#include "el512_panel.h"

#include <driver/gpio.h>
#include <esp_attr.h>
#include <esp_cpu.h>
#include <esp_heap_caps.h>
#include <esp_intr_alloc.h>
#include <esp_rom_gpio.h>
#include <freertos/FreeRTOS.h>
#include <freertos/task.h>
#include <soc/gpio_sig_map.h>
#include <soc/i2s_reg.h>
#include <soc/i2s_struct.h>
#include <soc/periph_defs.h>
#include <string.h>

#if __has_include(<esp_private/periph_ctrl.h>)
#include <esp_private/periph_ctrl.h>
#else
#include <driver/periph_ctrl.h>
#endif
#if __has_include(<esp32/rom/lldesc.h>)
#include <esp32/rom/lldesc.h>
#else
#include <rom/lldesc.h>
#endif

namespace {

// I2S0 clock source when clka_en = 0: PLL_D2_CLK, 160 MHz whatever the CPU clock.
constexpr uint32_t I2S_SOURCE_CLOCK_HZ = 160000000;
constexpr uint32_t I2S_BYTE_RATE_HZ = EL512Panel::PIXEL_CLOCK_HZ * EL512Panel::BYTES_PER_CLOCK;
constexpr uint32_t I2S_CLKM_DIV = I2S_SOURCE_CLOCK_HZ / I2S_BYTE_RATE_HZ;
static_assert(I2S_SOURCE_CLOCK_HZ % I2S_BYTE_RATE_HZ == 0,
              "byte rate must divide 160 MHz exactly");
static_assert(I2S_CLKM_DIV >= 2 && I2S_CLKM_DIV <= 255, "clkm_div_num out of range");

// In 8-bit LCD mode I2S0 puts its byte on DATA_OUT16..23 (bitluni ESP32Lib).
constexpr uint32_t I2S0_DATA_SIGNAL_BASE = I2S0O_DATA_OUT16_IDX;

// BCK runs at half the byte rate (measured: 2.5 MHz BCK at 5 MHz bytes), so each
// VCLK period spans two identical bytes. Uninverted, the rising edge landed on the
// byte-pair boundary and the image jittered; inverted, it sits mid-pair (verified
// on the panel).
constexpr bool INVERT_VCLK = true;

static_assert(EL512Panel::LINES_PER_CHUNK * EL512Panel::ROW_BYTES <= 4092, "descriptor too long");
static_assert(EL512Panel::ROW_BYTES % 4 == 0, "rows must be whole DMA words");
static_assert(EL512Panel::RING_LINES < EL512Panel::LINES_PER_FRAME, "ring longer than a frame");

constexpr uint16_t FB_ROW_BYTES = EL512Panel::WIDTH / 8;
constexpr uint16_t DATA_BYTES = EL512Panel::DATA_CLOCKS_PER_ROW * EL512Panel::BYTES_PER_CLOCK;

// One framebuffer byte (8 pixels = 4 clocks) -> 8 stream bytes as two DMA words,
// already in I2S byte order. Built once in begin(); read from the ISR.
DRAM_ATTR uint32_t s_lut[256][2];

void buildLut() {
  for (int b = 0; b < 256; b++) {
    uint8_t bytes[8];
    for (int c = 0; c < 4; c++) {
      // Clock c carries x = 2c and 2c+1. The datasheet numbers columns from 1, so
      // x = 2c is column 2c+1 (odd -> TVID) and x = 2c+1 is column 2c+2 (even -> VID).
      const uint8_t bits = (uint8_t)(b << (c * 2));
      uint8_t v = EL512Panel::BIT_HS;
      if (bits & 0x80) v |= EL512Panel::BIT_TVID;
      if (bits & 0x40) v |= EL512Panel::BIT_VID;
      bytes[(2 * c) ^ 2] = v;
      bytes[(2 * c + 1) ^ 2] = v;
    }
    memcpy(s_lut[b], bytes, sizeof(bytes));
  }
}

inline uint32_t IRAM_ATTR repeat4(uint8_t v) { return v * 0x01010101u; }

void resetI2S0() {
  const uint32_t lcReset = I2S_IN_RST_M | I2S_OUT_RST_M | I2S_AHBM_RST_M | I2S_AHBM_FIFO_RST_M;
  I2S0.lc_conf.val |= lcReset;
  I2S0.lc_conf.val &= ~lcReset;
  const uint32_t confReset =
      I2S_RX_RESET_M | I2S_RX_FIFO_RESET_M | I2S_TX_RESET_M | I2S_TX_FIFO_RESET_M;
  I2S0.conf.val |= confReset;
  I2S0.conf.val &= ~confReset;
}

void routeOutput(int8_t pin, uint32_t signal, bool invert) {
  if (pin < 0) return;
  esp_rom_gpio_pad_select_gpio(pin);
  gpio_set_direction((gpio_num_t)pin, GPIO_MODE_OUTPUT);
  esp_rom_gpio_connect_out_signal(pin, signal, invert, false);
}

void setupGpioOutput(int8_t pin, bool level) {
  if (pin < 0) return;
  esp_rom_gpio_pad_select_gpio(pin);
  gpio_set_direction((gpio_num_t)pin, GPIO_MODE_OUTPUT);
  gpio_set_level((gpio_num_t)pin, level);
}

}  // namespace

EL512Panel::EL512Panel(const EL512Pins& pins) : _pins(pins) {}

EL512Panel::~EL512Panel() { end(); }

bool EL512Panel::begin() {
  if (_running) return true;

  setupGpioOutput(_pins.enable, false);
  setupGpioOutput(_pins.lowpwr_n, true);
  _enabled = false;
  _lowPower = false;

  buildLut();
  // Both framebuffers are read from the ISR, so they must be internal RAM.
  _fb = (uint8_t*)heap_caps_calloc(1, FRAMEBUFFER_BYTES, MALLOC_CAP_INTERNAL | MALLOC_CAP_8BIT);
  _front = (uint8_t*)heap_caps_calloc(1, FRAMEBUFFER_BYTES, MALLOC_CAP_INTERNAL | MALLOC_CAP_8BIT);
  _ring = (uint8_t*)heap_caps_malloc(RING_BYTES, MALLOC_CAP_DMA | MALLOC_CAP_8BIT);
  if (!_fb || !_front || !_ring || !startI2S()) {
    end();
    return false;
  }
  _running = true;
  setEnabled(true);
  return true;
}

void EL512Panel::end() {
  if (_pins.enable >= 0) gpio_set_level((gpio_num_t)_pins.enable, 0);
  _enabled = false;
  stopI2S();
  _running = false;
  heap_caps_free(_fb);
  _fb = nullptr;
  heap_caps_free(_front);
  _front = nullptr;
  heap_caps_free(_ring);
  _ring = nullptr;
}

void EL512Panel::setEnabled(bool on) {
  _enabled = on;
  if (_pins.enable >= 0) gpio_set_level((gpio_num_t)_pins.enable, on);
}

void EL512Panel::setLowPower(bool on) {
  _lowPower = on;
  if (_pins.lowpwr_n >= 0) gpio_set_level((gpio_num_t)_pins.lowpwr_n, !on);
}

void EL512Panel::setPixel(int16_t x, int16_t y, bool on) {
  if (!_fb || x < 0 || y < 0 || x >= WIDTH || y >= HEIGHT) return;
  uint8_t& b = _fb[(y * WIDTH + x) >> 3];
  const uint8_t mask = 0x80 >> (x & 7);
  if (on)
    b |= mask;
  else
    b &= ~mask;
}

bool EL512Panel::getPixel(int16_t x, int16_t y) const {
  if (!_fb || x < 0 || y < 0 || x >= WIDTH || y >= HEIGHT) return false;
  return _fb[(y * WIDTH + x) >> 3] & (0x80 >> (x & 7));
}

void EL512Panel::clear(bool on) {
  if (_fb) memset(_fb, on ? 0xFF : 0x00, FRAMEBUFFER_BYTES);
}

void EL512Panel::show() {
  if (!_fb || !_front) return;
  if (!_running) {
    memcpy(_front, _fb, FRAMEBUFFER_BYTES);
    return;
  }
  // The ISR copies _fb to _front just before it encodes line 0 of a frame.
  _swapPending = true;
  for (int i = 0; i < 40 && _swapPending; i++) vTaskDelay(1);
  if (_swapPending) _showTimeouts++;
}

EL512Panel::Stats EL512Panel::takeStats() {
  Stats s{_refills, _maxGap, _swaps, _showTimeouts, _underruns};
  _underruns = 0;
  _refills = 0;
  _maxGap = 0;
  _swaps = 0;
  _showTimeouts = 0;
  return s;
}

void IRAM_ATTR EL512Panel::encodeRow(uint16_t line, uint8_t* out) const {
  // VS is held high during the last blanking line so that the first HS high
  // time ending after its rising edge is row 0 (datasheet: Video Data Input).
  const uint32_t vs = repeat4(line == LINES_PER_FRAME - 1 ? BIT_VS : 0);
  uint32_t* o = (uint32_t*)out;

  if (line >= HEIGHT) {
    for (uint16_t i = 0; i < ROW_BYTES / 4; i++) o[i] = vs;
    return;
  }
  const uint8_t* src = _front + (size_t)line * FB_ROW_BYTES;
  for (uint16_t k = 0; k < FB_ROW_BYTES; k++) {
    o[2 * k] = s_lut[src[k]][0];
    o[2 * k + 1] = s_lut[src[k]][1];
  }
  // Blanking clocks: HS low. Uniform and word-aligned, so no byte swap needed.
  for (uint16_t i = DATA_BYTES / 4; i < ROW_BYTES / 4; i++) o[i] = vs;
}

void IRAM_ATTR EL512Panel::fillChunk(uint16_t chunk) {
  uint16_t line = _chunkLine[chunk];
  uint8_t* out = _ring + (size_t)chunk * LINES_PER_CHUNK * ROW_BYTES;
  for (uint16_t i = 0; i < LINES_PER_CHUNK; i++) {
    if (line == 0 && _swapPending) {
      const uint32_t* s = (const uint32_t*)_fb;
      uint32_t* d = (uint32_t*)_front;
      for (size_t w = 0; w < FRAMEBUFFER_BYTES / 4; w++) d[w] = s[w];
      _swapPending = false;
      _swaps = _swaps + 1;
    }
    encodeRow(line, out + (size_t)i * ROW_BYTES);
    line = (line + 1 == LINES_PER_FRAME) ? 0 : line + 1;
  }
}

void IRAM_ATTR EL512Panel::isr(void* arg) {
  auto* self = static_cast<EL512Panel*>(arg);
  const uint32_t now = esp_cpu_get_cycle_count();
  if (self->_lastIsrCycle != 0 && now - self->_lastIsrCycle > self->_maxGap) self->_maxGap = now - self->_lastIsrCycle;
  self->_lastIsrCycle = now;
  self->_refills = self->_refills + 1;
  // Health counters (see takeStats): TX FIFO ran empty / transmitter hung.
  const uint32_t raw = I2S0.int_raw.val;
  if (raw & (I2S_TX_REMPTY_INT_RAW_M | I2S_TX_HUNG_INT_RAW_M)) {
    self->_underruns = self->_underruns + 1;
    I2S0.int_clr.val = I2S_TX_REMPTY_INT_CLR_M | I2S_TX_HUNG_INT_CLR_M;
  }
  const uint32_t status = I2S0.int_st.val;
  I2S0.int_clr.val = status;
  if (!(status & I2S_OUT_EOF_INT_ST_M)) return;

  // The descriptor that just finished plays again RING_LINES lines later.
  const lldesc_t* done = (const lldesc_t*)I2S0.out_eof_des_addr;
  const uint16_t chunk = done - (const lldesc_t*)self->_desc;
  if (chunk >= CHUNK_COUNT) return;
  uint16_t next = self->_chunkLine[chunk] + RING_LINES;
  if (next >= LINES_PER_FRAME) next -= LINES_PER_FRAME;
  self->_chunkLine[chunk] = next;
  self->fillChunk(chunk);
}

bool EL512Panel::startI2S() {
  _desc = (lldesc_t*)heap_caps_calloc(CHUNK_COUNT, sizeof(lldesc_t), MALLOC_CAP_DMA);
  if (!_desc) return false;
  for (uint16_t i = 0; i < CHUNK_COUNT; i++) {
    _chunkLine[i] = i * LINES_PER_CHUNK;
    fillChunk(i);
    lldesc_t& d = _desc[i];
    d.size = LINES_PER_CHUNK * ROW_BYTES;
    d.length = d.size;
    d.owner = 1;
    d.eof = 1;  // EOF interrupt after every chunk
    d.buf = _ring + (size_t)i * LINES_PER_CHUNK * ROW_BYTES;
    d.qe.stqe_next = &_desc[(i + 1) % CHUNK_COUNT];
  }

  periph_module_enable(PERIPH_I2S0_MODULE);
  resetI2S0();

  // 8-bit parallel LCD mode (bitluni ESP32Lib setup; the i ^ 2 byte order in
  // the lookup table depends on these FIFO/channel modes).
  I2S0.conf2.val = 0;
  I2S0.conf2.lcd_en = 1;
  I2S0.conf2.lcd_tx_wrx2_en = 1;
  I2S0.conf2.lcd_tx_sdx2_en = 0;

  I2S0.sample_rate_conf.val = 0;
  I2S0.sample_rate_conf.tx_bits_mod = 8;
  I2S0.sample_rate_conf.tx_bck_div_num = 1;

  // PLL_D2_CLK / (N + b/a) with b = 0: 160 MHz / 16 = 10 MHz bytes -> 5 MHz BCK.
  I2S0.clkm_conf.val = 0;
  I2S0.clkm_conf.clka_en = 0;
  I2S0.clkm_conf.clkm_div_num = I2S_CLKM_DIV;
  I2S0.clkm_conf.clkm_div_a = 1;
  I2S0.clkm_conf.clkm_div_b = 0;

  I2S0.fifo_conf.val = 0;
  I2S0.fifo_conf.tx_fifo_mod_force_en = 1;
  I2S0.fifo_conf.tx_fifo_mod = 1;
  I2S0.fifo_conf.tx_data_num = 32;
  I2S0.fifo_conf.dscr_en = 1;

  I2S0.conf1.val = 0;
  I2S0.conf1.tx_stop_en = 0;
  I2S0.conf1.tx_pcm_bypass = 1;

  I2S0.conf_chan.val = 0;
  I2S0.conf_chan.tx_chan_mod = 1;

  I2S0.conf.tx_right_first = 1;
  I2S0.conf.tx_msb_right = 0;
  I2S0.conf.tx_msb_shift = 0;
  I2S0.conf.tx_mono = 0;
  I2S0.conf.tx_short_sync = 0;
  I2S0.timing.val = 0;

  // Stream bit n -> DATA_OUT(16 + n); BCK -> VCLK.
  const int8_t dataPins[] = {_pins.vid, _pins.tvid, _pins.hs, _pins.vs};
  for (uint32_t bit = 0; bit < sizeof(dataPins); bit++) {
    routeOutput(dataPins[bit], I2S0_DATA_SIGNAL_BASE + bit, false);
  }
  routeOutput(_pins.vclk, I2S0O_BCK_OUT_IDX, INVERT_VCLK);

  // EOF interrupt refills the ring. IRAM so it keeps running during flash writes.
  I2S0.int_ena.val = 0;
  I2S0.int_clr.val = I2S0.int_raw.val;
  if (esp_intr_alloc(ETS_I2S0_INTR_SOURCE, ESP_INTR_FLAG_IRAM, &EL512Panel::isr, this,
                     &_intr) != ESP_OK) {
    _intr = nullptr;
    return false;
  }
  I2S0.int_ena.out_eof = 1;

  // Start the DMA; it loops over the chain until stopI2S().
  I2S0.lc_conf.val = I2S_OUT_DATA_BURST_EN_M | I2S_OUTDSCR_BURST_EN_M;
  I2S0.out_link.addr = (uint32_t)&_desc[0];
  I2S0.out_link.start = 1;
  I2S0.conf.tx_start = 1;
  return true;
}

void EL512Panel::stopI2S() {
  if (!_desc) return;  // never started

  I2S0.int_ena.val = 0;
  if (_intr) {
    esp_intr_free(_intr);
    _intr = nullptr;
  }
  I2S0.conf.tx_start = 0;
  I2S0.out_link.stop = 1;
  resetI2S0();
  periph_module_disable(PERIPH_I2S0_MODULE);

  // Hand the pins back to GPIO and park them low.
  const int8_t pins[] = {_pins.vclk, _pins.vid, _pins.tvid, _pins.hs, _pins.vs};
  for (int8_t pin : pins) {
    if (pin < 0) continue;
    esp_rom_gpio_connect_out_signal(pin, SIG_GPIO_OUT_IDX, false, false);
    gpio_set_level((gpio_num_t)pin, 0);
  }

  heap_caps_free(_desc);
  _desc = nullptr;
}