Bill Yi | 4e213d5 | 2015-06-23 13:53:11 -0700 | [diff] [blame] | 1 | /* K=9 r=1/3 Viterbi decoder for x86 SSE |
| 2 | * Copyright Aug 2006, Phil Karn, KA9Q |
| 3 | * May be used under the terms of the GNU Lesser General Public License (LGPL) |
| 4 | */ |
| 5 | #include <xmmintrin.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[8]; unsigned char c[32];} decision_t; |
| 13 | typedef union { signed short s[256]; __m64 v[64];} metric_t; |
| 14 | |
| 15 | static union branchtab39 { unsigned short s[128]; __m64 v[32];} Branchtab39[3]; |
| 16 | static int Init = 0; |
| 17 | |
| 18 | /* State info for instance of Viterbi decoder */ |
| 19 | struct v39 { |
| 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 */ |
| 28 | int init_viterbi39_sse(void *p,int starting_state){ |
| 29 | struct v39 *vp = p; |
| 30 | int i; |
| 31 | |
| 32 | if(p == NULL) |
| 33 | return -1; |
| 34 | for(i=0;i<256;i++) |
| 35 | vp->metrics1.s[i] = (SHRT_MIN+1000); |
| 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 & 255] = SHRT_MIN; /* Bias known start state */ |
| 41 | return 0; |
| 42 | } |
| 43 | |
| 44 | /* Create a new instance of a Viterbi decoder */ |
| 45 | void *create_viterbi39_sse(int len){ |
| 46 | struct v39 *vp; |
| 47 | |
| 48 | if(!Init){ |
| 49 | int polys[3] = { V39POLYA, V39POLYB, V39POLYC }; |
| 50 | |
| 51 | set_viterbi39_polynomial_sse(polys); |
| 52 | } |
| 53 | if((vp = (struct v39 *)malloc(sizeof(struct v39))) == NULL){ |
| 54 | return NULL; |
| 55 | } |
| 56 | if((vp->decisions = malloc((len+8)*sizeof(decision_t))) == NULL){ |
| 57 | free(vp); |
| 58 | return NULL; |
| 59 | } |
| 60 | init_viterbi39_sse(vp,0); |
| 61 | return vp; |
| 62 | } |
| 63 | |
| 64 | void set_viterbi39_polynomial_sse(int polys[3]){ |
| 65 | int state; |
| 66 | |
| 67 | for(state=0;state < 128;state++){ |
| 68 | Branchtab39[0].s[state] = (polys[0] < 0) ^ parity((2*state) & polys[0]) ? 255:0; |
| 69 | Branchtab39[1].s[state] = (polys[1] < 0) ^ parity((2*state) & polys[1]) ? 255:0; |
| 70 | Branchtab39[2].s[state] = (polys[2] < 0) ^ parity((2*state) & polys[2]) ? 255:0; |
| 71 | } |
| 72 | Init++; |
| 73 | } |
| 74 | |
| 75 | /* Viterbi chainback */ |
| 76 | int chainback_viterbi39_sse( |
| 77 | void *p, |
| 78 | unsigned char *data, /* Decoded output data */ |
| 79 | unsigned int nbits, /* Number of data bits */ |
| 80 | unsigned int endstate){ /* Terminal encoder state */ |
| 81 | struct v39 *vp = p; |
| 82 | decision_t *d; |
| 83 | int path_metric; |
| 84 | |
| 85 | if(p == NULL) |
| 86 | return -1; |
| 87 | d = (decision_t *)vp->decisions; |
| 88 | endstate %= 256; |
| 89 | |
| 90 | path_metric = vp->old_metrics->s[endstate]; |
| 91 | |
| 92 | /* The store into data[] only needs to be done every 8 bits. |
| 93 | * But this avoids a conditional branch, and the writes will |
| 94 | * combine in the cache anyway |
| 95 | */ |
| 96 | d += 8; /* Look past tail */ |
| 97 | while(nbits-- != 0){ |
| 98 | int k; |
| 99 | |
| 100 | /* k = (d[nbits].w[endstate/32] >> (endstate%32)) & 1;*/ |
| 101 | k = (d[nbits].c[endstate/8] >> (endstate%8)) & 1; |
| 102 | endstate = (k << 7) | (endstate >> 1); |
| 103 | data[nbits>>3] = endstate; |
| 104 | } |
| 105 | return path_metric - SHRT_MIN; |
| 106 | } |
| 107 | |
| 108 | /* Delete instance of a Viterbi decoder */ |
| 109 | void delete_viterbi39_sse(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_sse(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 | d = (decision_t *)vp->dp; |
| 127 | while(nbits--){ |
| 128 | __m64 sym0v,sym1v,sym2v; |
| 129 | void *tmp; |
| 130 | int i; |
| 131 | |
| 132 | /* Splat the 0th symbol across sym0v, the 1st symbol across sym1v, etc */ |
| 133 | sym0v = _mm_set1_pi16(syms[0]); |
| 134 | sym1v = _mm_set1_pi16(syms[1]); |
| 135 | sym2v = _mm_set1_pi16(syms[2]); |
| 136 | syms += 3; |
| 137 | |
| 138 | for(i=0;i<32;i++){ |
| 139 | __m64 decision0,decision1,metric,m_metric,m0,m1,m2,m3,survivor0,survivor1; |
| 140 | |
| 141 | /* Form branch metrics |
| 142 | * Because Branchtab takes on values 0 and 255, and the values of sym?v are offset binary in the range 0-255, |
| 143 | * the XOR operations constitute conditional negation. |
| 144 | * metric and m_metric (-metric) are in the range 0-765 |
| 145 | */ |
| 146 | m0 = _mm_add_pi16(_mm_xor_si64(Branchtab39[0].v[i],sym0v),_mm_xor_si64(Branchtab39[1].v[i],sym1v)); |
| 147 | metric = _mm_add_pi16(_mm_xor_si64(Branchtab39[2].v[i],sym2v),m0); |
| 148 | m_metric = _mm_sub_pi16(_mm_set1_pi16(765),metric); |
| 149 | |
| 150 | /* Add branch metrics to path metrics */ |
| 151 | m0 = _mm_adds_pi16(vp->old_metrics->v[i],metric); |
| 152 | m3 = _mm_adds_pi16(vp->old_metrics->v[32+i],metric); |
| 153 | m1 = _mm_adds_pi16(vp->old_metrics->v[32+i],m_metric); |
| 154 | m2 = _mm_adds_pi16(vp->old_metrics->v[i],m_metric); |
| 155 | |
| 156 | /* Compare and select */ |
| 157 | survivor0 = _mm_min_pi16(m0,m1); |
| 158 | survivor1 = _mm_min_pi16(m2,m3); |
| 159 | decision0 = _mm_cmpeq_pi16(survivor0,m1); |
| 160 | decision1 = _mm_cmpeq_pi16(survivor1,m3); |
| 161 | |
| 162 | /* Pack decisions into 8 bits and store */ |
| 163 | d->c[i] = _mm_movemask_pi8(_mm_unpacklo_pi8(_mm_packs_pi16(decision0,_mm_setzero_si64()),_mm_packs_pi16(decision1,_mm_setzero_si64()))); |
| 164 | |
| 165 | /* Store surviving metrics */ |
| 166 | vp->new_metrics->v[2*i] = _mm_unpacklo_pi16(survivor0,survivor1); |
| 167 | vp->new_metrics->v[2*i+1] = _mm_unpackhi_pi16(survivor0,survivor1); |
| 168 | } |
| 169 | /* See if we need to renormalize |
| 170 | * Max metric spread for this code with 0-255 branch metrics is 12750 |
| 171 | */ |
| 172 | if(vp->new_metrics->s[0] >= SHRT_MAX-5000){ |
| 173 | int i,adjust; |
| 174 | __m64 adjustv; |
| 175 | union { __m64 v; signed short w[4]; } t; |
| 176 | |
| 177 | /* Find smallest metric and set adjustv to bring it down to SHRT_MIN */ |
| 178 | adjustv = vp->new_metrics->v[0]; |
| 179 | for(i=1;i<64;i++) |
| 180 | adjustv = _mm_min_pi16(adjustv,vp->new_metrics->v[i]); |
| 181 | |
| 182 | adjustv = _mm_min_pi16(adjustv,_mm_srli_si64(adjustv,32)); |
| 183 | adjustv = _mm_min_pi16(adjustv,_mm_srli_si64(adjustv,16)); |
| 184 | t.v = adjustv; |
| 185 | adjust = t.w[0] - SHRT_MIN; |
| 186 | path_metric += adjust; |
| 187 | adjustv = _mm_set1_pi16(adjust); |
| 188 | |
| 189 | for(i=0;i<64;i++) |
| 190 | vp->new_metrics->v[i] = _mm_sub_pi16(vp->new_metrics->v[i],adjustv); |
| 191 | } |
| 192 | d++; |
| 193 | /* Swap pointers to old and new metrics */ |
| 194 | tmp = vp->old_metrics; |
| 195 | vp->old_metrics = vp->new_metrics; |
| 196 | vp->new_metrics = tmp; |
| 197 | } |
| 198 | vp->dp = d; |
| 199 | _mm_empty(); |
| 200 | return path_metric; |
| 201 | } |