[Qt for WebAssembly] WebGL1 support

[birkskyum]

Ticket to demonstrate utilizing Qt for WebAssembly, to run the OpenGL2 legacy branch of MapLibre Native as a WebGL1 build in the browsers.

Update

This is working now, and the findings are documented in a separate repo:

Working demo here

Build instructions and binaries here

MapLibre Native WASM overview here

[ntadej]

The code already builds with emscripten (using Qt platform). The problem is threading is generally not supported by wasm yet and the codebase uses a lot of it.

[birkskyum]

Do you have any pointers on how I can run this build? The GL JS is also really slow without the web workers, but it runs.

[ntadej]

No, Qt wasm does not run unfortunately. It crashes somewhere where threading wants to be used as Qt is nominally not built with the support for that (but maybe this has changed since).

You can build it the same as for other Qt platforms, you just use the wasm build of Qt. See https://doc.qt.io/qt-6/wasm.html

[ankurvdev]

FWIW I managed to get it working with a few patches by statically linking with Qt6-wasm built via vcpkg (with pthread and asyncify)
The only major issue was getting the QEventLoop to play well with the JS event loop (which arguably is a Qt issue)

[amirtu]

Indeed, it is possible to make it work with the above-mentioned way.

[wipfli]

@amirtu do you have a public demo of this?

[birkskyum]

so many opportunities here - let's add a compilation target for this in the repo

[birkskyum]

Qt has some serious implications in terms of licensing (LGPLv3 or commercial license). What paths can we take to end up with a more permission license?

[birkskyum]

Update: In order to get multithreading working properly, we're awaiting bugfix in Qt that'll allow the QNetworkAccessManager to make requests from within threads in wasm. https://bugreports.qt.io/browse/QTBUG-109396

[amirtu]

@amirtu do you have a public demo of this?

https://sophomore.solutions/app.html

UPD.: https://amirtu.netlify.app/mapbox-qt-wasm/app.htm

[birkskyum]

@amirtu do you have a public demo of this?

https://sophomore.solutions/app.html

Tiles are found by zooming to Almaty in Kazakhstan

[birkskyum]

Tried again with latest versions. It says that this check returns false / that vertexArray doesn't exist:

bool Context::supportsVertexArrays() const {
    return vertexArray && vertexArray->genVertexArrays && vertexArray->bindVertexArray &&
           vertexArray->deleteVertexArrays;
}

This check was removed from MapLibre Native a year ago, as part of the shift to OpenGL 3+

maplibre/maplibre-native@8ff6704#diff-25e92576a5706afc614df2db9c0a00c25bc1636aefb081dde7d2b35aceb6b069L216-L228

Would be interesting to see how this faires after the maplibre-native-qt has been reconciled with latest maplibre-native

• Update to the latest version of MapLibre Core #59

[birkskyum]

When the style is downloaded, the callback that should fire isn't firing. That's where things are now.

[andreamancuso]

Hello folks - are there plans to add support for emscripten - no Qt dependencies? I'd like to embed Maplibre Native with Dear Imgui + emscripten.

Looks like this is already achievable as part of https://github.com/maplibre/maplibre-rs ?

[birkskyum]

After two years of working on this, I finally have a POC running!

More info here

Screen.Recording.2024-07-27.at.00.34.26.mov

[ntadej]

Can you open a draft PR with your fixes/hacks?

[birkskyum]

@ntadej , what gave a visual output in the end was changing this reply-> url in the opengl2 branch of Native. After that the map kept crashing due to duplicate requests, but then changing deleteLater(); to abort(); fixed that.

void HTTPFileSource::Impl::onReplyFinished() {
    QNetworkReply* reply = qobject_cast<QNetworkReply*>(sender());
--     const QUrl& url = reply->request().url();
++     const QUrl& url = reply->url();

    auto it = m_pending.find(url);
    if (it == m_pending.end()) {
--        reply->deleteLater();
++        reply->abort();
        return;
    }

    QByteArray data = reply->readAll();
    QVector<HTTPRequest*>& requestsVector = it.value().second;

    // Cannot use the iterator to walk the requestsVector
    // because calling handleNetworkReply() might get
    // requests added to the requestsVector.
    while (!requestsVector.isEmpty()) {
        requestsVector.takeFirst()->handleNetworkReply(reply, data);
    }

    m_pending.erase(it);
--    reply->deleteLater();
++    reply->abort();
}

Apart from that I just remove everything else from the example than the map, and added a defaultPitch function, but those changes are less interesting.

[ntadej]

Let's keep this open until it properly works

[birkskyum]

Regarding performance, I find it more feasible to invest in getting a webgl2 build going using the latest OpenGL 3/4 backend from MapLibre Native Core, rather than spending a lot of time optimizing on the opengl2 legacy branch.

[birkskyum]

@ntadej Can you elaborate on what you find is not working here? Is it the 3 line changes to the opengl-2 branch in MapLibre Native? I've made a pr here to fix those with conditionals and reduce the steps needed for compilation.

Since this opengl2 work will be obsoleted by #59 / #147 rather soon, I suggest we focus on getting that updated version in the CI rather than this webgl1 version which has little-to-no practical use anyway due to resource usage.