Commit 070e2096 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Updated Makefile based on tnl-mhfem

parent 60ce7e64

Makefile

+2 −9
Original line number Diff line number Diff line
@@ -19,8 +19,8 @@ host: tnl-benchmark-mesh 
tnl-benchmark-mesh: tnl-benchmark-mesh.o $(MESH_BENCHMARK_TEMPLATES_CPP:%.cpp=%.o)

	$(CXX) $(LDFLAGS) -o $@ $^ $(LDLIBS)

cuda: tnl-benchmark-mesh-cuda

tnl-benchmark-mesh-cuda: tnl-benchmark-mesh-cuda.cu.o $(MESH_BENCHMARK_TEMPLATES_CU:%.cu=%.cu.o)

	$(CUDA_COMPILER) $(CUDA_LDFLAGS) -o $@ $^ $(CUDA_LDLIBS)

tnl-benchmark-mesh-cuda: tnl-benchmark-mesh-cuda.cuo $(MESH_BENCHMARK_TEMPLATES_CU:%.cu=%.cuo)

	$(CXX) $(CUDA_LDFLAGS) -o $@ $^ $(CUDA_LDLIBS)



clean: clean_templates

.PHONY: clean_templates
@@ -29,11 +29,4 @@ clean_templates: 
	$(RM) -r MeshBenchmarks.templates/



-include $(SOURCES:%.cpp=%.d)



ifeq ($(CUDA_COMPILER),nvcc)

# nvcc creates .cu.d with rubbish and .d with the content we need

-include $(CUDA_SOURCES:%.cu=%.d)
 
else

# clang creates .cu.d

-include $(CUDA_SOURCES:%.cu=%.cu.d)

endif

Makefile.base

+134 −105
Original line number Diff line number Diff line
@@ -3,18 +3,6 @@ 
#

# vim: ft=make



# variables configurable from command line (e.g. make TNL_CXX=clang++ )

# or from environment (e.g. TNL_CXX=clang++ make )

# Note that we have to use TNL_CXX instead of CXX to set the default value,

# because CXX is an implicit variable defined by GNU make itself. (Thank you!)

TNL_CXX ?= g++

CXX := $(TNL_CXX)

CUDA_COMPILER ?= nvcc

CUDA_HOST_COMPILER ?= $(CXX)

# BUILD can be 'Release' or 'Debug'

BUILD ?= Release





# relative paths used in the messages

mkfile_path := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))

current_dir := ./$(subst $(mkfile_path),,$(CURDIR)/)
@@ -23,6 +11,14 @@ current_dir := $(current_dir:./%/=%) 
# include colors

include $(mkfile_path)/Makefile.colors



# include global and local configs

include $(mkfile_path)/config.mk

ifneq ($(current_dir),./)  # to skip double inclusion of the main config.mk

ifneq ("$(wildcard $(CURDIR)/config.mk)","")  # to check if the file exists

include $(CURDIR)/config.mk

endif

endif



# make does not have an OR statement...

ifneq ($(filter $(BUILD),Release Debug),)

    $(info $(shell printf "$(COLOR_BOLD)### Building directory $(current_dir) in $(BUILD) mode$(COLOR_ALL_OFF)" ))
@@ -31,14 +27,17 @@ $(error Wrong BUILD: $(BUILD)) 
endif





## essential host options

CPPFLAGS = -MD -MP

CXXFLAGS = -std=c++14 -Wall -Wno-unused-local-typedefs -Wno-unused-variable  -Wno-deprecated -Wno-deprecated-declarations

LDFLAGS = -pthread

LDLIBS = -lstdc++fs

ifneq ($(CXX),clang++)

## essential host compiler flags

CPPFLAGS := -MD -MP

CXXFLAGS := -std=c++14 -Wall -Wno-unused-local-typedefs

ifeq ($(CXX),g++)

    CXXFLAGS += -Wno-maybe-uninitialized

endif

## essential linker flags

LDFLAGS := -pthread

CUDA_LDFLAGS := $(LDFLAGS)

LDLIBS := -lstdc++fs

CUDA_LDLIBS := $(LDLIBS)



ifeq ($(BUILD),Release)

    CXXFLAGS += -O3 -march=native -mtune=native -DNDEBUG
@@ -47,6 +46,9 @@ CXXFLAGS += -O3 -march=native -mtune=native -DNDEBUG 
    ifeq ($(CXX),clang++)

        CXXFLAGS += -flto=thin

        LDFLAGS += -O3 -march=native -mtune=native -flto=thin

    else ifeq ($(CXX),g++)

        CXXFLAGS += -flto

        LDFLAGS += -O3 -march=native -mtune=native -flto

    endif

else ifeq ($(BUILD),Debug)

    CXXFLAGS += -Og -g
@@ -60,21 +62,26 @@ LDFLAGS += -fuse-ld=lld 
endif





## add flags for linking CUDA with the host compiler

CUDA_LDFLAGS += -L $(CUDA_PATH)/lib64/

ifeq ($(STATIC_LINKING),True)

    CUDA_LDLIBS += -lcudart_static

else

    CUDA_LDLIBS += -lcudart

endif

CUDA_LDLIBS += -ldl -lrt



## options for nvcc

# automatic dependency generation for nvcc (gcc has it automated in the pre-processor phase with the -MD flag)

# note that $@ is expanded only when $(CUDA_CPPFLAGS) is used

CUDA_CPPFLAGS = --compiler-options -MD,-MP,-MT$@

CUDA_CXXFLAGS = -std=c++14 --compiler-bindir $(CUDA_HOST_COMPILER)

CUDA_LDLIBS = -lstdc++fs

CUDA_CPPFLAGS := -MD -MP

CUDA_CXXFLAGS := -std=c++14 --compiler-bindir $(CUDA_HOST_COMPILER)

# enable nvcc features

CUDA_CXXFLAGS += --expt-relaxed-constexpr --expt-extended-lambda --default-stream per-thread

# disable false compiler warnings

#   reference for the -Xcudafe flag: http://stackoverflow.com/questions/14831051/how-to-disable-compiler-warnings-with-nvcc/17095910#17095910

#   list of possible tokens: http://www.ssl.berkeley.edu/~jimm/grizzly_docs/SSL/opt/intel/cc/9.0/lib/locale/en_US/mcpcom.msg

CUDA_CXXFLAGS += --expt-relaxed-constexpr --expt-extended-lambda -Xcudafe "--diag_suppress=code_is_unreachable --diag_suppress=loop_not_reachable --diag_suppress=implicit_return_from_non_void_function --diag_suppress=unsigned_compare_with_zero --display_error_number" -Xcompiler -Wno-vla

# This diagnostic is just plain wrong in CUDA 9

# See https://github.com/kokkos/kokkos/issues/1470

CUDA_CXXFLAGS += -Xcudafe "--diag_suppress=code_is_unreachable --diag_suppress=loop_not_reachable --diag_suppress=implicit_return_from_non_void_function --diag_suppress=unsigned_compare_with_zero --display_error_number" -Xcompiler -Wno-vla

# This diagnostic is just plain wrong in CUDA 9 and later, see https://github.com/kokkos/kokkos/issues/1470

CUDA_CXXFLAGS += -Xcudafe --diag_suppress=esa_on_defaulted_function_ignored

# disable deprecation warnings

CUDA_CXXFLAGS += -Xcudafe --diag_suppress=1444



ifeq ($(BUILD),Release)

    CUDA_CXXFLAGS += -O3 -DNDEBUG -Xcompiler -march=native,-mtune=native
@@ -84,20 +91,9 @@ endif 


## options for clang as CUDA compiler

ifeq ($(CUDA_COMPILER),clang++)

CUDA_CPPFLAGS = $(CPPFLAGS)

# immediate expansion of the following two vars

    CUDA_CPPFLAGS := $(CPPFLAGS)

    CUDA_CXXFLAGS := $(CXXFLAGS)

CUDA_LDFLAGS := $(LDFLAGS)

# FIXME: better detection

_CUDA_PATH = $(dir $(subst -I,,$(shell pkg-config --cflags-only-I cuda)))

CUDA_CXXFLAGS += --cuda-path=$(_CUDA_PATH)

CUDA_LDFLAGS += -L $(_CUDA_PATH)/lib64/

ifeq ($(STATIC_LINKING),True)

CUDA_LDLIBS += -lcudart_static

else

CUDA_LDLIBS += -lcudart

endif

CUDA_LDLIBS += -ldl -lrt

    CUDA_CXXFLAGS += --cuda-path=$(CUDA_PATH)

    # include info for profiling

    #CUDA_CXXFLAGS += -Xcuda-fatbinary --cmdline -Xcuda-fatbinary " --generate-line-info " -Xcuda-ptxas --generate-line-info -g

    # debugging register spills
@@ -108,32 +104,36 @@ CUDA_LDLIBS += -ldl -lrt 
    #CUDA_CXXFLAGS += -Xcuda-ptxas -dlcm=ca

endif



# select CUDA GPU architecture

# set the CUDA GPU architecture

ifeq ($(CUDA_COMPILER),clang++)

    ifeq ($(CUDA_GPU_ARCH),auto)

        CUDA_CXXFLAGS += $(shell tnl-cuda-arch --clang)

    else

        CUDA_CXXFLAGS += --cuda-gpu-arch=$(CUDA_GPU_ARCH)

    endif

else

    ifeq ($(CUDA_GPU_ARCH),auto)

        CUDA_CXXFLAGS += $(shell tnl-cuda-arch)

    else

        CUDA_CXXFLAGS += --gpu-architecture=$(CUDA_GPU_ARCH)

    endif

endif





# append flags for TNL

PKGS = tnl

PKGS_CUDA = tnl-cuda

CXXFLAGS += $(shell pkg-config --cflags $(PKGS))

CUDA_CXXFLAGS += $(shell pkg-config --cflags $(PKGS_CUDA))

#ifeq ($(STATIC_LINKING),True)

## hack because cmake does not support shared and static libraries with the same name

#LDLIBS += $(subst tnl,tnl_static,$(shell pkg-config --libs $(PKGS)))

#CUDA_LDLIBS += $(subst tnl,tnl-cuda_static,$(shell pkg-config --libs $(PKGS_CUDA)))

#else

#LDLIBS += $(shell pkg-config --libs $(PKGS))

#CUDA_LDLIBS += $(shell pkg-config --libs $(PKGS_CUDA))

#endif

# TODO: add TNL as a git submodule

#TNL_INCLUDE_DIRS := -I $(mkfile_path)/libs/tnl/src/ -I $(mkfile_path)/libs/tnl/src/3rdparty/

TNL_INCLUDE_DIRS := -I ~/.local/include

CPPFLAGS += $(TNL_INCLUDE_DIRS)

CUDA_CPPFLAGS += $(TNL_INCLUDE_DIRS) -DHAVE_CUDA



# OpenMP

ifeq ($(CXX),clang++)

    OPENMP_CXXFLAGS := -fopenmp=libomp -DHAVE_OPENMP

    OPENMP_LDLIBS := -lomp

else ifeq ($(CXX),icpc)

    OPENMP_CXXFLAGS := -fopenmp -DHAVE_OPENMP

    OPENMP_LDLIBS := -liomp5

else

    OPENMP_CXXFLAGS := -fopenmp -DHAVE_OPENMP

    OPENMP_LDLIBS := -lgomp
@@ -158,10 +158,13 @@ endif 




# base targets

.PHONY: all cuda host

all: cuda host

.PHONY: all cuda host mpi_host mpi_cuda mpi

all: cuda host mpi

host: $(TARGETS) subdirs

cuda: $(CUDA_TARGETS)

mpi_host: $(MPI_TARGETS)

mpi_cuda: $(MPI_CUDA_TARGETS)

mpi: mpi_host mpi_cuda



# descend into $(SUBDIRS) (when defined)

.PHONY: subdirs $(SUBDIRS)
@@ -177,7 +180,7 @@ ifdef SOURCES 
	$(RM) $(SOURCES:%.cpp=%.o) $(SOURCES:%.cpp=%.d)

endif

ifdef CUDA_SOURCES

	$(RM) $(CUDA_SOURCES:%.cu=%.cu.o) $(CUDA_SOURCES:%.cu=%.d) $(CUDA_SOURCES:%.cu=%.cu.d)

	$(RM) $(CUDA_SOURCES:%.cu=%.cuo) $(CUDA_SOURCES:%.cu=%.d) $(CUDA_SOURCES:%.cu=%.cu.d)

endif

ifdef TARGETS

	$(RM) $(TARGETS)
@@ -185,7 +188,14 @@ endif 
ifdef CUDA_TARGETS

	$(RM) $(CUDA_TARGETS)

endif

	$(RM) tmpxft_*.d

ifdef MPI_TARGETS

	$(RM) $(MPI_SOURCES:%.cpp=%_mpi.o) $(MPI_SOURCES:%.cpp=%_mpi.d)

	$(RM) $(MPI_TARGETS)

endif

ifdef MPI_CUDA_TARGETS

	$(RM) $(MPI_CUDA_TARGETS)

	$(RM) $(MPI_CUDA_SOURCES:%.cu=%_mpi.cuo) $(MPI_CUDA_SOURCES:%.cu=%_mpi.d) $(MPI_CUDA_SOURCES:%.cu=%_mpi.cu.d)

endif



# recursive cleanup

.PHONY: distclean
@@ -197,11 +207,30 @@ distclean: clean 
$(TARGETS): % : %.o

	$(CXX) $(LDFLAGS) -o $@ $< $(LDLIBS)



$(CUDA_TARGETS): % : %.cu.o

	$(CUDA_COMPILER) $(CUDA_LDFLAGS) -o $@ $< $(CUDA_LDLIBS)

# use .cuo instead of .cu.o to avoid problems with the implicit rules: https://stackoverflow.com/q/62967939

# (and use the host compiler for linking CUDA, nvcc does not understand that .cuo is an object file)

$(CUDA_TARGETS): % : %.cuo

	$(CXX) $(CUDA_LDFLAGS) -o $@ $< $(CUDA_LDLIBS)



$(SOURCES:%.cpp=%.o): %.o: %.cpp

	$(CXX) $(CPPFLAGS) $(CXXFLAGS) -c -o $@ $<



$(CUDA_SOURCES:%.cu=%.cu.o): %.cu.o : %.cu

$(CUDA_SOURCES:%.cu=%.cuo): %.cuo : %.cu

	$(CUDA_COMPILER) $(CUDA_CPPFLAGS) $(CUDA_CXXFLAGS) -c -o $@ $<



# rules for MPI targets

$(MPI_TARGETS): % : %.o

	$(CXX) $(LDFLAGS) -o $@ $< $(LDLIBS) $(MPI_LDLIBS)



$(MPI_CUDA_TARGETS): % : %.cuo

	$(CXX) $(CUDA_LDFLAGS) -o $@ $< $(CUDA_LDLIBS) $(MPI_LDLIBS)



$(MPI_SOURCES:%.cpp=%_mpi.o): %_mpi.o: %.cpp

	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(MPI_CXXFLAGS) -c -o $@ $<



$(MPI_CUDA_SOURCES:%.cu=%_mpi.cuo): %_mpi.cuo : %.cu

ifeq ($(CUDA_COMPILER),nvcc)

	$(CUDA_COMPILER) $(CUDA_CPPFLAGS) $(CUDA_CXXFLAGS) $(MPI_NVCC_CXXFLAGS) -c -o $@ $<

else

	$(CUDA_COMPILER) $(CUDA_CPPFLAGS) $(CUDA_CXXFLAGS) $(MPI_CXXFLAGS) -c -o $@ $<

endif

config.mk

0 → 100644
+30 −0
Original line number Diff line number Diff line 
# All variables are also configurable from the command line (e.g. make TNL_CXX=clang++ )

# or from the environment (e.g. TNL_CXX=clang++ make ).

#

# If you are building targets in a subdirectory, you can create a local config.mk

# file in that subdirectory to override options from the global config.mk file.



# build type (can be 'Release' or 'Debug')

BUILD ?= Release



# compiler for the .cpp files

# Note that we have to use TNL_CXX instead of CXX to set the default value,

# because CXX is an implicit variable defined by GNU make itself. (Thank you!)

TNL_CXX ?= g++

CXX := $(TNL_CXX)



# compiler for the .cu files

CUDA_COMPILER ?= nvcc



# compiler for the host/CPU code used by nvcc

CUDA_HOST_COMPILER ?= $(CXX)



# path to the CUDA toolkit installation

# (autodetection is attempted, set it manually if it fails)

CUDA_PATH ?= $(abspath $(dir $(shell command -v nvcc))/..)

#$(info Detected CUDA_PATH: $(CUDA_PATH))



# CUDA GPU architecture (e.g. "sm_61" or "auto")

# (if you use "auto", tnl-cuda-arch must be installed in your $PATH)

CUDA_GPU_ARCH ?= auto

#CUDA_GPU_ARCH ?= sm_70