/*
* 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.
*/
#ifndef ANDROID_ML_NN_COMMON_UTILS_H
#define ANDROID_ML_NN_COMMON_UTILS_H
#include "HalInterfaces.h"
#include "NeuralNetworks.h"
#include "ValidateHal.h"
#include <android-base/logging.h>
#include <optional>
#include <set>
#include <vector>
namespace android {
namespace nn {
// The number of data types (OperandCode) defined in NeuralNetworks.h.
const int kNumberOfDataTypes = 14;
// The number of operation types (OperationCode) defined in NeuralNetworks.h.
const int kNumberOfOperationTypes = 95;
// The number of execution preferences defined in NeuralNetworks.h.
const int kNumberOfPreferences = 3;
// The number of data types (OperandCode) defined in NeuralNetworksOEM.h.
const int kNumberOfDataTypesOEM = 2;
// The number of operation types (OperationCode) defined in NeuralNetworksOEM.h.
const int kNumberOfOperationTypesOEM = 1;
// The lowest number assigned to any OEM Code in NeuralNetworksOEM.h.
const int kOEMCodeBase = 10000;
/* IMPORTANT: if you change the following list, don't
* forget to update the corresponding 'tags' table in
* the initVlogMask() function implemented in Utils.cpp.
*/
enum VLogFlags {
MODEL = 0,
COMPILATION,
EXECUTION,
CPUEXE,
MANAGER,
DRIVER
};
#define VLOG_IS_ON(TAG) \
((vLogMask & (1 << (TAG))) != 0)
#define VLOG(TAG) \
if (LIKELY(!VLOG_IS_ON(TAG))) \
; \
else \
LOG(INFO)
extern int vLogMask;
void initVLogMask();
#ifdef NN_DEBUGGABLE
#define SHOW_IF_DEBUG(msg) msg
#else
#define SHOW_IF_DEBUG(msg) ""
#endif
// DEPRECATED(b/118737105). Use CHECK.
#define nnAssert(v) CHECK(v)
#define NN_RETURN_IF_ERROR(expr) \
do { \
int _errorCode = (expr); \
if (_errorCode != ANEURALNETWORKS_NO_ERROR) { \
return _errorCode; \
} \
} while (0)
// The NN_RET_CHECK family of macros defined below is similar to the CHECK family defined in
// system/core/base/include/android-base/logging.h
//
// The difference is that NN_RET_CHECK macros use LOG(ERROR) instead of LOG(FATAL)
// and return false instead of aborting.
// Logs an error and returns false. Append context using << after. For example:
//
// NN_RET_CHECK_FAIL() << "Something went wrong";
//
// The containing function must return a bool.
#define NN_RET_CHECK_FAIL() \
return ::android::nn::FalseyErrorStream() \
<< "NN_RET_CHECK failed (" << __FILE__ << ":" << __LINE__ << "): "
// Logs an error and returns false if condition is false. Extra logging can be appended using <<
// after. For example:
//
// NN_RET_CHECK(false) << "Something went wrong";
//
// The containing function must return a bool.
#define NN_RET_CHECK(condition) \
while (UNLIKELY(!(condition))) NN_RET_CHECK_FAIL() << #condition << " "
// Helper for NN_CHECK_xx(x, y) macros.
#define NN_RET_CHECK_OP(LHS, RHS, OP) \
for (auto _values = ::android::base::MakeEagerEvaluator(LHS, RHS); \
UNLIKELY(!(_values.lhs OP _values.rhs)); \
/* empty */) \
NN_RET_CHECK_FAIL() << #LHS << " " << #OP << " " << #RHS << " (" << #LHS << " = " \
<< _values.lhs << ", " << #RHS << " = " << _values.rhs << ") "
// Logs an error and returns false if a condition between x and y does not hold. Extra logging can
// be appended using << after. For example:
//
// NN_RET_CHECK_EQ(a, b) << "Something went wrong";
//
// The values must implement the appropriate comparison operator as well as
// `operator<<(std::ostream&, ...)`.
// The containing function must return a bool.
#define NN_RET_CHECK_EQ(x, y) NN_RET_CHECK_OP(x, y, ==)
#define NN_RET_CHECK_NE(x, y) NN_RET_CHECK_OP(x, y, !=)
#define NN_RET_CHECK_LE(x, y) NN_RET_CHECK_OP(x, y, <=)
#define NN_RET_CHECK_LT(x, y) NN_RET_CHECK_OP(x, y, <)
#define NN_RET_CHECK_GE(x, y) NN_RET_CHECK_OP(x, y, >=)
#define NN_RET_CHECK_GT(x, y) NN_RET_CHECK_OP(x, y, >)
// A wrapper around LOG(ERROR) that can be implicitly converted to bool (always evaluates to false).
// Used to implement stream logging in NN_RET_CHECK.
class FalseyErrorStream {
DISALLOW_COPY_AND_ASSIGN(FalseyErrorStream);
public:
FalseyErrorStream() {}
template <typename T>
FalseyErrorStream& operator<<(const T& value) {
mBuffer << value;
return *this;
}
~FalseyErrorStream() { LOG(ERROR) << mBuffer.str(); }
operator bool() const { return false; }
private:
std::ostringstream mBuffer;
};
// Return a vector with one entry for each non extension OperandType, set to the
// specified PerformanceInfo value. The vector will be sorted by OperandType.
hidl_vec<Capabilities::OperandPerformance> nonExtensionOperandPerformance(PerformanceInfo perf);
// Update the vector entry corresponding to the specified OperandType with the
// specified PerformanceInfo value. The vector must already have an entry for
// that OperandType, and must be sorted by OperandType.
void update(hidl_vec<Capabilities::OperandPerformance>* operandPerformance, OperandType type,
PerformanceInfo perf);
// Look for a vector entry corresponding to the specified OperandType. If
// found, return the associated PerformanceInfo. If not, return a pessimistic
// PerformanceInfo (FLT_MAX). The vector must be sorted by OperandType.
PerformanceInfo lookup(const hidl_vec<Capabilities::OperandPerformance>& operandPerformance,
OperandType type);
// Returns true if an operand type is an extension type.
bool isExtensionOperandType(OperandType type);
// Returns true if an operation type is an extension type.
bool isExtensionOperationType(OperationType type);
// Returns the amount of space needed to store a value of the specified
// dimensions and type. For a tensor with unspecified rank or at least one
// unspecified dimension, returns zero.
//
// Aborts if the specified type is an extension type.
//
// See also TypeManager::getSizeOfData(OperandType, const std::vector<uint32_t>&).
uint32_t nonExtensionOperandSizeOfData(OperandType type, const std::vector<uint32_t>& dimensions);
// Returns the amount of space needed to store a value of the dimensions and
// type of this operand. For a tensor with unspecified rank or at least one
// unspecified dimension, returns zero.
//
// Aborts if the specified type is an extension type.
//
// See also TypeManager::getSizeOfData(const Operand&).
inline uint32_t nonExtensionOperandSizeOfData(const Operand& operand) {
return nonExtensionOperandSizeOfData(operand.type, operand.dimensions);
}
// Returns true if a non-extension operand type is a scalar type.
//
// Aborts if the specified type is an extension type.
//
// See also TypeManager::isTensorType(OperandType).
bool nonExtensionOperandTypeIsScalar(int type);
// Returns the name of the operation type in ASCII.
std::string getOperationName(OperationType opCode);
// Returns the name of the operand type in ASCII.
std::string getOperandTypeName(OperandType type);
// Whether an operand of tensor type has unspecified dimensions.
//
// Undefined behavior if the operand type is a scalar type.
bool tensorHasUnspecifiedDimensions(int type, const uint32_t* dim, uint32_t dimCount);
bool tensorHasUnspecifiedDimensions(const Operand& operand);
bool tensorHasUnspecifiedDimensions(const ANeuralNetworksOperandType* type);
// Memory is unmapped.
// Memory is reference counted by hidl_memory instances, and is deallocated
// once there are no more references.
hidl_memory allocateSharedMemory(int64_t size);
// Returns the number of padding bytes needed to align data of the
// specified length. It aligns object of length:
// 2, 3 on a 2 byte boundary,
// 4+ on a 4 byte boundary.
// We may want to have different alignments for tensors.
// TODO: This is arbitrary, more a proof of concept. We need
// to determine what this should be.
uint32_t alignBytesNeeded(uint32_t index, size_t length);
// Does a detailed LOG(INFO) of the model
void logModelToInfo(const V1_0::Model& model);
void logModelToInfo(const V1_1::Model& model);
void logModelToInfo(const V1_2::Model& model);
inline std::string toString(uint32_t obj) {
return std::to_string(obj);
}
template <typename Type>
std::string toString(const std::vector<Type>& range) {
std::string os = "[";
for (size_t i = 0; i < range.size(); ++i) {
os += (i == 0 ? "" : ", ") + toString(range[i]);
}
return os += "]";
}
inline std::string toString(HalVersion halVersion) {
switch (halVersion) {
case HalVersion::UNKNOWN:
return "UNKNOWN HAL version";
case HalVersion::V1_0:
return "HAL version 1.0";
case HalVersion::V1_1:
return "HAL version 1.1";
case HalVersion::V1_2:
return "HAL version 1.2";
}
}
inline bool validCode(uint32_t codeCount, uint32_t codeCountOEM, uint32_t code) {
return (code < codeCount) || (code >= kOEMCodeBase && (code - kOEMCodeBase) < codeCountOEM);
}
bool validateOperandSymmPerChannelQuantParams(
const Operand& halOperand, const ANeuralNetworksSymmPerChannelQuantParams& channelQuant,
const char* tag);
// Validates an operand type.
//
// extensionOperandTypeInfo must be nullptr iff the type is not an extension type.
//
// If allowPartial is true, the dimensions may be underspecified.
int validateOperandType(const ANeuralNetworksOperandType& type,
const Extension::OperandTypeInformation* const extensionOperandTypeInfo,
const char* tag, bool allowPartial);
int validateOperandList(uint32_t count, const uint32_t* list, uint32_t operandCount,
const char* tag);
// Returns ANEURALNETWORKS_NO_ERROR if the corresponding operation is defined and can handle the
// provided operand types in the given HAL version, otherwise returns ANEURALNETWORKS_BAD_DATA.
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);
inline size_t getSizeFromInts(int lower, int higher) {
return (uint32_t)(lower) + ((uint64_t)(uint32_t)(higher) << 32);
}
// Convert ANEURALNETWORKS_* result code to ErrorStatus.
// Not guaranteed to be a 1-to-1 mapping.
ErrorStatus convertResultCodeToErrorStatus(int resultCode);
// Convert ErrorStatus to ANEURALNETWORKS_* result code.
// Not guaranteed to be a 1-to-1 mapping.
int convertErrorStatusToResultCode(ErrorStatus status);
// Versioning
bool compliantWithV1_0(const V1_0::Capabilities& capabilities);
bool compliantWithV1_0(const V1_1::Capabilities& capabilities);
bool compliantWithV1_0(const V1_2::Capabilities& capabilities);
bool compliantWithV1_1(const V1_0::Capabilities& capabilities);
bool compliantWithV1_1(const V1_1::Capabilities& capabilities);
bool compliantWithV1_1(const V1_2::Capabilities& capabilities);
bool compliantWithV1_2(const V1_0::Capabilities& capabilities);
bool compliantWithV1_2(const V1_1::Capabilities& capabilities);
bool compliantWithV1_2(const V1_2::Capabilities& capabilities);
bool compliantWithV1_0(const V1_2::Operand& operand);
// If noncompliantOperations != nullptr, then
// precondition: noncompliantOperations->empty()
// postcondition: *noncompliantOperations consists of the indices of the noncompliant
// operations; if the compliance check fails for some reason
// other than a noncompliant operation,
// *noncompliantOperations consists of the indices of all operations
bool compliantWithV1_0(const V1_0::Model& model);
bool compliantWithV1_0(const V1_1::Model& model);
bool compliantWithV1_0(const V1_2::Model& model,
std::set<uint32_t>* noncompliantOperations = nullptr);
bool compliantWithV1_1(const V1_0::Model& model);
bool compliantWithV1_1(const V1_1::Model& model);
bool compliantWithV1_1(const V1_2::Model& model,
std::set<uint32_t>* noncompliantOperations = nullptr);
V1_0::Capabilities convertToV1_0(const V1_0::Capabilities& capabilities);
V1_0::Capabilities convertToV1_0(const V1_1::Capabilities& capabilities);
V1_0::Capabilities convertToV1_0(const V1_2::Capabilities& capabilities);
V1_1::Capabilities convertToV1_1(const V1_0::Capabilities& capabilities);
V1_1::Capabilities convertToV1_1(const V1_1::Capabilities& capabilities);
V1_1::Capabilities convertToV1_1(const V1_2::Capabilities& capabilities);
V1_2::Capabilities convertToV1_2(const V1_0::Capabilities& capabilities);
V1_2::Capabilities convertToV1_2(const V1_1::Capabilities& capabilities);
V1_2::Capabilities convertToV1_2(const V1_2::Capabilities& capabilities);
V1_0::Model convertToV1_0(const V1_0::Model& model);
V1_0::Model convertToV1_0(const V1_1::Model& model);
V1_0::Model convertToV1_0(const V1_2::Model& model);
V1_1::Model convertToV1_1(const V1_0::Model& model);
V1_1::Model convertToV1_1(const V1_1::Model& model);
V1_1::Model convertToV1_1(const V1_2::Model& model);
V1_2::Model convertToV1_2(const V1_0::Model& model);
V1_2::Model convertToV1_2(const V1_1::Model& model);
V1_2::Model convertToV1_2(const V1_2::Model& model);
// The IModelSlicer abstract class provides methods to create from an original
// model a "slice" of that model consisting of the subset of operations that is
// compliant with a particular HAL version, and a mechanism for mapping
// operations from the slice back to operations of the original model. The
// slice is intended to be passed to getSupportedOperations*(), with the mapping
// used to translate the results of that call from the slice's operations to the
// original model's operations. The slice has no other purpose (for example, it
// is not guaranteed to have the same topology as a subgraph of the original
// model).
//
// Note that the original model is not part of the ModelSlicer specification --
// an instance of a class derived from ModelSlicer is responsible for knowing
// the original model. getSlice*() methods may be called multiple times on a
// given instance; the intention is that the instance cache slices internally.
//
// The meaning of the return value of the getSlice*() methods is explained by
// the following example:
//
// IModelSlicer* slicer = ...;
// auto ret = slicer->getSliceV1_0(); // getSliceV1_1() is similar
// if (ret.has_value()) {
// const V1_0::Model model = ret->first; // the slice
// auto mapper = ret->second;
// // mapper is a functor that takes an operation index in the
// // slice and returns the corresponding operation index in the
// // original model. The functor must remain valid for the lifetime
// // of *slicer.
// } else {
// // Could not obtain a slice. For example, perhaps none of the
// // original model's operations are compliant with V1_0.
// }
//
class IModelSlicer {
public:
virtual std::optional<std::pair<V1_0::Model, std::function<uint32_t(uint32_t)>>>
getSliceV1_0() = 0;
virtual std::optional<std::pair<V1_1::Model, std::function<uint32_t(uint32_t)>>>
getSliceV1_1() = 0;
virtual ~IModelSlicer() = default;
};
V1_0::OperationType uncheckedConvertToV1_0(V1_2::OperationType type);
V1_1::OperationType uncheckedConvertToV1_1(V1_2::OperationType type);
V1_0::Operand convertToV1_0(const V1_2::Operand& operand);
V1_2::Operand convertToV1_2(const V1_0::Operand& operand);
V1_2::Operand convertToV1_2(const V1_2::Operand& operand);
hidl_vec<V1_2::Operand> convertToV1_2(const hidl_vec<V1_0::Operand>& operands);
hidl_vec<V1_2::Operand> convertToV1_2(const hidl_vec<V1_2::Operand>& operands);
#ifdef NN_DEBUGGABLE
uint32_t getProp(const char* str, uint32_t defaultValue = 0);
#endif // NN_DEBUGGABLE
} // namespace nn
} // namespace android
#endif // ANDROID_ML_NN_COMMON_UTILS_H