CI: Generalize PhySim tests, run PhySims on Gracehopper
This changeset generalizes/harmonizes the CI infrastructure for physim
testing, and adds a new pipeline for physim tests on Gracehopper
machines.
Specifically:
- Refactor the existing CI code for running and analyzing physical
simulators. The existing code assumed those tests are deployed using
OpenShift. Instead, OpenShift/Helm-specificities (e.g., running a
"release") have been removed, and instead of hardcoding the deployment
of the physical simulators using a script for Helm, the function takes
any script that can deploy using Helm/docker/from source. The overall
appearance is the same.
- Add a new pipeline RAN-PhySim-GraceHopper-5G that executes physical
simulators on gracehopper. Reporting is done in the same way as for
existing physical simulators (e.g., like in the RAN-PhySim-Cluster-5G
pipeline), but deployed in docker instead of helm.
-> some tests don't work reliably on ARM yet and are skipped
- the RAN-gNB-N300-Timing-Phytest-LDPC pipeline has been reworked to use
the same code for testing T2 offload, with ctest, and the same
reporting.
- the CUDA LDPC test XML has been removed, as these tests don't work
reliably.
- add a new macro add_timed_physim_test() that allows to define
thresholds to be checked, so that ctest will check for them. see the
documentation for more information
Most (but not all) physical simulators are defined under
openair1/SIMULATION/. Defining physical simulators below there seems to
be the more "obvious" directory anyway, as the physical simulators are
openair1/SIMULATION/tests
Phy(L1) Sim tests => PhySim (tests)
gNB: adaptive DL/UL MCS based on reported SSB-SINR and measured PUSCH SNR
Currently this is used esp. if the number of rounds is limited to 1
(e.g. disable_harq).
feat(ldpc_aal): Do not try EAL init if FHI 7.2 is enabled
When FHI 7.2 is enabled, the EAL initialization is done by the FHI and
the attempt to init EAL by ldpc_aal will fail. In that case, it is not
necessary neither to provide a DPDK core list for BBDEV.
This MR offers to skip EAL init in ldpc_aal when FHI 7.2 is enabled in
order to avoid error messages in the log and to allow dropping the
useless DPDK cores argument in that case.
Also add some fixes to the LDPC offload documentation.
When FHI 7.2 is enabled, the EAL initialization is done by the FHI and
the attempt to init EAL by ldpc_aal will fail.
In that case, it is not necessary neither to provide a DPDK core list
for BBDEV.
This commit offers to skip EAL initialization if it is already initialized
in order to allow dropping the useless DPDK cores argument in that case.
The change relies on `rte_dev_probe` to detect if EAL was already
initialized.
Ldpc encoder zc384 optimization
This is a modification for the ldpc encoder which improves performance for
high-throughput cases. It contains a newly-generated file for innermost part of
the parity bit generation which, for AVX512, uses the mm512_permutex2var_epi8
instruction and for 128-bit (aarch64) the mm_alignr_epi8 (vextq_s8) in the
computation corresponding to each non-zero bit of the H matrix. This allows for
removing a very large memcpy and reduces the memory demand for the parity matrix
computation. For the moment this is limited to largest case in BG1, namely
Zc=384. This is the only format used when the number of segments after codeblock
segmentation is large (>8). This improves the timing required for ldpc parity
bits generation by more than a factor 2 with AVX512. This significantly reduces
the total time to generate the DLSCH/PDSCH.
See the merge request description for numbers.
Speedup complex rotate for aarch64
This is a simple optimization for aarch64 for the "rotate_cpx_vector" used for
the phase_rotation in 5G NR. It provides a speedup of around 1.5 on high-end
aarch64 targets compared to the SIMDe translated version. It is implemented with
native NEON intrinsics and was largely inspired by suggestions from ChatGPT 5 :-)
newly-generated files for innermost part of the parity bit
generation which, for AVX512, uses the mm512_permutex2var_epi8
instruction and for 128-bit (aarch64) the mm_alignr_epi8 (vextq_s8) in
the computation corresponding to each non-zero bit of the H matrix.
It is generated with ldpctest -n0 -g1. the changes for this are in a
future MR. I did these changes to compare with the GPU version
(ldpc_encoder_cuda), so that the CPU version is as fast as possible. The
generator will come when we merge the GPU encoder/decoder which also has
a generator for the CUDA version. it was difficult to extract the
changes in ldpctest/ldpc_encoder.c so I just took the generated file.
This allows for removing a very large memcpy and reduces the memory
demand for the parity matrix computation. For the moment this is limited
to largest case in BG1, namely Zc=384. This is the only format used when
the number of segments after codeblock segmentation is large (>8). This
improves the timing required for ldpc parity bits generation by more
than a factor 2 with AVX512. This significantly reduces the total time
to generate the DLSCH/PDSCH
RFSim test for 10 UEs/process
This MR adds a testcase and changes required to run 10 UEs in a single
process with RFSim
- Added deregistration for all connected UEs
- Refactored MAC/RRC instance handling
- made map_current_symbol thread safe
- made opeanir0_cfg thread safe and local
- Added 10UE/process testcase
RF emulator and improvements of the timers
RF emulator:
Add an RF library emulating a real device without the need of a real
device. This is thought to enable using phy-test without requiring a
radio device or channel simulation. The library synchronizes on the
real time clock, drops TX and generate noise on RX. Noise generation
can be disabled and the noise level configured.
Timers improvements:
- Fix standard deviation for merged timers
- Add display of ULSCH decoding timers
- Add decoding timer in AAL
CI: speed up and simplify build, remove unnecessary Python code, remove --eNB* parameter list
This cleans up the CI framework for running end-to-end tests, and
improves the integration with OAI code. Specifically
- cleanup build_oai and helpers to output all text, which is captured
better by the CI
- build some images in the Cluster-Build in parallel to increase speed
- remove all --eNB{,1,2}{IPAddress,UserName,Password} variables. Those
were indexed through "instance" and "server ID" and are replaced with
the actual node name. I consider further cleanup to introduce a
"global" XML file node name so we don't need this in each step and
instead define this once.
- individual cleanup
Introduce support for Cat-B RU using Predefined-Beam BF with Aerial L1
This merge introduces the support for BeamID based Beamforming in OAI L2 using FAPI and leveraging Aerial L1. Support for OAI L1 is not in the scope of this MR.
Main changes:
- Introduced a third state for the analog_beamforming parameter.
0=disabled,1=analog_bf,2=oran_bf
- Correctly populated the FAPI BF PDUs (prg_size, num_prgs,
dig_bf_interfaces)
Cleanup documentation and interface of the LDPC segment coding libraries
Fix interface: The LDPC segment decoder interface had variables
'harq_pid', 'ulsch_id' and 'C' that are not useful to the segment
decoder. This MR removes these variables from the interface.
Fix documentation: libconfig log do not display anymore by default so we
may better not show them in the documentation to avoid any
misunderstanding.
Remove ldpc_cl library since the library does not work.
Workaround for handling 2 SS0 occasions per slot
Search Space 0 monitoring might depend on SSB. In !3332 this dependence
was removed by moving the configuration of SS0 earlier in a workaround
to avoid a segfault in handover. Now, a new workaround is needed to
handle 2 search space occasions per slot.
rmagueta