1 /* K=9 r=1/2 Viterbi decoder for PowerPC G4/G5 Altivec
2 * Copyright Feb 2004, Phil Karn, KA9Q
3 * May be used under the terms of the GNU Lesser General Public License (LGPL)
4 */
5 #include <stdio.h>
6 #include <stdlib.h>
7 #include <memory.h>
8 #include <sys/sysctl.h>
9 #include "fec.h"
10
11 typedef union { unsigned char c[256]; vector bool char v[16]; } decision_t;
12 typedef union { unsigned char c[256]; vector unsigned char v[16]; } metric_t;
13
14 static union branchtab29 { unsigned char c[128]; vector unsigned char v[8]; } Branchtab29[2];
15 static int Init = 0;
16
17 /* State info for instance of Viterbi decoder */
18 struct v29 {
19 metric_t metrics1; /* path metric buffer 1 */
20 metric_t metrics2; /* path metric buffer 2 */
21 decision_t *dp; /* Pointer to current decision */
22 metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */
23 decision_t *decisions; /* Beginning of decisions for block */
24 };
25
26 /* Initialize Viterbi decoder for start of new frame */
init_viterbi29_av(void * p,int starting_state)27 int init_viterbi29_av(void *p,int starting_state){
28 struct v29 *vp = p;
29 int i;
30
31 if(p == NULL)
32 return -1;
33 for(i=0;i<16;i++)
34 vp->metrics1.v[i] = (vector unsigned char)(63);
35
36 vp->old_metrics = &vp->metrics1;
37 vp->new_metrics = &vp->metrics2;
38 vp->dp = vp->decisions;
39 vp->old_metrics->c[starting_state & 255] = 0; /* Bias known start state */
40 return 0;
41 }
42
set_viterbi29_polynomial_av(int polys[2])43 void set_viterbi29_polynomial_av(int polys[2]){
44 int state;
45
46 for(state=0;state < 128;state++){
47 Branchtab29[0].c[state] = (polys[0] < 0) ^ parity((2*state) & abs(polys[0])) ? 255 : 0;
48 Branchtab29[1].c[state] = (polys[1] < 0) ^ parity((2*state) & abs(polys[1])) ? 255 : 0;
49 }
50 Init++;
51 }
52
53 /* Create a new instance of a Viterbi decoder */
create_viterbi29_av(int len)54 void *create_viterbi29_av(int len){
55 struct v29 *vp;
56
57 if(!Init){
58 int polys[2] = { V29POLYA,V29POLYB };
59 set_viterbi29_polynomial_av(polys);
60 }
61 if((vp = (struct v29 *)malloc(sizeof(struct v29))) == NULL)
62 return NULL;
63 if((vp->decisions = (decision_t *)malloc((len+8)*sizeof(decision_t))) == NULL){
64 free(vp);
65 return NULL;
66 }
67 init_viterbi29_av(vp,0);
68 return vp;
69 }
70
71 /* Viterbi chainback */
chainback_viterbi29_av(void * p,unsigned char * data,unsigned int nbits,unsigned int endstate)72 int chainback_viterbi29_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 v29 *vp = p;
78 decision_t *d;
79
80 if(p == NULL)
81 return -1;
82 d = (decision_t *)vp->decisions;
83 /* Make room beyond the end of the encoder register so we can
84 * accumulate a full byte of decoded data
85 */
86 endstate %= 256;
87
88 /* The store into data[] only needs to be done every 8 bits.
89 * But this avoids a conditional branch, and the writes will
90 * combine in the cache anyway
91 */
92 d += 8; /* Look past tail */
93 while(nbits-- != 0){
94 int k;
95
96 k = d[nbits].c[endstate] & 1;
97 data[nbits>>3] = endstate = (endstate >> 1) | (k << 7);
98 }
99 return 0;
100 }
101
102
103 /* Delete instance of a Viterbi decoder */
delete_viterbi29_av(void * p)104 void delete_viterbi29_av(void *p){
105 struct v29 *vp = p;
106
107 if(vp != NULL){
108 free(vp->decisions);
109 free(vp);
110 }
111 }
112
113
update_viterbi29_blk_av(void * p,unsigned char * syms,int nbits)114 int update_viterbi29_blk_av(void *p,unsigned char *syms,int nbits){
115 struct v29 *vp = p;
116 decision_t *d;
117 int i;
118
119 if(p == NULL)
120 return -1;
121 d = (decision_t *)vp->dp;
122
123 while(nbits--){
124 vector unsigned char sym1v,sym2v;
125 void *tmp;
126
127 /* All this seems necessary just to load a byte into all elements of a vector! */
128 sym1v = vec_perm(vec_ld(0,syms),vec_ld(1,syms),vec_lvsl(0,syms)); /* sym1v.0 = syms[0]; sym1v.1 = syms[1] */
129 sym2v = vec_splat(sym1v,1); /* Splat syms[1] across sym2v */
130 sym1v = vec_splat(sym1v,0); /* Splat syms[0] across sym1v */
131 syms += 2;
132
133 for(i=0;i<8;i++){
134 vector bool char decision0,decision1;
135 vector unsigned char metric,m_metric,m0,m1,m2,m3,survivor0,survivor1;
136
137 /* Form branch metrics */
138 metric = vec_avg(vec_xor(Branchtab29[0].v[i],sym1v),vec_xor(Branchtab29[1].v[i],sym2v));
139 metric = vec_sr(metric,(vector unsigned char)(3));
140 m_metric = (vector unsigned char)(31) - metric;
141
142 /* Add branch metrics to path metrics */
143 m0 = vec_adds(vp->old_metrics->v[i],metric);
144 m3 = vec_adds(vp->old_metrics->v[8+i],metric);
145 m1 = vec_adds(vp->old_metrics->v[8+i],m_metric);
146 m2 = vec_adds(vp->old_metrics->v[i],m_metric);
147
148 /* Compare and select first set */
149 decision0 = vec_cmpgt(m0,m1);
150 decision1 = vec_cmpgt(m2,m3);
151 survivor0 = vec_min(m0,m1);
152 survivor1 = vec_min(m2,m3);
153
154 /* Interleave and store decisions and survivors */
155 d->v[2*i] = vec_mergeh(decision0,decision1);
156 d->v[2*i+1] = vec_mergel(decision0,decision1);
157 vp->new_metrics->v[2*i] = vec_mergeh(survivor0,survivor1);
158 vp->new_metrics->v[2*i+1] = vec_mergel(survivor0,survivor1);
159 }
160 d++;
161 /* renormalize if necessary */
162 if(vp->new_metrics->c[0] >= 50){
163 int i;
164 vector unsigned char scale0,scale1;
165
166 /* Find smallest metric and splat */
167 scale0 = vp->new_metrics->v[0];
168 scale1 = vp->new_metrics->v[1];
169 for(i=2;i<16;i+=2){
170 scale0 = vec_min(scale0,vp->new_metrics->v[i]);
171 scale1 = vec_min(scale1,vp->new_metrics->v[i+1]);
172 }
173 scale0 = vec_min(scale0,scale1);
174 scale0 = vec_min(scale0,vec_sld(scale0,scale0,8));
175 scale0 = vec_min(scale0,vec_sld(scale0,scale0,4));
176 scale0 = vec_min(scale0,vec_sld(scale0,scale0,2));
177 scale0 = vec_min(scale0,vec_sld(scale0,scale0,1));
178
179 /* Now subtract from all metrics */
180 for(i=0;i<16;i++)
181 vp->new_metrics->v[i] = vec_subs(vp->new_metrics->v[i],scale0);
182 }
183 /* Swap pointers to old and new metrics */
184 tmp = vp->old_metrics;
185 vp->old_metrics = vp->new_metrics;
186 vp->new_metrics = tmp;
187 }
188 vp->dp = d;
189 return 0;
190 }
191