Bill Yi | 4e213d5 | 2015-06-23 13:53:11 -0700 | [diff] [blame] | 1 | /* K=9 r=1/2 Viterbi decoder for SSE2 |
| 2 | * Copyright Feb 2004, Phil Karn, KA9Q |
| 3 | * May be used under the terms of the GNU Lesser General Public License (LGPL) |
| 4 | */ |
| 5 | #include <stdio.h> |
| 6 | #include <stdlib.h> |
| 7 | #include <memory.h> |
| 8 | #include <emmintrin.h> |
| 9 | #include "fec.h" |
| 10 | |
| 11 | typedef union { unsigned char c[256]; __m128i v[16];} metric_t; |
| 12 | typedef union { unsigned long w[8]; unsigned char c[32];} decision_t; |
| 13 | |
| 14 | union branchtab29 { unsigned char c[128]; } Branchtab29_sse2[2]; |
| 15 | static int Init = 0; |
| 16 | |
| 17 | /* State info for instance of Viterbi decoder |
| 18 | * Don't change this without also changing references in sse2bfly29.s! |
| 19 | */ |
| 20 | struct v29 { |
| 21 | metric_t metrics1; /* path metric buffer 1 */ |
| 22 | metric_t metrics2; /* path metric buffer 2 */ |
| 23 | decision_t *dp; /* Pointer to current decision */ |
| 24 | metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */ |
| 25 | decision_t *decisions; /* Beginning of decisions for block */ |
| 26 | }; |
| 27 | |
| 28 | /* Initialize Viterbi decoder for start of new frame */ |
| 29 | int init_viterbi29_sse2(void *p,int starting_state){ |
| 30 | struct v29 *vp = p; |
| 31 | int i; |
| 32 | |
| 33 | for(i=0;i<256;i++) |
| 34 | vp->metrics1.c[i] = 63; |
| 35 | |
| 36 | vp->old_metrics = &vp->metrics1; |
| 37 | vp->new_metrics = &vp->metrics2; |
| 38 | vp->dp = vp->decisions; |
| 39 | vp->old_metrics->c[starting_state & 255] = 0; /* Bias known start state */ |
| 40 | return 0; |
| 41 | } |
| 42 | |
| 43 | void set_viterbi29_polynomial_sse2(int polys[2]){ |
| 44 | int state; |
| 45 | |
| 46 | for(state=0;state < 128;state++){ |
| 47 | Branchtab29_sse2[0].c[state] = (polys[0] < 0) ^ parity((2*state) & abs(polys[0])) ? 255 : 0; |
| 48 | Branchtab29_sse2[1].c[state] = (polys[1] < 0) ^ parity((2*state) & abs(polys[1])) ? 255 : 0; |
| 49 | } |
| 50 | Init++; |
| 51 | } |
| 52 | |
| 53 | |
| 54 | /* Create a new instance of a Viterbi decoder */ |
| 55 | void *create_viterbi29_sse2(int len){ |
| 56 | void *p; |
| 57 | struct v29 *vp; |
| 58 | |
| 59 | if(!Init){ |
| 60 | int polys[2] = {V29POLYA,V29POLYB}; |
| 61 | |
| 62 | set_viterbi29_polynomial(polys); |
| 63 | } |
| 64 | /* Ordinary malloc() only returns 8-byte alignment, we need 16 */ |
| 65 | if(posix_memalign(&p, sizeof(__m128i),sizeof(struct v29))) |
| 66 | return NULL; |
| 67 | vp = (struct v29 *)p; |
| 68 | if((p = malloc((len+8)*sizeof(decision_t))) == NULL){ |
| 69 | free(vp); |
| 70 | return NULL; |
| 71 | } |
| 72 | vp->decisions = (decision_t *)p; |
| 73 | init_viterbi29_sse2(vp,0); |
| 74 | return vp; |
| 75 | } |
| 76 | |
| 77 | |
| 78 | /* Viterbi chainback */ |
| 79 | int chainback_viterbi29_sse2( |
| 80 | void *p, |
| 81 | unsigned char *data, /* Decoded output data */ |
| 82 | unsigned int nbits, /* Number of data bits */ |
| 83 | unsigned int endstate){ /* Terminal encoder state */ |
| 84 | struct v29 *vp = p; |
| 85 | decision_t *d; |
| 86 | |
| 87 | if(p == NULL) |
| 88 | return -1; |
| 89 | d = vp->decisions; |
| 90 | |
| 91 | /* Make room beyond the end of the encoder register so we can |
| 92 | * accumulate a full byte of decoded data |
| 93 | */ |
| 94 | endstate %= 256; |
| 95 | |
| 96 | /* The store into data[] only needs to be done every 8 bits. |
| 97 | * But this avoids a conditional branch, and the writes will |
| 98 | * combine in the cache anyway |
| 99 | */ |
| 100 | d += 8; /* Look past tail */ |
| 101 | while(nbits-- != 0){ |
| 102 | int k; |
| 103 | |
| 104 | k = (d[nbits].c[endstate/8] >> (endstate%8)) & 1; |
| 105 | data[nbits>>3] = endstate = (endstate >> 1) | (k << 7); |
| 106 | } |
| 107 | return 0; |
| 108 | } |
| 109 | |
| 110 | |
| 111 | /* Delete instance of a Viterbi decoder */ |
| 112 | void delete_viterbi29_sse2(void *p){ |
| 113 | struct v29 *vp = p; |
| 114 | |
| 115 | if(vp != NULL){ |
| 116 | free(vp->decisions); |
| 117 | free(vp); |
| 118 | } |
| 119 | } |