Merge changes from topic "aosp-nnapi-reusable-execution-canonical" am: 43ae2ecfe4

Original change: https://android-review.googlesource.com/c/platform/hardware/interfaces/+/1701932

Change-Id: If503fe78b3281a7a6feffdceacff45fb1b0c93bb

neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Burst.h

@@ -48,6 +48,10 @@ class Burst final : public nn::IBurst {
             const nn::OptionalTimePoint& deadline,
             const nn::OptionalDuration& loopTimeoutDuration) const override;
 
+    nn::GeneralResult<nn::SharedExecution> createReusableExecution(
+            const nn::Request& request, nn::MeasureTiming measure,
+            const nn::OptionalDuration& loopTimeoutDuration) const override;
+
   private:
     const nn::SharedPreparedModel kPreparedModel;
 };

neuralnetworks/1.0/utils/src/Burst.cpp

@@ -55,4 +55,10 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::
     return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration);
 }
 
+nn::GeneralResult<nn::SharedExecution> Burst::createReusableExecution(
+        const nn::Request& request, nn::MeasureTiming measure,
+        const nn::OptionalDuration& loopTimeoutDuration) const {
+    return kPreparedModel->createReusableExecution(request, measure, loopTimeoutDuration);
+}
+
 }  // namespace android::hardware::neuralnetworks::V1_0::utils

neuralnetworks/1.2/utils/include/nnapi/hal/1.2/ExecutionBurstController.h

@@ -28,9 +28,11 @@
 #include <fmq/MessageQueue.h>
 #include <hidl/MQDescriptor.h>
 #include <nnapi/IBurst.h>
+#include <nnapi/IExecution.h>
 #include <nnapi/IPreparedModel.h>
 #include <nnapi/Result.h>
 #include <nnapi/Types.h>
+#include <nnapi/hal/CommonUtils.h>
 #include <nnapi/hal/ProtectCallback.h>
 
 #include <atomic>
@@ -51,14 +53,14 @@ namespace android::hardware::neuralnetworks::V1_2::utils {
  * across FMQ, making it appear to the runtime as a regular synchronous inference. Additionally,
  * this class manages the burst's memory cache.
  */
-class ExecutionBurstController final : public nn::IBurst {
+class ExecutionBurstController final
+    : public nn::IBurst,
+      public std::enable_shared_from_this<ExecutionBurstController> {
     struct PrivateConstructorTag {};
 
   public:
-    using FallbackFunction =
-            std::function<nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>(
-                    const nn::Request&, nn::MeasureTiming, const nn::OptionalTimePoint&,
-                    const nn::OptionalDuration&)>;
+    using FallbackFunction = std::function<
+            nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>()>;
 
     /**
      * NN runtime memory cache.
@@ -154,10 +156,10 @@ class ExecutionBurstController final : public nn::IBurst {
      * @return ExecutionBurstController Execution burst controller object.
      */
     static nn::GeneralResult<std::shared_ptr<const ExecutionBurstController>> create(
-            const sp<IPreparedModel>& preparedModel, FallbackFunction fallback,
+            nn::SharedPreparedModel preparedModel, const sp<IPreparedModel>& hidlPreparedModel,
             std::chrono::microseconds pollingTimeWindow);
 
-    ExecutionBurstController(PrivateConstructorTag tag, FallbackFunction fallback,
+    ExecutionBurstController(PrivateConstructorTag tag, nn::SharedPreparedModel preparedModel,
                              std::unique_ptr<RequestChannelSender> requestChannelSender,
                              std::unique_ptr<ResultChannelReceiver> resultChannelReceiver,
                              sp<ExecutionBurstCallback> callback, sp<IBurstContext> burstContext,
@@ -173,9 +175,21 @@ class ExecutionBurstController final : public nn::IBurst {
             const nn::OptionalTimePoint& deadline,
             const nn::OptionalDuration& loopTimeoutDuration) const override;
 
+    // See IBurst::createReusableExecution for information on this method.
+    nn::GeneralResult<nn::SharedExecution> createReusableExecution(
+            const nn::Request& request, nn::MeasureTiming measure,
+            const nn::OptionalDuration& loopTimeoutDuration) const override;
+
+    // If fallback is not nullptr, this method will invoke the fallback function to try another
+    // execution path if the packet could not be sent. Otherwise, failing to send the packet will
+    // result in an error.
+    nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> executeInternal(
+            const std::vector<FmqRequestDatum>& requestPacket,
+            const hal::utils::RequestRelocation& relocation, FallbackFunction fallback) const;
+
   private:
     mutable std::atomic_flag mExecutionInFlight = ATOMIC_FLAG_INIT;
-    const FallbackFunction kFallback;
+    const nn::SharedPreparedModel kPreparedModel;
     const std::unique_ptr<RequestChannelSender> mRequestChannelSender;
     const std::unique_ptr<ResultChannelReceiver> mResultChannelReceiver;
     const sp<ExecutionBurstCallback> mBurstCallback;

neuralnetworks/1.2/utils/src/ExecutionBurstController.cpp

@@ -28,6 +28,7 @@
 #include <nnapi/Types.h>
 #include <nnapi/Validation.h>
 #include <nnapi/hal/1.0/Conversions.h>
+#include <nnapi/hal/CommonUtils.h>
 #include <nnapi/hal/HandleError.h>
 #include <nnapi/hal/ProtectCallback.h>
 #include <nnapi/hal/TransferValue.h>
@@ -50,6 +51,35 @@
 namespace android::hardware::neuralnetworks::V1_2::utils {
 namespace {
 
+class BurstExecution final : public nn::IExecution,
+                             public std::enable_shared_from_this<BurstExecution> {
+    struct PrivateConstructorTag {};
+
+  public:
+    static nn::GeneralResult<std::shared_ptr<const BurstExecution>> create(
+            std::shared_ptr<const ExecutionBurstController> controller,
+            std::vector<FmqRequestDatum> request, hal::utils::RequestRelocation relocation,
+            std::vector<ExecutionBurstController::OptionalCacheHold> cacheHolds);
+
+    BurstExecution(PrivateConstructorTag tag,
+                   std::shared_ptr<const ExecutionBurstController> controller,
+                   std::vector<FmqRequestDatum> request, hal::utils::RequestRelocation relocation,
+                   std::vector<ExecutionBurstController::OptionalCacheHold> cacheHolds);
+
+    nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> compute(
+            const nn::OptionalTimePoint& deadline) const override;
+
+    nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> computeFenced(
+            const std::vector<nn::SyncFence>& waitFor, const nn::OptionalTimePoint& deadline,
+            const nn::OptionalDuration& timeoutDurationAfterFence) const override;
+
+  private:
+    const std::shared_ptr<const ExecutionBurstController> kController;
+    const std::vector<FmqRequestDatum> kRequest;
+    const hal::utils::RequestRelocation kRelocation;
+    const std::vector<ExecutionBurstController::OptionalCacheHold> kCacheHolds;
+};
+
 nn::GeneralResult<sp<IBurstContext>> executionBurstResultCallback(
         V1_0::ErrorStatus status, const sp<IBurstContext>& burstContext) {
     HANDLE_HAL_STATUS(status) << "IPreparedModel::configureExecutionBurst failed with status "
@@ -209,10 +239,10 @@ Return<void> ExecutionBurstController::ExecutionBurstCallback::getMemories(
 // ExecutionBurstController methods
 
 nn::GeneralResult<std::shared_ptr<const ExecutionBurstController>> ExecutionBurstController::create(
-        const sp<V1_2::IPreparedModel>& preparedModel, FallbackFunction fallback,
+        nn::SharedPreparedModel preparedModel, const sp<V1_2::IPreparedModel>& hidlPreparedModel,
         std::chrono::microseconds pollingTimeWindow) {
     // check inputs
-    if (preparedModel == nullptr) {
+    if (preparedModel == nullptr || hidlPreparedModel == nullptr) {
         return NN_ERROR() << "ExecutionBurstController::create passed a nullptr";
     }
 
@@ -236,7 +266,7 @@ nn::GeneralResult<std::shared_ptr<const ExecutionBurstController>> ExecutionBurs
     auto cb = hal::utils::CallbackValue(executionBurstResultCallback);
 
     // configure burst
-    const Return<void> ret = preparedModel->configureExecutionBurst(
+    const Return<void> ret = hidlPreparedModel->configureExecutionBurst(
             burstCallback, *requestChannelDescriptor, *resultChannelDescriptor, cb);
     HANDLE_TRANSPORT_FAILURE(ret);
 
@@ -250,18 +280,18 @@ nn::GeneralResult<std::shared_ptr<const ExecutionBurstController>> ExecutionBurs
 
     // make and return controller
     return std::make_shared<const ExecutionBurstController>(
-            PrivateConstructorTag{}, std::move(fallback), std::move(requestChannelSender),
+            PrivateConstructorTag{}, std::move(preparedModel), std::move(requestChannelSender),
             std::move(resultChannelReceiver), std::move(burstCallback), std::move(burstContext),
             std::move(memoryCache), std::move(deathHandler));
 }
 
 ExecutionBurstController::ExecutionBurstController(
-        PrivateConstructorTag /*tag*/, FallbackFunction fallback,
+        PrivateConstructorTag /*tag*/, nn::SharedPreparedModel preparedModel,
         std::unique_ptr<RequestChannelSender> requestChannelSender,
         std::unique_ptr<ResultChannelReceiver> resultChannelReceiver,
         sp<ExecutionBurstCallback> callback, sp<IBurstContext> burstContext,
         std::shared_ptr<MemoryCache> memoryCache, neuralnetworks::utils::DeathHandler deathHandler)
-    : kFallback(std::move(fallback)),
+    : kPreparedModel(std::move(preparedModel)),
       mRequestChannelSender(std::move(requestChannelSender)),
       mResultChannelReceiver(std::move(resultChannelReceiver)),
       mBurstCallback(std::move(callback)),
@@ -283,26 +313,96 @@ ExecutionBurstController::execute(const nn::Request& request, nn::MeasureTiming
     // systraces. Note that the first point we can begin collecting systraces in
     // ExecutionBurstServer is when the RequestChannelReceiver realizes there is data in the FMQ, so
     // ExecutionBurstServer collects systraces at different points in the code.
-    NNTRACE_FULL(NNTRACE_LAYER_IPC, NNTRACE_PHASE_EXECUTION, "ExecutionBurstController::execute");
+    NNTRACE_RT(NNTRACE_PHASE_EXECUTION, "ExecutionBurstController::execute");
 
     // if the request is valid but of a higher version than what's supported in burst execution,
     // fall back to another execution path
     if (const auto version = NN_TRY(hal::utils::makeExecutionFailure(nn::validate(request)));
         version > nn::Version::ANDROID_Q) {
         // fallback to another execution path if the packet could not be sent
-        if (kFallback) {
-            return kFallback(request, measure, deadline, loopTimeoutDuration);
-        }
-        return NN_ERROR() << "Request object has features not supported by IBurst::execute";
+        return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration);
     }
 
+    // ensure that request is ready for IPC
+    std::optional<nn::Request> maybeRequestInShared;
+    hal::utils::RequestRelocation relocation;
+    const nn::Request& requestInShared =
+            NN_TRY(hal::utils::makeExecutionFailure(hal::utils::convertRequestFromPointerToShared(
+                    &request, &maybeRequestInShared, &relocation)));
+
     // clear pools field of request, as they will be provided via slots
-    const auto requestWithoutPools =
-            nn::Request{.inputs = request.inputs, .outputs = request.outputs, .pools = {}};
+    const auto requestWithoutPools = nn::Request{
+            .inputs = requestInShared.inputs, .outputs = requestInShared.outputs, .pools = {}};
     auto hidlRequest = NN_TRY(
             hal::utils::makeExecutionFailure(V1_0::utils::unvalidatedConvert(requestWithoutPools)));
     const auto hidlMeasure = NN_TRY(hal::utils::makeExecutionFailure(convert(measure)));
 
+    std::vector<int32_t> slots;
+    std::vector<OptionalCacheHold> holds;
+    slots.reserve(requestInShared.pools.size());
+    holds.reserve(requestInShared.pools.size());
+    for (const auto& memoryPool : requestInShared.pools) {
+        auto [slot, hold] = mMemoryCache->cacheMemory(std::get<nn::SharedMemory>(memoryPool));
+        slots.push_back(slot);
+        holds.push_back(std::move(hold));
+    }
+
+    // send request packet
+    const auto requestPacket = serialize(hidlRequest, hidlMeasure, slots);
+    const auto fallback = [this, &request, measure, &deadline, &loopTimeoutDuration] {
+        return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration);
+    };
+    return executeInternal(requestPacket, relocation, fallback);
+}
+
+// See IBurst::createReusableExecution for information on this method.
+nn::GeneralResult<nn::SharedExecution> ExecutionBurstController::createReusableExecution(
+        const nn::Request& request, nn::MeasureTiming measure,
+        const nn::OptionalDuration& loopTimeoutDuration) const {
+    NNTRACE_RT(NNTRACE_PHASE_EXECUTION, "ExecutionBurstController::createReusableExecution");
+
+    // if the request is valid but of a higher version than what's supported in burst execution,
+    // fall back to another execution path
+    if (const auto version = NN_TRY(hal::utils::makeGeneralFailure(nn::validate(request)));
+        version > nn::Version::ANDROID_Q) {
+        // fallback to another execution path if the packet could not be sent
+        return kPreparedModel->createReusableExecution(request, measure, loopTimeoutDuration);
+    }
+
+    // ensure that request is ready for IPC
+    std::optional<nn::Request> maybeRequestInShared;
+    hal::utils::RequestRelocation relocation;
+    const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
+            &request, &maybeRequestInShared, &relocation));
+
+    // clear pools field of request, as they will be provided via slots
+    const auto requestWithoutPools = nn::Request{
+            .inputs = requestInShared.inputs, .outputs = requestInShared.outputs, .pools = {}};
+    auto hidlRequest = NN_TRY(V1_0::utils::unvalidatedConvert(requestWithoutPools));
+    const auto hidlMeasure = NN_TRY(convert(measure));
+
+    std::vector<int32_t> slots;
+    std::vector<OptionalCacheHold> holds;
+    slots.reserve(requestInShared.pools.size());
+    holds.reserve(requestInShared.pools.size());
+    for (const auto& memoryPool : requestInShared.pools) {
+        auto [slot, hold] = mMemoryCache->cacheMemory(std::get<nn::SharedMemory>(memoryPool));
+        slots.push_back(slot);
+        holds.push_back(std::move(hold));
+    }
+
+    const auto requestPacket = serialize(hidlRequest, hidlMeasure, slots);
+    return BurstExecution::create(shared_from_this(), std::move(requestPacket),
+                                  std::move(relocation), std::move(holds));
+}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>
+ExecutionBurstController::executeInternal(const std::vector<FmqRequestDatum>& requestPacket,
+                                          const hal::utils::RequestRelocation& relocation,
+                                          FallbackFunction fallback) const {
+    NNTRACE_FULL(NNTRACE_LAYER_IPC, NNTRACE_PHASE_EXECUTION,
+                 "ExecutionBurstController::executeInternal");
+
     // Ensure that at most one execution is in flight at any given time.
     const bool alreadyInFlight = mExecutionInFlight.test_and_set();
     if (alreadyInFlight) {
@@ -310,22 +410,16 @@ ExecutionBurstController::execute(const nn::Request& request, nn::MeasureTiming
     }
     const auto guard = base::make_scope_guard([this] { mExecutionInFlight.clear(); });
 
-    std::vector<int32_t> slots;
-    std::vector<OptionalCacheHold> holds;
-    slots.reserve(request.pools.size());
-    holds.reserve(request.pools.size());
-    for (const auto& memoryPool : request.pools) {
-        auto [slot, hold] = mMemoryCache->cacheMemory(std::get<nn::SharedMemory>(memoryPool));
-        slots.push_back(slot);
-        holds.push_back(std::move(hold));
+    if (relocation.input) {
+        relocation.input->flush();
     }
 
     // send request packet
-    const auto sendStatus = mRequestChannelSender->send(hidlRequest, hidlMeasure, slots);
+    const auto sendStatus = mRequestChannelSender->sendPacket(requestPacket);
     if (!sendStatus.ok()) {
         // fallback to another execution path if the packet could not be sent
-        if (kFallback) {
-            return kFallback(request, measure, deadline, loopTimeoutDuration);
+        if (fallback) {
+            return fallback();
         }
         return NN_ERROR() << "Error sending FMQ packet: " << sendStatus.error();
     }
@@ -333,7 +427,47 @@ ExecutionBurstController::execute(const nn::Request& request, nn::MeasureTiming
     // get result packet
     const auto [status, outputShapes, timing] =
             NN_TRY(hal::utils::makeExecutionFailure(mResultChannelReceiver->getBlocking()));
+
+    if (relocation.output) {
+        relocation.output->flush();
+    }
     return executionCallback(status, outputShapes, timing);
 }
 
+nn::GeneralResult<std::shared_ptr<const BurstExecution>> BurstExecution::create(
+        std::shared_ptr<const ExecutionBurstController> controller,
+        std::vector<FmqRequestDatum> request, hal::utils::RequestRelocation relocation,
+        std::vector<ExecutionBurstController::OptionalCacheHold> cacheHolds) {
+    if (controller == nullptr) {
+        return NN_ERROR() << "V1_2::utils::BurstExecution::create must have non-null controller";
+    }
+
+    return std::make_shared<const BurstExecution>(PrivateConstructorTag{}, std::move(controller),
+                                                  std::move(request), std::move(relocation),
+                                                  std::move(cacheHolds));
+}
+
+BurstExecution::BurstExecution(PrivateConstructorTag /*tag*/,
+                               std::shared_ptr<const ExecutionBurstController> controller,
+                               std::vector<FmqRequestDatum> request,
+                               hal::utils::RequestRelocation relocation,
+                               std::vector<ExecutionBurstController::OptionalCacheHold> cacheHolds)
+    : kController(std::move(controller)),
+      kRequest(std::move(request)),
+      kRelocation(std::move(relocation)),
+      kCacheHolds(std::move(cacheHolds)) {}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> BurstExecution::compute(
+        const nn::OptionalTimePoint& /*deadline*/) const {
+    return kController->executeInternal(kRequest, kRelocation, /*fallback=*/nullptr);
+}
+
+nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
+BurstExecution::computeFenced(const std::vector<nn::SyncFence>& /*waitFor*/,
+                              const nn::OptionalTimePoint& /*deadline*/,
+                              const nn::OptionalDuration& /*timeoutDurationAfterFence*/) const {
+    return NN_ERROR(nn::ErrorStatus::GENERAL_FAILURE)
+           << "IExecution::computeFenced is not supported on burst object";
+}
+
 }  // namespace android::hardware::neuralnetworks::V1_2::utils

neuralnetworks/1.2/utils/src/PreparedModel.cpp

@@ -158,7 +158,7 @@ nn::GeneralResult<nn::SharedBurst> PreparedModel::configureExecutionBurst() cons
         return preparedModel->execute(request, measure, deadline, loopTimeoutDuration);
     };
     const auto pollingTimeWindow = getBurstControllerPollingTimeWindow();
-    return ExecutionBurstController::create(kPreparedModel, std::move(fallback), pollingTimeWindow);
+    return ExecutionBurstController::create(shared_from_this(), kPreparedModel, pollingTimeWindow);
 }
 
 std::any PreparedModel::getUnderlyingResource() const {

neuralnetworks/1.3/utils/src/PreparedModel.cpp

@@ -255,7 +255,7 @@ nn::GeneralResult<nn::SharedBurst> PreparedModel::configureExecutionBurst() cons
         return preparedModel->execute(request, measure, deadline, loopTimeoutDuration);
     };
     const auto pollingTimeWindow = V1_2::utils::getBurstControllerPollingTimeWindow();
-    return V1_2::utils::ExecutionBurstController::create(kPreparedModel, std::move(fallback),
+    return V1_2::utils::ExecutionBurstController::create(shared_from_this(), kPreparedModel,
                                                          pollingTimeWindow);
 }
 

neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Burst.h

@@ -38,7 +38,7 @@
 namespace aidl::android::hardware::neuralnetworks::utils {
 
 // Class that adapts aidl_hal::IBurst to nn::IBurst.
-class Burst final : public nn::IBurst {
+class Burst final : public nn::IBurst, public std::enable_shared_from_this<Burst> {
     struct PrivateConstructorTag {};
 
   public:
@@ -100,6 +100,16 @@ class Burst final : public nn::IBurst {
             const nn::OptionalTimePoint& deadline,
             const nn::OptionalDuration& loopTimeoutDuration) const override;
 
+    // See IBurst::createReusableExecution for information.
+    nn::GeneralResult<nn::SharedExecution> createReusableExecution(
+            const nn::Request& request, nn::MeasureTiming measure,
+            const nn::OptionalDuration& loopTimeoutDuration) const override;
+
+    nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> executeInternal(
+            const aidl_hal::Request& request, const std::vector<int64_t>& memoryIdentifierTokens,
+            bool measure, int64_t deadline, int64_t loopTimeoutDuration,
+            const hal::utils::RequestRelocation& relocation) const;
+
   private:
     mutable std::atomic_flag mExecutionInFlight = ATOMIC_FLAG_INIT;
     const std::shared_ptr<aidl_hal::IBurst> kBurst;

neuralnetworks/aidl/utils/src/Burst.cpp

@@ -22,6 +22,7 @@
 #include <android-base/logging.h>
 #include <android/binder_auto_utils.h>
 #include <nnapi/IBurst.h>
+#include <nnapi/IExecution.h>
 #include <nnapi/Result.h>
 #include <nnapi/TypeUtils.h>
 #include <nnapi/Types.h>
@@ -35,6 +36,39 @@
 namespace aidl::android::hardware::neuralnetworks::utils {
 namespace {
 
+class BurstExecution final : public nn::IExecution,
+                             public std::enable_shared_from_this<BurstExecution> {
+    struct PrivateConstructorTag {};
+
+  public:
+    static nn::GeneralResult<std::shared_ptr<const BurstExecution>> create(
+            std::shared_ptr<const Burst> burst, Request request,
+            std::vector<int64_t> memoryIdentifierTokens, bool measure, int64_t loopTimeoutDuration,
+            hal::utils::RequestRelocation relocation,
+            std::vector<Burst::OptionalCacheHold> cacheHolds);
+
+    BurstExecution(PrivateConstructorTag tag, std::shared_ptr<const Burst> burst, Request request,
+                   std::vector<int64_t> memoryIdentifierTokens, bool measure,
+                   int64_t loopTimeoutDuration, hal::utils::RequestRelocation relocation,
+                   std::vector<Burst::OptionalCacheHold> cacheHolds);
+
+    nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> compute(
+            const nn::OptionalTimePoint& deadline) const override;
+
+    nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> computeFenced(
+            const std::vector<nn::SyncFence>& waitFor, const nn::OptionalTimePoint& deadline,
+            const nn::OptionalDuration& timeoutDurationAfterFence) const override;
+
+  private:
+    const std::shared_ptr<const Burst> kBurst;
+    const Request kRequest;
+    const std::vector<int64_t>& kMemoryIdentifierTokens;
+    const bool kMeasure;
+    const int64_t kLoopTimeoutDuration;
+    const hal::utils::RequestRelocation kRelocation;
+    const std::vector<Burst::OptionalCacheHold> kCacheHolds;
+};
+
 nn::GeneralResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> convertExecutionResults(
         const std::vector<OutputShape>& outputShapes, const Timing& timing) {
     return std::make_pair(NN_TRY(nn::convert(outputShapes)), NN_TRY(nn::convert(timing)));
@@ -139,13 +173,6 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::
         const nn::Request& request, nn::MeasureTiming measure,
         const nn::OptionalTimePoint& deadline,
         const nn::OptionalDuration& loopTimeoutDuration) const {
-    // Ensure that at most one execution is in flight at any given time.
-    const bool alreadyInFlight = mExecutionInFlight.test_and_set();
-    if (alreadyInFlight) {
-        return NN_ERROR() << "IBurst already has an execution in flight";
-    }
-    const auto guard = ::android::base::make_scope_guard([this] { mExecutionInFlight.clear(); });
-
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
@@ -161,9 +188,9 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::
 
     std::vector<int64_t> memoryIdentifierTokens;
     std::vector<OptionalCacheHold> holds;
-    memoryIdentifierTokens.reserve(request.pools.size());
-    holds.reserve(request.pools.size());
-    for (const auto& memoryPool : request.pools) {
+    memoryIdentifierTokens.reserve(requestInShared.pools.size());
+    holds.reserve(requestInShared.pools.size());
+    for (const auto& memoryPool : requestInShared.pools) {
         if (const auto* memory = std::get_if<nn::SharedMemory>(&memoryPool)) {
             if (auto cached = kMemoryCache->getMemoryIfAvailable(*memory)) {
                 auto& [identifier, hold] = *cached;
@@ -174,16 +201,30 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::
         }
         memoryIdentifierTokens.push_back(-1);
     }
-    CHECK_EQ(request.pools.size(), memoryIdentifierTokens.size());
+    CHECK_EQ(requestInShared.pools.size(), memoryIdentifierTokens.size());
+
+    return executeInternal(aidlRequest, memoryIdentifierTokens, aidlMeasure, aidlDeadline,
+                           aidlLoopTimeoutDuration, relocation);
+}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::executeInternal(
+        const Request& request, const std::vector<int64_t>& memoryIdentifierTokens, bool measure,
+        int64_t deadline, int64_t loopTimeoutDuration,
+        const hal::utils::RequestRelocation& relocation) const {
+    // Ensure that at most one execution is in flight at any given time.
+    const bool alreadyInFlight = mExecutionInFlight.test_and_set();
+    if (alreadyInFlight) {
+        return NN_ERROR() << "IBurst already has an execution in flight";
+    }
+    const auto guard = ::android::base::make_scope_guard([this] { mExecutionInFlight.clear(); });
 
     if (relocation.input) {
         relocation.input->flush();
     }
 
     ExecutionResult executionResult;
-    const auto ret =
-            kBurst->executeSynchronously(aidlRequest, memoryIdentifierTokens, aidlMeasure,
-                                         aidlDeadline, aidlLoopTimeoutDuration, &executionResult);
+    const auto ret = kBurst->executeSynchronously(request, memoryIdentifierTokens, measure,
+                                                  deadline, loopTimeoutDuration, &executionResult);
     HANDLE_ASTATUS(ret) << "execute failed";
     if (!executionResult.outputSufficientSize) {
         auto canonicalOutputShapes =
@@ -200,4 +241,82 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::
     return std::make_pair(std::move(outputShapes), timing);
 }
 
+nn::GeneralResult<nn::SharedExecution> Burst::createReusableExecution(
+        const nn::Request& request, nn::MeasureTiming measure,
+        const nn::OptionalDuration& loopTimeoutDuration) const {
+    // Ensure that request is ready for IPC.
+    std::optional<nn::Request> maybeRequestInShared;
+    hal::utils::RequestRelocation relocation;
+    const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
+            &request, &maybeRequestInShared, &relocation));
+
+    auto aidlRequest = NN_TRY(convert(requestInShared));
+    const auto aidlMeasure = NN_TRY(convert(measure));
+    const auto aidlLoopTimeoutDuration = NN_TRY(convert(loopTimeoutDuration));
+
+    std::vector<int64_t> memoryIdentifierTokens;
+    std::vector<OptionalCacheHold> holds;
+    memoryIdentifierTokens.reserve(requestInShared.pools.size());
+    holds.reserve(requestInShared.pools.size());
+    for (const auto& memoryPool : requestInShared.pools) {
+        if (const auto* memory = std::get_if<nn::SharedMemory>(&memoryPool)) {
+            if (auto cached = kMemoryCache->getMemoryIfAvailable(*memory)) {
+                auto& [identifier, hold] = *cached;
+                memoryIdentifierTokens.push_back(identifier);
+                holds.push_back(std::move(hold));
+                continue;
+            }
+        }
+        memoryIdentifierTokens.push_back(-1);
+    }
+    CHECK_EQ(requestInShared.pools.size(), memoryIdentifierTokens.size());
+
+    return BurstExecution::create(shared_from_this(), std::move(aidlRequest),
+                                  std::move(memoryIdentifierTokens), aidlMeasure,
+                                  aidlLoopTimeoutDuration, std::move(relocation), std::move(holds));
+}
+
+nn::GeneralResult<std::shared_ptr<const BurstExecution>> BurstExecution::create(
+        std::shared_ptr<const Burst> burst, Request request,
+        std::vector<int64_t> memoryIdentifierTokens, bool measure, int64_t loopTimeoutDuration,
+        hal::utils::RequestRelocation relocation,
+        std::vector<Burst::OptionalCacheHold> cacheHolds) {
+    if (burst == nullptr) {
+        return NN_ERROR() << "aidl::utils::BurstExecution::create must have non-null burst";
+    }
+
+    return std::make_shared<const BurstExecution>(
+            PrivateConstructorTag{}, std::move(burst), std::move(request),
+            std::move(memoryIdentifierTokens), measure, loopTimeoutDuration, std::move(relocation),
+            std::move(cacheHolds));
+}
+
+BurstExecution::BurstExecution(PrivateConstructorTag /*tag*/, std::shared_ptr<const Burst> burst,
+                               Request request, std::vector<int64_t> memoryIdentifierTokens,
+                               bool measure, int64_t loopTimeoutDuration,
+                               hal::utils::RequestRelocation relocation,
+                               std::vector<Burst::OptionalCacheHold> cacheHolds)
+    : kBurst(std::move(burst)),
+      kRequest(std::move(request)),
+      kMemoryIdentifierTokens(std::move(memoryIdentifierTokens)),
+      kMeasure(measure),
+      kLoopTimeoutDuration(loopTimeoutDuration),
+      kRelocation(std::move(relocation)),
+      kCacheHolds(std::move(cacheHolds)) {}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> BurstExecution::compute(
+        const nn::OptionalTimePoint& deadline) const {
+    const auto aidlDeadline = NN_TRY(hal::utils::makeExecutionFailure(convert(deadline)));
+    return kBurst->executeInternal(kRequest, kMemoryIdentifierTokens, kMeasure, aidlDeadline,
+                                   kLoopTimeoutDuration, kRelocation);
+}
+
+nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
+BurstExecution::computeFenced(const std::vector<nn::SyncFence>& /*waitFor*/,
+                              const nn::OptionalTimePoint& /*deadline*/,
+                              const nn::OptionalDuration& /*timeoutDurationAfterFence*/) const {
+    return NN_ERROR(nn::ErrorStatus::GENERAL_FAILURE)
+           << "IExecution::computeFenced is not supported on burst object";
+}
+
 }  // namespace aidl::android::hardware::neuralnetworks::utils

neuralnetworks/utils/common/include/nnapi/hal/InvalidBurst.h

@@ -35,6 +35,10 @@ class InvalidBurst final : public nn::IBurst {
             const nn::Request& request, nn::MeasureTiming measure,
             const nn::OptionalTimePoint& deadline,
             const nn::OptionalDuration& loopTimeoutDuration) const override;
+
+    nn::GeneralResult<nn::SharedExecution> createReusableExecution(
+            const nn::Request& request, nn::MeasureTiming measure,
+            const nn::OptionalDuration& loopTimeoutDuration) const override;
 };
 
 }  // namespace android::hardware::neuralnetworks::utils

neuralnetworks/utils/common/include/nnapi/hal/ResilientBurst.h

@@ -51,7 +51,16 @@ class ResilientBurst final : public nn::IBurst,
             const nn::OptionalTimePoint& deadline,
             const nn::OptionalDuration& loopTimeoutDuration) const override;
 
+    nn::GeneralResult<nn::SharedExecution> createReusableExecution(
+            const nn::Request& request, nn::MeasureTiming measure,
+            const nn::OptionalDuration& loopTimeoutDuration) const override;
+
   private:
+    bool isValidInternal() const EXCLUDES(mMutex);
+    nn::GeneralResult<nn::SharedExecution> createReusableExecutionInternal(
+            const nn::Request& request, nn::MeasureTiming measure,
+            const nn::OptionalDuration& loopTimeoutDuration) const;
+
     const Factory kMakeBurst;
     mutable std::mutex mMutex;
     mutable nn::SharedBurst mBurst GUARDED_BY(mMutex);

neuralnetworks/utils/common/src/InvalidBurst.cpp

@@ -38,4 +38,10 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Invalid
     return NN_ERROR() << "InvalidBurst";
 }
 
+nn::GeneralResult<nn::SharedExecution> InvalidBurst::createReusableExecution(
+        const nn::Request& /*request*/, nn::MeasureTiming /*measure*/,
+        const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
+    return NN_ERROR() << "InvalidBurst";
+}
+
 }  // namespace android::hardware::neuralnetworks::utils

neuralnetworks/utils/common/src/ResilientBurst.cpp

@@ -19,6 +19,7 @@
 #include <android-base/logging.h>
 #include <android-base/thread_annotations.h>
 #include <nnapi/IBurst.h>
+#include <nnapi/IPreparedModel.h>
 #include <nnapi/Result.h>
 #include <nnapi/TypeUtils.h>
 #include <nnapi/Types.h>
@@ -29,6 +30,9 @@
 #include <optional>
 #include <utility>
 
+#include "InvalidExecution.h"
+#include "ResilientExecution.h"
+
 namespace android::hardware::neuralnetworks::utils {
 namespace {
 
@@ -46,11 +50,11 @@ auto protect(const ResilientBurst& resilientBurst, const FnType& fn)
     // Attempt recovery and return if it fails.
     auto maybeBurst = resilientBurst.recover(burst.get());
     if (!maybeBurst.has_value()) {
-        auto [resultErrorMessage, resultErrorCode, resultOutputShapes] = std::move(result).error();
-        const auto& [recoveryErrorMessage, recoveryErrorCode] = maybeBurst.error();
-        return nn::error(resultErrorCode, std::move(resultOutputShapes))
-               << resultErrorMessage << ", and failed to recover dead burst object with error "
-               << recoveryErrorCode << ": " << recoveryErrorMessage;
+        const auto& [message, code] = maybeBurst.error();
+        std::ostringstream oss;
+        oss << ", and failed to recover dead burst object with error " << code << ": " << message;
+        result.error().message += oss.str();
+        return result;
     }
     burst = std::move(maybeBurst).value();
 
@@ -109,4 +113,35 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Resilie
     return protect(*this, fn);
 }
 
+nn::GeneralResult<nn::SharedExecution> ResilientBurst::createReusableExecution(
+        const nn::Request& request, nn::MeasureTiming measure,
+        const nn::OptionalDuration& loopTimeoutDuration) const {
+#if 0
+    auto self = shared_from_this();
+    ResilientExecution::Factory makeExecution =
+            [burst = std::move(self), request, measure, loopTimeoutDuration] {
+        return burst->createReusableExecutionInternal(request, measure, loopTimeoutDuration);
+    };
+    return ResilientExecution::create(std::move(makeExecution));
+#else
+    return createReusableExecutionInternal(request, measure, loopTimeoutDuration);
+#endif
+}
+
+nn::GeneralResult<nn::SharedExecution> ResilientBurst::createReusableExecutionInternal(
+        const nn::Request& request, nn::MeasureTiming measure,
+        const nn::OptionalDuration& loopTimeoutDuration) const {
+    if (!isValidInternal()) {
+        return std::make_shared<const InvalidExecution>();
+    }
+    const auto fn = [&request, measure, &loopTimeoutDuration](const nn::IBurst& burst) {
+        return burst.createReusableExecution(request, measure, loopTimeoutDuration);
+    };
+    return protect(*this, fn);
+}
+
+bool ResilientBurst::isValidInternal() const {
+    return true;
+}
+
 }  // namespace android::hardware::neuralnetworks::utils