linking in library that contains reference to cuda kernel
include .h
into .cu
as well. nvcc is c++ compiler and mangles names.
compile as:
nvcc -c file.cu // compile cuda kernel
nvcc file.o main.cpp // compile and link
I would change you code as:
.hpp
#ifndef __reductions2d_H_
#define __reductions2d_H_
void getMean_wrapper(); // c++ linkga
#endif
.cu:
#include "...hpp"
#include <cuda.h>
#include <cuda_runtime.h>
__global__ void getMean(float *devDataPtr, size_t pitch, int rows, int cols)
{
which then you compile as
nvcc -c file.cu // compile cuda kernel
g++ -lcudart file.o main.cpp // no cuda stuff needed save for lib
Comments
-
Derek over 1 year
I am trying to figure out how to use a CUDA kernel as part of a library so that I can just add the library to my existing C++ source files, and be able to use the cuda kernel.
So how do you go about doing this? I tried to create a wrapper, like so:
.h file:
#ifndef __reductions2d_H_ #define __reductions2d_H_ #include <stdio.h> #include <cuda.h> #include <cuda_runtime.h> extern "C" void getMean_wrapper(); #endif
.cu
__global__ void getMean(float *devDataPtr, size_t pitch, int rows, int cols) { for (int r = 0; r < height; ++r) { float* row = (float*)((char*)devPtr + r * pitch); for (int c = 0; c < width; ++c) { printf("Row[%i][%i]: %4.3f \n",r,c row[c]); } } } void getMean_wrapper() { // Host code int width = 3, height = 3; int N = width*height; float* devData; size_t pitch; cudaMallocPitch(&devData, &pitch, width * sizeof(float),height); int blockSize = 4; int nBlocks = N/blockSize + (N%blockSize == 0?0:1); getMean<<<nBlocks, blockSize>>>(devData, pitch, width,height); }
main.cpp
#include "reductions2d.h" int main(void){ getMean_wrapper(); return 0; }
However, when I compile this with nvcc *.cpp, it tells me it cant find getMean_wrapper(), and when I try to just compile with g++ -c main.cpp, it tells me it cant find cuda.h and cuda_runtime.h
Is the best approach to specify the location of the cuda libraries with my G++ command line, build those objects, build the .cu objects, then link them? Seems like a hassle to have to have a 3 step process to add in some cuda functionality
Thanks
edit:
it seems like when I try to do it individually,t hen link wtih
g++ -o runme *.o -lcuda
I get
$ g++ -o runme *.o -lcuda reductions2d.o: In function `__sti____cudaRegisterAll_47_tmpxft_00007643_00000000_4_reductions2d_cpp1_ii_4ef 611a7()': tmpxft_00007643_00000000-1_reductions2d.cudafe1.cpp:(.text+0x15e): undefined reference to `__cudaRegisterFatBinary' tmpxft_00007643_00000000-1_reductions2d.cudafe1.cpp:(.text+0x1b9): undefined reference to `__cudaRegisterFunction' reductions2d.o: In function `__cudaUnregisterBinaryUtil()': tmpxft_00007643_00000000-1_reductions2d.cudafe1.cpp:(.text+0x1d8): undefined reference to `__cudaUnregisterFatBinary' reductions2d.o: In function `__device_stub__Z7getMeanPfmii(float*, unsigned long, int, int)': tmpxft_00007643_00000000-1_reductions2d.cudafe1.cpp:(.text+0x20d): undefined reference to `cudaSetupArgument' tmpxft_00007643_00000000-1_reductions2d.cudafe1.cpp:(.text+0x22f): undefined reference to `cudaSetupArgument' tmpxft_00007643_00000000-1_reductions2d.cudafe1.cpp:(.text+0x251): undefined reference to `cudaSetupArgument' tmpxft_00007643_00000000-1_reductions2d.cudafe1.cpp:(.text+0x273): undefined reference to `cudaSetupArgument' reductions2d.o: In function `getMean_wrapper': tmpxft_00007643_00000000-1_reductions2d.cudafe1.cpp:(.text+0x164c): undefined reference to `cudaConfigureCall' reductions2d.o: In function `cudaError cudaLaunch<char>(char*)': tmpxft_00007643_00000000-1_reductions2d.cudafe1.cpp:(.text._Z10cudaLaunchIcE9cudaErrorPT_[cudaError cudaLaunch<char>(char*)]+0x11): undefined reference to `cudaLaunch' reductions2d.o: In function `cudaError cudaMallocPitch<float>(float**, unsigned long*, unsigned long, unsigned long)': tmpxft_00007643_00000000-1_reductions2d.cudafe1.cpp:(.text._Z15cudaMallocPitchIfE9cudaErrorPPT_Pmmm[cudaError cudaMallocPitch<float>(float**, unsigned long*, unsigned long, unsigned long)]+0x29): undefined reference to `cudaMallocPitch'
I read that i need to include the cuda runtime libraries, so I did
ldconfig -p | grep cudart and included /usr/local/cuda/lib64 in my LD_LIBRARY_PATH and it still cant find cudart
-
Derek over 13 yearsthanks for the edited response - that did the trick! I just edited question with some linker errors I was getting by trying to use g++ to do the linking, and this saved me some time!