/* p = amd(A+A') if symmetric is true, or amd(A'A) otherwise */
int *csr_amd (int order, const csr *A) /* order 0:natural, 1:Chol, 2:LU, 3:QR */
{
csr *C, *A2, *AT ;
int *Cp, *Cj, *last, *W, *len, *nv, *next, *P, *head, *elen, *degree, *w,
*hhead, *ATp, *ATj, d, dk, dext, lemax = 0, e, elenk, eln, i, j, k, k1,
k2, k3, jlast, ln, dense, nzmax, mindeg = 0, nvi, nvj, nvk, mark, wnvi,
ok, cnz, nel = 0, p, p1, p2, p3, p4, pj, pk, pk1, pk2, pn, q, n, m, t ;
unsigned int h ;
/* --- Construct matrix C ----------------------------------------------- */
if (!CS_CSC (A) || order <= 0 || order > 3) return (NULL) ; /* check */
AT = csr_transpose (A, 0) ; /* compute A' */
if (!AT) return (NULL) ;
m = A->m ; n = A->n ;
if ( n != m) return(NULL); /* For rectangular matrices, use csr_amd */
dense = (int)CS_MAX (16, 10 * sqrt ((double) n)) ; /* find dense threshold */
dense = CS_MIN (n-2, dense) ;
if (order == 1 && n == m)
{
C = csr_add (A, AT, 0, 0) ; /* C = A+A' */
}
else if (order == 2)
{
ATp = AT->p ; /* drop dense columns from AT */
ATj = AT->j ;
for (p2 = 0, j = 0 ; j < m ; j++)
{
p = ATp [j] ; /* column j of AT starts here */
ATp [j] = p2 ; /* new column j starts here */
if (ATp [j+1] - p > dense) continue ; /* skip dense col j */
for ( ; p < ATp [j+1] ; p++) ATj [p2++] = ATj [p] ;
}
ATp [m] = p2 ; /* finalize AT */
A2 = csr_transpose (AT, 0) ; /* A2 = AT' */
C = A2 ? csr_multiply (AT, A2) : NULL ; /* C=A'*A with no dense rows */
csr_spfree (A2) ;
}
else
{
C = csr_multiply (AT, A) ; /* C=A'*A */
}
csr_spfree (AT) ;
if (!C) return (NULL) ;
csr_fkeep (C, &csr_diag, NULL) ; /* drop diagonal entries */
Cp = C->p ;
cnz = Cp [n] ;
P = (int*)malloc (CS_MAX(n+1,1) * sizeof (int)) ; /* allocate result */
W = (int*)malloc (CS_MAX(8*(n+1),1) * sizeof (int)) ; /* get workspace */
t = cnz + cnz/5 + 2*n ; /* add elbow room to C */
if (!P || !W || !csr_sprealloc (C, t)) return (csr_idone (P, C, W, 0)) ;
len = W ; nv = W + (n+1) ; next = W + 2*(n+1) ;
head = W + 3*(n+1) ; elen = W + 4*(n+1) ; degree = W + 5*(n+1) ;
w = W + 6*(n+1) ; hhead = W + 7*(n+1) ;
last = P ; /* use P as workspace for last */
/* --- Initialize quotient graph ---------------------------------------- */
for (k = 0 ; k < n ; k++) len [k] = Cp [k+1] - Cp [k] ;
len [n] = 0 ;
nzmax = C->nzmax ;
Cj = C->j ;
for (i = 0 ; i <= n ; i++)
{
head [i] = -1 ; /* degree list i is empty */
last [i] = -1 ;
next [i] = -1 ;
hhead [i] = -1 ; /* hash list i is empty */
nv [i] = 1 ; /* node i is just one node */
w [i] = 1 ; /* node i is alive */
elen [i] = 0 ; /* Ek of node i is empty */
degree [i] = len [i] ; /* degree of node i */
}
mark = csr_wclear (0, 0, w, n) ; /* clear w */
elen [n] = -2 ; /* n is a dead element */
Cp [n] = -1 ; /* n is a root of assembly tree */
w [n] = 0 ; /* n is a dead element */
/* --- Initialize degree lists ------------------------------------------ */
for (i = 0 ; i < n ; i++)
{
d = degree [i] ;
if (d == 0) /* node i is empty */
{
elen [i] = -2 ; /* element i is dead */
nel++ ;
Cp [i] = -1 ; /* i is a root of assemby tree */
w [i] = 0 ;
}
else if (d > dense) /* node i is dense */
{
nv [i] = 0 ; /* absorb i into element n */
elen [i] = -1 ; /* node i is dead */
nel++ ;
Cp [i] = CS_FLIP (n) ;
nv [n]++ ;
}
else
{
if (head [d] != -1) last [head [d]] = i ;
next [i] = head [d] ; /* put node i in degree list d */
head [d] = i ;
}
}
while (nel < n) /* while (selecting pivots) do */
{
/* --- Select node of minimum approximate degree -------------------- */
for (k = -1 ; mindeg < n && (k = head [mindeg]) == -1 ; mindeg++) ;
if (next [k] != -1) last [next [k]] = -1 ;
head [mindeg] = next [k] ; /* remove k from degree list */
elenk = elen [k] ; /* elenk = |Ek| */
nvk = nv [k] ; /* # of nodes k represents */
nel += nvk ; /* nv[k] nodes of A eliminated */
/* --- Garbage collection ------------------------------------------- */
if (elenk > 0 && cnz + mindeg >= nzmax)
{
for (j = 0 ; j < n ; j++)
{
if ((p = Cp [j]) >= 0) /* j is a live node or element */
{
Cp [j] = Cj [p] ; /* save first entry of object */
Cj [p] = CS_FLIP (j) ; /* first entry is now CS_FLIP(j) */
}
}
for (q = 0, p = 0 ; p < cnz ; ) /* scan all of memory */
{
if ((j = CS_FLIP (Cj [p++])) >= 0) /* found object j */
{
Cj [q] = Cp [j] ; /* restore first entry of object */
Cp [j] = q++ ; /* new pointer to object j */
for (k3 = 0 ; k3 < len [j]-1 ; k3++) Cj [q++] = Cj [p++] ;
}
}
cnz = q ; /* Cj [cnz...nzmax-1] now free */
}
/* --- Construct new element ---------------------------------------- */
dk = 0 ;
nv [k] = -nvk ; /* flag k as in Lk */
p = Cp [k] ;
pk1 = (elenk == 0) ? p : cnz ; /* do in place if elen[k] == 0 */
pk2 = pk1 ;
for (k1 = 1 ; k1 <= elenk + 1 ; k1++)
{
if (k1 > elenk)
{
e = k ; /* search the nodes in k */
pj = p ; /* list of nodes starts at Cj[pj]*/
ln = len [k] - elenk ; /* length of list of nodes in k */
}
else
{
e = Cj [p++] ; /* search the nodes in e */
pj = Cp [e] ;
ln = len [e] ; /* length of list of nodes in e */
}
for (k2 = 1 ; k2 <= ln ; k2++)
{
i = Cj [pj++] ;
if ((nvi = nv [i]) <= 0) continue ; /* node i dead, or seen */
dk += nvi ; /* degree[Lk] += size of node i */
nv [i] = -nvi ; /* negate nv[i] to denote i in Lk*/
Cj [pk2++] = i ; /* place i in Lk */
if (next [i] != -1) last [next [i]] = last [i] ;
if (last [i] != -1) /* remove i from degree list */
{
next [last [i]] = next [i] ;
}
else
{
head [degree [i]] = next [i] ;
}
}
if (e != k)
{
Cp [e] = CS_FLIP (k) ; /* absorb e into k */
w [e] = 0 ; /* e is now a dead element */
}
}
if (elenk != 0) cnz = pk2 ; /* Cj [cnz...nzmax] is free */
degree [k] = dk ; /* external degree of k - |Lk\i| */
Cp [k] = pk1 ; /* element k is in Cj[pk1..pk2-1] */
len [k] = pk2 - pk1 ;
elen [k] = -2 ; /* k is now an element */
/* --- Find set differences ----------------------------------------- */
mark = csr_wclear (mark, lemax, w, n) ; /* clear w if necessary */
for (pk = pk1 ; pk < pk2 ; pk++) /* scan 1: find |Le\Lk| */
{
i = Cj [pk] ;
if ((eln = elen [i]) <= 0) continue ;/* skip if elen[i] empty */
nvi = -nv [i] ; /* nv [i] was negated */
wnvi = mark - nvi ;
for (p = Cp [i] ; p <= Cp [i] + eln - 1 ; p++) /* scan Ei */
{
e = Cj [p] ;
if (w [e] >= mark)
{
w [e] -= nvi ; /* decrement |Le\Lk| */
}
else if (w [e] != 0) /* ensure e is a live element */
{
w [e] = degree [e] + wnvi ; /* 1st time e seen in scan 1 */
}
}
}
/* --- Degree update ------------------------------------------------ */
for (pk = pk1 ; pk < pk2 ; pk++) /* scan2: degree update */
{
i = Cj [pk] ; /* consider node i in Lk */
p1 = Cp [i] ;
p2 = p1 + elen [i] - 1 ;
pn = p1 ;
for (h = 0, d = 0, p = p1 ; p <= p2 ; p++) /* scan Ei */
{
e = Cj [p] ;
if (w [e] != 0) /* e is an unabsorbed element */
{
dext = w [e] - mark ; /* dext = |Le\Lk| */
if (dext > 0)
{
d += dext ; /* sum up the set differences */
Cj [pn++] = e ; /* keep e in Ei */
h += e ; /* compute the hash of node i */
}
else
{
Cp [e] = CS_FLIP (k) ; /* aggressive absorb. e->k */
w [e] = 0 ; /* e is a dead element */
}
}
}
elen [i] = pn - p1 + 1 ; /* elen[i] = |Ei| */
p3 = pn ;
p4 = p1 + len [i] ;
for (p = p2 + 1 ; p < p4 ; p++) /* prune edges in Ai */
{
j = Cj [p] ;
if ((nvj = nv [j]) <= 0) continue ; /* node j dead or in Lk */
d += nvj ; /* degree(i) += |j| */
Cj [pn++] = j ; /* place j in node list of i */
h += j ; /* compute hash for node i */
}
if (d == 0) /* check for mass elimination */
{
Cp [i] = CS_FLIP (k) ; /* absorb i into k */
nvi = -nv [i] ;
dk -= nvi ; /* |Lk| -= |i| */
nvk += nvi ; /* |k| += nv[i] */
nel += nvi ;
nv [i] = 0 ;
elen [i] = -1 ; /* node i is dead */
}
else
{
degree [i] = CS_MIN (degree [i], d) ; /* update degree(i) */
Cj [pn] = Cj [p3] ; /* move first node to end */
Cj [p3] = Cj [p1] ; /* move 1st el. to end of Ei */
Cj [p1] = k ; /* add k as 1st element in of Ei */
len [i] = pn - p1 + 1 ; /* new len of adj. list of node i */
h %= n ; /* finalize hash of i */
next [i] = hhead [h] ; /* place i in hash bucket */
hhead [h] = i ;
last [i] = h ; /* save hash of i in last[i] */
}
} /* scan2 is done */
degree [k] = dk ; /* finalize |Lk| */
lemax = CS_MAX (lemax, dk) ;
mark = csr_wclear (mark+lemax, lemax, w, n) ; /* clear w */
/* --- Supernode detection ------------------------------------------ */
for (pk = pk1 ; pk < pk2 ; pk++)
{
i = Cj [pk] ;
if (nv [i] >= 0) continue ; /* skip if i is dead */
h = last [i] ; /* scan hash bucket of node i */
i = hhead [h] ;
hhead [h] = -1 ; /* hash bucket will be empty */
for ( ; i != -1 && next [i] != -1 ; i = next [i], mark++)
{
ln = len [i] ;
eln = elen [i] ;
for (p = Cp [i]+1 ; p <= Cp [i] + ln-1 ; p++) w [Cj [p]] = mark;
jlast = i ;
for (j = next [i] ; j != -1 ; ) /* compare i with all j */
{
ok = (len [j] == ln) && (elen [j] == eln) ;
for (p = Cp [j] + 1 ; ok && p <= Cp [j] + ln - 1 ; p++)
{
if (w [Cj [p]] != mark) ok = 0 ; /* compare i and j*/
}
if (ok) /* i and j are identical */
{
Cp [j] = CS_FLIP (i) ; /* absorb j into i */
nv [i] += nv [j] ;
nv [j] = 0 ;
elen [j] = -1 ; /* node j is dead */
j = next [j] ; /* delete j from hash bucket */
next [jlast] = j ;
}
else
{
jlast = j ; /* j and i are different */
j = next [j] ;
}
}
}
}
/* --- Finalize new element------------------------------------------ */
for (p = pk1, pk = pk1 ; pk < pk2 ; pk++) /* finalize Lk */
{
i = Cj [pk] ;
if ((nvi = -nv [i]) <= 0) continue ;/* skip if i is dead */
nv [i] = nvi ; /* restore nv[i] */
d = degree [i] + dk - nvi ; /* compute external degree(i) */
d = CS_MIN (d, n - nel - nvi) ;
if (head [d] != -1) last [head [d]] = i ;
next [i] = head [d] ; /* put i back in degree list */
last [i] = -1 ;
head [d] = i ;
mindeg = CS_MIN (mindeg, d) ; /* find new minimum degree */
degree [i] = d ;
Cj [p++] = i ; /* place i in Lk */
}
nv [k] = nvk ; /* # nodes absorbed into k */
if ((len [k] = p-pk1) == 0) /* length of adj list of element k*/
{
Cp [k] = -1 ; /* k is a root of the tree */
w [k] = 0 ; /* k is now a dead element */
}
if (elenk != 0) cnz = p ; /* free unused space in Lk */
}
/* --- Postordering ----------------------------------------------------- */
for (i = 0 ; i < n ; i++) Cp [i] = CS_FLIP (Cp [i]) ;/* fix assembly tree */
for (j = 0 ; j <= n ; j++) head [j] = -1 ;
for (j = n ; j >= 0 ; j--) /* place unordered nodes in lists */
{
if (nv [j] > 0) continue ; /* skip if j is an element */
next [j] = head [Cp [j]] ; /* place j in list of its parent */
head [Cp [j]] = j ;
}
for (e = n ; e >= 0 ; e--) /* place elements in lists */
{
if (nv [e] <= 0) continue ; /* skip unless e is an element */
if (Cp [e] != -1)
{
next [e] = head [Cp [e]] ; /* place e in list of its parent */
head [Cp [e]] = e ;
}
}
for (k = 0, i = 0 ; i <= n ; i++) /* postorder the assembly tree */
{
if (Cp [i] == -1) k = csr_tdfs (i, k, head, next, P, w) ;
}
return (csr_idone (P, C, W, 1)) ;
}
int main(int argc, char *argv[])
{
if (argc != 2)
{
fprintf(stderr, "usage: %s csr_matrix_file\n", argv[0]);
exit(0);
}
int i, j, k;
struct timespec start, end;
int num_threads = 1;
#pragma omp parallel
{
#pragma omp master
{
num_threads = omp_get_num_threads();
}
}
printf("Thread number: %d.\n", num_threads);
#pragma omp parallel for
for (i = 0; i < num_threads; i++)
{
int cpu = omp_get_thread_num();
thread_bind(cpu);
}
FILE *fp;
struct csr_mat_t csr, csr_re, csr_t, csr_t_re, csr_t_t;
struct blk_mat_t blk;
struct csr_cont_t csr_h, csr_v;
struct blk_cont_t blk_h, blk_t_h;
read_csr_mat(argv[1], &csr);
int rows = csr.rows;
int cols = csr.cols;
INT64 non_zeros = csr.non_zeros;
csr_transpose(&csr, &csr_t);
release_csr_mat(&csr);
int *reorder_map = (int*)malloc(cols * sizeof(int));
csr_reorder(&csr_t, &csr_re, reorder_map);
release_csr_mat(&csr_t);
csr_transpose(&csr_re, &csr_t_t);
release_csr_mat(&csr_re);
split_csr_lb_nz(&csr_t_t, &csr_h, num_threads, SPLIT_HORIZON);
release_csr_mat(&csr_t_t);
csr_cont_to_blk_cont(&csr_h, &blk_h);
release_csr_cont(&csr_h);
printf("Notify: finished the preprocessing.\n");
FLOAT *x = (FLOAT*)numa_alloc(cols * sizeof(FLOAT));
FLOAT *y = (FLOAT*)numa_alloc(rows * sizeof(FLOAT));
for (i = 0; i < cols; i++)
{
x[i] = 1.0;
}
// warm up
spmv_blks(&blk_h, x, y, NULL);
printf("Notify: begin csr spmv.\n");
clock_gettime(CLOCK_MONOTONIC_RAW, &start);
for (i = 0; i < LOOP_TIME; i++)
{
spmv_blks(&blk_h, x, y, NULL);
}
clock_gettime(CLOCK_MONOTONIC_RAW, &end);
double time = get_sec(&start, &end) / LOOP_TIME;
double gflops = 2.0 * non_zeros / time * 1e-9;
printf("Notify: blk spmv time = %lfs, perf = %lf GFLOPS.\n", time, gflops);
// result_file(y, rows);
return 0;
}
