Add new Dockerfiles to enable building nrUE and gNB with CUDA support on Jetson.
Compared to the standard build, these images use Ubuntu 22.04 instead of Ubuntu
24.04. This change is required because CUDA 12.2.0 runtime and development
images are only available for Ubuntu 22.04.
To build the nrUE and gNB image with CUDA support, run the following commands:
docker build --target ran-base-cuda --tag ran-base-cuda:latest --file docker/Dockerfile.base.cuda.ubuntu .
docker build --target ran-build-cuda --tag ran-build-cuda:latest --file docker/Dockerfile.build.cuda.ubuntu .
docker build --target oai-gnb --tag oai-gnb:test --file docker/Dockerfile.gNB.cuda.ubuntu .
docker build --target oai-nr-ue --tag oai-nr-ue:test --file docker/Dockerfile.nrUE.cuda.ubuntu .
A new XML file was added to build images on Jetson3 and push images with CUDA
support, using the prefix "armv8_cuda".
Remove double nr_dlsim eval
Remove CUDA from normal LDPC coding, make standalone
nr_dlsim seems to work.
ldpctest: need to implement memcpy() in _cuda.c
nr_ulsim fails
fixup ldpctest: make it run
Initial version of ldpc_decoder_cuda done. Change configuration to GH200
modify -h option in dlsim and ulsim. First scratch for nrLDPC GPU decoder tutorial
Change code
add missing initialization on jetson for nr_ulsim. Bug fixed
add buffer memory management; remove an extra synchronization in cuda decoder
fix for ulsim via cuda encoder on jetson
fix for 256QAM
Move LDPC CUDA to separate subdirectory
Only functional change: remove USE_CUDA
Rename constant for clarity
ldpc_cuda: add tests to verify it works
Signed-off-by: Robert Schmidt <robert.schmidt@openairinterface.org>
Implement fallback CPU
Remove use_gpu
Signed-off-by: Robert Schmidt <robert.schmidt@openairinterface.org>
Report errors found in the build logs, but do not fail the job if some are
present. Based on observations, certain errors (for example):
update-alternatives: error: alternative path /usr/share/man/man7/bash-builtins.7.gz doesn't exist
do not actually cause the build to fail. Instead of relying on log parsing, we should maybe
check the return code of the build process to determine whether the build truly failed?
Signed-off-by: Robert Schmidt <robert.schmidt@openairinterface.org>
an incorrect assertion that was enforcing symmetric antenna
configurations only. I removed the assertion and updated the logging for
UE antenna dimensions and the channel model antenna dimensions used to
search in CIRDB.
Signed-off-by: Merkebu Girmay <merkebu.girmay@openairinterface.org>
Preparatory changes for LDPC CUDA integration
This changeset is preparatory work to merge LDPC CUDA offload in !4097.
Notably, it includes changes for cleaning up some defines, updates to
the LDPC interface, and CI specific changes to prepare for the other MR.
The goal is to merge changes in a first step such that, in a second
step, only the actual LDPC CUDA implementation and ancillary changes
(CUDA memory allocations, CI files, ...) need to be merged.
Reviewed-By: Bartosz Podrygajlo <bartosz.podrygajlo@openairinterface.org>
This memset() seems to be superfluous: following the code flow,
d_to_be_cleared is set on the same condition which guards the memset.
Following nrLDPC_coding_decoder() -> nrLDPC_prepare_TB_decoding() ->
nr_process_decode_segment() -> nr_rate_matching_ldpc_rx(), another
memset() is done using d_to_be_cleared.
Reported-by: Bartosz Podrygajlo <bartosz.podrygajlo@openairinterface.org>
Signed-off-by: Robert Schmidt <robert.schmidt@openairinterface.org>
CI: Clone from internal git repo
This MR refactors the CI pipeline as a preparatory step for the GitLab
--> GitHub migration. The core idea is to use a dedicated stage
RAN-Local-Repo-Push (introduced in !4017 (merged)) to clone source and
target branches, merges them and push a resulting branch into internal
git mirror. All downstream test jobs then check out exclusively from
that mirror. This makes the CI subpipelines independent of the remote
repository and improves stability in case of disruptions - test jobs
always check out from the internal mirror, not from GitLab/GitHub
directly.
Changes:
- doGitLabMerge.sh removed - merge is now done once in
RAN-Local-Repo-Push, slave jobs check out the pre-merged branch from
the internal mirror via SCM RAN-Local-Repo-Push extended:
- mergeWithTarget boolean parameter makes the merge optional (skipped on
direct push events) source/target remotes replace reliance on origin,
- targetRepo is now a configurable parameter to support fork-based MRs
- CreateTag() removed from Python - testBranch and testRepository
(default: internal mirror) are computed in the Jenkinsfile and passed
directly, two optional override parameters (customBranch,
customRepository) allow targeting a non-default branch or repo
- create_workspace.sh is reduced to git clone --depth=1 --branch
- all eNB_/ran-prefixed Jenkinsfile parameters and Python CLI arguments
renamed to neutral names (sourceBranch, sourceCommit, workspace,
repository,...), both old and new names accepted during the transition
period
Reviewed-by: Robert Schmidt <robert.schmidt@openairinterface.org>
A later commit (for LDPC CUDA) will reuse add_physim_test(), but is in a
completely different location than the existing physim test definitions.
Hence, move it to global scope for later reuse.
Signed-off-by: Robert Schmidt <robert.schmidt@openairinterface.org>
Provide the memory buffer size for storing LLRs across HARQ rounds. then
we don't need to check for this in the API users.
Signed-off-by: Robert Schmidt <robert.schmidt@openairinterface.org>
Rename this option from CUDA_ENABLE, as it might otherwise conflict with
another option for LDPC CUDA. For consistency, name the compile option
to CHANNEL_SIM_CUDA. This way, the LDPC CUDA option (ENABLE_LDPC_CUDA
and compile definition LDPC_CUDA) will not conflict, and the
corresponding code is clearly "labelled".
A further change is the use of target_sources to simplify the definition
of the channel_pipeline executable.
Signed-off-by: Robert Schmidt <robert.schmidt@openairinterface.org>
Information about sourceBranch and sourceCommit is already encoded in
the testBranch name, making it redundant in the HTML header.
Also remove the Job Trigger and Target Branch lines, as these parameters
provide limited value and are not particularly relevant for the report.
Signed-off-by: Jaroslava Fiedlerova <jaroslava.fiedlerova@openairinterface.org>
"Create new Workspace for server 0/1" no longer makes sense now that
workspace creation is not tied to a specific server index.
Signed-off-by: Jaroslava Fiedlerova <jaroslava.fiedlerova@openairinterface.org>
NR_MAC/PHY: align SRS time_start_position semantics between L2 and L1
Problem
nfapi_nr_srs_pdu_t::time_start_position was being interpreted
inconsistently between OAI MAC (the producer) and the various L1
consumers:
- MAC populated the field with SRS-ResourceMapping.startPosition from
the RRC config — the value 3GPP defines as an offset counted backwards
from the end of the slot (TS 38.331).
- L1 consumers (openair1/PHY/NR_TRANSPORT/srs_rx.c,
openair1/SCHED_NR/phy_procedures_nr_gNB.c) and the MAC's own VRB-map
consumer (gNB_scheduler_srs.c) all re-derived the absolute symbol
index with NR_SYMBOLS_PER_SLOT - 1 - time_start_position.
This works for OAI's in-tree soft L1 (every consumer applies the same
inversion), but external FAPI L1s (NVIDIA Aerial / cuBB) treat
time_start_position as the absolute symbol index per the SCF FAPI 222
PDU semantics. The MAC and external L1 therefore disagreed by one symbol
inversion. Net effect: SRS was scheduled on the wrong OFDM symbol, the
RU sampled noise instead of the SRS, and the channel-matrix returned in
SRS.indication was all zeros for almost every occasion. Fix
Move the inversion to the producer side. MAC populates the PDU with the
absolute symbol index (NR_SYMBOLS_PER_SLOT - 1 - startPosition), and
every consumer uses srs_pdu->time_start_position directly as l0.
The two semantics are mathematically equivalent for OAI's soft L1 (l0 =
NR_SYMBOLS_PER_SLOT - 1 - startPosition either way), so the symbol index
used internally is unchanged on non-Aerial builds. External L1 consumers
now receive the value they expect. Testing status
- [x] OAI L1 + L2
- [x] Aerial L1 in CAT-B / mMIMO mode
- [x] Aerial L1 in CAT-A mode
- [x] RFsim / nrUE end-to-end loopback:
Reviewed-By: Rúben Soares Silva <rsilva@allbesmart.pt>
Add Fedora 44/Ubuntu 26 to list of supported distributions
See commits for more details. Includes a minor cleanup. What I tested
for U22/U24/U26 and F44 is
./build_oai --ninja -c -I -w USRP --gNB --nrUE
Reviewed-By: Luis Pereira <lpereira@allbesmart.pt>
changes in cudaMemadvise/cudaPrefetch API for v13.
Build fails on DGX spark or any Blackwell GPU target without this fix.
It just checks the CUDA version in a couple of places where GPU offload
is used for channel simulation (nr_dlsim,nr_ulsim and
channel_pipeline.cu) and adapts the usage of the CUDA API accordingly.
Reviewed-By: Bartosz Podrygajlo <bartosz.podrygajlo@openairinterface.org>
Remove CreateTag() and the merge/targetBranch parameters from
CreateWorkspace, passing the final branch name and repository directly
from Jenkins instead. Jenkinsfile computes testBranch and testRepository
(with INTERNAL_REPO as default). create_workspace.sh now does a simple
shallow clone by branch name.
Introduce two optional Jenkins parameters that allow overriding the
default branch reference and git repository URL used across the CI
pipeline.
- customBranch: when set, replaces the entire computed
{sourceBranch}-{commitID}
- customRepository: when set, replaces INTERNAL_REPO as the git remote
used by Create_Workspace to clone the source code onto the test node.
Signed-off-by: Jaroslava Fiedlerova <jaroslava.fiedlerova@openairinterface.org>
Use the current OAI CI default UHD version. Update the documentation to
show how to install a recent version.
Signed-off-by: Robert Schmidt <robert.schmidt@openairinterface.org>
The existing patch (generated through sed) works also in UHD versions
beyond 4.8, so allow to use that.
Further, since sed is confusing, store the patch directly. Update the
dockerfiles so that docker copies the right file for UHD 4.8.
Signed-off-by: Robert Schmidt <robert.schmidt@openairinterface.org>
Also, for UHD install from packages, add the right version for both
Ubuntu 24 and Ubuntu 26: in U24, use the currently recommended 4.8, and
bump to latest in U26.
Signed-off-by: Robert Schmidt <robert.schmidt@openairinterface.org>
The CI does not use installation of UHD from package, hence this
mechanism is not needed (and should not be, anyway).
Signed-off-by: Robert Schmidt <robert.schmidt@openairinterface.org>
