BF16 vector conversion and round-to-odd about riscv-bfloat16 HOT 2 CLOSED

Thanks for the question, Nicolas.

Is there a particular use for round-to-odd conversion targeting BF16?

As I understand it, round to odd is used for an intermediate conversion, when the ultimate goal is to round to an even smaller format. Since we don't have a smaller format, it doesn't seem like we need a round-to-odd flavor when converting to BF16.

This is further clarified by the following non-normative text from the RVV spec:

A full set of floating-point narrowing conversions is not supported as single instructions. Conversions can be implemented in a sequence of halving steps. Results are equivalently rounded and the same exception flags are raised if all but the last halving step use round-towards-odd (vfncvt.rod.f.f.w). Only the final step should use the desired rounding mode.

from riscv-bfloat16.

Is there a particular use for round-to-odd conversion targeting BF16?

I did not have a particular example in mind and you explanation makes me think there might not be one (I cannot think of a use-case FP32 -> BF16 -> FP8 where someone would want to avoid double-rounding issue through intermediary use of round-to-odd and as you said FP8 is not even a legal format yet).

I am going to ask around to user more familiar with this matter.

from riscv-bfloat16.