integrating SIMD optimized LDPC encoder

Conflicts:
	openair1/PHY/LTE_TRANSPORT/dlsch_coding.c

openair1/PHY/CODING/defs.h

@@ -573,6 +573,7 @@ void phy_viterbi_dot11(int8_t *,uint8_t *,uint16_t);
 
 short *ldpc_decoder(short *msgChannel,short block_length,short No_iteration,double rate);
 int encode_parity_check_part(uint16_t *c,uint16_t *d, short BG,short Zc,short Kb);
+int encode_parity_check_part_orig(unsigned char *c,unsigned char *d, short BG,short Zc,short Kb,short block_length);
 int ldpc_encoder(unsigned char *test_input,unsigned char *channel_input,short block_length,int nom_rate,int denom_rate);
 int ldpc_encoder_orig(unsigned char *test_input,unsigned char *channel_input,short block_length,int nom_rate,int denom_rate,unsigned char gen_code);
 int ldpc_encoder_multi_segment(unsigned char **test_input,unsigned char **channel_input,short block_length,double rate,uint8_t n_segments);

openair1/PHY/CODING/ldpc_encoder2.c

@@ -94,30 +94,28 @@ int ldpc_encoder_optim(unsigned char *test_input,unsigned char *channel_input,sh
   int simd_size;
 
   //determine number of bits in codeword
-  //if (block_length>3840)
-  //{
-  AssertFatal(block_length>3840,"Block length < 3840 not supported yet\n");
-  BG=1;
-  Kb = 22;
-  nrows=46; //parity check bits
-  ncols=22; //info bits
-  
-  //}
-  /* else if (block_length<=3840)
-  {
-    BG=2;
-    nrows=42; //parity check bits
-    ncols=10; // info bits
-
-    if (block_length>640)
-      Kb = 10;
-    else if (block_length>560)
-      Kb = 9;
-    else if (block_length>192)
+  if (block_length>3840)
+    {
+      BG=1;
+      Kb = 22;
+      nrows=46; //parity check bits
+      ncols=22; //info bits
+    }
+  else if (block_length<=3840)
+    {
+      BG=2;
+      nrows=42; //parity check bits
+      ncols=10; // info bits
+      
+      if (block_length>640)
+	Kb = 10;
+      else if (block_length>560)
+	Kb = 9;
+      else if (block_length>192)
       Kb = 8;
     else
       Kb = 6;
-  } */
+    }
 
   //find minimum value in all sets of lifting size
   Zc=0;
@@ -155,28 +153,42 @@ int ldpc_encoder_optim(unsigned char *test_input,unsigned char *channel_input,sh
     //c[i]=c[i]>>7&1;
     c[i]=(test_input[i/8]&(1<<(i&7)))>>(i&7);
   }
-  stop_meas(tinput); 
-  // extend matrix
-  start_meas(tprep);
-  for (i1=0; i1 < ncols; i1++)
-  {
-    memcpy(&c_extension[2*i1*Zc], &c[i1*Zc], Zc*sizeof(unsigned char));
-    memcpy(&c_extension[(2*i1+1)*Zc], &c[i1*Zc], Zc*sizeof(unsigned char));
+  stop_meas(tinput);
+
+  if (BG==1) { 
+    // extend matrix
+    start_meas(tprep);
+    for (i1=0; i1 < ncols; i1++)
+      {
+	memcpy(&c_extension[2*i1*Zc], &c[i1*Zc], Zc*sizeof(unsigned char));
+	memcpy(&c_extension[(2*i1+1)*Zc], &c[i1*Zc], Zc*sizeof(unsigned char));
+      }
+    for (i1=1;i1<simd_size;i1++) {
+      memcpy(&c_extension[(2*ncols*Zc*i1)], &c_extension[i1], (2*ncols*Zc*sizeof(unsigned char))-i1);
+      //    memset(&c_extension[(2*ncols*Zc*i1)],0,i1);
+      /*
+	printf("shift %d: ",i1);
+	for (int j=0;j<64;j++) printf("%d ",c_extension[(2*ncols*Zc*i1)+j]);
+	printf("\n");
+      */
+    }
+    stop_meas(tprep);
+    //parity check part
+    start_meas(tparity);
+    encode_parity_check_part_optim(c_extension, d, BG, Zc, Kb);
+    stop_meas(tparity);
   }
-  for (i1=1;i1<simd_size;i1++) {
-    memcpy(&c_extension[(2*ncols*Zc*i1)], &c_extension[i1], (2*ncols*Zc*sizeof(unsigned char))-i1);
-    //    memset(&c_extension[(2*ncols*Zc*i1)],0,i1);
-    /*
-    printf("shift %d: ",i1);
-    for (int j=0;j<64;j++) printf("%d ",c_extension[(2*ncols*Zc*i1)+j]);
-    printf("\n");
-    */
+  else if (BG==2) {
+    if (encode_parity_check_part_orig(c, d, BG, Zc, Kb, block_length)!=0) {
+      printf("Problem with encoder\n");
+      return(-1);
+    }
   }
-  stop_meas(tprep);
-  //parity check part
-  start_meas(tparity);
-  encode_parity_check_part_optim(c_extension, d, BG, Zc, Kb);
-  stop_meas(tparity);
+  else {
+      printf("Problem with encoder\n");
+      return(-1);
+  }
+
   start_meas(toutput);
   // information part and puncture columns
   memcpy(&channel_input[0], &c[2*Zc], (block_length-2*Zc)*sizeof(unsigned char));

openair1/PHY/LTE_TRANSPORT/dlsch_coding.c

@@ -709,6 +709,7 @@ int dlsch_encoding(PHY_VARS_eNB *eNB,
     else
     	Kr_int = Kr;
 
+    start_meas(te_stats);
     for (r=0; r<dlsch->harq_processes[harq_pid]->C; r++) {
       d_tmp[r] = &dlsch->harq_processes[harq_pid]->d[r][96];
 #ifdef DEBUG_DLSCH_CODING
@@ -716,9 +717,9 @@ int dlsch_encoding(PHY_VARS_eNB *eNB,
       printf("input %d %d %d %d %d \n", dlsch->harq_processes[harq_pid]->c[r][0], dlsch->harq_processes[harq_pid]->c[r][1], dlsch->harq_processes[harq_pid]->c[r][2],dlsch->harq_processes[harq_pid]->c[r][3], dlsch->harq_processes[harq_pid]->c[r][4]);
 #endif
       //ldpc_encoder((unsigned char*)dlsch->harq_processes[harq_pid]->c[r],&dlsch->harq_processes[harq_pid]->d[r][96],Kr,rate);
+      ldpc_encoder_optim((unsigned char*)dlsch->harq_processes[harq_pid]->c[r],(unsigned char*)&dlsch->harq_processes[harq_pid]->d[r][96],Kr,1,3,NULL,NULL,NULL,NULL);
     }
-    start_meas(te_stats);
-    ldpc_encoder_multi_segment(dlsch->harq_processes[harq_pid]->c,d_tmp,Kr,rate,dlsch->harq_processes[harq_pid]->C);
+    //ldpc_encoder_multi_segment(dlsch->harq_processes[harq_pid]->c,d_tmp,Kr,rate,dlsch->harq_processes[harq_pid]->C);
     stop_meas(te_stats);
 #endif
       /*printf("end ldpc encoder -- output\n");

openair1/PHY/TOOLS/time_meas.c

@@ -43,7 +43,7 @@ double get_cpu_freq_GHz(void) {
 
 void print_meas_now(time_stats_t *ts, const char* name, FILE* file_name){
 
-  if (opp_enabled) {
+  if (opp_enabled && ts) {
 
     //static double cpu_freq_GHz = 3.2;
 
@@ -62,7 +62,7 @@ void print_meas_now(time_stats_t *ts, const char* name, FILE* file_name){
 void print_meas(time_stats_t *ts, const char* name, time_stats_t * total_exec_time, time_stats_t * sf_exec_time)
 {
 
-  if (opp_enabled) {
+  if (opp_enabled && ts) {
 
     static int first_time = 0;
     static double cpu_freq_GHz = 0.0;

openair1/PHY/TOOLS/time_meas.h

@@ -98,7 +98,7 @@ static inline uint32_t rdtsc_oai(void)
 static inline void start_meas(time_stats_t *ts)
 {
 
-  if (opp_enabled) {
+  if (opp_enabled && ts) {
     if (ts->meas_flag==0) {
       ts->trials++;
       ts->in = rdtsc_oai();
@@ -113,7 +113,7 @@ static inline void start_meas(time_stats_t *ts)
 static inline void stop_meas(time_stats_t *ts)
 {
 
-  if (opp_enabled) {
+  if (opp_enabled && ts) {
     long long out = rdtsc_oai();
     
     ts->diff_now = (out-ts->in);
@@ -133,6 +133,8 @@ static inline void stop_meas(time_stats_t *ts)
 
 static inline void reset_meas(time_stats_t *ts) {
 
+  if (ts) {
+
   ts->trials=0;
   ts->diff=0;
   ts->diff_now=0;
@@ -141,6 +143,7 @@ static inline void reset_meas(time_stats_t *ts) {
   ts->max=0;
   ts->meas_flag=0;
   
+  }
 }
 
 static inline void copy_meas(time_stats_t *dst_ts,time_stats_t *src_ts)