1/* K=15 r=1/6 Viterbi decoder for PowerPC G4/G5 Altivec vector instructions
2 * 8-bit offset-binary soft decision samples
3 * Copyright Mar 2004, Phil Karn, KA9Q
4 * May be used under the terms of the GNU Lesser General Public License (LGPL)
5 */
6#include <stdio.h>
7#include <stdlib.h>
8#include <memory.h>
9#include <limits.h>
10#include "fec.h"
11
12typedef union { unsigned char c[128][16]; vector unsigned char v[128]; } decision_t;
13typedef union { unsigned short s[16384]; vector unsigned short v[2048]; } metric_t;
14
15static union branchtab615 { unsigned short s[8192]; vector unsigned short v[1024];} Branchtab615[6];
16static int Init = 0;
17
18/* State info for instance of Viterbi decoder */
19struct v615 {
20  metric_t metrics1; /* path metric buffer 1 */
21  metric_t metrics2; /* path metric buffer 2 */
22  void *dp;          /* Pointer to current decision */
23  metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */
24  void *decisions;   /* Beginning of decisions for block */
25};
26
27/* Initialize Viterbi decoder for start of new frame */
28int init_viterbi615_av(void *p,int starting_state){
29  struct v615 *vp = p;
30  int i;
31
32  if(p == NULL)
33    return -1;
34
35  for(i=0;i<2048;i++)
36    vp->metrics1.v[i] = (vector unsigned short)(5000);
37
38  vp->old_metrics = &vp->metrics1;
39  vp->new_metrics = &vp->metrics2;
40  vp->dp = vp->decisions;
41  vp->old_metrics->s[starting_state & 16383] = 0; /* Bias known start state */
42  return 0;
43}
44
45/* Create a new instance of a Viterbi decoder */
46void *create_viterbi615_av(int len){
47  struct v615 *vp;
48
49  if(!Init){
50    int polys[6] = { V615POLYA,V615POLYB,V615POLYC,V615POLYD,V615POLYE,V615POLYF };
51    set_viterbi615_polynomial_av(polys);
52  }
53  vp = (struct v615 *)malloc(sizeof(struct v615));
54  vp->decisions = malloc(sizeof(decision_t)*(len+14));
55  init_viterbi615_av(vp,0);
56  return vp;
57}
58
59void set_viterbi615_polynomial_av(int polys[6]){
60  int state;
61  int i;
62
63  for(state=0;state < 8192;state++){
64    for(i=0;i<6;i++)
65      Branchtab615[i].s[state] = (polys[i] < 0) ^ parity((2*state) & abs(polys[i])) ? 255 : 0;
66  }
67  Init++;
68}
69
70
71/* Viterbi chainback */
72int chainback_viterbi615_av(
73      void *p,
74      unsigned char *data, /* Decoded output data */
75      unsigned int nbits, /* Number of data bits */
76      unsigned int endstate){ /* Terminal encoder state */
77  struct v615 *vp = p;
78  decision_t *d = (decision_t *)vp->decisions;
79  int path_metric;
80
81  endstate %= 16384;
82
83  path_metric = vp->old_metrics->s[endstate];
84
85  /* The store into data[] only needs to be done every 8 bits.
86   * But this avoids a conditional branch, and the writes will
87   * combine in the cache anyway
88   */
89  d += 14; /* Look past tail */
90  while(nbits-- != 0){
91    int k;
92
93    k = (d[nbits].c[endstate >> 7][endstate & 15] & (0x80 >> ((endstate>>4)&7)) ) ? 1 : 0;
94    endstate = (k << 13) | (endstate >> 1);
95    data[nbits>>3] = endstate >> 6;
96  }
97  return path_metric;
98}
99
100/* Delete instance of a Viterbi decoder */
101void delete_viterbi615_av(void *p){
102  struct v615 *vp = p;
103
104  if(vp != NULL){
105    free(vp->decisions);
106    free(vp);
107  }
108}
109
110int update_viterbi615_blk_av(void *p,unsigned char *syms,int nbits){
111  struct v615 *vp = p;
112  decision_t *d = (decision_t *)vp->dp;
113  int path_metric = 0;
114  vector unsigned char decisions = (vector unsigned char)(0);
115
116  while(nbits--){
117    vector unsigned short symv,sym0v,sym1v,sym2v,sym3v,sym4v,sym5v;
118    vector unsigned char s;
119    void *tmp;
120    int i;
121
122    /* Splat the 0th symbol across sym0v, the 1st symbol across sym1v, etc */
123    s = (vector unsigned char)vec_perm(vec_ld(0,syms),vec_ld(5,syms),vec_lvsl(0,syms));
124
125    symv = (vector unsigned short)vec_mergeh((vector unsigned char)(0),s);    /* Unsigned byte->word unpack */
126    sym0v = vec_splat(symv,0);
127    sym1v = vec_splat(symv,1);
128    sym2v = vec_splat(symv,2);
129    sym3v = vec_splat(symv,3);
130    sym4v = vec_splat(symv,4);
131    sym5v = vec_splat(symv,5);
132    syms += 6;
133
134    for(i=0;i<1024;i++){
135      vector bool short decision0,decision1;
136      vector unsigned short metric,m_metric,m0,m1,m2,m3,survivor0,survivor1;
137
138      /* Form branch metrics
139       * Because Branchtab takes on values 0 and 255, and the values of sym?v are offset binary in the range 0-255,
140       * the XOR operations constitute conditional negation.
141       * metric and m_metric (-metric) are in the range 0-1530
142       */
143      m0 = vec_add(vec_xor(Branchtab615[0].v[i],sym0v),vec_xor(Branchtab615[1].v[i],sym1v));
144      m1 = vec_add(vec_xor(Branchtab615[2].v[i],sym2v),vec_xor(Branchtab615[3].v[i],sym3v));
145      m2 = vec_add(vec_xor(Branchtab615[4].v[i],sym4v),vec_xor(Branchtab615[5].v[i],sym5v));
146      metric = vec_add(m0,m1);
147      metric = vec_add(metric,m2);
148      m_metric = vec_sub((vector unsigned short)(1530),metric);
149
150      /* Add branch metrics to path metrics */
151      m0 = vec_adds(vp->old_metrics->v[i],metric);
152      m3 = vec_adds(vp->old_metrics->v[1024+i],metric);
153      m1 = vec_adds(vp->old_metrics->v[1024+i],m_metric);
154      m2 = vec_adds(vp->old_metrics->v[i],m_metric);
155
156      /* Compare and select */
157      decision0 = vec_cmpgt(m0,m1);
158      decision1 = vec_cmpgt(m2,m3);
159      survivor0 = vec_min(m0,m1);
160      survivor1 = vec_min(m2,m3);
161
162      /* Store decisions and survivors.
163       * To save space without SSE2's handy PMOVMSKB instruction, we pack and store them in
164       * a funny interleaved fashion that we undo in the chainback function.
165       */
166      decisions = vec_add(decisions,decisions); /* Shift each byte 1 bit to the left */
167
168      /* Booleans are either 0xff or 0x00. Subtracting 0x00 leaves the lsb zero; subtracting
169       * 0xff is equivalent to adding 1, which sets the lsb.
170       */
171      decisions = vec_sub(decisions,(vector unsigned char)vec_pack(vec_mergeh(decision0,decision1),vec_mergel(decision0,decision1)));
172
173      vp->new_metrics->v[2*i] = vec_mergeh(survivor0,survivor1);
174      vp->new_metrics->v[2*i+1] = vec_mergel(survivor0,survivor1);
175
176      if((i % 8) == 7){
177	/* We've accumulated a total of 128 decisions, stash and start again */
178	d->v[i>>3] = decisions; /* No need to clear, the new bits will replace the old */
179      }
180    }
181#if 0
182    /* Experimentally determine metric spread
183     * The results are fixed for a given code and input symbol size
184     */
185    {
186      int i;
187      vector unsigned short min_metric;
188      vector unsigned short max_metric;
189      union { vector unsigned short v; unsigned short s[8];} t;
190      int minimum,maximum;
191      static int max_spread = 0;
192
193      min_metric = max_metric = vp->new_metrics->v[0];
194      for(i=1;i<2048;i++){
195	min_metric = vec_min(min_metric,vp->new_metrics->v[i]);
196	max_metric = vec_max(max_metric,vp->new_metrics->v[i]);
197      }
198      min_metric = vec_min(min_metric,vec_sld(min_metric,min_metric,8));
199      max_metric = vec_max(max_metric,vec_sld(max_metric,max_metric,8));
200      min_metric = vec_min(min_metric,vec_sld(min_metric,min_metric,4));
201      max_metric = vec_max(max_metric,vec_sld(max_metric,max_metric,4));
202      min_metric = vec_min(min_metric,vec_sld(min_metric,min_metric,2));
203      max_metric = vec_max(max_metric,vec_sld(max_metric,max_metric,2));
204
205      t.v = min_metric;
206      minimum = t.s[0];
207      t.v = max_metric;
208      maximum = t.s[0];
209      if(maximum-minimum > max_spread){
210	max_spread = maximum-minimum;
211	printf("metric spread = %d\n",max_spread);
212      }
213    }
214#endif
215
216    /* Renormalize if necessary. This deserves some explanation.
217
218     * The maximum possible spread, found by experiment, for 4-bit symbols is 405; for 8 bit symbols, it's 12750.
219     * So by looking at one arbitrary metric we can tell if any of them have possibly saturated.
220     * However, this is very conservative. Large spreads occur only at very high Eb/No, where
221     * saturating a bad path metric doesn't do much to increase its chances of being erroneously chosen as a survivor.
222
223     * At more interesting (low) Eb/No ratios, the spreads are much smaller so our chances of saturating a metric
224     * by not not normalizing when we should are extremely low. So either way, the risk to performance is small.
225
226     * All this is borne out by experiment.
227     */
228    if(vp->new_metrics->s[0] >= USHRT_MAX-12750){
229      vector unsigned short scale;
230      union { vector unsigned short v; unsigned short s[8];} t;
231
232      /* Find smallest metric and splat */
233      scale = vp->new_metrics->v[0];
234      for(i=1;i<2048;i++)
235	scale = vec_min(scale,vp->new_metrics->v[i]);
236
237      scale = vec_min(scale,vec_sld(scale,scale,8));
238      scale = vec_min(scale,vec_sld(scale,scale,4));
239      scale = vec_min(scale,vec_sld(scale,scale,2));
240
241      /* Subtract it from all metrics
242       * Work backwards to try to improve the cache hit ratio, assuming LRU
243       */
244      for(i=2047;i>=0;i--)
245	vp->new_metrics->v[i] = vec_subs(vp->new_metrics->v[i],scale);
246      t.v = scale;
247      path_metric += t.s[0];
248    }
249    d++;
250    /* Swap pointers to old and new metrics */
251    tmp = vp->old_metrics;
252    vp->old_metrics = vp->new_metrics;
253    vp->new_metrics = tmp;
254  }
255  vp->dp = d;
256  return path_metric;
257}
258