adapter/src/PreparedModel.cpp

/*
 * 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.
 */

#include "PreparedModel.h"

#include <ExecutionBurstServer.h>
#include <android-base/logging.h>
#include <android/hardware/neuralnetworks/1.0/IExecutionCallback.h>
#include <android/hardware/neuralnetworks/1.0/types.h>
#include <android/hardware/neuralnetworks/1.2/IBurstCallback.h>
#include <android/hardware/neuralnetworks/1.2/IExecutionCallback.h>
#include <android/hardware/neuralnetworks/1.2/types.h>
#include <android/hardware/neuralnetworks/1.3/IExecutionCallback.h>
#include <android/hardware/neuralnetworks/1.3/IFencedExecutionCallback.h>
#include <android/hardware/neuralnetworks/1.3/IPreparedModel.h>
#include <android/hardware/neuralnetworks/1.3/types.h>
#include <hwbinder/IPCThreadState.h>
#include <nnapi/IPreparedModel.h>
#include <nnapi/TypeUtils.h>
#include <nnapi/Types.h>
#include <nnapi/Validation.h>
#include <nnapi/hal/1.0/Utils.h>
#include <nnapi/hal/1.2/Utils.h>
#include <nnapi/hal/1.3/Conversions.h>
#include <nnapi/hal/1.3/Utils.h>
#include <nnapi/hal/HandleError.h>
#include <sys/types.h>

#include <memory>
#include <thread>

// See hardware/interfaces/neuralnetworks/utils/README.md for more information on HIDL interface
// lifetimes across processes and for protecting asynchronous calls across HIDL.

namespace android::hardware::neuralnetworks::adapter {
namespace {

template <typename Type>
auto convertInput(const Type& object) -> decltype(nn::convert(std::declval<Type>())) {
    auto result = nn::convert(object);
    if (!result.has_value()) {
        result.error().code = nn::ErrorStatus::INVALID_ARGUMENT;
    }
    return result;
}

class FencedExecutionCallback final : public V1_3::IFencedExecutionCallback {
  public:
    explicit FencedExecutionCallback(const nn::ExecuteFencedInfoCallback& callback)
        : kCallback(callback) {
        CHECK(callback != nullptr);
    }

    Return<void> getExecutionInfo(getExecutionInfo_cb cb) override {
        const auto result = kCallback();
        if (!result.has_value()) {
            const auto& [message, code] = result.error();
            const auto status =
                    V1_3::utils::convert(code).value_or(V1_3::ErrorStatus::GENERAL_FAILURE);
            LOG(ERROR) << message;
            cb(status, V1_2::utils::kNoTiming, V1_2::utils::kNoTiming);
            return Void();
        }
        const auto [timingLaunched, timingFenced] = result.value();
        const auto hidlTimingLaunched = V1_3::utils::convert(timingLaunched).value();
        const auto hidlTimingFenced = V1_3::utils::convert(timingFenced).value();
        cb(V1_3::ErrorStatus::NONE, hidlTimingLaunched, hidlTimingFenced);
        return Void();
    }

  private:
    const nn::ExecuteFencedInfoCallback kCallback;
};

using ExecutionResult = nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>;

void notify(V1_0::IExecutionCallback* callback, nn::ErrorStatus status,
            const std::vector<nn::OutputShape>& /*outputShapes*/, const nn::Timing& /*timing*/) {
    if (callback != nullptr) {
        const auto hidlStatus = V1_0::utils::convert(status).value();
        const auto ret = callback->notify(hidlStatus);
        if (!ret.isOk()) {
            LOG(ERROR) << "V1_0::IExecutionCallback::notify failed with " << ret.description();
        }
    }
}

void notify(V1_2::IExecutionCallback* callback, nn::ErrorStatus status,
            const std::vector<nn::OutputShape>& outputShapes, const nn::Timing& timing) {
    if (callback != nullptr) {
        const auto hidlStatus = V1_2::utils::convert(status).value();
        const auto hidlOutputShapes = V1_2::utils::convert(outputShapes).value();
        const auto hidlTiming = V1_2::utils::convert(timing).value();
        const auto ret = callback->notify_1_2(hidlStatus, hidlOutputShapes, hidlTiming);
        if (!ret.isOk()) {
            LOG(ERROR) << "V1_2::IExecutionCallback::notify_1_2 failed with " << ret.description();
        }
    }
}

void notify(V1_3::IExecutionCallback* callback, nn::ErrorStatus status,
            const std::vector<nn::OutputShape>& outputShapes, const nn::Timing& timing) {
    if (callback != nullptr) {
        const auto hidlStatus = V1_3::utils::convert(status).value();
        const auto hidlOutputShapes = V1_3::utils::convert(outputShapes).value();
        const auto hidlTiming = V1_3::utils::convert(timing).value();
        const auto ret = callback->notify_1_3(hidlStatus, hidlOutputShapes, hidlTiming);
        if (!ret.isOk()) {
            LOG(ERROR) << "V1_3::IExecutionCallback::notify_1_3 failed with " << ret.description();
        }
    }
}

template <typename CallbackType>
void notify(CallbackType* callback, ExecutionResult result) {
    if (!result.has_value()) {
        const auto [message, status, outputShapes] = std::move(result).error();
        LOG(ERROR) << message;
        notify(callback, status, outputShapes, {});
    } else {
        const auto [outputShapes, timing] = std::move(result).value();
        notify(callback, nn::ErrorStatus::NONE, outputShapes, timing);
    }
}

nn::GeneralResult<void> execute(const nn::SharedPreparedModel& preparedModel, uid_t userId,
                                const Executor& executor, const V1_0::Request& request,
                                const sp<V1_0::IExecutionCallback>& callback) {
    if (callback.get() == nullptr) {
        return NN_ERROR(nn::ErrorStatus::INVALID_ARGUMENT) << "Invalid callback";
    }

    auto nnRequest = NN_TRY(convertInput(request));

    const std::any resource = preparedModel->getUnderlyingResource();
    if (const auto* model = std::any_cast<const nn::Model*>(&resource)) {
        CHECK(*model != nullptr);
        NN_TRY(utils::makeGeneralFailure(nn::validateRequestForModel(nnRequest, **model),
                                         nn::ErrorStatus::INVALID_ARGUMENT));
    }

    Task task = [preparedModel, nnRequest = std::move(nnRequest), callback] {
        auto result = preparedModel->execute(nnRequest, nn::MeasureTiming::NO, {}, {});
        notify(callback.get(), std::move(result));
    };
    executor(std::move(task), userId, {});

    return {};
}

nn::GeneralResult<void> execute_1_2(const nn::SharedPreparedModel& preparedModel, uid_t userId,
                                    const Executor& executor, const V1_0::Request& request,
                                    V1_2::MeasureTiming measure,
                                    const sp<V1_2::IExecutionCallback>& callback) {
    if (callback.get() == nullptr) {
        return NN_ERROR(nn::ErrorStatus::INVALID_ARGUMENT) << "Invalid callback";
    }

    auto nnRequest = NN_TRY(convertInput(request));
    const auto nnMeasure = NN_TRY(convertInput(measure));

    const std::any resource = preparedModel->getUnderlyingResource();
    if (const auto* model = std::any_cast<const nn::Model*>(&resource)) {
        CHECK(*model != nullptr);
        NN_TRY(utils::makeGeneralFailure(nn::validateRequestForModel(nnRequest, **model),
                                         nn::ErrorStatus::INVALID_ARGUMENT));
    }

    Task task = [preparedModel, nnRequest = std::move(nnRequest), nnMeasure, callback] {
        auto result = preparedModel->execute(nnRequest, nnMeasure, {}, {});
        notify(callback.get(), std::move(result));
    };
    executor(std::move(task), userId, {});

    return {};
}

nn::GeneralResult<void> execute_1_3(const nn::SharedPreparedModel& preparedModel, uid_t userId,
                                    const Executor& executor, const V1_3::Request& request,
                                    V1_2::MeasureTiming measure,
                                    const V1_3::OptionalTimePoint& deadline,
                                    const V1_3::OptionalTimeoutDuration& loopTimeoutDuration,
                                    const sp<V1_3::IExecutionCallback>& callback) {
    if (callback.get() == nullptr) {
        return NN_ERROR(nn::ErrorStatus::INVALID_ARGUMENT) << "Invalid callback";
    }

    auto nnRequest = NN_TRY(convertInput(request));
    const auto nnMeasure = NN_TRY(convertInput(measure));
    const auto nnDeadline = NN_TRY(convertInput(deadline));
    const auto nnLoopTimeoutDuration = NN_TRY(convertInput(loopTimeoutDuration));

    const std::any resource = preparedModel->getUnderlyingResource();
    if (const auto* model = std::any_cast<const nn::Model*>(&resource)) {
        CHECK(*model != nullptr);
        NN_TRY(utils::makeGeneralFailure(nn::validateRequestForModel(nnRequest, **model),
                                         nn::ErrorStatus::INVALID_ARGUMENT));
    }

    Task task = [preparedModel, nnRequest = std::move(nnRequest), nnMeasure, nnDeadline,
                 nnLoopTimeoutDuration, callback] {
        auto result =
                preparedModel->execute(nnRequest, nnMeasure, nnDeadline, nnLoopTimeoutDuration);
        notify(callback.get(), std::move(result));
    };
    executor(std::move(task), userId, nnDeadline);

    return {};
}

nn::ExecutionResult<std::pair<hidl_vec<V1_2::OutputShape>, V1_2::Timing>> executeSynchronously(
        const nn::SharedPreparedModel& preparedModel, const V1_0::Request& request,
        V1_2::MeasureTiming measure) {
    const auto nnRequest = NN_TRY(utils::makeExecutionFailure(convertInput(request)));
    const auto nnMeasure = NN_TRY(utils::makeExecutionFailure(convertInput(measure)));

    const auto [outputShapes, timing] =
            NN_TRY(preparedModel->execute(nnRequest, nnMeasure, {}, {}));

    auto hidlOutputShapes = NN_TRY(utils::makeExecutionFailure(V1_2::utils::convert(outputShapes)));
    const auto hidlTiming = NN_TRY(utils::makeExecutionFailure(V1_2::utils::convert(timing)));
    return std::make_pair(std::move(hidlOutputShapes), hidlTiming);
}

nn::ExecutionResult<std::pair<hidl_vec<V1_2::OutputShape>, V1_2::Timing>> executeSynchronously_1_3(
        const nn::SharedPreparedModel& preparedModel, const V1_3::Request& request,
        V1_2::MeasureTiming measure, const V1_3::OptionalTimePoint& deadline,
        const V1_3::OptionalTimeoutDuration& loopTimeoutDuration) {
    const auto nnRequest = NN_TRY(utils::makeExecutionFailure(convertInput(request)));
    const auto nnMeasure = NN_TRY(utils::makeExecutionFailure(convertInput(measure)));
    const auto nnDeadline = NN_TRY(utils::makeExecutionFailure(convertInput(deadline)));
    const auto nnLoopTimeoutDuration =
            NN_TRY(utils::makeExecutionFailure(convertInput(loopTimeoutDuration)));

    const auto [outputShapes, timing] =
            NN_TRY(preparedModel->execute(nnRequest, nnMeasure, nnDeadline, nnLoopTimeoutDuration));

    auto hidlOutputShapes = NN_TRY(utils::makeExecutionFailure(V1_3::utils::convert(outputShapes)));
    const auto hidlTiming = NN_TRY(utils::makeExecutionFailure(V1_3::utils::convert(timing)));
    return std::make_pair(std::move(hidlOutputShapes), hidlTiming);
}

nn::GeneralResult<std::vector<nn::SyncFence>> convertSyncFences(
        const hidl_vec<hidl_handle>& handles) {
    std::vector<nn::SyncFence> syncFences;
    syncFences.reserve(handles.size());
    for (const auto& handle : handles) {
        auto nativeHandle = NN_TRY(convertInput(handle));
        auto syncFence = NN_TRY(utils::makeGeneralFailure(
                nn::SyncFence::create(std::move(nativeHandle)), nn::ErrorStatus::INVALID_ARGUMENT));
        syncFences.push_back(std::move(syncFence));
    }
    return syncFences;
}

nn::GeneralResult<std::pair<hidl_handle, sp<V1_3::IFencedExecutionCallback>>> executeFenced(
        const nn::SharedPreparedModel& preparedModel, const V1_3::Request& request,
        const hidl_vec<hidl_handle>& waitFor, V1_2::MeasureTiming measure,
        const V1_3::OptionalTimePoint& deadline,
        const V1_3::OptionalTimeoutDuration& loopTimeoutDuration,
        const V1_3::OptionalTimeoutDuration& duration) {
    const auto nnRequest = NN_TRY(convertInput(request));
    const auto nnWaitFor = NN_TRY(convertSyncFences(waitFor));
    const auto nnMeasure = NN_TRY(convertInput(measure));
    const auto nnDeadline = NN_TRY(convertInput(deadline));
    const auto nnLoopTimeoutDuration = NN_TRY(convertInput(loopTimeoutDuration));
    const auto nnDuration = NN_TRY(convertInput(duration));

    auto [syncFence, executeFencedCallback] = NN_TRY(preparedModel->executeFenced(
            nnRequest, nnWaitFor, nnMeasure, nnDeadline, nnLoopTimeoutDuration, nnDuration));

    auto hidlSyncFence = NN_TRY(V1_3::utils::convert(syncFence.getSharedHandle()));
    auto hidlExecuteFencedCallback = sp<FencedExecutionCallback>::make(executeFencedCallback);
    return std::make_pair(std::move(hidlSyncFence), std::move(hidlExecuteFencedCallback));
}

}  // namespace

PreparedModel::PreparedModel(nn::SharedPreparedModel preparedModel, Executor executor, uid_t userId)
    : kPreparedModel(std::move(preparedModel)), kExecutor(std::move(executor)), kUserId(userId) {
    CHECK(kPreparedModel != nullptr);
    CHECK(kExecutor != nullptr);
}

nn::SharedPreparedModel PreparedModel::getUnderlyingPreparedModel() const {
    return kPreparedModel;
}

Return<V1_0::ErrorStatus> PreparedModel::execute(const V1_0::Request& request,
                                                 const sp<V1_0::IExecutionCallback>& callback) {
    auto result = adapter::execute(kPreparedModel, kUserId, kExecutor, request, callback);
    if (!result.has_value()) {
        auto [message, code] = std::move(result).error();
        LOG(ERROR) << "adapter::PreparedModel::execute failed with " << code << ": " << message;
        notify(callback.get(), code, {}, {});
        return V1_0::utils::convert(code).value();
    }
    return V1_0::ErrorStatus::NONE;
}

Return<V1_0::ErrorStatus> PreparedModel::execute_1_2(const V1_0::Request& request,
                                                     V1_2::MeasureTiming measure,
                                                     const sp<V1_2::IExecutionCallback>& callback) {
    auto result =
            adapter::execute_1_2(kPreparedModel, kUserId, kExecutor, request, measure, callback);
    if (!result.has_value()) {
        auto [message, code] = std::move(result).error();
        LOG(ERROR) << "adapter::PreparedModel::execute_1_2 failed with " << code << ": " << message;
        notify(callback.get(), code, {}, {});
        return V1_2::utils::convert(code).value();
    }
    return V1_0::ErrorStatus::NONE;
}

Return<V1_3::ErrorStatus> PreparedModel::execute_1_3(
        const V1_3::Request& request, V1_2::MeasureTiming measure,
        const V1_3::OptionalTimePoint& deadline,
        const V1_3::OptionalTimeoutDuration& loopTimeoutDuration,
        const sp<V1_3::IExecutionCallback>& callback) {
    auto result = adapter::execute_1_3(kPreparedModel, kUserId, kExecutor, request, measure,
                                       deadline, loopTimeoutDuration, callback);
    if (!result.has_value()) {
        auto [message, code] = std::move(result).error();
        LOG(ERROR) << "adapter::PreparedModel::execute_1_3 failed with " << code << ": " << message;
        notify(callback.get(), code, {}, {});
        return V1_3::utils::convert(code).value();
    }
    return V1_3::ErrorStatus::NONE;
}

Return<void> PreparedModel::executeSynchronously(const V1_0::Request& request,
                                                 V1_2::MeasureTiming measure,
                                                 executeSynchronously_cb cb) {
    auto result = adapter::executeSynchronously(kPreparedModel, request, measure);
    if (!result.has_value()) {
        auto [message, code, outputShapes] = std::move(result).error();
        LOG(ERROR) << "adapter::PreparedModel::executeSynchronously failed with " << code << ": "
                   << message;
        cb(V1_2::utils::convert(code).value(), V1_2::utils::convert(outputShapes).value(),
           V1_2::utils::kNoTiming);
        return Void();
    }
    auto [outputShapes, timing] = std::move(result).value();
    cb(V1_0::ErrorStatus::NONE, outputShapes, timing);
    return Void();
}

Return<void> PreparedModel::executeSynchronously_1_3(
        const V1_3::Request& request, V1_2::MeasureTiming measure,
        const V1_3::OptionalTimePoint& deadline,
        const V1_3::OptionalTimeoutDuration& loopTimeoutDuration, executeSynchronously_1_3_cb cb) {
    auto result = adapter::executeSynchronously_1_3(kPreparedModel, request, measure, deadline,
                                                    loopTimeoutDuration);
    if (!result.has_value()) {
        auto [message, code, outputShapes] = std::move(result).error();
        LOG(ERROR) << "adapter::PreparedModel::executeSynchronously_1_3 failed with " << code
                   << ": " << message;
        cb(V1_3::utils::convert(code).value(), V1_3::utils::convert(outputShapes).value(),
           V1_2::utils::kNoTiming);
        return Void();
    }
    auto [outputShapes, timing] = std::move(result).value();
    cb(V1_3::ErrorStatus::NONE, outputShapes, timing);
    return Void();
}

Return<void> PreparedModel::configureExecutionBurst(
        const sp<V1_2::IBurstCallback>& callback,
        const MQDescriptorSync<V1_2::FmqRequestDatum>& requestChannel,
        const MQDescriptorSync<V1_2::FmqResultDatum>& resultChannel,
        configureExecutionBurst_cb cb) {
    const sp<V1_2::IBurstContext> burst = nn::ExecutionBurstServer::create(
            callback, requestChannel, resultChannel, this, std::chrono::microseconds{0});

    if (burst == nullptr) {
        cb(V1_0::ErrorStatus::GENERAL_FAILURE, {});
    } else {
        cb(V1_0::ErrorStatus::NONE, burst);
    }
    return Void();
}

Return<void> PreparedModel::executeFenced(const V1_3::Request& request,
                                          const hidl_vec<hidl_handle>& waitFor,
                                          V1_2::MeasureTiming measure,
                                          const V1_3::OptionalTimePoint& deadline,
                                          const V1_3::OptionalTimeoutDuration& loopTimeoutDuration,
                                          const V1_3::OptionalTimeoutDuration& duration,
                                          executeFenced_cb callback) {
    auto result = adapter::executeFenced(kPreparedModel, request, waitFor, measure, deadline,
                                         loopTimeoutDuration, duration);
    if (!result.has_value()) {
        auto [message, code] = std::move(result).error();
        LOG(ERROR) << "adapter::PreparedModel::executeFenced failed with " << code << ": "
                   << message;
        callback(V1_3::utils::convert(code).value(), {}, nullptr);
        return Void();
    }
    auto [syncFence, executeFencedCallback] = std::move(result).value();
    callback(V1_3::ErrorStatus::NONE, syncFence, executeFencedCallback);
    return Void();
}

}  // namespace android::hardware::neuralnetworks::adapter