Accumulate GUS frames accross each voice instead of per-sample (#2696) #2698

kcgen · 2023-07-22T23:38:11Z

Previously when rendering frames in the GUS, we iterated one sample at a time, gathering the next panned sample from each voice into the resulting frame before moving onto the next frame (and again, gathering the next sample from all the voices).

         start
          i=0,     i=1,    i=2, .. for N frames
           |        |       |
           V        V       V
voice 1  [v1s0]   [v1s1]   ...
voice 2 +[v2s0]  +[v2s1]   ...
voice 3 +[v3s0]  +[v3s1]   ...
...
output = {summed-s0, summed-s1, summed-s2 ... }

Rendering samples from the GUS changes its state, and in the case of Duke 3D, the above approach doesn't traverse all of the available samples in all the populated voices before a state change halts one or more of the voices, resulting in essentially a truncated sequence.

This change switches back to how the original SVN and X GUS rendering loops work, which is to gather samples per-voice first, summing all of the samples across an array. This ensures that we gather all the requested samples from each voice before moving onto the next voice.

start
i=1 ->  voice 1   [v1s0]   [v1s1]   ... for N frames
i=2,    voice 2  +[v2s0]  +[v2s1]   ... for N frames
i=3,    voice 3  +[v3s0]  +[v3s1]   ... for N frames
...
output = {summed-s0, summed-s1, summed-s2 ... }

The sub-millisecond rendering prior to port and register writes is still in place, as is the use of the mixer's floating-point compressor.

Fixes #2696 ; thanks for reporting @x73rmin8r!

src/hardware/gus.cpp

johnnovak · 2023-07-23T00:03:35Z

Nice quick fix @kcgen , will test shortly with a few demos and mod players as well 👍🏻

Previously when rendering frames in the GUS, we iterated one sample at a time, gathering the next panned sample from each voice into the resulting frame before moving onto the next frame (and again, gathering the next sample from all the voices). start i=0, i=1, i=2, .. for N frames | | | V V V voice 1 [v1s0] [v1s1] ... voice 2 +[v2s0] +[v2s1] ... voice 3 +[v3s0] +[v3s1] ... ... output = {summed-s0, summed-s1, summed-s2 ... } Rendering samples from the GUS changes its state, and in the case of Duke 3D, the above approach doesn't traverse all of the available samples in all the populated voices before a state change halts one or more of the voices, resulting in essentially a truncated sequence. This change switches back to how the original SVN and X GUS rendering loops work, which is to gather samples per-voice first, summing all of the samples across an array. This ensures that we gather all the requested samples from each voice before moving onto the next voice. start i=1 -> voice 1 [v1s0] [v1s1] ... for N frames i=2, voice 2 +[v2s0] +[v2s1] ... for N frames i=3, voice 3 +[v3s0] +[v3s1] ... for N frames ... output = {summed-s0, summed-s1, summed-s2 ... } The sub-millisecond rendering prior to port and register writes is still in place, as is the use of the mixer's floating-point compressor.

johnnovak

The issue in Duke3D is now gone. I've also tried FastTracker II, CapaMod and a few demos—all good 🎉 Feel free to merge @kcgen once you've addressed those minor issues.

kcgen · 2023-07-23T01:20:29Z

The issue in Duke3D is now gone. I've also tried FastTracker II, CapaMod and a few demos—all good tada Feel free to merge @kcgen once you've addressed those minor issues.

Thanks for the bisect, review, and test feedback @johnnovak !

johnnovak · 2023-07-23T01:23:31Z

The issue in Duke3D is now gone. I've also tried FastTracker II, CapaMod and a few demos—all good tada Feel free to merge @kcgen once you've addressed those minor issues.

Thanks for the bisect, review, and test feedback @johnnovak !

Nice one! You're back and you're solving issues in a swift manner on day 1! 🚀 😎 🤘🏻

kcgen added regression We broke something 😊 audio Audio related issues or enhancements labels Jul 22, 2023

kcgen requested a review from johnnovak July 22, 2023 23:38

kcgen self-assigned this Jul 22, 2023