// a multi pin uart that send 8 bits at a time by treating the buffer as
// transposed (i.e. byte0 provides LSB0 of all pins, byte1 -> LSB1 etc.)
// buffer is 32 bits, so for 8 x 8-bit pushes, we need two full buffers
// to fill a single UART byte (start-b7-b6-b5-b4-b3-b2-b1-b0-stop)

// only tested at https://wokwi.com/tools/pioasm so far, not on a real
// device see also https://wokwi.com/projects/344410676217774675
    
.program uart_tx
    pull
    mov pins, null [1] ; start bit (0)
    out pins, 8 [1]
    out pins, 8 [1]
    out pins, 8 [1]
    out pins, 8
    pull
    out pins, 8 [1]
    out pins, 8 [1]
    out pins, 8 [1]
    out pins, 8 [1]
    mov pins, !null   ; stop bit (1)

% c-sdk {
static inline void uart_tx_program_init(PIO pio, uint sm, uint offset, uint pin, uint baud) {
    const int PIN_COUNT = 8;
    const int PIN_MASK = 0xFF;
    // the number of PIO cycles that it takes to output a single bit
    const uint cycles_per_bit = 2;
    // the clock divider (number of CPU cycles that each PIO cycle takes)
    float div = (float)clock_get_hz(clk_sys) / (baud * cycles_per_bit);

    // initialize the pins to high indicating the stop bit
    pio_sm_set_set_pins(pio, sm, pin, PIN_COUNT);
    pio_sm_set_pindirs_with_mask(pio, sm, 0xFFFFFFFF, PIN_MASK << pin);
    pio_sm_set_pins_with_mask(pio, sm, 0xFFFFFFFF, PIN_MASK << pin);
    for (int i = 0; i < PIN_COUNT; i++) {
        pio_gpio_init(pio, pin + i);
    }
    
    pio_sm_config c = uart_tx_program_get_default_config(offset);
    sm_config_set_out_shift(&c, true, false, 32); // shift to right, no autopull
    sm_config_set_out_pins(&c, pin, PIN_COUNT); // Set 8 pins starting at `pin`
    sm_config_set_set_pins(&c, pin, PIN_COUNT); // Set 8 pins starting at `pin`
    sm_config_set_clkdiv(&c, div);
    pio_sm_init(pio, sm, offset, &c); // Load config
    pio_sm_set_enabled(pio, sm, true); // Start state machine
}

static inline void uart_tx_program_putc(PIO pio, uint sm, uint32_t b1, uint32_t b2) {
    pio_sm_put_blocking(pio, sm, b1);
    pio_sm_put_blocking(pio, sm, b2);
}
%}