Remove global state #68

jcmoyer · 2024-05-02T03:47:30Z

Hello, first of all thank you for releasing this project. The SC-55 is legendary and it is great to see such an important piece of history preserved.

The main goal of this PR is to remove global state from the emulator so that multiple instances can be created in a single process. This PR splits the emulator into a frontend (the SDL application) and a backend (the SC-55 emulator). A frontend can create and manage as many backends as it sees fit. As a side effect this also turns the backend into a library that can be linked into new frontends.

Global state is blocking a couple of issues:

Outputting MIDIs to WAV/MP3 possible? #11 - Polyphony limits on prototype midi-wav renderer
proposal to make standalone library and add multi-instance capability #12 - Proposal for a frontend/backend split with linkable backend
combine 2 emus to increase polyphony #29 - Polyphony limits; closed by this PR

In an effort to make this PR as mergeable as possible, I tried to make a minimum number of changes to existing code. New code attempts to follow existing patterns.

Most variable names were preserved
- One exception: mcu_timer had some variable shadowing (global mcu_timer_t timer + local frt_t* timer) - this was resolved by renaming the local frt_t* timer to ftimer
I tried to stick to using C features since the project seems to be doing that already. Two exceptions: references for parameters (done to avoid changing . to -> everywhere, but these can be changed to pointers on request), and two math function templates (these can also be changed, but I'd like to avoid it).
All globals that were previously initialized at file scope are now initialized in their respective <module>_Init function. For example, sw_pos is initialized to 3 in MCU_Init. They are not initialized inline because doing so would cause the compiler to generate a default constructor, which makes the type non-trivial so it cannot be memset.

That said, the changes weren't all simple and some problems took a bit more effort to solve. Other contributors might like to be aware of these points:

All global variables were moved into their respective component structures. For example, lcd_width is now a field in lcd_t.
Some structures now contain RAM/ROM inline and pointers to other components. They cannot be memset to reset their state.
TIMER_Reset wasn't used anywhere. I renamed it TIMER_Init and use it to initialize a timer and link it to its MCU.
LCD size is now only set in LCD_Init instead of two places.
LCD m_back_path was removed because moving a std::string into lcd_t would make it non-trivial. LCD_SetBackPath was replaced with LCD_LoadBack and should be called after LCD_Init.
Global lcd_init was removed. If LCD_Init succeeded you have an LCD.
LCD event handling was extracted from LCD_Update into LCD_HandleEvent. Since multiple LCDs are supported, each LCD needs to handle only events targeted at it.
The current frontend SDL application was extracted from the MCU module. main is now located in main_frontend.cpp.
The MCU now posts sample data to the frontend as it is produced via callback. This approach was chosen because it allows the frontend to decide how to handle sample data in a zero-copy manner.
A new emu_t type was introduced that is conceptually a single, complete backend instance. It contains all of the SC-55 emulator state and links the individual modules together. Romset detection and loading has been moved here since romsets need to be loaded into multiple emulator components.
The ringbuffer in the current MCU module was extracted into a new type ringbuffer_t. The frontend creates a ringbuffer for each emu instance and fills it as the frontend receives sample data. All instance buffers are then read and mixed in the SDL audio callback. Right now ringbuffer_t is specialized for signed 16-bit sample data but I would like for it to eventually become a template type to handle different types of sample data (useful for Output 32-bit float samples to preserve dynamic range and prevent clipping #59 for instance).

As an example of what this PR enables I have modified the frontend SDL application so that you can pass -instances:<n> to open n instances of the emulator. Non-sysex MIDI events are routed to their midi_channel_id % n (so if n=2, even channels go to the first emulator, and odd channels go to the second one). Sysex events are broadcast to all emulators. This is a quick and easy way to solve polyphony problems at the expense of twice the computing power.

Performance impact

Performance seems to be unchanged or marginally better. I benchmarked running the emulator as fast as possible without locking or waiting for the ringbuffer to drain for 100_000_000 iterations. These are the time measurements for 5 consecutive runs:

	master	multi-instance -instances:1	multi-instance -instances:2
1	30411 ms	29287 ms	29872 ms
2	30009 ms	29393 ms	29709 ms
3	30972 ms	28519 ms	28786 ms
4	30259 ms	29483 ms	29867 ms
5	30463 ms	29288 ms	29466 ms
average	30423 ms	29194 ms	29540 ms
%change		-4.04%	-2.90%

nikitalita · 2024-05-02T21:18:57Z

This is really cool; I've tested this with -instances:2 on midi files that had polyphony >28 such that parts would cut out or go flat on a real sc-55 mk2, and this plays those beautifully.

pachuco · 2024-05-02T21:27:18Z

Yes please, make it easier to use as lib.

johnnovak · 2024-05-04T00:00:32Z

Very nice @jcmoyer, this makes it a lot easier to turn this into a plugin that can be instantiated multiple times. Then clear separation and boundaries between the processing engine and the frontend is just good design. You should be able to run the processing engine completely "headless", without any UI, just from the command line.

johnnovak · 2024-05-04T04:11:05Z

src/emu.h

+// Should be called after loading roms
+void EMU_Reset(emu_t& emu);
+
+void EMU_SetSampleCallback(emu_t& emu, mcu_sample_callback callback, void* userdata);


EMU_SetRenderCallback and mcu_render_callback would be better.

johnnovak · 2024-05-04T04:11:24Z

src/mcu.cpp

    }
 }

-void MCU_Init(void)
+void MCU_DefaultSampleCallback(void* userdata, int* sample)


Suggest MCU_DefaultRenderCallback

johnnovak · 2024-05-04T04:15:34Z

src/main_frontend.cpp

+    }
+}
+
+void FE_ReceiveSample(void* userdata, int *sample)


Passing something like an AudioFrame struct we use in DOSBox Staging would be nicer than these raw unsafe pointers:

https://github.com/dosbox-staging/dosbox-staging/blob/main/include/audio_frame.h

Another general comment: using std::array or std::vector is zero-cost compared to C arrays but safer and there's potential to handle them in a nicer way (e.g., using iterators instead of error prone indexing and oldschool for loops).

These are all zero-cost and make for readable and more maintainable code. On MSVC on Windows you even get bounds checking for free with the standard STL containers in debug builds which is great.

Yes, I think this function should be changed (as well as MCU_PostSample) but I left it as I found it. The int* is a performance pessimization and it's not an intuitive API.

using std::array or std::vector is zero-cost compared to C arrays but safer

I totally agree with you and I would have written it that way myself, but I didn't want to introduce a bunch of C++ features since the current code is not using them. As far as I can tell the only STL type used anywhere is std::string, and std::vector made it into the rtmidi module only because it's required for the rtmidi callback.

performance pessimization and it's not an intuitive API

Haha, okay, I can see I'm among friends—we're 100% on the same page 😏 I get it, you didn't want to change too much.

johnnovak · 2024-05-04T04:19:39Z

src/main_frontend.cpp

+    else if (sample[1] < INT16_MIN)
+        sample[1] = INT16_MIN;
+    RB_Write(fe.sample_buffer, sample[0], sample[1]);
+}


IMO, you're overthinking this a bit and micro-optimising early... You could just simply pass AudioFrame structs around (or similar). Doing something like this is perfectly fine, so just using simple values:

AudioFrame RenderNextSample() { ... }

The compiler is more than capable of getting rid unnecessary copies, plus you'll get guaranteed copy elision and return value optimisation unless you're using some archaic compiler.

https://en.wikipedia.org/wiki/Copy_elision

So no need to complicate life with pointers and such (which can make things run slower in fact, because unsafe pointers can point to well anything, which makes the compiler's job a lot harder than just expressing intent by returning simple structs/values).

Also, you could simply try it and re-run your nice test suite that you used for the measurements. My guess is you won't see any performance degradation.

Nice improvement by introducing audio_frame_t 👍🏻

johnnovak · 2024-05-04T04:23:16Z

src/ringbuffer.h

+struct ringbuffer_t {
+    int16_t* samples;
+    size_t frame_count;
+    size_t sample_count;


My suggestion is to always use the terms frames, always deal with frames and frame counts, unless you absolutely need to do things at sample level. Simplifies the code a bit and makes terms less ambiguous.

In my own code, the only exception to this is using "sample rate" because that's such an established term (instead of the technically correct "frame rate").

So std::vector<AudioFrame> would make this a lot simpler, then you don't need separate count vars which can be error-prone to maintain.

I think this is a good suggestion, and it fixes some parts of the API I was kind of unhappy with. I will try to get around to simplifying this tomorrow.

johnnovak

Very nice job @jcmoyer , left a few comments. You might take them onboard, or completely ignore them 😄 I'll probably use your code for my own stuff and do the suggested refactorings anyway (and I will credit you, of course).

This PR is an important step to take @nukeykt 's excellent work to the next level and introduce proper separation of concerns and proper multi-instance support by getting rid of globals.

johnnovak · 2024-05-04T04:28:20Z

src/emu.h

+    mcu_timer_t* timer;
+    lcd_t*       lcd;
+    pcm_t*       pcm;
+};


Another general comment: could use std::unique_ptr for these, and in general it would be good to get rid of pointers and use references and standard containers instead. All that would be zero-cost, no performance impact, or even could make things faster by expressing intent to the compiler instead of attempting to outsmart it by direct pointer access (doesn't always end well).

I saw you started using emu_t& emu and similar stuff, so probably you have the same views as myself, just didn't want to blast the whole codebase apart in a single PR 😎

johnnovak · 2024-05-04T04:30:46Z

src/mcu.h

+static const int SRAM_SIZE = 0x8000;
+static const int NVRAM_SIZE = 0x8000; // JV880 only
+static const int CARDRAM_SIZE = 0x8000; // JV880 only
+static const int ROMSM_SIZE = 0x1000;


Minor: could be static constexpr uint32_t. Using uint32_t (or uint16_t) for hex addresses probably is a bit more correct.

Agreed, but I just moved these constants from elsewhere. I didn't change anything that didn't need to be.

johnnovak · 2024-05-04T04:33:58Z

src/mcu.h

+    void* callback_userdata;
+    mcu_sample_callback sample_callback;
+
+    SDL_mutex *work_thread_lock;


While you're at it, you could take it further and initialise every var. So for example:

uint16_t ad_val[4] = {}; uint8_t ad_nibble = {}; uint8_t sw_pos = {}; uint8_t io_sd = {}; submcu_t* sm = nullptr; pcm_t* pcm = nullptr; mcu_timer_t* timer = nullptr; lcd_t* LCD = nullptr;

Not initialising can lead to obscure bugs, and I wasted too much time on that, so these days I always initialise everything, it's cheap to do so and eliminated weird bugs that manifest only sometime, depending on the concrete memory layout (e.g. obscure bugs that crash your program every 10th run or something because some var you forgot to init is sometimes zero, and sometimes some random number).

Same comment for every var everywhere—always initialising is just good practice to save some headache later.

I agree, but since memset has been used throughout the codebase to clear state I don't want to do anything that would make these types non-trivial. I thought it would be best to submit a PR as more-or-less C-style C++ in case @nukeykt doesn't approve of C++ features, and if he does then these things can be pretty easily cleaned up later.

I agree, but since memset has been used throughout the codebase to clear state

Right, I actually missed that 😄 Not used to seeing memset much these days 😄

johnnovak · 2024-05-04T04:35:25Z

src/mcu.h

-void MCU_WorkThread_Lock(void);
-void MCU_WorkThread_Unlock(void);
+void MCU_WorkThread_Lock(mcu_t& mcu);
+void MCU_WorkThread_Unlock(mcu_t& mcu);


Not sure how you feel about const-correctness, but a lot of args and vars could be made const.

giulioz · 2024-05-04T11:21:59Z

src/main_frontend.cpp

+
+    int romset = ROM_SET_MK2;
+
+    frontend_t frontend;


If I'm reading this correctly, this would end up allocating a giant chunk of memory in stack, which could cause problems. I did this exact refactor in my branch (https://github.com/giulioz/Nuked-SC55/tree/audiounit_jv880) and I started to have some stack overflows in some conditions by using this method. I think using a malloc here would be a bit safer.

This structure is only 1688 bytes and you'll get 1-2mb of stack space by default on windows depending on the compiler, usually more on linux. At worst it's 0.16% of the total stack reservation. I did run into some stack overflow problems with the larger structures but those are malloc'd now and this type mostly contains pointers to those ones.

I misunderstood how the structures are created! You are actually creating the mcu and pcm data structures with a malloc, sorry :)

johnnovak · 2024-05-05T07:47:02Z

src/ringbuffer.h

+struct audio_frame_t {
+    int16_t left;
+    int16_t right;
+};


I approve 😎

Axis4s · 2024-05-13T12:38:15Z

Hello, first of all thank you for releasing this project. The SC-55 is legendary and it is great to see such an important piece of history preserved.

The main goal of this PR is to remove global state from the emulator so that multiple instances can be created in a single process. This PR splits the emulator into a frontend (the SDL application) and a backend (the SC-55 emulator). A frontend can create and manage as many backends as it sees fit. As a side effect this also turns the backend into a library that can be linked into new frontends.

Global state is blocking a couple of issues:
* [Outputting MIDIs to WAV/MP3 possible? #11](https://github.com/nukeykt/Nuked-SC55/issues/11) - Polyphony limits on prototype midi-wav renderer

* [proposal to make standalone library and add multi-instance capability #12](https://github.com/nukeykt/Nuked-SC55/issues/12) - Proposal for a frontend/backend split with linkable backend

* [combine 2 emus to increase polyphony #29](https://github.com/nukeykt/Nuked-SC55/issues/29) - Polyphony limits; closed by this PR
In an effort to make this PR as mergeable as possible, I tried to make a minimum number of changes to existing code. New code attempts to follow existing patterns.
* Most variable names were preserved
  
  * One exception: `mcu_timer` had some variable shadowing (global `mcu_timer_t timer` + local `frt_t* timer`) - this was resolved by renaming the local `frt_t* timer` to `ftimer`

* I tried to stick to using C features since the project seems to be doing that already. Two exceptions: references for parameters (done to avoid changing . to -> everywhere, but these can be changed to pointers on request), and two math function templates (these can also be changed, but I'd like to avoid it).

* All globals that were previously initialized at file scope are now initialized in their respective `<module>_Init` function. For example, `sw_pos` is initialized to 3 in `MCU_Init`. They are not initialized inline because doing so would cause the compiler to generate a default constructor, which makes the type non-trivial so it cannot be `memset`.
That said, the changes weren't all simple and some problems took a bit more effort to solve. Other contributors might like to be aware of these points:
* All global variables were moved into their respective component structures. For example, `lcd_width` is now a field in `lcd_t`.

* Some structures now contain RAM/ROM _inline_ and pointers to other components. They cannot be `memset` to reset their state.

* `TIMER_Reset` wasn't used anywhere. I renamed it `TIMER_Init` and use it to initialize a timer and link it to its MCU.

* LCD size is now only set in `LCD_Init` instead of two places.

* LCD `m_back_path` was removed because moving a `std::string` into `lcd_t` would make it non-trivial. `LCD_SetBackPath` was replaced with `LCD_LoadBack` and should be called after `LCD_Init`.

* Global `lcd_init` was removed. If `LCD_Init` succeeded you have an LCD.

* LCD event handling was extracted from `LCD_Update` into `LCD_HandleEvent`. Since multiple LCDs are supported, each LCD needs to handle only events targeted at it.

* The current frontend SDL application was extracted from the MCU module. `main` is now located in `main_frontend.cpp`.

* The MCU now posts sample data to the frontend as it is produced via callback. This approach was chosen because it allows the frontend to decide how to handle sample data in a zero-copy manner.

* A new `emu_t` type was introduced that is conceptually a single, complete backend instance. It contains all of the SC-55 emulator state and links the individual modules together. Romset detection and loading has been moved here since romsets need to be loaded into multiple emulator components.

* The ringbuffer in the current MCU module was extracted into a new type `ringbuffer_t`. The frontend creates a ringbuffer for each emu instance and fills it as the frontend receives sample data. All instance buffers are then read and mixed in the SDL audio callback. Right now `ringbuffer_t` is specialized for signed 16-bit sample data but I would like for it to eventually become a template type to handle different types of sample data (useful for [Output 32-bit float samples to preserve dynamic range and prevent clipping #59](https://github.com/nukeykt/Nuked-SC55/pull/59) for instance).
As an example of what this PR enables I have modified the frontend SDL application so that you can pass -instances:<n> to open n instances of the emulator. Non-sysex MIDI events are routed to their midi_channel_id % n (so if n=2, even channels go to the first emulator, and odd channels go to the second one). Sysex events are broadcast to all emulators. This is a quick and easy way to solve polyphony problems at the expense of twice the computing power.

Performance impact

Performance seems to be unchanged or marginally better. I benchmarked running the emulator as fast as possible without locking or waiting for the ringbuffer to drain for 100_000_000 iterations. These are the time measurements for 5 consecutive runs:
master multi-instance -instances:1 multi-instance -instances:2
1 30411 ms 29287 ms 29872 ms
2 30009 ms 29393 ms 29709 ms
3 30972 ms 28519 ms 28786 ms
4 30259 ms 29483 ms 29867 ms
5 30463 ms 29288 ms 29466 ms
average 30423 ms 29194 ms 29540 ms
%change -4.04% -2.90%

As a side effect this also turns the backend into a library that can be linked into new frontends.

Does it mean it can be more easily integrated as a VST form?

johnnovak · 2024-05-13T21:34:46Z

Does it mean it can be more easily integrated as a VST form?

Yes.

Falcosoft · 2024-05-14T16:25:12Z

Hi,
I have tried with the latest updates using VS 2019 and I have noticed 2 problems:

It does not work with SC-55 MK1 roms (and supposedly neither with any rom sets where the 2nd rom is not 512k). The problem seems to be EMU_ReadStreamUpTo() and the error is "FATAL ERROR: Failed to read the mcu ROM2".
It works this way:
std::streamsize rom2_read = EMU_ReadStreamUpTo(s_rf[1], m_mcu->rom2, m_mcu->mcu_mk1 ? ROM2_SIZE / 2 : ROM2_SIZE);
If enable_lcd = false in EMU_Options then LCD_Update() throws a null pointer exception (not with the wave renderer).
It works with this check:
if (!lcd.mcu) return;

jcmoyer · 2024-05-14T21:38:57Z

Hi @Falcosoft I think you are building the modernize branch which isn't related to this PR, but I'll take a look.

EDIT:

1 is a bug, thanks for reporting it.

As for 2, enable_lcd isn't complete yet. Right now the idea is that when a frontend sets enable_lcd it either commits to calling LCD_* functions (true), or never calling LCD_* functions (false). Eventually, I would like to have a version of the backend that doesn't depend on SDL at all, and it's not clear yet how enable_lcd will fit in. That said I think it's desirable to be able to run the main frontend headless today, so I implemented this feature keeping the above design in mind. You can simply pass --no-lcd on the command line.

Falcosoft · 2024-05-15T06:11:15Z

Hi @Falcosoft I think you are building the modernize branch which isn't related to this PR, but I'll take a look.

EDIT:

1 is a bug, thanks for reporting it.

As for 2, enable_lcd isn't complete yet. Right now the idea is that when a frontend sets enable_lcd it either commits to calling LCD_* functions (true), or never calling LCD_* functions (false). Eventually, I would like to have a version of the backend that doesn't depend on SDL at all, and it's not clear yet how enable_lcd will fit in. That said I think it's desirable to be able to run the main frontend headless today, so I implemented this feature keeping the above design in mind. You can simply pass --no-lcd on the command line.

OK, thanks!

This also reorders main() so that the mcu_t structure isn't zeroed after loading ROMs. ROM loading will probably need to be extracted from main at some point, especially for library usage.

The existing SDL application becomes the frontend. The backend consists of all the emulator modules (mcu, timer, pcm, etc). The goal here is to make the backend reusable between different frontend applications as a library.

emu is a collection of all the emulator subsystems that simplifies initialization and allows multiple emulators to be easily instantiated

Pass `-instances:<n>` to run `n` backend emulators. Up to 16 emulators are supported in a single process. Events are routed to the emulator modulo the number of instances, so if `n=2` even MIDI channels will go to the first emulator and odd MIDI channels to the second. - Rom loading code has been moved into emu module - LCDs now only handle events for their own window - MIDI module now forwards events to the frontend which in turn routes it to the correct backend emulator

This commit routes sysex messages to all emulators and fixes a bug in the win32 midi input where sysex messages would be duplicated.

- Proper error handling - Deal with all the warnings these changes introduced

It is not always zero when LCD_Init is called with memory from malloc()

This allows roms to be loaded on multiple threads simultaneously.

The midi input callback might be invoked after MIDI_Quit(). If this happens, midi_frontend is null and is not valid for FE_RouteMidi.

Missed this one, this state is now part of fe_emu_instance_t.

begin_step and wait callbacks are needlessly complicated, and controlling how often to step the MCU is really the responsibility of the frontend. These two callbacks have been removed and the logic baked into FE_RunInstance. The sample callback is fine since it provides a zero-copy way to get sample data out of the emulator.

johnnovak · 2024-05-24T23:12:57Z

@jcmoyer Please read your email 😏 (the one linked to your GitHub account)

johnnovak reviewed May 4, 2024

View reviewed changes

src/mcu.cpp

}

}

void MCU_Init(void)

void MCU_DefaultSampleCallback(void* userdata, int* sample)

Copy link

johnnovak May 4, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggest MCU_DefaultRenderCallback

johnnovak reviewed May 4, 2024

View reviewed changes

giulioz reviewed May 4, 2024

View reviewed changes

johnnovak reviewed May 5, 2024

View reviewed changes

src/ringbuffer.h

struct audio_frame_t {

int16_t left;

int16_t right;

};

Copy link

johnnovak May 5, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I approve 😎

jcmoyer mentioned this pull request May 8, 2024

Outputting MIDIs to WAV/MP3 possible? #11

Open

mmontag mentioned this pull request May 9, 2024

RIAA filter skjelten/emusc#39

Closed

jcmoyer mentioned this pull request May 11, 2024

Threading and overrun/underrun #72

Open

jcmoyer force-pushed the multi-instance branch from 26d9ba6 to cf0d60d Compare May 13, 2024 12:01

jcmoyer added 6 commits May 21, 2024 08:41

WIP: Remove global mcu

580a9da

WIP: move global memory into mcu_t

a299648

This also reorders main() so that the mcu_t structure isn't zeroed after loading ROMs. ROM loading will probably need to be extracted from main at some point, especially for library usage.

Move more mcu state into mcu_t

123ca81

Remove sm global

5ddb771

Finish removing globals from submcu and uart

d23df55

Remove romset and pcm globals

6213222

jcmoyer added 27 commits May 21, 2024 08:42

Remove mcu_timer global state

2778362

Move ga_* globals into mcu_t

3889ef7

Move mcu_button_pressed into mcu_t

c77cfe9

Move mcu_p0_data, mcu_p1_data into mcu_t

3a5a4cf

Make midi_rtmidi conform to new interface

93bf4fc

Move global lcd state into lcd_t

65630d5

Move adf_rd, analog_end_time, ssr_rd into mcu_t

2b1d521

Move operand globals into mcu_t

a89ddb1

Move window, renderer, texture into lcd_t

c7d8f25

Extract event loop from LCD update in preparation for multi-LCD

17d2429

Split into frontend and backend

94f248b

The existing SDL application becomes the frontend. The backend consists of all the emulator modules (mcu, timer, pcm, etc). The goal here is to make the backend reusable between different frontend applications as a library.

Introduce emu abstraction

e5314a9

emu is a collection of all the emulator subsystems that simplifies initialization and allows multiple emulators to be easily instantiated

Turn global sample buffer into proper ringbuffer

c65d338

Fix sysex handling

73e9824

This commit routes sysex messages to all emulators and fixes a bug in the win32 midi input where sysex messages would be duplicated.

Remove global_fe and the last of the global state

5fae2bc

Last minute cleanup

d5a6382

- Proper error handling - Deal with all the warnings these changes introduced

Remove lcd_init

6a5255e

It is not always zero when LCD_Init is called with memory from malloc()

Remove global state from rom loader

aa214ce

This allows roms to be loaded on multiple threads simultaneously.

Check EMU_Init return

109fa7f

Add license headers

4059d32

Shorten emu_backend_t to emu_t

75836b3

Fix race condition that was causing UB

51e204e

The midi input callback might be invoked after MIDI_Quit(). If this happens, midi_frontend is null and is not valid for FE_RouteMidi.

Remove work_thread_run global

6d2a6c6

Missed this one, this state is now part of fe_emu_instance_t.

Fix UAF freeing things in the wrong order

5166d2e

Work in frames instead of samples

4733550

jcmoyer force-pushed the multi-instance branch from cf0d60d to 4733550 Compare May 24, 2024 22:35

Create window after loading roms

ca10a55

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Remove global state #68

Remove global state #68

jcmoyer commented May 2, 2024

nikitalita commented May 2, 2024

pachuco commented May 2, 2024

johnnovak commented May 4, 2024 •

edited

Loading

johnnovak May 4, 2024

johnnovak May 4, 2024

johnnovak May 4, 2024 •

edited

Loading

jcmoyer May 4, 2024

johnnovak May 4, 2024

johnnovak May 4, 2024 •

edited

Loading

johnnovak May 4, 2024

johnnovak May 5, 2024

johnnovak May 4, 2024

jcmoyer May 4, 2024

johnnovak left a comment •

edited

Loading

johnnovak May 4, 2024 •

edited

Loading

johnnovak May 4, 2024 •

edited

Loading

jcmoyer May 4, 2024

johnnovak May 4, 2024 •

edited

Loading

jcmoyer May 4, 2024

johnnovak May 4, 2024

johnnovak May 4, 2024

giulioz May 4, 2024

jcmoyer May 4, 2024

giulioz May 4, 2024

johnnovak May 5, 2024

Axis4s commented May 13, 2024

Performance impact

johnnovak commented May 13, 2024

Falcosoft commented May 14, 2024

jcmoyer commented May 14, 2024 •

edited

Loading

Falcosoft commented May 15, 2024

johnnovak commented May 24, 2024 •

edited

Loading

Remove global state #68

Are you sure you want to change the base?

Remove global state #68

Conversation

jcmoyer commented May 2, 2024

Performance impact

nikitalita commented May 2, 2024

pachuco commented May 2, 2024

johnnovak commented May 4, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

johnnovak May 4, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

johnnovak May 4, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

johnnovak left a comment • edited Loading

Choose a reason for hiding this comment

johnnovak May 4, 2024 • edited Loading

Choose a reason for hiding this comment

johnnovak May 4, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

johnnovak May 4, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Axis4s commented May 13, 2024

Performance impact

johnnovak commented May 13, 2024

Falcosoft commented May 14, 2024

jcmoyer commented May 14, 2024 • edited Loading

Falcosoft commented May 15, 2024

johnnovak commented May 24, 2024 • edited Loading

johnnovak commented May 4, 2024 •

edited

Loading

johnnovak May 4, 2024 •

edited

Loading

johnnovak May 4, 2024 •

edited

Loading

johnnovak left a comment •

edited

Loading

johnnovak May 4, 2024 •

edited

Loading

johnnovak May 4, 2024 •

edited

Loading

johnnovak May 4, 2024 •

edited

Loading

jcmoyer commented May 14, 2024 •

edited

Loading

johnnovak commented May 24, 2024 •

edited

Loading