Linux build fails with GCC 10.3.0 (compiler bug)
f-fusco opened this issue · 9 comments
Not related to xmrig-cuda per se but it's worth pointing out for people like me that have to go through the error-prone process of building from source.
A workaround is to compile with GCC 9:
cmake -DCMAKE_C_COMPILER=gcc-9 -DCMAKE_CXX_COMPILER=g++-9 -DCUDA_TOOLKIT_ROOT_DIR="/usr/lib/cuda" .
cmake --build .
Same issue in another repo: kokkos/kokkos#4334
That didn't work for me though.
Err:-
CMake Error at CMakeLists.txt:2 (project):
The CMAKE_C_COMPILER:
gcc-9
is not a full path and was not found in the PATH.
Tell CMake where to find the compiler by setting either the environment
variable "CC" or the CMake cache entry CMAKE_C_COMPILER to the full path to
the compiler, or to the compiler name if it is in the PATH.
CMake Error at CMakeLists.txt:2 (project):
The CMAKE_CXX_COMPILER:
g++-9
is not a full path and was not found in the PATH.
Tell CMake where to find the compiler by setting either the environment
variable "CXX" or the CMake cache entry CMAKE_CXX_COMPILER to the full path
to the compiler, or to the compiler name if it is in the PATH.
Of course you need to have gcc 9 installed in your system for this to work.
You can also try clang alternatively.
@TeamRazen you have to have g++-9 package installed for it to be available...
CUDA Toolkit always has specific compiler requirements, which is the whole reason it's a plugin. So you can use the old garbage CUDA requires for the CUDA parts, while also compiling the main miner with a legitimate current compiler.
CUDA Toolkits before 11.4.1 did not support newer than gcc-9.
You can find the supported maximum gcc version listed in the Installation Guide for Linux within the versioned documentation of each release. That is how I just went back and found where the docs said gcc-9 and then at 11.4.1 it jumps ahead to gcc-11. If you are using an older CUDA due to an older GPU family (Fermi/Kepler) being unsupported in newer toolkits then you are limited by what the maximum compiler was "back then". CUDA 8.0GA2 requires gcc-5 for example.
Thanks for the help, anyways, I fixed my install with just adding #include <cstddef> to the Algorithms.h
I am running Debian testing (bookworm) and is already gcc 12 although I have gcc 9/10/11 also still installed, so I was able to compile with -DCMAKE_C_COMPILER=gcc-10 -DCMAKE_CXX_COMPILER=g++-10
Is this already merged in the dev branch for gcc 11?
Never mind I see #include is already added
For those who have the same problem, there is a solution to install (non-destructive) older versions of tools like GCC.
The solution I use is "Spack".
# install spack in local directory
mkdir -p ~/.local/share
git clone -c feature.manyFiles=true https://github.com/spack/spack.git ~/.local/share/spack
# type this, you can also add this in your bashrc or bashrc.d
source ~/.local/share/spack/share/spack/setup-env.sh
# now, spack command exists, install gcc@11
# because the build is a bit heavy, change your TMPDIR to not fill your RAM
TMPDIR=/var/tmp spack install gcc@11
# it takes a long time...
# then, when you need to use gcc v11, you can type (session)
spack load gcc@11
# Then build the plugin
git clone git@github.com:xmrig/xmrig-cuda.git
cd xmrig-cuda
mkdir build
cd build
cmake .. && makeThis works on Fedora 38.
Due to the Kepler arch's support are dropped in CUDA 12.x (which is compute_35 I believe), just remove these flags in cmake file. Following is what I did to my /cmake/CUDA.cmake file. And it do work.
set(DEFAULT_CUDA_ARCH "75")
# Fermi GPUs are only supported with CUDA < 9.0
#if (CUDA_VERSION VERSION_LESS 9.0)
# list(APPEND DEFAULT_CUDA_ARCH "20;21")
#endif()
# Kepler GPUs are only supported with CUDA < 11.0
#if (CUDA_VERSION VERSION_LESS 11.0)
#list(APPEND DEFAULT_CUDA_ARCH "30")
#else()
#list(APPEND DEFAULT_CUDA_ARCH "35")
#endif()
# add Pascal support for CUDA >= 8.0
#if (NOT CUDA_VERSION VERSION_LESS 8.0)
# list(APPEND DEFAULT_CUDA_ARCH "60")
#endif()
# add Volta support for CUDA >= 9.0
#if (NOT CUDA_VERSION VERSION_LESS 9.0)
# list(APPEND DEFAULT_CUDA_ARCH "70")
#endif()
# add Turing support for CUDA >= 10.0
if (NOT CUDA_VERSION VERSION_LESS 10.0)
list(APPEND DEFAULT_CUDA_ARCH "75")
endif()
# add Ampere support for CUDA >= 11.0
if (NOT CUDA_VERSION VERSION_LESS 11.0)
list(APPEND DEFAULT_CUDA_ARCH "80")
endif()
list(SORT DEFAULT_CUDA_ARCH)
BTW, I'm using fedora39, and it has gcc13 by default which is not supported by nvcc. I added set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS}" -allow-unsupported-compiler) in the same file to by pass the error.
For those who have the same problem, there is a solution to install (non-destructive) older versions of tools like GCC.
The solution I use is "Spack".
# install spack in local directory mkdir -p ~/.local/share git clone -c feature.manyFiles=true https://github.com/spack/spack.git ~/.local/share/spack # type this, you can also add this in your bashrc or bashrc.d source ~/.local/share/spack/share/spack/setup-env.sh # now, spack command exists, install gcc@11 # because the build is a bit heavy, change your TMPDIR to not fill your RAM TMPDIR=/var/tmp spack install gcc@11 # it takes a long time... # then, when you need to use gcc v11, you can type (session) spack load gcc@11 # Then build the plugin git clone git@github.com:xmrig/xmrig-cuda.git cd xmrig-cuda mkdir build cd build cmake .. && makeThis works on Fedora 38.
After following all of this, it ended up not working with debian bookworm. Rest is in xmrig/xmrig-nvidia#348