1 | #include <clock.hfa>
|
---|
2 | #include <kernel.hfa>
|
---|
3 | #include <parseargs.hfa>
|
---|
4 | #include <stdio.h>
|
---|
5 | #include <stdlib.hfa>
|
---|
6 | #include <stats.hfa>
|
---|
7 | #include <thread.hfa>
|
---|
8 | #include <time.hfa>
|
---|
9 | #include <unistd.h>
|
---|
10 |
|
---|
11 | volatile bool stop = false;
|
---|
12 | bool clock_mode;
|
---|
13 | double duration = -1;
|
---|
14 | unsigned long long stop_count = 0;
|
---|
15 | unsigned nprocs = 1;
|
---|
16 | unsigned nthreads = 1;
|
---|
17 |
|
---|
18 | volatile unsigned long long threads_left;
|
---|
19 |
|
---|
20 | #define thread_loop for(this.count = 0;; this.count++)
|
---|
21 |
|
---|
22 | #define BENCH_OPT \
|
---|
23 | {'d', "duration", "Duration of the experiments in seconds", duration }, \
|
---|
24 | {'i', "iterations", "Number of iterations of the experiments", stop_count }, \
|
---|
25 | {'t', "nthreads", "Number of threads to use", nthreads }, \
|
---|
26 | {'p', "nprocs", "Number of processors to use", nprocs }
|
---|
27 |
|
---|
28 | #define BENCH_OPT_PARSE(name) \
|
---|
29 | { \
|
---|
30 | int opt_cnt = sizeof(opt) / sizeof(cfa_option); \
|
---|
31 | char **left; \
|
---|
32 | parse_args( argc, argv, opt, opt_cnt, "[OPTIONS]...\n" name, left ); \
|
---|
33 | if(duration > 0 && stop_count > 0) { \
|
---|
34 | fprintf(stderr, "--duration and --iterations cannot be used together\n"); \
|
---|
35 | print_args_usage(argc, argv, opt, opt_cnt, "[OPTIONS]...\n" name, true); \
|
---|
36 | } else if(duration > 0) { \
|
---|
37 | clock_mode = true; \
|
---|
38 | stop_count = 0xFFFFFFFFFFFFFFFF; \
|
---|
39 | printf("Running for %lf seconds\n", duration); \
|
---|
40 | } else if(stop_count > 0) { \
|
---|
41 | clock_mode = false; \
|
---|
42 | printf("Running for %llu iterations\n", stop_count); \
|
---|
43 | } else { \
|
---|
44 | duration = 5; clock_mode = true;\
|
---|
45 | printf("Running for %lf seconds\n", duration); \
|
---|
46 | } \
|
---|
47 | }
|
---|
48 |
|
---|
49 | struct cluster & bench_cluster;
|
---|
50 |
|
---|
51 | struct BenchCluster {
|
---|
52 | cluster cl;
|
---|
53 | processor * procs;
|
---|
54 | unsigned nprocs;
|
---|
55 | };
|
---|
56 |
|
---|
57 | void ?{}( BenchCluster & this, unsigned nprocs ) {
|
---|
58 | (this.cl){ "Benchmark Cluster" };
|
---|
59 | &bench_cluster = &this.cl;
|
---|
60 | this.nprocs = nprocs;
|
---|
61 | this.procs = alloc( this.nprocs );
|
---|
62 | for(i; this.nprocs){
|
---|
63 | processor * p = &this.procs[i];
|
---|
64 | (*p){ "Benchmark Processor", this.cl };
|
---|
65 | }
|
---|
66 | #if !defined(__CFA_NO_STATISTICS__)
|
---|
67 | print_stats_at_exit( this.cl, CFA_STATS_READY_Q );
|
---|
68 | #endif
|
---|
69 | }
|
---|
70 |
|
---|
71 | void ^?{}( BenchCluster & this ) {
|
---|
72 | adelete( this.procs );
|
---|
73 | ^(this.cl){};
|
---|
74 | }
|
---|
75 |
|
---|
76 | void wait(const Time & start, bool is_tty) {
|
---|
77 | for() {
|
---|
78 | sleep(100`ms);
|
---|
79 | Time end = timeHiRes();
|
---|
80 | Duration delta = end - start;
|
---|
81 | if(is_tty) {
|
---|
82 | printf(" %.1f\r", delta`ds);
|
---|
83 | fflush(stdout);
|
---|
84 | }
|
---|
85 | if( clock_mode && delta >= duration`s ) {
|
---|
86 | break;
|
---|
87 | }
|
---|
88 | else if( !clock_mode && threads_left == 0 ) {
|
---|
89 | break;
|
---|
90 | }
|
---|
91 | }
|
---|
92 | }
|
---|
93 |
|
---|
94 | struct __attribute__((aligned(128))) bench_sem {
|
---|
95 | struct thread$ * volatile ptr;
|
---|
96 | };
|
---|
97 |
|
---|
98 | static inline {
|
---|
99 | void ?{}(bench_sem & this) {
|
---|
100 | this.ptr = 0p;
|
---|
101 | }
|
---|
102 |
|
---|
103 | void ^?{}(bench_sem & this) {}
|
---|
104 |
|
---|
105 | bool wait(bench_sem & this) {
|
---|
106 | for() {
|
---|
107 | struct thread$ * expected = this.ptr;
|
---|
108 | if(expected == 1p) {
|
---|
109 | if(__atomic_compare_exchange_n(&this.ptr, &expected, 0p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
|
---|
110 | return false;
|
---|
111 | }
|
---|
112 | }
|
---|
113 | else {
|
---|
114 | /* paranoid */ verify( expected == 0p );
|
---|
115 | if(__atomic_compare_exchange_n(&this.ptr, &expected, active_thread(), false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
|
---|
116 | park();
|
---|
117 | return true;
|
---|
118 | }
|
---|
119 | }
|
---|
120 |
|
---|
121 | }
|
---|
122 | }
|
---|
123 |
|
---|
124 | bool post(bench_sem & this) {
|
---|
125 | for() {
|
---|
126 | struct thread$ * expected = this.ptr;
|
---|
127 | if(expected == 1p) return false;
|
---|
128 | if(expected == 0p) {
|
---|
129 | if(__atomic_compare_exchange_n(&this.ptr, &expected, 1p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
|
---|
130 | return false;
|
---|
131 | }
|
---|
132 | }
|
---|
133 | else {
|
---|
134 | if(__atomic_compare_exchange_n(&this.ptr, &expected, 0p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
|
---|
135 | unpark( expected );
|
---|
136 | return true;
|
---|
137 | }
|
---|
138 | }
|
---|
139 | }
|
---|
140 | }
|
---|
141 | }
|
---|