cgra-systemc

SystemC implementation of a parameterizable CGRA

Build

The architecture is designed in SystemC. It is compiled with different Linux distributions like Debian, Ubuntu or OpenSuse. The CMakeLists.txt contains an automatic mechanism to find the SystemC dependencies. Therefore please provide SYSTEMC_ROOT environment variable to point to the root directory of your SystemC installation.

Architecture variants

• VCGRA: A processing system acts directly with an accelerator (VCGRA). The virtual CGRA consists of processing elements and virtual channels. The processing system needs to control the calculation the whole runtime.
• VCGRA & Pre-Fetchers: A processing systems is has access to configuration pre-fetchers for processing elements and virtual channels to prepare runtime configurations for the VCGRA before switching.
• Full Architecture: The full architecture is an accelerator with a Central Control Unit, a Memory Management Unit, pre-fetchers for configurations, input and output data and a shared scratchpad memory between the accelerator and the processing system. The Central Control Unit is programmable with a special assembler and can work autark with data stored at the shared memory. The processing system does not need to perform any control during processing of data of the VCGRA accelerator architecture. The VCGRA assembler is available here.

Evaluation

Demo

Evaluation is performed processing an edge detection of an image of Lena. The configuration files for McPAT are stored in the repository in the mcpat directory. An adapted version of McPAT is necessary to overcome the limitations of McPAT that is meanly created to simulate big multicore systems. Because of this limitations this is a very rough estimation for area and power.

Evaluation Results