/*  Plays the birthday music on a buzzer attached to ATmega168 pin 24      */
/*  Written by Thanassis Tsiodras, June 2009                               */

/*  I've dropped the CKDIV8 in lfuse, so my ATmega168 runs at 8MHz         */
#define F_CPU 8000000

#include <avr/io.h>          /*  this contains all the IO port definitions */
#include <util/delay.h>      /*  for _delay_us and _delay_ms               */

#define UPPER_LIMIT_IN_DELAY_US (768000000/F_CPU)

/*  The song, in pairs of frequency and duration (in milliseconds)         */
struct _Note {
    int frequency;
    int durationMS;
} g_Notes[] = {
    {587,500}, {659,500}, {587,500}, {784,500}, {740,700}, {0,200},
    {587,500}, {659,500}, {587,500}, {880,500}, {784,700}, {0,200},
    {587,500}, {1175,500}, {988,500}, {784,500}, {740,500}, {659,500},
    {1046,500}, {988,500}, {784,500}, {880,500}, {784,700},
};

/*  The calculation "cache" that helps us avoid floating point in the loop */
/*  (read further below)                                                   */
unsigned g_remainderTicks[ sizeof(g_Notes)/sizeof(struct _Note) ];

/*  Create a specific frequency for a specified amount of ms.              */
void buzz(int idx)
{
    /*  From the desired frequency, we want the half-period, since we will */
    /*  set the buzzer on, sleep for half a period, set it off, sleep for  */
    /*  half a period, and then repeat...                                  */

    /*  <--Period-->                                                       */
    /*   ____       _____       _                                          */
    /*  |    |     |     |     |                                           */
    /*  |    |_____|     |_____|                                           */

    /*  <-H->                                                              */

    /*  H=half period */

    /*  We want to use _delay_us to sleep for half a period, but...        */
    /*  _delay_us has a limitation: the largest value it can sleep for is  */

    /*     Upper limit in microSeconds = 768 / FrequencyInMHz              */

    /*  If using a 1MHz clock, this value is 768us... The lowest note in   */
    /*  the "Happy birthday" tune is however 587Hz, so it requires a half  */
    /*  period of 1000000/(2*587) = 851us, larger than the limit...        */

    /*  Not to mention that I have already removed the CKDIV8 fuse, so my  */
    /*  ATmega168 runs at 8MHz: the limit there is 96us !                  */

    /*  We therefore sleep the half periods in steps of "upper limit",     */
    /*  until we have a remainder less than UPPER_LIMIT_IN_DELAY_US.       */

    int freq = g_Notes[idx].frequency, ms = g_Notes[idx].durationMS;

    /*  We first calculate the total micro-seconds for this note.          */
    int32_t total_us = (((int32_t) ms)*1000LL);

    register unsigned desiredHalfPeriod_us = 500000/freq;

    /* How many loops will we run, delaying one Period each time?          */
    int loops = total_us/(2*desiredHalfPeriod_us);

    /*  Originally, at this point the following code existed: */

    /*  while(loops--) {                                                   */
    /*    PORTC ^= 0x02;                 // Buzzer on                      */
    /*    _delay_us(desiredHalfPeriod);  // sleep for...                   */
    /*    PORTC ^= 0x02                  // Buzzer off;                    */
    /*    _delay_us(desiredHalfPeriod);  // sleep for...                   */
    /*  } */

    /*  It turned out however, that this code did not work...              */
    /*  The implementation of _delay_us does floating point arithmetic     */
    /*  to calculate the number of ticks to wait, and since _delay_us is   */
    /*  called all the time, we end up doing lots of floating point work   */
    /*  which takes too much time and  messes up our timing...             */

    /*  Two things, after reading the code in util/delay.h :               */

    /*  First, we know that the UPPER_LIMIT_IN_DELAY_US is actually the    */
    /*  same as a _delay_loop_1(255), so we will be "eating up" the        */
    /*  desiredHalfPeriod calling this directly - and knowing that we      */
    /*  "eat" UPPER_LIMIT_IN_DELAY_US each time we do it.                  */

    /*  Second, this will eventually get us to a stage where the remaining */
    /*  microseconds will be less than UPPER_LIMIT_IN_DELAY_US; We can't   */
    /*  use _delay_us for them (floating point is slow...), so we...       */
    /*  pre-calculate the "remainder" for each note, at the beginning of   */
    /*  the main() function below, and store it inside g_remainderTicks... */

    /*  We therefore prepare a "calculation" cache, that we use per loop...*/

    /*  And here, in all its glory, the main oscillator loop, without any  */
    /*  slow floating point anywhere...                                    */

    while(loops--) {
        register unsigned counter = desiredHalfPeriod_us;
        
        PORTC ^= 0x02;                  /*  Buzzer on                      */
                                        /*  Now sleep for half a period    */
        while(counter) {
            if (counter >= UPPER_LIMIT_IN_DELAY_US) {
                _delay_loop_1(255);
                counter -= UPPER_LIMIT_IN_DELAY_US;
            } else {
                _delay_loop_1(g_remainderTicks[idx]);
                break;
            }
        }

        PORTC ^= 0x02;                  /*  Buzzer off                     */
                                        /*  Now sleep for half a period    */
        counter = desiredHalfPeriod_us;
        while(counter) {
            if (counter >= UPPER_LIMIT_IN_DELAY_US) {
                _delay_loop_1(255);
                counter -= UPPER_LIMIT_IN_DELAY_US;
            } else {
                _delay_loop_1(g_remainderTicks[idx]);
                break;
            }
        }
    }
}

int main()
{
    int idx = 0;

    DDRC  = 0x0F;  /*  PC0..PC3 as output                                  */
    PORTC = 0x00;  /*  all PORTC output pins Off                           */

    /* First, update the "remainder" table for each note (a "cache",       */
    /* read the comments inside the function "buzz" above)                 */

    while(idx<sizeof(g_Notes)/sizeof(struct _Note)) {
        unsigned desiredHalfPeriod_us = 500000/g_Notes[idx].frequency;
        uint8_t __ticks;
        double __tmp = ((F_CPU) / 3e6) * 
            (desiredHalfPeriod_us % UPPER_LIMIT_IN_DELAY_US);
        if (__tmp < 1.0)
            __ticks = 1;
        else if (__tmp > 255) {
            __ticks = 255;
        } else
            __ticks = (uint8_t)__tmp;
        g_remainderTicks[idx++] = __ticks;
    }

    /* Then, play the notes (and the silences), over and over again...     */

    while(1) {
        idx = 0;
        while(idx<sizeof(g_Notes)/sizeof(struct _Note)) {
            PORTC ^= 0x01;
            if (g_Notes[idx].frequency == 0)
                _delay_ms(g_Notes[idx].durationMS);
            else
                buzz(idx);
            idx ++;
        }
    }
    return 0;
}