sqr_toom3.c 文件参考

#include "../../include/lammp/lmmpn.h"
#include "../../include/lammp/impl/toom_interp.h"

sqr_toom3.c 的引用(Include)关系图:

浏览源代码.

宏定义
#define	a0   numa
#define	a1   (numa + n)
#define	a2   (numa + 2 * n)
#define	am1   (dst)
#define	ap1   tp
#define	ap2   ap1
#define	v0   dst
#define	v1   (dst + 2 * n)
#define	v2   (tp + 2 * n + 2)
#define	vinf   (dst + 4 * n)
#define	vm1   tp

函数
void	lmmp_sqr_toom3_ (mp_ptr restrict dst, mp_srcptr restrict numa, mp_size_t na)

宏定义说明

a0

#define a0   numa

a1

#define a1   (numa + n)

a2

#define a2   (numa + 2 * n)

am1

#define am1   (dst)

ap1

#define ap1   tp

ap2

#define ap2   ap1

v0

#define v0   dst

v1

#define v1   (dst + 2 * n)

v2

#define v2   (tp + 2 * n + 2)

vinf

#define vinf   (dst + 4 * n)

vm1

#define vm1   tp

函数说明

lmmp_sqr_toom3_()

void lmmp_sqr_toom3_	(	mp_ptr restrict	dst,
		mp_srcptr restrict	numa,
		mp_size_t	na
	)

在文件 sqr_toom3.c 第 23 行定义.

                                                                                 {
    lmmp_param_assert(na > 0);
    lmmp_param_assert(numa != NULL);
    lmmp_param_assert(dst != NULL);
    TEMP_S_DECL;
    mp_size_t n = (na + 2) / 3, s = na - 2 * n;
    mp_limb_t cy, cy2, vinf0, am1h;
    mp_limb_t* restrict tp = SALLOC_TYPE(4 * n + 4, mp_limb_t);
 
#define a0 numa
#define a1 (numa + n)
#define a2 (numa + 2 * n)
 
#define v0 dst               //[dst,2*n]
#define v1 (dst + 2 * n)     //[dst+2*n,2*n+1]
#define vinf (dst + 4 * n)   //[dst+4*n,s+t]
#define vm1 tp               //[tp,2*n+1]
#define v2 (tp + 2 * n + 2)  //[tp+2*n+2,2*n+1]
 
#define am1 (dst)  //[dst,n]
#define ap1 tp     //[tp,n+1]
#define ap2 ap1    // same space
 
    // ap1, am1
    cy = lmmp_add_(ap1, a0, n, a2, s);
    if (cy == 0 && lmmp_cmp_(ap1, a1, n) < 0) {
        cy = lmmp_add_n_sub_n_(ap1, am1, a1, ap1, n);
        ap1[n] = cy >> 1;
        am1h = 0;
    } else {
        cy2 = lmmp_add_n_sub_n_(ap1, am1, ap1, a1, n);
        ap1[n] = cy + (cy2 >> 1);
        am1h = cy - (cy2 & 1);
    }
 
    // vinf
    lmmp_sqr_(vinf, a2, s);
    vinf0 = vinf[0];  // overlap with v1
    cy = vinf[1];     // overlap with v1
 
    // v1
    lmmp_sqr_(v1, ap1, n + 1);
    vinf[1] = cy;  // restore, since v1[2*n+1]==0.
 
    // ap2
    cy = lmmp_addshl1_n_(ap2, a1, a2, s);
    if (s != n)
        cy = lmmp_add_1_(ap2 + s, a1 + s, n - s, cy);
    cy = 2 * cy + lmmp_addshl1_n_(ap2, a0, ap2, n);
    ap2[n] = cy;
 
    // v2
    lmmp_sqr_(v2, ap2, n + 1);
 
    // vm1
    lmmp_sqr_(vm1, am1, n);
    if (am1h)
        am1h += lmmp_addshl1_n_(vm1 + n, vm1 + n, am1, n);
    vm1[2 * n] = am1h;
 
    // v0
    lmmp_sqr_(v0, a0, n);
 
    lmmp_toom_interp5_(dst, v2, vm1, n, s + s, 0, vinf0);
    TEMP_S_FREE;
}

引用了 a0, a1, a2, am1, ap1, ap2, lmmp_add_(), lmmp_add_1_(), lmmp_add_n_sub_n_(), lmmp_addshl1_n_(), lmmp_cmp_(), lmmp_param_assert, lmmp_sqr_(), lmmp_toom_interp5_(), SALLOC_TYPE, TEMP_S_DECL, TEMP_S_FREE, tp, v0, v1, v2, vinf , 以及 vm1.

函数调用图: