Allow setting thread count from Java

Summary: Allow passing in an explicit thread count through the Java bindings.

Differential Revision: D74676049

Author: GregoryComer

.gitignore

@@ -40,3 +40,6 @@ xcuserdata/
 .swiftpm/
 *.xcworkspace/
 *.xcframework/
+
+# Android
+*.aar

extension/android/executorch_android/src/main/java/org/pytorch/executorch/Module.java

@@ -52,7 +52,23 @@ public static Module load(final String modelPath, int loadMode) {
     if (!NativeLoader.isInitialized()) {
       NativeLoader.init(new SystemDelegate());
     }
-    return new Module(new NativePeer(modelPath, loadMode));
+    return new Module(new NativePeer(modelPath, loadMode, 0));
+  }
+
+  /**
+   * Loads a serialized ExecuTorch module from the specified path on the disk.
+   *
+   * @param modelPath path to file that contains the serialized ExecuTorch module.
+   * @param loadMode load mode for the module. See constants in {@link Module}.
+   * @param numThreads the number of threads to use for inference. A value of 0 defaults to a
+   *     hardware-specific default.
+   * @return new {@link org.pytorch.executorch.Module} object which owns the model module.
+   */
+  public static Module load(final String modelPath, int loadMode, int numThreads) {
+    if (!NativeLoader.isInitialized()) {
+      NativeLoader.init(new SystemDelegate());
+    }
+    return new Module(new NativePeer(modelPath, loadMode, numThreads));
   }
 
   /**

extension/android/executorch_android/src/main/java/org/pytorch/executorch/NativePeer.java

@@ -28,10 +28,11 @@ class NativePeer {
   private final HybridData mHybridData;
 
   @DoNotStrip
-  private static native HybridData initHybrid(String moduleAbsolutePath, int loadMode);
+  private static native HybridData initHybrid(
+      String moduleAbsolutePath, int loadMode, int numThreads);
 
-  NativePeer(String moduleAbsolutePath, int loadMode) {
-    mHybridData = initHybrid(moduleAbsolutePath, loadMode);
+  NativePeer(String moduleAbsolutePath, int loadMode, int numThreads) {
+    mHybridData = initHybrid(moduleAbsolutePath, loadMode, numThreads);
   }
 
   /** Clean up the native resources associated with this instance */

extension/android/jni/jni_layer.cpp

@@ -222,11 +222,15 @@ class ExecuTorchJni : public facebook::jni::HybridClass<ExecuTorchJni> {
   static facebook::jni::local_ref<jhybriddata> initHybrid(
       facebook::jni::alias_ref<jclass>,
       facebook::jni::alias_ref<jstring> modelPath,
-      jint loadMode) {
-    return makeCxxInstance(modelPath, loadMode);
+      jint loadMode,
+      jint numThreads) {
+    return makeCxxInstance(modelPath, loadMode, numThreads);
   }
 
-  ExecuTorchJni(facebook::jni::alias_ref<jstring> modelPath, jint loadMode) {
+  ExecuTorchJni(
+      facebook::jni::alias_ref<jstring> modelPath,
+      jint loadMode,
+      jint numThreads) {
     Module::LoadMode load_mode = Module::LoadMode::Mmap;
     if (loadMode == 0) {
       load_mode = Module::LoadMode::File;
@@ -248,11 +252,10 @@ class ExecuTorchJni : public facebook::jni::HybridClass<ExecuTorchJni> {
     // Based on testing, this is almost universally faster than using all
     // cores, as efficiency cores can be quite slow. In extreme cases, using
     // all cores can be 10x slower than using cores/2.
-    //
-    // TODO Allow overriding this default from Java.
     auto threadpool = executorch::extension::threadpool::get_threadpool();
     if (threadpool) {
-      int thread_count = cpuinfo_get_processors_count() / 2;
+      int thread_count =
+          numThreads != 0 ? numThreads : cpuinfo_get_processors_count() / 2;
       if (thread_count > 0) {
         threadpool->_unsafe_reset_threadpool(thread_count);
       }