[fluid-dev] Re: Patch for channel_type, also XG drum-channel autoswitch

[jimmy]

--- On Sat, 5 Feb 2011, "Pedro Lopez-Cabanillas" <address@hidden> wrote:

> > And to the rest of you - this bank select handling
> seems to be a never 
> > ending debate, and it's not my area of expertise, so
> let me know if this 
> > change screwed something up for you.
> 
> Looks OK, and also works fine in my tests with a few MIDI
> files created for XG synths that I've found in my HD.
> 
> You are perceiving that there is debate around this matter,
> when in fact the controversy comes from many years ago, when
> the MIDI manufacturers defined those different standards
> (GM, GS and XG) and people used the synths to  create
> MIDI files following each one. Now, to render those old MIDI
> files using FluidSynth we would like that FS  follow
> each one of the conventions as closely as possible. I've
> never been a XG user, so changeset #340 included several
> TODOs and FIXMEs that this patch has finally addressed. But
> there is still one (FIXME: pending implementation of SYSEX
> midi mode changes) that somebody has to work on, and there
> may be a few issues more than bank changes involved.
> 
> Regards,
> Pedro



The Fluidsynth MSB, LSB handling code is not easy to read the first few dozen 
times for me, involving several bit masks, bit inversion, bit shift and can 
easily clear out other fields.  Perhaps we can put in some comments to help 
explain it quickly.  Please do it, keep this with the code, probably above 
fluid_channel_set_bank_msb(), or fluid_channel_set_bank_lsb():


/* The MSB, LSB calculation has 2 parts:
 *       "(oldval & ~BANKLSB_MASKVAL)" , and
 *       " | (someval << BANK_SHIFTVAL) 
 * where:
 * (oldval & ~BANKLSB_MASKVAL) means zeroes out 7-bit LSB field 
 * (oldval & ~BANKMSB_MASKVAL) means zeroes out 7-bit MSB field 
 * (oldval & ~BANK_MASKVAL) means zeroes out 14-bit field (both MSB,LSB)  
 * and:
 * " | (someval << BANK_SHIFTVAL) means add in 7-bit variable "someval" as LSB 
field
 * " | (someval << (BANK_SHIFTVAL + 7)) means add in 7-bit variable "someval" 
as MSB field
 */
/* Set bank MSB 7 bits. */
void
fluid_channel_set_bank_msb(fluid_channel_t* chan, int bankmsb)
{ . . . }


After reading such comments in the code, it is so easy and much quicker to make 
changes. Isn't it?

-----

I haven't read much on GS, so if someone who knows about GS bank handling, can 
you check and see if GS bank number can be calculated MMA-style?  It is 
basically convert the combined MSB, LSB as a decimal number, as us in Cakewalk 
.ins instrument definition files.

MSB, LSB calculation is only confusing because of the way it was "assumed" by 
programmers, especially on older devices.

I believe MSB, LSB are straight forward.  The calculation is always MMA style.  
It is basically a 14-bit unsigned interger split up as 2 7-bit unsigned 
integers (because MIDI is operating on 8-bits, but the highest bit is already 
used for other things).  First 7-bits (left to right) of that 14-bit is called 
the MSB, the rest is the LSB.  So to calculate it as a decimal number it is 
(128*MSB) + LSB.  In C, multiply by 128 is the same as bit-shift to the left 7 
places.

In ancient cultures 8-) , hardware was limited.  So, some of them only use one 
byte instead of both bytes, "assuming" the other byte was 0 and perhaps didn't 
have the hardware component for it.  For example, LSB is "optional" in older XG 
instruments.  It probably means that some of their old equipments didn't use 
LSB (memory was expensive), while their newer equipments do, and they want them 
to still work together.

Programmers, especially "reversed engineers" of those days would see only one 
byte (one CC message), never see the other CC-message, and may claim that there 
was no use at all for that other byte, just ignore it altogether in their 
programs (and/or midi files).  When in fact, optional really means it should be 
counted as 0, and not "short circuited" from program flow altogether (the way 
drum channel is assumed in Fluidsynth calculation of MSB, and especially LSB).

I guess originally MSB=128 was usable on any channels, because some (drum, or 
voice generator) hardwares probably didn't even have 16 channels.  So some midi 
hardwares just work fine by itself, but would mess up big time when hooked up 
with other midi equipments.  Midi specs came out and recommend certain things, 
including drum channel being #10.  Many people assume that is the only thing to 
do, no exception, and hard-code drum related thing to channel 10, just because 
most hardwares were made to only satisfy the GM specs and nothing more.  The 
way I see it, both GS and XG have done it right by allowing drums on any 
channel.  Although, I'm not too familiar with GS in detail.

The old "assumptions" of how GS does it, how XG does it, resulted in code that 
assume too much and cut out the full calculation.  I think both of them are 
calculated the same as MMA-style.  For example, in XG drum bank usage, 
currently only MSB is used.  But it should not be assumed that LSB will never 
be used with drum banks.  The "return if channel is drum-channel" check in both 
Fluidsynth functions should not be there, I believe.

-----

The way I see it, the check "if channel is drum-channel" in both 
fluid_channel_set_bank_msb(), and fluid_channel_set_bank_lsb() should be ripped 
out.

For example, in GM mode it is "commonly" assumed that it can use any bank 
number except bank number 128.  And that channel 10, defaults to bank number 
128 cannot be changed to any other number.

That's just wrong.  I think any channel should still be able to change to use 
bank number 128 (which has drum sounds for notes).  Of course, XG uses MSB=128 
(that is 128*128, or 16384 in MMA-style decimal number, 16383 for 0-offset), I 
believe GS uses LSB=128, that's all.  

On the other hand, I think channel 10 should still be able to use any 
banknumber, too.  Not being tied to bank number 128.

Regardless, once a channel is determined to be drum, or melodic, the only 
difference between them is the bank fall-back lookup scheme.  Melodic channels 
fall-back eventually to bank number 0.  Drum channels fall-back to bank number 
128 (XG use of MSB=120, or 126, for example).

2-3 years ago I put in the patch for melodic bank number fall-back scheme.  
Back then, if a bank could not be found, the channel went completely silent.  
Since then, for XG midi files that uses drum bank 126, I will hear Piano.  It 
doesn't sound good, or right, but at least now I know something needed to be 
changed (the midi file), or fixed (in the soft synth), rather than not knowing 
anything is there at all.

Now with drum channel auto switching (at least in XG mode), I can say that drum 
channels finally sound like drums.  If it doesn't sound good, maybe a better 
soundfont could help.  I don't have to hear Piano thumping on drum channels (on 
channel other than 10) any more.


Jimmy