SampleDrivers/Matrix_Compression_and_Recovery/ADOL-C/11_Compression_and_direct_recovery_for_Hessian_return_Row_Compressed_Format__unmanaged_usermem.cpp

Go to the documentation of this file.

// An example of (Column) compression (by using Star coloring) and direct recovery for Hessian
/* How to compile this driver manually:
        Please make sure that "baseDir" point to the directory (folder) containing the input matrix file, and
                s_InputFile should point to the input file that you want to use
        To compile the code, replace the Main.cpp file in Main directory with this file
                and run "make" in ColPack installation directory. Make will generate "ColPack.exe" executable
        Run "ColPack.exe"

Note: If you got "symbol lookup error ... undefined symbol "
  Please make sure that your LD_LIBRARY_PATH contains libColPack.so

//*/

#include "ColPackHeaders.h"

using namespace ColPack;
using namespace std;

#ifndef TOP_DIR
#define TOP_DIR "."
#endif

// baseDir should point to the directory (folder) containing the input file
string baseDir=TOP_DIR;

#include "extra.h" //This .h file contains functions that are used in the below examples:
                                        //ReadMM(), MatrixMultiplication...(), Times2Plus1point5(), displayMatrix() and displayCompressedRowMatrix()

int main()
{
        // s_InputFile = baseDir + <name of the input file>
        string s_InputFile; //path of the input file
        s_InputFile = baseDir;
        s_InputFile += DIR_SEPARATOR; s_InputFile += "Graphs"; s_InputFile += DIR_SEPARATOR; s_InputFile += "mtx-spear-head.mtx";

        // Step 1: Determine sparsity structure of the Jacobian.
        // This step is done by an AD tool. For the purpose of illustration here, we read the structure from a file,
        // and store the structure in a Compressed Row Format.
        unsigned int *** uip3_SparsityPattern = new unsigned int **;
        double*** dp3_Value = new double**;
        int rowCount, columnCount;
        ConvertMatrixMarketFormat2RowCompressedFormat(s_InputFile, uip3_SparsityPattern, dp3_Value,rowCount, columnCount);

        cout<<"just for debugging purpose, display the 2 matrices: the matrix with SparsityPattern only and the matrix with Value"<<endl;
        cout<<fixed<<showpoint<<setprecision(2); //formatting output
        cout<<"(*uip3_SparsityPattern)"<<endl;
        displayCompressedRowMatrix((*uip3_SparsityPattern),rowCount);
        cout<<"(*dp3_Value)"<<endl;
        displayCompressedRowMatrix((*dp3_Value),rowCount);
        cout<<"Finish ConvertMatrixMarketFormat2RowCompressedFormat()"<<endl;
        Pause();

        //Step 2: Obtain the seed matrix via coloring.
        double*** dp3_Seed = new double**;
        int *ip1_SeedRowCount = new int;
        int *ip1_SeedColumnCount = new int;
        int *ip1_ColorCount = new int; //The number of distinct colors used to color the graph

        //Step 2.1: Read the sparsity pattern of the given Hessian matrix (compressed sparse rows format)
        //and create the corresponding graph
        GraphColoringInterface *g = new GraphColoringInterface(SRC_MEM_ADOLC,*uip3_SparsityPattern, rowCount);

        //Step 2.2: Color the bipartite graph with the specified ordering
        g->Coloring("SMALLEST_LAST", "STAR");

        //Step 2.3 (Option 1): From the coloring information, create and return the seed matrix
        (*dp3_Seed) = g->GetSeedMatrix(ip1_SeedRowCount, ip1_SeedColumnCount);
        /* Notes:
        Step 2.3 (Option 2): From the coloring information, you can also get the vector of colorIDs of vertices
                vector<int> vi_VertexColors;
                g->GetVertexColors(vi_VertexColors);
        */
        cout<<"Finish GenerateSeed()"<<endl;
        *ip1_ColorCount = *ip1_SeedColumnCount;

        displayMatrix(*dp3_Seed,columnCount,*ip1_SeedColumnCount);
        Pause();

        // Step 3: Obtain the Hessian-seed matrix product.
        // This step will also be done by an AD tool. For the purpose of illustration here, the orginial matrix V
        // (for Values) is multiplied with the seed matrix S. The resulting matrix is stored in dp3_CompressedMatrix.
        double*** dp3_CompressedMatrix = new double**;
        cout<<"Start MatrixMultiplication()"<<endl;
        MatrixMultiplication_VxS(*uip3_SparsityPattern, *dp3_Value, rowCount, columnCount, *dp3_Seed, *ip1_ColorCount, dp3_CompressedMatrix);
        cout<<"Finish MatrixMultiplication()"<<endl;

        displayMatrix(*dp3_CompressedMatrix,rowCount,*ip1_ColorCount);
        Pause();

        //Step 4: Recover the numerical values of the original matrix from the compressed representation.
        // The new values are store in "dp3_NewValue"
        double*** dp3_NewValue = new double**;
        HessianRecovery* hr = new HessianRecovery;
        int rowCount_for_dp3_NewValue = hr->DirectRecover_RowCompressedFormat_unmanaged(g, *dp3_CompressedMatrix, *uip3_SparsityPattern, dp3_NewValue);
        /* RecoverD2Cln_RowCompressedFormat_unmanaged is called instead of RecoverD2Cln_RowCompressedFormat so that
        we could manage the memory deallocation for dp3_NewValue by ourselves.
        This way, we can reuse (*dp3_NewValue) to store new values if RecoverD2Cln_RowCompressedFormat...() need to be called again.
        //*/
        
        cout<<"Finish Recover()"<<endl;

        displayCompressedRowMatrix(*dp3_NewValue,rowCount);
        Pause();

        //Check for consistency, make sure the values in the 2 matrices are the same.
        if (CompressedRowMatricesAreEqual(*dp3_Value, *dp3_NewValue, rowCount,0,1)) cout<< "*dp3_Value == dp3_NewValue"<<endl;
        else cout<< "*dp3_Value != dp3_NewValue"<<endl;

        Pause();
        
        /* Let say that we have new matrices with the same sparsity structure (only the values changed),
         We can take advantage of the memory already allocated to (*dp3_NewValue) and use (*dp3_NewValue) to stored the new values
         by calling DirectRecover_RowCompressedFormat_usermem recovery function.
         This function works in the same way as DirectRecover_RowCompressedFormat_unmanaged except that the memory for (*dp3_NewValue)
         is reused and therefore, takes less time than the _unmanaged counterpart.
         //*/
        for(int i=0; i<3;i++) {
          hr->DirectRecover_RowCompressedFormat_usermem(g, *dp3_CompressedMatrix, *uip3_SparsityPattern, dp3_NewValue);
        }
        
        //Deallocate memory for 2-dimensional array (*dp3_NewValue)
        for(int i=0; i<rowCount_for_dp3_NewValue;i++) {
          free((*dp3_NewValue)[i]);
        }
        free(*dp3_NewValue);

        //Deallocate memory using functions in Utilities/MatrixDeallocation.h

        free_2DMatrix(uip3_SparsityPattern, rowCount);
        uip3_SparsityPattern=NULL;

        free_2DMatrix(dp3_Value, rowCount);
        dp3_Value=NULL;

        delete dp3_Seed;
        dp3_Seed = NULL;

        delete ip1_SeedRowCount;
        ip1_SeedRowCount=NULL;

        delete ip1_SeedColumnCount;
        ip1_SeedColumnCount = NULL;

        free_2DMatrix(dp3_CompressedMatrix, rowCount);
        dp3_CompressedMatrix = NULL;

        delete ip1_ColorCount;
        ip1_ColorCount = NULL;

        delete hr;
        hr = NULL;

        delete dp3_NewValue;
        dp3_NewValue=NULL;

        delete g;
        g=NULL;

        return 0;
}