xref: /netbsd-src/external/lgpl3/gmp/dist/mpn/x86_64/skylake/gmp-mparam.h (revision 87d689fb734c654d2486f87f7be32f1b53ecdbec)
1 /* Skylake gmp-mparam.h -- Compiler/machine parameter header file.
2 
3 Copyright 1991, 1993, 1994, 2000-2015 Free Software Foundation, Inc.
4 
5 This file is part of the GNU MP Library.
6 
7 The GNU MP Library is free software; you can redistribute it and/or modify
8 it under the terms of either:
9 
10   * the GNU Lesser General Public License as published by the Free
11     Software Foundation; either version 3 of the License, or (at your
12     option) any later version.
13 
14 or
15 
16   * the GNU General Public License as published by the Free Software
17     Foundation; either version 2 of the License, or (at your option) any
18     later version.
19 
20 or both in parallel, as here.
21 
22 The GNU MP Library is distributed in the hope that it will be useful, but
23 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
24 or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
25 for more details.
26 
27 You should have received copies of the GNU General Public License and the
28 GNU Lesser General Public License along with the GNU MP Library.  If not,
29 see https://www.gnu.org/licenses/.  */
30 
31 #define GMP_LIMB_BITS 64
32 #define GMP_LIMB_BYTES 8
33 
34 /* Disable use of slow functions.  FIXME: We should disable lib inclusion.  */
35 #undef HAVE_NATIVE_mpn_mul_2
36 #undef HAVE_NATIVE_mpn_addmul_2
37 
38 /* 3500 MHz i5-6600K Skylake */
39 /* FFT tuning limit = 300 M */
40 /* Generated by tuneup.c, 2015-10-10, gcc 5.2 */
41 
42 #define MOD_1_NORM_THRESHOLD                 0  /* always */
43 #define MOD_1_UNNORM_THRESHOLD               0  /* always */
44 #define MOD_1N_TO_MOD_1_1_THRESHOLD          4
45 #define MOD_1U_TO_MOD_1_1_THRESHOLD          3
46 #define MOD_1_1_TO_MOD_1_2_THRESHOLD        13
47 #define MOD_1_2_TO_MOD_1_4_THRESHOLD        34
48 #define PREINV_MOD_1_TO_MOD_1_THRESHOLD      9
49 #define USE_PREINV_DIVREM_1                  1  /* native */
50 #define DIV_QR_1_NORM_THRESHOLD              1
51 #define DIV_QR_1_UNNORM_THRESHOLD        MP_SIZE_T_MAX  /* never */
52 #define DIV_QR_2_PI2_THRESHOLD           MP_SIZE_T_MAX  /* never */
53 #define DIVEXACT_1_THRESHOLD                 0  /* always (native) */
54 #define BMOD_1_TO_MOD_1_THRESHOLD           21
55 
56 #define MUL_TOOM22_THRESHOLD                26
57 #define MUL_TOOM33_THRESHOLD                73
58 #define MUL_TOOM44_THRESHOLD               208
59 #define MUL_TOOM6H_THRESHOLD               366
60 #define MUL_TOOM8H_THRESHOLD               430
61 
62 #define MUL_TOOM32_TO_TOOM43_THRESHOLD      73
63 #define MUL_TOOM32_TO_TOOM53_THRESHOLD     162
64 #define MUL_TOOM42_TO_TOOM53_THRESHOLD     137
65 #define MUL_TOOM42_TO_TOOM63_THRESHOLD     151
66 #define MUL_TOOM43_TO_TOOM54_THRESHOLD     203
67 
68 #define SQR_BASECASE_THRESHOLD               0  /* always (native) */
69 #define SQR_TOOM2_THRESHOLD                 34
70 #define SQR_TOOM3_THRESHOLD                109
71 #define SQR_TOOM4_THRESHOLD                450
72 #define SQR_TOOM6_THRESHOLD                562
73 #define SQR_TOOM8_THRESHOLD                  0  /* always */
74 
75 #define MULMID_TOOM42_THRESHOLD             46
76 
77 #define MULMOD_BNM1_THRESHOLD               15
78 #define SQRMOD_BNM1_THRESHOLD               17
79 
80 #define MUL_FFT_MODF_THRESHOLD             404  /* k = 5 */
81 #define MUL_FFT_TABLE3                                      \
82   { {    404, 5}, {     21, 6}, {     11, 5}, {     23, 6}, \
83     {     25, 7}, {     13, 6}, {     28, 7}, {     25, 8}, \
84     {     13, 7}, {     28, 8}, {     15, 7}, {     32, 8}, \
85     {     17, 7}, {     35, 8}, {     19, 7}, {     39, 8}, \
86     {     21, 9}, {     11, 8}, {     23, 7}, {     47, 8}, \
87     {     27, 9}, {     15, 8}, {     35, 9}, {     19, 8}, \
88     {     41, 9}, {     23, 8}, {     49, 9}, {     27,10}, \
89     {     15, 9}, {     43,10}, {     23, 9}, {     51,11}, \
90     {     15,10}, {     31, 9}, {     67,10}, {     39, 9}, \
91     {     83,10}, {     47, 9}, {     99,10}, {     55,11}, \
92     {     31,10}, {     79,11}, {     47,10}, {     95,12}, \
93     {     31,11}, {     63,10}, {    127, 6}, {   2047, 5}, \
94     {   4223, 6}, {   2175, 5}, {   4351, 7}, {   1215, 8}, \
95     {    639, 9}, {    335, 7}, {   1343, 8}, {    703,11}, \
96     {     95,10}, {    191, 9}, {    383, 8}, {    799,11}, \
97     {    111,12}, {     63,11}, {    127,10}, {    255, 9}, \
98     {    543,11}, {    143, 9}, {    607,11}, {    159, 9}, \
99     {    639,12}, {     95,10}, {    383, 9}, {    767,10}, \
100     {    415,13}, {     63,11}, {    255,10}, {    543,11}, \
101     {    287,10}, {    575,11}, {    303,12}, {    159,11}, \
102     {    319,10}, {    639,11}, {    335,10}, {    671,11}, \
103     {    351,10}, {    703,11}, {    415,10}, {    831,12}, \
104     {    223,11}, {    447,13}, {    127,11}, {    543,10}, \
105     {   1087,11}, {    607,12}, {    319,11}, {    671,12}, \
106     {    351,11}, {    703,12}, {    383,11}, {    767,12}, \
107     {    415,11}, {    831,12}, {    479,14}, {    127,12}, \
108     {    607,13}, {    319,12}, {    703,13}, {    383,12}, \
109     {    831,13}, {    447,12}, {    959,13}, {    511,12}, \
110     {   1087,13}, {    575,12}, {   1151,13}, {    639,12}, \
111     {   1279,13}, {    703,14}, {    383,13}, {    767,12}, \
112     {   1535,13}, {    831,12}, {   1727,13}, {    959,14}, \
113     {    511,13}, {   1087,12}, {   2175,13}, {   1151,14}, \
114     {    639,13}, {   1343,12}, {   2687,13}, {   1407,12}, \
115     {   2815,14}, {    767,13}, {   1599,12}, {   3199,13}, \
116     {   1727,14}, {    895,15}, {    511,14}, {   1023,13}, \
117     {   2175,14}, {   1151,13}, {   2431,12}, {   4863,14}, \
118     {   1279,13}, {   2687,14}, {   1407,13}, {   2815,15}, \
119     {    767,14}, {   1535,13}, {   3199,14}, {   1663,13}, \
120     {   3455,12}, {   6911,14}, {   1791,16}, {    511,15}, \
121     {   1023,14}, {   2175,13}, {   4351,14}, {   2431,13}, \
122     {   4863,15}, {   1279,14}, {   2943,13}, {   5887,15}, \
123     {   1535,14}, {   3455,15}, {   1791,14}, {   3839,16}, \
124     {   1023,15}, {   2047,14}, {   4351,15}, {   2303,14}, \
125     {   4863,15}, {   2559,14}, {   5247,15}, {   2815,14}, \
126     {   5887,16}, {   1535,15}, {   3327,14}, {   6911,15}, \
127     {   3839,17}, {   1023,16}, {   2047,15}, {   4863,16}, \
128     {   2559,15}, {   5887,14}, {  11775,16}, {   3071,15}, \
129     {   6911,16}, {   3583,15}, {   7679,14}, {  15359,17}, \
130     {   2047,16}, {   4095,15}, {   8191,16}, {  65536,17}, \
131     { 131072,18}, { 262144,19}, { 524288,20}, {1048576,21}, \
132     {2097152,22}, {4194304,23}, {8388608,24} }
133 #define MUL_FFT_TABLE3_SIZE 203
134 #define MUL_FFT_THRESHOLD                 4736
135 
136 #define SQR_FFT_MODF_THRESHOLD             368  /* k = 5 */
137 #define SQR_FFT_TABLE3                                      \
138   { {    368, 5}, {     23, 6}, {     25, 7}, {     13, 6}, \
139     {     27, 7}, {     25, 8}, {     13, 7}, {     28, 8}, \
140     {     15, 7}, {     31, 8}, {     21, 9}, {     11, 8}, \
141     {     27, 9}, {     15, 8}, {     35, 9}, {     19, 8}, \
142     {     39, 9}, {     23, 8}, {     47, 9}, {     27,10}, \
143     {     15, 9}, {     39,10}, {     23, 9}, {     51,11}, \
144     {     15,10}, {     31, 9}, {     67,10}, {     47, 9}, \
145     {     95,10}, {     55,11}, {     31,10}, {     79,11}, \
146     {     47,10}, {     95,12}, {     31,11}, {     63,10}, \
147     {    127, 9}, {    255,10}, {    135,11}, {     79,10}, \
148     {    159, 9}, {    319,11}, {     95,10}, {    191, 6}, \
149     {   3199, 7}, {   1727, 8}, {    895, 9}, {    479,10}, \
150     {    255, 8}, {   1023, 9}, {    543,10}, {    287, 9}, \
151     {    575, 8}, {   1151, 9}, {    607,11}, {    159, 9}, \
152     {    671,12}, {     95,11}, {    191, 9}, {    767,11}, \
153     {    207,13}, {     63,11}, {    255,10}, {    511, 9}, \
154     {   1023,10}, {    543, 9}, {   1087,10}, {    671,11}, \
155     {    351,10}, {    735,11}, {    383,10}, {    767,11}, \
156     {    415,10}, {    831,12}, {    223,11}, {    447,10}, \
157     {    895,11}, {    479,13}, {    127,12}, {    255,11}, \
158     {    607,12}, {    319,11}, {    735,12}, {    383,11}, \
159     {    767,12}, {    415,11}, {    831,12}, {    479,14}, \
160     {    127,12}, {    735,13}, {    383,12}, {    831,13}, \
161     {    447,12}, {    959,13}, {    511,12}, {   1023,13}, \
162     {    575,12}, {   1151,13}, {    639,12}, {   1279,13}, \
163     {    703,14}, {    383,13}, {    767,12}, {   1535,13}, \
164     {    831,12}, {   1663,13}, {    959,14}, {    511,13}, \
165     {   1087,12}, {   2175,13}, {   1151,14}, {    639,13}, \
166     {   1343,12}, {   2687,13}, {   1471,14}, {    767,13}, \
167     {   1599,12}, {   3199,13}, {   1663,14}, {    895,15}, \
168     {    511,14}, {   1023,13}, {   2175,14}, {   1151,13}, \
169     {   2431,12}, {   4863,14}, {   1279,13}, {   2687,14}, \
170     {   1407,15}, {    767,14}, {   1535,13}, {   3199,14}, \
171     {   1663,13}, {   3455,12}, {   6911,16}, {    511,15}, \
172     {   1023,14}, {   2303,13}, {   4607,14}, {   2431,13}, \
173     {   4863,15}, {   1279,14}, {   2943,13}, {   5887,15}, \
174     {   1535,14}, {   3455,13}, {   6911,15}, {   1791,14}, \
175     {   3839,16}, {   1023,15}, {   2047,14}, {   4351,15}, \
176     {   2303,14}, {   4863,15}, {   2559,14}, {   5119,15}, \
177     {   2815,14}, {   5887,16}, {   1535,15}, {   3071,14}, \
178     {   6143,15}, {   3327,14}, {   6911,15}, {   3839,17}, \
179     {   1023,16}, {   2047,15}, {   4863,16}, {   2559,15}, \
180     {   5887,14}, {  11775,16}, {   3071,15}, {   6911,16}, \
181     {   3583,15}, {   7679,14}, {  15359,17}, {   2047,16}, \
182     {   4095,15}, {   8191,16}, {  65536,17}, { 131072,18}, \
183     { 262144,19}, { 524288,20}, {1048576,21}, {2097152,22}, \
184     {4194304,23}, {8388608,24} }
185 #define SQR_FFT_TABLE3_SIZE 186
186 #define SQR_FFT_THRESHOLD                 3712
187 
188 #define MULLO_BASECASE_THRESHOLD             0  /* always */
189 #define MULLO_DC_THRESHOLD                  27
190 #define MULLO_MUL_N_THRESHOLD             9174
191 #define SQRLO_BASECASE_THRESHOLD             8
192 #define SQRLO_DC_THRESHOLD                 132
193 #define SQRLO_SQR_THRESHOLD               7269
194 
195 #define DC_DIV_QR_THRESHOLD                 52
196 #define DC_DIVAPPR_Q_THRESHOLD             173
197 #define DC_BDIV_QR_THRESHOLD                87
198 #define DC_BDIV_Q_THRESHOLD                156
199 
200 #define INV_MULMOD_BNM1_THRESHOLD           50
201 #define INV_NEWTON_THRESHOLD               165
202 #define INV_APPR_THRESHOLD                 165
203 
204 #define BINV_NEWTON_THRESHOLD              294
205 #define REDC_1_TO_REDC_2_THRESHOLD          30
206 #define REDC_2_TO_REDC_N_THRESHOLD          75
207 
208 #define MU_DIV_QR_THRESHOLD               1620
209 #define MU_DIVAPPR_Q_THRESHOLD            1620
210 #define MUPI_DIV_QR_THRESHOLD               62
211 #define MU_BDIV_QR_THRESHOLD              1589
212 #define MU_BDIV_Q_THRESHOLD               1718
213 
214 #define POWM_SEC_TABLE  1,9,194,960,1133,1730
215 
216 #define GET_STR_DC_THRESHOLD                10
217 #define GET_STR_PRECOMPUTE_THRESHOLD        17
218 #define SET_STR_DC_THRESHOLD              1065
219 #define SET_STR_PRECOMPUTE_THRESHOLD      2024
220 
221 #define FAC_DSC_THRESHOLD                  767
222 #define FAC_ODD_THRESHOLD                    0  /* always */
223 
224 #define MATRIX22_STRASSEN_THRESHOLD         15
225 #define HGCD_THRESHOLD                      71
226 #define HGCD_APPR_THRESHOLD                 55
227 #define HGCD_REDUCE_THRESHOLD             3014
228 #define GCD_DC_THRESHOLD                   706
229 #define GCDEXT_DC_THRESHOLD                393
230 #define JACOBI_BASE_METHOD                   3
231