),
    207         32 => return expf(x),
    208         64 => return exp(x),
    209         80 => return __expx(x),
    210         128 => return expq(x),
    211         else => @compileError("unreachable"),
    212     }
    213 }
    214 
    215 test "expf() special" {
    216     try expectEqual(expf(0.0), 1.0);
    217     try expectEqual(expf(-0.0), 1.0);
    218     try expectEqual(expf(1.0), math.e);
    219     try expectEqual(expf(math.ln2), 2.0);
    220     try expectEqual(expf(math.inf(f32)), math.inf(f32));
    221     try expect(math.isPositiveZero(expf(-math.inf(f32))));
    222     try expect(math.isNan(expf(math.nan(f32))));
    223     try expect(math.isNan(expf(math.snan(f32))));
    224 }
    225 
    226 test "expf() sanity" {
    227     try expectEqual(expf(-0x1.0223a0p+3), 0x1.490320p-12);
    228     try expectEqual(expf(0x1.161868p+2), 0x1.34712ap+6);
    229     try expectEqual(expf(-0x1.0c34b4p+3), 0x1.e06b1ap-13);
    230     try expectEqual(expf(-0x1.a206f0p+2), 0x1.7dd484p-10);
    231     try expectEqual(expf(0x1.288bbcp+3), 0x1.4abc80p+13);
    232     try expectEqual(expf(0x1.52efd0p-1), 0x1.f04a9cp+0);
    233     try expectEqual(expf(-0x1.a05cc8p-2), 0x1.54f1e0p-1);
    234     try expectEqual(expf(0x1.1f9efap-1), 0x1.c0f628p+0);
    235     try expectEqual(expf(0x1.8c5db0p-1), 0x1.1599b2p+1);
    236     try expectEqual(expf(-0x1.5b86eap-1), 0x1.03b572p-1);
    237     try expectEqual(expf(-0x1.57f25cp+2), 0x1.2fbea2p-8);
    238     try expectEqual(expf(0x1.c7d310p+3), 0x1.76eefp+20);
    239     try expectEqual(expf(0x1.19be70p+4), 0x1.52d3dep+25);
    240     try expectEqual(expf(-0x1.ab6d70p+3), 0x1.a88adep-20);
    241     try expectEqual(expf(-0x1.5ac18ep+2), 0x1.22b328p-8);
    242     try expectEqual(expf(-0x1.925982p-1), 0x1.d2acc0p-2);
    243     try expectEqual(expf(0x1.7221cep+3), 0x1.9c2ceap+16);
    244     try expectEqual(expf(0x1.11a0d4p+4), 0x1.980ee6p+24);
    245     try expectEqual(expf(-0x1.ae41a2p+1), 0x1.1c28d0p-5);
    246     try expectEqual(expf(-0x1.329154p+4), 0x1.47ef94p-28);
    247 }
    248 
    249 test "expf() boundary" {
    250     try expectEqual(expf(0x1.62e42ep+6), 0x1.ffff08p+127); // The last value before the result gets infinite
    251     try expectEqual(expf(0x1.62e430p+6), math.inf(f32)); // The first value that gives inf
    252     try expectEqual(expf(0x1.fffffep+127), math.inf(f32)); // Max input value
    253     try expectEqual(expf(0x1p-149), 1.0); // Min positive input value
    254     try expectEqual(expf(-0x1p-149), 1.0); // Min negative input value
    255     try expectEqual(expf(0x1p-126), 1.0); // First positive subnormal input
    256     try expectEqual(expf(-0x1p-126), 1.0); // First negative subnormal input
    257     try expectEqual(expf(-0x1.9fe368p+6), 0x1p-149); // The last value before the result flushes to zero
    258     try expectEqual(expf(-0x1.9fe36ap+6), 0.0); // The first value at which the result flushes to zero
    259     try expectEqual(expf(-0x1.5d589ep+6), 0x1.00004cp-126); // The last value before the result flushes to subnormal
    260     try expectEqual(expf(-0x1.5d58a0p+6), 0x1.ffff98p-127); // The first value for which the result flushes to subnormal
    261 
    262 }
    263 
    264 test "exp() special" {
    265     try expectEqual(exp(0.0), 1.0);
    266     try expectEqual(exp(-0.0), 1.0);
    267     // TODO: Accuracy error - off in the last bit in 64-bit, disagreeing with GCC
    268     // try expectEqual(exp(1.0), math.e);
    269     try expectEqual(exp(math.ln2), 2.0);
    270     try expectEqual(exp(math.inf(f64)), math.inf(f64));
    271     try expect(math.isPositiveZero(exp(-math.inf(f64))));
    272     try expect(math.isNan(exp(math.nan(f64))));
    273     try expect(math.isNan(exp(math.snan(f64))));
    274 }
    275 
    276 test "exp() sanity" {
    277     try expectEqual(exp(-0x1.02239f3c6a8f1p+3), 0x1.490327ea61235p-12);
    278     try expectEqual(exp(0x1.161868e18bc67p+2), 0x1.34712ed238c04p+6);
    279     try expectEqual(exp(-0x1.0c34b3e01e6e7p+3), 0x1.e06b1b6c18e64p-13);
    280     try expectEqual(exp(-0x1.a206f0a19dcc4p+2), 0x1.7dd47f810e68cp-10);
    281     try expectEqual(exp(0x1.288bbb0d6a1e6p+3), 0x1.4abc77496e07ep+13);
    282     try expectEqual(exp(0x1.52efd0cd80497p-1), 0x1.f04a9c1080500p+0);
    283     try expectEqual(exp(-0x1.a05cc754481d1p-2), 0x1.54f1e0fd3ea0dp-1);
    284     try expectEqual(exp(0x1.1f9ef934745cbp-1), 0x1.c0f6266a6a547p+0);
    285     try expectEqual(exp(0x1.8c5db097f7442p-1), 0x1.1599b1d4a25fbp+1);
    286     try expectEqual(exp(-0x1.5b86ea8118a0ep-1), 0x1.03b5728a00229p-1);
    287     try expectEqual(exp(-0x1.57f25b2b5006dp+2), 0x1.2fbea6a01cab9p-8);
    288     try expectEqual(exp(0x1.c7d30fb825911p+3), 0x1.76eeed45a0634p+20);
    289     try expectEqual(exp(0x1.19be709de7505p+4), 0x1.52d3eb7be6844p+25);
    290     try expectEqual(exp(-0x1.ab6d6fba96889p+3), 0x1.a88ae12f985d6p-20);
    291     try expectEqual(exp(-0x1.5ac18e27084ddp+2), 0x1.22b327da9cca6p-8);
    292     try expectEqual(exp(-0x1.925981b093c41p-1), 0x1.d2acc046b55f7p-2);
    293     try expectEqual(exp(0x1.7221cd18455f5p+3), 0x1.9c2cde8699cfbp+16);
    294     try expectEqual(exp(0x1.11a0d4a51b239p+4), 0x1.980ef612ff182p+24);
    295     try expectEqual(exp(-0x1.ae41a1079de4dp+1), 0x1.1c28d16bb3222p-5);
    296     try expectEqual(exp(-0x1.329153103b871p+4), 0x1.47efa6ddd0d22p-28);
    297 }
    298 
    299 test "exp() boundary" {
    300     try expectEqual(exp(0x1.62e42fefa39efp+9), 0x1.fffffffffff2ap+1023); // The last value before the result gets infinite
    301     try expectEqual(exp(0x1.62e42fefa39f0p+9), math.inf(f64)); // The first value that gives inf
    302     try expectEqual(exp(0x1.fffffffffffffp+1023), math.inf(f64)); // Max input value
    303     try expectEqual(exp(0x1p-1074), 1.0); // Min positive input value
    304     try expectEqual(exp(-0x1p-1074), 1.0); // Min negative input value
    305     try expectEqual(exp(0x1p-1022), 1.0); // First positive subnormal input
    306     try expectEqual(exp(-0x1p-1022), 1.0); // First negative subnormal input
    307     try expectEqual(exp(-0x1.74910d52d3051p+9), 0x1p-1074); // The last value before the result flushes to zero
    308     try expectEqual(exp(-0x1.74910d52d3052p+9), 0.0); // The first value at which the result flushes to zero
    309     try expectEqual(exp(-0x1.6232bdd7abcd2p+9), 0x1.000000000007cp-1022); // The last value before the result flushes to subnormal
    310     try expectEqual(exp(-0x1.6232bdd7abcd3p+9), 0x1.ffffffffffcf8p-1023); // The first value for which the result flushes to subnormal
    311 }