xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/sve/exp2f.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1 /*
2  * Single-precision SVE 2^x function.
3  *
4  * Copyright (c) 2023-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "sv_math.h"
9 #include "test_sig.h"
10 #include "test_defs.h"
11 
12 #define Thres 0x1.5d5e2ap+6f
13 
14 static const struct data
15 {
16   float c0, c2, c4, c1, c3;
17   float shift, thres;
18 } data = {
19   /* Coefficients copied from the polynomial in AdvSIMD variant.  */
20   .c0 = 0x1.62e422p-1f,
21   .c1 = 0x1.ebf9bcp-3f,
22   .c2 = 0x1.c6bd32p-5f,
23   .c3 = 0x1.3ce9e4p-7f,
24   .c4 = 0x1.59977ap-10f,
25   /* 1.5*2^17 + 127.  */
26   .shift = 0x1.803f8p17f,
27   /* Roughly 87.3. For x < -Thres, the result is subnormal and not handled
28      correctly by FEXPA.  */
29   .thres = Thres,
30 };
31 
32 static inline svfloat32_t
sv_exp2f_inline(svfloat32_t x,const svbool_t pg,const struct data * d)33 sv_exp2f_inline (svfloat32_t x, const svbool_t pg, const struct data *d)
34 {
35   /* exp2(x) = 2^n (1 + poly(r)), with 1 + poly(r) in [1/sqrt(2),sqrt(2)]
36     x = n + r, with r in [-1/2, 1/2].  */
37   svfloat32_t z = svadd_x (svptrue_b32 (), x, d->shift);
38   svfloat32_t n = svsub_x (svptrue_b32 (), z, d->shift);
39   svfloat32_t r = svsub_x (svptrue_b32 (), x, n);
40 
41   svfloat32_t scale = svexpa (svreinterpret_u32 (z));
42 
43   /* Polynomial evaluation: poly(r) ~ exp2(r)-1.
44      Evaluate polynomial use hybrid scheme - offset ESTRIN by 1 for
45      coefficients 1 to 4, and apply most significant coefficient directly.  */
46   svfloat32_t even_coeffs = svld1rq (svptrue_b32 (), &d->c0);
47   svfloat32_t r2 = svmul_x (svptrue_b32 (), r, r);
48   svfloat32_t p12 = svmla_lane (sv_f32 (d->c1), r, even_coeffs, 1);
49   svfloat32_t p34 = svmla_lane (sv_f32 (d->c3), r, even_coeffs, 2);
50   svfloat32_t p14 = svmla_x (pg, p12, r2, p34);
51   svfloat32_t p0 = svmul_lane (r, even_coeffs, 0);
52   svfloat32_t poly = svmla_x (pg, p0, r2, p14);
53 
54   return svmla_x (pg, scale, scale, poly);
55 }
56 
57 static svfloat32_t NOINLINE
special_case(svfloat32_t x,svbool_t special,const struct data * d)58 special_case (svfloat32_t x, svbool_t special, const struct data *d)
59 {
60   return sv_call_f32 (exp2f, x, sv_exp2f_inline (x, svptrue_b32 (), d),
61 		      special);
62 }
63 
64 /* Single-precision SVE exp2f routine. Implements the same algorithm
65    as AdvSIMD exp2f.
66    Worst case error is 1.04 ULPs.
67    _ZGVsMxv_exp2f(-0x1.af994ap-3) got 0x1.ba6a66p-1
68 				 want 0x1.ba6a64p-1.  */
SV_NAME_F1(exp2)69 svfloat32_t SV_NAME_F1 (exp2) (svfloat32_t x, const svbool_t pg)
70 {
71   const struct data *d = ptr_barrier (&data);
72   svbool_t special = svacgt (pg, x, d->thres);
73   if (unlikely (svptest_any (special, special)))
74     return special_case (x, special, d);
75   return sv_exp2f_inline (x, pg, d);
76 }
77 
78 TEST_SIG (SV, F, 1, exp2, -9.9, 9.9)
79 TEST_ULP (SV_NAME_F1 (exp2), 0.54)
80 TEST_DISABLE_FENV (SV_NAME_F1 (exp2))
81 TEST_SYM_INTERVAL (SV_NAME_F1 (exp2), 0, Thres, 50000)
82 TEST_SYM_INTERVAL (SV_NAME_F1 (exp2), Thres, inf, 50000)
83 CLOSE_SVE_ATTR
84