1/* Function powf 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_s_powf_data.h"
21#include "svml_s_wrapper_impl.h"
22
23/*
24 ALGORITHM DESCRIPTION:
25
26 We are using the next identity : pow(x,y) = 2^(y * log2(x)).
27
28 1) log2(x) calculation
29 Here we use the following formula.
30 Let |x|=2^k1*X1, where k1 is integer, 1<=X1<2.
31 Let C ~= 1/ln(2),
32 Rcp1 ~= 1/X1, X2=Rcp1*X1,
33 Rcp2 ~= 1/X2, X3=Rcp2*X2,
34 Rcp3 ~= 1/X3, Rcp3C ~= C/X3.
35 Then
36 log2|x| = k1 + log2(1/Rcp1) + log2(1/Rcp2) + log2(C/Rcp3C) +
37 log2(X1*Rcp1*Rcp2*Rcp3C/C),
38 where X1*Rcp1*Rcp2*Rcp3C = C*(1+q), q is very small.
39
40 The values of Rcp1, log2(1/Rcp1), Rcp2, log2(1/Rcp2),
41 Rcp3C, log2(C/Rcp3C) are taken from tables.
42 Values of Rcp1, Rcp2, Rcp3C are such that RcpC=Rcp1*Rcp2*Rcp3C
43 is exactly represented in target precision.
44
45 log2(X1*Rcp1*Rcp2*Rcp3C/C) = log2(1+q) = ln(1+q)/ln2 =
46 = 1/(ln2)*q - 1/(2ln2)*q^2 + 1/(3ln2)*q^3 - ... =
47 = 1/(C*ln2)*cq - 1/(2*C^2*ln2)*cq^2 + 1/(3*C^3*ln2)*cq^3 - ... =
48 = (1 + a1)*cq + a2*cq^2 + a3*cq^3 + ...,
49 where
50 cq=X1*Rcp1*Rcp2*Rcp3C-C,
51 a1=1/(C*ln(2))-1 is small,
52 a2=1/(2*C^2*ln2),
53 a3=1/(3*C^3*ln2),
54 ...
55 Log2 result is split by three parts: HH+HL+HLL
56
57 2) Calculation of y*log2(x)
58 Split y into YHi+YLo.
59 Get high PH and medium PL parts of y*log2|x|.
60 Get low PLL part of y*log2|x|.
61 Now we have PH+PL+PLL ~= y*log2|x|.
62
63 3) Calculation of 2^(y*log2(x))
64 Let's represent PH+PL+PLL in the form N + j/2^expK + Z,
65 where expK=7 in this implementation, N and j are integers,
66 0<=j<=2^expK-1, |Z|<2^(-expK-1). Hence
67 2^(PH+PL+PLL) ~= 2^N * 2^(j/2^expK) * 2^Z,
68 where 2^(j/2^expK) is stored in a table, and
69 2^Z ~= 1 + B1*Z + B2*Z^2 ... + B5*Z^5.
70 We compute 2^(PH+PL+PLL) as follows:
71 Break PH into PHH + PHL, where PHH = N + j/2^expK.
72 Z = PHL + PL + PLL
73 Exp2Poly = B1*Z + B2*Z^2 ... + B5*Z^5
74 Get 2^(j/2^expK) from table in the form THI+TLO.
75 Now we have 2^(PH+PL+PLL) ~= 2^N * (THI + TLO) * (1 + Exp2Poly).
76 Get significand of 2^(PH+PL+PLL) in the form ResHi+ResLo:
77 ResHi := THI
78 ResLo := THI * Exp2Poly + TLO
79 Get exponent ERes of the result:
80 Res := ResHi + ResLo:
81 Result := ex(Res) + N. */
82
83 .text
84ENTRY (_ZGVeN16vv_powf_knl)
85#ifndef HAVE_AVX512DQ_ASM_SUPPORT
86WRAPPER_IMPL_AVX512_ff _ZGVdN8vv_powf
87#else
88 pushq %rbp
89 cfi_adjust_cfa_offset (8)
90 cfi_rel_offset (%rbp, 0)
91 movq %rsp, %rbp
92 cfi_def_cfa_register (%rbp)
93 andq $-64, %rsp
94 subq $1344, %rsp
95 movq __svml_spow_data@GOTPCREL(%rip), %rdx
96 vmovaps %zmm1, %zmm9
97 vshuff32x4 $238, %zmm0, %zmm0, %zmm7
98 kxnorw %k3, %k3, %k3
99 vcvtps2pd %ymm0, %zmm14
100 vcvtps2pd %ymm7, %zmm10
101 movl $-1, %eax
102 movq $-1, %rcx
103 vpandd _ABSMASK(%rdx), %zmm9, %zmm4
104 vmovups _ExpMask(%rdx), %zmm6
105
106/* exponent bits selection */
107 vpsrlq $20, %zmm14, %zmm13
108 vshuff32x4 $238, %zmm9, %zmm9, %zmm8
109 vpcmpd $5, _INF(%rdx), %zmm4, %k2
110 vpsrlq $32, %zmm13, %zmm15
111 vcvtps2pd %ymm8, %zmm2
112 vmovups _Two10(%rdx), %zmm4
113 vpmovqd %zmm15, %ymm12
114 vcvtps2pd %ymm9, %zmm1
115 vpsubd _NMINNORM(%rdx), %zmm0, %zmm3
116 vpbroadcastd %eax, %zmm8{%k2}{z}
117 vpcmpd $5, _NMAXVAL(%rdx), %zmm3, %k1
118
119/* preserve mantissa, set input exponent to 2^(-10) */
120 vmovaps %zmm6, %zmm3
121 vpternlogq $248, %zmm6, %zmm10, %zmm4
122 vpsrlq $20, %zmm10, %zmm10
123 vpternlogq $234, _Two10(%rdx), %zmm14, %zmm3
124
125/* reciprocal approximation good to at least 11 bits */
126 vrcp28pd %zmm4, %zmm11
127 vpsrlq $32, %zmm10, %zmm14
128 vpbroadcastd %eax, %zmm7{%k1}{z}
129 kxnorw %k1, %k1, %k1
130 vrcp28pd %zmm3, %zmm5
131 vpmovqd %zmm14, %ymm6
132 vshufi32x4 $68, %zmm6, %zmm12, %zmm13
133 vmovups _One(%rdx), %zmm6
134
135/* round reciprocal to nearest integer, will have 1+9 mantissa bits */
136 vrndscalepd $8, %zmm5, %zmm14
137
138/* biased exponent in DP format */
139 vshuff32x4 $238, %zmm13, %zmm13, %zmm5
140 vrndscalepd $8, %zmm11, %zmm11
141 vcmppd $30, _Threshold(%rdx), %zmm14, %k2
142 vcvtdq2pd %ymm13, %zmm10
143 vcvtdq2pd %ymm5, %zmm15
144
145/* table lookup */
146 vpsrlq $40, %zmm14, %zmm13
147 vpxord %zmm5, %zmm5, %zmm5
148 vgatherqpd _Log2Rcp_lookup(%rdx,%zmm13), %zmm5{%k3}
149 vfmsub213pd %zmm6, %zmm14, %zmm3
150 vfmsub213pd %zmm6, %zmm11, %zmm4
151 vcmppd $30, _Threshold(%rdx), %zmm11, %k3
152 vpbroadcastq %rcx, %zmm14{%k2}{z}
153
154/* dpP= _dbT+lJ*T_ITEM_GRAN */
155 kxnorw %k2, %k2, %k2
156 vpsrlq $40, %zmm11, %zmm12
157 vpxord %zmm6, %zmm6, %zmm6
158 vpbroadcastq %rcx, %zmm11{%k3}{z}
159 kxnorw %k3, %k3, %k3
160 vgatherqpd _Log2Rcp_lookup(%rdx,%zmm12), %zmm6{%k1}
161 vmovups _Bias1(%rdx), %zmm12
162 vpternlogq $236, _Bias(%rdx), %zmm12, %zmm14
163 vpternlogq $248, _Bias(%rdx), %zmm11, %zmm12
164 vsubpd %zmm14, %zmm10, %zmm13
165 vsubpd %zmm12, %zmm15, %zmm10
166 vmovups _poly_coeff_3(%rdx), %zmm11
167 vmovups _poly_coeff_4(%rdx), %zmm15
168 vfmadd213pd %zmm15, %zmm4, %zmm11
169 vmulpd %zmm4, %zmm4, %zmm12
170 vmovaps %zmm15, %zmm14
171 vmulpd %zmm3, %zmm3, %zmm15
172 vfmadd231pd _poly_coeff_3(%rdx), %zmm3, %zmm14
173
174/* reconstruction */
175 vfmadd213pd %zmm4, %zmm12, %zmm11
176 vfmadd213pd %zmm3, %zmm15, %zmm14
177 vaddpd %zmm6, %zmm11, %zmm11
178 vaddpd %zmm5, %zmm14, %zmm3
179 vfmadd231pd _L2(%rdx), %zmm10, %zmm11
180 vfmadd132pd _L2(%rdx), %zmm3, %zmm13
181 vmulpd %zmm2, %zmm11, %zmm12
182 vmulpd %zmm1, %zmm13, %zmm10
183 vmulpd __dbInvLn2(%rdx), %zmm12, %zmm6
184
185/* hi bits */
186 vpsrlq $32, %zmm12, %zmm12
187 vmulpd __dbInvLn2(%rdx), %zmm10, %zmm1
188
189/* to round down; if dR is an integer we will get R = 1, which is ok */
190 vsubpd __dbHALF(%rdx), %zmm6, %zmm4
191 vpsrlq $32, %zmm10, %zmm11
192 vpmovqd %zmm11, %ymm3
193 vsubpd __dbHALF(%rdx), %zmm1, %zmm2
194 vaddpd __dbShifter(%rdx), %zmm4, %zmm14
195 vpmovqd %zmm12, %ymm4
196 vshufi32x4 $68, %zmm4, %zmm3, %zmm5
197 vpxord %zmm4, %zmm4, %zmm4
198 vaddpd __dbShifter(%rdx), %zmm2, %zmm2
199
200/* iAbsX = iAbsX&iAbsMask; */
201 vpandd __iAbsMask(%rdx), %zmm5, %zmm11
202 vpxord %zmm5, %zmm5, %zmm5
203 vsubpd __dbShifter(%rdx), %zmm14, %zmm13
204
205/* iRangeMask = (iAbsX>iDomainRange) */
206 vpcmpgtd __iDomainRange(%rdx), %zmm11, %k1
207 vsubpd __dbShifter(%rdx), %zmm2, %zmm15
208 vpbroadcastd %eax, %zmm10{%k1}{z}
209 vpternlogd $254, %zmm8, %zmm7, %zmm10
210
211/* [0..1) */
212 vsubpd %zmm15, %zmm1, %zmm1
213
214/* low K bits */
215 vpandq __lbLOWKBITS(%rdx), %zmm14, %zmm11
216 vgatherqpd 13952(%rdx,%zmm11,8), %zmm5{%k3}
217 vsubpd %zmm13, %zmm6, %zmm7
218 vptestmd %zmm10, %zmm10, %k0
219 vpandq __lbLOWKBITS(%rdx), %zmm2, %zmm10
220 vmulpd __dbC1(%rdx), %zmm1, %zmm1
221 vmulpd __dbC1(%rdx), %zmm7, %zmm3
222 vpsrlq $11, %zmm2, %zmm8
223 vpsrlq $11, %zmm14, %zmm2
224
225/* NB : including +/- sign for the exponent!! */
226 vpsllq $52, %zmm8, %zmm8
227 kmovw %k0, %ecx
228 vpsllq $52, %zmm2, %zmm6
229 vfmadd213pd %zmm5, %zmm3, %zmm5
230 vgatherqpd 13952(%rdx,%zmm10,8), %zmm4{%k2}
231 vfmadd213pd %zmm4, %zmm1, %zmm4
232 vpaddq %zmm6, %zmm5, %zmm10
233 vcvtpd2ps %zmm10, %ymm12
234 vpaddq %zmm8, %zmm4, %zmm7
235 vcvtpd2ps %zmm7, %ymm11
236 vshuff32x4 $68, %zmm12, %zmm11, %zmm1
237 testl %ecx, %ecx
238 jne .LBL_1_3
239
240.LBL_1_2:
241 cfi_remember_state
242 vmovaps %zmm1, %zmm0
243 movq %rbp, %rsp
244 cfi_def_cfa_register (%rsp)
245 popq %rbp
246 cfi_adjust_cfa_offset (-8)
247 cfi_restore (%rbp)
248 ret
249
250.LBL_1_3:
251 cfi_restore_state
252 vmovups %zmm0, 1152(%rsp)
253 vmovups %zmm9, 1216(%rsp)
254 vmovups %zmm1, 1280(%rsp)
255 je .LBL_1_2
256
257 xorb %dl, %dl
258 kmovw %k4, 1048(%rsp)
259 xorl %eax, %eax
260 kmovw %k5, 1040(%rsp)
261 kmovw %k6, 1032(%rsp)
262 kmovw %k7, 1024(%rsp)
263 vmovups %zmm16, 960(%rsp)
264 vmovups %zmm17, 896(%rsp)
265 vmovups %zmm18, 832(%rsp)
266 vmovups %zmm19, 768(%rsp)
267 vmovups %zmm20, 704(%rsp)
268 vmovups %zmm21, 640(%rsp)
269 vmovups %zmm22, 576(%rsp)
270 vmovups %zmm23, 512(%rsp)
271 vmovups %zmm24, 448(%rsp)
272 vmovups %zmm25, 384(%rsp)
273 vmovups %zmm26, 320(%rsp)
274 vmovups %zmm27, 256(%rsp)
275 vmovups %zmm28, 192(%rsp)
276 vmovups %zmm29, 128(%rsp)
277 vmovups %zmm30, 64(%rsp)
278 vmovups %zmm31, (%rsp)
279 movq %rsi, 1064(%rsp)
280 movq %rdi, 1056(%rsp)
281 movq %r12, 1096(%rsp)
282 cfi_offset_rel_rsp (12, 1096)
283 movb %dl, %r12b
284 movq %r13, 1088(%rsp)
285 cfi_offset_rel_rsp (13, 1088)
286 movl %ecx, %r13d
287 movq %r14, 1080(%rsp)
288 cfi_offset_rel_rsp (14, 1080)
289 movl %eax, %r14d
290 movq %r15, 1072(%rsp)
291 cfi_offset_rel_rsp (15, 1072)
292 cfi_remember_state
293
294.LBL_1_6:
295 btl %r14d, %r13d
296 jc .LBL_1_12
297
298.LBL_1_7:
299 lea 1(%r14), %esi
300 btl %esi, %r13d
301 jc .LBL_1_10
302
303.LBL_1_8:
304 addb $1, %r12b
305 addl $2, %r14d
306 cmpb $16, %r12b
307 jb .LBL_1_6
308
309 kmovw 1048(%rsp), %k4
310 movq 1064(%rsp), %rsi
311 kmovw 1040(%rsp), %k5
312 movq 1056(%rsp), %rdi
313 kmovw 1032(%rsp), %k6
314 movq 1096(%rsp), %r12
315 cfi_restore (%r12)
316 movq 1088(%rsp), %r13
317 cfi_restore (%r13)
318 kmovw 1024(%rsp), %k7
319 vmovups 960(%rsp), %zmm16
320 vmovups 896(%rsp), %zmm17
321 vmovups 832(%rsp), %zmm18
322 vmovups 768(%rsp), %zmm19
323 vmovups 704(%rsp), %zmm20
324 vmovups 640(%rsp), %zmm21
325 vmovups 576(%rsp), %zmm22
326 vmovups 512(%rsp), %zmm23
327 vmovups 448(%rsp), %zmm24
328 vmovups 384(%rsp), %zmm25
329 vmovups 320(%rsp), %zmm26
330 vmovups 256(%rsp), %zmm27
331 vmovups 192(%rsp), %zmm28
332 vmovups 128(%rsp), %zmm29
333 vmovups 64(%rsp), %zmm30
334 vmovups (%rsp), %zmm31
335 movq 1080(%rsp), %r14
336 cfi_restore (%r14)
337 movq 1072(%rsp), %r15
338 cfi_restore (%r15)
339 vmovups 1280(%rsp), %zmm1
340 jmp .LBL_1_2
341
342.LBL_1_10:
343 cfi_restore_state
344 movzbl %r12b, %r15d
345 vmovss 1156(%rsp,%r15,8), %xmm0
346 vmovss 1220(%rsp,%r15,8), %xmm1
347 call JUMPTARGET(powf)
348 vmovss %xmm0, 1284(%rsp,%r15,8)
349 jmp .LBL_1_8
350
351.LBL_1_12:
352 movzbl %r12b, %r15d
353 vmovss 1152(%rsp,%r15,8), %xmm0
354 vmovss 1216(%rsp,%r15,8), %xmm1
355 call JUMPTARGET(powf)
356 vmovss %xmm0, 1280(%rsp,%r15,8)
357 jmp .LBL_1_7
358#endif
359END (_ZGVeN16vv_powf_knl)
360
361ENTRY (_ZGVeN16vv_powf_skx)
362#ifndef HAVE_AVX512DQ_ASM_SUPPORT
363WRAPPER_IMPL_AVX512_ff _ZGVdN8vv_powf
364#else
365 pushq %rbp
366 cfi_adjust_cfa_offset (8)
367 cfi_rel_offset (%rbp, 0)
368 movq %rsp, %rbp
369 cfi_def_cfa_register (%rbp)
370 andq $-64, %rsp
371 subq $1344, %rsp
372 movq __svml_spow_data@GOTPCREL(%rip), %rax
373 vextractf32x8 $1, %zmm1, %ymm14
374 vextractf32x8 $1, %zmm0, %ymm15
375 vpsubd _NMINNORM(%rax), %zmm0, %zmm9
376 vmovups %zmm26, 1280(%rsp)
377 vmovups _ExpMask(%rax), %zmm6
378 vpcmpd $1, _NMAXVAL(%rax), %zmm9, %k1
379 vcvtps2pd %ymm0, %zmm5
380 vcvtps2pd %ymm1, %zmm12
381 kxnorw %k3, %k3, %k3
382
383/* exponent bits selection */
384 vpsrlq $20, %zmm5, %zmm3
385 vpsrlq $32, %zmm3, %zmm2
386 vpmovqd %zmm2, %ymm11
387 vcvtps2pd %ymm14, %zmm13
388 vmovups .L_2il0floatpacket.23(%rip), %zmm14
389 vmovaps %zmm14, %zmm26
390 vpandd _ABSMASK(%rax), %zmm1, %zmm8
391 vpcmpd $1, _INF(%rax), %zmm8, %k2
392 vpandnd %zmm9, %zmm9, %zmm26{%k1}
393 vmovups _Two10(%rax), %zmm9
394 kxnorw %k1, %k1, %k1
395 vcvtps2pd %ymm15, %zmm4
396 vmovaps %zmm14, %zmm15
397
398/* preserve mantissa, set input exponent to 2^(-10) */
399 vpternlogq $248, %zmm6, %zmm4, %zmm9
400 vpsrlq $20, %zmm4, %zmm4
401
402/* reciprocal approximation good to at least 11 bits */
403 vrcp14pd %zmm9, %zmm10
404
405/* round reciprocal to nearest integer, will have 1+9 mantissa bits */
406 vrndscalepd $8, %zmm10, %zmm3
407 vmovups _One(%rax), %zmm10
408 vfmsub213pd %zmm10, %zmm3, %zmm9
409 vpandnd %zmm8, %zmm8, %zmm15{%k2}
410 vmovaps %zmm6, %zmm8
411 vpternlogq $234, _Two10(%rax), %zmm5, %zmm8
412 vpsrlq $32, %zmm4, %zmm5
413 vrcp14pd %zmm8, %zmm7
414 vpmovqd %zmm5, %ymm6
415 vrndscalepd $8, %zmm7, %zmm2
416 vfmsub213pd %zmm10, %zmm2, %zmm8
417
418/* table lookup */
419 vpsrlq $40, %zmm2, %zmm10
420 vinserti32x8 $1, %ymm6, %zmm11, %zmm4
421 vpsrlq $40, %zmm3, %zmm11
422
423/* biased exponent in DP format */
424 vextracti32x8 $1, %zmm4, %ymm7
425 vcvtdq2pd %ymm4, %zmm6
426 vpmovqd %zmm10, %ymm4
427 vpmovqd %zmm11, %ymm5
428 vpxord %zmm10, %zmm10, %zmm10
429 vgatherdpd _Log2Rcp_lookup(%rax,%ymm4), %zmm10{%k3}
430 vpbroadcastq .L_2il0floatpacket.24(%rip), %zmm4
431 vpxord %zmm11, %zmm11, %zmm11
432 vcvtdq2pd %ymm7, %zmm7
433 vgatherdpd _Log2Rcp_lookup(%rax,%ymm5), %zmm11{%k1}
434 vmovups _Threshold(%rax), %zmm5
435 vcmppd $21, %zmm2, %zmm5, %k2
436 vcmppd $21, %zmm3, %zmm5, %k3
437 vmovups _Bias1(%rax), %zmm3
438 vmovaps %zmm4, %zmm2
439 vpandnq %zmm5, %zmm5, %zmm2{%k2}
440 vpternlogq $236, _Bias(%rax), %zmm3, %zmm2
441
442/* dpP= _dbT+lJ*T_ITEM_GRAN */
443 kxnorw %k2, %k2, %k2
444 vpandnq %zmm5, %zmm5, %zmm4{%k3}
445 vpternlogq $248, _Bias(%rax), %zmm4, %zmm3
446 vsubpd %zmm2, %zmm6, %zmm4
447 vmovups _poly_coeff_3(%rax), %zmm6
448 vmovups _poly_coeff_4(%rax), %zmm2
449 vsubpd %zmm3, %zmm7, %zmm5
450 vmulpd %zmm8, %zmm8, %zmm7
451 vfmadd213pd %zmm2, %zmm9, %zmm6
452 kxnorw %k3, %k3, %k3
453 vmovaps %zmm2, %zmm3
454 vmulpd %zmm9, %zmm9, %zmm2
455 vfmadd231pd _poly_coeff_3(%rax), %zmm8, %zmm3
456
457/* reconstruction */
458 vfmadd213pd %zmm9, %zmm2, %zmm6
459 vfmadd213pd %zmm8, %zmm7, %zmm3
460 vaddpd %zmm11, %zmm6, %zmm8
461 vaddpd %zmm10, %zmm3, %zmm9
462 vfmadd231pd _L2(%rax), %zmm5, %zmm8
463 vfmadd132pd _L2(%rax), %zmm9, %zmm4
464 vmulpd %zmm13, %zmm8, %zmm13
465 vmulpd %zmm12, %zmm4, %zmm3
466 vmulpd __dbInvLn2(%rax), %zmm13, %zmm10
467 vmulpd __dbInvLn2(%rax), %zmm3, %zmm8
468
469/* hi bits */
470 vpsrlq $32, %zmm3, %zmm4
471 vpsrlq $32, %zmm13, %zmm13
472
473/* to round down; if dR is an integer we will get R = 1, which is ok */
474 vsubpd __dbHALF(%rax), %zmm8, %zmm12
475 vpmovqd %zmm4, %ymm5
476 vpmovqd %zmm13, %ymm2
477 vsubpd __dbHALF(%rax), %zmm10, %zmm9
478 vaddpd __dbShifter(%rax), %zmm12, %zmm7
479 vaddpd __dbShifter(%rax), %zmm9, %zmm9
480 vsubpd __dbShifter(%rax), %zmm7, %zmm11
481 vsubpd __dbShifter(%rax), %zmm9, %zmm12
482 vinserti32x8 $1, %ymm2, %zmm5, %zmm3
483
484/* iAbsX = iAbsX&iAbsMask */
485 vpandd __iAbsMask(%rax), %zmm3, %zmm4
486
487/* iRangeMask = (iAbsX>iDomainRange) */
488 vpcmpd $2, __iDomainRange(%rax), %zmm4, %k1
489 vpandnd %zmm4, %zmm4, %zmm14{%k1}
490 vpternlogd $254, %zmm15, %zmm26, %zmm14
491
492/* [0..1) */
493 vsubpd %zmm11, %zmm8, %zmm15
494 vsubpd %zmm12, %zmm10, %zmm26
495 vptestmd %zmm14, %zmm14, %k0
496 vpsrlq $11, %zmm7, %zmm8
497 vpsrlq $11, %zmm9, %zmm10
498 vmulpd __dbC1(%rax), %zmm26, %zmm26
499 vmulpd __dbC1(%rax), %zmm15, %zmm15
500
501/* NB : including +/- sign for the exponent!! */
502 vpsllq $52, %zmm10, %zmm13
503 vpsllq $52, %zmm8, %zmm12
504 kmovw %k0, %ecx
505
506/* low K bits */
507 vpandq __lbLOWKBITS(%rax), %zmm9, %zmm14
508 vpandq __lbLOWKBITS(%rax), %zmm7, %zmm6
509 vpmovqd %zmm14, %ymm7
510 vpmovqd %zmm6, %ymm9
511 vpxord %zmm2, %zmm2, %zmm2
512 vgatherdpd 13952(%rax,%ymm7,8), %zmm2{%k3}
513 vfmadd213pd %zmm2, %zmm26, %zmm2
514 vpaddq %zmm13, %zmm2, %zmm2
515 vcvtpd2ps %zmm2, %ymm4
516 vpxord %zmm11, %zmm11, %zmm11
517 vgatherdpd 13952(%rax,%ymm9,8), %zmm11{%k2}
518 vfmadd213pd %zmm11, %zmm15, %zmm11
519 vpaddq %zmm12, %zmm11, %zmm3
520 vcvtpd2ps %zmm3, %ymm5
521 vinsertf32x8 $1, %ymm4, %zmm5, %zmm2
522 testl %ecx, %ecx
523 jne .LBL_2_3
524
525.LBL_2_2:
526 cfi_remember_state
527 vmovups 1280(%rsp), %zmm26
528 vmovaps %zmm2, %zmm0
529 movq %rbp, %rsp
530 cfi_def_cfa_register (%rsp)
531 popq %rbp
532 cfi_adjust_cfa_offset (-8)
533 cfi_restore (%rbp)
534 ret
535
536.LBL_2_3:
537 cfi_restore_state
538 vmovups %zmm0, 1088(%rsp)
539 vmovups %zmm1, 1152(%rsp)
540 vmovups %zmm2, 1216(%rsp)
541 je .LBL_2_2
542
543 xorb %dl, %dl
544 xorl %eax, %eax
545 kmovw %k4, 984(%rsp)
546 kmovw %k5, 976(%rsp)
547 kmovw %k6, 968(%rsp)
548 kmovw %k7, 960(%rsp)
549 vmovups %zmm16, 896(%rsp)
550 vmovups %zmm17, 832(%rsp)
551 vmovups %zmm18, 768(%rsp)
552 vmovups %zmm19, 704(%rsp)
553 vmovups %zmm20, 640(%rsp)
554 vmovups %zmm21, 576(%rsp)
555 vmovups %zmm22, 512(%rsp)
556 vmovups %zmm23, 448(%rsp)
557 vmovups %zmm24, 384(%rsp)
558 vmovups %zmm25, 320(%rsp)
559 vmovups %zmm27, 256(%rsp)
560 vmovups %zmm28, 192(%rsp)
561 vmovups %zmm29, 128(%rsp)
562 vmovups %zmm30, 64(%rsp)
563 vmovups %zmm31, (%rsp)
564 movq %rsi, 1000(%rsp)
565 movq %rdi, 992(%rsp)
566 movq %r12, 1032(%rsp)
567 cfi_offset_rel_rsp (12, 1032)
568 movb %dl, %r12b
569 movq %r13, 1024(%rsp)
570 cfi_offset_rel_rsp (13, 1024)
571 movl %ecx, %r13d
572 movq %r14, 1016(%rsp)
573 cfi_offset_rel_rsp (14, 1016)
574 movl %eax, %r14d
575 movq %r15, 1008(%rsp)
576 cfi_offset_rel_rsp (15, 1008)
577 cfi_remember_state
578
579.LBL_2_6:
580 btl %r14d, %r13d
581 jc .LBL_2_12
582
583.LBL_2_7:
584 lea 1(%r14), %esi
585 btl %esi, %r13d
586 jc .LBL_2_10
587
588.LBL_2_8:
589 incb %r12b
590 addl $2, %r14d
591 cmpb $16, %r12b
592 jb .LBL_2_6
593
594 kmovw 984(%rsp), %k4
595 kmovw 976(%rsp), %k5
596 kmovw 968(%rsp), %k6
597 kmovw 960(%rsp), %k7
598 vmovups 896(%rsp), %zmm16
599 vmovups 832(%rsp), %zmm17
600 vmovups 768(%rsp), %zmm18
601 vmovups 704(%rsp), %zmm19
602 vmovups 640(%rsp), %zmm20
603 vmovups 576(%rsp), %zmm21
604 vmovups 512(%rsp), %zmm22
605 vmovups 448(%rsp), %zmm23
606 vmovups 384(%rsp), %zmm24
607 vmovups 320(%rsp), %zmm25
608 vmovups 256(%rsp), %zmm27
609 vmovups 192(%rsp), %zmm28
610 vmovups 128(%rsp), %zmm29
611 vmovups 64(%rsp), %zmm30
612 vmovups (%rsp), %zmm31
613 vmovups 1216(%rsp), %zmm2
614 movq 1000(%rsp), %rsi
615 movq 992(%rsp), %rdi
616 movq 1032(%rsp), %r12
617 cfi_restore (%r12)
618 movq 1024(%rsp), %r13
619 cfi_restore (%r13)
620 movq 1016(%rsp), %r14
621 cfi_restore (%r14)
622 movq 1008(%rsp), %r15
623 cfi_restore (%r15)
624 jmp .LBL_2_2
625
626.LBL_2_10:
627 cfi_restore_state
628 movzbl %r12b, %r15d
629 vmovss 1156(%rsp,%r15,8), %xmm1
630 vzeroupper
631 vmovss 1092(%rsp,%r15,8), %xmm0
632 call JUMPTARGET(powf)
633 vmovss %xmm0, 1220(%rsp,%r15,8)
634 jmp .LBL_2_8
635
636.LBL_2_12:
637 movzbl %r12b, %r15d
638 vmovss 1152(%rsp,%r15,8), %xmm1
639 vzeroupper
640 vmovss 1088(%rsp,%r15,8), %xmm0
641 call JUMPTARGET(powf)
642 vmovss %xmm0, 1216(%rsp,%r15,8)
643 jmp .LBL_2_7
644#endif
645END (_ZGVeN16vv_powf_skx)
646
647 .section .rodata, "a"
648.L_2il0floatpacket.23:
649 .long 0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff
650 .type .L_2il0floatpacket.23,@object
651.L_2il0floatpacket.24:
652 .long 0xffffffff,0xffffffff
653 .type .L_2il0floatpacket.24,@object
654