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

Commit a5a663c6 authored by Miao Wang's avatar Miao Wang
Browse files

Remove the data type from the OperationType enum. 

  - Now the driver needs to report a list of tuple
    {OperationType, OperandType} for capabilities.
  - Any time Operation information is passed across HIDL,
    it should be passed as the tuple {OperationType, OperandType}

Bug: 63905942
Test: mm
Change-Id: I909b5acf4936f65c242ee0925d3a1ac665f46131
parent 7dcc345b

neuralnetworks/1.0/types.hal

+44 −37
Original line number Diff line number Diff line
@@ -32,7 +32,7 @@ enum OperandType : uint32_t { 
    UINT32                    = 7,

    TENSOR_FLOAT16            = 8,

    TENSOR_FLOAT32            = 9,

    TENSOR_SYMMETRICAL_QUANT8 = 10,

    TENSOR_QUANT8_ASYMM       = 10,

};



// The type of operations.  Unlike the operation types found in
@@ -41,39 +41,39 @@ enum OperandType : uint32_t { 
// TODO: Currently they are the same.  Add a conversion when finalizing the model.

// When modifying, be sure to update HAL_NUM_OPERATION_TYPES in HalIntefaces.h.

enum OperationType : uint32_t {

    AVERAGE_POOL_FLOAT32                 = 0,

    CONCATENATION_FLOAT32                = 1,

    CONV_FLOAT32                         = 2,

    DEPTHWISE_CONV_FLOAT32               = 3,

    MAX_POOL_FLOAT32                     = 4,

    L2_POOL_FLOAT32                      = 5,

    DEPTH_TO_SPACE_FLOAT32               = 6,

    SPACE_TO_DEPTH_FLOAT32               = 7,

    LOCAL_RESPONSE_NORMALIZATION_FLOAT32 = 8,

    SOFTMAX_FLOAT32                      = 9,

    RESHAPE_FLOAT32                      = 10,

    SPLIT_FLOAT32                        = 11,

    FAKE_QUANT_FLOAT32                   = 12,

    ADD_FLOAT32                          = 13,

    FULLY_CONNECTED_FLOAT32              = 14,

    CAST_FLOAT32                         = 15,

    MUL_FLOAT32                          = 16,

    L2_NORMALIZATION_FLOAT32             = 17,

    LOGISTIC_FLOAT32                     = 18,

    RELU_FLOAT32                         = 19,

    RELU6_FLOAT32                        = 20,

    RELU1_FLOAT32                        = 21,

    TANH_FLOAT32                         = 22,

    DEQUANTIZE_FLOAT32                   = 23,

    FLOOR_FLOAT32                        = 24,

    GATHER_FLOAT32                       = 25,

    RESIZE_BILINEAR_FLOAT32              = 26,

    LSH_PROJECTION_FLOAT32               = 27,

    LSTM_FLOAT32                         = 28,

    SVDF_FLOAT32                         = 29,

    RNN_FLOAT32                          = 30,

    N_GRAM_FLOAT32                       = 31,

    LOOKUP_FLOAT32                       = 32,

    AVERAGE_POOL                 = 0,

    CONCATENATION                = 1,

    CONV                         = 2,

    DEPTHWISE_CONV               = 3,

    MAX_POOL                     = 4,

    L2_POOL                      = 5,

    DEPTH_TO_SPACE               = 6,

    SPACE_TO_DEPTH               = 7,

    LOCAL_RESPONSE_NORMALIZATION = 8,

    SOFTMAX                      = 9,

    RESHAPE                      = 10,

    SPLIT                        = 11,

    FAKE_QUANT                   = 12,

    ADD                          = 13,

    FULLY_CONNECTED              = 14,

    CAST                         = 15,

    MUL                          = 16,

    L2_NORMALIZATION             = 17,

    LOGISTIC                     = 18,

    RELU                         = 19,

    RELU6                        = 20,

    RELU1                        = 21,

    TANH                         = 22,

    DEQUANTIZE                   = 23,

    FLOOR                        = 24,

    GATHER                       = 25,

    RESIZE_BILINEAR              = 26,

    LSH_PROJECTION               = 27,

    LSTM                         = 28,

    SVDF                         = 29,

    RNN                          = 30,

    N_GRAM                       = 31,

    LOOKUP                       = 32,

};



// Two special values that can be used instead of a regular poolIndex.
@@ -102,9 +102,16 @@ struct PerformanceInfo { 
    float powerUsage;  // in picoJoules

};



struct OperationTuple {

    // The type of operation.

    OperationType operationType;

    // The input data type of operation.

    OperandType operandType;

};



// The capabilities of a driver.

struct Capabilities {

    vec<OperationType> supportedOperationTypes;

    vec<OperationTuple> supportedOperationTuples;

    // TODO Do the same for baseline model IDs

    bool cachesCompilation;

    // TODO revisit the data types and scales.
@@ -142,8 +149,8 @@ struct Operand { 


// Describes one operation of the graph.

struct Operation {

    // The type of operation.

    OperationType type;

    // The tuple describing the operation type and input type.

    OperationTuple opTuple;

    // Describes the table that contains the indexes of the inputs of the

    // operation. The offset is the index in the operandIndexes table.

    vec<uint32_t> inputs;

neuralnetworks/1.0/vts/functional/VtsHalNeuralnetworksV1_0TargetTest.cpp

+8 −6
Original line number Diff line number Diff line
@@ -66,8 +66,8 @@ TEST_F(NeuralnetworksHidlTest, StatusTest) { 
// initialization

TEST_F(NeuralnetworksHidlTest, InitializeTest) {

    Return<void> ret = device->initialize([](const Capabilities& capabilities) {

        EXPECT_NE(nullptr, capabilities.supportedOperationTypes.data());

        EXPECT_NE(0ull, capabilities.supportedOperationTypes.size());

        EXPECT_NE(nullptr, capabilities.supportedOperationTuples.data());

        EXPECT_NE(0ull, capabilities.supportedOperationTuples.size());

        EXPECT_EQ(0u, static_cast<uint32_t>(capabilities.cachesCompilation) & ~0x1);

        EXPECT_LT(0.0f, capabilities.bootupTime);

        EXPECT_LT(0.0f, capabilities.float16Performance.execTime);
@@ -92,7 +92,7 @@ Model createTestModel() { 


    const std::vector<Operand> operands = {

        {

            .type = OperandType::FLOAT32,

            .type = OperandType::TENSOR_FLOAT32,

            .dimensions = {1, 2, 2, 1},

            .numberOfConsumers = 1,

            .scale = 0.0f,
@@ -102,7 +102,7 @@ Model createTestModel() { 
                         .length = 0},

        },

        {

            .type = OperandType::FLOAT32,

            .type = OperandType::TENSOR_FLOAT32,

            .dimensions = {1, 2, 2, 1},

            .numberOfConsumers = 1,

            .scale = 0.0f,
@@ -112,7 +112,7 @@ Model createTestModel() { 
                         .length = size},

        },

        {

            .type = OperandType::FLOAT32,

            .type = OperandType::TENSOR_FLOAT32,

            .dimensions = {1, 2, 2, 1},

            .numberOfConsumers = 0,

            .scale = 0.0f,
@@ -124,7 +124,9 @@ Model createTestModel() { 
    };



    const std::vector<Operation> operations = {{

        .type = OperationType::ADD_FLOAT32, .inputs = {operand1, operand2}, .outputs = {operand3},

        .opTuple = {OperationType::ADD, OperandType::TENSOR_FLOAT32},

        .inputs = {operand1, operand2},

        .outputs = {operand3},

    }};



    const std::vector<uint32_t> inputIndexes = {operand1};