"Why is there no artificial intelligence yet?" Or, analysis of CNTK tool kit from Microsoft Research

• Introduction
• Those misprints...
• Working with files
• Working with memory
• Miscellaneous errors
• Conclusion:

Microsoft have given open access to the source code of a tool kit that is used in the company to speed up the development of artificial intelligence: Computational Network Toolkit is now available at GitHub. The developers had to create their own custom solution, because the existing tools did not work fast enough. Let's have a look at the analysis results of the source code of this project, as done by our static code analyzer.

Introduction

Computational Network Toolkit (CNTK) is a set of tools for designing and projecting nets of different types, that can be used for image detection, speech recognition, text analysis, and much more.

PVS-Studio is a static analyzer for bug detection in the source code of programs, written in C, C++ and C#. The PVS-Studio tool is made for developers of contemporary applications, and integrates into the Visual Studio environments of 2010-2015.

Preparing an article on an open source project check, we can only report , of course, on a limited number of all warnings issued by the analyzer, therefore we recommend the authors of the project run the analyzer on their code themselves, and study the complete analysis results. We also provide developers of open source projects with a temporary key.

I should say right away that there weren't many bugs found, which was as expected. Having checked several Microsoft projects, we can say that their code really is of very high quality. But we shouldn't forget that the benefit of a static code analyzer is in its regular use, not random checks.

Those misprints...

Typos are a very unpleasant thing. They've penetrated social networks, books, online publications, and even TV. In simple texts they can be eliminated by means of the spell check functions in text editors; in programming it can be done with the help of static code analyzers.

V501 There are identical sub-expressions '!Input(0)->HasMBLayout()' to the left and to the right of the '||' operator. trainingnodes.h 1416

virtual void Validate(bool isFinalValidationPass) override
{
  ....
  if (isFinalValidationPass &&
      !(Input(0)->GetSampleMatrixNumRows() ==
        Input(2)->GetSampleMatrixNumRows() &&
       (Input(0)->GetMBLayout() ==
        Input(2)->GetMBLayout() ||
       !Input(0)->HasMBLayout() ||            // <=
       !Input(0)->HasMBLayout())))            // <=
  {
    LogicError(..., NodeName().c_str(),OperationName().c_str());
  }
   ....
}

Formatting of this fragment is modified for clarity. Only after that it became apparent that there are two similar "! Input (0)-> HasMBLayout () "checks in the condition. Most likely, it is impossible to use an element with index '2' in one of the cases.

V501 There are identical sub-expressions to the left and to the right of the '-' operator: i0 - i0 ssematrix.h 564

void assignpatch(const ssematrixbase &patch,
                 const size_t i0,
                 const size_t i1,
                 const size_t j0,
                 const size_t j1)
{
  ....
  for (size_t j = j0; j < j1; j++)
  {
    const float *pcol = &patch(i0 - i0, j - j0);      // <=
    float *qcol = &us(i0, j);
    const size_t colbytes = (i1 - i0) * sizeof(*pcol);
    memcpy(qcol, pcol, colbytes);
  }
  ....
}

Because of the misprint, the condition "i0-i0" is always equal to zero. Perhaps "i1-i0" or "j-i1" or something else was meant here. Developers should definitely recheck this place.

V596 The object was created but it is not being used. The 'throw' keyword could be missing: throw runtime_error(FOO); simplenetworkbuilder.cpp 1578

template <class ElemType>
ComputationNetworkPtr SimpleNetworkBuilder<ElemType>::
  BuildNetworkFromDbnFile(const std::wstring& dbnModelFileName)
{
  ....
  if (this->m_outputLayerSize >= 0)
    outputLayerSize = this->m_outputLayerSize;
  else if (m_layerSizes.size() > 0)
    m_layerSizes[m_layerSizes.size() - 1];
  else
    std::runtime_error("Output layer size must be...");     // <=
  ....
}

The error, is that the keyword 'throw' was accidentally forgotten. As a result, this code does not generate an exception in case of an error. The correct code variant should be:

....
else
  throw std::runtime_error("Output layer size must be...");
....

Working with files

V739 EOF should not be compared with a value of the 'char' type. The 'c' should be of the 'int' type. fileutil.cpp 852

string fgetstring(FILE* f)
{
  string res;
  for (;;)
  {
    char c = (char) fgetc(f);        // <=
    if (c == EOF)                    // <=
      RuntimeError("error reading .... 0: %s", strerror(errno));
    if (c == 0)
      break;
    res.push_back(c);
  }
  return res;
}

The analyzer detected that the EOF constant is compared with a variable of 'char' type. This shows that some symbols will be processed incorrectly.

Let's look at the way EOF is declared:

#define EOF (-1)

As you can see, the EOF is nothing more than '-1 ' of 'int' type. Fgetc() function returns a value of 'int' type. Specifically, it can return a number from 0 to 255 or -1 (EOF). The values read are placed into a variable of 'char' type. Because of this, a symbol with the 0xFF (255) value turns into -1, and then handled in the same way as the file end (EOF).

Users that use Extended ASCII Codes, may encounter an error when one of the symbols of their alphabet is handled incorrectly by the program.

For example in the Windows 1251 code page, the last letter of Russian alphabet has the 0xFF code, and so is interpreted by the program as the end-of-file character.

Correct code fragment:

int c = fgetc(f);
if (c == EOF)
  RuntimeError(....);

V547 Expression 'val[0] == 0xEF' is always false. The value range of char type: [-128, 127]. file.cpp 462

bool File::IsUnicodeBOM(bool skip)
{
  ....
  else if (m_options & fileOptionsText)
  {
    char val[3];
    file.ReadString(val, 3);
    found = (val[0] == 0xEF && val[1] == 0xBB && val[2] == 0xBF);
  }
  // restore pointer if no BOM or we aren't skipping it
  if (!found || !skip)
  {
    SetPosition(pos);
  }
  ....
}

By default the 'char' type has a range of values equal to [-127;127]. Using the compilation flag /J, we can specify to the compiler to use the range [0; 255]. But there is no such flag for this source file, and so this code will never determine that this file contains BOM.

Working with memory

V595 The 'm_rowIndices' pointer was utilized before it was verified against nullptr. Check lines: 171, 175. libsvmbinaryreader.cpp 171

template <class ElemType>
void SparseBinaryMatrix<ElemType>::ResizeArrays(size_t newNNz)
{
  ....
  if (m_nnz > 0)
  {
    memcpy(rowIndices, m_rowIndices, sizeof(int32_t)....);  // <=
    memcpy(values, this->m_values, sizeof(ElemType)....);   // <=
  }

  if (m_rowIndices != nullptr)
  {
    // free(m_rowIndices);
    CUDAPageLockedMemAllocator::Free(this->m_rowIndices, ....);
  }
  if (this->m_values != nullptr)
  {
    // free(this->m_values);
    CUDAPageLockedMemAllocator::Free(this->m_values, ....);
  }
  ....
}

The analyzer detected integer dereference of a null pointer.

If there is a comparison with null in the code, when at an earlier time this pointer was used without a check, then this code is suspicious, and so, could be dangerous.

The memcpy() function copies the bytes located at "m_rowIndices" and "m_values", at the same time there is dereference of this pointer and in the given code it can potentially be equal to zero.

V510 The 'sprintf_s' function is not expected to receive class-type variable as third actual argument. binaryfile.cpp 501

const std::wstring& GetName()
{
  return m_name;
}

Section* Section::ReadSection(....)
{
  ....
  char message[256];
  sprintf_s(message,"Invalid header in file %ls, in header %s\n",
              m_file->GetName(), section->GetName());       // <=
  RuntimeError(message);
  ....
}

Only POD types can serve as actual parameters of sprint_s() function. POD is an abbreviation of "Plain Old Data", which can be interpreted as "Simple data in C style".

"std::wstring" doesn't belong to POD types. Instead of a pointer, the content of the object will go to the stack. This code will lead to some rubbish in the buffer or a program crash.

Correct variant:

sprintf_s(message,"Invalid header in file %ls, in header %s\n",
          m_file->GetName().c_str(), section->GetName().c_str());

V630 The 'malloc' function is used to allocate memory for an array of objects which are classes containing constructors. latticeforwardbackward.cpp 912

void lattice::forwardbackwardalign()
{
  ....
  aligninfo *refinfo;
  unsigned short *refalign;

  refinfo = (aligninfo *) malloc(sizeof(aligninfo) * 1);    // <=
  refalign = (unsigned short *) malloc(sizeof(....) * framenum);

  array_ref<aligninfo> refunits(refinfo, 1);
  array_ref<unsigned short> refedgealignmentsj(....);
  ....
}

In this code fragment, the analyzer detected an incorrect allocation of dynamic memory for a structure of "aligninfo" type. The thing here, is that there are constructors in the structure definition, but the constructor won't be called with this method of memory allocation. Also the destructor won't be called during the freeing of the memory with the help of free() function.

Here you can see a code fragment with the description of "aligninfo" type.

struct aligninfo // phonetic alignment
{
  unsigned int unit : 19;   // triphone index
  unsigned int frames : 11; // duration in frames
  unsigned int unused : 1; // (for future use)
  unsigned int last : 1;   // set for last entry
  aligninfo(size_t punit, size_t pframes)
      : unit((unsigned int) punit),
        frames((unsigned int) pframes), unused(0), last(0)
  {
    checkoverflow(unit, punit, "aligninfo::unit");
    checkoverflow(frames, pframes, "aligninfo::frames");
  }
  aligninfo() // [v-hansu] initialize to impossible values
  {
#ifdef INITIAL_STRANGE
    unit = unsigned int(-1);
    frames = unsigned int(-1);
    unused = unsigned int(-1);
    last = unsigned int(-1);
#endif
  }
  template <class IDMAP>
  void updateunit(const IDMAP& idmap /*[unit] -> new unit*/)
  {
    const size_t mappedunit = idmap[unit];
    unit = (unsigned int) mappedunit;
    checkoverflow(unit, mappedunit, "aligninfo::unit");
  }
};

Correct variant:

aligninfo *refinfo = new aligninfo();

And of course, you will need to call the 'delete' operator to free the memory.

V599 The virtual destructor is not present, although the 'IDataWriter' class contains virtual functions. datawriter.cpp 47

IDataWriter<ElemType>* m_dataWriter;
....
template <class ElemType>
void DataWriter<ElemType>::Destroy()
{
    delete m_dataWriter; // <= V599 warning
    m_dataWriter = NULL;
}

The analyzer warning shows that a base type of the object to be destroyed has no virtual destructor. In this case the destruction of the object of the derived class will cause undefined behavior of the program. In practice this can lead to memory leaks, and to a situation wherein other resources will not be released. Let's try to understand what caused this warning to appear.

template <class ElemType>
class DATAWRITER_API IDataWriter
{
public:
    typedef std::string LabelType;
    typedef unsigned int LabelIdType;

    virtual void Init(....) = 0;
    virtual void Init(....) = 0;
    virtual void Destroy() = 0;
    virtual void GetSections(....) = 0;
    virtual bool SaveData(....) = 0;
    virtual void SaveMapping(....) = 0;
};

This is a definition of a base class, as we can see, it has virtual functions, but a virtual destructor is missing.

m_dataWriter = new HTKMLFWriter<ElemType>();

Thus the memory is allocated for the object of the derived class "HTKMLFWriter". It's description:

template <class ElemType>
class HTKMLFWriter : public IDataWriter<ElemType>
{
private:
    std::vector<size_t> outputDims;
    std::vector<std::vector<std::wstring>> outputFiles;

    std::vector<size_t> udims;
    std::map<std::wstring, size_t> outputNameToIdMap;
    std::map<std::wstring, size_t> outputNameToDimMap;
    std::map<std::wstring, size_t> outputNameToTypeMap;
    unsigned int sampPeriod;
    size_t outputFileIndex;
    void Save(std::wstring& outputFile, ....);
    ElemType* m_tempArray;
    size_t m_tempArraySize;
    ....
};

Due to the missing virtual destructor in the base class, this object will not be properly destroyed. For outputDims, outputFiles objects the destructors will also not be called. However, in general it is impossible to predict all the aftereffects, this is why we use the term "undefined behavior".

Miscellaneous errors

V502 Perhaps the '?:' operator works in a different way than it was expected. The '?:' operator has a lower priority than the '|' operator. sequenceparser.h 338

enum SequenceFlags
{
    seqFlagNull = 0,
    seqFlagLineBreak = 1, // line break on the parsed line
    seqFlagEmptyLine = 2, // empty line
    seqFlagStartLabel = 4,
    seqFlagStopLabel = 8
};

long Parse(....)
{
  ....
  // sequence state machine variables
  bool m_beginSequence;
  bool m_endSequence;
  ....
  if (seqPos)
  {
    SequencePosition sequencePos(numbers->size(), labels->size(),
      m_beginSequence ? seqFlagStartLabel : 0 | m_endSequence ?
      seqFlagStopLabel : 0 | seqFlagLineBreak);
    // add a sequence element to the list
    seqPos->push_back(sequencePos);
    sequencePositionLast = sequencePos;
  }
  
  // end of sequence determines record separation
  if (m_endSequence)
      recordCount = (long) labels->size();
  ....
}

The priority of a ternary operator ':?' is lower than of a bitwise OR '|' operator. Lets' have a closer look at the fragment containing an error:

0 | m_endSequence ? seqFlagStopLabel : 0 | seqFlagLineBreak

It is expected that only bitwise operations with specified flags will be executed, however, because of an unexpected execution order, "0 | m_endSequence" will be executed first, instead of "m_endSequence ? seqFlagStopLabel : 0 | seqFlagLineBreak".

In fact, this is an interesting case. In spite of the error, the code works correctly. Bitwise OR with 0 doesn't affect anything.

Nevertheless, it is better to correct the error.

There are two more similar fragments:

• V502 Perhaps the '?:' operator works in a different way than it was expected. The '?:' operator has a lower priority than the '|' operator. sequenceparser.h 433
• V502 Perhaps the '?:' operator works in a different way than it was expected. The '?:' operator has a lower priority than the '|' operator. sequenceparser.h 598

V530 The return value of function 'size' is required to be utilized. basics.h 428

// TODO: merge this with todouble(const char*) above
static inline double todouble(const std::string& s)
{
  s.size(); // just used to remove the unreferenced warning

  double value = 0.0;
  ....
}

There is no error here, which we can see in the comment, but this example is given here for two reasons:

Firstly, to disable a compiler warning, there is an UNREFERENCED_PARAMETER macro, whose name makes it clear that the parameter of the function is not used deliberately:

#define UNREFERENCED_PARAMETER(P) (P)

static inline double todouble(const std::string& s)
{
  UNREFERENCED_PARAMETER(s);
  ....
}

Secondly, we would like to show another compiler warning that most likely indicates an error.

V530 The return value of function 'empty' is required to be utilized. utterancesourcemulti.h 340

template <class UTTREF>
std::vector<shiftedvector<....>>getclassids(const UTTREF &uttref)
{
  std::vector<shiftedvector<....>> allclassids;
  allclassids.empty();  // <=
  ....
}

There is no point in non-use of the result of the empty() function.

Perhaps the vector was to be cleared with the clear() function.

A similar fragment:

• V530 The return value of function 'empty' is required to be utilized. utterancesourcemulti.h 364

V688 The 'm_file' local variable possesses the same name as one of the class members, which can result in confusion. sequencereader.cpp 552

template <class ElemType>
class SequenceReader : public IDataReader<ElemType>
{
protected:
  bool m_idx2clsRead;
  bool m_clsinfoRead;

  bool m_idx2probRead;
  std::wstring m_file;                               // <=
  ....
}

template <class ElemType>
template <class ConfigRecordType>
void SequenceReader<ElemType>::InitFromConfig(....)
{
  ....
  std::wstring m_file = readerConfig(L"file");       // <=
  if (m_traceLevel > 0)
  {
    fprintf(stderr, "....", m_file.c_str());

  }
  ....
}

Using variables of the same name in the class, class functions, and class parameters, is a very bad programming style. For instance: was the variable declaration "std::wstring m_file = readerConfig(L"file");" supposed to be here, or was it added temporarily for debugging and then, was left forgotten?

Developers should also review the following fragments:

• V688 The 'm_file' local variable possesses the same name as one of the class members, which can result in a confusion. sequencereader.cpp 1554
• V688 The 'm_mbStartSample' function argument possesses the same name as one of the class members, which can result in a confusion. sequencereader.cpp 2062
• V688 The 'm_file' local variable possesses the same name as one of the class members, which can result in a confusion. lusequencereader.cpp 417

Conclusion:

Computational Network Toolkit (CNTK), being a relatively small project, turned out to be a pretty interesting piece of software. As the CNTK project has just recently been made open, we are looking forward to seeing new ideas for its usage, and of course, other open source projects by Microsoft.