Add etdump to android

.github/workflows/android-perf.yml

@@ -367,7 +367,7 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
 
         mkdir -p aar-out
-        PYTHON_EXECUTABLE=python ANDROID_ABIS="arm64-v8a" BUILD_AAR_DIR=aar-out EXECUTORCH_BUILD_QNN=ON QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029 bash scripts/build_android_library.sh
+        PYTHON_EXECUTABLE=python ANDROID_ABIS="arm64-v8a" BUILD_AAR_DIR=aar-out EXECUTORCH_BUILD_QNN=ON QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029 EXECUTORCH_ANDROID_PROFILING=ON bash scripts/build_android_library.sh
         mkdir -p extension/benchmark/android/benchmark/app/libs
         cp aar-out/executorch.aar extension/benchmark/android/benchmark/app/libs
         pushd extension/benchmark/android/benchmark

extension/android/CMakeLists.txt

@@ -80,6 +80,16 @@ list(
   fbjni
 )
 
+if(EXECUTORCH_ANDROID_PROFILING)
+  list(
+    APPEND
+    link_libraries
+    etdump
+    flatccrt
+  )
+  target_compile_definitions(executorch_jni PUBLIC EXECUTORCH_ANDROID_PROFILING=1)
+endif()
+
 if(TARGET optimized_native_cpu_ops_lib)
   list(
     APPEND

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

@@ -152,6 +152,18 @@ public String[] readLogBuffer() {
     return mNativePeer.readLogBuffer();
   }
 
+  /**
+   * Dump the ExecuTorch ETRecord file to /data/local/tmp/result.etdump.
+   *
+   * <p>Currently for internal (minibench) use only.
+   *
+   * @return true if the etdump was successfully written, false otherwise.
+   */
+  @Experimental
+  public boolean etdump() {
+    return mNativePeer.etdump();
+  }
+
   /**
    * Explicitly destroys the native Module object. Calling this method is not required, as the
    * native object will be destroyed when this object is garbage-collected. However, the timing of

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

@@ -62,4 +62,7 @@ public void resetNative() {
   /** Retrieve the in-memory log buffer, containing the most recent ExecuTorch log entries. */
   @DoNotStrip
   public native String[] readLogBuffer();
+
+  @DoNotStrip
+  public native boolean etdump();
 }

extension/android/jni/jni_layer.cpp

@@ -15,7 +15,6 @@
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
-
 #include "jni_layer_constants.h"
 
 #include <executorch/extension/android/jni/log.h>
@@ -32,6 +31,12 @@
 #include <executorch/extension/threadpool/threadpool.h>
 #endif
 
+#ifdef EXECUTORCH_ANDROID_PROFILING
+#include <executorch/devtools/etdump/etdump_flatcc.h>
+#include <fcntl.h>
+#include <unistd.h>
+#endif
+
 #include <fbjni/ByteBuffer.h>
 #include <fbjni/fbjni.h>
 
@@ -238,8 +243,13 @@ class ExecuTorchJni : public facebook::jni::HybridClass<ExecuTorchJni> {
     } else if (loadMode == 3) {
       load_mode = Module::LoadMode::MmapUseMlockIgnoreErrors;
     }
-
-    module_ = std::make_unique<Module>(modelPath->toStdString(), load_mode);
+#ifdef EXECUTORCH_ANDROID_PROFILING
+    auto etdump_gen = std::make_unique<executorch::etdump::ETDumpGen>();
+#else
+    auto etdump_gen = nullptr;
+#endif
+    module_ = std::make_unique<Module>(
+        modelPath->toStdString(), load_mode, std::move(etdump_gen));
 
 #ifdef ET_USE_THREADPOOL
     // Default to using cores/2 threadpool threads. The long-term plan is to
@@ -362,7 +372,6 @@ class ExecuTorchJni : public facebook::jni::HybridClass<ExecuTorchJni> {
       auto jevalue = JEValue::newJEValueFromEValue(result.get()[i]);
       jresult->setElement(i, *jevalue);
     }
-
     return jresult;
   }
 
@@ -396,6 +405,37 @@ class ExecuTorchJni : public facebook::jni::HybridClass<ExecuTorchJni> {
 #endif
   }
 
+  jboolean etdump() {
+#ifdef EXECUTORCH_ANDROID_PROFILING
+    executorch::etdump::ETDumpGen* etdumpgen =
+        (executorch::etdump::ETDumpGen*)module_->event_tracer();
+    auto etdump_data = etdumpgen->get_etdump_data();
+
+    if (etdump_data.buf != nullptr && etdump_data.size > 0) {
+      int etdump_file =
+          open("/data/local/tmp/result.etdump", O_WRONLY | O_CREAT, 0644);
+      if (etdump_file == -1) {
+        ET_LOG(Error, "Cannot create result.etdump error: %d", errno);
+        return false;
+      }
+      ssize_t bytes_written =
+          write(etdump_file, (uint8_t*)etdump_data.buf, etdump_data.size);
+      if (bytes_written == -1) {
+        ET_LOG(Error, "Cannot write result.etdump error: %d", errno);
+        return false;
+      } else {
+        ET_LOG(Info, "ETDump written %d bytes to file.", bytes_written);
+      }
+      close(etdump_file);
+      free(etdump_data.buf);
+      return true;
+    } else {
+      ET_LOG(Error, "No ETDump data available!");
+    }
+#endif
+    return false;
+  }
+
   facebook::jni::local_ref<facebook::jni::JArrayClass<jstring>> getUsedBackends(
       facebook::jni::alias_ref<jstring> methodName) {
     auto methodMeta = module_->method_meta(methodName->toStdString()).get();
@@ -423,6 +463,7 @@ class ExecuTorchJni : public facebook::jni::HybridClass<ExecuTorchJni> {
         makeNativeMethod("execute", ExecuTorchJni::execute),
         makeNativeMethod("loadMethod", ExecuTorchJni::load_method),
         makeNativeMethod("readLogBuffer", ExecuTorchJni::readLogBuffer),
+        makeNativeMethod("etdump", ExecuTorchJni::etdump),
         makeNativeMethod("getUsedBackends", ExecuTorchJni::getUsedBackends),
     });
   }

extension/benchmark/android/benchmark/android-llm-device-farm-test-spec.yml.j2

@@ -107,6 +107,7 @@ phases:
 
       - echo "Run benchmark"
       - |
+        adb -s $DEVICEFARM_DEVICE_UDID shell touch /data/local/tmp/result.etdump
         adb -s $DEVICEFARM_DEVICE_UDID shell am force-stop org.pytorch.minibench
 
         adb -s $DEVICEFARM_DEVICE_UDID shell dumpsys deviceidle force-idle
@@ -147,6 +148,8 @@ phases:
         # Trying to pull the file using adb ends up with permission error, but this works too, so why not
         echo "${BENCHMARK_RESULTS}" > $DEVICEFARM_LOG_DIR/benchmark_results.json
 
+        adb -s $DEVICEFARM_DEVICE_UDID pull /data/local/tmp/result.etdump $DEVICEFARM_LOG_DIR/result.etdump
+
 artifacts:
   # By default, Device Farm will collect your artifacts from the $DEVICEFARM_LOG_DIR directory.
   - $DEVICEFARM_LOG_DIR

extension/benchmark/android/benchmark/app/src/main/java/org/pytorch/minibench/ModelRunner.java

@@ -41,6 +41,8 @@ public void runBenchmark(
       latency.add(forwardMs);
     }
 
+    module.etdump();
+
     final BenchmarkMetric.BenchmarkModel benchmarkModel =
         BenchmarkMetric.extractBackendAndQuantization(model.getName().replace(".pte", ""));
     // The list of metrics we have atm includes:

scripts/build_android_library.sh

@@ -40,6 +40,8 @@ build_android_native_library() {
     -DANDROID_PLATFORM=android-26 \
     -DBUILD_TESTING=OFF \
     -DEXECUTORCH_ENABLE_LOGGING=ON \
+    -DEXECUTORCH_BUILD_DEVTOOLS=ON \
+    -DEXECUTORCH_ENABLE_EVENT_TRACER="${EXECUTORCH_ANDROID_PROFILING:-OFF}" \
     -DEXECUTORCH_LOG_LEVEL=Info \
     -DEXECUTORCH_BUILD_XNNPACK=ON \
     -DEXECUTORCH_XNNPACK_SHARED_WORKSPACE=ON \
@@ -75,6 +77,7 @@ build_android_native_library() {
     -DEXECUTORCH_ENABLE_LOGGING=ON \
     -DEXECUTORCH_LOG_LEVEL=Info \
     -DCMAKE_FIND_ROOT_PATH_MODE_PACKAGE=BOTH \
+    -DEXECUTORCH_ANDROID_PROFILING="${EXECUTORCH_ANDROID_PROFILING:-OFF}" \
     -DNEURON_BUFFER_ALLOCATOR_LIB="$NEURON_BUFFER_ALLOCATOR_LIB" \
     -DEXECUTORCH_BUILD_KERNELS_CUSTOM="${EXECUTORCH_BUILD_EXTENSION_LLM:-ON}" \
     -DEXECUTORCH_BUILD_LLAMA_JNI="${EXECUTORCH_BUILD_EXTENSION_LLM:-ON}" \