diff options
author | Szabolcs Nagy <szabolcs.nagy@arm.com> | 2018-06-26 16:25:12 +0100 |
---|---|---|
committer | Corinna Vinschen <corinna@vinschen.de> | 2018-06-27 15:40:49 +0200 |
commit | b99d49e506f595c4be91072fe0169a405ae16007 (patch) | |
tree | 57ff07b7cd9110b22e64526ea5dc93ae734a0be5 | |
parent | 07e2c328285e63c78d23ded96fd333afad35961d (diff) | |
download | newlib-b99d49e506f595c4be91072fe0169a405ae16007.zip newlib-b99d49e506f595c4be91072fe0169a405ae16007.tar.gz newlib-b99d49e506f595c4be91072fe0169a405ae16007.tar.bz2 |
New pow implementation
The new implementation is provided under !__OBSOLETE_MATH, it uses
ISO C99 code. With default settings the worst case error in nearest
rounding mode is 0.54 ULP with inlined fma and fma contraction. It uses
a 4 KB lookup table in addition to the table in exp_data.c, on aarch64
.text+.rodata size of libm.a is increased by 2295 bytes.
Improvements on Cortex-A72:
latency: 3.3x
thruput: 4.9x
-rw-r--r-- | newlib/libm/common/Makefile.am | 3 | ||||
-rw-r--r-- | newlib/libm/common/Makefile.in | 20 | ||||
-rw-r--r-- | newlib/libm/common/fdlibm.h | 1 | ||||
-rw-r--r-- | newlib/libm/common/math_config.h | 19 | ||||
-rw-r--r-- | newlib/libm/common/pow.c | 377 | ||||
-rw-r--r-- | newlib/libm/common/pow_log_data.c | 187 | ||||
-rw-r--r-- | newlib/libm/math/e_pow.c | 2 | ||||
-rw-r--r-- | newlib/libm/math/w_pow.c | 3 |
8 files changed, 607 insertions, 5 deletions
diff --git a/newlib/libm/common/Makefile.am b/newlib/libm/common/Makefile.am index 7fb3cc2..b085671 100644 --- a/newlib/libm/common/Makefile.am +++ b/newlib/libm/common/Makefile.am @@ -13,7 +13,8 @@ src = s_finite.c s_copysign.c s_modf.c s_scalbn.c \ s_lrint.c s_llrint.c \ s_lround.c s_llround.c s_nearbyint.c s_remquo.c s_round.c s_scalbln.c \ s_signbit.c s_trunc.c \ - exp.c exp2.c exp_data.c math_err.c log.c log_data.c log2.c log2_data.c + exp.c exp2.c exp_data.c math_err.c log.c log_data.c log2.c log2_data.c \ + pow.c pow_log_data.c fsrc = sf_finite.c sf_copysign.c sf_modf.c sf_scalbn.c \ sf_cbrt.c sf_exp10.c sf_expm1.c sf_ilogb.c \ diff --git a/newlib/libm/common/Makefile.in b/newlib/libm/common/Makefile.in index 97dddbd..a777a34 100644 --- a/newlib/libm/common/Makefile.in +++ b/newlib/libm/common/Makefile.in @@ -97,7 +97,8 @@ am__objects_1 = lib_a-s_finite.$(OBJEXT) lib_a-s_copysign.$(OBJEXT) \ lib_a-exp2.$(OBJEXT) lib_a-exp_data.$(OBJEXT) \ lib_a-math_err.$(OBJEXT) lib_a-log.$(OBJEXT) \ lib_a-log_data.$(OBJEXT) lib_a-log2.$(OBJEXT) \ - lib_a-log2_data.$(OBJEXT) + lib_a-log2_data.$(OBJEXT) lib_a-pow.$(OBJEXT) \ + lib_a-pow_log_data.$(OBJEXT) am__objects_2 = lib_a-sf_finite.$(OBJEXT) lib_a-sf_copysign.$(OBJEXT) \ lib_a-sf_modf.$(OBJEXT) lib_a-sf_scalbn.$(OBJEXT) \ lib_a-sf_cbrt.$(OBJEXT) lib_a-sf_exp10.$(OBJEXT) \ @@ -168,7 +169,7 @@ am__objects_5 = s_finite.lo s_copysign.lo s_modf.lo s_scalbn.lo \ s_lround.lo s_llround.lo s_nearbyint.lo s_remquo.lo s_round.lo \ s_scalbln.lo s_signbit.lo s_trunc.lo exp.lo exp2.lo \ exp_data.lo math_err.lo log.lo log_data.lo log2.lo \ - log2_data.lo + log2_data.lo pow.lo pow_log_data.lo am__objects_6 = sf_finite.lo sf_copysign.lo sf_modf.lo sf_scalbn.lo \ sf_cbrt.lo sf_exp10.lo sf_expm1.lo sf_ilogb.lo sf_infinity.lo \ sf_isinf.lo sf_isinff.lo sf_isnan.lo sf_isnanf.lo sf_log1p.lo \ @@ -359,7 +360,8 @@ src = s_finite.c s_copysign.c s_modf.c s_scalbn.c \ s_lrint.c s_llrint.c \ s_lround.c s_llround.c s_nearbyint.c s_remquo.c s_round.c s_scalbln.c \ s_signbit.c s_trunc.c \ - exp.c exp2.c exp_data.c math_err.c log.c log_data.c log2.c log2_data.c + exp.c exp2.c exp_data.c math_err.c log.c log_data.c log2.c log2_data.c \ + pow.c pow_log_data.c fsrc = sf_finite.c sf_copysign.c sf_modf.c sf_scalbn.c \ sf_cbrt.c sf_exp10.c sf_expm1.c sf_ilogb.c \ @@ -751,6 +753,18 @@ lib_a-log2_data.o: log2_data.c lib_a-log2_data.obj: log2_data.c $(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) $(AM_CPPFLAGS) $(CPPFLAGS) $(lib_a_CFLAGS) $(CFLAGS) -c -o lib_a-log2_data.obj `if test -f 'log2_data.c'; then $(CYGPATH_W) 'log2_data.c'; else $(CYGPATH_W) '$(srcdir)/log2_data.c'; fi` +lib_a-pow.o: pow.c + $(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) $(AM_CPPFLAGS) $(CPPFLAGS) $(lib_a_CFLAGS) $(CFLAGS) -c -o lib_a-pow.o `test -f 'pow.c' || echo '$(srcdir)/'`pow.c + +lib_a-pow.obj: pow.c + $(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) $(AM_CPPFLAGS) $(CPPFLAGS) $(lib_a_CFLAGS) $(CFLAGS) -c -o lib_a-pow.obj `if test -f 'pow.c'; then $(CYGPATH_W) 'pow.c'; else $(CYGPATH_W) '$(srcdir)/pow.c'; fi` + +lib_a-pow_log_data.o: pow_log_data.c + $(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) $(AM_CPPFLAGS) $(CPPFLAGS) $(lib_a_CFLAGS) $(CFLAGS) -c -o lib_a-pow_log_data.o `test -f 'pow_log_data.c' || echo '$(srcdir)/'`pow_log_data.c + +lib_a-pow_log_data.obj: pow_log_data.c + $(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) $(AM_CPPFLAGS) $(CPPFLAGS) $(lib_a_CFLAGS) $(CFLAGS) -c -o lib_a-pow_log_data.obj `if test -f 'pow_log_data.c'; then $(CYGPATH_W) 'pow_log_data.c'; else $(CYGPATH_W) '$(srcdir)/pow_log_data.c'; fi` + lib_a-sf_finite.o: sf_finite.c $(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) $(AM_CPPFLAGS) $(CPPFLAGS) $(lib_a_CFLAGS) $(CFLAGS) -c -o lib_a-sf_finite.o `test -f 'sf_finite.c' || echo '$(srcdir)/'`sf_finite.c diff --git a/newlib/libm/common/fdlibm.h b/newlib/libm/common/fdlibm.h index ae85569..2155f51 100644 --- a/newlib/libm/common/fdlibm.h +++ b/newlib/libm/common/fdlibm.h @@ -233,6 +233,7 @@ extern float __ieee754_scalbf __P((float,float)); and there are no observable error handling side effects. */ # define __ieee754_exp(x) exp(x) # define __ieee754_log(x) log(x) +# define __ieee754_pow(x,y) pow(x,y) # define __ieee754_expf(x) expf(x) # define __ieee754_logf(x) logf(x) # define __ieee754_powf(x,y) powf(x,y) diff --git a/newlib/libm/common/math_config.h b/newlib/libm/common/math_config.h index 32439f7..3a8cebb 100644 --- a/newlib/libm/common/math_config.h +++ b/newlib/libm/common/math_config.h @@ -147,6 +147,15 @@ issignalingf_inline (float x) return 2 * (ix ^ 0x00400000) > 2u * 0x7fc00000; } +static inline int +issignaling_inline (double x) +{ + uint64_t ix = asuint64 (x); + if (!IEEE_754_2008_SNAN) + return (ix & 0x7ff8000000000000) == 0x7ff8000000000000; + return 2 * (ix ^ 0x0008000000000000) > 2 * 0x7ff8000000000000ULL; +} + #if __aarch64__ && __GNUC__ /* Prevent the optimization of a floating-point expression. */ static inline float @@ -354,4 +363,14 @@ extern const struct log2_data { #endif } __log2_data HIDDEN; +#define POW_LOG_TABLE_BITS 7 +#define POW_LOG_POLY_ORDER 8 +extern const struct pow_log_data { + double ln2hi; + double ln2lo; + double poly[POW_LOG_POLY_ORDER - 1]; /* First coefficient is 1. */ + /* Note: the pad field is unused, but allows slightly faster indexing. */ + struct {double invc, pad, logc, logctail;} tab[1 << POW_LOG_TABLE_BITS]; +} __pow_log_data HIDDEN; + #endif diff --git a/newlib/libm/common/pow.c b/newlib/libm/common/pow.c new file mode 100644 index 0000000..da7081c --- /dev/null +++ b/newlib/libm/common/pow.c @@ -0,0 +1,377 @@ +/* Double-precision x^y function. + Copyright (c) 2018 Arm Ltd. All rights reserved. + + SPDX-License-Identifier: BSD-3-Clause + + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions + are met: + 1. Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + 2. Redistributions in binary form must reproduce the above copyright + notice, this list of conditions and the following disclaimer in the + documentation and/or other materials provided with the distribution. + 3. The name of the company may not be used to endorse or promote + products derived from this software without specific prior written + permission. + + THIS SOFTWARE IS PROVIDED BY Arm LTD ``AS IS AND ANY EXPRESS OR IMPLIED + WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF + MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + IN NO EVENT SHALL Arm LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED + TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR + PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF + LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING + NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ + +#include "fdlibm.h" +#if !__OBSOLETE_MATH + +#include <math.h> +#include <stdint.h> +#include "math_config.h" + +/* +Worst-case error: 0.54 ULP (~= ulperr_exp + 1024*Ln2*relerr_log*2^53) +relerr_log: 1.3 * 2^-68 (Relative error of log, 1.5 * 2^-68 without fma) +ulperr_exp: 0.509 ULP (ULP error of exp, 0.511 ULP without fma) +*/ + +#define T __pow_log_data.tab +#define A __pow_log_data.poly +#define Ln2hi __pow_log_data.ln2hi +#define Ln2lo __pow_log_data.ln2lo +#define N (1 << POW_LOG_TABLE_BITS) +#define OFF 0x3fe6955500000000 + +static inline uint32_t +top12 (double x) +{ + return asuint64 (x) >> 52; +} + +static inline double_t +log_inline (uint64_t ix, double_t *tail) +{ + /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */ + double_t z, r, y, invc, logc, logctail, kd, hi, t1, t2, lo, lo1, lo2, p; + uint64_t iz, tmp; + int k, i; + + /* x = 2^k z; where z is in range [OFF,2*OFF) and exact. + The range is split into N subintervals. + The ith subinterval contains z and c is near its center. */ + tmp = ix - OFF; + i = (tmp >> (52 - POW_LOG_TABLE_BITS)) % N; + k = (int64_t) tmp >> 52; /* arithmetic shift */ + iz = ix - (tmp & 0xfffULL << 52); + z = asdouble (iz); + kd = (double_t) k; + + /* log(x) = k*Ln2 + log(c) + log1p(z/c-1). */ + invc = T[i].invc; + logc = T[i].logc; + logctail = T[i].logctail; + + /* r = z/c - 1, arranged to be exact. */ +#if __HAVE_FAST_FMA + r = fma (z, invc, -1.0); +#else + double_t zhi = asdouble (iz & (-1ULL << 32)); + double_t zlo = z - zhi; + double_t rhi = zhi * invc - 1.0; + double_t rlo = zlo * invc; + r = rhi + rlo; +#endif + + /* k*Ln2 + log(c) + r. */ + t1 = kd * Ln2hi + logc; + t2 = t1 + r; + lo1 = kd * Ln2lo + logctail; + lo2 = t1 - t2 + r; + + /* Evaluation is optimized assuming superscalar pipelined execution. */ + double_t ar, ar2, ar3, lo3, lo4; + ar = A[0] * r; /* A[0] = -0.5. */ + ar2 = r * ar; + ar3 = r * ar2; + /* k*Ln2 + log(c) + r + A[0]*r*r. */ +#if __HAVE_FAST_FMA + hi = t2 + ar2; + lo3 = fma (ar, r, -ar2); + lo4 = t2 - hi + ar2; +#else + double_t arhi = A[0] * rhi; + double_t arhi2 = rhi * arhi; + hi = t2 + arhi2; + lo3 = rlo * (ar + arhi); + lo4 = t2 - hi + arhi2; +#endif + /* p = log1p(r) - r - A[0]*r*r. */ +#if POW_LOG_POLY_ORDER == 8 + p = ar3*(A[1] + r*A[2] + ar2*(A[3] + r*A[4] + ar2*(A[5] + r*A[6]))); +#endif + lo = lo1 + lo2 + lo3 + lo4 + p; + y = hi + lo; + *tail = hi - y + lo; + return y; +} + +#undef N +#undef T +#define N (1 << EXP_TABLE_BITS) +#define InvLn2N __exp_data.invln2N +#define NegLn2hiN __exp_data.negln2hiN +#define NegLn2loN __exp_data.negln2loN +#define Shift __exp_data.shift +#define T __exp_data.tab +#define C2 __exp_data.poly[5 - EXP_POLY_ORDER] +#define C3 __exp_data.poly[6 - EXP_POLY_ORDER] +#define C4 __exp_data.poly[7 - EXP_POLY_ORDER] +#define C5 __exp_data.poly[8 - EXP_POLY_ORDER] +#define C6 __exp_data.poly[9 - EXP_POLY_ORDER] + +static inline double +specialcase (double_t tmp, uint64_t sbits, uint64_t ki) +{ + double_t scale, y; + + if ((ki & 0x80000000) == 0) + { + /* k > 0, the exponent of scale might have overflowed by <= 460. */ + sbits -= 1009ull << 52; + scale = asdouble (sbits); + y = 0x1p1009 * (scale + scale * tmp); + return check_oflow (y); + } + /* k < 0, need special care in the subnormal range. */ + sbits += 1022ull << 52; + /* Note: sbits is signed scale. */ + scale = asdouble (sbits); + y = scale + scale * tmp; + if (fabs (y) < 1.0) + { + /* Round y to the right precision before scaling it into the subnormal + range to avoid double rounding that can cause 0.5+E/2 ulp error where + E is the worst-case ulp error outside the subnormal range. So this + is only useful if the goal is better than 1 ulp worst-case error. */ + double_t hi, lo, one = 1.0; + if (y < 0.0) + one = -1.0; + lo = scale - y + scale * tmp; + hi = one + y; + lo = one - hi + y + lo; + y = eval_as_double (hi + lo) - one; + /* Fix the sign of 0. */ + if (y == 0.0) + y = asdouble (sbits & 0x8000000000000000); + /* The underflow exception needs to be signaled explicitly. */ + force_eval_double (opt_barrier_double (0x1p-1022) * 0x1p-1022); + } + y = 0x1p-1022 * y; + return check_uflow (y); +} + +#define SIGN_BIAS (0x800 << EXP_TABLE_BITS) + +static inline double +exp_inline (double x, double xtail, uint32_t sign_bias) +{ + uint32_t abstop; + uint64_t ki, idx, top, sbits; + /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */ + double_t kd, z, r, r2, scale, tail, tmp; + + abstop = top12 (x) & 0x7ff; + if (unlikely (abstop - top12 (0x1p-54) >= top12 (512.0) - top12 (0x1p-54))) + { + if (abstop - top12 (0x1p-54) >= 0x80000000) + { + /* Avoid spurious underflow for tiny x. */ + /* Note: 0 is common input. */ + double_t one = WANT_ROUNDING ? 1.0 + x : 1.0; + return sign_bias ? -one : one; + } + if (abstop >= top12 (1024.0)) + { + /* Note: inf and nan are already handled. */ + if (asuint64 (x) >> 63) + return __math_uflow (sign_bias); + else + return __math_oflow (sign_bias); + } + /* Large x is special cased below. */ + abstop = 0; + } + + /* exp(x) = 2^(k/N) * exp(r), with exp(r) in [2^(-1/2N),2^(1/2N)]. */ + /* x = ln2/N*k + r, with int k and r in [-ln2/2N, ln2/2N]. */ + z = InvLn2N * x; +#if TOINT_INTRINSICS + kd = roundtoint (z); + ki = converttoint (z); +#elif EXP_USE_TOINT_NARROW + /* z - kd is in [-0.5-2^-16, 0.5] in all rounding modes. */ + kd = eval_as_double (z + Shift); + ki = asuint64 (kd) >> 16; + kd = (double_t) (int32_t) ki; +#else + /* z - kd is in [-1, 1] in non-nearest rounding modes. */ + kd = eval_as_double (z + Shift); + ki = asuint64 (kd); + kd -= Shift; +#endif + r = x + kd*NegLn2hiN + kd*NegLn2loN; + /* The code assumes 2^-200 < |xtail| < 2^-8/N. */ + r += xtail; + /* 2^(k/N) ~= scale * (1 + tail). */ + idx = 2*(ki % N); + top = (ki + sign_bias) << (52 - EXP_TABLE_BITS); + tail = asdouble (T[idx]); + /* This is only a valid scale when -1023*N < k < 1024*N. */ + sbits = T[idx + 1] + top; + /* exp(x) = 2^(k/N) * exp(r) ~= scale + scale * (tail + exp(r) - 1). */ + /* Evaluation is optimized assuming superscalar pipelined execution. */ + r2 = r*r; + /* Without fma the worst case error is 0.25/N ulp larger. */ + /* Worst case error is less than 0.5+1.11/N+(abs poly error * 2^53) ulp. */ +#if EXP_POLY_ORDER == 4 + tmp = tail + r + r2*C2 + r*r2*(C3 + r*C4); +#elif EXP_POLY_ORDER == 5 + tmp = tail + r + r2*(C2 + r*C3) + r2*r2*(C4 + r*C5); +#elif EXP_POLY_ORDER == 6 + tmp = tail + r + r2*(0.5 + r*C3) + r2*r2*(C4 + r*C5 + r2*C6); +#endif + if (unlikely (abstop == 0)) + return specialcase (tmp, sbits, ki); + scale = asdouble (sbits); + /* Note: tmp == 0 or |tmp| > 2^-200 and scale > 2^-739, so there + is no spurious underflow here even without fma. */ + return scale + scale * tmp; +} + +/* Returns 0 if not int, 1 if odd int, 2 if even int. */ +static inline int +checkint (uint64_t iy) +{ + int e = iy >> 52 & 0x7ff; + if (e < 0x3ff) + return 0; + if (e > 0x3ff + 52) + return 2; + if (iy & ((1ULL << (0x3ff + 52 - e)) - 1)) + return 0; + if (iy & (1ULL << (0x3ff + 52 - e))) + return 1; + return 2; +} + +static inline int +zeroinfnan (uint64_t i) +{ + return 2 * i - 1 >= 2 * asuint64 (INFINITY) - 1; +} + +double +pow (double x, double y) +{ + uint32_t sign_bias = 0; + uint64_t ix, iy; + uint32_t topx, topy; + + ix = asuint64 (x); + iy = asuint64 (y); + topx = top12 (x); + topy = top12 (y); + if (unlikely (topx - 0x001 >= 0x7ff - 0x001 + || (topy & 0x7ff) - 0x3be >= 0x43e - 0x3be)) + { + /* Note: if |y| > 1075 * ln2 * 2^53 ~= 0x1.749p62 then pow(x,y) = inf/0 + and if |y| < 2^-54 / 1075 ~= 0x1.e7b6p-65 then pow(x,y) = +-1. */ + /* Special cases: (x < 0x1p-126 or inf or nan) or + (|y| < 0x1p-65 or |y| >= 0x1p63 or nan). */ + if (unlikely (zeroinfnan (iy))) + { + if (2 * iy == 0) + return issignaling_inline (x) ? x + y : 1.0; + if (ix == asuint64 (1.0)) + return issignaling_inline (y) ? x + y : 1.0; + if (2 * ix > 2 * asuint64 (INFINITY) + || 2 * iy > 2 * asuint64 (INFINITY)) + return x + y; + if (2 * ix == 2 * asuint64 (1.0)) + return 1.0; + if ((2 * ix < 2 * asuint64 (1.0)) == !(iy >> 63)) + return 0.0; /* |x|<1 && y==inf or |x|>1 && y==-inf. */ + return y * y; + } + if (unlikely (zeroinfnan (ix))) + { + double_t x2 = x * x; + if (ix >> 63 && checkint (iy) == 1) + { + x2 = -x2; + sign_bias = 1; + } + if (WANT_ERRNO && 2 * ix == 0 && iy >> 63) + return __math_divzero (sign_bias); + /* Without the barrier some versions of clang hoist the 1/x2 and + thus division by zero exception can be signaled spuriously. */ + return iy >> 63 ? opt_barrier_double (1 / x2) : x2; + } + /* Here x and y are non-zero finite. */ + if (ix >> 63) + { + /* Finite x < 0. */ + int yint = checkint (iy); + if (yint == 0) + return __math_invalid (x); + if (yint == 1) + sign_bias = SIGN_BIAS; + ix &= 0x7fffffffffffffff; + topx &= 0x7ff; + } + if ((topy & 0x7ff) - 0x3be >= 0x43e - 0x3be) + { + /* Note: sign_bias == 0 here because y is not odd. */ + if (ix == asuint64 (1.0)) + return 1.0; + if ((topy & 0x7ff) < 0x3be) + { + /* |y| < 2^-65, x^y ~= 1 + y*log(x). */ + if (WANT_ROUNDING) + return ix > asuint64 (1.0) ? 1.0 + y : 1.0 - y; + else + return 1.0; + } + return (ix > asuint64 (1.0)) == (topy < 0x800) ? __math_oflow (0) + : __math_uflow (0); + } + if (topx == 0) + { + /* Normalize subnormal x so exponent becomes negative. */ + ix = asuint64 (x * 0x1p52); + ix &= 0x7fffffffffffffff; + ix -= 52ULL << 52; + } + } + + double_t lo; + double_t hi = log_inline (ix, &lo); + double_t ehi, elo; +#if __HAVE_FAST_FMA + ehi = y * hi; + elo = y * lo + fma (y, hi, -ehi); +#else + double_t yhi = asdouble (iy & -1ULL << 27); + double_t ylo = y - yhi; + double_t lhi = asdouble (asuint64 (hi) & -1ULL << 27); + double_t llo = hi - lhi + lo; + ehi = yhi * lhi; + elo = ylo * lhi + y * llo; /* |elo| < |ehi| * 2^-25. */ +#endif + return exp_inline (ehi, elo, sign_bias); +} +#endif diff --git a/newlib/libm/common/pow_log_data.c b/newlib/libm/common/pow_log_data.c new file mode 100644 index 0000000..9b94f5e --- /dev/null +++ b/newlib/libm/common/pow_log_data.c @@ -0,0 +1,187 @@ +/* Data for the log part of pow. + Copyright (c) 2018 Arm Ltd. All rights reserved. + + SPDX-License-Identifier: BSD-3-Clause + + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions + are met: + 1. Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + 2. Redistributions in binary form must reproduce the above copyright + notice, this list of conditions and the following disclaimer in the + documentation and/or other materials provided with the distribution. + 3. The name of the company may not be used to endorse or promote + products derived from this software without specific prior written + permission. + + THIS SOFTWARE IS PROVIDED BY Arm LTD ``AS IS AND ANY EXPRESS OR IMPLIED + WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF + MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + IN NO EVENT SHALL Arm LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED + TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR + PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF + LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING + NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ + +#include "fdlibm.h" +#if !__OBSOLETE_MATH + +#include "math_config.h" + +#define N (1 << POW_LOG_TABLE_BITS) + +const struct pow_log_data __pow_log_data = { +.ln2hi = 0x1.62e42fefa3800p-1, +.ln2lo = 0x1.ef35793c76730p-45, +.poly = { +#if N == 128 && POW_LOG_POLY_ORDER == 8 +// relative error: 0x1.11922ap-70 +// in -0x1.6bp-8 0x1.6bp-8 +// Coefficients are scaled to match the scaling during evaluation. +-0x1p-1, +0x1.555555555556p-2 * -2, +-0x1.0000000000006p-2 * -2, +0x1.999999959554ep-3 * 4, +-0x1.555555529a47ap-3 * 4, +0x1.2495b9b4845e9p-3 * -8, +-0x1.0002b8b263fc3p-3 * -8, +#endif +}, +.tab = { +#if N == 128 +#define A(a,b,c) {a,0,b,c}, +A(0x1.6a00000000000p+0, -0x1.62c82f2b9c800p-2, 0x1.ab42428375680p-48) +A(0x1.6800000000000p+0, -0x1.5d1bdbf580800p-2, -0x1.ca508d8e0f720p-46) +A(0x1.6600000000000p+0, -0x1.5767717455800p-2, -0x1.362a4d5b6506dp-45) +A(0x1.6400000000000p+0, -0x1.51aad872df800p-2, -0x1.684e49eb067d5p-49) +A(0x1.6200000000000p+0, -0x1.4be5f95777800p-2, -0x1.41b6993293ee0p-47) +A(0x1.6000000000000p+0, -0x1.4618bc21c6000p-2, 0x1.3d82f484c84ccp-46) +A(0x1.5e00000000000p+0, -0x1.404308686a800p-2, 0x1.c42f3ed820b3ap-50) +A(0x1.5c00000000000p+0, -0x1.3a64c55694800p-2, 0x1.0b1c686519460p-45) +A(0x1.5a00000000000p+0, -0x1.347dd9a988000p-2, 0x1.5594dd4c58092p-45) +A(0x1.5800000000000p+0, -0x1.2e8e2bae12000p-2, 0x1.67b1e99b72bd8p-45) +A(0x1.5600000000000p+0, -0x1.2895a13de8800p-2, 0x1.5ca14b6cfb03fp-46) +A(0x1.5600000000000p+0, -0x1.2895a13de8800p-2, 0x1.5ca14b6cfb03fp-46) +A(0x1.5400000000000p+0, -0x1.22941fbcf7800p-2, -0x1.65a242853da76p-46) +A(0x1.5200000000000p+0, -0x1.1c898c1699800p-2, -0x1.fafbc68e75404p-46) +A(0x1.5000000000000p+0, -0x1.1675cababa800p-2, 0x1.f1fc63382a8f0p-46) +A(0x1.4e00000000000p+0, -0x1.1058bf9ae4800p-2, -0x1.6a8c4fd055a66p-45) +A(0x1.4c00000000000p+0, -0x1.0a324e2739000p-2, -0x1.c6bee7ef4030ep-47) +A(0x1.4a00000000000p+0, -0x1.0402594b4d000p-2, -0x1.036b89ef42d7fp-48) +A(0x1.4a00000000000p+0, -0x1.0402594b4d000p-2, -0x1.036b89ef42d7fp-48) +A(0x1.4800000000000p+0, -0x1.fb9186d5e4000p-3, 0x1.d572aab993c87p-47) +A(0x1.4600000000000p+0, -0x1.ef0adcbdc6000p-3, 0x1.b26b79c86af24p-45) +A(0x1.4400000000000p+0, -0x1.e27076e2af000p-3, -0x1.72f4f543fff10p-46) +A(0x1.4200000000000p+0, -0x1.d5c216b4fc000p-3, 0x1.1ba91bbca681bp-45) +A(0x1.4000000000000p+0, -0x1.c8ff7c79aa000p-3, 0x1.7794f689f8434p-45) +A(0x1.4000000000000p+0, -0x1.c8ff7c79aa000p-3, 0x1.7794f689f8434p-45) +A(0x1.3e00000000000p+0, -0x1.bc286742d9000p-3, 0x1.94eb0318bb78fp-46) +A(0x1.3c00000000000p+0, -0x1.af3c94e80c000p-3, 0x1.a4e633fcd9066p-52) +A(0x1.3a00000000000p+0, -0x1.a23bc1fe2b000p-3, -0x1.58c64dc46c1eap-45) +A(0x1.3a00000000000p+0, -0x1.a23bc1fe2b000p-3, -0x1.58c64dc46c1eap-45) +A(0x1.3800000000000p+0, -0x1.9525a9cf45000p-3, -0x1.ad1d904c1d4e3p-45) +A(0x1.3600000000000p+0, -0x1.87fa06520d000p-3, 0x1.bbdbf7fdbfa09p-45) +A(0x1.3400000000000p+0, -0x1.7ab890210e000p-3, 0x1.bdb9072534a58p-45) +A(0x1.3400000000000p+0, -0x1.7ab890210e000p-3, 0x1.bdb9072534a58p-45) +A(0x1.3200000000000p+0, -0x1.6d60fe719d000p-3, -0x1.0e46aa3b2e266p-46) +A(0x1.3000000000000p+0, -0x1.5ff3070a79000p-3, -0x1.e9e439f105039p-46) +A(0x1.3000000000000p+0, -0x1.5ff3070a79000p-3, -0x1.e9e439f105039p-46) +A(0x1.2e00000000000p+0, -0x1.526e5e3a1b000p-3, -0x1.0de8b90075b8fp-45) +A(0x1.2c00000000000p+0, -0x1.44d2b6ccb8000p-3, 0x1.70cc16135783cp-46) +A(0x1.2c00000000000p+0, -0x1.44d2b6ccb8000p-3, 0x1.70cc16135783cp-46) +A(0x1.2a00000000000p+0, -0x1.371fc201e9000p-3, 0x1.178864d27543ap-48) +A(0x1.2800000000000p+0, -0x1.29552f81ff000p-3, -0x1.48d301771c408p-45) +A(0x1.2600000000000p+0, -0x1.1b72ad52f6000p-3, -0x1.e80a41811a396p-45) +A(0x1.2600000000000p+0, -0x1.1b72ad52f6000p-3, -0x1.e80a41811a396p-45) +A(0x1.2400000000000p+0, -0x1.0d77e7cd09000p-3, 0x1.a699688e85bf4p-47) +A(0x1.2400000000000p+0, -0x1.0d77e7cd09000p-3, 0x1.a699688e85bf4p-47) +A(0x1.2200000000000p+0, -0x1.fec9131dbe000p-4, -0x1.575545ca333f2p-45) +A(0x1.2000000000000p+0, -0x1.e27076e2b0000p-4, 0x1.a342c2af0003cp-45) +A(0x1.2000000000000p+0, -0x1.e27076e2b0000p-4, 0x1.a342c2af0003cp-45) +A(0x1.1e00000000000p+0, -0x1.c5e548f5bc000p-4, -0x1.d0c57585fbe06p-46) +A(0x1.1c00000000000p+0, -0x1.a926d3a4ae000p-4, 0x1.53935e85baac8p-45) +A(0x1.1c00000000000p+0, -0x1.a926d3a4ae000p-4, 0x1.53935e85baac8p-45) +A(0x1.1a00000000000p+0, -0x1.8c345d631a000p-4, 0x1.37c294d2f5668p-46) +A(0x1.1a00000000000p+0, -0x1.8c345d631a000p-4, 0x1.37c294d2f5668p-46) +A(0x1.1800000000000p+0, -0x1.6f0d28ae56000p-4, -0x1.69737c93373dap-45) +A(0x1.1600000000000p+0, -0x1.51b073f062000p-4, 0x1.f025b61c65e57p-46) +A(0x1.1600000000000p+0, -0x1.51b073f062000p-4, 0x1.f025b61c65e57p-46) +A(0x1.1400000000000p+0, -0x1.341d7961be000p-4, 0x1.c5edaccf913dfp-45) +A(0x1.1400000000000p+0, -0x1.341d7961be000p-4, 0x1.c5edaccf913dfp-45) +A(0x1.1200000000000p+0, -0x1.16536eea38000p-4, 0x1.47c5e768fa309p-46) +A(0x1.1000000000000p+0, -0x1.f0a30c0118000p-5, 0x1.d599e83368e91p-45) +A(0x1.1000000000000p+0, -0x1.f0a30c0118000p-5, 0x1.d599e83368e91p-45) +A(0x1.0e00000000000p+0, -0x1.b42dd71198000p-5, 0x1.c827ae5d6704cp-46) +A(0x1.0e00000000000p+0, -0x1.b42dd71198000p-5, 0x1.c827ae5d6704cp-46) +A(0x1.0c00000000000p+0, -0x1.77458f632c000p-5, -0x1.cfc4634f2a1eep-45) +A(0x1.0c00000000000p+0, -0x1.77458f632c000p-5, -0x1.cfc4634f2a1eep-45) +A(0x1.0a00000000000p+0, -0x1.39e87b9fec000p-5, 0x1.502b7f526feaap-48) +A(0x1.0a00000000000p+0, -0x1.39e87b9fec000p-5, 0x1.502b7f526feaap-48) +A(0x1.0800000000000p+0, -0x1.f829b0e780000p-6, -0x1.980267c7e09e4p-45) +A(0x1.0800000000000p+0, -0x1.f829b0e780000p-6, -0x1.980267c7e09e4p-45) +A(0x1.0600000000000p+0, -0x1.7b91b07d58000p-6, -0x1.88d5493faa639p-45) +A(0x1.0400000000000p+0, -0x1.fc0a8b0fc0000p-7, -0x1.f1e7cf6d3a69cp-50) +A(0x1.0400000000000p+0, -0x1.fc0a8b0fc0000p-7, -0x1.f1e7cf6d3a69cp-50) +A(0x1.0200000000000p+0, -0x1.fe02a6b100000p-8, -0x1.9e23f0dda40e4p-46) +A(0x1.0200000000000p+0, -0x1.fe02a6b100000p-8, -0x1.9e23f0dda40e4p-46) +A(0x1.0000000000000p+0, 0x0.0000000000000p+0, 0x0.0000000000000p+0) +A(0x1.0000000000000p+0, 0x0.0000000000000p+0, 0x0.0000000000000p+0) +A(0x1.fc00000000000p-1, 0x1.0101575890000p-7, -0x1.0c76b999d2be8p-46) +A(0x1.f800000000000p-1, 0x1.0205658938000p-6, -0x1.3dc5b06e2f7d2p-45) +A(0x1.f400000000000p-1, 0x1.8492528c90000p-6, -0x1.aa0ba325a0c34p-45) +A(0x1.f000000000000p-1, 0x1.0415d89e74000p-5, 0x1.111c05cf1d753p-47) +A(0x1.ec00000000000p-1, 0x1.466aed42e0000p-5, -0x1.c167375bdfd28p-45) +A(0x1.e800000000000p-1, 0x1.894aa149fc000p-5, -0x1.97995d05a267dp-46) +A(0x1.e400000000000p-1, 0x1.ccb73cdddc000p-5, -0x1.a68f247d82807p-46) +A(0x1.e200000000000p-1, 0x1.eea31c006c000p-5, -0x1.e113e4fc93b7bp-47) +A(0x1.de00000000000p-1, 0x1.1973bd1466000p-4, -0x1.5325d560d9e9bp-45) +A(0x1.da00000000000p-1, 0x1.3bdf5a7d1e000p-4, 0x1.cc85ea5db4ed7p-45) +A(0x1.d600000000000p-1, 0x1.5e95a4d97a000p-4, -0x1.c69063c5d1d1ep-45) +A(0x1.d400000000000p-1, 0x1.700d30aeac000p-4, 0x1.c1e8da99ded32p-49) +A(0x1.d000000000000p-1, 0x1.9335e5d594000p-4, 0x1.3115c3abd47dap-45) +A(0x1.cc00000000000p-1, 0x1.b6ac88dad6000p-4, -0x1.390802bf768e5p-46) +A(0x1.ca00000000000p-1, 0x1.c885801bc4000p-4, 0x1.646d1c65aacd3p-45) +A(0x1.c600000000000p-1, 0x1.ec739830a2000p-4, -0x1.dc068afe645e0p-45) +A(0x1.c400000000000p-1, 0x1.fe89139dbe000p-4, -0x1.534d64fa10afdp-45) +A(0x1.c000000000000p-1, 0x1.1178e8227e000p-3, 0x1.1ef78ce2d07f2p-45) +A(0x1.be00000000000p-1, 0x1.1aa2b7e23f000p-3, 0x1.ca78e44389934p-45) +A(0x1.ba00000000000p-1, 0x1.2d1610c868000p-3, 0x1.39d6ccb81b4a1p-47) +A(0x1.b800000000000p-1, 0x1.365fcb0159000p-3, 0x1.62fa8234b7289p-51) +A(0x1.b400000000000p-1, 0x1.4913d8333b000p-3, 0x1.5837954fdb678p-45) +A(0x1.b200000000000p-1, 0x1.527e5e4a1b000p-3, 0x1.633e8e5697dc7p-45) +A(0x1.ae00000000000p-1, 0x1.6574ebe8c1000p-3, 0x1.9cf8b2c3c2e78p-46) +A(0x1.ac00000000000p-1, 0x1.6f0128b757000p-3, -0x1.5118de59c21e1p-45) +A(0x1.aa00000000000p-1, 0x1.7898d85445000p-3, -0x1.c661070914305p-46) +A(0x1.a600000000000p-1, 0x1.8beafeb390000p-3, -0x1.73d54aae92cd1p-47) +A(0x1.a400000000000p-1, 0x1.95a5adcf70000p-3, 0x1.7f22858a0ff6fp-47) +A(0x1.a000000000000p-1, 0x1.a93ed3c8ae000p-3, -0x1.8724350562169p-45) +A(0x1.9e00000000000p-1, 0x1.b31d8575bd000p-3, -0x1.c358d4eace1aap-47) +A(0x1.9c00000000000p-1, 0x1.bd087383be000p-3, -0x1.d4bc4595412b6p-45) +A(0x1.9a00000000000p-1, 0x1.c6ffbc6f01000p-3, -0x1.1ec72c5962bd2p-48) +A(0x1.9600000000000p-1, 0x1.db13db0d49000p-3, -0x1.aff2af715b035p-45) +A(0x1.9400000000000p-1, 0x1.e530effe71000p-3, 0x1.212276041f430p-51) +A(0x1.9200000000000p-1, 0x1.ef5ade4dd0000p-3, -0x1.a211565bb8e11p-51) +A(0x1.9000000000000p-1, 0x1.f991c6cb3b000p-3, 0x1.bcbecca0cdf30p-46) +A(0x1.8c00000000000p-1, 0x1.07138604d5800p-2, 0x1.89cdb16ed4e91p-48) +A(0x1.8a00000000000p-1, 0x1.0c42d67616000p-2, 0x1.7188b163ceae9p-45) +A(0x1.8800000000000p-1, 0x1.1178e8227e800p-2, -0x1.c210e63a5f01cp-45) +A(0x1.8600000000000p-1, 0x1.16b5ccbacf800p-2, 0x1.b9acdf7a51681p-45) +A(0x1.8400000000000p-1, 0x1.1bf99635a6800p-2, 0x1.ca6ed5147bdb7p-45) +A(0x1.8200000000000p-1, 0x1.214456d0eb800p-2, 0x1.a87deba46baeap-47) +A(0x1.7e00000000000p-1, 0x1.2bef07cdc9000p-2, 0x1.a9cfa4a5004f4p-45) +A(0x1.7c00000000000p-1, 0x1.314f1e1d36000p-2, -0x1.8e27ad3213cb8p-45) +A(0x1.7a00000000000p-1, 0x1.36b6776be1000p-2, 0x1.16ecdb0f177c8p-46) +A(0x1.7800000000000p-1, 0x1.3c25277333000p-2, 0x1.83b54b606bd5cp-46) +A(0x1.7600000000000p-1, 0x1.419b423d5e800p-2, 0x1.8e436ec90e09dp-47) +A(0x1.7400000000000p-1, 0x1.4718dc271c800p-2, -0x1.f27ce0967d675p-45) +A(0x1.7200000000000p-1, 0x1.4c9e09e173000p-2, -0x1.e20891b0ad8a4p-45) +A(0x1.7000000000000p-1, 0x1.522ae0738a000p-2, 0x1.ebe708164c759p-45) +A(0x1.6e00000000000p-1, 0x1.57bf753c8d000p-2, 0x1.fadedee5d40efp-46) +A(0x1.6c00000000000p-1, 0x1.5d5bddf596000p-2, -0x1.a0b2a08a465dcp-47) +#endif +}, +}; +#endif /* __OBSOLETE_MATH */ diff --git a/newlib/libm/math/e_pow.c b/newlib/libm/math/e_pow.c index cad6737..6d2a501 100644 --- a/newlib/libm/math/e_pow.c +++ b/newlib/libm/math/e_pow.c @@ -58,6 +58,7 @@ */ #include "fdlibm.h" +#if __OBSOLETE_MATH #ifndef _DOUBLE_IS_32BITS @@ -313,3 +314,4 @@ ivln2_l = 1.92596299112661746887e-08; /* 0x3E54AE0B, 0xF85DDF44 =1/ln2 tail*/ } #endif /* defined(_DOUBLE_IS_32BITS) */ +#endif /* __OBSOLETE_MATH */ diff --git a/newlib/libm/math/w_pow.c b/newlib/libm/math/w_pow.c index fc38f68..9d1e396 100644 --- a/newlib/libm/math/w_pow.c +++ b/newlib/libm/math/w_pow.c @@ -52,6 +52,7 @@ PORTABILITY */ #include "fdlibm.h" +#if __OBSOLETE_MATH #include <errno.h> #ifndef _DOUBLE_IS_32BITS @@ -207,7 +208,7 @@ PORTABILITY } #endif /* defined(_DOUBLE_IS_32BITS) */ - +#endif /* __OBSOLETE_MATH */ |