arma：：Mat给出了GCC和NVCC的不同结果。

Question

#include <iostream>
#include <armadillo>
using namespace std;

int main()
{
    arma::Mat<float> a;
    cout << sizeof(a) << "\n";
    return 0;
}

当我将NVCC用于CUDA时，上面的代码给出了不同的结果。

$ g++ -o main test.cu.cpp -O3 -larmadillo
$ ./main
112
$ nvcc -o main test.cu.cpp -O3 -larmadillo
$ ./main
104

我希望使NVCC版本与GCC版本一样。。

这种差异是从何而来的？我的项目需要单独编译不同的部分，由于遗留部分必须使用GCC，所以不可能将所有内容转换为NVCC。

编辑：这里是NVCC和GCC之间的编译日志，我不知道要找什么

[huyduc@ny5-dtlgpu06 test]$ nvcc -o main test.cu.cpp -O3 -larmadillo --verbose
#$ _NVVM_BRANCH_=nvvm
#$ _SPACE_= 
#$ _CUDART_=cudart
#$ _HERE_=/usr/local/cuda-11.4/bin
#$ _THERE_=/usr/local/cuda-11.4/bin
#$ _TARGET_SIZE_=
#$ _TARGET_DIR_=
#$ _TARGET_DIR_=targets/x86_64-linux
#$ TOP=/usr/local/cuda-11.4/bin/..
#$ NVVMIR_LIBRARY_DIR=/usr/local/cuda-11.4/bin/../nvvm/libdevice
#$ LD_LIBRARY_PATH=/usr/local/cuda-11.4/bin/../lib:/local/export/scratch/pulse_packages/lib/:/local/export/scratch/pulse_packages/lib64/:/local/export/scratch/pulse_packages/usr/lib:/local/export/scratch/pulse_packages/usr/local/lib/:/local/export/scratch/pulse_packages/mkl/lib:/local/export/scratch/pulse_packages/mkl/mkl/lib/intel64:/local/export/scratch/pulse_packages/mods/libtorch/lib::/usr/local/cuda-11.4/lib64:/usr/lib/x86_64-linux-gnu
#$ PATH=/usr/local/cuda-11.4/bin/../nvvm/bin:/usr/local/cuda-11.4/bin:/local/export/scratch/pulse_packages/bin:/local/export/scratch/pulse_packages/usr/bin:/local/export/scratch/pulse_packages/mods/libtorch/bin:/usr/lib64/ccache:/usr/local/bin:/usr/bin:/usr/local/sbin:/usr/sbin:/opt/ibutils/bin:/home/huyduc/.local/bin:/home/huyduc/bin:/usr/local/cuda-11.4/bin
#$ INCLUDES="-I/usr/local/cuda-11.4/bin/../targets/x86_64-linux/include"  
#$ LIBRARIES=  "-L/usr/local/cuda-11.4/bin/../targets/x86_64-linux/lib/stubs" "-L/usr/local/cuda-11.4/bin/../targets/x86_64-linux/lib"
#$ CUDAFE_FLAGS=
#$ PTXAS_FLAGS=
#$ gcc -c -x c++ -D__NVCC__  -O3 "-I/usr/local/cuda-11.4/bin/../targets/x86_64-linux/include"    -D__CUDACC_VER_MAJOR__=11 -D__CUDACC_VER_MINOR__=4 -D__CUDACC_VER_BUILD__=100 -D__CUDA_API_VER_MAJOR__=11 -D__CUDA_API_VER_MINOR__=4 -m64 "test.cu.cpp" -o "/tmp/tmpxft_000140a2_00000000-5_test.cu.o" 
#$ nvlink --arch=sm_52 --register-link-binaries="/tmp/tmpxft_000140a2_00000000-3_main_dlink.reg.c"  -m64 -larmadillo   "-L/usr/local/cuda-11.4/bin/../targets/x86_64-linux/lib/stubs" "-L/usr/local/cuda-11.4/bin/../targets/x86_64-linux/lib" -cpu-arch=X86_64 "/tmp/tmpxft_000140a2_00000000-5_test.cu.o"  -lcudadevrt  -o "/tmp/tmpxft_000140a2_00000000-6_main_dlink.sm_52.cubin"
#$ fatbinary -64 -link "--image3=kind=elf,sm=52,file=/tmp/tmpxft_000140a2_00000000-6_main_dlink.sm_52.cubin" --embedded-fatbin="/tmp/tmpxft_000140a2_00000000-4_main_dlink.fatbin.c" 
#$ rm /tmp/tmpxft_000140a2_00000000-4_main_dlink.fatbin
#$ gcc -c -x c++ -DFATBINFILE="\"/tmp/tmpxft_000140a2_00000000-4_main_dlink.fatbin.c\"" -DREGISTERLINKBINARYFILE="\"/tmp/tmpxft_000140a2_00000000-3_main_dlink.reg.c\"" -I. -D__NV_EXTRA_INITIALIZATION= -D__NV_EXTRA_FINALIZATION= -D__CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__  -O3 "-I/usr/local/cuda-11.4/bin/../targets/x86_64-linux/include"    -D__CUDACC_VER_MAJOR__=11 -D__CUDACC_VER_MINOR__=4 -D__CUDACC_VER_BUILD__=100 -D__CUDA_API_VER_MAJOR__=11 -D__CUDA_API_VER_MINOR__=4 -m64 "/usr/local/cuda-11.4/bin/crt/link.stub" -o "/tmp/tmpxft_000140a2_00000000-7_main_dlink.o" 
#$ g++ -O3 -m64 -Wl,--start-group "/tmp/tmpxft_000140a2_00000000-7_main_dlink.o" "/tmp/tmpxft_000140a2_00000000-5_test.cu.o" -larmadillo   "-L/usr/local/cuda-11.4/bin/../targets/x86_64-linux/lib/stubs" "-L/usr/local/cuda-11.4/bin/../targets/x86_64-linux/lib"  -lcudadevrt  -lcudart_static  -lrt -lpthread  -ldl  -Wl,--end-group -o "main" 

[huyduc@ny5-dtlgpu06 test]$ g++ -o gccc test.cu.cpp -O3 -larmadillo --verbose
Using built-in specs.
COLLECT_GCC=g++
COLLECT_LTO_WRAPPER=/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: ../gcc-10.2.0/configure --prefix=/local/export/scratch/pulse_packages --libexecdir=/local/export/scratch/pulse_packages/lib --enable-shared --enable-threads=posix --enable-__cxa_atexit --disable-multilib --enable-bootstrap --enable-clocale=gnu --enable-languages=c,c++,fortran --with-zstd=no
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 10.2.0 (GCC) 
COLLECT_GCC_OPTIONS='-o' 'gccc' '-O3' '-v' '-shared-libgcc' '-mtune=generic' '-march=x86-64'
 /local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/cc1plus -quiet -v -D_GNU_SOURCE test.cu.cpp -quiet -dumpbase test.cu.cpp -mtune=generic -march=x86-64 -auxbase test.cu -O3 -version -o /tmp/cctgH2WC.s
GNU C++14 (GCC) version 10.2.0 (x86_64-pc-linux-gnu)
    compiled by GNU C version 10.2.0, GMP version 6.1.0, MPFR version 3.1.4, MPC version 1.0.3, isl version isl-0.18-GMP

GGC heuristics: --param ggc-min-expand=100 --param ggc-min-heapsize=131072
ignoring nonexistent directory "/local/export/scratch/pulse_packages/usr/include/"
ignoring nonexistent directory "/local/export/scratch/pulse_packages/usr/local/include"
ignoring nonexistent directory "/usr/include/x86_64-linux-gnu"
ignoring duplicate directory "/local/export/scratch/pulse_packages/mods/libtorch/include"
ignoring duplicate directory "/local/export/scratch/pulse_packages/include"
ignoring nonexistent directory "/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../../../x86_64-pc-linux-gnu/include"
#include "..." search starts here:
#include <...> search starts here:
 /local/export/scratch/pulse_packages/include/
 /local/export/scratch/pulse_packages/mods/libtorch/include
 /local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../../../include/c++/10.2.0
 /local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../../../include/c++/10.2.0/x86_64-pc-linux-gnu
 /local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../../../include/c++/10.2.0/backward
 /local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/include
 /usr/local/include
 /local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/include-fixed
 /usr/include
End of search list.
GNU C++14 (GCC) version 10.2.0 (x86_64-pc-linux-gnu)
    compiled by GNU C version 10.2.0, GMP version 6.1.0, MPFR version 3.1.4, MPC version 1.0.3, isl version isl-0.18-GMP

GGC heuristics: --param ggc-min-expand=100 --param ggc-min-heapsize=131072
Compiler executable checksum: fcab5cdad8fab5c0a9dfd14a10ab3fb4
COLLECT_GCC_OPTIONS='-o' 'gccc' '-O3' '-v' '-shared-libgcc' '-mtune=generic' '-march=x86-64'
 /local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../../../x86_64-pc-linux-gnu/bin/as -v --64 -o /tmp/ccPwcVUC.o /tmp/cctgH2WC.s
GNU assembler version 2.36.1 (x86_64-pc-linux-gnu) using BFD version (GNU Binutils) 2.36.1
COMPILER_PATH=/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/:/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/:/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/:/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/:/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/:/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../../../x86_64-pc-linux-gnu/bin/
LIBRARY_PATH=/local/export/scratch/pulse_packages/lib/../lib64/:/local/export/scratch/pulse_packages/lib64/../lib64/:/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/:/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../../../lib64/:/lib/../lib64/:/usr/lib/../lib64/:/local/export/scratch/pulse_packages/lib/:/local/export/scratch/pulse_packages/lib64/:/local/export/scratch/pulse_packages/mkl/lib/:/local/export/scratch/pulse_packages/mkl/mkl/lib/intel64/:/local/export/scratch/pulse_packages/mods/libtorch/lib/:./:/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../../../x86_64-pc-linux-gnu/lib/:/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../../:/lib/:/usr/lib/
COLLECT_GCC_OPTIONS='-o' 'gccc' '-O3' '-v' '-shared-libgcc' '-mtune=generic' '-march=x86-64'
 /local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/collect2 -plugin /local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/liblto_plugin.so -plugin-opt=/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/lto-wrapper -plugin-opt=-fresolution=/tmp/ccd147SC.res -plugin-opt=-pass-through=-lgcc_s -plugin-opt=-pass-through=-lgcc -plugin-opt=-pass-through=-lc -plugin-opt=-pass-through=-lgcc_s -plugin-opt=-pass-through=-lgcc --eh-frame-hdr -m elf_x86_64 -dynamic-linker /lib64/ld-linux-x86-64.so.2 -o gccc /lib/../lib64/crt1.o /lib/../lib64/crti.o /local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/crtbegin.o -L/local/export/scratch/pulse_packages/lib/../lib64 -L/local/export/scratch/pulse_packages/lib64/../lib64 -L/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0 -L/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../../../lib64 -L/lib/../lib64 -L/usr/lib/../lib64 -L/local/export/scratch/pulse_packages/lib -L/local/export/scratch/pulse_packages/lib64 -L/local/export/scratch/pulse_packages/mkl/lib -L/local/export/scratch/pulse_packages/mkl/mkl/lib/intel64 -L/local/export/scratch/pulse_packages/mods/libtorch/lib -L. -L/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../../../x86_64-pc-linux-gnu/lib -L/local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/../../.. /tmp/ccPwcVUC.o -larmadillo -lstdc++ -lm -lgcc_s -lgcc -lc -lgcc_s -lgcc /local/export/scratch/pulse_packages/lib/gcc/x86_64-pc-linux-gnu/10.2.0/crtend.o /lib/../lib64/crtn.o
COLLECT_GCC_OPTIONS='-o' 'gccc' '-O3' '-v' '-shared-libgcc' '-mtune=generic' '-march=x86-64'

Answer 1

我希望使NVCC版本的行为与GCC版本相同。

我不知道你用的是什么版本。我安装了10.8，因为这样做对我来说很容易。(我使用的是数据自动化系统11.4，与您相同。)当我运行您的测试用例时，我得到的大小在nvcc和g++之间确实是不同的，尽管它们并不完全匹配您的大小，但是我也可能从armadillo开发人员那里得到了有用的输出，您注意到了吗？

$ cat t2106.cpp
#include <iostream>
#ifdef USE_SUGGESTION
#define ARMA_ALLOW_FAKE_GCC
#endif
#include <armadillo>
using namespace std;

int main()
{
    arma::arma_version ver;
    cout << "ARMA version: "<< ver.as_string() << std::endl;

    arma::Mat<float> a;
    cout << sizeof(a) << "\n";
    return 0;
}
$ g++ t2106.cpp -o t2106g++ -larmadillo
$ ./t2106g++
ARMA version: 10.8.2 (Realm Raider)
128
$ nvcc t2106.cpp -o t2106nvcc -larmadillo
In file included from /usr/include/armadillo:68:0,
                 from t2106.cpp:5:
/usr/include/armadillo_bits/compiler_setup.hpp:151:106: note: #pragma message: WARNING: this compiler is pretending to be GCC but it may not be fully compatible;
 NG: this compiler is pretending to be GCC but it may not be fully compatible;"

/usr/include/armadillo_bits/compiler_setup.hpp:152:110: note: #pragma message: WARNING: to allow this compiler to use GCC features such as data alignment attributes,
 to allow this compiler to use GCC features such as data alignment attributes,"

/usr/include/armadillo_bits/compiler_setup.hpp:153:88: note: #pragma message: WARNING: #define ARMA_ALLOW_FAKE_GCC before #include <armadillo>
 ma message ("WARNING: #define ARMA_ALLOW_FAKE_GCC before #include <armadillo>"

$ ./t2106nvcc
ARMA version: 10.8.2 (Realm Raider)
112
$ nvcc t2106.cpp -o t2106nvcc -larmadillo -DUSE_SUGGESTION
$ ./t2106nvcc
ARMA version: 10.8.2 (Realm Raider)
128
$

因此，开发人员在10.8中已经指出的，允许在这两种情况下大小相等，这似乎解决了您所提供的测试用例中的问题。

我不知道这里的所有后果是什么，当然也没有做过任何详尽的测试。但这些大小似乎与代码中的定义相同。

不可能将所有内容转换为NVCC。

另一种选择是，您不必转换所有东西才能使用nvcc。在用或g++编译的.c/.cpp文件中只包括，然后使用包装器函数将编译的文件与gcc / g++编译的文件连接起来。这样，您就可以完全避免这个问题，而不必使用特殊的#define。

补充讨论：

相关的对象定义是这里。看看这段代码，再加上对警告消息的观察，我猜arma_aligned和/或arma_aligned_mem装饰器基于开发人员已经实现的编译器检测而表现得不同。但除此之外我还没研究过。

Answer 2

我已经找到了一个解决方案，它适用于较早版本的Armadillo (当您无法更改遗留代码时非常有用)。

这个问题是由于usr/include/armadillo_bits/compiler_setup.hpp中的这些行，Arma7.500.2中的150->156号行(其他版本不同，您可以通过CTRL +F文本#define ARMA_GOOD_COMPILER来查看它附近)。

#if (defined(__GNUG__) || defined(__GNUC__)) && (defined(__clang__) || defined(__INTEL_COMPILER) || defined(__NVCC__) || defined(__CUDACC__) || defined(__PGI) || defined(__PATHSCALE__) || defined(__ARMCC_VERSION) || defined(__IBMCPP__))
  #undef  ARMA_FAKE_GCC
  #define ARMA_FAKE_GCC
#endif


#if defined(__GNUG__) && !defined(ARMA_FAKE_GCC)
... define arma_aligned and arma_align_mem here
#endif

新版本有一个标志#define ARMA_ALLOW_FAKE_GCC，禁止检查ARMA_FAKE_GCC，因此它与CUDA一起工作。但是由于旧版本没有它，所以您必须自己编辑/usr/include/armadillo_bits/compiler_setup.hpp来添加它(这条路径在PC机之间可能有所不同)

#if defined(__GNUG__) && (!defined(ARMA_FAKE_GCC) || defined(__NVCC__) || defined(__CUDACC__))
... define arma_aligned and arma_align_mem here
#endif

示例程序来测试下面的差异。在我的机器上，使用ARMA7.500.2，nvcc在更改前给出104个，在修改后给出112个(与g++相同)。而且，在更改之前，arma_aligned和arma_align_mem是未定义的(空字符串)，而在它们变成__attribute__((__aligned__))和__attribute__((__aligned__(16)))之后(同样，与g++相同)。

#include <iostream>
#define ARMA_ALLOW_FAKE_GCC
#define ARMA_ALLOW_FAKE_CLANG
#include <armadillo>
#include <cstdint>
using namespace std;

#define STRINGIFY(x) #x
#define STRINGIFYMACRO(y) STRINGIFY(y)

int main()
{
    arma::arma_version ver;
        std::cout << "ARMA version: "<< ver.as_string() << std::endl;
    
    arma::Mat<float> a;
    cout << sizeof(a) << "\n";
    
    cout << STRINGIFYMACRO(arma_aligned) << "\n";
    cout << STRINGIFYMACRO(arma_align_mem) << "\n";    

    return 0;
}