1/* K=15 r=1/6 Viterbi decoder for x86 MMX
2 * Mar 2004, Phil Karn, KA9Q
3 * May be used under the terms of the GNU Lesser General Public License (LGPL)
4 */
5#include <mmintrin.h>
6#include <stdio.h>
7#include <stdlib.h>
8#include <memory.h>
9#include "fec.h"
10
11typedef union { unsigned char c[16384]; __m64 v[2048];} decision_t;
12typedef union { unsigned short s[16384]; __m64 v[4096];} metric_t;
13
14static union branchtab615 { unsigned short s[8192]; __m64 v[2048];} Branchtab615[6];
15static int Init = 0;
16
17/* State info for instance of Viterbi decoder */
18struct v615 {
19  metric_t metrics1; /* path metric buffer 1 */
20  metric_t metrics2; /* path metric buffer 2 */
21  void *dp;          /* Pointer to current decision */
22  metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */
23  void *decisions;   /* Beginning of decisions for block */
24};
25
26/* Initialize Viterbi decoder for start of new frame */
27int init_viterbi615_mmx(void *p,int starting_state){
28  struct v615 *vp = p;
29  int i;
30
31  if(p == NULL)
32    return -1;
33  for(i=0;i<16384;i++)
34    vp->metrics1.s[i] = 5000;
35
36  vp->old_metrics = &vp->metrics1;
37  vp->new_metrics = &vp->metrics2;
38  vp->dp = vp->decisions;
39  vp->old_metrics->s[starting_state & 16383] = 0; /* Bias known start state */
40  return 0;
41}
42
43/* Create a new instance of a Viterbi decoder */
44void *create_viterbi615_mmx(int len){
45  struct v615 *vp;
46
47  if(!Init){
48    int polys[6] = { V615POLYA,V615POLYB,V615POLYC,V615POLYD,V615POLYE,V615POLYF };
49    set_viterbi615_polynomial_mmx(polys);
50  }
51
52  if((vp = (struct v615 *)malloc(sizeof(struct v615))) == NULL)
53    return NULL;
54  if((vp->decisions = malloc((len+14)*sizeof(decision_t))) == NULL){
55    free(vp);
56    return NULL;
57  }
58  init_viterbi615_mmx(vp,0);
59  return vp;
60}
61
62void set_viterbi615_polynomial_mmx(int polys[6]){
63  int state;
64  int i;
65
66  for(state=0;state < 8192;state++){
67    for(i=0;i<6;i++)
68      Branchtab615[i].s[state] = (polys[i] < 0) ^ parity((2*state) & abs(polys[i])) ? 255 : 0;
69  }
70  Init++;
71}
72
73/* Viterbi chainback */
74int chainback_viterbi615_mmx(
75      void *p,
76      unsigned char *data, /* Decoded output data */
77      unsigned int nbits, /* Number of data bits */
78      unsigned int endstate){ /* Terminal encoder state */
79  struct v615 *vp = p;
80  decision_t *d;
81
82  if(p == NULL)
83    return -1;
84
85  d = (decision_t *)vp->decisions;
86
87  endstate %= 16384;
88
89  /* The store into data[] only needs to be done every 8 bits.
90   * But this avoids a conditional branch, and the writes will
91   * combine in the cache anyway
92   */
93  d += 14; /* Look past tail */
94  while(nbits-- != 0){
95    int k;
96
97    k = d[nbits].c[endstate] & 1;
98    endstate = (k << 13) | (endstate >> 1);
99    data[nbits>>3] = endstate >> 6;
100  }
101  return 0;
102}
103
104/* Delete instance of a Viterbi decoder */
105void delete_viterbi615_mmx(void *p){
106  struct v615 *vp = p;
107
108  if(vp != NULL){
109    free(vp->decisions);
110    free(vp);
111  }
112}
113
114
115int update_viterbi615_blk_mmx(void *p,unsigned char *syms,int nbits){
116  struct v615 *vp = p;
117  decision_t *d;
118
119  if(p == NULL)
120    return -1;
121
122  d = (decision_t *)vp->dp;
123
124  while(nbits--){
125    __m64 sym0v,sym1v,sym2v,sym3v,sym4v,sym5v;
126    void *tmp;
127    int i;
128
129    /* Splat the 0th symbol across sym0v, the 1st symbol across sym1v, etc */
130    sym0v = _mm_set1_pi16(syms[0]);
131    sym1v = _mm_set1_pi16(syms[1]);
132    sym2v = _mm_set1_pi16(syms[2]);
133    sym3v = _mm_set1_pi16(syms[3]);
134    sym4v = _mm_set1_pi16(syms[4]);
135    sym5v = _mm_set1_pi16(syms[5]);
136    syms += 6;
137
138    for(i=0;i<2048;i++){
139      __m64 decision0,decision1,metric,m_metric,m0,m1,m2,m3,survivor0,survivor1;
140
141      /* Form branch metrics
142       * Because Branchtab takes on values 0 and 255, and the values of sym?v are offset binary in the range 0-255,
143       * the XOR operations constitute conditional negation.
144       * metric and m_metric (-metric) are in the range 0-1530
145       */
146      m0 = _mm_add_pi16(_mm_xor_si64(Branchtab615[0].v[i],sym0v),_mm_xor_si64(Branchtab615[1].v[i],sym1v));
147      m1 = _mm_add_pi16(_mm_xor_si64(Branchtab615[2].v[i],sym2v),_mm_xor_si64(Branchtab615[3].v[i],sym3v));
148      m2 = _mm_add_pi16(_mm_xor_si64(Branchtab615[4].v[i],sym4v),_mm_xor_si64(Branchtab615[5].v[i],sym5v));
149      metric = _mm_add_pi16(m0,_mm_add_pi16(m1,m2));
150      m_metric = _mm_sub_pi16(_mm_set1_pi16(1530),metric);
151
152      /* Add branch metrics to path metrics */
153      m0 = _mm_add_pi16(vp->old_metrics->v[i],metric);
154      m3 = _mm_add_pi16(vp->old_metrics->v[2048+i],metric);
155      m1 = _mm_add_pi16(vp->old_metrics->v[2048+i],m_metric);
156      m2 = _mm_add_pi16(vp->old_metrics->v[i],m_metric);
157
158      /* Compare and select
159       * There's no packed min instruction in MMX, so we use modulo arithmetic
160       * to form the decisions and then do the select the hard way
161       */
162      decision0 = _mm_cmpgt_pi16(_mm_sub_pi16(m0,m1),_mm_setzero_si64());
163      decision1 = _mm_cmpgt_pi16(_mm_sub_pi16(m2,m3),_mm_setzero_si64());
164      survivor0 = _mm_or_si64(_mm_and_si64(decision0,m1),_mm_andnot_si64(decision0,m0));
165      survivor1 = _mm_or_si64(_mm_and_si64(decision1,m3),_mm_andnot_si64(decision1,m2));
166
167      /* Merge decisions and store as bytes */
168      d->v[i] = _mm_unpacklo_pi8(_mm_packs_pi16(decision0,_mm_setzero_si64()),_mm_packs_pi16(decision1,_mm_setzero_si64()));
169
170      /* Store surviving metrics */
171      vp->new_metrics->v[2*i] = _mm_unpacklo_pi16(survivor0,survivor1);
172      vp->new_metrics->v[2*i+1] = _mm_unpackhi_pi16(survivor0,survivor1);
173    }
174    d++;
175    /* Swap pointers to old and new metrics */
176    tmp = vp->old_metrics;
177    vp->old_metrics = vp->new_metrics;
178    vp->new_metrics = tmp;
179  }
180  vp->dp = d;
181  _mm_empty();
182  return 0;
183}
184