1d25e02daSmrg /* Nehalem gmp-mparam.h -- Compiler/machine parameter header file. 2d25e02daSmrg 3*671ea119Smrg Copyright 2019 Free Software Foundation, Inc. 4d25e02daSmrg 5d25e02daSmrg This file is part of the GNU MP Library. 6d25e02daSmrg 7d25e02daSmrg The GNU MP Library is free software; you can redistribute it and/or modify 8f81b1c5bSmrg it under the terms of either: 9f81b1c5bSmrg 10f81b1c5bSmrg * the GNU Lesser General Public License as published by the Free 11f81b1c5bSmrg Software Foundation; either version 3 of the License, or (at your 12d25e02daSmrg option) any later version. 13d25e02daSmrg 14f81b1c5bSmrg or 15f81b1c5bSmrg 16f81b1c5bSmrg * the GNU General Public License as published by the Free Software 17f81b1c5bSmrg Foundation; either version 2 of the License, or (at your option) any 18f81b1c5bSmrg later version. 19f81b1c5bSmrg 20f81b1c5bSmrg or both in parallel, as here. 21f81b1c5bSmrg 22d25e02daSmrg The GNU MP Library is distributed in the hope that it will be useful, but 23d25e02daSmrg WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY 24f81b1c5bSmrg or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 25f81b1c5bSmrg for more details. 26d25e02daSmrg 27f81b1c5bSmrg You should have received copies of the GNU General Public License and the 28f81b1c5bSmrg GNU Lesser General Public License along with the GNU MP Library. If not, 29f81b1c5bSmrg see https://www.gnu.org/licenses/. */ 30d25e02daSmrg 31d25e02daSmrg #define GMP_LIMB_BITS 64 32f81b1c5bSmrg #define GMP_LIMB_BYTES 8 33d25e02daSmrg 34*671ea119Smrg /* 2933-3200 MHz Intel Xeon X3470 Nehalem */ 35*671ea119Smrg /* FFT tuning limit = 468,424,931 */ 36*671ea119Smrg /* Generated by tuneup.c, 2019-10-18, gcc 8.3 */ 37d25e02daSmrg 38d25e02daSmrg #define MOD_1_NORM_THRESHOLD 0 /* always */ 39d25e02daSmrg #define MOD_1_UNNORM_THRESHOLD 0 /* always */ 40d25e02daSmrg #define MOD_1N_TO_MOD_1_1_THRESHOLD 3 41f81b1c5bSmrg #define MOD_1U_TO_MOD_1_1_THRESHOLD 2 42d25e02daSmrg #define MOD_1_1_TO_MOD_1_2_THRESHOLD 11 43d25e02daSmrg #define MOD_1_2_TO_MOD_1_4_THRESHOLD 16 44*671ea119Smrg #define PREINV_MOD_1_TO_MOD_1_THRESHOLD 7 45d25e02daSmrg #define USE_PREINV_DIVREM_1 1 /* native */ 46f81b1c5bSmrg #define DIV_QR_1_NORM_THRESHOLD 1 47f81b1c5bSmrg #define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ 48*671ea119Smrg #define DIV_QR_2_PI2_THRESHOLD 10 49d25e02daSmrg #define DIVEXACT_1_THRESHOLD 0 /* always (native) */ 50*671ea119Smrg #define BMOD_1_TO_MOD_1_THRESHOLD 17 51*671ea119Smrg 52*671ea119Smrg #define DIV_1_VS_MUL_1_PERCENT 301 53d25e02daSmrg 54d25e02daSmrg #define MUL_TOOM22_THRESHOLD 18 55*671ea119Smrg #define MUL_TOOM33_THRESHOLD 59 56d25e02daSmrg #define MUL_TOOM44_THRESHOLD 169 57*671ea119Smrg #define MUL_TOOM6H_THRESHOLD 230 58*671ea119Smrg #define MUL_TOOM8H_THRESHOLD 333 59d25e02daSmrg 60*671ea119Smrg #define MUL_TOOM32_TO_TOOM43_THRESHOLD 97 61*671ea119Smrg #define MUL_TOOM32_TO_TOOM53_THRESHOLD 110 62*671ea119Smrg #define MUL_TOOM42_TO_TOOM53_THRESHOLD 104 63*671ea119Smrg #define MUL_TOOM42_TO_TOOM63_THRESHOLD 101 64*671ea119Smrg #define MUL_TOOM43_TO_TOOM54_THRESHOLD 147 65d25e02daSmrg 66d25e02daSmrg #define SQR_BASECASE_THRESHOLD 0 /* always (native) */ 67f81b1c5bSmrg #define SQR_TOOM2_THRESHOLD 28 68*671ea119Smrg #define SQR_TOOM3_THRESHOLD 98 69d25e02daSmrg #define SQR_TOOM4_THRESHOLD 250 70*671ea119Smrg #define SQR_TOOM6_THRESHOLD 351 71*671ea119Smrg #define SQR_TOOM8_THRESHOLD 478 72d25e02daSmrg 73f81b1c5bSmrg #define MULMID_TOOM42_THRESHOLD 28 74d25e02daSmrg 75*671ea119Smrg #define MULMOD_BNM1_THRESHOLD 13 76d25e02daSmrg #define SQRMOD_BNM1_THRESHOLD 13 77d25e02daSmrg 78*671ea119Smrg #define MUL_FFT_MODF_THRESHOLD 372 /* k = 5 */ 79d25e02daSmrg #define MUL_FFT_TABLE3 \ 80*671ea119Smrg { { 372, 5}, { 17, 6}, { 9, 5}, { 19, 6}, \ 81*671ea119Smrg { 10, 5}, { 21, 6}, { 21, 7}, { 11, 6}, \ 82*671ea119Smrg { 23, 7}, { 21, 8}, { 11, 7}, { 24, 8}, \ 83*671ea119Smrg { 13, 7}, { 27, 8}, { 15, 7}, { 31, 8}, \ 84*671ea119Smrg { 21, 9}, { 11, 8}, { 27, 9}, { 15, 8}, \ 85*671ea119Smrg { 33, 9}, { 19, 8}, { 39, 9}, { 23, 8}, \ 86*671ea119Smrg { 47, 9}, { 27,10}, { 15, 9}, { 39,10}, \ 87*671ea119Smrg { 23, 9}, { 51,11}, { 15,10}, { 31, 9}, \ 88*671ea119Smrg { 67,10}, { 39, 9}, { 83,10}, { 47, 9}, \ 89*671ea119Smrg { 95,10}, { 55,11}, { 31,10}, { 79,11}, \ 90*671ea119Smrg { 47,10}, { 95,12}, { 31, 8}, { 511,10}, \ 91f81b1c5bSmrg { 135,11}, { 79,10}, { 159, 9}, { 319,11}, \ 92*671ea119Smrg { 95,10}, { 191, 9}, { 383,12}, { 63,11}, \ 93*671ea119Smrg { 127,10}, { 255, 9}, { 511,11}, { 143,10}, \ 94*671ea119Smrg { 287, 9}, { 575,10}, { 303,11}, { 159,10}, \ 95*671ea119Smrg { 319,12}, { 95,11}, { 191,10}, { 383,13}, \ 96*671ea119Smrg { 63,12}, { 127,11}, { 255,10}, { 511,11}, \ 97*671ea119Smrg { 271,10}, { 543,11}, { 287,10}, { 575,11}, \ 98*671ea119Smrg { 303,10}, { 607,11}, { 319,10}, { 639,11}, \ 99*671ea119Smrg { 351,10}, { 703,12}, { 191,11}, { 383,10}, \ 100*671ea119Smrg { 767,11}, { 415,10}, { 831,12}, { 223,11}, \ 101*671ea119Smrg { 447,10}, { 895,13}, { 127,12}, { 255,11}, \ 102*671ea119Smrg { 543,12}, { 287,11}, { 607,12}, { 319,11}, \ 103*671ea119Smrg { 639,12}, { 351,11}, { 703,13}, { 191,12}, \ 104*671ea119Smrg { 383,11}, { 767,12}, { 415,11}, { 831,12}, \ 105*671ea119Smrg { 447,11}, { 895,12}, { 479,14}, { 127,13}, \ 106*671ea119Smrg { 255,12}, { 543,11}, { 1087,12}, { 607,13}, \ 107*671ea119Smrg { 319,12}, { 703,13}, { 383,12}, { 831,13}, \ 108*671ea119Smrg { 447,12}, { 959,14}, { 255,13}, { 511,12}, \ 109*671ea119Smrg { 1087,13}, { 575,12}, { 1215,11}, { 2431,13}, \ 110*671ea119Smrg { 639,12}, { 1279,13}, { 703,12}, { 1407,14}, \ 111*671ea119Smrg { 383,13}, { 831,12}, { 1663,13}, { 959,14}, \ 112*671ea119Smrg { 511,13}, { 1087,12}, { 2175,13}, { 1215,12}, \ 113*671ea119Smrg { 2431,14}, { 639,13}, { 1343,12}, { 2687,13}, \ 114*671ea119Smrg { 1407,12}, { 2815,13}, { 1471,14}, { 767,13}, \ 115*671ea119Smrg { 1663,14}, { 895,13}, { 1791,15}, { 511,14}, \ 116*671ea119Smrg { 1023,13}, { 2175,14}, { 1151,13}, { 2431,12}, \ 117*671ea119Smrg { 4863,14}, { 1279,13}, { 2687,14}, { 1407,13}, \ 118*671ea119Smrg { 2815,15}, { 767,14}, { 1663,13}, { 3455,14}, \ 119*671ea119Smrg { 1919,16}, { 511,15}, { 1023,14}, { 2431,13}, \ 120f81b1c5bSmrg { 4863,15}, { 1279,14}, { 2943,13}, { 5887,15}, \ 121f81b1c5bSmrg { 1535,14}, { 3455,15}, { 1791,14}, { 3839,16}, \ 122f81b1c5bSmrg { 1023,15}, { 2047,14}, { 4223,15}, { 2303,14}, \ 123f81b1c5bSmrg { 4863,15}, { 2815,14}, { 5887,16}, { 1535,15}, \ 124f81b1c5bSmrg { 3327,14}, { 6911,15}, { 3839,17}, { 1023,16}, \ 125f81b1c5bSmrg { 2047,15}, { 4863,16}, { 2559,15}, { 5887,14}, \ 126*671ea119Smrg { 11775,16}, { 3071,15}, { 6911,16}, { 3583,15}, \ 127*671ea119Smrg { 7679,14}, { 15359,17}, { 2047,16}, { 4607,15}, \ 128*671ea119Smrg { 9983,16}, { 5631,15}, { 11775,17}, { 3071,16}, \ 129*671ea119Smrg { 65536,17}, { 131072,18}, { 262144,19}, { 524288,20}, \ 130*671ea119Smrg {1048576,21}, {2097152,22}, {4194304,23}, {8388608,24} } 131*671ea119Smrg #define MUL_FFT_TABLE3_SIZE 204 132f81b1c5bSmrg #define MUL_FFT_THRESHOLD 4224 133d25e02daSmrg 134*671ea119Smrg #define SQR_FFT_MODF_THRESHOLD 336 /* k = 5 */ 135d25e02daSmrg #define SQR_FFT_TABLE3 \ 136*671ea119Smrg { { 336, 5}, { 19, 6}, { 10, 5}, { 21, 6}, \ 137d25e02daSmrg { 21, 7}, { 11, 6}, { 23, 7}, { 21, 8}, \ 138*671ea119Smrg { 11, 7}, { 25, 8}, { 13, 7}, { 27, 8}, \ 139d25e02daSmrg { 15, 7}, { 31, 8}, { 21, 9}, { 11, 8}, \ 140d25e02daSmrg { 27, 9}, { 15, 8}, { 33, 9}, { 19, 8}, \ 141*671ea119Smrg { 39, 9}, { 23, 8}, { 47, 9}, { 27,10}, \ 142d25e02daSmrg { 15, 9}, { 39,10}, { 23, 9}, { 47,11}, \ 143*671ea119Smrg { 15,10}, { 31, 9}, { 67,10}, { 39, 9}, \ 144*671ea119Smrg { 79,10}, { 47,11}, { 31,10}, { 79,11}, \ 145*671ea119Smrg { 47,10}, { 95,12}, { 31,11}, { 63,10}, \ 146*671ea119Smrg { 127, 9}, { 255, 8}, { 511,10}, { 135,11}, \ 147*671ea119Smrg { 79, 9}, { 319, 6}, { 2687, 7}, { 1407, 9}, \ 148*671ea119Smrg { 383,12}, { 63,11}, { 127,10}, { 255, 9}, \ 149*671ea119Smrg { 511,10}, { 271, 9}, { 543,11}, { 143,10}, \ 150*671ea119Smrg { 287, 9}, { 575,10}, { 303, 9}, { 607,10}, \ 151*671ea119Smrg { 319,12}, { 95,11}, { 191,10}, { 383,13}, \ 152*671ea119Smrg { 63,12}, { 127,11}, { 255,10}, { 511,11}, \ 153*671ea119Smrg { 271,10}, { 543,11}, { 287,10}, { 575,11}, \ 154*671ea119Smrg { 303,10}, { 607,11}, { 319,10}, { 639,11}, \ 155*671ea119Smrg { 351,10}, { 703,12}, { 191,11}, { 383,10}, \ 156*671ea119Smrg { 767,11}, { 415,10}, { 831,12}, { 223,11}, \ 157*671ea119Smrg { 447,10}, { 895,11}, { 479,13}, { 127,12}, \ 158*671ea119Smrg { 255,11}, { 511,10}, { 1023,11}, { 543,12}, \ 159*671ea119Smrg { 287,11}, { 607,12}, { 319,11}, { 671,12}, \ 160*671ea119Smrg { 351,11}, { 703,13}, { 191,12}, { 383,11}, \ 161*671ea119Smrg { 767,12}, { 415,11}, { 831,12}, { 447,11}, \ 162*671ea119Smrg { 895,12}, { 479,14}, { 127,13}, { 255,12}, \ 163*671ea119Smrg { 511,11}, { 1023,12}, { 543,11}, { 1087,12}, \ 164*671ea119Smrg { 575,11}, { 1151,12}, { 607,13}, { 319,12}, \ 165*671ea119Smrg { 671,11}, { 1343,12}, { 703,13}, { 383,12}, \ 166*671ea119Smrg { 767,11}, { 1535,12}, { 831,13}, { 447,12}, \ 167*671ea119Smrg { 959,13}, { 511,12}, { 1087,13}, { 575,12}, \ 168*671ea119Smrg { 1215,11}, { 2431,13}, { 639,12}, { 1343,13}, \ 169*671ea119Smrg { 703,14}, { 383,13}, { 767,12}, { 1535,13}, \ 170*671ea119Smrg { 831,12}, { 1663,13}, { 959,14}, { 511,13}, \ 171*671ea119Smrg { 1087,12}, { 2175,13}, { 1215,12}, { 2431,14}, \ 172f81b1c5bSmrg { 639,13}, { 1343,12}, { 2687,13}, { 1407,12}, \ 173*671ea119Smrg { 2815,13}, { 1471,14}, { 767,13}, { 1663,14}, \ 174*671ea119Smrg { 895,13}, { 1791,15}, { 511,14}, { 1023,13}, \ 175*671ea119Smrg { 2175,14}, { 1151,13}, { 2431,12}, { 4863,14}, \ 176*671ea119Smrg { 1279,13}, { 2687,14}, { 1407,13}, { 2815,15}, \ 177*671ea119Smrg { 767,14}, { 1535,13}, { 3071,14}, { 1663,13}, \ 178*671ea119Smrg { 3455,14}, { 1919,16}, { 511,15}, { 1023,14}, \ 179*671ea119Smrg { 2431,13}, { 4863,15}, { 1279,14}, { 2943,13}, \ 180*671ea119Smrg { 5887,15}, { 1535,14}, { 3455,15}, { 1791,14}, \ 181*671ea119Smrg { 3839,16}, { 1023,15}, { 2047,14}, { 4223,15}, \ 182*671ea119Smrg { 2303,14}, { 4863,15}, { 2815,14}, { 5887,16}, \ 183*671ea119Smrg { 1535,15}, { 3327,14}, { 6911,15}, { 3839,17}, \ 184*671ea119Smrg { 1023,16}, { 2047,15}, { 4863,16}, { 2559,15}, \ 185*671ea119Smrg { 5887,14}, { 11775,16}, { 3071,15}, { 6655,16}, \ 186*671ea119Smrg { 3583,15}, { 7679,14}, { 15359,17}, { 2047,16}, \ 187*671ea119Smrg { 4607,15}, { 9983,14}, { 19967,16}, { 5631,15}, \ 188*671ea119Smrg { 11775,17}, { 3071,16}, { 65536,17}, { 131072,18}, \ 189*671ea119Smrg { 262144,19}, { 524288,20}, {1048576,21}, {2097152,22}, \ 190*671ea119Smrg {4194304,23}, {8388608,24} } 191*671ea119Smrg #define SQR_FFT_TABLE3_SIZE 218 192f81b1c5bSmrg #define SQR_FFT_THRESHOLD 3520 193d25e02daSmrg 194d25e02daSmrg #define MULLO_BASECASE_THRESHOLD 0 /* always */ 195*671ea119Smrg #define MULLO_DC_THRESHOLD 49 196f81b1c5bSmrg #define MULLO_MUL_N_THRESHOLD 8397 197*671ea119Smrg #define SQRLO_BASECASE_THRESHOLD 10 198*671ea119Smrg #define SQRLO_DC_THRESHOLD 11 199f81b1c5bSmrg #define SQRLO_SQR_THRESHOLD 7035 200d25e02daSmrg 201*671ea119Smrg #define DC_DIV_QR_THRESHOLD 47 202*671ea119Smrg #define DC_DIVAPPR_Q_THRESHOLD 151 203*671ea119Smrg #define DC_BDIV_QR_THRESHOLD 40 204f81b1c5bSmrg #define DC_BDIV_Q_THRESHOLD 30 205d25e02daSmrg 206f81b1c5bSmrg #define INV_MULMOD_BNM1_THRESHOLD 34 207*671ea119Smrg #define INV_NEWTON_THRESHOLD 199 208*671ea119Smrg #define INV_APPR_THRESHOLD 157 209d25e02daSmrg 210*671ea119Smrg #define BINV_NEWTON_THRESHOLD 254 211*671ea119Smrg #define REDC_1_TO_REDC_N_THRESHOLD 48 212d25e02daSmrg 213*671ea119Smrg #define MU_DIV_QR_THRESHOLD 1334 214f81b1c5bSmrg #define MU_DIVAPPR_Q_THRESHOLD 1334 215*671ea119Smrg #define MUPI_DIV_QR_THRESHOLD 83 216f81b1c5bSmrg #define MU_BDIV_QR_THRESHOLD 1142 217*671ea119Smrg #define MU_BDIV_Q_THRESHOLD 1308 218d25e02daSmrg 219*671ea119Smrg #define POWM_SEC_TABLE 1,64,66,452,1486 220d25e02daSmrg 221*671ea119Smrg #define GET_STR_DC_THRESHOLD 11 222*671ea119Smrg #define GET_STR_PRECOMPUTE_THRESHOLD 18 223*671ea119Smrg #define SET_STR_DC_THRESHOLD 141 224*671ea119Smrg #define SET_STR_PRECOMPUTE_THRESHOLD 1023 225d25e02daSmrg 226*671ea119Smrg #define FAC_DSC_THRESHOLD 182 227*671ea119Smrg #define FAC_ODD_THRESHOLD 0 /* always */ 228f81b1c5bSmrg 229*671ea119Smrg #define MATRIX22_STRASSEN_THRESHOLD 19 230*671ea119Smrg #define HGCD2_DIV1_METHOD 5 /* 2.91% faster than 3 */ 231*671ea119Smrg #define HGCD_THRESHOLD 116 232*671ea119Smrg #define HGCD_APPR_THRESHOLD 164 233f81b1c5bSmrg #define HGCD_REDUCE_THRESHOLD 2205 234*671ea119Smrg #define GCD_DC_THRESHOLD 321 235*671ea119Smrg #define GCDEXT_DC_THRESHOLD 358 236*671ea119Smrg #define JACOBI_BASE_METHOD 4 /* 0.12% faster than 1 */ 237*671ea119Smrg 238*671ea119Smrg /* Tuneup completed successfully, took 452116 seconds */ 239