New PIC, PIT, OPL3, Adlib Gold, DMA, many SB improvements

[codengine]

Description of Changes

New PIC, PIC, DMA, SIMD Instructions for audio processing, new OPL3, Adlib Gold and many, many other changes...

Currently missing or broken:

• Time Scaling
• Instruction-based limiting
• Breakpoint test is mostly broken (perhaps because of missing Instruction-based limiting)

Rationale behind Changes

I guess Max would be happy ;)

Suggested Testing Steps

Test games thoroughly.

[github-advanced-security]

CodeQL found more than 20 potential problems in the proposed changes. Check the Files changed tab for more details.

[maximilien-noal]

here is a test with the Adlib Gold and Dune:

https://www.youtube.com/watch?v=_cEEsm5jE7c

Dune is one of the very few games that can use not only the Adlib Gold, but also its AdLib Gold Surround Module, which is also emulated now.

Here is more information about the Adlib Gold card itself:

https://pcmidi.eu/goldlib.html

[maximilien-noal]

• it has lots of unit tests
• the code is clean
• it has lots of documentation
• methods are public, enable ASM overrides of C# code to use it.
• the music of Dune are 100% correct
• Performance is good
• PCM doesn't make the GC crazy anymore
• 80% of code changes is in the sound stack.
• The PicEventQueue is a separate class from the PIC chip, but it's still a field of the PIC chip class and instantiated by it

However, PCM still has dropouts and cutoffs. Tested with CFGCPU, in Debug mode, with fixed cycles at 3000.

Tested with commit 1b68ec0

This at least has to be fixed.

[maximilien-noal]

PCM still has cutoffs in debug mode with cfgcpu

[codengine]

PCM still has cutoffs in debug mode with cfgcpu

It should be much better now. I saw an IRQ storm in the logs because of cyclic Pump requests and very small chunks being read per pump.

One remark regarding the slowness: The time source for OPL and Adlib gold is the PIC Ticker. If the emulation runs much slower than the target cycles, it will be played back too slowly.

Another thing is an open to do that I left in code: Time Scaling is unimplemented right now, the property is there but has no effect.

[codengine]

Another issue is that the Breakpoint Tests won't work reliably and will lockup my computer... I have no clue yet what the cause is.

[maximilien-noal]

test with commit:

940cdebf7476670df1234d500ecff567906d575a

and arguments:

--CfgCpu -e "C:\Jeux\ABWFR\DUNECDVF\C\DUNECD\DNCDPRG.EXE" -a "ADG388 SBP2227"

In Debug Mode with debugger attached.

I had several issues:

• Irulan tends to desync in the intro video
• Adlib Gold music was too fast
• Allowing more than 3000 cycles with the UI (tested up to 10000) made emulation and music slower

IRULAN_DESYNC.mp4

Time multiplier value wasn't modified in the UI.

Notes:

• Cycles/ms reported often drops below 3000 cycles (down to 750, or even 180), but it's the same on master branch.

@codengine we can look at these issues together if you prefer.

[Copilot]

Pull Request Overview

This pull request introduces a major overhaul of the sound and timing subsystems, replacing significant portions of the audio and hardware emulation infrastructure. The changes include:

• New OPL3 implementation: Full Nuked-OPL3 emulator port with AdLib Gold support
• New PIT architecture: Complete rewrite of the Programmable Interval Timer with event-driven scheduling
• New PIC/DMA systems: Dual PIC controller and DMA subsystem replacements
• I/O abstraction layer: New delegate-based I/O system (IoSystem, IoReadHandler, IoWriteHandler)
• Audio processing: AdLib Gold stereo processing, surround processing, and supporting filters

Reviewed Changes

Copilot reviewed 129 out of 184 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/Spice86.Libs/Sound/Devices/NukedOpl3/*	Nuked-OPL3 emulator port including chip core, operators, channels, envelopes, LFO, I/O layer, and lookup tables
src/Spice86.Libs/Sound/Devices/AdlibGold/*	AdLib Gold device implementation with stereo/surround processors
src/Spice86.Libs/Sound/Devices/YM7128B/LICENSE	BSD 2-Clause license for YM7128B surround chip emulator
src/Spice86.Libs/Sound/Common/*	Audio frame structure and math utilities for decibel/gain conversions
src/Spice86.Core/Emulator/Devices/Timer/PitTimer.cs	Complete PIT reimplementation with deterministic event scheduling
src/Spice86.Core/Emulator/VM/PicPitCpuState.cs	Shared execution slice state for PIC/PIT scheduler integration
src/Spice86.Core/Emulator/VM/EmulationLoop.cs	Refactored emulation loop using slice-based scheduling and PIC queue
src/Spice86.Core/Emulator/IOPorts/IoSystem.cs	New delegate-based I/O port management system
src/Spice86.Core/Emulator/VM/CpuSpeedLimit/*	Simplified cycle limiting (removed NullCycleLimiter, CycleLimiterBase)
src/Spice86.Core/Emulator/InterruptHandlers/Bios/DefaultIrqHandler.cs	Default BIOS IRQ handler for unhandled hardware interrupts

Comments suppressed due to low confidence (2)

src/Spice86.Core/Emulator/VM/EmulationLoop.cs:1

• The _timeMultiplier field is declared but never read. The TODO comment at line 89 in PitTimer.cs indicates time multiplier support is not yet implemented. Either implement the feature or remove the field to avoid maintaining unused code.

namespace Spice86.Core.Emulator.VM;

src/Spice86.Core/Emulator/Devices/Timer/PitTimer.cs:90

• The _timeMultiplier field is set in SetTimeMultiplier but never applied to any timing calculations. This contradicts the documented purpose of the method. Either implement the multiplier in the delay/frequency calculations or remove the field and method.

    // TODO: Implement time Multiplier
    private double _timeMultiplier = 1.0;

[maximilien-noal]

No issues with the performance, code quality, code clarity, or the sound effects or music.

Just two requests for changes.

[maximilien-noal]

Breakpoint test is mostly broken (perhaps because of missing Instruction-based limiting) needs to be fixed before merging.

The license of the emulator needs to be compatible too.

To resolve the issue, replace the explicit foreach loop over indices with a loop over the mapped operator objects by applying .Select(index => _chip.Slots[index]) to the array of indices. Thus, instead of looping over integers and doing the mapping inside the loop, we loop directly over the mapped Opl3Operator objects. The only change required is to replace:

foreach (int index in fourOp) {
    Opl3Operator slot = _chip.Slots[index];
    ...
}

with

foreach (Opl3Operator slot in fourOp.Select(index => _chip.Slots[index])) {
    ...
}

This requires using System.Linq; for the Select extension method. Since the code snippet does not currently include that import, add it at the top among the other using statements.

Files/regions to change:

• Add using System.Linq; at the top (after the existing using block).
• Change the loop on line 102–114 to the LINQ form.

---

To fix this issue, replace the foreach (int index in twoOp) loop (lines 117–128) with a foreach over a LINQ Select, mapping each index in twoOp to the corresponding slot object (_chip.Slots[index]). This is best achieved by replacing the variable index entirely with a direct loop over the mapped sequence. The code performing the assignments can then proceed without modification to the slot variable. To support this, ensure that the required using System.Linq; directive is present at the top of the file.

The edit applies only to the loop at lines 117–128, and possibly the imports at the top if LINQ (System.Linq) is not present.

---

The best way to fix this issue is to replace the four ref parameters with a single struct (or named tuple) that encapsulates all four channel outputs. This struct can be returned from the method, eliminating the need for multiple ref outputs. Specifically:

• Define a private struct (e.g., ResampledChannels) with four short fields/properties: Channel0, Channel1, Channel2, and Channel3.
• Change the signature of Generate4ChResampledCore (lines 689-724) to return this struct, with no ref parameters.
• In the calling code (lines 680–685), call the method, capture the returned struct, and write its four values back into the span, preserving the previous assignment behavior.
• No new imports are needed; defining a simple struct suffices.
• Do not alter method logic or channel ordering.

---

[maximilien-noal]

performance is great.

only tests need to pass.