1 | #include "stdio.h" |
---|
2 | |
---|
3 | //////////////////////////////////// |
---|
4 | // Image parameters |
---|
5 | |
---|
6 | #define PIXEL_SIZE 2 |
---|
7 | #define NL 1024 |
---|
8 | #define NP 1024 |
---|
9 | #define BLOCK_SIZE 1512 |
---|
10 | |
---|
11 | #define PRINTF if(lid==0) tty_printf |
---|
12 | |
---|
13 | #define TA(c,l,p) (A[c][((NP)*(l))+(p)]) |
---|
14 | #define TB(c,p,l) (B[c][((NL)*(p))+(l)]) |
---|
15 | #define TC(c,l,p) (C[c][((NP)*(l))+(p)]) |
---|
16 | #define TD(c,l,p) (D[c][((NP)*(l))+(p)]) |
---|
17 | |
---|
18 | #define max(x,y) ((x) > (y) ? (x) : (y)) |
---|
19 | #define min(x,y) ((x) < (y) ? (x) : (y)) |
---|
20 | |
---|
21 | /////////////////////////////////////////// |
---|
22 | // tricks to read parameters from ldscript |
---|
23 | /////////////////////////////////////////// |
---|
24 | |
---|
25 | struct plaf; |
---|
26 | |
---|
27 | extern struct plaf seg_heap_base; |
---|
28 | extern struct plaf NB_PROCS; |
---|
29 | extern struct plaf NB_CLUSTERS; |
---|
30 | |
---|
31 | ///////////// |
---|
32 | void main() |
---|
33 | { |
---|
34 | |
---|
35 | ////////////////////////////////// |
---|
36 | // convolution kernel parameters |
---|
37 | // The content of this section is |
---|
38 | // Philips proprietary information. |
---|
39 | /////////////////////////////////// |
---|
40 | |
---|
41 | int vrange = 17; |
---|
42 | int vnorm = 115; |
---|
43 | int vf[35]; |
---|
44 | vf[0] = 1; |
---|
45 | vf[1] = 1; |
---|
46 | vf[2] = 2; |
---|
47 | vf[3] = 2; |
---|
48 | vf[4] = 2; |
---|
49 | vf[5] = 2; |
---|
50 | vf[6] = 3; |
---|
51 | vf[7] = 3; |
---|
52 | vf[8] = 3; |
---|
53 | vf[9] = 4; |
---|
54 | vf[10] = 4; |
---|
55 | vf[11] = 4; |
---|
56 | vf[12] = 4; |
---|
57 | vf[13] = 5; |
---|
58 | vf[14] = 5; |
---|
59 | vf[15] = 5; |
---|
60 | vf[16] = 5; |
---|
61 | vf[17] = 5; |
---|
62 | vf[18] = 5; |
---|
63 | vf[19] = 5; |
---|
64 | vf[20] = 5; |
---|
65 | vf[21] = 5; |
---|
66 | vf[22] = 4; |
---|
67 | vf[23] = 4; |
---|
68 | vf[24] = 4; |
---|
69 | vf[25] = 4; |
---|
70 | vf[26] = 3; |
---|
71 | vf[27] = 3; |
---|
72 | vf[28] = 3; |
---|
73 | vf[29] = 2; |
---|
74 | vf[30] = 2; |
---|
75 | vf[31] = 2; |
---|
76 | vf[32] = 2; |
---|
77 | vf[33] = 1; |
---|
78 | vf[34] = 1; |
---|
79 | |
---|
80 | int hrange = 100; |
---|
81 | int hnorm = 201; |
---|
82 | |
---|
83 | unsigned int date = 0; |
---|
84 | unsigned int delta = 0; |
---|
85 | |
---|
86 | int c; // cluster index for loops |
---|
87 | int l; // line index for loops |
---|
88 | int p; // pixel index for loops |
---|
89 | int x; // filter index for loops |
---|
90 | |
---|
91 | int pid = procid(); // processor id |
---|
92 | int nprocs = (unsigned int)&NB_PROCS; // number of processors per cluster |
---|
93 | int nclusters = (unsigned int)&NB_CLUSTERS; // number of clusters |
---|
94 | int lid = pid%nprocs; // local processor id |
---|
95 | int cid = pid/nprocs; // local processor id |
---|
96 | int base = (unsigned int)&seg_heap_base; // base address for shared buffers |
---|
97 | int increment = (0x80000000 / nclusters) * 2; // cluster increment |
---|
98 | int ntasks = nclusters * nprocs; // number of tasks |
---|
99 | int nblocks = (NP*NL*PIXEL_SIZE)/BLOCK_SIZE; // number of blocks per image |
---|
100 | |
---|
101 | int lines_per_task = NL/ntasks; // number of lines per task |
---|
102 | int lines_per_cluster = NL/nclusters; // number of lines per cluster |
---|
103 | int columns_per_task = NP/ntasks; // number of columns per task |
---|
104 | int columns_per_cluster = NP/nclusters; // number of columns per cluster |
---|
105 | |
---|
106 | PRINTF("\n *** Processor %d entering main at cycle %d ***\n\n", pid, proctime()); |
---|
107 | |
---|
108 | ////////////////////////// |
---|
109 | // parameters checking |
---|
110 | if( (nprocs != 1) && (nprocs != 2) && (nprocs != 4) ) |
---|
111 | { |
---|
112 | PRINTF("NB_PROCS must be 1, 2 or 4\n"); |
---|
113 | while(1); |
---|
114 | } |
---|
115 | if( (nclusters != 4) && (nclusters != 8) && (nclusters != 16) && |
---|
116 | (nclusters != 32) && (nclusters != 64) && (nclusters !=128) && (nclusters != 256) ) |
---|
117 | { |
---|
118 | PRINTF("NB_CLUSTERS must be a power of 2 between 4 and 256\n"); |
---|
119 | while(1); |
---|
120 | } |
---|
121 | if( pid >= ntasks ) |
---|
122 | { |
---|
123 | PRINTF("processor id %d larger than NB_CLUSTERS*NB_PROCS\n", pid); |
---|
124 | while(1); |
---|
125 | } |
---|
126 | if ( NL % nclusters != 0 ) |
---|
127 | { |
---|
128 | PRINTF("NB_CLUSTERS must be a divider of NL"); |
---|
129 | while(1); |
---|
130 | } |
---|
131 | if( NP % nclusters != 0 ) |
---|
132 | { |
---|
133 | PRINTF("NB_CLUSTERS must be a divider of NP"); |
---|
134 | while(1); |
---|
135 | } |
---|
136 | |
---|
137 | ////////////////////////////////////////////////////////////////// |
---|
138 | // Arrays of pointers on the shared, distributed buffers |
---|
139 | // containing the images (sized for the worst case : 256 clusters) |
---|
140 | unsigned short* A[256]; |
---|
141 | int* B[256]; |
---|
142 | int* C[256]; |
---|
143 | int* D[256]; |
---|
144 | |
---|
145 | // The shared, distributed buffers addresses are computed |
---|
146 | // from the seg_heap_base value defined in the ldscript file |
---|
147 | // and from the cluster increment = 4Gbytes/nclusters. |
---|
148 | // These arrays of pointers are identical and |
---|
149 | // replicated in the stack of each task |
---|
150 | for( c=0 ; c<nclusters ; c++) |
---|
151 | { |
---|
152 | A[c] = (unsigned short*)(base + increment*c); |
---|
153 | B[c] = (int*)(base + 4*NP*NL/nclusters + increment*c); |
---|
154 | C[c] = (int*)(base + 8*NP*NL/nclusters + increment*c); |
---|
155 | D[c] = (int*)(base + 12*NP*NL/nclusters + increment*c); |
---|
156 | } |
---|
157 | |
---|
158 | unsigned char* line_buf = (unsigned char*)(base + 2*NP*NL/nclusters + increment*c); |
---|
159 | |
---|
160 | PRINTF("NCLUSTERS = %d\n", nclusters); |
---|
161 | PRINTF("NPROCS = %d\n\n", nprocs); |
---|
162 | |
---|
163 | PRINTF("*** starting barrier init at cycle %d ***\n", proctime()); |
---|
164 | |
---|
165 | // barriers initialization |
---|
166 | barrier_init(0, ntasks); |
---|
167 | barrier_init(1, ntasks); |
---|
168 | barrier_init(2, ntasks); |
---|
169 | |
---|
170 | PRINTF("*** completing barrier init at cycle %d ***\n", proctime()); |
---|
171 | |
---|
172 | //////////////////////////////////////////////////////// |
---|
173 | // pseudo parallel load from disk to A[c] buffers |
---|
174 | // only task running on processor with (lid==0) does it |
---|
175 | // nblocks/nclusters are loaded in each cluster |
---|
176 | |
---|
177 | if ( lid == 0 ) |
---|
178 | { |
---|
179 | delta = proctime() - date; |
---|
180 | date = date + delta; |
---|
181 | PRINTF("\n *** Starting load at cycle %d (%d)\n", date, delta); |
---|
182 | |
---|
183 | if( ioc_read(nblocks*cid/nclusters, |
---|
184 | A[cid] , |
---|
185 | nblocks/nclusters) ) |
---|
186 | { |
---|
187 | PRINTF("echec ioc_read\n"); |
---|
188 | while(1); |
---|
189 | } |
---|
190 | if ( ioc_completed() ) |
---|
191 | { |
---|
192 | PRINTF("echec ioc_completed\n"); |
---|
193 | while(1); |
---|
194 | } |
---|
195 | |
---|
196 | delta = proctime() - date; |
---|
197 | date = date + delta; |
---|
198 | PRINTF(" *** Completing load at cycle %d (%d)\n", date, delta); |
---|
199 | } |
---|
200 | |
---|
201 | barrier_wait(0); |
---|
202 | |
---|
203 | ////////////////////////////////////////////////////////// |
---|
204 | // parallel horizontal filter : |
---|
205 | // B <= transpose(FH(A)) |
---|
206 | // D <= A - FH(A) |
---|
207 | // each task computes (NL/ntasks) lines |
---|
208 | |
---|
209 | delta = proctime() - date; |
---|
210 | date = date + delta; |
---|
211 | PRINTF("\n *** starting horizontal filter at cycle %d (%d)\n", date, delta); |
---|
212 | |
---|
213 | // l = line index in the cluster / p = pixel index |
---|
214 | for ( l = lines_per_task*lid ; l < lines_per_task*(lid+1) ; l++) |
---|
215 | { |
---|
216 | // The image must be extended : |
---|
217 | // if (p<0) TA(cid,l,p) == TA(cid,l,0) |
---|
218 | // if (p>NP-1) TA(cid,l,p) == TA(cid,l,NL-1) |
---|
219 | // We use the spécific values of the horizontal ep-filter for optimisation: |
---|
220 | // sum(p) = sum(p-1) + TA[p+hrange] - TA[p-hrange-1] |
---|
221 | // To minimize the number of tests, the loop on pixels is split in three domains |
---|
222 | |
---|
223 | int sum = (hrange+2)*TA(cid,l,0); |
---|
224 | for ( x = 1 ; x < hrange ; x++) sum = sum + TA(cid,l,x); |
---|
225 | |
---|
226 | // first domain : from 0 to hrange |
---|
227 | for ( p = 0 ; p < hrange+1 ; p++) |
---|
228 | { |
---|
229 | sum = sum + TA(cid,l,p+hrange) - TA(cid,l,0); |
---|
230 | TB((p/columns_per_cluster),(p%columns_per_cluster),(cid*lines_per_cluster+l)) = sum/hnorm; |
---|
231 | TD(cid,l,p) = TA(cid,l,p) - sum/hnorm; |
---|
232 | } |
---|
233 | // second domain : from (hrange+1) to (NP-hrange-1) |
---|
234 | for ( p = hrange+1 ; p < NP-hrange ; p++) |
---|
235 | { |
---|
236 | sum = sum + TA(cid,l,p+hrange) - TA(cid,l,p-hrange-1); |
---|
237 | TB((p/columns_per_cluster),(p%columns_per_cluster),(cid*lines_per_cluster+l)) = sum/hnorm; |
---|
238 | TD(cid,l,p) = TA(cid,l,p) - sum/hnorm; |
---|
239 | } |
---|
240 | // third domain : from (NP-hrange) to (NP-1) |
---|
241 | for ( p = NP-hrange ; p < NP ; p++) |
---|
242 | { |
---|
243 | sum = sum + TA(cid,l,NP-1) - TA(cid,l,p-hrange-1); |
---|
244 | TB((p/columns_per_cluster),(p%columns_per_cluster),(cid*lines_per_cluster+l)) = sum/hnorm; |
---|
245 | TD(cid,l,p) = TA(cid,l,p) - sum/hnorm; |
---|
246 | } |
---|
247 | |
---|
248 | PRINTF(" - line %d computed at cycle %d\n", l, proctime()); |
---|
249 | } |
---|
250 | |
---|
251 | delta = proctime() - date; |
---|
252 | date = date + delta; |
---|
253 | PRINTF(" *** completing horizontal filter at cycle %d (%d)\n", date, delta); |
---|
254 | |
---|
255 | barrier_wait(1); |
---|
256 | |
---|
257 | ////////////////////////////////////////////////////////// |
---|
258 | // parallel vertical filter : |
---|
259 | // C <= transpose(FV(B)) |
---|
260 | // each processor computes (NP/ntasks) columns |
---|
261 | |
---|
262 | delta = proctime() - date; |
---|
263 | date = date + delta; |
---|
264 | PRINTF("\n *** starting vertical filter at cycle %d (%d)\n", date, delta); |
---|
265 | |
---|
266 | // l = line index / p = column index in the cluster |
---|
267 | for ( p = columns_per_task*lid ; p < columns_per_task*(lid+1) ; p++) |
---|
268 | { |
---|
269 | unsigned int sum = 0; |
---|
270 | |
---|
271 | // The image must be extended : |
---|
272 | // if (l<0) TB(cid,p,x) == TB(cid,p,0) |
---|
273 | // if (l>NL-1) TB(cid,p,x) == TB(cid,p,NL-1) |
---|
274 | // We use the spécific values of the vertical ep-filter |
---|
275 | // To minimize the number of tests, the NL lines are split in three domains |
---|
276 | |
---|
277 | // first domain |
---|
278 | for ( l = 0 ; l < vrange ; l++) |
---|
279 | { |
---|
280 | for ( x = 0 ; x < (2*vrange+1) ; x++ ) |
---|
281 | { |
---|
282 | sum = sum + vf[x] * TB(cid,p,max(l-vrange+x,0)); |
---|
283 | } |
---|
284 | TC((l/lines_per_cluster),(l%lines_per_cluster),(cid*columns_per_cluster+p)) = sum/vnorm; |
---|
285 | } |
---|
286 | // second domain |
---|
287 | for ( l = vrange ; l < NL-vrange ; l++ ) |
---|
288 | { |
---|
289 | sum = sum + TB(cid,p,l+4) |
---|
290 | + TB(cid,p,l+8) |
---|
291 | + TB(cid,p,l+11) |
---|
292 | + TB(cid,p,l+15) |
---|
293 | + TB(cid,p,l+17) |
---|
294 | - TB(cid,p,l-5) |
---|
295 | - TB(cid,p,l-9) |
---|
296 | - TB(cid,p,l-12) |
---|
297 | - TB(cid,p,l-16) |
---|
298 | - TB(cid,p,max(l-18,0)); |
---|
299 | TC((l/lines_per_cluster),(l%lines_per_cluster),(cid*columns_per_cluster+p)) = sum/vnorm; |
---|
300 | } |
---|
301 | // third domain |
---|
302 | for ( l = NL-vrange ; l < NL ; l++ ) |
---|
303 | { |
---|
304 | sum = sum + TB(cid,p,min(l+5,NL-1)) |
---|
305 | + TB(cid,p,min(l+9,NL-1)) |
---|
306 | + TB(cid,p,min(l+12,NL-1)) |
---|
307 | + TB(cid,p,min(l+16,NL-1)) |
---|
308 | + TB(cid,p,min(l+18,NL-1)) |
---|
309 | - TB(cid,p,l-4) |
---|
310 | - TB(cid,p,l-8) |
---|
311 | - TB(cid,p,l-11) |
---|
312 | - TB(cid,p,l-15) |
---|
313 | - TB(cid,p,l-17); |
---|
314 | TC((l/lines_per_cluster),(l%lines_per_cluster),(cid*columns_per_cluster+p)) = sum/vnorm; |
---|
315 | } |
---|
316 | |
---|
317 | PRINTF(" - column %d computed at cycle %d\n", p, proctime()); |
---|
318 | } |
---|
319 | |
---|
320 | delta = proctime() - date; |
---|
321 | date = date + delta; |
---|
322 | PRINTF(" *** completing vertical filter at cycle %d (%d)\n", date, delta); |
---|
323 | |
---|
324 | barrier_wait(2); |
---|
325 | |
---|
326 | //////////////////////////////////////////////////////////////////////////// |
---|
327 | // final computation and parallel display using the distributed DMA |
---|
328 | // D <= D + C |
---|
329 | // Each processor use its private DMA channel to display |
---|
330 | // the resulting image, line per line (one byte per pixel). |
---|
331 | // Eah processor computes & displays (NL/ntasks) lines. |
---|
332 | |
---|
333 | delta = proctime() - date; |
---|
334 | date = date + delta; |
---|
335 | PRINTF("\n *** final computation and display at cycle %d (%d)\n", date, delta); |
---|
336 | |
---|
337 | for ( l = 0 ; l < lines_per_task ; l++) |
---|
338 | { |
---|
339 | for ( p = 0 ; p < NP ; p++) |
---|
340 | { |
---|
341 | TD(cid,l,p) = TD(cid,l,p) + TC(cid,l,p); |
---|
342 | line_buf[p] = (unsigned char)(TD(cid,l,p)); |
---|
343 | } |
---|
344 | int xxx = ( fb_write( NP*(cid*lines_per_cluster+lid*lines_per_task+l), line_buf, NP) ); |
---|
345 | if ( xxx ) |
---|
346 | { |
---|
347 | PRINTF("echec fb_write = %d\n", xxx); |
---|
348 | while(1); |
---|
349 | } |
---|
350 | if ( fb_completed() ) |
---|
351 | { |
---|
352 | PRINTF("echec fb_completed\n"); |
---|
353 | while(1); |
---|
354 | } |
---|
355 | PRINTF(" - line %d displayed at cycle %d\n", l, proctime()); |
---|
356 | } |
---|
357 | |
---|
358 | delta = proctime() - date; |
---|
359 | date = date + delta; |
---|
360 | PRINTF(" *** completing display at cycle %d (%d)\n", date, delta); |
---|
361 | |
---|
362 | while(1); |
---|
363 | |
---|
364 | } // end main() |
---|
365 | |
---|