Vela Bare-Metal Example Generator

This repo contains:

• a script that converts the raw output provided by Vela into C that directly invokes the Ethos driver
• A proof-of-concept example output for the MLPerf Tiny KWS model
• A snapshot of the ethos driver and a customer toolchain file to compile it for Ambiqsuite-compatible binaries
• A neuralspot-compatible wrapper that gets that compiling

Limitations

• Only single-command-stream models are supported. Any model that requires CPU fallback will fail to convert.
• Only GCC (arm-none-eabi-gcc, specifically) is supported
• This has not been tested on real hardware. We only know that it compiles.

Process

1. Compile the model using Vela
2. Convert the resulting raw output file (a serialized numpy file) into C
3. Compile the resulting C as needed (at minimum you need a main() and a pointer to the Ethos register base)

Pipeline Runner (Recommended)

The run_vela_pipeline.py script automates the entire process: running Vela, converting raw output to C, generating reference input/output arrays, and extracting them to text files.

Basic Usage

python3 run_vela_pipeline.py example_models/ic/ic.tflite

This will generate all files in a default output directory example_models/ic/ic_output/.

Custom Configuration

python3 run_vela_pipeline.py ic.tflite \
    --output-dir output/ic \
    --accelerator-config ethos-u85-256 \
    --system-config AmbiqLP_SRAM \
    --memory-mode Sram_Only_256KB

Options

• --output-dir: Custom output directory for all generated files.
• --vela-config: Path to Vela configuration file (default: config/ambiq_final.ini).
• --skip-array-to-txt: Skip generating the .txt files in src/.
• --clean: Clean the output directory before running.

Manual Steps

The following sections describe how to run each step of the pipeline manually.

Compiling your model using Vela

Vela needs to be told about memory and other HW configuration in a *.ini file. See /configs for several examples. Here we'll use bobby.ini:

pip install ethos-u-vela
vela --accelerator-config ethos-u85-256 mobilenet_v3_sm_min_s8_md.tflite --output-format raw --config ambiq_final.ini --system-config AmbiqLP_SRAM --memory-mode Sram_Only_256KB

Converting model to C

python3 vela_raw_to_c.py ../output/kws_ref_model_aligned_vela.npz --out-dir . --prefix foo

Run generate_c_arrays.py

Run generate_c_arrays.py to get the reference inputs and outputs as a header file. If -o,--output is not set will generate <model_name>_data.h in the same path as the tflite model.

python3 python/generate_c_arrays.py ../<path to tflite>/kws_ref_model.tflite

Converting C to txt

The array_2_txt.py script automatically extracts arrays from C header files and generates text files:

• input.txt (from *_input array)
• golden_output.txt (from *_output array)
• weights.txt (from *_weights array)
• cmd_data.txt (from *_cmd_data array)

The script is automatically run as part of the pipeline, but can also be run manually:

python3 python/array_2_txt.py model_data.h model_cmd_data.h model_weights.h -o src/

Slicing Models

The slice_tflite.py script can split a TFLite model into multiple slices (prefixes of the operator sequence) for incremental testing or debugging. Each slice contains the first N operators of the model.

Basic Usage

# Slice model into chunks of 5 operators each
python3 python/slice_tflite.py example_models/resnet_v1_8_32_tfs_int8/resnet_v1_8_32_tfs_int8_17.tflite --step 5

This creates subfolders slice_1/, slice_2/, etc., each containing a .tflite file with progressively more operators.

Running Pipeline on Each Slice

To automatically run the full vela pipeline on each slice:

python3 python/slice_tflite.py example_models/resnet_v1_8_32_tfs_int8/resnet_v1_8_32_tfs_int8_17.tflite --step 5 --run-pipeline

This will:

1. Create slices in slice_1/, slice_2/, etc.
2. Run the full vela pipeline for each slice
3. Generate all outputs (C files, txt files) in slice_N/output/ for each slice

Custom Options

# Custom step size and output directory
python3 python/slice_tflite.py model.tflite --step 10 --output-dir slices/

# With custom vela configuration
python3 python/slice_tflite.py model.tflite --step 5 --run-pipeline \
    --accelerator-config ethos-u55-128 \
    --system-config AmbiqLP \
    --memory-mode Sram_Only_256KB \
    --vela-verbose

# Skip certain pipeline steps
python3 python/slice_tflite.py model.tflite --step 5 --run-pipeline \
    --skip-c-arrays \
    --skip-array-to-txt

Output Structure

With --run-pipeline, each slice gets its own output directory:

slice_1/
  ├── model_1.tflite
  └── output/
      ├── model_1_vela.npz
      ├── model_1_cmd_data.h
      ├── model_1_weights.h
      ├── model_1_data.h
      └── src/
          ├── input.txt
          ├── golden_output.txt
          ├── weights.txt
          └── cmd_data.txt
slice_2/
  └── ...

Compiling Ethos Driver

The repo contains contains a pre-built ethos static lib ready for linking into a C project, but if you need to modify something, here is how to compile

For neuralSPOT and Ambiqsuite

$> rm -rf build
$> cmake -B build  -DCMAKE_TOOLCHAIN_FILE=./arm-none-eabi-gcc.cmake  -DTARGET_CPU=cortex-m55  -DETHOSU_TARGET_NPU_CONFIG=ethos-u85-256  -DCMSIS_PATH=../../clean/neuralSPOT/extern/CMSIS/CMSIS_5-5.9.0 && cmake --build build -j

For softfp

$> rm -rf build
$> cmake -B build  -DCMAKE_TOOLCHAIN_FILE=./arm-none-eabi-gcc.cmake  -DTARGET_CPU=cortex-m55+nofp  -DETHOSU_TARGET_NPU_CONFIG=ethos-u85-256  -DCMSIS_PATH=../../clean/neuralSPOT/extern/CMSIS/CMSIS_5-5.9.0 && cmake --build build -j

For building inside vela_for_neuralspot example

$> cd example/vela_for_neuralspot/src/ethos-u-core-driver-nogithub
$> rm -rf build
$> cmake -B build  -DCMAKE_TOOLCHAIN_FILE=./arm-none-eabi-gcc.cmake  -DTARGET_CPU=cortex-m55  -DETHOSU_TARGET_NPU_CONFIG=ethos-u85-256  -DCMSIS_PATH=../../../../../neuralSPOT/extern/CMSIS/CMSIS_5-5.9.0 && cmake --build build -j

Compiling the example

This example is designed to compile in neuralSPOT, but it should be pretty straightforward to compile into bare metal C. To compile the example:

$> cp -R example/vela_for_neuralspot ../neuralSPOT/apps/experiments
$> cd ../neuralSPOT
$> make EXAMPLE=experiments/vela_for_neuralspot