Intel Corporation

Blaize, Inc.

Thinci, Inc.

Real 3-d

Blaze, Inc.

Val Gene Cook

Satyaki Koneru

Ke Yin

Thomas A Piazza

Dinakar C Munagala

Hong Jiang

Michael Mantor

Peter L Doyle

Kam Leung

Wing Hang Wong

Kim A Meinerth

Ralph Clayton Taylor

Indraneel Ghosh

David W Watson

Michael K Dwyer

Bishara Shomar

Dinakar Munagala

Vineet Goel

Scott R Janus

Steven Spangler

Allen H Simon

John T Pedicone

Nicolas I Kacevas

Brian D Rauchfuss

Sam W Jensen

Tom A Piazza

Adam P Geringer

Jun Tian

Stuart Krupnik

Steven Manno

Methods, systems, and apparatuses for unsupervised data drift detection for classification neural networks are disclosed. One method includes providing the data stream of images to a neural network, generating, by the neural network, class wise probabilities, storing each image of the data stream of images, storing the class wise probabilities generated by the neural network, comparing artifacts of images of the data stream at a first time with artifacts of images of the data stream at a second time, comparing artifacts produced by the class wise probabilities of the data stream retrieved from the stored class wise probabilities at a third time with artifacts produced by the class wise probabilities of the data stream retrieved from the stored class wise probabilities at a fourth time, and generating an informative communication based on the comparisons.

Unsupervised data drift detection for classification neural networks

Systems and methods are disclosures for scheduling code in a multiprocessor system. Code is portioned into code blocks by a compiler. The compiler schedules execution of code blocks in nodes. The nodes are connected in a directed acyclical graph with a top node, terminal node and a plurality of intermediate nodes. Execution of the top node is initiated by the compiler. After executing at least one instance of the top node, an instruction in the code block indicates to the scheduler to initiate at least one intermediary node. The scheduler schedules a thread for execution of the intermediary node. The data for the nodes resides in a plurality of data buffers; the index to the data buffer is stored in a command buffer.

Configurable scheduler for graph processing on multi-processor computing systems

The described embodiments include systems, methods, and apparatuses for increased efficiency processing flow. One method includes a plurality of stages configured to process an execution graph that includes a plurality of logical nodes with defined properties and resources associated with each logical node of the plurality of logical nodes, a recirculating ring buffer, wherein the recirculating ring buffer is configured to holding only any one of a control information, input, and, or out data necessary to stream a temporary data between each logical node of the execution graph, and a data producer, wherein the data producer is configured to stall from writing control information into a command buffer upon the command buffer being full, preventing command buffer over-writing.

Node topology employing command and data buffers for executing thread scheduling

Systems, apparatuses and methods are disclosed for scheduling threads comprising of code blocks in a graph streaming processor (GSP) system. One system includes a scheduler for scheduling plurality of threads, the plurality of threads includes a set of instructions operating on the graph streaming processors of GSP system. The scheduler comprises a plurality of stages where each stage is coupled to an input command buffer and an output command buffer. A portion of the scheduler is implemented in hardware and comprises of a command parser operative to interpret commands within a corresponding input command buffer, a thread generator coupled to the command parser operate to generate the plurality of threads, and a thread scheduler coupled to the thread generator for dispatching the plurality of threads for operating on the plurality of graph streaming processors.

Configurable scheduler in a graph streaming processing system

Node topology employing recirculating ring command and data buffers for executing thread scheduling

The claimed invention discloses system comprising a plurality of logical nodes comprised in a single or plurality of stages, with defined properties and resources associated with each node, for reducing compute resources, said system further comprising: at least a recirculating ring buffer holding only any one of a control information, input, and, or out data necessary to stream a temporary data between node and, or nodes in an execution graph, thereby reducing size of said recirculating ring buffer; said recirculating ring buffer being sufficiently reduced in size to reside in an on-chip cache, such that any one of the control information, input, and, or out data between node and, or nodes need not be stored in memory; wherein the control information further comprises a command related to invalidating any one of the input and, or out data held in a recirculating ring data buffer, clearing the buffer of tasked data; and wherein a producer is stalled from writing any more control information into a recirculating ring command buffer upon the buffer being full, preventing command buffer over-writing, and thereby reducing compute resources associated with a DRAM memory transaction.

Embodiments disclosed include a mechanism in a system and method for significantly reducing power consumption by reducing computation and bandwidth. This mechanism is particularly applicable for modern 3D synthetic images which contain high pixel overdraw and dynamically generated intermediates images. Only blocks of computation which contribute to the final image are performed. This is accomplished by rendering in reverse order and by performing multiple visibility sort in a streaming fashion through the pipeline. Rendering of dynamically generated intermediate images is performed sparsely by projecting texture coordinates from a current image back into one or more dependent images in a recursive manner. The newly computed pixel values are then filtered and control is returned to the sampling shader of the current image. When only visible pixels are projected optimal computation is performed. Several implementations are presented with increasing efficiency. An acceleration structure, termed a Draw Buffer, simplifies the process of projecting backward and utilizes a hardware managed dynamic memory object. This mechanism reduces computation by 50%, with significant bandwidth and power savings.

Mechanism for minimal computation and power consumption for rendering synthetic 3D images, containing pixel overdraw and dynamically generated intermediate images

A machine readable storage media containing executable program instructions which when executed cause a digital processing system to set a plurality of operands and operators. A plurality of texture maps are sent to a processor for texture compositing. Operands are selected for a texture operation. A first logic is used wherein if the circulation of blend stages is equal to a number of blend stages, then a color saturation is performed, and a second logic that if the circulation number of blend stages does not equal the number of blend stages then at least one operand is selected for another texture compositing operation.

Multiple texture compositing

A context-free (stateless) dataport may allow multiple processors to perform read and write operations on a shared memory. The operations may include, for example, structured data operations such as image and video operations. The dataport may perform addressing computations associated with block memory operations. Therefore, the dataport may be able, for example, to relieve the processors that it serves from this duty. The dataport may be accessed using a message interface that may be implemented in a standard and generalized manner and that may therefore be easily transportable between different types of processors.

Dataport and methods thereof

A configurable filter module for providing shared filter resource between an overlay engine and a texture mapping engine of a graphics system. The configurable filter may comprise a plurality of linear blend units each of which receives data input from one of the overlay engine and a mapping engine cache, and generates a linear blend filter output respectively; and a filter output multiplexer which receives data output from the linear blend units and selects a proper byte ordering output, wherein the linear blend units serve as an overlay interpolator filter to perform linear blending of the data input from the overlay engine during a linear blend mode, and serve as a texture bilinear filter to perform bilinear filtering of the data input from the mapping engine cache during a bilinear filtering mode.

Pixel filtering using shared filter resource between overlay and texture mapping engines

Growing community of inventors

Shingle Springs, CA, United States of America

Val Gene Cook