[29185fc] | 1 | #include "thrdlib/thread.hpp" |
---|
[73fbe91] | 2 | |
---|
| 3 | #include <cassert> |
---|
| 4 | |
---|
| 5 | #include <algorithm> |
---|
| 6 | #include <atomic> |
---|
[d035cf7] | 7 | #include <iostream> |
---|
[73fbe91] | 8 | #include <memory> |
---|
| 9 | #include <vector> |
---|
| 10 | |
---|
[6750bcd] | 11 | #include <getopt.h> |
---|
[29185fc] | 12 | using thrdlib::thread_t; |
---|
| 13 | |
---|
[6750bcd] | 14 | |
---|
[4e39f51] | 15 | extern __attribute__((aligned(128))) thread_local struct { |
---|
| 16 | void * volatile this_thread; |
---|
| 17 | void * volatile this_processor; |
---|
| 18 | void * volatile this_stats; |
---|
| 19 | |
---|
| 20 | struct { |
---|
| 21 | volatile unsigned short disable_count; |
---|
| 22 | volatile bool enabled; |
---|
| 23 | volatile bool in_progress; |
---|
| 24 | } preemption_state; |
---|
| 25 | |
---|
| 26 | #if defined(__SIZEOF_INT128__) |
---|
| 27 | __uint128_t rand_seed; |
---|
| 28 | #else |
---|
| 29 | uint64_t rand_seed; |
---|
| 30 | #endif |
---|
| 31 | struct { |
---|
| 32 | uint64_t fwd_seed; |
---|
| 33 | uint64_t bck_seed; |
---|
| 34 | } ready_rng; |
---|
| 35 | } kernelTLS __attribute__ ((tls_model ( "initial-exec" ))); |
---|
| 36 | |
---|
[73fbe91] | 37 | //-------------------- |
---|
| 38 | // Constants |
---|
| 39 | unsigned nframes; |
---|
| 40 | unsigned fsize; |
---|
| 41 | unsigned nproduce; |
---|
| 42 | |
---|
| 43 | //-------------------- |
---|
| 44 | // Frame management |
---|
| 45 | |
---|
| 46 | class Frame { |
---|
| 47 | static const thread_t reset; |
---|
| 48 | static const thread_t set; |
---|
| 49 | std::atomic<thread_t> rdy_state = { reset }; |
---|
| 50 | std::atomic<thread_t> rnd_state = { set }; |
---|
| 51 | public: |
---|
| 52 | unsigned number; |
---|
| 53 | std::unique_ptr<unsigned char[]> data; |
---|
| 54 | |
---|
| 55 | private: |
---|
| 56 | inline bool wait( thread_t self, std::atomic<thread_t> & state, std::atomic<thread_t> & other ) { |
---|
| 57 | bool ret; |
---|
| 58 | while(true) { |
---|
| 59 | thread_t expected = state; |
---|
| 60 | if( expected == set ) { ret = false; goto END; } |
---|
| 61 | assert( expected == reset ); |
---|
| 62 | if( std::atomic_compare_exchange_strong( &state, &expected, self) ) { |
---|
[29185fc] | 63 | thrdlib::park( self ); |
---|
[73fbe91] | 64 | ret = true; |
---|
| 65 | goto END; |
---|
| 66 | } |
---|
| 67 | } |
---|
| 68 | END: |
---|
| 69 | assert( state == set ); |
---|
| 70 | assert( other != set ); |
---|
| 71 | state = reset; |
---|
| 72 | return ret; |
---|
| 73 | } |
---|
| 74 | |
---|
| 75 | inline bool publish( std::atomic<thread_t> & state ) { |
---|
| 76 | thread_t got = std::atomic_exchange( &state, set ); |
---|
| 77 | assert( got != set ); |
---|
| 78 | |
---|
| 79 | if( got == reset ) return false; |
---|
| 80 | |
---|
[29185fc] | 81 | thrdlib::unpark( got ); |
---|
[73fbe91] | 82 | return true; |
---|
| 83 | } |
---|
| 84 | |
---|
| 85 | public: |
---|
| 86 | inline bool wait_rendered( thread_t self ) { |
---|
| 87 | return wait( self, rnd_state, rdy_state ); |
---|
| 88 | } |
---|
| 89 | |
---|
| 90 | inline bool wait_ready ( thread_t self ) { |
---|
| 91 | return wait( self, rdy_state, rnd_state ); |
---|
| 92 | } |
---|
| 93 | |
---|
| 94 | inline bool publish() { |
---|
| 95 | return publish( rdy_state ); |
---|
| 96 | } |
---|
| 97 | |
---|
| 98 | inline bool release() { |
---|
| 99 | return publish( rnd_state ); |
---|
| 100 | } |
---|
| 101 | }; |
---|
| 102 | |
---|
| 103 | const thread_t Frame::reset = nullptr; |
---|
| 104 | const thread_t Frame::set = reinterpret_cast<thread_t>(1); |
---|
| 105 | |
---|
| 106 | std::unique_ptr<Frame[]> frames; |
---|
| 107 | volatile unsigned last_produced = 0; |
---|
| 108 | |
---|
| 109 | //-------------------- |
---|
| 110 | // Threads |
---|
| 111 | thread_t volatile the_stats_thread = nullptr; |
---|
| 112 | |
---|
| 113 | inline void fence(void) { |
---|
| 114 | std::atomic_thread_fence(std::memory_order_seq_cst); |
---|
| 115 | } |
---|
| 116 | |
---|
| 117 | struct { |
---|
| 118 | struct { |
---|
| 119 | volatile unsigned long long parks = 0; |
---|
| 120 | volatile unsigned long long unparks = 0; |
---|
| 121 | } sim; |
---|
| 122 | struct { |
---|
| 123 | volatile unsigned long long parks = 0; |
---|
| 124 | volatile unsigned long long unparks = 0; |
---|
| 125 | } rend; |
---|
| 126 | |
---|
| 127 | struct { |
---|
| 128 | volatile unsigned long long ran = 0; |
---|
| 129 | volatile unsigned long long saw = 0; |
---|
| 130 | } stats; |
---|
| 131 | } thrd_stats; |
---|
| 132 | |
---|
| 133 | void Stats( thread_t self ) { |
---|
| 134 | the_stats_thread = self; |
---|
| 135 | fence(); |
---|
[29185fc] | 136 | thrdlib::park( self ); |
---|
[73fbe91] | 137 | |
---|
| 138 | std::vector<bool> seen; |
---|
| 139 | seen.resize(nproduce, false); |
---|
| 140 | |
---|
| 141 | while(last_produced < nproduce) { |
---|
[29185fc] | 142 | thrdlib::yield(); |
---|
[73fbe91] | 143 | thrd_stats.stats.ran++; |
---|
| 144 | if( last_produced > 0 ) seen.at(last_produced - 1) = true; |
---|
| 145 | } |
---|
| 146 | |
---|
| 147 | thrd_stats.stats.saw = std::count(seen.begin(), seen.end(), true); |
---|
| 148 | } |
---|
| 149 | |
---|
| 150 | void Simulator( thread_t self ) { |
---|
| 151 | for(unsigned i = 0; i < nproduce; i++) { |
---|
| 152 | auto & frame = frames[i % nframes]; |
---|
| 153 | // Wait for the frames to be rendered |
---|
| 154 | if( frame.wait_rendered( self ) ) { |
---|
| 155 | thrd_stats.sim.parks++; |
---|
| 156 | } |
---|
| 157 | |
---|
| 158 | // Write the frame information |
---|
| 159 | frame.number = i; |
---|
| 160 | for( unsigned x = 0; x < fsize; x++ ) { |
---|
| 161 | frame.data[x] = i; |
---|
| 162 | } |
---|
| 163 | std::cout << "Simulated " << i << std::endl; |
---|
| 164 | last_produced = i+1; |
---|
| 165 | |
---|
| 166 | // Publish it |
---|
| 167 | if( frame.publish() ) { |
---|
| 168 | thrd_stats.sim.unparks++; |
---|
| 169 | } |
---|
| 170 | } |
---|
| 171 | } |
---|
| 172 | |
---|
| 173 | void Renderer( thread_t self ) { |
---|
[29185fc] | 174 | thrdlib::unpark( the_stats_thread ); |
---|
[73fbe91] | 175 | for(unsigned i = 0; i < nproduce; i++) { |
---|
| 176 | auto & frame = frames[i % nframes]; |
---|
| 177 | // Wait for the frames to be ready |
---|
| 178 | if( frame.wait_ready( self ) ) { |
---|
| 179 | thrd_stats.rend.parks++; |
---|
| 180 | } |
---|
| 181 | |
---|
| 182 | // Render the frame |
---|
| 183 | unsigned total = 0; |
---|
| 184 | for( unsigned x = 0; x < fsize; x++ ) { |
---|
| 185 | total += frame.data[x]; |
---|
| 186 | } |
---|
| 187 | |
---|
| 188 | std::cout << "Rendered " << i << std::endl; |
---|
| 189 | assert(total == i * fsize); |
---|
| 190 | |
---|
| 191 | // Release |
---|
| 192 | if( frame.release() ) { |
---|
| 193 | thrd_stats.rend.unparks++; |
---|
| 194 | } |
---|
| 195 | } |
---|
| 196 | |
---|
| 197 | } |
---|
| 198 | |
---|
| 199 | |
---|
| 200 | |
---|
[6750bcd] | 201 | int main(int argc, char * argv[]) { |
---|
[73fbe91] | 202 | nframes = 3; |
---|
| 203 | fsize = 1000; |
---|
| 204 | nproduce = 60; |
---|
| 205 | |
---|
[4e39f51] | 206 | const char * framework; |
---|
| 207 | |
---|
[6750bcd] | 208 | for(;;) { |
---|
| 209 | static struct option options[] = { |
---|
| 210 | {"buff", required_argument, 0, 'b'}, |
---|
| 211 | {"nprod", required_argument, 0, 'p'}, |
---|
| 212 | {"fsize", required_argument, 0, 'f'}, |
---|
| 213 | {0, 0, 0, 0} |
---|
| 214 | }; |
---|
| 215 | |
---|
| 216 | int idx = 0; |
---|
| 217 | int opt = getopt_long(argc, argv, "b:p:f:", options, &idx); |
---|
| 218 | |
---|
| 219 | std::string arg = optarg ? optarg : ""; |
---|
| 220 | size_t len = 0; |
---|
| 221 | switch(opt) { |
---|
| 222 | // Exit Case |
---|
| 223 | case -1: |
---|
| 224 | /* paranoid */ assert(optind <= argc); |
---|
[4e39f51] | 225 | if( optind == argc ) { |
---|
| 226 | std::cerr << "Must specify a framework" << std::endl; |
---|
| 227 | goto usage; |
---|
| 228 | |
---|
| 229 | } |
---|
| 230 | framework = argv[optind]; |
---|
[6750bcd] | 231 | goto run; |
---|
| 232 | case 'b': |
---|
| 233 | try { |
---|
| 234 | nframes = std::stoul(optarg, &len); |
---|
| 235 | if(nframes == 0 || len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 236 | } catch(std::invalid_argument &) { |
---|
| 237 | std::cerr << "Number of buffered frames must be at least 1, was" << arg << std::endl; |
---|
| 238 | goto usage; |
---|
| 239 | } |
---|
| 240 | break; |
---|
| 241 | case 'p': |
---|
| 242 | try { |
---|
| 243 | nproduce = std::stoul(optarg, &len); |
---|
| 244 | if(nproduce == 0 || len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 245 | } catch(std::invalid_argument &) { |
---|
| 246 | std::cerr << "Number of produced frames must be at least 1, was" << arg << std::endl; |
---|
| 247 | goto usage; |
---|
| 248 | } |
---|
| 249 | break; |
---|
| 250 | case 'f': |
---|
| 251 | try { |
---|
| 252 | fsize = std::stoul(optarg, &len); |
---|
| 253 | if(fsize == 0 || len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 254 | } catch(std::invalid_argument &) { |
---|
| 255 | std::cerr << "Size of produced frames must be at least 1, was" << arg << std::endl; |
---|
| 256 | goto usage; |
---|
| 257 | } |
---|
| 258 | break; |
---|
| 259 | // Other cases |
---|
| 260 | default: /* ? */ |
---|
| 261 | std::cerr << opt << std::endl; |
---|
| 262 | usage: |
---|
[4e39f51] | 263 | std::cerr << "Usage: " << argv[0] << " [options] framework" << std::endl; |
---|
[6750bcd] | 264 | std::cerr << std::endl; |
---|
| 265 | std::cerr << " -b, --buff=COUNT Number of frames to buffer" << std::endl; |
---|
| 266 | std::cerr << " -p, --nprod=COUNT Number of frames to produce" << std::endl; |
---|
| 267 | std::cerr << " -f, --fsize=SIZE Size of each frame in bytes" << std::endl; |
---|
| 268 | std::exit(1); |
---|
| 269 | } |
---|
| 270 | } |
---|
| 271 | run: |
---|
[4e39f51] | 272 | assert( framework ); |
---|
[6750bcd] | 273 | |
---|
[73fbe91] | 274 | frames.reset(new Frame[nframes]); |
---|
| 275 | for(unsigned i = 0; i < nframes; i++) { |
---|
| 276 | frames[i].number = 0; |
---|
| 277 | frames[i].data.reset(new unsigned char[fsize]); |
---|
| 278 | } |
---|
[6750bcd] | 279 | std::cout << "Created frames of " << fsize << " bytes" << std::endl; |
---|
| 280 | std::cout << "(Buffering " << nframes << ")" << std::endl; |
---|
[73fbe91] | 281 | |
---|
[4e39f51] | 282 | thrdlib::init( framework, 2 ); |
---|
[73fbe91] | 283 | |
---|
[29185fc] | 284 | thread_t stats = thrdlib::create( Stats ); |
---|
[73fbe91] | 285 | std::cout << "Created Stats Thread" << std::endl; |
---|
[29185fc] | 286 | while( the_stats_thread == nullptr ) thrdlib::yield(); |
---|
| 287 | |
---|
[73fbe91] | 288 | std::cout << "Creating Main Threads" << std::endl; |
---|
[29185fc] | 289 | thread_t renderer = thrdlib::create( Renderer ); |
---|
| 290 | thread_t simulator = thrdlib::create( Simulator ); |
---|
[73fbe91] | 291 | |
---|
| 292 | std::cout << "Running" << std::endl; |
---|
| 293 | |
---|
[29185fc] | 294 | thrdlib::join( simulator ); |
---|
| 295 | thrdlib::join( renderer ); |
---|
| 296 | thrdlib::join( stats ); |
---|
[73fbe91] | 297 | |
---|
[29185fc] | 298 | thrdlib::clean(); |
---|
[d035cf7] | 299 | |
---|
[73fbe91] | 300 | std::cout << "----------" << std::endl; |
---|
| 301 | std::cout << "# Parks" << std::endl; |
---|
| 302 | std::cout << " Renderer park: " << thrd_stats. sim. parks << std::endl; |
---|
| 303 | std::cout << " Renderer unpark: " << thrd_stats. sim.unparks << std::endl; |
---|
| 304 | std::cout << " Simulator park: " << thrd_stats.rend. parks << std::endl; |
---|
| 305 | std::cout << " Simulator unpark: " << thrd_stats.rend.unparks << std::endl; |
---|
| 306 | |
---|
| 307 | std::cout << "Stats thread" << std::endl; |
---|
| 308 | std::cout << " Ran : " << thrd_stats.stats.ran << " times" << std::endl; |
---|
| 309 | std::cout << " Saw : " << thrd_stats.stats.saw << " (" << ((100.f * thrd_stats.stats.saw) / nproduce) << "%)" << std::endl; |
---|
| 310 | } |
---|