alpaka::WorkDivMembers generates incomplete type errors

[frobnitzem]

I get the following error message when trying to invoke a kernel created with WorkDivMembers (to manually set launch parameters). I can't determine what this error means, but I suspect the WorkDivMembers might not be used as much as getValidWorkDiv, so perhaps the former is less advanced?

/inst/include/alpaka/dim/Traits.hpp(27): error #70: incomplete type is not allowed
          detected during:
            instantiation of type "alpaka::Dim<alpaka::DevUniformCudaHipRt>" 
/inst/include/alpaka/kernel/Traits.hpp(275): here
            instantiation of "auto alpaka::createTaskKernel<TAcc,TWorkDiv,TKernelFnObj,TArgs...>(const TWorkDiv &, const TKernelFnObj &, TArgs &&...) [with TAcc=alpaka::DevUniformCudaHipRt, TWorkDiv=alpaka::WorkDivMembers<Dim, Idx>, TKernelFnObj=VectorAddKernel, TArgs=<>]" 
/proj/src/vecMain.cpp(52): here
            instantiation of "int vecAdd(const Acc &, TQueue &, const TVec &) [with Acc=alpaka::DevUniformCudaHipRt, TQueue=alpaka::Queue<alpaka::ExampleDefaultAcc<Dim, Idx>, alpaka::property::Blocking>, TVec=Vec]" 
/proj/src/vecMain.cpp(73): here

/inst/include/alpaka/kernel/Traits.hpp(275): error #276: name followed by "::" must be a class or namespace name
          detected during:
            instantiation of "auto alpaka::createTaskKernel<TAcc,TWorkDiv,TKernelFnObj,TArgs...>(const TWorkDiv &, const TKernelFnObj &, TArgs &&...) [with TAcc=alpaka::DevUniformCudaHipRt, TWorkDiv=alpaka::WorkDivMembers<Dim, Idx>, TKernelFnObj=VectorAddKernel, TArgs=<>]

...

The reproducing code:

#include <alpaka/alpaka.hpp>
#include <alpaka/example/ExampleDefaultAcc.hpp>

#include <iostream>
#include <random>
#include <typeinfo>

using Dim = alpaka::DimInt<1u>;
using Idx = std::uint32_t;
using Vec = alpaka::Vec<Dim, Idx>;

//#############################################################################
//! A vector addition kernel.
class VectorAddKernel
{
public:
    ALPAKA_NO_HOST_ACC_WARNING
    template<typename TAcc>
    ALPAKA_FN_ACC auto operator()(TAcc const& acc) const -> void { }
};

/* Declare and run the VectorAddKernel */
template<typename Acc, typename TQueue, typename TVec>
int vecAdd(const Acc &devAcc, TQueue &queue, const TVec &numElements) {
    // Launch with one warp per thread block:
    Idx const warpExtent = alpaka::getWarpSize(devAcc);
    TVec const gridBlockExtent = TVec::all(Idx(numElements[0] / warpExtent));
    TVec blockThreadExtent = TVec::all(Idx(warpExtent));
    TVec elemExtent = TVec::all(Idx(1));

    alpaka::WorkDivMembers<Dim, Idx> workDiv{gridBlockExtent, blockThreadExtent, elemExtent};

    // Create the kernel execution task.
    VectorAddKernel K{};
    auto const taskKernel = alpaka::createTaskKernel<Acc>(workDiv, K);

    // Enqueue the kernel execution task
    alpaka::enqueue(queue, taskKernel);

    return 0;
}

int main() {
    using Acc = alpaka::ExampleDefaultAcc<Dim, Idx>;
    auto const devAcc = alpaka::getDevByIdx<Acc>(0u);

    std::cout << "Using alpaka accelerator: " << alpaka::getAccName<Acc>() << std::endl;

    using QueueAcc = alpaka::Queue<Acc, alpaka::Blocking>;
    QueueAcc queue(devAcc);

    // Define the work division
    Idx const numElements(123456);
    const Vec extents(numElements);

    vecAdd(devAcc, queue, extents);

    return 0;
}

It was compiled with stock gcc 8.4.0-1ubuntu1~18.04 and cuda toolkit V10.2.89 using flags -DALPAKA_ACC_GPU_CUDA_ENABLE=ON -DALPAKA_ACC_GPU_CUDA_ONLY_MODE=ON and:

cmake_minimum_required(VERSION 3.15)

set(_TARGET_NAME pairEn)

project(${_TARGET_NAME})

#-------------------------------------------------------------------------------
# Find alpaka.

find_package(alpaka REQUIRED)

#-------------------------------------------------------------------------------
# Add executable.

alpaka_add_executable(
    ${_TARGET_NAME}
    src/vecMain.cpp)
target_compile_features(${_TARGET_NAME} PUBLIC cxx_std_14)
target_link_libraries(
    ${_TARGET_NAME}
    PUBLIC alpaka::alpaka)
install(TARGETS ${_TARGET_NAME} DESTINATION bin)

I also tried with cxx_std_11 and cxx_std_17 and got the same error.

[bernhardmgruber]

You are mixing up accelerators and devices.

vecAdd(devAcc, queue, extents);

template<typename Acc, typename TQueue, typename TVec>
int vecAdd(const Acc &devAcc, TQueue &queue, const TVec &numElements) {
...

You pass in devAcc which is a device and not an accelerator. Acc is deduced to the type of a device. When you call alpaka::createTaskKernel<Acc>(workDiv, K);, you instantiate the function with the type of a device and not with an accelerator. And the Dim trait seems to not be specialized for devices.

You can make a simple change:

vecAdd<Acc>(devAcc, queue, extents);

template<typename Acc, typename Dev, typename TQueue, typename TVec>
int vecAdd(const Dev&devAcc, TQueue &queue, const TVec &numElements) {
...

and it should work then (works for me).

[frobnitzem]

Confirmed - this fixes the issue, and solves related troubles I've been having.

I guess I assumed that devAcc would have the type Acc
in auto const devAcc = alpaka::getDevByIdx<Acc>(0u) but I should have checked.