GPU subdivision: Bad performance with viewport statistics #96434

System Information
Operating system: Windows-10-10.0.19044-SP0 64 Bits
Graphics card: NVIDIA GeForce GTX 1070/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 511.23

Blender Version
Broken: version: 3.2.0 Alpha, branch: master, commit date: 2022-03-14 00:08, hash: b959f603da.
Also broken in 3.1.0 Stable release.
Worked: None i can find.

I'm getting bad perfomance when using Modifier stacks and GPU Subdivision are enabled.
the video are recorded on 3.1 stable release, issue also occurred on 3.2 alpha.
Repro scene are attached.
ModifierStack_GPUSubdiv_BadPerformance.blend
blender_ksRjvVXVIX.mp4

Exact steps for others to reproduce the error

Open the scene file and try to change the Bend Angle on Simple Deform modifier.
Now Toggle the GPU Subdiv settings and repeat step one.
There's should be a performance difference when GPU Subdiv are enabled.

or just watch the attached video of a flailing magazine

Added subscriber: @reanimate

#97182 was marked as duplicate of this issue

Added subscribers: @kevindietrich, @mano-wii

Changed status from 'Needs Triage' to: 'Needs Developer To Reproduce'

I can confirm the delay in the example presented in the file.
The GPU subdivision update is 1/5 of the CPU subdivision performance.
@kevindietrich, any idea why this is happening?

Operating system: Windows-10-10.0.22000-SP0 64 Bits
Graphics card: Radeon (TM) RX 480 Graphics ATI Technologies Inc. 4.5.14802 Core Profile/Debug Context 22.2.3 30.0.14029.5006

This is because the viewport statistics require a the final mesh for the number of triangles/vertices/etc. so the subdivision is evaluated twice: on the CPU and on the GPU. I'm not sure yet how we can improve the situation, maybe by caching the stats when evaluating on the GPU.

Changed status from 'Needs Developer To Reproduce' to: 'Confirmed'

In #96434#1326519, @kevindietrich wrote:
This is because the viewport statistics require a the final mesh for the number of triangles/vertices/etc. so the subdivision is evaluated twice: on the CPU and on the GPU. I'm not sure yet how we can improve the situation, maybe by caching the stats when evaluating on the GPU.

Oooh interesting, turning off statistics did bring back the performance.
Hmm how about delaying the update of the statistics? ie on mouse click up event?

In #96434#1328333, @reanimate wrote:
Oooh interesting, turning off statistics did bring back the performance.
Hmm how about delaying the update of the statistics? ie on mouse click up event?

The statistics update is part of the 3D view draw code, they will be refreshed on each redraw event, it is not really possible to discriminate those events and know which one to delay statistics, and how the delay should be defined.

In #96434#1328661, @kevindietrich wrote:

In #96434#1328333, @reanimate wrote:
Oooh interesting, turning off statistics did bring back the performance.
Hmm how about delaying the update of the statistics? ie on mouse click up event?

The statistics update is part of the 3D view draw code, they will be refreshed on each redraw event, it is not really possible to discriminate those events and know which one to delay statistics, and how the delay should be defined.

I see, this is kinda problematic. Thanks for the information

Added subscriber: @Smjert

Added subscriber: @lichtwerk

Talked to @kevindietrich , setting to High prio (since it is a bug in a new feature according to the playbook

Added subscriber: @brecht

Could we compute the stats without actually performing the subdivision? Or even just provide a rough estimate? I don't think the exact number is even that important, some guess based on the subdivision level could be enough.

Actually we can just get the counters from the draw code's subdivision grid traversal, they are cached there anyway.

This issue was referenced by 163f6a17e4

Changed status from 'Confirmed' to: 'Resolved'

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

GPU subdivision: Bad performance with viewport statistics #96434