On 10 April 2017 at 17:38, Michael Jones <michael.jo...@gmail.com> wrote:
> It is not crazy in any way to use the special NaN payload--it was designed > just for you. Dr. Kahan and others provided the broad set of NaNs in '754 > for just this purpose, for example, to encode missing data values and N/A > data values separately for statistical processing, or to allow great > freedom to encode provenance or intent in the lower bits. > Great, thanks for the information. > > One reservation is that FP operations with quiet NaNs and signalling NaNs > are not inherently full speed, so benchmark to make sure you are satisfied > about performance in your specific context. > At the level we care about it's all bit twiddling (e.g. XOR-based compression), so this shouldn't affect us. It'll all be benchmarked anyway. Yours, Brian > > On Mon, Apr 10, 2017 at 7:27 AM, <brian.bra...@robustperception.io> wrote: > >> Hi, >> >> I'm working on a feature for the Prometheus monitoring system that >> requires adding special entries in our database. >> Our values are float64, and I plan on taking advantage of the fact that >> there's 2^52-1 valid bit representations for NaN. >> So I'll choose one of those representations for an actual NaN (as we >> support that), and others for the special entries I need. >> >> What I'd like to do is change the few bits of code that need to know >> about the special values, and continue to use float64 types as-is elsewhere. >> >> I worry though that with the complexity of modern systems that depending >> on the bit representation of a NaN not to change as a float64 is passed >> around may not be wise. >> Not having to change everything to [8]byte throughout the stack would be >> handy, but I am willing to switch all the relevant code if needed. >> >> What I'm wondering is: >> 1) Is it safe to presume Go itself won't change the bit representation of >> a NaN? >> 2) It is safe to presume FPUs won't change the bit representation of a >> NaN? >> 3) If 1) and not 2), is there a way control/know when a float64 could >> have ended up in a FPU? At the level where it matters, we're almost >> entirely doing bit manipulation rather than floating point operations. >> >> As this is a mildly crazy idea, I've unfortunately not been able to find >> any documentation indicating one way or other. >> >> Thanks, >> Brian >> >> -- >> You received this message because you are subscribed to the Google Groups >> "golang-nuts" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to golang-nuts+unsubscr...@googlegroups.com. >> For more options, visit https://groups.google.com/d/optout. >> > > > > -- > Michael T. Jones > michael.jo...@gmail.com > -- Brian Brazil www.robustperception.io -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
