Widen tanh acceptable range (output values) (gpuweb#5199)

petermcneeleychromium · Peter McNeeley · mehmetoguzderin · web-flow · commit 9cd0b020c35e · 2025-07-30T13:27:02.000-04:00
After a bit of investigation it has been determined that tanh has a max
Absolute error of 1e-5 for some devices (nvidia)

We discussed polyfilling this function (sinh/cosh) but this inaccuracy
is likely intentional as tanh is commonly used in sigmoid functions for
ML.

crbug.com/390221422

Co-authored-by: Peter McNeeley &lt;petermcneeley@google.com&gt;
Co-authored-by: Mehmet Oguz Derin &lt;mehmetoguzderin@mehmetoguzderin.com&gt;
diff --git a/wgsl/index.bs b/wgsl/index.bs
@@ -12938,7 +12938,10 @@ the rules in [[#floating-point-rounding-and-overflow]] apply.
   <tr><td>`sqrt(x)`<td colspan=2 style="text-align:left;">Inherited from `1.0 / inverseSqrt(x)`
   <tr><td>`step(edge, x)`<td colspan=2 style="text-align:left;">Correctly rounded
   <tr><td>`tan(x)`<td colspan=2 style="text-align:left;">Inherited from `sin(x) / cos(x)`
-  <tr><td>`tanh(x)`<td colspan=2 style="text-align:left;">Inherited from `sinh(x) / cosh(x)`
+  <tr><td>`tanh(x)`<td colspan=2 style="text-align:left;">
+       The worse of:
+       * Absolute error 1.0&times;10<sup>-5</sup>
+       * Inherited from `sinh(x) / cosh(x)`
   <tr><td>`transpose(x)`<td colspan=2 style="text-align:left;">Correctly rounded
   <tr><td>`trunc(x)`<td colspan=2 style="text-align:left;">Correctly rounded
 
