On Thursday, 7 March 2013 at 09:43:21 UTC, jerro wrote:
Are errors of this magnitude to be expected using doubles, or
is this a compiler bug?
Errors of this magnitude are to be expected. the value of accum
in your example is somewhere around 3e+08, so the relative
error is around 1e-15, and double.epsilon is 2.22045e-16.
This.
You can use reals to store the intermediary results. A real has
the largest hardware supported size, which is 80 bits for x87.
It's not a silver bullet but can be useful in cases like this.
