SHELL = csh
##################### intel compiler #############
CC      = icc
FC      = ifort 
CFLAGS  = -w -O3 -ip  -c
FCFLAGS =  -align all -FR -convert big_endian -fno-alias -fno-fnalias
OPT     =  -O0 -g -C -traceback
OPT     =  -O3
LD      = $(FC)
LDOPT   =  -g
XXX = 32
YYY = 8
##################### gcc/gfortran ###############
#CC      = gcc
#FC      = gfortran 
#FCFLAGS = -fconvert=big-endian -frecord-marker=4 -ffree-form
#OPT     =  -O3
#LD      = $(FC)
#LDOPT   = 
##################################################
# do not change this definition, change the one further down
FLOAT   = float
RWORDSIZE = 4
##################### cuda location ##############
# eces-shell
#CUDALIBPATH =  ~/emu/cuda/lib
# ncsa
CUDALIBPATH =  /usr/local/cuda/lib
LIBCUBLAS = $(CUDALIBPATH)/libcublas.so
LIBCUBLASEMU = $(CUDALIBPATH)/libcublasemu.so


########################  THIS SECTION YOU CAN CHANGE ##################
#
# Hard coded number of levels  (35 for conus, 28 for jan00)
MKX = 28

# uncomment this to use FLOAT4 data type (optimization)
#FLOAT4 = -DFLOAT_4=4 

# this must always be defined but they do not do anything unless
# other settings are set
DEBUG_I = 59
DEBUG_J = 45
DEBUG_K = 1
DEBUGOUTPUT = -DDEBUG_I=$(DEBUG_I) -DDEBUG_J=$(DEBUG_J) -DDEBUG_K=$(DEBUG_K) $(FLOAT4)

# uncomment to run on emulator instead of the device
#DEVICEEMU       = -DDEVICEEMU
#DEVICEEMU_NVCC  = -deviceemu $(DEVICEEMU)
#LIBCUBLAS = $(LIBCUBLASEMU)

# uncomment to output detailed debug data output
# must have DEVICEEMU settings uncommented above
#DEBUGOUTPUT = -DDEBUGOUTPUT -DDEBUG_I=$(DEBUG_I) -DDEBUG_J=$(DEBUG_J) -DDEBUG_K=$(DEBUG_K) $(FLOAT4)

# uncomment to allow settings to force closer agreement
#DEBUGDEBUG = -DDEBUGDEBUG

# uncomment for to promote to 8 byte floats
# note, if you do this without DEVICE EMU above, compiler will complain it does not have enough shared mem
#PROMOTE = -DPROMOTE
#FLOAT = double
#FCFLAGS = -fconvert=big-endian -frecord-marker=4 -fdefault-real-8
#RWORDSIZE = 8
########################  END OF SECTION YOU CAN CHANGE ################

#NVCC    = nvcc -DCUDA
NVCC    = nvcc -DCUDA -ccbin /usr/bin #--ptxas-options=-v
PHASE   = -ptx
PHASE   = -cuda
NVOPT   = $(DEVICEEMU_NVCC) $(PROMOTE) $(DEBUGDEBUG) $(DEBUGOUTPUT) -DXXX=$(XXX) -DYYY=$(YYY) -DMKX=$(MKX) --use_fast_math

.SUFFIXES :

#all : vanilla chocolate compare_snaps
all : vanilla compare_snaps

swrad.cu.o : swrad.cu
	m4 swrad.cu | sed "s/float/$(FLOAT)/g" > y.cu
	$(NVCC) $(PHASE) $(NVOPT) y.cu
	/bin/mv y.cu.cpp swrad.cu.cpp
	$(CC) $(CFLAGS) -c swrad.cu.cpp

swrad_gpu.cu.o : swrad_gpu.cu spt.h
	m4 swrad_gpu.cu | spt.pl | sed "s/float/$(FLOAT)/g" > x.cu
	$(NVCC) $(PHASE) $(NVOPT) x.cu
	/bin/mv x.cu.cpp swrad_gpu.cu.cpp
	$(CC) $(CFLAGS) -c swrad_gpu.cu.cpp

microclock.o : microclock.c
	$(CC) -c $(CFLAGS) -DMKX=$(MKX) microclock.c

module_ra_sw.o : module_ra_sw.F
	/lib/cpp -C -P $(DEBUGDEBUG) $(DEVICEEMU) $(DEBUGOUTPUT) -DRWORDSIZE=$(RWORDSIZE) module_ra_sw.F > module_ra_sw.f90
	$(FC) -c $(OPT) $(FCFLAGS) module_ra_sw.f90

driver_swrad.o : driver_swrad.F module_ra_sw.o
	/lib/cpp -C -P $(DEBUGDEBUG) $(DEVICEEMU) $(DEBUGOUTPUT) -DRWORDSIZE=$(RWORDSIZE) driver_swrad.F > driver_swrad.f90
	$(FC) -c $(OPT) $(FCFLAGS) driver_swrad.f90

vanilla : module_ra_sw.o microclock.o driver_swrad.o
	$(LD) -o swrad_driver_vanilla $(LDOPT) driver_swrad.o module_ra_sw.o microclock.o $(LIBCUBLASEMU)

chocolate : module_ra_sw.F swrad.cu.o swrad_gpu.cu.o driver_swrad.o microclock.o
	/lib/cpp -C -P $(DEBUGDEBUG) $(DEVICEEMU) $(DEBUGOUTPUT) -DRUN_ON_GPU -DRWORDSIZE=$(RWORDSIZE) module_ra_sw.F > module_ra_sw.f90
	$(FC) -c $(OPT) $(FCFLAGS) module_ra_sw.f90
	$(LD) -o swrad_driver_chocolate $(LDOPT) driver_swrad.o module_ra_sw.o swrad.cu.o swrad_gpu.cu.o microclock.o $(LIBCUBLAS) 

compare_snaps : compare_snaps.F
	\cp compare_snaps.F compare_snaps.f90
	$(FC) -o compare_snaps $(FCFLAGS) compare_snaps.f90
	\rm compare_snaps.f90

clean clena :
	\rm -f *.o *.cu.c x.cu y.cu swrad_driver_* *.mod *.f90 x.ptx