cuSat

GPU-accelerated DVB-S2/S2X software transceiver in C++17/CUDA.

The self-contained DSP library (lib/cusat_dsp/) provides a CUDA pipeline and standalone applications for bidirectional links on NVIDIA GPUs and UHD radios.

Start with the documentation: prepare each Orin, Thor, or Spark Docker host once, then use the two-radio CCM guide. ACM/orbital channels, OAI backhaul, and OpenShift each have a separate operator guide.

Features

• All 28 DVB-S2 MODCODs (QPSK 1/4 through 32APSK 9/10) in normal and short FECFRAMEs, plus 38 DVB-S2X normal-FECFRAME MODCODs (QPSK 13/45 through 256APSK 3/4, including the -L variants)
• Full-GPU TX pipeline: BCH + LDPC (IRA) encode, bit interleave, constellation mapping, PL framing/scrambling, RRC pulse shaping — batched with double buffering
• Full-GPU RX pipeline: AGC + rotation, polyphase symbol timing recovery, PLHEADER search, pilot-aided phase tracking, batched soft demapping, LDPC decode, BCH syndrome check
• ACM (adaptive coding and modulation) with in-band QPSK 1/4 signaling, per-MODCOD SNR thresholds, FER feedback
• GSE encapsulation with TAP networking (IP over DVB-S2)
• Configurable symbol rates: TX Farrow resampling to the hardware rate; RX processes at the hardware rate directly via a fractional-stride timing NCO
• Pilot-based SNR estimation (modcod-independent, accurate down to LDPC threshold)
• YAML configuration with CLI override
• TX chain bit-exact against the DVB.org DVB-S2X V&V test vectors at every tap through the transmitted PLFRAME (dvbs2x-vv-check, scripts/run_vv.sh; the vectors are distributed by DVB and not included)

Docker deployment

The supported physical-host workflow requires Docker Engine, the Docker Compose plugin, the NVIDIA Container Toolkit, a supported NVIDIA GPU, and a UHD-supported USRP accessible from the host. OTA validation currently covers USRP X410s; other UHD radios require device-specific channel, rate, and transport configuration. Compiler, CUDA toolkit, UHD, Boost, and CMake dependencies are installed and pinned inside the image build.

Prepare each host through its platform guide; each guide installs the reviewed kernel/IRQ profile, builds cusat:local, and runs the container tests:

• Jetson Orin NX
• Jetson Thor
• DGX Spark

After preparing any two hosts, the CCM guide assigns node0/node1 roles and uses the same Compose command on both.

OpenShift on GH200

From a workstation authenticated with oc, build the exact local checkout and deploy two long-running GH200/USRP test pods. The binary build avoids repository credentials and private branch names. The cluster must provide an ARM64 CUDA/UHD 4.8 base image and a network attachment for the radios; both are passed to the release script:

BUILD_NS=my-builds RUN_NS=my-runtime NAME=cusat \
RADIO_NETWORK=my-radio-network BASE_IMAGE=<cuda-uhd-image> \
  ./deployment/cusat/kubernetes/release.sh

Applications

Binary	Purpose
dvbs2-uhd-trx	Full FDD transceiver on one USRP: simultaneous TX+RX, ACM, TAP networking
dvbs2-gse-tx	PDU/GSE → IQ samples (stdin/stdout pipe transmitter)
dvbs2-gse-rx	IQ samples → PDU/GSE (stdin/stdout pipe receiver)
dvbs2-gse-acm	Launches dvbs2-uhd-trx in ACM mode over the TCP radio driver (no hardware)
bench-ldpc-decode	Standalone LDPC decoder throughput benchmark
bench-coded-modulation	Coded-modulation AWGN FER curves through the real mapper, demapper and decoder
bench-timing-recovery	Host launch-cost micro-benchmark of the GPU timing recovery
iq-farrow-resample	Resample an IQ file between samples-per-symbol ratios
dvbs2x-vv-check	Conformance check of the TX chain against a DVB.org V&V test case

Transceiver (USRP)

Docker Compose starts /build/apps/dvbs2-uhd-trx with the reviewed local YAML mounted at /etc/cusat/node.yaml. With a TAP configured, IP packets are GSE-encapsulated and transmitted; received frames are decapsulated onto the peer TAP. The CCM guide explains the primary radio, RF-chain, rate, MODCOD, gain, and channel-model fields before first launch.

Testing

• sudo docker run --rm --runtime=nvidia cusat:local ctest --test-dir /build --output-on-failure — unit tests in the release image
• scripts/loopback-test.sh — GPU file loopback (dvbs2-gse-tx → dvbs2-gse-rx)
• scripts/run_vv.sh <build> <vv-dir> — conformance against the DVB.org DVB-S2X V&V package (obtained from DVB)
• scripts/check_ldpc_tables.py — LDPC table checksums; --etsi verifies them against the ETSI PDFs

Signal Flow

TX:  PDUs (TAP/stdin)
       → GSE encapsulation → BBHEADER + BB scramble       [CPU]
       → BCH encode (GPU ≥16-frame batches, else CPU)
       → LDPC encode → bit interleave + constellation map
         → PL frame + pilots + PL scramble → RRC ×SPS     [GPU, fused batch]
       → Farrow resample to hardware rate                 [GPU]
       → UHD send

RX:  UHD recv → SPSC ring (2 s)
       → AGC + frequency rotation                         [GPU]
       → symbol timing recovery at hardware rate          [GPU]
         (fractional-stride NCO; polyphase energy search)
       → PLHEADER search (acquisition)                    [GPU]
       → PLS decode, pilot SNR/phase tracking             [CPU]
       → soft demap + deinterleave (int16 LLRs)           [GPU]
       → LDPC decode                                      [GPU]
       → BCH syndrome check                               [GPU]
       → BB descramble + deheader → GSE decap → TAP       [CPU]

Project Structure

apps/
  dvbs2-uhd-trx.cc        # UHD FDD transceiver (TX+RX, ACM, TAP)
  dvbs2-gse-tx.cc         # Pipe transmitter
  dvbs2-gse-rx.cc         # Pipe receiver
  bench_ldpc_decode.cc    # LDPC benchmark
  dvbs2x-vv-check.cc      # DVB.org V&V conformance check
  config_yaml.h           # YAML config parser
  sidecar.py              # Serves the --spectrum FIFO and status over HTTP
lib/cusat_dsp/
  cuda/                   # All CUDA kernels (TX pipeline, timing, header search, demap, LDPC, BCH, ...)
  cuda/ldpc_schedules/    # Conflict-free LDPC layer schedules (generated, scripts/gen_blossom_schedules.py)
  ldpc_tables/            # LDPC parity-check tables (ETSI EN 302 307-1/-2 annexes)
  framing/                # PLFRAME layout, PLHEADER/PLS code, PL scrambling
  modulation/             # Constellations, bit interleaver, data-aided SNR
  fec/                    # BCH code, CRC-8, code dimensions
  plrx/                   # PL receiver: frame sync, PLS, carrier/phase, metrics
  receiver_frontend.*     # RX orchestration: AGC, timing, header search, PL receiver
  transmitter_frontend.*  # TX orchestration
  tx_fec_chain.* / rx_fec_chain.*  # FEC chains (GPU batching, ACM MODCOD pools)
  acm_config.h / acm_controller.h  # MODCOD table + ACM decision logic
  gse.* / bb_*.{h,cc}     # GSE encapsulation, BBFRAME handling
lib/qa_*.cc               # Unit tests (Boost.UTF)
configs/                  # Example node0/node1 YAML configurations
scripts/                  # Setup, loopback and conformance helpers
deployment/               # Docker and Kubernetes/OpenShift workflows
docs/                     # Operator guides and implementation notes

License

Licensed under the Apache License 2.0. See LICENSE.