Changeset 641 for trunk/user/fft/fft.c

Timestamp:

Oct 10, 2019, 1:42:04 PM (5 years ago)

Author:

alain

Message:

• Fix several bugs.
• Introduce the "stat" command in KSH.

This almos-mkh version sucessfully executed the FFT application
(65536 complex points) on the TSAR architecture from 1 to 64 cores.

File:

• trunk/user/fft/fft.c (modified) (6 diffs)

trunk/user/fft/fft.c

@@ -92 +92 @@
 // parameters
 
-#define DEFAULT_M               14              // 16384 data points
+#define DEFAULT_M               16              // 16384 data points
 #define USE_DQT_BARRIER         1               // use DDT barrier if non zero
 #define MODE                    COSIN           // DATA array initialisation mode
@@ -135 +135 @@
 
 // instrumentation counters
+unsigned int   pgfault_nr[THREADS_MAX];    // total number of page faults (per thread)
+unsigned int   pgfault_cost[THREADS_MAX];  // total page faults cost (per thread)
+unsigned int   pgfault_max[THREADS_MAX];   // max page faults cost (per thread)
 unsigned int   parallel_time[THREADS_MAX]; // total computation time (per thread)
 unsigned int   sync_time[THREADS_MAX];     // cumulated waiting time in barriers (per thread)
@@ -458 +461 @@
     }
 
-    // get instrumentation results for each thread
+    // initializes global (all threads) instrumentation values
+    unsigned int time_para      = parallel_time[0];
+    unsigned int time_sync      = sync_time[0];
+    unsigned int pgfaults_nr    = 0;
+    unsigned int pgfaults_cost  = 0;
+    unsigned int pgfaults_max   = pgfault_max[0];
+
+    // loop on threads to compute global instrumentation results
     for (tid = 0 ; tid < nthreads ; tid++) 
     {
-        snprintf( string , 256 , "- tid %d : Sequencial %d / Parallel %d / Barrier %d\n",
-        tid, init_time, parallel_time[tid], sync_time[tid] );
+        snprintf( string , 256 ,
+        "- tid %d : Seq %d / Para %d / Sync %d / Pgfaults %d ( cost %d / max %d )\n",
+        tid, init_time, parallel_time[tid], sync_time[tid], 
+        pgfault_nr[tid], (pgfault_cost[tid] / pgfault_nr[tid]) , pgfault_max[tid] );
 
         // save  to instrumentation file
@@ -468 +480 @@
         if( ret < 0 )
         {
-            printf("\n[fft error] cannot write thread %d to file <%s>\n", tid, path );
+            printf("\n[fft error] cannot save thread %d results to file <%s>\n", tid, path );
             printf("%s", string );
             exit(0);
         }
-    }
-
-    // compute min/max values
-    unsigned int min_para = parallel_time[0];
-    unsigned int max_para = parallel_time[0];
-    unsigned int min_sync = sync_time[0];
-    unsigned int max_sync = sync_time[0];
-
-    for (tid = 0 ; tid < nthreads ; tid++) 
-    {
-        if (parallel_time[tid] > max_para)  max_para = parallel_time[tid];
-        if (parallel_time[tid] < min_para)  min_para = parallel_time[tid];
-        if (sync_time[tid]     > max_sync)  max_sync = sync_time[tid];
-        if (sync_time[tid]     < min_sync)  min_sync = sync_time[tid];
-    }
-
-    // display MIN/MAX values on terminal and save to file
-    snprintf( string , 256 , "\n      Sequencial  Parallel       Barrier\n"
-                             "MIN : %d\t | %d\t | %d\t   (cycles)\n" 
-                             "MAX : %d\t | %d\t | %d\t   (cycles)\n",
-                             (int)init_time, (int)min_para, (int)min_sync,
-                             (int)init_time, (int)max_para, (int)max_sync );
+
+        // compute global values
+        if (parallel_time[tid] > time_para)    time_para      = parallel_time[tid];
+        if (sync_time[tid]     > time_sync)    time_sync      = sync_time[tid];
+                                               pgfaults_nr   += pgfault_nr[tid];
+                                               pgfaults_cost += pgfault_cost[tid];
+        if (pgfault_max[tid]   > pgfaults_max) pgfaults_max   = pgfault_max[tid];
+    }
+
+    // display global values on terminal and save to file
+    snprintf( string , 256 ,
+    "\nSeq %d / Para %d / Sync %d / Pgfaults %d ( cost %d / max %d )\n",
+    init_time, time_para, time_sync, pgfaults_nr, (pgfaults_cost / pgfaults_nr), pgfaults_max );
+
     printf("%s", string );
+
+    // save global values to file
     ret = fprintf( f , "%s", string );
+
     if( ret < 0 )
     {
-        printf("\n[fft error] cannot write MIN/MAX to file <%s>\n", path );
+        printf("\n[fft error] cannot save global results to file <%s>\n", path );
+        exit(0);
+    }
+
+    // close instrumentation file
+    ret = fclose( f );
+
+    if( ret < 0 )
+    {
+        printf("\n[fft error] cannot close file <%s>\n", path );
         exit(0);
     }
@@ -504 +520 @@
 #if DEBUG_MAIN
 get_cycle( &debug_cycle );
-printf("\n[fft] main close file <%s> at cycle %d\n",
+printf("\n[fft] main exit <%s> at cycle %d\n",
 path, (unsigned int)debug_cycle );
 #endif
@@ -645 +661 @@
     get_cycle( &parallel_stop );
 
-    // register parallel time
+    // register parallel time in instrumentation counters
     parallel_time[tid] = (unsigned int)(parallel_stop - parallel_start);
 
+    // get work thread info for page faults
+    thread_info_t info;
+    get_thread_info( &info );
+    
+    // register page faults in instrumentation counters
+    pgfault_nr[tid]   = info.false_pgfault_nr + 
+                        info.local_pgfault_nr + 
+                        info.global_pgfault_nr;
+    pgfault_cost[tid] = info.false_pgfault_cost + 
+                        info.local_pgfault_cost + 
+                        info.global_pgfault_cost;
+    pgfault_max[tid]  = info.false_pgfault_max + 
+                        info.local_pgfault_max + 
+                        info.global_pgfault_max;
 #if DEBUG_WORK
 printf("\n[fft] %s : thread %d completes fft / p_start %d / p_stop %d\n",