[Android benchmarking] Cool down the device before a run

This could reduce the variance.
Example: 25.44 → 24.74 which is good variance.

Author: kirklandsign

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

@@ -109,6 +109,10 @@ phases:
       - |
         adb -s $DEVICEFARM_DEVICE_UDID shell am force-stop org.pytorch.minibench
 
+        adb -s $DEVICEFARM_DEVICE_UDID shell dumpsys deviceidle force-idle
+        adb -s $DEVICEFARM_DEVICE_UDID shell dumpsys deviceidle unforce
+        adb -s $DEVICEFARM_DEVICE_UDID shell sleep 10
+
         if [ -n "$BIN_FOUND" ]; then
           adb -s $DEVICEFARM_DEVICE_UDID shell am start -W -n org.pytorch.minibench/.LlmBenchmarkActivity \
             --es "model_dir" "/data/local/tmp/minibench" \

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

@@ -95,6 +95,12 @@ protected void onPostExecute(Void aVoid) {
             new BenchmarkMetric(
                 benchmarkModel,
                 "avg_inference_latency(ms)",
+                stats.latency.stream().mapToDouble(l -> l).average().orElse(0.0f),
+                0.0f));
+        results.add(
+            new BenchmarkMetric(
+                benchmarkModel,
+                "trimmean_inference_latency(ms)",
                 usedLatencyResults.stream().mapToDouble(l -> l).average().orElse(0.0f),
                 0.0f));
         // Model load time