A from-scratch C++17 OFDM transceiver with QPSK / 16-QAM, AWGN, cyclic-prefix timing synchronization, and Rayleigh multipath experiments.
The complete communication chain is implemented in C++: bit generation, QAM mapping, custom FFT/IFFT, cyclic-prefix insertion/removal, noise generation, timing acquisition, demapping, and BER measurement. MATLAB reads exported C++ measurements for figures and numerical verification. Python summarizes and checks the measurement files. No MATLAB runtime or external communications library is needed to run the simulator.
C++ transmitter C++ receiver
Random bits → QPSK / 16-QAM → 64-point IFFT → CP → channel
↓
BER ← bit decisions ← QAM demapping ← 64-point FFT ← timing / CP removal
The base configuration uses 64 subcarriers, a 16-sample CP, normalized constellations, and a custom complex Gaussian noise generator. Timing acquisition evaluates a cyclic-prefix autocorrelation metric with a configurable accumulation length M. A separate extension implements continuous linear FIR Rayleigh channels and per-subcarrier zero-forcing equalization.
Requirements: a C++17 compiler and Make. Python 3 is optional for inspecting retained evidence; MATLAB is optional for reproducing figures. The Python helpers use the standard library only.
make all test test-multipath
./bin/ofdm --help
# Short C++ end-to-end run: both modulations, nine SNR points each.
./bin/ofdm smoke 262144 64The executable accepts ofdm [output_directory] [bits] [M]. Its default bit count is 33,554,432 per modulation/SNR point, and default M is 64. Bit counts must be at least 262,144 and divisible by 256.
The C++ tests cover direct-DFT agreement, mapping/demapping, CP samples, all 80 noiseless timing offsets, noise statistics, rolling/direct correlation agreement, frequency-offset phase invariance, invalid inputs, linear FIR boundaries, and zero-forcing behavior. The existing suites contain 4,717 OFDM checks and 892 multipath checks.
make sanitize sanitize-multipath # AddressSanitizer + UndefinedBehaviorSanitizerGitHub Actions builds and tests the C++ programs on Linux and macOS, runs a reduced-bit smoke simulation, and checks the retained data separately. A smoke run is not a full experiment rerun.
| Experiment | Configuration / retained evidence |
|---|---|
| AWGN BER | QPSK and 16-QAM, SNR 0:3:24 dB, 33,554,432 bits per point, paired M = 40 / 64 |
| Waveforms and constellations | Exported C++ samples at 3 and 15 dB for both modulations |
| Timing accumulation | Coarse M sweep, followed by every integer M from 32 to 64; 2,000 trials per modulation / SNR / M condition at 0, 3, and 15 dB |
| Rayleigh CP / equalization | Delay profiles with maximum delays 12 and 24; CP lengths 0, 8, 12, 16, 24, 32; 1,024 channel packets |
| Multipath timing | Paired M = 40 / 64 acquisition diagnostics on the same channel packets |
SNR in the base AWGN experiment is nominal frequency-domain symbol energy over noise density, Es/N0, rather than Eb/N0. Seeds, bit counts, measured noise statistics, timing estimates, and exact error counts are exported with the measurements.
M = 40 is a finite-dataset reliability / accumulation-cost tradeoff; it is not a universal optimum. M = 64 remains the executable's default and a comparison baseline. In the retained full BER sweeps, both selected the correct boundary and produced equal error counts; this does not establish equal timing reliability in other trials.
The Rayleigh CP / ZF BER comparison uses known timing and perfect CSI to isolate the effects of channel memory and CP length. The separate timing diagnostic does not establish a practical blind receiver with channel estimation. In multipath CSVs, snr_db = -1 is a sentinel for zero noise, not a measured -1 dB operating point. Zero observed BER remains zero in CSVs and is omitted on logarithmic plot lines.
Large raw timing CSVs are stored as lossless .csv.gz files. Expand them before running the summary helper; existing modified CSVs are protected from replacement.
python3 scripts/unpack_evidence.py
python3 scripts/summarize_results.pyTo rerun the C++ simulations, write to rerun/ to preserve the retained measurements:
./bin/ofdm rerun 33554432 64
./bin/ofdm rerun/m40 33554432 40
./bin/sync_study rerun/sync_selection_trials.csv selection 64 2000
./bin/sync_study rerun/sync_fine_trials.csv fine 64 2000
./bin/multipath_study rerun/multipath
./bin/timing_compare rerun/m40The full simulation commands can take substantially longer than the tests. Summary helpers and figure scripts read results/; to analyze a new full run, use a separate working copy with its own results/ dataset. Runtime fields can vary between machines.
To reproduce and verify all 14 retained figures in MATLAB R2024a (no Communications Toolbox needed):
addpath('matlab');
plot_ofdm_figures;
plot_multipath_figures;
verify_ofdm_figures;Figures are included as 300 dpi PNG, SVG, and editable MATLAB .fig files. The MATLAB verifier compares saved figure series directly with C++ measurement CSVs.
| Path | Contents |
|---|---|
src/ |
C++ FFT/IFFT, transceiver, timing sweeps, Rayleigh channel, and equalization |
tests/ |
Existing C++ numerical and signal-processing tests |
results/ |
C++ measurement CSVs, summaries, and compressed raw trials |
scripts/ |
Standard-library Python data expansion and consistency checks |
matlab/ |
Figure generation and saved-series verification |
figures_matlab/ |
PNG / SVG / FIG outputs |
The transform implementation is also presented independently in fft-ifft-cpp.
這是以 C++17 從零實作的 OFDM 基頻收發模擬器,涵蓋 QPSK / 16-QAM、64 點 FFT / IFFT、CP、AWGN、時間同步、BER 與 Rayleigh 多路徑實驗。核心通訊運算全由 C++ 完成;MATLAB 負責讀取實測資料、繪圖與核對曲線,Python 僅負責整理與檢查資料。C++ 程式可獨立編譯執行。
Author: Jin-Teng Sung · Joe05520


