diff --git a/ci-scripts/cls_containerize.py b/ci-scripts/cls_containerize.py
index ee9f726c11..8092febe39 100644
--- a/ci-scripts/cls_containerize.py
+++ b/ci-scripts/cls_containerize.py
@@ -757,9 +757,13 @@ class Containerize():
 		cmd = 'cd ' + self.yamlPath[0] + ' && docker-compose -f docker-compose-ci.yml config | grep com.docker.network.bridge.name | sed -e "s@^.*name: @@"'
 		networkNames = subprocess.check_output(cmd, shell=True, stderr=subprocess.STDOUT, universal_newlines=True, timeout=10)
 		if re.search('4g.*rfsimulator', self.yamlPath[0]) is not None:
+			# Excluding any traffic from LTE-UE container (192.168.61.30)
+			# From the trf-gen, keeping only PING traffic
 			cmd = 'sudo nohup tshark -f "(host 192.168.61.11 and icmp) or (not host 192.168.61.11 and not host 192.168.61.30 and not arp and not port 53 and not port 2152)"'
 		elif re.search('5g.*rfsimulator', self.yamlPath[0]) is not None:
-			cmd = 'sudo nohup tshark -f "(host 192.168.72.135 and icmp) or (not host 192.168.72.135 and not host 192.168.71.150 and not arp and not port 53 and not port 2152 and not port 2153)"'
+			# Excluding any traffic from NR-UE containers (192.168.71.150 and 192.168.71.151)
+			# From the ext-dn, keeping only PING traffic
+			cmd = 'sudo nohup tshark -f "(host 192.168.72.135 and icmp) or (not host 192.168.72.135 and not host 192.168.71.150 and not host 192.168.71.151 and not arp and not port 53 and not port 2152 and not port 2153)"'
 		elif re.search('5g_l2sim', self.yamlPath[0]) is not None:
 			cmd = 'sudo nohup tshark -f "(host 192.168.72.135 and icmp) or (not host 192.168.72.135 and not arp and not port 53 and not port 2152 and not port 2153)"'
 		else:
@@ -817,46 +821,48 @@ class Containerize():
 			# Analyzing log file(s)!
 			listOfPossibleRanContainers = ['enb', 'gnb', 'cu', 'du']
 			for container in listOfPossibleRanContainers:
-				filename = self.yamlPath[0] + '/*-oai-' + container + '.log'
-				cmd = 'ls ' + filename
+				filenames = self.yamlPath[0] + '/*-oai-' + container + '.log'
+				cmd = 'ls ' + filenames
 				containerStatus = True
 				try:
 					lsStatus = subprocess.check_output(cmd, shell=True, stderr=subprocess.STDOUT, universal_newlines=True, timeout=10)
-					filename = str(lsStatus).strip()
+					filenames = str(lsStatus).strip()
 				except:
 					containerStatus = False
 				if not containerStatus:
 					continue
 
-				logging.debug('\u001B[1m Analyzing xNB logfile ' + filename + ' \u001B[0m')
-				logStatus = RAN.AnalyzeLogFile_eNB(filename, HTML, self.ran_checkers)
-				if (logStatus < 0):
-					fullStatus = False
-					self.exitStatus = 1
-					HTML.CreateHtmlTestRow(RAN.runtime_stats, 'KO', logStatus)
-				else:
-					HTML.CreateHtmlTestRow(RAN.runtime_stats, 'OK', CONST.ALL_PROCESSES_OK)
+				for filename in filenames.split('\n'):
+					logging.debug('\u001B[1m Analyzing xNB logfile ' + filename + ' \u001B[0m')
+					logStatus = RAN.AnalyzeLogFile_eNB(filename, HTML, self.ran_checkers)
+					if (logStatus < 0):
+						fullStatus = False
+						self.exitStatus = 1
+						HTML.CreateHtmlTestRow(RAN.runtime_stats, 'KO', logStatus)
+					else:
+						HTML.CreateHtmlTestRow(RAN.runtime_stats, 'OK', CONST.ALL_PROCESSES_OK)
 
 			listOfPossibleUeContainers = ['lte-ue*', 'nr-ue*']
 			for container in listOfPossibleUeContainers:
-				filename = self.yamlPath[0] + '/*-oai-' + container + '.log'
-				cmd = 'ls ' + filename
+				filenames = self.yamlPath[0] + '/*-oai-' + container + '.log'
+				cmd = 'ls ' + filenames
 				containerStatus = True
 				try:
 					lsStatus = subprocess.check_output(cmd, shell=True, stderr=subprocess.STDOUT, universal_newlines=True, timeout=10)
-					filename = str(lsStatus).strip()
+					filenames = str(lsStatus).strip()
 				except:
 					containerStatus = False
 				if not containerStatus:
 					continue
 
-				logging.debug('\u001B[1m Analyzing UE logfile ' + filename + ' \u001B[0m')
-				logStatus = UE.AnalyzeLogFile_UE(filename, HTML, RAN)
-				if (logStatus < 0):
-					fullStatus = False
-					HTML.CreateHtmlTestRow('UE log Analysis', 'KO', logStatus)
-				else:
-					HTML.CreateHtmlTestRow('UE log Analysis', 'OK', CONST.ALL_PROCESSES_OK)
+				for filename in filenames.split('\n'):
+					logging.debug('\u001B[1m Analyzing UE logfile ' + filename + ' \u001B[0m')
+					logStatus = UE.AnalyzeLogFile_UE(filename, HTML, RAN)
+					if (logStatus < 0):
+						fullStatus = False
+						HTML.CreateHtmlTestRow('UE log Analysis', 'KO', logStatus)
+					else:
+						HTML.CreateHtmlTestRow('UE log Analysis', 'OK', CONST.ALL_PROCESSES_OK)
 
 			cmd = 'rm ' + self.yamlPath[0] + '/*.log'
 			logging.debug(cmd)
diff --git a/ci-scripts/conf_files/gNB_SA_CU.conf b/ci-scripts/conf_files/gNB_SA_CU.conf
index 39c9a43ad4..9d64ae1976 100644
--- a/ci-scripts/conf_files/gNB_SA_CU.conf
+++ b/ci-scripts/conf_files/gNB_SA_CU.conf
@@ -45,10 +45,7 @@ gNBs =
     remote_s_portc  = 500;
     remote_s_portd  = 2152;
     ssb_SubcarrierOffset                                      = 0;
-    pdsch_AntennaPorts                                        = 1;
-    pusch_AntennaPorts                                        = 1;
     min_rxtxtime                                              = 6;
-    sib1_tda                                                  = 0;
 
      pdcch_ConfigSIB1 = (
       {
diff --git a/ci-scripts/conf_files/gNB_SA_DU.conf b/ci-scripts/conf_files/gNB_SA_DU.conf
index 256d724f3f..f95bdf91ae 100644
--- a/ci-scripts/conf_files/gNB_SA_DU.conf
+++ b/ci-scripts/conf_files/gNB_SA_DU.conf
@@ -38,7 +38,6 @@ gNBs =
 
     ssb_SubcarrierOffset                                      = 0;
     min_rxtxtime                                              = 6;
-    sib1_tda                                                  = 0;
 
     pdcch_ConfigSIB1 = (
       {
diff --git a/ci-scripts/conf_files/gnb.band78.sa.fr1.106PRB.2x2.usrpn310.asue.conf b/ci-scripts/conf_files/gnb.band78.sa.fr1.106PRB.2x2.usrpn310.asue.conf
index 0feb0bb811..60fb16318f 100644
--- a/ci-scripts/conf_files/gnb.band78.sa.fr1.106PRB.2x2.usrpn310.asue.conf
+++ b/ci-scripts/conf_files/gnb.band78.sa.fr1.106PRB.2x2.usrpn310.asue.conf
@@ -42,7 +42,6 @@ gNBs =
     ////////// Physical parameters:
 
     ssb_SubcarrierOffset                                      = 0;
-    pdsch_AntennaPorts                                        = 1;
     pusch_AntennaPorts                                        = 2;
     ul_prbblacklist                                           = "51,52,53,54"
     do_SRS                                                    = 1;
diff --git a/ci-scripts/conf_files/gnb.sa.band66.fr1.106PRB.usrpn300.conf b/ci-scripts/conf_files/gnb.sa.band66.fr1.106PRB.usrpn300.conf
index 7f5a9d4343..e26fc3030d 100644
--- a/ci-scripts/conf_files/gnb.sa.band66.fr1.106PRB.usrpn300.conf
+++ b/ci-scripts/conf_files/gnb.sa.band66.fr1.106PRB.usrpn300.conf
@@ -33,10 +33,7 @@ gNBs =
     ////////// Physical parameters:
 
     ssb_SubcarrierOffset                                      = 0;
-    pdsch_AntennaPorts                                        = 1;
-    pusch_AntennaPorts                                        = 1;
     min_rxtxtime                                              = 6;
-    sib1_tda                                                  = 0;
 
      pdcch_ConfigSIB1 = (
       {
diff --git a/ci-scripts/ran.py b/ci-scripts/ran.py
index 1551c1ea9a..287a0a3c09 100644
--- a/ci-scripts/ran.py
+++ b/ci-scripts/ran.py
@@ -997,8 +997,18 @@ class RANManagement():
 			for k in keys:
 				result = re.search(k, line)
 				if result is not None:
+					ue_prefix = 'ue0'
+					ue_res = re.search('UE ID 1|UE 1:', line)
+					if ue_res is not None:
+						ue_prefix = 'ue1'
+					ue_res = re.search('UE ID 2|UE 2:', line)
+					if ue_res is not None:
+						ue_prefix = 'ue2'
+					ue_res = re.search('UE ID 3|UE 3:', line)
+					if ue_res is not None:
+						ue_prefix = 'ue3'
 					#remove 1- all useless char before relevant info (ulsch or dlsch) 2- trailing char
-					dlsch_ulsch_stats[k]=re.sub(r'^.*\]\s+', r'' , line.rstrip())
+					dlsch_ulsch_stats[ue_prefix+k]=re.sub(r'^.*\]\s+', r'' , line.rstrip())
 
 			result = re.search('Received NR_RRCReconfigurationComplete from UE', str(line))
 			if result is not None:
diff --git a/ci-scripts/xml_files/container_5g_rfsim.xml b/ci-scripts/xml_files/container_5g_rfsim.xml
index bab7f9637a..a1f50948a5 100644
--- a/ci-scripts/xml_files/container_5g_rfsim.xml
+++ b/ci-scripts/xml_files/container_5g_rfsim.xml
@@ -31,8 +31,11 @@
  000001
  000002
  000003
+ 000004
  020001
  020002
+ 020003
+ 020004
  030001
  030002
  100001
@@ -71,6 +74,14 @@
                 <nb_healthy>8</nb_healthy>
         </testCase>
 
+        <testCase id="000004">
+                <class>DeployGenObject</class>
+                <desc>Deploy Second OAI 5G NR-UE RF sim SA</desc>
+                <yaml_path>yaml_files/5g_rfsimulator</yaml_path>
+                <services>oai-nr-ue2</services>
+                <nb_healthy>9</nb_healthy>
+        </testCase>
+
         <testCase id="020001">
                 <class>PingFromContainer</class>
                 <desc>Ping ext-dn from NR-UE</desc>
@@ -87,6 +98,22 @@
                 <loss_threshold>5</loss_threshold>
         </testCase>
 
+        <testCase id="020003">
+                <class>PingFromContainer</class>
+                <desc>Ping ext-dn from Second NR-UE</desc>
+                <container_name>rfsim5g-oai-nr-ue2</container_name>
+                <options>-I oaitun_ue1 -c 20 192.168.72.135</options>
+                <loss_threshold>5</loss_threshold>
+        </testCase>
+
+        <testCase id="020004">
+                <class>PingFromContainer</class>
+                <desc>Ping Second NR-UE from ext-dn</desc>
+		<container_name>rfsim5g-oai-ext-dn</container_name>
+                <options>-c 20 12.1.1.3</options>
+                <loss_threshold>5</loss_threshold>
+        </testCase>
+
         <testCase id="030001">
                 <class>IperfFromContainer</class>
                 <desc>Iperf UDP Downlink</desc>
diff --git a/doc/TESTING_5GSA_setup.md b/doc/TESTING_5GSA_setup.md
index f8c85bf2ca..9b27549df4 100644
--- a/doc/TESTING_5GSA_setup.md
+++ b/doc/TESTING_5GSA_setup.md
@@ -36,6 +36,8 @@ At the moment of writing this document interoperability with the following COTS
  - [Quectel RM500Q-GL](https://www.quectel.com/product/5g-rm500q-gl/)
  - [Simcom SIMCOM8200EA](https://www.simcom.com/product/SIM8200EA_M2.html)
  - Huawei Mate 30 Pro
+ - Oneplus 8
+ - Google Pixel 5
 
  End-to-end control plane signaling to achieve a 5G SA connection, UE registration and PDU session establishment with the CN, as well as some basic user-plane traffic tests have been validated so far using SIMCOM/Quectel modules and Huawei Mate 30 pro. In terms of interoperability with different 5G Core Networks, so far this setup has been tested with:
  
diff --git a/nfapi/open-nFAPI/nfapi/public_inc/fapi_nr_ue_interface.h b/nfapi/open-nFAPI/nfapi/public_inc/fapi_nr_ue_interface.h
index a943a2462b..725c389070 100644
--- a/nfapi/open-nFAPI/nfapi/public_inc/fapi_nr_ue_interface.h
+++ b/nfapi/open-nFAPI/nfapi/public_inc/fapi_nr_ue_interface.h
@@ -464,7 +464,8 @@ typedef struct {
   uint8_t nEpreRatioOfPDSCHToPTRS;
   /// MCS table for this DLSCH
   uint8_t mcs_table;
-
+  uint8_t nscid;
+  uint16_t dlDmrsScramblingId;
   uint16_t pduBitmap;
 } fapi_nr_dl_config_dlsch_pdu_rel15_t;
 
diff --git a/openair1/PHY/INIT/nr_init.c b/openair1/PHY/INIT/nr_init.c
index 26f25cec6c..ab38278207 100644
--- a/openair1/PHY/INIT/nr_init.c
+++ b/openair1/PHY/INIT/nr_init.c
@@ -536,6 +536,7 @@ int phy_init_nr_gNB(PHY_VARS_gNB *gNB,
   }
 
   nr_generate_modulation_table();
+  gNB->pdcch_gold_init = cfg->cell_config.phy_cell_id.value;
   nr_init_pdcch_dmrs(gNB, cfg->cell_config.phy_cell_id.value);
   nr_init_pbch_interleaver(gNB->nr_pbch_interleaver);
 
@@ -549,28 +550,30 @@ int phy_init_nr_gNB(PHY_VARS_gNB *gNB,
     pdsch_dmrs[slot] = (uint32_t ***)malloc16(fp->symbols_per_slot*sizeof(uint32_t **));
     AssertFatal(pdsch_dmrs[slot]!=NULL, "NR init: pdsch_dmrs for slot %d - malloc failed\n", slot);
 
-    int nb_codewords = NR_MAX_NB_LAYERS > 4 ? 2 : 1;
     for (int symb=0; symb<fp->symbols_per_slot; symb++) {
-      pdsch_dmrs[slot][symb] = (uint32_t **)malloc16(nb_codewords*sizeof(uint32_t *));
+      pdsch_dmrs[slot][symb] = (uint32_t **)malloc16(NR_NB_NSCID*sizeof(uint32_t *));
       AssertFatal(pdsch_dmrs[slot][symb]!=NULL, "NR init: pdsch_dmrs for slot %d symbol %d - malloc failed\n", slot, symb);
 
-      for (int q=0; q<nb_codewords; q++) {
+      for (int q=0; q<NR_NB_NSCID; q++) {
         pdsch_dmrs[slot][symb][q] = (uint32_t *)malloc16(pdsch_dmrs_init_length*sizeof(uint32_t));
-        AssertFatal(pdsch_dmrs[slot][symb][q]!=NULL, "NR init: pdsch_dmrs for slot %d symbol %d codeword %d - malloc failed\n", slot, symb, q);
+        AssertFatal(pdsch_dmrs[slot][symb][q]!=NULL, "NR init: pdsch_dmrs for slot %d symbol %d nscid %d - malloc failed\n", slot, symb, q);
       }
     }
   }
 
-  nr_init_pdsch_dmrs(gNB, cfg->cell_config.phy_cell_id.value);
+
+  for (int nscid = 0; nscid < NR_NB_NSCID; nscid++) {
+    gNB->pdsch_gold_init[nscid] = cfg->cell_config.phy_cell_id.value;
+    nr_init_pdsch_dmrs(gNB, nscid, cfg->cell_config.phy_cell_id.value);
+  }
 
   //PUSCH DMRS init
-  gNB->nr_gold_pusch_dmrs = (uint32_t ****)malloc16(2*sizeof(uint32_t ***));
+  gNB->nr_gold_pusch_dmrs = (uint32_t ****)malloc16(NR_NB_NSCID*sizeof(uint32_t ***));
 
   uint32_t ****pusch_dmrs = gNB->nr_gold_pusch_dmrs;
 
-  // ceil(((NB_RB*6(k)*2(QPSK)/32) // 3 RE *2(QPSK)
   int pusch_dmrs_init_length =  ((fp->N_RB_UL*12)>>5)+1;
-  for(int nscid=0; nscid<2; nscid++) {
+  for(int nscid=0; nscid<NR_NB_NSCID; nscid++) {
     pusch_dmrs[nscid] = (uint32_t ***)malloc16(fp->slots_per_frame*sizeof(uint32_t **));
     AssertFatal(pusch_dmrs[nscid]!=NULL, "NR init: pusch_dmrs for nscid %d - malloc failed\n", nscid);
 
@@ -585,9 +588,10 @@ int phy_init_nr_gNB(PHY_VARS_gNB *gNB,
     }
   }
 
-  uint32_t Nid_pusch[2] = {cfg->cell_config.phy_cell_id.value,cfg->cell_config.phy_cell_id.value};
-  LOG_D(PHY,"Initializing PUSCH DMRS Gold sequence with (%x,%x)\n",Nid_pusch[0],Nid_pusch[1]);
-  nr_gold_pusch(gNB, &Nid_pusch[0]);
+  for (int nscid=0; nscid<NR_NB_NSCID; nscid++) {
+    gNB->pusch_gold_init[nscid] = cfg->cell_config.phy_cell_id.value;
+    nr_gold_pusch(gNB, nscid, gNB->pusch_gold_init[nscid]);
+  }
 
   // CSI RS init
   // ceil((NB_RB*8(max allocation per RB)*2(QPSK))/32)
@@ -603,6 +607,8 @@ int phy_init_nr_gNB(PHY_VARS_gNB *gNB,
       AssertFatal(gNB->nr_csi_rs_info->nr_gold_csi_rs[slot][symb]!=NULL, "NR init: csi reference signal for slot %d symbol %d - malloc failed\n", slot, symb);
     }
   }
+
+  gNB->csi_gold_init = cfg->cell_config.phy_cell_id.value;
   nr_init_csi_rs(&gNB->frame_parms, gNB->nr_csi_rs_info->nr_gold_csi_rs, cfg->cell_config.phy_cell_id.value);
 
   for (int id=0; id<NUMBER_OF_NR_SRS_MAX; id++) {
@@ -640,12 +646,12 @@ int phy_init_nr_gNB(PHY_VARS_gNB *gNB,
   common_vars->beam_id = (uint8_t **)malloc16(Ptx*sizeof(uint8_t*));
 
   for (i=0;i<Ptx;i++){
-      common_vars->txdataF[i] = (int32_t*)malloc16_clear(fp->samples_per_frame_wCP*sizeof(int32_t)); // [hna] samples_per_frame without CP
-      LOG_D(PHY,"[INIT] common_vars->txdataF[%d] = %p (%lu bytes)\n",
-	    i,common_vars->txdataF[i],
-	    fp->samples_per_frame_wCP*sizeof(int32_t));
-      common_vars->beam_id[i] = (uint8_t*)malloc16_clear(fp->symbols_per_slot*fp->slots_per_frame*sizeof(uint8_t));
-      memset(common_vars->beam_id[i],255,fp->symbols_per_slot*fp->slots_per_frame);
+    common_vars->txdataF[i] = (int32_t*)malloc16_clear(fp->samples_per_frame_wCP*sizeof(int32_t)); // [hna] samples_per_frame without CP
+    LOG_D(PHY,"[INIT] common_vars->txdataF[%d] = %p (%lu bytes)\n",
+          i,common_vars->txdataF[i],
+          fp->samples_per_frame_wCP*sizeof(int32_t));
+    common_vars->beam_id[i] = (uint8_t*)malloc16_clear(fp->symbols_per_slot*fp->slots_per_frame*sizeof(uint8_t));
+    memset(common_vars->beam_id[i],255,fp->symbols_per_slot*fp->slots_per_frame);
   }
   for (i=0;i<Prx;i++){
     common_vars->rxdataF[i] = (int32_t*)malloc16_clear(fp->samples_per_frame_wCP*sizeof(int32_t));
@@ -736,10 +742,9 @@ void phy_free_nr_gNB(PHY_VARS_gNB *gNB)
   free_and_zero(pdcch_dmrs);
 
   uint32_t ****pdsch_dmrs = gNB->nr_gold_pdsch_dmrs;
-  int nb_codewords = NR_MAX_NB_LAYERS > 4 ? 2 : 1;
   for (int slot = 0; slot < fp->slots_per_frame; slot++) {
     for (int symb = 0; symb < fp->symbols_per_slot; symb++) {
-      for (int q = 0; q < nb_codewords; q++)
+      for (int q = 0; q < NR_NB_NSCID; q++)
         free_and_zero(pdsch_dmrs[slot][symb][q]);
       free_and_zero(pdsch_dmrs[slot][symb]);
     }
@@ -768,7 +773,7 @@ void phy_free_nr_gNB(PHY_VARS_gNB *gNB)
   free_and_zero(gNB->nr_csi_rs_info);
 
   for (int id = 0; id < NUMBER_OF_NR_SRS_MAX; id++) {
-    for (int i = 0; i < Prx; i++){
+    for (int i = 0; i < Prx; i++) {
       free_and_zero(gNB->nr_srs_info[id]->srs_received_signal[i]);
       free_and_zero(gNB->nr_srs_info[id]->srs_ls_estimated_channel[i]);
       free_and_zero(gNB->nr_srs_info[id]->srs_estimated_channel_freq[i]);
diff --git a/openair1/PHY/INIT/nr_init_ue.c b/openair1/PHY/INIT/nr_init_ue.c
index 8892cd8b23..c5407af37f 100644
--- a/openair1/PHY/INIT/nr_init_ue.c
+++ b/openair1/PHY/INIT/nr_init_ue.c
@@ -212,18 +212,21 @@ int init_nr_ue_signal(PHY_VARS_NR_UE *ue, int nb_connected_gNB)
 
   // ceil(((NB_RB*6(k)*2(QPSK)/32) // 3 RE *2(QPSK)
   int pusch_dmrs_init_length =  ((fp->N_RB_UL*12)>>5)+1;
-
-  ue->nr_gold_pusch_dmrs = (uint32_t ***)malloc16(fp->slots_per_frame*sizeof(uint32_t **));
-  uint32_t ***pusch_dmrs = ue->nr_gold_pusch_dmrs;
+  ue->nr_gold_pusch_dmrs = (uint32_t ****)malloc16(fp->slots_per_frame*sizeof(uint32_t ***));
+  uint32_t ****pusch_dmrs = ue->nr_gold_pusch_dmrs;
 
   for (slot=0; slot<fp->slots_per_frame; slot++) {
-    pusch_dmrs[slot] = (uint32_t **)malloc16(fp->symbols_per_slot*sizeof(uint32_t *));
+    pusch_dmrs[slot] = (uint32_t ***)malloc16(fp->symbols_per_slot*sizeof(uint32_t **));
     AssertFatal(pusch_dmrs[slot]!=NULL, "init_nr_ue_signal: pusch_dmrs for slot %d - malloc failed\n", slot);
 
     for (symb=0; symb<fp->symbols_per_slot; symb++) {
-      pusch_dmrs[slot][symb] = (uint32_t *)malloc16(pusch_dmrs_init_length*sizeof(uint32_t));
+      pusch_dmrs[slot][symb] = (uint32_t **)malloc16(NR_NB_NSCID*sizeof(uint32_t *));
       AssertFatal(pusch_dmrs[slot][symb]!=NULL, "init_nr_ue_signal: pusch_dmrs for slot %d symbol %d - malloc failed\n", slot, symb);
 
+      for (int q=0; q<NR_NB_NSCID; q++) {
+        pusch_dmrs[slot][symb][q] = (uint32_t *)malloc16(pusch_dmrs_init_length*sizeof(uint32_t));
+        AssertFatal(pusch_dmrs[slot][symb][q]!=NULL, "init_nr_ue_signal: pusch_dmrs for slot %d symbol %d nscid %d - malloc failed\n", slot, symb, q);
+      }
     }
   }
 
@@ -300,24 +303,38 @@ int init_nr_ue_signal(PHY_VARS_NR_UE *ue, int nb_connected_gNB)
     pdsch_dmrs[slot] = (uint32_t ***)malloc16(fp->symbols_per_slot*sizeof(uint32_t **));
     AssertFatal(pdsch_dmrs[slot]!=NULL, "NR init: pdsch_dmrs for slot %d - malloc failed\n", slot);
 
-    int nb_codewords = NR_MAX_NB_LAYERS > 4 ? 2 : 1;
     for (int symb=0; symb<fp->symbols_per_slot; symb++) {
-      pdsch_dmrs[slot][symb] = (uint32_t **)malloc16(nb_codewords*sizeof(uint32_t *));
+      pdsch_dmrs[slot][symb] = (uint32_t **)malloc16(NR_NB_NSCID*sizeof(uint32_t *));
       AssertFatal(pdsch_dmrs[slot][symb]!=NULL, "NR init: pdsch_dmrs for slot %d symbol %d - malloc failed\n", slot, symb);
 
-      for (int q=0; q<nb_codewords; q++) {
+      for (int q=0; q<NR_NB_NSCID; q++) {
         pdsch_dmrs[slot][symb][q] = (uint32_t *)malloc16(pdsch_dmrs_init_length*sizeof(uint32_t));
-        AssertFatal(pdsch_dmrs[slot][symb][q]!=NULL, "NR init: pdsch_dmrs for slot %d symbol %d codeword %d - malloc failed\n", slot, symb, q);
+        AssertFatal(pdsch_dmrs[slot][symb][q]!=NULL, "NR init: pdsch_dmrs for slot %d symbol %d nscid %d - malloc failed\n", slot, symb, q);
       }
     }
   }
 
   // DLSCH
-  for (gNB_id = 0; gNB_id < ue->n_connected_gNB; gNB_id++) {
+  for (gNB_id = 0; gNB_id < ue->n_connected_gNB+1; gNB_id++) {
     for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
       ue->pdsch_vars[th_id][gNB_id] = (NR_UE_PDSCH *)malloc16_clear(sizeof(NR_UE_PDSCH));
     }
+    for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
+      phy_init_nr_ue_PDSCH( ue->pdsch_vars[th_id][gNB_id], fp );
+    }
 
+    int nb_codewords = NR_MAX_NB_LAYERS > 4 ? 2 : 1;
+    for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
+      for (i=0; i<nb_codewords; i++) {
+        ue->pdsch_vars[th_id][gNB_id]->llr[i] = (int16_t *)malloc16_clear( (8*(3*8*8448))*sizeof(int16_t) );//Q_m = 8 bits/Sym, Code_Rate=3, Number of Segments =8, Circular Buffer K_cb = 8448
+      }
+      for (i=0; i<NR_MAX_NB_LAYERS; i++) {
+        ue->pdsch_vars[th_id][gNB_id]->layer_llr[i] = (int16_t *)malloc16_clear( (8*(3*8*8448))*sizeof(int16_t) );//Q_m = 8 bits/Sym, Code_Rate=3, Number of Segments =8, Circular Buffer K_cb = 8448
+      }
+    }
+  }
+
+  for (gNB_id = 0; gNB_id < ue->n_connected_gNB; gNB_id++) {
     for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
       ue->pdcch_vars[th_id][gNB_id] = (NR_UE_PDCCH *)malloc16_clear(sizeof(NR_UE_PDCCH));
     }
@@ -380,21 +397,6 @@ int init_nr_ue_signal(PHY_VARS_NR_UE *ue, int nb_connected_gNB)
       ue->nr_srs_info->srs_estimated_channel_time_shifted[i] = (int32_t *) malloc16_clear(fp->ofdm_symbol_size*MAX_NUM_NR_SRS_SYMBOLS*sizeof(int32_t));
     }
 
-
-    for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
-      phy_init_nr_ue_PDSCH( ue->pdsch_vars[th_id][gNB_id], fp );
-    }
-
-    int nb_codewords = NR_MAX_NB_LAYERS > 4 ? 2 : 1;
-    for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
-      for (i=0; i<nb_codewords; i++) {
-        ue->pdsch_vars[th_id][gNB_id]->llr[i] = (int16_t *)malloc16_clear( (8*(3*8*8448))*sizeof(int16_t) );//Q_m = 8 bits/Sym, Code_Rate=3, Number of Segments =8, Circular Buffer K_cb = 8448
-      }
-      for (i=0; i<NR_MAX_NB_LAYERS; i++) {
-        ue->pdsch_vars[th_id][gNB_id]->layer_llr[i] = (int16_t *)malloc16_clear( (8*(3*8*8448))*sizeof(int16_t) );//Q_m = 8 bits/Sym, Code_Rate=3, Number of Segments =8, Circular Buffer K_cb = 8448
-      }
-    }
-
     // 100 PRBs * 12 REs/PRB * 4 PDCCH SYMBOLS * 2 LLRs/RE
     for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
       ue->pdcch_vars[th_id][gNB_id]->llr                 = (int16_t *)malloc16_clear( 2*4*100*12*sizeof(uint16_t) );
@@ -462,6 +464,8 @@ void term_nr_ue_signal(PHY_VARS_NR_UE *ue, int nb_connected_gNB)
 
   for (int slot = 0; slot < fp->slots_per_frame; slot++) {
     for (int symb = 0; symb < fp->symbols_per_slot; symb++) {
+      for (int q=0; q<NR_NB_NSCID; q++)
+        free_and_zero(ue->nr_gold_pusch_dmrs[slot][symb][q]);
       free_and_zero(ue->nr_gold_pusch_dmrs[slot][symb]);
     }
     free_and_zero(ue->nr_gold_pusch_dmrs[slot]);
@@ -495,10 +499,9 @@ void term_nr_ue_signal(PHY_VARS_NR_UE *ue, int nb_connected_gNB)
   }
   free_and_zero(ue->nr_gold_pdcch[0]);
 
-  int nb_codewords = NR_MAX_NB_LAYERS > 4 ? 2 : 1;
   for (int slot=0; slot<fp->slots_per_frame; slot++) {
     for (int symb=0; symb<fp->symbols_per_slot; symb++) {
-      for (int q=0; q<nb_codewords; q++) 
+      for (int q=0; q<NR_NB_NSCID; q++)
         free_and_zero(ue->nr_gold_pdsch[0][slot][symb][q]);
       free_and_zero(ue->nr_gold_pdsch[0][slot][symb]);
     }
@@ -506,16 +509,18 @@ void term_nr_ue_signal(PHY_VARS_NR_UE *ue, int nb_connected_gNB)
   }
   free_and_zero(ue->nr_gold_pdsch[0]);
 
-  for (int gNB_id = 0; gNB_id < ue->n_connected_gNB; gNB_id++) {
+  for (int gNB_id = 0; gNB_id < ue->n_connected_gNB+1; gNB_id++) {
 
     // PDSCH
     for (int th_id = 0; th_id < RX_NB_TH_MAX; th_id++) {
-
       free_and_zero(ue->pdsch_vars[th_id][gNB_id]->llr_shifts);
       free_and_zero(ue->pdsch_vars[th_id][gNB_id]->llr128_2ndstream);
       phy_term_nr_ue__PDSCH(ue->pdsch_vars[th_id][gNB_id], fp);
       free_and_zero(ue->pdsch_vars[th_id][gNB_id]);
     }
+  }
+  
+  for (int gNB_id = 0; gNB_id < ue->n_connected_gNB; gNB_id++) {
 
     for (int th_id = 0; th_id < RX_NB_TH_MAX; th_id++) {
       for (int i = 0; i < fp->nb_antennas_rx; i++) {
diff --git a/openair1/PHY/LTE_UE_TRANSPORT/transport_ue.h b/openair1/PHY/LTE_UE_TRANSPORT/transport_ue.h
index 6b84a5f969..b4ae44f6ea 100644
--- a/openair1/PHY/LTE_UE_TRANSPORT/transport_ue.h
+++ b/openair1/PHY/LTE_UE_TRANSPORT/transport_ue.h
@@ -155,8 +155,6 @@ typedef struct {
   uint8_t h[MAX_NUM_CHANNEL_BITS];
   /// Scrambled "b"-sequences (for definition see 36-211 V8.6 2009-03, p.14)
   uint8_t b_tilde[MAX_NUM_CHANNEL_BITS];
-  /// Modulated "d"-sequences (for definition see 36-211 V8.6 2009-03, p.14)
-  int32_t d[MAX_NUM_RE];
   /// Transform-coded "z"-sequences (for definition see 36-211 V8.6 2009-03, p.14-15)
   int32_t z[MAX_NUM_RE];
   /// "q" sequences for CQI/PMI (for definition see 36-212 V8.6 2009-03, p.27)
diff --git a/openair1/PHY/LTE_UE_TRANSPORT/ulsch_modulation.c b/openair1/PHY/LTE_UE_TRANSPORT/ulsch_modulation.c
index 7ec3e7bfb6..6a663e8940 100644
--- a/openair1/PHY/LTE_UE_TRANSPORT/ulsch_modulation.c
+++ b/openair1/PHY/LTE_UE_TRANSPORT/ulsch_modulation.c
@@ -42,7 +42,7 @@
 
 //#define DEBUG_ULSCH_MODULATION
 
-void dft_lte(int32_t *z,int32_t *d, int32_t Msc_PUSCH, uint8_t Nsymb)
+void dft_lte(int32_t *z,struct complex16 *input, int32_t Msc_PUSCH, uint8_t Nsymb)
 {
 
 #if defined(__x86_64__) || defined(__i386__)
@@ -66,7 +66,7 @@ void dft_lte(int32_t *z,int32_t *d, int32_t Msc_PUSCH, uint8_t Nsymb)
 #endif
   //  printf("Doing lte_dft for Msc_PUSCH %d\n",Msc_PUSCH);
 
-  d0 = (uint32_t *)d;
+  d0 = (uint32_t *)input;
   d1 = d0+Msc_PUSCH;
   d2 = d1+Msc_PUSCH;
   d3 = d2+Msc_PUSCH;
@@ -476,7 +476,8 @@ void ulsch_modulation(int32_t **txdataF,
   // Modulation
 
   ulsch_Msymb = G/Q_m;
-
+  /// Modulated "d"-sequences (for definition see 36-211 V8.6 2009-03, p.14)
+  struct complex16 d[MAX_NUM_RE] __attribute__((aligned(32)));
   if(ulsch->cooperation_flag == 2)
     // For Distributed Alamouti Scheme in Collabrative Communication
   {
@@ -488,14 +489,14 @@ void ulsch_modulation(int32_t **txdataF,
 
 
         //UE1, -x1*
-        ((int16_t*)&ulsch->d[i])[0] = (ulsch->b_tilde[j] == 1)  ? (gain_lin_QPSK) : -gain_lin_QPSK;
-        ((int16_t*)&ulsch->d[i])[1] = (ulsch->b_tilde[j+1] == 1)? (-gain_lin_QPSK) : gain_lin_QPSK;
+        d[i].r = (ulsch->b_tilde[j] == 1)  ? (gain_lin_QPSK) : -gain_lin_QPSK;
+        d[i].i = (ulsch->b_tilde[j+1] == 1)? (-gain_lin_QPSK) : gain_lin_QPSK;
         //      if (i<Msc_PUSCH)
         //  printf("input %d (%p): %d,%d\n", i,&ulsch->d[i],((int16_t*)&ulsch->d[i])[0],((int16_t*)&ulsch->d[i])[1]);
 
         // UE1, x0*
-        ((int16_t*)&ulsch->d[i+1])[0] = (ulsch->b_tilde[j-2] == 1)  ? (-gain_lin_QPSK) : gain_lin_QPSK;
-        ((int16_t*)&ulsch->d[i+1])[1] = (ulsch->b_tilde[j-1] == 1)? (gain_lin_QPSK) : -gain_lin_QPSK;
+        d[i+1].r = (ulsch->b_tilde[j-2] == 1)  ? (-gain_lin_QPSK) : gain_lin_QPSK;
+        d[i+1].i = (ulsch->b_tilde[j-1] == 1)? (gain_lin_QPSK) : -gain_lin_QPSK;
 
         break;
 
@@ -521,8 +522,8 @@ void ulsch_modulation(int32_t **txdataF,
           qam16_table_offset_im+=1;
 
 
-        ((int16_t*)&ulsch->d[i])[0]=-(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i])[1]=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
+        d[i].r =-(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
+        d[i].i =(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
 
         //UE1,x0*
         qam16_table_offset_re = 0;
@@ -544,8 +545,8 @@ void ulsch_modulation(int32_t **txdataF,
 
         //    ((int16_t*)&ulsch->d[i+1])[0]=-(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
         //    ((int16_t*)&ulsch->d[i+1])[1]=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
-        ((int16_t*)&ulsch->d[i+1])[0]=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i+1])[1]=-(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
+        d[i+1].r=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
+        d[i+1].i=-(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
 
 
         break;
@@ -578,8 +579,8 @@ void ulsch_modulation(int32_t **txdataF,
           qam64_table_offset_im+=1;
 
 
-        ((int16_t*)&ulsch->d[i])[0]=-(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i])[1]=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
+        d[i].r=-(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
+        d[i].i=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
 
         //UE1,x0*
         qam64_table_offset_re = 0;
@@ -605,8 +606,8 @@ void ulsch_modulation(int32_t **txdataF,
           qam64_table_offset_im+=1;
 
 
-        ((int16_t*)&ulsch->d[i+1])[0]=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i+1])[1]=-(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
+        d[i+1].r=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
+        d[i+1].i=-(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
 
         break;
 
@@ -621,8 +622,8 @@ void ulsch_modulation(int32_t **txdataF,
       case 2:
         // TODO: this has to be updated!!!
 
-        ((int16_t*)&ulsch->d[i])[0] = (ulsch->b_tilde[j] == 1)  ? (-gain_lin_QPSK) : gain_lin_QPSK;
-        ((int16_t*)&ulsch->d[i])[1] = (ulsch->b_tilde[j+1] == 1)? (-gain_lin_QPSK) : gain_lin_QPSK;
+        d[i].r = (ulsch->b_tilde[j] == 1)  ? (-gain_lin_QPSK) : gain_lin_QPSK;
+        d[i].i = (ulsch->b_tilde[j+1] == 1)? (-gain_lin_QPSK) : gain_lin_QPSK;
         //        if (i<Msc_PUSCH)
         //    printf("input %d/%d Msc_PUSCH %d (%p): %d,%d\n", i,Msymb,Msc_PUSCH,&ulsch->d[i],((int16_t*)&ulsch->d[i])[0],((int16_t*)&ulsch->d[i])[1]);
 
@@ -646,8 +647,8 @@ void ulsch_modulation(int32_t **txdataF,
           qam16_table_offset_im+=1;
 
 
-        ((int16_t*)&ulsch->d[i])[0]=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i])[1]=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
+        d[i].r=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
+        d[i].i=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
         //      printf("input(16qam) %d (%p): %d,%d\n", i,&ulsch->d[i],((int16_t*)&ulsch->d[i])[0],((int16_t*)&ulsch->d[i])[1]);
         break;
 
@@ -676,8 +677,8 @@ void ulsch_modulation(int32_t **txdataF,
           qam64_table_offset_im+=1;
 
 
-        ((int16_t*)&ulsch->d[i])[0]=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i])[1]=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
+        d[i].r=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
+        d[i].i=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
 
         break;
 
@@ -688,7 +689,7 @@ void ulsch_modulation(int32_t **txdataF,
 
   // Transform Precoding
 
-  dft_lte(ulsch->z,ulsch->d,Msc_PUSCH,ulsch->Nsymb_pusch);
+  dft_lte(ulsch->z,d,Msc_PUSCH,ulsch->Nsymb_pusch);
 
   DevAssert(txdataF);
 
diff --git a/openair1/PHY/MODULATION/slot_fep.c b/openair1/PHY/MODULATION/slot_fep.c
index 8f41735f10..b2c7d934c6 100644
--- a/openair1/PHY/MODULATION/slot_fep.c
+++ b/openair1/PHY/MODULATION/slot_fep.c
@@ -91,7 +91,7 @@ int slot_fep(PHY_VARS_UE *ue,
   }
 
   //  subframe_offset_F = frame_parms->ofdm_symbol_size * frame_parms->symbols_per_tti * (Ns>>1);
-
+ 
   if (l<0 || l>=7-frame_parms->Ncp) {
     printf("slot_fep: l must be between 0 and %d\n",7-frame_parms->Ncp);
     return(-1);
diff --git a/openair1/PHY/NR_ESTIMATION/nr_ul_channel_estimation.c b/openair1/PHY/NR_ESTIMATION/nr_ul_channel_estimation.c
index 9aaba7fb66..4a6f8ef9af 100644
--- a/openair1/PHY/NR_ESTIMATION/nr_ul_channel_estimation.c
+++ b/openair1/PHY/NR_ESTIMATION/nr_ul_channel_estimation.c
@@ -165,6 +165,11 @@ int nr_pusch_channel_estimation(PHY_VARS_gNB *gNB,
 
   //------------------generate DMRS------------------//
 
+  if(pusch_pdu->ul_dmrs_scrambling_id != gNB->pusch_gold_init[pusch_pdu->scid])  {
+    gNB->pusch_gold_init[pusch_pdu->scid] = pusch_pdu->ul_dmrs_scrambling_id;
+    nr_gold_pusch(gNB, pusch_pdu->scid, pusch_pdu->ul_dmrs_scrambling_id);
+  }
+
   // transform precoding = 1 means disabled
   if (pusch_pdu->transform_precoding == 1) {
     nr_pusch_dmrs_rx(gNB, Ns, gNB->nr_gold_pusch_dmrs[pusch_pdu->scid][Ns][symbol], &pilot[0], 1000, 0, nb_rb_pusch,
@@ -1247,4 +1252,4 @@ int nr_srs_channel_estimation(PHY_VARS_gNB *gNB,
 #endif
 
   return 0;
-}
\ No newline at end of file
+}
diff --git a/openair1/PHY/NR_REFSIG/nr_gold.c b/openair1/PHY/NR_REFSIG/nr_gold.c
index 14f62b8033..9b4cc50953 100644
--- a/openair1/PHY/NR_REFSIG/nr_gold.c
+++ b/openair1/PHY/NR_REFSIG/nr_gold.c
@@ -78,57 +78,47 @@ void nr_init_pdcch_dmrs(PHY_VARS_gNB* gNB, uint32_t Nid)
 }
 
 
-void nr_init_pdsch_dmrs(PHY_VARS_gNB* gNB, uint32_t Nid)
-{
+void nr_init_pdsch_dmrs(PHY_VARS_gNB* gNB, uint8_t nscid, uint32_t Nid) {
   
   uint32_t x1, x2;
   uint8_t reset;
   NR_DL_FRAME_PARMS *fp = &gNB->frame_parms;
   uint32_t ****pdsch_dmrs = gNB->nr_gold_pdsch_dmrs;
   int pdsch_dmrs_init_length =  ((fp->N_RB_DL*12)>>5)+1;
-  int nb_codewords = NR_MAX_NB_LAYERS > 4 ? 2 : 1;
-  uint16_t N_n_scid[2]={Nid, Nid};
-  uint8_t n_scid=0; // again works only for 1_0
+
   for (uint8_t slot=0; slot<fp->slots_per_frame; slot++) {
 
     for (uint8_t symb=0; symb<fp->symbols_per_slot; symb++) {
-        reset = 1;
-        x2 = ((1<<17) * (fp->symbols_per_slot*slot+symb+1) * ((N_n_scid[n_scid]<<1)+1) +((N_n_scid[n_scid]<<1)+n_scid));
-	      LOG_D(PHY,"PDSCH DMRS slot %d, symb %d x2 %x, N_n_scid %d,n_scid %d\n",slot,symb,x2,N_n_scid[n_scid],n_scid);
-        for (uint32_t n=0; n<pdsch_dmrs_init_length; n++) {
-          pdsch_dmrs[slot][symb][0][n] = lte_gold_generic(&x1, &x2, reset);
-          reset = 0;
-        }
-
-        if(nb_codewords>1)
-          memcpy(pdsch_dmrs[slot][symb][1],pdsch_dmrs[slot][symb][0],sizeof(uint32_t)*pdsch_dmrs_init_length);
+      reset = 1;
+      x2 = ((1<<17) * (fp->symbols_per_slot*slot+symb+1) * ((Nid<<1)+1) +((Nid<<1)+nscid));
+      LOG_D(PHY,"PDSCH DMRS slot %d, symb %d x2 %x, Nid %d,nscid %d\n",slot,symb,x2,Nid,nscid);
+      for (uint32_t n=0; n<pdsch_dmrs_init_length; n++) {
+        pdsch_dmrs[slot][symb][nscid][n] = lte_gold_generic(&x1, &x2, reset);
+        reset = 0;
+      }
     }
   }
 }
 
 
-void nr_gold_pusch(PHY_VARS_gNB* gNB, uint32_t *Nid) {
+void nr_gold_pusch(PHY_VARS_gNB* gNB, int nscid, uint32_t nid) {
 
   unsigned char ns;
   unsigned int n,x1,x2;
-  int nscid, reset;
-  unsigned int nid;
+  int reset;
   NR_DL_FRAME_PARMS *fp = &gNB->frame_parms;
   unsigned short l;
   int pusch_dmrs_init_length =  ((fp->N_RB_UL*12)>>5)+1;
 
-  for (nscid=0; nscid<2; nscid++) {
-    nid = Nid[nscid];
-    for (ns=0; ns<fp->slots_per_frame; ns++) {
-      for (l=0; l<fp->symbols_per_slot; l++) {
-        reset = 1;
-        x2 = ((1<<17) * (fp->symbols_per_slot*ns+l+1) * ((nid<<1)+1) +((nid<<1)+nscid));
-        LOG_D(PHY,"DMRS slot %d, symb %d x2 %x\n",ns,l,x2);
+  for (ns=0; ns<fp->slots_per_frame; ns++) {
+    for (l=0; l<fp->symbols_per_slot; l++) {
+      reset = 1;
+      x2 = ((1<<17) * (fp->symbols_per_slot*ns+l+1) * ((nid<<1)+1) +((nid<<1)+nscid));
+      LOG_D(PHY,"DMRS slot %d, symb %d x2 %x\n",ns,l,x2);
 
-        for (n=0; n<pusch_dmrs_init_length; n++) {
-          gNB->nr_gold_pusch_dmrs[nscid][ns][l][n] = lte_gold_generic(&x1, &x2, reset);
-          reset = 0;
-        }
+      for (n=0; n<pusch_dmrs_init_length; n++) {
+        gNB->nr_gold_pusch_dmrs[nscid][ns][l][n] = lte_gold_generic(&x1, &x2, reset);
+        reset = 0;
       }
     }
   }
diff --git a/openair1/PHY/NR_REFSIG/nr_gold_ue.c b/openair1/PHY/NR_REFSIG/nr_gold_ue.c
index 932cdba538..203d6c8625 100644
--- a/openair1/PHY/NR_REFSIG/nr_gold_ue.c
+++ b/openair1/PHY/NR_REFSIG/nr_gold_ue.c
@@ -76,46 +76,38 @@ void nr_gold_pdcch(PHY_VARS_NR_UE* ue,
 }
 
 void nr_gold_pdsch(PHY_VARS_NR_UE* ue,
-                   unsigned short *n_idDMRS)
-{
+                   int nscid,
+                   uint32_t nid) {
 
   unsigned int x1,x2,x2tmp0;
-  unsigned int nid;
   uint8_t reset;
   int pdsch_dmrs_init_length =  ((ue->frame_parms.N_RB_DL*12)>>5)+1;
-  int nb_codewords = NR_MAX_NB_LAYERS > 4 ? 2 : 1;
-  /// to be updated from higher layer
-  //unsigned short lbar = 0;
 
-  for (int nscid=0; nscid<nb_codewords; nscid++) {
-    for (int ns=0; ns<ue->frame_parms.slots_per_frame; ns++) {
+  for (int ns=0; ns<ue->frame_parms.slots_per_frame; ns++) {
 
-      nid = n_idDMRS[nscid];
+    for (int l=0; l<ue->frame_parms.symbols_per_slot; l++) {
 
-      for (int l=0; l<ue->frame_parms.symbols_per_slot; l++) {
+      reset = 1;
+      x2tmp0 = ((ue->frame_parms.symbols_per_slot*ns+l+1)*((nid<<1)+1))<<17;
+      x2 = (x2tmp0+(nid<<1)+nscid)%(1U<<31);  //cinit
+      LOG_D(PHY,"UE DMRS slot %d, symb %d, x2 %x, nscid %d\n",ns,l,x2,nscid);
 
-        reset = 1;
-        x2tmp0 = ((ue->frame_parms.symbols_per_slot*ns+l+1)*((nid<<1)+1))<<17;
-        x2 = (x2tmp0+(nid<<1)+nscid)%(1U<<31);  //cinit
-        LOG_D(PHY,"UE DMRS slot %d, symb %d, x2 %x, nscid %d\n",ns,l,x2,nscid);
-
-        for (int n=0; n<pdsch_dmrs_init_length; n++) {
-          ue->nr_gold_pdsch[0][ns][l][nscid][n] = lte_gold_generic(&x1, &x2, reset);
-          reset = 0;
-        }
+      for (int n=0; n<pdsch_dmrs_init_length; n++) {
+        ue->nr_gold_pdsch[0][ns][l][nscid][n] = lte_gold_generic(&x1, &x2, reset);
+        reset = 0;
       }
     }
   }
 }
 
 void nr_init_pusch_dmrs(PHY_VARS_NR_UE* ue,
-                        uint16_t *N_n_scid,
+                        uint16_t N_n_scid,
                         uint8_t n_scid)
 {
   uint32_t x1, x2, n;
   uint8_t reset, slot, symb;
   NR_DL_FRAME_PARMS *fp = &ue->frame_parms;
-  uint32_t ***pusch_dmrs = ue->nr_gold_pusch_dmrs;
+  uint32_t ****pusch_dmrs = ue->nr_gold_pusch_dmrs;
   int pusch_dmrs_init_length =  ((fp->N_RB_UL*12)>>5)+1;
 
   for (slot=0; slot<fp->slots_per_frame; slot++) {
@@ -123,10 +115,10 @@ void nr_init_pusch_dmrs(PHY_VARS_NR_UE* ue,
     for (symb=0; symb<fp->symbols_per_slot; symb++) {
 
       reset = 1;
-      x2 = ((1<<17) * (fp->symbols_per_slot*slot+symb+1) * ((N_n_scid[n_scid]<<1)+1) +((N_n_scid[n_scid]<<1)+n_scid));
-
+      x2 = ((1<<17) * (fp->symbols_per_slot*slot+symb+1) * ((N_n_scid<<1)+1) +((N_n_scid<<1)+n_scid));
+      LOG_D(PHY,"DMRS slot %d, symb %d x2 %x\n",slot,symb,x2);
       for (n=0; n<pusch_dmrs_init_length; n++) {
-        pusch_dmrs[slot][symb][n] = lte_gold_generic(&x1, &x2, reset);
+        pusch_dmrs[slot][symb][n_scid][n] = lte_gold_generic(&x1, &x2, reset);
         reset = 0;
       }
     }
diff --git a/openair1/PHY/NR_REFSIG/nr_refsig.h b/openair1/PHY/NR_REFSIG/nr_refsig.h
index cbbb0b4899..2cac72eeb7 100644
--- a/openair1/PHY/NR_REFSIG/nr_refsig.h
+++ b/openair1/PHY/NR_REFSIG/nr_refsig.h
@@ -37,10 +37,10 @@ void nr_init_pbch_dmrs(PHY_VARS_gNB* gNB);
 @param Nid is used for the initialization of x2, Physical cell Id by default or upper layer configured pdcch_scrambling_ID
  */
 void nr_init_pdcch_dmrs(PHY_VARS_gNB* gNB, uint32_t Nid);
-void nr_init_pdsch_dmrs(PHY_VARS_gNB* gNB, uint32_t Nid);
+void nr_init_pdsch_dmrs(PHY_VARS_gNB* gNB, uint8_t nscid, uint32_t Nid);
 void nr_init_csi_rs(NR_DL_FRAME_PARMS *fp, uint32_t ***csi_rs, uint32_t Nid);
 
-void nr_gold_pusch(PHY_VARS_gNB* gNB, uint32_t *Nid);
+void nr_gold_pusch(PHY_VARS_gNB* gNB, int nscid, uint32_t nid);
 
 int nr_pusch_dmrs_rx(PHY_VARS_gNB *gNB,
                      unsigned int Ns,
diff --git a/openair1/PHY/NR_REFSIG/refsig_defs_ue.h b/openair1/PHY/NR_REFSIG/refsig_defs_ue.h
index 6b14a971b3..eb5c573368 100644
--- a/openair1/PHY/NR_REFSIG/refsig_defs_ue.h
+++ b/openair1/PHY/NR_REFSIG/refsig_defs_ue.h
@@ -61,10 +61,11 @@ void nr_gold_pdcch(PHY_VARS_NR_UE* ue,
                    unsigned short n_idDMRS);
 
 void nr_gold_pdsch(PHY_VARS_NR_UE* ue,
-                   unsigned short *n_idDMRS);
+                   int nscid,
+                   uint32_t nid);
 
 void nr_init_pusch_dmrs(PHY_VARS_NR_UE* ue,
-                        uint16_t *N_n_scid,
+                        uint16_t N_n_scid,
                         uint8_t n_scid);
 
 void nr_init_csi_rs(NR_DL_FRAME_PARMS *fp, uint32_t ***csi_rs, uint32_t Nid);
diff --git a/openair1/PHY/NR_TRANSPORT/nr_csi_rs.c b/openair1/PHY/NR_TRANSPORT/nr_csi_rs.c
index 1aa082e90e..94fe532258 100644
--- a/openair1/PHY/NR_TRANSPORT/nr_csi_rs.c
+++ b/openair1/PHY/NR_TRANSPORT/nr_csi_rs.c
@@ -48,7 +48,6 @@ void nr_generate_csi_rs(NR_DL_FRAME_PARMS frame_parms,
                         int16_t amp,
                         nr_csi_rs_info_t *nr_csi_rs_info,
                         nfapi_nr_dl_tti_csi_rs_pdu_rel15_t *csi_params,
-                        uint16_t cell_id,
                         int slot){
 
 #ifdef NR_CSIRS_DEBUG
@@ -86,20 +85,10 @@ void nr_generate_csi_rs(NR_DL_FRAME_PARMS frame_parms,
 
   AssertFatal(b!=0, "Invalid CSI frequency domain mapping: no bit selected in bitmap\n");
 
-  // pre-computed for scrambling id equel to cell id
-  // if the scrambling id is not the cell id we need to re-initialize the rs
-  if (csi_params->scramb_id != cell_id) {
-    uint8_t reset;
-    uint32_t x1, x2;
-    uint32_t Nid = csi_params->scramb_id;
-    for (uint8_t symb=0; symb<frame_parms.symbols_per_slot; symb++) {
-      reset = 1;
-      x2 = ((1<<10) * (frame_parms.symbols_per_slot*slot+symb+1) * ((Nid<<1)+1) + (Nid));
-      for (uint32_t n=0; n<(csi_rs_length>>5)+1; n++) {
-        nr_gold_csi_rs[symb][n] = lte_gold_generic(&x1, &x2, reset);
-        reset = 0;
-      }
-    }
+  // if the scrambling id is not the one previously used to initialize we need to re-initialize the rs
+  if (csi_params->scramb_id != gNB->csi_gold_init) {
+    gNB->csi_gold_init = csi_params->scramb_id;
+    nr_init_csi_rs(&frame_parms, nr_csi_rs_info->nr_gold_csi_rs, csi_params->scramb_id);
   }
 
   switch (csi_params->row) {
diff --git a/openair1/PHY/NR_TRANSPORT/nr_dci.c b/openair1/PHY/NR_TRANSPORT/nr_dci.c
index 0af56ba706..ba7e539429 100644
--- a/openair1/PHY/NR_TRANSPORT/nr_dci.c
+++ b/openair1/PHY/NR_TRANSPORT/nr_dci.c
@@ -66,11 +66,12 @@ void nr_pdcch_scrambling(uint32_t *in,
   }
 }
 
-void nr_generate_dci(nfapi_nr_dl_tti_pdcch_pdu_rel15_t *pdcch_pdu_rel15,
-                     uint32_t **gold_pdcch_dmrs,
+void nr_generate_dci(PHY_VARS_gNB *gNB,
+                     nfapi_nr_dl_tti_pdcch_pdu_rel15_t *pdcch_pdu_rel15,
                      int32_t *txdataF,
                      int16_t amp,
-                     NR_DL_FRAME_PARMS *frame_parms) {
+                     NR_DL_FRAME_PARMS *frame_parms,
+                     int slot) {
 
   int16_t mod_dmrs[NR_MAX_CSET_DURATION][NR_MAX_PDCCH_DMRS_LENGTH>>1] __attribute__((aligned(16))); // 3 for the max coreset duration
   uint16_t cset_start_sc;
@@ -81,7 +82,7 @@ void nr_generate_dci(nfapi_nr_dl_tti_pdcch_pdu_rel15_t *pdcch_pdu_rel15,
     
   int rb_offset;
   int n_rb;
-  
+
   // compute rb_offset and n_prb based on frequency allocation
   nr_cce_t cce_list[MAX_DCI_CORESET][NR_MAX_PDCCH_AGG_LEVEL];
   nr_fill_cce_list(cce_list,0,pdcch_pdu_rel15);
@@ -95,6 +96,13 @@ void nr_generate_dci(nfapi_nr_dl_tti_pdcch_pdu_rel15_t *pdcch_pdu_rel15,
      * in time: by its first slot and its first symbol*/
     const nfapi_nr_dl_dci_pdu_t *dci_pdu = &pdcch_pdu_rel15->dci_pdu[d];
 
+    if(dci_pdu->ScramblingId != gNB->pdcch_gold_init) {
+      gNB->pdcch_gold_init = dci_pdu->ScramblingId;
+      nr_init_pdcch_dmrs(gNB, dci_pdu->ScramblingId);
+    }
+
+    uint32_t **gold_pdcch_dmrs = gNB->nr_gold_pdcch_dmrs[slot];
+
     cset_start_symb = pdcch_pdu_rel15->StartSymbolIndex;
     cset_nsymb = pdcch_pdu_rel15->DurationSymbols;
     dci_idx = 0;
@@ -104,8 +112,8 @@ void nr_generate_dci(nfapi_nr_dl_tti_pdcch_pdu_rel15_t *pdcch_pdu_rel15,
     uint32_t dmrs_length = n_rb*6; //2(QPSK)*3(per RB)*6(REG per CCE)
     uint32_t encoded_length = dci_pdu->AggregationLevel*108; //2(QPSK)*9(per RB)*6(REG per CCE)
     LOG_D(PHY, "DL_DCI : rb_offset %d, nb_rb %d, DMRS length per symbol %d\t DCI encoded length %d (precoder_granularity %d,reg_mapping %d),Scrambling_Id %d,ScramblingRNTI %x,PayloadSizeBits %d\n",
-    rb_offset, n_rb,dmrs_length, encoded_length,pdcch_pdu_rel15->precoderGranularity,pdcch_pdu_rel15->CceRegMappingType,
-    dci_pdu->ScramblingId,dci_pdu->ScramblingRNTI,dci_pdu->PayloadSizeBits);
+          rb_offset, n_rb,dmrs_length, encoded_length,pdcch_pdu_rel15->precoderGranularity,pdcch_pdu_rel15->CceRegMappingType,
+          dci_pdu->ScramblingId,dci_pdu->ScramblingRNTI,dci_pdu->PayloadSizeBits);
     dmrs_length += rb_offset*6; // To accommodate more DMRS symbols in case of rb offset
       
     /// DMRS QPSK modulation
@@ -115,10 +123,10 @@ void nr_generate_dci(nfapi_nr_dl_tti_pdcch_pdu_rel15_t *pdcch_pdu_rel15,
       
 #ifdef DEBUG_PDCCH_DMRS
        if(dci_pdu->RNTI!=0xFFFF) {      
-      for (int i=0; i<dmrs_length>>1; i++)
-	printf("symb %d i %d %p gold seq 0x%08x mod_dmrs %d %d\n", symb, i,
-	       &gold_pdcch_dmrs[symb][i>>5],gold_pdcch_dmrs[symb][i>>5], mod_dmrs[symb][i<<1], mod_dmrs[symb][(i<<1)+1] );
-    }  
+         for (int i=0; i<dmrs_length>>1; i++)
+	   printf("symb %d i %d %p gold seq 0x%08x mod_dmrs %d %d\n", symb, i,
+	          &gold_pdcch_dmrs[symb][i>>5],gold_pdcch_dmrs[symb][i>>5], mod_dmrs[symb][i<<1], mod_dmrs[symb][(i<<1)+1] );
+       }
 #endif
     }
     
@@ -248,15 +256,16 @@ void nr_generate_dci(nfapi_nr_dl_tti_pdcch_pdu_rel15_t *pdcch_pdu_rel15,
 }
 
 void nr_generate_dci_top(processingData_L1tx_t *msgTx,
-                         uint32_t **gold_pdcch_dmrs,
+                         int slot,
                          int32_t *txdataF,
                          int16_t amp,
                          NR_DL_FRAME_PARMS *frame_parms) {
 
+
   for (int i=0; i<msgTx->num_ul_pdcch; i++)
-    nr_generate_dci(&msgTx->ul_pdcch_pdu[i].pdcch_pdu.pdcch_pdu_rel15,gold_pdcch_dmrs,txdataF,amp,frame_parms);
+    nr_generate_dci(msgTx->gNB,&msgTx->ul_pdcch_pdu[i].pdcch_pdu.pdcch_pdu_rel15,txdataF,amp,frame_parms,slot);
   for (int i=0; i<msgTx->num_dl_pdcch; i++)
-    nr_generate_dci(&msgTx->pdcch_pdu[i].pdcch_pdu_rel15,gold_pdcch_dmrs,txdataF,amp,frame_parms);
+    nr_generate_dci(msgTx->gNB,&msgTx->pdcch_pdu[i].pdcch_pdu_rel15,txdataF,amp,frame_parms,slot);
 
 }
 
diff --git a/openair1/PHY/NR_TRANSPORT/nr_dci.h b/openair1/PHY/NR_TRANSPORT/nr_dci.h
index bdc8509d12..490a8b5286 100644
--- a/openair1/PHY/NR_TRANSPORT/nr_dci.h
+++ b/openair1/PHY/NR_TRANSPORT/nr_dci.h
@@ -30,7 +30,7 @@ uint16_t nr_get_dci_size(nfapi_nr_dci_format_e format,
                          uint16_t N_RB);
 
 void nr_generate_dci_top(processingData_L1tx_t *msgTx,
-                         uint32_t **gold_pdcch_dmrs,
+                         int slot,
                          int32_t *txdataF,
                          int16_t amp,
                          NR_DL_FRAME_PARMS *frame_parms);
diff --git a/openair1/PHY/NR_TRANSPORT/nr_dlsch.c b/openair1/PHY/NR_TRANSPORT/nr_dlsch.c
index 07f046e940..2ac660f999 100644
--- a/openair1/PHY/NR_TRANSPORT/nr_dlsch.c
+++ b/openair1/PHY/NR_TRANSPORT/nr_dlsch.c
@@ -58,7 +58,6 @@ void nr_generate_pdsch(processingData_L1tx_t *msgTx,
 
   PHY_VARS_gNB *gNB = msgTx->gNB;
   NR_gNB_DLSCH_t *dlsch;
-  uint32_t ***pdsch_dmrs = gNB->nr_gold_pdsch_dmrs[slot];
   int32_t** txdataF = gNB->common_vars.txdataF;
   int16_t amp = AMP;
   int xOverhead = 0;
@@ -93,12 +92,17 @@ void nr_generate_pdsch(processingData_L1tx_t *msgTx,
     }
     n_dmrs = (rel15->BWPStart+rel15->rbStart+rel15->rbSize)*nb_re_dmrs;
 
+    if(rel15->dlDmrsScramblingId != gNB->pdsch_gold_init[rel15->SCID])  {
+      gNB->pdsch_gold_init[rel15->SCID] = rel15->dlDmrsScramblingId;
+      nr_init_pdsch_dmrs(gNB, rel15->SCID, rel15->dlDmrsScramblingId);
+    }
+
+    uint32_t ***pdsch_dmrs = gNB->nr_gold_pdsch_dmrs[slot];
     uint16_t dmrs_symbol_map = rel15->dlDmrsSymbPos;//single DMRS: 010000100 Double DMRS 110001100
     uint8_t dmrs_len = get_num_dmrs(rel15->dlDmrsSymbPos);
     uint16_t nb_re = ((12*rel15->NrOfSymbols)-nb_re_dmrs*dmrs_len-xOverhead)*rel15->rbSize*rel15->nrOfLayers;
     uint8_t Qm = rel15->qamModOrder[0];
     uint32_t encoded_length = nb_re*Qm;
-    uint32_t scrambled_output[rel15->NrOfCodewords][(encoded_length>>5)+1];
     int16_t mod_dmrs[n_dmrs<<1] __attribute__ ((aligned(16)));
 
     /* PTRS */
@@ -147,63 +151,61 @@ void nr_generate_pdsch(processingData_L1tx_t *msgTx,
     printf("\n");
 #endif
 
-    /// scrambling
-    start_meas(dlsch_scrambling_stats);
     for (int q=0; q<rel15->NrOfCodewords; q++) {
-      memset((void*)scrambled_output[q], 0, ((encoded_length>>5)+1)*sizeof(uint32_t));
+      /// scrambling
+      start_meas(dlsch_scrambling_stats);
+      uint32_t scrambled_output[(encoded_length>>5)+4]; // modulator acces by 4 bytes in some cases
+      memset(scrambled_output, 0, sizeof(scrambled_output));
+      if ( encoded_length > rel15->rbSize * NR_SYMBOLS_PER_SLOT * NR_NB_SC_PER_RB * Qm * rel15->nrOfLayers) abort();
       nr_pdsch_codeword_scrambling(output,
                                    encoded_length,
                                    q,
                                    rel15->dataScramblingId,
                                    rel15->rnti,
-                                   scrambled_output[q]);
-    }
-    
-    stop_meas(dlsch_scrambling_stats);
+                                   scrambled_output);
 #ifdef DEBUG_DLSCH
-    printf("PDSCH scrambling:\n");
-    for (int i=0; i<encoded_length>>8; i++) {
-      for (int j=0; j<8; j++)
-	printf("0x%08x\t", scrambled_output[0][(i<<3)+j]);
-      printf("\n");
-    }
-#endif
-    
-    /// Modulation
-    start_meas(dlsch_modulation_stats);
-    VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_gNB_PDSCH_MODULATION, 1);
-    for (int q=0; q<rel15->NrOfCodewords; q++)
-      nr_modulation(scrambled_output[q],
-		    encoded_length,
-		    Qm,
-		    mod_symbs[q]);
-    VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_gNB_PDSCH_MODULATION, 0);
-    stop_meas(dlsch_modulation_stats);
-#ifdef DEBUG_DLSCH
-    printf("PDSCH Modulation: Qm %d(%d)\n", Qm, nb_re);
-    for (int i=0; i<nb_re>>3; i++) {
-      for (int j=0; j<8; j++) {
-	printf("%d %d\t", mod_symbs[0][((i<<3)+j)<<1], mod_symbs[0][(((i<<3)+j)<<1)+1]);
+      printf("PDSCH scrambling:\n");
+      for (int i=0; i<encoded_length>>8; i++) {
+        for (int j=0; j<8; j++)
+          printf("0x%08x\t", scrambled_output[(i<<3)+j]);
+        printf("\n");
       }
-      printf("\n");
-    }
 #endif
+      stop_meas(dlsch_scrambling_stats);
+      /// Modulation
+      start_meas(dlsch_modulation_stats);
+      nr_modulation(scrambled_output,
+                    encoded_length,
+                    Qm,
+                    mod_symbs[q]);
+      VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_gNB_PDSCH_MODULATION, 0);
+      stop_meas(dlsch_modulation_stats);
+#ifdef DEBUG_DLSCH
+      printf("PDSCH Modulation: Qm %d(%d)\n", Qm, nb_re);
+      for (int i=0; i<nb_re>>3; i++) {
+        for (int j=0; j<8; j++) {
+          printf("%d %d\t", mod_symbs[0][((i<<3)+j)<<1], mod_symbs[0][(((i<<3)+j)<<1)+1]);
+        }
+        printf("\n");
+      }
+#endif
+    }
     
     start_meas(&gNB->dlsch_layer_mapping_stats); 
     /// Layer mapping
     nr_layer_mapping(mod_symbs,
-		     rel15->nrOfLayers,
-		     nb_re,
-		     tx_layers);
+                     rel15->nrOfLayers,
+                     nb_re,
+                     tx_layers);
 #ifdef DEBUG_DLSCH
     printf("Layer mapping (%d layers):\n", rel15->nrOfLayers);
     for (int l=0; l<rel15->nrOfLayers; l++)
       for (int i=0; i<(nb_re/rel15->nrOfLayers)>>3; i++) {
-	printf("layer %d, Re %d..%d : ",l,i<<3,(i<<3)+7);
-	for (int j=0; j<8; j++) {
-	  printf("l%d %d\t", tx_layers[l][((i<<3)+j)<<1], tx_layers[l][(((i<<3)+j)<<1)+1]);
-	}
-	printf("\n");
+        printf("layer %d, Re %d..%d : ",l,i<<3,(i<<3)+7);
+        for (int j=0; j<8; j++) {
+          printf("l%d %d\t", tx_layers[l][((i<<3)+j)<<1], tx_layers[l][(((i<<3)+j)<<1)+1]);
+        }
+        printf("\n");
       }
 #endif
 
@@ -223,7 +225,7 @@ void nr_generate_pdsch(processingData_L1tx_t *msgTx,
 
 #ifdef DEBUG_DLSCH_MAPPING
     printf("PDSCH resource mapping started (start SC %d\tstart symbol %d\tN_PRB %d\tnb_re %d,nb_layers %d)\n",
-	   start_sc, rel15->StartSymbolIndex, rel15->rbSize, nb_re,rel15->nrOfLayers);
+           start_sc, rel15->StartSymbolIndex, rel15->rbSize, nb_re,rel15->nrOfLayers);
 #endif
 
     start_meas(&gNB->dlsch_resource_mapping_stats);
@@ -277,7 +279,7 @@ void nr_generate_pdsch(processingData_L1tx_t *msgTx,
             l_prime = 0;
           }
           /// DMRS QPSK modulation
-          nr_modulation(pdsch_dmrs[l][0], n_dmrs*2, DMRS_MOD_ORDER, mod_dmrs); // currently only codeword 0 is modulated. Qm = 2 as DMRS is QPSK modulated
+          nr_modulation(pdsch_dmrs[l][rel15->SCID], n_dmrs*2, DMRS_MOD_ORDER, mod_dmrs); // Qm = 2 as DMRS is QPSK modulated
 
 #ifdef DEBUG_DLSCH
           printf("DMRS modulation (symbol %d, %d symbols, type %d):\n", l, n_dmrs, dmrs_Type);
@@ -298,7 +300,7 @@ void nr_generate_pdsch(processingData_L1tx_t *msgTx,
           if(ptrs_symbol) {
             /* PTRS QPSK Modulation for each OFDM symbol in a slot */
             LOG_D(PHY,"Doing ptrs modulation for symbol %d, n_ptrs %d\n",l,n_ptrs);
-            nr_modulation(pdsch_dmrs[l][0], (n_ptrs<<1), DMRS_MOD_ORDER, mod_ptrs);
+            nr_modulation(pdsch_dmrs[l][rel15->SCID], (n_ptrs<<1), DMRS_MOD_ORDER, mod_ptrs);
           }
         }
         uint16_t k = start_sc;
diff --git a/openair1/PHY/NR_UE_ESTIMATION/nr_dl_channel_estimation.c b/openair1/PHY/NR_UE_ESTIMATION/nr_dl_channel_estimation.c
index 55deb686ad..5c415e7d40 100644
--- a/openair1/PHY/NR_UE_ESTIMATION/nr_dl_channel_estimation.c
+++ b/openair1/PHY/NR_UE_ESTIMATION/nr_dl_channel_estimation.c
@@ -466,6 +466,7 @@ int nr_pdcch_channel_estimation(PHY_VARS_NR_UE *ue,
                                 uint8_t gNB_id,
                                 unsigned char Ns,
                                 unsigned char symbol,
+                                unsigned short scrambling_id,
                                 unsigned short coreset_start_subcarrier,
                                 unsigned short nb_rb_coreset)
 {
@@ -493,13 +494,11 @@ int nr_pdcch_channel_estimation(PHY_VARS_NR_UE *ue,
   fm = filt16a_m1;
   fr = filt16a_r1;
 
-
   // checking if re-initialization of scrambling IDs is needed (should be done here but scrambling ID for PDCCH is not taken from RRC)
-/*  if (( != ue->scramblingID_pdcch){
-    ue->scramblingID_pdcch=;
-    nr_gold_pdsch(ue,ue->scramblingID_pdcch);
-  }*/
-
+  if (scrambling_id != ue->scramblingID_pdcch){
+    ue->scramblingID_pdcch = scrambling_id;
+    nr_gold_pdcch(ue,ue->scramblingID_pdcch);
+  }
 
   // generate pilot
   int pilot[nb_rb_coreset * 3] __attribute__((aligned(16))); 
@@ -666,6 +665,8 @@ int nr_pdsch_channel_estimation(PHY_VARS_NR_UE *ue,
                                 unsigned char Ns,
                                 unsigned short p,
                                 unsigned char symbol,
+                                unsigned char nscid,
+                                unsigned short scrambling_id,
                                 unsigned short BWPStart,
                                 uint8_t config_type,
                                 unsigned short bwp_start_subcarrier,
@@ -703,11 +704,10 @@ int nr_pdsch_channel_estimation(PHY_VARS_NR_UE *ue,
   int8_t delta = get_delta(p, config_type);
 
   // checking if re-initialization of scrambling IDs is needed
-  /*if ((XXX.scramblingID0 != ue->scramblingID[0]) || (XXX.scramblingID1 != ue->scramblingID[1])){
-    ue->scramblingID[0] = XXX.scramblingID0;
-    ue->scramblingID[1] = XXX.scramblingID1;
-    nr_gold_pdsch(ue,ue->scramblingID);
-  }*/
+  if (scrambling_id != ue->scramblingID_dlsch[nscid]){
+    ue->scramblingID_dlsch[nscid] = scrambling_id;
+    nr_gold_pdsch(ue, nscid, scrambling_id);
+  }
 
   nr_pdsch_dmrs_rx(ue, Ns, ue->nr_gold_pdsch[gNB_id][Ns][symbol][0], &pilot[0], 1000+p, 0, nb_rb_pdsch+rb_offset, config_type);
 
diff --git a/openair1/PHY/NR_UE_ESTIMATION/nr_estimation.h b/openair1/PHY/NR_UE_ESTIMATION/nr_estimation.h
index aa4b78e145..963a845876 100644
--- a/openair1/PHY/NR_UE_ESTIMATION/nr_estimation.h
+++ b/openair1/PHY/NR_UE_ESTIMATION/nr_estimation.h
@@ -44,6 +44,7 @@ int nr_pdcch_channel_estimation(PHY_VARS_NR_UE *ue,
                                 uint8_t gNB_id,
                                 unsigned char Ns,
                                 unsigned char symbol,
+                                unsigned short scrambling_id,
                                 unsigned short coreset_start_subcarrier,
                                 unsigned short nb_rb_coreset);
 
@@ -74,6 +75,8 @@ int nr_pdsch_channel_estimation(PHY_VARS_NR_UE *ue,
                                 unsigned char Ns,
                                 unsigned short p,
                                 unsigned char symbol,
+                                unsigned char nscid,
+                                unsigned short scrambling_id,
                                 unsigned short BWPStart,
                                 uint8_t config_type,
                                 unsigned short bwp_start_subcarrier,
diff --git a/openair1/PHY/NR_UE_TRANSPORT/nr_initial_sync.c b/openair1/PHY/NR_UE_TRANSPORT/nr_initial_sync.c
index 30f87709c4..d2a7035948 100644
--- a/openair1/PHY/NR_UE_TRANSPORT/nr_initial_sync.c
+++ b/openair1/PHY/NR_UE_TRANSPORT/nr_initial_sync.c
@@ -352,17 +352,19 @@ int nr_initial_sync(UE_nr_rxtx_proc_t *proc,
         nr_gold_pdcch(ue,fp->Nid_cell);
 
         // compute the scrambling IDs for PDSCH DMRS
-        for (int i=0; i<2; i++)
-          ue->scramblingID[i]=fp->Nid_cell;
-
-        nr_gold_pdsch(ue,ue->scramblingID);
+        for (int i=0; i<NR_NB_NSCID; i++) {
+          ue->scramblingID_dlsch[i]=fp->Nid_cell;
+          nr_gold_pdsch(ue, i, ue->scramblingID_dlsch[i]);
+        }
 
         nr_init_csi_rs(fp, ue->nr_csi_rs_info->nr_gold_csi_rs, fp->Nid_cell);
 
         // initialize the pusch dmrs
-        uint16_t N_n_scid[2] = {fp->Nid_cell,fp->Nid_cell};
-        int n_scid = 0; // This quantity is indicated by higher layer parameter dmrs-SeqInitialization
-        nr_init_pusch_dmrs(ue, N_n_scid, n_scid);
+        for (int i=0; i<NR_NB_NSCID; i++) {
+          ue->scramblingID_ulsch[i]=fp->Nid_cell;
+          nr_init_pusch_dmrs(ue, ue->scramblingID_ulsch[i], i);
+        }
+
 
         // we also need to take into account the shift by samples_per_frame in case the if is true
         if (ue->ssb_offset < sync_pos_frame){
@@ -553,6 +555,7 @@ int nr_initial_sync(UE_nr_rxtx_proc_t *proc,
                                       0,
                                       pdcch_vars->slot,
                                       l,
+                                      fp->Nid_cell,
                                       fp->first_carrier_offset+(pdcch_vars->pdcch_config[n_ss].BWPStart + coreset_start_rb)*12,
                                       coreset_nb_rb);
 
diff --git a/openair1/PHY/NR_UE_TRANSPORT/nr_transport_ue.h b/openair1/PHY/NR_UE_TRANSPORT/nr_transport_ue.h
index e479271a5c..7b5019edfe 100644
--- a/openair1/PHY/NR_UE_TRANSPORT/nr_transport_ue.h
+++ b/openair1/PHY/NR_UE_TRANSPORT/nr_transport_ue.h
@@ -259,6 +259,8 @@ typedef struct {
   uint16_t ptrs_symbols;
   // PTRS symbol index, to be updated every PTRS symbol within a slot.
   uint8_t ptrs_symbol_index;
+  uint8_t nscid;
+  uint16_t dlDmrsScramblingId;
   /// PDU BITMAP 
   uint16_t pduBitmap;
 } NR_DL_UE_HARQ_t;
diff --git a/openair1/PHY/NR_UE_TRANSPORT/nr_ulsch_ue.c b/openair1/PHY/NR_UE_TRANSPORT/nr_ulsch_ue.c
index 802cb55e03..bab0c75aff 100644
--- a/openair1/PHY/NR_UE_TRANSPORT/nr_ulsch_ue.c
+++ b/openair1/PHY/NR_UE_TRANSPORT/nr_ulsch_ue.c
@@ -32,6 +32,7 @@
 #include <stdint.h>
 #include "PHY/NR_REFSIG/dmrs_nr.h"
 #include "PHY/NR_REFSIG/ptrs_nr.h"
+#include "PHY/NR_REFSIG/refsig_defs_ue.h"
 #include "PHY/NR_UE_TRANSPORT/nr_transport_ue.h"
 #include "PHY/NR_UE_TRANSPORT/nr_transport_proto_ue.h"
 #include "PHY/MODULATION/nr_modulation.h"
@@ -215,9 +216,16 @@ void nr_ue_ulsch_procedures(PHY_VARS_NR_UE *UE,
 
   /////////////////////////DMRS Modulation/////////////////////////
   ///////////
-  uint32_t **pusch_dmrs = UE->nr_gold_pusch_dmrs[slot];
+
+  if(pusch_pdu->ul_dmrs_scrambling_id != UE->scramblingID_ulsch[pusch_pdu->scid])  {
+    UE->scramblingID_ulsch[pusch_pdu->scid] = pusch_pdu->ul_dmrs_scrambling_id;
+    nr_init_pusch_dmrs(UE, pusch_pdu->scid, pusch_pdu->ul_dmrs_scrambling_id);
+  }
+
+  uint32_t ***pusch_dmrs = UE->nr_gold_pusch_dmrs[slot];
   uint16_t n_dmrs = (pusch_pdu->bwp_start + start_rb + nb_rb)*((dmrs_type == pusch_dmrs_type1) ? 6:4);
   int16_t mod_dmrs[n_dmrs<<1] __attribute((aligned(16)));
+
   ///////////
   ////////////////////////////////////////////////////////////////////////
 
@@ -382,7 +390,7 @@ void nr_ue_ulsch_procedures(PHY_VARS_NR_UE *UE,
           // TODO: performance improvement, we can skip the modulation of DMRS symbols outside the bandwidth part
           // Perform this on gold sequence, not required when SC FDMA operation is done,
           LOG_D(PHY,"DMRS in symbol %d\n",l);
-          nr_modulation(pusch_dmrs[l], n_dmrs*2, DMRS_MOD_ORDER, mod_dmrs); // Qm = 2 as DMRS is QPSK modulated
+          nr_modulation(pusch_dmrs[l][pusch_pdu->scid], n_dmrs*2, DMRS_MOD_ORDER, mod_dmrs); // currently only codeword 0 is modulated. Qm = 2 as DMRS is QPSK modulated
         } else {
           dmrs_idx = 0;
         }
@@ -392,7 +400,7 @@ void nr_ue_ulsch_procedures(PHY_VARS_NR_UE *UE,
 
         if(is_ptrs_symbol(l, ulsch_ue->ptrs_symbols)) {
           is_ptrs_sym = 1;
-          nr_modulation(pusch_dmrs[l], nb_rb, DMRS_MOD_ORDER, mod_ptrs);
+          nr_modulation(pusch_dmrs[l][pusch_pdu->scid], nb_rb, DMRS_MOD_ORDER, mod_ptrs);
         }
       }
 
diff --git a/openair1/PHY/NR_UE_TRANSPORT/sss_nr.c b/openair1/PHY/NR_UE_TRANSPORT/sss_nr.c
index dc73fe0add..88585e6094 100644
--- a/openair1/PHY/NR_UE_TRANSPORT/sss_nr.c
+++ b/openair1/PHY/NR_UE_TRANSPORT/sss_nr.c
@@ -558,6 +558,9 @@ int rx_sss_nr(PHY_VARS_NR_UE *ue, UE_nr_rxtx_proc_t *proc, int32_t *tot_metric,
   }
   //#endif
 
+  if (Nid1==N_ID_1_NUMBER)
+    return -1;
+  
   int re = 0;
   int im = 0;
   d = (int16_t *)&d_sss[Nid2][Nid1];
diff --git a/openair1/PHY/TOOLS/oai_dfts.c b/openair1/PHY/TOOLS/oai_dfts.c
index 83207426fd..c2b6a55210 100644
--- a/openair1/PHY/TOOLS/oai_dfts.c
+++ b/openair1/PHY/TOOLS/oai_dfts.c
@@ -7082,43 +7082,40 @@ static inline void dft12f(simd_q15_t *x0,
 
   simd_q15_t tmp_dft12[12];
 
-  simd_q15_t *tmp_dft12_ptr = &tmp_dft12[0];
-
   // msg("dft12\n");
 
   bfly4_tw1(x0,
             x3,
             x6,
             x9,
-            tmp_dft12_ptr,
-            tmp_dft12_ptr+3,
-            tmp_dft12_ptr+6,
-            tmp_dft12_ptr+9);
-
+            tmp_dft12,
+            tmp_dft12+3,
+            tmp_dft12+6,
+            tmp_dft12+9);
 
   bfly4_tw1(x1,
             x4,
             x7,
             x10,
-            tmp_dft12_ptr+1,
-            tmp_dft12_ptr+4,
-            tmp_dft12_ptr+7,
-            tmp_dft12_ptr+10);
+            tmp_dft12+1,
+            tmp_dft12+4,
+            tmp_dft12+7,
+            tmp_dft12+10);
 
 
   bfly4_tw1(x2,
             x5,
             x8,
             x11,
-            tmp_dft12_ptr+2,
-            tmp_dft12_ptr+5,
-            tmp_dft12_ptr+8,
-            tmp_dft12_ptr+11);
+            tmp_dft12+2,
+            tmp_dft12+5,
+            tmp_dft12+8,
+            tmp_dft12+11);
 
   //  k2=0;
-  bfly3_tw1(tmp_dft12_ptr,
-            tmp_dft12_ptr+1,
-            tmp_dft12_ptr+2,
+  bfly3_tw1(tmp_dft12,
+            tmp_dft12+1,
+            tmp_dft12+2,
             y0,
             y4,
             y8);
@@ -7126,9 +7123,9 @@ static inline void dft12f(simd_q15_t *x0,
 
 
   //  k2=1;
-  bfly3(tmp_dft12_ptr+3,
-        tmp_dft12_ptr+4,
-        tmp_dft12_ptr+5,
+  bfly3(tmp_dft12+3,
+        tmp_dft12+4,
+        tmp_dft12+5,
         y1,
         y5,
         y9,
@@ -7138,9 +7135,9 @@ static inline void dft12f(simd_q15_t *x0,
 
 
   //  k2=2;
-  bfly3(tmp_dft12_ptr+6,
-        tmp_dft12_ptr+7,
-        tmp_dft12_ptr+8,
+  bfly3(tmp_dft12+6,
+        tmp_dft12+7,
+        tmp_dft12+8,
         y2,
         y6,
         y10,
@@ -7148,9 +7145,9 @@ static inline void dft12f(simd_q15_t *x0,
         W4_12);
 
   //  k2=3;
-  bfly3(tmp_dft12_ptr+9,
-        tmp_dft12_ptr+10,
-        tmp_dft12_ptr+11,
+  bfly3(tmp_dft12+9,
+        tmp_dft12+10,
+        tmp_dft12+11,
         y3,
         y7,
         y11,
@@ -10606,15 +10603,43 @@ int dfts_autoinit(void)
 
 #ifndef MR_MAIN
 
-void dft(uint8_t sizeidx, int16_t *sigF,int16_t *sig,unsigned char scale_flag){
-	AssertFatal((sizeidx>=0 && sizeidx<(int)DFT_SIZE_IDXTABLESIZE),"Invalid dft size index %i\n",sizeidx);
-	dft_ftab[sizeidx](sigF,sig,scale_flag);
+void dft(uint8_t sizeidx, int16_t *input,int16_t *output,unsigned char scale_flag){
+	AssertFatal((sizeidx >= 0 && sizeidx<DFT_SIZE_IDXTABLESIZE),"Invalid dft size index %i\n",sizeidx);
+        int algn=0xF;
+        #ifdef __AVX2__
+        if ( (dft_ftab[sizeidx].size%3) != 0 ) // there is no AVX2 implementation for multiples of 3 DFTs
+          algn=0x1F;
+        #endif 
+        AssertFatal(((intptr_t)output&algn)==0,"Buffers should be aligned %p",output);
+        if (((intptr_t)input)&algn) {
+          LOG_D(PHY, "DFT called with input not aligned, add a memcpy, size %d\n", sizeidx);
+          int sz=dft_ftab[sizeidx].size;
+          if (sizeidx==DFT_12) // This case does 8 DFTs in //
+            sz*=8;
+          int16_t tmp[sz*2] __attribute__ ((aligned(32))); // input and output are not in right type (int16_t instead of c16_t)
+          memcpy(tmp, input, sizeof tmp);
+          dft_ftab[sizeidx].func(tmp,output,scale_flag);
+        } else
+          dft_ftab[sizeidx].func(input,output,scale_flag);
 };
 
-void idft(uint8_t sizeidx, int16_t *sigF,int16_t *sig,unsigned char scale_flag){
-	AssertFatal((sizeidx>=0 && sizeidx<(int)IDFT_SIZE_IDXTABLESIZE),"Invalid idft size index %i\n",sizeidx);
-	idft_ftab[sizeidx](sigF,sig,scale_flag);
+void idft(uint8_t sizeidx, int16_t *input,int16_t *output,unsigned char scale_flag){
+	AssertFatal((sizeidx>=0 && sizeidx<DFT_SIZE_IDXTABLESIZE),"Invalid idft size index %i\n",sizeidx);
+        int algn=0xF;
+        #ifdef __AVX2__
+          algn=0x1F;
+        #endif
+        AssertFatal( ((intptr_t)output&algn)==0,"Buffers should be 16 bytes aligned %p",output);
+        if (((intptr_t)input)&algn ) {  
+          LOG_D(PHY, "DFT called with input not aligned, add a memcpy\n");
+          int sz=idft_ftab[sizeidx].size;
+          int16_t tmp[sz*2] __attribute__ ((aligned(32))); // input and output are not in right type (int16_t instead of c16_t)
+          memcpy(tmp, input, sizeof tmp);
+          dft_ftab[sizeidx].func(tmp,output,scale_flag);
+        } else
+          idft_ftab[sizeidx].func(input,output,scale_flag);
 };
+
 #endif
 
 /*---------------------------------------------------------------------------------------*/
diff --git a/openair1/PHY/TOOLS/tools_defs.h b/openair1/PHY/TOOLS/tools_defs.h
index 966e93440e..86bce9614f 100644
--- a/openair1/PHY/TOOLS/tools_defs.h
+++ b/openair1/PHY/TOOLS/tools_defs.h
@@ -181,159 +181,156 @@ This function performs optimized fixed-point radix-2 FFT/IFFT.
         );
 */
 
+#define FOREACH_DFTSZ(SZ_DEF) \
+  SZ_DEF(12) \
+  SZ_DEF(24) \
+  SZ_DEF(36) \
+  SZ_DEF(48) \
+  SZ_DEF(60) \
+  SZ_DEF(64) \
+  SZ_DEF(72) \
+  SZ_DEF(96) \
+  SZ_DEF(108) \
+  SZ_DEF(120) \
+  SZ_DEF(128) \
+  SZ_DEF(144) \
+  SZ_DEF(180) \
+  SZ_DEF(192) \
+  SZ_DEF(216) \
+  SZ_DEF(240) \
+  SZ_DEF(256) \
+  SZ_DEF(288) \
+  SZ_DEF(300) \
+  SZ_DEF(324) \
+  SZ_DEF(360) \
+  SZ_DEF(384) \
+  SZ_DEF(432) \
+  SZ_DEF(480) \
+  SZ_DEF(512) \
+  SZ_DEF(540) \
+  SZ_DEF(576) \
+  SZ_DEF(600) \
+  SZ_DEF(648) \
+  SZ_DEF(720) \
+  SZ_DEF(768) \
+  SZ_DEF(864) \
+  SZ_DEF(900) \
+  SZ_DEF(960) \
+  SZ_DEF(972) \
+  SZ_DEF(1024) \
+  SZ_DEF(1080) \
+  SZ_DEF(1152) \
+  SZ_DEF(1200) \
+  SZ_DEF(1296) \
+  SZ_DEF(1440) \
+  SZ_DEF(1500) \
+  SZ_DEF(1536) \
+  SZ_DEF(1620) \
+  SZ_DEF(1728) \
+  SZ_DEF(1800) \
+  SZ_DEF(1920) \
+  SZ_DEF(1944) \
+  SZ_DEF(2048) \
+  SZ_DEF(2160) \
+  SZ_DEF(2304) \
+  SZ_DEF(2400) \
+  SZ_DEF(2592) \
+  SZ_DEF(2700) \
+  SZ_DEF(2880) \
+  SZ_DEF(2916) \
+  SZ_DEF(3000) \
+  SZ_DEF(3072) \
+  SZ_DEF(3240) \
+  SZ_DEF(4096) \
+  SZ_DEF(6144) \
+  SZ_DEF(8192) \
+  SZ_DEF(9216) \
+  SZ_DEF(12288) \
+  SZ_DEF(18432) \
+  SZ_DEF(24576) \
+  SZ_DEF(36864) \
+  SZ_DEF(49152) \
+  SZ_DEF(73728) \
+  SZ_DEF(98304)
 
+#define FOREACH_IDFTSZ(SZ_DEF)\
+  SZ_DEF(64) \
+  SZ_DEF(128) \
+  SZ_DEF(256) \
+  SZ_DEF(512) \
+  SZ_DEF(1024) \
+  SZ_DEF(1536) \
+  SZ_DEF(2048) \
+  SZ_DEF(3072) \
+  SZ_DEF(4096) \
+  SZ_DEF(6144) \
+  SZ_DEF(8192) \
+  SZ_DEF(9216) \
+  SZ_DEF(12288) \
+  SZ_DEF(18432) \
+  SZ_DEF(24576) \
+  SZ_DEF(36864) \
+  SZ_DEF(49152) \
+  SZ_DEF(73728) \
+  SZ_DEF(98304)
 
 #ifdef OAIDFTS_MAIN
-typedef  void(*adftfunc_t)(int16_t *sigF,int16_t *sig,unsigned char scale_flag);  
-typedef  void(*aidftfunc_t)(int16_t *sigF,int16_t *sig,unsigned char scale_flag);     
+typedef  void(*adftfunc_t)(int16_t *sigF,int16_t *sig,unsigned char scale_flag);
+typedef  void(*aidftfunc_t)(int16_t *sigF,int16_t *sig,unsigned char scale_flag);
 
-void dft12(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft24(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft36(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft48(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft60(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft64(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft72(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft96(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft108(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft120(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft128(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft144(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft180(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft192(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft216(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft240(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft256(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft288(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft300(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft324(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft360(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft384(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft432(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft480(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft512(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft540(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft576(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft600(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft648(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft720(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft768(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft864(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft900(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft960(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft972(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1024(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1080(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1152(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1200(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1296(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1440(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1500(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1536(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void dft1620(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1728(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1800(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1920(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1944(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2048(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2160(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2304(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2400(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2592(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2700(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2880(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2916(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft3000(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft3072(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void dft3240(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft4096(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft6144(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void dft8192(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft9216(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft12288(int16_t *x,int16_t *y,uint8_t scale_flag);  
-void dft18432(int16_t *x,int16_t *y,uint8_t scale_flag); 
-void dft24576(int16_t *x,int16_t *y,uint8_t scale_flag); 
-void dft36864(int16_t *x,int16_t *y,uint8_t scale_flag); 
-void dft49152(int16_t *x,int16_t *y,uint8_t scale_flag); 
-void dft73728(int16_t *x,int16_t *y,uint8_t scale_flag); 
-void dft98304(int16_t *x,int16_t *y,uint8_t scale_flag);
-
-
-void idft64(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft128(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft256(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft512(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft1024(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft1536(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft2048(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft3072(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft4096(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft6144(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft8192(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft9216(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft12288(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft18432(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft24576(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft36864(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft49152(int16_t *sigF,int16_t *sig,uint8_t scale_flag); 
-void idft73728(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft98304(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
+#define SZ_FUNC(Sz) void dft ## Sz(int16_t *x,int16_t *y,uint8_t scale_flag);
 
+FOREACH_DFTSZ(SZ_FUNC)
 
+#define SZ_iFUNC(Sz) void idft ## Sz(int16_t *x,int16_t *y,uint8_t scale_flag);
 
+FOREACH_IDFTSZ(SZ_iFUNC)
 
 #else
-  typedef  void(*dftfunc_t)(uint8_t sizeidx,int16_t *sigF,int16_t *sig,unsigned char scale_flag);  
-  typedef  void(*idftfunc_t)(uint8_t sizeidx,int16_t *sigF,int16_t *sig,unsigned char scale_flag);  
+typedef  void(*dftfunc_t)(uint8_t sizeidx,int16_t *sigF,int16_t *sig,unsigned char scale_flag);
+typedef  void(*idftfunc_t)(uint8_t sizeidx,int16_t *sigF,int16_t *sig,unsigned char scale_flag);
 #  ifdef OAIDFTS_LOADER
-  dftfunc_t dft;
-  idftfunc_t idft;
+dftfunc_t dft;
+idftfunc_t idft;
 #  else
-  extern dftfunc_t dft;
-  extern idftfunc_t idft;
-  extern int load_dftslib(void);
+extern dftfunc_t dft;
+extern idftfunc_t idft;
+extern int load_dftslib(void);
 #  endif
 #endif
 
-typedef enum DFT_size_idx {
-	DFT_12,    DFT_24,    DFT_36,   DFT_48,     DFT_60,   DFT_72,   DFT_96,
-	DFT_108,   DFT_120,   DFT_128,  DFT_144,    DFT_180,  DFT_192,  DFT_216,   DFT_240,
-	DFT_256,   DFT_288,   DFT_300,  DFT_324,    DFT_360,  DFT_384,  DFT_432,   DFT_480,
-	DFT_512,   DFT_540,   DFT_576,  DFT_600,    DFT_648,  DFT_720,  DFT_768,   DFT_864,
-	DFT_900,   DFT_960,   DFT_972,  DFT_1024,   DFT_1080, DFT_1152, DFT_1200,  DFT_1296,
-	DFT_1440,  DFT_1500,  DFT_1536, DFT_1620,   DFT_1728, DFT_1800, DFT_1920,  DFT_1944,
-	DFT_2048,  DFT_2160,  DFT_2304, DFT_2400,   DFT_2592, DFT_2700, DFT_2880,  DFT_2916,
-	DFT_3000,  DFT_3072,  DFT_3240, DFT_4096,   DFT_6144, DFT_8192, DFT_9216,  DFT_12288,
-	DFT_18432, DFT_24576, DFT_36864, DFT_49152, DFT_73728, DFT_98304,
-	DFT_SIZE_IDXTABLESIZE
-} dft_size_idx_t;
+#define SZ_ENUM(Sz) DFT_ ## Sz,
 
-#ifdef OAIDFTS_MAIN
-adftfunc_t dft_ftab[]={
-	dft12,    dft24,    dft36,    dft48,    dft60,   dft72,   dft96,
-	dft108,   dft120,   dft128,   dft144,   dft180,  dft192,  dft216,   dft240,
-	dft256,   dft288,   dft300,   dft324,   dft360,  dft384,  dft432,   dft480,
-	dft512,   dft540,   dft576,   dft600,   dft648,  dft720,  dft768,   dft864,
-	dft900,   dft960,   dft972,   dft1024,  dft1080, dft1152, dft1200,  dft1296,
-	dft1440,  dft1500,  dft1536,  dft1620,  dft1728, dft1800, dft1920,  dft1944,
-	dft2048,  dft2160,  dft2304,  dft2400,  dft2592, dft2700, dft2880,  dft2916,
-	dft3000,  dft3072,  dft3240,  dft4096,  dft6144, dft8192, dft9216,  dft12288,
-	dft18432, dft24576, dft36864, dft49152, dft73728, dft98304
-};
-#endif
+typedef enum dft_size_idx {
+  FOREACH_DFTSZ(SZ_ENUM)
+  DFT_SIZE_IDXTABLESIZE
+}  dft_size_idx_t;
+
+#define SZ_iENUM(Sz) IDFT_ ## Sz,
 
 typedef enum idft_size_idx {
-	IDFT_128,   IDFT_256,  IDFT_512,   IDFT_1024,  IDFT_1536,  IDFT_2048,  IDFT_3072,  IDFT_4096,
-	IDFT_6144,  IDFT_8192, IDFT_9216,  IDFT_12288, IDFT_18432, IDFT_24576, IDFT_36864, IDFT_49152, 
-	IDFT_73728, IDFT_98304, 
-	IDFT_SIZE_IDXTABLESIZE
-} idft_size_idx_t;
+  FOREACH_IDFTSZ(SZ_iENUM)
+  IDFT_SIZE_IDXTABLESIZE
+}  idft_size_idx_t;
+
 #ifdef OAIDFTS_MAIN
-aidftfunc_t idft_ftab[]={
-	idft128,   idft256,  idft512,   idft1024,  idft1536,  idft2048,  idft3072,  idft4096,
-	idft6144,  idft8192, idft9216,  idft12288, idft18432, idft24576, idft36864, idft49152,
-	idft73728, idft98304
+
+#define SZ_PTR(Sz) {dft ## Sz,Sz},
+struct {
+  adftfunc_t func;
+  int size;
+} dft_ftab[]= {
+  FOREACH_DFTSZ(SZ_PTR)
 };
+
+#define SZ_iPTR(Sz)  {idft ## Sz,Sz},
+struct {
+  adftfunc_t func;
+  int size;
+} idft_ftab[]= {
+  FOREACH_IDFTSZ(SZ_iPTR)
+};
+
 #endif
 
 
@@ -387,7 +384,7 @@ int32_t add_real_vector64(int16_t *x,
                           uint32_t N);
 
 int32_t sub_real_vector64(int16_t *x,
-                          int16_t* y,
+                          int16_t *y,
                           int16_t *z,
                           uint32_t N);
 
@@ -444,7 +441,7 @@ int32_t signal_energy_amp_shift(int32_t *input, uint32_t length);
 /*!\fn int32_t signal_energy(int *,uint32_t);
 \brief Computes the signal energy per subcarrier
 */
-int32_t subcarrier_energy(int32_t *,uint32_t, int32_t* subcarrier_energy, uint16_t rx_power_correction);
+int32_t subcarrier_energy(int32_t *,uint32_t, int32_t *subcarrier_energy, uint16_t rx_power_correction);
 #endif
 
 /*!\fn int32_t signal_energy_nodc(int32_t *,uint32_t);
diff --git a/openair1/PHY/defs_gNB.h b/openair1/PHY/defs_gNB.h
index 100101c38e..24d813ebc1 100644
--- a/openair1/PHY/defs_gNB.h
+++ b/openair1/PHY/defs_gNB.h
@@ -850,9 +850,15 @@ typedef struct PHY_VARS_gNB_s {
   /// counter to average prach energh over first 100 prach opportunities
   int prach_energy_counter;
 
+  int csi_gold_init;
+  int pdcch_gold_init;
+  int pdsch_gold_init[2];
+  int pusch_gold_init[2];
+
   int ap_N1;
   int ap_N2;
   int ap_XP;
+
   int pucch0_thres;
   int pusch_thres;
   int prach_thres;
diff --git a/openair1/PHY/defs_nr_UE.h b/openair1/PHY/defs_nr_UE.h
index 4175d96780..5436c35db0 100644
--- a/openair1/PHY/defs_nr_UE.h
+++ b/openair1/PHY/defs_nr_UE.h
@@ -801,7 +801,10 @@ typedef struct {
   uint32_t ****nr_gold_pdsch[NUMBER_OF_CONNECTED_eNB_MAX];
 
   // Scrambling IDs used in PDSCH DMRS
-  uint16_t scramblingID[2];
+  uint16_t scramblingID_dlsch[2];
+
+  // Scrambling IDs used in PUSCH DMRS
+  uint16_t scramblingID_ulsch[2];
 
   /// PDCCH DMRS
   uint32_t ***nr_gold_pdcch[NUMBER_OF_CONNECTED_eNB_MAX];
@@ -810,7 +813,7 @@ typedef struct {
   uint16_t scramblingID_pdcch;
 
   /// PUSCH DMRS sequence
-  uint32_t ***nr_gold_pusch_dmrs;
+  uint32_t ****nr_gold_pusch_dmrs;
 
   uint32_t X_u[64][839];
 
diff --git a/openair1/PHY/defs_nr_common.h b/openair1/PHY/defs_nr_common.h
index aab9601f80..f436d249c4 100644
--- a/openair1/PHY/defs_nr_common.h
+++ b/openair1/PHY/defs_nr_common.h
@@ -95,6 +95,8 @@
 #define NR_RX_NB_TH 1
 #define NR_NB_TH_SLOT 2
 
+#define NR_NB_NSCID 2
+
 extern const uint8_t nr_rv_round_map[4]; 
 
 static inline
diff --git a/openair1/SCHED_NR/phy_procedures_nr_gNB.c b/openair1/SCHED_NR/phy_procedures_nr_gNB.c
index fb689cf83e..9eb63898ad 100644
--- a/openair1/SCHED_NR/phy_procedures_nr_gNB.c
+++ b/openair1/SCHED_NR/phy_procedures_nr_gNB.c
@@ -158,8 +158,7 @@ void phy_procedures_gNB_TX(processingData_L1tx_t *msgTx,
   
     VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_PHY_gNB_PDCCH_TX,1);
 
-    nr_generate_dci_top(msgTx,
-			gNB->nr_gold_pdcch_dmrs[slot],
+    nr_generate_dci_top(msgTx, slot,
 			&gNB->common_vars.txdataF[0][txdataF_offset],
 			AMP, fp);
 
diff --git a/openair1/SCHED_NR_UE/fapi_nr_ue_l1.c b/openair1/SCHED_NR_UE/fapi_nr_ue_l1.c
index bac3fa33c9..98c4213aeb 100644
--- a/openair1/SCHED_NR_UE/fapi_nr_ue_l1.c
+++ b/openair1/SCHED_NR_UE/fapi_nr_ue_l1.c
@@ -315,6 +315,8 @@ void configure_dlsch(NR_UE_DLSCH_t *dlsch0,
   dlsch0_harq->mcs = dlsch_config_pdu->mcs;
   dlsch0_harq->rvidx = dlsch_config_pdu->rv;
   dlsch0->g_pucch = dlsch_config_pdu->accumulated_delta_PUCCH;
+  dlsch0_harq->nscid = dlsch_config_pdu->nscid;
+  dlsch0_harq->dlDmrsScramblingId = dlsch_config_pdu->dlDmrsScramblingId;
   //get nrOfLayers from DCI info
   uint8_t Nl = 0;
   for (int i = 0; i < 12; i++) { // max 12 ports
diff --git a/openair1/SCHED_NR_UE/phy_procedures_nr_ue.c b/openair1/SCHED_NR_UE/phy_procedures_nr_ue.c
index 7e7581ecc3..262acdec3a 100644
--- a/openair1/SCHED_NR_UE/phy_procedures_nr_ue.c
+++ b/openair1/SCHED_NR_UE/phy_procedures_nr_ue.c
@@ -580,6 +580,8 @@ int nr_ue_pdsch_procedures(PHY_VARS_NR_UE *ue, UE_nr_rxtx_proc_t *proc, int gNB_
                                       nr_slot_rx,
                                       get_dmrs_port(aatx,dlsch0_harq->dmrs_ports),
                                       m,
+                                      dlsch0_harq->nscid,
+                                      dlsch0_harq->dlDmrsScramblingId,
                                       BWPStart,
                                       dlsch0_harq->dmrsConfigType,
                                       ue->frame_parms.first_carrier_offset+(BWPStart + pdsch_start_rb)*12,
@@ -1481,6 +1483,7 @@ int phy_procedures_nrUE_RX(PHY_VARS_NR_UE *ue,
                                     gNB_id,
                                     nr_slot_rx,
                                     l,
+                                    pdcch_vars->pdcch_config[n_ss].coreset.pdcch_dmrs_scrambling_id,
                                     fp->first_carrier_offset+(pdcch_vars->pdcch_config[n_ss].BWPStart + coreset_start_rb)*12,
                                     coreset_nb_rb);
 
diff --git a/openair1/SIMULATION/NR_PHY/dlsim.c b/openair1/SIMULATION/NR_PHY/dlsim.c
index 8f6a453ee5..412ab1a6ea 100644
--- a/openair1/SIMULATION/NR_PHY/dlsim.c
+++ b/openair1/SIMULATION/NR_PHY/dlsim.c
@@ -961,10 +961,10 @@ int main(int argc, char **argv)
   nr_gold_pdcch(UE, frame_parms->Nid_cell);
 
   // compute the scrambling IDs for PDSCH DMRS
-  for (int i = 0; i < 2; i++)
-    UE->scramblingID[i] = frame_parms->Nid_cell;
-
-  nr_gold_pdsch(UE, UE->scramblingID);
+  for (int i = 0; i < 2; i++) {
+    UE->scramblingID_dlsch[i] = frame_parms->Nid_cell;
+    nr_gold_pdsch(UE, i, UE->scramblingID_dlsch[i]);
+  }
 
   nr_l2_init_ue(NULL);
   UE_mac = get_mac_inst(0);
diff --git a/openair1/SIMULATION/NR_PHY/ulsim.c b/openair1/SIMULATION/NR_PHY/ulsim.c
index 05dec6c6be..bf635cbc85 100644
--- a/openair1/SIMULATION/NR_PHY/ulsim.c
+++ b/openair1/SIMULATION/NR_PHY/ulsim.c
@@ -782,10 +782,10 @@ int main(int argc, char **argv)
 
   init_nr_ue_transport(UE);
 
-  // initialize the pusch dmrs
-  uint16_t N_n_scid[2] = {frame_parms->Nid_cell,frame_parms->Nid_cell};
-  int n_scid = 0; // This quantity is indicated by higher layer parameter dmrs-SeqInitialization
-  nr_init_pusch_dmrs(UE, N_n_scid, n_scid);
+  for(int n_scid = 0; n_scid<2; n_scid++) {
+    UE->scramblingID_ulsch[n_scid] = frame_parms->Nid_cell;
+    nr_init_pusch_dmrs(UE, frame_parms->Nid_cell, n_scid);
+  }
 
   //Configure UE
   NR_UE_RRC_INST_t rrcue;
diff --git a/openair2/GNB_APP/gnb_paramdef.h b/openair2/GNB_APP/gnb_paramdef.h
index eb6b2bdccd..8cd831970f 100644
--- a/openair2/GNB_APP/gnb_paramdef.h
+++ b/openair2/GNB_APP/gnb_paramdef.h
@@ -154,7 +154,7 @@ typedef enum {
 {GNB_CONFIG_STRING_PDSCHANTENNAPORTS_N2, "vert. log. antenna ports", 0, iptr:NULL,  defintval:1,                 TYPE_INT,       0},  \
 {GNB_CONFIG_STRING_PDSCHANTENNAPORTS_XP, "XP log. antenna ports",   0, iptr:NULL,   defintval:1,                 TYPE_INT,       0},  \
 {GNB_CONFIG_STRING_PUSCHANTENNAPORTS,            NULL,   0,            iptr:NULL,   defintval:1,                 TYPE_INT,       0},  \
-{GNB_CONFIG_STRING_SIB1TDA,                      NULL,   0,            iptr:NULL,   defintval:0,                 TYPE_INT,       0},  \
+{GNB_CONFIG_STRING_SIB1TDA,                      NULL,   0,            iptr:NULL,   defintval:1,                 TYPE_INT,       0},  \
 {GNB_CONFIG_STRING_DOCSIRS,                      NULL,   0,            iptr:NULL,   defintval:0,                 TYPE_INT,       0},  \
 {GNB_CONFIG_STRING_DOSRS,                        NULL,   0,            iptr:NULL,   defintval:0,                 TYPE_INT,       0},  \
 {GNB_CONFIG_STRING_NRCELLID,                     NULL,   0,            u64ptr:NULL, defint64val:1,               TYPE_UINT64,    0},  \
diff --git a/openair2/LAYER2/NR_MAC_UE/nr_ue_procedures.c b/openair2/LAYER2/NR_MAC_UE/nr_ue_procedures.c
index 4b7635929c..192fb25405 100644
--- a/openair2/LAYER2/NR_MAC_UE/nr_ue_procedures.c
+++ b/openair2/LAYER2/NR_MAC_UE/nr_ue_procedures.c
@@ -281,7 +281,6 @@ int get_rnti_type(NR_UE_MAC_INST_t *mac, uint16_t rnti){
     LOG_D(MAC, "In %s: returning rnti_type %s \n", __FUNCTION__, rnti_types[rnti_type]);
 
     return rnti_type;
-
 }
 
 
@@ -874,6 +873,18 @@ int8_t nr_ue_process_dci(module_id_t module_id, int cc_id, uint8_t gNB_index, fr
     if(pdsch_TimeDomainAllocationList && rnti!=SI_RNTI)
       mappingtype = pdsch_TimeDomainAllocationList->list.array[dci->time_domain_assignment.val]->mappingType;
 
+    struct NR_DMRS_DownlinkConfig *dl_dmrs_config = NULL;
+    if(mac->DLbwp[0])
+      dl_dmrs_config = (mappingtype == typeA) ?
+                       mac->DLbwp[0]->bwp_Dedicated->pdsch_Config->choice.setup->dmrs_DownlinkForPDSCH_MappingTypeA->choice.setup :
+                       mac->DLbwp[0]->bwp_Dedicated->pdsch_Config->choice.setup->dmrs_DownlinkForPDSCH_MappingTypeB->choice.setup;
+
+    dlsch_config_pdu_1_0->nscid = 0;
+    if(dl_dmrs_config && dl_dmrs_config->scramblingID0)
+      dlsch_config_pdu_1_0->dlDmrsScramblingId = *dl_dmrs_config->scramblingID0;
+    else
+      dlsch_config_pdu_1_0->dlDmrsScramblingId = mac->physCellId;
+
     /* dmrs symbol positions*/
     dlsch_config_pdu_1_0->dlDmrsSymbPos = fill_dmrs_mask(pdsch_config,
                                                          (get_softmodem_params()->nsa) ? mac->scc->dmrs_TypeA_Position : mac->mib->dmrs_TypeA_Position,
@@ -881,8 +892,9 @@ int8_t nr_ue_process_dci(module_id_t module_id, int cc_id, uint8_t gNB_index, fr
                                                          dlsch_config_pdu_1_0->start_symbol,
                                                          mappingtype,
                                                          1);
-    dlsch_config_pdu_1_0->dmrsConfigType = (mac->DLbwp[0] != NULL) ?
-                                           (mac->DLbwp[0]->bwp_Dedicated->pdsch_Config->choice.setup->dmrs_DownlinkForPDSCH_MappingTypeA->choice.setup->dmrs_Type == NULL ? 0 : 1) : 0;
+    dlsch_config_pdu_1_0->dmrsConfigType = (dl_dmrs_config != NULL) ?
+                                           (dl_dmrs_config->dmrs_Type == NULL ? 0 : 1) : 0;
+
     /* number of DM-RS CDM groups without data according to subclause 5.1.6.2 of 3GPP TS 38.214 version 15.9.0 Release 15 */
     if (dlsch_config_pdu_1_0->number_symbols == 2)
       dlsch_config_pdu_1_0->n_dmrs_cdm_groups = 1;
@@ -1075,7 +1087,30 @@ int8_t nr_ue_process_dci(module_id_t module_id, int cc_id, uint8_t gNB_index, fr
     if(pdsch_TimeDomainAllocationList)
       mappingtype = pdsch_TimeDomainAllocationList->list.array[dci->time_domain_assignment.val]->mappingType;
 
-    dlsch_config_pdu_1_1->dmrsConfigType = pdsch_Config->dmrs_DownlinkForPDSCH_MappingTypeA->choice.setup->dmrs_Type == NULL ? NFAPI_NR_DMRS_TYPE1 : NFAPI_NR_DMRS_TYPE2;
+    struct NR_DMRS_DownlinkConfig *dl_dmrs_config = (mappingtype == typeA) ?
+                                                    pdsch_Config->dmrs_DownlinkForPDSCH_MappingTypeA->choice.setup :
+                                                    pdsch_Config->dmrs_DownlinkForPDSCH_MappingTypeB->choice.setup;
+
+    switch (dci->dmrs_sequence_initialization.val) {
+      case 0:
+        dlsch_config_pdu_1_1->nscid = 0;
+        if(dl_dmrs_config->scramblingID0)
+          dlsch_config_pdu_1_1->dlDmrsScramblingId = *dl_dmrs_config->scramblingID0;
+        else
+          dlsch_config_pdu_1_1->dlDmrsScramblingId = mac->physCellId;
+        break;
+      case 1:
+        dlsch_config_pdu_1_1->nscid = 1;
+        if(dl_dmrs_config->scramblingID1)
+          dlsch_config_pdu_1_1->dlDmrsScramblingId = *dl_dmrs_config->scramblingID1;
+        else
+          dlsch_config_pdu_1_1->dlDmrsScramblingId = mac->physCellId;
+        break;
+      default:
+        AssertFatal(1==0,"Invalid dmrs sequence initialization value\n");
+    }
+
+    dlsch_config_pdu_1_1->dmrsConfigType = dl_dmrs_config->dmrs_Type == NULL ? NFAPI_NR_DMRS_TYPE1 : NFAPI_NR_DMRS_TYPE2;
 
     /* TODO: fix number of DM-RS CDM groups without data according to subclause 5.1.6.2 of 3GPP TS 38.214,
              using tables 7.3.1.2.2-1, 7.3.1.2.2-2, 7.3.1.2.2-3, 7.3.1.2.2-4 of 3GPP TS 38.212 */
@@ -1136,17 +1171,11 @@ int8_t nr_ue_process_dci(module_id_t module_id, int cc_id, uint8_t gNB_index, fr
 
     /* ANTENNA_PORTS */
     uint8_t n_codewords = 1; // FIXME!!!
-    long *max_length = NULL;
-    long *dmrs_type = NULL;
+    long *max_length = dl_dmrs_config->maxLength;
+    long *dmrs_type = dl_dmrs_config->dmrs_Type;
+
     dlsch_config_pdu_1_1->n_front_load_symb = 1; // default value
-    if (pdsch_Config->dmrs_DownlinkForPDSCH_MappingTypeA) {
-      max_length = pdsch_Config->dmrs_DownlinkForPDSCH_MappingTypeA->choice.setup->maxLength;
-      dmrs_type = pdsch_Config->dmrs_DownlinkForPDSCH_MappingTypeA->choice.setup->dmrs_Type;
-    }
-    if (pdsch_Config->dmrs_DownlinkForPDSCH_MappingTypeB) {
-      max_length = pdsch_Config->dmrs_DownlinkForPDSCH_MappingTypeB->choice.setup->maxLength;
-      dmrs_type = pdsch_Config->dmrs_DownlinkForPDSCH_MappingTypeB->choice.setup->dmrs_Type;
-    }
+
     if ((dmrs_type == NULL) && (max_length == NULL)){
       // Table 7.3.1.2.2-1: Antenna port(s) (1000 + DMRS port), dmrs-Type=1, maxLength=1
       dlsch_config_pdu_1_1->n_dmrs_cdm_groups = table_7_3_2_3_3_1[dci->antenna_ports.val][0];
diff --git a/openair2/LAYER2/NR_MAC_UE/nr_ue_scheduler.c b/openair2/LAYER2/NR_MAC_UE/nr_ue_scheduler.c
index d2532cfff8..ff2e08ac36 100644
--- a/openair2/LAYER2/NR_MAC_UE/nr_ue_scheduler.c
+++ b/openair2/LAYER2/NR_MAC_UE/nr_ue_scheduler.c
@@ -734,6 +734,11 @@ int nr_config_pusch_pdu(NR_UE_MAC_INST_t *mac,
                          ? pusch_Config->dmrs_UplinkForPUSCH_MappingTypeA->choice.setup : pusch_Config->dmrs_UplinkForPUSCH_MappingTypeB->choice.setup;
     }
 
+    pusch_config_pdu->scid = 0;
+    pusch_config_pdu->ul_dmrs_scrambling_id = mac->physCellId;
+    if(*dci_format == NR_UL_DCI_FORMAT_0_1)
+      pusch_config_pdu->scid = dci->dmrs_sequence_initialization.val;
+
     /* TRANSFORM PRECODING ------------------------------------------------------------------------------------------*/
 
     if (pusch_config_pdu->transform_precoding == NR_PUSCH_Config__transformPrecoder_enabled) {
@@ -759,6 +764,14 @@ int nr_config_pusch_pdu(NR_UE_MAC_INST_t *mac,
       LOG_D(NR_MAC,"TRANSFORM PRECODING IS ENABLED. CDM groups: %d, U: %d \n", pusch_config_pdu->num_dmrs_cdm_grps_no_data,
                 pusch_config_pdu->dfts_ofdm.low_papr_group_number);
     }
+    else {
+      if (pusch_config_pdu->scid == 0 &&
+          NR_DMRS_ulconfig->transformPrecodingDisabled->scramblingID0)
+        pusch_config_pdu->ul_dmrs_scrambling_id = *NR_DMRS_ulconfig->transformPrecodingDisabled->scramblingID0;
+      if (pusch_config_pdu->scid == 1 &&
+          NR_DMRS_ulconfig->transformPrecodingDisabled->scramblingID1)
+        pusch_config_pdu->ul_dmrs_scrambling_id = *NR_DMRS_ulconfig->transformPrecodingDisabled->scramblingID1;
+    }
 
     /* TRANSFORM PRECODING --------------------------------------------------------------------------------------------------------*/
 
diff --git a/openair2/LAYER2/NR_MAC_gNB/gNB_scheduler_bch.c b/openair2/LAYER2/NR_MAC_gNB/gNB_scheduler_bch.c
index 8c9d7e8bc8..4b654e2ad1 100644
--- a/openair2/LAYER2/NR_MAC_gNB/gNB_scheduler_bch.c
+++ b/openair2/LAYER2/NR_MAC_gNB/gNB_scheduler_bch.c
@@ -621,6 +621,8 @@ void schedule_nr_sib1(module_id_t module_idP, frame_t frameP, sub_frame_t slotP)
                                1, &is_typeA,
                                &startSymbolIndex, &nrOfSymbols);
 
+      AssertFatal((startSymbolIndex+nrOfSymbols)<14,"SIB1 TDA %d would cause overlap with CSI-RS. Please select a different SIB1 TDA.\n",time_domain_allocation);
+
       int mappingtype = is_typeA? typeA: typeB;
       uint16_t dlDmrsSymbPos = fill_dmrs_mask(NULL, gNB_mac->common_channels->ServingCellConfigCommon->dmrs_TypeA_Position, nrOfSymbols, startSymbolIndex, mappingtype, 1);
 
diff --git a/openair2/LAYER2/NR_MAC_gNB/mac_proto.h b/openair2/LAYER2/NR_MAC_gNB/mac_proto.h
index 784a71f23f..399033abd0 100644
--- a/openair2/LAYER2/NR_MAC_gNB/mac_proto.h
+++ b/openair2/LAYER2/NR_MAC_gNB/mac_proto.h
@@ -519,4 +519,7 @@ bool nr_find_nb_rb(uint16_t Qm,
 void nr_sr_reporting(int Mod_idP, frame_t frameP, sub_frame_t slotP);
 
 void dump_mac_stats(gNB_MAC_INST *gNB, char *output, int strlen, bool reset_rsrp);
+
+void process_CellGroup(NR_CellGroupConfig_t *CellGroup, NR_UE_sched_ctrl_t *sched_ctrl);
+
 #endif /*__LAYER2_NR_MAC_PROTO_H__*/
diff --git a/openair2/RRC/NR/MESSAGES/asn1_msg.c b/openair2/RRC/NR/MESSAGES/asn1_msg.c
index a972f2a946..7f3f4cd51d 100755
--- a/openair2/RRC/NR/MESSAGES/asn1_msg.c
+++ b/openair2/RRC/NR/MESSAGES/asn1_msg.c
@@ -1713,15 +1713,14 @@ void fill_mastercellGroupConfig(NR_CellGroupConfig_t *cellGroupConfig, NR_CellGr
 }
 
 void update_cellGroupConfig(NR_CellGroupConfig_t *cellGroupConfig,
-                            rrc_gNB_carrier_data_t *carrier,
                             int uid,
                             NR_UE_NR_Capability_t *uecap,
-                            const gNB_RrcConfigurationReq* configuration)
-{
+                            const gNB_RrcConfigurationReq* configuration) {
+
   NR_SpCellConfig_t *SpCellConfig = cellGroupConfig->spCellConfig;
   if (SpCellConfig == NULL) return;
 
-  NR_ServingCellConfigCommon_t *scc = carrier->servingcellconfigcommon;
+  NR_ServingCellConfigCommon_t *scc = configuration->scc;
 
   // Config CSI-RS
   if (configuration->do_CSIRS) {
@@ -1752,7 +1751,7 @@ void update_cellGroupConfig(NR_CellGroupConfig_t *cellGroupConfig,
     for (int i=0; i<DL_BWP_list->list.count; i++){
       NR_BWP_Downlink_t *bwp = DL_BWP_list->list.array[i];
       int scs = bwp->bwp_Common->genericParameters.subcarrierSpacing;
-      set_dl_mcs_table(scs, configuration->force_256qam_off ? NULL : uecap, bwp->bwp_Dedicated, carrier->servingcellconfigcommon);
+      set_dl_mcs_table(scs, configuration->force_256qam_off ? NULL : uecap, bwp->bwp_Dedicated, scc);
     }
   }
 }
@@ -2231,10 +2230,10 @@ int16_t do_RRCReconfiguration(
 
     if(cellGroupConfig!=NULL){
       update_cellGroupConfig(cellGroupConfig,
-                             carrier,
                              ue_context_pP->local_uid,
                              ue_context_pP->ue_context.UE_Capability_nr,
                              configuration);
+
       enc_rval = uper_encode_to_buffer(&asn_DEF_NR_CellGroupConfig,
           NULL,
           (void *)cellGroupConfig,
diff --git a/openair2/RRC/NR/MESSAGES/asn1_msg.h b/openair2/RRC/NR/MESSAGES/asn1_msg.h
index 1275b463ae..2b11d170f4 100644
--- a/openair2/RRC/NR/MESSAGES/asn1_msg.h
+++ b/openair2/RRC/NR/MESSAGES/asn1_msg.h
@@ -112,7 +112,6 @@ void fill_initial_cellGroupConfig(int uid,
                                   const gNB_RrcConfigurationReq *configuration);
 
 void update_cellGroupConfig(NR_CellGroupConfig_t *cellGroupConfig,
-                            rrc_gNB_carrier_data_t *carrier,
                             int uid,
                             NR_UE_NR_Capability_t *uecap,
                             const gNB_RrcConfigurationReq *configuration);
diff --git a/openair2/RRC/NR/nr_rrc_config.c b/openair2/RRC/NR/nr_rrc_config.c
index 2dd47a80f8..373e2f3417 100644
--- a/openair2/RRC/NR/nr_rrc_config.c
+++ b/openair2/RRC/NR/nr_rrc_config.c
@@ -38,6 +38,7 @@ void set_csirs_periodicity(NR_NZP_CSI_RS_Resource_t *nzpcsi0, int uid, int nb_sl
 
   nzpcsi0->periodicityAndOffset = calloc(1,sizeof(*nzpcsi0->periodicityAndOffset));
   int ideal_period = nb_slots_per_period*MAX_MOBILES_PER_GNB;
+
   if (ideal_period<5) {
     nzpcsi0->periodicityAndOffset->present = NR_CSI_ResourcePeriodicityAndOffset_PR_slots4;
     nzpcsi0->periodicityAndOffset->choice.slots4 = nb_slots_per_period*uid;