error: redefinition of ‘constexpr const bool stdgpu::numeric_limits<T>::is_specialized’ about stdgpu

[stotko]

Does this error also occur with different host compilers (GCC, Clang) and different CUDA versions (11.3, etc.)? Furthermore, is this the only error you are getting? Or are there also errors regarding the is_signed, is_integer, etc. member variables? If possible, could you attach a (minimal) reproducable example?

[geotyper]

Try only GCC and CUDA 11.4, this is my work setup.
Have 5 errors, as list of this variables:

/**
   * \brief Whether the type is signed
   */
  static constexpr bool is_signed = false;
  /**
   * \brief Whether the type is an integer
   */
  static constexpr bool is_integer = true;
  /**
   * \brief Whether the type is exact
   */
  static constexpr bool is_exact = true;
  /**
   * \brief Number of radix digits
   */
  static constexpr int digits = 1;
  /**
   * \brief Integer base
   */
  static constexpr int radix = 2;

Add in work project #include header from example vector.cu

#include <iostream>
#include <thrust/copy.h>
#include <thrust/reduce.h>
#include <thrust/sequence.h>

#include <stdgpu/memory.h>          // createDeviceArray, destroyDeviceArray
#include <stdgpu/iterator.h>        // device_begin, device_end
#include <stdgpu/platform.h>        // STDGPU_HOST_DEVICE
#include <stdgpu/vector.cuh>        // stdgpu::vector

and two lines of code

stdgpu::index_t n = 100;
int* d_input = createDeviceArray<int>(n);

In cmakelists.txt add (CUDA is active and work in project):
find_package(stdgpu REQUIRED)
target_link_libraries(${EXECUTABLE_NAME} PUBLIC stdgpu::stdgpu)

can't understand what to write in add_library(foo ... ) so don't add it
if add include_directory with path to include directory of stdgpu, errors more than 200

[geotyper]

this errors if add include_directory with headers

[stotko]

if add include_directory with path to include directory of stdgpu, errors more than 200

Could you elaborate what you exactly mean by that? The lines

find_package(stdgpu REQUIRED)
target_link_libraries(${EXECUTABLE_NAME} PUBLIC stdgpu::stdgpu)

are sufficient to add all necessary properties stdgpu needs add to your project, including the include directory paths.

It seems like this is an error/incompatibility in your project. stdgpu requires compilation with C++14 and should enforce this by its imported cxx_std_14 property. Did you set the C++ standard somewhere in your project? Could you provide more details on your project? In particular, the build configuration, i.e. the CMakeLists.txt file(s), would be helpful to know (if it is possible for you to attach them).

[geotyper]

Made small project to solve problem, under some changes in cmakelists.txt all work as expected:

cmake_minimum_required(VERSION 3.5)

if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
  set(CMAKE_CUDA_ARCHITECTURES 61)
  message(CMAKE_CUDA_ARCHITECTURES)
endif()

project(CellInstancing LANGUAGES CUDA CXX)

set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)

set(CMAKE_CUDA_STANDARD 14)
set(CMAKE_CUDA_STANDARD_REQUIRED ON)
set(CMAKE_CUDA_SEPARABLE_COMPILATION ON)


set(EXECUTABLE_NAME "CUDA_exp")

find_package(CUDA REQUIRED)


SET(TARGET_H
)

SET(TARGET_SRC
main.cu
)

SET(TARGET_CUDA_H
cuda_exp.cuh
)

SET(TARGET_CUDA_SRC
cuda_exp.cu
)

find_package(stdgpu REQUIRED)

set(Cuda_common_include "/usr/local/cuda/samples/common/inc")
include_directories(${Cuda_common_include})

message(${CUDA_INCLUDE_DIRS})
include_directories(${CUDA_INCLUDE_DIRS})

add_executable(${EXECUTABLE_NAME} ${TARGET_SRC} ${TARGET_H} ${TARGET_CUDA_H} ${TARGET_CUDA_SRC} )

set_target_properties(${EXECUTABLE_NAME} PROPERTIES CUDA_ARCHITECTURES "61")
target_link_libraries(${EXECUTABLE_NAME} ${CUDA_LIBRARIES})
target_link_libraries(${EXECUTABLE_NAME} stdgpu::stdgpu)