1 /* AMD Zen 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 /* 3700-4300 MHz Pinnacle Ridge */
39 /* FFT tuning limit = 468,514,360 */
40 /* Generated by tuneup.c, 2019-10-18, 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 13
47 #define MOD_1_2_TO_MOD_1_4_THRESHOLD 18
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 32
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 338
57
58 #define MUL_TOOM22_THRESHOLD 16
59 #define MUL_TOOM33_THRESHOLD 107
60 #define MUL_TOOM44_THRESHOLD 190
61 #define MUL_TOOM6H_THRESHOLD 230
62 #define MUL_TOOM8H_THRESHOLD 272
63
64 #define MUL_TOOM32_TO_TOOM43_THRESHOLD 97
65 #define MUL_TOOM32_TO_TOOM53_THRESHOLD 110
66 #define MUL_TOOM42_TO_TOOM53_THRESHOLD 106
67 #define MUL_TOOM42_TO_TOOM63_THRESHOLD 117
68 #define MUL_TOOM43_TO_TOOM54_THRESHOLD 136
69
70 #define SQR_BASECASE_THRESHOLD 0 /* always (native) */
71 #define SQR_TOOM2_THRESHOLD 32
72 #define SQR_TOOM3_THRESHOLD 114
73 #define SQR_TOOM4_THRESHOLD 422
74 #define SQR_TOOM6_THRESHOLD 0 /* always */
75 #define SQR_TOOM8_THRESHOLD 0 /* always */
76
77 #define MULMID_TOOM42_THRESHOLD 40
78
79 #define MULMOD_BNM1_THRESHOLD 12
80 #define SQRMOD_BNM1_THRESHOLD 17
81
82 #define MUL_FFT_MODF_THRESHOLD 540 /* k = 5 */
83 #define MUL_FFT_TABLE3 \
84 { { 540, 5}, { 22, 6}, { 12, 5}, { 25, 6}, \
85 { 25, 7}, { 13, 6}, { 29, 7}, { 15, 6}, \
86 { 31, 7}, { 21, 8}, { 11, 7}, { 25, 8}, \
87 { 13, 7}, { 29, 8}, { 15, 7}, { 32, 8}, \
88 { 17, 7}, { 35, 8}, { 19, 7}, { 39, 8}, \
89 { 21, 7}, { 43, 9}, { 11, 8}, { 29, 9}, \
90 { 15, 8}, { 35, 9}, { 19, 8}, { 43, 9}, \
91 { 23, 8}, { 49, 9}, { 27,10}, { 15, 9}, \
92 { 31, 8}, { 63, 9}, { 43,10}, { 23, 9}, \
93 { 55,11}, { 15,10}, { 31, 9}, { 67,10}, \
94 { 39, 9}, { 83,10}, { 47, 9}, { 99,10}, \
95 { 55,11}, { 31,10}, { 79,11}, { 47,10}, \
96 { 103,12}, { 31,11}, { 63,10}, { 135,11}, \
97 { 79,10}, { 167,11}, { 95,10}, { 191,12}, \
98 { 63,11}, { 159,12}, { 95,11}, { 191,13}, \
99 { 63,12}, { 127,11}, { 255,10}, { 511,11}, \
100 { 271,10}, { 543,11}, { 287,12}, { 159,11}, \
101 { 319,10}, { 639,11}, { 335,10}, { 671, 9}, \
102 { 1343,11}, { 351,12}, { 191,11}, { 383,10}, \
103 { 767,11}, { 415,10}, { 831,12}, { 223,11}, \
104 { 447,13}, { 127,12}, { 255,11}, { 543,10}, \
105 { 1087,12}, { 287,11}, { 575,10}, { 1151,11}, \
106 { 607,10}, { 1215,12}, { 319,11}, { 639,10}, \
107 { 1279,11}, { 671,10}, { 1343, 9}, { 2687,12}, \
108 { 351,11}, { 703,13}, { 191,12}, { 383,11}, \
109 { 767,12}, { 415,11}, { 831,10}, { 1663,12}, \
110 { 447,14}, { 127,13}, { 255,12}, { 511,11}, \
111 { 1023,12}, { 543,11}, { 1087,12}, { 575,11}, \
112 { 1151,12}, { 607,11}, { 1215,13}, { 319,12}, \
113 { 639,11}, { 1279,12}, { 671,11}, { 1343,10}, \
114 { 2687,12}, { 703,11}, { 1407,13}, { 383,12}, \
115 { 799,11}, { 1599,12}, { 831,11}, { 1663,13}, \
116 { 447,12}, { 895,11}, { 1791,12}, { 927,11}, \
117 { 1855,12}, { 959,11}, { 1919,10}, { 3839,13}, \
118 { 511,12}, { 1087,11}, { 2175,13}, { 575,12}, \
119 { 1215,11}, { 2431,13}, { 639,12}, { 1343,11}, \
120 { 2687,13}, { 703,12}, { 1407,14}, { 383,13}, \
121 { 767,12}, { 1599,13}, { 831,12}, { 1727,11}, \
122 { 3455,13}, { 895,12}, { 1855,13}, { 959,12}, \
123 { 1919,11}, { 3839,14}, { 511,13}, { 1087,12}, \
124 { 2175,13}, { 1215,12}, { 2431,14}, { 639,13}, \
125 { 1343,12}, { 2687,13}, { 1471,12}, { 2943,14}, \
126 { 767,13}, { 1599,12}, { 3199,13}, { 1727,12}, \
127 { 3455,14}, { 895,13}, { 1855,12}, { 3711,13}, \
128 { 1919,15}, { 511,14}, { 1023,13}, { 2175,14}, \
129 { 1151,13}, { 2431,12}, { 4863,14}, { 1279,13}, \
130 { 2687,14}, { 1407,13}, { 2815,15}, { 767,14}, \
131 { 1535,13}, { 3199,14}, { 1663,13}, { 3455,12}, \
132 { 6911,14}, { 1791,13}, { 3583,14}, { 1919,16}, \
133 { 511,15}, { 1023,14}, { 2175,13}, { 4479,14}, \
134 { 2431,13}, { 4863,15}, { 1279,14}, { 2687,13}, \
135 { 5375,14}, { 2943,13}, { 5887,15}, { 1535,14}, \
136 { 3455,13}, { 6911,15}, { 1791,14}, { 3839,13}, \
137 { 7679,16}, { 1023,15}, { 2047,14}, { 4479,15}, \
138 { 2303,14}, { 4991,15}, { 2559,14}, { 5247,15}, \
139 { 2815,14}, { 5887,16}, { 1535,15}, { 3327,14}, \
140 { 6911,15}, { 3839,14}, { 7679,17}, { 1023,16}, \
141 { 2047,15}, { 4095,14}, { 8191,15}, { 4351,14}, \
142 { 8959,15}, { 4863,16}, { 2559,15}, { 5375,14}, \
143 { 11007,15}, { 5887,14}, { 11775,16}, { 3071,15}, \
144 { 6911,16}, { 3583,15}, { 7167,14}, { 14335,15}, \
145 { 7679,14}, { 15359,15}, { 7935,14}, { 15871,17}, \
146 { 2047,16}, { 4095,15}, { 8959,16}, { 4607,15}, \
147 { 9215,14}, { 18431,15}, { 9727,14}, { 19455,15}, \
148 { 9983,14}, { 19967,16}, { 5119,15}, { 11007,16}, \
149 { 5631,15}, { 11775,17}, { 3071,16}, { 65536,17}, \
150 { 131072,18}, { 262144,19}, { 524288,20}, {1048576,21}, \
151 {2097152,22}, {4194304,23}, {8388608,24} }
152 #define MUL_FFT_TABLE3_SIZE 271
153 #define MUL_FFT_THRESHOLD 6272
154
155 #define SQR_FFT_MODF_THRESHOLD 404 /* k = 5 */
156 #define SQR_FFT_TABLE3 \
157 { { 404, 5}, { 13, 4}, { 27, 5}, { 21, 6}, \
158 { 11, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \
159 { 14, 5}, { 29, 6}, { 29, 7}, { 15, 6}, \
160 { 31, 7}, { 17, 6}, { 35, 7}, { 25, 8}, \
161 { 13, 7}, { 29, 8}, { 15, 7}, { 33, 8}, \
162 { 17, 7}, { 35, 8}, { 19, 7}, { 39, 8}, \
163 { 29, 9}, { 15, 8}, { 35, 9}, { 19, 8}, \
164 { 41, 9}, { 23, 8}, { 49, 9}, { 27,10}, \
165 { 15, 9}, { 31, 8}, { 63, 9}, { 43,10}, \
166 { 23, 9}, { 51,11}, { 15,10}, { 31, 9}, \
167 { 67,10}, { 39, 9}, { 79,10}, { 47, 9}, \
168 { 95,10}, { 55,11}, { 31,10}, { 79,11}, \
169 { 47,10}, { 95,12}, { 31,11}, { 63,10}, \
170 { 135,11}, { 79,10}, { 159,11}, { 95,12}, \
171 { 63,11}, { 127,10}, { 255, 9}, { 511,11}, \
172 { 143,10}, { 287, 9}, { 575,11}, { 159,12}, \
173 { 95,11}, { 191,13}, { 63,12}, { 127,11}, \
174 { 255,10}, { 511,11}, { 271,10}, { 543,11}, \
175 { 287,10}, { 575,11}, { 303,12}, { 159,11}, \
176 { 319,10}, { 639,11}, { 335,10}, { 671, 9}, \
177 { 1343,11}, { 351,10}, { 703,11}, { 367,10}, \
178 { 735,12}, { 191,11}, { 383,10}, { 767,11}, \
179 { 399,10}, { 799,11}, { 415,10}, { 831,12}, \
180 { 223,11}, { 447,10}, { 895,13}, { 127,12}, \
181 { 255,11}, { 511,10}, { 1023,11}, { 543,10}, \
182 { 1087,12}, { 287,11}, { 575,10}, { 1151,11}, \
183 { 607,10}, { 1215,12}, { 319,11}, { 639,10}, \
184 { 1279,11}, { 671,10}, { 1343,12}, { 351,11}, \
185 { 703,10}, { 1407,11}, { 735,10}, { 1471,13}, \
186 { 191,12}, { 383,11}, { 767,10}, { 1535,11}, \
187 { 799,12}, { 415,11}, { 831,10}, { 1663,12}, \
188 { 447,11}, { 895,14}, { 127,13}, { 255,12}, \
189 { 511,11}, { 1023,12}, { 543,11}, { 1087,12}, \
190 { 575,11}, { 1151,12}, { 607,11}, { 1215,13}, \
191 { 319,12}, { 639,11}, { 1279,12}, { 671,11}, \
192 { 1343,12}, { 703,11}, { 1407,12}, { 735,11}, \
193 { 1471,13}, { 383,12}, { 767,11}, { 1535,12}, \
194 { 799,11}, { 1599,12}, { 831,11}, { 1663,13}, \
195 { 447,12}, { 895,11}, { 1791,12}, { 959,14}, \
196 { 255,13}, { 511,12}, { 1023,11}, { 2047,12}, \
197 { 1087,11}, { 2175,13}, { 575,12}, { 1215,11}, \
198 { 2431,13}, { 639,12}, { 1343,13}, { 703,12}, \
199 { 1471,11}, { 2943,14}, { 383,13}, { 767,12}, \
200 { 1599,13}, { 831,12}, { 1727,11}, { 3455,13}, \
201 { 895,12}, { 1855,13}, { 959,15}, { 255,14}, \
202 { 511,13}, { 1023,12}, { 2047,13}, { 1087,12}, \
203 { 2175,13}, { 1215,12}, { 2431,14}, { 639,13}, \
204 { 1343,12}, { 2687,13}, { 1471,12}, { 2943,14}, \
205 { 767,13}, { 1599,12}, { 3199,13}, { 1727,12}, \
206 { 3455,14}, { 895,13}, { 1855,12}, { 3711,13}, \
207 { 1919,15}, { 511,14}, { 1023,13}, { 2175,14}, \
208 { 1151,13}, { 2431,12}, { 4863,14}, { 1279,13}, \
209 { 2687,14}, { 1407,13}, { 2943,15}, { 767,14}, \
210 { 1535,13}, { 3199,14}, { 1663,13}, { 3455,12}, \
211 { 6911,14}, { 1791,13}, { 3583,14}, { 1919,16}, \
212 { 511,15}, { 1023,14}, { 2047,13}, { 4095,14}, \
213 { 2175,13}, { 4479,12}, { 8959,14}, { 2431,13}, \
214 { 4863,15}, { 1279,14}, { 2943,13}, { 5887,12}, \
215 { 11775,15}, { 1535,14}, { 3455,13}, { 6911,15}, \
216 { 1791,14}, { 3839,13}, { 7679,14}, { 3967,16}, \
217 { 1023,15}, { 2047,14}, { 4479,15}, { 2303,14}, \
218 { 4991,15}, { 2559,14}, { 5247,15}, { 2815,14}, \
219 { 5887,13}, { 11775,16}, { 1535,15}, { 3071,14}, \
220 { 6143,15}, { 3327,14}, { 6911,15}, { 3839,14}, \
221 { 7679,17}, { 1023,16}, { 2047,15}, { 4095,14}, \
222 { 8191,15}, { 4351,14}, { 8959,15}, { 4863,14}, \
223 { 9727,16}, { 2559,15}, { 5887,14}, { 11775,16}, \
224 { 3071,15}, { 6911,16}, { 3583,15}, { 7167,14}, \
225 { 14335,15}, { 7679,14}, { 15359,15}, { 7935,14}, \
226 { 15871,17}, { 2047,16}, { 4095,15}, { 8959,16}, \
227 { 4607,15}, { 9215,14}, { 18431,15}, { 9727,14}, \
228 { 19455,15}, { 9983,14}, { 19967,16}, { 5119,15}, \
229 { 10239,16}, { 5631,15}, { 11775,17}, { 3071,16}, \
230 { 6655,15}, { 13311,16}, { 65536,17}, { 131072,18}, \
231 { 262144,19}, { 524288,20}, {1048576,21}, {2097152,22}, \
232 {4194304,23}, {8388608,24} }
233 #define SQR_FFT_TABLE3_SIZE 302
234 #define SQR_FFT_THRESHOLD 4224
235
236 #define MULLO_BASECASE_THRESHOLD 0 /* always */
237 #define MULLO_DC_THRESHOLD 69
238 #define MULLO_MUL_N_THRESHOLD 11278
239 #define SQRLO_BASECASE_THRESHOLD 12
240 #define SQRLO_DC_THRESHOLD 82
241 #define SQRLO_SQR_THRESHOLD 8207
242
243 #define DC_DIV_QR_THRESHOLD 76
244 #define DC_DIVAPPR_Q_THRESHOLD 232
245 #define DC_BDIV_QR_THRESHOLD 76
246 #define DC_BDIV_Q_THRESHOLD 104
247
248 #define INV_MULMOD_BNM1_THRESHOLD 37
249 #define INV_NEWTON_THRESHOLD 274
250 #define INV_APPR_THRESHOLD 230
251
252 #define BINV_NEWTON_THRESHOLD 372
253 #define REDC_1_TO_REDC_N_THRESHOLD 68
254
255 #define MU_DIV_QR_THRESHOLD 1499
256 #define MU_DIVAPPR_Q_THRESHOLD 1718
257 #define MUPI_DIV_QR_THRESHOLD 108
258 #define MU_BDIV_QR_THRESHOLD 1470
259 #define MU_BDIV_Q_THRESHOLD 1787
260
261 #define POWM_SEC_TABLE 3,22,81,494
262
263 #define GET_STR_DC_THRESHOLD 12
264 #define GET_STR_PRECOMPUTE_THRESHOLD 20
265 #define SET_STR_DC_THRESHOLD 486
266 #define SET_STR_PRECOMPUTE_THRESHOLD 1264
267
268 #define FAC_DSC_THRESHOLD 187
269 #define FAC_ODD_THRESHOLD 0 /* always */
270
271 #define MATRIX22_STRASSEN_THRESHOLD 23
272 #define HGCD2_DIV1_METHOD 1 /* 9.20% faster than 3 */
273 #define HGCD_THRESHOLD 109
274 #define HGCD_APPR_THRESHOLD 104
275 #define HGCD_REDUCE_THRESHOLD 3014
276 #define GCD_DC_THRESHOLD 566
277 #define GCDEXT_DC_THRESHOLD 382
278 #define JACOBI_BASE_METHOD 1 /* 15.55% faster than 3 */
279
280 /* Tuneup completed successfully, took 281243 seconds */