YMF278.cc

Go to the documentation of this file.

// Based on ymf278b.c written by R. Belmont and O. Galibert

// Improved by Valley Bell, 2018
// Thanks to niekniek and l_oliveira for providing recordings from OPL4 hardware.
// Thanks to superctr and wouterv for discussing changes.
//
// Improvements:
// - added TL interpolation, recordings show that internal TL levels are 0x00..0xff
// - fixed ADSR speeds, attack rate 15 is now instant
// - correct clamping of intermediate Rate Correction values
// - emulation of "loop glitch" (going out-of-bounds by playing a sample faster than it the loop is long)
// - made calculation of sample position cleaner and closer to how the HW works
// - increased output resolution from TL (0.375dB) to envelope (0.09375dB)
// - fixed volume table -6dB steps are done using bit shifts, steps in between are multiplicators
// - made octave -8 freeze the sample
// - verified that TL and envelope levels are applied separately, both go silent at -60dB
// - implemented pseudo-reverb and damping according to manual
// - made pseudo-reverb ignore Rate Correction (real hardware ignores it)
// - reimplemented LFO, speed exactly matches the formulas that were probably used when creating the manual
// - fixed LFO (tremolo) amplitude modulation
// - made LFO vibrato and tremolo accurate to hardware
//
// Known issues:
// - Octave -8 was only tested with fnum 0. Other fnum values might behave differently.

// This class doesn't model a full YMF278b chip. Instead it only models the
// wave part. The FM part in modeled in YMF262 (it's almost 100% compatible,
// the small differences are handled in YMF262). The status register and
// interaction with the FM registers (e.g. the NEW2 bit) is currently handled
// in the MSXMoonSound class.

#include "YMF278.hh"
#include "DeviceConfig.hh"
#include "MSXMotherBoard.hh"
#include "MSXException.hh"
#include "Math.hh"
#include "likely.hh"
#include "outer.hh"
#include "ranges.hh"
#include "serialize.hh"

namespace openmsx {

// envelope output entries
// fixed to match recordings from actual OPL4 -Valley Bell
static const int MAX_ATT_INDEX = 0x280; // makes attack phase right and also goes well with "envelope stops at -60dB"
static const int MIN_ATT_INDEX = 0;
static const int TL_SHIFT      = 2; // envelope values are 4x as fine as TL levels

static const unsigned LFO_SHIFT = 18; // LFO period of up to 0x40000 sample
static const unsigned LFO_PERIOD = 1 << LFO_SHIFT;

// Envelope Generator phases
static const int EG_ATT = 4;
static const int EG_DEC = 3;
static const int EG_SUS = 2;
static const int EG_REL = 1;
static const int EG_OFF = 0;
// these 2 are only used in old savestates (and are converted to EG_REL on load)
static const int EG_REV = 5; // pseudo reverb
static const int EG_DMP = 6; // damp

// Pan values, units are -3dB, i.e. 8.
static const uint8_t pan_left[16]  = {
        0, 8, 16, 24, 32, 40, 48, 255, 255,   0,  0,  0,  0,  0,  0, 0
};
static const uint8_t pan_right[16] = {
        0, 0,  0,  0,  0,  0,  0,   0, 255, 255, 48, 40, 32, 24, 16, 8
};

// decay level table (3dB per step)
// 0 - 15: 0, 3, 6, 9,12,15,18,21,24,27,30,33,36,39,42,93 (dB)
#define SC(dB) int16_t((dB) / 3 * 0x20)
static const int16_t dl_tab[16] = {
 SC( 0), SC( 3), SC( 6), SC( 9), SC(12), SC(15), SC(18), SC(21),
 SC(24), SC(27), SC(30), SC(33), SC(36), SC(39), SC(42), SC(93)
};
#undef SC

static const byte RATE_STEPS = 8;
static const byte eg_inc[15 * RATE_STEPS] = {
//cycle:0  1   2  3   4  5   6  7
        0, 1,  0, 1,  0, 1,  0, 1, //  0  rates 00..12 0 (increment by 0 or 1)
        0, 1,  0, 1,  1, 1,  0, 1, //  1  rates 00..12 1
        0, 1,  1, 1,  0, 1,  1, 1, //  2  rates 00..12 2
        0, 1,  1, 1,  1, 1,  1, 1, //  3  rates 00..12 3

        1, 1,  1, 1,  1, 1,  1, 1, //  4  rate 13 0 (increment by 1)
        1, 1,  1, 2,  1, 1,  1, 2, //  5  rate 13 1
        1, 2,  1, 2,  1, 2,  1, 2, //  6  rate 13 2
        1, 2,  2, 2,  1, 2,  2, 2, //  7  rate 13 3

        2, 2,  2, 2,  2, 2,  2, 2, //  8  rate 14 0 (increment by 2)
        2, 2,  2, 4,  2, 2,  2, 4, //  9  rate 14 1
        2, 4,  2, 4,  2, 4,  2, 4, // 10  rate 14 2
        2, 4,  4, 4,  2, 4,  4, 4, // 11  rate 14 3

        4, 4,  4, 4,  4, 4,  4, 4, // 12  rates 15 0, 15 1, 15 2, 15 3 for decay
        8, 8,  8, 8,  8, 8,  8, 8, // 13  rates 15 0, 15 1, 15 2, 15 3 for attack (zero time)
        0, 0,  0, 0,  0, 0,  0, 0, // 14  infinity rates for attack and decay(s)
};

#define O(a) ((a) * RATE_STEPS)
static const byte eg_rate_select[64] = {
        O(14),O(14),O(14),O(14), // inf rate
        O( 0),O( 1),O( 2),O( 3),
        O( 0),O( 1),O( 2),O( 3),
        O( 0),O( 1),O( 2),O( 3),
        O( 0),O( 1),O( 2),O( 3),
        O( 0),O( 1),O( 2),O( 3),
        O( 0),O( 1),O( 2),O( 3),
        O( 0),O( 1),O( 2),O( 3),
        O( 0),O( 1),O( 2),O( 3),
        O( 0),O( 1),O( 2),O( 3),
        O( 0),O( 1),O( 2),O( 3),
        O( 0),O( 1),O( 2),O( 3),
        O( 0),O( 1),O( 2),O( 3),
        O( 4),O( 5),O( 6),O( 7),
        O( 8),O( 9),O(10),O(11),
        O(12),O(12),O(12),O(12),
};
#undef O

// rate  0,    1,    2,    3,   4,   5,   6,  7,  8,  9,  10, 11, 12, 13, 14, 15
// shift 12,   11,   10,   9,   8,   7,   6,  5,  4,  3,  2,  1,  0,  0,  0,  0
// mask  4095, 2047, 1023, 511, 255, 127, 63, 31, 15, 7,  3,  1,  0,  0,  0,  0
#define O(a) (a)
static const byte eg_rate_shift[64] = {
        O(12),O(12),O(12),O(12),
        O(11),O(11),O(11),O(11),
        O(10),O(10),O(10),O(10),
        O( 9),O( 9),O( 9),O( 9),
        O( 8),O( 8),O( 8),O( 8),
        O( 7),O( 7),O( 7),O( 7),
        O( 6),O( 6),O( 6),O( 6),
        O( 5),O( 5),O( 5),O( 5),
        O( 4),O( 4),O( 4),O( 4),
        O( 3),O( 3),O( 3),O( 3),
        O( 2),O( 2),O( 2),O( 2),
        O( 1),O( 1),O( 1),O( 1),
        O( 0),O( 0),O( 0),O( 0),
        O( 0),O( 0),O( 0),O( 0),
        O( 0),O( 0),O( 0),O( 0),
        O( 0),O( 0),O( 0),O( 0),
};
#undef O


// number of steps the LFO counter advances per sample
#define O(a) int((LFO_PERIOD * (a)) / 44100.0 + 0.5)  // LFO frequency (Hz) -> LFO counter steps per sample
static const int lfo_period[8] = {
        O(0.168), // step:  1, period: 262144 samples
        O(2.019), // step: 12, period:  21845 samples
        O(3.196), // step: 19, period:  13797 samples
        O(4.206), // step: 25, period:  10486 samples
        O(5.215), // step: 31, period:   8456 samples
        O(5.888), // step: 35, period:   7490 samples
        O(6.224), // step: 37, period:   7085 samples
        O(7.066), // step: 42, period:   6242 samples
};
#undef O


// formula used by Yamaha docs:
//     vib_depth_cents(x) = (log2(0x400 + x) - 10) * 1200
static const int16_t vib_depth[8] = {
        0,      //  0.000 cents
        2,      //  3.378 cents
        3,      //  5.065 cents
        4,      //  6.750 cents
        6,      // 10.114 cents
        12,     // 20.170 cents
        24,     // 40.106 cents
        48,     // 79.307 cents
};


// formula used by Yamaha docs:
//     am_depth_db(x) = (x-1) / 0x40 * 6.0
//     They use (x-1), because the depth is multiplied with the AM counter, which has a range of 0..0x7F.
//     Thus the maximum attenuation with x=0x80 is (0x7F * 0x80) >> 7 = 0x7F.
// reversed formula:
//     am_depth(dB) = round(dB / 6.0 * 0x40) + 1
static const uint8_t am_depth[8] = {
        0x00,   //  0.000 dB
        0x14,   //  1.781 dB
        0x20,   //  2.906 dB
        0x28,   //  3.656 dB
        0x30,   //  4.406 dB
        0x40,   //  5.906 dB
        0x50,   //  7.406 dB
        0x80,   // 11.910 dB
};


YMF278::Slot::Slot()
{
        reset();
}

// Sign extend a 4-bit value to int (32-bit)
// require: x in range [0..15]
static int sign_extend_4(int x)
{
        return (x ^ 8) - 8;
}

// Params: oct in [-8 ..   +7]
//         fn  in [ 0 .. 1023]
// We want to interpret oct as a signed 4-bit number and calculate
//    ((fn | 1024) + vib) << (5 + sign_extend_4(oct))
// Though in this formula the shift can go over a negative distance (in that
// case we should shift in the other direction).
static unsigned calcStep(int8_t oct, uint16_t fn, int16_t vib = 0)
{
        if (oct == -8) return 0;
        unsigned t = (fn + 1024 + vib) << (8 + oct); // use '+' iso '|' (generates slightly better code)
        return t >> 3; // was shifted 3 positions too far
}

void YMF278::Slot::reset()
{
        wave = FN = OCT = TLdest = TL = pan = vib = AM = 0;
        DL = AR = D1R = D2R = RC = RR = 0;
        PRVB = keyon = DAMP = false;
        stepptr = 0;
        step = calcStep(OCT, FN);
        bits = startaddr = loopaddr = endaddr = 0;
        env_vol = MAX_ATT_INDEX;

        lfo_active = false;
        lfo_cnt = 0;

        state = EG_OFF;

        // not strictly needed, but avoid UMR on savestate
        pos = sample1 = sample2 = 0;
}

int YMF278::Slot::compute_rate(int val) const
{
        if (val == 0) {
                return 0;
        } else if (val == 15) {
                return 63;
        }
        int res = val * 4;
        if (RC != 15) {
                // clamping verified with HW tests -Valley Bell
                res += 2 * Math::clip<0, 15>(OCT + RC);
                res += (FN & 0x200) ? 1 : 0;
        }
        return Math::clip<0, 63>(res);
}

int YMF278::Slot::compute_decay_rate(int val) const
{
        if (DAMP) {
                // damping
                // The manual lists these values for time and attenuation: (44100 samples/second)
                // -12dB at  5.8ms, sample 256
                // -48dB at  8.0ms, sample 352
                // -72dB at  9.4ms, sample 416
                // -96dB at 10.9ms, sample 480
                // This results in these durations and rate values for the respective phases:
                //   0dB .. -12dB: 256 samples (5.80ms) -> 128 samples per -6dB = rate 48
                // -12dB .. -48dB:  96 samples (2.18ms) ->  16 samples per -6dB = rate 63
                // -48dB .. -72dB:  64 samples (1.45ms) ->  16 samples per -6dB = rate 63
                // -72dB .. -96dB:  64 samples (1.45ms) ->  16 samples per -6dB = rate 63
                // Damping was verified to ignore rate correction.
                if (env_vol < dl_tab[4]) {
                        return 48; //   0dB .. -12dB
                } else {
                        return 63; // -12dB .. -96dB
                }
        }
        if (PRVB) {
                // pseudo reverb
                // activated when reaching -18dB, overrides D1R/D2R/RR with reverb rate 5
                //
                // The manual is actually a bit unclear and just says "RATE=5",
                // referring to the D1R/D2R/RR register value. However, later
                // pages use "RATE" to refer to the "internal" rate, which is
                // (register * 4) + rate correction. HW recordings prove that
                // Rate Correction is ignored, so pseudo reverb just sets the
                // "internal" rate to a value of 4*5 = 20.
                if (env_vol >= dl_tab[6]) {
                        return 20;
                }
        }
        return compute_rate(val);
}

int16_t YMF278::Slot::compute_vib() const
{
        // verified via hardware recording:
        //  With LFO speed 0 (period 262144 samples), each vibrato step takes
        //  4096 samples.
        //  -> 64 steps total
        //  Also, with vibrato depth 7 (80 cents) and an F-Num of 0x400, the
        //  final F-Nums are: 0x400 .. 0x43C, 0x43C .. 0x400, 0x400 .. 0x3C4,
        //  0x3C4 .. 0x400
        int16_t lfo_fm = lfo_cnt / (LFO_PERIOD / 0x40);
        // results in +0x00..+0x0F, +0x0F..+0x00, -0x00..-0x0F, -0x0F..-0x00
        if (lfo_fm & 0x10) lfo_fm ^= 0x1F;
        if (lfo_fm & 0x20) lfo_fm = -(lfo_fm & 0x0F);

        return (lfo_fm * vib_depth[vib]) / 12;
}

uint16_t YMF278::Slot::compute_am() const
{
        // verified via hardware recording:
        //  With LFO speed 0 (period 262144 samples), each tremolo step takes
        //  1024 samples.
        //  -> 256 steps total
        uint16_t lfo_am = lfo_cnt / (LFO_PERIOD / 0x100);
        // results in 0x00..0x7F, 0x7F..0x00
        if (lfo_am >= 0x80) lfo_am ^= 0xFF;

        return (lfo_am * am_depth[AM]) >> 7;
}


void YMF278::advance()
{
        eg_cnt++;

        // modulo counters for volume interpolation
        int tl_int_cnt  =  eg_cnt % 9;      // 0 .. 8
        int tl_int_step = (eg_cnt / 9) % 3; // 0 .. 2

        for (auto& op : slots) {
                // volume interpolation
                if (tl_int_cnt == 0) {
                        if (tl_int_step == 0) {
                                // decrease volume by one step every 27 samples
                                if (op.TL < op.TLdest) ++op.TL;
                        } else {
                                // increase volume by one step every 13.5 samples
                                if (op.TL > op.TLdest) --op.TL;
                        }
                }

                if (op.lfo_active) {
                        op.lfo_cnt = (op.lfo_cnt + lfo_period[op.lfo]) & (LFO_PERIOD - 1);
                }

                // Envelope Generator
                switch (op.state) {
                case EG_ATT: { // attack phase
                        uint8_t rate = op.compute_rate(op.AR);
                        // Verified by HW recording (and matches Nemesis' tests of the YM2612):
                        // AR = 0xF during KeyOn results in instant switch to EG_DEC. (see keyOnHelper)
                        // Setting AR = 0xF while the attack phase is in progress freezes the envelope.
                        if (rate >= 63) {
                                break;
                        }
                        uint8_t shift = eg_rate_shift[rate];
                        if (!(eg_cnt & ((1 << shift) - 1))) {
                                uint8_t select = eg_rate_select[rate];
                                // >>4 makes the attack phase's shape match the actual chip -Valley Bell
                                op.env_vol += (~op.env_vol * eg_inc[select + ((eg_cnt >> shift) & 7)]) >> 4;
                                if (op.env_vol <= MIN_ATT_INDEX) {
                                        op.env_vol = MIN_ATT_INDEX;
                                        // TODO does the real HW skip EG_DEC completely,
                                        //      or is it active for 1 sample?
                                        op.state = op.DL ? EG_DEC : EG_SUS;
                                }
                        }
                        break;
                }
                case EG_DEC: { // decay phase
                        uint8_t rate = op.compute_decay_rate(op.D1R);
                        uint8_t shift = eg_rate_shift[rate];
                        if (!(eg_cnt & ((1 << shift) - 1))) {
                                uint8_t select = eg_rate_select[rate];
                                op.env_vol += eg_inc[select + ((eg_cnt >> shift) & 7)];
                                if (op.env_vol >= op.DL) {
                                        op.state = (op.env_vol < MAX_ATT_INDEX) ? EG_SUS : EG_OFF;
                                }
                        }
                        break;
                }
                case EG_SUS: { // sustain phase
                        uint8_t rate = op.compute_decay_rate(op.D2R);
                        uint8_t shift = eg_rate_shift[rate];
                        if (!(eg_cnt & ((1 << shift) - 1))) {
                                uint8_t select = eg_rate_select[rate];
                                op.env_vol += eg_inc[select + ((eg_cnt >> shift) & 7)];
                                if (op.env_vol >= MAX_ATT_INDEX) {
                                        op.env_vol = MAX_ATT_INDEX;
                                        op.state = EG_OFF;
                                }
                        }
                        break;
                }
                case EG_REL: { // release phase
                        uint8_t rate = op.compute_decay_rate(op.RR);
                        uint8_t shift = eg_rate_shift[rate];
                        if (!(eg_cnt & ((1 << shift) - 1))) {
                                uint8_t select = eg_rate_select[rate];
                                op.env_vol += eg_inc[select + ((eg_cnt >> shift) & 7)];
                                if (op.env_vol >= MAX_ATT_INDEX) {
                                        op.env_vol = MAX_ATT_INDEX;
                                        op.state = EG_OFF;
                                }
                        }
                        break;
                }
                case EG_OFF:
                        // nothing
                        break;

                default:
                        UNREACHABLE;
                }
        }
}

int16_t YMF278::getSample(Slot& op)
{
        // TODO How does this behave when R#2 bit 0 = 1?
        //      As-if read returns 0xff? (Like for CPU memory reads.) Or is
        //      sound generation blocked at some higher level?
        int16_t sample;
        switch (op.bits) {
        case 0: {
                // 8 bit
                sample = readMem(op.startaddr + op.pos) << 8;
                break;
        }
        case 1: {
                // 12 bit
                unsigned addr = op.startaddr + ((op.pos / 2) * 3);
                if (op.pos & 1) {
                        sample = (readMem(addr + 2) << 8) |
                                 ((readMem(addr + 1) << 4) & 0xF0);
                } else {
                        sample = (readMem(addr + 0) << 8) |
                                 (readMem(addr + 1) & 0xF0);
                }
                break;
        }
        case 2: {
                // 16 bit
                unsigned addr = op.startaddr + (op.pos * 2);
                sample = (readMem(addr + 0) << 8) |
                         (readMem(addr + 1));
                break;
        }
        default:
                // TODO unspecified
                sample = 0;
        }
        return sample;
}

bool YMF278::anyActive()
{
        return ranges::any_of(slots, [](auto& op) { return op.state != EG_OFF; });
}

// In: 'envVol', 0=max volume, others -> -3/32 = -0.09375 dB/step
// Out: 'x' attenuated by the corresponding factor.
// Note: microbenchmarks have shown that re-doing this calculation is about the
// same speed as using a 4kB lookup table.
static int vol_factor(int x, unsigned envVol)
{
        if (envVol >= MAX_ATT_INDEX) return 0; // hardware clips to silence below -60dB
        int vol_mul = 0x80 - (envVol & 0x3F); // 0x40 values per 6dB
        int vol_shift = 7 + (envVol >> 6);
        return (x * ((0x8000 * vol_mul) >> vol_shift)) >> 15;
}

void YMF278::setMixLevel(uint8_t x, EmuTime::param time)
{
        static const float level[8] = {
                (1.00f / 1), //   0dB
                (0.75f / 1), //  -3dB (approx)
                (1.00f / 2), //  -6dB
                (0.75f / 2), //  -9dB (approx)
                (1.00f / 4), // -12dB
                (0.75f / 4), // -15dB (approx)
                (1.00f / 8), // -18dB
                (0.00f    ), // -inf dB
        };
        setSoftwareVolume(level[x & 7], level[(x >> 3) & 7], time);
}

void YMF278::generateChannels(float** bufs, unsigned num)
{
        if (!anyActive()) {
                // TODO update internal state, even if muted
                // TODO also mute individual channels
                for (int i = 0; i < 24; ++i) {
                        bufs[i] = nullptr;
                }
                return;
        }

        for (unsigned j = 0; j < num; ++j) {
                for (int i = 0; i < 24; ++i) {
                        auto& sl = slots[i];
                        if (sl.state == EG_OFF) {
                                //bufs[i][2 * j + 0] += 0;
                                //bufs[i][2 * j + 1] += 0;
                                continue;
                        }

                        int16_t sample = (sl.sample1 * (0x10000 - sl.stepptr) +
                                          sl.sample2 * sl.stepptr) >> 16;
                        // TL levels are 00..FF internally (TL register value 7F is mapped to TL level FF)
                        // Envelope levels have 4x the resolution (000..3FF)
                        // Volume levels are approximate logarithmic. -6dB result in half volume. Steps in between use linear interpolation.
                        // A volume of -60dB or lower results in silence. (value 0x280..0x3FF).
                        // Recordings from actual hardware indicate that TL level and envelope level are applied separarely.
                        // Each of them is clipped to silence below -60dB, but TL+envelope might result in a lower volume. -Valley Bell
                        uint16_t envVol = std::min(sl.env_vol + ((sl.lfo_active && sl.AM) ? sl.compute_am() : 0),
                                                   MAX_ATT_INDEX);
                        int smplOut = vol_factor(vol_factor(sample, envVol), sl.TL << TL_SHIFT);

                        // Panning is also done separately. (low-volume TL + low-volume panning goes below -60dB)
                        // I'll be taking wild guess and assume that -3dB is approximated with 75%. (same as with TL and envelope levels)
                        // The same applies to the PCM mix level.
                        int32_t volLeft  = pan_left [sl.pan]; // note: register 0xF9 is handled externally
                        int32_t volRight = pan_right[sl.pan];
                        // 0 -> 0x20, 8 -> 0x18, 16 -> 0x10, 24 -> 0x0C, etc. (not using vol_factor here saves array boundary checks)
                        volLeft  = (0x20 - (volLeft  & 0x0f)) >> (volLeft  >> 4);
                        volRight = (0x20 - (volRight & 0x0f)) >> (volRight >> 4);

                        bufs[i][2 * j + 0] += (smplOut * volLeft ) >> 5;
                        bufs[i][2 * j + 1] += (smplOut * volRight) >> 5;

                        unsigned step = (sl.lfo_active && sl.vib)
                                      ? calcStep(sl.OCT, sl.FN, sl.compute_vib())
                                      : sl.step;
                        sl.stepptr += step;

                        // If there is a 4-sample loop and you advance 12 samples per step,
                        // it may exceed the end offset.
                        // This is abused by the "Lizard Star" song to generate noise at 0:52. -Valley Bell
                        if (sl.stepptr >= 0x10000) {
                                sl.sample1 = sl.sample2;
                                sl.sample2 = getSample(sl);
                                sl.pos += (sl.stepptr >> 16);
                                sl.stepptr &= 0xffff;
                                if ((uint32_t(sl.pos) + sl.endaddr) >= 0x10000) { // check position >= (negated) end address
                                        sl.pos += sl.endaddr + sl.loopaddr; // This is how the actual chip does it.
                                }
                        }
                }
                advance();
        }
}

void YMF278::keyOnHelper(YMF278::Slot& slot)
{
        // Unlike FM, the envelope level is reset. (And it makes sense, because you restart the sample.)
        slot.env_vol = MAX_ATT_INDEX;
        if (slot.compute_rate(slot.AR) < 63) {
                slot.state = EG_ATT;
        } else {
                // Nuke.YKT verified that the FM part does it exactly this way,
                // and the OPL4 manual says it's instant as well.
                slot.env_vol = MIN_ATT_INDEX;
                // see comment in 'case EG_ATT' in YMF278::advance()
                slot.state = slot.DL ? EG_DEC : EG_SUS;
        }
        slot.stepptr = 0;
        slot.pos = 0;
        slot.sample1 = getSample(slot);
        slot.pos = 1;
        slot.sample2 = getSample(slot);
}

void YMF278::writeReg(byte reg, byte data, EmuTime::param time)
{
        updateStream(time); // TODO optimize only for regs that directly influence sound
        writeRegDirect(reg, data, time);
}

void YMF278::writeRegDirect(byte reg, byte data, EmuTime::param time)
{
        // Handle slot registers specifically
        if (reg >= 0x08 && reg <= 0xF7) {
                int snum = (reg - 8) % 24;
                auto& slot = slots[snum];
                switch ((reg - 8) / 24) {
                case 0: {
                        slot.wave = (slot.wave & 0x100) | data;
                        int wavetblhdr = (regs[2] >> 2) & 0x7;
                        int base = (slot.wave < 384 || !wavetblhdr) ?
                                   (slot.wave * 12) :
                                   (wavetblhdr * 0x80000 + ((slot.wave - 384) * 12));
                        byte buf[12];
                        for (int i = 0; i < 12; ++i) {
                                // TODO What if R#2 bit 0 = 1?
                                //      See also getSample()
                                buf[i] = readMem(base + i);
                        }
                        slot.bits = (buf[0] & 0xC0) >> 6;
                        slot.startaddr = buf[2] | (buf[1] << 8) | ((buf[0] & 0x3F) << 16);
                        slot.loopaddr = buf[4] | (buf[3] << 8);
                        slot.endaddr  = buf[6] | (buf[5] << 8);
                        for (int i = 7; i < 12; ++i) {
                                // Verified on real YMF278:
                                // After tone loading, if you read these
                                // registers, their value actually has changed.
                                writeRegDirect(8 + snum + (i - 2) * 24, buf[i], time);
                        }
                        if (slot.keyon) {
                                keyOnHelper(slot);
                        }
                        break;
                }
                case 1: {
                        slot.wave = (slot.wave & 0xFF) | ((data & 0x1) << 8);
                        slot.FN = (slot.FN & 0x380) | (data >> 1);
                        slot.step = calcStep(slot.OCT, slot.FN);
                        break;
                }
                case 2: {
                        slot.FN = (slot.FN & 0x07F) | ((data & 0x07) << 7);
                        slot.PRVB = (data & 0x08) != 0;
                        slot.OCT = sign_extend_4((data & 0xF0) >> 4);
                        slot.step = calcStep(slot.OCT, slot.FN);
                        break;
                }
                case 3: {
                        uint8_t t = data >> 1;
                        slot.TLdest = (t != 0x7f) ? t : 0xff; // verified on HW via volume interpolation
                        if (data & 1) {
                                // directly change volume
                                slot.TL = slot.TLdest;
                        } else {
                                // interpolate volume
                        }
                        break;
                }
                case 4:
                        if (data & 0x10) {
                                // output to DO1 pin:
                                // this pin is not used in moonsound
                                // we emulate this by muting the sound
                                slot.pan = 8; // both left/right -inf dB
                        } else {
                                slot.pan = data & 0x0F;
                        }

                        if (data & 0x20) {
                                // LFO reset
                                slot.lfo_active = false;
                                slot.lfo_cnt = 0;
                        } else {
                                // LFO activate
                                slot.lfo_active = true;
                        }

                        slot.DAMP = (data & 0x40) != 0;

                        if (data & 0x80) {
                                if (!slot.keyon) {
                                        slot.keyon = true;
                                        keyOnHelper(slot);
                                }
                        } else {
                                if (slot.keyon) {
                                        slot.keyon = false;
                                        slot.state = EG_REL;
                                }
                        }
                        break;
                case 5:
                        slot.lfo = (data >> 3) & 0x7;
                        slot.vib = data & 0x7;
                        break;
                case 6:
                        slot.AR  = data >> 4;
                        slot.D1R = data & 0xF;
                        break;
                case 7:
                        slot.DL  = dl_tab[data >> 4];
                        slot.D2R = data & 0xF;
                        break;
                case 8:
                        slot.RC = data >> 4;
                        slot.RR = data & 0xF;
                        break;
                case 9:
                        slot.AM = data & 0x7;
                        break;
                }
        } else {
                // All non-slot registers
                switch (reg) {
                case 0x00: // TEST
                case 0x01:
                        break;

                case 0x02:
                        // wave-table-header / memory-type / memory-access-mode
                        // Simply store in regs[2]
                        break;

                case 0x03:
                        // Verified on real YMF278:
                        // * Don't update the 'memadr' variable on writes to
                        //   reg 3 and 4. Only store the value in the 'regs'
                        //   array for later use.
                        // * The upper 2 bits are not used to address the
                        //   external memories (so from a HW pov they don't
                        //   matter). But if you read back this register, the
                        //   upper 2 bits always read as '0' (even if you wrote
                        //   '1'). So we mask the bits here already.
                        data &= 0x3F;
                        break;

                case 0x04:
                        // See reg 3.
                        break;

                case 0x05:
                        // Verified on real YMF278: (see above)
                        // Only writes to reg 5 change the (full) 'memadr'.
                        memadr = (regs[3] << 16) | (regs[4] << 8) | data;
                        break;

                case 0x06:  // memory data
                        if (regs[2] & 1) {
                                writeMem(memadr, data);
                                ++memadr; // no need to mask (again) here
                        } else {
                                // Verified on real YMF278:
                                //  - writes are ignored
                                //  - memadr is NOT increased
                        }
                        break;

                case 0xf8: // These are implemented in MSXMoonSound.cc
                case 0xf9:
                        break;
                }
        }

        regs[reg] = data;
}

byte YMF278::readReg(byte reg)
{
        // no need to call updateStream(time)
        byte result = peekReg(reg);
        if (reg == 6) {
                // Memory Data Register
                if (regs[2] & 1) {
                        // Verified on real YMF278:
                        // memadr is only increased when 'regs[2] & 1'
                        ++memadr; // no need to mask (again) here
                }
        }
        return result;
}

byte YMF278::peekReg(byte reg) const
{
        byte result;
        switch (reg) {
                case 2: // 3 upper bits are device ID
                        result = (regs[2] & 0x1F) | 0x20;
                        break;

                case 6: // Memory Data Register
                        if (regs[2] & 1) {
                                result = readMem(memadr);
                        } else {
                                // Verified on real YMF278
                                result = 0xff;
                        }
                        break;

                default:
                        result = regs[reg];
                        break;
        }
        return result;
}

YMF278::YMF278(const std::string& name_, int ramSize_,
               const DeviceConfig& config)
        : ResampledSoundDevice(config.getMotherBoard(), name_, "MoonSound wave-part",
                               24, true)
        , motherBoard(config.getMotherBoard())
        , debugRegisters(motherBoard, getName())
        , debugMemory   (motherBoard, getName())
        , rom(getName() + " ROM", "rom", config)
        , ram(config, getName() + " RAM", "YMF278 sample RAM",
              ramSize_ * 1024) // size in kB
{
        if (rom.getSize() != 0x200000) { // 2MB
                throw MSXException(
                        "Wrong ROM for MoonSound (YMF278). The ROM (usually "
                        "called yrw801.rom) should have a size of exactly 2MB.");
        }
        if ((ramSize_ !=    0) &&  //   -     -
            (ramSize_ !=  128) &&  // 128kB   -
            (ramSize_ !=  256) &&  // 128kB  128kB
            (ramSize_ !=  512) &&  // 512kB   -
            (ramSize_ !=  640) &&  // 512kB  128kB
            (ramSize_ != 1024) &&  // 512kB  512kB
            (ramSize_ != 2048)) {  // 512kB  512kB  512kB  512kB
                throw MSXException(
                        "Wrong sampleram size for MoonSound (YMF278). "
                        "Got ", ramSize_, ", but must be one of "
                        "0, 128, 256, 512, 640, 1024 or 2048.");
        }

        memadr = 0; // avoid UMR

        setInputRate(44100);

        registerSound(config);
        reset(motherBoard.getCurrentTime()); // must come after registerSound() because of call to setSoftwareVolume() via setMixLevel()
}

YMF278::~YMF278()
{
        unregisterSound();
}

void YMF278::clearRam()
{
        ram.clear(0);
}

void YMF278::reset(EmuTime::param time)
{
        updateStream(time);

        eg_cnt = 0;

        for (auto& op : slots) {
                op.reset();
        }
        regs[2] = 0; // avoid UMR
        for (int i = 0xf7; i >= 0; --i) { // reverse order to avoid UMR
                writeRegDirect(i, 0, time);
        }
        memadr = 0;
        setMixLevel(0, time);
}

// This routine translates an address from the (upper) MoonSound address space
// to an address inside the (linearized) SRAM address space.
//
// The following info is based on measurements on a real MoonSound (v2.0)
// PCB. This PCB can have several possible SRAM configurations:
//   128kB:
//    1 SRAM chip of 128kB, chip enable (/CE) of this SRAM chip is connected to
//    the 1Y0 output of a 74LS139 (2-to-4 decoder). The enable input of the
//    74LS139 is connected to YMF278 pin /MCS6 and the 74LS139 1B:1A inputs are
//    connected to YMF278 pins MA18:MA17. So the SRAM is selected when /MC6 is
//    active and MA18:MA17 == 0:0.
//   256kB:
//    2 SRAM chips of 128kB. First one connected as above. Second one has /CE
//    connected to 74LS139 pin 1Y1. So SRAM2 is selected when /MSC6 is active
//    and MA18:MA17 == 0:1.
//   512kB:
//    1 SRAM chip of 512kB, /CE connected to /MCS6
//   640kB:
//    1 SRAM chip of 512kB, /CE connected to /MCS6
//    1 SRAM chip of 128kB, /CE connected to /MCS7.
//      (This means SRAM2 is potentially mirrored over a 512kB region)
//  1024kB:
//    1 SRAM chip of 512kB, /CE connected to /MCS6
//    1 SRAM chip of 512kB, /CE connected to /MCS7
//  2048kB:
//    1 SRAM chip of 512kB, /CE connected to /MCS6
//    1 SRAM chip of 512kB, /CE connected to /MCS7
//    1 SRAM chip of 512kB, /CE connected to /MCS8
//    1 SRAM chip of 512kB, /CE connected to /MCS9
//      This configuration is not so easy to create on the v2.0 PCB. So it's
//      very rare.
//
// So the /MCS6 and /MCS7 (and /MCS8 and /MCS9 in case of 2048kB) signals are
// used to select the different SRAM chips. The meaning of these signals
// depends on the 'memory access mode'. This mode can be changed at run-time
// via bit 1 in register 2. The following table indicates for which regions
// these signals are active (normally MoonSound should be used with mode=0):
//              mode=0              mode=1
//  /MCS6   0x200000-0x27FFFF   0x380000-0x39FFFF
//  /MCS7   0x280000-0x2FFFFF   0x3A0000-0x3BFFFF
//  /MCS8   0x300000-0x37FFFF   0x3C0000-0x3DFFFF
//  /MCS9   0x380000-0x3FFFFF   0x3E0000-0x3FFFFF
//
// (For completeness) MoonSound also has 2MB ROM (YRW801), /CE of this ROM is
// connected to YMF278 /MCS0. In both mode=0 and mode=1 this signal is active
// for the region 0x000000-0x1FFFFF. (But this routine does not handle ROM).
unsigned YMF278::getRamAddress(unsigned addr) const
{
        addr -= 0x200000; // RAM starts at 0x200000
        if (unlikely(regs[2] & 2)) {
                // Normally MoonSound is used in 'memory access mode = 0'. But
                // in the rare case that mode=1 we adjust the address.
                if ((0x180000 <= addr) && (addr <= 0x1FFFFF)) {
                        addr -= 0x180000;
                        switch (addr & 0x060000) {
                        case 0x000000: // [0x380000-0x39FFFF]
                                // 1st 128kB of SRAM1
                                break;
                        case 0x020000: // [0x3A0000-0x3BFFFF]
                                if (ram.getSize() == 256 * 1024) {
                                        // 2nd 128kB SRAM chip
                                } else {
                                        // 2nd block of 128kB in SRAM2
                                        // In case of 512+128, we use mirroring
                                        addr += 0x080000;
                                }
                                break;
                        case 0x040000: // [0x3C0000-0x3DFFFF]
                                // 3rd 128kB block in SRAM3
                                addr += 0x100000;
                                break;
                        case 0x060000: // [0x3EFFFF-0x3FFFFF]
                                // 4th 128kB block in SRAM4
                                addr += 0x180000;
                                break;
                        }
                } else {
                        addr = unsigned(-1); // unmapped
                }
        }
        if (ram.getSize() == 640 * 1024) {
                // Verified on real MoonSound cartridge (v2.0): In case of
                // 640kB (1x512kB + 1x128kB), the 128kB SRAM chip is 4 times
                // visible. None of the other SRAM configurations show similar
                // mirroring (because the others are powers of two).
                if (addr > 0x080000) {
                        addr &= ~0x060000;
                }
        }
        return addr;
}

byte YMF278::readMem(unsigned address) const
{
        // Verified on real YMF278: address space wraps at 4MB.
        address &= 0x3FFFFF;
        if (address < 0x200000) {
                // ROM connected to /MCS0
                return rom[address];
        } else {
                unsigned ramAddr = getRamAddress(address);
                if (ramAddr < ram.getSize()) {
                        return ram[ramAddr];
                } else {
                        // unmapped region
                        return 255; // TODO check
                }
        }
}

void YMF278::writeMem(unsigned address, byte value)
{
        address &= 0x3FFFFF;
        if (address < 0x200000) {
                // can't write to ROM
        } else {
                unsigned ramAddr = getRamAddress(address);
                if (ramAddr < ram.getSize()) {
                        ram.write(ramAddr, value);
                } else {
                        // can't write to unmapped memory
                }
        }
}

// version 1: initial version, some variables were saved as char
// version 2: serialization framework was fixed to save/load chars as numbers
//            but for backwards compatibility we still load old savestates as
//            characters
// version 3: 'step' is no longer stored (it is recalculated)
// version 4:
//  - removed members: 'lfo', 'LD', 'active'
//  - new members 'TLdest', 'keyon', 'DAMP' restored from registers instead of serialized
//  - store 'OCT' sign-extended
//  - store 'endaddr' as 2s complement
//  - removed EG_DMP and EG_REV enum values from 'state'
// version 5:
//  - re-added 'lfo' member. This is not stored in the savestate, instead it's
//    restored from register values in YMF278::serialize()
//  - removed members 'lfo_step' and ' 'lfo_max'
//  - 'lfo_cnt' has changed meaning (but we don't try to translate old to new meaning)
template<typename Archive>
void YMF278::Slot::serialize(Archive& ar, unsigned version)
{
        // TODO restore more state from registers
        ar.serialize("startaddr", startaddr);
        ar.serialize("loopaddr", loopaddr);
        ar.serialize("stepptr", stepptr);
        ar.serialize("pos", pos);
        ar.serialize("sample1", sample1);
        ar.serialize("sample2", sample2);
        ar.serialize("env_vol", env_vol);
        ar.serialize("lfo_cnt", lfo_cnt);
        ar.serialize("DL", DL);
        ar.serialize("wave", wave);
        ar.serialize("FN", FN);
        if (ar.versionAtLeast(version, 4)) {
                ar.serialize("endaddr", endaddr);
                ar.serialize("OCT", OCT);
        } else {
                unsigned e = 0; ar.serialize("endaddr", e); endaddr = (e ^ 0xffff) + 1;

                char O = 0;
                if (ar.versionAtLeast(version, 2)) {
                        ar.serialize("OCT", O);
                } else {
                        ar.serializeChar("OCT", O);
                }
                OCT = sign_extend_4(O);
        }

        if (ar.versionAtLeast(version, 2)) {
                ar.serialize("PRVB", PRVB);
                ar.serialize("TL", TL);
                ar.serialize("pan", pan);
                ar.serialize("vib", vib);
                ar.serialize("AM", AM);
                ar.serialize("AR", AR);
                ar.serialize("D1R", D1R);
                ar.serialize("D2R", D2R);
                ar.serialize("RC", RC);
                ar.serialize("RR", RR);
        } else {
                // for backwards compatibility with old savestates
                char PRVB_ = 0; ar.serializeChar("PRVB", PRVB_); PRVB = PRVB_;
                char TL_  = 0; ar.serializeChar("TL",  TL_ ); TL  = TL_;
                char pan_ = 0; ar.serializeChar("pan", pan_); pan = pan_;
                char vib_ = 0; ar.serializeChar("vib", vib_); vib = vib_;
                char AM_  = 0; ar.serializeChar("AM",  AM_ ); AM  = AM_;
                char AR_  = 0; ar.serializeChar("AR",  AR_ ); AR  = AR_;
                char D1R_ = 0; ar.serializeChar("D1R", D1R_); D1R = D1R_;
                char D2R_ = 0; ar.serializeChar("D2R", D2R_); D2R = D2R_;
                char RC_  = 0; ar.serializeChar("RC",  RC_ ); RC  = RC_;
                char RR_  = 0; ar.serializeChar("RR",  RR_ ); RR  = RR_;
        }
        ar.serialize("bits", bits);
        ar.serialize("lfo_active", lfo_active);

        ar.serialize("state", state);
        if (ar.versionBelow(version, 4)) {
                assert(ar.isLoader());
                if ((state == EG_REV) || (state == EG_DMP)) {
                        state = EG_REL;
                }
        }

        // Recalculate redundant state
        if (ar.isLoader()) {
                step = calcStep(OCT, FN);
        }

        // This old comment is NOT completely true:
        //    Older version also had "env_vol_step" and "env_vol_lim" but those
        //    members were nowhere used, so removed those in the current
        //    version (it's ok to remove members from the savestate without
        //    updating the version number).
        // When you remove member variables without increasing the version
        // number, new openMSX executables can still read old savestates. And
        // if you try to load a new savestate in an old openMSX version you do
        // get a (cryptic) error message. But if the version number is
        // increased the error message is much clearer.
}

// version 1: initial version
// version 2: loadTime and busyTime moved to MSXMoonSound class
// version 3: memadr cannot be restored from register values
// version 4: implement ram via Ram class
template<typename Archive>
void YMF278::serialize(Archive& ar, unsigned version)
{
        ar.serialize("slots", slots);
        ar.serialize("eg_cnt", eg_cnt);
        if (ar.versionAtLeast(version, 4)) {
                ar.serialize("ram", ram);
        } else {
                ar.serialize_blob("ram", ram.getWriteBackdoor(), ram.getSize());
        }
        ar.serialize_blob("registers", regs, sizeof(regs));
        if (ar.versionAtLeast(version, 3)) { // must come after 'regs'
                ar.serialize("memadr", memadr);
        } else {
                assert(ar.isLoader());
                // Old formats didn't store 'memadr' so we also can't magically
                // restore the correct value. The best we can do is restore the
                // last set address.
                regs[3] &= 0x3F; // mask upper two bits
                memadr = (regs[3] << 16) | (regs[4] << 8) | regs[5];
        }

        // TODO restore more state from registers
        if (ar.isLoader()) {
                for (int i = 0; i < 24; ++i) {
                        Slot& sl = slots[i];

                        auto t = regs[0x50 + i] >> 1;
                        sl.TLdest = (t != 0x7f) ? t : 0xff;

                        sl.keyon = (regs[0x68 + i] & 0x80) != 0;
                        sl.DAMP  = (regs[0x68 + i] & 0x40) != 0;
                        sl.lfo   = (regs[0x80 + i] >> 3) & 7;
                }
        }
}
INSTANTIATE_SERIALIZE_METHODS(YMF278);


// class DebugRegisters

YMF278::DebugRegisters::DebugRegisters(MSXMotherBoard& motherBoard_,
                                       const std::string& name_)
        : SimpleDebuggable(motherBoard_, name_ + " regs",
                           "OPL4 registers", 0x100)
{
}

byte YMF278::DebugRegisters::read(unsigned address)
{
        auto& ymf278 = OUTER(YMF278, debugRegisters);
        return ymf278.peekReg(address);
}

void YMF278::DebugRegisters::write(unsigned address, byte value, EmuTime::param time)
{
        auto& ymf278 = OUTER(YMF278, debugRegisters);
        ymf278.writeReg(address, value, time);
}


// class DebugMemory

YMF278::DebugMemory::DebugMemory(MSXMotherBoard& motherBoard_,
                                 const std::string& name_)
        : SimpleDebuggable(motherBoard_, name_ + " mem",
                           "OPL4 memory (includes both ROM and RAM)", 0x400000) // 4MB
{
}

byte YMF278::DebugMemory::read(unsigned address)
{
        auto& ymf278 = OUTER(YMF278, debugMemory);
        return ymf278.readMem(address);
}

void YMF278::DebugMemory::write(unsigned address, byte value)
{
        auto& ymf278 = OUTER(YMF278, debugMemory);
        ymf278.writeMem(address, value);
}

} // namespace openmsx