1 /* Function exp vectorized with AVX-512. KNL and SKX versions.
2 Copyright (C) 2014-2023 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
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.
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.
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/>. */
20 #include "svml_d_exp_data.h"
21 #include "svml_d_wrapper_impl.h"
23 .section .text.evex512, "ax", @progbits
24 ENTRY (_ZGVeN8v_exp_knl)
26 ALGORITHM DESCRIPTION:
28 Argument representation:
29 N = rint(X*2^k/ln2) = 2^k*M+j
30 X = N*ln2/2^k + r = M*ln2 + ln2*(j/2^k) + r
31 then -ln2/2^(k+1) < r < ln2/2^(k+1)
37 exp(X) = exp(M*ln2 + ln2*(j/2^k) + r)
38 = 2^M * 2^(j/2^k) * exp(r)
39 2^M is calculated by bit manipulation
40 2^(j/2^k) is stored in table
41 exp(r) is approximated by polynomial
43 The table lookup is skipped if k = 0. */
46 cfi_adjust_cfa_offset (8)
47 cfi_rel_offset (%rbp, 0)
49 cfi_def_cfa_register (%rbp)
52 movq __svml_dexp_data@GOTPCREL(%rip), %rax
54 /* dR = X - dN*dbLn2hi, dbLn2hi is 52-8-k hi bits of ln2/2^k */
57 /* iAbsX = (int)(lX>>32), lX = *(longlong*)&X */
58 vpsrlq $32, %zmm0, %zmm1
60 /* iAbsX = iAbsX&iAbsMask */
65 /* iRangeMask = (iAbsX>iDomainRange) */
68 /* table lookup for dT[j] = 2^(j/2^k) */
69 vpxord %zmm11, %zmm11, %zmm11
70 vmovups __dbInvLn2(%rax), %zmm5
71 vmovups __dbLn2hi(%rax), %zmm7
74 /* dM = X*dbInvLn2+dbShifter, dbInvLn2 = 2^k/Ln2 */
75 vfmadd213pd __dbShifter(%rax), %zmm0, %zmm5
76 vmovups __dPC2(%rax), %zmm12
78 /* dN = dM-dbShifter, dN = rint(X*2^k/Ln2) */
79 vsubpd __dbShifter(%rax), %zmm5, %zmm9
80 vmovups __lIndexMask(%rax), %zmm4
81 vfnmadd231pd %zmm9, %zmm7, %zmm8
82 vpandd __iAbsMask(%rax), %zmm2, %zmm2{%k2}
84 /* lIndex = (*(longlong*)&dM)&lIndexMask, lIndex is the lower K bits of lM */
85 vpandq %zmm4, %zmm5, %zmm10
86 vgatherqpd (%rax,%zmm10,8), %zmm11{%k3}
87 vpcmpgtd __iDomainRange(%rax), %zmm2, %k1{%k2}
89 /* lM = (*(longlong*)&dM)&(~lIndexMask) */
90 vpandnq %zmm5, %zmm4, %zmm6
91 vpbroadcastd %ecx, %zmm3{%k1}{z}
93 /* lM = lM<<(52-K), 2^M */
94 vpsllq $42, %zmm6, %zmm14
96 /* dR = dR - dN*dbLn2lo, dbLn2lo is 40..94 bits of lo part of ln2/2^k */
97 vfnmadd132pd __dbLn2lo(%rax), %zmm8, %zmm9
99 /* Mask = iRangeMask?1:0, set mask for overflow/underflow */
100 vptestmd %zmm3, %zmm3, %k0{%k2}
102 /* exp(r) = b0+r*(b0+r*(b1+r*b2)) */
103 vfmadd213pd __dPC1(%rax), %zmm9, %zmm12
106 vfmadd213pd __dPC0(%rax), %zmm9, %zmm12
107 vfmadd213pd __dPC0(%rax), %zmm9, %zmm12
109 /* 2^(j/2^k) * exp(r) */
110 vmulpd %zmm12, %zmm11, %zmm13
112 /* multiply by 2^M through integer add */
113 vpaddq %zmm14, %zmm13, %zmm1
121 cfi_def_cfa_register (%rsp)
123 cfi_adjust_cfa_offset (-8)
129 vmovups %zmm0, 1152(%rsp)
130 vmovups %zmm1, 1216(%rsp)
134 kmovw %k4, 1048(%rsp)
136 kmovw %k5, 1040(%rsp)
137 kmovw %k6, 1032(%rsp)
138 kmovw %k7, 1024(%rsp)
139 vmovups %zmm16, 960(%rsp)
140 vmovups %zmm17, 896(%rsp)
141 vmovups %zmm18, 832(%rsp)
142 vmovups %zmm19, 768(%rsp)
143 vmovups %zmm20, 704(%rsp)
144 vmovups %zmm21, 640(%rsp)
145 vmovups %zmm22, 576(%rsp)
146 vmovups %zmm23, 512(%rsp)
147 vmovups %zmm24, 448(%rsp)
148 vmovups %zmm25, 384(%rsp)
149 vmovups %zmm26, 320(%rsp)
150 vmovups %zmm27, 256(%rsp)
151 vmovups %zmm28, 192(%rsp)
152 vmovups %zmm29, 128(%rsp)
153 vmovups %zmm30, 64(%rsp)
154 vmovups %zmm31, (%rsp)
155 movq %rsi, 1064(%rsp)
156 movq %rdi, 1056(%rsp)
157 movq %r12, 1096(%rsp)
158 cfi_offset_rel_rsp (12, 1096)
160 movq %r13, 1088(%rsp)
161 cfi_offset_rel_rsp (13, 1088)
163 movq %r14, 1080(%rsp)
164 cfi_offset_rel_rsp (14, 1080)
166 movq %r15, 1072(%rsp)
167 cfi_offset_rel_rsp (15, 1072)
185 kmovw 1048(%rsp), %k4
186 movq 1064(%rsp), %rsi
187 kmovw 1040(%rsp), %k5
188 movq 1056(%rsp), %rdi
189 kmovw 1032(%rsp), %k6
190 movq 1096(%rsp), %r12
192 movq 1088(%rsp), %r13
194 kmovw 1024(%rsp), %k7
195 vmovups 960(%rsp), %zmm16
196 vmovups 896(%rsp), %zmm17
197 vmovups 832(%rsp), %zmm18
198 vmovups 768(%rsp), %zmm19
199 vmovups 704(%rsp), %zmm20
200 vmovups 640(%rsp), %zmm21
201 vmovups 576(%rsp), %zmm22
202 vmovups 512(%rsp), %zmm23
203 vmovups 448(%rsp), %zmm24
204 vmovups 384(%rsp), %zmm25
205 vmovups 320(%rsp), %zmm26
206 vmovups 256(%rsp), %zmm27
207 vmovups 192(%rsp), %zmm28
208 vmovups 128(%rsp), %zmm29
209 vmovups 64(%rsp), %zmm30
210 vmovups (%rsp), %zmm31
211 movq 1080(%rsp), %r14
213 movq 1072(%rsp), %r15
215 vmovups 1216(%rsp), %zmm1
222 vmovsd 1160(%rsp,%r15), %xmm0
224 vmovsd %xmm0, 1224(%rsp,%r15)
230 vmovsd 1152(%rsp,%r15), %xmm0
232 vmovsd %xmm0, 1216(%rsp,%r15)
234 END (_ZGVeN8v_exp_knl)
236 ENTRY (_ZGVeN8v_exp_skx)
238 ALGORITHM DESCRIPTION:
240 Argument representation:
241 N = rint(X*2^k/ln2) = 2^k*M+j
242 X = N*ln2/2^k + r = M*ln2 + ln2*(j/2^k) + r
243 then -ln2/2^(k+1) < r < ln2/2^(k+1)
249 exp(X) = exp(M*ln2 + ln2*(j/2^k) + r)
250 = 2^M * 2^(j/2^k) * exp(r)
251 2^M is calculated by bit manipulation
252 2^(j/2^k) is stored in table
253 exp(r) is approximated by polynomial
255 The table lookup is skipped if k = 0. */
258 cfi_adjust_cfa_offset (8)
259 cfi_rel_offset (%rbp, 0)
261 cfi_def_cfa_register (%rbp)
264 movq __svml_dexp_data@GOTPCREL(%rip), %rax
266 /* table lookup for dT[j] = 2^(j/2^k) */
269 /* iAbsX = (int)(lX>>32), lX = *(longlong*)&X */
270 vpsrlq $32, %zmm0, %zmm1
271 vmovups __dbInvLn2(%rax), %zmm7
272 vmovups __dbShifter(%rax), %zmm5
273 vmovups __lIndexMask(%rax), %zmm6
274 vmovups __dbLn2hi(%rax), %zmm9
275 vmovups __dPC0(%rax), %zmm12
277 /* dM = X*dbInvLn2+dbShifter, dbInvLn2 = 2^k/Ln2 */
278 vfmadd213pd %zmm5, %zmm0, %zmm7
281 /* dN = dM-dbShifter, dN = rint(X*2^k/Ln2) */
282 vsubpd %zmm5, %zmm7, %zmm11
284 /* iAbsX = iAbsX&iAbsMask */
285 vpand __iAbsMask(%rax), %ymm2, %ymm3
287 /* dR = X - dN*dbLn2hi, dbLn2hi is 52-8-k hi bits of ln2/2^k */
288 vmovaps %zmm0, %zmm10
289 vfnmadd231pd %zmm11, %zmm9, %zmm10
290 vmovups __dPC2(%rax), %zmm9
292 /* dR = dR - dN*dbLn2lo, dbLn2lo is 40..94 bits of lo part of ln2/2^k */
293 vfnmadd132pd __dbLn2lo(%rax), %zmm10, %zmm11
295 /* exp(r) = b0+r*(b0+r*(b1+r*b2)) */
296 vfmadd213pd __dPC1(%rax), %zmm11, %zmm9
297 vfmadd213pd %zmm12, %zmm11, %zmm9
298 vfmadd213pd %zmm12, %zmm11, %zmm9
300 /* iRangeMask = (iAbsX>iDomainRange) */
301 vpcmpgtd __iDomainRange(%rax), %ymm3, %ymm4
303 /* Mask = iRangeMask?1:0, set mask for overflow/underflow */
304 vmovmskps %ymm4, %ecx
306 /* lIndex = (*(longlong*)&dM)&lIndexMask, lIndex is the lower K bits of lM */
307 vpandq %zmm6, %zmm7, %zmm13
308 vpmovqd %zmm13, %ymm14
309 vpxord %zmm15, %zmm15, %zmm15
310 vgatherdpd (%rax,%ymm14,8), %zmm15{%k1}
312 /* 2^(j/2^k) * exp(r) */
313 vmulpd %zmm9, %zmm15, %zmm10
315 /* lM = (*(longlong*)&dM)&(~lIndexMask) */
316 vpandnq %zmm7, %zmm6, %zmm8
318 /* lM = lM<<(52-K), 2^M */
319 vpsllq $42, %zmm8, %zmm1
321 /* multiply by 2^M through integer add */
322 vpaddq %zmm1, %zmm10, %zmm1
330 cfi_def_cfa_register (%rsp)
332 cfi_adjust_cfa_offset (-8)
338 vmovups %zmm0, 1152(%rsp)
339 vmovups %zmm1, 1216(%rsp)
344 kmovw %k4, 1048(%rsp)
345 kmovw %k5, 1040(%rsp)
346 kmovw %k6, 1032(%rsp)
347 kmovw %k7, 1024(%rsp)
348 vmovups %zmm16, 960(%rsp)
349 vmovups %zmm17, 896(%rsp)
350 vmovups %zmm18, 832(%rsp)
351 vmovups %zmm19, 768(%rsp)
352 vmovups %zmm20, 704(%rsp)
353 vmovups %zmm21, 640(%rsp)
354 vmovups %zmm22, 576(%rsp)
355 vmovups %zmm23, 512(%rsp)
356 vmovups %zmm24, 448(%rsp)
357 vmovups %zmm25, 384(%rsp)
358 vmovups %zmm26, 320(%rsp)
359 vmovups %zmm27, 256(%rsp)
360 vmovups %zmm28, 192(%rsp)
361 vmovups %zmm29, 128(%rsp)
362 vmovups %zmm30, 64(%rsp)
363 vmovups %zmm31, (%rsp)
364 movq %rsi, 1064(%rsp)
365 movq %rdi, 1056(%rsp)
366 movq %r12, 1096(%rsp)
367 cfi_offset_rel_rsp (12, 1096)
369 movq %r13, 1088(%rsp)
370 cfi_offset_rel_rsp (13, 1088)
372 movq %r14, 1080(%rsp)
373 cfi_offset_rel_rsp (14, 1080)
375 movq %r15, 1072(%rsp)
376 cfi_offset_rel_rsp (15, 1072)
394 kmovw 1048(%rsp), %k4
395 kmovw 1040(%rsp), %k5
396 kmovw 1032(%rsp), %k6
397 kmovw 1024(%rsp), %k7
398 vmovups 960(%rsp), %zmm16
399 vmovups 896(%rsp), %zmm17
400 vmovups 832(%rsp), %zmm18
401 vmovups 768(%rsp), %zmm19
402 vmovups 704(%rsp), %zmm20
403 vmovups 640(%rsp), %zmm21
404 vmovups 576(%rsp), %zmm22
405 vmovups 512(%rsp), %zmm23
406 vmovups 448(%rsp), %zmm24
407 vmovups 384(%rsp), %zmm25
408 vmovups 320(%rsp), %zmm26
409 vmovups 256(%rsp), %zmm27
410 vmovups 192(%rsp), %zmm28
411 vmovups 128(%rsp), %zmm29
412 vmovups 64(%rsp), %zmm30
413 vmovups (%rsp), %zmm31
414 vmovups 1216(%rsp), %zmm1
415 movq 1064(%rsp), %rsi
416 movq 1056(%rsp), %rdi
417 movq 1096(%rsp), %r12
419 movq 1088(%rsp), %r13
421 movq 1080(%rsp), %r14
423 movq 1072(%rsp), %r15
431 vmovsd 1160(%rsp,%r15), %xmm0
433 vmovsd 1160(%rsp,%r15), %xmm0
435 vmovsd %xmm0, 1224(%rsp,%r15)
441 vmovsd 1152(%rsp,%r15), %xmm0
443 vmovsd 1152(%rsp,%r15), %xmm0
445 vmovsd %xmm0, 1216(%rsp,%r15)
448 END (_ZGVeN8v_exp_skx)