[502] | 1 | ////////////////////////////////////////////////////////////////////////////////// |
---|
[509] | 2 | // File : main.c (for gameoflife) |
---|
| 3 | // Date : November 2013 / February 2015 |
---|
| 4 | // Authors : Alexandre Joannou <alexandre.joannou@lip6.fr> november 2013 |
---|
| 5 | // Alain Greiner <alain.greiner@lip6.fr> february 2015 |
---|
[708] | 6 | ////////////////////////////////////////////////////////////////////////////////// |
---|
[669] | 7 | // This multi-threaded application is an emulation of the Game of Life automaton. |
---|
[502] | 8 | // The world size is defined by the HEIGHT and WIDTH parameters. |
---|
[509] | 9 | // |
---|
[708] | 10 | // There is at most one thread per processor in the platform. |
---|
| 11 | // - If the number of processors is larger than the number of lines, |
---|
| 12 | // the number of threads is equal to the number of lines, and |
---|
| 13 | // each thread process one single line. |
---|
| 14 | // - if the number of processors is not larger than the number of lines, |
---|
| 15 | // the number of threads is equal to the number of processors, and |
---|
| 16 | // each thread process HEIGHT/nthreads (or HEIGHT/nthreads + 1) lines. |
---|
| 17 | // |
---|
| 18 | // Thread running on processor P(0,0,0) execute the main() function, |
---|
| 19 | // that initialises the barrier, the TTY terminal, the CMA controler, |
---|
| 20 | // and launch the other threads, before calling the execute function. |
---|
| 21 | // Other threads are just running the execute() function. |
---|
| 22 | // |
---|
| 23 | // The total number of processors cannot be larger than 1024 = 16 * 16 * 4 |
---|
[502] | 24 | ////////////////////////////////////////////////////////////////////////////////// |
---|
[251] | 25 | |
---|
| 26 | #include "stdio.h" |
---|
| 27 | #include "limits.h" |
---|
[502] | 28 | #include "user_barrier.h" |
---|
[251] | 29 | #include "mapping_info.h" |
---|
[669] | 30 | #include "hard_config.h" |
---|
[708] | 31 | #include "malloc.h" |
---|
[251] | 32 | |
---|
[708] | 33 | #define WIDTH FBUF_X_SIZE |
---|
| 34 | #define HEIGHT FBUF_Y_SIZE |
---|
[251] | 35 | |
---|
[708] | 36 | #define VERBOSE 1 |
---|
[251] | 37 | |
---|
[509] | 38 | typedef unsigned char uint8_t; |
---|
[251] | 39 | |
---|
[708] | 40 | typedef struct |
---|
| 41 | { |
---|
| 42 | unsigned int index; // index of first line to be processed |
---|
| 43 | unsigned int lines; // number of lines to be processed |
---|
| 44 | } arguments_t; |
---|
[295] | 45 | |
---|
[708] | 46 | arguments_t args[1024]; // at most 1024 threads |
---|
[251] | 47 | |
---|
[708] | 48 | uint8_t world[2][HEIGHT][WIDTH] __attribute__((aligned(64))); |
---|
| 49 | |
---|
| 50 | uint8_t display[2][HEIGHT][WIDTH] __attribute__((aligned(64))); |
---|
| 51 | |
---|
[669] | 52 | unsigned int status0[16]; |
---|
| 53 | unsigned int status1[16]; |
---|
| 54 | |
---|
[509] | 55 | giet_sqt_barrier_t barrier; |
---|
[251] | 56 | |
---|
[509] | 57 | //////////////////////////////////// |
---|
| 58 | void init_world( unsigned int phase, |
---|
| 59 | unsigned int base_line, |
---|
| 60 | unsigned int nb_line ) |
---|
[251] | 61 | { |
---|
[509] | 62 | unsigned int x,y; |
---|
| 63 | for (y = base_line ; y < base_line + nb_line ; y++) |
---|
[502] | 64 | { |
---|
[509] | 65 | for(x = 0 ; x < WIDTH ; x++) |
---|
[502] | 66 | { |
---|
[708] | 67 | world[phase][y][x] = (giet_rand() >> (x % 8)) & 0x1; |
---|
[251] | 68 | } |
---|
| 69 | } |
---|
| 70 | } |
---|
| 71 | |
---|
[509] | 72 | ////////////////////////////////////////////////////// |
---|
| 73 | uint8_t number_of_alive_neighbour( unsigned int phase, |
---|
| 74 | unsigned int x, |
---|
| 75 | unsigned int y ) |
---|
[251] | 76 | { |
---|
| 77 | uint8_t nb = 0; |
---|
| 78 | |
---|
[708] | 79 | nb += world[phase][(y - 1) % HEIGHT][(x - 1) % WIDTH]; |
---|
| 80 | nb += world[phase][ y ][(x - 1) % WIDTH]; |
---|
| 81 | nb += world[phase][(y + 1) % HEIGHT][(x - 1) % WIDTH]; |
---|
| 82 | nb += world[phase][(y - 1) % HEIGHT][ x ]; |
---|
| 83 | nb += world[phase][(y + 1) % HEIGHT][ x ]; |
---|
| 84 | nb += world[phase][(y - 1) % HEIGHT][(x + 1) % WIDTH]; |
---|
| 85 | nb += world[phase][ y ][(x + 1) % WIDTH]; |
---|
| 86 | nb += world[phase][(y + 1) % HEIGHT][(x + 1) % WIDTH]; |
---|
[251] | 87 | |
---|
| 88 | return nb; |
---|
| 89 | } |
---|
| 90 | |
---|
[509] | 91 | ///////////////////////////////////////// |
---|
| 92 | uint8_t compute_cell( unsigned int phase, |
---|
| 93 | unsigned int x, |
---|
| 94 | unsigned int y ) |
---|
[251] | 95 | { |
---|
[509] | 96 | uint8_t nb_neighbours_alive = number_of_alive_neighbour( phase, x , y ); |
---|
| 97 | |
---|
[708] | 98 | if (world[phase][y][x] == 1) |
---|
[502] | 99 | { |
---|
| 100 | if (nb_neighbours_alive == 2 || nb_neighbours_alive == 3) return 1; |
---|
[251] | 101 | } |
---|
[502] | 102 | else |
---|
| 103 | { |
---|
| 104 | if (nb_neighbours_alive == 3) return 1; |
---|
[708] | 105 | else return world[phase][y][x]; |
---|
[251] | 106 | } |
---|
| 107 | return 0; |
---|
| 108 | } |
---|
| 109 | |
---|
[509] | 110 | ///////////////////////////////////////// |
---|
| 111 | void compute_new_gen( unsigned int phase, |
---|
| 112 | unsigned int base_line, |
---|
| 113 | unsigned int nb_line ) |
---|
[251] | 114 | { |
---|
[509] | 115 | unsigned int x,y; |
---|
[295] | 116 | for (y = base_line; y < base_line + nb_line; y++) |
---|
| 117 | { |
---|
| 118 | for(x = 0; x < WIDTH ; x++) |
---|
| 119 | { |
---|
[708] | 120 | world[phase][y][x] = compute_cell( 1 - phase , x , y ); |
---|
[251] | 121 | } |
---|
| 122 | } |
---|
| 123 | } |
---|
| 124 | |
---|
[509] | 125 | //////////////////////////////////// |
---|
| 126 | void copy_world( unsigned int phase, |
---|
| 127 | unsigned int base_line, |
---|
| 128 | unsigned int nb_line ) |
---|
[251] | 129 | { |
---|
[509] | 130 | unsigned int x,y; |
---|
[502] | 131 | for (y = base_line; y < base_line + nb_line; y++) |
---|
| 132 | { |
---|
| 133 | for(x = 0; x < WIDTH ; x++) |
---|
| 134 | { |
---|
[708] | 135 | display[phase][y][x] = world[phase][y][x]*255; |
---|
[251] | 136 | } |
---|
| 137 | } |
---|
| 138 | } |
---|
| 139 | |
---|
[708] | 140 | |
---|
| 141 | |
---|
| 142 | /////////////////////////////////////////////////////////////// |
---|
| 143 | __attribute__((constructor)) void execute( arguments_t* pargs ) |
---|
| 144 | /////////////////////////////////////////////////////////////// |
---|
| 145 | { |
---|
| 146 | unsigned int nb_lines = pargs->lines; |
---|
| 147 | unsigned int base_line = pargs->index; |
---|
| 148 | |
---|
| 149 | ///////////// parallel world initialization |
---|
| 150 | |
---|
| 151 | // All processors initialize world[0] |
---|
| 152 | init_world( 0 , base_line , nb_lines ); |
---|
| 153 | |
---|
| 154 | // copy world[0] to display[0] |
---|
| 155 | copy_world( 0 , base_line , nb_lines ); |
---|
| 156 | |
---|
| 157 | // synchronise with other procs |
---|
| 158 | sqt_barrier_wait( &barrier ); |
---|
| 159 | |
---|
| 160 | // main() makes display[0] |
---|
| 161 | if ( base_line == 0 ) giet_fbf_cma_display ( 0 ); |
---|
| 162 | |
---|
| 163 | //////////// evolution : 2 steps per iteration |
---|
| 164 | |
---|
| 165 | unsigned int i = 0; |
---|
| 166 | while( 1 ) |
---|
| 167 | { |
---|
| 168 | // compute world[1] from world[0] |
---|
| 169 | compute_new_gen( 1 , base_line , nb_lines ); |
---|
| 170 | |
---|
| 171 | // copy world[1] to display[1] |
---|
| 172 | copy_world( 1 , base_line , nb_lines ); |
---|
| 173 | |
---|
| 174 | // synchronise with other procs |
---|
| 175 | sqt_barrier_wait( &barrier ); |
---|
| 176 | |
---|
| 177 | // main makes display[1] |
---|
| 178 | if ( base_line == 0 ) giet_fbf_cma_display ( 1 ); |
---|
| 179 | |
---|
| 180 | #if VERBOSE |
---|
| 181 | if ( base_line == 0 ) giet_tty_printf(" - step %d\n", 2*i ); |
---|
| 182 | #endif |
---|
| 183 | |
---|
| 184 | // compute world[0] from world[1] |
---|
| 185 | compute_new_gen( 0 , base_line , nb_lines ); |
---|
| 186 | |
---|
| 187 | // copy world[0] to display[0] |
---|
| 188 | copy_world( 0 , base_line , nb_lines ); |
---|
| 189 | |
---|
| 190 | // synchronise with other procs |
---|
| 191 | sqt_barrier_wait( &barrier ); |
---|
| 192 | |
---|
| 193 | // main makes display[0] |
---|
| 194 | if ( base_line == 0 ) giet_fbf_cma_display ( 0 ); |
---|
| 195 | |
---|
| 196 | #if VERBOSE |
---|
| 197 | if ( base_line == 0 ) giet_tty_printf(" - step %d\n", 2*i + 1 ); |
---|
| 198 | #endif |
---|
| 199 | |
---|
| 200 | i++; |
---|
| 201 | |
---|
| 202 | } // end evolution loop |
---|
| 203 | |
---|
| 204 | giet_pthread_exit("Completed"); |
---|
| 205 | |
---|
| 206 | } // end main() |
---|
| 207 | |
---|
| 208 | |
---|
| 209 | |
---|
[263] | 210 | //////////////////////////////////////// |
---|
[251] | 211 | __attribute__((constructor)) void main() |
---|
[669] | 212 | //////////////////////////////////////// |
---|
[251] | 213 | { |
---|
[432] | 214 | // get processor identifier |
---|
| 215 | unsigned int x; |
---|
| 216 | unsigned int y; |
---|
| 217 | unsigned int p; |
---|
| 218 | giet_proc_xyp( &x, &y, &p ); |
---|
| 219 | |
---|
[708] | 220 | // get platform parameters |
---|
[502] | 221 | unsigned int x_size; |
---|
| 222 | unsigned int y_size; |
---|
[669] | 223 | unsigned int nprocs; |
---|
| 224 | giet_procs_number( &x_size, &y_size, &nprocs ); |
---|
[502] | 225 | |
---|
[708] | 226 | giet_pthread_assert( (x_size <= 16) , "x_size no larger than 16" ); |
---|
| 227 | giet_pthread_assert( (y_size <= 16) , "y_size no larger than 16" ); |
---|
| 228 | giet_pthread_assert( (nprocs <= 4) , "nprocs no larger than 16" ); |
---|
[432] | 229 | |
---|
[708] | 230 | // compute number of threads and min number of lines per thread |
---|
| 231 | // extra is the number of threads that must process one extra line |
---|
| 232 | unsigned int total_procs = x_size * y_size * nprocs; |
---|
| 233 | unsigned int nthreads; |
---|
| 234 | unsigned int nlines; |
---|
| 235 | unsigned int extra; |
---|
| 236 | if ( total_procs > HEIGHT ) |
---|
| 237 | { |
---|
| 238 | nthreads = HEIGHT; |
---|
| 239 | nlines = 1; |
---|
| 240 | extra = 0; |
---|
| 241 | } |
---|
| 242 | else |
---|
| 243 | { |
---|
| 244 | nthreads = total_procs; |
---|
| 245 | nlines = HEIGHT / total_procs; |
---|
| 246 | extra = HEIGHT % total_procs; |
---|
| 247 | } |
---|
[251] | 248 | |
---|
[708] | 249 | // get a shared TTY |
---|
| 250 | giet_tty_alloc( 1 ); |
---|
[251] | 251 | |
---|
[708] | 252 | // get a Chained Buffer DMA channel |
---|
| 253 | giet_fbf_cma_alloc(); |
---|
[509] | 254 | |
---|
[708] | 255 | // initializes the source and destination buffers |
---|
| 256 | giet_fbf_cma_init_buf( &display[0][0][0] , |
---|
| 257 | &display[1][0][0] , |
---|
| 258 | status0 , |
---|
| 259 | status1 ); |
---|
[669] | 260 | |
---|
[708] | 261 | // activates CMA channel |
---|
| 262 | giet_fbf_cma_start( HEIGHT * WIDTH ); |
---|
[669] | 263 | |
---|
[708] | 264 | // initializes distributed heap |
---|
| 265 | unsigned int cx; |
---|
| 266 | unsigned int cy; |
---|
| 267 | for ( cx = 0 ; cx < x_size ; cx++ ) |
---|
[295] | 268 | { |
---|
[708] | 269 | for ( cy = 0 ; cy < y_size ; cy++ ) |
---|
[669] | 270 | { |
---|
[708] | 271 | heap_init( cx , cy ); |
---|
[669] | 272 | } |
---|
[708] | 273 | } |
---|
[669] | 274 | |
---|
[708] | 275 | // initialises barrier |
---|
| 276 | sqt_barrier_init( &barrier , x_size , y_size , nprocs ); |
---|
[509] | 277 | |
---|
[708] | 278 | giet_tty_printf("\n[GAMEOFLIFE] P[%d,%d,%d] completes initialisation at cycle %d\n" |
---|
| 279 | " nprocs = %d / nlines = %d / nthreads = %d\n", |
---|
| 280 | x, y, p, giet_proctime() , total_procs , HEIGHT , nthreads ); |
---|
[669] | 281 | |
---|
[708] | 282 | // compute arguments (index, nlines) for all threads |
---|
| 283 | unsigned int n; // thread index |
---|
| 284 | unsigned int index; // first line index |
---|
| 285 | for ( n = 0 , index = 0 ; n < nthreads ; n++ ) |
---|
[295] | 286 | { |
---|
[708] | 287 | if ( extra ) |
---|
| 288 | { |
---|
| 289 | args[n].index = index; |
---|
| 290 | args[n].lines = nlines + 1; |
---|
| 291 | index = index + nlines + 1; |
---|
| 292 | } |
---|
| 293 | else |
---|
| 294 | { |
---|
| 295 | args[n].index = index; |
---|
| 296 | args[n].lines = nlines; |
---|
| 297 | index = index + nlines; |
---|
| 298 | } |
---|
| 299 | #if VERBOSE |
---|
| 300 | giet_tty_printf("[GAMEOFLIFE] Thread %d : first = %d / nlines = %d\n", |
---|
| 301 | n , args[n].index , args[n].lines ); |
---|
| 302 | #endif |
---|
[295] | 303 | } |
---|
| 304 | |
---|
[708] | 305 | // launch all other threads |
---|
| 306 | pthread_t trdid; // unused because no pthread_join() |
---|
| 307 | for ( n = 1 ; n < nthreads ; n++ ) |
---|
[251] | 308 | { |
---|
[708] | 309 | if ( giet_pthread_create( &trdid, |
---|
| 310 | NULL, // no attribute |
---|
| 311 | &execute, |
---|
| 312 | &args[n] ) ) |
---|
| 313 | { |
---|
| 314 | giet_tty_printf("\n[TRANSPOSE ERROR] creating thread %x\n", n ); |
---|
| 315 | giet_pthread_exit( NULL ); |
---|
| 316 | } |
---|
| 317 | } |
---|
[502] | 318 | |
---|
[708] | 319 | // run execute function |
---|
| 320 | execute( &args[0] ); |
---|
[509] | 321 | |
---|
[708] | 322 | giet_pthread_exit( "completed" ); |
---|
| 323 | |
---|
| 324 | } // end main() |
---|
[502] | 325 | |
---|
[251] | 326 | |
---|
[509] | 327 | |
---|
[251] | 328 | // Local Variables: |
---|
| 329 | // tab-width: 3 |
---|
| 330 | // c-basic-offset: 3 |
---|
| 331 | // c-file-offsets:((innamespace . 0)(inline-open . 0)) |
---|
| 332 | // indent-tabs-mode: nil |
---|
| 333 | // End: |
---|
| 334 | |
---|
| 335 | // vim: filetype=cpp:expandtab:shiftwidth=3:tabstop=3:softtabstop=3 |
---|
| 336 | |
---|
| 337 | |
---|
| 338 | |
---|