#include <map>
#include <deque>
#include <pthread.h>
#include "sort.hpp"

// A set of precomputed WaksmanNetworks of a given size
struct SizedWNs {
    pthread_mutex_t mutex;
    std::deque<WaksmanNetwork> wns;
};

// A (mutexed) map mapping sizes to SizedWNs
struct PrecompWNs {
    pthread_mutex_t mutex;
    std::map<uint32_t,SizedWNs> sized_wns;
};

static PrecompWNs precomp_wns;

// A (mutexed) map mapping (N, nthreads) pairs to WNEvalPlans
struct EvalPlans {
    pthread_mutex_t mutex;
    std::map<std::pair<uint32_t,threadid_t>,WNEvalPlan> eval_plans;
};

static EvalPlans precomp_eps;

void sort_precompute(uint32_t N)
{
    uint32_t *random_permutation = new uint32_t[N];
    if (!random_permutation) {
        printf("Allocating memory failed in sort_precompute\n");
    }
    assert(random_permutation);
    for (uint32_t i=0;i<N;++i) {
        random_permutation[i] = i;
    }
    RecursiveShuffle_M2((unsigned char *) random_permutation, N, sizeof(uint32_t));

    WaksmanNetwork wnet(N);
    wnet.setPermutation(random_permutation);

    // Note that sized_wns[N] creates a map entry for N if it doesn't yet exist
    pthread_mutex_lock(&precomp_wns.mutex);
    SizedWNs& szwn = precomp_wns.sized_wns[N];
    pthread_mutex_unlock(&precomp_wns.mutex);
    pthread_mutex_lock(&szwn.mutex);
    szwn.wns.push_back(std::move(wnet));
    pthread_mutex_unlock(&szwn.mutex);
}

void sort_precompute_evalplan(uint32_t N, threadid_t nthreads)
{
    std::pair<uint32_t,threadid_t> idx = {N, nthreads};
    pthread_mutex_lock(&precomp_eps.mutex);
    if (!precomp_eps.eval_plans.count(idx)) {
        precomp_eps.eval_plans.try_emplace(idx, N, nthreads);
    }
    pthread_mutex_unlock(&precomp_eps.mutex);
}

// Perform the sort using up to nthreads threads.  The items to sort are
// byte arrays of size msg_size.  The key is the first 4 bytes of each
// item.
void sort_mtobliv(threadid_t nthreads, uint8_t* items, uint16_t msg_size,
    uint32_t Nr, uint32_t Na,
    // the arguments to the callback are nthreads, items, the sorted
    // indices, and the number of non-padding items
    std::function<void(threadid_t, const uint8_t*, const uint64_t*,
        uint32_t Nr)> cb)
{
    // Find the smallest Nw for which we have a precomputed
    // WaksmanNetwork with Nr <= Nw <= Na
    pthread_mutex_lock(&precomp_wns.mutex);
    std::optional<WaksmanNetwork> wn;
    uint32_t Nw;
    for (auto& N : precomp_wns.sized_wns) {
        if (N.first > Na) {
            printf("No precomputed WaksmanNetworks of size at most %u\n", Na);
            assert(false);
        }
        if (N.first < Nr) {
            continue;
        }
        // We're in the right range, but see if we have an actual
        // precomputed WaksmanNetwork
        pthread_mutex_lock(&N.second.mutex);
        if (N.second.wns.size() == 0) {
            pthread_mutex_unlock(&N.second.mutex);
            continue;
        }
        wn = std::move(N.second.wns.front());
        N.second.wns.pop_front();
        Nw = N.first;
        pthread_mutex_unlock(&N.second.mutex);
        break;
    }
    pthread_mutex_unlock(&precomp_wns.mutex);
    if (!wn) {
        printf("No precomputed WaksmanNetwork of size range [%u,%u] found.\n",
            Nr, Na);
        assert(wn);
    }
    std::pair<uint32_t,threadid_t> epidx = {Nw, nthreads};
    pthread_mutex_lock(&precomp_eps.mutex);
    if (!precomp_eps.eval_plans.count(epidx)) {
        printf("No precomputed WNEvalPlan with N=%u, nthreads=%hu\n",
            Nw, nthreads);
        assert(false);
    }
    const WNEvalPlan &eval_plan = precomp_eps.eval_plans.at(epidx);
    pthread_mutex_unlock(&precomp_eps.mutex);
    // Mark Nw-Nr items as padding (Nr, Na, and Nw are _not_ private)
    for (uint32_t i=Nr; i<Nw; ++i) {
        (*(uint32_t*)(items+msg_size*i)) = uint32_t(-1);
    }
    // Shuffle Nw items
    wn.value().applyInversePermutation<OSWAP_16X>(
        items, msg_size, eval_plan);

    // Create the indices
    uint64_t *idx = new uint64_t[Nr];
    uint64_t *nextidx = idx;
    for (uint32_t i=0; i<Nw; ++i) {
        uint64_t key = (*(uint32_t*)(items+msg_size*i));
        if (key != uint32_t(-1)) {
            *nextidx = (key<<32) + i;
            ++nextidx;
        }
    }
    if (nextidx != idx + Nr) {
        printf("Found %u non-padding items, expected %u\n",
            nextidx-idx, Nr);
        assert(nextidx == idx + Nr);
    }
    // Sort the keys and indices
    uint64_t *backingidx = new uint64_t[Nr];
    bool whichbuf = mtmergesort<uint64_t>(idx, Nr, backingidx, nthreads);
    uint64_t *sortedidx = whichbuf ? backingidx : idx;
    for (uint32_t i=0; i<Nr; ++i) {
        sortedidx[i] &= uint64_t(0xffffffff);
    }
    cb(nthreads, items, sortedidx, Nr);

    delete[] idx;
    delete[] backingidx;
}