1 | /* Function log2f 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 | * log2(x) = k - log2(Rcp) + poly_approximation(R) |
25 | * log2(Rcp) is tabulated |
26 | * |
27 | * |
28 | */ |
29 | |
30 | /* Offsets for data table __svml_slog2_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 | |
38 | #include <sysdep.h> |
39 | |
40 | .section .text.evex512, "ax" , @progbits |
41 | ENTRY(_ZGVeN16v_log2f_skx) |
42 | pushq %rbp |
43 | cfi_def_cfa_offset(16) |
44 | movq %rsp, %rbp |
45 | cfi_def_cfa(6, 16) |
46 | cfi_offset(6, -16) |
47 | andq $-64, %rsp |
48 | subq $192, %rsp |
49 | vgetmantps $11, {sae}, %zmm0, %zmm3 |
50 | vmovups __svml_slog2_data_internal_avx512(%rip), %zmm1 |
51 | vgetexpps {sae}, %zmm0, %zmm5 |
52 | |
53 | /* x<=0? */ |
54 | vfpclassps $94, %zmm0, %k0 |
55 | vsubps {rn-sae}, %zmm1, %zmm3, %zmm9 |
56 | vpsrld $19, %zmm3, %zmm7 |
57 | vgetexpps {sae}, %zmm3, %zmm6 |
58 | vpermps coeff4+__svml_slog2_data_internal_avx512(%rip), %zmm7, %zmm1 |
59 | vpermps coeff3+__svml_slog2_data_internal_avx512(%rip), %zmm7, %zmm2 |
60 | vpermps coeff2+__svml_slog2_data_internal_avx512(%rip), %zmm7, %zmm4 |
61 | vpermps coeff1+__svml_slog2_data_internal_avx512(%rip), %zmm7, %zmm8 |
62 | vsubps {rn-sae}, %zmm6, %zmm5, %zmm10 |
63 | vfmadd213ps {rn-sae}, %zmm2, %zmm9, %zmm1 |
64 | kmovw %k0, %edx |
65 | vfmadd213ps {rn-sae}, %zmm4, %zmm9, %zmm1 |
66 | vfmadd213ps {rn-sae}, %zmm8, %zmm9, %zmm1 |
67 | vfmadd213ps {rn-sae}, %zmm10, %zmm9, %zmm1 |
68 | testl %edx, %edx |
69 | |
70 | /* Go to special inputs processing branch */ |
71 | jne L(SPECIAL_VALUES_BRANCH) |
72 | # LOE rbx r12 r13 r14 r15 edx zmm0 zmm1 |
73 | |
74 | /* Restore registers |
75 | * and exit the function |
76 | */ |
77 | |
78 | L(EXIT): |
79 | vmovaps %zmm1, %zmm0 |
80 | movq %rbp, %rsp |
81 | popq %rbp |
82 | cfi_def_cfa(7, 8) |
83 | cfi_restore(6) |
84 | ret |
85 | cfi_def_cfa(6, 16) |
86 | cfi_offset(6, -16) |
87 | |
88 | /* Branch to process |
89 | * special inputs |
90 | */ |
91 | |
92 | L(SPECIAL_VALUES_BRANCH): |
93 | vmovups %zmm0, 64(%rsp) |
94 | vmovups %zmm1, 128(%rsp) |
95 | # LOE rbx r12 r13 r14 r15 edx zmm1 |
96 | |
97 | xorl %eax, %eax |
98 | # LOE rbx r12 r13 r14 r15 eax edx |
99 | |
100 | vzeroupper |
101 | movq %r12, 16(%rsp) |
102 | /* 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) */ |
103 | .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22 |
104 | movl %eax, %r12d |
105 | movq %r13, 8(%rsp) |
106 | /* 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) */ |
107 | .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22 |
108 | movl %edx, %r13d |
109 | movq %r14, (%rsp) |
110 | /* 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) */ |
111 | .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22 |
112 | # LOE rbx r15 r12d r13d |
113 | |
114 | /* Range mask |
115 | * bits check |
116 | */ |
117 | |
118 | L(RANGEMASK_CHECK): |
119 | btl %r12d, %r13d |
120 | |
121 | /* Call scalar math function */ |
122 | jc L(SCALAR_MATH_CALL) |
123 | # LOE rbx r15 r12d r13d |
124 | |
125 | /* Special inputs |
126 | * processing loop |
127 | */ |
128 | |
129 | L(SPECIAL_VALUES_LOOP): |
130 | incl %r12d |
131 | cmpl $16, %r12d |
132 | |
133 | /* Check bits in range mask */ |
134 | jl L(RANGEMASK_CHECK) |
135 | # LOE rbx r15 r12d r13d |
136 | |
137 | movq 16(%rsp), %r12 |
138 | cfi_restore(12) |
139 | movq 8(%rsp), %r13 |
140 | cfi_restore(13) |
141 | movq (%rsp), %r14 |
142 | cfi_restore(14) |
143 | vmovups 128(%rsp), %zmm1 |
144 | |
145 | /* Go to exit */ |
146 | jmp L(EXIT) |
147 | /* 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) */ |
148 | .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22 |
149 | /* 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) */ |
150 | .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22 |
151 | /* 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) */ |
152 | .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22 |
153 | # LOE rbx r12 r13 r14 r15 zmm1 |
154 | |
155 | /* Scalar math function call |
156 | * to process special input |
157 | */ |
158 | |
159 | L(SCALAR_MATH_CALL): |
160 | movl %r12d, %r14d |
161 | vmovss 64(%rsp, %r14, 4), %xmm0 |
162 | call log2f@PLT |
163 | # LOE rbx r14 r15 r12d r13d xmm0 |
164 | |
165 | vmovss %xmm0, 128(%rsp, %r14, 4) |
166 | |
167 | /* Process special inputs in loop */ |
168 | jmp L(SPECIAL_VALUES_LOOP) |
169 | # LOE rbx r15 r12d r13d |
170 | END(_ZGVeN16v_log2f_skx) |
171 | |
172 | .section .rodata, "a" |
173 | .align 64 |
174 | |
175 | #ifdef __svml_slog2_data_internal_avx512_typedef |
176 | typedef unsigned int VUINT32; |
177 | typedef struct { |
178 | __declspec(align(64)) VUINT32 One[16][1]; |
179 | __declspec(align(64)) VUINT32 coeff4[16][1]; |
180 | __declspec(align(64)) VUINT32 coeff3[16][1]; |
181 | __declspec(align(64)) VUINT32 coeff2[16][1]; |
182 | __declspec(align(64)) VUINT32 coeff1[16][1]; |
183 | } __svml_slog2_data_internal_avx512; |
184 | #endif |
185 | __svml_slog2_data_internal_avx512: |
186 | /* One */ |
187 | .long 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 |
188 | // c4 |
189 | .align 64 |
190 | .long 0xbea77e4a, 0xbe8aae3d |
191 | .long 0xbe67fe32, 0xbe43d1b6 |
192 | .long 0xbe26a589, 0xbe0ee09b |
193 | .long 0xbdf6a8a1, 0xbdd63b49 |
194 | .long 0xbf584e51, 0xbf3e80a1 |
195 | .long 0xbf2892f0, 0xbf15d377 |
196 | .long 0xbf05b525, 0xbeef8e30 |
197 | .long 0xbed75c8f, 0xbec24184 |
198 | // c3 |
199 | .align 64 |
200 | .long 0x3ef5910c, 0x3ef045a1 |
201 | .long 0x3ee7d87e, 0x3eddbb84 |
202 | .long 0x3ed2d6df, 0x3ec7bbd2 |
203 | .long 0x3ebcc42f, 0x3eb22616 |
204 | .long 0x3e8f3399, 0x3eb1223e |
205 | .long 0x3ec9db4a, 0x3edb7a09 |
206 | .long 0x3ee79a1a, 0x3eef77cb |
207 | .long 0x3ef407a4, 0x3ef607b4 |
208 | // c2 |
209 | .align 64 |
210 | .long 0xbf38a934, 0xbf387de6 |
211 | .long 0xbf37f6f0, 0xbf37048b |
212 | .long 0xbf35a88a, 0xbf33ed04 |
213 | .long 0xbf31df56, 0xbf2f8d82 |
214 | .long 0xbf416814, 0xbf3daf58 |
215 | .long 0xbf3b5c08, 0xbf39fa2a |
216 | .long 0xbf393713, 0xbf38d7e1 |
217 | .long 0xbf38b2cd, 0xbf38aa62 |
218 | // c1 |
219 | .align 64 |
220 | .long 0x3fb8aa3b, 0x3fb8a9c0 |
221 | .long 0x3fb8a6e8, 0x3fb89f4e |
222 | .long 0x3fb890cb, 0x3fb879b1 |
223 | .long 0x3fb858d8, 0x3fb82d90 |
224 | .long 0x3fb8655e, 0x3fb8883a |
225 | .long 0x3fb89aea, 0x3fb8a42f |
226 | .long 0x3fb8a848, 0x3fb8a9c9 |
227 | .long 0x3fb8aa2f, 0x3fb8aa3b |
228 | .align 64 |
229 | .type __svml_slog2_data_internal_avx512, @object |
230 | .size __svml_slog2_data_internal_avx512, .-__svml_slog2_data_internal_avx512 |
231 | |