1 | /* Function atanf vectorized with AVX-512. |
2 | Copyright (C) 2021-2024 Free Software Foundation, Inc. |
3 | This file is part of the GNU C Library. |
4 | |
5 | The GNU C Library is free software; you can redistribute it and/or |
6 | modify it under the terms of the GNU Lesser General Public |
7 | License as published by the Free Software Foundation; either |
8 | version 2.1 of the License, or (at your option) any later version. |
9 | |
10 | The GNU C Library is distributed in the hope that it will be useful, |
11 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
13 | Lesser General Public License for more details. |
14 | |
15 | You should have received a copy of the GNU Lesser General Public |
16 | License along with the GNU C Library; if not, see |
17 | https://www.gnu.org/licenses/. */ |
18 | |
19 | /* |
20 | * ALGORITHM DESCRIPTION: |
21 | * |
22 | * For 0.0 <= x <= 7.0/16.0: atan(x) = atan(0.0) + atan(s), where s=(x-0.0)/(1.0+0.0*x) |
23 | * For 7.0/16.0 <= x <= 11.0/16.0: atan(x) = atan(0.5) + atan(s), where s=(x-0.5)/(1.0+0.5*x) |
24 | * For 11.0/16.0 <= x <= 19.0/16.0: atan(x) = atan(1.0) + atan(s), where s=(x-1.0)/(1.0+1.0*x) |
25 | * For 19.0/16.0 <= x <= 39.0/16.0: atan(x) = atan(1.5) + atan(s), where s=(x-1.5)/(1.0+1.5*x) |
26 | * For 39.0/16.0 <= x <= inf : atan(x) = atan(inf) + atan(s), where s=-1.0/x |
27 | * Where atan(s) ~= s+s^3*Poly11(s^2) on interval |s|<7.0/0.16. |
28 | * |
29 | */ |
30 | |
31 | /* Offsets for data table __svml_satan_data_internal_avx512 |
32 | */ |
33 | #define AbsMask 0 |
34 | #define Shifter 64 |
35 | #define MaxThreshold 128 |
36 | #define MOne 192 |
37 | #define One 256 |
38 | #define LargeX 320 |
39 | #define Zero 384 |
40 | #define Tbl_H 448 |
41 | #define Pi2 576 |
42 | #define coeff_1 640 |
43 | #define coeff_2 704 |
44 | #define coeff_3 768 |
45 | |
46 | #include <sysdep.h> |
47 | |
48 | .section .text.evex512, "ax" , @progbits |
49 | ENTRY(_ZGVeN16v_atanf_skx) |
50 | vandps __svml_satan_data_internal_avx512(%rip), %zmm0, %zmm7 |
51 | vmovups MaxThreshold+__svml_satan_data_internal_avx512(%rip), %zmm3 |
52 | vmovups One+__svml_satan_data_internal_avx512(%rip), %zmm8 |
53 | |
54 | /* round to 2 bits after binary point */ |
55 | vreduceps $40, {sae}, %zmm7, %zmm5 |
56 | |
57 | /* saturate X range */ |
58 | vmovups LargeX+__svml_satan_data_internal_avx512(%rip), %zmm6 |
59 | vmovups Shifter+__svml_satan_data_internal_avx512(%rip), %zmm2 |
60 | vcmpps $29, {sae}, %zmm3, %zmm7, %k1 |
61 | |
62 | /* table lookup sequence */ |
63 | vmovups Tbl_H+__svml_satan_data_internal_avx512(%rip), %zmm3 |
64 | vsubps {rn-sae}, %zmm5, %zmm7, %zmm4 |
65 | vaddps {rn-sae}, %zmm2, %zmm7, %zmm1 |
66 | vxorps %zmm0, %zmm7, %zmm0 |
67 | vfmadd231ps {rn-sae}, %zmm7, %zmm4, %zmm8 |
68 | vmovups coeff_2+__svml_satan_data_internal_avx512(%rip), %zmm4 |
69 | |
70 | /* if|X|>=MaxThreshold, set DiffX=-1 */ |
71 | vblendmps MOne+__svml_satan_data_internal_avx512(%rip), %zmm5, %zmm9{%k1} |
72 | vmovups coeff_3+__svml_satan_data_internal_avx512(%rip), %zmm5 |
73 | |
74 | /* if|X|>=MaxThreshold, set Y=X */ |
75 | vminps {sae}, %zmm7, %zmm6, %zmm8{%k1} |
76 | |
77 | /* R+Rl = DiffX/Y */ |
78 | vgetmantps $0, {sae}, %zmm9, %zmm12 |
79 | vgetexpps {sae}, %zmm9, %zmm10 |
80 | vpermt2ps Tbl_H+64+__svml_satan_data_internal_avx512(%rip), %zmm1, %zmm3 |
81 | vgetmantps $0, {sae}, %zmm8, %zmm15 |
82 | vgetexpps {sae}, %zmm8, %zmm11 |
83 | vmovups coeff_1+__svml_satan_data_internal_avx512(%rip), %zmm1 |
84 | |
85 | /* set table value to Pi/2 for large X */ |
86 | vblendmps Pi2+__svml_satan_data_internal_avx512(%rip), %zmm3, %zmm9{%k1} |
87 | vrcp14ps %zmm15, %zmm13 |
88 | vsubps {rn-sae}, %zmm11, %zmm10, %zmm2 |
89 | vmulps {rn-sae}, %zmm13, %zmm12, %zmm14 |
90 | vfnmadd213ps {rn-sae}, %zmm12, %zmm14, %zmm15 |
91 | vfmadd213ps {rn-sae}, %zmm14, %zmm13, %zmm15 |
92 | vscalefps {rn-sae}, %zmm2, %zmm15, %zmm7 |
93 | |
94 | /* polynomial evaluation */ |
95 | vmulps {rn-sae}, %zmm7, %zmm7, %zmm8 |
96 | vmulps {rn-sae}, %zmm7, %zmm8, %zmm6 |
97 | vfmadd231ps {rn-sae}, %zmm8, %zmm1, %zmm4 |
98 | vfmadd213ps {rn-sae}, %zmm5, %zmm4, %zmm8 |
99 | vfmadd213ps {rn-sae}, %zmm7, %zmm6, %zmm8 |
100 | vaddps {rn-sae}, %zmm9, %zmm8, %zmm10 |
101 | vxorps %zmm0, %zmm10, %zmm0 |
102 | ret |
103 | |
104 | END(_ZGVeN16v_atanf_skx) |
105 | |
106 | .section .rodata, "a" |
107 | .align 64 |
108 | |
109 | #ifdef __svml_satan_data_internal_avx512_typedef |
110 | typedef unsigned int VUINT32; |
111 | typedef struct { |
112 | __declspec(align(64)) VUINT32 AbsMask[16][1]; |
113 | __declspec(align(64)) VUINT32 Shifter[16][1]; |
114 | __declspec(align(64)) VUINT32 MaxThreshold[16][1]; |
115 | __declspec(align(64)) VUINT32 MOne[16][1]; |
116 | __declspec(align(64)) VUINT32 One[16][1]; |
117 | __declspec(align(64)) VUINT32 LargeX[16][1]; |
118 | __declspec(align(64)) VUINT32 Zero[16][1]; |
119 | __declspec(align(64)) VUINT32 Tbl_H[32][1]; |
120 | __declspec(align(64)) VUINT32 Pi2[16][1]; |
121 | __declspec(align(64)) VUINT32 coeff[3][16][1]; |
122 | } __svml_satan_data_internal_avx512; |
123 | #endif |
124 | __svml_satan_data_internal_avx512: |
125 | /* AbsMask */ |
126 | .long 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff |
127 | /* Shifter */ |
128 | .align 64 |
129 | .long 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000, 0x4a000000 |
130 | /* MaxThreshold */ |
131 | .align 64 |
132 | .long 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000, 0x40F80000 |
133 | /* MOne */ |
134 | .align 64 |
135 | .long 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000 |
136 | /* One */ |
137 | .align 64 |
138 | .long 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 |
139 | /* LargeX */ |
140 | .align 64 |
141 | .long 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000, 0x4f800000 |
142 | /* Zero */ |
143 | .align 64 |
144 | .long 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000 |
145 | /* Tbl_H */ |
146 | .align 64 |
147 | .long 0x00000000, 0x3e7adbb0 |
148 | .long 0x3eed6338, 0x3f24bc7d |
149 | .long 0x3f490fdb, 0x3f6563e3 |
150 | .long 0x3f7b985f, 0x3f869c79 |
151 | .long 0x3f8db70d, 0x3f93877b |
152 | .long 0x3f985b6c, 0x3f9c6b53 |
153 | .long 0x3f9fe0bb, 0x3fa2daa4 |
154 | .long 0x3fa57088, 0x3fa7b46f |
155 | .long 0x3fa9b465, 0x3fab7b7a |
156 | .long 0x3fad1283, 0x3fae809e |
157 | .long 0x3fafcb99, 0x3fb0f836 |
158 | .long 0x3fb20a6a, 0x3fb30581 |
159 | .long 0x3fb3ec43, 0x3fb4c10a |
160 | .long 0x3fb585d7, 0x3fb63c64 |
161 | .long 0x3fb6e62c, 0x3fb78478 |
162 | .long 0x3fb81868, 0x3fb8a2f5 |
163 | /* Pi2 */ |
164 | .align 64 |
165 | .long 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB, 0x3fc90FDB |
166 | /* coeff3 */ |
167 | .align 64 |
168 | .long 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de, 0xbe0fa8de |
169 | .long 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2, 0x3e4cc8e2 |
170 | .long 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa, 0xbeaaaaaa |
171 | .align 64 |
172 | .type __svml_satan_data_internal_avx512, @object |
173 | .size __svml_satan_data_internal_avx512, .-__svml_satan_data_internal_avx512 |
174 | |