#define _POSIX_C_SOURCE 199309L // for CLOCK_MONOTONIC

#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <sys/types.h>
#include <unistd.h> // for getopt
#include <getopt.h> // for getopt_long
#include <string.h> // for strcpy
#include <time.h> // for benchmarking

// All implementations
#include "tricorn.h" // Final implmementation
#include "tricorn_V1.h"
#include "tricorn_V2.h"
#include "tricorn_V3.h"


const uint16_t TESTING_AMOUNT = 10;
const uint16_t NUMBER_OF_VERSIONS = 4;


void save_bmp(const char *filename, unsigned char *img, size_t width, ssize_t height,_Bool color);
void benchmarking(size_t benchmarking_amount, int version_to_use, float complex complexA, size_t width, ssize_t height, float resolution, unsigned n, _Bool colored);
unsigned tests(void);
void checkInputs(_Bool n,_Bool r,_Bool o,_Bool d,_Bool s);


// Example test command
// ./project -s -2.0,-1.5 -d 1200,900 -r 0.003 -n 256 -o tricorn.bmp -C
const char* help_messsage = "Help Message\n\n"
                "Usage:\n"
                "   ./tricorn [OPTIONS]\n\n"
                "Options:\n"
                "-V <number>\n"
                "   Selects the implementation to use.\n"
                "   -V 0 uses the main implementation.\n"
                "   If no argument is given, -V 0 is used.\n"
                "   Example: -V 1\n\n"
                "-B <number>\n"
                "   Benchmarks the selected implementation.\n"
                "   The argument chooses how many times the function is called.\n"
                "   Example: -B 5\n\n"
                "-s <real>,<imaginary>\n"
                "   Sets the starting point of the function.\n"
                "   Example: -s -2.0,1.5\n\n"
                "-d <width>,<height>\n"
                "   Sets the width and the height of the generated image.\n"
                "   Example: -d 800,600\n\n"
                "-n <number>\n"
                "   Sets the maximum number of iterations per pixel.\n"
                "   Example: -n 255\n\n"
                "-r <floating-point number>\n"
                "   Sets the resolution per pixel.\n"
                "   Example: -r 0.01\n\n"
                "-o <filename>\n"
                "   Sets the output file name.\n"
                "   The file name must end with .bmp\n"
                "   Example: -o image.bmp\n\n"
                "-C, --color\n"
                "   Enables colored output.\n\n"
                "-h, --help\n"
                "   Displays the help message and exits.\n\n"
                "-t, --test\n"
                "   Runs the tests, prints the results, and exists.\n";

int main(int argc, char** argv)
{
    int opt_long; 
    // will be the returned optional character, 
    // example:  ./main -c     ->  opt_long here will be the optional 'c'

    const char* optstring = "V:B:s:d:n:r:o:Cht" ;
    // valid options that can be given to the programme. options with ":" after them expect an argument, which we'll later store in <char *optarg>
    // example:  B: <number>   ->  optional B expects an argument after it, and is valid    -> ./main -B 67 (optarg is now "67")
    // example:  h -> optional h doesn't expect an argument -> ./main -h

    // optind -> keeps track of the next index. initialised as 1 in the beginning. calling opt_long updates it every time

    // optarg -> is a charptr. stores the arguments given to the optionals

    static struct option long_options[] = {
        {"color",0,0,'C'},
        {"help",0,0,'h'},
        {"test",0,0,'t'},
        {0,0,0,0}
    };
    // these are the long optionals, which we can call with double "--".  "--help" or "--test". they are converted to the optionals "h" and "t" respectively
    // first argument is "name" -> name of the long. "--<name>"
    // second is "has_arg" -> 0 if no arg, 1 if arg, 2 if optional arg
    // third is "*flag" -> sets a flag so that "--help" simply returns "h" and does nothing else
    // fourth is "val" -> what to return when "--help" is given, returns "h"


    int version_to_use = 0; // actual initial value
    _Bool colored = 0; // actual initial value
    _Bool will_benchmark = false; // actual initial value
    _Bool will_test = false;


    _Bool nGiven = false;
    _Bool rGiven = false;
    _Bool oGiven = false;
    _Bool dGiven = false;
    _Bool sGiven = false;
    size_t amount_function_called = 0; // initial values, so that the compiler doesn't complain
    unsigned n = 0;
    float starting_point_real = -1;
    float starting_point_imaginary = -1;
    size_t width = 0;
    ssize_t height = -1;
    float resolution = -1;
    char* output_data = NULL;


    opt_long = getopt_long(argc,argv,optstring, long_options, NULL); 
    // first four arguments are self-explanatory, fifth argument is the index of which long_option we return in the struct "long_options", but it's irrelevant so we set it to NULL
    while (opt_long != -1)
    {   
        switch (opt_long)
        {
        case 'V': {
            char *end;
            version_to_use = strtol(optarg,&end,10);


            if (version_to_use < 0 || version_to_use > 5 || end == optarg || *end != '\0')
            {
                fprintf(stderr,"Invalid version to use: V%s. Available versions are : V0,V1,V2,V3\n", optarg);
                return EXIT_FAILURE;
            }
            break;
        }
        
        case 'B': {
            char *end;
            long value = strtol(optarg,&end,10);

            if (value <= 0 || end == optarg || *end != '\0') {
                fprintf(stderr, "Invalid benchmarking input: %s\n", optarg);
                return EXIT_FAILURE;
            }

            amount_function_called = (size_t)value;
            will_benchmark = true;
            break;
        }

        case 's': {
            sGiven = true;
            char *imaginary_string;
            char *end;

            starting_point_real = strtof(optarg,&imaginary_string); // split into two
            if (imaginary_string == optarg || *imaginary_string != ',')
            {
                fprintf(stderr, "Invalid first number or missing comma: %s\n",optarg);
                return EXIT_FAILURE;
            }

            starting_point_imaginary = strtof(imaginary_string + 1, &end); // second part, the imaginary part
            if (end == imaginary_string + 1 || *end != '\0')
            {
                fprintf(stderr, "Invalid second number or extra characters: %s\n",optarg);
                return EXIT_FAILURE;
            }
            break;
        }
            
        case 'd': {
            dGiven = true;
            char* height_string;
            char *end;
            long widthValue = strtol(optarg,&height_string,10);

            if (height_string == optarg || *height_string != ',')
            {
                fprintf(stderr, "Invalid width or missing comma: %s\n",optarg);
                return EXIT_FAILURE;
            }

            long heightValue = strtol(height_string + 1,&end,10);

            if (end == height_string + 1 || *end != '\0')
            {
                fprintf(stderr, "Invalid height or extra characters: %s\n",optarg);
                return EXIT_FAILURE;
            }

            if (widthValue <= 0 || heightValue == 0)
            {
                fprintf(stderr, "Width must be > 0 and height must not be 0: %s\n", optarg);
                return EXIT_FAILURE;
            }

            width = (size_t)widthValue;
            height = (ssize_t)heightValue;
            break;
        }

        case 'n': {
            nGiven = true;
            char* end;
            long nValue = strtol(optarg,&end,10);

            if (end == optarg || *end != '\0' || nValue <= 0)
            {
                fprintf(stderr, "Invalid n value: %s\n", optarg);
                return EXIT_FAILURE;
            }
            
            n = (unsigned) nValue;
            break;
        }

        case 'r': {
            rGiven = true;
            char *end;
            resolution = strtof(optarg,&end);
            if (end == optarg || *end != '\0' || resolution <= 0.0f) 
            {
                fprintf(stderr, "Invalid floating point number: %s\n",optarg);
                return EXIT_FAILURE;
            }
            break;
        }

        case 'o':
            oGiven = true;
            if (optarg[0] == '-')
            {
                fprintf(stderr, "Invalid output filename: %s\n",optarg);
                return EXIT_FAILURE;
            }
            
            output_data = malloc(strlen(optarg) + 1);
            if (output_data == NULL)
            {
                fprintf(stderr,"Memory allocation failed.\n");
                return EXIT_FAILURE;
            }
            
            strcpy(output_data,optarg);
            break;

        case 'C':
            colored = true;
            break;

        case 'h':
            printf("%s",help_messsage);
            return EXIT_SUCCESS;

        case 't':
            will_test = true;
            break;

        default:
            // case '?' -> when the argument is not valid
            fprintf(stderr,"ERROR: Invalid argument -%c:\n",optopt);
            return EXIT_FAILURE;
        }

        opt_long = getopt_long(argc,argv,optstring, long_options, NULL); 
    }


    // Run 
    float complex start = (starting_point_real + starting_point_imaginary*I);
    int bytesPerColour = colored ? 3 : 1;
    ssize_t absolute_height = height >= 0 ? height : height * - 1;
    unsigned char *imageArr = malloc(width*absolute_height*bytesPerColour);// allocate image array on the heap to support large images as the stack is usually only 8 MB big

    if(imageArr == NULL)
        exit(EXIT_FAILURE);
    
    if (will_test)
        return tests() ? EXIT_FAILURE : EXIT_SUCCESS;
    
    if (!(nGiven && rGiven && oGiven && dGiven && sGiven))
    {
        checkInputs(nGiven,rGiven,oGiven,dGiven,sGiven);
        free(imageArr);
        free(output_data);
        return EXIT_FAILURE;
    }

    if (will_benchmark)
    {   
        printf("\nDoing benchmarking. No images will be created\n");
        benchmarking(amount_function_called,version_to_use,start,width,height,resolution,n,colored);
        free(imageArr);
        free(output_data);
        return EXIT_SUCCESS;
    }
    
    switch (version_to_use)
    {
        case 0:
            // Uses highly optimized SIMD - main implementation
            tricorn(start,width,height,resolution,n,colored,imageArr);
            break;
        case 1: 
            // Uses scalar comparison implementation
            tricorn_V1(start,width,height,resolution,n,colored,imageArr);
            break;
        case 2:
            // Uses unoptimized SIMD implementation
            tricorn_V2(start,width,height,resolution,n,colored,imageArr);
            break;
        case 3:
            // Uses highly optimized SIMD + Threading implementation
            tricorn_V3(start,width,height,resolution,n,colored,imageArr);
            break;
        default:
            break;
    }

    save_bmp(output_data,imageArr,width,height,colored);
    
    free(output_data);
    free(imageArr);
    return 0;
}

void benchmarking(size_t benchmarking_amount, int version_to_use, float complex complexA, size_t width, ssize_t height, float resolution, unsigned n, _Bool colored) {
    struct timespec start;
    struct timespec end;
    void (*tricorn_implementation)(float complex, size_t, ssize_t, float, unsigned, _Bool, unsigned char*);

    unsigned int bytePerColor = colored ? 3 : 1;
    unsigned char *tempArr = malloc(width*height*bytePerColor);
    if(tempArr == NULL) {
        fprintf(stderr,"Could not allocate memory for buffer");
        exit(EXIT_FAILURE);
    }

    switch (version_to_use)
    {
        case 0: // Uses highly optimized SIMD and threaded version / main implementation
            printf("Benchmarking V0 - highly optimized SIMD\n");
            tricorn_implementation = tricorn;
            break;
        case 1: // Uses unoptimised scalar comparison implementation
            printf("Benchmarking V1 - scalar comparison implementation\n");
            tricorn_implementation = tricorn_V1;
            break;
        case 2: // Uses optimized SISD implementation
            printf("Benchmarking V2 - unoptimized SIMD implementation\n");
            tricorn_implementation = tricorn_V2;
            break;
        case 3: // Uses unoptimized SIMD implementation
            printf("Benchmarking V3 - highly optimized SIMD + Threading implementation\n");
            tricorn_implementation = tricorn_V3;
            break;
        default:
            fprintf(stderr,"Invalid version to use: V%d. Available versions are : V0,V1,V2,V3,V4,V5\n",version_to_use);
            return;
    }

    printf("\n");

    clock_gettime(CLOCK_MONOTONIC, &start);
    for (size_t i = 0; i < benchmarking_amount; i++)
    {
        tricorn_implementation(complexA,width,height,resolution,n,colored,tempArr);
    }
    clock_gettime(CLOCK_MONOTONIC, &end);

    double time = end.tv_sec- start.tv_sec + 1e-9 * (end.tv_nsec- start.tv_nsec);
    printf("The total amount of time spent for %zu function calls: %lf seconds\n", benchmarking_amount,time);
    double average = time / benchmarking_amount;
    printf("Average time per function: %lf seconds\n", average);

    free(tempArr);
}

void run_benchmarks(float complex start,size_t width, size_t height,float res,int n, _Bool colored);
unsigned tests(void) {

    printf("Doing testing\n\n");


    printf("\n\n\nTest 1: Default image, coloured\n");
    float complex start = -2.0f - 0.5f*I;
    size_t width = 1920;
    size_t height = 1080;
    float res = 0.001f;
    int n = 200;
    _Bool colored = true;
    run_benchmarks(start,width,height,res,n,colored);


    printf("\n\n\nTest 2: Ultra high resolution low iterations\n");
    start = -2.0f - 0.5f*I;
    width = 7680;
    height = 4320;
    res = 0.001f;
    n = 5;
    run_benchmarks(start,width,height,res,n,colored);


    printf("\n\n\nTest 3: Zoomed on fractal edge\n");
    start = -1.431f - 0.250f*I;
    width = 1024;
    height = 1024;
    res = 0.0005f;
    n = 3000;
    run_benchmarks(start,width,height,res,n,colored);


    printf("\nTest 4: Pure Black\n");
    start = 0;
    width = 1920;
    height = 1080;
    res = 0.00001f;
    n = 1000;
    run_benchmarks(start,width,height,res,n,colored);


    return 0;
}
void run_benchmarks(float complex start,size_t width, size_t height,float res,int n, _Bool colored){
    printf("Start: %f%+fi\n", crealf(start), cimagf(start));
    printf("Width: %li\n",width);
    printf("Height: %li\n", height);
    printf("Resolution: %f\n", res);
    printf("Iterations: %i\n", n);
    printf("Colored: %s\n", colored ? "true" : "false");
    for(int i = 0; i < NUMBER_OF_VERSIONS;i++) {
        printf("\n\n");
        benchmarking(TESTING_AMOUNT,i,start, width, height, res, n, colored);
    }
}

void checkInputs(_Bool n,_Bool r,_Bool o,_Bool d,_Bool s) {
    if (!n)
    {
        fprintf(stderr, "ERROR: Input n not given.\n");
    }
    if (!r)
    {
        fprintf(stderr, "ERROR: Input r not given.\n");
    }
    if (!o)
    {
        fprintf(stderr, "ERROR: Input o not given.\n");
    }
    if (!d)
    {
        fprintf(stderr, "ERROR: Input d not given.\n");
    }
    if (!s)
    {
        fprintf(stderr, "ERROR: Input s not given.\n");
    }
    
}


void save_bmp(const char *filename, unsigned char *img, size_t width, ssize_t height,_Bool color) {
    // Open a file at the filepath given as a stream in the program
    FILE *fptr = fopen(filename,"wb");

    if(fptr == NULL) { // Chech for errors with opening the file stream
        fprintf(stderr,"ERROR: Could not open output file at %s\n.",filename);
        exit(EXIT_FAILURE);
    }

    size_t bytesPerColor = color ? 3 : 1;


    int actual_height = (height < 0) ? -height : height; // Calculate absolute height for iteration and calculation purposes
    size_t row_padding = (4 - (width * bytesPerColor) % 4) % 4; // Calculate padding for each row as .bmp is 4 bytes aligned
    size_t row_width = width * bytesPerColor + row_padding; // Calculate row width with padding

    size_t image_size = actual_height * row_width;
    size_t header_offset = color ? 54 : 1078; // If colored then offset = 54 if not we add offset = 54 + 1024; 1024 Bytes for color pallete
    size_t file_size = image_size + header_offset; // File size = Image Size + Headers



    // File header - contains information about the file
    unsigned char file_header[14] = { 
        'B', 'M',           // BM - identifies file as a .bmp file
        0, 0, 0, 0,         // 4 bytes for file size - will be set later
        0, 0, 0, 0,         // Reserved
        0, 0, 0, 0         // 4 bytes that specify the offset where the image array begins
    };

    // Serialize the size into 4 bytes - (unsigned char) takes only bottom 4 bytes 1 2 3 4
    file_header[2] = (unsigned char)(file_size); 
    file_header[3] = (unsigned char)(file_size >> 8);
    file_header[4] = (unsigned char)(file_size >> 16);
    file_header[5] = (unsigned char)(file_size >> 24);

    file_header[10] = (unsigned char)(header_offset);
    file_header[11] = (unsigned char)(header_offset >> 8);

    // Info header - contains information about the image itself
    unsigned char info_header[40] = { 
        40, 0, 0, 0,        // Size of this header    
        0, 0, 0, 0,         // Width in pixels - set later
        0, 0, 0, 0,         // Height in pixels - set later
        1, 0,               // Number of color planes - always 1
        bytesPerColor * 8, 0,              // Number of bits per pixel or color depth
        0, 0, 0, 0,         // Compression method - 0 for no compression
        0, 0, 0, 0,         // Image size - set later
        0x13, 0x0B, 0, 0,   // Horizontal resolution in pixels per metre - Set to standart 72 pixels per inch
        0x13, 0x0B, 0, 0,   // Vertical resolution in pixels per metre - Set to standart 72 pixels per inch
        0, 0, 0, 0,         // Number of colors in the color pallete - 0 for default of 2^n where n is bits per pixel
        0, 0, 0, 0          // Number of important colors - 0 for all colors important
    };

    // Setting width
    info_header[4] = (unsigned char)(width);
    info_header[5] = (unsigned char)(width >> 8);
    info_header[6] = (unsigned char)(width >> 16);
    info_header[7] = (unsigned char)(width >> 24);

    // Setting height
    uint32_t h = (uint32_t)height; 
    info_header[8] = (unsigned char)(h);
    info_header[9] = (unsigned char)(h >> 8);
    info_header[10] = (unsigned char)(h >> 16);
    info_header[11] = (unsigned char)(h >> 24);

    // Setting image size
    info_header[20] = (unsigned char)(image_size);
    info_header[21] = (unsigned char)(image_size >> 8);
    info_header[22] = (unsigned char)(image_size >> 16);
    info_header[23] = (unsigned char)(image_size >> 24);

    // Write headers into file
    //size_t fwrite(const void ptr[restrict .size * .n], size_t size, size_t n, FILE *restrict stream);
    fwrite(file_header,1,14,fptr); 
    fwrite(info_header,1,40,fptr);


    //If image is grayscale write color dictionary to tell reader what color to visualise for each value
    // Row 0 - Value 0: [R=0;B=0;G=0]
    // ...
    //Row i - Value i: [R=i;B=i;G=i]
    if (!color) {
        unsigned char palette[1024];
        for (int i = 0; i < 256; i++) {
            palette[i * 4 + 0] = i; // B
            palette[i * 4 + 1] = i; // G
            palette[i * 4 + 2] = i; // R
            palette[i * 4 + 3] = 0; // Reserved (Padding)
        }
        fwrite(palette, 1, 1024, fptr);
    }

    // Save padding bytes as array so we can write from it easily to the file later
    unsigned char padding_bytes[3] = {0, 0, 0}; 
    
    for(int i = 0; i < actual_height; i++) {
        // Write whole row of pixels - img + (i * width * 3) is the position of the row in the image array
        // size = 3 as we are interpreting every pixel as an object
        fwrite(img + (i * width * bytesPerColor), bytesPerColor, width, fptr); 
        if(row_padding > 0) {
            fwrite(padding_bytes,1,row_padding,fptr); // Write row_padding number of padding bytes to the file stream
        }
    }
    
    fclose(fptr);
}