1. repair Windows build script
2. Add tuning for nehalem on Windows
This commit is contained in:
gladman
parent
48edb06e9f
commit
0ac38fd71c
@ -66,7 +66,7 @@ if %BITS% == ? (
|
||||
set ARCH=x86
|
||||
set ARCHW=Win32
|
||||
if %BITS% == 64 (
|
||||
set ARCH=k8
|
||||
set ARCH=amd64
|
||||
set ARCHW=x64
|
||||
)
|
||||
cl config.guess.c > nul 2>&1
|
||||
|
||||
79
mpn/x86_64w/nehalem/gmp-mparam.h
Normal file
79
mpn/x86_64w/nehalem/gmp-mparam.h
Normal file
@ -0,0 +1,79 @@
|
||||
/* Generated by tuneup.c, 2010-01-06, system compiler */
|
||||
|
||||
#define MUL_KARATSUBA_THRESHOLD 20
|
||||
#define MUL_TOOM3_THRESHOLD 66
|
||||
#define MUL_TOOM4_THRESHOLD 292
|
||||
#define MUL_TOOM7_THRESHOLD 577
|
||||
|
||||
#define SQR_BASECASE_THRESHOLD 0 /* always (native) */
|
||||
#define SQR_KARATSUBA_THRESHOLD 28
|
||||
#define SQR_TOOM3_THRESHOLD 90
|
||||
#define SQR_TOOM4_THRESHOLD 90
|
||||
#define SQR_TOOM7_THRESHOLD 90
|
||||
|
||||
#define MULLOW_BASECASE_THRESHOLD 21
|
||||
#define MULLOW_DC_THRESHOLD 21
|
||||
#define MULLOW_MUL_THRESHOLD 21
|
||||
|
||||
#define MULHIGH_BASECASE_THRESHOLD 0 /* always */
|
||||
#define MULHIGH_DC_THRESHOLD 16
|
||||
#define MULHIGH_MUL_THRESHOLD 16
|
||||
|
||||
#define MULMOD_2EXPM1_THRESHOLD 1
|
||||
|
||||
#define DIV_SB_PREINV_THRESHOLD 4
|
||||
#define DIV_DC_THRESHOLD 49
|
||||
#define POWM_THRESHOLD 162
|
||||
#define FAC_UI_THRESHOLD 16187
|
||||
|
||||
#define GCD_ACCEL_THRESHOLD 890
|
||||
#define GCDEXT_THRESHOLD MP_SIZE_T_MAX /* never */
|
||||
#define JACOBI_BASE_METHOD 1
|
||||
|
||||
#define DIVREM_1_NORM_THRESHOLD MP_SIZE_T_MAX /* never */
|
||||
#define DIVREM_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */
|
||||
#define MOD_1_NORM_THRESHOLD 0 /* always */
|
||||
#define MOD_1_UNNORM_THRESHOLD 0 /* always */
|
||||
#define USE_PREINV_DIVREM_1 0
|
||||
#define USE_PREINV_MOD_1 1
|
||||
#define DIVREM_2_THRESHOLD MP_SIZE_T_MAX /* never */
|
||||
#define DIVEXACT_1_THRESHOLD 0 /* always (native) */
|
||||
#define MODEXACT_1_ODD_THRESHOLD 0 /* always (native) */
|
||||
#define MOD_1_1_THRESHOLD 116
|
||||
#define MOD_1_2_THRESHOLD 117
|
||||
#define MOD_1_3_THRESHOLD 117
|
||||
#define DIVREM_HENSEL_QR_1_THRESHOLD 12
|
||||
#define RSH_DIVREM_HENSEL_QR_1_THRESHOLD 3
|
||||
#define DIVREM_EUCLID_HENSEL_THRESHOLD 8
|
||||
|
||||
#define ROOTREM_THRESHOLD 8
|
||||
|
||||
#define GET_STR_DC_THRESHOLD 10
|
||||
#define GET_STR_PRECOMPUTE_THRESHOLD 12
|
||||
#define SET_STR_THRESHOLD 7059
|
||||
|
||||
#define MUL_FFT_TABLE { 464, 1056, 2880, 3328, 9216, 28672, 0 }
|
||||
#define MUL_FFT_MODF_THRESHOLD 464
|
||||
#define MUL_FFT_THRESHOLD 4864
|
||||
|
||||
#define SQR_FFT_TABLE { 496, 1056, 2752, 3328, 9216, 28672, 0 }
|
||||
#define SQR_FFT_MODF_THRESHOLD 464
|
||||
#define SQR_FFT_THRESHOLD 3200
|
||||
|
||||
/* Tuneup completed successfully, took 6 seconds */
|
||||
|
||||
#if 1
|
||||
|
||||
#define MUL_FFT_TABLE2 {{1, 4}, {377, 5}, {386, 4}, {404, 5}, {743, 6}, {777, 5}, {795, 6}, {850, 5}, {869, 6}, {2144, 7}, {2191, 6}, {2289, 7}, {2445, 6}, {2499, 7}, {2610, 8}, {2849, 7}, {3109, 8}, {3393, 7}, {3622, 8}, {3867, 7}, {4039, 8}, {5598, 9}, {5721, 8}, {6957, 9}, {7756, 8}, {8460, 9}, {8836, 8}, {9030, 9}, {9850, 8}, {10066, 9}, {11976, 8}, {12239, 9}, {13939, 10}, {15202, 9}, {18084, 10}, {19722, 9}, {20154, 10}, {23970, 9}, {24495, 10}, {27898, 9}, {28509, 10}, {30424, 11}, {31091, 10}, {31772, 9}, {32468, 10}, {40326, 11}, {41210, 10}, {42113, 11}, {47962, 10}, {53451, 11}, {62207, 12}, {63570, 10}, {73983, 11}, {80681, 10}, {82448, 11}, {98051, 12}, {100199, 11}, {102394, 12}, {127165, 11}, {164920, 12}, {196129, 13}, {254354, 12}, {259924, 11}, {277377, 12}, {392279, 13}, {519871, 12}, {531255, 11}, {566925, 12}, {592026, 11}, {631776, 12}, {659749, 13}, {784582, 11}, {995682, 12}, {MP_SIZE_T_MAX,0}}
|
||||
|
||||
#define MUL_FFTM_TABLE2 {{1, 4}, {278, 5}, {285, 4}, {344, 5}, {360, 4}, {377, 5}, {743, 6}, {777, 5}, {813, 6}, {1612, 7}, {4039, 8}, {4128, 7}, {4312, 8}, {5598, 9}, {5721, 8}, {6518, 9}, {6661, 8}, {7266, 9}, {7756, 8}, {8100, 9}, {14245, 10}, {14557, 9}, {14876, 10}, {15535, 9}, {16944, 10}, {17696, 9}, {18084, 10}, {19722, 9}, {20154, 10}, {23970, 9}, {24495, 10}, {29772, 11}, {30424, 10}, {36184, 11}, {39461, 10}, {42113, 11}, {47962, 10}, {49013, 11}, {58292, 12}, {59569, 13}, {62207, 12}, {63570, 10}, {64963, 11}, {80681, 12}, {84254, 11}, {87985, 12}, {95949, 11}, {109271, 12}, {111664, 11}, {114110, 12}, {127165, 11}, {147989, 12}, {161386, 11}, {164920, 12}, {196129, 13}, {200424, 12}, {218566, 13}, {223352, 12}, {228243, 13}, {254354, 12}, {329862, 13}, {392279, 14}, {508731, 13}, {519871, 12}, {554776, 13}, {913042, 14}, {MP_SIZE_T_MAX,0}}
|
||||
|
||||
#define SQR_FFT_TABLE2 {{1, 4}, {344, 5}, {360, 4}, {386, 5}, {743, 6}, {777, 5}, {813, 6}, {2009, 7}, {2053, 6}, {2340, 7}, {2727, 8}, {2849, 7}, {3109, 8}, {3393, 7}, {3468, 8}, {3867, 7}, {4039, 8}, {5478, 9}, {5721, 8}, {6957, 9}, {7756, 8}, {8460, 9}, {8836, 8}, {9030, 9}, {9850, 8}, {10066, 9}, {11976, 8}, {12239, 9}, {13939, 10}, {15535, 9}, {18084, 10}, {18885, 9}, {19299, 10}, {19722, 9}, {20154, 10}, {23970, 9}, {24495, 10}, {27898, 11}, {29134, 10}, {29772, 11}, {31091, 10}, {31772, 9}, {32468, 10}, {41210, 11}, {42113, 10}, {44943, 11}, {47962, 10}, {52305, 11}, {62207, 12}, {63570, 10}, {69326, 11}, {80681, 10}, {82448, 11}, {98051, 12}, {127165, 11}, {164920, 12}, {196129, 13}, {254354, 12}, {259924, 11}, {265616, 12}, {277377, 10}, {283451, 11}, {289658, 12}, {392279, 13}, {519871, 12}, {531255, 11}, {579340, 12}, {592026, 11}, {719463, 10}, {735218, 11}, {784582, 10}, {801763, 11}, {953467, 12}, {974346, 13}, {995682, 14}, {MP_SIZE_T_MAX,0}}
|
||||
|
||||
#define SQR_FFTM_TABLE2 {{1, 4}, {248, 5}, {254, 4}, {278, 5}, {285, 4}, {306, 5}, {313, 4}, {320, 5}, {328, 4}, {344, 5}, {695, 6}, {711, 5}, {743, 6}, {1476, 7}, {1543, 6}, {1612, 7}, {4039, 8}, {4128, 7}, {4312, 8}, {4407, 7}, {4504, 8}, {6518, 9}, {6661, 8}, {7110, 9}, {7756, 8}, {8100, 9}, {13939, 10}, {14557, 9}, {14876, 10}, {15535, 9}, {16944, 10}, {17696, 9}, {18084, 10}, {19722, 9}, {20154, 10}, {23970, 9}, {24495, 10}, {29772, 11}, {31091, 10}, {36977, 11}, {37787, 10}, {38615, 11}, {39461, 10}, {42113, 11}, {47962, 10}, {49013, 11}, {50087, 10}, {51184, 11}, {55819, 12}, {57042, 11}, {59569, 12}, {60874, 13}, {62207, 12}, {63570, 10}, {64963, 11}, {89912, 12}, {95949, 11}, {104637, 12}, {127165, 11}, {138676, 12}, {161386, 11}, {164920, 12}, {196129, 13}, {254354, 12}, {329862, 13}, {392279, 14}, {508731, 13}, {519871, 12}, {554776, 11}, {566925, 12}, {579340, 13}, {784582, 14}, {MP_SIZE_T_MAX,0}}
|
||||
|
||||
#define MUL_FFT_FULL_TABLE2 {{16, 1}, {1045, 2}, {1068, 1}, {1092, 2}, {1116, 1}, {1141, 2}, {1166, 1}, {1219, 2}, {1274, 1}, {1811, 2}, {1851, 1}, {1934, 2}, {1977, 1}, {2066, 4}, {2112, 6}, {2256, 5}, {2306, 4}, {2409, 1}, {2462, 3}, {2572, 1}, {2629, 4}, {2746, 2}, {2932, 4}, {3063, 1}, {3131, 2}, {3271, 1}, {3343, 4}, {3417, 3}, {3492, 2}, {3569, 3}, {3648, 4}, {3728, 5}, {3810, 1}, {3894, 2}, {3980, 4}, {4068, 1}, {4158, 4}, {4250, 6}, {4344, 2}, {4440, 5}, {4740, 4}, {4844, 1}, {4951, 3}, {5171, 2}, {5285, 1}, {5401, 2}, {5765, 1}, {5892, 2}, {6022, 4}, {6154, 1}, {6568, 3}, {6712, 1}, {7010, 4}, {7482, 1}, {7986, 4}, {8161, 2}, {8523, 4}, {8710, 2}, {8901, 5}, {9296, 1}, {9922, 2}, {10140, 3}, {10363, 2}, {10590, 1}, {11059, 2}, {11302, 1}, {11803, 2}, {12062, 4}, {12327, 1}, {13155, 3}, {13444, 1}, {14040, 4}, {14985, 1}, {15993, 2}, {17824, 1}, {18215, 3}, {18614, 1}, {20300, 2}, {20745, 1}, {23628, 2}, {24146, 4}, {24675, 1}, {28104, 2}, {28720, 4}, {29349, 1}, {32007, 2}, {35670, 1}, {36452, 4}, {38901, 1}, {40624, 3}, {41514, 2}, {44303, 1}, {45274, 2}, {46266, 1}, {47280, 4}, {50456, 3}, {53845, 1}, {60007, 4}, {61321, 3}, {62664, 6}, {64037, 8}, {65440, 11}, {66873, 13}, {68338, 10}, {69835, 7}, {71365, 8}, {72928, 5}, {74525, 4}, {79530, 1}, {81272, 3}, {83052, 2}, {90571, 1}, {94582, 4}, {96654, 2}, {98771, 4}, {100934, 1}, {103145, 2}, {107713, 1}, {112483, 4}, {114947, 1}, {130905, 2}, {142755, 5}, {149076, 6}, {152341, 3}, {155677, 1}, {162570, 3}, {166130, 2}, {169768, 3}, {173486, 2}, {197567, 4}, {201894, 3}, {215450, 1}, {261830, 2}, {267564, 5}, {273423, 7}, {279411, 9}, {291783, 7}, {298173, 6}, {304703, 4}, {325163, 3}, {346997, 2}, {395159, 4}, {403812, 7}, {412655, 10}, {430925, 6}, {450005, 5}, {459859, 1}, {523686, 2}, {583592, 4}, {596372, 1}, {650350, 2}, {664591, 3}, {694016, 1}, {790340, 4}, {825333, 6}, {843406, 3}, {861875, 2}, {880748, 1}, {919743, 2}, {939883, 1}, {981496, 3}, {MP_SIZE_T_MAX,0}}
|
||||
|
||||
#define SQR_FFT_FULL_TABLE2 {{16, 1}, {1092, 2}, {1116, 1}, {1811, 2}, {1851, 1}, {1934, 2}, {2021, 1}, {2066, 4}, {2112, 1}, {2207, 2}, {2256, 5}, {2357, 4}, {2409, 1}, {2462, 3}, {2572, 4}, {2746, 2}, {2932, 4}, {3063, 1}, {3131, 2}, {3271, 4}, {3417, 2}, {3648, 4}, {3728, 1}, {3894, 2}, {4068, 3}, {4158, 6}, {4344, 2}, {4440, 5}, {4638, 4}, {4844, 1}, {4951, 3}, {5171, 2}, {5285, 1}, {5401, 4}, {5520, 3}, {5641, 2}, {5765, 1}, {5892, 4}, {6154, 6}, {6289, 2}, {6568, 4}, {6859, 1}, {7010, 4}, {7482, 1}, {7986, 4}, {8161, 2}, {8523, 4}, {8710, 2}, {8901, 5}, {9296, 1}, {9922, 3}, {10363, 2}, {10590, 1}, {10822, 2}, {11550, 1}, {11803, 2}, {12062, 4}, {12327, 1}, {12873, 2}, {13155, 3}, {13444, 1}, {14040, 4}, {14985, 1}, {15993, 2}, {16344, 1}, {16702, 2}, {17824, 1}, {18215, 3}, {18614, 1}, {19022, 4}, {19439, 1}, {20300, 2}, {20745, 1}, {23628, 2}, {24146, 4}, {24675, 2}, {25216, 1}, {25769, 3}, {26911, 1}, {28104, 2}, {29349, 4}, {29992, 1}, {32007, 2}, {32708, 1}, {34157, 2}, {35670, 1}, {36452, 4}, {38901, 1}, {40624, 3}, {41514, 2}, {46266, 1}, {47280, 4}, {50456, 3}, {53845, 1}, {61321, 3}, {62664, 6}, {64037, 8}, {66873, 10}, {68338, 8}, {69835, 6}, {71365, 8}, {72928, 4}, {79530, 1}, {81272, 3}, {83052, 2}, {92555, 1}, {94582, 4}, {96654, 2}, {98771, 1}, {103145, 2}, {107713, 1}, {130905, 2}, {142755, 5}, {149076, 4}, {152341, 3}, {166130, 2}, {197567, 4}, {201894, 1}, {261830, 2}, {291783, 5}, {298173, 6}, {304703, 4}, {325163, 3}, {346997, 2}, {395159, 4}, {412655, 1}, {523686, 2}, {546873, 1}, {558848, 2}, {583592, 4}, {596372, 6}, {609431, 4}, {650350, 3}, {694016, 2}, {709213, 1}, {724743, 2}, {790340, 1}, {MP_SIZE_T_MAX,0}}
|
||||
|
||||
#endif
|
||||
Loading…
Reference in New Issue
Block a user