1 /* AMD Zen2 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 /* 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 /* 3600-4400 MHz Matisse */
39 /* FFT tuning limit = 703,392,483 */
40 /* Generated by tuneup.c, 2019-10-19, gcc 8.3 */
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 8
47 #define MOD_1_2_TO_MOD_1_4_THRESHOLD 27
48 #define PREINV_MOD_1_TO_MOD_1_THRESHOLD 1
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 13
53 #define DIVEXACT_1_THRESHOLD 0 /* always (native) */
54 #define BMOD_1_TO_MOD_1_THRESHOLD 22
55
56 #define DIV_1_VS_MUL_1_PERCENT 385
57
58 #define MUL_TOOM22_THRESHOLD 19
59 #define MUL_TOOM33_THRESHOLD 125
60 #define MUL_TOOM44_THRESHOLD 196
61 #define MUL_TOOM6H_THRESHOLD 276
62 #define MUL_TOOM8H_THRESHOLD 369
63
64 #define MUL_TOOM32_TO_TOOM43_THRESHOLD 121
65 #define MUL_TOOM32_TO_TOOM53_THRESHOLD 138
66 #define MUL_TOOM42_TO_TOOM53_THRESHOLD 129
67 #define MUL_TOOM42_TO_TOOM63_THRESHOLD 132
68 #define MUL_TOOM43_TO_TOOM54_THRESHOLD 185
69
70 #define SQR_BASECASE_THRESHOLD 0 /* always (native) */
71 #define SQR_TOOM2_THRESHOLD 30
72 #define SQR_TOOM3_THRESHOLD 117
73 #define SQR_TOOM4_THRESHOLD 315
74 #define SQR_TOOM6_THRESHOLD 446
75 #define SQR_TOOM8_THRESHOLD 527
76
77 #define MULMID_TOOM42_THRESHOLD 38
78
79 #define MULMOD_BNM1_THRESHOLD 14
80 #define SQRMOD_BNM1_THRESHOLD 20
81
82 #define MUL_FFT_MODF_THRESHOLD 436 /* k = 5 */
83 #define MUL_FFT_TABLE3 \
84 { { 436, 5}, { 25, 6}, { 25, 7}, { 13, 6}, \
85 { 27, 7}, { 15, 6}, { 31, 7}, { 25, 8}, \
86 { 13, 7}, { 28, 8}, { 15, 7}, { 32, 8}, \
87 { 17, 7}, { 35, 8}, { 19, 7}, { 39, 8}, \
88 { 27, 9}, { 15, 8}, { 35, 9}, { 19, 8}, \
89 { 41, 9}, { 23, 8}, { 49, 9}, { 27,10}, \
90 { 15, 9}, { 31, 8}, { 63, 9}, { 39,10}, \
91 { 23, 9}, { 51,11}, { 15,10}, { 31, 9}, \
92 { 71,10}, { 39, 9}, { 83,10}, { 47, 9}, \
93 { 95,10}, { 55,11}, { 31,10}, { 79,11}, \
94 { 47,10}, { 95,12}, { 31,11}, { 63,10}, \
95 { 135,11}, { 79,10}, { 159,11}, { 95,10}, \
96 { 191,12}, { 63,11}, { 127,10}, { 255, 9}, \
97 { 511,11}, { 143,10}, { 287, 9}, { 575,11}, \
98 { 159,12}, { 95,11}, { 191,13}, { 63,12}, \
99 { 127,11}, { 255,10}, { 511,11}, { 271,10}, \
100 { 543, 9}, { 1087,11}, { 287,10}, { 575,12}, \
101 { 159,11}, { 319,10}, { 639,11}, { 335,10}, \
102 { 671,11}, { 351,10}, { 703,11}, { 367,12}, \
103 { 191,11}, { 383,10}, { 767,11}, { 415,10}, \
104 { 831,12}, { 223,11}, { 447,13}, { 127,12}, \
105 { 255,11}, { 543,10}, { 1087,12}, { 287,11}, \
106 { 575,10}, { 1151,11}, { 607,10}, { 1215,12}, \
107 { 319,11}, { 639,10}, { 1279,11}, { 671,10}, \
108 { 1343,12}, { 351,11}, { 703,10}, { 1407,11}, \
109 { 735,13}, { 191,12}, { 383,11}, { 767,10}, \
110 { 1535,11}, { 799,12}, { 415,11}, { 831,10}, \
111 { 1663,12}, { 447,11}, { 895,12}, { 479,14}, \
112 { 127,13}, { 255,12}, { 543,11}, { 1087,10}, \
113 { 2175,12}, { 575,11}, { 1151,12}, { 607,11}, \
114 { 1215,10}, { 2431,13}, { 319,12}, { 639,11}, \
115 { 1279,12}, { 671,11}, { 1343,10}, { 2687,12}, \
116 { 703,11}, { 1471,10}, { 2943,13}, { 383,12}, \
117 { 767,11}, { 1535,12}, { 799,11}, { 1599,12}, \
118 { 831,11}, { 1663,13}, { 447,12}, { 959,11}, \
119 { 1919,10}, { 3839,14}, { 255,13}, { 511,12}, \
120 { 1087,11}, { 2175,13}, { 575,12}, { 1215,11}, \
121 { 2431,13}, { 639,12}, { 1343,11}, { 2687,13}, \
122 { 703,12}, { 1471,11}, { 2943,14}, { 383,13}, \
123 { 767,12}, { 1599,11}, { 3199,13}, { 831,12}, \
124 { 1727,13}, { 895,12}, { 1791,13}, { 959,12}, \
125 { 1919,11}, { 3839,14}, { 511,13}, { 1087,12}, \
126 { 2175,13}, { 1215,12}, { 2431,14}, { 639,13}, \
127 { 1343,12}, { 2687,13}, { 1471,12}, { 2943,11}, \
128 { 5887,14}, { 767,13}, { 1599,12}, { 3199,13}, \
129 { 1727,12}, { 3455,14}, { 895,13}, { 1919,12}, \
130 { 3839,11}, { 7679,15}, { 511,14}, { 1023,13}, \
131 { 2175,14}, { 1151,13}, { 2431,12}, { 4863,14}, \
132 { 1279,13}, { 2687,14}, { 1407,13}, { 2943,12}, \
133 { 5887,15}, { 767,14}, { 1535,13}, { 3199,14}, \
134 { 1663,13}, { 3455,12}, { 6911,14}, { 1919,13}, \
135 { 3839,16}, { 511,15}, { 1023,14}, { 2175,13}, \
136 { 4479,12}, { 8959,14}, { 2431,13}, { 4863,15}, \
137 { 1279,14}, { 2943,13}, { 5887,12}, { 11775,15}, \
138 { 1535,14}, { 3455,15}, { 1791,14}, { 3839,13}, \
139 { 7679,14}, { 3967,16}, { 1023,15}, { 2047,14}, \
140 { 4479,15}, { 2303,14}, { 4863,15}, { 2559,14}, \
141 { 5247,15}, { 2815,14}, { 5887,16}, { 1535,15}, \
142 { 3327,14}, { 6911,15}, { 3839,14}, { 7679,13}, \
143 { 15359,17}, { 1023,16}, { 2047,15}, { 4351,14}, \
144 { 8959,15}, { 4863,16}, { 2559,15}, { 5887,14}, \
145 { 11775,16}, { 3071,15}, { 6911,16}, { 3583,15}, \
146 { 7679,14}, { 15359,15}, { 7935,17}, { 2047,16}, \
147 { 4095,15}, { 8959,16}, { 4607,15}, { 9983,14}, \
148 { 19967,16}, { 5631,15}, { 11775,17}, { 3071,16}, \
149 { 7679,15}, { 15871,18}, { 2047,17}, { 4095,16}, \
150 { 9727,15}, { 19967,17}, { 5119,16}, { 65536,17}, \
151 { 131072,18}, { 262144,19}, { 524288,20}, {1048576,21}, \
152 {2097152,22}, {4194304,23}, {8388608,24} }
153 #define MUL_FFT_TABLE3_SIZE 275
154 #define MUL_FFT_THRESHOLD 4736
155
156 #define SQR_FFT_MODF_THRESHOLD 396 /* k = 5 */
157 #define SQR_FFT_TABLE3 \
158 { { 396, 5}, { 25, 6}, { 25, 7}, { 13, 6}, \
159 { 27, 7}, { 15, 6}, { 31, 7}, { 25, 8}, \
160 { 13, 7}, { 28, 8}, { 15, 7}, { 32, 8}, \
161 { 17, 7}, { 35, 8}, { 19, 7}, { 39, 8}, \
162 { 27, 9}, { 15, 8}, { 35, 9}, { 19, 8}, \
163 { 41, 9}, { 23, 8}, { 47, 9}, { 27,10}, \
164 { 15, 9}, { 39,10}, { 23, 9}, { 51,11}, \
165 { 15,10}, { 31, 9}, { 67,10}, { 39, 9}, \
166 { 79,10}, { 55,11}, { 31,10}, { 79,11}, \
167 { 47,10}, { 95,12}, { 31,11}, { 63,10}, \
168 { 135,11}, { 79,10}, { 159,11}, { 95,12}, \
169 { 63,11}, { 127,10}, { 255, 9}, { 511,11}, \
170 { 143,10}, { 287, 9}, { 575,10}, { 303,11}, \
171 { 159,12}, { 95,13}, { 63,12}, { 127,11}, \
172 { 255,10}, { 511,11}, { 271,10}, { 543,11}, \
173 { 287,10}, { 575,11}, { 303,12}, { 159,11}, \
174 { 319,10}, { 639,11}, { 335,10}, { 671, 9}, \
175 { 1343,11}, { 351,10}, { 703,11}, { 367,10}, \
176 { 735,11}, { 383,10}, { 767,11}, { 415,10}, \
177 { 831,12}, { 223,11}, { 447,13}, { 127,12}, \
178 { 255,11}, { 511,10}, { 1023,11}, { 543,10}, \
179 { 1087,12}, { 287,11}, { 575,10}, { 1151,11}, \
180 { 607,10}, { 1215,12}, { 319,11}, { 639,10}, \
181 { 1279,11}, { 671,10}, { 1343,12}, { 351,11}, \
182 { 703,10}, { 1407,11}, { 735,10}, { 1471,12}, \
183 { 383,11}, { 767,10}, { 1535,11}, { 799,12}, \
184 { 415,11}, { 831,10}, { 1663,12}, { 447,11}, \
185 { 895,12}, { 479,11}, { 959,14}, { 127,12}, \
186 { 511,11}, { 1023,12}, { 543,11}, { 1087,10}, \
187 { 2175,12}, { 575,11}, { 1151,12}, { 607,11}, \
188 { 1215,10}, { 2431,12}, { 639,11}, { 1279,12}, \
189 { 671,11}, { 1343,10}, { 2687,12}, { 703,11}, \
190 { 1407,12}, { 735,11}, { 1471,10}, { 2943,13}, \
191 { 383,12}, { 767,11}, { 1535,12}, { 799,11}, \
192 { 1599,12}, { 831,11}, { 1663,13}, { 447,12}, \
193 { 959,11}, { 1919,10}, { 3839,13}, { 511,12}, \
194 { 1087,11}, { 2175,13}, { 575,12}, { 1215,11}, \
195 { 2431,13}, { 639,12}, { 1343,11}, { 2687,13}, \
196 { 703,12}, { 1471,11}, { 2943,14}, { 383,13}, \
197 { 767,12}, { 1599,13}, { 831,12}, { 1727,11}, \
198 { 3455,13}, { 959,12}, { 1919,11}, { 3839,14}, \
199 { 511,13}, { 1023,12}, { 2047,13}, { 1087,12}, \
200 { 2175,13}, { 1215,12}, { 2431,11}, { 4863,14}, \
201 { 639,13}, { 1343,12}, { 2687,13}, { 1471,12}, \
202 { 2943,11}, { 5887,14}, { 767,13}, { 1599,12}, \
203 { 3199,13}, { 1727,12}, { 3455,14}, { 895,13}, \
204 { 1919,12}, { 3839,15}, { 511,14}, { 1023,13}, \
205 { 2175,14}, { 1151,13}, { 2431,12}, { 4863,14}, \
206 { 1279,13}, { 2687,14}, { 1407,13}, { 2943,12}, \
207 { 5887,15}, { 767,14}, { 1535,13}, { 3199,14}, \
208 { 1663,13}, { 3455,12}, { 6911,14}, { 1919,13}, \
209 { 3839,12}, { 7679,16}, { 511,15}, { 1023,14}, \
210 { 2175,13}, { 4479,14}, { 2431,13}, { 4863,15}, \
211 { 1279,14}, { 2943,13}, { 5887,12}, { 11775,15}, \
212 { 1535,14}, { 3455,13}, { 6911,15}, { 1791,14}, \
213 { 3839,13}, { 7679,14}, { 3967,16}, { 1023,15}, \
214 { 2047,14}, { 4479,15}, { 2303,14}, { 4863,15}, \
215 { 2559,14}, { 5247,15}, { 2815,14}, { 5887,13}, \
216 { 11775,16}, { 1535,15}, { 3071,14}, { 6143,15}, \
217 { 3327,14}, { 6911,15}, { 3839,14}, { 7679,17}, \
218 { 1023,16}, { 2047,15}, { 4095,14}, { 8191,15}, \
219 { 4351,14}, { 8959,15}, { 4863,16}, { 2559,15}, \
220 { 5887,14}, { 11775,16}, { 3071,15}, { 6911,16}, \
221 { 3583,15}, { 7679,14}, { 15359,15}, { 7935,17}, \
222 { 2047,16}, { 4095,15}, { 8959,16}, { 4607,15}, \
223 { 9983,14}, { 19967,16}, { 5119,15}, { 10239,16}, \
224 { 5631,15}, { 11775,17}, { 3071,16}, { 7679,15}, \
225 { 15359,18}, { 2047,17}, { 4095,16}, { 9727,15}, \
226 { 19967,17}, { 5119,16}, { 65536,17}, { 131072,18}, \
227 { 262144,19}, { 524288,20}, {1048576,21}, {2097152,22}, \
228 {4194304,23}, {8388608,24} }
229 #define SQR_FFT_TABLE3_SIZE 282
230 #define SQR_FFT_THRESHOLD 3264
231
232 #define MULLO_BASECASE_THRESHOLD 0 /* always */
233 #define MULLO_DC_THRESHOLD 57
234 #define MULLO_MUL_N_THRESHOLD 8907
235 #define SQRLO_BASECASE_THRESHOLD 8
236 #define SQRLO_DC_THRESHOLD 0 /* never mpn_sqrlo_basecase */
237 #define SQRLO_SQR_THRESHOLD 6440
238
239 #define DC_DIV_QR_THRESHOLD 43
240 #define DC_DIVAPPR_Q_THRESHOLD 154
241 #define DC_BDIV_QR_THRESHOLD 46
242 #define DC_BDIV_Q_THRESHOLD 93
243
244 #define INV_MULMOD_BNM1_THRESHOLD 36
245 #define INV_NEWTON_THRESHOLD 141
246 #define INV_APPR_THRESHOLD 149
247
248 #define BINV_NEWTON_THRESHOLD 264
249 #define REDC_1_TO_REDC_N_THRESHOLD 47
250
251 #define MU_DIV_QR_THRESHOLD 1470
252 #define MU_DIVAPPR_Q_THRESHOLD 1528
253 #define MUPI_DIV_QR_THRESHOLD 47
254 #define MU_BDIV_QR_THRESHOLD 1187
255 #define MU_BDIV_Q_THRESHOLD 1589
256
257 #define POWM_SEC_TABLE 3,22,194,579
258
259 #define GET_STR_DC_THRESHOLD 12
260 #define GET_STR_PRECOMPUTE_THRESHOLD 19
261 #define SET_STR_DC_THRESHOLD 195
262 #define SET_STR_PRECOMPUTE_THRESHOLD 1752
263
264 #define FAC_DSC_THRESHOLD 345
265 #define FAC_ODD_THRESHOLD 0 /* always */
266
267 #define MATRIX22_STRASSEN_THRESHOLD 24
268 #define HGCD2_DIV1_METHOD 1 /* 11.29% faster than 3 */
269 #define HGCD_THRESHOLD 89
270 #define HGCD_APPR_THRESHOLD 96
271 #define HGCD_REDUCE_THRESHOLD 2681
272 #define GCD_DC_THRESHOLD 465
273 #define GCDEXT_DC_THRESHOLD 233
274 #define JACOBI_BASE_METHOD 1 /* 25.56% faster than 4 */
275
276 /* Tuneup completed successfully, took 294200 seconds */