In this paper we present a convolutional neural network architecture that supports user customization through incremental transfer learning. The architecture consists of a large basic inference engine and a small augmenting engine. After training the basic inference engine and augmenting engine on a large general dataset, the basic inference engine is fixed. For user customization, only the augmenting engine is re-trained on-device using a small user specific dataset provided by the user. To accelerate the training of the augmenting engine we map this to a coarsegrained reconfigurable array processor. The complete network architecture is evaluated using the Caffe framework, and a C-code equivalent network is implemented and tested on a CGRA processor. Experiments with NIST'19 and our user-specific datasets show an increase in accuracy of the system from 76.3% to 93.2% after user customization. Mapping this code to a CGRA gives us a speed up of 45x and a 49-and 3-fold reduced energy consumption over an ARMv7 processor and a 3-way VLIW processor, respectively, showing the potential of CGRAs as DNN processors.

This paper presents a convolutional neural network architecture that supports transfer learning for user customization. The architecture consists of a large basic inference engine and a small augmenting engine. Initially, both engines are trained using a large dataset. Only the augmenting engine is tuned to the user-specific dataset. To preserve the accuracy for the original dataset, the novel concept of quality factor is proposed. The final network is evaluated with the Caffe framework, and our own implementation on a coarse-grained reconfigurable array (CGRA) processor. Experiments with MNIST, NIST'19, and our user-specific datasets show the effectiveness of the proposed approach and the potential of CGRAs as DNN processors.

This paper presents methods for automatic generation of models of Transducer, a highly flexible communication module for interfacing Multiprocessor System-on-Chip (MPSoC) components. We describe the transducer architecture, comprising the bus interface, high-level communication controllers and buffer management blocks. The well-defined architecture of the transducer enables automatic generation of its Transaction-level and Register-transfer level (RTL) models. Moreover, the simple interface of the transducer provides for a well-defined software interface, making it easy to update the software after changes in MPSoC platform. Our experimental results show that MPSoC design for industrial-size applications, such as MP3 decoder and JPEG encoder, greatly benefits from automatic generation of transducer models. We found productivity gains of 9-23× due to significant savings in modeling effort. On the quality axis, we show that MPSoC communication design using automatically generated transducers has very little overhead in communication delay over a fully connected point-to-point communication architecture. Finally, we show that our automatically generated TLMs greatly reduce the system-level modeling time and provide a fast executable model for early functional validation.

This paper presents synthesis of hardware dependent software (HdS) for multicore and many-core designs using embedded system environment (ESE). ESE is a tool set, developed at UC Irvine, for transaction level design of multicore embedded systems. HdS synthesis is a key component of ESE back-end design flow. We follow a design process that starts with an application model consisting of C processes communicating via abstract message passing channels. The application model is mapped to a platform net-list of SW and HW cores, buses and buffers. A high speed transaction level model (TLM) is generated to validate abstract communication between processes mapped to different cores. The TLM is further refined into a pin-cycle accurate model (PCAM) for board implementation. The PCAM includes C code for all the HdS layers including routing, packeting, synchronization and bus transfer. The generated HdS methods provide a library of application level services to the C processes on individual SW cores. Therefore, the application developer does not need to write low level HdS for board implementation. Synthesis results for an multi-core MP3 decoder design, using ESE, show that the HdS is generated in order of seconds, compared to hours of manual coding. The quality of synthesized code is comparable to manually written code in terms of performance and code size.

This paper presents a tool for automatic synthesis of RTL interfaces for heterogeneous MPSoC from transaction level models (TLMs). The tool captures the communication parameters in the platform and generates interface modules called universal bridges between buses in the design. The design and configuration of the bridges depend on several platform components including heterogeneity of the components, traffic on the bus, size of messages and so on. We define these parameters and show how the synthesizable RTL code for the bridge can be automatically derived based on these parameters. We use industrial strength design drivers such as an MP3 decoder to test our automatically generated bridges for a variety of platforms and compare them to manually designed bridges on different quality metrics. Our experimental results show that performance of automatically generated bridges are within 5% of manual design for simple platforms but surpasses them for more complex platforms. The area and RTL code size is consistently better than manual design while giving 5 orders of improvement in development time.

Communication between components, with different interface protocols, requires an extra component that must translate one protocol to another. This component is referred to as a transducer. In this paper we describe the design and implementation of a transducer between AMBA bus and TMS DSP bus. The transducer allows system designers to send data from AMBA compliant components to TMS compliant ones, and vice versa. The transducer was modeled in Verilog and implemented on Xilinx VirtexII FPGA board