Merge remote-tracking branch...

Merge remote-tracking branch 'origin/373-parallelization-not-working-correctly-in-phy-simulators-and-bad-implementation-2' into develop_integration_w48
Signed-off-by: Raphael Defosseux <raphael.defosseux@eurecom.fr>

common/utils/LOG/log.c

@@ -39,9 +39,8 @@
 #include "vcd_signal_dumper.h"
 #include "assertions.h"
 
-
-# include <pthread.h>
-# include <string.h>
+#include <pthread.h>
+#include <string.h>
 #include <linux/prctl.h>
 #include "common/config/config_userapi.h"
 // main log variables

nfapi/oai_integration/nfapi_vnf.c

@@ -196,14 +196,12 @@ void oai_create_enb(void) {
   int bodge_counter=0;
   PHY_VARS_eNB *eNB = RC.eNB[0][0];
 
-  printf("[VNF] RC.eNB[0][0]. Mod_id:%d CC_id:%d nb_CC[0]:%d abstraction_flag:%d single_thread_flag:%d td:%p te:%p if_inst:%p\n", eNB->Mod_id, eNB->CC_id, RC.nb_CC[0], eNB->abstraction_flag, eNB->single_thread_flag, eNB->td, eNB->te, eNB->if_inst);
+  printf("[VNF] RC.eNB[0][0]. Mod_id:%d CC_id:%d nb_CC[0]:%d abstraction_flag:%d single_thread_flag:%d if_inst:%p\n", eNB->Mod_id, eNB->CC_id, RC.nb_CC[0], eNB->abstraction_flag, eNB->single_thread_flag, eNB->if_inst);
 
   eNB->Mod_id  = bodge_counter;
   eNB->CC_id   = bodge_counter;
   eNB->abstraction_flag   = 0;
   eNB->single_thread_flag = 0;//single_thread_flag;
-  eNB->td                   = ulsch_decoding_data_all;//(single_thread_flag==1) ? ulsch_decoding_data_2thread : ulsch_decoding_data;
-  eNB->te                   = dlsch_encoding_all;//(single_thread_flag==1) ? dlsch_encoding_2threads : dlsch_encoding;
 
   RC.nb_CC[bodge_counter] = 1;
 

openair1/PHY/LTE_TRANSPORT/ulsch_decoding.c

@@ -1505,7 +1505,7 @@ unsigned int  ulsch_decoding(PHY_VARS_eNB *eNB,eNB_rxtx_proc_t *proc,
 
   // Do ULSCH Decoding for data portion
 
-  ret = eNB->td(eNB,UE_id,harq_pid,llr8_flag);
+  ret = ulsch_decoding_data_all(eNB,UE_id,harq_pid,llr8_flag);
 
   VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_PHY_ENB_ULSCH_DECODING0+harq_pid,0);
 

openair1/PHY/LTE_UE_TRANSPORT/dlsch_demodulation.c

@@ -1298,7 +1298,7 @@ void dlsch_channel_compensation(int **rxdataF_ext,
   unsigned short rb;
   unsigned char aatx,aarx,symbol_mod,pilots=0;
   __m128i *dl_ch128,*dl_ch128_2,*dl_ch_mag128,*dl_ch_mag128b,*rxdataF128,*rxdataF_comp128,*rho128;
-  __m128i mmtmpD0,mmtmpD1,mmtmpD2,mmtmpD3,QAM_amp128,QAM_amp128b;
+  __m128i mmtmpD0,mmtmpD1,mmtmpD2,mmtmpD3,QAM_amp128;
 
   symbol_mod = (symbol>=(7-frame_parms->Ncp)) ? symbol-(7-frame_parms->Ncp) : symbol;
 
@@ -1311,9 +1311,9 @@ void dlsch_channel_compensation(int **rxdataF_ext,
   }
 
   for (aatx=0; aatx<frame_parms->nb_antenna_ports_eNB; aatx++) {
+    __m128i QAM_amp128b = _mm_setzero_si128();
     if (mod_order == 4) {
       QAM_amp128 = _mm_set1_epi16(QAM16_n1);  // 2/sqrt(10)
-      QAM_amp128b = _mm_setzero_si128();
     } else if (mod_order == 6) {
       QAM_amp128  = _mm_set1_epi16(QAM64_n1); //
       QAM_amp128b = _mm_set1_epi16(QAM64_n2);
@@ -1766,11 +1766,11 @@ void dlsch_channel_compensation_core(int **rxdataF_ext,
   int length_mod8 = 0;
   int length2;
   __m128i *dl_ch128,*dl_ch_mag128,*dl_ch_mag128b, *dl_ch128_2, *rxdataF128,*rxdataF_comp128,*rho128;
-  __m128i mmtmpD0,mmtmpD1,mmtmpD2,mmtmpD3,QAM_amp128,QAM_amp128b;
+  __m128i mmtmpD0,mmtmpD1,mmtmpD2,mmtmpD3,QAM_amp128;
   int aatx = 0, aarx = 0;
 
   for (aatx=0; aatx<n_tx; aatx++) {
-
+    __m128i QAM_amp128b;
     if (mod_order == 4) {
       QAM_amp128 = _mm_set1_epi16(QAM16_n1);  // 2/sqrt(10)
       QAM_amp128b = _mm_setzero_si128();
@@ -2158,7 +2158,7 @@ void dlsch_channel_compensation_TM56(int **rxdataF_ext,
   __m128i *dl_ch0_128,*dl_ch1_128,*dl_ch_mag128,*dl_ch_mag128b,*rxdataF128,*rxdataF_comp128;
   unsigned char aarx=0,symbol_mod,pilots=0;
   int precoded_signal_strength=0;
-  __m128i mmtmpD0,mmtmpD1,mmtmpD2,mmtmpD3,QAM_amp128,QAM_amp128b;
+  __m128i mmtmpD0,mmtmpD1,mmtmpD2,mmtmpD3,QAM_amp128;
 
   symbol_mod = (symbol>=(7-frame_parms->Ncp)) ? symbol-(7-frame_parms->Ncp) : symbol;
 
@@ -2167,10 +2167,9 @@ void dlsch_channel_compensation_TM56(int **rxdataF_ext,
 
 
   //printf("comp prec: symbol %d, pilots %d\n",symbol, pilots);
-
+  __m128i QAM_amp128b = _mm_setzero_si128();
   if (mod_order == 4) {
     QAM_amp128 = _mm_set1_epi16(QAM16_n1);
-    QAM_amp128b = _mm_setzero_si128();
   } else if (mod_order == 6) {
     QAM_amp128  = _mm_set1_epi16(QAM64_n1);
     QAM_amp128b = _mm_set1_epi16(QAM64_n2);
@@ -2621,7 +2620,7 @@ void dlsch_channel_compensation_TM34(LTE_DL_FRAME_PARMS *frame_parms,
   int **rxdataF_comp0         = pdsch_vars->rxdataF_comp0;
   int **rxdataF_comp1         = pdsch_vars->rxdataF_comp1[harq_pid][round];
   unsigned char *pmi_ext      = pdsch_vars->pmi_ext;
-  __m128i mmtmpD0,mmtmpD1,mmtmpD2,mmtmpD3,QAM_amp0_128,QAM_amp0_128b,QAM_amp1_128,QAM_amp1_128b;
+  __m128i mmtmpD0,mmtmpD1,mmtmpD2,mmtmpD3,QAM_amp0_128,QAM_amp1_128;
 
   symbol_mod = (symbol>=(7-frame_parms->Ncp)) ? symbol-(7-frame_parms->Ncp) : symbol;
 
@@ -2632,17 +2631,17 @@ void dlsch_channel_compensation_TM34(LTE_DL_FRAME_PARMS *frame_parms,
 
  // printf("comp prec: symbol %d, pilots %d\n",symbol, pilots);
 
+  __m128i  QAM_amp0_128b = _mm_setzero_si128();
   if (mod_order0 == 4) {
     QAM_amp0_128  = _mm_set1_epi16(QAM16_n1);
-    QAM_amp0_128b = _mm_setzero_si128();
   } else if (mod_order0 == 6) {
     QAM_amp0_128  = _mm_set1_epi16(QAM64_n1);
     QAM_amp0_128b = _mm_set1_epi16(QAM64_n2);
   }
 
+  __m128i  QAM_amp1_128b = _mm_setzero_si128();
   if (mod_order1 == 4) {
     QAM_amp1_128  = _mm_set1_epi16(QAM16_n1);
-    QAM_amp1_128b = _mm_setzero_si128();
   } else if (mod_order1 == 6) {
     QAM_amp1_128  = _mm_set1_epi16(QAM64_n1);
     QAM_amp1_128b = _mm_set1_epi16(QAM64_n2);

openair1/PHY/defs_common.h

@@ -58,6 +58,7 @@
 #include <math.h>
 #include "common_lib.h"
 #include "msc.h"
+#include <common/utils/LOG/log.h>
 
 
 //#include <complex.h>
@@ -862,22 +863,60 @@ typedef enum {
   RESYNCH=4
 } UE_MODE_t;
 
-/// Threading Parameter
+#define FOREACH_PARALLEL(GEN)			\
+  GEN(PARALLEL_SINGLE_THREAD)			\
+  GEN(PARALLEL_RU_L1_SPLIT)			\
+  GEN(PARALLEL_RU_L1_TRX_SPLIT)
+
+#define GENERATE_ENUM(N) N,
+#define GENERATE_ENUMTXT(N) {(char*)#N, N},
+
 typedef enum {
-  PARALLEL_SINGLE_THREAD    =0,
-  PARALLEL_RU_L1_SPLIT      =1,
-  PARALLEL_RU_L1_TRX_SPLIT  =2
-}PARALLEL_CONF_t;
+  FOREACH_PARALLEL(GENERATE_ENUM)
+} PARALLEL_CONF_t;
 
+#define FOREACH_WORKER(GEN) GEN(WORKER_DISABLE) GEN(WORKER_ENABLE)
 typedef enum {
-  WORKER_DISABLE            =0,
-  WORKER_ENABLE             =1
+  FOREACH_WORKER(GENERATE_ENUM)
 }WORKER_CONF_t;
 
 typedef struct THREAD_STRUCT_s {
   PARALLEL_CONF_t  parallel_conf;
   WORKER_CONF_t    worker_conf;
 } THREAD_STRUCT;
+extern THREAD_STRUCT  thread_struct;
+
+static inline void set_parallel_conf(char *parallel_conf) {
+  mapping config[]= {
+    FOREACH_PARALLEL(GENERATE_ENUMTXT)
+    {NULL,-1}
+  };
+  thread_struct.parallel_conf = (PARALLEL_CONF_t)map_str_to_int(config, parallel_conf);
+  if (thread_struct.parallel_conf == -1 ) {
+    LOG_E(ENB_APP,"Impossible value: %s\n", parallel_conf);
+    thread_struct.parallel_conf = PARALLEL_SINGLE_THREAD;
+  }
+}
+
+static inline void set_worker_conf(char *worker_conf) {
+  mapping config[]={
+    FOREACH_WORKER(GENERATE_ENUMTXT)
+    {NULL, -1}
+  };
+  thread_struct.worker_conf =  (WORKER_CONF_t)map_str_to_int(config, worker_conf);
+  if (thread_struct.worker_conf == -1 ) {
+    LOG_E(ENB_APP,"Impossible value: %s\n", worker_conf);
+    thread_struct.worker_conf = WORKER_DISABLE ;
+  }
+}
+
+static inline PARALLEL_CONF_t get_thread_parallel_conf(void) {
+  return thread_struct.parallel_conf;
+}
+
+static inline WORKER_CONF_t get_thread_worker_conf(void) {
+  return thread_struct.worker_conf;
+}
 
 typedef enum {SF_DL, SF_UL, SF_S} lte_subframe_t;
 

openair1/PHY/defs_eNB.h

@@ -904,8 +904,6 @@ typedef struct PHY_VARS_eNB_s {
   /// Ethernet parameters for fronthaul interface
   eth_params_t         eth_params;
   int                  rx_total_gain_dB;
-  int                  (*td)(struct PHY_VARS_eNB_s *eNB,int UE_id,int harq_pid,int llr8_flag);
-  int                  (*te)(struct PHY_VARS_eNB_s *,uint8_t *,uint8_t,LTE_eNB_DLSCH_t *,int,uint8_t,time_stats_t *,time_stats_t *,time_stats_t *,time_stats_t *,time_stats_t *,time_stats_t *,time_stats_t *);
   int                  (*start_if)(struct RU_t_s *ru,struct PHY_VARS_eNB_s *eNB);
   uint8_t              local_flag;
   LTE_DL_FRAME_PARMS   frame_parms;
@@ -1080,6 +1078,7 @@ typedef struct PHY_VARS_eNB_s {
 
   time_stats_t ofdm_mod_stats;
   time_stats_t dlsch_common_and_dci;
+  time_stats_t dlsch_ue_specific;
   time_stats_t dlsch_encoding_stats;
   time_stats_t dlsch_modulation_stats;
   time_stats_t dlsch_scrambling_stats;

openair1/SCHED/phy_procedures_lte_eNb.c

@@ -400,7 +400,7 @@ void pdsch_procedures(PHY_VARS_eNB *eNB,
 
     start_meas(&eNB->dlsch_encoding_stats);
 
-    eNB->te(eNB,
+    dlsch_encoding_all(eNB,
 	  dlsch_harq->pdu,
 	  dlsch_harq->pdsch_start,
 	  dlsch,
@@ -582,6 +582,7 @@ void phy_procedures_eNB_TX(PHY_VARS_eNB *eNB,
   }
 
   if (do_meas==1) stop_meas(&eNB->dlsch_common_and_dci);
+  if (do_meas==1) start_meas(&eNB->dlsch_ue_specific);
 
   VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_PHY_ENB_PDCCH_TX,0);
 
@@ -649,6 +650,7 @@ void phy_procedures_eNB_TX(PHY_VARS_eNB *eNB,
   VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_GENERATE_PHICH,0);
 
   VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_PHY_PROCEDURES_ENB_TX+(eNB->CC_id),0);
+  if (do_meas==1) stop_meas(&eNB->dlsch_ue_specific);
   if (do_meas==1) stop_meas(&eNB->phy_proc_tx);
   
 }

openair1/SIMULATION/LTE_PHY/common_sim.h

@@ -49,38 +49,46 @@ void sumUpStatsSlot(time_stats_t *res, time_stats_t src[RX_NB_TH][2], int lastAc
 	res->p_time=src[lastActive][last].p_time;
 }
 
+double squareRoot(time_stats_t *ptr) {
+  double timeBase=1/(1000*cpu_freq_GHz);
+  return sqrt((double)ptr->diff_square*pow(timeBase,2)/ptr->trials -
+              pow((double)ptr->diff/ptr->trials*timeBase,2));
+}
+
+void printDistribution(time_stats_t *ptr, varArray_t *sortedList, char *txt) {
+  double timeBase=1/(1000*cpu_freq_GHz);
+  printf("%-43s %6.2f us (%d trials)\n",
+         txt,
+         (double)ptr->diff/ptr->trials*timeBase,
+         ptr->trials);
+  printf(" Statistics std=%.2f, median=%.2f, q1=%.2f, q3=%.2f µs (on %ld trials)\n",
+         squareRoot(ptr), median(sortedList),q1(sortedList),q3(sortedList), sortedList->size);
+}
+
 void printStatIndent(time_stats_t *ptr, char *txt) {
-  printf("|__ %-50s %.2f us (%d trials)\n",
+  printf("|__ %-38s %6.2f us (%3d trials)\n",
          txt,
          ptr->trials?inMicroS(ptr->diff/ptr->trials):0,
          ptr->trials);
 }
 
-void printStatIndent2(time_stats_t *ptr, char *txt, int turbo_iter) {
+void printStatIndent2(time_stats_t *ptr, char *txt) {
   double timeBase=1/(1000*cpu_freq_GHz);
-  printf("    |__ %-45s %.2f us (cycles/block %ld, %5d trials)\n",
+  printf("    |__ %-34s %6.2f us (%3d trials)\n",
          txt,
          ptr->trials?((double)ptr->diff)/ptr->trials*timeBase:0,
-         turbo_iter?(uint64_t)round(((double)ptr->diff)/turbo_iter):0,
 	 ptr->trials);
 }
 
-double squareRoot(time_stats_t *ptr) {
-  double timeBase=1/(1000*cpu_freq_GHz);
-  return sqrt((double)ptr->diff_square*pow(timeBase,2)/ptr->trials -
-              pow((double)ptr->diff/ptr->trials*timeBase,2));
-}
-
-void printDistribution(time_stats_t *ptr, varArray_t *sortedList, char *txt) {
+void printStatIndent3(time_stats_t *ptr, char *txt) {
   double timeBase=1/(1000*cpu_freq_GHz);
-  printf("%-50s             :%.2f us (%d trials)\n",
+  printf("        |__ %-30s %6.2f us (%3d trials)\n",
          txt,
-         (double)ptr->diff/ptr->trials*timeBase,
+         ptr->trials?((double)ptr->diff)/ptr->trials*timeBase:0,
 	 ptr->trials);
-  printf("|__ Statistics std=%.2f, median=%.2f, q1=%.2f, q3=%.2f µs (on %ld trials)\n",
-         squareRoot(ptr), median(sortedList),q1(sortedList),q3(sortedList), sortedList->size);
 }
 
+
 void logDistribution(FILE* fd, time_stats_t *ptr, varArray_t *sortedList, int dropped) {
   fprintf(fd,"%f;%f;%f;%f;%f;%f;%d;",
 	  squareRoot(ptr),

openair1/SIMULATION/LTE_PHY/dlsim.c

@@ -85,30 +85,7 @@ double t_rx_min = 1000000000; /*!< \brief initial min process time for rx */
 int n_tx_dropped = 0; /*!< \brief initial max process time for tx */
 int n_rx_dropped = 0; /*!< \brief initial max process time for rx */
 
-char *parallel_config = NULL;
-char *worker_config = NULL;
-static THREAD_STRUCT thread_struct;
-void set_parallel_conf(char *parallel_conf)
-{
-  if(strcmp(parallel_conf,"PARALLEL_SINGLE_THREAD")==0)           thread_struct.parallel_conf = PARALLEL_SINGLE_THREAD;
-  else if(strcmp(parallel_conf,"PARALLEL_RU_L1_SPLIT")==0)        thread_struct.parallel_conf = PARALLEL_RU_L1_SPLIT;
-  else if(strcmp(parallel_conf,"PARALLEL_RU_L1_TRX_SPLIT")==0)    thread_struct.parallel_conf = PARALLEL_RU_L1_TRX_SPLIT;
-  printf("[CONFIG] parallel conf is set to %d\n",thread_struct.parallel_conf);
-} 
-void set_worker_conf(char *worker_conf)
-{
-  if(strcmp(worker_conf,"WORKER_DISABLE")==0)                     thread_struct.worker_conf = WORKER_DISABLE;
-  else if(strcmp(worker_conf,"WORKER_ENABLE")==0)                 thread_struct.worker_conf = WORKER_ENABLE;
-  printf("[CONFIG] worker conf is set to %d\n",thread_struct.worker_conf);
-} 
-PARALLEL_CONF_t get_thread_parallel_conf(void)
-{
-  return thread_struct.parallel_conf;
-} 
-WORKER_CONF_t get_thread_worker_conf(void)
-{
-  return thread_struct.worker_conf;
-} 
+THREAD_STRUCT thread_struct;
 
 int emulate_rf = 0;
 
@@ -715,7 +692,7 @@ int main(int argc, char **argv)
 
   DL_req.dl_config_request_body.dl_config_pdu_list = dl_config_pdu_list;
   TX_req.tx_request_body.tx_pdu_list = tx_pdu_list;
-
+  set_parallel_conf("PARALLEL_SINGLE_THREAD");
   cpuf = cpu_freq_GHz;
 
   //signal(SIGSEGV, handler);
@@ -758,6 +735,7 @@ int main(int argc, char **argv)
     { "Subframe", "subframe ",0, iptr:&subframe,  defintval:7, TYPE_INT, 0 },
     { "Trnti", "rnti",0, u16ptr:&n_rnti,  defuintval:0x1234, TYPE_UINT16, 0 },
     { "vi_mod", "i_mod",0, iptr:NULL,  defintval:0, TYPE_INT, 0 },
+    { "Qparallel", "Enable parallel execution",0, strptr:NULL,  defstrval:NULL, TYPE_STRING,  0 },
     { "Performance", "Display CPU perfomance of each L1 piece", PARAMFLAG_BOOL,  iptr:&print_perf,  defintval:0, TYPE_INT, 0 },
     { "q_tx_port", "Number of TX antennas ports used in eNB",0, iptr:NULL,  defintval:0, TYPE_INT, 0 },
     { "uEdual", "Enables the Interference Aware Receiver for TM5 (default is normal receiver)",0, iptr:NULL,  defintval:0, TYPE_INT, 0 },
@@ -965,6 +943,10 @@ int main(int argc, char **argv)
 
       break;
 
+    case 'Q':
+      set_parallel_conf(optarg);
+      break;
+      
     default:
       printf("Wrong option: %s\n",long_options[option_index].name);
       exit(1);
@@ -982,7 +964,7 @@ int main(int argc, char **argv)
   if (help)
     exit(0);
   
-  set_parallel_conf("PARALLEL_RU_L1_TRX_SPLIT");
+  if (thread_struct.parallel_conf != PARALLEL_SINGLE_THREAD)
     set_worker_conf("WORKER_ENABLE");
 
   if (transmission_mode>1) pa=dBm3;
@@ -1079,11 +1061,7 @@ int main(int argc, char **argv)
      ru->do_precoding=1;
 
   eNB->mac_enabled=1;
-  if (two_thread_flag == 0) {
-    eNB->te = dlsch_encoding;
-  }
-  else {
-    eNB->te = dlsch_encoding_2threads;
+  if(get_thread_worker_conf() == WORKER_ENABLE) {
     extern void init_td_thread(PHY_VARS_eNB *);
     extern void init_te_thread(PHY_VARS_eNB *);
     init_td_thread(eNB);
@@ -1442,6 +1420,8 @@ int main(int argc, char **argv)
       reset_meas(&eNB->dlsch_interleaving_stats);
       reset_meas(&eNB->dlsch_rate_matching_stats);
       reset_meas(&eNB->dlsch_turbo_encoding_stats);
+      reset_meas(&eNB->dlsch_common_and_dci);
+      reset_meas(&eNB->dlsch_ue_specific);
       for (int i=0; i<RX_NB_TH; i++) {
 	reset_meas(&UE->phy_proc_rx[i]); // total UE rx
 	reset_meas(&UE->ue_front_end_stat[i]);
@@ -1783,9 +1763,6 @@ int main(int argc, char **argv)
 
 	  }
 
-
-
-
           if (UE->dlsch[UE->current_thread_id[subframe]][eNB_id][0]->harq_ack[subframe].ack == 1) {
 
             avg_iter += UE->dlsch[UE->current_thread_id[subframe]][eNB_id][0]->last_iteration_cnt;
@@ -2011,21 +1988,22 @@ int main(int argc, char **argv)
         printf("\neNB TX function statistics (per 1ms subframe)\n");
 	printDistribution(&eNB->phy_proc_tx,table_tx,"PHY proc tx");
 	printStatIndent(&eNB->dlsch_common_and_dci,"DL common channels and dci time");
-	printStatIndent(&eNB->dlsch_encoding_stats,"DLSCH encoding time");
-	printStatIndent2(&eNB->dlsch_rate_matching_stats,"DLSCH rate matching time",eNB->dlsch_rate_matching_stats.trials);
-	printStatIndent2(&eNB->dlsch_turbo_encoding_stats,"DLSCH turbo encoding time", eNB->dlsch_turbo_encoding_stats.trials);
-	printStatIndent2(&eNB->dlsch_interleaving_stats,"DLSCH interleaving time", eNB->dlsch_interleaving_stats.trials);
-	printStatIndent(&eNB->dlsch_scrambling_stats,  "DLSCH scrambling time");
-	printStatIndent(&eNB->dlsch_modulation_stats, "DLSCH modulation time");
+	printStatIndent(&eNB->dlsch_ue_specific,"DL per ue part time");
+	printStatIndent2(&eNB->dlsch_encoding_stats,"DLSCH encoding time");
+	printStatIndent3(&eNB->dlsch_rate_matching_stats,"DLSCH rate matching time");
+	printStatIndent3(&eNB->dlsch_turbo_encoding_stats,"DLSCH turbo encoding time");
+	printStatIndent3(&eNB->dlsch_interleaving_stats,"DLSCH interleaving time");
+	printStatIndent2(&eNB->dlsch_scrambling_stats,  "DLSCH scrambling time");
+	printStatIndent2(&eNB->dlsch_modulation_stats, "DLSCH modulation time");
 	printDistribution(&eNB->ofdm_mod_stats,table_tx_ifft,"OFDM_mod (idft) time");
 
         printf("\nUE RX function statistics (per 1ms subframe)\n");
 	printDistribution(&phy_proc_rx_tot, table_rx,"Total PHY proc rx");
 	printStatIndent(&ue_front_end_tot,"Front end processing");
 	printStatIndent(&dlsch_llr_tot,"rx_pdsch processing");
-	printStatIndent2(&pdsch_procedures_tot,"pdsch processing", pdsch_procedures_tot.trials);
-	printStatIndent2(&dlsch_procedures_tot,"dlsch processing", dlsch_procedures_tot.trials);
-	printStatIndent2(&UE->crnti_procedures_stats,"C-RNTI processing", UE->crnti_procedures_stats.trials);
+	printStatIndent2(&pdsch_procedures_tot,"pdsch processing");
+	printStatIndent2(&dlsch_procedures_tot,"dlsch processing");
+	printStatIndent2(&UE->crnti_procedures_stats,"C-RNTI processing");
 	printStatIndent(&UE->ofdm_demod_stats,"ofdm demodulation");
 	printStatIndent(&UE->dlsch_channel_estimation_stats,"DLSCH channel estimation time");
 	printStatIndent(&UE->dlsch_freq_offset_estimation_stats,"DLSCH frequency offset estimation time");
@@ -2041,13 +2019,13 @@ int main(int argc, char **argv)
                (double)UE->dlsch_turbo_decoding_stats.diff/UE->dlsch_turbo_decoding_stats.trials*timeBase,
                (int)((double)UE->dlsch_turbo_decoding_stats.diff/UE->dlsch_turbo_decoding_stats.trials),
                UE->dlsch_turbo_decoding_stats.trials);
-        printStatIndent2(&UE->dlsch_tc_init_stats,"init", UE->dlsch_tc_init_stats.trials);
-        printStatIndent2(&UE->dlsch_tc_alpha_stats,"alpha", UE->dlsch_tc_init_stats.trials);
-        printStatIndent2(&UE->dlsch_tc_beta_stats,"beta", UE->dlsch_tc_init_stats.trials);
-        printStatIndent2(&UE->dlsch_tc_gamma_stats,"gamma", UE->dlsch_tc_init_stats.trials);
-        printStatIndent2(&UE->dlsch_tc_ext_stats,"ext", UE->dlsch_tc_init_stats.trials);
-        printStatIndent2(&UE->dlsch_tc_intl1_stats,"turbo internal interleaver", UE->dlsch_tc_init_stats.trials);
-        printStatIndent2(&UE->dlsch_tc_intl2_stats,"intl2+HardDecode+CRC", UE->dlsch_tc_init_stats.trials);
+        printStatIndent2(&UE->dlsch_tc_init_stats,"init");
+        printStatIndent2(&UE->dlsch_tc_alpha_stats,"alpha");
+        printStatIndent2(&UE->dlsch_tc_beta_stats,"beta");
+        printStatIndent2(&UE->dlsch_tc_gamma_stats,"gamma");
+        printStatIndent2(&UE->dlsch_tc_ext_stats,"ext");
+        printStatIndent2(&UE->dlsch_tc_intl1_stats,"turbo internal interleaver");
+        printStatIndent2(&UE->dlsch_tc_intl2_stats,"intl2+HardDecode+CRC");
 	
       }
 

openair1/SIMULATION/LTE_PHY/ulsim.c

@@ -96,30 +96,7 @@ nfapi_tx_request_pdu_t tx_pdu_list[MAX_NUM_TX_REQUEST_PDU];
 nfapi_tx_request_t TX_req;
 Sched_Rsp_t sched_resp;
 
-char *parallel_config = NULL;
-char *worker_config = NULL;
-static THREAD_STRUCT thread_struct;
-void set_parallel_conf(char *parallel_conf)
-{
-  if(strcmp(parallel_conf,"PARALLEL_SINGLE_THREAD")==0)           thread_struct.parallel_conf = PARALLEL_SINGLE_THREAD;
-  else if(strcmp(parallel_conf,"PARALLEL_RU_L1_SPLIT")==0)        thread_struct.parallel_conf = PARALLEL_RU_L1_SPLIT;
-  else if(strcmp(parallel_conf,"PARALLEL_RU_L1_TRX_SPLIT")==0)    thread_struct.parallel_conf = PARALLEL_RU_L1_TRX_SPLIT;
-  printf("[CONFIG] parallel conf is set to %d\n",thread_struct.parallel_conf);
-} 
-void set_worker_conf(char *worker_conf)
-{
-  if(strcmp(worker_conf,"WORKER_DISABLE")==0)                     thread_struct.worker_conf = WORKER_DISABLE;
-  else if(strcmp(worker_conf,"WORKER_ENABLE")==0)                 thread_struct.worker_conf = WORKER_ENABLE;
-  printf("[CONFIG] worker conf is set to %d\n",thread_struct.worker_conf);
-} 
-PARALLEL_CONF_t get_thread_parallel_conf(void)
-{
-  return thread_struct.parallel_conf;
-} 
-WORKER_CONF_t get_thread_worker_conf(void)
-{
-  return thread_struct.worker_conf;
-} 
+THREAD_STRUCT thread_struct;
 
 void
 fill_nfapi_ulsch_config_request(nfapi_ul_config_request_pdu_t *ul_config_pdu,
@@ -410,7 +387,6 @@ int main(int argc, char **argv) {
   double effective_rate=0.0;
   char channel_model_input[10]= {0};
   static int max_turbo_iterations=4;
-  static int parallel_flag=0;
   int nb_rb_set = 0;
   int sf;
   static int threequarter_fs=0;
@@ -428,6 +404,8 @@ int main(int argc, char **argv) {
   TX_req.tx_request_body.tx_pdu_list = tx_pdu_list;
   cpu_freq_GHz = (double)get_cpu_freq_GHz();
   cpuf = cpu_freq_GHz;
+  set_parallel_conf("PARALLEL_SINGLE_THREAD");
+
   printf("Detected cpu_freq %f GHz\n",cpu_freq_GHz);
   AssertFatal(load_configmodule(argc,argv,CONFIG_ENABLECMDLINEONLY) != NULL,
               "cannot load configuration module, exiting\n");
@@ -455,7 +433,7 @@ int main(int argc, char **argv) {
     { "Doppler", "Maximum doppler shift",0, dblptr:&maxDoppler,  defdblval:0.0, TYPE_DOUBLE, 0 },
     { "Zdump", "dump table",PARAMFLAG_BOOL,  iptr:&dump_table, defintval:0, TYPE_INT, 0 },
     { "Forms", "Display the soft scope", PARAMFLAG_BOOL, iptr:&xforms,  defintval:0, TYPE_INT, 0 },
-    { "Lparallel", "Enable parallel execution", PARAMFLAG_BOOL, iptr:&parallel_flag,  defintval:0, TYPE_INT, 0 },
+    { "Lparallel", "Enable parallel execution",0, strptr:NULL,  defstrval:NULL, TYPE_STRING,  0 },
     { "Iterations", "Number of iterations of turbo decoder", 0, iptr:&max_turbo_iterations,  defintval:4, TYPE_INT, 0 },
     { "Performance", "Display CPU perfomance of each L1 piece", PARAMFLAG_BOOL,  iptr:NULL,  defintval:0, TYPE_INT, 0 },
     { "Q_cqi", "Enable CQI", PARAMFLAG_BOOL, iptr:&cqi_flag,  defintval:0, TYPE_INT, 0 },
@@ -600,6 +578,10 @@ int main(int argc, char **argv) {
       opp_enabled=1;
       break;
 
+    case 'L':
+      set_parallel_conf(optarg);
+      break;
+      
     default:
       printf("Wrong option: %s\n",long_options[option_index].name);
       exit(1);
@@ -618,7 +600,7 @@ int main(int argc, char **argv) {
   if (help)
     exit(0);
   
-  set_parallel_conf("PARALLEL_RU_L1_TRX_SPLIT");
+  if (thread_struct.parallel_conf != PARALLEL_SINGLE_THREAD)
     set_worker_conf("WORKER_ENABLE");
   RC.nb_L1_inst = 1;
   RC.nb_RU = 1;
@@ -767,7 +749,7 @@ int main(int argc, char **argv) {
   UE->ulsch[0]   = new_ue_ulsch(N_RB_DL,0);
   printf("ULSCH %p\n",UE->ulsch[0]);
 
-  if (parallel_flag == 1) {
+  if(get_thread_worker_conf() == WORKER_ENABLE) {
     extern void init_fep_thread(PHY_VARS_eNB *, pthread_attr_t *);
     extern void init_td_thread(PHY_VARS_eNB *);
     init_fep_thread(eNB,NULL);
@@ -1175,8 +1157,7 @@ int main(int argc, char **argv) {
           }
 
 	  start_meas(&eNB->phy_proc_rx);
-          ru->feprx = (parallel_flag == 1) ? ru_fep_full_2thread        : fep_full;
-          eNB->td  = (parallel_flag == 1) ? ulsch_decoding_data_2thread : ulsch_decoding_data;
+          ru->feprx = (get_thread_worker_conf() == WORKER_ENABLE) ? ru_fep_full_2thread        : fep_full;
           ru->feprx(ru);
           phy_procedures_eNB_uespec_RX(eNB,proc_rxtx);
 	  stop_meas(&eNB->phy_proc_rx);
@@ -1377,20 +1358,18 @@ int main(int argc, char **argv) {
         printStatIndent(&UE->ulsch_rate_matching_stats,"ULSCH rate-matching time");
         printStatIndent(&UE->ulsch_interleaving_stats,"ULSCH sub-block interleaving");
         printStatIndent(&UE->ulsch_multiplexing_stats,"ULSCH multiplexing time");
-        printDistribution(&eNB->phy_proc_rx,table_rx,"\nTotal PHY proc rx subframe");
-        printDistribution(&ru->ofdm_demod_stats,table_rx_fft,"OFDM_demod time");
-        printDistribution(&eNB->ulsch_demodulation_stats,table_rx_demod,"ULSCH demodulation time");
-        printf("ULSCH Decoding time (%.2f Mbit/s, avg iter %.2f)      :%.2f us (%d trials, max %.2f)\n",
-               UE->ulsch[0]->harq_processes[harq_pid]->TBS/1000.0,(double)iter_trials,
-               (double)eNB->ulsch_decoding_stats.diff/eNB->ulsch_decoding_stats.trials*timeBase,
-               eNB->ulsch_decoding_stats.trials,
+	printf("\n");
+        printDistribution(&eNB->phy_proc_rx,table_rx,"Total PHY proc rx subframe");
+        printDistribution(&ru->ofdm_demod_stats,table_rx_fft,"|__ OFDM_demod time");
+        printDistribution(&eNB->ulsch_demodulation_stats,table_rx_demod,"|__ ULSCH demodulation time");
+	printDistribution(&eNB->ulsch_decoding_stats,table_rx_dec,"|__ ULSCH Decoding time");
+        printf("     (%.2f Mbit/s, avg iter %.2f, max %.2f)\n",
+               UE->ulsch[0]->harq_processes[harq_pid]->TBS/1000.0,
+	       (double)iter_trials,
 	       (double)eNB->ulsch_decoding_stats.max*timeBase);
-        printf("|__ Statistics                           std: %.2fus median %.2fus q1 %.2fus q3 %.2fus \n",
-               squareRoot(&eNB->ulsch_decoding_stats),
-               median(table_rx_dec), q1(table_rx_dec), q3(table_rx_dec));
-        printStatIndent(&eNB->ulsch_deinterleaving_stats,"sub-block interleaving" );
-        printStatIndent(&eNB->ulsch_demultiplexing_stats,"sub-block demultiplexing" );
-        printStatIndent(&eNB->ulsch_rate_unmatching_stats,"sub-block rate-matching" );
+        printStatIndent2(&eNB->ulsch_deinterleaving_stats,"sub-block interleaving" );
+        printStatIndent2(&eNB->ulsch_demultiplexing_stats,"sub-block demultiplexing" );
+        printStatIndent2(&eNB->ulsch_rate_unmatching_stats,"sub-block rate-matching" );
         printf("|__ turbo_decoder(%d bits), avg iterations: %.1f       %.2f us (%d cycles, %d trials)\n",
                eNB->ulsch[0]->harq_processes[harq_pid]->Cminus ?
                eNB->ulsch[0]->harq_processes[harq_pid]->Kminus :
@@ -1399,13 +1378,13 @@ int main(int argc, char **argv) {
                (double)eNB->ulsch_turbo_decoding_stats.diff/eNB->ulsch_turbo_decoding_stats.trials*timeBase,
                (int)((double)eNB->ulsch_turbo_decoding_stats.diff/eNB->ulsch_turbo_decoding_stats.trials),
                eNB->ulsch_turbo_decoding_stats.trials);
-        printStatIndent2(&eNB->ulsch_tc_init_stats,"init", eNB->ulsch_tc_init_stats.trials);
-        printStatIndent2(&eNB->ulsch_tc_alpha_stats,"alpha", eNB->ulsch_tc_init_stats.trials);
-        printStatIndent2(&eNB->ulsch_tc_beta_stats,"beta", eNB->ulsch_tc_init_stats.trials);
-        printStatIndent2(&eNB->ulsch_tc_gamma_stats,"gamma", eNB->ulsch_tc_init_stats.trials);
-        printStatIndent2(&eNB->ulsch_tc_ext_stats,"ext", eNB->ulsch_tc_init_stats.trials);
-        printStatIndent2(&eNB->ulsch_tc_intl1_stats,"turbo internal interleaver", eNB->ulsch_tc_init_stats.trials);
-        printStatIndent2(&eNB->ulsch_tc_intl2_stats,"intl2+HardDecode+CRC", eNB->ulsch_tc_init_stats.trials);
+        printStatIndent3(&eNB->ulsch_tc_init_stats,"init");
+        printStatIndent3(&eNB->ulsch_tc_alpha_stats,"alpha");
+        printStatIndent3(&eNB->ulsch_tc_beta_stats,"beta");
+        printStatIndent3(&eNB->ulsch_tc_gamma_stats,"gamma");
+        printStatIndent3(&eNB->ulsch_tc_ext_stats,"ext");
+        printStatIndent3(&eNB->ulsch_tc_intl1_stats,"turbo internal interleaver");
+        printStatIndent3(&eNB->ulsch_tc_intl2_stats,"intl2+HardDecode+CRC");
       }
 
       if(abstx) { //ABSTRACTION

targets/RT/USER/lte-enb.c

@@ -149,8 +149,6 @@ void wakeup_prach_eNB(PHY_VARS_eNB *eNB,RU_t *ru,int frame,int subframe);
 #if (LTE_RRC_VERSION >= MAKE_VERSION(14, 0, 0))
   void wakeup_prach_eNB_br(PHY_VARS_eNB *eNB,RU_t *ru,int frame,int subframe);
 #endif
-extern PARALLEL_CONF_t get_thread_parallel_conf(void);
-extern WORKER_CONF_t   get_thread_worker_conf(void);
 
 extern uint8_t nfapi_mode;
 extern void oai_subframe_ind(uint16_t sfn, uint16_t sf);
@@ -800,9 +798,10 @@ static void *process_stats_thread(void *param) {
     sleep(1);
 
     if (opp_enabled == 1) {
-      if (eNB->td) print_meas(&eNB->ulsch_decoding_stats,"ulsch_decoding",NULL,NULL);
+      if ( eNB->ulsch_decoding_stats.trials>0)
+        print_meas(&eNB->ulsch_decoding_stats,"ulsch_decoding",NULL,NULL);
 
-      if (eNB->te) {
+      if (eNB->dlsch_encoding_stats.trials >0) {
         print_meas(&eNB->dlsch_turbo_encoding_preperation_stats,"dlsch_coding_crc",NULL,NULL);
         print_meas(&eNB->dlsch_turbo_encoding_segmentation_stats,"dlsch_segmentation",NULL,NULL);
         print_meas(&eNB->dlsch_encoding_stats,"dlsch_encoding",NULL,NULL);
@@ -1210,7 +1209,7 @@ void init_eNB(int single_thread_flag,int wait_for_sync) {
   PHY_VARS_eNB *eNB;
   LOG_I(PHY,"[lte-softmodem.c] eNB structure about to allocated RC.nb_L1_inst:%d RC.nb_L1_CC[0]:%d\n",RC.nb_L1_inst,RC.nb_L1_CC[0]);
 
-  if (RC.eNB == NULL) RC.eNB = (PHY_VARS_eNB ***) malloc(RC.nb_L1_inst*sizeof(PHY_VARS_eNB **));
+  if (RC.eNB == NULL) RC.eNB = (PHY_VARS_eNB ** *) malloc(RC.nb_L1_inst*sizeof(PHY_VARS_eNB **));
 
   LOG_I(PHY,"[lte-softmodem.c] eNB structure RC.eNB allocated\n");
 
@@ -1227,8 +1226,6 @@ void init_eNB(int single_thread_flag,int wait_for_sync) {
 #ifndef OCP_FRAMEWORK
       LOG_I(PHY,"Initializing eNB %d CC_id %d\n",inst,CC_id);
 #endif
-      eNB->td                   = ulsch_decoding_data_all;
-      eNB->te                   = dlsch_encoding_all;
       LOG_I(PHY,"Registering with MAC interface module\n");
       AssertFatal((eNB->if_inst         = IF_Module_init(inst))!=NULL,"Cannot register interface");
       eNB->if_inst->schedule_response   = schedule_response;