C++ OptionParser::getOptionVecInt方法代码示例

// ****************************************************************************
// Function: GPUSetup
//
// Purpose:
//  do the necessary OpenCL setup for GPU part of the test
//
// Arguments:
//   op: the options parser / parameter database
//   mympirank: for printing errors in case of failure
//   mynoderank: this is typically the device ID (the mapping done in main)
//
// Returns: success/failure
//
// Creation: 2009
//
// Modifications:
//
// ****************************************************************************
//
int GPUSetup(OptionParser &op, int mympirank, int mynoderank)
{
    addBenchmarkSpecOptions(op);
    
    if (op.getOptionBool("infoDevices"))
    {
        OpenCLNodePlatformContainer ndc1;
        ndc1.Print (cout);
        return (0);
    }
    
    // The device option supports specifying more than one device
    int platform = op.getOptionInt("platform");
    int deviceIdx = mynoderank;
    if( deviceIdx >= op.getOptionVecInt( "device" ).size() )
    {
        std::ostringstream estr;
        estr << "Warning: not enough devices specified with --device flag for task "
            << mympirank
            << " ( node rank " << mynoderank
            << ") to claim its own device; forcing to use first device ";
        std::cerr << estr.str() << std::endl;
        deviceIdx = 0;
    }
    int device = op.getOptionVecInt("device")[deviceIdx];

    // Initialization
    _mpicontention_ocldev = new cl::Device( ListDevicesAndGetDevice(platform, device) );
    std::vector<cl::Device> ctxDevices;
    ctxDevices.push_back( *_mpicontention_ocldev );
    _mpicontention_ocldriver_ctx   = new cl::Context( ctxDevices );
    _mpicontention_ocldriver_queue = new cl::CommandQueue( *_mpicontention_ocldriver_ctx, *_mpicontention_ocldev, CL_QUEUE_PROFILING_ENABLE );
    _mpicontention_gpuop = op;
    return 0;
}

// validate stencil-independent values
void
CheckOptions( const OptionParser& opts )
{
    // check matrix dimensions - must be 2d, must be positive
    std::vector<long long> arrayDims = opts.getOptionVecInt( "customSize" );
    if( arrayDims.size() != 2 )
    {
        throw InvalidArgValue( "overall size must have two dimensions" );
    }
    if( (arrayDims[0] < 0) || (arrayDims[1] < 0) )
    {
        throw InvalidArgValue( "each size dimension must be positive" );
    }

    // validation error threshold must be positive
    float valThreshold = opts.getOptionFloat( "val-threshold" );
    if( valThreshold <= 0.0f )
    {
        throw InvalidArgValue( "validation threshold must be positive" );
    }

    // number of validation errors to print must be non-negative
    int nErrsToPrint = opts.getOptionInt( "val-print-limit" );
    if( nErrsToPrint < 0 )
    {
        throw InvalidArgValue( "number of validation errors to print must be non-negative" );
    }

    int nWarmupPasses = opts.getOptionInt( "warmupPasses" );
    if( nWarmupPasses < 0 )
    {
        throw InvalidArgValue( "number of warmup passes must be non-negative" );
    }
}

void 
CommonMICStencilFactory<T>::CheckOptions( const OptionParser& opts ) const
{
    // let base class check its options first
    StencilFactory<T>::CheckOptions( opts );

    // check our options
    std::vector<long long> arrayDims = opts.getOptionVecInt( "customSize" );
    assert( arrayDims.size() == 2 );

    // If both of these are zero, we're using a non-custom size, skip this test
    if (arrayDims[0] == 0 && arrayDims[0] == 0)
    {
        return;
    }

    size_t gRows = (size_t)arrayDims[0];
    size_t gCols = (size_t)arrayDims[1];
    size_t lRows = LROWS;
    size_t lCols = LCOLS;

    // verify that local dimensions evenly divide global dimensions
    if( ((gRows % lRows) != 0) || (lRows > gRows) )
    {
        throw InvalidArgValue( "number of rows must be even multiple of lsize rows" );
    }
    if( ((gCols % lCols) != 0) || (lCols > gCols) )
    {
        throw InvalidArgValue( "number of columns must be even multiple of lsize columns" );
    }

    // TODO ensure local dims are smaller than CUDA implementation limits
}

void
MPICUDAStencilFactory<T>::CheckOptions( const OptionParser& opts ) const
{
    // let base class check its options first
    CommonCUDAStencilFactory<T>::CheckOptions( opts );

    // check our options
    std::vector<long long> shDims = opts.getOptionVecInt( "lsize" );
    std::vector<long long> arrayDims = opts.getOptionVecInt( "customSize" );
    if( arrayDims[0] == 0 )
    {
        // custom size was not specified - we are using a standard size
        int sizeClass = opts.getOptionInt("size");
        arrayDims = StencilFactory<T>::GetStandardProblemSize( sizeClass );
    }
    assert( shDims.size() == 2 );
    assert( arrayDims.size() == 2 );

    size_t gRows = (size_t)arrayDims[0];
    size_t gCols = (size_t)arrayDims[1];
    size_t lRows = shDims[0];
    size_t lCols = shDims[1];

    unsigned int haloWidth = (unsigned int)opts.getOptionInt( "iters-per-exchange" );

    // verify that MPI halo width will result in a matrix being passed
    // to the kernel that also has its global size as a multiple of
    // the local work size
    //
    // Because the MPI halo width is arbitrary, and the kernel halo width
    // is always 1, we have to ensure that:
    //   ((size + 2*halo) - 2) % lsize == 0
    if( (((gRows + 2*haloWidth) - 2) % lRows) != 0 )
    {
        throw InvalidArgValue( "rows including halo must be even multiple of lsize (e.g., lsize rows evenly divides ((rows + 2*halo) - 2) )" );
    }
    if( (((gCols + 2*haloWidth) - 2) % lCols) != 0 )
    {
        throw InvalidArgValue( "columns including halo must be even multiple of lsize (e.g., lsize cols evenly divides ((cols + 2*halo) - 2) )" );
    }
}

void
CommonMICStencilFactory<T>::ExtractOptions( const OptionParser& options,
                                            T& wCenter,
                                            T& wCardinal,
                                            T& wDiagonal,
                                            std::vector<long long>& devices )
{
    // let base class extract its options
    StencilFactory<T>::ExtractOptions( options, wCenter, wCardinal, wDiagonal );

    // extract our options
    // with hardcoded lsize, we no longer have any to extract

    // determine which device to use
    // We would really prefer this to be done in main() but 
    // since BuildStencil is a virtual function, we cannot change its
    // signature, and OptionParser provides no way to override an
    // option's value after it is set during parsing.
    devices = options.getOptionVecInt("device");
}

//.........这里部分代码省略.........
            mnc2.Print (cout);
            mnc1.merge (mnc2);
            cout << "==============\nObject1 after merging 1:\n";
            mnc1.Print (cout);
            mnc1.merge (mnc2);
            cout << "==============\nObject1 after merging 2:\n";
            mnc1.Print (cout);
#endif  // DEBUG
#endif  // PARALLEL
            return (0);
        }

        bool verbose = op.getOptionBool("verbose");

        // The device option supports specifying more than one device
        // for now, just choose the first one.
        int platform = op.getOptionInt("platform");

#ifdef PARALLEL
        NodeInfo ni;
        int myNodeRank = ni.nodeRank();
        if (verbose)
        cout << "Global rank "<<rank<<" is local rank "<<myNodeRank << endl;
#else
        int myNodeRank = 0;
#endif

        // If they haven't specified any devices, assume they
        // want the process with in-node rank N to use device N
        int deviceIdx = myNodeRank;

        // If they have, then round-robin the list of devices
        // among the processes on a node.
        vector<long long> deviceVec = op.getOptionVecInt("device");
        if (deviceVec.size() > 0)
        {
        int len = deviceVec.size();
            deviceIdx = deviceVec[myNodeRank % len];
        }

        // Check for an erroneous device
        if (deviceIdx >= GetNumOclDevices(platform)) {
            cerr << "Warning: device index: " << deviceIdx
                 << " out of range, defaulting to device 0.\n";
            deviceIdx = 0;
        }

        // Initialization
        if (verbose) cout << ">> initializing\n";
        cl_device_id devID = ListDevicesAndGetDevice(platform, deviceIdx);
        cl_int clErr;
        cl_context ctx = clCreateContext( NULL,     // properties
                                            1,      // number of devices
                                            &devID, // device
                                            NULL,   // notification function
                                            NULL,
                                            &clErr );
        CL_CHECK_ERROR(clErr);
        cl_command_queue queue = clCreateCommandQueue( ctx,
                                                        devID,
                                                        CL_QUEUE_PROFILING_ENABLE,
                                                        &clErr );
        CL_CHECK_ERROR(clErr);
        ResultDatabase resultDB;

        // Run the benchmark

int main(int argc, char *argv[])
{
    int numdev=0, totalnumdev=0, numtasks, mympirank, dest, source, rc,
        mypair=0, count, tag=2, mynoderank,myclusterrank,nodenprocs;
    int *grp1, *grp2;
    int mygrprank,grpnumtasks;
    MPI_Group orig_group,bmgrp;
    MPI_Comm bmcomm,nlrcomm;
    ResultDatabase resultDB,resultDBWU,resultDB1;
    OptionParser op;
    ParallelResultDatabase pardb, pardb1;
    bool amGPUTask = false;
    volatile unsigned long long *mpidone;
    int i,shmid;

    /* Allocate System V shared memory */

    MPI_Init(&argc,&argv);
    MPI_Comm_size(MPI_COMM_WORLD, &numtasks);
    MPI_Comm_rank(MPI_COMM_WORLD, &mympirank);
    MPI_Comm_group(MPI_COMM_WORLD, &orig_group);


    //Add shared options to the parser
    op.addOption("device", OPT_VECINT, "0", "specify device(s) to run on",
		    'd');
    op.addOption("verbose", OPT_BOOL, "", "enable verbose output", 'v');
    op.addOption("quiet", OPT_BOOL, "",
		    "write minimum necessary to standard output", 'q');
    op.addOption("passes", OPT_INT, "10", "specify number of passes", 'z');
    op.addOption("size", OPT_VECINT, "1", "specify problem size", 's');
    op.addOption("time", OPT_INT, "5", "specify running time in miuntes", 't');
    op.addOption("outputFile", OPT_STRING, "output.txt", "specify output file",
       'o');
    op.addOption("infoDevices", OPT_BOOL, "", "show summary info for available devices",
       'i');
    op.addOption("fullInfoDevices", OPT_BOOL, "", "show full info for available devices");
    op.addOption("MPIminmsg", OPT_INT, "0", "specify minimum MPI message size");
    op.addOption("MPImaxmsg", OPT_INT, "16384",
                    "specify maximum MPI message size");
    op.addOption("MPIiter", OPT_INT, "1000",
                    "specify number of MPI benchmark iterations for each size");
    op.addOption("platform", OPT_INT, "0", "specify platform for device selection", 'y');

    if (!op.parse(argc, argv))
    {
        if (mympirank == 0)
            op.usage();
        MPI_Finalize();
        return 0;
    }

    int npasses = op.getOptionInt("passes");

    //our simple mapping
    NodeInfo NI;
    mynoderank = NI.nodeRank();         // rank of my process within the node
    myclusterrank = NI.clusterRank();   // cluster (essentially, node) id
    MPI_Comm smpcomm = NI.getSMPComm();

    if(mynoderank==0){
        shmid = shmget(IPC_PRIVATE,
                 sizeof(unsigned long long),
                 (IPC_CREAT | 0600));
    }

    MPI_Bcast(&shmid, 1, MPI_INT, 0, NI.getSMPComm());

    mpidone = ((volatile unsigned long long*) shmat(shmid, 0, 0));
    if (mynoderank == 0)
        shmctl(shmid, IPC_RMID, 0);
    *mpidone = 0;

    nlrcomm = NI.getNLRComm(); // communcator of all the lowest rank processes
                               // on all the nodes
    int numnodes = NI.numNodes();
    if ( numnodes%2!=0 )
    {
        if(mympirank==0)
            printf("\nThis test needs an even number of nodes\n");
        MPI_Finalize();
	exit(0);
    }
    int nodealr = NI.nodeALR();

    nodenprocs=NI.nodeNprocs();

    // determine how many GPU devices we are to use
    int devsPerNode = op.getOptionVecInt( "device" ).size();
    //cout<<mympirank<<":numgpus="<<devsPerNode<<endl;

    // if there are as many or more devices as the nprocs, only use half of
    // the nproc
    if ( devsPerNode >= nodenprocs ) devsPerNode = nodenprocs/2;

    numdev = (mynoderank == 0) ? devsPerNode : 0;
    MPI_Allreduce(&numdev, &totalnumdev, 1, MPI_INT, MPI_SUM,
                    MPI_COMM_WORLD);
    numdev = devsPerNode;

//.........这里部分代码省略.........

void
DoTest( const char* timerDesc, ResultDatabase& resultDB, OptionParser& opts )
{
    StencilFactory<T>* stdStencilFactory = NULL;
    Stencil<T>* stdStencil = NULL;
    StencilFactory<T>* testStencilFactory = NULL;
    Stencil<T>* testStencil = NULL;

    //try
    {
        stdStencilFactory = new HostStencilFactory<T>;
        testStencilFactory = new MICStencilFactory<T>;
        assert( (stdStencilFactory != NULL) && (testStencilFactory != NULL) );

        // do a sanity check on option values
        CheckOptions( opts );
        stdStencilFactory->CheckOptions( opts );
        testStencilFactory->CheckOptions( opts );

        // extract and validate options
        std::vector<long long> arrayDims = opts.getOptionVecInt( "customSize" );
        if( arrayDims.size() != 2 )
        {
            cerr << "Dim size: " << arrayDims.size() << "\n";
            //throw InvalidArgValue( "all overall dimensions must be positive" );
        }
        if (arrayDims[0] == 0) // User has not specified a custom size
        {
            const int probSizes[4] = { 768, 1408, 2048, 4096 };
            int sizeClass = opts.getOptionInt("size");
            if (!(sizeClass >= 0 && sizeClass < 5))
            {
                //throw InvalidArgValue( "Size class must be between 1-4" );
            }
            arrayDims[0] = arrayDims[1] =probSizes[sizeClass - 1];
        }

        long int seed = (long)opts.getOptionInt( "seed" );
        bool beVerbose = opts.getOptionBool( "verbose" );
        unsigned int nIters = (unsigned int)opts.getOptionInt( "num-iters" );
        double valErrThreshold = (double)opts.getOptionFloat( "val-threshold" );
        unsigned int nValErrsToPrint = (unsigned int)opts.getOptionInt( "val-print-limit" );

#if defined(PARALLEL)
        unsigned int haloWidth = (unsigned int)opts.getOptionInt( "iters-per-exchange" );
#else
        unsigned int haloWidth = 1;
#endif // defined(PARALLEL)

        float haloVal = (float)opts.getOptionFloat( "haloVal" );

        // build a description of this experiment
        std::ostringstream experimentDescriptionStr;
        experimentDescriptionStr 
            << nIters << ':'
            << arrayDims[0] << 'x' << arrayDims[1] << ':'
            << LROWS << 'x' << LCOLS;

        unsigned int nPasses =(unsigned int)opts.getOptionInt( "passes" );
         unsigned long npts = (arrayDims[0] + 2*haloWidth - 2) * 
                                     (arrayDims[1] + 2*haloWidth - 2); 

unsigned long nflops = npts * 11 * nIters;
cout<<"flops are = "<<nflops<<endl;

        // compute the expected result on the host
#if defined(PARALLEL)
        int cwrank;
        MPI_Comm_rank( MPI_COMM_WORLD, &cwrank );
        if( cwrank == 0 )
        {
#endif // defined(PARALLEL)
        std::cout << "\nPerforming stencil operation on host for later comparison with MIC output\n"
            << "Depending on host capabilities, this may take a while."
            << std::endl;
#if defined(PARALLEL)
        }
#endif // defined(PARALLEL)
        Matrix2D<T> exp( arrayDims[0] + 2*haloWidth, 
                            arrayDims[1] + 2*haloWidth );
        Initialize<T> init( seed,
                        haloWidth,
                        haloVal );	

        init( exp );
        if( beVerbose )
        {
            std::cout << "initial state:\n" << exp << std::endl;
        }
        Stencil<T>* stdStencil = stdStencilFactory->BuildStencil( opts );
        (*stdStencil)( exp, nIters );
        if( beVerbose )
        {
            std::cout << "expected result:\n" << exp << std::endl;
        }
	

        // compute the result on the MIC device
        Matrix2D<T> data( arrayDims[0] + 2*haloWidth, 
                            arrayDims[1] + 2*haloWidth );
//.........这里部分代码省略.........

// ****************************************************************************
// Function: main
//
// Purpose:
//   The main function takes care of initialization (device and MPI),  then
//   performs the benchmark and prints results.
//
// Arguments:
//
//
// Programmer: Jeremy Meredith
// Creation:
//
// Modifications:
//   Jeremy Meredith, Wed Nov 10 14:20:47 EST 2010
//   Split timing reports into detailed and summary.  For serial code, we
//   report all trial values, and for parallel, skip the per-process vals.
//   Also detect and print outliers from parallel runs.
//
// ****************************************************************************
int main(int argc, char *argv[])
{
    int ret = 0;
    bool noprompt = false;

    try
    {
#ifdef PARALLEL
        int rank, size;
        MPI_Init(&argc,&argv);
        MPI_Comm_size(MPI_COMM_WORLD, &size);
        MPI_Comm_rank(MPI_COMM_WORLD, &rank);
        cerr << "MPI Task " << rank << "/" << size - 1 << " starting....\n";
#endif

        // Get args
        OptionParser op;
       
        //Add shared options to the parser
        op.addOption("device", OPT_VECINT, "0",
                "specify device(s) to run on", 'd');
        op.addOption("verbose", OPT_BOOL, "", "enable verbose output", 'v');
        op.addOption("passes", OPT_INT, "10", "specify number of passes", 'n');
        op.addOption("size", OPT_INT, "1", "specify problem size", 's');
        op.addOption("infoDevices", OPT_BOOL, "",
                "show info for available platforms and devices", 'i');
        op.addOption("quiet", OPT_BOOL, "", "write minimum necessary to standard output", 'q');
#ifdef _WIN32
        op.addOption("noprompt", OPT_BOOL, "", "don't wait for prompt at program exit");
#endif

        addBenchmarkSpecOptions(op);

        if (!op.parse(argc, argv))
        {
#ifdef PARALLEL
            if (rank == 0)
                op.usage();
            MPI_Finalize();
#else
            op.usage();
#endif
            return (op.HelpRequested() ? 0 : 1);
        }
        
        bool verbose = op.getOptionBool("verbose");
        bool infoDev = op.getOptionBool("infoDevices");
#ifdef _WIN32
        noprompt = op.getOptionBool("noprompt");
#endif

        int device;
#ifdef PARALLEL
        NodeInfo ni;
        int myNodeRank = ni.nodeRank();
        vector<long long> deviceVec = op.getOptionVecInt("device");
        if (myNodeRank >= deviceVec.size()) {
            // Default is for task i to test device i
            device = myNodeRank;
        } else {
            device = deviceVec[myNodeRank];
        }
#else
        device = op.getOptionVecInt("device")[0];
#endif
        int deviceCount;
        cudaGetDeviceCount(&deviceCount);
        if (device >= deviceCount) {
            cerr << "Warning: device index: " << device <<
            " out of range, defaulting to device 0.\n";
            device = 0;
        }

        // Initialization
        EnumerateDevicesAndChoose(device, infoDev);
        if( infoDev )
        {
            return 0;
        }
        ResultDatabase resultDB;
//.........这里部分代码省略.........

void 
init(OptionParser& op, bool _do_dp)
{
    cl_int err;

    do_dp = _do_dp;
    
    if (!fftCtx) {
        // first get the device
        int device, platform = op.getOptionInt("platform");
        if (op.getOptionVecInt("device").size() > 0) {
            device = op.getOptionVecInt("device")[0];
        }
        else {
            device = 0;
        }
        fftDev = ListDevicesAndGetDevice(platform, device);

        // now get the context
        fftCtx = clCreateContext(NULL, 1, &fftDev, NULL, NULL, &err);
        CL_CHECK_ERROR(err);
    }

    if (!fftQueue) {
        // get a queue
        fftQueue = clCreateCommandQueue(fftCtx, fftDev, CL_QUEUE_PROFILING_ENABLE,
                                        &err);
        CL_CHECK_ERROR(err);
    }
    
    // create the program...
    fftProg = clCreateProgramWithSource(fftCtx, 1, &cl_source_fft, NULL, &err);
    CL_CHECK_ERROR(err);
    
    // ...and build it
    string args = " -cl-mad-enable ";
    if (op.getOptionBool("use-native")) {
        args += " -cl-fast-relaxed-math ";
    }
    if (!do_dp) {
        args += " -DSINGLE_PRECISION ";
    }
    else if (checkExtension(fftDev, "cl_khr_fp64")) {
        args += " -DK_DOUBLE_PRECISION ";
    }
    else if (checkExtension(fftDev, "cl_amd_fp64")) {
        args += " -DAMD_DOUBLE_PRECISION ";
    }
    err = clBuildProgram(fftProg, 0, NULL, args.c_str(), NULL, NULL);
    {
        char* log = NULL;
        size_t bytesRequired = 0;
        err = clGetProgramBuildInfo(fftProg, 
                                    fftDev, 
                                    CL_PROGRAM_BUILD_LOG,
                                    0,
                                    NULL,
                                    &bytesRequired );
        log = (char*)malloc( bytesRequired + 1 );
        err = clGetProgramBuildInfo(fftProg, 
                                    fftDev, 
                                    CL_PROGRAM_BUILD_LOG,
                                    bytesRequired,
                                    log,
                                    NULL );
        std::cout << log << std::endl;
        free( log );
    }
    if (err != CL_SUCCESS) {
        char log[50000];
        size_t retsize = 0;
        err = clGetProgramBuildInfo(fftProg, fftDev, CL_PROGRAM_BUILD_LOG,
                                    50000*sizeof(char),  log, &retsize);
        CL_CHECK_ERROR(err);
        cout << "Retsize: " << retsize << endl;
        cout << "Log: " << log << endl;
        dumpPTXCode(fftCtx, fftProg, "oclFFT");
        exit(-1);
    }
    else {
        // dumpPTXCode(fftCtx, fftProg, "oclFFT");
    }

    // Create kernel for forward FFT
    fftKrnl = clCreateKernel(fftProg, "fft1D_512", &err);
    CL_CHECK_ERROR(err);
    // Create kernel for inverse FFT
    ifftKrnl = clCreateKernel(fftProg, "ifft1D_512", &err);
    CL_CHECK_ERROR(err);
    // Create kernel for check
    chkKrnl = clCreateKernel(fftProg, "chk1D_512", &err);
    CL_CHECK_ERROR(err);
}