1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
|
#include <pthread.h>
#include <stdlib.h>
#include <limits.h>
/************
* kt_for() *
************/
struct kt_for_t;
typedef struct {
struct kt_for_t *t;
long i;
} ktf_worker_t;
typedef struct kt_for_t {
int n_threads;
long n;
ktf_worker_t *w;
void (*func)(void*,long,int);
void *data;
} kt_for_t;
static inline long steal_work(kt_for_t *t)
{
int i, min_i = -1;
long k, min = LONG_MAX;
for (i = 0; i < t->n_threads; ++i)
if (min > t->w[i].i) min = t->w[i].i, min_i = i;
k = __sync_fetch_and_add(&t->w[min_i].i, t->n_threads);
return k >= t->n? -1 : k;
}
static void *ktf_worker(void *data)
{
ktf_worker_t *w = (ktf_worker_t*)data;
long i;
for (;;) {
i = __sync_fetch_and_add(&w->i, w->t->n_threads);
if (i >= w->t->n) break;
w->t->func(w->t->data, i, w - w->t->w);
}
while ((i = steal_work(w->t)) >= 0)
w->t->func(w->t->data, i, w - w->t->w);
pthread_exit(0);
}
void kt_for(int n_threads, void (*func)(void*,long,int), void *data, long n)
{
int i;
kt_for_t t;
pthread_t *tid;
t.func = func, t.data = data, t.n_threads = n_threads, t.n = n;
t.w = (ktf_worker_t*)alloca(n_threads * sizeof(ktf_worker_t));
tid = (pthread_t*)alloca(n_threads * sizeof(pthread_t));
for (i = 0; i < n_threads; ++i)
t.w[i].t = &t, t.w[i].i = i;
for (i = 0; i < n_threads; ++i) pthread_create(&tid[i], 0, ktf_worker, &t.w[i]);
for (i = 0; i < n_threads; ++i) pthread_join(tid[i], 0);
}
/*****************
* kt_pipeline() *
*****************/
struct ktp_t;
typedef struct {
struct ktp_t *pl;
int64_t index;
int step;
void *data;
} ktp_worker_t;
typedef struct ktp_t {
void *shared;
void *(*func)(void*, int, void*);
int64_t index;
int n_workers, n_steps;
ktp_worker_t *workers;
pthread_mutex_t mutex;
pthread_cond_t cv;
} ktp_t;
static void *ktp_worker(void *data)
{
ktp_worker_t *w = (ktp_worker_t*)data;
ktp_t *p = w->pl;
while (w->step < p->n_steps) {
// test whether we can kick off the job with this worker
pthread_mutex_lock(&p->mutex);
for (;;) {
int i;
// test whether another worker is doing the same step
for (i = 0; i < p->n_workers; ++i) {
if (w == &p->workers[i]) continue; // ignore itself
if (p->workers[i].step <= w->step && p->workers[i].index < w->index)
break;
}
if (i == p->n_workers) break; // no workers with smaller indices are doing w->step or the previous steps
pthread_cond_wait(&p->cv, &p->mutex);
}
pthread_mutex_unlock(&p->mutex);
// working on w->step
w->data = p->func(p->shared, w->step, w->step? w->data : 0); // for the first step, input is NULL
// update step and let other workers know
pthread_mutex_lock(&p->mutex);
w->step = w->step == p->n_steps - 1 || w->data? (w->step + 1) % p->n_steps : p->n_steps;
if (w->step == 0) w->index = p->index++;
pthread_cond_broadcast(&p->cv);
pthread_mutex_unlock(&p->mutex);
}
pthread_exit(0);
}
void kt_pipeline(int n_threads, void *(*func)(void*, int, void*), void *shared_data, int n_steps)
{
ktp_t aux;
pthread_t *tid;
int i;
if (n_threads < 1) n_threads = 1;
aux.n_workers = n_threads;
aux.n_steps = n_steps;
aux.func = func;
aux.shared = shared_data;
aux.index = 0;
pthread_mutex_init(&aux.mutex, 0);
pthread_cond_init(&aux.cv, 0);
aux.workers = (ktp_worker_t*)alloca(n_threads * sizeof(ktp_worker_t));
for (i = 0; i < n_threads; ++i) {
ktp_worker_t *w = &aux.workers[i];
w->step = 0; w->pl = &aux; w->data = 0;
w->index = aux.index++;
}
tid = (pthread_t*)alloca(n_threads * sizeof(pthread_t));
for (i = 0; i < n_threads; ++i) pthread_create(&tid[i], 0, ktp_worker, &aux.workers[i]);
for (i = 0; i < n_threads; ++i) pthread_join(tid[i], 0);
pthread_mutex_destroy(&aux.mutex);
pthread_cond_destroy(&aux.cv);
}
|
