CUDA Range-Doppler Processing

This project provides CUDA-based source code to process radar data from an FMCW-radar into range and range-Doppler maps. Furthermore, the project includes various benchmarks that measure the execution time of the present algorithms, the used CUDA-kernel, 1D FFTs based on the cufft library and memory bandwidth.

Getting Started

To get started with the project some requirements are necessary to install.

Requirements

Just tested in Ubuntu 18.04 LTS
CUDA 9.1 or later versions and a CUDA capable GPU - installation guide for the latest version
GCC/G++ Compiler 6.3 or later versions - installation guide
CSNET GPUJPEG library - source code and installation guide can be found here

Installing

If the requirements are met, the project can be installed as follows.

1.) If the project is not downloaded yet, open a terminal and point to a desired folder ($DIR), otherwise proceed with step 3.)

cd $DIR

2.) clone the latest release from github

git clone https://github.com/NiclasEsser1/CUDARangeDopplerProcessing.git

3.) Point the console to the root folder of the project

cd CUDARangeDopplerProcessing

4.) Compile & build

make

The project should be compiled. If not, take a closer look into makefile and verify that the paths of libraries and header are adjusted to your system.

Running

The project provides an example for offline processing bin/offline_processing. If the file is executed, processing configurations must be passed to the program. First, a raw data file must be specified. To verify the installation process, a sample file is given.

$DIR/samples/raw_sample.dat

After you passed the remaining processing arguments, the algorithm is executed on the GPU. You can find the results of the processing in $DIR/results/img/

Read-out your own raw data

The raw data file corresponds to the default format shown in the Figure below.

If your raw data does not match this format, feel free to add your own read-out method (e.g SignalGenerator::loadMyFile).

Benchmarks

The project also contains different benchmarks:

Benchmark the algorithm (results are displayed within the terminal)

$DIR/bin/benchmark_algorithm

Benchmark single CUDA kernels (results can be found in $DIR/results/benchmarks/benchmark_kernels.csv)

$DIR/bin/benchmark_kernels

Benchmark memory bandwidth (results can be found in $DIR/results/benchmarks/benchmark_memory_bandwidth.csv)

$DIR/bin/benchmark_memory_bandwidth

Benchmark 1D FFTs (results can be found in $DIR/results/benchmarks/benchmark_fft.csv)

$DIR/bin/benchmark_fft

Algorithms Results (tested on Geforce 960m GTX)

width	height	Range-Doppler	Range
256	256	2.98 ms	2.90 ms
512	512	3.85 ms	3.02 ms
1024	1024	7.17 ms	4.9 ms
2048	2048	17.92 ms	10.34 ms
4096	4096	68.56 ms	40.192 ms
8192	4096	140.84 ms	82.14 ms

More Information

More information can be found in the header files.

Authors

Niclas Esser - Initial work - NiclasEsser1

See also the list of contributors who participated in this project.

License

No license

TODOs

Implement a test class which verifies calculations of GPU-Kernels
Optimize CudaBase::mapColors, CudaBase::fftshift, CudaBase::absolute and CudaBase::window
Add optimized matrix-vector and matrix-matrix multiplication
Commenting all classes and functions

NiclasEsser1 / CUDARangeDopplerProcessing