chages in AVX512 code generator. Yields small improvement compared to AVX2, BLER performance is still degraded.

cmake_targets/CMakeLists.txt

@@ -138,8 +138,8 @@ if (CMAKE_SYSTEM_PROCESSOR STREQUAL "armv7l")
 else (CMAKE_SYSTEM_PROCESSOR STREQUAL "armv7l")
   if(EXISTS  "/proc/cpuinfo")
     file(STRINGS "/proc/cpuinfo" CPUINFO REGEX flags LIMIT_COUNT 1)
-    if (CPUINFO MATCHES "avx512f")
+    if (CPUINFO MATCHES "avx512bw")
-      set(C_FLAGS_PROCESSOR "${C_FLAGS_PROCESSOR} -march=skylake-avx512 -mtune=skylake-avx512")
+      set(C_FLAGS_PROCESSOR "${C_FLAGS_PROCESSOR} -mavx512bw")
       set(COMPILATION_AVX2 "True")
     else()
       if (CPUINFO MATCHES "avx2")

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_decoder.c

@@ -333,8 +333,8 @@ static inline uint32_t nrLDPC_decoder_core(int8_t* p_llr, int8_t* p_out, t_nrLDP
         {
         case 384:
         {
-           // nrLDPC_cnProc_BG1_Z384_13_AVX2(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);
+            //nrLDPC_cnProc_BG1_Z384_13_AVX2(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);
-              nrLDPC_cnProc_BG1_Z384_13_AVX512(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);  //we test  here 
+            nrLDPC_cnProc_BG1_Z384_13_AVX512(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);  //we test  here 
             break;
 	}                                                                           
         case 352:
@@ -1475,7 +1475,7 @@ static inline uint32_t nrLDPC_decoder_core(int8_t* p_llr, int8_t* p_out, t_nrLDP
         case 384:
         {
             nrLDPC_cnProc_BG2_Z384_13_AVX2(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);
-           // nrLDPC_cnProc_BG2_Z384_13_AVX512(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);
+            //nrLDPC_cnProc_BG2_Z384_13_AVX512(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);
             break;
         }
         case 352:
@@ -1922,7 +1922,7 @@ static inline uint32_t nrLDPC_decoder_core(int8_t* p_llr, int8_t* p_out, t_nrLDP
         {
         case 384:
         {
-           // nrLDPC_cnProc_BG1_Z384_13_AVX2(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);
+            //nrLDPC_cnProc_BG1_Z384_13_AVX2(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);
             nrLDPC_cnProc_BG1_Z384_13_AVX512(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes); //we test here
             break;
 	}
@@ -2273,7 +2273,7 @@ static inline uint32_t nrLDPC_decoder_core(int8_t* p_llr, int8_t* p_out, t_nrLDP
         case 384:
         {
             nrLDPC_cnProc_BG2_Z384_13_AVX2(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);
-           // nrLDPC_cnProc_BG2_Z384_13_AVX512(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);
+            //nrLDPC_cnProc_BG2_Z384_13_AVX512(p_procBuf->cnProcBuf,p_procBuf->cnProcBufRes);
             break;
         }
         case 352:

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_init_mem.h

@@ -42,7 +42,7 @@
 */
 static inline void* malloc32_clear(size_t size)
 {
-    void* ptr = (void*) memalign(32, size+32);
+    void* ptr = (void*) memalign(64, size+64);
     memset(ptr, 0, size);
     return ptr;
 }

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/generator_avx2/Makefile

@@ -8,12 +8,12 @@ OBJ= $(SRC:.c=.o)
 all: $(EXEC)
 
 cnProc_gen_avx2: $(OBJ)
-	@$(CC) -o $@ $^ $(LDFLAGS) -O2  
+	@$(CC) -o $@ $^ $(LDFLAGS) -O2 
 
 #main.o: cnProc_gen_avx2.h
 
 %.o: %.c
-	@$(CC) -o $@ -c $< $(CFLAGS) -I ${OPENAIR_HOME}/openair1 -g   
+	@$(CC) -o $@ -c $< $(CFLAGS) -I ${OPENAIR_HOME}/openair1 -g -std=c99  
 
 .PHONY: clean mrproper
 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/generator_avx2/cnProc_gen_BG1_avx2.c

@@ -21,7 +21,6 @@ void nrLDPC_cnProc_BG1_generator_AVX2(uint16_t Z,int R)
 
   fprintf(fd,"#include <stdint.h>\n");
   fprintf(fd,"#include <immintrin.h>\n");
- fprintf(fd,"#include "\"../include/avx512fintrin.h>\"\n");
 
   fprintf(fd,"void nrLDPC_cnProc_BG1_Z%d_%s_AVX2(int8_t* cnProcBuf,int8_t* cnProcBufRes) {\n",Z,ratestr[R]);
 
@@ -411,10 +410,10 @@ void nrLDPC_cnProc_BG1_generator_AVX2(uint16_t Z,int R)
   // Process group with 8 BNs
   fprintf(fd,"//Process group with 8 BNs\n");
   // Offset is 2*384/32 = 24
-  const uint8_t lut_idxCnProcG8[8][7] = {{24,48,72,96,120,144,168}, {0,48,72,96,120,144,168},
+  const uint8_t lut_idxCnProcG8[8][7] = {{24,48,72,96,112,144,168}, {0,48,72,96,112,144,168},
-					 {0,24,72,96,120,144,168}, {0,24,48,96,120,144,168},
+					 {0,24,72,96,112,144,168}, {0,24,48,96,112,144,168},
-					 {0,24,48,72,120,144,168}, {0,24,48,72,96,144,168},
+					 {0,24,48,72,112,144,168}, {0,24,48,72,96,144,168},
-					 {0,24,48,72,96,120,168}, {0,24,48,72,96,120,144}};
+					 {0,24,48,72,96,112,168},  {0,24,48,72,96,112,144}};
 
 
 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/generator_avx512/Makefile

@@ -13,7 +13,7 @@ cnProc_gen_avx512: $(OBJ)
 #main.o: cnProc_gen_avx512.h
 
 %.o: %.c
-	@$(CC) -o $@ -c $< $(CFLAGS) -I ${OPENAIR_HOME}/openair1 -g   
+	@$(CC) -o $@ -c $< $(CFLAGS) -I ${OPENAIR_HOME}/openair1 -g    -std=c99
 
 .PHONY: clean mrproper
 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/generator_avx512/cnProc_gen_BG1_avx512.c

@@ -1,8 +1,6 @@
+#include <stdio.h>
 #include <stdint.h>
-#include <immintrin.h>
 #include "../../nrLDPCdecoder_defs.h"
-#include "../../nrLDPC_types.h"                                                                                            
-#include "../../nrLDPC_bnProc.h"
 
 void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
 {
@@ -22,12 +20,7 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
   fprintf(fd,"#include <immintrin.h>\n");
 
 
-  fprintf(fd,	"__m512i _mm512_sign_epi16(__m512i a, __m512i b){ \n");     /* Emulate _mm512_sign_epi16() with instructions that exist in the AVX-512 instruction set      */
+  fprintf(fd,	"#define conditional_negate(a,b,z) _mm512_mask_sub_epi8(a,_mm512_movepi8_mask(b),z,a)\n");
-  fprintf(fd,	"b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 \n" );
-  fprintf(fd,	"b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 \n" );
-  fprintf(fd,	" a = _mm512_mullo_epi16(a, b);\n");
-  fprintf(fd,	"return a;\n");
-  fprintf(fd,  "}\n" );
 
 
   fprintf(fd,"void nrLDPC_cnProc_BG1_Z%d_%s_AVX512(int8_t* cnProcBuf,int8_t* cnProcBufRes) {\n",Z,ratestr[R]);
@@ -69,21 +62,21 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
   // Offsets are in units of bitOffsetInGroup (1*384/64)=6
      // Offsets are in units of bitOffsetInGroup (1*384/64)=6
 
-    const uint8_t lut_idxCnProcG3[3][2] = {{6,12}, {0,12}, {0,6}};
+  const uint8_t lut_idxCnProcG3[3][2] = {{12,24}, {0,24}, {0,12}};
 
 
-  fprintf(fd,"                __m512i zmm0, min, sgn,ones,maxLLR;\n");
+  fprintf(fd,"                __m512i zmm0, min, sgn,zeros,maxLLR;\n");
-  fprintf(fd,"                ones   = _mm512_set1_epi8((char)1);\n");
+  fprintf(fd,"                zeros  = _mm512_setzero_si512();\n");
   fprintf(fd,"                maxLLR = _mm512_set1_epi8((char)127);\n");
 
   if (lut_numCnInCnGroups[0] > 0)
     {
       // Number of groups of 64  CNs for parallel processing
       // Ceil for values not divisible by 64
-      M = (lut_numCnInCnGroups[0]*Z + 63)>>5;
+      M = (lut_numCnInCnGroups[0]*Z + 63)>>6;
 
       // Set the offset to each bit within a group in terms of 64  Byte
-      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[0]*NR_LDPC_ZMAX)>>5;
+      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[0]*NR_LDPC_ZMAX)>>6;
 
 
       // Set pointers to start of group 3
@@ -104,53 +97,51 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
 	  fprintf(fd,"            for (int i=0;i<%d;i+=2) {\n",M);
 	  // Abs and sign of 64  CNs (first BN)
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[0]>>5)+lut_idxCnProcG3[j][0]);
+	  fprintf(fd,"                sgn = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[0]>>6)+lut_idxCnProcG3[j][0]/2);
-	  //                sgn  = _mm512_sign_epi16(ones, zmm0);
+	  //                sgn  = _mm512_sign_epi8(ones, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(ones, zmm0);\n");
 	  //                min  = _mm512_abs_epi8(zmm0);
-	  fprintf(fd,"                min  = _mm512_abs_epi8(zmm0);\n");
+	  fprintf(fd,"                min  = _mm512_abs_epi8(sgn);\n");
 	  
 	  // 32 CNs of second BN
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][1] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[0]>>5)+lut_idxCnProcG3[j][1]);
+	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[0]>>6)+lut_idxCnProcG3[j][1]/2);
 	  
 	  //                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));
 	  fprintf(fd,"                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));\n");
 	  
-	  //                sgn  = _mm512_sign_epi16(sgn, zmm0);
+	  //                sgn  = _mm512_sign_epi8(sgn, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(sgn, zmm0);\n");
+	  fprintf(fd,"                sgn  = _mm512_xor_si512(sgn, zmm0);\n");
 	  
 	  // Store result
 	  //                min = _mm512_min_epu8(min, maxLLR); // 128 in epi8 is -127
 	  fprintf(fd,"                min = _mm512_min_epu8(min, maxLLR);\n");
-	  //                *p_cnProcBufResBit = _mm512_sign_epi16(min, sgn);
+	  //                *p_cnProcBufResBit = _mm512_sign_epi8(min, sgn);
 	  //                p_cnProcBufResBit++;
-	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = _mm512_sign_epi16(min, sgn);\n",(lut_startAddrCnGroups[0]>>5)+(j*bitOffsetInGroup));
+	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = conditional_negate(min, sgn,zeros);\n",(lut_startAddrCnGroups[0]>>6)+(j*bitOffsetInGroup));
 
 	  // Abs and sign of 64  CNs (first BN)
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[0]>>5)+lut_idxCnProcG3[j][0]+1);
+	  fprintf(fd,"                sgn = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[0]>>6)+(lut_idxCnProcG3[j][0]/2)+1);
-	  //                sgn  = _mm512_sign_epi16(ones, zmm0);
+	  //                sgn  = _mm512_sign_epi8(ones, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(ones, zmm0);\n");
 	  //                min  = _mm512_abs_epi8(zmm0);
-	  fprintf(fd,"                min  = _mm512_abs_epi8(zmm0);\n");
+	  fprintf(fd,"                min  = _mm512_abs_epi8(sgn);\n");
 	  
 	  // 32 CNs of second BN
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][1] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[0]>>5)+lut_idxCnProcG3[j][1]+1);
+	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[0]>>6)+(lut_idxCnProcG3[j][1]/2)+1);
 	  
 	  //                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));
 	  fprintf(fd,"                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));\n");
 	  
-	  //                sgn  = _mm512_sign_epi16(sgn, zmm0);
+	  //                sgn  = _mm512_sign_epi8(sgn, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(sgn, zmm0);\n");
+	  fprintf(fd,"                sgn  = _mm512_xor_si512(sgn, zmm0);\n");
 	  
 	  // Store result
 	  //                min = _mm512_min_epu8(min, maxLLR); // 128 in epi8 is -127
 	  fprintf(fd,"                min = _mm512_min_epu8(min, maxLLR);\n");
-	  //                *p_cnProcBufResBit = _mm512_sign_epi16(min, sgn);
+	  //                *p_cnProcBufResBit = _mm512_sign_epi8(min, sgn);
 	  //                p_cnProcBufResBit++;
-	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = _mm512_sign_epi16(min, sgn);\n",(lut_startAddrCnGroups[0]>>5)+(j*bitOffsetInGroup)+1);
+	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = conditional_negate(min, sgn,zeros);\n",(lut_startAddrCnGroups[0]>>6)+(j*bitOffsetInGroup)+1);
 
 	  fprintf(fd,"            }\n");
         }
@@ -160,16 +151,16 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
   // Process group with 4 BNs
   fprintf(fd,"//Process group with 4 BNs\n");
     // Offset is 5*384/64 = 30
-    const uint8_t lut_idxCnProcG4[4][3] = {{30,60,90}, {0,60,90}, {0,30,90}, {0,30,60}};
+  const uint8_t lut_idxCnProcG4[4][3] = {{60,120,180}, {0,120,180}, {0,60,180}, {0,60,120}};
 
   if (lut_numCnInCnGroups[1] > 0)
     {
       // Number of groups of 64  CNs for parallel processing
       // Ceil for values not divisible by 64
-      M = (lut_numCnInCnGroups[1]*Z + 63)>>5;
+      M = (lut_numCnInCnGroups[1]*Z + 63)>>6;
 
       // Set the offset to each bit within a group in terms of 64  Byte
-      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[1]*NR_LDPC_ZMAX)>>5;
+      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[1]*NR_LDPC_ZMAX)>>6;
 
 
       // Set pointers to start of group 4
@@ -188,31 +179,30 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
 	  fprintf(fd,"            for (int i=0;i<%d;i++) {\n",M);
 	  // Abs and sign of 64  CNs (first BN)
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[1]>>5)+lut_idxCnProcG4[j][0]);
+	  fprintf(fd,"                sgn = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[1]>>6)+lut_idxCnProcG4[j][0]/2);
-	  //                sgn  = _mm512_sign_epi16(ones, zmm0);
+	  //                sgn  = _mm512_sign_epi8(ones, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(ones, zmm0);\n");
 	  //                min  = _mm512_abs_epi8(zmm0);
-	  fprintf(fd,"                min  = _mm512_abs_epi8(zmm0);\n");
+	  fprintf(fd,"                min  = _mm512_abs_epi8(sgn);\n");
 	  
 	  
 	  // Loop over BNs
 	  for (k=1; k<3; k++)
 	    {
-	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[1]>>5)+lut_idxCnProcG4[j][k]);
+	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[1]>>6)+lut_idxCnProcG4[j][k]/2);
 	      
 	      //                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));
 	      fprintf(fd,"                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));\n");
 	      
-	      //                sgn  = _mm512_sign_epi16(sgn, zmm0);
+	      //                sgn  = _mm512_sign_epi8(sgn, zmm0);
-	      fprintf(fd,"                sgn  = _mm512_sign_epi16(sgn, zmm0);\n");
+	      fprintf(fd,"                sgn  = _mm512_xor_si512(sgn, zmm0);\n");
 	    }
 	  
 	  // Store result
 	  //                min = _mm512_min_epu8(min, maxLLR); // 128 in epi8 is -127
 	  fprintf(fd,"                min = _mm512_min_epu8(min, maxLLR);\n");
-	  //                *p_cnProcBufResBit = _mm512_sign_epi16(min, sgn);
+	  //                *p_cnProcBufResBit = _mm512_sign_epi8(min, sgn);
 	      //                p_cnProcBufResBit++;
-	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = _mm512_sign_epi16(min, sgn);\n",(lut_startAddrCnGroups[1]>>5)+(j*bitOffsetInGroup));
+	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = conditional_negate(min, sgn,zeros);\n",(lut_startAddrCnGroups[1]>>6)+(j*bitOffsetInGroup));
 	  fprintf(fd,"            }\n");
         }
     }
@@ -222,18 +212,18 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
   // Process group with 5 BNs
   fprintf(fd,"//Process group with 5 BNs\n");
   // Offset is 18*384/64 = 216
-	const uint16_t lut_idxCnProcG5[5][4] = {{108,216,324,432}, {0,216,324,432},
+  const uint16_t lut_idxCnProcG5[5][4] = {{216,432,648,864}, {0,432,648,864},
-                                            {0,108,324,432}, {0,108,216,432}, {0,108,216,324}};
+                                          {0,216,648,864}, {0,216,432,864}, {0,216,432,648}};
 
 
   if (lut_numCnInCnGroups[2] > 0)
     {
       // Number of groups of 64  CNs for parallel processing
       // Ceil for values not divisible by 64
-      M = (lut_numCnInCnGroups[2]*Z + 63)>>5;
+      M = (lut_numCnInCnGroups[2]*Z + 63)>>6;
 
       // Set the offset to each bit within a group in terms of 64  Byte
-      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[2]*NR_LDPC_ZMAX)>>5;
+      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[2]*NR_LDPC_ZMAX)>>6;
 
 
       // Set pointers to start of group 4
@@ -253,31 +243,30 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
 	  fprintf(fd,"            for (int i=0;i<%d;i++) {\n",M);
 	  // Abs and sign of 64  CNs (first BN)
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[2]>>5)+lut_idxCnProcG5[j][0]);
+	  fprintf(fd,"                sgn = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[2]>>6)+lut_idxCnProcG5[j][0]/2);
-	  //                sgn  = _mm512_sign_epi16(ones, zmm0);
+	  //                sgn  = _mm512_sign_epi8(ones, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(ones, zmm0);\n");
 	  //                min  = _mm512_abs_epi8(zmm0);
-	  fprintf(fd,"                min  = _mm512_abs_epi8(zmm0);\n");
+	  fprintf(fd,"                min  = _mm512_abs_epi8(sgn);\n");
 	  
 	  
 	  // Loop over BNs
 	  for (k=1; k<4; k++)
 	    {
-	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[2]>>5)+lut_idxCnProcG5[j][k]);
+	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[2]>>6)+lut_idxCnProcG5[j][k]/2);
 	      
 	      //                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));
 	      fprintf(fd,"                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));\n");
 	      
-	      //                sgn  = _mm512_sign_epi16(sgn, zmm0);
+	      //                sgn  = _mm512_sign_epi8(sgn, zmm0);
-	      fprintf(fd,"                sgn  = _mm512_sign_epi16(sgn, zmm0);\n");
+	      fprintf(fd,"                sgn  = _mm512_xor_si512(sgn, zmm0);\n");
 	    }
 	  
 	  // Store result
 	  //                min = _mm512_min_epu8(min, maxLLR); // 128 in epi8 is -127
 	  fprintf(fd,"                min = _mm512_min_epu8(min, maxLLR);\n");
-	  //                *p_cnProcBufResBit = _mm512_sign_epi16(min, sgn);
+	  //                *p_cnProcBufResBit = _mm512_sign_epi8(min, sgn);
 	  //                p_cnProcBufResBit++;
-	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = _mm512_sign_epi16(min, sgn);\n",(lut_startAddrCnGroups[2]>>5)+(j*bitOffsetInGroup));
+	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = conditional_negate(min, sgn,zeros);\n",(lut_startAddrCnGroups[2]>>6)+(j*bitOffsetInGroup));
 	  fprintf(fd,"           }\n");
         }
     }
@@ -286,19 +275,19 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
   // Process group with 6 BNs
   fprintf(fd,"//Process group with 6 BNs\n");
     // Offset is 8*384/64 = 48
-    const uint16_t lut_idxCnProcG6[6][5] = {{48,96,144,192,240}, {0,96,144,192,240},
+  const uint16_t lut_idxCnProcG6[6][5] = {{96,192,288,384,480}, {0,192,288,384,480},
-                                            {0,48,144,192,240}, {0,48,96,192,240},
+                                          {0,96,288,384,480}, {0,96,192,384,480},
-                                            {0,48,96,144,240}, {0,48,96,144,192}};
+                                          {0,96,192,288,480}, {0,96,192,288,384}};
 
 
   if (lut_numCnInCnGroups[3] > 0)
     {
       // Number of groups of 64  CNs for parallel processing
       // Ceil for values not divisible by 64
-      M = (lut_numCnInCnGroups[3]*Z + 63)>>5;
+      M = (lut_numCnInCnGroups[3]*Z + 63)>>6;
 
       // Set the offset to each bit within a group in terms of 64  Byte
-      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[3]*NR_LDPC_ZMAX)>>5;
+      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[3]*NR_LDPC_ZMAX)>>6;
 
 
       // Set pointers to start of group 4
@@ -318,31 +307,30 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
 	  fprintf(fd,"            for (int i=0;i<%d;i++) {\n",M);
 	  // Abs and sign of 64  CNs (first BN)
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[3]>>5)+lut_idxCnProcG6[j][0]);
+	  fprintf(fd,"                sgn = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[3]>>6)+lut_idxCnProcG6[j][0]/2);
-	  //                sgn  = _mm512_sign_epi16(ones, zmm0);
+	  //                sgn  = _mm512_sign_epi8(ones, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(ones, zmm0);\n");
 	  //                min  = _mm512_abs_epi8(zmm0);
-	  fprintf(fd,"                min  = _mm512_abs_epi8(zmm0);\n");
+	  fprintf(fd,"                min  = _mm512_abs_epi8(sgn);\n");
 	  
 	  
 	  // Loop over BNs
 	  for (k=1; k<5; k++)
 	    {
-	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[3]>>5)+lut_idxCnProcG6[j][k]);
+	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[3]>>6)+lut_idxCnProcG6[j][k]/2);
 	      
 	      //                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));
 	      fprintf(fd,"                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));\n");
 	      
-	      //                sgn  = _mm512_sign_epi16(sgn, zmm0);
+	      //                sgn  = _mm512_sign_epi8(sgn, zmm0);
-	      fprintf(fd,"                sgn  = _mm512_sign_epi16(sgn, zmm0);\n");
+	      fprintf(fd,"                sgn  = _mm512_xor_si512(sgn, zmm0);\n");
 	    }
 	  
 	  // Store result
 	  //                min = _mm512_min_epu8(min, maxLLR); // 128 in epi8 is -127
 	  fprintf(fd,"                min = _mm512_min_epu8(min, maxLLR);\n");
-	  //                *p_cnProcBufResBit = _mm512_sign_epi16(min, sgn);
+	  //                *p_cnProcBufResBit = _mm512_sign_epi8(min, sgn);
 	  //                p_cnProcBufResBit++;
-	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = _mm512_sign_epi16(min, sgn);\n",(lut_startAddrCnGroups[3]>>5)+(j*bitOffsetInGroup));
+	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = conditional_negate(min, sgn,zeros);\n",(lut_startAddrCnGroups[3]>>6)+(j*bitOffsetInGroup));
 	  fprintf(fd,"            }\n");
 	}
     }
@@ -352,20 +340,20 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
   // Process group with 7 BNs
   fprintf(fd,"//Process group with 7 BNs\n");
     // Offset is 5*384/64 = 30
-    const uint16_t lut_idxCnProcG7[7][6] = {{30,60,90,120,150,180}, {0,60,90,120,150,180},
+  const uint16_t lut_idxCnProcG7[7][6] = {{60,120,180,240,300,360}, {0,120,180,240,300,360},
-                                            {0,30,90,120,150,180},   {0,30,60,120,150,180},
+                                          {0,60,180,240,300,360},   {0,60,120,240,300,360},
-                                            {0,30,60,90,150,180},   {0,30,60,90,120,180},
+                                          {0,60,120,180,300,360},   {0,60,120,180,240,360},
-                                            {0,30,60,90,120,150}};
+                                          {0,60,120,180,240,300}};
 
 
   if (lut_numCnInCnGroups[4] > 0)
     {
       // Number of groups of 64  CNs for parallel processing
       // Ceil for values not divisible by 64
-      M = (lut_numCnInCnGroups[4]*Z + 63)>>5;
+      M = (lut_numCnInCnGroups[4]*Z + 63)>>6;
 
       // Set the offset to each bit within a group in terms of 64  Byte
-      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[4]*NR_LDPC_ZMAX)>>5;
+      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[4]*NR_LDPC_ZMAX)>>6;
 
 
       // Set pointers to start of group 4
@@ -385,31 +373,30 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
 	  fprintf(fd,"            for (int i=0;i<%d;i++) {\n",M);
 	  // Abs and sign of 64  CNs (first BN)
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[4]>>5)+lut_idxCnProcG7[j][0]);
+	  fprintf(fd,"                sgn= ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[4]>>6)+lut_idxCnProcG7[j][0]/2);
-	  //                sgn  = _mm512_sign_epi16(ones, zmm0);
+	  //                sgn  = _mm512_sign_epi816(ones, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(ones, zmm0);\n");
 	  //                min  = _mm512_abs_epi8(zmm0);
-	  fprintf(fd,"                min  = _mm512_abs_epi8(zmm0);\n");
+	  fprintf(fd,"                min  = _mm512_abs_epi8(sgn);\n");
 	  
 	  
 	  // Loop over BNs
 	  for (k=1; k<6; k++)
 	    {
-	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[4]>>5)+lut_idxCnProcG7[j][k]);
+	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[4]>>6)+lut_idxCnProcG7[j][k]/2);
 	      
 	      //                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));
 	      fprintf(fd,"                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));\n");
 	      
-	      //                sgn  = _mm512_sign_epi16(sgn, zmm0);
+	      //                sgn  = _mm512_sign_epi8(sgn, zmm0);
-	      fprintf(fd,"                sgn  = _mm512_sign_epi16(sgn, zmm0);\n");
+	      fprintf(fd,"                sgn  = _mm512_xor_si512(sgn, zmm0);\n");
 	    }
 	  
 	  // Store result
 	  //                min = _mm512_min_epu8(min, maxLLR); // 128 in epi8 is -127
 	  fprintf(fd,"                min = _mm512_min_epu8(min, maxLLR);\n");
-	  //                *p_cnProcBufResBit = _mm512_sign_epi16(min, sgn);
+	  //                *p_cnProcBufResBit = _mm512_sign_epi8(min, sgn);
 	  //                p_cnProcBufResBit++;
-	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = _mm512_sign_epi16(min, sgn);\n",(lut_startAddrCnGroups[4]>>5)+(j*bitOffsetInGroup));
+	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = conditional_negate(min, sgn,zeros);\n",(lut_startAddrCnGroups[4]>>6)+(j*bitOffsetInGroup));
 	  fprintf(fd,"            }\n");
 	}
     }
@@ -419,10 +406,10 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
   // Process group with 8 BNs
   fprintf(fd,"//Process group with 8 BNs\n");
     // Offset is 2*384/64 = 12
-    const uint8_t lut_idxCnProcG8[8][7] = {{12,24,36,48,56,72,84}, {0,24,36,48,56,72,84},
+  const uint8_t lut_idxCnProcG8[8][7] = {{24,48,72,96,112,144,168}, {0,48,72,96,112,144,168},
-                                           {0,12,36,48,56,72,84}, {0,12,24,48,56,72,84},
+                                         {0,24,72,96,112,144,168}, {0,24,48,96,112,144,168},
-                                           {0,12,24,36,56,72,84}, {0,12,24,36,48,72,84},
+                                         {0,24,48,72,112,144,168}, {0,24,48,72,96,144,168},
-                                           {0,12,24,36,48,56,84}, {0,12,24,36,48,120,72}};
+                                         {0,24,48,72,96,112,168},  {0,24,48,72,96,112,144}};
 
 
 
@@ -430,10 +417,10 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
     {
       // Number of groups of 64  CNs for parallel processing
       // Ceil for values not divisible by 64
-      M = (lut_numCnInCnGroups[5]*Z + 63)>>5;
+      M = (lut_numCnInCnGroups[5]*Z + 63)>>6;
 
       // Set the offset to each bit within a group in terms of 64  Byte
-      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[5]*NR_LDPC_ZMAX)>>5;
+      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[5]*NR_LDPC_ZMAX)>>6;
 
 
       // Set pointers to start of group 4
@@ -453,31 +440,30 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
 	  fprintf(fd,"            for (int i=0;i<%d;i++) {\n",M);
 	  // Abs and sign of 64  CNs (first BN)
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[5]>>5)+lut_idxCnProcG8[j][0]);
+	  fprintf(fd,"                sgn = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[5]>>6)+lut_idxCnProcG8[j][0]/2);
-	  //                sgn  = _mm512_sign_epi16(ones, zmm0);
+	  //                sgn  = _mm512_sign_epi8(ones, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(ones, zmm0);\n");
 	  //                min  = _mm512_abs_epi8(zmm0);
-	  fprintf(fd,"                min  = _mm512_abs_epi8(zmm0);\n");
+	  fprintf(fd,"                min  = _mm512_abs_epi8(sgn);\n");
 	  
 	  
 	  // Loop over BNs
 	  for (k=1; k<7; k++)
 	    {
-	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[5]>>5)+lut_idxCnProcG8[j][k]);
+	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[5]>>6)+lut_idxCnProcG8[j][k]/2);
 	      
 	      //                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));
 	      fprintf(fd,"                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));\n");
 	      
-	      //                sgn  = _mm512_sign_epi16(sgn, zmm0);
+	      //                sgn  = _mm512_sign_epi8(sgn, zmm0);
-	      fprintf(fd,"                sgn  = _mm512_sign_epi16(sgn, zmm0);\n");
+	      fprintf(fd,"                sgn  = _mm512_xor_si512(sgn, zmm0);\n");
 	    }
 	  
 	  // Store result
 	  //                min = _mm512_min_epu8(min, maxLLR); // 128 in epi8 is -127
 	  fprintf(fd,"                min = _mm512_min_epu8(min, maxLLR);\n");
-	  //                *p_cnProcBufResBit = _mm512_sign_epi16(min, sgn);
+	  //                *p_cnProcBufResBit = _mm512_sign_epi8(min, sgn);
 	  //                p_cnProcBufResBit++;
-	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = _mm512_sign_epi16(min, sgn);\n",(lut_startAddrCnGroups[5]>>5)+(j*bitOffsetInGroup));
+	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = conditional_negate(min, sgn,zeros);\n",(lut_startAddrCnGroups[5]>>6)+(j*bitOffsetInGroup));
 	  fprintf(fd,"              }\n");
         }
     }
@@ -486,11 +472,11 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
   // Process group with 9 BNs
   fprintf(fd,"//Process group with 9 BNs\n");
     // Offset is 2*384/64 = 12
-    const uint8_t lut_idxCnProcG9[9][8] = {{12,24,36,48,60,72,84,96}, {0,24,36,48,60,72,84,96},
+  const uint8_t lut_idxCnProcG9[9][8] = {{24,48,72,96,120,144,168,192}, {0,48,72,96,120,144,168,192},
-                                           {0,12,36,48,60,72,84,96}, {0,12,24,48,60,72,84,96},
+                                         {0,24,72,96,120,144,168,192}, {0,24,48,96,120,144,168,192},
-                                           {0,12,24,36,60,72,84,96}, {0,12,24,36,48,72,84,96},
+                                         {0,24,48,72,120,144,168,192}, {0,24,48,72,96,144,168,192},
-                                           {0,12,24,36,48,60,84,96}, {0,12,24,36,48,60,72,96},
+                                         {0,24,48,72,96,120,168,192}, {0,24,48,72,96,120,144,192},
-                                           {0,12,24,36,48,60,72,84}};
+                                         {0,24,48,72,96,120,144,168}};
 
 
 
@@ -499,10 +485,10 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
     {
       // Number of groups of 64  CNs for parallel processing
       // Ceil for values not divisible by 64
-      M = (lut_numCnInCnGroups[6]*Z + 63)>>5;
+      M = (lut_numCnInCnGroups[6]*Z + 63)>>6;
 
       // Set the offset to each bit within a group in terms of 64  Byte
-      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[6]*NR_LDPC_ZMAX)>>5;
+      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[6]*NR_LDPC_ZMAX)>>6;
 
 
       // Set pointers to start of group 9
@@ -522,31 +508,30 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
 	  fprintf(fd,"            for (int i=0;i<%d;i++) {\n",M);
 	  // Abs and sign of 64  CNs (first BN)
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[6]>>5)+lut_idxCnProcG9[j][0]);
+	  fprintf(fd,"                sgn = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[6]>>6)+lut_idxCnProcG9[j][0]/2);
-	  //                sgn  = _mm512_sign_epi16(ones, zmm0);
+	  //                sgn  = _mm512_sign_epi8(ones, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(ones, zmm0);\n");
 	  //                min  = _mm512_abs_epi8(zmm0);
-	  fprintf(fd,"                min  = _mm512_abs_epi8(zmm0);\n");
+	  fprintf(fd,"                min  = _mm512_abs_epi8(sgn);\n");
 	  
 	  
 	  // Loop over BNs
 	  for (k=1; k<8; k++)
 	    {
-	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[6]>>5)+lut_idxCnProcG9[j][k]);
+	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[6]>>6)+lut_idxCnProcG9[j][k]/2);
 	      
 	      //                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));
 	      fprintf(fd,"                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));\n");
 	      
-	      //                sgn  = _mm512_sign_epi16(sgn, zmm0);
+	      //                sgn  = _mm512_sign_epi8(sgn, zmm0);
-	      fprintf(fd,"                sgn  = _mm512_sign_epi16(sgn, zmm0);\n");
+	      fprintf(fd,"                sgn  = _mm512_xor_si512(sgn, zmm0);\n");
 	    }
 	  
 	  // Store result
 	  //                min = _mm512_min_epu8(min, maxLLR); // 128 in epi8 is -127
 	  fprintf(fd,"                min = _mm512_min_epu8(min, maxLLR);\n");
-	  //                *p_cnProcBufResBit = _mm512_sign_epi16(min, sgn);
+	  //                *p_cnProcBufResBit = _mm512_sign_epi8(min, sgn);
 	  //                p_cnProcBufResBit++;
-	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = _mm512_sign_epi16(min, sgn);\n",(lut_startAddrCnGroups[6]>>5)+(j*bitOffsetInGroup));
+	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = conditional_negate(min, sgn, zeros);\n",(lut_startAddrCnGroups[6]>>6)+(j*bitOffsetInGroup));
 	  fprintf(fd,"            }\n");
 	}
     }
@@ -555,11 +540,12 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
   // Process group with 10 BNs
   fprintf(fd,"//Process group with 10 BNs\n");
   	// Offset is 1*384/64 = 6
-    const uint8_t lut_idxCnProcG10[10][9] = {{6,12,18,24,30,36,42,48,54}, {0,12,18,24,30,36,42,48,54},
+  const uint8_t lut_idxCnProcG10[10][9] = {{12,24,36,48,60,72,84,96,108}, {0,24,36,48,60,72,84,96,108},
-                                             {0,6,18,24,30,36,42,48,54}, {0,6,12,24,30,36,42,48,54},
+                                           {0,12,36,48,60,72,84,96,108}, {0,12,24,48,60,72,84,96,108},
-                                             {0,6,12,18,30,36,42,48,54}, {0,6,12,18,24,36,42,48,54},
+                                           {0,12,24,36,60,72,84,96,108}, {0,12,24,36,48,72,84,96,108},
-                                             {0,6,12,18,24,30,42,48,54}, {0,6,12,18,24,30,36,48,54},
+                                           {0,12,24,36,48,60,84,96,108}, {0,12,24,36,48,60,72,96,108},
-                                             {0,6,12,18,24,30,36,42,54}, {0,6,12,36,24,30,36,42,48}};
+                                           {0,12,24,36,48,60,72,84,108}, {0,12,24,36,48,60,72,84,96}};
+
 
 
 
@@ -569,10 +555,10 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
     {
       // Number of groups of 64  CNs for parallel processing
       // Ceil for values not divisible by 64
-      M = (lut_numCnInCnGroups[7]*Z + 63)>>5;
+      M = (lut_numCnInCnGroups[7]*Z + 63)>>6;
 
       // Set the offset to each bit within a group in terms of 64  Byte
-      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[7]*NR_LDPC_ZMAX)>>5;
+      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[7]*NR_LDPC_ZMAX)>>6;
 
 
       // Set pointers to start of group 10
@@ -592,31 +578,30 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
 	  fprintf(fd,"            for (int i=0;i<%d;i++) {\n",M);
 	  // Abs and sign of 64  CNs (first BN)
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[7]>>5)+lut_idxCnProcG10[j][0]);
+	  fprintf(fd,"                sgn = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[7]>>6)+lut_idxCnProcG10[j][0]/2);
-	  //                sgn  = _mm512_sign_epi16(ones, zmm0);
+	  //                sgn  = _mm512_sign_epi8(ones, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(ones, zmm0);\n");
 	  //                min  = _mm512_abs_epi8(zmm0);
-	  fprintf(fd,"                min  = _mm512_abs_epi8(zmm0);\n");
+	  fprintf(fd,"                min  = _mm512_abs_epi8(sgn);\n");
 	  
 	  
 	  // Loop over BNs
 	  for (k=1; k<9; k++)
 	    {
-	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[7]>>5)+lut_idxCnProcG10[j][k]);
+	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[7]>>6)+lut_idxCnProcG10[j][k]/2);
 	      
 	      //                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));
 	      fprintf(fd,"                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));\n");
 	      
-	      //                sgn  = _mm512_sign_epi16(sgn, zmm0);
+	      //                sgn  = _mm512_sign_epi8(sgn, zmm0);
-	      fprintf(fd,"                sgn  = _mm512_sign_epi16(sgn, zmm0);\n");
+	      fprintf(fd,"                sgn  = _mm512_xor_si512(sgn, zmm0);\n");
 	    }
 	  
 	  // Store result
 	  //                min = _mm512_min_epu8(min, maxLLR); // 128 in epi8 is -127
 	  fprintf(fd,"                min = _mm512_min_epu8(min, maxLLR);\n");
-	  //                *p_cnProcBufResBit = _mm512_sign_epi16(min, sgn);
+	  //                *p_cnProcBufResBit = _mm512_sign_epi8(min, sgn);
 	  //                p_cnProcBufResBit++;
-	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = _mm512_sign_epi16(min, sgn);\n",(lut_startAddrCnGroups[7]>>5)+(j*bitOffsetInGroup));
+	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = conditional_negate(min,sgn,zeros);\n",(lut_startAddrCnGroups[7]>>6)+(j*bitOffsetInGroup));
 	  fprintf(fd,"            }\n");
         }
     }
@@ -626,26 +611,26 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
   // Process group with 19 BNs
   fprintf(fd,"//Process group with 19 BNs\n");
   // Offset is 4*384/64 = 24
-    const uint16_t lut_idxCnProcG19[19][18] = {{24,48,72,96,120,144,168,192,216,240,264,288,312,336,360,384,408,432}, {0,48,72,96,120,144,168,192,216,240,264,288,312,336,360,384,408,432},
+  const uint16_t lut_idxCnProcG19[19][18] = {{48,96,144,192,240,288,336,384,432,480,528,576,624,672,720,768,816,864}, {0,96,144,192,240,288,336,384,432,480,528,576,624,672,720,768,816,864},
-                                               {0,24,72,96,120,144,168,192,216,240,264,288,312,336,360,384,408,432}, {0,24,48,96,120,144,168,192,216,240,264,288,312,336,360,384,408,432},
+                                             {0,48,144,192,240,288,336,384,432,480,528,576,624,672,720,768,816,864}, {0,48,96,192,240,288,336,384,432,480,528,576,624,672,720,768,816,864},
-                                               {0,24,48,72,120,144,168,192,216,240,264,288,312,336,360,384,408,432}, {0,24,48,72,96,144,168,192,216,240,264,288,312,336,360,384,408,432},
+                                             {0,48,96,144,240,288,336,384,432,480,528,576,624,672,720,768,816,864}, {0,48,96,144,192,288,336,384,432,480,528,576,624,672,720,768,816,864},
-                                               {0,24,48,72,96,120,168,192,216,240,264,288,312,336,360,384,408,432}, {0,24,48,72,96,120,144,192,216,240,264,288,312,336,360,384,408,432},
+                                             {0,48,96,144,192,240,336,384,432,480,528,576,624,672,720,768,816,864}, {0,48,96,144,192,240,288,384,432,480,528,576,624,672,720,768,816,864},
-                                               {0,24,48,72,96,120,144,168,216,240,264,288,312,336,360,384,408,432}, {0,24,48,72,96,120,144,168,192,240,264,288,312,336,360,384,408,432},
+                                             {0,48,96,144,192,240,288,336,432,480,528,576,624,672,720,768,816,864}, {0,48,96,144,192,240,288,336,384,480,528,576,624,672,720,768,816,864},
-                                               {0,24,48,72,96,120,144,168,192,216,264,288,312,336,360,384,408,432}, {0,24,48,72,96,120,144,168,192,216,240,288,312,336,360,384,408,432},
+                                             {0,48,96,144,192,240,288,336,384,432,528,576,624,672,720,768,816,864}, {0,48,96,144,192,240,288,336,384,432,480,576,624,672,720,768,816,864},
-                                               {0,24,48,72,96,120,144,168,192,216,240,264,312,336,360,384,408,432}, {0,24,48,72,96,120,144,168,192,216,240,264,288,336,360,384,408,432},
+                                             {0,48,96,144,192,240,288,336,384,432,480,528,624,672,720,768,816,864}, {0,48,96,144,192,240,288,336,384,432,480,528,576,672,720,768,816,864},
-                                               {0,24,48,72,96,120,144,168,192,216,240,264,288,312,360,384,408,432}, {0,24,48,72,96,120,144,168,192,216,240,264,288,312,336,384,408,432},
+                                             {0,48,96,144,192,240,288,336,384,432,480,528,576,624,720,768,816,864}, {0,48,96,144,192,240,288,336,384,432,480,528,576,624,672,768,816,864},
-                                               {0,24,48,72,96,120,144,168,192,216,240,264,288,312,336,360,408,432}, {0,24,48,72,96,120,144,168,192,216,240,264,288,312,336,360,384,432},
+                                             {0,48,96,144,192,240,288,336,384,432,480,528,576,624,672,720,816,864}, {0,48,96,144,192,240,288,336,384,432,480,528,576,624,672,720,768,864},
-                                               {0,24,48,72,96,120,144,168,192,216,240,264,288,312,336,360,384,408}};
+                                             {0,48,96,144,192,240,288,336,384,432,480,528,576,624,672,720,768,816}};
 
 
   if (lut_numCnInCnGroups[8] > 0)
     {
       // Number of groups of 64  CNs for parallel processing
       // Ceil for values not divisible by 64
-      M = (lut_numCnInCnGroups[8]*Z + 63)>>5;
+      M = (lut_numCnInCnGroups[8]*Z + 63)>>6;
 
       // Set the offset to each bit within a group in terms of 64  Byte
-      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[8]*NR_LDPC_ZMAX)>>5;
+      bitOffsetInGroup = (lut_numCnInCnGroups_BG1_R13[8]*NR_LDPC_ZMAX)>>6;
 
 
       // Set pointers to start of group 19
@@ -665,31 +650,30 @@ void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t Z,int R)
 	  fprintf(fd,"            for (int i=0;i<%d;i++) {\n",M);
 	  // Abs and sign of 64  CNs (first BN)
 	  //                zmm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
-	  fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[8]>>5)+lut_idxCnProcG19[j][0]);
+	  fprintf(fd,"                sgn = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[8]>>6)+lut_idxCnProcG19[j][0]/2);
-	  //                sgn  = _mm512_sign_epi16(ones, zmm0);
+	  //                sgn  = _mm512_sign_epi8(ones, zmm0);
-	  fprintf(fd,"                sgn  = _mm512_sign_epi16(ones, zmm0);\n");
 	  //                min  = _mm512_abs_epi8(zmm0);
-	  fprintf(fd,"                min  = _mm512_abs_epi8(zmm0);\n");
+	  fprintf(fd,"                min  = _mm512_abs_epi8(sgn);\n");
 	  
 	  
 	  // Loop over BNs
 	  for (k=1; k<18; k++)
 	    {
-	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[8]>>5)+lut_idxCnProcG19[j][k]);
+	      fprintf(fd,"                zmm0 = ((__m512i*)cnProcBuf)[%d+i];\n",(lut_startAddrCnGroups[8]>>6)+lut_idxCnProcG19[j][k]/2);
 	      
 	      //                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));
 	      fprintf(fd,"                min  = _mm512_min_epu8(min, _mm512_abs_epi8(zmm0));\n");
 	      
-	      //                sgn  = _mm512_sign_epi16(sgn, zmm0);
+	      //                sgn  = _mm512_sign_epi8(sgn, zmm0);
-	      fprintf(fd,"                sgn  = _mm512_sign_epi16(sgn, zmm0);\n");
+	      fprintf(fd,"                sgn  = _mm512_xor_si512(sgn, zmm0);\n");
 	    }
 	  
 	  // Store result
 	  //                min = _mm512_min_epu8(min, maxLLR); // 128 in epi8 is -127
 	  fprintf(fd,"                min = _mm512_min_epu8(min, maxLLR);\n");
-	  //                *p_cnProcBufResBit = _mm512_sign_epi16(min, sgn);
+	  //                *p_cnProcBufResBit = _mm512_sign_epi8(min, sgn);
 	  //                p_cnProcBufResBit++;
-	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = _mm512_sign_epi16(min, sgn);\n",(lut_startAddrCnGroups[8]>>5)+(j*bitOffsetInGroup));
+	  fprintf(fd,"                ((__m512i*)cnProcBufRes)[%d+i] = conditional_negate(min, sgn,zeros);\n",(lut_startAddrCnGroups[8]>>6)+(j*bitOffsetInGroup));
 	  fprintf(fd,"            }\n");
         }
     }

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/generator_avx512/cnProc_gen_BG2_avx512.c

@@ -1,8 +1,7 @@
+#include <stdio.h>
 #include <stdint.h>
-#include <immintrin.h>
 #include "../../nrLDPCdecoder_defs.h"
-#include "../../nrLDPC_types.h"                                                                                            
+
-#include "../../nrLDPC_bnProc.h"
 
 void nrLDPC_cnProc_BG2_generator_AVX512(uint16_t Z,int R)
 {

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/generator_avx512/main.c

@@ -1,9 +1,6 @@
 
 #include <stdio.h>
-#include <immintrin.h>
+#include <stdint.h>
-#include "../../nrLDPC_types.h"
-#include "../../nrLDPC_init.h"
-#include "../../nrLDPC_bnProc.h"
 #define NB_Z  51
 void nrLDPC_cnProc_BG1_generator_AVX512(uint16_t,int);
 void nrLDPC_cnProc_BG2_generator_AVX512(uint16_t,int);

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z104_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z10_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z112_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z11_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z120_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z128_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z12_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z13_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z144_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z14_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z15_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z160_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z16_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z176_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z18_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z192_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z208_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z20_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z224_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z22_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z240_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z24_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z256_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z26_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z288_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z28_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z2_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z30_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z320_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z32_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z352_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z36_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z384_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z3_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z40_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z44_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z48_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1)); 	 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/nrLDPC_cnProc_BG2_Z4_13_AVX512.c

@@ -1,5 +1,6 @@
 #include <stdint.h>
 #include <immintrin.h>
+#include  "../include/avx512fintrin.h"
 __m512i _mm512_sign_epi16(__m512i a, __m512i b){ 
 b = _mm512_min_epi16(b, _mm512_set1_epi16(1)); 	 
 b = _mm512_max_epi16(b, _mm512_set1_epi16(-1));