1 | /* Function coshf vectorized with SSE4. |
2 | Copyright (C) 2021-2024 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 | /* |
20 | * ALGORITHM DESCRIPTION: |
21 | * |
22 | * Compute cosh(x) as (exp(x)+exp(-x))/2, |
23 | * where exp is calculated as |
24 | * exp(M*ln2 + ln2*(j/2^k) + r) = 2^M * 2^(j/2^k) * exp(r) |
25 | * |
26 | * Special cases: |
27 | * |
28 | * cosh(NaN) = quiet NaN, and raise invalid exception |
29 | * cosh(INF) = that INF |
30 | * cosh(0) = 1 |
31 | * cosh(x) overflows for big x and returns MAXLOG+log(2) |
32 | * |
33 | */ |
34 | |
35 | /* Offsets for data table __svml_scosh_data_internal |
36 | */ |
37 | #define _sInvLn2 0 |
38 | #define _sLn2hi 16 |
39 | #define _sLn2lo 32 |
40 | #define _sSign 48 |
41 | #define _sShifter 64 |
42 | #define _iDomainRange 80 |
43 | #define _sPC1 96 |
44 | #define _sPC2 112 |
45 | #define _sPC3 128 |
46 | #define _sPC4 144 |
47 | #define _sPC5 160 |
48 | #define _sPC6 176 |
49 | #define _iHalf 192 |
50 | |
51 | #include <sysdep.h> |
52 | |
53 | .section .text.sse4, "ax" , @progbits |
54 | ENTRY(_ZGVbN4v_coshf_sse4) |
55 | subq $72, %rsp |
56 | cfi_def_cfa_offset(80) |
57 | |
58 | /* |
59 | * Implementation |
60 | * Abs argument |
61 | */ |
62 | movups _sSign+__svml_scosh_data_internal(%rip), %xmm1 |
63 | |
64 | /* |
65 | * Load argument |
66 | * dM = x/log(2) + RShifter |
67 | */ |
68 | movups _sInvLn2+__svml_scosh_data_internal(%rip), %xmm9 |
69 | andnps %xmm0, %xmm1 |
70 | mulps %xmm1, %xmm9 |
71 | |
72 | /* Check for overflow\underflow */ |
73 | movaps %xmm1, %xmm3 |
74 | movups _sShifter+__svml_scosh_data_internal(%rip), %xmm4 |
75 | movups _sLn2hi+__svml_scosh_data_internal(%rip), %xmm5 |
76 | addps %xmm4, %xmm9 |
77 | |
78 | /* |
79 | * R |
80 | * sN = sM - RShifter |
81 | */ |
82 | movaps %xmm9, %xmm6 |
83 | |
84 | /* |
85 | * G1, G2 2^N, 2^(-N) |
86 | * iM now is an EXP(2^N) |
87 | */ |
88 | pslld $23, %xmm9 |
89 | movups _sLn2lo+__svml_scosh_data_internal(%rip), %xmm7 |
90 | subps %xmm4, %xmm6 |
91 | |
92 | /* sR = sX - sN*Log2_hi */ |
93 | mulps %xmm6, %xmm5 |
94 | |
95 | /* sR = (sX - sN*Log2_hi) - sN*Log2_lo */ |
96 | mulps %xmm6, %xmm7 |
97 | movdqu _iDomainRange+__svml_scosh_data_internal(%rip), %xmm2 |
98 | pcmpgtd %xmm2, %xmm3 |
99 | pcmpeqd %xmm1, %xmm2 |
100 | |
101 | /* |
102 | * sinh(r) = r*((a1=1)+r^2*(a3+r^2*(a5+{v1 r^2*a7})))) = r + r*(r^2*(a3+r^2*(a5+r^2*a7))) .... |
103 | * sSinh_r = (a3+r^2*a5) |
104 | */ |
105 | movups _sPC5+__svml_scosh_data_internal(%rip), %xmm10 |
106 | por %xmm2, %xmm3 |
107 | |
108 | /* |
109 | * sinh(X) = sG2 + sG1*sinh(dR) + sG2*sR2*(a2+sR2*(a4+a6*sR2) |
110 | * sOut = (a4 +a6*sR2) |
111 | */ |
112 | movups _sPC6+__svml_scosh_data_internal(%rip), %xmm11 |
113 | subps %xmm5, %xmm1 |
114 | movmskps %xmm3, %edx |
115 | movdqu _iHalf+__svml_scosh_data_internal(%rip), %xmm8 |
116 | subps %xmm7, %xmm1 |
117 | |
118 | /* sR2 = sR^2, shaffled */ |
119 | movaps %xmm1, %xmm13 |
120 | movdqa %xmm8, %xmm2 |
121 | mulps %xmm1, %xmm13 |
122 | paddd %xmm9, %xmm2 |
123 | mulps %xmm13, %xmm10 |
124 | psubd %xmm9, %xmm8 |
125 | mulps %xmm13, %xmm11 |
126 | addps _sPC3+__svml_scosh_data_internal(%rip), %xmm10 |
127 | addps _sPC4+__svml_scosh_data_internal(%rip), %xmm11 |
128 | |
129 | /* sSinh_r = r^2*(a3+r^2*a5) */ |
130 | mulps %xmm13, %xmm10 |
131 | |
132 | /* sOut = a2+sR2*(a4+a6*sR2) */ |
133 | mulps %xmm13, %xmm11 |
134 | |
135 | /* sSinh_r = r + r*(r^2*(a3+r^2*a5)) */ |
136 | mulps %xmm1, %xmm10 |
137 | addps _sPC2+__svml_scosh_data_internal(%rip), %xmm11 |
138 | addps %xmm10, %xmm1 |
139 | |
140 | /* sOut = sR2*(a2+sR2*(a4+a6*sR2) */ |
141 | mulps %xmm11, %xmm13 |
142 | |
143 | /* sG1 = 2^(N-1)-2^(-N-1) */ |
144 | movdqa %xmm2, %xmm12 |
145 | |
146 | /* sG2 = 2^(N-1)+2^(-N-1) */ |
147 | addps %xmm8, %xmm2 |
148 | subps %xmm8, %xmm12 |
149 | |
150 | /* sOut = sG2*sR2*(a2+sR2*(a4+a6*sR2) */ |
151 | mulps %xmm2, %xmm13 |
152 | |
153 | /* sOut = sG1*sinh(dR)+sG2*sR2*(a2+sR2*(a4+a6*sR2) */ |
154 | mulps %xmm1, %xmm12 |
155 | addps %xmm12, %xmm13 |
156 | |
157 | /* sOut = sG2 + sG1*sinh(dR) + sG2*sR2*(a2+sR2*(a4+a6*sR2) */ |
158 | addps %xmm13, %xmm2 |
159 | |
160 | /* Ret H */ |
161 | testl %edx, %edx |
162 | |
163 | /* Go to special inputs processing branch */ |
164 | jne L(SPECIAL_VALUES_BRANCH) |
165 | # LOE rbx rbp r12 r13 r14 r15 edx xmm0 xmm2 |
166 | |
167 | /* Restore registers |
168 | * and exit the function |
169 | */ |
170 | |
171 | L(EXIT): |
172 | movaps %xmm2, %xmm0 |
173 | addq $72, %rsp |
174 | cfi_def_cfa_offset(8) |
175 | ret |
176 | cfi_def_cfa_offset(80) |
177 | |
178 | /* Branch to process |
179 | * special inputs |
180 | */ |
181 | |
182 | L(SPECIAL_VALUES_BRANCH): |
183 | movups %xmm0, 32(%rsp) |
184 | movups %xmm2, 48(%rsp) |
185 | # LOE rbx rbp r12 r13 r14 r15 edx |
186 | |
187 | xorl %eax, %eax |
188 | movq %r12, 16(%rsp) |
189 | cfi_offset(12, -64) |
190 | movl %eax, %r12d |
191 | movq %r13, 8(%rsp) |
192 | cfi_offset(13, -72) |
193 | movl %edx, %r13d |
194 | movq %r14, (%rsp) |
195 | cfi_offset(14, -80) |
196 | # LOE rbx rbp r15 r12d r13d |
197 | |
198 | /* Range mask |
199 | * bits check |
200 | */ |
201 | |
202 | L(RANGEMASK_CHECK): |
203 | btl %r12d, %r13d |
204 | |
205 | /* Call scalar math function */ |
206 | jc L(SCALAR_MATH_CALL) |
207 | # LOE rbx rbp r15 r12d r13d |
208 | |
209 | /* Special inputs |
210 | * processing loop |
211 | */ |
212 | |
213 | L(SPECIAL_VALUES_LOOP): |
214 | incl %r12d |
215 | cmpl $4, %r12d |
216 | |
217 | /* Check bits in range mask */ |
218 | jl L(RANGEMASK_CHECK) |
219 | # LOE rbx rbp r15 r12d r13d |
220 | |
221 | movq 16(%rsp), %r12 |
222 | cfi_restore(12) |
223 | movq 8(%rsp), %r13 |
224 | cfi_restore(13) |
225 | movq (%rsp), %r14 |
226 | cfi_restore(14) |
227 | movups 48(%rsp), %xmm2 |
228 | |
229 | /* Go to exit */ |
230 | jmp L(EXIT) |
231 | cfi_offset(12, -64) |
232 | cfi_offset(13, -72) |
233 | cfi_offset(14, -80) |
234 | # LOE rbx rbp r12 r13 r14 r15 xmm2 |
235 | |
236 | /* Scalar math function call |
237 | * to process special input |
238 | */ |
239 | |
240 | L(SCALAR_MATH_CALL): |
241 | movl %r12d, %r14d |
242 | movss 32(%rsp, %r14, 4), %xmm0 |
243 | call coshf@PLT |
244 | # LOE rbx rbp r14 r15 r12d r13d xmm0 |
245 | |
246 | movss %xmm0, 48(%rsp, %r14, 4) |
247 | |
248 | /* Process special inputs in loop */ |
249 | jmp L(SPECIAL_VALUES_LOOP) |
250 | # LOE rbx rbp r15 r12d r13d |
251 | END(_ZGVbN4v_coshf_sse4) |
252 | |
253 | .section .rodata, "a" |
254 | .align 16 |
255 | |
256 | #ifdef __svml_scosh_data_internal_typedef |
257 | typedef unsigned int VUINT32; |
258 | typedef struct { |
259 | __declspec(align(16)) VUINT32 _sInvLn2[4][1]; |
260 | __declspec(align(16)) VUINT32 _sLn2hi[4][1]; |
261 | __declspec(align(16)) VUINT32 _sLn2lo[4][1]; |
262 | __declspec(align(16)) VUINT32 _sSign[4][1]; |
263 | __declspec(align(16)) VUINT32 _sShifter[4][1]; |
264 | __declspec(align(16)) VUINT32 _iDomainRange[4][1]; |
265 | __declspec(align(16)) VUINT32 _sPC1[4][1]; |
266 | __declspec(align(16)) VUINT32 _sPC2[4][1]; |
267 | __declspec(align(16)) VUINT32 _sPC3[4][1]; |
268 | __declspec(align(16)) VUINT32 _sPC4[4][1]; |
269 | __declspec(align(16)) VUINT32 _sPC5[4][1]; |
270 | __declspec(align(16)) VUINT32 _sPC6[4][1]; |
271 | __declspec(align(16)) VUINT32 _iHalf[4][1]; |
272 | } __svml_scosh_data_internal; |
273 | #endif |
274 | __svml_scosh_data_internal: |
275 | .long 0x3FB8AA3B, 0x3FB8AA3B, 0x3FB8AA3B, 0x3FB8AA3B /* _sInvLn2 */ // k=0 |
276 | .align 16 |
277 | .long 0x3F317000, 0x3F317000, 0x3F317000, 0x3F317000 /* _sLn2hi */ |
278 | .align 16 |
279 | .long 0x3805fdf4, 0x3805fdf4, 0x3805fdf4, 0x3805fdf4 /* _sLn2lo */ |
280 | .align 16 |
281 | .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 /* _sSign */ |
282 | .align 16 |
283 | .long 0x4b400000, 0x4b400000, 0x4b400000, 0x4b400000 /* _sShifter */ |
284 | .align 16 |
285 | .long 0x42AEAC4E, 0x42AEAC4E, 0x42AEAC4E, 0x42AEAC4E /* _iDomainRange */ |
286 | .align 16 |
287 | .long 0x3F800000, 0x3F800000, 0x3F800000, 0x3F800000 /* _sPC1=1 */ |
288 | .align 16 |
289 | .long 0x3f000000, 0x3f000000, 0x3f000000, 0x3f000000 /* _sPC2 */ |
290 | .align 16 |
291 | .long 0x3e2aaa57, 0x3e2aaa57, 0x3e2aaa57, 0x3e2aaa57 /* _sPC3 */ |
292 | .align 16 |
293 | .long 0x3d2aaa72, 0x3d2aaa72, 0x3d2aaa72, 0x3d2aaa72 /* _sPC4 */ |
294 | .align 16 |
295 | .long 0x3c091461, 0x3c091461, 0x3c091461, 0x3c091461 /* _sPC5 */ |
296 | .align 16 |
297 | .long 0x3ab6a8a3, 0x3ab6a8a3, 0x3ab6a8a3, 0x3ab6a8a3 /* _sPC6 */ |
298 | // Integer constants |
299 | .align 16 |
300 | .long 0x3f000000, 0x3f000000, 0x3f000000, 0x3f000000 /* _iHalf */ |
301 | .align 16 |
302 | .type __svml_scosh_data_internal, @object |
303 | .size __svml_scosh_data_internal, .-__svml_scosh_data_internal |
304 | |