Changing the make flow so all the generated files, both object and code generated go in a directory independent of the source directory

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 16503]

12 files changed, 151 insertions, 156 deletions

diff --git a/CHANGES b/CHANGES
index 0f41234..cbc929b 100644
--- a/CHANGES
+++ b/CHANGES
@@ -23,6 +23,8 @@ Version 3.2.1+edits (development branch) versus 3.2.1
   simulation liveness messages printed by the timing model (defaults to 
   1 message per second in wall clock time). 
 - Added breakdown for interconnection traffic based on memory access type. 
+- Changes to the makefiles s.t. all intermediate files get output to the build
+  directory, and nothing is written to the same directory as the source code
 - Bug Fixes:
     - Fixed the flit count sent to GPUWattch for atomic operations. 
     - Fix for Bug 51 - Updated the function declaration of
diff --git a/Makefile b/Makefile
index 7a9c9bd..9bf8f83 100644
--- a/Makefile
+++ b/Makefile
@@ -31,22 +31,23 @@
 
 include version_detection.mk
 
-ifeq ($(GPGPUSIM_CONFIG), $(CUDART_VERSION)/debug)
+ifeq ($(GPGPUSIM_CONFIG), gcc-$(CC_VERSION)/cuda-$(CUDART_VERSION)/debug)
 	export DEBUG=1
 else
 	export DEBUG=0
 endif
 
+BUILD_ROOT?=$(shell pwd)
 export TRACE?=1
 
 NVCC_PATH=$(shellwhich nvcc)
 ifneq ($(shell which nvcc), "")
 	ifeq ($(DEBUG), 1)
-		export SIM_LIB_DIR=lib/$(CUDART_VERSION)/debug
-		export SIM_OBJ_FILES_DIR=build/$(CUDART_VERSION)/debug
+		export SIM_LIB_DIR=lib/gcc-$(CC_VERSION)/cuda-$(CUDART_VERSION)/debug
+		export SIM_OBJ_FILES_DIR=$(BUILD_ROOT)/build/gcc-$(CC_VERSION)/cuda-$(CUDART_VERSION)/debug
 	else
-		export SIM_LIB_DIR=lib/$(CUDART_VERSION)/release
-		export SIM_OBJ_FILES_DIR=build/$(CUDART_VERSION)/release
+		export SIM_LIB_DIR=lib/gcc-$(CC_VERSION)/cuda-$(CUDART_VERSION)/release
+		export SIM_OBJ_FILES_DIR=$(BUILD_ROOT)/build/gcc-$(CC_VERSION)/cuda-$(CUDART_VERSION)/release
 	endif
 endif
 
@@ -77,12 +78,11 @@ ifneq ($(GPGPUSIM_POWER_MODEL),)
 
 
 	ifeq ($(DEBUG), 1)
-		MCPAT_OBJ_DIR = $(GPGPUSIM_POWER_MODEL)/obj_dbg
 		MCPAT_DBG_FLAG = dbg
-	else
-		MCPAT_OBJ_DIR = $(GPGPUSIM_POWER_MODEL)/obj_opt
 	endif
 
+	MCPAT_OBJ_DIR = $(SIM_OBJ_FILES_DIR)/gpuwattch
+
 	MCPAT = $(MCPAT_OBJ_DIR)/*.o
 endif
 
@@ -136,7 +136,7 @@ check_power:
 no_opencl_support:
 	@echo "Warning: gpgpu-sim is building without opencl support. Make sure NVOPENCL_LIBDIR and NVOPENCL_INCDIR are set"
 
-$(SIM_LIB_DIR)/libcudart.so: $(LIBS) cudalib
+$(SIM_LIB_DIR)/libcudart.so: makedirs $(LIBS) cudalib
 	g++ -shared -Wl,-soname,libcudart.so \
 			$(SIM_OBJ_FILES_DIR)/libcuda/*.o \
 			$(SIM_OBJ_FILES_DIR)/cuda-sim/*.o \
@@ -150,7 +150,7 @@ $(SIM_LIB_DIR)/libcudart.so: $(LIBS) cudalib
 	if [ ! -f $(SIM_LIB_DIR)/libcudart.so.3 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.3; fi
 	if [ ! -f $(SIM_LIB_DIR)/libcudart.so.4 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.4; fi
 
-$(SIM_LIB_DIR)/libcudart.dylib: $(LIBS) cudalib
+$(SIM_LIB_DIR)/libcudart.dylib: makedirs $(LIBS) cudalib
 	g++ -dynamiclib -Wl,-headerpad_max_install_names,-undefined,dynamic_lookup,-compatibility_version,1.1,-current_version,1.1\
 			$(SIM_OBJ_FILES_DIR)/libcuda/*.o \
 			$(SIM_OBJ_FILES_DIR)/cuda-sim/*.o \
@@ -161,7 +161,7 @@ $(SIM_LIB_DIR)/libcudart.dylib: $(LIBS) cudalib
 			$(MCPAT) \
 			-o $(SIM_LIB_DIR)/libcudart.dylib
 
-$(SIM_LIB_DIR)/libOpenCL.so: $(LIBS) opencllib
+$(SIM_LIB_DIR)/libOpenCL.so: makedirs $(LIBS) opencllib
 	g++ -shared -Wl,-soname,libOpenCL.so \
 			$(SIM_OBJ_FILES_DIR)/libopencl/*.o \
 			$(SIM_OBJ_FILES_DIR)/cuda-sim/*.o \
@@ -174,37 +174,37 @@ $(SIM_LIB_DIR)/libOpenCL.so: $(LIBS) opencllib
 	if [ ! -f $(SIM_LIB_DIR)/libOpenCL.so.1 ]; then ln -s libOpenCL.so $(SIM_LIB_DIR)/libOpenCL.so.1; fi
 	if [ ! -f $(SIM_LIB_DIR)/libOpenCL.so.1.1 ]; then ln -s libOpenCL.so $(SIM_LIB_DIR)/libOpenCL.so.1.1; fi
 
-cudalib: cuda-sim
+cudalib: makedirs cuda-sim
 	$(MAKE) -C ./libcuda/ depend
 	$(MAKE) -C ./libcuda/
 
 ifneq ($(GPGPUSIM_POWER_MODEL),)
-mcpat:
+mcpat: makedirs
 	$(MAKE) -C $(GPGPUSIM_POWER_MODEL) depend
 	$(MAKE) -C $(GPGPUSIM_POWER_MODEL) $(MCPAT_DBG_FLAG)
 endif
 
-cuda-sim:
+cuda-sim: makedirs
 	$(MAKE) -C ./src/cuda-sim/ depend
 	$(MAKE) -C ./src/cuda-sim/
 
-gpgpu-sim_uarch: cuda-sim
+gpgpu-sim_uarch: makedirs cuda-sim
 	$(MAKE) -C ./src/gpgpu-sim/ depend
 	$(MAKE) -C ./src/gpgpu-sim/
 
-intersim: cuda-sim gpgpu-sim_uarch
+intersim: makedirs cuda-sim gpgpu-sim_uarch
 	$(MAKE) "CREATELIBRARY=1" "DEBUG=$(DEBUG)" -C ./src/intersim	
 
-gpgpusimlib: cuda-sim gpgpu-sim_uarch intersim
+gpgpusimlib: makedirs cuda-sim gpgpu-sim_uarch intersim
 	$(MAKE) -C ./src/ depend
 	$(MAKE) -C ./src/
 
-opencllib: cuda-sim
+opencllib: makedirs cuda-sim
 	$(MAKE) -C ./libopencl/ depend
 	$(MAKE) -C ./libopencl/
 
 .PHONY: cuobjdump_to_ptxplus/cuobjdump_to_ptxplus
-cuobjdump_to_ptxplus/cuobjdump_to_ptxplus:
+cuobjdump_to_ptxplus/cuobjdump_to_ptxplus: makedirs
 	$(MAKE) -C ./cuobjdump_to_ptxplus/ depend
 	$(MAKE) -C ./cuobjdump_to_ptxplus/
 
@@ -215,9 +215,11 @@ makedirs:
 	if [ ! -d $(SIM_OBJ_FILES_DIR)/cuda-sim/decuda_pred_table ]; then mkdir -p $(SIM_OBJ_FILES_DIR)/cuda-sim/decuda_pred_table; fi;
 	if [ ! -d $(SIM_OBJ_FILES_DIR)/gpgpu-sim ]; then mkdir -p $(SIM_OBJ_FILES_DIR)/gpgpu-sim; fi;
 	if [ ! -d $(SIM_OBJ_FILES_DIR)/libopencl ]; then mkdir -p $(SIM_OBJ_FILES_DIR)/libopencl; fi;
+	if [ ! -d $(SIM_OBJ_FILES_DIR)/libopencl/bin ]; then mkdir -p $(SIM_OBJ_FILES_DIR)/libopencl/bin; fi;
 	if [ ! -d $(SIM_OBJ_FILES_DIR)/intersim ]; then mkdir -p $(SIM_OBJ_FILES_DIR)/intersim; fi;
 	if [ ! -d $(SIM_OBJ_FILES_DIR)/cuobjdump_to_ptxplus ]; then mkdir -p $(SIM_OBJ_FILES_DIR)/cuobjdump_to_ptxplus; fi;
-	if [ ! -d $(GPGPUSIM_POWER_MODEL)/obj_opt ] && [ ! -n $(GPGPUSIM_POWER_MODEL) ]; then mkdir -p $(GPGPUSIM_POWER_MODEL)/obj_opt; fi;
+	if [ ! -d $(SIM_OBJ_FILES_DIR)/gpuwattch ]; then mkdir -p $(SIM_OBJ_FILES_DIR)/gpuwattch; fi;
+	if [ ! -d $(SIM_OBJ_FILES_DIR)/gpuwattch/cacti ]; then mkdir -p $(SIM_OBJ_FILES_DIR)/gpuwattch/cacti; fi;
 
 all:
 	$(MAKE) gpgpusim
@@ -228,21 +230,9 @@ docs:
 cleandocs:
 	$(MAKE) clean -C doc/doxygen/
 
-clean: 
+clean: makedirs
 	$(MAKE) cleangpgpusim
 
 cleangpgpusim: cleandocs
-	$(MAKE) clean -C ./libcuda/
-ifneq  ($(NVOPENCL_LIBDIR),)
-	$(MAKE) clean -C ./libopencl/
-endif
-	$(MAKE) clean -C ./src/intersim/
-	$(MAKE) clean -C ./src/cuda-sim/
-	$(MAKE) clean -C ./src/gpgpu-sim/
-	$(MAKE) clean -C ./src/
-	$(MAKE) clean -C ./cuobjdump_to_ptxplus/
-ifneq ($(GPGPUSIM_POWER_MODEL),)
-	$(MAKE) clean -C $(GPGPUSIM_POWER_MODEL)
-endif
 	rm -rf $(SIM_LIB_DIR)
 	rm -rf $(SIM_OBJ_FILES_DIR)
diff --git a/cuobjdump_to_ptxplus/Makefile b/cuobjdump_to_ptxplus/Makefile
index 14fcc68..40b8b48 100644
--- a/cuobjdump_to_ptxplus/Makefile
+++ b/cuobjdump_to_ptxplus/Makefile
@@ -7,7 +7,7 @@ CXX			= g++
 CXXFLAGS	= ${CCFLAGS}
 LD			= g++
 LDFLAGS		= ${CCFLAGS}
-OUTPUT_DIR=../$(SIM_OBJ_FILES_DIR)/cuobjdump_to_ptxplus
+OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/cuobjdump_to_ptxplus
 
 CXXFLAGS += -I $(OUTPUT_DIR) -I . -I ../src/cuda-sim/
 
@@ -64,11 +64,11 @@ $(OUTPUT_DIR)/%.o: %.c %.h
 
 SRCS = $(shell ls *.cc)
 
-Makefile.makedepend: depend
+$(OUTPUT_DIR)/Makefile.makedepend: depend
 
 depend:
-	touch Makefile.makedepend
-	makedepend -fMakefile.makedepend $(SRCS) 2> /dev/null
+	touch $(OUTPUT_DIR)/Makefile.makedepend
+	makedepend -f$(OUTPUT_DIR)/Makefile.makedepend $(SRCS) 2> /dev/null
 
 clean:
 	rm -f lex.ptx_.c ptx.tab.c ptx.tab.h ptx.output
@@ -76,7 +76,7 @@ clean:
 	rm -f sass_lexer.cc  sass_parser.cc  sass_parser.hh  sass_parser.output
 	rm -f header_lexer.cc  header_parser.cc  header_parser.hh  header_parser.output
 	rm -rf $(OUTPUT_DIR)
-	rm -f Makefile.makedepend Makefile.makedepend.bak
+	rm -f $(OUTPUT_DIR)/Makefile.makedepend $(OUTPUT_DIR)/Makefile.makedepend.bak
 
-include Makefile.makedepend
+include $(OUTPUT_DIR)/Makefile.makedepend
 
diff --git a/libcuda/Makefile b/libcuda/Makefile
index 399a533..b1a6552 100644
--- a/libcuda/Makefile
+++ b/libcuda/Makefile
@@ -96,7 +96,7 @@ YFLAGS		= -t -d -v --report=all
 
 .PHONY: clean
 
-OUTPUT_DIR=../$(SIM_OBJ_FILES_DIR)/libcuda
+OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/libcuda
 
 OBJS = $(CXX_SRCS:%.cc=$(OUTPUT_DIR)/%.o)
 OBJS += $(LEX_SRCS:%.l=$(OUTPUT_DIR)/%_lexer.o)
@@ -123,11 +123,11 @@ cuobjdump_lexer.c: cuobjdump.l cuobjdump_parser.h
 # in other words, get parser.c and do nothing to get parser.h
 %_parser.h: %_parser.c
 	:
-Makefile.makedepend: depend
+$(OUTPUT_DIR)/Makefile.makedepend: depend
 
 depend:
-	touch Makefile.makedepend
-	makedepend -fMakefile.makedepend -p$(OUTPUT_DIR)/ $(CXX_SRCS) 2> /dev/null
+	touch $(OUTPUT_DIR)/Makefile.makedepend
+	makedepend -f$(OUTPUT_DIR)/Makefile.makedepend -p$(OUTPUT_DIR)/ $(CXX_SRCS) 2> /dev/null
 
 clean:
 	rm -f $(PROG)
@@ -137,5 +137,5 @@ clean:
 	rm -f *_lexer.*
 	rm -f Makefile.makedepend Makefile.makedepend.bak
 
-include Makefile.makedepend
+include $(OUTPUT_DIR)/Makefile.makedepend
 
diff --git a/libopencl/Makefile b/libopencl/Makefile
index d597a0d..f849225 100644
--- a/libopencl/Makefile
+++ b/libopencl/Makefile
@@ -87,28 +87,27 @@ CCFLAGS += -DCUDART_VERSION=$(CUDART_VERSION)
 
 .PHONY: clean
 
-OUTPUT_DIR=../$(SIM_OBJ_FILES_DIR)/libopencl
+OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/libopencl
 
 OBJS = $(CXX_SRCS:%.cc=$(OUTPUT_DIR)/%.o)
 
 #--- Make rules ---
 
-all: lib$(PROG).a nvopencl_wrapper
+all: lib$(PROG).a $(OUTPUT_DIR)/nvopencl_wrapper
 
 lib$(PROG).a: $(OBJS)
 	ar rcs $(OUTPUT_DIR)/lib$(PROG).a $(OBJS)
 
-Makefile.makedepend: depend
+$(OUTPUT_DIR)/Makefile.makedepend: depend
 
 depend:
-	touch Makefile.makedepend
-	makedepend -fMakefile.makedepend -p$(OUTPUT_DIR)/ $(CXX_SRCS) 2> /dev/null
+	touch $(OUTPUT_DIR)/Makefile.makedepend
+	makedepend -f$(OUTPUT_DIR)/Makefile.makedepend -p$(OUTPUT_DIR)/ $(CXX_SRCS) 2> /dev/null
 
-nvopencl_wrapper: nvopencl_wrapper.cc
+$(OUTPUT_DIR)/nvopencl_wrapper: nvopencl_wrapper.cc
 	if [ ! -d bin ]; then mkdir bin; fi
-	g++ $(CCFLAGS) nvopencl_wrapper.cc -I./ -I$(NVOPENCL_INCDIR)/ -L $(NVOPENCL_LIBDIR) -lOpenCL  -o bin/nvopencl_wrapper
+	g++ $(CCFLAGS) nvopencl_wrapper.cc -I./ -I$(NVOPENCL_INCDIR)/ -L $(NVOPENCL_LIBDIR) -lOpenCL  -o $(OUTPUT_DIR)/bin/nvopencl_wrapper
 	
-OUTPUT_DIR=../$(SIM_OBJ_FILES_DIR)/libopencl
 
 $(OUTPUT_DIR)/%.o: %.cc
 	$(CPP) $(CCFLAGS) -I./ -I$(CUDA_INSTALL_PATH)/include  -c $< -o $@
@@ -118,7 +117,7 @@ clean:
 	rm -f *.o
 	rm -f lib$(PROG).a
 	rm -f bin/nvopencl_wrapper
-	rm -f Makefile.makedepend Makefile.makedepend.bak
+	rm -f $(OUTPUT_DIR)/Makefile.makedepend $(OUTPUT_DIR)/Makefile.makedepend.bak
 
-include Makefile.makedepend
+include $(OUTPUT_DIR)/Makefile.makedepend
 
diff --git a/setup_environment b/setup_environment
index cb39553..b4bcd1b 100644
--- a/setup_environment
+++ b/setup_environment
@@ -1,6 +1,6 @@
 # see README before running this
 
-ps -p $$ | awk '/bash/ {exit 1;}' && echo "ERROR ** source setup_environment must be run in a bash shell; see README" && exit
+ps -p $$ | awk '/bash/ || / sh/ {exit 1;}' && echo "ERROR ** source setup_environment must be run in a bash or sh shell; see README" && exit
 
 export GPGPUSIM_SETUP_ENVIRONMENT_WAS_RUN=
 export GPGPUSIM_ROOT="$( cd "$( dirname "$BASH_SOURCE" )" && pwd )"
@@ -39,6 +39,8 @@ if [ $? = 1 ]; then
 	return;
 fi
 
+CC_VERSION=`gcc --version | head -1 | sed -re "s/gcc\s+\(.*\)\s+([0-9.]+)/\1/"`
+
 CUDA_VERSION_STRING=`$CUDA_INSTALL_PATH/bin/nvcc --version | awk '/release/ {print $5;}' | sed 's/,//'`;
 CUDA_VERSION_NUMBER=`echo $CUDA_VERSION_STRING | sed 's/\./ /' | awk '{printf("%02u%02u", 10*int($1), 10*$2);}'`
 if [ $CUDA_VERSION_NUMBER -gt 4020 -o $CUDA_VERSION_NUMBER -lt 2030  ]; then
@@ -48,9 +50,9 @@ fi
 
 if [ $# == '1' ] ;
 then
-    export GPGPUSIM_CONFIG=$CUDA_VERSION_NUMBER/$1
+    export GPGPUSIM_CONFIG=gcc-$CC_VERSION/cuda-$CUDA_VERSION_NUMBER/$1
 else
-    export GPGPUSIM_CONFIG=$CUDA_VERSION_NUMBER/release
+    export GPGPUSIM_CONFIG=gcc-$CC_VERSION/cuda-$CUDA_VERSION_NUMBER/release
 fi
 
 export QTINC=/usr/include
diff --git a/src/Makefile b/src/Makefile
index cc31764..6001669 100644
--- a/src/Makefile
+++ b/src/Makefile
@@ -56,7 +56,7 @@ OPTFLAGS += -g3 -fPIC
 CPP = g++ $(SNOW)
 OEXT = o
 
-OUTPUT_DIR=../$(SIM_OBJ_FILES_DIR)
+OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)
 SRCS = $(shell ls *.cc)
 OBJS = $(SRCS:%.cc=$(OUTPUT_DIR)/%.$(OEXT))
 
@@ -66,19 +66,19 @@ $(OUTPUT_DIR)/libgpgpusim.a:	$(OBJS) gpu_uarch_simlib
 gpu_uarch_simlib:
 	make   -C ./gpgpu-sim
 	
-Makefile.makedepend: depend
+$(OUTPUT_DIR)/Makefile.makedepend: depend
 
 depend:
-	touch Makefile.makedepend
-	makedepend -fMakefile.makedepend -p$(OUTPUT_DIR)/ $(SRCS) 2> /dev/null
-
-$(OUTPUT_DIR)/%.$(OEXT): %.cc
-	$(CPP) $(OPTFLAGS) $(CXXFLAGS) -o $(OUTPUT_DIR)/$*.$(OEXT) -c $*.cc
+	touch $(OUTPUT_DIR)/Makefile.makedepend
+	makedepend -f$(OUTPUT_DIR)/Makefile.makedepend -p$(OUTPUT_DIR)/ $(SRCS) 2> /dev/null
 
 clean:
 	rm -f *.o core *~ *.a Makefile.makedepend Makefile.makedepend.bak
 
+$(OUTPUT_DIR)/%.$(OEXT): %.cc
+	$(CPP) $(OPTFLAGS) $(CXXFLAGS) -o $(OUTPUT_DIR)/$*.$(OEXT) -c $*.cc
+
 option_parser.$(OEXT): option_parser.h
 
-include Makefile.makedepend
+include $(OUTPUT_DIR)/Makefile.makedepend
 
diff --git a/src/cuda-sim/Makefile b/src/cuda-sim/Makefile
index 51a9d8e..166e256 100644
--- a/src/cuda-sim/Makefile
+++ b/src/cuda-sim/Makefile
@@ -40,11 +40,13 @@ endif
 
 include ../../version_detection.mk
 
+OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/cuda-sim
+
 OPT	:=  -O3 -g3 -Wall -Wno-unused-function -Wno-sign-compare
 ifeq ($(DEBUG),1)
 	OPT := -g3 -Wall  -Wno-unused-function -Wno-sign-compare
 endif
-OPT += -I$(CUDA_INSTALL_PATH)/include 
+OPT += -I$(CUDA_INSTALL_PATH)/include  -I$(OUTPUT_DIR)/ -I.
 OPT += -fPIC 
 
 ifeq ($(TRACE),1)
@@ -60,8 +62,6 @@ ifeq ($(GNUC_CPP0X),1)
 endif
 endif
 
-OUTPUT_DIR=../../$(SIM_OBJ_FILES_DIR)/cuda-sim
-
 OBJS	:= $(OUTPUT_DIR)/ptx_parser.o $(OUTPUT_DIR)/ptx_loader.o $(OUTPUT_DIR)/cuda_device_printf.o $(OUTPUT_DIR)/instructions.o $(OUTPUT_DIR)/cuda-sim.o $(OUTPUT_DIR)/ptx_ir.o $(OUTPUT_DIR)/ptx_sim.o  $(OUTPUT_DIR)/memory.o $(OUTPUT_DIR)/ptx-stats.o $(OUTPUT_DIR)/decuda_pred_table/decuda_pred_table.o $(OUTPUT_DIR)/ptx.tab.o $(OUTPUT_DIR)/lex.ptx_.o $(OUTPUT_DIR)/ptxinfo.tab.o $(OUTPUT_DIR)/lex.ptxinfo_.o
 
 
@@ -69,53 +69,53 @@ OPT += -DCUDART_VERSION=$(CUDART_VERSION)
 
 SRCS = $(shell ls *.cc)
 
-Makefile.makedepend: depend
+$(OUTPUT_DIR)/Makefile.makedepend: depend
 
 depend:
-	touch Makefile.makedepend
-	makedepend -fMakefile.makedepend -p$(OUTPUT_DIR)/ $(SRCS) 2> /dev/null
+	touch $(OUTPUT_DIR)/Makefile.makedepend
+	makedepend -f$(OUTPUT_DIR)/Makefile.makedepend -p$(OUTPUT_DIR)/ $(SRCS) 2> /dev/null
 
 libgpgpu_ptx_sim.a: $(OBJS) 
 	ar rcs $(OUTPUT_DIR)/libgpgpu_ptx_sim.a $(OUTPUT_DIR)/ptx.tab.o $(OUTPUT_DIR)/lex.ptx_.o $(OUTPUT_DIR)/ptxinfo.tab.o $(OUTPUT_DIR)/lex.ptxinfo_.o $(OBJS)
 
-$(OUTPUT_DIR)/ptx.tab.o: ptx.tab.c
-	$(CPP) -c $(OPT) -DYYDEBUG ptx.tab.c -o $(OUTPUT_DIR)/ptx.tab.o
+$(OUTPUT_DIR)/ptx.tab.o: $(OUTPUT_DIR)/ptx.tab.c
+	$(CPP) -c $(OPT) -DYYDEBUG $(OUTPUT_DIR)/ptx.tab.c -o $(OUTPUT_DIR)/ptx.tab.o
 
-$(OUTPUT_DIR)/lex.ptx_.o: lex.ptx_.c 
-	$(CPP) -c $(OPT) lex.ptx_.c -o $(OUTPUT_DIR)/lex.ptx_.o
+$(OUTPUT_DIR)/lex.ptx_.o: $(OUTPUT_DIR)/lex.ptx_.c 
+	$(CPP) -c $(OPT) $(OUTPUT_DIR)/lex.ptx_.c -o $(OUTPUT_DIR)/lex.ptx_.o
 
-$(OUTPUT_DIR)/ptxinfo.tab.o: ptxinfo.tab.c
-	$(CPP) -c $(OPT) -DYYDEBUG ptxinfo.tab.c -o $(OUTPUT_DIR)/ptxinfo.tab.o
+$(OUTPUT_DIR)/ptxinfo.tab.o: $(OUTPUT_DIR)/ptxinfo.tab.c
+	$(CPP) -c $(OPT) -DYYDEBUG $(OUTPUT_DIR)/ptxinfo.tab.c -o $(OUTPUT_DIR)/ptxinfo.tab.o
 
-$(OUTPUT_DIR)/lex.ptxinfo_.o: lex.ptxinfo_.c ptxinfo.tab.c
-	$(CPP) -c $(OPT) lex.ptxinfo_.c -o $(OUTPUT_DIR)/lex.ptxinfo_.o
+$(OUTPUT_DIR)/lex.ptxinfo_.o: $(OUTPUT_DIR)/lex.ptxinfo_.c $(OUTPUT_DIR)/ptxinfo.tab.c
+	$(CPP) -c $(OPT) $(OUTPUT_DIR)/lex.ptxinfo_.c -o $(OUTPUT_DIR)/lex.ptxinfo_.o
 
-ptx.tab.c: ptx.y
-	bison --name-prefix=ptx_ -v -d ptx.y
+$(OUTPUT_DIR)/ptx.tab.c: ptx.y
+	bison --name-prefix=ptx_ -v -d ptx.y --file-prefix=$(OUTPUT_DIR)/ptx
 
-ptxinfo.tab.c: ptxinfo.y
-	bison --name-prefix=ptxinfo_ -v -d ptxinfo.y
+$(OUTPUT_DIR)/ptxinfo.tab.c: ptxinfo.y
+	bison --name-prefix=ptxinfo_ -v -d ptxinfo.y --file-prefix=$(OUTPUT_DIR)/ptxinfo
 
-lex.ptx_.c: ptx.l
-	flex ptx.l
+$(OUTPUT_DIR)/lex.ptx_.c: ptx.l
+	flex --outfile=$(OUTPUT_DIR)/lex.ptx_.c ptx.l 
 
-lex.ptxinfo_.c: ptxinfo.l
-	flex ptxinfo.l
+$(OUTPUT_DIR)/lex.ptxinfo_.c: ptxinfo.l
+	flex --outfile=$(OUTPUT_DIR)/lex.ptxinfo_.c ptxinfo.l 
 
 clean:
 	rm -f *~ *.o *.gcda *.gcno *.gcov libgpgpu_ptx_sim.a \
 		ptx.tab.h ptx.tab.c ptx.output lex.ptx_.c \
 		ptxinfo.tab.h ptxinfo.tab.c ptxinfo.output lex.ptxinfo_.c \
 		instructions.h ptx_parser_decode.def directed_tests.log 
-	rm -f decuda_pred_table/*.o
-	rm -f Makefile.makedepend Makefile.makedepend.bak
+	rm -f $(OUTPUT_DIR)/decuda_pred_table/*.o
+	rm -f $(OUTPUT_DIR)/Makefile.makedepend $(OUTPUT_DIR)/Makefile.makedepend.bak
 
 $(OUTPUT_DIR)/%.o: %.c
 	$(CPP) -c $(OPT) $< -o $(OUTPUT_DIR)/$*.o
 $(OUTPUT_DIR)/%.o: %.cc
 	$(CPP) -c $(CXX_OPT) $< -o $(OUTPUT_DIR)/$*.o
 
-instructions.h: instructions.cc
+$(OUTPUT_DIR)/instructions.h: instructions.cc
 	@touch $*.h
 	@chmod +w $*.h
 	@echo "// DO NOT EDIT THIS FILE! IT IS AUTOMATICALLY GENERATED BY THE MAKEFILE (see target for instructions.h)" > $*.h
@@ -125,25 +125,25 @@ instructions.h: instructions.cc
 	@cat $< | grep "_impl(" | sed 's/{.*//' | sed 's/$$/;/' >> $*.h
 	@echo "#endif" >> $*.h
 	@chmod -w $*.h
-	@echo "created instructions.h"
+	@echo "created $(OUTPUT_DIR)/instructions.h"
 
-ptx_parser_decode.def: ptx.tab.c
+$(OUTPUT_DIR)/ptx_parser_decode.def: $(OUTPUT_DIR)/ptx.tab.c
 ifeq ($(shell uname),Linux)
-	cat ptx.tab.h | grep "=" | sed 's/^[ ]\+//' | sed 's/[=,]//g' | sed 's/\([_A-Z1-9]\+\)[ ]\+\([0-9]\+\)/\1 \1/' | sed 's/^/DEF(/' | sed 's/ /,"/' | sed 's/$$/")/' > ptx_parser_decode.def
+	cat $(OUTPUT_DIR)/ptx.tab.h | grep "=" | sed 's/^[ ]\+//' | sed 's/[=,]//g' | sed 's/\([_A-Z1-9]\+\)[ ]\+\([0-9]\+\)/\1 \1/' | sed 's/^/DEF(/' | sed 's/ /,"/' | sed 's/$$/")/' > $(OUTPUT_DIR)/ptx_parser_decode.def
 else
-	cat ptx.tab.h | grep "=" | sed -E 's/^ +//' | sed 's/[=,]//g' | sed -E 's/([_A-Z1-9]+).*/\1 \1/' | sed 's/^/DEF(/' | sed 's/ /,"/' | sed 's/$$/")/' > ptx_parser_decode.def
+	cat $(OUTPUT_DIR)/ptx.tab.h | grep "=" | sed -E 's/^ +//' | sed 's/[=,]//g' | sed -E 's/([_A-Z1-9]+).*/\1 \1/' | sed 's/^/DEF(/' | sed 's/ /,"/' | sed 's/$$/")/' > $(OUTPUT_DIR)/ptx_parser_decode.def
 endif
 
-$(OUTPUT_DIR)/instructions.o: instructions.h ptx.tab.c
-$(OUTPUT_DIR)/cuda_device_printf.o: ptx.tab.c
-$(OUTPUT_DIR)/ptx_ir.o: ptx.tab.c ptx_parser_decode.def
-$(OUTPUT_DIR)/ptx_loader.o: ptx.tab.c ptx_parser_decode.def
-$(OUTPUT_DIR)/ptx_parser.o: ptx.tab.c ptx_parser_decode.def
-$(OUTPUT_DIR)/ptxinfo.tab.o: ptx.tab.c
-$(OUTPUT_DIR)/ptx-stats.o: ptx.tab.c
-$(OUTPUT_DIR)/ptx_sim.o: ptx.tab.c
-$(OUTPUT_DIR)/cuda-sim.o: ptx.tab.c
-$(OUTPUT_DIR)/lex.ptxinfo_.o: ptx.tab.c
-$(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
-
+$(OUTPUT_DIR)/instructions.o: $(OUTPUT_DIR)/instructions.h $(OUTPUT_DIR)/ptx.tab.c
+$(OUTPUT_DIR)/cuda_device_printf.o: $(OUTPUT_DIR)/ptx.tab.c
+$(OUTPUT_DIR)/ptx_ir.o: $(OUTPUT_DIR)/ptx.tab.c $(OUTPUT_DIR)/ptx_parser_decode.def
+$(OUTPUT_DIR)/ptx_loader.o: $(OUTPUT_DIR)/ptx.tab.c $(OUTPUT_DIR)/ptx_parser_decode.def
+$(OUTPUT_DIR)/ptx_parser.o: $(OUTPUT_DIR)/ptx.tab.c $(OUTPUT_DIR)/ptx_parser_decode.def
+$(OUTPUT_DIR)/ptxinfo.tab.o: $(OUTPUT_DIR)/ptx.tab.c
+$(OUTPUT_DIR)/ptx-stats.o: $(OUTPUT_DIR)/ptx.tab.c
+$(OUTPUT_DIR)/ptx_sim.o: $(OUTPUT_DIR)/ptx.tab.c
+$(OUTPUT_DIR)/cuda-sim.o: $(OUTPUT_DIR)/ptx.tab.c
+$(OUTPUT_DIR)/lex.ptxinfo_.o: $(OUTPUT_DIR)/ptx.tab.c
+$(OUTPUT_DIR)/lex.ptx_.o: $(OUTPUT_DIR)/ptx.tab.c
 
+include $(OUTPUT_DIR)/Makefile.makedepend
diff --git a/src/gpgpu-sim/Makefile b/src/gpgpu-sim/Makefile
index b9749e5..e355b59 100644
--- a/src/gpgpu-sim/Makefile
+++ b/src/gpgpu-sim/Makefile
@@ -67,7 +67,7 @@ OPTFLAGS += -g3 -fPIC
 CPP = g++ $(SNOW)
 OEXT = o
 
-OUTPUT_DIR=../../$(SIM_OBJ_FILES_DIR)/gpgpu-sim
+OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/gpgpu-sim
 
 SRCS = $(shell ls *.cc)
 
@@ -84,11 +84,11 @@ OBJS = $(CSRCS:%.cc=$(OUTPUT_DIR)/%.$(OEXT))
 libgpu_uarch_sim.a:$(OBJS)
 	ar rcs  $(OUTPUT_DIR)/libgpu_uarch_sim.a $(OBJS)
 
-Makefile.makedepend: depend
+$(OUTPUT_DIR)/Makefile.makedepend: depend
 
 depend:
-	touch Makefile.makedepend
-	makedepend -fMakefile.makedepend -p$(OUTPUT_DIR)/ $(CSRCS) 2> /dev/null
+	touch $(OUTPUT_DIR)/Makefile.makedepend
+	makedepend -f$(OUTPUT_DIR)/Makefile.makedepend -p$(OUTPUT_DIR)/ $(CSRCS) 2> /dev/null
 
 $(OUTPUT_DIR)/l2cache.$(OEXT): l2cache.cc
 	$(CPP) $(OPTFLAGS) $(CXXFLAGS_L2CACHE) -o $*.$(OEXT) -c l2cache.cc
@@ -102,10 +102,10 @@ clean:
 
 $(OUTPUT_DIR)/option_parser.$(OEXT): option_parser.h
 
-$(OUTPUT_DIR)/dram_sched.$(OEXT): ../cuda-sim/ptx.tab.h
+$(OUTPUT_DIR)/dram_sched.$(OEXT): $(OUTPUT_DIR)/../cuda-sim/ptx.tab.h
 
-../cuda-sim/ptx.tab.h:
-	make -C ../cuda-sim/ ptx.tab.c
+$(OUTPUT_DIR)/../cuda-sim/ptx.tab.h:
+	make -C ../cuda-sim/ $(OUTPUT_DIR)/../cuda-sim/ptx.tab.c
 
-include Makefile.makedepend
+include $(OUTPUT_DIR)/Makefile.makedepend
 
diff --git a/src/gpuwattch/cacti/cacti.mk b/src/gpuwattch/cacti/cacti.mk
index c8c2192..1de6e9e 100644
--- a/src/gpuwattch/cacti/cacti.mk
+++ b/src/gpuwattch/cacti/cacti.mk
@@ -1,3 +1,5 @@
+
+OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/gpuwattch/cacti
 TARGET = cacti
 SHELL = /bin/sh
 .PHONY: all depend clean
@@ -35,30 +37,29 @@ SRCS  = area.cc bank.cc mat.cc main.cc Ucache.cc io.cc technology.cc basic_circu
 		decoder.cc component.cc uca.cc subarray.cc wire.cc htree2.cc \
 		cacti_interface.cc router.cc nuca.cc crossbar.cc arbiter.cc 
 
-OBJS = $(patsubst %.cc,obj_$(TAG)/%.o,$(SRCS))
-PYTHONLIB_SRCS = $(patsubst main.cc, ,$(SRCS)) obj_$(TAG)/cacti_wrap.cc
+OBJS = $(patsubst %.cc,$(OUTPUT_DIR)/%.o,$(SRCS))
+PYTHONLIB_SRCS = $(patsubst main.cc, ,$(SRCS)) $(OUTPUT_DIR)/cacti_wrap.cc
 PYTHONLIB_OBJS = $(patsubst %.cc,%.o,$(PYTHONLIB_SRCS)) 
 INCLUDES       = -I /usr/include/python2.4 -I /usr/lib/python2.4/config
 
-all: obj_$(TAG)/$(TARGET)
-	cp -f obj_$(TAG)/$(TARGET) $(TARGET)
+all: $(OUTPUT_DIR)/$(TARGET)
 
-obj_$(TAG)/$(TARGET) : $(OBJS)
+$(OUTPUT_DIR)/$(TARGET) : $(OBJS)
 	$(CXX) $(OBJS) -o $@ $(INCS) $(CXXFLAGS) $(LIBS) -pthread
 
 #obj_$(TAG)/%.o : %.cc
 #	$(CXX) -c $(CXXFLAGS) $(INCS) -o $@ $<
 
-Makefile.makedepend: depend
+$(OUTPUT_DIR)/Makefile.makedepend: depend
 
 depend:
-	touch Makefile.makedepend
-	makedepend -fMakefile.makedepend -pobj_$(TAG)/ -a $(SRCS) 2> /dev/null
+	touch $(OUTPUT_DIR)/Makefile.makedepend
+	makedepend -f$(OUTPUT_DIR)/Makefile.makedepend -p$(OUTPUT_DIR)/ -a $(SRCS) 2> /dev/null
 
-obj_$(TAG)/%.o : %.cc
+$(OUTPUT_DIR)/%.o : %.cc
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 clean:
 	-rm -f *.o _cacti.so cacti.py $(TARGET)
 
-include Makefile.makedepend
+include $(OUTPUT_DIR)/Makefile.makedepend
diff --git a/src/gpuwattch/mcpat.mk b/src/gpuwattch/mcpat.mk
index 64db479..a09c23b 100644
--- a/src/gpuwattch/mcpat.mk
+++ b/src/gpuwattch/mcpat.mk
@@ -1,3 +1,5 @@
+
+OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/gpuwattch
 TARGET = mcpat
 SHELL = /bin/sh
 .PHONY: all depend clean
@@ -76,29 +78,28 @@ SRCS  = \
 
 
 
-OBJS = $(patsubst %.cc,obj_$(TAG)/%.o,$(SRCS))
+OBJS = $(patsubst %.cc,$(OUTPUT_DIR)/%.o,$(SRCS))
 
-all: obj_$(TAG)/$(TARGET)
-	cp -f obj_$(TAG)/$(TARGET) $(TARGET)
+all: $(OUTPUT_DIR)/$(TARGET)
 
-obj_$(TAG)/$(TARGET) : $(OBJS)
+$(OUTPUT_DIR)/$(TARGET) : $(OBJS)
 	$(CXX) $(OBJS) -o $@ $(INCS) $(CXXFLAGS) $(LIBS) -pthread
 
 #obj_$(TAG)/%.o : %.cc
 #	$(CXX) -c $(CXXFLAGS) $(INCS) -o $@ $<
 
-obj_$(TAG)/%.o : %.cc
+$(OUTPUT_DIR)/%.o : %.cc
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-Makefile.makedepend: depend
+$(OUTPUT_DIR)/Makefile.makedepend: depend
 
 depend:
-	touch Makefile.makedepend
-	makedepend -fMakefile.makedepend -pobj_$(TAG)/ $(SRCS) 2> /dev/null
+	touch $(OUTPUT_DIR)/Makefile.makedepend
+	makedepend -f$(OUTPUT_DIR)/Makefile.makedepend -p$(OUTPUT_DIR)/ $(SRCS) 2> /dev/null
 	$(MAKE) -C ./cacti/ depend
 
 clean:
 	-rm -f *.o $(TARGET)
 	rm -f Makefile.makedepend Makefile.makedepend.bak
 
-include Makefile.makedepend
+include $(OUTPUT_DIR)/Makefile.makedepend
diff --git a/src/intersim/Makefile b/src/intersim/Makefile
index 0613274..cf42bb5 100644
--- a/src/intersim/Makefile
+++ b/src/intersim/Makefile
@@ -13,16 +13,19 @@ LEX    = flex
 PURIFY = /usr/bin/purify
 QUANT  = /usr/bin/quantify
 
+OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/intersim
 
 INTERFACE = interconnect_interface.cpp
 DEBUG = 0 
-CPPFLAGS = -g -Wall 
+CPPFLAGS = -g -Wall
 ifneq ($(DEBUG),1)
 CPPFLAGS = -O3 -g 
 else
 CPPFLAGS += 
 endif
 
+CPPFLAGS += -I$(OUTPUT_DIR) -I.
+
 TEST = -DUNIT_TEST 
 ifneq ($(UNIT_TEST),1)
 TEST = 
@@ -32,7 +35,6 @@ endif
 
 PROG     = intersim
 
-OUTPUT_DIR=../../$(SIM_OBJ_FILES_DIR)/intersim
 
 CPP_SRCS = $(INTERFACE) \
 	config_utils.cpp \
@@ -69,23 +71,18 @@ CPP_SRCS = $(INTERFACE) \
 	rng_double_wrapper.cpp \
 	statwraper.cpp
 
-#LEX_OBJS  = $(OUTPUT_DIR)/configlex.o
-LEX_OBJS  = configlex.o
-#YACC_OBJS = $(OUTPUT_DIR)/config_tab.o
-YACC_OBJS = config_tab.o
+LEX_OBJS  = $(OUTPUT_DIR)/configlex.o
+YACC_OBJS = $(OUTPUT_DIR)/config_tab.o
 
 #--- Make rules ---
 
 OBJS = $(CPP_SRCS:%.cpp=$(OUTPUT_DIR)/%.o) $(LEX_OBJS) $(YACC_OBJS)
 
 .PHONY: clean
-.PRECIOUS: %_tab.cpp %_tab.hpp %lex.cpp
+.PRECIOUS: $(OUTPUT_DIR)/%_tab.cpp $(OUTPUT_DIR)/%_tab.hpp $(OUTPUT_DIR)/%lex.cpp
 
 lib$(PROG).a: $(OBJS)
 	ar rcs $(OUTPUT_DIR)/lib$(PROG).a $(OBJS)
-	mv $(LEX_OBJS) $(OUTPUT_DIR)/
-	mv $(YACC_OBJS) $(OUTPUT_DIR)/
-	
 
 purify: $(OBJS)
 	$(PURIFY) -always-use-cache-dir $(CPP) $(OBJS) -o $(PROG) -L/usr/lib
@@ -93,26 +90,29 @@ purify: $(OBJS)
 quantify: $(OBJS)
 	$(QUANT) -always-use-cache-dir $(CPP) $(OBJS) -o $(PROG) -L/usr/lib
 
-%lex.o: %lex.cpp %_tab.hpp
+$(OUTPUT_DIR)/%lex.o: $(OUTPUT_DIR)/%lex.cpp $(OUTPUT_DIR)/%_tab.hpp
 	$(CPP) $(CPPFLAGS) -c $< -o $@
 
 $(OUTPUT_DIR)/%.o: %.cpp
 	$(CPP) $(EXTRA) $(CPPFLAGS) $(TEST) -c $< -o $@
+
+$(OUTPUT_DIR)/%.o: $(OUTPUT_DIR)/%.cpp
+	$(CPP) $(EXTRA) $(CPPFLAGS) $(TEST) -c $< -o $@
  
 $(OUTPUT_DIR)/%.o: %.c
 	$(CPP) $(CPPFLAGS) $(TEST) $(VCSFLAGS) -c $< -o $@
 
-%_tab.cpp: %.y
-	$(YACC) -b$* -p$* $<
-	cp -f $*.tab.c $*_tab.cpp
+$(OUTPUT_DIR)/%_tab.cpp: %.y
+	$(YACC) -b$* -p$* $< --file-prefix=$(OUTPUT_DIR)/$*
+	cp -f $(OUTPUT_DIR)/$*.tab.c $(OUTPUT_DIR)/$*_tab.cpp
 
-%_tab.hpp: %_tab.cpp
-	cp -f $*.tab.h $*_tab.hpp
+$(OUTPUT_DIR)/%_tab.hpp: $(OUTPUT_DIR)/%_tab.cpp
+	cp -f $(OUTPUT_DIR)/$*.tab.h $(OUTPUT_DIR)/$*_tab.hpp
 
-%lex.cpp: %.l
-	$(LEX) -P$* -o$@ $<
-	cp configlex.cpp configlex.cpp.orig
-	awk '/configlineno = 1/ {if (line == 0) {line = 1; print} next;} //{print}' configlex.cpp.orig > configlex.cpp
+$(OUTPUT_DIR)/%lex.cpp: %.l
+	$(LEX) --outfile=$@ -P$* $<
+	cp $(OUTPUT_DIR)/configlex.cpp $(OUTPUT_DIR)/configlex.cpp.orig
+	awk '/configlineno = 1/ {if (line == 0) {line = 1; print} next;} //{print}' $(OUTPUT_DIR)/configlex.cpp.orig > $(OUTPUT_DIR)/configlex.cpp
 
 clean:
 	rm -f $(OBJS) *_tab.cpp *_tab.hpp *.tab.c *.tab.h *lex.cpp
@@ -121,7 +121,7 @@ clean:
 	rm -f lib$(PROG).so
 	rm -f $(OBJS) *.o 
 
-interconnect_interface.o: ../cuda-sim/ptx.tab.h
+$(OUTPUT_DIR)/interconnect_interface.o: $(OUTPUT_DIR)/../cuda-sim/ptx.tab.h
 
-../cuda-sim/ptx.tab.h:
-	        make -C ../cuda-sim/ ptx.tab.c
+$(OUTPUT_DIR)/../cuda-sim/ptx.tab.h:
+	        make -C ../cuda-sim/ $(OUTPUT_DIR)/../cuda-sim/ptx.tab.c