1*638691a0SAndroid Build Coastguard Worker /* K=9 r=1/3 Viterbi decoder for x86 SSE
2*638691a0SAndroid Build Coastguard Worker * Copyright Aug 2006, 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 <xmmintrin.h>
6*638691a0SAndroid Build Coastguard Worker #include <stdio.h>
7*638691a0SAndroid Build Coastguard Worker #include <stdlib.h>
8*638691a0SAndroid Build Coastguard Worker #include <memory.h>
9*638691a0SAndroid Build Coastguard Worker #include <limits.h>
10*638691a0SAndroid Build Coastguard Worker #include "fec.h"
11*638691a0SAndroid Build Coastguard Worker
12*638691a0SAndroid Build Coastguard Worker typedef union { unsigned long w[8]; unsigned char c[32];} decision_t;
13*638691a0SAndroid Build Coastguard Worker typedef union { signed short s[256]; __m64 v[64];} metric_t;
14*638691a0SAndroid Build Coastguard Worker
15*638691a0SAndroid Build Coastguard Worker static union branchtab39 { unsigned short s[128]; __m64 v[32];} Branchtab39[3];
16*638691a0SAndroid Build Coastguard Worker static int Init = 0;
17*638691a0SAndroid Build Coastguard Worker
18*638691a0SAndroid Build Coastguard Worker /* State info for instance of Viterbi decoder */
19*638691a0SAndroid Build Coastguard Worker struct v39 {
20*638691a0SAndroid Build Coastguard Worker metric_t metrics1; /* path metric buffer 1 */
21*638691a0SAndroid Build Coastguard Worker metric_t metrics2; /* path metric buffer 2 */
22*638691a0SAndroid Build Coastguard Worker void *dp; /* Pointer to current decision */
23*638691a0SAndroid Build Coastguard Worker metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */
24*638691a0SAndroid Build Coastguard Worker void *decisions; /* Beginning of decisions for block */
25*638691a0SAndroid Build Coastguard Worker };
26*638691a0SAndroid Build Coastguard Worker
27*638691a0SAndroid Build Coastguard Worker /* Initialize Viterbi decoder for start of new frame */
init_viterbi39_sse(void * p,int starting_state)28*638691a0SAndroid Build Coastguard Worker int init_viterbi39_sse(void *p,int starting_state){
29*638691a0SAndroid Build Coastguard Worker struct v39 *vp = p;
30*638691a0SAndroid Build Coastguard Worker int i;
31*638691a0SAndroid Build Coastguard Worker
32*638691a0SAndroid Build Coastguard Worker if(p == NULL)
33*638691a0SAndroid Build Coastguard Worker return -1;
34*638691a0SAndroid Build Coastguard Worker for(i=0;i<256;i++)
35*638691a0SAndroid Build Coastguard Worker vp->metrics1.s[i] = (SHRT_MIN+1000);
36*638691a0SAndroid Build Coastguard Worker
37*638691a0SAndroid Build Coastguard Worker vp->old_metrics = &vp->metrics1;
38*638691a0SAndroid Build Coastguard Worker vp->new_metrics = &vp->metrics2;
39*638691a0SAndroid Build Coastguard Worker vp->dp = vp->decisions;
40*638691a0SAndroid Build Coastguard Worker vp->old_metrics->s[starting_state & 255] = SHRT_MIN; /* Bias known start state */
41*638691a0SAndroid Build Coastguard Worker return 0;
42*638691a0SAndroid Build Coastguard Worker }
43*638691a0SAndroid Build Coastguard Worker
44*638691a0SAndroid Build Coastguard Worker /* Create a new instance of a Viterbi decoder */
create_viterbi39_sse(int len)45*638691a0SAndroid Build Coastguard Worker void *create_viterbi39_sse(int len){
46*638691a0SAndroid Build Coastguard Worker struct v39 *vp;
47*638691a0SAndroid Build Coastguard Worker
48*638691a0SAndroid Build Coastguard Worker if(!Init){
49*638691a0SAndroid Build Coastguard Worker int polys[3] = { V39POLYA, V39POLYB, V39POLYC };
50*638691a0SAndroid Build Coastguard Worker
51*638691a0SAndroid Build Coastguard Worker set_viterbi39_polynomial_sse(polys);
52*638691a0SAndroid Build Coastguard Worker }
53*638691a0SAndroid Build Coastguard Worker if((vp = (struct v39 *)malloc(sizeof(struct v39))) == NULL){
54*638691a0SAndroid Build Coastguard Worker return NULL;
55*638691a0SAndroid Build Coastguard Worker }
56*638691a0SAndroid Build Coastguard Worker if((vp->decisions = malloc((len+8)*sizeof(decision_t))) == NULL){
57*638691a0SAndroid Build Coastguard Worker free(vp);
58*638691a0SAndroid Build Coastguard Worker return NULL;
59*638691a0SAndroid Build Coastguard Worker }
60*638691a0SAndroid Build Coastguard Worker init_viterbi39_sse(vp,0);
61*638691a0SAndroid Build Coastguard Worker return vp;
62*638691a0SAndroid Build Coastguard Worker }
63*638691a0SAndroid Build Coastguard Worker
set_viterbi39_polynomial_sse(int polys[3])64*638691a0SAndroid Build Coastguard Worker void set_viterbi39_polynomial_sse(int polys[3]){
65*638691a0SAndroid Build Coastguard Worker int state;
66*638691a0SAndroid Build Coastguard Worker
67*638691a0SAndroid Build Coastguard Worker for(state=0;state < 128;state++){
68*638691a0SAndroid Build Coastguard Worker Branchtab39[0].s[state] = (polys[0] < 0) ^ parity((2*state) & polys[0]) ? 255:0;
69*638691a0SAndroid Build Coastguard Worker Branchtab39[1].s[state] = (polys[1] < 0) ^ parity((2*state) & polys[1]) ? 255:0;
70*638691a0SAndroid Build Coastguard Worker Branchtab39[2].s[state] = (polys[2] < 0) ^ parity((2*state) & polys[2]) ? 255:0;
71*638691a0SAndroid Build Coastguard Worker }
72*638691a0SAndroid Build Coastguard Worker Init++;
73*638691a0SAndroid Build Coastguard Worker }
74*638691a0SAndroid Build Coastguard Worker
75*638691a0SAndroid Build Coastguard Worker /* Viterbi chainback */
chainback_viterbi39_sse(void * p,unsigned char * data,unsigned int nbits,unsigned int endstate)76*638691a0SAndroid Build Coastguard Worker int chainback_viterbi39_sse(
77*638691a0SAndroid Build Coastguard Worker void *p,
78*638691a0SAndroid Build Coastguard Worker unsigned char *data, /* Decoded output data */
79*638691a0SAndroid Build Coastguard Worker unsigned int nbits, /* Number of data bits */
80*638691a0SAndroid Build Coastguard Worker unsigned int endstate){ /* Terminal encoder state */
81*638691a0SAndroid Build Coastguard Worker struct v39 *vp = p;
82*638691a0SAndroid Build Coastguard Worker decision_t *d;
83*638691a0SAndroid Build Coastguard Worker int path_metric;
84*638691a0SAndroid Build Coastguard Worker
85*638691a0SAndroid Build Coastguard Worker if(p == NULL)
86*638691a0SAndroid Build Coastguard Worker return -1;
87*638691a0SAndroid Build Coastguard Worker d = (decision_t *)vp->decisions;
88*638691a0SAndroid Build Coastguard Worker endstate %= 256;
89*638691a0SAndroid Build Coastguard Worker
90*638691a0SAndroid Build Coastguard Worker path_metric = vp->old_metrics->s[endstate];
91*638691a0SAndroid Build Coastguard Worker
92*638691a0SAndroid Build Coastguard Worker /* The store into data[] only needs to be done every 8 bits.
93*638691a0SAndroid Build Coastguard Worker * But this avoids a conditional branch, and the writes will
94*638691a0SAndroid Build Coastguard Worker * combine in the cache anyway
95*638691a0SAndroid Build Coastguard Worker */
96*638691a0SAndroid Build Coastguard Worker d += 8; /* Look past tail */
97*638691a0SAndroid Build Coastguard Worker while(nbits-- != 0){
98*638691a0SAndroid Build Coastguard Worker int k;
99*638691a0SAndroid Build Coastguard Worker
100*638691a0SAndroid Build Coastguard Worker /* k = (d[nbits].w[endstate/32] >> (endstate%32)) & 1;*/
101*638691a0SAndroid Build Coastguard Worker k = (d[nbits].c[endstate/8] >> (endstate%8)) & 1;
102*638691a0SAndroid Build Coastguard Worker endstate = (k << 7) | (endstate >> 1);
103*638691a0SAndroid Build Coastguard Worker data[nbits>>3] = endstate;
104*638691a0SAndroid Build Coastguard Worker }
105*638691a0SAndroid Build Coastguard Worker return path_metric - SHRT_MIN;
106*638691a0SAndroid Build Coastguard Worker }
107*638691a0SAndroid Build Coastguard Worker
108*638691a0SAndroid Build Coastguard Worker /* Delete instance of a Viterbi decoder */
delete_viterbi39_sse(void * p)109*638691a0SAndroid Build Coastguard Worker void delete_viterbi39_sse(void *p){
110*638691a0SAndroid Build Coastguard Worker struct v39 *vp = p;
111*638691a0SAndroid Build Coastguard Worker
112*638691a0SAndroid Build Coastguard Worker if(vp != NULL){
113*638691a0SAndroid Build Coastguard Worker free(vp->decisions);
114*638691a0SAndroid Build Coastguard Worker free(vp);
115*638691a0SAndroid Build Coastguard Worker }
116*638691a0SAndroid Build Coastguard Worker }
117*638691a0SAndroid Build Coastguard Worker
118*638691a0SAndroid Build Coastguard Worker
update_viterbi39_blk_sse(void * p,unsigned char * syms,int nbits)119*638691a0SAndroid Build Coastguard Worker int update_viterbi39_blk_sse(void *p,unsigned char *syms,int nbits){
120*638691a0SAndroid Build Coastguard Worker struct v39 *vp = p;
121*638691a0SAndroid Build Coastguard Worker decision_t *d;
122*638691a0SAndroid Build Coastguard Worker int path_metric = 0;
123*638691a0SAndroid Build Coastguard Worker
124*638691a0SAndroid Build Coastguard Worker if(p == NULL)
125*638691a0SAndroid Build Coastguard Worker return -1;
126*638691a0SAndroid Build Coastguard Worker d = (decision_t *)vp->dp;
127*638691a0SAndroid Build Coastguard Worker while(nbits--){
128*638691a0SAndroid Build Coastguard Worker __m64 sym0v,sym1v,sym2v;
129*638691a0SAndroid Build Coastguard Worker void *tmp;
130*638691a0SAndroid Build Coastguard Worker int i;
131*638691a0SAndroid Build Coastguard Worker
132*638691a0SAndroid Build Coastguard Worker /* Splat the 0th symbol across sym0v, the 1st symbol across sym1v, etc */
133*638691a0SAndroid Build Coastguard Worker sym0v = _mm_set1_pi16(syms[0]);
134*638691a0SAndroid Build Coastguard Worker sym1v = _mm_set1_pi16(syms[1]);
135*638691a0SAndroid Build Coastguard Worker sym2v = _mm_set1_pi16(syms[2]);
136*638691a0SAndroid Build Coastguard Worker syms += 3;
137*638691a0SAndroid Build Coastguard Worker
138*638691a0SAndroid Build Coastguard Worker for(i=0;i<32;i++){
139*638691a0SAndroid Build Coastguard Worker __m64 decision0,decision1,metric,m_metric,m0,m1,m2,m3,survivor0,survivor1;
140*638691a0SAndroid Build Coastguard Worker
141*638691a0SAndroid Build Coastguard Worker /* Form branch metrics
142*638691a0SAndroid Build Coastguard Worker * Because Branchtab takes on values 0 and 255, and the values of sym?v are offset binary in the range 0-255,
143*638691a0SAndroid Build Coastguard Worker * the XOR operations constitute conditional negation.
144*638691a0SAndroid Build Coastguard Worker * metric and m_metric (-metric) are in the range 0-765
145*638691a0SAndroid Build Coastguard Worker */
146*638691a0SAndroid Build Coastguard Worker m0 = _mm_add_pi16(_mm_xor_si64(Branchtab39[0].v[i],sym0v),_mm_xor_si64(Branchtab39[1].v[i],sym1v));
147*638691a0SAndroid Build Coastguard Worker metric = _mm_add_pi16(_mm_xor_si64(Branchtab39[2].v[i],sym2v),m0);
148*638691a0SAndroid Build Coastguard Worker m_metric = _mm_sub_pi16(_mm_set1_pi16(765),metric);
149*638691a0SAndroid Build Coastguard Worker
150*638691a0SAndroid Build Coastguard Worker /* Add branch metrics to path metrics */
151*638691a0SAndroid Build Coastguard Worker m0 = _mm_adds_pi16(vp->old_metrics->v[i],metric);
152*638691a0SAndroid Build Coastguard Worker m3 = _mm_adds_pi16(vp->old_metrics->v[32+i],metric);
153*638691a0SAndroid Build Coastguard Worker m1 = _mm_adds_pi16(vp->old_metrics->v[32+i],m_metric);
154*638691a0SAndroid Build Coastguard Worker m2 = _mm_adds_pi16(vp->old_metrics->v[i],m_metric);
155*638691a0SAndroid Build Coastguard Worker
156*638691a0SAndroid Build Coastguard Worker /* Compare and select */
157*638691a0SAndroid Build Coastguard Worker survivor0 = _mm_min_pi16(m0,m1);
158*638691a0SAndroid Build Coastguard Worker survivor1 = _mm_min_pi16(m2,m3);
159*638691a0SAndroid Build Coastguard Worker decision0 = _mm_cmpeq_pi16(survivor0,m1);
160*638691a0SAndroid Build Coastguard Worker decision1 = _mm_cmpeq_pi16(survivor1,m3);
161*638691a0SAndroid Build Coastguard Worker
162*638691a0SAndroid Build Coastguard Worker /* Pack decisions into 8 bits and store */
163*638691a0SAndroid Build Coastguard Worker d->c[i] = _mm_movemask_pi8(_mm_unpacklo_pi8(_mm_packs_pi16(decision0,_mm_setzero_si64()),_mm_packs_pi16(decision1,_mm_setzero_si64())));
164*638691a0SAndroid Build Coastguard Worker
165*638691a0SAndroid Build Coastguard Worker /* Store surviving metrics */
166*638691a0SAndroid Build Coastguard Worker vp->new_metrics->v[2*i] = _mm_unpacklo_pi16(survivor0,survivor1);
167*638691a0SAndroid Build Coastguard Worker vp->new_metrics->v[2*i+1] = _mm_unpackhi_pi16(survivor0,survivor1);
168*638691a0SAndroid Build Coastguard Worker }
169*638691a0SAndroid Build Coastguard Worker /* See if we need to renormalize
170*638691a0SAndroid Build Coastguard Worker * Max metric spread for this code with 0-255 branch metrics is 12750
171*638691a0SAndroid Build Coastguard Worker */
172*638691a0SAndroid Build Coastguard Worker if(vp->new_metrics->s[0] >= SHRT_MAX-5000){
173*638691a0SAndroid Build Coastguard Worker int i,adjust;
174*638691a0SAndroid Build Coastguard Worker __m64 adjustv;
175*638691a0SAndroid Build Coastguard Worker union { __m64 v; signed short w[4]; } t;
176*638691a0SAndroid Build Coastguard Worker
177*638691a0SAndroid Build Coastguard Worker /* Find smallest metric and set adjustv to bring it down to SHRT_MIN */
178*638691a0SAndroid Build Coastguard Worker adjustv = vp->new_metrics->v[0];
179*638691a0SAndroid Build Coastguard Worker for(i=1;i<64;i++)
180*638691a0SAndroid Build Coastguard Worker adjustv = _mm_min_pi16(adjustv,vp->new_metrics->v[i]);
181*638691a0SAndroid Build Coastguard Worker
182*638691a0SAndroid Build Coastguard Worker adjustv = _mm_min_pi16(adjustv,_mm_srli_si64(adjustv,32));
183*638691a0SAndroid Build Coastguard Worker adjustv = _mm_min_pi16(adjustv,_mm_srli_si64(adjustv,16));
184*638691a0SAndroid Build Coastguard Worker t.v = adjustv;
185*638691a0SAndroid Build Coastguard Worker adjust = t.w[0] - SHRT_MIN;
186*638691a0SAndroid Build Coastguard Worker path_metric += adjust;
187*638691a0SAndroid Build Coastguard Worker adjustv = _mm_set1_pi16(adjust);
188*638691a0SAndroid Build Coastguard Worker
189*638691a0SAndroid Build Coastguard Worker for(i=0;i<64;i++)
190*638691a0SAndroid Build Coastguard Worker vp->new_metrics->v[i] = _mm_sub_pi16(vp->new_metrics->v[i],adjustv);
191*638691a0SAndroid Build Coastguard Worker }
192*638691a0SAndroid Build Coastguard Worker d++;
193*638691a0SAndroid Build Coastguard Worker /* Swap pointers to old and new metrics */
194*638691a0SAndroid Build Coastguard Worker tmp = vp->old_metrics;
195*638691a0SAndroid Build Coastguard Worker vp->old_metrics = vp->new_metrics;
196*638691a0SAndroid Build Coastguard Worker vp->new_metrics = tmp;
197*638691a0SAndroid Build Coastguard Worker }
198*638691a0SAndroid Build Coastguard Worker vp->dp = d;
199*638691a0SAndroid Build Coastguard Worker _mm_empty();
200*638691a0SAndroid Build Coastguard Worker return path_metric;
201*638691a0SAndroid Build Coastguard Worker }
202