srsRAN_4G/srslte/lib/fec/viterbi37_avx2.c

/* Adapted Phil Karn's r=1/3 k=9 viterbi decoder to r=1/3 k=7
 * 
 * K=15 r=1/6 Viterbi decoder for x86 SSE2
 * Copyright Mar 2004, Phil Karn, KA9Q
 * May be used under the terms of the GNU Lesser General Public License (LGPL)
 */

#include <stdio.h>
#include <stdint.h>
#include <stdlib.h>
#include <memory.h>
#include <limits.h>
#include "parity.h"

//#define DEBUG

#ifdef LV_HAVE_SSE

#include <emmintrin.h>
#include <tmmintrin.h>
#include <immintrin.h>
#include <emmintrin.h>
#define _mm256_set_m128i(v0, v1)  _mm256_insertf128_si256(_mm256_castsi128_si256(v1), (v0), 1)

#define _mm256_setr_m128i(v0, v1) _mm256_set_m128i((v1), (v0))

typedef union { 
  unsigned char c[64]; 
  __m128i       v[4];   
} metric_t;
typedef union { 
  unsigned int  w[2]; 
  unsigned char  c[8]; 
  unsigned short s[4]; 
  __m64 v;  
} decision_t;

union branchtab27 { 
  unsigned char c[32]; 
  __m256i       v;  
} Branchtab37_sse2[3];

int firstGo;
/* State info for instance of Viterbi decoder */
struct v37 {
  metric_t metrics1; /* path metric buffer 1 */
  metric_t metrics2; /* path metric buffer 2 */
  decision_t *dp; /* Pointer to current decision */
  metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */
  decision_t *decisions; /* Beginning of decisions for block */
  uint32_t len; 
};

void set_viterbi37_polynomial_avx2(int polys[3]) {
  int state;

  for(state=0;state < 32;state++){
    Branchtab37_sse2[0].c[state] = (polys[0] < 0) ^ parity((2*state) & polys[0]) ? 255:0;
    Branchtab37_sse2[1].c[state] = (polys[1] < 0) ^ parity((2*state) & polys[1]) ? 255:0;
    Branchtab37_sse2[2].c[state] = (polys[2] < 0) ^ parity((2*state) & polys[2]) ? 255:0;
  }
}

void clear_v37_avx2(struct v37 *vp) {
  bzero(vp->decisions, sizeof(decision_t)*vp->len);
  vp->dp = NULL; 
  bzero(&vp->metrics1, sizeof(metric_t));
  bzero(&vp->metrics2, sizeof(metric_t));
  vp->old_metrics = NULL; 
  vp->new_metrics = NULL; 
}


/* Initialize Viterbi decoder for start of new frame */
int init_viterbi37_avx2(void *p, int starting_state) {
  struct v37 *vp = p;
  uint32_t i;
  firstGo = 1;
  for(i=0;i<64;i++)
    vp->metrics1.c[i] = 63;

  clear_v37_avx2(vp);

  vp->old_metrics = &vp->metrics1;
  vp->new_metrics = &vp->metrics2;
  vp->dp = vp->decisions;
  if (starting_state != -1) {
    vp->old_metrics->c[starting_state & 63] = 0; /* Bias known start state */
  }
  return 0;
}

/* Create a new instance of a Viterbi decoder */
void *create_viterbi37_avx2(int polys[3], uint32_t len) {
  void *p;
  struct v37 *vp;

  set_viterbi37_polynomial_avx2(polys);
  
  /* Ordinary malloc() only returns 8-byte alignment, we need 16 */
  if(posix_memalign(&p, sizeof(__m128i),sizeof(struct v37)))
    return NULL;

  vp = (struct v37 *)p;
  if(posix_memalign(&p, sizeof(__m128i),(len+6)*sizeof(decision_t))) {
    free(vp);
    return NULL;
  }
  vp->decisions = (decision_t *)p;
  vp->len = len+6;
  return vp;
}


/* Viterbi chainback */
int chainback_viterbi37_avx2(
      void *p,
      uint8_t *data, /* Decoded output data */
      uint32_t nbits, /* Number of data bits */
      uint32_t endstate) { /* Terminal encoder state */
  struct v37 *vp = p;
  
  if (p == NULL)
    return -1;

  decision_t *d = (decision_t *)vp->decisions;

  /* Make room beyond the end of the encoder register so we can
   * accumulate a full byte of decoded data
   */
  endstate %= 64;
  endstate <<= 2;
  
  /* The store into data[] only needs to be done every 8 bits.
   * But this avoids a conditional branch, and the writes will
   * combine in the cache anyway
   */
  d += 6; /* Look past tail */
  while(nbits--) {
    int k;

    k = (d[nbits].c[(endstate>>2)/8] >> ((endstate>>2)%8)) & 1;
    endstate = (endstate >> 1) | (k << 7);
    data[nbits] = k;
    //printf("nbits=%d, endstate=%3d, k=%d, w[0]=%d, w[1]=%d, c=%d\n", nbits, endstate, k, d[nbits].s[1]&1, d[nbits].s[2]&1, d[nbits].c[(endstate>>2)/8]&1);
  }
  return 0;
}

/* Delete instance of a Viterbi decoder */
void delete_viterbi37_avx2(void *p){
  struct v37 *vp = p;

  if(vp != NULL){
    free(vp->decisions);
    free(vp);
  }
}
void printer_256i(char *s, __m256i val) {
  
  printf("%s: ", s);
  
  uint8_t *x = (uint8_t*) &val; 
  for (int i=0;i<32;i++) {
    printf("%3d, ", x[i]);
  }
  printf("\n");
}

void printer_128i(char *s, __m128i val) {
  
  printf("%s: ", s);
  
  uint8_t *x = (uint8_t*) &val; 
  for (int i=0;i<16;i++) {
    printf("%3d, ", x[i]);
  }
  printf("\n");
}

void printer_m64(char *s, __m64 val) {
  
  printf("%s: ", s);
  
  uint8_t *x = (uint8_t*) &val; 
  for (int i=0;i<8;i++) {
    printf("%3d, ", x[i]);
  }
  printf("\n");
}


void update_viterbi37_blk_avx2(void *p,unsigned char *syms,int nbits, uint32_t *best_state) {
  struct v37 *vp = p;
  decision_t *d;

  if(p == NULL)
    return;
  
#ifdef DEBUG
  printf("[");
#endif
  
  d = (decision_t *) vp->dp;
  
  for (int s=0;s<nbits;s++) {
    memset(d+s,0,sizeof(decision_t));
  }
  
  while(nbits--) {
    __m256i sym0v,sym1v,sym2v;
    
    void *tmp;

    sym0v = _mm256_set1_epi8(syms[0]);
    sym1v = _mm256_set1_epi8(syms[1]);
    sym2v = _mm256_set1_epi8(syms[2]);

    syms += 3;
    
     
    __m256i decision0,decision1,survivor0,survivor1,metric,m_metric,m0,m1,m2,m3;

    /* Form branch metrics */
    m0 = _mm256_avg_epu8(_mm256_xor_si256(Branchtab37_sse2[0].v,sym0v),_mm256_xor_si256(Branchtab37_sse2[1].v,sym1v));
    metric = _mm256_avg_epu8(_mm256_xor_si256(Branchtab37_sse2[2].v,sym2v),m0);
    
#ifdef DEBUG
      print_128i("metric_initial", metric);
#endif
      /* There's no packed bytes right shift in SSE2, so we use the word version and mask
       */
    metric = _mm256_srli_epi16(metric,3);
    metric = _mm256_and_si256(metric,_mm256_set1_epi8(31));
    m_metric = _mm256_sub_epi8(_mm256_set1_epi8(31),metric);
      
#ifdef DEBUG
      print_128i("metric        ", metric);
      print_128i("m_metric      ", m_metric);
#endif

    __m256i temp = _mm256_set_m128i( vp->old_metrics->v[1], vp->old_metrics->v[0]);
    m0 = _mm256_add_epi8(temp,metric); 
    m2 = _mm256_add_epi8(temp,m_metric);
     
    temp = _mm256_set_m128i( vp->old_metrics->v[3], vp->old_metrics->v[2]);
    m3 = _mm256_add_epi8(temp,metric);
    m1 = _mm256_add_epi8(temp,m_metric);
     
     /* Compare and select, using modulo arithmetic */
    decision0 = _mm256_cmpgt_epi8(_mm256_sub_epi8(m0,m1),_mm256_setzero_si256());
    decision1 =_mm256_cmpgt_epi8(_mm256_sub_epi8(m2,m3),_mm256_setzero_si256());
    survivor0 = _mm256_or_si256(_mm256_and_si256(decision0,m1),_mm256_andnot_si256(decision0,m0));
    survivor1 = _mm256_or_si256(_mm256_and_si256(decision1,m3),_mm256_andnot_si256(decision1,m2));
 
    unsigned int x  = _mm256_movemask_epi8(_mm256_unpackhi_epi8(decision0,decision1));
    unsigned int y = _mm256_movemask_epi8(_mm256_unpacklo_epi8(decision0,decision1));
  
    d->s[0] = (short) y;
    d->s[1] = (short) x;
    d->s[2] = (short) (y >>16);
    d->s[3] = (short)(x>> 16);

      
    __m256i unpack;
    unpack =  _mm256_unpacklo_epi8(survivor0,survivor1);
    vp->new_metrics->v[0] =_mm256_castsi256_si128(unpack);

    vp->new_metrics->v[1] = _mm256_extractf128_si256(unpack,1);

    unpack = _mm256_unpackhi_epi8(survivor0,survivor1);

    vp->new_metrics->v[2] =_mm256_castsi256_si128(unpack);
    vp->new_metrics->v[3] = _mm256_extractf128_si256(unpack,1);
       
    __m128i temp1 = vp->new_metrics->v[1];

    vp->new_metrics->v[1] = vp->new_metrics->v[2];
    vp->new_metrics->v[2] = temp1;

    // See if we need to normalize
    if (vp->new_metrics->c[0] > 100) {
      int i;
      uint8_t adjust; 
      __m128i adjustv;
      union { __m128i v; signed short w[8]; } t;
      
      adjustv = vp->new_metrics->v[0];
      for(i=1;i<4;i++) {
        adjustv = _mm_min_epu8(adjustv,vp->new_metrics->v[i]);
      }
      
      adjustv = _mm_min_epu8(adjustv,_mm_srli_si128(adjustv,8));
      adjustv = _mm_min_epu8(adjustv,_mm_srli_si128(adjustv,4));
      adjustv = _mm_min_epu8(adjustv,_mm_srli_si128(adjustv,2));
      
      t.v = adjustv;
      adjust = t.w[0];
      adjustv = _mm_set1_epi8(adjust);
      
      /* We cannot use a saturated subtract, because we often have to adjust by more than SHRT_MAX
       * This is okay since it can't overflow anyway
       */
      for(i=0;i<4;i++)
        vp->new_metrics->v[i] = _mm_sub_epi8(vp->new_metrics->v[i],adjustv);
     
    }

    firstGo = 0;
    d++;
    /* Swap pointers to old and new metrics */
    tmp = vp->old_metrics;
    vp->old_metrics = vp->new_metrics;
    vp->new_metrics = tmp;
  }
  
  if (best_state) {
    uint32_t i, bst=0;
    uint8_t minmetric=UINT8_MAX;
    for (i=0;i<64;i++) {
      if (vp->old_metrics->c[i] <= minmetric) {
        bst = i;
        minmetric = vp->old_metrics->c[i];
      }
    }
    *best_state = bst;
  }

  #ifdef DEBUG
  printf("];\n===========================================\n");
#endif

  vp->dp = d;
}

#endif