Real-time GPU-based software beamformer designed for advanced imagingmethods research

Abstract

High computational demand is known to be a technical hurdle for real-timeimplementation of advanced methods like synthetic aperture imaging (SAI) andplane wave imaging (PWI) that work with the pre-beamform data of each arrayelement. In this paper, we present the development of a software beamformer forSAI and PWI with real-time parallel processing capacity. Our beamformer designcomprises a pipelined group of graphics processing units (GPU) that are hostedwithin the same computer workstation. During operation, each available GPU isassigned to perform demodulation and beamforming for one frame of pre-beamformdata acquired from one transmit firing (e.g. point firing for SAI). Tofacilitate parallel computation, the GPUs have been programmed to treat thecalculation of depth pixels from the same image scanline as a block ofprocessing threads that can be executed concurrently, and it would repeat thisprocess for all scanlines to obtain the entire frame of image data i.e.low-resolution image (LRI). To reduce processing latency due to repeated accessof each GPU's global memory, we have made use of each thread block's fast-sharedmemory (to store an entire line of pre-beamform data during demodulation),created texture memory pointers, and utilized global memory caches (to streamrepeatedly used data samples during beamforming). Based on this beamformerarchitecture, a prototype platform has been implemented for SAI and PWI, and itsLRI processing throughput has been measured for test datasets with 40 MHzsampling rate, 32 receive channels, and imaging depths between 5-15 cm. Whenusing two Fermi-class GPUs (GTX-470), our beamformer can compute LRIs of512-by-255 pixels at over 3200 fps and 1300 fps respectively for imaging depthsof 5 cm and 15 cm. This processing throughput is roughly 3.2 times higher than aTesla-class GPU (GTX-275). © 2010 IEEE.