Merge branch 'develop' into feature-131-new-license

cmake_targets/autotests/run_exec_lte-softmodem_tests.py

@@ -885,7 +885,8 @@ def handle_testcaseclass_softmodem (testcase, oldprogramList, logdirOAI5GRepo ,
     task_EPC  = task_EPC + ' ) > ' + logfile_task_EPC_out + ' 2>&1 ' 
     write_file(logfile_task_EPC, task_EPC, mode="w")
     
-    #first we compile all the programs
+    #first we compile all the programs but only for run_0
+    if run == 0:
        thread_EPC = oaiThread(1, "EPC_thread", EPCMachine, user, password , task_EPC_compile, False, timeout_thread)
        thread_eNB = oaiThread(2, "eNB_thread", eNBMachine, user, password , task_eNB_compile, False, timeout_thread)
        thread_UE = oaiThread(3, "UE_thread", UEMachine, user, password  , task_UE_compile, False, timeout_thread) 

cmake_targets/build_oai

@@ -74,7 +74,7 @@ Options
 --install-optional-packages
    Install useful but not mandatory packages such as valgrind
 -g | --run-with-gdb
-   Add debugging symbols to compilation directives. It also disables any compiler optimization.
+   Add debugging symbols to compilation directives. It also disables any compiler optimization. Only for debugging. Do not use in normal operation!
 -h | --help
    Print this help
 --eNB
@@ -128,9 +128,9 @@ Options
 --disable-hardware-dependency
    Disable HW dependency during installation
 Usage (first build):
- oaisim (eNB + UE): ./build_oai -I -g --oaisim -x --install-system-files
- Eurecom EXMIMO + COTS UE : ./build_oai -I -g --eNB -x --install-system-files
- NI/ETTUS B201  + COTS UE : ./build_oai -I -g --eNB -x --install-system-files -w USRP
+ oaisim (eNB + UE): ./build_oai -I  --oaisim -x --install-system-files
+ Eurecom EXMIMO + COTS UE : ./build_oai -I  --eNB -x --install-system-files
+ NI/ETTUS B201  + COTS UE : ./build_oai -I  --eNB -x --install-system-files -w USRP
 Usage (Regular):
  oaisim : ./build_oai --oaisim -x 
  Eurecom EXMIMO + OAI ENB : ./build_oai --eNB -x 

openair1/PHY/CODING/3gpplte_turbo_decoder_sse_8bit.c

@@ -486,6 +486,11 @@ void compute_beta8(llr_t* alpha,llr_t* beta,llr_t *m_11,llr_t* m_10,unsigned sho
 
 #endif
 
+  if (frame_length > 6143) {
+    LOG_E(PHY,"compute_beta: frame_length %d\n",frame_length);
+    return;
+  }
+
   // we are supposed to run compute_alpha just before compute_beta
   // so the initial states of backward computation can be set from last value of alpha states (forward computation)
 
@@ -505,9 +510,11 @@ void compute_beta8(llr_t* alpha,llr_t* beta,llr_t *m_11,llr_t* m_10,unsigned sho
   beta_ptr[6] = alpha128[6+(frame_length>>1)];
   beta_ptr[7] = alpha128[7+(frame_length>>1)];
 
+  int overlap = (frame_length>>4)> L ? (frame_length>>4)-L : 0 ;
+
   for (rerun_flag=0, loopval=0;
        rerun_flag<2 ;
-       loopval=(frame_length>>4)-L,rerun_flag++) {
+       loopval=overlap,rerun_flag++) {
 
     if (offset8_flag==0) {
       // FIXME! beta0-beta7 are used uninitialized. FIXME!

openair1/PHY/LTE_TRANSPORT/dci.c

@@ -2857,6 +2857,7 @@ void dci_decoding_procedure0(LTE_UE_PDCCH **pdcch_vars,
           }
         }
 
+        LOG_D(PHY,"DCI decoding CRNTI  [format_c: %d, nCCE[subframe: %d]: %d ]\n",format_c, subframe, pdcch_vars[eNB_id]->nCCE[subframe]);
         //  memcpy(&dci_alloc[*dci_cnt].dci_pdu[0],dci_decoded_output,sizeof_bytes);
 
 

openair1/PHY/LTE_TRANSPORT/if5_tools.c

@@ -97,8 +97,8 @@ void send_IF5(PHY_VARS_eNB *eNB, openair0_timestamp proc_timestamp, int subframe
     __m128i t0, t1;
 
     tx_buffer = memalign(16, MAC_HEADER_SIZE_BYTES + sizeof_IF5_mobipass_header_t + db_fulllength*sizeof(int16_t));
-    IF5_mobipass_header_t *header = (IF5_mobipass_header_t *)(tx_buffer + MAC_HEADER_SIZE_BYTES);
-    data_block_head = (__m128i *)(tx_buffer + MAC_HEADER_SIZE_BYTES + sizeof_IF5_mobipass_header_t + 4);
+    IF5_mobipass_header_t *header = (IF5_mobipass_header_t *)((uint8_t *)tx_buffer + MAC_HEADER_SIZE_BYTES);
+    data_block_head = (__m128i *)((uint8_t *)tx_buffer + MAC_HEADER_SIZE_BYTES + sizeof_IF5_mobipass_header_t + 4);
     
     header->flags = 0;
     header->fifo_status = 0;  
@@ -109,11 +109,11 @@ void send_IF5(PHY_VARS_eNB *eNB, openair0_timestamp proc_timestamp, int subframe
     txp[0] = (void*)&eNB->common_vars.txdata[0][0][subframe*eNB->frame_parms.samples_per_tti];
     txp128 = (__m128i *) txp[0];
               
-    for (packet_id=0; packet_id<(fp->samples_per_tti*2)/db_fulllength; packet_id++) {
-      header->time_stamp = proc_timestamp + packet_id*db_fulllength; 
+    for (packet_id=0; packet_id<fp->samples_per_tti/db_fulllength; packet_id++) {
+      header->time_stamp = (uint32_t)(proc_timestamp + packet_id*db_fulllength);
       data_block = data_block_head; 
     
-      for (i=0; i<db_fulllength>>3; i+=2) {
+      for (i=0; i<db_fulllength>>2; i+=2) {
         t0 = _mm_srai_epi16(*txp128++, 4);
         t1 = _mm_srai_epi16(*txp128++, 4);   
         

openair1/PHY/LTE_TRANSPORT/phich.c

@@ -427,9 +427,10 @@ void generate_phich(LTE_DL_FRAME_PARMS *frame_parms,
   if (HI>0)
     HI=1;
 
+
   //  c = (1-(2*HI))*SSS_AMP;
   // x1 is set in lte_gold_generic
-  x2 = (((subframe+1)*(frame_parms->Nid_cell+1))<<9) + frame_parms->Nid_cell;
+  x2 = (((subframe+1)*((frame_parms->Nid_cell<<1)+1))<<9) + frame_parms->Nid_cell;
 
   s = lte_gold_generic(&x1, &x2, reset);
 
@@ -1111,7 +1112,7 @@ void rx_phich(PHY_VARS_UE *ue,
   phich_d_ptr = phich_d;
 
   // x1 is set in lte_gold_generic
-  x2 = (((subframe+1)*(frame_parms->Nid_cell+1))<<9) + frame_parms->Nid_cell;
+  x2 = (((subframe+1)*((frame_parms->Nid_cell<<1)+1))<<9) + frame_parms->Nid_cell;
 
   s = lte_gold_generic(&x1, &x2, reset);
 

openair1/PHY/LTE_TRANSPORT/prach.c

@@ -810,11 +810,14 @@ int32_t generate_prach( PHY_VARS_UE *ue, uint8_t eNB_id, uint8_t subframe, uint1
     break;
 
   case 75:
-    memset((void*)prachF,0,4*19432);
+    memset((void*)prachF,0,4*18432);
     break;
 
   case 100:
+    if (ue->frame_parms.threequarter_fs == 0)
       memset((void*)prachF,0,4*24576);
+    else
+      memset((void*)prachF,0,4*18432);
     break;
   }
 
@@ -878,6 +881,9 @@ int32_t generate_prach( PHY_VARS_UE *ue, uint8_t eNB_id, uint8_t subframe, uint1
     break;
   }
 
+  if (ue->frame_parms.threequarter_fs == 1)
+    Ncp=(Ncp*3)>>2;
+
   prach2 = prach+(Ncp<<1);
 
   // do IDFT
@@ -978,6 +984,7 @@ int32_t generate_prach( PHY_VARS_UE *ue, uint8_t eNB_id, uint8_t subframe, uint1
     break;
 
   case 100:
+    if (ue->frame_parms.threequarter_fs == 0) { 
       if (prach_fmt == 4) {
 	idft4096(prachF,prach2,1);
 	memmove( prach, prach+8192, Ncp<<2 );
@@ -992,6 +999,24 @@ int32_t generate_prach( PHY_VARS_UE *ue, uint8_t eNB_id, uint8_t subframe, uint1
 	  prach_len = 2* 24576+Ncp;
 	}
       }
+    }
+    else {
+      if (prach_fmt == 4) {
+	idft3072(prachF,prach2);
+	//TODO: account for repeated format in dft output
+	memmove( prach, prach+6144, Ncp<<2 );
+	prach_len = 3072+Ncp;
+      } else {
+	idft18432(prachF,prach2);
+	memmove( prach, prach+36864, Ncp<<2 );
+	prach_len = 18432+Ncp;
+	printf("Generated prach for 100 PRB, 3/4 sampling\n");
+	if (prach_fmt>1) {
+	  memmove( prach2+36834, prach2, 73728 );
+	  prach_len = 2*18432+Ncp;
+	}
+      } 
+    }
 
     break;
   }
@@ -1488,7 +1513,7 @@ void rx_prach(PHY_VARS_eNB *eNB,
 
 #ifdef PRACH_DEBUG
 
-      if (en>40) {
+      //      if (en>40) {
 	k = (12*n_ra_prb) - 6*eNB->frame_parms.N_RB_UL;
 	
 	if (k<0)
@@ -1502,7 +1527,7 @@ void rx_prach(PHY_VARS_eNB *eNB,
 	write_output("prach_rxF_comp0.m","prach_rxF_comp0",prachF,1024,1,1);
 	write_output("prach_ifft0.m","prach_t0",prach_ifft[0],1024,1,1);
 	exit(-1);
-      }
+	//      }
 #endif
     } // new dft
     

openair1/PHY/LTE_TRANSPORT/pucch.c

@@ -146,6 +146,9 @@ void generate_pucch1x(int32_t **txdataF,
   uint8_t Ncs1                      = frame_parms->pucch_config_common.nCS_AN;
   uint8_t Ncs1_div_deltaPUCCH_Shift = Ncs1/deltaPUCCH_Shift;
 
+  LOG_D(PHY,"generate_pucch Start [deltaPUCCH_Shift %d, NRB2 %d, Ncs1_div_deltaPUCCH_Shift %d, n1_pucch %d]\n", deltaPUCCH_Shift, NRB2, Ncs1_div_deltaPUCCH_Shift,n1_pucch);
+
+
   uint32_t u0 = (frame_parms->Nid_cell + frame_parms->pusch_config_common.ul_ReferenceSignalsPUSCH.grouphop[subframe<<1]) % 30;
   uint32_t u1 = (frame_parms->Nid_cell + frame_parms->pusch_config_common.ul_ReferenceSignalsPUSCH.grouphop[1+(subframe<<1)]) % 30;
   uint32_t v0=frame_parms->pusch_config_common.ul_ReferenceSignalsPUSCH.seqhop[subframe<<1];
@@ -170,6 +173,11 @@ void generate_pucch1x(int32_t **txdataF,
   printf("[PHY] PUCCH: cNcs1/deltaPUCCH_Shift %d, Nprime %d, n1_pucch %d\n",thres,Nprime,n1_pucch);
 #endif
 
+  LOG_D(PHY,"[PHY] PUCCH: n1_pucch %d, thres %d Ncs1_div_deltaPUCCH_Shift %d (12/deltaPUCCH_Shift) %d Nprime_div_deltaPUCCH_Shift %d \n",
+		              n1_pucch, thres, Ncs1_div_deltaPUCCH_Shift, (int)(12/deltaPUCCH_Shift), Nprime_div_deltaPUCCH_Shift);
+  LOG_D(PHY,"[PHY] PUCCH: deltaPUCCH_Shift %d, Nprime %d\n",deltaPUCCH_Shift,Nprime);
+
+
   N_UL_symb = (frame_parms->Ncp==0) ? 7 : 6;
 
   if (n1_pucch < thres)
@@ -182,6 +190,9 @@ void generate_pucch1x(int32_t **txdataF,
   else {
     d = (frame_parms->Ncp==0) ? 2 : 0;
     h= (nprime0+d)%(c*Nprime_div_deltaPUCCH_Shift);
+#ifdef DEBUG_PUCCH_TX
+    printf("[PHY] PUCCH: h %d, d %d\n",h,d);
+#endif
     nprime1 = (h/c) + (h%c)*Nprime_div_deltaPUCCH_Shift;
   }
 
@@ -200,7 +211,7 @@ void generate_pucch1x(int32_t **txdataF,
     n_oc1<<=1;
 
 #ifdef DEBUG_PUCCH_TX
-  printf("[PHY] PUCCH: noc0 %d noc11 %d\n",n_oc0,n_oc1);
+  printf("[PHY] PUCCH: noc0 %d noc1 %d\n",n_oc0,n_oc1);
 #endif
 
   nprime=nprime0;

openair1/PHY/LTE_TRANSPORT/ulsch_decoding.c

@@ -1192,6 +1192,7 @@ unsigned int  ulsch_decoding(PHY_VARS_eNB *eNB,eNB_rxtx_proc_t *proc,
         y[i2] = c*ulsch_llr[i++];
         i2=(i2+(Cmux<<2)-3);
 	*/
+	// slightly more optimized version (equivalent to above) for 16QAM to improve computational performance
 	*(__m64 *)&y[i2] = _mm_sign_pi16(*(__m64*)&ulsch_llr[i],*(__m64*)&cseq[i]);i+=4;i2+=(Cmux<<2);
 
 
@@ -1523,6 +1524,7 @@ unsigned int  ulsch_decoding(PHY_VARS_eNB *eNB,eNB_rxtx_proc_t *proc,
 
   // RI
 
+  // rank 1
   if ((ulsch_harq->O_RI == 1) && (Qprime_RI > 0)) {
     ulsch_harq->o_RI[0] = ((ulsch_harq->q_RI[0] + ulsch_harq->q_RI[Q_m/2]) > 0) ? 0 : 1;
   }

openair1/PHY/MODULATION/slot_fep_mbsfn.c

@@ -108,10 +108,8 @@ int slot_fep_mbsfn(PHY_VARS_UE *ue,
       sample_offset);
 #endif
 
-
   for (aa=0; aa<frame_parms->nb_antennas_rx; aa++) {
-    memset(&common_vars->rxdataF[aa][2*frame_parms->ofdm_symbol_size*l],0,2*frame_parms->ofdm_symbol_size*sizeof(int));
-
+    memset(&common_vars->rxdataF[aa][frame_parms->ofdm_symbol_size*l],0,frame_parms->ofdm_symbol_size*sizeof(int));
     if (l==0) {
       start_meas(&ue->rx_dft_stats);
       dft((int16_t *)&common_vars->rxdata[aa][(sample_offset +
@@ -139,9 +137,10 @@ int slot_fep_mbsfn(PHY_VARS_UE *ue,
           (int16_t *)&common_vars->rxdataF[aa][frame_parms->ofdm_symbol_size*l],1);
       stop_meas(&ue->rx_dft_stats);
     }
-
   }
 
+
+
   //if ((l==0) || (l==(4-frame_parms->Ncp))) {
   // changed to invoke MBSFN channel estimation in symbols 2,6,10
   if ((l==2)||(l==6)||(l==10)) {

openair1/PHY/TOOLS/lte_dfts.c

@@ -5222,7 +5222,7 @@ void dft8192(int16_t *x,int16_t *y,int scale)
   
   xtmpp = xtmp;
-  for (i=0; i<32; i++) {
+  for (i=0; i<16; i++) {
     transpose4_ooff_simd256(x256  ,xtmpp,512);
     transpose4_ooff_simd256(x256+2,xtmpp+1,512);
     transpose4_ooff_simd256(x256+4,xtmpp+2,512);
@@ -5260,7 +5260,7 @@ void dft8192(int16_t *x,int16_t *y,int scale)
   }
   dft4096((int16_t*)(xtmp),(int16_t*)ytmp,1);
-  dft4096((int16_t*)(xtmp+1024),(int16_t*)(ytmp+512),1);
+  dft4096((int16_t*)(xtmp+512),(int16_t*)(ytmp+512),1);
   for (i=0; i<512; i++) {
@@ -5312,7 +5312,7 @@ void idft8192(int16_t *x,int16_t *y,int scale)
   
   xtmpp = xtmp;
-  for (i=0; i<32; i++) {
+  for (i=0; i<16; i++) {
     transpose4_ooff_simd256(x256  ,xtmpp,512);
     transpose4_ooff_simd256(x256+2,xtmpp+1,512);
     transpose4_ooff_simd256(x256+4,xtmpp+2,512);
@@ -5350,7 +5350,7 @@ void idft8192(int16_t *x,int16_t *y,int scale)
   }
   idft4096((int16_t*)(xtmp),(int16_t*)ytmp,1);
-  idft4096((int16_t*)(xtmp+1024),(int16_t*)(ytmp+512),1);
+  idft4096((int16_t*)(xtmp+512),(int16_t*)(ytmp+512),1);
   for (i=0; i<512; i++) {
@@ -5673,15 +5673,58 @@ void idft12288(int16_t *input, int16_t *output)
   //  write_output("out.m","out",output,6144,1,1);
 }
+#include "twiddle18432.h"
 // 6144 x 3
-void dft18432(int16_t *input, int16_t *output)
-{
+void dft18432(int16_t *input, int16_t *output) {
+  int i,i2,j;
+  uint32_t tmp[3][6144] __attribute__((aligned(32)));
+  uint32_t tmpo[3][6144] __attribute__((aligned(32)));
+
+  for (i=0,j=0; i<6144; i++) {
+    tmp[0][i] = ((uint32_t *)input)[j++];
+    tmp[1][i] = ((uint32_t *)input)[j++];
+    tmp[2][i] = ((uint32_t *)input)[j++];
+  }
+
+  dft6144((int16_t*)(tmp[0]),(int16_t*)(tmpo[0]));
+  dft6144((int16_t*)(tmp[1]),(int16_t*)(tmpo[1]));
+  dft6144((int16_t*)(tmp[2]),(int16_t*)(tmpo[2]));
+
+  for (i=0,i2=0; i<12288; i+=8,i2+=4)  {
+    bfly3((simd_q15_t*)(&tmpo[0][i2]),(simd_q15_t*)(&tmpo[1][i2]),(simd_q15_t*)(&tmpo[2][i2]),
+          (simd_q15_t*)(output+i),(simd_q15_t*)(output+12288+i),(simd_q15_t*)(output+24576+i),
+          (simd_q15_t*)(twa18432+i),(simd_q15_t*)(twb18432+i));
+  }
+
+  _mm_empty();
+  _m_empty();
 }
-void idft18432(int16_t *input, int16_t *output)
-{
+void idft18432(int16_t *input, int16_t *output) {
+  int i,i2,j;
+  uint32_t tmp[3][6144] __attribute__((aligned(32)));
+  uint32_t tmpo[3][6144] __attribute__((aligned(32)));
+
+  for (i=0,j=0; i<6144; i++) {
+    tmp[0][i] = ((uint32_t *)input)[j++];
+    tmp[1][i] = ((uint32_t *)input)[j++];
+    tmp[2][i] = ((uint32_t *)input)[j++];
+  }
+
+  idft6144((int16_t*)(tmp[0]),(int16_t*)(tmpo[0]));
+  idft6144((int16_t*)(tmp[1]),(int16_t*)(tmpo[1]));
+  idft6144((int16_t*)(tmp[2]),(int16_t*)(tmpo[2]));
+
+  for (i=0,i2=0; i<12288; i+=8,i2+=4)  {
+    ibfly3((simd_q15_t*)(&tmpo[0][i2]),(simd_q15_t*)(&tmpo[1][i2]),(simd_q15_t*)(&tmpo[2][i2]),
+	   (simd_q15_t*)(output+i),(simd_q15_t*)(output+12288+i),(simd_q15_t*)(output+24576+i),
+	   (simd_q15_t*)(twa18432+i),(simd_q15_t*)(twb18432+i));
+  }
+
+  _mm_empty();
+  _m_empty();
 }
 #include "twiddle24576.h"
@@ -5726,8 +5769,8 @@ void dft24576(int16_t *input, int16_t *output)
 void idft24576(int16_t *input, int16_t *output)
 {
   int i,i2,j;
-  uint32_t tmp[3][16384] __attribute__((aligned(32)));
-  uint32_t tmpo[3][16384] __attribute__((aligned(32)));
+  uint32_t tmp[3][8192] __attribute__((aligned(32)));
+  uint32_t tmpo[3][8192] __attribute__((aligned(32)));
   for (i=0,j=0; i<8192; i++) {
     tmp[0][i] = ((uint32_t *)input)[j++];
@@ -5739,13 +5782,6 @@ void idft24576(int16_t *input, int16_t *output)
   idft8192((int16_t*)(tmp[1]),(int16_t*)(tmpo[1]),1);
   idft8192((int16_t*)(tmp[2]),(int16_t*)(tmpo[2]),1);
   
-  /*
-  for (i=1; i<8192; i++) {
-    tmpo[0][i] = tmpo[0][i<<1];
-    tmpo[1][i] = tmpo[1][i<<1];
-    tmpo[2][i] = tmpo[2][i<<1];
-    }*/
-
   /*
     write_output("in.m","in",input,24576,1,1);
     write_output("out0.m","o0",tmpo[0],8192,1,1);
@@ -19122,6 +19158,55 @@ int main(int argc, char**argv)
   write_output("y2048.m","y2048",y,2048,1,1);
   write_output("x2048.m","x2048",x,2048,1,1);
+  memset((void*)x,0,2048*sizeof(int32_t));
+  for (i=2;i<2402;i++) {
+    if ((taus() & 1)==0)
+      ((int16_t*)x)[i] = 364;
+    else
+      ((int16_t*)x)[i] = -364;
+  }
+  for (i=2*(4096-1200);i<8192;i++) {
+    if ((taus() & 1)==0)
+      ((int16_t*)x)[i] = 364;
+    else
+      ((int16_t*)x)[i] = -364;
+  }
+  reset_meas(&ts);
+
+  for (i=0; i<10000; i++) {
+    start_meas(&ts);
+    idft4096((int16_t *)x,(int16_t *)y,1);
+    stop_meas(&ts);
+  }
+
+  printf("\n\n4096-point(%f cycles)\n",(double)ts.diff/(double)ts.trials);
+  write_output("y4096.m","y4096",y,4096,1,1);
+  write_output("x4096.m","x4096",x,4096,1,1);
+
+  memset((void*)x,0,8192*sizeof(int32_t));
+  for (i=2;i<4802;i++) {
+    if ((taus() & 1)==0)
+      ((int16_t*)x)[i] = 364;
+    else
+      ((int16_t*)x)[i] = -364;
+  }
+  for (i=2*(4096-1200);i<8192;i++) {
+    if ((taus() & 1)==0)
+      ((int16_t*)x)[i] = 364;
+    else
+      ((int16_t*)x)[i] = -364;
+  }
+  reset_meas(&ts);
+  for (i=0; i<10000; i++) {
+    start_meas(&ts);
+    idft8192((int16_t *)x,(int16_t *)y,1);
+    stop_meas(&ts);
+  }
+
+  printf("\n\n8192-point(%f cycles)\n",(double)ts.diff/(double)ts.trials);
+  write_output("y8192.m","y8192",y,8192,1,1);
+  write_output("x8192.m","x8192",x,8192,1,1);
+
   return(0);
 }