Run MobileSAM prompt segmentation on Arm Ethos-U85 with ExecuTorch
Introduction
Understand the MobileSAM deployment workflow
Prepare the ExecuTorch and Arm environment
Export MobileSAM for Ethos-U85
Build and run MobileSAM on the Corstone-320 FVP
Validate the MobileSAM segmentation result
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

Build the Arm executor runner

Run the remaining commands from the ExecuTorch repository root.

Load the Arm tools installed during environment setup:

    

        
        
source examples/arm/arm-scratch/setup_path.sh

Build the standard Arm executor runner with the exported MobileSAM program:

    

        
        
backends/arm/scripts/build_executor_runner.sh \
  --pte=arm_test/mobilesam/export/mobilesam.pte \
  --target=ethos-u85-256 \
  --output=arm_test/mobilesam/runner \
  '--extra_build_flags=-DSEMIHOSTING=ON -DET_COMPILED_PTE=ON'

The script builds the required ExecuTorch libraries and the runner for Corstone-320. ET_COMPILED_PTE=ON embeds mobilesam.pte in the Executable and Linkable Format (ELF) file named arm_executor_runner, under arm_test/mobilesam/runner/.

SEMIHOSTING=ON lets the simulated target read input tensors from host files and write its output tensors back to the host.

Prepare the target input

Copy the preprocessed image tensor into the directory used for target input and output. Remove any output from an earlier run so that validation uses a fresh result:

    

        
        
mkdir -p arm_test/mobilesam/io
cp arm_test/mobilesam/export/input.bin arm_test/mobilesam/io/input.bin
rm -f arm_test/mobilesam/io/output-0.bin

The runner reads input.bin at runtime. The image is not compiled into the ELF.

Run the application on the FVP

On macOS, the FVPs-on-Mac wrapper needs access to the repository files inside its container. Set its mount and working directories from the ExecuTorch repository root, preserving any existing configuration:

    

        
        
if [[ "$(uname -s)" == "Darwin" ]]; then
  export FVP_MOUNT_DIR="${FVP_MOUNT_DIR:-$PWD}"
  export FVP_WORKDIR="${FVP_WORKDIR:-$PWD}"
fi

Enable pipeline failure reporting, then run the ELF on the Corstone-320 Fixed Virtual Platform (FVP):

    

        
        
set -o pipefail

backends/arm/scripts/run_fvp.sh \
  --elf=arm_test/mobilesam/runner/arm_executor_runner \
  --target=ethos-u85-256 \
  --timeout=300 \
  --semihosting-cwd=arm_test/mobilesam/io \
  '--semihosting-cmd-line=executor_runner -i input.bin -o output' \
  --fast 2>&1 | tee arm_test/mobilesam/fvp.log

The runner reads the tensor specified by -i input.bin and writes its first output tensor to output-0.bin under arm_test/mobilesam/io/. The console transcript is saved to arm_test/mobilesam/fvp.log; validation reads the tensor file.

Confirm that the run created the output file before continuing:

    

        
        
test -s arm_test/mobilesam/io/output-0.bin

The check exits successfully without printing anything when the file exists and is not empty. The next page checks its shape and segmentation agreement.

What you’ve accomplished and what’s next

You’ve built the Arm executor runner with MobileSAM and run it on a virtual Cortex-M85 and Ethos-U85 system.

Next, you’ll reconstruct the target mask and compare it with the host result.

Back
Next