Makefile.kokkos

KOKKOS_PATH = ${HOME}/Kokkos/upstream/kokkos
SRC = su3_nn_bench.cpp
KOKKOS_DEVICES=Cuda
#KOKKOS_DEVICES=OpenMPTarget


default: build
	echo "Start Build"

CXXFLAGS= -O3 -g
LINKFLAGS=

# ARCH = V100 | A100
ifeq ($(ARCH),A100)
	KOKKOS_ARCH = AMPERE80
else
	KOKKOS_ARCH = VOLTA70
endif


CXXFLAGS+= -DUSE_KOKKOS
ifeq ($(KOKKOS_DEVICES),Cuda)
	KOKKOS_CUDA_OPTIONS=enable_lambda
	CXX = $(KOKKOS_PATH)/bin/nvcc_wrapper
	KOKKOS_CXX_STANDARD=c++14
	EXE32 = bench_f32_kokkos_cuda.exe
	EXE64 = bench_f64_kokkos_cuda.exe
else
	KOKKOS_CXX_STANDARD=c++17
	CXX = clang++
	# To avoid warnings
	CXXFLAGS += -Wno-unknown-cuda-version -Wno-undefined-internal -Wno-pass-failed -Wno-openmp-mapping -Wno-extra-tokens
	LINKFLAGS +=-Wno-unknown-cuda-version
	EXE32 = bench_f32_kokkos_ompt.exe
	EXE64 = bench_f64_kokkos_ompt.exe
endif

DEPENDS = su3.hpp lattice.hpp mat_nn_kokkos.hpp

LINK = ${CXX}

DEPFLAGS = -M

OBJ32= $(SRC:.cpp=.o)
OBJ64= $(SRC:.cpp=.o)
LIB =

include $(KOKKOS_PATH)/Makefile.kokkos

build: $(EXE32) $(EXE64)

$(EXE32): $(OBJ32) $(KOKKOS_LINK_DEPENDS)
	$(LINK) $(KOKKOS_LDFLAGS) $(LINKFLAGS) $(EXTRA_PATH) $(OBJ32) $(KOKKOS_LIBS) $(LIB) -o $(EXE32)

$(EXE64): $(OBJ64) $(KOKKOS_LINK_DEPENDS)
	$(LINK) $(KOKKOS_LDFLAGS) $(LINKFLAGS) $(EXTRA_PATH) $(OBJ64) $(KOKKOS_LIBS) $(LIB) -o $(EXE64)

# Compilation rules

$(OBJ32):$(SRC) $(KOKKOS_CPP_DEPENDS) $(DEPENDS)
	$(CXX) -DPRECISION=1 $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) $(EXTRA_INC) -c $<

$(OBJ64):$(SRC) $(KOKKOS_CPP_DEPENDS) $(DEPENDS)
	$(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) $(EXTRA_INC) -c $<

clean: kokkos-clean
	rm -f *.o *kokkos*.exe