[rfsim] CUDA LDPC acceleration fails at Msg3 (ULSCH with no signal) (exceeded RA window)

[wdnmdwdnmdwdnmdwdnmd]

when I enable the CUDA LDPC decoder (--loader.ldpc.shlibversion _cuda), the RACH procedure fails at Msg3.

Initializing LDPC runtime 51
Initializing LDPC context (TID 51)
Initialized CUDA LDPC decoder
43338.126859 [PHY] I Attaching RU 0 antenna 0 to gNB antenna 0
43338.126884 [UTIL] I threadCreate() for Tpool0_-1: creating thread with affinity ffffffff, priority 97
43338.127888 [UTIL] I threadCreate() for Tpool1_-1: creating thread with affinity ffffffff, priority 97
43338.128924 [UTIL] I threadCreate() for Tpool2_-1: creating thread with affinity ffffffff, priority 97
Initializing LDPC context (TID 75)
Initializing LDPC context (TID 76)
Initialized new CUDA LDPC decoder context
Initialized new CUDA LDPC decoder context
43338.129929 [UTIL] I threadCreate() for Tpool3_-1: creating thread with affinity ffffffff, priority 97
Initializing LDPC context (TID 77)
Initialized new CUDA LDPC decoder context
43338.130918 [UTIL] I threadCreate() for Tpool4_-1: creating thread with affinity ffffffff, priority 97
43338.131921 [UTIL] I threadCreate() for Tpool5_-1: creating thread with affinity ffffffff, priority 97
Initializing LDPC context (TID 78)
Initialized new CUDA LDPC decoder context
43338.132900 [UTIL] I threadCreate() for Tpool6_-1: creating thread with affinity ffffffff, priority 97
Initializing LDPC context (TID 79)
Initialized new CUDA LDPC decoder context
Initializing LDPC context (TID 80)
Initialized new CUDA LDPC decoder context
43338.133950 [UTIL] I threadCreate() for Tpool7_-1: creating thread with affinity ffffffff, priority 97
Initializing LDPC context (TID 81)
Initialized new CUDA LDPC decoder context
Initializing LDPC context (TID 82)
Initialized new CUDA LDPC decoder context
43338.135043 [UTIL] I threadCreate() for L1_rx_thread: creating thread with affinity 1, priority 97
43338.135961 [UTIL] I threadCreate() for L1_tx_thread: creating thread with affinity 2, priority 97
43338.138131 [UTIL] I threadCreate() for L1_stats: creating thread with affinity ffffffff, priority 1
43338.139083 [PHY] I got sync (L1_stats_thread)
TYPE TO TERMINATE
43338.139089 [PHY] I got sync (ru_thread)
43338.139143 [PHY] I RU 0 rf device ready
43338.139145 [PHY] I RU 0 RF started cpu_meas_enabled 0
43338.139151 [HW] I No connected device, generating void samples...
43340.160056 [PHY] A Command line parameters for OAI UE: -C 3319680000 -r 106 --numerology 1 --ssb 516
43340.727250 [NR_MAC] I Frame.Slot 128.0

43343.320684 [NR_MAC] I Frame.Slot 256.0

43345.915954 [NR_MAC] I Frame.Slot 384.0

43347.350176 [HW] I Client connects from ::ffff:192.168.71.150:53419
43347.387424 [NR_MAC] I Frame.Slot 512.0

43347.523883 [NR_PHY] A [RAPROC] 563.19 Initiating RA procedure with preamble 14, energy 44.3 dB (I0 0, thres 120), delay 0 start symbol 0 freq index 0
43347.523938 [NR_MAC] A 563.19 UE RA-RNTI 010b TC-RNTI 7987: initiating RA procedure
43347.523955 [NR_MAC] I UE 7987: Msg3 scheduled at 564.19 (564.10 TDA 0) start 0 RBs 8
43347.523961 [NR_MAC] A UE 7987: 564.10 Generating RA-Msg2 DCI, RA RNTI 0x10b, state 1, preamble_index(RAPID) 14, timing_offset = 0 (estimated distance 0.0 [m])
43347.523973 [NR_MAC] A 564.10 Send RAR to RA-RNTI 010b
43347.528537 [NR_MAC] W MSG3 ULSCH with no signal
.
.
.
.
.
43495.734598 [NR_MAC] E UE RA-RNTI 010f TC-RNTI c45f: exceeded RA window, cannot schedule Msg2
43495.734599 [NR_MAC] I Remove NR rnti 0xc45f
43495.735334 [NR_PHY] A [RAPROC] 451.19 Initiating RA procedure with preamble 44, energy 31.7 dB (I0 102, thres 120), delay 0 start symbol 0 freq index 0
43495.735393 [NR_MAC] A 451.19 UE RA-RNTI 010b TC-RNTI cda5: initiating RA procedure

[SebastianCa]

Hi @wdnmdwdnmdwdnmdwdnmd,

on which hardware do you run the setup? Can you please run the unittests of the LDPC plugin?
Did you modify anything in the config or is it the SionnaRK default configuration?

[wdnmdwdnmdwdnmdwdnmd]

Hi @SebastianCa
my environment:

• CPU: AMD Ryzen 7 9700X, 8 cores / 16 threads
• GPU: NVIDIA GeForce RTX 3060 LHR
• OS: Ubuntu 22.04
• USRP-B210 + quectel-RM500U-CNV
  I use default configuration. Without enabling the LDPC CUDA plugin, both the rfsim setup and the B210/USRP setup can run successfully.
  I use pytorch-based-sionna environment, so the unit test for ldpc plugin report ModuleNotFoundError: No module named 'tensorflow'
  Thank you.

[SebastianCa]

Please note that SionnaRK targets DGX Spark and Jetson platforms; we cannot provide detailed support for other hardware configurations.

A few comments:

• It seems you have already migrated to Sionna 2.0. However, the unit tests still depend on TensorFlow. It may be worth to check if the unittests pass with Tensorflow.
• Did you manually change the Dockerfiles? The ones we provide are for Spark/Jetson.
• Please check the Docker build logs to verify that the CUDA decoder was built successfully.

[wdnmdwdnmdwdnmdwdnmd]

ok thank you！