lib/IDS/DeReKoVecs/derekovecs-server.c

#include <collocatordb.h>
#include <math.h>
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mman.h>
#include <fcntl.h>
#include <unistd.h>
#include <perl.h>

#define max_size 2000
#define max_w 50
#define MAX_NEIGHBOURS 1000
#define MAX_WORDS -1
#define MAX_THREADS 100
#define MAX_CC 50
#define EXP_TABLE_SIZE 1000
#define MAX_EXP 6
#define MIN_RESP 0.50

//the thread function
void *connection_handler(void *);

typedef struct {
  long long wordi;
  long position;
  float activation;
  float average;
  float cprobability;  // column wise probability
  float cprobability_sum;
  float probability;
  float activation_sum;
  float max_activation;
  float heat[16];
} collocator;

typedef struct {
  collocator *best;
  int length;
} knn;

typedef struct {
  long long wordi[MAX_NEIGHBOURS];
  char sep[MAX_NEIGHBOURS];
  int length;
} wordlist;

typedef struct {
  long cutoff;
  wordlist *wl;
  char *token;
  int N;
  long from;
  unsigned long upto;
  collocator *best;
  float *target_sums;
  float *window_sums;
  float threshold;
} knnpars;

typedef struct {
  uint32_t index;
  float value;
} sparse_t;

typedef struct {
  uint32_t len;
  sparse_t nbr[100];
} profile_t;

float *M, *M2 = 0L, *syn1neg_window, *expTable;
char *vocab;
char *garbage = NULL;
COLLOCATORDB *cdb = NULL;
profile_t *sprofiles = NULL;
size_t sprofiles_qty = 0;

long long words, size, merged_end;
long long merge_words = 0;
int num_threads = 20;
int latin_enc = 0;
int window;

/* load collocation profiles if file exists */
int load_sprofiles(char *vecsname) {
  char *basename = strdup(vecsname);
  char *pos = strstr(basename, ".vecs");
  if (pos)
    *pos = 0;

  char binsprofiles_fname[256];
  strcpy(binsprofiles_fname, basename);
  strcat(binsprofiles_fname, ".sprofiles.bin");
  FILE *fp = fopen(binsprofiles_fname, "rb");
  if (fp == NULL) {
    printf("Collocation profiles %s not found. No problem.\n", binsprofiles_fname);
    return 0;
  }
  fseek(fp, 0L, SEEK_END);
  size_t sz = ftell(fp);
  fclose(fp);

  int fd = open(binsprofiles_fname, O_RDONLY);
  sprofiles = mmap(0, sz, PROT_READ, MAP_SHARED, fd, 0);
  if (sprofiles == MAP_FAILED) {
    close(fd);
    fprintf(stderr, "Cannot mmap %s\n", binsprofiles_fname);
    sprofiles = NULL;
    return 0;
  } else {
    sprofiles_qty = sz / sizeof(profile_t);
    fprintf(stderr, "Successfully mmaped %s containing similar profiles for %ld word forms.\n", binsprofiles_fname, sprofiles_qty);
  }
  return 1;
}

char *removeExtension(char* myStr) {
    char *retStr;
    char *lastExt;
    if (myStr == NULL) return NULL;
    if ((retStr = malloc (strlen (myStr) + 1)) == NULL) return NULL;
    strcpy (retStr, myStr);
    lastExt = strrchr (retStr, '.');
    if (lastExt != NULL)
        *lastExt = '\0';
    return retStr;
}

int init_net(char *file_name, char *net_name, int latin, int do_open_cdb) {
  FILE *f, *binvecs, *binwords;
  int binwords_fd, binvecs_fd, net_fd, i;
  long long a, b;
  float len;
  double val;

  char binvecs_fname[1024], binwords_fname[1024];

  if (strstr(file_name, ".txt")) {
    strcpy(binwords_fname, removeExtension(file_name));
  } else {
    strcpy(binwords_fname, file_name);
  }
  strcat(binwords_fname, ".words");
  strcpy(binvecs_fname, file_name);
  strcat(binvecs_fname, ".vecs");

  latin_enc = latin;
  f = fopen(file_name, "rb");
  if (f == NULL) {
    printf("Input file %s not found\n", file_name);
    return -1;
  }
  fscanf(f, "%lld", &words);
  if (MAX_WORDS > 0 && words > MAX_WORDS) words = MAX_WORDS;
  fscanf(f, "%lld", &size);
  if ((binvecs_fd = open(binvecs_fname, O_RDONLY)) < 0 || (binwords_fd = open(binwords_fname, O_RDONLY)) < 0) {
    printf("Converting %s to memory mappable structures\n", file_name);
    vocab = (char *)malloc((long long)words * max_w * sizeof(char));
    M = (float *)malloc((long long)words * (long long)size * sizeof(float));
    if (M == NULL) {
      printf("Cannot allocate memory: %lld MB    %lld  %lld\n", (long long)words * size * sizeof(float) / 1048576, words, size);
      return -1;
    }
    if (strstr(file_name, ".txt")) {
      printf("%lld words in ascii vector file with vector size %lld\n", words, size);
      for (b = 0; b < words; b++) {
        a = 0;
        while (1) {
          vocab[b * max_w + a] = fgetc(f);
          if (feof(f) || (vocab[b * max_w + a] == ' ')) break;
          if ((a < max_w) && (vocab[b * max_w + a] != '\n')) a++;
        }
        vocab[b * max_w + a] = 0;
        len = 0;
        for (a = 0; a < size; a++) {
          fscanf(f, "%lf", &val);
          M[a + b * size] = val;
          len += val * val;
        }
        len = sqrt(len);
        for (a = 0; a < size; a++) M[a + b * size] /= len;
      }
    } else {
      for (b = 0; b < words; b++) {
        a = 0;
        while (1) {
          vocab[b * max_w + a] = fgetc(f);
          if (feof(f) || (vocab[b * max_w + a] == ' ')) break;
          if ((a < max_w) && (vocab[b * max_w + a] != '\n')) a++;
        }
        vocab[b * max_w + a] = 0;
        fread(&M[b * size], sizeof(float), size, f);
        len = 0;
        for (a = 0; a < size; a++) len += M[a + b * size] * M[a + b * size];
        len = sqrt(len);
        for (a = 0; a < size; a++) M[a + b * size] /= len;
      }
    }
    if ((binvecs = fopen(binvecs_fname, "wb")) != NULL && (binwords = fopen(binwords_fname, "wb")) != NULL) {
      fwrite(M, sizeof(float), (long long)words * (long long)size, binvecs);
      fclose(binvecs);
      fwrite(vocab, sizeof(char), (long long)words * max_w, binwords);
      fclose(binwords);
    }
  }
  if ((binvecs_fd = open(binvecs_fname, O_RDONLY)) >= 0 && (binwords_fd = open(binwords_fname, O_RDONLY)) >= 0) {
    M = mmap(0, sizeof(float) * (long long)words * (long long)size, PROT_READ, MAP_SHARED, binvecs_fd, 0);
    vocab = mmap(0, sizeof(char) * (long long)words * max_w, PROT_READ, MAP_SHARED, binwords_fd, 0);
    if (M == MAP_FAILED || vocab == MAP_FAILED) {
      close(binvecs_fd);
      close(binwords_fd);
      fprintf(stderr, "Cannot mmap %s or %s\n", binwords_fname, binvecs_fname);
      exit(-1);
    }
  } else {
    fprintf(stderr, "Cannot open %s or %s\n", binwords_fname, binvecs_fname);
    exit(-1);
  }
  fclose(f);

  if (net_name && strlen(net_name) > 0) {
    if ((net_fd = open(net_name, O_RDONLY)) >= 0) {
      window = (lseek(net_fd, 0, SEEK_END) - sizeof(float) * words * size) / words / size / sizeof(float) / 2;
      //      lseek(net_fd, sizeof(float) * words * size, SEEK_SET);
      // munmap(M,  sizeof(float) * words * size);
      M2 = mmap(0, sizeof(float) * words * size + sizeof(float) * 2 * window * size * words, PROT_READ, MAP_SHARED, net_fd, 0);
      if (M2 == MAP_FAILED) {
        close(net_fd);
        fprintf(stderr, "Cannot mmap %s\n", net_name);
        exit(-1);
      }
      syn1neg_window = M2 + words * size;
    } else {
      fprintf(stderr, "Cannot open %s\n", net_name);
      exit(-1);
    }
    fprintf(stderr, "Successfully memmaped %s. Determined window size: %d\n", net_name, window);

    if (do_open_cdb) {
      char collocatordb_name[2048];
      strcpy(collocatordb_name, net_name);
      char *ext = rindex(collocatordb_name, '.');
      if (ext) {
        strcpy(ext, ".rocksdb");
        if (access(collocatordb_name, R_OK) == 0) {
          *ext = 0;
          fprintf(stderr, "Opening collocator DB	%s\n", collocatordb_name);
          cdb = open_collocatordb(collocatordb_name);
        } else {
           fprintf(stderr, "Cannot open collocator DB	%s\n", collocatordb_name);
        }
      }
    }
  }

  expTable = (float *)malloc((EXP_TABLE_SIZE + 1) * sizeof(float));
  for (i = 0; i < EXP_TABLE_SIZE; i++) {
    expTable[i] = exp((i / (float)EXP_TABLE_SIZE * 2 - 1) * MAX_EXP);  // Precompute the exp() table
    expTable[i] = expTable[i] / (expTable[i] + 1);                     // Precompute f(x) = x / (x + 1)
  }

  return 0;
}

/* The collocator threads of one request accumulate their activation sums per
   window position here. This must not be shared between requests, so every
   request allocates its own array. Only valid once init_net() has determined
   the window size, i.e. if M2 is set. */
float *new_window_sums() {
  return calloc((window + 1) * 2, sizeof(float));
}

long mergeVectors(char *file_name) {
  FILE *f;
  int binwords_fd, binvecs_fd;
  float *merge_vecs;
  char *merge_vocab;
  /*  long long merge_words, merge_size; */
  long long merge_size;

  char binvecs_fname[256], binwords_fname[256];


  strcpy(binwords_fname, file_name);
  strcat(binwords_fname, ".words");
  strcpy(binvecs_fname, file_name);
  strcat(binvecs_fname, ".vecs");

  f = fopen(file_name, "rb");
  if (f == NULL) {
    printf("Input file %s not found\n", file_name);
    exit(-1);
  }
  fscanf(f, "%lld", &merge_words);
  fscanf(f, "%lld", &merge_size);
  if (merge_size != size) {
    fprintf(stderr, "vectors must have the same length\n");
    exit(-1);
  }
  if ((binvecs_fd = open(binvecs_fname, O_RDONLY)) >= 0 && (binwords_fd = open(binwords_fname, O_RDONLY)) >= 0) {
    merge_vecs = malloc(sizeof(float) * (words + merge_words) * size);
    merge_vocab = malloc(sizeof(char) * (words + merge_words) * max_w);
    if (merge_vecs == NULL || merge_vocab == NULL) {
      close(binvecs_fd);
      close(binwords_fd);
      fprintf(stderr, "Cannot reserve memory for %s or %s\n", binwords_fname, binvecs_fname);
      exit(-1);
    }
    read(binvecs_fd, merge_vecs, merge_words * size * sizeof(float));
    read(binwords_fd, merge_vocab, merge_words * max_w);
  } else {
    fprintf(stderr, "Cannot open %s or %s\n", binwords_fname, binvecs_fname);
    exit(-1);
  }
  printf("Successfully reallocated memory\nMerging...\n");
  fflush(stdout);
  memcpy(merge_vecs + merge_words * size, M, words * size * sizeof(float));
  memcpy(merge_vocab + merge_words * max_w, vocab, words * max_w);
  munmap(M, words * size * sizeof(float));
  munmap(vocab, words * max_w);
  M = merge_vecs;
  vocab = merge_vocab;
  merged_end = merge_words;
  words += merge_words;
  fclose(f);
  printf("merged_end: %lld, words: %lld\n", merged_end, words);
  //printBiggestMergedDifferences();
  return ((long)merged_end);
}

void filter_garbage() {
  long i;
  unsigned char *w, previous, c;
  garbage = malloc(words);
  memset(garbage, 0, words);
  for (i = 0; i < words; i++) {
    w = (unsigned char *) vocab + i * max_w;
    previous = 0;
    if (strncmp("quot", (const char *)w, 4) == 0) {
      garbage[i] = 1;
      //      printf("Gargabe: %s\n", vocab + i * max_w);
    } else {
      while ((c = *w++) && !garbage[i]) {
        if (((c <= 90 && c >= 65) && (previous >= 97 && previous <= 122)) ||
            (previous == '-' && (c & 32)) ||
            (previous == 0xc2 && (c == 0xa4 || c == 0xb6)) ||
            (previous == 'q' && c == 'u' && *(w) == 'o' && *(w + 1) == 't') || /* quot */
            c == '<') {
          garbage[i] = 1;
          continue;
        }
        previous = c;
      }
    }
  }
  return;
}

knn *simpleGetCollocators(int word, int number, long cutoff, int *result) {
  knnpars *pars = calloc(sizeof(knnpars), 1);
  float *target_sums = NULL;
  float *my_window_sums = malloc(sizeof(float) * (window + 1) * 2);
  pars->cutoff = (cutoff ? cutoff : 300000);
  long a;
  for (a = 0; a < cutoff; a++)
    target_sums[a] = 0;
  pars->target_sums = target_sums;
  pars->window_sums = my_window_sums;
  pars->N = (number ? number : 20);
  pars->from = 0;
  pars->upto = window * 2 - 1;
  knn *syn_nbs = NULL;  // = (knn*) getCollocators(pars);
  free(pars);
  free(my_window_sums);
  free(target_sums);
  return syn_nbs;
}

/* Frees a knn result as returned by getCollocators() via pthread_exit(). */
void free_knn(knn *nbs) {
  if (nbs == NULL)
    return;
  free(nbs->best);
  free(nbs);
}

void *getCollocators(void *args) {
  knnpars *pars = args;
  int N = pars->N;

  int cc = pars->wl->wordi[0];
  knn *nbs = NULL;
  long window_layer_size = size * window * 2;
  long a, b, c, d, window_offset, target, max_target = 0, maxmax_target;
  float f, max_f, maxmax_f;
  float *target_sums = NULL, worstbest, wpos_sum;
  collocator *best;

  if (M2 == NULL || cc == -1)
    return NULL;

  a = posix_memalign((void **)&target_sums, 128, pars->cutoff * sizeof(float));
  memset(target_sums, 0, pars->cutoff * sizeof(float));
  best = malloc((N > 200 ? N : 200) * sizeof(collocator));
  memset(best, 0, (N > 200 ? N : 200) * sizeof(collocator));
  worstbest = pars->threshold;

  for (b = 0; b < pars->cutoff; b++)
    target_sums[b] = 0;
  for (b = 0; b < N; b++) {
    best[b].wordi = -1;
    best[b].probability = 1;
    best[b].activation = worstbest;
  }

  d = cc;
  maxmax_f = -1;
  maxmax_target = 0;

  for (a = pars->from; a < pars->upto; a++) {
    if (a >= window)
      a++;
    wpos_sum = 0;
    printf("window pos: %ld\n", a);
    if (a != window) {
      max_f = -1;
      window_offset = a * size;
      if (a > window)
        window_offset -= size;
      for (target = 0; target < pars->cutoff; target++) {
        if (garbage && garbage[target]) continue;
        if (target == d)
          continue;
        f = 0;
        for (c = 0; c < size; c++)
          f += M2[d * size + c] * syn1neg_window[target * window_layer_size + window_offset + c];
        if (f < -MAX_EXP)
          continue;
        else if (f > MAX_EXP)
          continue;
        else
          f = expTable[(int)((f + MAX_EXP) * (EXP_TABLE_SIZE / MAX_EXP / 2))];
        wpos_sum += f;

        target_sums[target] += f;
        if (f > worstbest) {
          for (b = 0; b < N; b++) {
            if (f > best[b].activation) {
              memmove(best + b + 1, best + b, (N - b - 1) * sizeof(collocator));
              best[b].activation = f;
              best[b].wordi = target;
              best[b].position = window - a;
              break;
            }
          }
          if (b == N - 1)
            worstbest = best[N - 1].activation;
        }
      }
      printf("%ld %.2f\n", max_target, max_f);
      printf("%s (%.2f) ", &vocab[max_target * max_w], max_f);
      if (max_f > maxmax_f) {
        maxmax_f = max_f;
        maxmax_target = max_target;
      }
      for (b = 0; b < N; b++)
        if (best[b].position == window - a)
          best[b].cprobability = best[b].activation / wpos_sum;
    } else {
      printf("\x1b[1m%s\x1b[0m ", &vocab[d * max_w]);
    }
    pars->window_sums[a] = wpos_sum;
  }
  for (b = 0; b < pars->cutoff; b++)
    pars->target_sums[b] += target_sums[b];  //(target_sums[b] / wpos_sum ) / (window * 2);

  free(target_sums);
  for (b = 0; b < N && best[b].wordi >= 0; b++)
    ;
     // THIS LOOP IS NEEDED (b...)
     //		printf("%d: best syn: %s %.2f %.5f\n", b, &vocab[best[b].wordi*max_w], best[b].activation, best[b].probability);
     //	printf("\n");
  nbs = malloc(sizeof(knn));
  nbs->best = best;
  nbs->length = b - 1;
  pthread_exit(nbs);
}

float getOutputWeight(int hidden, long target, int window_position) {
  const long window_layer_size = size * window * 2;
  int a;

  if (window_position == 0 || window_position > window || window_position < -window) {
    fprintf(stderr, "window_position: %d - assert: -%d <= window_position <= %d && window_position != 0 failed.\n", window_position, window, window);
    exit(-1);
  }

  if (hidden >= size) {
    fprintf(stderr, "hidden: %d - assert: hidden < %lld failed.\n", hidden, size);
    exit(-1);
  }

  if (target >= words) {
    fprintf(stderr, "target: %ld - assert: target < %lld failed.\n", target, words);
    exit(-1);
  }

  a = window_position + window;
  if (a > window) {
    --a;
  }
  long window_offset = a * size;
  return syn1neg_window[target * window_layer_size + window_offset + hidden];
}

/* Returns an SV* (not an AV*) on purpose: for an AV* return value Inline::C
   generates newRV(), which leaves the array itself with a reference count of
   one after the mortal reference is gone, i.e. it leaks the whole array on
   every call. */
SV *getVecs(AV *array) {
  int i, b;
  AV *result = newAV();
  for (i = 0; i <= av_len(array); i++) {
    SV **elem = av_fetch(array, i, 0);
    if (elem != NULL) {
      long j = (long)SvNV(*elem);
      AV *vector = newAV();
      for (b = 0; b < size; b++) {
        av_push(vector, newSVnv(M[b + j * size]));
      }
      av_push(result, newRV_noinc((SV *)vector));
    }
  }
  return newRV_noinc((SV *)result);
}

/* All functions handing a string back to perl return an SV*, because for a
   char* return value Inline::C only copies the string into the return SV and
   never frees the buffer we allocated here. */
SV *getSimilarProfiles(long node) {
  int i;
  char buffer[120000];
  char pair_buffer[2048];
  buffer[0] = '[';
  buffer[1] = 0;
  if (node >= sprofiles_qty) {
    printf("Not available in precomputed profile\n");
    return newSVpv("[{\"w\":\"not available\", \"v\":0}]\n", 0);
  }

  printf("******* %s ******\n", &vocab[max_w * node]);

  for (i = 0; i < 100 && i < sprofiles[node].len; i++) {
    sprintf(pair_buffer, "{\"w\":\"%s\", \"v\":%f},", &vocab[max_w * (sprofiles[node].nbr[i].index)], sprofiles[node].nbr[i].value);
    strcat(buffer, pair_buffer);
  }
  if (i > 0)
    buffer[strlen(buffer) - 1] = ']';
  else
    strcat(buffer, "]");
  strcat(buffer, "\n");
  printf("%s", buffer);
  return newSVpv(buffer, 0);
}

/* get_collocat*_as_json() hand out strdup()ed buffers that we own. */
SV *getCollocationScores(long node, long collocate) {
  char *json = (cdb ? (char *)get_collocation_scores_as_json(cdb, node, collocate) : NULL);
  SV *res = newSVpv(json ? json : "[]", 0);
  free(json);
  return res;
}

SV *getClassicCollocators(long node) {
  char *json = (cdb ? (char *)get_collocators_as_json(cdb, node) : NULL);
  SV *res = newSVpv(json ? json : "[]", 0);
  free(json);
  return res;
}

wordlist *getTargetWords(char *st1, int search_backw) {
  wordlist *wl = malloc(sizeof(wordlist));
  char st[100][max_size];
  long a, b = 0, c = 0, cn = 0;

  while (1) {
    st[cn][b] = st1[c];
    b++;
    c++;
    st[cn][b] = 0;
    if (st1[c] == 0) break;
    if (st1[c] == ' ' /*|| st1[c] == '-'*/) {
      b = 0;
      c++;
    }
  }
  cn++;
  for (a = 0; a < cn; a++) {
    if (search_backw) {
      for (b = words - 1; b >= (merge_words ? merge_words : 0) && strcmp(&vocab[b * max_w], st[a]) != 0; b--)
        ;
    } else {
      for (b = 0; b < (merge_words ? merge_words : words) && strcmp(&vocab[b * max_w], st[a]) != 0; b++)
        ;
    }
    if (b == words) b = -1;
    wl->wordi[a] = b;
    if (b == -1) {
      fprintf(stderr, "Out of dictionary word!\n");
      cn--;
    } else {
      fprintf(stderr, "Word: \"%s\"  Position in vocabulary: %lld\n", &vocab[wl->wordi[a] * max_w], wl->wordi[a]);
    }
  }
  wl->length = cn;
  return (wl);
}

long getWordNumber(char *word) {
  wordlist *wl = getTargetWords(word, 0);
  long res = 0;
  if (wl == NULL)
    return(0);
  if(wl->length > 0)
    res = wl->wordi[0];
  free(wl);
  return(res);
}

float get_distance(long b, long c) {
  long a;
  float dist = 0;
  for (a = 0; a < size; a++) dist += M[a + c * size] * M[a + b * size];
  return dist;
}

/* The result is computed once and then kept in a static buffer for the
   lifetime of the process. */
SV *getBiggestMergedDifferences() {
  static char *result = NULL;
  float dist;
  long long a, c;
  int N = 1000;

  if (merged_end == 0)
    result = "[]";

  if (result != NULL)
    return newSVpv(result, 0);

  printf("Looking for biggest distances between main and merged vectors ...\n");
  collocator *best;
  best = malloc(N * sizeof(collocator));
  memset(best, 0, N * sizeof(collocator));

  float worstbest = 1000000;

  for (a = 0; a < N; a++) best[a].activation = worstbest;

  for (c = 0; c < 500000; c++) {
    if (garbage && garbage[c]) continue;
    dist = 0;
    for (a = 0; a < size; a++) dist += M[a + c * size] * M[a + (c + merged_end) * size];
    if (dist < worstbest) {
      for (a = 0; a < N; a++) {
        if (dist < best[a].activation) {
          memmove(best + a + 1, best + a, (N - a - 1) * sizeof(collocator));
          best[a].activation = dist;
          best[a].wordi = c;
          break;
        }
      }
      worstbest = best[N - 1].activation;
    }
  }

  result = malloc(N * (max_w + 64));
  char *p = (char *) result;
  *p++ = '[';
  *p = 0;
  for (a = 0; a < N; a++) {
    p += sprintf(p, "{\"rank\":%lld,\"word\":\"%s\",\"dist\":%.3f},", a, &vocab[best[a].wordi * max_w], 1 - best[a].activation);
  }
  *--p = ']';
  free(best);
  return newSVpv(result, 0);
}

float cos_similarity(long b, long c) {
  float dist = 0;
  long a;
  for (a = 0; a < size; a++) dist += M[b * size + a] * M[c * size + a];
  return dist;
}

SV *cos_similarity_as_json(char *w1, char *w2) {
  wordlist *a, *b;
  float res;
  char json[32];
  a = getTargetWords(w1, 0);
  b = getTargetWords(w2, 0);
  if (a == NULL || b == NULL || a->length != 1 || b->length != 1)
    res = -1;
  else {
    res = cos_similarity(a->wordi[0], b->wordi[0]);
    fprintf(stderr, "a: %lld b: %lld res:%f\n", a->wordi[0], b->wordi[0], res);
  }
  free(a);
  free(b);
  sprintf(json, "%.5f", res);
  return newSVpv(json, 0);
}

void *_get_neighbours(void *arg) {
  knnpars *pars = arg;
  int N = pars->N;
  long from = pars->from;
  unsigned long upto = pars->upto;
  char *sep;
  float dist, len, vec[max_size];
  long long a, b, c, cn, *bi;
  knn *nbs = NULL;
  wordlist *wl = pars->wl;

  collocator *best = pars->best;

  float worstbest = -1;

  for (a = 0; a < N; a++) best[a].activation = 0;
  a = 0;
  bi = wl->wordi;
  cn = wl->length;
  sep = wl->sep;
  b = bi[0];
  if (b == -1) {
    goto end;
  }
  for (a = 0; a < size; a++) vec[a] = 0;
  for (b = 0; b < cn; b++) {
    if (bi[b] == -1) continue;
    if (b > 0 && sep[b - 1] == '-')
      for (a = 0; a < size; a++) vec[a] -= M[a + bi[b] * size];
    else
      for (a = 0; a < size; a++) vec[a] += M[a + bi[b] * size];
  }
  len = 0;
  for (a = 0; a < size; a++) len += vec[a] * vec[a];
  len = sqrt(len);
  for (a = 0; a < size; a++) vec[a] /= len;
  for (a = 0; a < N; a++) best[a].activation = -1;
  for (c = from; c < upto; c++) {
    if (garbage && garbage[c]) continue;
    a = 0;
    // do not skip taget word
    //		for (b = 0; b < cn; b++) if (bi[b] == c) a = 1;
    //		if (a == 1) continue;
    dist = 0;
    for (a = 0; a < size; a++) dist += vec[a] * M[a + c * size];
    if (dist > worstbest) {
      for (a = 0; a < N; a++) {
        if (dist > best[a].activation) {
          memmove(best + a + 1, best + a, (N - a - 1) * sizeof(collocator));
          best[a].activation = dist;
          best[a].wordi = c;
          break;
        }
      }
      worstbest = best[N - 1].activation;
    }
  }

end:
  pthread_exit(nbs);
}

int cmp_activation(const void *a, const void *b) {
  float fb = ((collocator *)a)->activation;
  float fa = ((collocator *)b)->activation;
  return (fa > fb) - (fa < fb);
}

int cmp_probability(const void *a, const void *b) {
  float fb = ((collocator *)a)->probability;
  float fa = ((collocator *)b)->probability;
  return (fa > fb) - (fa < fb);
}

SV *getPosWiseW2VCollocators(char *word, long maxPerPos, long cutoff, float threshold, const char *format) {
  float *target_sums = NULL;
  float *window_sums = NULL;
  long a, b, entries = 0;
  knn *syn_nbs[MAX_THREADS];
  knnpars pars[MAX_THREADS];
  pthread_t *pt = NULL;
  wordlist *wl = NULL;
  int syn_threads = (M2 ? window * 2 : 0);
  int search_backw = 0;
  char *result = NULL;
  SV *res_sv;

  for (a = 0; a < MAX_THREADS; a++) syn_nbs[a] = NULL;

  if (cutoff < 1 || cutoff > words)
    cutoff = words;

  wl = getTargetWords(word, search_backw);
  if (wl == NULL || wl->length < 1 || wl->wordi[0] < 0 || syn_threads < 1) {
    free(wl);
    return newSVpv("", 0);
  }

  pt = (pthread_t *)malloc((num_threads + 1) * sizeof(pthread_t));
  window_sums = new_window_sums();
  a = posix_memalign((void **)&target_sums, 128, cutoff * sizeof(float));
  memset(target_sums, 0, cutoff * sizeof(float));

  printf("Starting %d threads\n", syn_threads);
  fflush(stdout);
  for (a = 0; a < syn_threads; a++) {
    pars[a].cutoff = cutoff;
    pars[a].target_sums = target_sums;
    pars[a].window_sums = window_sums;
    pars[a].wl = wl;
    pars[a].N = maxPerPos;
    pars[a].best = NULL;  /* getCollocators() allocates its own result array */
    pars[a].threshold = threshold;
    pars[a].from = a;
    pars[a].upto = a + 1;
    pthread_create(&pt[a], NULL, getCollocators, (void *)&pars[a]);
  }
  printf("Waiting for syn threads to join\n");
  fflush(stdout);
  for (a = 0; a < syn_threads; a++) pthread_join(pt[a], (void *)&syn_nbs[a]);
  printf("Syn threads joint\n");
  fflush(stdout);
  result = malloc((maxPerPos > 0 ? maxPerPos : 1) * (max_w + 96) * syn_threads + 16);
  char *p = (char *) result;
  *p = 0;
  if (strcmp(format, "tsv") == 0) {
    for (a = syn_threads - 1; a >= 0; a--) {
      if (syn_nbs[a] == NULL) continue;
      for (b = 0; b < syn_nbs[a]->length; b++, entries++) {
        p += sprintf(p, "%ld\t%s\t%f\n", syn_nbs[a]->best[b].position, &vocab[syn_nbs[a]->best[b].wordi * max_w], syn_nbs[a]->best[b].activation);
      }
    }
  } else {
    p += sprintf(p, "[");
      for (a = syn_threads - 1; a >= 0; a--) {
        if (syn_nbs[a] == NULL) continue;
        for (b = 0; b < syn_nbs[a]->length; b++, entries++) {
          p += sprintf(p, "{\"pos\": %ld, \"word\":\"%s\",\"activation\": %f},\n", syn_nbs[a]->best[b].position, &vocab[syn_nbs[a]->best[b].wordi * max_w], syn_nbs[a]->best[b].activation);
        }
      }
      if (entries > 0)
        p -= 2;  /* drop the trailing ",\n" */
      p += sprintf(p, "\n]");
  }

  res_sv = newSVpv(result, 0);

  free(result);
  free(target_sums);
  free(window_sums);
  free(pt);
  free(wl);
  for (a = 0; a < syn_threads; a++) free_knn(syn_nbs[a]);

  return res_sv;
}

SV *getPosWiseW2VCollocatorsAsTsv(char *word, long maxPerPos, long cutoff, float threshold) {
  return getPosWiseW2VCollocators(word, maxPerPos, cutoff, threshold, "tsv");
}

SV *getPosWiseW2VCollocatorsAsJson(char *word, long maxPerPos, long cutoff, float threshold) {
  return getPosWiseW2VCollocators(word, maxPerPos, cutoff, threshold, "json");
}

SV *get_neighbours(char *st1, int N, int sort_by, int search_backw, long cutoff, int dedupe, int no_similar_profiles) {
  HV *result = newHV();
  float *target_sums = NULL;
  float *window_sums = NULL;
  long a, b, c, d, slice;
  knn *para_nbs[MAX_THREADS];
  knn *syn_nbs[MAX_THREADS];
  knnpars pars[MAX_THREADS];
  pthread_t *pt = (pthread_t *)malloc((num_threads + 1) * sizeof(pthread_t));
  wordlist *wl = NULL;
  int syn_threads = (M2 ? window * 2 : 0);
  int para_threads = (no_similar_profiles ? 0 : num_threads - syn_threads);

  for (a = 0; a < MAX_THREADS; a++) para_nbs[a] = syn_nbs[a] = NULL;

  collocator *best = NULL;
  posix_memalign((void **)&best, 128, 10 * (N >= 200 ? N : 200) * sizeof(collocator));
  memset(best, 0, (N >= 200 ? N : 200) * sizeof(collocator));

  if (N > MAX_NEIGHBOURS) N = MAX_NEIGHBOURS;

  if (cutoff < 1 || cutoff > words)
    cutoff = words;

  wl = getTargetWords(st1, search_backw);
  if (wl == NULL || wl->length < 1)
    goto end;

  slice = (para_threads > 0 ? cutoff / para_threads : cutoff);

  a = posix_memalign((void **)&target_sums, 128, cutoff * sizeof(float));
  memset(target_sums, 0, cutoff * sizeof(float));

  printf("Starting %d threads for paradigmatic search\n", para_threads);
  fflush(stdout);
  for (a = 0; a < para_threads; a++) {
    pars[a].cutoff = cutoff;
    pars[a].token = st1;
    pars[a].wl = wl;
    pars[a].N = N;
    pars[a].best = &best[N * a];
    if (merge_words == 0 || search_backw == 0) {
      pars[a].from = a * slice;
      pars[a].upto = ((a + 1) * slice > cutoff ? cutoff : (a + 1) * slice);
    } else {
      pars[a].from = merge_words + a * slice;
      pars[a].upto = merge_words + ((a + 1) * slice > cutoff ? cutoff : (a + 1) * slice);
    }
    printf("From: %ld, Upto: %ld\n", pars[a].from, pars[a].upto);
    pthread_create(&pt[a], NULL, _get_neighbours, (void *)&pars[a]);
  }
  if (M2) {
    window_sums = new_window_sums();
    for (a = 0; a < syn_threads; a++) {
      pars[a + para_threads].cutoff = cutoff;
      pars[a + para_threads].target_sums = target_sums;
      pars[a + para_threads].window_sums = window_sums;
      pars[a + para_threads].wl = wl;
      pars[a + para_threads].N = N;
      pars[a + para_threads].threshold = MIN_RESP;
      pars[a + para_threads].from = a;
      pars[a + para_threads].upto = a + 1;
      pthread_create(&pt[a + para_threads], NULL, getCollocators, (void *)&pars[a + para_threads]);
    }
  }
  printf("Waiting for para threads to join\n");
  fflush(stdout);
  for (a = 0; a < para_threads; a++) pthread_join(pt[a], (void *)&para_nbs[a]);
  printf("Para threads joint\n");
  fflush(stdout);

  /* if(!syn_nbs[0]) */
  /* 	goto end; */

  qsort(best, N * para_threads, sizeof(collocator), cmp_activation);

  long long chosen[MAX_NEIGHBOURS];
  printf("N: %d\n", N);

  AV *array = newAV();
  int i, j;
  int l1_words = 0, l2_words = 0;

  for (a = 0, i = 0; i < N && a < N * para_threads; a++) {
    int filtered = 0;
    long long c = best[a].wordi;
    if ((merge_words && dedupe && i > 1) || (!merge_words && dedupe && i > 0)) {
      for (j = 0; j < i && !filtered; j++)
        if (strcasestr(&vocab[c * max_w], &vocab[chosen[j] * max_w]) ||
            strcasestr(&vocab[chosen[j] * max_w], &vocab[c * max_w])) {
          printf("filtering %s %s\n", &vocab[chosen[j] * max_w], &vocab[c * max_w]);
          filtered = 1;
        }
      if (filtered)
        continue;
    }

    if (0 && merge_words > 0) {
      if (c >= merge_words) {
        if (l1_words > N / 2)
          continue;
        else
          l1_words++;
      } else {
        if (l2_words > N / 2)
          continue;
        else
          l2_words++;
      }
    }

    //    printf("%s l1:%d l2:%d i:%d a:%ld\n", &vocab[c * max_w], l1_words, l2_words, i, a);
    //    fflush(stdout);
    HV *hash = newHV();
    SV *word = newSVpvf(&vocab[c * max_w], 0);
    chosen[i] = c;
    if (latin_enc == 0) SvUTF8_on(word);
    fflush(stdout);
    hv_store(hash, "word", strlen("word"), word, 0);
    hv_store(hash, "dist", strlen("dist"), newSVnv(best[a].activation), 0);
    hv_store(hash, "rank", strlen("rank"), newSVuv(best[a].wordi), 0);
    AV *vector = newAV();
    for (b = 0; b < size; b++) {
      av_push(vector, newSVnv(M[b + best[a].wordi * size]));
    }
    hv_store(hash, "vector", strlen("vector"), newRV_noinc((SV *)vector), 0);
    av_push(array, newRV_noinc((SV *)hash));
    i++;
  }
  hv_store(result, "paradigmatic", strlen("paradigmatic"), newRV_noinc((SV *)array), 0);

  for (b = 0; b < MAX_NEIGHBOURS; b++) {
    best[b].wordi = -1L;
    best[b].activation = 0;
    best[b].probability = 0;
    best[b].position = 0;
    best[b].activation_sum = 0;
    memset(best[b].heat, 0, sizeof(float) * 16);
  }

  float total_activation = 0;

  if (M2) {
    printf("Waiting for syn threads to join\n");
    fflush(stdout);
    for (a = 0; a < syn_threads; a++) pthread_join(pt[a + para_threads], (void *)&syn_nbs[a]);
    for (a = 0; a <= syn_threads; a++) {
      if (a == window) continue;
      total_activation += window_sums[a];
      printf("window pos: %ld, sum: %f\n", a, window_sums[a]);
    }
    printf("syn threads joint\n");
    fflush(stdout);

    for (b = 0; b < syn_nbs[0]->length; b++) {
      memcpy(best + b, &syn_nbs[0]->best[b], sizeof(collocator));
      best[b].position = -1;  //  syn_nbs[0]->pos[b];
      best[b].activation_sum = target_sums[syn_nbs[0]->best[b].wordi];
      best[b].max_activation = 0.0;
      best[b].average = 0.0;
      best[b].probability = 0.0;
      best[b].cprobability = syn_nbs[0]->best[b].cprobability;
      memset(best[b].heat, 0, sizeof(float) * 16);
    }

    float best_window_sum[MAX_NEIGHBOURS];
    int found_index = 0, i = 0, w;
    for (a = 0; a < syn_threads; a++) {
      for (b = 0; b < syn_nbs[a]->length; b++) {
        for (i = 0; i < found_index; i++)
          if (best[i].wordi == syn_nbs[a]->best[b].wordi)
            break;
        if (i >= found_index) {
          best[found_index].max_activation = 0.0;
          best[found_index].average = 0.0;
          best[found_index].probability = 0.0;
          memset(best[found_index].heat, 0, sizeof(float) * 16);
          best[found_index].cprobability = syn_nbs[a]->best[b].cprobability;
          best[found_index].activation_sum = target_sums[syn_nbs[a]->best[b].wordi];  // syn_nbs[a]->best[b].activation_sum;
          best[found_index++].wordi = syn_nbs[a]->best[b].wordi;
          //						printf("found: %s\n", &vocab[syn_nbs[a]->index[b] * max_w]);
        }
      }
    }
    sort_by = 0;                         // ALWAYS AUTO-FOCUS
    if (sort_by != 1 && sort_by != 2) {  // sort by auto focus mean
      printf("window: %d  -  syn_threads: %d, %d\n", window, syn_threads, (1 << syn_threads) - 1);
      int wpos;
      int bits_set = 0;
      for (i = 0; i < found_index; i++) {
        best[i].activation = best[i].probability = best[i].average = best[i].cprobability_sum = 0;
        for (w = 1; w < (1 << syn_threads); w++) {  // loop through all possible windows
          float word_window_sum = 0, word_window_average = 0, word_cprobability_sum = 0, word_activation_sum = 0, total_window_sum = 0;
          bits_set = 0;
          for (a = 0; a < syn_threads; a++) {
            if ((1 << a) & w) {
              wpos = (a >= window ? a + 1 : a);
              total_window_sum += window_sums[wpos];
            }
          }
          //					printf("%d window-sum %f\n", w, total_window_sum);
          for (a = 0; a < syn_threads; a++) {
            if ((1 << a) & w) {
              wpos = (a >= window ? a + 1 : a);
              bits_set++;
              for (b = 0; b < syn_nbs[a]->length; b++)
                if (best[i].wordi == syn_nbs[a]->best[b].wordi) {
                  //									float acti = syn_nbs[a]->best[b].activation / total_window_sum;
                  //                  word_window_sum += syn_nbs[a]->dist[b] *  syn_nbs[a]->norm[b]; // / window_sums[wpos];  // syn_nbs[a]->norm[b];
                  //                    word_window_sum += syn_nbs[a]->norm[b]; // / window_sums[wpos];  // syn_nbs[a]->norm[b];
                  //                  word_window_sum = (word_window_sum + syn_nbs[a]->norm[b]) - (word_window_sum * syn_nbs[a]->norm[b]);  // syn_nbs[a]->norm[b];

                  word_window_sum += syn_nbs[a]->best[b].activation;  // / window_sums[wpos];  // syn_nbs[a]->norm[b];
                                                                      //                  word_window_sum += acti - (word_window_sum * acti); syn_nbs[a]->best[b].activation; // / window_sums[wpos];  // syn_nbs[a]->norm[b];

                  word_window_average += syn_nbs[a]->best[b].activation;                                                                 // - word_window_average * syn_nbs[a]->best[b].activation;  // conormalied activation sum
                  word_cprobability_sum += syn_nbs[a]->best[b].cprobability - word_cprobability_sum * syn_nbs[a]->best[b].cprobability;  // conormalied column probability sum
                  word_activation_sum += syn_nbs[a]->best[b].activation;
                  if (syn_nbs[a]->best[b].activation > best[i].max_activation)
                    best[i].max_activation = syn_nbs[a]->best[b].activation;
                  if (syn_nbs[a]->best[b].activation > best[i].heat[wpos])
                    best[i].heat[wpos] = syn_nbs[a]->best[b].activation;
                }
            }
          }
          if (bits_set) {
            word_window_average /= bits_set;
            //						word_activation_sum /= bits_set;
            //						word_window_sum /= bits_set;
          }

          word_window_sum /= total_window_sum;

          if (word_window_sum > best[i].probability) {
            //						best[i].position = w;
            best[i].probability = word_window_sum;
          }

          if (word_cprobability_sum > best[i].cprobability_sum) {
            best[i].position = w;
            best[i].cprobability_sum = word_cprobability_sum;
          }

          best[i].average = word_window_average;
          //						best[i].activation = word_activation_sum;
        }
      }
      qsort(best, found_index, sizeof(collocator), cmp_probability);
      //      for(i=0; i < found_index; i++) {
      //				printf("found: %s - sum: %f - window: %d\n", &vocab[best[i].wordi * max_w], best[i].activation, best[i].position);
      //			}

    } else if (sort_by == 1) {  // responsiveness any window position
      int wpos;
      for (i = 0; i < found_index; i++) {
        float word_window_sum = 0, word_activation_sum = 0, total_window_sum = 0;
        for (a = 0; a < syn_threads; a++) {
          wpos = (a >= window ? a + 1 : a);
          for (b = 0; b < syn_nbs[a]->length; b++)
            if (best[i].wordi == syn_nbs[a]->best[b].wordi) {
              best[i].probability += syn_nbs[a]->best[b].probability;
              if (syn_nbs[a]->best[b].activation > 0.25)
                best[i].position |= 1 << wpos;
              if (syn_nbs[a]->best[b].activation > best[i].activation) {
                best[i].activation = syn_nbs[a]->best[b].activation;
              }
            }
        }
      }
      qsort(best, found_index, sizeof(collocator), cmp_activation);
    } else if (sort_by == 2) {  // single window position
      for (a = 1; a < syn_threads; a++) {
        for (b = 0; b < syn_nbs[a]->length; b++) {
          for (c = 0; c < MAX_NEIGHBOURS; c++) {
            if (syn_nbs[a]->best[b].activation > best[c].activation) {
              for (d = MAX_NEIGHBOURS - 1; d > c; d--) {
                memmove(best + d, best + d - 1, sizeof(collocator));
              }
              memcpy(best + c, &syn_nbs[a]->best[b], sizeof(collocator));
              best[c].position = 1 << (-syn_nbs[a]->best[b].position + window - (syn_nbs[a]->best[b].position < 0 ? 1 : 0));
              break;
            }
          }
        }
      }
    } else {  // sort by mean p
      for (a = 1; a < syn_threads; a++) {
        for (b = 0; b < syn_nbs[a]->length; b++) {
          for (c = 0; c < MAX_NEIGHBOURS; c++) {
            if (target_sums[syn_nbs[a]->best[b].wordi] > best[c].activation_sum) {
              for (d = MAX_NEIGHBOURS - 1; d > c; d--) {
                memmove(best + d, best + d - 1, sizeof(collocator));
              }
              memcpy(best + c, &syn_nbs[a]->best[b], sizeof(collocator));
              best[c].position = (1 << 2 * window) - 1;  // syn_nbs[a]->pos[b];
              best[c].activation_sum = target_sums[syn_nbs[a]->best[b].wordi];
              break;
            }
          }
        }
      }
    }
    array = newAV();
    for (a = 0, i = 0; a < MAX_NEIGHBOURS && best[a].wordi >= 0; a++) {
      long long c = best[a].wordi;
      /*
      if (dedupe) {
	  		int filtered=0;
        for (j=0; j<i; j++)
          if (strcasestr(&vocab[c * max_w], chosen[j]) ||
              strcasestr(chosen[j], &vocab[c * max_w])) {
						printf("filtering %s %s\n", chosen[j], &vocab[c * max_w]);
						filtered = 1;
					}
				if(filtered)
					continue;
			}
*/
      chosen[i++] = c;
      HV *hash = newHV();
      SV *word = newSVpvf(&vocab[best[a].wordi * max_w], 0);
      AV *heat = newAV();
      if (latin_enc == 0) SvUTF8_on(word);
      hv_store(hash, "word", strlen("word"), word, 0);
      hv_store(hash, "rank", strlen("rank"), newSVuv(best[a].wordi), 0);
      hv_store(hash, "average", strlen("average"), newSVnv(best[a].average), 0);
      hv_store(hash, "prob", strlen("prob"), newSVnv(best[a].probability), 0);
      hv_store(hash, "cprob", strlen("cprob"), newSVnv(best[a].cprobability_sum), 0);
      hv_store(hash, "max", strlen("max"), newSVnv(best[a].max_activation), 0);                             // newSVnv(target_sums[best[a].wordi]), 0);
      hv_store(hash, "overall", strlen("overall"), newSVnv(best[a].activation_sum / total_activation), 0);  // newSVnv(target_sums[best[a].wordi]), 0);
      hv_store(hash, "pos", strlen("pos"), newSVnv(best[a].position), 0);
      best[a].heat[5] = 0;
      for (i = 10; i >= 0; i--) av_push(heat, newSVnv(best[a].heat[i]));
      hv_store(hash, "heat", strlen("heat"), newRV_noinc((SV *)heat), 0);
      av_push(array, newRV_noinc((SV *)hash));
    }
    hv_store(result, "syntagmatic", strlen("syntagmatic"), newRV_noinc((SV *)array), 0);
  }
end:
  free(best);
  free(target_sums);
  free(window_sums);
  free(pt);
  free(wl);
  for (a = 0; a < MAX_THREADS; a++) {
    free_knn(para_nbs[a]);
    free_knn(syn_nbs[a]);
  }
  return newRV_noinc((SV *)result);
}

int dump_vecs(char *fname) {
  long i, j;
  FILE *f;
  /*  if(words>100000)
	 	words=100000;
*/
  if ((f = fopen(fname, "w")) == NULL) {
    fprintf(stderr, "cannot open %s for writing\n", fname);
    return (-1);
  }
  fprintf(f, "%lld %lld\n", words, size);
  for (i = 0; i < words; i++) {
    fprintf(f, "%s ", &vocab[i * max_w]);
    for (j = 0; j < size - 1; j++)
      fprintf(f, "%f ", M[i * size + j]);
    fprintf(f, "%f\n", M[i * size + j]);
  }
  fclose(f);
  return (0);
}

int dump_for_numpy(char *fname) {
  long i, j;
  FILE *f;
  int max = words; // 300000;

  if ((f = fopen(fname, "w")) == NULL) {
    fprintf(stderr, "cannot open %s for writing\n", fname);
    return (-1);
  }
  for (i = 0; i < max; i++) {
    for (j = 0; j < size - 1; j++)
      fprintf(f, "%f\t", M[i * size + j]);
    fprintf(f, "%f\n", M[i * size + j]);
    printf("%s\r\n", &vocab[i * max_w]);
  }
  if (merged_end > 0) {
    for (i = 0; i < max; i++) {
      for (j = 0; j < size - 1; j++)
        fprintf(f, "%f\t", M[(merged_end + i) * size + j]);
      fprintf(f, "%f\n", M[(merged_end + i) * size + j]);
      printf("_%s\r\n", &vocab[i * max_w]);
    }
  }
  fclose(f);
  return (0);
}

unsigned long getVocabSize() {
  return (unsigned long) words;
}