| 1 |
32 |
alirezamon |
#include <stdlib.h>
|
| 2 |
|
|
#include <stdio.h>
|
| 3 |
|
|
#include <unistd.h>
|
| 4 |
|
|
#include <string.h>
|
| 5 |
41 |
alirezamon |
#include <limits.h>
|
| 6 |
32 |
alirezamon |
#include <ctype.h>
|
| 7 |
|
|
#include <stdint.h>
|
| 8 |
|
|
#include <inttypes.h>
|
| 9 |
|
|
#include <verilated.h> // Defines common routines
|
| 10 |
48 |
alirezamon |
|
| 11 |
32 |
alirezamon |
#include "Vtraffic.h"
|
| 12 |
48 |
alirezamon |
#include "Vpck_inj.h"
|
| 13 |
|
|
#include <thread>
|
| 14 |
|
|
#include <vector>
|
| 15 |
|
|
#include <atomic>
|
| 16 |
32 |
alirezamon |
|
| 17 |
|
|
|
| 18 |
48 |
alirezamon |
#include <cstdint>
|
| 19 |
|
|
#include <cstdlib>
|
| 20 |
|
|
#include <iostream>
|
| 21 |
32 |
alirezamon |
|
| 22 |
48 |
alirezamon |
#include "simulator.h"
|
| 23 |
32 |
alirezamon |
|
| 24 |
48 |
alirezamon |
int main(int argc, char** argv) {
|
| 25 |
|
|
char change_injection_ratio=0;
|
| 26 |
|
|
int i,j,x,y;//,report_delay_counter=0;
|
| 27 |
32 |
alirezamon |
|
| 28 |
48 |
alirezamon |
char deafult_out[] = {"result"};
|
| 29 |
|
|
NEw=Log2(NE);
|
| 30 |
|
|
for(i=0;i<NE;i++) custom_traffic_table[i]=INJECT_OFF; //off
|
| 31 |
|
|
Verilated::commandArgs(argc, argv); // Remember args
|
| 32 |
|
|
processArgs ( argc, argv );
|
| 33 |
|
|
|
| 34 |
|
|
if (class_percentage==NULL) {
|
| 35 |
|
|
class_percentage = (int *) malloc(sizeof(int));
|
| 36 |
|
|
class_percentage[0]=100;
|
| 37 |
42 |
alirezamon |
}
|
| 38 |
32 |
alirezamon |
|
| 39 |
|
|
|
| 40 |
48 |
alirezamon |
Vrouter_new();
|
| 41 |
|
|
if( TRAFFIC_TYPE == NETRACE) for(i=0;i<NE;i++) pck_inj[i] = new Vpck_inj;
|
| 42 |
|
|
else for(i=0;i<NE;i++) traffic[i] = new Vtraffic;
|
| 43 |
32 |
alirezamon |
|
| 44 |
|
|
|
| 45 |
48 |
alirezamon |
if( TRAFFIC_TYPE == NETRACE) netrace_init(netrace_file);
|
| 46 |
|
|
|
| 47 |
|
|
FIXED_SRC_DST_PAIR = strcmp (TRAFFIC,"RANDOM") & strcmp(TRAFFIC,"HOTSPOT") & strcmp(TRAFFIC,"random") & strcmp(TRAFFIC,"hot spot") & strcmp(TRAFFIC,"TASK");
|
| 48 |
|
|
|
| 49 |
|
|
|
| 50 |
|
|
/********************
|
| 51 |
|
|
* initialize input
|
| 52 |
|
|
*********************/
|
| 53 |
|
|
|
| 54 |
|
|
reset=1;
|
| 55 |
|
|
reset_all_register();
|
| 56 |
|
|
start_i=0;
|
| 57 |
|
|
topology_init();
|
| 58 |
|
|
if( TRAFFIC_TYPE == NETRACE) pck_inj_init();
|
| 59 |
|
|
else traffic_gen_init();
|
| 60 |
|
|
main_time=0;
|
| 61 |
|
|
print_parameter();
|
| 62 |
|
|
if( thread_num>1) initial_threads();
|
| 63 |
|
|
|
| 64 |
|
|
while (!Verilated::gotFinish()) {
|
| 65 |
|
|
|
| 66 |
|
|
if (main_time-saved_time >= 10 ) {
|
| 67 |
|
|
reset = 0;
|
| 68 |
|
|
}
|
| 69 |
|
|
|
| 70 |
|
|
if(main_time == saved_time+21){ count_en=1; start_i=1;}
|
| 71 |
|
|
if(main_time == saved_time+23) start_i=0;
|
| 72 |
|
|
|
| 73 |
|
|
if(TRAFFIC_TYPE==NETRACE) netrace_posedge_event();
|
| 74 |
|
|
else traffic_clk_posedge_event();
|
| 75 |
|
|
//The valus of all registers and input ports valuse change @ posedge of the clock. Once clk is deasserted, as multiple modules are connected inside the testbench we need several eval for propogating combinational logic values
|
| 76 |
|
|
//between modules when the clock .
|
| 77 |
|
|
for (i=0;i<SMART_MAX+2;i++) {
|
| 78 |
|
|
if(TRAFFIC_TYPE==NETRACE) netrace_clk_negedge_event( );
|
| 79 |
|
|
else traffic_clk_negedge_event( );
|
| 80 |
|
|
}
|
| 81 |
|
|
|
| 82 |
|
|
if(simulation_done){
|
| 83 |
|
|
if( TRAFFIC_TYPE == NETRACE) netrace_final_report();
|
| 84 |
|
|
else traffic_gen_final_report();
|
| 85 |
|
|
sim_final_all();
|
| 86 |
|
|
return 0;
|
| 87 |
|
|
}
|
| 88 |
|
|
|
| 89 |
|
|
main_time++;
|
| 90 |
|
|
|
| 91 |
|
|
}//Simulating is done
|
| 92 |
|
|
|
| 93 |
|
|
sim_final_all();
|
| 94 |
|
|
return 0;
|
| 95 |
|
|
|
| 96 |
32 |
alirezamon |
}
|
| 97 |
|
|
|
| 98 |
|
|
|
| 99 |
48 |
alirezamon |
#define __FILENAME__ (__FILE__ + SOURCE_PATH_SIZE)
|
| 100 |
|
|
|
| 101 |
|
|
void usage(char * bin_name){
|
| 102 |
|
|
printf("Usage:\n"
|
| 103 |
|
|
" %s -t <synthetic Traffic Pattern name> [synthetic Traffic options]\n"
|
| 104 |
|
|
" %s -f <Task file> [Task options] or\n"
|
| 105 |
|
|
" %s -F <netrace file> [Netrace options] \n\n"
|
| 106 |
|
|
"synthetic Traffic options:\n"
|
| 107 |
|
|
" -t <Traffic Pattern> \"HOTSPOT\", \"RANDOM\", \"BIT_COMPLEMENT\" , \"BIT_REVERSE\",\n "
|
| 108 |
|
|
" \"TORNADO\", \"TRANSPOSE1\", \"TRANSPOSE2\", \"SHUFFEL\", \"CUSTOM\"\n"
|
| 109 |
|
|
" -m <Packet size info> packet size format Random-Range or Random-discrete:\n"
|
| 110 |
|
|
" Random-Range : \"R,MIN,MAX\" : The injected packets' size in flits are\n"
|
| 111 |
|
|
" randomly selected in range MIN <= PCK_size <=MAX \n"
|
| 112 |
|
|
" Random-discrete: \"D,S1,S2,..Sn,P,P1,P2,P3,...Pn\": Si are the discrete\n"
|
| 113 |
|
|
" set of numbers representing packet size. The injected packet size is\n"
|
| 114 |
|
|
" randomly selected among these discrete values according to associated\n"
|
| 115 |
|
|
" probability values.\n"
|
| 116 |
|
|
" -c <sim_end_clk_num> Simulation will stop when simulation clock number reach this value\n"
|
| 117 |
|
|
" -n <sim_end_pck_num> Simulation will stop when total of sent packets to the noc reaches this number\n"
|
| 118 |
|
|
" -i <injection ratio> flit injection ratio in percentage\n"
|
| 119 |
|
|
" -p <class traffic ratios> The percentage of traffic injected for each class. represented in string\n"
|
| 120 |
|
|
" each class ratio is separated by comma. \"n0,n1,n2..\" \n"
|
| 121 |
|
|
" -h <HOTSPOT traffic format> represented in a string with following format:\n"
|
| 122 |
|
|
" total number of hotspot nodes, first hotspot node ID, first hotspot node\n"
|
| 123 |
|
|
" send enable(1 or 0),first hotspot node percentage x10,second hotspot node ...\n"
|
| 124 |
|
|
" -H <custom traffic pattern> custom traffic pattern: represented in a string with following format:\n"
|
| 125 |
|
|
" \"SRC1,DEST1, SRC2,DEST2, .., SRCn, DESTn\" \n"
|
| 126 |
|
|
" -T <thread-num> total number of threads. The default is one (no-thread). \n"
|
| 127 |
|
|
//" -Q Quick (fast) simulation. ignore evaluating non-active routers \n"
|
| 128 |
|
|
" to speed up simulation time"
|
| 129 |
|
|
"\nTrace options:\n"
|
| 130 |
|
|
" -f <Task file> path to the task file. any custom task file can be generated using ProNoC gui\n"
|
| 131 |
|
|
" -c <sim_end_clk_num> Simulation will stop when simulation clock number reach this value \n"
|
| 132 |
|
|
" -T <thread-num> total number of threads. The default is one (no-thread). \n"
|
| 133 |
|
|
//" -Q Quick (fast) simulation. ignore evaluating non-active routers \n"
|
| 134 |
|
|
" to speed up simulation time"
|
| 135 |
|
|
"\nNetrace options:\n"
|
| 136 |
|
|
" -F <Netrace file> path to the task file. any custom task file can be generated using ProNoC gui\n"
|
| 137 |
|
|
" -n <sim_end_pck_num> Simulation will stop when total of sent packets to the NoC reaches this number\n"
|
| 138 |
|
|
" -d ignore dependencies\n"
|
| 139 |
|
|
" -r <start region> start region\n"
|
| 140 |
|
|
" -l reader throttling\n"
|
| 141 |
|
|
" -v <level> Verbosity level. 0: off, 1:display a live number of injected packet,\n"
|
| 142 |
|
|
" 3: print injected/ejected packets details, default is 1\n"
|
| 143 |
|
|
" -T <thread-num> total number of threads. The default is one (no-thread). \n"
|
| 144 |
|
|
" -s <speed-up-num> the speed-up-num is the ratio of netrace frequency to pronoc.The higher value\n"
|
| 145 |
|
|
" results in higher injection ratio to the NoC. Default is one"
|
| 146 |
|
|
//" -Q Quick (fast) simulation. ignore evaluating non-active routers \n"
|
| 147 |
|
|
" to speed up simulation time" ,
|
| 148 |
|
|
bin_name,bin_name,bin_name
|
| 149 |
|
|
);
|
| 150 |
|
|
|
| 151 |
|
|
}
|
| 152 |
|
|
|
| 153 |
|
|
|
| 154 |
|
|
|
| 155 |
|
|
void netrace_processArgs (int argc, char **argv )
|
| 156 |
|
|
{
|
| 157 |
|
|
char c;
|
| 158 |
|
|
|
| 159 |
|
|
/* don't want getopt to moan - I can do that just fine thanks! */
|
| 160 |
|
|
opterr = 0;
|
| 161 |
|
|
if (argc < 2) usage(argv[0]);
|
| 162 |
|
|
while ((c = getopt (argc, argv, "F:dr:lv:T:n:s:")) != -1)
|
| 163 |
|
|
{
|
| 164 |
|
|
switch (c)
|
| 165 |
|
|
{
|
| 166 |
|
|
case 'F':
|
| 167 |
|
|
TRAFFIC_TYPE=NETRACE;
|
| 168 |
|
|
TRAFFIC=(char *) "NETRACE";
|
| 169 |
|
|
netrace_file = optarg;
|
| 170 |
|
|
break;
|
| 171 |
|
|
case 'd':
|
| 172 |
|
|
ignore_dependencies=1;
|
| 173 |
|
|
break;
|
| 174 |
|
|
case 'r':
|
| 175 |
|
|
start_region=atoi(optarg);
|
| 176 |
|
|
break;
|
| 177 |
|
|
case 'l':
|
| 178 |
|
|
reader_throttling=1;
|
| 179 |
|
|
break;
|
| 180 |
|
|
case 'v':
|
| 181 |
|
|
verbosity= atoi(optarg);
|
| 182 |
|
|
break;
|
| 183 |
|
|
case 'T':
|
| 184 |
|
|
thread_num = atoi(optarg);
|
| 185 |
|
|
break;
|
| 186 |
|
|
case 'n':
|
| 187 |
|
|
end_sim_pck_num=atoi(optarg);
|
| 188 |
|
|
break;
|
| 189 |
|
|
case 's':
|
| 190 |
|
|
netrace_speed_up=atoi(optarg);
|
| 191 |
|
|
|
| 192 |
|
|
break;
|
| 193 |
|
|
case '?':
|
| 194 |
|
|
if (isprint (optopt))
|
| 195 |
|
|
fprintf (stderr, "Unknown option `-%c'.\n", optopt);
|
| 196 |
|
|
else
|
| 197 |
|
|
fprintf (stderr, "Unknown option character `\\x%x'.\n", optopt);
|
| 198 |
|
|
default:
|
| 199 |
|
|
usage(argv[0]);
|
| 200 |
|
|
exit(1);
|
| 201 |
|
|
}
|
| 202 |
|
|
}
|
| 203 |
|
|
}
|
| 204 |
|
|
|
| 205 |
|
|
|
| 206 |
|
|
|
| 207 |
|
|
void synthetic_task_processArgs (int argc, char **argv )
|
| 208 |
|
|
{
|
| 209 |
|
|
char c;
|
| 210 |
|
|
int p;
|
| 211 |
|
|
int array[10];
|
| 212 |
|
|
float f;
|
| 213 |
|
|
|
| 214 |
|
|
/* don't want getopt to moan - I can do that just fine thanks! */
|
| 215 |
|
|
opterr = 0;
|
| 216 |
|
|
if (argc < 2) usage(argv[0]);
|
| 217 |
|
|
while ((c = getopt (argc, argv, "t:m:n:c:i:p:h:H:f:T:Q")) != -1)
|
| 218 |
|
|
{
|
| 219 |
|
|
switch (c)
|
| 220 |
|
|
{
|
| 221 |
|
|
case 'f':
|
| 222 |
|
|
TRAFFIC_TYPE=TASK;
|
| 223 |
|
|
TRAFFIC=(char *) "TASK";
|
| 224 |
|
|
task_traffic_init(optarg);
|
| 225 |
|
|
break;
|
| 226 |
|
|
case 't':
|
| 227 |
|
|
TRAFFIC=optarg;
|
| 228 |
|
|
total_active_routers=-1;
|
| 229 |
|
|
break;
|
| 230 |
|
|
case 's':
|
| 231 |
|
|
MIN_PACKET_SIZE=atoi(optarg);
|
| 232 |
|
|
break;
|
| 233 |
|
|
case 'n':
|
| 234 |
|
|
end_sim_pck_num=atoi(optarg);
|
| 235 |
|
|
break;
|
| 236 |
|
|
case 'c':
|
| 237 |
|
|
sim_end_clk_num=atoi(optarg);
|
| 238 |
|
|
break;
|
| 239 |
|
|
case 'i':
|
| 240 |
|
|
f=atof(optarg);
|
| 241 |
|
|
f*=(MAX_RATIO/100);
|
| 242 |
|
|
ratio= (int) f;
|
| 243 |
|
|
break;
|
| 244 |
|
|
case 'p':
|
| 245 |
|
|
p= parse_string (optarg, array);
|
| 246 |
|
|
if (p==0) {
|
| 247 |
|
|
printf("Warning: class setting is ignored!\n");
|
| 248 |
|
|
break;
|
| 249 |
|
|
}
|
| 250 |
|
|
class_percentage = (int *) malloc( p * sizeof(int));
|
| 251 |
|
|
for(int k=0;k<p;k++){
|
| 252 |
|
|
class_percentage[k]=array[k];
|
| 253 |
|
|
}
|
| 254 |
|
|
if(p >1 && p>C){
|
| 255 |
|
|
printf("Warning: the number of given class %u is larger than the number of message classes in ProNoC (C=%u)!\n",p,C);
|
| 256 |
|
|
}
|
| 257 |
|
|
break;
|
| 258 |
|
|
case 'm':
|
| 259 |
|
|
update_pck_size(optarg);
|
| 260 |
|
|
|
| 261 |
|
|
break;
|
| 262 |
|
|
case 'H':
|
| 263 |
|
|
update_custom_traffic(optarg);
|
| 264 |
|
|
break;
|
| 265 |
|
|
case 'h':
|
| 266 |
|
|
update_hotspot(optarg);
|
| 267 |
|
|
break;
|
| 268 |
|
|
case 'T':
|
| 269 |
|
|
thread_num = atoi(optarg);
|
| 270 |
|
|
break;
|
| 271 |
|
|
case 'Q':
|
| 272 |
|
|
//Quick_sim_en=1;
|
| 273 |
|
|
fprintf (stderr, "Unknown option `-%c'.\n", optopt);
|
| 274 |
|
|
usage(argv[0]);
|
| 275 |
|
|
exit(1);
|
| 276 |
|
|
break;
|
| 277 |
|
|
case '?':
|
| 278 |
|
|
if (isprint (optopt))
|
| 279 |
|
|
fprintf (stderr, "Unknown option `-%c'.\n", optopt);
|
| 280 |
|
|
else
|
| 281 |
|
|
fprintf (stderr, "Unknown option character `\\x%x'.\n", optopt);
|
| 282 |
|
|
default:
|
| 283 |
|
|
usage(argv[0]);
|
| 284 |
|
|
exit(1);
|
| 285 |
|
|
}
|
| 286 |
|
|
}
|
| 287 |
|
|
}
|
| 288 |
|
|
|
| 289 |
|
|
|
| 290 |
|
|
|
| 291 |
|
|
|
| 292 |
32 |
alirezamon |
int parse_string ( char * str, int * array)
|
| 293 |
|
|
{
|
| 294 |
|
|
int i=0;
|
| 295 |
|
|
char *pt;
|
| 296 |
|
|
pt = strtok (str,",");
|
| 297 |
|
|
while (pt != NULL) {
|
| 298 |
|
|
int a = atoi(pt);
|
| 299 |
|
|
array[i]=a;
|
| 300 |
|
|
i++;
|
| 301 |
|
|
pt = strtok (NULL, ",");
|
| 302 |
|
|
}
|
| 303 |
|
|
return i;
|
| 304 |
|
|
}
|
| 305 |
|
|
|
| 306 |
43 |
alirezamon |
|
| 307 |
48 |
alirezamon |
|
| 308 |
|
|
|
| 309 |
|
|
|
| 310 |
|
|
|
| 311 |
43 |
alirezamon |
unsigned int pck_dst_gen ( unsigned int core_num) {
|
| 312 |
48 |
alirezamon |
if(TRAFFIC_TYPE==TASK) return pck_dst_gen_task_graph ( core_num);
|
| 313 |
43 |
alirezamon |
if((strcmp (TOPOLOGY,"MESH")==0)||(strcmp (TOPOLOGY,"TORUS")==0)) return pck_dst_gen_2D (core_num);
|
| 314 |
|
|
return pck_dst_gen_1D (core_num);
|
| 315 |
|
|
}
|
| 316 |
|
|
|
| 317 |
|
|
|
| 318 |
|
|
|
| 319 |
38 |
alirezamon |
void update_hotspot(char * str){
|
| 320 |
|
|
int i;
|
| 321 |
|
|
int array[1000];
|
| 322 |
|
|
int p;
|
| 323 |
|
|
int acuum=0;
|
| 324 |
|
|
hotspot_st * new_node;
|
| 325 |
|
|
p= parse_string (str, array);
|
| 326 |
|
|
if (p<4){
|
| 327 |
48 |
alirezamon |
fprintf(stderr,"ERROR: in hotspot traffic parameters. 4 value should be given as hotspot parameter\n");
|
| 328 |
38 |
alirezamon |
exit(1);
|
| 329 |
|
|
}
|
| 330 |
|
|
HOTSPOT_NUM=array[0];
|
| 331 |
|
|
if (p<1+HOTSPOT_NUM*3){
|
| 332 |
48 |
alirezamon |
fprintf(stderr,"ERROR: in hotspot traffic parameters \n");
|
| 333 |
38 |
alirezamon |
exit(1);
|
| 334 |
|
|
}
|
| 335 |
|
|
new_node = (hotspot_st *) malloc( HOTSPOT_NUM * sizeof(hotspot_st));
|
| 336 |
|
|
if( new_node == NULL){
|
| 337 |
48 |
alirezamon |
fprintf(stderr,"ERROR: cannot allocate memory for hotspot traffic\n");
|
| 338 |
38 |
alirezamon |
exit(1);
|
| 339 |
|
|
}
|
| 340 |
|
|
for (i=1;i<3*HOTSPOT_NUM; i+=3){
|
| 341 |
|
|
new_node[i/3]. ip_num = array[i];
|
| 342 |
|
|
new_node[i/3]. send_enable=array[i+1];
|
| 343 |
|
|
new_node[i/3]. percentage = acuum + array[i+2];
|
| 344 |
|
|
acuum= new_node[i/3]. percentage;
|
| 345 |
|
|
|
| 346 |
|
|
}
|
| 347 |
|
|
if(acuum> 1000){
|
| 348 |
48 |
alirezamon |
printf("Warning: The hotspot traffic summation %f exceed than 100 percent. \n", (float) acuum /10);
|
| 349 |
43 |
alirezamon |
}
|
| 350 |
38 |
alirezamon |
hotspots=new_node;
|
| 351 |
|
|
}
|
| 352 |
|
|
|
| 353 |
48 |
alirezamon |
void update_custom_traffic (char * str){
|
| 354 |
|
|
int i;
|
| 355 |
|
|
int array[10000];
|
| 356 |
|
|
int p;
|
| 357 |
|
|
p= parse_string (str, array);
|
| 358 |
|
|
for (i=0;i<p; i+=2){
|
| 359 |
|
|
custom_traffic_table[array[i]] = array[i+1];
|
| 360 |
|
|
}
|
| 361 |
32 |
alirezamon |
}
|
| 362 |
|
|
|
| 363 |
48 |
alirezamon |
void update_pck_size(char *str){
|
| 364 |
|
|
int i;
|
| 365 |
|
|
int array[1000];
|
| 366 |
|
|
char substring[1000];
|
| 367 |
|
|
int p;
|
| 368 |
|
|
char *pt,*pt2;
|
| 369 |
|
|
MIN_PACKET_SIZE=100000;
|
| 370 |
|
|
MAX_PACKET_SIZE=1;
|
| 371 |
32 |
alirezamon |
|
| 372 |
43 |
alirezamon |
|
| 373 |
48 |
alirezamon |
pt = strtok (str,",");
|
| 374 |
|
|
if(*pt=='R'){//random range
|
| 375 |
|
|
p= parse_string (str+2, array);
|
| 376 |
|
|
if(p<2){
|
| 377 |
|
|
fprintf(stderr,"ERROR: Wrong Packet size format %s. It should be \"R,min,max\" : \n",str);
|
| 378 |
|
|
exit(1);
|
| 379 |
|
|
}
|
| 380 |
43 |
alirezamon |
|
| 381 |
48 |
alirezamon |
MIN_PACKET_SIZE=array[0];
|
| 382 |
|
|
MAX_PACKET_SIZE=array[1];
|
| 383 |
|
|
AVG_PACKET_SIZE=(MIN_PACKET_SIZE+MAX_PACKET_SIZE)/2;// average packet size
|
| 384 |
|
|
}else if(*pt=='D'){//random discrete
|
| 385 |
|
|
pck_size_sel = RANDOM_discrete;
|
| 386 |
|
|
pt = strtok (str+2,"P");
|
| 387 |
|
|
pt2 = strtok (NULL,"P");
|
| 388 |
|
|
if (pt == NULL || pt2==NULL) {
|
| 389 |
|
|
fprintf(stderr,"ERROR: Wrong Packet size format %s. It should be \"D,s1,s2..sn,P,p1,p2..pn\". missing letter \"P\" in format \n",str);
|
| 390 |
|
|
exit(1);
|
| 391 |
|
|
}
|
| 392 |
|
|
p= parse_string (pt, array);
|
| 393 |
|
|
if (p==0){
|
| 394 |
|
|
fprintf(stderr,"ERROR: Wrong Packet size format %s. It should be \"D,s1,s2..sn,P,p1,p2..pn\". missing si values after letter \"D\" \"P\" in format \n",str);
|
| 395 |
|
|
exit(1);
|
| 396 |
|
|
}
|
| 397 |
|
|
int in=p;
|
| 398 |
|
|
//alocate mmeory for pck size
|
| 399 |
|
|
discrete_size = (int*)malloc((p) * sizeof(int));
|
| 400 |
|
|
discrete_prob = (int*)malloc((p) * sizeof(int));
|
| 401 |
|
|
// Check if the memory has been successfully allocated
|
| 402 |
|
|
if (discrete_size == NULL || discrete_prob==NULL) {
|
| 403 |
|
|
printf("ERROR: Memory not allocated.\n");
|
| 404 |
|
|
exit(1);
|
| 405 |
|
|
}
|
| 406 |
43 |
alirezamon |
|
| 407 |
48 |
alirezamon |
for (i=0; i<p; i++){
|
| 408 |
32 |
alirezamon |
|
| 409 |
48 |
alirezamon |
//printf("I[%u]=%u,\n",i,array[i]);
|
| 410 |
|
|
discrete_size[i] = array[i];
|
| 411 |
|
|
if(MIN_PACKET_SIZE > array[i]) MIN_PACKET_SIZE = array[i];
|
| 412 |
|
|
if(MAX_PACKET_SIZE < array[i]) MAX_PACKET_SIZE = array[i];
|
| 413 |
|
|
}
|
| 414 |
32 |
alirezamon |
|
| 415 |
48 |
alirezamon |
p= parse_string (pt2+1, array);
|
| 416 |
|
|
int sum=0;
|
| 417 |
|
|
AVG_PACKET_SIZE=0;
|
| 418 |
|
|
for (i=0; i<p; i++){
|
| 419 |
|
|
//printf("P[%u]=%u,\n",i,array[i]);
|
| 420 |
|
|
if(i<in){
|
| 421 |
|
|
sum+=array[i];
|
| 422 |
|
|
discrete_prob[i]=sum;
|
| 423 |
|
|
AVG_PACKET_SIZE+=discrete_size[i] * array[i];
|
| 424 |
32 |
alirezamon |
|
| 425 |
48 |
alirezamon |
}
|
| 426 |
|
|
}
|
| 427 |
|
|
AVG_PACKET_SIZE/=100;
|
| 428 |
32 |
alirezamon |
|
| 429 |
48 |
alirezamon |
if(sum!=100){
|
| 430 |
|
|
fprintf(stderr,"ERROR: The accumulatio of the first %u probebility values is %u which is not equal to 100\n",in,sum);
|
| 431 |
|
|
exit(1);
|
| 432 |
32 |
alirezamon |
}
|
| 433 |
|
|
|
| 434 |
48 |
alirezamon |
}else {
|
| 435 |
|
|
fprintf(stderr,"ERROR: Wrong Packet size format %s. It should start with one of \"D\" or \"R\" letter\n",str);
|
| 436 |
|
|
exit(1);
|
| 437 |
|
|
}
|
| 438 |
|
|
p=(MAX_PACKET_SIZE-MIN_PACKET_SIZE)+1;
|
| 439 |
|
|
rsv_size_array = (unsigned int*) calloc ( p , sizeof(int));
|
| 440 |
|
|
if (rsv_size_array==NULL){
|
| 441 |
|
|
fprintf(stderr,"ERROR: cannot allocate memory for rsv_size_array\n");
|
| 442 |
|
|
exit(1);
|
| 443 |
|
|
}
|
| 444 |
43 |
alirezamon |
|
| 445 |
48 |
alirezamon |
}
|
| 446 |
43 |
alirezamon |
|
| 447 |
32 |
alirezamon |
|
| 448 |
48 |
alirezamon |
void task_traffic_init (char * str) {
|
| 449 |
|
|
load_traffic_file(str,task_graph_data,task_graph_abstract);
|
| 450 |
|
|
end_sim_pck_num=task_graph_total_pck_num;
|
| 451 |
|
|
MIN_PACKET_SIZE = task_graph_min_pck_size;
|
| 452 |
|
|
MAX_PACKET_SIZE = task_graph_max_pck_size;
|
| 453 |
|
|
AVG_PACKET_SIZE=(MIN_PACKET_SIZE+MAX_PACKET_SIZE)/2;// average packet size
|
| 454 |
|
|
int p=(MAX_PACKET_SIZE-MIN_PACKET_SIZE)+1;
|
| 455 |
|
|
rsv_size_array = (unsigned int*) calloc ( p , sizeof(int));
|
| 456 |
|
|
if (rsv_size_array==NULL){
|
| 457 |
|
|
fprintf(stderr,"ERROR: cannot allocate (%d x int) memory for rsv_size_array. \n",p);
|
| 458 |
|
|
exit(1);
|
| 459 |
|
|
}
|
| 460 |
|
|
}
|
| 461 |
32 |
alirezamon |
|
| 462 |
|
|
|
| 463 |
|
|
|
| 464 |
|
|
|
| 465 |
|
|
|
| 466 |
38 |
alirezamon |
|
| 467 |
32 |
alirezamon |
|
| 468 |
48 |
alirezamon |
void processArgs (int argc, char **argv ){
|
| 469 |
|
|
int i;
|
| 470 |
|
|
for( i = 1; i < argc; ++i ) {
|
| 471 |
|
|
if( strcmp(argv[i], "-t") == 0 ) {
|
| 472 |
|
|
synthetic_task_processArgs ( argc, argv );
|
| 473 |
|
|
return;
|
| 474 |
|
|
} else if( strcmp(argv[i], "-f") == 0 ) {
|
| 475 |
|
|
synthetic_task_processArgs ( argc, argv );
|
| 476 |
|
|
return;
|
| 477 |
32 |
alirezamon |
|
| 478 |
48 |
alirezamon |
} else if( strcmp(argv[i], "-F") == 0 ) {
|
| 479 |
|
|
netrace_processArgs (argc, argv );
|
| 480 |
|
|
return;
|
| 481 |
|
|
}
|
| 482 |
|
|
}
|
| 483 |
|
|
fprintf (stderr, "you should define one one of Synthetic,Task or nettrace based simulation. \n");
|
| 484 |
|
|
usage(argv[0]);
|
| 485 |
|
|
exit(1);
|
| 486 |
|
|
}
|
| 487 |
32 |
alirezamon |
|
| 488 |
43 |
alirezamon |
|
| 489 |
48 |
alirezamon |
int get_new_pck_size(){
|
| 490 |
|
|
if(pck_size_sel == RANDOM_discrete){
|
| 491 |
|
|
int rnd = rand() % 100; // 0~99
|
| 492 |
|
|
int i=0;
|
| 493 |
|
|
while( rnd > discrete_prob[i] ) i++;
|
| 494 |
|
|
return discrete_size [i];
|
| 495 |
32 |
alirezamon |
}
|
| 496 |
48 |
alirezamon |
//random range
|
| 497 |
|
|
return rnd_between(MIN_PACKET_SIZE,MAX_PACKET_SIZE);
|
| 498 |
|
|
}
|
| 499 |
32 |
alirezamon |
|
| 500 |
|
|
|
| 501 |
|
|
|
| 502 |
|
|
|
| 503 |
48 |
alirezamon |
|
| 504 |
|
|
|
| 505 |
|
|
void traffic_gen_final_report(){
|
| 506 |
|
|
int i;
|
| 507 |
|
|
for (i=0;i<NE;i++) if(traffic[i]->pck_number>0) total_active_endp = total_active_endp +1;
|
| 508 |
|
|
printf("\nsimulation results-------------------\n");
|
| 509 |
|
|
printf("\tSimulation clock cycles:%d\n",clk_counter);
|
| 510 |
|
|
printf("\n\tTotal injected packet in different size:\n");
|
| 511 |
|
|
printf("\tflit_size,");
|
| 512 |
|
|
for (i=0;i<=(MAX_PACKET_SIZE - MIN_PACKET_SIZE);i++){
|
| 513 |
|
|
if(rsv_size_array[i]>0) printf("%u,",i+ MIN_PACKET_SIZE);
|
| 514 |
|
|
}
|
| 515 |
|
|
printf("\n\t#pck,");
|
| 516 |
|
|
for (i=0;i<=(MAX_PACKET_SIZE - MIN_PACKET_SIZE);i++){
|
| 517 |
|
|
if(rsv_size_array[i]>0) printf("%u,",rsv_size_array[i]);
|
| 518 |
|
|
}
|
| 519 |
|
|
printf("\n");
|
| 520 |
|
|
|
| 521 |
|
|
// printf(" total received flits:%d\n",total_rsv_flit_number);
|
| 522 |
|
|
// printf(" total sent flits:%d\n",total_sent_flit_number);
|
| 523 |
|
|
print_statistic_new (clk_counter);
|
| 524 |
|
|
|
| 525 |
32 |
alirezamon |
}
|
| 526 |
|
|
|
| 527 |
|
|
|
| 528 |
48 |
alirezamon |
void traffic_gen_init( void ){
|
| 529 |
|
|
int i;
|
| 530 |
|
|
unsigned int dest_e_addr;
|
| 531 |
|
|
for (i=0;i<NE;i++){
|
| 532 |
|
|
random_var[i] = 100;
|
| 533 |
|
|
traffic[i]->current_e_addr = endp_addr_encoder(i);
|
| 534 |
|
|
traffic[i]->start=0;
|
| 535 |
|
|
traffic[i]->pck_class_in= pck_class_in_gen( i);
|
| 536 |
|
|
traffic[i]->pck_size_in=get_new_pck_size();
|
| 537 |
|
|
dest_e_addr=pck_dst_gen (i);
|
| 538 |
|
|
traffic[i]->dest_e_addr= dest_e_addr;
|
| 539 |
|
|
if(dest_e_addr == INJECT_OFF) traffic[i]->stop=1;
|
| 540 |
|
|
//printf("src=%u, des_eaddr=%x, dest=%x\n", i,dest_e_addr, endp_addr_decoder(dest_e_addr));
|
| 541 |
|
|
if(inject_done) traffic[i]->stop=1;
|
| 542 |
|
|
traffic[i]->start_delay=rnd_between(1,4*NE-2);
|
| 543 |
|
|
if(TRAFFIC_TYPE==SYNTHETIC){
|
| 544 |
|
|
//traffic[i]->avg_pck_size_in=AVG_PACKET_SIZE;
|
| 545 |
|
|
traffic[i]->ratio=ratio;
|
| 546 |
|
|
traffic[i]->init_weight=1;
|
| 547 |
|
|
}
|
| 548 |
|
|
}
|
| 549 |
|
|
}
|
| 550 |
32 |
alirezamon |
|
| 551 |
48 |
alirezamon |
void pck_inj_init (void){
|
| 552 |
|
|
int i;
|
| 553 |
|
|
for (i=0;i<NE;i++){
|
| 554 |
|
|
pck_inj[i]->current_e_addr = endp_addr_encoder(i);
|
| 555 |
|
|
//TODO mapping should be done according to number of NE and should be set by the user larer
|
| 556 |
|
|
netrace_to_pronoc_map[i]=(int)((i* NE)/header->num_nodes);
|
| 557 |
|
|
pck_inj[i]->pck_injct_in_ready= (0x1<<V)-1;
|
| 558 |
|
|
pck_inj[i]->pck_injct_in_pck_wr=0;
|
| 559 |
|
|
//pronoc_to_netrace_map[i]=i;
|
| 560 |
32 |
alirezamon |
|
| 561 |
48 |
alirezamon |
}
|
| 562 |
|
|
|
| 563 |
|
|
}
|
| 564 |
|
|
|
| 565 |
38 |
alirezamon |
/*************
|
| 566 |
|
|
* sc_time_stamp
|
| 567 |
|
|
*
|
| 568 |
|
|
* **********/
|
| 569 |
32 |
alirezamon |
double sc_time_stamp () { // Called by $time in Verilog
|
| 570 |
|
|
return main_time;
|
| 571 |
|
|
}
|
| 572 |
|
|
|
| 573 |
|
|
int pow2( int num){
|
| 574 |
|
|
int pw;
|
| 575 |
|
|
pw= (0x1 << num);
|
| 576 |
|
|
return pw;
|
| 577 |
|
|
}
|
| 578 |
|
|
|
| 579 |
48 |
alirezamon |
/*
|
| 580 |
|
|
volatile int * lock;
|
| 581 |
|
|
unsigned int nr_per_thread=0;
|
| 582 |
|
|
unsigned int ne_per_thread=0;
|
| 583 |
32 |
alirezamon |
|
| 584 |
48 |
alirezamon |
void thread_function (int n){
|
| 585 |
|
|
int i;
|
| 586 |
|
|
unsigned int node=0;
|
| 587 |
|
|
while(1){
|
| 588 |
|
|
while(lock[n]==0) std::this_thread::yield();
|
| 589 |
|
|
for(i=0;i<nr_per_thread;i++){
|
| 590 |
|
|
node= (n * nr_per_thread)+i;
|
| 591 |
|
|
if (node >= NR) break;
|
| 592 |
|
|
single_router_eval(node);
|
| 593 |
|
|
}
|
| 594 |
|
|
for(i=0;i<ne_per_thread;i++){
|
| 595 |
|
|
node= (n * ne_per_thread)+i;
|
| 596 |
|
|
if (node >= NE) break;
|
| 597 |
|
|
if( TRAFFIC_TYPE == NETRACE) pck_inj[node]->eval();
|
| 598 |
|
|
else traffic[node]->eval();
|
| 599 |
|
|
}
|
| 600 |
|
|
|
| 601 |
|
|
//router1[n]->eval();
|
| 602 |
|
|
//if( TRAFFIC_TYPE == NETRACE) pck_inj[n]->eval();
|
| 603 |
|
|
//else traffic[n]->eval();
|
| 604 |
|
|
|
| 605 |
|
|
lock[n]=0;
|
| 606 |
|
|
if(n==0) break;//first thread is the main process
|
| 607 |
|
|
}
|
| 608 |
|
|
}
|
| 609 |
|
|
*/
|
| 610 |
32 |
alirezamon |
|
| 611 |
48 |
alirezamon |
class alignas(64) Vthread
|
| 612 |
|
|
{
|
| 613 |
|
|
// Access specifier
|
| 614 |
|
|
public:
|
| 615 |
|
|
std::atomic<bool> ready;
|
| 616 |
|
|
// Data Members
|
| 617 |
|
|
int n;//thread num
|
| 618 |
|
|
int nr_per_thread;
|
| 619 |
|
|
int ne_per_thread;
|
| 620 |
|
|
// Member Functions()
|
| 621 |
|
|
//Parameterized Constructor
|
| 622 |
|
|
|
| 623 |
|
|
|
| 624 |
|
|
void function ( ){
|
| 625 |
|
|
int i;
|
| 626 |
|
|
unsigned int node=0;
|
| 627 |
|
|
while(1){
|
| 628 |
|
|
while(!ready) std::this_thread::yield();
|
| 629 |
|
|
for(i=0;i<nr_per_thread;i++){
|
| 630 |
|
|
node= (n * nr_per_thread)+i;
|
| 631 |
|
|
if (node >= NR) break;
|
| 632 |
|
|
//if(router_is_active[node] | (Quick_sim_en==0))
|
| 633 |
|
|
single_router_eval(node);
|
| 634 |
|
|
}
|
| 635 |
|
|
for(i=0;i<ne_per_thread;i++){
|
| 636 |
|
|
node= (n * ne_per_thread)+i;
|
| 637 |
|
|
if (node >= NE) break;
|
| 638 |
|
|
if( TRAFFIC_TYPE == NETRACE) pck_inj[node]->eval();
|
| 639 |
|
|
else traffic[node]->eval();
|
| 640 |
|
|
}
|
| 641 |
|
|
|
| 642 |
|
|
//router1[n]->eval();
|
| 643 |
|
|
//if( TRAFFIC_TYPE == NETRACE) pck_inj[n]->eval();
|
| 644 |
|
|
//else traffic[n]->eval();
|
| 645 |
|
|
|
| 646 |
|
|
ready=false;
|
| 647 |
|
|
if(n==0) break;//first thread is the main process
|
| 648 |
|
|
}
|
| 649 |
|
|
}
|
| 650 |
|
|
|
| 651 |
|
|
Vthread(int x,int r,int e)
|
| 652 |
|
|
{
|
| 653 |
|
|
n=x; nr_per_thread=r; ne_per_thread=e;
|
| 654 |
|
|
ready=false;
|
| 655 |
|
|
if(n!=0) {
|
| 656 |
|
|
std::thread th {&Vthread::function,this};
|
| 657 |
|
|
th.detach();
|
| 658 |
|
|
}
|
| 659 |
|
|
}
|
| 660 |
|
|
|
| 661 |
|
|
|
| 662 |
|
|
};
|
| 663 |
|
|
|
| 664 |
|
|
Vthread ** thread;
|
| 665 |
|
|
|
| 666 |
|
|
void initial_threads (void){
|
| 667 |
|
|
int i;
|
| 668 |
|
|
//devide nodes equally between threads
|
| 669 |
|
|
unsigned int nr_per_thread=0;
|
| 670 |
|
|
unsigned int ne_per_thread=0;
|
| 671 |
|
|
nr_per_thread = (NR % thread_num)? (unsigned int)(NR/thread_num) + 1 : (unsigned int)(NR/thread_num);
|
| 672 |
|
|
ne_per_thread = (NE % thread_num)? (unsigned int)(NE/thread_num) + 1 : (unsigned int)(NE/thread_num);
|
| 673 |
|
|
|
| 674 |
|
|
//std::vector<std::thread> threads(thread_num-1);
|
| 675 |
|
|
//lock = new int[thread_num];
|
| 676 |
|
|
//for(i=0;i<thread_num;i++) lock [i]=0;
|
| 677 |
|
|
|
| 678 |
|
|
//Dynamically Allocating Memory
|
| 679 |
|
|
thread = (Vthread **) new Vthread * [thread_num];
|
| 680 |
|
|
for(i=0;i<thread_num;i++) thread[i] = new Vthread(i,nr_per_thread,ne_per_thread) ;
|
| 681 |
|
|
|
| 682 |
|
|
//initiates (thread_num-1) number of live thread
|
| 683 |
|
|
//for(i=0;i<thread_num-1;i++) threads[i] = std::thread(&thread_function, (i+1));
|
| 684 |
|
|
//for (auto& th : threads) th.detach();
|
| 685 |
|
|
|
| 686 |
|
|
unsigned maxThreads = std::thread::hardware_concurrency();
|
| 687 |
|
|
printf("Thread is initiated as following:\n"
|
| 688 |
|
|
"\tMax hardware supported threads:%u\n"
|
| 689 |
|
|
"\tthread_num:%u\n"
|
| 690 |
|
|
"\trouter per thread:%u\n"
|
| 691 |
|
|
"\tendpoint per thread:%u\n"
|
| 692 |
|
|
,maxThreads,thread_num,nr_per_thread,ne_per_thread);
|
| 693 |
|
|
}
|
| 694 |
|
|
|
| 695 |
|
|
|
| 696 |
|
|
|
| 697 |
|
|
|
| 698 |
|
|
void sim_eval_all (void){
|
| 699 |
|
|
int i;
|
| 700 |
|
|
if(thread_num>1) {
|
| 701 |
|
|
for(i=0;i<thread_num;i++) thread[i]->ready=true;
|
| 702 |
|
|
//thread_function (0);
|
| 703 |
|
|
thread[0]->function();
|
| 704 |
|
|
for(i=0;i<thread_num;i++)while(thread[i]->ready);
|
| 705 |
|
|
}else{// no thread
|
| 706 |
|
|
//routers_eval();
|
| 707 |
|
|
for(i=0;i<NR;i++){
|
| 708 |
|
|
//if(router_is_active[i] | (Quick_sim_en==0))
|
| 709 |
|
|
single_router_eval(i);
|
| 710 |
|
|
}
|
| 711 |
|
|
if( TRAFFIC_TYPE == NETRACE) for(i=0;i<NE;i++) pck_inj[i]->eval();
|
| 712 |
|
|
else for(i=0;i<NE;i++) traffic[i]->eval();
|
| 713 |
|
|
}
|
| 714 |
|
|
}
|
| 715 |
|
|
|
| 716 |
|
|
void sim_final_all (void){
|
| 717 |
|
|
int i;
|
| 718 |
|
|
routers_final();
|
| 719 |
|
|
if( TRAFFIC_TYPE == NETRACE) for(i=0;i<NE;i++) pck_inj[i]->final();
|
| 720 |
|
|
else for(i=0;i<NE;i++) traffic[i]->final();
|
| 721 |
|
|
//noc->final();
|
| 722 |
|
|
}
|
| 723 |
|
|
|
| 724 |
|
|
void connect_clk_reset_start_all(void){
|
| 725 |
|
|
int i;
|
| 726 |
|
|
//noc-> clk = clk;
|
| 727 |
|
|
//noc-> reset = reset;
|
| 728 |
|
|
if( TRAFFIC_TYPE == NETRACE) {
|
| 729 |
|
|
for(i=0;i<NE;i++) {
|
| 730 |
|
|
pck_inj[i]->reset= reset;
|
| 731 |
|
|
pck_inj[i]->clk = clk;
|
| 732 |
|
|
}
|
| 733 |
|
|
}else {
|
| 734 |
|
|
for(i=0;i<NE;i++) {
|
| 735 |
|
|
start_o[i]=start_i;
|
| 736 |
|
|
traffic[i]->start= start_o[i];
|
| 737 |
|
|
traffic[i]->reset= reset;
|
| 738 |
|
|
traffic[i]->clk = clk;
|
| 739 |
|
|
}
|
| 740 |
|
|
}
|
| 741 |
|
|
connect_routers_reset_clk();
|
| 742 |
|
|
}
|
| 743 |
|
|
|
| 744 |
|
|
|
| 745 |
|
|
void traffic_clk_negedge_event(void){
|
| 746 |
|
|
int i;
|
| 747 |
|
|
clk = 0;
|
| 748 |
|
|
//for (i=0;i<NR;i++) router_is_active [i]=0;
|
| 749 |
|
|
topology_connect_all_nodes ();
|
| 750 |
|
|
|
| 751 |
|
|
for (i=0;i<NE;i++){
|
| 752 |
|
|
if(inject_done) traffic[i]->stop=1;
|
| 753 |
|
|
}
|
| 754 |
|
|
connect_clk_reset_start_all();
|
| 755 |
|
|
sim_eval_all();
|
| 756 |
|
|
}
|
| 757 |
|
|
|
| 758 |
|
|
|
| 759 |
|
|
|
| 760 |
|
|
|
| 761 |
|
|
void traffic_clk_posedge_event(void) {
|
| 762 |
|
|
int i;
|
| 763 |
|
|
unsigned int dest_e_addr;
|
| 764 |
|
|
clk = 1; // Toggle clock
|
| 765 |
|
|
if(count_en) clk_counter++;
|
| 766 |
|
|
inject_done= ((total_sent_pck_num >= end_sim_pck_num) || (clk_counter>= sim_end_clk_num) || total_active_routers == 0);
|
| 767 |
|
|
//if(inject_done) printf("clk_counter=========%d\n",clk_counter);
|
| 768 |
|
|
total_rsv_flit_number_old=total_rsv_flit_number;
|
| 769 |
|
|
for (i=0;i<NE;i++){
|
| 770 |
|
|
// a packet has been received
|
| 771 |
|
|
if(traffic[i]->update & ~reset){
|
| 772 |
|
|
update_noc_statistic (i) ;
|
| 773 |
|
|
}
|
| 774 |
|
|
// the header flit has been sent out
|
| 775 |
|
|
if(traffic[i]->hdr_flit_sent ){
|
| 776 |
|
|
traffic[i]->pck_class_in= pck_class_in_gen( i);
|
| 777 |
|
|
sent_core_total_pck_num[i]++;
|
| 778 |
|
|
traffic[i]->pck_size_in=get_new_pck_size();
|
| 779 |
|
|
if(!FIXED_SRC_DST_PAIR){
|
| 780 |
|
|
dest_e_addr=pck_dst_gen (i);
|
| 781 |
|
|
traffic[i]->dest_e_addr= dest_e_addr;
|
| 782 |
|
|
if(dest_e_addr == INJECT_OFF) traffic[i]->stop=1;
|
| 783 |
|
|
//printf("src=%u, dest=%x\n", i,endp_addr_decoder(dest_e_addr));
|
| 784 |
|
|
}
|
| 785 |
|
|
}
|
| 786 |
|
|
|
| 787 |
|
|
if(traffic[i]->flit_out_wr==1){
|
| 788 |
|
|
total_sent_flit_number++;
|
| 789 |
|
|
#if (C>1)
|
| 790 |
|
|
sent_stat [i][traffic[i]->flit_out_class].flit_num++;
|
| 791 |
|
|
#else
|
| 792 |
|
|
sent_stat [i].flit_num++;
|
| 793 |
|
|
#endif
|
| 794 |
|
|
}
|
| 795 |
|
|
if(traffic[i]->flit_in_wr==1){
|
| 796 |
|
|
total_rsv_flit_number++;
|
| 797 |
|
|
#if (C>1)
|
| 798 |
|
|
rsvd_stat [i][traffic[i]->pck_class_out].flit_num++;
|
| 799 |
|
|
#else
|
| 800 |
|
|
rsvd_stat [i].flit_num++;
|
| 801 |
|
|
#endif
|
| 802 |
|
|
}
|
| 803 |
|
|
if(traffic[i]->hdr_flit_sent==1){
|
| 804 |
|
|
total_sent_pck_num++;
|
| 805 |
|
|
#if (C>1)
|
| 806 |
|
|
sent_stat [i][traffic[i]->flit_out_class].pck_num++;
|
| 807 |
|
|
#else
|
| 808 |
|
|
sent_stat [i].pck_num++;
|
| 809 |
|
|
#endif
|
| 810 |
|
|
}
|
| 811 |
|
|
|
| 812 |
|
|
}//for
|
| 813 |
|
|
|
| 814 |
|
|
|
| 815 |
|
|
if(inject_done){
|
| 816 |
|
|
if(total_rsv_flit_number_old == total_rsv_flit_number){
|
| 817 |
|
|
ideal_rsv_cnt++;
|
| 818 |
|
|
if(ideal_rsv_cnt >= 100){
|
| 819 |
|
|
print_statistic( );
|
| 820 |
|
|
fprintf(stderr,"ERROR: The number of sent (%u) & received flits (%u) were not equal at the end of simulation\n",total_sent_flit_number, total_rsv_flit_number);
|
| 821 |
|
|
exit(1);
|
| 822 |
|
|
}
|
| 823 |
|
|
}
|
| 824 |
|
|
if(total_sent_flit_number == total_rsv_flit_number ) simulation_done=1;
|
| 825 |
|
|
}
|
| 826 |
|
|
connect_clk_reset_start_all();
|
| 827 |
|
|
sim_eval_all();
|
| 828 |
|
|
|
| 829 |
|
|
}
|
| 830 |
|
|
|
| 831 |
|
|
|
| 832 |
32 |
alirezamon |
/**********************************
|
| 833 |
|
|
*
|
| 834 |
|
|
* update_noc_statistic
|
| 835 |
|
|
*
|
| 836 |
|
|
*
|
| 837 |
|
|
*********************************/
|
| 838 |
|
|
|
| 839 |
48 |
alirezamon |
|
| 840 |
|
|
|
| 841 |
|
|
void update_statistic_at_ejection (
|
| 842 |
|
|
int core_num,
|
| 843 |
|
|
unsigned int clk_num_h2h,
|
| 844 |
|
|
unsigned int clk_num_h2t,
|
| 845 |
|
|
unsigned int distance,
|
| 846 |
|
|
unsigned int class_num,
|
| 847 |
|
|
unsigned int src ){
|
| 848 |
|
|
|
| 849 |
|
|
total_rsv_pck_num+=1;
|
| 850 |
|
|
|
| 851 |
|
|
if( TRAFFIC_TYPE != NETRACE){
|
| 852 |
|
|
|
| 853 |
|
|
//old st
|
| 854 |
|
|
if((total_rsv_pck_num & 0Xffff )==0 ) printf(" packet sent total=%d\n",total_rsv_pck_num);
|
| 855 |
38 |
alirezamon |
sum_clk_h2h+=clk_num_h2h;
|
| 856 |
|
|
sum_clk_h2t+=clk_num_h2t;
|
| 857 |
32 |
alirezamon |
#if (STND_DEV_EN)
|
| 858 |
38 |
alirezamon |
sum_clk_pow2+=(double)clk_num_h2h * (double) clk_num_h2h;
|
| 859 |
|
|
sum_clk_pow2_per_class[class_num]+=(double)clk_num_h2h * (double) clk_num_h2h;
|
| 860 |
43 |
alirezamon |
#endif
|
| 861 |
38 |
alirezamon |
sum_clk_per_hop+= ((double)clk_num_h2h/(double)distance);
|
| 862 |
48 |
alirezamon |
//printf("sum_clk_per_hop(%f)+= clk_num_h2h(%u)/distance(%u)\n",sum_clk_per_hop,clk_num_h2h,distance);
|
| 863 |
|
|
total_rsv_pck_num_per_class[class_num]+=1;
|
| 864 |
38 |
alirezamon |
sum_clk_h2h_per_class[class_num]+=clk_num_h2h ;
|
| 865 |
|
|
sum_clk_h2t_per_class[class_num]+=clk_num_h2t ;
|
| 866 |
|
|
sum_clk_per_hop_per_class[class_num]+= ((double)clk_num_h2h/(double)distance);
|
| 867 |
|
|
rsvd_core_total_pck_num[core_num]=rsvd_core_total_pck_num[core_num]+1;
|
| 868 |
48 |
alirezamon |
|
| 869 |
|
|
if (rsvd_core_worst_delay[core_num] < clk_num_h2t) rsvd_core_worst_delay[core_num] = (strcmp (AVG_LATENCY_METRIC,"HEAD_2_TAIL")==0)? clk_num_h2t : clk_num_h2h;
|
| 870 |
|
|
if (sent_core_worst_delay[src] < clk_num_h2t) sent_core_worst_delay[src] = (strcmp (AVG_LATENCY_METRIC,"HEAD_2_TAIL")==0)? clk_num_h2t : clk_num_h2h;
|
| 871 |
|
|
|
| 872 |
|
|
if( traffic[core_num]->pck_size_o >= MIN_PACKET_SIZE && traffic[core_num]->pck_size_o <=MAX_PACKET_SIZE){
|
| 873 |
|
|
if(rsv_size_array!=NULL) rsv_size_array[traffic[core_num]->pck_size_o-MIN_PACKET_SIZE]++;
|
| 874 |
|
|
}
|
| 875 |
|
|
}
|
| 876 |
|
|
//new one TODO remove old st
|
| 877 |
|
|
|
| 878 |
|
|
if(verbosity==0 && TRAFFIC_TYPE == NETRACE) if((total_rsv_pck_num & 0X1FFFF )==0 ) printf(" packet sent total=%d\n",total_rsv_pck_num);
|
| 879 |
|
|
unsigned int latency = (strcmp (AVG_LATENCY_METRIC,"HEAD_2_TAIL")==0)? clk_num_h2t : clk_num_h2h;
|
| 880 |
|
|
#if(C>1)
|
| 881 |
|
|
|
| 882 |
|
|
rsvd_stat[core_num][class_num].pck_num ++;
|
| 883 |
|
|
rsvd_stat[core_num][class_num].sum_clk_h2h +=clk_num_h2h;
|
| 884 |
|
|
rsvd_stat[core_num][class_num].sum_clk_h2t +=clk_num_h2t;
|
| 885 |
|
|
rsvd_stat[core_num][class_num].sum_clk_per_hop+= ((double)clk_num_h2h/(double)distance);
|
| 886 |
|
|
if (rsvd_stat[core_num][class_num].worst_latency < latency ) rsvd_stat[core_num][class_num].worst_latency =latency;
|
| 887 |
|
|
if (rsvd_stat[core_num][class_num].min_latency==0 ) rsvd_stat[core_num][class_num].min_latency =latency;
|
| 888 |
|
|
if (rsvd_stat[core_num][class_num].min_latency > latency ) rsvd_stat[core_num][class_num].min_latency =latency;
|
| 889 |
|
|
if (sent_stat[src ][class_num].worst_latency < latency ) sent_stat[src ][class_num].worst_latency =latency;
|
| 890 |
|
|
if (sent_stat[src ][class_num].min_latency==0 ) sent_stat[src ][class_num].min_latency =latency;
|
| 891 |
|
|
if (sent_stat[src ][class_num].min_latency > latency ) sent_stat[src ][class_num].min_latency =latency;
|
| 892 |
|
|
|
| 893 |
|
|
#if (STND_DEV_EN)
|
| 894 |
|
|
rsvd_stat[core_num][class_num].sum_clk_pow2 += (double)clk_num_h2h * (double) clk_num_h2h;
|
| 895 |
|
|
#endif
|
| 896 |
|
|
#else
|
| 897 |
|
|
rsvd_stat[core_num].pck_num ++;
|
| 898 |
|
|
rsvd_stat[core_num].sum_clk_h2h +=(double)clk_num_h2h;
|
| 899 |
|
|
rsvd_stat[core_num].sum_clk_h2t +=(double)clk_num_h2t;
|
| 900 |
|
|
rsvd_stat[core_num].sum_clk_per_hop+= ((double)clk_num_h2h/(double)distance);
|
| 901 |
|
|
if (rsvd_stat[core_num].worst_latency < latency ) rsvd_stat[core_num].worst_latency=latency;
|
| 902 |
|
|
if (rsvd_stat[core_num].min_latency==0 ) rsvd_stat[core_num].min_latency =latency;
|
| 903 |
|
|
if (rsvd_stat[core_num].min_latency > latency ) rsvd_stat[core_num].min_latency =latency;
|
| 904 |
|
|
if (sent_stat[src ].worst_latency < latency ) sent_stat[src ].worst_latency=latency;
|
| 905 |
|
|
if (sent_stat[src ].min_latency==0 ) sent_stat[src ].min_latency =latency;
|
| 906 |
|
|
if (sent_stat[src ].min_latency > latency ) sent_stat[src ].min_latency =latency;
|
| 907 |
|
|
|
| 908 |
|
|
#if (STND_DEV_EN)
|
| 909 |
|
|
rsvd_stat[core_num].sum_clk_pow2 += (double)clk_num_h2h * (double) clk_num_h2h;
|
| 910 |
|
|
#endif
|
| 911 |
|
|
#endif
|
| 912 |
|
|
|
| 913 |
|
|
|
| 914 |
|
|
|
| 915 |
|
|
|
| 916 |
|
|
|
| 917 |
38 |
alirezamon |
}
|
| 918 |
32 |
alirezamon |
|
| 919 |
48 |
alirezamon |
void update_noc_statistic ( int core_num){
|
| 920 |
|
|
unsigned int clk_num_h2h =traffic[core_num]->time_stamp_h2h;
|
| 921 |
|
|
unsigned int clk_num_h2t =traffic[core_num]->time_stamp_h2t;
|
| 922 |
|
|
unsigned int distance=traffic[core_num]->distance;
|
| 923 |
|
|
unsigned int class_num=traffic[core_num]->pck_class_out;
|
| 924 |
|
|
unsigned int src_e_addr=traffic[core_num]->src_e_addr;
|
| 925 |
|
|
unsigned int src = endp_addr_decoder (src_e_addr);
|
| 926 |
32 |
alirezamon |
|
| 927 |
48 |
alirezamon |
update_statistic_at_ejection ( core_num, clk_num_h2h, clk_num_h2t, distance, class_num, src);
|
| 928 |
32 |
alirezamon |
|
| 929 |
48 |
alirezamon |
|
| 930 |
|
|
}
|
| 931 |
|
|
|
| 932 |
|
|
avg_st_t finilize_statistic (unsigned long int total_clk, statistic_t rsvd_stat){
|
| 933 |
|
|
|
| 934 |
|
|
avg_st_t avg_statistic;
|
| 935 |
|
|
avg_statistic.avg_throughput= ((double)(rsvd_stat.flit_num*100)/NE )/total_clk;
|
| 936 |
|
|
avg_statistic.avg_latency_flit = rsvd_stat.sum_clk_h2h/rsvd_stat.pck_num;
|
| 937 |
|
|
avg_statistic.avg_latency_pck = rsvd_stat.sum_clk_h2t/rsvd_stat.pck_num;
|
| 938 |
|
|
avg_statistic.avg_latency_per_hop = ( rsvd_stat.pck_num==0)? 0 : rsvd_stat.sum_clk_per_hop/rsvd_stat.pck_num;
|
| 939 |
|
|
avg_statistic.avg_pck_siz = ( rsvd_stat.pck_num==0)? 0 : (double)(rsvd_stat.flit_num / rsvd_stat.pck_num);
|
| 940 |
|
|
#if (STND_DEV_EN)
|
| 941 |
|
|
avg_statistic.std_dev =standard_dev( rsvd_stat.sum_clk_pow2,rsvd_stat.pck_num, avg_statistic.avg_latency_flit);
|
| 942 |
|
|
#endif
|
| 943 |
|
|
return avg_statistic;
|
| 944 |
|
|
}
|
| 945 |
|
|
|
| 946 |
|
|
template<typename T>
|
| 947 |
|
|
void myout(T value)
|
| 948 |
|
|
{
|
| 949 |
|
|
std::cout << value << std::endl;
|
| 950 |
|
|
}
|
| 951 |
|
|
template<typename First, typename ... Rest>
|
| 952 |
|
|
void myout(First first, Rest ... rest)
|
| 953 |
|
|
{
|
| 954 |
|
|
std::cout << first << ",";
|
| 955 |
|
|
myout(rest...);
|
| 956 |
|
|
}
|
| 957 |
|
|
|
| 958 |
|
|
void print_st_single (unsigned long int total_clk, statistic_t rsvd_stat, statistic_t sent_stat){
|
| 959 |
|
|
|
| 960 |
|
|
|
| 961 |
|
|
|
| 962 |
|
|
avg_st_t avg;
|
| 963 |
|
|
avg=finilize_statistic (total_clk, rsvd_stat);
|
| 964 |
|
|
|
| 965 |
|
|
myout(
|
| 966 |
|
|
sent_stat.pck_num,
|
| 967 |
|
|
rsvd_stat.pck_num,
|
| 968 |
|
|
sent_stat.flit_num,
|
| 969 |
|
|
rsvd_stat.flit_num,
|
| 970 |
|
|
sent_stat.worst_latency,
|
| 971 |
|
|
rsvd_stat.worst_latency,
|
| 972 |
|
|
sent_stat.min_latency,
|
| 973 |
|
|
rsvd_stat.min_latency,
|
| 974 |
|
|
avg.avg_latency_per_hop,
|
| 975 |
|
|
avg.avg_latency_flit,
|
| 976 |
|
|
avg.avg_latency_pck,
|
| 977 |
|
|
avg.avg_throughput,
|
| 978 |
|
|
avg.avg_pck_siz,
|
| 979 |
|
|
#if (STND_DEV_EN)
|
| 980 |
|
|
avg.std_dev
|
| 981 |
|
|
#endif
|
| 982 |
|
|
);
|
| 983 |
|
|
// printf("\n");
|
| 984 |
|
|
|
| 985 |
|
|
}
|
| 986 |
|
|
|
| 987 |
|
|
|
| 988 |
|
|
void merge_statistic (statistic_t * merge_stat, statistic_t stat_in){
|
| 989 |
|
|
merge_stat->pck_num+=stat_in.pck_num;
|
| 990 |
|
|
merge_stat->flit_num+=stat_in.flit_num;
|
| 991 |
|
|
if(merge_stat->worst_latency < stat_in.worst_latency) merge_stat->worst_latency= stat_in.worst_latency;
|
| 992 |
|
|
if(merge_stat->min_latency == 0 ) merge_stat->min_latency = stat_in.min_latency;
|
| 993 |
|
|
if(merge_stat->min_latency > stat_in.min_latency && stat_in.min_latency!=0 ) merge_stat->min_latency = stat_in.min_latency;
|
| 994 |
|
|
merge_stat->sum_clk_h2h +=stat_in.sum_clk_h2h ;
|
| 995 |
|
|
merge_stat->sum_clk_h2t +=stat_in.sum_clk_h2t ;
|
| 996 |
|
|
merge_stat->sum_clk_per_hop +=stat_in.sum_clk_per_hop;
|
| 997 |
|
|
#if (STND_DEV_EN)
|
| 998 |
|
|
merge_stat->sum_clk_pow2 +=stat_in.sum_clk_pow2;
|
| 999 |
|
|
#endif
|
| 1000 |
|
|
|
| 1001 |
|
|
}
|
| 1002 |
|
|
|
| 1003 |
|
|
void print_statistic_new (unsigned long int total_clk){
|
| 1004 |
32 |
alirezamon |
int i;
|
| 1005 |
48 |
alirezamon |
printf("\n\t#node,"
|
| 1006 |
|
|
"sent_stat.pck_num,"
|
| 1007 |
|
|
"rsvd_stat.pck_num,"
|
| 1008 |
|
|
"sent_stat.flit_num,"
|
| 1009 |
|
|
"rsvd_stat.flit_num,"
|
| 1010 |
|
|
"sent_stat.worst_latency,"
|
| 1011 |
|
|
"rsvd_stat.worst_latency,"
|
| 1012 |
|
|
"sent_stat.min_latency,"
|
| 1013 |
|
|
"rsvd_stat.min_latency,"
|
| 1014 |
|
|
"avg_latency_per_hop,"
|
| 1015 |
|
|
"avg_latency_flit,"
|
| 1016 |
|
|
"avg_latency_pck,"
|
| 1017 |
|
|
"avg_throughput(%%),"
|
| 1018 |
|
|
"avg_pck_size,"
|
| 1019 |
|
|
#if (STND_DEV_EN)
|
| 1020 |
|
|
"avg.std_dev"
|
| 1021 |
|
|
#endif
|
| 1022 |
|
|
"\n");
|
| 1023 |
|
|
|
| 1024 |
|
|
|
| 1025 |
|
|
|
| 1026 |
|
|
#if(C>1)
|
| 1027 |
|
|
int c;
|
| 1028 |
|
|
statistic_t sent_stat_class [NE];
|
| 1029 |
|
|
statistic_t rsvd_stat_class [NE];
|
| 1030 |
|
|
statistic_t sent_stat_per_class [C];
|
| 1031 |
|
|
statistic_t rsvd_stat_per_class [C];
|
| 1032 |
|
|
|
| 1033 |
|
|
memset (&rsvd_stat_class,0,sizeof(statistic_t)*NE);
|
| 1034 |
|
|
memset (&sent_stat_class,0,sizeof(statistic_t)*NE);
|
| 1035 |
|
|
memset (&rsvd_stat_per_class,0,sizeof(statistic_t)*C);
|
| 1036 |
|
|
memset (&sent_stat_per_class,0,sizeof(statistic_t)*C);
|
| 1037 |
|
|
|
| 1038 |
|
|
|
| 1039 |
|
|
for (i=0; i<NE;i++){
|
| 1040 |
|
|
for (c=0; c<C;c++){
|
| 1041 |
|
|
merge_statistic (&rsvd_stat_class[i],rsvd_stat[i][c]);
|
| 1042 |
|
|
merge_statistic (&sent_stat_class[i],sent_stat[i][c]);
|
| 1043 |
|
|
merge_statistic (&rsvd_stat_per_class[c],rsvd_stat[i][c]);
|
| 1044 |
|
|
merge_statistic (&sent_stat_per_class[c],sent_stat[i][c]);
|
| 1045 |
|
|
}
|
| 1046 |
|
|
}
|
| 1047 |
|
|
|
| 1048 |
|
|
|
| 1049 |
|
|
|
| 1050 |
|
|
|
| 1051 |
|
|
#else
|
| 1052 |
|
|
#define sent_stat_class sent_stat
|
| 1053 |
|
|
#define rsvd_stat_class rsvd_stat
|
| 1054 |
32 |
alirezamon |
#endif
|
| 1055 |
|
|
|
| 1056 |
48 |
alirezamon |
|
| 1057 |
|
|
|
| 1058 |
|
|
|
| 1059 |
|
|
|
| 1060 |
|
|
statistic_t rsvd_stat_total, sent_stat_total;
|
| 1061 |
|
|
memset (&rsvd_stat_total,0,sizeof(statistic_t));
|
| 1062 |
|
|
memset (&sent_stat_total,0,sizeof(statistic_t));
|
| 1063 |
|
|
for (i=0; i<NE;i++){
|
| 1064 |
|
|
merge_statistic (&rsvd_stat_total,rsvd_stat_class[i]);
|
| 1065 |
|
|
merge_statistic (&sent_stat_total,sent_stat_class[i]);
|
| 1066 |
|
|
}
|
| 1067 |
|
|
printf("\ttotal,");
|
| 1068 |
|
|
print_st_single (total_clk, rsvd_stat_total,sent_stat_total);
|
| 1069 |
|
|
|
| 1070 |
|
|
#if(C>1)
|
| 1071 |
|
|
for (c=0; c<C;c++){
|
| 1072 |
|
|
printf("\ttotal_class%u,",c);
|
| 1073 |
|
|
print_st_single (total_clk, rsvd_stat_per_class[c],sent_stat_per_class[c]);
|
| 1074 |
|
|
}
|
| 1075 |
32 |
alirezamon |
#endif
|
| 1076 |
48 |
alirezamon |
|
| 1077 |
|
|
for (i=0; i<NE;i++){
|
| 1078 |
|
|
printf("\t%u,",i);
|
| 1079 |
|
|
print_st_single (total_clk, rsvd_stat_class[i],sent_stat_class[i] );
|
| 1080 |
|
|
}
|
| 1081 |
32 |
alirezamon |
}
|
| 1082 |
|
|
|
| 1083 |
|
|
|
| 1084 |
|
|
|
| 1085 |
48 |
alirezamon |
void print_statistic (void){
|
| 1086 |
|
|
double avg_latency_per_hop, avg_latency_flit, avg_latency_pck, avg_throughput,min_avg_latency_per_class;
|
| 1087 |
|
|
int i;
|
| 1088 |
|
|
#if (STND_DEV_EN)
|
| 1089 |
|
|
double std_dev;
|
| 1090 |
|
|
#endif
|
| 1091 |
32 |
alirezamon |
|
| 1092 |
48 |
alirezamon |
avg_throughput= ((double)(total_sent_flit_number*100)/total_active_endp )/clk_counter;
|
| 1093 |
|
|
printf(" Total active Endpoint: %d \n",total_active_endp);
|
| 1094 |
|
|
printf(" Avg throughput is: %f (flits/clk/Total active Endpoint %%)\n", avg_throughput);
|
| 1095 |
|
|
avg_latency_flit = (double)sum_clk_h2h/total_rsv_pck_num;
|
| 1096 |
|
|
avg_latency_pck = (double)sum_clk_h2t/total_rsv_pck_num;
|
| 1097 |
|
|
if(ratio==RATIO_INIT) first_avg_latency_flit=avg_latency_flit;
|
| 1098 |
32 |
alirezamon |
#if (STND_DEV_EN)
|
| 1099 |
48 |
alirezamon |
std_dev= standard_dev( sum_clk_pow2,total_rsv_pck_num, avg_latency_flit);
|
| 1100 |
|
|
printf(" standard_dev = %f\n",std_dev);
|
| 1101 |
|
|
#endif
|
| 1102 |
|
|
avg_latency_per_hop = (double)sum_clk_per_hop/total_rsv_pck_num;
|
| 1103 |
|
|
printf("\nall : \n");
|
| 1104 |
|
|
printf(" Total number of packet = %d \n average latency per hop = %f \n",total_rsv_pck_num,avg_latency_per_hop);
|
| 1105 |
|
|
printf(" average packet latency = %f \n average flit latency = %f \n",avg_latency_pck, avg_latency_flit);
|
| 1106 |
|
|
min_avg_latency_per_class=1000000;
|
| 1107 |
|
|
printf(" Total injected packet in different size:\n");
|
| 1108 |
|
|
for (i=0;i<=(MAX_PACKET_SIZE - MIN_PACKET_SIZE);i++){
|
| 1109 |
|
|
if(rsv_size_array[i]>0) printf("\t %u flit_sized pck = %u\n",i+ MIN_PACKET_SIZE, rsv_size_array[i]);
|
| 1110 |
|
|
}
|
| 1111 |
|
|
printf("\n");
|
| 1112 |
32 |
alirezamon |
|
| 1113 |
48 |
alirezamon |
for(i=0;i<C;i++){
|
| 1114 |
|
|
avg_throughput = (total_rsv_pck_num_per_class[i]>0)? ((double)(total_rsv_pck_num_per_class[i]*AVG_PACKET_SIZE*100)/total_active_endp )/clk_counter:0;
|
| 1115 |
|
|
avg_latency_flit = (total_rsv_pck_num_per_class[i]>0)? (double)sum_clk_h2h_per_class[i]/total_rsv_pck_num_per_class[i]:0;
|
| 1116 |
|
|
avg_latency_pck = (total_rsv_pck_num_per_class[i]>0)? (double)sum_clk_h2t_per_class[i]/total_rsv_pck_num_per_class[i]:0;
|
| 1117 |
|
|
avg_latency_per_hop = (total_rsv_pck_num_per_class[i]>0)? (double)sum_clk_per_hop_per_class[i]/total_rsv_pck_num_per_class[i]:0;
|
| 1118 |
|
|
printf ("\nclass : %d \n",i);
|
| 1119 |
|
|
printf (" Total number of packet = %d \n avg_throughput = %f \n average latency per hop = %f \n ",total_rsv_pck_num_per_class[i],avg_throughput,avg_latency_per_hop);
|
| 1120 |
|
|
printf (" average packet latency = %f \n average flit latency = %f \n",avg_latency_pck,avg_latency_flit);
|
| 1121 |
|
|
if(min_avg_latency_per_class > avg_latency_flit) min_avg_latency_per_class=avg_latency_flit;
|
| 1122 |
|
|
|
| 1123 |
|
|
#if (STND_DEV_EN)
|
| 1124 |
|
|
std_dev= (total_rsv_pck_num_per_class[i]>0)? standard_dev( sum_clk_pow2_per_class[i],total_rsv_pck_num_per_class[i], avg_latency_flit):0;
|
| 1125 |
|
|
printf(" standard_dev = %f\n",std_dev);
|
| 1126 |
32 |
alirezamon |
#endif
|
| 1127 |
48 |
alirezamon |
}//for
|
| 1128 |
|
|
current_avg_latency_flit=min_avg_latency_per_class;
|
| 1129 |
43 |
alirezamon |
for (i=0;i<NE;i++) {
|
| 1130 |
48 |
alirezamon |
printf ("\n\nEnd_point %d\n",i);
|
| 1131 |
|
|
printf ("\n\ttotal number of received packets: %u\n",rsvd_core_total_pck_num[i]);
|
| 1132 |
|
|
printf ("\n\tworst-case-delay of received packets (clks): %u\n",rsvd_core_worst_delay[i] );
|
| 1133 |
|
|
printf ("\n\ttotal number of sent packets: %u\n",traffic[i]->pck_number);
|
| 1134 |
|
|
printf ("\n\tworst-case-delay of sent packets (clks): %u\n",sent_core_worst_delay[i] );
|
| 1135 |
38 |
alirezamon |
}
|
| 1136 |
48 |
alirezamon |
|
| 1137 |
|
|
print_statistic_new (clk_counter);
|
| 1138 |
|
|
|
| 1139 |
|
|
|
| 1140 |
43 |
alirezamon |
}
|
| 1141 |
32 |
alirezamon |
|
| 1142 |
|
|
|
| 1143 |
|
|
void print_parameter (){
|
| 1144 |
48 |
alirezamon |
printf ("NoC parameters:---------------- \n");
|
| 1145 |
32 |
alirezamon |
printf ("\tTopology: %s\n",TOPOLOGY);
|
| 1146 |
|
|
printf ("\tRouting algorithm: %s\n",ROUTE_NAME);
|
| 1147 |
|
|
printf ("\tVC_per port: %d\n", V);
|
| 1148 |
48 |
alirezamon |
printf ("\tNon-local port buffer_width per VC: %d\n", B);
|
| 1149 |
|
|
printf ("\tLocal port buffer_width per VC: %d\n", LB);
|
| 1150 |
43 |
alirezamon |
if((strcmp (TOPOLOGY,"MESH")==0)||(strcmp (TOPOLOGY,"TORUS")==0)){
|
| 1151 |
|
|
printf ("\tRouter num in row: %d \n",T1);
|
| 1152 |
|
|
printf ("\tRouter num in column: %d \n",T2);
|
| 1153 |
48 |
alirezamon |
printf ("\tEndpoint num per router: %d\n",T3);
|
| 1154 |
43 |
alirezamon |
}else if ((strcmp (TOPOLOGY,"RING")==0)||(strcmp (TOPOLOGY,"LINE")==0)){
|
| 1155 |
48 |
alirezamon |
printf ("\tTotal Router num: %d \n",T1);
|
| 1156 |
|
|
printf ("\tEndpoint num per router: %d\n",T3);
|
| 1157 |
43 |
alirezamon |
}
|
| 1158 |
48 |
alirezamon |
else if ((strcmp (TOPOLOGY,"TREE")==0)||(strcmp (TOPOLOGY,"FATTREE")==0)){
|
| 1159 |
43 |
alirezamon |
printf ("\tK: %d \n",T1);
|
| 1160 |
|
|
printf ("\tL: %d \n",T2);
|
| 1161 |
48 |
alirezamon |
} else{ //CUSTOM
|
| 1162 |
|
|
printf ("\tTotal Endpoints number: %d \n",T1);
|
| 1163 |
|
|
printf ("\tTotal Routers number: %d \n",T2);
|
| 1164 |
43 |
alirezamon |
}
|
| 1165 |
32 |
alirezamon |
printf ("\tNumber of Class: %d\n", C);
|
| 1166 |
|
|
printf ("\tFlit data width: %d \n", Fpay);
|
| 1167 |
|
|
printf ("\tVC reallocation mechanism: %s \n", VC_REALLOCATION_TYPE);
|
| 1168 |
|
|
printf ("\tVC/sw combination mechanism: %s \n", COMBINATION_TYPE);
|
| 1169 |
|
|
printf ("\tAVC_ATOMIC_EN:%d \n", AVC_ATOMIC_EN);
|
| 1170 |
|
|
printf ("\tCongestion Index:%d \n",CONGESTION_INDEX);
|
| 1171 |
|
|
printf ("\tADD_PIPREG_AFTER_CROSSBAR:%d\n",ADD_PIPREG_AFTER_CROSSBAR);
|
| 1172 |
38 |
alirezamon |
printf ("\tSSA_EN enabled:%s \n",SSA_EN);
|
| 1173 |
|
|
printf ("\tSwitch allocator arbitration type:%s \n",SWA_ARBITER_TYPE);
|
| 1174 |
43 |
alirezamon |
printf ("\tMinimum supported packet size:%d flit(s) \n",MIN_PCK_SIZE);
|
| 1175 |
48 |
alirezamon |
printf ("\tLoop back is enabled:%s",SELF_LOOP_EN);
|
| 1176 |
|
|
printf ("\tNumber of multihop bypass (SMART max):%d \n",SMART_MAX);
|
| 1177 |
|
|
printf ("NoC parameters:---------------- \n");
|
| 1178 |
|
|
printf ("\nSimulation parameters-------------\n");
|
| 1179 |
32 |
alirezamon |
#if(DEBUG_EN)
|
| 1180 |
38 |
alirezamon |
printf ("\tDebuging is enabled\n");
|
| 1181 |
32 |
alirezamon |
#else
|
| 1182 |
38 |
alirezamon |
printf ("\tDebuging is disabled\n");
|
| 1183 |
32 |
alirezamon |
#endif
|
| 1184 |
48 |
alirezamon |
//if(strcmp (AVG_LATENCY_METRIC,"HEAD_2_TAIL")==0)printf ("\tOutput is the average latency on sending the packet header until receiving tail\n");
|
| 1185 |
|
|
//else printf ("\tOutput is the average latency on sending the packet header until receiving header flit at destination node\n");
|
| 1186 |
32 |
alirezamon |
printf ("\tTraffic pattern:%s\n",TRAFFIC);
|
| 1187 |
48 |
alirezamon |
size_t n = sizeof(class_percentage)/sizeof(class_percentage[0]);
|
| 1188 |
|
|
for(int p=0;p<n; p++){
|
| 1189 |
|
|
printf ("\ttraffic percentage of class %u is : %d\n",p, class_percentage[p]);
|
| 1190 |
|
|
}
|
| 1191 |
38 |
alirezamon |
if(strcmp (TRAFFIC,"HOTSPOT")==0){
|
| 1192 |
|
|
//printf ("\tHot spot percentage: %u\n", HOTSPOT_PERCENTAGE);
|
| 1193 |
32 |
alirezamon |
printf ("\tNumber of hot spot cores: %d\n", HOTSPOT_NUM);
|
| 1194 |
|
|
|
| 1195 |
|
|
}
|
| 1196 |
|
|
//printf ("\tTotal packets sent by one router: %u\n", TOTAL_PKT_PER_ROUTER);
|
| 1197 |
48 |
alirezamon |
if(sim_end_clk_num!=0) printf ("\tSimulation timeout =%d\n", sim_end_clk_num);
|
| 1198 |
|
|
if(end_sim_pck_num!=0) printf ("\tSimulation ends on total packet num of =%d\n", end_sim_pck_num);
|
| 1199 |
|
|
if(TRAFFIC_TYPE!=NETRACE){
|
| 1200 |
|
|
printf ("\tPacket size (min,max,average) in flits: (%u,%u,%u)\n",MIN_PACKET_SIZE,MAX_PACKET_SIZE,AVG_PACKET_SIZE);
|
| 1201 |
|
|
printf ("\tPacket injector FIFO width in flit:%u \n",TIMSTMP_FIFO_NUM);
|
| 1202 |
|
|
}
|
| 1203 |
|
|
if( TRAFFIC_TYPE == SYNTHETIC) printf("\tFlit injection ratio per router is =%f (flits/clk/Total Endpoint %%)\n",(float)ratio*100/MAX_RATIO);
|
| 1204 |
|
|
printf ("Simulation parameters-------------\n");
|
| 1205 |
32 |
alirezamon |
}
|
| 1206 |
|
|
|
| 1207 |
|
|
|
| 1208 |
38 |
alirezamon |
|
| 1209 |
|
|
|
| 1210 |
|
|
|
| 1211 |
32 |
alirezamon |
/************************
|
| 1212 |
|
|
*
|
| 1213 |
|
|
* reset system
|
| 1214 |
|
|
*
|
| 1215 |
|
|
*
|
| 1216 |
|
|
* *******************/
|
| 1217 |
|
|
|
| 1218 |
|
|
void reset_all_register (void){
|
| 1219 |
|
|
int i;
|
| 1220 |
48 |
alirezamon |
total_active_endp=0;
|
| 1221 |
|
|
total_rsv_pck_num=0;
|
| 1222 |
|
|
total_sent_pck_num=0;
|
| 1223 |
32 |
alirezamon |
sum_clk_h2h=0;
|
| 1224 |
|
|
sum_clk_h2t=0;
|
| 1225 |
48 |
alirezamon |
ideal_rsv_cnt=0;
|
| 1226 |
32 |
alirezamon |
#if (STND_DEV_EN)
|
| 1227 |
|
|
sum_clk_pow2=0;
|
| 1228 |
|
|
#endif
|
| 1229 |
|
|
|
| 1230 |
|
|
sum_clk_per_hop=0;
|
| 1231 |
|
|
count_en=0;
|
| 1232 |
|
|
clk_counter=0;
|
| 1233 |
|
|
|
| 1234 |
|
|
for(i=0;i<C;i++)
|
| 1235 |
|
|
{
|
| 1236 |
48 |
alirezamon |
total_rsv_pck_num_per_class[i]=0;
|
| 1237 |
32 |
alirezamon |
sum_clk_h2h_per_class[i]=0;
|
| 1238 |
|
|
sum_clk_h2t_per_class[i]=0;
|
| 1239 |
|
|
sum_clk_per_hop_per_class[i]=0;
|
| 1240 |
|
|
#if (STND_DEV_EN)
|
| 1241 |
|
|
sum_clk_pow2_per_class[i]=0;
|
| 1242 |
|
|
#endif
|
| 1243 |
|
|
|
| 1244 |
|
|
} //for
|
| 1245 |
48 |
alirezamon |
total_sent_flit_number=0;
|
| 1246 |
|
|
|
| 1247 |
|
|
|
| 1248 |
32 |
alirezamon |
}
|
| 1249 |
|
|
|
| 1250 |
|
|
|
| 1251 |
|
|
|
| 1252 |
|
|
|
| 1253 |
|
|
/***********************
|
| 1254 |
|
|
*
|
| 1255 |
|
|
* standard_dev
|
| 1256 |
|
|
*
|
| 1257 |
|
|
* ******************/
|
| 1258 |
|
|
|
| 1259 |
|
|
#if (STND_DEV_EN)
|
| 1260 |
42 |
alirezamon |
/************************
|
| 1261 |
|
|
* std_dev = sqrt[(B-A^2/N)/N] = sqrt [(B/N)- (A/N)^2] = sqrt [B/N - mean^2]
|
| 1262 |
|
|
* A = sum of the values
|
| 1263 |
|
|
* B = sum of the squarded values
|
| 1264 |
|
|
* *************/
|
| 1265 |
32 |
alirezamon |
|
| 1266 |
|
|
double standard_dev( double sum_pow2, unsigned int total_num, double average){
|
| 1267 |
|
|
double std_dev;
|
| 1268 |
42 |
alirezamon |
|
| 1269 |
|
|
/*
|
| 1270 |
|
|
double A, B, N;
|
| 1271 |
|
|
N= total_num;
|
| 1272 |
|
|
A= average * N;
|
| 1273 |
|
|
B= sum_pow2;
|
| 1274 |
32 |
alirezamon |
|
| 1275 |
42 |
alirezamon |
A=(A*A)/N;
|
| 1276 |
|
|
std_dev = (B-A)/N;
|
| 1277 |
32 |
alirezamon |
std_dev = sqrt(std_dev);
|
| 1278 |
42 |
alirezamon |
*/
|
| 1279 |
48 |
alirezamon |
if(total_num==0) return 0;
|
| 1280 |
32 |
alirezamon |
|
| 1281 |
42 |
alirezamon |
std_dev = sum_pow2/(double)total_num; //B/N
|
| 1282 |
|
|
std_dev -= (average*average);// (B/N) - mean^2
|
| 1283 |
|
|
std_dev = sqroot(std_dev);// sqrt [B/N - mean^2]
|
| 1284 |
|
|
|
| 1285 |
32 |
alirezamon |
return std_dev;
|
| 1286 |
|
|
|
| 1287 |
|
|
}
|
| 1288 |
|
|
|
| 1289 |
|
|
#endif
|
| 1290 |
|
|
|
| 1291 |
|
|
|
| 1292 |
|
|
|
| 1293 |
|
|
/**********************
|
| 1294 |
|
|
*
|
| 1295 |
|
|
* pck_class_in_gen
|
| 1296 |
|
|
*
|
| 1297 |
|
|
* *****************/
|
| 1298 |
|
|
|
| 1299 |
|
|
unsigned char pck_class_in_gen(
|
| 1300 |
|
|
unsigned int core_num
|
| 1301 |
|
|
|
| 1302 |
|
|
) {
|
| 1303 |
|
|
unsigned char pck_class_in;
|
| 1304 |
|
|
unsigned char rnd=rand()%100;
|
| 1305 |
48 |
alirezamon |
int c=0;
|
| 1306 |
|
|
int sum=class_percentage[0];
|
| 1307 |
|
|
size_t n = sizeof(class_percentage)/sizeof(class_percentage[0]);
|
| 1308 |
|
|
for(;;){
|
| 1309 |
|
|
if( rnd < sum) return c;
|
| 1310 |
|
|
if( c==n-1 ) return c;
|
| 1311 |
|
|
c++;
|
| 1312 |
|
|
sum+=class_percentage[c];
|
| 1313 |
|
|
}
|
| 1314 |
|
|
return 0;
|
| 1315 |
32 |
alirezamon |
}
|
| 1316 |
|
|
|
| 1317 |
38 |
alirezamon |
|
| 1318 |
32 |
alirezamon |
|
| 1319 |
|
|
|
| 1320 |
38 |
alirezamon |
void update_injct_var(unsigned int src, unsigned int injct_var){
|
| 1321 |
|
|
//printf("before%u=%u\n",src,random_var[src]);
|
| 1322 |
|
|
random_var[src]= rnd_between(100-injct_var, 100+injct_var);
|
| 1323 |
|
|
//printf("after=%u\n",random_var[src]);
|
| 1324 |
|
|
}
|
| 1325 |
|
|
|
| 1326 |
|
|
unsigned int pck_dst_gen_task_graph ( unsigned int src){
|
| 1327 |
|
|
task_t task;
|
| 1328 |
|
|
float f,v;
|
| 1329 |
|
|
|
| 1330 |
|
|
int index = task_graph_abstract[src].active_index;
|
| 1331 |
|
|
|
| 1332 |
|
|
if(index == DISABLE){
|
| 1333 |
|
|
traffic[src]->ratio=0;
|
| 1334 |
|
|
traffic[src]->stop=1;
|
| 1335 |
48 |
alirezamon |
return INJECT_OFF; //disable sending
|
| 1336 |
38 |
alirezamon |
}
|
| 1337 |
|
|
|
| 1338 |
|
|
if( read(task_graph_data[src],index,&task)==0){
|
| 1339 |
|
|
traffic[src]->ratio=0;
|
| 1340 |
|
|
traffic[src]->stop=1;
|
| 1341 |
48 |
alirezamon |
return INJECT_OFF; //disable sending
|
| 1342 |
38 |
alirezamon |
|
| 1343 |
|
|
}
|
| 1344 |
|
|
|
| 1345 |
|
|
if(sent_core_total_pck_num[src] & 0xFF){//sent 255 packets
|
| 1346 |
|
|
//printf("uu=%u\n",task.jnjct_var);
|
| 1347 |
|
|
update_injct_var(src, task.jnjct_var);
|
| 1348 |
|
|
|
| 1349 |
|
|
}
|
| 1350 |
|
|
|
| 1351 |
|
|
task_graph_total_pck_num++;
|
| 1352 |
|
|
task.pck_sent = task.pck_sent +1;
|
| 1353 |
|
|
task.burst_sent= task.burst_sent+1;
|
| 1354 |
|
|
task.byte_sent = task.byte_sent + (task.avg_pck_size * (Fpay/8) );
|
| 1355 |
|
|
|
| 1356 |
|
|
traffic[src]->pck_class_in= pck_class_in_gen(src);
|
| 1357 |
48 |
alirezamon |
//traffic[src]->avg_pck_size_in=task.avg_pck_size;
|
| 1358 |
38 |
alirezamon |
traffic[src]->pck_size_in=rnd_between(task.min_pck_size,task.max_pck_size);
|
| 1359 |
|
|
|
| 1360 |
|
|
f= task.injection_rate;
|
| 1361 |
|
|
v= random_var[src];
|
| 1362 |
|
|
f*= (v /100);
|
| 1363 |
|
|
if(f>100) f= 100;
|
| 1364 |
|
|
f= f * MAX_RATIO / 100;
|
| 1365 |
|
|
|
| 1366 |
|
|
traffic[src]->ratio=(unsigned int)f;
|
| 1367 |
|
|
traffic[src]->init_weight=task.initial_weight;
|
| 1368 |
|
|
|
| 1369 |
|
|
if (task.burst_sent >= task.burst_size){
|
| 1370 |
|
|
task.burst_sent=0;
|
| 1371 |
|
|
task_graph_abstract[src].active_index=task_graph_abstract[src].active_index+1;
|
| 1372 |
|
|
if(task_graph_abstract[src].active_index>=task_graph_abstract[src].total_index) task_graph_abstract[src].active_index=0;
|
| 1373 |
|
|
|
| 1374 |
|
|
}
|
| 1375 |
|
|
|
| 1376 |
|
|
update_by_index(task_graph_data[src],index,task);
|
| 1377 |
|
|
|
| 1378 |
|
|
if (task.byte_sent >= task.bytes){ // This task is done remove it from the queue
|
| 1379 |
|
|
remove_by_index(&task_graph_data[src],index);
|
| 1380 |
|
|
task_graph_abstract[src].total_index = task_graph_abstract[src].total_index-1;
|
| 1381 |
|
|
if(task_graph_abstract[src].total_index==0){ //all tasks are done turned off the core
|
| 1382 |
|
|
task_graph_abstract[src].active_index=-1;
|
| 1383 |
|
|
traffic[src]->ratio=0;
|
| 1384 |
|
|
traffic[src]->stop=1;
|
| 1385 |
|
|
if(total_active_routers!=0) total_active_routers--;
|
| 1386 |
48 |
alirezamon |
return INJECT_OFF;
|
| 1387 |
38 |
alirezamon |
}
|
| 1388 |
|
|
if(task_graph_abstract[src].active_index>=task_graph_abstract[src].total_index) task_graph_abstract[src].active_index=0;
|
| 1389 |
|
|
}
|
| 1390 |
|
|
|
| 1391 |
48 |
alirezamon |
return endp_addr_encoder(task.dst);
|
| 1392 |
38 |
alirezamon |
}
|
| 1393 |
|
|
|
| 1394 |
|
|
|
| 1395 |
|
|
|
| 1396 |
|
|
|
| 1397 |
|
|
|
| 1398 |
|
|
|
