From bf211c34993921eccbc074f82cfbb8e9a16d850c Mon Sep 17 00:00:00 2001 From: Adhemerval Zanella Date: Thu, 13 Nov 2025 09:58:20 -0300 Subject: [PATCH] math: New generic fma implementation MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The current implementation relies on setting the rounding mode for different calculations (first to FE_TONEAREST and then to FE_TOWARDZERO) to obtain correctly rounded results. For most CPUs, this adds a significant performance overhead since it requires executing a typically slow instruction (to get/set the floating-point status), it necessitates flushing the pipeline, and breaks some compiler assumptions/optimizations. This patch introduces a new implementation originally written by Szabolcs for musl, which utilizes mostly integer arithmetic. Floating-point arithmetic is used to raise the expected exceptions, without the need for fenv.h operations. I added some changes compared to the original code: * Fixed some signaling NaN issues when the 3-argument is NaN. * Use math_uint128.h for the 64-bit multiplication operation. It allows the compiler to use 128-bit types where available, which enables some optimizations on certain targets (for instance, MIPS64). * Fixed an arm32 issue where the libgcc routine might not respect the rounding mode [1]. This can also be used on other targets to optimize the conversion from int64_t to double. * Use -fexcess-precision=standard on i686. I tested this implementation on various targets (x86_64, i686, arm, aarch64, powerpc), including some by manually disabling the compiler instructions. Performance-wise, it shows large improvements: reciprocal-throughput master patched improvement x86_64 [2] 289.4640 22.4396 12.90x i686 [2] 636.8660 169.3640 3.76x aarch64 [3] 46.0020 11.3281 4.06x armhf [3] 63.989 26.5056 2.41x powerpc [4] 23.9332 6.40205 3.74x latency master patched improvement x86_64 293.7360 38.1478 7.70x i686 658.4160 187.9940 3.50x aarch64 44.5166 14.7157 3.03x armhf 63.7678 28.4116 2.24x power10 23.8561 11.4250 2.09x Checked on x86_64-linux-gnu and i686-linux-gnu with —disable-multi-arch, and on arm-linux-gnueabihf. [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91970 [2] gcc 15.2.1, Zen3 [3] gcc 15.2.1, Neoverse N1 [4] gcc 15.2.1, POWER10 Signed-off-by: Szabolcs Nagy Co-authored-by: Adhemerval Zanella Reviewed-by: Wilco Dijkstra  --- sysdeps/arm/fpu/math_private.h | 32 ++ sysdeps/i386/Makefile | 1 + sysdeps/ieee754/dbl-64/math_config.h | 18 ++ sysdeps/ieee754/dbl-64/s_fma.c | 450 ++++++++++++--------------- 4 files changed, 243 insertions(+), 258 deletions(-) create mode 100644 sysdeps/arm/fpu/math_private.h diff --git a/sysdeps/arm/fpu/math_private.h b/sysdeps/arm/fpu/math_private.h new file mode 100644 index 0000000000..b66ce65c76 --- /dev/null +++ b/sysdeps/arm/fpu/math_private.h @@ -0,0 +1,32 @@ +/* Configure optimized libm functions. AArch64 version. + Copyright (C) 2017-2025 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library 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 + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + . */ + +#ifndef ARM_MATH_PRIVATE_H +#define ARM_MATH_PRIVATE_H 1 + +#include + +/* For int64_t to double conversion, libgcc might not respect the rounding + mode [1]. + + [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91970 */ +#define TOINT64_INTRINSICS 0 + +#include_next + +#endif diff --git a/sysdeps/i386/Makefile b/sysdeps/i386/Makefile index bb4f59094c..11ddbd402d 100644 --- a/sysdeps/i386/Makefile +++ b/sysdeps/i386/Makefile @@ -13,6 +13,7 @@ CFLAGS-e_gamma_r.c += -fexcess-precision=standard CFLAGS-s_erf.c += -fexcess-precision=standard CFLAGS-s_erfc.c += -fexcess-precision=standard CFLAGS-s_erf_common.c += -fexcess-precision=standard +CFLAGS-s_fma.c += -fexcess-precision=standard endif ifeq ($(subdir),gmon) diff --git a/sysdeps/ieee754/dbl-64/math_config.h b/sysdeps/ieee754/dbl-64/math_config.h index 36a47ae7db..6a7b98e1f0 100644 --- a/sysdeps/ieee754/dbl-64/math_config.h +++ b/sysdeps/ieee754/dbl-64/math_config.h @@ -85,6 +85,24 @@ static inline int32_t converttoint (double x); #endif +#ifndef TOINT64_INTRINSICS +# define TOINT64_INTRINSICS 1 +#endif + +static inline double convertfromint64 (int64_t a) +{ +#if !TOINT64_INTRINSICS + union { int64_t x; double d; } low = {.d = 0x1.0p52}; + + double high = (int32_t)(a >> 32) * 0x1.0p32; + low.x |= a & INT64_C(0x00000000ffffffff); + + return (high - 0x1.0p52) + low.d; +#else + return a; +#endif +} + static inline uint64_t asuint64 (double f) { diff --git a/sysdeps/ieee754/dbl-64/s_fma.c b/sysdeps/ieee754/dbl-64/s_fma.c index c7ff3cf447..f89b3bb0cc 100644 --- a/sysdeps/ieee754/dbl-64/s_fma.c +++ b/sysdeps/ieee754/dbl-64/s_fma.c @@ -23,17 +23,51 @@ #include #undef dfmal #undef f32xfmaf64 -#include -#include -#include -#include #include #include -#include +#include -/* This implementation uses rounding to odd to avoid problems with - double rounding. See a paper by Boldo and Melquiond: - http://www.lri.fr/~melquion/doc/08-tc.pdf */ + +#if !USE_FMA_BUILTIN +# include +# include "math_config.h" +# include + +# define ZEROINFNAN (0x7ff - EXPONENT_BIAS - MANTISSA_WIDTH - 1) + +struct num +{ + uint64_t m; + int e; + int sign; +}; + +static inline struct num normalize (double x) +{ + uint64_t ix = asuint64 (x); + int e = ix >> MANTISSA_WIDTH; + int sign = e & 0x800; + e &= 0x7ff; + if (!e) + { + ix = asuint64 (x * 0x1p63); + e = ix >> MANTISSA_WIDTH & 0x7ff; + e = e ? e-63 : 0x800; + } + ix &= (UINT64_C(1) << MANTISSA_WIDTH) - 1; + ix |= UINT64_C(1) << MANTISSA_WIDTH; + ix <<= 1; + e -= EXPONENT_BIAS + MANTISSA_WIDTH + 1; + return (struct num){ix,e,sign}; +} + +static void mul (uint64_t *hi, uint64_t *lo, uint64_t x, uint64_t y) +{ + u128 r = u128_mul (u128_from_u64 (x), u128_from_u64 (y)); + *hi = u128_high (r); + *lo = u128_low (r); +} +#endif double __fma (double x, double y, double z) @@ -41,271 +75,171 @@ __fma (double x, double y, double z) #if USE_FMA_BUILTIN return __builtin_fma (x, y, z); #else - /* Use generic implementation. */ - union ieee754_double u, v, w; - int adjust = 0; - u.d = x; - v.d = y; - w.d = z; - if (__builtin_expect (u.ieee.exponent + v.ieee.exponent - >= 0x7ff + IEEE754_DOUBLE_BIAS - DBL_MANT_DIG, 0) - || __builtin_expect (u.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0) - || __builtin_expect (v.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0) - || __builtin_expect (w.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0) - || __builtin_expect (u.ieee.exponent + v.ieee.exponent - <= IEEE754_DOUBLE_BIAS + DBL_MANT_DIG, 0)) + /* Normalize so top 10bits and last bit are 0 */ + struct num nx, ny, nz; + nx = normalize (x); + ny = normalize (y); + nz = normalize (z); + + if (nx.e >= ZEROINFNAN || ny.e >= ZEROINFNAN) + return x * y + z; + if (nz.e >= ZEROINFNAN) { - /* If z is Inf, but x and y are finite, the result should be - z rather than NaN. */ - if (w.ieee.exponent == 0x7ff - && u.ieee.exponent != 0x7ff - && v.ieee.exponent != 0x7ff) - return (z + x) + y; - /* If z is zero and x are y are nonzero, compute the result - as x * y to avoid the wrong sign of a zero result if x * y - underflows to 0. */ - if (z == 0 && x != 0 && y != 0) + if (nz.e > ZEROINFNAN) /* z==0 */ return x * y; - /* If x or y or z is Inf/NaN, or if x * y is zero, compute as - x * y + z. */ - if (u.ieee.exponent == 0x7ff - || v.ieee.exponent == 0x7ff - || w.ieee.exponent == 0x7ff - || x == 0 - || y == 0) - return x * y + z; - /* If fma will certainly overflow, compute as x * y. */ - if (u.ieee.exponent + v.ieee.exponent > 0x7ff + IEEE754_DOUBLE_BIAS) - return x * y; - /* If x * y is less than 1/4 of DBL_TRUE_MIN, neither the - result nor whether there is underflow depends on its exact - value, only on its sign. */ - if (u.ieee.exponent + v.ieee.exponent - < IEEE754_DOUBLE_BIAS - DBL_MANT_DIG - 2) + else if (isnan (z)) + return __builtin_nan (""); + return z; + } + + /* mul: r = x*y */ + uint64_t rhi, rlo, zhi, zlo; + mul (&rhi, &rlo, nx.m, ny.m); + /* Either top 20 or 21 bits of rhi and last 2 bits of rlo are 0 */ + + /* align exponents */ + int e = nx.e + ny.e; + int d = nz.e - e; + /* Shift bits z<<=kz, r>>=kr, so kz+kr == d, set e = e+kr (== ez-kz). */ + if (d > 0) + { + if (d < 64) { - int neg = u.ieee.negative ^ v.ieee.negative; - double tiny = neg ? -0x1p-1074 : 0x1p-1074; - if (w.ieee.exponent >= 3) - return tiny + z; - /* Scaling up, adding TINY and scaling down produces the - correct result, because in round-to-nearest mode adding - TINY has no effect and in other modes double rounding is - harmless. But it may not produce required underflow - exceptions. */ - v.d = z * 0x1p54 + tiny; - if (TININESS_AFTER_ROUNDING - ? v.ieee.exponent < 55 - : (w.ieee.exponent == 0 - || (w.ieee.exponent == 1 - && w.ieee.negative != neg - && w.ieee.mantissa1 == 0 - && w.ieee.mantissa0 == 0))) + zlo = nz.m << d; + zhi = nz.m >> (64 - d); + } + else + { + zlo = 0; + zhi = nz.m; + e = nz.e - 64; + d -= 64; + if (d < 64) { - double force_underflow = x * y; - math_force_eval (force_underflow); + rlo = rhi << (64 - d) | rlo >> d | !!(rlo << (64 - d)); + rhi = rhi >> d; } - return v.d * 0x1p-54; - } - if (u.ieee.exponent + v.ieee.exponent - >= 0x7ff + IEEE754_DOUBLE_BIAS - DBL_MANT_DIG) - { - /* Compute 1p-53 times smaller result and multiply - at the end. */ - if (u.ieee.exponent > v.ieee.exponent) - u.ieee.exponent -= DBL_MANT_DIG; else - v.ieee.exponent -= DBL_MANT_DIG; - /* If x + y exponent is very large and z exponent is very small, - it doesn't matter if we don't adjust it. */ - if (w.ieee.exponent > DBL_MANT_DIG) - w.ieee.exponent -= DBL_MANT_DIG; - adjust = 1; - } - else if (w.ieee.exponent >= 0x7ff - DBL_MANT_DIG) - { - /* Similarly. - If z exponent is very large and x and y exponents are - very small, adjust them up to avoid spurious underflows, - rather than down. */ - if (u.ieee.exponent + v.ieee.exponent - <= IEEE754_DOUBLE_BIAS + 2 * DBL_MANT_DIG) { - if (u.ieee.exponent > v.ieee.exponent) - u.ieee.exponent += 2 * DBL_MANT_DIG + 2; - else - v.ieee.exponent += 2 * DBL_MANT_DIG + 2; + rlo = 1; + rhi = 0; } - else if (u.ieee.exponent > v.ieee.exponent) - { - if (u.ieee.exponent > DBL_MANT_DIG) - u.ieee.exponent -= DBL_MANT_DIG; - } - else if (v.ieee.exponent > DBL_MANT_DIG) - v.ieee.exponent -= DBL_MANT_DIG; - w.ieee.exponent -= DBL_MANT_DIG; - adjust = 1; } - else if (u.ieee.exponent >= 0x7ff - DBL_MANT_DIG) - { - u.ieee.exponent -= DBL_MANT_DIG; - if (v.ieee.exponent) - v.ieee.exponent += DBL_MANT_DIG; - else - v.d *= 0x1p53; - } - else if (v.ieee.exponent >= 0x7ff - DBL_MANT_DIG) - { - v.ieee.exponent -= DBL_MANT_DIG; - if (u.ieee.exponent) - u.ieee.exponent += DBL_MANT_DIG; - else - u.d *= 0x1p53; - } - else /* if (u.ieee.exponent + v.ieee.exponent - <= IEEE754_DOUBLE_BIAS + DBL_MANT_DIG) */ - { - if (u.ieee.exponent > v.ieee.exponent) - u.ieee.exponent += 2 * DBL_MANT_DIG + 2; - else - v.ieee.exponent += 2 * DBL_MANT_DIG + 2; - if (w.ieee.exponent <= 4 * DBL_MANT_DIG + 6) - { - if (w.ieee.exponent) - w.ieee.exponent += 2 * DBL_MANT_DIG + 2; - else - w.d *= 0x1p108; - adjust = -1; - } - /* Otherwise x * y should just affect inexact - and nothing else. */ - } - x = u.d; - y = v.d; - z = w.d; - } - - /* Ensure correct sign of exact 0 + 0. */ - if (__glibc_unlikely ((x == 0 || y == 0) && z == 0)) - { - x = math_opt_barrier (x); - return x * y + z; - } - - fenv_t env; - libc_feholdexcept_setround (&env, FE_TONEAREST); - - /* Multiplication m1 + m2 = x * y using Dekker's algorithm. */ -#define C ((1 << (DBL_MANT_DIG + 1) / 2) + 1) - double x1 = x * C; - double y1 = y * C; - double m1 = x * y; - x1 = (x - x1) + x1; - y1 = (y - y1) + y1; - double x2 = x - x1; - double y2 = y - y1; - double m2 = (((x1 * y1 - m1) + x1 * y2) + x2 * y1) + x2 * y2; - - /* Addition a1 + a2 = z + m1 using Knuth's algorithm. */ - double a1 = z + m1; - double t1 = a1 - z; - double t2 = a1 - t1; - t1 = m1 - t1; - t2 = z - t2; - double a2 = t1 + t2; - /* Ensure the arithmetic is not scheduled after feclearexcept call. */ - math_force_eval (m2); - math_force_eval (a2); - __feclearexcept (FE_INEXACT); - - /* If the result is an exact zero, ensure it has the correct sign. */ - if (a1 == 0 && m2 == 0) - { - libc_feupdateenv (&env); - /* Ensure that round-to-nearest value of z + m1 is not reused. */ - z = math_opt_barrier (z); - return z + m1; - } - - libc_fesetround (FE_TOWARDZERO); - - /* Perform m2 + a2 addition with round to odd. */ - u.d = a2 + m2; - - if (__glibc_unlikely (adjust < 0)) - { - if ((u.ieee.mantissa1 & 1) == 0) - u.ieee.mantissa1 |= libc_fetestexcept (FE_INEXACT) != 0; - v.d = a1 + u.d; - /* Ensure the addition is not scheduled after fetestexcept call. */ - math_force_eval (v.d); - } - - /* Reset rounding mode and test for inexact simultaneously. */ - int j = libc_feupdateenv_test (&env, FE_INEXACT) != 0; - - /* Ensure value of a1 + u.d is not reused. */ - a1 = math_opt_barrier (a1); - - if (__glibc_likely (adjust == 0)) - { - if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7ff) - u.ieee.mantissa1 |= j; - /* Result is a1 + u.d. */ - return a1 + u.d; - } - else if (__glibc_likely (adjust > 0)) - { - if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7ff) - u.ieee.mantissa1 |= j; - /* Result is a1 + u.d, scaled up. */ - return (a1 + u.d) * 0x1p53; } else { - /* If a1 + u.d is exact, the only rounding happens during - scaling down. */ - if (j == 0) - return v.d * 0x1p-108; - /* If result rounded to zero is not subnormal, no double - rounding will occur. */ - if (v.ieee.exponent > 108) - return (a1 + u.d) * 0x1p-108; - /* If v.d * 0x1p-108 with round to zero is a subnormal above - or equal to DBL_MIN / 2, then v.d * 0x1p-108 shifts mantissa - down just by 1 bit, which means v.ieee.mantissa1 |= j would - change the round bit, not sticky or guard bit. - v.d * 0x1p-108 never normalizes by shifting up, - so round bit plus sticky bit should be already enough - for proper rounding. */ - if (v.ieee.exponent == 108) - { - /* If the exponent would be in the normal range when - rounding to normal precision with unbounded exponent - range, the exact result is known and spurious underflows - must be avoided on systems detecting tininess after - rounding. */ - if (TININESS_AFTER_ROUNDING) - { - w.d = a1 + u.d; - if (w.ieee.exponent == 109) - return w.d * 0x1p-108; - } - /* v.ieee.mantissa1 & 2 is LSB bit of the result before rounding, - v.ieee.mantissa1 & 1 is the round bit and j is our sticky - bit. */ - w.d = 0.0; - w.ieee.mantissa1 = ((v.ieee.mantissa1 & 3) << 1) | j; - w.ieee.negative = v.ieee.negative; - v.ieee.mantissa1 &= ~3U; - v.d *= 0x1p-108; - w.d *= 0x1p-2; - return v.d + w.d; - } - v.ieee.mantissa1 |= j; - return v.d * 0x1p-108; + zhi = 0; + d = -d; + if (d == 0) + zlo = nz.m; + else if (d < 64) + zlo = nz.m >> d | !!(nz.m << (64 - d)); + else + zlo = 1; } + + /* add */ + int sign = nx.sign ^ ny.sign; + bool samesign = !(sign ^ nz.sign); + bool nonzero = true; + if (samesign) + { + /* r += z */ + rlo += zlo; + rhi += zhi + (rlo < zlo); + } + else + { + /* r -= z */ + uint64_t t = rlo; + rlo -= zlo; + rhi = rhi - zhi - (t < rlo); + if (rhi >> 63) + { + rlo = -rlo; + rhi = -rhi - !!rlo; + sign = !sign; + } + nonzero = !!rhi; + } + + /* Set rhi to top 63bit of the result (last bit is sticky). */ + if (nonzero) + { + e += 64; + d = stdc_leading_zeros (rhi) - 1; + /* note: d > 0 */ + rhi = rhi << d | rlo >> (64 - d) | !!(rlo << d); + } + else if (rlo) + { + d = stdc_leading_zeros (rlo) - 1; + if (d < 0) + rhi = rlo >> 1 | (rlo & 1); + else + rhi = rlo << d; + } + else + { + /* Exact +-0 */ + return x * y + z; + } + e -= d; + + /* Convert to double. */ + int64_t i = rhi; /* i is in [1<<62,(1<<63)-1] */ + if (sign) + i = -i; + double r = convertfromint64 (i); /* |r| is in [0x1p62,0x1p63] */ + + if (e < -1022 - 62) + { + /* Result is subnormal before rounding. */ + if (e == -1022 - 63) + { + double c = 0x1p63; + if (sign) + c = -c; + if (r == c) + { + /* Min normal after rounding, underflow depends + on arch behaviour which can be imitated by + a double to float conversion. */ + float fltmin = 0x0.ffffff8p-63 * FLT_MIN * r; + return DBL_MIN / FLT_MIN * fltmin; + } + /* One bit is lost when scaled, add another top bit to + only round once at conversion if it is inexact. */ + if (rhi << 53) + { + i = rhi >> 1 | (rhi & 1) | 1ull << 62; + if (sign) + i = -i; + r = convertfromint64 (i); + r = 2 * r - c; /* remove top bit */ + + /* Raise underflow portably, such that it + cannot be optimized away. */ + { + double_t tiny = DBL_MIN / FLT_MIN * r; + r += (double) (tiny * tiny) * (r - r); + } + } + } + else + { + /* Only round once when scaled. */ + d = 10; + i = (rhi >> d | !!(rhi << (64 - d))) << d; + if (sign) + i = -i; + r = convertfromint64 (i); + } + } + return __scalbn (r, e); #endif /* ! USE_FMA_BUILTIN */ } + #ifndef __fma libm_alias_double (__fma, fma) libm_alias_double_narrow (__fma, fma)