Google Inc.

Nvidia Corporation

Other

Intel Corporation

William R Mark

Albert Meixner

Jason Rupert Redgrave

Ofer Shacham

Daniel Finchelstein

David Patterson

Neeti Desai

Qiuling Zhu

Edward Y Chang

Hyunchul Park

Donald C Stark

Dilan Manatunga

Timothy James Knight

Shahriar Rabii

Don Stark

Masumi Reynders

Fabrizio Basso

Michelle Tomasko

Mark J Kilgard

Cass Watson Everitt

Robert Steven Glanville

Rev Lebaredian

Kurt Barton Akeley

Matthew Milton Pharr

Gregory F Johnson

Craig E Kolb

Edward T Chang

Gordon W Stoll

Chris Burns

Daniel Frederic Finchelstein

A method is described that includes executing a convolutional neural network layer on an image processor having an array of execution lanes and a two-dimensional shift register. The two-dimensional shift register provides local respective register space for the execution lanes. The executing of the convolutional neural network includes loading a plane of image data of a three-dimensional block of image data into the two-dimensional shift register. The executing of the convolutional neural network also includes performing a two-dimensional convolution of the plane of image data with an array of coefficient values by sequentially: concurrently multiplying within the execution lanes respective pixel and coefficient values to produce an array of partial products; concurrently summing within the execution lanes the partial products with respective accumulations of partial products being kept within the two dimensional register for different stencils within the image data; and, effecting alignment of values for the two-dimensional convolution within the execution lanes by shifting content within the two-dimensional shift register array.

Convolutional neural network on programmable two dimensional image processor

An image processor unit is described. The image processor unit includes a plurality of inputs to receive at least one input image. The image processor unit includes a plurality of outputs to provide at least one output image. The image processor unit includes a network coupled to the plurality of inputs and the plurality of outputs. The network is to couple at least one of the inputs to at least one of the outputs. The image processor unit includes an image processor circuit coupled to the network. The network to route an input image that is received at one of the inputs to the image processor circuit. The image processor circuit is to execute image signal processing program code to generate a processed output image from the input image. The network is to route the processed output image to at least one of the outputs.

Configurable and programmable image processor unit

A method is described that includes, on an image processor having a two dimensional execution lane array and a two dimensional shift register array, repeatedly shifting first content of multiple rows or columns of the two dimensional shift register array and repeatedly executing at least one instruction between shifts that operates on the shifted first content and/or second content that is resident in respective locations of the two dimensional shift register array that the shifted first content has been shifted into.

Block operations for an image processor having a two-dimensional execution lane array and a two-dimensional shift register

A method is described that includes executing a convolutional neural network layer on an image processor having an array of execution lanes and a two-dimensional shift register. The executing of the convolutional neural network includes loading a plane of image data of a three-dimensional block of image data into the two-dimensional shift register. The executing of the convolutional neural network also includes performing a two-dimensional convolution of the plane of image data with an array of coefficient values by sequentially: concurrently multiplying within the execution lanes respective pixel and coefficient values to produce an array of partial products; concurrently summing within the execution lanes the partial products with respective accumulations of partial products being kept within the two dimensional register for different stencils within the image data; and, effecting alignment of values for the two-dimensional convolution within the execution lanes by shifting content within the two-dimensional shift register array.

An image processor is described. The image processor includes an I/O unit to read input image data from external memory for processing by the image processor and to write output image data from the image processor into the external memory. The I/O unit includes multiple logical channel units. Each logical channel unit is to form a logical channel between the external memory and a respective producing or consuming component within the image processor. Each logical channel unit is designed to utilize reformatting circuitry and addressing circuitry. The addressing circuitry is to control addressing schemes applied to the external memory and reformatting of image data between external memory and the respective producing or consuming component. The reformatting circuitry is to perform the reformatting.

Macro I/O unit for image processor

An apparatus is described. The apparatus includes an image processing unit. The image processing unit includes a plurality of stencil processor circuits each comprising an array of execution unit lanes coupled to a two-dimensional shift register array structure to simultaneously process multiple overlapping stencils through execution of program code. The image processing unit includes a plurality of sheet generators respectively coupled between the plurality of stencil processors and the network. The sheet generators are to parse input line groups of image data into input sheets of image data for processing by the stencil processors, and, to form output line groups of image data from output sheets of image data received from the stencil processors. The image processing unit includes a plurality of line buffer units coupled to the network to pass line groups in a direction from producing stencil processors to consuming stencil processors to implement an overall program flow.

Architecture for high performance, power efficient, programmable image processing

A method is described. The method includes simulating execution of an image processing application software program. The simulating includes intercepting kernel-to-kernel communications with simulated line buffer memories that store and forward lines of image data communicated from models of producing kernels to models of consuming kernels. The simulating further includes tracking respective amounts of image data stored in the respective line buffer memories over a simulation runtime. The method also includes determining respective hardware memory allocations for corresponding hardware line buffer memories from the tracked respective amounts of image data. The method also includes generating configuration information for an image processor to execute the image processing application software program. The configuration information describes the hardware memory allocations for the hardware line buffer memories of the image processor.

Determination of per line buffer unit memory allocation

A method is described. The method includes compiling program code targeted for an image processor having programmable stencil processors composed of respective two-dimensional execution lane and shift register circuit structures. The program code is to implement a directed acyclic graph and is composed of multiple kernels that are to execute on respective ones of the stencil processors, wherein the compiling includes any of: recognizing there are a different number of kernels in the program code than stencil processors in the image processor; recognizing that at least one of the kernels is more computationally intensive than another one of the kernels; and, recognizing that the program code has resource requirements that exceed the image processor's memory capacity. The compiling further includes in response to any of the recognizing above performing any of: horizontal fusion of kernels; vertical fusion of kernels; fission of one of the kernels into multiple kernels; spatial partitioning of a kernel into multiple spatially partitioned kernels; splitting the directed acyclic graph into smaller graphs.

Compiler techniques for mapping program code to a high performance, power efficient, programmable image processing hardware platform

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for restructuring an image processing pipeline. The method includes compiling program code targeted for an image processor having programmable stencil processors composed of respective two-dimensional execution lane and shift register circuit structures. The program code is to implement a directed acyclic graph and is composed of multiple kernels that are to execute on respective ones of the stencil processors, wherein the compiling includes performing any of: horizontal fusion of kernels; vertical fusion of kernels; fission of one of the kernels into multiple kernels; spatial partitioning of a kernel into multiple spatially partitioned kernels; or splitting the directed acyclic graph into smaller graphs.

Growing community of inventors

Austin, TX, United States of America

William R Mark