xref: /freebsd/contrib/arm-optimized-routines/pl/math/sv_exp_1u5.c (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 /*
2  * Double-precision vector e^x function.
3  *
4  * Copyright (c) 2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "sv_math.h"
9 #include "pl_sig.h"
10 #include "pl_test.h"
11 
12 static const struct data
13 {
14   double poly[4];
15   double ln2_hi, ln2_lo, inv_ln2, shift, thres;
16 } data = {
17   .poly = { /* ulp error: 0.53.  */
18 	    0x1.fffffffffdbcdp-2, 0x1.555555555444cp-3, 0x1.555573c6a9f7dp-5,
19 	    0x1.1111266d28935p-7 },
20   .ln2_hi = 0x1.62e42fefa3800p-1,
21   .ln2_lo = 0x1.ef35793c76730p-45,
22   /* 1/ln2.  */
23   .inv_ln2 = 0x1.71547652b82fep+0,
24   /* 1.5*2^46+1023. This value is further explained below.  */
25   .shift = 0x1.800000000ffc0p+46,
26   .thres = 704.0,
27 };
28 
29 #define C(i) sv_f64 (d->poly[i])
30 #define SpecialOffset 0x6000000000000000 /* 0x1p513.  */
31 /* SpecialBias1 + SpecialBias1 = asuint(1.0).  */
32 #define SpecialBias1 0x7000000000000000 /* 0x1p769.  */
33 #define SpecialBias2 0x3010000000000000 /* 0x1p-254.  */
34 
35 /* Update of both special and non-special cases, if any special case is
36    detected.  */
37 static inline svfloat64_t
38 special_case (svbool_t pg, svfloat64_t s, svfloat64_t y, svfloat64_t n)
39 {
40   /* s=2^n may overflow, break it up into s=s1*s2,
41      such that exp = s + s*y can be computed as s1*(s2+s2*y)
42      and s1*s1 overflows only if n>0.  */
43 
44   /* If n<=0 then set b to 0x6, 0 otherwise.  */
45   svbool_t p_sign = svcmple (pg, n, 0.0); /* n <= 0.  */
46   svuint64_t b
47       = svdup_u64_z (p_sign, SpecialOffset); /* Inactive lanes set to 0.  */
48 
49   /* Set s1 to generate overflow depending on sign of exponent n.  */
50   svfloat64_t s1 = svreinterpret_f64 (
51       svsubr_x (pg, b, SpecialBias1)); /* 0x70...0 - b.  */
52   /* Offset s to avoid overflow in final result if n is below threshold.  */
53   svfloat64_t s2 = svreinterpret_f64 (
54       svadd_x (pg, svsub_x (pg, svreinterpret_u64 (s), SpecialBias2),
55 	       b)); /* as_u64 (s) - 0x3010...0 + b.  */
56 
57   /* |n| > 1280 => 2^(n) overflows.  */
58   svbool_t p_cmp = svacgt (pg, n, 1280.0);
59 
60   svfloat64_t r1 = svmul_x (pg, s1, s1);
61   svfloat64_t r2 = svmla_x (pg, s2, s2, y);
62   svfloat64_t r0 = svmul_x (pg, r2, s1);
63 
64   return svsel (p_cmp, r1, r0);
65 }
66 
67 /* SVE exp algorithm. Maximum measured error is 1.01ulps:
68    SV_NAME_D1 (exp)(0x1.4619d7b04da41p+6) got 0x1.885d9acc41da7p+117
69 					 want 0x1.885d9acc41da6p+117.  */
70 svfloat64_t SV_NAME_D1 (exp) (svfloat64_t x, const svbool_t pg)
71 {
72   const struct data *d = ptr_barrier (&data);
73 
74   svbool_t special = svacgt (pg, x, d->thres);
75 
76   /* Use a modifed version of the shift used for flooring, such that x/ln2 is
77      rounded to a multiple of 2^-6=1/64, shift = 1.5 * 2^52 * 2^-6 = 1.5 *
78      2^46.
79 
80      n is not an integer but can be written as n = m + i/64, with i and m
81      integer, 0 <= i < 64 and m <= n.
82 
83      Bits 5:0 of z will be null every time x/ln2 reaches a new integer value
84      (n=m, i=0), and is incremented every time z (or n) is incremented by 1/64.
85      FEXPA expects i in bits 5:0 of the input so it can be used as index into
86      FEXPA hardwired table T[i] = 2^(i/64) for i = 0:63, that will in turn
87      populate the mantissa of the output. Therefore, we use u=asuint(z) as
88      input to FEXPA.
89 
90      We add 1023 to the modified shift value in order to set bits 16:6 of u to
91      1, such that once these bits are moved to the exponent of the output of
92      FEXPA, we get the exponent of 2^n right, i.e. we get 2^m.  */
93   svfloat64_t z = svmla_x (pg, sv_f64 (d->shift), x, d->inv_ln2);
94   svuint64_t u = svreinterpret_u64 (z);
95   svfloat64_t n = svsub_x (pg, z, d->shift);
96 
97   /* r = x - n * ln2, r is in [-ln2/(2N), ln2/(2N)].  */
98   svfloat64_t ln2 = svld1rq (svptrue_b64 (), &d->ln2_hi);
99   svfloat64_t r = svmls_lane (x, n, ln2, 0);
100   r = svmls_lane (r, n, ln2, 1);
101 
102   /* y = exp(r) - 1 ~= r + C0 r^2 + C1 r^3 + C2 r^4 + C3 r^5.  */
103   svfloat64_t r2 = svmul_x (pg, r, r);
104   svfloat64_t p01 = svmla_x (pg, C (0), C (1), r);
105   svfloat64_t p23 = svmla_x (pg, C (2), C (3), r);
106   svfloat64_t p04 = svmla_x (pg, p01, p23, r2);
107   svfloat64_t y = svmla_x (pg, r, p04, r2);
108 
109   /* s = 2^n, computed using FEXPA. FEXPA does not propagate NaNs, so for
110      consistent NaN handling we have to manually propagate them. This comes at
111      significant performance cost.  */
112   svfloat64_t s = svexpa (u);
113 
114   /* Assemble result as exp(x) = 2^n * exp(r).  If |x| > Thresh the
115      multiplication may overflow, so use special case routine.  */
116 
117   if (unlikely (svptest_any (pg, special)))
118     {
119       /* FEXPA zeroes the sign bit, however the sign is meaningful to the
120 	 special case function so needs to be copied.
121 	 e = sign bit of u << 46.  */
122       svuint64_t e = svand_x (pg, svlsl_x (pg, u, 46), 0x8000000000000000);
123       /* Copy sign to s.  */
124       s = svreinterpret_f64 (svadd_x (pg, e, svreinterpret_u64 (s)));
125       return special_case (pg, s, y, n);
126     }
127 
128   /* No special case.  */
129   return svmla_x (pg, s, s, y);
130 }
131 
132 PL_SIG (SV, D, 1, exp, -9.9, 9.9)
133 PL_TEST_ULP (SV_NAME_D1 (exp), 1.46)
134 PL_TEST_SYM_INTERVAL (SV_NAME_D1 (exp), 0, 0x1p-23, 40000)
135 PL_TEST_SYM_INTERVAL (SV_NAME_D1 (exp), 0x1p-23, 1, 50000)
136 PL_TEST_SYM_INTERVAL (SV_NAME_D1 (exp), 1, 0x1p23, 50000)
137 PL_TEST_SYM_INTERVAL (SV_NAME_D1 (exp), 0x1p23, inf, 50000)
138