OracularSlabCache.hpp

Go to the documentation of this file.

#ifndef TATAMI_CHUNKED_ORACULAR_SLAB_CACHE_HPP
#define TATAMI_CHUNKED_ORACULAR_SLAB_CACHE_HPP
 
#include "utils.hpp"
 
#include <unordered_map>
#include <vector>
#include <list>
#include <type_traits>
#include <memory>
#include <cstddef>
 
#include "tatami/tatami.hpp"
#include "sanisizer/sanisizer.hpp"
 
namespace tatami_chunked {
 
template<typename Id_, typename Index_, class Slab_, bool track_reuse_ = false> 
class OracularSlabCache {
private:
    std::shared_ptr<const tatami::Oracle<Index_> > my_oracle;
    tatami::PredictionIndex my_total;
    tatami::PredictionIndex my_counter = 0;
 
    Id_ my_last_slab_id = 0;
    Slab_* my_last_slab = NULL;
 
    typedef std::vector<Slab_> SlabPool;
    typename SlabPool::size_type my_max_slabs;
    SlabPool my_all_slabs;
 
    std::unordered_map<Id_, Slab_*> my_current_cache, my_future_cache;
    std::vector<std::pair<Id_, Slab_*> > my_to_populate;
    std::vector<Id_> my_in_need;
    tatami::PredictionIndex my_refresh_point = 0;
 
    typename std::conditional<track_reuse_, std::vector<std::pair<Id_, Slab_*> >, bool>::type my_to_reuse;
 
public:
    template<typename MaxSlabs_>
    OracularSlabCache(std::shared_ptr<const tatami::Oracle<Index_> > oracle, MaxSlabs_ max_slabs) : 
        my_oracle(std::move(oracle)), 
        my_total(my_oracle->total()),
        my_max_slabs(sanisizer::cast<I<decltype(my_max_slabs)> >(max_slabs)) 
    {
        my_all_slabs.reserve(max_slabs);
        my_current_cache.reserve(max_slabs);
        my_future_cache.reserve(max_slabs);
    } 
 
    OracularSlabCache(const OracularSlabCache&) = delete;
 
    OracularSlabCache& operator=(const OracularSlabCache&) = delete;
 
    // Move operators are still okay as pointers still point to the moved vectors.
    // see https://stackoverflow.com/questions/43988553/stdvector-stdmove-and-pointer-invalidation.
    OracularSlabCache& operator=(OracularSlabCache&&) = default;
    OracularSlabCache(OracularSlabCache&&) = default;
 
    // Might as well define this.
    ~OracularSlabCache() = default;
public:
    Index_ next() {
        return my_oracle->get(my_counter++);
    }
 
public:
    template<class Ifunction_, class Cfunction_, class Pfunction_>
    std::pair<const Slab_*, Index_> next(Ifunction_ identify, Cfunction_ create, Pfunction_ populate) {
        Index_ index = this->next(); 
        auto slab_info = identify(index);
        if (slab_info.first == my_last_slab_id && my_last_slab) {
            return std::make_pair(my_last_slab, slab_info.second);
        }
        my_last_slab_id = slab_info.first;
 
        // Updating the cache if we hit the refresh point.
        if (my_counter - 1 == my_refresh_point) {
            // Note that, for any given populate cycle, the first prediction's
            // slab cannot already be in the cache, otherwise it would have
            // incorporated into the previous cycle. So we can skip some code.
            my_future_cache[slab_info.first] = NULL;
            my_in_need.push_back(slab_info.first);
            I<decltype(my_max_slabs)> used_slabs = 1;
            auto last_future_slab_id = slab_info.first;
 
            while (++my_refresh_point < my_total) {
                auto future_index = my_oracle->get(my_refresh_point);
                auto future_slab_info = identify(future_index);
                if (last_future_slab_id == future_slab_info.first) {
                    continue;
                }
 
                last_future_slab_id = future_slab_info.first;
                if (my_future_cache.find(future_slab_info.first) != my_future_cache.end()) {
                    continue;
                }
 
                if (used_slabs == my_max_slabs) {
                    break;
                } 
                ++used_slabs;
 
                auto ccIt = my_current_cache.find(future_slab_info.first);
                if (ccIt == my_current_cache.end()) {
                    my_future_cache[future_slab_info.first] = NULL;
                    my_in_need.push_back(future_slab_info.first);
 
                } else {
                    auto slab_ptr = ccIt->second;
                    my_future_cache[future_slab_info.first] = slab_ptr;
                    my_current_cache.erase(ccIt);
                    if constexpr(track_reuse_) {
                        my_to_reuse.emplace_back(future_slab_info.first, slab_ptr);
                    }
                }
            }
 
            auto cIt = my_current_cache.begin();
            for (auto a : my_in_need) {
                if (cIt != my_current_cache.end()) {
                    my_to_populate.emplace_back(a, cIt->second);
                    my_future_cache[a] = cIt->second;
                    ++cIt;
                } else {
                    // We reserved my_all_slabs so further push_backs() should not 
                    // trigger any reallocation or invalidation of the pointers.
                    my_all_slabs.push_back(create());
                    auto slab_ptr = &(my_all_slabs.back());
                    my_to_populate.emplace_back(a, slab_ptr);
                    my_future_cache[a] = slab_ptr;
                }
            }
            my_in_need.clear();
 
            if constexpr(track_reuse_) {
                populate(my_to_populate, my_to_reuse);
            } else {
                populate(my_to_populate);
            }
 
            my_to_populate.clear();
            if constexpr(track_reuse_) {
                my_to_reuse.clear();
            }
 
            // We always fill my_future_cache to the brim so every entry of
            // my_all_slabs should be referenced by a pointer in
            // my_future_cache.  There shouldn't be any free cache entries
            // remaining in my_current_cache i.e., at this point, cIt should
            // equal my_current_cache.end(), as we transferred everything to
            // my_future_cache. Thus it is safe to clear my_current_cache
            // without worrying about leaking memory. The only exception is if
            // we run out of predictions, in which case it doesn't matter.
            my_current_cache.clear();
            my_current_cache.swap(my_future_cache);
        }
 
        // We know it must exist, so no need to check ccIt's validity.
        auto ccIt = my_current_cache.find(slab_info.first);
        my_last_slab = ccIt->second;
        return std::make_pair(my_last_slab, slab_info.second);
    }
 
public:
    auto get_max_slabs() const {
        return my_max_slabs;
    }
 
    auto get_num_slabs() const {
        return my_current_cache.size();
    }
};
 
}
 
#endif