Help making a simple Makefile

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

I just want to make a simple makefile for a CUDA project that I am working on. I’ve read some of the documentation for nvcc but I’m still confused about what order everything needs to be in and what all of the suffixes mean.

For example, suppose I have some files: main.cu, header.h, compare.c. Could I make a simple makefile like this:

Project: main.cu.o compare.o
cc main.cu.o compare.o -o Project
main.cu.o: header.h main.cu.c
gcc -c main.cu.c
main.cu.c: header.h main.cu.c
nvcc main.cu -cuda
compare.o: header.h compare.c
cc -c compare.c

I know that this is probably wrong, but any explanations of why would be greatly appreciated.

2

I’m also fairly new to this, but what I’m currently doing is keeping the kernels separate from the rest of my application, compiling them to PTX code, and using the CUDA Driver API to access them, i.e., something like:

[codebox] struct stat file_info;

stat (name.c_str (), &file_info);

char * file_buffer = new char[file_info.st_size+1];

int fd = open (name.c_str (), O_RDONLY);

read (fd, file_buffer, file_info.st_size);

close (fd);

file_buffer[file_info.st_size] = 0;

CUjit_option options[1] = { CU_JIT_MAX_REGISTERS };

void * optionValues[1] = { &max_registers };

cuModuleLoadDataEx (&cuModule, file_buffer, 1, options, optionValues);

[/codebox]

where “name” is the PTX file I’m loading.

Here is the Makefile I am using with my little test project - maybe it will answer some questions:

[codebox]CC= /Developer/usr/bin/llvm-gcc-4.2

CPP= /Developer/usr/bin/llvm-cpp-4.2

CXX= /Developer/usr/bin/llvm-g+±4.2

NVCC= /usr/local/cuda/bin/nvcc

INCLUDES= -I. -I/usr/local/include -I/usr/local/cuda/include -framework OpenGL

NVCCFLAGS= -m32 --ptx --compiler-options -fno-strict-aliasing

QUADRO=

#QUADRO= -arch sm_13

CXXFLAGS= -m32 -fno-strict-aliasing -O3

CFLAGS= -m32 -fno-strict-aliasing -O3

LDFLAGS= -L/usr/local/lib -L/usr/local/cuda/lib -fPIC

LIBS= -lcuda

RTLIBS= -lcudart -lcuda

#LIBS= -lcufft -lcublas

all: dQ dQ.ptx CASS.ptx

clean:

rm *.cubin *.ptx

dQ: dQ.cpp CUDA_Q.cpp CUDA_Q.h CUDA_Q_private.h

$(CXX) -o dQ dQ.cpp CUDA_Q.cpp $(INCLUDES) $(CXXFLAGS) $(LDFLAGS) $(RTLIBS)

dQ.ptx: dQ.cu

$(NVCC) -o dQ.ptx dQ.cu $(QUADRO) $(NVCCFLAGS)

CASS.ptx: CASS.cu

$(NVCC) -o CASS.ptx CASS.cu $(QUADRO) $(NVCCFLAGS)

info: dQ.ptx CASS.ptx

ptxas -o dQ.cubin dQ.ptx -v -mem

ptxas -o CASS.cubin CASS.ptx -v -mem --maxrregcount=16

[/codebox]