Message ID | 20180514221219.7091-14-richard.henderson@linaro.org |
---|---|
State | Superseded |
Headers | show |
Series | softfloat patch roundup | expand |
On 14 May 2018 at 23:12, Richard Henderson <richard.henderson@linaro.org> wrote: > From: Alex Bennée <alex.bennee@linaro.org> > > For float16 ARM supports an alternative half-precision format which > sacrifices the ability to represent NaN/Inf in return for a higher > dynamic range. 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. > > Usage of this new flag waits until we re-factor float-to-float conversions. > > Signed-off-by: Alex Bennée <alex.bennee@linaro.org> > Signed-off-by: Richard Henderson <richard.henderson@linaro.org> > > --- > 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 > v5 > - split out canonicalize and round_canonical changes from the rest > --- > fpu/softfloat.c | 19 ++++++++++++++++--- > 1 file changed, 16 insertions(+), 3 deletions(-) > > diff --git a/fpu/softfloat.c b/fpu/softfloat.c > index 41253c6749..55d0d01ec3 100644 > --- a/fpu/softfloat.c > +++ b/fpu/softfloat.c > @@ -220,8 +220,10 @@ 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" > * round_mask/roundeven_mask: masks used for rounding > + * The following optional modifiers are available: > + * arm_althp: handle ARM Alternative Half Precision > */ Otherwise Reviewed-by: Peter Maydell <peter.maydell@linaro.org> thanks -- PMM
diff --git a/fpu/softfloat.c b/fpu/softfloat.c index 41253c6749..55d0d01ec3 100644 --- a/fpu/softfloat.c +++ b/fpu/softfloat.c @@ -220,8 +220,10 @@ 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) * round_mask/roundeven_mask: masks used for rounding + * The following optional modifiers are available: + * arm_althp: handle ARM Alternative Half Precision */ typedef struct { int exp_size; @@ -233,6 +235,7 @@ typedef struct { uint64_t frac_lsbm1; uint64_t round_mask; uint64_t roundeven_mask; + bool arm_althp; } FloatFmt; /* Expand fields based on the size of exponent and fraction */ @@ -324,7 +327,7 @@ static inline float64 float64_pack_raw(FloatParts p) static FloatParts canonicalize(FloatParts part, const FloatFmt *parm, float_status *status) { - if (part.exp == parm->exp_max) { + if (part.exp == parm->exp_max && !parm->arm_althp) { if (part.frac == 0) { part.cls = float_class_inf; } else { @@ -413,7 +416,15 @@ static FloatParts round_canonical(FloatParts p, float_status *s, } frac >>= frac_shift; - if (unlikely(exp >= exp_max)) { + if (parm->arm_althp) { + /* ARM Alt HP eschews Inf and NaN for a wider exponent. */ + if (unlikely(exp > exp_max)) { + /* Overflow. Return the maximum normal. */ + flags = float_flag_invalid; + exp = exp_max; + frac = -1; + } + } else if (unlikely(exp >= exp_max)) { flags |= float_flag_overflow | float_flag_inexact; if (overflow_norm) { exp = exp_max - 1; @@ -464,12 +475,14 @@ static FloatParts round_canonical(FloatParts p, float_status *s, case float_class_inf: do_inf: + assert(!parm->arm_althp); exp = exp_max; frac = 0; break; case float_class_qnan: case float_class_snan: + assert(!parm->arm_althp); exp = exp_max; frac >>= parm->frac_shift; break;