Search

We propose a system for dynamic triangle mesh processing entirely on the GPU. Our system offers an efficient data structure that allows fast updates of the underlying mesh connectivity and attributes. Our data structure partitions the mesh into small patches which allows processing all dynamic updates for each patch within the GPU's fast shared memory. This allows us to rely on speculative processing for conflict handling, which has low rollback cost while maximizing parallelism and reducing the cost of locking. Our system also introduces a new programming model for dynamic mesh processing. The programming model offers concise semantics for dynamic updates, relieving the user from having to worry about conflicting updates in the context of parallel execution. Our programming model relies on the cavity operator, which is a general mesh update operator that formulates any dynamic operation as an element reinsertion by removing a set of mesh elements and inserting others in the created void. We used our system to implement Delaunay edge flips and isotropic remeshing applications on the GPU. Our system achieves a 3—18x speedup on large models compared to multithreaded CPU solutions. Despite our additional dynamic features, our data structure also outperforms state-of-the-art GPU static data structures in terms of speed and memory requirements.

BackgroundGene expression may be regulated by the DNA methylation of regulatory elements in cis, distal, and trans regions. One method to evaluate the relationship between DNA methylation and gene expression is the mapping of expression quantitative trait methylation (eQTM) loci (also called expression associated CpG loci, eCpG). However, no open-source tools are available to provide eQTM mapping. In addition, eQTM mapping can involve a large number of comparisons which may prevent the analyses due to limitations of computational resources. Here, we describe Torch-eCpG, an open-source tool to perform eQTM mapping that includes an optimized implementation that can use the graphical processing unit (GPU) to reduce runtime.ResultsWe demonstrate the analyses using the tool are reproducible, up to 18 × faster using the GPU, and scale linearly with increasing methylation loci.ConclusionsTorch-eCpG is a fast, reliable, and scalable tool to perform eQTM mapping. Source code for Torch-eCpG is available at https://github.com/kordk/torch-ecpg .

X-ray-based computed tomography is a well established technique for determining the three-dimensional structure of an object from its two-dimensional projections. In the past few decades, there have been significant advancements in the brightness and detector technology of tomography instruments at synchrotron sources. These advancements have led to the emergence of new observations and discoveries, with improved capabilities such as faster frame rates, larger fields of view, higher resolution and higher dimensionality. These advancements have enabled the material science community to expand the scope of tomographic measurements towards increasingly in situ and in operando measurements. In these new experiments, samples can be rapidly evolving, have complex geometries and restrictions on the field of view, limiting the number of projections that can be collected. In such cases, standard filtered back-projection often results in poor quality reconstructions. Iterative reconstruction algorithms, such as model-based iterative reconstructions (MBIR), have demonstrated considerable success in producing high-quality reconstructions under such restrictions, but typically require high-performance computing resources with hundreds of compute nodes to solve the problem in a reasonable time. Here, tomoCAM, is introduced, a new GPU-accelerated implementation of model-based iterative reconstruction that leverages non-uniform fast Fourier transforms to efficiently compute Radon and back-projection operators and asynchronous memory transfers to maximize the throughput to the GPU memory. The resulting code is significantly faster than traditional MBIR codes and delivers the reconstructive improvement offered by MBIR with affordable computing time and resources. tomoCAM has a Python front-end, allowing access from Jupyter-based frameworks, providing straightforward integration into existing workflows at synchrotron facilities.

Nearly 20 years after the birth of general-purpose GPU computing, the HPC landscape is now dominated by GPUs. After years of undisputed dominance by NVIDIA, new players have entered the arena in a convincing manner, namely AMD and more recently Intel, whose devices currently power the first two clusters in the Top500 ranking. Unfortunately, code porting is still a major problem, even more due to the presence of different vendors, but at the same time the emergence of simplified standard paradigms suggests an encouraging prospect for developers. In this work, we provide a detailed OpenMP porting strategy of STREAmS, a community code for the compressible fluid dynamics. The proposed porting technique is based on the offload functionality of the OpenMP 5.x paradigm and in particular on a hybrid directives/APIs approach that fits seamlessly into the multi-backend software ecosystem of STREAmS. We further carry out a comprehensive performance analysis on the Intel® Data Center GPU Max 1550 (formerly called Ponte Vecchio or PVC). In addition, we analyze the performance of the code on two benchmark clusters powered by PVC, including the exascale Aurora cluster. The performance is evaluated at different levels of parallelism involved, i.e., the intrinsic parallelism of the PVC tile, the inter-tile parallelism within the GPU configuration, between the GPUs within the node and between the nodes within the cluster. The analysis shows that although the implementation complexity of the OpenMP porting is limited, it is necessary to follow some important guidelines to achieve satisfactory performance. The PVC GPU shows about 40% higher performance than the NVIDIA A100 or AMD MI250X GPUs, which, however, were released about 3 years earlier. Both intra-node and internode scalability show good results. Overall, the introduction of PVC into the GPU computing HPC landscape represents a positive step forward for the diversification and competitiveness of the sector. [ABSTRACT FROM AUTHOR]

This paper presents a novel implementation of the Coral Reef Optimization with Substrate Layers (CRO‐SL) algorithm. Our approach, which we call TensorCRO, takes advantage of the TensorFlow framework to represent CRO‐SL as a series of tensor operations, allowing it to run on GPU and search for solutions in a faster and more efficient way. We evaluate the performance of the proposed implementation across a wide range of benchmark functions commonly used in optimization research (such as the Rastrigin, Rosenbrock, Ackley, and Griewank functions), and we show that GPU execution leads to considerable speedups when compared to its CPU counterpart. Then, when comparing TensorCRO to other state‐of‐the‐art optimization algorithms (such as the Genetic Algorithm, Simulated Annealing, and Particle Swarm Optimization), the results show that TensorCRO can achieve better convergence rates and solutions than other algorithms within a fixed execution time, given that the fitness functions are also implemented on TensorFlow. Furthermore, we also evaluate the proposed approach in a real‐world problem of optimizing power production in wind farms by selecting the locations of turbines; in every evaluated scenario, TensorCRO outperformed the other meta‐heuristics and achieved solutions close to the best known in the literature. Overall, our implementation of the CRO‐SL algorithm in TensorFlow GPU provides a new, fast, and efficient approach to solving optimization problems, and we believe that the proposed implementation has significant potential to be applied in various domains, such as engineering, finance, and machine learning, where optimization is often used to solve complex problems. Furthermore, we propose that this implementation can be used to optimize models that cannot propagate an error gradient, which is an excellent choice for non‐gradient‐based optimizers. [ABSTRACT FROM AUTHOR]

The unprecedented increase in interconnected data has driven the development of efficient graph analytics for extensive data analysis, resulting in improvements across various domains. Prior work has focused on optimizing graph execution for both CPUs and GPUs while overlooking the scalability of graph applications and the selection of an ideal architecture. Thus, we propose Allok, a flexible graph processing framework that aids in selecting the optimal processing architecture (CPU or GPU) for a batch of graph applications while also optimizing number of threads on CPUs. Allok relies solely on high-level graph features to make decisions without the need for further application execution. Our experiments on an HPC system with 4 CPUs and 3 GPUs running 5 algorithms over 25 input graphs show that Allok reduces application execution time by an average of 67.26 × and reduces energy consumption and energy-delay product by an average of 18.06 × and 1237.96 × , respectively. [ABSTRACT FROM AUTHOR]

When dealing with large bipartite graphs, butterfly counting is a crucial and time-consuming operation. Graphics processing units (GPUs) are widely used parallel heterogeneous devices that can significantly boost performance for data science programs. However, currently no work enables efficient butterfly counting on GPU. To fill this gap, we propose a GPU-based butterfly counting method, called G-BFC. G-BFC solves three significant technical problems. First, butterfly counting involves massive serial operations, which leads to severe synchronization overheads and performance degradation. We unlock the serial region and utilize the shared memory on GPU to efficiently handle it. Second, butterfly counting on GPU faces the workload imbalance problem. To maximize efficiency, we develop a novel adaptive strategy to balance the workload among threads. Third, the large number of two-hop paths, also known as wedges, in bipartite graphs make parallel butterfly counting difficult to traverse. We develop an innovative preprocessing strategy that can significantly cut down on the required number of wedges. We conduct comprehensive experiments on both server-grade and edge-grade GPU platforms, and experiments show that G-BFC brings significant performance benefits. G-BFC achieves 4.84 × performance speedup over the state-of-the-art solution on eleven real-world datasets. [ABSTRACT FROM AUTHOR]

Meeting the rising global demand for healthcare diagnostic tools is crucial, especially with a shortage of medical professionals. This issue has increased interest in utilizing deep learning (DL) and telemedicine technologies. DL, a branch of artificial intelligence, has progressed due to advancements in digital technology and data availability and has proven to be effective in solving previously challenging learning problems. Convolutional neural networks (CNNs) show potential in image detection and recognition, particularly in healthcare applications. However, due to their resource-intensiveness, they surpass the capabilities of general-purpose CPUs. Therefore, hardware accelerators such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), and graphics processing units (GPUs) have been developed. With their parallelism efficiency and energy-saving capabilities, FPGAs have gained popularity for DL networks. This research aims to automate the classification of normal and abnormal (specifically Diabetic Foot Ulcer—DFU) classes using various parallel hardware accelerators. The study introduces two CNN models, namely DFU_FNet and DFU_TFNet. DFU_FNet is a simple model that extracts features used to train classifiers like SVM and KNN. On the other hand, DFU_TFNet is a deeper model that employs transfer learning to test hardware efficiency on both shallow and deep models. DFU_TFNet has outperformed AlexNet, VGG16, and GoogleNet benchmarks with an accuracy 99.81%, precision 99.38% and F1-Score 99.25%. In addition, the study evaluated two high-performance computing platforms, GPUs and FPGAs, for real-time system requirements. The comparison of processing time and power consumption revealed that while GPUs outpace FPGAs in processing speed, FPGAs exhibit significantly lower power consumption than GPUs. [ABSTRACT FROM AUTHOR]

Gene co-expression networks are valuable tools for discovering biologically relevant information within gene expression data. However, analysing large datasets presents challenges due to the identification of nonlinear gene–gene associations and the need to process an ever-growing number of gene pairs and their potential network connections. These challenges mean that some experiments are discarded because the techniques do not support these intense workloads. This paper presents pyEnGNet, a Python library that can generate gene co-expression networks in High-performance computing environments. To do this, pyEnGNet harnesses CPU and multi-GPU parallel computing resources, efficiently handling large datasets. These implementations have optimised memory management and processing, delivering timely results. We have used synthetic datasets to prove the runtime and intensive workload improvements. In addition, pyEnGNet was used in a real-life study of patients after allogeneic stem cell transplantation with invasive aspergillosis and was able to detect biological perspectives in the study. [ABSTRACT FROM AUTHOR]

Shear strain localization refers to the phenomenon of accumulation of material deformation in narrow slip zones. Many materials exhibit strain localization under different spatial and temporal scales, particularly rocks, metals, soils, and concrete. In the Earth's crust, irreversible deformation can occur in brittle as well as in ductile regimes. Modeling of shear zones is essential in the geodynamic framework. Numerical modeling of strain localization remains challenging due to the non‐linearity and multi‐scale nature of the problem. We develop a numerical approach based on graphical processing units (GPU) to resolve the strain localization in two and three dimensions of a (visco)‐hypoelastic‐perfectly plastic medium. Our approach allows modeling both the compressible and incompressible visco‐elasto‐plastic flows. In contrast to symmetric shear bands frequently observed in the literature, we demonstrate that using sufficiently small strain or strain rate increments, a non‐symmetric strain localization pattern is resolved in two‐ and three‐dimensions, highlighting the importance of high spatial and temporal resolution. We show that elasto‐plastic and visco‐plastic models yield similar strain localization patterns for material properties relevant to applications in geodynamics. We achieve fast computations using three‐dimensional high‐resolution models involving more than 1.3 billion degrees of freedom. We propose a new physics‐based approach explaining spontaneous stress drops in a deforming medium. Plain Language Summary: Strain localization is the accumulation of strain in narrow regions of rocks and other materials like metals, soils, and concrete, occurring at different scales. The strength of most geomaterials, particularly rocks, is strongly pressure‐dependent, with strength increasing with increasing pressure. We developed efficient numerical algorithms using High‐Performance Computing (HPC) and graphical processing units (GPUs) to model strain localization in 2D and 3D for applications in geodynamics and earthquake physics. Unlike previous models, our method reveals non‐symmetrical patterns by using very small strain increments, highlighting the need for high‐detail modeling. We found that elasto‐plastic and visco‐plastic models show similar strain patterns for relevant materials. Our method also achieves fast, detailed computations with over 1.3 billion variables and offers a new explanation for sudden stress drops in deforming materials. Key Points: We resolve material instability during deformation resulting in a non‐symmetric pattern of strain localizationWe demonstrate the similarity in patterns of strain localization between frictional and time‐dependent plasticity modelsWe achieve fast numerical simulations in high‐resolution model setups in three dimensions involving more than 500 million degrees of freedom [ABSTRACT FROM AUTHOR]

The paper describes a sparse direct solver for the linear systems that arise from the discretization of an elliptic PDE on a two-dimensional domain. The scheme decomposes the domain into thin subdomains, or “slabs” and uses a two-level approach that is designed with parallelization in mind. The scheme takes advantage of H 2 -matrix structure emerging during factorization and utilizes randomized algorithms to efficiently recover this structure. As opposed to multi-level nested dissection schemes that incorporate the use of H or H 2 matrices for a hierarchy of front sizes, SlabLU is a two-level scheme which only uses H 2 -matrix algebra for fronts of roughly the same size. The simplicity allows the scheme to be easily tuned for performance on modern architectures and GPUs. The solver described is compatible with a range of different local discretizations, and numerical experiments demonstrate its performance for regular discretizations of rectangular and curved geometries. The technique becomes particularly efficient when combined with very high-order accurate multidomain spectral collocation schemes. With this discretization, a Helmholtz problem on a domain of size 1000 λ × 1000 λ (for which N = 100 M ) is solved in 15 min to 6 correct digits on a high-powered desktop with GPU acceleration. [ABSTRACT FROM AUTHOR]

The Monte Carlo Computational Summit was held on the campus of the University of Notre Dame in South Bend, Indiana, USA on 25–26 October 2023. The goals of the summit were to discuss algorithmic and software alterations required for successfully porting respective code bases to exascale-class computing hardware, compare software engineering techniques used by various code teams, and consider the adoption of industry-standard benchmark problems to better facilitate code-to-code performance comparisons. Participants reported that identifying and implementing suitable Monte Carlo algorithms for GPUs continues to be a sticking point. They also report significant difficulty porting existing algorithms between GPU APIs (specifically Nvidia CUDA to AMD ROCm). To better compare code-to-code performance, participants decided to design a C5G7-like benchmark problem with a defined figure of merit, with the expectation of adding more benchmarks in the future. The participants also identified the need to explore the intermediate and long-term future of the Monte Carlo neutron transport community and how best to modernize and contextualize Monte Carlo as a useful tool in modern industry. Overall the summit was considered to be a success by the organizers and participants, and the group shared a strong desire for future, potentially larger, Monte Carlo summits. [ABSTRACT FROM AUTHOR]

The Intermittent Small Baseline Subset approach to Interferometric Synthetic Aperture Radar data was originally devised as a way to recover information from regions with intermittent coherence, making it particularly useful in agricultural regions or those featuring significant vegetation. However, as modern data products grow in size, the increased computational complexity that this methodology demands makes processing more daunting. Here, we present a solution: leveraging the Bifrost data processing framework and GPUs, we analyze Sentinel-1 data covering a large region of northern California and are able to achieve dramatic speed-ups on the order of 300–400 times faster than CPU-bound implementations of ISBAS, processing the entire dataset in only 5 h. [ABSTRACT FROM AUTHOR]

Field Programmable Gate Arrays (FPGAs) have witnessed an increase in space applications in the last years, mainly due to their cost-effective high-performances and flexibility. However, the susceptibility of these devices to radiation-induced effects when working in such an environment is well known. When common mitigation techniques are not sufficient to ensure the correct completion of a task, radiation-hardened FPGAs represent one of the most effective solutions. NanoXplore, in this context, is the first European developer of rad-hard FPGAs, which embed intrinsic high complexity in their architectures preventing the user from using or developing custom placement and routing algorithms. In this paper, we overcame these issues by proposing the first tool tailored to NanoXplore devices which allows the exploration of NanoXplore device architectures and routing of points through a Python interface. We developed a model that reflects the one used by the vendor, allowing the user to extract info about routes, nets and additional logic, otherwise unavailable. The tool also performs routing of points in the programmable logic, computing the optimal path. An implementation of the router on Graphic Processing Unit (GPU) is proposed to exploit the highly parallelizable nature of the problem. Finally, routing timing analyses on different benchmarks have been performed, improving the routing routine time. [ABSTRACT FROM AUTHOR]

Massive MIMO (Multiple Input Multiple Output) systems impose significant processing burdens along with strict latency requirements. The combination of large-scale antenna arrays and wide bandwidth requirements for next-generation wireless systems creates an exponential increase in frontend to backend data. Balancing the processing latency and reliability is critical for baseband processing tasks such as QAM detection. While linear detection algorithms have low computational complexity, their use in Massive MIMO scenario has heavy degradation in error performance. Nonlinear detection methods such as Maximum Likelihood and Sphere Decoding have good error performance, but they suffer from high, variable, and uncontrollable computational complexity. For such cases, the K-best QAM detection algorithm can provide required control over the system performance while maintaining near-ML error performance. In this paper, hard-output, as well as soft-output K-best QAM detection, is implemented in a CPU by utilizing the multiple cores combined with vector processing. Similarly, hard-output detection in a GPU is implemented by leveraging the SIMD (Single Instruction, Multiple Data) architecture and Warp-based execution model. The processing time per bit and the energy consumption per bit are compared for CPU and GPU implementations for QAM constellation density and MIMO array size. The GPU implementation shows up to 5× processing latency per bit improvement and up to 120× energy consumption per bit improvement over the CPU implementation for typical QAM constellations such as 4, 16, and 64 QAM. GPU implementation also shows up to 125× improvement over CPU implementation in energy consumption per bit for larger MIMO configurations such as 24 × 24 and 32 × 32. Finally, the soft-output detector is combined with a LDPC (Low-Density Parity Check) decoder to obtain the FER (Frame Error Rate) performance for CPU implementation. The FER is then combined with frame processing latency to form a Goodput metric to demonstrate the latency and reliability tradeoff. [ABSTRACT FROM AUTHOR]

This article provides an enhanced parallelization of the WSM7 microphysics scheme for the Weather Research and Forecasting Model (WRF). The parallelization is designed to maximize the utilization of a heterogeneous computing system consisting of CPUs, GPUs or both. Therefore the reference implementation of the WSM7 scheme is re-implemented for the heterogeneous execution model. For each time step, a dynamic load distribution is introduced which balances the computational load between the two components aiming for an overall minimum execution time. The evaluation of the parallelized implementation is done for a specific weather situation. Specifically, the precipitation of the low-pressure zone "Bernd" from July 2021 is simulated using an Intel Core i7-7700 CPU and a NVIDIA GTX 1070 GPU. The results show a speedup of up to 28.51 for the GPU version in comparison with the reference implementation. The heterogeneous dynamic load balancing increases the speedup achieved even further by introducing a distribution factor that is updated for each time step. [ABSTRACT FROM AUTHOR]

Sparse matrix-vector multiplication (SpMV) plays a critical role in a wide range of linear algebra computations, particularly in scientific and engineering disciplines. However, the irregular memory access patterns, extensive memory usage, high bandwidth requirements, and underutilization of parallelism hinder the computational efficiency of SpMV on GPUs. In this paper, we propose a novel approach called block-wise dynamic mixed-precision (BDMP) to address these challenges. Our methodology involves partitioning the original matrix into uniformly sized blocks, with each block's size determined by considering architectural characteristics and accuracy requirements. Additionally, we dynamically assign precision to each block using a precision selection method that takes into account the value distribution of the original sparse matrix. We develop two distinct SpMV computation algorithms for BDMP: BDMP-PBP (Precision-based partitioning) and BDMP-TCKI (Tailored compression and kernel implementation). BDMP-PBP partitions the matrix into two independent matrices for separate computations based on block precision, offering flexibility for integration with other optimization techniques. Meanwhile, BDMP-TCKI focuses on achieving significant thread-level parallelism and memory utilization by tailoring an appropriate compressed storage format and kernel implementation for each block. We compare BDMP with NVIDIA's cuSPARSE library and three state-of-the-art SpMV methods, including SELLP, MergeBase, and BalanceCSR, using matrices from the University of Florida's SuiteSparse dataset collection. BDMP-PBP and BDMP-TCKI show average speedups up to 2.64 × and 2.91 × on Turing RTX 2080Ti, and up to 2.99 × and 3.22 × on Ampere A100. The results demonstrate that BDMP enables the optimization of computation speed without compromising the precision necessary for reliable results. [ABSTRACT FROM AUTHOR]

Despite the increased brilliance of the new generation synchrotron sources, there is still a challenge with high‐resolution scanning of very thick and absorbing samples, such as a whole mouse brain stained with heavy elements, and, extending further, brains of primates. Samples are typically cut into smaller parts, to ensure a sufficient X‐ray transmission, and scanned separately. Compared with the standard tomography setup where the sample would be cut into many pillars, the laminographic geometry operates with slab‐shaped sections significantly reducing the number of sample parts to be prepared, the cutting damage and data stitching problems. In this work, a laminography pipeline for imaging large samples (>1 cm) at micrometre resolution is presented. The implementation includes a low‐cost instrument setup installed at the 2‐BM micro‐CT beamline of the Advanced Photon Source. Additionally, sample mounting, scanning techniques, data stitching procedures, a fast reconstruction algorithm with low computational complexity, and accelerated reconstruction on multi‐GPU systems for processing large‐scale datasets are presented. The applicability of the whole laminography pipeline was demonstrated by imaging four sequential slabs throughout an entire mouse brain sample stained with osmium, in total generating approximately 12 TB of raw data for reconstruction. [ABSTRACT FROM AUTHOR]

The world is becoming increasingly dependant in computing intensive applications. The appearance of new paradigms, such as Internet of Things (IoT), and advances in technologies such as Computer Vision (CV) and Artificial Intelligence (AI) are creating a demand for high-performance applications. In this regard, Graphics Processing Units (GPUs) have the ability to provide better performance by allowing a high degree of data parallelism. These devices are also beneficial in specialized fields of manufacturing industry such as CAD/CAM. For all these applications, there is a recent tendency to offload these computations to the Cloud, using a computing offloading Cloud architecture. However, the use of GPUs in the Cloud presents some inefficiencies, where GPU virtualization is still not fully resolved, as our research on what main Cloud providers currently offer in terms of GPU Cloud instances shows. To address these problems, this paper first makes a review of current GPU technologies and programming techniques that increase concurrency, to then propose a Cloud computing outsourcing architecture to make more efficient use of these devices in the Cloud. [ABSTRACT FROM AUTHOR]

This study focuses on training a custom, small Convolutional Neural Network (CNN) using a limited dataset through data augmentation that is aimed at developing weights for subsequent fine-tuning on specific defects, namely improperly polished aluminum surfaces. The objective is to adapt the network for use in computationally restricted environments. The methodology involves using two computers—a low-performance PC for network creation and initial testing and a more powerful PC for network training using the Darknet framework—after which the network is transferred back to the initial low-performance PC. The results demonstrate that the custom lightweight network suited for a low-performance PC effectively performs object detection under the described conditions. These findings suggest that using tailored lightweight networks for recognizing specific types of defects is feasible and warrants further investigation to enhance the industrial defect detection processes in limited computational settings. This approach highlights the potential for deploying AI-driven quality control in environments with constrained hardware capabilities. [ABSTRACT FROM AUTHOR]

Tomographic imaging has benefited from advances in X-ray sources, detectors and optics to enable novel observations in science, engineering and medicine. These advances have come with a dramatic increase of input data in the form of faster frame rates, larger fields of view or higher resolution, so high performance solutions are currently widely used for analysis. Tomographic instruments can vary significantly from one to another, including the hardware employed for reconstruction: from single CPU workstations to large scale hybrid CPU/GPU supercomputers. Flexibility on the software interfaces and reconstruction engines are also highly valued to allow for easy development and prototyping. This paper presents a novel software framework for tomographic analysis that tackles all aforementioned requirements. The proposed solution capitalizes on the increased performance of sparse matrix-vector multiplication and exploits multi-CPU and GPU reconstruction over MPI. The solution is implemented in Python and relies on CuPy for fast GPU operators and CUDA kernel integration, and on SciPy for CPU sparse matrix computation. As opposed to previous tomography solutions that are tailor-made for specific use cases or hardware, the proposed software is designed to provide flexible, portable and high-performance operators that can be used for continuous integration at different production environments, but also for prototyping new experimental settings or for algorithmic development. The experimental results demonstrate how our implementation can even outperform state-of-the-art software packages used at advanced X-ray sources worldwide.

Despite the increased brilliance of the new generation synchrotron sources, there is still a challenge with high-resolution scanning of very thick and absorbing samples, such as a whole mouse brain stained with heavy elements, and, extending further, brains of primates. Samples are typically cut into smaller parts, to ensure a sufficient X-ray transmission, and scanned separately. Compared with the standard tomography setup where the sample would be cut into many pillars, the laminographic geometry operates with slab-shaped sections significantly reducing the number of sample parts to be prepared, the cutting damage and data stitching problems. In this work, a laminography pipeline for imaging large samples (>1 cm) at micrometre resolution is presented. The implementation includes a low-cost instrument setup installed at the 2-BM micro-CT beamline of the Advanced Photon Source. Additionally, sample mounting, scanning techniques, data stitching procedures, a fast reconstruction algorithm with low computational complexity, and accelerated reconstruction on multi-GPU systems for processing large-scale datasets are presented. The applicability of the whole laminography pipeline was demonstrated by imaging four sequential slabs throughout an entire mouse brain sample stained with osmium, in total generating approximately 12 TB of raw data for reconstruction.