Function TailImpl

catboost/cuda/cuda_util/helpers.cpp:125–139

template <class T>
static TDistributedObject<std::remove_const_t<T>> TailImpl(const TCudaBuffer<T, TStripeMapping>& data, ui32 stream) {
    Y_ASSERT(data.GetObjectsSlice().Size());
    auto result = TCudaBuffer<std::remove_const_t<T>, TStripeMapping, EPtrType::CudaHost>::Create(data.GetMapping().RepeatOnAllDevices(1, data.GetMapping().SingleObjectSize()));

    using TKernel = TTailKernel<std::remove_const_t<T>, EPtrType::CudaHost>;
    LaunchKernels<TKernel>(result.NonEmptyDevices(), stream, data, result);

    TVector<std::remove_const_t<T>> resultVec;
    result.Read(resultVec, stream);
    auto res = CreateDistributedObject<std::remove_const_t<T>>(0);
    for (ui32 i = 0; i < NCudaLib::GetCudaManager().GetDeviceCount(); ++i) {
        res.Set(i, resultVec[i]);
    }
    return res;
}

#define Y_CATBOOST_CUDA_F_IMPL(T)                                                                                 \
    template <>                                                                                                   \