1/* Function logf vectorized with AVX-512. KNL and SKX versions.
2 Copyright (C) 2014-2024 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#include <sysdep.h>
20#include "svml_s_logf_data.h"
21#include "svml_s_wrapper_impl.h"
22
23 .section .text.evex512, "ax", @progbits
24ENTRY (_ZGVeN16v_logf_knl)
25/*
26 ALGORITHM DESCRIPTION:
27
28 log(x) = exponent_x*log(2) + log(mantissa_x), if mantissa_x<4/3
29 log(x) = (exponent_x+1)*log(2) + log(0.5*mantissa_x), if mantissa_x>4/3
30
31 R = mantissa_x - 1, if mantissa_x<4/3
32 R = 0.5*mantissa_x - 1, if mantissa_x>4/3
33 |R|< 1/3
34
35 log(1+R) is approximated as a polynomial: degree 9 for 1-ulp,
36 degree 7 for 4-ulp, degree 3 for half-precision. */
37
38 pushq %rbp
39 cfi_adjust_cfa_offset (8)
40 cfi_rel_offset (%rbp, 0)
41 movq %rsp, %rbp
42 cfi_def_cfa_register (%rbp)
43 andq $-64, %rsp
44 subq $1280, %rsp
45 movq __svml_slog_data@GOTPCREL(%rip), %rax
46 movl $-1, %ecx
47
48/* reduction: compute r,n */
49 vpsubd _iBrkValue(%rax), %zmm0, %zmm2
50 vmovups _sPoly_7(%rax), %zmm7
51 vpandd _iOffExpoMask(%rax), %zmm2, %zmm3
52
53/* exponent_x (mantissa_x<4/3) or exponent_x+1 (mantissa_x>4/3) */
54 vpsrad $23, %zmm2, %zmm4
55
56/* check for working range,
57 set special argument mask (denormals/zero/Inf/NaN)
58 */
59 vpaddd _iHiDelta(%rax), %zmm0, %zmm1
60
61/* mantissa_x (mantissa_x<4/3), or 0.5*mantissa_x (mantissa_x>4/3) */
62 vpaddd _iBrkValue(%rax), %zmm3, %zmm6
63 vpcmpd $1, _iLoRange(%rax), %zmm1, %k1
64 vcvtdq2ps {rn-sae}, %zmm4, %zmm1
65
66/* reduced argument R */
67 vsubps _sOne(%rax), %zmm6, %zmm8
68 vpbroadcastd %ecx, %zmm5{%k1}{z}
69
70/* polynomial evaluation starts here */
71 vfmadd213ps _sPoly_6(%rax), %zmm8, %zmm7
72 vptestmd %zmm5, %zmm5, %k0
73 kmovw %k0, %ecx
74 vfmadd213ps _sPoly_5(%rax), %zmm8, %zmm7
75 vfmadd213ps _sPoly_4(%rax), %zmm8, %zmm7
76 vfmadd213ps _sPoly_3(%rax), %zmm8, %zmm7
77 vfmadd213ps _sPoly_2(%rax), %zmm8, %zmm7
78 vfmadd213ps _sPoly_1(%rax), %zmm8, %zmm7
79 vmulps %zmm8, %zmm7, %zmm9
80
81/* polynomial evaluation end */
82 vfmadd213ps %zmm8, %zmm8, %zmm9
83
84/*
85 final reconstruction:
86 add exponent_value*log2 to polynomial result
87 */
88 vfmadd132ps _sLn2(%rax), %zmm9, %zmm1
89 testl %ecx, %ecx
90 jne .LBL_1_3
91
92.LBL_1_2:
93 cfi_remember_state
94 vmovaps %zmm1, %zmm0
95 movq %rbp, %rsp
96 cfi_def_cfa_register (%rsp)
97 popq %rbp
98 cfi_adjust_cfa_offset (-8)
99 cfi_restore (%rbp)
100 ret
101
102.LBL_1_3:
103 cfi_restore_state
104 vmovups %zmm0, 1152(%rsp)
105 vmovups %zmm1, 1216(%rsp)
106 je .LBL_1_2
107
108 xorb %dl, %dl
109 kmovw %k4, 1048(%rsp)
110 xorl %eax, %eax
111 kmovw %k5, 1040(%rsp)
112 kmovw %k6, 1032(%rsp)
113 kmovw %k7, 1024(%rsp)
114 vmovups %zmm16, 960(%rsp)
115 vmovups %zmm17, 896(%rsp)
116 vmovups %zmm18, 832(%rsp)
117 vmovups %zmm19, 768(%rsp)
118 vmovups %zmm20, 704(%rsp)
119 vmovups %zmm21, 640(%rsp)
120 vmovups %zmm22, 576(%rsp)
121 vmovups %zmm23, 512(%rsp)
122 vmovups %zmm24, 448(%rsp)
123 vmovups %zmm25, 384(%rsp)
124 vmovups %zmm26, 320(%rsp)
125 vmovups %zmm27, 256(%rsp)
126 vmovups %zmm28, 192(%rsp)
127 vmovups %zmm29, 128(%rsp)
128 vmovups %zmm30, 64(%rsp)
129 vmovups %zmm31, (%rsp)
130 movq %rsi, 1064(%rsp)
131 movq %rdi, 1056(%rsp)
132 movq %r12, 1096(%rsp)
133 cfi_offset_rel_rsp (12, 1096)
134 movb %dl, %r12b
135 movq %r13, 1088(%rsp)
136 cfi_offset_rel_rsp (13, 1088)
137 movl %ecx, %r13d
138 movq %r14, 1080(%rsp)
139 cfi_offset_rel_rsp (14, 1080)
140 movl %eax, %r14d
141 movq %r15, 1072(%rsp)
142 cfi_offset_rel_rsp (15, 1072)
143 cfi_remember_state
144
145.LBL_1_6:
146 btl %r14d, %r13d
147 jc .LBL_1_12
148
149.LBL_1_7:
150 lea 1(%r14), %esi
151 btl %esi, %r13d
152 jc .LBL_1_10
153
154.LBL_1_8:
155 addb $1, %r12b
156 addl $2, %r14d
157 cmpb $16, %r12b
158 jb .LBL_1_6
159
160 kmovw 1048(%rsp), %k4
161 movq 1064(%rsp), %rsi
162 kmovw 1040(%rsp), %k5
163 movq 1056(%rsp), %rdi
164 kmovw 1032(%rsp), %k6
165 movq 1096(%rsp), %r12
166 cfi_restore (%r12)
167 movq 1088(%rsp), %r13
168 cfi_restore (%r13)
169 kmovw 1024(%rsp), %k7
170 vmovups 960(%rsp), %zmm16
171 vmovups 896(%rsp), %zmm17
172 vmovups 832(%rsp), %zmm18
173 vmovups 768(%rsp), %zmm19
174 vmovups 704(%rsp), %zmm20
175 vmovups 640(%rsp), %zmm21
176 vmovups 576(%rsp), %zmm22
177 vmovups 512(%rsp), %zmm23
178 vmovups 448(%rsp), %zmm24
179 vmovups 384(%rsp), %zmm25
180 vmovups 320(%rsp), %zmm26
181 vmovups 256(%rsp), %zmm27
182 vmovups 192(%rsp), %zmm28
183 vmovups 128(%rsp), %zmm29
184 vmovups 64(%rsp), %zmm30
185 vmovups (%rsp), %zmm31
186 movq 1080(%rsp), %r14
187 cfi_restore (%r14)
188 movq 1072(%rsp), %r15
189 cfi_restore (%r15)
190 vmovups 1216(%rsp), %zmm1
191 jmp .LBL_1_2
192
193.LBL_1_10:
194 cfi_restore_state
195 movzbl %r12b, %r15d
196 vmovss 1156(%rsp,%r15,8), %xmm0
197 call JUMPTARGET(logf)
198 vmovss %xmm0, 1220(%rsp,%r15,8)
199 jmp .LBL_1_8
200
201.LBL_1_12:
202 movzbl %r12b, %r15d
203 vmovss 1152(%rsp,%r15,8), %xmm0
204 call JUMPTARGET(logf)
205 vmovss %xmm0, 1216(%rsp,%r15,8)
206 jmp .LBL_1_7
207END (_ZGVeN16v_logf_knl)
208
209ENTRY (_ZGVeN16v_logf_skx)
210/*
211 ALGORITHM DESCRIPTION:
212
213 log(x) = exponent_x*log(2) + log(mantissa_x), if mantissa_x<4/3
214 log(x) = (exponent_x+1)*log(2) + log(0.5*mantissa_x), if mantissa_x>4/3
215
216 R = mantissa_x - 1, if mantissa_x<4/3
217 R = 0.5*mantissa_x - 1, if mantissa_x>4/3
218 |R|< 1/3
219
220 log(1+R) is approximated as a polynomial: degree 9 for 1-ulp,
221 degree 7 for 4-ulp, degree 3 for half-precision. */
222
223 pushq %rbp
224 cfi_adjust_cfa_offset (8)
225 cfi_rel_offset (%rbp, 0)
226 movq %rsp, %rbp
227 cfi_def_cfa_register (%rbp)
228 andq $-64, %rsp
229 subq $1280, %rsp
230 movq __svml_slog_data@GOTPCREL(%rip), %rax
231 vpternlogd $0xff, %zmm6, %zmm6, %zmm6
232 vmovups _iBrkValue(%rax), %zmm4
233 vmovups _sPoly_7(%rax), %zmm8
234
235/*
236 check for working range,
237 set special argument mask (denormals/zero/Inf/NaN)
238 */
239 vpaddd _iHiDelta(%rax), %zmm0, %zmm1
240
241/* reduction: compute r,n */
242 vpsubd %zmm4, %zmm0, %zmm2
243 vpcmpd $5, _iLoRange(%rax), %zmm1, %k1
244
245/* exponent_x (mantissa_x<4/3) or exponent_x+1 (mantissa_x>4/3) */
246 vpsrad $23, %zmm2, %zmm5
247 vpandd _iOffExpoMask(%rax), %zmm2, %zmm3
248
249/* mantissa_x (mantissa_x<4/3), or 0.5*mantissa_x (mantissa_x>4/3) */
250 vpaddd %zmm4, %zmm3, %zmm7
251
252/* reduced argument R */
253 vsubps _sOne(%rax), %zmm7, %zmm9
254
255/* polynomial evaluation starts here */
256 vfmadd213ps _sPoly_6(%rax), %zmm9, %zmm8
257 vfmadd213ps _sPoly_5(%rax), %zmm9, %zmm8
258 vfmadd213ps _sPoly_4(%rax), %zmm9, %zmm8
259 vfmadd213ps _sPoly_3(%rax), %zmm9, %zmm8
260 vfmadd213ps _sPoly_2(%rax), %zmm9, %zmm8
261 vfmadd213ps _sPoly_1(%rax), %zmm9, %zmm8
262 vmulps %zmm9, %zmm8, %zmm10
263
264/* polynomial evaluation end */
265 vfmadd213ps %zmm9, %zmm9, %zmm10
266 vpandnd %zmm1, %zmm1, %zmm6{%k1}
267 vptestmd %zmm6, %zmm6, %k0
268 vcvtdq2ps {rn-sae}, %zmm5, %zmm1
269 kmovw %k0, %ecx
270
271/*
272 final reconstruction:
273 add exponent_value*log2 to polynomial result
274 */
275 vfmadd132ps _sLn2(%rax), %zmm10, %zmm1
276 testl %ecx, %ecx
277 jne .LBL_2_3
278
279.LBL_2_2:
280 cfi_remember_state
281 vmovaps %zmm1, %zmm0
282 movq %rbp, %rsp
283 cfi_def_cfa_register (%rsp)
284 popq %rbp
285 cfi_adjust_cfa_offset (-8)
286 cfi_restore (%rbp)
287 ret
288
289.LBL_2_3:
290 cfi_restore_state
291 vmovups %zmm0, 1152(%rsp)
292 vmovups %zmm1, 1216(%rsp)
293 je .LBL_2_2
294
295 xorb %dl, %dl
296 xorl %eax, %eax
297 kmovw %k4, 1048(%rsp)
298 kmovw %k5, 1040(%rsp)
299 kmovw %k6, 1032(%rsp)
300 kmovw %k7, 1024(%rsp)
301 vmovups %zmm16, 960(%rsp)
302 vmovups %zmm17, 896(%rsp)
303 vmovups %zmm18, 832(%rsp)
304 vmovups %zmm19, 768(%rsp)
305 vmovups %zmm20, 704(%rsp)
306 vmovups %zmm21, 640(%rsp)
307 vmovups %zmm22, 576(%rsp)
308 vmovups %zmm23, 512(%rsp)
309 vmovups %zmm24, 448(%rsp)
310 vmovups %zmm25, 384(%rsp)
311 vmovups %zmm26, 320(%rsp)
312 vmovups %zmm27, 256(%rsp)
313 vmovups %zmm28, 192(%rsp)
314 vmovups %zmm29, 128(%rsp)
315 vmovups %zmm30, 64(%rsp)
316 vmovups %zmm31, (%rsp)
317 movq %rsi, 1064(%rsp)
318 movq %rdi, 1056(%rsp)
319 movq %r12, 1096(%rsp)
320 cfi_offset_rel_rsp (12, 1096)
321 movb %dl, %r12b
322 movq %r13, 1088(%rsp)
323 cfi_offset_rel_rsp (13, 1088)
324 movl %ecx, %r13d
325 movq %r14, 1080(%rsp)
326 cfi_offset_rel_rsp (14, 1080)
327 movl %eax, %r14d
328 movq %r15, 1072(%rsp)
329 cfi_offset_rel_rsp (15, 1072)
330 cfi_remember_state
331
332.LBL_2_6:
333 btl %r14d, %r13d
334 jc .LBL_2_12
335
336.LBL_2_7:
337 lea 1(%r14), %esi
338 btl %esi, %r13d
339 jc .LBL_2_10
340
341.LBL_2_8:
342 incb %r12b
343 addl $2, %r14d
344 cmpb $16, %r12b
345 jb .LBL_2_6
346
347 kmovw 1048(%rsp), %k4
348 kmovw 1040(%rsp), %k5
349 kmovw 1032(%rsp), %k6
350 kmovw 1024(%rsp), %k7
351 vmovups 960(%rsp), %zmm16
352 vmovups 896(%rsp), %zmm17
353 vmovups 832(%rsp), %zmm18
354 vmovups 768(%rsp), %zmm19
355 vmovups 704(%rsp), %zmm20
356 vmovups 640(%rsp), %zmm21
357 vmovups 576(%rsp), %zmm22
358 vmovups 512(%rsp), %zmm23
359 vmovups 448(%rsp), %zmm24
360 vmovups 384(%rsp), %zmm25
361 vmovups 320(%rsp), %zmm26
362 vmovups 256(%rsp), %zmm27
363 vmovups 192(%rsp), %zmm28
364 vmovups 128(%rsp), %zmm29
365 vmovups 64(%rsp), %zmm30
366 vmovups (%rsp), %zmm31
367 vmovups 1216(%rsp), %zmm1
368 movq 1064(%rsp), %rsi
369 movq 1056(%rsp), %rdi
370 movq 1096(%rsp), %r12
371 cfi_restore (%r12)
372 movq 1088(%rsp), %r13
373 cfi_restore (%r13)
374 movq 1080(%rsp), %r14
375 cfi_restore (%r14)
376 movq 1072(%rsp), %r15
377 cfi_restore (%r15)
378 jmp .LBL_2_2
379
380.LBL_2_10:
381 cfi_restore_state
382 movzbl %r12b, %r15d
383 vmovss 1156(%rsp,%r15,8), %xmm0
384 vzeroupper
385 vmovss 1156(%rsp,%r15,8), %xmm0
386
387 call JUMPTARGET(logf)
388
389 vmovss %xmm0, 1220(%rsp,%r15,8)
390 jmp .LBL_2_8
391
392.LBL_2_12:
393 movzbl %r12b, %r15d
394 vmovss 1152(%rsp,%r15,8), %xmm0
395 vzeroupper
396 vmovss 1152(%rsp,%r15,8), %xmm0
397
398 call JUMPTARGET(logf)
399
400 vmovss %xmm0, 1216(%rsp,%r15,8)
401 jmp .LBL_2_7
402
403END (_ZGVeN16v_logf_skx)
404

source code of glibc/sysdeps/x86_64/fpu/multiarch/svml_s_logf16_core_avx512.S