C++ CUDA 和 libpng 链接错误

Question

已解决:请看下面我的评论

我一直在研究 GPU 分形生成器，并一直在使用 CUDA 访问 GPU，并使用 libpng 生成输出图像。我正在运行 OSX 10.7.4、Cuda 版本 4.2、V0.2.1221、libpng15(我认为是 1.5.11)

第一个版本(真的很乱很乱)代码只有一个源文件， GPUkernel.cu，它具有使用 CUDA 库计算分形的功能，还可以使用 libpng 生成 png。使用命令可以很好地编译

all:    src/GPUkernel.cu
        nvcc -o base src/GPUkernel.cu -lpng15 -lz

生成可执行库，一切运行良好。

新版本:

现在，为了在项目变大时尝试使代码更易于管理，我将源代码分成两个文件， GPU内核.cu和 GPU分形.cpp. GPU内核.cu拥有所有 CUDA GPU 功能，并且 GPU分形.cpp包含 main() 和 libpng 函数。

当我尝试编译此代码时遇到一些问题，以处理 libpng 和 CUDA 使用的不同架构(我认为)。我的生成文件是:

CUDA_INSTALL_PATH ?= /Developer/GPU\ Computing/CUDALibraries/common

ARCH=-arch i386 -arch x86_64

CXX := g++
CC := gcc
LINK := g++ -fPIC $(ARCH)
NVCC := nvcc -ccbin /usr/bin

# Includes
INCLUDES = -I. -I$(CUDA_INSTALL_PATH)/inc -I/usr/local/cuda/include 

# Common flags
COMMONFLAGS += $(INCLUDES)
NVCCFLAGS += $(COMMONFLAGS)
CXXFLAGS += $(COMMONFLAGS)
CFLAGS += $(COMMONFLAGS)

LIB_CUDA := -L$(CUDA_INSTALL_PATH)/lib `libpng-config --cflags --ldflags`

OBJS = GPUkernel.cu.o GPUfractal.cpp.o
TARGET = base
LINKLINE = $(LINK) -o $(TARGET) $(OBJS) $(LIB_CUDA)

.SUFFIXES: .c .cpp .cu .o

%.c.o: %.c
    $(CC) $(CFLAGS) -c $< -o $@

%.cu.o: %.cu
    $(NVCC) $(NVCCFLAGS) -c $< -o $@

%.cpp.o: %.cpp
    $(CXX) $(CXXFLAGS) -c $< -o $@

$(TARGET): $(OBJS) Makefile
    $(LINKLINE)

这是我的终端输出:

pikachu:GPUfractal tom$ make
nvcc -ccbin /usr/bin -I. -I/Developer/GPU\ Computing/CUDALibraries/common/inc -I/usr/local/cuda/include  -c GPUkernel.cu -o GPUkernel.cu.o
ptxas /tmp/tmpxft_00000847_00000000-2_GPUkernel.ptx, line 105; warning : Double is not supported. Demoting to float
g++ -I. -I/Developer/GPU\ Computing/CUDALibraries/common/inc -I/usr/local/cuda/include  -c GPUfractal.cpp -o GPUfractal.cpp.o
GPUfractal.cpp: In function ‘int renderImage(int*, float*, int, int, std::string, bool, int)’:
GPUfractal.cpp:180: warning: deprecated conversion from string constant to ‘char*’
GPUfractal.cpp:232: warning: deprecated conversion from string constant to ‘char*’
g++ -fPIC -arch i386 -arch x86_64 -o base GPUkernel.cu.o GPUfractal.cpp.o -L/Developer/GPU\ Computing/CUDALibraries/common/lib `libpng-config --cflags --ldflags`
ld: warning: ignoring file GPUfractal.cpp.o, file was built for unsupported file format which is not the architecture being linked (i386)
Undefined symbols for architecture i386:
  "_main", referenced from:
      start in crt1.10.6.o
  "_cudaGetLastError", referenced from:
      calculateFractal(int*, float*, int, int, float, float, float, float, int, bool, float, float)in GPUkernel.cu.o
  "_cudaGetErrorString", referenced from:
      calculateFractal(int*, float*, int, int, float, float, float, float, int, bool, float, float)in GPUkernel.cu.o
  "_cudaMemset", referenced from:
      calculateFractal(int*, float*, int, int, float, float, float, float, int, bool, float, float)in GPUkernel.cu.o
  "_cudaConfigureCall", referenced from:
      calculateFractal(int*, float*, int, int, float, float, float, float, int, bool, float, float)in GPUkernel.cu.o
  "_cudaMemcpy", referenced from:
      calculateFractal(int*, float*, int, int, float, float, float, float, int, bool, float, float)in GPUkernel.cu.o
  "_cudaFree", referenced from:
      calculateFractal(int*, float*, int, int, float, float, float, float, int, bool, float, float)in GPUkernel.cu.o
  "___cudaRegisterFatBinary", referenced from:
      __sti____cudaRegisterAll_44_tmpxft_00000847_00000000_4_GPUkernel_cpp1_ii_ad246568()      in GPUkernel.cu.o
  "___cudaRegisterFunction", referenced from:
      __sti____cudaRegisterAll_44_tmpxft_00000847_00000000_4_GPUkernel_cpp1_ii_ad246568()      in GPUkernel.cu.o
  "___cudaUnregisterFatBinary", referenced from:
      __cudaUnregisterBinaryUtil()      in GPUkernel.cu.o
  "_cudaLaunch", referenced from:
      cudaError cudaLaunch<char>(char*)in GPUkernel.cu.o
  "_cudaSetupArgument", referenced from:
      __device_stub__Z19calculateMandlebrotPiPfffffibff(int*, float*, float, float, float, float, int, bool, float, float)in GPUkernel.cu.o
  "_cudaMalloc", referenced from:
      cudaError cudaMalloc<int>(int**, unsigned long)in GPUkernel.cu.o
      cudaError cudaMalloc<float>(float**, unsigned long)in GPUkernel.cu.o
ld: symbol(s) not found for architecture i386
collect2: ld returned 1 exit status
ld: warning: ignoring file GPUkernel.cu.o, file was built for i386 which is not the architecture being linked (x86_64)
ld: warning: ignoring file /usr/local/lib/libpng15.a, file was built for archive which is not the architecture being linked (x86_64)
Undefined symbols for architecture x86_64:
  "_png_create_write_struct", referenced from:
      renderImage(int*, float*, int, int, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, int)in GPUfractal.cpp.o
  "_png_create_info_struct", referenced from:
      renderImage(int*, float*, int, int, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, int)in GPUfractal.cpp.o
  "_png_set_longjmp_fn", referenced from:
      renderImage(int*, float*, int, int, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, int)in GPUfractal.cpp.o
  "_png_init_io", referenced from:
      renderImage(int*, float*, int, int, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, int)in GPUfractal.cpp.o
  "_png_set_IHDR", referenced from:
      renderImage(int*, float*, int, int, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, int)in GPUfractal.cpp.o
  "_png_set_text", referenced from:
      renderImage(int*, float*, int, int, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, int)in GPUfractal.cpp.o
  "_png_write_info", referenced from:
      renderImage(int*, float*, int, int, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, int)in GPUfractal.cpp.o
  "_png_write_row", referenced from:
      renderImage(int*, float*, int, int, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, int)in GPUfractal.cpp.o
  "_png_write_end", referenced from:
      renderImage(int*, float*, int, int, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, int)in GPUfractal.cpp.o
  "_png_free_data", referenced from:
      renderImage(int*, float*, int, int, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, int)in GPUfractal.cpp.o
  "_png_destroy_write_struct", referenced from:
      renderImage(int*, float*, int, int, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, int)in GPUfractal.cpp.o
ld: symbol(s) not found for architecture x86_64
collect2: ld returned 1 exit status
lipo: can't open input file: /var/folders/rw/344w_6js7d7dhg8pch7zws180000gn/T//ccZYSoRL.out (No such file or directory)
make: *** [base] Error 1

我已经上传最新版代码到github ，您还可以在此处的存档文件夹中看到代码的旧(工作)版本(由于 noob 用户限制，链接已删除，但您可以在 github 上的“archive/0.1”内找到它)。

更多信息:我在代码的第一个版本中遇到了与 undefined symbol 类似的问题，但通过大量谷歌搜索，发现 this post通过添加解决了我的问题 libpng-config --cflags --ldflags到生成文件

感谢阅读到这里，希望第一篇文章不会太糟糕 :)

感谢您的帮助!

Answer 1

您似乎正在为 32 位和 64 位构建主机代码。在 OS X 上，如果你想为 64 位编译你的主机代码 (.cpp)，你需要 -arch x86_64，或者如果你想要 32 位，你需要 -arch i386 ，但我认为您不应该同时指定两者。

对于您的设备代码，如果您想要 64 位，那么您需要将 -m64 添加到 nvcc 命令行。对于 32 位，使用 -m32。

您还需要链接到适当的 CUDA 库(32 位或 64 位)。如果您查看 CUDA 安装路径，您会看到有 32 位和 64 位目录，以及相应的库。 只包含一个这些链接路径(使用 -L)，具体取决于您构建的是 32 位还是 64 位。然后，如果您正在使用 CUDA 运行时 API(可能)，则还必须链接 libcuda (-lcuda) 和 libcudart (-lcudart)。

您还需要对您链接的任何其他库(libpng 或其他)执行相同的操作。