#include <string>
#include "base/kaldi-common.h"
#include "util/common-utils.h"
#include "tree/context-dep.h"
#include "hmm/transition-model.h"
#include "fstext/fstext-lib.h"
#include "decoder/faster-decoder.h"
#include "decoder/decodable-matrix.h"
#include "lat/kaldi-lattice.h"
#include "lat/lattice-functions.h"

#include "nnet/nnet-trnopts.h"
#include "nnet/nnet-component.h"
#include "nnet/nnet-activation.h"
#include "nnet/nnet-nnet.h"
#include "nnet/nnet-pdf-prior.h"
#include "nnet/nnet-utils.h"
#include "base/timer.h"
#include "cudamatrix/cu-device.h"

typedef kaldi::BaseFloat BaseFloat;
typedef struct Matrix NervMatrix;

namespace kaldi {
    namespace nnet1 {

        void LatticeAcousticRescore(const Matrix<BaseFloat> &log_like,
                const TransitionModel &trans_model,
                const std::vector<int32> &state_times,
                Lattice *lat) {
            kaldi::uint64 props = lat->Properties(fst::kFstProperties, false);
            if (!(props & fst::kTopSorted))
                KALDI_ERR << "Input lattice must be topologically sorted.";

            KALDI_ASSERT(!state_times.empty());
            std::vector<std::vector<int32> > time_to_state(log_like.NumRows());
            for (size_t i = 0; i < state_times.size(); i++) {
                KALDI_ASSERT(state_times[i] >= 0);
                if (state_times[i] < log_like.NumRows())  // end state may be past this..
                    time_to_state[state_times[i]].push_back(i);
                else
                    KALDI_ASSERT(state_times[i] == log_like.NumRows()
                            && "There appears to be lattice/feature mismatch.");
            }

            for (int32 t = 0; t < log_like.NumRows(); t++) {
                for (size_t i = 0; i < time_to_state[t].size(); i++) {
                    int32 state = time_to_state[t][i];
                    for (fst::MutableArcIterator<Lattice> aiter(lat, state); !aiter.Done();
                            aiter.Next()) {
                        LatticeArc arc = aiter.Value();
                        int32 trans_id = arc.ilabel;
                        if (trans_id != 0) {  // Non-epsilon input label on arc
                            int32 pdf_id = trans_model.TransitionIdToPdf(trans_id);
                            arc.weight.SetValue2(-log_like(t, pdf_id) + arc.weight.Value2());
                            aiter.SetValue(arc);
                        }
                    }
                }
            }
        }

    }  // namespace nnet1
}  // namespace kaldi


extern "C" {
#include "kaldi_mpe.h"
#include "string.h"
#include "assert.h"
#include "nerv/lib/common.h"
#include "nerv/lib/matrix/mmatrix.h"

    extern NervMatrix *nerv_matrix_host_float_create(long nrow, long ncol, MContext *context, Status *status);
    extern void nerv_matrix_host_float_copy_fromd(NervMatrix *mat, const NervMatrix *cumat, int, int, int, Status *);
    using namespace kaldi;
    using namespace kaldi::nnet1;
    typedef kaldi::int32 int32;

    struct KaldiMPE {
        TransitionModel *trans_model;
        RandomAccessLatticeReader *den_lat_reader;
        RandomAccessInt32VectorReader *ref_ali_reader;

        Lattice den_lat;
        vector<int32> state_times;

        PdfPriorOptions *prior_opts;
        PdfPrior *log_prior;

        std::vector<int32> silence_phones;
        std::vector<int32> ref_ali;

        Timer *time;
        double time_now;

        int32 num_done, num_no_ref_ali, num_no_den_lat, num_other_error;

        kaldi::int64 total_frames;
        int32 num_frames;
        double total_frame_acc, utt_frame_acc;

        bool binary;
        bool one_silence_class;
        BaseFloat acoustic_scale, lm_scale, old_acoustic_scale;
        kaldi::int32 max_frames;
        bool do_smbr;
        std::string use_gpu;
    };

    KaldiMPE * new_KaldiMPE(const char* arg, const char* mdl, const char* lat, const char* ali)
    {
        KaldiMPE * mpe = new KaldiMPE;

        const char *usage =
            "Perform iteration of Neural Network MPE/sMBR training by stochastic "
            "gradient descent.\n"
            "The network weights are updated on each utterance.\n"
            "Usage:  nnet-train-mpe-sequential [options] <model-in> <transition-model-in> "
            "<feature-rspecifier> <den-lat-rspecifier> <ali-rspecifier> [<model-out>]\n"
            "e.g.: \n"
            " nnet-train-mpe-sequential nnet.init trans.mdl scp:train.scp scp:denlats.scp ark:train.ali "
            "nnet.iter1\n";

        ParseOptions po(usage);

        NnetTrainOptions trn_opts; trn_opts.learn_rate=0.00001;
        trn_opts.Register(&po);

        mpe->binary = true;
        po.Register("binary", &(mpe->binary), "Write output in binary mode");

        std::string feature_transform;
        po.Register("feature-transform", &feature_transform,
                "Feature transform in Nnet format");
        std::string silence_phones_str;
        po.Register("silence-phones", &silence_phones_str, "Colon-separated list "
                "of integer id's of silence phones, e.g. 46:47");

        mpe->prior_opts = new PdfPriorOptions;
        PdfPriorOptions &prior_opts = *(mpe->prior_opts);
        prior_opts.Register(&po);

        mpe->one_silence_class = false;
        mpe->acoustic_scale = 1.0,
            mpe->lm_scale = 1.0,
            mpe->old_acoustic_scale = 0.0;
        po.Register("acoustic-scale", &(mpe->acoustic_scale),
                "Scaling factor for acoustic likelihoods");
        po.Register("lm-scale", &(mpe->lm_scale),
                "Scaling factor for \"graph costs\" (including LM costs)");
        po.Register("old-acoustic-scale", &(mpe->old_acoustic_scale),
                "Add in the scores in the input lattices with this scale, rather "
                "than discarding them.");
        po.Register("one-silence-class", &(mpe->one_silence_class), "If true, newer "
                "behavior which will tend to reduce insertions.");
        mpe->max_frames = 6000; // Allow segments maximum of one minute by default
        po.Register("max-frames",&(mpe->max_frames), "Maximum number of frames a segment can have to be processed");
        mpe->do_smbr = false;
        po.Register("do-smbr", &(mpe->do_smbr), "Use state-level accuracies instead of "
                "phone accuracies.");

        mpe->use_gpu=std::string("yes");
        po.Register("use-gpu", &(mpe->use_gpu), "yes|no|optional, only has effect if compiled with CUDA");

        int narg = 0;
        char args[64][1024];
        char *token;
        char *saveptr = NULL;
        char tmpstr[1024];

        strcpy(tmpstr, arg);
        strcpy(args[0], "nnet-train-mpe-sequential");
        for(narg = 1, token = strtok_r(tmpstr, " ", &saveptr); token; token = strtok_r(NULL, " ", &saveptr))
            strcpy(args[narg++], token);
        strcpy(args[narg++], "0.nnet");
        strcpy(args[narg++], mdl);
        strcpy(args[narg++], "feat");
        strcpy(args[narg++], lat);
        strcpy(args[narg++], ali);
        strcpy(args[narg++], "1.nnet");

        char **argsv = new char*[narg];
        for(int _i = 0; _i < narg; _i++)
            argsv[_i] = args[_i];

        po.Read(narg, argsv);
        delete [] argsv;

        if (po.NumArgs() != 6) {
            po.PrintUsage();
            exit(1);
        }

        std::string transition_model_filename = po.GetArg(2),
            den_lat_rspecifier = po.GetArg(4),
            ref_ali_rspecifier = po.GetArg(5);

        std::vector<int32> &silence_phones = mpe->silence_phones;
        if (!kaldi::SplitStringToIntegers(silence_phones_str, ":", false,
                    &silence_phones))
            KALDI_ERR << "Invalid silence-phones string " << silence_phones_str;
        kaldi::SortAndUniq(&silence_phones);
        if (silence_phones.empty())
            KALDI_LOG << "No silence phones specified.";

        // Select the GPU
#if HAVE_CUDA == 1
        CuDevice::Instantiate().SelectGpuId(mpe->use_gpu);
#endif

        // Read the class-frame-counts, compute priors
        mpe->log_prior = new PdfPrior(prior_opts);

        // Read transition model
        mpe->trans_model = new TransitionModel;
        ReadKaldiObject(transition_model_filename, mpe->trans_model);

        mpe->den_lat_reader = new RandomAccessLatticeReader(den_lat_rspecifier);
        mpe->ref_ali_reader = new RandomAccessInt32VectorReader(ref_ali_rspecifier);

        mpe->time = new Timer;
        mpe->time_now = 0;
        mpe->num_done =0;
        mpe->num_no_ref_ali = 0;
        mpe->num_no_den_lat = 0;
        mpe->num_other_error = 0;
        mpe->total_frames = 0;
        mpe->total_frame_acc = 0.0;
        mpe->utt_frame_acc = 0.0;

        return mpe;
    }

    void destroy_KaldiMPE(KaldiMPE *mpe)
    {
        delete mpe->trans_model;
        delete mpe->den_lat_reader;
        delete mpe->ref_ali_reader;
        delete mpe->time;
        delete mpe->prior_opts;
        delete mpe->log_prior;
    }

    int check_mpe(KaldiMPE *mpe, const NervMatrix* mat, const char *key)
    {
        std::string utt(key);
        if (!mpe->den_lat_reader->HasKey(utt)) {
            KALDI_WARN << "Utterance " << utt << ": found no lattice.";
            mpe->num_no_den_lat++;
            return 0;
        }
        if (!mpe->ref_ali_reader->HasKey(utt)) {
            KALDI_WARN << "Utterance " << utt << ": found no reference alignment.";
            mpe->num_no_ref_ali++;
            return 0;
        }

        //assert(sizeof(BaseFloat) == sizeof(float));
        // 1) get the features, numerator alignment
        mpe->ref_ali = mpe->ref_ali_reader->Value(utt);
        long mat_nrow = mat->nrow, mat_ncol = mat->ncol;
        // check for temporal length of numerator alignments
        if (static_cast<MatrixIndexT>(mpe->ref_ali.size()) != mat_nrow) {
            KALDI_WARN << "Numerator alignment has wrong length "
                << mpe->ref_ali.size() << " vs. "<< mat_nrow;
            mpe->num_other_error++;
            return 0;
        }
        if (mat_nrow > mpe->max_frames) {
            KALDI_WARN << "Utterance " << utt << ": Skipped because it has " << mat_nrow <<
                " frames, which is more than " << mpe->max_frames << ".";
            mpe->num_other_error++;
            return 0;
        }
        // 2) get the denominator lattice, preprocess
        mpe->den_lat = mpe->den_lat_reader->Value(utt);
        Lattice &den_lat = mpe->den_lat;
        if (den_lat.Start() == -1) {
            KALDI_WARN << "Empty lattice for utt " << utt;
            mpe->num_other_error++;
            return 0;
        }
        if (mpe->old_acoustic_scale != 1.0) {
            fst::ScaleLattice(fst::AcousticLatticeScale(mpe->old_acoustic_scale),
                    &den_lat);
        }
        // optional sort it topologically
        kaldi::uint64 props = den_lat.Properties(fst::kFstProperties, false);
        if (!(props & fst::kTopSorted)) {
            if (fst::TopSort(&den_lat) == false)
                KALDI_ERR << "Cycles detected in lattice.";
        }
        // get the lattice length and times of states
        mpe->state_times.clear();
        vector<int32> &state_times = mpe->state_times;
        int32 max_time = kaldi::LatticeStateTimes(den_lat, &state_times);
        // check for temporal length of denominator lattices
        if (max_time != mat_nrow) {
            KALDI_WARN << "Denominator lattice has wrong length "
                << max_time << " vs. " << mat_nrow;
            mpe->num_other_error++;
            return 0;
        }

        return 1;
    }

    NervMatrix * calc_diff_mpe(KaldiMPE * mpe, NervMatrix * mat, const char * key)
    {
        std::string utt(key);
        //assert(sizeof(BaseFloat) == sizeof(float));

        CuMatrix<BaseFloat> nnet_diff;
        kaldi::Matrix<BaseFloat> nnet_out_h;
        nnet_out_h.Resize(mat->nrow, mat->ncol, kUndefined);

        size_t stride = mat->stride;
        for (int i = 0; i < mat->nrow; i++)
        {
            const BaseFloat *nerv_row = (BaseFloat *)((char *)mat->data.f + i * stride);
            BaseFloat *row = nnet_out_h.RowData(i);
            memmove(row, nerv_row, sizeof(BaseFloat) * mat->ncol);
        }

        mpe->num_frames = nnet_out_h.NumRows();

        PdfPriorOptions &prior_opts = *(mpe->prior_opts);
        if (prior_opts.class_frame_counts != "") {
            CuMatrix<BaseFloat> nnet_out;
            nnet_out.Resize(nnet_out_h.NumRows(), nnet_out_h.NumCols(), kUndefined);
            nnet_out.CopyFromMat(nnet_out_h);
            mpe->log_prior->SubtractOnLogpost(&nnet_out);
            nnet_out_h.Resize(nnet_out.NumRows(), nnet_out.NumCols(), kUndefined);
            nnet_out.CopyToMat(&nnet_out_h);
            nnet_out.Resize(0,0);
        }

        // 4) rescore the latice
        LatticeAcousticRescore(nnet_out_h, *(mpe->trans_model), mpe->state_times, &(mpe->den_lat));
        if (mpe->acoustic_scale != 1.0 || mpe->lm_scale != 1.0)
            fst::ScaleLattice(fst::LatticeScale(mpe->lm_scale, mpe->acoustic_scale), &(mpe->den_lat));

        kaldi::Posterior post;
        std::vector<int32> &silence_phones = mpe->silence_phones;

        if (mpe->do_smbr) {  // use state-level accuracies, i.e. sMBR estimation
            mpe->utt_frame_acc = LatticeForwardBackwardMpeVariants(
                    *(mpe->trans_model), silence_phones, mpe->den_lat, mpe->ref_ali, "smbr",
                    mpe->one_silence_class, &post);
        } else {  // use phone-level accuracies, i.e. MPFE (minimum phone frame error)
            mpe->utt_frame_acc = LatticeForwardBackwardMpeVariants(
                    *(mpe->trans_model), silence_phones, mpe->den_lat, mpe->ref_ali, "mpfe",
                    mpe->one_silence_class, &post);
        }

        // 6) convert the Posterior to a matrix,
        PosteriorToMatrixMapped(post, *(mpe->trans_model), &nnet_diff);
        nnet_diff.Scale(-1.0); // need to flip the sign of derivative,

        KALDI_VLOG(1) << "Lattice #" << mpe->num_done + 1 << " processed"
            << " (" << utt << "): found " << mpe->den_lat.NumStates()
            << " states and " << fst::NumArcs(mpe->den_lat) << " arcs.";

        KALDI_VLOG(1) << "Utterance " << utt << ": Average frame accuracy = "
            << (mpe->utt_frame_acc/mpe->num_frames) << " over " << mpe->num_frames
            << " frames,"
            << " diff-range(" << nnet_diff.Min() << "," << nnet_diff.Max() << ")";

        nnet_out_h.Resize(nnet_diff.NumRows(), nnet_diff.NumCols(), kUndefined);
        nnet_diff.CopyToMat(&nnet_out_h);
        nnet_diff.Resize(0,0); // release GPU memory,

        assert(mat->nrow == nnet_out_h.NumRows() && mat->ncol == nnet_out_h.NumCols());
        stride = mat->stride;
        for (int i = 0; i < mat->nrow; i++)
        {
            const BaseFloat *row = nnet_out_h.RowData(i);
            BaseFloat *nerv_row = (BaseFloat *)((char *)mat->data.f + i * stride);
            memmove(nerv_row, row, sizeof(BaseFloat) * mat->ncol);
        }
        nnet_out_h.Resize(0,0);

        // increase time counter
        mpe->total_frame_acc += mpe->utt_frame_acc;
        mpe->total_frames += mpe->num_frames;
        mpe->num_done++;

        if (mpe->num_done % 100 == 0) {
            mpe->time_now = mpe->time->Elapsed();
            KALDI_VLOG(1) << "After " << mpe->num_done << " utterances: time elapsed = "
                << mpe->time_now/60 << " min; processed " << mpe->total_frames/mpe->time_now
                << " frames per second.";
#if HAVE_CUDA==1
            // check the GPU is not overheated
            CuDevice::Instantiate().CheckGpuHealth();
#endif
        }
        return mat;
    }

    double get_num_frames_mpe(const KaldiMPE *mpe)
    {
        return (double)mpe->num_frames;
    }

    double get_utt_frame_acc_mpe(const KaldiMPE *mpe)
    {
        return (double)mpe->utt_frame_acc;
    }

}