Distributed Sparse Block Grids on GPUs

A portable C++ library for memory and compute abstraction on multi‐core CPUs and GPUs

“…The abstract data structures provided by openfpm_data are composable and can be used as building blocks for more complex data structures, such as distributed sparse block grids, 8 and for domain‐specific data structures 9 . The memory layout capabilities are inherited, as well as the memory double‐mapping and absorbing capabilities, allowing the same data structure to simultaneously be mapped to host and device.…”

“…A vector is a one-dimensional array, a Compressed Sparse Row (CSR) graph is stored in an encapsulated vector of vertices and edges, a map is a sorted vector, and a sparse grid is an n-dimensional map. 8 All sub-classes inherit the layout reconfigurability of the base class as defined by the four template parameters (distinguished by different colors) shown in the right box. Every container in the hierarchy can override every layout parameter, leading to a combinatorial diversity of possible implementations.…”

“…The UML diagram on the left shows the composition of the available containers, starting from the base class “multi‐dimensional array”. A vector is a one‐dimensional array, a Compressed Sparse Row (CSR) graph is stored in an encapsulated vector of vertices and edges, a map is a sorted vector, and a sparse grid is an $$$ n $$$‐dimensional map 8 . All sub‐classes inherit the layout reconfigurability of the base class as defined by the four template parameters (distinguished by different colors) shown in the right box.…”

“…All benchmarks are performed on the hardware and using the compilers listed in Table 3. Benchmarks for sparse data structures are available elsewhere 8 . We only benchmark the OpenMP, CUDA, and HIP backends of openfpm_data; SEQUENTIAL is always slower and only intended for debugging or porting purposes.…”

Morphogen gradients are regulated by porous media characteristics of the developing tissue

“…5a– e for an exemplary z-slice. The steps were: 1) extracting volumes of individual time points from the light-sheet video; 2) image processing and segmentation using Fiji [99] and ilastik [9]; 3) processing of the segmentation result and conversion to an indicator function; 4) computing a numerical surface description as a signed distance function (SDF) using the level-set method; 5) generating a geometry-adapted discretization of the ECS using distributed sparse block grids [47]. We next describe these five image-based modeling steps in more detail.…”