diff --git a/lte/phy/lib/fec/src/turbodecoder.c b/lte/phy/lib/fec/src/turbodecoder.c
index 87ff32b50..89c784966 100644
--- a/lte/phy/lib/fec/src/turbodecoder.c
+++ b/lte/phy/lib/fec/src/turbodecoder.c
@@ -33,6 +33,7 @@
 #include <math.h>
 
 #include "liblte/phy/fec/turbodecoder.h"
+#include "liblte/phy/utils/vector.h"
 
 /************************************************
  *
@@ -157,9 +158,9 @@ void map_gen_alpha(map_gen_t * s, llr_t * input, llr_t * parity, llr_t * output,
 int map_gen_init(map_gen_t * h, int max_long_cb)
 {
   bzero(h, sizeof(map_gen_t));
-  h->beta = malloc(sizeof(llr_t) * (max_long_cb + TOTALTAIL + 1) * NUMSTATES);
+  h->beta = vec_malloc(sizeof(llr_t) * (max_long_cb + TOTALTAIL + 1) * NUMSTATES);
   if (!h->beta) {
-    perror("malloc");
+    perror("vec_malloc");
     return -1;
   }
   h->max_long_cb = max_long_cb;
@@ -200,29 +201,29 @@ int tdec_init(tdec_t * h, uint32_t max_long_cb)
 
   h->max_long_cb = max_long_cb;
 
-  h->llr1 = malloc(sizeof(llr_t) * len);
+  h->llr1 = vec_malloc(sizeof(llr_t) * len);
   if (!h->llr1) {
-    perror("malloc");
+    perror("vec_malloc");
     goto clean_and_exit;
   }
-  h->llr2 = malloc(sizeof(llr_t) * len);
+  h->llr2 = vec_malloc(sizeof(llr_t) * len);
   if (!h->llr2) {
-    perror("malloc");
+    perror("vec_malloc");
     goto clean_and_exit;
   }
-  h->w = malloc(sizeof(llr_t) * len);
+  h->w = vec_malloc(sizeof(llr_t) * len);
   if (!h->w) {
-    perror("malloc");
+    perror("vec_malloc");
     goto clean_and_exit;
   }
-  h->syst = malloc(sizeof(llr_t) * len);
+  h->syst = vec_malloc(sizeof(llr_t) * len);
   if (!h->syst) {
-    perror("malloc");
+    perror("vec_malloc");
     goto clean_and_exit;
   }
-  h->parity = malloc(sizeof(llr_t) * len);
+  h->parity = vec_malloc(sizeof(llr_t) * len);
   if (!h->parity) {
-    perror("malloc");
+    perror("vec_malloc");
     goto clean_and_exit;
   }
 
diff --git a/lte/phy/lib/phch/src/pbch.c b/lte/phy/lib/phch/src/pbch.c
index 4e0965e61..4dd07bae9 100644
--- a/lte/phy/lib/phch/src/pbch.c
+++ b/lte/phy/lib/phch/src/pbch.c
@@ -166,34 +166,34 @@ int pbch_init(pbch_t *q, lte_cell_t cell) {
     q->encoder.tail_biting = true;
     memcpy(q->encoder.poly, poly, 3 * sizeof(int));
 
-    q->pbch_d = malloc(sizeof(cf_t) * q->nof_symbols);
+    q->pbch_d = vec_malloc(sizeof(cf_t) * q->nof_symbols);
     if (!q->pbch_d) {
       goto clean;
     }
     int i;
     for (i = 0; i < q->cell.nof_ports; i++) {
-      q->ce[i] = malloc(sizeof(cf_t) * q->nof_symbols);
+      q->ce[i] = vec_malloc(sizeof(cf_t) * q->nof_symbols);
       if (!q->ce[i]) {
         goto clean;
       }
-      q->pbch_x[i] = malloc(sizeof(cf_t) * q->nof_symbols);
+      q->pbch_x[i] = vec_malloc(sizeof(cf_t) * q->nof_symbols);
       if (!q->pbch_x[i]) {
         goto clean;
       }
-      q->pbch_symbols[i] = malloc(sizeof(cf_t) * q->nof_symbols);
+      q->pbch_symbols[i] = vec_malloc(sizeof(cf_t) * q->nof_symbols);
       if (!q->pbch_symbols[i]) {
         goto clean;
       }
     }
-    q->pbch_llr = malloc(sizeof(float) * q->nof_symbols * 4 * 2);
+    q->pbch_llr = vec_malloc(sizeof(float) * q->nof_symbols * 4 * 2);
     if (!q->pbch_llr) {
       goto clean;
     }
-    q->temp = malloc(sizeof(float) * q->nof_symbols * 4 * 2);
+    q->temp = vec_malloc(sizeof(float) * q->nof_symbols * 4 * 2);
     if (!q->temp) {
       goto clean;
     }
-    q->pbch_rm_b = malloc(sizeof(float) * q->nof_symbols * 4 * 2);
+    q->pbch_rm_b = vec_malloc(sizeof(float) * q->nof_symbols * 4 * 2);
     if (!q->pbch_rm_b) {
       goto clean;
     }
diff --git a/lte/phy/lib/phch/src/pdsch.c b/lte/phy/lib/phch/src/pdsch.c
index d2f3f10e8..8f7184819 100644
--- a/lte/phy/lib/phch/src/pdsch.c
+++ b/lte/phy/lib/phch/src/pdsch.c
@@ -232,26 +232,26 @@ int pdsch_init(pdsch_t *q, lte_cell_t cell) {
     q->rnti_is_set = false; 
 
     // Allocate floats for reception (LLRs)
-    q->pdsch_e = malloc(sizeof(float) * q->max_symbols * lte_mod_bits_x_symbol(LTE_QAM64));
+    q->pdsch_e = vec_malloc(sizeof(float) * q->max_symbols * lte_mod_bits_x_symbol(LTE_QAM64));
     if (!q->pdsch_e) {
       goto clean;
     }
     
-    q->pdsch_d = malloc(sizeof(cf_t) * q->max_symbols);
+    q->pdsch_d = vec_malloc(sizeof(cf_t) * q->max_symbols);
     if (!q->pdsch_d) {
       goto clean;
     }
 
     for (i = 0; i < q->cell.nof_ports; i++) {
-      q->ce[i] = malloc(sizeof(cf_t) * q->max_symbols);
+      q->ce[i] = vec_malloc(sizeof(cf_t) * q->max_symbols);
       if (!q->ce[i]) {
         goto clean;
       }
-      q->pdsch_x[i] = malloc(sizeof(cf_t) * q->max_symbols);
+      q->pdsch_x[i] = vec_malloc(sizeof(cf_t) * q->max_symbols);
       if (!q->pdsch_x[i]) {
         goto clean;
       }
-      q->pdsch_symbols[i] = malloc(sizeof(cf_t) * q->max_symbols);
+      q->pdsch_symbols[i] = vec_malloc(sizeof(cf_t) * q->max_symbols);
       if (!q->pdsch_symbols[i]) {
         goto clean;
       }
diff --git a/lte/phy/lib/phch/src/pusch.c b/lte/phy/lib/phch/src/pusch.c
index 2d94f3949..a14d2bf1d 100644
--- a/lte/phy/lib/phch/src/pusch.c
+++ b/lte/phy/lib/phch/src/pusch.c
@@ -121,42 +121,42 @@ int pusch_init(pusch_t *q, lte_cell_t cell) {
     q->rnti_is_set = false; 
 
     // Allocate floats for reception (LLRs). Buffer casted to uint8_t for transmission
-    q->pusch_q = malloc(sizeof(float) * q->max_symbols * lte_mod_bits_x_symbol(LTE_QAM64));
+    q->pusch_q = vec_malloc(sizeof(float) * q->max_symbols * lte_mod_bits_x_symbol(LTE_QAM64));
     if (!q->pusch_q) {
       goto clean;
     }
 
     // Allocate floats for reception (LLRs). Buffer casted to uint8_t for transmission
-    q->pusch_g = malloc(sizeof(float) * q->max_symbols * lte_mod_bits_x_symbol(LTE_QAM64));
+    q->pusch_g = vec_malloc(sizeof(float) * q->max_symbols * lte_mod_bits_x_symbol(LTE_QAM64));
     if (!q->pusch_g) {
       goto clean;
     }
 
     // Allocate buffers for q bits for coded RI and ACK bits 
-    q->pusch_g_ack = malloc(sizeof(uint8_t) * 4 * q->cell.nof_prb * lte_mod_bits_x_symbol(LTE_QAM64));
+    q->pusch_g_ack = vec_malloc(sizeof(uint8_t) * 4 * q->cell.nof_prb * lte_mod_bits_x_symbol(LTE_QAM64));
     if (!q->pusch_g_ack) {
       goto clean;
     }
-    q->pusch_g_ri = malloc(sizeof(uint8_t) * 4 * q->cell.nof_prb * lte_mod_bits_x_symbol(LTE_QAM64));
+    q->pusch_g_ri = vec_malloc(sizeof(uint8_t) * 4 * q->cell.nof_prb * lte_mod_bits_x_symbol(LTE_QAM64));
     if (!q->pusch_g_ri) {
       goto clean;
     }
 
-    q->pusch_d = malloc(sizeof(cf_t) * q->max_symbols);
+    q->pusch_d = vec_malloc(sizeof(cf_t) * q->max_symbols);
     if (!q->pusch_d) {
       goto clean;
     }
 
     for (i = 0; i < q->cell.nof_ports; i++) {
-      q->ce[i] = malloc(sizeof(cf_t) * q->max_symbols);
+      q->ce[i] = vec_malloc(sizeof(cf_t) * q->max_symbols);
       if (!q->ce[i]) {
         goto clean;
       }
-      q->pusch_x[i] = malloc(sizeof(cf_t) * q->max_symbols);
+      q->pusch_x[i] = vec_malloc(sizeof(cf_t) * q->max_symbols);
       if (!q->pusch_x[i]) {
         goto clean;
       }
-      q->pusch_symbols[i] = malloc(sizeof(cf_t) * q->max_symbols);
+      q->pusch_symbols[i] = vec_malloc(sizeof(cf_t) * q->max_symbols);
       if (!q->pusch_symbols[i]) {
         goto clean;
       }
@@ -180,6 +180,9 @@ void pusch_free(pusch_t *q) {
   if (q->pusch_d) {
     free(q->pusch_d);
   }
+  if (q->pusch_g) {
+    free(q->pusch_g);
+  }
   if (q->pusch_g_ack) {
     free(q->pusch_g_ack);
   }
diff --git a/lte/phy/lib/phch/src/sch.c b/lte/phy/lib/phch/src/sch.c
index 1788655c4..e46aa11ef 100644
--- a/lte/phy/lib/phch/src/sch.c
+++ b/lte/phy/lib/phch/src/sch.c
@@ -67,12 +67,12 @@ int sch_init(sch_t *q) {
     }
 
     // Allocate floats for reception (LLRs)
-    q->cb_in = malloc(sizeof(uint8_t) * MAX_LONG_CB);
+    q->cb_in = vec_malloc(sizeof(uint8_t) * MAX_LONG_CB);
     if (!q->cb_in) {
       goto clean;
     }
     
-    q->cb_out = malloc(sizeof(float) * (3 * MAX_LONG_CB + 12));
+    q->cb_out = vec_malloc(sizeof(float) * (3 * MAX_LONG_CB + 12));
     if (!q->cb_out) {
       goto clean;
     }  
@@ -407,7 +407,7 @@ uint8_t ulsch_y_idx[10000];
 uint8_t ulsch_y_mat[10000];
 
 /* UL-SCH channel interleaver according to 5.5.2.8 of 36.212 */
-void ulsch_interleave(uint8_t *q_bits, uint32_t nb_q, 
+void ulsch_interleave2(uint8_t *q_bits, uint32_t nb_q, 
                       uint8_t q_bits_ack[6], uint32_t Q_prime_ack, 
                       uint8_t q_bits_ri[6], uint32_t Q_prime_ri,
                       uint32_t Q_m, uint8_t *g_bits) 
@@ -505,6 +505,92 @@ void ulsch_interleave(uint8_t *q_bits, uint32_t nb_q,
   
 }
 
+/* UL-SCH channel interleaver according to 5.5.2.8 of 36.212 */
+void ulsch_interleave(uint8_t *g_bits, uint32_t nb_q, 
+                      uint8_t g_bits_ack[6], uint32_t Q_prime_ack, 
+                      uint8_t g_bits_ri[6], uint32_t Q_prime_ri,
+                      uint32_t Q_m,
+                     uint8_t *q_bits) 
+{
+  uint32_t C_mux;
+  uint32_t H_prime;
+  uint32_t H_prime_total;
+  uint32_t R_mux;
+  uint32_t R_prime_mux;
+  uint32_t i;
+  uint32_t j;
+  uint32_t k;
+  uint32_t r;
+  uint32_t idx;
+  uint32_t ri_column_set[4]  = {1, 4, 7, 10};
+  uint32_t ack_column_set[4] = {2, 3, 8, 9};
+  uint32_t C_ri;
+  uint32_t C_ack;
+  uint32_t N_pusch_symbs = 12; 
+  
+  // Step 1: Define C_mux
+  C_mux = N_pusch_symbs;
+
+  // Step 2: Define R_mux and R_prime_mux
+  H_prime       = nb_q;
+  H_prime_total = H_prime + Q_prime_ri;
+  R_mux         = (H_prime_total*Q_m)/C_mux;
+  R_prime_mux   = R_mux/Q_m;
+
+  
+  
+  // ACK insertion can be done at uci.c 
+  
+  // Step 5: Interleave the ACK control bits
+  i = 0;
+  j = 0;
+  r = R_prime_mux-1;
+  while(i < Q_prime_ack) {
+    C_ack = ack_column_set[j];
+    for(k=0; k<Q_m; k++) {
+      g_bits[(C_mux*r*Q_m) + C_ack*Q_m + k] = g_bits_ack[Q_m*i+k];
+    }
+    i++;
+    r = R_prime_mux - 1 - i/4;
+    j = (j + 3) % 4;
+  }
+
+  // Step 3: Interleave the RI control bits
+  i = 0;
+  j = 0;
+  r = R_prime_mux-1;
+  while(i < Q_prime_ri) {
+    C_ri = ri_column_set[j];
+    ulsch_y_idx[r*C_mux + C_ri] = 1;
+    for(k=0; k<Q_m; k++) {
+      q_bits[(r*Q_m) + C_mux*C_ri*Q_m + k] = 10+g_bits_ri[Q_m*i+k];
+    }
+    i++;
+    r = R_prime_mux - 1 - i/4;
+    j = (j + 3) % 4;
+  }
+  
+  // Step 6: Read out the bits
+  idx = 0;
+  printf("go for C_mux: %d, R_prime: %d, Q_m: %d\n", C_mux, R_prime_mux, Q_m);
+  for(i=0; i<C_mux; i++) {
+    for(j=0; j<R_prime_mux; j++) {
+      for(k=0; k<Q_m; k++) {
+        if (q_bits[idx] >= 10) {
+          printf("10 at %d is %d\n",idx, q_bits[idx]);
+          //q_bits[idx] -= 10;          
+        } else {          
+          printf("reading %d\n", j*C_mux*Q_m + i*Q_m + k);
+          q_bits[idx] = g_bits[j*C_mux*Q_m + i*Q_m + k];                       
+        }
+        idx++;
+      }
+    }
+  }
+  
+  
+}
+
 int ulsch_encode(sch_t *q, uint8_t *data, uint8_t *g_bits, 
                  harq_t *harq_process, uint32_t rv_idx, uint8_t *q_bits) 
 {
diff --git a/lte/phy/lib/phch/test/pusch_test.c b/lte/phy/lib/phch/test/pusch_test.c
index c0184e938..06e4ffe53 100644
--- a/lte/phy/lib/phch/test/pusch_test.c
+++ b/lte/phy/lib/phch/test/pusch_test.c
@@ -207,6 +207,9 @@ int main(int argc, char **argv) {
     uint32_t nof_symbols = 12*harq_process.prb_alloc.slot[0].nof_prb*RE_X_RB;
     uint32_t nof_bits_e = nof_symbols * lte_mod_bits_x_symbol(harq_process.mcs.mod);
 
+    bzero(pusch.pusch_q, nof_bits_e*sizeof(uint8_t));
+    
+
     if (ulsch_uci_encode(&pusch.dl_sch, data, uci_data, pusch.pusch_g, 
       pusch.pusch_g_ack, pusch.pusch_g_ri, &harq_process, rv, pusch.pusch_q)) 
     {