'J': is not a valid type for non-type template parameter 'J'

[BennyRod]

Hello, I have been facing an issue using Boost Odeint with CUDA in Windows 11then I am going to include a MVCE using a .cu file:

Configurations: -Microsoft Visual Studio Community 2022 (64-bit) - Current Version 17.11.0 -Nvidia Driver Version: 560.81 -CUDA Toolkit 12.6 -BOOST 1.86 -MVCE.cu

MVCE.cu:

#include <boost/numeric/odeint.hpp>
int main()
{

}

Output:

Error C2334 unexpected token(s) preceding '{'; skipping apparent function body CudaRuntime1 C:\Users\benny\Downloads\boost_1_86_0\boost\fusion\container\vector\vector.hpp 246 2.Error C2993 'J': is not a valid type for non-type template parameter 'J' CudaRuntime1 C:\Users\benny\Downloads\boost_1_86_0\boost\fusion\container\vector\vector.hpp 236

Error C2062 type 'unknown-type' unexpected CudaRuntime1 C:\Users\benny\Downloads\boost_1_86_0\boost\fusion\container\vector\vector.hpp 236

Error C2334 unexpected token(s) preceding '{'; skipping apparent function body CudaRuntime1 C:\Users\benny\Downloads\boost_1_86_0\boost\fusion\container\vector\vector.hpp 239

5.Error C2993 'J': is not a valid type for non-type template parameter 'J' CudaRuntime1 C:\Users\benny\Downloads\boost_1_86_0\boost\fusion\container\vector\vector.hpp 243

6.Error C4430 missing type specifier - int assumed. Note: C++ does not support default-int CudaRuntime1 C:\Users\benny\Downloads\boost_1_86_0\boost\fusion\container\vector\vector.hpp 243

7.Error C2062 type 'unknown-type' unexpected CudaRuntime1 C:\Users\benny\Downloads\boost_1_86_0\boost\fusion\container\vector\vector.hpp 243

The first error refers to the use of a C++ version prior to C++ 20 but I have used C++ 20 and it is the same error.

On the other hand, if it is replaced "odeint" by "interval" then the code compiles properly.

#include <boost/numeric/odeint.hpp>
int main()
{

}

Output: ========== Build: 1 succeeded, 0 failed, 0 up-to-date, 0 skipped ========== ========== Build completed at 11:07 AM and took 01.917 seconds ==========

Finally, this code works properly in Ubuntu using CMake then it is not a general incompatibility between CUDA and BOOST.

*I have already published this error in fusion but I was referred to Odeint: https://github.com/boostorg/fusion/issues/269

I hope you can help us with this issue. Thank you so much in advance! Let me know if there is something I can do to help you to fix this error.

[mborland]

You will have to use the thrust adaptor: https://www.boost.org/doc/libs/1_86_0/libs/numeric/odeint/doc/html/boost_numeric_odeint/tutorial/using_cuda__or_openmp__tbb_______via_thrust.html. You cannot directly use NVCC since that would require all dependent libraries to support CUDA (which is the issue you are seeing pop up from Fusion).

[BennyRod]

Thank you so much for your response! Could you explain me why this code works in Ubuntu using a CMakeLists.txt? I have included the CMakeLists that I am using. Thank you so much in advance.

# A CMake version 3.10 or higher is required 
cmake_minimum_required(VERSION 3.10 FATAL_ERROR)

# set the project language to C++
project(main LANGUAGES CUDA CXX)

# set all targets to use C++14
set(CMAKE_CXX_STANDARD 14)

# it is an error if C++14 is not available 
set(CMAKE_CXX_STANDARD_REQUIRED ON)

# set the CUDA architectures
#set(CMAKE_CUDA_ARCHITECTURES 60 70 75 80)

# set the output directory for executable files
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY "${CMAKE_CURRENT_LIST_DIR}/../bin")

#store the local location of Boost 
set(BOOST_ROOT "/usr/local/apps/boost/1.84/include/")

# add the local include directory to the include path
include_directories(${BOOST_ROOT})

# find Boost library
find_package(Boost 1.84.0 REQUIRED)

# Add CUDA files to the project
set(CUDA_FILES
     main.cu
)

# Tell CMake to compile the CUDA files
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-O3)

#create the executable
add_executable(main ${CUDA_FILES} )

# allow separate compilation. it is neccessary for multiple .cu files
set_target_properties(main PROPERTIES CUDA_SEPARABLE_COMPILATION ON)

[mborland]

Presumably Ubuntu is using GCC as the host compiler. Sometimes invalid code (e.g. not marked with __device__) will compile just fine, but then at run time you'll get all kinds of failures. Since you're not calling any functions it's like it's not instantiating any templates for you to see an error.

[BennyRod]

I have been testing one of the Odeint examples, the Harmonic Oscillator: https://github.com/headmyshoulder/odeint-v2/blob/master/examples/harmonic_oscillator.cpp It was compiled using the C++ and nvcc compiler individually but in both cases, it compiled and runned properly. Also, the numeric results were the same! I have included the used CMakeLists.txt in both cases and the outputs. Please, letme know what do you think about this, I am really interested in using Odeint in Windows.

HarmonicOscillator:

/*
Copyright 2010-2012 Karsten Ahnert
Copyright 2011-2013 Mario Mulansky
Copyright 2013 Pascal Germroth

Distributed under the Boost Software License, Version 1.0.
(See accompanying file LICENSE_1_0.txt or
copy at http://www.boost.org/LICENSE_1_0.txt)
*/

#include <iostream>
#include <vector>

#include <boost/numeric/odeint.hpp>

//[ rhs_function
/* The type of container used to hold the state vector */
typedef std::vector<double> state_type;

const double gam = 0.15;

/* The rhs of x' = f(x) */
void harmonic_oscillator(const state_type &x, state_type &dxdt, const double /* t */)
{
   dxdt[0] = x[1];
   dxdt[1] = -x[0] - gam * x[1];
}
//]

//[ rhs_class
/* The rhs of x' = f(x) defined as a class */
class harm_osc
{

   double m_gam;

public:
   harm_osc(double gam) : m_gam(gam) {}

   void operator()(const state_type &x, state_type &dxdt, const double /* t */)
   {
       dxdt[0] = x[1];
       dxdt[1] = -x[0] - m_gam * x[1];
   }
};
//]

//[ integrate_observer
struct push_back_state_and_time
{
   std::vector<state_type> &m_states;
   std::vector<double> &m_times;

   push_back_state_and_time(std::vector<state_type> &states, std::vector<double> &times)
       : m_states(states), m_times(times) {}

   void operator()(const state_type &x, double t)
   {
       m_states.push_back(x);
       m_times.push_back(t);
   }
};
//]

struct write_state
{
   void operator()(const state_type &x) const
   {
       std::cout << x[0] << "\t" << x[1] << "\n";
   }
};

int main(int /* argc */, char ** /* argv */)
{
   using namespace std;
   using namespace boost::numeric::odeint;

   //[ state_initialization
   state_type x(2);
   x[0] = 1.0; // start at x=1.0, p=0.0
   x[1] = 0.0;
   //]

   //[ integration
   size_t steps = integrate(harmonic_oscillator,
                            x, 0.0, 10.0, 0.1);
   //]

   //[ integration_class
   harm_osc ho(0.15);
   steps = integrate(ho,
                     x, 0.0, 10.0, 0.1);
   //]

   //[ integrate_observ
   vector<state_type> x_vec;
   vector<double> times;

   steps = integrate(harmonic_oscillator,
                     x, 0.0, 10.0, 0.1,
                     push_back_state_and_time(x_vec, times));

   /* output */
   for (size_t i = 0; i <= steps; i++)
   {
       cout << times[i] << '\t' << x_vec[i][0] << '\t' << x_vec[i][1] << '\n';
   }
   //]

   //[ define_const_stepper
   runge_kutta4<state_type> stepper;
   integrate_const(stepper, harmonic_oscillator, x, 0.0, 10.0, 0.01);
   //]

   //[ integrate_const_loop
   const double dt = 0.01;
   for (double t = 0.0; t < 10.0; t += dt)
       stepper.do_step(harmonic_oscillator, x, t, dt);
   //]

   //[ define_adapt_stepper
   typedef runge_kutta_cash_karp54<state_type> error_stepper_type;
   //]

   //[ integrate_adapt
   typedef controlled_runge_kutta<error_stepper_type> controlled_stepper_type;
   controlled_stepper_type controlled_stepper;
   integrate_adaptive(controlled_stepper, harmonic_oscillator, x, 0.0, 10.0, 0.01);
   //]

   {
       //[integrate_adapt_full
       double abs_err = 1.0e-10, rel_err = 1.0e-6, a_x = 1.0, a_dxdt = 1.0;
       controlled_stepper_type controlled_stepper(
           default_error_checker<double, range_algebra, default_operations>(abs_err, rel_err, a_x, a_dxdt));
       integrate_adaptive(controlled_stepper, harmonic_oscillator, x, 0.0, 10.0, 0.01);
       //]
   }

   //[integrate_adapt_make_controlled
   integrate_adaptive(make_controlled<error_stepper_type>(1.0e-10, 1.0e-6),
                      harmonic_oscillator, x, 0.0, 10.0, 0.01);
   //]

   //[integrate_adapt_make_controlled_alternative
   integrate_adaptive(make_controlled(1.0e-10, 1.0e-6, error_stepper_type()),
                      harmonic_oscillator, x, 0.0, 10.0, 0.01);
   //]

#ifdef BOOST_NUMERIC_ODEINT_CXX11
   //[ define_const_stepper_cpp11
   {
       runge_kutta4<state_type> stepper;
       integrate_const(stepper, [](const state_type &x, state_type &dxdt, double t)
                       {
           dxdt[0] = x[1]; dxdt[1] = -x[0] - gam*x[1]; }, x, 0.0, 10.0, 0.01);
   }
   //]

   //[ harm_iterator_const_step]
   std::for_each(make_const_step_time_iterator_begin(stepper, harmonic_oscillator, x, 0.0, 0.1, 10.0),
                 make_const_step_time_iterator_end(stepper, harmonic_oscillator, x),
                 [](std::pair<const state_type &, const double &> x)
                 { cout << x.second << " " << x.first[0] << " " << x.first[1] << "\n"; });
//]
#endif
}

C++:

# A CMake version 3.10 or higher is required 
cmake_minimum_required(VERSION 3.10 FATAL_ERROR)

# set the project language to C++
project(main LANGUAGES CUDA CXX)

# set all targets to use C++14
set(CMAKE_CXX_STANDARD 14)

# it is an error if C++14 is not available 
set(CMAKE_CXX_STANDARD_REQUIRED ON)

# set the CUDA architectures
#set(CMAKE_CUDA_ARCHITECTURES 60 70 75 80)

# set the output directory for executable files
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY "${CMAKE_CURRENT_LIST_DIR}/../bin")

#store the local location of Boost 
set(BOOST_ROOT "/usr/local/apps/boost/1.84/include/")

# add the local include directory to the include path
include_directories(${BOOST_ROOT})

# find Boost library
find_package(Boost 1.84.0 REQUIRED)

# Add CUDA files to the project
set(CPP_FILES
     main.cpp
)

# Tell CMake to compile the CUDA files
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-O3)

#create the executable
add_executable(main ${CPP_FILES} )

# allow separate compilation. it is neccessary for multiple .cu files
set_target_properties(main PROPERTIES CUDA_SEPARABLE_COMPILATION ON)

Output:

CUDA:

# A CMake version 3.10 or higher is required 
cmake_minimum_required(VERSION 3.10 FATAL_ERROR)

# set the project language to C++
project(main LANGUAGES CUDA CXX)

# set all targets to use C++14
set(CMAKE_CXX_STANDARD 14)

# it is an error if C++14 is not available 
set(CMAKE_CXX_STANDARD_REQUIRED ON)

# set the CUDA architectures
#set(CMAKE_CUDA_ARCHITECTURES 60 70 75 80)

# set the output directory for executable files
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY "${CMAKE_CURRENT_LIST_DIR}/../bin")

#store the local location of Boost 
set(BOOST_ROOT "/usr/local/apps/boost/1.84/include/")

# add the local include directory to the include path
include_directories(${BOOST_ROOT})

# find Boost library
find_package(Boost 1.84.0 REQUIRED)

# Add CUDA files to the project
set(CUDA_FILES
     main.cu
)

# Tell CMake to compile the CUDA files
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-O3)

#create the executable
add_executable(main ${CUDA_FILES} )

# allow separate compilation. it is neccessary for multiple .cu files
set_target_properties(main PROPERTIES CUDA_SEPARABLE_COMPILATION ON)

[mborland]

You need to run one of the examples out of this folder: https://github.com/headmyshoulder/odeint-v2/tree/master/examples/thrust. The results are the same since no CUDA was used.

[BennyRod]

Thank you for your continuous responses! My point is that even not using any CUDA code, the example does not compile on Window, but it does on Ubuntu. As you pointed out before, the host code must be compiled using the C++ compiler (this is why both outputs are the same) then I don't need to use Thrust because I don't have any CUDA code. I hope my point is clearer now.

[mborland]

Thank you for your continuous responses! My point is that even not using any CUDA code, the example does not compile on Window, but it does on Ubuntu. As you pointed out before, the host code must be compiled using the C++ compiler (this is why both outputs are the same) then I don't need to use Thrust because I don't have any CUDA code. I hope my point is clearer now.

Unfortunately I have no experience with CUDA on windows. I have only ever run it on Ubuntu and Red Hat platforms.

[BennyRod]

Ok no problem. I would like to thank you because you have been very kind! Have a great day!

[BennyRod]

I have a brute force solution: https://github.com/boostorg/fusion/issues/269