How to use a linker flag with cuRAND? And implementation of cuRAND

Question

CURAND Library - Compiling Error - Undefined reference to functions

I suppose my understanding of linker flags is less than substantial as I don't understand how to use them. Do I need to put the code in my makefile for the program I am running it with? Can someone please explain by writing out the syntax and where I would need to put it? Also, can I simply put the curand.lib file from the CUDA library in the program folder I want to use it with and reference it from there?

Here is the code I am attempting to use to generate a random number for a raytracer program. Any help is greatly appreciated.

float drandGPU()
{

  std::size_t n = 1;
  std::size_t i;
  curandGenerator_t gen;
  float *devData, *hostData;

  hostData = (float*) calloc(n, sizeof(float));
  cudaMalloc((void**) &devData, n*sizeof(float));
  curandCreateGenerator(&gen, CURAND_RNG_PSEUDO_MTGP32);
  curandSetPseudoRandomGeneratorSeed(gen, 1234ULL);
  curandGenerateUniform(gen,devData,n);
  cudaMemcpy(hostData, devData, n*sizeof(float), cudaMemcpyDeviceToHost);

  float f = hostData[0];
  return f;
}

EDIT: I am adding the makefile here:

all: exe

exe: pipe
    g++ -o rayTracer -D__PIKOC_HOST__ -I/usr/local/cuda/include -I../../api/include -I../.. -I. main.cpp -L/usr/local/cuda/lib -lcuda -lGL -lglut

pipe:
    ../../bin/pikoc --numRuns=10 --timer main.cpp

clean:
    rm -f rayTracer __pikoDefines.h __pikoCompiledPipe.h __pikoCompiledPipe.ptx

Answer 1

On the Linux command line, you can use a Makefile to easily build your code. Assuming that your file is named main.cpp, here is a Makefile that should compile your code:

CUDA_ROOT=/usr/local/cuda
NVCC=$(CUDA_ROOT)/bin/nvcc
NVCC_FLAGS=-ccbin g++ -m64
NVCC_ARCH_FLAGS=-gencode arch=compute_20,code=sm_20 -gencode arch=compute_30,code=sm_30
NVCC_IFLAGS=-I$(CUDA_ROOT)/include -I$(CUDA_ROOT)/samples/common/inc
LFLAGS=-lcurand

all: main.exe

main.o: main.cpp
    $(NVCC) $(NVCC_FLAGS) $(NVCC_IFLAGS) $(NVCC_ARCH_FLAGS) -c $< -o $@

main.exe: main.o
    $(NVCC) $(NVCC_FLAGS) $(NVCC_ARCH_FLAGS) $< -o $@ $(LFLAGS)

clean:
    rm -f *.o

clobber: clean
    rm -f main.exe

All you need to do is to put the snippet above in a file named Makefile in the same directory with your code and type make.

A brief summary of what is going on here is:

The top part of the Makefile that has all-uppercase variables like CUDA_ROOT and NVCC are variable definitions. These variables are used below in compilation and linking using the dollar sign such as $(NVCC).

Here, the variable CUDA_ROOT is especially critical and it may be different depending on the Linux/CUDA installation you have. Mine is actually /usr/local/cuda-6.5 but it is also symlinked to /usr/local/cuda which I am using in the Makefile. You may need to change this variable. If you have the nvcc command available, typing which nvcc should give you this path with /bin/nvcc directory added to it.

Typing make calls the first target rule, which is all: main.exe, and main.exe's rule needs main.o, so it goes to main.o's rule. Thus compiling of the object file, main.o is the first step. There is no linking there, just compiling the .cpp file into .o object file. Then, once the main.o is compiled, main.exe is linked with it and the cuRAND library (hence the LFLAGS variable is included in the main.exe recipe).

I also changed the code you pasted to get some output as:

#include <curand_kernel.h>
#include <helper_cuda.h>
#include <iostream>

float drandGPU()
{
  std::size_t n = 1;
  std::size_t i;
  curandGenerator_t gen;
  float *devData, *hostData;

  hostData = (float*) calloc(n, sizeof(float));
  cudaMalloc((void**) &devData, n*sizeof(float));
  curandCreateGenerator(&gen, CURAND_RNG_PSEUDO_MTGP32);
  curandSetPseudoRandomGeneratorSeed(gen, 1234ULL);
  curandGenerateUniform(gen,devData,n);
  cudaMemcpy(hostData, devData, n*sizeof(float), cudaMemcpyDeviceToHost);

  float f = hostData[0];
  return f;
}


int main()
{
  std::cout << drandGPU() << std::endl;
}

and named it main.cpp.