softfloat: Factor out RoundAndPackFloat16 and NormalizeFloat16Subnormal
In preparation for adding conversions between float16 and float64, factor out code currently done inline in the float16<=>float32 conversion functions into functions RoundAndPackFloat16 and NormalizeFloat16Subnormal along the lines of the existing versions for the other float types. Note that we change the handling of zExp from the inline code to match the API of the other RoundAndPackFloat functions; however we leave the positioning of the binary point between bits 22 and 23 rather than shifting it up to the high end of the word. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net>
This commit is contained in:
parent
e9ef30c079
commit
84c6b75d85
209
fpu/softfloat.c
209
fpu/softfloat.c
|
@ -3086,6 +3086,127 @@ static float16 packFloat16(flag zSign, int_fast16_t zExp, uint16_t zSig)
|
|||
(((uint32_t)zSign) << 15) + (((uint32_t)zExp) << 10) + zSig);
|
||||
}
|
||||
|
||||
/*----------------------------------------------------------------------------
|
||||
| Takes an abstract floating-point value having sign `zSign', exponent `zExp',
|
||||
| and significand `zSig', and returns the proper half-precision floating-
|
||||
| point value corresponding to the abstract input. Ordinarily, the abstract
|
||||
| value is simply rounded and packed into the half-precision format, with
|
||||
| the inexact exception raised if the abstract input cannot be represented
|
||||
| exactly. However, if the abstract value is too large, the overflow and
|
||||
| inexact exceptions are raised and an infinity or maximal finite value is
|
||||
| returned. If the abstract value is too small, the input value is rounded to
|
||||
| a subnormal number, and the underflow and inexact exceptions are raised if
|
||||
| the abstract input cannot be represented exactly as a subnormal half-
|
||||
| precision floating-point number.
|
||||
| The `ieee' flag indicates whether to use IEEE standard half precision, or
|
||||
| ARM-style "alternative representation", which omits the NaN and Inf
|
||||
| encodings in order to raise the maximum representable exponent by one.
|
||||
| The input significand `zSig' has its binary point between bits 22
|
||||
| and 23, which is 13 bits to the left of the usual location. This shifted
|
||||
| significand must be normalized or smaller. If `zSig' is not normalized,
|
||||
| `zExp' must be 0; in that case, the result returned is a subnormal number,
|
||||
| and it must not require rounding. In the usual case that `zSig' is
|
||||
| normalized, `zExp' must be 1 less than the ``true'' floating-point exponent.
|
||||
| Note the slightly odd position of the binary point in zSig compared with the
|
||||
| other roundAndPackFloat functions. This should probably be fixed if we
|
||||
| need to implement more float16 routines than just conversion.
|
||||
| The handling of underflow and overflow follows the IEC/IEEE Standard for
|
||||
| Binary Floating-Point Arithmetic.
|
||||
*----------------------------------------------------------------------------*/
|
||||
|
||||
static float32 roundAndPackFloat16(flag zSign, int_fast16_t zExp,
|
||||
uint32_t zSig, flag ieee STATUS_PARAM)
|
||||
{
|
||||
int maxexp = ieee ? 29 : 30;
|
||||
uint32_t mask;
|
||||
uint32_t increment;
|
||||
int8 roundingMode;
|
||||
bool rounding_bumps_exp;
|
||||
bool is_tiny = false;
|
||||
|
||||
/* Calculate the mask of bits of the mantissa which are not
|
||||
* representable in half-precision and will be lost.
|
||||
*/
|
||||
if (zExp < 1) {
|
||||
/* Will be denormal in halfprec */
|
||||
mask = 0x00ffffff;
|
||||
if (zExp >= -11) {
|
||||
mask >>= 11 + zExp;
|
||||
}
|
||||
} else {
|
||||
/* Normal number in halfprec */
|
||||
mask = 0x00001fff;
|
||||
}
|
||||
|
||||
roundingMode = STATUS(float_rounding_mode);
|
||||
switch (roundingMode) {
|
||||
case float_round_nearest_even:
|
||||
increment = (mask + 1) >> 1;
|
||||
if ((zSig & mask) == increment) {
|
||||
increment = zSig & (increment << 1);
|
||||
}
|
||||
break;
|
||||
case float_round_up:
|
||||
increment = zSign ? 0 : mask;
|
||||
break;
|
||||
case float_round_down:
|
||||
increment = zSign ? mask : 0;
|
||||
break;
|
||||
default: /* round_to_zero */
|
||||
increment = 0;
|
||||
break;
|
||||
}
|
||||
|
||||
rounding_bumps_exp = (zSig + increment >= 0x01000000);
|
||||
|
||||
if (zExp > maxexp || (zExp == maxexp && rounding_bumps_exp)) {
|
||||
if (ieee) {
|
||||
float_raise(float_flag_overflow | float_flag_inexact STATUS_VAR);
|
||||
return packFloat16(zSign, 0x1f, 0);
|
||||
} else {
|
||||
float_raise(float_flag_invalid STATUS_VAR);
|
||||
return packFloat16(zSign, 0x1f, 0x3ff);
|
||||
}
|
||||
}
|
||||
|
||||
if (zExp < 0) {
|
||||
/* Note that flush-to-zero does not affect half-precision results */
|
||||
is_tiny =
|
||||
(STATUS(float_detect_tininess) == float_tininess_before_rounding)
|
||||
|| (zExp < -1)
|
||||
|| (!rounding_bumps_exp);
|
||||
}
|
||||
if (zSig & mask) {
|
||||
float_raise(float_flag_inexact STATUS_VAR);
|
||||
if (is_tiny) {
|
||||
float_raise(float_flag_underflow STATUS_VAR);
|
||||
}
|
||||
}
|
||||
|
||||
zSig += increment;
|
||||
if (rounding_bumps_exp) {
|
||||
zSig >>= 1;
|
||||
zExp++;
|
||||
}
|
||||
|
||||
if (zExp < -10) {
|
||||
return packFloat16(zSign, 0, 0);
|
||||
}
|
||||
if (zExp < 0) {
|
||||
zSig >>= -zExp;
|
||||
zExp = 0;
|
||||
}
|
||||
return packFloat16(zSign, zExp, zSig >> 13);
|
||||
}
|
||||
|
||||
static void normalizeFloat16Subnormal(uint32_t aSig, int_fast16_t *zExpPtr,
|
||||
uint32_t *zSigPtr)
|
||||
{
|
||||
int8_t shiftCount = countLeadingZeros32(aSig) - 21;
|
||||
*zSigPtr = aSig << shiftCount;
|
||||
*zExpPtr = 1 - shiftCount;
|
||||
}
|
||||
|
||||
/* Half precision floats come in two formats: standard IEEE and "ARM" format.
|
||||
The latter gains extra exponent range by omitting the NaN/Inf encodings. */
|
||||
|
||||
|
@ -3106,15 +3227,12 @@ float32 float16_to_float32(float16 a, flag ieee STATUS_PARAM)
|
|||
return packFloat32(aSign, 0xff, 0);
|
||||
}
|
||||
if (aExp == 0) {
|
||||
int8 shiftCount;
|
||||
|
||||
if (aSig == 0) {
|
||||
return packFloat32(aSign, 0, 0);
|
||||
}
|
||||
|
||||
shiftCount = countLeadingZeros32( aSig ) - 21;
|
||||
aSig = aSig << shiftCount;
|
||||
aExp = -shiftCount;
|
||||
normalizeFloat16Subnormal(aSig, &aExp, &aSig);
|
||||
aExp--;
|
||||
}
|
||||
return packFloat32( aSign, aExp + 0x70, aSig << 13);
|
||||
}
|
||||
|
@ -3124,12 +3242,6 @@ float16 float32_to_float16(float32 a, flag ieee STATUS_PARAM)
|
|||
flag aSign;
|
||||
int_fast16_t aExp;
|
||||
uint32_t aSig;
|
||||
uint32_t mask;
|
||||
uint32_t increment;
|
||||
int8 roundingMode;
|
||||
int maxexp = ieee ? 15 : 16;
|
||||
bool rounding_bumps_exp;
|
||||
bool is_tiny = false;
|
||||
|
||||
a = float32_squash_input_denormal(a STATUS_VAR);
|
||||
|
||||
|
@ -3164,80 +3276,9 @@ float16 float32_to_float16(float32 a, flag ieee STATUS_PARAM)
|
|||
* codepath.
|
||||
*/
|
||||
aSig |= 0x00800000;
|
||||
aExp -= 0x7f;
|
||||
/* Calculate the mask of bits of the mantissa which are not
|
||||
* representable in half-precision and will be lost.
|
||||
*/
|
||||
if (aExp < -14) {
|
||||
/* Will be denormal in halfprec */
|
||||
mask = 0x00ffffff;
|
||||
if (aExp >= -24) {
|
||||
mask >>= 25 + aExp;
|
||||
}
|
||||
} else {
|
||||
/* Normal number in halfprec */
|
||||
mask = 0x00001fff;
|
||||
}
|
||||
aExp -= 0x71;
|
||||
|
||||
roundingMode = STATUS(float_rounding_mode);
|
||||
switch (roundingMode) {
|
||||
case float_round_nearest_even:
|
||||
increment = (mask + 1) >> 1;
|
||||
if ((aSig & mask) == increment) {
|
||||
increment = aSig & (increment << 1);
|
||||
}
|
||||
break;
|
||||
case float_round_up:
|
||||
increment = aSign ? 0 : mask;
|
||||
break;
|
||||
case float_round_down:
|
||||
increment = aSign ? mask : 0;
|
||||
break;
|
||||
default: /* round_to_zero */
|
||||
increment = 0;
|
||||
break;
|
||||
}
|
||||
|
||||
rounding_bumps_exp = (aSig + increment >= 0x01000000);
|
||||
|
||||
if (aExp > maxexp || (aExp == maxexp && rounding_bumps_exp)) {
|
||||
if (ieee) {
|
||||
float_raise(float_flag_overflow | float_flag_inexact STATUS_VAR);
|
||||
return packFloat16(aSign, 0x1f, 0);
|
||||
} else {
|
||||
float_raise(float_flag_invalid STATUS_VAR);
|
||||
return packFloat16(aSign, 0x1f, 0x3ff);
|
||||
}
|
||||
}
|
||||
|
||||
if (aExp < -14) {
|
||||
/* Note that flush-to-zero does not affect half-precision results */
|
||||
is_tiny =
|
||||
(STATUS(float_detect_tininess) == float_tininess_before_rounding)
|
||||
|| (aExp < -15)
|
||||
|| (!rounding_bumps_exp);
|
||||
}
|
||||
if (aSig & mask) {
|
||||
float_raise(float_flag_inexact STATUS_VAR);
|
||||
if (is_tiny) {
|
||||
float_raise(float_flag_underflow STATUS_VAR);
|
||||
}
|
||||
}
|
||||
|
||||
aSig += increment;
|
||||
if (rounding_bumps_exp) {
|
||||
aSig >>= 1;
|
||||
aExp++;
|
||||
}
|
||||
|
||||
if (aExp < -24) {
|
||||
return packFloat16(aSign, 0, 0);
|
||||
}
|
||||
if (aExp < -14) {
|
||||
aSig >>= -14 - aExp;
|
||||
aExp = -14;
|
||||
}
|
||||
return packFloat16(aSign, aExp + 14, aSig >> 13);
|
||||
return roundAndPackFloat16(aSign, aExp, aSig, ieee STATUS_VAR);
|
||||
}
|
||||
|
||||
/*----------------------------------------------------------------------------
|
||||
|
|
Loading…
Reference in a new issue