From 1487ddca4ac765245614e7790acb637674cb95fc Mon Sep 17 00:00:00 2001 From: jasonmoxham Date: Sat, 3 Dec 2011 15:15:19 +0000 Subject: [PATCH] sparc64 tuning --- mpn/sparc64/gmp-mparam.h | 106 +++++++++++++++++++-------------------- 1 file changed, 53 insertions(+), 53 deletions(-) diff --git a/mpn/sparc64/gmp-mparam.h b/mpn/sparc64/gmp-mparam.h index 084bc4c6..8fe91525 100644 --- a/mpn/sparc64/gmp-mparam.h +++ b/mpn/sparc64/gmp-mparam.h @@ -1,76 +1,76 @@ -/* Generated by tuneup.c, 2010-03-24, gcc 4.4 */ +/* Generated by tuneup.c, 2011-12-03, gcc 4.6 */ -#define MUL_KARATSUBA_THRESHOLD 22 -#define MUL_TOOM3_THRESHOLD 78 -#define MUL_TOOM4_THRESHOLD 396 -#define MUL_TOOM8H_THRESHOLD 396 +#define MUL_KARATSUBA_THRESHOLD 34 +#define MUL_TOOM3_THRESHOLD 97 +#define MUL_TOOM4_THRESHOLD 438 +#define MUL_TOOM8H_THRESHOLD 438 -#define SQR_BASECASE_THRESHOLD 8 -#define SQR_KARATSUBA_THRESHOLD 55 -#define SQR_TOOM3_THRESHOLD 81 -#define SQR_TOOM4_THRESHOLD 414 -#define SQR_TOOM8_THRESHOLD 414 +#define SQR_BASECASE_THRESHOLD 11 +#define SQR_KARATSUBA_THRESHOLD 68 +#define SQR_TOOM3_THRESHOLD 95 +#define SQR_TOOM4_THRESHOLD 494 +#define SQR_TOOM8_THRESHOLD 494 #define POWM_THRESHOLD 984 -#define GCD_ACCEL_THRESHOLD 30 -#define GCD_THRESHOLD 29 -#define GCDEXT_THRESHOLD 490 -#define JACOBI_BASE_METHOD 2 +#define HGCD_THRESHOLD 156 +#define GCD_DC_THRESHOLD 527 +#define GCDEXT_DC_THRESHOLD 454 +#define JACOBI_BASE_METHOD 3 -#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 MP_SIZE_T_MAX /* never */ -#define MOD_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVREM_1_NORM_THRESHOLD 3 +#define DIVREM_1_UNNORM_THRESHOLD 3 +#define MOD_1_NORM_THRESHOLD 3 +#define MOD_1_UNNORM_THRESHOLD 3 #define USE_PREINV_DIVREM_1 1 -#define USE_PREINV_MOD_1 0 -#define DIVREM_2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define USE_PREINV_MOD_1 1 +#define DIVREM_2_THRESHOLD 0 /* always */ #define DIVEXACT_1_THRESHOLD 0 /* always */ #define MODEXACT_1_ODD_THRESHOLD 0 /* always */ -#define MOD_1_1_THRESHOLD 24 -#define MOD_1_2_THRESHOLD 24 -#define MOD_1_3_THRESHOLD 115 +#define MOD_1_1_THRESHOLD 10 +#define MOD_1_2_THRESHOLD 10 +#define MOD_1_3_THRESHOLD 23 #define DIVREM_HENSEL_QR_1_THRESHOLD 996 #define RSH_DIVREM_HENSEL_QR_1_THRESHOLD 996 -#define DIVREM_EUCLID_HENSEL_THRESHOLD 95 +#define DIVREM_EUCLID_HENSEL_THRESHOLD 40 #define ROOTREM_THRESHOLD 3 -#define GET_STR_DC_THRESHOLD 15 -#define GET_STR_PRECOMPUTE_THRESHOLD 29 -#define SET_STR_DC_THRESHOLD 244 -#define SET_STR_PRECOMPUTE_THRESHOLD 282 +#define GET_STR_DC_THRESHOLD 13 +#define GET_STR_PRECOMPUTE_THRESHOLD 18 +#define SET_STR_DC_THRESHOLD 290 +#define SET_STR_PRECOMPUTE_THRESHOLD 1394 -#define MUL_FFT_TABLE { 240, 352, 960, 1280, 5120, 20480, 81920, 196608, 1310720, 5242880, 12582912, 0 } -#define MUL_FFT_MODF_THRESHOLD 216 -#define MUL_FFT_FULL_THRESHOLD 1056 +#define MUL_FFT_TABLE { 240, 480, 1216, 1792, 7168, 20480, 81920, 327680, 1835008, 5242880, 0 } +#define MUL_FFT_MODF_THRESHOLD 232 +#define MUL_FFT_FULL_THRESHOLD 1312 -#define SQR_FFT_TABLE { 240, 416, 1088, 1792, 5120, 20480, 81920, 196608, 1310720, 5242880, 0 } -#define SQR_FFT_MODF_THRESHOLD 216 -#define SQR_FFT_FULL_THRESHOLD 1184 +#define SQR_FFT_TABLE { 304, 608, 1344, 2304, 7168, 20480, 81920, 196608, 1835008, 5242880, 0 } +#define SQR_FFT_MODF_THRESHOLD 264 +#define SQR_FFT_FULL_THRESHOLD 1440 -#define MULLOW_BASECASE_THRESHOLD 0 /* always */ -#define MULLOW_DC_THRESHOLD 12 -#define MULLOW_MUL_THRESHOLD 1623 +#define MULLOW_BASECASE_THRESHOLD 30 +#define MULLOW_DC_THRESHOLD 30 +#define MULLOW_MUL_THRESHOLD 1877 -#define MULHIGH_BASECASE_THRESHOLD 4 -#define MULHIGH_DC_THRESHOLD 11 -#define MULHIGH_MUL_THRESHOLD 1623 +#define MULHIGH_BASECASE_THRESHOLD 42 +#define MULHIGH_DC_THRESHOLD 42 +#define MULHIGH_MUL_THRESHOLD 1877 -#define MULMOD_2EXPM1_THRESHOLD 10 +#define MULMOD_2EXPM1_THRESHOLD 14 -#define FAC_UI_THRESHOLD 2687 -#define DC_DIV_QR_THRESHOLD 10 -#define DC_DIVAPPR_Q_N_THRESHOLD 41 -#define INV_DIV_QR_THRESHOLD 1652 -#define INV_DIVAPPR_Q_N_THRESHOLD 41 -#define DC_DIV_Q_THRESHOLD 75 -#define INV_DIV_Q_THRESHOLD 1652 -#define DC_DIVAPPR_Q_THRESHOLD 55 -#define INV_DIVAPPR_Q_THRESHOLD 3679 -#define DC_BDIV_QR_THRESHOLD 18 -#define DC_BDIV_Q_THRESHOLD 14 -/* Tuneup completed successfully, took 8333 seconds */ +#define FAC_UI_THRESHOLD 12757 +#define DC_DIV_QR_THRESHOLD 17 +#define DC_DIVAPPR_Q_N_THRESHOLD 122 +#define INV_DIV_QR_THRESHOLD 693 +#define INV_DIVAPPR_Q_N_THRESHOLD 122 +#define DC_DIV_Q_THRESHOLD 180 +#define INV_DIV_Q_THRESHOLD 595 +#define DC_DIVAPPR_Q_THRESHOLD 118 +#define INV_DIVAPPR_Q_THRESHOLD 1975 +#define DC_BDIV_QR_THRESHOLD 28 +#define DC_BDIV_Q_THRESHOLD 84 +/* Tuneup completed successfully, took 2566 seconds */ #define MUL_FFT_TABLE2 {{1, 4}, {278, 5}, {292, 4}, {306, 5}, {622, 6}, {650, 5}, {680, 6}, {1476, 7}, {1543, 6}, {1612, 7}, {2727, 8}, {2787, 7}, {3248, 8}, {6957, 9}, {7756, 8}, {8278, 9}, {9850, 8}, {10066, 9}, {13939, 10}, {15535, 9}, {18084, 10}, {19722, 9}, {20154, 10}, {20596, 9}, {21047, 10}, {23970, 9}, {24495, 10}, {27898, 11}, {31091, 10}, {40326, 11}, {47962, 10}, {52305, 11}, {53451, 12}, {63570, 10}, {66386, 11}, {67840, 10}, {69326, 11}, {80681, 10}, {82448, 11}, {98051, 12}, {100199, 11}, {102394, 12}, {127165, 11}, {164920, 12}, {196129, 13}, {254354, 12}, {259924, 11}, {277377, 12}, {329862, 11}, {359720, 12}, {392279, 11}, {427785, 12}, {437153, 13}, {519871, 12}, {531255, 10}, {566925, 11}, {592026, 12}, {719463, 13}, {784582, 12}, {933036, 11}, {995682, 12}, {MP_SIZE_T_MAX,0}}