From 556346dd1987b23abe676c1620ee841c4e8d4ee3 Mon Sep 17 00:00:00 2001
From: Raymond Knopp <raymond.knopp@eurecom.fr>
Date: Tue, 27 May 2014 13:34:26 +0000
Subject: [PATCH] git-svn-id: http://svn.eurecom.fr/openair4G/trunk@5351
 818b1a75-f10b-46b9-bf7c-635c3b92a50f

---
 openair1/PHY/CODING/TESTBENCH/Makefile        |   5 +-
 openair1/PHY/CODING/TESTBENCH/ltetest.c       | 129 ++++++----------
 openair1/PHY/CODING/lte_rate_matching.c       |   5 +-
 openair1/PHY/LTE_REFSIG/defs.h                |  23 +++
 openair1/PHY/LTE_REFSIG/lte_dl_cell_spec.c    |  68 +--------
 openair1/PHY/LTE_REFSIG/lte_gold.c            |  46 ++++++
 openair1/PHY/LTE_TRANSPORT/dci.c              |   8 +-
 openair1/PHY/LTE_TRANSPORT/defs.h             |   7 +
 openair1/PHY/LTE_TRANSPORT/dlsch_coding.c     |   4 +-
 openair1/PHY/LTE_TRANSPORT/dlsch_decoding.c   |   2 +-
 .../PHY/LTE_TRANSPORT/dlsch_demodulation.c    |  42 +++---
 .../PHY/LTE_TRANSPORT/dlsch_llr_computation.c |   7 +-
 openair1/PHY/LTE_TRANSPORT/dlsch_modulation.c | 142 ++++++++++++++----
 openair1/PHY/LTE_TRANSPORT/initial_sync.c     |  16 +-
 openair1/PHY/LTE_TRANSPORT/lte_mcs.c          |  16 +-
 openair1/PHY/LTE_TRANSPORT/pbch.c             |   2 +
 openair1/PHY/LTE_TRANSPORT/pmch.c             |   2 +-
 openair1/PHY/LTE_TRANSPORT/proto.h            |  40 ++---
 openair1/PHY/LTE_TRANSPORT/pss.c              |   2 +-
 openair1/PHY/LTE_TRANSPORT/sss.c              |  12 +-
 openair1/PHY/LTE_TRANSPORT/ulsch_coding.c     |   2 +-
 openair1/PHY/LTE_TRANSPORT/ulsch_decoding.c   |   2 +-
 .../PHY/LTE_TRANSPORT/ulsch_demodulation.c    |  68 +--------
 openair1/PHY/Makefile.inc                     |   1 +
 openair1/PHY/TOOLS/Makefile                   |   2 +-
 openair1/PHY/TOOLS/lte_dfts.c                 | 105 +++++++------
 openair1/PHY/defs.h                           |  58 +++----
 openair1/PHY/impl_defs_lte.h                  |   4 +-
 openair1/SCHED/phy_procedures_lte_eNb.c       |   4 +
 openair1/SCHED/phy_procedures_lte_ue.c        |  13 +-
 openair1/SIMULATION/LTE_PHY/dlsim.c           |   6 +-
 openair1/SIMULATION/LTE_PHY/mbmssim.c         |   2 +-
 32 files changed, 449 insertions(+), 396 deletions(-)

diff --git a/openair1/PHY/CODING/TESTBENCH/Makefile b/openair1/PHY/CODING/TESTBENCH/Makefile
index 0e6aa83348..ac9905ed23 100644
--- a/openair1/PHY/CODING/TESTBENCH/Makefile
+++ b/openair1/PHY/CODING/TESTBENCH/Makefile
@@ -6,7 +6,7 @@ CPUFLAGS = -mmmx -msse -msse2 -mssse3 -msse4
 #CPUFLAGS += $(shell if [ -z $(SSE3PROC)]; then echo "" ; else echo "-mssse3"; fi)
 #CPUFLAGS += $(shell if [ -z $(SSE4PROC)]; then echo "" ; else echo "-msse4"; fi)
 
-
+COMMON_UTILS_DIR = $(OPENAIR_HOME)/common/utils
 TOP_DIR = $(OPENAIR1_DIR)
 OPENAIR1_TOP = $(OPENAIR1_DIR)
 OPENAIR2_TOP = $(OPENAIR2_DIR)
@@ -40,6 +40,7 @@ ifeq "$(GCCVERSION)" "4.6.1"
     CFLAGS += -Wno-packed-bitfield-compat
 endif
 
+include $(COMMON_UTILS_DIR)/Makefile.inc
 include $(TOP_DIR)/PHY/Makefile.inc
 #include $(TOP_DIR)/SCHED/Makefile.inc
 SCHED_OBJS = $(TOP_DIR)/SCHED/phy_procedures_lte_common.o
@@ -53,7 +54,7 @@ LAYER2_OBJ = $(OPENAIR2_DIR)/LAYER2/MAC/lte_transport_init.o
 
 OBJ = $(PHY_OBJS) $(SIMULATION_OBJS) $(TOOLS_OBJS) $(SCHED_OBJS) $(LAYER2_OBJ) $(LOG_OBJS) #$(ASN1_MSG_OBJS) 
 
-CFLAGS += -I$(TOP_DIR) $(L2_incl) -I$(ASN1_MSG_INC) -I$(OPENAIR3) $(UTIL_incl)
+CFLAGS += -I$(TOP_DIR) $(L2_incl) -I$(ASN1_MSG_INC) -I$(OPENAIR3) $(UTIL_incl) $(UTILS_incl)
 CFLAGS += -DNO_RRM
 
 all: ltetest
diff --git a/openair1/PHY/CODING/TESTBENCH/ltetest.c b/openair1/PHY/CODING/TESTBENCH/ltetest.c
index 4d3e88d81b..18a6334b96 100644
--- a/openair1/PHY/CODING/TESTBENCH/ltetest.c
+++ b/openair1/PHY/CODING/TESTBENCH/ltetest.c
@@ -143,8 +143,6 @@ char quantize(double D,double x,unsigned char B) {
 }
 
 #define MAX_BLOCK_LENGTH 6000
-static char channel_output[2*MAX_BLOCK_LENGTH]__attribute__ ((aligned(16)));
-static unsigned char decoded_output[MAX_BLOCK_LENGTH/8];
 
 int test_logmap8(LTE_eNB_DLSCH_t *dlsch_eNB,
 		 LTE_UE_DLSCH_t *dlsch_ue,
@@ -163,8 +161,6 @@ int test_logmap8(LTE_eNB_DLSCH_t *dlsch_eNB,
 		 unsigned int subframe) {
 
   unsigned char test_input[block_length+1];
-  //_declspec(align(16))  char channel_output[512];
-  //_declspec(align(16))  unsigned char output[512],decoded_output[16], *inPtr, *outPtr;
 
   short *channel_output;
 
@@ -186,7 +182,7 @@ int test_logmap8(LTE_eNB_DLSCH_t *dlsch_eNB,
 
 
 
-  //  printf("dlsch_eNB->TBS= %d\n",dlsch_eNB->harq_processes[0]->TBS);
+  //  printf("dlsch_eNB->TBS= %d, block_length %d\n",dlsch_eNB->harq_processes[0]->TBS,block_length);
 
   while (trial++ < ntrials) {
 
@@ -201,6 +197,7 @@ int test_logmap8(LTE_eNB_DLSCH_t *dlsch_eNB,
 		   &PHY_vars_eNB->lte_frame_parms,
 		   num_pdcch_symbols,
 		   PHY_vars_eNB->dlsch_eNB[0][0],
+		   0,
 		   subframe,
 		   &PHY_vars_eNB->dlsch_rate_matching_stats,
 		   &PHY_vars_eNB->dlsch_turbo_encoding_stats,
@@ -216,6 +213,8 @@ int test_logmap8(LTE_eNB_DLSCH_t *dlsch_eNB,
       printf("%d.",PHY_vars_eNB->dlsch_eNB[0][0]->e[i]);
 #endif
       channel_output[i] = (short)quantize(sigma/4.0,(2.0*PHY_vars_eNB->dlsch_eNB[0][0]->e[i]) - 1.0 + sigma*gaussdouble(0.0,1.0),qbits);
+      //            printf("input %d, output %f\n",(2*PHY_vars_eNB->dlsch_eNB[0][0]->e[i]) - 1,
+      //      	     (2.0*PHY_vars_eNB->dlsch_eNB[0][0]->e[i]) - 1.0 + sigma*gaussdouble(0.0,1.0));
     }
 #ifdef DEBUG_CODER
     printf("\n");
@@ -223,13 +222,15 @@ int test_logmap8(LTE_eNB_DLSCH_t *dlsch_eNB,
 #endif
   
     
-    //    memset(decoded_output,0,16);
-    //    printf("decoding %d\n",trial);
+
+    PHY_vars_UE->dlsch_ue[0][0]->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->G = coded_bits;
     ret = dlsch_decoding(PHY_vars_UE,
 			 channel_output,
 			 &PHY_vars_UE->lte_frame_parms,
 			 PHY_vars_UE->dlsch_ue[0][0],
+			 PHY_vars_UE->dlsch_ue[0][0]->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid],
 			 subframe,
+			 PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid,
 			 num_pdcch_symbols,1);
 
     /*    int diffs = 0,puncts=0;
@@ -246,15 +247,15 @@ int test_logmap8(LTE_eNB_DLSCH_t *dlsch_eNB,
 	printf("%d same (%d,%d)\n",i,dlsch_ue->harq_processes[0]->d[0][96+i],dlsch_eNb->harq_processes[0]->d[0][96+i]);
     }
     printf("diffs %d puncts %d(%d,%d,%d,%d,%d)\n",diffs,puncts,dlsch_ue->harq_processes[0]->F,coded_bits,3*(block_length<<3),3*dlsch_ue->harq_processes[0]->Kplus,3*dlsch_ue->harq_processes[0]->F+3*(block_length<<3)-coded_bits);
+    
+        
+    printf("ret %d (max %d)\n",ret,dlsch_ue->max_turbo_iterations);
+        printf("trial %d : i %d/%d : Input %x, Output %x (%x, F %d)\n",trial,0,block_length,test_input[0],
+    	   dlsch_ue->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->b[0],
+    	   dlsch_ue->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->c[0][0],
+    	   (dlsch_ue->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->F>>3));
     */
-    
-    //    printf("ret %d\n",ret);
-    //    printf("trial %d : i %d/%d : Input %x, Output %x (%x, F %d)\n",trial,0,block_length,test_input[0],
-    //	   dlsch_ue->harq_processes[0]->b[0],
-    //	   dlsch_ue->harq_processes[0]->c[0][0],
-    //	   (dlsch_ue->harq_processes[0]->F>>3));
-    
-    if (ret < MAX_TURBO_ITERATIONS+1) {
+    if (ret < dlsch_ue->max_turbo_iterations+1) {
       *iterations = (*iterations) + ret;
       //      if (ret>1)
       //	printf("ret %d\n",ret);
@@ -267,20 +268,21 @@ int test_logmap8(LTE_eNB_DLSCH_t *dlsch_eNB,
     
     for (i=0;i<block_length;i++) {
             
-      if (dlsch_ue->harq_processes[0]->b[i] != test_input[i]) {
-/*		
+      if (dlsch_ue->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->c[0][i] != test_input[i]) {
+	/*			
 		printf("i %d/%d : Input %x, Output %x (%x, F %d)\n",i,block_length,test_input[i],
-		       dlsch_ue->harq_processes[0]->b[i],
-		       dlsch_ue->harq_processes[0]->c[0][i],
-		       (dlsch_ue->harq_processes[0]->F>>3));
-*/	
+		       dlsch_ue->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->b[i],
+		       dlsch_ue->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->c[0][i],
+		       (dlsch_ue->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->F>>3));
+	*/
 	*errors = (*errors) + 1;
 	//	printf("*%d, ret %d\n",*errors,ret);	
 
 	
 	
-	if (ret < MAX_TURBO_ITERATIONS+1)
+	if (ret < dlsch_ue->max_turbo_iterations+1)
 	  *crc_misses = (*crc_misses)+1;
+
 	break;
 
       }
@@ -288,10 +290,12 @@ int test_logmap8(LTE_eNB_DLSCH_t *dlsch_eNB,
     
     }
     
-    if (ret == MAX_TURBO_ITERATIONS+1) {
+    if (ret == dlsch_ue->max_turbo_iterations+1) {
       //      exit(-1);
     }
-      /*
+    /*
+    else {  
+    
       for (i=0;i<block_length;i++) {
 	
 	if (dlsch_ue->harq_processes[0]->b[i] != test_input[i]) {
@@ -303,10 +307,10 @@ int test_logmap8(LTE_eNB_DLSCH_t *dlsch_eNB,
 	}
 	
       }
-    }
-    */
+      }*/
+    
     if (*errors == 100) {
-      //printf("\n");
+      printf("trials %d\n",trial);
       break;
     }
   }
@@ -316,7 +320,7 @@ int test_logmap8(LTE_eNB_DLSCH_t *dlsch_eNB,
   return(0);
 }
 
-#define NTRIALS 1000
+#define NTRIALS 10000
 #define DLSCH_RB_ALLOC 0x1fff//0x1fbf // igore DC component,RB13
 
 int main(int argc, char *argv[]) {
@@ -340,10 +344,10 @@ int main(int argc, char *argv[]) {
   logInit();
   lte_param_init(1,1,1,0,0,3);
 
-  PHY_vars_eNB->dlsch_eNB[0][0] = new_eNB_dlsch(1,8,0);
-  PHY_vars_UE->dlsch_ue[0][0]  = new_ue_dlsch(1,8,0);
-  PHY_vars_eNB->dlsch_eNB[0][1] = new_eNB_dlsch(1,8,0);
-  PHY_vars_UE->dlsch_ue[0][1]  = new_ue_dlsch(1,8,0);
+  PHY_vars_eNB->dlsch_eNB[0][0] = new_eNB_dlsch(1,8,NB_RB,0);
+  PHY_vars_UE->dlsch_ue[0][0]  = new_ue_dlsch(1,8,4,NB_RB,0);
+  PHY_vars_eNB->dlsch_eNB[0][1] = new_eNB_dlsch(1,8,NB_RB,0);
+  PHY_vars_UE->dlsch_ue[0][1]  = new_ue_dlsch(1,8,4,NB_RB,0);
 
   if (argc>1)
     mcs = atoi(argv[1]);
@@ -390,8 +394,14 @@ int main(int argc, char *argv[]) {
 				    0,
 				    P_RNTI);
   
-  coded_bits = 	get_G(&PHY_vars_eNB->lte_frame_parms,NB_RB,PHY_vars_eNB->dlsch_eNB[0][0]->rb_alloc,
-		      get_Qm(mcs),num_pdcch_symbols,subframe);
+  coded_bits = 	get_G(&PHY_vars_eNB->lte_frame_parms,
+		      PHY_vars_eNB->dlsch_eNB[0][0]->nb_rb,
+		      PHY_vars_eNB->dlsch_eNB[0][0]->rb_alloc,
+		      get_Qm(mcs),
+		      1,
+		      num_pdcch_symbols,
+		      0,
+		      subframe);
 
   printf("Coded_bits (G) = %d\n",coded_bits);
 
@@ -400,22 +410,13 @@ int main(int argc, char *argv[]) {
 	 dlsch_tbs25[get_I_TBS(mcs)][NB_RB-1],(double)dlsch_tbs25[get_I_TBS(mcs)][NB_RB-1]/coded_bits,
 	 mcs,get_I_TBS(mcs),PHY_vars_eNB->dlsch_eNB[0][0]->harq_processes[0]->F,NB_RB);
 
-  // Test Openair0 3GPP encoder
-/*
-  test_encoder(block_length,
-	       f1f2mat[(block_length-5)*2],   // f1 (see 36121-820, page 14)
-	       f1f2mat[((block_length-5)*2)+1],  // f2 (see 36121-820, page 14)
-	       3);
- */ //  exit(0);
-
-
-
   for (SNR=-5;SNR<5;SNR+=.1) {
 
 
-    //    printf("\n\nSNR %f dB\n",SNR);
+    
 
     sigma = pow(10.0,-.05*SNR);
+    printf("\n\nSNR %f dB => sigma %f\n",SNR,sigma);
 
     errors=0;
     crc_misses=0;
@@ -445,49 +446,11 @@ int main(int argc, char *argv[]) {
 		       subframe);
 
     if (ret>=0)
-      //      printf("ref: Errors %d (%f), Uerrors %d (%f), CRC Misses %d (%f), Avg iterations %f\n",errors,(double)errors/trials,uerrors,(double)uerrors/trials,crc_misses,(double)crc_misses/trials,(double)iterations/trials);
       printf("%f,%f,%f,%f\n",SNR,(double)errors/trials,(double)crc_misses/trials,(double)iterations/trials);
     if (((double)errors/trials) < 1e-2)
       done0=1;
     } 
-    /*    
-    if (done1 == 0) { 
 
-      printf("exmimo\n");
-      ret = test_logmapexmimo(rate,    // code rate
-			      sigma,   // noise standard deviation
-			      qbits,
-			      block_length,   // block length bytes
-			      f1f2mat[iind*2],   // f1 (see 36121-820, page 14)
-			      f1f2mat[(iind*2)+1],  // f2 (see 36121-820, page 14)
-			      3,
-			      NTRIALS,
-			      &errors3,
-			      &trials3);
-
-      if (ret>=0)
-	printf("exmimo : Errors %d (%f)\n",errors3,(double)errors3/trials3);
-      if (((double)errors3/trials3) < 1e-3)
-	done1=1;
-    }
-    
-
-    if (done2 == 0) {  
-    
-      printf("Viterbi ...\n");
-      ret2 = test_viterbi(sigma,
-			  8*block_length,
-			  NTRIALS,
-			  &errors2,
-			  &trials2,
-			  rate);
-      
-      if (ret2>=0)
-	printf("viterbi : Errors %d (%f)\n",errors2,(double)errors2/trials2);
-      if (((double)errors2/trials2) < 1e-3)
-	done2=1;
-    } 
-    */
     if ((done0==1) && (done1==1) && (done2==1)) {
       printf("done\n");
       break;
diff --git a/openair1/PHY/CODING/lte_rate_matching.c b/openair1/PHY/CODING/lte_rate_matching.c
index 123af99b70..73bebe557a 100644
--- a/openair1/PHY/CODING/lte_rate_matching.c
+++ b/openair1/PHY/CODING/lte_rate_matching.c
@@ -691,8 +691,7 @@ int lte_rate_matching_turbo_rx(uint32_t RTC,
   if (clear==1)
     memset(w,0,Ncb*sizeof(int16_t));
  
-  //soft_input2 = soft_input + (r*E);
-  soft_input2 = soft_input; // navid
+  soft_input2 = soft_input;// + (r*E);
   k=0;
 
   for (;(ind<Ncb)&&(k<E);ind++) {
@@ -719,7 +718,7 @@ int lte_rate_matching_turbo_rx(uint32_t RTC,
       else {
 	printf("RM_RX Ind: %d NULL %d\n",ind,nulled);
 	nulled++;
-      }
+      } 
 #endif
     }
   }
diff --git a/openair1/PHY/LTE_REFSIG/defs.h b/openair1/PHY/LTE_REFSIG/defs.h
index bfdec19011..b24aeea85e 100644
--- a/openair1/PHY/LTE_REFSIG/defs.h
+++ b/openair1/PHY/LTE_REFSIG/defs.h
@@ -24,6 +24,13 @@ unsigned int lte_gold_generic(unsigned int *x1, unsigned int *x2, unsigned char
 
 void lte_gold(LTE_DL_FRAME_PARMS *frame_parms,uint32_t lte_gold_table[20][2][14],uint16_t Nid_cell);
 
+/*!\brief This function generates the LTE Gold sequence (36-211, Sec 7.2), specifically for DL UE-specific reference signals for antenna ports 7..14.
+@param frame_parms LTE DL Frame parameters
+@param lte_gold_uespec_table pointer to table where sequences are stored
+@param Nid_cell Cell Id (to compute sequences for local and adjacent cells) 
+@param n_idDMRS Scrambling identity for TM10*/
+void lte_gold_ue_spec(LTE_DL_FRAME_PARMS *frame_parms,uint32_t lte_gold_uespec_table[2][20][2][21],uint16_t Nid_cell, uint16_t *n_idDMRS);
+
 void lte_gold_mbsfn(LTE_DL_FRAME_PARMS *frame_parms,uint32_t lte_gold_mbsfn_table[10][3][42],uint16_t Nid_MBSFN);
 
 
@@ -44,6 +51,22 @@ int lte_dl_cell_spec(PHY_VARS_eNB *phy_vars_eNB,
 		     unsigned char l,
 		     unsigned char p);
 
+/*! \brief This function generates the UE-specific reference signal sequence (36-211, Sec 6.10.3.2)
+@param phy_vars_eNB Pointer to eNB variables
+@param output Output vector for OFDM symbol (Frequency Domain)
+@param amp Q15 amplitude
+@param Ns Slot number (0..19)
+@param p antenna index
+@param SS_flag Flag to indicate special subframe
+*/
+int lte_dl_ue_spec(PHY_VARS_eNB *phy_vars_eNB,
+		   uint8_t UE_id,
+		   mod_sym_t *output,
+		   short amp,
+		   uint8_t Ns,
+		   uint8_t p,
+		   int SS_flag );
+
 /*! \brief This function generates the MBSFN reference signal sequence (36-211, Sec 6.10.1.2)
 @param phy_vars_eNB Pointer to eNB variables
 @param output Output vector for OFDM symbol (Frequency Domain)
diff --git a/openair1/PHY/LTE_REFSIG/lte_dl_cell_spec.c b/openair1/PHY/LTE_REFSIG/lte_dl_cell_spec.c
index 4253c8e522..624dd8222b 100644
--- a/openair1/PHY/LTE_REFSIG/lte_dl_cell_spec.c
+++ b/openair1/PHY/LTE_REFSIG/lte_dl_cell_spec.c
@@ -23,36 +23,16 @@ int lte_dl_cell_spec_SS(PHY_VARS_eNB *phy_vars_eNB,
   unsigned short k,a;
   mod_sym_t qpsk[4];
 
-#ifdef IFFT_FPGA
-  // new mod table
-  qpsk[0] = 1;
-  qpsk[1] = 3;
-  qpsk[2] = 2;
-  qpsk[3] = 4;
-  /* old mod table
-  qpsk[0] = 4;
-  qpsk[1] = 2;
-  qpsk[2] = 3;
-  qpsk[3] = 1;
-  */
-#else
   a = (amp*ONE_OVER_SQRT2_Q15)>>15;
   ((short *)&qpsk[0])[0] = a;
   ((short *)&qpsk[0])[1] = a;
-  // new mod table
   ((short *)&qpsk[1])[0] = -a;
   ((short *)&qpsk[1])[1] = a;
   ((short *)&qpsk[2])[0] = a;
   ((short *)&qpsk[2])[1] = -a;
-  /* old mod table
-  ((short *)&qpsk[1])[0] = a;
-  ((short *)&qpsk[1])[1] = -a;
-  ((short *)&qpsk[2])[0] = -a;
-  ((short *)&qpsk[2])[1] = a;
-  */
   ((short *)&qpsk[3])[0] = -a;
   ((short *)&qpsk[3])[1] = -a;
-#endif
+
 
   if ((p==0) && (l==0) )
     nu = 0;
@@ -73,12 +53,9 @@ int lte_dl_cell_spec_SS(PHY_VARS_eNB *phy_vars_eNB,
   if (k > 6)//b
     k -=6;//b
 
-#ifdef IFFT_FPGA
-  k+=phy_vars_eNB->lte_frame_parms.N_RB_DL*6;
-#else  
   k+=phy_vars_eNB->lte_frame_parms.first_carrier_offset;
-#endif
-  for (m=0;m<phy_vars_eNB->lte_frame_parms.N_RB_DL<<1;m++) {
+
+  for (m=0;m<phy_vars_eNB->lte_frame_parms.N_RB_DL<<1;m++) { // loop over pilots in one slot/symbol, 2*N_RB_DL pilots
 
     mprime_dword     = mprime>>4;
     mprime_qpsk_symb = mprime&0xf;
@@ -98,16 +75,10 @@ int lte_dl_cell_spec_SS(PHY_VARS_eNB *phy_vars_eNB,
     printf("Ns %d, l %d output[%d] = (%d,%d)\n",Ns,l,k,((short *)&output[k])[0],((short *)&output[k])[1]);
 #endif
     k+=6;//b
-#ifdef IFFT_FPGA
-    if (k >= phy_vars_eNB->lte_frame_parms.N_RB_DL*12) {
-      k-=phy_vars_eNB->lte_frame_parms.N_RB_DL*12;
-    }
-#else
     if (k >= phy_vars_eNB->lte_frame_parms.ofdm_symbol_size) {
       k++;  // skip DC carrier
       k-=phy_vars_eNB->lte_frame_parms.ofdm_symbol_size;
     }
-#endif
     //    printf("** k %d\n",k);
   }
   return(0);
@@ -125,36 +96,15 @@ int lte_dl_cell_spec(PHY_VARS_eNB *phy_vars_eNB,
   unsigned short k,a;
   mod_sym_t qpsk[4];
 
-#ifdef IFFT_FPGA
-  // new mod table
-  qpsk[0] = 1;
-  qpsk[1] = 3;
-  qpsk[2] = 2;
-  qpsk[3] = 4;
-  /* old mod table
-  qpsk[0] = 4;
-  qpsk[1] = 2;
-  qpsk[2] = 3;
-  qpsk[3] = 1;
-  */
-#else
   a = (amp*ONE_OVER_SQRT2_Q15)>>15;
   ((short *)&qpsk[0])[0] = a;
   ((short *)&qpsk[0])[1] = a;
-  // new mod table
   ((short *)&qpsk[1])[0] = -a;
   ((short *)&qpsk[1])[1] = a;
   ((short *)&qpsk[2])[0] = a;
   ((short *)&qpsk[2])[1] = -a;
-  /* old mod table
-  ((short *)&qpsk[1])[0] = a;
-  ((short *)&qpsk[1])[1] = -a;
-  ((short *)&qpsk[2])[0] = -a;
-  ((short *)&qpsk[2])[1] = a;
-  */
   ((short *)&qpsk[3])[0] = -a;
   ((short *)&qpsk[3])[1] = -a;
-#endif
 
   if ((p==0) && (l==0) )
     nu = 0;
@@ -175,11 +125,8 @@ int lte_dl_cell_spec(PHY_VARS_eNB *phy_vars_eNB,
   if (k > 5)
     k -=6;
 
-#ifdef IFFT_FPGA
-  k+=phy_vars_eNB->lte_frame_parms.N_RB_DL*6;
-#else  
   k+=phy_vars_eNB->lte_frame_parms.first_carrier_offset;
-#endif
+
   for (m=0;m<phy_vars_eNB->lte_frame_parms.N_RB_DL<<1;m++) {
 
     mprime_dword     = mprime>>4;
@@ -200,16 +147,11 @@ int lte_dl_cell_spec(PHY_VARS_eNB *phy_vars_eNB,
     printf("Ns %d, l %d output[%d] = (%d,%d)\n",Ns,l,k,((short *)&output[k])[0],((short *)&output[k])[1]);
 #endif
     k+=6;
-#ifdef IFFT_FPGA
-    if (k >= phy_vars_eNB->lte_frame_parms.N_RB_DL*12) {
-      k-=phy_vars_eNB->lte_frame_parms.N_RB_DL*12;
-    }
-#else
     if (k >= phy_vars_eNB->lte_frame_parms.ofdm_symbol_size) {
       k++;  // skip DC carrier
       k-=phy_vars_eNB->lte_frame_parms.ofdm_symbol_size;
     }
-#endif
+
     //    printf("** k %d\n",k);
   }
   return(0);
diff --git a/openair1/PHY/LTE_REFSIG/lte_gold.c b/openair1/PHY/LTE_REFSIG/lte_gold.c
index ee32b02dd6..d447bdf1d2 100644
--- a/openair1/PHY/LTE_REFSIG/lte_gold.c
+++ b/openair1/PHY/LTE_REFSIG/lte_gold.c
@@ -61,6 +61,52 @@ void lte_gold(LTE_DL_FRAME_PARMS *frame_parms,uint32_t lte_gold_table[20][2][14]
   }
 }
 
+void lte_gold_ue_spec(LTE_DL_FRAME_PARMS *frame_parms,uint32_t lte_gold_uespec_table[2][20][2][21],uint16_t Nid_cell, uint16_t *n_idDMRS) {
+
+  unsigned char ns,l;
+  unsigned int n,x1,x2;//,x1tmp,x2tmp;
+  int nscid;
+  int nid;
+  for (nscid=0;nscid<2;nscid++) {
+    if (n_idDMRS)
+      nid = n_idDMRS[nscid];
+    else
+      nid = Nid_cell;
+    for (ns=0;ns<20;ns++) {
+      
+      for (l=0;l<2;l++) {
+	
+	
+	x2 = ((((ns>>1)+1)*((nid<<1)+1))<<16) + nscid;
+	//x2 = frame_parms->Ncp + (Nid_cell<<1) + (1+(Nid_cell<<1))*(1 + (3*l) + (7*(1+ns))); //cinit
+	//n = 0
+	//      printf("cinit (ns %d, l %d) => %d\n",ns,l,x2);
+	x1 = 1+ (1<<31);
+	x2=x2 ^ ((x2 ^ (x2>>1) ^ (x2>>2) ^ (x2>>3))<<31);
+	// skip first 50 double words (1600 bits)
+	//printf("n=0 : x1 %x, x2 %x\n",x1,x2);
+	for (n=1;n<50;n++) {
+	  x1 = (x1>>1) ^ (x1>>4);
+	  x1 = x1 ^ (x1<<31) ^ (x1<<28);
+	  x2 = (x2>>1) ^ (x2>>2) ^ (x2>>3) ^ (x2>>4);
+	  x2 = x2 ^ (x2<<31) ^ (x2<<30) ^ (x2<<29) ^ (x2<<28);
+	  //	printf("x1 : %x, x2 : %x\n",x1,x2);
+	}
+	
+	for (n=0;n<14;n++) {
+	  x1 = (x1>>1) ^ (x1>>4);
+	  x1 = x1 ^ (x1<<31) ^ (x1<<28);
+	  x2 = (x2>>1) ^ (x2>>2) ^ (x2>>3) ^ (x2>>4);
+	  x2 = x2 ^ (x2<<31) ^ (x2<<30) ^ (x2<<29) ^ (x2<<28);
+	  lte_gold_uespec_table[nscid][ns][l][n] = x1^x2;
+	  //	printf("n=%d : c %x\n",n,x1^x2);	
+	}
+	
+      }
+    }
+  }
+}
+
 /*! \brief gold sequenquence generator
 \param x1 
 \param x2 this should be set to c_init if reset=1
diff --git a/openair1/PHY/LTE_TRANSPORT/dci.c b/openair1/PHY/LTE_TRANSPORT/dci.c
index c951518c46..aa21cf6d8f 100644
--- a/openair1/PHY/LTE_TRANSPORT/dci.c
+++ b/openair1/PHY/LTE_TRANSPORT/dci.c
@@ -1954,7 +1954,7 @@ uint8_t get_num_pdcch_symbols(uint8_t num_dci,
 
   // check pdcch duration imposed by PHICH duration (Section 6.9 of 36-211)
   if (frame_parms->Ncp==1) { // extended prefix
-    if ((frame_parms->frame_type == 1) && 
+    if ((frame_parms->frame_type == TDD) && 
 	((frame_parms->tdd_config<3)||(frame_parms->tdd_config==6)) &&
 	((subframe==1) || (subframe==6))) // subframes 1 and 6 (S-subframes) for 5ms switching periodicity are 2 symbols
       nCCEmin = 2;
@@ -2726,6 +2726,7 @@ uint16_t dci_decoding_procedure(PHY_VARS_UE *phy_vars_ue,
     }
     break;
   case 25:
+  default:
     if (frame_type == TDD) {
       format1A_size_bits  = sizeof_DCI1A_5MHz_TDD_1_6_t;
       format1A_size_bytes = sizeof(DCI1A_5MHz_TDD_1_6_t);
@@ -2780,11 +2781,6 @@ uint16_t dci_decoding_procedure(PHY_VARS_UE *phy_vars_ue,
       format1_size_bytes = sizeof(DCI1_20MHz_FDD_t);
     }
     break;
-  default:
-    format1_size_bits = 0;
-    format1_size_bytes = 0;
-    DevParam(frame_parms->N_RB_DL, frame_type, 0);
-    break;
   }
 
   if (do_common == 1) { 
diff --git a/openair1/PHY/LTE_TRANSPORT/defs.h b/openair1/PHY/LTE_TRANSPORT/defs.h
index 90cf49ad27..5f4b37cfc4 100644
--- a/openair1/PHY/LTE_TRANSPORT/defs.h
+++ b/openair1/PHY/LTE_TRANSPORT/defs.h
@@ -109,6 +109,10 @@ typedef struct {
   uint8_t rvidx;
   /// MIMO mode for this DLSCH
   MIMO_mode_t mimo_mode;
+  /// Number of layers for this PDSCH transmission
+  uint8_t Nlayers;
+  /// First layer for this PSCH transmission
+  uint8_t first_layer;
   /// Turbo-code outputs (36-212 V8.6 2009-03, p.12 
   uint8_t d[MAX_NUM_DLSCH_SEGMENTS][(96+3+(3*6144))];  
   /// Sub-block interleaver outputs (36-212 V8.6 2009-03, p.16-17)
@@ -199,6 +203,8 @@ typedef struct {
 } LTE_UL_UE_HARQ_t;
 
 typedef struct {
+  /// TX buffers for UE-spec transmission (antenna ports 5 or 7..14, prior to precoding)
+  uint32_t *txdataF[8]; 
   /// Allocated RNTI (0 means DLSCH_t is not currently used)
   uint16_t rnti; 
   /// Active flag for baseband transmitter processing
@@ -247,6 +253,7 @@ typedef struct {
   int16_t sqrt_rho_a;
   /// amplitude of PDSCH (compared to RS) in symbols containing pilots
   int16_t sqrt_rho_b;
+
 } LTE_eNB_DLSCH_t;
 
 #define PUSCH_x 2
diff --git a/openair1/PHY/LTE_TRANSPORT/dlsch_coding.c b/openair1/PHY/LTE_TRANSPORT/dlsch_coding.c
index 14972080ce..4908b26b1e 100644
--- a/openair1/PHY/LTE_TRANSPORT/dlsch_coding.c
+++ b/openair1/PHY/LTE_TRANSPORT/dlsch_coding.c
@@ -243,7 +243,7 @@ int dlsch_encoding(unsigned char *a,
   // printf("Encoder: A: %d\n",A);
   mod_order = get_Qm(dlsch->harq_processes[harq_pid]->mcs);
 
-  G = get_G(frame_parms,nb_rb,dlsch->rb_alloc,mod_order,num_pdcch_symbols,frame,subframe);
+  G = get_G(frame_parms,nb_rb,dlsch->rb_alloc,mod_order,dlsch->harq_processes[harq_pid]->Nl,num_pdcch_symbols,frame,subframe);
 
    
   //  if (dlsch->harq_processes[harq_pid]->Ndi == 1) {  // this is a new packet
@@ -353,7 +353,7 @@ int dlsch_encoding(unsigned char *a,
     r_offset += lte_rate_matching_turbo(dlsch->harq_processes[harq_pid]->RTC[r],
 					G,  //G
 					dlsch->harq_processes[harq_pid]->w[r],
-					&dlsch->e[r_offset],
+					dlsch->e+r_offset,
 					dlsch->harq_processes[harq_pid]->C, // C
 					NSOFT,                    // Nsoft,
 					dlsch->Mdlharq,
diff --git a/openair1/PHY/LTE_TRANSPORT/dlsch_decoding.c b/openair1/PHY/LTE_TRANSPORT/dlsch_decoding.c
index e5799a075a..8c378e1f1d 100644
--- a/openair1/PHY/LTE_TRANSPORT/dlsch_decoding.c
+++ b/openair1/PHY/LTE_TRANSPORT/dlsch_decoding.c
@@ -304,7 +304,7 @@ uint32_t  dlsch_decoding(PHY_VARS_UE *phy_vars_ue,
 				   G,
 				   harq_process->w[r],
 				   (uint8_t*)&dummy_w[r][0],
-				   dlsch_llr,
+				   dlsch_llr+r_offset,
 				   harq_process->C,
 				   NSOFT,
 				   dlsch->Mdlharq,
diff --git a/openair1/PHY/LTE_TRANSPORT/dlsch_demodulation.c b/openair1/PHY/LTE_TRANSPORT/dlsch_demodulation.c
index 35aaf3c29a..a809a4e579 100644
--- a/openair1/PHY/LTE_TRANSPORT/dlsch_demodulation.c
+++ b/openair1/PHY/LTE_TRANSPORT/dlsch_demodulation.c
@@ -247,14 +247,14 @@ int rx_pdsch(PHY_VARS_UE *phy_vars_ue,
     msg("dlsch_demodulation.c: nb_rb=0\n");
     return(-1);
   }
-
+  /*
   // DL power control: Scaling of Channel estimates for PDSCH
   dlsch_scale_channel(lte_ue_pdsch_vars[eNB_id]->dl_ch_estimates_ext,
 		      frame_parms,
 		      dlsch_ue,
 		      symbol,
 		      nb_rb);
-
+  */
   if (first_symbol_flag==1) {
     dlsch_channel_level(lte_ue_pdsch_vars[eNB_id]->dl_ch_estimates_ext,
 			frame_parms,
@@ -1353,7 +1353,7 @@ void dlsch_scale_channel(int **dl_ch_estimates_ext,
   // Determine scaling amplitude based the symbol
   ch_amp = ((pilots) ? (dlsch_ue[0]->sqrt_rho_b) : (dlsch_ue[0]->sqrt_rho_a));
 
-  //    msg("Scaling PDSCH Chest in OFDM symbol %d by %d\n",symbol_mod,ch_amp);
+  //  msg("Scaling PDSCH Chest in OFDM symbol %d by %d\n",symbol_mod,ch_amp);
 
   ch_amp128 = _mm_set1_epi16(ch_amp); // Q3.13
 
@@ -1657,9 +1657,9 @@ unsigned short dlsch_extract_rbs_single(int **rxdataF,
   symbol_mod = (symbol>=(7-frame_parms->Ncp)) ? symbol-(7-frame_parms->Ncp) : symbol;
   pilots = ((symbol_mod==0)||(symbol_mod==(4-frame_parms->Ncp))) ? 1 : 0;
   l=symbol;
-  nsymb = (frame_parms->Ncp==0) ? 14:12;
+  nsymb = (frame_parms->Ncp==NORMAL) ? 14:12;
 
-  if (frame_parms->frame_type == 1) {  // TDD
+  if (frame_parms->frame_type == TDD) {  // TDD
     sss_symb = nsymb-1;
     pss_symb = 2;
   }
@@ -1803,7 +1803,7 @@ unsigned short dlsch_extract_rbs_single(int **rxdataF,
 	  skip_half=2;
 
 	//PSS in subframe 0/5 if FDD
-	if (frame_parms->frame_type == 0) {  //FDD
+	if (frame_parms->frame_type == FDD) {  //FDD
 	  if (((subframe==0)||(subframe==5)) && (rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
 	    rb_alloc_ind = 0;
 	  }
@@ -1813,7 +1813,7 @@ unsigned short dlsch_extract_rbs_single(int **rxdataF,
 	    skip_half=2;
 	}
 	
-	if ((frame_parms->frame_type == 1) &&
+	if ((frame_parms->frame_type == TDD) &&
 	    (subframe==6)){  //TDD Subframe 6
 	  if ((rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
 	    rb_alloc_ind = 0;
@@ -1951,14 +1951,14 @@ unsigned short dlsch_extract_rbs_single(int **rxdataF,
       if (((subframe==0)||(subframe==5)) && (rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==sss_symb) ) {
 	rb_alloc_ind = 0;
       }
-      if (frame_parms->frame_type == 0) {
+      if (frame_parms->frame_type == FDD) {
 	//PSS
 	if (((subframe==0)||(subframe==5)) && (rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
 	  rb_alloc_ind = 0;
 	}
       }
 
-      if ((frame_parms->frame_type == 1) &&
+      if ((frame_parms->frame_type == TDD) &&
 	  (subframe==6)){
 	//PSS
 	if ((rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
@@ -2078,7 +2078,7 @@ unsigned short dlsch_extract_rbs_single(int **rxdataF,
 	else if (((subframe==0)||(subframe==5)) && (rb==((frame_parms->N_RB_DL>>1)+3)) && (l==sss_symb))
 	  skip_half=2;
       
-	if (frame_parms->frame_type == 0) {
+	if (frame_parms->frame_type == FDD) {
 	  //PSS
 	  if (((subframe==0)||(subframe==5)) && (rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
 	    rb_alloc_ind = 0;
@@ -2090,7 +2090,7 @@ unsigned short dlsch_extract_rbs_single(int **rxdataF,
 	    skip_half=2;
 	}
 
-	if ((frame_parms->frame_type == 1) &&
+	if ((frame_parms->frame_type == TDD) &&
 	    (subframe==6)){  //TDD Subframe 6
 	  if ((rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
 	    rb_alloc_ind = 0;
@@ -2238,10 +2238,10 @@ unsigned short dlsch_extract_rbs_dual(int **rxdataF,
 
   if ((symbol_mod == 0) || (symbol_mod == (4-frame_parms->Ncp)))
     pilots=1;
-  nsymb = (frame_parms->Ncp==0) ? 14:12;
+  nsymb = (frame_parms->Ncp==NORMAL) ? 14:12;
   l=symbol;
 
-  if (frame_parms->frame_type == 1) {  // TDD
+  if (frame_parms->frame_type == TDD) {  // TDD
     sss_symb = nsymb-1;
     pss_symb = 2;
   }
@@ -2320,7 +2320,7 @@ unsigned short dlsch_extract_rbs_dual(int **rxdataF,
 	  skip_half=2;
 
 	//PSS in subframe 0/5 if FDD
-	if (frame_parms->frame_type == 0) {  //FDD
+	if (frame_parms->frame_type == FDD) {  //FDD
 	  if (((subframe==0)||(subframe==5)) && (rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
 	    rb_alloc_ind = 0;
 	  }
@@ -2330,7 +2330,7 @@ unsigned short dlsch_extract_rbs_dual(int **rxdataF,
 	    skip_half=2;
 	}
 	
-	if ((frame_parms->frame_type == 1) &&
+	if ((frame_parms->frame_type == TDD) &&
 	    (subframe==6)){  //TDD Subframe 6
 	  if ((rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
 	    rb_alloc_ind = 0;
@@ -2451,7 +2451,7 @@ unsigned short dlsch_extract_rbs_dual(int **rxdataF,
 	  skip_half=2;
 	
 	//PSS in subframe 0/5 if FDD
-	if (frame_parms->frame_type == 0) {  //FDD
+	if (frame_parms->frame_type == FDD) {  //FDD
 	  if (((subframe==0)||(subframe==5)) && (rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
 	    rb_alloc_ind = 0;
 	  }
@@ -2461,7 +2461,7 @@ unsigned short dlsch_extract_rbs_dual(int **rxdataF,
 	    skip_half=2;
 	}
 	
-	if ((frame_parms->frame_type == 1) &&
+	if ((frame_parms->frame_type == TDD) &&
 	    (subframe==6)){  //TDD Subframe 6
 	  if ((rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
 	    rb_alloc_ind = 0;
@@ -2621,14 +2621,14 @@ unsigned short dlsch_extract_rbs_dual(int **rxdataF,
       if (((subframe==0)||(subframe==5)) && (rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==sss_symb) ) {
 	rb_alloc_ind = 0;
       }
-      if (frame_parms->frame_type == 0) {
+      if (frame_parms->frame_type == FDD) {
 	//PSS
 	if (((subframe==0)||(subframe==5)) && (rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
 	  rb_alloc_ind = 0;
 	}
       }
       
-      if ((frame_parms->frame_type == 1) &&
+      if ((frame_parms->frame_type == TDD) &&
 	  (subframe==6)){
 	//PSS
 	if ((rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
@@ -2770,7 +2770,7 @@ unsigned short dlsch_extract_rbs_dual(int **rxdataF,
 		 (l==sss_symb))
 	  skip_half=2;
 
-	if (frame_parms->frame_type == 0) {
+	if (frame_parms->frame_type == FDD) {
 	  //PSS
 	  if (((subframe==0)||(subframe==5)) && (rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
 	    rb_alloc_ind = 0;
@@ -2782,7 +2782,7 @@ unsigned short dlsch_extract_rbs_dual(int **rxdataF,
 	    skip_half=2;
 	}
 	
-	if ((frame_parms->frame_type == 1) &&
+	if ((frame_parms->frame_type == TDD) &&
 	    (subframe==6)){
 	  //PSS
 	  if ((rb>((frame_parms->N_RB_DL>>1)-3)) && (rb<((frame_parms->N_RB_DL>>1)+3)) && (l==pss_symb) ) {
diff --git a/openair1/PHY/LTE_TRANSPORT/dlsch_llr_computation.c b/openair1/PHY/LTE_TRANSPORT/dlsch_llr_computation.c
index 9bfc35ce4e..b8e29b5d9f 100644
--- a/openair1/PHY/LTE_TRANSPORT/dlsch_llr_computation.c
+++ b/openair1/PHY/LTE_TRANSPORT/dlsch_llr_computation.c
@@ -56,7 +56,7 @@
 
 #ifndef __SSE3__
 //#define _mm_abs_epi16(xmmx) _mm_xor_si128((xmmx),_mm_cmpgt_epi16(zero,(xmmx)))
-#define _mm_abs_epi16(xmmx) _mm_add_epi16(_mm_xor_si128((xmmx),_mm_cmpgt_epi16((*(__m128i*)&zero[0]),(xmmx))),_mm_srli_epi16(_mm_cmpgt_epi16((*(__m128i*)&zero[0]),(xmmx)),15))
+//abs_#define _mm_abs_epi16(xmmx) _mm_add_epi16(_mm_xor_si128((xmmx),_mm_cmpgt_epi16((*(__m128i*)&zero[0]),(xmmx))),_mm_srli_epi16(_mm_cmpgt_epi16((*(__m128i*)&zero[0]),(xmmx)),15))
 //#define _mm_sign_epi16(xmmx,xmmy) _mm_xor_si128((xmmx),_mm_cmpgt_epi16(zero,(xmmy)))
 #endif
 
@@ -720,7 +720,8 @@ void dlsch_16qam_llr(LTE_DL_FRAME_PARMS *frame_parms,
     int i,len;
     unsigned char symbol_mod,len_mod4=0;
     uint32_t *llr32;
-    
+
+
     if (first_symbol_flag==1) {
         llr32 = (uint32_t*)dlsch_llr;
     }
@@ -826,7 +827,7 @@ void dlsch_64qam_llr(LTE_DL_FRAME_PARMS *frame_parms,
     for (i=0;i<len2;i++) {
         
         xmm1 = _mm_abs_epi16(rxF[i]);
-        xmm1  = _mm_subs_epi16(ch_mag[i],xmm1);
+        xmm1 = _mm_subs_epi16(ch_mag[i],xmm1);
         xmm2 = _mm_abs_epi16(xmm1);
         xmm2 = _mm_subs_epi16(ch_magb[i],xmm2);
         
diff --git a/openair1/PHY/LTE_TRANSPORT/dlsch_modulation.c b/openair1/PHY/LTE_TRANSPORT/dlsch_modulation.c
index 68c272df90..2b377d68b9 100644
--- a/openair1/PHY/LTE_TRANSPORT/dlsch_modulation.c
+++ b/openair1/PHY/LTE_TRANSPORT/dlsch_modulation.c
@@ -69,6 +69,11 @@ uint8_t is_not_pilot(uint8_t pilots, uint8_t re, uint8_t nushift, uint8_t use2nd
   return(0);
 }
 
+uint8_t is_not_UEspecRS(int first_layer,int re) {
+
+  return(1);
+}
+
 void generate_64qam_table(void) {
 
   int a,b,c,index;
@@ -142,6 +147,8 @@ int allocate_REs_in_RB(mod_sym_t **txdataF,
 		       uint8_t skip_dc,
 		       uint8_t skip_half,
 		       uint8_t use2ndpilots,
+		       uint8_t Nlayers,
+		       uint8_t first_layer,
 		       LTE_DL_FRAME_PARMS *frame_parms) {
 
   uint32_t tti_offset,aa;
@@ -156,7 +163,7 @@ int allocate_REs_in_RB(mod_sym_t **txdataF,
   uint8_t first_re,last_re;
   int32_t tmp_sample1,tmp_sample2;
   int16_t tmp_amp=amp;
-
+  uint8_t layer;
   /*
   switch (mod_order) {
   case 2:
@@ -197,24 +204,27 @@ int allocate_REs_in_RB(mod_sym_t **txdataF,
     
     tti_offset = symbol_offset + re_off + re;
     
-    // check that re is not a pilot
+    // check that RE is not from Cell-specific RS
+
     if (is_not_pilot(pilots,re,frame_parms->nushift,use2ndpilots)==1) { 
       //     printf("re %d (jj %d)\n",re,*jj);
-      *re_allocated = *re_allocated + 1;
+
       
       if (mimo_mode == SISO) {  //SISO mapping
-	
+	*re_allocated = *re_allocated + 1;	
 	switch (mod_order) {
 	case 2:  //QPSK
-	  //	  printf("%d : %d,%d => ",tti_offset,((int16_t*)&txdataF[0][tti_offset])[0],((int16_t*)&txdataF[0][tti_offset])[1]);
-	  for (aa=0; aa<frame_parms->nb_antennas_tx; aa++)
+	  //	  printf("%d(%d) : %d,%d => ",tti_offset,*jj,((int16_t*)&txdataF[0][tti_offset])[0],((int16_t*)&txdataF[0][tti_offset])[1]);
+	  for (aa=0; aa<frame_parms->nb_antennas_tx; aa++) {
 	    ((int16_t*)&txdataF[aa][tti_offset])[0] += (output[*jj]==1) ? (-gain_lin_QPSK) : gain_lin_QPSK; //I //b_i
+	  }
 	  *jj = *jj + 1;
-	  for (aa=0; aa<frame_parms->nb_antennas_tx; aa++)
+	  for (aa=0; aa<frame_parms->nb_antennas_tx; aa++) {
 	    ((int16_t*)&txdataF[aa][tti_offset])[1] += (output[*jj]==1) ? (-gain_lin_QPSK) : gain_lin_QPSK; //Q //b_{i+1}
+	  }
 	  *jj = *jj + 1;
-
-	  //	  printf("%d,%d\n",((int16_t*)&txdataF[0][tti_offset])[0],((int16_t*)&txdataF[0][tti_offset])[1]);
+	  
+	  // printf("%d,%d\n",((int16_t*)&txdataF[0][tti_offset])[0],((int16_t*)&txdataF[0][tti_offset])[1]);
 	  break;
 	  
 	case 4:  //16QAM
@@ -250,7 +260,7 @@ int allocate_REs_in_RB(mod_sym_t **txdataF,
 	  
 	  qam64_table_offset_re = 0;
 	  qam64_table_offset_im = 0;
-
+	  
 	  if (output[*jj] == 1)
 	    qam64_table_offset_re+=4;
 	  *jj=*jj+1;
@@ -280,16 +290,16 @@ int allocate_REs_in_RB(mod_sym_t **txdataF,
       }
             
       else if (mimo_mode == ALAMOUTI){
-
+	*re_allocated = *re_allocated + 1;	
           // normalization for 2 tx antennas
-          amp = (int16_t)(((int32_t)tmp_amp*ONE_OVER_SQRT2_Q15)>>15);
+	amp = (int16_t)(((int32_t)tmp_amp*ONE_OVER_SQRT2_Q15)>>15);
 
 	switch (mod_order) {
 	case 2:  //QPSK
 	  
 	  // first antenna position n -> x0
 	  
-      ((int16_t*)&tmp_sample1)[0] = (output[*jj]==1) ? (-gain_lin_QPSK) : gain_lin_QPSK;
+	  ((int16_t*)&tmp_sample1)[0] = (output[*jj]==1) ? (-gain_lin_QPSK) : gain_lin_QPSK;
 	  *jj=*jj+1;
 	  ((int16_t*)&tmp_sample1)[1] = (output[*jj]==1) ? (-gain_lin_QPSK) : gain_lin_QPSK;
 	  *jj=*jj+1;
@@ -428,6 +438,7 @@ int allocate_REs_in_RB(mod_sym_t **txdataF,
 	}
       }
       else if (mimo_mode == ANTCYCLING ) {
+	*re_allocated = *re_allocated + 1;	
 	switch (mod_order) {
 	case 2:  //QPSK
 	  
@@ -495,8 +506,8 @@ int allocate_REs_in_RB(mod_sym_t **txdataF,
        
       else if ((mimo_mode >= UNIFORM_PRECODING11)&&(mimo_mode <= PUSCH_PRECODING1)) {
 	// this is for transmission modes 4-6 (1 layer)
-         
-          amp = (int16_t)(((int32_t)tmp_amp*ONE_OVER_SQRT2_Q15)>>15);
+	*re_allocated = *re_allocated + 1;	         
+	amp = (int16_t)(((int32_t)tmp_amp*ONE_OVER_SQRT2_Q15)>>15);
 
 	switch (mod_order) {
 	case 2:  //QPSK
@@ -592,14 +603,6 @@ int allocate_REs_in_RB(mod_sym_t **txdataF,
 	  
 	}
       }
-      
-      else {
-	msg("allocate_REs_in_RB() [dlsch.c] : ERROR, unknown mimo_mode %d\n",mimo_mode);
-	return(-1);
-      }
-      
-     
-    
       if (mimo_mode == ALAMOUTI) {
 	re++;  // adjacent carriers are taken care of by precoding
 	*re_allocated = *re_allocated + 1;
@@ -608,8 +611,92 @@ int allocate_REs_in_RB(mod_sym_t **txdataF,
 	  *re_allocated = *re_allocated + 1;
 	}
       }
+    
+      if (mimo_mode >= TM8) { //TM8,TM9,TM10
+
+	if (is_not_UEspecRS(first_layer,re)) {	  	  	
+	  switch (mod_order) {
+	  case 2:  //QPSK
+	    //	  printf("%d : %d,%d => ",tti_offset,((int16_t*)&txdataF[0][tti_offset])[0],((int16_t*)&txdataF[0][tti_offset])[1]);
+	    for (layer=first_layer;layer<=(first_layer+Nlayers);layer++) {
+	      
+	      ((int16_t*)&txdataF[layer][tti_offset])[0] = (output[*jj]==1) ? (-gain_lin_QPSK) : gain_lin_QPSK; //I //b_i
+	      *jj = *jj + 1;
+	      ((int16_t*)&txdataF[layer][tti_offset])[1] = (output[*jj]==1) ? (-gain_lin_QPSK) : gain_lin_QPSK; //Q //b_{i+1}
+	      *jj = *jj + 1;
+	    }
+	    break;
+	  
+	  case 4:  //16QAM
+	    if (is_not_UEspecRS(layer,re)) {	  	  
+	      qam16_table_offset_re = 0;
+	      qam16_table_offset_im = 0;
+	      if (output[*jj] == 1)
+		qam16_table_offset_re+=2;
+	      *jj=*jj+1;
+	      if (output[*jj] == 1)
+		qam16_table_offset_im+=2;
+	      *jj=*jj+1;
+	      
+	      
+	      if (output[*jj] == 1)
+		qam16_table_offset_re+=1;
+	      *jj=*jj+1;
+	      if (output[*jj] == 1)
+		qam16_table_offset_im+=1;
+	      *jj=*jj+1;
+	      
+	    
+	      for (layer=first_layer;layer<=(first_layer+Nlayers);layer++) {
+		((int16_t*)&txdataF[layer][tti_offset])[0] = qam_table_s[qam16_table_offset_re];
+		((int16_t*)&txdataF[layer][tti_offset])[1] = qam_table_s[qam16_table_offset_im];
+	      }
+	    }
+	    break;
+	    
+	  case 6:  //64QAM
+
+
+	    qam64_table_offset_re = 0;
+	    qam64_table_offset_im = 0;
+	    
+	    if (output[*jj] == 1)
+	      qam64_table_offset_re+=4;
+	    *jj=*jj+1;
+	    if (output[*jj] == 1)
+	      qam64_table_offset_im+=4;
+	    *jj=*jj+1;
+	    if (output[*jj] == 1)
+	      qam64_table_offset_re+=2;
+	    *jj=*jj+1;
+	    if (output[*jj] == 1)
+	      qam64_table_offset_im+=2;
+	    *jj=*jj+1;
+	    if (output[*jj] == 1)
+	      qam64_table_offset_re+=1;
+	    *jj=*jj+1;
+	    if (output[*jj] == 1)
+	      qam64_table_offset_im+=1;
+	    *jj=*jj+1;
+	    
+	    for (layer=first_layer;layer<=(first_layer+Nlayers);layer++) {
+	      ((int16_t*)&txdataF[layer][tti_offset])[0] = qam_table_s[qam64_table_offset_re];
+	      ((int16_t*)&txdataF[layer][tti_offset])[1] = qam_table_s[qam64_table_offset_im];
+	    }	  
+	    break;
+	    
+	  }
+	}
+      }
+      else if (mimo_mode>=TM9_10) {
+	msg("allocate_REs_in_RB() [dlsch.c] : ERROR, unknown mimo_mode %d\n",mimo_mode);
+	return(-1);
+      }
     }
+      
+     
   }
+
   return(0);
 }
 
@@ -768,6 +855,7 @@ int dlsch_modulation(mod_sym_t **txdataF,
   int16_t amp_rho_a, amp_rho_b;
   int16_t qam16_table_a[4],qam64_table_a[8],qam16_table_b[4],qam64_table_b[8];
   int16_t *qam_table_s;
+  uint8_t Nl = dlsch->harq_processes[harq_pid]->Nl;
 
   vcd_signal_dumper_dump_function_by_name(VCD_SIGNAL_DUMPER_FUNCTIONS_ENB_DLSCH_MODULATION, VCD_FUNCTION_IN);
 
@@ -959,7 +1047,7 @@ int dlsch_modulation(mod_sym_t **txdataF,
 
 	if (rb_alloc_ind > 0)
         {
-	  //printf("Allocated rb %d, subframe_offset %d, symbol_offset %d, re_offset %d\n",rb,subframe_offset,symbol_offset,re_offset);
+	  //	  printf("Allocated rb %d, subframe_offset %d, symbol_offset %d, re_offset %d\n",rb,subframe_offset,symbol_offset,re_offset);
 	  allocate_REs_in_RB(txdataF,
 			     &jj,
 			     re_offset,
@@ -976,6 +1064,8 @@ int dlsch_modulation(mod_sym_t **txdataF,
 			     skip_dc,
 			     skip_half,
 			     (frame_parms->mode1_flag==1)?1:0,
+			     dlsch->harq_processes[harq_pid]->Nlayers, 
+			     dlsch->harq_processes[harq_pid]->first_layer, 
 			     frame_parms);
       }
 	re_offset+=12; // go to next RB
@@ -995,7 +1085,7 @@ int dlsch_modulation(mod_sym_t **txdataF,
 
 
 #ifdef DEBUG_DLSCH_MODULATION
-  msg("generate_dlsch : jj = %d,re_allocated = %d (G %d)\n",jj,re_allocated,get_G(frame_parms,dlsch->nb_rb,dlsch->rb_alloc,mod_order,2,0,subframe_offset));
+  msg("generate_dlsch : jj = %d,re_allocated = %d (G %d)\n",jj,re_allocated,get_G(frame_parms,dlsch->nb_rb,dlsch->rb_alloc,mod_order,Nl,2,0,subframe_offset));
 #endif
   
   vcd_signal_dumper_dump_function_by_name(VCD_SIGNAL_DUMPER_FUNCTIONS_ENB_DLSCH_MODULATION, VCD_FUNCTION_OUT);
@@ -1092,7 +1182,7 @@ int mch_modulation(mod_sym_t **txdataF,
 
 
 #ifdef DEBUG_DLSCH_MODULATION
-    msg("generate_dlsch(MCH) : jj = %d,re_allocated = %d (G %d)\n",jj,re_allocated,get_G(frame_parms,dlsch->nb_rb,dlsch->rb_alloc,mod_order,2,0,subframe_offset));
+  msg("generate_dlsch(MCH) : jj = %d,re_allocated = %d (G %d)\n",jj,re_allocated,get_G(frame_parms,dlsch->nb_rb,dlsch->rb_alloc,mod_order,1,2,0,subframe_offset));
 #endif
     
     return (re_allocated);
diff --git a/openair1/PHY/LTE_TRANSPORT/initial_sync.c b/openair1/PHY/LTE_TRANSPORT/initial_sync.c
index c427cf27de..8611cb9f1d 100644
--- a/openair1/PHY/LTE_TRANSPORT/initial_sync.c
+++ b/openair1/PHY/LTE_TRANSPORT/initial_sync.c
@@ -244,8 +244,8 @@ int initial_sync(PHY_VARS_UE *phy_vars_ue, runmode_t mode) {
 
   //  LOG_I(PHY,"**************************************************************\n");
   // First try FDD normal prefix
-  frame_parms->Ncp=0;
-  frame_parms->frame_type=0;
+  frame_parms->Ncp=NORMAL;
+  frame_parms->frame_type=FDD;
   init_frame_parms(frame_parms,1);
 
   sync_pos = lte_sync_time(phy_vars_ue->lte_ue_common_vars.rxdata, 
@@ -316,8 +316,8 @@ int initial_sync(PHY_VARS_UE *phy_vars_ue, runmode_t mode) {
   if (ret==-1) {
 
     // Now FDD extended prefix
-    frame_parms->Ncp=1;
-    frame_parms->frame_type=0;
+    frame_parms->Ncp=EXTENDED;
+    frame_parms->frame_type=FDD;
     init_frame_parms(frame_parms,1);
     if (sync_pos < frame_parms->nb_prefix_samples)
       sync_pos2 = sync_pos + FRAME_LENGTH_COMPLEX_SAMPLES - frame_parms->nb_prefix_samples;
@@ -357,8 +357,8 @@ int initial_sync(PHY_VARS_UE *phy_vars_ue, runmode_t mode) {
 
     if (ret==-1) {
       // Now TDD normal prefix
-      frame_parms->Ncp=0;
-      frame_parms->frame_type=1;
+      frame_parms->Ncp=NORMAL;
+      frame_parms->frame_type=TDD;
       init_frame_parms(frame_parms,1);
 
       if (sync_pos >= frame_parms->nb_prefix_samples)
@@ -405,8 +405,8 @@ int initial_sync(PHY_VARS_UE *phy_vars_ue, runmode_t mode) {
 
       if (ret==-1) {
 	// Now TDD extended prefix
-	frame_parms->Ncp=1;
-	frame_parms->frame_type=1;
+	frame_parms->Ncp=EXTENDED;
+	frame_parms->frame_type=TDD;
 	init_frame_parms(frame_parms,1);
 	sync_pos2 = sync_pos - frame_parms->nb_prefix_samples;
 	if (sync_pos >= frame_parms->nb_prefix_samples)
diff --git a/openair1/PHY/LTE_TRANSPORT/lte_mcs.c b/openair1/PHY/LTE_TRANSPORT/lte_mcs.c
index 3d79f9ead5..9de315de6b 100644
--- a/openair1/PHY/LTE_TRANSPORT/lte_mcs.c
+++ b/openair1/PHY/LTE_TRANSPORT/lte_mcs.c
@@ -150,7 +150,7 @@ int adjust_G2(LTE_DL_FRAME_PARMS *frame_parms,uint32_t *rb_alloc,uint8_t mod_ord
   int rb,re_pbch_sss=0;
   int rb_alloc_ind,nsymb;
 
-  nsymb = (frame_parms->Ncp==0) ? 14 : 12;
+  nsymb = (frame_parms->Ncp==NORMAL) ? 14 : 12;
 
   //      printf("adjust_G2 : symbol %d, subframe %d\n",symbol,subframe);
   if ((subframe!=0) && (subframe!=5) && (subframe!=6))  // if not PBCH/SSS or SSS
@@ -296,7 +296,7 @@ int adjust_G(LTE_DL_FRAME_PARMS *frame_parms,uint32_t *rb_alloc,uint8_t mod_orde
   }
   //    printf("re_pbch_sss %d\n",re_pbch_sss);
   if (subframe==0) {  //PBCH+SSS+PSS
-    if (frame_parms->frame_type == 1) { // TDD
+    if (frame_parms->frame_type == TDD) { // TDD
       if (frame_parms->mode1_flag==0)
 	//2ant so PBCH 3+2/3 symbols, SSS 1 symbol * REs => (14/3)*re_pbch_sss for normal CP,  
 	// 2+4/3 symbols, SSS 1 symbol => (13/3)*re_pbch_sss for ext. CP
@@ -319,14 +319,14 @@ int adjust_G(LTE_DL_FRAME_PARMS *frame_parms,uint32_t *rb_alloc,uint8_t mod_orde
     }
   }
   else if (subframe == 5)  // SSS+PSS for FDD, SSS for TDD
-    return(((frame_parms->frame_type==0)?2:1)*re_pbch_sss * 1 * mod_order);
-  else if ((subframe == 6)&&(frame_parms->frame_type == 1)) // PSS for TDD
+    return(((frame_parms->frame_type==FDD)?2:1)*re_pbch_sss * 1 * mod_order);
+  else if ((subframe == 6)&&(frame_parms->frame_type == TDD)) // PSS for TDD
     return(re_pbch_sss * 1 * mod_order);
 
   return(0);
 }
 
-int get_G(LTE_DL_FRAME_PARMS *frame_parms,uint16_t nb_rb,uint32_t *rb_alloc,uint8_t mod_order,uint8_t num_pdcch_symbols,int frame,uint8_t subframe) {
+int get_G(LTE_DL_FRAME_PARMS *frame_parms,uint16_t nb_rb,uint32_t *rb_alloc,uint8_t mod_order,uint8_t Nl,uint8_t num_pdcch_symbols,int frame,uint8_t subframe) {
 
   
 
@@ -336,12 +336,12 @@ int get_G(LTE_DL_FRAME_PARMS *frame_parms,uint16_t nb_rb,uint32_t *rb_alloc,uint
     G_adj= adjust_G(frame_parms,rb_alloc,mod_order,subframe);
     
     //    printf("get_G subframe %d mod_order %d, nb_rb %d: rb_alloc %x,%x,%x,%x, G_adj %d\n",subframe,mod_order,nb_rb,rb_alloc[3],rb_alloc[2],rb_alloc[1],rb_alloc[0], G_adj);
-    if (frame_parms->Ncp==0) { // normal prefix
+    if (frame_parms->Ncp==NORMAL) { // normal prefix
       // PDDDPDD PDDDPDD - 13 PDSCH symbols, 10 full, 3 w/ pilots = 10*12 + 3*8
       // PCDDPDD PDDDPDD - 12 PDSCH symbols, 9 full, 3 w/ pilots = 9*12 + 3*8
       // PCCDPDD PDDDPDD - 11 PDSCH symbols, 8 full, 3 w/pilots = 8*12 + 3*8
       if (frame_parms->mode1_flag==0) // SISO 
-	return(((int)nb_rb * mod_order * ((11-num_pdcch_symbols)*12 + 3*8)) - G_adj);
+	return((((int)nb_rb * mod_order * ((11-num_pdcch_symbols)*12 + 3*8)) - G_adj)*Nl);
       else
 	return(((int)nb_rb * mod_order * ((11-num_pdcch_symbols)*12 + 3*10)) - G_adj);
     }
@@ -350,7 +350,7 @@ int get_G(LTE_DL_FRAME_PARMS *frame_parms,uint16_t nb_rb,uint32_t *rb_alloc,uint
       // PCDPDD PDDPDD - 10 PDSCH symbols, 7 full, 3 w/ pilots = 7*12 + 3*8
       // PCCPDD PDDPDD - 9 PDSCH symbols, 6 full, 3 w/pilots = 6*12 + 3*8
       if (frame_parms->mode1_flag==0)
-	return(((int)nb_rb * mod_order * ((9-num_pdcch_symbols)*12 + 3*8)) - G_adj);
+	return((((int)nb_rb * mod_order * ((9-num_pdcch_symbols)*12 + 3*8)) - G_adj)*Nl);
       else
 	return(((int)nb_rb * mod_order * ((9-num_pdcch_symbols)*12 + 3*10)) - G_adj);
     }
diff --git a/openair1/PHY/LTE_TRANSPORT/pbch.c b/openair1/PHY/LTE_TRANSPORT/pbch.c
index 6197605154..ed9edec8c3 100755
--- a/openair1/PHY/LTE_TRANSPORT/pbch.c
+++ b/openair1/PHY/LTE_TRANSPORT/pbch.c
@@ -293,6 +293,8 @@ int generate_pbch(LTE_eNB_PBCH *eNB_pbch,
 			 0,
 			 0,
 			 0,
+			 1,
+			 0,
 			 frame_parms);
       
       re_offset+=12; // go to next RB
diff --git a/openair1/PHY/LTE_TRANSPORT/pmch.c b/openair1/PHY/LTE_TRANSPORT/pmch.c
index 9a7f75cd94..e243421329 100644
--- a/openair1/PHY/LTE_TRANSPORT/pmch.c
+++ b/openair1/PHY/LTE_TRANSPORT/pmch.c
@@ -245,7 +245,7 @@ void fill_UE_dlsch_MCH(PHY_VARS_UE *phy_vars_ue,int mcs,int ndi,int rvidx,int eN
     G = get_G(&phy_vars_eNB->lte_frame_parms,
 	      phy_vars_eNB->lte_frame_parms.N_RB_DL,
 	      phy_vars_eNB->dlsch_eNB_MCH->rb_alloc,
-	      get_Qm(phy_vars_eNB->dlsch_eNB_MCH->harq_processes[0]->mcs),
+	      get_Qm(phy_vars_eNB->dlsch_eNB_MCH->harq_processes[0]->mcs),1,
 	      2,phy_vars_eNB->frame,subframe);
     
     generate_mbsfn_pilot(phy_vars_eNB,
diff --git a/openair1/PHY/LTE_TRANSPORT/proto.h b/openair1/PHY/LTE_TRANSPORT/proto.h
index b5d0e46d30..1dc5182ed6 100644
--- a/openair1/PHY/LTE_TRANSPORT/proto.h
+++ b/openair1/PHY/LTE_TRANSPORT/proto.h
@@ -142,6 +142,8 @@ void dlsch_encoding_emul(PHY_VARS_eNB *phy_vars_eNB,
     uint8_t skip_dc,
     uint8_t skip_half,
     uint8_t use2ndpilots,
+    uint8_t Nlayers,
+    uint8_t firstlayer,
     LTE_DL_FRAME_PARMS *frame_parms);
 
     \brief Fills RB with data
@@ -161,26 +163,30 @@ void dlsch_encoding_emul(PHY_VARS_eNB *phy_vars_eNB,
     \param skip_dc offset for positive RBs
     \param skip_half indicate that first or second half of RB must be skipped for PBCH/PSS/SSS
     \param use2ndpilots Set to use the pilots from antenna port 1 for PDSCH
+    \param Nlayers Number of layers for this codeword
+    \param firstlayer Index of first layer (minus 7, i.e. 0..7 <-> p=7,...,14
     \param frame_parms Frame parameter descriptor
 */
 
 int32_t allocate_REs_in_RB(mod_sym_t **txdataF,
-		       uint32_t *jj,
-		       uint16_t re_offset,
-		       uint32_t symbol_offset,
-		       uint8_t *output,
-		       MIMO_mode_t mimo_mode,
-		       uint8_t nu,
-		       uint8_t pilots,
-		       uint8_t mod_order,
-		       uint8_t precoder_index,
-		       int16_t amp,
-		       int16_t *qam_table_s,
-		       uint32_t *re_allocated,
-		       uint8_t skip_dc,
-		       uint8_t skip_half,
-		       uint8_t use2ndpilots,
-		       LTE_DL_FRAME_PARMS *frame_parms);
+			   uint32_t *jj,
+			   uint16_t re_offset,
+			   uint32_t symbol_offset,
+			   uint8_t *output,
+			   MIMO_mode_t mimo_mode,
+			   uint8_t nu,
+			   uint8_t pilots,
+			   uint8_t mod_order,
+			   uint8_t precoder_index,
+			   int16_t amp,
+			   int16_t *qam_table_s,
+			   uint32_t *re_allocated,
+			   uint8_t skip_dc,
+			   uint8_t skip_half,
+			   uint8_t use2ndpilots,
+			   uint8_t Nlayers,
+			   uint8_t firstlayer,
+			   LTE_DL_FRAME_PARMS *frame_parms);
 
 /** \fn int32_t dlsch_modulation(mod_sym_t **txdataF,
     int16_t amp,
@@ -1225,7 +1231,7 @@ uint8_t get_transmission_mode(uint16_t Mod_id, uint16_t rnti);
 */
 uint32_t conv_nprb(uint8_t ra_header,uint32_t rb_alloc,int N_RB_DL);
 
-int get_G(LTE_DL_FRAME_PARMS *frame_parms,uint16_t nb_rb,uint32_t *rb_alloc,uint8_t mod_order,uint8_t num_pdcch_symbols,int frame,uint8_t subframe);
+int get_G(LTE_DL_FRAME_PARMS *frame_parms,uint16_t nb_rb,uint32_t *rb_alloc,uint8_t mod_order,uint8_t Nl,uint8_t num_pdcch_symbols,int frame,uint8_t subframe);
 
 int adjust_G(LTE_DL_FRAME_PARMS *frame_parms,uint32_t *rb_alloc,uint8_t mod_order,uint8_t subframe);
 int adjust_G2(LTE_DL_FRAME_PARMS *frame_parms,uint32_t *rb_alloc,uint8_t mod_order,uint8_t subframe,uint8_t symbol);
diff --git a/openair1/PHY/LTE_TRANSPORT/pss.c b/openair1/PHY/LTE_TRANSPORT/pss.c
index 2e92da3691..6b162b017c 100644
--- a/openair1/PHY/LTE_TRANSPORT/pss.c
+++ b/openair1/PHY/LTE_TRANSPORT/pss.c
@@ -94,7 +94,7 @@ int generate_pss(mod_sym_t **txdataF,
   //a = (amp*ONE_OVER_SQRT2_Q15)>>15;
   //printf("[PSS] amp=%d, a=%d\n",amp,a);
 
-  Nsymb = (frame_parms->Ncp==0)?14:12;
+  Nsymb = (frame_parms->Ncp==NORMAL)?14:12;
 
   //for (aa=0;aa<frame_parms->nb_antennas_tx;aa++) {
   aa = 0;
diff --git a/openair1/PHY/LTE_TRANSPORT/sss.c b/openair1/PHY/LTE_TRANSPORT/sss.c
index 5b5f98b6ad..597cbf2777 100644
--- a/openair1/PHY/LTE_TRANSPORT/sss.c
+++ b/openair1/PHY/LTE_TRANSPORT/sss.c
@@ -64,7 +64,7 @@ int generate_sss(mod_sym_t **txdataF,
   else
     d = &d5_sss[62*(Nid2 + (Nid1*3))];
 
-  Nsymb = (frame_parms->Ncp==0)?14:12;
+  Nsymb = (frame_parms->Ncp==NORMAL)?14:12;
   k = frame_parms->ofdm_symbol_size-3*12+5;
   for (i=0;i<62;i++) {
     //for (aa=0;aa<frame_parms->nb_antennas_tx;aa++) {
@@ -166,7 +166,7 @@ int pss_sss_extract(PHY_VARS_UE *phy_vars_ue,
 
   int32_t **rxdataF =  phy_vars_ue->lte_ue_common_vars.rxdataF;
 
-  if (frame_parms->frame_type == 0) {
+  if (frame_parms->frame_type == FDD) {
     pss_symb = 6-frame_parms->Ncp;
     sss_symb = pss_symb-1;
   }
@@ -247,9 +247,9 @@ int rx_sss(PHY_VARS_UE *phy_vars_ue,int32_t *tot_metric,uint8_t *flip_max,uint8_
   int32_t metric;
   int16_t *d0,*d5;
 
-  if (phy_vars_ue->lte_frame_parms.frame_type == 0) { // FDD 
+  if (phy_vars_ue->lte_frame_parms.frame_type == FDD) { 
 #ifdef DEBUG_SSS
-    if (phy_vars_ue->lte_frame_parms.Ncp == 0)
+    if (phy_vars_ue->lte_frame_parms.Ncp == NORMAL)
       msg("[PHY][UE%d] Doing SSS for FDD Normal Prefix\n",phy_vars_ue->Mod_id);
     else
       msg("[PHY][UE%d] Doing SSS for FDD Extended Prefix\n",phy_vars_ue->Mod_id);
@@ -270,7 +270,7 @@ int rx_sss(PHY_VARS_UE *phy_vars_ue,int32_t *tot_metric,uint8_t *flip_max,uint8_
   }
   else {   // TDD
 #ifdef DEBUG_SSS
-    if (phy_vars_ue->lte_frame_parms.Ncp == 0)
+    if (phy_vars_ue->lte_frame_parms.Ncp == NORMAL)
       msg("[PHY][UE%d] Doing SSS for TDD Normal Prefix\n",phy_vars_ue->Mod_id);
     else
       msg("[PHY][UE%d] Doing SSS for TDD Extended Prefix\n",phy_vars_ue->Mod_id);
@@ -308,7 +308,7 @@ int rx_sss(PHY_VARS_UE *phy_vars_ue,int32_t *tot_metric,uint8_t *flip_max,uint8_
   
   //  write_output("sss0_comp0.m","sss0comp0",sss0_ext,72,1,1);
 
-  if (phy_vars_ue->lte_frame_parms.frame_type == 0) { // FDD 
+  if (phy_vars_ue->lte_frame_parms.frame_type == FDD) { // FDD 
 
     // SSS
     slot_fep(phy_vars_ue,
diff --git a/openair1/PHY/LTE_TRANSPORT/ulsch_coding.c b/openair1/PHY/LTE_TRANSPORT/ulsch_coding.c
index 596a39eb28..e1b3078a30 100644
--- a/openair1/PHY/LTE_TRANSPORT/ulsch_coding.c
+++ b/openair1/PHY/LTE_TRANSPORT/ulsch_coding.c
@@ -510,7 +510,7 @@ uint32_t ulsch_encoding(uint8_t *a,
       r_offset += lte_rate_matching_turbo(ulsch->harq_processes[harq_pid]->RTC[r],
 					  G,
 					  ulsch->harq_processes[harq_pid]->w[r],
-					  &ulsch->e[0],
+					  ulsch->e+r_offset,
 					  ulsch->harq_processes[harq_pid]->C, // C
 					  NSOFT,                    // Nsoft,
 					  ulsch->Mdlharq,
diff --git a/openair1/PHY/LTE_TRANSPORT/ulsch_decoding.c b/openair1/PHY/LTE_TRANSPORT/ulsch_decoding.c
index 8bb3a91139..e345121b07 100644
--- a/openair1/PHY/LTE_TRANSPORT/ulsch_decoding.c
+++ b/openair1/PHY/LTE_TRANSPORT/ulsch_decoding.c
@@ -1161,7 +1161,7 @@ unsigned int  ulsch_decoding(PHY_VARS_eNB *phy_vars_eNB,
 				   G,
 				   ulsch->harq_processes[harq_pid]->w[r],
 				   (uint8_t*) &dummy_w[r][0],
-				   ulsch->e,
+				   ulsch->e+r_offset,
 				   ulsch->harq_processes[harq_pid]->C,
 				   NSOFT,
 				   ulsch->Mdlharq,
diff --git a/openair1/PHY/LTE_TRANSPORT/ulsch_demodulation.c b/openair1/PHY/LTE_TRANSPORT/ulsch_demodulation.c
index ade4091a19..2707f8e206 100644
--- a/openair1/PHY/LTE_TRANSPORT/ulsch_demodulation.c
+++ b/openair1/PHY/LTE_TRANSPORT/ulsch_demodulation.c
@@ -588,53 +588,30 @@ void ulsch_extract_rbs_single(int32_t **rxdataF,
     msg("ulsch_extract_rbs_single: 2*nb_rb1 = %d, 2*nb_rb2 = %d\n",nb_rb1,nb_rb2);
 #endif
 
-#ifndef NEW_FFT
-    rxF_ext   = &rxdataF_ext[aarx][(symbol*frame_parms->N_RB_UL*12)*2];
-#else
     rxF_ext   = &rxdataF_ext[aarx][(symbol*frame_parms->N_RB_UL*12)];
-#endif    
     if (nb_rb1) {
-#ifndef NEW_FFT
-      rxF = &rxdataF[aarx][(first_rb*12 + frame_parms->first_carrier_offset + symbol*frame_parms->ofdm_symbol_size)*2];
-      memcpy(rxF_ext, rxF, nb_rb1*12*sizeof(int));
-      rxF_ext += nb_rb1*12;
-#else
       rxF = &rxdataF[aarx][(first_rb*12 + frame_parms->first_carrier_offset + symbol*frame_parms->ofdm_symbol_size)];
       memcpy(rxF_ext, rxF, nb_rb1*6*sizeof(int));
       rxF_ext += nb_rb1*6;
-#endif    
+
       if (nb_rb2)  {
 	//#ifdef OFDMA_ULSCH
 	//	rxF = &rxdataF[aarx][(1 + symbol*frame_parms->ofdm_symbol_size)*2];
 	//#else
-#ifndef NEW_FFT
-	rxF = &rxdataF[aarx][(symbol*frame_parms->ofdm_symbol_size)*2];
-	//#endif
-	memcpy(rxF_ext, rxF, nb_rb2*12*sizeof(int));
-	rxF_ext += nb_rb2*12;
-#else
 	rxF = &rxdataF[aarx][(symbol*frame_parms->ofdm_symbol_size)];
 	//#endif
 	memcpy(rxF_ext, rxF, nb_rb2*6*sizeof(int));
 	rxF_ext += nb_rb2*6;
-#endif
       } 
     }
     else { //there is only data in the second half
       //#ifdef OFDMA_ULSCH
       //      rxF = &rxdataF[aarx][(1 + 6*(2*first_rb - frame_parms->N_RB_UL) + symbol*frame_parms->ofdm_symbol_size)*2];
       //#else
-#ifndef NEW_FFT
-      rxF = &rxdataF[aarx][(6*(2*first_rb - frame_parms->N_RB_UL) + symbol*frame_parms->ofdm_symbol_size)*2];
-      //#endif
-      memcpy(rxF_ext, rxF, nb_rb2*12*sizeof(int));
-      rxF_ext += nb_rb2*12;
-#else
       rxF = &rxdataF[aarx][(6*(2*first_rb - frame_parms->N_RB_UL) + symbol*frame_parms->ofdm_symbol_size)];
       //#endif
       memcpy(rxF_ext, rxF, nb_rb2*6*sizeof(int));
       rxF_ext += nb_rb2*6;
-#endif
     }
   }
 
@@ -773,7 +750,7 @@ void ulsch_channel_compensation(int32_t **rxdataF_ext,
 	mmtmpU1 = _mm_packs_epi32(mmtmpU0,mmtmpU0);
 	ul_ch_mag128[2] = _mm_unpacklo_epi16(mmtmpU1,mmtmpU1);
 
-	//	printf("comp: symbol %d rb %d => %d,%d,%d (output_shift %d)\n",symbol,rb,*((int16_t*)&ul_ch_mag128[0]),*((int16_t*)&ul_ch_mag128[1]),*((int16_t*)&ul_ch_mag128[2]),output_shift);	
+	// printf("comp: symbol %d rb %d => %d,%d,%d (output_shift %d)\n",symbol,rb,*((int16_t*)&ul_ch_mag128[0]),*((int16_t*)&ul_ch_mag128[1]),*((int16_t*)&ul_ch_mag128[2]),output_shift);	
 #endif  
               
       // multiply by conjugated channel
@@ -1311,15 +1288,7 @@ void rx_ulsch(PHY_VARS_eNB *phy_vars_eNB,
   msg("rx_ulsch: eNB_id %d, harq_pid %d, nb_rb %d first_rb %d, cooperation %d\n",eNB_id,harq_pid,ulsch[UE_id]->harq_processes[harq_pid]->nb_rb,ulsch[UE_id]->harq_processes[harq_pid]->first_rb, cooperation_flag);
 #endif //DEBUG_ULSCH
 
-#ifndef NEW_FFT 
-  if ( (frame_parms->ofdm_symbol_size == 128) ||
-       (frame_parms->ofdm_symbol_size == 512) )
-    rx_power_correction = 2;
-  else
-    rx_power_correction = 1;
-#else
-    rx_power_correction = 1;
-#endif
+  rx_power_correction = 1;
 
   for (l=0;l<(frame_parms->symbols_per_tti-ulsch[UE_id]->srs_active);l++) {
           
@@ -1347,14 +1316,6 @@ void rx_ulsch(PHY_VARS_eNB *phy_vars_eNB,
 			      l/(frame_parms->symbols_per_tti/2),
 			      cooperation_flag);
 
-#ifndef NEW_FFT
-    ulsch_correct_ext(eNB_pusch_vars->rxdataF_ext[eNB_id],
-		      eNB_pusch_vars->rxdataF_ext2[eNB_id],
-		      l,
-		      frame_parms,
-		      ulsch[UE_id]->harq_processes[harq_pid]->nb_rb);  
-#endif
-
     if(cooperation_flag == 2)
       {
 	for (i=0;i<frame_parms->nb_antennas_rx;i++){
@@ -1425,11 +1386,9 @@ void rx_ulsch(PHY_VARS_eNB *phy_vars_eNB,
 	avgs = cmax(avgs,avgU[(aarx<<1)]);
   
       //      log2_maxh = 4+(log2_approx(avgs)/2);
-#ifdef NEW_FFT
+
       log2_maxh = (log2_approx(avgs)/2)+ log2_approx(frame_parms->nb_antennas_rx-1)+4;
-#else
-      log2_maxh = (log2_approx(avgs)/2)+ log2_approx(frame_parms->nb_antennas_rx-1)+3;
-#endif
+
 #ifdef DEBUG_ULSCH
       msg("[ULSCH] log2_maxh = %d (%d,%d)\n",log2_maxh,avgU[0],avgs);
 #endif
@@ -1446,11 +1405,7 @@ void rx_ulsch(PHY_VARS_eNB *phy_vars_eNB,
       {
 
 	ulsch_channel_compensation_alamouti(
-#ifndef NEW_FFT
-					    eNB_pusch_vars->rxdataF_ext2[eNB_id],
-#else
 					    eNB_pusch_vars->rxdataF_ext[eNB_id],
-#endif
 					    eNB_pusch_vars->drs_ch_estimates_0[eNB_id],
 					    eNB_pusch_vars->drs_ch_estimates_1[eNB_id],
 					    eNB_pusch_vars->ul_ch_mag_0[eNB_id],
@@ -1481,11 +1436,7 @@ void rx_ulsch(PHY_VARS_eNB *phy_vars_eNB,
     else
       {
 	ulsch_channel_compensation(
-#ifndef NEW_FFT
-				   eNB_pusch_vars->rxdataF_ext2[eNB_id],
-#else
 				   eNB_pusch_vars->rxdataF_ext[eNB_id],
-#endif				   
 				   eNB_pusch_vars->drs_ch_estimates[eNB_id],
 				   eNB_pusch_vars->ul_ch_mag[eNB_id],
 				   eNB_pusch_vars->ul_ch_magb[eNB_id],
@@ -1613,14 +1564,6 @@ void dump_ulsch(PHY_VARS_eNB *PHY_vars_eNB,uint8_t subframe, uint8_t UE_id) {
   write_output("/tmp/rxsig0.m","rxs0", &PHY_vars_eNB->lte_eNB_common_vars.rxdata[0][0][0],PHY_vars_eNB->lte_frame_parms.samples_per_tti*10,1,1);
   if (PHY_vars_eNB->lte_frame_parms.nb_antennas_rx>1)
     write_output("/tmp/rxsig1.m","rxs1", &PHY_vars_eNB->lte_eNB_common_vars.rxdata[0][1][0],PHY_vars_eNB->lte_frame_parms.samples_per_tti*10,1,1);
-#ifndef NEW_FFT
-  write_output("/tmp/rxsigF0.m","rxsF0", &PHY_vars_eNB->lte_eNB_common_vars.rxdataF[0][0][0],PHY_vars_eNB->lte_frame_parms.ofdm_symbol_size*nsymb*2,2,1);
-  if (PHY_vars_eNB->lte_frame_parms.nb_antennas_rx>1)
-    write_output("/tmp/rxsigF1.m","rxsF1", &PHY_vars_eNB->lte_eNB_common_vars.rxdataF[0][1][0],PHY_vars_eNB->lte_frame_parms.ofdm_symbol_size*nsymb*2,2,1);
-  write_output("/tmp/rxsigF0_ext.m","rxsF0_ext", &PHY_vars_eNB->lte_eNB_pusch_vars[UE_id]->rxdataF_ext[0][0][0],PHY_vars_eNB->lte_frame_parms.N_RB_UL*12*nsymb*2,2,1);
-  if (PHY_vars_eNB->lte_frame_parms.nb_antennas_rx>1)
-    write_output("/tmp/rxsigF1_ext.m","rxsF1_ext", &PHY_vars_eNB->lte_eNB_pusch_vars[UE_id]->rxdataF_ext[1][0][0],PHY_vars_eNB->lte_frame_parms.N_RB_UL*12*nsymb*2,2,1);
-#else
   write_output("/tmp/rxsigF0.m","rxsF0", &PHY_vars_eNB->lte_eNB_common_vars.rxdataF[0][0][0],PHY_vars_eNB->lte_frame_parms.ofdm_symbol_size*nsymb,1,1);
   if (PHY_vars_eNB->lte_frame_parms.nb_antennas_rx>1)
     write_output("/tmp/rxsigF1.m","rxsF1", &PHY_vars_eNB->lte_eNB_common_vars.rxdataF[0][1][0],PHY_vars_eNB->lte_frame_parms.ofdm_symbol_size*nsymb,1,1);
@@ -1641,6 +1584,5 @@ void dump_ulsch(PHY_VARS_eNB *PHY_vars_eNB,uint8_t subframe, uint8_t UE_id) {
   write_output("/tmp/ulsch_ch_mag.m","ulsch_ch_mag",&PHY_vars_eNB->lte_eNB_pusch_vars[UE_id]->ul_ch_mag[0][0][0],PHY_vars_eNB->lte_frame_parms.N_RB_UL*12*nsymb,1,1);	  
   //  write_output("ulsch_ch_mag1.m","ulsch_ch_mag1",&PHY_vars_eNB->lte_eNB_pusch_vars[UE_id]->ul_ch_mag[0][1][0],PHY_vars_eNB->lte_frame_parms.N_RB_UL*12*nsymb,1,1);	  
 #endif	 
-#endif 
 }
 
diff --git a/openair1/PHY/Makefile.inc b/openair1/PHY/Makefile.inc
index b924b02f6d..e16c244688 100644
--- a/openair1/PHY/Makefile.inc
+++ b/openair1/PHY/Makefile.inc
@@ -46,6 +46,7 @@ PHY_OBJS += $(TOP_DIR)/PHY/LTE_ESTIMATION/lte_ue_measurements.o
 PHY_OBJS += $(TOP_DIR)/PHY/LTE_ESTIMATION/lte_eNB_measurements.o
 PHY_OBJS += $(TOP_DIR)/PHY/LTE_ESTIMATION/adjust_gain.o
 PHY_OBJS += $(TOP_DIR)/PHY/LTE_REFSIG/lte_dl_cell_spec.o
+PHY_OBJS += $(TOP_DIR)/PHY/LTE_REFSIG/lte_dl_uespec.o
 PHY_OBJS += $(TOP_DIR)/PHY/LTE_REFSIG/lte_gold.o
 PHY_OBJS += $(TOP_DIR)/PHY/LTE_REFSIG/lte_gold_mbsfn.o
 PHY_OBJS += $(TOP_DIR)/PHY/LTE_REFSIG/lte_dl_mbsfn.o
diff --git a/openair1/PHY/TOOLS/Makefile b/openair1/PHY/TOOLS/Makefile
index d7ea1bf76f..a4ce7dbdc5 100644
--- a/openair1/PHY/TOOLS/Makefile
+++ b/openair1/PHY/TOOLS/Makefile
@@ -1,5 +1,5 @@
 lte_dfts: lte_dfts.c
-	gcc -O3 -msse4.1 -o lte_dfts lte_dfts.c fft.c ../../SIMULATION/TOOLS/taus.c -I$$OPENAIR1_DIR -I$$OPENAIR_TARGETS -DUSER_MODE -DMR_MAIN -DNB_ANTENNAS_RX=1 # -DD256STATS #-DD64STATS
+	gcc -O3 -msse4.1 -o lte_dfts lte_dfts.c fft.c ../../SIMULATION/TOOLS/taus.c -I$$OPENAIR1_DIR -I$$OPENAIR_TARGETS -I$$OPENAIR2_DIR/COMMON -DUSER_MODE -DMR_MAIN -DNB_ANTENNAS_RX=1 # -DD256STATS #-DD64STATS
 
 dft_cycles: lte_dfts
 	./lte_dfts | egrep cycles
diff --git a/openair1/PHY/TOOLS/lte_dfts.c b/openair1/PHY/TOOLS/lte_dfts.c
index 20250f50ec..8f0d6e2981 100644
--- a/openair1/PHY/TOOLS/lte_dfts.c
+++ b/openair1/PHY/TOOLS/lte_dfts.c
@@ -51,9 +51,9 @@ static inline void cmac(__m128i a,__m128i b, __m128i *re32, __m128i *im32) {
   cmac_tmp_re32  = _mm_madd_epi16(a,cmac_tmp);
 
 
-  cmac_tmp    = _mm_shufflelo_epi16(b,_MM_SHUFFLE(2,3,0,1));
-  cmac_tmp    = _mm_shufflehi_epi16(cmac_tmp,_MM_SHUFFLE(2,3,0,1));
-
+  //  cmac_tmp    = _mm_shufflelo_epi16(b,_MM_SHUFFLE(2,3,0,1));
+  //  cmac_tmp    = _mm_shufflehi_epi16(cmac_tmp,_MM_SHUFFLE(2,3,0,1));
+  cmac_tmp = _mm_shuffle_epi8(b,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   cmac_tmp_im32  = _mm_madd_epi16(cmac_tmp,a);
 
   *re32 = _mm_add_epi32(*re32,cmac_tmp_re32);
@@ -71,8 +71,9 @@ static inline void cmult(__m128i a,__m128i b, __m128i *re32, __m128i *im32) {
 
   mmtmpb    = _mm_sign_epi16(b,*(__m128i*)reflip);
   *re32     = _mm_madd_epi16(a,mmtmpb);
-  mmtmpb    = _mm_shufflelo_epi16(b,_MM_SHUFFLE(2,3,0,1));
-  mmtmpb    = _mm_shufflehi_epi16(mmtmpb,_MM_SHUFFLE(2,3,0,1));
+  //  mmtmpb    = _mm_shufflelo_epi16(b,_MM_SHUFFLE(2,3,0,1));
+  //  mmtmpb    = _mm_shufflehi_epi16(mmtmpb,_MM_SHUFFLE(2,3,0,1));
+  mmtmpb        = _mm_shuffle_epi8(b,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   *im32  = _mm_madd_epi16(a,mmtmpb);
 
 }
@@ -85,8 +86,9 @@ static inline void cmultc(__m128i a,__m128i b, __m128i *re32, __m128i *im32) {
 
   *re32     = _mm_madd_epi16(a,b);
   mmtmpb    = _mm_sign_epi16(b,*(__m128i*)reflip);
-  mmtmpb    = _mm_shufflelo_epi16(mmtmpb,_MM_SHUFFLE(2,3,0,1));
-  mmtmpb    = _mm_shufflehi_epi16(mmtmpb,_MM_SHUFFLE(2,3,0,1));
+  //  mmtmpb    = _mm_shufflelo_epi16(mmtmpb,_MM_SHUFFLE(2,3,0,1));
+  //  mmtmpb    = _mm_shufflehi_epi16(mmtmpb,_MM_SHUFFLE(2,3,0,1));
+  mmtmpb    = _mm_shuffle_epi8(mmtmpb,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   *im32  = _mm_madd_epi16(a,mmtmpb);
 
 }
@@ -414,11 +416,13 @@ static inline void bfly4_tw1(__m128i *x0,__m128i *x1,__m128i *x2,__m128i *x3,
   *(y0) = _mm_adds_epi16(*(x0),_mm_adds_epi16(*(x1),_mm_adds_epi16(*(x2),*(x3)))); 
 
   x1_flip = _mm_sign_epi16(*(x1),*(__m128i*)conjugatedft);
-  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
-  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  x1_flip = _mm_shuffle_epi8(x1_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x3_flip = _mm_sign_epi16(*(x3),*(__m128i*)conjugatedft);
-  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  x3_flip = _mm_shuffle_epi8(x3_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   *(y1)   = _mm_adds_epi16(*(x0),_mm_subs_epi16(x1_flip,_mm_adds_epi16(*(x2),x3_flip)));
   *(y2)   = _mm_subs_epi16(*(x0),_mm_subs_epi16(*(x1),_mm_subs_epi16(*(x2),*(x3))));
   *(y3)   = _mm_subs_epi16(*(x0),_mm_adds_epi16(x1_flip,_mm_subs_epi16(*(x2),x3_flip)));
@@ -435,11 +439,13 @@ static inline void ibfly4_tw1(__m128i *x0,__m128i *x1,__m128i *x2,__m128i *x3,
   *(y0) = _mm_adds_epi16(*(x0),_mm_adds_epi16(*(x1),_mm_adds_epi16(*(x2),*(x3)))); 
 
   x1_flip = _mm_sign_epi16(*(x1),*(__m128i*)conjugatedft);
-  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
-  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  x1_flip = _mm_shuffle_epi8(x1_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x3_flip = _mm_sign_epi16(*(x3),*(__m128i*)conjugatedft);
-  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  x3_flip = _mm_shuffle_epi8(x3_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   *(y1)   = _mm_subs_epi16(*(x0),_mm_adds_epi16(x1_flip,_mm_subs_epi16(*(x2),x3_flip)));
   *(y2)   = _mm_subs_epi16(*(x0),_mm_subs_epi16(*(x1),_mm_subs_epi16(*(x2),*(x3))));
   *(y3)   = _mm_adds_epi16(*(x0),_mm_subs_epi16(x1_flip,_mm_adds_epi16(*(x2),x3_flip)));
@@ -474,12 +480,13 @@ static inline void bfly4_16(__m128i *x0,__m128i *x1,__m128i *x2,__m128i *x3,
   *(y2)   = _mm_subs_epi16(x02t,x13t);
 
   x1_flip = _mm_sign_epi16(x1t,*(__m128i*)conjugatedft);
-  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
-  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  x1_flip = _mm_shuffle_epi8(x1_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x3_flip = _mm_sign_epi16(x3t,*(__m128i*)conjugatedft);
-  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-
+  //  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  x3_flip = _mm_shuffle_epi8(x3_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x02t  = _mm_subs_epi16(*(x0),x2t);
   x13t  = _mm_subs_epi16(x1_flip,x3_flip);
   /*
@@ -520,12 +527,13 @@ static inline void ibfly4_16(__m128i *x0,__m128i *x1,__m128i *x2,__m128i *x3,
   *(y2)   = _mm_subs_epi16(x02t,x13t);
 
   x1_flip = _mm_sign_epi16(x1t,*(__m128i*)conjugatedft);
-  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
-  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  x1_flip = _mm_shuffle_epi8(x1_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x3_flip = _mm_sign_epi16(x3t,*(__m128i*)conjugatedft);
-  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-
+  //  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  x3_flip = _mm_shuffle_epi8(x3_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x02t  = _mm_subs_epi16(*(x0),x2t);
   x13t  = _mm_subs_epi16(x1_flip,x3_flip);
   /*
@@ -713,12 +721,13 @@ static inline void dft16(int16_t *x,int16_t *y) {
   xtmp2   = _mm_subs_epi16(x128[0],_mm_subs_epi16(x128[1],_mm_subs_epi16(x128[2],x128[3])));
   */
   x1_flip = _mm_sign_epi16(x128[1],*(__m128i*)conjugatedft);
-  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
-  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  x1_flip = _mm_shuffle_epi8(x1_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x3_flip = _mm_sign_epi16(x128[3],*(__m128i*)conjugatedft);
-  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-
+  //  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  x3_flip = _mm_shuffle_epi8(x3_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x02t    = _mm_subs_epi16(x128[0],x128[2]);
   x13t    = _mm_subs_epi16(x1_flip,x3_flip);
   xtmp1   = _mm_adds_epi16(x02t,x13t);  // x0 + x1f - x2 - x3f
@@ -753,12 +762,13 @@ static inline void dft16(int16_t *x,int16_t *y) {
   */
 
   x1_flip = _mm_sign_epi16(xtmp1,*(__m128i*)conjugatedft);
-  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
-  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  x1_flip = _mm_shuffle_epi8(x1_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x3_flip = _mm_sign_epi16(xtmp3,*(__m128i*)conjugatedft);
-  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-
+  //  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  x3_flip = _mm_shuffle_epi8(x3_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x02t    = _mm_subs_epi16(xtmp0,xtmp2);
   x13t    = _mm_subs_epi16(x1_flip,x3_flip);
   y128[1] = _mm_adds_epi16(x02t,x13t);  // x0 + x1f - x2 - x3f
@@ -801,12 +811,13 @@ static inline void idft16(int16_t *x,int16_t *y) {
   xtmp2   = _mm_subs_epi16(x128[0],_mm_subs_epi16(x128[1],_mm_subs_epi16(x128[2],x128[3])));
   */
   x1_flip = _mm_sign_epi16(x128[1],*(__m128i*)conjugatedft);
-  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
-  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  x1_flip = _mm_shuffle_epi8(x1_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x3_flip = _mm_sign_epi16(x128[3],*(__m128i*)conjugatedft);
-  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-
+  //  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  x3_flip = _mm_shuffle_epi8(x3_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x02t    = _mm_subs_epi16(x128[0],x128[2]);
   x13t    = _mm_subs_epi16(x1_flip,x3_flip);
   xtmp3   = _mm_adds_epi16(x02t,x13t);  // x0 + x1f - x2 - x3f
@@ -841,12 +852,13 @@ static inline void idft16(int16_t *x,int16_t *y) {
   */
 
   x1_flip = _mm_sign_epi16(xtmp1,*(__m128i*)conjugatedft);
-  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
-  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflelo_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x1_flip = _mm_shufflehi_epi16(x1_flip,_MM_SHUFFLE(2,3,0,1));
+  x1_flip = _mm_shuffle_epi8(x1_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x3_flip = _mm_sign_epi16(xtmp3,*(__m128i*)conjugatedft);
-  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
-
+  //  x3_flip = _mm_shufflelo_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  //  x3_flip = _mm_shufflehi_epi16(x3_flip,_MM_SHUFFLE(2,3,0,1));
+  x3_flip = _mm_shuffle_epi8(x3_flip,_mm_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2));
   x02t    = _mm_subs_epi16(xtmp0,xtmp2);
   x13t    = _mm_subs_epi16(x1_flip,x3_flip);
   y128[3] = _mm_adds_epi16(x02t,x13t);  // x0 + x1f - x2 - x3f
@@ -14851,6 +14863,7 @@ void dft1200(int16_t *x,int16_t *y,unsigned char scale_flag){
 
 #ifdef MR_MAIN
 
+int opp_enabled=1;
 int main(int argc, char**argv) {
 
 
@@ -15188,7 +15201,7 @@ int main(int argc, char**argv) {
   dft64((int16_t *)x,(int16_t *)y,1);
   dft64((int16_t *)x,(int16_t *)y,1); 
   reset_meas(&ts);
-  for (i=0;i<1000;i++) {
+  for (i=0;i<10000000;i++) {
     start_meas(&ts);
     dft64((int16_t *)x,(int16_t *)y,1);
     stop_meas(&ts);
diff --git a/openair1/PHY/defs.h b/openair1/PHY/defs.h
index fa1f154829..4ac5b6f867 100755
--- a/openair1/PHY/defs.h
+++ b/openair1/PHY/defs.h
@@ -118,9 +118,9 @@
 
 
 #ifdef CBMIMO1
-#define openair_get_mbox() (*(unsigned int *)mbox)
+#define openair_get_mbox() (*(uint32_t *)mbox)
 #else //CBMIMO1
-#define openair_get_mbox() (*(unsigned int *)PHY_vars->mbox>>1)
+#define openair_get_mbox() (*(uint32_t *)PHY_vars->mbox>>1)
 #endif //CBMIMO1
 
 #endif // USERMODE
@@ -183,7 +183,7 @@ typedef struct PHY_VARS_eNB_s{
   /// Module ID indicator for this instance
   module_id_t               Mod_id;
   uint8_t                   local_flag;
-  unsigned int         rx_total_gain_eNB_dB;
+  uint32_t         rx_total_gain_eNB_dB;
   frame_t              frame;
   LTE_DL_FRAME_PARMS   lte_frame_parms;
   PHY_MEASUREMENTS_eNB PHY_measurements_eNB[NUMBER_OF_eNB_SECTORS_MAX]; /// Measurement variables 
@@ -201,10 +201,13 @@ typedef struct PHY_VARS_eNB_s{
   LTE_eNB_UE_stats    *eNB_UE_stats_ptr[NUMBER_OF_UE_MAX];
 
   /// cell-specific reference symbols
-  unsigned int         lte_gold_table[20][2][14];
+  uint32_t         lte_gold_table[20][2][14];
+
+  /// UE-specific reference symbols (p=7...14), TM 8/9/10
+  uint32_t         lte_gold_uespec_table[2][20][2][21];
   
   /// mbsfn reference symbols
-  unsigned int         lte_gold_mbsfn_table[10][3][42];
+  uint32_t         lte_gold_mbsfn_table[10][3][42];
   
   uint32_t X_u[64][839];
 
@@ -214,7 +217,7 @@ typedef struct PHY_VARS_eNB_s{
   /// Indicator set to 0 after first SR
   uint8_t first_sr[NUMBER_OF_UE_MAX];
 
-  unsigned int max_peak_val; 
+  uint32_t max_peak_val; 
   int max_eNB_id, max_sync_pos;
 
 
@@ -286,21 +289,21 @@ typedef struct PHY_VARS_eNB_s{
 
 
   ///check for Total Transmissions
-  unsigned int check_for_total_transmissions;
+  uint32_t check_for_total_transmissions;
 
   ///check for MU-MIMO Transmissions
-  unsigned int check_for_MUMIMO_transmissions;
+  uint32_t check_for_MUMIMO_transmissions;
 
   ///check for SU-MIMO Transmissions
-  unsigned int check_for_SUMIMO_transmissions;
+  uint32_t check_for_SUMIMO_transmissions;
 
   ///check for FULL MU-MIMO Transmissions
-  unsigned int  FULL_MUMIMO_transmissions;
+  uint32_t  FULL_MUMIMO_transmissions;
 
   /// Counter for total bitrate, bits and throughput in downlink
-  unsigned int total_dlsch_bitrate;
-  unsigned int total_transmitted_bits;
-  unsigned int total_system_throughput;
+  uint32_t total_dlsch_bitrate;
+  uint32_t total_transmitted_bits;
+  uint32_t total_system_throughput;
   
   time_stats_t phy_proc;
   time_stats_t phy_proc_tx;
@@ -356,12 +359,12 @@ typedef struct
   /// Module ID indicator for this instance
   uint8_t Mod_id;
   uint8_t local_flag;
-  unsigned int tx_total_gain_dB;
-  unsigned int rx_total_gain_dB; ///this is a function of rx_gain_mode (and the corresponding gain) and the rx_gain of the card
+  uint32_t tx_total_gain_dB;
+  uint32_t rx_total_gain_dB; ///this is a function of rx_gain_mode (and the corresponding gain) and the rx_gain of the card
   rx_gain_t rx_gain_mode[4];
-  unsigned int rx_gain_max[4];
-  unsigned int rx_gain_med[4];
-  unsigned int rx_gain_byp[4];
+  uint32_t rx_gain_max[4];
+  uint32_t rx_gain_med[4];
+  uint32_t rx_gain_byp[4];
   int8_t tx_power_dBm;
   int8_t tx_power_max_dBm;
   uint32_t frame;
@@ -395,26 +398,29 @@ typedef struct
   UE_MODE_t        UE_mode[NUMBER_OF_CONNECTED_eNB_MAX];
   int8_t               g_pucch[NUMBER_OF_CONNECTED_eNB_MAX];
   /// cell-specific reference symbols
-  unsigned int lte_gold_table[7][20][2][14];
+  uint32_t lte_gold_table[7][20][2][14];
 
-/// mbsfn reference symbols
-  unsigned int lte_gold_mbsfn_table[10][3][42];
+  /// ue-specific reference symbols
+  uint32_t         lte_gold_uespec_table[2][20][2][21];
+
+  /// mbsfn reference symbols
+  uint32_t lte_gold_mbsfn_table[10][3][42];
   
   uint32_t X_u[64][839];
 
   char ulsch_no_allocation_counter[NUMBER_OF_CONNECTED_eNB_MAX];
 
   unsigned char ulsch_ue_Msg3_active[NUMBER_OF_CONNECTED_eNB_MAX];
-  unsigned int  ulsch_ue_Msg3_frame[NUMBER_OF_CONNECTED_eNB_MAX];
+  uint32_t  ulsch_ue_Msg3_frame[NUMBER_OF_CONNECTED_eNB_MAX];
   unsigned char ulsch_ue_Msg3_subframe[NUMBER_OF_CONNECTED_eNB_MAX];
   //  unsigned char Msg3_timer[NUMBER_OF_CONNECTED_eNB_MAX];
   //unsigned char *Msg3_ptr[NUMBER_OF_CONNECTED_eNB_MAX];
   PRACH_RESOURCES_t *prach_resources[NUMBER_OF_CONNECTED_eNB_MAX];
   int turbo_iterations, turbo_cntl_iterations;
-  unsigned int total_TBS[NUMBER_OF_CONNECTED_eNB_MAX];
-  unsigned int total_TBS_last[NUMBER_OF_CONNECTED_eNB_MAX];
-  unsigned int bitrate[NUMBER_OF_CONNECTED_eNB_MAX];
-  unsigned int total_received_bits[NUMBER_OF_CONNECTED_eNB_MAX];
+  uint32_t total_TBS[NUMBER_OF_CONNECTED_eNB_MAX];
+  uint32_t total_TBS_last[NUMBER_OF_CONNECTED_eNB_MAX];
+  uint32_t bitrate[NUMBER_OF_CONNECTED_eNB_MAX];
+  uint32_t total_received_bits[NUMBER_OF_CONNECTED_eNB_MAX];
   int dlsch_errors[NUMBER_OF_CONNECTED_eNB_MAX];
   int dlsch_errors_last[NUMBER_OF_CONNECTED_eNB_MAX];
   int dlsch_received[NUMBER_OF_CONNECTED_eNB_MAX];
diff --git a/openair1/PHY/impl_defs_lte.h b/openair1/PHY/impl_defs_lte.h
index 7138d2890c..18707c54b6 100644
--- a/openair1/PHY/impl_defs_lte.h
+++ b/openair1/PHY/impl_defs_lte.h
@@ -534,7 +534,9 @@ typedef enum {
   PUSCH_PRECODING1=8,
   DUALSTREAM_UNIFORM_PRECODING1=9,
   DUALSTREAM_UNIFORM_PRECODINGj=10,
-  DUALSTREAM_PUSCH_PRECODING=11
+  DUALSTREAM_PUSCH_PRECODING=11,
+  TM8=12,
+  TM9_10=13
 } MIMO_mode_t;
 
 typedef struct{
diff --git a/openair1/SCHED/phy_procedures_lte_eNb.c b/openair1/SCHED/phy_procedures_lte_eNb.c
index c18ff68cb1..02c5a0ab28 100755
--- a/openair1/SCHED/phy_procedures_lte_eNb.c
+++ b/openair1/SCHED/phy_procedures_lte_eNb.c
@@ -1791,6 +1791,7 @@ void phy_procedures_eNB_TX(unsigned char next_slot,PHY_VARS_eNB *phy_vars_eNB,ui
 				 phy_vars_eNB->dlsch_eNB_SI->nb_rb,
 				 phy_vars_eNB->dlsch_eNB_SI->rb_alloc,
 				 get_Qm(phy_vars_eNB->dlsch_eNB_SI->harq_processes[0]->mcs),
+				 1,
 				 num_pdcch_symbols,phy_vars_eNB->frame,next_slot>>1),
 			   0,
 			   next_slot);      
@@ -1909,6 +1910,7 @@ void phy_procedures_eNB_TX(unsigned char next_slot,PHY_VARS_eNB *phy_vars_eNB,ui
 				   phy_vars_eNB->dlsch_eNB_ra->nb_rb,
 				   phy_vars_eNB->dlsch_eNB_ra->rb_alloc,
 				   get_Qm(phy_vars_eNB->dlsch_eNB_ra->harq_processes[0]->mcs),
+				   1,
 				   num_pdcch_symbols,phy_vars_eNB->frame,next_slot>>1),
 			     0,
 			     next_slot);
@@ -1958,6 +1960,7 @@ void phy_procedures_eNB_TX(unsigned char next_slot,PHY_VARS_eNB *phy_vars_eNB,ui
 		    phy_vars_eNB->dlsch_eNB[(uint8_t)UE_id][0]->nb_rb,
 		    phy_vars_eNB->dlsch_eNB[(uint8_t)UE_id][0]->rb_alloc,
 		    get_Qm(phy_vars_eNB->dlsch_eNB[(uint8_t)UE_id][0]->harq_processes[harq_pid]->mcs),
+		    phy_vars_eNB->dlsch_eNB[(uint8_t)UE_id][0]->harq_processes[harq_pid]->Nl,
 		    num_pdcch_symbols,phy_vars_eNB->frame,next_slot>>1),
 	      phy_vars_eNB->dlsch_eNB[(uint8_t)UE_id][0]->nb_rb,
 	      phy_vars_eNB->dlsch_eNB[(uint8_t)UE_id][0]->harq_processes[harq_pid]->mcs,
@@ -2031,6 +2034,7 @@ void phy_procedures_eNB_TX(unsigned char next_slot,PHY_VARS_eNB *phy_vars_eNB,ui
 				   phy_vars_eNB->dlsch_eNB[(uint8_t)UE_id][0]->nb_rb,
 				   phy_vars_eNB->dlsch_eNB[(uint8_t)UE_id][0]->rb_alloc,
 				   get_Qm(phy_vars_eNB->dlsch_eNB[(uint8_t)UE_id][0]->harq_processes[harq_pid]->mcs),
+				   phy_vars_eNB->dlsch_eNB[(uint8_t)UE_id][0]->harq_processes[harq_pid]->Nl,
 				   num_pdcch_symbols,phy_vars_eNB->frame,next_slot>>1),
 			     0,
 			     next_slot);      
diff --git a/openair1/SCHED/phy_procedures_lte_ue.c b/openair1/SCHED/phy_procedures_lte_ue.c
index 9ba21bf93f..fc75549179 100755
--- a/openair1/SCHED/phy_procedures_lte_ue.c
+++ b/openair1/SCHED/phy_procedures_lte_ue.c
@@ -149,7 +149,8 @@ void dump_dlsch(PHY_VARS_UE *phy_vars_ue,uint8_t eNB_id,uint8_t subframe,uint8_t
   coded_bits_per_codeword = get_G(&phy_vars_ue->lte_frame_parms,
                                   phy_vars_ue->dlsch_ue[eNB_id][0]->harq_processes[harq_pid]->nb_rb,
                                   phy_vars_ue->dlsch_ue[eNB_id][0]->harq_processes[harq_pid]->rb_alloc,
-                                  get_Qm(phy_vars_ue->dlsch_ue[eNB_id][0]->harq_processes[harq_pid]->mcs),  
+                                  get_Qm(phy_vars_ue->dlsch_ue[eNB_id][0]->harq_processes[harq_pid]->mcs),
+				  phy_vars_ue->dlsch_ue[eNB_id][0]->harq_processes[harq_pid]->Nl,    
                                   phy_vars_ue->lte_ue_pdcch_vars[eNB_id]->num_pdcch_symbols,
                                   phy_vars_ue->frame,subframe);
 
@@ -176,7 +177,8 @@ void dump_dlsch_SI(PHY_VARS_UE *phy_vars_ue,uint8_t eNB_id,uint8_t subframe) {
   coded_bits_per_codeword = get_G(&phy_vars_ue->lte_frame_parms,
                                   phy_vars_ue->dlsch_ue_SI[eNB_id]->harq_processes[0]->nb_rb,
                                   phy_vars_ue->dlsch_ue_SI[eNB_id]->harq_processes[0]->rb_alloc,
-                                  get_Qm(phy_vars_ue->dlsch_ue_SI[eNB_id]->harq_processes[0]->mcs),  
+                                  get_Qm(phy_vars_ue->dlsch_ue_SI[eNB_id]->harq_processes[0]->mcs),
+				  1,
                                   phy_vars_ue->lte_ue_pdcch_vars[eNB_id]->num_pdcch_symbols,
                                   phy_vars_ue->frame,subframe);
   LOG_D(PHY,"[UE %d] Dumping dlsch_SI : nb_rb %d, mcs %d, nb_rb %d, num_pdcch_symbols %d,G %d\n",
@@ -230,6 +232,7 @@ void dump_dlsch_ra(PHY_VARS_UE *phy_vars_ue,uint8_t eNB_id,uint8_t subframe) {
                                   phy_vars_ue->dlsch_ue_ra[eNB_id]->harq_processes[0]->nb_rb,
                                   phy_vars_ue->dlsch_ue_ra[eNB_id]->harq_processes[0]->rb_alloc,
                                   get_Qm(phy_vars_ue->dlsch_ue_ra[eNB_id]->harq_processes[0]->mcs),  
+				  1,
                                   phy_vars_ue->lte_ue_pdcch_vars[eNB_id]->num_pdcch_symbols,
                                   phy_vars_ue->frame,subframe);
   LOG_D(PHY,"[UE %d] Dumping dlsch_ra : nb_rb %d, mcs %d, nb_rb %d, num_pdcch_symbols %d,G %d\n",
@@ -2465,6 +2468,7 @@ int lte_ue_pdcch_procedures(uint8_t eNB_id,uint8_t last_slot, PHY_VARS_UE *phy_v
                                                                                   phy_vars_ue->dlsch_ue[eNB_id][0]->harq_processes[harq_pid]->nb_rb,
                                                                                   phy_vars_ue->dlsch_ue[eNB_id][0]->harq_processes[harq_pid]->rb_alloc,
                                                                                   get_Qm(phy_vars_ue->dlsch_ue[eNB_id][0]->harq_processes[harq_pid]->mcs),
+										  phy_vars_ue->dlsch_ue[eNB_id][0]->harq_processes[harq_pid]->Nl,
                                                                                   phy_vars_ue->lte_ue_pdcch_vars[eNB_id]->num_pdcch_symbols,
                                                                                   phy_vars_ue->frame,(((last_slot>>1)==0) ? 9 : ((last_slot>>1)-1)));
 	    start_meas(&phy_vars_ue->dlsch_unscrambling_stats);
@@ -2614,7 +2618,7 @@ int lte_ue_pdcch_procedures(uint8_t eNB_id,uint8_t last_slot, PHY_VARS_UE *phy_v
 	      get_G(&phy_vars_ue->lte_frame_parms,
 		    phy_vars_ue->dlsch_ue_SI[eNB_id]->harq_processes[0]->nb_rb,
 		    phy_vars_ue->dlsch_ue_SI[eNB_id]->harq_processes[0]->rb_alloc,
-		    get_Qm(phy_vars_ue->dlsch_ue_SI[eNB_id]->harq_processes[0]->mcs),
+		    get_Qm(phy_vars_ue->dlsch_ue_SI[eNB_id]->harq_processes[0]->mcs),1,
 		    phy_vars_ue->lte_ue_pdcch_vars[eNB_id]->num_pdcch_symbols,
 		    phy_vars_ue->frame,(((last_slot>>1)==0) ? 9 : ((last_slot>>1)-1)));
 	    
@@ -2749,7 +2753,7 @@ int lte_ue_pdcch_procedures(uint8_t eNB_id,uint8_t last_slot, PHY_VARS_UE *phy_v
         phy_vars_ue->dlsch_ue_ra[eNB_id]->harq_processes[0]->G = get_G(&phy_vars_ue->lte_frame_parms,
                                                                        phy_vars_ue->dlsch_ue_ra[eNB_id]->harq_processes[0]->nb_rb,
                                                                        phy_vars_ue->dlsch_ue_ra[eNB_id]->harq_processes[0]->rb_alloc,
-                                                                       get_Qm(phy_vars_ue->dlsch_ue_ra[eNB_id]->harq_processes[0]->mcs),
+                                                                       get_Qm(phy_vars_ue->dlsch_ue_ra[eNB_id]->harq_processes[0]->mcs),1,
                                                                        phy_vars_ue->lte_ue_pdcch_vars[eNB_id]->num_pdcch_symbols,
                                                                        phy_vars_ue->frame,
                                                                        (((last_slot>>1)==0) ? 9 : ((last_slot>>1)-1)));
@@ -3097,6 +3101,7 @@ int lte_ue_pdcch_procedures(uint8_t eNB_id,uint8_t last_slot, PHY_VARS_UE *phy_v
 								   phy_vars_ue->dlsch_ue_MCH[0]->harq_processes[0]->nb_rb,
 								     phy_vars_ue->dlsch_ue_MCH[0]->harq_processes[0]->rb_alloc,
 								     get_Qm(phy_vars_ue->dlsch_ue_MCH[0]->harq_processes[0]->mcs),
+								     1,
 								     2,
 								     ((last_slot>>1)==9?-1:0)+phy_vars_ue->frame,last_slot>>1);
 	  
diff --git a/openair1/SIMULATION/LTE_PHY/dlsim.c b/openair1/SIMULATION/LTE_PHY/dlsim.c
index 1cb4fc8492..6f29c6eac1 100644
--- a/openair1/SIMULATION/LTE_PHY/dlsim.c
+++ b/openair1/SIMULATION/LTE_PHY/dlsim.c
@@ -1264,6 +1264,7 @@ int main(int argc, char **argv) {
 				  PHY_vars_eNB->dlsch_eNB[0][0]->nb_rb,
 				  PHY_vars_eNB->dlsch_eNB[0][0]->rb_alloc,
 				  get_Qm(PHY_vars_eNB->dlsch_eNB[0][0]->harq_processes[0]->mcs),
+				  PHY_vars_eNB->dlsch_eNB[0][0]->harq_processes[0]->Nl,
 				  num_pdcch_symbols,
 				  0,subframe);
   
@@ -1528,6 +1529,7 @@ int main(int argc, char **argv) {
 					      PHY_vars_eNB->dlsch_eNB[k][0]->nb_rb,
 					      PHY_vars_eNB->dlsch_eNB[k][0]->rb_alloc,
 					      get_Qm(PHY_vars_eNB->dlsch_eNB[k][0]->harq_processes[0]->mcs),
+					      PHY_vars_eNB->dlsch_eNB[k][0]->harq_processes[0]->Nl,
 					      num_pdcch_symbols,
 					      0,subframe);
 
@@ -1995,7 +1997,8 @@ int main(int argc, char **argv) {
 						      PHY_vars_UE->dlsch_ue[0][0]->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->nb_rb,
 						      PHY_vars_UE->dlsch_ue[0][0]->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->rb_alloc,
 						      get_Qm(PHY_vars_UE->dlsch_ue[0][0]->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->mcs),
-						      PHY_vars_UE->lte_ue_pdcch_vars[0]->num_pdcch_symbols,
+						      PHY_vars_UE->dlsch_ue[0][0]->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->Nl,
+PHY_vars_UE->lte_ue_pdcch_vars[0]->num_pdcch_symbols,
 						      0,subframe);
 		      /*
 			rate = (double)dlsch_tbs25[get_I_TBS(PHY_vars_UE->dlsch_ue[0][0]->harq_processes[PHY_vars_UE->dlsch_ue[0][0]->current_harq_pid]->mcs)][PHY_vars_UE->dlsch_ue[0][0]->nb_rb-1]/(coded_bits_per_codeword);
@@ -2210,6 +2213,7 @@ int main(int argc, char **argv) {
 					  PHY_vars_eNB->dlsch_eNB[0][0]->nb_rb,
 					  PHY_vars_eNB->dlsch_eNB[0][0]->rb_alloc,
 					  get_Qm(PHY_vars_eNB->dlsch_eNB[0][0]->harq_processes[0]->mcs),
+					  PHY_vars_eNB->dlsch_eNB[0][0]->harq_processes[0]->Nl,
 					  num_pdcch_symbols,
 					  0,subframe);
 
diff --git a/openair1/SIMULATION/LTE_PHY/mbmssim.c b/openair1/SIMULATION/LTE_PHY/mbmssim.c
index 6befae9516..e175d9f74e 100644
--- a/openair1/SIMULATION/LTE_PHY/mbmssim.c
+++ b/openair1/SIMULATION/LTE_PHY/mbmssim.c
@@ -498,7 +498,7 @@ int main(int argc, char **argv) {
 								 PHY_vars_UE->dlsch_ue_MCH[0]->harq_processes[0]->nb_rb,
 								 PHY_vars_UE->dlsch_ue_MCH[0]->harq_processes[0]->rb_alloc,
 								 get_Qm(PHY_vars_UE->dlsch_ue_MCH[0]->harq_processes[0]->mcs),
-								 2,
+								 1,2,
 								 PHY_vars_UE->frame,subframe);
       dlsch_unscrambling(&PHY_vars_UE->lte_frame_parms,1,PHY_vars_UE->dlsch_ue_MCH[0],
 			 PHY_vars_UE->dlsch_ue_MCH[0]->harq_processes[0]->G,