Cleaned viterbi. Reduced number of symbols in CP-based CFO estmation

matlab/tests/viterbi_bler.m

@@ -2,7 +2,7 @@
 clear
 blen=40;
 SNR_values_db=linspace(-3,4,8);
-Nrealizations=6000;
+Nrealizations=50000;
 
 addpath('../../build/srslte/lib/fec/test')
 

srslte/lib/fec/src/viterbi.c

@@ -38,11 +38,12 @@
 
 #define DEB 0
 
+#define TB_ITER 3
+
 //#undef LV_HAVE_SSE
 
 int decode37(void *o, uint8_t *symbols, uint8_t *data, uint32_t frame_length) {
   srslte_viterbi_t *q = o;
-  uint32_t i;
 
   uint32_t best_state;
 
@@ -56,23 +57,18 @@ int decode37(void *o, uint8_t *symbols, uint8_t *data, uint32_t frame_length) {
   init_viterbi37_port(q->ptr, q->tail_biting ? -1 : 0);
 
   /* Decode block */
-  uint8_t *tmp = q->tmp;
   if (q->tail_biting) {
-    memcpy(tmp, symbols, 3 * frame_length * sizeof(uint8_t));
-    for (i = 0; i < 3 * (q->K - 1); i++) {
-      q->tmp[i + 3 * frame_length] = q->tmp[i];
+    for (int i=0;i<TB_ITER;i++) {
+      memcpy(&q->tmp[i*3*frame_length], symbols, 3*frame_length*sizeof(uint8_t));      
     }
+    update_viterbi37_blk_port(q->ptr, q->tmp, TB_ITER*frame_length, &best_state);
+    chainback_viterbi37_port(q->ptr,  q->tmp, TB_ITER*frame_length, best_state);
+    memcpy(data, &q->tmp[((int) (TB_ITER/2))*frame_length], frame_length*sizeof(uint8_t));
   } else {
-    tmp = symbols;
+    update_viterbi37_blk_port(q->ptr, symbols, frame_length + q->K - 1, NULL);
+    chainback_viterbi37_port(q->ptr, data, frame_length, 0);
   }
 
-  update_viterbi37_blk_port(q->ptr, tmp, frame_length + q->K - 1,
-      q->tail_biting ? &best_state : NULL);
-
-  /* Do Viterbi chainback */
-  chainback_viterbi37_port(q->ptr, data, frame_length,
-      q->tail_biting ? best_state : 0);
-
   return q->framebits;
 }
 
@@ -90,22 +86,21 @@ int decode37_sse(void *o, uint8_t *symbols, uint8_t *data, uint32_t frame_length
   }
 
   /* Initialize Viterbi decoder */
-  init_viterbi37_sse(q->ptr, -1);
+  init_viterbi37_sse(q->ptr, q->tail_biting?-1:0);
 
   /* Decode block */
   if (q->tail_biting) {
-    memcpy(q->tmp,                  symbols, 3*frame_length*sizeof(uint8_t));
-    memcpy(&q->tmp[3*frame_length], symbols, 3*frame_length*sizeof(uint8_t));    
-    memcpy(&q->tmp[6*frame_length], symbols, 3*frame_length*sizeof(uint8_t));    
-    update_viterbi37_blk_sse(q->ptr, q->tmp, 3*frame_length, &best_state);
-    chainback_viterbi37_sse(q->ptr,  q->tmp, 3*frame_length, best_state);
-    memcpy(data, &q->tmp[frame_length], frame_length*sizeof(uint8_t));
+    for (int i=0;i<TB_ITER;i++) {
+      memcpy(&q->tmp[i*3*frame_length], symbols, 3*frame_length*sizeof(uint8_t));      
+    }
+    update_viterbi37_blk_sse(q->ptr, q->tmp, TB_ITER*frame_length, &best_state);
+    chainback_viterbi37_sse(q->ptr,  q->tmp, TB_ITER*frame_length, best_state);
+    memcpy(data, &q->tmp[((int) (TB_ITER/2))*frame_length], frame_length*sizeof(uint8_t));
   } else {
     update_viterbi37_blk_sse(q->ptr, symbols, frame_length+q->K-1, &best_state);
     chainback_viterbi37_sse(q->ptr, data, frame_length, best_state);
   }
   
-  
   return q->framebits;
 }
 
@@ -148,7 +143,7 @@ int init37(srslte_viterbi_t *q, uint32_t poly[3], uint32_t framebits, bool tail_
     return -1;
   }
   if (q->tail_biting) {
-    q->tmp = srslte_vec_malloc(3 * (q->framebits + q->K - 1) * sizeof(uint8_t));
+    q->tmp = srslte_vec_malloc(TB_ITER * 3 * (q->framebits + q->K - 1) * sizeof(uint8_t));
     bzero(q->tmp, 3 * (q->framebits + q->K - 1) * sizeof(uint8_t));
     if (!q->tmp) {
       perror("malloc");
@@ -159,7 +154,7 @@ int init37(srslte_viterbi_t *q, uint32_t poly[3], uint32_t framebits, bool tail_
     q->tmp = NULL;
   }
   
-  if ((q->ptr = create_viterbi37_port(poly, framebits)) == NULL) {
+  if ((q->ptr = create_viterbi37_port(poly, TB_ITER*framebits)) == NULL) {
     fprintf(stderr, "create_viterbi37 failed\n");
     free37(q);
     return -1;
@@ -184,7 +179,7 @@ int init37_sse(srslte_viterbi_t *q, uint32_t poly[3], uint32_t framebits, bool t
     return -1;
   }
   if (q->tail_biting) {
-    q->tmp = srslte_vec_malloc(10 * (q->framebits + q->K - 1) * sizeof(uint8_t));
+    q->tmp = srslte_vec_malloc(TB_ITER*3*(q->framebits + q->K - 1) * sizeof(uint8_t));
     if (!q->tmp) {
       perror("malloc");
       free37(q);
@@ -194,7 +189,7 @@ int init37_sse(srslte_viterbi_t *q, uint32_t poly[3], uint32_t framebits, bool t
     q->tmp = NULL;
   }
   
-  if ((q->ptr = create_viterbi37_sse(poly, 3*framebits)) == NULL) {
+  if ((q->ptr = create_viterbi37_sse(poly, TB_ITER*framebits)) == NULL) {
     fprintf(stderr, "create_viterbi37 failed\n");
     free37(q);
     return -1;

srslte/lib/fec/src/viterbi37.h

@@ -51,6 +51,9 @@ void *create_viterbi37_sse(uint32_t polys[3],
 int init_viterbi37_sse(void *p, 
                         int starting_state);
 
+
+void reset_blk_sse(void *p, int nbits);
+
 int chainback_viterbi37_sse(void *p, 
                              uint8_t *data, 
                              uint32_t nbits, 

srslte/lib/fec/test/viterbi_test.c

@@ -213,7 +213,7 @@ int main(int argc, char **argv) {
       gettimeofday(&t[1], NULL);
       int M = 1; 
       
-      //srslte_vec_fprint_b(stdout, data_tx, frame_length);
+      srslte_vec_fprint_b(stdout, data_tx, frame_length);
       
       for (int i=0;i<M;i++) {
         srslte_viterbi_decode_uc(&dec, llr_c, data_rx, frame_length);      

srslte/lib/sync/src/pss.c

@@ -306,7 +306,6 @@ int srslte_pss_synch_find_pss(srslte_pss_synch_t *q, cf_t *input, float *corr_pe
     if (q->frame_size >= q->fft_size) {
     #ifdef CONVOLUTION_FFT
       memcpy(q->tmp_input, input, q->frame_size * sizeof(cf_t));
-            
       conv_output_len = srslte_conv_fft_cc_run(&q->conv_fft, q->tmp_input, q->pss_signal_time[q->N_id_2], q->conv_output);
     #else
       conv_output_len = srslte_conv_cc(input, q->pss_signal_time[q->N_id_2], q->conv_output, q->frame_size, q->fft_size);

srslte/lib/sync/src/sync.c

@@ -350,7 +350,7 @@ srslte_pss_synch_t* srslte_sync_get_cur_pss_obj(srslte_sync_t *q) {
 
 static float cfo_estimate(srslte_sync_t *q, cf_t *input) {
   uint32_t cp_offset = 0; 
-  cp_offset = srslte_cp_synch(&q->cp_synch, input, q->max_offset, q->nof_symbols, SRSLTE_CP_LEN_NORM(1,q->fft_size));
+  cp_offset = srslte_cp_synch(&q->cp_synch, input, q->max_offset, 2, SRSLTE_CP_LEN_NORM(1,q->fft_size));
   cf_t cp_corr_max = srslte_cp_synch_corr_output(&q->cp_synch, cp_offset);
   float cfo = -carg(cp_corr_max) / M_PI / 2; 
   return cfo; 
@@ -419,7 +419,6 @@ srslte_sync_find_ret_t srslte_sync_find(srslte_sync_t *q, cf_t *input, uint32_t
       
       /* Correct CFO with the averaged CFO estimation */
       srslte_cfo_correct(&q->cfocorr, input, input, -q->mean_cfo / q->fft_size);                             
-      
     }
     
     /* If integer CFO is enabled, find max PSS correlation for shifted +1/0/-1 integer versions */

srslte/lib/ue/src/ue_sync.c

@@ -528,6 +528,7 @@ int srslte_ue_sync_zerocopy(srslte_ue_sync_t *q, cf_t *input_buffer) {
               break;
             case SRSLTE_SYNC_FOUND_NOSPACE:
               /* If a peak was found but there is not enough space for SSS/CP detection, discard a few samples */
+              printf("No space for SSS/CP detection. Realigning frame...\n");
               q->recv_callback(q->stream, dummy_offset_buffer, q->frame_len/2, NULL); 
               srslte_sync_reset(&q->sfind);
               ret = SRSLTE_SUCCESS;