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 <pthread.h> |
---|
27 | #include <almosmkh.h> |
---|
28 | #include <hal_macros.h> |
---|
29 | |
---|
30 | #define ARRAY_LENGTH 0x100 // 256 values |
---|
31 | |
---|
32 | #define MAX_THREADS 1024 // 16 * 16 * 4 |
---|
33 | |
---|
34 | #define DISPLAY_ARRAY 0 |
---|
35 | #define INTERACTIVE_MODE 0 |
---|
36 | |
---|
37 | ///////////////////////////////////////////////////////////// |
---|
38 | // argument for the sort() function (one thread per core) |
---|
39 | ///////////////////////////////////////////////////////////// |
---|
40 | |
---|
41 | typedef struct |
---|
42 | { |
---|
43 | unsigned int threads; // total number of threads |
---|
44 | unsigned int thread_uid; // thread user index (0 to threads -1) |
---|
45 | unsigned int main_uid; // main thread user index |
---|
46 | } |
---|
47 | args_t; |
---|
48 | |
---|
49 | ////////////////////////////////////////// |
---|
50 | // Global variables |
---|
51 | ////////////////////////////////////////// |
---|
52 | |
---|
53 | int array0[ARRAY_LENGTH]; // values to sort |
---|
54 | int array1[ARRAY_LENGTH]; |
---|
55 | |
---|
56 | pthread_barrier_t barrier; // synchronisation variables |
---|
57 | |
---|
58 | pthread_attr_t attr[MAX_THREADS]; // thread attributes (one per thread) |
---|
59 | args_t arg[MAX_THREADS]; // sort function arguments (one per thread) |
---|
60 | |
---|
61 | //////////////////////////////////// |
---|
62 | void bubbleSort( int * array, |
---|
63 | unsigned int length, |
---|
64 | unsigned int init_pos ) |
---|
65 | { |
---|
66 | int i; |
---|
67 | int j; |
---|
68 | int aux; |
---|
69 | |
---|
70 | for(i = 0; i < length; i++) |
---|
71 | { |
---|
72 | for(j = init_pos; j < (init_pos + length - i - 1); j++) |
---|
73 | { |
---|
74 | if(array[j] > array[j + 1]) |
---|
75 | { |
---|
76 | aux = array[j + 1]; |
---|
77 | array[j + 1] = array[j]; |
---|
78 | array[j] = aux; |
---|
79 | } |
---|
80 | } |
---|
81 | } |
---|
82 | } // end bubbleSort() |
---|
83 | |
---|
84 | |
---|
85 | ///////////////////////// |
---|
86 | void merge( int * src, |
---|
87 | int * dst, |
---|
88 | int length, |
---|
89 | int init_pos_src_a, |
---|
90 | int init_pos_src_b, |
---|
91 | int init_pos_dst ) |
---|
92 | { |
---|
93 | int i; |
---|
94 | int j; |
---|
95 | int k; |
---|
96 | |
---|
97 | i = 0; |
---|
98 | j = 0; |
---|
99 | k = init_pos_dst; |
---|
100 | |
---|
101 | while((i < length) || (j < length)) |
---|
102 | { |
---|
103 | if((i < length) && (j < length)) |
---|
104 | { |
---|
105 | if(src[init_pos_src_a + i] < src[init_pos_src_b + j]) |
---|
106 | { |
---|
107 | dst[k++] = src[init_pos_src_a + i]; |
---|
108 | i++; |
---|
109 | } |
---|
110 | else |
---|
111 | { |
---|
112 | dst[k++] = src[init_pos_src_b + j]; |
---|
113 | j++; |
---|
114 | } |
---|
115 | } |
---|
116 | else if(i < length) |
---|
117 | { |
---|
118 | dst[k++] = src[init_pos_src_a + i]; |
---|
119 | i++; |
---|
120 | } |
---|
121 | else |
---|
122 | { |
---|
123 | dst[k++] = src[init_pos_src_b + j]; |
---|
124 | j++; |
---|
125 | } |
---|
126 | } |
---|
127 | } // end merge() |
---|
128 | |
---|
129 | ///////////////////////// |
---|
130 | void sort( args_t * ptr ) |
---|
131 | { |
---|
132 | unsigned int i; |
---|
133 | unsigned long long cycle; |
---|
134 | unsigned int cxy; |
---|
135 | unsigned int lid; |
---|
136 | |
---|
137 | int * src_array = NULL; |
---|
138 | int * dst_array = NULL; |
---|
139 | |
---|
140 | // get core coordinates an date |
---|
141 | get_core( &cxy , &lid ); |
---|
142 | get_cycle( &cycle ); |
---|
143 | |
---|
144 | unsigned int thread_uid = ptr->thread_uid; |
---|
145 | unsigned int threads = ptr->threads; |
---|
146 | unsigned int main_uid = ptr->main_uid; |
---|
147 | |
---|
148 | unsigned int items = ARRAY_LENGTH / threads; |
---|
149 | unsigned int stages = __builtin_ctz( threads ) + 1; |
---|
150 | |
---|
151 | printf("\n[SORT] thread[%d] : start\n", thread_uid ); |
---|
152 | |
---|
153 | bubbleSort( array0, items, items * thread_uid ); |
---|
154 | |
---|
155 | printf("\n[SORT] thread[%d] : stage 0 completed\n", thread_uid ); |
---|
156 | |
---|
157 | ///////////////////////////////// |
---|
158 | pthread_barrier_wait( &barrier ); |
---|
159 | printf("\n[SORT] thread[%d] exit barrier\n", thread_uid ); |
---|
160 | |
---|
161 | // the number of threads contributing to sort |
---|
162 | // is divided by 2 at each next stage |
---|
163 | for ( i = 1 ; i < stages ; i++ ) |
---|
164 | { |
---|
165 | pthread_barrier_wait( &barrier ); |
---|
166 | |
---|
167 | if( (thread_uid & ((1<<i)-1)) == 0 ) |
---|
168 | { |
---|
169 | printf("\n[SORT] thread[%d] : stage %d start\n", thread_uid , i ); |
---|
170 | |
---|
171 | if((i % 2) == 1) // odd stage |
---|
172 | { |
---|
173 | src_array = array0; |
---|
174 | dst_array = array1; |
---|
175 | } |
---|
176 | else // even stage |
---|
177 | { |
---|
178 | src_array = array1; |
---|
179 | dst_array = array0; |
---|
180 | } |
---|
181 | |
---|
182 | merge( src_array, |
---|
183 | dst_array, |
---|
184 | items << i, |
---|
185 | items * thread_uid, |
---|
186 | items * (thread_uid + (1 << (i-1))), |
---|
187 | items * thread_uid ); |
---|
188 | |
---|
189 | printf("\n[SORT] thread[%d] : stage %d completed\n", thread_uid , i ); |
---|
190 | } |
---|
191 | |
---|
192 | ///////////////////////////////// |
---|
193 | pthread_barrier_wait( &barrier ); |
---|
194 | printf("\n[SORT] thread[%d] exit barrier\n", thread_uid ); |
---|
195 | |
---|
196 | } |
---|
197 | |
---|
198 | // all threads but the main thread exit |
---|
199 | if( thread_uid != main_uid ) pthread_exit( NULL ); |
---|
200 | |
---|
201 | } // end sort() |
---|
202 | |
---|
203 | |
---|
204 | /////////// |
---|
205 | void main() |
---|
206 | { |
---|
207 | unsigned int x_size; // number of rows |
---|
208 | unsigned int y_size; // number of columns |
---|
209 | unsigned int ncores; // number of cores per cluster |
---|
210 | unsigned int threads; // total number of threads |
---|
211 | unsigned int thread_uid; // user defined thread index |
---|
212 | unsigned int main_cxy; // cluster identifier for main |
---|
213 | unsigned int main_x; // X coordinate for main thread |
---|
214 | unsigned int main_y; // Y coordinate for main thread |
---|
215 | unsigned int main_lid; // core local index for main thread |
---|
216 | unsigned int main_uid; // thread user index for main thread |
---|
217 | unsigned int x; // X coordinate for a thread |
---|
218 | unsigned int y; // Y coordinate for a thread |
---|
219 | unsigned int lid; // core local index for a thread |
---|
220 | unsigned int n; // index in array to sort |
---|
221 | unsigned long long cycle; // current date for log |
---|
222 | pthread_t trdid; // kernel allocated thread index (unused) |
---|
223 | pthread_barrierattr_t barrier_attr; // barrier attributes |
---|
224 | |
---|
225 | // compute number of threads (one thread per proc) |
---|
226 | get_config( &x_size , &y_size , &ncores ); |
---|
227 | threads = x_size * y_size * ncores; |
---|
228 | |
---|
229 | // get core coordinates and user index for the main thread |
---|
230 | get_core( &main_cxy , & main_lid ); |
---|
231 | main_x = HAL_X_FROM_CXY( main_cxy ); |
---|
232 | main_y = HAL_Y_FROM_CXY( main_cxy ); |
---|
233 | main_uid = (((main_x * y_size) + main_y) * ncores) + main_lid; |
---|
234 | |
---|
235 | // checks number of threads |
---|
236 | if ( (threads != 1) && (threads != 2) && (threads != 4) && |
---|
237 | (threads != 8) && (threads != 16 ) && (threads != 32) && |
---|
238 | (threads != 64) && (threads != 128) && (threads != 256) && |
---|
239 | (threads != 512) && (threads != 1024) ) |
---|
240 | { |
---|
241 | printf("\n[SORT ERROR] number of cores must be power of 2\n"); |
---|
242 | exit( 0 ); |
---|
243 | } |
---|
244 | |
---|
245 | // check array size |
---|
246 | if ( ARRAY_LENGTH % threads) |
---|
247 | { |
---|
248 | printf("\n[SORT ERROR] array size must be multiple of number of threads\n"); |
---|
249 | exit( 0 ); |
---|
250 | } |
---|
251 | |
---|
252 | printf("\n[SORT] main starts on core[%x,%d] / %d thread(s) / %d values\n", |
---|
253 | main_cxy, main_lid, threads, ARRAY_LENGTH ); |
---|
254 | |
---|
255 | // Barrier initialization |
---|
256 | barrier_attr.x_size = x_size; |
---|
257 | barrier_attr.y_size = y_size; |
---|
258 | barrier_attr.nthreads = ncores; |
---|
259 | if( pthread_barrier_init( &barrier, &barrier_attr , threads ) ) |
---|
260 | { |
---|
261 | printf("\n[SORT ERROR] cannot initialise barrier\n" ); |
---|
262 | exit( 0 ); |
---|
263 | } |
---|
264 | |
---|
265 | get_cycle( &cycle ); |
---|
266 | printf("\n[SORT] main completes barrier init at cycle %d\n", (unsigned int)cycle ); |
---|
267 | |
---|
268 | // Array to sort initialization |
---|
269 | for ( n = 0 ; n < ARRAY_LENGTH ; n++ ) |
---|
270 | { |
---|
271 | array0[n] = rand(); |
---|
272 | } |
---|
273 | |
---|
274 | #if DISPLAY_ARRAY |
---|
275 | printf("\n*** array before sort\n"); |
---|
276 | for( n=0; n<ARRAY_LENGTH; n++) printf("array[%d] = %d\n", n , array0[n] ); |
---|
277 | #endif |
---|
278 | |
---|
279 | get_cycle( &cycle ); |
---|
280 | printf("\n[SORT] main completes array init at cycle %d\n", (unsigned int)cycle ); |
---|
281 | |
---|
282 | // launch other threads to execute sort() function |
---|
283 | // on cores other than the core running the main thread |
---|
284 | for ( x=0 ; x<x_size ; x++ ) |
---|
285 | { |
---|
286 | for ( y=0 ; y<y_size ; y++ ) |
---|
287 | { |
---|
288 | for ( lid=0 ; lid<ncores ; lid++ ) |
---|
289 | { |
---|
290 | thread_uid = (((x * y_size) + y) * ncores) + lid; |
---|
291 | |
---|
292 | // set sort arguments for all threads |
---|
293 | arg[thread_uid].threads = threads; |
---|
294 | arg[thread_uid].thread_uid = thread_uid; |
---|
295 | arg[thread_uid].main_uid = main_uid; |
---|
296 | |
---|
297 | // set thread attributes for all threads |
---|
298 | attr[thread_uid].attributes = PT_ATTR_CLUSTER_DEFINED | PT_ATTR_CORE_DEFINED; |
---|
299 | attr[thread_uid].cxy = HAL_CXY_FROM_XY( x , y ); |
---|
300 | attr[thread_uid].lid = lid; |
---|
301 | |
---|
302 | if( thread_uid != main_uid ) |
---|
303 | { |
---|
304 | if ( pthread_create( &trdid, // not used because no join |
---|
305 | &attr[thread_uid], // thread attributes |
---|
306 | &sort, // entry function |
---|
307 | &arg[thread_uid] ) ) // sort arguments |
---|
308 | { |
---|
309 | printf("\n[SORT ERROR] main cannot create thread %x \n", thread_uid ); |
---|
310 | exit( 0 ); |
---|
311 | } |
---|
312 | else |
---|
313 | { |
---|
314 | printf("\n[SORT] main created thread %x \n", thread_uid ); |
---|
315 | } |
---|
316 | } |
---|
317 | |
---|
318 | #if INTERACTIVE_MODE |
---|
319 | idbg(); |
---|
320 | #endif |
---|
321 | } |
---|
322 | } |
---|
323 | } |
---|
324 | |
---|
325 | get_cycle( &cycle ); |
---|
326 | printf("\n[SORT] main completes threads create at cycle %d\n", (unsigned int)cycle ); |
---|
327 | |
---|
328 | #if INTERACTIVE_MODE |
---|
329 | idbg(); |
---|
330 | #endif |
---|
331 | |
---|
332 | // the main thread run also the sort() function |
---|
333 | sort( &arg[main_uid] ); |
---|
334 | |
---|
335 | // Check result |
---|
336 | int success = 1; |
---|
337 | int* res_array = ( (threads== 2) || |
---|
338 | (threads== 8) || |
---|
339 | (threads== 32) || |
---|
340 | (threads==128) || |
---|
341 | (threads==512) ) ? array1 : array0; |
---|
342 | |
---|
343 | for( n=0 ; n<(ARRAY_LENGTH-2) ; n++ ) |
---|
344 | { |
---|
345 | if ( res_array[n] > res_array[n+1] ) |
---|
346 | { |
---|
347 | printf("\n[SORT] array[%d] = %d > array[%d] = %d\n", |
---|
348 | n , res_array[n] , n+1 , res_array[n+1] ); |
---|
349 | success = 0; |
---|
350 | break; |
---|
351 | } |
---|
352 | } |
---|
353 | |
---|
354 | #if DISPLAY_ARRAY |
---|
355 | printf("\n*** array after sort\n"); |
---|
356 | for( n=0; n<ARRAY_LENGTH; n++) printf("array[%d] = %d\n", n , res_array[n] ); |
---|
357 | #endif |
---|
358 | |
---|
359 | get_cycle( &cycle ); |
---|
360 | |
---|
361 | if ( success ) |
---|
362 | { |
---|
363 | printf("\n[SORT] success at cycle %d\n", (unsigned int)cycle ); |
---|
364 | exit( 0 ); |
---|
365 | } |
---|
366 | else |
---|
367 | { |
---|
368 | printf("\n[SORT] failure at cycle %d\n", (unsigned int)cycle ); |
---|
369 | exit( 1 ); |
---|
370 | } |
---|
371 | |
---|
372 | } // end main() |
---|
373 | |
---|
374 | |
---|
375 | /* |
---|
376 | vim: tabstop=4 : shiftwidth=4 : expandtab |
---|
377 | */ |
---|