Back to results
US 11,960,566 B1Granted
Reducing computations for data including padding
AssigneeAmazon Technologies, Inc.
InventorsDana Michelle Vantrease, Ron Diamant
Life of this patent
Estimating the term…
Abstract
Systems and methods are provided to eliminate multiplication operations with zero padding data for convolution computations. A multiplication matrix is generated from an input feature map matrix with padding by adjusting coordinates and dimensions of the input feature map matrix to exclude padding data. The multiplication matrix is used to perform matrix multiplications with respective weight values which results in fewer computations as compared to matrix multiplications which include the zero padding data.
Claims20
A method comprising:
- receiving, by a compiler, a padded input data element matrix that includes an input data element matrix at least partially surrounded by padding elements, wherein the input data element matrix stores input data elements;
- receiving, by the compiler, a weight matrix that stores weight values to be used with the input data elements;
- for each weight value from the weight values of the weight matrix:
- determining, by the compiler, coordinates and dimensions for a multiplication matrix within the padded input data element matrix to be used with the weight value;
- and
- reassigning, by the compiler, the coordinates and the dimensions for the multiplication matrix within the padded input data element matrix to exclude the padding elements;
- generating, by the compiler, compiled code to be executed by a computation engine having a systolic array having columns and rows of processing elements, the compiled code incorporating the reassigned coordinates and the reassigned dimensions;
- and
- performing, at the computation engine and based on executing the compiled code, matrix multiplications using the reassigned coordinates and the reassigned dimensions.
6 more elements
A system comprising:
- one or more processors;
- and
- a non-transitory computer-readable medium having instructions stored therein that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
- receiving, by a compiler, a padded input data element matrix that includes an input data element matrix at least partially surrounded by padding elements, wherein the input data element matrix stores input data elements;
- receiving, by the compiler, a weight matrix that stores weight values to be used with the input data elements;
- for each weight value from the weight values of the weight matrix:
- determining, by the compiler, coordinates and dimensions for a multiplication matrix within the padded input data element matrix to be used with the weight value;
- and
- reassigning, by the compiler, the coordinates and the dimensions for the multiplication matrix within the padded input data element matrix to exclude the padding elements;
- and
- generating, by the compiler, compiled code to be executed by a computation engine having a systolic array having columns and rows of processing elements, the compiled code incorporating the reassigned coordinates and the reassigned dimensions;
- wherein matrix multiplications using the reassigned coordinates and the reassigned dimensions are to be performed at the computation engine and based on executing the compiled code.
9 more elements
A non-transitory computer-readable medium having instructions stored therein that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
- receiving, by a compiler, a padded input data element matrix that includes an input data element matrix at least partially surrounded by padding elements, wherein the input data element matrix stores input data elements;
- receiving, by the compiler, a weight matrix that stores weight values to be used with the input data elements;
- for each weight value from the weight values of the weight matrix:
- determining, by the compiler, coordinates and dimensions for a multiplication matrix within the padded input data element matrix to be used with the weight value;
- and
- reassigning, by the compiler, the coordinates and the dimensions for the multiplication matrix within the padded input data element matrix to exclude the padding elements;
- and
- generating, by the compiler, compiled code to be executed by a computation engine having a systolic array having columns and rows of processing elements, the compiled code incorporating the reassigned coordinates and the reassigned dimensions;
- wherein matrix multiplications using the reassigned coordinates and the reassigned dimensions are to be performed at the computation engine and based on executing the compiled code.
6 more elements
Dependencies are inferred from the claim text.