[417] | 1 | /////////////////////////////////////////////////////////////////////////////// |
---|
| 2 | // File : sort.c |
---|
| 3 | // Date : November 2013 |
---|
| 4 | // Author : Cesar Fuguet Tortolero <cesar.fuguet-tortolero@lip6.fr> |
---|
| 5 | /////////////////////////////////////////////////////////////////////////////// |
---|
| 6 | // This multi-threaded application implement a multi-stage sort application. |
---|
| 7 | // The various stages are separated by synchronisation barriers. |
---|
| 8 | // There is one thread per physical cores. |
---|
| 9 | // Computation is organised as a binary tree: |
---|
| 10 | // - All threads execute in parallel a buble sort on a sub-array during the |
---|
| 11 | // the first stage of parallel sort, |
---|
| 12 | // - The number of participating threads is divided by 2 at each next stage, |
---|
| 13 | // to make a merge sort, on two subsets of previous stage. |
---|
| 14 | // |
---|
| 15 | // Number_of_stages = number of barriers = log2(Number_of_threads) |
---|
| 16 | // |
---|
| 17 | // Constraints : |
---|
| 18 | // - It supports up to 1024 cores: x_size, y_size, and ncores must be |
---|
| 19 | // power of 2 (max 16*16 clusters / max 4 cores per cluster) |
---|
| 20 | // _ The array of values to be sorted (ARRAY_LENGTH) must be power of 2 |
---|
| 21 | // larger than the number of cores. |
---|
| 22 | /////////////////////////////////////////////////////////////////////////////// |
---|
| 23 | |
---|
| 24 | #include <stdio.h> |
---|
| 25 | #include <stdlib.h> |
---|
| 26 | #include <malloc.h> |
---|
| 27 | #include <pthread.h> |
---|
| 28 | |
---|
[442] | 29 | #define ARRAY_LENGTH 0x400 // 1024 values |
---|
[417] | 30 | |
---|
[440] | 31 | #define MAX_THREADS 1024 // 16 * 16 * 4 |
---|
| 32 | |
---|
| 33 | #define DISPLAY_ARRAY 0 |
---|
[442] | 34 | #define INTERACTIVE_MODE 1 |
---|
[440] | 35 | |
---|
[417] | 36 | /////////////////////////////////////////////////////// |
---|
| 37 | // macros for fixed format cxy <=> (x,y) translation |
---|
[440] | 38 | // TODO these macros are only for TSAR architecture... |
---|
[417] | 39 | /////////////////////////////////////////////////////// |
---|
| 40 | |
---|
| 41 | #define CXY_FROM_XY( x , y ) ((x<<4) + y) |
---|
| 42 | |
---|
| 43 | #define X_FROM_CXY( cxy ) ((cxy>>4) & 0xF) |
---|
| 44 | |
---|
| 45 | #define Y_FROM_CXY( cxy ) (cxy & 0xF) |
---|
| 46 | |
---|
| 47 | ///////////////////////////////////////////////////////////// |
---|
| 48 | // argument for the sort() function (one thread per core) |
---|
| 49 | ///////////////////////////////////////////////////////////// |
---|
| 50 | |
---|
| 51 | typedef struct |
---|
| 52 | { |
---|
| 53 | unsigned int threads; // total number of threads |
---|
| 54 | unsigned int thread_uid; // thread user index (0 to threads -1) |
---|
| 55 | unsigned int main_uid; // main thread user index |
---|
| 56 | } |
---|
| 57 | args_t; |
---|
| 58 | |
---|
| 59 | ////////////////////////////////////////// |
---|
| 60 | // Global variables |
---|
| 61 | ////////////////////////////////////////// |
---|
| 62 | |
---|
| 63 | int array0[ARRAY_LENGTH]; // values to sort |
---|
| 64 | int array1[ARRAY_LENGTH]; |
---|
| 65 | |
---|
| 66 | pthread_barrier_t barrier; // synchronisation variables |
---|
| 67 | |
---|
[440] | 68 | pthread_attr_t attr[MAX_THREADS]; // thread attributes (one per thread) |
---|
| 69 | args_t arg[MAX_THREADS]; // sort function arguments (one per thread) |
---|
[417] | 70 | |
---|
| 71 | //////////////////////////////////// |
---|
| 72 | void bubbleSort( int * array, |
---|
| 73 | unsigned int length, |
---|
| 74 | unsigned int init_pos ) |
---|
| 75 | { |
---|
| 76 | int i; |
---|
| 77 | int j; |
---|
| 78 | int aux; |
---|
| 79 | |
---|
| 80 | for(i = 0; i < length; i++) |
---|
| 81 | { |
---|
| 82 | for(j = init_pos; j < (init_pos + length - i - 1); j++) |
---|
| 83 | { |
---|
| 84 | if(array[j] > array[j + 1]) |
---|
| 85 | { |
---|
| 86 | aux = array[j + 1]; |
---|
| 87 | array[j + 1] = array[j]; |
---|
| 88 | array[j] = aux; |
---|
| 89 | } |
---|
| 90 | } |
---|
| 91 | } |
---|
| 92 | } // end bubbleSort() |
---|
| 93 | |
---|
| 94 | |
---|
| 95 | ///////////////////////// |
---|
| 96 | void merge( int * src, |
---|
| 97 | int * dst, |
---|
| 98 | int length, |
---|
| 99 | int init_pos_src_a, |
---|
| 100 | int init_pos_src_b, |
---|
| 101 | int init_pos_dst ) |
---|
| 102 | { |
---|
| 103 | int i; |
---|
| 104 | int j; |
---|
| 105 | int k; |
---|
| 106 | |
---|
| 107 | i = 0; |
---|
| 108 | j = 0; |
---|
| 109 | k = init_pos_dst; |
---|
| 110 | |
---|
| 111 | while((i < length) || (j < length)) |
---|
| 112 | { |
---|
| 113 | if((i < length) && (j < length)) |
---|
| 114 | { |
---|
| 115 | if(src[init_pos_src_a + i] < src[init_pos_src_b + j]) |
---|
| 116 | { |
---|
| 117 | dst[k++] = src[init_pos_src_a + i]; |
---|
| 118 | i++; |
---|
| 119 | } |
---|
| 120 | else |
---|
| 121 | { |
---|
| 122 | dst[k++] = src[init_pos_src_b + j]; |
---|
| 123 | j++; |
---|
| 124 | } |
---|
| 125 | } |
---|
| 126 | else if(i < length) |
---|
| 127 | { |
---|
| 128 | dst[k++] = src[init_pos_src_a + i]; |
---|
| 129 | i++; |
---|
| 130 | } |
---|
| 131 | else |
---|
| 132 | { |
---|
| 133 | dst[k++] = src[init_pos_src_b + j]; |
---|
| 134 | j++; |
---|
| 135 | } |
---|
| 136 | } |
---|
| 137 | } // end merge() |
---|
| 138 | |
---|
| 139 | ///////////////////////// |
---|
| 140 | void sort( args_t * ptr ) |
---|
| 141 | { |
---|
| 142 | unsigned int i; |
---|
| 143 | unsigned long long cycle; |
---|
[440] | 144 | unsigned int cxy; |
---|
| 145 | unsigned int lid; |
---|
[417] | 146 | |
---|
| 147 | int * src_array = NULL; |
---|
| 148 | int * dst_array = NULL; |
---|
| 149 | |
---|
[440] | 150 | // get core coordinates an date |
---|
| 151 | get_core( &cxy , &lid ); |
---|
| 152 | get_cycle( &cycle ); |
---|
| 153 | |
---|
[417] | 154 | unsigned int thread_uid = ptr->thread_uid; |
---|
| 155 | unsigned int threads = ptr->threads; |
---|
| 156 | unsigned int main_uid = ptr->main_uid; |
---|
| 157 | |
---|
| 158 | unsigned int items = ARRAY_LENGTH / threads; |
---|
| 159 | unsigned int stages = __builtin_ctz( threads ) + 1; |
---|
| 160 | |
---|
[442] | 161 | printf("\n[SORT] thread[%d] : start\n", thread_uid ); |
---|
| 162 | |
---|
[417] | 163 | bubbleSort( array0, items, items * thread_uid ); |
---|
| 164 | |
---|
| 165 | printf("\n[SORT] thread[%d] / stage 0 completed\n", thread_uid ); |
---|
| 166 | |
---|
| 167 | ///////////////////////////////// |
---|
| 168 | pthread_barrier_wait( &barrier ); |
---|
[440] | 169 | printf("\n[SORT] thread[%d] exit barrier\n", thread_uid ); |
---|
| 170 | |
---|
[417] | 171 | // the number of threads contributing to sort |
---|
| 172 | // is divided by 2 at each next stage |
---|
| 173 | for ( i = 1 ; i < stages ; i++ ) |
---|
| 174 | { |
---|
| 175 | pthread_barrier_wait( &barrier ); |
---|
| 176 | |
---|
| 177 | if( (thread_uid & ((1<<i)-1)) == 0 ) |
---|
| 178 | { |
---|
| 179 | printf("\n[SORT] thread[%d] / stage %d start\n", thread_uid , i ); |
---|
| 180 | |
---|
| 181 | if((i % 2) == 1) // odd stage |
---|
| 182 | { |
---|
| 183 | src_array = array0; |
---|
| 184 | dst_array = array1; |
---|
| 185 | } |
---|
| 186 | else // even stage |
---|
| 187 | { |
---|
| 188 | src_array = array1; |
---|
| 189 | dst_array = array0; |
---|
| 190 | } |
---|
| 191 | |
---|
| 192 | merge( src_array, |
---|
| 193 | dst_array, |
---|
| 194 | items << i, |
---|
| 195 | items * thread_uid, |
---|
| 196 | items * (thread_uid + (1 << (i-1))), |
---|
| 197 | items * thread_uid ); |
---|
| 198 | |
---|
| 199 | printf("\n[SORT] thread[%d] / stage %d completed\n", thread_uid , i ); |
---|
| 200 | } |
---|
| 201 | |
---|
| 202 | ///////////////////////////////// |
---|
| 203 | pthread_barrier_wait( &barrier ); |
---|
[442] | 204 | printf("\n[SORT] thread[%d] exit barrier\n", thread_uid ); |
---|
[417] | 205 | |
---|
| 206 | } |
---|
| 207 | |
---|
| 208 | // all threads but the main thread exit |
---|
| 209 | if( thread_uid != main_uid ) pthread_exit( NULL ); |
---|
| 210 | |
---|
| 211 | } // end sort() |
---|
| 212 | |
---|
| 213 | |
---|
| 214 | /////////// |
---|
| 215 | void main() |
---|
| 216 | { |
---|
| 217 | unsigned int x_size; // number of rows |
---|
| 218 | unsigned int y_size; // number of columns |
---|
| 219 | unsigned int ncores; // number of cores per cluster |
---|
| 220 | unsigned int threads; // total number of threads |
---|
| 221 | unsigned int thread_uid; // user defined thread index |
---|
| 222 | unsigned int main_cxy; // cluster identifier for main |
---|
| 223 | unsigned int main_x; // X coordinate for main thread |
---|
| 224 | unsigned int main_y; // Y coordinate for main thread |
---|
| 225 | unsigned int main_lid; // core local index for main thread |
---|
| 226 | unsigned int main_uid; // thread user index for main thread |
---|
| 227 | unsigned int x; // X coordinate for a thread |
---|
| 228 | unsigned int y; // Y coordinate for a thread |
---|
| 229 | unsigned int lid; // core local index for a thread |
---|
| 230 | unsigned int n; // index in array to sort |
---|
| 231 | unsigned long long cycle; // current date for log |
---|
| 232 | pthread_t trdid; // kernel allocated thread index (unused) |
---|
| 233 | pthread_barrierattr_t barrier_attr; // barrier attributes |
---|
| 234 | |
---|
| 235 | // compute number of threads (one thread per proc) |
---|
| 236 | get_config( &x_size , &y_size , &ncores ); |
---|
| 237 | threads = x_size * y_size * ncores; |
---|
| 238 | |
---|
| 239 | // get core coordinates and user index for the main thread |
---|
| 240 | get_core( &main_cxy , & main_lid ); |
---|
| 241 | main_x = X_FROM_CXY( main_cxy ); |
---|
| 242 | main_y = Y_FROM_CXY( main_cxy ); |
---|
| 243 | main_uid = (((main_x * y_size) + main_y) * ncores) + main_lid; |
---|
| 244 | |
---|
| 245 | // checks number of threads |
---|
| 246 | if ( (threads != 1) && (threads != 2) && (threads != 4) && |
---|
| 247 | (threads != 8) && (threads != 16 ) && (threads != 32) && |
---|
| 248 | (threads != 64) && (threads != 128) && (threads != 256) && |
---|
| 249 | (threads != 512) && (threads != 1024) ) |
---|
| 250 | { |
---|
| 251 | printf("\n[SORT ERROR] number of cores must be power of 2\n"); |
---|
[434] | 252 | exit( 0 ); |
---|
[417] | 253 | } |
---|
| 254 | |
---|
| 255 | // check array size |
---|
| 256 | if ( ARRAY_LENGTH % threads) |
---|
| 257 | { |
---|
| 258 | printf("\n[SORT ERROR] array size must be multiple of number of threads\n"); |
---|
[434] | 259 | exit( 0 ); |
---|
[417] | 260 | } |
---|
| 261 | |
---|
| 262 | get_cycle( &cycle ); |
---|
[440] | 263 | printf("\n\n[SORT] main starts on core[%x,%d] / %d threads / %d values / cycle %d\n", |
---|
| 264 | main_cxy, main_lid, threads, ARRAY_LENGTH, (unsigned int)cycle ); |
---|
[417] | 265 | |
---|
| 266 | // Barrier initialization |
---|
| 267 | barrier_attr.x_size = x_size; |
---|
| 268 | barrier_attr.y_size = y_size; |
---|
| 269 | barrier_attr.nthreads = ncores; |
---|
| 270 | if( pthread_barrier_init( &barrier, &barrier_attr , threads ) ) |
---|
| 271 | { |
---|
| 272 | printf("\n[SORT ERROR] cannot initialise barrier\n" ); |
---|
[434] | 273 | exit( 0 ); |
---|
[417] | 274 | } |
---|
| 275 | |
---|
| 276 | get_cycle( &cycle ); |
---|
[440] | 277 | printf("\n[SORT] main completes barrier init at cycle %d\n", (unsigned int)cycle ); |
---|
[417] | 278 | |
---|
| 279 | // Array to sort initialization |
---|
| 280 | for ( n = 0 ; n < ARRAY_LENGTH ; n++ ) |
---|
| 281 | { |
---|
| 282 | array0[n] = rand(); |
---|
| 283 | } |
---|
| 284 | |
---|
[440] | 285 | #if DISPLAY_ARRAY |
---|
[417] | 286 | printf("\n*** array before sort\n"); |
---|
| 287 | for( n=0; n<ARRAY_LENGTH; n++) printf("array[%d] = %d\n", n , array0[n] ); |
---|
| 288 | #endif |
---|
| 289 | |
---|
| 290 | get_cycle( &cycle ); |
---|
[440] | 291 | printf("\n[SORT] main completes array init at cycle %d\n", (unsigned int)cycle ); |
---|
[417] | 292 | |
---|
| 293 | // launch other threads to execute sort() function |
---|
| 294 | // on cores other than the core running the main thread |
---|
| 295 | for ( x=0 ; x<x_size ; x++ ) |
---|
| 296 | { |
---|
| 297 | for ( y=0 ; y<y_size ; y++ ) |
---|
| 298 | { |
---|
| 299 | for ( lid=0 ; lid<ncores ; lid++ ) |
---|
| 300 | { |
---|
| 301 | thread_uid = (((x * y_size) + y) * ncores) + lid; |
---|
| 302 | |
---|
| 303 | // set sort arguments for all threads |
---|
| 304 | arg[thread_uid].threads = threads; |
---|
| 305 | arg[thread_uid].thread_uid = thread_uid; |
---|
| 306 | arg[thread_uid].main_uid = main_uid; |
---|
| 307 | |
---|
| 308 | // set thread attributes for all threads |
---|
| 309 | attr[thread_uid].attributes = PT_ATTR_CLUSTER_DEFINED | PT_ATTR_CORE_DEFINED; |
---|
| 310 | attr[thread_uid].cxy = CXY_FROM_XY( x , y ); |
---|
| 311 | attr[thread_uid].lid = lid; |
---|
| 312 | |
---|
| 313 | if( thread_uid != main_uid ) |
---|
| 314 | { |
---|
| 315 | if ( pthread_create( &trdid, // not used because no join |
---|
| 316 | &attr[thread_uid], // thread attributes |
---|
| 317 | &sort, // entry function |
---|
| 318 | &arg[thread_uid] ) ) // sort arguments |
---|
| 319 | { |
---|
[442] | 320 | printf("\n[SORT ERROR] main cannot create thread %x \n", thread_uid ); |
---|
[434] | 321 | exit( 0 ); |
---|
[417] | 322 | } |
---|
[442] | 323 | else |
---|
| 324 | { |
---|
| 325 | printf("\n[SORT] main created thread %x \n", thread_uid ); |
---|
| 326 | } |
---|
[440] | 327 | } |
---|
[442] | 328 | |
---|
| 329 | #if INTERACTIVE_MODE |
---|
| 330 | idbg(); |
---|
[440] | 331 | #endif |
---|
[417] | 332 | } |
---|
| 333 | } |
---|
| 334 | } |
---|
[442] | 335 | |
---|
[417] | 336 | get_cycle( &cycle ); |
---|
[440] | 337 | printf("\n[SORT] main completes threads create at cycle %d\n", (unsigned int)cycle ); |
---|
[417] | 338 | |
---|
[440] | 339 | // main run also the sort() function |
---|
[417] | 340 | sort( &arg[main_uid] ); |
---|
| 341 | |
---|
[442] | 342 | #if INTERACTIVE_MODE |
---|
| 343 | idbg(); |
---|
| 344 | #endif |
---|
[441] | 345 | |
---|
[417] | 346 | // Check result |
---|
| 347 | int success = 1; |
---|
| 348 | int* res_array = ( (threads== 2) || |
---|
| 349 | (threads== 8) || |
---|
| 350 | (threads== 32) || |
---|
| 351 | (threads==128) || |
---|
| 352 | (threads==512) ) ? array1 : array0; |
---|
| 353 | |
---|
[441] | 354 | for( n=0 ; n<(ARRAY_LENGTH-2) ; n++ ) |
---|
[417] | 355 | { |
---|
| 356 | if ( res_array[n] > res_array[n+1] ) |
---|
| 357 | { |
---|
[440] | 358 | printf("\n[SORT] array[%d] = %d > array[%d] = %d\n", |
---|
| 359 | n , res_array[n] , n+1 , res_array[n+1] ); |
---|
[417] | 360 | success = 0; |
---|
| 361 | break; |
---|
| 362 | } |
---|
| 363 | } |
---|
| 364 | |
---|
[440] | 365 | #if DISPLAY_ARRAY |
---|
[417] | 366 | printf("\n*** array after sort\n"); |
---|
| 367 | for( n=0; n<ARRAY_LENGTH; n++) printf("array[%d] = %d\n", n , res_array[n] ); |
---|
| 368 | #endif |
---|
| 369 | |
---|
| 370 | get_cycle( &cycle ); |
---|
| 371 | |
---|
| 372 | if ( success ) |
---|
| 373 | { |
---|
| 374 | printf("\n[SORT] success at cycle %d\n", (unsigned int)cycle ); |
---|
[434] | 375 | exit( 0 ); |
---|
[417] | 376 | } |
---|
| 377 | else |
---|
| 378 | { |
---|
| 379 | printf("\n[SORT] failure at cycle %d\n", (unsigned int)cycle ); |
---|
[436] | 380 | exit( 1 ); |
---|
[417] | 381 | } |
---|
| 382 | |
---|
| 383 | } // end main() |
---|
| 384 | |
---|
| 385 | |
---|
| 386 | /* |
---|
| 387 | vim: tabstop=4 : shiftwidth=4 : expandtab |
---|
| 388 | */ |
---|