Skip to content

Commit

Permalink
Performance improvement: ~13% faster on ESP32 and ~21% faster on ESP8…
Browse files Browse the repository at this point in the history 
…266 (empirically tested) (#554)

* Performance improvement
- new tsf_note_on_fast returning playIndex and removing lowpass filter setup since it is not being processed in tsf_voice_render_fast anyway but is consuming a lot of cpu
- new tsf_note_off_fast using playIndex to stop voices faster
- Removing lowpass filter setup from tsf_voice_render_fast since it is not being processed
- Math tweaks trying to avoid complex math operations (specially conversions) at voice render stage
- Precalculating *F32P32 values to avoid doing them at voice render stage
* Fixing data type
* Initializing play indexes
---------
Co-authored-by: Edgar <edgar.coria@xema.tech>
  • Loading branch information
@ecoriag
ecoriag authored Oct 9, 2024
1 parent 235bcb1 commit d32d8b4
Show file tree
Hide file tree
Showing 3 changed files with 150 additions and 60 deletions.
5 changes: 3 additions & 2 deletions src/AudioGeneratorMIDI.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -443,8 +443,9 @@ int AudioGeneratorMIDI::PlayMIDI()
for (tgnum = 0; tgnum < num_tonegens; ++tgnum) { /* find which generator is playing it */
tg = &tonegen[tgnum];
if (tg->playing && tg->track == tracknum && tg->note == trk->note) {
tsf_note_off (g_tsf, tg->instrument, tg->note);
tsf_note_off_fast (g_tsf, tg->instrument, tg->note, tg->playIndex);
tg->playing = false;
tg->playIndex = -1;
trk->tonegens[tgnum] = false;
}
}
Expand Down Expand Up @@ -476,7 +477,7 @@ int AudioGeneratorMIDI::PlayMIDI()
if (tg->instrument != midi_chan_instrument[trk->chan]) { /* new instrument for this generator */
tg->instrument = midi_chan_instrument[trk->chan];
}
tsf_note_on (g_tsf, tg->instrument, tg->note, trk->velocity / 127.0); // velocity = 0...127
tg->playIndex = tsf_note_on_fast (g_tsf, tg->instrument, tg->note, trk->velocity / 127.0); // velocity = 0...127
} else {
++notes_skipped;
}
Expand Down
5 changes: 4 additions & 1 deletion src/AudioGeneratorMIDI.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -94,10 +94,13 @@ class AudioGeneratorMIDI : public AudioGenerator
unsigned long earliest_time = 0;

struct tonegen_status { /* current status of a tone generator */
bool playing; /* is it playing? */
bool playing; /* is it playing? */
char track; /* if so, which track is the note from? */
char note; /* what note is playing? */
char instrument; /* what instrument? */
int playIndex; /* is index provided?
Unique identifier generated when note starts playing.
This help us to turn the note off faster */
} tonegen[MAX_TONEGENS];

struct track_status { /* current processing point of a MIDI track */
Expand Down
200 changes: 143 additions & 57 deletions src/libtinysoundfont/tsf.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -612,11 +612,11 @@ struct tsf_voice
{
int playingPreset, playingKey, playingChannel;
struct tsf_region* region;
double pitchInputTimecents, pitchOutputFactor;
double pitchRatio;
double sourceSamplePosition;
fixed32p32 sourceSamplePositionF32P32;
float noteGainDB, panFactorLeft, panFactorRight;
unsigned int playIndex, loopStart, loopEnd;
fixed32p32 sourceSamplePositionF32P32, loopStartF32P32, loopEndF32P32, loopSizeF32P32, pitchRatioF32P32;
float noteGain, panFactorLeft, panFactorRight;
int playIndex, loopStart, loopEnd;
struct tsf_voice_envelope ampenv, modenv;
struct tsf_voice_lowpass lowpass;
struct tsf_voice_lfo modlfo, viblfo;
Expand All @@ -635,6 +635,14 @@ struct tsf_channels
int channelNum, activeChannel;
};

// Math properties of tsf_timecents2Secs* (tc2s):
// # tc2s(a + b) = tc2s(a) * tc2s(b)
// # tc2s(a - b) = tc2s(a) / tc2s(b)
// # 1 / tc2s(a) = tc2s(-a)
// # The same applies to tsf_cents2Hertz and tsf_decibelsToGain
// Path properties of tsf_decibelsToGain (db2g) and tsf_gainToDecibels (g2db)
// # db2g(g2db(a)) = a
// # g2db(db2g(a)) = a
static double tsf_timecents2Secsd(double timecents) { return TSF_POW(2.0, timecents / 1200.0); }
static float tsf_timecents2Secsf(float timecents) { return TSF_POWF(2.0f, timecents / 1200.0f); }
static float tsf_cents2Hertz(float cents) { return 8.176f * TSF_POWF(2.0f, cents / 1200.0f); }
Expand Down Expand Up @@ -1230,6 +1238,7 @@ static void tsf_voice_end(struct tsf_voice* v, float outSampleRate)
{
// Continue playing, but stop looping.
v->loopEnd = v->loopStart;
v->loopSizeF32P32 = 0;
}
}

Expand All @@ -1242,10 +1251,9 @@ static void tsf_voice_endquick(struct tsf_voice* v, float outSampleRate)
static void tsf_voice_calcpitchratio(struct tsf_voice* v, float pitchShift, float outSampleRate)
{
double note = v->playingKey + v->region->transpose + v->region->tune / 100.0;
double adjustedPitch = v->region->pitch_keycenter + (note - v->region->pitch_keycenter) * (v->region->pitch_keytrack / 100.0);
if (pitchShift) adjustedPitch += pitchShift;
v->pitchInputTimecents = adjustedPitch * 100.0;
v->pitchOutputFactor = v->region->sample_rate / (tsf_timecents2Secsd(v->region->pitch_keycenter * 100.0) * outSampleRate);
double adjustedPitch = (note - v->region->pitch_keycenter) * v->region->pitch_keytrack;
if (pitchShift) adjustedPitch += adjustedPitch * 100.0f;
v->pitchRatio = v->region->sample_rate * tsf_timecents2Secsd(adjustedPitch) / outSampleRate;
}

short tsf_read_short_cached(tsf *f, int pos)
Expand Down Expand Up @@ -1313,10 +1321,10 @@ static void tsf_voice_render(tsf* f, struct tsf_voice* v, float* outputBuffer, i
else tmpInitialFilterFc = 0, tmpModLfoToFilterFc = 0, tmpModEnvToFilterFc = 0;

if (dynamicPitchRatio) pitchRatio = 0, tmpModLfoToPitch = (float)region->modLfoToPitch, tmpVibLfoToPitch = (float)region->vibLfoToPitch, tmpModEnvToPitch = (float)region->modEnvToPitch;
else pitchRatio = tsf_timecents2Secsd(v->pitchInputTimecents) * v->pitchOutputFactor, tmpModLfoToPitch = 0, tmpVibLfoToPitch = 0, tmpModEnvToPitch = 0;
else pitchRatio = v->pitchRatio, tmpModLfoToPitch = 0, tmpVibLfoToPitch = 0, tmpModEnvToPitch = 0;

if (dynamicGain) tmpModLfoToVolume = (float)region->modLfoToVolume * 0.1f;
else noteGain = tsf_decibelsToGain(v->noteGainDB), tmpModLfoToVolume = 0;
else noteGain = v->noteGain, tmpModLfoToVolume = 0;

while (numSamples)
{
Expand All @@ -1333,10 +1341,10 @@ static void tsf_voice_render(tsf* f, struct tsf_voice* v, float* outputBuffer, i
}

if (dynamicPitchRatio)
pitchRatio = tsf_timecents2Secsd(v->pitchInputTimecents + (v->modlfo.level * tmpModLfoToPitch + v->viblfo.level * tmpVibLfoToPitch + v->modenv.level * tmpModEnvToPitch)) * v->pitchOutputFactor;
pitchRatio = v->pitchRatio * tsf_timecents2Secsd(v->modlfo.level * tmpModLfoToPitch + v->viblfo.level * tmpVibLfoToPitch + v->modenv.level * tmpModEnvToPitch);

if (dynamicGain)
noteGain = tsf_decibelsToGain(v->noteGainDB + (v->modlfo.level * tmpModLfoToVolume));
noteGain = v->noteGain * tsf_decibelsToGain(v->modlfo.level * tmpModLfoToVolume);

gainMono = noteGain * v->ampenv.level;

Expand Down Expand Up @@ -1445,60 +1453,40 @@ static void tsf_voice_render_fast(tsf* f, struct tsf_voice* v, short* outputBuff
TSF_BOOL updateModEnv = (region->modEnvToPitch || region->modEnvToFilterFc);
TSF_BOOL updateModLFO = (v->modlfo.delta && (region->modLfoToPitch || region->modLfoToFilterFc || region->modLfoToVolume));
TSF_BOOL updateVibLFO = (v->viblfo.delta && (region->vibLfoToPitch));
TSF_BOOL isLooping = (v->loopStart < v->loopEnd);
unsigned int tmpLoopStart = v->loopStart, tmpLoopEnd = v->loopEnd;
//double tmpSampleEndDbl = (double)v->sampleEnd, tmpLoopEndDbl = (double)tmpLoopEnd + 1.0;
//double tmpSourceSamplePosition = v->sourceSamplePosition;
fixed32p32 tmpSampleEndF32P32 = ((fixed32p32)(region->end)) << 32;
fixed32p32 tmpLoopStartF32P32 = ((fixed32p32)(tmpLoopStart + 1)) << 32;
fixed32p32 tmpLoopEndF32P32 = ((fixed32p32)(tmpLoopEnd + 1)) << 32;
TSF_BOOL isLooping = (v->loopSizeF32P32 > 0);
fixed32p32 tmpSampleEndF32P32 = (fixed32p32)(region->end) << 32;
fixed32p32 tmpSourceSamplePositionF32P32 = v->sourceSamplePositionF32P32;
struct tsf_voice_lowpass tmpLowpass = v->lowpass;

TSF_BOOL dynamicLowpass = (region->modLfoToFilterFc || region->modEnvToFilterFc);
float tmpSampleRate = f->outSampleRate, tmpInitialFilterFc, tmpModLfoToFilterFc, tmpModEnvToFilterFc;

TSF_BOOL dynamicPitchRatio = (region->modLfoToPitch || region->modEnvToPitch || region->vibLfoToPitch);
//double pitchRatio;
fixed32p32 pitchRatioF32P32;
fixed32p32 pitchRatioF32P32, tmpPitchRatioF32P32;
float tmpModLfoToPitch, tmpVibLfoToPitch, tmpModEnvToPitch;

TSF_BOOL dynamicGain = (region->modLfoToVolume != 0);
float noteGain, tmpModLfoToVolume;
float noteGain, tmpNoteGain, tmpModLfoToVolume;

if (dynamicLowpass) tmpInitialFilterFc = (float)region->initialFilterFc, tmpModLfoToFilterFc = (float)region->modLfoToFilterFc, tmpModEnvToFilterFc = (float)region->modEnvToFilterFc;
else tmpInitialFilterFc = 0, tmpModLfoToFilterFc = 0, tmpModEnvToFilterFc = 0;
if (dynamicPitchRatio) tmpPitchRatioF32P32 = v->pitchRatioF32P32, pitchRatioF32P32 = 0, tmpModLfoToPitch = (float)region->modLfoToPitch, tmpVibLfoToPitch = (float)region->vibLfoToPitch, tmpModEnvToPitch = (float)region->modEnvToPitch;
else pitchRatioF32P32 = v->pitchRatioF32P32, tmpModLfoToPitch = 0, tmpVibLfoToPitch = 0, tmpModEnvToPitch = 0;

if (dynamicPitchRatio) pitchRatioF32P32 = 0, tmpModLfoToPitch = (float)region->modLfoToPitch, tmpVibLfoToPitch = (float)region->vibLfoToPitch, tmpModEnvToPitch = (float)region->modEnvToPitch;
else {
double pr = tsf_timecents2Secsd(v->pitchInputTimecents) * v->pitchOutputFactor;
fixed32p32 adj = 1LL<<32;
pr *= adj;
pitchRatioF32P32 = (int64_t)pr, tmpModLfoToPitch = 0, tmpVibLfoToPitch = 0, tmpModEnvToPitch = 0;
}

if (dynamicGain) noteGain = 0, tmpModLfoToVolume = (float)region->modLfoToVolume * 0.1f;
else noteGain = tsf_decibelsToGain(v->noteGainDB), tmpModLfoToVolume = 0;
if (dynamicGain) noteGain = 0, tmpNoteGain = v->noteGain, tmpModLfoToVolume = (float)region->modLfoToVolume * 0.1f;
else noteGain = v->noteGain, tmpModLfoToVolume = 0;

while (numSamples)
{
float gainMono;
int blockSamples = (numSamples > TSF_RENDER_EFFECTSAMPLEBLOCK ? TSF_RENDER_EFFECTSAMPLEBLOCK : numSamples);
numSamples -= blockSamples;

if (dynamicLowpass)
{
float fres = tmpInitialFilterFc + v->modlfo.level * tmpModLfoToFilterFc + v->modenv.level * tmpModEnvToFilterFc;
tmpLowpass.active = (fres <= 13500.0f);
if (tmpLowpass.active) tsf_voice_lowpass_setup(&tmpLowpass, tsf_cents2Hertz(fres) / tmpSampleRate);
}

if (dynamicPitchRatio) {
pitchRatioF32P32 = tsf_timecents2Secsd(v->pitchInputTimecents + (v->modlfo.level * tmpModLfoToPitch + v->viblfo.level * tmpVibLfoToPitch + v->modenv.level * tmpModEnvToPitch)) * v->pitchOutputFactor * (1LL<<32);
}
if (v->modlfo.level || v->viblfo.level || v->modenv.level)
pitchRatioF32P32 = tmpPitchRatioF32P32 * tsf_timecents2Secsd(v->modlfo.level * tmpModLfoToPitch + v->viblfo.level * tmpVibLfoToPitch + v->modenv.level * tmpModEnvToPitch);
else
pitchRatioF32P32 = tmpPitchRatioF32P32; // If all levels are 0, just bypass conversion function
}

if (dynamicGain)
noteGain = tsf_decibelsToGain(v->noteGainDB + (v->modlfo.level * tmpModLfoToVolume));
if (dynamicGain) {
if (v->modlfo.level) noteGain = tmpNoteGain * tsf_decibelsToGain(v->modlfo.level * tmpModLfoToVolume);
else noteGain = tmpNoteGain; // If level is 0, just bypass conversion function
}

gainMono = noteGain * v->ampenv.level;
short gainMonoFP = gainMono * 32767;
Expand All @@ -1511,7 +1499,7 @@ static void tsf_voice_render_fast(tsf* f, struct tsf_voice* v, short* outputBuff
if (updateModLFO) tsf_voice_lfo_process(&v->modlfo, blockSamples);
if (updateVibLFO) tsf_voice_lfo_process(&v->viblfo, blockSamples);

while (blockSamples-- && tmpSourceSamplePositionF32P32 < tmpSampleEndF32P32)
while (blockSamples--)
{
unsigned int pos = (unsigned int)(tmpSourceSamplePositionF32P32>>32);
if (pos == 0xffffffff) pos = 0;
Expand All @@ -1523,8 +1511,14 @@ static void tsf_voice_render_fast(tsf* f, struct tsf_voice* v, short* outputBuff

// Next sample.
tmpSourceSamplePositionF32P32 += pitchRatioF32P32;
if (tmpSourceSamplePositionF32P32 >= tmpLoopEndF32P32 && isLooping)
tmpSourceSamplePositionF32P32 -= (tmpLoopEndF32P32 - tmpLoopStartF32P32 + (1LL<<32));

if (tmpSourceSamplePositionF32P32 >= v->loopEndF32P32)
{
if (isLooping)
tmpSourceSamplePositionF32P32 -= v->loopSizeF32P32;
else if (tmpSourceSamplePositionF32P32 >= tmpSampleEndF32P32)
break;
}
}

if (tmpSourceSamplePositionF32P32 >= tmpSampleEndF32P32 || v->ampenv.segment == TSF_SEGMENT_DONE)
Expand All @@ -1535,7 +1529,6 @@ static void tsf_voice_render_fast(tsf* f, struct tsf_voice* v, short* outputBuff
}

v->sourceSamplePositionF32P32 = tmpSourceSamplePositionF32P32;
if (tmpLowpass.active || dynamicLowpass) v->lowpass = tmpLowpass;
}


Expand Down Expand Up @@ -1730,13 +1723,14 @@ TSFDEF void tsf_note_on(tsf* f, int preset_index, int key, float vel)
}
voice = &f->voices[f->voiceNum - 4];
voice[1].playingPreset = voice[2].playingPreset = voice[3].playingPreset = -1;
voice[1].playIndex = voice[2].playIndex = voice[3].playIndex = -1;
}

voice->region = region;
voice->playingPreset = preset_index;
voice->playingKey = key;
voice->playIndex = voicePlayIndex;
voice->noteGainDB = f->globalGainDB - region->attenuation - tsf_gainToDecibels(1.0f / vel);
voice->noteGain = vel * tsf_decibelsToGain(f->globalGainDB - region->attenuation);

if (f->channels)
{
Expand All @@ -1752,7 +1746,6 @@ TSFDEF void tsf_note_on(tsf* f, int preset_index, int key, float vel)

// Offset/end.
voice->sourceSamplePosition = region->offset;
voice->sourceSamplePositionF32P32 = ((int64_t)region->offset)<< 32;

// Loop.
doLoop = (region->loop_mode != TSF_LOOPMODE_NONE && region->loop_start < region->loop_end);
Expand All @@ -1777,6 +1770,86 @@ TSFDEF void tsf_note_on(tsf* f, int preset_index, int key, float vel)
}
}

/**
* Returns the generated playing index that can be used to identify all voices related to this note.
*/
TSFDEF int tsf_note_on_fast(tsf* f, int preset_index, int key, float vel)
{
if (preset_index < 0 || preset_index >= f->presetNum) return -1;
if (!vel) { tsf_note_off(f, preset_index, key); return -1; }
if (f->presets[preset_index].regions == NULL) tsf_load_preset(f, f->hydra, preset_index);

short midiVelocity = (short)(vel * 127);
struct tsf_region *region, *regionEnd;

// Play all matching regions.
int voicePlayIndex = f->voicePlayIndex++;
for (region = f->presets[preset_index].regions, regionEnd = region + f->presets[preset_index].regionNum; region != regionEnd; region++)
{
if (key < region->lokey || key > region->hikey || midiVelocity < region->lovel || midiVelocity > region->hivel) continue;

struct tsf_voice *voice, *v, *vEnd; TSF_BOOL doLoop;
voice = TSF_NULL, v = f->voices, vEnd = v + f->voiceNum;
if (region->group)
{
for (; v != vEnd; v++)
if (v->playingPreset == preset_index && v->region->group == region->group) tsf_voice_endquick(v, f->outSampleRate);
else if (v->playingPreset == -1 && !voice) voice = v;
}
else for (; v != vEnd; v++) if (v->playingPreset == -1) { voice = v; break; }

if (!voice)
{
f->voiceNum += 4;
struct tsf_voice *saveVoice = f->voices;
f->voices = (struct tsf_voice*)TSF_REALLOC(f->voices, f->voiceNum * sizeof(struct tsf_voice));
if (!f->voices) {
f->voices = saveVoice;
printf("OOM, no room for new voice. Ignoring note_on\n");
return -1;
}
voice = &f->voices[f->voiceNum - 4];
voice[1].playingPreset = voice[2].playingPreset = voice[3].playingPreset = -1;
voice[1].playIndex = voice[2].playIndex = voice[3].playIndex = -1;
}

voice->region = region;
voice->playingPreset = preset_index;
voice->playingKey = key;
voice->playIndex = voicePlayIndex;
voice->noteGain = vel * tsf_decibelsToGain(f->globalGainDB - region->attenuation);

if (f->channels)
{
f->channels->setupVoice(f, voice);
}
else
{
tsf_voice_calcpitchratio(voice, 0, f->outSampleRate);
voice->pitchRatioF32P32 = voice->pitchRatio * (1LL << 32);
}

// Offset/end.
voice->sourceSamplePositionF32P32 = ((int64_t)region->offset)<< 32;

// Loop.
doLoop = (region->loop_mode != TSF_LOOPMODE_NONE && region->loop_start < region->loop_end);
voice->loopStartF32P32 = doLoop ? ((int64_t)region->loop_start + 1) << 32 : 0;
voice->loopEndF32P32 = doLoop ? ((int64_t)region->loop_end + 1) << 32 : 0;
voice->loopSizeF32P32 = doLoop ? voice->loopEndF32P32 - voice->loopStartF32P32 + (1LL << 32) : 0;

// Setup envelopes.
tsf_voice_envelope_setup(&voice->ampenv, &region->ampenv, key, midiVelocity, TSF_TRUE, f->outSampleRate);
tsf_voice_envelope_setup(&voice->modenv, &region->modenv, key, midiVelocity, TSF_FALSE, f->outSampleRate);

// Setup LFO filters.
tsf_voice_lfo_setup(&voice->modlfo, region->delayModLFO, region->freqModLFO, f->outSampleRate);
tsf_voice_lfo_setup(&voice->viblfo, region->delayVibLFO, region->freqVibLFO, f->outSampleRate);
}

return voicePlayIndex;
}

TSFDEF int tsf_bank_note_on(tsf* f, int bank, int preset_number, int key, float vel)
{
int preset_index = tsf_get_presetindex(f, bank, preset_number);
Expand Down Expand Up @@ -1805,6 +1878,19 @@ TSFDEF void tsf_note_off(tsf* f, int preset_index, int key)
}
}

/**
* Stops all voices with the given playing index. If no key provided or if -1, fallbacks to tsf_note_off
*/
TSFDEF void tsf_note_off_fast(tsf* f, int preset_index, int key, int playIndex = -1)
{
if (playIndex < 0)
tsf_note_off(f, preset_index, key);
else
for (struct tsf_voice *v = f->voices, *vEnd = v + f->voiceNum; v != vEnd; v++)
if (v->playIndex == playIndex && v->playingPreset == preset_index && v->playingKey == key && v->ampenv.segment < TSF_SEGMENT_RELEASE)
tsf_voice_end(v, f->outSampleRate);
}

TSFDEF int tsf_bank_note_off(tsf* f, int bank, int preset_number, int key)
{
int preset_index = tsf_get_presetindex(f, bank, preset_number);
Expand Down Expand Up @@ -1884,7 +1970,7 @@ static void tsf_channel_setup_voice(tsf* f, struct tsf_voice* v)
struct tsf_channel* c = &f->channels->channels[f->channels->activeChannel];
float newpan = v->region->pan + c->panOffset;
v->playingChannel = f->channels->activeChannel;
v->noteGainDB += c->gainDB;
v->noteGain *= tsf_decibelsToGain(c->gainDB);
tsf_voice_calcpitchratio(v, (c->pitchWheel == 8192 ? c->tuning : ((c->pitchWheel / 16383.0f * c->pitchRange * 2.0f) - c->pitchRange + c->tuning)), f->outSampleRate);
if (newpan <= -0.5f) { v->panFactorLeft = 1.0f; v->panFactorRight = 0.0f; }
else if (newpan >= 0.5f) { v->panFactorLeft = 0.0f; v->panFactorRight = 1.0f; }
Expand Down Expand Up @@ -1994,7 +2080,7 @@ TSFDEF void tsf_channel_set_volume(tsf* f, int channel, float volume)
if (gainDBChange == 0) return;
for (v = f->voices, vEnd = v + f->voiceNum; v != vEnd; v++)
if (v->playingChannel == channel && v->playingPreset != -1)
v->noteGainDB += gainDBChange;
v->noteGain *= tsf_decibelsToGain(gainDBChange);
c->gainDB = gainDB;
}

Expand Down

0 comments on commit d32d8b4

Please sign in to comment.