/*
 * Combined WS281x LED Driver for PRU0
 * Handles BOTH mood strip AND jaw LEDs from PRU0
 * 
 * Why combined? All PRU1 native pins (P8_27-46) are locked by HDMI on stock image.
 * Solution: Use two free PRU0 pins instead.
 * 
 * Pins (both free and available):
 *   P9_27 = pr1_pru0_pru_r30_5 (Mode 5) - Mood strip (56 LEDs)
 *   P9_25 = pr1_pru0_pru_r30_7 (Mode 5) - Jaw LEDs (24 LEDs)
 * 
 * Shared Memory Layout (0x00010000):
 *   [0]:       mood_num_leds (max 56)
 *   [1]:       mood_trigger (write 1 to update)
 *   [2]:       jaw_num_leds (max 24)
 *   [3]:       jaw_trigger (write 1 to update)
 *   [4-171]:   mood LED data (56 * 3 bytes, GRB format)
 *   [172-243]: jaw LED data (24 * 3 bytes, GRB format)
 */

#include <stdint.h>
#include <pru_cfg.h>

#define MOOD_PIN        5       /* P9_27 */
#define JAW_PIN         7       /* P9_25 */
#define MAX_MOOD_LEDS   56
#define MAX_JAW_LEDS    24

/* WS2812 timing at 200MHz (5ns per cycle) */
#define T0H     70
#define T0L     160
#define T1H     140
#define T1L     120

#define SHARED_MEM      0x00010000
#define MOOD_DATA_OFF   4
#define JAW_DATA_OFF    172

volatile uint8_t *shared = (volatile uint8_t *)SHARED_MEM;
volatile register uint32_t __R30;

static inline void delay_cycles(uint32_t cycles) {
    while (cycles--) {
        __asm(" NOP");
    }
}

static void send_byte_on_pin(uint8_t byte, uint8_t pin) {
    for (int i = 7; i >= 0; i--) {
        if ((byte >> i) & 1) {
            __R30 |= (1 << pin);
            delay_cycles(T1H);
            __R30 &= ~(1 << pin);
            delay_cycles(T1L);
        } else {
            __R30 |= (1 << pin);
            delay_cycles(T0H);
            __R30 &= ~(1 << pin);
            delay_cycles(T0L);
        }
    }
}

static void send_strip(uint8_t pin, uint8_t *data, uint8_t num_leds) {
    for (int i = 0; i < num_leds; i++) {
        send_byte_on_pin(data[i * 3 + 0], pin);  /* G */
        send_byte_on_pin(data[i * 3 + 1], pin);  /* R */
        send_byte_on_pin(data[i * 3 + 2], pin);  /* B */
    }
    /* Reset pulse */
    __R30 &= ~(1 << pin);
    delay_cycles(10000);
}

void main(void) {
    /* Enable OCP master port */
    CT_CFG.SYSCFG_bit.STANDBY_INIT = 0;
    
    /* Clear output pins */
    __R30 &= ~((1 << MOOD_PIN) | (1 << JAW_PIN));
    
    while (1) {
        /* Check mood strip trigger */
        if (shared[1] == 1) {
            uint8_t num = shared[0];
            if (num > MAX_MOOD_LEDS) num = MAX_MOOD_LEDS;
            send_strip(MOOD_PIN, (uint8_t *)&shared[MOOD_DATA_OFF], num);
            shared[1] = 0;
        }
        
        /* Check jaw trigger */
        if (shared[3] == 1) {
            uint8_t num = shared[2];
            if (num > MAX_JAW_LEDS) num = MAX_JAW_LEDS;
            send_strip(JAW_PIN, (uint8_t *)&shared[JAW_DATA_OFF], num);
            shared[3] = 0;
        }
    }
}