root/Makefile

# adapt to your installation

TOPDIR?=..

include $(TOPDIR)/Makefile.local

#link some uncommon libs static (boost png z)
STATIC=-Wl,-Bstatic

# only use -DTMTO_USE_COMPRESSED_PAIR with g++-4.1
GXX?=g++-4.1 # -DTMTO_USE_BOOST_COMPRESSED_PAIR

GXX_PPE=ppu-g++ -m64

# lines below should not have to be edited

#SDK_LIBS=-L$(SDK_DIR)/lib
#SDK_INC=-I$(SDK_DIR)/common/inc
CUDA_INC=-I$(CUDA_DIR)/include
STXXL_INC=-I$(STXXL_DIR)/include
LIBSPE2_INC=-I$(LIBSPE2_DIR) -I$(LIBSPE2_DIR)/spebase
STXXL_LIBS=-L$(STXXL_DIR)/lib
NVCC=$(CUDA_DIR)/bin/nvcc -keep --ptxas-options="-v --maxrregcount=128" $(NVCC_CFLAGS)

DIFF=diff --strip-trailing-cr

ifneq ($(BOOST_DIR),)
BOOST_INC=-I$(BOOST_DIR)
else
BOOST_INC=
endif

ifneq ($(BOOST_DIR),)
BOOST_LIBDIR=-L$(BOOST_DIR)/stage/lib
else
BOOST_LIBDIR=
endif

all: c

sinclude .depend

INC=-I$(TOPDIR) $(CUDA_INC) $(STXXL_INC) $(BOOST_INC) $(LIBSPE2_INC)
CFLAGS=$(M32) -DSTXXL_BOOST_THREADS -DSTXXL_BOOST_CONFIG -DSTXXL_BOOST_FILESYSTEM -DSTXXL_BOOST_RANDOM -DSTXXL_BOOST_TIMESTAMP

LIBS=$(STXXL_LIBS) -L$(LIBSPE2_DIR) -L$(CUDA_DIR)/lib -L$(CUDA_DIR)/lib64 -L$(CUDA_DIR)/lib $(BOOST_LIBDIR) $(STATIC) -lboost_program_options$(BOOST_INFIX)-mt -lstxxl -lboost_regex$(BOOST_INFIX)-mt -lboost_thread$(BOOST_INFIX)-mt -lboost_system$(BOOST_INFIX)-mt -lboost_filesystem$(BOOST_INFIX)-mt -lboost_iostreams$(BOOST_INFIX)-mt -Wl,-Bdynamic -lgcc_s -lpthread

.PHONY: calculate_chain_cuda.E test

clean:
        PATH=.:$$PATH $(NVCC) -clean -keep -c $(INC) -o calculate_chain_cuda.o $(TOPDIR)/calculate_chain_cuda.cu
        PATH=.:$$PATH $(NVCC) -clean -keep -c $(INC) -o cuda_localmem_kernel.o $(TOPDIR)/plugins/A51/cuda_localmem_kernel.cu
        PATH=.:$$PATH $(NVCC) -clean -keep -c $(INC) -o cuda_bitslice_kernel.o $(TOPDIR)/plugins/A51/cuda_bitslice_kernel.cu
        rm c *.o refa51 c_cuda c_sse

calculate_chain_methods.cu:
        echo

nvcc_gcc_script: gcc
        (echo "#!/bin/sh"; echo "exec $(GXX) \"\$$@\"") > gcc
        chmod 755 gcc

calculate_chain_cuda.o: $(TOPDIR)/calculate_chain_cuda.cu
        PATH=.:$$PATH $(NVCC) -c -keep $(CFLAGS) $(INC) -o $@ $<

calculate_chain_cuda_bitsliced: $(TOPDIR)/calculate_chain_cuda_bitsliced.cu
        (echo "#!/bin/sh"; echo "exec $(GXX) \"\$$@\"") > gcc
        chmod 755 gcc
        PATH=.:$$PATH $(NVCC) -D_DEBUG -I/home/sascha/NVIDIA_CUDA_SDK/common/inc -keep $(CFLAGS) $(INC) -o $@ $<

calculate_chain_cuda_bitslice.E: $(TOPDIR)/calculate_chain_cuda_bitsliced.cu
        (echo "#!/bin/sh"; echo "exec $(GXX) \"\$$@\"") > gcc
        chmod 755 gcc
        PATH=.:$$PATH $(NVCC) -E -keep $(CFLAGS) $(INC) $<

calculate_chain.o: $(TOPDIR)/calculate_chain.cpp
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

calculate_chain_ppe.o: $(TOPDIR)/calculate_chain.cpp
        $(GXX_PPE) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

calculate_chain_spe.o: $(TOPDIR)/calculate_chain_spe.cpp
        spu-g++ -c $(HOST_CFLAGS) $(CFLAGS) -save-temps $(INC) -o $@ -c $<

calculate_chain_spe.elf: calculate_chain_spe.o
        spu-g++ $(HOST_CFLAGS) $(CFLAGS) -o $@ $<

calculate_chain_spe_embed.o: calculate_chain_spe.elf
        ppu-embedspu spe_program $< $@

refa51: $(TOPDIR)/reference/a51.cpp
        $(GXX) -O3 -o $@ $^

readahead: $(TOPDIR)/tools/readahead.cpp
        g++ -o $@ $^ -lpthread

main_plugin.o: $(TOPDIR)/plugins/A51/main.cpp
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

work_generators_plugin.o: $(TOPDIR)/plugins/A51/work_generators.cpp
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

work_consumers_plugin.o: $(TOPDIR)/plugins/A51/work_consumers.cpp
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

work_intermediates_plugin.o: $(TOPDIR)/plugins/A51/work_intermediates.cpp
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

main_plugin_ppe.o: $(TOPDIR)/plugins/A51/main.cpp
        $(GXX_PPE) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

spe_bitslice_plugin.o: $(TOPDIR)/plugins/A51/spe_bitslice.cpp
        $(GXX_PPE) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

sse_bitslice_plugin.o: $(TOPDIR)/plugins/A51/sse_bitslice.cpp
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

sse_bitslice_kernel.o: $(TOPDIR)/plugins/A51/sse_bitslice_kernel.cpp
        $(GXX) $(SSE_KERNEL_CFLAGS) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

cuda_localmem_plugin.o: $(TOPDIR)/plugins/A51/cuda_localmem.cpp
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

cuda_bitslice_plugin.o: $(TOPDIR)/plugins/A51/cuda_bitslice.cpp
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

cuda_bitslice_kernel.o: $(TOPDIR)/plugins/A51/cuda_bitslice_kernel.cu
        (echo "#!/bin/sh"; echo "exec $(GXX) \"\$$@\"") > gcc
        chmod 755 gcc
        PATH=.:$$PATH $(NVCC) $(INC) $(CFLAGS) -c -o $@ $< 2>&1 | grep -v 'Warning: Cannot tell what pointer points to'

cuda_bitslice_kernel.E: $(TOPDIR)/plugins/A51/cuda_bitslice_kernel.cu
        (echo "#!/bin/sh"; echo "exec $(GXX) \"\$$@\"") > gcc
        chmod 755 gcc
        PATH=.:$$PATH $(NVCC) $(INC) $(CFLAGS) -E $< 2>&1 | grep -v 'Warning: Cannot tell what pointer points to'

cuda_localmem_kernel.o: $(TOPDIR)/plugins/A51/cuda_localmem_kernel.cu
        (echo "#!/bin/sh"; echo "exec $(GXX) \"\$$@\"") > gcc
        chmod 755 gcc
        PATH=.:$$PATH $(NVCC) $(INC) $(CFLAGS) -c -o $@ $<

cuda_device_plugin.o: $(TOPDIR)/plugins/A51/cuda.cpp
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) $(INC) -c -o $@ $<

depend:
        $(GXX) $(INC) -M -MT calculate_chain.o $(TOPDIR)/calculate_chain.cpp > .depend
        $(GXX_PPE) $(INC) -M -MT calculate_chain_ppe.o $(TOPDIR)/calculate_chain.cpp >> .depend
        $(GXX) $(INC) -M -MT main_plugin.o $(TOPDIR)/plugins/A51/main.cpp >> .depend
        $(GXX) $(INC) -M -MT work_generators_plugin.o $(TOPDIR)/plugins/A51/work_generators.cpp >> .depend
        $(GXX) $(INC) -M -MT work_consumers_plugin.o $(TOPDIR)/plugins/A51/work_consumers.cpp >> .depend
        $(GXX) $(INC) -M -MT work_intermediates_plugin.o $(TOPDIR)/plugins/A51/work_intermediates.cpp >> .depend
        $(GXX_PPE) $(INC) -M -MT main_plugin_ppe.o $(TOPDIR)/plugins/A51/main.cpp >> .depend
        $(GXX) $(INC) -M -MT cuda_localmem_plugin.o $(TOPDIR)/plugins/A51/cuda_localmem.cpp >> .depend
        $(GXX) $(INC) -M -MT cuda_bitslice_plugin.o $(TOPDIR)/plugins/A51/cuda_bitslice.cpp >> .depend
        $(GXX) $(INC) -M -MT cuda_device_plugin.o $(TOPDIR)/plugins/A51/cuda.cpp >> .depend
        $(GXX) $(INC) -M -MT sse_bitslice_plugin.o $(TOPDIR)/plugins/A51/sse_bitslice.cpp >> .depend
        $(GXX) $(INC) -M -MT spe_bitslice_plugin.o $(TOPDIR)/plugins/A51/spe_bitslice.cpp >> .depend
        $(GXX) $(INC) -M -MT sse_bitslice_kernel.o $(TOPDIR)/plugins/A51/sse_bitslice_kernel.cpp >> .depend
        for k in cuda_bitslice_kernel cuda_localmem_kernel; do \
                $(GXX) -x c++ $(INC) -M -MT $$k.o $(TOPDIR)/plugins/A51/$$k.cu >> .depend; done
        spu-g++ $(INC) -M $(TOPDIR)/calculate_chain_spe.cpp >> .depend

c_ppe: calculate_chain_ppe.o main_plugin_ppe.o calculate_chain_spe_embed.o spe_bitslice_plugin.o
        $(GXX_PPE) $(HOST_CFLAGS) $(CFLAGS) -o $@ $^ $(LIBS) -lgcc_s -lstxxl -lspe2 

c_sse: calculate_chain.o main_plugin.o work_generators_plugin.o work_consumers_plugin.o work_intermediates_plugin.o sse_bitslice_plugin.o sse_bitslice_kernel.o
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) -o $@ $^ $(LIBS) -lgcc_s -lstxxl

c_cuda: calculate_chain.o main_plugin.o cuda_device_plugin.o cuda_bitslice_plugin.o cuda_bitslice_kernel.o work_generators_plugin.o work_consumers_plugin.o work_intermediates_plugin.o cuda_localmem_plugin.o cuda_localmem_kernel.o
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) -o $@ $^ $(LIBS) -lcudart -lgcc_s -lstxxl

c_raw: calculate_chain.o main_plugin.o
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) -o $@ $^ $(LIBS) -lgcc_s -lstxxl

c: calculate_chain.o main_plugin.o cuda_device_plugin.o cuda_bitslice_plugin.o cuda_bitslice_kernel.o sse_bitslice_plugin.o sse_bitslice_kernel.o work_generators_plugin.o work_consumers_plugin.o work_intermediates_plugin.o
        $(GXX) $(HOST_CFLAGS) $(CFLAGS) -o $@ $^ $(LIBS) -lcudart -lgcc_s -lstxxl

# not used or for debugging

calculate_chain.E: $(TOPDIR)/calculate_chain.cpp
        $(GXX) $(INC) -E $(TOPDIR)/calculate_chain.cpp

source_files_list:
        @find $(TOPDIR)/ -type f ! -name '*~' ! -path '$(TOPDIR)/doc/*' ! -path '*/.svn/*' ! -path '$(TOPDIR)/fpga/*' ! -path $(TOPDIR)/obj/\*

test:
        ./c --nodefault --operations 512 --algorithm A51 --condition rounds:rounds=32 --implementation sharedmem --device cuda --roundfunc xor:condition=distinguished_point::bits=15:generator=lfsr::tablesize=32::advance=0::force --work increment --consume print --logger verbose generate --chains 1024 --chainlength 3072 > testdata
        @echo -n +++ Test " "; if $(DIFF) testdata $(TOPDIR)/data/testdata; then echo -n OK; else echo -n FAILED; fi; echo " +++"

test2:
        ./c --nodefault --algorithm A51 --condition rounds:rounds=32 --implementation sharedmem --device cuda --roundfunc xor:condition=distinguished_point::bits=15:generator=lfsr::tablesize=32::advance=0::force --work increment --consume print --logger verbose generate --chains 1024 --chainlength 3072000 > testdata2
        @echo -n +++ Test " "; if $(DIFF) testdata2 $(TOPDIR)/data/testdata2; then echo -n OK; else echo -n FAILED; fi; echo " +++"

test3:
        ./c --nodefault --algorithm A51 --condition rounds:rounds=32 --implementation sharedmem --device cuda --roundfunc xor:condition=distinguished_point::bits=15:generator=lfsr::tablesize=32::advance=0::force --work increment --consume print --logger normal generate --chains 1024 --chainlength 3072000 --intermediate filter:runlength=512 > testdata3
        @echo -n +++ Test " "; if $(DIFF) testdata3 $(TOPDIR)/data/testdata3; then echo -n OK; else echo -n FAILED; fi; echo " +++"

test4:
        ./c --nodefault --operations 512 --algorithm A51 --condition rounds:rounds=32 --implementation sharedmem --device cuda --roundfunc xor:condition=distinguished_point::bits=15:generator=lfsr2::tablesize=32::advance=0 --work increment --consume print --logger verbose generate --chains 16 --chainlength 3072 > testdata4
        @echo -n +++ Test " "; if $(DIFF) testdata4 $(TOPDIR)/data/testdata4; then echo -n OK; else echo -n FAILED; fi; echo " +++"

test5:
        ./c --nodefault --algorithm A51 --condition rounds:rounds=32 --implementation sharedmem --device cuda --roundfunc xor:condition=distinguished_point::bits=15:generator=lfsr2::tablesize=32::advance=0 --work increment --consume print --logger verbose generate --chains 16 --chainlength 3072000 > testdata5
        @echo -n +++ Test " "; if $(DIFF) testdata5 $(TOPDIR)/data/testdata5; then echo -n OK; else echo -n FAILED; fi; echo " +++"

test6:
        ./c --nodefault --algorithm A51 --condition rounds:rounds=32 --implementation sharedmem --device cuda --roundfunc xor:condition=distinguished_point::bits=15:generator=lfsr2::tablesize=32::advance=0 --work increment --consume print --logger normal generate --chains 16 --chainlength 3072000 --intermediate filter:runlength=512 > testdata6
        @echo -n +++ Test " "; if $(DIFF) testdata6 $(TOPDIR)/data/testdata6; then echo -n OK; else echo -n FAILED; fi; echo " +++"

test23: refa51
        ./c_cuda --work sort:source=random::limit=1K:ram=1 --icondition distinguished_point:bits=15 --condition rounds:rounds=2 --advance 2048 --device cuda:implementation=bitslice:blocks=4:threads=128 --consume print:all --operations 512 generate --chainlength 1000000 --chains 8192 --intermediate sort:parts=1:ram=300 2>&1 | grep results | awk '{ print $$3 " " $$7 }' | while read s e; do echo -n "$$e "; ./refa51 -c $$e -o 1000000 $$s 15 2 2048; done

test24: refa51
        ./c_cuda --work sort:source=random::limit=1K:ram=1 --icondition distinguished_point:bits=15 --condition rounds:rounds=2 --advance 2048 --device cuda:implementation=sharedmem:blocks=4:threads=256 --consume print:all --operations 512 generate --chainlength 1000000 --chains 8192 --intermediate sort:parts=1:ram=300 2>&1 | grep results | awk '{ print $$3 " " $$7 }' | while read s e; do echo -n "$$e "; ./refa51 -c $$e -o 1000000 $$s 15 2 2048; done