Abstract: A method, computer readable medium, and system are disclosed for neural network pruning. The method includes the steps of receiving first-order gradients of a cost function relative to layer parameters for a trained neural network and computing a pruning criterion for each layer parameter based on the first-order gradient corresponding to the layer parameter, where the pruning criterion indicates an importance of each neuron that is included in the trained neural network and is associated with the layer parameter. The method includes the additional steps of identifying at least one neuron having a lowest importance and removing the at least one neuron from the trained neural network to produce a pruned neural network.

Abstract: A method and apparatus for enhancing motion estimation based on user input are provided. The motion estimation apparatus used for video encoding comprises a receiver operable to receive a user based input and an input analysis module operable to analyzed the user based input. The apparatus also comprises an encoder that is operable to compute displacement coordinates from the analyzed user based input for a current block in a target frame of a video stream and operable to determine a search area in a reference frame to search for a best match for the current block using the displacement coordinates. The encoder can also comprise a block match module operable to find a best match block for the current block in the search area of the reference frame using a block matching procedure.

Abstract: Motion adaptive shading increases rendering performance for real-time animation in graphics systems while maintaining dynamic image quality. Each frame of an animation is statically displayed within a refresh interval, while a viewer's eyes move continuously relative to the image when actively tracking a moving object being displayed. As a result, a statically displayed frame is essentially smeared across the viewer's continuously moving retina over the lifetime of the frame, causing a perception of blur referred to as an eye-tracking motion blur effect. A region of an image depicting a moving object may be rendered at a lower shading rate because eye-tracking motion blur will substantially mask any blur introduced by reducing the shading rate. Reducing an average shading rate for rendering frames reduces computational effort per frame and may advantageously allow a rendering system to operate at a higher frame rate to provide a smoother, clearer visual experience.

Abstract: Systems and methods are provided for efficiently performing processing intensive operations, such as those involving large volumes of data, that enable accelerated processing time of these operations. In at least one embodiment, a system includes a graphics processor unit (GPU) including a memory and a plurality of cores. The plurality of cores perform a plurality of data analytics operations on a respectively allocated portion of a dataset, each of the plurality of cores using only the memory to store data input for each of the plurality of data analytics operations performed by the plurality of cores. The data storage for the plurality of data analytics operations performed by the plurality of cores is also provided solely by the memory.

Abstract: In various examples, a virtual light meter may be implemented along with ray tracing techniques in order to determine incident light values—e.g., incoming irradiance, incident radiance, etc.—for adjusting auto exposure values of rendered frames. For example, one or more rays may be used to sample incident light over a sampling pattern—such as a hemispherical sampling pattern—for any position in a virtual game environment. As a result, the incident light values may be sampled near a subject of interest in a scene or frame such that exposure values are consistent or stable regardless of the composition of the rendered frames.

Abstract: Embodiments of the present invention set forth techniques for allocating execution resources to groups of threads within a graphics processing unit. A compute work distributor included in the graphics processing unit receives an indication from a process that a first group of threads is to be launched. The compute work distributor determines that a first subcontext associated with the process has at least one processor credit. In some embodiments, CTAs may be launched even when there are no processor credits, if one of the TPCs that was already acquired has sufficient space. The compute work distributor identifies a first processor included in a plurality of processors that has a processing load that is less than or equal to the processor loads associated with all other processors included in the plurality of processors. The compute work distributor launches the first group of threads to execute on the first processor.

Abstract: In various examples, a deep neural network (DNN) is trained to accurately predict, in deployment, distances to objects and obstacles using image data alone. The DNN may be trained with ground truth data that is generated and encoded using sensor data from any number of depth predicting sensors, such as, without limitation, RADAR sensors, LIDAR sensors, and/or SONAR sensors. Camera adaptation algorithms may be used in various embodiments to adapt the DNN for use with image data generated by cameras with varying parameters—such as varying fields of view. In some examples, a post-processing safety bounds operation may be executed on the predictions of the DNN to ensure that the predictions fall within a safety-permissible range.

Abstract: Determining the occlusions or shadows for an area light within a scene is difficult, especially realistic shadowing in large and dynamic scenes. The disclosure provides an adaptive occlusion sampling process that uses voxel cone tracing to distribute the voxel tracing cones on the surface of area lights to obtain samples for shadowing in computer generated images or scenes. A method of adaptive occlusion sampling from a rectangular area light is disclosed that can be used to provide realistic shadowing in a computer generated scene. A process to compute a shadow of an area light within a scene is also disclosed herein that includes obtaining samples, employing voxel cone tracing, from a light surface of the area light based on sample points of a sampling grid created from sample patterns that are based on a determined number of cones.

Abstract: Various types of systems or technologies can be used to collect data in a 3D space. For example, LiDAR (light detection and ranging) and RADAR (radio detection and ranging) systems are commonly used to generate point cloud data for 3D space around vehicles, for such functions as localization, mapping, and tracking. This disclosure provides improved techniques for processing the point cloud data that has been collected. The improved techniques include mapping 3D point cloud data points into a 2D depth map, fetching a group of the mapped 3D point cloud data points that are within a bounded window of the 2D depth map; and generating geometric space parameters based on the group of the mapped 3D point cloud data points. The generated geometric space parameters may be used for object motion, obstacle detection, freespace detection, and/or landmark detection for an area surrounding a vehicle.

Abstract: Ray tracing hardware accelerators supporting multiple specifiers for controlling the traversal of a ray tracing acceleration data structure are disclosed. For example, traversal efficiency and complex ray tracing effects can be achieved by specifying traversals through such data structures using both programmable ray operations and explicit node masking. The explicit node masking utilizes dedicated fields in the ray and in nodes of the acceleration data structure to control traversals. Ray operations, however, are programmable per ray using opcodes and additional parameters to control traversals. Traversal efficiency is improved by enabling more aggressive culling of parts of the data structure based on the combination of explicit node masking and programmable ray operations. More complex ray tracing effects are enabled by providing for dynamic selection of nodes based on individual ray characteristics.

Abstract: Inverse rendering estimates physical scene attributes (e.g., reflectance, geometry, and lighting) from image(s) and is used for gaming, virtual reality, augmented reality, and robotics. An inverse rendering network (IRN) receives a single input image of a 3D scene and generates the physical scene attributes for the image. The IRN is trained by using the estimated physical scene attributes generated by the IRN to reproduce the input image and updating parameters of the IRN to reduce differences between the reproduced input image and the input image. A direct renderer and a residual appearance renderer (RAR) reproduce the input image. The RAR predicts a residual image representing complex appearance effects of the real (not synthetic) image based on features extracted from the image and the reflectance and geometry properties. The residual image represents near-field illumination, cast shadows, inter-reflections, and realistic shading that are not provided by the direct renderer.

Abstract: The present invention facilitates efficient and effective image processing. A network can comprise: a first system configured to perform a first portion of lighting calculations for an image and combing results of the first portion of lighting calculations for the image with results of a second portion of lighting calculations; and a second system configured to perform the second portion of lighting calculations and forward the results of the second portion of the lighting calculations to the first system. The first and second portion of lighting calculations can be associated with indirect lighting calculations and direct lighting calculations respectively. The first system can be a client in a local location and the second system can be a server in a remote location (e.g., a cloud computing environment). The first system and second system can be in a cloud and a video is transmitted to a local system.

Abstract: A bounding volume is used to approximate the space an object occupies. If a more precise understanding beyond an approximation is required, the object itself is then inspected to determine what space it occupies. Often, a simple volume (such as an axis-aligned box) is used as bounding volume to approximate the space occupied by an object. But objects can be arbitrary, complicated shapes. So a simple volume often does not fit the object very well. That causes a lot of space that is not occupied by the object to be included in the approximation of the space being occupied by the object. Hardware-based techniques are disclosed herein, for example, for efficiently using multiple bounding volumes (such as axis-aligned bounding boxes) to represent, in effect, an arbitrarily shaped bounding volume to better fit the object, and for using such arbitrary bounding volumes to improve performance in applications such as ray tracing.

Abstract: Apparatuses, systems, and techniques to parallelize operations in one or more programs with data copies from global memory to shared memory in each of the one or more programs. In at least one embodiment, a program performs operations on shared data and then asynchronously copies shared data to shared memory, and continues performing additional operations in parallel while the shared data is copied to shared memory until an indicator provided by an application programming interface to facilitate parallel computing, such as CUDA, informs said program that shared data has been copied to shared memory.

Abstract: An adder circuit that includes an operand input and a second operand input to an XNOR cell. The XNOR cell is configured to provide the operand input and the second operand input to both a NAND gate and a first OAI cell. A second OAI cell transforms the output of the XNOR cell into a carry out signal.

Abstract: In various embodiments, rail decoupling circuits that are powered by an always on voltage rail allow a core voltage rail to power up independently of an I/O voltage rail without jeopardizing I/O pad circuits that are powered by the I/O voltage rail. In an embodiment, when the always on voltage rail is powered-up and a chip reset signal is asserted, the rail decoupling circuits drive control inputs of the I/O pad circuits based on default values. When the chip reset signal is de-asserted, the rail decoupling circuits drive the control inputs of the I/O pad circuits based on signals received from circuits powered by the core voltage rail. Because the rail decoupling circuits maintain control of the I/O pad circuits until the chip-reset is de-asserted, the core voltage rail can power up at any time before the chip-reset signal is de-asserted irrespective of when the I/O voltage rail powers up.

Abstract: A communication method between a source device and a target device utilizes speculative connection setup between the source device and the target device, target-device-side packet ordering, and fine-grained ordering to remove packet dependencies.

Abstract: A high-definition map system receives sensor data from vehicles travelling along routes and combines the data to generate a high definition map for use in driving vehicles, for example, for guiding autonomous vehicles. A pose graph is built from the collected data, each pose representing location and orientation of a vehicle. The pose graph is optimized to minimize constraints between poses. Points associated with surface are assigned a confidence measure determined using a measure of hardness/softness of the surface. A machine-learning-based result filter detects bad alignment results and prevents them from being entered in the subsequent global pose optimization. The alignment framework is parallelizable for execution using a parallel/distributed architecture. Alignment hot spots are detected for further verification and improvement. The system supports incremental updates, thereby allowing refinements of subgraphs for incrementally improving the high-definition map for keeping it up to date.

Abstract: Enhanced techniques applicable to a ray tracing hardware accelerator for traversing a hierarchical acceleration structure are disclosed. The traversal efficiency of such hardware accelerators are improved, for example, by transforming a ray, in hardware, from the ray's coordinate space to two or more coordinate spaces at respective points in traversing the hierarchical acceleration structure. In one example, the hardware accelerator is configured to transform a ray, received from a processor, from the world space to at least one alternate world space and then to an object space in hardware before a corresponding ray-primitive intersection results are returned to the processor. The techniques disclosed herein facilitate the use of additional coordinate spaces to orient acceleration structures in a manner that more efficiently approximate the space occupied by the underlying primitives being ray-traced.

Abstract: A method for remotely provisioning resources for running a computer application is described. The method includes: receiving a request to execute a computer application using a virtual machine, the computer application having a static video portion and a user interactive video portion; while preparing the user interactive video portion, providing the static video portion to a user device remotely positioned relative to a server hosting the virtual machine; and streaming the user interactive video portion to the user device at an end of the static video portion. A start time of the static video portion is adjusted to accommodate a display time of the static video portion being different than a time required to prepare the user interactive video portion. A server and a system that are capable of performing the above method are also described.