sysdeps/x86_64/fpu/multiarch/svml_d_log28_core_avx512.S

   1 /* Function log2 vectorized with AVX-512.
   2    Copyright (C) 2021-2022 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  *    Get short reciprocal approximation Rcp ~ 1/mantissa(x)
  23  *    R = Rcp*x - 1.0
  24  *    log2(x) = k - log2(Rcp) + poly_approximation(R)
  25  *       log2(Rcp) is tabulated
  26  *
  27  *
  28  */
  29
  30 /* Offsets for data table __svml_dlog2_data_internal_avx512
  31  */
  32 #define Log_tbl                         0
  33 #define One                             128
  34 #define C075                            192
  35 #define poly_coeff9                     256
  36 #define poly_coeff8                     320
  37 #define poly_coeff7                     384
  38 #define poly_coeff6                     448
  39 #define poly_coeff5                     512
  40 #define poly_coeff4                     576
  41 #define poly_coeff3                     640
  42 #define poly_coeff2                     704
  43 #define poly_coeff1                     768
  44
  45 #include <sysdep.h>
  46
  47         .text
  48         .section .text.evex512,"ax",@progbits
  49 ENTRY(_ZGVeN8v_log2_skx)
  50         pushq     %rbp
  51         cfi_def_cfa_offset(16)
  52         movq      %rsp, %rbp
  53         cfi_def_cfa(6, 16)
  54         cfi_offset(6, -16)
  55         andq      $-64, %rsp
  56         subq      $192, %rsp
  57         vmovaps   %zmm0, %zmm7
  58         vgetmantpd $8, {sae}, %zmm7, %zmm6
  59         vmovups   One+__svml_dlog2_data_internal_avx512(%rip), %zmm2
  60         vmovups   poly_coeff5+__svml_dlog2_data_internal_avx512(%rip), %zmm12
  61         vmovups   poly_coeff3+__svml_dlog2_data_internal_avx512(%rip), %zmm13
  62
  63 /* Start polynomial evaluation */
  64         vmovups   poly_coeff9+__svml_dlog2_data_internal_avx512(%rip), %zmm10
  65         vmovups   poly_coeff8+__svml_dlog2_data_internal_avx512(%rip), %zmm0
  66         vmovups   poly_coeff7+__svml_dlog2_data_internal_avx512(%rip), %zmm11
  67         vmovups   poly_coeff6+__svml_dlog2_data_internal_avx512(%rip), %zmm14
  68
  69 /* Prepare exponent correction: DblRcp<0.75? */
  70         vmovups   C075+__svml_dlog2_data_internal_avx512(%rip), %zmm1
  71
  72 /* Table lookup */
  73         vmovups   __svml_dlog2_data_internal_avx512(%rip), %zmm4
  74
  75 /* GetExp(x) */
  76         vgetexppd {sae}, %zmm7, %zmm5
  77
  78 /* DblRcp ~ 1/Mantissa */
  79         vrcp14pd  %zmm6, %zmm8
  80
  81 /* x<=0? */
  82         vfpclasspd $94, %zmm7, %k0
  83
  84 /* round DblRcp to 4 fractional bits (RN mode, no Precision exception) */
  85         vrndscalepd $88, {sae}, %zmm8, %zmm3
  86         vmovups   poly_coeff4+__svml_dlog2_data_internal_avx512(%rip), %zmm8
  87         kmovw     %k0, %edx
  88
  89 /* Reduced argument: R = DblRcp*Mantissa - 1 */
  90         vfmsub213pd {rn-sae}, %zmm2, %zmm3, %zmm6
  91         vcmppd    $17, {sae}, %zmm1, %zmm3, %k1
  92         vfmadd231pd {rn-sae}, %zmm6, %zmm12, %zmm8
  93         vmovups   poly_coeff2+__svml_dlog2_data_internal_avx512(%rip), %zmm12
  94         vfmadd231pd {rn-sae}, %zmm6, %zmm10, %zmm0
  95         vfmadd231pd {rn-sae}, %zmm6, %zmm11, %zmm14
  96         vmovups   poly_coeff1+__svml_dlog2_data_internal_avx512(%rip), %zmm1
  97
  98 /* R^2 */
  99         vmulpd    {rn-sae}, %zmm6, %zmm6, %zmm15
 100         vfmadd231pd {rn-sae}, %zmm6, %zmm13, %zmm12
 101
 102 /* Prepare table index */
 103         vpsrlq    $48, %zmm3, %zmm9
 104
 105 /* add 1 to Expon if DblRcp<0.75 */
 106         vaddpd    {rn-sae}, %zmm2, %zmm5, %zmm5{%k1}
 107         vmulpd    {rn-sae}, %zmm15, %zmm15, %zmm13
 108         vfmadd213pd {rn-sae}, %zmm14, %zmm15, %zmm0
 109         vfmadd213pd {rn-sae}, %zmm12, %zmm15, %zmm8
 110         vpermt2pd Log_tbl+64+__svml_dlog2_data_internal_avx512(%rip), %zmm9, %zmm4
 111
 112 /* polynomial */
 113         vfmadd213pd {rn-sae}, %zmm8, %zmm13, %zmm0
 114         vfmadd213pd {rn-sae}, %zmm1, %zmm6, %zmm0
 115         vfmadd213pd {rn-sae}, %zmm4, %zmm0, %zmm6
 116         vaddpd    {rn-sae}, %zmm6, %zmm5, %zmm0
 117         testl     %edx, %edx
 118
 119 /* Go to special inputs processing branch */
 120         jne       L(SPECIAL_VALUES_BRANCH)
 121                                 # LOE rbx r12 r13 r14 r15 edx zmm0 zmm7
 122
 123 /* Restore registers
 124  * and exit the function
 125  */
 126
 127 L(EXIT):
 128         movq      %rbp, %rsp
 129         popq      %rbp
 130         cfi_def_cfa(7, 8)
 131         cfi_restore(6)
 132         ret
 133         cfi_def_cfa(6, 16)
 134         cfi_offset(6, -16)
 135
 136 /* Branch to process
 137  * special inputs
 138  */
 139
 140 L(SPECIAL_VALUES_BRANCH):
 141         vmovups   %zmm7, 64(%rsp)
 142         vmovups   %zmm0, 128(%rsp)
 143                                 # LOE rbx r12 r13 r14 r15 edx zmm0
 144
 145         xorl      %eax, %eax
 146                                 # LOE rbx r12 r13 r14 r15 eax edx
 147
 148         vzeroupper
 149         movq      %r12, 16(%rsp)
 150         /*  DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -176; DW_OP_plus)  */
 151         .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22
 152         movl      %eax, %r12d
 153         movq      %r13, 8(%rsp)
 154         /*  DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -184; DW_OP_plus)  */
 155         .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22
 156         movl      %edx, %r13d
 157         movq      %r14, (%rsp)
 158         /*  DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -192; DW_OP_plus)  */
 159         .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22
 160                                 # LOE rbx r15 r12d r13d
 161
 162 /* Range mask
 163  * bits check
 164  */
 165
 166 L(RANGEMASK_CHECK):
 167         btl       %r12d, %r13d
 168
 169 /* Call scalar math function */
 170         jc        L(SCALAR_MATH_CALL)
 171                                 # LOE rbx r15 r12d r13d
 172
 173 /* Special inputs
 174  * processing loop
 175  */
 176
 177 L(SPECIAL_VALUES_LOOP):
 178         incl      %r12d
 179         cmpl      $8, %r12d
 180
 181 /* Check bits in range mask */
 182         jl        L(RANGEMASK_CHECK)
 183                                 # LOE rbx r15 r12d r13d
 184
 185         movq      16(%rsp), %r12
 186         cfi_restore(12)
 187         movq      8(%rsp), %r13
 188         cfi_restore(13)
 189         movq      (%rsp), %r14
 190         cfi_restore(14)
 191         vmovups   128(%rsp), %zmm0
 192
 193 /* Go to exit */
 194         jmp       L(EXIT)
 195         /*  DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -176; DW_OP_plus)  */
 196         .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22
 197         /*  DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -184; DW_OP_plus)  */
 198         .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22
 199         /*  DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -192; DW_OP_plus)  */
 200         .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22
 201                                 # LOE rbx r12 r13 r14 r15 zmm0
 202
 203 /* Scalar math fucntion call
 204  * to process special input
 205  */
 206
 207 L(SCALAR_MATH_CALL):
 208         movl      %r12d, %r14d
 209         movsd     64(%rsp,%r14,8), %xmm0
 210         call      log2@PLT
 211                                 # LOE rbx r14 r15 r12d r13d xmm0
 212
 213         movsd     %xmm0, 128(%rsp,%r14,8)
 214
 215 /* Process special inputs in loop */
 216         jmp       L(SPECIAL_VALUES_LOOP)
 217                                 # LOE rbx r15 r12d r13d
 218 END(_ZGVeN8v_log2_skx)
 219
 220         .section .rodata, "a"
 221         .align 64
 222
 223 #ifdef __svml_dlog2_data_internal_avx512_typedef
 224 typedef unsigned int VUINT32;
 225 typedef struct {
 226         __declspec(align(64)) VUINT32 Log_tbl[16][2];
 227         __declspec(align(64)) VUINT32 One[8][2];
 228         __declspec(align(64)) VUINT32 C075[8][2];
 229         __declspec(align(64)) VUINT32 poly_coeff9[8][2];
 230         __declspec(align(64)) VUINT32 poly_coeff8[8][2];
 231         __declspec(align(64)) VUINT32 poly_coeff7[8][2];
 232         __declspec(align(64)) VUINT32 poly_coeff6[8][2];
 233         __declspec(align(64)) VUINT32 poly_coeff5[8][2];
 234         __declspec(align(64)) VUINT32 poly_coeff4[8][2];
 235         __declspec(align(64)) VUINT32 poly_coeff3[8][2];
 236         __declspec(align(64)) VUINT32 poly_coeff2[8][2];
 237         __declspec(align(64)) VUINT32 poly_coeff1[8][2];
 238    } __svml_dlog2_data_internal_avx512;
 239 #endif
 240 __svml_dlog2_data_internal_avx512:
 241         /*== Log_tbl ==*/
 242         .quad 0x0000000000000000
 243         .quad 0xbfb663f6fac91316
 244         .quad 0xbfc5c01a39fbd688
 245         .quad 0xbfcfbc16b902680a
 246         .quad 0xbfd49a784bcd1b8b
 247         .quad 0xbfd91bba891f1709
 248         .quad 0xbfdd6753e032ea0f
 249         .quad 0xbfe0c10500d63aa6
 250         .quad 0x3fda8ff971810a5e
 251         .quad 0x3fd6cb0f6865c8ea
 252         .quad 0x3fd32bfee370ee68
 253         .quad 0x3fcf5fd8a9063e35
 254         .quad 0x3fc8a8980abfbd32
 255         .quad 0x3fc22dadc2ab3497
 256         .quad 0x3fb7d60496cfbb4c
 257         .quad 0x3fa77394c9d958d5
 258         /*== One ==*/
 259         .align 64
 260         .quad 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000
 261         /*== C075 0.75 ==*/
 262         .align 64
 263         .quad 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000, 0x3fe8000000000000
 264         /*== poly_coeff9 ==*/
 265         .align 64
 266         .quad 0x3fc4904bda0e1d12, 0x3fc4904bda0e1d12, 0x3fc4904bda0e1d12, 0x3fc4904bda0e1d12, 0x3fc4904bda0e1d12, 0x3fc4904bda0e1d12, 0x3fc4904bda0e1d12, 0x3fc4904bda0e1d12
 267         /*== poly_coeff8 ==*/
 268         .align 64
 269         .quad 0xbfc71fb84deb5cce, 0xbfc71fb84deb5cce, 0xbfc71fb84deb5cce, 0xbfc71fb84deb5cce, 0xbfc71fb84deb5cce, 0xbfc71fb84deb5cce, 0xbfc71fb84deb5cce, 0xbfc71fb84deb5cce
 270         /*== poly_coeff7 ==*/
 271         .align 64
 272         .quad 0x3fca617351818613, 0x3fca617351818613, 0x3fca617351818613, 0x3fca617351818613, 0x3fca617351818613, 0x3fca617351818613, 0x3fca617351818613, 0x3fca617351818613
 273         /*== poly_coeff6 ==*/
 274         .align 64
 275         .quad 0xbfcec707e4e3144c, 0xbfcec707e4e3144c, 0xbfcec707e4e3144c, 0xbfcec707e4e3144c, 0xbfcec707e4e3144c, 0xbfcec707e4e3144c, 0xbfcec707e4e3144c, 0xbfcec707e4e3144c
 276         /*== poly_coeff5 ==*/
 277         .align 64
 278         .quad 0x3fd2776c5114d91a, 0x3fd2776c5114d91a, 0x3fd2776c5114d91a, 0x3fd2776c5114d91a, 0x3fd2776c5114d91a, 0x3fd2776c5114d91a, 0x3fd2776c5114d91a, 0x3fd2776c5114d91a
 279         /*== poly_coeff4 ==*/
 280         .align 64
 281         .quad 0xbfd71547653d0f8d, 0xbfd71547653d0f8d, 0xbfd71547653d0f8d, 0xbfd71547653d0f8d, 0xbfd71547653d0f8d, 0xbfd71547653d0f8d, 0xbfd71547653d0f8d, 0xbfd71547653d0f8d
 282         /*== poly_coeff3 ==*/
 283         .align 64
 284         .quad 0x3fdec709dc3a029f, 0x3fdec709dc3a029f, 0x3fdec709dc3a029f, 0x3fdec709dc3a029f, 0x3fdec709dc3a029f, 0x3fdec709dc3a029f, 0x3fdec709dc3a029f, 0x3fdec709dc3a029f
 285         /*== poly_coeff2 ==*/
 286         .align 64
 287         .quad 0xbfe71547652b82d4, 0xbfe71547652b82d4, 0xbfe71547652b82d4, 0xbfe71547652b82d4, 0xbfe71547652b82d4, 0xbfe71547652b82d4, 0xbfe71547652b82d4, 0xbfe71547652b82d4
 288         /*== poly_coeff1 ==*/
 289         .align 64
 290         .quad 0x3ff71547652b82fe, 0x3ff71547652b82fe, 0x3ff71547652b82fe, 0x3ff71547652b82fe, 0x3ff71547652b82fe, 0x3ff71547652b82fe, 0x3ff71547652b82fe, 0x3ff71547652b82fe
 291         .align 64
 292         .type   __svml_dlog2_data_internal_avx512,@object
 293         .size   __svml_dlog2_data_internal_avx512,.-__svml_dlog2_data_internal_avx512