GPU implementation of the STA algorithm on I/Q data

GPU computing is a new paradigm in high performance signal and image processing. Massive parallel processing offered by the GPUs provides high acceleration of computations when they are properly implemented. Ultrasound image reconstruction is one of these highly parallel classes of algorithms. Massive amount of multichannel input data and deterministic order of execution makes US applications good candidates for high performance GPU implementation. Our goal is to design a versatile ultrasound platform with GPU real-time processing. The project is based on a new system architecture allowing for bandwidth and scalability of processing power. We implemented and optimized SAFT image reconstruction algorithms on CUDA. The study shows that a single GTX-580 GPU card is capable of reconstructing the 128-channel I/Q data into 256×256 High Resolution Images (HRI) at a frame-rate of 44.64 fps which is an equivalent of 5700 Low Resolution Images (LRI) per second.