diff options
Diffstat (limited to 'fpu')
| -rw-r--r-- | fpu/softfloat.c | 168 |
1 files changed, 168 insertions, 0 deletions
diff --git a/fpu/softfloat.c b/fpu/softfloat.c index 64de0b4..77d81ab 100644 --- a/fpu/softfloat.c +++ b/fpu/softfloat.c @@ -554,6 +554,10 @@ static const FloatFmt float16_params_ahp = { .arm_althp = true }; +static const FloatFmt bfloat16_params = { + FLOAT_PARAMS(8, 7) +}; + static const FloatFmt float32_params = { FLOAT_PARAMS(8, 23) }; @@ -580,6 +584,11 @@ static inline FloatParts float16_unpack_raw(float16 f) return unpack_raw(float16_params, f); } +static inline FloatParts bfloat16_unpack_raw(bfloat16 f) +{ + return unpack_raw(bfloat16_params, f); +} + static inline FloatParts float32_unpack_raw(float32 f) { return unpack_raw(float32_params, f); @@ -603,6 +612,11 @@ static inline float16 float16_pack_raw(FloatParts p) return make_float16(pack_raw(float16_params, p)); } +static inline bfloat16 bfloat16_pack_raw(FloatParts p) +{ + return pack_raw(bfloat16_params, p); +} + static inline float32 float32_pack_raw(FloatParts p) { return make_float32(pack_raw(float32_params, p)); @@ -820,6 +834,11 @@ static FloatParts float16_unpack_canonical(float16 f, float_status *s) return float16a_unpack_canonical(f, s, &float16_params); } +static FloatParts bfloat16_unpack_canonical(bfloat16 f, float_status *s) +{ + return sf_canonicalize(bfloat16_unpack_raw(f), &bfloat16_params, s); +} + static float16 float16a_round_pack_canonical(FloatParts p, float_status *s, const FloatFmt *params) { @@ -831,6 +850,11 @@ static float16 float16_round_pack_canonical(FloatParts p, float_status *s) return float16a_round_pack_canonical(p, s, &float16_params); } +static bfloat16 bfloat16_round_pack_canonical(FloatParts p, float_status *s) +{ + return bfloat16_pack_raw(round_canonical(p, s, &bfloat16_params)); +} + static FloatParts float32_unpack_canonical(float32 f, float_status *s) { return sf_canonicalize(float32_unpack_raw(f), &float32_params, s); @@ -1159,6 +1183,28 @@ float64_sub(float64 a, float64 b, float_status *s) } /* + * Returns the result of adding or subtracting the bfloat16 + * values `a' and `b'. + */ +bfloat16 QEMU_FLATTEN bfloat16_add(bfloat16 a, bfloat16 b, float_status *status) +{ + FloatParts pa = bfloat16_unpack_canonical(a, status); + FloatParts pb = bfloat16_unpack_canonical(b, status); + FloatParts pr = addsub_floats(pa, pb, false, status); + + return bfloat16_round_pack_canonical(pr, status); +} + +bfloat16 QEMU_FLATTEN bfloat16_sub(bfloat16 a, bfloat16 b, float_status *status) +{ + FloatParts pa = bfloat16_unpack_canonical(a, status); + FloatParts pb = bfloat16_unpack_canonical(b, status); + FloatParts pr = addsub_floats(pa, pb, true, status); + + return bfloat16_round_pack_canonical(pr, status); +} + +/* * Returns the result of multiplying the floating-point values `a' and * `b'. The operation is performed according to the IEC/IEEE Standard * for Binary Floating-Point Arithmetic. @@ -1261,6 +1307,20 @@ float64_mul(float64 a, float64 b, float_status *s) } /* + * Returns the result of multiplying the bfloat16 + * values `a' and `b'. + */ + +bfloat16 QEMU_FLATTEN bfloat16_mul(bfloat16 a, bfloat16 b, float_status *status) +{ + FloatParts pa = bfloat16_unpack_canonical(a, status); + FloatParts pb = bfloat16_unpack_canonical(b, status); + FloatParts pr = mul_floats(pa, pb, status); + + return bfloat16_round_pack_canonical(pr, status); +} + +/* * Returns the result of multiplying the floating-point values `a' and * `b' then adding 'c', with no intermediate rounding step after the * multiplication. The operation is performed according to the @@ -1643,6 +1703,23 @@ float64_muladd(float64 xa, float64 xb, float64 xc, int flags, float_status *s) } /* + * Returns the result of multiplying the bfloat16 values `a' + * and `b' then adding 'c', with no intermediate rounding step after the + * multiplication. + */ + +bfloat16 QEMU_FLATTEN bfloat16_muladd(bfloat16 a, bfloat16 b, bfloat16 c, + int flags, float_status *status) +{ + FloatParts pa = bfloat16_unpack_canonical(a, status); + FloatParts pb = bfloat16_unpack_canonical(b, status); + FloatParts pc = bfloat16_unpack_canonical(c, status); + FloatParts pr = muladd_floats(pa, pb, pc, flags, status); + + return bfloat16_round_pack_canonical(pr, status); +} + +/* * Returns the result of dividing the floating-point value `a' by the * corresponding value `b'. The operation is performed according to * the IEC/IEEE Standard for Binary Floating-Point Arithmetic. @@ -1809,6 +1886,20 @@ float64_div(float64 a, float64 b, float_status *s) } /* + * Returns the result of dividing the bfloat16 + * value `a' by the corresponding value `b'. + */ + +bfloat16 bfloat16_div(bfloat16 a, bfloat16 b, float_status *status) +{ + FloatParts pa = bfloat16_unpack_canonical(a, status); + FloatParts pb = bfloat16_unpack_canonical(b, status); + FloatParts pr = div_floats(pa, pb, status); + + return bfloat16_round_pack_canonical(pr, status); +} + +/* * Float to Float conversions * * Returns the result of converting one float format to another. The @@ -2881,6 +2972,25 @@ MINMAX(64, maxnummag, false, true, true) #undef MINMAX +#define BF16_MINMAX(name, ismin, isiee, ismag) \ +bfloat16 bfloat16_ ## name(bfloat16 a, bfloat16 b, float_status *s) \ +{ \ + FloatParts pa = bfloat16_unpack_canonical(a, s); \ + FloatParts pb = bfloat16_unpack_canonical(b, s); \ + FloatParts pr = minmax_floats(pa, pb, ismin, isiee, ismag, s); \ + \ + return bfloat16_round_pack_canonical(pr, s); \ +} + +BF16_MINMAX(min, true, false, false) +BF16_MINMAX(minnum, true, true, false) +BF16_MINMAX(minnummag, true, true, true) +BF16_MINMAX(max, false, false, false) +BF16_MINMAX(maxnum, false, true, false) +BF16_MINMAX(maxnummag, false, true, true) + +#undef BF16_MINMAX + /* Floating point compare */ static FloatRelation compare_floats(FloatParts a, FloatParts b, bool is_quiet, float_status *s) @@ -3042,6 +3152,24 @@ FloatRelation float64_compare_quiet(float64 a, float64 b, float_status *s) return f64_compare(a, b, true, s); } +static FloatRelation QEMU_FLATTEN +soft_bf16_compare(bfloat16 a, bfloat16 b, bool is_quiet, float_status *s) +{ + FloatParts pa = bfloat16_unpack_canonical(a, s); + FloatParts pb = bfloat16_unpack_canonical(b, s); + return compare_floats(pa, pb, is_quiet, s); +} + +FloatRelation bfloat16_compare(bfloat16 a, bfloat16 b, float_status *s) +{ + return soft_bf16_compare(a, b, false, s); +} + +FloatRelation bfloat16_compare_quiet(bfloat16 a, bfloat16 b, float_status *s) +{ + return soft_bf16_compare(a, b, true, s); +} + /* Multiply A by 2 raised to the power N. */ static FloatParts scalbn_decomposed(FloatParts a, int n, float_status *s) { @@ -3081,6 +3209,13 @@ float64 float64_scalbn(float64 a, int n, float_status *status) return float64_round_pack_canonical(pr, status); } +bfloat16 bfloat16_scalbn(bfloat16 a, int n, float_status *status) +{ + FloatParts pa = bfloat16_unpack_canonical(a, status); + FloatParts pr = scalbn_decomposed(pa, n, status); + return bfloat16_round_pack_canonical(pr, status); +} + /* * Square Root * @@ -3231,6 +3366,13 @@ float64 QEMU_FLATTEN float64_sqrt(float64 xa, float_status *s) return soft_f64_sqrt(ua.s, s); } +bfloat16 QEMU_FLATTEN bfloat16_sqrt(bfloat16 a, float_status *status) +{ + FloatParts pa = bfloat16_unpack_canonical(a, status); + FloatParts pr = sqrt_float(pa, status, &bfloat16_params); + return bfloat16_round_pack_canonical(pr, status); +} + /*---------------------------------------------------------------------------- | The pattern for a default generated NaN. *----------------------------------------------------------------------------*/ @@ -3273,6 +3415,13 @@ float128 float128_default_nan(float_status *status) return r; } +bfloat16 bfloat16_default_nan(float_status *status) +{ + FloatParts p = parts_default_nan(status); + p.frac >>= bfloat16_params.frac_shift; + return bfloat16_pack_raw(p); +} + /*---------------------------------------------------------------------------- | Returns a quiet NaN from a signalling NaN for the floating point value `a'. *----------------------------------------------------------------------------*/ @@ -3304,6 +3453,14 @@ float64 float64_silence_nan(float64 a, float_status *status) return float64_pack_raw(p); } +bfloat16 bfloat16_silence_nan(bfloat16 a, float_status *status) +{ + FloatParts p = bfloat16_unpack_raw(a); + p.frac <<= bfloat16_params.frac_shift; + p = parts_silence_nan(p, status); + p.frac >>= bfloat16_params.frac_shift; + return bfloat16_pack_raw(p); +} /*---------------------------------------------------------------------------- | If `a' is denormal and we are in flush-to-zero mode then set the @@ -3353,6 +3510,17 @@ float64 float64_squash_input_denormal(float64 a, float_status *status) return a; } +bfloat16 bfloat16_squash_input_denormal(bfloat16 a, float_status *status) +{ + if (status->flush_inputs_to_zero) { + FloatParts p = bfloat16_unpack_raw(a); + if (parts_squash_denormal(p, status)) { + return bfloat16_set_sign(bfloat16_zero, p.sign); + } + } + return a; +} + /*---------------------------------------------------------------------------- | Takes a 64-bit fixed-point value `absZ' with binary point between bits 6 | and 7, and returns the properly rounded 32-bit integer corresponding to the |