| Message ID | 20171211125705.16120-6-alex.bennee@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | re-factor softfloat and add fp16 functions | expand |
Alex Bennée <alex.bennee@linaro.org> writes: > This defines the same set of common constants for float 16 as defined > for 32 and 64 bit floats. These are often used by target helper > functions. > > Signed-off-by: Alex Bennée <alex.bennee@linaro.org> > --- > include/fpu/softfloat.h | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/include/fpu/softfloat.h b/include/fpu/softfloat.h > index 17dfe60dbd..5a9258c57c 100644 > --- a/include/fpu/softfloat.h > +++ b/include/fpu/softfloat.h > @@ -395,6 +395,13 @@ static inline float16 float16_set_sign(float16 a, int sign) > return make_float16((float16_val(a) & 0x7fff) | (sign << 15)); > } > > +#define float16_zero make_float16(0) > +#define float16_one make_float16(0x3a00) > +#define float16_ln2 make_float16(0x34d1) > +#define float16_pi make_float16(0x4448) > +#define float16_half make_float16(0x3800) > +#define float16_infinity make_float16(0x7a00) And: #define float16_infinity make_float16(0x7c00) > + > /*---------------------------------------------------------------------------- > | The pattern for a default generated half-precision NaN. > *----------------------------------------------------------------------------*/ -- Alex Bennée
Hi Alex, On 12/11/2017 09:56 AM, Alex Bennée wrote: > This defines the same set of common constants for float 16 as defined > for 32 and 64 bit floats. These are often used by target helper > functions. > > Signed-off-by: Alex Bennée <alex.bennee@linaro.org> > --- > include/fpu/softfloat.h | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/include/fpu/softfloat.h b/include/fpu/softfloat.h > index 17dfe60dbd..5a9258c57c 100644 > --- a/include/fpu/softfloat.h > +++ b/include/fpu/softfloat.h > @@ -395,6 +395,13 @@ static inline float16 float16_set_sign(float16 a, int sign) > return make_float16((float16_val(a) & 0x7fff) | (sign << 15)); > } > > +#define float16_zero make_float16(0) ok > +#define float16_one make_float16(0x3a00) I'm a bit confused... >>> [np.fromstring(struct.pack("<H", x), dtype=np.float16)[0] for x in [0, 0x3a00, 0x34d1, 0x4448, 0x3800, 0x7a00]] [0.0, 0.75, 0.30103, 4.2812, 0.5, 49152.0] However: >>> ['0x' + binascii.hexlify(np.array([x], '>f2').tostring()) for x in [0, 1, math.log(2), np.pi, 0.5, np.inf]] ['0x0000', '0x3c00', '0x398c', '0x4248', '0x3800', '0x7c00'] It seems the MSB bit of the mantissa got shifted as the LSB of the biased exponent... > +#define float16_ln2 make_float16(0x34d1) incorrect? 0x398c > +#define float16_pi make_float16(0x4448) incorrect? 0x4248 > +#define float16_half make_float16(0x3800) ok > +#define float16_infinity make_float16(0x7a00) incorrect? 0x7c00 > + > /*---------------------------------------------------------------------------- > | The pattern for a default generated half-precision NaN. > *----------------------------------------------------------------------------*/ >
On 12/15/2017 05:37 AM, Philippe Mathieu-Daudé wrote: > Hi Alex, > > On 12/11/2017 09:56 AM, Alex Bennée wrote: >> This defines the same set of common constants for float 16 as defined >> for 32 and 64 bit floats. These are often used by target helper >> functions. >> >> Signed-off-by: Alex Bennée <alex.bennee@linaro.org> >> --- >> include/fpu/softfloat.h | 7 +++++++ >> 1 file changed, 7 insertions(+) >> >> diff --git a/include/fpu/softfloat.h b/include/fpu/softfloat.h >> index 17dfe60dbd..5a9258c57c 100644 >> --- a/include/fpu/softfloat.h >> +++ b/include/fpu/softfloat.h >> @@ -395,6 +395,13 @@ static inline float16 float16_set_sign(float16 a, int sign) >> return make_float16((float16_val(a) & 0x7fff) | (sign << 15)); >> } >> >> +#define float16_zero make_float16(0) > > ok > >> +#define float16_one make_float16(0x3a00) > > I'm a bit confused... > >>>> [np.fromstring(struct.pack("<H", x), dtype=np.float16)[0] for x in > [0, 0x3a00, 0x34d1, 0x4448, 0x3800, 0x7a00]] > [0.0, 0.75, 0.30103, 4.2812, 0.5, 49152.0] > > However: > >>>> ['0x' + binascii.hexlify(np.array([x], '>f2').tostring()) for x in > [0, 1, math.log(2), np.pi, 0.5, np.inf]] > ['0x0000', '0x3c00', '0x398c', '0x4248', '0x3800', '0x7c00'] > > It seems the MSB bit of the mantissa got shifted as the LSB of the > biased exponent... > >> +#define float16_ln2 make_float16(0x34d1) > > incorrect? 0x398c > >> +#define float16_pi make_float16(0x4448) > > incorrect? 0x4248 > >> +#define float16_half make_float16(0x3800) > > ok > >> +#define float16_infinity make_float16(0x7a00) > > incorrect? 0x7c00 All of Phil's numbers are correct -- I double-checked with gcc. Other than 0, 1 and +inf, I doubt any of the others will actually be used. Perhaps we should just leave them out? r~
Richard Henderson <richard.henderson@linaro.org> writes: > On 12/15/2017 05:37 AM, Philippe Mathieu-Daudé wrote: >> Hi Alex, >> >> On 12/11/2017 09:56 AM, Alex Bennée wrote: >>> This defines the same set of common constants for float 16 as defined >>> for 32 and 64 bit floats. These are often used by target helper >>> functions. >>> >>> Signed-off-by: Alex Bennée <alex.bennee@linaro.org> >>> --- >>> include/fpu/softfloat.h | 7 +++++++ >>> 1 file changed, 7 insertions(+) >>> >>> diff --git a/include/fpu/softfloat.h b/include/fpu/softfloat.h >>> index 17dfe60dbd..5a9258c57c 100644 >>> --- a/include/fpu/softfloat.h >>> +++ b/include/fpu/softfloat.h >>> @@ -395,6 +395,13 @@ static inline float16 float16_set_sign(float16 a, int sign) >>> return make_float16((float16_val(a) & 0x7fff) | (sign << 15)); >>> } >>> >>> +#define float16_zero make_float16(0) >> >> ok >> >>> +#define float16_one make_float16(0x3a00) >> >> I'm a bit confused... >> >>>>> [np.fromstring(struct.pack("<H", x), dtype=np.float16)[0] for x in >> [0, 0x3a00, 0x34d1, 0x4448, 0x3800, 0x7a00]] >> [0.0, 0.75, 0.30103, 4.2812, 0.5, 49152.0] >> >> However: >> >>>>> ['0x' + binascii.hexlify(np.array([x], '>f2').tostring()) for x in >> [0, 1, math.log(2), np.pi, 0.5, np.inf]] >> ['0x0000', '0x3c00', '0x398c', '0x4248', '0x3800', '0x7c00'] >> >> It seems the MSB bit of the mantissa got shifted as the LSB of the >> biased exponent... >> >>> +#define float16_ln2 make_float16(0x34d1) >> >> incorrect? 0x398c >> >>> +#define float16_pi make_float16(0x4448) >> >> incorrect? 0x4248 >> >>> +#define float16_half make_float16(0x3800) >> >> ok >> >>> +#define float16_infinity make_float16(0x7a00) >> >> incorrect? 0x7c00 > > All of Phil's numbers are correct -- I double-checked with gcc. > > Other than 0, 1 and +inf, I doubt any of the others will actually be used. > Perhaps we should just leave them out? I was following the other sizes. It seems it's x80 which uses the additional ones in helpers: floatx80_ln2 floatx80_pi Should I delete all the unused constants from the other sizes as well? -- Alex Bennée
On 01/04/2018 06:09 AM, Alex Bennée wrote: > I was following the other sizes. It seems it's x80 which uses the > additional ones in helpers: > > floatx80_ln2 > floatx80_pi > > Should I delete all the unused constants from the other sizes as well? Yeah, let's do that. r~
diff --git a/include/fpu/softfloat.h b/include/fpu/softfloat.h index 17dfe60dbd..5a9258c57c 100644 --- a/include/fpu/softfloat.h +++ b/include/fpu/softfloat.h @@ -395,6 +395,13 @@ static inline float16 float16_set_sign(float16 a, int sign) return make_float16((float16_val(a) & 0x7fff) | (sign << 15)); } +#define float16_zero make_float16(0) +#define float16_one make_float16(0x3a00) +#define float16_ln2 make_float16(0x34d1) +#define float16_pi make_float16(0x4448) +#define float16_half make_float16(0x3800) +#define float16_infinity make_float16(0x7a00) + /*---------------------------------------------------------------------------- | The pattern for a default generated half-precision NaN. *----------------------------------------------------------------------------*/
This defines the same set of common constants for float 16 as defined for 32 and 64 bit floats. These are often used by target helper functions. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> --- include/fpu/softfloat.h | 7 +++++++ 1 file changed, 7 insertions(+) -- 2.15.1