1 | /* Function log10 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_dlog10_data_internal_avx512 |
31 | */ |
32 | #define Log_tbl 0 |
33 | #define One 128 |
34 | #define C075 192 |
35 | #define poly_coeff9 256 |
36 | #define poly_coeff8 320 |
37 | #define poly_coeff7 384 |
38 | #define poly_coeff6 448 |
39 | #define poly_coeff5 512 |
40 | #define poly_coeff4 576 |
41 | #define poly_coeff3 640 |
42 | #define poly_coeff2 704 |
43 | #define poly_coeff1 768 |
44 | #define L2 832 |
45 | |
46 | #include <sysdep.h> |
47 | |
48 | .section .text.evex512, "ax" , @progbits |
49 | ENTRY(_ZGVeN8v_log10_skx) |
50 | pushq %rbp |
51 | cfi_def_cfa_offset(16) |
52 | movq %rsp, %rbp |
53 | cfi_def_cfa(6, 16) |
54 | cfi_offset(6, -16) |
55 | andq $-64, %rsp |
56 | subq $192, %rsp |
57 | vmovaps %zmm0, %zmm7 |
58 | vgetmantpd $8, {sae}, %zmm7, %zmm6 |
59 | vmovups One+__svml_dlog10_data_internal_avx512(%rip), %zmm3 |
60 | vmovups poly_coeff5+__svml_dlog10_data_internal_avx512(%rip), %zmm12 |
61 | vmovups poly_coeff3+__svml_dlog10_data_internal_avx512(%rip), %zmm13 |
62 | |
63 | /* Start polynomial evaluation */ |
64 | vmovups poly_coeff9+__svml_dlog10_data_internal_avx512(%rip), %zmm10 |
65 | vmovups poly_coeff8+__svml_dlog10_data_internal_avx512(%rip), %zmm1 |
66 | vmovups poly_coeff7+__svml_dlog10_data_internal_avx512(%rip), %zmm11 |
67 | vmovups poly_coeff6+__svml_dlog10_data_internal_avx512(%rip), %zmm14 |
68 | |
69 | /* Prepare exponent correction: DblRcp<0.75? */ |
70 | vmovups C075+__svml_dlog10_data_internal_avx512(%rip), %zmm2 |
71 | |
72 | /* Table lookup */ |
73 | vmovups __svml_dlog10_data_internal_avx512(%rip), %zmm5 |
74 | |
75 | /* GetExp(x) */ |
76 | vgetexppd {sae}, %zmm7, %zmm0 |
77 | |
78 | /* DblRcp ~ 1/Mantissa */ |
79 | vrcp14pd %zmm6, %zmm8 |
80 | |
81 | /* x<=0? */ |
82 | vfpclasspd $94, %zmm7, %k0 |
83 | |
84 | /* round DblRcp to 4 fractional bits (RN mode, no Precision exception) */ |
85 | vrndscalepd $88, {sae}, %zmm8, %zmm4 |
86 | vmovups poly_coeff4+__svml_dlog10_data_internal_avx512(%rip), %zmm8 |
87 | kmovw %k0, %edx |
88 | |
89 | /* Reduced argument: R = DblRcp*Mantissa - 1 */ |
90 | vfmsub213pd {rn-sae}, %zmm3, %zmm4, %zmm6 |
91 | vcmppd $17, {sae}, %zmm2, %zmm4, %k1 |
92 | vfmadd231pd {rn-sae}, %zmm6, %zmm12, %zmm8 |
93 | vmovups poly_coeff2+__svml_dlog10_data_internal_avx512(%rip), %zmm12 |
94 | vfmadd231pd {rn-sae}, %zmm6, %zmm10, %zmm1 |
95 | vfmadd231pd {rn-sae}, %zmm6, %zmm11, %zmm14 |
96 | vmovups poly_coeff1+__svml_dlog10_data_internal_avx512(%rip), %zmm2 |
97 | |
98 | /* R^2 */ |
99 | vmulpd {rn-sae}, %zmm6, %zmm6, %zmm15 |
100 | vfmadd231pd {rn-sae}, %zmm6, %zmm13, %zmm12 |
101 | |
102 | /* Prepare table index */ |
103 | vpsrlq $48, %zmm4, %zmm9 |
104 | |
105 | /* add 1 to Expon if DblRcp<0.75 */ |
106 | vaddpd {rn-sae}, %zmm3, %zmm0, %zmm0{%k1} |
107 | vmulpd {rn-sae}, %zmm15, %zmm15, %zmm13 |
108 | vfmadd213pd {rn-sae}, %zmm14, %zmm15, %zmm1 |
109 | vfmadd213pd {rn-sae}, %zmm12, %zmm15, %zmm8 |
110 | vpermt2pd Log_tbl+64+__svml_dlog10_data_internal_avx512(%rip), %zmm9, %zmm5 |
111 | |
112 | /* polynomial */ |
113 | vfmadd213pd {rn-sae}, %zmm8, %zmm13, %zmm1 |
114 | vfmadd213pd {rn-sae}, %zmm2, %zmm6, %zmm1 |
115 | vfmadd213pd {rn-sae}, %zmm5, %zmm1, %zmm6 |
116 | vmovups L2+__svml_dlog10_data_internal_avx512(%rip), %zmm1 |
117 | vfmadd213pd {rn-sae}, %zmm6, %zmm1, %zmm0 |
118 | testl %edx, %edx |
119 | |
120 | /* Go to special inputs processing branch */ |
121 | jne L(SPECIAL_VALUES_BRANCH) |
122 | # LOE rbx r12 r13 r14 r15 edx zmm0 zmm7 |
123 | |
124 | /* Restore registers |
125 | * and exit the function |
126 | */ |
127 | |
128 | L(EXIT): |
129 | movq %rbp, %rsp |
130 | popq %rbp |
131 | cfi_def_cfa(7, 8) |
132 | cfi_restore(6) |
133 | ret |
134 | cfi_def_cfa(6, 16) |
135 | cfi_offset(6, -16) |
136 | |
137 | /* Branch to process |
138 | * special inputs |
139 | */ |
140 | |
141 | L(SPECIAL_VALUES_BRANCH): |
142 | vmovups %zmm7, 64(%rsp) |
143 | vmovups %zmm0, 128(%rsp) |
144 | # LOE rbx r12 r13 r14 r15 edx zmm0 |
145 | |
146 | xorl %eax, %eax |
147 | # LOE rbx r12 r13 r14 r15 eax edx |
148 | |
149 | vzeroupper |
150 | movq %r12, 16(%rsp) |
151 | /* 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) */ |
152 | .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22 |
153 | movl %eax, %r12d |
154 | movq %r13, 8(%rsp) |
155 | /* 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) */ |
156 | .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22 |
157 | movl %edx, %r13d |
158 | movq %r14, (%rsp) |
159 | /* 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) */ |
160 | .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22 |
161 | # LOE rbx r15 r12d r13d |
162 | |
163 | /* Range mask |
164 | * bits check |
165 | */ |
166 | |
167 | L(RANGEMASK_CHECK): |
168 | btl %r12d, %r13d |
169 | |
170 | /* Call scalar math function */ |
171 | jc L(SCALAR_MATH_CALL) |
172 | # LOE rbx r15 r12d r13d |
173 | |
174 | /* Special inputs |
175 | * processing loop |
176 | */ |
177 | |
178 | L(SPECIAL_VALUES_LOOP): |
179 | incl %r12d |
180 | cmpl $8, %r12d |
181 | |
182 | /* Check bits in range mask */ |
183 | jl L(RANGEMASK_CHECK) |
184 | # LOE rbx r15 r12d r13d |
185 | |
186 | movq 16(%rsp), %r12 |
187 | cfi_restore(12) |
188 | movq 8(%rsp), %r13 |
189 | cfi_restore(13) |
190 | movq (%rsp), %r14 |
191 | cfi_restore(14) |
192 | vmovups 128(%rsp), %zmm0 |
193 | |
194 | /* Go to exit */ |
195 | jmp L(EXIT) |
196 | /* 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) */ |
197 | .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22 |
198 | /* 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) */ |
199 | .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22 |
200 | /* 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) */ |
201 | .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22 |
202 | # LOE rbx r12 r13 r14 r15 zmm0 |
203 | |
204 | /* Scalar math function call |
205 | * to process special input |
206 | */ |
207 | |
208 | L(SCALAR_MATH_CALL): |
209 | movl %r12d, %r14d |
210 | vmovsd 64(%rsp, %r14, 8), %xmm0 |
211 | call log10@PLT |
212 | # LOE rbx r14 r15 r12d r13d xmm0 |
213 | |
214 | vmovsd %xmm0, 128(%rsp, %r14, 8) |
215 | |
216 | /* Process special inputs in loop */ |
217 | jmp L(SPECIAL_VALUES_LOOP) |
218 | # LOE rbx r15 r12d r13d |
219 | END(_ZGVeN8v_log10_skx) |
220 | |
221 | .section .rodata, "a" |
222 | .align 64 |
223 | |
224 | #ifdef __svml_dlog10_data_internal_avx512_typedef |
225 | typedef unsigned int VUINT32; |
226 | typedef struct { |
227 | __declspec(align(64)) VUINT32 Log_tbl[16][2]; |
228 | __declspec(align(64)) VUINT32 One[8][2]; |
229 | __declspec(align(64)) VUINT32 C075[8][2]; |
230 | __declspec(align(64)) VUINT32 poly_coeff9[8][2]; |
231 | __declspec(align(64)) VUINT32 poly_coeff8[8][2]; |
232 | __declspec(align(64)) VUINT32 poly_coeff7[8][2]; |
233 | __declspec(align(64)) VUINT32 poly_coeff6[8][2]; |
234 | __declspec(align(64)) VUINT32 poly_coeff5[8][2]; |
235 | __declspec(align(64)) VUINT32 poly_coeff4[8][2]; |
236 | __declspec(align(64)) VUINT32 poly_coeff3[8][2]; |
237 | __declspec(align(64)) VUINT32 poly_coeff2[8][2]; |
238 | __declspec(align(64)) VUINT32 poly_coeff1[8][2]; |
239 | __declspec(align(64)) VUINT32 L2[8][2]; |
240 | } __svml_dlog10_data_internal_avx512; |
241 | #endif |
242 | __svml_dlog10_data_internal_avx512: |
243 | /* Log_tbl */ |
244 | .quad 0x0000000000000000 |
245 | .quad 0xbf9af5f92b00e610 |
246 | .quad 0xbfaa30a9d609efea |
247 | .quad 0xbfb31b3055c47118 |
248 | .quad 0xbfb8cf183886480d |
249 | .quad 0xbfbe3bc1ab0e19fe |
250 | .quad 0xbfc1b3e71ec94f7b |
251 | .quad 0xbfc42c7e7fe3fc02 |
252 | .quad 0x3fbffbfc2bbc7803 |
253 | .quad 0x3fbb721cd17157e3 |
254 | .quad 0x3fb715d0ce367afc |
255 | .quad 0x3fb2e3a740b7800f |
256 | .quad 0x3fadb11ed766abf4 |
257 | .quad 0x3fa5e3966b7e9295 |
258 | .quad 0x3f9cb38fccd8bfdb |
259 | .quad 0x3f8c3d0837784c41 |
260 | /* One */ |
261 | .align 64 |
262 | .quad 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000 |
263 | /* 0.75 */ |
264 | .align 64 |
265 | .quad 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000 |
266 | /* poly_coeff9 */ |
267 | .align 64 |
268 | .quad 0x3fa8c2d828480370, 0x3fa8c2d828480370, 0x3fa8c2d828480370, 0x3fa8c2d828480370, 0x3fa8c2d828480370, 0x3fa8c2d828480370, 0x3fa8c2d828480370, 0x3fa8c2d828480370 |
269 | /* poly_coeff8 */ |
270 | .align 64 |
271 | .quad 0xbfabd80d96029814, 0xbfabd80d96029814, 0xbfabd80d96029814, 0xbfabd80d96029814, 0xbfabd80d96029814, 0xbfabd80d96029814, 0xbfabd80d96029814, 0xbfabd80d96029814 |
272 | /* poly_coeff7 */ |
273 | .align 64 |
274 | .quad 0x3fafc3f6f38b58a2, 0x3fafc3f6f38b58a2, 0x3fafc3f6f38b58a2, 0x3fafc3f6f38b58a2, 0x3fafc3f6f38b58a2, 0x3fafc3f6f38b58a2, 0x3fafc3f6f38b58a2, 0x3fafc3f6f38b58a2 |
275 | /* poly_coeff6 */ |
276 | .align 64 |
277 | .quad 0xbfb287a63464dc80, 0xbfb287a63464dc80, 0xbfb287a63464dc80, 0xbfb287a63464dc80, 0xbfb287a63464dc80, 0xbfb287a63464dc80, 0xbfb287a63464dc80, 0xbfb287a63464dc80 |
278 | /* poly_coeff5 */ |
279 | .align 64 |
280 | .quad 0x3fb63c62777f27d9, 0x3fb63c62777f27d9, 0x3fb63c62777f27d9, 0x3fb63c62777f27d9, 0x3fb63c62777f27d9, 0x3fb63c62777f27d9, 0x3fb63c62777f27d9, 0x3fb63c62777f27d9 |
281 | /* poly_coeff4 */ |
282 | .align 64 |
283 | .quad 0xbfbbcb7b153c06a3, 0xbfbbcb7b153c06a3, 0xbfbbcb7b153c06a3, 0xbfbbcb7b153c06a3, 0xbfbbcb7b153c06a3, 0xbfbbcb7b153c06a3, 0xbfbbcb7b153c06a3, 0xbfbbcb7b153c06a3 |
284 | /* poly_coeff3 */ |
285 | .align 64 |
286 | .quad 0x3fc287a7636f428c, 0x3fc287a7636f428c, 0x3fc287a7636f428c, 0x3fc287a7636f428c, 0x3fc287a7636f428c, 0x3fc287a7636f428c, 0x3fc287a7636f428c, 0x3fc287a7636f428c |
287 | /* poly_coeff2 */ |
288 | .align 64 |
289 | .quad 0xbfcbcb7b1526e4db, 0xbfcbcb7b1526e4db, 0xbfcbcb7b1526e4db, 0xbfcbcb7b1526e4db, 0xbfcbcb7b1526e4db, 0xbfcbcb7b1526e4db, 0xbfcbcb7b1526e4db, 0xbfcbcb7b1526e4db |
290 | /* poly_coeff1 */ |
291 | .align 64 |
292 | .quad 0x3fdbcb7b1526e50e, 0x3fdbcb7b1526e50e, 0x3fdbcb7b1526e50e, 0x3fdbcb7b1526e50e, 0x3fdbcb7b1526e50e, 0x3fdbcb7b1526e50e, 0x3fdbcb7b1526e50e, 0x3fdbcb7b1526e50e |
293 | /* L2 */ |
294 | .align 64 |
295 | .quad 0x3fd34413509f79ff, 0x3fd34413509f79ff, 0x3fd34413509f79ff, 0x3fd34413509f79ff, 0x3fd34413509f79ff, 0x3fd34413509f79ff, 0x3fd34413509f79ff, 0x3fd34413509f79ff |
296 | .align 64 |
297 | .type __svml_dlog10_data_internal_avx512, @object |
298 | .size __svml_dlog10_data_internal_avx512, .-__svml_dlog10_data_internal_avx512 |
299 | |