Implement nosync baseline version for test-cdeque (+ tests and plot)

Author: nhatminhle

libworkstream_df/Makefile

@@ -9,7 +9,7 @@ WSTREAM_DEPS+=		cdeque-native.c.h cdeque-c11.c.h cdeque-native.h cdeque-c11.h
 
 TEST_DESTDIR=		tests
 
-TEST_KINDS=		-c11 -dumbc11 -llscopt -nolwcas -nofences
+TEST_KINDS=		-c11 -dumbc11 -llscopt -nolwcas -nofences -nosync
 TEST_EXP=		$(TEST_DESTDIR)/__TEST__
 TEST_EXP+=		$(foreach K,$(TEST_KINDS),$(TEST_DESTDIR)/__TEST__$(K))
 TEST_PROGS.cdeque=		$(subst __TEST__,test-cdeque,$(TEST_EXP))
@@ -36,6 +36,7 @@ $(TEST_DESTDIR)/test-%-dumbc11: WSTREAM_CFLAGS+=-DUSE_SEQ_CST_STDATOMIC=1
 $(TEST_DESTDIR)/test-%-llscopt: WSTREAM_CFLAGS+=-DLLSC_OPTIMIZATION=1
 $(TEST_DESTDIR)/test-%-nolwcas: WSTREAM_CFLAGS+=-DNO_LIGHTWEIGHT_CAS=1
 $(TEST_DESTDIR)/test-%-nofences: WSTREAM_CFLAGS+=-DNO_FENCES=1
+$(TEST_DESTDIR)/test-%-nosync: WSTREAM_CFLAGS+=-DNO_FENCES=1 -DNO_SYNC=1
 
 $(TEST_DESTDIR)/test-%: error.c $(WSTREAM_DEPS) $(TEST_DEPS)
 	mkdir -p $(TEST_DESTDIR)

libworkstream_df/benchmark-cdeque.sh

@@ -36,6 +36,17 @@ BEGIN {
 }
 ' >rand.$$
 
+t=test-cdeque-nosync
+echo $t prerun >&2
+for i in $(seq 1 10); do
+	[ $dry ] || ./$t $testargs -f 0 | tail -n 1 >/dev/null
+done
+log=$t.$nthread.$b.$d.log
+for i in $(seq 1 30); do
+	echo $t $testargs -f 0 >&2 \>\> $log
+	[ $dry ] || ./$t $testargs -f 0 | tail -n 1 >$log
+done
+
 for t in $tests; do
 	echo $t prerun >&2
 	for i in $(seq 1 10); do

libworkstream_df/plot-benchmark-cdeque.sh

@@ -70,7 +70,18 @@ if [ -n "$(find . -maxdepth 1 -name "test-*.$n.$b.$d.log")" ]; then
 		sort -n >tput_eff_tput.$t.$n.$b.$d.data
 	done
 	plot tput_eff_tput.test-cdeque points	\
+	    'Critical path throughput'		\
+	    'effective steal throughput (s⁻¹)'	\
+	    'push/take throughput (s⁻¹)'
+	norm=$(awk '{x+=2*$(NF-3)/$1;++n} END{print x/n}'	\
+	    test-cdeque-nosync.$n.$b.$d.log)
+	for t in $tests; do
+		awk -v norm=$norm '{print $NF/$1, 2*$(NF-3)/$1/norm}'	\
+		    $t.$n.$b.$d.log |
+		sort -n >norm_tput_eff_tput.$t.$n.$b.$d.data
+	done
+	plot norm_tput_eff_tput.test-cdeque points	\
 	    'Critical path throughput'			\
 	    'effective steal throughput (s⁻¹)'		\
-	    'push/take throughput (s⁻¹)'
+	    'normalized push/take throughput'
 fi

libworkstream_df/wstream_df.h

@@ -107,6 +107,7 @@ store_conditional (volatile size_t *ptr, size_t value)
 static inline bool
 compare_and_swap (volatile size_t *ptr, size_t oldval, size_t newval)
 {
+#if !NO_FENCES || !NO_SYNC
 #if !NO_LIGHTWEIGHT_CAS && defined(__arm__)
   int status = 1;
   __asm__ __volatile__ ("0: ldrex r0, [%1]\n\t"
@@ -123,11 +124,21 @@ compare_and_swap (volatile size_t *ptr, size_t oldval, size_t newval)
 #else
   return __sync_bool_compare_and_swap (ptr, oldval, newval);
 #endif
+#else
+  if (*ptr == oldval)
+    {
+      *ptr = newval;
+      return true;
+    }
+  else
+    return false;
+#endif
 }
 
 static inline bool
 weak_compare_and_swap (volatile size_t *ptr, size_t oldval, size_t newval)
 {
+#if !NO_FENCES || !NO_SYNC
 #if !NO_LIGHTWEIGHT_CAS && defined(__arm__)
   int status = 1;
   __asm__ __volatile__ ("ldrex r0, [%1]\n\t"
@@ -140,6 +151,9 @@ weak_compare_and_swap (volatile size_t *ptr, size_t oldval, size_t newval)
 #else
   return __sync_bool_compare_and_swap (ptr, oldval, newval);
 #endif
+#else
+  return compare_and_swap (ptr, oldval, newval);
+#endif
 }