1 /* K=9 r=1/3 Viterbi decoder for x86 MMX 2 * Aug 2006, Phil Karn, KA9Q 3 * May be used under the terms of the GNU Lesser General Public License (LGPL) 4 */ 5 #include <mmintrin.h> 6 #include <stdio.h> 7 #include <stdlib.h> 8 #include <memory.h> 9 #include "fec.h" 10 11 typedef union { unsigned char c[256]; __m64 v[32];} decision_t; 12 typedef union { unsigned short s[256]; __m64 v[64];} metric_t; 13 14 static union branchtab39 { unsigned short s[128]; __m64 v[32];} Branchtab39[3]; 15 static int Init = 0; 16 17 /* State info for instance of Viterbi decoder */ 18 struct v39 { 19 metric_t metrics1; /* path metric buffer 1 */ 20 metric_t metrics2; /* path metric buffer 2 */ 21 void *dp; /* Pointer to current decision */ 22 metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */ 23 void *decisions; /* Beginning of decisions for block */ 24 }; 25 26 /* Initialize Viterbi decoder for start of new frame */ 27 int init_viterbi39_mmx(void *p,int starting_state){ 28 struct v39 *vp = p; 29 int i; 30 31 if(p == NULL) 32 return -1; 33 for(i=0;i<256;i++) 34 vp->metrics1.s[i] = 1000; 35 36 vp->old_metrics = &vp->metrics1; 37 vp->new_metrics = &vp->metrics2; 38 vp->dp = vp->decisions; 39 vp->old_metrics->s[starting_state & 255] = 0; /* Bias known start state */ 40 return 0; 41 } 42 43 void set_viterbi39_polynomial_mmx(int polys[3]){ 44 int state; 45 46 for(state=0;state < 128;state++){ 47 Branchtab39[0].s[state] = (polys[0] < 0) ^ parity((2*state) & polys[0]) ? 255:0; 48 Branchtab39[1].s[state] = (polys[1] < 0) ^ parity((2*state) & polys[1]) ? 255:0; 49 Branchtab39[2].s[state] = (polys[2] < 0) ^ parity((2*state) & polys[2]) ? 255:0; 50 } 51 Init++; 52 } 53 54 /* Create a new instance of a Viterbi decoder */ 55 void *create_viterbi39_mmx(int len){ 56 struct v39 *vp; 57 58 if(!Init){ 59 int polys[3] = { V39POLYA,V39POLYB,V39POLYC }; 60 set_viterbi39_polynomial_mmx(polys); 61 } 62 if((vp = (struct v39 *)malloc(sizeof(struct v39))) == NULL) 63 return NULL; 64 if((vp->decisions = malloc((len+8)*sizeof(decision_t))) == NULL){ 65 free(vp); 66 return NULL; 67 } 68 init_viterbi39_mmx(vp,0); 69 return vp; 70 } 71 72 73 74 /* Viterbi chainback */ 75 int chainback_viterbi39_mmx( 76 void *p, 77 unsigned char *data, /* Decoded output data */ 78 unsigned int nbits, /* Number of data bits */ 79 unsigned int endstate){ /* Terminal encoder state */ 80 struct v39 *vp = p; 81 decision_t *d; 82 int path_metric; 83 84 if(p == NULL) 85 return -1; 86 87 d = (decision_t *)vp->decisions; 88 89 endstate %= 256; 90 91 path_metric = vp->old_metrics->s[endstate]; 92 93 /* The store into data[] only needs to be done every 8 bits. 94 * But this avoids a conditional branch, and the writes will 95 * combine in the cache anyway 96 */ 97 d += 8; /* Look past tail */ 98 while(nbits-- != 0){ 99 int k; 100 101 k = d[nbits].c[endstate] & 1; 102 endstate = (k << 7) | (endstate >> 1); 103 data[nbits>>3] = endstate; 104 } 105 return path_metric; 106 } 107 108 /* Delete instance of a Viterbi decoder */ 109 void delete_viterbi39_mmx(void *p){ 110 struct v39 *vp = p; 111 112 if(vp != NULL){ 113 free(vp->decisions); 114 free(vp); 115 } 116 } 117 118 119 int update_viterbi39_blk_mmx(void *p,unsigned char *syms,int nbits){ 120 struct v39 *vp = p; 121 decision_t *d; 122 int path_metric = 0; 123 124 if(p == NULL) 125 return -1; 126 127 d = (decision_t *)vp->dp; 128 129 while(nbits--){ 130 __m64 sym0v,sym1v,sym2v; 131 void *tmp; 132 int i; 133 134 /* Splat the 0th symbol across sym0v, the 1st symbol across sym1v, etc */ 135 sym0v = _mm_set1_pi16(syms[0]); 136 sym1v = _mm_set1_pi16(syms[1]); 137 sym2v = _mm_set1_pi16(syms[2]); 138 syms += 3; 139 140 for(i=0;i<32;i++){ 141 __m64 decision0,decision1,metric,m_metric,m0,m1,m2,m3,survivor0,survivor1; 142 143 /* Form branch metrics 144 * Because Branchtab takes on values 0 and 255, and the values of sym?v are offset binary in the range 0-255, 145 * the XOR operations constitute conditional negation. 146 * metric and m_metric (-metric) are in the range 0-1530 147 */ 148 m0 = _mm_add_pi16(_mm_xor_si64(Branchtab39[0].v[i],sym0v),_mm_xor_si64(Branchtab39[1].v[i],sym1v)); 149 metric = _mm_add_pi16(_mm_xor_si64(Branchtab39[2].v[i],sym2v),m0); 150 m_metric = _mm_sub_pi16(_mm_set1_pi16(765),metric); 151 152 /* Add branch metrics to path metrics */ 153 m0 = _mm_add_pi16(vp->old_metrics->v[i],metric); 154 m3 = _mm_add_pi16(vp->old_metrics->v[32+i],metric); 155 m1 = _mm_add_pi16(vp->old_metrics->v[32+i],m_metric); 156 m2 = _mm_add_pi16(vp->old_metrics->v[i],m_metric); 157 158 /* Compare and select 159 * There's no packed min instruction in MMX, so we use modulo arithmetic 160 * to form the decisions and then do the select the hard way 161 */ 162 decision0 = _mm_cmpgt_pi16(_mm_sub_pi16(m0,m1),_mm_setzero_si64()); 163 decision1 = _mm_cmpgt_pi16(_mm_sub_pi16(m2,m3),_mm_setzero_si64()); 164 survivor0 = _mm_or_si64(_mm_and_si64(decision0,m1),_mm_andnot_si64(decision0,m0)); 165 survivor1 = _mm_or_si64(_mm_and_si64(decision1,m3),_mm_andnot_si64(decision1,m2)); 166 167 /* Merge decisions and store as bytes */ 168 d->v[i] = _mm_unpacklo_pi8(_mm_packs_pi16(decision0,_mm_setzero_si64()),_mm_packs_pi16(decision1,_mm_setzero_si64())); 169 170 /* Store surviving metrics */ 171 vp->new_metrics->v[2*i] = _mm_unpacklo_pi16(survivor0,survivor1); 172 vp->new_metrics->v[2*i+1] = _mm_unpackhi_pi16(survivor0,survivor1); 173 } 174 if(vp->new_metrics->s[0] < vp->old_metrics->s[0]) 175 path_metric += 65536; /* Hack: wraparound probably occured */ 176 d++; 177 /* Swap pointers to old and new metrics */ 178 tmp = vp->old_metrics; 179 vp->old_metrics = vp->new_metrics; 180 vp->new_metrics = tmp; 181 } 182 vp->dp = d; 183 _mm_empty(); 184 return path_metric; 185 } 186