/*************************************************************************/
/*                                                                       */
/*  Copyright (c) 1994 Stanford University                               */
/*                                                                       */
/*  All rights reserved.                                                 */
/*                                                                       */
/*  Permission is given to use, copy, and modify this software for any   */
/*  non-commercial purpose as long as this copyright notice is not       */
/*  removed.  All other uses, including redistribution in whole or in    */
/*  part, are forbidden without prior written permission.                */
/*                                                                       */
/*  This software is provided with absolutely no warranty and no         */
/*  support.                                                             */
/*                                                                       */
/*************************************************************************/

/*************************************************************************/
/*                                                                       */
/*  SPLASH Ocean Code                                                    */
/*                                                                       */
/*  This application studies the role of eddy and boundary currents in   */
/*  influencing large-scale ocean movements.  This implementation uses   */
/*  dynamically allocated four-dimensional arrays for grid data storage. */
/*                                                                       */
/*  Main parameters are :                                                */
/*                                                                       */
/*     - M : Simulate MxM ocean. M must be (power of 2) +2.              */
/*     - N : N = number of threads. N must be power of 4.                */
/*     - E : E = error tolerance for iterative relaxation.               */
/*     - R : R = distance between grid points in meters.                 */
/*     - T : T = timestep in seconds.                                    */
/*                                                                       */
/*************************************************************************/

///////////////////////////////////////////////////////////////////////////
// This is the porting of the SLASH Ocean application on the GIET-VM
// operating system, for the TSAR manycores architecture.
// Done by Alain greiner (march 2016).
///////////////////////////////////////////////////////////////////////////

#define DEFAULT_M        258
#define DEFAULT_E        1e-7
#define DEFAULT_T    28800.0
#define DEFAULT_R    20000.0
#define UP               0
#define DOWN             1
#define LEFT             2
#define RIGHT            3
#define UPLEFT           4
#define UPRIGHT          5
#define DOWNLEFT         6
#define DOWNRIGHT        7
#define PAGE_SIZE     4096
#define MAX_THREADS   1024

#include "decs.h"

#include <user_lock.h>
#include <user_barrier.h>
#include <stdio.h>
#include <stdlib.h>
#include <malloc.h>

// GIET specific global variables

pthread_t                thread_kernel[MAX_THREADS];  // identifier defined by the kernel
long                     thread_user[MAX_THREADS];    // user index = x*yprocs + y

user_lock_t              tty_lock;

giet_sqt_barrier_t       barrier;

sqt_lock_t               id_lock;
sqt_lock_t               psiai_lock;
sqt_lock_t               psibi_lock;
sqt_lock_t               done_lock;
sqt_lock_t               error_lock;
sqt_lock_t               bar_lock;

// OCEAN global variables

struct multi_struct*     multi;
struct global_struct*    global;
struct Global_Private**  gps;              // array of pointers[nprocs]

double ****              psi;
double ****              psim;
double ***               psium;
double ***               psilm;
double ***               psib;
double ***               ga;
double ***               gb;
double ****              work1;
double ***               work2;
double ***               work3;
double ****              work4;
double ****              work5;
double ***               work6;
double ****              work7;
double ****              temparray;
double ***               tauz;
double ***               oldga;
double ***               oldgb;

double *                 f;
double ****              q_multi;
double ****              rhs_multi;

long                     xprocs;          // number of blocks in a row (one block per proc)
long                     yprocs;          // number of blocks in a col (one block per proc)
long                     nprocs;          // total number of blocks (number of procs)

const double             h1 = 1000.0;
const double             h3 = 4000.0;
const double             h = 5000.0;
const double             lf = -5.12e11;
double                   res = DEFAULT_R;
double                   dtau = DEFAULT_T;
const double             f0 = 8.3e-5;
const double             beta = 2.0e-11;
const double             gpsr = 0.02;

long                     im = DEFAULT_M;  // number of grid points in a row
long                     jm = DEFAULT_M;  // number of grid points in a column

double                   tolerance = DEFAULT_E;
double                   eig2;
double                   ysca;
long                     jmm1;
const double             pi = 3.141592653589793;
const double             dt0 = 0.5e-4;
const double             outday0 = 1.0;
const double             outday1 = 2.0;
const double             outday2 = 2.0;
const double             outday3 = 2.0;
double                   factjacob;
double                   factlap;
long                     numlev;
long *                   imx;             // array[numlev]
long *                   jmx;             // array[numlev]
double *                 lev_res;         // array[numlev]
double *                 lev_tol;         // array[numlev]
const double             maxwork = 10000.0;

double *                 i_int_coeff;
double *                 j_int_coeff;
long *                   xpts_per_proc;
long *                   ypts_per_proc;
long                     minlevel;

///////////////////////////////////////////
__attribute__ ((constructor)) int main()
///////////////////////////////////////////
{
    long      x;               // index to scan xprocs
    long      y;               // index to scan yprocs
    long      i;               // index to scan numlev
    long      j;               // index to scan phases
    long      x_part;
    long      y_part;
    long      d_size;
    long      itemp;
    long      jtemp;
    long      start_time;
    long      init_time;

    start_time = giet_proctime();

    // compute number of threads : nprocs
    // as we want one thread per processor, it depends on the
    // hardware architecture x_size / y_size / procs per cluster
    unsigned int mesh_x_size;
    unsigned int mesh_y_size;
    unsigned int procs_per_cluster;

    giet_procs_number(&mesh_x_size, &mesh_y_size, &procs_per_cluster);
    nprocs = mesh_x_size * mesh_y_size * procs_per_cluster;

    giet_pthread_assert( (procs_per_cluster == 1) || (procs_per_cluster == 4),
                         "[OCEAN ERROR] number of procs per cluster must be 1 or 4");

    giet_pthread_assert( (mesh_x_size == 1) || (mesh_x_size == 2) || (mesh_x_size == 4) ||
                         (mesh_x_size == 8) || (mesh_y_size == 16),
                         "[OCEAN ERROR] mesh_x_size must be 1,2,4,8,16");

    giet_pthread_assert( (mesh_y_size == 1) || (mesh_y_size == 2) || (mesh_y_size == 4) ||
                         (mesh_y_size == 8) || (mesh_y_size == 16),
                         "[OCEAN ERROR] mesh_y_size must be 1,2,4,8,16");

    giet_pthread_assert( (mesh_y_size == mesh_y_size ),
                         "[OCEAN ERROR] mesh_y_size and mesh_y_size must be equal");

    // check the ocean size : M*M grid / (M-2) mut be power of 2

    giet_pthread_assert( (im == 34)  || (im == 66)  || (im == 130) ||
                         (im == 258) || (im == 514) || (im == 1026),
                         "[OCEAN ERROR] grid side must be 34,66,130,258,514,1026");
    
    // initialise distributed heap
    for ( x = 0 ; x < mesh_x_size ; x++ ) 
    {
        for ( y = 0 ; y < mesh_y_size ; y++ ) 
        {
            heap_init( x , y );
        }
    }

    // allocate shared TTY & initialise tty_lock 
    giet_tty_alloc( 1 );     
    lock_init( &tty_lock);

    giet_tty_printf("\n[OCEAN] simulation with W-cycle multigrid solver\n"
                    "    Processors                         : %d x %d x %d\n"
                    "    Grid size                          : %d x %d\n",
                    mesh_x_size , mesh_y_size , procs_per_cluster, im , jm );

    // initialise distributed barrier
    sqt_barrier_init( &barrier , mesh_x_size , mesh_y_size , procs_per_cluster );

    // initialize distributed locks
    sqt_lock_init( &id_lock    , mesh_x_size , mesh_y_size , procs_per_cluster );
    sqt_lock_init( &psiai_lock , mesh_x_size , mesh_y_size , procs_per_cluster );
    sqt_lock_init( &psibi_lock , mesh_x_size , mesh_y_size , procs_per_cluster );
    sqt_lock_init( &done_lock  , mesh_x_size , mesh_y_size , procs_per_cluster );
    sqt_lock_init( &error_lock , mesh_x_size , mesh_y_size , procs_per_cluster );
    sqt_lock_init( &bar_lock   , mesh_x_size , mesh_y_size , procs_per_cluster );

    // allocate thread_kernel[] array : thread identidiers defined by the kernel
    pthread_t* thread_kernel = malloc( nprocs * sizeof(pthread_t) );

    // allocate thread_user[] array : continuous thread index defined by the user
    long* thread_user = malloc( nprocs * sizeof(unsigned int) );

    // compute number of blocks per row and per column: nprocs = xprocs * yprocs
    if (procs_per_cluster == 1)
    {
        xprocs = mesh_x_size;
        yprocs = mesh_y_size;
    }
    else
    {
        xprocs = mesh_x_size*2;
        yprocs = mesh_y_size*2;
    }

    // compute numlev 
    minlevel = 0;
    itemp = 1;
    jtemp = 1;
    numlev = 0;
    minlevel = 0;
    while (itemp < (im - 2)) 
    {
        itemp = itemp * 2;
        jtemp = jtemp * 2;
        if ((itemp / yprocs > 1) && (jtemp / xprocs > 1)) 
        {
            numlev++;
        }
    }

    giet_pthread_assert( (numlev > 0),
                         "[OCEAN ERROR] at least 2*2 grid points per processor");

    // allocate in cluster(0,0) arrays indexed by numlev
    imx           = (long *)  malloc(numlev * sizeof(long));
    jmx           = (long *)  malloc(numlev * sizeof(long));
    lev_res       = (double *)malloc(numlev * sizeof(double));
    lev_tol       = (double *)malloc(numlev * sizeof(double));
    i_int_coeff   = (double *)malloc(numlev * sizeof(double));
    j_int_coeff   = (double *)malloc(numlev * sizeof(double));
    xpts_per_proc = (long *)  malloc(numlev * sizeof(long));
    ypts_per_proc = (long *)  malloc(numlev * sizeof(long));
    
    // initialize these arrays
    imx[numlev - 1]     = im;
    jmx[numlev - 1]     = jm;
    lev_res[numlev - 1] = res;
    lev_tol[numlev - 1] = tolerance;

    for (i = numlev - 2; i >= 0; i--) 
    {
        imx[i] = ((imx[i + 1] - 2) / 2) + 2;
        jmx[i] = ((jmx[i + 1] - 2) / 2) + 2;
        lev_res[i] = lev_res[i + 1] * 2;
    }

    for (i = 0; i < numlev; i++) 
    {
        xpts_per_proc[i] = (jmx[i] - 2) / xprocs;
        ypts_per_proc[i] = (imx[i] - 2) / yprocs;
    }
    for (i = numlev - 1; i >= 0; i--) 
    {
        if ((xpts_per_proc[i] < 2) || (ypts_per_proc[i] < 2)) 
        {
            minlevel = i + 1;
            break;
        }
    }

    // allocate in cluster(0,0) arrays of pointers ****
    d_size     = nprocs * sizeof(double ***);
    psi        = (double ****)malloc(d_size);
    psim       = (double ****)malloc(d_size);
    work1      = (double ****)malloc(d_size);
    work4      = (double ****)malloc(d_size);
    work5      = (double ****)malloc(d_size);
    work7      = (double ****)malloc(d_size);
    temparray  = (double ****)malloc(d_size);

    // allocate in each cluster(x,y) arrays of pointers ****
    d_size = 2 * sizeof(double **);
    for (x = 0;  x < xprocs; x++) 
    {
        for (y = 0; y < yprocs; y++) 
        {
            long procid = y * xprocs + x;
            long cx     = (procs_per_cluster == 1) ? x : x>>1;
            long cy     = (procs_per_cluster == 1) ? y : y>>1;

            psi[procid]       = (double ***)remote_malloc(d_size , cx , cy);
            psim[procid]      = (double ***)remote_malloc(d_size , cx , cy);
            work1[procid]     = (double ***)remote_malloc(d_size , cx , cy);
            work4[procid]     = (double ***)remote_malloc(d_size , cx , cy);
            work5[procid]     = (double ***)remote_malloc(d_size , cx , cy);
            work7[procid]     = (double ***)remote_malloc(d_size , cx , cy);
            temparray[procid] = (double ***)remote_malloc(d_size , cx , cy);
        }
    }

    // allocate in cluster(0,0) arrays of pointers ***
    d_size = nprocs * sizeof(double **);
    psium = (double ***)malloc(d_size);
    psilm = (double ***)malloc(d_size);
    psib  = (double ***)malloc(d_size);
    ga    = (double ***)malloc(d_size);
    gb    = (double ***)malloc(d_size);
    work2 = (double ***)malloc(d_size);
    work3 = (double ***)malloc(d_size);
    work6 = (double ***)malloc(d_size);
    tauz  = (double ***)malloc(d_size);
    oldga = (double ***)malloc(d_size);
    oldgb = (double ***)malloc(d_size);

    // allocate in cluster(0,0) array of pointers gps[nprocs]
    gps = (struct Global_Private**)malloc((nprocs+1)*sizeof(struct Global_Private*));

    // allocate in each cluster(x,y) the gps[procid] structures
    for (x = 0; x < xprocs; x++) 
    {
        for (y = 0; y < yprocs; y++) 
        {
            long procid = y * xprocs + x;
            long cx     = (procs_per_cluster == 1) ? x : x>>1;
            long cy     = (procs_per_cluster == 1) ? y : y>>1;

            gps[procid] = (struct Global_Private*)remote_malloc(
                             sizeof(struct Global_Private) , cx , cy);

            gps[procid]->rel_num_x = (long *)remote_malloc(numlev * sizeof(long) , cx , cy);
            gps[procid]->rel_num_y = (long *)remote_malloc(numlev * sizeof(long) , cx , cy);
            gps[procid]->eist      = (long *)remote_malloc(numlev * sizeof(long) , cx , cy);
            gps[procid]->ejst      = (long *)remote_malloc(numlev * sizeof(long) , cx , cy);
            gps[procid]->oist      = (long *)remote_malloc(numlev * sizeof(long) , cx , cy);
            gps[procid]->ojst      = (long *)remote_malloc(numlev * sizeof(long) , cx , cy);
            gps[procid]->rlist     = (long *)remote_malloc(numlev * sizeof(long) , cx , cy);
            gps[procid]->rljst     = (long *)remote_malloc(numlev * sizeof(long) , cx , cy);
            gps[procid]->rlien     = (long *)remote_malloc(numlev * sizeof(long) , cx , cy);
            gps[procid]->rljen     = (long *)remote_malloc(numlev * sizeof(long) , cx , cy);
            gps[procid]->multi_time    = 0;
            gps[procid]->total_time    = 0;
            gps[procid]->sync_time     = 0;
            gps[procid]->steps_time[0] = 0;
            gps[procid]->steps_time[1] = 0;
            gps[procid]->steps_time[2] = 0;
            gps[procid]->steps_time[3] = 0;
            gps[procid]->steps_time[4] = 0;
            gps[procid]->steps_time[5] = 0;
            gps[procid]->steps_time[6] = 0;
            gps[procid]->steps_time[7] = 0;
            gps[procid]->steps_time[8] = 0;
            gps[procid]->steps_time[9] = 0;
        }
    }

    ////////////
    subblock();

    x_part = (jm - 2) / xprocs + 2;   // nunber of grid points in block row
    y_part = (im - 2) / yprocs + 2;   // nunber of grid points in block column

    d_size = x_part * y_part * sizeof(double) + y_part * sizeof(double *);

    global = (struct global_struct *)malloc(sizeof(struct global_struct));

    // allocate in each cluster(x,y) the arrays of pointers ** 
    for (x = 0; x < xprocs; x++) 
    {
        for (y = 0; y < yprocs; y++) 
        {
            long procid = y * xprocs + x;
            long cx     = (procs_per_cluster == 1) ? x : x>>1;
            long cy     = (procs_per_cluster == 1) ? y : y>>1;

            psi[procid][0]       = (double **)remote_malloc(d_size , cx , cy);
            psi[procid][1]       = (double **)remote_malloc(d_size , cx , cy);
            psim[procid][0]      = (double **)remote_malloc(d_size , cx , cy);
            psim[procid][1]      = (double **)remote_malloc(d_size , cx , cy);
            psium[procid]        = (double **)remote_malloc(d_size , cx , cy);
            psilm[procid]        = (double **)remote_malloc(d_size , cx , cy);
            psib[procid]         = (double **)remote_malloc(d_size , cx , cy);
            ga[procid]           = (double **)remote_malloc(d_size , cx , cy);
            gb[procid]           = (double **)remote_malloc(d_size , cx , cy);
            work1[procid][0]     = (double **)remote_malloc(d_size , cx , cy);
            work1[procid][1]     = (double **)remote_malloc(d_size , cx , cy);
            work2[procid]        = (double **)remote_malloc(d_size , cx , cy);
            work3[procid]        = (double **)remote_malloc(d_size , cx , cy);
            work4[procid][0]     = (double **)remote_malloc(d_size , cx , cy);
            work4[procid][1]     = (double **)remote_malloc(d_size , cx , cy);
            work5[procid][0]     = (double **)remote_malloc(d_size , cx , cy);
            work5[procid][1]     = (double **)remote_malloc(d_size , cx , cy);
            work6[procid]        = (double **)remote_malloc(d_size , cx , cy);
            work7[procid][0]     = (double **)remote_malloc(d_size , cx , cy);
            work7[procid][1]     = (double **)remote_malloc(d_size , cx , cy);
            temparray[procid][0] = (double **)remote_malloc(d_size , cx , cy);
            temparray[procid][1] = (double **)remote_malloc(d_size , cx , cy);
            tauz[procid]         = (double **)remote_malloc(d_size , cx , cy);
            oldga[procid]        = (double **)remote_malloc(d_size , cx , cy);
            oldgb[procid]        = (double **)remote_malloc(d_size , cx , cy);
        }
    }

    f = (double *)malloc(im*sizeof(double));

    multi = (struct multi_struct *)malloc(sizeof(struct multi_struct));
    
    // allocate memory for q_multi and rhs_multi 
    d_size = numlev * sizeof(double **);
    if (numlev % 2 == 1) // add an extra pointer for double word alignment
    {
        d_size += sizeof(double **);
    }

    for (i = 0; i < numlev; i++) 
    {
        d_size += ((imx[i] - 2) / yprocs + 2) * ((jmx[i] - 2) / xprocs + 2) * sizeof(double) + 
                  ((imx[i] - 2) / yprocs + 2) * sizeof(double *);
    }
    d_size *= nprocs;
    if (nprocs % 2 == 1) // add an extra pointer for double word alignment
    {      
        d_size += sizeof(double ***);
    }

    d_size += nprocs * sizeof(double ***);
    q_multi   = (double ****)malloc( d_size );
    rhs_multi = (double ****)malloc( d_size );

    //////////
    link_all();

    multi->err_multi = 0.0;
    i_int_coeff[0] = 0.0;
    j_int_coeff[0] = 0.0;

    for (i = 0; i < numlev; i++)
    {
        i_int_coeff[i] = 1.0 / (imx[i] - 1);
        j_int_coeff[i] = 1.0 / (jmx[i] - 1);
    }

    global->psibi      = 0.0;

    factjacob = -1. / (12. * res * res);
    factlap   = 1. / (res * res);
    eig2 = -h * f0 * f0 / (h1 * h3 * gpsr);

    jmm1 = jm - 1;
    ysca = ((double) jmm1) * res;

    im = (imx[numlev-1]-2)/yprocs + 2;
    jm = (jmx[numlev-1]-2)/xprocs + 2;

    init_time = giet_proctime() - start_time;

    printf("\n[OCEAN] initialisation completed / start parallel execution\n");

    ///////////////////////////////////////////////////
    // launch (N-1) other threads to execute slave()
    ///////////////////////////////////////////////////

    for (i = 1 ; i < nprocs ; i++) 
    {
        thread_user[i] = i;
        if (giet_pthread_create( &thread_kernel[i], 
                                 NULL, 
                                 &slave, 
                                 &thread_user[i] ))
        {
            giet_pthread_exit("[OCEAN ERROR] in giet_pthread_create()\n");
        }
    }

    // main itself execute slave()
    thread_user[0] = 0;
    slave( &thread_user[0] );

    // wait other threads completion
    for ( i = 1 ; i < nprocs ; i++ )
    {
        if ( giet_pthread_join( thread_kernel[i], NULL ) )
        {
            giet_pthread_exit( "[OCEAN ERROR] in giet_pthread_join()\n" );
        }
    }

    ///////////////////////////////////////////////
    // instrumentation (display & save on disk)
    ///////////////////////////////////////////////
 
    char string[256];

    snprintf( string , 256 , "/home/ocean_%d_%d_%d" , 
              mesh_x_size , mesh_y_size , procs_per_cluster );

    // open instrumentation file
    unsigned int fd = giet_fat_open( string , O_CREAT );
    if ( fd < 0 ) 
    { 
        printf("\n[OCEAN ERROR] cannot open instrumentation file %s\n", string );
        giet_pthread_exit( NULL );
    }

    snprintf( string , 256 , "\n--- ocean : (%dx%dx%d) procs on (%dx%d) grid ---\n",
                             mesh_x_size, mesh_y_size, procs_per_cluster , 
                             DEFAULT_M , DEFAULT_M );

    giet_tty_printf( "%s" , string );
    giet_fat_fprintf( fd , "%s" , string );

    // compute instrumentation results
    long min_total = gps[0]->total_time;
    long max_total = gps[0]->total_time;
    long min_multi = gps[0]->multi_time;
    long max_multi = gps[0]->multi_time;
    long min_sync  = gps[0]->sync_time;
    long max_sync  = gps[0]->sync_time;

    for (i = 1 ; i < nprocs ; i++) 
    {
        if (gps[i]->total_time > max_total)  max_total = (gps[i]->total_time);
        if (gps[i]->total_time < min_total)  min_total = (gps[i]->total_time);
        if (gps[i]->multi_time > max_multi)  max_multi = (gps[i]->multi_time);
        if (gps[i]->multi_time < min_multi)  min_multi = (gps[i]->multi_time);
        if (gps[i]->sync_time  > max_sync )  max_sync  = (gps[i]->sync_time );
        if (gps[i]->sync_time  < min_sync )  min_sync  = (gps[i]->sync_time );
    }

    snprintf( string , 256 , "\n    Init Time    Total Time    Multi Time    Sync Time\n"
                             "MIN : %d | %d | %d | %d  (cycles)\n" 
                             "MAX : %d | %d | %d | %d  (cycles)\n",
                             (int)init_time, (int)min_total, (int)min_multi, (int)min_sync,
                             (int)init_time, (int)max_total, (int)max_multi, (int)max_sync );

    giet_tty_printf("%s" , string );
    giet_fat_fprintf( fd , "%s" , string );

    for (i = 0; i < 10; i++) 
    {
        long phase_time = 0;
        for (j = 0; j < nprocs; j++) 
        {
            phase_time += gps[j]->steps_time[i];
        }
        snprintf( string , 256 , " - Phase %d : %d cycles\n",
                                 (int)i , (int)(phase_time/nprocs) );
        giet_tty_printf("%s" , string );
        giet_fat_fprintf( fd , "%s" , string );
    }

    giet_pthread_exit("main completed");

    return 0;
    
}  // end main()


// Local Variables:
// tab-width: 4
// c-basic-offset: 4
// c-file-offsets:((innamespace . 0)(inline-open . 0))
// indent-tabs-mode: nil
// End:

// vim: filetype=cpp:expandtab:shiftwidth=4:tabstop=4:softtabstop=4