From 1b5f3084b8c02079fdf0349c5516faa0c16bef27 Mon Sep 17 00:00:00 2001 From: Richard Kreckel Date: Mon, 14 Jan 2008 14:52:09 +0000 Subject: [PATCH] * Clarify that GMP and intDsize cannot be tuned separately. --- src/TUNING | 88 ++++++++++++++++++++++++++++-------------------------- 1 file changed, 45 insertions(+), 43 deletions(-) diff --git a/src/TUNING b/src/TUNING index fc24b7c..2519510 100644 --- a/src/TUNING +++ b/src/TUNING @@ -1,50 +1,52 @@ Tips for performance tuning on a specific architecture: -1. Choose the optimal limb size (intDsize). This is fundamental. On 32-bit - platforms intDsize=32 is best. On 64-bit platforms intDsize=64 may be - better, especially if there is a 64x64-bit multiplication in hardware. +1a. Choose the optimal digit size (intDsize). This is fundamental. On 32-bit + platforms intDsize=32 is best. On 64-bit platforms intDsize=64 may be + better, especially if there is a 64x64-bit multiplication in hardware. -2. Tune GMP. +1b. Alternatively, tune GMP. When GMP is used, CLN's digit size (intDsize) has + to match GMP's limb size (sizeof(mp_limb_t)). There is nothing to do at the + CLN side: The configure script will take care of intDsize automatically. -3. The break-even points between several algorithms for the same task - have to be determined experimentally, in the order given below: +2. The break-even points between several algorithms for the same task + have to be determined experimentally, in the order given below: - multiplication: - cl_DS_mul.cc karatsuba_threshold - cl_DS_mul.cc function cl_fftm_suitable - division: - cl_DS_div.cc function cl_recip_suitable - 2-adic reciprocal: - cl_2DS_recip.cc recip2adic_threshold - 2-adic division: - cl_2DS_div.cc function cl_recip_suitable - square root: - cl_DS_sqrt.cc function cl_recipsqrt_suitable - cl_LF_sqrt.cc "if (len > ...)" - gcd: - cl_I_gcd.cc cl_gcd_double_threshold - binary->decimal conversion: - cl_I_to_digits.cc cl_digits_div_threshold - pi: - cl_LF_pi.cc best of 4 algorithms - exp, log: - cl_F_expx.cc factor limit_slope of isqrt(d) - cl_R_exp.cc inside function exp - cl_R_ln.cc inside function ln - eulerconst: - cl_LF_eulerconst.cc function compute_eulerconst - sin, cos, sinh, cosh: - cl_F_sinx.cc factor limit_slope of isqrt(d) - cl_R_sin.cc inside function sin - cl_R_cos.cc inside function cos - cl_R_cossin.cc inside function cl_cos_sin - cl_F_sinhx.cc factor limit_slope of isqrt(d) - cl_R_sinh.cc inside function sinh - cl_R_cosh.cc inside function cosh - cl_R_coshsinh.cc inside function cl_cosh_sinh - cl_F_atanx.cc factor limit_slope of isqrt(d) - cl_F_atanx.cc inside function atanx - cl_F_atanhx.cc factor limit_slope of isqrt(d) - cl_F_atanhx.cc inside function atanhx + multiplication: + cl_DS_mul.cc karatsuba_threshold + cl_DS_mul.cc function cl_fftm_suitable + division: + cl_DS_div.cc function cl_recip_suitable + 2-adic reciprocal: + cl_2DS_recip.cc recip2adic_threshold + 2-adic division: + cl_2DS_div.cc function cl_recip_suitable + square root: + cl_DS_sqrt.cc function cl_recipsqrt_suitable + cl_LF_sqrt.cc "if (len > ...)" + gcd: + cl_I_gcd.cc cl_gcd_double_threshold + binary->decimal conversion: + cl_I_to_digits.cc cl_digits_div_threshold + pi: + cl_LF_pi.cc best of 4 algorithms + exp, log: + cl_F_expx.cc factor limit_slope of isqrt(d) + cl_R_exp.cc inside function exp + cl_R_ln.cc inside function ln + eulerconst: + cl_LF_eulerconst.cc function compute_eulerconst + sin, cos, sinh, cosh: + cl_F_sinx.cc factor limit_slope of isqrt(d) + cl_R_sin.cc inside function sin + cl_R_cos.cc inside function cos + cl_R_cossin.cc inside function cl_cos_sin + cl_F_sinhx.cc factor limit_slope of isqrt(d) + cl_R_sinh.cc inside function sinh + cl_R_cosh.cc inside function cosh + cl_R_coshsinh.cc inside function cl_cosh_sinh + cl_F_atanx.cc factor limit_slope of isqrt(d) + cl_F_atanx.cc inside function atanx + cl_F_atanhx.cc factor limit_slope of isqrt(d) + cl_F_atanhx.cc inside function atanhx