/*
 * WS281x LED Strip Driver for PRU0
 * Vixy's Mood Strip - 56 LEDs
 * 
 * Pin: P8_11 = pr1_pru0_pru_r30_15 (Mode 6)
 * Alternative: P9_27 = pr1_pru0_pru_r30_5 (Mode 5) - FREE PIN!
 * 
 * Shared Memory Layout (0x00010000):
 *   [0]:     num_leds (max 56)
 *   [1]:     trigger (write 1 to update)
 *   [4-171]: LED data (56 * 3 bytes, GRB format)
 */

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

/* Use P9_27 (R30 bit 5) - it's free! */
#define LED_PIN     5
#define NUM_LEDS    56

/* WS2812 timing at 200MHz (5ns per cycle) */
#define T0H     70      /* 350ns high for 0 bit */
#define T0L     160     /* 800ns low for 0 bit */
#define T1H     140     /* 700ns high for 1 bit */
#define T1L     120     /* 600ns low for 1 bit */

/* Shared memory base */
#define SHARED_MEM  0x00010000

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_bit(uint8_t bit) {
    if (bit) {
        __R30 |= (1 << LED_PIN);
        delay_cycles(T1H);
        __R30 &= ~(1 << LED_PIN);
        delay_cycles(T1L);
    } else {
        __R30 |= (1 << LED_PIN);
        delay_cycles(T0H);
        __R30 &= ~(1 << LED_PIN);
        delay_cycles(T0L);
    }
}

static void send_byte(uint8_t byte) {
    for (int i = 7; i >= 0; i--) {
        send_bit((byte >> i) & 1);
    }
}

static void send_led(uint8_t g, uint8_t r, uint8_t b) {
    send_byte(g);
    send_byte(r);
    send_byte(b);
}

void main(void) {
    /* Enable OCP master port */
    CT_CFG.SYSCFG_bit.STANDBY_INIT = 0;
    
    /* Clear output pin */
    __R30 &= ~(1 << LED_PIN);
    
    while (1) {
        /* Check trigger flag */
        if (shared[1] == 1) {
            uint8_t num = shared[0];
            if (num > NUM_LEDS) num = NUM_LEDS;
            
            /* Send LED data (GRB format) */
            for (int i = 0; i < num; i++) {
                uint8_t g = shared[4 + i * 3 + 0];
                uint8_t r = shared[4 + i * 3 + 1];
                uint8_t b = shared[4 + i * 3 + 2];
                send_led(g, r, b);
            }
            
            /* Reset pulse (>50us) */
            __R30 &= ~(1 << LED_PIN);
            delay_cycles(10000);
            
            /* Clear trigger */
            shared[1] = 0;
        }
    }
}