maki/openairinterface5g
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/resource_mapping_optim' into integration_2025_w07 (!3127)

Browse Source 
Optimizations of PDSCH Resource Mapping in nr_dlsch.c/nr_modulation.c

These changes add SIMD optimizations for Neon/AVX2/AVX512 in the PDSCH
transmit path. The timing improvements are listed here based on the

    nr_dlsim -e25 -R273 -b273 -s30 -x "layers" -y 4 -z 4 -P

benchmark with "layers" 2,3,4 and comparing "PHY proc tx":

273 PRBS, mcs25, 64QAM

peafowl (gcc11,AMD EPYC 9374F)

    2-layer, 4 TX : 431 us (develop 565 us)
    3-layer, 4 TX : 692 us (develop 849 us)
    4-layer, 4 TX : 963 us (develop 1172 us)

stupix (gcc10, Xeon Gold 6354)

    2-layer, 4 TX : 568 us (develop 652 us)
    3-layer, 4 TX : 901 us (develop 1030 us)
    4-layer, 4 TX : 1250 us (develop 1396 us)

matix (gcc14, Ryzen 9 PRO 7945)

    2-layer, 4 TX : 317 us (develop 505 us)
    3-layer, 4 TX : 538 us (develop 779 us)
    4-layer, 4 TX : 767 us (develop 1233 us)
This commit is contained in:
Robert Schmidt
2025-02-17 22:09:26 +01:00
parent f0d6d42aec 8a63b01387
commit 651ef3daee
7 changed files with 1255 additions and 791 deletions
Download Patch File Download Diff File Expand all files Collapse all files
openair1/SIMULATION/NR_PHY/dlsim.c
+6 -2
View File
@@ -965,10 +965,12 @@ printf("%d\n", slot);
reset_meas(&gNB->dlsch_segmentation_stats);
reset_meas(&gNB->dlsch_modulation_stats);
reset_meas(&gNB->dlsch_encoding_stats);
reset_meas(&gNB->dci_generation_stats);
reset_meas(&gNB->tinput);
reset_meas(&gNB->tprep);
reset_meas(&gNB->tparity);
reset_meas(&gNB->toutput);
reset_meas(&gNB->phase_comp_stats);
uint32_t errors_scrambling[16] = {0};
int n_errors[16] = {0};
@@ -1264,6 +1266,7 @@ printf("%d\n", slot);
UE->dl_harq_processes[0][slot].C,
msgDataTx->dlsch[0][0].harq_process.pdsch_pdu.pdsch_pdu_rel15.TBSize[0] << 3);
printDistribution(&gNB->phy_proc_tx,table_tx,"PHY proc tx");
printStatIndent2(&gNB->dci_generation_stats, "DCI encoding time");
printStatIndent2(&gNB->dlsch_encoding_stats,"DLSCH encoding time");
printStatIndent3(&gNB->dlsch_segmentation_stats,"DLSCH segmentation time");
printStatIndent3(&gNB->tinput,"DLSCH LDPC input processing time");
@@ -1274,8 +1277,9 @@ printf("%d\n", slot);
printStatIndent3(&gNB->dlsch_interleaving_stats, "DLSCH Interleaving time");
printStatIndent2(&gNB->dlsch_modulation_stats,"DLSCH modulation time");
printStatIndent2(&gNB->dlsch_scrambling_stats, "DLSCH scrambling time");
printStatIndent2(&gNB->dlsch_resource_mapping_stats, "DLSCH Resource Mapping time");
printStatIndent2(&gNB->dlsch_precoding_stats,"DLSCH Layer Precoding time");
printStatIndent2(&gNB->dlsch_precoding_stats,"DLSCH Mapping/Precoding time");
if (gNB->phase_comp)
printStatIndent2(&gNB->phase_comp_stats, "Phase Compensation");
printf("\nUE function statistics (per %d us slot)\n", 1000 >> *scc->ssbSubcarrierSpacing);
for (int i = RX_PDSCH_STATS; i <= DLSCH_PROCEDURES_STATS; i++) {