| 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/tanhf.c |
| 5 | // https://git.musl-libc.org/cgit/musl/tree/src/math/tanh.c |
| 6 | |
| 7 | const std = @import("../std.zig"); |
| 8 | const math = std.math; |
| 9 | const mem = std.mem; |
| 10 | const expect = std.testing.expect; |
| 11 | const expo2 = @import("expo2.zig").expo2; |
| 12 | const maxInt = std.math.maxInt; |
| 13 | |
| 14 | /// Returns the hyperbolic tangent of x. |
| 15 | /// |
| 16 | /// Special Cases: |
| 17 | /// - tanh(+-0) = +-0 |
| 18 | /// - tanh(+-inf) = +-1 |
| 19 | /// - tanh(nan) = nan |
| 20 | pub fn tanh(x: anytype) @TypeOf(x) { |
| 21 | const T = @TypeOf(x); |
| 22 | return switch (T) { |
| 23 | f32 => tanh32(x), |
| 24 | f64 => tanh64(x), |
| 25 | else => @compileError("tanh not implemented for " ++ @typeName(T)), |
| 26 | }; |
| 27 | } |
| 28 | |
| 29 | // tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x)) |
| 30 | // = (exp(2x) - 1) / (exp(2x) - 1 + 2) |
| 31 | // = (1 - exp(-2x)) / (exp(-2x) - 1 + 2) |
| 32 | fn tanh32(x: f32) f32 { |
| 33 | const u = @as(u32, @bitCast(x)); |
| 34 | const ux = u & 0x7FFFFFFF; |
| 35 | const ax = @as(f32, @bitCast(ux)); |
| 36 | const sign = (u >> 31) != 0; |
| 37 | |
| 38 | var t: f32 = undefined; |
| 39 | |
| 40 | // |x| < log(3) / 2 ~= 0.5493 or nan |
| 41 | if (ux > 0x3F0C9F54) { |
| 42 | // |x| > 10 |
| 43 | if (ux > 0x41200000) { |
| 44 | t = 1.0 + 0 / x; |
| 45 | } else { |
| 46 | t = math.expm1(2 * ax); |
| 47 | t = 1 - 2 / (t + 2); |
| 48 | } |
| 49 | } |
| 50 | // |x| > log(5 / 3) / 2 ~= 0.2554 |
| 51 | else if (ux > 0x3E82C578) { |
| 52 | t = math.expm1(2 * ax); |
| 53 | t = t / (t + 2); |
| 54 | } |
| 55 | // |x| >= 0x1.0p-126 |
| 56 | else if (ux >= 0x00800000) { |
| 57 | t = math.expm1(-2 * ax); |
| 58 | t = -t / (t + 2); |
| 59 | } |
| 60 | // |x| is subnormal |
| 61 | else { |
| 62 | mem.doNotOptimizeAway(ax * ax); |
| 63 | t = ax; |
| 64 | } |
| 65 | |
| 66 | return if (sign) -t else t; |
| 67 | } |
| 68 | |
| 69 | fn tanh64(x: f64) f64 { |
| 70 | const u = @as(u64, @bitCast(x)); |
| 71 | const ux = u & 0x7FFFFFFFFFFFFFFF; |
| 72 | const w = @as(u32, @intCast(ux >> 32)); |
| 73 | const ax = @as(f64, @bitCast(ux)); |
| 74 | const sign = (u >> 63) != 0; |
| 75 | |
| 76 | var t: f64 = undefined; |
| 77 | |
| 78 | // |x| < log(3) / 2 ~= 0.5493 or nan |
| 79 | if (w > 0x3FE193EA) { |
| 80 | // |x| > 20 or nan |
| 81 | if (w > 0x40340000) { |
| 82 | t = 1.0 - 0 / ax; |
| 83 | } else { |
| 84 | t = math.expm1(2 * ax); |
| 85 | t = 1 - 2 / (t + 2); |
| 86 | } |
| 87 | } |
| 88 | // |x| > log(5 / 3) / 2 ~= 0.2554 |
| 89 | else if (w > 0x3FD058AE) { |
| 90 | t = math.expm1(2 * ax); |
| 91 | t = t / (t + 2); |
| 92 | } |
| 93 | // |x| >= 0x1.0p-1022 |
| 94 | else if (w >= 0x00100000) { |
| 95 | t = math.expm1(-2 * ax); |
| 96 | t = -t / (t + 2); |
| 97 | } |
| 98 | // |x| is subnormal |
| 99 | else { |
| 100 | mem.doNotOptimizeAway(@as(f32, @floatCast(ax))); |
| 101 | t = ax; |
| 102 | } |
| 103 | |
| 104 | return if (sign) -t else t; |
| 105 | } |
| 106 | |
| 107 | test tanh { |
| 108 | try expect(tanh(@as(f32, 1.5)) == tanh32(1.5)); |
| 109 | try expect(tanh(@as(f64, 1.5)) == tanh64(1.5)); |
| 110 | } |
| 111 | |
| 112 | test tanh32 { |
| 113 | const epsilon = 0.000001; |
| 114 | |
| 115 | try expect(math.approxEqAbs(f32, tanh32(0.0), 0.0, epsilon)); |
| 116 | try expect(math.approxEqAbs(f32, tanh32(0.2), 0.197375, epsilon)); |
| 117 | try expect(math.approxEqAbs(f32, tanh32(0.8923), 0.712528, epsilon)); |
| 118 | try expect(math.approxEqAbs(f32, tanh32(1.5), 0.905148, epsilon)); |
| 119 | try expect(math.approxEqAbs(f32, tanh32(37.45), 1.0, epsilon)); |
| 120 | try expect(math.approxEqAbs(f32, tanh32(-0.8923), -0.712528, epsilon)); |
| 121 | try expect(math.approxEqAbs(f32, tanh32(-1.5), -0.905148, epsilon)); |
| 122 | try expect(math.approxEqAbs(f32, tanh32(-37.45), -1.0, epsilon)); |
| 123 | } |
| 124 | |
| 125 | test tanh64 { |
| 126 | const epsilon = 0.000001; |
| 127 | |
| 128 | try expect(math.approxEqAbs(f64, tanh64(0.0), 0.0, epsilon)); |
| 129 | try expect(math.approxEqAbs(f64, tanh64(0.2), 0.197375, epsilon)); |
| 130 | try expect(math.approxEqAbs(f64, tanh64(0.8923), 0.712528, epsilon)); |
| 131 | try expect(math.approxEqAbs(f64, tanh64(1.5), 0.905148, epsilon)); |
| 132 | try expect(math.approxEqAbs(f64, tanh64(37.45), 1.0, epsilon)); |
| 133 | try expect(math.approxEqAbs(f64, tanh64(-0.8923), -0.712528, epsilon)); |
| 134 | try expect(math.approxEqAbs(f64, tanh64(-1.5), -0.905148, epsilon)); |
| 135 | try expect(math.approxEqAbs(f64, tanh64(-37.45), -1.0, epsilon)); |
| 136 | } |
| 137 | |
| 138 | test "tanh32.special" { |
| 139 | try expect(math.isPositiveZero(tanh32(0.0))); |
| 140 | try expect(math.isNegativeZero(tanh32(-0.0))); |
| 141 | try expect(tanh32(math.inf(f32)) == 1.0); |
| 142 | try expect(tanh32(-math.inf(f32)) == -1.0); |
| 143 | try expect(math.isNan(tanh32(math.nan(f32)))); |
| 144 | } |
| 145 | |
| 146 | test "tanh64.special" { |
| 147 | try expect(math.isPositiveZero(tanh64(0.0))); |
| 148 | try expect(math.isNegativeZero(tanh64(-0.0))); |
| 149 | try expect(tanh64(math.inf(f64)) == 1.0); |
| 150 | try expect(tanh64(-math.inf(f64)) == -1.0); |
| 151 | try expect(math.isNan(tanh64(math.nan(f64)))); |
| 152 | } |