1 | /* Function logf 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_logf_data.h" |
21 | #include "svml_s_wrapper_impl.h" |
22 | |
23 | .text |
24 | ENTRY (_ZGVeN16v_logf_knl) |
25 | #ifndef HAVE_AVX512DQ_ASM_SUPPORT |
26 | WRAPPER_IMPL_AVX512 _ZGVdN8v_logf |
27 | #else |
28 | /* |
29 | ALGORITHM DESCRIPTION: |
30 | |
31 | log(x) = exponent_x*log(2) + log(mantissa_x), if mantissa_x<4/3 |
32 | log(x) = (exponent_x+1)*log(2) + log(0.5*mantissa_x), if mantissa_x>4/3 |
33 | |
34 | R = mantissa_x - 1, if mantissa_x<4/3 |
35 | R = 0.5*mantissa_x - 1, if mantissa_x>4/3 |
36 | |R|< 1/3 |
37 | |
38 | log(1+R) is approximated as a polynomial: degree 9 for 1-ulp, |
39 | degree 7 for 4-ulp, degree 3 for half-precision. */ |
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_slog_data@GOTPCREL(%rip), %rax |
49 | movl $-1, %ecx |
50 | |
51 | /* reduction: compute r,n */ |
52 | vpsubd _iBrkValue(%rax), %zmm0, %zmm2 |
53 | vmovups _sPoly_7(%rax), %zmm7 |
54 | vpandd _iOffExpoMask(%rax), %zmm2, %zmm3 |
55 | |
56 | /* exponent_x (mantissa_x<4/3) or exponent_x+1 (mantissa_x>4/3) */ |
57 | vpsrad $23, %zmm2, %zmm4 |
58 | |
59 | /* check for working range, |
60 | set special argument mask (denormals/zero/Inf/NaN) |
61 | */ |
62 | vpaddd _iHiDelta(%rax), %zmm0, %zmm1 |
63 | |
64 | /* mantissa_x (mantissa_x<4/3), or 0.5*mantissa_x (mantissa_x>4/3) */ |
65 | vpaddd _iBrkValue(%rax), %zmm3, %zmm6 |
66 | vpcmpd $1, _iLoRange(%rax), %zmm1, %k1 |
67 | vcvtdq2ps {rn-sae}, %zmm4, %zmm1 |
68 | |
69 | /* reduced argument R */ |
70 | vsubps _sOne(%rax), %zmm6, %zmm8 |
71 | vpbroadcastd %ecx, %zmm5{%k1}{z} |
72 | |
73 | /* polynomial evaluation starts here */ |
74 | vfmadd213ps _sPoly_6(%rax), %zmm8, %zmm7 |
75 | vptestmd %zmm5, %zmm5, %k0 |
76 | kmovw %k0, %ecx |
77 | vfmadd213ps _sPoly_5(%rax), %zmm8, %zmm7 |
78 | vfmadd213ps _sPoly_4(%rax), %zmm8, %zmm7 |
79 | vfmadd213ps _sPoly_3(%rax), %zmm8, %zmm7 |
80 | vfmadd213ps _sPoly_2(%rax), %zmm8, %zmm7 |
81 | vfmadd213ps _sPoly_1(%rax), %zmm8, %zmm7 |
82 | vmulps %zmm8, %zmm7, %zmm9 |
83 | |
84 | /* polynomial evaluation end */ |
85 | vfmadd213ps %zmm8, %zmm8, %zmm9 |
86 | |
87 | /* |
88 | final reconstruction: |
89 | add exponent_value*log2 to polynomial result |
90 | */ |
91 | vfmadd132ps _sLn2(%rax), %zmm9, %zmm1 |
92 | testl %ecx, %ecx |
93 | jne .LBL_1_3 |
94 | |
95 | .LBL_1_2: |
96 | cfi_remember_state |
97 | vmovaps %zmm1, %zmm0 |
98 | movq %rbp, %rsp |
99 | cfi_def_cfa_register (%rsp) |
100 | popq %rbp |
101 | cfi_adjust_cfa_offset (-8) |
102 | cfi_restore (%rbp) |
103 | ret |
104 | |
105 | .LBL_1_3: |
106 | cfi_restore_state |
107 | vmovups %zmm0, 1152(%rsp) |
108 | vmovups %zmm1, 1216(%rsp) |
109 | je .LBL_1_2 |
110 | |
111 | xorb %dl, %dl |
112 | kmovw %k4, 1048(%rsp) |
113 | xorl %eax, %eax |
114 | kmovw %k5, 1040(%rsp) |
115 | kmovw %k6, 1032(%rsp) |
116 | kmovw %k7, 1024(%rsp) |
117 | vmovups %zmm16, 960(%rsp) |
118 | vmovups %zmm17, 896(%rsp) |
119 | vmovups %zmm18, 832(%rsp) |
120 | vmovups %zmm19, 768(%rsp) |
121 | vmovups %zmm20, 704(%rsp) |
122 | vmovups %zmm21, 640(%rsp) |
123 | vmovups %zmm22, 576(%rsp) |
124 | vmovups %zmm23, 512(%rsp) |
125 | vmovups %zmm24, 448(%rsp) |
126 | vmovups %zmm25, 384(%rsp) |
127 | vmovups %zmm26, 320(%rsp) |
128 | vmovups %zmm27, 256(%rsp) |
129 | vmovups %zmm28, 192(%rsp) |
130 | vmovups %zmm29, 128(%rsp) |
131 | vmovups %zmm30, 64(%rsp) |
132 | vmovups %zmm31, (%rsp) |
133 | movq %rsi, 1064(%rsp) |
134 | movq %rdi, 1056(%rsp) |
135 | movq %r12, 1096(%rsp) |
136 | cfi_offset_rel_rsp (12, 1096) |
137 | movb %dl, %r12b |
138 | movq %r13, 1088(%rsp) |
139 | cfi_offset_rel_rsp (13, 1088) |
140 | movl %ecx, %r13d |
141 | movq %r14, 1080(%rsp) |
142 | cfi_offset_rel_rsp (14, 1080) |
143 | movl %eax, %r14d |
144 | movq %r15, 1072(%rsp) |
145 | cfi_offset_rel_rsp (15, 1072) |
146 | cfi_remember_state |
147 | |
148 | .LBL_1_6: |
149 | btl %r14d, %r13d |
150 | jc .LBL_1_12 |
151 | |
152 | .LBL_1_7: |
153 | lea 1(%r14), %esi |
154 | btl %esi, %r13d |
155 | jc .LBL_1_10 |
156 | |
157 | .LBL_1_8: |
158 | addb $1, %r12b |
159 | addl $2, %r14d |
160 | cmpb $16, %r12b |
161 | jb .LBL_1_6 |
162 | |
163 | kmovw 1048(%rsp), %k4 |
164 | movq 1064(%rsp), %rsi |
165 | kmovw 1040(%rsp), %k5 |
166 | movq 1056(%rsp), %rdi |
167 | kmovw 1032(%rsp), %k6 |
168 | movq 1096(%rsp), %r12 |
169 | cfi_restore (%r12) |
170 | movq 1088(%rsp), %r13 |
171 | cfi_restore (%r13) |
172 | kmovw 1024(%rsp), %k7 |
173 | vmovups 960(%rsp), %zmm16 |
174 | vmovups 896(%rsp), %zmm17 |
175 | vmovups 832(%rsp), %zmm18 |
176 | vmovups 768(%rsp), %zmm19 |
177 | vmovups 704(%rsp), %zmm20 |
178 | vmovups 640(%rsp), %zmm21 |
179 | vmovups 576(%rsp), %zmm22 |
180 | vmovups 512(%rsp), %zmm23 |
181 | vmovups 448(%rsp), %zmm24 |
182 | vmovups 384(%rsp), %zmm25 |
183 | vmovups 320(%rsp), %zmm26 |
184 | vmovups 256(%rsp), %zmm27 |
185 | vmovups 192(%rsp), %zmm28 |
186 | vmovups 128(%rsp), %zmm29 |
187 | vmovups 64(%rsp), %zmm30 |
188 | vmovups (%rsp), %zmm31 |
189 | movq 1080(%rsp), %r14 |
190 | cfi_restore (%r14) |
191 | movq 1072(%rsp), %r15 |
192 | cfi_restore (%r15) |
193 | vmovups 1216(%rsp), %zmm1 |
194 | jmp .LBL_1_2 |
195 | |
196 | .LBL_1_10: |
197 | cfi_restore_state |
198 | movzbl %r12b, %r15d |
199 | vmovss 1156(%rsp,%r15,8), %xmm0 |
200 | call JUMPTARGET(logf) |
201 | vmovss %xmm0, 1220(%rsp,%r15,8) |
202 | jmp .LBL_1_8 |
203 | |
204 | .LBL_1_12: |
205 | movzbl %r12b, %r15d |
206 | vmovss 1152(%rsp,%r15,8), %xmm0 |
207 | call JUMPTARGET(logf) |
208 | vmovss %xmm0, 1216(%rsp,%r15,8) |
209 | jmp .LBL_1_7 |
210 | #endif |
211 | END (_ZGVeN16v_logf_knl) |
212 | |
213 | ENTRY (_ZGVeN16v_logf_skx) |
214 | #ifndef HAVE_AVX512DQ_ASM_SUPPORT |
215 | WRAPPER_IMPL_AVX512 _ZGVdN8v_logf |
216 | #else |
217 | /* |
218 | ALGORITHM DESCRIPTION: |
219 | |
220 | log(x) = exponent_x*log(2) + log(mantissa_x), if mantissa_x<4/3 |
221 | log(x) = (exponent_x+1)*log(2) + log(0.5*mantissa_x), if mantissa_x>4/3 |
222 | |
223 | R = mantissa_x - 1, if mantissa_x<4/3 |
224 | R = 0.5*mantissa_x - 1, if mantissa_x>4/3 |
225 | |R|< 1/3 |
226 | |
227 | log(1+R) is approximated as a polynomial: degree 9 for 1-ulp, |
228 | degree 7 for 4-ulp, degree 3 for half-precision. */ |
229 | |
230 | pushq %rbp |
231 | cfi_adjust_cfa_offset (8) |
232 | cfi_rel_offset (%rbp, 0) |
233 | movq %rsp, %rbp |
234 | cfi_def_cfa_register (%rbp) |
235 | andq $-64, %rsp |
236 | subq $1280, %rsp |
237 | movq __svml_slog_data@GOTPCREL(%rip), %rax |
238 | vmovups .L_2il0floatpacket.7(%rip), %zmm6 |
239 | vmovups _iBrkValue(%rax), %zmm4 |
240 | vmovups _sPoly_7(%rax), %zmm8 |
241 | |
242 | /* |
243 | check for working range, |
244 | set special argument mask (denormals/zero/Inf/NaN) |
245 | */ |
246 | vpaddd _iHiDelta(%rax), %zmm0, %zmm1 |
247 | |
248 | /* reduction: compute r,n */ |
249 | vpsubd %zmm4, %zmm0, %zmm2 |
250 | vpcmpd $5, _iLoRange(%rax), %zmm1, %k1 |
251 | |
252 | /* exponent_x (mantissa_x<4/3) or exponent_x+1 (mantissa_x>4/3) */ |
253 | vpsrad $23, %zmm2, %zmm5 |
254 | vpandd _iOffExpoMask(%rax), %zmm2, %zmm3 |
255 | |
256 | /* mantissa_x (mantissa_x<4/3), or 0.5*mantissa_x (mantissa_x>4/3) */ |
257 | vpaddd %zmm4, %zmm3, %zmm7 |
258 | |
259 | /* reduced argument R */ |
260 | vsubps _sOne(%rax), %zmm7, %zmm9 |
261 | |
262 | /* polynomial evaluation starts here */ |
263 | vfmadd213ps _sPoly_6(%rax), %zmm9, %zmm8 |
264 | vfmadd213ps _sPoly_5(%rax), %zmm9, %zmm8 |
265 | vfmadd213ps _sPoly_4(%rax), %zmm9, %zmm8 |
266 | vfmadd213ps _sPoly_3(%rax), %zmm9, %zmm8 |
267 | vfmadd213ps _sPoly_2(%rax), %zmm9, %zmm8 |
268 | vfmadd213ps _sPoly_1(%rax), %zmm9, %zmm8 |
269 | vmulps %zmm9, %zmm8, %zmm10 |
270 | |
271 | /* polynomial evaluation end */ |
272 | vfmadd213ps %zmm9, %zmm9, %zmm10 |
273 | vpandnd %zmm1, %zmm1, %zmm6{%k1} |
274 | vptestmd %zmm6, %zmm6, %k0 |
275 | vcvtdq2ps {rn-sae}, %zmm5, %zmm1 |
276 | kmovw %k0, %ecx |
277 | |
278 | /* |
279 | final reconstruction: |
280 | add exponent_value*log2 to polynomial result |
281 | */ |
282 | vfmadd132ps _sLn2(%rax), %zmm10, %zmm1 |
283 | testl %ecx, %ecx |
284 | jne .LBL_2_3 |
285 | |
286 | .LBL_2_2: |
287 | cfi_remember_state |
288 | vmovaps %zmm1, %zmm0 |
289 | movq %rbp, %rsp |
290 | cfi_def_cfa_register (%rsp) |
291 | popq %rbp |
292 | cfi_adjust_cfa_offset (-8) |
293 | cfi_restore (%rbp) |
294 | ret |
295 | |
296 | .LBL_2_3: |
297 | cfi_restore_state |
298 | vmovups %zmm0, 1152(%rsp) |
299 | vmovups %zmm1, 1216(%rsp) |
300 | je .LBL_2_2 |
301 | |
302 | xorb %dl, %dl |
303 | xorl %eax, %eax |
304 | kmovw %k4, 1048(%rsp) |
305 | kmovw %k5, 1040(%rsp) |
306 | kmovw %k6, 1032(%rsp) |
307 | kmovw %k7, 1024(%rsp) |
308 | vmovups %zmm16, 960(%rsp) |
309 | vmovups %zmm17, 896(%rsp) |
310 | vmovups %zmm18, 832(%rsp) |
311 | vmovups %zmm19, 768(%rsp) |
312 | vmovups %zmm20, 704(%rsp) |
313 | vmovups %zmm21, 640(%rsp) |
314 | vmovups %zmm22, 576(%rsp) |
315 | vmovups %zmm23, 512(%rsp) |
316 | vmovups %zmm24, 448(%rsp) |
317 | vmovups %zmm25, 384(%rsp) |
318 | vmovups %zmm26, 320(%rsp) |
319 | vmovups %zmm27, 256(%rsp) |
320 | vmovups %zmm28, 192(%rsp) |
321 | vmovups %zmm29, 128(%rsp) |
322 | vmovups %zmm30, 64(%rsp) |
323 | vmovups %zmm31, (%rsp) |
324 | movq %rsi, 1064(%rsp) |
325 | movq %rdi, 1056(%rsp) |
326 | movq %r12, 1096(%rsp) |
327 | cfi_offset_rel_rsp (12, 1096) |
328 | movb %dl, %r12b |
329 | movq %r13, 1088(%rsp) |
330 | cfi_offset_rel_rsp (13, 1088) |
331 | movl %ecx, %r13d |
332 | movq %r14, 1080(%rsp) |
333 | cfi_offset_rel_rsp (14, 1080) |
334 | movl %eax, %r14d |
335 | movq %r15, 1072(%rsp) |
336 | cfi_offset_rel_rsp (15, 1072) |
337 | cfi_remember_state |
338 | |
339 | .LBL_2_6: |
340 | btl %r14d, %r13d |
341 | jc .LBL_2_12 |
342 | |
343 | .LBL_2_7: |
344 | lea 1(%r14), %esi |
345 | btl %esi, %r13d |
346 | jc .LBL_2_10 |
347 | |
348 | .LBL_2_8: |
349 | incb %r12b |
350 | addl $2, %r14d |
351 | cmpb $16, %r12b |
352 | jb .LBL_2_6 |
353 | |
354 | kmovw 1048(%rsp), %k4 |
355 | kmovw 1040(%rsp), %k5 |
356 | kmovw 1032(%rsp), %k6 |
357 | kmovw 1024(%rsp), %k7 |
358 | vmovups 960(%rsp), %zmm16 |
359 | vmovups 896(%rsp), %zmm17 |
360 | vmovups 832(%rsp), %zmm18 |
361 | vmovups 768(%rsp), %zmm19 |
362 | vmovups 704(%rsp), %zmm20 |
363 | vmovups 640(%rsp), %zmm21 |
364 | vmovups 576(%rsp), %zmm22 |
365 | vmovups 512(%rsp), %zmm23 |
366 | vmovups 448(%rsp), %zmm24 |
367 | vmovups 384(%rsp), %zmm25 |
368 | vmovups 320(%rsp), %zmm26 |
369 | vmovups 256(%rsp), %zmm27 |
370 | vmovups 192(%rsp), %zmm28 |
371 | vmovups 128(%rsp), %zmm29 |
372 | vmovups 64(%rsp), %zmm30 |
373 | vmovups (%rsp), %zmm31 |
374 | vmovups 1216(%rsp), %zmm1 |
375 | movq 1064(%rsp), %rsi |
376 | movq 1056(%rsp), %rdi |
377 | movq 1096(%rsp), %r12 |
378 | cfi_restore (%r12) |
379 | movq 1088(%rsp), %r13 |
380 | cfi_restore (%r13) |
381 | movq 1080(%rsp), %r14 |
382 | cfi_restore (%r14) |
383 | movq 1072(%rsp), %r15 |
384 | cfi_restore (%r15) |
385 | jmp .LBL_2_2 |
386 | |
387 | .LBL_2_10: |
388 | cfi_restore_state |
389 | movzbl %r12b, %r15d |
390 | vmovss 1156(%rsp,%r15,8), %xmm0 |
391 | vzeroupper |
392 | vmovss 1156(%rsp,%r15,8), %xmm0 |
393 | |
394 | call JUMPTARGET(logf) |
395 | |
396 | vmovss %xmm0, 1220(%rsp,%r15,8) |
397 | jmp .LBL_2_8 |
398 | |
399 | .LBL_2_12: |
400 | movzbl %r12b, %r15d |
401 | vmovss 1152(%rsp,%r15,8), %xmm0 |
402 | vzeroupper |
403 | vmovss 1152(%rsp,%r15,8), %xmm0 |
404 | |
405 | call JUMPTARGET(logf) |
406 | |
407 | vmovss %xmm0, 1216(%rsp,%r15,8) |
408 | jmp .LBL_2_7 |
409 | |
410 | #endif |
411 | END (_ZGVeN16v_logf_skx) |
412 | |
413 | .section .rodata, "a" |
414 | .L_2il0floatpacket.7: |
415 | .long 0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff,0xffffffff |
416 | .type .L_2il0floatpacket.7,@object |
417 | |