Skip to content

Latest commit

 

History

History
88 lines (61 loc) · 3.12 KB

README.md

File metadata and controls

88 lines (61 loc) · 3.12 KB
Raw

Description

This folder contains the testbench to validate the functionality of our hardware, including butterfly matrix multiplication, fast fourier transform, layer nomalizationa and shortcut addition. All these layers are run on one unifed hardware engine as mentioned in our paper.

Test Data Generation

Before running our testbench, you will need to generate test data, including inputs, weights and golden outputs. The training script is in the subfolder data_gen. You will need to install the same environment of running our software.

Butterfly Test Data

You can generate the test data for butterfly matrix multiplication by running:

python data_gen/torch_float16_bfly.py --length 512

You can specify your length from {128, 256, 512, 1024} depending on the testbench you wan to run.

To evalute butterfly with short-cut addition enabled, run:

python data_gen/torch_float16_bfly_sc.py --length 512

To evalute butterfly with layer normalization and short-cut addition enabled, run:

python data_gen/torch_float16_bfly_ln_sc.py --length 512

FFT Test Data

python data_gen/torch_float16_fft.py --length 512

You can specify your length from {128, 256, 512, 1024} accordingly, depending on the testbench you wan to run.

To evalute fft with short-cut addition enabled, run:

python data_gen/torch_float16_fft_sc.py --length 512

To evalute fft with layer normalization and short-cut addition enabled, run:

python data_gen/torch_float16_fft_ln_sc.py --length 512

Run Test Testbench on Vivado Behavior Simulation

To run our test tenbench, you will need to

  1. Change the file path of input, weight and output in the system verilog testbench.
  2. Set the testbench you want to run as the top module in 'sim' file (Right click and select Set As Top).
  3. Click the Run All button to kick off the simulation.

Butterfly Layer

For instance, to evaluate the Butterfly Matrix Multiplication layer, you could select the following system verilog files as top module in sim folder.

tb_butterfly_be32_length128.sv # Sequence Length = 128, Number of BE = 32
tb_butterfly_be32_length256.sv # Sequence Length = 256, Number of BE = 32
tb_butterfly_be32_length512.sv # Sequence Length = 512, Number of BE = 32
tb_butterfly_be32_length1024.sv # Sequence Length = 1024, Number of BE = 32

FFT Layer

Similarly, to evaluate FFT layer, you could select tb_fft_be\*_length\*.sv

tb_fft_be32_length256.sv # Sequence Length = 256, Number of BE = 32
tb_fft_be32_length1024.sv # Sequence Length = 1024, Number of BE = 32

You could also create your own testbench to evaluate different lengths and numbers of BEs. In this case, the test data will need to be generated accordingly.

Butterfly and FFT with post processing (LN and SC)

We also include the Test bench to evaluate the functionality of layer normalization (LN) and shortcut addition (SC).

For instance, you can run butterfly layer with SC enabled using the system verilog file.

tb_butterfly_sc_be32_length256.sv

To run butterfly layer with LN and SC enabled using:

tb_butterfly_ln_sc_be32_length256.sv