How to Set Up a Java/Android Dev Environment

Overview

The Coral NPU stack uses two language ecosystems:

Kotlin/JVM (SKaiNET): model definition, tensor DSL, HLO compilation
Python (iree-tools): MLIR transpilation, simulator integration

This guide sets up both, using jenv for Java version management and uv for Python.

Install jenv (Java Version Manager)

jenv lets you switch between JDK versions per-project without modifying system settings.

# Clone jenv
git clone https://github.com/jenv/jenv.git ~/.jenv

# Add to shell (zsh)
echo 'export PATH="$HOME/.jenv/bin:$PATH"' >> ~/.zshrc
echo 'eval "$(jenv init -)"' >> ~/.zshrc
source ~/.zshrc

# Enable Gradle plugin (sets JAVA_HOME automatically)
jenv enable-plugin gradle
jenv enable-plugin export

Install JDK 21

SKaiNET requires JDK 21+ for the JDK Vector API (SIMD operations).

Ubuntu/Debian

sudo apt update
sudo apt install openjdk-21-jdk

# Register with jenv
jenv add /usr/lib/jvm/java-21-openjdk-amd64

macOS (Homebrew)

brew install openjdk@21
jenv add /opt/homebrew/opt/openjdk@21/libexec/openjdk.jdk/Contents/Home

SDKMAN Alternative

curl -s "https://get.sdkman.io" | bash
sdk install java 21.0.4-tem
jenv add ~/.sdkman/candidates/java/21.0.4-tem

Set JDK 21 as Default

# Set globally
jenv global 21

# Or set per-project (creates .java-version file)
cd coral/SKaiNET
jenv local 21

# Verify
java -version
# openjdk version "21.x.x" ...

Install uv (Python Environment Manager)

uv manages Python versions and virtual environments. No system Python installation needed.

curl -LsSf https://astral.sh/uv/install.sh | sh

# Verify
uv --version

Set Up iree-tools

cd coral/iree-tools
uv sync  # Creates .venv and installs all dependencies

This installs:

Python 3.11+ (managed by uv)
numpy (for simulator data handling)
iree-compiler and iree-runtime (for host verification)

Install Bazel

Bazel builds the Coral NPU cross-compilation toolchain and simulator.

# Via Bazelisk (recommended — auto-downloads correct Bazel version)
npm install -g @aspect-bazel/bazelisk

# Verify
bazel --version
# bazel 7.4.1

Build SKaiNET

cd coral/SKaiNET

# Ensure JDK 21 is active
jenv local 21

# Build all modules
./gradlew build

# Build specific modules
./gradlew :skainet-lang:skainet-lang-core:build
./gradlew :skainet-compile:skainet-compile-hlo:build
./gradlew :skainet-apps:skainet-grayscale-cli:build

# Run tests
./gradlew test

Key Gradle Tasks for Android/Java Developers

Task Purpose

Task	Purpose
`./gradlew build`	Build all modules
`./gradlew test`	Run all tests
`./gradlew :skainet-apps:skainet-grayscale-cli:run --args="--help"`	Run the grayscale CLI
`./gradlew publishToMavenLocal`	Publish to local Maven repo (for Android project integration)
`./gradlew :skainet-compile:skainet-compile-hlo:generateHlo`	Export StableHLO MLIR
`./gradlew koverHtmlReport`	Generate test coverage report

./gradlew build

Build all modules

./gradlew test

Run all tests

./gradlew :skainet-apps:skainet-grayscale-cli:run --args="--help"

Run the grayscale CLI

./gradlew publishToMavenLocal

Publish to local Maven repo (for Android project integration)

./gradlew :skainet-compile:skainet-compile-hlo:generateHlo

Export StableHLO MLIR

./gradlew koverHtmlReport

Generate test coverage report

Build the MPACT Simulator

cd coral/coralnpu
bazel build //sim:coralnpu_v2_sim

The simulator requires the host Clang toolchain constraint (//toolchain/host_clang:clang_compiler). If Clang is not installed, sudo apt install clang (Ubuntu/Debian).

Project Structure for Android Integration

To use SKaiNET in an Android project, add the Maven dependency after publishing locally:

// build.gradle.kts (Android module)
dependencies {
    implementation("sk.ainet:skainet-lang-core:0.1.0")
    implementation("sk.ainet:skainet-compile-hlo:0.1.0")
    implementation("sk.ainet:skainet-io-image:0.1.0")
}

Android-Specific Considerations

JDK Vector API: Not available on Android. Use skainet-backend-cpu with the non-SIMD fallback, or target the native Kotlin Multiplatform target for computation-heavy workloads.
Model Deployment: For on-device inference, export to StableHLO and compile offline. Ship the compiled artifact (ELF for NPU, or native library for CPU) with the APK.
Memory: Android devices have more RAM than the Coral NPU, but you should still profile memory usage when working with large models.

Verify the Full Stack

Run the end-to-end pipeline to confirm everything works:

# 1. Build SKaiNET
cd coral/SKaiNET && jenv local 21 && ./gradlew build

# 2. Run grayscale CLI
./gradlew :skainet-apps:skainet-grayscale-cli:run \
  --args="--input test.jpg --verbose"

# 3. Run iree-tools pipeline
cd ../iree-tools && uv run python main.py run-all rgb2grayscale.mlir

# 4. Build and run on simulator
cd ../coralnpu && bazel run //sim:coralnpu_v2_sim -- \
  $(pwd)/bazel-bin/examples/generated/rgb2grayscale/coralnpu_v2_rgb2grayscale.elf