Re: [Qemu-devel] [PATCH v2 11/27] fpu/softfloat: support ARM Alternative half-precision

[Peter Maydell]

On 12 May 2018 at 01:42, Richard Henderson <address@hidden> wrote:
> From: Alex Bennée <address@hidden>
>
> For float16 ARM supports an alternative half-precision format which
> sacrifices the ability to represent NaN/Inf in return for a higher
> dynamic range. To support this I've added an additional
> FloatFmt (float16_params_ahp).
>
> The new FloatFmt flag (arm_althp) is then used to modify the behaviour
> of canonicalize and round_canonical with respect to representation and
> exception raising.
>
> Finally the float16_to_floatN and floatN_to_float16 conversion
> routines select the new alternative FloatFmt when !ieee.
>
> Signed-off-by: Alex Bennée <address@hidden>
> Signed-off-by: Richard Henderson <address@hidden>

If we do this like this then we have a regression for the
alternate-halfprec conversions between the patch where we change
over the conversions functions and this one.

To fix that we could put the parts of this patch that change
existing-in-master code like round_canonical() first, and then
fold the parts that fix up the conversion functions into the
patch which is currently 10/27.

> ---
> v3
>   - squash NaN to 0 if destination is AHP F16
> v4
>   - handle inf -> ahp max in float_to_float not round_canonical
>   - assert no nan and inf for ahp in round_canonical
>   - check ahp before snan in float_to_float
> ---
>  fpu/softfloat.c | 95 +++++++++++++++++++++++++++++++++++++++++--------
>  1 file changed, 81 insertions(+), 14 deletions(-)
>
> diff --git a/fpu/softfloat.c b/fpu/softfloat.c
> index aa219223ff..15a272759d 100644
> --- a/fpu/softfloat.c
> +++ b/fpu/softfloat.c
> @@ -211,8 +211,10 @@ typedef struct {
>   *   frac_shift: shift to normalise the fraction with DECOMPOSED_BINARY_POINT
>   * The following are computed based the size of fraction
>   *   frac_lsb: least significant bit of fraction
> - *   fram_lsbm1: the bit bellow the least significant bit (for rounding)
> + *   frac_lsbm1: the bit bellow the least significant bit (for rounding)

"below" (I see this typo slipped past us the first time around)

thanks
-- PMM