1 | /* Function sin 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_trig_data.h" |
21 | #include "svml_d_wrapper_impl.h" |
22 | |
23 | .text |
24 | ENTRY (_ZGVeN8v_sin_knl) |
25 | #ifndef HAVE_AVX512DQ_ASM_SUPPORT |
26 | WRAPPER_IMPL_AVX512 _ZGVdN4v_sin |
27 | #else |
28 | /* |
29 | ALGORITHM DESCRIPTION: |
30 | |
31 | ( low accuracy ( < 4ulp ) or enhanced performance |
32 | ( half of correct mantissa ) implementation ) |
33 | |
34 | Argument representation: |
35 | arg = N*Pi + R |
36 | |
37 | Result calculation: |
38 | sin(arg) = sin(N*Pi + R) = (-1)^N * sin(R) |
39 | sin(R) is approximated by corresponding polynomial |
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_d_trig_data@GOTPCREL(%rip), %rax |
49 | movq $-1, %rdx |
50 | vmovups __dAbsMask(%rax), %zmm6 |
51 | vmovups __dInvPI(%rax), %zmm1 |
52 | |
53 | /* |
54 | ARGUMENT RANGE REDUCTION: |
55 | X' = |X| |
56 | */ |
57 | vpandq %zmm6, %zmm0, %zmm12 |
58 | vmovups __dPI1_FMA(%rax), %zmm2 |
59 | vmovups __dC7_sin(%rax), %zmm7 |
60 | |
61 | /* SignX - sign bit of X */ |
62 | vpandnq %zmm0, %zmm6, %zmm11 |
63 | |
64 | /* R = X' - N*Pi1 */ |
65 | vmovaps %zmm12, %zmm3 |
66 | |
67 | /* Y = X'*InvPi + RS : right shifter add */ |
68 | vfmadd213pd __dRShifter(%rax), %zmm12, %zmm1 |
69 | vcmppd $22, __dRangeVal(%rax), %zmm12, %k1 |
70 | vpbroadcastq %rdx, %zmm13{%k1}{z} |
71 | |
72 | /* N = Y - RS : right shifter sub */ |
73 | vsubpd __dRShifter(%rax), %zmm1, %zmm4 |
74 | |
75 | /* SignRes = Y<<63 : shift LSB to MSB place for result sign */ |
76 | vpsllq $63, %zmm1, %zmm5 |
77 | vptestmq %zmm13, %zmm13, %k0 |
78 | vfnmadd231pd %zmm4, %zmm2, %zmm3 |
79 | kmovw %k0, %ecx |
80 | movzbl %cl, %ecx |
81 | |
82 | /* R = R - N*Pi2 */ |
83 | vfnmadd231pd __dPI2_FMA(%rax), %zmm4, %zmm3 |
84 | |
85 | /* R = R - N*Pi3 */ |
86 | vfnmadd132pd __dPI3_FMA(%rax), %zmm3, %zmm4 |
87 | |
88 | /* |
89 | POLYNOMIAL APPROXIMATION: |
90 | R2 = R*R |
91 | */ |
92 | vmulpd %zmm4, %zmm4, %zmm8 |
93 | |
94 | /* R = R^SignRes : update sign of reduced argument */ |
95 | vpxorq %zmm5, %zmm4, %zmm9 |
96 | vfmadd213pd __dC6_sin(%rax), %zmm8, %zmm7 |
97 | vfmadd213pd __dC5_sin(%rax), %zmm8, %zmm7 |
98 | vfmadd213pd __dC4_sin(%rax), %zmm8, %zmm7 |
99 | |
100 | /* Poly = C3+R2*(C4+R2*(C5+R2*(C6+R2*C7))) */ |
101 | vfmadd213pd __dC3_sin(%rax), %zmm8, %zmm7 |
102 | |
103 | /* Poly = R2*(C1+R2*(C2+R2*Poly)) */ |
104 | vfmadd213pd __dC2_sin(%rax), %zmm8, %zmm7 |
105 | vfmadd213pd __dC1_sin(%rax), %zmm8, %zmm7 |
106 | vmulpd %zmm8, %zmm7, %zmm10 |
107 | |
108 | /* Poly = Poly*R + R */ |
109 | vfmadd213pd %zmm9, %zmm9, %zmm10 |
110 | |
111 | /* |
112 | RECONSTRUCTION: |
113 | Final sign setting: Res = Poly^SignX |
114 | */ |
115 | vpxorq %zmm11, %zmm10, %zmm1 |
116 | testl %ecx, %ecx |
117 | jne .LBL_1_3 |
118 | |
119 | .LBL_1_2: |
120 | cfi_remember_state |
121 | vmovaps %zmm1, %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 %zmm1, 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), %zmm1 |
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(sin) |
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(sin) |
234 | vmovsd %xmm0, 1216(%rsp,%r15) |
235 | jmp .LBL_1_7 |
236 | #endif |
237 | END (_ZGVeN8v_sin_knl) |
238 | |
239 | ENTRY (_ZGVeN8v_sin_skx) |
240 | #ifndef HAVE_AVX512DQ_ASM_SUPPORT |
241 | WRAPPER_IMPL_AVX512 _ZGVdN4v_sin |
242 | #else |
243 | /* |
244 | ALGORITHM DESCRIPTION: |
245 | |
246 | ( low accuracy ( < 4ulp ) or enhanced performance |
247 | ( half of correct mantissa ) implementation ) |
248 | |
249 | Argument representation: |
250 | arg = N*Pi + R |
251 | |
252 | Result calculation: |
253 | sin(arg) = sin(N*Pi + R) = (-1)^N * sin(R) |
254 | sin(R) is approximated by corresponding polynomial |
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_d_trig_data@GOTPCREL(%rip), %rax |
264 | vpbroadcastq .L_2il0floatpacket.14(%rip), %zmm14 |
265 | vmovups __dAbsMask(%rax), %zmm7 |
266 | vmovups __dInvPI(%rax), %zmm2 |
267 | vmovups __dRShifter(%rax), %zmm1 |
268 | vmovups __dPI1_FMA(%rax), %zmm3 |
269 | vmovups __dC7_sin(%rax), %zmm8 |
270 | |
271 | /* |
272 | ARGUMENT RANGE REDUCTION: |
273 | X' = |X| |
274 | */ |
275 | vandpd %zmm7, %zmm0, %zmm13 |
276 | |
277 | /* SignX - sign bit of X */ |
278 | vandnpd %zmm0, %zmm7, %zmm12 |
279 | |
280 | /* Y = X'*InvPi + RS : right shifter add */ |
281 | vfmadd213pd %zmm1, %zmm13, %zmm2 |
282 | vcmppd $18, __dRangeVal(%rax), %zmm13, %k1 |
283 | |
284 | /* SignRes = Y<<63 : shift LSB to MSB place for result sign */ |
285 | vpsllq $63, %zmm2, %zmm6 |
286 | |
287 | /* N = Y - RS : right shifter sub */ |
288 | vsubpd %zmm1, %zmm2, %zmm5 |
289 | |
290 | /* R = X' - N*Pi1 */ |
291 | vmovaps %zmm13, %zmm4 |
292 | vfnmadd231pd %zmm5, %zmm3, %zmm4 |
293 | |
294 | /* R = R - N*Pi2 */ |
295 | vfnmadd231pd __dPI2_FMA(%rax), %zmm5, %zmm4 |
296 | |
297 | /* R = R - N*Pi3 */ |
298 | vfnmadd132pd __dPI3_FMA(%rax), %zmm4, %zmm5 |
299 | |
300 | /* |
301 | POLYNOMIAL APPROXIMATION: |
302 | R2 = R*R |
303 | */ |
304 | vmulpd %zmm5, %zmm5, %zmm9 |
305 | |
306 | /* R = R^SignRes : update sign of reduced argument */ |
307 | vxorpd %zmm6, %zmm5, %zmm10 |
308 | vfmadd213pd __dC6_sin(%rax), %zmm9, %zmm8 |
309 | vfmadd213pd __dC5_sin(%rax), %zmm9, %zmm8 |
310 | vfmadd213pd __dC4_sin(%rax), %zmm9, %zmm8 |
311 | |
312 | /* Poly = C3+R2*(C4+R2*(C5+R2*(C6+R2*C7))) */ |
313 | vfmadd213pd __dC3_sin(%rax), %zmm9, %zmm8 |
314 | |
315 | /* Poly = R2*(C1+R2*(C2+R2*Poly)) */ |
316 | vfmadd213pd __dC2_sin(%rax), %zmm9, %zmm8 |
317 | vfmadd213pd __dC1_sin(%rax), %zmm9, %zmm8 |
318 | vmulpd %zmm9, %zmm8, %zmm11 |
319 | |
320 | /* Poly = Poly*R + R */ |
321 | vfmadd213pd %zmm10, %zmm10, %zmm11 |
322 | |
323 | /* |
324 | RECONSTRUCTION: |
325 | Final sign setting: Res = Poly^SignX |
326 | */ |
327 | vxorpd %zmm12, %zmm11, %zmm1 |
328 | vpandnq %zmm13, %zmm13, %zmm14{%k1} |
329 | vcmppd $3, %zmm14, %zmm14, %k0 |
330 | kmovw %k0, %ecx |
331 | testl %ecx, %ecx |
332 | jne .LBL_2_3 |
333 | |
334 | .LBL_2_2: |
335 | cfi_remember_state |
336 | vmovaps %zmm1, %zmm0 |
337 | movq %rbp, %rsp |
338 | cfi_def_cfa_register (%rsp) |
339 | popq %rbp |
340 | cfi_adjust_cfa_offset (-8) |
341 | cfi_restore (%rbp) |
342 | ret |
343 | |
344 | .LBL_2_3: |
345 | cfi_restore_state |
346 | vmovups %zmm0, 1152(%rsp) |
347 | vmovups %zmm1, 1216(%rsp) |
348 | je .LBL_2_2 |
349 | |
350 | xorb %dl, %dl |
351 | xorl %eax, %eax |
352 | kmovw %k4, 1048(%rsp) |
353 | kmovw %k5, 1040(%rsp) |
354 | kmovw %k6, 1032(%rsp) |
355 | kmovw %k7, 1024(%rsp) |
356 | vmovups %zmm16, 960(%rsp) |
357 | vmovups %zmm17, 896(%rsp) |
358 | vmovups %zmm18, 832(%rsp) |
359 | vmovups %zmm19, 768(%rsp) |
360 | vmovups %zmm20, 704(%rsp) |
361 | vmovups %zmm21, 640(%rsp) |
362 | vmovups %zmm22, 576(%rsp) |
363 | vmovups %zmm23, 512(%rsp) |
364 | vmovups %zmm24, 448(%rsp) |
365 | vmovups %zmm25, 384(%rsp) |
366 | vmovups %zmm26, 320(%rsp) |
367 | vmovups %zmm27, 256(%rsp) |
368 | vmovups %zmm28, 192(%rsp) |
369 | vmovups %zmm29, 128(%rsp) |
370 | vmovups %zmm30, 64(%rsp) |
371 | vmovups %zmm31, (%rsp) |
372 | movq %rsi, 1064(%rsp) |
373 | movq %rdi, 1056(%rsp) |
374 | movq %r12, 1096(%rsp) |
375 | cfi_offset_rel_rsp (12, 1096) |
376 | movb %dl, %r12b |
377 | movq %r13, 1088(%rsp) |
378 | cfi_offset_rel_rsp (13, 1088) |
379 | movl %ecx, %r13d |
380 | movq %r14, 1080(%rsp) |
381 | cfi_offset_rel_rsp (14, 1080) |
382 | movl %eax, %r14d |
383 | movq %r15, 1072(%rsp) |
384 | cfi_offset_rel_rsp (15, 1072) |
385 | cfi_remember_state |
386 | |
387 | .LBL_2_6: |
388 | btl %r14d, %r13d |
389 | jc .LBL_2_12 |
390 | |
391 | .LBL_2_7: |
392 | lea 1(%r14), %esi |
393 | btl %esi, %r13d |
394 | jc .LBL_2_10 |
395 | |
396 | .LBL_2_8: |
397 | incb %r12b |
398 | addl $2, %r14d |
399 | cmpb $16, %r12b |
400 | jb .LBL_2_6 |
401 | |
402 | kmovw 1048(%rsp), %k4 |
403 | kmovw 1040(%rsp), %k5 |
404 | kmovw 1032(%rsp), %k6 |
405 | kmovw 1024(%rsp), %k7 |
406 | vmovups 960(%rsp), %zmm16 |
407 | vmovups 896(%rsp), %zmm17 |
408 | vmovups 832(%rsp), %zmm18 |
409 | vmovups 768(%rsp), %zmm19 |
410 | vmovups 704(%rsp), %zmm20 |
411 | vmovups 640(%rsp), %zmm21 |
412 | vmovups 576(%rsp), %zmm22 |
413 | vmovups 512(%rsp), %zmm23 |
414 | vmovups 448(%rsp), %zmm24 |
415 | vmovups 384(%rsp), %zmm25 |
416 | vmovups 320(%rsp), %zmm26 |
417 | vmovups 256(%rsp), %zmm27 |
418 | vmovups 192(%rsp), %zmm28 |
419 | vmovups 128(%rsp), %zmm29 |
420 | vmovups 64(%rsp), %zmm30 |
421 | vmovups (%rsp), %zmm31 |
422 | vmovups 1216(%rsp), %zmm1 |
423 | movq 1064(%rsp), %rsi |
424 | movq 1056(%rsp), %rdi |
425 | movq 1096(%rsp), %r12 |
426 | cfi_restore (%r12) |
427 | movq 1088(%rsp), %r13 |
428 | cfi_restore (%r13) |
429 | movq 1080(%rsp), %r14 |
430 | cfi_restore (%r14) |
431 | movq 1072(%rsp), %r15 |
432 | cfi_restore (%r15) |
433 | jmp .LBL_2_2 |
434 | |
435 | .LBL_2_10: |
436 | cfi_restore_state |
437 | movzbl %r12b, %r15d |
438 | shlq $4, %r15 |
439 | vmovsd 1160(%rsp,%r15), %xmm0 |
440 | vzeroupper |
441 | vmovsd 1160(%rsp,%r15), %xmm0 |
442 | |
443 | call JUMPTARGET(sin) |
444 | |
445 | vmovsd %xmm0, 1224(%rsp,%r15) |
446 | jmp .LBL_2_8 |
447 | |
448 | .LBL_2_12: |
449 | movzbl %r12b, %r15d |
450 | shlq $4, %r15 |
451 | vmovsd 1152(%rsp,%r15), %xmm0 |
452 | vzeroupper |
453 | vmovsd 1152(%rsp,%r15), %xmm0 |
454 | |
455 | call JUMPTARGET(sin) |
456 | |
457 | vmovsd %xmm0, 1216(%rsp,%r15) |
458 | jmp .LBL_2_7 |
459 | #endif |
460 | END (_ZGVeN8v_sin_skx) |
461 | |
462 | .section .rodata, "a" |
463 | .L_2il0floatpacket.14: |
464 | .long 0xffffffff,0xffffffff |
465 | .type .L_2il0floatpacket.14,@object |
466 | |