High-Performance Multi-Mode Ptychography Reconstruction on Distributed GPUs

Abstract

Ptychography is an emerging imaging technique that is able to provide wavelength-limited spatial resolution from specimen with extended lateral dimensions. As a scanning microscopy method, a typical two-dimensional image requires a number of data frames. As a diffraction-based imaging technique, the realspace image has to be recovered through iterative reconstruction algorithms. Due to these two inherent aspects, a ptychographic reconstruction is generally a computation-intensive and timeconsuming process, which limits the throughput of this method. We report an accelerated version of the multi-mode difference map algorithm for ptychography reconstruction using multiple distributed GPUs. This approach leverages available scientific computing packages in Python, including mpi4py and PyCUDA, with the core computation functions implemented in CUDA C. We find that interestingly even with MPI collective communications, the weak scaling in the number of GPU nodes can still remain nearly constant. Most importantly, for realistic diffraction measurements, we observe a speedup ranging from a factor of 10 to 10 3 depending on the data size, which reduces the reconstruction time remarkably from hours to typically about 1 minute and is thus critical for real-time data processing and visualization.

Metrology of a Focusing Capillary Using Optical Ptychography

“…For each data frame, a array was cropped and fed to the ptychography reconstruction engine. The reconstruction was performed with the Difference Map algorithm [ 19 ] implemented with multiple GPU acceleration [ 20 ]. It took 31 seconds to run 500 iterations.…”

PyNX: high-performance computing toolkit for coherent X-ray imaging based on operators

“…The technique relies on the coherent scattering of an extended object with a shifting illumination. Exploitation of the redundancy of the overlapped illuminated areas yields robust reconstructions with a variety of algorithms (Thibault et al, 2009;Maiden & Rodenburg, 2009;Qian et al, 2014;Marchesini, Tu & Wu, 2016;Odstrčil et al, 2018), for which several software packages are available (Nashed et al, 2014;Marchesini, Krishnan et al, 2016;Enders & Thibault, 2016;Dong et al, 2018).…”

“…However that is quickly changing with new synchrotron sources, which can produce a much higher coherent flux (Johansson et al, 2016;Raimondi, 2016;Schroer et al, 2016). It is therefore useful to exploit multiple GPU or computing nodes (Nashed et al, 2014;Enders & Thibault, 2016;Dong et al, 2018) to handle larger ptychography data sets.…”

“…Data analysis often remains a bottleneck for these experiments, either from the complexity of the algorithms or simply because of the computing requirements. A variety of software is readily available for ptychography (Nashed et al, 2014(Nashed et al, , 2017Enders & Thibault, 2016;Marchesini, Krishnan et al, 2016;Odstrčil et al, 2018;Dong et al, 2018;Wakonig et al, 2020), and less for CDI (Maia et al, 2010;Newton et al, 2012). Several limitations remain: (i) software packages are not always publicly distributed, (ii) high-performance computing -generally based on graphical processing units (GPUs) -is not always available or can be complicated to set up, and (iii) the software can be difficult to maintain or improve owing to the complexity of algorithms or their GPU implementation.…”

Ptychography with multiple wavelength illumination

“…nonlinear optimization methods [34,35] and momentum-based methods [36]) or take an advantage of modern computational devices (e.g. GPU-based parallel computing [37][38][39]). In our numerical experiments, 1000 iterations were applied for each single simulation to ensure the convergence.…”