| 1 | #include "rq_bench.hfa"
|
|---|
| 2 |
|
|---|
| 3 | thread Partner {
|
|---|
| 4 | Partner * partner;
|
|---|
| 5 | unsigned long long count;
|
|---|
| 6 | bool first;
|
|---|
| 7 | };
|
|---|
| 8 |
|
|---|
| 9 | void ?{}( Partner & this ) {
|
|---|
| 10 | ((thread&)this){ bench_cluster };
|
|---|
| 11 | }
|
|---|
| 12 |
|
|---|
| 13 | void main( Partner & this ) {
|
|---|
| 14 | thread_loop {
|
|---|
| 15 | park();
|
|---|
| 16 | unpark( *this.partner );
|
|---|
| 17 | }
|
|---|
| 18 |
|
|---|
| 19 | __atomic_fetch_add(&threads_left, -1, __ATOMIC_SEQ_CST);
|
|---|
| 20 |
|
|---|
| 21 | if(this.first) park();
|
|---|
| 22 | }
|
|---|
| 23 |
|
|---|
| 24 | int main(int argc, char * argv[]) {
|
|---|
| 25 | unsigned ring_size = 2;
|
|---|
| 26 | cfa_option opt[] = {
|
|---|
| 27 | BENCH_OPT,
|
|---|
| 28 | { 'r', "ringsize", "Number of threads in a cycle", ring_size }
|
|---|
| 29 | };
|
|---|
| 30 | BENCH_OPT_PARSE("cforall cycle benchmark");
|
|---|
| 31 |
|
|---|
| 32 | {
|
|---|
| 33 | unsigned long long global_counter = 0;
|
|---|
| 34 | unsigned tthreads = nthreads * ring_size;
|
|---|
| 35 | Time start, end;
|
|---|
| 36 | BenchCluster bc = { nprocs };
|
|---|
| 37 | {
|
|---|
| 38 | threads_left = tthreads;
|
|---|
| 39 | Partner threads[tthreads];
|
|---|
| 40 | for(i; tthreads) {
|
|---|
| 41 | unsigned pi = (i + nthreads) % tthreads;
|
|---|
| 42 | threads[i].partner = &threads[pi];
|
|---|
| 43 | threads[i].first = i < nthreads;
|
|---|
| 44 | }
|
|---|
| 45 | printf("Starting\n");
|
|---|
| 46 |
|
|---|
| 47 | bool is_tty = isatty(STDOUT_FILENO);
|
|---|
| 48 | start = getTimeNsec();
|
|---|
| 49 |
|
|---|
| 50 | for(i; nthreads) {
|
|---|
| 51 | unpark( threads[i] );
|
|---|
| 52 | }
|
|---|
| 53 | wait(start, end, is_tty);
|
|---|
| 54 |
|
|---|
| 55 | stop = true;
|
|---|
| 56 | end = getTimeNsec();
|
|---|
| 57 | printf("\nDone\n");
|
|---|
| 58 |
|
|---|
| 59 | for(i; tthreads) {
|
|---|
| 60 | global_counter += join( threads[i] ).count;
|
|---|
| 61 | }
|
|---|
| 62 | }
|
|---|
| 63 |
|
|---|
| 64 | printf("Took %'ld ms\n", (end - start)`ms);
|
|---|
| 65 | printf("Yields per second : %'18.2lf\n", ((double)global_counter) / (end - start)`s);
|
|---|
| 66 | printf("ns per yields : %'18.2lf\n", ((double)(end - start)`ns) / global_counter);
|
|---|
| 67 | printf("Total yields : %'15llu\n", global_counter);
|
|---|
| 68 | printf("Yields per threads : %'15llu\n", global_counter / tthreads);
|
|---|
| 69 | printf("Yields per procs : %'15llu\n", global_counter / nprocs);
|
|---|
| 70 | printf("Yields/sec/procs : %'18.2lf\n", (((double)global_counter) / nprocs) / (end - start)`s);
|
|---|
| 71 | printf("ns per yields/procs : %'18.2lf\n", ((double)(end - start)`ns) / (global_counter / nprocs));
|
|---|
| 72 | fflush(stdout);
|
|---|
| 73 | }
|
|---|
| 74 |
|
|---|
| 75 | return 0;
|
|---|
| 76 | }
|
|---|