Amazon Technologies, Inc.

Marvell Asia Pte., Ltd.

Cavium, Inc.

Jeffrey T Huynh

Ron Diamant

Randy Renfu Huang

Drazen Borkovic

Sundeep Amirineni

Jindrich Zejda

Weihuang Wang

Richard John Heaton

Tsahi Daniel

Hongbin Zheng

Yida Wang

Yizhi Liu

Animesh Jain

Haichen Li

Se Jong Oh

Anh T Tran

Vignesh Vivekraja

Srinath Atluri

Yu Zhou

Vinod Sharma

Gerald Schmidt

Mohan Balan

Taemin Kim

Tobias Joseph Kastulus Edler Von Koch

Zhi Chen

Generating instructions for programming a processing element array to implement a convolution operation can include determining that the convolution operation under-utilizes the processing element array. The convolution operation involves using the processing element array to perform a series of matrix multiplications between a set of filters and a set of input matrices. Each filter comprises a weight matrix. Each input matrix is assigned to a respective row in the processing element array. Under-utilization can be determined through detecting that less than a threshold number of rows would be used concurrently. In response to determining that the convolution operation under-utilizes the processing element array, instructions can be added for modifying the convolution operation to increase the number of rows used concurrently. The added instructions are executable to cause at least one input matrix to be processed in parallel across more rows compared to processing without modifying the convolution operation.

Efficient utilization of processing element array

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

A neural network accelerator executes instructions to: load a first weight data element of an array of weight data elements from a memory into a systolic array; extract, from the instructions, information indicating a first number of input data elements to be obtained from a first address of the memory and a second number of input data elements to be skipped between adjacent input data elements to be obtained, the first address being based on first coordinates of the first weight data element, and the first and second numbers being based on a stride of a convolution operation; based on the information, obtain first input data elements from the first address of the memory; and control the systolic array to perform first computations based on the first weight data element and the first input data elements to generate first output data elements of an output data array.

Memory operation for systolic array

In one example, a neural network accelerator can execute a set of instructions to: load a first weight data element from a memory into a systolic array, the first weight data element having first coordinates; extract, from the instructions, information indicating a first subset of input data elements to be obtained from the memory, the first subset being based on a stride of a transposed convolution operation and second coordinates of first weight data element in a rotated array of weight data elements; based on the information, obtain the first subset of input data elements from the memory; load the first subset of input data elements into the systolic array; and control the systolic array to perform first computations based on the first weight data element and the first subset of input data elements to generate output data elements of an array of output data elements.

Transposed convolution using systolic array

Embodiments of the present invention are directed to a wildcard matching solution that uses a combination of static random access memories (SRAMs) and ternary content addressable memories (TCAMs) in a hybrid solution. In particular, the wildcard matching solution uses a plurality of SRAM pools for lookup and a spillover TCAM pool for unresolved hash conflicts.

Hybrid wildcard match table

Provided are systems and methods for operating a neural network processor, wherein the processor includes an input selector circuit that can be configured to select the data that will be input into the processor's computational array. In various implementations, the selector circuit can determine, for a row of the array, whether the row input will be the output from a buffer memory or data that the input selector circuit has selected for a different row. The row can receive an input feature map from a set of input data or an input feature map that was selected for inputting into a different row, such that the input feature map is input into more than one row at a time. The selector circuit can also include a delay circuit, so that the duplicated input feature map can be input into the computational array later than the original input feature map.

Data selection circuit

In one example, a non-transitory computer readable medium stores instructions that, when executed by one or more hardware processors, cause the one or more hardware processors to: load a first weight data element of an array of weight data elements from a memory into a systolic array; select a subset of input data elements from the memory into the systolic array to perform first computations of a dilated convolution operation, the subset being selected based on a rate of the dilated convolution operation and coordinates of the weight data element within the array of weight data elements; and control the systolic array to perform the first computations based on the first weight data element and the subset to generate first output data elements of an output data array. An example of a compiler that generates the instructions is also provided.

Dilated convolution using systolic array

A computer-implemented method includes receiving a neural network model for implementation using a processing element array, where the neural network model includes a convolution operation on a set of input feature maps and a set of filters. The method also includes determining, based on the neural network model, that the convolution operation utilizes less than a threshold number of rows in the processing element array for applying a set of filter elements to the set of input feature maps, where the set of filter elements includes one filter element in each filter of the set of filters. The method further includes generating, for the convolution operation and based on the neural network model, a first instruction and a second instruction for execution by respective rows in the processing element array, where the first instruction and the second instruction use different filter elements of a filter in the set of filters.

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

In one example, a neural network accelerator executes instructions to: load a first weight data element of an array of weight data elements from a memory into a systolic array; extract, from the instructions, information indicating a first number of input data elements to be obtained from a first address of the memory and a second number of input data elements to be skipped between adjacent input data elements to be obtained, the first address being based on first coordinates of the first weight data element, and the first and second numbers being based on a stride of a convolution operation; based on the information, obtain first input data elements from the first address of the memory; and control the systolic array to perform first computations based on the first weight data element and the first input data elements to generate first output data elements of an output data array.

Provided are integrated circuits and methods for transposing a tensor using processing element array operations. In some cases, it may be necessary to transpose elements of a tensor to perform a matrix operation. The tensor may be decomposed into blocks of data elements having dimensions consistent with the dimensions of a systolic array. An identity multiplication may be performed on each block of data elements loaded into a systolic array and the multiplication products summed in column partitions of a results buffer. The data elements in the column partitions of results buffer can then be mapped to row partitions of a buffer memory for further processing.

Transpose operations using processing element array

Techniques are disclosed for debugging a neural network execution on a target processor. A reference processor may generate a plurality of first reference tensors for the neural network. The neural network may be repeatedly reduced to produce a plurality of lengths. For each of the lengths, a compiler converts the neural network into first machine instructions, the target processor executes the first machine instructions to generate a first device tensor, and the debugger program determines whether the first device tensor matches a first reference tensor. A shortest length is identified for which the first device tensor does not match the first reference tensor. Tensor output is enabled for a lower-level intermediate representation of the shortest neural network, and the neural network is converted into second machine instructions, which are executed by the target processor to generate a second device tensor.

Growing community of inventors

San Jose, CA, United States of America

Jeffrey T Huynh