Details
-
Bug
-
Resolution: Fixed
-
P2: Important
-
None
-
6.4
-
None
-
i.MX 8M Plus with Vivante GC7000UL running Wayland/Weston, wayland-egl backend, Vivante EGL drivers
2 GB RAM
Display: 1280x720 LVDS
-
Linux/Wayland
Description
We have observed a performance regression from Qt5 to Qt 6 running the following example on i.MX 8M Plus with Vivante GC7000UL.
// Omit the version numbers for Qt6, otherwise use the same code import QtQuick 2.15 import QtQuick.Window 2.15 import QtQuick.Layouts 1.15 Window { id: window width: 720 height: 1280 visible: true Flickable { z: 100 anchors.fill: parent contentWidth: layout.width contentHeight: layout.implicitHeight flickableDirection: Flickable.VerticalFlick ColumnLayout { id: layout opacity: 0.5 width: parent.width anchors.top: parent.top anchors.left: parent.left spacing: 10 Repeater { model: 2500 Rectangle { height: 70 width: window.width color: "#555" Text { anchors.centerIn: parent text: model.index color: "#fff" font.pixelSize: 15 } } } } } }
Running the code on Qt5 yields good performance up to 12500 Repeater child elements.
Qt6 however starts lagging already with about 1000 Repeater child nodes.
I am aware of the existence of ListView with its virtualized nodes, but this is not the point. As proven by Qt5, it is obviously possible to efficiently handle 10000 nodes. So Qt6 should also support this.
When setting QSG_RENDERER_DEBUG=render, we can observe the following issue (output for Qt6):
-> Opaque: 0 nodes in 0 batches... -> Alpha: 5000 nodes in 2501 batches... - 0xffff5cbf2d60 [ upload] [noclip] [ alpha] [ merged] Nodes: 2500 Vertices: 10000 Indices: 15000 root: 0x0 opacity: 0.5 - 0xffff5cb62d70 [ upload] [noclip] [ alpha] [ merged] Nodes: 1 Vertices: 4 Indices: 6 root: 0x0 opacity: 0.5 - 0xffff5cc0bd00 [ upload] [noclip] [ alpha] [ merged] Nodes: 1 Vertices: 4 Indices: 6 root: 0x0 opacity: 0.5 - 0xffff5cb62750 [ upload] [noclip] [ alpha] [ merged] Nodes: 1 Vertices: 4 Indices: 6 root: 0x0 opacity: 0.5 [last line repeated 2500 times]
Whereas on Qt5, it only prints:
-> Alpha: 5000 nodes in 2 batches... - 0xffff607dea00 [ upload] [ clip] [ alpha] [ merged] Nodes: 2500 Vertices: 10000 Indices: 15000 root: 0xffff60035488 opacity: 0.5 - 0xffff607deaf0 [ upload] [ clip] [ alpha] [ merged] Nodes: 2500 Vertices: 35560 Indices: 53340 root: 0xffff60035488 opacity: 0.5
This shows that Qt6 is creating rendering batches less efficiently.
The earliest version I could reproduce the bug with is Qt 5.15.5 (when setting QSG_RHI=1). There, the critical change was introduced in qsgbatchrenderer.cpp, method checkOverlap:
where
if (!e)
was replaced by
if (!e || e->batch)
Undoing this change made the GUI run smoothly again, however in Qt6 this did not help (the change is already guarded away by the undefined macro QSGBATCHRENDERER_INVALIDATE_WEDGED_NODES)