1/* Function log vectorized with AVX-512. KNL and SKX versions.
2 Copyright (C) 2014-2020 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_d_log_data.h"
21#include "svml_d_wrapper_impl.h"
22
23 .text
24ENTRY (_ZGVeN8v_log_knl)
25#ifndef HAVE_AVX512DQ_ASM_SUPPORT
26WRAPPER_IMPL_AVX512 _ZGVdN4v_log
27#else
28/*
29 ALGORITHM DESCRIPTION:
30
31 log(x) = -log(Rcp) + log(Rcp*x),
32 where Rcp ~ 1/x (accuracy ~9 bits, obtained by
33 rounding HW approximation to 1+9 mantissa bits)
34
35 Reduced argument R=Rcp*x-1 is used to approximate log(1+R) as polynomial
36
37 log(Rcp) = exponent_Rcp*log(2) + log(mantissa_Rcp)
38 -log(mantissa_Rcp) is obtained from a lookup table,
39 accessed by a 9-bit index
40 */
41 pushq %rbp
42 cfi_adjust_cfa_offset (8)
43 cfi_rel_offset (%rbp, 0)
44 movq %rsp, %rbp
45 cfi_def_cfa_register (%rbp)
46 andq $-64, %rsp
47 subq $1280, %rsp
48 movq __svml_dlog_data@GOTPCREL(%rip), %rdx
49 movq $-1, %rax
50
51/* isolate exponent bits */
52 vpsrlq $20, %zmm0, %zmm2
53 vpsrlq $32, %zmm2, %zmm3
54 vpxord %zmm2, %zmm2, %zmm2
55 kxnorw %k3, %k3, %k3
56 vmovups _Two10(%rdx), %zmm1
57 vmovups _One(%rdx), %zmm9
58 vpmovqd %zmm3, %ymm4
59
60/* convert biased exponent to DP format */
61 vcvtdq2pd %ymm4, %zmm13
62
63/* preserve mantissa, set input exponent to 2^(-10) */
64 vpternlogq $248, _ExpMask(%rdx), %zmm0, %zmm1
65 vcmppd $17, _MinNorm(%rdx), %zmm0, %k1
66
67/* reciprocal approximation good to at least 11 bits */
68 vrcp28pd %zmm1, %zmm5
69 vpbroadcastq %rax, %zmm6{%k1}{z}
70 vmovups _poly_coeff_3(%rdx), %zmm15
71 vcmppd $22, _MaxNorm(%rdx), %zmm0, %k2
72 vmovups _Bias1(%rdx), %zmm14
73
74/* round reciprocal to nearest integer, will have 1+9 mantissa bits */
75 vrndscalepd $8, %zmm5, %zmm11
76 vpbroadcastq %rax, %zmm7{%k2}{z}
77
78/* argument reduction started: R = Mantissa*Rcp - 1 */
79 vfmsub213pd %zmm9, %zmm11, %zmm1
80
81/* calculate index for table lookup */
82 vpsrlq $40, %zmm11, %zmm10
83 vgatherqpd _LogRcp_lookup(%rdx,%zmm10), %zmm2{%k3}
84 vcmppd $30, _Threshold(%rdx), %zmm11, %k1
85
86/* combine and get argument value range mask */
87 vporq %zmm7, %zmm6, %zmm8
88
89/* exponent*log(2.0) */
90 vmovups _poly_coeff_1(%rdx), %zmm11
91 vmulpd %zmm1, %zmm1, %zmm10
92 vptestmq %zmm8, %zmm8, %k0
93 vfmadd213pd _poly_coeff_4(%rdx), %zmm1, %zmm15
94 kmovw %k0, %ecx
95
96/* polynomial computation */
97 vfmadd213pd _poly_coeff_2(%rdx), %zmm1, %zmm11
98 movzbl %cl, %ecx
99 vpbroadcastq %rax, %zmm12{%k1}{z}
100 vfmadd213pd %zmm15, %zmm10, %zmm11
101 vpternlogq $248, _Bias(%rdx), %zmm12, %zmm14
102
103/*
104 Table stores -log(0.5*mantissa) for larger mantissas,
105 adjust exponent accordingly
106 */
107 vsubpd %zmm14, %zmm13, %zmm3
108
109/*
110 reconstruction:
111 (exponent*log(2)) + (LogRcp + (R+poly))
112 */
113 vfmadd213pd %zmm1, %zmm10, %zmm11
114 vaddpd %zmm2, %zmm11, %zmm1
115 vfmadd132pd _L2(%rdx), %zmm1, %zmm3
116 testl %ecx, %ecx
117 jne .LBL_1_3
118
119.LBL_1_2:
120 cfi_remember_state
121 vmovaps %zmm3, %zmm0
122 movq %rbp, %rsp
123 cfi_def_cfa_register (%rsp)
124 popq %rbp
125 cfi_adjust_cfa_offset (-8)
126 cfi_restore (%rbp)
127 ret
128
129.LBL_1_3:
130 cfi_restore_state
131 vmovups %zmm0, 1152(%rsp)
132 vmovups %zmm3, 1216(%rsp)
133 je .LBL_1_2
134
135 xorb %dl, %dl
136 kmovw %k4, 1048(%rsp)
137 xorl %eax, %eax
138 kmovw %k5, 1040(%rsp)
139 kmovw %k6, 1032(%rsp)
140 kmovw %k7, 1024(%rsp)
141 vmovups %zmm16, 960(%rsp)
142 vmovups %zmm17, 896(%rsp)
143 vmovups %zmm18, 832(%rsp)
144 vmovups %zmm19, 768(%rsp)
145 vmovups %zmm20, 704(%rsp)
146 vmovups %zmm21, 640(%rsp)
147 vmovups %zmm22, 576(%rsp)
148 vmovups %zmm23, 512(%rsp)
149 vmovups %zmm24, 448(%rsp)
150 vmovups %zmm25, 384(%rsp)
151 vmovups %zmm26, 320(%rsp)
152 vmovups %zmm27, 256(%rsp)
153 vmovups %zmm28, 192(%rsp)
154 vmovups %zmm29, 128(%rsp)
155 vmovups %zmm30, 64(%rsp)
156 vmovups %zmm31, (%rsp)
157 movq %rsi, 1064(%rsp)
158 movq %rdi, 1056(%rsp)
159 movq %r12, 1096(%rsp)
160 cfi_offset_rel_rsp (12, 1096)
161 movb %dl, %r12b
162 movq %r13, 1088(%rsp)
163 cfi_offset_rel_rsp (13, 1088)
164 movl %ecx, %r13d
165 movq %r14, 1080(%rsp)
166 cfi_offset_rel_rsp (14, 1080)
167 movl %eax, %r14d
168 movq %r15, 1072(%rsp)
169 cfi_offset_rel_rsp (15, 1072)
170 cfi_remember_state
171
172.LBL_1_6:
173 btl %r14d, %r13d
174 jc .LBL_1_12
175
176.LBL_1_7:
177 lea 1(%r14), %esi
178 btl %esi, %r13d
179 jc .LBL_1_10
180
181.LBL_1_8:
182 addb $1, %r12b
183 addl $2, %r14d
184 cmpb $16, %r12b
185 jb .LBL_1_6
186
187 kmovw 1048(%rsp), %k4
188 movq 1064(%rsp), %rsi
189 kmovw 1040(%rsp), %k5
190 movq 1056(%rsp), %rdi
191 kmovw 1032(%rsp), %k6
192 movq 1096(%rsp), %r12
193 cfi_restore (%r12)
194 movq 1088(%rsp), %r13
195 cfi_restore (%r13)
196 kmovw 1024(%rsp), %k7
197 vmovups 960(%rsp), %zmm16
198 vmovups 896(%rsp), %zmm17
199 vmovups 832(%rsp), %zmm18
200 vmovups 768(%rsp), %zmm19
201 vmovups 704(%rsp), %zmm20
202 vmovups 640(%rsp), %zmm21
203 vmovups 576(%rsp), %zmm22
204 vmovups 512(%rsp), %zmm23
205 vmovups 448(%rsp), %zmm24
206 vmovups 384(%rsp), %zmm25
207 vmovups 320(%rsp), %zmm26
208 vmovups 256(%rsp), %zmm27
209 vmovups 192(%rsp), %zmm28
210 vmovups 128(%rsp), %zmm29
211 vmovups 64(%rsp), %zmm30
212 vmovups (%rsp), %zmm31
213 movq 1080(%rsp), %r14
214 cfi_restore (%r14)
215 movq 1072(%rsp), %r15
216 cfi_restore (%r15)
217 vmovups 1216(%rsp), %zmm3
218 jmp .LBL_1_2
219
220.LBL_1_10:
221 cfi_restore_state
222 movzbl %r12b, %r15d
223 shlq $4, %r15
224 vmovsd 1160(%rsp,%r15), %xmm0
225 call JUMPTARGET(log)
226 vmovsd %xmm0, 1224(%rsp,%r15)
227 jmp .LBL_1_8
228
229.LBL_1_12:
230 movzbl %r12b, %r15d
231 shlq $4, %r15
232 vmovsd 1152(%rsp,%r15), %xmm0
233 call JUMPTARGET(log)
234 vmovsd %xmm0, 1216(%rsp,%r15)
235 jmp .LBL_1_7
236#endif
237END (_ZGVeN8v_log_knl)
238
239ENTRY (_ZGVeN8v_log_skx)
240#ifndef HAVE_AVX512DQ_ASM_SUPPORT
241WRAPPER_IMPL_AVX512 _ZGVdN4v_log
242#else
243/*
244 ALGORITHM DESCRIPTION:
245
246 log(x) = -log(Rcp) + log(Rcp*x),
247 where Rcp ~ 1/x (accuracy ~9 bits,
248 obtained by rounding HW approximation to 1+9 mantissa bits)
249
250 Reduced argument R=Rcp*x-1 is used to approximate log(1+R) as polynomial
251
252 log(Rcp) = exponent_Rcp*log(2) + log(mantissa_Rcp)
253 -log(mantissa_Rcp) is obtained from a lookup table,
254 accessed by a 9-bit index
255 */
256 pushq %rbp
257 cfi_adjust_cfa_offset (8)
258 cfi_rel_offset (%rbp, 0)
259 movq %rsp, %rbp
260 cfi_def_cfa_register (%rbp)
261 andq $-64, %rsp
262 subq $1280, %rsp
263 movq __svml_dlog_data@GOTPCREL(%rip), %rax
264 vmovaps %zmm0, %zmm3
265 kxnorw %k3, %k3, %k3
266 vmovups _Two10(%rax), %zmm2
267 vmovups _Threshold(%rax), %zmm14
268 vmovups _One(%rax), %zmm11
269 vcmppd $21, _MinNorm(%rax), %zmm3, %k1
270 vcmppd $18, _MaxNorm(%rax), %zmm3, %k2
271
272/* isolate exponent bits */
273 vpsrlq $20, %zmm3, %zmm4
274
275/* preserve mantissa, set input exponent to 2^(-10) */
276 vpternlogq $248, _ExpMask(%rax), %zmm3, %zmm2
277 vpbroadcastq .L_2il0floatpacket.12(%rip), %zmm1
278 vpsrlq $32, %zmm4, %zmm6
279
280/* reciprocal approximation good to at least 11 bits */
281 vrcp14pd %zmm2, %zmm5
282
283/* exponent*log(2.0) */
284 vmovups _poly_coeff_1(%rax), %zmm4
285 vpmovqd %zmm6, %ymm7
286
287/* round reciprocal to nearest integer, will have 1+9 mantissa bits */
288 vrndscalepd $8, %zmm5, %zmm0
289
290/* calculate index for table lookup */
291 vpsrlq $40, %zmm0, %zmm12
292
293/* argument reduction started: R = Mantissa*Rcp - 1 */
294 vfmsub213pd %zmm11, %zmm0, %zmm2
295 vpmovqd %zmm12, %ymm13
296
297/* polynomial computation */
298 vfmadd213pd _poly_coeff_2(%rax), %zmm2, %zmm4
299 vmovaps %zmm1, %zmm8
300 vmovaps %zmm1, %zmm9
301 vpxord %zmm5, %zmm5, %zmm5
302 vgatherdpd _LogRcp_lookup(%rax,%ymm13), %zmm5{%k3}
303 vmovups _Bias1(%rax), %zmm13
304 vpandnq %zmm3, %zmm3, %zmm8{%k1}
305 vcmppd $21, %zmm0, %zmm14, %k1
306 vpandnq %zmm14, %zmm14, %zmm1{%k1}
307 vmulpd %zmm2, %zmm2, %zmm14
308 vpternlogq $248, _Bias(%rax), %zmm1, %zmm13
309 vmovups _poly_coeff_3(%rax), %zmm1
310 vfmadd213pd _poly_coeff_4(%rax), %zmm2, %zmm1
311 vfmadd213pd %zmm1, %zmm14, %zmm4
312
313/*
314 reconstruction:
315 (exponent*log(2)) + (LogRcp + (R+poly))
316 */
317 vfmadd213pd %zmm2, %zmm14, %zmm4
318 vaddpd %zmm5, %zmm4, %zmm2
319 vpandnq %zmm3, %zmm3, %zmm9{%k2}
320
321/* combine and get argument value range mask */
322 vorpd %zmm9, %zmm8, %zmm10
323 vcmppd $3, %zmm10, %zmm10, %k0
324 kmovw %k0, %ecx
325
326/* convert biased exponent to DP format */
327 vcvtdq2pd %ymm7, %zmm15
328
329/*
330 Table stores -log(0.5*mantissa) for larger mantissas,
331 adjust exponent accordingly
332 */
333 vsubpd %zmm13, %zmm15, %zmm0
334 vfmadd132pd _L2(%rax), %zmm2, %zmm0
335 testl %ecx, %ecx
336 jne .LBL_2_3
337
338.LBL_2_2:
339 cfi_remember_state
340 movq %rbp, %rsp
341 cfi_def_cfa_register (%rsp)
342 popq %rbp
343 cfi_adjust_cfa_offset (-8)
344 cfi_restore (%rbp)
345 ret
346
347.LBL_2_3:
348 cfi_restore_state
349 vmovups %zmm3, 1152(%rsp)
350 vmovups %zmm0, 1216(%rsp)
351 je .LBL_2_2
352
353 xorb %dl, %dl
354 xorl %eax, %eax
355 kmovw %k4, 1048(%rsp)
356 kmovw %k5, 1040(%rsp)
357 kmovw %k6, 1032(%rsp)
358 kmovw %k7, 1024(%rsp)
359 vmovups %zmm16, 960(%rsp)
360 vmovups %zmm17, 896(%rsp)
361 vmovups %zmm18, 832(%rsp)
362 vmovups %zmm19, 768(%rsp)
363 vmovups %zmm20, 704(%rsp)
364 vmovups %zmm21, 640(%rsp)
365 vmovups %zmm22, 576(%rsp)
366 vmovups %zmm23, 512(%rsp)
367 vmovups %zmm24, 448(%rsp)
368 vmovups %zmm25, 384(%rsp)
369 vmovups %zmm26, 320(%rsp)
370 vmovups %zmm27, 256(%rsp)
371 vmovups %zmm28, 192(%rsp)
372 vmovups %zmm29, 128(%rsp)
373 vmovups %zmm30, 64(%rsp)
374 vmovups %zmm31, (%rsp)
375 movq %rsi, 1064(%rsp)
376 movq %rdi, 1056(%rsp)
377 movq %r12, 1096(%rsp)
378 cfi_offset_rel_rsp (12, 1096)
379 movb %dl, %r12b
380 movq %r13, 1088(%rsp)
381 cfi_offset_rel_rsp (13, 1088)
382 movl %ecx, %r13d
383 movq %r14, 1080(%rsp)
384 cfi_offset_rel_rsp (14, 1080)
385 movl %eax, %r14d
386 movq %r15, 1072(%rsp)
387 cfi_offset_rel_rsp (15, 1072)
388 cfi_remember_state
389
390.LBL_2_6:
391 btl %r14d, %r13d
392 jc .LBL_2_12
393
394.LBL_2_7:
395 lea 1(%r14), %esi
396 btl %esi, %r13d
397 jc .LBL_2_10
398
399.LBL_2_8:
400 incb %r12b
401 addl $2, %r14d
402 cmpb $16, %r12b
403 jb .LBL_2_6
404
405 kmovw 1048(%rsp), %k4
406 kmovw 1040(%rsp), %k5
407 kmovw 1032(%rsp), %k6
408 kmovw 1024(%rsp), %k7
409 vmovups 960(%rsp), %zmm16
410 vmovups 896(%rsp), %zmm17
411 vmovups 832(%rsp), %zmm18
412 vmovups 768(%rsp), %zmm19
413 vmovups 704(%rsp), %zmm20
414 vmovups 640(%rsp), %zmm21
415 vmovups 576(%rsp), %zmm22
416 vmovups 512(%rsp), %zmm23
417 vmovups 448(%rsp), %zmm24
418 vmovups 384(%rsp), %zmm25
419 vmovups 320(%rsp), %zmm26
420 vmovups 256(%rsp), %zmm27
421 vmovups 192(%rsp), %zmm28
422 vmovups 128(%rsp), %zmm29
423 vmovups 64(%rsp), %zmm30
424 vmovups (%rsp), %zmm31
425 vmovups 1216(%rsp), %zmm0
426 movq 1064(%rsp), %rsi
427 movq 1056(%rsp), %rdi
428 movq 1096(%rsp), %r12
429 cfi_restore (%r12)
430 movq 1088(%rsp), %r13
431 cfi_restore (%r13)
432 movq 1080(%rsp), %r14
433 cfi_restore (%r14)
434 movq 1072(%rsp), %r15
435 cfi_restore (%r15)
436 jmp .LBL_2_2
437
438.LBL_2_10:
439 cfi_restore_state
440 movzbl %r12b, %r15d
441 shlq $4, %r15
442 vmovsd 1160(%rsp,%r15), %xmm0
443 vzeroupper
444 vmovsd 1160(%rsp,%r15), %xmm0
445
446 call JUMPTARGET(log)
447
448 vmovsd %xmm0, 1224(%rsp,%r15)
449 jmp .LBL_2_8
450
451.LBL_2_12:
452 movzbl %r12b, %r15d
453 shlq $4, %r15
454 vmovsd 1152(%rsp,%r15), %xmm0
455 vzeroupper
456 vmovsd 1152(%rsp,%r15), %xmm0
457
458 call JUMPTARGET(log)
459
460 vmovsd %xmm0, 1216(%rsp,%r15)
461 jmp .LBL_2_7
462#endif
463END (_ZGVeN8v_log_skx)
464
465 .section .rodata, "a"
466.L_2il0floatpacket.12:
467 .long 0xffffffff,0xffffffff
468 .type .L_2il0floatpacket.12,@object
469