The OneWare VHDL Documentation

When exporting your trained model you have the option to export as a VHDL project for an FPGA. Aside from the model itself this project contains also the VHDL code necessary to run your model. This page gives you the necessary information to integrate your OneAI model into your FPGA project.

Export Project Content

The VHDL export gives you a folder with the following content:

• Core Contains helper modules concerning CNN-architecture.
• Filters Contains helper modules for data preprocessing.
• Quartus_IP The One Ware IP for Quartus (e.g. to connect the CNN with a CPU)
• ONEAI_Confic_Package.vhd Configuration settings.
• ONEAI_Data_Package.vhd The model weights.
• ONEAI_Simulation.vhd A VHDL file you can use to simulate the model.
• ONEAI_Top.vhd The top level design entity to integrate into your project.
• Test_Data_Package.vhd Test data for simulation.

Core Concept: Stream

All modules of the OneAI-code make use of streams. These are a signal type, used to pass data into and throughout the CNN in VHDL. The stream type CNN_Stream_T is defined in the file ONEAI_Confic_Package.vhd. Streams implement the following signals:

• Data_CLK Design clock
• Data_Valid Indicator that data is valid
• Column Current pixel column
• Row Current pixel row
• Filter Current pixel filter (relevant for convolution; for the top-level input of the CNN usually leave this at 0)

This is paired with a Data signal consisting of n-dimensions depending on the images colour channels. Each channel is a 7-bit unsigned signal encoding the respecting pixel value in the range 0...127.

In case of multiple input Data signals (e.g. multi image analysis) all Data signals need to adhere to the same input stream, i.e. they must share clock, valid, column, row and filter signals.

Simulation

For simulation purposes you can simply use the ONEAI_Simulation module as the top level entity.

This simulation module will take example data included in the Test_Data_Package.vhd and run it through the model. Here you have the oportunity to verify correct inputs (iStream) and the output produced by the CNN (oStream).

Implementation

When including the OneAI model into your FPGA project simply include the ONEAI_Top.vhd module CNN into your design. The module has one input stream paired with one or more data signals. E.g. for standard RGB-image classification with one image, there is one data signal iData_1 with three channels. A multi image input would require multiple iData_x signals.

• iStream Input stream
• iData_1 Input data (n-channel)
• oStream_1 Output stream
• oData_1 Output data (n-channel)
• oCycle_1 Current output cycle

The output of the CNN-module adheres to the stream format as well. The result of the CNN is provided via the oData_1 signal. For classification models the output consists of the confidence values of the CNN for all of the classes. A model with 10 classes will produce 10 confidence values. Each of those will be produced in one clock cycle. The current class is indicated via the oCycle_1 signal. The confidence is encoded as a 7-bit unsigned integer. A large value indicates a high confidence. A percentage based confidence (range 0...100) can be derived by dividing the value of the oData_1 signal by 1.27 in the postprocessing.

Cycle	0	1	2
CNN Output	0x12	0x01	0xFE
Probabilities	0.14	0.01	0.99

Note that these are the raw confidence values produced by the model. It is customary to transform these into probabilities by using the softmax function. This post-processing step is not included in the provided VHDL-implementation and needs to be added by the user.

Also note that the output will be generated a few clock cycles after the last input pixel is reached. The exact latency depends on the number of layers of the CNN and thus varies between models. Be sure to leave the clock running after providing the last pixel data. It is possible to already start the next inference before the output is generated. Just start again with the (0,0) pixel after the last pixel of the former image was provided.

Example instantiation of the CNN-module (assuming singular image classification with 3-channel RBG input):

COMPONENT CNN
	PORT (
		iStream       : IN CNN_Stream_T;
		iData_1       : IN CNN_Values_T(2 downto 0);
		oStream_1     : OUT CNN_Stream_T;
		oData_1       : OUT CNN_Values_T(0 downto 0);
		oCycle_1      : OUT NATURAL
	);
END COMPONENT;