1 /* Intel Goldmont gmp-mparam.h -- Compiler/machine parameter header file.
2
3 Copyright 2019 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 /* 2200 MHz Intel Atom C3758 Goldmont/Denverton */
35 /* FFT tuning limit = 468,030,122 */
36 /* Generated by tuneup.c, 2019-10-12, gcc 8.3 */
37
38 #define MOD_1_NORM_THRESHOLD 3
39 #define MOD_1_UNNORM_THRESHOLD 5
40 #define MOD_1N_TO_MOD_1_1_THRESHOLD 5
41 #define MOD_1U_TO_MOD_1_1_THRESHOLD 4
42 #define MOD_1_1_TO_MOD_1_2_THRESHOLD 13
43 #define MOD_1_2_TO_MOD_1_4_THRESHOLD 38
44 #define PREINV_MOD_1_TO_MOD_1_THRESHOLD 9
45 #define USE_PREINV_DIVREM_1 1 /* native */
46 #define DIV_QR_1_NORM_THRESHOLD 3
47 #define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */
48 #define DIV_QR_2_PI2_THRESHOLD 17
49 #define DIVEXACT_1_THRESHOLD 0 /* always (native) */
50 #define BMOD_1_TO_MOD_1_THRESHOLD 19
51
52 #define DIV_1_VS_MUL_1_PERCENT 301
53
54 #define MUL_TOOM22_THRESHOLD 23
55 #define MUL_TOOM33_THRESHOLD 65
56 #define MUL_TOOM44_THRESHOLD 178
57 #define MUL_TOOM6H_THRESHOLD 258
58 #define MUL_TOOM8H_THRESHOLD 357
59
60 #define MUL_TOOM32_TO_TOOM43_THRESHOLD 121
61 #define MUL_TOOM32_TO_TOOM53_THRESHOLD 131
62 #define MUL_TOOM42_TO_TOOM53_THRESHOLD 121
63 #define MUL_TOOM42_TO_TOOM63_THRESHOLD 129
64 #define MUL_TOOM43_TO_TOOM54_THRESHOLD 178
65
66 #define SQR_BASECASE_THRESHOLD 0 /* always (native) */
67 #define SQR_TOOM2_THRESHOLD 30
68 #define SQR_TOOM3_THRESHOLD 113
69 #define SQR_TOOM4_THRESHOLD 290
70 #define SQR_TOOM6_THRESHOLD 351
71 #define SQR_TOOM8_THRESHOLD 0 /* always */
72
73 #define MULMID_TOOM42_THRESHOLD 36
74
75 #define MULMOD_BNM1_THRESHOLD 14
76 #define SQRMOD_BNM1_THRESHOLD 16
77
78 #define MUL_FFT_MODF_THRESHOLD 440 /* k = 5 */
79 #define MUL_FFT_TABLE3 \
80 { { 440, 5}, { 19, 6}, { 10, 5}, { 21, 6}, \
81 { 11, 5}, { 23, 6}, { 21, 7}, { 11, 6}, \
82 { 24, 7}, { 13, 6}, { 27, 7}, { 15, 6}, \
83 { 31, 7}, { 21, 8}, { 11, 7}, { 24, 8}, \
84 { 13, 7}, { 28, 8}, { 15, 7}, { 31, 8}, \
85 { 17, 7}, { 35, 8}, { 19, 7}, { 39, 8}, \
86 { 21, 9}, { 11, 8}, { 27, 9}, { 15, 8}, \
87 { 35, 9}, { 19, 8}, { 41, 9}, { 23, 8}, \
88 { 47, 9}, { 27,10}, { 15, 9}, { 39,10}, \
89 { 23, 9}, { 51,11}, { 15,10}, { 31, 9}, \
90 { 71,10}, { 39, 9}, { 83,10}, { 47, 9}, \
91 { 95,10}, { 55,11}, { 31,10}, { 79,11}, \
92 { 47,10}, { 103,12}, { 31,11}, { 63,10}, \
93 { 135,11}, { 79,10}, { 159,11}, { 95,10}, \
94 { 191,11}, { 111,12}, { 63,11}, { 127,10}, \
95 { 255,11}, { 143,10}, { 287, 9}, { 575,11}, \
96 { 159,10}, { 319,12}, { 95,11}, { 191,10}, \
97 { 383, 9}, { 767,11}, { 207,10}, { 415,13}, \
98 { 63,12}, { 127,11}, { 255,10}, { 511,11}, \
99 { 271,10}, { 543,11}, { 287,10}, { 575,11}, \
100 { 303,12}, { 159,11}, { 319,10}, { 639,11}, \
101 { 367,12}, { 191,11}, { 383,10}, { 767,11}, \
102 { 415,12}, { 223,11}, { 479,13}, { 127,12}, \
103 { 255,11}, { 543,12}, { 287,11}, { 607,12}, \
104 { 319,11}, { 639,12}, { 351,11}, { 703,13}, \
105 { 191,12}, { 383,11}, { 767,12}, { 415,11}, \
106 { 831,12}, { 479,14}, { 127,13}, { 255,12}, \
107 { 543,11}, { 1087,12}, { 607,13}, { 319,12}, \
108 { 671,11}, { 1343,12}, { 703,11}, { 1407,12}, \
109 { 735,13}, { 383,12}, { 767,11}, { 1535,12}, \
110 { 831,13}, { 447,12}, { 959,14}, { 255,13}, \
111 { 511,12}, { 1023,11}, { 2047,12}, { 1087,13}, \
112 { 575,12}, { 1215,11}, { 2431,10}, { 4863,13}, \
113 { 639,12}, { 1279,11}, { 2559,12}, { 1343,13}, \
114 { 703,12}, { 1407,14}, { 383,13}, { 767,12}, \
115 { 1535,13}, { 831,12}, { 1727,13}, { 959,15}, \
116 { 255,14}, { 511,13}, { 1023,12}, { 2047,13}, \
117 { 1087,12}, { 2175,13}, { 1151,12}, { 2303,13}, \
118 { 1215,12}, { 2431,11}, { 4863,14}, { 639,13}, \
119 { 1279,12}, { 2559,13}, { 1343,12}, { 2687,13}, \
120 { 1407,12}, { 2815,13}, { 1471,12}, { 2943,11}, \
121 { 5887,14}, { 767,13}, { 1535,12}, { 3071,13}, \
122 { 1727,14}, { 895,13}, { 1791,12}, { 3583,13}, \
123 { 1919,15}, { 511,14}, { 1023,13}, { 2175,14}, \
124 { 1151,13}, { 2303,12}, { 4607,13}, { 2431,12}, \
125 { 4863,14}, { 1279,13}, { 2687,14}, { 1407,13}, \
126 { 2943,12}, { 5887,15}, { 767,14}, { 1535,13}, \
127 { 3071,14}, { 1663,13}, { 3455,12}, { 6911,14}, \
128 { 1791,13}, { 3583,14}, { 1919,13}, { 3839,16}, \
129 { 511,15}, { 1023,14}, { 2431,13}, { 4863,15}, \
130 { 1279,14}, { 2943,13}, { 5887,12}, { 11775,15}, \
131 { 1535,14}, { 3455,13}, { 6911,15}, { 1791,14}, \
132 { 3839,13}, { 7679,12}, { 15359,14}, { 3967,16}, \
133 { 1023,15}, { 2047,14}, { 4351,15}, { 2303,14}, \
134 { 4863,15}, { 2815,14}, { 5887,13}, { 11775,16}, \
135 { 1535,15}, { 3071,14}, { 6143,15}, { 3327,14}, \
136 { 6911,15}, { 3839,14}, { 7679,13}, { 15359,17}, \
137 { 1023,16}, { 2047,15}, { 4351,14}, { 8703,15}, \
138 { 4863,16}, { 2559,15}, { 5887,14}, { 11775,16}, \
139 { 3071,15}, { 6911,16}, { 3583,15}, { 7679,14}, \
140 { 15359,15}, { 7935,17}, { 2047,16}, { 4095,15}, \
141 { 8703,16}, { 4607,15}, { 9983,14}, { 19967,16}, \
142 { 5119,15}, { 10239,16}, { 5631,15}, { 11775,17}, \
143 { 3071,16}, { 65536,17}, { 131072,18}, { 262144,19}, \
144 { 524288,20}, {1048576,21}, {2097152,22}, {4194304,23}, \
145 {8388608,24} }
146 #define MUL_FFT_TABLE3_SIZE 261
147 #define MUL_FFT_THRESHOLD 4544
148
149 #define SQR_FFT_MODF_THRESHOLD 380 /* k = 5 */
150 #define SQR_FFT_TABLE3 \
151 { { 380, 5}, { 19, 6}, { 10, 5}, { 21, 6}, \
152 { 11, 5}, { 23, 6}, { 12, 5}, { 25, 6}, \
153 { 25, 7}, { 13, 6}, { 27, 7}, { 15, 6}, \
154 { 31, 7}, { 25, 8}, { 13, 7}, { 27, 8}, \
155 { 15, 7}, { 31, 8}, { 21, 9}, { 11, 8}, \
156 { 27, 9}, { 15, 8}, { 33, 9}, { 19, 8}, \
157 { 41, 9}, { 23, 8}, { 47, 9}, { 27,10}, \
158 { 15, 9}, { 39,10}, { 23, 9}, { 51,11}, \
159 { 15,10}, { 31, 9}, { 67,10}, { 39, 9}, \
160 { 79,10}, { 55,11}, { 31,10}, { 79,11}, \
161 { 47,10}, { 95,12}, { 31,11}, { 63,10}, \
162 { 127, 9}, { 255,10}, { 135,11}, { 79,10}, \
163 { 159, 9}, { 319,11}, { 95,10}, { 191, 9}, \
164 { 383,12}, { 63,11}, { 127,10}, { 255, 9}, \
165 { 511,10}, { 271, 9}, { 543,10}, { 287, 9}, \
166 { 575,10}, { 303, 9}, { 607,10}, { 319, 9}, \
167 { 639,12}, { 95,11}, { 191,10}, { 383,11}, \
168 { 207,13}, { 63,12}, { 127,11}, { 255,10}, \
169 { 511,11}, { 271,10}, { 543,11}, { 287,10}, \
170 { 575,11}, { 303,10}, { 607,11}, { 319,10}, \
171 { 639,11}, { 351,10}, { 703,11}, { 367,12}, \
172 { 191,11}, { 415,12}, { 223,11}, { 479,13}, \
173 { 127,12}, { 255,11}, { 543,12}, { 287,11}, \
174 { 607,12}, { 319,11}, { 639,12}, { 351,11}, \
175 { 703,10}, { 1407,13}, { 191,12}, { 383,11}, \
176 { 767,12}, { 479,14}, { 127,13}, { 255,12}, \
177 { 607,13}, { 319,12}, { 703,11}, { 1407,12}, \
178 { 735,13}, { 383,12}, { 767,11}, { 1535,12}, \
179 { 799,13}, { 447,12}, { 895,11}, { 1791,14}, \
180 { 255,13}, { 511,12}, { 1023,13}, { 575,12}, \
181 { 1151,11}, { 2303,12}, { 1215,13}, { 639,12}, \
182 { 1279,13}, { 703,12}, { 1407,14}, { 383,13}, \
183 { 767,12}, { 1535,13}, { 831,12}, { 1663,13}, \
184 { 895,12}, { 1791,13}, { 959,15}, { 255,14}, \
185 { 511,13}, { 1023,12}, { 2047,13}, { 1087,12}, \
186 { 2175,13}, { 1151,12}, { 2303,13}, { 1215,12}, \
187 { 2431,14}, { 639,13}, { 1279,12}, { 2559,13}, \
188 { 1343,12}, { 2687,13}, { 1407,12}, { 2815,13}, \
189 { 1471,12}, { 2943,11}, { 5887,14}, { 767,13}, \
190 { 1535,12}, { 3071,13}, { 1599,12}, { 3199,13}, \
191 { 1663,14}, { 895,13}, { 1791,12}, { 3583,15}, \
192 { 511,14}, { 1023,13}, { 2175,14}, { 1151,13}, \
193 { 2303,12}, { 4607,13}, { 2431,12}, { 4863,14}, \
194 { 1279,13}, { 2687,14}, { 1407,13}, { 2943,12}, \
195 { 5887,15}, { 767,14}, { 1535,13}, { 3199,14}, \
196 { 1663,13}, { 3455,12}, { 6911,14}, { 1791,13}, \
197 { 3583,14}, { 1919,16}, { 511,15}, { 1023,14}, \
198 { 2303,13}, { 4607,14}, { 2431,13}, { 4863,15}, \
199 { 1279,14}, { 2943,13}, { 5887,12}, { 11775,15}, \
200 { 1535,14}, { 3455,13}, { 6911,15}, { 1791,14}, \
201 { 3583,13}, { 7167,14}, { 3839,13}, { 7679,12}, \
202 { 15359,16}, { 1023,15}, { 2047,14}, { 4223,15}, \
203 { 2303,14}, { 4863,15}, { 2559,14}, { 5119,15}, \
204 { 2815,14}, { 5887,13}, { 11775,16}, { 1535,15}, \
205 { 3071,14}, { 6143,15}, { 3327,14}, { 6911,15}, \
206 { 3583,14}, { 7167,15}, { 3839,14}, { 7679,13}, \
207 { 15359,17}, { 1023,16}, { 2047,15}, { 4095,14}, \
208 { 8191,15}, { 4863,16}, { 2559,15}, { 5887,14}, \
209 { 11775,16}, { 3071,15}, { 6911,16}, { 3583,15}, \
210 { 7679,14}, { 15359,15}, { 7935,14}, { 15871,17}, \
211 { 2047,16}, { 4095,15}, { 8447,16}, { 4607,15}, \
212 { 9983,14}, { 19967,16}, { 5119,15}, { 10239,16}, \
213 { 5631,15}, { 11775,17}, { 3071,16}, { 65536,17}, \
214 { 131072,18}, { 262144,19}, { 524288,20}, {1048576,21}, \
215 {2097152,22}, {4194304,23}, {8388608,24} }
216 #define SQR_FFT_TABLE3_SIZE 259
217 #define SQR_FFT_THRESHOLD 3520
218
219 #define MULLO_BASECASE_THRESHOLD 0 /* always */
220 #define MULLO_DC_THRESHOLD 62
221 #define MULLO_MUL_N_THRESHOLD 8907
222 #define SQRLO_BASECASE_THRESHOLD 10
223 #define SQRLO_DC_THRESHOLD 13
224 #define SQRLO_SQR_THRESHOLD 7035
225
226 #define DC_DIV_QR_THRESHOLD 51
227 #define DC_DIVAPPR_Q_THRESHOLD 183
228 #define DC_BDIV_QR_THRESHOLD 47
229 #define DC_BDIV_Q_THRESHOLD 88
230
231 #define INV_MULMOD_BNM1_THRESHOLD 46
232 #define INV_NEWTON_THRESHOLD 226
233 #define INV_APPR_THRESHOLD 204
234
235 #define BINV_NEWTON_THRESHOLD 264
236 #define REDC_1_TO_REDC_2_THRESHOLD 28
237 #define REDC_2_TO_REDC_N_THRESHOLD 54
238
239 #define MU_DIV_QR_THRESHOLD 1589
240 #define MU_DIVAPPR_Q_THRESHOLD 1620
241 #define MUPI_DIV_QR_THRESHOLD 83
242 #define MU_BDIV_QR_THRESHOLD 1334
243 #define MU_BDIV_Q_THRESHOLD 1470
244
245 #define POWM_SEC_TABLE 1,16,194,642
246
247 #define GET_STR_DC_THRESHOLD 10
248 #define GET_STR_PRECOMPUTE_THRESHOLD 17
249 #define SET_STR_DC_THRESHOLD 381
250 #define SET_STR_PRECOMPUTE_THRESHOLD 1042
251
252 #define FAC_DSC_THRESHOLD 218
253 #define FAC_ODD_THRESHOLD 25
254
255 #define MATRIX22_STRASSEN_THRESHOLD 21
256 #define HGCD2_DIV1_METHOD 1 /* 6.58% faster than 3 */
257 #define HGCD_THRESHOLD 136
258 #define HGCD_APPR_THRESHOLD 168
259 #define HGCD_REDUCE_THRESHOLD 3014
260 #define GCD_DC_THRESHOLD 416
261 #define GCDEXT_DC_THRESHOLD 393
262 #define JACOBI_BASE_METHOD 4 /* 1.17% faster than 3 */
263
264 /* Tuneup completed successfully, took 800192 seconds */