Proper way to define C++ standard in CUDA project

[tabedzki]

Hello there,

I am trying to compile on some outdated hardware that requires me to include the -std=c++11 flag with the nvcc command. This works nicely in the old makefile I have where I write all the flags explicitly.

Right now, my meson setup is a top level meson.build and then one in src/meson.build and the explicit inclusion of the flag brings in a warning.

What is the proper way of specifying the C++ standard that the CUDA project should use?
Thanks in advance

src/meson.build:1: WARNING: Consider using the built-in option for language standard version instead of using "-std=c++11".

# meson.build
project('cuda-example', 'cuda',
    version         : '0.1.0',
    license         : 'MIT',
    meson_version   : '>=0.50.0',
    default_options : ['cpp_std=c++11'])

cuda = import('unstable-cuda')
# cuda.nvcc_arch_flags('Auto')

subdir('src')

# src/meson.build
add_project_arguments(['-rdc=true', '-lm','-std=c++11'], language : ['cuda'])
dep = dependency('cuda', version : '>=10', modules : ['cufft'])

executable('main',
'pair_style.cu', 'pair_style_erf.cu',
'main.cu', dependencies: dep,
install: true)

[eli-schwartz]

Did you try cuda_std=c++11 in your default options? :)

[tabedzki]

Hi Eli, I did try that and apologies for neglecting to mention that. When I do remove the flag from src/meson.build and modify the top level "meson.build" to include that flag, I get the error at the bottom. It doesn't have the c++11 flag as denoted by the compiler; I also get a bunch of errors when attempting to use the range based for loop.

#meson.build
project('cuda-example', 'cuda',
    version         : '0.1.0',
    license         : 'MIT',
    meson_version   : '>=0.50.0',
    default_options : ['cuda_std=c++11'])

cuda = import('unstable-cuda')
# cuda.nvcc_arch_flags('Auto')

subdir('src')

=>meson build --buildtype release --reconfigure -Dprefix=$HOME
The Meson build system
Version: 0.63.0
Source dir: /mnt/io1/home/tabedzki/code/meson-test
Build dir: /mnt/io1/home/tabedzki/code/meson-test/build
Build type: native build
Project name: cuda-example
Project version: 0.1.0
Cuda compiler for the host machine: nvcc (nvcc 10.2.89)
Cuda linker for the host machine: nvcc nvlink 10.2.89
Host machine cpu family: x86_64
Host machine cpu: x86_64
meson.build:7: WARNING: Module unstable-cuda has no backwards or forwards compatibility and might not exist in future releases.
Dependency cuda (modules: cufft) found: YES 10.2 (cached)
Build targets in project: 1

cuda-example 0.1.0

  User defined options
    buildtype: release
    prefix   : /home/tabedzki

=> meson install
ninja: Entering directory `/mnt/io1/home/tabedzki/code/meson-test/build'
[2/45] Compiling Cuda object src/main.p/bonds.cu.o
FAILED: src/main.p/bonds.cu.o 
nvcc -Isrc/main.p -Xcompiler=-Wall,-Winvalid-pch,-Wnon-virtual-dtor -O3 -rdc=true -lm -I../src -Isrc -Isrc/main.p -o src/main.p/bonds.cu.o -c ../src/bonds.cu
../src/pair_style.cuh(23): warning: field initializers are a C++11 feature

../src/Compute.cuh(26): error: space required between adjacent ">" delimiters of nested template argument lists (">>" is the right shift operator)

[eli-schwartz]

That is indeed very relevant.

It sounds like a bug and we're not actually applying the setting when writing out the build file. Perhaps @obilaniu as a frequent contributor to our cuda support can confirm.

[tabedzki]

I think I found the change that works and can be useful for others. I put the multiple language requirements for project within brackets [] and then also have all the default_options. I thought I had tried that last night but evidently not. Thank you developers for including the meson init --language=cuda. That showed me where I was going wrong.

# meson.build
project('cuda-example', ['cuda', 'cpp'],
    version         : '0.1.0',
    license         : 'MIT',
    meson_version   : '>=0.50.0',
    default_options : ['cuda_std=c++11','c_std=c11', 'cpp_std=c++11'])

cuda = import('unstable-cuda')
# cuda.nvcc_arch_flags('Auto')

subdir('src')

# src/meson.build
add_project_arguments(['-rdc=true'], language : 'cuda')
dep = dependency('cuda', version : '>=10', modules : ['cufft'])

executable('main', 'main.cu', dependencies: dep,
install: true,
)

[tabedzki]

I think I found the change that works and can be useful for others. I put the multiple language requirements for project within brackets [] and then also have all the default_options. I thought I had tried that last night but evidently not. Thank you developers for including the meson init --language=cuda. That showed me where I was going wrong.

# meson.build
project('cuda-example', ['cuda', 'cpp'],
    version         : '0.1.0',
    license         : 'MIT',
    meson_version   : '>=0.50.0',
    default_options : ['cuda_std=c++11','c_std=c11', 'cpp_std=c++11'])

cuda = import('unstable-cuda')
# cuda.nvcc_arch_flags('Auto')

subdir('src')

# src/meson.build
add_project_arguments(['-rdc=true'], language : 'cuda')
dep = dependency('cuda', version : '>=10', modules : ['cufft'])

executable('main', 'main.cu', dependencies: dep,
install: true,
)