Fixed sst_view_parse.pl incorrect detection of correlation IDs

Modifications to instrument_file_cpp.py for SYCL support and new requirement for explicit perfstubs finalize call

Author: giltirn

app/sst_view_parse.pl

@@ -62,7 +62,7 @@
     #FOUND NEW ATTRIBUTE: counter 0 {Elements:1 Type:string Value:"Correlation ID" }    
     }elsif($line=~m/counter\s(\d+)\s.*Value:\"Correlation ID\"/){
 	$corid_cidx=$1;
-
+	print "Correlation id maps to counter $corid_cidx\n";
 
     #Parse GPU device/context
     #FOUND NEW ATTRIBUTE: MetaData:0:7:CUDA Context {Elements:1 Type:string Value:"1" }
@@ -160,15 +160,17 @@
 	$cval = $3;
 	$ts = $4;
 
-	if(exists($corids{$cval})){
-	    print "Found matching corid $cval\n";
-	    delete $corids{$cval};
-	    $ncorid_matched++;
-	}else{
-	    print "Found new corid $cval\n";
-	    $corids{$cval} = [$thread, $cid, $cval, $ts];
+	if($cid == $corid_cidx){
+	    if(exists($corids{$cval})){
+		print "Found matching corid $cval\n";
+		delete $corids{$cval};
+		$ncorid_matched++;
+	    }else{
+		print "Found new corid $cval\n";
+		$corids{$cval} = [$thread, $cid, $cval, $ts];
+	    }
 	}
-
+	
 	$counter_n++;
 	if($counter_n == $counter_nrows){
 	    $counter_active = 0;

scripts/instrument_file_cpp.py

@@ -265,7 +265,7 @@ def func_keyword_check(start, tokens, keywords):
     #Use perfstubs, which automatically collects the function info
     toins = "PERFSTUBS_SCOPED_TIMER_FUNC();"
     if func.name == "main":
-        toins = "PERFSTUBS_INITIALIZE(); " + toins
+        toins = "PS_SCOPED_INIT(); " + toins
 
     orig = flines[iline]   
     #print("Line %d char %d is %s,  expect {" % (iline,icol,orig[icol]))
@@ -279,11 +279,22 @@ def func_keyword_check(start, tokens, keywords):
         line_offsets[line] += len(toins)
     else:
         line_offsets[line] = len(toins)
-
+        
 
 f = open(filename + '.inst', 'w')
 f.write('#include "perfstubs_api/timer.h"\n')
-f.write('#ifdef __HIP_DEVICE_COMPILE__\n#undef PERFSTUBS_SCOPED_TIMER_FUNC\n#define PERFSTUBS_SCOPED_TIMER_FUNC()\n#endif\n')
+f.write('#if defined(__HIP_DEVICE_COMPILE__) || defined(__SYCL_DEVICE_ONLY__)\n#undef PERFSTUBS_SCOPED_TIMER_FUNC\n#define PERFSTUBS_SCOPED_TIMER_FUNC()\n#endif\n')
+f.write('''
+struct _ps_scoped_init{
+    _ps_scoped_init(){
+      PERFSTUBS_INITIALIZE();
+    }
+    ~_ps_scoped_init(){
+      PERFSTUBS_FINALIZE();
+    }
+};
+#define PS_SCOPED_INIT() _ps_scoped_init ___ps_scoped_init
+''')
 for line in flines:
     f.write(line)
 f.close()