diff --git a/Benchmarks/ChiBenchmark/Makefile.GPU.intel b/Benchmarks/ChiBenchmark/Makefile.GPU.intel
index 0a690af..3abea5b 100644
--- a/Benchmarks/ChiBenchmark/Makefile.GPU.intel
+++ b/Benchmarks/ChiBenchmark/Makefile.GPU.intel
@@ -1,85 +1,87 @@
 PROGRAM_NAME := ChiBenchmark
 #CXX=g++ -lm -ffast-math -ftree-loop-vectorize 
-CXX=icpc
+CXX=mpiicpc
 
 
 program_CXX_SRCS := $(wildcard *.cpp)
 program_CXX_OBJS := ${program_CXX_SRCS:.cpp=.o}
 
 #program_CXX_SRCS += $(wildcard ../../*.c) #Find C source files from additonal directories
 program_C_OBJS := ${program_C_SRCS:.c=.o}
 #
 program_CU_SRCS := $(wildcard *.cu)
 #program_CU_SRCS += $(wildcard ../../*.cu) #Find CUDA source files from additional directories
 #program_CU_HEADERS := $(wildcard *.cuh) #Optional: Include .cuh files as dependencies
 #program_CU_HEADERS += $(wildcard ../../*.cuh) #Find .cuh files from additional directories
 program_CU_OBJS := ${program_CU_SRCS:.cu=.cuo}
 #
 program_INCLUDE_DIRS := . /usr/local/cuda/include/ #C++ Include directories
 program_INCLUDE_DIRS += $(CFITSIO_ROOT)/include
 #program_INCLUDE_DIRS += /users/fgilles/bin/iaca-lin32/include
 program_INCLUDE_DIRS += $(LENSTOOL_ROOT)/include
 program_INCLUDE_DIRS += $(LENSTOOL_ROOT)/src
 program_INCLUDE_DIRS += $(LENSTOOL_ROOT)/liblit
 
 program_INCLUDE_DIRS += $(GSL_ROOT)
 program_INCLUDE_DIRS += $(LENSTOOLHPC_ROOT)/src
 #
 #program_CU_INCLUDE_DIRS := /home/users/amclaugh/CUB/cub-1.3.2/ #CUDA Include directories
 #
 #program_INCLUDE_LIBS := ./ #Include libraries
 program_INCLUDE_LIBS += $(CFITSIO_ROOT)/lib #Include libraries
 program_INCLUDE_LIBS += $(LENSTOOL_ROOT)/src
 program_INCLUDE_LIBS += $(LENSTOOL_ROOT)/liblt
 program_INCLUDE_LIBS += $(LENSTOOL_ROOT)/include
 program_INCLUDE_LIBS += $(LENSTOOLHPC_ROOT)/src/
 program_INCLUDE_LIBS += $(GSL_ROOT)/lib
 program_INCLUDE_LIBS += $(WCSTOOL_ROOT)
 #
 #
 # Compiler flags
 CPPFLAGS += $(foreach includedir,$(program_INCLUDE_DIRS),-I$(includedir))
 CPPFLAGS += $(foreach includelib,$(program_INCLUDE_LIBS),-L$(includelib))
 CPPFLAGS += -D__WITH_LENSTOOL
+CPPFLAGS += -D__WITH_MPI
+CPPFLAGS += -D__WITH_GPU
 CPPFLAGS += -qopenmp -xHost -g -O3 -std=c++0x -Wall -pedantic
 #CPPFLAGS += -qopenmp -march=core-avx2 -g -O3 -std=c++0x -Wall -pedantic
 #CPPFLAGS += -llenstoolhpc -qopenmp -xHost -g -O3 -std=c++0x -Wall -pedantic
 #CPPFLAGS += -llenstoolhpc -qopenmp -axMIC-AVX512,CORE-AVX2 -g -O3 -std=c++0x -Wall -pedantic
 #CPPFLAGS += -qopt-prefetch-distance=64,8 -qopt-streaming-cache-evict=0 -llenstoolhpc -qopenmp -xMIC-AVX512 -g -O3 -std=c++0x -Wall -pedantic
 LDFLAGS := -llenstoolhpc -llenstool -llt -lgsl -lgslcblas -lm -lwcs -lcfitsio
 
 NVFLAGS := -O3 -g -G -arch=sm_35 -rdc=true -ccbin icpc -Xcompiler -qopenmp #rdc=true needed for separable compilation
 NVFLAGS += -D__WITH_LENSTOOL 
 NVFLAGS += $(foreach includedir,$(program_INCLUDE_DIRS),-I$(includedir))
 NVFLAGS += $(foreach includelib,$(program_INCLUDE_LIBS),-L$(includelib)) 
 
 CUO_O_OBJECTS := ${program_CU_OBJS:.cuo=.cuo.o}
 
 
 OBJECTS = $(program_CXX_OBJS) $(program_C_OBJS) $(program_CU_OBJS)
 
 .PHONY: all clean distclean
 
 all: $(PROGRAM_NAME) 
 #
 debug: CXXFLAGS = -g -O0 -std=c++0x -Wall -pedantic -DDEBUG $(EXTRA_FLAGS)
 debug: $(PROGRAM_NAME)
 %.cuo: %.cu %.cuh
 	nvcc $(NVFLAGS) -o $@ -dc $<
 #nvcc $(NVFLAGS)  -Xptxas -v,-abi=no -ptx -src-in-ptx -dc $<
 
 $(PROGRAM_NAME): $(OBJECTS) 
 	@ for cu_obj in $(program_CU_OBJS); \
 	do        \
 		mv $$cu_obj $$cu_obj.o; \
 	done #append a .o suffix for nvcc
 	nvcc $(NVFLAGS) -o $@ $(program_CXX_OBJS) $(program_C_OBJS) $(CUO_O_OBJECTS) $(LDFLAGS)
 	@ for cu_obj in $(CUO_O_OBJECTS);  \
 	do        \
 		mv $$cu_obj $${cu_obj%.*} ;  \
 	done #remove the .o for make
 
 clean:
 	@- $(RM) $(PROGRAM_NAME) $(OBJECTS) *~ *.o *.optrpt
 
-distclean: clean
\ No newline at end of file
+distclean: clean