1 | #include "rq_bench.hfa"
|
---|
2 |
|
---|
3 | thread __attribute__((aligned(128))) Yielder {
|
---|
4 | unsigned long long count;
|
---|
5 | };
|
---|
6 | void ?{}( Yielder & this ) {
|
---|
7 | ((thread&)this){ "Yielder Thread", bench_cluster };
|
---|
8 | this.count = 0;
|
---|
9 | }
|
---|
10 |
|
---|
11 | void main( Yielder & this ) {
|
---|
12 | park();
|
---|
13 | for() {
|
---|
14 | yield();
|
---|
15 | this.count++;
|
---|
16 | if( clock_mode && stop) break;
|
---|
17 | if(!clock_mode && this.count >= stop_count) break;
|
---|
18 | }
|
---|
19 |
|
---|
20 | __atomic_fetch_add(&threads_left, -1, __ATOMIC_SEQ_CST);
|
---|
21 | }
|
---|
22 |
|
---|
23 | int main(int argc, char * argv[]) {
|
---|
24 | cfa_option opt[] = {
|
---|
25 | BENCH_OPT
|
---|
26 | };
|
---|
27 | BENCH_OPT_PARSE("cforall yield benchmark");
|
---|
28 |
|
---|
29 | {
|
---|
30 | unsigned long long global_counter = 0;
|
---|
31 |
|
---|
32 | Time start, end;
|
---|
33 | BenchCluster bc = { nprocs };
|
---|
34 | {
|
---|
35 | threads_left = nthreads;
|
---|
36 | Yielder * threads = alloc(nthreads);
|
---|
37 | for(i; nthreads) {
|
---|
38 | (threads[i]){};
|
---|
39 | }
|
---|
40 |
|
---|
41 | printf("Starting\n");
|
---|
42 |
|
---|
43 | bool is_tty = isatty(STDOUT_FILENO);
|
---|
44 | start = timeHiRes();
|
---|
45 |
|
---|
46 | for(i; nthreads) {
|
---|
47 | unpark( threads[i] );
|
---|
48 | }
|
---|
49 | wait(start, is_tty);
|
---|
50 |
|
---|
51 | stop = true;
|
---|
52 | end = timeHiRes();
|
---|
53 | printf("\nDone\n");
|
---|
54 |
|
---|
55 | for(i; nthreads) {
|
---|
56 | Yielder & y = join( threads[i] );
|
---|
57 | global_counter += y.count;
|
---|
58 | ^(threads[i]){};
|
---|
59 | }
|
---|
60 | free(threads);
|
---|
61 | }
|
---|
62 |
|
---|
63 | printf("Duration (ms) : %'ld\n", (end - start)`dms);
|
---|
64 | printf("Number of processors : %'d\n", nprocs);
|
---|
65 | printf("Number of threads : %'d\n", nthreads);
|
---|
66 | printf("Total Operations(ops): %'15llu\n", global_counter);
|
---|
67 | printf("Ops per second : %'18.2lf\n", ((double)global_counter) / (end - start)`s);
|
---|
68 | printf("ns per ops : %'18.2lf\n", (end - start)`dns / global_counter);
|
---|
69 | printf("Ops per threads : %'15llu\n", global_counter / nthreads);
|
---|
70 | printf("Ops per procs : %'15llu\n", global_counter / nprocs);
|
---|
71 | printf("Ops/sec/procs : %'18.2lf\n", (((double)global_counter) / nprocs) / (end - start)`s);
|
---|
72 | printf("ns per ops/procs : %'18.2lf\n", (end - start)`dns / (global_counter / nprocs));
|
---|
73 | fflush(stdout);
|
---|
74 | }
|
---|
75 | }
|
---|