diff --git a/neuralnetworks/aidl/utils/Android.bp b/neuralnetworks/aidl/utils/Android.bp
index 37ad6d6..9437d5c 100644
--- a/neuralnetworks/aidl/utils/Android.bp
+++ b/neuralnetworks/aidl/utils/Android.bp
@@ -26,7 +26,14 @@
 cc_defaults {
     name: "neuralnetworks_utils_hal_aidl_defaults",
     defaults: ["neuralnetworks_utils_defaults"],
-    srcs: ["src/*"],
+    srcs: [
+        // AIDL utils that a driver may depend on.
+        "src/BufferTracker.cpp",
+        "src/Conversions.cpp",
+        "src/HalUtils.cpp",
+        "src/Utils.cpp",
+        "src/ValidateHal.cpp",
+    ],
     local_include_dirs: ["include/nnapi/hal/aidl/"],
     export_include_dirs: ["include"],
     cflags: ["-Wthread-safety"],
@@ -47,6 +54,7 @@
     },
 }
 
+// Deprecated. Remove once all modules depending on this are migrated away.
 cc_library_static {
     name: "neuralnetworks_utils_hal_aidl_v1",
     defaults: ["neuralnetworks_utils_hal_aidl_defaults"],
@@ -56,19 +64,25 @@
 }
 
 cc_library_static {
-    name: "neuralnetworks_utils_hal_aidl_v2",
-    defaults: ["neuralnetworks_utils_hal_aidl_defaults"],
-    shared_libs: [
-        "android.hardware.neuralnetworks-V2-ndk",
-    ],
-}
-
-cc_library_static {
     name: "neuralnetworks_utils_hal_aidl",
     defaults: ["neuralnetworks_utils_hal_aidl_defaults"],
-    shared_libs: [
-        "android.hardware.neuralnetworks-V3-ndk",
+    srcs: [
+        // Additional AIDL utils for the runtime.
+        "src/Assertions.cpp",
+        "src/Buffer.cpp",
+        "src/Burst.cpp",
+        "src/Callbacks.cpp",
+        "src/Device.cpp",
+        "src/Execution.cpp",
+        "src/InvalidDevice.cpp",
+        "src/PreparedModel.cpp",
+        "src/ProtectCallback.cpp",
+        "src/Service.cpp",
     ],
+    shared_libs: [
+        "android.hardware.neuralnetworks-V4-ndk",
+    ],
+    cflags: ["-DNN_AIDL_V4_OR_ABOVE"],
 }
 
 // A cc_defaults that includes the latest non-experimental AIDL utilities and other AIDL libraries
@@ -79,9 +93,10 @@
     static_libs: [
         "android.hardware.common-V2-ndk",
         "android.hardware.graphics.common-V2-ndk",
-        "android.hardware.neuralnetworks-V3-ndk",
+        "android.hardware.neuralnetworks-V4-ndk",
         "neuralnetworks_utils_hal_aidl",
     ],
+    cflags: ["-DNN_AIDL_V4_OR_ABOVE"],
 }
 
 cc_test {
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Callbacks.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Callbacks.h
index 168264b..960be2b 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Callbacks.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Callbacks.h
@@ -36,6 +36,8 @@
   public:
     using Data = nn::GeneralResult<nn::SharedPreparedModel>;
 
+    PreparedModelCallback(nn::Version featureLevel) : kFeatureLevel(featureLevel) {}
+
     ndk::ScopedAStatus notify(ErrorStatus status,
                               const std::shared_ptr<IPreparedModel>& preparedModel) override;
 
@@ -44,6 +46,7 @@
     Data get();
 
   private:
+    const nn::Version kFeatureLevel;
     hal::utils::TransferValue<Data> mData;
 };
 
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Execution.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Execution.h
index a77ea98..14802b9 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Execution.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Execution.h
@@ -17,6 +17,8 @@
 #ifndef ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_AIDL_UTILS_EXECUTION_H
 #define ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_AIDL_UTILS_EXECUTION_H
 
+#include <aidl/android/hardware/neuralnetworks/IExecution.h>
+
 #include <nnapi/IExecution.h>
 #include <nnapi/Result.h>
 #include <nnapi/Types.h>
@@ -33,17 +35,22 @@
 
 namespace aidl::android::hardware::neuralnetworks::utils {
 
-class Execution final : public nn::IExecution, public std::enable_shared_from_this<Execution> {
+// A reusable execution implementation with a cached Request, internally it is still passing the
+// request to the driver in every computation.
+class ExecutionWithCachedRequest final
+    : public nn::IExecution,
+      public std::enable_shared_from_this<ExecutionWithCachedRequest> {
     struct PrivateConstructorTag {};
 
   public:
-    static nn::GeneralResult<std::shared_ptr<const Execution>> create(
+    static nn::GeneralResult<std::shared_ptr<const ExecutionWithCachedRequest>> create(
             std::shared_ptr<const PreparedModel> preparedModel, Request request,
             hal::utils::RequestRelocation relocation, bool measure, int64_t loopTimeoutDuration);
 
-    Execution(PrivateConstructorTag tag, std::shared_ptr<const PreparedModel> preparedModel,
-              Request request, hal::utils::RequestRelocation relocation, bool measure,
-              int64_t loopTimeoutDuration);
+    ExecutionWithCachedRequest(PrivateConstructorTag tag,
+                               std::shared_ptr<const PreparedModel> preparedModel, Request request,
+                               hal::utils::RequestRelocation relocation, bool measure,
+                               int64_t loopTimeoutDuration);
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> compute(
             const nn::OptionalTimePoint& deadline) const override;
@@ -60,6 +67,30 @@
     const int64_t kLoopTimeoutDuration;
 };
 
+// A reusable execution implementation that is backed by an actual AIDL IExecution object.
+class Execution final : public nn::IExecution, public std::enable_shared_from_this<Execution> {
+    struct PrivateConstructorTag {};
+
+  public:
+    static nn::GeneralResult<std::shared_ptr<const Execution>> create(
+            std::shared_ptr<aidl_hal::IExecution> execution,
+            hal::utils::RequestRelocation relocation);
+
+    Execution(PrivateConstructorTag tag, std::shared_ptr<aidl_hal::IExecution> execution,
+              hal::utils::RequestRelocation relocation);
+
+    nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> compute(
+            const nn::OptionalTimePoint& deadline) const override;
+
+    nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> computeFenced(
+            const std::vector<nn::SyncFence>& waitFor, const nn::OptionalTimePoint& deadline,
+            const nn::OptionalDuration& timeoutDurationAfterFence) const override;
+
+  private:
+    const std::shared_ptr<aidl_hal::IExecution> kExecution;
+    const hal::utils::RequestRelocation kRelocation;
+};
+
 }  // namespace aidl::android::hardware::neuralnetworks::utils
 
 #endif  // ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_AIDL_UTILS_EXECUTION_H
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/HalInterfaces.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/HalInterfaces.h
index 3fb443c..205d428 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/HalInterfaces.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/HalInterfaces.h
@@ -61,6 +61,11 @@
 #include <aidl/android/hardware/neuralnetworks/SymmPerChannelQuantParams.h>
 #include <aidl/android/hardware/neuralnetworks/Timing.h>
 
+#ifdef NN_AIDL_V4_OR_ABOVE
+#include <aidl/android/hardware/neuralnetworks/BnExecution.h>
+#include <aidl/android/hardware/neuralnetworks/IExecution.h>
+#endif  // NN_AIDL_V4_OR_ABOVE
+
 namespace android::nn {
 
 namespace aidl_hal = ::aidl::android::hardware::neuralnetworks;
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/PreparedModel.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/PreparedModel.h
index 4035764..24cd681 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/PreparedModel.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/PreparedModel.h
@@ -41,10 +41,11 @@
 
   public:
     static nn::GeneralResult<std::shared_ptr<const PreparedModel>> create(
-            std::shared_ptr<aidl_hal::IPreparedModel> preparedModel);
+            std::shared_ptr<aidl_hal::IPreparedModel> preparedModel, nn::Version featureLevel);
 
     PreparedModel(PrivateConstructorTag tag,
-                  std::shared_ptr<aidl_hal::IPreparedModel> preparedModel);
+                  std::shared_ptr<aidl_hal::IPreparedModel> preparedModel,
+                  nn::Version featureLevel);
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
@@ -78,6 +79,7 @@
 
   private:
     const std::shared_ptr<aidl_hal::IPreparedModel> kPreparedModel;
+    const nn::Version kFeatureLevel;
 };
 
 }  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Utils.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Utils.h
index a27487e..beca38b 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Utils.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Utils.h
@@ -38,6 +38,8 @@
             return nn::kVersionFeatureLevel6;
         case 3:
             return nn::kVersionFeatureLevel7;
+        case 4:
+            return nn::kVersionFeatureLevel8;
         default:
             return std::nullopt;
     }
diff --git a/neuralnetworks/aidl/utils/src/Callbacks.cpp b/neuralnetworks/aidl/utils/src/Callbacks.cpp
index 8084970..554f3fa 100644
--- a/neuralnetworks/aidl/utils/src/Callbacks.cpp
+++ b/neuralnetworks/aidl/utils/src/Callbacks.cpp
@@ -38,16 +38,17 @@
 // nn::kVersionFeatureLevel5. On failure, this function returns with the appropriate
 // nn::GeneralError.
 nn::GeneralResult<nn::SharedPreparedModel> prepareModelCallback(
-        ErrorStatus status, const std::shared_ptr<IPreparedModel>& preparedModel) {
+        ErrorStatus status, const std::shared_ptr<IPreparedModel>& preparedModel,
+        nn::Version featureLevel) {
     HANDLE_STATUS_AIDL(status) << "model preparation failed with " << toString(status);
-    return NN_TRY(PreparedModel::create(preparedModel));
+    return NN_TRY(PreparedModel::create(preparedModel, featureLevel));
 }
 
 }  // namespace
 
 ndk::ScopedAStatus PreparedModelCallback::notify(
         ErrorStatus status, const std::shared_ptr<IPreparedModel>& preparedModel) {
-    mData.put(prepareModelCallback(status, preparedModel));
+    mData.put(prepareModelCallback(status, preparedModel, kFeatureLevel));
     return ndk::ScopedAStatus::ok();
 }
 
diff --git a/neuralnetworks/aidl/utils/src/Device.cpp b/neuralnetworks/aidl/utils/src/Device.cpp
index 5b7ec4e..bad10ed 100644
--- a/neuralnetworks/aidl/utils/src/Device.cpp
+++ b/neuralnetworks/aidl/utils/src/Device.cpp
@@ -229,7 +229,7 @@
     const auto aidlDataCache = NN_TRY(convert(dataCache));
     const auto aidlToken = NN_TRY(convert(token));
 
-    const auto cb = ndk::SharedRefBase::make<PreparedModelCallback>();
+    const auto cb = ndk::SharedRefBase::make<PreparedModelCallback>(kFeatureLevel);
     const auto scoped = kDeathHandler.protectCallback(cb.get());
 
     const auto ret = kDevice->prepareModel(aidlModel, aidlPreference, aidlPriority, aidlDeadline,
@@ -247,7 +247,7 @@
     const auto aidlDataCache = NN_TRY(convert(dataCache));
     const auto aidlToken = NN_TRY(convert(token));
 
-    const auto cb = ndk::SharedRefBase::make<PreparedModelCallback>();
+    const auto cb = ndk::SharedRefBase::make<PreparedModelCallback>(kFeatureLevel);
     const auto scoped = kDeathHandler.protectCallback(cb.get());
 
     const auto ret = kDevice->prepareModelFromCache(aidlDeadline, aidlModelCache, aidlDataCache,
diff --git a/neuralnetworks/aidl/utils/src/Execution.cpp b/neuralnetworks/aidl/utils/src/Execution.cpp
index 94edd90..c4add63 100644
--- a/neuralnetworks/aidl/utils/src/Execution.cpp
+++ b/neuralnetworks/aidl/utils/src/Execution.cpp
@@ -35,36 +35,39 @@
 
 namespace aidl::android::hardware::neuralnetworks::utils {
 
-nn::GeneralResult<std::shared_ptr<const Execution>> Execution::create(
-        std::shared_ptr<const PreparedModel> preparedModel, Request request,
-        hal::utils::RequestRelocation relocation, bool measure, int64_t loopTimeoutDuration) {
+nn::GeneralResult<std::shared_ptr<const ExecutionWithCachedRequest>>
+ExecutionWithCachedRequest::create(std::shared_ptr<const PreparedModel> preparedModel,
+                                   Request request, hal::utils::RequestRelocation relocation,
+                                   bool measure, int64_t loopTimeoutDuration) {
     if (preparedModel == nullptr) {
-        return NN_ERROR() << "aidl::utils::Execution::create must have non-null preparedModel";
+        return NN_ERROR() << "aidl::utils::ExecutionWithCachedRequest::create must have non-null "
+                             "preparedModel";
     }
 
-    return std::make_shared<const Execution>(PrivateConstructorTag{}, std::move(preparedModel),
-                                             std::move(request), std::move(relocation), measure,
-                                             loopTimeoutDuration);
+    return std::make_shared<const ExecutionWithCachedRequest>(
+            PrivateConstructorTag{}, std::move(preparedModel), std::move(request),
+            std::move(relocation), measure, loopTimeoutDuration);
 }
 
-Execution::Execution(PrivateConstructorTag /*tag*/,
-                     std::shared_ptr<const PreparedModel> preparedModel, Request request,
-                     hal::utils::RequestRelocation relocation, bool measure,
-                     int64_t loopTimeoutDuration)
+ExecutionWithCachedRequest::ExecutionWithCachedRequest(
+        PrivateConstructorTag /*tag*/, std::shared_ptr<const PreparedModel> preparedModel,
+        Request request, hal::utils::RequestRelocation relocation, bool measure,
+        int64_t loopTimeoutDuration)
     : kPreparedModel(std::move(preparedModel)),
       kRequest(std::move(request)),
       kRelocation(std::move(relocation)),
       kMeasure(measure),
       kLoopTimeoutDuration(loopTimeoutDuration) {}
 
-nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Execution::compute(
-        const nn::OptionalTimePoint& deadline) const {
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>
+ExecutionWithCachedRequest::compute(const nn::OptionalTimePoint& deadline) const {
     const auto aidlDeadline = NN_TRY(convert(deadline));
     return kPreparedModel->executeInternal(kRequest, kMeasure, aidlDeadline, kLoopTimeoutDuration,
                                            kRelocation);
 }
 
-nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> Execution::computeFenced(
+nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
+ExecutionWithCachedRequest::computeFenced(
         const std::vector<nn::SyncFence>& waitFor, const nn::OptionalTimePoint& deadline,
         const nn::OptionalDuration& timeoutDurationAfterFence) const {
     const auto aidlWaitFor = NN_TRY(convert(waitFor));
@@ -75,4 +78,18 @@
                                                  aidlTimeoutDurationAfterFence, kRelocation);
 }
 
+nn::GeneralResult<std::shared_ptr<const Execution>> Execution::create(
+        std::shared_ptr<aidl_hal::IExecution> execution, hal::utils::RequestRelocation relocation) {
+    if (execution == nullptr) {
+        return NN_ERROR() << "aidl::utils::Execution::create must have non-null execution";
+    }
+
+    return std::make_shared<const Execution>(PrivateConstructorTag{}, std::move(execution),
+                                             std::move(relocation));
+}
+
+Execution::Execution(PrivateConstructorTag /*tag*/, std::shared_ptr<aidl_hal::IExecution> execution,
+                     hal::utils::RequestRelocation relocation)
+    : kExecution(std::move(execution)), kRelocation(std::move(relocation)) {}
+
 }  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/utils/src/PreparedModel.cpp b/neuralnetworks/aidl/utils/src/PreparedModel.cpp
index f25c2c8..6d1de56 100644
--- a/neuralnetworks/aidl/utils/src/PreparedModel.cpp
+++ b/neuralnetworks/aidl/utils/src/PreparedModel.cpp
@@ -54,21 +54,77 @@
     return std::make_pair(NN_TRY(nn::convert(timingLaunched)), NN_TRY(nn::convert(timingFenced)));
 }
 
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> handleExecutionResult(
+        const ExecutionResult& result, const hal::utils::RequestRelocation& relocation) {
+    if (!result.outputSufficientSize) {
+        auto canonicalOutputShapes =
+                nn::convert(result.outputShapes).value_or(std::vector<nn::OutputShape>{});
+        return NN_ERROR(nn::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE, std::move(canonicalOutputShapes))
+               << "execution failed with " << nn::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE;
+    }
+    auto [outputShapes, timing] =
+            NN_TRY(convertExecutionResults(result.outputShapes, result.timing));
+
+    if (relocation.output) {
+        relocation.output->flush();
+    }
+    return std::make_pair(std::move(outputShapes), timing);
+}
+
+nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
+handleFencedExecutionResult(const FencedExecutionResult& result,
+                            const hal::utils::RequestRelocation& relocation) {
+    auto resultSyncFence = nn::SyncFence::createAsSignaled();
+    if (result.syncFence.get() != -1) {
+        resultSyncFence = nn::SyncFence::create(NN_TRY(nn::convert(result.syncFence))).value();
+    }
+
+    auto callback = result.callback;
+    if (callback == nullptr) {
+        return NN_ERROR(nn::ErrorStatus::GENERAL_FAILURE) << "callback is null";
+    }
+
+    // If computeFenced required the request memory to be moved into shared memory, block here until
+    // the fenced execution has completed and flush the memory back.
+    if (relocation.output) {
+        const auto state = resultSyncFence.syncWait({});
+        if (state != nn::SyncFence::FenceState::SIGNALED) {
+            return NN_ERROR() << "syncWait failed with " << state;
+        }
+        relocation.output->flush();
+    }
+
+    // Create callback which can be used to retrieve the execution error status and timings.
+    nn::ExecuteFencedInfoCallback resultCallback =
+            [callback]() -> nn::GeneralResult<std::pair<nn::Timing, nn::Timing>> {
+        ErrorStatus errorStatus;
+        Timing timingLaunched;
+        Timing timingFenced;
+        const auto ret = callback->getExecutionInfo(&timingLaunched, &timingFenced, &errorStatus);
+        HANDLE_ASTATUS(ret) << "fenced execution callback getExecutionInfo failed";
+        return convertFencedExecutionResults(errorStatus, timingLaunched, timingFenced);
+    };
+
+    return std::make_pair(std::move(resultSyncFence), std::move(resultCallback));
+}
+
 }  // namespace
 
 nn::GeneralResult<std::shared_ptr<const PreparedModel>> PreparedModel::create(
-        std::shared_ptr<aidl_hal::IPreparedModel> preparedModel) {
+        std::shared_ptr<aidl_hal::IPreparedModel> preparedModel, nn::Version featureLevel) {
     if (preparedModel == nullptr) {
         return NN_ERROR()
                << "aidl_hal::utils::PreparedModel::create must have non-null preparedModel";
     }
 
-    return std::make_shared<const PreparedModel>(PrivateConstructorTag{}, std::move(preparedModel));
+    return std::make_shared<const PreparedModel>(PrivateConstructorTag{}, std::move(preparedModel),
+                                                 featureLevel);
 }
 
 PreparedModel::PreparedModel(PrivateConstructorTag /*tag*/,
-                             std::shared_ptr<aidl_hal::IPreparedModel> preparedModel)
-    : kPreparedModel(std::move(preparedModel)) {}
+                             std::shared_ptr<aidl_hal::IPreparedModel> preparedModel,
+                             nn::Version featureLevel)
+    : kPreparedModel(std::move(preparedModel)), kFeatureLevel(featureLevel) {}
 
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> PreparedModel::execute(
         const nn::Request& request, nn::MeasureTiming measure,
@@ -101,19 +157,7 @@
     const auto ret = kPreparedModel->executeSynchronously(request, measure, deadline,
                                                           loopTimeoutDuration, &executionResult);
     HANDLE_ASTATUS(ret) << "executeSynchronously failed";
-    if (!executionResult.outputSufficientSize) {
-        auto canonicalOutputShapes =
-                nn::convert(executionResult.outputShapes).value_or(std::vector<nn::OutputShape>{});
-        return NN_ERROR(nn::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE, std::move(canonicalOutputShapes))
-               << "execution failed with " << nn::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE;
-    }
-    auto [outputShapes, timing] =
-            NN_TRY(convertExecutionResults(executionResult.outputShapes, executionResult.timing));
-
-    if (relocation.output) {
-        relocation.output->flush();
-    }
-    return std::make_pair(std::move(outputShapes), timing);
+    return handleExecutionResult(executionResult, relocation);
 }
 
 nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
@@ -154,39 +198,7 @@
             kPreparedModel->executeFenced(request, waitFor, measure, deadline, loopTimeoutDuration,
                                           timeoutDurationAfterFence, &result);
     HANDLE_ASTATUS(ret) << "executeFenced failed";
-
-    auto resultSyncFence = nn::SyncFence::createAsSignaled();
-    if (result.syncFence.get() != -1) {
-        resultSyncFence = nn::SyncFence::create(NN_TRY(nn::convert(result.syncFence))).value();
-    }
-
-    auto callback = result.callback;
-    if (callback == nullptr) {
-        return NN_ERROR(nn::ErrorStatus::GENERAL_FAILURE) << "callback is null";
-    }
-
-    // If executeFenced required the request memory to be moved into shared memory, block here until
-    // the fenced execution has completed and flush the memory back.
-    if (relocation.output) {
-        const auto state = resultSyncFence.syncWait({});
-        if (state != nn::SyncFence::FenceState::SIGNALED) {
-            return NN_ERROR() << "syncWait failed with " << state;
-        }
-        relocation.output->flush();
-    }
-
-    // Create callback which can be used to retrieve the execution error status and timings.
-    nn::ExecuteFencedInfoCallback resultCallback =
-            [callback]() -> nn::GeneralResult<std::pair<nn::Timing, nn::Timing>> {
-        ErrorStatus errorStatus;
-        Timing timingLaunched;
-        Timing timingFenced;
-        const auto ret = callback->getExecutionInfo(&timingLaunched, &timingFenced, &errorStatus);
-        HANDLE_ASTATUS(ret) << "fenced execution callback getExecutionInfo failed";
-        return convertFencedExecutionResults(errorStatus, timingLaunched, timingFenced);
-    };
-
-    return std::make_pair(std::move(resultSyncFence), std::move(resultCallback));
+    return handleFencedExecutionResult(result, relocation);
 }
 
 nn::GeneralResult<nn::SharedExecution> PreparedModel::createReusableExecution(
@@ -202,8 +214,18 @@
     auto aidlRequest = NN_TRY(convert(requestInShared));
     auto aidlMeasure = NN_TRY(convert(measure));
     auto aidlLoopTimeoutDuration = NN_TRY(convert(loopTimeoutDuration));
-    return Execution::create(shared_from_this(), std::move(aidlRequest), std::move(relocation),
-                             aidlMeasure, aidlLoopTimeoutDuration);
+
+    if (kFeatureLevel.level >= nn::Version::Level::FEATURE_LEVEL_8) {
+        std::shared_ptr<IExecution> execution;
+        const auto ret = kPreparedModel->createReusableExecution(
+                aidlRequest, aidlMeasure, aidlLoopTimeoutDuration, &execution);
+        HANDLE_ASTATUS(ret) << "createReusableExecution failed";
+        return Execution::create(std::move(execution), std::move(relocation));
+    }
+
+    return ExecutionWithCachedRequest::create(shared_from_this(), std::move(aidlRequest),
+                                              std::move(relocation), aidlMeasure,
+                                              aidlLoopTimeoutDuration);
 }
 
 nn::GeneralResult<nn::SharedBurst> PreparedModel::configureExecutionBurst() const {
@@ -218,4 +240,36 @@
     return resource;
 }
 
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Execution::compute(
+        const nn::OptionalTimePoint& deadline) const {
+    const auto aidlDeadline = NN_TRY(convert(deadline));
+
+    if (kRelocation.input) {
+        kRelocation.input->flush();
+    }
+
+    ExecutionResult executionResult;
+    auto ret = kExecution->executeSynchronously(aidlDeadline, &executionResult);
+    HANDLE_ASTATUS(ret) << "executeSynchronously failed";
+    return handleExecutionResult(executionResult, kRelocation);
+}
+
+nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> Execution::computeFenced(
+        const std::vector<nn::SyncFence>& waitFor, const nn::OptionalTimePoint& deadline,
+        const nn::OptionalDuration& timeoutDurationAfterFence) const {
+    const auto aidlWaitFor = NN_TRY(convert(waitFor));
+    const auto aidlDeadline = NN_TRY(convert(deadline));
+    const auto aidlTimeoutDurationAfterFence = NN_TRY(convert(timeoutDurationAfterFence));
+
+    if (kRelocation.input) {
+        kRelocation.input->flush();
+    }
+
+    FencedExecutionResult result;
+    const auto ret = kExecution->executeFenced(aidlWaitFor, aidlDeadline,
+                                               aidlTimeoutDurationAfterFence, &result);
+    HANDLE_ASTATUS(ret) << "executeFenced failed";
+    return handleFencedExecutionResult(result, kRelocation);
+}
+
 }  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/utils/test/DeviceTest.cpp b/neuralnetworks/aidl/utils/test/DeviceTest.cpp
index 0366e7d..fb13af8 100644
--- a/neuralnetworks/aidl/utils/test/DeviceTest.cpp
+++ b/neuralnetworks/aidl/utils/test/DeviceTest.cpp
@@ -17,6 +17,7 @@
 #include "MockBuffer.h"
 #include "MockDevice.h"
 #include "MockPreparedModel.h"
+#include "TestUtils.h"
 
 #include <aidl/android/hardware/neuralnetworks/BnDevice.h>
 #include <android/binder_auto_utils.h>
@@ -146,26 +147,7 @@
     return ndk::ScopedAStatus::fromStatus(STATUS_DEAD_OBJECT);
 };
 
-class DeviceTest : public ::testing::TestWithParam<nn::Version> {
-  protected:
-    const nn::Version kVersion = GetParam();
-};
-
-std::string printDeviceTest(const testing::TestParamInfo<nn::Version>& info) {
-    const nn::Version version = info.param;
-    CHECK(!version.runtimeOnlyFeatures);
-    switch (version.level) {
-        case nn::Version::Level::FEATURE_LEVEL_5:
-            return "v1";
-        case nn::Version::Level::FEATURE_LEVEL_6:
-            return "v2";
-        case nn::Version::Level::FEATURE_LEVEL_7:
-            return "v3";
-        default:
-            LOG(FATAL) << "Invalid AIDL version: " << version;
-            return "invalid";
-    }
-}
+class DeviceTest : public VersionedAidlUtilsTestBase {};
 
 }  // namespace
 
@@ -894,9 +876,6 @@
     EXPECT_EQ(result.error().code, nn::ErrorStatus::DEAD_OBJECT);
 }
 
-INSTANTIATE_TEST_SUITE_P(TestDevice, DeviceTest,
-                         ::testing::Values(nn::kVersionFeatureLevel5, nn::kVersionFeatureLevel6,
-                                           nn::kVersionFeatureLevel7),
-                         printDeviceTest);
+INSTANTIATE_VERSIONED_AIDL_UTILS_TEST(DeviceTest, kAllAidlVersions);
 
 }  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/utils/test/ExecutionTest.cpp b/neuralnetworks/aidl/utils/test/ExecutionTest.cpp
new file mode 100644
index 0000000..8519290
--- /dev/null
+++ b/neuralnetworks/aidl/utils/test/ExecutionTest.cpp
@@ -0,0 +1,254 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "MockExecution.h"
+#include "MockFencedExecutionCallback.h"
+
+#include <aidl/android/hardware/neuralnetworks/IFencedExecutionCallback.h>
+#include <gmock/gmock.h>
+#include <gtest/gtest.h>
+#include <nnapi/IExecution.h>
+#include <nnapi/TypeUtils.h>
+#include <nnapi/Types.h>
+#include <nnapi/hal/aidl/Execution.h>
+
+#include <functional>
+#include <memory>
+
+namespace aidl::android::hardware::neuralnetworks::utils {
+namespace {
+
+using ::testing::_;
+using ::testing::DoAll;
+using ::testing::Invoke;
+using ::testing::InvokeWithoutArgs;
+using ::testing::SetArgPointee;
+
+const std::shared_ptr<IExecution> kInvalidExecution;
+constexpr auto kNoTiming = Timing{.timeOnDeviceNs = -1, .timeInDriverNs = -1};
+
+constexpr auto makeStatusOk = [] { return ndk::ScopedAStatus::ok(); };
+
+constexpr auto makeGeneralFailure = [] {
+    return ndk::ScopedAStatus::fromServiceSpecificError(
+            static_cast<int32_t>(ErrorStatus::GENERAL_FAILURE));
+};
+constexpr auto makeGeneralTransportFailure = [] {
+    return ndk::ScopedAStatus::fromStatus(STATUS_NO_MEMORY);
+};
+constexpr auto makeDeadObjectFailure = [] {
+    return ndk::ScopedAStatus::fromStatus(STATUS_DEAD_OBJECT);
+};
+
+auto makeFencedExecutionResult(const std::shared_ptr<MockFencedExecutionCallback>& callback) {
+    return [callback](const std::vector<ndk::ScopedFileDescriptor>& /*waitFor*/,
+                      int64_t /*deadline*/, int64_t /*duration*/,
+                      FencedExecutionResult* fencedExecutionResult) {
+        *fencedExecutionResult = FencedExecutionResult{.callback = callback,
+                                                       .syncFence = ndk::ScopedFileDescriptor(-1)};
+        return ndk::ScopedAStatus::ok();
+    };
+}
+
+}  // namespace
+
+TEST(ExecutionTest, invalidExecution) {
+    // run test
+    const auto result = Execution::create(kInvalidExecution, {});
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST(ExecutionTest, executeSync) {
+    // setup call
+    const auto mockExecution = MockExecution::create();
+    const auto execution = Execution::create(mockExecution, {}).value();
+    const auto mockExecutionResult = ExecutionResult{
+            .outputSufficientSize = true,
+            .outputShapes = {},
+            .timing = kNoTiming,
+    };
+    EXPECT_CALL(*mockExecution, executeSynchronously(_, _))
+            .Times(1)
+            .WillOnce(
+                    DoAll(SetArgPointee<1>(mockExecutionResult), InvokeWithoutArgs(makeStatusOk)));
+
+    // run test
+    const auto result = execution->compute({});
+
+    // verify result
+    EXPECT_TRUE(result.has_value())
+            << "Failed with " << result.error().code << ": " << result.error().message;
+}
+
+TEST(ExecutionTest, executeSyncError) {
+    // setup test
+    const auto mockExecution = MockExecution::create();
+    const auto execution = Execution::create(mockExecution, {}).value();
+    EXPECT_CALL(*mockExecution, executeSynchronously(_, _))
+            .Times(1)
+            .WillOnce(Invoke(makeGeneralFailure));
+
+    // run test
+    const auto result = execution->compute({});
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST(ExecutionTest, executeSyncTransportFailure) {
+    // setup test
+    const auto mockExecution = MockExecution::create();
+    const auto execution = Execution::create(mockExecution, {}).value();
+    EXPECT_CALL(*mockExecution, executeSynchronously(_, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
+
+    // run test
+    const auto result = execution->compute({});
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST(ExecutionTest, executeSyncDeadObject) {
+    // setup test
+    const auto mockExecution = MockExecution::create();
+    const auto execution = Execution::create(mockExecution, {}).value();
+    EXPECT_CALL(*mockExecution, executeSynchronously(_, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
+
+    // run test
+    const auto result = execution->compute({});
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::DEAD_OBJECT);
+}
+
+TEST(ExecutionTest, executeFenced) {
+    // setup call
+    const auto mockExecution = MockExecution::create();
+    const auto execution = Execution::create(mockExecution, {}).value();
+    const auto mockCallback = MockFencedExecutionCallback::create();
+    EXPECT_CALL(*mockCallback, getExecutionInfo(_, _, _))
+            .Times(1)
+            .WillOnce(DoAll(SetArgPointee<0>(kNoTiming), SetArgPointee<1>(kNoTiming),
+                            SetArgPointee<2>(ErrorStatus::NONE), Invoke(makeStatusOk)));
+    EXPECT_CALL(*mockExecution, executeFenced(_, _, _, _))
+            .Times(1)
+            .WillOnce(Invoke(makeFencedExecutionResult(mockCallback)));
+
+    // run test
+    const auto result = execution->computeFenced({}, {}, {});
+
+    // verify result
+    ASSERT_TRUE(result.has_value())
+            << "Failed with " << result.error().code << ": " << result.error().message;
+    const auto& [syncFence, callback] = result.value();
+    EXPECT_EQ(syncFence.syncWait({}), nn::SyncFence::FenceState::SIGNALED);
+    ASSERT_NE(callback, nullptr);
+
+    // get results from callback
+    const auto callbackResult = callback();
+    ASSERT_TRUE(callbackResult.has_value()) << "Failed with " << callbackResult.error().code << ": "
+                                            << callbackResult.error().message;
+}
+
+TEST(ExecutionTest, executeFencedCallbackError) {
+    // setup call
+    const auto mockExecution = MockExecution::create();
+    const auto execution = Execution::create(mockExecution, {}).value();
+    const auto mockCallback = MockFencedExecutionCallback::create();
+    EXPECT_CALL(*mockCallback, getExecutionInfo(_, _, _))
+            .Times(1)
+            .WillOnce(Invoke(DoAll(SetArgPointee<0>(kNoTiming), SetArgPointee<1>(kNoTiming),
+                                   SetArgPointee<2>(ErrorStatus::GENERAL_FAILURE),
+                                   Invoke(makeStatusOk))));
+    EXPECT_CALL(*mockExecution, executeFenced(_, _, _, _))
+            .Times(1)
+            .WillOnce(Invoke(makeFencedExecutionResult(mockCallback)));
+
+    // run test
+    const auto result = execution->computeFenced({}, {}, {});
+
+    // verify result
+    ASSERT_TRUE(result.has_value())
+            << "Failed with " << result.error().code << ": " << result.error().message;
+    const auto& [syncFence, callback] = result.value();
+    EXPECT_NE(syncFence.syncWait({}), nn::SyncFence::FenceState::ACTIVE);
+    ASSERT_NE(callback, nullptr);
+
+    // verify callback failure
+    const auto callbackResult = callback();
+    ASSERT_FALSE(callbackResult.has_value());
+    EXPECT_EQ(callbackResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST(ExecutionTest, executeFencedError) {
+    // setup test
+    const auto mockExecution = MockExecution::create();
+    const auto execution = Execution::create(mockExecution, {}).value();
+    EXPECT_CALL(*mockExecution, executeFenced(_, _, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeGeneralFailure));
+
+    // run test
+    const auto result = execution->computeFenced({}, {}, {});
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST(ExecutionTest, executeFencedTransportFailure) {
+    // setup test
+    const auto mockExecution = MockExecution::create();
+    const auto execution = Execution::create(mockExecution, {}).value();
+    EXPECT_CALL(*mockExecution, executeFenced(_, _, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
+
+    // run test
+    const auto result = execution->computeFenced({}, {}, {});
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST(ExecutionTest, executeFencedDeadObject) {
+    // setup test
+    const auto mockExecution = MockExecution::create();
+    const auto execution = Execution::create(mockExecution, {}).value();
+    EXPECT_CALL(*mockExecution, executeFenced(_, _, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
+
+    // run test
+    const auto result = execution->computeFenced({}, {}, {});
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::DEAD_OBJECT);
+}
+
+}  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/utils/test/MockExecution.h b/neuralnetworks/aidl/utils/test/MockExecution.h
new file mode 100644
index 0000000..216f569
--- /dev/null
+++ b/neuralnetworks/aidl/utils/test/MockExecution.h
@@ -0,0 +1,47 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_AIDL_UTILS_TEST_MOCK_EXECUTION_H
+#define ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_AIDL_UTILS_TEST_MOCK_EXECUTION_H
+
+#include <aidl/android/hardware/neuralnetworks/BnExecution.h>
+#include <android/binder_interface_utils.h>
+#include <gmock/gmock.h>
+#include <gtest/gtest.h>
+#include <hidl/HidlSupport.h>
+#include <hidl/Status.h>
+
+namespace aidl::android::hardware::neuralnetworks::utils {
+
+class MockExecution final : public BnExecution {
+  public:
+    static std::shared_ptr<MockExecution> create();
+
+    MOCK_METHOD(ndk::ScopedAStatus, executeSynchronously,
+                (int64_t deadline, ExecutionResult* executionResult), (override));
+    MOCK_METHOD(ndk::ScopedAStatus, executeFenced,
+                (const std::vector<ndk::ScopedFileDescriptor>& waitFor, int64_t deadline,
+                 int64_t duration, FencedExecutionResult* fencedExecutionResult),
+                (override));
+};
+
+inline std::shared_ptr<MockExecution> MockExecution::create() {
+    return ndk::SharedRefBase::make<MockExecution>();
+}
+
+}  // namespace aidl::android::hardware::neuralnetworks::utils
+
+#endif  // ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_AIDL_UTILS_TEST_MOCK_EXECUTION_H
diff --git a/neuralnetworks/aidl/utils/test/MockPreparedModel.h b/neuralnetworks/aidl/utils/test/MockPreparedModel.h
index a4ae2b7..0ed9af9 100644
--- a/neuralnetworks/aidl/utils/test/MockPreparedModel.h
+++ b/neuralnetworks/aidl/utils/test/MockPreparedModel.h
@@ -17,6 +17,7 @@
 #ifndef ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_AIDL_UTILS_TEST_MOCK_PREPARED_MODEL_H
 #define ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_AIDL_UTILS_TEST_MOCK_PREPARED_MODEL_H
 
+#include <aidl/android/hardware/neuralnetworks/BnExecution.h>
 #include <aidl/android/hardware/neuralnetworks/BnPreparedModel.h>
 #include <android/binder_interface_utils.h>
 #include <gmock/gmock.h>
@@ -41,6 +42,10 @@
                 (override));
     MOCK_METHOD(ndk::ScopedAStatus, configureExecutionBurst, (std::shared_ptr<IBurst> * burst),
                 (override));
+    MOCK_METHOD(ndk::ScopedAStatus, createReusableExecution,
+                (const Request& request, bool measureTiming, int64_t loopTimeoutDuration,
+                 std::shared_ptr<IExecution>* execution),
+                (override));
 };
 
 inline std::shared_ptr<MockPreparedModel> MockPreparedModel::create() {
diff --git a/neuralnetworks/aidl/utils/test/PreparedModelTest.cpp b/neuralnetworks/aidl/utils/test/PreparedModelTest.cpp
index 8bb5c90..8cfb7c1 100644
--- a/neuralnetworks/aidl/utils/test/PreparedModelTest.cpp
+++ b/neuralnetworks/aidl/utils/test/PreparedModelTest.cpp
@@ -15,8 +15,10 @@
  */
 
 #include "MockBurst.h"
+#include "MockExecution.h"
 #include "MockFencedExecutionCallback.h"
 #include "MockPreparedModel.h"
+#include "TestUtils.h"
 
 #include <aidl/android/hardware/neuralnetworks/IFencedExecutionCallback.h>
 #include <gmock/gmock.h>
@@ -66,21 +68,23 @@
     };
 }
 
+class PreparedModelTest : public VersionedAidlUtilsTestBase {};
+
 }  // namespace
 
-TEST(PreparedModelTest, invalidPreparedModel) {
+TEST_P(PreparedModelTest, invalidPreparedModel) {
     // run test
-    const auto result = PreparedModel::create(kInvalidPreparedModel);
+    const auto result = PreparedModel::create(kInvalidPreparedModel, kVersion);
 
     // verify result
     ASSERT_FALSE(result.has_value());
     EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, executeSync) {
+TEST_P(PreparedModelTest, executeSync) {
     // setup call
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     const auto mockExecutionResult = ExecutionResult{
             .outputSufficientSize = true,
             .outputShapes = {},
@@ -99,10 +103,10 @@
             << "Failed with " << result.error().code << ": " << result.error().message;
 }
 
-TEST(PreparedModelTest, executeSyncError) {
+TEST_P(PreparedModelTest, executeSyncError) {
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeSynchronously(_, _, _, _, _))
             .Times(1)
             .WillOnce(Invoke(makeGeneralFailure));
@@ -115,10 +119,10 @@
     EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, executeSyncTransportFailure) {
+TEST_P(PreparedModelTest, executeSyncTransportFailure) {
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeSynchronously(_, _, _, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
@@ -131,10 +135,10 @@
     EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, executeSyncDeadObject) {
+TEST_P(PreparedModelTest, executeSyncDeadObject) {
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeSynchronously(_, _, _, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
@@ -147,10 +151,10 @@
     EXPECT_EQ(result.error().code, nn::ErrorStatus::DEAD_OBJECT);
 }
 
-TEST(PreparedModelTest, executeFenced) {
+TEST_P(PreparedModelTest, executeFenced) {
     // setup call
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     const auto mockCallback = MockFencedExecutionCallback::create();
     EXPECT_CALL(*mockCallback, getExecutionInfo(_, _, _))
             .Times(1)
@@ -176,10 +180,10 @@
                                             << callbackResult.error().message;
 }
 
-TEST(PreparedModelTest, executeFencedCallbackError) {
+TEST_P(PreparedModelTest, executeFencedCallbackError) {
     // setup call
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     const auto mockCallback = MockFencedExecutionCallback::create();
     EXPECT_CALL(*mockCallback, getExecutionInfo(_, _, _))
             .Times(1)
@@ -206,10 +210,10 @@
     EXPECT_EQ(callbackResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, executeFencedError) {
+TEST_P(PreparedModelTest, executeFencedError) {
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeGeneralFailure));
@@ -222,10 +226,10 @@
     EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, executeFencedTransportFailure) {
+TEST_P(PreparedModelTest, executeFencedTransportFailure) {
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
@@ -238,10 +242,10 @@
     EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, executeFencedDeadObject) {
+TEST_P(PreparedModelTest, executeFencedDeadObject) {
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
@@ -254,11 +258,13 @@
     EXPECT_EQ(result.error().code, nn::ErrorStatus::DEAD_OBJECT);
 }
 
-TEST(PreparedModelTest, reusableExecuteSync) {
+TEST_P(PreparedModelTest, reusableExecuteSync) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup call
     const uint32_t kNumberOfComputations = 2;
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     const auto mockExecutionResult = ExecutionResult{
             .outputSufficientSize = true,
             .outputShapes = {},
@@ -283,10 +289,12 @@
     }
 }
 
-TEST(PreparedModelTest, reusableExecuteSyncError) {
+TEST_P(PreparedModelTest, reusableExecuteSyncError) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeSynchronously(_, _, _, _, _))
             .Times(1)
             .WillOnce(Invoke(makeGeneralFailure));
@@ -303,10 +311,12 @@
     EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, reusableExecuteSyncTransportFailure) {
+TEST_P(PreparedModelTest, reusableExecuteSyncTransportFailure) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeSynchronously(_, _, _, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
@@ -323,10 +333,12 @@
     EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, reusableExecuteSyncDeadObject) {
+TEST_P(PreparedModelTest, reusableExecuteSyncDeadObject) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeSynchronously(_, _, _, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
@@ -343,11 +355,13 @@
     EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::DEAD_OBJECT);
 }
 
-TEST(PreparedModelTest, reusableExecuteFenced) {
+TEST_P(PreparedModelTest, reusableExecuteFenced) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup call
     const uint32_t kNumberOfComputations = 2;
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     const auto mockCallback = MockFencedExecutionCallback::create();
     EXPECT_CALL(*mockCallback, getExecutionInfo(_, _, _))
             .Times(kNumberOfComputations)
@@ -379,10 +393,12 @@
     }
 }
 
-TEST(PreparedModelTest, reusableExecuteFencedCallbackError) {
+TEST_P(PreparedModelTest, reusableExecuteFencedCallbackError) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup call
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     const auto mockCallback = MockFencedExecutionCallback::create();
     EXPECT_CALL(*mockCallback, getExecutionInfo(_, _, _))
             .Times(1)
@@ -413,10 +429,12 @@
     EXPECT_EQ(callbackResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, reusableExecuteFencedError) {
+TEST_P(PreparedModelTest, reusableExecuteFencedError) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeGeneralFailure));
@@ -433,10 +451,12 @@
     EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, reusableExecuteFencedTransportFailure) {
+TEST_P(PreparedModelTest, reusableExecuteFencedTransportFailure) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
@@ -453,10 +473,12 @@
     EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, reusableExecuteFencedDeadObject) {
+TEST_P(PreparedModelTest, reusableExecuteFencedDeadObject) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
     EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
@@ -473,14 +495,14 @@
     EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::DEAD_OBJECT);
 }
 
-TEST(PreparedModelTest, configureExecutionBurst) {
+TEST_P(PreparedModelTest, configureExecutionBurst) {
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
     const auto mockBurst = ndk::SharedRefBase::make<MockBurst>();
     EXPECT_CALL(*mockPreparedModel, configureExecutionBurst(_))
             .Times(1)
             .WillOnce(DoAll(SetArgPointee<0>(mockBurst), Invoke(makeStatusOk)));
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
 
     // run test
     const auto result = preparedModel->configureExecutionBurst();
@@ -491,13 +513,13 @@
     EXPECT_NE(result.value(), nullptr);
 }
 
-TEST(PreparedModelTest, configureExecutionBurstError) {
+TEST_P(PreparedModelTest, configureExecutionBurstError) {
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
     EXPECT_CALL(*mockPreparedModel, configureExecutionBurst(_))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeGeneralFailure));
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
 
     // run test
     const auto result = preparedModel->configureExecutionBurst();
@@ -507,13 +529,13 @@
     EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, configureExecutionBurstTransportFailure) {
+TEST_P(PreparedModelTest, configureExecutionBurstTransportFailure) {
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
     EXPECT_CALL(*mockPreparedModel, configureExecutionBurst(_))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
 
     // run test
     const auto result = preparedModel->configureExecutionBurst();
@@ -523,13 +545,13 @@
     EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
 }
 
-TEST(PreparedModelTest, configureExecutionBurstDeadObject) {
+TEST_P(PreparedModelTest, configureExecutionBurstDeadObject) {
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
     EXPECT_CALL(*mockPreparedModel, configureExecutionBurst(_))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
 
     // run test
     const auto result = preparedModel->configureExecutionBurst();
@@ -539,10 +561,84 @@
     EXPECT_EQ(result.error().code, nn::ErrorStatus::DEAD_OBJECT);
 }
 
-TEST(PreparedModelTest, getUnderlyingResource) {
+TEST_P(PreparedModelTest, createReusableExecution) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+    const auto mockExecution = ndk::SharedRefBase::make<MockExecution>();
+    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _, _))
+            .Times(1)
+            .WillOnce(DoAll(SetArgPointee<3>(mockExecution), Invoke(makeStatusOk)));
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+
+    // run test
+    const auto result = preparedModel->createReusableExecution({}, {}, {});
+
+    // verify result
+    ASSERT_TRUE(result.has_value())
+            << "Failed with " << result.error().code << ": " << result.error().message;
+    EXPECT_NE(result.value(), nullptr);
+}
+
+TEST_P(PreparedModelTest, createReusableExecutionError) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup test
+    const auto mockPreparedModel = MockPreparedModel::create();
+    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeGeneralFailure));
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+
+    // run test
+    const auto result = preparedModel->createReusableExecution({}, {}, {});
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(PreparedModelTest, createReusableExecutionTransportFailure) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup test
+    const auto mockPreparedModel = MockPreparedModel::create();
+    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+
+    // run test
+    const auto result = preparedModel->createReusableExecution({}, {}, {});
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(PreparedModelTest, createReusableExecutionDeadObject) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup test
+    const auto mockPreparedModel = MockPreparedModel::create();
+    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+
+    // run test
+    const auto result = preparedModel->createReusableExecution({}, {}, {});
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::DEAD_OBJECT);
+}
+
+TEST_P(PreparedModelTest, getUnderlyingResource) {
+    // setup test
+    const auto mockPreparedModel = MockPreparedModel::create();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
 
     // run test
     const auto resource = preparedModel->getUnderlyingResource();
@@ -554,4 +650,6 @@
     EXPECT_EQ(maybeMock->get(), mockPreparedModel.get());
 }
 
+INSTANTIATE_VERSIONED_AIDL_UTILS_TEST(PreparedModelTest, kAllAidlVersions);
+
 }  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/utils/test/TestUtils.cpp b/neuralnetworks/aidl/utils/test/TestUtils.cpp
new file mode 100644
index 0000000..9abec88
--- /dev/null
+++ b/neuralnetworks/aidl/utils/test/TestUtils.cpp
@@ -0,0 +1,44 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "TestUtils.h"
+
+#include <android-base/logging.h>
+#include <gtest/gtest.h>
+#include <nnapi/TypeUtils.h>
+#include <nnapi/Types.h>
+#include <nnapi/hal/CommonUtils.h>
+#include <string>
+
+namespace aidl::android::hardware::neuralnetworks::utils {
+
+std::string printTestVersion(const testing::TestParamInfo<nn::Version>& info) {
+    switch (info.param.level) {
+        case nn::Version::Level::FEATURE_LEVEL_5:
+            return "v1";
+        case nn::Version::Level::FEATURE_LEVEL_6:
+            return "v2";
+        case nn::Version::Level::FEATURE_LEVEL_7:
+            return "v3";
+        case nn::Version::Level::FEATURE_LEVEL_8:
+            return "v4";
+        default:
+            LOG(FATAL) << "Invalid AIDL version: " << info.param;
+            return "invalid";
+    }
+}
+
+}  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/utils/test/TestUtils.h b/neuralnetworks/aidl/utils/test/TestUtils.h
new file mode 100644
index 0000000..23f734a
--- /dev/null
+++ b/neuralnetworks/aidl/utils/test/TestUtils.h
@@ -0,0 +1,43 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_AIDL_UTILS_TEST_TEST_UTILS_H
+#define ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_AIDL_UTILS_TEST_TEST_UTILS_H
+
+#include <gtest/gtest.h>
+#include <nnapi/Types.h>
+#include <nnapi/hal/CommonUtils.h>
+#include <string>
+
+namespace aidl::android::hardware::neuralnetworks::utils {
+
+class VersionedAidlUtilsTestBase : public ::testing::TestWithParam<nn::Version> {
+  protected:
+    const nn::Version kVersion = GetParam();
+};
+
+std::string printTestVersion(const testing::TestParamInfo<nn::Version>& info);
+
+inline const auto kAllAidlVersions =
+        ::testing::Values(nn::kVersionFeatureLevel5, nn::kVersionFeatureLevel6,
+                          nn::kVersionFeatureLevel7, nn::kVersionFeatureLevel8);
+
+#define INSTANTIATE_VERSIONED_AIDL_UTILS_TEST(TestSuite, versions) \
+    INSTANTIATE_TEST_SUITE_P(Versioned, TestSuite, versions, printTestVersion)
+
+}  // namespace aidl::android::hardware::neuralnetworks::utils
+
+#endif  // ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_AIDL_UTILS_TEST_TEST_UTILS_H
diff --git a/neuralnetworks/aidl/vts/functional/MemoryDomainTests.cpp b/neuralnetworks/aidl/vts/functional/MemoryDomainTests.cpp
index cd5475c..b3e9c63 100644
--- a/neuralnetworks/aidl/vts/functional/MemoryDomainTests.cpp
+++ b/neuralnetworks/aidl/vts/functional/MemoryDomainTests.cpp
@@ -208,6 +208,11 @@
         return ndk::ScopedAStatus::fromServiceSpecificError(
                 static_cast<int32_t>(ErrorStatus::GENERAL_FAILURE));
     }
+    ndk::ScopedAStatus createReusableExecution(const aidl_hal::Request&, bool, int64_t,
+                                               std::shared_ptr<aidl_hal::IExecution>*) override {
+        return ndk::ScopedAStatus::fromServiceSpecificError(
+                static_cast<int32_t>(ErrorStatus::GENERAL_FAILURE));
+    }
 };
 
 template <typename... Args>
diff --git a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Execution.h b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Execution.h
new file mode 100644
index 0000000..6a9ac57
--- /dev/null
+++ b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Execution.h
@@ -0,0 +1,56 @@
+/*
+ * Copyright (C) 2022 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_UTILS_ADAPTER_AIDL_EXECUTION_H
+#define ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_UTILS_ADAPTER_AIDL_EXECUTION_H
+
+#include "nnapi/hal/aidl/Adapter.h"
+
+#include <aidl/android/hardware/neuralnetworks/BnExecution.h>
+#include <aidl/android/hardware/neuralnetworks/ExecutionResult.h>
+#include <aidl/android/hardware/neuralnetworks/FencedExecutionResult.h>
+#include <aidl/android/hardware/neuralnetworks/IExecution.h>
+#include <aidl/android/hardware/neuralnetworks/Request.h>
+#include <android/binder_auto_utils.h>
+#include <nnapi/IExecution.h>
+#include <nnapi/Types.h>
+
+#include <memory>
+#include <vector>
+
+// See hardware/interfaces/neuralnetworks/utils/README.md for more information on AIDL interface
+// lifetimes across processes and for protecting asynchronous calls across AIDL.
+
+namespace aidl::android::hardware::neuralnetworks::adapter {
+
+// Class that adapts nn::IExecution to BnExecution.
+class Execution : public BnExecution {
+  public:
+    explicit Execution(::android::nn::SharedExecution execution);
+
+    ndk::ScopedAStatus executeSynchronously(int64_t deadlineNs,
+                                            ExecutionResult* executionResult) override;
+    ndk::ScopedAStatus executeFenced(const std::vector<ndk::ScopedFileDescriptor>& waitFor,
+                                     int64_t deadlineNs, int64_t durationNs,
+                                     FencedExecutionResult* fencedExecutionResult) override;
+
+  protected:
+    const ::android::nn::SharedExecution kExecution;
+};
+
+}  // namespace aidl::android::hardware::neuralnetworks::adapter
+
+#endif  // ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_UTILS_ADAPTER_AIDL_EXECUTION_H
diff --git a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h
index 93e0427..f92b0bc 100644
--- a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h
+++ b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h
@@ -23,6 +23,7 @@
 #include <aidl/android/hardware/neuralnetworks/ExecutionResult.h>
 #include <aidl/android/hardware/neuralnetworks/FencedExecutionResult.h>
 #include <aidl/android/hardware/neuralnetworks/IBurst.h>
+#include <aidl/android/hardware/neuralnetworks/IExecution.h>
 #include <aidl/android/hardware/neuralnetworks/Request.h>
 #include <android/binder_auto_utils.h>
 #include <nnapi/IPreparedModel.h>
@@ -50,6 +51,9 @@
                                      int64_t loopTimeoutDurationNs, int64_t durationNs,
                                      FencedExecutionResult* executionResult) override;
     ndk::ScopedAStatus configureExecutionBurst(std::shared_ptr<IBurst>* burst) override;
+    ndk::ScopedAStatus createReusableExecution(const Request& request, bool measureTiming,
+                                               int64_t loopTimeoutDurationNs,
+                                               std::shared_ptr<IExecution>* execution) override;
 
     ::android::nn::SharedPreparedModel getUnderlyingPreparedModel() const;
 
diff --git a/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp b/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp
index 71ed1a8..5cab62c 100644
--- a/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp
+++ b/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp
@@ -17,6 +17,7 @@
 #include "PreparedModel.h"
 
 #include "Burst.h"
+#include "Execution.h"
 
 #include <aidl/android/hardware/neuralnetworks/BnFencedExecutionCallback.h>
 #include <aidl/android/hardware/neuralnetworks/BnPreparedModel.h>
@@ -26,6 +27,7 @@
 #include <aidl/android/hardware/neuralnetworks/Request.h>
 #include <android-base/logging.h>
 #include <android/binder_auto_utils.h>
+#include <nnapi/IExecution.h>
 #include <nnapi/IPreparedModel.h>
 #include <nnapi/Result.h>
 #include <nnapi/SharedMemory.h>
@@ -167,6 +169,56 @@
                                  .syncFence = std::move(fileDescriptor)};
 }
 
+nn::GeneralResult<nn::SharedExecution> createReusableExecution(
+        const nn::IPreparedModel& preparedModel, const Request& request, bool measureTiming,
+        int64_t loopTimeoutDurationNs) {
+    const auto nnRequest = NN_TRY(convertInput(request));
+    const auto nnMeasureTiming = measureTiming ? nn::MeasureTiming::YES : nn::MeasureTiming::NO;
+    const auto nnLoopTimeoutDuration = NN_TRY(makeOptionalDuration(loopTimeoutDurationNs));
+    return preparedModel.createReusableExecution(nnRequest, nnMeasureTiming, nnLoopTimeoutDuration);
+}
+
+nn::ExecutionResult<ExecutionResult> executeSynchronously(const nn::IExecution& execution,
+                                                          int64_t deadlineNs) {
+    const auto nnDeadline = NN_TRY(makeOptionalTimePoint(deadlineNs));
+
+    const auto result = execution.compute(nnDeadline);
+
+    if (!result.ok() && result.error().code == nn::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE) {
+        const auto& [message, code, outputShapes] = result.error();
+        LOG(ERROR) << "executeSynchronously failed with " << code << ": " << message;
+        return ExecutionResult{.outputSufficientSize = false,
+                               .outputShapes = utils::convert(outputShapes).value(),
+                               .timing = {.timeInDriverNs = -1, .timeOnDeviceNs = -1}};
+    }
+
+    const auto& [outputShapes, timing] = NN_TRY(result);
+    return ExecutionResult{.outputSufficientSize = true,
+                           .outputShapes = utils::convert(outputShapes).value(),
+                           .timing = utils::convert(timing).value()};
+}
+
+nn::GeneralResult<FencedExecutionResult> executeFenced(
+        const nn::IExecution& execution, const std::vector<ndk::ScopedFileDescriptor>& waitFor,
+        int64_t deadlineNs, int64_t durationNs) {
+    const auto nnWaitFor = NN_TRY(convertSyncFences(waitFor));
+    const auto nnDeadline = NN_TRY(makeOptionalTimePoint(deadlineNs));
+    const auto nnDuration = NN_TRY(makeOptionalDuration(durationNs));
+
+    auto [syncFence, executeFencedInfoCallback] =
+            NN_TRY(execution.computeFenced(nnWaitFor, nnDeadline, nnDuration));
+
+    ndk::ScopedFileDescriptor fileDescriptor;
+    if (syncFence.hasFd()) {
+        auto uniqueFd = NN_TRY(nn::dupFd(syncFence.getFd()));
+        fileDescriptor = ndk::ScopedFileDescriptor(uniqueFd.release());
+    }
+
+    return FencedExecutionResult{.callback = ndk::SharedRefBase::make<FencedExecutionCallback>(
+                                         std::move(executeFencedInfoCallback)),
+                                 .syncFence = std::move(fileDescriptor)};
+}
+
 }  // namespace
 
 PreparedModel::PreparedModel(nn::SharedPreparedModel preparedModel)
@@ -222,4 +274,51 @@
     return kPreparedModel;
 }
 
+ndk::ScopedAStatus PreparedModel::createReusableExecution(const Request& request,
+                                                          bool measureTiming,
+                                                          int64_t loopTimeoutDurationNs,
+                                                          std::shared_ptr<IExecution>* execution) {
+    auto result = adapter::createReusableExecution(*kPreparedModel, request, measureTiming,
+                                                   loopTimeoutDurationNs);
+    if (!result.has_value()) {
+        const auto& [message, code] = result.error();
+        const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
+        return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage(
+                static_cast<int32_t>(aidlCode), message.c_str());
+    }
+    *execution = ndk::SharedRefBase::make<Execution>(std::move(result).value());
+    return ndk::ScopedAStatus::ok();
+}
+
+Execution::Execution(nn::SharedExecution execution) : kExecution(std::move(execution)) {
+    CHECK(kExecution != nullptr);
+}
+
+ndk::ScopedAStatus Execution::executeSynchronously(int64_t deadlineNs,
+                                                   ExecutionResult* executionResult) {
+    auto result = adapter::executeSynchronously(*kExecution, deadlineNs);
+    if (!result.has_value()) {
+        const auto& [message, code, _] = result.error();
+        const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
+        return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage(
+                static_cast<int32_t>(aidlCode), message.c_str());
+    }
+    *executionResult = std::move(result).value();
+    return ndk::ScopedAStatus::ok();
+}
+
+ndk::ScopedAStatus Execution::executeFenced(const std::vector<ndk::ScopedFileDescriptor>& waitFor,
+                                            int64_t deadlineNs, int64_t durationNs,
+                                            FencedExecutionResult* executionResult) {
+    auto result = adapter::executeFenced(*kExecution, waitFor, deadlineNs, durationNs);
+    if (!result.has_value()) {
+        const auto& [message, code] = result.error();
+        const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
+        return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage(
+                static_cast<int32_t>(aidlCode), message.c_str());
+    }
+    *executionResult = std::move(result).value();
+    return ndk::ScopedAStatus::ok();
+}
+
 }  // namespace aidl::android::hardware::neuralnetworks::adapter