The idea is that we make just one OS signals handler on the C (CGO) side instead of Go and handle everything there from both, which is needed for proper termination process of multithreaded Libretro cores. But it leads to strange segfaults.
For some reason the Libretro Mupen core breaks the pseudo-random order of channel selection leading to constant skip of the done channel on ticker aka an infinite loop.
- Fixed broken image cache for the first stage RGBA frames. It was not a thread-safe one, which led to image tearing (parts of old images in multiple consecutive frames).
- 180 flip function for the OpenGL coordinate system has been moved into the rotation part.
- Optimized YUV converter.
- Optimized color converters:
- Use __restrict pointers.
- Draw image pixels with the faster bitwise operators and pointer arithmetic as 32/16bit LE numbers (may break on ARM devices like RPi).
- Pass uints for less num conversions.
- Much faster XRGB -> RGBA conversion with Go's stdlib 32bit flip.
- Wrapped RGBA images into a custom struct in order to bypass opacity tests for the standard Go png functions, which needed for the PNG file export. Before that we set RGBx opacity byte explicitly during the pixel format conversions (much slower).
- Made Libretro core shutdown more deterministic. When we run a C core separately from the main Go process we have to make sure that the C core is not doing anything in its syscall while we stopping the emulator. Basically, a blocking call may be suspended on the Go's side while the other goroutines have no knowledge of that.
- Less info level logs.
- Added recording user label.
- Enabled RTCP sender reports by default, which may help with A/V sync.
- Check onMessage webrtc handler if it's set. May crash the program if not.
- Fixed some make dirs permissions.
- Enabled console colors (since MS has finally fixed their bloody Terminal).
- Disable log in some tests.
- Updated deps.
This PR contains refactored code.
**Changelog**
- Added new net code (the communication architecture was left intact).
- All network client IDs now have custom type `network.Uid` backed by github.com/rs/xid lib.
```
The string representation of a UUID takes 32 bytes, and the new type will take just 16.
Because of Golang JSON serialization problems with omitting zero-length empty slices (it can't)
and the need to use UID values as map keys (maps don't support slices as keys),
IDs are stored as strings (for now).
```
- A whole new WebSocket client/server implementation was added, as well as a new communication layer with synchronous and async call handlers.
- WebSocket connections now support dedicated Ping/Pong frames as opposed to original ping text messages.
- Used Gorilla WebSocket library doesn't allow concurrent (simultaneous) reads and writes, so this part was handled via send channel synchronization.
- New API structures can be found in the `pkg/api` folder.
- New communication protocol can be found in the `pkg/com/*` folder.
- Updated communication protocol is based on JSON-encoded messaging through WebSocket and has the following structure:
```
Packet
[id] string — a globally unique identification tag for the packet to track it trough a chain of requests.
t uint8 — contains packet type information (i.e. INIT_PACKET, SDP_OFFER_PACKET, ...).
[p] any — contains packet data (any type).
Each packet is a text message in JSON-serialized form (WebSocket control frames obviously not).
```
```
The main principle of this protocol and the duplex data exchange is:
the one who initializes connection is called a client, and
the one who is being connected to is called a server.
With the current architecture, the coordinator is the server, the user browsers and workers are the clients.
____ ____
↓ ↑ ↑ ↓
browser ⟶ coordinator ⟵ worker
(c) (s) (c)
One of the most crucial performance vise parts of these interactions is that
all the server-initiated calls to clients should be asynchronous!
```
- In order to track synchronous calls (packets) with an asynchronous protocol, such as WebSocket, each packet may have an `id` that should be copied in all subsequent requests/responses.
- The old `sessionID` param was replaced by `id` that should be stored inside the `p` (payload) part of the packet.
- It is possible to skip the default ping check for all connected workers on every user connection and just pick the first available with the new roundRobin param in the coordinator config file `coordinator.roundRobin: true/false`.
- Added a dedicated package for the system API (pkg/api/*).
- Added structured logging system (zerolog) for better logging and cloud services integration.
- Added a visual representation of the network message exchange in logs:
```
...
01:00:01.1078 3f98 INF w → c Handshake ws://localhost:8000/wso
01:00:01.1138 994 INF c ← w Handshake localhost:8000
01:00:01.1148 994 INF c ← w Connect cid=cep.hrg
01:00:01.1158 994 DBG c connection id has been changed to cepl7obdrc3jv66kp2ug cid=cep.hrg
01:00:01.1158 3f98 INF w → c Connect cid=cep.2ug
01:00:01.1158 994 INF c New worker / addr: localhost, ...
01:00:01.1158 3f98 INF w Connected to the coordinator localhost:8000 cid=cep.2ug
01:00:02.5834 994 INF c ← u Handshake localhost:8000
01:00:02.6175 994 INF c ← u Connect cid=cep.hs0
01:00:02.6209 994 INF c Search available workers cid=cep.hs0
01:00:02.6214 994 INF c Found next free worker cid=cep.hs0
01:00:02.6220 994 INF c → u InitSession cid=cep.hs0
01:00:02.6527 994 INF c ← u WebrtcInit cid=cep.hs0
01:00:02.6527 994 INF c → w ᵇWebrtcInit cid=cep.hrg
01:00:02.6537 3f98 INF w ← c WebrtcInit cid=cep.2ug
01:00:02.6537 3f98 INF w WebRTC start cid=cep.2ug
...
```
- Replaced a monstrous Prometheus metrics lib.
- Removed spflag dependency.
- Added new `version` config file param/constant for compatibility reasons.
- Bump the minimum required version for Go to 1.18 due to use of generics.
- Opus encoder now is cached and the default config is 96Kbps, complexity 5 (was 196Kbps, 8).
- Changed the default x264 quality parameters to `crf 23 / superfast / baseline` instead of `crf 17 / veryfast / main`.
- Added a separate WebRTC logging config param `webrtc.logLevel`.
- Worker now allocates much less memory.
- Optimized and fixed RGB to YUV converter.
- `--v=5` logging cmd flag was removed and replaced with the `debug` config parameter.
**Breaking changes (migration to v3)**
- Coordinator server API changes, see web/js/api/api.js.
- Coordinator client event API changes:
- c `GAME_PLAYER_IDX_CHANGE` (string) -> `GAME_PLAYER_IDX` (number)
- c `GAME_PLAYER_IDX` -> `GAME_PLAYER_IDX_SET`
- c `MEDIA_STREAM_INITIALIZED` -> `WEBRTC_NEW_CONNECTION`
- c `MEDIA_STREAM_SDP_AVAILABLE` -> `WEBRTC_SDP_OFFER`
- c `MEDIA_STREAM_CANDIDATE_ADD` -> `WEBRTC_ICE_CANDIDATE_RECEIVED`
- c `MEDIA_STREAM_CANDIDATE_FLUSH` -> `WEBRTC_ICE_CANDIDATES_FLUSH`
- x `MEDIA_STREAM_READY` -> **removed**
- c `CONNECTION_READY` -> `WEBRTC_CONNECTION_READY`
- c `CONNECTION_CLOSED` -> `WEBRTC_CONNECTION_CLOSED`
- c `GET_SERVER_LIST` -> `WORKER_LIST_FETCHED`
- x `KEY_STATE_UPDATED` -> **removed**
- n `WEBRTC_ICE_CANDIDATE_FOUND`
- n `WEBRTC_SDP_ANSWER`
- n `MESSAGE`
- `rtcp` module renamed to `webrtc`.
- Controller state equals Libretro controller state (changed order of bits), see: web/js/input/input.js.
- Added new `coordintaor.selector` config param that changes the selection algorithm for workers. By default it will select any free worker. Set this param to `ping` for the old behavior.
- Changed the name of the `webrtc.iceServers.url` config param to `webrtc.iceServers.urls`.
Added the optional saveCompression configuration parameter which enables ZIP compression of the emulator states, reducing files sizes by the factor of 10x (most of these state files store zeroes).
We add a new option for manual worker or machine (server with multiple workers) select, depending on the new coordinator option `coordinator.debug` which by default allows machine selection.
