1 | /* Function log 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_d_log_data.h" |
21 | #include "svml_d_wrapper_impl.h" |
22 | |
23 | .section .text.evex512, "ax" , @progbits |
24 | ENTRY (_ZGVeN8v_log_knl) |
25 | /* |
26 | ALGORITHM DESCRIPTION: |
27 | |
28 | log(x) = -log(Rcp) + log(Rcp*x), |
29 | where Rcp ~ 1/x (accuracy ~9 bits, obtained by |
30 | rounding HW approximation to 1+9 mantissa bits) |
31 | |
32 | Reduced argument R=Rcp*x-1 is used to approximate log(1+R) as polynomial |
33 | |
34 | log(Rcp) = exponent_Rcp*log(2) + log(mantissa_Rcp) |
35 | -log(mantissa_Rcp) is obtained from a lookup table, |
36 | accessed by a 9-bit index |
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_dlog_data@GOTPCREL(%rip), %rdx |
46 | movq $-1, %rax |
47 | |
48 | /* isolate exponent bits */ |
49 | vpsrlq $20, %zmm0, %zmm2 |
50 | vpsrlq $32, %zmm2, %zmm3 |
51 | vpxord %zmm2, %zmm2, %zmm2 |
52 | kxnorw %k3, %k3, %k3 |
53 | vmovups _Two10(%rdx), %zmm1 |
54 | vmovups _One(%rdx), %zmm9 |
55 | vpmovqd %zmm3, %ymm4 |
56 | |
57 | /* convert biased exponent to DP format */ |
58 | vcvtdq2pd %ymm4, %zmm13 |
59 | |
60 | /* preserve mantissa, set input exponent to 2^(-10) */ |
61 | vpternlogq $248, _ExpMask(%rdx), %zmm0, %zmm1 |
62 | vcmppd $17, _MinNorm(%rdx), %zmm0, %k1 |
63 | |
64 | /* reciprocal approximation good to at least 11 bits */ |
65 | vrcp28pd %zmm1, %zmm5 |
66 | vpbroadcastq %rax, %zmm6{%k1}{z} |
67 | vmovups _poly_coeff_3(%rdx), %zmm15 |
68 | vcmppd $22, _MaxNorm(%rdx), %zmm0, %k2 |
69 | vmovups _Bias1(%rdx), %zmm14 |
70 | |
71 | /* round reciprocal to nearest integer, will have 1+9 mantissa bits */ |
72 | vrndscalepd $8, %zmm5, %zmm11 |
73 | vpbroadcastq %rax, %zmm7{%k2}{z} |
74 | |
75 | /* argument reduction started: R = Mantissa*Rcp - 1 */ |
76 | vfmsub213pd %zmm9, %zmm11, %zmm1 |
77 | |
78 | /* calculate index for table lookup */ |
79 | vpsrlq $40, %zmm11, %zmm10 |
80 | vgatherqpd _LogRcp_lookup(%rdx,%zmm10), %zmm2{%k3} |
81 | vcmppd $30, _Threshold(%rdx), %zmm11, %k1 |
82 | |
83 | /* combine and get argument value range mask */ |
84 | vporq %zmm7, %zmm6, %zmm8 |
85 | |
86 | /* exponent*log(2.0) */ |
87 | vmovups _poly_coeff_1(%rdx), %zmm11 |
88 | vmulpd %zmm1, %zmm1, %zmm10 |
89 | vptestmq %zmm8, %zmm8, %k0 |
90 | vfmadd213pd _poly_coeff_4(%rdx), %zmm1, %zmm15 |
91 | kmovw %k0, %ecx |
92 | |
93 | /* polynomial computation */ |
94 | vfmadd213pd _poly_coeff_2(%rdx), %zmm1, %zmm11 |
95 | movzbl %cl, %ecx |
96 | vpbroadcastq %rax, %zmm12{%k1}{z} |
97 | vfmadd213pd %zmm15, %zmm10, %zmm11 |
98 | vpternlogq $248, _Bias(%rdx), %zmm12, %zmm14 |
99 | |
100 | /* |
101 | Table stores -log(0.5*mantissa) for larger mantissas, |
102 | adjust exponent accordingly |
103 | */ |
104 | vsubpd %zmm14, %zmm13, %zmm3 |
105 | |
106 | /* |
107 | reconstruction: |
108 | (exponent*log(2)) + (LogRcp + (R+poly)) |
109 | */ |
110 | vfmadd213pd %zmm1, %zmm10, %zmm11 |
111 | vaddpd %zmm2, %zmm11, %zmm1 |
112 | vfmadd132pd _L2(%rdx), %zmm1, %zmm3 |
113 | testl %ecx, %ecx |
114 | jne .LBL_1_3 |
115 | |
116 | .LBL_1_2: |
117 | cfi_remember_state |
118 | vmovaps %zmm3, %zmm0 |
119 | movq %rbp, %rsp |
120 | cfi_def_cfa_register (%rsp) |
121 | popq %rbp |
122 | cfi_adjust_cfa_offset (-8) |
123 | cfi_restore (%rbp) |
124 | ret |
125 | |
126 | .LBL_1_3: |
127 | cfi_restore_state |
128 | vmovups %zmm0, 1152(%rsp) |
129 | vmovups %zmm3, 1216(%rsp) |
130 | je .LBL_1_2 |
131 | |
132 | xorb %dl, %dl |
133 | kmovw %k4, 1048(%rsp) |
134 | xorl %eax, %eax |
135 | kmovw %k5, 1040(%rsp) |
136 | kmovw %k6, 1032(%rsp) |
137 | kmovw %k7, 1024(%rsp) |
138 | vmovups %zmm16, 960(%rsp) |
139 | vmovups %zmm17, 896(%rsp) |
140 | vmovups %zmm18, 832(%rsp) |
141 | vmovups %zmm19, 768(%rsp) |
142 | vmovups %zmm20, 704(%rsp) |
143 | vmovups %zmm21, 640(%rsp) |
144 | vmovups %zmm22, 576(%rsp) |
145 | vmovups %zmm23, 512(%rsp) |
146 | vmovups %zmm24, 448(%rsp) |
147 | vmovups %zmm25, 384(%rsp) |
148 | vmovups %zmm26, 320(%rsp) |
149 | vmovups %zmm27, 256(%rsp) |
150 | vmovups %zmm28, 192(%rsp) |
151 | vmovups %zmm29, 128(%rsp) |
152 | vmovups %zmm30, 64(%rsp) |
153 | vmovups %zmm31, (%rsp) |
154 | movq %rsi, 1064(%rsp) |
155 | movq %rdi, 1056(%rsp) |
156 | movq %r12, 1096(%rsp) |
157 | cfi_offset_rel_rsp (12, 1096) |
158 | movb %dl, %r12b |
159 | movq %r13, 1088(%rsp) |
160 | cfi_offset_rel_rsp (13, 1088) |
161 | movl %ecx, %r13d |
162 | movq %r14, 1080(%rsp) |
163 | cfi_offset_rel_rsp (14, 1080) |
164 | movl %eax, %r14d |
165 | movq %r15, 1072(%rsp) |
166 | cfi_offset_rel_rsp (15, 1072) |
167 | cfi_remember_state |
168 | |
169 | .LBL_1_6: |
170 | btl %r14d, %r13d |
171 | jc .LBL_1_12 |
172 | |
173 | .LBL_1_7: |
174 | lea 1(%r14), %esi |
175 | btl %esi, %r13d |
176 | jc .LBL_1_10 |
177 | |
178 | .LBL_1_8: |
179 | addb $1, %r12b |
180 | addl $2, %r14d |
181 | cmpb $16, %r12b |
182 | jb .LBL_1_6 |
183 | |
184 | kmovw 1048(%rsp), %k4 |
185 | movq 1064(%rsp), %rsi |
186 | kmovw 1040(%rsp), %k5 |
187 | movq 1056(%rsp), %rdi |
188 | kmovw 1032(%rsp), %k6 |
189 | movq 1096(%rsp), %r12 |
190 | cfi_restore (%r12) |
191 | movq 1088(%rsp), %r13 |
192 | cfi_restore (%r13) |
193 | kmovw 1024(%rsp), %k7 |
194 | vmovups 960(%rsp), %zmm16 |
195 | vmovups 896(%rsp), %zmm17 |
196 | vmovups 832(%rsp), %zmm18 |
197 | vmovups 768(%rsp), %zmm19 |
198 | vmovups 704(%rsp), %zmm20 |
199 | vmovups 640(%rsp), %zmm21 |
200 | vmovups 576(%rsp), %zmm22 |
201 | vmovups 512(%rsp), %zmm23 |
202 | vmovups 448(%rsp), %zmm24 |
203 | vmovups 384(%rsp), %zmm25 |
204 | vmovups 320(%rsp), %zmm26 |
205 | vmovups 256(%rsp), %zmm27 |
206 | vmovups 192(%rsp), %zmm28 |
207 | vmovups 128(%rsp), %zmm29 |
208 | vmovups 64(%rsp), %zmm30 |
209 | vmovups (%rsp), %zmm31 |
210 | movq 1080(%rsp), %r14 |
211 | cfi_restore (%r14) |
212 | movq 1072(%rsp), %r15 |
213 | cfi_restore (%r15) |
214 | vmovups 1216(%rsp), %zmm3 |
215 | jmp .LBL_1_2 |
216 | |
217 | .LBL_1_10: |
218 | cfi_restore_state |
219 | movzbl %r12b, %r15d |
220 | shlq $4, %r15 |
221 | vmovsd 1160(%rsp,%r15), %xmm0 |
222 | call JUMPTARGET(log) |
223 | vmovsd %xmm0, 1224(%rsp,%r15) |
224 | jmp .LBL_1_8 |
225 | |
226 | .LBL_1_12: |
227 | movzbl %r12b, %r15d |
228 | shlq $4, %r15 |
229 | vmovsd 1152(%rsp,%r15), %xmm0 |
230 | call JUMPTARGET(log) |
231 | vmovsd %xmm0, 1216(%rsp,%r15) |
232 | jmp .LBL_1_7 |
233 | END (_ZGVeN8v_log_knl) |
234 | |
235 | ENTRY (_ZGVeN8v_log_skx) |
236 | /* |
237 | ALGORITHM DESCRIPTION: |
238 | |
239 | log(x) = -log(Rcp) + log(Rcp*x), |
240 | where Rcp ~ 1/x (accuracy ~9 bits, |
241 | obtained by rounding HW approximation to 1+9 mantissa bits) |
242 | |
243 | Reduced argument R=Rcp*x-1 is used to approximate log(1+R) as polynomial |
244 | |
245 | log(Rcp) = exponent_Rcp*log(2) + log(mantissa_Rcp) |
246 | -log(mantissa_Rcp) is obtained from a lookup table, |
247 | accessed by a 9-bit index |
248 | */ |
249 | pushq %rbp |
250 | cfi_adjust_cfa_offset (8) |
251 | cfi_rel_offset (%rbp, 0) |
252 | movq %rsp, %rbp |
253 | cfi_def_cfa_register (%rbp) |
254 | andq $-64, %rsp |
255 | subq $1280, %rsp |
256 | movq __svml_dlog_data@GOTPCREL(%rip), %rax |
257 | vmovaps %zmm0, %zmm3 |
258 | kxnorw %k3, %k3, %k3 |
259 | vmovups _Two10(%rax), %zmm2 |
260 | vmovups _Threshold(%rax), %zmm14 |
261 | vmovups _One(%rax), %zmm11 |
262 | vcmppd $21, _MinNorm(%rax), %zmm3, %k1 |
263 | vcmppd $18, _MaxNorm(%rax), %zmm3, %k2 |
264 | |
265 | /* isolate exponent bits */ |
266 | vpsrlq $20, %zmm3, %zmm4 |
267 | |
268 | /* preserve mantissa, set input exponent to 2^(-10) */ |
269 | vpternlogq $248, _ExpMask(%rax), %zmm3, %zmm2 |
270 | vpternlogd $0xff, %zmm1, %zmm1, %zmm1 |
271 | vpsrlq $32, %zmm4, %zmm6 |
272 | |
273 | /* reciprocal approximation good to at least 11 bits */ |
274 | vrcp14pd %zmm2, %zmm5 |
275 | |
276 | /* exponent*log(2.0) */ |
277 | vmovups _poly_coeff_1(%rax), %zmm4 |
278 | vpmovqd %zmm6, %ymm7 |
279 | |
280 | /* round reciprocal to nearest integer, will have 1+9 mantissa bits */ |
281 | vrndscalepd $8, %zmm5, %zmm0 |
282 | |
283 | /* calculate index for table lookup */ |
284 | vpsrlq $40, %zmm0, %zmm12 |
285 | |
286 | /* argument reduction started: R = Mantissa*Rcp - 1 */ |
287 | vfmsub213pd %zmm11, %zmm0, %zmm2 |
288 | vpmovqd %zmm12, %ymm13 |
289 | |
290 | /* polynomial computation */ |
291 | vfmadd213pd _poly_coeff_2(%rax), %zmm2, %zmm4 |
292 | vmovaps %zmm1, %zmm8 |
293 | vmovaps %zmm1, %zmm9 |
294 | vpxord %zmm5, %zmm5, %zmm5 |
295 | vgatherdpd _LogRcp_lookup(%rax,%ymm13), %zmm5{%k3} |
296 | vmovups _Bias1(%rax), %zmm13 |
297 | vpandnq %zmm3, %zmm3, %zmm8{%k1} |
298 | vcmppd $21, %zmm0, %zmm14, %k1 |
299 | vpandnq %zmm14, %zmm14, %zmm1{%k1} |
300 | vmulpd %zmm2, %zmm2, %zmm14 |
301 | vpternlogq $248, _Bias(%rax), %zmm1, %zmm13 |
302 | vmovups _poly_coeff_3(%rax), %zmm1 |
303 | vfmadd213pd _poly_coeff_4(%rax), %zmm2, %zmm1 |
304 | vfmadd213pd %zmm1, %zmm14, %zmm4 |
305 | |
306 | /* |
307 | reconstruction: |
308 | (exponent*log(2)) + (LogRcp + (R+poly)) |
309 | */ |
310 | vfmadd213pd %zmm2, %zmm14, %zmm4 |
311 | vaddpd %zmm5, %zmm4, %zmm2 |
312 | vpandnq %zmm3, %zmm3, %zmm9{%k2} |
313 | |
314 | /* combine and get argument value range mask */ |
315 | vorpd %zmm9, %zmm8, %zmm10 |
316 | vcmppd $3, %zmm10, %zmm10, %k0 |
317 | kmovw %k0, %ecx |
318 | |
319 | /* convert biased exponent to DP format */ |
320 | vcvtdq2pd %ymm7, %zmm15 |
321 | |
322 | /* |
323 | Table stores -log(0.5*mantissa) for larger mantissas, |
324 | adjust exponent accordingly |
325 | */ |
326 | vsubpd %zmm13, %zmm15, %zmm0 |
327 | vfmadd132pd _L2(%rax), %zmm2, %zmm0 |
328 | testl %ecx, %ecx |
329 | jne .LBL_2_3 |
330 | |
331 | .LBL_2_2: |
332 | cfi_remember_state |
333 | movq %rbp, %rsp |
334 | cfi_def_cfa_register (%rsp) |
335 | popq %rbp |
336 | cfi_adjust_cfa_offset (-8) |
337 | cfi_restore (%rbp) |
338 | ret |
339 | |
340 | .LBL_2_3: |
341 | cfi_restore_state |
342 | vmovups %zmm3, 1152(%rsp) |
343 | vmovups %zmm0, 1216(%rsp) |
344 | je .LBL_2_2 |
345 | |
346 | xorb %dl, %dl |
347 | xorl %eax, %eax |
348 | kmovw %k4, 1048(%rsp) |
349 | kmovw %k5, 1040(%rsp) |
350 | kmovw %k6, 1032(%rsp) |
351 | kmovw %k7, 1024(%rsp) |
352 | vmovups %zmm16, 960(%rsp) |
353 | vmovups %zmm17, 896(%rsp) |
354 | vmovups %zmm18, 832(%rsp) |
355 | vmovups %zmm19, 768(%rsp) |
356 | vmovups %zmm20, 704(%rsp) |
357 | vmovups %zmm21, 640(%rsp) |
358 | vmovups %zmm22, 576(%rsp) |
359 | vmovups %zmm23, 512(%rsp) |
360 | vmovups %zmm24, 448(%rsp) |
361 | vmovups %zmm25, 384(%rsp) |
362 | vmovups %zmm26, 320(%rsp) |
363 | vmovups %zmm27, 256(%rsp) |
364 | vmovups %zmm28, 192(%rsp) |
365 | vmovups %zmm29, 128(%rsp) |
366 | vmovups %zmm30, 64(%rsp) |
367 | vmovups %zmm31, (%rsp) |
368 | movq %rsi, 1064(%rsp) |
369 | movq %rdi, 1056(%rsp) |
370 | movq %r12, 1096(%rsp) |
371 | cfi_offset_rel_rsp (12, 1096) |
372 | movb %dl, %r12b |
373 | movq %r13, 1088(%rsp) |
374 | cfi_offset_rel_rsp (13, 1088) |
375 | movl %ecx, %r13d |
376 | movq %r14, 1080(%rsp) |
377 | cfi_offset_rel_rsp (14, 1080) |
378 | movl %eax, %r14d |
379 | movq %r15, 1072(%rsp) |
380 | cfi_offset_rel_rsp (15, 1072) |
381 | cfi_remember_state |
382 | |
383 | .LBL_2_6: |
384 | btl %r14d, %r13d |
385 | jc .LBL_2_12 |
386 | |
387 | .LBL_2_7: |
388 | lea 1(%r14), %esi |
389 | btl %esi, %r13d |
390 | jc .LBL_2_10 |
391 | |
392 | .LBL_2_8: |
393 | incb %r12b |
394 | addl $2, %r14d |
395 | cmpb $16, %r12b |
396 | jb .LBL_2_6 |
397 | |
398 | kmovw 1048(%rsp), %k4 |
399 | kmovw 1040(%rsp), %k5 |
400 | kmovw 1032(%rsp), %k6 |
401 | kmovw 1024(%rsp), %k7 |
402 | vmovups 960(%rsp), %zmm16 |
403 | vmovups 896(%rsp), %zmm17 |
404 | vmovups 832(%rsp), %zmm18 |
405 | vmovups 768(%rsp), %zmm19 |
406 | vmovups 704(%rsp), %zmm20 |
407 | vmovups 640(%rsp), %zmm21 |
408 | vmovups 576(%rsp), %zmm22 |
409 | vmovups 512(%rsp), %zmm23 |
410 | vmovups 448(%rsp), %zmm24 |
411 | vmovups 384(%rsp), %zmm25 |
412 | vmovups 320(%rsp), %zmm26 |
413 | vmovups 256(%rsp), %zmm27 |
414 | vmovups 192(%rsp), %zmm28 |
415 | vmovups 128(%rsp), %zmm29 |
416 | vmovups 64(%rsp), %zmm30 |
417 | vmovups (%rsp), %zmm31 |
418 | vmovups 1216(%rsp), %zmm0 |
419 | movq 1064(%rsp), %rsi |
420 | movq 1056(%rsp), %rdi |
421 | movq 1096(%rsp), %r12 |
422 | cfi_restore (%r12) |
423 | movq 1088(%rsp), %r13 |
424 | cfi_restore (%r13) |
425 | movq 1080(%rsp), %r14 |
426 | cfi_restore (%r14) |
427 | movq 1072(%rsp), %r15 |
428 | cfi_restore (%r15) |
429 | jmp .LBL_2_2 |
430 | |
431 | .LBL_2_10: |
432 | cfi_restore_state |
433 | movzbl %r12b, %r15d |
434 | shlq $4, %r15 |
435 | vmovsd 1160(%rsp,%r15), %xmm0 |
436 | vzeroupper |
437 | vmovsd 1160(%rsp,%r15), %xmm0 |
438 | |
439 | call JUMPTARGET(log) |
440 | |
441 | vmovsd %xmm0, 1224(%rsp,%r15) |
442 | jmp .LBL_2_8 |
443 | |
444 | .LBL_2_12: |
445 | movzbl %r12b, %r15d |
446 | shlq $4, %r15 |
447 | vmovsd 1152(%rsp,%r15), %xmm0 |
448 | vzeroupper |
449 | vmovsd 1152(%rsp,%r15), %xmm0 |
450 | |
451 | call JUMPTARGET(log) |
452 | |
453 | vmovsd %xmm0, 1216(%rsp,%r15) |
454 | jmp .LBL_2_7 |
455 | END (_ZGVeN8v_log_skx) |
456 | |