Includes all the changes to the makefiles, adding new targets for CUDA/cppavx builds

epochX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/MG5aMC_patches/PROD/patch.common

@@ -82,11 +82,11 @@ index 348c283be..74db44d84 100644
 +###$(info CUDACPP_MAKEENV=$(CUDACPP_MAKEENV))
 +###$(info $(foreach v,$(CUDACPP_MAKEENV),$(v)="$($(v))"))
 +CUDACPP_BUILDDIR:=$(shell $(MAKE) $(foreach v,$(CUDACPP_MAKEENV),$(v)="$($(v))") -f $(CUDACPP_MAKEFILE) -pn 2>/dev/null | awk '/Building/{print $$3}' | sed s/BUILDDIR=//)
-+#ifeq ($(CUDACPP_BUILDDIR),)
-+#$(error CUDACPP_BUILDDIR='$(CUDACPP_BUILDDIR)' should not be empty!)
-+#else
++ifeq ($(CUDACPP_BUILDDIR),)
++$(error CUDACPP_BUILDDIR='$(CUDACPP_BUILDDIR)' should not be empty!)
++else
 +$(info CUDACPP_BUILDDIR='$(CUDACPP_BUILDDIR)')
-+#endif
++endif
 +CUDACPP_COMMONLIB=mg5amc_common
 +CUDACPP_CXXLIB=mg5amc_$(processid_short)_cpp
 +CUDACPP_CULIB=mg5amc_$(processid_short)_cuda
@@ -340,3 +340,248 @@ index a056d3861..b70b548e5 100755
                      alljobs = misc.glob('ajob*', Pdir)
 
                      #remove associated results.dat (ensure to not mix with all data)
+diff --git b/epochX/cudacpp/gg_tt.mad/SubProcesses/makefile a/epochX/cudacpp/gg_tt.mad/SubProcesses/makefile
+index 74db44d84..7ae599f03 100644
+--- b/epochX/cudacpp/gg_tt.mad/SubProcesses/makefile
++++ a/epochX/cudacpp/gg_tt.mad/SubProcesses/makefile
+@@ -9,6 +9,33 @@ FFLAGS+= -cpp
+ # Compile counters with -O3 as in the cudacpp makefile (avoid being "unfair" to Fortran #740)
+ CXXFLAGS = -O3 -Wall -Wshadow -Wextra
+
++# Sets correct target based on MAKECMDGOALS
++ifeq ($(MAKECMDGOALS),)
++    TARGET := undefined
++else
++    ifeq ($(shell grep -q madevent_ <<< $(MAKECMDGOALS)a),)
++        TARGET := $(word 2,($(subst _, ,$(MAKECMDGOALS))))
++    else
++        TARGET := $(MAKECMDGOALS)
++    endif
++endif
++
++# Sets correct cppavx target as dependency when compiling Fortran
++ifeq ($(shell grep -q cpp <<< $(MAKECMDGOALS)a),)
++	ifeq ($(TARGET),cpp)
++		CPPTARGET := cppavx2
++	else
++		CPPTARGET := $(TARGET)
++	endif
++endif
++
++# Default goal
++.DEFAULT_GOAL := usage
++
++# Target if user does not specify target
++usage:
++	$(error Unknown target='$(TARGET)': only 'cppnone', 'cppsse4', 'cppavx2', 'cpp512y', 'cpp512z' and 'cuda' are supported!)
++
+ # Enable ccache if USECCACHE=1
+ ifeq ($(USECCACHE)$(shell echo $(CXX) | grep ccache),1)
+   override CXX:=ccache $(CXX)
+@@ -48,10 +75,10 @@ CUDACPP_MAKEFILE=cudacpp.mk
+ # NB2 Use '|&' in CUDACPP_BUILDDIR to avoid confusing errors about googletest #507
+ # NB3 Do not add a comment inlined "CUDACPP_BUILDDIR=$(shell ...) # comment" as otherwise a trailing space is included...
+ # NB4 The variables relevant to the cudacpp Makefile must be explicitly passed to $(shell...)
+-CUDACPP_MAKEENV:=$(shell echo '$(.VARIABLES)' | tr " " "\n" | egrep "(USEBUILDDIR|AVX|FPTYPE|HELINL|HRDCOD)")
++CUDACPP_MAKEENV:=$(shell echo '$(.VARIABLES)' | tr " " "\n" | egrep "(USEBUILDDIR|FPTYPE|HELINL|HRDCOD)")
+ ###$(info CUDACPP_MAKEENV=$(CUDACPP_MAKEENV))
+ ###$(info $(foreach v,$(CUDACPP_MAKEENV),$(v)="$($(v))"))
+-CUDACPP_BUILDDIR:=$(shell $(MAKE) $(foreach v,$(CUDACPP_MAKEENV),$(v)="$($(v))") -f $(CUDACPP_MAKEFILE) -pn 2>/dev/null | awk '/Building/{print $$3}' | sed s/BUILDDIR=//)
++CUDACPP_BUILDDIR:=$(shell $(MAKE) $(foreach v,$(CUDACPP_MAKEENV),$(v)="$($(v))") -f $(CUDACPP_MAKEFILE) -pn $(TARGET) 2>/dev/null | awk '/Building/{print $$3}' | sed s/BUILDDIR=//)
+ ifeq ($(CUDACPP_BUILDDIR),)
+ $(error CUDACPP_BUILDDIR='$(CUDACPP_BUILDDIR)' should not be empty!)
+ else
+@@ -115,10 +142,6 @@ $(LIBS): .libs
+ 	cd ../../Source; make
+ 	touch $@
+
+-$(CUDACPP_BUILDDIR)/.cudacpplibs:
+-	$(MAKE) -f $(CUDACPP_MAKEFILE)
+-	touch $@
+-
+ # On Linux, set rpath to LIBDIR to make it unnecessary to use LD_LIBRARY_PATH
+ # Use relative paths with respect to the executables ($ORIGIN on Linux)
+ # On Darwin, building libraries with absolute paths in LIBDIR makes this unnecessary
+@@ -130,26 +153,47 @@ else
+   override LIBFLAGSRPATH = -Wl,-rpath,'$$ORIGIN/$(LIBDIR)'
+ endif
+
+-.PHONY: madevent_fortran_link madevent_cuda_link madevent_cpp_link
++.PHONY: madevent_fortran_link madevent_cuda_link madevent_cpp_link madevent_cppnone_link madevent_cppsse4_link madevent_cppavx2_link madevent_cpp512y_link madevent_cpp512z_link
+
+ madevent_fortran_link: $(PROG)_fortran
+ 	rm -f $(PROG)
+ 	ln -s $(PROG)_fortran $(PROG)
+
++# Only used for testing, should be removed in the future
+ madevent_cpp_link: $(CUDACPP_BUILDDIR)/$(PROG)_cpp
+ 	rm -f $(PROG)
+ 	ln -s $(CUDACPP_BUILDDIR)/$(PROG)_cpp $(PROG)
+
++madevent_cppnone_link: $(CUDACPP_BUILDDIR)/$(PROG)_cpp
++	rm -f $(PROG)
++	ln -s $(CUDACPP_BUILDDIR)/$(PROG)_cpp $(PROG)
++
++madevent_cppsse4_link: $(CUDACPP_BUILDDIR)/$(PROG)_cpp
++	rm -f $(PROG)
++	ln -s $(CUDACPP_BUILDDIR)/$(PROG)_cpp $(PROG)
++
++madevent_cppavx2_link: $(CUDACPP_BUILDDIR)/$(PROG)_cpp
++	rm -f $(PROG)
++	ln -s $(CUDACPP_BUILDDIR)/$(PROG)_cpp $(PROG)
++
++madevent_cpp512y_link: $(CUDACPP_BUILDDIR)/$(PROG)_cpp
++	rm -f $(PROG)
++	ln -s $(CUDACPP_BUILDDIR)/$(PROG)_cpp $(PROG)
++
++madevent_cpp512z_link: $(CUDACPP_BUILDDIR)/$(PROG)_cpp
++	rm -f $(PROG)
++	ln -s $(CUDACPP_BUILDDIR)/$(PROG)_cpp $(PROG)
++
+ madevent_cuda_link: $(CUDACPP_BUILDDIR)/$(PROG)_cuda
+ 	rm -f $(PROG)
+ 	ln -s $(CUDACPP_BUILDDIR)/$(PROG)_cuda $(PROG)
+
+ # Building $(PROG)_cpp also builds $(PROG)_cuda if $(CUDACPP_CULIB) exists (improved patch for cpp-only builds #503)
+-$(CUDACPP_BUILDDIR)/$(PROG)_cpp: $(PROCESS) $(DSIG_cudacpp) auto_dsig.o $(LIBS) $(MATRIX) counters.o ompnumthreads.o $(CUDACPP_BUILDDIR)/.cudacpplibs
++$(CUDACPP_BUILDDIR)/$(PROG)_cpp: $(PROCESS) $(DSIG_cudacpp) auto_dsig.o $(LIBS) $(MATRIX) counters.o ompnumthreads.o $(CPPTARGET)
+ 	$(FC) -o $(CUDACPP_BUILDDIR)/$(PROG)_cpp $(PROCESS) $(DSIG_cudacpp) auto_dsig.o $(MATRIX) $(LINKLIBS) $(BIASDEPENDENCIES) $(OMPFLAGS) counters.o ompnumthreads.o -L$(LIBDIR)/$(CUDACPP_BUILDDIR) -l$(CUDACPP_COMMONLIB) -l$(CUDACPP_CXXLIB) $(LIBFLAGSRPATH) $(LDFLAGS)
+-	if [ -f $(LIBDIR)/$(CUDACPP_BUILDDIR)/lib$(CUDACPP_CULIB).* ]; then $(FC) -o $(CUDACPP_BUILDDIR)/$(PROG)_cuda $(PROCESS) $(DSIG_cudacpp) auto_dsig.o $(MATRIX) $(LINKLIBS) $(BIASDEPENDENCIES) $(OMPFLAGS) counters.o ompnumthreads.o -L$(LIBDIR)/$(CUDACPP_BUILDDIR) -l$(CUDACPP_COMMONLIB) -l$(CUDACPP_CULIB) $(LIBFLAGSRPATH) $(LDFLAGS); fi
+
+-$(CUDACPP_BUILDDIR)/$(PROG)_cuda: $(CUDACPP_BUILDDIR)/$(PROG)_cpp
++$(CUDACPP_BUILDDIR)/$(PROG)_cuda: $(PROCESS) $(DSIG_cudacpp) auto_dsig.o $(LIBS) $(MATRIX) counters.o ompnumthreads.o cuda
++	$(FC) -o $(CUDACPP_BUILDDIR)/$(PROG)_cuda $(PROCESS) $(DSIG_cudacpp) auto_dsig.o $(MATRIX) $(LINKLIBS) $(BIASDEPENDENCIES) $(OMPFLAGS) counters.o ompnumthreads.o -L$(LIBDIR)/$(CUDACPP_BUILDDIR) -l$(CUDACPP_COMMONLIB) -l$(CUDACPP_CULIB) $(LIBFLAGSRPATH) $(LDFLAGS)
+
+ counters.o: counters.cc timer.h
+ 	$(CXX) $(CXXFLAGS) -c $< -o $@
+@@ -215,49 +259,105 @@ genps.o: .libs
+
+ UNAME_P := $(shell uname -p)
+ ifeq ($(UNAME_P),ppc64le)
+-avxall: avxnone avxsse4
++cppall:
++	$(MAKE) USEBUILDDIR=1 cppnone
++	$(MAKE) USEBUILDDIR=1 cppsse4
+ else ifeq ($(UNAME_P),arm)
+-avxall: avxnone avxsse4
++cppall:
++	$(MAKE) USEBUILDDIR=1 cppnone
++	$(MAKE) USEBUILDDIR=1 cppsse4
+ else
+-avxall: avxnone avxsse4 avxavx2 avx512y avx512z
++cppall:
++	$(MAKE) USEBUILDDIR=1 cppnone
++	$(MAKE) USEBUILDDIR=1 cppsse4
++	$(MAKE) USEBUILDDIR=1 cppavx2
++	$(MAKE) USEBUILDDIR=1 cpp512y
++	$(MAKE) USEBUILDDIR=1 cpp512z
+ endif
+
+-avxnone: $(PROG)_fortran $(DSIG_cudacpp)
+-	@echo
+-	$(MAKE) USEBUILDDIR=1 AVX=none
++ALL:
++	$(MAKE) USEBUILDDIR=1 cppnone
++	$(MAKE) USEBUILDDIR=1 cppsse4
++	$(MAKE) USEBUILDDIR=1 cppavx2
++	$(MAKE) USEBUILDDIR=1 cpp512y
++	$(MAKE) USEBUILDDIR=1 cpp512z
++	$(MAKE) USEBUILDDIR=1 cuda
+
+-avxsse4: $(PROG)_fortran $(DSIG_cudacpp)
+-	@echo
+-	$(MAKE) USEBUILDDIR=1 AVX=sse4
++cppnone:
++	$(MAKE) -f $(CUDACPP_MAKEFILE) cppnone
+
+-avxavx2: $(PROG)_fortran $(DSIG_cudacpp)
+-	@echo
+-	$(MAKE) USEBUILDDIR=1 AVX=avx2
++cppsse4:
++	$(MAKE) -f $(CUDACPP_MAKEFILE) cppsse4
+
+-avx512y: $(PROG)_fortran $(DSIG_cudacpp)
+-	@echo
+-	$(MAKE) USEBUILDDIR=1 AVX=512y
++cppavx2:
++	$(MAKE) -f $(CUDACPP_MAKEFILE) cppavx2
+
+-avx512z: $(PROG)_fortran $(DSIG_cudacpp)
+-	@echo
+-	$(MAKE) USEBUILDDIR=1 AVX=512z
++cpp512y:
++	$(MAKE) -f $(CUDACPP_MAKEFILE) cpp512y
+
+-###endif
++cpp512z:
++	$(MAKE) -f $(CUDACPP_MAKEFILE) cpp512z
++
++cuda:
++	$(MAKE) -f $(CUDACPP_MAKEFILE) cuda
+
+ # Clean (NB: 'make clean' in Source calls 'make clean' in all P*)
+
+-clean: # Clean builds: fortran in this Pn; cudacpp executables for one AVX in this Pn
+-	$(RM) *.o gensym $(PROG) $(PROG)_fortran $(PROG)_forhel $(CUDACPP_BUILDDIR)/$(PROG)_cpp $(CUDACPP_BUILDDIR)/$(PROG)_cuda
++BUILD_DIRS := $(wildcard build.*)
++BUILD_TARGET_DIR := $(shell echo $(BUILD_DIRS) | awk -F '[._]' '{print $$2}')
++NUM_BUILD_DIRS := $(words $(BUILD_DIRS))
++
++clean:
++ifeq ($(USEBUILDDIR),1)
++ifeq ($(NUM_BUILD_DIRS),1)
++	$(info USEBUILDDIR=1, only cleaning build dir.)
++	$(RM) -r $(BUILD_DIRS)
++	$(MAKE) -f $(CUDACPP_MAKEFILE) clean$(BUILD_TARGET_DIR)
++else ifeq ($(NUM_BUILD_DIRS),0)
++	$(error USEBUILDDIR=1, but no build directories are found.)
++else
++	$(error Multiple BUILDDIR's found! Use 'cleannone', 'cleansse4', 'cleanavx2', 'clean512y','clean512z', 'cleancuda' or 'cleanall'.)
++endif
++else
++	$(info USEBUILDDIR=0 only cleaning build with no dir.)
++	$(RM) *.o gensym $(PROG) $(PROG)_fortran $(PROG)_forhel ./$(PROG)_*
++	$(MAKE) -f $(CUDACPP_MAKEFILE) cleandir
++	$(RM) $(CUDACPP_BUILDDIR)/.*libs
++endif
++
++cleannone: # Clean builds: fortran in this Pn; cpp executables for one AVX in this Pn
++	$(RM) *.o gensym $(PROG) $(PROG)_fortran $(PROG)_forhel build.none_*/$(PROG)_cpp
++	$(MAKE) -f $(CUDACPP_MAKEFILE) cleannone
++
++cleansse4: # Clean build: fortran in this Pn; cpp executables for cppsse4 in this Pn
++	$(RM) *.o gensym $(PROG) $(PROG)_fortran $(PROG)_forhel build.sse4_*/$(PROG)_cpp
++	$(MAKE) -f $(CUDACPP_MAKEFILE) cleansse4
++
++cleanavx2: # Clean build: fortran in this Pn; cpp executables for cppavx2 in this Pn
++	$(RM) *.o gensym $(PROG) $(PROG)_fortran $(PROG)_forhel build.avx2_*/$(PROG)_cpp
++	$(MAKE) -f $(CUDACPP_MAKEFILE) cleanavx2
++
++clean512y: # Clean build: fortran in this Pn; cpp executables for cpp512y in this Pn
++	$(RM) *.o gensym $(PROG) $(PROG)_fortran $(PROG)_forhel build.512y_*/$(PROG)_cpp
++	$(MAKE) -f $(CUDACPP_MAKEFILE) clean512y
++
++clean512z: # Clean build: fortran in this Pn; cpp executables for cpp512z in this Pn
++	$(RM) *.o gensym $(PROG) $(PROG)_fortran $(PROG)_forhel build.512z_*/$(PROG)_cpp
++	$(MAKE) -f $(CUDACPP_MAKEFILE) clean512z
++
++cleancuda: # Clean build: fortran in this Pn; cuda executables in this Pn
++	$(RM) *.o gensym $(PROG) $(PROG)_fortran $(PROG)_forhel build.cuda_*/$(PROG)_cuda
++	$(MAKE) -f $(CUDACPP_MAKEFILE) cleancuda
+
+ cleanavxs: clean # Clean builds: fortran in this Pn; cudacpp for all AVX in this Pn and in src
+ 	$(MAKE) -f $(CUDACPP_MAKEFILE) cleanall
+-	rm -f $(CUDACPP_BUILDDIR)/.cudacpplibs
+-	rm -f .libs
++	$(RM) $(CUDACPP_BUILDDIR)/.*libs
++	$(RM) .libs
+
+ cleanall: # Clean builds: fortran in all P* and in Source; cudacpp for all AVX in all P* and in src
+ 	make -C ../../Source cleanall
+-	rm -rf $(LIBDIR)libbias.$(libext)
+-	rm -f ../../Source/*.mod ../../Source/*/*.mod
++	$(RM) -r $(LIBDIR)libbias.$(libext)
++	$(RM) ../../Source/*.mod ../../Source/*/*.mod
+
+ distclean: cleanall # Clean all fortran and cudacpp builds as well as the googletest installation
+ 	$(MAKE) -f $(CUDACPP_MAKEFILE) distclean