#define _CRT_SECURE_NO_WARNINGS
#include <stdio.h>
#include <conio.h>

#include <stdint.h>
#include <portaudio.h>
#include <math.h>
#include <vector>
#include <fstream>
#include <queue>

#include "wavehead.h"
#include "opmplay.h"
#include "vgm.h"
#include "rss.h"                    // YM2608 internal ADPCM-A ROM dump
#include "ymfm/src/ymfm_opn.h"

#define WIN32_LEAN_AND_MEAN
#define WIN32_EXTRA_LEAN
#define NOMINMAX
#include <Windows.h>

enum {
    CHANNELS = 2,
    FRAMES_PER_BUFFER = 512,
    MAX_FRAMES_PER_BUFFER = 4096,
};

const double OPN_CLOCK_RATE = 3500000;
//const double OPN_CLOCK_RATE = 3540000;

PaStreamParameters outputParameters;
PaStream* stream;

// console stuff
struct {
    HANDLE hStdout, hScreenBuffer;
    COORD bufcoord, bufsize;
    SMALL_RECT bufDestRect;
    CHAR_INFO* buffer;          // the main buffer to write to
} console;

struct opm_context_t {
    // full context
    opmplay_context_t opm;
    opmplay_io_t      io;

    // delay count relative to sample rate
    int32_t           delay_count;
    int32_t           delay_count_reload;
};

struct vgm_context_t {
    std::vector<uint8_t> vgmfile;
    std::vector<uint8_t>::iterator vgmfile_it;
    VGMHeader* header;
    uint32_t loop_pos;
    uint32_t start, end;        // offsets

    // delay count
    int32_t  delay_count;

    // rescaler for 44100hz delays
    double   rescaler;
};

vgm_context_t vgmctx;
opm_context_t opmctx;

class opna_interface_t : public ymfm::ymfm_interface {
protected:
    int32_t timer_a_count, timer_a_reload;
public:
    opna_interface_t() : timer_a_count(0), timer_a_reload(0) {};
    uint8_t ymfm_external_read(ymfm::access_class type, uint32_t address) {
        switch (type) {
        case ymfm::ACCESS_ADPCM_A: return YM2608_ADPCM_ROM[address & 0x1FFF];
        default: return 0;
        }
    }
    void ymfm_set_timer(uint32_t tnum, int32_t duration_in_clocks) {
        if (tnum == 0) {
            timer_a_reload = duration_in_clocks;
        }
    }
    void timer_advance(int32_t ticks) {
        // ripped from furnace :grins:
        if (timer_a_reload >= 0) {
            timer_a_count -= ticks;
            if (timer_a_count < 0) {
                m_engine->engine_timer_expired(0);
                timer_a_count += timer_a_reload;
            }
        }
    }
};

// register write queue
class opnx_register_queue_t {
private:
    uint64_t write_delay;
    uint64_t clock;
    struct reg_entry_t {
        uint64_t clock;
        int reg, data;
    };
    std::queue<reg_entry_t> reg_queue;
    ymfm::ym2203* chip;
public:
    opnx_register_queue_t() : opnx_register_queue_t(nullptr) {}
    opnx_register_queue_t(ymfm::ym2203* _chip) : clock(0), write_delay(0), chip(_chip) {}
    void set_chip(ymfm::ym2203* _chip) { chip = _chip; }
    void reset() {
        clock = 0;
        write_delay = 0;
        while (!reg_queue.empty()) reg_queue.pop(); // flush queue
    }
    void add(int chip, int reg, int data, uint64_t delay) {
        reg_entry_t r;
        r.clock = clock + write_delay;
        r.reg = reg;
        r.data = data;
        reg_queue.push(r);
        write_delay += delay;
    }
    void pop_clock() {
        if (!reg_queue.empty() && (reg_queue.front().clock <= clock)) {
            auto &r = reg_queue.front();
            if (chip) {
                chip->write_address(r.reg);
                chip->write_data(r.data);
            }
            reg_queue.pop();
        }
        clock++; write_delay = 0;
    }
};

opna_interface_t opna_interface[2];
ymfm::ym2203 *opnachip[2];
opnx_register_queue_t opna_regqueue[2];
ymfm::ym2203::output_data opna_out[MAX_FRAMES_PER_BUFFER][2];

// hack af
uint8_t opn_reg_view[2][256];

// generic output routine
void opn_write_reg(int chip, int reg, int data) {
    if (chip >= 2) return;
    opn_reg_view[chip][reg] = data;
    opna_regqueue[chip].add(0, reg, data, 4);
}

// ------------------

// draw plain string
void drawstring(const char* str, unsigned long x, unsigned long y, unsigned char attr) {
    CHAR_INFO* p = (CHAR_INFO*)console.buffer + (console.bufsize.X * y) + x;

    while (*str != '\0') {
        p->Char.AsciiChar = *str++;
        p->Attributes = attr;
        p++;
    }
}

// draw string with attributes
// '\0' - end, '\xFF\xaa' - set attribute byte 'aa'
void drawastring(const char* str, unsigned long x, unsigned long y) {
    CHAR_INFO* p = (CHAR_INFO*)console.buffer + (console.bufsize.X * y) + x;

    unsigned short attr = 0x07;

    while (*str != '\0') if (*str == '\xFF') {
        attr = (*++str); str++;
    }
    else {
        p->Char.AsciiChar = *str++;
        p->Attributes = attr;
        p++;
    }
}

// printf 
int tprintf(uint32_t x, uint32_t y, const char* format, ...) {
    char buffer[1024];      // large enough
    va_list arglist;

    va_start(arglist, format);
    int rtn = vsnprintf(buffer, sizeof(buffer), format, arglist);
    drawastring(buffer, x, y);
    va_end(arglist);

    return rtn;
};

// -------------------


int console_open() {
    // Get a handle to the STDOUT screen buffer to copy from and
    // create a new screen buffer to copy to.

    console.hStdout = GetStdHandle(STD_OUTPUT_HANDLE);
    console.hScreenBuffer = CreateConsoleScreenBuffer(
        GENERIC_READ |           // read/write access
        GENERIC_WRITE,
        FILE_SHARE_READ |
        FILE_SHARE_WRITE,        // shared
        NULL,                    // default security attributes
        CONSOLE_TEXTMODE_BUFFER, // must be TEXTMODE
        NULL);                   // reserved; must be NULL
    if (console.hStdout == INVALID_HANDLE_VALUE ||
        console.hScreenBuffer == INVALID_HANDLE_VALUE)
    {
        printf("CreateConsoleScreenBuffer failed - (%d)\n", GetLastError());
        return 1;
    }

    // resize
    console.bufsize.X = 132;
    console.bufsize.Y = 40;
    SetConsoleScreenBufferSize(console.hScreenBuffer, console.bufsize);

    // allocate console buffer
    console.buffer = new CHAR_INFO[console.bufsize.X * console.bufsize.Y];
    memset(console.buffer, 0, sizeof(CHAR_INFO) * console.bufsize.X * console.bufsize.Y);

    // Make the new screen buffer the active screen buffer.
    if (!SetConsoleActiveScreenBuffer(console.hScreenBuffer))
    {
        printf("SetConsoleActiveScreenBuffer failed - (%d)\n", GetLastError());
        return 1;
    }

    return 0;
}

void console_update() {
    console.bufDestRect.Top = 0;
    console.bufDestRect.Left = 0;
    console.bufDestRect.Bottom = console.bufsize.Y - 1;
    console.bufDestRect.Right = console.bufsize.X - 1;

    console.bufcoord.X = console.bufcoord.Y = 0;

    WriteConsoleOutput(
        console.hScreenBuffer,  // screen buffer to write to
        console.buffer,         // buffer to copy from
        console.bufsize,        // col-row size of chiBuffer
        console.bufcoord,       // top left src cell in chiBuffer
        &console.bufDestRect);  // dest. screen buffer rectangle
}

void console_done() {
    SetConsoleActiveScreenBuffer(console.hStdout);
}

// -------------------
// synth render
int synth_render(int16_t* buffer, uint32_t num_samples) {
    int samples_to_render = num_samples;

    memset(buffer, 0, sizeof(int16_t) * 2 * num_samples);

    while (samples_to_render > 0) {
        if (samples_to_render < opmctx.delay_count) {
            for (int i = 0; i < samples_to_render; i++) {
                for (int chip = 0; chip < 2; chip++) {
                    opna_regqueue[chip].pop_clock();
                    opnachip[chip]->generate(opna_out[chip] + i, 1);
                    opna_interface[chip].timer_advance(24);
                    opna_out[chip][i].clamp16();
                    buffer[2*i+0] += 0.5*(0.5 * opna_out[chip][i].data[0] + 0.5 * (1.0*opna_out[chip][i].data[1] + 0.5*opna_out[chip][i].data[2] + 0.0*opna_out[chip][i].data[3]));       // mix FM and SSG
                    buffer[2*i+1] += 0.5*(0.5 * opna_out[chip][i].data[0] + 0.5 * (0.0*opna_out[chip][i].data[1] + 0.5*opna_out[chip][i].data[2] + 1.0*opna_out[chip][i].data[3]));
                }
            }
            
            opmctx.delay_count -= samples_to_render;
            buffer += 2 * samples_to_render;
            break;
        }
        else {
            // calculate new delay
            for (int i = 0; i < opmctx.delay_count; i++) {
                for (int chip = 0; chip < 2; chip++) {
                    opna_regqueue[chip].pop_clock();
                    opnachip[chip]->generate(opna_out[chip] + i, 1);
                    opna_interface[chip].timer_advance(24);
                    opna_out[chip][i].clamp16();
                    buffer[2*i+0] += 0.5*(0.5 * opna_out[chip][i].data[0] + 0.5 * (1.0*opna_out[chip][i].data[1] + 0.5*opna_out[chip][i].data[2] + 0.0*opna_out[chip][i].data[3]));       // mix FM and SSG
                    buffer[2*i+1] += 0.5*(0.5 * opna_out[chip][i].data[0] + 0.5 * (0.0*opna_out[chip][i].data[1] + 0.5*opna_out[chip][i].data[2] + 1.0*opna_out[chip][i].data[3]));
                }
            }
            samples_to_render -= opmctx.delay_count;
            buffer += 2 * opmctx.delay_count;

            opmplay_tick(&opmctx.opm);
            opmctx.delay_count = opmctx.delay_count_reload;
        }
    }

    return 0;
}


int pa_init(double sample_rate) {
    PaError err;

    // init portaudio
    err = Pa_Initialize();
    if (err != paNoError) return 1;

    outputParameters.device = Pa_GetDefaultOutputDevice(); /* default output device */
    if (outputParameters.device == paNoDevice) {
        fprintf(stderr, "Error: No default output device.\n");
        return 1;
    }
    outputParameters.channelCount = CHANNELS;
    outputParameters.sampleFormat = paInt16;
    outputParameters.suggestedLatency = 0.04;
    outputParameters.hostApiSpecificStreamInfo = NULL;

    err = Pa_OpenStream(
        &stream,
        NULL, /* no input */
        &outputParameters,
        sample_rate,
        FRAMES_PER_BUFFER,
        0,     /* we won't output out of range samples so don't bother clipping them */
        NULL, /* no callback, use blocking API */
        NULL); /* no callback, so no callback userData */
    if (err != paNoError) return 1;

    err = Pa_StartStream(stream);
    if (err != paNoError) return 1;

    return 0;
}

int pa_write(void* data, int32_t count) {
    PaError err;
    err = Pa_WriteStream(stream, data, count);
    return 0;
}

int pa_done() {
    PaError err;
    err = Pa_StopStream(stream);
    if (err != paNoError) return 1;

    // deinit portaudio
    err = Pa_CloseStream(stream);
    if (err != paNoError) return 1;

    Pa_Terminate();
    return 0;
}

int main(int argc, char* argv[])
{
    bool render_to_wave = (argc >= 3);

    uint32_t sample_rate;
    for (int chip = 0; chip < 2; chip++) {
        opnachip[chip] = new ymfm::ym2203(opna_interface[chip]);
        if (opnachip[chip] == nullptr) {
            printf("error: unable to init ymfm!\n");
            return 1;
        }
        opnachip[chip]->reset();
        opnachip[chip]->set_fidelity(ymfm::OPN_FIDELITY_MIN);
        sample_rate = opnachip[chip]->sample_rate(OPN_CLOCK_RATE);
        vgmctx.rescaler = ((double)sample_rate / 44100.0);
        printf("sample rate - %d hz\n", sample_rate);

        opna_regqueue[chip].reset();
        opna_regqueue[chip].set_chip(opnachip[chip]);
    }

    if (!render_to_wave) {
        if (pa_init(sample_rate) != 0) {
            printf("error: unable to init PortAudio!\n");
            return 1;
        }
    }

    FILE* f = fopen(argv[1], "rb");
    if (f == NULL) {
        printf("error: unable to open file!\n");
        return 1;
    }

#if 1
    opmctx.io.type = OPMPLAY_IO_FILE;
    opmctx.io.io = f;

    int rtn;

    if ((rtn = opmplay_init(&opmctx.opm)) != OPMPLAY_ERR_OK) {
        printf("unable to init OPMPlay (error = %d)\n", rtn);
        return 1;
    }
    if ((rtn = opmplay_load_header(&opmctx.opm, &opmctx.io)) != OPMPLAY_ERR_OK) {
        printf("unable to load OPM header (error = %d)\n", rtn);
        return 1;
    };
    if ((rtn = opmplay_load_module(&opmctx.opm, &opmctx.io)) != OPMPLAY_ERR_OK) {
        printf("unable to load OPM module (error = %d)\n", rtn);
        return 1;
    };
    opmctx.delay_count_reload = opmctx.delay_count = ((double)sample_rate / 50.0);
#else
    // open VGM file, ready to parse
    std::ifstream infile(argv[1], std::ios::in | std::ios::binary);
    infile.unsetf(std::ios::skipws);

    // get filesize
    infile.seekg(0, std::ios::end);
    uint64_t fsize = infile.tellg();
    infile.seekg(0, std::ios::beg);

    // read whole file
    vgmctx.vgmfile.reserve(fsize);
    vgmctx.vgmfile.insert(vgmctx.vgmfile.begin(), std::istream_iterator<uint8_t>(infile), std::istream_iterator<uint8_t>());

    // get header
    vgmctx.header = reinterpret_cast<VGMHeader*>(vgmctx.vgmfile.data());

    // check header
    if (memcmp(vgmctx.header->id, "Vgm\x20", sizeof(vgmctx.header->id)) != 0) {
        printf("not a vaild VGM file!\n");
        return 1;
    }

    // parse basic VGM structure
    printf("VGM %d.%d file found\n", (vgmctx.header->version >> 8) & 0xFF, vgmctx.header->version & 0xFF);
    if (vgmctx.header->loopOffset != 0) vgmctx.loop_pos = vgmctx.header->loopOffset + offsetof(VGMHeader, loopOffset);
    vgmctx.end = vgmctx.header->eofOffset + offsetof(VGMHeader, eofOffset);
    vgmctx.start = ((vgmctx.header->version < 0x150) ? 0x40 : vgmctx.header->dataOffset + offsetof(VGMHeader, dataOffset));
    vgmctx.vgmfile_it = vgmctx.vgmfile.begin() + vgmctx.start;
    vgmctx.delay_count = 0;
#endif

    console_open();
    std::vector<int16_t> wavedata;

    int ff_pos = 0, ff_counter = 0;
    int16_t buf[FRAMES_PER_BUFFER * CHANNELS] = { 0 };
    while (1) {
        int rtn = synth_render(buf, FRAMES_PER_BUFFER);
        if (render_to_wave) {
            wavedata.insert(wavedata.end(), buf, buf + FRAMES_PER_BUFFER * CHANNELS);
        }
        else {
            pa_write(buf, FRAMES_PER_BUFFER);
        }
        ff_pos += FRAMES_PER_BUFFER;

        // update console
        memset(console.buffer, 0, sizeof(CHAR_INFO) * console.bufsize.X * console.bufsize.Y);

        tprintf(0, 0, "frame = %d", opmctx.opm.pos.frame);
        
        {
            int yy = 2;
            for (int ch = 0; ch < 6; ch++) {
                int cc = ch / 3;
                int co = ch % 3;
                tprintf(0, yy, "FM%d: [%02X %02X %02X %02X %02X %02X %02X] [%02X %02X %02X %02X %02X %02X %02X] [%02X %02X %02X %02X %02X %02X %02X] [%02X %02X %02X %02X %02X %02X %02X] - %02X %02X %02X",
                    ch,
                    opn_reg_view[cc][0x30 + co],
                    opn_reg_view[cc][0x40 + co],
                    opn_reg_view[cc][0x50 + co],
                    opn_reg_view[cc][0x60 + co],
                    opn_reg_view[cc][0x70 + co],
                    opn_reg_view[cc][0x80 + co],
                    opn_reg_view[cc][0x90 + co],

                    opn_reg_view[cc][0x34 + co],
                    opn_reg_view[cc][0x44 + co],
                    opn_reg_view[cc][0x54 + co],
                    opn_reg_view[cc][0x64 + co],
                    opn_reg_view[cc][0x74 + co],
                    opn_reg_view[cc][0x84 + co],
                    opn_reg_view[cc][0x94 + co],

                    opn_reg_view[cc][0x38 + co],
                    opn_reg_view[cc][0x48 + co],
                    opn_reg_view[cc][0x58 + co],
                    opn_reg_view[cc][0x68 + co],
                    opn_reg_view[cc][0x78 + co],
                    opn_reg_view[cc][0x88 + co],
                    opn_reg_view[cc][0x98 + co],

                    opn_reg_view[cc][0x3C + co],
                    opn_reg_view[cc][0x4C + co],
                    opn_reg_view[cc][0x5C + co],
                    opn_reg_view[cc][0x6C + co],
                    opn_reg_view[cc][0x7C + co],
                    opn_reg_view[cc][0x8C + co],
                    opn_reg_view[cc][0x9C + co],

                    opn_reg_view[cc][0xA0 + co],
                    opn_reg_view[cc][0xA4 + co],
                    opn_reg_view[cc][0xB0 + co]
                );

                yy++;

                if ((ch % 3) == 2) {
                    tprintf(0, yy, "AY%d: [%02X %02X] [%02X %02X] [%02X %02X] %02X %02X [%02X %02X %02X] [%02X %02X] <- %02X",
                        cc,
                        opn_reg_view[cc][0],
                        opn_reg_view[cc][1],
                        opn_reg_view[cc][2],
                        opn_reg_view[cc][3],
                        opn_reg_view[cc][4],
                        opn_reg_view[cc][5],
                        opn_reg_view[cc][6],
                        opn_reg_view[cc][7],
                        opn_reg_view[cc][8],
                        opn_reg_view[cc][9],
                        opn_reg_view[cc][10],
                        opn_reg_view[cc][11],
                        opn_reg_view[cc][12],
                        opn_reg_view[cc][13]
                    );
                    yy++;
                    tprintf(0, yy, "EC%d: [%02X %02X %02X %02X %02X %02X %02X %02X]", 
                        cc,
                        opn_reg_view[cc][0xAD],
                        opn_reg_view[cc][0xA9],
                        opn_reg_view[cc][0xAC],
                        opn_reg_view[cc][0xA8],
                        opn_reg_view[cc][0xAE],
                        opn_reg_view[cc][0xAA],
                        opn_reg_view[cc][0xA6],
                        opn_reg_view[cc][0xA2]
                    );
                    yy++;

                }
            }
        }

        console_update();

        if (_kbhit()) {
            _getch();
            break;
        }
    }

    // write wave file
    if (render_to_wave) {
        // create headers
        RIFF_Header riffHeader;
        memcpy(&riffHeader.id, "RIFF", sizeof(riffHeader.id));
        memcpy(&riffHeader.fourcc, "WAVE", sizeof(riffHeader.fourcc));
        riffHeader.size = sizeof(riffHeader.fourcc) + sizeof(fmt_Header) + sizeof(chunk_Header) + (wavedata.size() * sizeof(decltype(wavedata)::value_type));

        fmt_Header fmtHeader;
        memcpy(&fmtHeader.id, "fmt ", sizeof(fmtHeader.id));
        fmtHeader.size = sizeof(fmtHeader) - 8;
        fmtHeader.wFormatTag = 1;           // plain uncompressed PCM
        fmtHeader.nSamplesPerSec = sample_rate;
        fmtHeader.nBlockAlign = CHANNELS;
        fmtHeader.nAvgBytesPerSec = sample_rate * CHANNELS;
        fmtHeader.nChannels = CHANNELS;
        fmtHeader.wBitsPerSample = 8;

        chunk_Header dataHeader;
        memcpy(&dataHeader.id, "data", sizeof(dataHeader.id));
        dataHeader.size = (wavedata.size() * sizeof(decltype(wavedata)::value_type));

        // write wave file
        FILE* outfile = fopen("out.wav", "wb");

        fwrite(&riffHeader, sizeof(riffHeader), 1, outfile);
        fwrite(&fmtHeader, sizeof(fmtHeader), 1, outfile);
        fwrite(&dataHeader, sizeof(dataHeader), 1, outfile);
        fwrite(wavedata.data(), (wavedata.size() * sizeof(decltype(wavedata)::value_type)), 1, outfile);

        fclose(outfile);
    } else
    pa_done();

    console_done();

    return 0;
}