makefile

C++编写的高性能矩阵乘法的Stranssen算法

# Makefile for Strassen code

PRISM_BLAS_LIB=../../GotoBLAS/libgoto.a
PRISM_MPI_ARCH=P4

ARCH = $(PRISM_MPI_ARCH)

include ../make_include/Make-$(ARCH).include

CFLAGS = $(OPTFLAGS) $(PRISM_DEFINES) -I../include -DSTRAS_NON_SQ 

# choices for CFLAGS above for export versions:
# -DSTRAS_RANDOM -DSTRAS_NON_SQ -DSTRAS_ESSL -DSTRAS_CRAY  

# additional choices for CFLAGS above for CCS testing versions:
# -DSTRAS_PRISM_TIME -DSTRAS_PLOT   -DPRISM_MEM_CHK  
# -DSTRAS_TIME_PARTS -DSTRAS_DGEMMW -DINFILE

# choices to be used as PRISM_DEFINES in Make-$(ARCH).include:
# -DSTRAS_FORTRAN_ADD    -DSTRAS_FORTRAN_SUB
# _DSTRAS_ESSL_ADD       -DSTRAS_ESSL_SUB 
# -DSTRAS_BLAS_ADD       -DSTRAS_BLAS_SUB 
# -DPRISM_SP -DPRISM_SUN -DPRISM_CRAY -DPRISM_T3D

FFLAGS = $(FOPTFLAGS)

PROG = timefmm
EXECS = $(PROG)_$(ARCH)
LIB_PATH = ../lib
LIB = $(LIB_PATH)/stras_lib_$(ARCH).a

#default: stras_testing stras_lib
default: stras_lib

all: default

testing: stras_testing

.SUFFIXES: .f .c .o$(U_ARCH)

# Use the following to incorporate Gropp's TR_TOOLS - NOT FOR EXPORT
#OBJS_FMM = fmm_internal.o$(U_ARCH) fmm.o$(U_ARCH) DGEFMM.o$(U_ARCH) DGEFMM_MEM.o$(U_ARCH) gen_matrix.o$(U_ARCH) matrix.o$(U_ARCH) strassen_internal.o$(U_ARCH) matrix_arith.o$(U_ARCH) fixup_internal.o$(U_ARCH) timefmm.o$(U_ARCH) tr.o$(U_ARCH) $(OBJS_MACHINE) $(OBJS_CLOCK)

OBJS_FMM = fmm_internal.o$(U_ARCH) fmm.o$(U_ARCH) DGEFMM.o$(U_ARCH) DGEFMM_MEM.o$(U_ARCH) gen_matrix.o$(U_ARCH) matrix.o$(U_ARCH) strassen_internal.o$(U_ARCH) matrix_arith.o$(U_ARCH) fixup_internal.o$(U_ARCH) timefmm.o$(U_ARCH) $(OBJS_MACHINE) $(OBJS_CLOCK) 

OBJS_LIB = fmm_internal.o$(U_ARCH) lib_fmm.o$(U_ARCH) DGEFMM.o$(U_ARCH) DGEFMM_MEM.o$(U_ARCH) matrix.o$(U_ARCH) matrix_arith.o$(U_ARCH) fixup_internal.o$(U_ARCH) strassen_internal.o$(U_ARCH) $(OBJS_MACHINE)

LIBS = $(BLAS_LIB) -lm 

stras_testing:	$(OBJS_FMM)
		$(CFLINKER) $(LFLAGS) -o $(EXECS) $(OBJS_FMM) $(LIBS)

stras_lib:	$(LIB_PATH) $(OBJS_LIB) lib_fmm.o$(U_ARCH)
		$(AR) $(LIB) $(OBJS_LIB) 
		$(RANLIB) $(LIB)

# default rules
.c.o$(U_ARCH):
	$(CC) $(CFLAGS) -c $?
	if [ $(ARCH) != cray_t3d ] ; then \
          mv $*.o $@ ; \
	fi

matrix_arith.o$(U_ARCH):	matrix_arith.F
#	$(CPP) -P $(CPP_FLAGS) matrix_arith.F > matrix_arith.f
	$(F77) $(FFLAGS) -c matrix_arith.F
	if [ $(ARCH) != cray_t3d ] ; then \
          mv matrix_arith.o $@ ; \
	fi
	/bin/rm -f matrix_arith.f

# Need to take care of fmm.c separately for library version since we need 
# to compile in cutoff = STRAS_CUTOFF here
lib_fmm.o$(U_ARCH):	fmm.c
			cp fmm.c lib_fmm.c
			$(CC) $(CFLAGS) -DSTRAS_LIB -c lib_fmm.c
			if [ $(ARCH) != cray_t3d ] ; then \
		          mv lib_fmm.o lib_fmm.o$(U_ARCH); \
			fi
			/bin/rm -f lib_fmm.c

tr.o$(U_ARCH):	tr.c
# Won't compile with xlc on RS6000
	cc $(CFLAGS) -c tr.c
	if [ $(ARCH) != cray_t3d ] ; then \
          mv $*.o $@ ; \
	fi
	
$(LIB_PATH):
	mkdir -p $(LIB_PATH)

cleanobjs:	
	/bin/rm -f *.o$(U_ARCH)

clean:	
	/bin/rm -f $(EXECS) *.o *.o$(U_ARCH) *~