std: Fix tanh for negative inputs

It turns out the code was not ported correctly from C and produced wrong
results for negative input values. As a bonus fix the NaN codepath by
adding yet another missing piece of code.

Spotted in #9047
This commit is contained in:
LemonBoy
2021-06-11 09:23:48 +02:00
committed by Veikka Tuominen
parent 2ce033f415
commit 56c0a66ce3

View File

@@ -37,89 +37,75 @@ fn tanh32(x: f32) f32 {
const u = @bitCast(u32, x);
const ux = u & 0x7FFFFFFF;
const ax = @bitCast(f32, ux);
const sign = (u >> 31) != 0;
var t: f32 = undefined;
if (x == 0.0 or math.isNan(x)) {
return x;
}
// |x| < log(3) / 2 ~= 0.5493 or nan
if (ux > 0x3F0C9F54) {
// |x| > 10
if (ux > 0x41200000) {
t = 1.0;
t = 1.0 + 0 / x;
} else {
t = math.expm1(2 * x);
t = math.expm1(2 * ax);
t = 1 - 2 / (t + 2);
}
}
// |x| > log(5 / 3) / 2 ~= 0.2554
else if (ux > 0x3E82C578) {
t = math.expm1(2 * x);
t = math.expm1(2 * ax);
t = t / (t + 2);
}
// |x| >= 0x1.0p-126
else if (ux >= 0x00800000) {
t = math.expm1(-2 * x);
t = math.expm1(-2 * ax);
t = -t / (t + 2);
}
// |x| is subnormal
else {
math.doNotOptimizeAway(x * x);
t = x;
math.doNotOptimizeAway(ax * ax);
t = ax;
}
if (u >> 31 != 0) {
return -t;
} else {
return t;
}
return if (sign) -t else t;
}
fn tanh64(x: f64) f64 {
const u = @bitCast(u64, x);
const w = @intCast(u32, u >> 32);
const ax = @bitCast(f64, u & (maxInt(u64) >> 1));
const ux = u & 0x7FFFFFFFFFFFFFFF;
const w = @intCast(u32, ux >> 32);
const ax = @bitCast(f64, ux);
const sign = (u >> 63) != 0;
var t: f64 = undefined;
// TODO: Shouldn't need these checks.
if (x == 0.0 or math.isNan(x)) {
return x;
}
// |x| < log(3) / 2 ~= 0.5493 or nan
if (w > 0x3FE193EA) {
// |x| > 20 or nan
if (w > 0x40340000) {
t = 1.0;
t = 1.0 - 0 / ax;
} else {
t = math.expm1(2 * x);
t = math.expm1(2 * ax);
t = 1 - 2 / (t + 2);
}
}
// |x| > log(5 / 3) / 2 ~= 0.2554
else if (w > 0x3FD058AE) {
t = math.expm1(2 * x);
t = math.expm1(2 * ax);
t = t / (t + 2);
}
// |x| >= 0x1.0p-1022
else if (w >= 0x00100000) {
t = math.expm1(-2 * x);
t = math.expm1(-2 * ax);
t = -t / (t + 2);
}
// |x| is subnormal
else {
math.doNotOptimizeAway(@floatCast(f32, x));
t = x;
math.doNotOptimizeAway(@floatCast(f32, ax));
t = ax;
}
if (u >> 63 != 0) {
return -t;
} else {
return t;
}
return if (sign) -t else t;
}
test "math.tanh" {
@@ -135,6 +121,9 @@ test "math.tanh32" {
try expect(math.approxEqAbs(f32, tanh32(0.8923), 0.712528, epsilon));
try expect(math.approxEqAbs(f32, tanh32(1.5), 0.905148, epsilon));
try expect(math.approxEqAbs(f32, tanh32(37.45), 1.0, epsilon));
try expect(math.approxEqAbs(f32, tanh32(-0.8923), -0.712528, epsilon));
try expect(math.approxEqAbs(f32, tanh32(-1.5), -0.905148, epsilon));
try expect(math.approxEqAbs(f32, tanh32(-37.45), -1.0, epsilon));
}
test "math.tanh64" {
@@ -145,6 +134,9 @@ test "math.tanh64" {
try expect(math.approxEqAbs(f64, tanh64(0.8923), 0.712528, epsilon));
try expect(math.approxEqAbs(f64, tanh64(1.5), 0.905148, epsilon));
try expect(math.approxEqAbs(f64, tanh64(37.45), 1.0, epsilon));
try expect(math.approxEqAbs(f64, tanh64(-0.8923), -0.712528, epsilon));
try expect(math.approxEqAbs(f64, tanh64(-1.5), -0.905148, epsilon));
try expect(math.approxEqAbs(f64, tanh64(-37.45), -1.0, epsilon));
}
test "math.tanh32.special" {