1*638691a0SAndroid Build Coastguard Worker /* K=9 r=1/2 Viterbi decoder for PowerPC G4/G5 Altivec
2*638691a0SAndroid Build Coastguard Worker * Copyright Feb 2004, Phil Karn, KA9Q
3*638691a0SAndroid Build Coastguard Worker * May be used under the terms of the GNU Lesser General Public License (LGPL)
4*638691a0SAndroid Build Coastguard Worker */
5*638691a0SAndroid Build Coastguard Worker #include <stdio.h>
6*638691a0SAndroid Build Coastguard Worker #include <stdlib.h>
7*638691a0SAndroid Build Coastguard Worker #include <memory.h>
8*638691a0SAndroid Build Coastguard Worker #include <sys/sysctl.h>
9*638691a0SAndroid Build Coastguard Worker #include "fec.h"
10*638691a0SAndroid Build Coastguard Worker
11*638691a0SAndroid Build Coastguard Worker typedef union { unsigned char c[256]; vector bool char v[16]; } decision_t;
12*638691a0SAndroid Build Coastguard Worker typedef union { unsigned char c[256]; vector unsigned char v[16]; } metric_t;
13*638691a0SAndroid Build Coastguard Worker
14*638691a0SAndroid Build Coastguard Worker static union branchtab29 { unsigned char c[128]; vector unsigned char v[8]; } Branchtab29[2];
15*638691a0SAndroid Build Coastguard Worker static int Init = 0;
16*638691a0SAndroid Build Coastguard Worker
17*638691a0SAndroid Build Coastguard Worker /* State info for instance of Viterbi decoder */
18*638691a0SAndroid Build Coastguard Worker struct v29 {
19*638691a0SAndroid Build Coastguard Worker metric_t metrics1; /* path metric buffer 1 */
20*638691a0SAndroid Build Coastguard Worker metric_t metrics2; /* path metric buffer 2 */
21*638691a0SAndroid Build Coastguard Worker decision_t *dp; /* Pointer to current decision */
22*638691a0SAndroid Build Coastguard Worker metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */
23*638691a0SAndroid Build Coastguard Worker decision_t *decisions; /* Beginning of decisions for block */
24*638691a0SAndroid Build Coastguard Worker };
25*638691a0SAndroid Build Coastguard Worker
26*638691a0SAndroid Build Coastguard Worker /* Initialize Viterbi decoder for start of new frame */
init_viterbi29_av(void * p,int starting_state)27*638691a0SAndroid Build Coastguard Worker int init_viterbi29_av(void *p,int starting_state){
28*638691a0SAndroid Build Coastguard Worker struct v29 *vp = p;
29*638691a0SAndroid Build Coastguard Worker int i;
30*638691a0SAndroid Build Coastguard Worker
31*638691a0SAndroid Build Coastguard Worker if(p == NULL)
32*638691a0SAndroid Build Coastguard Worker return -1;
33*638691a0SAndroid Build Coastguard Worker for(i=0;i<16;i++)
34*638691a0SAndroid Build Coastguard Worker vp->metrics1.v[i] = (vector unsigned char)(63);
35*638691a0SAndroid Build Coastguard Worker
36*638691a0SAndroid Build Coastguard Worker vp->old_metrics = &vp->metrics1;
37*638691a0SAndroid Build Coastguard Worker vp->new_metrics = &vp->metrics2;
38*638691a0SAndroid Build Coastguard Worker vp->dp = vp->decisions;
39*638691a0SAndroid Build Coastguard Worker vp->old_metrics->c[starting_state & 255] = 0; /* Bias known start state */
40*638691a0SAndroid Build Coastguard Worker return 0;
41*638691a0SAndroid Build Coastguard Worker }
42*638691a0SAndroid Build Coastguard Worker
set_viterbi29_polynomial_av(int polys[2])43*638691a0SAndroid Build Coastguard Worker void set_viterbi29_polynomial_av(int polys[2]){
44*638691a0SAndroid Build Coastguard Worker int state;
45*638691a0SAndroid Build Coastguard Worker
46*638691a0SAndroid Build Coastguard Worker for(state=0;state < 128;state++){
47*638691a0SAndroid Build Coastguard Worker Branchtab29[0].c[state] = (polys[0] < 0) ^ parity((2*state) & abs(polys[0])) ? 255 : 0;
48*638691a0SAndroid Build Coastguard Worker Branchtab29[1].c[state] = (polys[1] < 0) ^ parity((2*state) & abs(polys[1])) ? 255 : 0;
49*638691a0SAndroid Build Coastguard Worker }
50*638691a0SAndroid Build Coastguard Worker Init++;
51*638691a0SAndroid Build Coastguard Worker }
52*638691a0SAndroid Build Coastguard Worker
53*638691a0SAndroid Build Coastguard Worker /* Create a new instance of a Viterbi decoder */
create_viterbi29_av(int len)54*638691a0SAndroid Build Coastguard Worker void *create_viterbi29_av(int len){
55*638691a0SAndroid Build Coastguard Worker struct v29 *vp;
56*638691a0SAndroid Build Coastguard Worker
57*638691a0SAndroid Build Coastguard Worker if(!Init){
58*638691a0SAndroid Build Coastguard Worker int polys[2] = { V29POLYA,V29POLYB };
59*638691a0SAndroid Build Coastguard Worker set_viterbi29_polynomial_av(polys);
60*638691a0SAndroid Build Coastguard Worker }
61*638691a0SAndroid Build Coastguard Worker if((vp = (struct v29 *)malloc(sizeof(struct v29))) == NULL)
62*638691a0SAndroid Build Coastguard Worker return NULL;
63*638691a0SAndroid Build Coastguard Worker if((vp->decisions = (decision_t *)malloc((len+8)*sizeof(decision_t))) == NULL){
64*638691a0SAndroid Build Coastguard Worker free(vp);
65*638691a0SAndroid Build Coastguard Worker return NULL;
66*638691a0SAndroid Build Coastguard Worker }
67*638691a0SAndroid Build Coastguard Worker init_viterbi29_av(vp,0);
68*638691a0SAndroid Build Coastguard Worker return vp;
69*638691a0SAndroid Build Coastguard Worker }
70*638691a0SAndroid Build Coastguard Worker
71*638691a0SAndroid Build Coastguard Worker /* Viterbi chainback */
chainback_viterbi29_av(void * p,unsigned char * data,unsigned int nbits,unsigned int endstate)72*638691a0SAndroid Build Coastguard Worker int chainback_viterbi29_av(
73*638691a0SAndroid Build Coastguard Worker void *p,
74*638691a0SAndroid Build Coastguard Worker unsigned char *data, /* Decoded output data */
75*638691a0SAndroid Build Coastguard Worker unsigned int nbits, /* Number of data bits */
76*638691a0SAndroid Build Coastguard Worker unsigned int endstate){ /* Terminal encoder state */
77*638691a0SAndroid Build Coastguard Worker struct v29 *vp = p;
78*638691a0SAndroid Build Coastguard Worker decision_t *d;
79*638691a0SAndroid Build Coastguard Worker
80*638691a0SAndroid Build Coastguard Worker if(p == NULL)
81*638691a0SAndroid Build Coastguard Worker return -1;
82*638691a0SAndroid Build Coastguard Worker d = (decision_t *)vp->decisions;
83*638691a0SAndroid Build Coastguard Worker /* Make room beyond the end of the encoder register so we can
84*638691a0SAndroid Build Coastguard Worker * accumulate a full byte of decoded data
85*638691a0SAndroid Build Coastguard Worker */
86*638691a0SAndroid Build Coastguard Worker endstate %= 256;
87*638691a0SAndroid Build Coastguard Worker
88*638691a0SAndroid Build Coastguard Worker /* The store into data[] only needs to be done every 8 bits.
89*638691a0SAndroid Build Coastguard Worker * But this avoids a conditional branch, and the writes will
90*638691a0SAndroid Build Coastguard Worker * combine in the cache anyway
91*638691a0SAndroid Build Coastguard Worker */
92*638691a0SAndroid Build Coastguard Worker d += 8; /* Look past tail */
93*638691a0SAndroid Build Coastguard Worker while(nbits-- != 0){
94*638691a0SAndroid Build Coastguard Worker int k;
95*638691a0SAndroid Build Coastguard Worker
96*638691a0SAndroid Build Coastguard Worker k = d[nbits].c[endstate] & 1;
97*638691a0SAndroid Build Coastguard Worker data[nbits>>3] = endstate = (endstate >> 1) | (k << 7);
98*638691a0SAndroid Build Coastguard Worker }
99*638691a0SAndroid Build Coastguard Worker return 0;
100*638691a0SAndroid Build Coastguard Worker }
101*638691a0SAndroid Build Coastguard Worker
102*638691a0SAndroid Build Coastguard Worker
103*638691a0SAndroid Build Coastguard Worker /* Delete instance of a Viterbi decoder */
delete_viterbi29_av(void * p)104*638691a0SAndroid Build Coastguard Worker void delete_viterbi29_av(void *p){
105*638691a0SAndroid Build Coastguard Worker struct v29 *vp = p;
106*638691a0SAndroid Build Coastguard Worker
107*638691a0SAndroid Build Coastguard Worker if(vp != NULL){
108*638691a0SAndroid Build Coastguard Worker free(vp->decisions);
109*638691a0SAndroid Build Coastguard Worker free(vp);
110*638691a0SAndroid Build Coastguard Worker }
111*638691a0SAndroid Build Coastguard Worker }
112*638691a0SAndroid Build Coastguard Worker
113*638691a0SAndroid Build Coastguard Worker
update_viterbi29_blk_av(void * p,unsigned char * syms,int nbits)114*638691a0SAndroid Build Coastguard Worker int update_viterbi29_blk_av(void *p,unsigned char *syms,int nbits){
115*638691a0SAndroid Build Coastguard Worker struct v29 *vp = p;
116*638691a0SAndroid Build Coastguard Worker decision_t *d;
117*638691a0SAndroid Build Coastguard Worker int i;
118*638691a0SAndroid Build Coastguard Worker
119*638691a0SAndroid Build Coastguard Worker if(p == NULL)
120*638691a0SAndroid Build Coastguard Worker return -1;
121*638691a0SAndroid Build Coastguard Worker d = (decision_t *)vp->dp;
122*638691a0SAndroid Build Coastguard Worker
123*638691a0SAndroid Build Coastguard Worker while(nbits--){
124*638691a0SAndroid Build Coastguard Worker vector unsigned char sym1v,sym2v;
125*638691a0SAndroid Build Coastguard Worker void *tmp;
126*638691a0SAndroid Build Coastguard Worker
127*638691a0SAndroid Build Coastguard Worker /* All this seems necessary just to load a byte into all elements of a vector! */
128*638691a0SAndroid Build Coastguard Worker sym1v = vec_perm(vec_ld(0,syms),vec_ld(1,syms),vec_lvsl(0,syms)); /* sym1v.0 = syms[0]; sym1v.1 = syms[1] */
129*638691a0SAndroid Build Coastguard Worker sym2v = vec_splat(sym1v,1); /* Splat syms[1] across sym2v */
130*638691a0SAndroid Build Coastguard Worker sym1v = vec_splat(sym1v,0); /* Splat syms[0] across sym1v */
131*638691a0SAndroid Build Coastguard Worker syms += 2;
132*638691a0SAndroid Build Coastguard Worker
133*638691a0SAndroid Build Coastguard Worker for(i=0;i<8;i++){
134*638691a0SAndroid Build Coastguard Worker vector bool char decision0,decision1;
135*638691a0SAndroid Build Coastguard Worker vector unsigned char metric,m_metric,m0,m1,m2,m3,survivor0,survivor1;
136*638691a0SAndroid Build Coastguard Worker
137*638691a0SAndroid Build Coastguard Worker /* Form branch metrics */
138*638691a0SAndroid Build Coastguard Worker metric = vec_avg(vec_xor(Branchtab29[0].v[i],sym1v),vec_xor(Branchtab29[1].v[i],sym2v));
139*638691a0SAndroid Build Coastguard Worker metric = vec_sr(metric,(vector unsigned char)(3));
140*638691a0SAndroid Build Coastguard Worker m_metric = (vector unsigned char)(31) - metric;
141*638691a0SAndroid Build Coastguard Worker
142*638691a0SAndroid Build Coastguard Worker /* Add branch metrics to path metrics */
143*638691a0SAndroid Build Coastguard Worker m0 = vec_adds(vp->old_metrics->v[i],metric);
144*638691a0SAndroid Build Coastguard Worker m3 = vec_adds(vp->old_metrics->v[8+i],metric);
145*638691a0SAndroid Build Coastguard Worker m1 = vec_adds(vp->old_metrics->v[8+i],m_metric);
146*638691a0SAndroid Build Coastguard Worker m2 = vec_adds(vp->old_metrics->v[i],m_metric);
147*638691a0SAndroid Build Coastguard Worker
148*638691a0SAndroid Build Coastguard Worker /* Compare and select first set */
149*638691a0SAndroid Build Coastguard Worker decision0 = vec_cmpgt(m0,m1);
150*638691a0SAndroid Build Coastguard Worker decision1 = vec_cmpgt(m2,m3);
151*638691a0SAndroid Build Coastguard Worker survivor0 = vec_min(m0,m1);
152*638691a0SAndroid Build Coastguard Worker survivor1 = vec_min(m2,m3);
153*638691a0SAndroid Build Coastguard Worker
154*638691a0SAndroid Build Coastguard Worker /* Interleave and store decisions and survivors */
155*638691a0SAndroid Build Coastguard Worker d->v[2*i] = vec_mergeh(decision0,decision1);
156*638691a0SAndroid Build Coastguard Worker d->v[2*i+1] = vec_mergel(decision0,decision1);
157*638691a0SAndroid Build Coastguard Worker vp->new_metrics->v[2*i] = vec_mergeh(survivor0,survivor1);
158*638691a0SAndroid Build Coastguard Worker vp->new_metrics->v[2*i+1] = vec_mergel(survivor0,survivor1);
159*638691a0SAndroid Build Coastguard Worker }
160*638691a0SAndroid Build Coastguard Worker d++;
161*638691a0SAndroid Build Coastguard Worker /* renormalize if necessary */
162*638691a0SAndroid Build Coastguard Worker if(vp->new_metrics->c[0] >= 50){
163*638691a0SAndroid Build Coastguard Worker int i;
164*638691a0SAndroid Build Coastguard Worker vector unsigned char scale0,scale1;
165*638691a0SAndroid Build Coastguard Worker
166*638691a0SAndroid Build Coastguard Worker /* Find smallest metric and splat */
167*638691a0SAndroid Build Coastguard Worker scale0 = vp->new_metrics->v[0];
168*638691a0SAndroid Build Coastguard Worker scale1 = vp->new_metrics->v[1];
169*638691a0SAndroid Build Coastguard Worker for(i=2;i<16;i+=2){
170*638691a0SAndroid Build Coastguard Worker scale0 = vec_min(scale0,vp->new_metrics->v[i]);
171*638691a0SAndroid Build Coastguard Worker scale1 = vec_min(scale1,vp->new_metrics->v[i+1]);
172*638691a0SAndroid Build Coastguard Worker }
173*638691a0SAndroid Build Coastguard Worker scale0 = vec_min(scale0,scale1);
174*638691a0SAndroid Build Coastguard Worker scale0 = vec_min(scale0,vec_sld(scale0,scale0,8));
175*638691a0SAndroid Build Coastguard Worker scale0 = vec_min(scale0,vec_sld(scale0,scale0,4));
176*638691a0SAndroid Build Coastguard Worker scale0 = vec_min(scale0,vec_sld(scale0,scale0,2));
177*638691a0SAndroid Build Coastguard Worker scale0 = vec_min(scale0,vec_sld(scale0,scale0,1));
178*638691a0SAndroid Build Coastguard Worker
179*638691a0SAndroid Build Coastguard Worker /* Now subtract from all metrics */
180*638691a0SAndroid Build Coastguard Worker for(i=0;i<16;i++)
181*638691a0SAndroid Build Coastguard Worker vp->new_metrics->v[i] = vec_subs(vp->new_metrics->v[i],scale0);
182*638691a0SAndroid Build Coastguard Worker }
183*638691a0SAndroid Build Coastguard Worker /* Swap pointers to old and new metrics */
184*638691a0SAndroid Build Coastguard Worker tmp = vp->old_metrics;
185*638691a0SAndroid Build Coastguard Worker vp->old_metrics = vp->new_metrics;
186*638691a0SAndroid Build Coastguard Worker vp->new_metrics = tmp;
187*638691a0SAndroid Build Coastguard Worker }
188*638691a0SAndroid Build Coastguard Worker vp->dp = d;
189*638691a0SAndroid Build Coastguard Worker return 0;
190*638691a0SAndroid Build Coastguard Worker }
191