#include <sys/types.h>
#include <sys/mman.h>
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <unistd.h>
#include <assert.h>
#include <fcntl.h>

#include "voodoo2.h"
#include "voodoo2_bits.h"
#include "voodoo2_config.h"
#include "voodoo2_config_bits.h"
#include "ics5342.h"

static inline bool _wait_graphics_busy(voodoo2_reg * voodoo2)
{
  int cnt = 0;
  for (int i = 0; i < 4096; i++) {
    if (voodoo2->status & STATUS__CHUCK_BUSY) {
      cnt = 0;
    } else {
      cnt += 1;
    }
    if (cnt >= 3)
      return true;
  }
  return false;
}

#define wait_graphics_busy(voodoo2) assert(_wait_graphics_busy(voodoo2));

static inline void dac_data_read(voodoo2_reg * voodoo2,
                                 int data,
                                 int rs)
{
  voodoo2->dacData
    = DACDATA__WRITE_DATA(data)
    | DACDATA__ADDRESS_2_0(rs >> 0)
    | DACDATA__READ_COMMAND
    | DACDATA__ADDRESS_4_3(rs >> 3)
    ;
}

static inline void dac_data_write(voodoo2_reg * voodoo2,
                                  int data,
                                  int rs)
{
  voodoo2->dacData
    = DACDATA__WRITE_DATA(data)
    | DACDATA__ADDRESS_2_0(rs >> 0)
    | DACDATA__ADDRESS_4_3(rs >> 3)
    ;
}

static inline void dac_write_pll_8(voodoo2_reg * voodoo2,
                                   int address,
                                   int parameter_m)
{
  dac_data_write(voodoo2, address, DAC__RS__PLL_ADDRESS_WRITE);
  wait_graphics_busy(voodoo2);

  dac_data_write(voodoo2, parameter_m, DAC__RS__PLL_PARAMETER);
  wait_graphics_busy(voodoo2);
}

static inline void dac_write_pll_16(voodoo2_reg * voodoo2,
                                    int address,
                                    int parameter_m,
                                    int parameter_n)
{
  dac_data_write(voodoo2, address, DAC__RS__PLL_ADDRESS_WRITE);
  wait_graphics_busy(voodoo2);

  dac_data_write(voodoo2, parameter_m, DAC__RS__PLL_PARAMETER);
  wait_graphics_busy(voodoo2);

  dac_data_write(voodoo2, parameter_n, DAC__RS__PLL_PARAMETER);
  wait_graphics_busy(voodoo2);
}

typedef struct mn {
  uint8_t m;
  uint8_t n;
} mn_t;

static inline mn_t dac_read_pll_16(voodoo2_reg * voodoo2,
                                   int address)
{
  dac_data_write(voodoo2, address, DAC__RS__PLL_ADDRESS_READ);
  wait_graphics_busy(voodoo2);

  dac_data_read(voodoo2, 0, DAC__RS__PLL_PARAMETER);
  wait_graphics_busy(voodoo2);
  int m = voodoo2->fbiInit2 & 0xff;

  dac_data_read(voodoo2, 0, DAC__RS__PLL_PARAMETER);
  wait_graphics_busy(voodoo2);
  int n = voodoo2->fbiInit2 & 0xff;

  return (mn_t){ m, n };
}

uint8_t read_fd_u8(int fd, uint32_t offset)
{
  off_t ret = lseek(fd, offset, SEEK_SET);
  assert(ret == offset);
  uint8_t value;
  ssize_t len = read(fd, &value, 1);
  assert(len == 1);
  return value;
}

uint16_t read_fd_u16(int fd, uint32_t offset)
{
  off_t ret = lseek(fd, offset, SEEK_SET);
  assert(ret == offset);
  uint16_t value;
  ssize_t len = read(fd, &value, 2);
  assert(len == 2);
  return value;
}

void write_fd_u8(int fd, uint32_t offset, uint8_t value)
{
  off_t ret = lseek(fd, offset, SEEK_SET);
  assert(ret == offset);
  ssize_t len = write(fd, &value, 1);
  assert(len == 1);
}

void write_fd_u32(int fd, uint32_t offset, uint32_t value)
{
  off_t ret = lseek(fd, offset, SEEK_SET);
  assert(ret == offset);
  ssize_t len = write(fd, &value, 4);
  assert(len == 4);
}

int main()
{
  const char * config_path = "/sys/bus/pci/devices/0000:02:0a.0/config";
  int config_fd = open(config_path, O_RDWR | O_SYNC);
  assert(config_fd >= 0);

  uint16_t vendor_id = read_fd_u16(config_fd, PCI__CONFIG__VENDOR_ID);
  uint16_t device_id = read_fd_u16(config_fd, PCI__CONFIG__DEVICE_ID);
  printf("Vendor ID: %04x\n", vendor_id);
  printf("Device ID: %04x\n", device_id);
  assert(vendor_id == 0x121a);
  assert(device_id == 0x0002);

  ////////////////////////////////////////////////////////////////////////
  // PCI enable
  ////////////////////////////////////////////////////////////////////////
  {
    const char * enable_path = "/sys/bus/pci/devices/0000:02:0a.0/enable";
    int enable_fd = open(enable_path, O_RDWR | O_SYNC);
    assert(enable_fd >= 0);

    if (read_fd_u8(enable_fd, 0) == '0') {
      write_fd_u8(enable_fd, 0, '1'); // enable
      assert(read_fd_u8(enable_fd, 0) != '0');
    }
    close(enable_fd);
  }

  ////////////////////////////////////////////////////////////////////////
  // PCI resource0
  ////////////////////////////////////////////////////////////////////////
  const char * resource0_path = "/sys/bus/pci/devices/0000:02:0a.0/resource0";
  int resource0_fd = open(resource0_path, O_RDWR | O_SYNC);
  assert(resource0_fd >= 0);

  uint32_t resource0_size = 16 * 1024 * 1024;
  void * resource0_base = mmap(0, resource0_size, PROT_READ | PROT_WRITE, MAP_SHARED, resource0_fd, 0);
  assert(resource0_base != MAP_FAILED);

  voodoo2_reg * voodoo2 = (voodoo2_reg *)resource0_base;

  int init_enable;

  ////////////////////////////////////////////////////////////////////////
  // reset
  ////////////////////////////////////////////////////////////////////////
  init_enable
    = INIT_ENABLE__ENABLE_WRITES_TO_HARDWARE_REGISTERS;
  write_fd_u32(config_fd, PCI__CONFIG__INIT_ENABLE, init_enable);
  wait_graphics_busy(voodoo2);

  voodoo2->fbiInit0
    = FBIINIT0__CHUCK_GRAPHICS_RESET
    | FBIINIT0__CHUCK_FIFO_RESET
    ;
  wait_graphics_busy(voodoo2);

  voodoo2->fbiInit1
    = FBIINIT1__VIDEO_TIMING_RESET
    ;
  wait_graphics_busy(voodoo2);

  voodoo2->fbiInit2
    = 0; // disable DRAM refresh
  wait_graphics_busy(voodoo2);

  ////////////////////////////////////////////////////////////////////////
  // DAC initialization
  ////////////////////////////////////////////////////////////////////////

  // enable fbiInit2/dacRead remap
  init_enable
    = INIT_ENABLE__ENABLE_WRITES_TO_HARDWARE_REGISTERS
    | INIT_ENABLE__REMAP_FBIINIT2_FBIINIT3_TO_DACREAD_VIDEOCHECKSUM;
  write_fd_u32(config_fd, PCI__CONFIG__INIT_ENABLE, init_enable);
  wait_graphics_busy(voodoo2);

  int command = DAC__COMMAND__COLOR_MODE__16BIT_DIRECT_COLOR_WITH_BYPASS;
  dac_data_write(voodoo2, command, DAC__RS__COMMAND);
  wait_graphics_busy(voodoo2);
  dac_data_read(voodoo2, command, DAC__RS__COMMAND);
  wait_graphics_busy(voodoo2);
  int command_read = voodoo2->fbiInit2 & 0xff;
  printf("dac read: DAC__RS__COMMAND: m: %02x\n", command_read);

  /*
    pixel clock: 40MHz
    graphics/memory clock: 44.6MHz
  */

  int m = 54;
  int n = 67;
  dac_write_pll_16(voodoo2, DAC__PLL_PARAMETER__CLK0_f0_PLL, m, n);
  mn_t clk0_f0_res = dac_read_pll_16(voodoo2, DAC__PLL_PARAMETER__CLK0_f0_PLL);
  printf("dac read: DAC__PLL_PARAMETER__CLK0_f0_PLL: m: %02x n: %02x\n",
         clk0_f0_res.m, clk0_f0_res.n);

  int pll_control
    = DAC__PLL_CONTROL__CLK0_SELECT(0) // f0
    | DAC__PLL_CONTROL__CLK1_SELECT(0) // fA
    | DAC__PLL_CONTROL__ENABLE_INTERNAL_CLOCK_SELECT;
  dac_write_pll_8(voodoo2, DAC__PLL_PARAMETER__PLL_CONTROL, pll_control);
  mn_t pll_control_res = dac_read_pll_16(voodoo2, DAC__PLL_PARAMETER__PLL_CONTROL);
  printf("dac read: DAC__PLL_PARAMETER__PLL_CONTROL: m: %02x\n",
         pll_control_res.m);

  // disable fbiInit2/dacRead remap
  init_enable
    = INIT_ENABLE__ENABLE_WRITES_TO_HARDWARE_REGISTERS;
  write_fd_u32(config_fd, PCI__CONFIG__INIT_ENABLE, init_enable);
  wait_graphics_busy(voodoo2);

  ////////////////////////////////////////////////////////////////////////
  // framebuffer initialization
  ////////////////////////////////////////////////////////////////////////
  voodoo2->fbiInit0
    = FBIINIT0__VGA_PASSTHROUGH // disable VGA passthrough
    | FBIINIT0__STALL_PCI_ENABLE_FOR_HIGH_WATER_MARK
    | FBIINIT0__PCI_FIFO_EMPTY_ENTRIES_LOW_WATER_MARK(0x10)
    ;
  wait_graphics_busy(voodoo2);

  int x_tiles = 832 / 32;
  int x_tiles_0 = (x_tiles >> 0) & 0b1;
  int x_tiles_4_1 = (x_tiles >> 1) & 0b1111;
  int x_tiles_5 = (x_tiles >> 5) & 0b1;

  voodoo2->fbiInit1
    = FBIINIT1__ENABLE_LINEAR_FRAME_BUFFER_READS
    | FBIINIT1__NUMBER_OF_32X32_VIDEO_TILES_IN_HORIZONTAL_DIMENSION_4_1(x_tiles_4_1)
    | FBIINIT1__DRIVE_VIDEO_TIMING_DATA_OUTPUTS
    | FBIINIT1__DRIVE_VIDEO_TIMING_BLANK_OUTPUTS
    | FBIINIT1__DRIVE_VIDEO_TIMING_HSYNC_VSYNC_OUTPUTS
    | FBIINIT1__DRIVE_VIDEO_TIMING_DCLK_OUTPUT
    | FBIINIT1__VIDEO_TIMING_VCLK_INPUT_SELECT(0)
    | FBIINIT1__VIDEO_TIMING_VCLK_SOURCE_SELECT__VID_CLK_2X_SEL
    | FBIINIT1__NUMBER_OF_32X32_VIDEO_TILES_IN_HORIZONTAL_DIMENSION_5(x_tiles_5)
    ;
  wait_graphics_busy(voodoo2);

  voodoo2->fbiInit2
    = FBIINIT2__DRAM_BANKING_CONFIGURATION
    | FBIINIT2__ENABLE_FAST_RAS_READ_CYCLES
    | FBIINIT2__ENABLE_GENERATED_DRAM_OE_SIGNAL
    | FBIINIT2__VIDEO_BUFFER_OFFSET(247)
    | FBIINIT2__ENABLE_DRAM_READ_AHEAD_FIFO
    | FBIINIT2__REFRESH_ENABLE
    | FBIINIT2__REFRESH_LOAD_VALUE(0x30)
    ;
  wait_graphics_busy(voodoo2);

  voodoo2->fbiInit3
    = FBIINIT3__DISABLE_TEXTURE_MAPPING
    ;
  wait_graphics_busy(voodoo2);

  voodoo2->fbiInit4
    = FBIINIT4__ENABLE_READ_AHEAD_LOGIC_FOR_LINEAR_FRAME_BUFFER_READS
    ;
  wait_graphics_busy(voodoo2);

  voodoo2->fbiInit5
    = FBIINIT5__INVERT_DAC_HSYNC_OUTPUT_TO_DAC
    | FBIINIT5__INVERT_DAC_VSYNC_OUTPUT_TO_DAC
    ;
  wait_graphics_busy(voodoo2);

  voodoo2->fbiInit6
    = FBIINIT6__NUMBER_OF_32X32_VIDEO_TILES_IN_HORIZONTAL_DIMENSION_0(x_tiles_0)
    ;
  wait_graphics_busy(voodoo2);

  ////////////////////////////////////////////////////////////////////////
  // video timing initialization
  ////////////////////////////////////////////////////////////////////////

  int y_height = 600;
  int x_width = 800;

  int v_back_porch = 628 - 605;
  int h_back_porch = 1056 - 968;

  int v_sync_on = 605 - 601;
  int h_sync_on = 968 - 840;

  int v_sync_off = 628 - v_sync_on;
  int h_sync_off = 1056 - h_sync_on;

  voodoo2->hSync
    = HSYNC__H_SYNC_ON(h_sync_on - 1)
    | HSYNC__H_SYNC_OFF(h_sync_off)
    ;
  wait_graphics_busy(voodoo2);

  voodoo2->vSync
    = VSYNC__V_SYNC_ON(v_sync_on)
    | VSYNC__V_SYNC_OFF(v_sync_off)
    ;
  wait_graphics_busy(voodoo2);

  voodoo2->backPorch
    = BACKPORCH__H_BACK_PORCH(h_back_porch - 2)
    | BACKPORCH__V_BACK_PORCH(v_back_porch)
    ;
  wait_graphics_busy(voodoo2);

  voodoo2->videoDimensions
    = VIDEODIMENSIONS__X_WIDTH(x_width - 1)
    | VIDEODIMENSIONS__Y_HEIGHT(y_height)
    ;
  wait_graphics_busy(voodoo2);

  ////////////////////////////////////////////////////////////////////////
  // panda
  ////////////////////////////////////////////////////////////////////////

  init_enable
    = INIT_ENABLE__ENABLE_WRITES_TO_PCI_FIFO;
  write_fd_u32(config_fd, PCI__CONFIG__INIT_ENABLE, init_enable);
  wait_graphics_busy(voodoo2);

  static const uint16_t panda[] __attribute__((aligned(4))) = {
    #include "panda.rgb565.inc"
  };
  static const uint16_t bear[] __attribute__((aligned(4))) = {
    #include "bear.rgb565.inc"
  };

  while (true) {
    volatile uint16_t * framebuffer
      = (volatile uint16_t * )(((ptrdiff_t)resource0_base) + 0x0400000);

    printf("panda\n");
    for (int y = 0; y < 600; y++) {
      for (int x = 0; x < 800; x++) {
        framebuffer[y * 1024 + x] = panda[y * 800 + x];
        wait_graphics_busy(voodoo2);
      }
    }
    sleep(1);
    printf("bear\n");
    for (int y = 0; y < 600; y++) {
      for (int x = 0; x < 800; x++) {
        framebuffer[y * 1024 + x] = bear[y * 800 + x];
        wait_graphics_busy(voodoo2);
      }
    }
    sleep(1);
  }

  ////////////////////////////////////////////////////////////////////////
  // cleanup
  ////////////////////////////////////////////////////////////////////////
  close(config_fd);
  close(resource0_fd);

  return 0;
}