[2649ff9] | 1 | #include <stdlib.h> |
---|
| 2 | #include <stdio.h> |
---|
| 3 | #include <string.h> |
---|
| 4 | #include <limits.h> |
---|
| 5 | |
---|
| 6 | extern "C" { |
---|
| 7 | #include <locale.h> |
---|
| 8 | #include <getopt.h> |
---|
| 9 | } |
---|
| 10 | |
---|
| 11 | #include <unistd.h> |
---|
| 12 | |
---|
| 13 | #include <clock.hfa> |
---|
| 14 | #include <time.hfa> |
---|
[566fde0] | 15 | #include <stats.hfa> |
---|
[2649ff9] | 16 | |
---|
| 17 | #include "../benchcltr.hfa" |
---|
| 18 | |
---|
| 19 | extern bool traceHeapOn(); |
---|
| 20 | |
---|
| 21 | |
---|
| 22 | volatile bool run = false; |
---|
| 23 | volatile unsigned long long global_counter; |
---|
| 24 | |
---|
[04b5cef] | 25 | thread __attribute__((aligned(128))) Yielder { |
---|
[2649ff9] | 26 | unsigned long long counter; |
---|
| 27 | }; |
---|
| 28 | void ?{}( Yielder & this ) { |
---|
| 29 | this.counter = 0; |
---|
| 30 | ((thread&)this){ "Yielder Thread", *the_benchmark_cluster }; |
---|
| 31 | } |
---|
| 32 | |
---|
| 33 | void main( Yielder & this ) { |
---|
[e235429] | 34 | park(); |
---|
[2649ff9] | 35 | /* paranoid */ assert( true == __atomic_load_n(&run, __ATOMIC_RELAXED) ); |
---|
| 36 | |
---|
| 37 | while(__atomic_load_n(&run, __ATOMIC_RELAXED)) { |
---|
| 38 | yield(); |
---|
| 39 | this.counter++; |
---|
| 40 | } |
---|
| 41 | __atomic_fetch_add(&global_counter, this.counter, __ATOMIC_SEQ_CST); |
---|
| 42 | } |
---|
| 43 | |
---|
| 44 | int main(int argc, char * argv[]) { |
---|
[01c6256] | 45 | unsigned num_io = 1; |
---|
| 46 | io_context_params params; |
---|
[2649ff9] | 47 | |
---|
[5bcdc8c] | 48 | cfa_option opt[] = { |
---|
| 49 | BENCH_OPT_CFA |
---|
| 50 | }; |
---|
| 51 | int opt_cnt = sizeof(opt) / sizeof(cfa_option); |
---|
[2649ff9] | 52 | |
---|
[5bcdc8c] | 53 | char **left; |
---|
| 54 | parse_args( argc, argv, opt, opt_cnt, "[OPTIONS]...\ncforall yield benchmark", left ); |
---|
[2649ff9] | 55 | |
---|
| 56 | { |
---|
[cb85603] | 57 | printf("Running %d threads on %d processors for %f seconds\n", nthreads, nprocs, duration); |
---|
[2649ff9] | 58 | |
---|
| 59 | Time start, end; |
---|
[01c6256] | 60 | BenchCluster cl = { num_io, params, CFA_STATS_READY_Q }; |
---|
[2649ff9] | 61 | { |
---|
| 62 | BenchProc procs[nprocs]; |
---|
| 63 | { |
---|
| 64 | Yielder threads[nthreads]; |
---|
| 65 | printf("Starting\n"); |
---|
[9791ab5] | 66 | |
---|
| 67 | bool is_tty = isatty(STDOUT_FILENO); |
---|
[e54d0c3] | 68 | start = timeHiRes(); |
---|
[2649ff9] | 69 | run = true; |
---|
| 70 | |
---|
| 71 | for(i; nthreads) { |
---|
[e235429] | 72 | unpark( threads[i] ); |
---|
[2649ff9] | 73 | } |
---|
[9791ab5] | 74 | wait(duration, start, end, is_tty); |
---|
[2649ff9] | 75 | |
---|
| 76 | run = false; |
---|
[e54d0c3] | 77 | end = timeHiRes(); |
---|
[9791ab5] | 78 | printf("\nDone\n"); |
---|
[2649ff9] | 79 | } |
---|
| 80 | } |
---|
| 81 | |
---|
[6dc2db9] | 82 | printf("Duration (ms) : %'ld\n", (end - start)`dms); |
---|
[2c7d0e8] | 83 | printf("Number of processors: %'d\n", nprocs); |
---|
| 84 | printf("Number of threads : %'d\n", nthreads); |
---|
| 85 | printf("Total yields : %'15llu\n", global_counter); |
---|
[564148f] | 86 | printf("Yields per second : %'18.2lf\n", ((double)global_counter) / (end - start)`s); |
---|
| 87 | printf("ns per yields : %'18.2lf\n", ((double)(end - start)`ns) / global_counter); |
---|
[2073d207] | 88 | printf("Yields per procs : %'15llu\n", global_counter / nprocs); |
---|
| 89 | printf("Yields/sec/procs : %'18.2lf\n", (((double)global_counter) / nprocs) / (end - start)`s); |
---|
| 90 | printf("ns per yields/procs : %'18.2lf\n", ((double)(end - start)`ns) / (global_counter / nprocs)); |
---|
[b813f53] | 91 | fflush(stdout); |
---|
[2649ff9] | 92 | } |
---|
[e54d0c3] | 93 | } |
---|