xref: /freebsd/contrib/arm-optimized-routines/pl/math/sv_cosf_2u1.c (revision a4e5e0106ac7145f56eb39a691e302cabb4635be)
1 /*
2  * Single-precision SVE cos(x) function.
3  *
4  * Copyright (c) 2019-2023, 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 #if SV_SUPPORTED
13 
14 #define NegPio2_1 (sv_f32 (-0x1.921fb6p+0f))
15 #define NegPio2_2 (sv_f32 (0x1.777a5cp-25f))
16 #define NegPio2_3 (sv_f32 (0x1.ee59dap-50f))
17 #define RangeVal (sv_f32 (0x1p20f))
18 #define InvPio2 (sv_f32 (0x1.45f306p-1f))
19 /* Original shift used in Neon cosf,
20    plus a contribution to set the bit #0 of q
21    as expected by trigonometric instructions.  */
22 #define Shift (sv_f32 (0x1.800002p+23f))
23 #define AbsMask (0x7fffffff)
24 
25 static NOINLINE sv_f32_t
26 __sv_cosf_specialcase (sv_f32_t x, sv_f32_t y, svbool_t cmp)
27 {
28   return sv_call_f32 (cosf, x, y, cmp);
29 }
30 
31 /* A fast SVE implementation of cosf based on trigonometric
32    instructions (FTMAD, FTSSEL, FTSMUL).
33    Maximum measured error: 2.06 ULPs.
34    __sv_cosf(0x1.dea2f2p+19) got 0x1.fffe7ap-6
35 			    want 0x1.fffe76p-6.  */
36 sv_f32_t
37 __sv_cosf_x (sv_f32_t x, const svbool_t pg)
38 {
39   sv_f32_t n, r, r2, y;
40   svbool_t cmp;
41 
42   r = sv_as_f32_u32 (svand_n_u32_x (pg, sv_as_u32_f32 (x), AbsMask));
43   cmp = svcmpge_u32 (pg, sv_as_u32_f32 (r), sv_as_u32_f32 (RangeVal));
44 
45   /* n = rint(|x|/(pi/2)).  */
46   sv_f32_t q = sv_fma_f32_x (pg, InvPio2, r, Shift);
47   n = svsub_f32_x (pg, q, Shift);
48 
49   /* r = |x| - n*(pi/2)  (range reduction into -pi/4 .. pi/4).  */
50   r = sv_fma_f32_x (pg, NegPio2_1, n, r);
51   r = sv_fma_f32_x (pg, NegPio2_2, n, r);
52   r = sv_fma_f32_x (pg, NegPio2_3, n, r);
53 
54   /* Final multiplicative factor: 1.0 or x depending on bit #0 of q.  */
55   sv_f32_t f = svtssel_f32 (r, sv_as_u32_f32 (q));
56 
57   /* cos(r) poly approx.  */
58   r2 = svtsmul_f32 (r, sv_as_u32_f32 (q));
59   y = sv_f32 (0.0f);
60   y = svtmad_f32 (y, r2, 4);
61   y = svtmad_f32 (y, r2, 3);
62   y = svtmad_f32 (y, r2, 2);
63   y = svtmad_f32 (y, r2, 1);
64   y = svtmad_f32 (y, r2, 0);
65 
66   /* Apply factor.  */
67   y = svmul_f32_x (pg, f, y);
68 
69   /* No need to pass pg to specialcase here since cmp is a strict subset,
70      guaranteed by the cmpge above.  */
71   if (unlikely (svptest_any (pg, cmp)))
72     return __sv_cosf_specialcase (x, y, cmp);
73   return y;
74 }
75 
76 PL_ALIAS (__sv_cosf_x, _ZGVsMxv_cosf)
77 
78 PL_SIG (SV, F, 1, cos, -3.1, 3.1)
79 PL_TEST_ULP (__sv_cosf, 1.57)
80 PL_TEST_INTERVAL (__sv_cosf, 0, 0xffff0000, 10000)
81 PL_TEST_INTERVAL (__sv_cosf, 0x1p-4, 0x1p4, 500000)
82 #endif
83