Implementing truncation rounding

[alexrp]

I just want to make sure I've got this one right.

When truncation rounding is enabled, should an implementation simply round all inputs towards zero before performing an FP computation and then round the result towards zero? Or is there something more subtle that I'm missing?

[timpart]

[alexrp]

Can you think of a good way to implement this with regular floating point types and operations in a language other than C? Trouble is that I can't access <fenv.h>, so I can't simply change rounding mode...

[timpart]

Can't think of a way offhand, but I'm not into that kind of computing. You might have to resort to doing it in C with integers (32 bit for most things, 64 bit for multiplication). Perhaps a web search will reveal someone has already done this. This looks interesting

Tim

[ysapir]

[ysapir]

If you are working with floats, then converting inputs to doubles and doing your calculation in doubles, then rounding back to floats should enable you to keep the intermediate resolution, I think. The casting back to floats (truncating rather than rounding) can be done by extracting the bit fields comprising an FP format and re-packing them in a 32-bit variable.