Donate to e Foundation | Murena handsets with /e/OS | Own a part of Murena! Learn more

Commit 2504c2fe authored by Michael Butler's avatar Michael Butler
Browse files

NNAPI: Add execution preference to prepareModel (HAL) 

A model can be prepared in different ways to optimize for different
use-cases. This CL propagates the execution preference across the HAL so
that the NN service can better fit the users needs.

Bug: 77864669
Test: mma
Test: NeuralNetworksTest_static
Test: VtsHalNeuralnetworksV1_1TargetTest
Change-Id: Ib928d510d462f36b6a87d5e81010513db7829fa8
parent 1f88e209

current.txt

+2 −2
Original line number Diff line number Diff line
@@ -338,8 +338,8 @@ e15ebdf1e0a326ff5b8a59668d4d8cd3852bd88388eae91de13f5f7e1af50ed1 android.hardwar 
b8c7ed58aa8740361e63d0ce9e7c94227572a629f356958840b34809d2393a7c android.hardware.media.bufferpool@1.0::IClientManager

4a2c0dc82780e6c90731725a103feab8ab6ecf85a64e049b9cbd2b2c61620fe1 android.hardware.media.bufferpool@1.0::IConnection

6aef1218e5949f867b0104752ac536c1b707222a403341720de90141df129e3e android.hardware.media.bufferpool@1.0::types

3e4d8e0085ebe8549efb8ad4b8b400a141a3fa3f47ae23696b3e05a1612eb003 android.hardware.neuralnetworks@1.1::IDevice

50db076b03a6760557fc60ef433ba9dd2ff983cf3305eeb504b0fff3eaa604ff android.hardware.neuralnetworks@1.1::types

7698dc2382a2eeb43541840e3ee624f34108efdfb976b2bfa7c13ef15fb8c4c4 android.hardware.neuralnetworks@1.1::IDevice

5604001029a255648a9e955de0a822a48d9ba7cc259b106fb8be0cd43dc8eece android.hardware.neuralnetworks@1.1::types

8d3d86da0bfa4bf070970d8303c659f67f35d670c287d45a3f542e4fedadd578 android.hardware.nfc@1.1::INfc

e85f566698d2a2c28100e264fcf2c691a066756ddf8dd341d009ff50cfe10614 android.hardware.nfc@1.1::INfcClientCallback

5e278fcaa3287d397d8eebe1c22aaa28150f5caae1cf9381cd6dc32cb37899c5 android.hardware.nfc@1.1::types

neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp

+2 −2
Original line number Diff line number Diff line
@@ -242,8 +242,8 @@ void Execute(const sp<V1_1::IDevice>& device, std::function<V1_1::Model(void)> c 
    // launch prepare model

    sp<PreparedModelCallback> preparedModelCallback = new PreparedModelCallback();

    ASSERT_NE(nullptr, preparedModelCallback.get());

    Return<ErrorStatus> prepareLaunchStatus =

        device->prepareModel_1_1(model, preparedModelCallback);

    Return<ErrorStatus> prepareLaunchStatus = device->prepareModel_1_1(

        model, ExecutionPreference::FAST_SINGLE_ANSWER, preparedModelCallback);

    ASSERT_TRUE(prepareLaunchStatus.isOk());

    ASSERT_EQ(ErrorStatus::NONE, static_cast<ErrorStatus>(prepareLaunchStatus));

neuralnetworks/1.1/IDevice.hal

+4 −1
Original line number Diff line number Diff line
@@ -102,6 +102,8 @@ interface IDevice extends @1.0::IDevice { 
     * Multiple threads can call prepareModel on the same model concurrently.

     *

     * @param model The model to be prepared for execution.

     * @param preference Indicates the intended execution behavior of a prepared

     *                   model.

     * @param callback A callback object used to return the error status of

     *                 preparing the model for execution and the prepared model

     *                 if successful, nullptr otherwise. The callback object's
@@ -115,6 +117,7 @@ interface IDevice extends @1.0::IDevice { 
     *                - INVALID_ARGUMENT if one of the input arguments is

     *                  invalid

     */

    prepareModel_1_1(Model model, IPreparedModelCallback callback)

    prepareModel_1_1(Model model, ExecutionPreference preference,

                     IPreparedModelCallback callback)

          generates (ErrorStatus status);

};

neuralnetworks/1.1/types.hal

+21 −0
Original line number Diff line number Diff line
@@ -382,3 +382,24 @@ struct Model { 
     */

    bool relaxComputationFloat32toFloat16;

};



/**

 * Execution preferences.

 */

enum ExecutionPreference : int32_t {

    /**

     * Prefer executing in a way that minimizes battery drain.

     * This is desirable for compilations that will be executed often.

     */

    LOW_POWER = 0,

    /**

     * Prefer returning a single answer as fast as possible, even if this causes

     * more power consumption.

     */

    FAST_SINGLE_ANSWER = 1,

    /**

     * Prefer maximizing the throughput of successive frames, for example when

     * processing successive frames coming from the camera.

     */

    SUSTAINED_SPEED = 2,

};

neuralnetworks/1.1/vts/functional/ValidateModel.cpp

+31 −5
Original line number Diff line number Diff line
@@ -50,13 +50,13 @@ static void validateGetSupportedOperations(const sp<IDevice>& device, const std: 
}



static void validatePrepareModel(const sp<IDevice>& device, const std::string& message,

                                 const V1_1::Model& model) {

                                 const V1_1::Model& model, ExecutionPreference preference) {

    SCOPED_TRACE(message + " [prepareModel_1_1]");



    sp<PreparedModelCallback> preparedModelCallback = new PreparedModelCallback();

    ASSERT_NE(nullptr, preparedModelCallback.get());

    Return<ErrorStatus> prepareLaunchStatus =

        device->prepareModel_1_1(model, preparedModelCallback);

        device->prepareModel_1_1(model, preference, preparedModelCallback);

    ASSERT_TRUE(prepareLaunchStatus.isOk());

    ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, static_cast<ErrorStatus>(prepareLaunchStatus));
@@ -67,15 +67,24 @@ static void validatePrepareModel(const sp<IDevice>& device, const std::string& m 
    ASSERT_EQ(nullptr, preparedModel.get());

}



static bool validExecutionPreference(ExecutionPreference preference) {

    return preference == ExecutionPreference::LOW_POWER ||

           preference == ExecutionPreference::FAST_SINGLE_ANSWER ||

           preference == ExecutionPreference::SUSTAINED_SPEED;

}



// Primary validation function. This function will take a valid model, apply a

// mutation to it to invalidate the model, then pass it to interface calls that

// use the model. Note that the model here is passed by value, and any mutation

// to the model does not leave this function.

static void validate(const sp<IDevice>& device, const std::string& message, V1_1::Model model,

                     const std::function<void(Model*)>& mutation) {

                     const std::function<void(Model*)>& mutation,

                     ExecutionPreference preference = ExecutionPreference::FAST_SINGLE_ANSWER) {

    mutation(&model);

    if (validExecutionPreference(preference)) {

        validateGetSupportedOperations(device, message, model);

    validatePrepareModel(device, message, model);

    }

    validatePrepareModel(device, message, model, preference);

}



// Delete element from hidl_vec. hidl_vec doesn't support a "remove" operation,
@@ -486,6 +495,22 @@ static void addOperationOutputTest(const sp<IDevice>& device, const V1_1::Model& 
    }

}



///////////////////////// 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 sp<IDevice>& device, const V1_1::Model& model) {

    for (int32_t preference : invalidExecutionPreferences) {

        const std::string message =

            "mutateExecutionPreferenceTest: preference " + std::to_string(preference);

        validate(device, message, model, [](Model*) {},

                 static_cast<ExecutionPreference>(preference));

    }

}



////////////////////////// ENTRY POINT //////////////////////////////



void ValidationTest::validateModel(const V1_1::Model& model) {
@@ -503,6 +528,7 @@ void ValidationTest::validateModel(const V1_1::Model& model) { 
    removeOperationOutputTest(device, model);

    addOperationInputTest(device, model);

    addOperationOutputTest(device, model);

    mutateExecutionPreferenceTest(device, model);

}



}  // namespace functional