By rendering into a 32b int we don't have to saturate check in the inner
loop of the voice render, which is called multiple times per actual output
block. That check is very expensive.
In tsf_render_short make the input `short *buffer` have 2x the space as
normal to allow the full 32b quantity for each sample to accumulate.
After all voices have been rendered, then resample with saturation into
int16_t values.
About 24% faster on the Pico while still saturating properly.
* Pull untouched upstream TinySoundFont
Start from the upstream TinySoundFont by @schellingb at
https://github.com/schellingb/TinySoundFont and rearchitect the MIDI
processing.
* Convert to PROGMEM-based SoundFont structures.
This makes it much faster to access the SoundFont and also removes all the
static RAM usage. The samples, the regions, the instruments are all now
in ROM. Only the dynamic voices and related info takes up RAM. No caching.
* (Re)convert to fixed point math.
TSF upstream is built for systems with real floating point units (and
produces great sound with it!), but processors like the Pico, ESP8266,
and several models of ESP32 don't have one so emulate FP in *extremely*
slow SW. Use fixed point (integer) representations for the inner loop
and wherever possible to avoid `float` math.
For processors like the Pico 2 or original ESP32 with hardware FPUs this
still increases performance since integer is still much faster.
* Redo the example and add default SF2s
Rewrite the MIDI examples to use PROGMEM for both SF2 and MID file.
It simplifies life for the casual user and me.
* Breaking change
The MIDI object now has a `midi::SetSoundFont(tsf *)` instead of a
`midi::SetSoundfont(AudioFileSource *)`. It no longer supports reading
SF2 from LittleFS or SD.
Move to Opus 1.5.1
Remove OpusFile and OGG libraries because we parse the OGG file in
the object directly with way less memory and overhead.
Update example for LittleFS and RP2040
Fixes#372