Arbitrary waveform generation on a GPU using the additive synthesis framework for waveform synthesis.
This is a software achitecture for real-time arbitrary waveform generation based on a CUDA GPU and PCIe DAC module, allowing for a high throughput static waveform generation as well as a flexible and low-latecy computation of complex waveforms.
Currently this project is implemented with two pathways of generating dynamic waveforms with chirping tones, allowing the waveforms to be computed in advance or in parallel of streaming.
The project is implemented with a primitive TCP based control logic that is subject to improvement or replacement according to users' need.
The program only runs on a Linux system and requires a NVIDIA GPU and a PCIe interfaced DAC card both with GPUDirect RDMA support.
First install NVIDIA CUDA Tool Kit, CUDA Driver, and Open GPU Kernel Modules following guide here, or following the guide provided by your DAC vendor. Remember to disable IOMMU to avoid RDMA errors; you should be able to do that in your BIOS setting.
The DAC we used is theSpectrum Instrumentation M4i.6622-x8, whose driver includes are included in the folder spcm_header. The drivers and their installation guide could be found here. If you are using other DACs, you would need to configure the device handle hCard and the pinned buffer pvDMABuffer_gpu accordingly.
To compile the code, you may also need to g++ installed as the host compiler.
With the drivers and compilers verified, you should be able to compile with the code. You need to modify the Makefile to make sure: 1. The CUDA driver directory is set properly; 2. The source file name matches the code you would like to compile, which by default should be one of the waveform_synthesis*.cu files. Upon the successful compilation, an executable waveform_synthesis should appear in the code directory.
Currently there is no GUI interface. The parameters of generated waveform could be editted in parameters files. For real-time interface, the server in waveform_synthesis*.cu, or other kinds of interruption needs to be implemented.
The Nvidia NSight Profile of the cuda function in this program is placed in the benchmark folder; the code used for testing is located in the benchmark\test_setting folder.
Please see the paper correlated with this project. Please contanct Juntian Tu (juntian"at"umd.edu) for issues related to this repository.