diff --git a/matlab/tests/ulsch_test.m b/matlab/tests/ulsch_test.m
index 8a4ada651..013ebca1d 100644
--- a/matlab/tests/ulsch_test.m
+++ b/matlab/tests/ulsch_test.m
@@ -2,13 +2,12 @@ clear
 ueConfig=struct('NCellID',1,'CyclicPrefixUL','Normal','NTxAnts',1);
 puschConfig=struct('NLayers',1,'OrthCover','Off','Shortened',0);
 
-addpath('../../build/srslte/lib/phch/test')
+addpath('/home/ismael/work/srsLTE/debug/srslte/lib/phch/test')
 
-
-cqilen=0;
-mods={'QPSK'};
+cqilen=[0 20];
+mods={'QPSK','16QAM', '64QAM'};
 rvs=0;
-betas=0;
+betas=[0 2.0 2.5 5.0, 20.0];
 for p=1:100
     for i=0:26
         for m=1:length(mods)
@@ -17,10 +16,11 @@ for p=1:100
                     for back=1:length(betas)          
                         for c=1:length(cqilen)
                             puschConfig.PRBSet=(0:p-1)';
+                            
                             TBs=lteTBS(length(puschConfig.PRBSet),i);
-                            %TBs=24496;
-                            trblkin=randi(2,TBs,1)-1;
-                            %trblkin=ones(1,TBs);
+                            %TBs=256;
+                            %trblkin=randi(2,TBs,1)-1;
+                            trblkin=ones(1,TBs);
                             
                             puschConfig.Modulation = mods{m};
                             puschConfig.RV = rvs(r);
@@ -28,24 +28,22 @@ for p=1:100
                             puschConfig.BetaRI = betas(bri);
                             puschConfig.BetaACK = betas(back);
 
-                            if (betas(bri)>0)
-                                ri_bit=randi(2,1,1)-1;
-                            else
-                                ri_bit=[];
-                            end
-                            if (betas(back)>0)
-                                ack_bit=randi(2,1,1)-1;
-                            else
-                                ack_bit=[];
-                            end
-
+                           if (betas(bri)>0)
+                               ri_bit=randi(2,1,1)-1;
+                           else
+                               ri_bit=[];
+                           end
+                           if (betas(back)>0)
+                               ack_bit=randi(2,1,1)-1;
+                           else
+                               ack_bit=[];
+                           end
+    
                             if (cqilen(c)>0 || TBs>0)
                                 [lib]=srslte_ulsch_encode(ueConfig,puschConfig,trblkin,ones(1,cqilen(c)),ri_bit,ack_bit);
-                                lib(lib==192)=3;
-                                lib(lib==48)=2;
                                 [mat, info]=lteULSCH(ueConfig,puschConfig,trblkin,ones(1,cqilen(c)),ri_bit,ack_bit,[]);
-                                mat(mat==-2)=3;
-                                mat(mat==-1)=2;
+                                mat(mat==-2)=0;
+                                mat(mat==-1)=0;
                                 err=sum(abs(double(mat)-double(lib)));
                                 if (err > 0)
                                   disp(err)    
diff --git a/srslte/include/srslte/common/sequence.h b/srslte/include/srslte/common/sequence.h
index 6a22f8876..6d33a7d3c 100644
--- a/srslte/include/srslte/common/sequence.h
+++ b/srslte/include/srslte/common/sequence.h
@@ -42,6 +42,7 @@
 
 typedef struct SRSLTE_API {
   uint8_t *c;
+  uint8_t *c_bytes;
   uint32_t len;
 } srslte_sequence_t;
 
diff --git a/srslte/include/srslte/phch/sch.h b/srslte/include/srslte/phch/sch.h
index 90811320a..99b4c76bd 100644
--- a/srslte/include/srslte/phch/sch.h
+++ b/srslte/include/srslte/phch/sch.h
@@ -65,17 +65,19 @@ typedef struct SRSLTE_API {
   
   /* buffers */
   uint8_t *cb_in; 
-  uint8_t *cb_temp; 
   void *cb_out;  
   void *e;
+  uint8_t *temp_g_bits;
+  uint32_t *ul_interleaver;
+  srslte_uci_bit_t ack_ri_bits[12*288];
+  uint32_t nof_ri_ack_bits; 
   
   srslte_tcod_t encoder;
   srslte_tdec_t decoder;  
   srslte_crc_t crc_tb;
   srslte_crc_t crc_cb;
   
-  srslte_uci_cqi_pusch_t uci_cqi; 
-  srslte_uci_pos_t uci_pos;
+  srslte_uci_cqi_pusch_t uci_cqi;
   
 } srslte_sch_t;
 
diff --git a/srslte/include/srslte/phch/uci.h b/srslte/include/srslte/phch/uci.h
index 78e6b5423..153b3ed4e 100644
--- a/srslte/include/srslte/phch/uci.h
+++ b/srslte/include/srslte/phch/uci.h
@@ -46,9 +46,6 @@
 #define SRSLTE_UCI_MAX_CQI_LEN_PUCCH       13
 #define SRSLTE_UCI_CQI_CODED_PUCCH_B       20
 
-#define SRSLTE_UCI_ACK_RI_PLACEHOLDER_REPETITION 0xC0
-#define SRSLTE_UCI_ACK_RI_PLACEHOLDER            0x30
-
 typedef struct SRSLTE_API {
   srslte_crc_t crc;
   uint8_t tmp_cqi[SRSLTE_UCI_MAX_CQI_LEN_PUSCH];
@@ -67,10 +64,14 @@ typedef struct SRSLTE_API {
   bool channel_selection; 
 } srslte_uci_data_t;
 
+typedef enum {
+  UCI_BIT_1 = 0, UCI_BIT_0, UCI_BIT_REPETITION, UCI_BIT_PLACEHOLDER
+} srslte_uci_bit_type_t;
+
 typedef struct {
-  uint32_t idx;
-  uint32_t pos[SRSLTE_UCI_MAX_CQI_LEN_PUSCH];  
-} srslte_uci_pos_t;
+  uint32_t position;
+  srslte_uci_bit_type_t type;
+} srslte_uci_bit_t;
 
 SRSLTE_API int srslte_uci_cqi_init(srslte_uci_cqi_pusch_t *q);
 
@@ -89,19 +90,18 @@ SRSLTE_API int srslte_uci_encode_cqi_pucch(uint8_t *cqi_data,
                                            uint8_t b_bits[SRSLTE_UCI_CQI_CODED_PUCCH_B]);
 
 SRSLTE_API int srslte_uci_encode_ack(srslte_pusch_cfg_t *cfg,
-                                     srslte_uci_pos_t *pos,
                                      uint8_t data, 
                                      uint32_t O_cqi, 
                                      float beta, 
                                      uint32_t H_prime_total, 
-                                     uint8_t *q_bits); 
+                                     srslte_uci_bit_t *ri_bits); 
 
 SRSLTE_API int srslte_uci_encode_ri(srslte_pusch_cfg_t *cfg,
                                     uint8_t data, 
                                     uint32_t O_cqi, 
                                     float beta, 
                                     uint32_t H_prime_total, 
-                                    uint8_t *q_bits); 
+                                    srslte_uci_bit_t *ri_bits); 
 
 
 #endif
diff --git a/srslte/include/srslte/scrambling/scrambling.h b/srslte/include/srslte/scrambling/scrambling.h
index 59418b4ea..e7129fa7b 100644
--- a/srslte/include/srslte/scrambling/scrambling.h
+++ b/srslte/include/srslte/scrambling/scrambling.h
@@ -49,6 +49,14 @@ SRSLTE_API void srslte_scrambling_b_offset(srslte_sequence_t *s,
                                            int offset, 
                                            int len);
 
+SRSLTE_API void srslte_scrambling_bytes(srslte_sequence_t *s, 
+                                        uint8_t *data); 
+
+SRSLTE_API void srslte_scrambling_bytes_offset(srslte_sequence_t *s, 
+                                               uint8_t *data, 
+                                               int offset, 
+                                               int len);
+
 SRSLTE_API void srslte_scrambling_f(srslte_sequence_t *s, 
                                     float *data);
 
diff --git a/srslte/lib/common/src/sequence.c b/srslte/lib/common/src/sequence.c
index d9bd485f4..109c036c0 100644
--- a/srslte/lib/common/src/sequence.c
+++ b/srslte/lib/common/src/sequence.c
@@ -32,6 +32,7 @@
 
 #include "srslte/common/sequence.h"
 #include "srslte/utils/vector.h"
+#include "srslte/utils/bit.h"
 
 #define Nc 1600
 
@@ -81,18 +82,26 @@ int srslte_sequence_LTE_pr(srslte_sequence_t *q, uint32_t len, uint32_t seed) {
   }
   q->len = len;
   srslte_sequence_set_LTE_pr(q, seed);
+  srslte_bit_pack_vector(q->c, q->c_bytes, len);
   return SRSLTE_SUCCESS;
 }
 
 int srslte_sequence_init(srslte_sequence_t *q, uint32_t len) {
   if (q->c && (q->len != len)) {
     free(q->c);
+    if (q->c_bytes) {
+      free(q->c_bytes);
+    }
   }
   if (!q->c) {
     q->c = srslte_vec_malloc(len * sizeof(uint8_t));
     if (!q->c) {
       return SRSLTE_ERROR;
     }
+    q->c_bytes = srslte_vec_malloc(len * sizeof(uint8_t)/8);
+    if (!q->c_bytes) {
+      return SRSLTE_ERROR;
+    }
     q->len = len;
   }
   return SRSLTE_SUCCESS;
@@ -102,6 +111,9 @@ void srslte_sequence_free(srslte_sequence_t *q) {
   if (q->c) {
     free(q->c);
   }
+  if (q->c_bytes) {
+    free(q->c_bytes);
+  }
   bzero(q, sizeof(srslte_sequence_t));
 }
 
diff --git a/srslte/lib/fec/src/cbsegm.c b/srslte/lib/fec/src/cbsegm.c
index e32a064d4..73fcfa874 100644
--- a/srslte/lib/fec/src/cbsegm.c
+++ b/srslte/lib/fec/src/cbsegm.c
@@ -89,7 +89,7 @@ int srslte_cbsegm(srslte_cbsegm_t *s, uint32_t tbs) {
             s->C1 = s->C - s->C2;
           }
           s->F = s->C1 * s->K1 + s->C2 * s->K2 - Bp;
-          printf("CB Segmentation: TBS: %d, C=%d, C+=%d K+=%d, C-=%d, K-=%d, F=%d, Bp=%d\n",
+          INFO("CB Segmentation: TBS: %d, C=%d, C+=%d K+=%d, C-=%d, K-=%d, F=%d, Bp=%d\n",
               tbs, s->C, s->C1, s->K1, s->C2, s->K2, s->F, Bp);         
           ret = SRSLTE_SUCCESS; 
         }
diff --git a/srslte/lib/phch/src/pusch.c b/srslte/lib/phch/src/pusch.c
index 0b28f23fb..a2066a236 100644
--- a/srslte/lib/phch/src/pusch.c
+++ b/srslte/lib/phch/src/pusch.c
@@ -546,12 +546,18 @@ int srslte_pusch_uci_encode_rnti(srslte_pusch_t *q, srslte_pusch_cfg_t *cfg, srs
     
     // Correct UCI placeholder bits    
     uint8_t *d = q->q; 
-    for (int i = 0; i < q->dl_sch.uci_pos.idx; i++) {     
-      if (d[q->dl_sch.uci_pos.pos[i]] & SRSLTE_UCI_ACK_RI_PLACEHOLDER) {
-        d[q->dl_sch.uci_pos.pos[i]] = 1; 
-      } else if (d[q->dl_sch.uci_pos.pos[i]] & SRSLTE_UCI_ACK_RI_PLACEHOLDER_REPETITION) {
-        if (q->dl_sch.uci_pos.pos[i] > 1) {
-          d[q->dl_sch.uci_pos.pos[i]] = d[q->dl_sch.uci_pos.pos[i]-1];        
+    for (int i = 0; i < q->dl_sch.nof_ri_ack_bits; i++) {     
+      if (q->dl_sch.ack_ri_bits[i].type == UCI_BIT_PLACEHOLDER) {
+        d[q->dl_sch.ack_ri_bits[i].position/8] |= (1<<(q->dl_sch.ack_ri_bits[i].position%8)); 
+      } else if (q->dl_sch.ack_ri_bits[i].type == UCI_BIT_REPETITION) {
+        if (q->dl_sch.ack_ri_bits[i].position > 1) {
+          uint32_t p=q->dl_sch.ack_ri_bits[i].position;
+          uint8_t bit = d[(p-1)/8] & (1<<((p-1)%8)); 
+          if (bit) {
+            d[p/8] |= 1<<(p%8);
+          } else {
+            d[p/8] &= ~(1<<(p%8));
+          }
         } 
       }
     }
diff --git a/srslte/lib/phch/src/sch.c b/srslte/lib/phch/src/sch.c
index 3420a0387..404786a01 100644
--- a/srslte/lib/phch/src/sch.c
+++ b/srslte/lib/phch/src/sch.c
@@ -113,15 +113,20 @@ int srslte_sch_init(srslte_sch_t *q) {
     if (!q->cb_in) {
       goto clean;
     }
-    q->cb_temp = srslte_vec_malloc(sizeof(uint8_t) * SRSLTE_TCOD_MAX_LEN_CB+4);
-    if (!q->cb_temp) {
-      goto clean;
-    }
     
     q->cb_out = srslte_vec_malloc(sizeof(float) * (3 * SRSLTE_TCOD_MAX_LEN_CB + 12));
     if (!q->cb_out) {
       goto clean;
     }  
+    q->temp_g_bits = srslte_vec_malloc(sizeof(uint8_t)*SRSLTE_MAX_PRB*12*12*12);
+    if (!q->temp_g_bits) {
+      goto clean; 
+    }
+    bzero(q->temp_g_bits, SRSLTE_MAX_PRB*12*12*12);
+    q->ul_interleaver = srslte_vec_malloc(sizeof(uint32_t)*SRSLTE_MAX_PRB*12*12*12);
+    if (!q->ul_interleaver) {
+      goto clean; 
+    }
     if (srslte_uci_cqi_init(&q->uci_cqi)) {
       goto clean;
     }
@@ -139,12 +144,15 @@ void srslte_sch_free(srslte_sch_t *q) {
   if (q->cb_in) {
     free(q->cb_in);
   }
-  if (q->cb_temp) {
-    free(q->cb_temp);
-  }
   if (q->cb_out) {
     free(q->cb_out);
   }
+  if (q->temp_g_bits) {
+    free(q->temp_g_bits);
+  }
+  if (q->ul_interleaver) {
+    free(q->ul_interleaver);
+  }
   srslte_tdec_free(&q->decoder);
   srslte_tcod_free(&q->encoder);
   srslte_uci_cqi_free(&q->uci_cqi);
@@ -161,15 +169,13 @@ uint32_t srslte_sch_last_noi(srslte_sch_t *q) {
 }
 
 
-uint8_t temp[64*1024];
-
 /* Encode a transport block according to 36.212 5.3.2
  *
  */
-static int encode_tb(srslte_sch_t *q, 
+static int encode_tb_off(srslte_sch_t *q, 
                      srslte_softbuffer_tx_t *soft_buffer, srslte_cbsegm_t *cb_segm, 
                      uint32_t Qm, uint32_t rv, uint32_t nof_e_bits,  
-                     uint8_t *data, uint8_t *e_bits) 
+                     uint8_t *data, uint8_t *e_bits, uint32_t w_offset) 
 {
   uint8_t parity[3] = {0, 0, 0};
   uint32_t par;
@@ -281,12 +287,11 @@ static int encode_tb(srslte_sch_t *q,
       
       /* Rate matching */
       if (3*cb_len+12 < soft_buffer->buff_size) {
-        if (srslte_rm_turbo_tx_lut(soft_buffer->buffer_b[i], q->cb_in, (uint8_t*) q->cb_out, &temp[wp/8], cblen_idx, n_e, wp%8, rv))
+        if (srslte_rm_turbo_tx_lut(soft_buffer->buffer_b[i], q->cb_in, (uint8_t*) q->cb_out, &e_bits[(wp+w_offset)/8], cblen_idx, n_e, (wp+w_offset)%8, rv))
         {
           fprintf(stderr, "Error in rate matching\n");
           return SRSLTE_ERROR;
         }
-        srslte_vec_fprint_byte(stdout, &temp[wp/8], (n_e-1)/8+1);
       } else {
         fprintf(stderr, "Encoded CB length exceeds RM buffer (%d>%d)\n",3*cb_len+12,soft_buffer->buff_size);
         return SRSLTE_ERROR; 
@@ -295,8 +300,6 @@ static int encode_tb(srslte_sch_t *q,
       rp += rlen;
       wp += n_e;
     }
-    srslte_bit_unpack_vector(temp, e_bits, nof_e_bits);
-    srslte_vec_fprint_b(stdout, e_bits, nof_e_bits);
     
     INFO("END CB#%d: wp: %d, rp: %d\n", i, wp, rp);
     ret = SRSLTE_SUCCESS;      
@@ -304,6 +307,16 @@ static int encode_tb(srslte_sch_t *q,
   return ret; 
 }
 
+
+static int encode_tb(srslte_sch_t *q, 
+                     srslte_softbuffer_tx_t *soft_buffer, srslte_cbsegm_t *cb_segm, 
+                     uint32_t Qm, uint32_t rv, uint32_t nof_e_bits,  
+                     uint8_t *data, uint8_t *e_bits) 
+{
+  return encode_tb_off(q, soft_buffer, cb_segm, Qm, rv, nof_e_bits, data, e_bits, 0);
+}
+
+
 /* Decode a transport block according to 36.212 5.3.2
  *
  */
@@ -498,34 +511,58 @@ int srslte_ulsch_decode(srslte_sch_t *q, srslte_pusch_cfg_t *cfg, srslte_softbuf
                    e_bits, data);
 }
 
-
 /* UL-SCH channel interleaver according to 5.2.2.8 of 36.212 */
-void ulsch_interleave(srslte_uci_pos_t *q, uint8_t *g_bits, uint32_t Qm, uint32_t H_prime_total, 
-                      uint32_t N_pusch_symbs, uint8_t *q_bits) 
+void ulsch_interleave(uint8_t *g_bits, uint32_t Qm, uint32_t H_prime_total, 
+                      uint32_t N_pusch_symbs, uint8_t *q_bits, srslte_uci_bit_t *ri_bits, uint32_t nof_ri_bits, 
+                      uint32_t *interleaver_buffer, uint8_t *temp_buffer, uint32_t buffer_sz) 
 {
   
   uint32_t rows = H_prime_total/N_pusch_symbs;
   uint32_t cols = N_pusch_symbs;
   
-  uint32_t idx = 0;
-  for(uint32_t j=0; j<rows; j++) {        
-    for(uint32_t i=0; i<cols; i++) {
-      for(uint32_t k=0; k<Qm; k++) {
-        if (q_bits[j*Qm + i*rows*Qm + k] >= 10) {
-          q_bits[j*Qm + i*rows*Qm + k] -= 10;
-          q->pos[q->idx%SRSLTE_UCI_MAX_CQI_LEN_PUSCH] = j*Qm + i*rows*Qm + k;
-          q->idx++;
-          if (q->idx >= SRSLTE_UCI_MAX_CQI_LEN_PUSCH) {
-            fprintf(stderr, "Error number of UCI bits exceeds SRSLTE_UCI_MAX_CQI_LEN_PUSCH\n");
-          }
-        } else {
-          q_bits[j*Qm + i*rows*Qm + k] = g_bits[idx];                                
-          idx++;                  
-        }
+  // Prepare ri_bits for fast search using temp_buffer
+  if (nof_ri_bits > 0) {
+    for (uint32_t i=0;i<nof_ri_bits;i++) {
+      if (ri_bits[i].position < buffer_sz) {
+        temp_buffer[ri_bits[i].position] = 1; 
+      } else {
+        fprintf(stderr, "Error computing ULSCH interleaver RI bits. Position %d exceeds buffer size %d\n", ri_bits[i].position, buffer_sz);
       }
     }
   }
   
+  /* Compute the interleaving function on-the-fly, because it depends on number of RI bits 
+   * Profiling show that the computation of this matrix is neglegible. 
+   */
+  uint32_t idx = 0;
+  for(uint32_t j=0; j<rows; j++) {        
+    for(uint32_t i=0; i<cols; i++) {
+      for(uint32_t k=0; k<Qm; k++) {
+        if (temp_buffer[j*Qm + i*rows*Qm + k]) {
+          interleaver_buffer[j*Qm + i*rows*Qm + k] = 0; 
+        } else {
+          if (j*Qm + i*rows*Qm + k < buffer_sz) {
+            interleaver_buffer[j*Qm + i*rows*Qm + k] = idx;
+            idx++;                  
+          } else {
+            fprintf(stderr, "Error computing ULSCH interleaver. Position %d exceeds buffer size %d\n", j*Qm + i*rows*Qm + k, buffer_sz);
+          }
+        }
+      }
+    }
+  }
+  srslte_bit_interleave(g_bits, q_bits, interleaver_buffer, H_prime_total*Qm);    
+  
+  // Reset temp_buffer because will be reused next time
+  if (nof_ri_bits > 0) {
+    for (uint32_t i=0;i<nof_ri_bits;i++) {
+      if (ri_bits[i].position < buffer_sz) {
+        temp_buffer[ri_bits[i].position] = 0; 
+      } else {
+        fprintf(stderr, "Error computing ULSCH interleaver RI bits. Position %d exceeds buffer size %d\n", ri_bits[i].position, buffer_sz);
+      }
+    }
+  }
 }
 
 int srslte_ulsch_encode(srslte_sch_t *q, srslte_pusch_cfg_t *cfg, srslte_softbuffer_tx_t *softbuffer,
@@ -551,15 +588,13 @@ int srslte_ulsch_uci_encode(srslte_sch_t *q,
   uint32_t nb_q = cfg->nbits.nof_bits; 
   uint32_t Qm = cfg->grant.Qm; 
   
-  bzero(q_bits, sizeof(uint8_t) * nb_q);
-
   // Encode RI
   if (uci_data.uci_ri_len > 0) {
     float beta = beta_ri_offset[cfg->uci_cfg.I_offset_ri]; 
     if (cfg->cb_segm.tbs == 0) {
         beta /= beta_cqi_offset[cfg->uci_cfg.I_offset_cqi];
     }
-    ret = srslte_uci_encode_ri(cfg, uci_data.uci_ri, uci_data.uci_cqi_len, beta, nb_q/Qm, q_bits);
+    ret = srslte_uci_encode_ri(cfg, uci_data.uci_ri, uci_data.uci_cqi_len, beta, nb_q/Qm, q->ack_ri_bits);
     if (ret < 0) {
       return ret; 
     }
@@ -571,43 +606,65 @@ int srslte_ulsch_uci_encode(srslte_sch_t *q,
     ret = srslte_uci_encode_cqi_pusch(&q->uci_cqi, cfg, 
                                       uci_data.uci_cqi, uci_data.uci_cqi_len, 
                                       beta_cqi_offset[cfg->uci_cfg.I_offset_cqi], 
-                                      Q_prime_ri, g_bits);
+                                      Q_prime_ri, q->temp_g_bits);
     if (ret < 0) {
       return ret; 
     }
     Q_prime_cqi = (uint32_t) ret; 
+    srslte_bit_pack_vector(q->temp_g_bits, g_bits, Q_prime_cqi*Qm);
+    // Reset the buffer because will be reused in ulsch_interleave
+    bzero(q->temp_g_bits, Q_prime_cqi*Qm);
   }
   
   e_offset += Q_prime_cqi*Qm;
- 
+
   // Encode UL-SCH
   if (cfg->cb_segm.tbs > 0) {
     uint32_t G = nb_q/Qm - Q_prime_ri - Q_prime_cqi;     
-    ret = encode_tb(q, softbuffer, &cfg->cb_segm, 
+    ret = encode_tb_off(q, softbuffer, &cfg->cb_segm, 
                     Qm, cfg->rv, G*Qm, 
-                    data, &g_bits[e_offset]);
+                    data, &g_bits[e_offset/8], e_offset%8);
     if (ret) {
       return ret; 
     }    
   } 
-   
+  
+  //srslte_bit_unpack_vector(g_bits, kk, nb_q);
+  //srslte_vec_fprint_b(stdout, kk, nb_q);
+      
   // Interleave UL-SCH (and RI and CQI)
-  q->uci_pos.idx=0;
-  ulsch_interleave(&q->uci_pos, g_bits, Qm, nb_q/Qm, cfg->nbits.nof_symb, q_bits);
-
+  ulsch_interleave(g_bits, Qm, nb_q/Qm, cfg->nbits.nof_symb, q_bits, q->ack_ri_bits, Q_prime_ri*Qm, 
+                   q->ul_interleaver, q->temp_g_bits, SRSLTE_MAX_PRB*12*12*12);
+  
    // Encode (and interleave) ACK
   if (uci_data.uci_ack_len > 0) {
     float beta = beta_harq_offset[cfg->uci_cfg.I_offset_ack]; 
     if (cfg->cb_segm.tbs == 0) {
         beta /= beta_cqi_offset[cfg->uci_cfg.I_offset_cqi];
     }
-    ret = srslte_uci_encode_ack(cfg, &q->uci_pos, uci_data.uci_ack, uci_data.uci_cqi_len, beta, nb_q/Qm, q_bits);
+    ret = srslte_uci_encode_ack(cfg, uci_data.uci_ack, uci_data.uci_cqi_len, beta, nb_q/Qm, &q->ack_ri_bits[Q_prime_ri*Qm]);
     if (ret < 0) {
       return ret; 
     }
     Q_prime_ack = (uint32_t) ret; 
   }
+  
+  q->nof_ri_ack_bits = (Q_prime_ack+Q_prime_ri)*Qm;
+  
+  for (uint32_t i=0;i<q->nof_ri_ack_bits;i++) {
+    uint32_t p = q->ack_ri_bits[i].position;
+    if (p < nb_q) {
+      if (q->ack_ri_bits[i].type == UCI_BIT_1) {
+        q_bits[p/8] |= (1<<(7-p%8));
+      } else {
+        q_bits[p/8] &= ~(1<<(7-p%8));
+      }
+    } else {
+      fprintf(stderr, "Invalid RI/ACK bit position %d. Max bits=%d\n", p, nb_q);
+    }    
+  }
     
+  
   INFO("Q_prime_ack=%d, Q_prime_cqi=%d, Q_prime_ri=%d\n",Q_prime_ack, Q_prime_cqi, Q_prime_ri);  
 
   return SRSLTE_SUCCESS;
diff --git a/srslte/lib/phch/src/uci.c b/srslte/lib/phch/src/uci.c
index 456c6c4b2..79a6027ef 100644
--- a/srslte/lib/phch/src/uci.c
+++ b/srslte/lib/phch/src/uci.c
@@ -241,10 +241,10 @@ int srslte_uci_encode_cqi_pusch(srslte_uci_cqi_pusch_t *q, srslte_pusch_cfg_t *c
   }
 }
 
-/* Inserts UCI-ACK bits into the correct positions in the g buffer before interleaving */
-static int uci_ulsch_interleave_ack(srslte_uci_pos_t *q, uint8_t ack_coded_bits[6], uint32_t ack_q_bit_idx, 
+/* Generates UCI-ACK bits and computes position in q bits */
+static int uci_ulsch_interleave_ack(srslte_uci_bit_type_t ack_coded_bits[6], uint32_t ack_q_bit_idx, 
                           uint32_t Qm, uint32_t H_prime_total, uint32_t N_pusch_symbs, srslte_cp_t cp,
-                          uint8_t *q_bits) {
+                          srslte_uci_bit_t *ack_bits) {
 
   const uint32_t ack_column_set_norm[4] = {2, 3, 8, 9};
   const uint32_t ack_column_set_ext[4] = {1, 2, 6, 7};
@@ -255,14 +255,8 @@ static int uci_ulsch_interleave_ack(srslte_uci_pos_t *q, uint8_t ack_coded_bits[
     uint32_t colidx = (3*ack_q_bit_idx)%4;
     uint32_t col = SRSLTE_CP_ISNORM(cp)?ack_column_set_norm[colidx]:ack_column_set_ext[colidx];
     for(uint32_t k=0; k<Qm; k++) {
-      q_bits[row *Qm + (H_prime_total/N_pusch_symbs)*col*Qm + k] = ack_coded_bits[k];
-      
-      // Save position for placeholder bits
-      q->pos[q->idx%SRSLTE_UCI_MAX_CQI_LEN_PUSCH] = row *Qm + (H_prime_total/N_pusch_symbs)*col*Qm + k;
-      q->idx++;
-      if (q->idx >= SRSLTE_UCI_MAX_CQI_LEN_PUSCH) {
-        fprintf(stderr, "Error number of UCI bits exceeds SRSLTE_UCI_MAX_CQI_LEN_PUSCH\n");
-      }
+      ack_bits[k].position = row *Qm + (H_prime_total/N_pusch_symbs)*col*Qm + k;
+      ack_bits[k].type = ack_coded_bits[k];
     }    
     return SRSLTE_SUCCESS;
   } else {
@@ -273,9 +267,9 @@ static int uci_ulsch_interleave_ack(srslte_uci_pos_t *q, uint8_t ack_coded_bits[
 }
 
 /* Inserts UCI-RI bits into the correct positions in the g buffer before interleaving */
-static int uci_ulsch_interleave_ri(uint8_t ri_coded_bits[6], uint32_t ri_q_bit_idx, 
+static int uci_ulsch_interleave_ri(srslte_uci_bit_type_t ri_coded_bits[6], uint32_t ri_q_bit_idx, 
                           uint32_t Qm, uint32_t H_prime_total, uint32_t N_pusch_symbs, srslte_cp_t cp,
-                          uint8_t *q_bits) {
+                          srslte_uci_bit_t *ri_bits) {
   
   static uint32_t ri_column_set_norm[4]  = {1, 4, 7, 10};
   static uint32_t ri_column_set_ext[4]  = {0, 3, 5, 8};
@@ -286,7 +280,8 @@ static int uci_ulsch_interleave_ri(uint8_t ri_coded_bits[6], uint32_t ri_q_bit_i
     uint32_t col = SRSLTE_CP_ISNORM(cp)?ri_column_set_norm[colidx]:ri_column_set_ext[colidx];
 
     for(uint32_t k=0; k<Qm; k++) {
-      q_bits[row *Qm + (H_prime_total/N_pusch_symbs)*col*Qm + k] = 10+ri_coded_bits[k];
+      ri_bits[k].position = row *Qm + (H_prime_total/N_pusch_symbs)*col*Qm + k;
+      ri_bits[k].type = ri_coded_bits[k];
     }    
     return SRSLTE_SUCCESS;
   } else {
@@ -323,21 +318,20 @@ static uint32_t Q_prime_ri_ack(srslte_pusch_cfg_t *cfg,
   return Q_prime; 
 }
 
-static void encode_ri_ack(uint8_t data, uint8_t q_encoded_bits[6], uint8_t Qm) {
-  q_encoded_bits[0] = data;
-  q_encoded_bits[1] = SRSLTE_UCI_ACK_RI_PLACEHOLDER_REPETITION; 
+static void encode_ri_ack(uint8_t data, srslte_uci_bit_type_t q_encoded_bits[6], uint8_t Qm) {
+  q_encoded_bits[0] = data?UCI_BIT_1:UCI_BIT_0;
+  q_encoded_bits[1] = UCI_BIT_REPETITION; 
   for (uint32_t i=2;i<Qm;i++) {
-    q_encoded_bits[i] = SRSLTE_UCI_ACK_RI_PLACEHOLDER;
+    q_encoded_bits[i] = UCI_BIT_PLACEHOLDER;
   }
 }
 
 /* Encode UCI HARQ/ACK bits as described in 5.2.2.6 of 36.212 
  *  Currently only supporting 1-bit HARQ
  */
-int srslte_uci_encode_ack(srslte_pusch_cfg_t *cfg, srslte_uci_pos_t *pos,
-                          uint8_t data, 
+int srslte_uci_encode_ack(srslte_pusch_cfg_t *cfg, uint8_t data, 
                           uint32_t O_cqi, float beta, uint32_t H_prime_total, 
-                          uint8_t *q_bits)
+                          srslte_uci_bit_t *ack_bits)
 {
   if (beta < 0) {
     fprintf(stderr, "Error beta is reserved\n");
@@ -345,12 +339,12 @@ int srslte_uci_encode_ack(srslte_pusch_cfg_t *cfg, srslte_uci_pos_t *pos,
   }
 
   uint32_t Qprime = Q_prime_ri_ack(cfg, 1, O_cqi, beta);
-  uint8_t q_encoded_bits[6];
+  srslte_uci_bit_type_t q_encoded_bits[6];
 
   encode_ri_ack(data, q_encoded_bits, cfg->grant.Qm);
   
   for (uint32_t i=0;i<Qprime;i++) {
-    uci_ulsch_interleave_ack(pos, q_encoded_bits, i, cfg->grant.Qm, H_prime_total, cfg->nbits.nof_symb, cfg->cp, q_bits);
+    uci_ulsch_interleave_ack(q_encoded_bits, i, cfg->grant.Qm, H_prime_total, cfg->nbits.nof_symb, cfg->cp, &ack_bits[cfg->grant.Qm*i]);
   }
   
   return (int) Qprime;
@@ -363,19 +357,19 @@ int srslte_uci_encode_ack(srslte_pusch_cfg_t *cfg, srslte_uci_pos_t *pos,
 int srslte_uci_encode_ri(srslte_pusch_cfg_t *cfg,
                          uint8_t data, 
                          uint32_t O_cqi, float beta, uint32_t H_prime_total, 
-                         uint8_t *q_bits)
+                         srslte_uci_bit_t *ri_bits)
 {
   if (beta < 0) {
     fprintf(stderr, "Error beta is reserved\n");
     return -1; 
   }
   uint32_t Qprime = Q_prime_ri_ack(cfg, 1, O_cqi, beta);
-  uint8_t q_encoded_bits[6];
+  srslte_uci_bit_type_t q_encoded_bits[6];
 
   encode_ri_ack(data, q_encoded_bits, cfg->grant.Qm);
   
   for (uint32_t i=0;i<Qprime;i++) {
-    uci_ulsch_interleave_ri(q_encoded_bits, i, cfg->grant.Qm, H_prime_total, cfg->nbits.nof_symb, cfg->cp, q_bits);
+    uci_ulsch_interleave_ri(q_encoded_bits, i, cfg->grant.Qm, H_prime_total, cfg->nbits.nof_symb, cfg->cp, &ri_bits[cfg->grant.Qm*i]);
   }
   
   return (int) Qprime;
diff --git a/srslte/lib/phch/test/pusch_test.c b/srslte/lib/phch/test/pusch_test.c
index 364f26adc..a08ef8f9e 100644
--- a/srslte/lib/phch/test/pusch_test.c
+++ b/srslte/lib/phch/test/pusch_test.c
@@ -167,7 +167,7 @@ int main(int argc, char **argv) {
   
   srslte_uci_data_t uci_data; 
   bzero(&uci_data, sizeof(srslte_uci_data_t));
-  uci_data.uci_cqi_len = 0; 
+  uci_data.uci_cqi_len = 20; 
   uci_data.uci_ri_len = 0; 
   uci_data.uci_ack_len = 0; 
 
@@ -202,7 +202,7 @@ int main(int argc, char **argv) {
     data[i] = 1;
   }
 
-  if (srslte_softbuffer_tx_init(&softbuffer, 4*cell.nof_prb)) {
+  if (srslte_softbuffer_tx_init(&softbuffer, 100)) {
     fprintf(stderr, "Error initiating soft buffer\n");
     goto quit;
   }
diff --git a/srslte/lib/phch/test/ulsch_encode_test_mex.c b/srslte/lib/phch/test/ulsch_encode_test_mex.c
index dd793721d..d2be9c19b 100644
--- a/srslte/lib/phch/test/ulsch_encode_test_mex.c
+++ b/srslte/lib/phch/test/ulsch_encode_test_mex.c
@@ -69,7 +69,7 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
   cell.id=1;
   cell.cp=SRSLTE_CP_NORM;
 
-  srslte_verbose = SRSLTE_VERBOSE_INFO;
+  srslte_verbose = SRSLTE_VERBOSE_NONE;
   
   if (srslte_softbuffer_tx_init(&softbuffer, cell.nof_prb)) {
     mexErrMsgTxt("Error initiating HARQ\n");
@@ -175,11 +175,15 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
     return;
   }
     
-  uint8_t *q_bits = srslte_vec_malloc(cfg.nbits.nof_bits * sizeof(uint8_t));
+  uint8_t *q_bits = srslte_vec_malloc(cfg.nbits.nof_bits * sizeof(uint8_t)/8);
   if (!q_bits) {
     return;
   }
-  uint8_t *g_bits = srslte_vec_malloc(cfg.nbits.nof_bits * sizeof(uint8_t));
+  uint8_t *q_bits_unpacked = srslte_vec_malloc(cfg.nbits.nof_bits * sizeof(uint8_t));
+  if (!q_bits_unpacked) {
+    return;
+  }
+  uint8_t *g_bits = srslte_vec_malloc(cfg.nbits.nof_bits * sizeof(uint8_t)/8);
   if (!g_bits) {
     return;
   }
@@ -196,9 +200,11 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
       return;
     }    
   }
-  
+
+  srslte_bit_unpack_vector(q_bits, q_bits_unpacked, cfg.nbits.nof_bits);
+
   if (nlhs >= 1) {
-    mexutils_write_uint8(q_bits, &plhs[0], cfg.nbits.nof_bits, 1);  
+    mexutils_write_uint8(q_bits_unpacked, &plhs[0], cfg.nbits.nof_bits, 1);  
   }
   
   srslte_sch_free(&ulsch);  
@@ -206,6 +212,7 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
   
   free(trblkin);
   free(g_bits);    
+  free(q_bits_unpacked);
   free(q_bits);    
   
   
diff --git a/srslte/lib/scrambling/src/scrambling.c b/srslte/lib/scrambling/src/scrambling.c
index fe4136ba7..3a2dded0d 100644
--- a/srslte/lib/scrambling/src/scrambling.c
+++ b/srslte/lib/scrambling/src/scrambling.c
@@ -58,37 +58,47 @@ void srslte_scrambling_c_offset(srslte_sequence_t *s, cf_t *data, int offset, in
   }
 }
 
-void srslte_scrambling_b(srslte_sequence_t *s, uint8_t *data) {
-  srslte_scrambling_b_offset(s, data, 0, s->len);  
-}
-
-void srslte_scrambling_b_offset(srslte_sequence_t *s, uint8_t *data, int offset, int len) {
+void scrambling_b(uint8_t *c, uint8_t *data, int offset, int len) {
   int i;
-  assert (len + offset <= s->len);
   // Do XOR on a word basis
   
   if (!(len%8)) {
     uint64_t *x = (uint64_t*) data; 
-    uint64_t *y = (uint64_t*) &s->c[offset];
+    uint64_t *y = (uint64_t*) &c[offset];
     for (int i=0;i<len/8;i++) {
       x[i] = (x[i] ^ y[i]);
     }
   } else if (!(len%4)) {
     uint32_t *x = (uint32_t*) data; 
-    uint32_t *y = (uint32_t*) &s->c[offset];
+    uint32_t *y = (uint32_t*) &c[offset];
     for (int i=0;i<len/4;i++) {
       x[i] = (x[i] ^ y[i]);
     }
   } else if (!(len%2)) {
     uint16_t *x = (uint16_t*) data; 
-    uint16_t *y = (uint16_t*) &s->c[offset];
+    uint16_t *y = (uint16_t*) &c[offset];
     for (int i=0;i<len/2;i++) {
       x[i] = (x[i] ^ y[i]);
     }
   } else {    
     for (i = 0; i < len; i++) {
-      data[i] = (data[i] + s->c[i + offset]) % 2;
+      data[i] = (data[i] ^ c[i + offset]);
     }
   }
 }
 
+void srslte_scrambling_b(srslte_sequence_t *s, uint8_t *data) {
+  scrambling_b(s->c, data, 0, s->len);  
+}
+
+void srslte_scrambling_b_offset(srslte_sequence_t *s, uint8_t *data, int offset, int len) {
+  scrambling_b(s->c, data, offset, len);  
+}
+
+void srslte_scrambling_bytes(srslte_sequence_t *s, uint8_t *data) {
+  scrambling_b(s->c_bytes, data, 0, s->len);  
+}
+
+void srslte_scrambling_bytes_offset(srslte_sequence_t *s, uint8_t *data, int offset, int len) {
+  scrambling_b(s->c_bytes, data, offset, len);  
+}
diff --git a/srslte/lib/utils/src/bit.c b/srslte/lib/utils/src/bit.c
index fe2f4be67..5dfe9bf44 100644
--- a/srslte/lib/utils/src/bit.c
+++ b/srslte/lib/utils/src/bit.c
@@ -212,23 +212,14 @@ bitarray_copy(const unsigned char *src_org, int src_offset, int src_len,
 
 void srslte_bit_copy(uint8_t *dst, uint32_t dst_offset, uint8_t *src, uint32_t src_offset, uint32_t nof_bits)
 {
-  static const uint8_t mask_src[] =
-        { 0x00, 0x1, 0x3, 0x7, 0xf, 0x1f, 0x3f, 0x7f, 0xff };
   static const uint8_t mask_dst[] =
         { 0x00, 0x80, 0xc0, 0xe0, 0xf0, 0xf8, 0xfc, 0xfe, 0xff };
-  if ((dst_offset%8) == (src_offset%8)) {
-    if (src_offset%8) {
-      // copy 1st word
-      dst[dst_offset/8] &= 0xf0;
-      dst[dst_offset/8] |= (src[src_offset/8] & mask_src[src_offset%8]);
-      dst_offset+=(src_offset%8);
-      src_offset+=(src_offset%8);
-    }
+  if ((dst_offset%8) == 0 && (src_offset%8) == 0) {
     // copy rest of words
     memcpy(&dst[dst_offset/8], &src[src_offset/8], nof_bits/8);
     // copy last word
-    if ((src_offset%8+nof_bits)%8) {
-      dst[dst_offset/8+nof_bits/8] = src[src_offset/8+nof_bits/8] & mask_dst[(src_offset%8+nof_bits)%8];
+    if (nof_bits%8) {
+      dst[dst_offset/8+nof_bits/8] = src[src_offset/8+nof_bits/8] & mask_dst[nof_bits%8];
     }
   } else {
     bitarray_copy(src, src_offset, nof_bits, dst, dst_offset);