GPU Channel Simulation: nr_dlsim/nr_ulsim Integration + CI Test
This MR is a follow-up to !3886 (merged) and extends CUDA-based channel
simulation support to nr_dlsim and nr_ulsim, together with the
corresponding CI tests.
The MR incorporates:
- Cherry-picked commits e0c204f7 and 5f344876 from !3588
- Integration of new CI physim tests using GPU-accelerated channel
simulation in nr_dlsim and nr_ulsim tests, see commit c727b3bb
New CI job: RAN-Channel-Simulation
https://jenkins-oai.eurecom.fr/job/RAN-Channel-Simulation/
Fix compilation error:
/oai-ran/openair1/SIMULATION/NR_PHY/dlsim.c:932:42: error: variable 'h_channel_coeffs' set but not used [-Werror,-Wunused-but-set-variable]
932 | float *h_channel_coeffs = NULL;
| ^
- Introduce usage of the -f option in ulsim to allow enabling specific
feature flags at runtime.
- Add support for the "cuda" flag to enable CUDA channel simulation when
compiled with CUDA support. If an unsupported flag is provided, nr_ulsim
reports an error.
This commit integrates the GPU-based channel simulation pipeline into
the ulsim.
A `--cuda` command-line flag is added to enable the GPU path at
runtime. When activated, the simulator will:
- Allocate all necessary GPU and pinned host memory via the
`init_cuda_chsim_buffers` helper function.
- In the main simulation loop, call the `run_channel_pipeline_cuda`
function to perform the channel simulation on the GPU, bypassing the
CPU-based functions.
- Clean up all CUDA resources on exit using the
`free_cuda_chsim_buffers` helper.
The cpu path is also updated to use multipath_channel_float and add_noise_float.
- Replace previous handling of -f with a generic feature flag option.
- Add support for the "cuda" flag to enable CUDA channel simulation when
compiled with ENABLE_CUDA.
- Remove unused SCGroupConfig file handling (scg_fd) and related code.
- Clean up deprecated code and variables associated with SCG reading - not
required anymore after b232be94ed
Remove L1 dependence from NR band number
L1 band information relied on get_band function that couldn't provide an
unambiguous result. Also there is no SCF compliant parameter for L1 and
generally we don't really need that information in L1.
The MR also remove a RU section that used nr_band but was never really
accessed in NR.
Fix error
openair1/SIMULATION/NR_PHY/dlsim.c:1057:37: runtime error: member access within misaligned address 0x7f8a67e01010 for type 'struct NR_Sched_Rsp_t', which requires 32 byte alignment
Prior to this commit, constants NB_ANTENNAS_RX/TX are provided on the
command line to each compiler invocation. This has the drawback (1) that
also files that don't need that get the constants, and (2) it is not
easy to find them using ctags or similar, as the only definition is in
CMakeLists.txt (where it is not found).
In this commit, use cmake configure_file() to write a header file with
these constants. The file will be generated during cmake invocation, and
written to <build-dir>/common/cmake_defs.h. The relevant files that
need this definition will pick it from the header.
This has been used previously also in a commit to write git information.
See-also: 8bffd1666d ("Avoid complete rebuild on cmake run with git info change")
This commit integrates the CUDA-based channel simulation pipeline into
the nr_dlsim.
A new `--cuda` command-line flag is added to enable the GPU path at
runtime. When activated, the simulator will:
- Allocate all necessary GPU and pinned host memory via the new
`init_cuda_chsim_buffers` helper function.
- In the main simulation loop, call the `run_channel_pipeline_cuda`
function to perform the channel simulation on the GPU, bypassing the
CPU-based functions.
- Clean up all CUDA resources on exit using the
`free_cuda_chsim_buffers` helper.
Additionally, dlsim is now using the float version of channel_multipath and add_noise functions.
Also, The input data is formatted as interleaved IQ, with separate antennas to ensure a fair comparison with the GPU operation.
Remove VCD signal dumper from NR code
We don't test VCD in 5G NR code (and haven't tested it in 4G for years).
Remove the VCD traces in 5G code to avoid any ambiguity. 4G code remains
in place, as well as the tools to analyze existing traces.
In this case, there are a number of linker errors:
/usr/bin/ld: libL2_NR.a(rrc_gNB.c.o): in function `nr_rrc_transfer_protected_rrc_message':
/home/richie/oai/openair2/RRC/NR/rrc_gNB.c:330:(.text+0x6f6): undefined reference to `signal_rrc_msg'
However, nr_srssim does not need RRC, while it is being linked in. Thus,
link without nr_rrc and L2. For the remaining two linker problems:
- provide a definition of NR_IF_Module_init() as in most other
simulators
- move rrc_get_max_nr_csrs() to nr_common.c. It is already declared in
the corresponding header, and itself does not have anything to do with
RRC radio configuration, therefore it does not belong to
nr_radio_config.c
For now, test in the following configuration:
- 16 parallel channels
- channel length of 32 and 128
- with 61440 and 122880 samples
- 12 trials per each test (due to CUDA kernel loading delay in first iterations)
- multiple MIMO configurations: 2TX - 2/4/8/32 RX, 2/4/8/32 TX - 2RX
This commit extends the CUDA API with new functions designed for
high-throughput scalability testing and advanced concurrency.
Key additions include:
- Batched versions of the kernels (`..._batched`) that use the
`blockIdx.z` grid dimension to process many independent channel
simulations in parallel.
- A C-callable wrapper, `run_channel_pipeline_cuda_batched`, to
orchestrate the execution of these new kernels.
- A "streamed" version of the pipeline,
`run_channel_pipeline_cuda_streamed`, which accepts a
`cudaStream_t` as an argument. This allows the calling
application to manage the execution stream, enabling overlap
with other GPU or CPU tasks.
This commit introduces the main C-callable pipeline function,
`run_channel_pipeline_cuda`, located in a new `channel_pipeline.cu`
file.
This function serves as the primary entry point for the GPU channel
simulation. It orchestrates the entire sequence of operations:
- Manages Host-to-Device data transfers for the input signal and
channel coefficients.
- Launches the `multipath_channel_kernel` and the
`add_noise_and_phase_noise_kernel` in the correct order.
- Manages the Device-to-Host data transfer for the final output signal.
- Synchronizes to ensure completion before returning.
This commit adds the CUDA implementation for adding Additive White
Gaussian Noise (AWGN) and phase noise to the signal. The code is
located in phase_noise.cu.
- The `add_noise_and_phase_noise_kernel` uses the NVIDIA cuRAND
library for efficient parallel random number generation.
- Helper functions are included to create, initialize, and destroy
the cuRAND states required by the kernel.
- Placeholder for a C-callable wrapper, `add_noise_cuda`, is also
provided.
This commit introduces the CUDA implementation of the multipath channel
model in `multipath_channel.cu`.
- Implements `multipath_channel_kernel` using shared memory for
efficient, high-performance convolution.
- Provides a C-callable wrapper, `multipath_channel_cuda`, to
orchestrate the full pipeline of data transfers and kernel
launches.
- The wrapper includes support for multiple memory models (Explicit
Copy, Unified Memory, ATS) via preprocessor directives to adapt to
different build configurations.
- Adds a `interleave_channel_output_cuda` utility to convert signal
data from the CPU's planar format to the GPU's interleaved format.
This commit refactors the CPU-based channel model for improved reusability and extends it with single-precision floating-point support.
This serves as a performance and correctness baseline for the new CUDA implementation.
Key changes include:
- The original double-based multipath_channel function is refactored to remove hardcoded antenna dimensions from its signature. It now accepts dynamically sized arrays, making it more flexible for use in simulators like dlsim and ulsim.
- A new multipath_channel_float function is introduced to operate on float data types, matching the GPU's native format.
- A corresponding add_noise_float function is also added to complete the floating-point CPU pipeline.
- The new float-based functions are designed to work with an interleaved IQ sample format ([I, Q, I, Q]), which is more suitable for vectorized (SSE) and GPU processing.
This commit adds the initial infrastructure for building accelerated simulation components.
- Adds the CUDA_ENABLE option and logic to find the CUDA toolkit.
- Adds the CHANNEL_SSE option to enable SSE-specific code paths.
- Sets required properties on the SIMU target to allow linking with CUDA-compiled objects.
- Creates the initial oai_cuda.h header for the public API.
semi-manual remove of not used #define commands
As i was upset to see all these #define with no code under, i made a
global pass to remove the obvious cases semi-manual remove of not used
bodies
manual script, then a lot of manual review
grep -IRs "#define" ../openair* ../common/* ../executables/ ../nfapi/ ../radio/ | \
fgrep -v '(' | fgrep -v '//' |awk '{print $2}' | egrep -v "#define *$n"| \
while read n ; do
grep -IRs $n ../openair* ../common/* ../executables/ ../nfapi/ ../radio/ | grep -v "#define *$n" | grep -q $n || echo $n;
done
will give a decent list of #define never used. Then, I think we should
keep lists of # define when it comes from standards, even if the code
doesn't use it right now
NR UE: add support for multiple RF frontends
This MR adds basic support for multiple RF-frontends at the UE.
There are multiple reasons why we might want this:
1. There might be different antennas needed for different bands (e.g. NTN or
TN antennas). These antennas might be connected to different RF-frontends.
In case of Handover, the UE needs to dynamically switch between these
RF-frontends.
2. Each rf-simulator connection carries baseband I/Q samples. To be able to
perform a handover between two cells at different frequencies, there must
be two rf-simulator connections, as the baseband channels should not
interfere and thus cannot simply be added.
3. When we want to support multiple UE simulation within one instance of
nr-uesoftmodem, and allow multiple cells and probably even handover between
these cells, we need either:
1. dedicated RF-frontends for each simulated UE, or
2. dedicated RF-frontends for each connected cell and dynamically assign
simulated UEs accordingly.
This MR implements what is needed for 1., 2. and 3.1.
Rename nr_slot_fep_ul to nr_symbol_fep_ul to relfect its function
Simplify logic so that the function does not require RU context and
the implementation is not dependent on the implementation of RU buffer management.
This makes the function easier to understand and moves the responsibility
to set the correct buffer offset on the caller.
Introduce SRS physical simulator and improvements
This MR consists of two parts:
- Introduces SRS physical simulator to test different configurations of
SRS
- Improvements in the SRS channel estimation (Addressing the comments
from old, unfinished MR !2875)
- rewrite nr_srs_channel_estimation() to loop over rx antennas and
antenna ports to be able to parallelize
- Improvements in timing_advance_offset and timing_advance_offset_nsec
computation
- Introduce SRS multi-symbol channel estimation
- The channel impulse response is estimated from the averaged
channel frequency response over multiple symbols
- The channel frequency estimates are left as they are to be able to
estimate doppler in the future
- Simplify SRS channel estimation to handle memory alignment in the
case of different SRS configurations (eg: comboffset)
To use multiple symbols when running gNB, the following parameters needs
to be edited in the function get_srs_resource()
- srs_res->resourceMapping.startPosition with a range from {0,1,..,5}
corresponds to the start symbol {13, 12, .. 8} with in a slot
- srs_res->resourceMapping.nrofSymbols indicates the number of symbols
{n1, n2, and n4}
Make sure the that start symbol + number of symbols is < 14
The timing results for SRS in develop and srs_phy_simulator are as follows:
Command used: sudo ./nr_ulsim -n1000 -s50 -S50 -R273 -W2 -y2 -z2 -E 1 -P
develop
|__ RX SRS time 124.94 us (1000 trials) (124.94 total [ms])
|__ Generate SRS sequence time 0.05 us (1000 trials) ( 0.05 total [ms])
|__ Get SRS signal time 19.41 us (1000 trials) ( 19.41 total [ms])
|__ SRS channel estimation time 89.11 us (1000 trials) ( 89.11 total [ms])
|__ SRS timing advance estimation time 13.10 us (1000 trials) ( 13.10 total [ms])
|__ SRS report TLV build time 2.80 us (1000 trials) ( 2.80 total [ms])
|__ SRS beam report build time 0.00 us ( 0 trials)
|__ SRS IQ matrix build time 2.73 us (1000 trials)
srs_phy_simulator
|__ RX SRS time 211.93 us (1000 trials) (211.93 total [ms])
|__ Generate SRS sequence time 0.04 us (1000 trials) ( 0.04 total [ms])
|__ Get SRS signal time 19.55 us (1000 trials) ( 19.55 total [ms])
|__ SRS channel estimation time 152.05 us (1000 trials) (152.05 total [ms])
|__ SRS timing advance estimation time 36.67 us (1000 trials) ( 36.67 total [ms])
|__ SRS report TLV build time 2.96 us (1000 trials) ( 2.96 total [ms])
|__ SRS beam report build time 0.00 us ( 0 trials)
|__ SRS IQ matrix build time 2.88 us (1000 trials)
Note that in srs_phy_simulator, to compute timing offset ns, an
oversampling factor of 2 is performed. Also, for peak detection, we
average the channel estimates for every port.
Update physim timing thresholds in gNB-N300-Timing-Phytest-LDPC
The lower bound for the tests are too tight and any slight improvement
in the timing can result in a test failure. For example, these tests
were failing due to the improvements from the MR !3718
The lower bound for the tests are too tight and any slight improvement in the timing can result in a test failure.
For example, these tests were failing due to the improvements from the MR 3718
Jaroslava Fiedlerova