Cannot open STL source file "concepts" when compiling stdexec example code using nvc++

Accelerated Computing HPC Compilers nvc, nvc++ and nvfortran
1

I’m trying to compile this example code posted on Nvidia’s GitHub page.
The compiler version is NVHPC SDK 22.11 23.3 and 23.9. However, both three of them cannot compile the code, here is the error message:
version 22.11:

$ nvc++ -std=c++20 --experimental-stdpar -o run sample.cc
"/work/opt/local/x86_64/cores/nvidia/22.11/Linux_x86_64/22.11/compilers/include-stdexec/experimental/stdexec/__detail/__config.hpp", line 19: catastrophic error: #error directive: This library requires the use of C++20.
  #error This library requires the use of C++20.
   ^

1 catastrophic error detected in the compilation of "sample.cc".
Compilation terminated.

version 23.3 and 23.9:

$ nvc++ --experimental-stdpar -std=c++20 -stdpar=gpu -o run sample.cc
"/work/opt/local/x86_64/cores/nvidia/23.9/Linux_x86_64/23.9/compilers/include-stdexec/experimental/stdexec/execution.hpp", line 20: catastrophic error: cannot open source file "concepts"
  #include <concepts>
                     ^
1 catastrophic error detected in the compilation of "sample.cc".

It seems that the nvc++ compiler cannot open the C++20 standard library header . But this sample code can be compiled in godbolt.org. Could you please help me to check this error?

2

In order to be object compatible with g++, nvc++ uses the g++ STL. Hence the problem is likely due to having nvc++ configured to use an older g++ STL. I believe “concepts” was first added in GNU 10.1.

By default, the system g++ STL is used, but you can use the flag “–gcc-toolchain=<path/to/gcc/install>” to have nvc++ use a different installation. Alternatively, you can configure the NVHPC install to always use a different install using the “makelocalrc” utility.

To use run “makelocalrc -d . -x -gcc=</full/path/to/bin/gcc> -gpp=</full/path/to/bin/g++ -g77=</full/path/to/bin/gfortran>”.

This will create a file name “localrc” which you can rename and move to any directory. Then set the environment variable “NVLOCALRC” to this file.

Hope this helps,
Mat

3

Thank you very much for your answer! Now I can compile the sample code and get the correct result.

4

Now I’m trying to compile other example programs on the GitHub page. However this method doesn’t work on reduce.cpp
The error message I got is:

$ nvc++ -std=c++20 --gcc-toolchain=/work/opt/local/x86_64/cores/gcc/12.2.0 --experimental-stdpar -stdpar=multicore -o run reduce.cpp
"/work/opt/local/x86_64/cores/nvidia/23.3/Linux_x86_64/23.3/compilers/include-stdexec/experimental/nvexec/detail/config.cuh", line 21: catastrophic error: #error directive: The NVIDIA schedulers and utilities require CUDA support
  #error The NVIDIA schedulers and utilities require CUDA support
   ^

1 catastrophic error detected in the compilation of "reduce.cpp".
Compilation terminated.

If I change the -stdpar flag from multicore to gpu, the error message will be:

$ nvc++ -std=c++20 --gcc-toolchain=/work/opt/local/x86_64/cores/gcc/12.2.0 --experimental-stdpar -stdpar=gpu -o run reduce.cpp
"reduce.cpp", line 38: error: no instance of overloaded function "stdexec::__sync_wait::sync_wait_t::operator()" matches the argument list
            argument types are: (nvexec::_strm::reduce_sender_t<nvexec::_strm::schedule_from_sender_t<nvexec::_strm::stream_scheduler, stdexec::__just::__sender<std::span<float, 18446744073709551615UL>>>, float>::__t)
            object type is: const stdexec::__sync_wait::sync_wait_t
    auto [result] = stdexec::sync_wait(std::move(snd)).value();
                    ^

1 error detected in the compilation of "reduce.cpp".

compiler version is nvidia/23.3

5

We didn’t add this support until the 23.9 release, so you’ll want to update your compiler version (https://developer.nvidia.com/hpc-sdk-downloads).

Also, add the flag “-cuda” to enable CUDA support.

6

Hi Mat, sorry for bothering you again. This time I wrote one test case and it works well in godbolt. but when I compiled it using NVHPC 23.11, I got the error message as shown below, could you please help me to check it?

nvc++ -std=c++20 -cuda --gcc-toolchain=/work/opt/local/x86_64/cores/gcc/12.2.0 --experimental-stdpar -stdpar=gpu -o run sample.cc
"/work/opt/local/x86_64/cores/nvidia/23.11/Linux_x86_64/23.11/compilers/include-stdexec/experimental/nvexec/stream/bulk.cuh", line 28: error: static assertion failed
        static_assert(trivially_copyable<Shape, Fun, As...>);
        ^
          detected during:
            instantiation of "void nvexec::_strm::_bulk::kernel<BlockThreads,As...,Shape,Fun>(Shape, Fun, As...) [with BlockThreads=256, As=<>, Shape=unsigned long, Fun=grid_initializer_t]" at line 62
            instantiation of "void nvexec::_strm::_bulk::tag_invoke(_Tag, nvexec::_strm::_bulk::receiver_t<stdexec::__minvoke_<stdexec::__id_<true>, nvexec::_strm::_transfer::operation_state_t<nvexec::_strm::bulk_sender_t<stdexec::__minvoke_<stdexec::__id_<true>, std::decay<nvexec::_strm::schedule_from_sender_t<nvexec::_strm::stream_scheduler, stdexec::__minvoke_<stdexec::__id_<true>, exec::__stl::__sender<exec::__on::__with_sched<stdexec::__id<stdexec::__decay_t<std::decay<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>>::type>>, stdexec::__mdefer_<stdexec::__q<stdexec::__call_result_>, stdexec::__queries::get_scheduler_t, stdexec::__mdefer_<stdexec::__q<stdexec::__call_result_>, stdexec::__env::get_env_t, exec::__stl::__receiver_placeholder<stdexec::__sync_wait::__env>>::__t>::__t>, stdexec::__t<stdexec::__minvoke_<stdexec::__id_<true>, std::decay<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>>::type>::__t>, exec::__on::__with_sched_kernel<stdexec::__mdefer_<stdexec::__q<stdexec::__call_result_>, stdexec::__queries::get_scheduler_t, stdexec::__mdefer_<stdexec::__q<stdexec::__call_result_>, stdexec::__env::get_env_t, exec::__stl::__receiver_placeholder<stdexec::__sync_wait::__env>>::__t>::__t>>::__t>::__t>::__t>::type>::__t, unsigned long, grid_initializer_t>, stdexec::__minvoke_<stdexec::__id_<false>, stdexec::__schedule_from::__receiver1<stdexec::__minvoke_<stdexec::__id_<true>, stdexec::__mdefer_<stdexec::__q<stdexec::__call_result_>, stdexec::__queries::get_scheduler_t, stdexec::__mdefer_<stdexec::__q<stdexec::__call_result_>, stdexec::__env::get_env_t, exec::__stl::__receiver_placeholder<stdexec::__sync_wait::__env>>::__t>::__t>::__t, stdexec::__minvoke_<stdexec::__id_<false>, stdexec::__minvoke_<stdexec::__minvoke_<stdexec::__mfold_right<stdexec::__munique<stdexec::__mbind_front_q<std::variant, std::monostate>>, stdexec::__mbind_front_q<stdexec::__schedule_from::__bind_completions_t, stdexec::__mtype<stdexec::__minvoke_<stdexec::__q<stdexec::__mfront>, stdexec::__minvoke_<stdexec::__detail::__mbc<nvexec::_strm::transfer_sender_th<std::decay<nvexec::_strm::bulk_sender_t<stdexec::__minvoke_<stdexec::__id_<true>, std::decay<nvexec::_strm::schedule_from_sender_t<nvexec::_strm::stream_scheduler, stdexec::__minvoke_<stdexec::__id_<true>, exec::__stl::__sender<exec::__on::__with_sched<stdexec::__id<stdexec::__decay_t<std::decay<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>>::type>>, stdexec::__mdefer_<stdexec::__q<stdexec::__call_result_>, stdexec::__queries::get_scheduler_t, stdexec::__mdefer_<stdexec::__q<stdexec::__call_result_>, stdexec::__env::get_env_t, exec::__stl::__receiver_placeholder<stdexec::__sync_wait::__env>>::__t>::__t>, stdexec::__t<stdexec::__minvoke_<stdexec::__id_<true>, std::decay<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>>::type>::__t>, exec::__on::__with_sched_kernel<stdexec::__mdefer_<stdexec::__q<stdexec::__call_result_>, stdexec::__queries::get_scheduler_t, stdexec::__mdefer_<stdexec::__q<stdexec::__call_result_>, stdexec::__env::get_env_t, exec::__stl::__receiver_placeholder<stdexec::__sync_wait::__env>>::__t>::__t>>::__t>::__t>::__t>::type>::__t, unsigned long, grid_initializer_t>::__t>::type>>, stdexec::__cp>::__t>::__t>::__t, stdexec::__mdefer_<stdexec::__q<stdexec::__call_result_>, stdexec::__env::get_env_t, exec::__stl::__operation<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))> &, exec::__on::__continue_on_kernel<nvexec::_strm::stream_scheduler, stdexec::__closure::__binder_back<stdexec::__bulk::bulk_t, std::size_t, grid_initializer_t>>, stdexec::__minvoke_<stdexec::__id_<true>, stdexec::__sync_wait::__receiver<>::__t>::__t>::__receiver_t>::__t>>, stdexec::__receivers::set_value_t, stdexec::__receivers::set_error_t, stdexec::__receivers::set_stopped_t>::__t>::__t>::__t, stdexec::__minvoke_<stdexec::__id_<true>, stdexec::__debug::__debug_receiver<stdexec::__cvref_id<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>, stdexec::__decay_t<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>>>, exec::__on::__with_sched_env<stdexec::__sync_wait::__env, nvexec::_strm::stream_scheduler>, stdexec::__tag_invoke::tag_invoke_result_t<stdexec::__get_completion_signatures::get_completion_signatures_t, stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>, exec::__on::__with_sched_env<stdexec::__sync_wait::__env, nvexec::_strm::stream_scheduler>>>>::__t>::__t>::__t>::__t::enqueue_receiver>::__t, unsigned long, grid_initializer_t>::__t &&, As &&...) noexcept [with _Tag=stdexec::__receivers::set_value_t, As=<>]" at line 169 of "/work/opt/local/x86_64/cores/nvidia/23.11/Linux_x86_64/23.11/compilers/include-stdexec/experimental/stdexec/functional.hpp"
            instantiation of "auto stdexec::__tag_invoke::tag_invoke_t::operator()(_Tag, _Args &&...) const->stdexec::__tag_invoke::tag_invoke_result_t<_Tag, _Args...> [with _Tag=stdexec::__receivers::set_value_t, _Args=<nvexec::_strm::_bulk::receiver_t<nvexec::_strm::stream_enqueue_receiver<stdexec::__env::__joined_env<stdexec::__env::__env_fn<lambda [](nvexec::_strm::get_stream_provider_t)->nvexec::_strm::stream_provider_t * noexcept>, stdexec::__env::__joined_env<stdexec::__env::__env_fn<lambda [](stdexec::__debug::__is_debug_env_t)->bool noexcept(true)>, exec::__on::__with_sched_env<stdexec::__sync_wait::__env, nvexec::_strm::stream_scheduler>>>, nvexec::variant_t<cuda::std::__4::tuple<nvexec::_strm::set_noop>, cuda::std::__4::tuple<stdexec::__receivers::set_error_t, cudaError_t>, cuda::std::__4::tuple<stdexec::__receivers::set_value_t>>>, unsigned long, grid_initializer_t>::__t>]" at line 656 of "/work/opt/local/x86_64/cores/nvidia/23.11/Linux_x86_64/23.11/compilers/include-stdexec/experimental/stdexec/execution.hpp"
            instantiation of "void stdexec::__receivers::set_value_t::operator()(_Receiver &&, _As &&...) const noexcept [with _Receiver=nvexec::_strm::_bulk::receiver_t<nvexec::_strm::stream_enqueue_receiver<stdexec::__env::__joined_env<stdexec::__env::__env_fn<lambda [](nvexec::_strm::get_stream_provider_t)->nvexec::_strm::stream_provider_t * noexcept>, stdexec::__env::__joined_env<stdexec::__env::__env_fn<lambda [](stdexec::__debug::__is_debug_env_t)->bool noexcept(true)>, exec::__on::__with_sched_env<stdexec::__sync_wait::__env, nvexec::_strm::stream_scheduler>>>, nvexec::variant_t<cuda::std::__4::tuple<nvexec::_strm::set_noop>, cuda::std::__4::tuple<stdexec::__receivers::set_error_t, cudaError_t>, cuda::std::__4::tuple<stdexec::__receivers::set_value_t>>>, unsigned long, grid_initializer_t>::__t, _As=<>]" at line 584 of "/work/opt/local/x86_64/cores/nvidia/23.11/Linux_x86_64/23.11/compilers/include-stdexec/experimental/nvexec/stream/common.cuh"
            instantiation of "void nvexec::_strm::operation_state_base_<OuterReceiverId>::__t::propagate_completion_signal(Tag, As &&...) noexcept [with OuterReceiverId=nvexec::_strm::_bulk::receiver_t<nvexec::_strm::stream_enqueue_receiver<stdexec::__env::__joined_env<stdexec::__env::__env_fn<lambda [](nvexec::_strm::get_stream_provider_t)->nvexec::_strm::stream_provider_t * noexcept>, stdexec::__env::__joined_env<stdexec::__env::__env_fn<lambda [](stdexec::__debug::__is_debug_env_t)->bool noexcept(true)>, exec::__on::__with_sched_env<stdexec::__sync_wait::__env, nvexec::_strm::stream_scheduler>>>, nvexec::variant_t<cuda::std::__4::tuple<nvexec::_strm::set_noop>, cuda::std::__4::tuple<stdexec::__receivers::set_error_t, cudaError_t>, cuda::std::__4::tuple<stdexec::__receivers::set_value_t>>>, unsigned long, grid_initializer_t>, Tag=stdexec::__receivers::set_value_t, As=<>]" at line 55 of "/work/opt/local/x86_64/cores/nvidia/23.11/Linux_x86_64/23.11/compilers/include-stdexec/experimental/nvexec/stream/schedule_from.cuh"
            [ 24 instantiation contexts not shown ]
            instantiation of "void stdexec::__start::start_t::operator()(_Op &) const noexcept [with _Op=nvexec::_strm::_transfer::operation_state_t<nvexec::_strm::bulk_sender_t<nvexec::_strm::schedule_from_sender_t<nvexec::_strm::stream_scheduler, exec::__on::__with_sched<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>, stdexec::__loop::run_loop::__scheduler>>, unsigned long, grid_initializer_t>, stdexec::_Yp<stdexec::__schedule_from::__receiver1<stdexec::__loop::run_loop::__scheduler::__id, stdexec::_Yp<std::variant<std::monostate, std::tuple<stdexec::__receivers::set_stopped_t>, std::tuple<stdexec::__receivers::set_error_t, cudaError_t>, std::tuple<stdexec::__receivers::set_value_t>>>, stdexec::__debug::__debug_receiver<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>, exec::__on::__with_sched_env<stdexec::__sync_wait::__env, nvexec::_strm::stream_scheduler>, stdexec::completion_signatures<stdexec::__receivers::set_error_t (std::__exception_ptr::exception_ptr), stdexec::__receivers::set_stopped_t (), stdexec::__receivers::set_value_t (), stdexec::__receivers::set_error_t (cudaError &&)>>>::__t>>::__t]" at line 5144 of "/work/opt/local/x86_64/cores/nvidia/23.11/Linux_x86_64/23.11/compilers/include-stdexec/experimental/stdexec/execution.hpp"
            instantiation of class "stdexec::__schedule_from::__operation1<_SchedulerId, _CvrefSenderId, _ReceiverId>::__t [with _SchedulerId=stdexec::__loop::run_loop::__scheduler::__id, _CvrefSenderId=nvexec::_strm::transfer_sender_t<nvexec::_strm::bulk_sender_t<nvexec::_strm::schedule_from_sender_t<nvexec::_strm::stream_scheduler, exec::__on::__with_sched<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>, stdexec::__loop::run_loop::__scheduler>>, unsigned long, grid_initializer_t>>, _ReceiverId=stdexec::__debug::__debug_receiver<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>, exec::__on::__with_sched_env<stdexec::__sync_wait::__env, nvexec::_strm::stream_scheduler>, stdexec::completion_signatures<stdexec::__receivers::set_error_t (std::__exception_ptr::exception_ptr), stdexec::__receivers::set_stopped_t (), stdexec::__receivers::set_value_t (), stdexec::__receivers::set_error_t (cudaError &&)>>]" at line 169 of "/work/opt/local/x86_64/cores/nvidia/23.11/Linux_x86_64/23.11/compilers/include-stdexec/experimental/stdexec/functional.hpp"
            instantiation of "auto stdexec::__tag_invoke::tag_invoke_t::operator()(_Tag, _Args &&...) const->stdexec::__tag_invoke::tag_invoke_result_t<_Tag, _Args...> [with _Tag=stdexec::__start::start_t, _Args=<stdexec::__schedule_from::__operation1<stdexec::__loop::run_loop::__scheduler::__id, nvexec::_strm::transfer_sender_t<nvexec::_strm::bulk_sender_t<nvexec::_strm::schedule_from_sender_t<nvexec::_strm::stream_scheduler, exec::__on::__with_sched<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>, stdexec::__loop::run_loop::__scheduler>>, unsigned long, grid_initializer_t>>, stdexec::__debug::__debug_receiver<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>, exec::__on::__with_sched_env<stdexec::__sync_wait::__env, nvexec::_strm::stream_scheduler>, stdexec::completion_signatures<stdexec::__receivers::set_error_t (std::__exception_ptr::exception_ptr), stdexec::__receivers::set_stopped_t (), stdexec::__receivers::set_value_t (), stdexec::__receivers::set_error_t (cudaError &&)>>>::__t &>]" at line 1675 of "/work/opt/local/x86_64/cores/nvidia/23.11/Linux_x86_64/23.11/compilers/include-stdexec/experimental/stdexec/execution.hpp"
            instantiation of "void stdexec::__start::start_t::operator()(_Op &) const noexcept [with _Op=stdexec::__schedule_from::__operation1<stdexec::__loop::run_loop::__scheduler::__id, nvexec::_strm::transfer_sender_t<nvexec::_strm::bulk_sender_t<nvexec::_strm::schedule_from_sender_t<nvexec::_strm::stream_scheduler, exec::__on::__with_sched<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>, stdexec::__loop::run_loop::__scheduler>>, unsigned long, grid_initializer_t>>, stdexec::__debug::__debug_receiver<stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>, exec::__on::__with_sched_env<stdexec::__sync_wait::__env, nvexec::_strm::stream_scheduler>, stdexec::completion_signatures<stdexec::__receivers::set_error_t (std::__exception_ptr::exception_ptr), stdexec::__receivers::set_stopped_t (), stdexec::__receivers::set_value_t (), stdexec::__receivers::set_error_t (cudaError &&)>>>::__t]" at line 1080 of "/work/opt/local/x86_64/cores/nvidia/23.11/Linux_x86_64/23.11/compilers/include-stdexec/experimental/stdexec/execution.hpp"
            instantiation of "void stdexec::__debug::__debug_sender<_Sigs,_Env,_Sender>(_Sender &&, const _Env &) [with _Sigs=stdexec::completion_signatures<stdexec::__receivers::set_error_t (std::__exception_ptr::exception_ptr), stdexec::__receivers::set_stopped_t (), stdexec::__receivers::set_value_t (), stdexec::__receivers::set_error_t (cudaError &&)>, _Env=exec::__on::__with_sched_env<stdexec::__sync_wait::__env, nvexec::_strm::stream_scheduler>, _Sender=stdexec::__basic_sender<lambda [](_Cvref, _Fun &&) mutable->decltype((<expression>))>]" at line 1306 of "/work/opt/local/x86_64/cores/nvidia/23.11/Linux_x86_64/23.11/compilers/include-stdexec/experimental/stdexec/execution.hpp"

1 error detected in the compilation of "sample.cc".
7

Looks like the code worked with 23.3 and up to 23.7. I’d need to check with engineering, but given this is experimental, there’s likely been updates the the header files.