par2cmdline/07-Optimization-by-inlining.patch

   1 --- a/reedsolomon.cpp
   2 +++ b/reedsolomon.cpp
   3 @@ -51,7 +51,7 @@ u32 gcd(u32 a, u32 b)
   4    }
   5  }
   6
   7 -bool ReedSolomon<Galois8>::SetInput(const vector<bool> &present)
   8 +template<> bool ReedSolomon<Galois8>::SetInput(const vector<bool> &present)
   9  {
  10    inputcount = (u32)present.size();
  11
  12 @@ -80,7 +80,7 @@ bool ReedSolomon<Galois8>::SetInput(const vector<bool> &present)
  13    return true;
  14  }
  15
  16 -bool ReedSolomon<Galois8>::SetInput(u32 count)
  17 +template<> bool ReedSolomon<Galois8>::SetInput(u32 count)
  18  {
  19    inputcount = count;
  20
  21 @@ -101,15 +101,8 @@ bool ReedSolomon<Galois8>::SetInput(u32 count)
  22    return true;
  23  }
  24
  25 -bool ReedSolomon<Galois8>::Process(size_t size, u32 inputindex, const void *inputbuffer, u32 outputindex, void *outputbuffer)
  26 +template<> bool ReedSolomon<Galois8>::InternalProcess(const Galois8 &factor, size_t size, const void *inputbuffer, void *outputbuffer)
  27  {
  28 -  // Look up the appropriate element in the RS matrix
  29 -  Galois8 factor = leftmatrix[outputindex * (datapresent + datamissing) + inputindex];
  30 -
  31 -  // Do nothing if the factor happens to be 0
  32 -  if (factor == 0)
  33 -    return eSuccess;
  34 -
  35  #ifdef LONGMULTIPLY
  36    // The 8-bit long multiplication tables
  37    Galois8 *table = glmt->tables;
  38 @@ -189,7 +182,7 @@ bool ReedSolomon<Galois8>::Process(size_t size, u32 inputindex, const void *inpu
  39
  40  // Set which of the source files are present and which are missing
  41  // and compute the base values to use for the vandermonde matrix.
  42 -bool ReedSolomon<Galois16>::SetInput(const vector<bool> &present)
  43 +template<> bool ReedSolomon<Galois16>::SetInput(const vector<bool> &present)
  44  {
  45    inputcount = (u32)present.size();
  46
  47 @@ -233,7 +226,7 @@ bool ReedSolomon<Galois16>::SetInput(const vector<bool> &present)
  48
  49  // Record that the specified number of source files are all present
  50  // and compute the base values to use for the vandermonde matrix.
  51 -bool ReedSolomon<Galois16>::SetInput(u32 count)
  52 +template<> bool ReedSolomon<Galois16>::SetInput(u32 count)
  53  {
  54    inputcount = count;
  55
  56 @@ -267,15 +260,8 @@ bool ReedSolomon<Galois16>::SetInput(u32 count)
  57    return true;
  58  }
  59
  60 -bool ReedSolomon<Galois16>::Process(size_t size, u32 inputindex, const void *inputbuffer, u32 outputindex, void *outputbuffer)
  61 +template<> bool ReedSolomon<Galois16>::InternalProcess(const Galois16 &factor, size_t size, const void *inputbuffer, void *outputbuffer)
  62  {
  63 -  // Look up the appropriate element in the RS matrix
  64 -
  65 -  Galois16 factor = leftmatrix[outputindex * (datapresent + datamissing) + inputindex];
  66 -  // Do nothing if the factor happens to be 0
  67 -  if (factor == 0)
  68 -    return eSuccess;
  69 -
  70  #ifdef LONGMULTIPLY
  71    // The 8-bit long multiplication tables
  72    Galois16 *table = glmt->tables;
  73 --- a/reedsolomon.h
  74 +++ b/reedsolomon.h
  75 @@ -64,7 +64,8 @@ public:
  76                 const void *inputbuffer, // Buffer containing input data
  77                 u32 outputindex,         // The row in the RS matrix
  78                 void *outputbuffer);     // Buffer containing output data
  79 -
  80 +private:
  81 +               bool InternalProcess(const g &factor, size_t size, const void *inputbuffer, void *outputbuffer);        // Optimization
  82  protected:
  83    // Perform Gaussian Elimination
  84    bool GaussElim(CommandLine::NoiseLevel noiselevel,
  85 @@ -146,6 +147,20 @@ inline ReedSolomon<g>::~ReedSolomon(void)
  86  #endif
  87  }
  88
  89 +template<class g>
  90 +inline bool ReedSolomon<g>::Process(size_t size, u32 inputindex, const void *inputbuffer, u32 outputindex, void *outputbuffer)
  91 +{
  92 +       // Optimization: it occurs frequently the function exits early on, so inline the start.
  93 +       // This resulted in a speed gain of approx. 8% in repairing.
  94 +
  95 +       // Look up the appropriate element in the RS matrix
  96 +       g factor = leftmatrix[outputindex * (datapresent + datamissing) + inputindex];
  97 +       // Do nothing if the factor happens to be 0
  98 +       if (factor == 0)
  99 +               return eSuccess;
 100 +       return this->InternalProcess (factor, size, inputbuffer, outputbuffer);
 101 +}
 102 +
 103  u32 gcd(u32 a, u32 b);
 104
 105  // Record whether the recovery block with the specified
 106 @@ -242,11 +257,14 @@ inline bool ReedSolomon<g>::Compute(CommandLine::NoiseLevel noiselevel)
 107    // One row for each present recovery block that will be used for a missing data block
 108    for (unsigned int row=0; row<datamissing; row++)
 109    {
 110 +    // Define MPDL to skip reporting and speed things up
 111 +#ifndef MPDL
 112      if (noiselevel > CommandLine::nlQuiet)
 113      {
 114        int progress = row * 1000 / (datamissing+parmissing);
 115        cout << "Constructing: " << progress/10 << '.' << progress%10 << "%\r" << flush;
 116      }
 117 +#endif
 118
 119      // Get the exponent of the next present recovery block
 120      while (!outputrow->present)
 121 @@ -286,11 +304,14 @@ inline bool ReedSolomon<g>::Compute(CommandLine::NoiseLevel noiselevel)
 122    outputrow = outputrows.begin();
 123    for (unsigned int row=0; row<parmissing; row++)
 124    {
 125 +    // Define MPDL to skip reporting and speed things up
 126 +#ifndef MPDL
 127      if (noiselevel > CommandLine::nlQuiet)
 128      {
 129        int progress = (row+datamissing) * 1000 / (datamissing+parmissing);
 130        cout << "Constructing: " << progress/10 << '.' << progress%10 << "%\r" << flush;
 131      }
 132 +#endif
 133
 134      // Get the exponent of the next missing recovery block
 135      while (outputrow->present)
 136 @@ -420,6 +441,8 @@ inline bool ReedSolomon<g>::GaussElim(CommandLine::NoiseLevel noiselevel, unsign
 137      // For every other row in the matrix
 138      for (unsigned int row2=0; row2<rows; row2++)
 139      {
 140 +      // Define MPDL to skip reporting and speed things up
 141 +#ifndef MPDL
 142        if (noiselevel > CommandLine::nlQuiet)
 143        {
 144          int newprogress = (row*rows+row2) * 1000 / (datamissing*rows);
 145 @@ -429,6 +452,7 @@ inline bool ReedSolomon<g>::GaussElim(CommandLine::NoiseLevel noiselevel, unsign
 146            cout << "Solving: " << progress/10 << '.' << progress%10 << "%\r" << flush;
 147          }
 148        }
 149 +#endif
 150
 151        if (row != row2)
 152        {