Recent advances in Deep Neural Networks (DNNs) have led to active development of specialized DNN accelerators, many of which feature a large number of processing elements laid out spatially, together with a multi-level memory hierarchy and flexible interconnect. While DNN accelerators can take advantage of data reuse and achieve high peak throughput, they also expose a large number of runtime parameters to the programmers who need to explicitly manage how computation is scheduled both spatially and temporally. In fact, different scheduling choices can lead to wide variations in performance and efficiency, motivating the need for a fast and efficient search strategy to navigate the vast scheduling space.To address this challenge, we present CoSA, a constrained-optimization-based approach for scheduling DNN accelerators. As opposed to existing approaches that either rely on designers’ heuristics or iterative methods to navigate the search space, CoSA expresses scheduling decisions as a constrained-optimization problem that can be deterministically solved using mathematical optimization techniques. Specifically, CoSA leverages the regularities in DNN operators and hardware to formulate the DNN scheduling space into a mixed-integer programming (MIP) problem with algorithmic and architectural constraints, which can be solved to automatically generate a highly efficient schedule in one shot. We demonstrate that CoSA-generated schedules significantly outperform state-of-the-art approaches by a geometric mean of up to 2.5× across a wide range of DNN networks while improving the time-to-solution by 90×.

Traditional geometry representations have focused on representing the details of the geometry in a deterministic fashion. In this article we propose a statistical representation of the geometry that leverages local coherence for very large datasets. We show how the statistical analysis of a densely sampled point model can be used to improve the geometry bandwidth bottleneck, both on the system bus and over the network as well as for randomized rendering, without sacrificing visual realism. Our statistical representation is built using a clustering-based hierarchical principal component analysis (PCA) of the point geometry. It gives us a hierarchical partitioning of the geometry into compact local nodes representing attributes such as spatial coordinates, normal, and color. We pack this information into a few bytes using classification and quantization. This allows our representation to directly render from compressed format for efficient remote as well as local rendering. Our representation supports both view-dependent and on-demand rendering. Our approach renders each node using quasi-random sampling utilizing the probability distribution derived from the PCA analysis. We show many benefits of our approach: (1) several-fold improvement in the storage and transmission complexity of point geometry; (2) direct rendering from compressed data; and (3) support for local and remote rendering on a variety of rendering platforms such as CPUs, GPUs, and PDAs.

We present a novel rendering primitive that combines the modeling brevity of points with the rasterization efficiency of polygons. The surface is represented by a sampled collection of Differential Points (DP), each with embedded curvature information that captures the local differential geometry in the vicinity of that point. This is a more general point representation that, for the cost of a few additional bytes, packs much more information per point than the traditional point-based models. This information is used to efficiently render the surface as a collection of local geometries. To use the hardware acceleration, the DPs are quantized into 256 different types and each sampled point is approximated by the closest quantized DP and is rendered as a normal-mapped rectangle. The advantages to this representation are: 1) The surface can be represented more sparsely compared to other point primitives, 2) it achieves a robust hardware accelerated per-pixel shading - even with no connectivity information, and 3) it offers a novel point-based simplification technique that factors in the complexity of the local geometry. The number of primitives being equal, DPs produce a much better quality of rendering than a pure splat-based approach. Visual appearances being similar, DPs are about two times faster and require about 75 percent less disk space in comparison to splatting primitives.