degradation of performance in BLER corrected | bnProc & bnProcPc unrolled

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_mPass.h

-
 /*
  * Licensed to the OpenAirInterface (OAI) Software Alliance under one or more
  * contributor license agreements.  See the NOTICE file distributed with
@@ -22,7 +21,7 @@
 
 /*!\file nrLDPC_mPass.h
  * \brief Defines the functions for message passing
- * 
+ *
 */
 #ifndef __NR_LDPC_MPASS__H__
 #define __NR_LDPC_MPASS__H__
@@ -42,7 +41,7 @@
    \param Z Lifting size
    \param cshift Circular shift
 */
-//more faster memcpy by using "rep movsb", which on modern processors is highly optimized 
+//more faster memcpy by using "rep movsb", which on modern processors is highly optimized
 
 void *memcpy1(void *dst, const void *src, size_t n)
 {
@@ -177,8 +176,7 @@ static inline void nrLDPC_llr2CnProcBuf_BG1(t_nrLDPC_lut* p_lut, int8_t* llr, t_
 
     bitOffsetInGroup = lut_numCnInCnGroups_BG1_R13[0]*NR_LDPC_ZMAX;
 
-//      #pragma omp  simd
-//      #pragma omp parallel for schedule(dynamic)
+
 
     for (j=0; j<3; j++)
     {
@@ -1026,7 +1024,7 @@ static inline void nrLDPC_bn2cnProcBuf_BG1(t_nrLDPC_lut* p_lut, t_nrLDPC_procBuf
     // CN group with 3 BNs
 
     bitOffsetInGroup = lut_numCnInCnGroups_BG1_R13[0]*NR_LDPC_ZMAX;
-//    #pragma omp simd
+
     for (j=0;j<2; j++)
     {
         p_cnProcBuf = &cnProcBuf[lut_startAddrCnGroups[0] + j*bitOffsetInGroup];
@@ -1037,11 +1035,11 @@ static inline void nrLDPC_bn2cnProcBuf_BG1(t_nrLDPC_lut* p_lut, t_nrLDPC_procBuf
     // CN group with 4 BNs
 
     bitOffsetInGroup = lut_numCnInCnGroups_BG1_R13[1]*NR_LDPC_ZMAX;
-  // #pragma omp simd
+  
     for (j=0; j<3; j++)
     {
         p_cnProcBuf = &cnProcBuf[lut_startAddrCnGroups[1] + j*bitOffsetInGroup];
-  //      #pragma omp simd
+
         for (i=0; i<lut_numCnInCnGroups[1]; i++)
         {
             idxBn = lut_startAddrBnProcBuf_CNG4[j][i] + lut_bnPosBnProcBuf_CNG4[j][i]*Z;
@@ -1053,11 +1051,11 @@ static inline void nrLDPC_bn2cnProcBuf_BG1(t_nrLDPC_lut* p_lut, t_nrLDPC_procBuf
     // CN group with 5 BNs
 
     bitOffsetInGroup = lut_numCnInCnGroups_BG1_R13[2]*NR_LDPC_ZMAX;
-//    #pragma omp simd
+
     for (j=0; j<4; j++)
     {
         p_cnProcBuf = &cnProcBuf[lut_startAddrCnGroups[2] + j*bitOffsetInGroup];
-    //    #pragma omp simd
+    
         for (i=0; i<lut_numCnInCnGroups[2]; i++)
         {
             idxBn = lut_startAddrBnProcBuf_CNG5[j][i] + lut_bnPosBnProcBuf_CNG5[j][i]*Z;
@@ -1070,11 +1068,11 @@ static inline void nrLDPC_bn2cnProcBuf_BG1(t_nrLDPC_lut* p_lut, t_nrLDPC_procBuf
     // CN group with 6 BNs
 
     bitOffsetInGroup = lut_numCnInCnGroups_BG1_R13[3]*NR_LDPC_ZMAX;
-    //#pragma omp simd
+    
     for (j=0; j<5; j++)
     {
         p_cnProcBuf = &cnProcBuf[lut_startAddrCnGroups[3] + j*bitOffsetInGroup];
-      //  #pragma omp simd
+      
         for (i=0; i<lut_numCnInCnGroups[3]; i++)
         {
             idxBn = lut_startAddrBnProcBuf_CNG6[j][i] + lut_bnPosBnProcBuf_CNG6[j][i]*Z;
@@ -1087,12 +1085,12 @@ static inline void nrLDPC_bn2cnProcBuf_BG1(t_nrLDPC_lut* p_lut, t_nrLDPC_procBuf
     // CN group with 7 BNs
 
     bitOffsetInGroup = lut_numCnInCnGroups_BG1_R13[4]*NR_LDPC_ZMAX;
-    //#pragma omp simd
+    
     for (j=0; j<6; j++)
     {
         p_cnProcBuf = &cnProcBuf[lut_startAddrCnGroups[4] + j*bitOffsetInGroup];
 
-       // #pragma omp simd
+    
         for (i=0; i<lut_numCnInCnGroups[4]; i++)
         {
             idxBn = lut_startAddrBnProcBuf_CNG7[j][i] + lut_bnPosBnProcBuf_CNG7[j][i]*Z;
@@ -1105,7 +1103,7 @@ static inline void nrLDPC_bn2cnProcBuf_BG1(t_nrLDPC_lut* p_lut, t_nrLDPC_procBuf
     // CN group with 8 BNs
 
     bitOffsetInGroup = lut_numCnInCnGroups_BG1_R13[5]*NR_LDPC_ZMAX;
-//   #pragma omp simd
+
     for (j=0; j<7; j++)
     {
         p_cnProcBuf = &cnProcBuf[lut_startAddrCnGroups[5] + j*bitOffsetInGroup];
@@ -1121,11 +1119,11 @@ static inline void nrLDPC_bn2cnProcBuf_BG1(t_nrLDPC_lut* p_lut, t_nrLDPC_procBuf
     // CN group with 9 BNs
 
     bitOffsetInGroup = lut_numCnInCnGroups_BG1_R13[6]*NR_LDPC_ZMAX;
-   // #pragma omp simd
+
     for (j=0; j<8; j++)
     {
         p_cnProcBuf = &cnProcBuf[lut_startAddrCnGroups[6] + j*bitOffsetInGroup];
-  //      #pragma omp simd
+
         for (i=0; i<lut_numCnInCnGroups[6]; i++)
         {
             idxBn = lut_startAddrBnProcBuf_CNG9[j][i] + lut_bnPosBnProcBuf_CNG9[j][i]*Z;
@@ -1138,11 +1136,11 @@ static inline void nrLDPC_bn2cnProcBuf_BG1(t_nrLDPC_lut* p_lut, t_nrLDPC_procBuf
     // CN group with 10 BNs
 
     bitOffsetInGroup = lut_numCnInCnGroups_BG1_R13[7]*NR_LDPC_ZMAX;
-    //#pragma omp simd
+    
     for (j=0; j<9; j++)
     {
         p_cnProcBuf = &cnProcBuf[lut_startAddrCnGroups[7] + j*bitOffsetInGroup];
-    //    #pragma omp simd
+    
         for (i=0; i<lut_numCnInCnGroups[7]; i++)
         {
             idxBn = lut_startAddrBnProcBuf_CNG10[j][i] + lut_bnPosBnProcBuf_CNG10[j][i]*Z;
@@ -1155,11 +1153,11 @@ static inline void nrLDPC_bn2cnProcBuf_BG1(t_nrLDPC_lut* p_lut, t_nrLDPC_procBuf
     // CN group with 19 BNs
 
     bitOffsetInGroup = lut_numCnInCnGroups_BG1_R13[8]*NR_LDPC_ZMAX;
-   //#pragma omp simd
+   
     for (j=0; j<19; j++)
     {
         p_cnProcBuf = &cnProcBuf[lut_startAddrCnGroups[8] + j*bitOffsetInGroup];
-      //  #pragma omp simd
+      
         for (i=0; i<lut_numCnInCnGroups[8]; i++)
         {
             idxBn = lut_startAddrBnProcBuf_CNG19[j][i] + lut_bnPosBnProcBuf_CNG19[j][i]*Z;
@@ -1211,3 +1209,5 @@ static inline void nrLDPC_llrRes2llrOut(t_nrLDPC_lut* p_lut, int8_t* llrOut, t_n
 #endif
 
 
+
+

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

@@ -386,10 +386,10 @@ void nrLDPC_cnProc_BG1_generator_AVX2(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] = {{24,48,72,96,112,144,168}, {0,48,72,96,112,144,168},
-					 {0,24,72,96,112,144,168}, {0,24,48,96,112,144,168},
-					 {0,24,48,72,112,144,168}, {0,24,48,72,96,144,168},
-					 {0,24,48,72,96,112,168},  {0,24,48,72,96,112,144}};
+  const uint8_t lut_idxCnProcG8[8][7] = {{24,48,72,96,120,144,168}, {0,48,72,96,120,144,168},
+					 {0,24,72,96,120,144,168}, {0,24,48,96,120,144,168},
+					 {0,24,48,72,120,144,168}, {0,24,48,72,96,144,168},
+					 {0,24,48,72,96,120,168},  {0,24,48,72,96,120,144}};
 
 
 

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/generator_cnProc_avx512/cnProc_gen_BG1_avx512.c

-
-
 #include <stdio.h>
 #include <stdint.h>
 #include "../../nrLDPCdecoder_defs.h"
@@ -327,14 +325,13 @@ void nrLDPC_cnProc_BG1_generator_AVX512(int R)
   // =====================================================================
   // Process group with 8 BNs
   fprintf(fd,"//Process group with 8 BNs\n");
-    // Offset is 2*384/32 = 12
-  const uint8_t lut_idxCnProcG8[8][7] = {{24,48,72,96,112,144,168}, {0,48,72,96,112,144,168},
-                                         {0,24,72,96,112,144,168}, {0,24,48,96,112,144,168},
-                                         {0,24,48,72,112,144,168}, {0,24,48,72,96,144,168},
-                                         {0,24,48,72,96,112,168},  {0,24,48,72,96,112,144}};
-
-
+    // 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},
+                                           {0,24,72,96,120,144,168}, {0,24,48,96,120,144,168},
+                                           {0,24,48,72,120,144,168}, {0,24,48,72,96,144,168},
+                                           {0,24,48,72,96,120,168}, {0,24,48,72,96,120,144}};
 
+ 
   if (lut_numCnInCnGroups[5] > 0)
     {
       // Number of groups of 64  CNs for parallel processing

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_tools/ldpc_gen_files/bnProc/nrLDPC_bnProc_BG1_R89_AVX2.h

@@ -11,16 +11,12 @@ static inline void nrLDPC_bnProc_BG1_R89_AVX2(int8_t* bnProcBuf,int8_t* bnProcBu
             p_res = &p_bnProcBufRes[0];
             p_llrRes = (__m256i*) &llrRes  [384];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[0 + i]);
-             p_res++;
-             p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[0 + i]);
 }
             p_res = &p_bnProcBufRes[36];
             p_llrRes = (__m256i*) &llrRes  [384];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[36 + i]);
-             p_res++;
-             p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[36 + i]);
 }
 // Process group with 3 CNs 
        M = (21*Z + 31)>>5;
@@ -29,23 +25,17 @@ static inline void nrLDPC_bnProc_BG1_R89_AVX2(int8_t* bnProcBuf,int8_t* bnProcBu
             p_res = &p_bnProcBufRes[0];
             p_llrRes = (__m256i*) &llrRes  [1536];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[0 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[0 + i]);
 }
             p_res = &p_bnProcBufRes[252];
             p_llrRes = (__m256i*) &llrRes  [1536];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[252 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[252 + i]);
 }
             p_res = &p_bnProcBufRes[504];
             p_llrRes = (__m256i*) &llrRes  [1536];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[504 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[504 + i]);
 }
 // Process group with 4 CNs 
        M = (1*Z + 31)>>5;
@@ -54,30 +44,22 @@ static inline void nrLDPC_bnProc_BG1_R89_AVX2(int8_t* bnProcBuf,int8_t* bnProcBu
             p_res = &p_bnProcBufRes[0];
             p_llrRes = (__m256i*) &llrRes  [9600];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[0 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[0 + i]);
 }
             p_res = &p_bnProcBufRes[12];
             p_llrRes = (__m256i*) &llrRes  [9600];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[12 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[12 + i]);
 }
             p_res = &p_bnProcBufRes[24];
             p_llrRes = (__m256i*) &llrRes  [9600];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[24 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[24 + i]);
 }
             p_res = &p_bnProcBufRes[36];
             p_llrRes = (__m256i*) &llrRes  [9600];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[36 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[36 + i]);
 }
 // Process group with 5 CNs 
        M = (1*Z + 31)>>5;
@@ -86,37 +68,27 @@ static inline void nrLDPC_bnProc_BG1_R89_AVX2(int8_t* bnProcBuf,int8_t* bnProcBu
             p_res = &p_bnProcBufRes[0];
             p_llrRes = (__m256i*) &llrRes  [9984];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[0 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[0 + i]);
 }
             p_res = &p_bnProcBufRes[12];
             p_llrRes = (__m256i*) &llrRes  [9984];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[12 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[12 + i]);
 }
             p_res = &p_bnProcBufRes[24];
             p_llrRes = (__m256i*) &llrRes  [9984];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[24 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[24 + i]);
 }
             p_res = &p_bnProcBufRes[36];
             p_llrRes = (__m256i*) &llrRes  [9984];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[36 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[36 + i]);
 }
             p_res = &p_bnProcBufRes[48];
             p_llrRes = (__m256i*) &llrRes  [9984];
             for (i=0;i<M;i++) {
-            *p_res = _mm256_subs_epi8(*p_llrRes, p_bnProcBuf[48 + i]);
-            p_res++;
-            p_llrRes++;
+            p_res[i] = _mm256_subs_epi8(p_llrRes[i], p_bnProcBuf[48 + i]);
 }
 // Process group with 6 CNs 
 // Process group with 7 CNs