1 | /////////////////////////////////////////////////////////////////////////////////////// |
---|
2 | // File : convol.c |
---|
3 | // Date : june 2014 |
---|
4 | // author : Alain Greiner |
---|
5 | /////////////////////////////////////////////////////////////////////////////////////// |
---|
6 | // This multi-threaded application implements a 2D convolution product. |
---|
7 | // It can run on a multi-processors, multi-clusters architecture, with one thread |
---|
8 | // per processor, and uses the POSIX threads API. |
---|
9 | // |
---|
10 | // The main() function can be launched on any processor P[x,y,l]. |
---|
11 | // It makes the initialisations, launch (N-1) threads to run the execute() function |
---|
12 | // on the (N-1) other processors than P[x,y,l], call himself the execute() function, |
---|
13 | // and finally call the instrument() function to display instrumentation results |
---|
14 | // when the parallel execution is completed. |
---|
15 | // |
---|
16 | // The convolution kernel is [201]*[35] pixels, but it can be factored in two |
---|
17 | // independant line and column convolution products. |
---|
18 | // The five buffers containing the image are distributed in clusters. |
---|
19 | // |
---|
20 | // The (1024 * 1024) pixels image is read from a file (2 bytes per pixel). |
---|
21 | // |
---|
22 | // - number of clusters containing processors must be power of 2 no larger than 256. |
---|
23 | // - number of processors per cluster must be power of 2 no larger than 8. |
---|
24 | /////////////////////////////////////////////////////////////////////////////////////// |
---|
25 | |
---|
26 | #include "stdio.h" |
---|
27 | #include "stdlib.h" |
---|
28 | #include "user_barrier.h" |
---|
29 | #include "malloc.h" |
---|
30 | |
---|
31 | #define USE_SQT_BARRIER 1 |
---|
32 | #define VERBOSE 1 |
---|
33 | #define SUPER_VERBOSE 0 |
---|
34 | |
---|
35 | #define X_SIZE_MAX 16 |
---|
36 | #define Y_SIZE_MAX 16 |
---|
37 | #define PROCS_MAX 8 |
---|
38 | #define CLUSTERS_MAX (X_SIZE_MAX * Y_SIZE_MAX) |
---|
39 | |
---|
40 | #define INITIAL_DISPLAY_ENABLE 0 |
---|
41 | #define FINAL_DISPLAY_ENABLE 1 |
---|
42 | |
---|
43 | #define PIXEL_SIZE 2 |
---|
44 | #define NL 1024 |
---|
45 | #define NP 1024 |
---|
46 | #define NB_PIXELS (NP * NL) |
---|
47 | #define FRAME_SIZE (NB_PIXELS * PIXEL_SIZE) |
---|
48 | |
---|
49 | #define SEEK_SET 0 |
---|
50 | |
---|
51 | #define TA(c,l,p) (A[c][((NP) * (l)) + (p)]) |
---|
52 | #define TB(c,p,l) (B[c][((NL) * (p)) + (l)]) |
---|
53 | #define TC(c,l,p) (C[c][((NP) * (l)) + (p)]) |
---|
54 | #define TD(c,l,p) (D[c][((NP) * (l)) + (p)]) |
---|
55 | #define TZ(c,l,p) (Z[c][((NP) * (l)) + (p)]) |
---|
56 | |
---|
57 | #define max(x,y) ((x) > (y) ? (x) : (y)) |
---|
58 | #define min(x,y) ((x) < (y) ? (x) : (y)) |
---|
59 | |
---|
60 | // macro to use a shared TTY |
---|
61 | #define printf(...); { lock_acquire( &tty_lock ); \ |
---|
62 | giet_tty_printf(__VA_ARGS__); \ |
---|
63 | lock_release( &tty_lock ); } |
---|
64 | |
---|
65 | ////////////////////////////////////////////////////////// |
---|
66 | // global variables stored in seg_data in cluster[0,0] |
---|
67 | ////////////////////////////////////////////////////////// |
---|
68 | |
---|
69 | // Instrumentation counters (cluster_id, lpid] |
---|
70 | |
---|
71 | unsigned int START[CLUSTERS_MAX][PROCS_MAX]; |
---|
72 | unsigned int H_BEG[CLUSTERS_MAX][PROCS_MAX]; |
---|
73 | unsigned int H_END[CLUSTERS_MAX][PROCS_MAX]; |
---|
74 | unsigned int V_BEG[CLUSTERS_MAX][PROCS_MAX]; |
---|
75 | unsigned int V_END[CLUSTERS_MAX][PROCS_MAX]; |
---|
76 | unsigned int D_BEG[CLUSTERS_MAX][PROCS_MAX]; |
---|
77 | unsigned int D_END[CLUSTERS_MAX][PROCS_MAX]; |
---|
78 | |
---|
79 | // global synchronization barrier |
---|
80 | |
---|
81 | #if USE_SQT_BARRIER |
---|
82 | giet_sqt_barrier_t barrier; |
---|
83 | #else |
---|
84 | giet_barrier_t barrier; |
---|
85 | #endif |
---|
86 | |
---|
87 | volatile unsigned int barrier_init_ok = 0; |
---|
88 | volatile unsigned int load_image_ok = 0; |
---|
89 | volatile unsigned int instrumentation_ok = 0; |
---|
90 | |
---|
91 | // lock protecting access to shared TTY |
---|
92 | user_lock_t tty_lock; |
---|
93 | |
---|
94 | // global pointers on distributed buffers in all clusters |
---|
95 | unsigned short * GA[CLUSTERS_MAX]; |
---|
96 | int * GB[CLUSTERS_MAX]; |
---|
97 | int * GC[CLUSTERS_MAX]; |
---|
98 | int * GD[CLUSTERS_MAX]; |
---|
99 | unsigned char * GZ[CLUSTERS_MAX]; |
---|
100 | |
---|
101 | |
---|
102 | |
---|
103 | //////////////////////////////////////////// |
---|
104 | __attribute__ ((constructor)) void execute() |
---|
105 | //////////////////////////////////////////// |
---|
106 | { |
---|
107 | ///////////////////////////////////////////////////////////////////////////////////// |
---|
108 | // Each thread[x,y,p] initialises the convolution kernel parameters in local stack. |
---|
109 | // The values defined in the next 12 lines are Philips proprietary information. |
---|
110 | ///////////////////////////////////////////////////////////////////////////////////// |
---|
111 | |
---|
112 | int vnorm = 115; |
---|
113 | int vf[35] = { 1, 1, 2, 2, 2, |
---|
114 | 2, 3, 3, 3, 4, |
---|
115 | 4, 4, 4, 5, 5, |
---|
116 | 5, 5, 5, 5, 5, |
---|
117 | 5, 5, 4, 4, 4, |
---|
118 | 4, 3, 3, 3, 2, |
---|
119 | 2, 2, 2, 1, 1 }; |
---|
120 | |
---|
121 | int hrange = 100; |
---|
122 | int hnorm = 201; |
---|
123 | |
---|
124 | // get plat-form parameters |
---|
125 | unsigned int x_size; // number of clusters in a row |
---|
126 | unsigned int y_size; // number of clusters in a column |
---|
127 | unsigned int nprocs; // number of processors per cluster |
---|
128 | giet_procs_number( &x_size , &y_size , &nprocs ); |
---|
129 | |
---|
130 | // get processor identifiers |
---|
131 | unsigned int x; // x coordinate |
---|
132 | unsigned int y; // y coordinate |
---|
133 | unsigned int lpid; // local proc index |
---|
134 | giet_proc_xyp( &x, &y, &lpid ); |
---|
135 | |
---|
136 | // indexes for loops |
---|
137 | int c; // cluster index |
---|
138 | int l; // line index |
---|
139 | int p; // pixel index |
---|
140 | int z; // vertical filter index |
---|
141 | |
---|
142 | int file = 0; // file descriptor |
---|
143 | unsigned int nclusters = x_size * y_size; // number of clusters |
---|
144 | unsigned int cluster_id = (x * y_size) + y; // continuous cluster index |
---|
145 | unsigned int thread_id = (cluster_id * nprocs) + lpid; // continuous thread index |
---|
146 | unsigned int nthreads = nclusters * nprocs; // number of threads |
---|
147 | unsigned int frame_size = FRAME_SIZE; // total size (bytes) |
---|
148 | unsigned int lines_per_thread = NL / nthreads; // lines per thread |
---|
149 | unsigned int lines_per_cluster = NL / nclusters; // lines per cluster |
---|
150 | unsigned int pixels_per_thread = NP / nthreads; // columns per thread |
---|
151 | unsigned int pixels_per_cluster = NP / nclusters; // columns per cluster |
---|
152 | |
---|
153 | int first, last; |
---|
154 | |
---|
155 | unsigned int date = giet_proctime(); |
---|
156 | START[cluster_id][lpid] = date; |
---|
157 | |
---|
158 | ///////////////////////////////////////////////////////////////////// |
---|
159 | // Each thread[x][y][0] allocate the global buffers in cluster(x,y) |
---|
160 | ///////////////////////////////////////////////////////////////////// |
---|
161 | if ( lpid == 0 ) |
---|
162 | { |
---|
163 | |
---|
164 | #if VERBOSE |
---|
165 | printf( "\n[CONVOL] thread[%d,%d,%d] enters malloc at cycle %d\n", |
---|
166 | x,y,lpid, date ); |
---|
167 | #endif |
---|
168 | |
---|
169 | GA[cluster_id] = remote_malloc( (FRAME_SIZE/nclusters) , x , y ); |
---|
170 | GB[cluster_id] = remote_malloc( (FRAME_SIZE/nclusters)*2 , x , y ); |
---|
171 | GC[cluster_id] = remote_malloc( (FRAME_SIZE/nclusters)*2 , x , y ); |
---|
172 | GD[cluster_id] = remote_malloc( (FRAME_SIZE/nclusters)*2 , x , y ); |
---|
173 | GZ[cluster_id] = remote_malloc( (FRAME_SIZE/nclusters)/2 , x , y ); |
---|
174 | |
---|
175 | #if VERBOSE |
---|
176 | printf( "\n[CONVOL] Shared Buffer Virtual Addresses in cluster(%d,%d)\n" |
---|
177 | "### GA = %x\n" |
---|
178 | "### GB = %x\n" |
---|
179 | "### GC = %x\n" |
---|
180 | "### GD = %x\n" |
---|
181 | "### GZ = %x\n", |
---|
182 | x, y, |
---|
183 | GA[cluster_id], |
---|
184 | GB[cluster_id], |
---|
185 | GC[cluster_id], |
---|
186 | GD[cluster_id], |
---|
187 | GZ[cluster_id] ); |
---|
188 | #endif |
---|
189 | } |
---|
190 | |
---|
191 | /////////////////////////////// |
---|
192 | #if USE_SQT_BARRIER |
---|
193 | sqt_barrier_wait( &barrier ); |
---|
194 | #else |
---|
195 | barrier_wait( &barrier ); |
---|
196 | #endif |
---|
197 | |
---|
198 | ////////////////////////////////////////////////////////////////////// |
---|
199 | // Each thread[x,y,p] initialise in its private stack a copy of the |
---|
200 | // arrays of pointers on the shared, distributed buffers. |
---|
201 | ////////////////////////////////////////////////////////////////////// |
---|
202 | |
---|
203 | unsigned short * A[CLUSTERS_MAX]; |
---|
204 | int * B[CLUSTERS_MAX]; |
---|
205 | int * C[CLUSTERS_MAX]; |
---|
206 | int * D[CLUSTERS_MAX]; |
---|
207 | unsigned char * Z[CLUSTERS_MAX]; |
---|
208 | |
---|
209 | for (c = 0; c < nclusters; c++) |
---|
210 | { |
---|
211 | A[c] = GA[c]; |
---|
212 | B[c] = GB[c]; |
---|
213 | C[c] = GC[c]; |
---|
214 | D[c] = GD[c]; |
---|
215 | Z[c] = GZ[c]; |
---|
216 | } |
---|
217 | |
---|
218 | ///////////////////////////////////////////////////////////////////////////// |
---|
219 | // Ech thread[x,y,0] open the file containing image, and load it from disk |
---|
220 | // to the local A[c] buffer (frame_size / nclusters loaded in each cluster). |
---|
221 | // Other threads are waiting on the init_ok condition. |
---|
222 | //////////////////////////////////////////////////////////////////////////// |
---|
223 | if ( lpid==0 ) |
---|
224 | { |
---|
225 | // open file |
---|
226 | file = giet_fat_open( "/misc/philips_1024.raw" , 0 ); |
---|
227 | if (file < 0 ) |
---|
228 | { |
---|
229 | printf("\n[CONVOL ERROR] thread[%d,%d,%d] " |
---|
230 | "cannot open file /misc/philips_1024.raw", |
---|
231 | x, y, lpid ); |
---|
232 | giet_pthread_exit( NULL ); |
---|
233 | } |
---|
234 | |
---|
235 | #if VERBOSE |
---|
236 | printf( "\n[CONVOL] thread[%d,%d,%d] open file /misc/philips_1024.raw at cycle %d\n", |
---|
237 | x, y, lpid, giet_proctime() ); |
---|
238 | #endif |
---|
239 | |
---|
240 | unsigned int offset = (frame_size/nclusters)*cluster_id; |
---|
241 | unsigned int size = frame_size/nclusters; |
---|
242 | |
---|
243 | if ( giet_fat_lseek( file, |
---|
244 | offset, |
---|
245 | SEEK_SET ) ) |
---|
246 | { |
---|
247 | printf("\n[CONVOL ERROR] thread[%d,%d,%d] " |
---|
248 | "cannot seek file /misc/philips_1024.raw", |
---|
249 | x, y, lpid ); |
---|
250 | giet_pthread_exit( NULL ); |
---|
251 | } |
---|
252 | |
---|
253 | if ( giet_fat_read( file, |
---|
254 | A[cluster_id], |
---|
255 | size ) != size ) |
---|
256 | { |
---|
257 | printf("\n[CONVOL ERROR] thread[%d,%d,%d] " |
---|
258 | "cannot read file /misc/philips_1024.raw", |
---|
259 | x, y, lpid ); |
---|
260 | giet_pthread_exit( NULL ); |
---|
261 | } |
---|
262 | |
---|
263 | #if VERBOSE |
---|
264 | printf( "\n[CONVOL] thread[%d,%d,%d] load file /misc/philips_1024.raw at cycle %d\n", |
---|
265 | x, y, lpid, giet_proctime() ); |
---|
266 | #endif |
---|
267 | |
---|
268 | } |
---|
269 | |
---|
270 | /////////////////////////////// |
---|
271 | #if USE_SQT_BARRIER |
---|
272 | sqt_barrier_wait( &barrier ); |
---|
273 | #else |
---|
274 | barrier_wait( &barrier ); |
---|
275 | #endif |
---|
276 | |
---|
277 | ///////////////////////////////////////////////////////////////////////////// |
---|
278 | // Optionnal parallel display of the initial image stored in A[c] buffers. |
---|
279 | // Eah thread[x,y,p] displays (NL/nthreads) lines. (one byte per pixel). |
---|
280 | ///////////////////////////////////////////////////////////////////////////// |
---|
281 | |
---|
282 | if ( INITIAL_DISPLAY_ENABLE ) |
---|
283 | { |
---|
284 | |
---|
285 | #if VERBOSE |
---|
286 | printf( "\n[CONVOL] thread[%d,%d,%d] starts initial display at cycle %d\n", |
---|
287 | x, y, lpid, giet_proctime() ); |
---|
288 | #endif |
---|
289 | |
---|
290 | unsigned int line; |
---|
291 | unsigned int offset = lines_per_thread * lpid; |
---|
292 | |
---|
293 | for ( l = 0 ; l < lines_per_thread ; l++ ) |
---|
294 | { |
---|
295 | line = offset + l; |
---|
296 | |
---|
297 | for ( p = 0 ; p < NP ; p++ ) |
---|
298 | { |
---|
299 | TZ(cluster_id, line, p) = (unsigned char)(TA(cluster_id, line, p) >> 8); |
---|
300 | } |
---|
301 | |
---|
302 | giet_fbf_sync_write( NP*(l + (thread_id * lines_per_thread) ), |
---|
303 | &TZ(cluster_id, line, 0), |
---|
304 | NP); |
---|
305 | } |
---|
306 | |
---|
307 | #if VERBOSE |
---|
308 | printf( "\n[CONVOL] thread[%d,%d,%d] completes initial display at cycle %d\n", |
---|
309 | x, y, lpid, giet_proctime() ); |
---|
310 | #endif |
---|
311 | |
---|
312 | //////////////////////////// |
---|
313 | #if USE_SQT_BARRIER |
---|
314 | sqt_barrier_wait( &barrier ); |
---|
315 | #else |
---|
316 | barrier_wait( &barrier ); |
---|
317 | #endif |
---|
318 | |
---|
319 | } |
---|
320 | |
---|
321 | //////////////////////////////////////////////////////// |
---|
322 | // parallel horizontal filter : |
---|
323 | // B <= transpose(FH(A)) |
---|
324 | // D <= A - FH(A) |
---|
325 | // Each thread computes (NL/nthreads) lines |
---|
326 | // The image must be extended : |
---|
327 | // if (z<0) TA(cluster_id,l,z) == TA(cluster_id,l,0) |
---|
328 | // if (z>NP-1) TA(cluster_id,l,z) == TA(cluster_id,l,NP-1) |
---|
329 | //////////////////////////////////////////////////////// |
---|
330 | |
---|
331 | date = giet_proctime(); |
---|
332 | H_BEG[cluster_id][lpid] = date; |
---|
333 | |
---|
334 | #if VERBOSE |
---|
335 | printf( "\n[CONVOL] thread[%d,%d,%d] starts horizontal filter" |
---|
336 | " at cycle %d\n", |
---|
337 | x, y, lpid, date ); |
---|
338 | #else |
---|
339 | if ( (x==0) && (y==0) && (lpid==0) ) |
---|
340 | printf( "\n[CONVOL] thread[0,0,0] starts horizontal filter" |
---|
341 | " at cycle %d\n", date ); |
---|
342 | #endif |
---|
343 | |
---|
344 | // l = absolute line index / p = absolute pixel index |
---|
345 | // first & last define which lines are handled by a given thread |
---|
346 | |
---|
347 | first = thread_id * lines_per_thread; |
---|
348 | last = first + lines_per_thread; |
---|
349 | |
---|
350 | for (l = first; l < last; l++) |
---|
351 | { |
---|
352 | // src_c and src_l are the cluster index and the line index for A & D |
---|
353 | int src_c = l / lines_per_cluster; |
---|
354 | int src_l = l % lines_per_cluster; |
---|
355 | |
---|
356 | // We use the specific values of the horizontal ep-filter for optimisation: |
---|
357 | // sum(p) = sum(p-1) + TA[p+hrange] - TA[p-hrange-1] |
---|
358 | // To minimize the number of tests, the loop on pixels is split in three domains |
---|
359 | |
---|
360 | int sum_p = (hrange + 2) * TA(src_c, src_l, 0); |
---|
361 | for (z = 1; z < hrange; z++) |
---|
362 | { |
---|
363 | sum_p = sum_p + TA(src_c, src_l, z); |
---|
364 | } |
---|
365 | |
---|
366 | // first domain : from 0 to hrange |
---|
367 | for (p = 0; p < hrange + 1; p++) |
---|
368 | { |
---|
369 | // dst_c and dst_p are the cluster index and the pixel index for B |
---|
370 | int dst_c = p / pixels_per_cluster; |
---|
371 | int dst_p = p % pixels_per_cluster; |
---|
372 | sum_p = sum_p + (int) TA(src_c, src_l, p + hrange) - (int) TA(src_c, src_l, 0); |
---|
373 | TB(dst_c, dst_p, l) = sum_p / hnorm; |
---|
374 | TD(src_c, src_l, p) = (int) TA(src_c, src_l, p) - sum_p / hnorm; |
---|
375 | } |
---|
376 | // second domain : from (hrange+1) to (NP-hrange-1) |
---|
377 | for (p = hrange + 1; p < NP - hrange; p++) |
---|
378 | { |
---|
379 | // dst_c and dst_p are the cluster index and the pixel index for B |
---|
380 | int dst_c = p / pixels_per_cluster; |
---|
381 | int dst_p = p % pixels_per_cluster; |
---|
382 | sum_p = sum_p + (int) TA(src_c, src_l, p + hrange) |
---|
383 | - (int) TA(src_c, src_l, p - hrange - 1); |
---|
384 | TB(dst_c, dst_p, l) = sum_p / hnorm; |
---|
385 | TD(src_c, src_l, p) = (int) TA(src_c, src_l, p) - sum_p / hnorm; |
---|
386 | } |
---|
387 | // third domain : from (NP-hrange) to (NP-1) |
---|
388 | for (p = NP - hrange; p < NP; p++) |
---|
389 | { |
---|
390 | // dst_c and dst_p are the cluster index and the pixel index for B |
---|
391 | int dst_c = p / pixels_per_cluster; |
---|
392 | int dst_p = p % pixels_per_cluster; |
---|
393 | sum_p = sum_p + (int) TA(src_c, src_l, NP - 1) |
---|
394 | - (int) TA(src_c, src_l, p - hrange - 1); |
---|
395 | TB(dst_c, dst_p, l) = sum_p / hnorm; |
---|
396 | TD(src_c, src_l, p) = (int) TA(src_c, src_l, p) - sum_p / hnorm; |
---|
397 | } |
---|
398 | |
---|
399 | #if SUPER_VERBOSE |
---|
400 | printf(" - line %d computed at cycle %d\n", l, giet_proctime() ); |
---|
401 | #endif |
---|
402 | |
---|
403 | } |
---|
404 | |
---|
405 | date = giet_proctime(); |
---|
406 | H_END[cluster_id][lpid] = date; |
---|
407 | |
---|
408 | #if VERBOSE |
---|
409 | printf( "\n[CONVOL] thread[%d,%d,%d] completes horizontal filter" |
---|
410 | " at cycle %d\n", |
---|
411 | x, y, lpid, date ); |
---|
412 | #else |
---|
413 | if ( (x==0) && (y==0) && (lpid==0) ) |
---|
414 | printf( "\n[CONVOL] thread[0,0,0] completes horizontal filter" |
---|
415 | " at cycle %d\n", date ); |
---|
416 | #endif |
---|
417 | |
---|
418 | ///////////////////////////// |
---|
419 | #if USE_SQT_BARRIER |
---|
420 | sqt_barrier_wait( &barrier ); |
---|
421 | #else |
---|
422 | barrier_wait( &barrier ); |
---|
423 | #endif |
---|
424 | |
---|
425 | |
---|
426 | /////////////////////////////////////////////////////////////// |
---|
427 | // parallel vertical filter : |
---|
428 | // C <= transpose(FV(B)) |
---|
429 | // Each thread computes (NP/nthreads) columns |
---|
430 | // The image must be extended : |
---|
431 | // if (l<0) TB(cluster_id,p,l) == TB(cluster_id,p,0) |
---|
432 | // if (l>NL-1) TB(cluster_id,p,l) == TB(cluster_id,p,NL-1) |
---|
433 | /////////////////////////////////////////////////////////////// |
---|
434 | |
---|
435 | date = giet_proctime(); |
---|
436 | V_BEG[cluster_id][lpid] = date; |
---|
437 | |
---|
438 | #if VERBOSE |
---|
439 | printf( "\n[CONVOL] thread[%d,%d,%d] starts vertical filter" |
---|
440 | " at cycle %d\n", |
---|
441 | x, y, lpid, date ); |
---|
442 | #else |
---|
443 | if ( (x==0) && (y==0) && (lpid==0) ) |
---|
444 | printf( "\n[CONVOL] thread[0,0,0] starts vertical filter" |
---|
445 | " at cycle %d\n", date ); |
---|
446 | #endif |
---|
447 | |
---|
448 | // l = absolute line index / p = absolute pixel index |
---|
449 | // first & last define which pixels are handled by a given thread |
---|
450 | |
---|
451 | first = thread_id * pixels_per_thread; |
---|
452 | last = first + pixels_per_thread; |
---|
453 | |
---|
454 | for (p = first; p < last; p++) |
---|
455 | { |
---|
456 | // src_c and src_p are the cluster index and the pixel index for B |
---|
457 | int src_c = p / pixels_per_cluster; |
---|
458 | int src_p = p % pixels_per_cluster; |
---|
459 | |
---|
460 | int sum_l; |
---|
461 | |
---|
462 | // We use the specific values of the vertical ep-filter |
---|
463 | // To minimize the number of tests, the NL lines are split in three domains |
---|
464 | |
---|
465 | // first domain : explicit computation for the first 18 values |
---|
466 | for (l = 0; l < 18; l++) |
---|
467 | { |
---|
468 | // dst_c and dst_l are the cluster index and the line index for C |
---|
469 | int dst_c = l / lines_per_cluster; |
---|
470 | int dst_l = l % lines_per_cluster; |
---|
471 | |
---|
472 | for (z = 0, sum_l = 0; z < 35; z++) |
---|
473 | { |
---|
474 | sum_l = sum_l + vf[z] * TB(src_c, src_p, max(l - 17 + z,0) ); |
---|
475 | } |
---|
476 | TC(dst_c, dst_l, p) = sum_l / vnorm; |
---|
477 | } |
---|
478 | // second domain |
---|
479 | for (l = 18; l < NL - 17; l++) |
---|
480 | { |
---|
481 | // dst_c and dst_l are the cluster index and the line index for C |
---|
482 | int dst_c = l / lines_per_cluster; |
---|
483 | int dst_l = l % lines_per_cluster; |
---|
484 | |
---|
485 | sum_l = sum_l + TB(src_c, src_p, l + 4) |
---|
486 | + TB(src_c, src_p, l + 8) |
---|
487 | + TB(src_c, src_p, l + 11) |
---|
488 | + TB(src_c, src_p, l + 15) |
---|
489 | + TB(src_c, src_p, l + 17) |
---|
490 | - TB(src_c, src_p, l - 5) |
---|
491 | - TB(src_c, src_p, l - 9) |
---|
492 | - TB(src_c, src_p, l - 12) |
---|
493 | - TB(src_c, src_p, l - 16) |
---|
494 | - TB(src_c, src_p, l - 18); |
---|
495 | |
---|
496 | TC(dst_c, dst_l, p) = sum_l / vnorm; |
---|
497 | } |
---|
498 | // third domain |
---|
499 | for (l = NL - 17; l < NL; l++) |
---|
500 | { |
---|
501 | // dst_c and dst_l are the cluster index and the line index for C |
---|
502 | int dst_c = l / lines_per_cluster; |
---|
503 | int dst_l = l % lines_per_cluster; |
---|
504 | |
---|
505 | sum_l = sum_l + TB(src_c, src_p, min(l + 4, NL - 1)) |
---|
506 | + TB(src_c, src_p, min(l + 8, NL - 1)) |
---|
507 | + TB(src_c, src_p, min(l + 11, NL - 1)) |
---|
508 | + TB(src_c, src_p, min(l + 15, NL - 1)) |
---|
509 | + TB(src_c, src_p, min(l + 17, NL - 1)) |
---|
510 | - TB(src_c, src_p, l - 5) |
---|
511 | - TB(src_c, src_p, l - 9) |
---|
512 | - TB(src_c, src_p, l - 12) |
---|
513 | - TB(src_c, src_p, l - 16) |
---|
514 | - TB(src_c, src_p, l - 18); |
---|
515 | |
---|
516 | TC(dst_c, dst_l, p) = sum_l / vnorm; |
---|
517 | } |
---|
518 | |
---|
519 | #if SUPER_VERBOSE |
---|
520 | printf(" - column %d computed at cycle %d\n", p, giet_proctime()); |
---|
521 | #endif |
---|
522 | |
---|
523 | } |
---|
524 | |
---|
525 | date = giet_proctime(); |
---|
526 | V_END[cluster_id][lpid] = date; |
---|
527 | |
---|
528 | #if VERBOSE |
---|
529 | printf( "\n[CONVOL] thread[%d,%d,%d] completes vertical filter" |
---|
530 | " at cycle %d\n", |
---|
531 | x, y, lpid, date ); |
---|
532 | #else |
---|
533 | if ( (x==0) && (y==0) && (lpid==0) ) |
---|
534 | printf( "\n[CONVOL] thread[0,0,0] completes vertical filter" |
---|
535 | " at cycle %d\n", date ); |
---|
536 | #endif |
---|
537 | |
---|
538 | //////////////////////////// |
---|
539 | #if USE_SQT_BARRIER |
---|
540 | sqt_barrier_wait( &barrier ); |
---|
541 | #else |
---|
542 | barrier_wait( &barrier ); |
---|
543 | #endif |
---|
544 | |
---|
545 | //////////////////////////////////////////////////////////////////////// |
---|
546 | // Optional parallel display of the final image Z <= D + C |
---|
547 | // Eah thread[x,y,p] displays (NL/nthreads) lines. (one byte per pixel). |
---|
548 | //////////////////////////////////////////////////////////////////////// |
---|
549 | |
---|
550 | if ( FINAL_DISPLAY_ENABLE ) |
---|
551 | { |
---|
552 | date = giet_proctime(); |
---|
553 | D_BEG[cluster_id][lpid] = date; |
---|
554 | |
---|
555 | #if VERBOSE |
---|
556 | printf( "\n[CONVOL] thread[%d,%d,%d] starts final display" |
---|
557 | " at cycle %d\n", |
---|
558 | x, y, lpid, date); |
---|
559 | #else |
---|
560 | if ( (x==0) && (y==0) && (lpid==0) ) |
---|
561 | printf( "\n[CONVOL] thread[0,0,0] starts final display" |
---|
562 | " at cycle %d\n", date ); |
---|
563 | #endif |
---|
564 | |
---|
565 | unsigned int line; |
---|
566 | unsigned int offset = lines_per_thread * lpid; |
---|
567 | |
---|
568 | for ( l = 0 ; l < lines_per_thread ; l++ ) |
---|
569 | { |
---|
570 | line = offset + l; |
---|
571 | |
---|
572 | for ( p = 0 ; p < NP ; p++ ) |
---|
573 | { |
---|
574 | TZ(cluster_id, line, p) = |
---|
575 | (unsigned char)( (TD(cluster_id, line, p) + |
---|
576 | TC(cluster_id, line, p) ) >> 8 ); |
---|
577 | } |
---|
578 | |
---|
579 | giet_fbf_sync_write( NP*(l + (thread_id * lines_per_thread) ), |
---|
580 | &TZ(cluster_id, line, 0), |
---|
581 | NP); |
---|
582 | } |
---|
583 | |
---|
584 | date = giet_proctime(); |
---|
585 | D_END[cluster_id][lpid] = date; |
---|
586 | |
---|
587 | #if VERBOSE |
---|
588 | printf( "\n[CONVOL] thread[%d,%d,%d] completes final display" |
---|
589 | " at cycle %d\n", |
---|
590 | x, y, lpid, date); |
---|
591 | #else |
---|
592 | if ( (x==0) && (y==0) && (lpid==0) ) |
---|
593 | printf( "\n[CONVOL] thread[0,0,0] completes final display" |
---|
594 | " at cycle %d\n", date ); |
---|
595 | #endif |
---|
596 | |
---|
597 | ////////////////////////////// |
---|
598 | #if USE_SQT_BARRIER |
---|
599 | sqt_barrier_wait( &barrier ); |
---|
600 | #else |
---|
601 | barrier_wait( &barrier ); |
---|
602 | #endif |
---|
603 | |
---|
604 | } |
---|
605 | |
---|
606 | // all threads (but the one executing main) exit |
---|
607 | if ( (x!=0) || (y!=0) || (lpid!=0) ) |
---|
608 | { |
---|
609 | giet_pthread_exit( "completed"); |
---|
610 | } |
---|
611 | |
---|
612 | } // end execute() |
---|
613 | |
---|
614 | |
---|
615 | |
---|
616 | ///////////////////////////////////////// |
---|
617 | void instrument( unsigned int nclusters, |
---|
618 | unsigned int nprocs ) |
---|
619 | ///////////////////////////////////////// |
---|
620 | { |
---|
621 | int cc, pp; |
---|
622 | |
---|
623 | unsigned int min_start = 0xFFFFFFFF; |
---|
624 | unsigned int max_start = 0; |
---|
625 | |
---|
626 | unsigned int min_h_beg = 0xFFFFFFFF; |
---|
627 | unsigned int max_h_beg = 0; |
---|
628 | |
---|
629 | unsigned int min_h_end = 0xFFFFFFFF; |
---|
630 | unsigned int max_h_end = 0; |
---|
631 | |
---|
632 | unsigned int min_v_beg = 0xFFFFFFFF; |
---|
633 | unsigned int max_v_beg = 0; |
---|
634 | |
---|
635 | unsigned int min_v_end = 0xFFFFFFFF; |
---|
636 | unsigned int max_v_end = 0; |
---|
637 | |
---|
638 | unsigned int min_d_beg = 0xFFFFFFFF; |
---|
639 | unsigned int max_d_beg = 0; |
---|
640 | |
---|
641 | unsigned int min_d_end = 0xFFFFFFFF; |
---|
642 | unsigned int max_d_end = 0; |
---|
643 | |
---|
644 | for (cc = 0; cc < nclusters; cc++) |
---|
645 | { |
---|
646 | for (pp = 0; pp < nprocs; pp++ ) |
---|
647 | { |
---|
648 | if (START[cc][pp] < min_start) min_start = START[cc][pp]; |
---|
649 | if (START[cc][pp] > max_start) max_start = START[cc][pp]; |
---|
650 | |
---|
651 | if (H_BEG[cc][pp] < min_h_beg) min_h_beg = H_BEG[cc][pp]; |
---|
652 | if (H_BEG[cc][pp] > max_h_beg) max_h_beg = H_BEG[cc][pp]; |
---|
653 | |
---|
654 | if (H_END[cc][pp] < min_h_end) min_h_end = H_END[cc][pp]; |
---|
655 | if (H_END[cc][pp] > max_h_end) max_h_end = H_END[cc][pp]; |
---|
656 | |
---|
657 | if (V_BEG[cc][pp] < min_v_beg) min_v_beg = V_BEG[cc][pp]; |
---|
658 | if (V_BEG[cc][pp] > max_v_beg) max_v_beg = V_BEG[cc][pp]; |
---|
659 | |
---|
660 | if (V_END[cc][pp] < min_v_end) min_v_end = V_END[cc][pp]; |
---|
661 | if (V_END[cc][pp] > max_v_end) max_v_end = V_END[cc][pp]; |
---|
662 | |
---|
663 | if (D_BEG[cc][pp] < min_d_beg) min_d_beg = D_BEG[cc][pp]; |
---|
664 | if (D_BEG[cc][pp] > max_d_beg) max_d_beg = D_BEG[cc][pp]; |
---|
665 | |
---|
666 | if (D_END[cc][pp] < min_d_end) min_d_end = D_END[cc][pp]; |
---|
667 | if (D_END[cc][pp] > max_d_end) max_d_end = D_END[cc][pp]; |
---|
668 | } |
---|
669 | } |
---|
670 | |
---|
671 | printf(" - START : min = %d / max = %d / med = %d / delta = %d\n", |
---|
672 | min_start, max_start, (min_start+max_start)/2, max_start-min_start); |
---|
673 | |
---|
674 | printf(" - H_BEG : min = %d / max = %d / med = %d / delta = %d\n", |
---|
675 | min_h_beg, max_h_beg, (min_h_beg+max_h_beg)/2, max_h_beg-min_h_beg); |
---|
676 | |
---|
677 | printf(" - H_END : min = %d / max = %d / med = %d / delta = %d\n", |
---|
678 | min_h_end, max_h_end, (min_h_end+max_h_end)/2, max_h_end-min_h_end); |
---|
679 | |
---|
680 | printf(" - V_BEG : min = %d / max = %d / med = %d / delta = %d\n", |
---|
681 | min_v_beg, max_v_beg, (min_v_beg+max_v_beg)/2, max_v_beg-min_v_beg); |
---|
682 | |
---|
683 | printf(" - V_END : min = %d / max = %d / med = %d / delta = %d\n", |
---|
684 | min_v_end, max_v_end, (min_v_end+max_v_end)/2, max_v_end-min_v_end); |
---|
685 | |
---|
686 | printf(" - D_BEG : min = %d / max = %d / med = %d / delta = %d\n", |
---|
687 | min_d_beg, max_d_beg, (min_d_beg+max_d_beg)/2, max_d_beg-min_d_beg); |
---|
688 | |
---|
689 | printf(" - D_END : min = %d / max = %d / med = %d / delta = %d\n", |
---|
690 | min_d_end, max_d_end, (min_d_end+max_d_end)/2, max_d_end-min_d_end); |
---|
691 | |
---|
692 | printf( "\n General Scenario (Kcycles for each step)\n" ); |
---|
693 | printf( " - BOOT OS = %d\n", (min_start )/1000 ); |
---|
694 | printf( " - LOAD IMAGE = %d\n", (min_h_beg - min_start)/1000 ); |
---|
695 | printf( " - H_FILTER = %d\n", (max_h_end - min_h_beg)/1000 ); |
---|
696 | printf( " - BARRIER HORI/VERT = %d\n", (min_v_beg - max_h_end)/1000 ); |
---|
697 | printf( " - V_FILTER = %d\n", (max_v_end - min_v_beg)/1000 ); |
---|
698 | printf( " - BARRIER VERT/DISP = %d\n", (min_d_beg - max_v_end)/1000 ); |
---|
699 | printf( " - DISPLAY = %d\n", (max_d_end - min_d_beg)/1000 ); |
---|
700 | |
---|
701 | } // end instrument() |
---|
702 | |
---|
703 | |
---|
704 | |
---|
705 | /////////////////////////////////////////// |
---|
706 | __attribute__ ((constructor)) void main() |
---|
707 | /////////////////////////////////////////// |
---|
708 | { |
---|
709 | // get plat-form parameters |
---|
710 | unsigned int x_size; // number of clusters in a row |
---|
711 | unsigned int y_size; // number of clusters in a column |
---|
712 | unsigned int nprocs; // number of processors per cluster |
---|
713 | giet_procs_number( &x_size , &y_size , &nprocs ); |
---|
714 | |
---|
715 | // processor identifiers |
---|
716 | unsigned int x; // x coordinate |
---|
717 | unsigned int y; // y coordinate |
---|
718 | unsigned int lpid; // local proc index |
---|
719 | giet_proc_xyp( &x, &y, &lpid ); |
---|
720 | |
---|
721 | // indexes for loops |
---|
722 | unsigned int cx; |
---|
723 | unsigned int cy; |
---|
724 | unsigned int n; |
---|
725 | |
---|
726 | unsigned int nclusters = x_size * y_size; |
---|
727 | unsigned int nthreads = nclusters * nprocs; |
---|
728 | |
---|
729 | // get a shared TTY |
---|
730 | giet_tty_alloc( 1 ); |
---|
731 | lock_init( &tty_lock ); |
---|
732 | |
---|
733 | // get FBF size |
---|
734 | unsigned int width; |
---|
735 | unsigned int height; |
---|
736 | giet_fbf_size( &width , &height ); |
---|
737 | |
---|
738 | // parameters checking |
---|
739 | if ( (width != NP) || (height != NL) ) |
---|
740 | giet_pthread_exit( "[CONVOL ERROR] FBF size must be NP * NL\n"); |
---|
741 | if ((nprocs != 1) && (nprocs != 2) && (nprocs != 4) && (nprocs != 8)) |
---|
742 | giet_pthread_exit( "[CONVOL ERROR] NB_PROCS_MAX must be 1, 2, 4 or 8\n"); |
---|
743 | |
---|
744 | if ((x_size!=1) && (x_size!=2) && (x_size!=4) && (x_size!=8) && (x_size!=16)) |
---|
745 | giet_pthread_exit( "[CONVOL ERROR] X_SIZE must be 1, 2, 4, 8, 16\n"); |
---|
746 | |
---|
747 | if ((y_size!=1) && (y_size!=2) && (y_size!=4) && (y_size!=8) && (y_size!=16)) |
---|
748 | giet_pthread_exit( "[CONVOL ERROR] Y_SIZE must be 1, 2, 4, 8, 16\n"); |
---|
749 | |
---|
750 | if ( NL % nclusters != 0 ) |
---|
751 | giet_pthread_exit( "[CONVOL ERROR] X_SIZE*Y_SIZE must be a divider of NL"); |
---|
752 | |
---|
753 | if ( NP % nclusters != 0 ) |
---|
754 | giet_pthread_exit( "[CONVOL ERROR] X_SIZE*Y_SIZE must be a divider of NP"); |
---|
755 | |
---|
756 | // get FBF ownership |
---|
757 | giet_fbf_alloc(); |
---|
758 | |
---|
759 | // initializes the distributed heap[x,y] |
---|
760 | for ( cx = 0 ; cx < x_size ; cx++ ) |
---|
761 | { |
---|
762 | for ( cy = 0 ; cy < y_size ; cy++ ) |
---|
763 | { |
---|
764 | heap_init( cx , cy ); |
---|
765 | } |
---|
766 | } |
---|
767 | |
---|
768 | // allocate trdid[] array |
---|
769 | pthread_t* trdid = malloc( nthreads * sizeof(pthread_t) ); |
---|
770 | |
---|
771 | // barrier initialisation |
---|
772 | #if USE_SQT_BARRIER |
---|
773 | sqt_barrier_init( &barrier, x_size , y_size , nprocs ); |
---|
774 | #else |
---|
775 | barrier_init( &barrier, nthreads ); |
---|
776 | #endif |
---|
777 | |
---|
778 | printf("\n[CONVOL] thread[0,0,0] completes initialisation at cycle %d\n" |
---|
779 | "- CLUSTERS = %d\n" |
---|
780 | "- PROCS = %d\n" |
---|
781 | "- THREADS = %d\n", |
---|
782 | giet_proctime(), nclusters, nprocs, nthreads ); |
---|
783 | |
---|
784 | // launch other threads to run execute() function |
---|
785 | for ( n = 1 ; n < nthreads ; n++ ) |
---|
786 | { |
---|
787 | if ( giet_pthread_create( &trdid[n], |
---|
788 | NULL, // no attribute |
---|
789 | &execute, |
---|
790 | NULL ) ) // no argument |
---|
791 | { |
---|
792 | printf("\n[TRANSPOSE ERROR] creating thread %x\n", trdid[n] ); |
---|
793 | giet_pthread_exit( NULL ); |
---|
794 | } |
---|
795 | } |
---|
796 | |
---|
797 | // run the execute() function |
---|
798 | execute(); |
---|
799 | |
---|
800 | // wait other threads completion |
---|
801 | for ( n = 1 ; n < nthreads ; n++ ) |
---|
802 | { |
---|
803 | if ( giet_pthread_join( trdid[n], NULL ) ) |
---|
804 | { |
---|
805 | printf("\n[TRANSPOSE ERROR] joining thread %x\n", trdid[n] ); |
---|
806 | giet_pthread_exit( NULL ); |
---|
807 | } |
---|
808 | else |
---|
809 | { |
---|
810 | printf("\n[TRANSPOSE] thread %x joined at cycle %d\n", |
---|
811 | trdid[n] , giet_proctime() ); |
---|
812 | } |
---|
813 | } |
---|
814 | |
---|
815 | // call the instrument() function |
---|
816 | instrument( nclusters , nprocs ); |
---|
817 | |
---|
818 | giet_pthread_exit( "completed" ); |
---|
819 | |
---|
820 | } // end main() |
---|
821 | |
---|
822 | |
---|
823 | |
---|
824 | // Local Variables: |
---|
825 | // tab-width: 3 |
---|
826 | // c-basic-offset: 3 |
---|
827 | // c-file-offsets:((innamespace . 0)(inline-open . 0)) |
---|
828 | // indent-tabs-mode: nil |
---|
829 | // End: |
---|
830 | |
---|
831 | // vim: filetype=cpp:expandtab:shiftwidth=3:tabstop=3:softtabstop=3 |
---|
832 | |
---|
833 | |
---|