using System;

namespace Core.Video
{
    public class SmsVdp
    {
        // The VDP's private memory! The CPU cannot touch these arrays directly.
        public byte[] VRAM { get; private set; } = new byte[0x4000]; // 16KB Video RAM
        public byte[] CRAM { get; private set; } = new byte[0x20];   // 32 Bytes Color Palette
        public byte[] Registers { get; private set; } = new byte[16]; // 11 Hardware Control Registers
        public int[] FrameBuffer { get; private set; } = new int[256 * 192];

        // The Control Port State Machine (Port 0xBF)
        private bool _isSecondControlByte = false;
        private ushort _controlWord = 0;
        private byte _readBuffer = 0;

        private int _tStateCounter = 0;
        private int _currentScanline = 0;
        private byte _statusRegister = 0x00;

        public bool InterruptPending => (_statusRegister & 0x80) != 0 && (Registers[1] & 0x20) != 0;

        public byte ReadDataPort() // Port 0xBE
        {
            _isSecondControlByte = false; // Reading data resets the control latch
            byte value = _readBuffer;
            _readBuffer = VRAM[_controlWord & 0x3FFF];
            _controlWord++;
            return value;
        }

        public byte ReadControlPort() // Port 0xBF
        {
            _isSecondControlByte = false;
            byte currentStatus = _statusRegister;

            // CRITICAL HARDWARE QUIRK: Reading the status port physically 
            // clears the flags inside the chip! If we don't clear this, 
            // the interrupt line gets stuck on forever.
            _statusRegister = 0x00;

            return currentStatus;
        }

        public void WriteDataPort(byte value) // Port 0xBE
        {
            _isSecondControlByte = false;
            _readBuffer = value; 

            int address = _controlWord & 0x3FFF;
            int command = (_controlWord >> 14) & 0x03;

            if (command == 3) // Code 3: Write to Color Palette (CRAM)
            {
                CRAM[address & 0x1F] = value;
            }
            else // Code 0, 1, 2: Write to VRAM
            {
                VRAM[address] = value;
            }
            _controlWord++; // Auto-increment so the Z80 can blast data fast
        }

        public void WriteControlPort(byte value) // Port 0xBF
        {
            if (!_isSecondControlByte)
            {
                // First byte arrives: Store it in the lower 8 bits
                _controlWord = (ushort)((_controlWord & 0xFF00) | value);
                _isSecondControlByte = true;
            }
            else
            {
                // Second byte arrives: Store it in the upper 8 bits and execute!
                _controlWord = (ushort)((_controlWord & 0x00FF) | (value << 8));
                _isSecondControlByte = false;

                int command = (_controlWord >> 14) & 0x03;
                
                if (command == 0) // Code 0: Prep for VRAM Read
                {
                    _readBuffer = VRAM[_controlWord & 0x3FFF];
                    _controlWord++;
                }
                else if (command == 2) // Code 2: Write to Internal VDP Register
                {
                    int regIndex = value & 0x0F;
                    byte regData = (byte)(_controlWord & 0xFF);
                    
                    if (regIndex < 16) Registers[regIndex] = regData;
                }
            }
        }

        public byte ReadVCounter()
        {
            // Note: On real NTSC hardware, the V-Counter jumps slightly around 
            // the VBlank period to keep the math 8-bit, but simply returning 
            // the raw current scanline is perfectly fine to get us booting!
            return (byte)_currentScanline;
        }

        public void Update(int tStates)
        {
            _tStateCounter += tStates;

            // 228 T-States per scanline
            if (_tStateCounter >= 228)
            {
                _tStateCounter -= 228;
                _currentScanline++;

                // Line 192 is the exact moment the screen finishes drawing!
                if (_currentScanline == 192)
                {
                    _statusRegister |= 0x80; // Set Bit 7 (VBlank Flag) to 1

                    RenderBackground(); // <--- DRAW THE FRAME!
                }

                // End of the NTSC frame (262 lines)
                if (_currentScanline > 261)
                {
                    _currentScanline = 0;
                }
            }
        }

        private void RenderBackground()
        {
            // The Name Table base address is stored in VDP Register 2.
            // It tells us where in VRAM the 32x24 screen grid starts.
            ushort nameTableBase = (ushort)((Registers[2] & 0x0E) << 10);

            // Loop through all 24 rows and 32 columns of the screen
            for (int row = 0; row < 24; row++)
            {
                for (int col = 0; col < 32; col++)
                {
                    // 1. Read the 16-bit Tile instruction from the Name Table
                    ushort nameTableAddr = (ushort)(nameTableBase + (row * 64) + (col * 2));
                    byte lowByte = VRAM[nameTableAddr];
                    byte highByte = VRAM[nameTableAddr + 1];
                    ushort tileData = (ushort)((highByte << 8) | lowByte);

                    // 2. Extract the Tile Index and Palette Info
                    int tileIndex = tileData & 0x01FF;
                    bool useSpritePalette = (tileData & 0x0800) != 0;

                    // 3. Find the actual pixel data for this tile in VRAM
                    // Each 8x8 tile takes exactly 32 bytes in memory
                    ushort tileAddress = (ushort)(tileIndex * 32);

                    // 4. Draw the 8x8 block of pixels!
                    for (int y = 0; y < 8; y++)
                    {
                        // The SMS uses 4 bitplanes to make a single row of pixels.
                        byte bp0 = VRAM[tileAddress + (y * 4) + 0];
                        byte bp1 = VRAM[tileAddress + (y * 4) + 1];
                        byte bp2 = VRAM[tileAddress + (y * 4) + 2];
                        byte bp3 = VRAM[tileAddress + (y * 4) + 3];

                        for (int x = 0; x < 8; x++)
                        {
                            // Combine 1 bit from each bitplane to get a color index (0-15)
                            int shift = 7 - x;
                            int colorIndex = ((bp0 >> shift) & 1) |
                                            (((bp1 >> shift) & 1) << 1) |
                                            (((bp2 >> shift) & 1) << 2) |
                                            (((bp3 >> shift) & 1) << 3);

                            // Find the raw SMS color in CRAM
                            int paletteOffset = useSpritePalette ? 16 : 0;
                            byte smsColor = CRAM[paletteOffset + colorIndex];

                            // Translate SMS 00BBGGRR format to Windows 32-bit ARGB
                            int r = (smsColor & 0x03) * 85;
                            int g = ((smsColor >> 2) & 0x03) * 85;
                            int b = ((smsColor >> 4) & 0x03) * 85;

                            // Calculate where this pixel goes on the final 256x192 screen
                            int pixelX = (col * 8) + x;
                            int pixelY = (row * 8) + y;

                            FrameBuffer[(pixelY * 256) + pixelX] = (255 << 24) | (r << 16) | (g << 8) | b;
                        }
                    }
                }
            }
        }

    }
}