shizhouxing / MIPS32_CPU

MIPS32 CPU on FPGA with a demo sequence-to-sequence chatbot (Computer Organization, Fall 2018)

MIPS32 CPU

Requirements

• THUCST Thinpad rev.3 (early 2018)
• Vivado 2018
• mips-mti-elfi toolchain
• Python 2.7
• Tensorflow 1.3.0

Monitor

To build and run the monitor:

1. Use Vivado to compile the project, yielding a .bit file
2. Run make ON_FPGA=y at thinpad_top.test directory, yielding the kernel program in binary format kernel.bin
3. Write kernel.bin into the Flash
4. Write the .bit file into the FPGA
5. Connect the terminal to the thinpad
6. Click the reset button

The screen should display "MONITOR for MIPS32 - initialized" after these steps.

For more information, please refer to https://github.com/z4yx/supervisor-mips32.

Chatbot

We've developed a sequence-to-sequence chatbot for our MIPS32 CPU. We implemented the model in Python 2.7 and Tensorflow first, and we then implemented a MIPS32 Assembly version of the inference code for the CPU we developed. The code of the chatbot is located at /app.

Based on the previously prepared FPGA, you may follow the steps below to get the chatbot work:

1. Prepare the datasets at the app/data directory. We use:

   • Cornell Movie-Dialogs Corpus
   • PersonaChat ConvAI2 Dataset

   And we also use the 50-dimensional GloVe word vectors.

2. Train the sequence-to-sequence model:

cd app/tf
python data_pre.py
python main.py --is_train

3. Dump the model parameters:

cd app/tf
python main.py --run_dump
cp params.S ../as

4. Build the chatbot for our MIPS32 CPU:

cd app/as
make

5. Write params.bin into the BaseRAM
6. Write kernel.bin into the Flash
7. Click the reset button
8. Now you can chat with the chatbot!