Function buffered

src/Common/examples/average.cpp:442–477

Float NO_INLINE buffered(const PODArray<UInt8> & keys, const PODArray<Float> & values)
{
    State map[256]{};

    static constexpr size_t BUF_SIZE = 16384 / 256 / sizeof(Float); /// Should fit in L1d.

    Float buffers[256 * BUF_SIZE];
    Float * ptrs[256];

    for (size_t i = 0; i < 256; ++i)
        ptrs[i] = &buffers[i * BUF_SIZE];

    size_t size = keys.size();
    const auto * key = keys.data();
    const auto * key_end = key + size;
    const auto * value = values.data();

    while (key < key_end)
    {
        *ptrs[*key] = *value;

        if (++ptrs[*key] == &buffers[(*key + 1) * BUF_SIZE])  /// Calculation is better than L1d load of cached end pointer.
        {
            ptrs[*key] -= BUF_SIZE;
            map[*key].addBatch<BUF_SIZE>(ptrs[*key], BUF_SIZE);
        }

        ++key;
        ++value;
    }

    for (size_t i = 0; i < 256; ++i)
        map[i].addBatch<4>(&buffers[i * BUF_SIZE], ptrs[i] - &buffers[i * BUF_SIZE]);

    return map[0].result();
}

#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wframe-larger-than"