Skip to content

O
OpenXG-RAN
Commit de4712e7 authored by hbilel's avatar hbilel
Browse files
Options

UE add new dlsch thread

parent 2528ed13
Hide whitespace changes
Inline Side-by-side
Showing with 1428 additions and 0 deletions
openair1/PHY/LTE_TRANSPORT/dlsch_decoding.c
View file @ de4712e7
......@@ -36,6 +36,7 @@
#include "PHY/CODING/extern.h"
#include "SCHED/extern.h"
#include "SIMULATION/TOOLS/defs.h"
#include "targets/RT/USER/lte-softmodem.h"
//#define DEBUG_DLSCH_DECODING
extern double cpuf;
......@@ -758,6 +759,1332 @@ uint32_t dlsch_decoding(PHY_VARS_UE *phy_vars_ue,
return(ret);
}
#ifdef UE_DLSCH_PARALLELISATION
uint32_t dlsch_decoding_mthread(PHY_VARS_UE *phy_vars_ue,
UE_rxtx_proc_t *proc,
int eNB_id,
short *dlsch_llr,
LTE_DL_FRAME_PARMS *frame_parms,
LTE_UE_DLSCH_t *dlsch,
LTE_DL_UE_HARQ_t *harq_process,
uint32_t frame,
uint8_t nr_tti_rx,
uint8_t harq_pid,
uint8_t is_crnti,
uint8_t llr8_flag)
{
#if UE_TIMING_TRACE
time_stats_t *dlsch_rate_unmatching_stats=&phy_vars_ue->dlsch_rate_unmatching_stats;
time_stats_t *dlsch_turbo_decoding_stats=&phy_vars_ue->dlsch_turbo_decoding_stats;
time_stats_t *dlsch_deinterleaving_stats=&phy_vars_ue->dlsch_deinterleaving_stats;
#endif
uint32_t A,E;
uint32_t G;
uint32_t ret,offset;
uint16_t iind;
// uint8_t dummy_channel_output[(3*8*block_length)+12];
short dummy_w[MAX_NUM_DLSCH_SEGMENTS][3*(6144+64)];
uint32_t r,r_offset=0,Kr,Kr_bytes,err_flag=0;
uint8_t crc_type;
//UE_rxtx_proc_t *proc = &phy_vars_ue->proc;
int Cby2;
#ifdef DEBUG_DLSCH_DECODING
uint16_t i;
#endif
//#ifdef __AVX2__
#if 0
int Kr_last,skipped_last=0;
uint8_t (*tc_2cw)(int16_t *y,
int16_t *y2,
uint8_t *,
uint8_t *,
uint16_t,
uint16_t,
uint16_t,
uint8_t,
uint8_t,
uint8_t,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *);
#endif
uint8_t (*tc)(int16_t *y,
uint8_t *,
uint16_t,
uint16_t,
uint16_t,
uint8_t,
uint8_t,
uint8_t,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *);
if (!dlsch_llr) {
printf("dlsch_decoding.c: NULL dlsch_llr pointer\n");
return(dlsch->max_turbo_iterations);
}
if (!harq_process) {
printf("dlsch_decoding.c: NULL harq_process pointer\n");
return(dlsch->max_turbo_iterations);
}
if (!frame_parms) {
printf("dlsch_decoding.c: NULL frame_parms pointer\n");
return(dlsch->max_turbo_iterations);
}
if (nr_tti_rx> (10*frame_parms->ttis_per_subframe-1)) {
printf("dlsch_decoding.c: Illegal subframe index %d\n",nr_tti_rx);
return(dlsch->max_turbo_iterations);
}
if (dlsch->harq_ack[nr_tti_rx].ack != 2) {
LOG_D(PHY, "[UE %d] DLSCH @ SF%d : ACK bit is %d instead of DTX even before PDSCH is decoded!\n",
phy_vars_ue->Mod_id, nr_tti_rx, dlsch->harq_ack[nr_tti_rx].ack);
}
if (llr8_flag == 0) {
//#ifdef __AVX2__
#if 0
tc_2cw = phy_threegpplte_turbo_decoder16avx2;
#endif
tc = phy_threegpplte_turbo_decoder16;
}
else
{
AssertFatal (harq_process->TBS >= 256 , "Mismatch flag nbRB=%d TBS=%d mcs=%d Qm=%d RIV=%d round=%d \n",
harq_process->nb_rb, harq_process->TBS,harq_process->mcs,harq_process->Qm,harq_process->rvidx,harq_process->round);
tc = phy_threegpplte_turbo_decoder8;
}
// nb_rb = dlsch->nb_rb;
/*
if (nb_rb > frame_parms->N_RB_DL) {
printf("dlsch_decoding.c: Illegal nb_rb %d\n",nb_rb);
return(max_turbo_iterations);
}*/
/*harq_pid = dlsch->current_harq_pid[phy_vars_ue->current_thread_id[subframe]];
if (harq_pid >= 8) {
printf("dlsch_decoding.c: Illegal harq_pid %d\n",harq_pid);
return(max_turbo_iterations);
}
*/
harq_process->trials[harq_process->round]++;
A = harq_process->TBS; //2072 for QPSK 1/3
ret = dlsch->max_turbo_iterations;
G = harq_process->G;
//get_G(frame_parms,nb_rb,dlsch->rb_alloc,mod_order,num_pdcch_symbols,phy_vars_ue->frame,subframe);
// printf("DLSCH Decoding, harq_pid %d Ndi %d\n",harq_pid,harq_process->Ndi);
if (harq_process->round == 0) {
// This is a new packet, so compute quantities regarding segmentation
harq_process->B = A+24;
lte_segmentation(NULL,
NULL,
harq_process->B,
&harq_process->C,
&harq_process->Cplus,
&harq_process->Cminus,
&harq_process->Kplus,
&harq_process->Kminus,
&harq_process->F);
// CLEAR LLR's HERE for first packet in process
}
/*
else {
printf("dlsch_decoding.c: Ndi>0 not checked yet!!\n");
return(max_turbo_iterations);
}
*/
err_flag = 0;
r_offset = 0;
unsigned char bw_scaling =1;
switch (frame_parms->N_RB_DL) {
case 6:
bw_scaling =16;
break;
case 25:
bw_scaling =4;
break;
case 50:
bw_scaling =2;
break;
default:
bw_scaling =1;
break;
}
if (harq_process->C > MAX_NUM_DLSCH_SEGMENTS/bw_scaling) {
LOG_E(PHY,"Illegal harq_process->C %d > %d\n",harq_process->C,MAX_NUM_DLSCH_SEGMENTS/bw_scaling);
return((1+dlsch->max_turbo_iterations));
}
#ifdef DEBUG_DLSCH_DECODING
printf("Segmentation: C %d, Cminus %d, Kminus %d, Kplus %d\n",harq_process->C,harq_process->Cminus,harq_process->Kminus,harq_process->Kplus);
#endif
opp_enabled=1;
if (harq_process->C>1) { // wakeup worker if more than 1 segment
if (pthread_mutex_lock(&proc->mutex_dlsch_td) != 0) {
LOG_E( PHY, "[SCHED][UE %d][Slot0] error locking mutex for UE dlsch td\n",phy_vars_ue->Mod_id );
exit_fun("nothing to add");
}
proc->instance_cnt_dlsch_td++;
proc->eNB_id = eNB_id;
proc->harq_pid = harq_pid;
proc->llr8_flag = llr8_flag;
if (proc->instance_cnt_dlsch_td == 0)
{
LOG_D(PHY,"unblock slot1 dl processing thread blocked on instance_cnt_dlsch_td : %d \n", proc->instance_cnt_dlsch_td );
if (pthread_cond_signal(&proc->cond_dlsch_td) != 0) {
LOG_E( PHY, "[SCHED][UE %d][Slot0] ERROR pthread_cond_signal for UE dlsch td\n", phy_vars_ue->Mod_id);
exit_fun("nothing to add");
}
if (pthread_mutex_unlock(&proc->mutex_dlsch_td) != 0) {
LOG_E( PHY, "[SCHED][UE %d][Slot0] error unlocking mutex for UE dlsch td \n",phy_vars_ue->Mod_id );
exit_fun("nothing to add");
}
} else
{
LOG_E( PHY, "[SCHED][UE %d] UE dlsch td thread busy (IC %d)!!\n", phy_vars_ue->Mod_id, proc->instance_cnt_dlsch_td);
if (proc->instance_cnt_dlsch_td > 2)
exit_fun("instance_cnt_dlsch_td > 2");
}
//AssertFatal(pthread_cond_signal(&proc->cond_slot1_dl_processing) ==0 ,"");
AssertFatal(pthread_mutex_unlock(&proc->mutex_dlsch_td) ==0,"");
/*
if (pthread_mutex_timedlock(&proc->mutex_td,&wait) != 0) {
printf("[eNB] ERROR pthread_mutex_lock for TD thread (IC %d)\n", proc->instance_cnt_td);
exit_fun( "error locking mutex_fep" );
return -1;
}
if (proc->instance_cnt_td==0) {
printf("[UE] TD thread busy\n");
exit_fun("TD thread busy");
pthread_mutex_unlock( &proc->mutex_td );
return -1;
}
++proc->instance_cnt_td;
proc->tdp.UE = phy_vars_ue;
proc->tdp.eNB_id = eNB_id;
proc->tdp.harq_pid = harq_pid;
proc->tdp.llr8_flag = llr8_flag;
printf("----- 2thread llr flag %d tdp flag %d\n",llr8_flag, proc->tdp.llr8_flag);
// wakeup worker to do second half segments
if (pthread_cond_signal(&proc->cond_td) != 0) {
printf("[UE] ERROR pthread_cond_signal for td thread exit\n");
exit_fun( "ERROR pthread_cond_signal" );
return (1+dlsch->last_iteration_cnt);
}
pthread_mutex_unlock( &proc->mutex_td );*/
Cby2 = harq_process->C/2;
}
else {
Cby2 = 1;
}
for (r=0; r<Cby2; r++) {
// Get Turbo interleaver parameters
if (r<harq_process->Cminus)
Kr = harq_process->Kminus;
else
Kr = harq_process->Kplus;
Kr_bytes = Kr>>3;
if (Kr_bytes<=64)
iind = (Kr_bytes-5);
else if (Kr_bytes <=128)
iind = 59 + ((Kr_bytes-64)>>1);
else if (Kr_bytes <= 256)
iind = 91 + ((Kr_bytes-128)>>2);
else if (Kr_bytes <= 768)
iind = 123 + ((Kr_bytes-256)>>3);
else {
printf("dlsch_decoding: Illegal codeword size %d!!!\n",Kr_bytes);
return(dlsch->max_turbo_iterations);
}
#ifdef DEBUG_DLSCH_DECODING
printf("f1 %d, f2 %d, F %d\n",f1f2mat_old[2*iind],f1f2mat_old[1+(2*iind)],(r==0) ? harq_process->F : 0);
#endif
#if UE_TIMING_TRACE
start_meas(dlsch_rate_unmatching_stats);
#endif
memset(&dummy_w[r][0],0,3*(6144+64)*sizeof(short));
harq_process->RTC[r] = generate_dummy_w(4+(Kr_bytes*8),
(uint8_t*) &dummy_w[r][0],
(r==0) ? harq_process->F : 0);
#ifdef DEBUG_DLSCH_DECODING
LOG_D(PHY,"HARQ_PID %d Rate Matching Segment %d (coded bits %d,unpunctured/repeated bits %d, TBS %d, mod_order %d, nb_rb %d, Nl %d, rv %d, round %d)...\n",
harq_pid,r, G,
Kr*3,
harq_process->TBS,
harq_process->Qm,
harq_process->nb_rb,
harq_process->Nl,
harq_process->rvidx,
harq_process->round);
#endif
#ifdef DEBUG_DLSCH_DECODING
printf(" in decoding dlsch->harq_processes[harq_pid]->rvidx = %d\n", dlsch->harq_processes[harq_pid]->rvidx);
#endif
if (lte_rate_matching_turbo_rx(harq_process->RTC[r],
G,
harq_process->w[r],
(uint8_t*)&dummy_w[r][0],
dlsch_llr+r_offset,
harq_process->C,
dlsch->Nsoft,
dlsch->Mdlharq,
dlsch->Kmimo,
harq_process->rvidx,
(harq_process->round==0)?1:0,
harq_process->Qm,
harq_process->Nl,
r,
&E)==-1) {
#if UE_TIMING_TRACE
stop_meas(dlsch_rate_unmatching_stats);
#endif
LOG_E(PHY,"dlsch_decoding.c: Problem in rate_matching\n");
return(dlsch->max_turbo_iterations);
} else
{
#if UE_TIMING_TRACE
stop_meas(dlsch_rate_unmatching_stats);
#endif
}
r_offset += E;
/*
printf("Subblock deinterleaving, d %p w %p\n",
harq_process->d[r],
harq_process->w);
*/
#if UE_TIMING_TRACE
start_meas(dlsch_deinterleaving_stats);
#endif
sub_block_deinterleaving_turbo(4+Kr,
&harq_process->d[r][96],
harq_process->w[r]);
#if UE_TIMING_TRACE
stop_meas(dlsch_deinterleaving_stats);
#endif
#ifdef DEBUG_DLSCH_DECODING
/*
if (r==0) {
write_output("decoder_llr.m","decllr",dlsch_llr,G,1,0);
write_output("decoder_in.m","dec",&harq_process->d[0][96],(3*8*Kr_bytes)+12,1,0);
}
printf("decoder input(segment %d) :",r);
int i; for (i=0;i<(3*8*Kr_bytes)+12;i++)
printf("%d : %d\n",i,harq_process->d[r][96+i]);
printf("\n");*/
#endif
// printf("Clearing c, %p\n",harq_process->c[r]);
memset(harq_process->c[r],0,Kr_bytes);
// printf("done\n");
if (harq_process->C == 1)
crc_type = CRC24_A;
else
crc_type = CRC24_B;
/*
printf("decoder input(segment %d)\n",r);
for (i=0;i<(3*8*Kr_bytes)+12;i++)
if ((harq_process->d[r][96+i]>7) ||
(harq_process->d[r][96+i] < -8))
printf("%d : %d\n",i,harq_process->d[r][96+i]);
printf("\n");
*/
//#ifndef __AVX2__
#if 1
if (err_flag == 0) {
/*
LOG_I(PHY, "turbo algo Kr=%d cb_cnt=%d C=%d nbRB=%d crc_type %d TBSInput=%d TBSHarq=%d TBSplus24=%d mcs=%d Qm=%d RIV=%d round=%d maxIter %d\n",
Kr,r,harq_process->C,harq_process->nb_rb,crc_type,A,harq_process->TBS,
harq_process->B,harq_process->mcs,harq_process->Qm,harq_process->rvidx,harq_process->round,dlsch->max_turbo_iterations);
*/
if (llr8_flag) {
AssertFatal (Kr >= 256, "turbo algo issue Kr=%d cb_cnt=%d C=%d nbRB=%d TBSInput=%d TBSHarq=%d TBSplus24=%d mcs=%d Qm=%d RIV=%d round=%d\n",
Kr,r,harq_process->C,harq_process->nb_rb,A,harq_process->TBS,harq_process->B,harq_process->mcs,harq_process->Qm,harq_process->rvidx,harq_process->round);
}
#if UE_TIMING_TRACE
start_meas(dlsch_turbo_decoding_stats);
#endif
LOG_D(PHY,"AbsSubframe %d.%d Start turbo segment %d/%d \n",frame%1024,nr_tti_rx,r,harq_process->C-1);
ret = tc
(&harq_process->d[r][96],
harq_process->c[r],
Kr,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
#if UE_TIMING_TRACE
stop_meas(dlsch_turbo_decoding_stats);
#endif
}
#else
if ((harq_process->C == 1) ||
((r==harq_process->C-1) && (skipped_last==0))) { // last segment with odd number of segments
#if UE_TIMING_TRACE
start_meas(dlsch_turbo_decoding_stats);
#endif
ret = tc
(&harq_process->d[r][96],
harq_process->c[r],
Kr,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
#if UE_TIMING_TRACE
stop_meas(dlsch_turbo_decoding_stats);
#endif
// printf("single decode, exit\n");
// exit(-1);
}
else {
// we can merge code segments
if ((skipped_last == 0) && (r<harq_process->C-1)) {
skipped_last = 1;
Kr_last = Kr;
}
else {
skipped_last=0;
if (Kr_last == Kr) { // decode 2 code segments with AVX2 version
#ifdef DEBUG_DLSCH_DECODING
printf("single decoding segment %d (%p)\n",r-1,&harq_process->d[r-1][96]);
#endif
#if UE_TIMING_TRACE
start_meas(dlsch_turbo_decoding_stats);
#endif
#ifdef DEBUG_DLSCH_DECODING
printf("double decoding segments %d,%d (%p,%p)\n",r-1,r,&harq_process->d[r-1][96],&harq_process->d[r][96]);
#endif
ret = tc_2cw
(&harq_process->d[r-1][96],
&harq_process->d[r][96],
harq_process->c[r-1],
harq_process->c[r],
Kr,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
/*
ret = tc
(&harq_process->d[r-1][96],
harq_process->c[r-1],
Kr_last,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
exit(-1);*/
#if UE_TIMING_TRACE
stop_meas(dlsch_turbo_decoding_stats);
#endif
}
else { // Kr_last != Kr
#if UE_TIMING_TRACE
start_meas(dlsch_turbo_decoding_stats);
#endif
ret = tc
(&harq_process->d[r-1][96],
harq_process->c[r-1],
Kr_last,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
#if UE_TIMING_TRACE
stop_meas(dlsch_turbo_decoding_stats);
start_meas(dlsch_turbo_decoding_stats);
#endif
ret = tc
(&harq_process->d[r][96],
harq_process->c[r],
Kr,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
#if UE_TIMING_TRACE
stop_meas(dlsch_turbo_decoding_stats);
/*printf("Segmentation: C %d r %d, dlsch_rate_unmatching_stats %5.3f dlsch_deinterleaving_stats %5.3f dlsch_turbo_decoding_stats %5.3f \n",
harq_process->C,
r,
dlsch_rate_unmatching_stats->p_time/(cpuf*1000.0),
dlsch_deinterleaving_stats->p_time/(cpuf*1000.0),
dlsch_turbo_decoding_stats->p_time/(cpuf*1000.0));*/
#endif
}
}
}
#endif
if ((err_flag == 0) && (ret>=(1+dlsch->max_turbo_iterations))) {// a Code segment is in error so break;
LOG_D(PHY,"AbsSubframe %d.%d CRC failed, segment %d/%d \n",frame%1024,nr_tti_rx,r,harq_process->C-1);
err_flag = 1;
}
}
int32_t frame_rx_prev = frame;
int32_t tti_rx_prev = nr_tti_rx - 1;
if (tti_rx_prev < 0) {
frame_rx_prev--;
tti_rx_prev += 10*frame_parms->ttis_per_subframe;
}
frame_rx_prev = frame_rx_prev%1024;
if (err_flag == 1) {
#if UE_DEBUG_TRACE
LOG_I(PHY,"[UE %d] DLSCH: Setting NAK for SFN/SF %d/%d (pid %d, status %d, round %d, TBS %d, mcs %d) Kr %d r %d harq_process->round %d\n",
phy_vars_ue->Mod_id, frame, nr_tti_rx, harq_pid,harq_process->status, harq_process->round,harq_process->TBS,harq_process->mcs,Kr,r,harq_process->round);
#endif
dlsch->harq_ack[nr_tti_rx].ack = 0;
dlsch->harq_ack[nr_tti_rx].harq_id = harq_pid;
dlsch->harq_ack[nr_tti_rx].send_harq_status = 1;
harq_process->errors[harq_process->round]++;
harq_process->round++;
// printf("Rate: [UE %d] DLSCH: Setting NACK for subframe %d (pid %d, round %d)\n",phy_vars_ue->Mod_id,subframe,harq_pid,harq_process->round);
if (harq_process->round >= dlsch->Mdlharq) {
harq_process->status = SCH_IDLE;
harq_process->round = 0;
}
if(is_crnti)
{
LOG_D(PHY,"[UE %d] DLSCH: Setting NACK for nr_tti_rx %d (pid %d, pid status %d, round %d/Max %d, TBS %d)\n",
phy_vars_ue->Mod_id,nr_tti_rx,harq_pid,harq_process->status,harq_process->round,dlsch->Mdlharq,harq_process->TBS);
}
return((1+dlsch->max_turbo_iterations));
} else {
#if UE_DEBUG_TRACE
LOG_I(PHY,"[UE %d] DLSCH: Setting ACK for nr_tti_rx %d TBS %d mcs %d nb_rb %d\n",
phy_vars_ue->Mod_id,nr_tti_rx,harq_process->TBS,harq_process->mcs,harq_process->nb_rb);
#endif
harq_process->status = SCH_IDLE;
harq_process->round = 0;
dlsch->harq_ack[nr_tti_rx].ack = 1;
dlsch->harq_ack[nr_tti_rx].harq_id = harq_pid;
dlsch->harq_ack[nr_tti_rx].send_harq_status = 1;
//LOG_I(PHY,"[UE %d] DLSCH: Setting ACK for SFN/SF %d/%d (pid %d, status %d, round %d, TBS %d, mcs %d)\n",
// phy_vars_ue->Mod_id, frame, subframe, harq_pid, harq_process->status, harq_process->round,harq_process->TBS,harq_process->mcs);
if(is_crnti)
{
LOG_D(PHY,"[UE %d] DLSCH: Setting ACK for nr_tti_rx %d (pid %d, round %d, TBS %d)\n",phy_vars_ue->Mod_id,nr_tti_rx,harq_pid,harq_process->round,harq_process->TBS);
}
//LOG_D(PHY,"[UE %d] DLSCH: Setting ACK for subframe %d (pid %d, round %d)\n",phy_vars_ue->Mod_id,subframe,harq_pid,harq_process->round);
}
// Reassembly of Transport block here
offset = 0;
/*
printf("harq_pid %d\n",harq_pid);
printf("F %d, Fbytes %d\n",harq_process->F,harq_process->F>>3);
printf("C %d\n",harq_process->C);
*/
for (r=0; r<harq_process->C; r++) {
if (r<harq_process->Cminus)
Kr = harq_process->Kminus;
else
Kr = harq_process->Kplus;
Kr_bytes = Kr>>3;
// printf("Segment %d : Kr= %d bytes\n",r,Kr_bytes);
if (r==0) {
memcpy(harq_process->b,
&harq_process->c[0][(harq_process->F>>3)],
Kr_bytes - (harq_process->F>>3)- ((harq_process->C>1)?3:0));
offset = Kr_bytes - (harq_process->F>>3) - ((harq_process->C>1)?3:0);
// printf("copied %d bytes to b sequence (harq_pid %d)\n",
// Kr_bytes - (harq_process->F>>3),harq_pid);
// printf("b[0] = %x,c[%d] = %x\n",
// harq_process->b[0],
// harq_process->F>>3,
// harq_process->c[0][(harq_process->F>>3)]);
} else {
memcpy(harq_process->b+offset,
harq_process->c[r],
Kr_bytes- ((harq_process->C>1)?3:0));
offset += (Kr_bytes - ((harq_process->C>1)?3:0));
}
}
dlsch->last_iteration_cnt = ret;
//wait for worker to finish
//wait_on_busy_condition(&proc->mutex_td,&proc->cond_td,&proc->instance_cnt_dlsch td,"dlsch td thread");
//return( (ret>proc->tdp.ret) ? ret : proc->tdp.ret );
return(ret);
}
#endif
#ifdef UE_DLSCH_PARALLELISATION
#define FIFO_PRIORITY 39
uint32_t dlsch_decoding_2thread0(void *arg)
{
static __thread int UE_dlsch_td_retval;
struct rx_tx_thread_data *rtd = arg;
UE_rxtx_proc_t *proc = rtd->proc;
PHY_VARS_UE *phy_vars_ue = rtd->UE;
proc->instance_cnt_dlsch_td=-1;
proc->nr_tti_rx=proc->sub_frame_start;
char threadname[256];
sprintf(threadname,"UE_thread_dlsch_td_%d", proc->sub_frame_start);
cpu_set_t cpuset;
CPU_ZERO(&cpuset);
if ( (proc->sub_frame_start+1)%RX_NB_TH == 0 && threads.dlsch_td_one != -1 )
CPU_SET(threads.dlsch_td_one, &cpuset);
if ( (proc->sub_frame_start+1)%RX_NB_TH == 1 && threads.dlsch_td_two != -1 )
CPU_SET(threads.dlsch_td_two, &cpuset);
if ( (proc->sub_frame_start+1)%RX_NB_TH == 2 && threads.dlsch_td_three != -1 )
CPU_SET(threads.dlsch_td_three, &cpuset);
//PHY_VARS_UE *phy_vars_ue = tdp->UE;
int eNB_id = proc->eNB_id;
int harq_pid = proc->harq_pid;
int llr8_flag1 = proc->llr8_flag;
//UE_rxtx_proc_t *proc = tdp->proc;
int frame = proc->frame_rx;
int subframe = proc->subframe_rx;
LTE_UE_DLSCH_t *dlsch = phy_vars_ue->dlsch[phy_vars_ue->current_thread_id[subframe]][eNB_id][0];
LTE_DL_UE_HARQ_t *harq_process = dlsch->harq_processes[harq_pid];
short *dlsch_llr = phy_vars_ue->pdsch_vars[phy_vars_ue->current_thread_id[subframe]][eNB_id]->llr[0];
//printf("2thread0 llr flag %d tdp flag %d\n",llr8_flag1, tdp->llr8_flag);
#if UE_TIMING_TRACE
time_stats_t *dlsch_rate_unmatching_stats=&phy_vars_ue->dlsch_rate_unmatching_stats;
time_stats_t *dlsch_turbo_decoding_stats=&phy_vars_ue->dlsch_turbo_decoding_stats;
time_stats_t *dlsch_deinterleaving_stats=&phy_vars_ue->dlsch_deinterleaving_stats;
#endif
uint32_t A,E;
uint32_t G;
uint32_t ret,offset;
uint16_t iind;
// uint8_t dummy_channel_output[(3*8*block_length)+12];
short dummy_w[MAX_NUM_DLSCH_SEGMENTS][3*(6144+64)];
uint32_t r,r_offset=0,Kr,Kr_bytes,err_flag=0;
uint8_t crc_type;
#ifdef DEBUG_DLSCH_DECODING
uint16_t i;
#endif
//#ifdef __AVX2__
uint8_t (*tc)(int16_t *y,
uint8_t *,
uint16_t,
uint16_t,
uint16_t,
uint8_t,
uint8_t,
uint8_t,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *,
time_stats_t *);
if (llr8_flag1 == 0) {
//#ifdef __AVX2__
#if 0
tc_2cw = phy_threegpplte_turbo_decoder16avx2;
#endif
tc = phy_threegpplte_turbo_decoder16;
}
else
{
//AssertFatal (harq_process->TBS >= 256 , "Mismatch flag nbRB=%d TBS=%d mcs=%d Qm=%d RIV=%d round=%d \n",
// harq_process->nb_rb, harq_process->TBS,harq_process->mcs,harq_process->Qm,harq_process->rvidx,harq_process->round);
tc = phy_threegpplte_turbo_decoder8;
}
init_thread(900000,1000000 , FIFO_PRIORITY-1, &cpuset, threadname);
while (!oai_exit) {
if (pthread_mutex_lock(&proc->mutex_dlsch_td) != 0) {
LOG_E( PHY, "[SCHED][UE] error locking mutex for UE dlsch td\n" );
exit_fun("nothing to add");
}
while (proc->instance_cnt_slot1_dl_processing < 0) {
// most of the time, the thread is waiting here
pthread_cond_wait( &proc->cond_dlsch_td, &proc->mutex_dlsch_td );
}
if (pthread_mutex_unlock(&proc->mutex_dlsch_td) != 0) {
LOG_E( PHY, "[SCHED][UE] error unlocking mutex for UE dlsch_td \n" );
exit_fun("nothing to add");
}
// nb_rb = dlsch->nb_rb;
/*
if (nb_rb > frame_parms->N_RB_DL) {
printf("dlsch_decoding.c: Illegal nb_rb %d\n",nb_rb);
return(max_turbo_iterations);
}*/
/*harq_pid = dlsch->current_harq_pid[phy_vars_ue->current_thread_id[subframe]];
if (harq_pid >= 8) {
printf("dlsch_decoding.c: Illegal harq_pid %d\n",harq_pid);
return(max_turbo_iterations);
}
*/
harq_process->trials[harq_process->round]++;
A = harq_process->TBS; //2072 for QPSK 1/3
ret = dlsch->max_turbo_iterations;
G = harq_process->G;
//get_G(frame_parms,nb_rb,dlsch->rb_alloc,mod_order,num_pdcch_symbols,phy_vars_ue->frame,subframe);
// printf("DLSCH Decoding, harq_pid %d Ndi %d\n",harq_pid,harq_process->Ndi);
if (harq_process->round == 0) {
// This is a new packet, so compute quantities regarding segmentation
harq_process->B = A+24;
lte_segmentation(NULL,
NULL,
harq_process->B,
&harq_process->C,
&harq_process->Cplus,
&harq_process->Cminus,
&harq_process->Kplus,
&harq_process->Kminus,
&harq_process->F);
// CLEAR LLR's HERE for first packet in process
}
/*
else {
printf("dlsch_decoding.c: Ndi>0 not checked yet!!\n");
return(max_turbo_iterations);
}
*/
err_flag = 0;
r_offset = 0;
/*
unsigned char bw_scaling =1;
switch (frame_parms->N_RB_DL) {
case 6:
bw_scaling =16;
break;
case 25:
bw_scaling =4;
break;
case 50:
bw_scaling =2;
break;
default:
bw_scaling =1;
break;
}
if (harq_process->C > MAX_NUM_DLSCH_SEGMENTS/bw_scaling) {
LOG_E(PHY,"Illegal harq_process->C %d > %d\n",harq_process->C,MAX_NUM_DLSCH_SEGMENTS/bw_scaling);
return((1+dlsch->max_turbo_iterations));
}*/
#ifdef DEBUG_DLSCH_DECODING
printf("Segmentation: C %d, Cminus %d, Kminus %d, Kplus %d\n",harq_process->C,harq_process->Cminus,harq_process->Kminus,harq_process->Kplus);
#endif
opp_enabled=1;
for (r=(harq_process->C)/2; r<harq_process->C; r++) {
// Get Turbo interleaver parameters
if (r<harq_process->Cminus)
Kr = harq_process->Kminus;
else
Kr = harq_process->Kplus;
Kr_bytes = Kr>>3;
if (Kr_bytes<=64)
iind = (Kr_bytes-5);
else if (Kr_bytes <=128)
iind = 59 + ((Kr_bytes-64)>>1);
else if (Kr_bytes <= 256)
iind = 91 + ((Kr_bytes-128)>>2);
else if (Kr_bytes <= 768)
iind = 123 + ((Kr_bytes-256)>>3);
else {
printf("dlsch_decoding: Illegal codeword size %d!!!\n",Kr_bytes);
return(dlsch->max_turbo_iterations);
}
#ifdef DEBUG_DLSCH_DECODING
printf("f1 %d, f2 %d, F %d\n",f1f2mat_old[2*iind],f1f2mat_old[1+(2*iind)],(r==0) ? harq_process->F : 0);
#endif
#if UE_TIMING_TRACE
start_meas(dlsch_rate_unmatching_stats);
#endif
memset(&dummy_w[r][0],0,3*(6144+64)*sizeof(short));
harq_process->RTC[r] = generate_dummy_w(4+(Kr_bytes*8),
(uint8_t*) &dummy_w[r][0],
(r==0) ? harq_process->F : 0);
#ifdef DEBUG_DLSCH_DECODING
LOG_D(PHY,"HARQ_PID %d Rate Matching Segment %d (coded bits %d,unpunctured/repeated bits %d, TBS %d, mod_order %d, nb_rb %d, Nl %d, rv %d, round %d)...\n",
harq_pid,r, G,
Kr*3,
harq_process->TBS,
harq_process->Qm,
harq_process->nb_rb,
harq_process->Nl,
harq_process->rvidx,
harq_process->round);
#endif
#ifdef DEBUG_DLSCH_DECODING
printf(" in decoding dlsch->harq_processes[harq_pid]->rvidx = %d\n", dlsch->harq_processes[harq_pid]->rvidx);
#endif
if (lte_rate_matching_turbo_rx(harq_process->RTC[r],
G,
harq_process->w[r],
(uint8_t*)&dummy_w[r][0],
dlsch_llr+r_offset,
harq_process->C,
dlsch->Nsoft,
dlsch->Mdlharq,
dlsch->Kmimo,
harq_process->rvidx,
(harq_process->round==0)?1:0,
harq_process->Qm,
harq_process->Nl,
r,
&E)==-1) {
#if UE_TIMING_TRACE
stop_meas(dlsch_rate_unmatching_stats);
#endif
LOG_E(PHY,"dlsch_decoding.c: Problem in rate_matching\n");
return(dlsch->max_turbo_iterations);
} else
{
#if UE_TIMING_TRACE
stop_meas(dlsch_rate_unmatching_stats);
#endif
}
r_offset += E;
/*
printf("Subblock deinterleaving, d %p w %p\n",
harq_process->d[r],
harq_process->w);
*/
#if UE_TIMING_TRACE
start_meas(dlsch_deinterleaving_stats);
#endif
sub_block_deinterleaving_turbo(4+Kr,
&harq_process->d[r][96],
harq_process->w[r]);
#if UE_TIMING_TRACE
stop_meas(dlsch_deinterleaving_stats);
#endif
#ifdef DEBUG_DLSCH_DECODING
/*
if (r==0) {
write_output("decoder_llr.m","decllr",dlsch_llr,G,1,0);
write_output("decoder_in.m","dec",&harq_process->d[0][96],(3*8*Kr_bytes)+12,1,0);
}
printf("decoder input(segment %d) :",r);
int i; for (i=0;i<(3*8*Kr_bytes)+12;i++)
printf("%d : %d\n",i,harq_process->d[r][96+i]);
printf("\n");*/
#endif
// printf("Clearing c, %p\n",harq_process->c[r]);
memset(harq_process->c[r],0,Kr_bytes);
// printf("done\n");
if (harq_process->C == 1)
crc_type = CRC24_A;
else
crc_type = CRC24_B;
/*
printf("decoder input(segment %d)\n",r);
for (i=0;i<(3*8*Kr_bytes)+12;i++)
if ((harq_process->d[r][96+i]>7) ||
(harq_process->d[r][96+i] < -8))
printf("%d : %d\n",i,harq_process->d[r][96+i]);
printf("\n");
*/
//#ifndef __AVX2__
#if 1
if (err_flag == 0) {
/*
LOG_I(PHY, "turbo algo Kr=%d cb_cnt=%d C=%d nbRB=%d crc_type %d TBSInput=%d TBSHarq=%d TBSplus24=%d mcs=%d Qm=%d RIV=%d round=%d maxIter %d\n",
Kr,r,harq_process->C,harq_process->nb_rb,crc_type,A,harq_process->TBS,
harq_process->B,harq_process->mcs,harq_process->Qm,harq_process->rvidx,harq_process->round,dlsch->max_turbo_iterations);
*/
if (llr8_flag1) {
AssertFatal (Kr >= 256, "turbo algo issue Kr=%d cb_cnt=%d C=%d nbRB=%d TBSInput=%d TBSHarq=%d TBSplus24=%d mcs=%d Qm=%d RIV=%d round=%d\n",
Kr,r,harq_process->C,harq_process->nb_rb,A,harq_process->TBS,harq_process->B,harq_process->mcs,harq_process->Qm,harq_process->rvidx,harq_process->round);
}
#if UE_TIMING_TRACE
start_meas(dlsch_turbo_decoding_stats);
#endif
// LOG_D(PHY,"AbsSubframe %d.%d Start turbo segment %d/%d \n",frame%1024,subframe,r,harq_process->C-1);
ret = tc
(&harq_process->d[r][96],
harq_process->c[r],
Kr,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
#if UE_TIMING_TRACE
stop_meas(dlsch_turbo_decoding_stats);
#endif
}
#else
if ((harq_process->C == 1) ||
((r==harq_process->C-1) && (skipped_last==0))) { // last segment with odd number of segments
#if UE_TIMING_TRACE
start_meas(dlsch_turbo_decoding_stats);
#endif
ret = tc
(&harq_process->d[r][96],
harq_process->c[r],
Kr,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
#if UE_TIMING_TRACE
stop_meas(dlsch_turbo_decoding_stats);
#endif
// printf("single decode, exit\n");
// exit(-1);
}
else {
// we can merge code segments
if ((skipped_last == 0) && (r<harq_process->C-1)) {
skipped_last = 1;
Kr_last = Kr;
}
else {
skipped_last=0;
if (Kr_last == Kr) { // decode 2 code segments with AVX2 version
#ifdef DEBUG_DLSCH_DECODING
printf("single decoding segment %d (%p)\n",r-1,&harq_process->d[r-1][96]);
#endif
#if UE_TIMING_TRACE
start_meas(dlsch_turbo_decoding_stats);
#endif
#ifdef DEBUG_DLSCH_DECODING
printf("double decoding segments %d,%d (%p,%p)\n",r-1,r,&harq_process->d[r-1][96],&harq_process->d[r][96]);
#endif
ret = tc_2cw
(&harq_process->d[r-1][96],
&harq_process->d[r][96],
harq_process->c[r-1],
harq_process->c[r],
Kr,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
/*
ret = tc
(&harq_process->d[r-1][96],
harq_process->c[r-1],
Kr_last,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
exit(-1);*/
#if UE_TIMING_TRACE
stop_meas(dlsch_turbo_decoding_stats);
#endif
}
else { // Kr_last != Kr
#if UE_TIMING_TRACE
start_meas(dlsch_turbo_decoding_stats);
#endif
ret = tc
(&harq_process->d[r-1][96],
harq_process->c[r-1],
Kr_last,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
#if UE_TIMING_TRACE
stop_meas(dlsch_turbo_decoding_stats);
start_meas(dlsch_turbo_decoding_stats);
#endif
ret = tc
(&harq_process->d[r][96],
harq_process->c[r],
Kr,
f1f2mat_old[iind*2],
f1f2mat_old[(iind*2)+1],
dlsch->max_turbo_iterations,
crc_type,
(r==0) ? harq_process->F : 0,
&phy_vars_ue->dlsch_tc_init_stats,
&phy_vars_ue->dlsch_tc_alpha_stats,
&phy_vars_ue->dlsch_tc_beta_stats,
&phy_vars_ue->dlsch_tc_gamma_stats,
&phy_vars_ue->dlsch_tc_ext_stats,
&phy_vars_ue->dlsch_tc_intl1_stats,
&phy_vars_ue->dlsch_tc_intl2_stats); //(is_crnti==0)?harq_pid:harq_pid+1);
#if UE_TIMING_TRACE
stop_meas(dlsch_turbo_decoding_stats);
/*printf("Segmentation: C %d r %d, dlsch_rate_unmatching_stats %5.3f dlsch_deinterleaving_stats %5.3f dlsch_turbo_decoding_stats %5.3f \n",
harq_process->C,
r,
dlsch_rate_unmatching_stats->p_time/(cpuf*1000.0),
dlsch_deinterleaving_stats->p_time/(cpuf*1000.0),
dlsch_turbo_decoding_stats->p_time/(cpuf*1000.0));*/
#endif
}
}
}
#endif
if ((err_flag == 0) && (ret>=(1+dlsch->max_turbo_iterations))) {// a Code segment is in error so break;
// LOG_D(PHY,"AbsSubframe %d.%d CRC failed, segment %d/%d \n",frame%1024,subframe,r,harq_process->C-1);
err_flag = 1;
}
}
/*int32_t frame_rx_prev = frame;
int32_t subframe_rx_prev = subframe - 1;
if (subframe_rx_prev < 0) {
frame_rx_prev--;
subframe_rx_prev += 10;
}
frame_rx_prev = frame_rx_prev%1024;*/
if (err_flag == 1) {
#if UE_DEBUG_TRACE
LOG_I(PHY,"[UE %d] DLSCH: Setting NAK for SFN/SF %d/%d (pid %d, status %d, round %d, TBS %d, mcs %d) Kr %d r %d harq_process->round %d\n",
phy_vars_ue->Mod_id, frame, subframe, harq_pid,harq_process->status, harq_process->round,harq_process->TBS,harq_process->mcs,Kr,r,harq_process->round);
#endif
dlsch->harq_ack[subframe].ack = 0;
dlsch->harq_ack[subframe].harq_id = harq_pid;
dlsch->harq_ack[subframe].send_harq_status = 1;
harq_process->errors[harq_process->round]++;
harq_process->round++;
// printf("Rate: [UE %d] DLSCH: Setting NACK for subframe %d (pid %d, round %d)\n",phy_vars_ue->Mod_id,subframe,harq_pid,harq_process->round);
if (harq_process->round >= dlsch->Mdlharq) {
harq_process->status = SCH_IDLE;
harq_process->round = 0;
}
/* if(is_crnti)
{
LOG_D(PHY,"[UE %d] DLSCH: Setting NACK for subframe %d (pid %d, pid status %d, round %d/Max %d, TBS %d)\n",
phy_vars_ue->Mod_id,subframe,harq_pid,harq_process->status,harq_process->round,dlsch->Mdlharq,harq_process->TBS);
}*/
return((1+dlsch->max_turbo_iterations));
} else {
#if UE_DEBUG_TRACE
LOG_I(PHY,"[UE %d] DLSCH: Setting ACK for subframe %d TBS %d mcs %d nb_rb %d\n",
phy_vars_ue->Mod_id,subframe,harq_process->TBS,harq_process->mcs,harq_process->nb_rb);
#endif
harq_process->status = SCH_IDLE;
harq_process->round = 0;
dlsch->harq_ack[subframe].ack = 1;
dlsch->harq_ack[subframe].harq_id = harq_pid;
dlsch->harq_ack[subframe].send_harq_status = 1;
//LOG_I(PHY,"[UE %d] DLSCH: Setting ACK for SFN/SF %d/%d (pid %d, status %d, round %d, TBS %d, mcs %d)\n",
// phy_vars_ue->Mod_id, frame, subframe, harq_pid, harq_process->status, harq_process->round,harq_process->TBS,harq_process->mcs);
/* if(is_crnti)
{
LOG_D(PHY,"[UE %d] DLSCH: Setting ACK for subframe %d (pid %d, round %d, TBS %d)\n",phy_vars_ue->Mod_id,subframe,harq_pid,harq_process->round,harq_process->TBS);
}
LOG_D(PHY,"[UE %d] DLSCH: Setting ACK for subframe %d (pid %d, round %d)\n",phy_vars_ue->Mod_id,subframe,harq_pid,harq_process->round);
}*/
// Reassembly of Transport block here
offset = 0;
/*
printf("harq_pid %d\n",harq_pid);
printf("F %d, Fbytes %d\n",harq_process->F,harq_process->F>>3);
printf("C %d\n",harq_process->C);
*/
for (r=0; r<harq_process->C; r++) {
if (r<harq_process->Cminus)
Kr = harq_process->Kminus;
else
Kr = harq_process->Kplus;
Kr_bytes = Kr>>3;
// printf("Segment %d : Kr= %d bytes\n",r,Kr_bytes);
if (r==0) {
memcpy(harq_process->b,
&harq_process->c[0][(harq_process->F>>3)],
Kr_bytes - (harq_process->F>>3)- ((harq_process->C>1)?3:0));
offset = Kr_bytes - (harq_process->F>>3) - ((harq_process->C>1)?3:0);
// printf("copied %d bytes to b sequence (harq_pid %d)\n",
// Kr_bytes - (harq_process->F>>3),harq_pid);
// printf("b[0] = %x,c[%d] = %x\n",
// harq_process->b[0],
// harq_process->F>>3,
// harq_process->c[0][(harq_process->F>>3)]);
} else {
memcpy(harq_process->b+offset,
harq_process->c[r],
Kr_bytes- ((harq_process->C>1)?3:0));
offset += (Kr_bytes - ((harq_process->C>1)?3:0));
}
}
dlsch->last_iteration_cnt = ret;
return(ret);
}
if (pthread_mutex_lock(&proc->mutex_dlsch_td) != 0) {
LOG_E( PHY, "[SCHED][UE] error locking mutex for UE RXTX\n" );
exit_fun("noting to add");
}
proc->instance_cnt_dlsch_td--;
if (pthread_mutex_unlock(&proc->mutex_dlsch_td) != 0) {
LOG_E( PHY, "[SCHED][UE] error unlocking mutex for UE td1\n" );
exit_fun("noting to add");
}
}
// thread finished
free(arg);
return &UE_dlsch_td_retval;
}
#endif
/*extern int oai_exit;
void *dlsch_td_thread(void *param) {
pthread_setname_np( pthread_self(), "dlsch td processing");
PHY_VARS_UE *UE = ((dlsch_td_params*)param)->UE;
UE_proc_t *proc = &UE->proc;
while (!oai_exit) {
if (wait_on_condition(&proc->mutex_td,&proc->cond_td,&proc->instance_cnt_td,"dlsch td thread")<0) break;
((dlsch_td_params*)param)->ret = dlsch_decoding_2thread0((dlsch_td_params*)param);
if (release_thread(&proc->mutex_td,&proc->instance_cnt_td,"dlsch td thread")<0) break;
if (pthread_cond_signal(&proc->cond_td) != 0) {
printf("[UE] ERROR pthread_cond_signal for dlsch td thread exit\n");
exit_fun( "ERROR pthread_cond_signal" );
return(NULL);
}
}
return(NULL);
}*/
#ifdef PHY_ABSTRACTION
#include "SIMULATION/TOOLS/defs.h"
#ifdef OPENAIR2
......
openair1/PHY/LTE_TRANSPORT/proto.h
View file @ de4712e7
......@@ -166,6 +166,8 @@ int32_t dlsch_encoding_2threads(PHY_VARS_eNB *eNB,
time_stats_t *te_stats,
time_stats_t *i_stats);
uint32_t dlsch_decoding_2thread0(void *arg);
void dlsch_encoding_emul(PHY_VARS_eNB *phy_vars_eNB,
uint8_t *DLSCH_pdu,
LTE_eNB_DLSCH_t *dlsch);
......@@ -1324,6 +1326,19 @@ uint32_t dlsch_decoding(PHY_VARS_UE *phy_vars_ue,
uint8_t harq_pid,
uint8_t is_crnti,
uint8_t llr8_flag);
/*
uint32_t dlsch_decoding_mthread(PHY_VARS_UE *phy_vars_ue,
UE_rxtx_proc_t *proc,
int eNB_id,
int16_t *dlsch_llr,
LTE_DL_FRAME_PARMS *lte_frame_parms,
LTE_UE_DLSCH_t *dlsch,
LTE_DL_UE_HARQ_t *harq_process,
uint32_t frame,
uint8_t subframe,
uint8_t harq_pid,
uint8_t is_crnti,
uint8_t llr8_flag);*/
uint32_t dlsch_decoding_emul(PHY_VARS_UE *phy_vars_ue,
uint8_t subframe,
......
openair1/PHY/defs.h
View file @ de4712e7
......@@ -424,6 +424,20 @@ typedef struct {
/// mutex for UE synch thread
//pthread_mutex_t mutex_slot0_dl_processing;
pthread_mutex_t mutex_slot1_dl_processing;
//int instance_cnt_slot0_dl_processing;
int instance_cnt_dlsch_td;
/// pthread descriptor fep_slot1 thread
//pthread_t pthread_slot0_dl_processing;
pthread_t pthread_dlsch_td;
/// pthread attributes for fep_slot1 processing thread
// pthread_attr_t attr_slot0_dl_processing;
pthread_attr_t attr_dlsch_td;
/// condition variable for UE fep_slot1 thread;
//pthread_cond_t cond_slot0_dl_processing;
pthread_cond_t cond_dlsch_td;
/// mutex for UE synch thread
//pthread_mutex_t mutex_slot0_dl_processing;
pthread_mutex_t mutex_dlsch_td;
//
uint8_t chan_est_pilot0_slot1_available;
uint8_t chan_est_slot1_available;
......@@ -431,6 +445,9 @@ typedef struct {
uint8_t dci_slot0_available;
uint8_t first_symbol_available;
//uint8_t channel_level;
int eNB_id;
int harq_pid;
int llr8_flag;
/// scheduling parameters for fep_slot1 thread
struct sched_param sched_param_fep_slot1;
......
openair1/SCHED/defs.h
View file @ de4712e7
......@@ -44,6 +44,7 @@ enum THREAD_INDEX { OPENAIR_THREAD_INDEX = 0,
#define OPENAIR_THREAD_STACK_SIZE PTHREAD_STACK_MIN //4096 //RTL_PTHREAD_STACK_MIN*6
//#define DLC_THREAD_STACK_SIZE 4096 //DLC stack size
//#define UE_SLOT_PARALLELISATION
//#define UE_DLSCH_PARALLELISATION
enum openair_SCHED_STATUS {
openair_SCHED_STOPPED=1,
......
openair1/SCHED/phy_procedures_lte_ue.c
View file @ de4712e7
......@@ -3553,6 +3553,17 @@ void ue_pmch_procedures(PHY_VARS_UE *ue, UE_rxtx_proc_t *proc,int eNB_id,int abs
ue->dlsch_MCH[0]->harq_processes[0]->G,
ue->pdsch_vars_MCH[0]->llr[0],0,nr_tti_rx<<1);
#ifdef UE_DLSCH_PARALLELISATION
ret = dlsch_decoding_mthread(ue,proc, eNB_id,
ue->pdsch_vars_MCH[0]->llr[0],
&ue->frame_parms,
ue->dlsch_MCH[0],
ue->dlsch_MCH[0]->harq_processes[0],
frame_rx,
nr_tti_rx,
0,
0,1);
#else
ret = dlsch_decoding(ue,
ue->pdsch_vars_MCH[0]->llr[0],
&ue->frame_parms,
......@@ -3562,6 +3573,8 @@ void ue_pmch_procedures(PHY_VARS_UE *ue, UE_rxtx_proc_t *proc,int eNB_id,int abs
nr_tti_rx,
0,
0,1);
printf("start pmch dlsch decoding\n");
#endif
} else { // abstraction
#ifdef PHY_ABSTRACTION
ret = dlsch_decoding_emul(ue,
......@@ -3973,6 +3986,19 @@ void ue_dlsch_procedures(PHY_VARS_UE *ue,
#if UE_TIMING_TRACE
start_meas(&ue->dlsch_decoding_stats[ue->current_thread_id[nr_tti_rx]]);
#endif
#ifdef UE_DLSCH_PARALLELISATION
ret = dlsch_decoding_mthread(ue,proc,eNB_id,
pdsch_vars->llr[0],
&ue->frame_parms,
dlsch0,
dlsch0->harq_processes[harq_pid],
frame_rx,
nr_tti_rx,
harq_pid,
pdsch==PDSCH?1:0,
dlsch0->harq_processes[harq_pid]->TBS>256?1:0);
#else
ret = dlsch_decoding(ue,
pdsch_vars->llr[0],
&ue->frame_parms,
......@@ -3983,6 +4009,8 @@ void ue_dlsch_procedures(PHY_VARS_UE *ue,
harq_pid,
pdsch==PDSCH?1:0,
dlsch0->harq_processes[harq_pid]->TBS>256?1:0);
printf("start cW0 dlsch decoding\n");
#endif
#if UE_TIMING_TRACE
stop_meas(&ue->dlsch_decoding_stats[ue->current_thread_id[nr_tti_rx]]);
......@@ -4039,6 +4067,19 @@ void ue_dlsch_procedures(PHY_VARS_UE *ue,
start_meas(&ue->dlsch_decoding_stats[ue->current_thread_id[nr_tti_rx]]);
#endif
#ifdef UE_DLSCH_PARALLELISATION
ret1 = dlsch_decoding_mthread(ue,proc, eNB_id,
pdsch_vars->llr[1],
&ue->frame_parms,
dlsch1,
dlsch1->harq_processes[harq_pid],
frame_rx,
nr_tti_rx,
harq_pid,
pdsch==PDSCH?1:0,
dlsch1->harq_processes[harq_pid]->TBS>256?1:0);
#else
ret1 = dlsch_decoding(ue,
pdsch_vars->llr[1],
&ue->frame_parms,
......@@ -4049,6 +4090,8 @@ void ue_dlsch_procedures(PHY_VARS_UE *ue,
harq_pid,
pdsch==PDSCH?1:0,
dlsch1->harq_processes[harq_pid]->TBS>256?1:0);
printf("start cw1 dlsch decoding\n");
#endif
#if UE_TIMING_TRACE
stop_meas(&ue->dlsch_decoding_stats[ue->current_thread_id[nr_tti_rx]]);
......
targets/COMMON/threads_t.h
View file @ de4712e7
......@@ -9,6 +9,9 @@ typedef struct threads_s {
int slot1_proc_one;
int slot1_proc_two;
int slot1_proc_three;
int dlsch_td_one;
int dlsch_td_two;
int dlsch_td_three;
} threads_t;
#endif /* _THREADS_T_H_ */
targets/RT/USER/lte-softmodem.c
View file @ de4712e7
......@@ -649,6 +649,9 @@ static void get_options (int argc, char **argv) {
LONG_OPTION_THREADSLOT1PROCONE,
LONG_OPTION_THREADSLOT1PROCTWO,
LONG_OPTION_THREADSLOT1PROCTHREE,
LONG_OPTION_THREADDLSCHTDONE,
LONG_OPTION_THREADDLSCHTDTWO,
LONG_OPTION_THREADDLSCHTDTHREE,
LONG_OPTION_DCIFORMAT,
LONG_OPTION_AGREGATIONLEVEL,
LONG_OPTION_DEMOD_SHIFT,
......@@ -691,6 +694,9 @@ static void get_options (int argc, char **argv) {
{"threadSlot1ProcOne", required_argument, NULL, LONG_OPTION_THREADSLOT1PROCONE},
{"threadSlot1ProcTwo", required_argument, NULL, LONG_OPTION_THREADSLOT1PROCTWO},
{"threadSlot1ProcThree", required_argument, NULL, LONG_OPTION_THREADSLOT1PROCTHREE},
{"threadDlschTdOne", required_argument, NULL, LONG_OPTION_THREADDLSCHTDONE},
{"threadDlschTdTwo", required_argument, NULL, LONG_OPTION_THREADDLSCHTDTWO},
{"threadDlschTdThree", required_argument, NULL, LONG_OPTION_THREADDLSCHTDTHREE},
{"DCIformat", required_argument, NULL, LONG_OPTION_DCIFORMAT},
{"AgregationLevel", required_argument, NULL, LONG_OPTION_AGREGATIONLEVEL},
{"dlsch-demod-shift", required_argument, NULL, LONG_OPTION_DEMOD_SHIFT},
......@@ -836,6 +842,15 @@ static void get_options (int argc, char **argv) {
case LONG_OPTION_THREADSLOT1PROCTHREE:
threads.slot1_proc_three=atoi(optarg);
break;
case LONG_OPTION_THREADDLSCHTDONE:
threads.dlsch_td_one=atoi(optarg);
break;
case LONG_OPTION_THREADDLSCHTDTWO:
threads.dlsch_td_two=atoi(optarg);
break;
case LONG_OPTION_THREADDLSCHTDTHREE:
threads.dlsch_td_three=atoi(optarg);
break;
case LONG_OPTION_DCIFORMAT:
dci_Format = atoi(optarg);
break;
......
targets/RT/USER/lte-ue.c
View file @ de4712e7
......@@ -46,6 +46,7 @@
#include "SCHED/extern.h"
#include "LAYER2/MAC/extern.h"
#include "LAYER2/MAC/proto.h"
#include "PHY/LTE_TRANSPORT/proto.h"
#include "UTIL/LOG/log_extern.h"
#include "UTIL/OTG/otg_tx.h"
......@@ -930,6 +931,12 @@ void init_UE_threads(PHY_VARS_UE *UE) {
printf("Init_UE_threads rtd %d proc %d nb_threads %d i %d\n",rtd->proc->sub_frame_start, UE->proc.proc_rxtx[i].sub_frame_start,nb_threads, i);
pthread_create(&UE->proc.proc_rxtx[i].pthread_rxtx, NULL, UE_thread_rxn_txnp4, rtd);
#ifdef UE_DLSCH_PARALLELISATION
pthread_mutex_init(&UE->proc.proc_rxtx[i].mutex_dlsch_td,NULL);
pthread_cond_init(&UE->proc.proc_rxtx[i].cond_dlsch_td,NULL);
pthread_create(&UE->proc.proc_rxtx[i].pthread_dlsch_td,NULL,dlsch_decoding_2thread0, rtd);
#endif
#ifdef UE_SLOT_PARALLELISATION
//pthread_mutex_init(&UE->proc.proc_rxtx[i].mutex_slot0_dl_processing,NULL);
//pthread_cond_init(&UE->proc.proc_rxtx[i].cond_slot0_dl_processing,NULL);
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment