Makefile

## compilation for different platforms and 
## compilers
ARCHS = knl hsw skl nec
ARCH ?= hsw

COMPILERs = icc gnu ncc
COMPILER ?= gnu

$(if $(filter $(COMPILERs),$(COMPILER)),,$(error COMPILER must be one of $(COMPILERs)))
$(if $(filter $(ARCHS),$(ARCH)),,$(error ARCH must be one of $(ARCHS)))

COMPILER_is_$(COMPILER) := yes
ARCH_is_$(ARCH)         := yes

## for openmp
-omp := $(if $(COMPILER_is_icc), -qopenmp, -fopenmp)
## for intel avx
-avx := $(if $(COMPILER_is_icc), $(if $(ARCH_is_knl), -xMIC-AVX512, $(if $(ARCH_is_skl), -xCORE-AVX512, -xCORE-AVX2) ),)
-rpt := $(if $(COMPILER_is_icc), -qopt-report=5, $(if $(COMPILER_is_ncc), , -ftree-vectorize -fopt-info-vec-missed))
## for nec mpi version
-nccflag := $(if $(COMPILER_is_ncc), -DNEC,)
-nccdep := $(if $(COMPILER_is_ncc), -M,)
## for knl MCDRAM
-knlflag := $(if $(ARCH_is_knl), -DSC_MEMKIND -I./memkind/include -L./memkind/lib -lmemkind, )
-mkl := $(if $(COMPILER_is_icc), -I/opt/intel/mkl/include, )
-mklld := $(if $(COMPILER_is_icc), -L/opt/intel/mkl/lib/intel64 -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core,)
## for intel prefetch
-prefetch := $(if $(COMPILER_is_icc), -qopt-prefetch=3,)
# -mpiheader := -I/opt/intel/compilers_and_libraries_2019.0.117/linux/mpi/intel64/include

## include the proprogation blocking codes
# VPATH=./radix
## ----------- start of compilation -----------

# output binary
BIN := sc-$(ARCH)-$(COMPILER).bin

# source files
SRCS := \
	$(wildcard *.cpp)
   # $(wildcard ./radix/*.cpp) \
	#$(wildcard ./radix/commons/*.cpp) \
	#$(wildcard ./SpBLAS/*.cpp) \
	#$(wildcard ./SpDM3/*.cpp) \
	#$(wildcard ./SpMP/*.cpp) \
	#$(wildcard ./SpMP/reordering/*.cpp) \
	#$(wildcard ./SpMP/synk/*.cpp)

# files included in the tarball generated by 'make dist' (e.g. add LICENSE file)
DISTFILES := $(BIN)

# filename of the tar archive generated by 'make dist'
DISTOUTPUT := $(BIN).tar.gz

# intermediate directory for generated object files
OBJDIR := .o
# intermediate directory for generated dependency files
DEPDIR := .d

# object files, auto generated from source files
OBJS := $(patsubst %,$(OBJDIR)/%.o,$(basename $(SRCS)))
# dependency files, auto generated from source files
DEPS := $(patsubst %,$(DEPDIR)/%.d,$(basename $(SRCS)))

# compilers (at least gcc and clang) don't create the subdirectories automatically
$(shell mkdir -p $(dir $(OBJS)) >/dev/null)
$(shell mkdir -p $(dir $(DEPS)) >/dev/null)

# C compiler
CC := $(if $(COMPILER_is_ncc), ncc, $(if $(COMPILER_is_icc), icc, gcc))
# C++ compiler
CXX := $(if $(COMPILER_is_ncc), nc++, $(if $(COMPILER_is_icc), icpc, g++))
# linker
LD := $(if $(COMPILER_is_ncc), nc++, $(if $(COMPILER_is_icc), icpc, g++))
# tar
TAR := tar

# C flags
CFLAGS := -std=c11 $(-nccflag) $(-omp) $(-avx) $(-rpt) $(-mkl) -O3
# C++ flags
CXXFLAGS := -std=c++11 $(-nccflag) $(-omp) $(-avx) $(-rpt) $(-mkl) $(-prefetch) -O3
# C/C++ flags
CPPFLAGS := -g -Wall -Wextra -pedantic -DVTUNE -DVERBOSE -DUSE_BLAS
# CPPFLAGS := -g -Wall -pedantic -DVERBOSE
# linker flags
LDFLAGS := $(-omp) $(-mklld)
# flags required for dependency generation; passed to compilers
DEPFLAGS = $(-nccdep) -MT $@ -MD -MP -MF  $(DEPDIR)/$*.Td

# compile C source files
COMPILE.c = $(CC) $(DEPFLAGS) $(CFLAGS) $(CPPFLAGS) -c -o $@
# compile C++ source files
#COMPILE.cc = $(CXX) $(DEPFLAGS) $(CXXFLAGS) $(CPPFLAGS) -c -o $@
COMPILE.cc = $(CXX) $(DEPFLAGS) $(CXXFLAGS) -c -o $@
# link object files to binary
LINK.o = $(LD) $(LDFLAGS) $(LDLIBS) -o $@
# precompile step
PRECOMPILE =
# postcompile step
POSTCOMPILE = mv -f $(DEPDIR)/$*.Td $(DEPDIR)/$*.d

all: $(BIN)

dist: $(DISTFILES)
	$(TAR) -cvzf $(DISTOUTPUT) $^

.PHONY: clean
clean:
	$(RM) -r $(OBJDIR) $(DEPDIR)

.PHONY: distclean
distclean: clean
	$(RM) $(BIN) $(DISTOUTPUT)

.PHONY: install
install:
	@echo no install tasks configured

.PHONY: uninstall
uninstall:
	@echo no uninstall tasks configured

.PHONY: check
check:
	@echo no tests configured

.PHONY: help
help:
	@echo available targets: all dist clean distclean install uninstall check

$(BIN): $(OBJS)
	$(LINK.o) $^

$(OBJDIR)/%.o: %.c
$(OBJDIR)/%.o: %.c $(DEPDIR)/%.d
	$(PRECOMPILE)
	$(COMPILE.c) $<
	$(POSTCOMPILE)

$(OBJDIR)/%.o: %.cpp
$(OBJDIR)/%.o: %.cpp $(DEPDIR)/%.d
	$(PRECOMPILE)
	$(COMPILE.cc) $<
	$(POSTCOMPILE)

$(OBJDIR)/%.o: %.cc
$(OBJDIR)/%.o: %.cc $(DEPDIR)/%.d
	$(PRECOMPILE)
	$(COMPILE.cc) $<
	$(POSTCOMPILE)

$(OBJDIR)/%.o: %.cxx
$(OBJDIR)/%.o: %.cxx $(DEPDIR)/%.d
	$(PRECOMPILE)
	$(COMPILE.cc) $<
	$(POSTCOMPILE)

.PRECIOUS = $(DEPDIR)/%.d
$(DEPDIR)/%.d: ;

-include $(DEPS)