#-------------------------------------------------------------------------------
# Programs and flags
# Define CXX and NVCC in environment or make.inc, if needed.

-include make.inc

NVCC     ?= nvcc

CXXFLAGS  = -std=c++11 -O3 -Wall -MMD
NVCCFLAGS = -std=c++11 -O3 -Xcompiler "-Wall -Wno-unused-function"
LDFLAGS   = -std=c++11

# Compile for CUDA architecture in ${GPU_TARGET}.
# (See SLATE or MAGMA Makefiles for more general code to compile for
# multiple architectures.)
# code=sm_XY is binary for CUDA architecture X.Y.
# code=compute_XY is PTX for forward compatability for CUDA architecture >= X.Y.
ifeq ("$(GPU_TARGET)","Kepler")
    NVCCFLAGS += -gencode arch=compute_30,code=sm_30 \
                 -gencode arch=compute_30,code=compute_30
endif
ifeq ("$(GPU_TARGET)","Maxwell")
    NVCCFLAGS += -gencode arch=compute_50,code=sm_50 \
                 -gencode arch=compute_50,code=compute_50
endif
ifeq ("$(GPU_TARGET)","Pascal")
    NVCCFLAGS += -gencode arch=compute_60,code=sm_60 \
                 -gencode arch=compute_60,code=compute_60
endif
ifeq ("$(GPU_TARGET)","Volta")
    NVCCFLAGS += -gencode arch=compute_70,code=sm_70 \
                 -gencode arch=compute_70,code=compute_70
endif
ifeq ("$(GPU_TARGET)","Turing")
    NVCCFLAGS += -gencode arch=compute_75,code=sm_75 \
                 -gencode arch=compute_75,code=compute_75
endif

# Extract CUDADIR, assuming nvcc is installed in ${CUDADIR}/bin/nvcc,
# with parallel ${CUDADIR}/include and ${CUDADIR}/lib directories.
nvcc_path := ${shell which ${NVCC}}
nvcc_dir  := ${dir ${nvcc_path}}
# %/=% strips trailing slash
CUDADIR   := ${dir ${nvcc_dir:%/=%}}
CUDADIR   := ${CUDADIR:%/=%}

# CUDA libraries
CPPFLAGS += -I${CUDADIR}/include
LDFLAGS  += -L${CUDADIR}/lib
LIBS     += -lcudart -lcublas -lnvToolsExt

#-------------------------------------------------------------------------------
# Makefile settings.

.SUFFIXES:
.DELETE_ON_ERROR:
.SECONDARY:

#-------------------------------------------------------------------------------
# Files

src = ${wildcard *.cc}
hdr = ${wildcard *.hh}
obj = ${addsuffix .o, ${basename ${src}}}
dep = ${addsuffix .d, ${basename ${src}}}
pch = ${addsuffix .pch, ${basename ${hdr}}}
exe = ${basename ${src}}

cu_src = ${wildcard *.cu}
cu_hdr = ${wildcard *.cuh}
cu_obj = ${addsuffix .o, ${basename ${cu_src}}}
cu_exe = ${basename ${cu_src}}

obj += ${cu_obj}
exe += ${cu_exe}

#-------------------------------------------------------------------------------
# Rules

all: ${exe} ${cu_exe}

%: %.o
	${CXX} ${LDFLAGS} ${LIBS} -o $@ $^

%.o: %.cc
	${CXX} ${CXXFLAGS} ${CPPFLAGS} -c -o $@ $<

%.o: %.cu
	${NVCC} ${NVCCFLAGS} ${CPPFLAGS} -c -o $@ $<

# Compiler-generated header dependencies.
-include ${dep}

# nvcc can't take -MMD due to tmp files, so manually add dependency.
${cu_obj}: util.hh

#----------
# Pre-compile headers to test for independence, not for inclusion.
pch: ${pch}

%.pch: %.hh
	${CXX} ${CXXFLAGS} ${CPPFLAGS} -c -o $@ $<

#----------
clean:
	-rm -f ${exe} ${cu_exe} ${obj} ${cu_obj} ${dep} ${pch}

# *.cc
push:
	rsync -av Makefile *.hh *.cu *.sh saturn:sc19/

ornl:
	rsync -av Makefile *.hh *.cu *.sh ornl:sc19/

#-------------------------------------------------------------------------------
# For Makefile debugging

echo:
	@echo "CXX        = '${CXX}'"
	@echo "NVCC       = '${NVCC}'"
	@echo
	@echo "nvcc_path  = '${nvcc_path}'"
	@echo "nvcc_dir   = '${nvcc_dir}'"
	@echo "CUDADIR    = '${CUDADIR}'"
	@echo "GPU_TARGET = '${GPU_TARGET}'"
	@echo
	@echo "CXXFLAGS   = '${CXXFLAGS}'"
	@echo "NVCCFLAGS  = '${NVCCFLAGS}'"
	@echo "CPPFLAGS   = '${CPPFLAGS}'"
	@echo "LDFLAGS    = '${LDFLAGS}'"
	@echo "LIBS       = '${LIBS}'"
	@echo
	@echo "src = ${src}"
	@echo "obj = ${obj}"
	@echo "dep = ${dep}"
	@echo "exe = ${exe}"
	@echo
	@echo "hdr = ${hdr}"
	@echo "pch = ${pch}"