fpu: Add conversion routines for OCP FP8 E5M2
Reviewed-by: Chao Liu <chao.liu.zevorn@gmail.com> Signed-off-by: Max Chou <max.chou@sifive.com> [rth: Split out of a larger patch] Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Max Chou committed
Feb 5, 2026 at 17:35 UTC
d8be49537658faef90722ca5fcd606cd09e91a7b
3 files changed
+98
fpu/softfloat.c
+85
@@ -572,6 +572,10 @@ typedef struct {
572
.frac_shift = (-F - 1) & 63, \
573
.round_mask = (1ull << ((-F - 1) & 63)) - 1
574
575
+static const FloatFmt float8_e5m2_params = {
576
+ FLOAT_PARAMS(5, 2)
577
+};
578
+
579
static const FloatFmt float16_params = {
580
FLOAT_PARAMS(5, 10)
581
};
@@ -627,6 +631,11 @@ static void unpack_raw64(FloatParts64 *r, const FloatFmt *fmt, uint64_t raw)
631
};
632
}
633
634
+static void QEMU_FLATTEN float8_e5m2_unpack_raw(FloatParts64 *p, float8_e5m2 f)
635
+{
636
+ unpack_raw64(p, &float8_e5m2_params, f);
637
+}
638
+
639
static void QEMU_FLATTEN float16_unpack_raw(FloatParts64 *p, float16 f)
640
{
641
unpack_raw64(p, &float16_params, f);
@@ -684,6 +693,11 @@ static uint64_t pack_raw64(const FloatParts64 *p, const FloatFmt *fmt)
693
return ret;
694
}
695
696
+static float8_e5m2 QEMU_FLATTEN float8_e5m2_pack_raw(const FloatParts64 *p)
697
+{
698
+ return pack_raw64(p, &float8_e5m2_params);
699
+}
700
+
701
static float16 QEMU_FLATTEN float16_pack_raw(const FloatParts64 *p)
702
{
703
return make_float16(pack_raw64(p, &float16_params));
@@ -1675,6 +1689,13 @@ static const uint16_t rsqrt_tab[128] = {
1689
* Pack/unpack routines with a specific FloatFmt.
1690
*/
1691
1692
+static void float8_e5m2_unpack_canonical(FloatParts64 *p, float8_e5m2 f,
1693
+ float_status *s)
1694
+{
1695
+ float8_e5m2_unpack_raw(p, f);
1696
+ parts_canonicalize(p, s, &float8_e5m2_params);
1697
+}
1698
+
1699
static void float16a_unpack_canonical(FloatParts64 *p, float16 f,
1700
float_status *s, const FloatFmt *params)
1701
{
@@ -1695,6 +1716,14 @@ static void bfloat16_unpack_canonical(FloatParts64 *p, bfloat16 f,
1716
parts_canonicalize(p, s, &bfloat16_params);
1717
}
1718
1719
+static float8_e5m2 float8_e5m2_round_pack_canonical(FloatParts64 *p,
1720
+ float_status *s,
1721
+ bool saturate)
1722
+{
1723
+ parts_uncanon(p, s, &float8_e5m2_params, saturate);
1724
+ return float8_e5m2_pack_raw(p);
1725
+}
1726
+
1727
static float16 float16a_round_pack_canonical(FloatParts64 *p,
1728
float_status *s,
1729
const FloatFmt *params)
@@ -2772,6 +2801,35 @@ static void parts_float_to_ahp(FloatParts64 *a, float_status *s)
2801
}
2802
}
2803
2804
+static void parts_float_to_e5m2(FloatParts64 *a, float_status *s, bool saturate)
2805
+{
2806
+ switch (a->cls) {
2807
+ case float_class_snan:
2808
+ case float_class_qnan:
2809
+ parts_return_nan(a, s);
2810
+ break;
2811
+
2812
+ case float_class_inf:
2813
+ /* Per OCP, conversion in SATURATE mode bounds Inf to MAX. */
2814
+ if (saturate) {
2815
+ a->cls = float_class_normal;
2816
+ a->exp = float8_e5m2_params.exp_max - 1;
2817
+ a->frac = MAKE_64BIT_MASK(float8_e5m2_params.frac_shift,
2818
+ float8_e5m2_params.frac_size + 1);
2819
+ }
2820
+ break;
2821
+
2822
+ case float_class_denormal:
2823
+ float_raise(float_flag_input_denormal_used, s);
2824
+ break;
2825
+ case float_class_normal:
2826
+ case float_class_zero:
2827
+ break;
2828
+ default:
2829
+ g_assert_not_reached();
2830
+ }
2831
+}
2832
+
2833
static void parts64_float_to_float(FloatParts64 *a, float_status *s)
2834
{
2835
if (is_nan(a->cls)) {
@@ -2836,6 +2894,15 @@ static void parts_float_to_float_widen(FloatParts128 *a, FloatParts64 *b,
2894
}
2895
}
2896
2897
+bfloat16 float8_e5m2_to_bfloat16(float8_e5m2 a, float_status *s)
2898
+{
2899
+ FloatParts64 p;
2900
+
2901
+ float8_e5m2_unpack_canonical(&p, a, s);
2902
+ parts_float_to_float(&p, s);
2903
+ return bfloat16_round_pack_canonical(&p, s);
2904
+}
2905
+
2906
float32 float16_to_float32(float16 a, bool ieee, float_status *s)
2907
{
2908
const FloatFmt *fmt16 = ieee ? &float16_params : &float16_params_ahp;
@@ -2856,6 +2923,15 @@ float64 float16_to_float64(float16 a, bool ieee, float_status *s)
2923
return float64_round_pack_canonical(&p, s);
2924
}
2925
2926
+float8_e5m2 float32_to_float8_e5m2(float32 a, bool saturate, float_status *s)
2927
+{
2928
+ FloatParts64 p;
2929
+
2930
+ float32_unpack_canonical(&p, a, s);
2931
+ parts_float_to_e5m2(&p, s, saturate);
2932
+ return float8_e5m2_round_pack_canonical(&p, s, saturate);
2933
+}
2934
+
2935
float16 float32_to_float16(float32 a, bool ieee, float_status *s)
2936
{
2937
FloatParts64 p;
@@ -2923,6 +2999,15 @@ float32 float64_to_float32(float64 a, float_status *s)
2999
return float32_round_pack_canonical(&p, s);
3000
}
3001
3002
+float8_e5m2 bfloat16_to_float8_e5m2(bfloat16 a, bool saturate, float_status *s)
3003
+{
3004
+ FloatParts64 p;
3005
+
3006
+ bfloat16_unpack_canonical(&p, a, s);
3007
+ parts_float_to_e5m2(&p, s, saturate);
3008
+ return float8_e5m2_round_pack_canonical(&p, s, saturate);
3009
+}
3010
+
3011
float32 bfloat16_to_float32(bfloat16 a, float_status *s)
3012
{
3013
FloatParts64 p;
include/fpu/softfloat-types.h
+5
@@ -119,6 +119,11 @@ typedef struct {
119
*/
120
typedef uint16_t bfloat16;
121
122
+/*
123
+ * Open Compute Project (OCP) Microscaling Formats
124
+ */
125
+typedef uint8_t float8_e5m2;
126
+
127
/*
128
* Software IEC/IEEE floating-point underflow tininess-detection mode.
129
*/
include/fpu/softfloat.h
+8
@@ -189,6 +189,14 @@ float128 int128_to_float128(Int128, float_status *status);
189
float128 uint64_to_float128(uint64_t, float_status *status);
190
float128 uint128_to_float128(Int128, float_status *status);
191
192
+/*----------------------------------------------------------------------------
193
+| OCP FP8 conversion routines.
194
+*----------------------------------------------------------------------------*/
195
+
196
+bfloat16 float8_e5m2_to_bfloat16(float8_e5m2, float_status *status);
197
+float8_e5m2 bfloat16_to_float8_e5m2(bfloat16, bool sat, float_status *status);
198
+float8_e5m2 float32_to_float8_e5m2(float32, bool sat, float_status *status);
199
+
200
/*----------------------------------------------------------------------------
201
| Software half-precision conversion routines.
202
*----------------------------------------------------------------------------*/