diff --git a/openair1/PHY/MODULATION/nr_modulation.c b/openair1/PHY/MODULATION/nr_modulation.c
index cba5e4ba53..9ad12c1a96 100644
--- a/openair1/PHY/MODULATION/nr_modulation.c
+++ b/openair1/PHY/MODULATION/nr_modulation.c
@@ -38,11 +38,7 @@ void nr_modulation(uint32_t *in,
   uint32_t bit_cnt;
   uint64_t x,x1,x2;
     
-#if 0//defined(__AVX2__)
-  uint16_t *in_2bytes = (uint16_t*) in;
-  __m256i *nr_mod_table256;
-  __m256i *out256;
-#elif defined(__SSE2__)
+#if defined(__SSE2__)
   __m128i *nr_mod_table128;
   __m128i *out128;
 #endif
@@ -51,6 +47,44 @@ void nr_modulation(uint32_t *in,
 
   switch (mod_order) {
 
+#if defined(__SSE2__)
+  case 2:
+    nr_mod_table128 = (__m128i*) nr_qpsk_byte_mod_table;
+    out128 = (__m128i*) out;
+    for (i=0; i<length/8; i++)
+      out128[i] = nr_mod_table128[in_bytes[i]];
+    // the bits that are left out
+    i = i*8/2;
+    nr_mod_table32 = (int32_t*) nr_qpsk_mod_table;
+    while (i<length/2){
+      idx = ((in_bytes[(i*2)/8]>>((i*2)&0x7)) & mask);
+      out32[i] = nr_mod_table32[idx];
+      i++;
+    }
+    return;
+#else
+  case 2:
+    nr_mod_table32 = (int32_t*) nr_qpsk_mod_table;
+    for (i=0; i<length/mod_order; i++) {
+      idx = ((in[i*2/32]>>((i*2)&0x1f)) & mask);
+      out32[i] = nr_mod_table32[idx];
+    }
+    return;
+#endif
+
+  case 4:
+    out64 = (int64_t*) out;
+    for (i=0; i<length/8; i++)
+      out64[i] = nr_16qam_byte_mod_table[in_bytes[i]];
+    // the bits that are left out
+    i = i*8/4;
+    while (i<length/4){
+      idx = ((in_bytes[(i*4)/8]>>((i*4)&0x7)) & mask);
+      out32[i] = nr_16qam_mod_table[idx];
+      i++;
+    }
+    return;
+
   case 6:
     j = 0;
     for (i=0; i<length/192; i++) {
@@ -113,46 +147,10 @@ void nr_modulation(uint32_t *in,
       out32[i] = nr_mod_table32[in_bytes[i]];
     return;
 
-#if 0//defined(__AVX2__) 
-// disabling this as it is slower than SSE2 
-// (maybe because L1 cache could not hold the large table)
-  case 2:
-    nr_mod_table256 = (__m256i*) nr_qpsk_2byte_mod_table;
-    out256 = (__m256i*) out;
-    for (i=0; i<length/16; i++)
-      out256[i] = nr_mod_table256[in_2bytes[i]];
-    if (length%16)
-      out256[i+1] = nr_mod_table256[in_2bytes[i]];
-    return;
-
-#elif defined(__SSE2__)
-  case 2:
-    nr_mod_table128 = (__m128i*) nr_qpsk_byte_mod_table;
-    out128 = (__m128i*) out;
-    for (i=0; i<length/8; i++)
-      out128[i] = nr_mod_table128[in_bytes[i]];
-    if (length%8)
-      out128[i+1] = nr_mod_table128[in_bytes[i]];
-    return;
-#endif
-
-  case 4:
-    out64 = (int64_t*) out;
-    for (int i=0; i<length/8; i++)
-      out64[i] = nr_16qam_byte_mod_table[in_bytes[i]];
-    return;
-
   default:
     break;
   }
 
-  nr_mod_table32 = (int32_t*) nr_qpsk_mod_table;
-  for (int i=0; i<length/mod_order; i++)
-  {
-    idx = ((in[i*mod_order/32]>>((i*mod_order)&0x1f)) & mask);
-
-    out32[i] = nr_mod_table32[idx];
-  }
 }
 
 void nr_layer_mapping(int16_t **mod_symbs,
diff --git a/openair1/PHY/NR_REFSIG/nr_gen_mod_table.c b/openair1/PHY/NR_REFSIG/nr_gen_mod_table.c
index 350b16d391..b0f9e6d03c 100644
--- a/openair1/PHY/NR_REFSIG/nr_gen_mod_table.c
+++ b/openair1/PHY/NR_REFSIG/nr_gen_mod_table.c
@@ -21,7 +21,6 @@
 
 #include "nr_refsig.h"
 #include "nr_mod_table.h"
-#include <math.h>
 
 void nr_generate_modulation_table() {
   float sqrt2 = 0.70711;
@@ -32,27 +31,14 @@ void nr_generate_modulation_table() {
   uint32_t i,j;
   short* table;
 
-nr_qpsk_mod_table[0] = 16384;
-nr_qpsk_mod_table[1] = 16384;
-nr_qpsk_mod_table[2] = -16384;
-nr_qpsk_mod_table[3] = 16384;
-nr_qpsk_mod_table[4] = 16384;
-nr_qpsk_mod_table[5] = -16384;
-nr_qpsk_mod_table[6] = -16384;
-nr_qpsk_mod_table[7] = -16384;
-
-#if 0//defined(__AVX2__)
-  //QPSK m256
-  table = (short*) nr_qpsk_2byte_mod_table;
-  for (i=0; i<65536; i++) {
-    for (j=0; j<8; j++) {
-      table[i*16+(j*2)]   = (short)(1-2*((i>>(j*2))&1))*val*sqrt2*sqrt2;
-      table[i*16+(j*2)+1] = (short)(1-2*((i>>(j*2+1))&1))*val*sqrt2*sqrt2;
-      //printf("%d j%d\n",nr_qpsk_byte_mod_table[i*8+(j*2)],nr_qpsk_byte_mod_table[i*8+(j*2)+1]);
-    }
+  // QPSK
+  for (i=0; i<4; i++) {
+    nr_qpsk_mod_table[i*2]   = (short)(1-2*(i&1))*val*sqrt2*sqrt2;
+    nr_qpsk_mod_table[i*2+1] = (short)(1-2*((i>>1)&1))*val*sqrt2*sqrt2;
+    //printf("%d j%d\n",nr_qpsk_mod_table[i*2],nr_qpsk_mod_table[i*2+1]);
   }
-#elif defined(__SSE2__)
 
+#if defined(__SSE2__)
   //QPSK m128
   table = (short*) nr_qpsk_byte_mod_table;
   for (i=0; i<256; i++) {
@@ -74,6 +60,13 @@ nr_qpsk_mod_table[7] = -16384;
     }
   }
 
+  table = (short*) nr_16qam_mod_table;
+  for (i=0; i<16; i++) {
+    table[i*2]   = (short)((1-2*(i&1))*(2-(1-2*((i>>2)&1))))*val*sqrt10*sqrt2;
+    table[i*2+1] = (short)((1-2*((i>>1)&1))*(2-(1-2*((i>>3)&1))))*val*sqrt10*sqrt2;
+    //printf("%d j%d\n",table[i*2],table[i*2+1]);
+  }
+
   //64QAM
   table = (short*) nr_64qam_mod_table;
   for (i=0; i<4096; i++) {
diff --git a/openair1/PHY/NR_REFSIG/nr_mod_table.h b/openair1/PHY/NR_REFSIG/nr_mod_table.h
index 659d63c8d0..aa8b99f584 100644
--- a/openair1/PHY/NR_REFSIG/nr_mod_table.h
+++ b/openair1/PHY/NR_REFSIG/nr_mod_table.h
@@ -31,9 +31,8 @@
 
 short nr_qpsk_mod_table[8];
 
-#if 0//defined(__AVX2__)
-__m256i nr_qpsk_2byte_mod_table[65536];
-#elif defined(__SSE2__)
+int32_t nr_16qam_mod_table[16];
+#if defined(__SSE2__)
 __m128i nr_qpsk_byte_mod_table[2048];
 #endif