Gaussian Splatting has emerged as a prominent model for constructing 3D representations from images across diverse domains. However, the efficiency of the 3D Gaussian Splatting rendering pipeline relies on several simplifications. Notably, reducing Gaussian to 2D splats with a single viewspace depth introduces popping and blending artifacts during view rotation. Addressing this issue requires accurate per-pixel depth computation, yet a full per-pixel sort proves excessively costly compared to a global sort operation. In this paper, we present a novel hierarchical rasterization approach that systematically resorts and culls splats with minimal processing overhead. Our software rasterizer effectively eliminates popping artifacts and view inconsistencies, as demonstrated through both quantitative and qualitative measurements. Simultaneously, our method mitigates the potential for cheating view-dependent effects with popping, ensuring a more authentic representation. Despite the elimination of cheating, our approach achieves comparable quantitative results for test images, while increasing the consistency for novel view synthesis in motion. Due to its design, our hierarchical approach is only 4% slower on average than the original Gaussian Splatting. Notably, enforcing consistency enables a reduction in the number of Gaussians by approximately half with nearly identical quality and view-consistency. Consequently, rendering performance is nearly doubled, making our approach 1.6x faster than the original Gaussian Splatting, with a 50% reduction in memory requirements. Our renderer is publicly available at https://github.com/r4dl/StopThePop.

Convolutional neural network inference on video input is computationally expensive and requires high memory bandwidth. Recently, DeltaCNN [26] managed to reduce the cost by only processing pixels with significant updates over the previous frame. However, DeltaCNN relies on static camera input. Moving cameras add new challenges in how to fuse newly unveiled image regions with already processed regions efficiently to minimize the update rate - without increasing memory overhead and without knowing the camera extrinsics of future frames. In this work, we propose MotionDeltaCNN, a sparse CNN inference framework that supports moving cameras. We introduce spherical buffers and padded convolutions to enable seamless fusion of newly unveiled regions and previously processed regions – without increasing memory footprint. Our evaluation shows that we outperform DeltaCNN by up to 90% for moving camera videos.

Convolutional neural network inference on video data requires powerful hardware for real-time processing. Given the inherent coherence across consecutive frames, large parts of a video typically change little. By skipping identical image regions and truncating insignificant pixel updates, computational redundancy can in theory be reduced significantly. However, these theoretical savings have been difficult to translate into practice, as sparse updates hamper computational consistency and memory access coherence; which are key for efficiency on real hardware. With DeltaCNN, we present a sparse convolutional neural network framework that enables sparse frame-by-frame updates to accelerate video inference in practice. We provide sparse implementations for all typical CNN layers and propagate sparse feature updates end-to-end – without accumulating errors over time. DeltaCNN is applicable to all convolutional neural networks without retraining. To the best of our knowledge, we are the first to significantly outperform the dense reference, cuDNN, in practical settings, achieving speedups of up to 7x with only marginal differences in accuracy. Our CUDA kernels and PyTorch extensions can be found at https://github.com/facebookresearch/DeltaCNN.

Tracking body and hand motions in 3D space is essential for social and self-presence in augmented and virtual environments. Unlike the popular 3D pose estimation setting, the problem is often formulated as egocentric tracking based on embodied perception (e.g., egocentric cameras, handheld sensors). In this article, we propose a new data-driven framework for egocentric body tracking, targeting challenges of omnipresent occlusions in optimization-based methods (e.g., inverse kinematics solvers). We first collect a large-scale motion capture dataset with both body and finger motions using optical markers and inertial sensors. This dataset focuses on social scenarios and captures ground truth poses under self-occlusions and body-hand interactions. We then simulate the occlusion patterns in head-mounted camera views on the captured ground truth using a ray casting algorithm and learn a deep neural network to infer the occluded body parts. Our experiments show that our method is able to generate high-fidelity embodied poses by applying the proposed method to the task of real-time egocentric body tracking, finger motion synthesis, and 3-point inverse kinematics.

Bandwidth reduction of sparse matrices is used to reduce fill-in of linear solvers and to increase performance of other sparse matrix operations, e.g., sparse matrix vector multiplication in iterative solvers. To compute a bandwidth reducing permutation, Reverse Cuthill-McKee (RCM) reordering is often applied, which is challenging to parallelize, as its core is inherently serial. As many-core architectures, like the GPU, offer subpar single-threading performance and are typically only connected to high-performance CPU cores via a slow memory bus, neither computing RCM on the GPU nor moving the data to the CPU are viable options. Nevertheless, reordering matrices, potentially multiple times in-between operations, might be essential for high throughput. Still, to the best of our knowledge, we are the first to propose an RCM implementation that can execute on multicore CPUs and many-core GPUs alike, moving the computation to the data rather than vice versa.Our algorithm parallelizes RCM into mostly independent batches of nodes. For every batch, a single CPU-thread/a GPU thread-block speculatively discovers child nodes and sorts them according to the RCM algorithm. Before writing their permutation, we re-evaluate the discovery and build new batches. To increase parallelism and reduce dependencies, we create a signaling chain along successive batches and introduce early signaling conditions. In combination with a parallel work queue, new batches are started in order and the resulting RCM permutation is identical to the ground-truth single-threaded algorithm.We propose the first RCM implementation that runs on the GPU. It achieves several orders of magnitude speed-up over NVIDIA's single-threaded cuSolver RCM implementation and is significantly faster than previous parallel CPU approaches. Our results are especially significant for many-core architectures, as it is now possible to include RCM reordering into sequences of sparse matrix operations without major performance loss.

Dynamic memory management on GPUs is generally understood to be a challenging topic. On current GPUs, hundreds of thousands of threads might concurrently allocate new memory or free previously allocated memory. This leads to problems with thread contention, synchronization overhead and fragmentation. Various approaches have been proposed in the last ten years and we set out to evaluate them on a level playing field on modern hardware to answer the question, if dynamic memory managers are as slow as commonly thought of. In this survey paper, we provide a consistent framework to evaluate all publicly available memory managers in a large set of scenarios. We summarize each approach and thoroughly evaluate allocation performance (thread-based as well as warp-based), and look at performance scaling, fragmentation and real-world performance considering a synthetic workload as well as updating dynamic graphs. We discuss the strengths and weaknesses of each approach and provide guidelines for the respective best usage scenario. We provide a unified interface to integrate any of the tested memory managers into an application and switch between them for benchmarking purposes. Given our results, we can dispel some of the dread associated with dynamic memory managers on the GPU.

Abstract The recent research explosion around implicit neural representations, such as NeRF, shows that there is immense potential for implicitly storing high‐quality scene and lighting information in compact neural networks. However, one major limitation preventing the use of NeRF in real‐time rendering applications is the prohibitive computational cost of excessive network evaluations along each view ray, requiring dozens of petaFLOPS. In this work, we bring compact neural representations closer to practical rendering of synthetic content in real‐time applications, such as games and virtual reality. We show that the number of samples required for each view ray can be significantly reduced when samples are placed around surfaces in the scene without compromising image quality. To this end, we propose a depth oracle network that predicts ray sample locations for each view ray with a single network evaluation. We show that using a classification network around logarithmically discretized and spherically warped depth values is essential to encode surface locations rather than directly estimating depth. The combination of these techniques leads to DONeRF, our compact dual network design with a depth oracle network as its first step and a locally sampled shading network for ray accumulation. With DONeRF, we reduce the inference costs by up to 48× compared to NeRF when conditioning on available ground truth depth information. Compared to concurrent acceleration methods for raymarching‐based neural representations, DONeRF does not require additional memory for explicit caching or acceleration structures, and can render interactively (20 frames per second) on a single GPU.

Dynamic memory allocation on a single instruction, multiple threads architecture, like the Graphics Processing Unit (GPU), is challenging and implementation guidelines caution against it. Data structures must rise to the challenge of thousands of concurrently active threads trying to allocate memory. Efficient queueing structures have been used in the past to allow for simple allocation and reuse of memory directly on the GPU but do not scale well to different allocation sizes, as each requires its own queue.

Sparse general matrix-matrix multiplication on GPUs is challenging due to the varying sparsity patterns of sparse matrices. Existing solutions achieve good performance for certain types of matrices, but fail to accelerate all kinds of matrices in the same manner. Our approach combines multiple strategies with dynamic parameter selection to dynamically choose and tune the best fitting algorithm for each row of the matrix. This choice is supported by a lightweight, multi-level matrix analysis, which carefully balances analysis cost and expected performance gains. Our evaluation on thousands of matrices with various characteristics shows that we outperform all currently available solutions in 79% over all matrices with >15k products and that we achieve the second best performance in 15%. For these matrices, our solution is on average 83% faster than the second best approach and up to 25X faster than other state-of-the-art GPU implementations. Using our approach, applications can expect great performance independent of the matrices they work on.

Abstract In this paper, we introduce the CPatch, a curved primitive that can be used to construct arbitrary vector graphics. A CPatch is a generalization of a 2D polygon: Any number of curves up to a cubic degree bound a primitive. We show that a CPatch can be rasterized efficiently in a hierarchical manner on the GPU, locally discarding irrelevant portions of the curves. Our rasterizer is fast and scalable, works on all patches in parallel, and does not require any approximations. We show a parallel implementation of our rasterizer, which naturally supports all kinds of color spaces, blending and super‐sampling. Additionally, we show how vector graphics input can efficiently be converted to a CPatch representation, solving challenges like patch self intersections and false inside‐outside classification. Results indicate that our approach is faster than the state‐of‐the‐art, more flexible and could potentially be implemented in hardware.

Having a virtual body can increase embodiment in virtual reality (VR) applications. However, comsumer-grade VR falls short of delivering sufficient sensory information for full-body motion capture. Consequently, most current VR applications do not even show arms, although they are often in the field of view. We address this shortcoming with a novel human upper-body inverse kinematics algorithm specifically targeted at tracking from head and hand sensors only. We present heuristics for elbow positioning depending on the shoulder-to-hand distance and for avoiding reaching unnatural joint limits. Our results show that our method increases the accuracy compared to general inverse kinematics applied to human arms with the same tracking input. In a user study, participants preferred our method over displaying disembodied hands without arms, but also over a more expensive motion capture system. In particular, our study shows that virtual arms animated with our inverse kinematics system can be used for applications involving heavy arm movement. We demonstrate that our method can not only be used to increase embodiment, but can also support interaction involving arms or shoulders, such as holding up a shield.