Amazon Technologies, Inc.

Xilinx, Inc.

Jindrich Zejda

Ashish Sirasao

Drazen Borkovic

Elliott Delaye

Ron Diamant

Aaron Ng

Randy Renfu Huang

Yongjun Wu

Jeffrey T Huynh

Richard John Heaton

Xiao Teng

Taemin Kim

Ehsan Ghasemi

Sean Settle

Yida Wang

Yizhi Liu

Animesh Jain

Tobias Joseph Kastulus Edler Von Koch

Ilya Minkin

Dana Michelle Vantrease

Khang Kim Dao

Vinod Sharma

Thomas A Volpe

Sonal Santan

Ilya K Ganusov

Robert Geva

Hongbin Zheng

Soren T Soe

Benjamin S Devlin

Satish B Sivaswamy

Walter A Manaker, Jr

Ling Liu

Haichen Li

Christopher J Case

Se Jong Oh

Atul Srinivasan

Zhi Chen

Zhao Ma

Yifei Zhou

Tiandong Zhao

A computer-implemented method includes receiving a neural network model that includes memory load operations and a plurality of computation operations, selecting a memory load operation having an arithmetic intensity factor (AIF) greater than a threshold value from the memory load operations, grouping computation operations associated with data loaded by the selected memory load operation into two or more clusters of computation operations, and incorporating an instance of the selected memory load operation before each cluster of the two or more clusters of computation operations in the neural network model.

Arithmetic-intensity based load cloning

A method of generating executable instructions for a computing system is provided. The method comprises: receiving a first set of instructions including a kernel of a first operator and a kernel of a second operator, the kernel of the first operator including instructions of the first operator and write instructions to a virtual data node, the kernel of the second operator including instructions of the second operator and read instructions to the virtual data node; determining, based on a mapping between the write instructions and read instructions, instructions of data transfer operations between the first operator and the second operator; and generating a second set of instructions representing a fused operator of the first operator and the second operator, the second set of instructions including the instructions of the first operator, the instructions of the second operator, and the instructions of the data transfer operations.

Performing hardware operator fusion

Methods and apparatuses for hierarchical partitioning of operators of a neural network for execution on an acceleration engine are provided. Neural networks are built in machine learning frameworks using neural network operators. The neural network operators are compiled into executable code for the acceleration engine. Development of new framework-level operators can exceed the capability to map the newly developed framework-level operators onto the acceleration engine. To enable neural networks to be executed on an acceleration engine, hierarchical partitioning can be used to partition the operators of the neural network. The hierarchical partitioning can identify operators that are supported by a compiler for execution on the acceleration engine, operators to be compiled for execution on a host processor, and operators to be executed on the machine learning framework.

Hierarchical partitioning of operators

Static memory allocation may be performed for weight values across multiple processing units executing a neural network. A neural network may be received for execution across multiple processing units. A partitioning scheme may be applied to divide the neural network into subgraphs. The subgraphs may be assigned to different processing units. The weights for the operations of the subgraph may be statically allocated in dedicated caches for the processing units as part of the instructions to execute the neural network across the processing units.

Static memory allocation for neural network inference

Embodiments herein describe techniques for expressing the layers of a neural network in a software model. In one embodiment, the software model includes a class that describes the various functional blocks (e.g., convolution units, max-pooling units, rectified linear units (ReLU), and scaling functions) used to execute the neural network layers. In turn, other classes in the software model can describe the operation of each of the functional blocks. In addition, the software model can include conditional logic for expressing how the data flows between the functional blocks since different layers in the neural network can process the data differently. A compiler can convert the high-level code in the software model (e.g., C++) into a hardware description language (e.g., register transfer level (RTL)) which is used to configure a hardware system to implement a neural network accelerator.

Software defined neural network layer pipelining

Embodiments herein describe techniques for static scheduling a neural network implemented in a massively parallel hardware system. The neural network may be scheduled using three different scheduling levels referred to herein as an upper level, an intermediate level, and a lower level. In one embodiment, the upper level includes a hardware or software model of the layers in the neural network that establishes a sequential order of functions that operate concurrently in the hardware system. In the intermediate level, identical processes in the functions defined in the upper level are connected to form a systolic array or mesh and balanced data flow channels are used to minimize latency. In the lower level, a compiler can assign the operations performed by the processing elements in the systolic array to different portions of the hardware system to provide a static schedule for the neural network.

Static block scheduling in massively parallel software defined hardware systems

Embodiments herein describe techniques for interfacing a neural network application with a neural network accelerator using a library. The neural network application may execute on a host computing system while the neural network accelerator executes on a massively parallel hardware system, e.g., a FPGA. The library operates a pipeline for submitting the tasks received from the neural network application to the neural network accelerator. In one embodiment, the pipeline includes a pre-processing stage, an FPGA execution stage, and a post-processing stage which each correspond to different threads. When receiving a task from the neural network application, the library generates a packet that includes the information required for the different stages in the pipeline to perform the tasks. Because the stages correspond to different threads, the library can process multiple packets in parallel which can increase the utilization of the neural network accelerator on the hardware system.

Machine learning runtime library for neural network acceleration

In the disclosed methods and systems for processing in a neural network system, a host computer system writes a plurality of weight matrices associated with a plurality of layers of a neural network to a memory shared with a neural network accelerator. The host computer system further assembles a plurality of per-layer instructions into an instruction package. Each per-layer instruction specifies processing of a respective layer of the plurality of layers of the neural network, and respective offsets of weight matrices in a shared memory. The host computer system writes input data and the instruction package to the shared memory. The neural network accelerator reads the instruction package from the shared memory and processes the plurality of per-layer instructions of the instruction package.

Multi-layer neural network processing by a neural network accelerator using host communicated merged weights and a package of per-layer instructions

Techniques for time-based memory allocation for a neural network inference are disclosed. A description of a neural network comprising a plurality of operations to be executed across a set of accelerators is received. A plurality of interconnect times at a plurality of partition points within the neural network are calculated. Each of the plurality of interconnect times corresponds to a duration of time for transferring an output feature map from one of the set of accelerators to another of the set of accelerators to be used as an input feature map. A partitioning scheme that divides the plurality of operations into a set of subgraphs is determined based on the plurality of interconnect times. Each of the set of subgraphs is assigned to a different accelerator of the set of accelerators in accordance with the partitioning scheme.

Time-based memory allocation for neural network inference

Techniques are disclosed for reordering operations of a neural network to improve runtime efficiency. In some examples, a compiler receives a description of the neural network comprising a plurality of operations. The compiler may determine which execution engine of a plurality of execution engines is to perform each of the plurality of operations. The compiler may determine an order of performance associated with the plurality of operations. The compiler may identify a runtime inefficiency based on the order of performance and a hardware usage for each of the plurality of operations. An operation may be reordered to reduce the runtime inefficiency. Instructions may be compiled based on the plurality of operations, which include the reordered operation.

Neural network operation reordering for parallel execution

A disclosed neural network processing system includes a host computer system, a RAMs coupled to the host computer system, and neural network accelerators coupled to the RAMs, respectively. The host computer system is configured with software that when executed causes the host computer system to write input data and work requests to the RAMS. Each work request specifies a subset of neural network operations to perform and memory locations in a RAM of the input data and parameters. A graph of dependencies among neural network operations is built and additional dependencies added. The operations are partitioned into coarse grain tasks and fine grain subtasks for optimal scheduling for parallel execution. The subtasks are scheduled to accelerator kernels of matching capabilities. Each neural network accelerator is configured to read a work request from the respective RAM and perform the subset of neural network operations on the input data using the parameters.

Neural network processing system having host controlled kernel acclerators

Techniques for operating a computing system to perform neural network operations are disclosed. In one example, a method comprises receiving a neural network model, determining a sequence of neural network operations based on data dependency in the neural network model, and determining a set of instructions to map the sequence of neural network operations to the processing resources of the neural network processor. The method further comprises determining, based on a set of memory access operations included in the set of instructions, a first set of memory references associated with a first location of an external memory to store the input data and a second set of memory references associated with a second location of the external memory to store the output data, and generating an instruction file including the set of instructions, the first set of memory references and the second set of memory references.

Allocation and placement of resources for network computation

In disclosed approaches of neural network processing, a host computer system copies an input data matrix from host memory to a shared memory for performing neural network operations of a first layer of a neural network by a neural network accelerator. The host instructs the neural network accelerator to perform neural network operations of each layer of the neural network beginning with the input data matrix. The neural network accelerator performs neural network operations of each layer in response to the instruction from the host. The host waits until the neural network accelerator signals completion of performing neural network operations of layer i before instructing the neural network accelerator to commence performing neural network operations of layer i+1, for i&#x2265;1. The host instructs the neural network accelerator to use a results data matrix in the shared memory from layer i as an input data matrix for layer i+1 for i&#x2265;1.

Host-directed multi-layer neural network processing via per-layer work requests

An example preprocessor circuit includes: a first buffer configured to store rows of image data and output a row thereof; a second buffer, coupled to the first buffer, including storage locations to store respective image samples of the row output by the first buffer; shift registers; an interconnect network including connections, each connection coupling a respective one of the shift registers to more than one of the storage locations, one or more of the storage locations being coupled to more than one of the connections; and a control circuit configured to load the shift registers with the image samples based on the connections and shift the shift registers to output streams of image samples.

Image preprocessing for generalized image processing

Growing community of inventors

Saratoga, CA, United States of America

Jindrich Zejda