1 /* K=15 r=1/6 Viterbi decoder for x86 SSE2
2  * Copyright Mar 2004, Phil Karn, KA9Q
3  * May be used under the terms of the GNU Lesser General Public License (LGPL)
4  */
5 #include <emmintrin.h>
6 #include <stdio.h>
7 #include <stdlib.h>
8 #include <memory.h>
9 #include <limits.h>
10 #include "fec.h"
11 
12 typedef union { unsigned long w[512]; unsigned short s[1024];} decision_t;
13 typedef union { signed short s[16384]; __m128i v[2048];} metric_t;
14 
15 static union branchtab615 { unsigned short s[8192]; __m128i v[1024];} Branchtab615[6];
16 static int Init = 0;
17 
18 /* State info for instance of Viterbi decoder */
19 struct v615 {
20   metric_t metrics1; /* path metric buffer 1 */
21   metric_t metrics2; /* path metric buffer 2 */
22   void *dp;          /* Pointer to current decision */
23   metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */
24   void *decisions;   /* Beginning of decisions for block */
25 };
26 
27 /* Initialize Viterbi decoder for start of new frame */
init_viterbi615_sse2(void * p,int starting_state)28 int init_viterbi615_sse2(void *p,int starting_state){
29   struct v615 *vp = p;
30   int i;
31 
32   if(p == NULL)
33     return -1;
34   for(i=0;i<16384;i++)
35     vp->metrics1.s[i] = (SHRT_MIN+5000);
36 
37   vp->old_metrics = &vp->metrics1;
38   vp->new_metrics = &vp->metrics2;
39   vp->dp = vp->decisions;
40   vp->old_metrics->s[starting_state & 16383] = SHRT_MIN; /* Bias known start state */
41   return 0;
42 }
43 
44 /* Create a new instance of a Viterbi decoder */
create_viterbi615_sse2(int len)45 void *create_viterbi615_sse2(int len){
46   void *p;
47   struct v615 *vp;
48 
49   if(!Init){
50     int polys[6] = { V615POLYA,V615POLYB,V615POLYC,V615POLYD,V615POLYE,V615POLYF };
51     set_viterbi615_polynomial_sse2(polys);
52   }
53 
54   /* Ordinary malloc() only returns 8-byte alignment, we need 16 */
55   if(posix_memalign(&p, sizeof(__m128i),sizeof(struct v615)))
56     return NULL;
57 
58   vp = (struct v615 *)p;
59   if((p = malloc((len+14)*sizeof(decision_t))) == NULL){
60     free(vp);
61     return NULL;
62   }
63   vp->decisions = (decision_t *)p;
64   init_viterbi615_sse2(vp,0);
65   return vp;
66 }
67 
set_viterbi615_polynomial_sse2(int polys[6])68 void set_viterbi615_polynomial_sse2(int polys[6]){
69   int state;
70   int i;
71 
72   for(state=0;state < 8192;state++){
73     for(i=0;i<6;i++)
74       Branchtab615[i].s[state] = (polys[i] < 0) ^ parity((2*state) & abs(polys[i])) ? 255 : 0;
75   }
76   Init++;
77 }
78 
79 /* Viterbi chainback */
chainback_viterbi615_sse2(void * p,unsigned char * data,unsigned int nbits,unsigned int endstate)80 int chainback_viterbi615_sse2(
81       void *p,
82       unsigned char *data, /* Decoded output data */
83       unsigned int nbits, /* Number of data bits */
84       unsigned int endstate){ /* Terminal encoder state */
85   struct v615 *vp = p;
86   decision_t *d = (decision_t *)vp->decisions;
87 
88   endstate %= 16384;
89 
90   /* The store into data[] only needs to be done every 8 bits.
91    * But this avoids a conditional branch, and the writes will
92    * combine in the cache anyway
93    */
94   d += 14; /* Look past tail */
95   while(nbits-- != 0){
96     int k;
97 
98     k = (d[nbits].w[endstate/32] >> (endstate%32)) & 1;
99     endstate = (k << 13) | (endstate >> 1);
100     data[nbits>>3] = endstate >> 6;
101   }
102   return 0;
103 }
104 
105 /* Delete instance of a Viterbi decoder */
delete_viterbi615_sse2(void * p)106 void delete_viterbi615_sse2(void *p){
107   struct v615 *vp = p;
108 
109   if(vp != NULL){
110     free(vp->decisions);
111     free(vp);
112   }
113 }
114 
115 
update_viterbi615_blk_sse2(void * p,unsigned char * syms,int nbits)116 int update_viterbi615_blk_sse2(void *p,unsigned char *syms,int nbits){
117   struct v615 *vp = p;
118   decision_t *d = (decision_t *)vp->dp;
119 
120   while(nbits--){
121     __m128i sym0v,sym1v,sym2v,sym3v,sym4v,sym5v;
122     void *tmp;
123     int i;
124 
125     /* Splat the 0th symbol across sym0v, the 1st symbol across sym1v, etc */
126     sym0v = _mm_set1_epi16(syms[0]);
127     sym1v = _mm_set1_epi16(syms[1]);
128     sym2v = _mm_set1_epi16(syms[2]);
129     sym3v = _mm_set1_epi16(syms[3]);
130     sym4v = _mm_set1_epi16(syms[4]);
131     sym5v = _mm_set1_epi16(syms[5]);
132     syms += 6;
133 
134     /* SSE2 doesn't support saturated adds on unsigned shorts, so we have to use signed shorts */
135     for(i=0;i<1024;i++){
136       __m128i decision0,decision1,metric,m_metric,m0,m1,m2,m3,survivor0,survivor1;
137 
138       /* Form branch metrics
139        * Because Branchtab takes on values 0 and 255, and the values of sym?v are offset binary in the range 0-255,
140        * the XOR operations constitute conditional negation.
141        * metric and m_metric (-metric) are in the range 0-1530
142        */
143       m0 = _mm_add_epi16(_mm_xor_si128(Branchtab615[0].v[i],sym0v),_mm_xor_si128(Branchtab615[1].v[i],sym1v));
144       m1 = _mm_add_epi16(_mm_xor_si128(Branchtab615[2].v[i],sym2v),_mm_xor_si128(Branchtab615[3].v[i],sym3v));
145       m2 = _mm_add_epi16(_mm_xor_si128(Branchtab615[4].v[i],sym4v),_mm_xor_si128(Branchtab615[5].v[i],sym5v));
146       metric = _mm_add_epi16(m0,_mm_add_epi16(m1,m2));
147       m_metric = _mm_sub_epi16(_mm_set1_epi16(1530),metric);
148 
149       /* Add branch metrics to path metrics */
150       m0 = _mm_adds_epi16(vp->old_metrics->v[i],metric);
151       m3 = _mm_adds_epi16(vp->old_metrics->v[1024+i],metric);
152       m1 = _mm_adds_epi16(vp->old_metrics->v[1024+i],m_metric);
153       m2 = _mm_adds_epi16(vp->old_metrics->v[i],m_metric);
154 
155       /* Compare and select */
156       survivor0 = _mm_min_epi16(m0,m1);
157       survivor1 = _mm_min_epi16(m2,m3);
158       decision0 = _mm_cmpeq_epi16(survivor0,m1);
159       decision1 = _mm_cmpeq_epi16(survivor1,m3);
160 
161       /* Pack each set of decisions into 8 8-bit bytes, then interleave them and compress into 16 bits */
162       d->s[i] = _mm_movemask_epi8(_mm_unpacklo_epi8(_mm_packs_epi16(decision0,_mm_setzero_si128()),_mm_packs_epi16(decision1,_mm_setzero_si128())));
163 
164       /* Store surviving metrics */
165       vp->new_metrics->v[2*i] = _mm_unpacklo_epi16(survivor0,survivor1);
166       vp->new_metrics->v[2*i+1] = _mm_unpackhi_epi16(survivor0,survivor1);
167     }
168     /* See if we need to renormalize
169      * Max metric spread for this code with 0-90 branch metrics is 405
170      */
171     if(vp->new_metrics->s[0] >= SHRT_MAX-12750){
172       int i,adjust;
173       __m128i adjustv;
174       union { __m128i v; signed short w[8]; } t;
175 
176       /* Find smallest metric and set adjustv to bring it down to SHRT_MIN */
177       adjustv = vp->new_metrics->v[0];
178       for(i=1;i<2048;i++)
179 	adjustv = _mm_min_epi16(adjustv,vp->new_metrics->v[i]);
180 
181       adjustv = _mm_min_epi16(adjustv,_mm_srli_si128(adjustv,8));
182       adjustv = _mm_min_epi16(adjustv,_mm_srli_si128(adjustv,4));
183       adjustv = _mm_min_epi16(adjustv,_mm_srli_si128(adjustv,2));
184       t.v = adjustv;
185       adjust = t.w[0] - SHRT_MIN;
186       adjustv = _mm_set1_epi16(adjust);
187 
188       /* We cannot use a saturated subtract, because we often have to adjust by more than SHRT_MAX
189        * This is okay since it can't overflow anyway
190        */
191       for(i=0;i<2048;i++)
192 	vp->new_metrics->v[i] = _mm_sub_epi16(vp->new_metrics->v[i],adjustv);
193     }
194     d++;
195     /* Swap pointers to old and new metrics */
196     tmp = vp->old_metrics;
197     vp->old_metrics = vp->new_metrics;
198     vp->new_metrics = tmp;
199   }
200   vp->dp = d;
201   return 0;
202 }
203 
204 
205