1 /* Function atan2 vectorized with AVX2.
2 Copyright (C) 2021-2023 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 * ALGORITHM DESCRIPTION:
21 * For 0.0 <= x <= 7.0/16.0: atan(x) = atan(0.0) + atan(s), where s=(x-0.0)/(1.0+0.0*x)
22 * For 7.0/16.0 <= x <= 11.0/16.0: atan(x) = atan(0.5) + atan(s), where s=(x-0.5)/(1.0+0.5*x)
23 * For 11.0/16.0 <= x <= 19.0/16.0: atan(x) = atan(1.0) + atan(s), where s=(x-1.0)/(1.0+1.0*x)
24 * For 19.0/16.0 <= x <= 39.0/16.0: atan(x) = atan(1.5) + atan(s), where s=(x-1.5)/(1.0+1.5*x)
25 * For 39.0/16.0 <= x <= inf : atan(x) = atan(inf) + atan(s), where s=-1.0/x
26 * Where atan(s) ~= s+s^3*Poly11(s^2) on interval |s|<7.0/0.16.
31 /* Offsets for data table __svml_datan2_data_internal
55 #define dSIGN_MASK 704
56 #define iCHK_WORK_SUB 736
57 #define iCHK_WORK_CMP 768
63 .section .text.avx2, "ax", @progbits
64 ENTRY(_ZGVdN4vv_atan2_avx2)
66 cfi_def_cfa_offset(16)
75 * #define NO_VECTOR_ZERO_ATAN2_ARGS
79 * The end of declarations
82 * Cannot be replaced by VQRCP(D, dR0, dB);
83 * Argument Absolute values
85 vmovupd dABS_MASK+__svml_datan2_data_internal(%rip), %ymm5
88 vmovupd dSIGN_MASK+__svml_datan2_data_internal(%rip), %ymm4
89 vmovups iCHK_WORK_SUB+__svml_datan2_data_internal(%rip), %xmm13
92 vandpd %ymm5, %ymm8, %ymm2
93 vandpd %ymm5, %ymm0, %ymm1
94 vcmpnlt_uqpd %ymm2, %ymm1, %ymm15
97 * 1) If y<x then a= y, b=x, PIO2=0
98 * 2) If y>x then a=-x, b=y, PIO2=Pi/2
100 vorpd %ymm4, %ymm2, %ymm6
101 vblendvpd %ymm15, %ymm6, %ymm1, %ymm3
102 vblendvpd %ymm15, %ymm1, %ymm2, %ymm6
103 vdivpd %ymm6, %ymm3, %ymm14
104 vmovups iCHK_WORK_CMP+__svml_datan2_data_internal(%rip), %xmm3
105 vmovupd %ymm6, 32(%rsp)
106 vandpd %ymm4, %ymm0, %ymm7
107 vandpd %ymm4, %ymm8, %ymm5
108 vandpd dPIO2+__svml_datan2_data_internal(%rip), %ymm15, %ymm4
110 /* Check if y and x are on main path. */
111 vextractf128 $1, %ymm2, %xmm9
112 vextractf128 $1, %ymm1, %xmm10
113 vshufps $221, %xmm9, %xmm2, %xmm11
114 vshufps $221, %xmm10, %xmm1, %xmm12
115 vpsubd %xmm13, %xmm11, %xmm0
116 vpsubd %xmm13, %xmm12, %xmm9
117 vpcmpgtd %xmm3, %xmm0, %xmm15
118 vpcmpeqd %xmm3, %xmm0, %xmm6
119 vpcmpgtd %xmm3, %xmm9, %xmm10
120 vpcmpeqd %xmm3, %xmm9, %xmm3
121 vpor %xmm6, %xmm15, %xmm11
122 vpor %xmm3, %xmm10, %xmm12
125 vmulpd %ymm14, %ymm14, %ymm10
126 vpor %xmm12, %xmm11, %xmm3
127 vmovupd dA18+__svml_datan2_data_internal(%rip), %ymm9
128 vmovupd dA17+__svml_datan2_data_internal(%rip), %ymm12
129 vmovupd dA16+__svml_datan2_data_internal(%rip), %ymm15
130 vmulpd %ymm10, %ymm10, %ymm11
132 /* if x<0, dPI = Pi, else dPI =0 */
133 vcmple_oqpd dZERO+__svml_datan2_data_internal(%rip), %ymm8, %ymm13
134 vmovmskps %xmm3, %eax
135 vmulpd %ymm11, %ymm11, %ymm0
136 vandpd __svml_datan2_data_internal(%rip), %ymm13, %ymm6
137 vmovupd dA19+__svml_datan2_data_internal(%rip), %ymm13
138 vfmadd213pd dA14+__svml_datan2_data_internal(%rip), %ymm0, %ymm9
139 vfmadd213pd dA13+__svml_datan2_data_internal(%rip), %ymm0, %ymm12
140 vfmadd213pd dA12+__svml_datan2_data_internal(%rip), %ymm0, %ymm15
141 vfmadd213pd dA15+__svml_datan2_data_internal(%rip), %ymm0, %ymm13
142 vfmadd213pd dA10+__svml_datan2_data_internal(%rip), %ymm0, %ymm9
143 vfmadd213pd dA09+__svml_datan2_data_internal(%rip), %ymm0, %ymm12
144 vfmadd213pd dA08+__svml_datan2_data_internal(%rip), %ymm0, %ymm15
145 vfmadd213pd dA11+__svml_datan2_data_internal(%rip), %ymm0, %ymm13
146 vfmadd213pd dA06+__svml_datan2_data_internal(%rip), %ymm0, %ymm9
147 vfmadd213pd dA05+__svml_datan2_data_internal(%rip), %ymm0, %ymm12
148 vfmadd213pd dA04+__svml_datan2_data_internal(%rip), %ymm0, %ymm15
149 vfmadd213pd dA07+__svml_datan2_data_internal(%rip), %ymm0, %ymm13
150 vfmadd213pd dA02+__svml_datan2_data_internal(%rip), %ymm0, %ymm9
151 vfmadd213pd dA01+__svml_datan2_data_internal(%rip), %ymm0, %ymm12
152 vfmadd213pd dA03+__svml_datan2_data_internal(%rip), %ymm0, %ymm13
154 /* A00=1.0, account for it later VQFMA(D, dP4, dP4, dR8, dA00); */
155 vmulpd %ymm15, %ymm0, %ymm0
156 vfmadd213pd %ymm9, %ymm10, %ymm13
157 vfmadd213pd %ymm0, %ymm10, %ymm12
158 vfmadd213pd %ymm12, %ymm11, %ymm13
162 * dP=(R+R*dP) + dPIO2
164 vfmadd213pd %ymm14, %ymm14, %ymm13
165 vaddpd %ymm13, %ymm4, %ymm14
166 vorpd %ymm5, %ymm14, %ymm0
167 vaddpd %ymm0, %ymm6, %ymm9
168 vorpd %ymm7, %ymm9, %ymm0
170 /* Special branch for fast (vector) processing of zero arguments */
173 /* Go to auxilary branch */
175 # LOE rbx r12 r13 r14 r15 edx xmm3 ymm0 ymm1 ymm2 ymm4 ymm5 ymm6 ymm7 ymm8
177 /* Return from auxilary branch
178 * for out of main path inputs
181 L(AUX_BRANCH_RETURN):
183 * Special branch for fast (vector) processing of zero arguments
184 * The end of implementation
188 /* Go to special inputs processing branch */
189 jne L(SPECIAL_VALUES_BRANCH)
190 # LOE rbx r12 r13 r14 r15 edx ymm0 ymm8
193 * and exit the function
209 L(SPECIAL_VALUES_BRANCH):
210 vmovupd (%rsp), %ymm1
211 vmovupd %ymm8, 64(%rsp)
212 vmovupd %ymm0, 96(%rsp)
213 vmovupd %ymm1, 32(%rsp)
214 # LOE rbx r12 r13 r14 r15 edx ymm0
217 # LOE rbx r12 r13 r14 r15 eax edx
221 /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -112; DW_OP_plus) */
222 .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x90, 0xff, 0xff, 0xff, 0x22
225 /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -120; DW_OP_plus) */
226 .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x88, 0xff, 0xff, 0xff, 0x22
229 /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -128; DW_OP_plus) */
230 .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x80, 0xff, 0xff, 0xff, 0x22
231 # LOE rbx r15 r12d r13d
240 /* Call scalar math function */
241 jc L(SCALAR_MATH_CALL)
242 # LOE rbx r15 r12d r13d
248 L(SPECIAL_VALUES_LOOP):
252 /* Check bits in range mask */
253 jl L(RANGEMASK_CHECK)
254 # LOE rbx r15 r12d r13d
262 vmovupd 96(%rsp), %ymm0
266 /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -112; DW_OP_plus) */
267 .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x90, 0xff, 0xff, 0xff, 0x22
268 /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -120; DW_OP_plus) */
269 .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x88, 0xff, 0xff, 0xff, 0x22
270 /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -128; DW_OP_plus) */
271 .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x80, 0xff, 0xff, 0xff, 0x22
272 # LOE rbx r12 r13 r14 r15 ymm0
274 /* Scalar math fucntion call
275 * to process special input
280 vmovsd 32(%rsp, %r14, 8), %xmm0
281 vmovsd 64(%rsp, %r14, 8), %xmm1
283 # LOE rbx r14 r15 r12d r13d xmm0
285 vmovsd %xmm0, 96(%rsp, %r14, 8)
287 /* Process special inputs in loop */
288 jmp L(SPECIAL_VALUES_LOOP)
292 # LOE rbx r15 r12d r13d
295 * for out of main path inputs
299 vmovupd (%rsp), %ymm11
301 /* Check if at least on of Y or Y is zero: iAXAYZERO */
302 vmovupd dZERO+__svml_datan2_data_internal(%rip), %ymm10
304 /* Check if both X & Y are not NaNs: iXYnotNAN */
305 vcmpordpd %ymm8, %ymm8, %ymm12
306 vcmpordpd %ymm11, %ymm11, %ymm13
307 vcmpeqpd %ymm10, %ymm2, %ymm2
308 vcmpeqpd %ymm10, %ymm1, %ymm1
309 vandpd %ymm13, %ymm12, %ymm14
310 vorpd %ymm1, %ymm2, %ymm2
311 vextractf128 $1, %ymm14, %xmm15
312 vextractf128 $1, %ymm2, %xmm11
313 vshufps $221, %xmm15, %xmm14, %xmm9
314 vshufps $221, %xmm11, %xmm2, %xmm12
317 * Path for zero arguments (at least one of both)
318 * Check if both args are zeros (den. is zero)
320 vcmpeqpd 32(%rsp), %ymm10, %ymm2
322 /* Check if at least on of Y or Y is zero and not NaN: iAXAYZEROnotNAN */
323 vpand %xmm9, %xmm12, %xmm1
325 /* Exclude from previous callout mask zero (and not NaN) arguments */
326 vpandn %xmm3, %xmm1, %xmm3
329 vmovmskps %xmm3, %edx
331 /* Set sPIO2 to zero if den. is zero */
332 vblendvpd %ymm2, %ymm10, %ymm4, %ymm4
333 vorpd %ymm5, %ymm4, %ymm5
335 /* Res = sign(Y)*(X<0)?(PIO2+PI):PIO2 */
336 vextractf128 $1, %ymm10, %xmm2
337 vextractf128 $1, %ymm8, %xmm3
338 vshufps $221, %xmm2, %xmm10, %xmm4
339 vshufps $221, %xmm3, %xmm8, %xmm9
340 vpcmpgtd %xmm9, %xmm4, %xmm12
341 vpshufd $80, %xmm12, %xmm11
342 vpshufd $250, %xmm12, %xmm13
343 vinsertf128 $1, %xmm13, %ymm11, %ymm14
344 vandpd %ymm6, %ymm14, %ymm6
345 vaddpd %ymm6, %ymm5, %ymm2
346 vorpd %ymm7, %ymm2, %ymm2
348 /* Merge results from main and spec path */
349 vpshufd $80, %xmm1, %xmm7
350 vpshufd $250, %xmm1, %xmm1
351 vinsertf128 $1, %xmm1, %ymm7, %ymm3
352 vblendvpd %ymm3, %ymm2, %ymm0, %ymm0
354 /* Return to main vector processing path */
355 jmp L(AUX_BRANCH_RETURN)
356 # LOE rbx r12 r13 r14 r15 edx ymm0 ymm8
357 END(_ZGVdN4vv_atan2_avx2)
359 .section .rodata, "a"
362 #ifdef __svml_datan2_data_internal_typedef
363 typedef unsigned int VUINT32;
365 __declspec(align(32)) VUINT32 dPI[4][2];
366 __declspec(align(32)) VUINT32 dPIO2[4][2];
367 __declspec(align(32)) VUINT32 dA19[4][2];
368 __declspec(align(32)) VUINT32 dA18[4][2];
369 __declspec(align(32)) VUINT32 dA17[4][2];
370 __declspec(align(32)) VUINT32 dA16[4][2];
371 __declspec(align(32)) VUINT32 dA15[4][2];
372 __declspec(align(32)) VUINT32 dA14[4][2];
373 __declspec(align(32)) VUINT32 dA13[4][2];
374 __declspec(align(32)) VUINT32 dA12[4][2];
375 __declspec(align(32)) VUINT32 dA11[4][2];
376 __declspec(align(32)) VUINT32 dA10[4][2];
377 __declspec(align(32)) VUINT32 dA09[4][2];
378 __declspec(align(32)) VUINT32 dA08[4][2];
379 __declspec(align(32)) VUINT32 dA07[4][2];
380 __declspec(align(32)) VUINT32 dA06[4][2];
381 __declspec(align(32)) VUINT32 dA05[4][2];
382 __declspec(align(32)) VUINT32 dA04[4][2];
383 __declspec(align(32)) VUINT32 dA03[4][2];
384 __declspec(align(32)) VUINT32 dA02[4][2];
385 __declspec(align(32)) VUINT32 dA01[4][2];
386 __declspec(align(32)) VUINT32 dA00[4][2];
387 __declspec(align(32)) VUINT32 dSIGN_MASK[4][2];
388 __declspec(align(32)) VUINT32 iCHK_WORK_SUB[8][1];
389 __declspec(align(32)) VUINT32 iCHK_WORK_CMP[8][1];
390 __declspec(align(32)) VUINT32 dABS_MASK[4][2];
391 __declspec(align(32)) VUINT32 dZERO[4][2];
392 } __svml_datan2_data_internal;
394 __svml_datan2_data_internal:
395 .quad 0x400921FB54442D18, 0x400921FB54442D18, 0x400921FB54442D18, 0x400921FB54442D18 // dPI
397 .quad 0x3FF921FB54442D18, 0x3FF921FB54442D18, 0x3FF921FB54442D18, 0x3FF921FB54442D18 // dPIO2
399 .quad 0xBEF4FDB537ABC7A3, 0xBEF4FDB537ABC7A3, 0xBEF4FDB537ABC7A3, 0xBEF4FDB537ABC7A3 // dA19
401 .quad 0x3F2CED0A36665209, 0x3F2CED0A36665209, 0x3F2CED0A36665209, 0x3F2CED0A36665209 // dA18
403 .quad 0xBF52E67C93954C23, 0xBF52E67C93954C23, 0xBF52E67C93954C23, 0xBF52E67C93954C23 // dA17
405 .quad 0x3F6F5A1DAE82AFB3, 0x3F6F5A1DAE82AFB3, 0x3F6F5A1DAE82AFB3, 0x3F6F5A1DAE82AFB3 // dA16
407 .quad 0xBF82B2EC618E4BAD, 0xBF82B2EC618E4BAD, 0xBF82B2EC618E4BAD, 0xBF82B2EC618E4BAD // dA15
409 .quad 0x3F914F4C661116A5, 0x3F914F4C661116A5, 0x3F914F4C661116A5, 0x3F914F4C661116A5 // dA14
411 .quad 0xBF9A5E83B081F69C, 0xBF9A5E83B081F69C, 0xBF9A5E83B081F69C, 0xBF9A5E83B081F69C // dA13
413 .quad 0x3FA169980CB6AD4F, 0x3FA169980CB6AD4F, 0x3FA169980CB6AD4F, 0x3FA169980CB6AD4F // dA12
415 .quad 0xBFA4EFA2E563C1BC, 0xBFA4EFA2E563C1BC, 0xBFA4EFA2E563C1BC, 0xBFA4EFA2E563C1BC // dA11
417 .quad 0x3FA7EC0FBC50683B, 0x3FA7EC0FBC50683B, 0x3FA7EC0FBC50683B, 0x3FA7EC0FBC50683B // dA10
419 .quad 0xBFAAD261EAA09954, 0xBFAAD261EAA09954, 0xBFAAD261EAA09954, 0xBFAAD261EAA09954 // dA09
421 .quad 0x3FAE1749BD612DCF, 0x3FAE1749BD612DCF, 0x3FAE1749BD612DCF, 0x3FAE1749BD612DCF // dA08
423 .quad 0xBFB11084009435E0, 0xBFB11084009435E0, 0xBFB11084009435E0, 0xBFB11084009435E0 // dA07
425 .quad 0x3FB3B12A49295651, 0x3FB3B12A49295651, 0x3FB3B12A49295651, 0x3FB3B12A49295651 // dA06
427 .quad 0xBFB745D009BADA94, 0xBFB745D009BADA94, 0xBFB745D009BADA94, 0xBFB745D009BADA94 // dA05
429 .quad 0x3FBC71C707F7D5B5, 0x3FBC71C707F7D5B5, 0x3FBC71C707F7D5B5, 0x3FBC71C707F7D5B5 // dA04
431 .quad 0xBFC2492491EE55C7, 0xBFC2492491EE55C7, 0xBFC2492491EE55C7, 0xBFC2492491EE55C7 // dA03
433 .quad 0x3FC999999997EE34, 0x3FC999999997EE34, 0x3FC999999997EE34, 0x3FC999999997EE34 // dA02
435 .quad 0xBFD55555555553C5, 0xBFD55555555553C5, 0xBFD55555555553C5, 0xBFD55555555553C5 // dA01
437 .quad 0x3FF0000000000000, 0x3FF0000000000000, 0x3FF0000000000000, 0x3FF0000000000000 // dA00
439 .quad 0x8000000000000000, 0x8000000000000000, 0x8000000000000000, 0x8000000000000000 // dSIGN_MASK
441 .long 0x80300000, 0x80300000, 0x80300000, 0x80300000, 0x80300000, 0x80300000, 0x80300000, 0x80300000 // iCHK_WORK_SUB
443 .long 0xfdd00000, 0xfdd00000, 0xfdd00000, 0xfdd00000, 0xfdd00000, 0xfdd00000, 0xfdd00000, 0xfdd00000 // iCHK_WORK_CMP
445 .quad 0x7fffffffffffffff, 0x7fffffffffffffff, 0x7fffffffffffffff, 0x7fffffffffffffff // dABS_MASK
447 .quad 0x0000000000000000, 0x0000000000000000, 0x0000000000000000, 0x0000000000000000 // dZERO
449 .type __svml_datan2_data_internal, @object
450 .size __svml_datan2_data_internal, .-__svml_datan2_data_internal