1 | /* Function log2f 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 | * 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 |
31 | */ |
32 | #define MinNorm 0 |
33 | #define MaxNorm 16 |
34 | #define iBrkValue 32 |
35 | #define iOffExpoMask 48 |
36 | #define One 64 |
37 | #define sPoly 80 |
38 | |
39 | #include <sysdep.h> |
40 | |
41 | .section .text.sse4, "ax" , @progbits |
42 | ENTRY(_ZGVbN4v_log2f_sse4) |
43 | subq $72, %rsp |
44 | cfi_def_cfa_offset(80) |
45 | movaps %xmm0, %xmm1 |
46 | |
47 | /* reduction: compute r, n */ |
48 | movdqu iBrkValue+__svml_slog2_data_internal(%rip), %xmm2 |
49 | movaps %xmm0, %xmm4 |
50 | movdqu iOffExpoMask+__svml_slog2_data_internal(%rip), %xmm10 |
51 | psubd %xmm2, %xmm1 |
52 | pand %xmm1, %xmm10 |
53 | movaps %xmm0, %xmm3 |
54 | paddd %xmm2, %xmm10 |
55 | psrad $23, %xmm1 |
56 | movups sPoly+__svml_slog2_data_internal(%rip), %xmm5 |
57 | movups sPoly+32+__svml_slog2_data_internal(%rip), %xmm6 |
58 | movups sPoly+64+__svml_slog2_data_internal(%rip), %xmm7 |
59 | movups sPoly+96+__svml_slog2_data_internal(%rip), %xmm9 |
60 | cmpltps MinNorm+__svml_slog2_data_internal(%rip), %xmm4 |
61 | cmpnleps MaxNorm+__svml_slog2_data_internal(%rip), %xmm3 |
62 | cvtdq2ps %xmm1, %xmm1 |
63 | subps One+__svml_slog2_data_internal(%rip), %xmm10 |
64 | mulps %xmm10, %xmm5 |
65 | movaps %xmm10, %xmm8 |
66 | mulps %xmm10, %xmm6 |
67 | mulps %xmm10, %xmm8 |
68 | addps sPoly+16+__svml_slog2_data_internal(%rip), %xmm5 |
69 | mulps %xmm10, %xmm7 |
70 | addps sPoly+48+__svml_slog2_data_internal(%rip), %xmm6 |
71 | mulps %xmm10, %xmm9 |
72 | mulps %xmm8, %xmm5 |
73 | addps sPoly+80+__svml_slog2_data_internal(%rip), %xmm7 |
74 | addps sPoly+112+__svml_slog2_data_internal(%rip), %xmm9 |
75 | addps %xmm5, %xmm6 |
76 | mulps %xmm8, %xmm6 |
77 | orps %xmm3, %xmm4 |
78 | |
79 | /* combine and get argument value range mask */ |
80 | movmskps %xmm4, %edx |
81 | addps %xmm6, %xmm7 |
82 | mulps %xmm7, %xmm8 |
83 | addps %xmm8, %xmm9 |
84 | mulps %xmm10, %xmm9 |
85 | addps sPoly+128+__svml_slog2_data_internal(%rip), %xmm9 |
86 | mulps %xmm9, %xmm10 |
87 | addps %xmm10, %xmm1 |
88 | testl %edx, %edx |
89 | |
90 | /* Go to special inputs processing branch */ |
91 | jne L(SPECIAL_VALUES_BRANCH) |
92 | # LOE rbx rbp r12 r13 r14 r15 edx xmm0 xmm1 |
93 | |
94 | /* Restore registers |
95 | * and exit the function |
96 | */ |
97 | |
98 | L(EXIT): |
99 | movaps %xmm1, %xmm0 |
100 | addq $72, %rsp |
101 | cfi_def_cfa_offset(8) |
102 | ret |
103 | cfi_def_cfa_offset(80) |
104 | |
105 | /* Branch to process |
106 | * special inputs |
107 | */ |
108 | |
109 | L(SPECIAL_VALUES_BRANCH): |
110 | movups %xmm0, 32(%rsp) |
111 | movups %xmm1, 48(%rsp) |
112 | # LOE rbx rbp r12 r13 r14 r15 edx |
113 | |
114 | xorl %eax, %eax |
115 | movq %r12, 16(%rsp) |
116 | cfi_offset(12, -64) |
117 | movl %eax, %r12d |
118 | movq %r13, 8(%rsp) |
119 | cfi_offset(13, -72) |
120 | movl %edx, %r13d |
121 | movq %r14, (%rsp) |
122 | cfi_offset(14, -80) |
123 | # LOE rbx rbp r15 r12d r13d |
124 | |
125 | /* Range mask |
126 | * bits check |
127 | */ |
128 | |
129 | L(RANGEMASK_CHECK): |
130 | btl %r12d, %r13d |
131 | |
132 | /* Call scalar math function */ |
133 | jc L(SCALAR_MATH_CALL) |
134 | # LOE rbx rbp r15 r12d r13d |
135 | |
136 | /* Special inputs |
137 | * processing loop |
138 | */ |
139 | |
140 | L(SPECIAL_VALUES_LOOP): |
141 | incl %r12d |
142 | cmpl $4, %r12d |
143 | |
144 | /* Check bits in range mask */ |
145 | jl L(RANGEMASK_CHECK) |
146 | # LOE rbx rbp r15 r12d r13d |
147 | |
148 | movq 16(%rsp), %r12 |
149 | cfi_restore(12) |
150 | movq 8(%rsp), %r13 |
151 | cfi_restore(13) |
152 | movq (%rsp), %r14 |
153 | cfi_restore(14) |
154 | movups 48(%rsp), %xmm1 |
155 | |
156 | /* Go to exit */ |
157 | jmp L(EXIT) |
158 | cfi_offset(12, -64) |
159 | cfi_offset(13, -72) |
160 | cfi_offset(14, -80) |
161 | # LOE rbx rbp r12 r13 r14 r15 xmm1 |
162 | |
163 | /* Scalar math function call |
164 | * to process special input |
165 | */ |
166 | |
167 | L(SCALAR_MATH_CALL): |
168 | movl %r12d, %r14d |
169 | movss 32(%rsp, %r14, 4), %xmm0 |
170 | call log2f@PLT |
171 | # LOE rbx rbp r14 r15 r12d r13d xmm0 |
172 | |
173 | movss %xmm0, 48(%rsp, %r14, 4) |
174 | |
175 | /* Process special inputs in loop */ |
176 | jmp L(SPECIAL_VALUES_LOOP) |
177 | # LOE rbx rbp r15 r12d r13d |
178 | END(_ZGVbN4v_log2f_sse4) |
179 | |
180 | .section .rodata, "a" |
181 | .align 16 |
182 | |
183 | #ifdef __svml_slog2_data_internal_typedef |
184 | typedef unsigned int VUINT32; |
185 | typedef struct { |
186 | __declspec(align(16)) VUINT32 MinNorm[4][1]; |
187 | __declspec(align(16)) VUINT32 MaxNorm[4][1]; |
188 | __declspec(align(16)) VUINT32 iBrkValue[4][1]; |
189 | __declspec(align(16)) VUINT32 iOffExpoMask[4][1]; |
190 | __declspec(align(16)) VUINT32 One[4][1]; |
191 | __declspec(align(16)) VUINT32 sPoly[9][4][1]; |
192 | } __svml_slog2_data_internal; |
193 | #endif |
194 | __svml_slog2_data_internal: |
195 | /* MinNorm */ |
196 | .long 0x00800000, 0x00800000, 0x00800000, 0x00800000 |
197 | /* MaxNorm */ |
198 | .align 16 |
199 | .long 0x7f7fffff, 0x7f7fffff, 0x7f7fffff, 0x7f7fffff |
200 | /* iBrkValue = SP 2/3 */ |
201 | .align 16 |
202 | .long 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab |
203 | /* iOffExpoMask = SP significand mask */ |
204 | .align 16 |
205 | .long 0x007fffff, 0x007fffff, 0x007fffff, 0x007fffff |
206 | /* sOne = SP 1.0 */ |
207 | .align 16 |
208 | .long 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 |
209 | /* spoly[9] */ |
210 | .align 16 |
211 | .long 0x3e554012, 0x3e554012, 0x3e554012, 0x3e554012 /* coeff9 */ |
212 | .long 0xbe638E14, 0xbe638E14, 0xbe638E14, 0xbe638E14 /* coeff8 */ |
213 | .long 0x3e4D660B, 0x3e4D660B, 0x3e4D660B, 0x3e4D660B /* coeff7 */ |
214 | .long 0xbe727824, 0xbe727824, 0xbe727824, 0xbe727824 /* coeff6 */ |
215 | .long 0x3e93DD07, 0x3e93DD07, 0x3e93DD07, 0x3e93DD07 /* coeff5 */ |
216 | .long 0xbeB8B969, 0xbeB8B969, 0xbeB8B969, 0xbeB8B969 /* coeff4 */ |
217 | .long 0x3eF637C0, 0x3eF637C0, 0x3eF637C0, 0x3eF637C0 /* coeff3 */ |
218 | .long 0xbf38AA2B, 0xbf38AA2B, 0xbf38AA2B, 0xbf38AA2B /* coeff2 */ |
219 | .long 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B /* coeff1 */ |
220 | .align 16 |
221 | .type __svml_slog2_data_internal, @object |
222 | .size __svml_slog2_data_internal, .-__svml_slog2_data_internal |
223 | |