1 | /* Function log10f 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 | * Get short reciprocal approximation Rcp ~ 1/mantissa(x) |
23 | * R = Rcp*x - 1.0 |
24 | * log10(x) = k*log10(2.0) - log10(Rcp) + poly_approximation(R) |
25 | * log10(Rcp) is tabulated |
26 | * |
27 | * |
28 | */ |
29 | |
30 | /* Offsets for data table __svml_slog10_data_internal_avx512 |
31 | */ |
32 | #define One 0 |
33 | #define coeff4 64 |
34 | #define coeff3 128 |
35 | #define coeff2 192 |
36 | #define coeff1 256 |
37 | #define L2 320 |
38 | |
39 | #include <sysdep.h> |
40 | |
41 | .section .text.evex512, "ax" , @progbits |
42 | ENTRY(_ZGVeN16v_log10f_skx) |
43 | pushq %rbp |
44 | cfi_def_cfa_offset(16) |
45 | movq %rsp, %rbp |
46 | cfi_def_cfa(6, 16) |
47 | cfi_offset(6, -16) |
48 | andq $-64, %rsp |
49 | subq $192, %rsp |
50 | vgetmantps $11, {sae}, %zmm0, %zmm3 |
51 | vmovups __svml_slog10_data_internal_avx512(%rip), %zmm1 |
52 | vgetexpps {sae}, %zmm0, %zmm5 |
53 | vmovups L2+__svml_slog10_data_internal_avx512(%rip), %zmm10 |
54 | vpsrld $19, %zmm3, %zmm7 |
55 | vgetexpps {sae}, %zmm3, %zmm6 |
56 | vsubps {rn-sae}, %zmm1, %zmm3, %zmm11 |
57 | vpermps coeff4+__svml_slog10_data_internal_avx512(%rip), %zmm7, %zmm1 |
58 | vpermps coeff3+__svml_slog10_data_internal_avx512(%rip), %zmm7, %zmm2 |
59 | vsubps {rn-sae}, %zmm6, %zmm5, %zmm9 |
60 | vpermps coeff2+__svml_slog10_data_internal_avx512(%rip), %zmm7, %zmm4 |
61 | vpermps coeff1+__svml_slog10_data_internal_avx512(%rip), %zmm7, %zmm8 |
62 | |
63 | /* x<=0? */ |
64 | vfpclassps $94, %zmm0, %k0 |
65 | vfmadd213ps {rn-sae}, %zmm2, %zmm11, %zmm1 |
66 | vmulps {rn-sae}, %zmm10, %zmm9, %zmm12 |
67 | vfmadd213ps {rn-sae}, %zmm4, %zmm11, %zmm1 |
68 | kmovw %k0, %edx |
69 | vfmadd213ps {rn-sae}, %zmm8, %zmm11, %zmm1 |
70 | vfmadd213ps {rn-sae}, %zmm12, %zmm11, %zmm1 |
71 | testl %edx, %edx |
72 | |
73 | /* Go to special inputs processing branch */ |
74 | jne L(SPECIAL_VALUES_BRANCH) |
75 | # LOE rbx r12 r13 r14 r15 edx zmm0 zmm1 |
76 | |
77 | /* Restore registers |
78 | * and exit the function |
79 | */ |
80 | |
81 | L(EXIT): |
82 | vmovaps %zmm1, %zmm0 |
83 | movq %rbp, %rsp |
84 | popq %rbp |
85 | cfi_def_cfa(7, 8) |
86 | cfi_restore(6) |
87 | ret |
88 | cfi_def_cfa(6, 16) |
89 | cfi_offset(6, -16) |
90 | |
91 | /* Branch to process |
92 | * special inputs |
93 | */ |
94 | |
95 | L(SPECIAL_VALUES_BRANCH): |
96 | vmovups %zmm0, 64(%rsp) |
97 | vmovups %zmm1, 128(%rsp) |
98 | # LOE rbx r12 r13 r14 r15 edx zmm1 |
99 | |
100 | xorl %eax, %eax |
101 | # LOE rbx r12 r13 r14 r15 eax edx |
102 | |
103 | vzeroupper |
104 | movq %r12, 16(%rsp) |
105 | /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -176; DW_OP_plus) */ |
106 | .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22 |
107 | movl %eax, %r12d |
108 | movq %r13, 8(%rsp) |
109 | /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -184; DW_OP_plus) */ |
110 | .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22 |
111 | movl %edx, %r13d |
112 | movq %r14, (%rsp) |
113 | /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -192; DW_OP_plus) */ |
114 | .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22 |
115 | # LOE rbx r15 r12d r13d |
116 | |
117 | /* Range mask |
118 | * bits check |
119 | */ |
120 | |
121 | L(RANGEMASK_CHECK): |
122 | btl %r12d, %r13d |
123 | |
124 | /* Call scalar math function */ |
125 | jc L(SCALAR_MATH_CALL) |
126 | # LOE rbx r15 r12d r13d |
127 | |
128 | /* Special inputs |
129 | * processing loop |
130 | */ |
131 | |
132 | L(SPECIAL_VALUES_LOOP): |
133 | incl %r12d |
134 | cmpl $16, %r12d |
135 | |
136 | /* Check bits in range mask */ |
137 | jl L(RANGEMASK_CHECK) |
138 | # LOE rbx r15 r12d r13d |
139 | |
140 | movq 16(%rsp), %r12 |
141 | cfi_restore(12) |
142 | movq 8(%rsp), %r13 |
143 | cfi_restore(13) |
144 | movq (%rsp), %r14 |
145 | cfi_restore(14) |
146 | vmovups 128(%rsp), %zmm1 |
147 | |
148 | /* Go to exit */ |
149 | jmp L(EXIT) |
150 | /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -176; DW_OP_plus) */ |
151 | .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22 |
152 | /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -184; DW_OP_plus) */ |
153 | .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22 |
154 | /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -192; DW_OP_plus) */ |
155 | .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22 |
156 | # LOE rbx r12 r13 r14 r15 zmm1 |
157 | |
158 | /* Scalar math function call |
159 | * to process special input |
160 | */ |
161 | |
162 | L(SCALAR_MATH_CALL): |
163 | movl %r12d, %r14d |
164 | vmovss 64(%rsp, %r14, 4), %xmm0 |
165 | call log10f@PLT |
166 | # LOE rbx r14 r15 r12d r13d xmm0 |
167 | |
168 | vmovss %xmm0, 128(%rsp, %r14, 4) |
169 | |
170 | /* Process special inputs in loop */ |
171 | jmp L(SPECIAL_VALUES_LOOP) |
172 | # LOE rbx r15 r12d r13d |
173 | END(_ZGVeN16v_log10f_skx) |
174 | |
175 | .section .rodata, "a" |
176 | .align 64 |
177 | |
178 | #ifdef __svml_slog10_data_internal_avx512_typedef |
179 | typedef unsigned int VUINT32; |
180 | typedef struct { |
181 | __declspec(align(64)) VUINT32 One[16][1]; |
182 | __declspec(align(64)) VUINT32 coeff4[16][1]; |
183 | __declspec(align(64)) VUINT32 coeff3[16][1]; |
184 | __declspec(align(64)) VUINT32 coeff2[16][1]; |
185 | __declspec(align(64)) VUINT32 coeff1[16][1]; |
186 | __declspec(align(64)) VUINT32 L2[16][1]; |
187 | } __svml_slog10_data_internal_avx512; |
188 | #endif |
189 | __svml_slog10_data_internal_avx512: |
190 | /* One */ |
191 | .long 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 |
192 | // c4 |
193 | .align 64 |
194 | .long 0xbdc9ae9b, 0xbda6fcf4 |
195 | .long 0xbd8bac76, 0xbd6bca30 |
196 | .long 0xbd48a99b, 0xbd2c0a9f |
197 | .long 0xbd1480db, 0xbd00faf2 |
198 | .long 0xbe823aa9, 0xbe656348 |
199 | .long 0xbe4afbb9, 0xbe346895 |
200 | .long 0xbe20ffff, 0xbe103a0b |
201 | .long 0xbe01a91c, 0xbde9e84e |
202 | // c3 |
203 | .align 64 |
204 | .long 0x3e13d888, 0x3e10a87c |
205 | .long 0x3e0b95c3, 0x3e057f0b |
206 | .long 0x3dfde038, 0x3df080d9 |
207 | .long 0x3de34c1e, 0x3dd68333 |
208 | .long 0x3dac6e8e, 0x3dd54a51 |
209 | .long 0x3df30f40, 0x3e04235d |
210 | .long 0x3e0b7033, 0x3e102c90 |
211 | .long 0x3e12ebad, 0x3e141ff8 |
212 | // c2 |
213 | .align 64 |
214 | .long 0xbe5e5a9b, 0xbe5e2677 |
215 | .long 0xbe5d83f5, 0xbe5c6016 |
216 | .long 0xbe5abd0b, 0xbe58a6fd |
217 | .long 0xbe562e02, 0xbe5362f8 |
218 | .long 0xbe68e27c, 0xbe646747 |
219 | .long 0xbe619a73, 0xbe5ff05a |
220 | .long 0xbe5f0570, 0xbe5e92d0 |
221 | .long 0xbe5e662b, 0xbe5e5c08 |
222 | // c1 |
223 | .align 64 |
224 | .long 0x3ede5bd8, 0x3ede5b45 |
225 | .long 0x3ede57d8, 0x3ede4eb1 |
226 | .long 0x3ede3d37, 0x3ede2166 |
227 | .long 0x3eddf9d9, 0x3eddc5bb |
228 | .long 0x3ede08ed, 0x3ede32e7 |
229 | .long 0x3ede4967, 0x3ede5490 |
230 | .long 0x3ede597f, 0x3ede5b50 |
231 | .long 0x3ede5bca, 0x3ede5bd9 |
232 | /* L2 */ |
233 | .align 64 |
234 | .long 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b |
235 | .align 64 |
236 | .type __svml_slog10_data_internal_avx512, @object |
237 | .size __svml_slog10_data_internal_avx512, .-__svml_slog10_data_internal_avx512 |
238 | |