Array crash using RelWithDebInfo with CUDA.

[BradWhitlock]

I think I was able to make a reproducer for an axom::Array crash. Build the following example program on rzansel. It will build in RelWithDebInfo mode and will crash in the axom::Array constructor trying to default-initialize memory with placement new on the host (but the memory should be on the GPU).

In thinking about this some more, having axom::Array constructors in a compiled .cpp file might be the real problem. In the original scenario, I did not do anything special in my CMakeLists.txt to indicate that the file has GPU code; I only added the source file to my library target. In this example, the bug.cpp file is marked as CUDA code so axom::CUDA_EXEC will be available. The execute.cpp file where the error happens is not marked as a GPU file, though some code in axom::Array might need it to be.

bug.cpp

#include <axom/config.hpp>
#include <axom/core.hpp>

// Forward declaration
namespace axom
{
namespace mir
{
namespace views
{
void buildShapeMap(axom::Array<axom::IndexType> &indices,
                   axom::Array<axom::IndexType> &values,
                   int allocatorID);
} // end namespace views
} // end namespace mir
} // end namespace axom

template <typename ExecSpace>
void execute()
{
  axom::Array<axom::IndexType> indices, values;
  int allocatorID = axom::execution_space<ExecSpace>::allocatorID();
  axom::mir::views::buildShapeMap(/*node,*/ indices, values, allocatorID);

  std::cout << "indices.size=" << indices.size() << std::endl;
  std::cout << "values.size=" << values.size() << std::endl;
}

int main(int argc, char *argv[])
{
#if defined(AXOM_USE_CUDA)
  constexpr int BLOCK_SIZE = 256;
  using cuda_exec = axom::CUDA_EXEC<BLOCK_SIZE>;
  execute<cuda_exec>();
#endif

  return 0;
}

execute.cpp

#include <axom/config.hpp>
#include <axom/core.hpp>

#include <vector>

// Forward declaration
namespace axom
{
namespace mir
{
namespace views
{
void buildShapeMap(axom::Array<axom::IndexType> &indices,
                   axom::Array<axom::IndexType> &values,
                   int allocatorID)
{
  std::vector<axom::IndexType> srcIndices{{0,1,2,3,4,5}};
  std::vector<axom::IndexType> srcValues{{5,4,3,2,1,0}};
  const axom::IndexType n = srcValues.size();

  // The error would manifest here when calling Array::Array(n, n, allocatorID).
  indices = axom::Array<axom::IndexType>(n, n, allocatorID);
  values = axom::Array<axom::IndexType>(n, n, allocatorID);
  axom::copy(indices.data(), srcIndices.data(), n * sizeof(axom::IndexType));
  axom::copy(values.data(), srcValues.data(), n * sizeof(axom::IndexType));
}

} // end namespace views
} // end namespace mir
} // end namespace axom

This file was hacked together from the QuickStart guide and I made some additions so it would work with CUDA. I would have hoped that Axom would take care of that but, it didn't.

CMakeLists.txt

cmake_minimum_required(VERSION 3.21)

set(CMAKE_C_COMPILER "/usr/tce/packages/clang/clang-ibm-10.0.1-gcc-8.3.1/bin/clang" CACHE PATH "")
set(CMAKE_CXX_COMPILER "/usr/tce/packages/clang/clang-ibm-10.0.1-gcc-8.3.1/bin/clang++" CACHE PATH "")
#set(CMAKE_BUILD_TYPE Debug)
#set(CMAKE_BUILD_TYPE Release)
set(CMAKE_BUILD_TYPE RelWithDebInfo)

project(bug)

# Point to an installed RelWithDebInfo Axom
set(AXOM_DIR /usr/WS2/whitlocb/Axom/axom_mir/axom/[email protected]_cuda-relwithdebinfo)

#------------------------------------------------------------------------------
# Check for AXOM_DIR and use CMake's find_package to import axom's targets
#------------------------------------------------------------------------------
if(NOT DEFINED AXOM_DIR OR NOT EXISTS ${AXOM_DIR}/lib/cmake/axom-config.cmake)
    message(FATAL_ERROR "Missing required 'AXOM_DIR' variable pointing to an installed axom")
endif()

if (ENABLE_CUDA)
    enable_language(CUDA)
endif()

if (ENABLE_HIP)
    if (NOT ROCM_PATH)
        find_path(ROCM_PATH
            hip
            ENV{ROCM_DIR}
            ENV{ROCM_PATH}
            ENV{HIP_PATH}
            ${HIP_PATH}/..
            ${HIP_ROOT_DIR}/../
            ${ROCM_ROOT_DIR}
            /opt/rocm)
    endif()
    set(CMAKE_PREFIX_PATH "${CMAKE_PREFIX_PATH};${ROCM_PATH}")
    find_package(hip REQUIRED CONFIG PATHS ${ROCM_PATH})
endif()

include(CMakeFindDependencyMacro)

# 70=Volta
set(CMAKE_CUDA_ARCHITECTURES "70")

find_dependency(axom REQUIRED
                NO_DEFAULT_PATH 
                PATHS ${AXOM_DIR}/lib/cmake)


add_executable(bug bug.cpp execute.cpp)

set_source_files_properties(bug.cpp PROPERTIES
    LANGUAGE CUDA
)

set_target_properties(bug PROPERTIES
    CUDA_SEPARABLE_COMPILATION ON
)
target_link_libraries(bug axom::core ${CUDA_LIBRARIES})