The Mathworks, Inc.

Yongfeng Gu

Girish Venkataramani

Rama P Kokku

Wang Chen

Zhihong Zhao

Partha Biswas

Purshottam Vishwakarma

Vibha Patil

Sanmukh Rao Kuppannagari

Bharathi Yogaraj

Anusha Vasantala

Yuteng Zhou

Systems and methods may configure a programmable logic device to efficiently run a deep learning (DL) network. Architecture code and algorithmic code may be generated. The architecture code may define convolutional and fully connected processor cores structured to run the layers of a Deep Neural Network (DNN). The processor cores may be interconnected by a First In First Out (FIFO) memory. The architecture code may also define stride-efficient memories for implementing convolution. The algorithmic code may include configuration instructions for running the DNN's layers at the processor cores. The algorithmic code may also include a schedule for executing the configuration instructions on the processor cores, for moving network parameters to the processor cores, and for transferring outputs between the layers.

Systems and methods for configuring programmable logic devices for deep learning networks

A system and method generates optimized code for a source model. The system may include a resource sharing optimizer that evaluates the source model and replaces multiple model elements of the source model that are functionally equivalent with a single shared model element. The model elements replaced with the single shared model element may have different fixed point data types. The resource sharing optimizer may convert some of the fixed point data types to a common fixed point data type.

Systems and methods for sharing resources having different data types

Systems and methods trace performance data generated by a hardware synthesis tool chain to model elements of a model. During code generation, an initial in-memory representation is generated for the model. The in-memory representation includes a plurality of nodes that correspond to the model elements. The in-memory representation is subjected to transformations and optimizations creating transitional in-memory representations and a final in-memory representation from which HDL code is generated. A graph builder constructs a genealogy graph that traces the transformations and optimizations. The genealogy graph includes graph objects corresponding to the nodes of the in-memory representations. The synthesis tool chain utilizes the HDL code to perform hardware synthesis. The synthesis tool chain also generates performance data. Utilizing the genealogy graph, the performance data is mapped to the nodes of the initial in-memory representation, and to the elements of the model.

Systems and methods for tracing performance information from hardware realizations to models

A device is configured to receive delay information associated with a model including a set of model elements and one or more delay elements. The delay information may identify a model element, of the set of model elements, and a quantity of delay to be associated with the model element. The model may be associated with a total quantity of delay. The device is configured to determine accumulated delay information based on the model, and to determine a set of retiming values associated with the set of model elements. The device is configured to redistribute the one or more delay elements associated with the model, based on the set of retiming values, to satisfy the quantity of delay to be associated with the model element, and to maintain the total quantity of delay associated with the model. The device is configured to provide the redistributed model.

User-constrained delay redistribution

Systems and methods automatically generate optimized hardware description language code for a model created in a modeling environment. A training tool selects and provides scripts to a hardware synthesis tool chain that direct the tool chain to synthesize hardware components for core components of the modeling environment. A report generated by the tool chain is evaluated to extract performance data for the core components, and the performance data is stored in a library. An optimization tool estimates the performance of the model using the performance data in the library. Based on the performance estimate and an analysis of the model, the optimization tool selects an optimization technique which it applies to the model generating a revised. Estimating performance, and selecting and applying optimizations may be repeated until a performance constraint is satisfied or a termination criterion is met.

Systems and methods for estimating performance characteristics of hardware implementations of executable models

A device generates a model associated with a multi-rate system. The multi-rate system includes a system associated with a clock rate and a sample rate, and the clock rate is greater than the sample rate. The device identifies the clock rate of the multi-rate system based on the model, and identifies a portion, of the model, associated with the sample rate. The device applies clock rate pipelining to adjust the sample rate associated with the portion of the model so that the sample rate substantially equals the clock rate, and generates code associated with the model and the applied clock rate pipelining.

Utilizing clock rate pipelining to generate code for multi-rate systems

Systems and methods automatically generate optimized hardware description language (HDL) code for an executable model. An intermediate representation is generated for the executable model, which includes model elements. The intermediate representation includes nodes corresponding to the model elements. The HDL code is generated from the intermediate representation. A synthesis tool chain performs hardware synthesis using the HDL code. The synthesis tool chain generates performance characteristics of hardware components defined by the synthesis tool chain. The performance characteristics are mapped to the nodes of the intermediate representation, and one or more performance bottlenecks are identified. At least one optimization technique is applied to the intermediate representation to produce a revised intermediate representation, which is then used to generate new HDL code. The process may be repeated until the performance bottlenecks are eliminated or a termination criterion is met.

Systems and methods for generating optimized hardware descriptions for models

A system and method tests for functional equivalence prior to automatically retiming a high-level specification. An Intermediate Representation (IR) includes one or more graphs or trees based on the high-level specification. A functional equivalence (FE) analyzer determines whether one or more components in the graph meet certain value and state conditions and thus is a candidate for retiming. A scheduler can use components that fail FE as a retiming boundary.

Model-based retiming with functional equivalence constraints

A system and method evaluates power information for a high-level model to be implemented in target hardware, and performs one or more power-reducing transmutations on the model. Transmutations may include moving one or more components from a fast rate region to a slow rate region, reducing bit width of data, signals, or other values, and replacing multiple instances of a resource with a shared instance of the resource. An in-memory representation of the model may be generated that reduces the model to a plurality of core components. A power score evaluation engine may assign power scores to the core components. Power scores may be retrieved from one or more power score database. The power scores may be non-dimensional scores representing power consumption relationships among the core components, and be target independent. Hints or alerts regarding suggested changes to the model to optimize power consumption may be presented to a user. A revised model incorporating the suggested changes may be constructed. The one or more transmutations resulting in a lowest total power score may be selected for hardware generation.

Model level power consumption optimization in hardware description generation

A system and method tests for functional equivalence prior to automatically retiming a high-level specification. An Intermediate Representation (IR) includes one or more graphs or trees based on the high-level specification. A functional equivalence (FE) analyzer determines whether one or more components in the graph meet certain value and state conditions and thus is a candidate for retiming. A bounded scheduler then retimes only those components that pass the FE analysis.

Growing community of inventors

Brookline, MA, United States of America

Yongfeng Gu