xref: /netbsd-src/external/lgpl3/gmp/dist/mpn/x86_64/k8/gmp-mparam.h (revision bdc22b2e01993381dcefeff2bc9b56ca75a4235c)
1 /* AMD K8 gmp-mparam.h -- Compiler/machine parameter header file.
2 
3 Copyright 1991, 1993, 1994, 2000-2012, 2014 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 // #undef mpn_sublsh_n
35 // #define mpn_sublsh_n(rp,up,vp,n,c)					\
36 //    (((rp) == (up)) ? mpn_submul_1 (rp, vp, n, CNST_LIMB(1) << (c))	\
37 //      : MPN(mpn_sublsh_n)(rp,up,vp,n,c))
38 
39 /* 2500 MHz K8 Brisbane */
40 /* FFT tuning limit = 25000000 */
41 /* Generated by tuneup.c, 2014-03-12, gcc 4.5 */
42 
43 #define MOD_1_NORM_THRESHOLD                 0  /* always */
44 #define MOD_1_UNNORM_THRESHOLD               0  /* always */
45 #define MOD_1N_TO_MOD_1_1_THRESHOLD          4
46 #define MOD_1U_TO_MOD_1_1_THRESHOLD          3
47 #define MOD_1_1_TO_MOD_1_2_THRESHOLD        13
48 #define MOD_1_2_TO_MOD_1_4_THRESHOLD        35
49 #define PREINV_MOD_1_TO_MOD_1_THRESHOLD      9
50 #define USE_PREINV_DIVREM_1                  1  /* native */
51 #define DIV_QR_1_NORM_THRESHOLD              1
52 #define DIV_QR_1_UNNORM_THRESHOLD        MP_SIZE_T_MAX  /* never */
53 #define DIV_QR_2_PI2_THRESHOLD           MP_SIZE_T_MAX  /* never */
54 #define DIVEXACT_1_THRESHOLD                 0  /* always (native) */
55 #define BMOD_1_TO_MOD_1_THRESHOLD           16
56 
57 #define MUL_TOOM22_THRESHOLD                28
58 #define MUL_TOOM33_THRESHOLD                81
59 #define MUL_TOOM44_THRESHOLD               242
60 #define MUL_TOOM6H_THRESHOLD               345
61 #define MUL_TOOM8H_THRESHOLD               482
62 
63 #define MUL_TOOM32_TO_TOOM43_THRESHOLD      97
64 #define MUL_TOOM32_TO_TOOM53_THRESHOLD     153
65 #define MUL_TOOM42_TO_TOOM53_THRESHOLD     161
66 #define MUL_TOOM42_TO_TOOM63_THRESHOLD     175
67 #define MUL_TOOM43_TO_TOOM54_THRESHOLD     166
68 
69 #define SQR_BASECASE_THRESHOLD               0  /* always (native) */
70 #define SQR_TOOM2_THRESHOLD                 34
71 #define SQR_TOOM3_THRESHOLD                129
72 #define SQR_TOOM4_THRESHOLD                527
73 #define SQR_TOOM6_THRESHOLD                562
74 #define SQR_TOOM8_THRESHOLD                  0  /* always */
75 
76 #define MULMID_TOOM42_THRESHOLD             36
77 
78 #define MULMOD_BNM1_THRESHOLD               18
79 #define SQRMOD_BNM1_THRESHOLD               22
80 
81 #define MUL_FFT_MODF_THRESHOLD             654  /* k = 5 */
82 #define MUL_FFT_TABLE3                                      \
83   { {    654, 5}, {     21, 6}, {     11, 5}, {     23, 6}, \
84     {     12, 5}, {     25, 6}, {     13, 5}, {     27, 6}, \
85     {     15, 5}, {     31, 6}, {     27, 7}, {     15, 6}, \
86     {     31, 7}, {     19, 6}, {     39, 7}, {     29, 8}, \
87     {     15, 7}, {     33, 8}, {     17, 7}, {     37, 8}, \
88     {     19, 7}, {     39, 8}, {     23, 7}, {     47, 8}, \
89     {     25, 7}, {     51, 8}, {     43, 9}, {     23, 8}, \
90     {     51, 9}, {     27, 8}, {     57, 9}, {     31, 8}, \
91     {     65, 9}, {     35, 8}, {     71, 9}, {     39, 8}, \
92     {     79, 9}, {     43,10}, {     23, 9}, {     59, 8}, \
93     {    119,10}, {     31, 8}, {    125, 9}, {     71,10}, \
94     {     39, 9}, {     87,10}, {     47, 9}, {     99,10}, \
95     {     55, 9}, {    123,11}, {     31,10}, {     63, 9}, \
96     {    131,10}, {     71, 9}, {    143,10}, {     79, 9}, \
97     {    159,10}, {     87,11}, {     47,10}, {    119,11}, \
98     {     63,10}, {    143,11}, {     79,10}, {    175,11}, \
99     {     95,10}, {    199,11}, {    111,10}, {    223,12}, \
100     {     63,11}, {    143,10}, {    287, 9}, {    575,10}, \
101     {    295,11}, {    159,10}, {    319,11}, {    175,12}, \
102     {     95,11}, {    191,10}, {    383,11}, {    207,10}, \
103     {    415, 9}, {    831,11}, {    223,10}, {    447,13}, \
104     {     63,12}, {    127,11}, {    255,10}, {    511,11}, \
105     {    271,10}, {    543, 9}, {   1087,11}, {    287,10}, \
106     {    575, 9}, {   1151,12}, {    159,11}, {    319,10}, \
107     {    639,11}, {    335,10}, {    671,11}, {    351,10}, \
108     {    703,11}, {    367,12}, {    191,11}, {    383,10}, \
109     {    767,11}, {    399,10}, {    799,11}, {    415,10}, \
110     {    831,12}, {    223,11}, {    447,10}, {    895,11}, \
111     {    479,13}, {    127,12}, {    255,11}, {    511,10}, \
112     {   1023,11}, {    543,10}, {   1087,12}, {    287,11}, \
113     {    575,10}, {   1151,11}, {    607,12}, {    319,11}, \
114     {    639,10}, {   1279,11}, {    671,12}, {    351,11}, \
115     {    703,10}, {   1407,13}, {    191,12}, {    383,11}, \
116     {    767,12}, {    415,11}, {    831,12}, {    447,11}, \
117     {    895,12}, {    479,11}, {    959,14}, {    127,13}, \
118     {    255,12}, {    511,11}, {   1023,12}, {    543,11}, \
119     {   1087,12}, {    575,11}, {   1151,12}, {    607,11}, \
120     {   1215,13}, {    319,12}, {    671,11}, {   1343,12}, \
121     {    735,13}, {    383,12}, {    799,11}, {   1599,12}, \
122     {    863,13}, {    447,12}, {    895,11}, {   1791,12}, \
123     {    991,13}, {    511,12}, {   1087,13}, {    575,12}, \
124     {   1215,13}, {    639,12}, {   1343,13}, {    703,12}, \
125     {   1471,14}, {    383,13}, {    767,12}, {   1599,13}, \
126     {    831,12}, {   1663,13}, {    895,12}, {   1791,13}, \
127     {    959,12}, {   1919,14}, {    511,13}, {   1087,12}, \
128     {   2175,13}, {   1215,14}, {    639,13}, {   1471,14}, \
129     {    767,13}, {   1663,14}, {    895,13}, {   1919,15}, \
130     {    511,14}, {   1023,13}, {   2175,14}, {   1151,13}, \
131     {   2367,14}, {   1279,13}, {   2559,14}, {   1407,15}, \
132     {  32768,16}, {  65536,17}, { 131072,18}, { 262144,19}, \
133     { 524288,20}, {1048576,21}, {2097152,22}, {4194304,23}, \
134     {8388608,24} }
135 #define MUL_FFT_TABLE3_SIZE 205
136 #define MUL_FFT_THRESHOLD                11520
137 
138 #define SQR_FFT_MODF_THRESHOLD             570  /* k = 5 */
139 #define SQR_FFT_TABLE3                                      \
140   { {    570, 5}, {     21, 6}, {     11, 5}, {     23, 6}, \
141     {     12, 5}, {     25, 6}, {     13, 5}, {     27, 6}, \
142     {     14, 5}, {     29, 6}, {     16, 5}, {     33, 6}, \
143     {     29, 7}, {     15, 6}, {     31, 7}, {     16, 6}, \
144     {     33, 7}, {     17, 6}, {     35, 7}, {     33, 8}, \
145     {     17, 7}, {     37, 8}, {     19, 7}, {     40, 8}, \
146     {     21, 7}, {     43, 8}, {     23, 7}, {     47, 8}, \
147     {     25, 7}, {     51, 8}, {     29, 9}, {     15, 8}, \
148     {     37, 9}, {     19, 8}, {     43, 9}, {     23, 8}, \
149     {     51, 9}, {     27, 8}, {     55, 9}, {     31, 8}, \
150     {     63, 9}, {     35, 8}, {     71, 9}, {     39, 8}, \
151     {     79, 9}, {     43,10}, {     23, 9}, {     55,10}, \
152     {     31, 9}, {     71,10}, {     39, 9}, {     83,10}, \
153     {     47, 9}, {     99,10}, {     55, 9}, {    123,11}, \
154     {     31,10}, {     63, 9}, {    127,10}, {     71, 9}, \
155     {    143,10}, {     87,11}, {     47,10}, {    111,12}, \
156     {     31,11}, {     63,10}, {    143,11}, {     79,10}, \
157     {    167,11}, {     95,10}, {    199,11}, {    111,12}, \
158     {     63,11}, {    127, 9}, {    511,11}, {    143,10}, \
159     {    287, 9}, {    575, 8}, {   1151,11}, {    159,10}, \
160     {    319, 9}, {    639,11}, {    175,12}, {     95,11}, \
161     {    191,10}, {    383, 9}, {    767,11}, {    207,10}, \
162     {    415, 9}, {    831,10}, {    431,11}, {    223,10}, \
163     {    447,13}, {     63,12}, {    127,10}, {    511, 9}, \
164     {   1023,10}, {    543, 9}, {   1087,11}, {    287,10}, \
165     {    575, 9}, {   1151,12}, {    159,11}, {    319,10}, \
166     {    639, 9}, {   1279,11}, {    335,10}, {    671,11}, \
167     {    351,10}, {    703,11}, {    367,12}, {    191,11}, \
168     {    383,10}, {    767,11}, {    399,10}, {    799,11}, \
169     {    415,10}, {    831,11}, {    431,12}, {    223,11}, \
170     {    447,10}, {    895,11}, {    463,13}, {    127,11}, \
171     {    511,10}, {   1023,11}, {    543,10}, {   1087,12}, \
172     {    287,11}, {    575,10}, {   1151,11}, {    607,10}, \
173     {   1215,12}, {    319,11}, {    639,10}, {   1279,11}, \
174     {    671,12}, {    351,11}, {    703,10}, {   1407,13}, \
175     {    191,12}, {    383,11}, {    767,10}, {   1535,11}, \
176     {    799,12}, {    415,11}, {    831,12}, {    447,11}, \
177     {    895,12}, {    479,14}, {    127,12}, {    511,11}, \
178     {   1023,12}, {    543,11}, {   1087,12}, {    575,11}, \
179     {   1151,12}, {    607,11}, {   1215,13}, {    319,12}, \
180     {    639,11}, {   1279,12}, {    671,11}, {   1343,12}, \
181     {    703,11}, {   1407,12}, {    735,13}, {    383,12}, \
182     {    799,11}, {   1599,12}, {    863,13}, {    447,12}, \
183     {    959,13}, {    511,12}, {   1087,13}, {    575,12}, \
184     {   1215,13}, {    639,12}, {   1343,13}, {    703,12}, \
185     {   1471,14}, {    383,13}, {    767,12}, {   1599,13}, \
186     {    831,12}, {   1663,13}, {    895,12}, {   1791,13}, \
187     {    959,14}, {    511,13}, {   1023,12}, {   2047,13}, \
188     {   1215,14}, {    639,13}, {   1471,14}, {    767,13}, \
189     {   1663,14}, {    895,13}, {   1855,15}, {    511,14}, \
190     {   1023,13}, {   2111,14}, {   1151,13}, {   2303,14}, \
191     {   1407,15}, {  32768,16}, {  65536,17}, { 131072,18}, \
192     { 262144,19}, { 524288,20}, {1048576,21}, {2097152,22}, \
193     {4194304,23}, {8388608,24} }
194 #define SQR_FFT_TABLE3_SIZE 214
195 #define SQR_FFT_THRESHOLD                 5760
196 
197 #define MULLO_BASECASE_THRESHOLD             0  /* always */
198 #define MULLO_DC_THRESHOLD                  61
199 #define MULLO_MUL_N_THRESHOLD            22906
200 
201 #define DC_DIV_QR_THRESHOLD                 51
202 #define DC_DIVAPPR_Q_THRESHOLD             264
203 #define DC_BDIV_QR_THRESHOLD                38
204 #define DC_BDIV_Q_THRESHOLD                170
205 
206 #define INV_MULMOD_BNM1_THRESHOLD           67
207 #define INV_NEWTON_THRESHOLD               246
208 #define INV_APPR_THRESHOLD                 244
209 
210 #define BINV_NEWTON_THRESHOLD              252
211 #define REDC_1_TO_REDC_2_THRESHOLD          35
212 #define REDC_2_TO_REDC_N_THRESHOLD          84
213 
214 #define MU_DIV_QR_THRESHOLD               2089
215 #define MU_DIVAPPR_Q_THRESHOLD            1752
216 #define MUPI_DIV_QR_THRESHOLD               93
217 #define MU_BDIV_QR_THRESHOLD              1718
218 #define MU_BDIV_Q_THRESHOLD               1895
219 
220 #define POWM_SEC_TABLE  2,16,194,904,2177
221 
222 #define MATRIX22_STRASSEN_THRESHOLD         21
223 #define HGCD_THRESHOLD                     148
224 #define HGCD_APPR_THRESHOLD                185
225 #define HGCD_REDUCE_THRESHOLD             4120
226 #define GCD_DC_THRESHOLD                   562
227 #define GCDEXT_DC_THRESHOLD                501
228 #define JACOBI_BASE_METHOD                   4
229 
230 #define GET_STR_DC_THRESHOLD                17
231 #define GET_STR_PRECOMPUTE_THRESHOLD        29
232 #define SET_STR_DC_THRESHOLD               268
233 #define SET_STR_PRECOMPUTE_THRESHOLD      1787
234 
235 #define FAC_DSC_THRESHOLD                 1240
236 #define FAC_ODD_THRESHOLD                    0  /* always */
237