This paper introduces the concept of Geometry Compression, lowing 3D triangle data to be represented with a factor of 6 to times fewer bits than conventional techniques, with only slight los es in object quality. The technique is amenable to rapid decomp sion in both software and hardware implementations; if 3D rend ing hardware contains a geometry decompression unit, applicat geometry can be stored in memory in compressed format. Geo try is first represented as a generalized triangle mesh, a data s ture that allows each instance of a vertex in a linear stream to sp ify an average of two triangles. Then a variable length compress is applied to individual positions, colors, and normals. Delta com pression followed by a modified Huffman compression is used f positions and colors; a novel table-based approach is used for mals. The table allows any useful normal to be represented by 18-bit index, many normals can be represented with index deltas 8 bits or less. Geometry compression is a general space-time tr off, and offers advantages at every level of the memory/interco nect hierarchy: less storage space is needed on disk, less trans sion time is needed on networks.

Geometry compression

This paper describes a system architecture that supports realtime generation of complex images, efficient generation of extremely high-quality images, and a smooth trade-off between the two.Based on the paradigm of integration, the architecture extends a state-of-the-art rendering system with an additional high-precision image buffer. This additional buffer, called the Accumulation Buffer, is used to integrate images that are rendered into the framebuffer. While originally conceived as a solution to the problem of aliasing, the Accumulation Buffer provides a general solution to the problems of motion blur and depth-of-field as well.Because the architecture is a direct extension of current workstation rendering technology, we begin by discussing the performance and quality characteristics of that technology. The problem of spatial aliasing is then discussed, and the Accumulation Buffer is shown to be a desirable solution. Finally the generality of the Accumulation Buffer is explored, concentrating on its application to the problems of motion blur, depth-of-field, and soft shadows.

/pdf/the-accumulation-buffer-hardware-support-for-high-quality-450nv0mgog.pdf

The accumulation buffer: hardware support for high-quality rendering

This paper introduces the architecture and initial algorithms for Pixel-Planes 5, a heterogeneous multi-computer designed both for high-speed polygon and sphere rendering (1M Phong-shaded triangles/second) and for supporting algorithm and application research in interactive 3D graphics. Techniques are described for volume rendering at multiple frames per second, font generation directly from conic spline descriptions, and rapid calculation of radiosity form-factors. The hardware consists of up to 32 math-oriented processors, up to 16 rendering units, and a conventional 1280 × 1024-pixel frame buffer, interconnected by a 5 gigabit ring network. Each rendering unit consists of a 128 × 128-pixel array of processors-with-memory with parallel quadratic expression evaluation for every pixel. Implemented on 1.6 micron CMOS chips designed to run at 40MHz, this array has 208 bits/pixel on-chip and is connected to a video RAM memory system that provides 4,096 bits of off-chip memory. Rendering units can be independently reasigned to any part of the screen or to non-screen-oriented computation. As of April 1989, both hardware and software are still under construction, with initial system operation scheduled for fall 1989.

Pixel-planes 5: a heterogeneous multiprocessor graphics system using processor-enhanced memories

We describe PixelFlow, an architecture for high-speed image generation that overcomes the transformationand frame-buffer– access bottlenecks of conventional hardware rendering architectures. PixelFlow uses the technique of image composition: it distributes the rendering task over an array of identical renderers, each of which computes a fill-screen image of a fraction of the primitives. A high-performance image-composition network composites these images in real time to produce an image of the entire scene. Image-composition architectures offer performance that scales linearly with the number of renderers; there is no fundamental limit to the maximum performance achievable using this approach. A single PixelFlow renderer rasterizes up to 1.4 million triangles per second, and an n-renderer system can rasterize at up to n times this basic rate. PixelFlow performs antialiasing by supersampling. It supports defemed shading with separate hardware shaders that operate on composite images containing intermediate pixel data. PixelFlow shaders compute complex shading algorithms and procedural and image-based textures in real-time. The shading rate is independent of scene complexity. A Pixel Flow system can be coupled to a parallel supercomputer to serve as an immediatemode graphics server, or it can maintain a display list for retainedmode rendering. The PixelFlow design has been simulated extensively at high level. Custom chip design is underway. We anticipate a working system by late 1993. CR

/pdf/pixelflow-high-speed-rendering-using-image-composition-16abtykuw5.pdf

PixelFlow: high-speed rendering using image composition

We describe a machine learning technique for reconstructing image sequences rendered using Monte Carlo methods. Our primary focus is on reconstruction of global illumination with extremely low sampling budgets at interactive rates. Motivated by recent advances in image restoration with deep convolutional networks, we propose a variant of these networks better suited to the class of noise present in Monte Carlo rendering. We allow for much larger pixel neighborhoods to be taken into account, while also improving execution speed by an order of magnitude. Our primary contribution is the addition of recurrent connections to the network in order to drastically improve temporal stability for sequences of sparsely sampled input images. Our method also has the desirable property of automatically modeling relationships based on auxiliary per-pixel input channels, such as depth and normals. We show significantly higher quality results compared to existing methods that run at comparable speeds, and furthermore argue a clear path for making our method run at realtime rates in the near future.

Interactive reconstruction of Monte Carlo image sequences using a recurrent denoising autoencoder

Current affordable architectures for high-speed display of shaded 3D objects operate orders of magnitude too slowly. Recent advances in floating point chip technology have outpaced polygon fill time, making the memory access bottleneck between the drawing processor and the frame buffer the most significant factor to be accelerated. Massively parallel VLSI system have the potential to bypass this bottleneck, but to date only at very high cost. We describe a new more affordable VLSI solution. A pipeline of triangle processors rasterizes the geometry, then a further pipeline of shading processors applies Phong shading with multiple light sources. The triangle processor pipeline performs 100 billion additions per second, and the shading pipeline performs two billion multiplies per second. This allows 3D graphics systems to be built capable of displaying more than one million triangles per second. We show the results of an anti-aliasing technique, and discuss extensions to texture mapping, shadows, and environment maps.

Chris Duffy

Papers

The triangle processor and normal vector shader: a VLSI system for high performance graphics