1 /* K=7 r=1/2 Viterbi decoder for PowerPC G4/G5 Altivec instructions
2 * Feb 2004, Phil Karn, KA9Q
3 */
4 #include <altivec.h>
5 #include <stdio.h>
6 #include <memory.h>
7 #include <stdlib.h>
8 #include "fec.h"
9
10 typedef union { long long p; unsigned char c[64]; vector bool char v[4]; } decision_t;
11 typedef union { long long p; unsigned char c[64]; vector unsigned char v[4]; } metric_t;
12
13 static union branchtab27 { unsigned char c[32]; vector unsigned char v[2];} Branchtab27[2];
14 static int Init = 0;
15
16 /* State info for instance of Viterbi decoder
17 * Don't change this without also changing references in [mmx|sse|sse2]bfly29.s!
18 */
19 struct v27 {
20 metric_t metrics1; /* path metric buffer 1 */
21 metric_t metrics2; /* path metric buffer 2 */
22 decision_t *dp; /* Pointer to current decision */
23 metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */
24 decision_t *decisions; /* Beginning of decisions for block */
25 };
26
27 /* Initialize Viterbi decoder for start of new frame */
init_viterbi27_av(void * p,int starting_state)28 int init_viterbi27_av(void *p,int starting_state){
29 struct v27 *vp = p;
30 int i;
31
32 if(p == NULL)
33 return -1;
34 for(i=0;i<4;i++)
35 vp->metrics1.v[i] = (vector unsigned char){63};
36 vp->old_metrics = &vp->metrics1;
37 vp->new_metrics = &vp->metrics2;
38 vp->dp = vp->decisions;
39 vp->old_metrics->c[starting_state & 63] = 0; /* Bias known start state */
40 return 0;
41 }
42
set_viterbi27_polynomial_av(int polys[2])43 void set_viterbi27_polynomial_av(int polys[2]){
44 int state;
45
46 for(state=0;state < 32;state++){
47 Branchtab27[0].c[state] = (polys[0] < 0) ^ parity((2*state) & abs(polys[0])) ? 255 : 0;
48 Branchtab27[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_viterbi27_av(int len)54 void *create_viterbi27_av(int len){
55 struct v27 *vp;
56
57 if(!Init){
58 int polys[2] = { V27POLYA,V27POLYB };
59 set_viterbi27_polynomial_av(polys);
60 }
61 if((vp = (struct v27 *)malloc(sizeof(struct v27))) == NULL)
62 return NULL;
63 if((vp->decisions = (decision_t *)malloc((len+6)*sizeof(decision_t))) == NULL){
64 free(vp);
65 return NULL;
66 }
67 init_viterbi27_av(vp,0);
68 return vp;
69 }
70
71 /* Viterbi chainback */
chainback_viterbi27_av(void * p,unsigned char * data,unsigned int nbits,unsigned int endstate)72 int chainback_viterbi27_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 v27 *vp = p;
78 decision_t *d = (decision_t *)vp->decisions;
79
80 if(p == NULL)
81 return -1;
82
83 /* Make room beyond the end of the encoder register so we can
84 * accumulate a full byte of decoded data
85 */
86 endstate %= 64;
87 endstate <<= 2;
88
89 /* The store into data[] only needs to be done every 8 bits.
90 * But this avoids a conditional branch, and the writes will
91 * combine in the cache anyway
92 */
93 d += 6; /* Look past tail */
94 while(nbits-- != 0){
95 int k;
96
97 k = d[nbits].c[endstate>>2] & 1;
98 data[nbits>>3] = endstate = (endstate >> 1) | (k << 7);
99 }
100 return 0;
101 }
102
103 /* Delete instance of a Viterbi decoder */
delete_viterbi27_av(void * p)104 void delete_viterbi27_av(void *p){
105 struct v27 *vp = p;
106
107 if(vp != NULL){
108 free(vp->decisions);
109 free(vp);
110 }
111 }
112
113 /* Process received symbols */
update_viterbi27_blk_av(void * p,unsigned char * syms,int nbits)114 int update_viterbi27_blk_av(void *p,unsigned char *syms,int nbits){
115 struct v27 *vp = p;
116 decision_t *d;
117
118 if(p == NULL)
119 return -1;
120 d = (decision_t *)vp->dp;
121 while(nbits--){
122 vector unsigned char survivor0,survivor1,sym0v,sym1v;
123 vector bool char decision0,decision1;
124 vector unsigned char metric,m_metric,m0,m1,m2,m3;
125 void *tmp;
126
127 /* sym0v.0 = syms[0]; sym0v.1 = syms[1] */
128 sym0v = vec_perm(vec_ld(0,syms),vec_ld(1,syms),vec_lvsl(0,syms));
129
130 sym1v = vec_splat(sym0v,1); /* Splat syms[1] across sym1v */
131 sym0v = vec_splat(sym0v,0); /* Splat syms[0] across sym0v */
132 syms += 2;
133
134 /* Do the 32 butterflies as two interleaved groups of 16 each to keep the pipes full */
135
136 /* Form first set of 16 branch metrics */
137 metric = vec_avg(vec_xor(Branchtab27[0].v[0],sym0v),vec_xor(Branchtab27[1].v[0],sym1v));
138 metric = vec_sr(metric,(vector unsigned char){3});
139 m_metric = vec_sub((vector unsigned char){31},metric);
140
141 /* Form first set of path metrics */
142 m0 = vec_adds(vp->old_metrics->v[0],metric);
143 m3 = vec_adds(vp->old_metrics->v[2],metric);
144 m1 = vec_adds(vp->old_metrics->v[2],m_metric);
145 m2 = vec_adds(vp->old_metrics->v[0],m_metric);
146
147 /* Form second set of 16 branch metrics */
148 metric = vec_avg(vec_xor(Branchtab27[0].v[1],sym0v),vec_xor(Branchtab27[1].v[1],sym1v));
149 metric = vec_sr(metric,(vector unsigned char){3});
150 m_metric = vec_sub((vector unsigned char){31},metric);
151
152 /* Compare and select first set */
153 decision0 = vec_cmpgt(m0,m1);
154 decision1 = vec_cmpgt(m2,m3);
155 survivor0 = vec_min(m0,m1);
156 survivor1 = vec_min(m2,m3);
157
158 /* Compute second set of path metrics */
159 m0 = vec_adds(vp->old_metrics->v[1],metric);
160 m3 = vec_adds(vp->old_metrics->v[3],metric);
161 m1 = vec_adds(vp->old_metrics->v[3],m_metric);
162 m2 = vec_adds(vp->old_metrics->v[1],m_metric);
163
164 /* Interleave and store first decisions and survivors */
165 d->v[0] = vec_mergeh(decision0,decision1);
166 d->v[1] = vec_mergel(decision0,decision1);
167 vp->new_metrics->v[0] = vec_mergeh(survivor0,survivor1);
168 vp->new_metrics->v[1] = vec_mergel(survivor0,survivor1);
169
170 /* Compare and select second set */
171 decision0 = vec_cmpgt(m0,m1);
172 decision1 = vec_cmpgt(m2,m3);
173 survivor0 = vec_min(m0,m1);
174 survivor1 = vec_min(m2,m3);
175
176 /* Interleave and store second set of decisions and survivors */
177 d->v[2] = vec_mergeh(decision0,decision1);
178 d->v[3] = vec_mergel(decision0,decision1);
179 vp->new_metrics->v[2] = vec_mergeh(survivor0,survivor1);
180 vp->new_metrics->v[3] = vec_mergel(survivor0,survivor1);
181
182 /* renormalize if necessary */
183 if(vp->new_metrics->c[0] >= 105){
184 vector unsigned char scale0,scale1;
185
186 /* Find smallest metric and splat */
187 scale0 = vec_min(vp->new_metrics->v[0],vp->new_metrics->v[1]);
188 scale1 = vec_min(vp->new_metrics->v[2],vp->new_metrics->v[3]);
189 scale0 = vec_min(scale0,scale1);
190 scale0 = vec_min(scale0,vec_sld(scale0,scale0,8));
191 scale0 = vec_min(scale0,vec_sld(scale0,scale0,4));
192 scale0 = vec_min(scale0,vec_sld(scale0,scale0,2));
193 scale0 = vec_min(scale0,vec_sld(scale0,scale0,1));
194
195 /* Now subtract from all metrics */
196 vp->new_metrics->v[0] = vec_subs(vp->new_metrics->v[0],scale0);
197 vp->new_metrics->v[1] = vec_subs(vp->new_metrics->v[1],scale0);
198 vp->new_metrics->v[2] = vec_subs(vp->new_metrics->v[2],scale0);
199 vp->new_metrics->v[3] = vec_subs(vp->new_metrics->v[3],scale0);
200 }
201 d++;
202 /* Swap pointers to old and new metrics */
203 tmp = vp->old_metrics;
204 vp->old_metrics = vp->new_metrics;
205 vp->new_metrics = tmp;
206 }
207 vp->dp = d;
208
209 return 0;
210 }
211
212