Tabla is an innovative framework that accelerates a class of statistical machine learning algorithms. It consists of a domain specific language, Design Builder, a predesigned template, and a model compiler.
If you would like to generate the lexer and parser, please refer to the "To generate lexer and parser directly" section below. Otherwise, you only need Python 3.4.3 or higher, in order to successfully run the compiler. If you would like to view the graphical representations of the compiler-generated dataflow graphs, Graphviz - graph visualization software - is needed. Please refer to the respective online resources in order to install them on your environment.
To run the compiler, run the following command:
$ python3 main.py <*.t file>
This generates a JSON representation of data flow graph and schedule each in a separate file. It also creates a Dot file for a visual representation of data flow graph. Note that this reflects the graph after scheduling is done; every node in the same horizontal level is the operations scheduled to execute in the same cycle. Run the following command to generate a jpeg file:
$ dot -Tjpeg <*.dot file> -o <filename>.jpeg
Dependencies The parser is implemented with ANTLR v4.5 parser generator, with Python as the target language. You also need Java Runtime Environment 1.6 or higher in order to run the compiler, since ANTLR is primarily written in Java.
Run the following command:
$ java -cp "/usr/local/lib/antlr-4.5-complete.jar:$CLASSPATH" org.antlr.v4.Tool -Dlanguage=Python3 Tabla.g
If you would like to see the lexer tokens, run:
$ python3 pygrun.py Tabla program --tokens TEST_FILE.t
This compiler was developed by Joon Kyung Kim and Chenkai Shao, both undergraduate students at Georgia Institute of Technology. For any inquiries, please contact jkkim@gatech.edu or cshao31@gatech.edu.
The design builder converts all the configuration provided by the compiler to customize the hardware template.
run:
$ cd design-builder
$ python builder.py
The TABLA template design is a clustered hierarchical architecture constituting
- Processing Units (PUs)
- Processing Engines (PEs).
This clustered template architecture is scalable, general, and highly customizable.
The code for the entire template, memory interface, FPGA wrapper is in fpga/hw-imp.
fpga/hw-imp/source -> source Verilog files fpga/hw-imp/source/mem_interface -> source files which have the memory interface Verilog files fpga/hw-imp/source/ALU -> compute Verilog modules that perform the arithmetic functions fpga/hw-imp/source/basic -> basic multiplexer and other files
To simulate this code. Change the top module in fpga/hw-imp/tb.list and run from fpga directory:
$ make test