[Cfp-interest 2052] Re: Subnormals

[Steve (Numerics) Canon]

> On Jul 5, 2021, at 6:48 PM, Jim Thomas <jaswthomas at sbcglobal.net> wrote:
> 
>> I believe that adding flushed subnormals shall be treated as zeros defines
>> their behaviour.
> 
> I believe ARM FTZ mode causes subnormal results to be set to zero. Saying it treats them as zero suggests that they are still subnormal (in some sense).

The effects of the FZ bit in the ARM FPCR are defined by the FPUnpack() and FPRound() operations in ARM’s pseudocode documentation. If the encoded exponent of an input is zero and fpcr.FZ is 1, then the unpacked input value is zero. If a result is tiny and fpcr.FZ is 1, then the result is zero.

These two aspects correspond to the x86 MXCSR bits DAZ and FTZ, respectively, except that x86 detects tininess after rounding and ARM detects it before rounding, so exactly which results are flushed are different for multiply and multiply-add.

– Steve
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.oakapple.net/pipermail/cfp-interest/attachments/20210706/5ac70c0d/attachment-0001.htm>