diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Utils.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Utils.h
index 802e703..79b511d 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Utils.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Utils.h
@@ -47,7 +47,10 @@
     return result.has_value();
 }
 
-nn::GeneralResult<Model> copyModel(const Model& model);
+nn::GeneralResult<Memory> clone(const Memory& memory);
+nn::GeneralResult<Request> clone(const Request& request);
+nn::GeneralResult<RequestMemoryPool> clone(const RequestMemoryPool& requestPool);
+nn::GeneralResult<Model> clone(const Model& model);
 
 }  // namespace aidl::android::hardware::neuralnetworks::utils
 
diff --git a/neuralnetworks/aidl/utils/src/Utils.cpp b/neuralnetworks/aidl/utils/src/Utils.cpp
index 04aa0e9..8d00e59 100644
--- a/neuralnetworks/aidl/utils/src/Utils.cpp
+++ b/neuralnetworks/aidl/utils/src/Utils.cpp
@@ -19,38 +19,77 @@
 #include <nnapi/Result.h>
 
 namespace aidl::android::hardware::neuralnetworks::utils {
+namespace {
 
 using ::android::nn::GeneralResult;
 
-GeneralResult<Model> copyModel(const Model& model) {
-    Model newModel{
+template <typename Type>
+nn::GeneralResult<std::vector<Type>> cloneVec(const std::vector<Type>& arguments) {
+    std::vector<Type> clonedObjects;
+    clonedObjects.reserve(arguments.size());
+    for (const auto& argument : arguments) {
+        clonedObjects.push_back(NN_TRY(clone(argument)));
+    }
+    return clonedObjects;
+}
+
+template <typename Type>
+GeneralResult<std::vector<Type>> clone(const std::vector<Type>& arguments) {
+    return cloneVec(arguments);
+}
+
+}  // namespace
+
+GeneralResult<Memory> clone(const Memory& memory) {
+    common::NativeHandle nativeHandle;
+    nativeHandle.ints = memory.handle.ints;
+    nativeHandle.fds.reserve(memory.handle.fds.size());
+    for (const auto& fd : memory.handle.fds) {
+        const int newFd = dup(fd.get());
+        if (newFd < 0) {
+            return NN_ERROR() << "Couldn't dup a file descriptor";
+        }
+        nativeHandle.fds.emplace_back(newFd);
+    }
+    return Memory{
+            .handle = std::move(nativeHandle),
+            .size = memory.size,
+            .name = memory.name,
+    };
+}
+
+GeneralResult<RequestMemoryPool> clone(const RequestMemoryPool& requestPool) {
+    using Tag = RequestMemoryPool::Tag;
+    switch (requestPool.getTag()) {
+        case Tag::pool:
+            return RequestMemoryPool::make<Tag::pool>(NN_TRY(clone(requestPool.get<Tag::pool>())));
+        case Tag::token:
+            return RequestMemoryPool::make<Tag::token>(requestPool.get<Tag::token>());
+    }
+    // Using explicit type conversion because std::variant inside the RequestMemoryPool confuses the
+    // compiler.
+    return (NN_ERROR() << "Unrecognized request pool tag: " << requestPool.getTag())
+            .
+            operator GeneralResult<RequestMemoryPool>();
+}
+
+GeneralResult<Request> clone(const Request& request) {
+    return Request{
+            .inputs = request.inputs,
+            .outputs = request.outputs,
+            .pools = NN_TRY(clone(request.pools)),
+    };
+}
+
+GeneralResult<Model> clone(const Model& model) {
+    return Model{
             .main = model.main,
             .referenced = model.referenced,
             .operandValues = model.operandValues,
-            .pools = {},
+            .pools = NN_TRY(clone(model.pools)),
             .relaxComputationFloat32toFloat16 = model.relaxComputationFloat32toFloat16,
             .extensionNameToPrefix = model.extensionNameToPrefix,
     };
-    newModel.pools.reserve(model.pools.size());
-    for (const auto& pool : model.pools) {
-        common::NativeHandle nativeHandle;
-        nativeHandle.ints = pool.handle.ints;
-        nativeHandle.fds.reserve(pool.handle.fds.size());
-        for (const auto& fd : pool.handle.fds) {
-            const int newFd = dup(fd.get());
-            if (newFd == -1) {
-                return NN_ERROR() << "Couldn't dup a file descriptor.";
-            }
-            nativeHandle.fds.emplace_back(newFd);
-        }
-        Memory memory = {
-                .handle = std::move(nativeHandle),
-                .size = pool.size,
-                .name = pool.name,
-        };
-        newModel.pools.push_back(std::move(memory));
-    }
-    return newModel;
 }
 
 }  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/vts/OWNERS b/neuralnetworks/aidl/vts/OWNERS
new file mode 100644
index 0000000..6719a5b
--- /dev/null
+++ b/neuralnetworks/aidl/vts/OWNERS
@@ -0,0 +1,12 @@
+# Neuralnetworks team
+butlermichael@google.com
+dgross@google.com
+jeanluc@google.com
+levp@google.com
+miaowang@google.com
+mikie@google.com
+mks@google.com
+pszczepaniak@google.com
+slavash@google.com
+vddang@google.com
+xusongw@google.com
diff --git a/neuralnetworks/aidl/vts/functional/Android.bp b/neuralnetworks/aidl/vts/functional/Android.bp
new file mode 100644
index 0000000..aa7afbf
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/Android.bp
@@ -0,0 +1,68 @@
+//
+// 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.
+//
+
+cc_test {
+    name: "VtsHalNeuralnetworksTargetTest",
+    defaults: [
+        "neuralnetworks_vts_functional_defaults",
+        "use_libaidlvintf_gtest_helper_static",
+    ],
+    srcs: [
+        "BasicTests.cpp",
+        "Callbacks.cpp",
+        "CompilationCachingTests.cpp",
+        "GeneratedTestHarness.cpp",
+        "MemoryDomainTests.cpp",
+        "QualityOfServiceTests.cpp",
+        "TestAssertions.cpp",
+        "TestMain.cpp",
+        "Utils.cpp",
+        "ValidateModel.cpp",
+        "ValidateRequest.cpp",
+        "VtsHalNeuralnetworks.cpp",
+    ],
+    shared_libs: [
+        "libbinder_ndk",
+        "libnativewindow",
+        "libvndksupport",
+    ],
+    static_libs: [
+        "android.hardware.common-V2-ndk_platform",
+        "android.hardware.neuralnetworks-V1-ndk_platform",
+        "android.hidl.allocator@1.0",
+        "android.hidl.memory@1.0",
+        "libgmock",
+        "libhidlmemory",
+        "libneuralnetworks_generated_test_harness",
+        "libneuralnetworks_utils",
+        "libsync",
+        "neuralnetworks_utils_hal_aidl",
+    ],
+    whole_static_libs: [
+        "neuralnetworks_generated_V1_0_example",
+        "neuralnetworks_generated_V1_1_example",
+        "neuralnetworks_generated_V1_2_example",
+        "neuralnetworks_generated_V1_3_example",
+    ],
+    header_libs: [
+        "libbase_headers",
+        "libneuralnetworks_headers",
+    ],
+    test_suites: [
+        "general-tests",
+        "vts",
+    ],
+}
diff --git a/neuralnetworks/aidl/vts/functional/AndroidTest.xml b/neuralnetworks/aidl/vts/functional/AndroidTest.xml
new file mode 100644
index 0000000..384d420
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/AndroidTest.xml
@@ -0,0 +1,33 @@
+<?xml version="1.0" encoding="utf-8"?>
+<!-- Copyright (C) 2020 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.
+-->
+<configuration description="Runs VtsHalNeuralnetworksTargetTest.">
+    <option name="test-suite-tag" value="apct" />
+    <option name="test-suite-tag" value="apct-native" />
+
+    <target_preparer class="com.android.tradefed.targetprep.RootTargetPreparer">
+    </target_preparer>
+
+    <target_preparer class="com.android.tradefed.targetprep.PushFilePreparer">
+        <option name="cleanup" value="true" />
+        <option name="push" value="VtsHalNeuralnetworksTargetTest->/data/local/tmp/VtsHalNeuralnetworksTargetTest" />
+    </target_preparer>
+
+    <test class="com.android.tradefed.testtype.GTest" >
+        <option name="native-test-device-path" value="/data/local/tmp" />
+        <option name="module-name" value="VtsHalNeuralnetworksTargetTest" />
+        <option name="native-test-timeout" value="20m" />
+    </test>
+</configuration>
diff --git a/neuralnetworks/aidl/vts/functional/BasicTests.cpp b/neuralnetworks/aidl/vts/functional/BasicTests.cpp
new file mode 100644
index 0000000..b2f4507
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/BasicTests.cpp
@@ -0,0 +1,193 @@
+/*
+ * 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.
+ */
+
+#define LOG_TAG "neuralnetworks_aidl_hal_test"
+
+#include <aidl/android/hardware/neuralnetworks/Capabilities.h>
+#include <aidl/android/hardware/neuralnetworks/IDevice.h>
+#include <aidl/android/hardware/neuralnetworks/Operand.h>
+#include <aidl/android/hardware/neuralnetworks/OperandType.h>
+#include <aidl/android/hardware/neuralnetworks/Priority.h>
+#include <android/binder_interface_utils.h>
+
+#include "Utils.h"
+#include "VtsHalNeuralnetworks.h"
+
+namespace aidl::android::hardware::neuralnetworks::vts::functional {
+
+using implementation::PreparedModelCallback;
+
+// create device test
+TEST_P(NeuralNetworksAidlTest, CreateDevice) {}
+
+// initialization
+TEST_P(NeuralNetworksAidlTest, GetCapabilitiesTest) {
+    Capabilities capabilities;
+    const auto retStatus = kDevice->getCapabilities(&capabilities);
+    ASSERT_TRUE(retStatus.isOk());
+
+    auto isPositive = [](const PerformanceInfo& perf) {
+        return perf.execTime > 0.0f && perf.powerUsage > 0.0f;
+    };
+
+    EXPECT_TRUE(isPositive(capabilities.relaxedFloat32toFloat16PerformanceScalar));
+    EXPECT_TRUE(isPositive(capabilities.relaxedFloat32toFloat16PerformanceTensor));
+    const auto& opPerf = capabilities.operandPerformance;
+    EXPECT_TRUE(
+            std::all_of(opPerf.begin(), opPerf.end(),
+                        [isPositive](const OperandPerformance& a) { return isPositive(a.info); }));
+    EXPECT_TRUE(std::is_sorted(opPerf.begin(), opPerf.end(),
+                               [](const OperandPerformance& a, const OperandPerformance& b) {
+                                   return a.type < b.type;
+                               }));
+    EXPECT_TRUE(std::all_of(opPerf.begin(), opPerf.end(), [](const OperandPerformance& a) {
+        return a.type != OperandType::SUBGRAPH;
+    }));
+    EXPECT_TRUE(isPositive(capabilities.ifPerformance));
+    EXPECT_TRUE(isPositive(capabilities.whilePerformance));
+}
+
+// detect cycle
+TEST_P(NeuralNetworksAidlTest, CycleTest) {
+    // opnd0 = TENSOR_FLOAT32            // model input
+    // opnd1 = TENSOR_FLOAT32            // model input
+    // opnd2 = INT32                     // model input
+    // opnd3 = ADD(opnd0, opnd4, opnd2)
+    // opnd4 = ADD(opnd1, opnd3, opnd2)
+    // opnd5 = ADD(opnd4, opnd0, opnd2)  // model output
+    //
+    //            +-----+
+    //            |     |
+    //            v     |
+    // 3 = ADD(0, 4, 2) |
+    // |                |
+    // +----------+     |
+    //            |     |
+    //            v     |
+    // 4 = ADD(1, 3, 2) |
+    // |                |
+    // +----------------+
+    // |
+    // |
+    // +-------+
+    //         |
+    //         v
+    // 5 = ADD(4, 0, 2)
+
+    const std::vector<Operand> operands = {
+            {
+                    // operands[0]
+                    .type = OperandType::TENSOR_FLOAT32,
+                    .dimensions = {1},
+                    .scale = 0.0f,
+                    .zeroPoint = 0,
+                    .lifetime = OperandLifeTime::SUBGRAPH_INPUT,
+                    .location = {.poolIndex = 0, .offset = 0, .length = 0},
+            },
+            {
+                    // operands[1]
+                    .type = OperandType::TENSOR_FLOAT32,
+                    .dimensions = {1},
+                    .scale = 0.0f,
+                    .zeroPoint = 0,
+                    .lifetime = OperandLifeTime::SUBGRAPH_INPUT,
+                    .location = {.poolIndex = 0, .offset = 0, .length = 0},
+            },
+            {
+                    // operands[2]
+                    .type = OperandType::INT32,
+                    .dimensions = {},
+                    .scale = 0.0f,
+                    .zeroPoint = 0,
+                    .lifetime = OperandLifeTime::SUBGRAPH_INPUT,
+                    .location = {.poolIndex = 0, .offset = 0, .length = 0},
+            },
+            {
+                    // operands[3]
+                    .type = OperandType::TENSOR_FLOAT32,
+                    .dimensions = {1},
+                    .scale = 0.0f,
+                    .zeroPoint = 0,
+                    .lifetime = OperandLifeTime::TEMPORARY_VARIABLE,
+                    .location = {.poolIndex = 0, .offset = 0, .length = 0},
+            },
+            {
+                    // operands[4]
+                    .type = OperandType::TENSOR_FLOAT32,
+                    .dimensions = {1},
+                    .scale = 0.0f,
+                    .zeroPoint = 0,
+                    .lifetime = OperandLifeTime::TEMPORARY_VARIABLE,
+                    .location = {.poolIndex = 0, .offset = 0, .length = 0},
+            },
+            {
+                    // operands[5]
+                    .type = OperandType::TENSOR_FLOAT32,
+                    .dimensions = {1},
+                    .scale = 0.0f,
+                    .zeroPoint = 0,
+                    .lifetime = OperandLifeTime::SUBGRAPH_OUTPUT,
+                    .location = {.poolIndex = 0, .offset = 0, .length = 0},
+            },
+    };
+
+    const std::vector<Operation> operations = {
+            {.type = OperationType::ADD, .inputs = {0, 4, 2}, .outputs = {3}},
+            {.type = OperationType::ADD, .inputs = {1, 3, 2}, .outputs = {4}},
+            {.type = OperationType::ADD, .inputs = {4, 0, 2}, .outputs = {5}},
+    };
+
+    Subgraph subgraph = {
+            .operands = operands,
+            .operations = operations,
+            .inputIndexes = {0, 1, 2},
+            .outputIndexes = {5},
+    };
+    const Model model = {
+            .main = std::move(subgraph),
+            .referenced = {},
+            .operandValues = {},
+            .pools = {},
+    };
+
+    // ensure that getSupportedOperations() checks model validity
+    std::vector<bool> supportedOps;
+    const auto supportedOpsStatus = kDevice->getSupportedOperations(model, &supportedOps);
+    ASSERT_FALSE(supportedOpsStatus.isOk());
+    ASSERT_EQ(supportedOpsStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
+    ASSERT_EQ(static_cast<ErrorStatus>(supportedOpsStatus.getServiceSpecificError()),
+              ErrorStatus::INVALID_ARGUMENT);
+
+    // ensure that prepareModel() checks model validity
+    auto preparedModelCallback = ndk::SharedRefBase::make<PreparedModelCallback>();
+    auto prepareLaunchStatus =
+            kDevice->prepareModel(model, ExecutionPreference::FAST_SINGLE_ANSWER, kDefaultPriority,
+                                  kNoDeadline, {}, {}, kEmptyCacheToken, preparedModelCallback);
+    //     Note that preparation can fail for reasons other than an
+    //     invalid model (invalid model should result in
+    //     INVALID_ARGUMENT) -- for example, perhaps not all
+    //     operations are supported, or perhaps the device hit some
+    //     kind of capacity limit.
+    ASSERT_FALSE(prepareLaunchStatus.isOk());
+    EXPECT_EQ(prepareLaunchStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
+    EXPECT_NE(static_cast<ErrorStatus>(prepareLaunchStatus.getServiceSpecificError()),
+              ErrorStatus::NONE);
+
+    EXPECT_NE(preparedModelCallback->getStatus(), ErrorStatus::NONE);
+    EXPECT_EQ(preparedModelCallback->getPreparedModel(), nullptr);
+}
+
+}  // namespace aidl::android::hardware::neuralnetworks::vts::functional
diff --git a/neuralnetworks/aidl/vts/functional/Callbacks.cpp b/neuralnetworks/aidl/vts/functional/Callbacks.cpp
new file mode 100644
index 0000000..ca2bb48
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/Callbacks.cpp
@@ -0,0 +1,59 @@
+/*
+ * 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.
+ */
+
+#define LOG_TAG "Callbacks"
+
+#include "Callbacks.h"
+
+#include <android-base/logging.h>
+#include <android/binder_auto_utils.h>
+#include <limits>
+
+namespace aidl::android::hardware::neuralnetworks::implementation {
+
+ndk::ScopedAStatus PreparedModelCallback::notify(
+        ErrorStatus errorStatus, const std::shared_ptr<IPreparedModel>& preparedModel) {
+    {
+        std::lock_guard<std::mutex> hold(mMutex);
+        // quick-return if object has already been notified
+        if (mNotified) {
+            return ndk::ScopedAStatus::ok();
+        }
+        // store results and mark as notified
+        mErrorStatus = errorStatus;
+        mPreparedModel = preparedModel;
+        mNotified = true;
+    }
+    mCondition.notify_all();
+    return ndk::ScopedAStatus::ok();
+}
+
+void PreparedModelCallback::wait() const {
+    std::unique_lock<std::mutex> lock(mMutex);
+    mCondition.wait(lock, [this] { return mNotified; });
+}
+
+ErrorStatus PreparedModelCallback::getStatus() const {
+    wait();
+    return mErrorStatus;
+}
+
+std::shared_ptr<IPreparedModel> PreparedModelCallback::getPreparedModel() const {
+    wait();
+    return mPreparedModel;
+}
+
+}  // namespace aidl::android::hardware::neuralnetworks::implementation
diff --git a/neuralnetworks/aidl/vts/functional/Callbacks.h b/neuralnetworks/aidl/vts/functional/Callbacks.h
new file mode 100644
index 0000000..0eb4d5f
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/Callbacks.h
@@ -0,0 +1,131 @@
+/*
+ * 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_NEURALNETWORKS_AIDL_CALLBACKS_H
+#define ANDROID_HARDWARE_NEURALNETWORKS_AIDL_CALLBACKS_H
+
+#include <android-base/thread_annotations.h>
+#include <condition_variable>
+#include <mutex>
+
+#include <aidl/android/hardware/neuralnetworks/BnPreparedModelCallback.h>
+#include <aidl/android/hardware/neuralnetworks/ErrorStatus.h>
+#include <aidl/android/hardware/neuralnetworks/IPreparedModel.h>
+
+/*
+ * The Callback classes are used internally by the NeuralNetworks runtime to
+ * synchronize between different threads. An asynchronous task is launched
+ * paired with a callback object. When a client thread requires the output being
+ * generated by the asynchronous task, the client thread can wait for the result
+ * and be blocked until it has completed. Any wait may safely be called
+ * concurrently, even on the same callback object. When the asynchronous task
+ * has finished its workload, it must immediately call "notify". If the
+ * asynchronous task has failed to launch, the function that tried to launch the
+ * asynchronous task must immediately call "notify". This "notify" call
+ * awakens any client threads waiting on the callback object.
+ *
+ * These classes exist to enable synchronization across AIDL. When
+ * synchronization is only required in the same process, consider using
+ * std::future, std::mutex, std::condition_variable, or std::experimental::latch
+ * instead.
+ */
+
+namespace aidl::android::hardware::neuralnetworks::implementation {
+
+/**
+ * The PreparedModelCallback class is used to receive the error status of
+ * preparing a model as well as the prepared model from a task executing
+ * asynchronously with respect to the runtime. If a calling thread calls wait
+ * or get* on a PreparedModelCallback object and the corresponding asynchronous
+ * task has not finished preparing the model, the calling thread will block
+ * until the asynchronous task has called notify.
+ *
+ * If the callback object is notified more than once, only the results of the
+ * first call to notify are used, and the results from subsequent calls are
+ * discarded.
+ *
+ * This callback object is passed as an argument to IDevice::prepareModel*.
+ */
+class PreparedModelCallback : public BnPreparedModelCallback {
+  public:
+    /**
+     * IPreparedModelCallback::notify marks the callback object with the return
+     * status of the asynchronous model preparation along with the prepared
+     * model, and allows all prior and future wait calls on the
+     * PreparedModelCallback object to proceed.
+     *
+     * IPreparedModelCallback::notify must be called on a given PreparedModelCallback object.
+     *
+     * If the callback object is notified more than once, only the results of
+     * the first call to notify are used, and the results from subsequent calls
+     * are discarded.
+     *
+     * @param status Error status returned from asynchronously preparing the
+     *     model; will be:
+     *     - NONE if the asynchronous preparation was successful
+     *     - DEVICE_UNAVAILABLE if driver is offline or busy
+     *     - GENERAL_FAILURE if there is an unspecified error
+     *     - INVALID_ARGUMENT if the input model is invalid
+     * @param preparedModel Returned model that has been prepared for execution,
+     *     nullptr if the model was unable to be prepared.
+     */
+    ndk::ScopedAStatus notify(ErrorStatus status,
+                              const std::shared_ptr<IPreparedModel>& preparedModel) override;
+
+    /**
+     * PreparedModelCallback::wait blocks until notify has been called on the
+     * callback object.
+     */
+    void wait() const;
+
+    /**
+     * Retrieves the error status returned from the asynchronous task launched
+     * by IDevice::prepareModel*. If IDevice::prepareModel* has not finished
+     * asynchronously preparing the model, this call will block until the
+     * asynchronous task notifies the object.
+     *
+     * @return status Error status returned from asynchronously preparing the
+     *     model; will be:
+     *     - NONE if the asynchronous preparation was successful
+     *     - DEVICE_UNAVAILABLE if driver is offline or busy
+     *     - GENERAL_FAILURE if there is an unspecified error
+     *     - INVALID_ARGUMENT if the input model is invalid
+     */
+    ErrorStatus getStatus() const;
+
+    /**
+     * Retrieves the model that has been prepared for execution from the
+     * asynchronous task launched by IDevice::prepareModel*. If
+     * IDevice::prepareModel* has not finished asynchronously preparing the
+     * model, this call will block until the asynchronous task notifies the
+     * object.
+     *
+     * @return preparedModel Returned model that has been prepared for
+     *     execution, nullptr if the model was unable to be prepared.
+     */
+    std::shared_ptr<IPreparedModel> getPreparedModel() const;
+
+  private:
+    mutable std::mutex mMutex;
+    mutable std::condition_variable mCondition;
+    bool mNotified GUARDED_BY(mMutex) = false;
+    ErrorStatus mErrorStatus = ErrorStatus::GENERAL_FAILURE;
+    std::shared_ptr<IPreparedModel> mPreparedModel;
+};
+
+}  // namespace aidl::android::hardware::neuralnetworks::implementation
+
+#endif  // ANDROID_HARDWARE_NEURALNETWORKS_AIDL_CALLBACKS_H
diff --git a/neuralnetworks/aidl/vts/functional/CompilationCachingTests.cpp b/neuralnetworks/aidl/vts/functional/CompilationCachingTests.cpp
new file mode 100644
index 0000000..e0b529f
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/CompilationCachingTests.cpp
@@ -0,0 +1,1177 @@
+/*
+ * 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.
+ */
+
+#define LOG_TAG "neuralnetworks_aidl_hal_test"
+
+#include <android-base/logging.h>
+#include <android/binder_auto_utils.h>
+#include <android/binder_interface_utils.h>
+#include <android/binder_status.h>
+#include <fcntl.h>
+#include <ftw.h>
+#include <gtest/gtest.h>
+#include <hidlmemory/mapping.h>
+#include <unistd.h>
+
+#include <cstdio>
+#include <cstdlib>
+#include <iterator>
+#include <random>
+#include <thread>
+
+#include "Callbacks.h"
+#include "GeneratedTestHarness.h"
+#include "MemoryUtils.h"
+#include "TestHarness.h"
+#include "Utils.h"
+#include "VtsHalNeuralnetworks.h"
+
+// Forward declaration of the mobilenet generated test models in
+// frameworks/ml/nn/runtime/test/generated/.
+namespace generated_tests::mobilenet_224_gender_basic_fixed {
+const test_helper::TestModel& get_test_model();
+}  // namespace generated_tests::mobilenet_224_gender_basic_fixed
+
+namespace generated_tests::mobilenet_quantized {
+const test_helper::TestModel& get_test_model();
+}  // namespace generated_tests::mobilenet_quantized
+
+namespace aidl::android::hardware::neuralnetworks::vts::functional {
+
+using namespace test_helper;
+using implementation::PreparedModelCallback;
+
+namespace float32_model {
+
+constexpr auto get_test_model = generated_tests::mobilenet_224_gender_basic_fixed::get_test_model;
+
+}  // namespace float32_model
+
+namespace quant8_model {
+
+constexpr auto get_test_model = generated_tests::mobilenet_quantized::get_test_model;
+
+}  // namespace quant8_model
+
+namespace {
+
+enum class AccessMode { READ_WRITE, READ_ONLY, WRITE_ONLY };
+
+// Creates cache handles based on provided file groups.
+// The outer vector corresponds to handles and the inner vector is for fds held by each handle.
+void createCacheFds(const std::vector<std::string>& files, const std::vector<AccessMode>& mode,
+                    std::vector<ndk::ScopedFileDescriptor>* fds) {
+    fds->clear();
+    fds->reserve(files.size());
+    for (uint32_t i = 0; i < files.size(); i++) {
+        const auto& file = files[i];
+        int fd;
+        if (mode[i] == AccessMode::READ_ONLY) {
+            fd = open(file.c_str(), O_RDONLY);
+        } else if (mode[i] == AccessMode::WRITE_ONLY) {
+            fd = open(file.c_str(), O_WRONLY | O_CREAT, S_IRUSR | S_IWUSR);
+        } else if (mode[i] == AccessMode::READ_WRITE) {
+            fd = open(file.c_str(), O_RDWR | O_CREAT, S_IRUSR | S_IWUSR);
+        } else {
+            FAIL();
+        }
+        ASSERT_GE(fd, 0);
+        fds->emplace_back(fd);
+    }
+}
+
+void createCacheFds(const std::vector<std::string>& files, AccessMode mode,
+                    std::vector<ndk::ScopedFileDescriptor>* fds) {
+    createCacheFds(files, std::vector<AccessMode>(files.size(), mode), fds);
+}
+
+// Create a chain of broadcast operations. The second operand is always constant tensor [1].
+// For simplicity, activation scalar is shared. The second operand is not shared
+// in the model to let driver maintain a non-trivial size of constant data and the corresponding
+// data locations in cache.
+//
+//                --------- activation --------
+//                ↓      ↓      ↓             ↓
+// E.g. input -> ADD -> ADD -> ADD -> ... -> ADD -> output
+//                ↑      ↑      ↑             ↑
+//               [1]    [1]    [1]           [1]
+//
+// This function assumes the operation is either ADD or MUL.
+template <typename CppType, TestOperandType operandType>
+TestModel createLargeTestModelImpl(TestOperationType op, uint32_t len) {
+    EXPECT_TRUE(op == TestOperationType::ADD || op == TestOperationType::MUL);
+
+    // Model operations and operands.
+    std::vector<TestOperation> operations(len);
+    std::vector<TestOperand> operands(len * 2 + 2);
+
+    // The activation scalar, value = 0.
+    operands[0] = {
+            .type = TestOperandType::INT32,
+            .dimensions = {},
+            .numberOfConsumers = len,
+            .scale = 0.0f,
+            .zeroPoint = 0,
+            .lifetime = TestOperandLifeTime::CONSTANT_COPY,
+            .data = TestBuffer::createFromVector<int32_t>({0}),
+    };
+
+    // The buffer value of the constant second operand. The logical value is always 1.0f.
+    CppType bufferValue;
+    // The scale of the first and second operand.
+    float scale1, scale2;
+    if (operandType == TestOperandType::TENSOR_FLOAT32) {
+        bufferValue = 1.0f;
+        scale1 = 0.0f;
+        scale2 = 0.0f;
+    } else if (op == TestOperationType::ADD) {
+        bufferValue = 1;
+        scale1 = 1.0f;
+        scale2 = 1.0f;
+    } else {
+        // To satisfy the constraint on quant8 MUL: input0.scale * input1.scale < output.scale,
+        // set input1 to have scale = 0.5f and bufferValue = 2, i.e. 1.0f in floating point.
+        bufferValue = 2;
+        scale1 = 1.0f;
+        scale2 = 0.5f;
+    }
+
+    for (uint32_t i = 0; i < len; i++) {
+        const uint32_t firstInputIndex = i * 2 + 1;
+        const uint32_t secondInputIndex = firstInputIndex + 1;
+        const uint32_t outputIndex = secondInputIndex + 1;
+
+        // The first operation input.
+        operands[firstInputIndex] = {
+                .type = operandType,
+                .dimensions = {1},
+                .numberOfConsumers = 1,
+                .scale = scale1,
+                .zeroPoint = 0,
+                .lifetime = (i == 0 ? TestOperandLifeTime::MODEL_INPUT
+                                    : TestOperandLifeTime::TEMPORARY_VARIABLE),
+                .data = (i == 0 ? TestBuffer::createFromVector<CppType>({1}) : TestBuffer()),
+        };
+
+        // The second operation input, value = 1.
+        operands[secondInputIndex] = {
+                .type = operandType,
+                .dimensions = {1},
+                .numberOfConsumers = 1,
+                .scale = scale2,
+                .zeroPoint = 0,
+                .lifetime = TestOperandLifeTime::CONSTANT_COPY,
+                .data = TestBuffer::createFromVector<CppType>({bufferValue}),
+        };
+
+        // The operation. All operations share the same activation scalar.
+        // The output operand is created as an input in the next iteration of the loop, in the case
+        // of all but the last member of the chain; and after the loop as a model output, in the
+        // case of the last member of the chain.
+        operations[i] = {
+                .type = op,
+                .inputs = {firstInputIndex, secondInputIndex, /*activation scalar*/ 0},
+                .outputs = {outputIndex},
+        };
+    }
+
+    // For TestOperationType::ADD, output = 1 + 1 * len = len + 1
+    // For TestOperationType::MUL, output = 1 * 1 ^ len = 1
+    CppType outputResult = static_cast<CppType>(op == TestOperationType::ADD ? len + 1u : 1u);
+
+    // The model output.
+    operands.back() = {
+            .type = operandType,
+            .dimensions = {1},
+            .numberOfConsumers = 0,
+            .scale = scale1,
+            .zeroPoint = 0,
+            .lifetime = TestOperandLifeTime::MODEL_OUTPUT,
+            .data = TestBuffer::createFromVector<CppType>({outputResult}),
+    };
+
+    return {
+            .main = {.operands = std::move(operands),
+                     .operations = std::move(operations),
+                     .inputIndexes = {1},
+                     .outputIndexes = {len * 2 + 1}},
+            .isRelaxed = false,
+    };
+}
+
+}  // namespace
+
+// Tag for the compilation caching tests.
+class CompilationCachingTestBase : public testing::Test {
+  protected:
+    CompilationCachingTestBase(std::shared_ptr<IDevice> device, OperandType type)
+        : kDevice(std::move(device)), kOperandType(type) {}
+
+    void SetUp() override {
+        testing::Test::SetUp();
+        ASSERT_NE(kDevice.get(), nullptr);
+
+        // Create cache directory. The cache directory and a temporary cache file is always created
+        // to test the behavior of prepareModelFromCache, even when caching is not supported.
+        char cacheDirTemp[] = "/data/local/tmp/TestCompilationCachingXXXXXX";
+        char* cacheDir = mkdtemp(cacheDirTemp);
+        ASSERT_NE(cacheDir, nullptr);
+        mCacheDir = cacheDir;
+        mCacheDir.push_back('/');
+
+        NumberOfCacheFiles numCacheFiles;
+        const auto ret = kDevice->getNumberOfCacheFilesNeeded(&numCacheFiles);
+        ASSERT_TRUE(ret.isOk());
+
+        mNumModelCache = numCacheFiles.numModelCache;
+        mNumDataCache = numCacheFiles.numDataCache;
+        ASSERT_GE(mNumModelCache, 0) << "Invalid numModelCache: " << mNumModelCache;
+        ASSERT_GE(mNumDataCache, 0) << "Invalid numDataCache: " << mNumDataCache;
+        mIsCachingSupported = mNumModelCache > 0 || mNumDataCache > 0;
+
+        // Create empty cache files.
+        mTmpCache = mCacheDir + "tmp";
+        for (uint32_t i = 0; i < mNumModelCache; i++) {
+            mModelCache.push_back({mCacheDir + "model" + std::to_string(i)});
+        }
+        for (uint32_t i = 0; i < mNumDataCache; i++) {
+            mDataCache.push_back({mCacheDir + "data" + std::to_string(i)});
+        }
+        // Placeholder handles, use AccessMode::WRITE_ONLY for createCacheFds to create files.
+        std::vector<ndk::ScopedFileDescriptor> modelHandle, dataHandle, tmpHandle;
+        createCacheFds(mModelCache, AccessMode::WRITE_ONLY, &modelHandle);
+        createCacheFds(mDataCache, AccessMode::WRITE_ONLY, &dataHandle);
+        createCacheFds({mTmpCache}, AccessMode::WRITE_ONLY, &tmpHandle);
+
+        if (!mIsCachingSupported) {
+            LOG(INFO) << "NN VTS: Early termination of test because vendor service does not "
+                         "support compilation caching.";
+            std::cout << "[          ]   Early termination of test because vendor service does not "
+                         "support compilation caching."
+                      << std::endl;
+        }
+    }
+
+    void TearDown() override {
+        // If the test passes, remove the tmp directory.  Otherwise, keep it for debugging purposes.
+        if (!testing::Test::HasFailure()) {
+            // Recursively remove the cache directory specified by mCacheDir.
+            auto callback = [](const char* entry, const struct stat*, int, struct FTW*) {
+                return remove(entry);
+            };
+            nftw(mCacheDir.c_str(), callback, 128, FTW_DEPTH | FTW_MOUNT | FTW_PHYS);
+        }
+        testing::Test::TearDown();
+    }
+
+    // Model and examples creators. According to kOperandType, the following methods will return
+    // either float32 model/examples or the quant8 variant.
+    TestModel createTestModel() {
+        if (kOperandType == OperandType::TENSOR_FLOAT32) {
+            return float32_model::get_test_model();
+        } else {
+            return quant8_model::get_test_model();
+        }
+    }
+
+    TestModel createLargeTestModel(OperationType op, uint32_t len) {
+        if (kOperandType == OperandType::TENSOR_FLOAT32) {
+            return createLargeTestModelImpl<float, TestOperandType::TENSOR_FLOAT32>(
+                    static_cast<TestOperationType>(op), len);
+        } else {
+            return createLargeTestModelImpl<uint8_t, TestOperandType::TENSOR_QUANT8_ASYMM>(
+                    static_cast<TestOperationType>(op), len);
+        }
+    }
+
+    // See if the service can handle the model.
+    bool isModelFullySupported(const Model& model) {
+        std::vector<bool> supportedOps;
+        const auto supportedCall = kDevice->getSupportedOperations(model, &supportedOps);
+        EXPECT_TRUE(supportedCall.isOk());
+        EXPECT_EQ(supportedOps.size(), model.main.operations.size());
+        if (!supportedCall.isOk() || supportedOps.size() != model.main.operations.size()) {
+            return false;
+        }
+        return std::all_of(supportedOps.begin(), supportedOps.end(),
+                           [](bool valid) { return valid; });
+    }
+
+    void saveModelToCache(const Model& model,
+                          const std::vector<ndk::ScopedFileDescriptor>& modelCache,
+                          const std::vector<ndk::ScopedFileDescriptor>& dataCache,
+                          std::shared_ptr<IPreparedModel>* preparedModel = nullptr) {
+        if (preparedModel != nullptr) *preparedModel = nullptr;
+
+        // Launch prepare model.
+        std::shared_ptr<PreparedModelCallback> preparedModelCallback =
+                ndk::SharedRefBase::make<PreparedModelCallback>();
+        std::vector<uint8_t> cacheToken(std::begin(mToken), std::end(mToken));
+        const auto prepareLaunchStatus = kDevice->prepareModel(
+                model, ExecutionPreference::FAST_SINGLE_ANSWER, kDefaultPriority, kNoDeadline,
+                modelCache, dataCache, cacheToken, preparedModelCallback);
+        ASSERT_TRUE(prepareLaunchStatus.isOk());
+
+        // Retrieve prepared model.
+        preparedModelCallback->wait();
+        ASSERT_EQ(preparedModelCallback->getStatus(), ErrorStatus::NONE);
+        if (preparedModel != nullptr) {
+            *preparedModel = preparedModelCallback->getPreparedModel();
+        }
+    }
+
+    bool checkEarlyTermination(ErrorStatus status) {
+        if (status == ErrorStatus::GENERAL_FAILURE) {
+            LOG(INFO) << "NN VTS: Early termination of test because vendor service cannot "
+                         "save the prepared model that it does not support.";
+            std::cout << "[          ]   Early termination of test because vendor service cannot "
+                         "save the prepared model that it does not support."
+                      << std::endl;
+            return true;
+        }
+        return false;
+    }
+
+    bool checkEarlyTermination(const Model& model) {
+        if (!isModelFullySupported(model)) {
+            LOG(INFO) << "NN VTS: Early termination of test because vendor service cannot "
+                         "prepare model that it does not support.";
+            std::cout << "[          ]   Early termination of test because vendor service cannot "
+                         "prepare model that it does not support."
+                      << std::endl;
+            return true;
+        }
+        return false;
+    }
+
+    void prepareModelFromCache(const std::vector<ndk::ScopedFileDescriptor>& modelCache,
+                               const std::vector<ndk::ScopedFileDescriptor>& dataCache,
+                               std::shared_ptr<IPreparedModel>* preparedModel,
+                               ErrorStatus* status) {
+        // Launch prepare model from cache.
+        std::shared_ptr<PreparedModelCallback> preparedModelCallback =
+                ndk::SharedRefBase::make<PreparedModelCallback>();
+        std::vector<uint8_t> cacheToken(std::begin(mToken), std::end(mToken));
+        const auto prepareLaunchStatus = kDevice->prepareModelFromCache(
+                kNoDeadline, modelCache, dataCache, cacheToken, preparedModelCallback);
+        ASSERT_TRUE(prepareLaunchStatus.isOk() ||
+                    prepareLaunchStatus.getExceptionCode() == EX_SERVICE_SPECIFIC)
+                << "prepareLaunchStatus: " << prepareLaunchStatus.getDescription();
+        if (!prepareLaunchStatus.isOk()) {
+            *preparedModel = nullptr;
+            *status = static_cast<ErrorStatus>(prepareLaunchStatus.getServiceSpecificError());
+            return;
+        }
+
+        // Retrieve prepared model.
+        preparedModelCallback->wait();
+        *status = preparedModelCallback->getStatus();
+        *preparedModel = preparedModelCallback->getPreparedModel();
+    }
+
+    // Absolute path to the temporary cache directory.
+    std::string mCacheDir;
+
+    // Groups of file paths for model and data cache in the tmp cache directory, initialized with
+    // size = mNum{Model|Data}Cache. The outer vector corresponds to handles and the inner vector is
+    // for fds held by each handle.
+    std::vector<std::string> mModelCache;
+    std::vector<std::string> mDataCache;
+
+    // A separate temporary file path in the tmp cache directory.
+    std::string mTmpCache;
+
+    uint8_t mToken[static_cast<uint32_t>(IDevice::BYTE_SIZE_OF_CACHE_TOKEN)] = {};
+    uint32_t mNumModelCache;
+    uint32_t mNumDataCache;
+    uint32_t mIsCachingSupported;
+
+    const std::shared_ptr<IDevice> kDevice;
+    // The primary data type of the testModel.
+    const OperandType kOperandType;
+};
+
+using CompilationCachingTestParam = std::tuple<NamedDevice, OperandType>;
+
+// A parameterized fixture of CompilationCachingTestBase. Every test will run twice, with the first
+// pass running with float32 models and the second pass running with quant8 models.
+class CompilationCachingTest : public CompilationCachingTestBase,
+                               public testing::WithParamInterface<CompilationCachingTestParam> {
+  protected:
+    CompilationCachingTest()
+        : CompilationCachingTestBase(getData(std::get<NamedDevice>(GetParam())),
+                                     std::get<OperandType>(GetParam())) {}
+};
+
+TEST_P(CompilationCachingTest, CacheSavingAndRetrieval) {
+    // Create test HIDL model and compile.
+    const TestModel& testModel = createTestModel();
+    const Model model = createModel(testModel);
+    if (checkEarlyTermination(model)) return;
+    std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+
+    // Save the compilation to cache.
+    {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        saveModelToCache(model, modelCache, dataCache);
+    }
+
+    // Retrieve preparedModel from cache.
+    {
+        preparedModel = nullptr;
+        ErrorStatus status;
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (!mIsCachingSupported) {
+            ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+            ASSERT_EQ(preparedModel, nullptr);
+            return;
+        } else if (checkEarlyTermination(status)) {
+            ASSERT_EQ(preparedModel, nullptr);
+            return;
+        } else {
+            ASSERT_EQ(status, ErrorStatus::NONE);
+            ASSERT_NE(preparedModel, nullptr);
+        }
+    }
+
+    // Execute and verify results.
+    EvaluatePreparedModel(kDevice, preparedModel, testModel, /*testKind=*/TestKind::GENERAL);
+}
+
+TEST_P(CompilationCachingTest, CacheSavingAndRetrievalNonZeroOffset) {
+    // Create test HIDL model and compile.
+    const TestModel& testModel = createTestModel();
+    const Model model = createModel(testModel);
+    if (checkEarlyTermination(model)) return;
+    std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+
+    // Save the compilation to cache.
+    {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        uint8_t placeholderBytes[] = {0, 0};
+        // Write a placeholder integer to the cache.
+        // The driver should be able to handle non-empty cache and non-zero fd offset.
+        for (uint32_t i = 0; i < modelCache.size(); i++) {
+            ASSERT_EQ(write(modelCache[i].get(), &placeholderBytes, sizeof(placeholderBytes)),
+                      sizeof(placeholderBytes));
+        }
+        for (uint32_t i = 0; i < dataCache.size(); i++) {
+            ASSERT_EQ(write(dataCache[i].get(), &placeholderBytes, sizeof(placeholderBytes)),
+                      sizeof(placeholderBytes));
+        }
+        saveModelToCache(model, modelCache, dataCache);
+    }
+
+    // Retrieve preparedModel from cache.
+    {
+        preparedModel = nullptr;
+        ErrorStatus status;
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        uint8_t placeholderByte = 0;
+        // Advance the offset of each handle by one byte.
+        // The driver should be able to handle non-zero fd offset.
+        for (uint32_t i = 0; i < modelCache.size(); i++) {
+            ASSERT_GE(read(modelCache[i].get(), &placeholderByte, 1), 0);
+        }
+        for (uint32_t i = 0; i < dataCache.size(); i++) {
+            ASSERT_GE(read(dataCache[i].get(), &placeholderByte, 1), 0);
+        }
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (!mIsCachingSupported) {
+            ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+            ASSERT_EQ(preparedModel, nullptr);
+            return;
+        } else if (checkEarlyTermination(status)) {
+            ASSERT_EQ(preparedModel, nullptr);
+            return;
+        } else {
+            ASSERT_EQ(status, ErrorStatus::NONE);
+            ASSERT_NE(preparedModel, nullptr);
+        }
+    }
+
+    // Execute and verify results.
+    EvaluatePreparedModel(kDevice, preparedModel, testModel, /*testKind=*/TestKind::GENERAL);
+}
+
+TEST_P(CompilationCachingTest, SaveToCacheInvalidNumCache) {
+    // Create test HIDL model and compile.
+    const TestModel& testModel = createTestModel();
+    const Model model = createModel(testModel);
+    if (checkEarlyTermination(model)) return;
+
+    // Test with number of model cache files greater than mNumModelCache.
+    {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        // Pass an additional cache file for model cache.
+        mModelCache.push_back({mTmpCache});
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        mModelCache.pop_back();
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        saveModelToCache(model, modelCache, dataCache, &preparedModel);
+        ASSERT_NE(preparedModel, nullptr);
+        // Execute and verify results.
+        EvaluatePreparedModel(kDevice, preparedModel, testModel, /*testKind=*/TestKind::GENERAL);
+        // Check if prepareModelFromCache fails.
+        preparedModel = nullptr;
+        ErrorStatus status;
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (status != ErrorStatus::INVALID_ARGUMENT) {
+            ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+        }
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+
+    // Test with number of model cache files smaller than mNumModelCache.
+    if (mModelCache.size() > 0) {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        // Pop out the last cache file.
+        auto tmp = mModelCache.back();
+        mModelCache.pop_back();
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        mModelCache.push_back(tmp);
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        saveModelToCache(model, modelCache, dataCache, &preparedModel);
+        ASSERT_NE(preparedModel, nullptr);
+        // Execute and verify results.
+        EvaluatePreparedModel(kDevice, preparedModel, testModel, /*testKind=*/TestKind::GENERAL);
+        // Check if prepareModelFromCache fails.
+        preparedModel = nullptr;
+        ErrorStatus status;
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (status != ErrorStatus::INVALID_ARGUMENT) {
+            ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+        }
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+
+    // Test with number of data cache files greater than mNumDataCache.
+    {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        // Pass an additional cache file for data cache.
+        mDataCache.push_back({mTmpCache});
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        mDataCache.pop_back();
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        saveModelToCache(model, modelCache, dataCache, &preparedModel);
+        ASSERT_NE(preparedModel, nullptr);
+        // Execute and verify results.
+        EvaluatePreparedModel(kDevice, preparedModel, testModel, /*testKind=*/TestKind::GENERAL);
+        // Check if prepareModelFromCache fails.
+        preparedModel = nullptr;
+        ErrorStatus status;
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (status != ErrorStatus::INVALID_ARGUMENT) {
+            ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+        }
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+
+    // Test with number of data cache files smaller than mNumDataCache.
+    if (mDataCache.size() > 0) {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        // Pop out the last cache file.
+        auto tmp = mDataCache.back();
+        mDataCache.pop_back();
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        mDataCache.push_back(tmp);
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        saveModelToCache(model, modelCache, dataCache, &preparedModel);
+        ASSERT_NE(preparedModel, nullptr);
+        // Execute and verify results.
+        EvaluatePreparedModel(kDevice, preparedModel, testModel, /*testKind=*/TestKind::GENERAL);
+        // Check if prepareModelFromCache fails.
+        preparedModel = nullptr;
+        ErrorStatus status;
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (status != ErrorStatus::INVALID_ARGUMENT) {
+            ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+        }
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+}
+
+TEST_P(CompilationCachingTest, PrepareModelFromCacheInvalidNumCache) {
+    // Create test HIDL model and compile.
+    const TestModel& testModel = createTestModel();
+    const Model model = createModel(testModel);
+    if (checkEarlyTermination(model)) return;
+
+    // Save the compilation to cache.
+    {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        saveModelToCache(model, modelCache, dataCache);
+    }
+
+    // Test with number of model cache files greater than mNumModelCache.
+    {
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        ErrorStatus status;
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        mModelCache.push_back({mTmpCache});
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        mModelCache.pop_back();
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (status != ErrorStatus::GENERAL_FAILURE) {
+            ASSERT_EQ(status, ErrorStatus::INVALID_ARGUMENT);
+        }
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+
+    // Test with number of model cache files smaller than mNumModelCache.
+    if (mModelCache.size() > 0) {
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        ErrorStatus status;
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        auto tmp = mModelCache.back();
+        mModelCache.pop_back();
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        mModelCache.push_back(tmp);
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (status != ErrorStatus::GENERAL_FAILURE) {
+            ASSERT_EQ(status, ErrorStatus::INVALID_ARGUMENT);
+        }
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+
+    // Test with number of data cache files greater than mNumDataCache.
+    {
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        ErrorStatus status;
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        mDataCache.push_back({mTmpCache});
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        mDataCache.pop_back();
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (status != ErrorStatus::GENERAL_FAILURE) {
+            ASSERT_EQ(status, ErrorStatus::INVALID_ARGUMENT);
+        }
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+
+    // Test with number of data cache files smaller than mNumDataCache.
+    if (mDataCache.size() > 0) {
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        ErrorStatus status;
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        auto tmp = mDataCache.back();
+        mDataCache.pop_back();
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        mDataCache.push_back(tmp);
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (status != ErrorStatus::GENERAL_FAILURE) {
+            ASSERT_EQ(status, ErrorStatus::INVALID_ARGUMENT);
+        }
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+}
+
+TEST_P(CompilationCachingTest, SaveToCacheInvalidAccessMode) {
+    // Create test HIDL model and compile.
+    const TestModel& testModel = createTestModel();
+    const Model model = createModel(testModel);
+    if (checkEarlyTermination(model)) return;
+    std::vector<AccessMode> modelCacheMode(mNumModelCache, AccessMode::READ_WRITE);
+    std::vector<AccessMode> dataCacheMode(mNumDataCache, AccessMode::READ_WRITE);
+
+    // Go through each handle in model cache, test with invalid access mode.
+    for (uint32_t i = 0; i < mNumModelCache; i++) {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        modelCacheMode[i] = AccessMode::READ_ONLY;
+        createCacheFds(mModelCache, modelCacheMode, &modelCache);
+        createCacheFds(mDataCache, dataCacheMode, &dataCache);
+        modelCacheMode[i] = AccessMode::READ_WRITE;
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        saveModelToCache(model, modelCache, dataCache, &preparedModel);
+        ASSERT_NE(preparedModel, nullptr);
+        // Execute and verify results.
+        EvaluatePreparedModel(kDevice, preparedModel, testModel, /*testKind=*/TestKind::GENERAL);
+        // Check if prepareModelFromCache fails.
+        preparedModel = nullptr;
+        ErrorStatus status;
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (status != ErrorStatus::INVALID_ARGUMENT) {
+            ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+        }
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+
+    // Go through each handle in data cache, test with invalid access mode.
+    for (uint32_t i = 0; i < mNumDataCache; i++) {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        dataCacheMode[i] = AccessMode::READ_ONLY;
+        createCacheFds(mModelCache, modelCacheMode, &modelCache);
+        createCacheFds(mDataCache, dataCacheMode, &dataCache);
+        dataCacheMode[i] = AccessMode::READ_WRITE;
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        saveModelToCache(model, modelCache, dataCache, &preparedModel);
+        ASSERT_NE(preparedModel, nullptr);
+        // Execute and verify results.
+        EvaluatePreparedModel(kDevice, preparedModel, testModel, /*testKind=*/TestKind::GENERAL);
+        // Check if prepareModelFromCache fails.
+        preparedModel = nullptr;
+        ErrorStatus status;
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        if (status != ErrorStatus::INVALID_ARGUMENT) {
+            ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+        }
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+}
+
+TEST_P(CompilationCachingTest, PrepareModelFromCacheInvalidAccessMode) {
+    // Create test HIDL model and compile.
+    const TestModel& testModel = createTestModel();
+    const Model model = createModel(testModel);
+    if (checkEarlyTermination(model)) return;
+    std::vector<AccessMode> modelCacheMode(mNumModelCache, AccessMode::READ_WRITE);
+    std::vector<AccessMode> dataCacheMode(mNumDataCache, AccessMode::READ_WRITE);
+
+    // Save the compilation to cache.
+    {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        saveModelToCache(model, modelCache, dataCache);
+    }
+
+    // Go through each handle in model cache, test with invalid access mode.
+    for (uint32_t i = 0; i < mNumModelCache; i++) {
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        ErrorStatus status;
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        modelCacheMode[i] = AccessMode::WRITE_ONLY;
+        createCacheFds(mModelCache, modelCacheMode, &modelCache);
+        createCacheFds(mDataCache, dataCacheMode, &dataCache);
+        modelCacheMode[i] = AccessMode::READ_WRITE;
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+
+    // Go through each handle in data cache, test with invalid access mode.
+    for (uint32_t i = 0; i < mNumDataCache; i++) {
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        ErrorStatus status;
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        dataCacheMode[i] = AccessMode::WRITE_ONLY;
+        createCacheFds(mModelCache, modelCacheMode, &modelCache);
+        createCacheFds(mDataCache, dataCacheMode, &dataCache);
+        dataCacheMode[i] = AccessMode::READ_WRITE;
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+}
+
+// Copy file contents between files.
+// The vector sizes must match.
+static void copyCacheFiles(const std::vector<std::string>& from,
+                           const std::vector<std::string>& to) {
+    constexpr size_t kBufferSize = 1000000;
+    uint8_t buffer[kBufferSize];
+
+    ASSERT_EQ(from.size(), to.size());
+    for (uint32_t i = 0; i < from.size(); i++) {
+        int fromFd = open(from[i].c_str(), O_RDONLY);
+        int toFd = open(to[i].c_str(), O_WRONLY | O_CREAT, S_IRUSR | S_IWUSR);
+        ASSERT_GE(fromFd, 0);
+        ASSERT_GE(toFd, 0);
+
+        ssize_t readBytes;
+        while ((readBytes = read(fromFd, &buffer, kBufferSize)) > 0) {
+            ASSERT_EQ(write(toFd, &buffer, readBytes), readBytes);
+        }
+        ASSERT_GE(readBytes, 0);
+
+        close(fromFd);
+        close(toFd);
+    }
+}
+
+// Number of operations in the large test model.
+constexpr uint32_t kLargeModelSize = 100;
+constexpr uint32_t kNumIterationsTOCTOU = 100;
+
+TEST_P(CompilationCachingTest, SaveToCache_TOCTOU) {
+    if (!mIsCachingSupported) return;
+
+    // Create test models and check if fully supported by the service.
+    const TestModel testModelMul = createLargeTestModel(OperationType::MUL, kLargeModelSize);
+    const Model modelMul = createModel(testModelMul);
+    if (checkEarlyTermination(modelMul)) return;
+    const TestModel testModelAdd = createLargeTestModel(OperationType::ADD, kLargeModelSize);
+    const Model modelAdd = createModel(testModelAdd);
+    if (checkEarlyTermination(modelAdd)) return;
+
+    // Save the modelMul compilation to cache.
+    auto modelCacheMul = mModelCache;
+    for (auto& cache : modelCacheMul) {
+        cache.append("_mul");
+    }
+    {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        createCacheFds(modelCacheMul, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        saveModelToCache(modelMul, modelCache, dataCache);
+    }
+
+    // Use a different token for modelAdd.
+    mToken[0]++;
+
+    // This test is probabilistic, so we run it multiple times.
+    for (uint32_t i = 0; i < kNumIterationsTOCTOU; i++) {
+        // Save the modelAdd compilation to cache.
+        {
+            std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+            createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+            createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+
+            // Spawn a thread to copy the cache content concurrently while saving to cache.
+            std::thread thread(copyCacheFiles, std::cref(modelCacheMul), std::cref(mModelCache));
+            saveModelToCache(modelAdd, modelCache, dataCache);
+            thread.join();
+        }
+
+        // Retrieve preparedModel from cache.
+        {
+            std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+            ErrorStatus status;
+            std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+            createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+            createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+            prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+
+            // The preparation may fail or succeed, but must not crash. If the preparation succeeds,
+            // the prepared model must be executed with the correct result and not crash.
+            if (status != ErrorStatus::NONE) {
+                ASSERT_EQ(preparedModel, nullptr);
+            } else {
+                ASSERT_NE(preparedModel, nullptr);
+                EvaluatePreparedModel(kDevice, preparedModel, testModelAdd,
+                                      /*testKind=*/TestKind::GENERAL);
+            }
+        }
+    }
+}
+
+TEST_P(CompilationCachingTest, PrepareFromCache_TOCTOU) {
+    if (!mIsCachingSupported) return;
+
+    // Create test models and check if fully supported by the service.
+    const TestModel testModelMul = createLargeTestModel(OperationType::MUL, kLargeModelSize);
+    const Model modelMul = createModel(testModelMul);
+    if (checkEarlyTermination(modelMul)) return;
+    const TestModel testModelAdd = createLargeTestModel(OperationType::ADD, kLargeModelSize);
+    const Model modelAdd = createModel(testModelAdd);
+    if (checkEarlyTermination(modelAdd)) return;
+
+    // Save the modelMul compilation to cache.
+    auto modelCacheMul = mModelCache;
+    for (auto& cache : modelCacheMul) {
+        cache.append("_mul");
+    }
+    {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        createCacheFds(modelCacheMul, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        saveModelToCache(modelMul, modelCache, dataCache);
+    }
+
+    // Use a different token for modelAdd.
+    mToken[0]++;
+
+    // This test is probabilistic, so we run it multiple times.
+    for (uint32_t i = 0; i < kNumIterationsTOCTOU; i++) {
+        // Save the modelAdd compilation to cache.
+        {
+            std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+            createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+            createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+            saveModelToCache(modelAdd, modelCache, dataCache);
+        }
+
+        // Retrieve preparedModel from cache.
+        {
+            std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+            ErrorStatus status;
+            std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+            createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+            createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+
+            // Spawn a thread to copy the cache content concurrently while preparing from cache.
+            std::thread thread(copyCacheFiles, std::cref(modelCacheMul), std::cref(mModelCache));
+            prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+            thread.join();
+
+            // The preparation may fail or succeed, but must not crash. If the preparation succeeds,
+            // the prepared model must be executed with the correct result and not crash.
+            if (status != ErrorStatus::NONE) {
+                ASSERT_EQ(preparedModel, nullptr);
+            } else {
+                ASSERT_NE(preparedModel, nullptr);
+                EvaluatePreparedModel(kDevice, preparedModel, testModelAdd,
+                                      /*testKind=*/TestKind::GENERAL);
+            }
+        }
+    }
+}
+
+TEST_P(CompilationCachingTest, ReplaceSecuritySensitiveCache) {
+    if (!mIsCachingSupported) return;
+
+    // Create test models and check if fully supported by the service.
+    const TestModel testModelMul = createLargeTestModel(OperationType::MUL, kLargeModelSize);
+    const Model modelMul = createModel(testModelMul);
+    if (checkEarlyTermination(modelMul)) return;
+    const TestModel testModelAdd = createLargeTestModel(OperationType::ADD, kLargeModelSize);
+    const Model modelAdd = createModel(testModelAdd);
+    if (checkEarlyTermination(modelAdd)) return;
+
+    // Save the modelMul compilation to cache.
+    auto modelCacheMul = mModelCache;
+    for (auto& cache : modelCacheMul) {
+        cache.append("_mul");
+    }
+    {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        createCacheFds(modelCacheMul, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        saveModelToCache(modelMul, modelCache, dataCache);
+    }
+
+    // Use a different token for modelAdd.
+    mToken[0]++;
+
+    // Save the modelAdd compilation to cache.
+    {
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        saveModelToCache(modelAdd, modelCache, dataCache);
+    }
+
+    // Replace the model cache of modelAdd with modelMul.
+    copyCacheFiles(modelCacheMul, mModelCache);
+
+    // Retrieve the preparedModel from cache, expect failure.
+    {
+        std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+        ErrorStatus status;
+        std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+        createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+        createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+        prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+        ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+        ASSERT_EQ(preparedModel, nullptr);
+    }
+}
+
+// TODO(b/179270601): restore kNamedDeviceChoices.
+static const auto kOperandTypeChoices =
+        testing::Values(OperandType::TENSOR_FLOAT32, OperandType::TENSOR_QUANT8_ASYMM);
+
+std::string printCompilationCachingTest(
+        const testing::TestParamInfo<CompilationCachingTestParam>& info) {
+    const auto& [namedDevice, operandType] = info.param;
+    const std::string type = (operandType == OperandType::TENSOR_FLOAT32 ? "float32" : "quant8");
+    return gtestCompliantName(getName(namedDevice) + "_" + type);
+}
+
+GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CompilationCachingTest);
+INSTANTIATE_TEST_SUITE_P(TestCompilationCaching, CompilationCachingTest,
+                         testing::Combine(testing::ValuesIn(getNamedDevices()),
+                                          kOperandTypeChoices),
+                         printCompilationCachingTest);
+
+using CompilationCachingSecurityTestParam = std::tuple<NamedDevice, OperandType, uint32_t>;
+
+class CompilationCachingSecurityTest
+    : public CompilationCachingTestBase,
+      public testing::WithParamInterface<CompilationCachingSecurityTestParam> {
+  protected:
+    CompilationCachingSecurityTest()
+        : CompilationCachingTestBase(getData(std::get<NamedDevice>(GetParam())),
+                                     std::get<OperandType>(GetParam())) {}
+
+    void SetUp() {
+        CompilationCachingTestBase::SetUp();
+        generator.seed(kSeed);
+    }
+
+    // Get a random integer within a closed range [lower, upper].
+    template <typename T>
+    T getRandomInt(T lower, T upper) {
+        std::uniform_int_distribution<T> dis(lower, upper);
+        return dis(generator);
+    }
+
+    // Randomly flip one single bit of the cache entry.
+    void flipOneBitOfCache(const std::string& filename, bool* skip) {
+        FILE* pFile = fopen(filename.c_str(), "r+");
+        ASSERT_EQ(fseek(pFile, 0, SEEK_END), 0);
+        long int fileSize = ftell(pFile);
+        if (fileSize == 0) {
+            fclose(pFile);
+            *skip = true;
+            return;
+        }
+        ASSERT_EQ(fseek(pFile, getRandomInt(0l, fileSize - 1), SEEK_SET), 0);
+        int readByte = fgetc(pFile);
+        ASSERT_NE(readByte, EOF);
+        ASSERT_EQ(fseek(pFile, -1, SEEK_CUR), 0);
+        ASSERT_NE(fputc(static_cast<uint8_t>(readByte) ^ (1U << getRandomInt(0, 7)), pFile), EOF);
+        fclose(pFile);
+        *skip = false;
+    }
+
+    // Randomly append bytes to the cache entry.
+    void appendBytesToCache(const std::string& filename, bool* skip) {
+        FILE* pFile = fopen(filename.c_str(), "a");
+        uint32_t appendLength = getRandomInt(1, 256);
+        for (uint32_t i = 0; i < appendLength; i++) {
+            ASSERT_NE(fputc(getRandomInt<uint8_t>(0, 255), pFile), EOF);
+        }
+        fclose(pFile);
+        *skip = false;
+    }
+
+    enum class ExpectedResult { GENERAL_FAILURE, NOT_CRASH };
+
+    // Test if the driver behaves as expected when given corrupted cache or token.
+    // The modifier will be invoked after save to cache but before prepare from cache.
+    // The modifier accepts one pointer argument "skip" as the returning value, indicating
+    // whether the test should be skipped or not.
+    void testCorruptedCache(ExpectedResult expected, std::function<void(bool*)> modifier) {
+        const TestModel& testModel = createTestModel();
+        const Model model = createModel(testModel);
+        if (checkEarlyTermination(model)) return;
+
+        // Save the compilation to cache.
+        {
+            std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+            createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+            createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+            saveModelToCache(model, modelCache, dataCache);
+        }
+
+        bool skip = false;
+        modifier(&skip);
+        if (skip) return;
+
+        // Retrieve preparedModel from cache.
+        {
+            std::shared_ptr<IPreparedModel> preparedModel = nullptr;
+            ErrorStatus status;
+            std::vector<ndk::ScopedFileDescriptor> modelCache, dataCache;
+            createCacheFds(mModelCache, AccessMode::READ_WRITE, &modelCache);
+            createCacheFds(mDataCache, AccessMode::READ_WRITE, &dataCache);
+            prepareModelFromCache(modelCache, dataCache, &preparedModel, &status);
+
+            switch (expected) {
+                case ExpectedResult::GENERAL_FAILURE:
+                    ASSERT_EQ(status, ErrorStatus::GENERAL_FAILURE);
+                    ASSERT_EQ(preparedModel, nullptr);
+                    break;
+                case ExpectedResult::NOT_CRASH:
+                    ASSERT_EQ(preparedModel == nullptr, status != ErrorStatus::NONE);
+                    break;
+                default:
+                    FAIL();
+            }
+        }
+    }
+
+    const uint32_t kSeed = std::get<uint32_t>(GetParam());
+    std::mt19937 generator;
+};
+
+TEST_P(CompilationCachingSecurityTest, CorruptedModelCache) {
+    if (!mIsCachingSupported) return;
+    for (uint32_t i = 0; i < mNumModelCache; i++) {
+        testCorruptedCache(ExpectedResult::GENERAL_FAILURE,
+                           [this, i](bool* skip) { flipOneBitOfCache(mModelCache[i], skip); });
+    }
+}
+
+TEST_P(CompilationCachingSecurityTest, WrongLengthModelCache) {
+    if (!mIsCachingSupported) return;
+    for (uint32_t i = 0; i < mNumModelCache; i++) {
+        testCorruptedCache(ExpectedResult::GENERAL_FAILURE,
+                           [this, i](bool* skip) { appendBytesToCache(mModelCache[i], skip); });
+    }
+}
+
+TEST_P(CompilationCachingSecurityTest, CorruptedDataCache) {
+    if (!mIsCachingSupported) return;
+    for (uint32_t i = 0; i < mNumDataCache; i++) {
+        testCorruptedCache(ExpectedResult::NOT_CRASH,
+                           [this, i](bool* skip) { flipOneBitOfCache(mDataCache[i], skip); });
+    }
+}
+
+TEST_P(CompilationCachingSecurityTest, WrongLengthDataCache) {
+    if (!mIsCachingSupported) return;
+    for (uint32_t i = 0; i < mNumDataCache; i++) {
+        testCorruptedCache(ExpectedResult::NOT_CRASH,
+                           [this, i](bool* skip) { appendBytesToCache(mDataCache[i], skip); });
+    }
+}
+
+TEST_P(CompilationCachingSecurityTest, WrongToken) {
+    if (!mIsCachingSupported) return;
+    testCorruptedCache(ExpectedResult::GENERAL_FAILURE, [this](bool* skip) {
+        // Randomly flip one single bit in mToken.
+        uint32_t ind =
+                getRandomInt(0u, static_cast<uint32_t>(IDevice::BYTE_SIZE_OF_CACHE_TOKEN) - 1);
+        mToken[ind] ^= (1U << getRandomInt(0, 7));
+        *skip = false;
+    });
+}
+
+std::string printCompilationCachingSecurityTest(
+        const testing::TestParamInfo<CompilationCachingSecurityTestParam>& info) {
+    const auto& [namedDevice, operandType, seed] = info.param;
+    const std::string type = (operandType == OperandType::TENSOR_FLOAT32 ? "float32" : "quant8");
+    return gtestCompliantName(getName(namedDevice) + "_" + type + "_" + std::to_string(seed));
+}
+
+GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CompilationCachingSecurityTest);
+INSTANTIATE_TEST_SUITE_P(TestCompilationCaching, CompilationCachingSecurityTest,
+                         testing::Combine(testing::ValuesIn(getNamedDevices()), kOperandTypeChoices,
+                                          testing::Range(0U, 10U)),
+                         printCompilationCachingSecurityTest);
+
+}  // namespace aidl::android::hardware::neuralnetworks::vts::functional
diff --git a/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.cpp
new file mode 100644
index 0000000..86d5f3f
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.cpp
@@ -0,0 +1,925 @@
+/*
+ * 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 "GeneratedTestHarness.h"
+
+#include <aidl/android/hardware/neuralnetworks/ErrorStatus.h>
+#include <android-base/logging.h>
+#include <android/binder_auto_utils.h>
+#include <android/sync.h>
+#include <gtest/gtest.h>
+
+#include <algorithm>
+#include <chrono>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include <MemoryUtils.h>
+#include <android/binder_status.h>
+#include <nnapi/Result.h>
+#include <nnapi/SharedMemory.h>
+#include <nnapi/Types.h>
+#include <nnapi/hal/aidl/Conversions.h>
+#include <nnapi/hal/aidl/Utils.h>
+
+#include "Callbacks.h"
+#include "TestHarness.h"
+#include "Utils.h"
+#include "VtsHalNeuralnetworks.h"
+
+namespace aidl::android::hardware::neuralnetworks::vts::functional {
+
+namespace nn = ::android::nn;
+using namespace test_helper;
+using implementation::PreparedModelCallback;
+
+namespace {
+
+enum class OutputType { FULLY_SPECIFIED, UNSPECIFIED, INSUFFICIENT, MISSED_DEADLINE };
+
+struct TestConfig {
+    Executor executor;
+    bool measureTiming;
+    OutputType outputType;
+    MemoryType memoryType;
+    // `reportSkipping` indicates if a test should print an info message in case
+    // it is skipped. The field is set to true by default and is set to false in
+    // quantization coupling tests to suppress skipping a test
+    bool reportSkipping;
+    TestConfig(Executor executor, bool measureTiming, OutputType outputType, MemoryType memoryType)
+        : executor(executor),
+          measureTiming(measureTiming),
+          outputType(outputType),
+          memoryType(memoryType),
+          reportSkipping(true) {}
+    TestConfig(Executor executor, bool measureTiming, OutputType outputType, MemoryType memoryType,
+               bool reportSkipping)
+        : executor(executor),
+          measureTiming(measureTiming),
+          outputType(outputType),
+          memoryType(memoryType),
+          reportSkipping(reportSkipping) {}
+};
+
+enum class IOType { INPUT, OUTPUT };
+
+class DeviceMemoryAllocator {
+  public:
+    DeviceMemoryAllocator(const std::shared_ptr<IDevice>& device,
+                          const std::shared_ptr<IPreparedModel>& preparedModel,
+                          const TestModel& testModel)
+        : kDevice(device), kPreparedModel(preparedModel), kTestModel(testModel) {}
+
+    // Allocate device memory for a target input/output operand.
+    // Return {IBuffer object, token} if successful.
+    // Return {nullptr, 0} if device memory is not supported.
+    template <IOType ioType>
+    std::pair<std::shared_ptr<IBuffer>, int32_t> allocate(uint32_t index) {
+        std::pair<std::shared_ptr<IBuffer>, int32_t> buffer;
+        allocateInternal<ioType>(index, &buffer);
+        return buffer;
+    }
+
+  private:
+    template <IOType ioType>
+    void allocateInternal(int32_t index, std::pair<std::shared_ptr<IBuffer>, int32_t>* result) {
+        ASSERT_NE(result, nullptr);
+
+        // Prepare arguments.
+        BufferRole role = {.modelIndex = 0, .ioIndex = index, .frequency = 1.0f};
+        std::vector<BufferRole> inputRoles, outputRoles;
+        if constexpr (ioType == IOType::INPUT) {
+            inputRoles = {role};
+        } else {
+            outputRoles = {role};
+        }
+
+        // Allocate device memory.
+        DeviceBuffer buffer;
+        IPreparedModelParcel parcel;
+        parcel.preparedModel = kPreparedModel;
+        const auto ret = kDevice->allocate({}, {parcel}, inputRoles, outputRoles, &buffer);
+
+        // Check allocation results.
+        if (ret.isOk()) {
+            ASSERT_NE(buffer.buffer, nullptr);
+            ASSERT_GT(buffer.token, 0);
+        } else {
+            ASSERT_EQ(ret.getExceptionCode(), EX_SERVICE_SPECIFIC);
+            ASSERT_EQ(static_cast<ErrorStatus>(ret.getServiceSpecificError()),
+                      ErrorStatus::GENERAL_FAILURE);
+            buffer.buffer = nullptr;
+            buffer.token = 0;
+        }
+
+        // Initialize input data from TestBuffer.
+        if constexpr (ioType == IOType::INPUT) {
+            if (buffer.buffer != nullptr) {
+                // TestBuffer -> Shared memory.
+                const auto& testBuffer =
+                        kTestModel.main.operands[kTestModel.main.inputIndexes[index]].data;
+                ASSERT_GT(testBuffer.size(), 0);
+                const auto sharedMemory = nn::createSharedMemory(testBuffer.size()).value();
+                const auto memory = utils::convert(sharedMemory).value();
+                const auto mapping = nn::map(sharedMemory).value();
+                uint8_t* inputPtr = static_cast<uint8_t*>(std::get<void*>(mapping.pointer));
+                ASSERT_NE(inputPtr, nullptr);
+                const uint8_t* begin = testBuffer.get<uint8_t>();
+                const uint8_t* end = begin + testBuffer.size();
+                std::copy(begin, end, inputPtr);
+
+                // Shared memory -> IBuffer.
+                auto ret = buffer.buffer->copyFrom(memory, {});
+                ASSERT_TRUE(ret.isOk());
+            }
+        }
+        *result = {std::move(buffer.buffer), buffer.token};
+    }
+
+    const std::shared_ptr<IDevice> kDevice;
+    const std::shared_ptr<IPreparedModel> kPreparedModel;
+    const TestModel& kTestModel;
+};
+
+Subgraph createSubgraph(const TestSubgraph& testSubgraph, uint32_t* constCopySize,
+                        std::vector<const TestBuffer*>* constCopies, uint32_t* constRefSize,
+                        std::vector<const TestBuffer*>* constReferences) {
+    CHECK(constCopySize != nullptr);
+    CHECK(constCopies != nullptr);
+    CHECK(constRefSize != nullptr);
+    CHECK(constReferences != nullptr);
+
+    // Operands.
+    std::vector<Operand> operands(testSubgraph.operands.size());
+    for (uint32_t i = 0; i < testSubgraph.operands.size(); i++) {
+        const auto& op = testSubgraph.operands[i];
+
+        DataLocation loc = {};
+        if (op.lifetime == TestOperandLifeTime::CONSTANT_COPY) {
+            loc = {
+                    .poolIndex = 0,
+                    .offset = *constCopySize,
+                    .length = static_cast<int64_t>(op.data.size()),
+            };
+            constCopies->push_back(&op.data);
+            *constCopySize += op.data.alignedSize();
+        } else if (op.lifetime == TestOperandLifeTime::CONSTANT_REFERENCE) {
+            loc = {
+                    .poolIndex = 0,
+                    .offset = *constRefSize,
+                    .length = static_cast<int64_t>(op.data.size()),
+            };
+            constReferences->push_back(&op.data);
+            *constRefSize += op.data.alignedSize();
+        } else if (op.lifetime == TestOperandLifeTime::SUBGRAPH) {
+            loc = {
+                    .poolIndex = 0,
+                    .offset = *op.data.get<uint32_t>(),
+                    .length = 0,
+            };
+        }
+
+        std::optional<OperandExtraParams> extraParams;
+        if (op.type == TestOperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
+            using Tag = OperandExtraParams::Tag;
+            extraParams = OperandExtraParams::make<Tag::channelQuant>(SymmPerChannelQuantParams{
+                    .scales = op.channelQuant.scales,
+                    .channelDim = static_cast<int32_t>(op.channelQuant.channelDim)});
+        }
+
+        operands[i] = {.type = static_cast<OperandType>(op.type),
+                       .dimensions = utils::toSigned(op.dimensions).value(),
+                       .scale = op.scale,
+                       .zeroPoint = op.zeroPoint,
+                       .lifetime = static_cast<OperandLifeTime>(op.lifetime),
+                       .location = loc,
+                       .extraParams = std::move(extraParams)};
+    }
+
+    // Operations.
+    std::vector<Operation> operations(testSubgraph.operations.size());
+    std::transform(testSubgraph.operations.begin(), testSubgraph.operations.end(),
+                   operations.begin(), [](const TestOperation& op) -> Operation {
+                       return {.type = static_cast<OperationType>(op.type),
+                               .inputs = utils::toSigned(op.inputs).value(),
+                               .outputs = utils::toSigned(op.outputs).value()};
+                   });
+
+    return {.operands = std::move(operands),
+            .operations = std::move(operations),
+            .inputIndexes = utils::toSigned(testSubgraph.inputIndexes).value(),
+            .outputIndexes = utils::toSigned(testSubgraph.outputIndexes).value()};
+}
+
+void copyTestBuffers(const std::vector<const TestBuffer*>& buffers, uint8_t* output) {
+    uint32_t offset = 0;
+    for (const TestBuffer* buffer : buffers) {
+        const uint8_t* begin = buffer->get<uint8_t>();
+        const uint8_t* end = begin + buffer->size();
+        std::copy(begin, end, output + offset);
+        offset += buffer->alignedSize();
+    }
+}
+
+}  // namespace
+
+void waitForSyncFence(int syncFd) {
+    constexpr int kInfiniteTimeout = -1;
+    ASSERT_GT(syncFd, 0);
+    int r = sync_wait(syncFd, kInfiniteTimeout);
+    ASSERT_GE(r, 0);
+}
+
+Model createModel(const TestModel& testModel) {
+    uint32_t constCopySize = 0;
+    uint32_t constRefSize = 0;
+    std::vector<const TestBuffer*> constCopies;
+    std::vector<const TestBuffer*> constReferences;
+
+    Subgraph mainSubgraph = createSubgraph(testModel.main, &constCopySize, &constCopies,
+                                           &constRefSize, &constReferences);
+    std::vector<Subgraph> refSubgraphs(testModel.referenced.size());
+    std::transform(testModel.referenced.begin(), testModel.referenced.end(), refSubgraphs.begin(),
+                   [&constCopySize, &constCopies, &constRefSize,
+                    &constReferences](const TestSubgraph& testSubgraph) {
+                       return createSubgraph(testSubgraph, &constCopySize, &constCopies,
+                                             &constRefSize, &constReferences);
+                   });
+
+    // Constant copies.
+    std::vector<uint8_t> operandValues(constCopySize);
+    copyTestBuffers(constCopies, operandValues.data());
+
+    // Shared memory.
+    std::vector<nn::Memory> pools = {};
+    if (constRefSize > 0) {
+        const auto pool = nn::createSharedMemory(constRefSize).value();
+        pools.push_back(pool);
+
+        // load data
+        const auto mappedMemory = nn::map(pool).value();
+        uint8_t* mappedPtr = static_cast<uint8_t*>(std::get<void*>(mappedMemory.pointer));
+        CHECK(mappedPtr != nullptr);
+
+        copyTestBuffers(constReferences, mappedPtr);
+    }
+
+    std::vector<Memory> aidlPools;
+    aidlPools.reserve(pools.size());
+    for (auto& pool : pools) {
+        auto aidlPool = utils::convert(pool).value();
+        aidlPools.push_back(std::move(aidlPool));
+    }
+
+    return {.main = std::move(mainSubgraph),
+            .referenced = std::move(refSubgraphs),
+            .operandValues = std::move(operandValues),
+            .pools = std::move(aidlPools),
+            .relaxComputationFloat32toFloat16 = testModel.isRelaxed};
+}
+
+static bool isOutputSizeGreaterThanOne(const TestModel& testModel, uint32_t index) {
+    const auto byteSize = testModel.main.operands[testModel.main.outputIndexes[index]].data.size();
+    return byteSize > 1u;
+}
+
+static void makeOutputInsufficientSize(uint32_t outputIndex, Request* request) {
+    auto& length = request->outputs[outputIndex].location.length;
+    ASSERT_GT(length, 1u);
+    length -= 1u;
+}
+
+static void makeOutputDimensionsUnspecified(Model* model) {
+    for (auto i : model->main.outputIndexes) {
+        auto& dims = model->main.operands[i].dimensions;
+        std::fill(dims.begin(), dims.end(), 0);
+    }
+}
+
+// Manages the lifetime of memory resources used in an execution.
+class ExecutionContext {
+  public:
+    ExecutionContext(std::shared_ptr<IDevice> device, std::shared_ptr<IPreparedModel> preparedModel)
+        : kDevice(std::move(device)), kPreparedModel(std::move(preparedModel)) {}
+
+    std::optional<Request> createRequest(const TestModel& testModel, MemoryType memoryType);
+    std::vector<TestBuffer> getOutputBuffers(const TestModel& testModel,
+                                             const Request& request) const;
+
+  private:
+    // Get a TestBuffer with data copied from an IBuffer object.
+    void getBuffer(const std::shared_ptr<IBuffer>& buffer, size_t size,
+                   TestBuffer* testBuffer) const;
+
+    static constexpr uint32_t kInputPoolIndex = 0;
+    static constexpr uint32_t kOutputPoolIndex = 1;
+    static constexpr uint32_t kDeviceMemoryBeginIndex = 2;
+
+    const std::shared_ptr<IDevice> kDevice;
+    const std::shared_ptr<IPreparedModel> kPreparedModel;
+    std::unique_ptr<TestMemoryBase> mInputMemory, mOutputMemory;
+    std::vector<std::shared_ptr<IBuffer>> mBuffers;
+};
+
+std::optional<Request> ExecutionContext::createRequest(const TestModel& testModel,
+                                                       MemoryType memoryType) {
+    // Memory pools are organized as:
+    // - 0: Input shared memory pool
+    // - 1: Output shared memory pool
+    // - [2, 2+i): Input device memories
+    // - [2+i, 2+i+o): Output device memories
+    DeviceMemoryAllocator allocator(kDevice, kPreparedModel, testModel);
+    std::vector<int32_t> tokens;
+    mBuffers.clear();
+
+    // Model inputs.
+    std::vector<RequestArgument> inputs(testModel.main.inputIndexes.size());
+    size_t inputSize = 0;
+    for (uint32_t i = 0; i < testModel.main.inputIndexes.size(); i++) {
+        const auto& op = testModel.main.operands[testModel.main.inputIndexes[i]];
+        if (op.data.size() == 0) {
+            // Omitted input.
+            inputs[i] = {.hasNoValue = true};
+            continue;
+        } else if (memoryType == MemoryType::DEVICE) {
+            SCOPED_TRACE("Input index = " + std::to_string(i));
+            auto [buffer, token] = allocator.allocate<IOType::INPUT>(i);
+            if (buffer != nullptr) {
+                DataLocation loc = {.poolIndex = static_cast<int32_t>(mBuffers.size() +
+                                                                      kDeviceMemoryBeginIndex)};
+                mBuffers.push_back(std::move(buffer));
+                tokens.push_back(token);
+                inputs[i] = {.hasNoValue = false, .location = loc, .dimensions = {}};
+                continue;
+            }
+        }
+
+        // Reserve shared memory for input.
+        DataLocation loc = {.poolIndex = kInputPoolIndex,
+                            .offset = static_cast<int64_t>(inputSize),
+                            .length = static_cast<int64_t>(op.data.size())};
+        inputSize += op.data.alignedSize();
+        inputs[i] = {.hasNoValue = false, .location = loc, .dimensions = {}};
+    }
+
+    // Model outputs.
+    std::vector<RequestArgument> outputs(testModel.main.outputIndexes.size());
+    size_t outputSize = 0;
+    for (uint32_t i = 0; i < testModel.main.outputIndexes.size(); i++) {
+        const auto& op = testModel.main.operands[testModel.main.outputIndexes[i]];
+        if (memoryType == MemoryType::DEVICE) {
+            SCOPED_TRACE("Output index = " + std::to_string(i));
+            auto [buffer, token] = allocator.allocate<IOType::OUTPUT>(i);
+            if (buffer != nullptr) {
+                DataLocation loc = {.poolIndex = static_cast<int32_t>(mBuffers.size() +
+                                                                      kDeviceMemoryBeginIndex)};
+                mBuffers.push_back(std::move(buffer));
+                tokens.push_back(token);
+                outputs[i] = {.hasNoValue = false, .location = loc, .dimensions = {}};
+                continue;
+            }
+        }
+
+        // In the case of zero-sized output, we should at least provide a one-byte buffer.
+        // This is because zero-sized tensors are only supported internally to the driver, or
+        // reported in output shapes. It is illegal for the client to pre-specify a zero-sized
+        // tensor as model output. Otherwise, we will have two semantic conflicts:
+        // - "Zero dimension" conflicts with "unspecified dimension".
+        // - "Omitted operand buffer" conflicts with "zero-sized operand buffer".
+        size_t bufferSize = std::max<size_t>(op.data.size(), 1);
+
+        // Reserve shared memory for output.
+        DataLocation loc = {.poolIndex = kOutputPoolIndex,
+                            .offset = static_cast<int64_t>(outputSize),
+                            .length = static_cast<int64_t>(bufferSize)};
+        outputSize += op.data.size() == 0 ? TestBuffer::kAlignment : op.data.alignedSize();
+        outputs[i] = {.hasNoValue = false, .location = loc, .dimensions = {}};
+    }
+
+    if (memoryType == MemoryType::DEVICE && mBuffers.empty()) {
+        return std::nullopt;
+    }
+
+    // Memory pools.
+    if (memoryType == MemoryType::BLOB_AHWB) {
+        mInputMemory = TestBlobAHWB::create(std::max<size_t>(inputSize, 1));
+        mOutputMemory = TestBlobAHWB::create(std::max<size_t>(outputSize, 1));
+    } else {
+        mInputMemory = TestAshmem::create(std::max<size_t>(inputSize, 1));
+        mOutputMemory = TestAshmem::create(std::max<size_t>(outputSize, 1));
+    }
+    CHECK_NE(mInputMemory, nullptr);
+    CHECK_NE(mOutputMemory, nullptr);
+    std::vector<RequestMemoryPool> pools;
+    pools.reserve(kDeviceMemoryBeginIndex + mBuffers.size());
+
+    auto copiedInputMemory = utils::clone(*mInputMemory->getAidlMemory());
+    CHECK(copiedInputMemory.has_value()) << copiedInputMemory.error().message;
+    auto copiedOutputMemory = utils::clone(*mOutputMemory->getAidlMemory());
+    CHECK(copiedOutputMemory.has_value()) << copiedOutputMemory.error().message;
+
+    pools.push_back(RequestMemoryPool::make<RequestMemoryPool::Tag::pool>(
+            std::move(copiedInputMemory).value()));
+    pools.push_back(RequestMemoryPool::make<RequestMemoryPool::Tag::pool>(
+            std::move(copiedOutputMemory).value()));
+    for (const auto& token : tokens) {
+        pools.push_back(RequestMemoryPool::make<RequestMemoryPool::Tag::token>(token));
+    }
+
+    // Copy input data to the input shared memory pool.
+    uint8_t* inputPtr = mInputMemory->getPointer();
+    for (uint32_t i = 0; i < testModel.main.inputIndexes.size(); i++) {
+        if (!inputs[i].hasNoValue && inputs[i].location.poolIndex == kInputPoolIndex) {
+            const auto& op = testModel.main.operands[testModel.main.inputIndexes[i]];
+            const uint8_t* begin = op.data.get<uint8_t>();
+            const uint8_t* end = begin + op.data.size();
+            std::copy(begin, end, inputPtr + inputs[i].location.offset);
+        }
+    }
+    return Request{
+            .inputs = std::move(inputs), .outputs = std::move(outputs), .pools = std::move(pools)};
+}
+
+std::vector<TestBuffer> ExecutionContext::getOutputBuffers(const TestModel& testModel,
+                                                           const Request& request) const {
+    // Copy out output results.
+    uint8_t* outputPtr = mOutputMemory->getPointer();
+    std::vector<TestBuffer> outputBuffers;
+    for (uint32_t i = 0; i < request.outputs.size(); i++) {
+        const auto& outputLoc = request.outputs[i].location;
+        if (outputLoc.poolIndex == kOutputPoolIndex) {
+            outputBuffers.emplace_back(outputLoc.length, outputPtr + outputLoc.offset);
+        } else {
+            const auto& op = testModel.main.operands[testModel.main.outputIndexes[i]];
+            if (op.data.size() == 0) {
+                outputBuffers.emplace_back(0, nullptr);
+            } else {
+                SCOPED_TRACE("Output index = " + std::to_string(i));
+                const uint32_t bufferIndex = outputLoc.poolIndex - kDeviceMemoryBeginIndex;
+                TestBuffer buffer;
+                getBuffer(mBuffers[bufferIndex], op.data.size(), &buffer);
+                outputBuffers.push_back(std::move(buffer));
+            }
+        }
+    }
+    return outputBuffers;
+}
+
+// Get a TestBuffer with data copied from an IBuffer object.
+void ExecutionContext::getBuffer(const std::shared_ptr<IBuffer>& buffer, size_t size,
+                                 TestBuffer* testBuffer) const {
+    // IBuffer -> Shared memory.
+    auto sharedMemory = nn::createSharedMemory(size).value();
+    auto aidlMemory = utils::convert(sharedMemory).value();
+    const auto ret = buffer->copyTo(aidlMemory);
+    ASSERT_TRUE(ret.isOk());
+
+    // Shared memory -> TestBuffer.
+    const auto outputMemory = nn::map(sharedMemory).value();
+    const uint8_t* outputPtr = std::visit(
+            [](auto* ptr) { return static_cast<const uint8_t*>(ptr); }, outputMemory.pointer);
+    ASSERT_NE(outputPtr, nullptr);
+    ASSERT_NE(testBuffer, nullptr);
+    *testBuffer = TestBuffer(size, outputPtr);
+}
+
+static bool hasZeroSizedOutput(const TestModel& testModel) {
+    return std::any_of(testModel.main.outputIndexes.begin(), testModel.main.outputIndexes.end(),
+                       [&testModel](uint32_t index) {
+                           return testModel.main.operands[index].data.size() == 0;
+                       });
+}
+
+void EvaluatePreparedModel(const std::shared_ptr<IDevice>& device,
+                           const std::shared_ptr<IPreparedModel>& preparedModel,
+                           const TestModel& testModel, const TestConfig& testConfig,
+                           bool* skipped = nullptr) {
+    if (skipped != nullptr) {
+        *skipped = false;
+    }
+    // If output0 does not have size larger than one byte, we can not test with insufficient buffer.
+    if (testConfig.outputType == OutputType::INSUFFICIENT &&
+        !isOutputSizeGreaterThanOne(testModel, 0)) {
+        return;
+    }
+
+    ExecutionContext context(device, preparedModel);
+    auto maybeRequest = context.createRequest(testModel, testConfig.memoryType);
+    // Skip if testing memory domain but no device memory has been allocated.
+    if (!maybeRequest.has_value()) {
+        return;
+    }
+
+    Request request = std::move(maybeRequest).value();
+
+    constexpr uint32_t kInsufficientOutputIndex = 0;
+    if (testConfig.outputType == OutputType::INSUFFICIENT) {
+        makeOutputInsufficientSize(kInsufficientOutputIndex, &request);
+    }
+
+    int64_t loopTimeoutDuration = kOmittedTimeoutDuration;
+    // OutputType::MISSED_DEADLINE is only used by
+    // TestKind::INTINITE_LOOP_TIMEOUT tests to verify that an infinite loop is
+    // aborted after a timeout.
+    if (testConfig.outputType == OutputType::MISSED_DEADLINE) {
+        // Override the default loop timeout duration with a small value to
+        // speed up test execution.
+        constexpr int64_t kMillisecond = 1'000'000;
+        loopTimeoutDuration = 1 * kMillisecond;
+    }
+
+    ErrorStatus executionStatus;
+    std::vector<OutputShape> outputShapes;
+    Timing timing = kNoTiming;
+    switch (testConfig.executor) {
+        case Executor::SYNC: {
+            SCOPED_TRACE("synchronous");
+
+            ExecutionResult executionResult;
+            // execute
+            const auto ret = preparedModel->executeSynchronously(request, testConfig.measureTiming,
+                                                                 kNoDeadline, loopTimeoutDuration,
+                                                                 &executionResult);
+            ASSERT_TRUE(ret.isOk() || ret.getExceptionCode() == EX_SERVICE_SPECIFIC)
+                    << ret.getDescription();
+            if (ret.isOk()) {
+                executionStatus = executionResult.outputSufficientSize
+                                          ? ErrorStatus::NONE
+                                          : ErrorStatus::OUTPUT_INSUFFICIENT_SIZE;
+                outputShapes = std::move(executionResult.outputShapes);
+                timing = executionResult.timing;
+            } else {
+                executionStatus = static_cast<ErrorStatus>(ret.getServiceSpecificError());
+            }
+            break;
+        }
+        case Executor::FENCED: {
+            SCOPED_TRACE("fenced");
+            ErrorStatus result = ErrorStatus::NONE;
+            ndk::ScopedFileDescriptor syncFenceFd;
+            std::shared_ptr<IFencedExecutionCallback> fencedCallback;
+            auto ret = preparedModel->executeFenced(request, {}, testConfig.measureTiming,
+                                                    kNoDeadline, loopTimeoutDuration, kNoDuration,
+                                                    &syncFenceFd, &fencedCallback);
+            ASSERT_TRUE(ret.isOk() || ret.getExceptionCode() == EX_SERVICE_SPECIFIC)
+                    << ret.getDescription();
+            if (!ret.isOk()) {
+                result = static_cast<ErrorStatus>(ret.getServiceSpecificError());
+                executionStatus = result;
+            } else if (syncFenceFd.get() != -1) {
+                std::vector<ndk::ScopedFileDescriptor> waitFor;
+                auto dupFd = dup(syncFenceFd.get());
+                ASSERT_NE(dupFd, -1);
+                waitFor.emplace_back(dupFd);
+                // If a sync fence is returned, try start another run waiting for the sync fence.
+                ret = preparedModel->executeFenced(request, waitFor, testConfig.measureTiming,
+                                                   kNoDeadline, loopTimeoutDuration, kNoDuration,
+                                                   &syncFenceFd, &fencedCallback);
+                ASSERT_TRUE(ret.isOk());
+                waitForSyncFence(syncFenceFd.get());
+            }
+            if (result == ErrorStatus::NONE) {
+                ASSERT_NE(fencedCallback, nullptr);
+                Timing timingFenced;
+                auto ret =
+                        fencedCallback->getExecutionInfo(&timing, &timingFenced, &executionStatus);
+                ASSERT_TRUE(ret.isOk());
+            }
+            break;
+        }
+        default: {
+            FAIL() << "Unsupported execution mode for AIDL interface.";
+        }
+    }
+
+    if (testConfig.outputType != OutputType::FULLY_SPECIFIED &&
+        executionStatus == ErrorStatus::GENERAL_FAILURE) {
+        if (skipped != nullptr) {
+            *skipped = true;
+        }
+        if (!testConfig.reportSkipping) {
+            return;
+        }
+        LOG(INFO) << "NN VTS: Early termination of test because vendor service cannot "
+                     "execute model that it does not support.";
+        std::cout << "[          ]   Early termination of test because vendor service cannot "
+                     "execute model that it does not support."
+                  << std::endl;
+        GTEST_SKIP();
+    }
+    if (!testConfig.measureTiming) {
+        EXPECT_EQ(timing, kNoTiming);
+    } else {
+        if (timing.timeOnDevice != -1 && timing.timeInDriver != -1) {
+            EXPECT_LE(timing.timeOnDevice, timing.timeInDriver);
+        }
+    }
+
+    switch (testConfig.outputType) {
+        case OutputType::FULLY_SPECIFIED:
+            if (testConfig.executor == Executor::FENCED && hasZeroSizedOutput(testModel)) {
+                // Executor::FENCED does not support zero-sized output.
+                ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, executionStatus);
+                return;
+            }
+            // If the model output operands are fully specified, outputShapes must be either
+            // either empty, or have the same number of elements as the number of outputs.
+            ASSERT_EQ(ErrorStatus::NONE, executionStatus);
+            ASSERT_TRUE(outputShapes.size() == 0 ||
+                        outputShapes.size() == testModel.main.outputIndexes.size());
+            break;
+        case OutputType::UNSPECIFIED:
+            if (testConfig.executor == Executor::FENCED) {
+                // For Executor::FENCED, the output shape must be fully specified.
+                ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, executionStatus);
+                return;
+            }
+            // If the model output operands are not fully specified, outputShapes must have
+            // the same number of elements as the number of outputs.
+            ASSERT_EQ(ErrorStatus::NONE, executionStatus);
+            ASSERT_EQ(outputShapes.size(), testModel.main.outputIndexes.size());
+            break;
+        case OutputType::INSUFFICIENT:
+            if (testConfig.executor == Executor::FENCED) {
+                // For Executor::FENCED, the output shape must be fully specified.
+                ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, executionStatus);
+                return;
+            }
+            ASSERT_EQ(ErrorStatus::OUTPUT_INSUFFICIENT_SIZE, executionStatus);
+            ASSERT_EQ(outputShapes.size(), testModel.main.outputIndexes.size());
+            // Check that all returned output dimensions are at least as fully specified as the
+            // union of the information about the corresponding operand in the model and in the
+            // request. In this test, all model outputs have known rank with all dimensions
+            // unspecified, and no dimensional information is provided in the request.
+            for (uint32_t i = 0; i < outputShapes.size(); i++) {
+                ASSERT_EQ(outputShapes[i].isSufficient, i != kInsufficientOutputIndex);
+                const auto& actual = outputShapes[i].dimensions;
+                const auto& golden =
+                        testModel.main.operands[testModel.main.outputIndexes[i]].dimensions;
+                ASSERT_EQ(actual.size(), golden.size());
+                for (uint32_t j = 0; j < actual.size(); j++) {
+                    if (actual[j] == 0) continue;
+                    EXPECT_EQ(actual[j], golden[j]) << "index: " << j;
+                }
+            }
+            return;
+        case OutputType::MISSED_DEADLINE:
+            ASSERT_TRUE(executionStatus == ErrorStatus::MISSED_DEADLINE_TRANSIENT ||
+                        executionStatus == ErrorStatus::MISSED_DEADLINE_PERSISTENT)
+                    << "executionStatus = " << executionStatus;
+            return;
+    }
+
+    // Go through all outputs, check returned output shapes.
+    for (uint32_t i = 0; i < outputShapes.size(); i++) {
+        EXPECT_TRUE(outputShapes[i].isSufficient);
+        const auto& expect = testModel.main.operands[testModel.main.outputIndexes[i]].dimensions;
+        const auto unsignedActual = nn::toUnsigned(outputShapes[i].dimensions);
+        ASSERT_TRUE(unsignedActual.has_value());
+        const std::vector<uint32_t>& actual = unsignedActual.value();
+        EXPECT_EQ(expect, actual);
+    }
+
+    // Retrieve execution results.
+    const std::vector<TestBuffer> outputs = context.getOutputBuffers(testModel, request);
+
+    // We want "close-enough" results.
+    checkResults(testModel, outputs);
+}
+
+void EvaluatePreparedModel(const std::shared_ptr<IDevice>& device,
+                           const std::shared_ptr<IPreparedModel>& preparedModel,
+                           const TestModel& testModel, TestKind testKind) {
+    std::vector<OutputType> outputTypesList;
+    std::vector<bool> measureTimingList;
+    std::vector<Executor> executorList;
+    std::vector<MemoryType> memoryTypeList;
+
+    switch (testKind) {
+        case TestKind::GENERAL: {
+            outputTypesList = {OutputType::FULLY_SPECIFIED};
+            measureTimingList = {false, true};
+            executorList = {Executor::SYNC};
+            memoryTypeList = {MemoryType::ASHMEM};
+        } break;
+        case TestKind::DYNAMIC_SHAPE: {
+            outputTypesList = {OutputType::UNSPECIFIED, OutputType::INSUFFICIENT};
+            measureTimingList = {false, true};
+            executorList = {Executor::SYNC, Executor::FENCED};
+            memoryTypeList = {MemoryType::ASHMEM};
+        } break;
+        case TestKind::MEMORY_DOMAIN: {
+            outputTypesList = {OutputType::FULLY_SPECIFIED};
+            measureTimingList = {false};
+            executorList = {Executor::SYNC, Executor::FENCED};
+            memoryTypeList = {MemoryType::BLOB_AHWB, MemoryType::DEVICE};
+        } break;
+        case TestKind::FENCED_COMPUTE: {
+            outputTypesList = {OutputType::FULLY_SPECIFIED};
+            measureTimingList = {false, true};
+            executorList = {Executor::FENCED};
+            memoryTypeList = {MemoryType::ASHMEM};
+        } break;
+        case TestKind::QUANTIZATION_COUPLING: {
+            LOG(FATAL) << "Wrong TestKind for EvaluatePreparedModel";
+            return;
+        } break;
+        case TestKind::INTINITE_LOOP_TIMEOUT: {
+            outputTypesList = {OutputType::MISSED_DEADLINE};
+            measureTimingList = {false, true};
+            executorList = {Executor::SYNC, Executor::FENCED};
+            memoryTypeList = {MemoryType::ASHMEM};
+        } break;
+    }
+
+    for (const OutputType outputType : outputTypesList) {
+        for (const bool measureTiming : measureTimingList) {
+            for (const Executor executor : executorList) {
+                for (const MemoryType memoryType : memoryTypeList) {
+                    const TestConfig testConfig(executor, measureTiming, outputType, memoryType);
+                    EvaluatePreparedModel(device, preparedModel, testModel, testConfig);
+                }
+            }
+        }
+    }
+}
+
+void EvaluatePreparedCoupledModels(const std::shared_ptr<IDevice>& device,
+                                   const std::shared_ptr<IPreparedModel>& preparedModel,
+                                   const TestModel& testModel,
+                                   const std::shared_ptr<IPreparedModel>& preparedCoupledModel,
+                                   const TestModel& coupledModel) {
+    const std::vector<OutputType> outputTypesList = {OutputType::FULLY_SPECIFIED};
+    const std::vector<bool> measureTimingList = {false, true};
+    const std::vector<Executor> executorList = {Executor::SYNC, Executor::FENCED};
+
+    for (const OutputType outputType : outputTypesList) {
+        for (const bool measureTiming : measureTimingList) {
+            for (const Executor executor : executorList) {
+                const TestConfig testConfig(executor, measureTiming, outputType, MemoryType::ASHMEM,
+                                            /*reportSkipping=*/false);
+                bool baseSkipped = false;
+                EvaluatePreparedModel(device, preparedModel, testModel, testConfig, &baseSkipped);
+                bool coupledSkipped = false;
+                EvaluatePreparedModel(device, preparedCoupledModel, coupledModel, testConfig,
+                                      &coupledSkipped);
+                ASSERT_EQ(baseSkipped, coupledSkipped);
+                if (baseSkipped) {
+                    LOG(INFO) << "NN VTS: Early termination of test because vendor service cannot "
+                                 "execute model that it does not support.";
+                    std::cout << "[          ]   Early termination of test because vendor service "
+                                 "cannot "
+                                 "execute model that it does not support."
+                              << std::endl;
+                    GTEST_SKIP();
+                }
+            }
+        }
+    }
+}
+
+void Execute(const std::shared_ptr<IDevice>& device, const TestModel& testModel,
+             TestKind testKind) {
+    Model model = createModel(testModel);
+    if (testKind == TestKind::DYNAMIC_SHAPE) {
+        makeOutputDimensionsUnspecified(&model);
+    }
+
+    std::shared_ptr<IPreparedModel> preparedModel;
+    switch (testKind) {
+        case TestKind::GENERAL:
+        case TestKind::DYNAMIC_SHAPE:
+        case TestKind::MEMORY_DOMAIN:
+        case TestKind::FENCED_COMPUTE:
+        case TestKind::INTINITE_LOOP_TIMEOUT: {
+            createPreparedModel(device, model, &preparedModel);
+            if (preparedModel == nullptr) return;
+            EvaluatePreparedModel(device, preparedModel, testModel, testKind);
+        } break;
+        case TestKind::QUANTIZATION_COUPLING: {
+            ASSERT_TRUE(testModel.hasQuant8CoupledOperands());
+            createPreparedModel(device, model, &preparedModel,
+                                /*reportSkipping*/ false);
+            TestModel signedQuantizedModel = convertQuant8AsymmOperandsToSigned(testModel);
+            std::shared_ptr<IPreparedModel> preparedCoupledModel;
+            createPreparedModel(device, createModel(signedQuantizedModel), &preparedCoupledModel,
+                                /*reportSkipping*/ false);
+            // If we couldn't prepare a model with unsigned quantization, we must
+            // fail to prepare a model with signed quantization as well.
+            if (preparedModel == nullptr) {
+                ASSERT_EQ(preparedCoupledModel, nullptr);
+                // If we failed to prepare both of the models, we can safely skip
+                // the test.
+                LOG(INFO) << "NN VTS: Early termination of test because vendor service cannot "
+                             "prepare model that it does not support.";
+                std::cout
+                        << "[          ]   Early termination of test because vendor service cannot "
+                           "prepare model that it does not support."
+                        << std::endl;
+                GTEST_SKIP();
+            }
+            ASSERT_NE(preparedCoupledModel, nullptr);
+            EvaluatePreparedCoupledModels(device, preparedModel, testModel, preparedCoupledModel,
+                                          signedQuantizedModel);
+        } break;
+    }
+}
+
+void GeneratedTestBase::SetUp() {
+    testing::TestWithParam<GeneratedTestParam>::SetUp();
+    ASSERT_NE(kDevice, nullptr);
+}
+
+std::vector<NamedModel> getNamedModels(const FilterFn& filter) {
+    return TestModelManager::get().getTestModels(filter);
+}
+
+std::vector<NamedModel> getNamedModels(const FilterNameFn& filter) {
+    return TestModelManager::get().getTestModels(filter);
+}
+
+std::string printGeneratedTest(const testing::TestParamInfo<GeneratedTestParam>& info) {
+    const auto& [namedDevice, namedModel] = info.param;
+    return gtestCompliantName(getName(namedDevice) + "_" + getName(namedModel));
+}
+
+// Tag for the generated tests
+class GeneratedTest : public GeneratedTestBase {};
+
+// Tag for the dynamic output shape tests
+class DynamicOutputShapeTest : public GeneratedTest {};
+
+// Tag for the memory domain tests
+class MemoryDomainTest : public GeneratedTest {};
+
+// Tag for the fenced compute tests
+class FencedComputeTest : public GeneratedTest {};
+
+// Tag for the dynamic output shape tests
+class QuantizationCouplingTest : public GeneratedTest {};
+
+// Tag for the loop timeout tests
+class InfiniteLoopTimeoutTest : public GeneratedTest {};
+
+TEST_P(GeneratedTest, Test) {
+    Execute(kDevice, kTestModel, TestKind::GENERAL);
+}
+
+TEST_P(DynamicOutputShapeTest, Test) {
+    Execute(kDevice, kTestModel, TestKind::DYNAMIC_SHAPE);
+}
+
+TEST_P(MemoryDomainTest, Test) {
+    Execute(kDevice, kTestModel, TestKind::MEMORY_DOMAIN);
+}
+
+TEST_P(FencedComputeTest, Test) {
+    Execute(kDevice, kTestModel, TestKind::FENCED_COMPUTE);
+}
+
+TEST_P(QuantizationCouplingTest, Test) {
+    Execute(kDevice, kTestModel, TestKind::QUANTIZATION_COUPLING);
+}
+
+TEST_P(InfiniteLoopTimeoutTest, Test) {
+    Execute(kDevice, kTestModel, TestKind::INTINITE_LOOP_TIMEOUT);
+}
+
+INSTANTIATE_GENERATED_TEST(GeneratedTest,
+                           [](const TestModel& testModel) { return !testModel.expectFailure; });
+
+INSTANTIATE_GENERATED_TEST(DynamicOutputShapeTest, [](const TestModel& testModel) {
+    return !testModel.expectFailure && !testModel.hasScalarOutputs();
+});
+
+INSTANTIATE_GENERATED_TEST(MemoryDomainTest,
+                           [](const TestModel& testModel) { return !testModel.expectFailure; });
+
+INSTANTIATE_GENERATED_TEST(FencedComputeTest,
+                           [](const TestModel& testModel) { return !testModel.expectFailure; });
+
+INSTANTIATE_GENERATED_TEST(QuantizationCouplingTest, [](const TestModel& testModel) {
+    return !testModel.expectFailure && testModel.hasQuant8CoupledOperands() &&
+           testModel.main.operations.size() == 1;
+});
+
+INSTANTIATE_GENERATED_TEST(InfiniteLoopTimeoutTest, [](const TestModel& testModel) {
+    return testModel.isInfiniteLoopTimeoutTest();
+});
+
+}  // namespace aidl::android::hardware::neuralnetworks::vts::functional
diff --git a/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.h b/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.h
new file mode 100644
index 0000000..ad40f06
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.h
@@ -0,0 +1,88 @@
+/*
+ * 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_NEURALNETWORKS_AIDL_GENERATED_TEST_HARNESS_H
+#define ANDROID_HARDWARE_NEURALNETWORKS_AIDL_GENERATED_TEST_HARNESS_H
+
+#include <functional>
+#include <vector>
+
+#include <TestHarness.h>
+#include "Utils.h"
+#include "VtsHalNeuralnetworks.h"
+
+namespace aidl::android::hardware::neuralnetworks::vts::functional {
+
+using NamedModel = Named<const test_helper::TestModel*>;
+using GeneratedTestParam = std::tuple<NamedDevice, NamedModel>;
+
+class GeneratedTestBase : public testing::TestWithParam<GeneratedTestParam> {
+  protected:
+    void SetUp() override;
+    const std::shared_ptr<IDevice> kDevice = getData(std::get<NamedDevice>(GetParam()));
+    const test_helper::TestModel& kTestModel = *getData(std::get<NamedModel>(GetParam()));
+};
+
+using FilterFn = std::function<bool(const test_helper::TestModel&)>;
+std::vector<NamedModel> getNamedModels(const FilterFn& filter);
+
+using FilterNameFn = std::function<bool(const std::string&)>;
+std::vector<NamedModel> getNamedModels(const FilterNameFn& filter);
+
+std::string printGeneratedTest(const testing::TestParamInfo<GeneratedTestParam>& info);
+
+#define INSTANTIATE_GENERATED_TEST(TestSuite, filter)                                     \
+    GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(TestSuite);                             \
+    INSTANTIATE_TEST_SUITE_P(TestGenerated, TestSuite,                                    \
+                             testing::Combine(testing::ValuesIn(getNamedDevices()),       \
+                                              testing::ValuesIn(getNamedModels(filter))), \
+                             printGeneratedTest)
+
+// Tag for the validation tests, instantiated in VtsHalNeuralnetworks.cpp.
+// TODO: Clean up the hierarchy for ValidationTest.
+class ValidationTest : public GeneratedTestBase {};
+
+Model createModel(const test_helper::TestModel& testModel);
+
+void PrepareModel(const std::shared_ptr<IDevice>& device, const Model& model,
+                  std::shared_ptr<IPreparedModel>* preparedModel);
+
+enum class TestKind {
+    // Runs a test model and compares the results to a golden data
+    GENERAL,
+    // Same as GENERAL but sets dimensions for the output tensors to zeros
+    DYNAMIC_SHAPE,
+    // Same as GENERAL but use device memories for inputs and outputs
+    MEMORY_DOMAIN,
+    // Same as GENERAL but use executeFenced for exeuction
+    FENCED_COMPUTE,
+    // Tests if quantized model with TENSOR_QUANT8_ASYMM produces the same result
+    // (OK/SKIPPED/FAILED) as the model with all such tensors converted to
+    // TENSOR_QUANT8_ASYMM_SIGNED.
+    QUANTIZATION_COUPLING,
+    // Runs a test model and verifies that MISSED_DEADLINE_* is returned.
+    INTINITE_LOOP_TIMEOUT
+};
+
+void EvaluatePreparedModel(const std::shared_ptr<IDevice>& device,
+                           const std::shared_ptr<IPreparedModel>& preparedModel,
+                           const test_helper::TestModel& testModel, TestKind testKind);
+
+void waitForSyncFence(int syncFd);
+
+}  // namespace aidl::android::hardware::neuralnetworks::vts::functional
+
+#endif  // ANDROID_HARDWARE_NEURALNETWORKS_AIDL_GENERATED_TEST_HARNESS_H
diff --git a/neuralnetworks/aidl/vts/functional/LogTestCaseToLogcat.h b/neuralnetworks/aidl/vts/functional/LogTestCaseToLogcat.h
new file mode 100644
index 0000000..c9fd432
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/LogTestCaseToLogcat.h
@@ -0,0 +1,40 @@
+/*
+ * 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_NEURALNETWORKS_AIDL_LOG_TEST_CASE_TO_LOGCAT_H
+#define ANDROID_HARDWARE_NEURALNETWORKS_AIDL_LOG_TEST_CASE_TO_LOGCAT_H
+
+#include <android-base/logging.h>
+#include <gtest/gtest.h>
+
+namespace aidl::android::hardware::neuralnetworks {
+
+class LogTestCaseToLogcat : public ::testing::EmptyTestEventListener {
+  public:
+    void OnTestStart(const ::testing::TestInfo& test_info) override {
+        LOG(INFO) << "[Test Case] " << test_info.test_suite_name() << "." << test_info.name()
+                  << " BEGIN";
+    }
+
+    void OnTestEnd(const ::testing::TestInfo& test_info) override {
+        LOG(INFO) << "[Test Case] " << test_info.test_suite_name() << "." << test_info.name()
+                  << " END";
+    }
+};
+
+}  // namespace aidl::android::hardware::neuralnetworks
+
+#endif  // ANDROID_HARDWARE_NEURALNETWORKS_AIDL_LOG_TEST_CASE_TO_LOGCAT_H
diff --git a/neuralnetworks/aidl/vts/functional/MemoryDomainTests.cpp b/neuralnetworks/aidl/vts/functional/MemoryDomainTests.cpp
new file mode 100644
index 0000000..a37a0ca
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/MemoryDomainTests.cpp
@@ -0,0 +1,1176 @@
+/*
+ * 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.
+ */
+
+#define LOG_TAG "neuralnetworks_aidl_hal_test"
+
+#include <android-base/logging.h>
+#include <android/binder_auto_utils.h>
+#include <android/binder_interface_utils.h>
+#include <android/binder_status.h>
+#include <gtest/gtest.h>
+
+#include <LegacyUtils.h>
+#include <TestHarness.h>
+#include <Utils.h>
+#include <nnapi/SharedMemory.h>
+#include <nnapi/hal/aidl/Conversions.h>
+#include <nnapi/hal/aidl/Utils.h>
+
+#include "AidlHalInterfaces.h"
+#include "Callbacks.h"
+#include "GeneratedTestHarness.h"
+#include "MemoryUtils.h"
+#include "Utils.h"
+#include "VtsHalNeuralnetworks.h"
+
+namespace aidl::android::hardware::neuralnetworks::vts::functional {
+
+using namespace test_helper;
+using implementation::PreparedModelCallback;
+
+namespace {
+
+// An AIDL driver is likely to support at least one of the following operand types.
+const std::vector<TestOperandType> kTestOperandTypeChoicesVector = {
+        TestOperandType::TENSOR_FLOAT32,
+        TestOperandType::TENSOR_FLOAT16,
+        TestOperandType::TENSOR_QUANT8_ASYMM,
+        TestOperandType::TENSOR_QUANT8_ASYMM_SIGNED,
+};
+const auto kTestOperandTypeChoices = testing::ValuesIn(kTestOperandTypeChoicesVector);
+// TODO(b/179270601): restore kNamedDeviceChoices
+
+bool isInChoices(TestOperandType type) {
+    return std::count(kTestOperandTypeChoicesVector.begin(), kTestOperandTypeChoicesVector.end(),
+                      type) > 0;
+}
+
+bool isFloat(TestOperandType type) {
+    CHECK(isInChoices(type));
+    return type == TestOperandType::TENSOR_FLOAT32 || type == TestOperandType::TENSOR_FLOAT16;
+}
+
+// Create placeholder buffers for model constants as well as inputs and outputs.
+// We only care about the size here because we will not check accuracy in validation tests.
+void createDummyData(TestModel* testModel) {
+    for (auto& operand : testModel->main.operands) {
+        if (operand.data != nullptr) continue;
+        switch (operand.lifetime) {
+            case TestOperandLifeTime::SUBGRAPH_INPUT:
+            case TestOperandLifeTime::SUBGRAPH_OUTPUT:
+            case TestOperandLifeTime::CONSTANT_COPY:
+            case TestOperandLifeTime::CONSTANT_REFERENCE: {
+                const uint32_t size = nn::nonExtensionOperandSizeOfData(
+                        static_cast<nn::OperandType>(operand.type), operand.dimensions);
+                operand.data = TestBuffer(size);
+            } break;
+            default:
+                break;
+        }
+    }
+}
+
+TestOperand createInt32Scalar(int32_t value) {
+    return {
+            .type = TestOperandType::INT32,
+            .dimensions = {},
+            .numberOfConsumers = 1,
+            .scale = 0.0f,
+            .zeroPoint = 0,
+            .lifetime = TestOperandLifeTime::CONSTANT_COPY,
+            .data = TestBuffer::createFromVector<int32_t>({value}),
+    };
+}
+
+// Construct a test model with multiple CONV_2D operations with the given operand as inputs.
+// The dimensions of the filters are chosen to ensure outputs has the same dimensions as inputs.
+// We choose CONV_2D operation because it is commonly supported by most drivers.
+TestModel createConvModel(const TestOperand& operand, uint32_t numOperations) {
+    CHECK(isInChoices(operand.type));
+
+    TestOperand weight = {.type = operand.type,
+                          .dimensions = {operand.dimensions[3], 3, 3, operand.dimensions[3]},
+                          .numberOfConsumers = 1,
+                          .scale = isFloat(operand.type) ? 0.0f : 1.0f,
+                          .zeroPoint = 0,
+                          .lifetime = TestOperandLifeTime::CONSTANT_COPY};
+
+    TestOperand bias = {
+            .type = isFloat(operand.type) ? operand.type : TestOperandType::TENSOR_INT32,
+            .dimensions = {operand.dimensions[3]},
+            .numberOfConsumers = 1,
+            .scale = operand.scale * weight.scale,
+            .zeroPoint = 0,
+            .lifetime = TestOperandLifeTime::CONSTANT_COPY};
+
+    TestOperand output = operand;
+    output.numberOfConsumers = 0;
+    output.lifetime = TestOperandLifeTime::SUBGRAPH_OUTPUT;
+
+    const std::vector<TestOperand> operands = {
+            operand,
+            std::move(weight),
+            std::move(bias),
+            createInt32Scalar(1),  // same padding
+            createInt32Scalar(1),  // width stride
+            createInt32Scalar(1),  // height stride
+            createInt32Scalar(0),  // activation = NONE
+            std::move(output),
+    };
+
+    TestModel model;
+    for (uint32_t i = 0; i < numOperations; i++) {
+        model.main.operands.insert(model.main.operands.end(), operands.begin(), operands.end());
+        const uint32_t inputIndex = operands.size() * i;
+        const uint32_t outputIndex = inputIndex + operands.size() - 1;
+        std::vector<uint32_t> inputs(operands.size() - 1);
+        std::iota(inputs.begin(), inputs.end(), inputIndex);
+        model.main.operations.push_back({.type = TestOperationType::CONV_2D,
+                                         .inputs = std::move(inputs),
+                                         .outputs = {outputIndex}});
+        model.main.inputIndexes.push_back(inputIndex);
+        model.main.outputIndexes.push_back(outputIndex);
+    }
+    createDummyData(&model);
+    return model;
+}
+
+// Construct a test model with a single ADD operation with the given operand as input0 and input1.
+// This is to cover additional cases that the CONV_2D model does not support, e.g. arbitrary input
+// operand rank, scalar input operand. We choose ADD operation because it is commonly supported by
+// most drivers.
+TestModel createSingleAddModel(const TestOperand& operand) {
+    CHECK(isInChoices(operand.type));
+
+    TestOperand act = {
+            .type = TestOperandType::INT32,
+            .dimensions = {},
+            .numberOfConsumers = 1,
+            .scale = 0.0f,
+            .zeroPoint = 0,
+            .lifetime = TestOperandLifeTime::SUBGRAPH_INPUT,
+    };
+
+    TestOperand output = operand;
+    output.numberOfConsumers = 0;
+    output.lifetime = TestOperandLifeTime::SUBGRAPH_OUTPUT;
+
+    TestModel model = {
+            .main =
+                    {
+                            .operands =
+                                    {
+                                            operand,
+                                            operand,
+                                            std::move(act),
+                                            output,
+                                    },
+                            .operations = {{.type = TestOperationType::ADD,
+                                            .inputs = {0, 1, 2},
+                                            .outputs = {3}}},
+                            .inputIndexes = {0, 1, 2},
+                            .outputIndexes = {3},
+                    },
+    };
+    createDummyData(&model);
+    return model;
+}
+
+// A placeholder invalid IPreparedModel class for MemoryDomainAllocateTest.InvalidPreparedModel
+class InvalidPreparedModel : public BnPreparedModel {
+  public:
+    ndk::ScopedAStatus executeSynchronously(const Request&, bool, int64_t, int64_t,
+                                            ExecutionResult*) override {
+        return ndk::ScopedAStatus::fromServiceSpecificError(
+                static_cast<int32_t>(ErrorStatus::GENERAL_FAILURE));
+    }
+    ndk::ScopedAStatus executeFenced(const Request&, const std::vector<ndk::ScopedFileDescriptor>&,
+                                     bool, int64_t, int64_t, int64_t, ndk::ScopedFileDescriptor*,
+                                     std::shared_ptr<IFencedExecutionCallback>*) override {
+        return ndk::ScopedAStatus::fromServiceSpecificError(
+                static_cast<int32_t>(ErrorStatus::GENERAL_FAILURE));
+    }
+};
+
+template <typename... Args>
+std::vector<RequestMemoryPool> createRequestMemoryPools(const Args&... pools) {
+    std::vector<RequestMemoryPool> memoryPools;
+    memoryPools.reserve(sizeof...(Args));
+    // This fold operator calls push_back on each of the function arguments.
+    (memoryPools.push_back(utils::clone(pools).value()), ...);
+    return memoryPools;
+};
+
+}  // namespace
+
+class MemoryDomainTestBase : public testing::Test {
+  protected:
+    MemoryDomainTestBase(std::shared_ptr<IDevice> device, TestOperandType type)
+        : kDevice(std::move(device)),
+          kTestOperandType(type),
+          kTestOperand(kTestOperandMap.at(type)),
+          kTestOperandDataSize(nn::nonExtensionOperandSizeOfData(static_cast<nn::OperandType>(type),
+                                                                 kTestOperand.dimensions)) {}
+
+    void SetUp() override {
+        testing::Test::SetUp();
+        ASSERT_NE(kDevice, nullptr);
+    }
+
+    std::shared_ptr<IPreparedModel> createConvPreparedModel(const TestOperand& testOperand,
+                                                            uint32_t numOperations = 1) {
+        const TestModel testModel = createConvModel(testOperand, numOperations);
+        const Model model = createModel(testModel);
+        std::shared_ptr<IPreparedModel> preparedModel;
+        createPreparedModel(kDevice, model, &preparedModel, /*reportSkipping=*/false);
+        return preparedModel;
+    }
+
+    std::shared_ptr<IPreparedModel> createAddPreparedModel(const TestOperand& testOperand) {
+        const TestModel testModel = createSingleAddModel(testOperand);
+        const Model model = createModel(testModel);
+        std::shared_ptr<IPreparedModel> preparedModel;
+        createPreparedModel(kDevice, model, &preparedModel, /*reportSkipping=*/false);
+        return preparedModel;
+    }
+
+    static const std::map<TestOperandType, TestOperand> kTestOperandMap;
+
+    const std::shared_ptr<IDevice> kDevice;
+    const TestOperandType kTestOperandType;
+    const TestOperand& kTestOperand;
+    const uint32_t kTestOperandDataSize;
+};
+
+const std::map<TestOperandType, TestOperand> MemoryDomainTestBase::kTestOperandMap = {
+        {TestOperandType::TENSOR_FLOAT32,
+         {
+                 .type = TestOperandType::TENSOR_FLOAT32,
+                 .dimensions = {1, 32, 32, 8},
+                 .numberOfConsumers = 1,
+                 .scale = 0.0f,
+                 .zeroPoint = 0,
+                 .lifetime = TestOperandLifeTime::SUBGRAPH_INPUT,
+         }},
+        {TestOperandType::TENSOR_FLOAT16,
+         {
+                 .type = TestOperandType::TENSOR_FLOAT16,
+                 .dimensions = {1, 32, 32, 8},
+                 .numberOfConsumers = 1,
+                 .scale = 0.0f,
+                 .zeroPoint = 0,
+                 .lifetime = TestOperandLifeTime::SUBGRAPH_INPUT,
+         }},
+        {TestOperandType::TENSOR_QUANT8_ASYMM,
+         {
+                 .type = TestOperandType::TENSOR_QUANT8_ASYMM,
+                 .dimensions = {1, 32, 32, 8},
+                 .numberOfConsumers = 1,
+                 .scale = 0.5f,
+                 .zeroPoint = 0,
+                 .lifetime = TestOperandLifeTime::SUBGRAPH_INPUT,
+         }},
+        {TestOperandType::TENSOR_QUANT8_ASYMM_SIGNED,
+         {
+                 .type = TestOperandType::TENSOR_QUANT8_ASYMM_SIGNED,
+                 .dimensions = {1, 32, 32, 8},
+                 .numberOfConsumers = 1,
+                 .scale = 0.5f,
+                 .zeroPoint = 0,
+                 .lifetime = TestOperandLifeTime::SUBGRAPH_INPUT,
+         }},
+};
+
+using MemoryDomainAllocateTestParam = std::tuple<NamedDevice, TestOperandType>;
+class MemoryDomainAllocateTest : public MemoryDomainTestBase,
+                                 public testing::WithParamInterface<MemoryDomainAllocateTestParam> {
+  protected:
+    MemoryDomainAllocateTest()
+        : MemoryDomainTestBase(getData(std::get<NamedDevice>(GetParam())),
+                               std::get<TestOperandType>(GetParam())) {}
+
+    struct AllocateTestArgs {
+        std::vector<int32_t> dimensions;
+        std::vector<std::shared_ptr<IPreparedModel>> preparedModels;
+        std::vector<BufferRole> inputRoles;
+        std::vector<BufferRole> outputRoles;
+    };
+
+    // Validation test for IDevice::allocate. The driver is expected to fail with INVALID_ARGUMENT,
+    // or GENERAL_FAILURE if memory domain is not supported.
+    void validateAllocate(AllocateTestArgs args) {
+        std::vector<IPreparedModelParcel> preparedModelParcels;
+        preparedModelParcels.reserve(args.preparedModels.size());
+        for (const auto& model : args.preparedModels) {
+            preparedModelParcels.push_back({.preparedModel = model});
+        }
+        DeviceBuffer buffer;
+        const auto ret =
+                kDevice->allocate({.dimensions = std::move(args.dimensions)}, preparedModelParcels,
+                                  args.inputRoles, args.outputRoles, &buffer);
+
+        ASSERT_EQ(ret.getExceptionCode(), EX_SERVICE_SPECIFIC);
+        ASSERT_TRUE(static_cast<ErrorStatus>(ret.getServiceSpecificError()) ==
+                            ErrorStatus::INVALID_ARGUMENT ||
+                    static_cast<ErrorStatus>(ret.getServiceSpecificError()) ==
+                            ErrorStatus::GENERAL_FAILURE);
+    }
+
+    void testConflictOperands(const std::shared_ptr<IPreparedModel>& model1,
+                              const std::shared_ptr<IPreparedModel>& model2) {
+        validateAllocate({
+                .preparedModels = {model1, model2},
+                .inputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f},
+                               {.modelIndex = 1, .ioIndex = 0, .frequency = 1.0f}},
+        });
+        validateAllocate({
+                .preparedModels = {model1, model2},
+                .inputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f}},
+                .outputRoles = {{.modelIndex = 1, .ioIndex = 0, .frequency = 1.0f}},
+        });
+        validateAllocate({
+                .preparedModels = {model1, model2},
+                .outputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f},
+                                {.modelIndex = 1, .ioIndex = 0, .frequency = 1.0f}},
+        });
+    }
+};
+
+TEST_P(MemoryDomainAllocateTest, EmptyRole) {
+    // Test with empty prepared models and roles.
+    validateAllocate({});
+
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    if (preparedModel == nullptr) return;
+
+    // Test again with non-empty prepared models but empty roles.
+    validateAllocate({
+            .preparedModels = {preparedModel},
+    });
+}
+
+TEST_P(MemoryDomainAllocateTest, NullptrPreparedModel) {
+    // Test with nullptr prepared model as input role.
+    validateAllocate({
+            .preparedModels = {nullptr},
+            .inputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f}},
+    });
+
+    // Test with nullptr prepared model as output role.
+    validateAllocate({
+            .preparedModels = {nullptr},
+            .outputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f}},
+    });
+}
+
+TEST_P(MemoryDomainAllocateTest, InvalidPreparedModel) {
+    std::shared_ptr<InvalidPreparedModel> invalidPreparedModel =
+            ndk::SharedRefBase::make<InvalidPreparedModel>();
+
+    // Test with invalid prepared model as input role.
+    validateAllocate({
+            .preparedModels = {invalidPreparedModel},
+            .inputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f}},
+    });
+
+    // Test with invalid prepared model as output role.
+    validateAllocate({
+            .preparedModels = {invalidPreparedModel},
+            .outputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f}},
+    });
+}
+
+TEST_P(MemoryDomainAllocateTest, InvalidModelIndex) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    if (preparedModel == nullptr) return;
+
+    // This should fail, because the model index is out of bound.
+    validateAllocate({
+            .preparedModels = {preparedModel},
+            .inputRoles = {{.modelIndex = 1, .ioIndex = 0, .frequency = 1.0f}},
+    });
+
+    // This should fail, because the model index is out of bound.
+    validateAllocate({
+            .preparedModels = {preparedModel},
+            .outputRoles = {{.modelIndex = 1, .ioIndex = 0, .frequency = 1.0f}},
+    });
+}
+
+TEST_P(MemoryDomainAllocateTest, InvalidIOIndex) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    if (preparedModel == nullptr) return;
+
+    // This should fail, because the model only has one input.
+    validateAllocate({
+            .preparedModels = {preparedModel},
+            .inputRoles = {{.modelIndex = 0, .ioIndex = 1, .frequency = 1.0f}},
+    });
+
+    // This should fail, because the model only has one output.
+    validateAllocate({
+            .preparedModels = {preparedModel},
+            .outputRoles = {{.modelIndex = 0, .ioIndex = 1, .frequency = 1.0f}},
+    });
+}
+
+TEST_P(MemoryDomainAllocateTest, InvalidFrequency) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    if (preparedModel == nullptr) return;
+
+    for (float invalidFreq : {10.0f, 0.0f, -0.5f}) {
+        // Test with invalid frequency for input roles.
+        validateAllocate({
+                .preparedModels = {preparedModel},
+                .inputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = invalidFreq}},
+        });
+        // Test with invalid frequency for output roles.
+        validateAllocate({
+                .preparedModels = {preparedModel},
+                .outputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = invalidFreq}},
+        });
+    }
+}
+
+TEST_P(MemoryDomainAllocateTest, SameRoleSpecifiedTwice) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    if (preparedModel == nullptr) return;
+
+    // Same role with same model index.
+    validateAllocate({
+            .preparedModels = {preparedModel},
+            .inputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f},
+                           {.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f}},
+    });
+    validateAllocate({
+            .preparedModels = {preparedModel},
+            .outputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f},
+                            {.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f}},
+    });
+
+    // Different model indexes, but logically referring to the same role.
+    validateAllocate({
+            .preparedModels = {preparedModel, preparedModel},
+            .inputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f},
+                           {.modelIndex = 1, .ioIndex = 0, .frequency = 1.0f}},
+    });
+    validateAllocate({
+            .preparedModels = {preparedModel, preparedModel},
+            .outputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f},
+                            {.modelIndex = 1, .ioIndex = 0, .frequency = 1.0f}},
+    });
+}
+
+TEST_P(MemoryDomainAllocateTest, ConflictOperandType) {
+    const std::map<TestOperandType, TestOperandType> conflictTypeMap = {
+            {TestOperandType::TENSOR_FLOAT32, TestOperandType::TENSOR_FLOAT16},
+            {TestOperandType::TENSOR_FLOAT16, TestOperandType::TENSOR_FLOAT32},
+            {TestOperandType::TENSOR_QUANT8_ASYMM, TestOperandType::TENSOR_QUANT8_ASYMM_SIGNED},
+            {TestOperandType::TENSOR_QUANT8_ASYMM_SIGNED, TestOperandType::TENSOR_QUANT8_ASYMM},
+    };
+
+    TestOperand conflictTestOperand = kTestOperand;
+    const auto it = conflictTypeMap.find(kTestOperandType);
+    ASSERT_FALSE(it == conflictTypeMap.end());
+    conflictTestOperand.type = it->second;
+
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    auto conflictPreparedModel = createConvPreparedModel(conflictTestOperand);
+    if (preparedModel == nullptr || conflictPreparedModel == nullptr) return;
+    testConflictOperands(preparedModel, conflictPreparedModel);
+}
+
+TEST_P(MemoryDomainAllocateTest, ConflictScale) {
+    if (isFloat(kTestOperandType)) return;
+
+    TestOperand conflictTestOperand = kTestOperand;
+    ASSERT_NE(conflictTestOperand.scale, 1.0f);
+    conflictTestOperand.scale = 1.0f;
+
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    auto conflictPreparedModel = createConvPreparedModel(conflictTestOperand);
+    if (preparedModel == nullptr || conflictPreparedModel == nullptr) return;
+    testConflictOperands(preparedModel, conflictPreparedModel);
+}
+
+TEST_P(MemoryDomainAllocateTest, ConflictZeroPoint) {
+    if (isFloat(kTestOperandType)) return;
+
+    TestOperand conflictTestOperand = kTestOperand;
+    ASSERT_NE(conflictTestOperand.zeroPoint, 10);
+    conflictTestOperand.zeroPoint = 10;
+
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    auto conflictPreparedModel = createConvPreparedModel(conflictTestOperand);
+    if (preparedModel == nullptr || conflictPreparedModel == nullptr) return;
+    testConflictOperands(preparedModel, conflictPreparedModel);
+}
+
+TEST_P(MemoryDomainAllocateTest, ConflictRankBetweenRoles) {
+    TestOperand conflictTestOperand = kTestOperand;
+    conflictTestOperand.dimensions.pop_back();
+
+    auto preparedModel = createAddPreparedModel(kTestOperand);
+    auto conflictPreparedModel = createAddPreparedModel(conflictTestOperand);
+    if (preparedModel == nullptr || conflictPreparedModel == nullptr) return;
+    testConflictOperands(preparedModel, conflictPreparedModel);
+}
+
+TEST_P(MemoryDomainAllocateTest, ConflictDimensionsBetweenRoles) {
+    TestOperand conflictTestOperand = kTestOperand;
+    conflictTestOperand.dimensions[0] = 4;
+
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    auto conflictPreparedModel = createConvPreparedModel(conflictTestOperand);
+    if (preparedModel == nullptr || conflictPreparedModel == nullptr) return;
+    testConflictOperands(preparedModel, conflictPreparedModel);
+}
+
+TEST_P(MemoryDomainAllocateTest, ConflictRankBetweenRoleAndDesc) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    if (preparedModel == nullptr) return;
+
+    auto badDimensions = utils::toSigned(kTestOperand.dimensions).value();
+    badDimensions.pop_back();
+
+    validateAllocate({
+            .dimensions = badDimensions,
+            .preparedModels = {preparedModel},
+            .inputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f}},
+    });
+    validateAllocate({
+            .dimensions = badDimensions,
+            .preparedModels = {preparedModel},
+            .outputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f}},
+    });
+}
+
+TEST_P(MemoryDomainAllocateTest, ConflictDimensionsBetweenRoleAndDesc) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    if (preparedModel == nullptr) return;
+
+    auto badDimensions = utils::toSigned(kTestOperand.dimensions).value();
+    badDimensions[0] = 4;
+
+    validateAllocate({
+            .dimensions = badDimensions,
+            .preparedModels = {preparedModel},
+            .inputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f}},
+    });
+    validateAllocate({
+            .dimensions = badDimensions,
+            .preparedModels = {preparedModel},
+            .outputRoles = {{.modelIndex = 0, .ioIndex = 0, .frequency = 1.0f}},
+    });
+}
+
+TEST_P(MemoryDomainAllocateTest, ConflictRankWithScalarRole) {
+    auto preparedModel = createAddPreparedModel(kTestOperand);
+    if (preparedModel == nullptr) return;
+
+    // This should fail, because the target operand is a scalar but a non-empty dimension is
+    // specified.
+    validateAllocate({
+            .dimensions = {1},
+            .preparedModels = {preparedModel},
+            .inputRoles = {{.modelIndex = 0, .ioIndex = 2, .frequency = 1.0f}},
+    });
+}
+
+std::string printMemoryDomainAllocateTest(
+        const testing::TestParamInfo<MemoryDomainAllocateTestParam>& info) {
+    const auto& [namedDevice, operandType] = info.param;
+    const std::string type = toString(static_cast<OperandType>(operandType));
+    return gtestCompliantName(getName(namedDevice) + "_" + type);
+}
+
+GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(MemoryDomainAllocateTest);
+INSTANTIATE_TEST_SUITE_P(TestMemoryDomain, MemoryDomainAllocateTest,
+                         testing::Combine(testing::ValuesIn(getNamedDevices()),
+                                          kTestOperandTypeChoices),
+                         printMemoryDomainAllocateTest);
+
+class MemoryDomainCopyTestBase : public MemoryDomainTestBase {
+  protected:
+    MemoryDomainCopyTestBase(std::shared_ptr<IDevice> device, TestOperandType type)
+        : MemoryDomainTestBase(std::move(device), type) {}
+
+    // Allocates device memory for roles of a single prepared model.
+    // Returns {IBuffer, token} if success; returns {nullptr, 0} if not supported.
+    DeviceBuffer allocateBuffer(const std::shared_ptr<IPreparedModel>& preparedModel,
+                                const std::vector<int32_t>& inputIndexes,
+                                const std::vector<int32_t>& outputIndexes,
+                                const std::vector<int32_t>& dimensions) {
+        if (preparedModel == nullptr) {
+            return {.buffer = nullptr, .token = 0};
+        }
+
+        std::vector<BufferRole> inputRoles(inputIndexes.size()), outputRoles(outputIndexes.size());
+        auto trans = [](int32_t ind) -> BufferRole {
+            return {.modelIndex = 0, .ioIndex = ind, .frequency = 1.0f};
+        };
+        std::transform(inputIndexes.begin(), inputIndexes.end(), inputRoles.begin(), trans);
+        std::transform(outputIndexes.begin(), outputIndexes.end(), outputRoles.begin(), trans);
+
+        IPreparedModelParcel parcel;
+        parcel.preparedModel = preparedModel;
+
+        DeviceBuffer buffer;
+
+        const auto ret = kDevice->allocate({.dimensions = dimensions}, {parcel}, inputRoles,
+                                           outputRoles, &buffer);
+
+        if (!ret.isOk()) {
+            EXPECT_EQ(ret.getExceptionCode(), EX_SERVICE_SPECIFIC);
+            EXPECT_EQ(static_cast<ErrorStatus>(ret.getServiceSpecificError()),
+                      ErrorStatus::GENERAL_FAILURE);
+            return DeviceBuffer{
+                    .buffer = nullptr,
+                    .token = 0,
+            };
+        }
+
+        EXPECT_NE(buffer.buffer, nullptr);
+        EXPECT_GT(buffer.token, 0);
+
+        return buffer;
+    }
+
+    DeviceBuffer allocateBuffer(const std::shared_ptr<IPreparedModel>& preparedModel,
+                                const std::vector<int32_t>& inputIndexes,
+                                const std::vector<int32_t>& outputIndexes) {
+        return allocateBuffer(preparedModel, inputIndexes, outputIndexes, {});
+    }
+
+    Memory allocateSharedMemory(uint32_t size) {
+        const auto sharedMemory = nn::createSharedMemory(size).value();
+        auto memory = utils::convert(sharedMemory).value();
+        EXPECT_EQ(memory.size, size);
+        return memory;
+    }
+
+    void testCopyFrom(const std::shared_ptr<IBuffer>& buffer, const Memory& memory,
+                      const std::vector<int32_t>& dimensions, ErrorStatus expectedStatus) {
+        const auto ret = buffer->copyFrom(memory, dimensions);
+        if (expectedStatus == ErrorStatus::NONE) {
+            ASSERT_TRUE(ret.isOk());
+        } else {
+            ASSERT_EQ(ret.getExceptionCode(), EX_SERVICE_SPECIFIC);
+            ASSERT_EQ(expectedStatus, static_cast<ErrorStatus>(ret.getServiceSpecificError()));
+        }
+    }
+
+    void testCopyTo(const std::shared_ptr<IBuffer>& buffer, const Memory& memory,
+                    ErrorStatus expectedStatus) {
+        const auto ret = buffer->copyTo(memory);
+        if (expectedStatus == ErrorStatus::NONE) {
+            ASSERT_TRUE(ret.isOk());
+        } else {
+            ASSERT_EQ(ret.getExceptionCode(), EX_SERVICE_SPECIFIC);
+            ASSERT_EQ(expectedStatus, static_cast<ErrorStatus>(ret.getServiceSpecificError()));
+        }
+    }
+
+    void initializeDeviceMemory(const std::shared_ptr<IBuffer>& buffer) {
+        Memory memory = allocateSharedMemory(kTestOperandDataSize);
+        ASSERT_EQ(memory.size, kTestOperandDataSize);
+        testCopyFrom(buffer, memory, utils::toSigned(kTestOperand.dimensions).value(),
+                     ErrorStatus::NONE);
+    }
+};
+
+using MemoryDomainCopyTestParam = std::tuple<NamedDevice, TestOperandType>;
+class MemoryDomainCopyTest : public MemoryDomainCopyTestBase,
+                             public testing::WithParamInterface<MemoryDomainCopyTestParam> {
+  protected:
+    MemoryDomainCopyTest()
+        : MemoryDomainCopyTestBase(getData(std::get<NamedDevice>(GetParam())),
+                                   std::get<TestOperandType>(GetParam())) {}
+};
+
+TEST_P(MemoryDomainCopyTest, CopyFrom_InvalidMemorySize) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    auto [buffer, token] = allocateBuffer(preparedModel, {0}, {0});
+    if (buffer == nullptr) return;
+
+    uint32_t badMemorySize1 = kTestOperandDataSize / 2, badMemorySize2 = kTestOperandDataSize * 2;
+    Memory badMemory1 = allocateSharedMemory(badMemorySize1);
+    Memory badMemory2 = allocateSharedMemory(badMemorySize2);
+    testCopyFrom(buffer, badMemory1, {}, ErrorStatus::INVALID_ARGUMENT);
+    testCopyFrom(buffer, badMemory2, {}, ErrorStatus::INVALID_ARGUMENT);
+}
+
+TEST_P(MemoryDomainCopyTest, CopyFrom_InvalidMemorySize_DynamicShape) {
+    TestOperand testOperand = kTestOperand;
+    testOperand.dimensions[0] = 0;
+    auto preparedModel = createConvPreparedModel(testOperand);
+    auto [buffer, token] = allocateBuffer(preparedModel, {0}, {0});
+    if (buffer == nullptr) return;
+
+    uint32_t badMemorySize1 = kTestOperandDataSize / 2, badMemorySize2 = kTestOperandDataSize * 2;
+    Memory badMemory1 = allocateSharedMemory(badMemorySize1);
+    Memory badMemory2 = allocateSharedMemory(badMemorySize2);
+    Memory goodMemory = allocateSharedMemory(kTestOperandDataSize);
+
+    const auto goodDimensions = utils::toSigned(kTestOperand.dimensions).value();
+    auto badDimensions = goodDimensions;
+    badDimensions[0] = 2;
+
+    testCopyFrom(buffer, badMemory1, goodDimensions, ErrorStatus::INVALID_ARGUMENT);
+    testCopyFrom(buffer, badMemory2, goodDimensions, ErrorStatus::INVALID_ARGUMENT);
+    testCopyFrom(buffer, goodMemory, goodDimensions, ErrorStatus::NONE);
+    testCopyFrom(buffer, goodMemory, badDimensions, ErrorStatus::INVALID_ARGUMENT);
+}
+
+TEST_P(MemoryDomainCopyTest, CopyFrom_InvalidDimensions) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    auto [buffer, token] = allocateBuffer(preparedModel, {0}, {0});
+    if (buffer == nullptr) return;
+
+    Memory memory = allocateSharedMemory(kTestOperandDataSize);
+
+    const auto goodDimensions = utils::toSigned(kTestOperand.dimensions).value();
+    std::vector<int32_t> badDimensions = goodDimensions;
+    badDimensions.pop_back();
+    testCopyFrom(buffer, memory, badDimensions, ErrorStatus::INVALID_ARGUMENT);
+
+    badDimensions = goodDimensions;
+    badDimensions[0] = 2;
+    testCopyFrom(buffer, memory, badDimensions, ErrorStatus::INVALID_ARGUMENT);
+
+    badDimensions = goodDimensions;
+    badDimensions[0] = 0;
+    testCopyFrom(buffer, memory, badDimensions, ErrorStatus::INVALID_ARGUMENT);
+
+    testCopyFrom(buffer, memory, {}, ErrorStatus::NONE);
+    testCopyFrom(buffer, memory, goodDimensions, ErrorStatus::NONE);
+}
+
+TEST_P(MemoryDomainCopyTest, CopyFrom_InvalidDimensions_DynamicShape) {
+    TestOperand testOperand = kTestOperand;
+    testOperand.dimensions[0] = 0;
+    auto preparedModel = createConvPreparedModel(testOperand);
+    auto [buffer, token] = allocateBuffer(preparedModel, {0}, {0});
+    if (buffer == nullptr) return;
+
+    Memory memory = allocateSharedMemory(kTestOperandDataSize);
+
+    const auto goodDimensions = utils::toSigned(kTestOperand.dimensions).value();
+    std::vector<int32_t> badDimensions = goodDimensions;
+    badDimensions.pop_back();
+    testCopyFrom(buffer, memory, badDimensions, ErrorStatus::INVALID_ARGUMENT);
+
+    badDimensions = goodDimensions;
+    badDimensions[0] = 2;
+    badDimensions[3] = 4;
+    testCopyFrom(buffer, memory, badDimensions, ErrorStatus::INVALID_ARGUMENT);
+
+    badDimensions = goodDimensions;
+    badDimensions[0] = 1;
+    badDimensions[3] = 0;
+    testCopyFrom(buffer, memory, badDimensions, ErrorStatus::INVALID_ARGUMENT);
+
+    testCopyFrom(buffer, memory, {}, ErrorStatus::INVALID_ARGUMENT);
+    testCopyFrom(buffer, memory, goodDimensions, ErrorStatus::NONE);
+}
+
+TEST_P(MemoryDomainCopyTest, CopyTo_UninitializedMemory) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    auto [buffer, token] = allocateBuffer(preparedModel, {0}, {0});
+    if (buffer == nullptr) return;
+
+    Memory memory = allocateSharedMemory(kTestOperandDataSize);
+    testCopyTo(buffer, memory, ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(MemoryDomainCopyTest, CopyTo_InvalidMemorySize) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    auto [buffer, token] = allocateBuffer(preparedModel, {0}, {0});
+    if (buffer == nullptr) return;
+
+    uint32_t badMemorySize1 = kTestOperandDataSize / 2, badMemorySize2 = kTestOperandDataSize * 2;
+    Memory badMemory1 = allocateSharedMemory(badMemorySize1);
+    Memory badMemory2 = allocateSharedMemory(badMemorySize2);
+    Memory goodMemory = allocateSharedMemory(kTestOperandDataSize);
+
+    initializeDeviceMemory(buffer);
+    testCopyTo(buffer, badMemory1, ErrorStatus::INVALID_ARGUMENT);
+    testCopyTo(buffer, badMemory2, ErrorStatus::INVALID_ARGUMENT);
+    testCopyTo(buffer, goodMemory, ErrorStatus::NONE);
+}
+
+TEST_P(MemoryDomainCopyTest, CopyTo_InvalidMemorySize_DynamicShape) {
+    TestOperand testOperand = kTestOperand;
+    testOperand.dimensions[0] = 0;
+    auto preparedModel = createConvPreparedModel(testOperand);
+    auto [buffer, token] = allocateBuffer(preparedModel, {0}, {0});
+    if (buffer == nullptr) return;
+
+    uint32_t badMemorySize1 = kTestOperandDataSize / 2, badMemorySize2 = kTestOperandDataSize * 2;
+    Memory badMemory1 = allocateSharedMemory(badMemorySize1);
+    Memory badMemory2 = allocateSharedMemory(badMemorySize2);
+    Memory goodMemory = allocateSharedMemory(kTestOperandDataSize);
+
+    initializeDeviceMemory(buffer);
+    testCopyTo(buffer, badMemory1, ErrorStatus::INVALID_ARGUMENT);
+    testCopyTo(buffer, badMemory2, ErrorStatus::INVALID_ARGUMENT);
+    testCopyTo(buffer, goodMemory, ErrorStatus::NONE);
+}
+
+std::string printMemoryDomainCopyTest(
+        const testing::TestParamInfo<MemoryDomainCopyTestParam>& info) {
+    const auto& [namedDevice, operandType] = info.param;
+    const std::string type = toString(static_cast<OperandType>(operandType));
+    return gtestCompliantName(getName(namedDevice) + "_" + type);
+}
+
+GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(MemoryDomainCopyTest);
+INSTANTIATE_TEST_SUITE_P(TestMemoryDomain, MemoryDomainCopyTest,
+                         testing::Combine(testing::ValuesIn(getNamedDevices()),
+                                          kTestOperandTypeChoices),
+                         printMemoryDomainCopyTest);
+
+using MemoryDomainExecutionTestParam = std::tuple<NamedDevice, TestOperandType, Executor>;
+class MemoryDomainExecutionTest
+    : public MemoryDomainCopyTestBase,
+      public testing::WithParamInterface<MemoryDomainExecutionTestParam> {
+  protected:
+    MemoryDomainExecutionTest()
+        : MemoryDomainCopyTestBase(getData(std::get<NamedDevice>(GetParam())),
+                                   std::get<TestOperandType>(GetParam())) {}
+
+    RequestMemoryPool createSharedMemoryPool(uint32_t size) {
+        return RequestMemoryPool(allocateSharedMemory(size));
+    }
+
+    RequestMemoryPool createDeviceMemoryPool(uint32_t token) {
+        return RequestMemoryPool(static_cast<int32_t>(token));
+    }
+
+    void testExecution(const std::shared_ptr<IPreparedModel>& preparedModel, const Request& request,
+                       ErrorStatus expectedStatus) {
+        switch (kExecutor) {
+            case Executor::SYNC:
+                EXPECT_EQ(executeSync(preparedModel, request), expectedStatus);
+                break;
+            case Executor::FENCED:
+                EXPECT_EQ(executeFenced(preparedModel, request), expectedStatus);
+                break;
+            default:
+                ASSERT_TRUE(false);
+        }
+    }
+
+    ErrorStatus executeSync(const std::shared_ptr<IPreparedModel>& preparedModel,
+                            const Request& request) {
+        ExecutionResult executionResult;
+        const auto ret = preparedModel->executeSynchronously(
+                request, false, kNoDeadline, kOmittedTimeoutDuration, &executionResult);
+
+        if (!ret.isOk()) {
+            EXPECT_EQ(ret.getExceptionCode(), EX_SERVICE_SPECIFIC);
+            return static_cast<ErrorStatus>(ret.getServiceSpecificError());
+        }
+        const ErrorStatus executionStatus = executionResult.outputSufficientSize
+                                                    ? ErrorStatus::NONE
+                                                    : ErrorStatus::OUTPUT_INSUFFICIENT_SIZE;
+        EXPECT_EQ(executionResult.timing, kNoTiming);
+        return executionStatus;
+    }
+
+    ErrorStatus executeFenced(const std::shared_ptr<IPreparedModel>& preparedModel,
+                              const Request& request) {
+        ndk::ScopedFileDescriptor syncFence;
+        std::shared_ptr<IFencedExecutionCallback> fencedCallback;
+        const auto ret = preparedModel->executeFenced(request, {}, false, kNoDeadline,
+                                                      kOmittedTimeoutDuration, kNoDuration,
+                                                      &syncFence, &fencedCallback);
+        if (!ret.isOk()) {
+            EXPECT_EQ(ret.getExceptionCode(), EX_SERVICE_SPECIFIC);
+            return static_cast<ErrorStatus>(ret.getServiceSpecificError());
+        }
+        if (syncFence.get() != -1) {
+            waitForSyncFence(syncFence.get());
+        }
+        EXPECT_NE(fencedCallback, nullptr);
+
+        ErrorStatus executionStatus = ErrorStatus::GENERAL_FAILURE;
+        Timing time = kNoTiming;
+        Timing timeFenced = kNoTiming;
+        const auto retExecutionInfo =
+                fencedCallback->getExecutionInfo(&time, &timeFenced, &executionStatus);
+        EXPECT_TRUE(retExecutionInfo.isOk());
+        EXPECT_EQ(time, kNoTiming);
+        return executionStatus;
+    }
+
+    const Executor kExecutor = std::get<Executor>(GetParam());
+};
+
+TEST_P(MemoryDomainExecutionTest, InvalidToken) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    if (preparedModel == nullptr) return;
+
+    RequestMemoryPool sharedMemory = createSharedMemoryPool(kTestOperandDataSize);
+    RequestMemoryPool badDeviceMemory1 = createDeviceMemoryPool(0);    // Invalid token.
+    RequestMemoryPool badDeviceMemory2 = createDeviceMemoryPool(100);  // Unknown token.
+    RequestArgument sharedMemoryArg = {
+            .location = {.poolIndex = 0, .offset = 0, .length = kTestOperandDataSize}};
+    RequestArgument deviceMemoryArg = {.location = {.poolIndex = 1}};
+
+    testExecution(preparedModel,
+                  {.inputs = {deviceMemoryArg},
+                   .outputs = {sharedMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, badDeviceMemory1)},
+                  ErrorStatus::INVALID_ARGUMENT);
+    testExecution(preparedModel,
+                  {.inputs = {deviceMemoryArg},
+                   .outputs = {sharedMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, badDeviceMemory2)},
+                  ErrorStatus::INVALID_ARGUMENT);
+    testExecution(preparedModel,
+                  {.inputs = {sharedMemoryArg},
+                   .outputs = {deviceMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, badDeviceMemory1)},
+                  ErrorStatus::INVALID_ARGUMENT);
+    testExecution(preparedModel,
+                  {.inputs = {sharedMemoryArg},
+                   .outputs = {deviceMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, badDeviceMemory2)},
+                  ErrorStatus::INVALID_ARGUMENT);
+}
+
+TEST_P(MemoryDomainExecutionTest, InvalidPreparedModel) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    auto [buffer, token] = allocateBuffer(preparedModel, {0}, {0});
+    if (buffer == nullptr) return;
+    auto badPreparedModel = createConvPreparedModel(kTestOperand);
+    if (badPreparedModel == nullptr) return;
+
+    RequestMemoryPool sharedMemory = createSharedMemoryPool(kTestOperandDataSize);
+    RequestMemoryPool deviceMemory = createDeviceMemoryPool(token);
+    RequestArgument sharedMemoryArg = {
+            .location = {.poolIndex = 0, .offset = 0, .length = kTestOperandDataSize}};
+    RequestArgument deviceMemoryArg = {.location = {.poolIndex = 1}};
+
+    // This should fail, because the buffer is not allocated for badPreparedModel.
+    initializeDeviceMemory(buffer);
+    testExecution(badPreparedModel,
+                  {.inputs = {deviceMemoryArg},
+                   .outputs = {sharedMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, deviceMemory)},
+                  ErrorStatus::INVALID_ARGUMENT);
+    testExecution(badPreparedModel,
+                  {.inputs = {sharedMemoryArg},
+                   .outputs = {deviceMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, deviceMemory)},
+                  ErrorStatus::INVALID_ARGUMENT);
+}
+
+TEST_P(MemoryDomainExecutionTest, InvalidIOIndex) {
+    auto preparedModel = createConvPreparedModel(kTestOperand, 2);
+    auto [buffer, token] = allocateBuffer(preparedModel, {0}, {});
+    if (buffer == nullptr) return;
+
+    RequestMemoryPool sharedMemory1 = createSharedMemoryPool(kTestOperandDataSize);
+    RequestMemoryPool sharedMemory2 = createSharedMemoryPool(kTestOperandDataSize);
+    RequestMemoryPool sharedMemory3 = createSharedMemoryPool(kTestOperandDataSize);
+    RequestMemoryPool deviceMemory = createDeviceMemoryPool(token);
+    RequestArgument sharedMemoryArg1 = {
+            .location = {.poolIndex = 0, .offset = 0, .length = kTestOperandDataSize}};
+    RequestArgument sharedMemoryArg2 = {
+            .location = {.poolIndex = 1, .offset = 0, .length = kTestOperandDataSize}};
+    RequestArgument sharedMemoryArg3 = {
+            .location = {.poolIndex = 2, .offset = 0, .length = kTestOperandDataSize}};
+    RequestArgument deviceMemoryArg = {.location = {.poolIndex = 3}};
+
+    // This should fail, because the device memory is not allocated for input 1.
+    initializeDeviceMemory(buffer);
+    testExecution(preparedModel,
+                  {.inputs = {sharedMemoryArg1, deviceMemoryArg},
+                   .outputs = {sharedMemoryArg2, sharedMemoryArg3},
+                   .pools = createRequestMemoryPools(sharedMemory1, sharedMemory2, sharedMemory3,
+                                                     deviceMemory)},
+                  ErrorStatus::INVALID_ARGUMENT);
+
+    // This should fail, because the device memory is not allocated for output 1.
+    testExecution(preparedModel,
+                  {.inputs = {sharedMemoryArg1, sharedMemoryArg2},
+                   .outputs = {sharedMemoryArg3, deviceMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory1, sharedMemory2, sharedMemory3,
+                                                     deviceMemory)},
+                  ErrorStatus::INVALID_ARGUMENT);
+}
+
+TEST_P(MemoryDomainExecutionTest, InvalidIOType) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    auto [inputBuffer, inputToken] = allocateBuffer(preparedModel, {0}, {});
+    auto [outputBuffer, outputToken] = allocateBuffer(preparedModel, {}, {0});
+    if (inputBuffer == nullptr || outputBuffer == nullptr) return;
+
+    RequestMemoryPool sharedMemory = createSharedMemoryPool(kTestOperandDataSize);
+    RequestMemoryPool deviceMemory = createDeviceMemoryPool(inputToken);
+    RequestArgument sharedMemoryArg = {
+            .location = {.poolIndex = 0, .offset = 0, .length = kTestOperandDataSize}};
+    RequestArgument deviceMemoryArg = {.location = {.poolIndex = 1}};
+
+    // This should fail, because the device memory is allocated for input but used as output.
+    testExecution(preparedModel,
+                  {.inputs = {sharedMemoryArg},
+                   .outputs = {deviceMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, deviceMemory)},
+                  ErrorStatus::INVALID_ARGUMENT);
+
+    // This should fail, because the device memory is allocated for output but used as input.
+    deviceMemory.set<RequestMemoryPool::Tag::token>(outputToken);
+    initializeDeviceMemory(outputBuffer);
+    testExecution(preparedModel,
+                  {.inputs = {deviceMemoryArg},
+                   .outputs = {sharedMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, deviceMemory)},
+                  ErrorStatus::INVALID_ARGUMENT);
+}
+
+TEST_P(MemoryDomainExecutionTest, UninitializedMemory) {
+    auto preparedModel = createConvPreparedModel(kTestOperand);
+    auto [buffer, token] = allocateBuffer(preparedModel, {0}, {0});
+    if (buffer == nullptr) return;
+
+    RequestMemoryPool sharedMemory = createSharedMemoryPool(kTestOperandDataSize);
+    RequestMemoryPool deviceMemory = createDeviceMemoryPool(token);
+    RequestArgument sharedMemoryArg = {
+            .location = {.poolIndex = 0, .offset = 0, .length = kTestOperandDataSize}};
+    RequestArgument deviceMemoryArg = {.location = {.poolIndex = 1}};
+
+    // This should fail, because the device memory is not initialized.
+    testExecution(preparedModel,
+                  {.inputs = {deviceMemoryArg},
+                   .outputs = {sharedMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, deviceMemory)},
+                  ErrorStatus::GENERAL_FAILURE);
+
+    // This should initialize the device memory.
+    testExecution(preparedModel,
+                  {.inputs = {sharedMemoryArg},
+                   .outputs = {deviceMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, deviceMemory)},
+                  ErrorStatus::NONE);
+
+    // Test again with initialized device memory.
+    testExecution(preparedModel,
+                  {.inputs = {deviceMemoryArg},
+                   .outputs = {sharedMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, deviceMemory)},
+                  ErrorStatus::NONE);
+}
+
+TEST_P(MemoryDomainExecutionTest, SameRequestMultipleRoles) {
+    auto preparedModel = createConvPreparedModel(kTestOperand, 2);
+    auto [buffer, token] = allocateBuffer(preparedModel, {0, 1}, {0, 1});
+    if (buffer == nullptr) return;
+
+    RequestMemoryPool sharedMemory1 = createSharedMemoryPool(kTestOperandDataSize);
+    RequestMemoryPool sharedMemory2 = createSharedMemoryPool(kTestOperandDataSize);
+    RequestMemoryPool deviceMemory = createDeviceMemoryPool(token);
+    RequestArgument sharedMemoryArg1 = {
+            .location = {.poolIndex = 0, .offset = 0, .length = kTestOperandDataSize}};
+    RequestArgument sharedMemoryArg2 = {
+            .location = {.poolIndex = 1, .offset = 0, .length = kTestOperandDataSize}};
+    RequestArgument deviceMemoryArg = {.location = {.poolIndex = 2}};
+
+    // This should fail, because the same device memory cannot be used for both input and output.
+    initializeDeviceMemory(buffer);
+    testExecution(preparedModel,
+                  {.inputs = {deviceMemoryArg, sharedMemoryArg1},
+                   .outputs = {deviceMemoryArg, sharedMemoryArg2},
+                   .pools = createRequestMemoryPools(sharedMemory1, sharedMemory2, deviceMemory)},
+                  ErrorStatus::INVALID_ARGUMENT);
+
+    // This should fail, because the same device memory cannot be used for multiple outputs.
+    testExecution(preparedModel,
+                  {.inputs = {sharedMemoryArg1, sharedMemoryArg2},
+                   .outputs = {deviceMemoryArg, deviceMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory1, sharedMemory2, deviceMemory)},
+                  ErrorStatus::INVALID_ARGUMENT);
+
+    // The same device memory can be used for multiple inputs.
+    initializeDeviceMemory(buffer);
+    testExecution(preparedModel,
+                  {.inputs = {deviceMemoryArg, deviceMemoryArg},
+                   .outputs = {sharedMemoryArg1, sharedMemoryArg2},
+                   .pools = createRequestMemoryPools(sharedMemory1, sharedMemory2, deviceMemory)},
+                  ErrorStatus::NONE);
+}
+
+TEST_P(MemoryDomainExecutionTest, InvalidDimensions) {
+    // FENCED execution does not support dynamic shape.
+    if (kExecutor == Executor::FENCED) return;
+
+    TestOperand testOperand = kTestOperand;
+    testOperand.dimensions[0] = 0;
+    auto preparedModel = createConvPreparedModel(testOperand);
+    auto deviceBuffer = allocateBuffer(preparedModel, {0}, {0},
+                                       utils::toSigned(kTestOperand.dimensions).value());
+    if (deviceBuffer.buffer == nullptr) return;
+
+    RequestMemoryPool sharedMemory = createSharedMemoryPool(kTestOperandDataSize);
+    RequestMemoryPool deviceMemory = createDeviceMemoryPool(deviceBuffer.token);
+    auto badDimensions = utils::toSigned(kTestOperand.dimensions).value();
+    badDimensions[0] = 2;
+    RequestArgument sharedMemoryArg = {
+            .location = {.poolIndex = 0, .offset = 0, .length = kTestOperandDataSize},
+            .dimensions = badDimensions};
+    RequestArgument deviceMemoryArg = {.location = {.poolIndex = 1}};
+    RequestArgument deviceMemoryArgWithBadDimensions = {.location = {.poolIndex = 1},
+                                                        .dimensions = badDimensions};
+
+    initializeDeviceMemory(deviceBuffer.buffer);
+    testExecution(preparedModel,
+                  {.inputs = {deviceMemoryArgWithBadDimensions},
+                   .outputs = {sharedMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, deviceMemory)},
+                  ErrorStatus::INVALID_ARGUMENT);
+
+    testExecution(preparedModel,
+                  {.inputs = {sharedMemoryArg},
+                   .outputs = {deviceMemoryArgWithBadDimensions},
+                   .pools = createRequestMemoryPools(sharedMemory, deviceMemory)},
+                  ErrorStatus::INVALID_ARGUMENT);
+
+    testExecution(preparedModel,
+                  {.inputs = {sharedMemoryArg},
+                   .outputs = {deviceMemoryArg},
+                   .pools = createRequestMemoryPools(sharedMemory, deviceMemory)},
+                  ErrorStatus::GENERAL_FAILURE);
+}
+
+const auto kExecutorChoices = testing::Values(Executor::SYNC, Executor::FENCED);
+
+std::string printMemoryDomainExecutionTest(
+        const testing::TestParamInfo<MemoryDomainExecutionTestParam>& info) {
+    const auto& [namedDevice, operandType, executor] = info.param;
+    const std::string type = toString(static_cast<OperandType>(operandType));
+    const std::string executorStr = toString(executor);
+    return gtestCompliantName(getName(namedDevice) + "_" + type + "_" + executorStr);
+}
+
+GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(MemoryDomainExecutionTest);
+INSTANTIATE_TEST_SUITE_P(TestMemoryDomain, MemoryDomainExecutionTest,
+                         testing::Combine(testing::ValuesIn(getNamedDevices()),
+                                          kTestOperandTypeChoices, kExecutorChoices),
+                         printMemoryDomainExecutionTest);
+
+}  // namespace aidl::android::hardware::neuralnetworks::vts::functional
diff --git a/neuralnetworks/aidl/vts/functional/QualityOfServiceTests.cpp b/neuralnetworks/aidl/vts/functional/QualityOfServiceTests.cpp
new file mode 100644
index 0000000..58db98f
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/QualityOfServiceTests.cpp
@@ -0,0 +1,270 @@
+/*
+ * 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 <android/binder_enums.h>
+#include <android/binder_interface_utils.h>
+#include <android/binder_status.h>
+
+#include <nnapi/hal/aidl/Conversions.h>
+
+#include "Callbacks.h"
+#include "GeneratedTestHarness.h"
+#include "Utils.h"
+
+namespace aidl::android::hardware::neuralnetworks::vts::functional {
+
+using implementation::PreparedModelCallback;
+using test_helper::TestBuffer;
+using test_helper::TestModel;
+
+enum class DeadlineBoundType { NOW, UNLIMITED, SHORT };
+constexpr std::array<DeadlineBoundType, 3> deadlineBounds = {
+        DeadlineBoundType::NOW, DeadlineBoundType::UNLIMITED, DeadlineBoundType::SHORT};
+std::string toString(DeadlineBoundType type) {
+    switch (type) {
+        case DeadlineBoundType::NOW:
+            return "NOW";
+        case DeadlineBoundType::UNLIMITED:
+            return "UNLIMITED";
+        case DeadlineBoundType::SHORT:
+            return "SHORT";
+    }
+    LOG(FATAL) << "Unrecognized DeadlineBoundType: " << static_cast<int>(type);
+    return {};
+}
+
+constexpr auto kShortDuration = std::chrono::milliseconds{5};
+
+using Results = std::tuple<ErrorStatus, std::vector<OutputShape>, Timing>;
+using MaybeResults = std::optional<Results>;
+
+static int64_t makeDeadline(DeadlineBoundType deadlineBoundType) {
+    const auto getNanosecondsSinceEpoch = [](const auto& time) -> int64_t {
+        const auto timeSinceEpoch = time.time_since_epoch();
+        return std::chrono::duration_cast<std::chrono::nanoseconds>(timeSinceEpoch).count();
+    };
+
+    std::chrono::steady_clock::time_point timePoint;
+    switch (deadlineBoundType) {
+        case DeadlineBoundType::NOW:
+            timePoint = std::chrono::steady_clock::now();
+            break;
+        case DeadlineBoundType::UNLIMITED:
+            timePoint = std::chrono::steady_clock::time_point::max();
+            break;
+        case DeadlineBoundType::SHORT:
+            timePoint = std::chrono::steady_clock::now() + kShortDuration;
+            break;
+    }
+
+    return getNanosecondsSinceEpoch(timePoint);
+}
+
+void runPrepareModelTest(const std::shared_ptr<IDevice>& device, const Model& model,
+                         Priority priority, std::optional<DeadlineBoundType> deadlineBound) {
+    int64_t deadline = kNoDeadline;
+    if (deadlineBound.has_value()) {
+        deadline = makeDeadline(deadlineBound.value());
+    }
+
+    // see if service can handle model
+    std::vector<bool> supportedOps;
+    const auto supportedCallStatus = device->getSupportedOperations(model, &supportedOps);
+    ASSERT_TRUE(supportedCallStatus.isOk());
+    ASSERT_NE(0ul, supportedOps.size());
+    const bool fullySupportsModel =
+            std::all_of(supportedOps.begin(), supportedOps.end(), [](bool valid) { return valid; });
+
+    // launch prepare model
+    const std::shared_ptr<PreparedModelCallback> preparedModelCallback =
+            ndk::SharedRefBase::make<PreparedModelCallback>();
+    const auto prepareLaunchStatus =
+            device->prepareModel(model, ExecutionPreference::FAST_SINGLE_ANSWER, priority, deadline,
+                                 {}, {}, kEmptyCacheToken, preparedModelCallback);
+    ASSERT_TRUE(prepareLaunchStatus.isOk())
+            << "prepareLaunchStatus: " << prepareLaunchStatus.getDescription();
+
+    // retrieve prepared model
+    preparedModelCallback->wait();
+    const ErrorStatus prepareReturnStatus = preparedModelCallback->getStatus();
+    const std::shared_ptr<IPreparedModel> preparedModel = preparedModelCallback->getPreparedModel();
+
+    // The getSupportedOperations call returns a list of operations that are guaranteed not to fail
+    // if prepareModel is called, and 'fullySupportsModel' is true i.f.f. the entire model is
+    // guaranteed. If a driver has any doubt that it can prepare an operation, it must return false.
+    // So here, if a driver isn't sure if it can support an operation, but reports that it
+    // successfully prepared the model, the test can continue.
+    if (!fullySupportsModel && prepareReturnStatus != ErrorStatus::NONE) {
+        ASSERT_EQ(nullptr, preparedModel.get());
+        return;
+    }
+
+    // verify return status
+    if (!deadlineBound.has_value()) {
+        EXPECT_EQ(ErrorStatus::NONE, prepareReturnStatus);
+    } else {
+        switch (deadlineBound.value()) {
+            case DeadlineBoundType::NOW:
+            case DeadlineBoundType::SHORT:
+                // Either the driver successfully completed the task or it
+                // aborted and returned MISSED_DEADLINE_*.
+                EXPECT_TRUE(prepareReturnStatus == ErrorStatus::NONE ||
+                            prepareReturnStatus == ErrorStatus::MISSED_DEADLINE_TRANSIENT ||
+                            prepareReturnStatus == ErrorStatus::MISSED_DEADLINE_PERSISTENT);
+                break;
+            case DeadlineBoundType::UNLIMITED:
+                // If an unlimited deadline is supplied, we expect the execution to
+                // proceed normally. In this case, check it normally by breaking out
+                // of the switch statement.
+                EXPECT_EQ(ErrorStatus::NONE, prepareReturnStatus);
+                break;
+        }
+    }
+    ASSERT_EQ(prepareReturnStatus == ErrorStatus::NONE, preparedModel.get() != nullptr);
+}
+
+void runPrepareModelTests(const std::shared_ptr<IDevice>& device, const Model& model) {
+    // test priority
+    for (auto priority : ndk::enum_range<Priority>{}) {
+        SCOPED_TRACE("priority: " + toString(priority));
+        if (priority == kDefaultPriority) continue;
+        runPrepareModelTest(device, model, priority, {});
+    }
+
+    // test deadline
+    for (auto deadlineBound : deadlineBounds) {
+        SCOPED_TRACE("deadlineBound: " + toString(deadlineBound));
+        runPrepareModelTest(device, model, kDefaultPriority, deadlineBound);
+    }
+}
+
+static MaybeResults executeSynchronously(const std::shared_ptr<IPreparedModel>& preparedModel,
+                                         const Request& request, int64_t deadline) {
+    SCOPED_TRACE("synchronous");
+    const bool measure = false;
+
+    // run execution
+    ExecutionResult executionResult;
+    const auto ret = preparedModel->executeSynchronously(request, measure, deadline,
+                                                         kOmittedTimeoutDuration, &executionResult);
+    EXPECT_TRUE(ret.isOk() || ret.getExceptionCode() == EX_SERVICE_SPECIFIC)
+            << ret.getDescription();
+    if (!ret.isOk()) {
+        if (ret.getExceptionCode() != EX_SERVICE_SPECIFIC) {
+            return std::nullopt;
+        }
+        return MaybeResults(
+                {static_cast<ErrorStatus>(ret.getServiceSpecificError()), {}, kNoTiming});
+    }
+
+    // return results
+    return MaybeResults({executionResult.outputSufficientSize
+                                 ? ErrorStatus::NONE
+                                 : ErrorStatus::OUTPUT_INSUFFICIENT_SIZE,
+                         std::move(executionResult.outputShapes), executionResult.timing});
+}
+
+void runExecutionTest(const std::shared_ptr<IPreparedModel>& preparedModel,
+                      const TestModel& testModel, const Request& request,
+                      const ExecutionContext& context, DeadlineBoundType deadlineBound) {
+    const auto deadline = makeDeadline(deadlineBound);
+
+    // Perform execution and unpack results.
+    const auto results = executeSynchronously(preparedModel, request, deadline);
+    if (!results.has_value()) return;
+    const auto& [status, outputShapes, timing] = results.value();
+
+    // Verify no timing information was returned
+    EXPECT_EQ(timing, kNoTiming);
+
+    // Validate deadline information if applicable.
+    switch (deadlineBound) {
+        case DeadlineBoundType::NOW:
+        case DeadlineBoundType::SHORT:
+            // Either the driver successfully completed the task or it
+            // aborted and returned MISSED_DEADLINE_*.
+            ASSERT_TRUE(status == ErrorStatus::NONE ||
+                        status == ErrorStatus::MISSED_DEADLINE_TRANSIENT ||
+                        status == ErrorStatus::MISSED_DEADLINE_PERSISTENT);
+            break;
+        case DeadlineBoundType::UNLIMITED:
+            // If an unlimited deadline is supplied, we expect the execution to
+            // proceed normally. In this case, check it normally by breaking out
+            // of the switch statement.
+            ASSERT_EQ(ErrorStatus::NONE, status);
+            break;
+    }
+
+    // If the model output operands are fully specified, outputShapes must be either
+    // either empty, or have the same number of elements as the number of outputs.
+    ASSERT_TRUE(outputShapes.size() == 0 ||
+                outputShapes.size() == testModel.main.outputIndexes.size());
+
+    // Go through all outputs, check returned output shapes.
+    for (uint32_t i = 0; i < outputShapes.size(); i++) {
+        EXPECT_TRUE(outputShapes[i].isSufficient);
+        const auto expect =
+                utils::toSigned(testModel.main.operands[testModel.main.outputIndexes[i]].dimensions)
+                        .value();
+        const std::vector<int32_t>& actual = outputShapes[i].dimensions;
+        EXPECT_EQ(expect, actual);
+    }
+
+    // Retrieve execution results.
+    const std::vector<TestBuffer> outputs = context.getOutputBuffers(request);
+
+    // We want "close-enough" results.
+    if (status == ErrorStatus::NONE) {
+        checkResults(testModel, outputs);
+    }
+}
+
+void runExecutionTests(const std::shared_ptr<IPreparedModel>& preparedModel,
+                       const TestModel& testModel, const Request& request,
+                       const ExecutionContext& context) {
+    for (auto deadlineBound : deadlineBounds) {
+        runExecutionTest(preparedModel, testModel, request, context, deadlineBound);
+    }
+}
+
+void runTests(const std::shared_ptr<IDevice>& device, const TestModel& testModel) {
+    // setup
+    const Model model = createModel(testModel);
+
+    // run prepare model tests
+    runPrepareModelTests(device, model);
+
+    // prepare model
+    std::shared_ptr<IPreparedModel> preparedModel;
+    createPreparedModel(device, model, &preparedModel);
+    if (preparedModel == nullptr) return;
+
+    // run execution tests
+    ExecutionContext context;
+    const Request request = context.createRequest(testModel);
+    runExecutionTests(preparedModel, testModel, request, context);
+}
+
+class DeadlineTest : public GeneratedTestBase {};
+
+TEST_P(DeadlineTest, Test) {
+    runTests(kDevice, kTestModel);
+}
+
+INSTANTIATE_GENERATED_TEST(DeadlineTest,
+                           [](const TestModel& testModel) { return !testModel.expectFailure; });
+
+}  // namespace aidl::android::hardware::neuralnetworks::vts::functional
diff --git a/neuralnetworks/aidl/vts/functional/TestAssertions.cpp b/neuralnetworks/aidl/vts/functional/TestAssertions.cpp
new file mode 100644
index 0000000..a9e9456
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/TestAssertions.cpp
@@ -0,0 +1,153 @@
+/*
+ * 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 <aidl/android/hardware/neuralnetworks/IPreparedModel.h>
+#include <aidl/android/hardware/neuralnetworks/OperandType.h>
+#include <aidl/android/hardware/neuralnetworks/OperationType.h>
+
+#include <ControlFlow.h>
+#include <TestHarness.h>
+
+namespace aidl::android::hardware::neuralnetworks {
+
+namespace nn = ::android::nn;
+
+static_assert(static_cast<uint64_t>(IPreparedModel::DEFAULT_LOOP_TIMEOUT_DURATION_NS) ==
+              nn::operation_while::kTimeoutNsDefault);
+static_assert(static_cast<uint64_t>(IPreparedModel::MAXIMUM_LOOP_TIMEOUT_DURATION_NS) ==
+              nn::operation_while::kTimeoutNsMaximum);
+
+// Make sure that the HIDL enums are compatible with the values defined in
+// frameworks/ml/nn/tools/test_generator/test_harness/include/TestHarness.h.
+using namespace test_helper;
+#define CHECK_TEST_ENUM(EnumType, enumValue) \
+    static_assert(static_cast<EnumType>(Test##EnumType::enumValue) == EnumType::enumValue)
+
+CHECK_TEST_ENUM(OperandType, FLOAT32);
+CHECK_TEST_ENUM(OperandType, INT32);
+CHECK_TEST_ENUM(OperandType, UINT32);
+CHECK_TEST_ENUM(OperandType, TENSOR_FLOAT32);
+CHECK_TEST_ENUM(OperandType, TENSOR_INT32);
+CHECK_TEST_ENUM(OperandType, TENSOR_QUANT8_ASYMM);
+CHECK_TEST_ENUM(OperandType, BOOL);
+CHECK_TEST_ENUM(OperandType, TENSOR_QUANT16_SYMM);
+CHECK_TEST_ENUM(OperandType, TENSOR_FLOAT16);
+CHECK_TEST_ENUM(OperandType, TENSOR_BOOL8);
+CHECK_TEST_ENUM(OperandType, FLOAT16);
+CHECK_TEST_ENUM(OperandType, TENSOR_QUANT8_SYMM_PER_CHANNEL);
+CHECK_TEST_ENUM(OperandType, TENSOR_QUANT16_ASYMM);
+CHECK_TEST_ENUM(OperandType, TENSOR_QUANT8_SYMM);
+CHECK_TEST_ENUM(OperandType, TENSOR_QUANT8_ASYMM_SIGNED);
+
+CHECK_TEST_ENUM(OperationType, ADD);
+CHECK_TEST_ENUM(OperationType, AVERAGE_POOL_2D);
+CHECK_TEST_ENUM(OperationType, CONCATENATION);
+CHECK_TEST_ENUM(OperationType, CONV_2D);
+CHECK_TEST_ENUM(OperationType, DEPTHWISE_CONV_2D);
+CHECK_TEST_ENUM(OperationType, DEPTH_TO_SPACE);
+CHECK_TEST_ENUM(OperationType, DEQUANTIZE);
+CHECK_TEST_ENUM(OperationType, EMBEDDING_LOOKUP);
+CHECK_TEST_ENUM(OperationType, FLOOR);
+CHECK_TEST_ENUM(OperationType, FULLY_CONNECTED);
+CHECK_TEST_ENUM(OperationType, HASHTABLE_LOOKUP);
+CHECK_TEST_ENUM(OperationType, L2_NORMALIZATION);
+CHECK_TEST_ENUM(OperationType, L2_POOL_2D);
+CHECK_TEST_ENUM(OperationType, LOCAL_RESPONSE_NORMALIZATION);
+CHECK_TEST_ENUM(OperationType, LOGISTIC);
+CHECK_TEST_ENUM(OperationType, LSH_PROJECTION);
+CHECK_TEST_ENUM(OperationType, LSTM);
+CHECK_TEST_ENUM(OperationType, MAX_POOL_2D);
+CHECK_TEST_ENUM(OperationType, MUL);
+CHECK_TEST_ENUM(OperationType, RELU);
+CHECK_TEST_ENUM(OperationType, RELU1);
+CHECK_TEST_ENUM(OperationType, RELU6);
+CHECK_TEST_ENUM(OperationType, RESHAPE);
+CHECK_TEST_ENUM(OperationType, RESIZE_BILINEAR);
+CHECK_TEST_ENUM(OperationType, RNN);
+CHECK_TEST_ENUM(OperationType, SOFTMAX);
+CHECK_TEST_ENUM(OperationType, SPACE_TO_DEPTH);
+CHECK_TEST_ENUM(OperationType, SVDF);
+CHECK_TEST_ENUM(OperationType, TANH);
+CHECK_TEST_ENUM(OperationType, BATCH_TO_SPACE_ND);
+CHECK_TEST_ENUM(OperationType, DIV);
+CHECK_TEST_ENUM(OperationType, MEAN);
+CHECK_TEST_ENUM(OperationType, PAD);
+CHECK_TEST_ENUM(OperationType, SPACE_TO_BATCH_ND);
+CHECK_TEST_ENUM(OperationType, SQUEEZE);
+CHECK_TEST_ENUM(OperationType, STRIDED_SLICE);
+CHECK_TEST_ENUM(OperationType, SUB);
+CHECK_TEST_ENUM(OperationType, TRANSPOSE);
+CHECK_TEST_ENUM(OperationType, ABS);
+CHECK_TEST_ENUM(OperationType, ARGMAX);
+CHECK_TEST_ENUM(OperationType, ARGMIN);
+CHECK_TEST_ENUM(OperationType, AXIS_ALIGNED_BBOX_TRANSFORM);
+CHECK_TEST_ENUM(OperationType, BIDIRECTIONAL_SEQUENCE_LSTM);
+CHECK_TEST_ENUM(OperationType, BIDIRECTIONAL_SEQUENCE_RNN);
+CHECK_TEST_ENUM(OperationType, BOX_WITH_NMS_LIMIT);
+CHECK_TEST_ENUM(OperationType, CAST);
+CHECK_TEST_ENUM(OperationType, CHANNEL_SHUFFLE);
+CHECK_TEST_ENUM(OperationType, DETECTION_POSTPROCESSING);
+CHECK_TEST_ENUM(OperationType, EQUAL);
+CHECK_TEST_ENUM(OperationType, EXP);
+CHECK_TEST_ENUM(OperationType, EXPAND_DIMS);
+CHECK_TEST_ENUM(OperationType, GATHER);
+CHECK_TEST_ENUM(OperationType, GENERATE_PROPOSALS);
+CHECK_TEST_ENUM(OperationType, GREATER);
+CHECK_TEST_ENUM(OperationType, GREATER_EQUAL);
+CHECK_TEST_ENUM(OperationType, GROUPED_CONV_2D);
+CHECK_TEST_ENUM(OperationType, HEATMAP_MAX_KEYPOINT);
+CHECK_TEST_ENUM(OperationType, INSTANCE_NORMALIZATION);
+CHECK_TEST_ENUM(OperationType, LESS);
+CHECK_TEST_ENUM(OperationType, LESS_EQUAL);
+CHECK_TEST_ENUM(OperationType, LOG);
+CHECK_TEST_ENUM(OperationType, LOGICAL_AND);
+CHECK_TEST_ENUM(OperationType, LOGICAL_NOT);
+CHECK_TEST_ENUM(OperationType, LOGICAL_OR);
+CHECK_TEST_ENUM(OperationType, LOG_SOFTMAX);
+CHECK_TEST_ENUM(OperationType, MAXIMUM);
+CHECK_TEST_ENUM(OperationType, MINIMUM);
+CHECK_TEST_ENUM(OperationType, NEG);
+CHECK_TEST_ENUM(OperationType, NOT_EQUAL);
+CHECK_TEST_ENUM(OperationType, PAD_V2);
+CHECK_TEST_ENUM(OperationType, POW);
+CHECK_TEST_ENUM(OperationType, PRELU);
+CHECK_TEST_ENUM(OperationType, QUANTIZE);
+CHECK_TEST_ENUM(OperationType, QUANTIZED_16BIT_LSTM);
+CHECK_TEST_ENUM(OperationType, RANDOM_MULTINOMIAL);
+CHECK_TEST_ENUM(OperationType, REDUCE_ALL);
+CHECK_TEST_ENUM(OperationType, REDUCE_ANY);
+CHECK_TEST_ENUM(OperationType, REDUCE_MAX);
+CHECK_TEST_ENUM(OperationType, REDUCE_MIN);
+CHECK_TEST_ENUM(OperationType, REDUCE_PROD);
+CHECK_TEST_ENUM(OperationType, REDUCE_SUM);
+CHECK_TEST_ENUM(OperationType, ROI_ALIGN);
+CHECK_TEST_ENUM(OperationType, ROI_POOLING);
+CHECK_TEST_ENUM(OperationType, RSQRT);
+CHECK_TEST_ENUM(OperationType, SELECT);
+CHECK_TEST_ENUM(OperationType, SIN);
+CHECK_TEST_ENUM(OperationType, SLICE);
+CHECK_TEST_ENUM(OperationType, SPLIT);
+CHECK_TEST_ENUM(OperationType, SQRT);
+CHECK_TEST_ENUM(OperationType, TILE);
+CHECK_TEST_ENUM(OperationType, TOPK_V2);
+CHECK_TEST_ENUM(OperationType, TRANSPOSE_CONV_2D);
+CHECK_TEST_ENUM(OperationType, UNIDIRECTIONAL_SEQUENCE_LSTM);
+CHECK_TEST_ENUM(OperationType, UNIDIRECTIONAL_SEQUENCE_RNN);
+CHECK_TEST_ENUM(OperationType, RESIZE_NEAREST_NEIGHBOR);
+
+#undef CHECK_TEST_ENUM
+
+}  // namespace aidl::android::hardware::neuralnetworks
diff --git a/neuralnetworks/aidl/vts/functional/TestMain.cpp b/neuralnetworks/aidl/vts/functional/TestMain.cpp
new file mode 100644
index 0000000..1d58608
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/TestMain.cpp
@@ -0,0 +1,27 @@
+/*
+ * 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 <android/binder_process.h>
+#include <gtest/gtest.h>
+#include "LogTestCaseToLogcat.h"
+
+int main(int argc, char** argv) {
+    testing::InitGoogleTest(&argc, argv);
+    testing::UnitTest::GetInstance()->listeners().Append(
+            new aidl::android::hardware::neuralnetworks::LogTestCaseToLogcat());
+    ABinderProcess_startThreadPool();
+    return RUN_ALL_TESTS();
+}
diff --git a/neuralnetworks/aidl/vts/functional/Utils.cpp b/neuralnetworks/aidl/vts/functional/Utils.cpp
new file mode 100644
index 0000000..14a496a
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/Utils.cpp
@@ -0,0 +1,252 @@
+/*
+ * 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 "Utils.h"
+
+#include <aidl/android/hardware/neuralnetworks/IPreparedModelParcel.h>
+#include <aidl/android/hardware/neuralnetworks/Operand.h>
+#include <aidl/android/hardware/neuralnetworks/OperandType.h>
+#include <android-base/logging.h>
+#include <android/binder_status.h>
+#include <android/hardware_buffer.h>
+
+#include <iostream>
+#include <limits>
+#include <numeric>
+
+#include <MemoryUtils.h>
+#include <nnapi/SharedMemory.h>
+#include <nnapi/hal/aidl/Conversions.h>
+#include <nnapi/hal/aidl/Utils.h>
+
+namespace aidl::android::hardware::neuralnetworks {
+
+using test_helper::TestBuffer;
+using test_helper::TestModel;
+
+uint32_t sizeOfData(OperandType type) {
+    switch (type) {
+        case OperandType::FLOAT32:
+        case OperandType::INT32:
+        case OperandType::UINT32:
+        case OperandType::TENSOR_FLOAT32:
+        case OperandType::TENSOR_INT32:
+            return 4;
+        case OperandType::TENSOR_QUANT16_SYMM:
+        case OperandType::TENSOR_FLOAT16:
+        case OperandType::FLOAT16:
+        case OperandType::TENSOR_QUANT16_ASYMM:
+            return 2;
+        case OperandType::TENSOR_QUANT8_ASYMM:
+        case OperandType::BOOL:
+        case OperandType::TENSOR_BOOL8:
+        case OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL:
+        case OperandType::TENSOR_QUANT8_SYMM:
+        case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
+            return 1;
+        case OperandType::SUBGRAPH:
+            return 0;
+        default:
+            CHECK(false) << "Invalid OperandType " << static_cast<uint32_t>(type);
+            return 0;
+    }
+}
+
+static bool isTensor(OperandType type) {
+    switch (type) {
+        case OperandType::FLOAT32:
+        case OperandType::INT32:
+        case OperandType::UINT32:
+        case OperandType::FLOAT16:
+        case OperandType::BOOL:
+        case OperandType::SUBGRAPH:
+            return false;
+        case OperandType::TENSOR_FLOAT32:
+        case OperandType::TENSOR_INT32:
+        case OperandType::TENSOR_QUANT16_SYMM:
+        case OperandType::TENSOR_FLOAT16:
+        case OperandType::TENSOR_QUANT16_ASYMM:
+        case OperandType::TENSOR_QUANT8_ASYMM:
+        case OperandType::TENSOR_BOOL8:
+        case OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL:
+        case OperandType::TENSOR_QUANT8_SYMM:
+        case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
+            return true;
+        default:
+            CHECK(false) << "Invalid OperandType " << static_cast<uint32_t>(type);
+            return false;
+    }
+}
+
+uint32_t sizeOfData(const Operand& operand) {
+    const uint32_t dataSize = sizeOfData(operand.type);
+    if (isTensor(operand.type) && operand.dimensions.size() == 0) return 0;
+    return std::accumulate(operand.dimensions.begin(), operand.dimensions.end(), dataSize,
+                           std::multiplies<>{});
+}
+
+std::unique_ptr<TestAshmem> TestAshmem::create(uint32_t size) {
+    auto ashmem = std::make_unique<TestAshmem>(size);
+    return ashmem->mIsValid ? std::move(ashmem) : nullptr;
+}
+
+void TestAshmem::initialize(uint32_t size) {
+    mIsValid = false;
+    ASSERT_GT(size, 0);
+    const auto sharedMemory = nn::createSharedMemory(size).value();
+    mMappedMemory = nn::map(sharedMemory).value();
+    mPtr = static_cast<uint8_t*>(std::get<void*>(mMappedMemory.pointer));
+    CHECK_NE(mPtr, nullptr);
+    mAidlMemory = utils::convert(sharedMemory).value();
+    mIsValid = true;
+}
+
+std::unique_ptr<TestBlobAHWB> TestBlobAHWB::create(uint32_t size) {
+    auto ahwb = std::make_unique<TestBlobAHWB>(size);
+    return ahwb->mIsValid ? std::move(ahwb) : nullptr;
+}
+
+void TestBlobAHWB::initialize(uint32_t size) {
+    mIsValid = false;
+    ASSERT_GT(size, 0);
+    const auto usage = AHARDWAREBUFFER_USAGE_CPU_READ_OFTEN | AHARDWAREBUFFER_USAGE_CPU_WRITE_OFTEN;
+    const AHardwareBuffer_Desc desc = {
+            .width = size,
+            .height = 1,
+            .layers = 1,
+            .format = AHARDWAREBUFFER_FORMAT_BLOB,
+            .usage = usage,
+            .stride = size,
+    };
+
+    ASSERT_EQ(AHardwareBuffer_allocate(&desc, &mAhwb), 0);
+    ASSERT_NE(mAhwb, nullptr);
+
+    const auto sharedMemory = nn::createSharedMemoryFromAHWB(*mAhwb).value();
+    mMapping = nn::map(sharedMemory).value();
+    mPtr = static_cast<uint8_t*>(std::get<void*>(mMapping.pointer));
+    CHECK_NE(mPtr, nullptr);
+    mAidlMemory = utils::convert(sharedMemory).value();
+
+    mIsValid = true;
+}
+
+TestBlobAHWB::~TestBlobAHWB() {
+    if (mAhwb) {
+        AHardwareBuffer_unlock(mAhwb, nullptr);
+        AHardwareBuffer_release(mAhwb);
+    }
+}
+
+std::string gtestCompliantName(std::string name) {
+    // gtest test names must only contain alphanumeric characters
+    std::replace_if(
+            name.begin(), name.end(), [](char c) { return !std::isalnum(c); }, '_');
+    return name;
+}
+
+::std::ostream& operator<<(::std::ostream& os, ErrorStatus errorStatus) {
+    return os << toString(errorStatus);
+}
+
+Request ExecutionContext::createRequest(const TestModel& testModel, MemoryType memoryType) {
+    CHECK(memoryType == MemoryType::ASHMEM || memoryType == MemoryType::BLOB_AHWB);
+
+    // Model inputs.
+    std::vector<RequestArgument> inputs(testModel.main.inputIndexes.size());
+    size_t inputSize = 0;
+    for (uint32_t i = 0; i < testModel.main.inputIndexes.size(); i++) {
+        const auto& op = testModel.main.operands[testModel.main.inputIndexes[i]];
+        if (op.data.size() == 0) {
+            // Omitted input.
+            inputs[i] = {.hasNoValue = true};
+        } else {
+            DataLocation loc = {.poolIndex = kInputPoolIndex,
+                                .offset = static_cast<int64_t>(inputSize),
+                                .length = static_cast<int64_t>(op.data.size())};
+            inputSize += op.data.alignedSize();
+            inputs[i] = {.hasNoValue = false, .location = loc, .dimensions = {}};
+        }
+    }
+
+    // Model outputs.
+    std::vector<RequestArgument> outputs(testModel.main.outputIndexes.size());
+    size_t outputSize = 0;
+    for (uint32_t i = 0; i < testModel.main.outputIndexes.size(); i++) {
+        const auto& op = testModel.main.operands[testModel.main.outputIndexes[i]];
+
+        // In the case of zero-sized output, we should at least provide a one-byte buffer.
+        // This is because zero-sized tensors are only supported internally to the driver, or
+        // reported in output shapes. It is illegal for the client to pre-specify a zero-sized
+        // tensor as model output. Otherwise, we will have two semantic conflicts:
+        // - "Zero dimension" conflicts with "unspecified dimension".
+        // - "Omitted operand buffer" conflicts with "zero-sized operand buffer".
+        size_t bufferSize = std::max<size_t>(op.data.size(), 1);
+
+        DataLocation loc = {.poolIndex = kOutputPoolIndex,
+                            .offset = static_cast<int64_t>(outputSize),
+                            .length = static_cast<int64_t>(bufferSize)};
+        outputSize += op.data.size() == 0 ? TestBuffer::kAlignment : op.data.alignedSize();
+        outputs[i] = {.hasNoValue = false, .location = loc, .dimensions = {}};
+    }
+
+    // Allocate memory pools.
+    if (memoryType == MemoryType::ASHMEM) {
+        mInputMemory = TestAshmem::create(inputSize);
+        mOutputMemory = TestAshmem::create(outputSize);
+    } else {
+        mInputMemory = TestBlobAHWB::create(inputSize);
+        mOutputMemory = TestBlobAHWB::create(outputSize);
+    }
+    CHECK_NE(mInputMemory, nullptr);
+    CHECK_NE(mOutputMemory, nullptr);
+
+    auto copiedInputMemory = utils::clone(*mInputMemory->getAidlMemory());
+    CHECK(copiedInputMemory.has_value()) << copiedInputMemory.error().message;
+    auto copiedOutputMemory = utils::clone(*mOutputMemory->getAidlMemory());
+    CHECK(copiedOutputMemory.has_value()) << copiedOutputMemory.error().message;
+
+    std::vector<RequestMemoryPool> pools;
+    pools.push_back(RequestMemoryPool::make<RequestMemoryPool::Tag::pool>(
+            std::move(copiedInputMemory).value()));
+    pools.push_back(RequestMemoryPool::make<RequestMemoryPool::Tag::pool>(
+            std::move(copiedOutputMemory).value()));
+
+    // Copy input data to the memory pool.
+    uint8_t* inputPtr = mInputMemory->getPointer();
+    for (uint32_t i = 0; i < testModel.main.inputIndexes.size(); i++) {
+        const auto& op = testModel.main.operands[testModel.main.inputIndexes[i]];
+        if (op.data.size() > 0) {
+            const uint8_t* begin = op.data.get<uint8_t>();
+            const uint8_t* end = begin + op.data.size();
+            std::copy(begin, end, inputPtr + inputs[i].location.offset);
+        }
+    }
+
+    return {.inputs = std::move(inputs), .outputs = std::move(outputs), .pools = std::move(pools)};
+}
+
+std::vector<TestBuffer> ExecutionContext::getOutputBuffers(const Request& request) const {
+    // Copy out output results.
+    uint8_t* outputPtr = mOutputMemory->getPointer();
+    std::vector<TestBuffer> outputBuffers;
+    for (const auto& output : request.outputs) {
+        outputBuffers.emplace_back(output.location.length, outputPtr + output.location.offset);
+    }
+    return outputBuffers;
+}
+
+}  // namespace aidl::android::hardware::neuralnetworks
diff --git a/neuralnetworks/aidl/vts/functional/Utils.h b/neuralnetworks/aidl/vts/functional/Utils.h
new file mode 100644
index 0000000..266301c
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/Utils.h
@@ -0,0 +1,153 @@
+/*
+ * 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_NEURALNETWORKS_AIDL_UTILS_H
+#define ANDROID_HARDWARE_NEURALNETWORKS_AIDL_UTILS_H
+
+#include <android-base/logging.h>
+#include <android/hardware_buffer.h>
+#include <gtest/gtest.h>
+
+#include <algorithm>
+#include <iosfwd>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <aidl/android/hardware/neuralnetworks/IDevice.h>
+#include <aidl/android/hardware/neuralnetworks/Memory.h>
+#include <aidl/android/hardware/neuralnetworks/Operand.h>
+#include <aidl/android/hardware/neuralnetworks/OperandType.h>
+#include <aidl/android/hardware/neuralnetworks/Priority.h>
+#include <aidl/android/hardware/neuralnetworks/Request.h>
+
+#include <TestHarness.h>
+#include <nnapi/SharedMemory.h>
+
+namespace aidl::android::hardware::neuralnetworks {
+
+namespace nn = ::android::nn;
+
+inline constexpr Priority kDefaultPriority = Priority::MEDIUM;
+
+inline constexpr Timing kNoTiming = {.timeOnDevice = -1, .timeInDriver = -1};
+inline constexpr int64_t kNoDeadline = -1;
+inline constexpr int64_t kOmittedTimeoutDuration = -1;
+inline constexpr int64_t kNoDuration = -1;
+inline const std::vector<uint8_t> kEmptyCacheToken(IDevice::BYTE_SIZE_OF_CACHE_TOKEN);
+
+// Returns the amount of space needed to store a value of the specified type.
+//
+// Aborts if the specified type is an extension type or OEM type.
+uint32_t sizeOfData(OperandType type);
+
+// Returns the amount of space needed to store a value of the dimensions and
+// type of this operand. For a non-extension, non-OEM tensor with unspecified
+// rank or at least one unspecified dimension, returns zero.
+//
+// Aborts if the specified type is an extension type or OEM type.
+uint32_t sizeOfData(const Operand& operand);
+
+// Convenience class to manage the lifetime of memory resources.
+class TestMemoryBase {
+    DISALLOW_COPY_AND_ASSIGN(TestMemoryBase);
+
+  public:
+    TestMemoryBase() = default;
+    virtual ~TestMemoryBase() = default;
+    uint8_t* getPointer() const { return mPtr; }
+    const Memory* getAidlMemory() const { return &mAidlMemory; }
+
+  protected:
+    uint8_t* mPtr = nullptr;
+    Memory mAidlMemory;
+    bool mIsValid = false;
+};
+
+class TestAshmem : public TestMemoryBase {
+  public:
+    static std::unique_ptr<TestAshmem> create(uint32_t size);
+
+    // Prefer TestAshmem::create.
+    // The constructor calls initialize, which constructs the memory resources. This is a workaround
+    // that gtest macros cannot be used directly in a constructor.
+    TestAshmem(uint32_t size) { initialize(size); }
+
+  private:
+    void initialize(uint32_t size);
+    nn::Mapping mMappedMemory;
+};
+
+class TestBlobAHWB : public TestMemoryBase {
+  public:
+    static std::unique_ptr<TestBlobAHWB> create(uint32_t size);
+
+    // Prefer TestBlobAHWB::create.
+    // The constructor calls initialize, which constructs the memory resources. This is a
+    // workaround that gtest macros cannot be used directly in a constructor.
+    TestBlobAHWB(uint32_t size) { initialize(size); }
+    ~TestBlobAHWB();
+
+  private:
+    void initialize(uint32_t size);
+    AHardwareBuffer* mAhwb = nullptr;
+    nn::Mapping mMapping;
+};
+
+enum class MemoryType { ASHMEM, BLOB_AHWB, DEVICE };
+
+// Manages the lifetime of memory resources used in an execution.
+class ExecutionContext {
+    DISALLOW_COPY_AND_ASSIGN(ExecutionContext);
+
+  public:
+    static constexpr uint32_t kInputPoolIndex = 0;
+    static constexpr uint32_t kOutputPoolIndex = 1;
+
+    ExecutionContext() = default;
+
+    // Create HIDL Request from the TestModel struct.
+    Request createRequest(const test_helper::TestModel& testModel,
+                          MemoryType memoryType = MemoryType::ASHMEM);
+
+    // After execution, copy out output results from the output memory pool.
+    std::vector<test_helper::TestBuffer> getOutputBuffers(const Request& request) const;
+
+  private:
+    std::unique_ptr<TestMemoryBase> mInputMemory, mOutputMemory;
+};
+
+template <typename Type>
+using Named = std::pair<std::string, Type>;
+
+template <typename Type>
+const std::string& getName(const Named<Type>& namedData) {
+    return namedData.first;
+}
+
+template <typename Type>
+const Type& getData(const Named<Type>& namedData) {
+    return namedData.second;
+}
+
+std::string gtestCompliantName(std::string name);
+
+// pretty-print values for error messages
+::std::ostream& operator<<(::std::ostream& os, ErrorStatus errorStatus);
+
+}  // namespace aidl::android::hardware::neuralnetworks
+
+#endif  // ANDROID_HARDWARE_NEURALNETWORKS_AIDL_UTILS_H
diff --git a/neuralnetworks/aidl/vts/functional/ValidateModel.cpp b/neuralnetworks/aidl/vts/functional/ValidateModel.cpp
new file mode 100644
index 0000000..b84d981
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/ValidateModel.cpp
@@ -0,0 +1,1338 @@
+/*
+ * 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.
+ */
+
+#define LOG_TAG "neuralnetworks_aidl_hal_test"
+
+#include <aidl/android/hardware/common/NativeHandle.h>
+#include <android/binder_auto_utils.h>
+#include <android/binder_enums.h>
+#include <android/binder_interface_utils.h>
+#include <nnapi/TypeUtils.h>
+#include <nnapi/hal/aidl/Conversions.h>
+#include <nnapi/hal/aidl/Utils.h>
+
+#include <optional>
+#include <type_traits>
+#include <utility>
+
+#include "Callbacks.h"
+#include "GeneratedTestHarness.h"
+#include "Utils.h"
+#include "VtsHalNeuralnetworks.h"
+
+namespace aidl::android::hardware::neuralnetworks::vts::functional {
+
+using common::NativeHandle;
+using implementation::PreparedModelCallback;
+
+using PrepareModelMutation = std::function<void(Model*, ExecutionPreference*, Priority*)>;
+
+///////////////////////// UTILITY FUNCTIONS /////////////////////////
+
+static void validateGetSupportedOperations(const std::shared_ptr<IDevice>& device,
+                                           const std::string& message, const Model& model) {
+    SCOPED_TRACE(message + " [getSupportedOperations]");
+
+    std::vector<bool> supported;
+    const auto retStatus = device->getSupportedOperations(model, &supported);
+
+    ASSERT_FALSE(retStatus.isOk());
+    ASSERT_EQ(retStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
+    ASSERT_EQ(static_cast<ErrorStatus>(retStatus.getServiceSpecificError()),
+              ErrorStatus::INVALID_ARGUMENT);
+}
+
+static void validatePrepareModel(const std::shared_ptr<IDevice>& device, const std::string& message,
+                                 const Model& model, ExecutionPreference preference,
+                                 Priority priority) {
+    SCOPED_TRACE(message + " [prepareModel]");
+
+    std::shared_ptr<PreparedModelCallback> preparedModelCallback =
+            ndk::SharedRefBase::make<PreparedModelCallback>();
+    const auto prepareLaunchStatus =
+            device->prepareModel(model, preference, priority, kNoDeadline, {}, {}, kEmptyCacheToken,
+                                 preparedModelCallback);
+    ASSERT_FALSE(prepareLaunchStatus.isOk());
+    ASSERT_EQ(prepareLaunchStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
+    ASSERT_EQ(static_cast<ErrorStatus>(prepareLaunchStatus.getServiceSpecificError()),
+              ErrorStatus::INVALID_ARGUMENT);
+
+    preparedModelCallback->wait();
+    ErrorStatus prepareReturnStatus = preparedModelCallback->getStatus();
+    ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, prepareReturnStatus);
+    std::shared_ptr<IPreparedModel> preparedModel = preparedModelCallback->getPreparedModel();
+    ASSERT_EQ(nullptr, preparedModel.get());
+}
+
+static bool validExecutionPreference(ExecutionPreference preference) {
+    return preference == ExecutionPreference::LOW_POWER ||
+           preference == ExecutionPreference::FAST_SINGLE_ANSWER ||
+           preference == ExecutionPreference::SUSTAINED_SPEED;
+}
+
+static bool validExecutionPriority(Priority priority) {
+    return priority == Priority::LOW || priority == Priority::MEDIUM || priority == Priority::HIGH;
+}
+
+// Primary validation function. This function will take a valid model, apply a
+// mutation to invalidate the model, the execution preference, or the priority,
+// then pass these to supportedOperations and/or prepareModel if that method is
+// called with an invalid argument.
+static void validate(const std::shared_ptr<IDevice>& device, const std::string& message,
+                     const Model& originalModel, const PrepareModelMutation& mutate) {
+    Model model = utils::clone(originalModel).value();
+    ExecutionPreference preference = ExecutionPreference::FAST_SINGLE_ANSWER;
+    Priority priority = kDefaultPriority;
+    mutate(&model, &preference, &priority);
+
+    if (validExecutionPreference(preference) && validExecutionPriority(priority)) {
+        validateGetSupportedOperations(device, message, model);
+    }
+
+    validatePrepareModel(device, message, model, preference, priority);
+}
+
+static uint32_t addOperand(Model* model) {
+    model->main.operands.push_back({
+            .type = OperandType::INT32,
+            .dimensions = {},
+            .scale = 0.0f,
+            .zeroPoint = 0,
+            .lifetime = OperandLifeTime::SUBGRAPH_INPUT,
+            .location = {.poolIndex = 0, .offset = 0, .length = 0},
+    });
+    return model->main.operands.size() - 1;
+}
+
+static uint32_t addOperand(Model* model, OperandLifeTime lifetime) {
+    uint32_t index = addOperand(model);
+    model->main.operands[index].lifetime = lifetime;
+    return index;
+}
+
+// If we introduce a CONSTANT_COPY for an operand of size operandSize,
+// how much will this increase the size of the model?  This assumes
+// that we can (re)use all of model.operandValues for the operand
+// value.
+static size_t constantCopyExtraSize(const Model& model, size_t operandSize) {
+    const size_t operandValuesSize = model.operandValues.size();
+    return (operandValuesSize < operandSize) ? (operandSize - operandValuesSize) : 0;
+}
+
+// Highly specialized utility routine for converting an operand to
+// CONSTANT_COPY lifetime.
+//
+// Expects that:
+// - operand has a known size
+// - operand->lifetime has already been set to CONSTANT_COPY
+// - operand->location has been zeroed out
+//
+// Does the following:
+// - initializes operand->location to point to the beginning of model->operandValues
+// - resizes model->operandValues (if necessary) to be large enough for the operand
+//   value, padding it with zeroes on the end
+//
+// Potential problem:
+// By changing the operand to CONSTANT_COPY lifetime, this function is effectively initializing the
+// operand with unspecified (but deterministic) data. This means that the model may be invalidated
+// in two ways: not only is the lifetime of CONSTANT_COPY invalid, but the operand's value in the
+// graph may also be invalid (e.g., if the operand is used as an activation code and has an invalid
+// value). For now, this should be fine because it just means we're not testing what we think we're
+// testing in certain cases; but we can handwave this and assume we're probabilistically likely to
+// exercise the validation code over the span of the entire test set and operand space.
+//
+// Aborts if the specified operand type is an extension type or OEM type.
+static void becomeConstantCopy(Model* model, Operand* operand) {
+    // sizeOfData will abort if the specified type is an extension type or OEM type.
+    const size_t sizeOfOperand = sizeOfData(*operand);
+    EXPECT_NE(sizeOfOperand, size_t(0));
+    operand->location.poolIndex = 0;
+    operand->location.offset = 0;
+    operand->location.length = sizeOfOperand;
+    if (model->operandValues.size() < sizeOfOperand) {
+        model->operandValues.resize(sizeOfOperand);
+    }
+}
+
+// The sizeForBinder() functions estimate the size of the
+// representation of a value when sent to binder.  It's probably a bit
+// of an under-estimate, because we don't know the size of the
+// metadata in the binder format (e.g., representation of the size of
+// a vector); but at least it adds up "big" things like vector
+// contents.  However, it doesn't treat inter-field or end-of-struct
+// padding in a methodical way -- there's no attempt to be consistent
+// in whether or not padding in the native (C++) representation
+// contributes to the estimated size for the binder representation;
+// and there's no attempt to understand what padding (if any) is
+// needed in the binder representation.
+//
+// This assumes that non-metadata uses a fixed length encoding (e.g.,
+// a uint32_t is always encoded in sizeof(uint32_t) bytes, rather than
+// using an encoding whose length is related to the magnitude of the
+// encoded value).
+
+template <typename Type>
+static size_t sizeForBinder(const Type& val) {
+    static_assert(std::is_trivially_copyable_v<std::remove_reference_t<Type>>,
+                  "expected a trivially copyable type");
+    return sizeof(val);
+}
+
+template <typename Type>
+static size_t sizeForBinder(const std::vector<Type>& vec) {
+    return std::accumulate(vec.begin(), vec.end(), 0,
+                           [](size_t acc, const Type& x) { return acc + sizeForBinder(x); });
+}
+
+template <>
+size_t sizeForBinder(const SymmPerChannelQuantParams& symmPerChannelQuantParams) {
+    size_t size = 0;
+
+    size += sizeForBinder(symmPerChannelQuantParams.scales);
+    size += sizeForBinder(symmPerChannelQuantParams.channelDim);
+
+    return size;
+}
+
+template <>
+size_t sizeForBinder(const std::optional<OperandExtraParams>& optionalExtraParams) {
+    if (!optionalExtraParams.has_value()) {
+        return 0;
+    }
+    const auto& extraParams = optionalExtraParams.value();
+    using Tag = OperandExtraParams::Tag;
+    switch (extraParams.getTag()) {
+        case Tag::channelQuant:
+            return sizeForBinder(extraParams.get<Tag::channelQuant>());
+        case Tag::extension:
+            return sizeForBinder(extraParams.get<Tag::extension>());
+    }
+    LOG(FATAL) << "Unrecognized extraParams tag: " << static_cast<int>(extraParams.getTag());
+    return 0;
+}
+
+template <>
+size_t sizeForBinder(const Operand& operand) {
+    size_t size = 0;
+
+    size += sizeForBinder(operand.type);
+    size += sizeForBinder(operand.dimensions);
+    size += sizeForBinder(operand.scale);
+    size += sizeForBinder(operand.zeroPoint);
+    size += sizeForBinder(operand.lifetime);
+    size += sizeForBinder(operand.location);
+    size += sizeForBinder(operand.extraParams);
+
+    return size;
+}
+
+template <>
+size_t sizeForBinder(const Operation& operation) {
+    size_t size = 0;
+
+    size += sizeForBinder(operation.type);
+    size += sizeForBinder(operation.inputs);
+    size += sizeForBinder(operation.outputs);
+
+    return size;
+}
+
+template <>
+size_t sizeForBinder(const std::string& name) {
+    return name.size();
+}
+
+template <>
+size_t sizeForBinder(const Memory& memory) {
+    // This is just a guess.
+
+    size_t size = 0;
+    const NativeHandle& handle = memory.handle;
+    size += sizeof(decltype(handle.fds)::value_type) * handle.fds.size();
+    size += sizeof(decltype(handle.ints)::value_type) * handle.ints.size();
+    size += sizeForBinder(memory.name);
+    size += sizeof(memory);
+
+    return size;
+}
+
+template <>
+size_t sizeForBinder(const Subgraph& subgraph) {
+    size_t size = 0;
+
+    size += sizeForBinder(subgraph.operands);
+    size += sizeForBinder(subgraph.operations);
+    size += sizeForBinder(subgraph.inputIndexes);
+    size += sizeForBinder(subgraph.outputIndexes);
+
+    return size;
+}
+
+template <>
+size_t sizeForBinder(const ExtensionNameAndPrefix& extensionNameToPrefix) {
+    size_t size = 0;
+
+    size += sizeForBinder(extensionNameToPrefix.name);
+    size += sizeForBinder(extensionNameToPrefix.prefix);
+
+    return size;
+}
+
+template <>
+size_t sizeForBinder(const Model& model) {
+    size_t size = 0;
+
+    size += sizeForBinder(model.main);
+    size += sizeForBinder(model.referenced);
+    size += sizeForBinder(model.operandValues);
+    size += sizeForBinder(model.pools);
+    size += sizeForBinder(model.relaxComputationFloat32toFloat16);
+    size += sizeForBinder(model.extensionNameToPrefix);
+
+    return size;
+}
+
+// https://developer.android.com/reference/android/os/TransactionTooLargeException.html
+//
+//     "The Binder transaction buffer has a limited fixed size,
+//     currently 1Mb, which is shared by all transactions in progress
+//     for the process."
+//
+// Will our representation fit under this limit?  There are two complications:
+// - Our representation size is just approximate (see sizeForBinder()).
+// - This object may not be the only occupant of the Binder transaction buffer.
+// So we'll be very conservative: We want the representation size to be no
+// larger than half the transaction buffer size.
+//
+// If our representation grows large enough that it still fits within
+// the transaction buffer but combined with other transactions may
+// exceed the buffer size, then we may see intermittent HAL transport
+// errors.
+static bool exceedsBinderSizeLimit(size_t representationSize) {
+    // Instead of using this fixed buffer size, we might instead be able to use
+    // ProcessState::self()->getMmapSize(). However, this has a potential
+    // problem: The binder/mmap size of the current process does not necessarily
+    // indicate the binder/mmap size of the service (i.e., the other process).
+    // The only way it would be a good indication is if both the current process
+    // and the service use the default size.
+    static const size_t kHalfBufferSize = 1024 * 1024 / 2;
+
+    return representationSize > kHalfBufferSize;
+}
+
+///////////////////////// VALIDATE EXECUTION ORDER ////////////////////////////
+
+static void mutateExecutionOrderTest(const std::shared_ptr<IDevice>& device, const Model& model,
+                                     const std::vector<uint32_t>& numberOfConsumers) {
+    for (size_t operation = 0; operation < model.main.operations.size(); ++operation) {
+        const Operation& operationObj = model.main.operations[operation];
+        for (uint32_t input : operationObj.inputs) {
+            if (model.main.operands[input].lifetime == OperandLifeTime::TEMPORARY_VARIABLE ||
+                model.main.operands[input].lifetime == OperandLifeTime::SUBGRAPH_OUTPUT) {
+                // This operation reads an operand written by some
+                // other operation.  Move this operation to the
+                // beginning of the sequence, ensuring that it reads
+                // the operand before that operand is written, thereby
+                // violating execution order rules.
+                const std::string message = "mutateExecutionOrderTest: operation " +
+                                            std::to_string(operation) + " is a reader";
+                validate(device, message, model,
+                         [operation](Model* model, ExecutionPreference*, Priority*) {
+                             auto& operations = model->main.operations;
+                             std::rotate(operations.begin(), operations.begin() + operation,
+                                         operations.begin() + operation + 1);
+                         });
+                break;  // only need to do this once per operation
+            }
+        }
+        for (uint32_t output : operationObj.outputs) {
+            if (numberOfConsumers[output] > 0) {
+                // This operation writes an operand read by some other
+                // operation.  Move this operation to the end of the
+                // sequence, ensuring that it writes the operand after
+                // that operand is read, thereby violating execution
+                // order rules.
+                const std::string message = "mutateExecutionOrderTest: operation " +
+                                            std::to_string(operation) + " is a writer";
+                validate(device, message, model,
+                         [operation](Model* model, ExecutionPreference*, Priority*) {
+                             auto& operations = model->main.operations;
+                             std::rotate(operations.begin() + operation,
+                                         operations.begin() + operation + 1, operations.end());
+                         });
+                break;  // only need to do this once per operation
+            }
+        }
+    }
+}
+
+///////////////////////// VALIDATE MODEL OPERAND TYPE /////////////////////////
+
+static const int32_t invalidOperandTypes[] = {
+        -1,
+        static_cast<int32_t>(*(ndk::enum_range<OperandType>().end() - 1)) + 1,
+};
+
+static void mutateOperandTypeTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    for (size_t operand = 0; operand < model.main.operands.size(); ++operand) {
+        for (int32_t invalidOperandType : invalidOperandTypes) {
+            const std::string message = "mutateOperandTypeTest: operand " +
+                                        std::to_string(operand) + " set to value " +
+                                        std::to_string(invalidOperandType);
+            validate(device, message, model,
+                     [operand, invalidOperandType](Model* model, ExecutionPreference*, Priority*) {
+                         model->main.operands[operand].type =
+                                 static_cast<OperandType>(invalidOperandType);
+                     });
+        }
+    }
+}
+
+///////////////////////// VALIDATE OPERAND RANK /////////////////////////
+
+static uint32_t getInvalidRank(OperandType type) {
+    switch (type) {
+        case OperandType::FLOAT16:
+        case OperandType::FLOAT32:
+        case OperandType::INT32:
+        case OperandType::UINT32:
+        case OperandType::BOOL:
+            return 1;
+        case OperandType::TENSOR_BOOL8:
+        case OperandType::TENSOR_FLOAT16:
+        case OperandType::TENSOR_FLOAT32:
+        case OperandType::TENSOR_INT32:
+        case OperandType::TENSOR_QUANT8_ASYMM:
+        case OperandType::TENSOR_QUANT8_SYMM:
+        case OperandType::TENSOR_QUANT16_ASYMM:
+        case OperandType::TENSOR_QUANT16_SYMM:
+        case OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL:
+            return 0;
+        default:
+            return 0;
+    }
+}
+
+static void mutateOperandRankTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    for (size_t operand = 0; operand < model.main.operands.size(); ++operand) {
+        const uint32_t invalidRank = getInvalidRank(model.main.operands[operand].type);
+        if (invalidRank == 0) {
+            continue;
+        }
+        const std::string message = "mutateOperandRankTest: operand " + std::to_string(operand) +
+                                    " has rank of " + std::to_string(invalidRank);
+        validate(device, message, model,
+                 [operand, invalidRank](Model* model, ExecutionPreference*, Priority*) {
+                     model->main.operands[operand].dimensions =
+                             std::vector<int32_t>(invalidRank, 0);
+                 });
+    }
+}
+
+///////////////////////// VALIDATE OPERAND SCALE /////////////////////////
+
+static float getInvalidScale(OperandType type) {
+    switch (type) {
+        case OperandType::FLOAT16:
+        case OperandType::FLOAT32:
+        case OperandType::INT32:
+        case OperandType::UINT32:
+        case OperandType::BOOL:
+        case OperandType::TENSOR_BOOL8:
+        case OperandType::TENSOR_FLOAT16:
+        case OperandType::TENSOR_FLOAT32:
+        case OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL:
+        case OperandType::SUBGRAPH:
+            return 1.0f;
+        case OperandType::TENSOR_INT32:
+            return -1.0f;
+        case OperandType::TENSOR_QUANT8_SYMM:
+        case OperandType::TENSOR_QUANT8_ASYMM:
+        case OperandType::TENSOR_QUANT16_ASYMM:
+        case OperandType::TENSOR_QUANT16_SYMM:
+            return 0.0f;
+        default:
+            return 0.0f;
+    }
+}
+
+static void mutateOperandScaleTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    for (size_t operand = 0; operand < model.main.operands.size(); ++operand) {
+        const float invalidScale = getInvalidScale(model.main.operands[operand].type);
+        const std::string message = "mutateOperandScaleTest: operand " + std::to_string(operand) +
+                                    " has scale of " + std::to_string(invalidScale);
+        validate(device, message, model,
+                 [operand, invalidScale](Model* model, ExecutionPreference*, Priority*) {
+                     model->main.operands[operand].scale = invalidScale;
+                 });
+    }
+}
+
+///////////////////////// VALIDATE OPERAND ZERO POINT /////////////////////////
+
+static std::vector<int32_t> getInvalidZeroPoints(OperandType type) {
+    switch (type) {
+        case OperandType::FLOAT16:
+        case OperandType::FLOAT32:
+        case OperandType::INT32:
+        case OperandType::UINT32:
+        case OperandType::BOOL:
+        case OperandType::TENSOR_BOOL8:
+        case OperandType::TENSOR_FLOAT16:
+        case OperandType::TENSOR_FLOAT32:
+        case OperandType::TENSOR_INT32:
+        case OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL:
+        case OperandType::SUBGRAPH:
+            return {1};
+        case OperandType::TENSOR_QUANT8_ASYMM:
+            return {-1, 256};
+        case OperandType::TENSOR_QUANT8_SYMM:
+            return {-129, -1, 1, 128};
+        case OperandType::TENSOR_QUANT16_ASYMM:
+            return {-1, 65536};
+        case OperandType::TENSOR_QUANT16_SYMM:
+            return {-32769, -1, 1, 32768};
+        default:
+            return {};
+    }
+}
+
+static void mutateOperandZeroPointTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    for (size_t operand = 0; operand < model.main.operands.size(); ++operand) {
+        const std::vector<int32_t> invalidZeroPoints =
+                getInvalidZeroPoints(model.main.operands[operand].type);
+        for (int32_t invalidZeroPoint : invalidZeroPoints) {
+            const std::string message = "mutateOperandZeroPointTest: operand " +
+                                        std::to_string(operand) + " has zero point of " +
+                                        std::to_string(invalidZeroPoint);
+            validate(device, message, model,
+                     [operand, invalidZeroPoint](Model* model, ExecutionPreference*, Priority*) {
+                         model->main.operands[operand].zeroPoint = invalidZeroPoint;
+                     });
+        }
+    }
+}
+
+///////////////////////// VALIDATE OPERAND LIFETIME /////////////////////////////////////////////
+
+static std::vector<OperandLifeTime> getInvalidLifeTimes(const Model& model, size_t modelSize,
+                                                        const Operand& operand) {
+    // TODO: Support OperandLifeTime::CONSTANT_REFERENCE as an invalid lifetime
+    // TODO: Support OperandLifeTime::NO_VALUE as an invalid lifetime
+
+    // Ways to get an invalid lifetime:
+    // - change whether a lifetime means an operand should have a writer
+    std::vector<OperandLifeTime> ret;
+    switch (operand.lifetime) {
+        case OperandLifeTime::SUBGRAPH_OUTPUT:
+        case OperandLifeTime::TEMPORARY_VARIABLE:
+            ret = {
+                    OperandLifeTime::SUBGRAPH_INPUT,
+                    OperandLifeTime::CONSTANT_COPY,
+            };
+            break;
+        case OperandLifeTime::CONSTANT_COPY:
+        case OperandLifeTime::CONSTANT_POOL:
+        case OperandLifeTime::SUBGRAPH_INPUT:
+            ret = {
+                    OperandLifeTime::TEMPORARY_VARIABLE,
+                    OperandLifeTime::SUBGRAPH_OUTPUT,
+            };
+            break;
+        case OperandLifeTime::NO_VALUE:
+            // Not enough information to know whether
+            // TEMPORARY_VARIABLE or CONSTANT_COPY would be invalid --
+            // is this operand written (then CONSTANT_COPY would be
+            // invalid) or not (then TEMPORARY_VARIABLE would be
+            // invalid)?
+            break;
+        case OperandLifeTime::SUBGRAPH:
+            break;
+        default:
+            ADD_FAILURE();
+            break;
+    }
+
+    const size_t operandSize = sizeOfData(operand);  // will be zero if shape is unknown
+    if (!operandSize ||
+        exceedsBinderSizeLimit(modelSize + constantCopyExtraSize(model, operandSize))) {
+        // Unknown size or too-large size
+        ret.erase(std::remove(ret.begin(), ret.end(), OperandLifeTime::CONSTANT_COPY), ret.end());
+    }
+
+    return ret;
+}
+
+static void mutateOperandLifeTimeTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    const size_t modelSize = sizeForBinder(model);
+    for (size_t operand = 0; operand < model.main.operands.size(); ++operand) {
+        const std::vector<OperandLifeTime> invalidLifeTimes =
+                getInvalidLifeTimes(model, modelSize, model.main.operands[operand]);
+        for (OperandLifeTime invalidLifeTime : invalidLifeTimes) {
+            const std::string message = "mutateOperandLifetimeTest: operand " +
+                                        std::to_string(operand) + " has lifetime " +
+                                        toString(invalidLifeTime) + " instead of lifetime " +
+                                        toString(model.main.operands[operand].lifetime);
+            validate(device, message, model,
+                     [operand, invalidLifeTime](Model* model, ExecutionPreference*, Priority*) {
+                         static const DataLocation kZeroDataLocation = {};
+                         Operand& operandObj = model->main.operands[operand];
+                         switch (operandObj.lifetime) {
+                             case OperandLifeTime::SUBGRAPH_INPUT: {
+                                 auto& inputs = model->main.inputIndexes;
+                                 inputs.erase(std::remove(inputs.begin(), inputs.end(), operand),
+                                              inputs.end());
+                                 break;
+                             }
+                             case OperandLifeTime::SUBGRAPH_OUTPUT: {
+                                 auto& outputs = model->main.outputIndexes;
+                                 outputs.erase(std::remove(outputs.begin(), outputs.end(), operand),
+                                               outputs.end());
+                                 break;
+                             }
+                             default:
+                                 break;
+                         }
+                         operandObj.lifetime = invalidLifeTime;
+                         operandObj.location = kZeroDataLocation;
+                         switch (invalidLifeTime) {
+                             case OperandLifeTime::CONSTANT_COPY: {
+                                 becomeConstantCopy(model, &operandObj);
+                                 break;
+                             }
+                             case OperandLifeTime::SUBGRAPH_INPUT:
+                                 model->main.inputIndexes.push_back(operand);
+                                 break;
+                             case OperandLifeTime::SUBGRAPH_OUTPUT:
+                                 model->main.outputIndexes.push_back(operand);
+                                 break;
+                             default:
+                                 break;
+                         }
+                     });
+        }
+    }
+}
+
+///////////////////////// VALIDATE OPERAND INPUT-or-OUTPUT //////////////////////////////////////
+
+static std::optional<OperandLifeTime> getInputOutputLifeTime(const Model& model, size_t modelSize,
+                                                             const Operand& operand) {
+    // Ways to get an invalid lifetime (with respect to model inputIndexes and outputIndexes):
+    // - change whether a lifetime means an operand is a model input, a model output, or neither
+    // - preserve whether or not a lifetime means an operand should have a writer
+    switch (operand.lifetime) {
+        case OperandLifeTime::CONSTANT_COPY:
+        case OperandLifeTime::CONSTANT_POOL:
+            return OperandLifeTime::SUBGRAPH_INPUT;
+        case OperandLifeTime::SUBGRAPH_INPUT: {
+            const size_t operandSize = sizeOfData(operand);  // will be zero if shape is unknown
+            if (!operandSize ||
+                exceedsBinderSizeLimit(modelSize + constantCopyExtraSize(model, operandSize))) {
+                // Unknown size or too-large size
+                break;
+            }
+            return OperandLifeTime::CONSTANT_COPY;
+        }
+        case OperandLifeTime::SUBGRAPH_OUTPUT:
+            return OperandLifeTime::TEMPORARY_VARIABLE;
+        case OperandLifeTime::TEMPORARY_VARIABLE:
+            return OperandLifeTime::SUBGRAPH_OUTPUT;
+        case OperandLifeTime::NO_VALUE:
+            // Not enough information to know whether
+            // TEMPORARY_VARIABLE or CONSTANT_COPY would be an
+            // appropriate choice -- is this operand written (then
+            // TEMPORARY_VARIABLE would be appropriate) or not (then
+            // CONSTANT_COPY would be appropriate)?
+            break;
+        case OperandLifeTime::SUBGRAPH:
+            break;
+        default:
+            ADD_FAILURE();
+            break;
+    }
+
+    return std::nullopt;
+}
+
+static void mutateOperandInputOutputTest(const std::shared_ptr<IDevice>& device,
+                                         const Model& model) {
+    const size_t modelSize = sizeForBinder(model);
+    for (size_t operand = 0; operand < model.main.operands.size(); ++operand) {
+        const std::optional<OperandLifeTime> changedLifeTime =
+                getInputOutputLifeTime(model, modelSize, model.main.operands[operand]);
+        if (changedLifeTime) {
+            const std::string message = "mutateOperandInputOutputTest: operand " +
+                                        std::to_string(operand) + " has lifetime " +
+                                        toString(*changedLifeTime) + " instead of lifetime " +
+                                        toString(model.main.operands[operand].lifetime);
+            validate(device, message, model,
+                     [operand, changedLifeTime](Model* model, ExecutionPreference*, Priority*) {
+                         static const DataLocation kZeroDataLocation = {};
+                         Operand& operandObj = model->main.operands[operand];
+                         operandObj.lifetime = *changedLifeTime;
+                         operandObj.location = kZeroDataLocation;
+                         if (*changedLifeTime == OperandLifeTime::CONSTANT_COPY) {
+                             becomeConstantCopy(model, &operandObj);
+                         }
+                     });
+        }
+    }
+}
+
+///////////////////////// VALIDATE OPERAND NUMBER OF WRITERS ////////////////////////////////////
+
+static void mutateOperandAddWriterTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    for (size_t operation = 0; operation < model.main.operations.size(); ++operation) {
+        for (size_t badOutputNum = 0;
+             badOutputNum < model.main.operations[operation].outputs.size(); ++badOutputNum) {
+            const uint32_t outputOperandIndex =
+                    model.main.operations[operation].outputs[badOutputNum];
+            const std::string message = "mutateOperandAddWriterTest: operation " +
+                                        std::to_string(operation) + " writes to " +
+                                        std::to_string(outputOperandIndex);
+            // We'll insert a copy of the operation, all of whose
+            // OTHER output operands are newly-created -- i.e.,
+            // there'll only be a duplicate write of ONE of that
+            // operation's output operands.
+            validate(device, message, model,
+                     [operation, badOutputNum](Model* model, ExecutionPreference*, Priority*) {
+                         Operation newOperation = model->main.operations[operation];
+                         for (size_t outputNum = 0; outputNum < newOperation.outputs.size();
+                              ++outputNum) {
+                             if (outputNum == badOutputNum) continue;
+
+                             Operand operandValue =
+                                     model->main.operands[newOperation.outputs[outputNum]];
+                             if (operandValue.lifetime == OperandLifeTime::SUBGRAPH_OUTPUT) {
+                                 operandValue.lifetime = OperandLifeTime::TEMPORARY_VARIABLE;
+                             } else {
+                                 ASSERT_EQ(operandValue.lifetime,
+                                           OperandLifeTime::TEMPORARY_VARIABLE);
+                             }
+                             newOperation.outputs[outputNum] = model->main.operands.size();
+                             model->main.operands.push_back(operandValue);
+                         }
+                         // Where do we insert the extra writer (a new
+                         // operation)?  It has to be later than all the
+                         // writers of its inputs.  The easiest thing to do
+                         // is to insert it at the end of the operation
+                         // sequence.
+                         model->main.operations.push_back(newOperation);
+                     });
+        }
+    }
+}
+
+///////////////////////// VALIDATE EXTRA ??? /////////////////////////
+
+// TODO: Operand::location
+
+///////////////////////// VALIDATE OPERATION OPERAND TYPE /////////////////////////
+
+static void mutateOperand(Operand* operand, OperandType type) {
+    Operand newOperand = *operand;
+    newOperand.type = type;
+    switch (type) {
+        case OperandType::FLOAT16:
+        case OperandType::FLOAT32:
+        case OperandType::INT32:
+        case OperandType::UINT32:
+        case OperandType::BOOL:
+            newOperand.dimensions = {};
+            newOperand.scale = 0.0f;
+            newOperand.zeroPoint = 0;
+            break;
+        case OperandType::TENSOR_BOOL8:
+        case OperandType::TENSOR_FLOAT16:
+        case OperandType::TENSOR_FLOAT32:
+            newOperand.dimensions = operand->dimensions.size() > 0 ? operand->dimensions
+                                                                   : std::vector<int32_t>({1});
+            newOperand.scale = 0.0f;
+            newOperand.zeroPoint = 0;
+            break;
+        case OperandType::TENSOR_INT32:
+            newOperand.dimensions = operand->dimensions.size() > 0 ? operand->dimensions
+                                                                   : std::vector<int32_t>({1});
+            newOperand.zeroPoint = 0;
+            break;
+        case OperandType::TENSOR_QUANT8_ASYMM:
+        case OperandType::TENSOR_QUANT8_SYMM:
+        case OperandType::TENSOR_QUANT16_ASYMM:
+        case OperandType::TENSOR_QUANT16_SYMM:
+            newOperand.dimensions = operand->dimensions.size() > 0 ? operand->dimensions
+                                                                   : std::vector<int32_t>({1});
+            newOperand.scale = operand->scale != 0.0f ? operand->scale : 1.0f;
+            break;
+        case OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL: {
+            newOperand.dimensions = operand->dimensions.size() > 0 ? operand->dimensions
+                                                                   : std::vector<int32_t>({1});
+            newOperand.scale = 0.0f;
+            newOperand.zeroPoint = 0;
+
+            SymmPerChannelQuantParams channelQuant;
+            channelQuant.channelDim = 0;
+            channelQuant.scales = std::vector<float>(
+                    operand->dimensions.size() > 0 ? static_cast<size_t>(operand->dimensions[0])
+                                                   : 0);
+            for (size_t i = 0; i < channelQuant.scales.size(); ++i) {
+                channelQuant.scales[i] = 1.0f;
+            }
+            newOperand.extraParams->set<OperandExtraParams::Tag::channelQuant>(
+                    std::move(channelQuant));
+        } break;
+        default:
+            break;
+    }
+    *operand = newOperand;
+}
+
+static bool mutateOperationOperandTypeSkip(size_t operand, OperandType type, const Model& model) {
+    if (type == model.main.operands[operand].type) {
+        return true;
+    }
+    for (const Operation& operation : model.main.operations) {
+        // Skip mutateOperationOperandTypeTest for the following operations.
+        // - LSH_PROJECTION's second argument is allowed to have any type.
+        // - ARGMIN and ARGMAX's first argument can be any of
+        // TENSOR_(FLOAT16|FLOAT32|INT32|QUANT8_ASYMM).
+        // - CAST's argument can be any of TENSOR_(FLOAT16|FLOAT32|INT32|QUANT8_ASYMM).
+        // - RANDOM_MULTINOMIAL's argument can be either TENSOR_FLOAT16 or TENSOR_FLOAT32.
+        // - DEQUANTIZE input can be any of
+        // TENSOR_(QUANT8_ASYMM|QUANT8_ASYMM_SIGNED|QUANT8_SYMM|QUANT8_SYMM_PER_CHANNEL),
+        // output can be of either TENSOR_FLOAT16 or TENSOR_FLOAT32.
+        // - QUANTIZE input can be either TENSOR_FLOAT16 or TENSOR_FLOAT32
+        // - CONV_2D filter type (arg 1) can be QUANT8_ASYMM or QUANT8_SYMM_PER_CHANNEL
+        // - DEPTHWISE_CONV_2D filter type (arg 1) can be QUANT8_ASYMM or QUANT8_SYMM_PER_CHANNEL
+        // - GROUPED_CONV_2D filter type (arg 1) can be QUANT8_ASYMM or QUANT8_SYMM_PER_CHANNEL
+        // - TRANSPOSE_CONV_2D filter type (arg 1) can be QUANT8_ASYMM or QUANT8_SYMM_PER_CHANNEL
+        // - AXIS_ALIGNED_BBOX_TRANSFORM bounding boxes (arg 1) can be of
+        //     TENSOR_QUANT8_ASYMM or TENSOR_QUANT8_ASYMM_SIGNED.
+        // - RANK's input can have any TENSOR_* type.
+        switch (operation.type) {
+            case OperationType::LSH_PROJECTION: {
+                if (operand == operation.inputs[1]) {
+                    return true;
+                }
+            } break;
+            case OperationType::CAST:
+            case OperationType::ARGMAX:
+            case OperationType::ARGMIN: {
+                if (type == OperandType::TENSOR_FLOAT16 || type == OperandType::TENSOR_FLOAT32 ||
+                    type == OperandType::TENSOR_INT32 || type == OperandType::TENSOR_QUANT8_ASYMM ||
+                    type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
+                    return true;
+                }
+            } break;
+            case OperationType::QUANTIZE: {
+                if (operand == operation.inputs[0] &&
+                    (type == OperandType::TENSOR_FLOAT16 || type == OperandType::TENSOR_FLOAT32)) {
+                    return true;
+                }
+                if (operand == operation.outputs[0] &&
+                    (type == OperandType::TENSOR_QUANT8_ASYMM ||
+                     type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED)) {
+                    return true;
+                }
+            } break;
+            case OperationType::RANDOM_MULTINOMIAL: {
+                if (operand == operation.inputs[0] &&
+                    (type == OperandType::TENSOR_FLOAT16 || type == OperandType::TENSOR_FLOAT32)) {
+                    return true;
+                }
+            } break;
+            case OperationType::DEQUANTIZE: {
+                if (operand == operation.inputs[0] &&
+                    (type == OperandType::TENSOR_QUANT8_ASYMM ||
+                     type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED ||
+                     type == OperandType::TENSOR_QUANT8_SYMM ||
+                     type == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL)) {
+                    return true;
+                }
+                if (operand == operation.outputs[0] &&
+                    (type == OperandType::TENSOR_FLOAT16 || type == OperandType::TENSOR_FLOAT32)) {
+                    return true;
+                }
+            } break;
+            case OperationType::TRANSPOSE_CONV_2D:
+            case OperationType::GROUPED_CONV_2D:
+            case OperationType::DEPTHWISE_CONV_2D:
+            case OperationType::CONV_2D: {
+                if (operand == operation.inputs[1] &&
+                    (type == OperandType::TENSOR_QUANT8_ASYMM ||
+                     type == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL)) {
+                    return true;
+                }
+            } break;
+            case OperationType::AXIS_ALIGNED_BBOX_TRANSFORM: {
+                if (operand == operation.inputs[1] &&
+                    (type == OperandType::TENSOR_QUANT8_ASYMM ||
+                     type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED)) {
+                    return true;
+                }
+            } break;
+            case OperationType::RANK: {
+                if (operand == operation.inputs[0] &&
+                    (type == OperandType::TENSOR_FLOAT16 || type == OperandType::TENSOR_FLOAT32 ||
+                     type == OperandType::TENSOR_INT32 ||
+                     type == OperandType::TENSOR_QUANT8_ASYMM ||
+                     type == OperandType::TENSOR_QUANT16_SYMM ||
+                     type == OperandType::TENSOR_BOOL8 ||
+                     type == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL ||
+                     type == OperandType::TENSOR_QUANT16_ASYMM ||
+                     type == OperandType::TENSOR_QUANT8_SYMM ||
+                     type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED)) {
+                    return true;
+                }
+            } break;
+            default:
+                break;
+        }
+    }
+    return false;
+}
+
+static void mutateOperationOperandTypeTest(const std::shared_ptr<IDevice>& device,
+                                           const Model& model) {
+    for (size_t operand = 0; operand < model.main.operands.size(); ++operand) {
+        for (OperandType invalidOperandType : ndk::enum_range<OperandType>()) {
+            if (mutateOperationOperandTypeSkip(operand, invalidOperandType, model)) {
+                continue;
+            }
+            const std::string message = "mutateOperationOperandTypeTest: operand " +
+                                        std::to_string(operand) + " set to type " +
+                                        toString(invalidOperandType);
+            validate(device, message, model,
+                     [operand, invalidOperandType](Model* model, ExecutionPreference*, Priority*) {
+                         mutateOperand(&model->main.operands[operand], invalidOperandType);
+                     });
+        }
+    }
+}
+
+///////////////////////// VALIDATE MODEL OPERATION TYPE /////////////////////////
+
+static const int32_t invalidOperationTypes[] = {
+        -1,
+        static_cast<int32_t>(*(ndk::enum_range<OperationType>().end() - 1)) + 1,
+};
+
+static void mutateOperationTypeTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    for (size_t operation = 0; operation < model.main.operations.size(); ++operation) {
+        for (int32_t invalidOperationType : invalidOperationTypes) {
+            const std::string message = "mutateOperationTypeTest: operation " +
+                                        std::to_string(operation) + " set to value " +
+                                        std::to_string(invalidOperationType);
+            validate(device, message, model,
+                     [operation, invalidOperationType](Model* model, ExecutionPreference*,
+                                                       Priority*) {
+                         model->main.operations[operation].type =
+                                 static_cast<OperationType>(invalidOperationType);
+                     });
+        }
+    }
+}
+
+///////////////////////// VALIDATE MODEL OPERATION INPUT OPERAND INDEX /////////////////////////
+
+static void mutateOperationInputOperandIndexTest(const std::shared_ptr<IDevice>& device,
+                                                 const Model& model) {
+    for (size_t operation = 0; operation < model.main.operations.size(); ++operation) {
+        const uint32_t invalidOperand = model.main.operands.size();
+        for (size_t input = 0; input < model.main.operations[operation].inputs.size(); ++input) {
+            const std::string message = "mutateOperationInputOperandIndexTest: operation " +
+                                        std::to_string(operation) + " input " +
+                                        std::to_string(input);
+            validate(device, message, model,
+                     [operation, input, invalidOperand](Model* model, ExecutionPreference*,
+                                                        Priority*) {
+                         model->main.operations[operation].inputs[input] = invalidOperand;
+                     });
+        }
+    }
+}
+
+///////////////////////// VALIDATE MODEL OPERATION OUTPUT OPERAND INDEX /////////////////////////
+
+static void mutateOperationOutputOperandIndexTest(const std::shared_ptr<IDevice>& device,
+                                                  const Model& model) {
+    for (size_t operation = 0; operation < model.main.operations.size(); ++operation) {
+        const uint32_t invalidOperand = model.main.operands.size();
+        for (size_t output = 0; output < model.main.operations[operation].outputs.size();
+             ++output) {
+            const std::string message = "mutateOperationOutputOperandIndexTest: operation " +
+                                        std::to_string(operation) + " output " +
+                                        std::to_string(output);
+            validate(device, message, model,
+                     [operation, output, invalidOperand](Model* model, ExecutionPreference*,
+                                                         Priority*) {
+                         model->main.operations[operation].outputs[output] = invalidOperand;
+                     });
+        }
+    }
+}
+
+///////////////////////// VALIDATE MODEL OPERANDS WRITTEN ///////////////////////////////////////
+
+static void mutateOperationRemoveWriteTest(const std::shared_ptr<IDevice>& device,
+                                           const Model& model,
+                                           const std::vector<uint32_t>& numberOfConsumers) {
+    for (size_t operation = 0; operation < model.main.operations.size(); ++operation) {
+        for (size_t outputNum = 0; outputNum < model.main.operations[operation].outputs.size();
+             ++outputNum) {
+            const uint32_t outputOperandIndex = model.main.operations[operation].outputs[outputNum];
+            if (numberOfConsumers[outputOperandIndex] > 0) {
+                const std::string message = "mutateOperationRemoveWriteTest: operation " +
+                                            std::to_string(operation) + " writes to " +
+                                            std::to_string(outputOperandIndex);
+                validate(device, message, model,
+                         [operation, outputNum](Model* model, ExecutionPreference*, Priority*) {
+                             int32_t& outputOperandIndex =
+                                     model->main.operations[operation].outputs[outputNum];
+                             Operand operandValue = model->main.operands[outputOperandIndex];
+                             if (operandValue.lifetime == OperandLifeTime::SUBGRAPH_OUTPUT) {
+                                 operandValue.lifetime = OperandLifeTime::TEMPORARY_VARIABLE;
+                             } else {
+                                 ASSERT_EQ(operandValue.lifetime,
+                                           OperandLifeTime::TEMPORARY_VARIABLE);
+                             }
+                             outputOperandIndex = model->main.operands.size();
+                             model->main.operands.push_back(operandValue);
+                         });
+            }
+        }
+    }
+}
+
+///////////////////////// REMOVE OPERAND FROM EVERYTHING /////////////////////////
+
+static void removeValueAndDecrementGreaterValues(std::vector<int32_t>* vec, uint32_t value) {
+    if (vec) {
+        // remove elements matching "value"
+        vec->erase(std::remove(vec->begin(), vec->end(), value), vec->end());
+
+        // decrement elements exceeding "value"
+        std::transform(vec->begin(), vec->end(), vec->begin(),
+                       [value](uint32_t v) { return v > value ? v-- : v; });
+    }
+}
+
+static void removeOperand(Model* model, uint32_t index) {
+    model->main.operands.erase(model->main.operands.begin() + index);
+    for (Operation& operation : model->main.operations) {
+        removeValueAndDecrementGreaterValues(&operation.inputs, index);
+        removeValueAndDecrementGreaterValues(&operation.outputs, index);
+    }
+    removeValueAndDecrementGreaterValues(&model->main.inputIndexes, index);
+    removeValueAndDecrementGreaterValues(&model->main.outputIndexes, index);
+}
+
+static bool removeOperandSkip(size_t operandIndex, const Model& model,
+                              const std::vector<uint32_t>& numberOfConsumers) {
+    if (numberOfConsumers[operandIndex] == 0) {
+        // Removing an unused operand has no effect.
+        return true;
+    }
+    for (const Operation& operation : model.main.operations) {
+        // Skip removeOperandTest for the following operations.
+        // - SPLIT's outputs are not checked during prepareModel.
+        if (operation.type == OperationType::SPLIT) {
+            for (const size_t index : operation.outputs) {
+                if (index == operandIndex) {
+                    return true;
+                }
+            }
+        }
+        // BIDIRECTIONAL_SEQUENCE_LSTM and BIDIRECTIONAL_SEQUENCE_RNN can have
+        // either one, two, three or four outputs depending on their
+        // mergeOutputs parameter and if state outputs are provided.
+        // UNIDIRECTIONAL_SEQUENCE_LSTM and UNIDIRECTIONAL_SEQUENCE_RNN can have
+        // either one or three outputs depending on whether state outputs are
+        // provided.
+        if (operation.type == OperationType::UNIDIRECTIONAL_SEQUENCE_LSTM ||
+            operation.type == OperationType::UNIDIRECTIONAL_SEQUENCE_RNN ||
+            operation.type == OperationType::BIDIRECTIONAL_SEQUENCE_LSTM ||
+            operation.type == OperationType::BIDIRECTIONAL_SEQUENCE_RNN) {
+            for (const size_t index : operation.outputs) {
+                if (index == operandIndex) {
+                    return true;
+                }
+            }
+        }
+    }
+    return false;
+}
+
+static void removeOperandTest(const std::shared_ptr<IDevice>& device, const Model& model,
+                              const std::vector<uint32_t>& numberOfConsumers) {
+    for (size_t operand = 0; operand < model.main.operands.size(); ++operand) {
+        if (removeOperandSkip(operand, model, numberOfConsumers)) {
+            continue;
+        }
+        const std::string message = "removeOperandTest: operand " + std::to_string(operand);
+        validate(device, message, model, [operand](Model* model, ExecutionPreference*, Priority*) {
+            removeOperand(model, operand);
+        });
+    }
+}
+
+///////////////////////// REMOVE OPERATION /////////////////////////
+
+static void removeOperation(Model* model, uint32_t index) {
+    auto& operations = model->main.operations;
+    operations.erase(operations.begin() + index);
+}
+
+static void removeOperationTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    for (size_t operation = 0; operation < model.main.operations.size(); ++operation) {
+        const std::string message = "removeOperationTest: operation " + std::to_string(operation);
+        validate(device, message, model,
+                 [operation](Model* model, ExecutionPreference*, Priority*) {
+                     removeOperation(model, operation);
+                 });
+    }
+}
+
+///////////////////////// REMOVE OPERATION INPUT /////////////////////////
+
+static bool removeOperationInputSkip(const Operation& op, size_t input) {
+    // Skip removeOperationInputTest for the following operations.
+    // - CONCATENATION has at least 2 inputs, with the last element being INT32.
+    // - CONV_2D, DEPTHWISE_CONV_2D, MAX_POOL_2D, AVERAGE_POOL_2D, L2_POOL_2D, RESIZE_BILINEAR,
+    //   SPACE_TO_DEPTH, SPACE_TO_DEPTH, SPACE_TO_BATCH_ND, BATCH_TO_SPACE_ND can have an optional
+    //   layout parameter.
+    //   RESIZE_BILINEAR and RESIZE_NEAREST_NEIGHBOR can have optional
+    //   align_corners and half_pixel_centers parameters.
+    // - L2_NORMALIZATION, LOCAL_RESPONSE_NORMALIZATION, SOFTMAX can have an optional axis
+    //   parameter.
+    switch (op.type) {
+        case OperationType::CONCATENATION: {
+            if (op.inputs.size() > 2 && input != op.inputs.size() - 1) {
+                return true;
+            }
+        } break;
+        case OperationType::DEPTHWISE_CONV_2D: {
+            if ((op.inputs.size() == 12 && input == 11) || (op.inputs.size() == 9 && input == 8)) {
+                return true;
+            }
+        } break;
+        case OperationType::CONV_2D:
+        case OperationType::AVERAGE_POOL_2D:
+        case OperationType::MAX_POOL_2D:
+        case OperationType::L2_POOL_2D: {
+            if ((op.inputs.size() == 11 && input == 10) || (op.inputs.size() == 8 && input == 7)) {
+                return true;
+            }
+        } break;
+        case OperationType::RESIZE_BILINEAR: {
+            if (op.inputs.size() >= 4 && input >= 3) {
+                return true;
+            }
+        } break;
+        case OperationType::RESIZE_NEAREST_NEIGHBOR: {
+            if (op.inputs.size() >= 5 && input >= 3) {
+                return true;
+            }
+        } break;
+        case OperationType::SPACE_TO_DEPTH:
+        case OperationType::DEPTH_TO_SPACE:
+        case OperationType::BATCH_TO_SPACE_ND: {
+            if (op.inputs.size() == 3 && input == 2) {
+                return true;
+            }
+        } break;
+        case OperationType::SPACE_TO_BATCH_ND: {
+            if (op.inputs.size() == 4 && input == 3) {
+                return true;
+            }
+        } break;
+        case OperationType::L2_NORMALIZATION: {
+            if (op.inputs.size() == 2 && input == 1) {
+                return true;
+            }
+        } break;
+        case OperationType::LOCAL_RESPONSE_NORMALIZATION: {
+            if (op.inputs.size() == 6 && input == 5) {
+                return true;
+            }
+        } break;
+        case OperationType::SOFTMAX: {
+            if (op.inputs.size() == 3 && input == 2) {
+                return true;
+            }
+        } break;
+        default:
+            break;
+    }
+    return false;
+}
+
+static void removeOperationInputTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    for (size_t operation = 0; operation < model.main.operations.size(); ++operation) {
+        for (size_t input = 0; input < model.main.operations[operation].inputs.size(); ++input) {
+            const Operation& op = model.main.operations[operation];
+            if (removeOperationInputSkip(op, input)) {
+                continue;
+            }
+            const std::string message = "removeOperationInputTest: operation " +
+                                        std::to_string(operation) + ", input " +
+                                        std::to_string(input);
+            validate(device, message, model,
+                     [operation, input](Model* model, ExecutionPreference*, Priority*) {
+                         auto& inputs = model->main.operations[operation].inputs;
+                         inputs.erase(inputs.begin() + input);
+                     });
+        }
+    }
+}
+
+///////////////////////// REMOVE OPERATION OUTPUT /////////////////////////
+
+static void removeOperationOutputTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    for (size_t operation = 0; operation < model.main.operations.size(); ++operation) {
+        for (size_t output = 0; output < model.main.operations[operation].outputs.size();
+             ++output) {
+            const std::string message = "removeOperationOutputTest: operation " +
+                                        std::to_string(operation) + ", output " +
+                                        std::to_string(output);
+            validate(device, message, model,
+                     [operation, output](Model* model, ExecutionPreference*, Priority*) {
+                         auto& outputs = model->main.operations[operation].outputs;
+                         outputs.erase(outputs.begin() + output);
+                     });
+        }
+    }
+}
+
+///////////////////////// MODEL VALIDATION /////////////////////////
+
+// TODO: remove model input
+// TODO: remove model output
+// TODO: add unused operation
+
+///////////////////////// ADD OPERATION INPUT /////////////////////////
+
+static bool addOperationInputSkip(const Operation& op) {
+    // Skip addOperationInputTest for the following operations.
+    // - L2_NORMALIZATION, LOCAL_RESPONSE_NORMALIZATION, SOFTMAX can have an optional INT32 axis
+    //   parameter.
+    if ((op.type == OperationType::L2_NORMALIZATION && op.inputs.size() == 1) ||
+        (op.type == OperationType::LOCAL_RESPONSE_NORMALIZATION && op.inputs.size() == 5) ||
+        (op.type == OperationType::SOFTMAX && op.inputs.size() == 2) ||
+        (op.type == OperationType::RESIZE_BILINEAR && op.inputs.size() < 6) ||
+        (op.type == OperationType::RESIZE_NEAREST_NEIGHBOR && op.inputs.size() < 6)) {
+        return true;
+    }
+    return false;
+}
+
+static void addOperationInputTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    for (size_t operation = 0; operation < model.main.operations.size(); ++operation) {
+        if (addOperationInputSkip(model.main.operations[operation])) {
+            continue;
+        }
+        const std::string message = "addOperationInputTest: operation " + std::to_string(operation);
+        validate(device, message, model,
+                 [operation](Model* model, ExecutionPreference*, Priority*) {
+                     uint32_t index = addOperand(model, OperandLifeTime::SUBGRAPH_INPUT);
+                     model->main.operations[operation].inputs.push_back(index);
+                     model->main.inputIndexes.push_back(index);
+                 });
+    }
+}
+
+///////////////////////// ADD OPERATION OUTPUT /////////////////////////
+
+static void addOperationOutputTest(const std::shared_ptr<IDevice>& device, const Model& model) {
+    for (size_t operation = 0; operation < model.main.operations.size(); ++operation) {
+        const std::string message =
+                "addOperationOutputTest: operation " + std::to_string(operation);
+        validate(device, message, model,
+                 [operation](Model* model, ExecutionPreference*, Priority*) {
+                     uint32_t index = addOperand(model, OperandLifeTime::SUBGRAPH_OUTPUT);
+                     model->main.operations[operation].outputs.push_back(index);
+                     model->main.outputIndexes.push_back(index);
+                 });
+    }
+}
+
+///////////////////////// VALIDATE EXECUTION PREFERENCE /////////////////////////
+
+static const int32_t invalidExecutionPreferences[] = {
+        static_cast<int32_t>(ExecutionPreference::LOW_POWER) - 1,        // lower bound
+        static_cast<int32_t>(ExecutionPreference::SUSTAINED_SPEED) + 1,  // upper bound
+};
+
+static void mutateExecutionPreferenceTest(const std::shared_ptr<IDevice>& device,
+                                          const Model& model) {
+    for (int32_t invalidPreference : invalidExecutionPreferences) {
+        const std::string message =
+                "mutateExecutionPreferenceTest: preference " + std::to_string(invalidPreference);
+        validate(device, message, model,
+                 [invalidPreference](Model*, ExecutionPreference* preference, Priority*) {
+                     *preference = static_cast<ExecutionPreference>(invalidPreference);
+                 });
+    }
+}
+
+///////////////////////// VALIDATE PRIORITY /////////////////////////
+
+static const int32_t invalidPriorities[] = {
+        static_cast<int32_t>(Priority::LOW) - 1,   // lower bound
+        static_cast<int32_t>(Priority::HIGH) + 1,  // upper bound
+};
+
+static void mutateExecutionPriorityTest(const std::shared_ptr<IDevice>& device,
+                                        const Model& model) {
+    for (int32_t invalidPriority : invalidPriorities) {
+        const std::string message =
+                "mutatePriorityTest: priority " + std::to_string(invalidPriority);
+        validate(device, message, model,
+                 [invalidPriority](Model*, ExecutionPreference*, Priority* priority) {
+                     *priority = static_cast<Priority>(invalidPriority);
+                 });
+    }
+}
+
+////////////////////////// ENTRY POINT //////////////////////////////
+
+void validateModel(const std::shared_ptr<IDevice>& device, const Model& model) {
+    const auto numberOfConsumers = nn::countNumberOfConsumers(
+            model.main.operands.size(), nn::convert(model.main.operations).value());
+    mutateExecutionOrderTest(device, model, numberOfConsumers);
+    mutateOperandTypeTest(device, model);
+    mutateOperandRankTest(device, model);
+    mutateOperandScaleTest(device, model);
+    mutateOperandZeroPointTest(device, model);
+    mutateOperandLifeTimeTest(device, model);
+    mutateOperandInputOutputTest(device, model);
+    mutateOperandAddWriterTest(device, model);
+    mutateOperationOperandTypeTest(device, model);
+    mutateOperationTypeTest(device, model);
+    mutateOperationInputOperandIndexTest(device, model);
+    mutateOperationOutputOperandIndexTest(device, model);
+    mutateOperationRemoveWriteTest(device, model, numberOfConsumers);
+    removeOperandTest(device, model, numberOfConsumers);
+    removeOperationTest(device, model);
+    removeOperationInputTest(device, model);
+    removeOperationOutputTest(device, model);
+    addOperationInputTest(device, model);
+    addOperationOutputTest(device, model);
+    mutateExecutionPreferenceTest(device, model);
+    mutateExecutionPriorityTest(device, model);
+}
+
+}  // namespace aidl::android::hardware::neuralnetworks::vts::functional
diff --git a/neuralnetworks/aidl/vts/functional/ValidateRequest.cpp b/neuralnetworks/aidl/vts/functional/ValidateRequest.cpp
new file mode 100644
index 0000000..db8f429
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/ValidateRequest.cpp
@@ -0,0 +1,126 @@
+/*
+ * 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.
+ */
+
+#define LOG_TAG "neuralnetworks_aidl_hal_test"
+
+#include <android/binder_auto_utils.h>
+
+#include <chrono>
+
+#include <TestHarness.h>
+#include <nnapi/hal/aidl/Utils.h>
+
+#include "Callbacks.h"
+#include "GeneratedTestHarness.h"
+#include "Utils.h"
+#include "VtsHalNeuralnetworks.h"
+
+namespace aidl::android::hardware::neuralnetworks::vts::functional {
+
+using ExecutionMutation = std::function<void(Request*)>;
+
+///////////////////////// UTILITY FUNCTIONS /////////////////////////
+
+// Primary validation function. This function will take a valid request, apply a
+// mutation to it to invalidate the request, then pass it to interface calls
+// that use the request.
+static void validate(const std::shared_ptr<IPreparedModel>& preparedModel,
+                     const std::string& message, const Request& originalRequest,
+                     const ExecutionMutation& mutate) {
+    Request request = utils::clone(originalRequest).value();
+    mutate(&request);
+
+    // We'd like to test both with timing requested and without timing
+    // requested. Rather than running each test both ways, we'll decide whether
+    // to request timing by hashing the message. We do not use std::hash because
+    // it is not guaranteed stable across executions.
+    char hash = 0;
+    for (auto c : message) {
+        hash ^= c;
+    };
+    bool measure = (hash & 1);
+
+    // synchronous
+    {
+        SCOPED_TRACE(message + " [executeSynchronously]");
+        ExecutionResult executionResult;
+        const auto executeStatus = preparedModel->executeSynchronously(
+                request, measure, kNoDeadline, kOmittedTimeoutDuration, &executionResult);
+        ASSERT_FALSE(executeStatus.isOk());
+        ASSERT_EQ(executeStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
+        ASSERT_EQ(static_cast<ErrorStatus>(executeStatus.getServiceSpecificError()),
+                  ErrorStatus::INVALID_ARGUMENT);
+    }
+
+    // fenced
+    {
+        SCOPED_TRACE(message + " [executeFenced]");
+        ndk::ScopedFileDescriptor syncFence;
+        std::shared_ptr<IFencedExecutionCallback> callback;
+        const auto executeStatus = preparedModel->executeFenced(request, {}, false, kNoDeadline,
+                                                                kOmittedTimeoutDuration,
+                                                                kNoDuration, &syncFence, &callback);
+        ASSERT_FALSE(executeStatus.isOk());
+        ASSERT_EQ(executeStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
+        ASSERT_EQ(static_cast<ErrorStatus>(executeStatus.getServiceSpecificError()),
+                  ErrorStatus::INVALID_ARGUMENT);
+    }
+}
+
+///////////////////////// REMOVE INPUT ////////////////////////////////////
+
+static void removeInputTest(const std::shared_ptr<IPreparedModel>& preparedModel,
+                            const Request& request) {
+    for (size_t input = 0; input < request.inputs.size(); ++input) {
+        const std::string message = "removeInput: removed input " + std::to_string(input);
+        validate(preparedModel, message, request, [input](Request* request) {
+            request->inputs.erase(request->inputs.begin() + input);
+        });
+    }
+}
+
+///////////////////////// REMOVE OUTPUT ////////////////////////////////////
+
+static void removeOutputTest(const std::shared_ptr<IPreparedModel>& preparedModel,
+                             const Request& request) {
+    for (size_t output = 0; output < request.outputs.size(); ++output) {
+        const std::string message = "removeOutput: removed Output " + std::to_string(output);
+        validate(preparedModel, message, request, [output](Request* request) {
+            request->outputs.erase(request->outputs.begin() + output);
+        });
+    }
+}
+
+///////////////////////////// ENTRY POINT //////////////////////////////////
+
+void validateRequest(const std::shared_ptr<IPreparedModel>& preparedModel, const Request& request) {
+    removeInputTest(preparedModel, request);
+    removeOutputTest(preparedModel, request);
+}
+
+void validateRequestFailure(const std::shared_ptr<IPreparedModel>& preparedModel,
+                            const Request& request) {
+    SCOPED_TRACE("Expecting request to fail [executeSynchronously]");
+    ExecutionResult executionResult;
+    const auto executeStatus = preparedModel->executeSynchronously(
+            request, false, kNoDeadline, kOmittedTimeoutDuration, &executionResult);
+
+    ASSERT_FALSE(executeStatus.isOk());
+    ASSERT_EQ(executeStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
+    ASSERT_NE(static_cast<ErrorStatus>(executeStatus.getServiceSpecificError()), ErrorStatus::NONE);
+}
+
+}  // namespace aidl::android::hardware::neuralnetworks::vts::functional
diff --git a/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.cpp b/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.cpp
new file mode 100644
index 0000000..2d91b8e
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.cpp
@@ -0,0 +1,194 @@
+/*
+ * 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.
+ */
+
+#define LOG_TAG "neuralnetworks_aidl_hal_test"
+#include "VtsHalNeuralnetworks.h"
+
+#include <android-base/logging.h>
+#include <android/binder_auto_utils.h>
+#include <android/binder_interface_utils.h>
+#include <android/binder_manager.h>
+#include <android/binder_status.h>
+#include <gtest/gtest.h>
+#include <memory>
+#include <string>
+#include <utility>
+
+#include <TestHarness.h>
+#include <aidl/Vintf.h>
+#include <nnapi/hal/aidl/Conversions.h>
+
+#include "Callbacks.h"
+#include "GeneratedTestHarness.h"
+#include "Utils.h"
+
+namespace aidl::android::hardware::neuralnetworks::vts::functional {
+
+using implementation::PreparedModelCallback;
+
+// internal helper function
+void createPreparedModel(const std::shared_ptr<IDevice>& device, const Model& model,
+                         std::shared_ptr<IPreparedModel>* preparedModel, bool reportSkipping) {
+    ASSERT_NE(nullptr, preparedModel);
+    *preparedModel = nullptr;
+
+    // see if service can handle model
+    std::vector<bool> supportedOperations;
+    const auto supportedCallStatus = device->getSupportedOperations(model, &supportedOperations);
+    ASSERT_TRUE(supportedCallStatus.isOk());
+    ASSERT_NE(0ul, supportedOperations.size());
+    const bool fullySupportsModel = std::all_of(
+            supportedOperations.begin(), supportedOperations.end(), [](bool v) { return v; });
+
+    // launch prepare model
+    const std::shared_ptr<PreparedModelCallback> preparedModelCallback =
+            ndk::SharedRefBase::make<PreparedModelCallback>();
+    const auto prepareLaunchStatus =
+            device->prepareModel(model, ExecutionPreference::FAST_SINGLE_ANSWER, kDefaultPriority,
+                                 kNoDeadline, {}, {}, kEmptyCacheToken, preparedModelCallback);
+    ASSERT_TRUE(prepareLaunchStatus.isOk()) << prepareLaunchStatus.getDescription();
+
+    // retrieve prepared model
+    preparedModelCallback->wait();
+    const ErrorStatus prepareReturnStatus = preparedModelCallback->getStatus();
+    *preparedModel = preparedModelCallback->getPreparedModel();
+
+    // The getSupportedOperations call returns a list of operations that are guaranteed not to fail
+    // if prepareModel is called, and 'fullySupportsModel' is true i.f.f. the entire model is
+    // guaranteed. If a driver has any doubt that it can prepare an operation, it must return false.
+    // So here, if a driver isn't sure if it can support an operation, but reports that it
+    // successfully prepared the model, the test can continue.
+    if (!fullySupportsModel && prepareReturnStatus != ErrorStatus::NONE) {
+        ASSERT_EQ(nullptr, preparedModel->get());
+        if (!reportSkipping) {
+            return;
+        }
+        LOG(INFO) << "NN VTS: Early termination of test because vendor service cannot prepare "
+                     "model that it does not support.";
+        std::cout << "[          ]   Early termination of test because vendor service cannot "
+                     "prepare model that it does not support."
+                  << std::endl;
+        GTEST_SKIP();
+    }
+
+    ASSERT_EQ(ErrorStatus::NONE, prepareReturnStatus);
+    ASSERT_NE(nullptr, preparedModel->get());
+}
+
+void NeuralNetworksAidlTest::SetUp() {
+    testing::TestWithParam<NeuralNetworksAidlTestParam>::SetUp();
+    ASSERT_NE(kDevice, nullptr);
+}
+
+static NamedDevice makeNamedDevice(const std::string& name) {
+    ndk::SpAIBinder binder(AServiceManager_getService(name.c_str()));
+    return {name, IDevice::fromBinder(binder)};
+}
+
+static std::vector<NamedDevice> getNamedDevicesImpl() {
+    // Retrieves the name of all service instances that implement IDevice,
+    // including any Lazy HAL instances.
+    const std::vector<std::string> names = ::android::getAidlHalInstanceNames(IDevice::descriptor);
+
+    // Get a handle to each device and pair it with its name.
+    std::vector<NamedDevice> namedDevices;
+    namedDevices.reserve(names.size());
+    std::transform(names.begin(), names.end(), std::back_inserter(namedDevices), makeNamedDevice);
+    return namedDevices;
+}
+
+const std::vector<NamedDevice>& getNamedDevices() {
+    const static std::vector<NamedDevice> devices = getNamedDevicesImpl();
+    return devices;
+}
+
+std::string printNeuralNetworksAidlTest(
+        const testing::TestParamInfo<NeuralNetworksAidlTestParam>& info) {
+    return gtestCompliantName(getName(info.param));
+}
+
+INSTANTIATE_DEVICE_TEST(NeuralNetworksAidlTest);
+
+// Forward declaration from ValidateModel.cpp
+void validateModel(const std::shared_ptr<IDevice>& device, const Model& model);
+// Forward declaration from ValidateRequest.cpp
+void validateRequest(const std::shared_ptr<IPreparedModel>& preparedModel, const Request& request);
+// Forward declaration from ValidateRequest.cpp
+void validateRequestFailure(const std::shared_ptr<IPreparedModel>& preparedModel,
+                            const Request& request);
+
+void validateEverything(const std::shared_ptr<IDevice>& device, const Model& model,
+                        const Request& request) {
+    validateModel(device, model);
+
+    // Create IPreparedModel.
+    std::shared_ptr<IPreparedModel> preparedModel;
+    createPreparedModel(device, model, &preparedModel);
+    if (preparedModel == nullptr) return;
+
+    validateRequest(preparedModel, request);
+    // HIDL also had test that expected executeFenced to fail on received null fd (-1). This is not
+    // allowed in AIDL and will result in EX_TRANSACTION_FAILED.
+}
+
+void validateFailure(const std::shared_ptr<IDevice>& device, const Model& model,
+                     const Request& request) {
+    // TODO: Should this always succeed?
+    //       What if the invalid input is part of the model (i.e., a parameter).
+    validateModel(device, model);
+
+    // Create IPreparedModel.
+    std::shared_ptr<IPreparedModel> preparedModel;
+    createPreparedModel(device, model, &preparedModel);
+    if (preparedModel == nullptr) return;
+
+    validateRequestFailure(preparedModel, request);
+}
+
+TEST_P(ValidationTest, Test) {
+    const Model model = createModel(kTestModel);
+    ExecutionContext context;
+    const Request request = context.createRequest(kTestModel);
+    if (kTestModel.expectFailure) {
+        validateFailure(kDevice, model, request);
+    } else {
+        validateEverything(kDevice, model, request);
+    }
+}
+
+INSTANTIATE_GENERATED_TEST(ValidationTest, [](const std::string& testName) {
+    // Skip validation for the "inputs_as_internal" and "all_tensors_as_inputs"
+    // generated tests.
+    return testName.find("inputs_as_internal") == std::string::npos &&
+           testName.find("all_tensors_as_inputs") == std::string::npos;
+});
+
+std::string toString(Executor executor) {
+    switch (executor) {
+        case Executor::ASYNC:
+            return "ASYNC";
+        case Executor::SYNC:
+            return "SYNC";
+        case Executor::BURST:
+            return "BURST";
+        case Executor::FENCED:
+            return "FENCED";
+        default:
+            CHECK(false);
+    }
+}
+
+}  // namespace aidl::android::hardware::neuralnetworks::vts::functional
diff --git a/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.h b/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.h
new file mode 100644
index 0000000..9b81ee1
--- /dev/null
+++ b/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.h
@@ -0,0 +1,61 @@
+/*
+ * 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_NEURALNETWORKS_AIDL_VTS_HAL_NEURALNETWORKS_H
+#define ANDROID_HARDWARE_NEURALNETWORKS_AIDL_VTS_HAL_NEURALNETWORKS_H
+
+#include <gtest/gtest.h>
+#include <vector>
+
+#include <aidl/android/hardware/neuralnetworks/IDevice.h>
+
+#include "Callbacks.h"
+#include "Utils.h"
+
+namespace aidl::android::hardware::neuralnetworks::vts::functional {
+
+using NamedDevice = Named<std::shared_ptr<IDevice>>;
+using NeuralNetworksAidlTestParam = NamedDevice;
+
+class NeuralNetworksAidlTest : public testing::TestWithParam<NeuralNetworksAidlTestParam> {
+  protected:
+    void SetUp() override;
+    const std::shared_ptr<IDevice> kDevice = getData(GetParam());
+};
+
+const std::vector<NamedDevice>& getNamedDevices();
+
+std::string printNeuralNetworksAidlTest(
+        const testing::TestParamInfo<NeuralNetworksAidlTestParam>& info);
+
+#define INSTANTIATE_DEVICE_TEST(TestSuite)                                                 \
+    GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(TestSuite);                              \
+    INSTANTIATE_TEST_SUITE_P(PerInstance, TestSuite, testing::ValuesIn(getNamedDevices()), \
+                             printNeuralNetworksAidlTest)
+
+// Create an IPreparedModel object. If the model cannot be prepared,
+// "preparedModel" will be nullptr instead.
+void createPreparedModel(const std::shared_ptr<IDevice>& device, const Model& model,
+                         std::shared_ptr<IPreparedModel>* preparedModel,
+                         bool reportSkipping = true);
+
+enum class Executor { ASYNC, SYNC, BURST, FENCED };
+
+std::string toString(Executor executor);
+
+}  // namespace aidl::android::hardware::neuralnetworks::vts::functional
+
+#endif  // ANDROID_HARDWARE_NEURALNETWORKS_AIDL_VTS_HAL_NEURALNETWORKS_H