Add threaded -O3 build mode

rtl/Makefile

@@ -21,13 +21,19 @@ LIB=
 
 CF=-CFLAGS '-std=c++11 -fms-extensions'
 
+CFRel=-CFLAGS '-std=c++11 -fms-extensions -O3 -DVL_THREADED'
+
 DEB=--trace --prof-cfuncs -DVL_DEBUG=1 
 
 
 MAKECPP=(cd obj_dir && make -j -f VVortex.mk OPT='-DVL_DEBUG' VL_DEBUG=1 DVL_DEBUG=1)
 
+MAKECPPRel=(cd obj_dir && make -j -f VVortex.mk)
+
 MAKEMULTICPP=(cd obj_dir && make -j -f VVortex_SOC.mk OPT='-DVL_DEBUG' VL_DEBUG=1 DVL_DEBUG=1)
 
+THREADS ?= $(shell python3 -c 'import multiprocessing as mp; print(max(1, mp.cpu_count() // 2))' )
+
 # -LDFLAGS '-lsystemc'
 VERILATOR:
 	echo "#define VCD_OFF" > simulate/tb_debug.h
@@ -37,6 +43,11 @@ VERILATORnoWarnings:
 	echo "#define VCD_OFF" > simulate/tb_debug.h
 	verilator $(COMP) -cc $(SINGLE_CORE) $(INCLUDE) $(EXE) $(LIB) $(CF)  $(WNO) $(DEB)
 
+VERILATORnoWarningsRel:
+	echo "#define VCD_OFF" > simulate/tb_debug.h
+	verilator $(COMP) -cc $(SINGLE_CORE) $(INCLUDE) $(EXE) $(LIB) $(CFRel)  $(WNO) --threads $(THREADS)
+
+
 VERILATORMULTInoWarnings:
 	echo "#define VCD_OFF" > simulate/tb_debug.h
 	verilator $(COMP) -cc $(MULTI_CORE) $(INCLUDE) $(MULTI_EXE) $(LIB) $(CF)  $(WNO) $(DEB)
@@ -58,6 +69,9 @@ debug: compdebug
 w: VERILATORnoWarnings
 	$(MAKECPP)
 
+wRel: VERILATORnoWarningsRel
+	$(MAKECPPRel)
+
 multicore: VERILATORMULTInoWarnings
 	$(MAKEMULTICPP)
 
@@ -67,5 +81,8 @@ dmulticore: compdebugmulti
 run: w
 	(cd obj_dir && ./VVortex)
 
+runRel: wRel
+	(cd obj_dir && ./VVortex)
+
 clean:
 	rm obj_dir/*