diff --git a/target/hexagon/fma_emu.c b/target/hexagon/fma_emu.c new file mode 100644 index 0000000000..842d903710 --- /dev/null +++ b/target/hexagon/fma_emu.c @@ -0,0 +1,702 @@ +/* + * Copyright(c) 2019-2021 Qualcomm Innovation Center, Inc. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation; either version 2 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, see . + */ + +#include "qemu/osdep.h" +#include "qemu/int128.h" +#include "fpu/softfloat.h" +#include "macros.h" +#include "conv_emu.h" +#include "fma_emu.h" + +#define DF_INF_EXP 0x7ff +#define DF_BIAS 1023 +#define DF_MANTBITS 52 +#define DF_NAN 0xffffffffffffffffULL +#define DF_INF 0x7ff0000000000000ULL +#define DF_MINUS_INF 0xfff0000000000000ULL +#define DF_MAXF 0x7fefffffffffffffULL +#define DF_MINUS_MAXF 0xffefffffffffffffULL + +#define SF_INF_EXP 0xff +#define SF_BIAS 127 +#define SF_MANTBITS 23 +#define SF_INF 0x7f800000 +#define SF_MINUS_INF 0xff800000 +#define SF_MAXF 0x7f7fffff +#define SF_MINUS_MAXF 0xff7fffff + +#define HF_INF_EXP 0x1f +#define HF_BIAS 15 + +#define WAY_BIG_EXP 4096 + +typedef union { + double f; + uint64_t i; + struct { + uint64_t mant:52; + uint64_t exp:11; + uint64_t sign:1; + }; +} Double; + +typedef union { + float f; + uint32_t i; + struct { + uint32_t mant:23; + uint32_t exp:8; + uint32_t sign:1; + }; +} Float; + +static inline uint64_t float64_getmant(float64 f64) +{ + Double a = { .i = f64 }; + if (float64_is_normal(f64)) { + return a.mant | 1ULL << 52; + } + if (float64_is_zero(f64)) { + return 0; + } + if (float64_is_denormal(f64)) { + return a.mant; + } + return ~0ULL; +} + +int32_t float64_getexp(float64 f64) +{ + Double a = { .i = f64 }; + if (float64_is_normal(f64)) { + return a.exp; + } + if (float64_is_denormal(f64)) { + return a.exp + 1; + } + return -1; +} + +static inline uint64_t float32_getmant(float32 f32) +{ + Float a = { .i = f32 }; + if (float32_is_normal(f32)) { + return a.mant | 1ULL << 23; + } + if (float32_is_zero(f32)) { + return 0; + } + if (float32_is_denormal(f32)) { + return a.mant; + } + return ~0ULL; +} + +int32_t float32_getexp(float32 f32) +{ + Float a = { .i = f32 }; + if (float32_is_normal(f32)) { + return a.exp; + } + if (float32_is_denormal(f32)) { + return a.exp + 1; + } + return -1; +} + +static inline uint32_t int128_getw0(Int128 x) +{ + return int128_getlo(x); +} + +static inline uint32_t int128_getw1(Int128 x) +{ + return int128_getlo(x) >> 32; +} + +static inline Int128 int128_mul_6464(uint64_t ai, uint64_t bi) +{ + Int128 a, b; + uint64_t pp0, pp1a, pp1b, pp1s, pp2; + + a = int128_make64(ai); + b = int128_make64(bi); + pp0 = (uint64_t)int128_getw0(a) * (uint64_t)int128_getw0(b); + pp1a = (uint64_t)int128_getw1(a) * (uint64_t)int128_getw0(b); + pp1b = (uint64_t)int128_getw1(b) * (uint64_t)int128_getw0(a); + pp2 = (uint64_t)int128_getw1(a) * (uint64_t)int128_getw1(b); + + pp1s = pp1a + pp1b; + if ((pp1s < pp1a) || (pp1s < pp1b)) { + pp2 += (1ULL << 32); + } + uint64_t ret_low = pp0 + (pp1s << 32); + if ((ret_low < pp0) || (ret_low < (pp1s << 32))) { + pp2 += 1; + } + + return int128_make128(ret_low, pp2 + (pp1s >> 32)); +} + +static inline Int128 int128_sub_borrow(Int128 a, Int128 b, int borrow) +{ + Int128 ret = int128_sub(a, b); + if (borrow != 0) { + ret = int128_sub(ret, int128_one()); + } + return ret; +} + +typedef struct { + Int128 mant; + int32_t exp; + uint8_t sign; + uint8_t guard; + uint8_t round; + uint8_t sticky; +} Accum; + +static inline void accum_init(Accum *p) +{ + p->mant = int128_zero(); + p->exp = 0; + p->sign = 0; + p->guard = 0; + p->round = 0; + p->sticky = 0; +} + +static inline Accum accum_norm_left(Accum a) +{ + a.exp--; + a.mant = int128_lshift(a.mant, 1); + a.mant = int128_or(a.mant, int128_make64(a.guard)); + a.guard = a.round; + a.round = a.sticky; + return a; +} + +static inline Accum accum_norm_right(Accum a, int amt) +{ + if (amt > 130) { + a.sticky |= + a.round | a.guard | int128_nz(a.mant); + a.guard = a.round = 0; + a.mant = int128_zero(); + a.exp += amt; + return a; + + } + while (amt >= 64) { + a.sticky |= a.round | a.guard | (int128_getlo(a.mant) != 0); + a.guard = (int128_getlo(a.mant) >> 63) & 1; + a.round = (int128_getlo(a.mant) >> 62) & 1; + a.mant = int128_make64(int128_gethi(a.mant)); + a.exp += 64; + amt -= 64; + } + while (amt > 0) { + a.exp++; + a.sticky |= a.round; + a.round = a.guard; + a.guard = int128_getlo(a.mant) & 1; + a.mant = int128_rshift(a.mant, 1); + amt--; + } + return a; +} + +/* + * On the add/sub, we need to be able to shift out lots of bits, but need a + * sticky bit for what was shifted out, I think. + */ +static Accum accum_add(Accum a, Accum b); + +static inline Accum accum_sub(Accum a, Accum b, int negate) +{ + Accum ret; + accum_init(&ret); + int borrow; + + if (a.sign != b.sign) { + b.sign = !b.sign; + return accum_add(a, b); + } + if (b.exp > a.exp) { + /* small - big == - (big - small) */ + return accum_sub(b, a, !negate); + } + if ((b.exp == a.exp) && (int128_gt(b.mant, a.mant))) { + /* small - big == - (big - small) */ + return accum_sub(b, a, !negate); + } + + while (a.exp > b.exp) { + /* Try to normalize exponents: shrink a exponent and grow mantissa */ + if (int128_gethi(a.mant) & (1ULL << 62)) { + /* Can't grow a any more */ + break; + } else { + a = accum_norm_left(a); + } + } + + while (a.exp > b.exp) { + /* Try to normalize exponents: grow b exponent and shrink mantissa */ + /* Keep around shifted out bits... we might need those later */ + b = accum_norm_right(b, a.exp - b.exp); + } + + if ((int128_gt(b.mant, a.mant))) { + return accum_sub(b, a, !negate); + } + + /* OK, now things should be normalized! */ + ret.sign = a.sign; + ret.exp = a.exp; + assert(!int128_gt(b.mant, a.mant)); + borrow = (b.round << 2) | (b.guard << 1) | b.sticky; + ret.mant = int128_sub_borrow(a.mant, b.mant, (borrow != 0)); + borrow = 0 - borrow; + ret.guard = (borrow >> 2) & 1; + ret.round = (borrow >> 1) & 1; + ret.sticky = (borrow >> 0) & 1; + if (negate) { + ret.sign = !ret.sign; + } + return ret; +} + +static Accum accum_add(Accum a, Accum b) +{ + Accum ret; + accum_init(&ret); + if (a.sign != b.sign) { + b.sign = !b.sign; + return accum_sub(a, b, 0); + } + if (b.exp > a.exp) { + /* small + big == (big + small) */ + return accum_add(b, a); + } + if ((b.exp == a.exp) && int128_gt(b.mant, a.mant)) { + /* small + big == (big + small) */ + return accum_add(b, a); + } + + while (a.exp > b.exp) { + /* Try to normalize exponents: shrink a exponent and grow mantissa */ + if (int128_gethi(a.mant) & (1ULL << 62)) { + /* Can't grow a any more */ + break; + } else { + a = accum_norm_left(a); + } + } + + while (a.exp > b.exp) { + /* Try to normalize exponents: grow b exponent and shrink mantissa */ + /* Keep around shifted out bits... we might need those later */ + b = accum_norm_right(b, a.exp - b.exp); + } + + /* OK, now things should be normalized! */ + if (int128_gt(b.mant, a.mant)) { + return accum_add(b, a); + }; + ret.sign = a.sign; + ret.exp = a.exp; + assert(!int128_gt(b.mant, a.mant)); + ret.mant = int128_add(a.mant, b.mant); + ret.guard = b.guard; + ret.round = b.round; + ret.sticky = b.sticky; + return ret; +} + +/* Return an infinity with requested sign */ +static inline float64 infinite_float64(uint8_t sign) +{ + if (sign) { + return make_float64(DF_MINUS_INF); + } else { + return make_float64(DF_INF); + } +} + +/* Return a maximum finite value with requested sign */ +static inline float64 maxfinite_float64(uint8_t sign) +{ + if (sign) { + return make_float64(DF_MINUS_MAXF); + } else { + return make_float64(DF_MAXF); + } +} + +/* Return a zero value with requested sign */ +static inline float64 zero_float64(uint8_t sign) +{ + if (sign) { + return make_float64(0x8000000000000000); + } else { + return float64_zero; + } +} + +/* Return an infinity with the requested sign */ +float32 infinite_float32(uint8_t sign) +{ + if (sign) { + return make_float32(SF_MINUS_INF); + } else { + return make_float32(SF_INF); + } +} + +/* Return a maximum finite value with the requested sign */ +static inline float32 maxfinite_float32(uint8_t sign) +{ + if (sign) { + return make_float32(SF_MINUS_MAXF); + } else { + return make_float32(SF_MAXF); + } +} + +/* Return a zero value with requested sign */ +static inline float32 zero_float32(uint8_t sign) +{ + if (sign) { + return make_float32(0x80000000); + } else { + return float32_zero; + } +} + +#define GEN_XF_ROUND(SUFFIX, MANTBITS, INF_EXP, INTERNAL_TYPE) \ +static inline SUFFIX accum_round_##SUFFIX(Accum a, float_status * fp_status) \ +{ \ + if ((int128_gethi(a.mant) == 0) && (int128_getlo(a.mant) == 0) \ + && ((a.guard | a.round | a.sticky) == 0)) { \ + /* result zero */ \ + switch (fp_status->float_rounding_mode) { \ + case float_round_down: \ + return zero_##SUFFIX(1); \ + default: \ + return zero_##SUFFIX(0); \ + } \ + } \ + /* Normalize right */ \ + /* We want MANTBITS bits of mantissa plus the leading one. */ \ + /* That means that we want MANTBITS+1 bits, or 0x000000000000FF_FFFF */ \ + /* So we need to normalize right while the high word is non-zero and \ + * while the low word is nonzero when masked with 0xffe0_0000_0000_0000 */ \ + while ((int128_gethi(a.mant) != 0) || \ + ((int128_getlo(a.mant) >> (MANTBITS + 1)) != 0)) { \ + a = accum_norm_right(a, 1); \ + } \ + /* \ + * OK, now normalize left \ + * We want to normalize left until we have a leading one in bit 24 \ + * Theoretically, we only need to shift a maximum of one to the left if we \ + * shifted out lots of bits from B, or if we had no shift / 1 shift sticky \ + * shoudl be 0 \ + */ \ + while ((int128_getlo(a.mant) & (1ULL << MANTBITS)) == 0) { \ + a = accum_norm_left(a); \ + } \ + /* \ + * OK, now we might need to denormalize because of potential underflow. \ + * We need to do this before rounding, and rounding might make us normal \ + * again \ + */ \ + while (a.exp <= 0) { \ + a = accum_norm_right(a, 1 - a.exp); \ + /* \ + * Do we have underflow? \ + * That's when we get an inexact answer because we ran out of bits \ + * in a denormal. \ + */ \ + if (a.guard || a.round || a.sticky) { \ + float_raise(float_flag_underflow, fp_status); \ + } \ + } \ + /* OK, we're relatively canonical... now we need to round */ \ + if (a.guard || a.round || a.sticky) { \ + float_raise(float_flag_inexact, fp_status); \ + switch (fp_status->float_rounding_mode) { \ + case float_round_to_zero: \ + /* Chop and we're done */ \ + break; \ + case float_round_up: \ + if (a.sign == 0) { \ + a.mant = int128_add(a.mant, int128_one()); \ + } \ + break; \ + case float_round_down: \ + if (a.sign != 0) { \ + a.mant = int128_add(a.mant, int128_one()); \ + } \ + break; \ + default: \ + if (a.round || a.sticky) { \ + /* round up if guard is 1, down if guard is zero */ \ + a.mant = int128_add(a.mant, int128_make64(a.guard)); \ + } else if (a.guard) { \ + /* exactly .5, round up if odd */ \ + a.mant = int128_add(a.mant, int128_and(a.mant, int128_one())); \ + } \ + break; \ + } \ + } \ + /* \ + * OK, now we might have carried all the way up. \ + * So we might need to shr once \ + * at least we know that the lsb should be zero if we rounded and \ + * got a carry out... \ + */ \ + if ((int128_getlo(a.mant) >> (MANTBITS + 1)) != 0) { \ + a = accum_norm_right(a, 1); \ + } \ + /* Overflow? */ \ + if (a.exp >= INF_EXP) { \ + /* Yep, inf result */ \ + float_raise(float_flag_overflow, fp_status); \ + float_raise(float_flag_inexact, fp_status); \ + switch (fp_status->float_rounding_mode) { \ + case float_round_to_zero: \ + return maxfinite_##SUFFIX(a.sign); \ + case float_round_up: \ + if (a.sign == 0) { \ + return infinite_##SUFFIX(a.sign); \ + } else { \ + return maxfinite_##SUFFIX(a.sign); \ + } \ + case float_round_down: \ + if (a.sign != 0) { \ + return infinite_##SUFFIX(a.sign); \ + } else { \ + return maxfinite_##SUFFIX(a.sign); \ + } \ + default: \ + return infinite_##SUFFIX(a.sign); \ + } \ + } \ + /* Underflow? */ \ + if (int128_getlo(a.mant) & (1ULL << MANTBITS)) { \ + /* Leading one means: No, we're normal. So, we should be done... */ \ + INTERNAL_TYPE ret; \ + ret.i = 0; \ + ret.sign = a.sign; \ + ret.exp = a.exp; \ + ret.mant = int128_getlo(a.mant); \ + return ret.i; \ + } \ + assert(a.exp == 1); \ + INTERNAL_TYPE ret; \ + ret.i = 0; \ + ret.sign = a.sign; \ + ret.exp = 0; \ + ret.mant = int128_getlo(a.mant); \ + return ret.i; \ +} + +GEN_XF_ROUND(float64, DF_MANTBITS, DF_INF_EXP, Double) +GEN_XF_ROUND(float32, SF_MANTBITS, SF_INF_EXP, Float) + +static bool is_inf_prod(float64 a, float64 b) +{ + return ((float64_is_infinity(a) && float64_is_infinity(b)) || + (float64_is_infinity(a) && is_finite(b) && (!float64_is_zero(b))) || + (float64_is_infinity(b) && is_finite(a) && (!float64_is_zero(a)))); +} + +static inline float64 special_fma(float64 a, float64 b, float64 c, + float_status *fp_status) +{ + float64 ret = make_float64(0); + + /* + * If A multiplied by B is an exact infinity and C is also an infinity + * but with the opposite sign, FMA returns NaN and raises invalid. + */ + uint8_t a_sign = float64_is_neg(a); + uint8_t b_sign = float64_is_neg(b); + uint8_t c_sign = float64_is_neg(c); + if (is_inf_prod(a, b) && float64_is_infinity(c)) { + if ((a_sign ^ b_sign) != c_sign) { + ret = make_float64(DF_NAN); + float_raise(float_flag_invalid, fp_status); + return ret; + } + } + if ((float64_is_infinity(a) && float64_is_zero(b)) || + (float64_is_zero(a) && float64_is_infinity(b))) { + ret = make_float64(DF_NAN); + float_raise(float_flag_invalid, fp_status); + return ret; + } + /* + * If none of the above checks are true and C is a NaN, + * a NaN shall be returned + * If A or B are NaN, a NAN shall be returned. + */ + if (float64_is_any_nan(a) || + float64_is_any_nan(b) || + float64_is_any_nan(c)) { + if (float64_is_any_nan(a) && (fGETBIT(51, a) == 0)) { + float_raise(float_flag_invalid, fp_status); + } + if (float64_is_any_nan(b) && (fGETBIT(51, b) == 0)) { + float_raise(float_flag_invalid, fp_status); + } + if (float64_is_any_nan(c) && (fGETBIT(51, c) == 0)) { + float_raise(float_flag_invalid, fp_status); + } + ret = make_float64(DF_NAN); + return ret; + } + /* + * We have checked for adding opposite-signed infinities. + * Other infinities return infinity with the correct sign + */ + if (float64_is_infinity(c)) { + ret = infinite_float64(c_sign); + return ret; + } + if (float64_is_infinity(a) || float64_is_infinity(b)) { + ret = infinite_float64(a_sign ^ b_sign); + return ret; + } + g_assert_not_reached(); +} + +static inline float32 special_fmaf(float32 a, float32 b, float32 c, + float_status *fp_status) +{ + float64 aa, bb, cc; + aa = float32_to_float64(a, fp_status); + bb = float32_to_float64(b, fp_status); + cc = float32_to_float64(c, fp_status); + return float64_to_float32(special_fma(aa, bb, cc, fp_status), fp_status); +} + +float32 internal_fmafx(float32 a, float32 b, float32 c, int scale, + float_status *fp_status) +{ + Accum prod; + Accum acc; + Accum result; + accum_init(&prod); + accum_init(&acc); + accum_init(&result); + + uint8_t a_sign = float32_is_neg(a); + uint8_t b_sign = float32_is_neg(b); + uint8_t c_sign = float32_is_neg(c); + if (float32_is_infinity(a) || + float32_is_infinity(b) || + float32_is_infinity(c)) { + return special_fmaf(a, b, c, fp_status); + } + if (float32_is_any_nan(a) || + float32_is_any_nan(b) || + float32_is_any_nan(c)) { + return special_fmaf(a, b, c, fp_status); + } + if ((scale == 0) && (float32_is_zero(a) || float32_is_zero(b))) { + float32 tmp = float32_mul(a, b, fp_status); + tmp = float32_add(tmp, c, fp_status); + return tmp; + } + + /* (a * 2**b) * (c * 2**d) == a*c * 2**(b+d) */ + prod.mant = int128_mul_6464(float32_getmant(a), float32_getmant(b)); + + /* + * Note: extracting the mantissa into an int is multiplying by + * 2**23, so adjust here + */ + prod.exp = float32_getexp(a) + float32_getexp(b) - SF_BIAS - 23; + prod.sign = a_sign ^ b_sign; + if (float32_is_zero(a) || float32_is_zero(b)) { + prod.exp = -2 * WAY_BIG_EXP; + } + if ((scale > 0) && float32_is_denormal(c)) { + acc.mant = int128_mul_6464(0, 0); + acc.exp = -WAY_BIG_EXP; + acc.sign = c_sign; + acc.sticky = 1; + result = accum_add(prod, acc); + } else if (!float32_is_zero(c)) { + acc.mant = int128_mul_6464(float32_getmant(c), 1); + acc.exp = float32_getexp(c); + acc.sign = c_sign; + result = accum_add(prod, acc); + } else { + result = prod; + } + result.exp += scale; + return accum_round_float32(result, fp_status); +} + +float32 internal_mpyf(float32 a, float32 b, float_status *fp_status) +{ + if (float32_is_zero(a) || float32_is_zero(b)) { + return float32_mul(a, b, fp_status); + } + return internal_fmafx(a, b, float32_zero, 0, fp_status); +} + +float64 internal_mpyhh(float64 a, float64 b, + unsigned long long int accumulated, + float_status *fp_status) +{ + Accum x; + unsigned long long int prod; + unsigned int sticky; + uint8_t a_sign, b_sign; + + sticky = accumulated & 1; + accumulated >>= 1; + accum_init(&x); + if (float64_is_zero(a) || + float64_is_any_nan(a) || + float64_is_infinity(a)) { + return float64_mul(a, b, fp_status); + } + if (float64_is_zero(b) || + float64_is_any_nan(b) || + float64_is_infinity(b)) { + return float64_mul(a, b, fp_status); + } + x.mant = int128_mul_6464(accumulated, 1); + x.sticky = sticky; + prod = fGETUWORD(1, float64_getmant(a)) * fGETUWORD(1, float64_getmant(b)); + x.mant = int128_add(x.mant, int128_mul_6464(prod, 0x100000000ULL)); + x.exp = float64_getexp(a) + float64_getexp(b) - DF_BIAS - 20; + if (!float64_is_normal(a) || !float64_is_normal(b)) { + /* crush to inexact zero */ + x.sticky = 1; + x.exp = -4096; + } + a_sign = float64_is_neg(a); + b_sign = float64_is_neg(b); + x.sign = a_sign ^ b_sign; + return accum_round_float64(x, fp_status); +} diff --git a/target/hexagon/fma_emu.h b/target/hexagon/fma_emu.h new file mode 100644 index 0000000000..e3b99a8cf4 --- /dev/null +++ b/target/hexagon/fma_emu.h @@ -0,0 +1,36 @@ +/* + * Copyright(c) 2019-2021 Qualcomm Innovation Center, Inc. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation; either version 2 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, see . + */ + +#ifndef HEXAGON_FMA_EMU_H +#define HEXAGON_FMA_EMU_H + +static inline bool is_finite(float64 x) +{ + return !float64_is_any_nan(x) && !float64_is_infinity(x); +} + +int32_t float64_getexp(float64 f64); +int32_t float32_getexp(float32 f32); +float32 infinite_float32(uint8_t sign); +float32 internal_fmafx(float32 a, float32 b, float32 c, + int scale, float_status *fp_status); +float32 internal_mpyf(float32 a, float32 b, float_status *fp_status); +float64 internal_mpyhh(float64 a, float64 b, + unsigned long long int accumulated, + float_status *fp_status); + +#endif