1 | /* Function log10f vectorized with SSE4. |
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 |
31 | */ |
32 | #define MinNorm 0 |
33 | #define MaxNorm 16 |
34 | #define L2H 32 |
35 | #define L2L 48 |
36 | #define iBrkValue 64 |
37 | #define iOffExpoMask 80 |
38 | #define One 96 |
39 | #define sPoly 112 |
40 | #define L2 256 |
41 | |
42 | #include <sysdep.h> |
43 | |
44 | .section .text.sse4, "ax" , @progbits |
45 | ENTRY(_ZGVbN4v_log10f_sse4) |
46 | subq $72, %rsp |
47 | cfi_def_cfa_offset(80) |
48 | movaps %xmm0, %xmm1 |
49 | |
50 | /* reduction: compute r, n */ |
51 | movdqu iBrkValue+__svml_slog10_data_internal(%rip), %xmm2 |
52 | movaps %xmm0, %xmm4 |
53 | movdqu iOffExpoMask+__svml_slog10_data_internal(%rip), %xmm10 |
54 | psubd %xmm2, %xmm1 |
55 | pand %xmm1, %xmm10 |
56 | psrad $23, %xmm1 |
57 | paddd %xmm2, %xmm10 |
58 | movaps %xmm0, %xmm3 |
59 | movups sPoly+__svml_slog10_data_internal(%rip), %xmm5 |
60 | movups sPoly+32+__svml_slog10_data_internal(%rip), %xmm6 |
61 | movups sPoly+64+__svml_slog10_data_internal(%rip), %xmm7 |
62 | movups sPoly+96+__svml_slog10_data_internal(%rip), %xmm9 |
63 | cvtdq2ps %xmm1, %xmm12 |
64 | cmpltps MinNorm+__svml_slog10_data_internal(%rip), %xmm4 |
65 | cmpnleps MaxNorm+__svml_slog10_data_internal(%rip), %xmm3 |
66 | subps One+__svml_slog10_data_internal(%rip), %xmm10 |
67 | mulps %xmm10, %xmm5 |
68 | movaps %xmm10, %xmm8 |
69 | mulps %xmm10, %xmm6 |
70 | mulps %xmm10, %xmm8 |
71 | addps sPoly+16+__svml_slog10_data_internal(%rip), %xmm5 |
72 | mulps %xmm10, %xmm7 |
73 | addps sPoly+48+__svml_slog10_data_internal(%rip), %xmm6 |
74 | mulps %xmm10, %xmm9 |
75 | mulps %xmm8, %xmm5 |
76 | addps sPoly+80+__svml_slog10_data_internal(%rip), %xmm7 |
77 | addps sPoly+112+__svml_slog10_data_internal(%rip), %xmm9 |
78 | addps %xmm5, %xmm6 |
79 | mulps %xmm8, %xmm6 |
80 | orps %xmm3, %xmm4 |
81 | |
82 | /* combine and get argument value range mask */ |
83 | movmskps %xmm4, %edx |
84 | movups L2L+__svml_slog10_data_internal(%rip), %xmm1 |
85 | addps %xmm6, %xmm7 |
86 | mulps %xmm12, %xmm1 |
87 | mulps %xmm7, %xmm8 |
88 | movups L2H+__svml_slog10_data_internal(%rip), %xmm11 |
89 | addps %xmm8, %xmm9 |
90 | mulps %xmm11, %xmm12 |
91 | mulps %xmm10, %xmm9 |
92 | addps sPoly+128+__svml_slog10_data_internal(%rip), %xmm9 |
93 | mulps %xmm9, %xmm10 |
94 | addps %xmm10, %xmm1 |
95 | addps %xmm12, %xmm1 |
96 | testl %edx, %edx |
97 | |
98 | /* Go to special inputs processing branch */ |
99 | jne L(SPECIAL_VALUES_BRANCH) |
100 | # LOE rbx rbp r12 r13 r14 r15 edx xmm0 xmm1 |
101 | |
102 | /* Restore registers |
103 | * and exit the function |
104 | */ |
105 | |
106 | L(EXIT): |
107 | movaps %xmm1, %xmm0 |
108 | addq $72, %rsp |
109 | cfi_def_cfa_offset(8) |
110 | ret |
111 | cfi_def_cfa_offset(80) |
112 | |
113 | /* Branch to process |
114 | * special inputs |
115 | */ |
116 | |
117 | L(SPECIAL_VALUES_BRANCH): |
118 | movups %xmm0, 32(%rsp) |
119 | movups %xmm1, 48(%rsp) |
120 | # LOE rbx rbp r12 r13 r14 r15 edx |
121 | |
122 | xorl %eax, %eax |
123 | movq %r12, 16(%rsp) |
124 | cfi_offset(12, -64) |
125 | movl %eax, %r12d |
126 | movq %r13, 8(%rsp) |
127 | cfi_offset(13, -72) |
128 | movl %edx, %r13d |
129 | movq %r14, (%rsp) |
130 | cfi_offset(14, -80) |
131 | # LOE rbx rbp r15 r12d r13d |
132 | |
133 | /* Range mask |
134 | * bits check |
135 | */ |
136 | |
137 | L(RANGEMASK_CHECK): |
138 | btl %r12d, %r13d |
139 | |
140 | /* Call scalar math function */ |
141 | jc L(SCALAR_MATH_CALL) |
142 | # LOE rbx rbp r15 r12d r13d |
143 | |
144 | /* Special inputs |
145 | * processing loop |
146 | */ |
147 | |
148 | L(SPECIAL_VALUES_LOOP): |
149 | incl %r12d |
150 | cmpl $4, %r12d |
151 | |
152 | /* Check bits in range mask */ |
153 | jl L(RANGEMASK_CHECK) |
154 | # LOE rbx rbp r15 r12d r13d |
155 | |
156 | movq 16(%rsp), %r12 |
157 | cfi_restore(12) |
158 | movq 8(%rsp), %r13 |
159 | cfi_restore(13) |
160 | movq (%rsp), %r14 |
161 | cfi_restore(14) |
162 | movups 48(%rsp), %xmm1 |
163 | |
164 | /* Go to exit */ |
165 | jmp L(EXIT) |
166 | cfi_offset(12, -64) |
167 | cfi_offset(13, -72) |
168 | cfi_offset(14, -80) |
169 | # LOE rbx rbp r12 r13 r14 r15 xmm1 |
170 | |
171 | /* Scalar math function call |
172 | * to process special input |
173 | */ |
174 | |
175 | L(SCALAR_MATH_CALL): |
176 | movl %r12d, %r14d |
177 | movss 32(%rsp, %r14, 4), %xmm0 |
178 | call log10f@PLT |
179 | # LOE rbx rbp r14 r15 r12d r13d xmm0 |
180 | |
181 | movss %xmm0, 48(%rsp, %r14, 4) |
182 | |
183 | /* Process special inputs in loop */ |
184 | jmp L(SPECIAL_VALUES_LOOP) |
185 | # LOE rbx rbp r15 r12d r13d |
186 | END(_ZGVbN4v_log10f_sse4) |
187 | |
188 | .section .rodata, "a" |
189 | .align 16 |
190 | |
191 | #ifdef __svml_slog10_data_internal_typedef |
192 | typedef unsigned int VUINT32; |
193 | typedef struct { |
194 | __declspec(align(16)) VUINT32 MinNorm[4][1]; |
195 | __declspec(align(16)) VUINT32 MaxNorm[4][1]; |
196 | __declspec(align(16)) VUINT32 L2H[4][1]; |
197 | __declspec(align(16)) VUINT32 L2L[4][1]; |
198 | __declspec(align(16)) VUINT32 iBrkValue[4][1]; |
199 | __declspec(align(16)) VUINT32 iOffExpoMask[4][1]; |
200 | __declspec(align(16)) VUINT32 One[4][1]; |
201 | __declspec(align(16)) VUINT32 sPoly[9][4][1]; |
202 | __declspec(align(16)) VUINT32 L2[4][1]; |
203 | } __svml_slog10_data_internal; |
204 | #endif |
205 | __svml_slog10_data_internal: |
206 | /* MinNorm */ |
207 | .long 0x00800000, 0x00800000, 0x00800000, 0x00800000 |
208 | /* MaxNorm */ |
209 | .align 16 |
210 | .long 0x7f7fffff, 0x7f7fffff, 0x7f7fffff, 0x7f7fffff |
211 | /* L2H */ |
212 | .align 16 |
213 | .long 0x3e9a2100, 0x3e9a2100, 0x3e9a2100, 0x3e9a2100 |
214 | /* L2L */ |
215 | .align 16 |
216 | .long 0xb64AF600, 0xb64AF600, 0xb64AF600, 0xb64AF600 |
217 | /* iBrkValue = SP 2/3 */ |
218 | .align 16 |
219 | .long 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab |
220 | /* iOffExpoMask = SP significand mask */ |
221 | .align 16 |
222 | .long 0x007fffff, 0x007fffff, 0x007fffff, 0x007fffff |
223 | /* sOne = SP 1.0 */ |
224 | .align 16 |
225 | .long 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 |
226 | /* spoly[9] */ |
227 | .align 16 |
228 | .long 0x3d8063B4, 0x3d8063B4, 0x3d8063B4, 0x3d8063B4 /* coeff9 */ |
229 | .long 0xbd890073, 0xbd890073, 0xbd890073, 0xbd890073 /* coeff8 */ |
230 | .long 0x3d775317, 0x3d775317, 0x3d775317, 0x3d775317 /* coeff7 */ |
231 | .long 0xbd91FB27, 0xbd91FB27, 0xbd91FB27, 0xbd91FB27 /* coeff6 */ |
232 | .long 0x3dB20B96, 0x3dB20B96, 0x3dB20B96, 0x3dB20B96 /* coeff5 */ |
233 | .long 0xbdDE6E20, 0xbdDE6E20, 0xbdDE6E20, 0xbdDE6E20 /* coeff4 */ |
234 | .long 0x3e143CE5, 0x3e143CE5, 0x3e143CE5, 0x3e143CE5 /* coeff3 */ |
235 | .long 0xbe5E5BC5, 0xbe5E5BC5, 0xbe5E5BC5, 0xbe5E5BC5 /* coeff2 */ |
236 | .long 0x3eDE5BD9, 0x3eDE5BD9, 0x3eDE5BD9, 0x3eDE5BD9 /* coeff1 */ |
237 | /* L2 */ |
238 | .align 16 |
239 | .long 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b |
240 | .align 16 |
241 | .type __svml_slog10_data_internal, @object |
242 | .size __svml_slog10_data_internal, .-__svml_slog10_data_internal |
243 | |