14e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi/* K=7 r=1/2 Viterbi decoder for SSE 24e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi * Feb 2004, Phil Karn, KA9Q 34e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi */ 44e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi#include <stdio.h> 54e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi#include <stdlib.h> 64e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi#include <memory.h> 74e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi#include <xmmintrin.h> 84e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi#include "fec.h" 94e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 104e213d510f437769f8a28578dd4f786fb7d16c4Bill Yitypedef union { unsigned char c[64]; } metric_t; 114e213d510f437769f8a28578dd4f786fb7d16c4Bill Yitypedef union { unsigned long w[2]; unsigned char c[8]; __m64 v[1];} decision_t; 124e213d510f437769f8a28578dd4f786fb7d16c4Bill Yiunion branchtab27 { unsigned char c[32]; __m64 v[4];} Branchtab27_sse[2]; 134e213d510f437769f8a28578dd4f786fb7d16c4Bill Yistatic int Init = 0; 144e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 154e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi/* State info for instance of Viterbi decoder 164e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi * Don't change this without also changing references in ssebfly27.s! 174e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi */ 184e213d510f437769f8a28578dd4f786fb7d16c4Bill Yistruct v27 { 194e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi metric_t metrics1; /* path metric buffer 1 */ 204e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi metric_t metrics2; /* path metric buffer 2 */ 214e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi decision_t *dp; /* Pointer to current decision */ 224e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi metric_t *old_metrics,*new_metrics; /* Pointers to path metrics, swapped on every bit */ 234e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi decision_t *decisions; /* Beginning of decisions for block */ 244e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi}; 254e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 264e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi/* Create a new instance of a Viterbi decoder */ 274e213d510f437769f8a28578dd4f786fb7d16c4Bill Yivoid *create_viterbi27_sse(int len){ 284e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi struct v27 *vp; 294e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 304e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi if(!Init){ 314e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi int polys[2] = { V27POLYA, V27POLYB }; 324e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 334e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi set_viterbi27_polynomial_sse(polys); 344e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi } 354e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi if((vp = malloc(sizeof(struct v27))) == NULL) 364e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi return NULL; 374e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi if((vp->decisions = malloc((len+6)*sizeof(decision_t))) == NULL){ 384e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi free(vp); 394e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi return NULL; 404e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi } 414e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi init_viterbi27(vp,0); 424e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi return vp; 434e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi} 444e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 454e213d510f437769f8a28578dd4f786fb7d16c4Bill Yivoid set_viterbi27_polynomial_sse(int polys[2]){ 464e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi int state; 474e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 484e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi for(state=0;state < 32;state++){ 494e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi Branchtab27_sse[0].c[state] = (polys[0] < 0) ^ parity((2*state) & abs(polys[0])) ? 255 : 0; 504e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi Branchtab27_sse[1].c[state] = (polys[1] < 0) ^ parity((2*state) & abs(polys[1])) ? 255 : 0; 514e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi } 524e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi Init++; 534e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi} 544e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 554e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi/* Initialize Viterbi decoder for start of new frame */ 564e213d510f437769f8a28578dd4f786fb7d16c4Bill Yiint init_viterbi27_sse(void *p,int starting_state){ 574e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi struct v27 *vp = p; 584e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi int i; 594e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 604e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi if(p == NULL) 614e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi return -1; 624e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi for(i=0;i<64;i++) 634e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi vp->metrics1.c[i] = 63; 644e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 654e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi vp->old_metrics = &vp->metrics1; 664e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi vp->new_metrics = &vp->metrics2; 674e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi vp->dp = vp->decisions; 684e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi vp->old_metrics->c[starting_state & 63] = 0; /* Bias known start state */ 694e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi return 0; 704e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi} 714e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 724e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi/* Viterbi chainback */ 734e213d510f437769f8a28578dd4f786fb7d16c4Bill Yiint chainback_viterbi27_sse( 744e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi void *p, 754e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi unsigned char *data, /* Decoded output data */ 764e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi unsigned int nbits, /* Number of data bits */ 774e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi unsigned int endstate){ /* Terminal encoder state */ 784e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi struct v27 *vp = p; 794e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi decision_t *d; 804e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 814e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi if(p == NULL) 824e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi return -1; 834e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 844e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi d = vp->decisions; 854e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi /* Make room beyond the end of the encoder register so we can 864e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi * accumulate a full byte of decoded data 874e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi */ 884e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi endstate %= 64; 894e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi endstate <<= 2; 904e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 914e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi /* The store into data[] only needs to be done every 8 bits. 924e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi * But this avoids a conditional branch, and the writes will 934e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi * combine in the cache anyway 944e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi */ 954e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi d += 6; /* Look past tail */ 964e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi while(nbits-- != 0){ 974e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi int k; 984e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 994e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi k = (d[nbits].c[(endstate>>2)/8] >> ((endstate>>2)%8)) & 1; 1004e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi data[nbits>>3] = endstate = (endstate >> 1) | (k << 7); 1014e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi } 1024e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi return 0; 1034e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi} 1044e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 1054e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi/* Delete instance of a Viterbi decoder */ 1064e213d510f437769f8a28578dd4f786fb7d16c4Bill Yivoid delete_viterbi27_sse(void *p){ 1074e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi struct v27 *vp = p; 1084e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi 1094e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi if(vp != NULL){ 1104e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi free(vp->decisions); 1114e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi free(vp); 1124e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi } 1134e213d510f437769f8a28578dd4f786fb7d16c4Bill Yi} 114