update_y_slave.cpp

#include "data.h"
#include "cmd_slave.h"
#include <fstream>
#include <vector>
#include <stdio.h>
#include <string.h>

#include "sagecal.h"
#include "utils.h"
#include "utils_dn.h"

using namespace std;
using namespace Data;

#ifndef LMCUT
#define LMCUT 40
#endif

int update_y_slave(dlg_app_info *app) {
    /*-----------------------------------------------input------------------------------------------------------------*/
    Data::IOData old_iodata, iodata;
    Data::LBeam beam;
    Data::MPIData mpiData;
    double *arho;
    openblas_set_num_threads(1);

    int flag = 0;
    app->inputs[0].read((char *)&flag, sizeof(int));

    if(flag==0) {
        load_iodata_dn_noprefix(&(app->inputs[0]), &old_iodata);
        load_mpidata_dn(&(app->inputs[1]), &mpiData);
    } else {
        load_iodata_dn(&(app->inputs[1]), &old_iodata);
        load_mpidata_dn_noprefix(&(app->inputs[0]), &mpiData);
    }

    load_share_iodata(Data::shareDir, old_iodata.msname, &iodata);
    cout << "[update_y_slave]========, " << iodata.msname <<",iodata.N/M/Mt/Nms:" << iodata.N << "/" << iodata.M  << "/" << iodata.Mt << "/" << iodata.Nms
         << ", iodata.freq0:" << iodata.freq0/ 1e6<< "Mhz" << endl;
    if (Data::doBeam) {
        load_share_beam(Data::shareDir, iodata.msname, &beam);
    }
    Data::freeData(old_iodata);

    int M = mpiData.Mo;
    int Mt = mpiData.M;

    if ((arho = (double *) calloc((size_t) mpiData.Mo, sizeof(double))) == 0) {
        fprintf(stderr, "%s: %d: no free memory\n", __FILE__, __LINE__);
        exit(1);
    }
    read_share_XYZ(Data::shareDir, iodata.msname, arho, mpiData.Mo, "arho");

    /*----------------------------------------------------------------------------------------------------------------*/
    clus_source_t *carr;
    read_sky_cluster(Data::SkyModel, Data::Clusters, &carr, &M, iodata.freq0, iodata.ra0, iodata.dec0, Data::format);
    if (M <= 0) {
        fprintf(stderr, "%s: %d: no clusters to solve\n", __FILE__, __LINE__);
        exit(1);
    } else {
        printf("%s:Got %d clusters\n", __FILE__, M);
    }
    /* update cluster array with correct pointers to parameters */
    int ci = 0, ck = 0,cj = 0;
    for (ci = 0; ci < M; ci++) {
        if ((carr[ci].p = (int *) calloc((size_t) carr[ci].nchunk, sizeof(int))) == 0) {
            fprintf(stderr, "%s: %d: no free memory\n", __FILE__, __LINE__);
            exit(1);
        }
        for (ck = 0; ck < carr[ci].nchunk; ck++) {
            carr[ci].p[ck] = cj * 8 * iodata.N;
            cj++;
        }
    }

    double *xbackup = 0;
    if (iodata.Nchan > 1 || Data::whiten) {
        if ((xbackup = (double *) calloc((size_t) iodata.Nbase * 8 * iodata.tilesz, sizeof(double))) == 0) {
            fprintf(stderr, "%s: %d: no free memory\n", __FILE__, __LINE__);
            exit(1);
        }
        read_share_XYZ(Data::shareDir, iodata.msname, xbackup, iodata.Nbase * 8 * iodata.tilesz, "xbackup");
    }

    /*----------------------------------------------------------------------------------------------------------------*/
    complex double *coh;

    if ((coh = (complex double *) calloc((size_t)(iodata.M * iodata.Nbase * iodata.tilesz * 4), sizeof(complex double)))==0) {
        fprintf(stderr, "%s: %d: no free memory\n", __FILE__, __LINE__);
        exit(1);
    }
    load_share_coh(Data::shareDir, iodata.msname, &iodata, coh);

    /* ADMM memory */
    double *Z, *Y;
    /* Z: (store B_f Z) 2Nx2 x M */
    if ((Z = (double *) calloc((size_t) iodata.N * 8 * Mt, sizeof(double))) == 0) {
        fprintf(stderr, "%s: %d: no free memory\n", __FILE__, __LINE__);
        exit(1);
    }
    /* Y, 2Nx2 , M times */
    if ((Y = (double *) calloc((size_t) iodata.N * 8 * Mt, sizeof(double))) == 0) {
        fprintf(stderr, "%s: %d: no free memory\n", __FILE__, __LINE__);
        exit(1);
    }
    /* primal residual J-BZ */
    double *pres;
    if ((pres = (double *) calloc((size_t) iodata.N * 8 * Mt, sizeof(double))) == 0) {
        fprintf(stderr, "%s: %d: no free memory\n", __FILE__, __LINE__);
        exit(1);
    }
    double *p;
    /* parameters 8*N*M ==> 8*N*Mt */
    if ((p = (double *) calloc((size_t) iodata.N * 8 * Mt, sizeof(double))) == 0) {
        fprintf(stderr, "%s: %d: no free memory\n", __FILE__, __LINE__);
        exit(1);
    }

    read_share_XYZ(Data::shareDir, iodata.msname, Y, iodata.N * 8 * Mt, "Y");
    read_share_XYZ(Data::shareDir, iodata.msname, Z, iodata.N * 8 * Mt, "Z");
    read_share_XYZ(Data::shareDir, iodata.msname, pres, iodata.N * 8 * Mt, "pres");
    read_share_XYZ(Data::shareDir, iodata.msname, p, iodata.N * 8 * Mt, "p");

    /*----------------------------------------------------------------------------------------------------------------*/
    double res_0, res_1, res_00, res_01;
    double mean_nu;
    res_0 = res_1 = res_00 = res_01 = 0.0;
    int start_iter = 0, tilex = 0;

    load_share_res(Data::shareDir, iodata.msname, &start_iter, &res_0, &res_1, &res_00, &res_01, &mean_nu, &tilex);

    /*------------------------------------------processing-------------------------------------------------------------*/
    int admm = get_last_iter(app->uid);
    /* update Y_i <= Y_i + rho (J_i-B_i Z)
        since we already have Y_i + rho J_i, only need -rho (B_i Z) */
    ck = 0;
    for (ci = 0; ci < M; ci++) {
        if (arho[ci] > 0.0) {
            my_daxpy(iodata.N * 8 * carr[ci].nchunk, &Z[ck], -arho[ci], &Y[ck]);
        }
        ck += iodata.N * 8 * carr[ci].nchunk;
    }

    /* calculate primal residual J-BZ */
    my_dcopy(iodata.N * 8 * Mt, p, 1, pres, 1);
    my_daxpy(iodata.N * 8 * Mt, Z, -1.0, pres);

    /* primal residual : per one real parameter */
    /* to remove a load of network traffic and screen output, disable this info */
    if (Data::verbose) {
        cout << iodata.msname << ": ADMM : " << admm << " residual: primal="
             << my_dnrm2(iodata.N * 8 * Mt, pres) / sqrt((double) 8 * iodata.N * Mt) << ", initial=" << res_0
             << ", final=" << res_1 << endl;
    }

    write_share_XYZ(Data::shareDir, iodata.msname, Y, iodata.N * 8 * Mt, "Y");
    write_share_XYZ(Data::shareDir, iodata.msname, Z, iodata.N * 8 * Mt, "Z");
    write_share_XYZ(Data::shareDir, iodata.msname, pres, iodata.N * 8 * Mt, "pres");
    write_share_XYZ(Data::shareDir, iodata.msname, p, iodata.N * 8 * Mt, "p");

    /*--------------------------------------------output---------------------------------------------------------------*/
    /* if most data are flagged, only send the original Y we got at the beginning */
    /* for initial ADMM iteration, get back Y with common unitary ambiguity */

    flag = 2;
    app->outputs[0].write((char *)&flag, sizeof(int));
    dump_iodata_dn(&(app->outputs[0]), &iodata);
    app->outputs[0].write((char *)Y, sizeof(double) * iodata.N * 8 * Mt);
    app->outputs[0].write((char *)Z, sizeof(double) * iodata.N * 8 * Mt);

    /*------------------------------------free -----------------------------------------------------------------------*/

    exinfo_gaussian *exg;
    exinfo_disk *exd;
    exinfo_ring *exr;
    exinfo_shapelet *exs;
    for (ci = 0; ci < M; ci++) {
        free(carr[ci].ll);
        free(carr[ci].mm);
        free(carr[ci].nn);
        free(carr[ci].sI);
        free(carr[ci].p);
        free(carr[ci].ra);
        free(carr[ci].dec);
        for (cj = 0; cj < carr[ci].N; cj++) {
            /* do a proper typecast before freeing */
            switch (carr[ci].stype[cj]) {
                case STYPE_GAUSSIAN:
                    exg = (exinfo_gaussian *) carr[ci].ex[cj];
                    if (exg) free(exg);
                    break;
                case STYPE_DISK:
                    exd = (exinfo_disk *) carr[ci].ex[cj];
                    if (exd) free(exd);
                    break;
                case STYPE_RING:
                    exr = (exinfo_ring *) carr[ci].ex[cj];
                    if (exr) free(exr);
                    break;
                case STYPE_SHAPELET:
                    exs = (exinfo_shapelet *) carr[ci].ex[cj];
                    if (exs) {
                        if (exs->modes) {
                            free(exs->modes);
                        }
                        free(exs);
                    }
                    break;
                default:
                    break;
            }
        }
        free(carr[ci].ex);
        free(carr[ci].stype);
        free(carr[ci].sI0);
        free(carr[ci].f0);
        free(carr[ci].spec_idx);
        free(carr[ci].spec_idx1);
        free(carr[ci].spec_idx2);
    }
    free(carr);
    free(p);
    if (iodata.Nchan > 1 || Data::whiten) {
        free(xbackup);
    }
    free(Z);
    free(Y);
    free(coh);
    free(arho);
    if (!doBeam) {
        Data::freeData(iodata);
    } else {
        Data::freeData(iodata, beam);
    }
    delete[] mpiData.freqs;
    cout << "[update_y_slave]========, Done." << endl;
    return 0;
}