Android ExecuTorch :Improve ET Dump (Prod / Release ready) with Runtime control

[psiddh]

ETDump Android Runtime Control - Feature Summary

🎯 What Changed

Before: Hardcoded profiling, always-on overhead, if turned on. Offers very less flexible code, not ready for Prod
After: Runtime control, zero overhead when disabled, production-ready

---

✨ Key Improvements

1. Runtime Control (Most Important!)

// OLD: Always profiling if built with flag (can't turn off)
Module module = Module.load("model.pte");
module.forward(inputs); // ALWAYS profiling = ALWAYS overhead

// NEW: Enable/disable on-demand
module.forward(inputs); // FAST! No profiling by default

ETDump.enableProfiling("/sdcard/profile.etdump");
module.forward(inputs); // NOW profiling
ETDump.disableProfiling();
module.forward(inputs); // FAST again!

2. Configurable Output Paths

// OLD: Hardcoded /data/local/tmp/result.etdump (needs root!)
module.etdump(); // ❌ Requires rooted device

// NEW: Any writable path
ETDump.enableProfiling(getCacheDir() + "/profile.etdump"); // ✅ App directory
ETDump.enableProfiling(getExternalFilesDir(null) + "/debug.etdump"); // ✅ External storage
ETDump.enableProfiling("/sdcard/Android/data/com.myapp/profile.etdump"); // ✅ App-specific

3. Zero Overhead

// OLD: 2-5% performance penalty always present
for (int i = 0; i < 1000; i++) {
module.forward(inputs); // ALL runs slowed down
}

// NEW: Zero overhead when disabled
for (int i = 0; i < 1000; i++) {
module.forward(inputs); // FULL SPEED!
}

4. Single APK Deployment

// OLD: Need TWO APKs
// my-app-debug.apk (with profiling, slow)
// my-app-release.apk (without profiling, can't debug)

// NEW: ONE APK with runtime control
if (BuildConfig.DEBUG || remoteConfig.enableProfiling) {
ETDump.enableProfiling(path);
}
// Same APK in production and debug! ✅

---

🎓 Use Case Examples

Use Case 1: Debug Slow Inferences in Production

long start = System.currentTimeMillis();
module.forward(inputs);
long duration = System.currentTimeMillis() - start;

if (duration > 500) { // Slower than expected
// Automatically capture profile for analysis
ETDump.enableProfiling(getCacheDir() + "/slow_run_" + duration + "ms.etdump");
module.forward(inputs); // Re-run with profiling
ETDump.writeETDump();
uploadToAnalyticsServer(ETDump.getETDumpData()); // Send to team
ETDump.disableProfiling();
}

Use Case 2: A/B Test Model Performance

// Compare two models side-by-side
ETDump.enableProfiling(getCacheDir() + "/model_v1.etdump");
Module modelV1 = Module.load("model_v1.pte");
long timeV1 = benchmarkModel(modelV1);
ETDump.writeETDump();

ETDump.enableProfiling(getCacheDir() + "/model_v2.etdump");
Module modelV2 = Module.load("model_v2.pte");
long timeV2 = benchmarkModel(modelV2);
ETDump.writeETDump();

// Analyze which is faster
Log.i(TAG, "Model V1: " + timeV1 + "ms, Model V2: " + timeV2 + "ms");
// Use Inspector to see operator-level breakdown

Use Case 3: Profile 1 of 1000 Inferences

// Only profile one specific run
for (int i = 0; i < 1000; i++) {
if (i == 42) {
ETDump.enableProfiling(getCacheDir() + "/inference_42.etdump");
}

  module.forward(inputs);  // 999 fast, 1 profiled

  if (i == 42) {
      ETDump.writeETDump();
      ETDump.disableProfiling();
  }

}

Use Case 4: Remote-Controlled Profiling

public class ModelRunner {
void runInference() {
// Check remote config (Firebase, etc.)
if (FirebaseRemoteConfig.getInstance().getBoolean("enable_profiling")) {
String userId = getCurrentUserId();
ETDump.enableProfiling(getCacheDir() + "/profile_" + userId + ".etdump");
}

      module.forward(inputs);

      if (ETDump.isProfilingEnabled()) {
          ETDump.writeETDump();
          uploadToServer(ETDump.getETDumpData());
          ETDump.disableProfiling();
      }
  }

}
// Enable profiling for specific users without rebuilding! ✅

Use Case 5: Conditional Profiling (Device-Specific)

// Only profile on high-end devices
boolean isHighEnd = Runtime.getRuntime().availableProcessors() >= 8;

if (isHighEnd && BuildConfig.DEBUG) {
ETDump.enableProfiling(getExternalFilesDir(null) + "/profile.etdump");
module.forward(inputs);
ETDump.writeETDump();
ETDump.disableProfiling();
} else {
module.forward(inputs); // Fast, no profiling
}

Use Case 6: CI/CD Performance Regression Tests

@test
public void testModelPerformanceRegression() {
String etdumpPath = new File(context.getCacheDir(), "ci_test.etdump").getAbsolutePath();

  ETDump.enableProfiling(etdumpPath);
  Module module = Module.load(modelPath);

  long start = System.nanoTime();
  module.forward(inputs);
  long duration = System.nanoTime() - start;

  ETDump.writeETDump();

  // Assert performance hasn't regressed
  assertTrue("Model too slow: " + duration + "ns", duration < 50_000_000); // 50ms

  // Can analyze ETDump file in CI for operator-level breakdowns
  byte[] profileData = ETDump.getETDumpData();
  assertNotNull(profileData);

}

Use Case 7: Compare Delegate Performance

// Profile XNNPACK vs default backend
ETDump.enableProfiling(getCacheDir() + "/xnnpack.etdump");
Module xnnpackModel = Module.load("model_xnnpack.pte");
xnnpackModel.forward(inputs);
ETDump.writeETDump();

ETDump.enableProfiling(getCacheDir() + "/cpu.etdump");
Module cpuModel = Module.load("model_cpu.pte");
cpuModel.forward(inputs);
ETDump.writeETDump();

// Use ExecuTorch Inspector to compare:
// python -m executorch.sdk.inspector.inspector_cli
// --etdump_path xnnpack.etdump --model_path model_xnnpack.pte

Use Case 8: Programmatic Upload (No File I/O)

// Get profiling data without writing to file
ETDump.enableProfiling("/tmp/dummy.etdump"); // Path required but not used
module.forward(inputs);

byte[] profileData = ETDump.getETDumpData(); // Get data directly
if (profileData != null) {
// Upload to analytics server
analyticsClient.upload("model_profile", profileData);

  // Or compress and send
  byte[] compressed = gzip(profileData);
  sendToServer(compressed);

}

ETDump.disableProfiling();

🚀 Complete Example

public class InferenceActivity extends Activity {
private Module module;

  @Override
  protected void onCreate(Bundle savedInstanceState) {
      super.onCreate(savedInstanceState);

      // Load model (profiling disabled by default)
      module = Module.load(getModelPath());
  }

  public void runInference(Tensor input) {
      // Normal fast inference
      EValue[] outputs = module.forward(EValue.from(input));

      // Display results
      updateUI(outputs);
  }

  public void debugSlowInference(Tensor input) {
      // Enable profiling for debugging
      String profilePath = getCacheDir() + "/debug_" + System.currentTimeMillis() + ".etdump";
      ETDump.enableProfiling(profilePath);

      long start = System.nanoTime();
      EValue[] outputs = module.forward(EValue.from(input));
      long duration = System.nanoTime() - start;

      // Save profile
      ETDump.writeETDump();
      ETDump.disableProfiling();

      Log.i(TAG, "Profiled inference took " + (duration / 1_000_000) + "ms");
      Log.i(TAG, "Profile saved to: " + profilePath);

      // Can now analyze with:
      // adb pull <profilePath> .
      // python -m executorch.sdk.inspector.inspector_cli --etdump_path profile.etdump
  }

}

---

✅ Bottom Line

One feature, massive improvements:

• ✅ Zero overhead in production
• ✅ Debug issues on-demand
• ✅ Single APK deployment
• ✅ Full ExecuTorch Inspector integration

Build once, profile when needed!

[psiddh]

Feature WIP : #16359