degradation of performance in BLER corrected | bnProc & bnProcPc unrolled
This commit is contained in:
Sy
@@ -1,4 +1,3 @@
|
||||
|
||||
/*
|
||||
* 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
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user