EE 290C Wellness Monitor: ExG

Brief Course Description

The Fall 2018 offering of EE290C, taught by Prof. Borivoje Nikolic, covers digital signal-processing algorithms, architectures, and their rapid implementation in advanced technology nodes. The class uses Chisel as a hardware construction language to design a modular system and connect it to a RISC-V system-on-a-chip. For our final project, we chose to implement a wellness monitor as described below.

Project Description

The ExG generators describe a tool box of important digital blocks for any arbitrary wellness monitor. The goal of this project is to develop a flow such that any designer could describe a wellness monitor by high level features (e.g. arrhythmia detection, EEG alpha wave monitoring, seizure detection, etc.), and the generator would build and connect the neccessary filters, FFT blocks, feature extractors, and learning mechanisms required to achieve the high level specification. More details can be found in reports.

An example datapath is shown below:

Team Members

Adelson Chua, Justin Doong, Ryan Kaveh, Cem Yalcin, and Rachel Zoll

Generators

1. Filters: FIR, IIR
2. FFT: FFT Buffer, FFT
3. Feature Extractors: Line Length, Bandpower, DWT
4. PCA
5. SVM
6. Memory Buffer
7. Wellness Monitor

Tape-in 1

For tape-in 1, generators and unit tests were written for FIR and IIR filters, FFT, bandpower, PCA and SVM. These blocks were individually tested and guaranteed working. For integration, because testing functionality would have been very difficult with some incomplete blocks, RocketChip integration was confirmed using just the FIR filter block.

Tape-in 2

Tape-in 2's primary goal was a more complete integration. Throughout the integration process, several supporting generators were written (Memory Buffer and FFT Buffer) that will connect to the input vectors and matrices for the PCA, SVM, and FFT blocks. Additional feature generators were also constructed (Line Length and a WIP Arythmia Detection). The original blocks were updated to support various new tests and functions (FIR, IIR, FFT, SVM). Finally, a top-level module was written that connects different blocks together for functionality verification. More detail regarding this is presented in the next section.

Tape-out

For tape-out, the top-level wellness monitor generator was written. The discrete wavelet transform block (DWT) was also written, but not integrated into the rest of the wellness monitor framework. The other feature extractors (line length, bandpower) help with classification of seizure (EEG) datasets, whereas the DWT provides analysis of cardiac (ECG) datasets, which we have not yet tested with SVM.

How To Set Up and Test The Current Generators

Setting up the generators is straightforward because everything is self-contained. Simply clone the repo and explore the different generators in the src folder. Refer to the doc directory doc for documentation on each generator.

There are multiple ways to test each block and whole designs: unit-tests, SBT-based integration tests, and C code tests that integrate the blocks with a RISC-V Rocket core. An application-based test setup was also developed which goes through SVM training in Python, integrated testing in Scala, and a C-based integration test on a simulated RISC-V environment.

Unit Tests

Each type-generic generator comes equipped with unit tests which feed random numbers (either integers or fixed points) to the generators and compares their outputs against golden Scala-based models (all of which are in the test directory).

To run a specific unit test, open SBT and use the 'testOnly' function:

testOnly packageName.specName

for example:

testOnly firFilter.FIRFilterSpec

Wellness Integration Tests

In order to test the entire flow, each block has been connected in a tester that generates an example system. The current test datapath involves a realistic assortment of different blocks for simple siezure detection. All data is pre-filtered and then split to separate frequency domain and time domain branches. The time domain branch consists of a single line length block (for now) while the frequency domain path involves a serial-to-parallel buffer, FFT, and different bandpower blocks. The outputs of the line length and bandpower blocks are fed into a PCA and lastly an SVM. A block diagram for this tester is shown below.

In parallel to the simulated hardware, the integration test also calculates an expected output using a Scala golden model of the entire system. The golden model uses all of the smaller unit golden models strung together to build a Scala version of the wellness datapath. Like the actual Chisel hardware, the golden model is also generated.

You can run it as part of the all encompassing SBT 'tests' or a similar testOnly as the unit tests:

testOnly wellness.wellnessGenIntegrationSpec

---

Application-specific and Comprehensive Test Setup

A test framework was developed to enable application-specific testing given some raw data that needs to be classified. To process covers testing starting from SVM training all the way to C tests in a simulated RISC-V environment.

SVM training in Python

A Python script performs SVM training from the raw data while also calculating the expected accuracies after the model has been created. This is based on the SVM modelling script discussed here. All parameters, configuration matrices, and input vectors are then written to different CSV files.

• You will need to extract the dataset to replicate this project. The data folder should be in the main tree fa18-wellness-monitor/data/.

• Set up the configuration parameters in the top level Python script top.py. There are comments in the code which explain what each parameter refers to. You can leave it 'as is' to replicate the results of this project.

• Run the top.py, all configuration files will be saved as CSV in fa18-wellness-monitor/scripts/generated_files/

Scala-based testing

The CSV files will then be read by the top-level tester. The tester sets all submodule generator parameters accordingly and passes an actual input from the test dataset through the datapath chain. Afterwards, the expected values from the SVM classifier and the configuration matrices are written to a C header file.

• Make sure that there are files in the fa18-wellness-monitor/scripts/generated_files/ directory. Else, you would need to run the Python script from the previous section.

• This test is the second test in wellnessGenIntegrationSpec.scala.

• Run the wellnessGenIntegrationSpec either as stand-alone or through sbt test. It will generate a C header file named fa18-wellness-monitor/tests/array.h

C-based integration test

The C header file will be used in the C integration test where the wellness datapath is already integrated with a RISC-V core. This checks that the expected output from the Python model are consistent and are synthesizeable in hardware.

• Make sure that the wellnessParams inside the fa18-wellness-monitor/src/test/scala/WellnessGen.scala is pointed to FixedPointModelWellnessGenParams. This gets the parameters from the same files used in the Scala-based test.

trait HasPeripheryWellness extends BaseSubsystem {

  val wellnessParams = FixedPointModelWellnessGenParams

• Build the project by running make inside the fa18-wellness-monitor/verisim folder. It will take a while.

• Create the executable for the C tester file wellness_IntegrationTest_FixedPoint.c by running make inside the fa18-wellness-monitor/tests folder. Make sure that the Makefile points to this C code.

PROGRAMS = wellness_IntegrationTest_FixedPoint

• Go back to the fa18-wellness-monitor/verisim folder and run the project:

./simulator-freechips.rocketchip.system-DefaultConfig ../tests/wellness_IntegrationTest_FixedPoint.riscv