anon (anon@anon.com) on 3/1/11 wrote:
---------------------------
[...]
>Interval arithmetic?
>
That wouldn't really solve my problem. No, just use two FP values as (supported by the hardware), combined to effectively double the mantissa width. The limited exponent range will not be a problem. But rather than the more common pair of doubles, it will be a pair of single precision values.

OpenCL mandates a fully precise fused multiply add, so the impact on speed should be quite bearable.