317 lines
7.4 KiB
Zig
317 lines
7.4 KiB
Zig
// Special Cases:
|
|
//
|
|
// - expm1(+inf) = +inf
|
|
// - expm1(-inf) = -1
|
|
// - expm1(nan) = nan
|
|
|
|
const math = @import("index.zig");
|
|
const assert = @import("../debug.zig").assert;
|
|
|
|
// TODO issue #393
|
|
pub const expm1 = expm1_workaround;
|
|
|
|
pub fn expm1_workaround(x: var) -> @typeOf(x) {
|
|
const T = @typeOf(x);
|
|
switch (T) {
|
|
f32 => @inlineCall(expm1_32, x),
|
|
f64 => @inlineCall(expm1_64, x),
|
|
else => @compileError("exp1m not implemented for " ++ @typeName(T)),
|
|
}
|
|
}
|
|
|
|
fn expm1_32(x_: f32) -> f32 {
|
|
const o_threshold: f32 = 8.8721679688e+01;
|
|
const ln2_hi: f32 = 6.9313812256e-01;
|
|
const ln2_lo: f32 = 9.0580006145e-06;
|
|
const invln2: f32 = 1.4426950216e+00;
|
|
const Q1: f32 = -3.3333212137e-2;
|
|
const Q2: f32 = 1.5807170421e-3;
|
|
|
|
var x = x_;
|
|
const ux = @bitCast(u32, x);
|
|
const hx = ux & 0x7FFFFFFF;
|
|
const sign = hx >> 31;
|
|
|
|
// TODO: Shouldn't need this check explicitly.
|
|
if (math.isNegativeInf(x)) {
|
|
return -1.0;
|
|
}
|
|
|
|
// |x| >= 27 * ln2
|
|
if (hx >= 0x4195B844) {
|
|
// nan
|
|
if (hx > 0x7F800000) {
|
|
return x;
|
|
}
|
|
if (sign != 0) {
|
|
return -1;
|
|
}
|
|
if (x > o_threshold) {
|
|
x *= 0x1.0p127;
|
|
return x;
|
|
}
|
|
}
|
|
|
|
var hi: f32 = undefined;
|
|
var lo: f32 = undefined;
|
|
var c: f32 = undefined;
|
|
var k: i32 = undefined;
|
|
|
|
// |x| > 0.5 * ln2
|
|
if (hx > 0x3EB17218) {
|
|
// |x| < 1.5 * ln2
|
|
if (hx < 0x3F851592) {
|
|
if (sign == 0) {
|
|
hi = x - ln2_hi;
|
|
lo = ln2_lo;
|
|
k = 1;
|
|
} else {
|
|
hi = x + ln2_hi;
|
|
lo = -ln2_lo;
|
|
k = -1;
|
|
}
|
|
} else {
|
|
var kf = invln2 * x;
|
|
if (sign != 0) {
|
|
kf -= 0.5;
|
|
} else {
|
|
kf += 0.5;
|
|
}
|
|
|
|
k = i32(kf);
|
|
const t = f32(k);
|
|
hi = x - t * ln2_hi;
|
|
lo = t * ln2_lo;
|
|
}
|
|
|
|
x = hi - lo;
|
|
c = (hi - x) - lo;
|
|
}
|
|
// |x| < 2^(-25)
|
|
else if (hx < 0x33000000) {
|
|
if (hx < 0x00800000) {
|
|
math.forceEval(x * x);
|
|
}
|
|
return x;
|
|
}
|
|
else {
|
|
k = 0;
|
|
}
|
|
|
|
const hfx = 0.5 * x;
|
|
const hxs = x * hfx;
|
|
const r1 = 1.0 + hxs * (Q1 + hxs * Q2);
|
|
const t = 3.0 - r1 * hfx;
|
|
var e = hxs * ((r1 - t) / (6.0 - x * t));
|
|
|
|
// c is 0
|
|
if (k == 0) {
|
|
return x - (x * e - hxs);
|
|
}
|
|
|
|
e = x * (e - c) - c;
|
|
e -= hxs;
|
|
|
|
// exp(x) ~ 2^k (x_reduced - e + 1)
|
|
if (k == -1) {
|
|
return 0.5 * (x - e) - 0.5;
|
|
}
|
|
if (k == 1) {
|
|
if (x < -0.25) {
|
|
return -2.0 * (e - (x + 0.5));
|
|
} else {
|
|
return 1.0 + 2.0 * (x - e);
|
|
}
|
|
}
|
|
|
|
const twopk = @bitCast(f32, u32((0x7F + k) <<% 23));
|
|
|
|
if (k < 0 or k > 56) {
|
|
var y = x - e + 1.0;
|
|
if (k == 128) {
|
|
y = y * 2.0 * 0x1.0p127;
|
|
} else {
|
|
y = y * twopk;
|
|
}
|
|
|
|
return y - 1.0;
|
|
}
|
|
|
|
const uf = @bitCast(f32, u32(0x7F - k) << 23);
|
|
if (k < 23) {
|
|
return (x - e + (1 - uf)) * twopk;
|
|
} else {
|
|
return (x - (e + uf) + 1) * twopk;
|
|
}
|
|
}
|
|
|
|
fn expm1_64(x_: f64) -> f64 {
|
|
const o_threshold: f64 = 7.09782712893383973096e+02;
|
|
const ln2_hi: f64 = 6.93147180369123816490e-01;
|
|
const ln2_lo: f64 = 1.90821492927058770002e-10;
|
|
const invln2: f64 = 1.44269504088896338700e+00;
|
|
const Q1: f64 = -3.33333333333331316428e-02;
|
|
const Q2: f64 = 1.58730158725481460165e-03;
|
|
const Q3: f64 = -7.93650757867487942473e-05;
|
|
const Q4: f64 = 4.00821782732936239552e-06;
|
|
const Q5: f64 = -2.01099218183624371326e-07;
|
|
|
|
var x = x_;
|
|
const ux = @bitCast(u64, x);
|
|
const hx = u32(ux >> 32) & 0x7FFFFFFF;
|
|
const sign = hx >> 63;
|
|
|
|
if (math.isNegativeInf(x)) {
|
|
return -1.0;
|
|
}
|
|
|
|
// |x| >= 56 * ln2
|
|
if (hx >= 0x4043687A) {
|
|
// exp1md(nan) = nan
|
|
if (hx > 0x7FF00000) {
|
|
return x;
|
|
}
|
|
// exp1md(-ve) = -1
|
|
if (sign != 0) {
|
|
return -1;
|
|
}
|
|
if (x > o_threshold) {
|
|
math.raiseOverflow();
|
|
return math.inf(f64);
|
|
}
|
|
}
|
|
|
|
var hi: f64 = undefined;
|
|
var lo: f64 = undefined;
|
|
var c: f64 = undefined;
|
|
var k: i32 = undefined;
|
|
|
|
// |x| > 0.5 * ln2
|
|
if (hx > 0x3FD62E42) {
|
|
// |x| < 1.5 * ln2
|
|
if (hx < 0x3FF0A2B2) {
|
|
if (sign == 0) {
|
|
hi = x - ln2_hi;
|
|
lo = ln2_lo;
|
|
k = 1;
|
|
} else {
|
|
hi = x + ln2_hi;
|
|
lo = -ln2_lo;
|
|
k = -1;
|
|
}
|
|
} else {
|
|
var kf = invln2 * x;
|
|
if (sign != 0) {
|
|
kf -= 0.5;
|
|
} else {
|
|
kf += 0.5;
|
|
}
|
|
|
|
k = i32(kf);
|
|
const t = f64(k);
|
|
hi = x - t * ln2_hi;
|
|
lo = t * ln2_lo;
|
|
}
|
|
|
|
x = hi - lo;
|
|
c = (hi - x) - lo;
|
|
}
|
|
// |x| < 2^(-54)
|
|
else if (hx < 0x3C900000) {
|
|
if (hx < 0x00100000) {
|
|
math.forceEval(f32(x));
|
|
}
|
|
return x;
|
|
}
|
|
else {
|
|
k = 0;
|
|
}
|
|
|
|
const hfx = 0.5 * x;
|
|
const hxs = x * hfx;
|
|
const r1 = 1.0 + hxs * (Q1 + hxs * (Q2 + hxs * (Q3 + hxs * (Q4 + hxs * Q5))));
|
|
const t = 3.0 - r1 * hfx;
|
|
var e = hxs * ((r1 - t) / (6.0 - x * t));
|
|
|
|
// c is 0
|
|
if (k == 0) {
|
|
return x - (x * e - hxs);
|
|
}
|
|
|
|
e = x * (e - c) - c;
|
|
e -= hxs;
|
|
|
|
// exp(x) ~ 2^k (x_reduced - e + 1)
|
|
if (k == -1) {
|
|
return 0.5 * (x - e) - 0.5;
|
|
}
|
|
if (k == 1) {
|
|
if (x < -0.25) {
|
|
return -2.0 * (e - (x + 0.5));
|
|
} else {
|
|
return 1.0 + 2.0 * (x - e);
|
|
}
|
|
}
|
|
|
|
const twopk = @bitCast(f64, u64(0x3FF + k) <<% 52);
|
|
|
|
if (k < 0 or k > 56) {
|
|
var y = x - e + 1.0;
|
|
if (k == 1024) {
|
|
y = y * 2.0 * 0x1.0p1022 * 10;
|
|
} else {
|
|
y = y * twopk;
|
|
}
|
|
|
|
return y - 1.0;
|
|
}
|
|
|
|
const uf = @bitCast(f64, u64(0x3FF - k) << 52);
|
|
if (k < 20) {
|
|
return (x - e + (1 - uf)) * twopk;
|
|
} else {
|
|
return (x - (e + uf) + 1) * twopk;
|
|
}
|
|
}
|
|
|
|
test "math.exp1m" {
|
|
assert(expm1(f32(0.0)) == expm1_32(0.0));
|
|
assert(expm1(f64(0.0)) == expm1_64(0.0));
|
|
}
|
|
|
|
test "math.expm1_32" {
|
|
const epsilon = 0.000001;
|
|
|
|
assert(expm1_32(0.0) == 0.0);
|
|
assert(math.approxEq(f32, expm1_32(0.0), 0.0, epsilon));
|
|
assert(math.approxEq(f32, expm1_32(0.2), 0.221403, epsilon));
|
|
assert(math.approxEq(f32, expm1_32(0.8923), 1.440737, epsilon));
|
|
assert(math.approxEq(f32, expm1_32(1.5), 3.481689, epsilon));
|
|
}
|
|
|
|
test "math.expm1_64" {
|
|
const epsilon = 0.000001;
|
|
|
|
assert(expm1_64(0.0) == 0.0);
|
|
assert(math.approxEq(f64, expm1_64(0.0), 0.0, epsilon));
|
|
assert(math.approxEq(f64, expm1_64(0.2), 0.221403, epsilon));
|
|
assert(math.approxEq(f64, expm1_64(0.8923), 1.440737, epsilon));
|
|
assert(math.approxEq(f64, expm1_64(1.5), 3.481689, epsilon));
|
|
}
|
|
|
|
test "math.expm1_32.special" {
|
|
const epsilon = 0.000001;
|
|
|
|
assert(math.isPositiveInf(expm1_32(math.inf(f32))));
|
|
assert(expm1_32(-math.inf(f32)) == -1.0);
|
|
assert(math.isNan(expm1_32(math.nan(f32))));
|
|
}
|
|
|
|
test "math.expm1_64.special" {
|
|
const epsilon = 0.000001;
|
|
|
|
assert(math.isPositiveInf(expm1_64(math.inf(f64))));
|
|
assert(expm1_64(-math.inf(f64)) == -1.0);
|
|
assert(math.isNan(expm1_64(math.nan(f64))));
|
|
}
|