1*638691a0SAndroid Build Coastguard Worker /* K=7 r=1/2 Viterbi decoder for SSE
2*638691a0SAndroid Build Coastguard Worker * Feb 2004, Phil Karn, KA9Q
3*638691a0SAndroid Build Coastguard Worker */
4*638691a0SAndroid Build Coastguard Worker #include <stdio.h>
5*638691a0SAndroid Build Coastguard Worker #include <stdlib.h>
6*638691a0SAndroid Build Coastguard Worker #include <memory.h>
7*638691a0SAndroid Build Coastguard Worker #include <xmmintrin.h>
8*638691a0SAndroid Build Coastguard Worker #include "fec.h"
9*638691a0SAndroid Build Coastguard Worker
10*638691a0SAndroid Build Coastguard Worker typedef union { unsigned char c[64]; } metric_t;
11*638691a0SAndroid Build Coastguard Worker typedef union { unsigned long w[2]; unsigned char c[8]; __m64 v[1];} decision_t;
12*638691a0SAndroid Build Coastguard Worker union branchtab27 { unsigned char c[32]; __m64 v[4];} Branchtab27_sse[2];
13*638691a0SAndroid Build Coastguard Worker static int Init = 0;
14*638691a0SAndroid Build Coastguard Worker
15*638691a0SAndroid Build Coastguard Worker /* State info for instance of Viterbi decoder
16*638691a0SAndroid Build Coastguard Worker * Don't change this without also changing references in ssebfly27.s!
17*638691a0SAndroid Build Coastguard Worker */
18*638691a0SAndroid Build Coastguard Worker struct v27 {
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 /* Create a new instance of a Viterbi decoder */
create_viterbi27_sse(int len)27*638691a0SAndroid Build Coastguard Worker void *create_viterbi27_sse(int len){
28*638691a0SAndroid Build Coastguard Worker struct v27 *vp;
29*638691a0SAndroid Build Coastguard Worker
30*638691a0SAndroid Build Coastguard Worker if(!Init){
31*638691a0SAndroid Build Coastguard Worker int polys[2] = { V27POLYA, V27POLYB };
32*638691a0SAndroid Build Coastguard Worker
33*638691a0SAndroid Build Coastguard Worker set_viterbi27_polynomial_sse(polys);
34*638691a0SAndroid Build Coastguard Worker }
35*638691a0SAndroid Build Coastguard Worker if((vp = malloc(sizeof(struct v27))) == NULL)
36*638691a0SAndroid Build Coastguard Worker return NULL;
37*638691a0SAndroid Build Coastguard Worker if((vp->decisions = malloc((len+6)*sizeof(decision_t))) == NULL){
38*638691a0SAndroid Build Coastguard Worker free(vp);
39*638691a0SAndroid Build Coastguard Worker return NULL;
40*638691a0SAndroid Build Coastguard Worker }
41*638691a0SAndroid Build Coastguard Worker init_viterbi27(vp,0);
42*638691a0SAndroid Build Coastguard Worker return vp;
43*638691a0SAndroid Build Coastguard Worker }
44*638691a0SAndroid Build Coastguard Worker
set_viterbi27_polynomial_sse(int polys[2])45*638691a0SAndroid Build Coastguard Worker void set_viterbi27_polynomial_sse(int polys[2]){
46*638691a0SAndroid Build Coastguard Worker int state;
47*638691a0SAndroid Build Coastguard Worker
48*638691a0SAndroid Build Coastguard Worker for(state=0;state < 32;state++){
49*638691a0SAndroid Build Coastguard Worker Branchtab27_sse[0].c[state] = (polys[0] < 0) ^ parity((2*state) & abs(polys[0])) ? 255 : 0;
50*638691a0SAndroid Build Coastguard Worker Branchtab27_sse[1].c[state] = (polys[1] < 0) ^ parity((2*state) & abs(polys[1])) ? 255 : 0;
51*638691a0SAndroid Build Coastguard Worker }
52*638691a0SAndroid Build Coastguard Worker Init++;
53*638691a0SAndroid Build Coastguard Worker }
54*638691a0SAndroid Build Coastguard Worker
55*638691a0SAndroid Build Coastguard Worker /* Initialize Viterbi decoder for start of new frame */
init_viterbi27_sse(void * p,int starting_state)56*638691a0SAndroid Build Coastguard Worker int init_viterbi27_sse(void *p,int starting_state){
57*638691a0SAndroid Build Coastguard Worker struct v27 *vp = p;
58*638691a0SAndroid Build Coastguard Worker int i;
59*638691a0SAndroid Build Coastguard Worker
60*638691a0SAndroid Build Coastguard Worker if(p == NULL)
61*638691a0SAndroid Build Coastguard Worker return -1;
62*638691a0SAndroid Build Coastguard Worker for(i=0;i<64;i++)
63*638691a0SAndroid Build Coastguard Worker vp->metrics1.c[i] = 63;
64*638691a0SAndroid Build Coastguard Worker
65*638691a0SAndroid Build Coastguard Worker vp->old_metrics = &vp->metrics1;
66*638691a0SAndroid Build Coastguard Worker vp->new_metrics = &vp->metrics2;
67*638691a0SAndroid Build Coastguard Worker vp->dp = vp->decisions;
68*638691a0SAndroid Build Coastguard Worker vp->old_metrics->c[starting_state & 63] = 0; /* Bias known start state */
69*638691a0SAndroid Build Coastguard Worker return 0;
70*638691a0SAndroid Build Coastguard Worker }
71*638691a0SAndroid Build Coastguard Worker
72*638691a0SAndroid Build Coastguard Worker /* Viterbi chainback */
chainback_viterbi27_sse(void * p,unsigned char * data,unsigned int nbits,unsigned int endstate)73*638691a0SAndroid Build Coastguard Worker int chainback_viterbi27_sse(
74*638691a0SAndroid Build Coastguard Worker void *p,
75*638691a0SAndroid Build Coastguard Worker unsigned char *data, /* Decoded output data */
76*638691a0SAndroid Build Coastguard Worker unsigned int nbits, /* Number of data bits */
77*638691a0SAndroid Build Coastguard Worker unsigned int endstate){ /* Terminal encoder state */
78*638691a0SAndroid Build Coastguard Worker struct v27 *vp = p;
79*638691a0SAndroid Build Coastguard Worker decision_t *d;
80*638691a0SAndroid Build Coastguard Worker
81*638691a0SAndroid Build Coastguard Worker if(p == NULL)
82*638691a0SAndroid Build Coastguard Worker return -1;
83*638691a0SAndroid Build Coastguard Worker
84*638691a0SAndroid Build Coastguard Worker d = vp->decisions;
85*638691a0SAndroid Build Coastguard Worker /* Make room beyond the end of the encoder register so we can
86*638691a0SAndroid Build Coastguard Worker * accumulate a full byte of decoded data
87*638691a0SAndroid Build Coastguard Worker */
88*638691a0SAndroid Build Coastguard Worker endstate %= 64;
89*638691a0SAndroid Build Coastguard Worker endstate <<= 2;
90*638691a0SAndroid Build Coastguard Worker
91*638691a0SAndroid Build Coastguard Worker /* The store into data[] only needs to be done every 8 bits.
92*638691a0SAndroid Build Coastguard Worker * But this avoids a conditional branch, and the writes will
93*638691a0SAndroid Build Coastguard Worker * combine in the cache anyway
94*638691a0SAndroid Build Coastguard Worker */
95*638691a0SAndroid Build Coastguard Worker d += 6; /* Look past tail */
96*638691a0SAndroid Build Coastguard Worker while(nbits-- != 0){
97*638691a0SAndroid Build Coastguard Worker int k;
98*638691a0SAndroid Build Coastguard Worker
99*638691a0SAndroid Build Coastguard Worker k = (d[nbits].c[(endstate>>2)/8] >> ((endstate>>2)%8)) & 1;
100*638691a0SAndroid Build Coastguard Worker data[nbits>>3] = endstate = (endstate >> 1) | (k << 7);
101*638691a0SAndroid Build Coastguard Worker }
102*638691a0SAndroid Build Coastguard Worker return 0;
103*638691a0SAndroid Build Coastguard Worker }
104*638691a0SAndroid Build Coastguard Worker
105*638691a0SAndroid Build Coastguard Worker /* Delete instance of a Viterbi decoder */
delete_viterbi27_sse(void * p)106*638691a0SAndroid Build Coastguard Worker void delete_viterbi27_sse(void *p){
107*638691a0SAndroid Build Coastguard Worker struct v27 *vp = p;
108*638691a0SAndroid Build Coastguard Worker
109*638691a0SAndroid Build Coastguard Worker if(vp != NULL){
110*638691a0SAndroid Build Coastguard Worker free(vp->decisions);
111*638691a0SAndroid Build Coastguard Worker free(vp);
112*638691a0SAndroid Build Coastguard Worker }
113*638691a0SAndroid Build Coastguard Worker }
114