/*
 * Josh Holtrop
 * 2008-12-11
 * CS677 Final Project
 * This program implements a fractal-image generator and viewer
 * that uses OpenMPI and OpenMP.
 */

#include <SDL/SDL.h>
#include <mpi.h>
#include <sys/time.h>               /* struct timeval, gettimeofday() */
#include <iostream>
#include <omp.h>
#include "Computation.h"
#include "NewtonComputation.h"
#include "FatouComputation.h"
using namespace std;

#define PROGNAME "Josh's CS677 Final : MPI Fractal Generator"
#define getXVirt(x) (((x) - (width >> 1)) * zoom + x_center)
#define getYVirt(y) ((-((y) - (height >> 1))) * zoom + y_center)

/**************************************************************************
 * Utility functions                                                      *
 *************************************************************************/
bool createWindow(int width, int height,
                  SDL_Surface ** screen, Uint32 ** pixels);
void getSizes(int * rank, int * size, int * nprocs);
void draw(int rank, int world_size, int nprocs, int width, int height,
    Uint32 * pixels, Uint32 * taskVals, Computation * computation);
void sendWindowVals(double * winVals, int world_size);
inline void taskAllocate(int total_tasks, int total_workers, int this_id,
                         int * first_task_id, int * num);

/**************************************************************************
 * Global variables                                                       *
 *************************************************************************/
static double x_center = 0.0;
static double y_center = 0.0;
static double zoom = 1/300.0;
/* a "task" will be processing task_size pixels */
static int task_size = 100;

/**************************************************************************
 * This is the main entry point for our program                           *
 *************************************************************************/
int main(int argc, char * argv[])
{
    int width = 600;
    int height = 600;
    int my_rank = 0;
    int world_size = 0;
    int nprocs = 0;
    Computation * computation = NULL;
    int fractal_type = 0;
    bool display_times = false;

    SDL_Surface * screen;
    Uint32 * pixels;

    MPI_Init(&argc, &argv);
    for (int i = 1; i < argc; i++)
    {
        if (!strncmp(argv[i], "-w", 2))
        {
            width = atoi(strlen(argv[i]) > 2 ? argv[i] + 2 : argv[++i]);
        }
        else if (!strncmp(argv[i], "-h", 2))
        {
            height = atoi(strlen(argv[i]) > 2 ? argv[i] + 2 : argv[++i]);
        }
        else if (!strncmp(argv[i], "-t", 2))
        {
            fractal_type = atoi(strlen(argv[i]) > 2 ? argv[i] + 2 : argv[++i]);
        }
        else if (!strcmp(argv[i], "--times"))
        {
            display_times = true;
        }
        else if (!strcmp(argv[i], "--no-threads"))
        {
            omp_set_num_threads(1);
        }
        else if (!strncmp(argv[i], "-s", 2))
        {
            task_size = atoi(strlen(argv[i]) > 2 ? argv[i] + 2 : argv[++i]);
        }
    }
    getSizes(&my_rank, &world_size, &nprocs);
    if (my_rank == 0)
    {
        char hostname[1000];
        gethostname(&hostname[0], 1000);
        cout << "Master hostname: " << hostname << endl;
    }

    switch (fractal_type)
    {
        case 0:
        default:
            computation = new NewtonComputation();
            x_center = 0.0;
            y_center = 0.0;
            zoom = 2.0 / width;
            break;
        case 1:
            computation = new FatouComputation();
            x_center = 3.001;
            y_center = 0.075975;
            zoom = 2.0 / width;
            break;
    }

    unsigned int * taskVals = new unsigned int[width * height];
    double window_vals[4];
    if (my_rank == 0)
    {
        SDL_Event event;
        bool going = true;
        bool window_success = createWindow(width, height, &screen, &pixels);
        bool redraw = true;

        if (!window_success)
            going = false;

        /* master loop */
        while (going && SDL_WaitEvent(&event) != 0)
        {
            if (redraw)
            {
                struct timeval before, after;
                window_vals[0] = 0.0;
                window_vals[1] = x_center;
                window_vals[2] = y_center;
                window_vals[3] = zoom;
                sendWindowVals(&window_vals[0], world_size);
                gettimeofday(&before, NULL);
                draw(my_rank, world_size, nprocs, width, height,
                    pixels, taskVals, computation);
                gettimeofday(&after, NULL);
                if (display_times)
                {
                    double time_before = before.tv_sec + before.tv_usec / 1000000.0;
                    double time_after = after.tv_sec + after.tv_usec / 1000000.0;
                    double diff = time_after - time_before;
                    cout << "Elapsed time: " << diff << " seconds." << endl;
                }
                redraw = false;
            }
            SDL_UpdateRect(screen, 0, 0, 0, 0);
            switch (event.type)
            {
                case SDL_QUIT:
                    going = false;
                    break;
                case SDL_KEYDOWN:
                    if (event.key.keysym.sym == SDLK_q)
                        going = false;
                    break;
                case SDL_MOUSEBUTTONDOWN:
                    switch (event.button.button)
                    {
                        case 1:         /* left-click to re-center and zoom in */
                            x_center = getXVirt(event.button.x);
                            y_center = getYVirt(event.button.y);
                            zoom /= 2.0;
                            redraw = true;
                            break;
                        case 2:         /* middle click just to re-center */
                            x_center = getXVirt(event.button.x);
                            y_center = getYVirt(event.button.y);
                            redraw = true;
                            break;
                        case 4:         /* zoom in */
                            zoom /= 2.0;
                            redraw = true;
                            break;
                        case 5:         /* zoom out */
                            zoom *= 2.0;
                            redraw = true;
                            break;
                    }
                    break;
            }
        }
        window_vals[0] = 1.0;
        sendWindowVals(&window_vals[0], world_size);
    }
    else
    {
        /* slave loop */
        for (;;)
        {
            // DEBUG:
//            cout << "MPI node " << my_rank << " waiting for command." << endl;
            /* wait for a redraw or quit command */
            MPI_Recv(&window_vals[0], 4, MPI_DOUBLE,
                     MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, NULL);
            if (window_vals[0] != 0.0)
                break;
            x_center = window_vals[1];
            y_center = window_vals[2];
            zoom = window_vals[3];
            // DEBUG:
//            cout << "MPI node " << my_rank << " received ("
//                << x_center << ", " << y_center << "), zoom " << zoom << endl;
            draw(my_rank, world_size, nprocs, width, height,
                NULL, taskVals, computation);
        }
    }
    delete[] taskVals;

    MPI_Finalize();

    delete computation;
    return 0;
}

/**************************************************************************
 * This utility function is used by the master MPI node to create a       *
 * window using SDL for displaying the fractal in and getting user input. *
 *************************************************************************/
bool createWindow(int width, int height,
                  SDL_Surface ** screen, Uint32 ** pixels)
{
    if (SDL_Init(SDL_INIT_VIDEO))
    {
        cerr << "Failed to initialize SDL!" << endl;
        return false;
    }

    atexit(SDL_Quit);

    if (!(*screen = SDL_SetVideoMode(width, height, 32, 0)))
    {
        cerr << "Failed to set video mode!" << endl;
        return false;
    }
    SDL_WM_SetCaption(PROGNAME, PROGNAME);

    *pixels = (Uint32 *) (*screen)->pixels;

    return true;
}

/**************************************************************************
 * This utility function returns the MPI node's rank, the total number    *
 * of MPI nodes, and the number of processing cores on the local node     *
 *************************************************************************/
void getSizes(int * rank, int * size, int * nprocs)
{
    MPI_Comm_rank(MPI_COMM_WORLD, rank);
    MPI_Comm_size(MPI_COMM_WORLD, size);

    *nprocs = sysconf(_SC_NPROCESSORS_CONF);

    int displs[*size];
    int counts[*size];
    for (int i = 0; i < *size; i++)
    {
        displs[i] = i;
        counts[i] = 1;
    }
    int all_nprocs[*size];
    MPI_Gatherv(nprocs, 1, MPI_INT,
                &all_nprocs[0], &counts[0], &displs[0], MPI_INT,
                0, MPI_COMM_WORLD);

    if (*rank == 0)
    {
        int total_nprocs = 0;
        cout << "Number of cores on each MPI node:" << endl;
        for (int i = 0; i < *size; i++)
        {
            cout << all_nprocs[i] << " ";
            total_nprocs += all_nprocs[i];
        }
        cout << endl;
        cout << "Total number of cores: " << total_nprocs << endl;
    }
}

/**************************************************************************
 * This function is executed by each MPI node every time a fractal        *
 * frame is to be drawn.                                                  *
 *************************************************************************/
void draw(int rank, int world_size, int nprocs, int width, int height,
    Uint32 * pixels, Uint32 * taskVals, Computation * computation)
{
    int num_pixels = width * height;
    int firstPixel;
    int numPixels;
    taskAllocate(num_pixels, world_size, rank, &firstPixel, &numPixels);

#pragma omp parallel for
    for (int i = 0; i < numPixels; i++)
    {
        int this_pixel_num = firstPixel + i;
        int x = this_pixel_num % width;
        int y = this_pixel_num / width;
        double x_virt = getXVirt(x);
        double y_virt = getYVirt(y);
        unsigned int color = computation->compute(x_virt, y_virt);
        if (rank == 0)
            pixels[this_pixel_num] = color;
        else
            taskVals[i] = color;
    }

    if (rank == 0)
    {
        for (int i = 1; i < world_size; i++)
        {
            taskAllocate(num_pixels, world_size, i, &firstPixel, &numPixels);
            MPI_Recv(pixels + firstPixel, numPixels, MPI_INT,
                i, MPI_ANY_TAG, MPI_COMM_WORLD, NULL);
        }
    }
    else
    {
        MPI_Send(taskVals, numPixels, MPI_INT, 0, 0, MPI_COMM_WORLD);
    }
}

/**************************************************************************
 * This utility function is used by the master process to update all      *
 * of the slave processes for the position and zoom-level of the view.    *
 *************************************************************************/
void sendWindowVals(double * winVals, int world_size)
{
    // DEBUG:
//    cout << "Master sending out new window values" << endl;
    for (int to_proc = 1; to_proc < world_size; to_proc++)
    {
        MPI_Send(winVals, 4, MPI_DOUBLE, to_proc,
            0, MPI_COMM_WORLD);
    }
}

/*
 * taskAllocate() will divide a set of total_tasks tasks into
 * total_workers groups, as evenly as possible
 * Parameters:
 *   total_tasks   : IN  : the total number of tasks to divide up
 *   total_workers : IN  : the total number of workers to allocate tasks to (>0)
 *   this_id       : IN  : the id (base 0) of the task calling us for work
 *   first_task_id : OUT : the id (base 0) of the first task for this worker
 *   num           : OUT : the number of tasks assigned to this worker
 */
inline void taskAllocate(int total_tasks, int total_workers, int this_id,
                         int * first_task_id, int * num)
{
    int l_num;
    int leftovers = total_tasks % total_workers; /* num of "leftover" tasks */
    if (this_id < leftovers)
    {
        l_num = total_tasks / total_workers + 1; /* do one of the leftovers */
        *first_task_id = l_num * this_id;
    }
    else
    {
        l_num = total_tasks / total_workers;
        *first_task_id = l_num * this_id + leftovers;
    }
    *num = l_num;
}