nn/common/Utils.cpp

/*
 * Copyright (C) 2017 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.
 */

#define LOG_TAG "Utils"

#include "Utils.h"

#include <android-base/logging.h>
#include <android-base/properties.h>
#include <android-base/strings.h>
#include <errno.h>
#include <poll.h>
#include <sys/system_properties.h>

#include <algorithm>
#include <limits>
#include <set>
#include <string>
#include <tuple>
#include <unordered_map>
#include <utility>
#include <vector>

#include "ControlFlow.h"
#include "NeuralNetworks.h"
#include "NeuralNetworksOEM.h"
#include "OperationResolver.h"
#include "ValidateHal.h"

namespace android {
namespace nn {

using namespace hal;

constexpr PerformanceInfo kNoPerformanceInfo = {.execTime = FLT_MAX, .powerUsage = FLT_MAX};

const char kVLogPropKey[] = "debug.nn.vlog";
int vLogMask = ~0;

// Split the space separated list of tags from verbose log setting and build the
// logging mask from it. note that '1' and 'all' are special cases to enable all
// verbose logging.
//
// NN API verbose logging setting comes from system property debug.nn.vlog.
// Example:
// setprop debug.nn.vlog 1 : enable all logging tags.
// setprop debug.nn.vlog "model compilation" : only enable logging for MODEL and
//                                             COMPILATION tags.
void initVLogMask() {
    vLogMask = 0;
    const std::string vLogSetting = android::base::GetProperty(kVLogPropKey, "");
    if (vLogSetting.empty()) {
        return;
    }

    std::unordered_map<std::string, int> vLogFlags = {{"1", -1},
                                                      {"all", -1},
                                                      {"model", MODEL},
                                                      {"compilation", COMPILATION},
                                                      {"execution", EXECUTION},
                                                      {"cpuexe", CPUEXE},
                                                      {"manager", MANAGER},
                                                      {"driver", DRIVER},
                                                      {"memory", MEMORY}};

    std::vector<std::string> elements = android::base::Split(vLogSetting, " ,:");
    for (const auto& elem : elements) {
        const auto& flag = vLogFlags.find(elem);
        if (flag == vLogFlags.end()) {
            LOG(ERROR) << "Unknown trace flag: " << elem;
            continue;
        }

        if (flag->second == -1) {
            // -1 is used for the special values "1" and "all" that enable all
            // tracing.
            vLogMask = ~0;
            return;
        } else {
            vLogMask |= 1 << flag->second;
        }
    }
}

Deadline makeDeadline(uint64_t duration) {
    const auto maxTime = Deadline::max();
    const auto currentTime = std::chrono::steady_clock::now();

    // Create Deadline. If there would be an overflow, use the max value.
    const uint64_t remainingNanoseconds =
            std::chrono::duration_cast<std::chrono::nanoseconds>(maxTime - currentTime).count();
    if (duration > remainingNanoseconds) {
        return maxTime;
    }
    return currentTime + std::chrono::nanoseconds{duration};
}

std::optional<Deadline> makeDeadline(std::optional<uint64_t> duration) {
    return duration.has_value() ? makeDeadline(*duration) : std::optional<Deadline>{};
}

static uint64_t getMaxNanosecondsSinceEpoch() {
    const auto maxTime =
            std::chrono::time_point<std::chrono::steady_clock, std::chrono::nanoseconds>::max();
    return maxTime.time_since_epoch().count();
}

std::optional<Deadline> makeDeadline(const OptionalTimePoint& timePoint) {
    using Discriminator = hal::OptionalTimePoint::hidl_discriminator;
    if (timePoint.getDiscriminator() == Discriminator::none) {
        return std::nullopt;
    }
    const uint64_t nanosecondsSinceEpoch = timePoint.nanosecondsSinceEpoch();
    const uint64_t maxNanosecondsSinceEpoch = getMaxNanosecondsSinceEpoch();

    // Clamp time point to max.
    if (nanosecondsSinceEpoch >= maxNanosecondsSinceEpoch) {
        return Deadline::max();
    }

    // Return provided time point.
    return Deadline{std::chrono::nanoseconds{nanosecondsSinceEpoch}};
}

bool hasDeadlinePassed(const std::optional<Deadline>& deadline) {
    if (!deadline.has_value()) {
        return false;
    }
    return std::chrono::steady_clock::now() >= *deadline;
}

static OptionalTimePoint makeTimePoint(const Deadline& deadline) {
    const auto timeSinceEpoch = deadline.time_since_epoch();
    const uint64_t nanosecondsSinceEpoch =
            std::chrono::duration_cast<std::chrono::nanoseconds>(timeSinceEpoch).count();
    OptionalTimePoint ret;
    ret.nanosecondsSinceEpoch(nanosecondsSinceEpoch);
    return ret;
}

OptionalTimePoint makeTimePoint(const std::optional<Deadline>& deadline) {
    return deadline.has_value() ? makeTimePoint(*deadline) : OptionalTimePoint{};
}

static bool isExtensionOperandType(int32_t type) {
    return static_cast<uint32_t>(type) > static_cast<uint32_t>(OperandTypeRange::BASE_MAX);
}

static bool isExtensionOperationType(ANeuralNetworksOperationType type) {
    return static_cast<uint32_t>(type) > static_cast<uint32_t>(OperationTypeRange::BASE_MAX);
}

bool isExtensionOperandType(OperandType type) {
    return isExtensionOperandType(static_cast<int32_t>(type));
}

bool isExtensionOperationType(OperationType type) {
    return isExtensionOperationType(static_cast<int32_t>(type));
}

namespace {

template <typename EntryType, uint32_t entryCount, uint32_t entryCountOEM>
EntryType tableLookup(const EntryType (&table)[entryCount],
                      const EntryType (&tableOEM)[entryCountOEM], uint32_t code) {
    if (code < entryCount) {
        return table[code];
    } else if (code >= kOEMCodeBase && (code - kOEMCodeBase) < entryCountOEM) {
        return tableOEM[code - kOEMCodeBase];
    } else {
        nnAssert(!"tableLookup: bad code");
        return EntryType();
    }
}

class OperationValidationContext : public IOperationValidationContext {
    DISALLOW_IMPLICIT_CONSTRUCTORS(OperationValidationContext);

   public:
    OperationValidationContext(const char* operationName, uint32_t inputCount,
                               const uint32_t* inputIndexes, uint32_t outputCount,
                               const uint32_t* outputIndexes, const Operand* operands,
                               HalVersion halVersion)
        : operationName(operationName),
          inputCount(inputCount),
          inputIndexes(inputIndexes),
          outputCount(outputCount),
          outputIndexes(outputIndexes),
          operands(operands),
          halVersion(halVersion) {}

    const char* getOperationName() const override;
    HalVersion getHalVersion() const override;

    uint32_t getNumInputs() const override;
    OperandType getInputType(uint32_t index) const override;
    Shape getInputShape(uint32_t index) const override;
    const OperandExtraParams getInputExtraParams(uint32_t index) const override;

    uint32_t getNumOutputs() const override;
    OperandType getOutputType(uint32_t index) const override;
    Shape getOutputShape(uint32_t index) const override;

   private:
    const Operand* getInputOperand(uint32_t index) const;
    const Operand* getOutputOperand(uint32_t index) const;

    const char* operationName;
    uint32_t inputCount;
    const uint32_t* inputIndexes;
    uint32_t outputCount;
    const uint32_t* outputIndexes;
    const Operand* operands;
    HalVersion halVersion;
};

const char* OperationValidationContext::getOperationName() const {
    return operationName;
}

HalVersion OperationValidationContext::getHalVersion() const {
    return halVersion;
}

const Operand* OperationValidationContext::getInputOperand(uint32_t index) const {
    CHECK(index < static_cast<uint32_t>(inputCount));
    return &operands[inputIndexes[index]];
}

const Operand* OperationValidationContext::getOutputOperand(uint32_t index) const {
    CHECK(index < static_cast<uint32_t>(outputCount));
    return &operands[outputIndexes[index]];
}

uint32_t OperationValidationContext::getNumInputs() const {
    return inputCount;
}

uint32_t OperationValidationContext::getNumOutputs() const {
    return outputCount;
}

OperandType OperationValidationContext::getInputType(uint32_t index) const {
    return getInputOperand(index)->type;
}

Shape OperationValidationContext::getInputShape(uint32_t index) const {
    const Operand* operand = getInputOperand(index);
    return {operand->type, operand->dimensions, operand->scale, operand->zeroPoint,
            operand->extraParams};
}

const OperandExtraParams OperationValidationContext::getInputExtraParams(uint32_t index) const {
    return getInputOperand(index)->extraParams;
}

OperandType OperationValidationContext::getOutputType(uint32_t index) const {
    return getOutputOperand(index)->type;
}

Shape OperationValidationContext::getOutputShape(uint32_t index) const {
    const Operand* operand = getOutputOperand(index);
    return {operand->type, operand->dimensions, operand->scale, operand->zeroPoint,
            operand->extraParams};
}

};  // anonymous namespace

#define COUNT(X) (sizeof(X) / sizeof(X[0]))

std::string getOperandTypeName(OperandType type) {
    return toString(type);
}

static std::string getOperationName(uint32_t code) {
    return getOperationName(static_cast<OperationType>(code));
}

std::string getOperationName(OperationType type) {
    return toString(type);
}

const uint32_t kSizeOfDataType[]{
        4,  // ANEURALNETWORKS_FLOAT32
        4,  // ANEURALNETWORKS_INT32
        4,  // ANEURALNETWORKS_UINT32
        4,  // ANEURALNETWORKS_TENSOR_FLOAT32
        4,  // ANEURALNETWORKS_TENSOR_INT32
        1,  // ANEURALNETWORKS_TENSOR_QUANT8_ASYMM
        1,  // ANEURALNETWORKS_BOOL
        2,  // ANEURALNETWORKS_TENSOR_QUANT16_SYMM
        2,  // ANEURALNETWORKS_TENSOR_FLOAT16
        1,  // ANEURALNETWORKS_TENSOR_BOOL8
        2,  // ANEURALNETWORKS_FLOAT16
        1,  // ANEURALNETWORKS_TENSOR_QUANT8_SYMM_PER_CHANNEL
        2,  // ANEURALNETWORKS_TENSOR_QUANT16_ASYMM
        1,  // ANEURALNETWORKS_TENSOR_QUANT8_SYMM
        1,  // ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED
        0,  // ANEURALNETWORKS_MODEL
};

static_assert(COUNT(kSizeOfDataType) == kNumberOfDataTypes, "kSizeOfDataType is incorrect");

const bool kScalarDataType[]{
        true,   // ANEURALNETWORKS_FLOAT32
        true,   // ANEURALNETWORKS_INT32
        true,   // ANEURALNETWORKS_UINT32
        false,  // ANEURALNETWORKS_TENSOR_FLOAT32
        false,  // ANEURALNETWORKS_TENSOR_INT32
        false,  // ANEURALNETWORKS_TENSOR_QUANT8_ASYMM
        true,   // ANEURALNETWORKS_BOOL
        false,  // ANEURALNETWORKS_TENSOR_QUANT16_SYMM
        false,  // ANEURALNETWORKS_TENSOR_FLOAT16
        false,  // ANEURALNETWORKS_TENSOR_BOOL8
        true,   // ANEURALNETWORKS_FLOAT16
        false,  // ANEURALNETWORKS_TENSOR_QUANT8_SYMM_PER_CHANNEL
        false,  // ANEURALNETWORKS_TENSOR_QUANT16_ASYMM
        false,  // ANEURALNETWORKS_TENSOR_QUANT8_SYMM
        false,  // ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED
        true,   // ANEURALNETWORKS_MODEL
};

static_assert(COUNT(kScalarDataType) == kNumberOfDataTypes, "kScalarDataType is incorrect");

const uint32_t kSizeOfDataTypeOEM[]{
        0,  // ANEURALNETWORKS_OEM
        1,  // ANEURALNETWORKS_TENSOR_OEM_BYTE
};

static_assert(COUNT(kSizeOfDataTypeOEM) == kNumberOfDataTypesOEM,
              "kSizeOfDataTypeOEM is incorrect");

const bool kScalarDataTypeOEM[]{
        true,   // ANEURALNETWORKS_OEM
        false,  // ANEURALNETWORKS_TENSOR_OEM_BYTE
};

static_assert(COUNT(kScalarDataTypeOEM) == kNumberOfDataTypesOEM,
              "kScalarDataTypeOEM is incorrect");

bool nonExtensionOperandTypeIsScalar(int type) {
    CHECK(!isExtensionOperandType(type)) << "Extension operand types are not supported";
    return tableLookup(kScalarDataType, kScalarDataTypeOEM, type);
}

uint32_t nonExtensionOperandSizeOfData(OperandType type, const std::vector<uint32_t>& dimensions) {
    CHECK(!isExtensionOperandType(type)) << "Size of extension operand data is unknown";
    int n = static_cast<int>(type);
    uint32_t sizeOfElement = tableLookup(kSizeOfDataType, kSizeOfDataTypeOEM, n);
    return tableLookup(kScalarDataType, kScalarDataTypeOEM, n)
                   ? sizeOfElement
                   : sizeOfTensorData(sizeOfElement, dimensions);
}

// Returns a pair of {false, size} on success, {true, 0} if size overflows uint32_t.
static std::pair<bool, uint32_t> sizeOfTensorDataHelper(uint32_t sizeOfElement,
                                                        const std::vector<uint32_t>& dimensions) {
    if (dimensions.empty()) {
        return {false, 0};
    }
    uint64_t size = static_cast<uint64_t>(sizeOfElement);
    constexpr uint64_t kMaxSize = static_cast<uint64_t>(std::numeric_limits<uint32_t>::max());
    for (uint32_t d : dimensions) {
        size *= d;
        if (size > kMaxSize) return {true, 0};
    }
    return {false, static_cast<uint32_t>(size)};
}

uint32_t sizeOfTensorData(uint32_t sizeOfElement, const std::vector<uint32_t>& dimensions) {
    const auto [overflow, size] = sizeOfTensorDataHelper(sizeOfElement, dimensions);
    CHECK(!overflow);
    return size;
}

bool nonExtensionOperandSizeOfDataOverflowsUInt32(hal::OperandType type,
                                                  const std::vector<uint32_t>& dimensions) {
    CHECK(!isExtensionOperandType(type)) << "Size of extension operand data is unknown";
    int n = static_cast<int>(type);
    uint32_t sizeOfElement = tableLookup(kSizeOfDataType, kSizeOfDataTypeOEM, n);
    return tableLookup(kScalarDataType, kScalarDataTypeOEM, n)
                   ? false
                   : sizeOfTensorDataOverflowsUInt32(sizeOfElement, dimensions);
}

bool sizeOfTensorDataOverflowsUInt32(uint32_t sizeOfElement,
                                     const std::vector<uint32_t>& dimensions) {
    return sizeOfTensorDataHelper(sizeOfElement, dimensions).first;
}

bool tensorHasUnspecifiedDimensions(int type, const uint32_t* dim, uint32_t dimCount) {
    if (!isExtensionOperandType(type)) {
        CHECK(!nonExtensionOperandTypeIsScalar(type))
                << "A scalar type can never have unspecified dimensions";
    }
    return dimCount == 0 || std::find(dim, dim + dimCount, 0) != (dim + dimCount);
}

bool tensorHasUnspecifiedDimensions(OperandType type, const std::vector<uint32_t>& dimensions) {
    return tensorHasUnspecifiedDimensions(static_cast<int>(type), dimensions.data(),
                                          dimensions.size());
}

bool tensorHasUnspecifiedDimensions(const ANeuralNetworksOperandType* type) {
    return tensorHasUnspecifiedDimensions(type->type, type->dimensions, type->dimensionCount);
}

bool tensorHasUnspecifiedDimensions(const Operand& operand) {
    return tensorHasUnspecifiedDimensions(static_cast<int>(operand.type), operand.dimensions.data(),
                                          operand.dimensions.size());
}

uint32_t alignBytesNeeded(uint32_t index, size_t length) {
    uint32_t pattern;
    if (length < 2) {
        pattern = 0;  // No alignment necessary
    } else if (length < 4) {
        pattern = 1;  // Align on 2-byte boundary
    } else {
        pattern = 3;  // Align on 4-byte boundary
    }
    uint32_t extra = (~(index - 1)) & pattern;
    return extra;
}

void logModelToInfo(const V1_0::Model& model) {
    LOG(INFO) << "V1_0::Model start";
    LOG(INFO) << "operands" << toString(model.operands);
    LOG(INFO) << "operations" << toString(model.operations);
    LOG(INFO) << "inputIndexes" << toString(model.inputIndexes);
    LOG(INFO) << "outputIndexes" << toString(model.outputIndexes);
    LOG(INFO) << "operandValues size" << model.operandValues.size();
    LOG(INFO) << "pools" << SHOW_IF_DEBUG(toString(model.pools));
}

void logModelToInfo(const V1_1::Model& model) {
    LOG(INFO) << "V1_1::Model start";
    LOG(INFO) << "operands" << toString(model.operands);
    LOG(INFO) << "operations" << toString(model.operations);
    LOG(INFO) << "inputIndexes" << toString(model.inputIndexes);
    LOG(INFO) << "outputIndexes" << toString(model.outputIndexes);
    LOG(INFO) << "operandValues size " << model.operandValues.size();
    LOG(INFO) << "pools" << SHOW_IF_DEBUG(toString(model.pools));
}

void logModelToInfo(const V1_2::Model& model) {
    LOG(INFO) << "V1_2::Model start";
    LOG(INFO) << "operands" << toString(model.operands);
    LOG(INFO) << "operations" << toString(model.operations);
    LOG(INFO) << "inputIndexes" << toString(model.inputIndexes);
    LOG(INFO) << "outputIndexes" << toString(model.outputIndexes);
    LOG(INFO) << "operandValues size" << model.operandValues.size();
    LOG(INFO) << "pools" << SHOW_IF_DEBUG(toString(model.pools));
    LOG(INFO) << "relaxComputationFloat32toFloat16" << model.relaxComputationFloat32toFloat16;
    LOG(INFO) << "extensionNameToPrefix" << toString(model.extensionNameToPrefix);
}

static void logSubgraphToInfo(std::string label, const V1_3::Subgraph& subgraph) {
    LOG(INFO) << label << ".operands" << toString(subgraph.operands);
    LOG(INFO) << label << ".operations" << toString(subgraph.operations);
    LOG(INFO) << label << ".inputIndexes" << toString(subgraph.inputIndexes);
    LOG(INFO) << label << ".outputIndexes" << toString(subgraph.outputIndexes);
}

void logModelToInfo(const V1_3::Model& model) {
    LOG(INFO) << "V1_3::Model start";
    logSubgraphToInfo("main", model.main);
    for (uint32_t i = 0, n = model.referenced.size(); i < n; ++i) {
        logSubgraphToInfo("referenced[" + std::to_string(i) + "]", model.referenced[i]);
    }
    LOG(INFO) << "operandValues size " << model.operandValues.size();
    LOG(INFO) << "pools" << SHOW_IF_DEBUG(toString(model.pools));
    LOG(INFO) << "relaxComputationFloat32toFloat16 " << model.relaxComputationFloat32toFloat16;
    LOG(INFO) << "extensionNameToPrefix" << toString(model.extensionNameToPrefix);
}

bool validateOperandSymmPerChannelQuantParams(
        const Operand& halOperand, const ANeuralNetworksSymmPerChannelQuantParams& channelQuant,
        const char* tag) {
    if (halOperand.type != OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
        return false;
    }

    NN_RET_CHECK_LT(channelQuant.channelDim, halOperand.dimensions.size()) << tag;
    NN_RET_CHECK(channelQuant.scales != nullptr) << tag;
    NN_RET_CHECK_EQ(channelQuant.scaleCount, halOperand.dimensions[channelQuant.channelDim]) << tag;
    NN_RET_CHECK_NE(halOperand.dimensions[channelQuant.channelDim], 0u)
            << tag << " channel dimension " << channelQuant.channelDim << " is underspecified";
    for (uint32_t i = 0; i < halOperand.dimensions[channelQuant.channelDim]; i++) {
        NN_RET_CHECK_GT(channelQuant.scales[i], 0.0f) << tag << " invalid scaleArray[" << i << "]";
    }
    return true;
}

static bool validateScalarDimensions(const ANeuralNetworksOperandType& type, const char* tag) {
    NN_RET_CHECK_EQ(type.dimensionCount, 0u) << tag << " invalid dimensions for scalar type";
    NN_RET_CHECK(type.dimensions == nullptr) << tag << " invalid dimensions for scalar type";
    return true;
}

static bool validateQuant8AsymmParams(const ANeuralNetworksOperandType& type, const char* tag) {
    NN_RET_CHECK(0 <= type.zeroPoint && type.zeroPoint <= 255)
            << tag << " invalid zeroPoint: " << type.zeroPoint;
    NN_RET_CHECK_GT(type.scale, 0.f) << tag << " invalid scale";
    return true;
}

static bool validateQuant8AsymmSignedParams(const ANeuralNetworksOperandType& type,
                                            const char* tag) {
    NN_RET_CHECK(-128 <= type.zeroPoint && type.zeroPoint <= 127)
            << tag << " invalid zeroPoint: " << type.zeroPoint;
    NN_RET_CHECK_GT(type.scale, 0.f) << tag << " invalid scale";
    return true;
}

static bool validateQuant8SymmParams(const ANeuralNetworksOperandType& type, const char* tag) {
    NN_RET_CHECK_EQ(type.zeroPoint, 0) << tag << " invalid zeroPoint: " << type.zeroPoint;
    NN_RET_CHECK_GT(type.scale, 0.f) << tag << " invalid scale";
    return true;
}

static bool validateQuant16AsymmParams(const ANeuralNetworksOperandType& type, const char* tag) {
    NN_RET_CHECK(0 <= type.zeroPoint && type.zeroPoint <= 65535)
            << tag << " invalid zeroPoint: " << type.zeroPoint;
    NN_RET_CHECK_GT(type.scale, 0.f) << tag << " invalid scale";
    return true;
}

static bool validateQuantSymmParams(const ANeuralNetworksOperandType& type, const char* tag) {
    NN_RET_CHECK_EQ(type.zeroPoint, 0) << tag << " zeroPoint is not zero";
    NN_RET_CHECK_GT(type.scale, 0.f) << tag << " invalid scale";
    return true;
}

static bool validateNoQuantParams(const ANeuralNetworksOperandType& type, const char* tag) {
    NN_RET_CHECK_EQ(type.zeroPoint, 0) << tag << " zeroPoint is not zero";
    NN_RET_CHECK_EQ(type.scale, 0.f) << tag << " scale is not zero";
    return true;
}

static bool validateTensorDimensions(
        const ANeuralNetworksOperandType& type,
        const Extension::OperandTypeInformation* const extensionOperandTypeInfo, const char* tag,
        bool allowPartial) {
    if (!allowPartial) {
        NN_RET_CHECK_GT(type.dimensionCount, 0u) << tag << " invalid operand dimensions";
    }
    uint64_t size =
            isExtensionOperandType(type.type)
                    ? extensionOperandTypeInfo->byteSize
                    : tableLookup(kSizeOfDataType, kSizeOfDataTypeOEM, static_cast<int>(type.type));
    constexpr uint64_t kMaxSize = std::numeric_limits<uint32_t>::max();
    for (uint32_t i = 0; i < type.dimensionCount; i++) {
        if (!allowPartial) {
            NN_RET_CHECK_NE(type.dimensions[i], 0u) << tag << " invalid operand dimensions";
        }
        if (type.dimensions[i] != 0) {
            size *= type.dimensions[i];
            NN_RET_CHECK_LE(size, kMaxSize) << tag << " operand byte size exceeds " << kMaxSize;
        }
    }
    return true;
}

static bool validateOperandTypeHelper(
        const ANeuralNetworksOperandType& type,
        const Extension::OperandTypeInformation* const extensionOperandTypeInfo, const char* tag,
        bool allowPartial) {
    NN_RET_CHECK_EQ(type.dimensionCount == 0, type.dimensions == nullptr);
    if (isExtensionOperandType(type.type)) {
        NN_RET_CHECK(extensionOperandTypeInfo != nullptr);
        if (extensionOperandTypeInfo->isTensor) {
            NN_RET_CHECK(
                    validateTensorDimensions(type, extensionOperandTypeInfo, tag, allowPartial));
        } else {
            NN_RET_CHECK(validateScalarDimensions(type, tag));
        }
        return validateNoQuantParams(type, tag);
    }

    NN_RET_CHECK(extensionOperandTypeInfo == nullptr);
    NN_RET_CHECK(validCode(kNumberOfDataTypes, kNumberOfDataTypesOEM, type.type))
            << tag << " invalid OperandType: " << type.type;

    bool isScalar = tableLookup(kScalarDataType, kScalarDataTypeOEM, type.type);
    if (isScalar) {
        NN_RET_CHECK(validateScalarDimensions(type, tag));
        if (type.type != ANEURALNETWORKS_OEM_SCALAR) {  // Historically, we have allowed OEM types
                                                        // to use quantization parameters.
            NN_RET_CHECK(validateNoQuantParams(type, tag));
        }
    } else {
        NN_RET_CHECK(validateTensorDimensions(type, extensionOperandTypeInfo, tag, allowPartial));
        if (type.type == ANEURALNETWORKS_TENSOR_QUANT8_ASYMM) {
            NN_RET_CHECK(validateQuant8AsymmParams(type, tag));
        } else if (type.type == ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED) {
            NN_RET_CHECK(validateQuant8AsymmSignedParams(type, tag));
        } else if (type.type == ANEURALNETWORKS_TENSOR_QUANT8_SYMM) {
            NN_RET_CHECK(validateQuant8SymmParams(type, tag));
        } else if (type.type == ANEURALNETWORKS_TENSOR_QUANT16_ASYMM) {
            NN_RET_CHECK(validateQuant16AsymmParams(type, tag));
        } else if (type.type == ANEURALNETWORKS_TENSOR_QUANT16_SYMM) {
            NN_RET_CHECK(validateQuantSymmParams(type, tag));
        } else if (type.type == ANEURALNETWORKS_TENSOR_INT32) {
            // TODO(b/119869082): TENSOR_INT32 should not use quantization parameters.
        } else if (type.type == ANEURALNETWORKS_TENSOR_OEM_BYTE) {
            // Historically, we have allowed OEM types to use quantization parameters.
        } else {
            NN_RET_CHECK(validateNoQuantParams(type, tag));
        }
    }

    return true;
}

int validateOperandType(const ANeuralNetworksOperandType& type,
                        const Extension::OperandTypeInformation* const extensionOperandTypeInfo,
                        const char* tag, bool allowPartial) {
    return validateOperandTypeHelper(type, extensionOperandTypeInfo, tag, allowPartial)
                   ? ANEURALNETWORKS_NO_ERROR
                   : ANEURALNETWORKS_BAD_DATA;
}

int validateOperandList(uint32_t count, const uint32_t* list, uint32_t operandCount,
                        const char* tag) {
    for (uint32_t i = 0; i < count; i++) {
        if (list[i] >= operandCount) {
            LOG(ERROR) << tag << " invalid operand index at " << i << " = " << list[i]
                       << ", operandCount " << operandCount;
            return ANEURALNETWORKS_BAD_DATA;
        }
    }
    return ANEURALNETWORKS_NO_ERROR;
}

int validateOperationOperandTypes(const std::vector<Operand>& operands, uint32_t inOperandCount,
                                  const uint32_t* inOperandIndexes,
                                  const std::vector<OperandType>& inExpectedTypes,
                                  uint32_t outOperandCount, const uint32_t* outOperandIndexes,
                                  const std::vector<OperandType>& outExpectedInTypes) {
    if (inOperandCount != static_cast<uint32_t>(inExpectedTypes.size()) ||
        outOperandCount != static_cast<uint32_t>(outExpectedInTypes.size())) {
        LOG(ERROR) << "Wrong operand count: expected " << inExpectedTypes.size() << " inputs and "
                   << outExpectedInTypes.size() << " outputs,"
                   << "got " << inOperandCount << " inputs and " << outOperandCount << " outputs";
        return ANEURALNETWORKS_BAD_DATA;
    }
    for (uint32_t i = 0; i < inOperandCount; i++) {
        if (operands[inOperandIndexes[i]].type != inExpectedTypes[i]) {
            LOG(ERROR) << "Invalid input tensor type "
                       << toString(operands[inOperandIndexes[i]].type) << " for input " << i
                       << ", expected " << toString(inExpectedTypes[i]);
            return ANEURALNETWORKS_BAD_DATA;
        }
    }
    for (uint32_t i = 0; i < outOperandCount; i++) {
        if (operands[outOperandIndexes[i]].type != outExpectedInTypes[i]) {
            LOG(ERROR) << "Invalid output tensor type "
                       << toString(operands[outOperandIndexes[i]].type) << " for input " << i
                       << ", expected " << toString(outExpectedInTypes[i]);
            return ANEURALNETWORKS_BAD_DATA;
        }
    }

    return ANEURALNETWORKS_NO_ERROR;
}

static int validateHalVersion(ANeuralNetworksOperationType opType, HalVersion halVersion,
                              HalVersion minSupportedHalVersion) {
    if (halVersion < minSupportedHalVersion) {
        LOG(ERROR) << "The given inputs and outputs for operation " << getOperationName(opType)
                   << " are only supported in " << toString(minSupportedHalVersion)
                   << " and later (validating using " << toString(halVersion) << ")";
        return ANEURALNETWORKS_BAD_DATA;
    }
    return ANEURALNETWORKS_NO_ERROR;
}

// Checks if two operands have the same types, shapes, and parameters.
// Omits lifetime, numberOfConsumers, and location.
static bool compatible(const Operand& a, const Operand& b) {
    NN_RET_CHECK(a.type == b.type) << toString(a.type) << " != " << toString(b.type);
    NN_RET_CHECK(a.dimensions == b.dimensions)
            << toString(a.dimensions) << " != " << toString(b.dimensions);
    NN_RET_CHECK_EQ(a.scale, b.scale);
    NN_RET_CHECK_EQ(a.zeroPoint, b.zeroPoint);
    NN_RET_CHECK(a.extraParams == b.extraParams)
            << toString(a.extraParams) << " != " << toString(b.extraParams);
    return true;
}

static bool validateConditionOperand(const Operand& operand) {
    NN_RET_CHECK(operand.type == OperandType::TENSOR_BOOL8)
            << "Unexpected condition operand type: " << toString(operand.type);
    NN_RET_CHECK_EQ(operand.dimensions.size(), 1u) << "Condition operand must be a singleton";
    NN_RET_CHECK_EQ(operand.dimensions[0], 1u) << "Condition operand must be a singleton";
    return true;
}

static void checkSubgraphValidationHelper(const SubgraphValidationHelper& helper) {
    CHECK(helper.isValidSubgraphReference != nullptr);
    CHECK(helper.getSubgraphInputCount != nullptr);
    CHECK(helper.getSubgraphOutputCount != nullptr);
    CHECK(helper.getSubgraphInputOperand != nullptr);
    CHECK(helper.getSubgraphOutputOperand != nullptr);
}

static bool validateIfOperation(uint32_t inputCount, const uint32_t* inputs, uint32_t outputCount,
                                const uint32_t* outputs, const std::vector<Operand>& operands,
                                const SubgraphValidationHelper& helper) {
    namespace op = operation_if;
    checkSubgraphValidationHelper(helper);
    NN_RET_CHECK_GE(inputCount, 3u) << "ANEURALNETWORKS_IF must have at least 3 inputs";
    NN_RET_CHECK_GE(outputCount, 1u) << "ANEURALNETWORKS_IF must have at least 1 output";
    auto validateBranchOperand = [&](const Operand& branchModelOperand) -> bool {
        NN_RET_CHECK(helper.isValidSubgraphReference(branchModelOperand))
                << "Operand is not a valid subgraph reference";
        const uint32_t branchModelInputCount = helper.getSubgraphInputCount(branchModelOperand);
        const uint32_t branchModelOutputCount = helper.getSubgraphOutputCount(branchModelOperand);
        NN_RET_CHECK_EQ(inputCount, op::kFirstInput + branchModelInputCount);
        NN_RET_CHECK_EQ(outputCount, branchModelOutputCount);
        for (uint32_t i = 0; i < branchModelInputCount; ++i) {
            const Operand& innerOperand = *helper.getSubgraphInputOperand(branchModelOperand, i);
            const Operand& outerOperand = operands[inputs[op::kFirstInput + i]];
            NN_RET_CHECK(compatible(innerOperand, outerOperand));
        }
        for (uint32_t i = 0; i < branchModelOutputCount; ++i) {
            const Operand& innerOperand = *helper.getSubgraphOutputOperand(branchModelOperand, i);
            const Operand& outerOperand = operands[outputs[i]];
            NN_RET_CHECK(compatible(innerOperand, outerOperand));
        }
        return true;
    };
    NN_RET_CHECK(validateConditionOperand(operands[inputs[op::kCondBoolOperand]]))
            << "Validation failed for IF condition operand";
    NN_RET_CHECK(validateBranchOperand(operands[inputs[op::kThenModelOperand]]))
            << "Validation failed for IF then model";
    NN_RET_CHECK(validateBranchOperand(operands[inputs[op::kElseModelOperand]]))
            << "Validation failed for IF else model";
    return true;
}

static bool validateWhileOperation(uint32_t inputCount, const uint32_t* inputs,
                                   uint32_t outputCount, const uint32_t* outputs,
                                   const std::vector<Operand>& operands,
                                   const SubgraphValidationHelper& helper) {
    // Let the loop have
    // - m >= 1 input-output operands,
    // - k >= 0 state-only operands, and
    // - n >= 0 input-only operands.
    // Then
    // - the WHILE loop operation has (2 + m + k + n) inputs and m outputs.
    // - the condition model has (m + k + n) inputs and 1 output.
    // - the body model has (m + k + n) inputs and (m + k) outputs.
    namespace op = operation_while;
    checkSubgraphValidationHelper(helper);
    NN_RET_CHECK_GE(inputCount, 3u) << "ANEURALNETWORKS_WHILE must have at least 3 inputs";
    NN_RET_CHECK_GE(outputCount, 1u) << "ANEURALNETWORKS_WHILE must have at least 1 output";
    auto validateCondOperand = [&](const Operand& condModelOperand) -> bool {
        NN_RET_CHECK(helper.isValidSubgraphReference(condModelOperand))
                << "Operand is not a valid subgraph reference";
        const uint32_t condModelInputCount = helper.getSubgraphInputCount(condModelOperand);
        const uint32_t condModelOutputCount = helper.getSubgraphOutputCount(condModelOperand);
        NN_RET_CHECK_EQ(inputCount, op::kFirstInput + condModelInputCount);
        NN_RET_CHECK_EQ(condModelOutputCount, 1u);
        for (uint32_t i = 0; i < condModelInputCount; ++i) {
            const Operand& innerOperand = *helper.getSubgraphInputOperand(condModelOperand, i);
            const Operand& outerOperand = operands[inputs[op::kFirstInput + i]];
            NN_RET_CHECK(compatible(innerOperand, outerOperand));
        }
        NN_RET_CHECK(
                validateConditionOperand(*helper.getSubgraphOutputOperand(condModelOperand, 0)));
        return true;
    };
    auto validateBodyOperand = [&](const Operand& bodyModelOperand) -> bool {
        NN_RET_CHECK(helper.isValidSubgraphReference(bodyModelOperand))
                << "Operand is not a valid subgraph reference";
        const uint32_t bodyModelInputCount = helper.getSubgraphInputCount(bodyModelOperand);
        const uint32_t bodyModelOutputCount = helper.getSubgraphOutputCount(bodyModelOperand);
        NN_RET_CHECK_EQ(inputCount, op::kFirstInput + bodyModelInputCount);
        NN_RET_CHECK_GE(bodyModelOutputCount, outputCount);
        NN_RET_CHECK_GE(bodyModelInputCount, bodyModelOutputCount);
        const uint32_t inputOutputCount = outputCount;
        const uint32_t stateOnlyCount = bodyModelOutputCount - inputOutputCount;
        const uint32_t inputOnlyCount = bodyModelInputCount - bodyModelOutputCount;
        for (uint32_t i = 0, n = inputOutputCount + stateOnlyCount + inputOnlyCount; i < n; ++i) {
            const Operand& innerOperand = *helper.getSubgraphInputOperand(bodyModelOperand, i);
            const Operand& outerOperand = operands[inputs[op::kFirstInput + i]];
            NN_RET_CHECK(compatible(innerOperand, outerOperand));
        }
        for (uint32_t i = 0; i < inputOutputCount; ++i) {
            const Operand& innerOperand = *helper.getSubgraphOutputOperand(bodyModelOperand, i);
            const Operand& outerOperand = operands[outputs[i]];
            NN_RET_CHECK(compatible(innerOperand, outerOperand));
        }
        for (uint32_t i = 0, n = inputOutputCount + stateOnlyCount; i < n; ++i) {
            const Operand& inputOperand = *helper.getSubgraphInputOperand(bodyModelOperand, i);
            const Operand& outputOperand = *helper.getSubgraphOutputOperand(bodyModelOperand, i);
            NN_RET_CHECK(compatible(inputOperand, outputOperand));
        }
        return true;
    };
    NN_RET_CHECK(validateCondOperand(operands[inputs[op::kCondModelOperand]]))
            << "Validation failed for WHILE condition model";
    NN_RET_CHECK(validateBodyOperand(operands[inputs[op::kBodyModelOperand]]))
            << "Validation failed for WHILE body model";
    return true;
}

static inline int validateOperation(ANeuralNetworksOperationType opType, uint32_t inputCount,
                                    const uint32_t* inputIndexes, uint32_t outputCount,
                                    const uint32_t* outputIndexes,
                                    const std::vector<hal::Operand>& operands,
                                    HalVersion halVersion) {
    if (opType == ANEURALNETWORKS_IF || opType == ANEURALNETWORKS_WHILE) {
        NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
        LOG(ERROR) << "This validateOperation() overload does not support control flow";
        return ANEURALNETWORKS_BAD_DATA;
    }
    return validateOperation(opType, inputCount, inputIndexes, outputCount, outputIndexes, operands,
                             halVersion, {});
}

int validateOperation(ANeuralNetworksOperationType opType, uint32_t inputCount,
                      const uint32_t* inputIndexes, uint32_t outputCount,
                      const uint32_t* outputIndexes, const std::vector<Operand>& operands,
                      HalVersion halVersion, const SubgraphValidationHelper& helper) {
    NN_RETURN_IF_ERROR(validateOperandList(inputCount, inputIndexes,
                                           static_cast<uint32_t>(operands.size()),
                                           "ANeuralNetworksModel_addOperation inputs"));
    NN_RETURN_IF_ERROR(validateOperandList(outputCount, outputIndexes,
                                           static_cast<uint32_t>(operands.size()),
                                           "ANeuralNetworksModel_addOperation outputs"));

    if (isExtensionOperationType(opType)) {
        if (halVersion < HalVersion::V1_2) {
            LOG(ERROR)
                    << "Extension operations are supported since HAL version 1.2, validating using "
                    << toString(halVersion);
            return ANEURALNETWORKS_BAD_DATA;
        }
        // There is no other validation we can do for an extension operation.
        return ANEURALNETWORKS_NO_ERROR;
    }

    auto logInvalidInOutNumber = [opType, inputCount, outputCount](int expIn, int expOut) {
        LOG(ERROR) << "Invalid number of input operands (" << inputCount << ", expected " << expIn
                   << ") or output operands (" << outputCount << ", expected " << expOut
                   << ") for operation " << getOperationName(opType);
    };

    switch (opType) {
        case ANEURALNETWORKS_OEM_OPERATION: {
            return ANEURALNETWORKS_NO_ERROR;
        }
        case ANEURALNETWORKS_RESHAPE: {
            if (inputCount != 2 || outputCount != 1) {
                logInvalidInOutNumber(2, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT32) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
                inExpectedTypes = {OperandType::TENSOR_FLOAT32, OperandType::TENSOR_INT32};
                outExpectedTypes = {OperandType::TENSOR_FLOAT32};
            } else if (inputType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes = {OperandType::TENSOR_FLOAT16, OperandType::TENSOR_INT32};
                outExpectedTypes = {OperandType::TENSOR_FLOAT16};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
                inExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM, OperandType::TENSOR_INT32};
                outExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
                inExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM_SIGNED,
                                   OperandType::TENSOR_INT32};
                outExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM_SIGNED};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            const auto inputRank = operands[inputIndexes[0]].dimensions.size();
            if (inputRank > 4) {
                LOG(ERROR) << "Unsupported input tensor rank for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_DEPTH_TO_SPACE: {
            if ((inputCount != 3 && inputCount != 2) || outputCount != 1) {
                LOG(ERROR) << "Invalid number of input operands (" << inputCount
                           << ", expected 3 or 2) or output operands (" << outputCount
                           << ", expected 1) for operation " << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT32) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
                inExpectedTypes = {OperandType::TENSOR_FLOAT32, OperandType::INT32};
                outExpectedTypes = {OperandType::TENSOR_FLOAT32};
            } else if (inputType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes = {OperandType::TENSOR_FLOAT16, OperandType::INT32};
                outExpectedTypes = {OperandType::TENSOR_FLOAT16};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
                inExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM, OperandType::INT32};
                outExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
                inExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM_SIGNED, OperandType::INT32};
                outExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM_SIGNED};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            if (inputCount == 3) {
                inExpectedTypes.push_back(OperandType::BOOL);
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_SPACE_TO_DEPTH: {
            if ((inputCount != 3 && inputCount != 2) || outputCount != 1) {
                LOG(ERROR) << "Invalid number of input operands (" << inputCount
                           << ", expected 3 or 2) or output operands (" << outputCount
                           << ", expected 1) for operation " << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT32) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
                inExpectedTypes = {OperandType::TENSOR_FLOAT32, OperandType::INT32};
                outExpectedTypes = {OperandType::TENSOR_FLOAT32};
            } else if (inputType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes = {OperandType::TENSOR_FLOAT16, OperandType::INT32};
                outExpectedTypes = {OperandType::TENSOR_FLOAT16};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
                inExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM, OperandType::INT32};
                outExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
                inExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM_SIGNED, OperandType::INT32};
                outExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM_SIGNED};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            if (inputCount == 3) {
                inExpectedTypes.push_back(OperandType::BOOL);
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_EMBEDDING_LOOKUP: {
            if (inputCount != 2 || outputCount != 1) {
                logInvalidInOutNumber(2, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[1]].type;
            if (inputType != OperandType::TENSOR_FLOAT16 &&
                inputType != OperandType::TENSOR_FLOAT32 &&
                inputType != OperandType::TENSOR_INT32 &&
                inputType != OperandType::TENSOR_QUANT8_ASYMM &&
                inputType != OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            std::vector<OperandType> inExpectedTypes = {OperandType::TENSOR_INT32, inputType};
            std::vector<OperandType> outExpectedTypes = {inputType};
            if (inputType == OperandType::TENSOR_FLOAT16 ||
                inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
            } else if (inputType == OperandType::TENSOR_INT32 ||
                       inputType == OperandType::TENSOR_QUANT8_ASYMM) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_HASHTABLE_LOOKUP: {
            if (inputCount != 3 || outputCount != 2) {
                logInvalidInOutNumber(3, 2);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[2]].type;
            if (inputType != OperandType::TENSOR_FLOAT32 &&
                inputType != OperandType::TENSOR_INT32 &&
                inputType != OperandType::TENSOR_QUANT8_ASYMM) {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            std::vector<OperandType> inExpectedTypes = {OperandType::TENSOR_INT32,
                                                        OperandType::TENSOR_INT32, inputType};
            std::vector<OperandType> outExpectedTypes = {inputType,
                                                         OperandType::TENSOR_QUANT8_ASYMM};
            NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_LSH_PROJECTION: {
            if (inputCount != 4 || outputCount != 1) {
                logInvalidInOutNumber(4, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[1]].type;
            if (inputType != OperandType::TENSOR_FLOAT16 &&
                inputType != OperandType::TENSOR_FLOAT32 &&
                inputType != OperandType::TENSOR_INT32 &&
                inputType != OperandType::TENSOR_QUANT8_ASYMM) {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto hashType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            if (hashType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT16,
                        inputType,
                        OperandType::TENSOR_FLOAT16,
                        OperandType::INT32,
                };
            } else if (hashType == OperandType::TENSOR_FLOAT32) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT32,
                        inputType,
                        OperandType::TENSOR_FLOAT32,
                        OperandType::INT32,
                };
            } else {
                LOG(ERROR) << "Unsupported hash tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            std::vector<OperandType> outExpectedTypes = {OperandType::TENSOR_INT32};
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_BIDIRECTIONAL_SEQUENCE_LSTM: {
            std::vector<OperandType> inExpectedTypes;
            auto inputType = operands[inputIndexes[0]].type;
            if (inputType != OperandType::TENSOR_FLOAT32 &&
                inputType != OperandType::TENSOR_FLOAT16) {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }

            inExpectedTypes = {};
            for (int i = 0; i < 48; ++i) {
                inExpectedTypes.push_back(inputType);
            }
            inExpectedTypes.push_back(OperandType::INT32);
            inExpectedTypes.push_back(inputType == OperandType::TENSOR_FLOAT32
                                              ? OperandType::FLOAT32
                                              : OperandType::FLOAT16);
            inExpectedTypes.push_back(inputType == OperandType::TENSOR_FLOAT32
                                              ? OperandType::FLOAT32
                                              : OperandType::FLOAT16);
            inExpectedTypes.push_back(OperandType::BOOL);
            inExpectedTypes.push_back(OperandType::BOOL);
            for (int i = 0; i < 8; ++i) {
                inExpectedTypes.push_back(inputType);
            }

            const uint32_t kNumOutputs = 2;
            const uint32_t kNumOutputsMerged = 1;
            const uint32_t kNumOutputsWithState = 6;
            const uint32_t kNumOutputsMergedWithState = 5;

            if (inputCount != 61 ||
                (outputCount != kNumOutputs && outputCount != kNumOutputsMerged &&
                 outputCount != kNumOutputsWithState &&
                 outputCount != kNumOutputsMergedWithState)) {
                LOG(ERROR) << "Invalid number of input operands (" << inputCount
                           << ", expected 61) or output operands (" << outputCount
                           << ", expected 1, 2, 5 or 6) for operation " << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            HalVersion minSupportedHalVersion = HalVersion::V1_2;
            if (outputCount == kNumOutputsWithState || outputCount == kNumOutputsMergedWithState) {
                minSupportedHalVersion = HalVersion::V1_3;
            }
            NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, minSupportedHalVersion));
            std::vector<OperandType> outExpectedTypes(outputCount, inputType);
            auto status = validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                        inExpectedTypes, outputCount, outputIndexes,
                                                        outExpectedTypes);
            return status;
        }
        case ANEURALNETWORKS_LSTM: {
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            auto inputType = operands[inputIndexes[0]].type;
            if (inputType != OperandType::TENSOR_FLOAT32 &&
                inputType != OperandType::TENSOR_FLOAT16) {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }

            inExpectedTypes = {inputType,         inputType, inputType, inputType, inputType,
                               inputType,         inputType, inputType, inputType, inputType,
                               inputType,         inputType, inputType, inputType, inputType,
                               inputType,         inputType, inputType, inputType, inputType,
                               OperandType::INT32};
            if (inputType == OperandType::TENSOR_FLOAT32) {
                inExpectedTypes.push_back(OperandType::FLOAT32);
                inExpectedTypes.push_back(OperandType::FLOAT32);
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes.push_back(OperandType::FLOAT16);
                inExpectedTypes.push_back(OperandType::FLOAT16);
            }

            outExpectedTypes = {inputType, inputType, inputType, inputType};
            if (inputCount == 23 && outputCount == 4) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
            } else if (inputCount == 27 && outputCount == 4) {
                for (int i = 0; i < 4; ++i) {
                    inExpectedTypes.push_back(inputType);
                }
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            } else {
                LOG(ERROR) << "Invalid number of input operands (" << inputCount
                           << ", expected 23 or 27) or output operands (" << outputCount
                           << ", expected 4) for operation " << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_QUANTIZED_16BIT_LSTM: {
            if (inputCount != 15 || outputCount != 2) {
                logInvalidInOutNumber(15, 2);
                return ANEURALNETWORKS_BAD_DATA;
            }
            NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            std::vector<OperandType> inExpectedTypes = {
                    OperandType::TENSOR_QUANT8_ASYMM, OperandType::TENSOR_QUANT8_ASYMM,
                    OperandType::TENSOR_QUANT8_ASYMM, OperandType::TENSOR_QUANT8_ASYMM,
                    OperandType::TENSOR_QUANT8_ASYMM, OperandType::TENSOR_QUANT8_ASYMM,
                    OperandType::TENSOR_QUANT8_ASYMM, OperandType::TENSOR_QUANT8_ASYMM,
                    OperandType::TENSOR_QUANT8_ASYMM, OperandType::TENSOR_INT32,
                    OperandType::TENSOR_INT32,        OperandType::TENSOR_INT32,
                    OperandType::TENSOR_INT32,        OperandType::TENSOR_QUANT16_SYMM,
                    OperandType::TENSOR_QUANT8_ASYMM};
            std::vector<OperandType> outExpectedTypes = {OperandType::TENSOR_QUANT16_SYMM,
                                                         OperandType::TENSOR_QUANT8_ASYMM};
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_RANDOM_MULTINOMIAL: {
            if (inputCount != 3 || outputCount != 1) {
                logInvalidInOutNumber(3, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            OperandType inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT32 ||
                inputType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes = {
                        inputType,
                        OperandType::INT32,
                        OperandType::TENSOR_INT32,
                };
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            std::vector<OperandType> outExpectedTypes = {OperandType::TENSOR_INT32};
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_RNN: {
            if (inputCount != 6 || outputCount != 2) {
                logInvalidInOutNumber(6, 2);
                return ANEURALNETWORKS_BAD_DATA;
            }
            OperandType inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT32) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT32, OperandType::TENSOR_FLOAT32,
                        OperandType::TENSOR_FLOAT32, OperandType::TENSOR_FLOAT32,
                        OperandType::TENSOR_FLOAT32, OperandType::INT32,
                };
                outExpectedTypes = {
                        OperandType::TENSOR_FLOAT32,
                        OperandType::TENSOR_FLOAT32,
                };
            } else if (inputType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT16, OperandType::TENSOR_FLOAT16,
                        OperandType::TENSOR_FLOAT16, OperandType::TENSOR_FLOAT16,
                        OperandType::TENSOR_FLOAT16, OperandType::INT32,
                };
                outExpectedTypes = {
                        OperandType::TENSOR_FLOAT16,
                        OperandType::TENSOR_FLOAT16,
                };
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_SVDF: {
            if (inputCount != 7 || outputCount != 2) {
                logInvalidInOutNumber(7, 2);
                return ANEURALNETWORKS_BAD_DATA;
            }
            OperandType inputType = operands[inputIndexes[0]].type;
            if (inputType == OperandType::TENSOR_FLOAT32) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_0));

            } else if (inputType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            std::vector<OperandType> inExpectedTypes = {
                    inputType, inputType,          inputType,          inputType,
                    inputType, OperandType::INT32, OperandType::INT32,
            };
            std::vector<OperandType> outExpectedTypes = {inputType, inputType};
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_BATCH_TO_SPACE_ND: {
            if ((inputCount != 3 && inputCount != 2) || outputCount != 1) {
                LOG(ERROR) << "Invalid number of input operands (" << inputCount
                           << ", expected 3 or 2) or output operands (" << outputCount
                           << ", expected 1) for operation " << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT32) {
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT32,
                        OperandType::TENSOR_INT32,
                };
                outExpectedTypes = {OperandType::TENSOR_FLOAT32};
            } else if (inputType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT16,
                        OperandType::TENSOR_INT32,
                };
                outExpectedTypes = {OperandType::TENSOR_FLOAT16};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM) {
                inExpectedTypes = {
                        OperandType::TENSOR_QUANT8_ASYMM,
                        OperandType::TENSOR_INT32,
                };
                outExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
                inExpectedTypes = {
                        OperandType::TENSOR_QUANT8_ASYMM_SIGNED,
                        OperandType::TENSOR_INT32,
                };
                outExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM_SIGNED};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            if (inputCount == 3) {
                inExpectedTypes.push_back(OperandType::BOOL);
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_1));
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_SPACE_TO_BATCH_ND: {
            if ((inputCount != 4 && inputCount != 3) || outputCount != 1) {
                LOG(ERROR) << "Invalid number of input operands (" << inputCount
                           << ", expected 4 or 3) or output operands (" << outputCount
                           << ", expected 1) for operation " << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT32) {
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT32,
                        OperandType::TENSOR_INT32,
                        OperandType::TENSOR_INT32,
                };
                outExpectedTypes = {OperandType::TENSOR_FLOAT32};
            } else if (inputType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT16,
                        OperandType::TENSOR_INT32,
                        OperandType::TENSOR_INT32,
                };
                outExpectedTypes = {OperandType::TENSOR_FLOAT16};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM) {
                if (operands[inputIndexes[0]].zeroPoint != 0) {
                    NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                }
                inExpectedTypes = {
                        OperandType::TENSOR_QUANT8_ASYMM,
                        OperandType::TENSOR_INT32,
                        OperandType::TENSOR_INT32,
                };
                outExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
                inExpectedTypes = {
                        OperandType::TENSOR_QUANT8_ASYMM_SIGNED,
                        OperandType::TENSOR_INT32,
                        OperandType::TENSOR_INT32,
                };
                outExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM_SIGNED};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            if (inputCount == 4) {
                inExpectedTypes.push_back(OperandType::BOOL);
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_1));
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_PAD: {
            if (inputCount != 2 || outputCount != 1) {
                logInvalidInOutNumber(2, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT32) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_1));
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT32,
                        OperandType::TENSOR_INT32,
                };
                outExpectedTypes = {OperandType::TENSOR_FLOAT32};
            } else if (inputType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT16,
                        OperandType::TENSOR_INT32,
                };
                outExpectedTypes = {OperandType::TENSOR_FLOAT16};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM ||
                       inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                    NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
                } else {
                    if (operands[inputIndexes[0]].zeroPoint == 0) {
                        NN_RETURN_IF_ERROR(
                                validateHalVersion(opType, halVersion, HalVersion::V1_1));
                    } else {
                        NN_RETURN_IF_ERROR(
                                validateHalVersion(opType, halVersion, HalVersion::V1_2));
                    }
                }
                inExpectedTypes = {
                        inputType,
                        OperandType::TENSOR_INT32,
                };
                outExpectedTypes = {inputType};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            const auto inputRank = operands[inputIndexes[0]].dimensions.size();
            if (inputRank > 4) {
                LOG(ERROR) << "Unsupported input tensor rank for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_PAD_V2: {
            if (inputCount != 3 || outputCount != 1) {
                logInvalidInOutNumber(3, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT32) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT32,
                        OperandType::TENSOR_INT32,
                        OperandType::FLOAT32,
                };
                outExpectedTypes = {OperandType::TENSOR_FLOAT32};
            } else if (inputType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                inExpectedTypes = {
                        OperandType::TENSOR_FLOAT16,
                        OperandType::TENSOR_INT32,
                        OperandType::FLOAT16,
                };
                outExpectedTypes = {OperandType::TENSOR_FLOAT16};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM ||
                       inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                    NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
                } else {
                    NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
                }
                inExpectedTypes = {
                        inputType,
                        OperandType::TENSOR_INT32,
                        OperandType::INT32,
                };  // TODO(b/116699425): Make it UINT8.
                outExpectedTypes = {inputType};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            const auto inputRank = operands[inputIndexes[0]].dimensions.size();
            if (inputRank > 4) {
                LOG(ERROR) << "Unsupported input tensor rank for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_CAST: {
            if (inputCount != 1 || outputCount != 1) {
                logInvalidInOutNumber(1, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            auto outputType = operands[outputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if ((inputType == OperandType::TENSOR_FLOAT16 ||
                 inputType == OperandType::TENSOR_FLOAT32 ||
                 inputType == OperandType::TENSOR_INT32 ||
                 inputType == OperandType::TENSOR_QUANT8_ASYMM) &&
                (outputType == OperandType::TENSOR_FLOAT16 ||
                 outputType == OperandType::TENSOR_FLOAT32 ||
                 outputType == OperandType::TENSOR_INT32 ||
                 outputType == OperandType::TENSOR_QUANT8_ASYMM)) {
                inExpectedTypes = {inputType};
                outExpectedTypes = {outputType};
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            } else if (inputType == OperandType::TENSOR_BOOL8 ||
                       inputType == OperandType::TENSOR_QUANT16_ASYMM ||
                       inputType == OperandType::TENSOR_QUANT16_SYMM ||
                       inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED ||
                       inputType == OperandType::TENSOR_QUANT8_SYMM) {
                inExpectedTypes = {inputType};
                outExpectedTypes = {inputType};  // Only identity CAST is supported.
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
            } else {
                LOG(ERROR) << "Unsupported data type for operation " << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_MEAN: {
            if (inputCount != 3 || outputCount != 1) {
                logInvalidInOutNumber(3, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            const auto inputRank = operands[inputIndexes[0]].dimensions.size();
            if (inputRank > 4) {
                LOG(ERROR) << "Unsupported input tensor rank for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            if (inputType == OperandType::TENSOR_FLOAT32) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_1));
            } else if (inputType == OperandType::TENSOR_FLOAT16) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_1));
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            std::vector<OperandType> inExpectedTypes = {inputType, OperandType::TENSOR_INT32,
                                                        OperandType::INT32};
            std::vector<OperandType> outExpectedTypes = {inputType};
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_ARGMAX:
        case ANEURALNETWORKS_ARGMIN: {
            if (inputCount != 2 || outputCount != 1) {
                logInvalidInOutNumber(2, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT16 ||
                inputType == OperandType::TENSOR_FLOAT32 ||
                inputType == OperandType::TENSOR_INT32 ||
                inputType == OperandType::TENSOR_QUANT8_ASYMM ||
                inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                inExpectedTypes = {inputType, OperandType::INT32};
                outExpectedTypes = {OperandType::TENSOR_INT32};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_EXPAND_DIMS: {
            if (inputCount != 2 || outputCount != 1) {
                logInvalidInOutNumber(2, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT16 ||
                inputType == OperandType::TENSOR_FLOAT32 ||
                inputType == OperandType::TENSOR_INT32 ||
                inputType == OperandType::TENSOR_QUANT8_ASYMM ||
                inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                inExpectedTypes = {inputType, OperandType::INT32};
                outExpectedTypes = {inputType};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_SPLIT: {
            if (inputCount != 3) {
                LOG(ERROR) << "Invalid number of input operands (" << inputCount << ", expected 3)"
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            if (inputType != OperandType::TENSOR_FLOAT16 &&
                inputType != OperandType::TENSOR_FLOAT32 &&
                inputType != OperandType::TENSOR_INT32 &&
                inputType != OperandType::TENSOR_QUANT8_ASYMM &&
                inputType != OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            }
            std::vector<OperandType> inExpectedTypes = {inputType, OperandType::INT32,
                                                        OperandType::INT32};
            std::vector<OperandType> outExpectedTypes(outputCount, inputType);
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_MAXIMUM:
        case ANEURALNETWORKS_MINIMUM: {
            if (inputCount != 2 || outputCount != 1) {
                logInvalidInOutNumber(2, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            OperandType inputType = operands[inputIndexes[0]].type;
            if (inputType == OperandType::TENSOR_FLOAT16 ||
                inputType == OperandType::TENSOR_FLOAT32 ||
                inputType == OperandType::TENSOR_INT32 ||
                inputType == OperandType::TENSOR_QUANT8_ASYMM ||
                inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                inExpectedTypes = {inputType, inputType};
                outExpectedTypes = {inputType};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_GROUPED_CONV_2D: {
            if ((inputCount != 12 && inputCount != 9) || outputCount != 1) {
                LOG(ERROR) << "Invalid number of input operands (" << inputCount
                           << ", expected 12 or 9) or output operands (" << outputCount
                           << ", expected 1) for operation " << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            auto filterType = operands[inputIndexes[1]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT32) {
                inExpectedTypes = {OperandType::TENSOR_FLOAT32, OperandType::TENSOR_FLOAT32,
                                   OperandType::TENSOR_FLOAT32, OperandType::INT32,
                                   OperandType::INT32,          OperandType::INT32,
                                   OperandType::INT32,          OperandType::INT32};
                outExpectedTypes = {OperandType::TENSOR_FLOAT32};
            } else if (inputType == OperandType::TENSOR_FLOAT16) {
                inExpectedTypes = {OperandType::TENSOR_FLOAT16, OperandType::TENSOR_FLOAT16,
                                   OperandType::TENSOR_FLOAT16, OperandType::INT32,
                                   OperandType::INT32,          OperandType::INT32,
                                   OperandType::INT32,          OperandType::INT32};
                outExpectedTypes = {OperandType::TENSOR_FLOAT16};
            } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM ||
                       inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                if (filterType != inputType &&
                    filterType != OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
                    LOG(ERROR) << "Unsupported filter tensor type for operation "
                               << getOperationName(opType);
                    return ANEURALNETWORKS_BAD_DATA;
                }

                if (filterType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL &&
                    operands[inputIndexes[1]].extraParams.channelQuant().channelDim != 0) {
                    LOG(ERROR) << "Unsupported filter tensor channel dimension for operation "
                               << getOperationName(opType);
                    return ANEURALNETWORKS_BAD_DATA;
                }

                inExpectedTypes = {
                        inputType,          filterType,         OperandType::TENSOR_INT32,
                        OperandType::INT32, OperandType::INT32, OperandType::INT32,
                        OperandType::INT32, OperandType::INT32};
                outExpectedTypes = {inputType};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }

            if (inputCount == 12) {
                std::vector<OperandType> explicitScalarTypes(3, OperandType::INT32);
                inExpectedTypes.insert(inExpectedTypes.end(), explicitScalarTypes.begin(),
                                       explicitScalarTypes.end());
            }
            inExpectedTypes.push_back(OperandType::BOOL);
            if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_TILE: {
            if (inputCount != 2 || outputCount != 1) {
                logInvalidInOutNumber(2, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT16 ||
                inputType == OperandType::TENSOR_FLOAT32 ||
                inputType == OperandType::TENSOR_INT32 ||
                inputType == OperandType::TENSOR_QUANT8_ASYMM ||
                inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                inExpectedTypes = {inputType, OperandType::TENSOR_INT32};
                outExpectedTypes = {inputType};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_POW: {
            if (inputCount != 2 || outputCount != 1) {
                logInvalidInOutNumber(2, 1);
                return ANEURALNETWORKS_BAD_DATA;
            }
            auto inputType = operands[inputIndexes[0]].type;
            std::vector<OperandType> inExpectedTypes;
            std::vector<OperandType> outExpectedTypes;
            if (inputType == OperandType::TENSOR_FLOAT16 ||
                inputType == OperandType::TENSOR_FLOAT32) {
                inExpectedTypes = {inputType, inputType};
                outExpectedTypes = {inputType};
            } else {
                LOG(ERROR) << "Unsupported input tensor type for operation "
                           << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
            } else {
                NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_2));
            }
            return validateOperationOperandTypes(operands, inputCount, inputIndexes,
                                                 inExpectedTypes, outputCount, outputIndexes,
                                                 outExpectedTypes);
        }
        case ANEURALNETWORKS_IF: {
            NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
            return validateIfOperation(inputCount, inputIndexes, outputCount, outputIndexes,
                                       operands, helper)
                           ? ANEURALNETWORKS_NO_ERROR
                           : ANEURALNETWORKS_BAD_DATA;
        }
        case ANEURALNETWORKS_WHILE: {
            NN_RETURN_IF_ERROR(validateHalVersion(opType, halVersion, HalVersion::V1_3));
            return validateWhileOperation(inputCount, inputIndexes, outputCount, outputIndexes,
                                          operands, helper)
                           ? ANEURALNETWORKS_NO_ERROR
                           : ANEURALNETWORKS_BAD_DATA;
        }
        default: {
            const OperationRegistration* operationRegistration =
                    BuiltinOperationResolver::get()->findOperation(
                            static_cast<OperationType>(opType));
            if (operationRegistration == nullptr) {
                if (0 <= opType && opType < kNumberOfOperationTypes) {
                    LOG(ERROR) << getOperationName(opType) << " not registered";
                } else {
                    LOG(ERROR) << "Operation type " << opType << " out of the range [0, "
                               << kNumberOfOperationTypes << ")";
                }
                return ANEURALNETWORKS_UNEXPECTED_NULL;
            }
            if (operationRegistration->validate == nullptr) {
                LOG(ERROR) << "Incomplete operation registration: " << getOperationName(opType);
                return ANEURALNETWORKS_UNEXPECTED_NULL;
            }
            OperationValidationContext context(operationRegistration->name, inputCount,
                                               inputIndexes, outputCount, outputIndexes,
                                               operands.data(), halVersion);
            if (!operationRegistration->validate(&context)) {
                LOG(ERROR) << "Validation failed for operation " << getOperationName(opType);
                return ANEURALNETWORKS_BAD_DATA;
            }
            return ANEURALNETWORKS_NO_ERROR;
        }
    }
}

ErrorStatus convertResultCodeToErrorStatus(int resultCode) {
    switch (resultCode) {
        case ANEURALNETWORKS_NO_ERROR:
            return ErrorStatus::NONE;

        case ANEURALNETWORKS_BAD_DATA:
        case ANEURALNETWORKS_UNEXPECTED_NULL:
            return ErrorStatus::INVALID_ARGUMENT;

        case ANEURALNETWORKS_OUTPUT_INSUFFICIENT_SIZE:
            return ErrorStatus::OUTPUT_INSUFFICIENT_SIZE;

        case ANEURALNETWORKS_UNAVAILABLE_DEVICE:
            return ErrorStatus::DEVICE_UNAVAILABLE;

        case ANEURALNETWORKS_BAD_STATE:
        case ANEURALNETWORKS_INCOMPLETE:
        case ANEURALNETWORKS_OP_FAILED:
        case ANEURALNETWORKS_OUT_OF_MEMORY:
        case ANEURALNETWORKS_UNMAPPABLE:
        case ANEURALNETWORKS_DEAD_OBJECT:
            return ErrorStatus::GENERAL_FAILURE;

        case ANEURALNETWORKS_MISSED_DEADLINE_TRANSIENT:
            return ErrorStatus::MISSED_DEADLINE_TRANSIENT;
        case ANEURALNETWORKS_MISSED_DEADLINE_PERSISTENT:
            return ErrorStatus::MISSED_DEADLINE_PERSISTENT;
        case ANEURALNETWORKS_RESOURCE_EXHAUSTED_TRANSIENT:
            return ErrorStatus::RESOURCE_EXHAUSTED_TRANSIENT;
        case ANEURALNETWORKS_RESOURCE_EXHAUSTED_PERSISTENT:
            return ErrorStatus::RESOURCE_EXHAUSTED_PERSISTENT;
    }
    LOG(ERROR) << "Unknown result code " << resultCode << " mapped to ErrorStatus::GENERAL_FAILURE";
    return ErrorStatus::GENERAL_FAILURE;
}

int convertErrorStatusToResultCode(ErrorStatus status) {
    switch (status) {
        case ErrorStatus::NONE:
            return ANEURALNETWORKS_NO_ERROR;
        case ErrorStatus::DEVICE_UNAVAILABLE:
            return ANEURALNETWORKS_UNAVAILABLE_DEVICE;
        case ErrorStatus::GENERAL_FAILURE:
            return ANEURALNETWORKS_OP_FAILED;
        case ErrorStatus::OUTPUT_INSUFFICIENT_SIZE:
            return ANEURALNETWORKS_OUTPUT_INSUFFICIENT_SIZE;
        case ErrorStatus::INVALID_ARGUMENT:
            return ANEURALNETWORKS_BAD_DATA;
        case ErrorStatus::MISSED_DEADLINE_TRANSIENT:
            return ANEURALNETWORKS_MISSED_DEADLINE_TRANSIENT;
        case ErrorStatus::MISSED_DEADLINE_PERSISTENT:
            return ANEURALNETWORKS_MISSED_DEADLINE_PERSISTENT;
        case ErrorStatus::RESOURCE_EXHAUSTED_TRANSIENT:
            return ANEURALNETWORKS_RESOURCE_EXHAUSTED_TRANSIENT;
        case ErrorStatus::RESOURCE_EXHAUSTED_PERSISTENT:
            return ANEURALNETWORKS_RESOURCE_EXHAUSTED_PERSISTENT;
    }
    LOG(ERROR) << "Unknown ErrorStatus " << toString(status)
               << " mapped to ANEURALNETWORKS_OP_FAILED";
    return ANEURALNETWORKS_OP_FAILED;
}

std::tuple<int, std::vector<OutputShape>, Timing> getExecutionResult(
        ErrorStatus status, std::vector<OutputShape> outputShapes, Timing timing) {
    constexpr Timing kNoTiming = {std::numeric_limits<uint64_t>::max(),
                                  std::numeric_limits<uint64_t>::max()};
    const int n = convertErrorStatusToResultCode(status);
    if (status != ErrorStatus::NONE && status != ErrorStatus::OUTPUT_INSUFFICIENT_SIZE &&
        !outputShapes.empty()) {
        LOG(ERROR) << "The driver returned OutputShapes when it shouldn't.";
        outputShapes.clear();
    }
    if (status != ErrorStatus::NONE && timing != kNoTiming) {
        LOG(ERROR) << "The driver returned Timing when it shouldn't.";
        timing = kNoTiming;
    }
    return {n, std::move(outputShapes), timing};
}

std::optional<std::vector<uint32_t>> combineDimensions(const std::vector<uint32_t>& lhs,
                                                       const std::vector<uint32_t>& rhs) {
    if (rhs.empty()) return lhs;
    if (lhs.empty()) return rhs;
    if (lhs.size() != rhs.size()) {
        LOG(ERROR) << "Incompatible ranks: " << toString(lhs) << " and " << toString(rhs);
        return std::nullopt;
    }
    std::vector<uint32_t> combined = lhs;
    for (uint32_t i = 0; i < lhs.size(); i++) {
        if (lhs[i] == 0) {
            combined[i] = rhs[i];
        } else if (rhs[i] != 0 && lhs[i] != rhs[i]) {
            LOG(ERROR) << "Incompatible dimensions: " << toString(lhs) << " and " << toString(rhs);
            return std::nullopt;
        }
    }
    return combined;
}

// Capabilities::operandPerformance utilities.
// The field Capabilities::operandPerformance is a vector sorted by the field
// Capabilities::OperandPerformance::type.

template <HalVersion version>
hidl_vec<VersionedOperandPerformance<version>> nonExtensionOperandPerformance(
        PerformanceInfo perf) {
    using OpPerf = VersionedOperandPerformance<version>;

    // Note: range presents enumerators in declaration order, not in numerical order.
    static constexpr hidl_enum_range<VersionedOperandType<version>> kOperandTypeRange;

    std::vector<OpPerf> ret;
    ret.reserve(kOperandTypeRange.end() - kOperandTypeRange.begin());
    for (VersionedOperandType<version> type : kOperandTypeRange) {
        if (static_cast<OperandType>(type) != OperandType::SUBGRAPH) {
            ret.push_back(OpPerf{type, perf});
        }
    }
    std::sort(ret.begin(), ret.end(),
              [](const OpPerf& a, const OpPerf& b) { return a.type < b.type; });

    return ret;
}

template hal::hidl_vec<V1_2::Capabilities::OperandPerformance>
nonExtensionOperandPerformance<HalVersion::V1_2>(PerformanceInfo perf);
template hal::hidl_vec<V1_3::Capabilities::OperandPerformance>
nonExtensionOperandPerformance<HalVersion::V1_3>(PerformanceInfo perf);

template <HalVersion version>
void update(hal::hidl_vec<VersionedOperandPerformance<version>>* operandPerformance,
            VersionedOperandType<version> type, hal::PerformanceInfo perf) {
    CHECK(operandPerformance != nullptr);
    const auto it =
            std::lower_bound(operandPerformance->begin(), operandPerformance->end(), type,
                             [](const VersionedOperandPerformance<version>& perf,
                                VersionedOperandType<version> type) { return perf.type < type; });
    CHECK(it != operandPerformance->end())
            << toString(type) << " not in " << toString(*operandPerformance);
    it->info = perf;
}

void update(hidl_vec<V1_2::Capabilities::OperandPerformance>* operandPerformance,
            V1_2::OperandType type, PerformanceInfo perf) {
    update<HalVersion::V1_2>(operandPerformance, type, perf);
}
void update(hidl_vec<V1_3::Capabilities::OperandPerformance>* operandPerformance,
            V1_3::OperandType type, PerformanceInfo perf) {
    update<HalVersion::V1_3>(operandPerformance, type, perf);
}

template <HalVersion version>
PerformanceInfo lookup(const hidl_vec<VersionedOperandPerformance<version>>& operandPerformance,
                       VersionedOperandType<version> type) {
    const auto it = std::lower_bound(operandPerformance.begin(), operandPerformance.end(), type,
                                     [](const VersionedOperandPerformance<version>& perf,
                                        VersionedOperandType<version> type) {
                                         return static_cast<OperandType>(perf.type) <
                                                static_cast<OperandType>(type);
                                     });
    if (it == operandPerformance.end()) {
        LOG(WARNING) << "No PerformanceInfo for " << toString(type);
        return kNoPerformanceInfo;
    } else {
        return it->info;
    }
}

PerformanceInfo lookup(const hidl_vec<V1_2::Capabilities::OperandPerformance>& operandPerformance,
                       V1_2::OperandType type) {
    return lookup<HalVersion::V1_2>(operandPerformance, type);
}
PerformanceInfo lookup(const hidl_vec<V1_3::Capabilities::OperandPerformance>& operandPerformance,
                       V1_3::OperandType type) {
    CHECK(type != V1_3::OperandType::SUBGRAPH)
            << "Use Capabilities::ifPerformance or Capabilities::whilePerformance";
    return lookup<HalVersion::V1_3>(operandPerformance, type);
}

// Versioning

// In Android P, most data types are treated as having the same performance as TENSOR_QUANT8_ASYMM.
// This array must be in sorted order.
static const OperandType kQuantized8PerformanceConsistentWithP[] = {
        OperandType::INT32, OperandType::UINT32, OperandType::TENSOR_INT32, OperandType::OEM,
        OperandType::TENSOR_OEM_BYTE};

static bool isQuantized8PerformanceConsistentWithP(const V1_2::Capabilities& capabilities) {
    const PerformanceInfo quantized8Performance =
            lookup(capabilities.operandPerformance, V1_2::OperandType::TENSOR_QUANT8_ASYMM);
    return std::all_of(std::begin(kQuantized8PerformanceConsistentWithP),
                       std::end(kQuantized8PerformanceConsistentWithP),
                       [quantized8Performance, &capabilities](OperandType type) {
                           return quantized8Performance ==
                                  lookup(capabilities.operandPerformance,
                                         static_cast<V1_2::OperandType>(type));
                       });
}

static bool isQuantized8PerformanceConsistentWithP(const V1_3::Capabilities& capabilities) {
    const PerformanceInfo quantized8Performance =
            lookup(capabilities.operandPerformance, OperandType::TENSOR_QUANT8_ASYMM);
    return std::all_of(std::begin(kQuantized8PerformanceConsistentWithP),
                       std::end(kQuantized8PerformanceConsistentWithP),
                       [quantized8Performance, &capabilities](OperandType type) {
                           return quantized8Performance ==
                                  lookup(capabilities.operandPerformance, type);
                       });
}

static hidl_vec<V1_2::Capabilities::OperandPerformance> makeQuantized8PerformanceConsistentWithP(
        PerformanceInfo quantized8Performance) {
    hidl_vec<V1_2::Capabilities::OperandPerformance> ret(
            std::size(kQuantized8PerformanceConsistentWithP));
    std::transform(
            std::begin(kQuantized8PerformanceConsistentWithP),
            std::end(kQuantized8PerformanceConsistentWithP), ret.begin(),
            [quantized8Performance](OperandType type) -> V1_2::Capabilities::OperandPerformance {
                return {static_cast<V1_2::OperandType>(type), quantized8Performance};
            });
    return ret;
}

bool compliantWithV1_0(const V1_0::Capabilities&) {
    return true;
}

bool compliantWithV1_0(const V1_1::Capabilities& capabilities) {
    return capabilities.relaxedFloat32toFloat16Performance == capabilities.float32Performance;
}

bool compliantWithV1_0(const V1_2::Capabilities& capabilities) {
    const PerformanceInfo perfTensorFloat32 =
            lookup(capabilities.operandPerformance, V1_2::OperandType::TENSOR_FLOAT32);
    const PerformanceInfo perfFloat32 =
            lookup(capabilities.operandPerformance, V1_2::OperandType::FLOAT32);
    if (perfTensorFloat32 != perfFloat32 ||
        perfTensorFloat32 != capabilities.relaxedFloat32toFloat16PerformanceTensor ||
        perfFloat32 != capabilities.relaxedFloat32toFloat16PerformanceScalar) {
        return false;
    }

    return isQuantized8PerformanceConsistentWithP(capabilities);
}

bool compliantWithV1_0(const V1_3::Capabilities& capabilities) {
    const PerformanceInfo perfTensorFloat32 =
            lookup(capabilities.operandPerformance, OperandType::TENSOR_FLOAT32);
    const PerformanceInfo perfFloat32 =
            lookup(capabilities.operandPerformance, OperandType::FLOAT32);
    if (perfTensorFloat32 != perfFloat32 ||
        perfTensorFloat32 != capabilities.relaxedFloat32toFloat16PerformanceTensor ||
        perfFloat32 != capabilities.relaxedFloat32toFloat16PerformanceScalar) {
        return false;
    }

    return isQuantized8PerformanceConsistentWithP(capabilities);
}

bool compliantWithV1_1(const V1_0::Capabilities&) {
    return true;
}

bool compliantWithV1_1(const V1_1::Capabilities&) {
    return true;
}

bool compliantWithV1_1(const V1_2::Capabilities& capabilities) {
    if ((capabilities.relaxedFloat32toFloat16PerformanceTensor !=
         capabilities.relaxedFloat32toFloat16PerformanceScalar) ||
        (lookup(capabilities.operandPerformance, V1_2::OperandType::TENSOR_FLOAT32) !=
         lookup(capabilities.operandPerformance, V1_2::OperandType::FLOAT32))) {
        return false;
    }

    return isQuantized8PerformanceConsistentWithP(capabilities);
}

bool compliantWithV1_1(const V1_3::Capabilities& capabilities) {
    if ((capabilities.relaxedFloat32toFloat16PerformanceTensor !=
         capabilities.relaxedFloat32toFloat16PerformanceScalar) ||
        (lookup(capabilities.operandPerformance, OperandType::TENSOR_FLOAT32) !=
         lookup(capabilities.operandPerformance, OperandType::FLOAT32))) {
        return false;
    }

    return isQuantized8PerformanceConsistentWithP(capabilities);
}

bool compliantWithV1_2(const V1_0::Capabilities&) {
    return true;
}

bool compliantWithV1_2(const V1_1::Capabilities&) {
    return true;
}

bool compliantWithV1_2(const V1_2::Capabilities&) {
    return true;
}

bool compliantWithV1_2(const V1_3::Capabilities&) {
    return true;
}

bool compliantWithV1_3(const V1_0::Capabilities&) {
    return true;
}

bool compliantWithV1_3(const V1_1::Capabilities&) {
    return true;
}

bool compliantWithV1_3(const V1_2::Capabilities&) {
    return true;
}

bool compliantWithV1_3(const V1_3::Capabilities&) {
    return true;
}

V1_0::ErrorStatus convertToV1_0(V1_0::ErrorStatus status) {
    return status;
}

V1_0::ErrorStatus convertToV1_0(V1_3::ErrorStatus status) {
    switch (status) {
        case V1_3::ErrorStatus::NONE:
            return V1_0::ErrorStatus::NONE;
        case V1_3::ErrorStatus::DEVICE_UNAVAILABLE:
            return V1_0::ErrorStatus::DEVICE_UNAVAILABLE;
        case V1_3::ErrorStatus::GENERAL_FAILURE:
            return V1_0::ErrorStatus::GENERAL_FAILURE;
        case V1_3::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE:
            return V1_0::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE;
        case V1_3::ErrorStatus::INVALID_ARGUMENT:
            return V1_0::ErrorStatus::INVALID_ARGUMENT;
        case V1_3::ErrorStatus::MISSED_DEADLINE_TRANSIENT:
            return V1_0::ErrorStatus::GENERAL_FAILURE;
        case V1_3::ErrorStatus::MISSED_DEADLINE_PERSISTENT:
            return V1_0::ErrorStatus::GENERAL_FAILURE;
        case V1_3::ErrorStatus::RESOURCE_EXHAUSTED_TRANSIENT:
            return V1_0::ErrorStatus::GENERAL_FAILURE;
        case V1_3::ErrorStatus::RESOURCE_EXHAUSTED_PERSISTENT:
            return V1_0::ErrorStatus::GENERAL_FAILURE;
    }
    LOG(ERROR) << "Unknown ErrorStatus: " << toString(status) << " mapped to GENERAL_FAILURE";
    return V1_0::ErrorStatus::GENERAL_FAILURE;
}

V1_3::ErrorStatus convertToV1_3(V1_0::ErrorStatus status) {
    return static_cast<V1_3::ErrorStatus>(status);
}

V1_3::ErrorStatus convertToV1_3(V1_3::ErrorStatus status) {
    return status;
}

static V1_0::OperationType uncheckedConvertToV1_0(V1_1::OperationType type) {
    return static_cast<V1_0::OperationType>(type);
}

static V1_0::OperationType uncheckedConvertToV1_0(V1_2::OperationType type) {
    return static_cast<V1_0::OperationType>(type);
}

V1_0::OperationType uncheckedConvertToV1_0(V1_3::OperationType type) {
    return static_cast<V1_0::OperationType>(type);
}

static V1_1::OperationType convertToV1_1(V1_0::OperationType type) {
    return static_cast<V1_1::OperationType>(type);
}

static V1_1::OperationType uncheckedConvertToV1_1(V1_2::OperationType type) {
    return static_cast<V1_1::OperationType>(type);
}

V1_1::OperationType uncheckedConvertToV1_1(V1_3::OperationType type) {
    return static_cast<V1_1::OperationType>(type);
}

static V1_2::OperationType convertToV1_2(V1_0::OperationType type) {
    return static_cast<V1_2::OperationType>(type);
}

static V1_2::OperationType convertToV1_2(V1_1::OperationType type) {
    return static_cast<V1_2::OperationType>(type);
}

V1_2::OperationType uncheckedConvertToV1_2(V1_3::OperationType type) {
    return static_cast<V1_2::OperationType>(type);
}

static V1_3::OperationType convertToV1_3(V1_0::OperationType type) {
    return static_cast<V1_3::OperationType>(type);
}

static V1_3::OperationType convertToV1_3(V1_1::OperationType type) {
    return static_cast<V1_3::OperationType>(type);
}

static V1_3::OperationType convertToV1_3(V1_2::OperationType type) {
    return static_cast<V1_3::OperationType>(type);
}

V1_0::Capabilities convertToV1_0(const V1_0::Capabilities& capabilities) {
    return capabilities;
}

V1_0::Capabilities convertToV1_0(const V1_1::Capabilities& capabilities) {
    if (!compliantWithV1_0(capabilities)) {
        LOG(ERROR) << "Upcasting non-compliant capabilities " << toString(capabilities)
                   << " from V1_1::Capabilities to V1_0::Capabilities";
    }
    return {.float32Performance = capabilities.float32Performance,
            .quantized8Performance = capabilities.quantized8Performance};
}

V1_0::Capabilities convertToV1_0(const V1_2::Capabilities& capabilities) {
    if (!compliantWithV1_0(capabilities)) {
        LOG(ERROR) << "Upcasting non-compliant capabilities " << toString(capabilities)
                   << " from V1_2::Capabilities to V1_0::Capabilities";
    }
    return {.float32Performance =
                    lookup(capabilities.operandPerformance, V1_2::OperandType::TENSOR_FLOAT32),
            .quantized8Performance = lookup(capabilities.operandPerformance,
                                            V1_2::OperandType::TENSOR_QUANT8_ASYMM)};
}

V1_0::Capabilities convertToV1_0(const V1_3::Capabilities& capabilities) {
    if (!compliantWithV1_0(capabilities)) {
        LOG(ERROR) << "Upcasting non-compliant capabilities " << toString(capabilities)
                   << " from V1_3::Capabilities to V1_0::Capabilities";
    }
    return {.float32Performance =
                    lookup(capabilities.operandPerformance, OperandType::TENSOR_FLOAT32),
            .quantized8Performance =
                    lookup(capabilities.operandPerformance, OperandType::TENSOR_QUANT8_ASYMM)};
}

V1_1::Capabilities convertToV1_1(const V1_0::Capabilities& capabilities) {
    return {.float32Performance = capabilities.float32Performance,
            .quantized8Performance = capabilities.quantized8Performance,
            .relaxedFloat32toFloat16Performance = capabilities.float32Performance};
}

V1_1::Capabilities convertToV1_1(const V1_1::Capabilities& capabilities) {
    return capabilities;
}

V1_1::Capabilities convertToV1_1(const V1_2::Capabilities& capabilities) {
    if (!compliantWithV1_1(capabilities)) {
        LOG(ERROR) << "Upcasting non-compliant capabilities " << toString(capabilities)
                   << " from V1_2::Capabilities to V1_1::Capabilities";
    }
    return {.float32Performance =
                    lookup(capabilities.operandPerformance, V1_2::OperandType::TENSOR_FLOAT32),
            .quantized8Performance =
                    lookup(capabilities.operandPerformance, V1_2::OperandType::TENSOR_QUANT8_ASYMM),
            .relaxedFloat32toFloat16Performance =
                    capabilities.relaxedFloat32toFloat16PerformanceTensor};
}

V1_1::Capabilities convertToV1_1(const V1_3::Capabilities& capabilities) {
    if (!compliantWithV1_1(capabilities)) {
        LOG(ERROR) << "Upcasting non-compliant capabilities " << toString(capabilities)
                   << " from V1_3::Capabilities to V1_1::Capabilities";
    }
    return {.float32Performance =
                    lookup(capabilities.operandPerformance, OperandType::TENSOR_FLOAT32),
            .quantized8Performance =
                    lookup(capabilities.operandPerformance, OperandType::TENSOR_QUANT8_ASYMM),
            .relaxedFloat32toFloat16Performance =
                    capabilities.relaxedFloat32toFloat16PerformanceTensor};
}

V1_2::Capabilities convertToV1_2(const V1_0::Capabilities& capabilities) {
    V1_2::Capabilities ret = {
            .relaxedFloat32toFloat16PerformanceScalar = capabilities.float32Performance,
            .relaxedFloat32toFloat16PerformanceTensor = capabilities.float32Performance,
            .operandPerformance =
                    makeQuantized8PerformanceConsistentWithP(capabilities.quantized8Performance)};
    auto& opPerf = ret.operandPerformance;
    opPerf.resize(opPerf.size() + 2);
    opPerf[opPerf.size() - 2] = {V1_2::OperandType::TENSOR_FLOAT32,
                                 capabilities.float32Performance};
    opPerf[opPerf.size() - 1] = {V1_2::OperandType::FLOAT32, capabilities.float32Performance};
    using OperandPerformance = V1_2::Capabilities::OperandPerformance;
    std::sort(opPerf.begin(), opPerf.end(),
              [](const OperandPerformance& a, const OperandPerformance& b) {
                  return a.type < b.type;
              });
    return ret;
}

V1_2::Capabilities convertToV1_2(const V1_1::Capabilities& capabilities) {
    V1_2::Capabilities ret = {.relaxedFloat32toFloat16PerformanceScalar =
                                      capabilities.relaxedFloat32toFloat16Performance,
                              .relaxedFloat32toFloat16PerformanceTensor =
                                      capabilities.relaxedFloat32toFloat16Performance,
                              .operandPerformance = makeQuantized8PerformanceConsistentWithP(
                                      capabilities.quantized8Performance)};
    auto& opPerf = ret.operandPerformance;
    opPerf.resize(opPerf.size() + 2);
    opPerf[opPerf.size() - 2] = {V1_2::OperandType::TENSOR_FLOAT32,
                                 capabilities.float32Performance};
    opPerf[opPerf.size() - 1] = {V1_2::OperandType::FLOAT32, capabilities.float32Performance};
    using OperandPerformance = V1_2::Capabilities::OperandPerformance;
    std::sort(opPerf.begin(), opPerf.end(),
              [](const OperandPerformance& a, const OperandPerformance& b) {
                  return a.type < b.type;
              });
    return ret;
}

V1_2::Capabilities convertToV1_2(const V1_2::Capabilities& capabilities) {
    return capabilities;
}

V1_2::Capabilities convertToV1_2(const V1_3::Capabilities& capabilities) {
    V1_2::Capabilities ret = {
            .relaxedFloat32toFloat16PerformanceScalar =
                    capabilities.relaxedFloat32toFloat16PerformanceScalar,
            .relaxedFloat32toFloat16PerformanceTensor =
                    capabilities.relaxedFloat32toFloat16PerformanceTensor,
    };
    const auto& inputOpPerf = capabilities.operandPerformance;
    hidl_vec<V1_3::Capabilities::OperandPerformance> opPerfSupported;
    opPerfSupported.resize(inputOpPerf.size());
    auto last =
            std::copy_if(inputOpPerf.begin(), inputOpPerf.end(), opPerfSupported.begin(),
                         [](V1_3::Capabilities::OperandPerformance opPerf) {
                             return validOperandType(static_cast<V1_2::OperandType>(opPerf.type));
                         });
    opPerfSupported.resize(std::distance(opPerfSupported.begin(), last));

    auto& convertedOpPerf = ret.operandPerformance;
    convertedOpPerf.resize(opPerfSupported.size());
    std::transform(opPerfSupported.begin(), opPerfSupported.end(), convertedOpPerf.begin(),
                   [](V1_3::Capabilities::OperandPerformance opPerf) {
                       return V1_2::Capabilities::OperandPerformance{
                               static_cast<V1_2::OperandType>(opPerf.type), opPerf.info};
                   });
    return ret;
}

V1_3::Capabilities convertToV1_3(const V1_0::Capabilities& capabilities) {
    return convertToV1_3(convertToV1_2(capabilities));
}

V1_3::Capabilities convertToV1_3(const V1_1::Capabilities& capabilities) {
    return convertToV1_3(convertToV1_2(capabilities));
}

V1_3::Capabilities convertToV1_3(const V1_2::Capabilities& capabilities) {
    V1_3::Capabilities ret = {
            .relaxedFloat32toFloat16PerformanceScalar =
                    capabilities.relaxedFloat32toFloat16PerformanceScalar,
            .relaxedFloat32toFloat16PerformanceTensor =
                    capabilities.relaxedFloat32toFloat16PerformanceTensor,
            .ifPerformance = kNoPerformanceInfo,
            .whilePerformance = kNoPerformanceInfo,
    };
    auto& opPerf = ret.operandPerformance;
    opPerf.resize(capabilities.operandPerformance.size());
    std::transform(capabilities.operandPerformance.begin(), capabilities.operandPerformance.end(),
                   opPerf.begin(), [](V1_2::Capabilities::OperandPerformance opPerf) {
                       return V1_3::Capabilities::OperandPerformance{
                               static_cast<V1_3::OperandType>(opPerf.type), opPerf.info};
                   });
    return ret;
}

V1_3::Capabilities convertToV1_3(const V1_3::Capabilities& capabilities) {
    return capabilities;
}

static V1_0::Operation uncheckedConvertToV1_0(const V1_1::Operation& operation) {
    return {.type = uncheckedConvertToV1_0(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static V1_1::Operation convertToV1_1(const V1_0::Operation& operation) {
    return {.type = convertToV1_1(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static hidl_vec<V1_0::Operation> uncheckedConvertToV1_0(
        const hidl_vec<V1_1::Operation>& operations) {
    hidl_vec<V1_0::Operation> result(operations.size());
    std::transform(
            operations.begin(), operations.end(), result.begin(),
            [](const V1_1::Operation& operation) { return uncheckedConvertToV1_0(operation); });
    return result;
}

static hidl_vec<V1_1::Operation> convertToV1_1(const hidl_vec<V1_0::Operation>& operations) {
    hidl_vec<V1_1::Operation> result(operations.size());
    std::transform(operations.begin(), operations.end(), result.begin(),
                   [](const V1_0::Operation& operation) { return convertToV1_1(operation); });
    return result;
}

bool compliantWithV1_0(const V1_3::Operand& operand) {
    return validOperandType(static_cast<V1_0::OperandType>(operand.type)) &&
           (nonExtensionOperandTypeIsScalar(static_cast<int>(operand.type)) ||
            operand.dimensions.size() != 0) &&
           compliantWithV1_0(operand.lifetime);
}

bool compliantWithV1_2(const V1_3::Operand& operand) {
    return validOperandType(static_cast<V1_2::OperandType>(operand.type)) &&
           compliantWithV1_0(operand.lifetime);
}

bool compliantWithV1_3(const V1_3::Operand& operand) {
    return true;
}

static bool compliantWith(HalVersion version, const V1_3::Model& model,
                          std::set<uint32_t>* noncompliantOperations) {
    // A boolean vector indicating whether each pool is compliant with the target HAL version.
    std::vector<bool> isPoolCompliant(model.pools.size(), false);
    std::transform(model.pools.begin(), model.pools.end(), isPoolCompliant.begin(),
                   [version](const hidl_memory& pool) { return validatePool(pool, version); });

    // A boolean vector indicating whether each operand is compliant with the target HAL version.
    std::vector<bool> isOperandCompliant(model.main.operands.size(), false);
    std::transform(model.main.operands.begin(), model.main.operands.end(),
                   isOperandCompliant.begin(), [&isPoolCompliant, version](const Operand& op) {
                       bool is_operand_compliant = false;
                       switch (version) {
                           case HalVersion::UNKNOWN:
                               is_operand_compliant = false;
                               break;
                           case HalVersion::V1_0:
                               is_operand_compliant = compliantWithV1_0(op);
                               break;
                           case HalVersion::V1_1:
                               // There is no V1_1::Operand -- both V1_0::Model
                               // and V1_1::Model use V1_0::Operand.
                               is_operand_compliant = compliantWithV1_0(op);
                               break;
                           case HalVersion::V1_2:
                               is_operand_compliant = compliantWithV1_2(op);
                               break;
                           case HalVersion::V1_3:
                               is_operand_compliant = compliantWithV1_3(op);
                               break;
                       }
                       return is_operand_compliant &&
                              !(op.lifetime == OperandLifeTime::CONSTANT_REFERENCE &&
                                !isPoolCompliant[op.location.poolIndex]);
                   });

    auto allOperandsCompliant = [&isOperandCompliant](const hidl_vec<uint32_t>& indices) {
        return std::all_of(
                indices.begin(), indices.end(),
                [&isOperandCompliant](const uint32_t ind) { return isOperandCompliant[ind]; });
    };

    auto localValidateOperation = [&model, version, &allOperandsCompliant](const Operation& op) {
        if (!allOperandsCompliant(op.inputs) || !allOperandsCompliant(op.outputs)) return false;
        int error = validateOperation(
                static_cast<int32_t>(op.type), op.inputs.size(),
                op.inputs.size() > 0 ? op.inputs.data() : nullptr, op.outputs.size(),
                op.outputs.size() > 0 ? op.outputs.data() : nullptr, model.main.operands, version);
        return error == ANEURALNETWORKS_NO_ERROR;
    };

    if (noncompliantOperations) {
        CHECK(noncompliantOperations->empty());
        for (uint32_t idx = 0; idx < model.main.operations.size(); ++idx) {
            if (!localValidateOperation(model.main.operations[idx])) {
                noncompliantOperations->insert(idx);
            }
        }
        return noncompliantOperations->empty();
    } else {
        return std::all_of(model.main.operations.begin(), model.main.operations.end(),
                           localValidateOperation);
    }
}

bool compliantWithV1_0(const V1_0::Model& model) {
    return true;
}

bool compliantWithV1_0(const V1_1::Model& model) {
    // In addition to new enumeration values being introduced in V1_1::Model, a
    // new flag was introduced to indicate whether or not float32 data can be
    // calculated using float16 units. This 'relaxComputationFloat32toFloat16'
    // flag is not relevant in whether a V1_1::Model is compliant with a
    // V1_0::Model because all 1.0 drivers require strict calculation by default
    // in the P NN runtime. Even if fp16 calculations are allowed, they can
    // still be computed by a strict fp32 driver.
    return std::all_of(
            model.operations.begin(), model.operations.end(), [&model](const V1_1::Operation& op) {
                int error = validateOperation(static_cast<int32_t>(op.type), op.inputs.size(),
                                              op.inputs.size() > 0 ? op.inputs.data() : nullptr,
                                              op.outputs.size(),
                                              op.outputs.size() > 0 ? op.outputs.data() : nullptr,
                                              convertToV1_3(model.operands), HalVersion::V1_0);
                return error == ANEURALNETWORKS_NO_ERROR;
            });
}

bool compliantWithV1_0(const V1_2::Model& model, std::set<uint32_t>* noncompliantOperations) {
    return compliantWith(HalVersion::V1_0, convertToV1_3(model), noncompliantOperations);
}

bool compliantWithV1_0(const V1_3::Model& model, std::set<uint32_t>* noncompliantOperations) {
    return compliantWith(HalVersion::V1_0, model, noncompliantOperations);
}

bool compliantWithV1_1(const V1_0::Model&) {
    return true;
}

bool compliantWithV1_1(const V1_1::Model&) {
    return true;
}

bool compliantWithV1_1(const V1_2::Model& model, std::set<uint32_t>* noncompliantOperations) {
    return compliantWith(HalVersion::V1_1, convertToV1_3(model), noncompliantOperations);
}

bool compliantWithV1_1(const V1_3::Model& model, std::set<uint32_t>* noncompliantOperations) {
    return compliantWith(HalVersion::V1_1, model, noncompliantOperations);
}

bool compliantWithV1_2(const V1_0::Model&) {
    return true;
}

bool compliantWithV1_2(const V1_1::Model&) {
    return true;
}

bool compliantWithV1_2(const V1_2::Model&, std::set<uint32_t>* noncompliantOperations) {
    return true;
}

bool compliantWithV1_2(const V1_3::Model& model, std::set<uint32_t>* noncompliantOperations) {
    return compliantWith(HalVersion::V1_2, model, noncompliantOperations);
}

static V1_0::Operation uncheckedConvertToV1_0(const V1_2::Operation& operation) {
    return {.type = uncheckedConvertToV1_0(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static V1_0::Operation uncheckedConvertToV1_0(const V1_3::Operation& operation) {
    return {.type = uncheckedConvertToV1_0(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static V1_1::Operation uncheckedConvertToV1_1(const V1_2::Operation& operation) {
    return {.type = uncheckedConvertToV1_1(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static V1_1::Operation uncheckedConvertToV1_1(const V1_3::Operation& operation) {
    return {.type = uncheckedConvertToV1_1(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static V1_2::Operation convertToV1_2(const V1_0::Operation& operation) {
    return {.type = convertToV1_2(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static V1_2::Operation convertToV1_2(const V1_1::Operation& operation) {
    return {.type = convertToV1_2(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static V1_2::Operation uncheckedConvertToV1_2(const V1_3::Operation& operation) {
    return {.type = uncheckedConvertToV1_2(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static V1_3::Operation convertToV1_3(const V1_0::Operation& operation) {
    return {.type = convertToV1_3(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static V1_3::Operation convertToV1_3(const V1_1::Operation& operation) {
    return {.type = convertToV1_3(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static V1_3::Operation convertToV1_3(const V1_2::Operation& operation) {
    return {.type = convertToV1_3(operation.type),
            .inputs = operation.inputs,
            .outputs = operation.outputs};
}

static hidl_vec<V1_0::Operation> uncheckedConvertToV1_0(
        const hidl_vec<V1_3::Operation>& operations) {
    hidl_vec<V1_0::Operation> result(operations.size());
    std::transform(
            operations.begin(), operations.end(), result.begin(),
            [](const V1_3::Operation& operation) { return uncheckedConvertToV1_0(operation); });
    return result;
}

static hidl_vec<V1_0::Operation> uncheckedConvertToV1_0(
        const hidl_vec<V1_2::Operation>& operations) {
    hidl_vec<V1_0::Operation> result(operations.size());
    std::transform(
            operations.begin(), operations.end(), result.begin(),
            [](const V1_2::Operation& operation) { return uncheckedConvertToV1_0(operation); });
    return result;
}

static hidl_vec<V1_2::Operation> uncheckedConvertToV1_2(
        const hidl_vec<V1_3::Operation>& operations) {
    hidl_vec<V1_2::Operation> result(operations.size());
    std::transform(
            operations.begin(), operations.end(), result.begin(),
            [](const V1_3::Operation& operation) { return uncheckedConvertToV1_2(operation); });
    return result;
}

static hidl_vec<V1_1::Operation> uncheckedConvertToV1_1(
        const hidl_vec<V1_2::Operation>& operations) {
    hidl_vec<V1_1::Operation> result(operations.size());
    std::transform(
            operations.begin(), operations.end(), result.begin(),
            [](const V1_2::Operation& operation) { return uncheckedConvertToV1_1(operation); });
    return result;
}

static hidl_vec<V1_1::Operation> uncheckedConvertToV1_1(
        const hidl_vec<V1_3::Operation>& operations) {
    hidl_vec<V1_1::Operation> result(operations.size());
    std::transform(
            operations.begin(), operations.end(), result.begin(),
            [](const V1_3::Operation& operation) { return uncheckedConvertToV1_1(operation); });
    return result;
}

static hidl_vec<V1_2::Operation> convertToV1_2(const hidl_vec<V1_0::Operation>& operations) {
    hidl_vec<V1_2::Operation> result(operations.size());
    std::transform(operations.begin(), operations.end(), result.begin(),
                   [](const V1_0::Operation& operation) { return convertToV1_2(operation); });
    return result;
}

static hidl_vec<V1_2::Operation> convertToV1_2(const hidl_vec<V1_1::Operation>& operations) {
    hidl_vec<V1_2::Operation> result(operations.size());
    std::transform(operations.begin(), operations.end(), result.begin(),
                   [](const V1_1::Operation& operation) { return convertToV1_2(operation); });
    return result;
}

static hidl_vec<V1_3::Operation> convertToV1_3(const hidl_vec<V1_0::Operation>& operations) {
    hidl_vec<V1_3::Operation> result(operations.size());
    std::transform(operations.begin(), operations.end(), result.begin(),
                   [](const V1_0::Operation& operation) { return convertToV1_3(operation); });
    return result;
}

static hidl_vec<V1_3::Operation> convertToV1_3(const hidl_vec<V1_1::Operation>& operations) {
    hidl_vec<V1_3::Operation> result(operations.size());
    std::transform(operations.begin(), operations.end(), result.begin(),
                   [](const V1_1::Operation& operation) { return convertToV1_3(operation); });
    return result;
}

static hidl_vec<V1_3::Operation> convertToV1_3(const hidl_vec<V1_2::Operation>& operations) {
    hidl_vec<V1_3::Operation> result(operations.size());
    std::transform(operations.begin(), operations.end(), result.begin(),
                   [](const V1_2::Operation& operation) { return convertToV1_3(operation); });
    return result;
}

static bool compliantWithV1_0(const V1_2::OperandType& operandType) {
    return validOperandType(static_cast<V1_0::OperandType>(operandType));
}

static bool compliantWithV1_0(const V1_3::OperandType& operandType) {
    return validOperandType(static_cast<V1_0::OperandType>(operandType));
}

static bool compliantWithV1_2(const V1_3::OperandType& operandType) {
    return validOperandType(static_cast<V1_2::OperandType>(operandType));
}

V1_0::OperandType convertToV1_0(const V1_2::OperandType& operandType) {
    if (!compliantWithV1_0(operandType)) {
        LOG(ERROR) << "Upcasting non-compliant operand type " << toString(operandType)
                   << " from V1_2::OperandType to V1_0::OperandType";
    }
    return static_cast<V1_0::OperandType>(operandType);
}

V1_2::OperandType convertToV1_2(const V1_0::OperandType& operandType) {
    return static_cast<V1_2::OperandType>(operandType);
}

V1_2::OperandType convertToV1_2(const V1_3::OperandType& operandType) {
    if (!compliantWithV1_2(operandType)) {
        LOG(ERROR) << "Upcasting non-compliant operand type " << toString(operandType)
                   << " from V1_3::OperandType to V1_2::OperandType";
    }
    return static_cast<V1_2::OperandType>(operandType);
}

V1_0::OperandType convertToV1_0(const V1_3::OperandType& operandType) {
    if (!compliantWithV1_0(operandType)) {
        LOG(ERROR) << "Upcasting non-compliant operand type " << toString(operandType)
                   << " from V1_3::Operand to V1_0::Operand";
    }
    return static_cast<V1_0::OperandType>(operandType);
}

bool compliantWithV1_0(hal::V1_0::OperandLifeTime lifetime) {
    return true;
}

bool compliantWithV1_0(hal::V1_3::OperandLifeTime lifetime) {
    return lifetime != V1_3::OperandLifeTime::SUBGRAPH;
}

bool compliantWithV1_3(hal::V1_0::OperandLifeTime lifetime) {
    return true;
}

bool compliantWithV1_3(hal::V1_3::OperandLifeTime lifetime) {
    return true;
}

V1_0::OperandLifeTime convertToV1_0(V1_0::OperandLifeTime lifetime) {
    return lifetime;
}

V1_0::OperandLifeTime convertToV1_0(V1_3::OperandLifeTime lifetime) {
    if (!compliantWithV1_0(lifetime)) {
        LOG(ERROR) << "Upcasting non-compliant lifetime " << toString(lifetime)
                   << " from V1_3 to V1_0";
    }
    return static_cast<V1_0::OperandLifeTime>(lifetime);
}

V1_3::OperandLifeTime convertToV1_3(V1_0::OperandLifeTime lifetime) {
    return static_cast<V1_3::OperandLifeTime>(lifetime);
}

V1_3::OperandLifeTime convertToV1_3(V1_3::OperandLifeTime lifetime) {
    return lifetime;
}

V1_0::Operand convertToV1_0(const V1_2::Operand& operand) {
    return {.type = convertToV1_0(operand.type),
            .dimensions = operand.dimensions,
            .numberOfConsumers = operand.numberOfConsumers,
            .scale = operand.scale,
            .zeroPoint = operand.zeroPoint,
            .lifetime = convertToV1_0(operand.lifetime),
            .location = operand.location};
}

V1_0::Operand convertToV1_0(const V1_3::Operand& operand) {
    return {.type = convertToV1_0(operand.type),
            .dimensions = operand.dimensions,
            .numberOfConsumers = operand.numberOfConsumers,
            .scale = operand.scale,
            .zeroPoint = operand.zeroPoint,
            .lifetime = convertToV1_0(operand.lifetime),
            .location = operand.location};
}

V1_2::Operand convertToV1_2(const V1_0::Operand& operand) {
    return {.type = convertToV1_2(operand.type),
            .dimensions = operand.dimensions,
            .numberOfConsumers = operand.numberOfConsumers,
            .scale = operand.scale,
            .zeroPoint = operand.zeroPoint,
            .lifetime = operand.lifetime,
            .location = operand.location};
}

V1_2::Operand convertToV1_2(const V1_3::Operand& operand) {
    return {.type = convertToV1_2(operand.type),
            .dimensions = operand.dimensions,
            .numberOfConsumers = operand.numberOfConsumers,
            .scale = operand.scale,
            .zeroPoint = operand.zeroPoint,
            .lifetime = static_cast<V1_0::OperandLifeTime>(operand.lifetime),
            .location = operand.location,
            .extraParams = operand.extraParams};
}

V1_3::Operand convertToV1_3(const V1_0::Operand& operand) {
    return {.type = static_cast<V1_3::OperandType>(operand.type),
            .dimensions = operand.dimensions,
            .numberOfConsumers = operand.numberOfConsumers,
            .scale = operand.scale,
            .zeroPoint = operand.zeroPoint,
            .lifetime = convertToV1_3(operand.lifetime),
            .location = operand.location};
}

V1_3::Operand convertToV1_3(const V1_2::Operand& operand) {
    return {.type = static_cast<V1_3::OperandType>(operand.type),
            .dimensions = operand.dimensions,
            .numberOfConsumers = operand.numberOfConsumers,
            .scale = operand.scale,
            .zeroPoint = operand.zeroPoint,
            .lifetime = convertToV1_3(operand.lifetime),
            .location = operand.location,
            .extraParams = operand.extraParams};
}

V1_3::Operand convertToV1_3(const V1_3::Operand& operand) {
    return operand;
}

hidl_vec<V1_0::Operand> convertToV1_0(const hidl_vec<V1_0::Operand>& operands) {
    return operands;
}

hidl_vec<V1_0::Operand> convertToV1_0(const hidl_vec<V1_2::Operand>& operands) {
    hidl_vec<V1_0::Operand> result(operands.size());
    std::transform(operands.begin(), operands.end(), result.begin(),
                   [](const V1_2::Operand& operand) { return convertToV1_0(operand); });
    return result;
}

hidl_vec<V1_0::Operand> convertToV1_0(const hidl_vec<V1_3::Operand>& operands) {
    hidl_vec<V1_0::Operand> result(operands.size());
    std::transform(operands.begin(), operands.end(), result.begin(),
                   [](const V1_3::Operand& operand) { return convertToV1_0(operand); });
    return result;
}

hidl_vec<V1_2::Operand> convertToV1_2(const hidl_vec<V1_0::Operand>& operands) {
    hidl_vec<V1_2::Operand> result(operands.size());
    std::transform(operands.begin(), operands.end(), result.begin(),
                   [](const V1_0::Operand& operand) { return convertToV1_2(operand); });
    return result;
}

hidl_vec<V1_2::Operand> convertToV1_2(const hidl_vec<V1_2::Operand>& operands) {
    return operands;
}

hidl_vec<V1_2::Operand> convertToV1_2(const hidl_vec<V1_3::Operand>& operands) {
    hidl_vec<V1_2::Operand> result(operands.size());
    std::transform(operands.begin(), operands.end(), result.begin(),
                   [](const V1_3::Operand& operand) { return convertToV1_2(operand); });
    return result;
}

hidl_vec<V1_3::Operand> convertToV1_3(const hidl_vec<V1_0::Operand>& operands) {
    hidl_vec<V1_3::Operand> result(operands.size());
    std::transform(operands.begin(), operands.end(), result.begin(),
                   [](const V1_0::Operand& operand) { return convertToV1_3(operand); });
    return result;
}

hidl_vec<V1_3::Operand> convertToV1_3(const hidl_vec<V1_2::Operand>& operands) {
    hidl_vec<V1_3::Operand> result(operands.size());
    std::transform(operands.begin(), operands.end(), result.begin(),
                   [](const V1_2::Operand& operand) { return convertToV1_3(operand); });
    return result;
}

hidl_vec<V1_3::Operand> convertToV1_3(const hidl_vec<V1_3::Operand>& operands) {
    return operands;
}

V1_0::Model convertToV1_0(const V1_0::Model& model) {
    return model;
}

V1_0::Model convertToV1_0(const V1_1::Model& model) {
    if (!compliantWithV1_0(model)) {
        LOG(ERROR) << "Upcasting non-compliant model " << SHOW_IF_DEBUG(toString(model))
                   << " from V1_1::Model to V1_0::Model";
    }
    return {.operands = model.operands,
            .operations = uncheckedConvertToV1_0(model.operations),
            .inputIndexes = model.inputIndexes,
            .outputIndexes = model.outputIndexes,
            .operandValues = model.operandValues,
            .pools = model.pools};
}

V1_0::Model convertToV1_0(const V1_2::Model& model) {
    if (!compliantWithV1_0(model)) {
        LOG(ERROR) << "Upcasting non-compliant model " << SHOW_IF_DEBUG(toString(model))
                   << " from V1_2::Model to V1_0::Model";
    }
    return {.operands = convertToV1_0(model.operands),
            .operations = uncheckedConvertToV1_0(model.operations),
            .inputIndexes = model.inputIndexes,
            .outputIndexes = model.outputIndexes,
            .operandValues = model.operandValues,
            .pools = model.pools};
}

V1_0::Model convertToV1_0(const V1_3::Model& model) {
    if (!compliantWithV1_0(model)) {
        LOG(ERROR) << "Upcasting non-compliant model " << SHOW_IF_DEBUG(toString(model))
                   << " from V1_3::Model to V1_0::Model";
    }
    return {.operands = convertToV1_0(model.main.operands),
            .operations = uncheckedConvertToV1_0(model.main.operations),
            .inputIndexes = model.main.inputIndexes,
            .outputIndexes = model.main.outputIndexes,
            .operandValues = model.operandValues,
            .pools = model.pools};
}

V1_1::Model convertToV1_1(const V1_0::Model& model) {
    return {.operands = model.operands,
            .operations = convertToV1_1(model.operations),
            .inputIndexes = model.inputIndexes,
            .outputIndexes = model.outputIndexes,
            .operandValues = model.operandValues,
            .pools = model.pools,
            .relaxComputationFloat32toFloat16 = false};
}

V1_1::Model convertToV1_1(const V1_1::Model& model) {
    return model;
}

V1_1::Model convertToV1_1(const V1_2::Model& model) {
    if (!compliantWithV1_1(model)) {
        LOG(ERROR) << "Upcasting non-compliant model " << SHOW_IF_DEBUG(toString(model))
                   << " from V1_2::Model to V1_1::Model";
    }
    return {.operands = convertToV1_0(model.operands),  // Operands in 1.1 and 1.0 are identical.
            .operations = uncheckedConvertToV1_1(model.operations),
            .inputIndexes = model.inputIndexes,
            .outputIndexes = model.outputIndexes,
            .operandValues = model.operandValues,
            .pools = model.pools,
            .relaxComputationFloat32toFloat16 = model.relaxComputationFloat32toFloat16};
}

V1_1::Model convertToV1_1(const V1_3::Model& model) {
    if (!compliantWithV1_1(model)) {
        LOG(ERROR) << "Upcasting non-compliant model " << SHOW_IF_DEBUG(toString(model))
                   << " from V1_3::Model to V1_1::Model";
    }
    return {// Operands in 1.1 and 1.0 are identical.
            .operands = convertToV1_0(model.main.operands),
            .operations = uncheckedConvertToV1_1(model.main.operations),
            .inputIndexes = model.main.inputIndexes,
            .outputIndexes = model.main.outputIndexes,
            .operandValues = model.operandValues,
            .pools = model.pools,
            .relaxComputationFloat32toFloat16 = model.relaxComputationFloat32toFloat16};
}

V1_2::Model convertToV1_2(const V1_0::Model& model) {
    return {.operands = convertToV1_2(model.operands),
            .operations = convertToV1_2(model.operations),
            .inputIndexes = model.inputIndexes,
            .outputIndexes = model.outputIndexes,
            .operandValues = model.operandValues,
            .pools = model.pools,
            .relaxComputationFloat32toFloat16 = false};
}

V1_2::Model convertToV1_2(const V1_1::Model& model) {
    return {.operands = convertToV1_2(model.operands),
            .operations = convertToV1_2(model.operations),
            .inputIndexes = model.inputIndexes,
            .outputIndexes = model.outputIndexes,
            .operandValues = model.operandValues,
            .pools = model.pools,
            .relaxComputationFloat32toFloat16 = model.relaxComputationFloat32toFloat16};
}

V1_2::Model convertToV1_2(const V1_2::Model& model) {
    return model;
}

V1_2::Model convertToV1_2(const V1_3::Model& model) {
    if (!compliantWithV1_2(model)) {
        LOG(ERROR) << "Upcasting non-compliant model " << SHOW_IF_DEBUG(toString(model))
                   << " from V1_3::Model to V1_2::Model";
    }
    return {.operands = convertToV1_2(model.main.operands),
            .operations = uncheckedConvertToV1_2(model.main.operations),
            .inputIndexes = model.main.inputIndexes,
            .outputIndexes = model.main.outputIndexes,
            .operandValues = model.operandValues,
            .pools = model.pools,
            .relaxComputationFloat32toFloat16 = model.relaxComputationFloat32toFloat16,
            .extensionNameToPrefix = model.extensionNameToPrefix};
}

V1_3::Model convertToV1_3(const V1_0::Model& model) {
    return {.main = {.operands = convertToV1_3(model.operands),
                     .operations = convertToV1_3(model.operations),
                     .inputIndexes = model.inputIndexes,
                     .outputIndexes = model.outputIndexes},
            .operandValues = model.operandValues,
            .pools = model.pools,
            .relaxComputationFloat32toFloat16 = false};
}

V1_3::Model convertToV1_3(const V1_1::Model& model) {
    return {.main = {.operands = convertToV1_3(model.operands),
                     .operations = convertToV1_3(model.operations),
                     .inputIndexes = model.inputIndexes,
                     .outputIndexes = model.outputIndexes},
            .operandValues = model.operandValues,
            .pools = model.pools,
            .relaxComputationFloat32toFloat16 = model.relaxComputationFloat32toFloat16};
}

V1_3::Model convertToV1_3(const V1_2::Model& model) {
    return {.main = {.operands = convertToV1_3(model.operands),
                     .operations = convertToV1_3(model.operations),
                     .inputIndexes = model.inputIndexes,
                     .outputIndexes = model.outputIndexes},
            .operandValues = model.operandValues,
            .pools = model.pools,
            .relaxComputationFloat32toFloat16 = model.relaxComputationFloat32toFloat16,
            .extensionNameToPrefix = model.extensionNameToPrefix};
}

V1_3::Model convertToV1_3(const V1_3::Model& model) {
    return model;
}

bool compliantWithV1_0(const V1_0::Request& request) {
    return true;
}

bool compliantWithV1_0(const V1_3::Request& request) {
    return std::all_of(request.pools.begin(), request.pools.end(), [](const auto& pool) {
        if (pool.getDiscriminator() != V1_3::Request::MemoryPool::hidl_discriminator::hidlMemory) {
            return false;
        }
        const auto& name = pool.hidlMemory().name();
        return name == "ashmem" || name == "mmap_fd";
    });
}

bool compliantWithV1_2(const V1_3::Request& request) {
    return std::all_of(request.pools.begin(), request.pools.end(), [](const auto& pool) {
        if (pool.getDiscriminator() != V1_3::Request::MemoryPool::hidl_discriminator::hidlMemory) {
            return false;
        }
        const auto& name = pool.hidlMemory().name();
        return name == "ashmem" || name == "mmap_fd" || name == "hardware_buffer_blob" ||
               name == "hardware_buffer";
    });
}

static hidl_memory convertToV1_0(const V1_3::Request::MemoryPool& pool) {
    switch (pool.getDiscriminator()) {
        case V1_3::Request::MemoryPool::hidl_discriminator::hidlMemory:
            return pool.hidlMemory();
        case V1_3::Request::MemoryPool::hidl_discriminator::token:
            return hidl_memory{};
    }
}

static V1_3::Request::MemoryPool convertToV1_3(const hidl_memory& pool) {
    V1_3::Request::MemoryPool ret;
    ret.hidlMemory(pool);
    return ret;
}

V1_0::Request convertToV1_0(const V1_0::Request& request) {
    return request;
}

static V1_0::Request uncheckedConvertToV1_0(const V1_3::Request& request) {
    hidl_vec<hidl_memory> pools(request.pools.size());
    std::transform(request.pools.begin(), request.pools.end(), pools.begin(),
                   [](const auto& pool) { return convertToV1_0(pool); });
    return {.inputs = request.inputs, .outputs = request.outputs, .pools = std::move(pools)};
}

V1_0::Request convertToV1_0(const V1_3::Request& request) {
    if (!compliantWithV1_0(request)) {
        LOG(ERROR) << "Upcasting non-compliant request " << SHOW_IF_DEBUG(toString(request))
                   << " from V1_3::Request to V1_0::Request of version 1.0";
    }
    return uncheckedConvertToV1_0(request);
}

V1_0::Request convertToV1_2(const V1_3::Request& request) {
    if (!compliantWithV1_2(request)) {
        LOG(ERROR) << "Upcasting non-compliant request " << SHOW_IF_DEBUG(toString(request))
                   << " from V1_3::Request to V1_0::Request of version 1.2";
    }
    return uncheckedConvertToV1_0(request);
}

V1_3::Request convertToV1_3(const V1_0::Request& request) {
    hidl_vec<V1_3::Request::MemoryPool> pools(request.pools.size());
    std::transform(request.pools.begin(), request.pools.end(), pools.begin(),
                   [](const auto& pool) { return convertToV1_3(pool); });
    return {.inputs = request.inputs, .outputs = request.outputs, .pools = std::move(pools)};
}

V1_3::Request convertToV1_3(const V1_3::Request& request) {
    return request;
}

FenceState syncWait(int fd, int timeout) {
    // This implementation is directly based on the ::sync_wait() implementation.

    struct pollfd fds;
    int ret;

    if (fd < 0) {
        errno = EINVAL;
        return FenceState::UNKNOWN;
    }

    fds.fd = fd;
    fds.events = POLLIN;

    do {
        ret = poll(&fds, 1, timeout);
        if (ret > 0) {
            if (fds.revents & POLLNVAL) {
                errno = EINVAL;
                return FenceState::UNKNOWN;
            }
            if (fds.revents & POLLERR) {
                errno = EINVAL;
                return FenceState::ERROR;
            }
            return FenceState::SIGNALED;
        } else if (ret == 0) {
            errno = ETIME;
            return FenceState::ACTIVE;
        }
    } while (ret == -1 && (errno == EINTR || errno == EAGAIN));

    return FenceState::UNKNOWN;
}

#ifdef NN_DEBUGGABLE
uint32_t getProp(const char* str, uint32_t defaultValue) {
    const std::string propStr = android::base::GetProperty(str, "");
    if (propStr.size() > 0) {
        return std::stoi(propStr);
    } else {
        return defaultValue;
    }
}
#endif  // NN_DEBUGGABLE

}  // namespace nn
}  // namespace android