diff --git a/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp
index e28605d..4ab228f 100644
--- a/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp
+++ b/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp
@@ -125,7 +125,9 @@
 // Test driver for those generated from ml/nn/runtime/test/spec
 void Execute(const sp<IDevice>& device, const TestModel& testModel) {
     const Model model = createModel(testModel);
-    const Request request = createRequest(testModel);
+
+    ExecutionContext context;
+    const Request request = context.createRequest(testModel);
 
     // Create IPreparedModel.
     sp<IPreparedModel> preparedModel;
@@ -143,7 +145,7 @@
     ASSERT_EQ(ErrorStatus::NONE, executionCallback->getStatus());
 
     // Retrieve execution results.
-    const std::vector<TestBuffer> outputs = getOutputBuffers(request);
+    const std::vector<TestBuffer> outputs = context.getOutputBuffers(request);
 
     // We want "close-enough" results.
     checkResults(testModel, outputs);
diff --git a/neuralnetworks/1.0/vts/functional/Utils.cpp b/neuralnetworks/1.0/vts/functional/Utils.cpp
index 0dba85a..3613e69 100644
--- a/neuralnetworks/1.0/vts/functional/Utils.cpp
+++ b/neuralnetworks/1.0/vts/functional/Utils.cpp
@@ -21,10 +21,13 @@
 
 #include <android-base/logging.h>
 #include <android/hardware/neuralnetworks/1.0/types.h>
+#include <android/hardware_buffer.h>
 #include <android/hidl/allocator/1.0/IAllocator.h>
 #include <android/hidl/memory/1.0/IMemory.h>
 #include <hidlmemory/mapping.h>
+#include <vndk/hardware_buffer.h>
 
+#include <gtest/gtest.h>
 #include <algorithm>
 #include <iostream>
 #include <vector>
@@ -37,10 +40,64 @@
 using V1_0::Request;
 using V1_0::RequestArgument;
 
-constexpr uint32_t kInputPoolIndex = 0;
-constexpr uint32_t kOutputPoolIndex = 1;
+std::unique_ptr<TestAshmem> TestAshmem::create(uint32_t size) {
+    auto ashmem = std::make_unique<TestAshmem>(size);
+    return ashmem->mIsValid ? std::move(ashmem) : nullptr;
+}
 
-Request createRequest(const TestModel& testModel) {
+void TestAshmem::initialize(uint32_t size) {
+    mIsValid = false;
+    ASSERT_GT(size, 0);
+    mHidlMemory = nn::allocateSharedMemory(size);
+    ASSERT_TRUE(mHidlMemory.valid());
+    mMappedMemory = mapMemory(mHidlMemory);
+    ASSERT_NE(mMappedMemory, nullptr);
+    mPtr = static_cast<uint8_t*>(static_cast<void*>(mMappedMemory->getPointer()));
+    ASSERT_NE(mPtr, nullptr);
+    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);
+
+    void* buffer = nullptr;
+    ASSERT_EQ(AHardwareBuffer_lock(mAhwb, usage, -1, nullptr, &buffer), 0);
+    ASSERT_NE(buffer, nullptr);
+    mPtr = static_cast<uint8_t*>(buffer);
+
+    const native_handle_t* handle = AHardwareBuffer_getNativeHandle(mAhwb);
+    ASSERT_NE(handle, nullptr);
+    mHidlMemory = hidl_memory("hardware_buffer_blob", handle, desc.width);
+    mIsValid = true;
+}
+
+TestBlobAHWB::~TestBlobAHWB() {
+    if (mAhwb) {
+        AHardwareBuffer_unlock(mAhwb, nullptr);
+        AHardwareBuffer_release(mAhwb);
+    }
+}
+
+Request ExecutionContext::createRequest(const TestModel& testModel, MemoryType memoryType) {
+    CHECK(memoryType == MemoryType::ASHMEM || memoryType == MemoryType::BLOB_AHWB);
+
     // Model inputs.
     hidl_vec<RequestArgument> inputs(testModel.main.inputIndexes.size());
     size_t inputSize = 0;
@@ -80,16 +137,19 @@
     }
 
     // Allocate memory pools.
-    hidl_vec<hidl_memory> pools = {nn::allocateSharedMemory(inputSize),
-                                   nn::allocateSharedMemory(outputSize)};
-    CHECK_NE(pools[kInputPoolIndex].size(), 0u);
-    CHECK_NE(pools[kOutputPoolIndex].size(), 0u);
-    sp<IMemory> inputMemory = mapMemory(pools[kInputPoolIndex]);
-    CHECK(inputMemory.get() != nullptr);
-    uint8_t* inputPtr = static_cast<uint8_t*>(static_cast<void*>(inputMemory->getPointer()));
-    CHECK(inputPtr != nullptr);
+    if (memoryType == MemoryType::ASHMEM) {
+        mInputMemory = TestAshmem::create(inputSize);
+        mOutputMemory = TestAshmem::create(outputSize);
+    } else {
+        mInputMemory = TestBlobAHWB::create(inputSize);
+        mOutputMemory = TestBlobAHWB::create(outputSize);
+    }
+    EXPECT_NE(mInputMemory, nullptr);
+    EXPECT_NE(mOutputMemory, nullptr);
+    hidl_vec<hidl_memory> pools = {mInputMemory->getHidlMemory(), mOutputMemory->getHidlMemory()};
 
     // 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) {
@@ -102,18 +162,13 @@
     return {.inputs = std::move(inputs), .outputs = std::move(outputs), .pools = std::move(pools)};
 }
 
-std::vector<TestBuffer> getOutputBuffers(const Request& request) {
-    sp<IMemory> outputMemory = mapMemory(request.pools[kOutputPoolIndex]);
-    CHECK(outputMemory.get() != nullptr);
-    uint8_t* outputPtr = static_cast<uint8_t*>(static_cast<void*>(outputMemory->getPointer()));
-    CHECK(outputPtr != nullptr);
-
+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;
 }
 
diff --git a/neuralnetworks/1.0/vts/functional/VtsHalNeuralnetworks.cpp b/neuralnetworks/1.0/vts/functional/VtsHalNeuralnetworks.cpp
index cb22250..7f7dac0 100644
--- a/neuralnetworks/1.0/vts/functional/VtsHalNeuralnetworks.cpp
+++ b/neuralnetworks/1.0/vts/functional/VtsHalNeuralnetworks.cpp
@@ -129,7 +129,8 @@
 
 TEST_P(ValidationTest, Test) {
     const Model model = createModel(kTestModel);
-    const Request request = createRequest(kTestModel);
+    ExecutionContext context;
+    const Request request = context.createRequest(kTestModel);
     ASSERT_FALSE(kTestModel.expectFailure);
     validateEverything(kDevice, model, request);
 }
diff --git a/neuralnetworks/1.0/vts/functional/include/1.0/Utils.h b/neuralnetworks/1.0/vts/functional/include/1.0/Utils.h
index 6d4534c..3292f79 100644
--- a/neuralnetworks/1.0/vts/functional/include/1.0/Utils.h
+++ b/neuralnetworks/1.0/vts/functional/include/1.0/Utils.h
@@ -19,6 +19,8 @@
 
 #include <android-base/logging.h>
 #include <android/hardware/neuralnetworks/1.0/types.h>
+#include <android/hardware_buffer.h>
+#include <android/hidl/memory/1.0/IMemory.h>
 #include <algorithm>
 #include <iosfwd>
 #include <string>
@@ -28,11 +30,73 @@
 
 namespace android::hardware::neuralnetworks {
 
-// Create HIDL Request from the TestModel struct.
-V1_0::Request createRequest(const test_helper::TestModel& testModel);
+// Convenience class to manage the lifetime of memory resources.
+class TestMemoryBase {
+    DISALLOW_COPY_AND_ASSIGN(TestMemoryBase);
 
-// After execution, copy out output results from the output memory pool.
-std::vector<::test_helper::TestBuffer> getOutputBuffers(const V1_0::Request& request);
+  public:
+    TestMemoryBase() = default;
+    virtual ~TestMemoryBase() = default;
+    uint8_t* getPointer() const { return mPtr; }
+    hidl_memory getHidlMemory() const { return mHidlMemory; }
+
+  protected:
+    uint8_t* mPtr = nullptr;
+    hidl_memory mHidlMemory;
+    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);
+    sp<hidl::memory::V1_0::IMemory> 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;
+};
+
+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.
+    V1_0::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 V1_0::Request& request) const;
+
+  private:
+    std::unique_ptr<TestMemoryBase> mInputMemory, mOutputMemory;
+};
 
 // Delete element from hidl_vec. hidl_vec doesn't support a "remove" operation,
 // so this is efficiently accomplished by moving the element to the end and
diff --git a/neuralnetworks/1.1/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/1.1/vts/functional/GeneratedTestHarness.cpp
index cee15a3..14d300d 100644
--- a/neuralnetworks/1.1/vts/functional/GeneratedTestHarness.cpp
+++ b/neuralnetworks/1.1/vts/functional/GeneratedTestHarness.cpp
@@ -133,7 +133,9 @@
 // Test driver for those generated from ml/nn/runtime/test/spec
 void Execute(const sp<IDevice>& device, const TestModel& testModel) {
     const Model model = createModel(testModel);
-    const Request request = createRequest(testModel);
+
+    ExecutionContext context;
+    const Request request = context.createRequest(testModel);
 
     // Create IPreparedModel.
     sp<IPreparedModel> preparedModel;
@@ -151,7 +153,7 @@
     ASSERT_EQ(ErrorStatus::NONE, executionCallback->getStatus());
 
     // Retrieve execution results.
-    const std::vector<TestBuffer> outputs = getOutputBuffers(request);
+    const std::vector<TestBuffer> outputs = context.getOutputBuffers(request);
 
     // We want "close-enough" results.
     checkResults(testModel, outputs);
diff --git a/neuralnetworks/1.1/vts/functional/VtsHalNeuralnetworks.cpp b/neuralnetworks/1.1/vts/functional/VtsHalNeuralnetworks.cpp
index d56d40b..04af6ec 100644
--- a/neuralnetworks/1.1/vts/functional/VtsHalNeuralnetworks.cpp
+++ b/neuralnetworks/1.1/vts/functional/VtsHalNeuralnetworks.cpp
@@ -132,7 +132,8 @@
 
 TEST_P(ValidationTest, Test) {
     const Model model = createModel(kTestModel);
-    const Request request = createRequest(kTestModel);
+    ExecutionContext context;
+    const Request request = context.createRequest(kTestModel);
     ASSERT_FALSE(kTestModel.expectFailure);
     validateEverything(kDevice, model, request);
 }
diff --git a/neuralnetworks/1.2/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/1.2/vts/functional/GeneratedTestHarness.cpp
index 3ab0135..aaaafc7 100644
--- a/neuralnetworks/1.2/vts/functional/GeneratedTestHarness.cpp
+++ b/neuralnetworks/1.2/vts/functional/GeneratedTestHarness.cpp
@@ -68,6 +68,7 @@
     Executor executor;
     MeasureTiming measureTiming;
     OutputType outputType;
+    MemoryType memoryType;
 };
 
 }  // namespace
@@ -216,7 +217,8 @@
         return;
     }
 
-    Request request = createRequest(testModel);
+    ExecutionContext context;
+    Request request = context.createRequest(testModel, testConfig.memoryType);
     if (testConfig.outputType == OutputType::INSUFFICIENT) {
         makeOutputInsufficientSize(/*outputIndex=*/0, &request);
     }
@@ -326,7 +328,7 @@
     }
 
     // Retrieve execution results.
-    const std::vector<TestBuffer> outputs = getOutputBuffers(request);
+    const std::vector<TestBuffer> outputs = context.getOutputBuffers(request);
 
     // We want "close-enough" results.
     checkResults(testModel, outputs);
@@ -337,24 +339,30 @@
     std::vector<OutputType> outputTypesList;
     std::vector<MeasureTiming> measureTimingList;
     std::vector<Executor> executorList;
+    std::vector<MemoryType> memoryTypeList;
 
     if (testDynamicOutputShape) {
         outputTypesList = {OutputType::UNSPECIFIED, OutputType::INSUFFICIENT};
         measureTimingList = {MeasureTiming::NO, MeasureTiming::YES};
         executorList = {Executor::ASYNC, Executor::SYNC, Executor::BURST};
+        memoryTypeList = {MemoryType::ASHMEM};
     } else {
         outputTypesList = {OutputType::FULLY_SPECIFIED};
         measureTimingList = {MeasureTiming::NO, MeasureTiming::YES};
         executorList = {Executor::ASYNC, Executor::SYNC, Executor::BURST};
+        memoryTypeList = {MemoryType::ASHMEM, MemoryType::BLOB_AHWB};
     }
 
     for (const OutputType outputType : outputTypesList) {
         for (const MeasureTiming measureTiming : measureTimingList) {
             for (const Executor executor : executorList) {
-                const TestConfig testConfig = {.executor = executor,
-                                               .measureTiming = measureTiming,
-                                               .outputType = outputType};
-                EvaluatePreparedModel(preparedModel, testModel, testConfig);
+                for (const MemoryType memoryType : memoryTypeList) {
+                    const TestConfig testConfig = {.executor = executor,
+                                                   .measureTiming = measureTiming,
+                                                   .outputType = outputType,
+                                                   .memoryType = memoryType};
+                    EvaluatePreparedModel(preparedModel, testModel, testConfig);
+                }
             }
         }
     }
diff --git a/neuralnetworks/1.2/vts/functional/VtsHalNeuralnetworks.cpp b/neuralnetworks/1.2/vts/functional/VtsHalNeuralnetworks.cpp
index 4fbd0e2..5853fa4 100644
--- a/neuralnetworks/1.2/vts/functional/VtsHalNeuralnetworks.cpp
+++ b/neuralnetworks/1.2/vts/functional/VtsHalNeuralnetworks.cpp
@@ -153,7 +153,8 @@
 
 TEST_P(ValidationTest, Test) {
     const Model model = createModel(kTestModel);
-    const Request request = createRequest(kTestModel);
+    ExecutionContext context;
+    const Request request = context.createRequest(kTestModel);
     if (kTestModel.expectFailure) {
         validateFailure(kDevice, model, request);
     } else {
diff --git a/neuralnetworks/1.3/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/1.3/vts/functional/GeneratedTestHarness.cpp
index aae58bf..45e0f82 100644
--- a/neuralnetworks/1.3/vts/functional/GeneratedTestHarness.cpp
+++ b/neuralnetworks/1.3/vts/functional/GeneratedTestHarness.cpp
@@ -76,8 +76,6 @@
 
 enum class OutputType { FULLY_SPECIFIED, UNSPECIFIED, INSUFFICIENT, MISSED_DEADLINE };
 
-enum class MemoryType { SHARED, DEVICE };
-
 enum class IOType { INPUT, OUTPUT };
 
 static void waitForSyncFence(int syncFd) {
@@ -338,21 +336,39 @@
     }
 }
 
-constexpr uint32_t kInputPoolIndex = 0;
-constexpr uint32_t kOutputPoolIndex = 1;
-constexpr uint32_t kDeviceMemoryBeginIndex = 2;
+class ExecutionContextV1_3 {
+  public:
+    ExecutionContextV1_3(sp<IDevice> device, sp<IPreparedModel> preparedModel)
+        : kDevice(std::move(device)), kPreparedModel(std::move(preparedModel)) {}
 
-static std::pair<Request, std::vector<sp<IBuffer>>> createRequest(
-        const sp<IDevice>& device, const sp<IPreparedModel>& preparedModel,
-        const TestModel& testModel, bool preferDeviceMemory) {
+    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 sp<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 sp<IDevice> kDevice;
+    const sp<IPreparedModel> kPreparedModel;
+    std::unique_ptr<TestMemoryBase> mInputMemory, mOutputMemory;
+    std::vector<sp<IBuffer>> mBuffers;
+};
+
+std::optional<Request> ExecutionContextV1_3::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(device, preparedModel, testModel);
-    std::vector<sp<IBuffer>> buffers;
+    DeviceMemoryAllocator allocator(kDevice, kPreparedModel, testModel);
     std::vector<uint32_t> tokens;
+    mBuffers.clear();
 
     // Model inputs.
     hidl_vec<RequestArgument> inputs(testModel.main.inputIndexes.size());
@@ -363,13 +379,13 @@
             // Omitted input.
             inputs[i] = {.hasNoValue = true};
             continue;
-        } else if (preferDeviceMemory) {
+        } 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<uint32_t>(buffers.size() +
+                DataLocation loc = {.poolIndex = static_cast<uint32_t>(mBuffers.size() +
                                                                        kDeviceMemoryBeginIndex)};
-                buffers.push_back(std::move(buffer));
+                mBuffers.push_back(std::move(buffer));
                 tokens.push_back(token);
                 inputs[i] = {.hasNoValue = false, .location = loc, .dimensions = {}};
                 continue;
@@ -389,13 +405,13 @@
     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 (preferDeviceMemory) {
+        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<uint32_t>(buffers.size() +
+                DataLocation loc = {.poolIndex = static_cast<uint32_t>(mBuffers.size() +
                                                                        kDeviceMemoryBeginIndex)};
-                buffers.push_back(std::move(buffer));
+                mBuffers.push_back(std::move(buffer));
                 tokens.push_back(token);
                 outputs[i] = {.hasNoValue = false, .location = loc, .dimensions = {}};
                 continue;
@@ -418,21 +434,29 @@
         outputs[i] = {.hasNoValue = false, .location = loc, .dimensions = {}};
     }
 
+    if (memoryType == MemoryType::DEVICE && mBuffers.empty()) {
+        return std::nullopt;
+    }
+
     // Memory pools.
-    hidl_vec<Request::MemoryPool> pools(kDeviceMemoryBeginIndex + buffers.size());
-    pools[kInputPoolIndex].hidlMemory(nn::allocateSharedMemory(std::max<size_t>(inputSize, 1)));
-    pools[kOutputPoolIndex].hidlMemory(nn::allocateSharedMemory(std::max<size_t>(outputSize, 1)));
-    CHECK_NE(pools[kInputPoolIndex].hidlMemory().size(), 0u);
-    CHECK_NE(pools[kOutputPoolIndex].hidlMemory().size(), 0u);
-    for (uint32_t i = 0; i < buffers.size(); i++) {
+    hidl_vec<Request::MemoryPool> pools(kDeviceMemoryBeginIndex + mBuffers.size());
+    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));
+    }
+    EXPECT_NE(mInputMemory, nullptr);
+    EXPECT_NE(mOutputMemory, nullptr);
+    pools[kInputPoolIndex].hidlMemory(mInputMemory->getHidlMemory());
+    pools[kOutputPoolIndex].hidlMemory(mOutputMemory->getHidlMemory());
+    for (uint32_t i = 0; i < mBuffers.size(); i++) {
         pools[kDeviceMemoryBeginIndex + i].token(tokens[i]);
     }
 
     // Copy input data to the input shared memory pool.
-    sp<IMemory> inputMemory = mapMemory(pools[kInputPoolIndex].hidlMemory());
-    CHECK(inputMemory.get() != nullptr);
-    uint8_t* inputPtr = static_cast<uint8_t*>(static_cast<void*>(inputMemory->getPointer()));
-    CHECK(inputPtr != nullptr);
+    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]];
@@ -441,14 +465,38 @@
             std::copy(begin, end, inputPtr + inputs[i].location.offset);
         }
     }
-
-    Request request = {
+    return Request{
             .inputs = std::move(inputs), .outputs = std::move(outputs), .pools = std::move(pools)};
-    return {std::move(request), std::move(buffers)};
+}
+
+std::vector<TestBuffer> ExecutionContextV1_3::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.
-static void getBuffer(const sp<IBuffer>& buffer, size_t size, TestBuffer* testBuffer) {
+void ExecutionContextV1_3::getBuffer(const sp<IBuffer>& buffer, size_t size,
+                                     TestBuffer* testBuffer) const {
     // IBuffer -> Shared memory.
     hidl_memory tmp = nn::allocateSharedMemory(size);
     const auto ret = buffer->copyTo(tmp);
@@ -464,35 +512,6 @@
     *testBuffer = TestBuffer(size, outputPtr);
 }
 
-static std::vector<TestBuffer> getOutputBuffers(const TestModel& testModel, const Request& request,
-                                                const std::vector<sp<IBuffer>>& buffers) {
-    sp<IMemory> outputMemory = mapMemory(request.pools[kOutputPoolIndex].hidlMemory());
-    CHECK(outputMemory.get() != nullptr);
-    uint8_t* outputPtr = static_cast<uint8_t*>(static_cast<void*>(outputMemory->getPointer()));
-    CHECK(outputPtr != nullptr);
-
-    // Copy out output results.
-    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();
-            } else {
-                SCOPED_TRACE("Output index = " + std::to_string(i));
-                const uint32_t bufferIndex = outputLoc.poolIndex - kDeviceMemoryBeginIndex;
-                TestBuffer buffer;
-                getBuffer(buffers[bufferIndex], op.data.size(), &buffer);
-                outputBuffers.push_back(std::move(buffer));
-            }
-        }
-    }
-    return outputBuffers;
-}
-
 static bool hasZeroSizedOutput(const TestModel& testModel) {
     return std::any_of(testModel.main.outputIndexes.begin(), testModel.main.outputIndexes.end(),
                        [&testModel](uint32_t index) {
@@ -543,13 +562,14 @@
         return;
     }
 
-    auto [request, buffers] =
-            createRequest(device, preparedModel, testModel,
-                          /*preferDeviceMemory=*/testConfig.memoryType == MemoryType::DEVICE);
+    ExecutionContextV1_3 context(device, preparedModel);
+    auto maybeRequest = context.createRequest(testModel, testConfig.memoryType);
     // Skip if testing memory domain but no device memory has been allocated.
-    if (testConfig.memoryType == MemoryType::DEVICE && buffers.empty()) {
+    if (!maybeRequest.has_value()) {
         return;
     }
+
+    Request request = std::move(maybeRequest.value());
     if (testConfig.outputType == OutputType::INSUFFICIENT) {
         makeOutputInsufficientSize(/*outputIndex=*/0, &request);
     }
@@ -744,7 +764,7 @@
     }
 
     // Retrieve execution results.
-    const std::vector<TestBuffer> outputs = getOutputBuffers(testModel, request, buffers);
+    const std::vector<TestBuffer> outputs = context.getOutputBuffers(testModel, request);
 
     // We want "close-enough" results.
     checkResults(testModel, outputs);
@@ -755,29 +775,32 @@
     std::vector<OutputType> outputTypesList;
     std::vector<MeasureTiming> measureTimingList;
     std::vector<Executor> executorList;
-    MemoryType memoryType = MemoryType::SHARED;
+    std::vector<MemoryType> memoryTypeList;
 
     switch (testKind) {
         case TestKind::GENERAL: {
             outputTypesList = {OutputType::FULLY_SPECIFIED};
             measureTimingList = {MeasureTiming::NO, MeasureTiming::YES};
             executorList = {Executor::ASYNC, Executor::SYNC, Executor::BURST};
+            memoryTypeList = {MemoryType::ASHMEM};
         } break;
         case TestKind::DYNAMIC_SHAPE: {
             outputTypesList = {OutputType::UNSPECIFIED, OutputType::INSUFFICIENT};
             measureTimingList = {MeasureTiming::NO, MeasureTiming::YES};
             executorList = {Executor::ASYNC, Executor::SYNC, Executor::BURST, Executor::FENCED};
+            memoryTypeList = {MemoryType::ASHMEM};
         } break;
         case TestKind::MEMORY_DOMAIN: {
             outputTypesList = {OutputType::FULLY_SPECIFIED};
             measureTimingList = {MeasureTiming::NO};
             executorList = {Executor::ASYNC, Executor::SYNC, Executor::FENCED};
-            memoryType = MemoryType::DEVICE;
+            memoryTypeList = {MemoryType::BLOB_AHWB, MemoryType::DEVICE};
         } break;
         case TestKind::FENCED_COMPUTE: {
             outputTypesList = {OutputType::FULLY_SPECIFIED};
             measureTimingList = {MeasureTiming::NO, MeasureTiming::YES};
             executorList = {Executor::FENCED};
+            memoryTypeList = {MemoryType::ASHMEM};
         } break;
         case TestKind::QUANTIZATION_COUPLING: {
             LOG(FATAL) << "Wrong TestKind for EvaluatePreparedModel";
@@ -788,14 +811,17 @@
             measureTimingList = {MeasureTiming::NO, MeasureTiming::YES};
             // Burst does not support V1_3 loop timeout.
             executorList = {Executor::ASYNC, Executor::SYNC, Executor::FENCED};
+            memoryTypeList = {MemoryType::ASHMEM};
         } break;
     }
 
     for (const OutputType outputType : outputTypesList) {
         for (const MeasureTiming measureTiming : measureTimingList) {
             for (const Executor executor : executorList) {
-                const TestConfig testConfig(executor, measureTiming, outputType, memoryType);
-                EvaluatePreparedModel(device, preparedModel, testModel, testConfig);
+                for (const MemoryType memoryType : memoryTypeList) {
+                    const TestConfig testConfig(executor, measureTiming, outputType, memoryType);
+                    EvaluatePreparedModel(device, preparedModel, testModel, testConfig);
+                }
             }
         }
     }
@@ -814,7 +840,7 @@
     for (const OutputType outputType : outputTypesList) {
         for (const MeasureTiming measureTiming : measureTimingList) {
             for (const Executor executor : executorList) {
-                const TestConfig testConfig(executor, measureTiming, outputType, MemoryType::SHARED,
+                const TestConfig testConfig(executor, measureTiming, outputType, MemoryType::ASHMEM,
                                             /*reportSkipping=*/false);
                 bool baseSkipped = false;
                 EvaluatePreparedModel(device, preparedModel, testModel, testConfig, &baseSkipped);
diff --git a/neuralnetworks/1.3/vts/functional/QualityOfServiceTests.cpp b/neuralnetworks/1.3/vts/functional/QualityOfServiceTests.cpp
index 879989e..2ef1e8f 100644
--- a/neuralnetworks/1.3/vts/functional/QualityOfServiceTests.cpp
+++ b/neuralnetworks/1.3/vts/functional/QualityOfServiceTests.cpp
@@ -214,7 +214,8 @@
 }
 
 void runExecutionTest(const sp<IPreparedModel>& preparedModel, const TestModel& testModel,
-                      const Request& request, bool synchronous, DeadlineBoundType deadlineBound) {
+                      const Request& request, const ExecutionContext& context, bool synchronous,
+                      DeadlineBoundType deadlineBound) {
     const ExecutionFunction execute = synchronous ? executeSynchronously : executeAsynchronously;
     const auto deadline = makeDeadline(deadlineBound);
 
@@ -261,7 +262,7 @@
     // Retrieve execution results.
     ASSERT_TRUE(nn::compliantWithV1_0(request));
     const V1_0::Request request10 = nn::convertToV1_0(request);
-    const std::vector<TestBuffer> outputs = getOutputBuffers(request10);
+    const std::vector<TestBuffer> outputs = context.getOutputBuffers(request10);
 
     // We want "close-enough" results.
     if (status == ErrorStatus::NONE) {
@@ -270,10 +271,11 @@
 }
 
 void runExecutionTests(const sp<IPreparedModel>& preparedModel, const TestModel& testModel,
-                       const Request& request) {
+                       const Request& request, const ExecutionContext& context) {
     for (bool synchronous : {false, true}) {
         for (auto deadlineBound : deadlineBounds) {
-            runExecutionTest(preparedModel, testModel, request, synchronous, deadlineBound);
+            runExecutionTest(preparedModel, testModel, request, context, synchronous,
+                             deadlineBound);
         }
     }
 }
@@ -291,8 +293,9 @@
     if (preparedModel == nullptr) return;
 
     // run execution tests
-    const Request request = nn::convertToV1_3(createRequest(testModel));
-    runExecutionTests(preparedModel, testModel, request);
+    ExecutionContext context;
+    const Request request = nn::convertToV1_3(context.createRequest(testModel));
+    runExecutionTests(preparedModel, testModel, request, context);
 }
 
 class DeadlineTest : public GeneratedTestBase {};
diff --git a/neuralnetworks/1.3/vts/functional/VtsHalNeuralnetworks.cpp b/neuralnetworks/1.3/vts/functional/VtsHalNeuralnetworks.cpp
index 5b07034..703a2f5 100644
--- a/neuralnetworks/1.3/vts/functional/VtsHalNeuralnetworks.cpp
+++ b/neuralnetworks/1.3/vts/functional/VtsHalNeuralnetworks.cpp
@@ -177,7 +177,8 @@
 
 TEST_P(ValidationTest, Test) {
     const Model model = createModel(kTestModel);
-    const Request request = nn::convertToV1_3(createRequest(kTestModel));
+    ExecutionContext context;
+    const Request request = nn::convertToV1_3(context.createRequest(kTestModel));
     if (kTestModel.expectFailure) {
         validateFailure(kDevice, model, request);
     } else {