1 | |
---|
2 | #include "stdio.h" |
---|
3 | #include "limits.h" |
---|
4 | #include "../giet_tsar/block_device.h" |
---|
5 | |
---|
6 | #define NL 512 |
---|
7 | #define NP 512 |
---|
8 | #define NB_IMAGES 1 |
---|
9 | #define NB_CLUSTER_MAX 256 |
---|
10 | |
---|
11 | #define PRINTF(...) ({ if (proc_id == 0) { tty_printf(__VA_ARGS__); } }) |
---|
12 | |
---|
13 | //#define DISPLAY_ONLY |
---|
14 | |
---|
15 | /////////////////////////////////////////// |
---|
16 | // tricks to read parameters from ldscript |
---|
17 | /////////////////////////////////////////// |
---|
18 | |
---|
19 | struct plaf; |
---|
20 | |
---|
21 | extern struct plouf seg_ioc_base; |
---|
22 | extern struct plaf seg_heap_base; |
---|
23 | extern struct plaf NB_PROCS; |
---|
24 | extern struct plaf NB_CLUSTERS; |
---|
25 | |
---|
26 | ///////////// |
---|
27 | void main(){ |
---|
28 | unsigned int frame = 0; |
---|
29 | unsigned int date = 0; |
---|
30 | |
---|
31 | unsigned int c; // cluster index for loops |
---|
32 | unsigned int l; // line index for loops |
---|
33 | unsigned int p; // pixel index for loops |
---|
34 | |
---|
35 | unsigned int proc_id = procid(); // processor id |
---|
36 | unsigned int nlocal_procs = (unsigned int) &NB_PROCS; // number of processors per cluster |
---|
37 | unsigned int nclusters = (unsigned int) &NB_CLUSTERS; // number of clusters |
---|
38 | unsigned int local_id = proc_id % nlocal_procs; // local processor id |
---|
39 | unsigned int cluster_id = proc_id / nlocal_procs; // cluster id |
---|
40 | unsigned int base = (unsigned int) &seg_heap_base; // base address for shared buffers |
---|
41 | unsigned int increment = 0x80000000 / nclusters * 2; // cluster increment |
---|
42 | unsigned int nglobal_procs = nclusters * nlocal_procs; // number of tasks |
---|
43 | unsigned int npixels = NP * NL; // number of pixel per frame |
---|
44 | |
---|
45 | unsigned int * ioc_address = (unsigned int *) &seg_ioc_base; |
---|
46 | unsigned int block_size = ioc_address[BLOCK_DEVICE_BLOCK_SIZE]; |
---|
47 | unsigned int nblocks = npixels / block_size; // number of blocks per frame |
---|
48 | |
---|
49 | PRINTF("\n *** Entering main at cycle %d ***\n\n", proctime()); |
---|
50 | |
---|
51 | // parameters checking |
---|
52 | if ((nlocal_procs != 1) && (nlocal_procs != 2) && (nlocal_procs != 4)){ |
---|
53 | PRINTF("NB_PROCS must be 1, 2 or 4\n"); |
---|
54 | exit(1); |
---|
55 | } |
---|
56 | if ((nclusters != 1) && (nclusters != 2) && (nclusters != 4) && (nclusters != 8) && |
---|
57 | (nclusters != 16) && (nclusters != 32) && (nclusters != 64) && (nclusters != 128) && |
---|
58 | (nclusters != 256)){ |
---|
59 | PRINTF("NB_CLUSTERS must be a power of 1 between 1 and 256\n"); |
---|
60 | exit(1); |
---|
61 | } |
---|
62 | if (nglobal_procs > 1024){ |
---|
63 | PRINTF("NB_PROCS * NB_CLUSTERS cannot be larger than 1024\n"); |
---|
64 | exit(1); |
---|
65 | } |
---|
66 | if (proc_id >= nglobal_procs){ |
---|
67 | PRINTF("processor id %d larger than NB_CLUSTERS*NB_PROCS\n", proc_id); |
---|
68 | exit(1); |
---|
69 | } |
---|
70 | |
---|
71 | // Arrays of pointers on the shared, distributed buffers containing the frames |
---|
72 | // These arrays are indexed by the cluster index (sized for the worst case : 256 clusters) |
---|
73 | unsigned char * A[NB_CLUSTER_MAX]; |
---|
74 | unsigned char * B[NB_CLUSTER_MAX]; |
---|
75 | |
---|
76 | // Arrays of pointers on the instrumentation arrays |
---|
77 | // These arrays are indexed by the cluster index (sized for the worst case : 256 clusters) |
---|
78 | // each pointer points on the base adress of an array of NPROCS unsigned int |
---|
79 | unsigned int * LOAD_START[NB_CLUSTER_MAX]; |
---|
80 | unsigned int * LOAD_END[NB_CLUSTER_MAX]; |
---|
81 | unsigned int * TRSP_START[NB_CLUSTER_MAX]; |
---|
82 | unsigned int * TRSP_END[NB_CLUSTER_MAX]; |
---|
83 | unsigned int * DISP_START[NB_CLUSTER_MAX]; |
---|
84 | unsigned int * DISP_END[NB_CLUSTER_MAX]; |
---|
85 | |
---|
86 | // shared buffers address definition |
---|
87 | // from the seg_heap_base and increment depending on the cluster index |
---|
88 | // These arrays of pointers are identical and replicated in the stack of each task |
---|
89 | for (c = 0; c < nclusters; c++){ |
---|
90 | A[c] = (unsigned char *) (base + increment * c); |
---|
91 | B[c] = (unsigned char *) (base + npixels + increment * c); |
---|
92 | LOAD_START[c] = (unsigned int *) (base + 2 * npixels + increment * c); |
---|
93 | LOAD_END[c] = (unsigned int *) (base + 2 * npixels + nlocal_procs + increment * c); |
---|
94 | TRSP_START[c] = (unsigned int *) (base + 2 * npixels + 2 * nlocal_procs + increment * c); |
---|
95 | TRSP_END[c] = (unsigned int *) (base + 2 * npixels + 3 * nlocal_procs + increment * c); |
---|
96 | DISP_START[c] = (unsigned int *) (base + 2 * npixels + 4 * nlocal_procs + increment * c); |
---|
97 | DISP_END[c] = (unsigned int *) (base + 2 * npixels + 5 * nlocal_procs + increment * c); |
---|
98 | } |
---|
99 | |
---|
100 | PRINTF("NB_CLUSTERS = %d\n", nclusters); |
---|
101 | PRINTF("NB_LOCAL_PROCS = %d\n", nlocal_procs); |
---|
102 | PRINTF("NB_GLOBAL_PROCS = %d\n", nglobal_procs); |
---|
103 | PRINTF("NB_PIXELS = %d\n", npixels); |
---|
104 | PRINTF("BLOCK_SIZE = %d\n", block_size); |
---|
105 | PRINTF("NB_BLOCKS = %d\n\n", nblocks); |
---|
106 | |
---|
107 | |
---|
108 | PRINTF("*** Starting barrier init at cycle %d ***\n", proctime()); |
---|
109 | |
---|
110 | // barriers initialization |
---|
111 | barrier_init(0, nglobal_procs); |
---|
112 | barrier_init(1, nglobal_procs); |
---|
113 | barrier_init(2, nglobal_procs); |
---|
114 | |
---|
115 | PRINTF("*** Completing barrier init at cycle %d ***\n", proctime()); |
---|
116 | |
---|
117 | // Main loop (on frames) |
---|
118 | while (frame < NB_IMAGES){ |
---|
119 | // pseudo parallel load from disk to A[c] buffer : nblocks/nclusters blocks |
---|
120 | // only task running on processor with (local_id == 0) does it |
---|
121 | |
---|
122 | if (local_id == 0){ |
---|
123 | int p; |
---|
124 | |
---|
125 | date = proctime(); |
---|
126 | PRINTF("\n*** Starting load for frame %d at cycle %d\n", frame, date); |
---|
127 | |
---|
128 | for (p = 0; p < nlocal_procs; p++){ |
---|
129 | LOAD_START[cluster_id][p] = date; |
---|
130 | } |
---|
131 | if (ioc_read(frame * nblocks + nblocks * cluster_id / nclusters, A[cluster_id], nblocks / nclusters)){ |
---|
132 | PRINTF("echec ioc_read\n"); |
---|
133 | exit(); |
---|
134 | } |
---|
135 | if (ioc_completed()){ |
---|
136 | PRINTF("echec ioc_completed\n"); |
---|
137 | exit(); |
---|
138 | } |
---|
139 | |
---|
140 | date = proctime(); |
---|
141 | PRINTF("*** Completing load for frame %d at cycle %d\n", frame, date); |
---|
142 | for (p = 0; p < nlocal_procs; p++){ |
---|
143 | LOAD_END[cluster_id][p] = date; |
---|
144 | } |
---|
145 | } |
---|
146 | |
---|
147 | barrier_wait(0); |
---|
148 | |
---|
149 | // parallel transpose from A to B buffers |
---|
150 | // each processor makes the transposition for (NL/nglobal_procs) lines |
---|
151 | // (p,l) are the (x,y) pixel coordinates in the source frame |
---|
152 | |
---|
153 | #ifndef DISPLAY_ONLY |
---|
154 | date = proctime(); |
---|
155 | PRINTF("\n*** Starting transpose for frame %d at cycle %d\n", frame, date); |
---|
156 | TRSP_START[cluster_id][local_id] = date; |
---|
157 | |
---|
158 | unsigned int nlt = NL / nglobal_procs; // Nombre de ligne à traiter par processeur |
---|
159 | unsigned int first = proc_id * nlt; // Index de la premiÚre ligne à traiter pour le proc courant (celui qui exécute le code) |
---|
160 | unsigned int last = first + nlt; // Index de la derniÚre ligne |
---|
161 | unsigned int nlines_clusters = NL / nclusters; // Nombre de lignes à traiter par cluster |
---|
162 | unsigned int npix_clusters = NP / nclusters; // Nombre de pixels par ligne à traiter par cluster |
---|
163 | |
---|
164 | for (l = first; l < last; l++){ |
---|
165 | PRINTF(" - processing line %d\n", l); |
---|
166 | for (p = 0; p < NP; p++){ |
---|
167 | unsigned int source_index = (l % nlines_clusters) * NP + p; |
---|
168 | unsigned int dest_cluster = p / npix_clusters; |
---|
169 | unsigned int dest_index = (p % npix_clusters) * NL + l; |
---|
170 | B[dest_cluster][dest_index] = A[cluster_id][source_index]; |
---|
171 | } |
---|
172 | } |
---|
173 | |
---|
174 | date = proctime(); |
---|
175 | PRINTF("*** Completing transpose for frame %d at cycle %d\n", frame, date); |
---|
176 | TRSP_END[cluster_id][local_id] = date; |
---|
177 | barrier_wait(1); |
---|
178 | #endif |
---|
179 | |
---|
180 | // parallel display from B[c] to frame buffer |
---|
181 | // each processor uses its private dma to display NL*NP/nglobal_procs pixels |
---|
182 | |
---|
183 | date = proctime(); |
---|
184 | PRINTF("\n*** Starting display for frame %d at cycle %d\n", frame, date); |
---|
185 | DISP_START[cluster_id][local_id] = date; |
---|
186 | |
---|
187 | unsigned int npxt = npixels / nglobal_procs; // number of pixels per proc |
---|
188 | |
---|
189 | #ifndef DISPLAY_ONLY |
---|
190 | if (fb_write(npxt * proc_id, B[cluster_id] + npxt * local_id, npxt)){ |
---|
191 | PRINTF("[%d]: echec fb_sync_write\n", proc_id); |
---|
192 | exit(); |
---|
193 | } |
---|
194 | #else |
---|
195 | if (fb_write(npxt * proc_id, A[cluster_id] + npxt * local_id, npxt)){ |
---|
196 | PRINTF("[%d]: echec fb_sync_write\n", proc_id); |
---|
197 | exit(); |
---|
198 | } |
---|
199 | #endif |
---|
200 | |
---|
201 | if (fb_completed()){ |
---|
202 | PRINTF("[%d]: echec fb_completed\n", proc_id); |
---|
203 | exit(); |
---|
204 | } |
---|
205 | |
---|
206 | date = proctime(); |
---|
207 | PRINTF("*** Completing display for frame %d at cycle %d\n", frame, date); |
---|
208 | DISP_END[cluster_id][local_id] = date; |
---|
209 | |
---|
210 | barrier_wait(2); |
---|
211 | |
---|
212 | // Instrumentation (done by processor 0 in cluster 0) |
---|
213 | if (local_id == 0){ |
---|
214 | date = proctime(); |
---|
215 | PRINTF("\n*** Starting Instrumentation for frame %d at cycle %d\n\n", frame, date); |
---|
216 | |
---|
217 | int cc, pp; |
---|
218 | unsigned int min_load_start = INT_MAX; |
---|
219 | unsigned int max_load_start = 0; |
---|
220 | unsigned int min_load_ended = INT_MAX; |
---|
221 | unsigned int max_load_ended = 0; |
---|
222 | unsigned int min_trsp_start = INT_MAX; |
---|
223 | unsigned int max_trsp_start = 0; |
---|
224 | unsigned int min_trsp_ended = INT_MAX; |
---|
225 | unsigned int max_trsp_ended = 0; |
---|
226 | unsigned int min_disp_start = INT_MAX; |
---|
227 | unsigned int max_disp_start = 0; |
---|
228 | unsigned int min_disp_ended = INT_MAX; |
---|
229 | unsigned int max_disp_ended = 0; |
---|
230 | |
---|
231 | for (cc = 0; cc < nclusters; cc++){ |
---|
232 | for (pp = 0; pp < nlocal_procs; pp++){ |
---|
233 | if (LOAD_START[cc][pp] < min_load_start){ |
---|
234 | min_load_start = LOAD_START[cc][pp]; |
---|
235 | } |
---|
236 | if (LOAD_START[cc][pp] > max_load_start){ |
---|
237 | max_load_start = LOAD_START[cc][pp]; |
---|
238 | } |
---|
239 | if (LOAD_END[cc][pp] < min_load_ended){ |
---|
240 | min_load_ended = LOAD_END[cc][pp]; |
---|
241 | } |
---|
242 | if (LOAD_END[cc][pp] > max_load_ended){ |
---|
243 | max_load_ended = LOAD_END[cc][pp]; |
---|
244 | } |
---|
245 | |
---|
246 | if (TRSP_START[cc][pp] < min_trsp_start){ |
---|
247 | min_trsp_start = TRSP_START[cc][pp]; |
---|
248 | } |
---|
249 | if (TRSP_START[cc][pp] > max_trsp_start){ |
---|
250 | max_trsp_start = TRSP_START[cc][pp]; |
---|
251 | } |
---|
252 | if (TRSP_END[cc][pp] < min_trsp_ended){ |
---|
253 | min_trsp_ended = TRSP_END[cc][pp]; |
---|
254 | } |
---|
255 | if (TRSP_END[cc][pp] > max_trsp_ended){ |
---|
256 | max_trsp_ended = TRSP_END[cc][pp]; |
---|
257 | } |
---|
258 | |
---|
259 | if (DISP_START[cc][pp] < min_disp_start){ |
---|
260 | min_disp_start = DISP_START[cc][pp]; |
---|
261 | } |
---|
262 | if (DISP_START[cc][pp] > max_disp_start){ |
---|
263 | max_disp_start = DISP_START[cc][pp]; |
---|
264 | } |
---|
265 | if (DISP_END[cc][pp] < min_disp_ended){ |
---|
266 | min_disp_ended = DISP_END[cc][pp]; |
---|
267 | } |
---|
268 | if (DISP_END[cc][pp] > max_disp_ended){ |
---|
269 | max_disp_ended = DISP_END[cc][pp]; |
---|
270 | } |
---|
271 | } |
---|
272 | } |
---|
273 | |
---|
274 | PRINTF(" - LOAD_START : min = %d / max = %d / med = %d / delta = %d\n", |
---|
275 | min_load_start, max_load_start, (min_load_start+max_load_start)/2, max_load_start-min_load_start); |
---|
276 | PRINTF(" - LOAD_END : min = %d / max = %d / med = %d / delta = %d\n", |
---|
277 | min_load_ended, max_load_ended, (min_load_ended+max_load_ended)/2, max_load_ended-min_load_ended); |
---|
278 | |
---|
279 | PRINTF(" - TRSP_START : min = %d / max = %d / med = %d / delta = %d\n", |
---|
280 | min_trsp_start, max_trsp_start, (min_trsp_start+max_trsp_start)/2, max_trsp_start-min_trsp_start); |
---|
281 | PRINTF(" - TRSP_END : min = %d / max = %d / med = %d / delta = %d\n", |
---|
282 | min_trsp_ended, max_trsp_ended, (min_trsp_ended+max_trsp_ended)/2, max_trsp_ended-min_trsp_ended); |
---|
283 | |
---|
284 | PRINTF(" - DISP_START : min = %d / max = %d / med = %d / delta = %d\n", |
---|
285 | min_disp_start, max_disp_start, (min_disp_start+max_disp_start)/2, max_disp_start-min_disp_start); |
---|
286 | PRINTF(" - DISP_END : min = %d / max = %d / med = %d / delta = %d\n", |
---|
287 | min_disp_ended, max_disp_ended, (min_disp_ended+max_disp_ended)/2, max_disp_ended-min_disp_ended); |
---|
288 | |
---|
289 | PRINTF(" - BARRIER TRSP/DISP = %d\n", min_disp_start - max_trsp_ended); |
---|
290 | } |
---|
291 | frame++; |
---|
292 | |
---|
293 | } // end while frame |
---|
294 | |
---|
295 | PRINTF("*** End of main ***\n"); |
---|
296 | |
---|
297 | while(1); |
---|
298 | } // end main() |
---|
299 | |
---|
300 | // Local Variables: |
---|
301 | // tab-width: 3 |
---|
302 | // c-basic-offset: 3 |
---|
303 | // c-file-offsets:((innamespace . 0)(inline-open . 0)) |
---|
304 | // indent-tabs-mode: nil |
---|
305 | // End: |
---|
306 | |
---|
307 | // vim: filetype=cpp:expandtab:shiftwidth=3:tabstop=3:softtabstop=3 |
---|
308 | |
---|
309 | |
---|
310 | |
---|