| 1 | ///////////////////////////////////////////////////////////////////////////////////////
|
|---|
| 2 | // File : transpose.c
|
|---|
| 3 | // Date : september 2015
|
|---|
| 4 | // author : Alain Greiner
|
|---|
| 5 | ///////////////////////////////////////////////////////////////////////////////////////
|
|---|
| 6 | // This multi-threaded aplication transposes a raw image (one pbyte per pixel).
|
|---|
| 7 | // It can run on a multi-processors, multi-clusters architecture, with one thread
|
|---|
| 8 | // per processor, and uses the POSIX threads API.
|
|---|
| 9 | // It does not use the CMA to display the result image.
|
|---|
| 10 | //
|
|---|
| 11 | // The main() function can be launched on any processor P[x,y,l].
|
|---|
| 12 | // It makes the initialisations, launch (N-1) threads to run the execute() function
|
|---|
| 13 | // on the (N-1) other processors than P[x,y,l], call himself the execute() function,
|
|---|
| 14 | // and finally call the instrument() function to display instrumentation results
|
|---|
| 15 | // when the parallel execution is completed.
|
|---|
| 16 | //
|
|---|
| 17 | // The input and output buffers containing the image are distributed in clusters.
|
|---|
| 18 | //
|
|---|
| 19 | // The execute() function read a set of lines from an input file on disk,
|
|---|
| 20 | // to the local buffer buf_in[x][y], transpose it, write the result to a remote buffer
|
|---|
| 21 | // buf_out[x'][y'], display the content of the local buffer buf_out[x][y] to the
|
|---|
| 22 | // frame buffer, and store it on disk to another output file.
|
|---|
| 23 | //
|
|---|
| 24 | // - The image size must fit the frame buffer size.
|
|---|
| 25 | // - The block size in block device must be 512 bytes.
|
|---|
| 26 | // - The number of clusters must be a power of 2 no larger than 256.
|
|---|
| 27 | // - The number of processors per cluster must be a power of 2 no larger than 4.
|
|---|
| 28 | ///////////////////////////////////////////////////////////////////////////////////////
|
|---|
| 29 |
|
|---|
| 30 | #include "stdio.h"
|
|---|
| 31 | #include "stdlib.h"
|
|---|
| 32 | #include "user_barrier.h"
|
|---|
| 33 | #include "malloc.h"
|
|---|
| 34 |
|
|---|
| 35 | #define BLOCK_SIZE 512 // block size on disk
|
|---|
| 36 | #define X_MAX 16 // max number of clusters in row
|
|---|
| 37 | #define Y_MAX 16 // max number of clusters in column
|
|---|
| 38 | #define PROCS_MAX 4 // max number of procs per cluster
|
|---|
| 39 | #define CLUSTER_MAX (X_MAX * Y_MAX) // max number of clusters
|
|---|
| 40 | #define IMAGE_SIZE 256 // default image size
|
|---|
| 41 | #define INPUT_FILE_PATH "/misc/lena_256.raw" // default input file pathname
|
|---|
| 42 | #define OUTPUT_FILE_PATH "/home/lena_transposed.raw" // default output file pathname
|
|---|
| 43 | #define INTERACTIVE 0 // interactive capture of filenames
|
|---|
| 44 | #define VERBOSE 1
|
|---|
| 45 |
|
|---|
| 46 |
|
|---|
| 47 | // macro to use a shared TTY
|
|---|
| 48 | #define printf(...); { lock_acquire( &tty_lock ); \
|
|---|
| 49 | giet_tty_printf(__VA_ARGS__); \
|
|---|
| 50 | lock_release( &tty_lock ); }
|
|---|
| 51 |
|
|---|
| 52 | ///////////////////////////////////////////////////////
|
|---|
| 53 | // global variables stored in seg_data in cluster(0,0)
|
|---|
| 54 | ///////////////////////////////////////////////////////
|
|---|
| 55 |
|
|---|
| 56 | // instrumentation counters for each processor in each cluster
|
|---|
| 57 | unsigned int LOAD_START[X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
|
|---|
| 58 | unsigned int LOAD_END [X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
|
|---|
| 59 | unsigned int TRSP_START[X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
|
|---|
| 60 | unsigned int TRSP_END [X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
|
|---|
| 61 | unsigned int DISP_START[X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
|
|---|
| 62 | unsigned int DISP_END [X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
|
|---|
| 63 | unsigned int STOR_START[X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
|
|---|
| 64 | unsigned int STOR_END [X_MAX][Y_MAX][PROCS_MAX] = {{{ 0 }}};
|
|---|
| 65 |
|
|---|
| 66 | // arrays of pointers on distributed buffers
|
|---|
| 67 | // one input buffer & one output buffer per cluster
|
|---|
| 68 | unsigned char* buf_in [CLUSTER_MAX];
|
|---|
| 69 | unsigned char* buf_out[CLUSTER_MAX];
|
|---|
| 70 |
|
|---|
| 71 | // checksum variables
|
|---|
| 72 | unsigned check_line_before[1024];
|
|---|
| 73 | unsigned check_line_after[1024];
|
|---|
| 74 |
|
|---|
| 75 | // lock protecting shared TTY
|
|---|
| 76 | user_lock_t tty_lock;
|
|---|
| 77 |
|
|---|
| 78 | // synchronisation barrier (all threads)
|
|---|
| 79 | giet_sqt_barrier_t barrier;
|
|---|
| 80 |
|
|---|
| 81 | // input & output files pathname and size
|
|---|
| 82 | char input_file_name[256];
|
|---|
| 83 | char output_file_name[256];
|
|---|
| 84 | unsigned int image_size;
|
|---|
| 85 |
|
|---|
| 86 | ////////////////////////////////////////////
|
|---|
| 87 | __attribute__ ((constructor)) void execute()
|
|---|
| 88 | ////////////////////////////////////////////
|
|---|
| 89 | {
|
|---|
| 90 | unsigned int l; // line index for loops
|
|---|
| 91 | unsigned int p; // pixel index for loops
|
|---|
| 92 |
|
|---|
| 93 | // get processor identifiers
|
|---|
| 94 | unsigned int x_id; // x cluster coordinate
|
|---|
| 95 | unsigned int y_id; // y cluster coordinate
|
|---|
| 96 | unsigned int p_id; // local processor index
|
|---|
| 97 |
|
|---|
| 98 | giet_proc_xyp( &x_id, &y_id, &p_id);
|
|---|
| 99 |
|
|---|
| 100 | // get & check plat-form parameters
|
|---|
| 101 | unsigned int x_size; // number of clusters in a row
|
|---|
| 102 | unsigned int y_size; // number of clusters in a column
|
|---|
| 103 | unsigned int nprocs; // number of processors per cluster
|
|---|
| 104 |
|
|---|
| 105 | giet_procs_number( &x_size , &y_size , &nprocs );
|
|---|
| 106 |
|
|---|
| 107 | unsigned int nclusters = x_size * y_size; // number of clusters
|
|---|
| 108 | unsigned int nthreads = x_size * y_size * nprocs; // number of threads
|
|---|
| 109 | unsigned int npixels = image_size * image_size; // pixels per image
|
|---|
| 110 | int fd_in = 0; // initial file descriptor
|
|---|
| 111 | int fd_out = 0; // output file descriptor
|
|---|
| 112 | unsigned int cluster_id = (x_id * y_size) + y_id; // "continuous" index
|
|---|
| 113 | unsigned int thread_id = (cluster_id * nprocs) + p_id; // "continuous" thread index
|
|---|
| 114 |
|
|---|
| 115 | // parallel load of image:
|
|---|
| 116 | // allocate buf_in and buf_out distributed buffers (one buf_in & one buf_out per cluster).
|
|---|
| 117 | // open input and output files, and load the relevant lines in local buf_in.
|
|---|
| 118 | // only thread running on processor[x,y,0] does it.
|
|---|
| 119 |
|
|---|
| 120 | LOAD_START[x_id][y_id][p_id] = giet_proctime();
|
|---|
| 121 |
|
|---|
| 122 | if ( p_id == 0 )
|
|---|
| 123 | {
|
|---|
| 124 | buf_in[cluster_id] = remote_malloc( npixels/nclusters, x_id, y_id );
|
|---|
| 125 | buf_out[cluster_id] = remote_malloc( npixels/nclusters, x_id, y_id );
|
|---|
| 126 |
|
|---|
| 127 | if ( (x_id==0) && (y_id==0) )
|
|---|
| 128 | {
|
|---|
| 129 | printf("\n[TRANSPOSE] Proc [%d,%d,%d] completes buffer allocation at cycle %d\n",
|
|---|
| 130 | x_id, y_id, p_id, giet_proctime() );
|
|---|
| 131 | }
|
|---|
| 132 |
|
|---|
| 133 | // open input file
|
|---|
| 134 | fd_in = giet_fat_open( input_file_name , O_RDONLY ); // read_only
|
|---|
| 135 | if ( fd_in < 0 )
|
|---|
| 136 | {
|
|---|
| 137 | printf("\n[TRANSPOSE ERROR] Proc [%d,%d,%d] cannot open file %s\n",
|
|---|
| 138 | x_id , y_id , p_id , input_file_name );
|
|---|
| 139 | giet_pthread_exit(" open() failure");
|
|---|
| 140 | }
|
|---|
| 141 | else if ( (x_id==0) && (y_id==0) )
|
|---|
| 142 | {
|
|---|
| 143 | printf("\n[TRANSPOSE] Proc [0,0,0] open file %s / fd = %d\n",
|
|---|
| 144 | input_file_name , fd_in );
|
|---|
| 145 | }
|
|---|
| 146 |
|
|---|
| 147 | // open output file
|
|---|
| 148 | fd_out = giet_fat_open( output_file_name , O_CREAT ); // create if required
|
|---|
| 149 | if ( fd_out < 0 )
|
|---|
| 150 | {
|
|---|
| 151 | printf("\n[TRANSPOSE ERROR] Proc [%d,%d,%d] cannot open file %s\n",
|
|---|
| 152 | x_id , y_id , p_id , output_file_name );
|
|---|
| 153 | giet_pthread_exit(" open() failure");
|
|---|
| 154 | }
|
|---|
| 155 | else if ( (x_id==0) && (y_id==0) )
|
|---|
| 156 | {
|
|---|
| 157 | printf("\n[TRANSPOSE] Proc [0,0,0] open file %s / fd = %d\n",
|
|---|
| 158 | output_file_name , fd_out );
|
|---|
| 159 | }
|
|---|
| 160 |
|
|---|
| 161 |
|
|---|
| 162 | unsigned int offset = ((npixels*cluster_id)/nclusters);
|
|---|
| 163 | if ( giet_fat_lseek( fd_in,
|
|---|
| 164 | offset,
|
|---|
| 165 | SEEK_SET ) != offset )
|
|---|
| 166 | {
|
|---|
| 167 | printf("\n[TRANSPOSE ERROR] Proc [%d,%d,%d] cannot seek fd = %d\n",
|
|---|
| 168 | x_id , y_id , p_id , fd_in );
|
|---|
| 169 | giet_pthread_exit(" seek() failure");
|
|---|
| 170 | }
|
|---|
| 171 |
|
|---|
| 172 | unsigned int pixels = npixels / nclusters;
|
|---|
| 173 | if ( giet_fat_read( fd_in,
|
|---|
| 174 | buf_in[cluster_id],
|
|---|
| 175 | pixels ) != pixels )
|
|---|
| 176 | {
|
|---|
| 177 | printf("\n[TRANSPOSE ERROR] Proc [%d,%d,%d] cannot read fd = %d\n",
|
|---|
| 178 | x_id , y_id , p_id , fd_in );
|
|---|
| 179 | giet_pthread_exit(" read() failure");
|
|---|
| 180 | }
|
|---|
| 181 |
|
|---|
| 182 | if ( (x_id==0) && (y_id==0) )
|
|---|
| 183 | {
|
|---|
| 184 | printf("\n[TRANSPOSE] Proc [%d,%d,%d] completes load at cycle %d\n",
|
|---|
| 185 | x_id, y_id, p_id, giet_proctime() );
|
|---|
| 186 | }
|
|---|
| 187 | }
|
|---|
| 188 |
|
|---|
| 189 | LOAD_END[x_id][y_id][p_id] = giet_proctime();
|
|---|
| 190 |
|
|---|
| 191 | /////////////////////////////
|
|---|
| 192 | sqt_barrier_wait( &barrier );
|
|---|
| 193 | /////////////////////////////
|
|---|
| 194 |
|
|---|
| 195 | // parallel transpose from buf_in to buf_out
|
|---|
| 196 | // each thread makes the transposition for nlt lines (nlt = image_size/nthreads)
|
|---|
| 197 | // from line [thread_id*nlt] to line [(thread_id + 1)*nlt - 1]
|
|---|
| 198 | // (p,l) are the absolute pixel coordinates in the source image
|
|---|
| 199 |
|
|---|
| 200 | TRSP_START[x_id][y_id][p_id] = giet_proctime();
|
|---|
| 201 |
|
|---|
| 202 | unsigned int nlt = image_size / nthreads; // number of lines per thread
|
|---|
| 203 | unsigned int nlc = image_size / nclusters; // number of lines per cluster
|
|---|
| 204 |
|
|---|
| 205 | unsigned int src_cluster;
|
|---|
| 206 | unsigned int src_index;
|
|---|
| 207 | unsigned int dst_cluster;
|
|---|
| 208 | unsigned int dst_index;
|
|---|
| 209 |
|
|---|
| 210 | unsigned char byte;
|
|---|
| 211 |
|
|---|
| 212 | unsigned int first = thread_id * nlt; // first line index for a given thread
|
|---|
| 213 | unsigned int last = first + nlt; // last line index for a given thread
|
|---|
| 214 |
|
|---|
| 215 | for ( l = first ; l < last ; l++ )
|
|---|
| 216 | {
|
|---|
| 217 | check_line_before[l] = 0;
|
|---|
| 218 |
|
|---|
| 219 | // in each iteration we transfer one byte
|
|---|
| 220 | for ( p = 0 ; p < image_size ; p++ )
|
|---|
| 221 | {
|
|---|
| 222 | // read one byte from local buf_in
|
|---|
| 223 | src_cluster = l / nlc;
|
|---|
| 224 | src_index = (l % nlc)*image_size + p;
|
|---|
| 225 | byte = buf_in[src_cluster][src_index];
|
|---|
| 226 |
|
|---|
| 227 | // compute checksum
|
|---|
| 228 | check_line_before[l] = check_line_before[l] + byte;
|
|---|
| 229 |
|
|---|
| 230 | // write one byte to remote buf_out
|
|---|
| 231 | dst_cluster = p / nlc;
|
|---|
| 232 | dst_index = (p % nlc)*image_size + l;
|
|---|
| 233 | buf_out[dst_cluster][dst_index] = byte;
|
|---|
| 234 | }
|
|---|
| 235 | }
|
|---|
| 236 |
|
|---|
| 237 | if ( (p_id == 0) && (x_id==0) && (y_id==0) )
|
|---|
| 238 | {
|
|---|
| 239 | printf("\n[TRANSPOSE] proc [%d,%d,%d] completes transpose at cycle %d\n",
|
|---|
| 240 | x_id, y_id, p_id, giet_proctime() );
|
|---|
| 241 | }
|
|---|
| 242 |
|
|---|
| 243 | TRSP_END[x_id][y_id][p_id] = giet_proctime();
|
|---|
| 244 |
|
|---|
| 245 | /////////////////////////////
|
|---|
| 246 | sqt_barrier_wait( &barrier );
|
|---|
| 247 | /////////////////////////////
|
|---|
| 248 |
|
|---|
| 249 | // parallel display from local buf_out to frame buffer
|
|---|
| 250 | // all threads contribute to display using memcpy...
|
|---|
| 251 |
|
|---|
| 252 | DISP_START[x_id][y_id][p_id] = giet_proctime();
|
|---|
| 253 |
|
|---|
| 254 | unsigned int npt = npixels / nthreads; // number of pixels per thread
|
|---|
| 255 |
|
|---|
| 256 | giet_fbf_sync_write( npt * thread_id,
|
|---|
| 257 | &buf_out[cluster_id][p_id*npt],
|
|---|
| 258 | npt );
|
|---|
| 259 |
|
|---|
| 260 | if ( (x_id==0) && (y_id==0) && (p_id==0) )
|
|---|
| 261 | {
|
|---|
| 262 | printf("\n[TRANSPOSE] Proc [%d,%d,%d] completes display at cycle %d\n",
|
|---|
| 263 | x_id, y_id, p_id, giet_proctime() );
|
|---|
| 264 | }
|
|---|
| 265 |
|
|---|
| 266 | DISP_END[x_id][y_id][p_id] = giet_proctime();
|
|---|
| 267 |
|
|---|
| 268 | /////////////////////////////
|
|---|
| 269 | sqt_barrier_wait( &barrier );
|
|---|
| 270 | /////////////////////////////
|
|---|
| 271 |
|
|---|
| 272 | // parallel store : buf_out buffers to disk
|
|---|
| 273 | // only thread running on processor(x,y,0) does it
|
|---|
| 274 |
|
|---|
| 275 | STOR_START[x_id][y_id][p_id] = giet_proctime();
|
|---|
| 276 |
|
|---|
| 277 | if ( p_id == 0 )
|
|---|
| 278 | {
|
|---|
| 279 | unsigned int offset = ((npixels*cluster_id)/nclusters);
|
|---|
| 280 | if ( giet_fat_lseek( fd_out,
|
|---|
| 281 | offset,
|
|---|
| 282 | SEEK_SET ) != offset )
|
|---|
| 283 | {
|
|---|
| 284 | printf("\n[TRANSPOSE ERROR] Proc [%d,%d,%d] cannot seek fr = %d\n",
|
|---|
| 285 | x_id , y_id , p_id , fd_out );
|
|---|
| 286 | giet_pthread_exit(" seek() failure");
|
|---|
| 287 | }
|
|---|
| 288 |
|
|---|
| 289 | unsigned int pixels = npixels / nclusters;
|
|---|
| 290 | if ( giet_fat_write( fd_out,
|
|---|
| 291 | buf_out[cluster_id],
|
|---|
| 292 | pixels ) != pixels )
|
|---|
| 293 | {
|
|---|
| 294 | printf("\n[TRANSPOSE ERROR] Proc [%d,%d,%d] cannot write fd = %d\n",
|
|---|
| 295 | x_id , y_id , p_id , fd_out );
|
|---|
| 296 | giet_pthread_exit(" write() failure");
|
|---|
| 297 | }
|
|---|
| 298 |
|
|---|
| 299 | if ( (x_id==0) && (y_id==0) )
|
|---|
| 300 | {
|
|---|
| 301 | printf("\n[TRANSPOSE] Proc [%d,%d,%d] completes store at cycle %d\n",
|
|---|
| 302 | x_id, y_id, p_id, giet_proctime() );
|
|---|
| 303 | }
|
|---|
| 304 | }
|
|---|
| 305 |
|
|---|
| 306 | STOR_END[x_id][y_id][p_id] = giet_proctime();
|
|---|
| 307 |
|
|---|
| 308 | // In each cluster, only thread running on Processor[x,y,0] releases
|
|---|
| 309 | // the distributed buffers and close the file descriptors.
|
|---|
| 310 |
|
|---|
| 311 | if ( p_id==0 )
|
|---|
| 312 | {
|
|---|
| 313 | free( buf_in[cluster_id] );
|
|---|
| 314 | free( buf_out[cluster_id] );
|
|---|
| 315 |
|
|---|
| 316 | giet_fat_close( fd_in );
|
|---|
| 317 | giet_fat_close( fd_out );
|
|---|
| 318 | }
|
|---|
| 319 |
|
|---|
| 320 | if ( (x_id != 0) || (y_id != 0) || (p_id != 0) )
|
|---|
| 321 | giet_pthread_exit( "completed" );
|
|---|
| 322 |
|
|---|
| 323 | } // end execute()
|
|---|
| 324 |
|
|---|
| 325 |
|
|---|
| 326 |
|
|---|
| 327 | //////////////////////////////////////
|
|---|
| 328 | void instrument( unsigned int x_size,
|
|---|
| 329 | unsigned int y_size,
|
|---|
| 330 | unsigned int nprocs )
|
|---|
| 331 | //////////////////////////////////////
|
|---|
| 332 | {
|
|---|
| 333 | unsigned int x, y, l;
|
|---|
| 334 |
|
|---|
| 335 | unsigned int min_load_start = 0xFFFFFFFF;
|
|---|
| 336 | unsigned int max_load_start = 0;
|
|---|
| 337 | unsigned int min_load_ended = 0xFFFFFFFF;
|
|---|
| 338 | unsigned int max_load_ended = 0;
|
|---|
| 339 | unsigned int min_trsp_start = 0xFFFFFFFF;
|
|---|
| 340 | unsigned int max_trsp_start = 0;
|
|---|
| 341 | unsigned int min_trsp_ended = 0xFFFFFFFF;
|
|---|
| 342 | unsigned int max_trsp_ended = 0;
|
|---|
| 343 | unsigned int min_disp_start = 0xFFFFFFFF;
|
|---|
| 344 | unsigned int max_disp_start = 0;
|
|---|
| 345 | unsigned int min_disp_ended = 0xFFFFFFFF;
|
|---|
| 346 | unsigned int max_disp_ended = 0;
|
|---|
| 347 | unsigned int min_stor_start = 0xFFFFFFFF;
|
|---|
| 348 | unsigned int max_stor_start = 0;
|
|---|
| 349 | unsigned int min_stor_ended = 0xFFFFFFFF;
|
|---|
| 350 | unsigned int max_stor_ended = 0;
|
|---|
| 351 |
|
|---|
| 352 | for (x = 0; x < x_size; x++)
|
|---|
| 353 | {
|
|---|
| 354 | for (y = 0; y < y_size; y++)
|
|---|
| 355 | {
|
|---|
| 356 | for ( l = 0 ; l < nprocs ; l++ )
|
|---|
| 357 | {
|
|---|
| 358 | if (LOAD_START[x][y][l] < min_load_start) min_load_start = LOAD_START[x][y][l];
|
|---|
| 359 | if (LOAD_START[x][y][l] > max_load_start) max_load_start = LOAD_START[x][y][l];
|
|---|
| 360 | if (LOAD_END[x][y][l] < min_load_ended) min_load_ended = LOAD_END[x][y][l];
|
|---|
| 361 | if (LOAD_END[x][y][l] > max_load_ended) max_load_ended = LOAD_END[x][y][l];
|
|---|
| 362 | if (TRSP_START[x][y][l] < min_trsp_start) min_trsp_start = TRSP_START[x][y][l];
|
|---|
| 363 | if (TRSP_START[x][y][l] > max_trsp_start) max_trsp_start = TRSP_START[x][y][l];
|
|---|
| 364 | if (TRSP_END[x][y][l] < min_trsp_ended) min_trsp_ended = TRSP_END[x][y][l];
|
|---|
| 365 | if (TRSP_END[x][y][l] > max_trsp_ended) max_trsp_ended = TRSP_END[x][y][l];
|
|---|
| 366 | if (DISP_START[x][y][l] < min_disp_start) min_disp_start = DISP_START[x][y][l];
|
|---|
| 367 | if (DISP_START[x][y][l] > max_disp_start) max_disp_start = DISP_START[x][y][l];
|
|---|
| 368 | if (DISP_END[x][y][l] < min_disp_ended) min_disp_ended = DISP_END[x][y][l];
|
|---|
| 369 | if (DISP_END[x][y][l] > max_disp_ended) max_disp_ended = DISP_END[x][y][l];
|
|---|
| 370 | if (STOR_START[x][y][l] < min_stor_start) min_stor_start = STOR_START[x][y][l];
|
|---|
| 371 | if (STOR_START[x][y][l] > max_stor_start) max_stor_start = STOR_START[x][y][l];
|
|---|
| 372 | if (STOR_END[x][y][l] < min_stor_ended) min_stor_ended = STOR_END[x][y][l];
|
|---|
| 373 | if (STOR_END[x][y][l] > max_stor_ended) max_stor_ended = STOR_END[x][y][l];
|
|---|
| 374 | }
|
|---|
| 375 | }
|
|---|
| 376 | }
|
|---|
| 377 |
|
|---|
| 378 | printf("\n ---------------- Instrumentation Results ---------------------\n");
|
|---|
| 379 |
|
|---|
| 380 | printf(" - LOAD_START : min = %d / max = %d / med = %d / delta = %d\n",
|
|---|
| 381 | min_load_start, max_load_start, (min_load_start+max_load_start)/2,
|
|---|
| 382 | max_load_start-min_load_start);
|
|---|
| 383 |
|
|---|
| 384 | printf(" - LOAD_END : min = %d / max = %d / med = %d / delta = %d\n",
|
|---|
| 385 | min_load_ended, max_load_ended, (min_load_ended+max_load_ended)/2,
|
|---|
| 386 | max_load_ended-min_load_ended);
|
|---|
| 387 |
|
|---|
| 388 | printf(" - TRSP_START : min = %d / max = %d / med = %d / delta = %d\n",
|
|---|
| 389 | min_trsp_start, max_trsp_start, (min_trsp_start+max_trsp_start)/2,
|
|---|
| 390 | max_trsp_start-min_trsp_start);
|
|---|
| 391 |
|
|---|
| 392 | printf(" - TRSP_END : min = %d / max = %d / med = %d / delta = %d\n",
|
|---|
| 393 | min_trsp_ended, max_trsp_ended, (min_trsp_ended+max_trsp_ended)/2,
|
|---|
| 394 | max_trsp_ended-min_trsp_ended);
|
|---|
| 395 |
|
|---|
| 396 | printf(" - DISP_START : min = %d / max = %d / med = %d / delta = %d\n",
|
|---|
| 397 | min_disp_start, max_disp_start, (min_disp_start+max_disp_start)/2,
|
|---|
| 398 | max_disp_start-min_disp_start);
|
|---|
| 399 |
|
|---|
| 400 | printf(" - DISP_END : min = %d / max = %d / med = %d / delta = %d\n",
|
|---|
| 401 | min_disp_ended, max_disp_ended, (min_disp_ended+max_disp_ended)/2,
|
|---|
| 402 | max_disp_ended-min_disp_ended);
|
|---|
| 403 |
|
|---|
| 404 | printf(" - STOR_START : min = %d / max = %d / med = %d / delta = %d\n",
|
|---|
| 405 | min_stor_start, max_stor_start, (min_stor_start+max_stor_start)/2,
|
|---|
| 406 | max_stor_start-min_stor_start);
|
|---|
| 407 |
|
|---|
| 408 | printf(" - STOR_END : min = %d / max = %d / med = %d / delta = %d\n",
|
|---|
| 409 | min_stor_ended, max_stor_ended, (min_stor_ended+max_stor_ended)/2,
|
|---|
| 410 | max_stor_ended-min_stor_ended);
|
|---|
| 411 |
|
|---|
| 412 | } // end instrument()
|
|---|
| 413 |
|
|---|
| 414 |
|
|---|
| 415 |
|
|---|
| 416 | //////////////////////////////////////////
|
|---|
| 417 | __attribute__ ((constructor)) void main()
|
|---|
| 418 | //////////////////////////////////////////
|
|---|
| 419 | {
|
|---|
| 420 | // indexes for loops
|
|---|
| 421 | unsigned int x , y , n;
|
|---|
| 422 |
|
|---|
| 423 | // get identifiers for proc executing main
|
|---|
| 424 | unsigned int x_id; // x cluster coordinate
|
|---|
| 425 | unsigned int y_id; // y cluster coordinate
|
|---|
| 426 | unsigned int p_id; // local processor index
|
|---|
| 427 |
|
|---|
| 428 | giet_proc_xyp( &x_id , &y_id , &p_id );
|
|---|
| 429 |
|
|---|
| 430 | // get & check plat-form parameters
|
|---|
| 431 | unsigned int x_size; // number of clusters in a row
|
|---|
| 432 | unsigned int y_size; // number of clusters in a column
|
|---|
| 433 | unsigned int nprocs; // number of processors per cluster
|
|---|
| 434 |
|
|---|
| 435 | giet_procs_number( &x_size , &y_size , &nprocs );
|
|---|
| 436 |
|
|---|
| 437 | giet_pthread_assert( ((nprocs == 1) || (nprocs == 2) || (nprocs == 4)),
|
|---|
| 438 | "[TRANSPOSE ERROR] number of procs per cluster must be 1, 2 or 4");
|
|---|
| 439 |
|
|---|
| 440 | giet_pthread_assert( ((x_size == 1) || (x_size == 2) || (x_size == 4) ||
|
|---|
| 441 | (x_size == 8) || (x_size == 16)),
|
|---|
| 442 | "[TRANSPOSE ERROR] x_size must be 1,2,4,8,16");
|
|---|
| 443 |
|
|---|
| 444 | giet_pthread_assert( ((y_size == 1) || (y_size == 2) || (y_size == 4) ||
|
|---|
| 445 | (y_size == 8) || (y_size == 16)),
|
|---|
| 446 | "[TRANSPOSE ERROR] y_size must be 1,2,4,8,16");
|
|---|
| 447 |
|
|---|
| 448 | // compute number of threads
|
|---|
| 449 | unsigned int nthreads = x_size * y_size * nprocs;
|
|---|
| 450 |
|
|---|
| 451 | // shared TTY allocation
|
|---|
| 452 | giet_tty_alloc( 1 );
|
|---|
| 453 | lock_init( &tty_lock);
|
|---|
| 454 |
|
|---|
| 455 | // get FBF ownership and FBF size
|
|---|
| 456 | unsigned int width;
|
|---|
| 457 | unsigned int height;
|
|---|
| 458 | giet_fbf_alloc();
|
|---|
| 459 | giet_fbf_size( &width , &height );
|
|---|
| 460 |
|
|---|
| 461 | printf("\n[TRANSPOSE] start at cycle %d on %d cores / FBF = %d * %d pixels\n",
|
|---|
| 462 | giet_proctime(), nthreads , width , height );
|
|---|
| 463 |
|
|---|
| 464 | if ( INTERACTIVE ) // input_file_name, output_file_name, and size acquisition
|
|---|
| 465 | {
|
|---|
| 466 | printf("\n[TRANSPOSE] enter path for input file / default is : %s\n> ", INPUT_FILE_PATH );
|
|---|
| 467 | giet_tty_gets( input_file_name , 256 );
|
|---|
| 468 | printf("\n");
|
|---|
| 469 | if ( strcmp( input_file_name , "" ) == 0 ) strcpy( input_file_name , INPUT_FILE_PATH );
|
|---|
| 470 |
|
|---|
| 471 | printf("\n[TRANSPOSE] enter path for output file / default is : %s\n> ", OUTPUT_FILE_PATH );
|
|---|
| 472 | giet_tty_gets( output_file_name , 256 );
|
|---|
| 473 | printf("\n");
|
|---|
| 474 | if ( strcmp( output_file_name , "" ) == 0 ) strcpy( output_file_name , OUTPUT_FILE_PATH );
|
|---|
| 475 |
|
|---|
| 476 | printf("\n[TRANSPOSE] enter image size / default is : %d\n> ", IMAGE_SIZE );
|
|---|
| 477 | giet_tty_getw( &image_size );
|
|---|
| 478 | printf("\n");
|
|---|
| 479 | if ( image_size == 0 ) image_size = IMAGE_SIZE;
|
|---|
| 480 | }
|
|---|
| 481 | else
|
|---|
| 482 | {
|
|---|
| 483 | strcpy( input_file_name , INPUT_FILE_PATH );
|
|---|
| 484 | strcpy( output_file_name , OUTPUT_FILE_PATH );
|
|---|
| 485 | image_size = IMAGE_SIZE;
|
|---|
| 486 | }
|
|---|
| 487 |
|
|---|
| 488 | // check image size / number of threads
|
|---|
| 489 | giet_pthread_assert( (nprocs * x_size * y_size <= image_size ),
|
|---|
| 490 | "[TRANSPOSE ERROR] number of threads larger than number of lines");
|
|---|
| 491 |
|
|---|
| 492 | printf("\n[TRANSPOSE] input = %s / output = %s / size = %d\n",
|
|---|
| 493 | input_file_name, output_file_name, image_size );
|
|---|
| 494 |
|
|---|
| 495 | // distributed heap initialisation
|
|---|
| 496 | for ( x = 0 ; x < x_size ; x++ )
|
|---|
| 497 | {
|
|---|
| 498 | for ( y = 0 ; y < y_size ; y++ )
|
|---|
| 499 | {
|
|---|
| 500 | heap_init( x , y );
|
|---|
| 501 | }
|
|---|
| 502 | }
|
|---|
| 503 |
|
|---|
| 504 | // allocate thread[] array
|
|---|
| 505 | pthread_t* thread = malloc( nthreads * sizeof(pthread_t) );
|
|---|
| 506 |
|
|---|
| 507 | // barrier initialisation
|
|---|
| 508 | sqt_barrier_init( &barrier, x_size , y_size , nprocs );
|
|---|
| 509 |
|
|---|
| 510 | // Initialisation completed
|
|---|
| 511 | printf("\n[TRANSPOSE] initialisation completed at cycle %d\n", giet_proctime() );
|
|---|
| 512 |
|
|---|
| 513 | // launch other threads to run execute() function
|
|---|
| 514 | for ( n = 1 ; n < nthreads ; n++ )
|
|---|
| 515 | {
|
|---|
| 516 | if ( giet_pthread_create( &thread[n],
|
|---|
| 517 | NULL, // no attribute
|
|---|
| 518 | &execute,
|
|---|
| 519 | NULL ) ) // no argument
|
|---|
| 520 | {
|
|---|
| 521 | printf("\n[TRANSPOSE ERROR] creating thread %x\n", thread[n] );
|
|---|
| 522 | giet_pthread_exit( NULL );
|
|---|
| 523 | }
|
|---|
| 524 | }
|
|---|
| 525 |
|
|---|
| 526 | // run the execute() function
|
|---|
| 527 | execute();
|
|---|
| 528 |
|
|---|
| 529 | // wait other threads completion
|
|---|
| 530 | for ( n = 1 ; n < nthreads ; n++ )
|
|---|
| 531 | {
|
|---|
| 532 | if ( giet_pthread_join( thread[n], NULL ) )
|
|---|
| 533 | {
|
|---|
| 534 | printf("\n[TRANSPOSE ERROR] joining thread %x\n", thread[n] );
|
|---|
| 535 | giet_pthread_exit( NULL );
|
|---|
| 536 | }
|
|---|
| 537 | else
|
|---|
| 538 | {
|
|---|
| 539 | printf("\n[TRANSPOSE] thread %x joined at cycle %d\n",
|
|---|
| 540 | thread[n] , giet_proctime() );
|
|---|
| 541 | }
|
|---|
| 542 | }
|
|---|
| 543 |
|
|---|
| 544 | // call the instrument() function
|
|---|
| 545 | instrument( x_size , y_size , nprocs );
|
|---|
| 546 |
|
|---|
| 547 | giet_pthread_exit( "completed" );
|
|---|
| 548 |
|
|---|
| 549 | } // end main()
|
|---|
| 550 |
|
|---|
![(please configure the [header_logo] section in trac.ini)](/trac/giet-vm/chrome/site/your_project_logo.png){kind=logo}
