Modern throughput processors such as GPUs employ thousands of threads to drive high-bandwidth, long-latency memory systems. These threads require substantial on-chip storage for registers, cache, and scratchpad memory. Existing designs hard-partition this local storage, fixing the capacities of these structures at design time. We evaluate modern GPU workloads and find that they have widely varying capacity needs across these different functions.

We describe a decomposition for in-place matrix transposition, with applications to Array of Structures memory accesses on SIMD processors. Traditional approaches to in-place matrix transposition involve cycle following, which is difficult to parallelize, and on matrices of dimension m by n require O(mn log mn) work when limited to less than O(mn) auxiliary space. Our decomposition allows the rows and columns to be operated on independently during in-place transposition, reducing work complexity to O(mn), given O(max(m, n)) auxiliary space.

We introduce a new method for computing two-level Layered Depth Images (LDIs) [Shade et al. 1998] that is designed for modern GPUs. The method is order-independent, can guarantee a mini- mum depth separation between the layers, operates within small, bounded memory, and requires no explicit sorting. Critically, it also operates in a single pass over scene geometry.

In this paper, we develop a new parallel auxiliary grid algebraic multigrid (AMG) method to leverage the power of graphic processing units (GPUs). In the construction of the hierarchical coarse grid, we use a simple and fixed coarsening procedure based on a region quadtree generated from an auxiliary grid. This allows us to explicitly control the sparsity patterns and operator complexities of the AMG solver.

Compression of Bayer pattern color filter array (CFA) data has gained a lot of attention during past years. Numerous algorithms have been proposed for lossless, near-lossless and lossy compression. The performance evaluation of compression methods is typically done only for artificial CFA data, obtained by sub-sampling full color images according to CFA pattern, without taking into account that CFA data are heavily processed before obtaining full color images. Therefore, some assumptions that are true for reconstructed images may not hold for real raw data.

High-quality motion blur is an increasingly important and pervasive effect in interactive graphics that, even in the context of offline rendering, is often approximated using a post process. Recent motion blur post-process filters (e.g., [MHBO12, Sou13]) efficiently generate plausible results suitable for modern interactive rendering pipelines. However, these approaches may produce distracting artifacts, for instance, when different motions overlap in depth or when both large- and fine-scale features undergo motion.

High-speed signaling over high density interconnect on organic package substrates or silicon interposers offers an attractive solution to the off-chip bandwidth limitation problem faced in modern digital systems. In this paper, we describe a signaling system co-designed with the interconnect to take advantage of the characteristics of this environment to enable a high-speed, low area, and low-power die to die link.

Laminated packages, silicon interposer substrates, and special-purpose package- to-package interconnect, together with 3D stacking of silicon components enable systems with greatly improved computational power, memory capacity and bandwidth. These package options offer very high-bandwidth channels between chips on the same substrate. We employ ground-referenced single- ended signaling, a charge-pump transmitter, and a co-designed channel to provide communication between multiple chips on one package with high bandwidth per pin and low energy per bit.

We present PixelPie, a highly parallel geometric formulation of the Poisson-disk sampling problem on the graphics pipeline. Traditionally, generating a distribution by throwing darts and removing conflicts has been viewed as an inherently sequential process. In this paper, we present an efficient Poisson-disk sampling algorithm that uses rasterization in a highly parallel manner. Our technique is an iterative two step process. The first step of each iteration involves rasterization of random darts at varying depths. The second step involves culling conflicted darts.

Mobile computer-vision technology will soon become as ubiquitous as touch interfaces.