1*638691a0SAndroid Build Coastguard Worker /* K=9 r=1/2 Viterbi decoder for SSE2
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 <emmintrin.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]; __m128i v[16];} metric_t;
12*638691a0SAndroid Build Coastguard Worker typedef union { unsigned long w[8]; unsigned char c[32];} decision_t;
13*638691a0SAndroid Build Coastguard Worker
14*638691a0SAndroid Build Coastguard Worker union branchtab29 { unsigned char c[128]; } Branchtab29_sse2[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 * Don't change this without also changing references in sse2bfly29.s!
19*638691a0SAndroid Build Coastguard Worker */
20*638691a0SAndroid Build Coastguard Worker struct v29 {
21*638691a0SAndroid Build Coastguard Worker metric_t metrics1; /* path metric buffer 1 */
22*638691a0SAndroid Build Coastguard Worker metric_t metrics2; /* path metric buffer 2 */
23*638691a0SAndroid Build Coastguard Worker decision_t *dp; /* Pointer to current decision */
24*638691a0SAndroid Build Coastguard Worker metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */
25*638691a0SAndroid Build Coastguard Worker decision_t *decisions; /* Beginning of decisions for block */
26*638691a0SAndroid Build Coastguard Worker };
27*638691a0SAndroid Build Coastguard Worker
28*638691a0SAndroid Build Coastguard Worker /* Initialize Viterbi decoder for start of new frame */
init_viterbi29_sse2(void * p,int starting_state)29*638691a0SAndroid Build Coastguard Worker int init_viterbi29_sse2(void *p,int starting_state){
30*638691a0SAndroid Build Coastguard Worker struct v29 *vp = p;
31*638691a0SAndroid Build Coastguard Worker int i;
32*638691a0SAndroid Build Coastguard Worker
33*638691a0SAndroid Build Coastguard Worker for(i=0;i<256;i++)
34*638691a0SAndroid Build Coastguard Worker vp->metrics1.c[i] = 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_sse2(int polys[2])43*638691a0SAndroid Build Coastguard Worker void set_viterbi29_polynomial_sse2(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_sse2[0].c[state] = (polys[0] < 0) ^ parity((2*state) & abs(polys[0])) ? 255 : 0;
48*638691a0SAndroid Build Coastguard Worker Branchtab29_sse2[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
54*638691a0SAndroid Build Coastguard Worker /* Create a new instance of a Viterbi decoder */
create_viterbi29_sse2(int len)55*638691a0SAndroid Build Coastguard Worker void *create_viterbi29_sse2(int len){
56*638691a0SAndroid Build Coastguard Worker void *p;
57*638691a0SAndroid Build Coastguard Worker struct v29 *vp;
58*638691a0SAndroid Build Coastguard Worker
59*638691a0SAndroid Build Coastguard Worker if(!Init){
60*638691a0SAndroid Build Coastguard Worker int polys[2] = {V29POLYA,V29POLYB};
61*638691a0SAndroid Build Coastguard Worker
62*638691a0SAndroid Build Coastguard Worker set_viterbi29_polynomial(polys);
63*638691a0SAndroid Build Coastguard Worker }
64*638691a0SAndroid Build Coastguard Worker /* Ordinary malloc() only returns 8-byte alignment, we need 16 */
65*638691a0SAndroid Build Coastguard Worker if(posix_memalign(&p, sizeof(__m128i),sizeof(struct v29)))
66*638691a0SAndroid Build Coastguard Worker return NULL;
67*638691a0SAndroid Build Coastguard Worker vp = (struct v29 *)p;
68*638691a0SAndroid Build Coastguard Worker if((p = malloc((len+8)*sizeof(decision_t))) == NULL){
69*638691a0SAndroid Build Coastguard Worker free(vp);
70*638691a0SAndroid Build Coastguard Worker return NULL;
71*638691a0SAndroid Build Coastguard Worker }
72*638691a0SAndroid Build Coastguard Worker vp->decisions = (decision_t *)p;
73*638691a0SAndroid Build Coastguard Worker init_viterbi29_sse2(vp,0);
74*638691a0SAndroid Build Coastguard Worker return vp;
75*638691a0SAndroid Build Coastguard Worker }
76*638691a0SAndroid Build Coastguard Worker
77*638691a0SAndroid Build Coastguard Worker
78*638691a0SAndroid Build Coastguard Worker /* Viterbi chainback */
chainback_viterbi29_sse2(void * p,unsigned char * data,unsigned int nbits,unsigned int endstate)79*638691a0SAndroid Build Coastguard Worker int chainback_viterbi29_sse2(
80*638691a0SAndroid Build Coastguard Worker void *p,
81*638691a0SAndroid Build Coastguard Worker unsigned char *data, /* Decoded output data */
82*638691a0SAndroid Build Coastguard Worker unsigned int nbits, /* Number of data bits */
83*638691a0SAndroid Build Coastguard Worker unsigned int endstate){ /* Terminal encoder state */
84*638691a0SAndroid Build Coastguard Worker struct v29 *vp = p;
85*638691a0SAndroid Build Coastguard Worker decision_t *d;
86*638691a0SAndroid Build Coastguard Worker
87*638691a0SAndroid Build Coastguard Worker if(p == NULL)
88*638691a0SAndroid Build Coastguard Worker return -1;
89*638691a0SAndroid Build Coastguard Worker d = vp->decisions;
90*638691a0SAndroid Build Coastguard Worker
91*638691a0SAndroid Build Coastguard Worker /* Make room beyond the end of the encoder register so we can
92*638691a0SAndroid Build Coastguard Worker * accumulate a full byte of decoded data
93*638691a0SAndroid Build Coastguard Worker */
94*638691a0SAndroid Build Coastguard Worker endstate %= 256;
95*638691a0SAndroid Build Coastguard Worker
96*638691a0SAndroid Build Coastguard Worker /* The store into data[] only needs to be done every 8 bits.
97*638691a0SAndroid Build Coastguard Worker * But this avoids a conditional branch, and the writes will
98*638691a0SAndroid Build Coastguard Worker * combine in the cache anyway
99*638691a0SAndroid Build Coastguard Worker */
100*638691a0SAndroid Build Coastguard Worker d += 8; /* Look past tail */
101*638691a0SAndroid Build Coastguard Worker while(nbits-- != 0){
102*638691a0SAndroid Build Coastguard Worker int k;
103*638691a0SAndroid Build Coastguard Worker
104*638691a0SAndroid Build Coastguard Worker k = (d[nbits].c[endstate/8] >> (endstate%8)) & 1;
105*638691a0SAndroid Build Coastguard Worker data[nbits>>3] = endstate = (endstate >> 1) | (k << 7);
106*638691a0SAndroid Build Coastguard Worker }
107*638691a0SAndroid Build Coastguard Worker return 0;
108*638691a0SAndroid Build Coastguard Worker }
109*638691a0SAndroid Build Coastguard Worker
110*638691a0SAndroid Build Coastguard Worker
111*638691a0SAndroid Build Coastguard Worker /* Delete instance of a Viterbi decoder */
delete_viterbi29_sse2(void * p)112*638691a0SAndroid Build Coastguard Worker void delete_viterbi29_sse2(void *p){
113*638691a0SAndroid Build Coastguard Worker struct v29 *vp = p;
114*638691a0SAndroid Build Coastguard Worker
115*638691a0SAndroid Build Coastguard Worker if(vp != NULL){
116*638691a0SAndroid Build Coastguard Worker free(vp->decisions);
117*638691a0SAndroid Build Coastguard Worker free(vp);
118*638691a0SAndroid Build Coastguard Worker }
119*638691a0SAndroid Build Coastguard Worker }
120